1/*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright 1998, 2000 Marshall Kirk McKusick. 5 * Copyright 2009, 2010 Jeffrey W. Roberson <jeff@FreeBSD.org> 6 * All rights reserved. 7 * 8 * The soft updates code is derived from the appendix of a University 9 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt, 10 * "Soft Updates: A Solution to the Metadata Update Problem in File 11 * Systems", CSE-TR-254-95, August 1995). 12 * 13 * Further information about soft updates can be obtained from: 14 * 15 * Marshall Kirk McKusick http://www.mckusick.com/softdep/ 16 * 1614 Oxford Street mckusick@mckusick.com 17 * Berkeley, CA 94709-1608 +1-510-843-9542 18 * USA 19 * 20 * Redistribution and use in source and binary forms, with or without 21 * modification, are permitted provided that the following conditions 22 * are met: 23 * 24 * 1. Redistributions of source code must retain the above copyright 25 * notice, this list of conditions and the following disclaimer. 26 * 2. Redistributions in binary form must reproduce the above copyright 27 * notice, this list of conditions and the following disclaimer in the 28 * documentation and/or other materials provided with the distribution. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE AUTHORS ``AS IS'' AND ANY EXPRESS OR 31 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 32 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 33 * IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY DIRECT, INDIRECT, 34 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, 35 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS 36 * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND 37 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR 38 * TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE 39 * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 40 * 41 * from: @(#)ffs_softdep.c 9.59 (McKusick) 6/21/00 42 */ 43 44#include <sys/cdefs.h> 45__FBSDID("$FreeBSD$"); 46 47#include "opt_ffs.h" 48#include "opt_quota.h" 49#include "opt_ddb.h" 50 51#include <sys/param.h> 52#include <sys/kernel.h> 53#include <sys/systm.h> 54#include <sys/bio.h> 55#include <sys/buf.h> 56#include <sys/kdb.h> 57#include <sys/kthread.h> 58#include <sys/ktr.h> 59#include <sys/limits.h> 60#include <sys/lock.h> 61#include <sys/malloc.h> 62#include <sys/mount.h> 63#include <sys/mutex.h> 64#include <sys/namei.h> 65#include <sys/priv.h> 66#include <sys/proc.h> 67#include <sys/racct.h> 68#include <sys/rwlock.h> 69#include <sys/stat.h> 70#include <sys/sysctl.h> 71#include <sys/syslog.h> 72#include <sys/vnode.h> 73#include <sys/conf.h> 74 75#include <ufs/ufs/dir.h> 76#include <ufs/ufs/extattr.h> 77#include <ufs/ufs/quota.h> 78#include <ufs/ufs/inode.h> 79#include <ufs/ufs/ufsmount.h> 80#include <ufs/ffs/fs.h> 81#include <ufs/ffs/softdep.h> 82#include <ufs/ffs/ffs_extern.h> 83#include <ufs/ufs/ufs_extern.h> 84 85#include <vm/vm.h> 86#include <vm/vm_extern.h> 87#include <vm/vm_object.h> 88 89#include <geom/geom.h> 90#include <geom/geom_vfs.h> 91 92#include <ddb/ddb.h> 93 94#define KTR_SUJ 0 /* Define to KTR_SPARE. */ 95 96#ifndef SOFTUPDATES 97 98int 99softdep_flushfiles(oldmnt, flags, td) 100 struct mount *oldmnt; 101 int flags; 102 struct thread *td; 103{ 104 105 panic("softdep_flushfiles called"); 106} 107 108int 109softdep_mount(devvp, mp, fs, cred) 110 struct vnode *devvp; 111 struct mount *mp; 112 struct fs *fs; 113 struct ucred *cred; 114{ 115 116 return (0); 117} 118 119void 120softdep_initialize() 121{ 122 123 return; 124} 125 126void 127softdep_uninitialize() 128{ 129 130 return; 131} 132 133void 134softdep_unmount(mp) 135 struct mount *mp; 136{ 137 138 panic("softdep_unmount called"); 139} 140 141void 142softdep_setup_sbupdate(ump, fs, bp) 143 struct ufsmount *ump; 144 struct fs *fs; 145 struct buf *bp; 146{ 147 148 panic("softdep_setup_sbupdate called"); 149} 150 151void 152softdep_setup_inomapdep(bp, ip, newinum, mode) 153 struct buf *bp; 154 struct inode *ip; 155 ino_t newinum; 156 int mode; 157{ 158 159 panic("softdep_setup_inomapdep called"); 160} 161 162void 163softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 164 struct buf *bp; 165 struct mount *mp; 166 ufs2_daddr_t newblkno; 167 int frags; 168 int oldfrags; 169{ 170 171 panic("softdep_setup_blkmapdep called"); 172} 173 174void 175softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 176 struct inode *ip; 177 ufs_lbn_t lbn; 178 ufs2_daddr_t newblkno; 179 ufs2_daddr_t oldblkno; 180 long newsize; 181 long oldsize; 182 struct buf *bp; 183{ 184 185 panic("softdep_setup_allocdirect called"); 186} 187 188void 189softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 190 struct inode *ip; 191 ufs_lbn_t lbn; 192 ufs2_daddr_t newblkno; 193 ufs2_daddr_t oldblkno; 194 long newsize; 195 long oldsize; 196 struct buf *bp; 197{ 198 199 panic("softdep_setup_allocext called"); 200} 201 202void 203softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 204 struct inode *ip; 205 ufs_lbn_t lbn; 206 struct buf *bp; 207 int ptrno; 208 ufs2_daddr_t newblkno; 209 ufs2_daddr_t oldblkno; 210 struct buf *nbp; 211{ 212 213 panic("softdep_setup_allocindir_page called"); 214} 215 216void 217softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 218 struct buf *nbp; 219 struct inode *ip; 220 struct buf *bp; 221 int ptrno; 222 ufs2_daddr_t newblkno; 223{ 224 225 panic("softdep_setup_allocindir_meta called"); 226} 227 228void 229softdep_journal_freeblocks(ip, cred, length, flags) 230 struct inode *ip; 231 struct ucred *cred; 232 off_t length; 233 int flags; 234{ 235 236 panic("softdep_journal_freeblocks called"); 237} 238 239void 240softdep_journal_fsync(ip) 241 struct inode *ip; 242{ 243 244 panic("softdep_journal_fsync called"); 245} 246 247void 248softdep_setup_freeblocks(ip, length, flags) 249 struct inode *ip; 250 off_t length; 251 int flags; 252{ 253 254 panic("softdep_setup_freeblocks called"); 255} 256 257void 258softdep_freefile(pvp, ino, mode) 259 struct vnode *pvp; 260 ino_t ino; 261 int mode; 262{ 263 264 panic("softdep_freefile called"); 265} 266 267int 268softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 269 struct buf *bp; 270 struct inode *dp; 271 off_t diroffset; 272 ino_t newinum; 273 struct buf *newdirbp; 274 int isnewblk; 275{ 276 277 panic("softdep_setup_directory_add called"); 278} 279 280void 281softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 282 struct buf *bp; 283 struct inode *dp; 284 caddr_t base; 285 caddr_t oldloc; 286 caddr_t newloc; 287 int entrysize; 288{ 289 290 panic("softdep_change_directoryentry_offset called"); 291} 292 293void 294softdep_setup_remove(bp, dp, ip, isrmdir) 295 struct buf *bp; 296 struct inode *dp; 297 struct inode *ip; 298 int isrmdir; 299{ 300 301 panic("softdep_setup_remove called"); 302} 303 304void 305softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 306 struct buf *bp; 307 struct inode *dp; 308 struct inode *ip; 309 ino_t newinum; 310 int isrmdir; 311{ 312 313 panic("softdep_setup_directory_change called"); 314} 315 316void 317softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 318 struct mount *mp; 319 struct buf *bp; 320 ufs2_daddr_t blkno; 321 int frags; 322 struct workhead *wkhd; 323{ 324 325 panic("%s called", __FUNCTION__); 326} 327 328void 329softdep_setup_inofree(mp, bp, ino, wkhd) 330 struct mount *mp; 331 struct buf *bp; 332 ino_t ino; 333 struct workhead *wkhd; 334{ 335 336 panic("%s called", __FUNCTION__); 337} 338 339void 340softdep_setup_unlink(dp, ip) 341 struct inode *dp; 342 struct inode *ip; 343{ 344 345 panic("%s called", __FUNCTION__); 346} 347 348void 349softdep_setup_link(dp, ip) 350 struct inode *dp; 351 struct inode *ip; 352{ 353 354 panic("%s called", __FUNCTION__); 355} 356 357void 358softdep_revert_link(dp, ip) 359 struct inode *dp; 360 struct inode *ip; 361{ 362 363 panic("%s called", __FUNCTION__); 364} 365 366void 367softdep_setup_rmdir(dp, ip) 368 struct inode *dp; 369 struct inode *ip; 370{ 371 372 panic("%s called", __FUNCTION__); 373} 374 375void 376softdep_revert_rmdir(dp, ip) 377 struct inode *dp; 378 struct inode *ip; 379{ 380 381 panic("%s called", __FUNCTION__); 382} 383 384void 385softdep_setup_create(dp, ip) 386 struct inode *dp; 387 struct inode *ip; 388{ 389 390 panic("%s called", __FUNCTION__); 391} 392 393void 394softdep_revert_create(dp, ip) 395 struct inode *dp; 396 struct inode *ip; 397{ 398 399 panic("%s called", __FUNCTION__); 400} 401 402void 403softdep_setup_mkdir(dp, ip) 404 struct inode *dp; 405 struct inode *ip; 406{ 407 408 panic("%s called", __FUNCTION__); 409} 410 411void 412softdep_revert_mkdir(dp, ip) 413 struct inode *dp; 414 struct inode *ip; 415{ 416 417 panic("%s called", __FUNCTION__); 418} 419 420void 421softdep_setup_dotdot_link(dp, ip) 422 struct inode *dp; 423 struct inode *ip; 424{ 425 426 panic("%s called", __FUNCTION__); 427} 428 429int 430softdep_prealloc(vp, waitok) 431 struct vnode *vp; 432 int waitok; 433{ 434 435 panic("%s called", __FUNCTION__); 436} 437 438int 439softdep_journal_lookup(mp, vpp) 440 struct mount *mp; 441 struct vnode **vpp; 442{ 443 444 return (ENOENT); 445} 446 447void 448softdep_change_linkcnt(ip) 449 struct inode *ip; 450{ 451 452 panic("softdep_change_linkcnt called"); 453} 454 455void 456softdep_load_inodeblock(ip) 457 struct inode *ip; 458{ 459 460 panic("softdep_load_inodeblock called"); 461} 462 463void 464softdep_update_inodeblock(ip, bp, waitfor) 465 struct inode *ip; 466 struct buf *bp; 467 int waitfor; 468{ 469 470 panic("softdep_update_inodeblock called"); 471} 472 473int 474softdep_fsync(vp) 475 struct vnode *vp; /* the "in_core" copy of the inode */ 476{ 477 478 return (0); 479} 480 481void 482softdep_fsync_mountdev(vp) 483 struct vnode *vp; 484{ 485 486 return; 487} 488 489int 490softdep_flushworklist(oldmnt, countp, td) 491 struct mount *oldmnt; 492 int *countp; 493 struct thread *td; 494{ 495 496 *countp = 0; 497 return (0); 498} 499 500int 501softdep_sync_metadata(struct vnode *vp) 502{ 503 504 panic("softdep_sync_metadata called"); 505} 506 507int 508softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 509{ 510 511 panic("softdep_sync_buf called"); 512} 513 514int 515softdep_slowdown(vp) 516 struct vnode *vp; 517{ 518 519 panic("softdep_slowdown called"); 520} 521 522int 523softdep_request_cleanup(fs, vp, cred, resource) 524 struct fs *fs; 525 struct vnode *vp; 526 struct ucred *cred; 527 int resource; 528{ 529 530 return (0); 531} 532 533int 534softdep_check_suspend(struct mount *mp, 535 struct vnode *devvp, 536 int softdep_depcnt, 537 int softdep_accdepcnt, 538 int secondary_writes, 539 int secondary_accwrites) 540{ 541 struct bufobj *bo; 542 int error; 543 544 (void) softdep_depcnt, 545 (void) softdep_accdepcnt; 546 547 bo = &devvp->v_bufobj; 548 ASSERT_BO_WLOCKED(bo); 549 550 MNT_ILOCK(mp); 551 while (mp->mnt_secondary_writes != 0) { 552 BO_UNLOCK(bo); 553 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), 554 (PUSER - 1) | PDROP, "secwr", 0); 555 BO_LOCK(bo); 556 MNT_ILOCK(mp); 557 } 558 559 /* 560 * Reasons for needing more work before suspend: 561 * - Dirty buffers on devvp. 562 * - Secondary writes occurred after start of vnode sync loop 563 */ 564 error = 0; 565 if (bo->bo_numoutput > 0 || 566 bo->bo_dirty.bv_cnt > 0 || 567 secondary_writes != 0 || 568 mp->mnt_secondary_writes != 0 || 569 secondary_accwrites != mp->mnt_secondary_accwrites) 570 error = EAGAIN; 571 BO_UNLOCK(bo); 572 return (error); 573} 574 575void 576softdep_get_depcounts(struct mount *mp, 577 int *softdepactivep, 578 int *softdepactiveaccp) 579{ 580 (void) mp; 581 *softdepactivep = 0; 582 *softdepactiveaccp = 0; 583} 584 585void 586softdep_buf_append(bp, wkhd) 587 struct buf *bp; 588 struct workhead *wkhd; 589{ 590 591 panic("softdep_buf_appendwork called"); 592} 593 594void 595softdep_inode_append(ip, cred, wkhd) 596 struct inode *ip; 597 struct ucred *cred; 598 struct workhead *wkhd; 599{ 600 601 panic("softdep_inode_appendwork called"); 602} 603 604void 605softdep_freework(wkhd) 606 struct workhead *wkhd; 607{ 608 609 panic("softdep_freework called"); 610} 611 612int 613softdep_prerename(fdvp, fvp, tdvp, tvp) 614 struct vnode *fdvp; 615 struct vnode *fvp; 616 struct vnode *tdvp; 617 struct vnode *tvp; 618{ 619 620 panic("softdep_prerename called"); 621} 622 623int 624softdep_prelink(dvp, vp) 625 struct vnode *dvp; 626 struct vnode *vp; 627{ 628 629 panic("softdep_prelink called"); 630} 631 632#else 633 634FEATURE(softupdates, "FFS soft-updates support"); 635 636static SYSCTL_NODE(_debug, OID_AUTO, softdep, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 637 "soft updates stats"); 638static SYSCTL_NODE(_debug_softdep, OID_AUTO, total, 639 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 640 "total dependencies allocated"); 641static SYSCTL_NODE(_debug_softdep, OID_AUTO, highuse, 642 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 643 "high use dependencies allocated"); 644static SYSCTL_NODE(_debug_softdep, OID_AUTO, current, 645 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 646 "current dependencies allocated"); 647static SYSCTL_NODE(_debug_softdep, OID_AUTO, write, 648 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 649 "current dependencies written"); 650 651unsigned long dep_current[D_LAST + 1]; 652unsigned long dep_highuse[D_LAST + 1]; 653unsigned long dep_total[D_LAST + 1]; 654unsigned long dep_write[D_LAST + 1]; 655 656#define SOFTDEP_TYPE(type, str, long) \ 657 static MALLOC_DEFINE(M_ ## type, #str, long); \ 658 SYSCTL_ULONG(_debug_softdep_total, OID_AUTO, str, CTLFLAG_RD, \ 659 &dep_total[D_ ## type], 0, ""); \ 660 SYSCTL_ULONG(_debug_softdep_current, OID_AUTO, str, CTLFLAG_RD, \ 661 &dep_current[D_ ## type], 0, ""); \ 662 SYSCTL_ULONG(_debug_softdep_highuse, OID_AUTO, str, CTLFLAG_RD, \ 663 &dep_highuse[D_ ## type], 0, ""); \ 664 SYSCTL_ULONG(_debug_softdep_write, OID_AUTO, str, CTLFLAG_RD, \ 665 &dep_write[D_ ## type], 0, ""); 666 667SOFTDEP_TYPE(PAGEDEP, pagedep, "File page dependencies"); 668SOFTDEP_TYPE(INODEDEP, inodedep, "Inode dependencies"); 669SOFTDEP_TYPE(BMSAFEMAP, bmsafemap, 670 "Block or frag allocated from cyl group map"); 671SOFTDEP_TYPE(NEWBLK, newblk, "New block or frag allocation dependency"); 672SOFTDEP_TYPE(ALLOCDIRECT, allocdirect, "Block or frag dependency for an inode"); 673SOFTDEP_TYPE(INDIRDEP, indirdep, "Indirect block dependencies"); 674SOFTDEP_TYPE(ALLOCINDIR, allocindir, "Block dependency for an indirect block"); 675SOFTDEP_TYPE(FREEFRAG, freefrag, "Previously used frag for an inode"); 676SOFTDEP_TYPE(FREEBLKS, freeblks, "Blocks freed from an inode"); 677SOFTDEP_TYPE(FREEFILE, freefile, "Inode deallocated"); 678SOFTDEP_TYPE(DIRADD, diradd, "New directory entry"); 679SOFTDEP_TYPE(MKDIR, mkdir, "New directory"); 680SOFTDEP_TYPE(DIRREM, dirrem, "Directory entry deleted"); 681SOFTDEP_TYPE(NEWDIRBLK, newdirblk, "Unclaimed new directory block"); 682SOFTDEP_TYPE(FREEWORK, freework, "free an inode block"); 683SOFTDEP_TYPE(FREEDEP, freedep, "track a block free"); 684SOFTDEP_TYPE(JADDREF, jaddref, "Journal inode ref add"); 685SOFTDEP_TYPE(JREMREF, jremref, "Journal inode ref remove"); 686SOFTDEP_TYPE(JMVREF, jmvref, "Journal inode ref move"); 687SOFTDEP_TYPE(JNEWBLK, jnewblk, "Journal new block"); 688SOFTDEP_TYPE(JFREEBLK, jfreeblk, "Journal free block"); 689SOFTDEP_TYPE(JFREEFRAG, jfreefrag, "Journal free frag"); 690SOFTDEP_TYPE(JSEG, jseg, "Journal segment"); 691SOFTDEP_TYPE(JSEGDEP, jsegdep, "Journal segment complete"); 692SOFTDEP_TYPE(SBDEP, sbdep, "Superblock write dependency"); 693SOFTDEP_TYPE(JTRUNC, jtrunc, "Journal inode truncation"); 694SOFTDEP_TYPE(JFSYNC, jfsync, "Journal fsync complete"); 695 696static MALLOC_DEFINE(M_SENTINEL, "sentinel", "Worklist sentinel"); 697 698static MALLOC_DEFINE(M_SAVEDINO, "savedino", "Saved inodes"); 699static MALLOC_DEFINE(M_JBLOCKS, "jblocks", "Journal block locations"); 700static MALLOC_DEFINE(M_MOUNTDATA, "softdep", "Softdep per-mount data"); 701 702#define M_SOFTDEP_FLAGS (M_WAITOK) 703 704/* 705 * translate from workitem type to memory type 706 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 707 */ 708static struct malloc_type *memtype[] = { 709 NULL, 710 M_PAGEDEP, 711 M_INODEDEP, 712 M_BMSAFEMAP, 713 M_NEWBLK, 714 M_ALLOCDIRECT, 715 M_INDIRDEP, 716 M_ALLOCINDIR, 717 M_FREEFRAG, 718 M_FREEBLKS, 719 M_FREEFILE, 720 M_DIRADD, 721 M_MKDIR, 722 M_DIRREM, 723 M_NEWDIRBLK, 724 M_FREEWORK, 725 M_FREEDEP, 726 M_JADDREF, 727 M_JREMREF, 728 M_JMVREF, 729 M_JNEWBLK, 730 M_JFREEBLK, 731 M_JFREEFRAG, 732 M_JSEG, 733 M_JSEGDEP, 734 M_SBDEP, 735 M_JTRUNC, 736 M_JFSYNC, 737 M_SENTINEL 738}; 739 740#define DtoM(type) (memtype[type]) 741 742/* 743 * Names of malloc types. 744 */ 745#define TYPENAME(type) \ 746 ((unsigned)(type) <= D_LAST && (unsigned)(type) >= D_FIRST ? \ 747 memtype[type]->ks_shortdesc : "???") 748/* 749 * End system adaptation definitions. 750 */ 751 752#define DOTDOT_OFFSET offsetof(struct dirtemplate, dotdot_ino) 753#define DOT_OFFSET offsetof(struct dirtemplate, dot_ino) 754 755/* 756 * Internal function prototypes. 757 */ 758static void check_clear_deps(struct mount *); 759static void softdep_error(char *, int); 760static int softdep_prerename_vnode(struct ufsmount *, struct vnode *); 761static int softdep_process_worklist(struct mount *, int); 762static int softdep_waitidle(struct mount *, int); 763static void drain_output(struct vnode *); 764static struct buf *getdirtybuf(struct buf *, struct rwlock *, int); 765static int check_inodedep_free(struct inodedep *); 766static void clear_remove(struct mount *); 767static void clear_inodedeps(struct mount *); 768static void unlinked_inodedep(struct mount *, struct inodedep *); 769static void clear_unlinked_inodedep(struct inodedep *); 770static struct inodedep *first_unlinked_inodedep(struct ufsmount *); 771static int flush_pagedep_deps(struct vnode *, struct mount *, 772 struct diraddhd *, struct buf *); 773static int free_pagedep(struct pagedep *); 774static int flush_newblk_dep(struct vnode *, struct mount *, ufs_lbn_t); 775static int flush_inodedep_deps(struct vnode *, struct mount *, ino_t); 776static int flush_deplist(struct allocdirectlst *, int, int *); 777static int sync_cgs(struct mount *, int); 778static int handle_written_filepage(struct pagedep *, struct buf *, int); 779static int handle_written_sbdep(struct sbdep *, struct buf *); 780static void initiate_write_sbdep(struct sbdep *); 781static void diradd_inode_written(struct diradd *, struct inodedep *); 782static int handle_written_indirdep(struct indirdep *, struct buf *, 783 struct buf**, int); 784static int handle_written_inodeblock(struct inodedep *, struct buf *, int); 785static int jnewblk_rollforward(struct jnewblk *, struct fs *, struct cg *, 786 uint8_t *); 787static int handle_written_bmsafemap(struct bmsafemap *, struct buf *, int); 788static void handle_written_jaddref(struct jaddref *); 789static void handle_written_jremref(struct jremref *); 790static void handle_written_jseg(struct jseg *, struct buf *); 791static void handle_written_jnewblk(struct jnewblk *); 792static void handle_written_jblkdep(struct jblkdep *); 793static void handle_written_jfreefrag(struct jfreefrag *); 794static void complete_jseg(struct jseg *); 795static void complete_jsegs(struct jseg *); 796static void jseg_write(struct ufsmount *ump, struct jseg *, uint8_t *); 797static void jaddref_write(struct jaddref *, struct jseg *, uint8_t *); 798static void jremref_write(struct jremref *, struct jseg *, uint8_t *); 799static void jmvref_write(struct jmvref *, struct jseg *, uint8_t *); 800static void jtrunc_write(struct jtrunc *, struct jseg *, uint8_t *); 801static void jfsync_write(struct jfsync *, struct jseg *, uint8_t *data); 802static void jnewblk_write(struct jnewblk *, struct jseg *, uint8_t *); 803static void jfreeblk_write(struct jfreeblk *, struct jseg *, uint8_t *); 804static void jfreefrag_write(struct jfreefrag *, struct jseg *, uint8_t *); 805static inline void inoref_write(struct inoref *, struct jseg *, 806 struct jrefrec *); 807static void handle_allocdirect_partdone(struct allocdirect *, 808 struct workhead *); 809static struct jnewblk *cancel_newblk(struct newblk *, struct worklist *, 810 struct workhead *); 811static void indirdep_complete(struct indirdep *); 812static int indirblk_lookup(struct mount *, ufs2_daddr_t); 813static void indirblk_insert(struct freework *); 814static void indirblk_remove(struct freework *); 815static void handle_allocindir_partdone(struct allocindir *); 816static void initiate_write_filepage(struct pagedep *, struct buf *); 817static void initiate_write_indirdep(struct indirdep*, struct buf *); 818static void handle_written_mkdir(struct mkdir *, int); 819static int jnewblk_rollback(struct jnewblk *, struct fs *, struct cg *, 820 uint8_t *); 821static void initiate_write_bmsafemap(struct bmsafemap *, struct buf *); 822static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *); 823static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *); 824static void handle_workitem_freefile(struct freefile *); 825static int handle_workitem_remove(struct dirrem *, int); 826static struct dirrem *newdirrem(struct buf *, struct inode *, 827 struct inode *, int, struct dirrem **); 828static struct indirdep *indirdep_lookup(struct mount *, struct inode *, 829 struct buf *); 830static void cancel_indirdep(struct indirdep *, struct buf *, 831 struct freeblks *); 832static void free_indirdep(struct indirdep *); 833static void free_diradd(struct diradd *, struct workhead *); 834static void merge_diradd(struct inodedep *, struct diradd *); 835static void complete_diradd(struct diradd *); 836static struct diradd *diradd_lookup(struct pagedep *, int); 837static struct jremref *cancel_diradd_dotdot(struct inode *, struct dirrem *, 838 struct jremref *); 839static struct jremref *cancel_mkdir_dotdot(struct inode *, struct dirrem *, 840 struct jremref *); 841static void cancel_diradd(struct diradd *, struct dirrem *, struct jremref *, 842 struct jremref *, struct jremref *); 843static void dirrem_journal(struct dirrem *, struct jremref *, struct jremref *, 844 struct jremref *); 845static void cancel_allocindir(struct allocindir *, struct buf *bp, 846 struct freeblks *, int); 847static int setup_trunc_indir(struct freeblks *, struct inode *, 848 ufs_lbn_t, ufs_lbn_t, ufs2_daddr_t); 849static void complete_trunc_indir(struct freework *); 850static void trunc_indirdep(struct indirdep *, struct freeblks *, struct buf *, 851 int); 852static void complete_mkdir(struct mkdir *); 853static void free_newdirblk(struct newdirblk *); 854static void free_jremref(struct jremref *); 855static void free_jaddref(struct jaddref *); 856static void free_jsegdep(struct jsegdep *); 857static void free_jsegs(struct jblocks *); 858static void rele_jseg(struct jseg *); 859static void free_jseg(struct jseg *, struct jblocks *); 860static void free_jnewblk(struct jnewblk *); 861static void free_jblkdep(struct jblkdep *); 862static void free_jfreefrag(struct jfreefrag *); 863static void free_freedep(struct freedep *); 864static void journal_jremref(struct dirrem *, struct jremref *, 865 struct inodedep *); 866static void cancel_jnewblk(struct jnewblk *, struct workhead *); 867static int cancel_jaddref(struct jaddref *, struct inodedep *, 868 struct workhead *); 869static void cancel_jfreefrag(struct jfreefrag *); 870static inline void setup_freedirect(struct freeblks *, struct inode *, 871 int, int); 872static inline void setup_freeext(struct freeblks *, struct inode *, int, int); 873static inline void setup_freeindir(struct freeblks *, struct inode *, int, 874 ufs_lbn_t, int); 875static inline struct freeblks *newfreeblks(struct mount *, struct inode *); 876static void freeblks_free(struct ufsmount *, struct freeblks *, int); 877static void indir_trunc(struct freework *, ufs2_daddr_t, ufs_lbn_t); 878static ufs2_daddr_t blkcount(struct fs *, ufs2_daddr_t, off_t); 879static int trunc_check_buf(struct buf *, int *, ufs_lbn_t, int, int); 880static void trunc_dependencies(struct inode *, struct freeblks *, ufs_lbn_t, 881 int, int); 882static void trunc_pages(struct inode *, off_t, ufs2_daddr_t, int); 883static int cancel_pagedep(struct pagedep *, struct freeblks *, int); 884static int deallocate_dependencies(struct buf *, struct freeblks *, int); 885static void newblk_freefrag(struct newblk*); 886static void free_newblk(struct newblk *); 887static void cancel_allocdirect(struct allocdirectlst *, 888 struct allocdirect *, struct freeblks *); 889static int check_inode_unwritten(struct inodedep *); 890static int free_inodedep(struct inodedep *); 891static void freework_freeblock(struct freework *, u_long); 892static void freework_enqueue(struct freework *); 893static int handle_workitem_freeblocks(struct freeblks *, int); 894static int handle_complete_freeblocks(struct freeblks *, int); 895static void handle_workitem_indirblk(struct freework *); 896static void handle_written_freework(struct freework *); 897static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *); 898static struct worklist *jnewblk_merge(struct worklist *, struct worklist *, 899 struct workhead *); 900static struct freefrag *setup_allocindir_phase2(struct buf *, struct inode *, 901 struct inodedep *, struct allocindir *, ufs_lbn_t); 902static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t, 903 ufs2_daddr_t, ufs_lbn_t); 904static void handle_workitem_freefrag(struct freefrag *); 905static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long, 906 ufs_lbn_t, u_long); 907static void allocdirect_merge(struct allocdirectlst *, 908 struct allocdirect *, struct allocdirect *); 909static struct freefrag *allocindir_merge(struct allocindir *, 910 struct allocindir *); 911static int bmsafemap_find(struct bmsafemap_hashhead *, int, 912 struct bmsafemap **); 913static struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *, 914 int cg, struct bmsafemap *); 915static int newblk_find(struct newblk_hashhead *, ufs2_daddr_t, int, 916 struct newblk **); 917static int newblk_lookup(struct mount *, ufs2_daddr_t, int, struct newblk **); 918static int inodedep_find(struct inodedep_hashhead *, ino_t, 919 struct inodedep **); 920static int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **); 921static int pagedep_lookup(struct mount *, struct buf *bp, ino_t, ufs_lbn_t, 922 int, struct pagedep **); 923static int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t, 924 struct pagedep **); 925static void pause_timer(void *); 926static int request_cleanup(struct mount *, int); 927static int softdep_request_cleanup_flush(struct mount *, struct ufsmount *); 928static void schedule_cleanup(struct mount *); 929static void softdep_ast_cleanup_proc(struct thread *); 930static struct ufsmount *softdep_bp_to_mp(struct buf *bp); 931static int process_worklist_item(struct mount *, int, int); 932static void process_removes(struct vnode *); 933static void process_truncates(struct vnode *); 934static void jwork_move(struct workhead *, struct workhead *); 935static void jwork_insert(struct workhead *, struct jsegdep *); 936static void add_to_worklist(struct worklist *, int); 937static void wake_worklist(struct worklist *); 938static void wait_worklist(struct worklist *, char *); 939static void remove_from_worklist(struct worklist *); 940static void softdep_flush(void *); 941static void softdep_flushjournal(struct mount *); 942static int softdep_speedup(struct ufsmount *); 943static void worklist_speedup(struct mount *); 944static int journal_mount(struct mount *, struct fs *, struct ucred *); 945static void journal_unmount(struct ufsmount *); 946static int journal_space(struct ufsmount *, int); 947static void journal_suspend(struct ufsmount *); 948static int journal_unsuspend(struct ufsmount *ump); 949static void add_to_journal(struct worklist *); 950static void remove_from_journal(struct worklist *); 951static bool softdep_excess_items(struct ufsmount *, int); 952static void softdep_process_journal(struct mount *, struct worklist *, int); 953static struct jremref *newjremref(struct dirrem *, struct inode *, 954 struct inode *ip, off_t, nlink_t); 955static struct jaddref *newjaddref(struct inode *, ino_t, off_t, int16_t, 956 uint16_t); 957static inline void newinoref(struct inoref *, ino_t, ino_t, off_t, nlink_t, 958 uint16_t); 959static inline struct jsegdep *inoref_jseg(struct inoref *); 960static struct jmvref *newjmvref(struct inode *, ino_t, off_t, off_t); 961static struct jfreeblk *newjfreeblk(struct freeblks *, ufs_lbn_t, 962 ufs2_daddr_t, int); 963static void adjust_newfreework(struct freeblks *, int); 964static struct jtrunc *newjtrunc(struct freeblks *, off_t, int); 965static void move_newblock_dep(struct jaddref *, struct inodedep *); 966static void cancel_jfreeblk(struct freeblks *, ufs2_daddr_t); 967static struct jfreefrag *newjfreefrag(struct freefrag *, struct inode *, 968 ufs2_daddr_t, long, ufs_lbn_t); 969static struct freework *newfreework(struct ufsmount *, struct freeblks *, 970 struct freework *, ufs_lbn_t, ufs2_daddr_t, int, int, int); 971static int jwait(struct worklist *, int); 972static struct inodedep *inodedep_lookup_ip(struct inode *); 973static int bmsafemap_backgroundwrite(struct bmsafemap *, struct buf *); 974static struct freefile *handle_bufwait(struct inodedep *, struct workhead *); 975static void handle_jwork(struct workhead *); 976static struct mkdir *setup_newdir(struct diradd *, ino_t, ino_t, struct buf *, 977 struct mkdir **); 978static struct jblocks *jblocks_create(void); 979static ufs2_daddr_t jblocks_alloc(struct jblocks *, int, int *); 980static void jblocks_free(struct jblocks *, struct mount *, int); 981static void jblocks_destroy(struct jblocks *); 982static void jblocks_add(struct jblocks *, ufs2_daddr_t, int); 983 984/* 985 * Exported softdep operations. 986 */ 987static void softdep_disk_io_initiation(struct buf *); 988static void softdep_disk_write_complete(struct buf *); 989static void softdep_deallocate_dependencies(struct buf *); 990static int softdep_count_dependencies(struct buf *bp, int); 991 992/* 993 * Global lock over all of soft updates. 994 */ 995static struct mtx lk; 996MTX_SYSINIT(softdep_lock, &lk, "global softdep", MTX_DEF); 997 998#define ACQUIRE_GBLLOCK(lk) mtx_lock(lk) 999#define FREE_GBLLOCK(lk) mtx_unlock(lk) 1000#define GBLLOCK_OWNED(lk) mtx_assert((lk), MA_OWNED) 1001 1002/* 1003 * Per-filesystem soft-updates locking. 1004 */ 1005#define LOCK_PTR(ump) (&(ump)->um_softdep->sd_fslock) 1006#define TRY_ACQUIRE_LOCK(ump) rw_try_wlock(&(ump)->um_softdep->sd_fslock) 1007#define ACQUIRE_LOCK(ump) rw_wlock(&(ump)->um_softdep->sd_fslock) 1008#define FREE_LOCK(ump) rw_wunlock(&(ump)->um_softdep->sd_fslock) 1009#define LOCK_OWNED(ump) rw_assert(&(ump)->um_softdep->sd_fslock, \ 1010 RA_WLOCKED) 1011 1012#define BUF_AREC(bp) lockallowrecurse(&(bp)->b_lock) 1013#define BUF_NOREC(bp) lockdisablerecurse(&(bp)->b_lock) 1014 1015/* 1016 * Worklist queue management. 1017 * These routines require that the lock be held. 1018 */ 1019#ifndef /* NOT */ INVARIANTS 1020#define WORKLIST_INSERT(head, item) do { \ 1021 (item)->wk_state |= ONWORKLIST; \ 1022 LIST_INSERT_HEAD(head, item, wk_list); \ 1023} while (0) 1024#define WORKLIST_REMOVE(item) do { \ 1025 (item)->wk_state &= ~ONWORKLIST; \ 1026 LIST_REMOVE(item, wk_list); \ 1027} while (0) 1028#define WORKLIST_INSERT_UNLOCKED WORKLIST_INSERT 1029#define WORKLIST_REMOVE_UNLOCKED WORKLIST_REMOVE 1030 1031#else /* INVARIANTS */ 1032static void worklist_insert(struct workhead *, struct worklist *, int, 1033 const char *, int); 1034static void worklist_remove(struct worklist *, int, const char *, int); 1035 1036#define WORKLIST_INSERT(head, item) \ 1037 worklist_insert(head, item, 1, __func__, __LINE__) 1038#define WORKLIST_INSERT_UNLOCKED(head, item)\ 1039 worklist_insert(head, item, 0, __func__, __LINE__) 1040#define WORKLIST_REMOVE(item)\ 1041 worklist_remove(item, 1, __func__, __LINE__) 1042#define WORKLIST_REMOVE_UNLOCKED(item)\ 1043 worklist_remove(item, 0, __func__, __LINE__) 1044 1045static void 1046worklist_insert(head, item, locked, func, line) 1047 struct workhead *head; 1048 struct worklist *item; 1049 int locked; 1050 const char *func; 1051 int line; 1052{ 1053 1054 if (locked) 1055 LOCK_OWNED(VFSTOUFS(item->wk_mp)); 1056 if (item->wk_state & ONWORKLIST) 1057 panic("worklist_insert: %p %s(0x%X) already on list, " 1058 "added in function %s at line %d", 1059 item, TYPENAME(item->wk_type), item->wk_state, 1060 item->wk_func, item->wk_line); 1061 item->wk_state |= ONWORKLIST; 1062 item->wk_func = func; 1063 item->wk_line = line; 1064 LIST_INSERT_HEAD(head, item, wk_list); 1065} 1066 1067static void 1068worklist_remove(item, locked, func, line) 1069 struct worklist *item; 1070 int locked; 1071 const char *func; 1072 int line; 1073{ 1074 1075 if (locked) 1076 LOCK_OWNED(VFSTOUFS(item->wk_mp)); 1077 if ((item->wk_state & ONWORKLIST) == 0) 1078 panic("worklist_remove: %p %s(0x%X) not on list, " 1079 "removed in function %s at line %d", 1080 item, TYPENAME(item->wk_type), item->wk_state, 1081 item->wk_func, item->wk_line); 1082 item->wk_state &= ~ONWORKLIST; 1083 item->wk_func = func; 1084 item->wk_line = line; 1085 LIST_REMOVE(item, wk_list); 1086} 1087#endif /* INVARIANTS */ 1088 1089/* 1090 * Merge two jsegdeps keeping only the oldest one as newer references 1091 * can't be discarded until after older references. 1092 */ 1093static inline struct jsegdep * 1094jsegdep_merge(struct jsegdep *one, struct jsegdep *two) 1095{ 1096 struct jsegdep *swp; 1097 1098 if (two == NULL) 1099 return (one); 1100 1101 if (one->jd_seg->js_seq > two->jd_seg->js_seq) { 1102 swp = one; 1103 one = two; 1104 two = swp; 1105 } 1106 WORKLIST_REMOVE(&two->jd_list); 1107 free_jsegdep(two); 1108 1109 return (one); 1110} 1111 1112/* 1113 * If two freedeps are compatible free one to reduce list size. 1114 */ 1115static inline struct freedep * 1116freedep_merge(struct freedep *one, struct freedep *two) 1117{ 1118 if (two == NULL) 1119 return (one); 1120 1121 if (one->fd_freework == two->fd_freework) { 1122 WORKLIST_REMOVE(&two->fd_list); 1123 free_freedep(two); 1124 } 1125 return (one); 1126} 1127 1128/* 1129 * Move journal work from one list to another. Duplicate freedeps and 1130 * jsegdeps are coalesced to keep the lists as small as possible. 1131 */ 1132static void 1133jwork_move(dst, src) 1134 struct workhead *dst; 1135 struct workhead *src; 1136{ 1137 struct freedep *freedep; 1138 struct jsegdep *jsegdep; 1139 struct worklist *wkn; 1140 struct worklist *wk; 1141 1142 KASSERT(dst != src, 1143 ("jwork_move: dst == src")); 1144 freedep = NULL; 1145 jsegdep = NULL; 1146 LIST_FOREACH_SAFE(wk, dst, wk_list, wkn) { 1147 if (wk->wk_type == D_JSEGDEP) 1148 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1149 else if (wk->wk_type == D_FREEDEP) 1150 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1151 } 1152 1153 while ((wk = LIST_FIRST(src)) != NULL) { 1154 WORKLIST_REMOVE(wk); 1155 WORKLIST_INSERT(dst, wk); 1156 if (wk->wk_type == D_JSEGDEP) { 1157 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1158 continue; 1159 } 1160 if (wk->wk_type == D_FREEDEP) 1161 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1162 } 1163} 1164 1165static void 1166jwork_insert(dst, jsegdep) 1167 struct workhead *dst; 1168 struct jsegdep *jsegdep; 1169{ 1170 struct jsegdep *jsegdepn; 1171 struct worklist *wk; 1172 1173 LIST_FOREACH(wk, dst, wk_list) 1174 if (wk->wk_type == D_JSEGDEP) 1175 break; 1176 if (wk == NULL) { 1177 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1178 return; 1179 } 1180 jsegdepn = WK_JSEGDEP(wk); 1181 if (jsegdep->jd_seg->js_seq < jsegdepn->jd_seg->js_seq) { 1182 WORKLIST_REMOVE(wk); 1183 free_jsegdep(jsegdepn); 1184 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1185 } else 1186 free_jsegdep(jsegdep); 1187} 1188 1189/* 1190 * Routines for tracking and managing workitems. 1191 */ 1192static void workitem_free(struct worklist *, int); 1193static void workitem_alloc(struct worklist *, int, struct mount *); 1194static void workitem_reassign(struct worklist *, int); 1195 1196#define WORKITEM_FREE(item, type) \ 1197 workitem_free((struct worklist *)(item), (type)) 1198#define WORKITEM_REASSIGN(item, type) \ 1199 workitem_reassign((struct worklist *)(item), (type)) 1200 1201static void 1202workitem_free(item, type) 1203 struct worklist *item; 1204 int type; 1205{ 1206 struct ufsmount *ump; 1207 1208#ifdef INVARIANTS 1209 if (item->wk_state & ONWORKLIST) 1210 panic("workitem_free: %s(0x%X) still on list, " 1211 "added in function %s at line %d", 1212 TYPENAME(item->wk_type), item->wk_state, 1213 item->wk_func, item->wk_line); 1214 if (item->wk_type != type && type != D_NEWBLK) 1215 panic("workitem_free: type mismatch %s != %s", 1216 TYPENAME(item->wk_type), TYPENAME(type)); 1217#endif 1218 if (item->wk_state & IOWAITING) 1219 wakeup(item); 1220 ump = VFSTOUFS(item->wk_mp); 1221 LOCK_OWNED(ump); 1222 KASSERT(ump->softdep_deps > 0, 1223 ("workitem_free: %s: softdep_deps going negative", 1224 ump->um_fs->fs_fsmnt)); 1225 if (--ump->softdep_deps == 0 && ump->softdep_req) 1226 wakeup(&ump->softdep_deps); 1227 KASSERT(dep_current[item->wk_type] > 0, 1228 ("workitem_free: %s: dep_current[%s] going negative", 1229 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1230 KASSERT(ump->softdep_curdeps[item->wk_type] > 0, 1231 ("workitem_free: %s: softdep_curdeps[%s] going negative", 1232 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1233 atomic_subtract_long(&dep_current[item->wk_type], 1); 1234 ump->softdep_curdeps[item->wk_type] -= 1; 1235#ifdef INVARIANTS 1236 LIST_REMOVE(item, wk_all); 1237#endif 1238 free(item, DtoM(type)); 1239} 1240 1241static void 1242workitem_alloc(item, type, mp) 1243 struct worklist *item; 1244 int type; 1245 struct mount *mp; 1246{ 1247 struct ufsmount *ump; 1248 1249 item->wk_type = type; 1250 item->wk_mp = mp; 1251 item->wk_state = 0; 1252 1253 ump = VFSTOUFS(mp); 1254 ACQUIRE_GBLLOCK(&lk); 1255 dep_current[type]++; 1256 if (dep_current[type] > dep_highuse[type]) 1257 dep_highuse[type] = dep_current[type]; 1258 dep_total[type]++; 1259 FREE_GBLLOCK(&lk); 1260 ACQUIRE_LOCK(ump); 1261 ump->softdep_curdeps[type] += 1; 1262 ump->softdep_deps++; 1263 ump->softdep_accdeps++; 1264#ifdef INVARIANTS 1265 LIST_INSERT_HEAD(&ump->softdep_alldeps[type], item, wk_all); 1266#endif 1267 FREE_LOCK(ump); 1268} 1269 1270static void 1271workitem_reassign(item, newtype) 1272 struct worklist *item; 1273 int newtype; 1274{ 1275 struct ufsmount *ump; 1276 1277 ump = VFSTOUFS(item->wk_mp); 1278 LOCK_OWNED(ump); 1279 KASSERT(ump->softdep_curdeps[item->wk_type] > 0, 1280 ("workitem_reassign: %s: softdep_curdeps[%s] going negative", 1281 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1282 ump->softdep_curdeps[item->wk_type] -= 1; 1283 ump->softdep_curdeps[newtype] += 1; 1284 KASSERT(dep_current[item->wk_type] > 0, 1285 ("workitem_reassign: %s: dep_current[%s] going negative", 1286 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1287 ACQUIRE_GBLLOCK(&lk); 1288 dep_current[newtype]++; 1289 dep_current[item->wk_type]--; 1290 if (dep_current[newtype] > dep_highuse[newtype]) 1291 dep_highuse[newtype] = dep_current[newtype]; 1292 dep_total[newtype]++; 1293 FREE_GBLLOCK(&lk); 1294 item->wk_type = newtype; 1295} 1296 1297/* 1298 * Workitem queue management 1299 */ 1300static int max_softdeps; /* maximum number of structs before slowdown */ 1301static int tickdelay = 2; /* number of ticks to pause during slowdown */ 1302static int proc_waiting; /* tracks whether we have a timeout posted */ 1303static int *stat_countp; /* statistic to count in proc_waiting timeout */ 1304static struct callout softdep_callout; 1305static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 1306static int req_clear_remove; /* syncer process flush some freeblks */ 1307static int softdep_flushcache = 0; /* Should we do BIO_FLUSH? */ 1308 1309/* 1310 * runtime statistics 1311 */ 1312static int stat_flush_threads; /* number of softdep flushing threads */ 1313static int stat_worklist_push; /* number of worklist cleanups */ 1314static int stat_delayed_inact; /* number of delayed inactivation cleanups */ 1315static int stat_blk_limit_push; /* number of times block limit neared */ 1316static int stat_ino_limit_push; /* number of times inode limit neared */ 1317static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 1318static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 1319static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 1320static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 1321static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 1322static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 1323static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 1324static int stat_jaddref; /* bufs redirtied as ino bitmap can not write */ 1325static int stat_jnewblk; /* bufs redirtied as blk bitmap can not write */ 1326static int stat_journal_min; /* Times hit journal min threshold */ 1327static int stat_journal_low; /* Times hit journal low threshold */ 1328static int stat_journal_wait; /* Times blocked in jwait(). */ 1329static int stat_jwait_filepage; /* Times blocked in jwait() for filepage. */ 1330static int stat_jwait_freeblks; /* Times blocked in jwait() for freeblks. */ 1331static int stat_jwait_inode; /* Times blocked in jwait() for inodes. */ 1332static int stat_jwait_newblk; /* Times blocked in jwait() for newblks. */ 1333static int stat_cleanup_high_delay; /* Maximum cleanup delay (in ticks) */ 1334static int stat_cleanup_blkrequests; /* Number of block cleanup requests */ 1335static int stat_cleanup_inorequests; /* Number of inode cleanup requests */ 1336static int stat_cleanup_retries; /* Number of cleanups that needed to flush */ 1337static int stat_cleanup_failures; /* Number of cleanup requests that failed */ 1338static int stat_emptyjblocks; /* Number of potentially empty journal blocks */ 1339 1340SYSCTL_INT(_debug_softdep, OID_AUTO, max_softdeps, CTLFLAG_RW, 1341 &max_softdeps, 0, ""); 1342SYSCTL_INT(_debug_softdep, OID_AUTO, tickdelay, CTLFLAG_RW, 1343 &tickdelay, 0, ""); 1344SYSCTL_INT(_debug_softdep, OID_AUTO, flush_threads, CTLFLAG_RD, 1345 &stat_flush_threads, 0, ""); 1346SYSCTL_INT(_debug_softdep, OID_AUTO, worklist_push, 1347 CTLFLAG_RW | CTLFLAG_STATS, &stat_worklist_push, 0,""); 1348SYSCTL_INT(_debug_softdep, OID_AUTO, delayed_inactivations, CTLFLAG_RD, 1349 &stat_delayed_inact, 0, ""); 1350SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_push, 1351 CTLFLAG_RW | CTLFLAG_STATS, &stat_blk_limit_push, 0,""); 1352SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_push, 1353 CTLFLAG_RW | CTLFLAG_STATS, &stat_ino_limit_push, 0,""); 1354SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_hit, 1355 CTLFLAG_RW | CTLFLAG_STATS, &stat_blk_limit_hit, 0, ""); 1356SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_hit, 1357 CTLFLAG_RW | CTLFLAG_STATS, &stat_ino_limit_hit, 0, ""); 1358SYSCTL_INT(_debug_softdep, OID_AUTO, sync_limit_hit, 1359 CTLFLAG_RW | CTLFLAG_STATS, &stat_sync_limit_hit, 0, ""); 1360SYSCTL_INT(_debug_softdep, OID_AUTO, indir_blk_ptrs, 1361 CTLFLAG_RW | CTLFLAG_STATS, &stat_indir_blk_ptrs, 0, ""); 1362SYSCTL_INT(_debug_softdep, OID_AUTO, inode_bitmap, 1363 CTLFLAG_RW | CTLFLAG_STATS, &stat_inode_bitmap, 0, ""); 1364SYSCTL_INT(_debug_softdep, OID_AUTO, direct_blk_ptrs, 1365 CTLFLAG_RW | CTLFLAG_STATS, &stat_direct_blk_ptrs, 0, ""); 1366SYSCTL_INT(_debug_softdep, OID_AUTO, dir_entry, 1367 CTLFLAG_RW | CTLFLAG_STATS, &stat_dir_entry, 0, ""); 1368SYSCTL_INT(_debug_softdep, OID_AUTO, jaddref_rollback, 1369 CTLFLAG_RW | CTLFLAG_STATS, &stat_jaddref, 0, ""); 1370SYSCTL_INT(_debug_softdep, OID_AUTO, jnewblk_rollback, 1371 CTLFLAG_RW | CTLFLAG_STATS, &stat_jnewblk, 0, ""); 1372SYSCTL_INT(_debug_softdep, OID_AUTO, journal_low, 1373 CTLFLAG_RW | CTLFLAG_STATS, &stat_journal_low, 0, ""); 1374SYSCTL_INT(_debug_softdep, OID_AUTO, journal_min, 1375 CTLFLAG_RW | CTLFLAG_STATS, &stat_journal_min, 0, ""); 1376SYSCTL_INT(_debug_softdep, OID_AUTO, journal_wait, 1377 CTLFLAG_RW | CTLFLAG_STATS, &stat_journal_wait, 0, ""); 1378SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_filepage, 1379 CTLFLAG_RW | CTLFLAG_STATS, &stat_jwait_filepage, 0, ""); 1380SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_freeblks, 1381 CTLFLAG_RW | CTLFLAG_STATS, &stat_jwait_freeblks, 0, ""); 1382SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_inode, 1383 CTLFLAG_RW | CTLFLAG_STATS, &stat_jwait_inode, 0, ""); 1384SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_newblk, 1385 CTLFLAG_RW | CTLFLAG_STATS, &stat_jwait_newblk, 0, ""); 1386SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_blkrequests, 1387 CTLFLAG_RW | CTLFLAG_STATS, &stat_cleanup_blkrequests, 0, ""); 1388SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_inorequests, 1389 CTLFLAG_RW | CTLFLAG_STATS, &stat_cleanup_inorequests, 0, ""); 1390SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_high_delay, 1391 CTLFLAG_RW | CTLFLAG_STATS, &stat_cleanup_high_delay, 0, ""); 1392SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_retries, 1393 CTLFLAG_RW | CTLFLAG_STATS, &stat_cleanup_retries, 0, ""); 1394SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_failures, 1395 CTLFLAG_RW | CTLFLAG_STATS, &stat_cleanup_failures, 0, ""); 1396 1397SYSCTL_INT(_debug_softdep, OID_AUTO, flushcache, CTLFLAG_RW, 1398 &softdep_flushcache, 0, ""); 1399SYSCTL_INT(_debug_softdep, OID_AUTO, emptyjblocks, CTLFLAG_RD, 1400 &stat_emptyjblocks, 0, ""); 1401 1402SYSCTL_DECL(_vfs_ffs); 1403 1404/* Whether to recompute the summary at mount time */ 1405static int compute_summary_at_mount = 0; 1406SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW, 1407 &compute_summary_at_mount, 0, "Recompute summary at mount"); 1408static int print_threads = 0; 1409SYSCTL_INT(_debug_softdep, OID_AUTO, print_threads, CTLFLAG_RW, 1410 &print_threads, 0, "Notify flusher thread start/stop"); 1411 1412/* List of all filesystems mounted with soft updates */ 1413static TAILQ_HEAD(, mount_softdeps) softdepmounts; 1414 1415static void 1416get_parent_vp_unlock_bp(struct mount *mp, struct buf *bp, 1417 struct diraddhd *diraddhdp, struct diraddhd *unfinishedp) 1418{ 1419 struct diradd *dap; 1420 1421 /* 1422 * Requeue unfinished dependencies before 1423 * unlocking buffer, which could make 1424 * diraddhdp invalid. 1425 */ 1426 ACQUIRE_LOCK(VFSTOUFS(mp)); 1427 while ((dap = LIST_FIRST(unfinishedp)) != NULL) { 1428 LIST_REMOVE(dap, da_pdlist); 1429 LIST_INSERT_HEAD(diraddhdp, dap, da_pdlist); 1430 } 1431 FREE_LOCK(VFSTOUFS(mp)); 1432 1433 bp->b_vflags &= ~BV_SCANNED; 1434 BUF_NOREC(bp); 1435 BUF_UNLOCK(bp); 1436} 1437 1438/* 1439 * This function fetches inode inum on mount point mp. We already 1440 * hold a locked vnode vp, and might have a locked buffer bp belonging 1441 * to vp. 1442 1443 * We must not block on acquiring the new inode lock as we will get 1444 * into a lock-order reversal with the buffer lock and possibly get a 1445 * deadlock. Thus if we cannot instantiate the requested vnode 1446 * without sleeping on its lock, we must unlock the vnode and the 1447 * buffer before doing a blocking on the vnode lock. We return 1448 * ERELOOKUP if we have had to unlock either the vnode or the buffer so 1449 * that the caller can reassess its state. 1450 * 1451 * Top-level VFS code (for syscalls and other consumers, e.g. callers 1452 * of VOP_FSYNC() in syncer) check for ERELOOKUP and restart at safe 1453 * point. 1454 * 1455 * Since callers expect to operate on fully constructed vnode, we also 1456 * recheck v_data after relock, and return ENOENT if NULL. 1457 * 1458 * If unlocking bp, we must unroll dequeueing its unfinished 1459 * dependencies, and clear scan flag, before unlocking. If unlocking 1460 * vp while it is under deactivation, we re-queue deactivation. 1461 */ 1462static int 1463get_parent_vp(struct vnode *vp, struct mount *mp, ino_t inum, struct buf *bp, 1464 struct diraddhd *diraddhdp, struct diraddhd *unfinishedp, 1465 struct vnode **rvp) 1466{ 1467 struct vnode *pvp; 1468 int error; 1469 bool bplocked; 1470 1471 ASSERT_VOP_ELOCKED(vp, "child vnode must be locked"); 1472 for (bplocked = true, pvp = NULL;;) { 1473 error = ffs_vgetf(mp, inum, LK_EXCLUSIVE | LK_NOWAIT, &pvp, 1474 FFSV_FORCEINSMQ); 1475 if (error == 0) { 1476 /* 1477 * Since we could have unlocked vp, the inode 1478 * number could no longer indicate a 1479 * constructed node. In this case, we must 1480 * restart the syscall. 1481 */ 1482 if (VTOI(pvp)->i_mode == 0 || !bplocked) { 1483 if (bp != NULL && bplocked) 1484 get_parent_vp_unlock_bp(mp, bp, 1485 diraddhdp, unfinishedp); 1486 if (VTOI(pvp)->i_mode == 0) 1487 vgone(pvp); 1488 error = ERELOOKUP; 1489 goto out2; 1490 } 1491 goto out1; 1492 } 1493 if (bp != NULL && bplocked) { 1494 get_parent_vp_unlock_bp(mp, bp, diraddhdp, unfinishedp); 1495 bplocked = false; 1496 } 1497 1498 /* 1499 * Do not drop vnode lock while inactivating during 1500 * vunref. This would result in leaks of the VI flags 1501 * and reclaiming of non-truncated vnode. Instead, 1502 * re-schedule inactivation hoping that we would be 1503 * able to sync inode later. 1504 */ 1505 if ((vp->v_iflag & VI_DOINGINACT) != 0 && 1506 (vp->v_vflag & VV_UNREF) != 0) { 1507 VI_LOCK(vp); 1508 vp->v_iflag |= VI_OWEINACT; 1509 VI_UNLOCK(vp); 1510 return (ERELOOKUP); 1511 } 1512 1513 VOP_UNLOCK(vp); 1514 error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &pvp, 1515 FFSV_FORCEINSMQ); 1516 if (error != 0) { 1517 MPASS(error != ERELOOKUP); 1518 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1519 break; 1520 } 1521 if (VTOI(pvp)->i_mode == 0) { 1522 vgone(pvp); 1523 vput(pvp); 1524 pvp = NULL; 1525 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1526 error = ERELOOKUP; 1527 break; 1528 } 1529 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); 1530 if (error == 0) 1531 break; 1532 vput(pvp); 1533 pvp = NULL; 1534 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1535 if (vp->v_data == NULL) { 1536 error = ENOENT; 1537 break; 1538 } 1539 } 1540 if (bp != NULL) { 1541 MPASS(!bplocked); 1542 error = ERELOOKUP; 1543 } 1544out2: 1545 if (error != 0 && pvp != NULL) { 1546 vput(pvp); 1547 pvp = NULL; 1548 } 1549out1: 1550 *rvp = pvp; 1551 ASSERT_VOP_ELOCKED(vp, "child vnode must be locked on return"); 1552 return (error); 1553} 1554 1555/* 1556 * This function cleans the worklist for a filesystem. 1557 * Each filesystem running with soft dependencies gets its own 1558 * thread to run in this function. The thread is started up in 1559 * softdep_mount and shutdown in softdep_unmount. They show up 1560 * as part of the kernel "bufdaemon" process whose process 1561 * entry is available in bufdaemonproc. 1562 */ 1563static int searchfailed; 1564extern struct proc *bufdaemonproc; 1565static void 1566softdep_flush(addr) 1567 void *addr; 1568{ 1569 struct mount *mp; 1570 struct thread *td; 1571 struct ufsmount *ump; 1572 1573 td = curthread; 1574 td->td_pflags |= TDP_NORUNNINGBUF; 1575 mp = (struct mount *)addr; 1576 ump = VFSTOUFS(mp); 1577 atomic_add_int(&stat_flush_threads, 1); 1578 ACQUIRE_LOCK(ump); 1579 ump->softdep_flags &= ~FLUSH_STARTING; 1580 wakeup(&ump->softdep_flushtd); 1581 FREE_LOCK(ump); 1582 if (print_threads) { 1583 if (stat_flush_threads == 1) 1584 printf("Running %s at pid %d\n", bufdaemonproc->p_comm, 1585 bufdaemonproc->p_pid); 1586 printf("Start thread %s\n", td->td_name); 1587 } 1588 for (;;) { 1589 while (softdep_process_worklist(mp, 0) > 0 || 1590 (MOUNTEDSUJ(mp) && 1591 VFSTOUFS(mp)->softdep_jblocks->jb_suspended)) 1592 kthread_suspend_check(); 1593 ACQUIRE_LOCK(ump); 1594 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1595 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM, 1596 "sdflush", hz / 2); 1597 ump->softdep_flags &= ~FLUSH_CLEANUP; 1598 /* 1599 * Check to see if we are done and need to exit. 1600 */ 1601 if ((ump->softdep_flags & FLUSH_EXIT) == 0) { 1602 FREE_LOCK(ump); 1603 continue; 1604 } 1605 ump->softdep_flags &= ~FLUSH_EXIT; 1606 FREE_LOCK(ump); 1607 wakeup(&ump->softdep_flags); 1608 if (print_threads) 1609 printf("Stop thread %s: searchfailed %d, did cleanups %d\n", td->td_name, searchfailed, ump->um_softdep->sd_cleanups); 1610 atomic_subtract_int(&stat_flush_threads, 1); 1611 kthread_exit(); 1612 panic("kthread_exit failed\n"); 1613 } 1614} 1615 1616static void 1617worklist_speedup(mp) 1618 struct mount *mp; 1619{ 1620 struct ufsmount *ump; 1621 1622 ump = VFSTOUFS(mp); 1623 LOCK_OWNED(ump); 1624 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1625 ump->softdep_flags |= FLUSH_CLEANUP; 1626 wakeup(&ump->softdep_flushtd); 1627} 1628 1629static void 1630softdep_send_speedup(struct ufsmount *ump, off_t shortage, u_int flags) 1631{ 1632 struct buf *bp; 1633 1634 if ((ump->um_flags & UM_CANSPEEDUP) == 0) 1635 return; 1636 1637 bp = malloc(sizeof(*bp), M_TRIM, M_WAITOK | M_ZERO); 1638 bp->b_iocmd = BIO_SPEEDUP; 1639 bp->b_ioflags = flags; 1640 bp->b_bcount = omin(shortage, LONG_MAX); 1641 g_vfs_strategy(ump->um_bo, bp); 1642 bufwait(bp); 1643 free(bp, M_TRIM); 1644} 1645 1646static int 1647softdep_speedup(ump) 1648 struct ufsmount *ump; 1649{ 1650 struct ufsmount *altump; 1651 struct mount_softdeps *sdp; 1652 1653 LOCK_OWNED(ump); 1654 worklist_speedup(ump->um_mountp); 1655 bd_speedup(); 1656 /* 1657 * If we have global shortages, then we need other 1658 * filesystems to help with the cleanup. Here we wakeup a 1659 * flusher thread for a filesystem that is over its fair 1660 * share of resources. 1661 */ 1662 if (req_clear_inodedeps || req_clear_remove) { 1663 ACQUIRE_GBLLOCK(&lk); 1664 TAILQ_FOREACH(sdp, &softdepmounts, sd_next) { 1665 if ((altump = sdp->sd_ump) == ump) 1666 continue; 1667 if (((req_clear_inodedeps && 1668 altump->softdep_curdeps[D_INODEDEP] > 1669 max_softdeps / stat_flush_threads) || 1670 (req_clear_remove && 1671 altump->softdep_curdeps[D_DIRREM] > 1672 (max_softdeps / 2) / stat_flush_threads)) && 1673 TRY_ACQUIRE_LOCK(altump)) 1674 break; 1675 } 1676 if (sdp == NULL) { 1677 searchfailed++; 1678 FREE_GBLLOCK(&lk); 1679 } else { 1680 /* 1681 * Move to the end of the list so we pick a 1682 * different one on out next try. 1683 */ 1684 TAILQ_REMOVE(&softdepmounts, sdp, sd_next); 1685 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next); 1686 FREE_GBLLOCK(&lk); 1687 if ((altump->softdep_flags & 1688 (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1689 altump->softdep_flags |= FLUSH_CLEANUP; 1690 altump->um_softdep->sd_cleanups++; 1691 wakeup(&altump->softdep_flushtd); 1692 FREE_LOCK(altump); 1693 } 1694 } 1695 return (speedup_syncer()); 1696} 1697 1698/* 1699 * Add an item to the end of the work queue. 1700 * This routine requires that the lock be held. 1701 * This is the only routine that adds items to the list. 1702 * The following routine is the only one that removes items 1703 * and does so in order from first to last. 1704 */ 1705 1706#define WK_HEAD 0x0001 /* Add to HEAD. */ 1707#define WK_NODELAY 0x0002 /* Process immediately. */ 1708 1709static void 1710add_to_worklist(wk, flags) 1711 struct worklist *wk; 1712 int flags; 1713{ 1714 struct ufsmount *ump; 1715 1716 ump = VFSTOUFS(wk->wk_mp); 1717 LOCK_OWNED(ump); 1718 if (wk->wk_state & ONWORKLIST) 1719 panic("add_to_worklist: %s(0x%X) already on list", 1720 TYPENAME(wk->wk_type), wk->wk_state); 1721 wk->wk_state |= ONWORKLIST; 1722 if (ump->softdep_on_worklist == 0) { 1723 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1724 ump->softdep_worklist_tail = wk; 1725 } else if (flags & WK_HEAD) { 1726 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1727 } else { 1728 LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list); 1729 ump->softdep_worklist_tail = wk; 1730 } 1731 ump->softdep_on_worklist += 1; 1732 if (flags & WK_NODELAY) 1733 worklist_speedup(wk->wk_mp); 1734} 1735 1736/* 1737 * Remove the item to be processed. If we are removing the last 1738 * item on the list, we need to recalculate the tail pointer. 1739 */ 1740static void 1741remove_from_worklist(wk) 1742 struct worklist *wk; 1743{ 1744 struct ufsmount *ump; 1745 1746 ump = VFSTOUFS(wk->wk_mp); 1747 if (ump->softdep_worklist_tail == wk) 1748 ump->softdep_worklist_tail = 1749 (struct worklist *)wk->wk_list.le_prev; 1750 WORKLIST_REMOVE(wk); 1751 ump->softdep_on_worklist -= 1; 1752} 1753 1754static void 1755wake_worklist(wk) 1756 struct worklist *wk; 1757{ 1758 if (wk->wk_state & IOWAITING) { 1759 wk->wk_state &= ~IOWAITING; 1760 wakeup(wk); 1761 } 1762} 1763 1764static void 1765wait_worklist(wk, wmesg) 1766 struct worklist *wk; 1767 char *wmesg; 1768{ 1769 struct ufsmount *ump; 1770 1771 ump = VFSTOUFS(wk->wk_mp); 1772 wk->wk_state |= IOWAITING; 1773 msleep(wk, LOCK_PTR(ump), PVM, wmesg, 0); 1774} 1775 1776/* 1777 * Process that runs once per second to handle items in the background queue. 1778 * 1779 * Note that we ensure that everything is done in the order in which they 1780 * appear in the queue. The code below depends on this property to ensure 1781 * that blocks of a file are freed before the inode itself is freed. This 1782 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 1783 * until all the old ones have been purged from the dependency lists. 1784 */ 1785static int 1786softdep_process_worklist(mp, full) 1787 struct mount *mp; 1788 int full; 1789{ 1790 int cnt, matchcnt; 1791 struct ufsmount *ump; 1792 long starttime; 1793 1794 KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp")); 1795 if (MOUNTEDSOFTDEP(mp) == 0) 1796 return (0); 1797 matchcnt = 0; 1798 ump = VFSTOUFS(mp); 1799 ACQUIRE_LOCK(ump); 1800 starttime = time_second; 1801 softdep_process_journal(mp, NULL, full ? MNT_WAIT : 0); 1802 check_clear_deps(mp); 1803 while (ump->softdep_on_worklist > 0) { 1804 if ((cnt = process_worklist_item(mp, 10, LK_NOWAIT)) == 0) 1805 break; 1806 else 1807 matchcnt += cnt; 1808 check_clear_deps(mp); 1809 /* 1810 * We do not generally want to stop for buffer space, but if 1811 * we are really being a buffer hog, we will stop and wait. 1812 */ 1813 if (should_yield()) { 1814 FREE_LOCK(ump); 1815 kern_yield(PRI_USER); 1816 bwillwrite(); 1817 ACQUIRE_LOCK(ump); 1818 } 1819 /* 1820 * Never allow processing to run for more than one 1821 * second. This gives the syncer thread the opportunity 1822 * to pause if appropriate. 1823 */ 1824 if (!full && starttime != time_second) 1825 break; 1826 } 1827 if (full == 0) 1828 journal_unsuspend(ump); 1829 FREE_LOCK(ump); 1830 return (matchcnt); 1831} 1832 1833/* 1834 * Process all removes associated with a vnode if we are running out of 1835 * journal space. Any other process which attempts to flush these will 1836 * be unable as we have the vnodes locked. 1837 */ 1838static void 1839process_removes(vp) 1840 struct vnode *vp; 1841{ 1842 struct inodedep *inodedep; 1843 struct dirrem *dirrem; 1844 struct ufsmount *ump; 1845 struct mount *mp; 1846 ino_t inum; 1847 1848 mp = vp->v_mount; 1849 ump = VFSTOUFS(mp); 1850 LOCK_OWNED(ump); 1851 inum = VTOI(vp)->i_number; 1852 for (;;) { 1853top: 1854 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1855 return; 1856 LIST_FOREACH(dirrem, &inodedep->id_dirremhd, dm_inonext) { 1857 /* 1858 * If another thread is trying to lock this vnode 1859 * it will fail but we must wait for it to do so 1860 * before we can proceed. 1861 */ 1862 if (dirrem->dm_state & INPROGRESS) { 1863 wait_worklist(&dirrem->dm_list, "pwrwait"); 1864 goto top; 1865 } 1866 if ((dirrem->dm_state & (COMPLETE | ONWORKLIST)) == 1867 (COMPLETE | ONWORKLIST)) 1868 break; 1869 } 1870 if (dirrem == NULL) 1871 return; 1872 remove_from_worklist(&dirrem->dm_list); 1873 FREE_LOCK(ump); 1874 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1875 panic("process_removes: suspended filesystem"); 1876 handle_workitem_remove(dirrem, 0); 1877 vn_finished_secondary_write(mp); 1878 ACQUIRE_LOCK(ump); 1879 } 1880} 1881 1882/* 1883 * Process all truncations associated with a vnode if we are running out 1884 * of journal space. This is called when the vnode lock is already held 1885 * and no other process can clear the truncation. This function returns 1886 * a value greater than zero if it did any work. 1887 */ 1888static void 1889process_truncates(vp) 1890 struct vnode *vp; 1891{ 1892 struct inodedep *inodedep; 1893 struct freeblks *freeblks; 1894 struct ufsmount *ump; 1895 struct mount *mp; 1896 ino_t inum; 1897 int cgwait; 1898 1899 mp = vp->v_mount; 1900 ump = VFSTOUFS(mp); 1901 LOCK_OWNED(ump); 1902 inum = VTOI(vp)->i_number; 1903 for (;;) { 1904 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1905 return; 1906 cgwait = 0; 1907 TAILQ_FOREACH(freeblks, &inodedep->id_freeblklst, fb_next) { 1908 /* Journal entries not yet written. */ 1909 if (!LIST_EMPTY(&freeblks->fb_jblkdephd)) { 1910 jwait(&LIST_FIRST( 1911 &freeblks->fb_jblkdephd)->jb_list, 1912 MNT_WAIT); 1913 break; 1914 } 1915 /* Another thread is executing this item. */ 1916 if (freeblks->fb_state & INPROGRESS) { 1917 wait_worklist(&freeblks->fb_list, "ptrwait"); 1918 break; 1919 } 1920 /* Freeblks is waiting on a inode write. */ 1921 if ((freeblks->fb_state & COMPLETE) == 0) { 1922 FREE_LOCK(ump); 1923 ffs_update(vp, 1); 1924 ACQUIRE_LOCK(ump); 1925 break; 1926 } 1927 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST)) == 1928 (ALLCOMPLETE | ONWORKLIST)) { 1929 remove_from_worklist(&freeblks->fb_list); 1930 freeblks->fb_state |= INPROGRESS; 1931 FREE_LOCK(ump); 1932 if (vn_start_secondary_write(NULL, &mp, 1933 V_NOWAIT)) 1934 panic("process_truncates: " 1935 "suspended filesystem"); 1936 handle_workitem_freeblocks(freeblks, 0); 1937 vn_finished_secondary_write(mp); 1938 ACQUIRE_LOCK(ump); 1939 break; 1940 } 1941 if (freeblks->fb_cgwait) 1942 cgwait++; 1943 } 1944 if (cgwait) { 1945 FREE_LOCK(ump); 1946 sync_cgs(mp, MNT_WAIT); 1947 ffs_sync_snap(mp, MNT_WAIT); 1948 ACQUIRE_LOCK(ump); 1949 continue; 1950 } 1951 if (freeblks == NULL) 1952 break; 1953 } 1954 return; 1955} 1956 1957/* 1958 * Process one item on the worklist. 1959 */ 1960static int 1961process_worklist_item(mp, target, flags) 1962 struct mount *mp; 1963 int target; 1964 int flags; 1965{ 1966 struct worklist sentinel; 1967 struct worklist *wk; 1968 struct ufsmount *ump; 1969 int matchcnt; 1970 int error; 1971 1972 KASSERT(mp != NULL, ("process_worklist_item: NULL mp")); 1973 /* 1974 * If we are being called because of a process doing a 1975 * copy-on-write, then it is not safe to write as we may 1976 * recurse into the copy-on-write routine. 1977 */ 1978 if (curthread->td_pflags & TDP_COWINPROGRESS) 1979 return (-1); 1980 PHOLD(curproc); /* Don't let the stack go away. */ 1981 ump = VFSTOUFS(mp); 1982 LOCK_OWNED(ump); 1983 matchcnt = 0; 1984 sentinel.wk_mp = NULL; 1985 sentinel.wk_type = D_SENTINEL; 1986 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, &sentinel, wk_list); 1987 for (wk = LIST_NEXT(&sentinel, wk_list); wk != NULL; 1988 wk = LIST_NEXT(&sentinel, wk_list)) { 1989 if (wk->wk_type == D_SENTINEL) { 1990 LIST_REMOVE(&sentinel, wk_list); 1991 LIST_INSERT_AFTER(wk, &sentinel, wk_list); 1992 continue; 1993 } 1994 if (wk->wk_state & INPROGRESS) 1995 panic("process_worklist_item: %p already in progress.", 1996 wk); 1997 wk->wk_state |= INPROGRESS; 1998 remove_from_worklist(wk); 1999 FREE_LOCK(ump); 2000 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 2001 panic("process_worklist_item: suspended filesystem"); 2002 switch (wk->wk_type) { 2003 case D_DIRREM: 2004 /* removal of a directory entry */ 2005 error = handle_workitem_remove(WK_DIRREM(wk), flags); 2006 break; 2007 2008 case D_FREEBLKS: 2009 /* releasing blocks and/or fragments from a file */ 2010 error = handle_workitem_freeblocks(WK_FREEBLKS(wk), 2011 flags); 2012 break; 2013 2014 case D_FREEFRAG: 2015 /* releasing a fragment when replaced as a file grows */ 2016 handle_workitem_freefrag(WK_FREEFRAG(wk)); 2017 error = 0; 2018 break; 2019 2020 case D_FREEFILE: 2021 /* releasing an inode when its link count drops to 0 */ 2022 handle_workitem_freefile(WK_FREEFILE(wk)); 2023 error = 0; 2024 break; 2025 2026 default: 2027 panic("%s_process_worklist: Unknown type %s", 2028 "softdep", TYPENAME(wk->wk_type)); 2029 /* NOTREACHED */ 2030 } 2031 vn_finished_secondary_write(mp); 2032 ACQUIRE_LOCK(ump); 2033 if (error == 0) { 2034 if (++matchcnt == target) 2035 break; 2036 continue; 2037 } 2038 /* 2039 * We have to retry the worklist item later. Wake up any 2040 * waiters who may be able to complete it immediately and 2041 * add the item back to the head so we don't try to execute 2042 * it again. 2043 */ 2044 wk->wk_state &= ~INPROGRESS; 2045 wake_worklist(wk); 2046 add_to_worklist(wk, WK_HEAD); 2047 } 2048 /* Sentinal could've become the tail from remove_from_worklist. */ 2049 if (ump->softdep_worklist_tail == &sentinel) 2050 ump->softdep_worklist_tail = 2051 (struct worklist *)sentinel.wk_list.le_prev; 2052 LIST_REMOVE(&sentinel, wk_list); 2053 PRELE(curproc); 2054 return (matchcnt); 2055} 2056 2057/* 2058 * Move dependencies from one buffer to another. 2059 */ 2060int 2061softdep_move_dependencies(oldbp, newbp) 2062 struct buf *oldbp; 2063 struct buf *newbp; 2064{ 2065 struct worklist *wk, *wktail; 2066 struct ufsmount *ump; 2067 int dirty; 2068 2069 if ((wk = LIST_FIRST(&oldbp->b_dep)) == NULL) 2070 return (0); 2071 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 2072 ("softdep_move_dependencies called on non-softdep filesystem")); 2073 dirty = 0; 2074 wktail = NULL; 2075 ump = VFSTOUFS(wk->wk_mp); 2076 ACQUIRE_LOCK(ump); 2077 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 2078 LIST_REMOVE(wk, wk_list); 2079 if (wk->wk_type == D_BMSAFEMAP && 2080 bmsafemap_backgroundwrite(WK_BMSAFEMAP(wk), newbp)) 2081 dirty = 1; 2082 if (wktail == NULL) 2083 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 2084 else 2085 LIST_INSERT_AFTER(wktail, wk, wk_list); 2086 wktail = wk; 2087 } 2088 FREE_LOCK(ump); 2089 2090 return (dirty); 2091} 2092 2093/* 2094 * Purge the work list of all items associated with a particular mount point. 2095 */ 2096int 2097softdep_flushworklist(oldmnt, countp, td) 2098 struct mount *oldmnt; 2099 int *countp; 2100 struct thread *td; 2101{ 2102 struct vnode *devvp; 2103 struct ufsmount *ump; 2104 int count, error; 2105 2106 /* 2107 * Alternately flush the block device associated with the mount 2108 * point and process any dependencies that the flushing 2109 * creates. We continue until no more worklist dependencies 2110 * are found. 2111 */ 2112 *countp = 0; 2113 error = 0; 2114 ump = VFSTOUFS(oldmnt); 2115 devvp = ump->um_devvp; 2116 while ((count = softdep_process_worklist(oldmnt, 1)) > 0) { 2117 *countp += count; 2118 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 2119 error = VOP_FSYNC(devvp, MNT_WAIT, td); 2120 VOP_UNLOCK(devvp); 2121 if (error != 0) 2122 break; 2123 } 2124 return (error); 2125} 2126 2127#define SU_WAITIDLE_RETRIES 20 2128static int 2129softdep_waitidle(struct mount *mp, int flags __unused) 2130{ 2131 struct ufsmount *ump; 2132 struct vnode *devvp; 2133 struct thread *td; 2134 int error, i; 2135 2136 ump = VFSTOUFS(mp); 2137 devvp = ump->um_devvp; 2138 td = curthread; 2139 error = 0; 2140 ACQUIRE_LOCK(ump); 2141 for (i = 0; i < SU_WAITIDLE_RETRIES && ump->softdep_deps != 0; i++) { 2142 ump->softdep_req = 1; 2143 KASSERT((flags & FORCECLOSE) == 0 || 2144 ump->softdep_on_worklist == 0, 2145 ("softdep_waitidle: work added after flush")); 2146 msleep(&ump->softdep_deps, LOCK_PTR(ump), PVM | PDROP, 2147 "softdeps", 10 * hz); 2148 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 2149 error = VOP_FSYNC(devvp, MNT_WAIT, td); 2150 VOP_UNLOCK(devvp); 2151 ACQUIRE_LOCK(ump); 2152 if (error != 0) 2153 break; 2154 } 2155 ump->softdep_req = 0; 2156 if (i == SU_WAITIDLE_RETRIES && error == 0 && ump->softdep_deps != 0) { 2157 error = EBUSY; 2158 printf("softdep_waitidle: Failed to flush worklist for %p\n", 2159 mp); 2160 } 2161 FREE_LOCK(ump); 2162 return (error); 2163} 2164 2165/* 2166 * Flush all vnodes and worklist items associated with a specified mount point. 2167 */ 2168int 2169softdep_flushfiles(oldmnt, flags, td) 2170 struct mount *oldmnt; 2171 int flags; 2172 struct thread *td; 2173{ 2174#ifdef QUOTA 2175 struct ufsmount *ump; 2176 int i; 2177#endif 2178 int error, early, depcount, loopcnt, retry_flush_count, retry; 2179 int morework; 2180 2181 KASSERT(MOUNTEDSOFTDEP(oldmnt) != 0, 2182 ("softdep_flushfiles called on non-softdep filesystem")); 2183 loopcnt = 10; 2184 retry_flush_count = 3; 2185retry_flush: 2186 error = 0; 2187 2188 /* 2189 * Alternately flush the vnodes associated with the mount 2190 * point and process any dependencies that the flushing 2191 * creates. In theory, this loop can happen at most twice, 2192 * but we give it a few extra just to be sure. 2193 */ 2194 for (; loopcnt > 0; loopcnt--) { 2195 /* 2196 * Do another flush in case any vnodes were brought in 2197 * as part of the cleanup operations. 2198 */ 2199 early = retry_flush_count == 1 || (oldmnt->mnt_kern_flag & 2200 MNTK_UNMOUNT) == 0 ? 0 : EARLYFLUSH; 2201 if ((error = ffs_flushfiles(oldmnt, flags | early, td)) != 0) 2202 break; 2203 if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 || 2204 depcount == 0) 2205 break; 2206 } 2207 /* 2208 * If we are unmounting then it is an error to fail. If we 2209 * are simply trying to downgrade to read-only, then filesystem 2210 * activity can keep us busy forever, so we just fail with EBUSY. 2211 */ 2212 if (loopcnt == 0) { 2213 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 2214 panic("softdep_flushfiles: looping"); 2215 error = EBUSY; 2216 } 2217 if (!error) 2218 error = softdep_waitidle(oldmnt, flags); 2219 if (!error) { 2220 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) { 2221 retry = 0; 2222 MNT_ILOCK(oldmnt); 2223 morework = oldmnt->mnt_nvnodelistsize > 0; 2224#ifdef QUOTA 2225 ump = VFSTOUFS(oldmnt); 2226 UFS_LOCK(ump); 2227 for (i = 0; i < MAXQUOTAS; i++) { 2228 if (ump->um_quotas[i] != NULLVP) 2229 morework = 1; 2230 } 2231 UFS_UNLOCK(ump); 2232#endif 2233 if (morework) { 2234 if (--retry_flush_count > 0) { 2235 retry = 1; 2236 loopcnt = 3; 2237 } else 2238 error = EBUSY; 2239 } 2240 MNT_IUNLOCK(oldmnt); 2241 if (retry) 2242 goto retry_flush; 2243 } 2244 } 2245 return (error); 2246} 2247 2248/* 2249 * Structure hashing. 2250 * 2251 * There are four types of structures that can be looked up: 2252 * 1) pagedep structures identified by mount point, inode number, 2253 * and logical block. 2254 * 2) inodedep structures identified by mount point and inode number. 2255 * 3) newblk structures identified by mount point and 2256 * physical block number. 2257 * 4) bmsafemap structures identified by mount point and 2258 * cylinder group number. 2259 * 2260 * The "pagedep" and "inodedep" dependency structures are hashed 2261 * separately from the file blocks and inodes to which they correspond. 2262 * This separation helps when the in-memory copy of an inode or 2263 * file block must be replaced. It also obviates the need to access 2264 * an inode or file page when simply updating (or de-allocating) 2265 * dependency structures. Lookup of newblk structures is needed to 2266 * find newly allocated blocks when trying to associate them with 2267 * their allocdirect or allocindir structure. 2268 * 2269 * The lookup routines optionally create and hash a new instance when 2270 * an existing entry is not found. The bmsafemap lookup routine always 2271 * allocates a new structure if an existing one is not found. 2272 */ 2273#define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 2274 2275/* 2276 * Structures and routines associated with pagedep caching. 2277 */ 2278#define PAGEDEP_HASH(ump, inum, lbn) \ 2279 (&(ump)->pagedep_hashtbl[((inum) + (lbn)) & (ump)->pagedep_hash_size]) 2280 2281static int 2282pagedep_find(pagedephd, ino, lbn, pagedeppp) 2283 struct pagedep_hashhead *pagedephd; 2284 ino_t ino; 2285 ufs_lbn_t lbn; 2286 struct pagedep **pagedeppp; 2287{ 2288 struct pagedep *pagedep; 2289 2290 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 2291 if (ino == pagedep->pd_ino && lbn == pagedep->pd_lbn) { 2292 *pagedeppp = pagedep; 2293 return (1); 2294 } 2295 } 2296 *pagedeppp = NULL; 2297 return (0); 2298} 2299/* 2300 * Look up a pagedep. Return 1 if found, 0 otherwise. 2301 * If not found, allocate if DEPALLOC flag is passed. 2302 * Found or allocated entry is returned in pagedeppp. 2303 */ 2304static int 2305pagedep_lookup(mp, bp, ino, lbn, flags, pagedeppp) 2306 struct mount *mp; 2307 struct buf *bp; 2308 ino_t ino; 2309 ufs_lbn_t lbn; 2310 int flags; 2311 struct pagedep **pagedeppp; 2312{ 2313 struct pagedep *pagedep; 2314 struct pagedep_hashhead *pagedephd; 2315 struct worklist *wk; 2316 struct ufsmount *ump; 2317 int ret; 2318 int i; 2319 2320 ump = VFSTOUFS(mp); 2321 LOCK_OWNED(ump); 2322 if (bp) { 2323 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 2324 if (wk->wk_type == D_PAGEDEP) { 2325 *pagedeppp = WK_PAGEDEP(wk); 2326 return (1); 2327 } 2328 } 2329 } 2330 pagedephd = PAGEDEP_HASH(ump, ino, lbn); 2331 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp); 2332 if (ret) { 2333 if (((*pagedeppp)->pd_state & ONWORKLIST) == 0 && bp) 2334 WORKLIST_INSERT(&bp->b_dep, &(*pagedeppp)->pd_list); 2335 return (1); 2336 } 2337 if ((flags & DEPALLOC) == 0) 2338 return (0); 2339 FREE_LOCK(ump); 2340 pagedep = malloc(sizeof(struct pagedep), 2341 M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO); 2342 workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp); 2343 ACQUIRE_LOCK(ump); 2344 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp); 2345 if (*pagedeppp) { 2346 /* 2347 * This should never happen since we only create pagedeps 2348 * with the vnode lock held. Could be an assert. 2349 */ 2350 WORKITEM_FREE(pagedep, D_PAGEDEP); 2351 return (ret); 2352 } 2353 pagedep->pd_ino = ino; 2354 pagedep->pd_lbn = lbn; 2355 LIST_INIT(&pagedep->pd_dirremhd); 2356 LIST_INIT(&pagedep->pd_pendinghd); 2357 for (i = 0; i < DAHASHSZ; i++) 2358 LIST_INIT(&pagedep->pd_diraddhd[i]); 2359 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 2360 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2361 *pagedeppp = pagedep; 2362 return (0); 2363} 2364 2365/* 2366 * Structures and routines associated with inodedep caching. 2367 */ 2368#define INODEDEP_HASH(ump, inum) \ 2369 (&(ump)->inodedep_hashtbl[(inum) & (ump)->inodedep_hash_size]) 2370 2371static int 2372inodedep_find(inodedephd, inum, inodedeppp) 2373 struct inodedep_hashhead *inodedephd; 2374 ino_t inum; 2375 struct inodedep **inodedeppp; 2376{ 2377 struct inodedep *inodedep; 2378 2379 LIST_FOREACH(inodedep, inodedephd, id_hash) 2380 if (inum == inodedep->id_ino) 2381 break; 2382 if (inodedep) { 2383 *inodedeppp = inodedep; 2384 return (1); 2385 } 2386 *inodedeppp = NULL; 2387 2388 return (0); 2389} 2390/* 2391 * Look up an inodedep. Return 1 if found, 0 if not found. 2392 * If not found, allocate if DEPALLOC flag is passed. 2393 * Found or allocated entry is returned in inodedeppp. 2394 */ 2395static int 2396inodedep_lookup(mp, inum, flags, inodedeppp) 2397 struct mount *mp; 2398 ino_t inum; 2399 int flags; 2400 struct inodedep **inodedeppp; 2401{ 2402 struct inodedep *inodedep; 2403 struct inodedep_hashhead *inodedephd; 2404 struct ufsmount *ump; 2405 struct fs *fs; 2406 2407 ump = VFSTOUFS(mp); 2408 LOCK_OWNED(ump); 2409 fs = ump->um_fs; 2410 inodedephd = INODEDEP_HASH(ump, inum); 2411 2412 if (inodedep_find(inodedephd, inum, inodedeppp)) 2413 return (1); 2414 if ((flags & DEPALLOC) == 0) 2415 return (0); 2416 /* 2417 * If the system is over its limit and our filesystem is 2418 * responsible for more than our share of that usage and 2419 * we are not in a rush, request some inodedep cleanup. 2420 */ 2421 if (softdep_excess_items(ump, D_INODEDEP)) 2422 schedule_cleanup(mp); 2423 else 2424 FREE_LOCK(ump); 2425 inodedep = malloc(sizeof(struct inodedep), 2426 M_INODEDEP, M_SOFTDEP_FLAGS); 2427 workitem_alloc(&inodedep->id_list, D_INODEDEP, mp); 2428 ACQUIRE_LOCK(ump); 2429 if (inodedep_find(inodedephd, inum, inodedeppp)) { 2430 WORKITEM_FREE(inodedep, D_INODEDEP); 2431 return (1); 2432 } 2433 inodedep->id_fs = fs; 2434 inodedep->id_ino = inum; 2435 inodedep->id_state = ALLCOMPLETE; 2436 inodedep->id_nlinkdelta = 0; 2437 inodedep->id_nlinkwrote = -1; 2438 inodedep->id_savedino1 = NULL; 2439 inodedep->id_savedsize = -1; 2440 inodedep->id_savedextsize = -1; 2441 inodedep->id_savednlink = -1; 2442 inodedep->id_bmsafemap = NULL; 2443 inodedep->id_mkdiradd = NULL; 2444 LIST_INIT(&inodedep->id_dirremhd); 2445 LIST_INIT(&inodedep->id_pendinghd); 2446 LIST_INIT(&inodedep->id_inowait); 2447 LIST_INIT(&inodedep->id_bufwait); 2448 TAILQ_INIT(&inodedep->id_inoreflst); 2449 TAILQ_INIT(&inodedep->id_inoupdt); 2450 TAILQ_INIT(&inodedep->id_newinoupdt); 2451 TAILQ_INIT(&inodedep->id_extupdt); 2452 TAILQ_INIT(&inodedep->id_newextupdt); 2453 TAILQ_INIT(&inodedep->id_freeblklst); 2454 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 2455 *inodedeppp = inodedep; 2456 return (0); 2457} 2458 2459/* 2460 * Structures and routines associated with newblk caching. 2461 */ 2462#define NEWBLK_HASH(ump, inum) \ 2463 (&(ump)->newblk_hashtbl[(inum) & (ump)->newblk_hash_size]) 2464 2465static int 2466newblk_find(newblkhd, newblkno, flags, newblkpp) 2467 struct newblk_hashhead *newblkhd; 2468 ufs2_daddr_t newblkno; 2469 int flags; 2470 struct newblk **newblkpp; 2471{ 2472 struct newblk *newblk; 2473 2474 LIST_FOREACH(newblk, newblkhd, nb_hash) { 2475 if (newblkno != newblk->nb_newblkno) 2476 continue; 2477 /* 2478 * If we're creating a new dependency don't match those that 2479 * have already been converted to allocdirects. This is for 2480 * a frag extend. 2481 */ 2482 if ((flags & DEPALLOC) && newblk->nb_list.wk_type != D_NEWBLK) 2483 continue; 2484 break; 2485 } 2486 if (newblk) { 2487 *newblkpp = newblk; 2488 return (1); 2489 } 2490 *newblkpp = NULL; 2491 return (0); 2492} 2493 2494/* 2495 * Look up a newblk. Return 1 if found, 0 if not found. 2496 * If not found, allocate if DEPALLOC flag is passed. 2497 * Found or allocated entry is returned in newblkpp. 2498 */ 2499static int 2500newblk_lookup(mp, newblkno, flags, newblkpp) 2501 struct mount *mp; 2502 ufs2_daddr_t newblkno; 2503 int flags; 2504 struct newblk **newblkpp; 2505{ 2506 struct newblk *newblk; 2507 struct newblk_hashhead *newblkhd; 2508 struct ufsmount *ump; 2509 2510 ump = VFSTOUFS(mp); 2511 LOCK_OWNED(ump); 2512 newblkhd = NEWBLK_HASH(ump, newblkno); 2513 if (newblk_find(newblkhd, newblkno, flags, newblkpp)) 2514 return (1); 2515 if ((flags & DEPALLOC) == 0) 2516 return (0); 2517 if (softdep_excess_items(ump, D_NEWBLK) || 2518 softdep_excess_items(ump, D_ALLOCDIRECT) || 2519 softdep_excess_items(ump, D_ALLOCINDIR)) 2520 schedule_cleanup(mp); 2521 else 2522 FREE_LOCK(ump); 2523 newblk = malloc(sizeof(union allblk), M_NEWBLK, 2524 M_SOFTDEP_FLAGS | M_ZERO); 2525 workitem_alloc(&newblk->nb_list, D_NEWBLK, mp); 2526 ACQUIRE_LOCK(ump); 2527 if (newblk_find(newblkhd, newblkno, flags, newblkpp)) { 2528 WORKITEM_FREE(newblk, D_NEWBLK); 2529 return (1); 2530 } 2531 newblk->nb_freefrag = NULL; 2532 LIST_INIT(&newblk->nb_indirdeps); 2533 LIST_INIT(&newblk->nb_newdirblk); 2534 LIST_INIT(&newblk->nb_jwork); 2535 newblk->nb_state = ATTACHED; 2536 newblk->nb_newblkno = newblkno; 2537 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 2538 *newblkpp = newblk; 2539 return (0); 2540} 2541 2542/* 2543 * Structures and routines associated with freed indirect block caching. 2544 */ 2545#define INDIR_HASH(ump, blkno) \ 2546 (&(ump)->indir_hashtbl[(blkno) & (ump)->indir_hash_size]) 2547 2548/* 2549 * Lookup an indirect block in the indir hash table. The freework is 2550 * removed and potentially freed. The caller must do a blocking journal 2551 * write before writing to the blkno. 2552 */ 2553static int 2554indirblk_lookup(mp, blkno) 2555 struct mount *mp; 2556 ufs2_daddr_t blkno; 2557{ 2558 struct freework *freework; 2559 struct indir_hashhead *wkhd; 2560 struct ufsmount *ump; 2561 2562 ump = VFSTOUFS(mp); 2563 wkhd = INDIR_HASH(ump, blkno); 2564 TAILQ_FOREACH(freework, wkhd, fw_next) { 2565 if (freework->fw_blkno != blkno) 2566 continue; 2567 indirblk_remove(freework); 2568 return (1); 2569 } 2570 return (0); 2571} 2572 2573/* 2574 * Insert an indirect block represented by freework into the indirblk 2575 * hash table so that it may prevent the block from being re-used prior 2576 * to the journal being written. 2577 */ 2578static void 2579indirblk_insert(freework) 2580 struct freework *freework; 2581{ 2582 struct jblocks *jblocks; 2583 struct jseg *jseg; 2584 struct ufsmount *ump; 2585 2586 ump = VFSTOUFS(freework->fw_list.wk_mp); 2587 jblocks = ump->softdep_jblocks; 2588 jseg = TAILQ_LAST(&jblocks->jb_segs, jseglst); 2589 if (jseg == NULL) 2590 return; 2591 2592 LIST_INSERT_HEAD(&jseg->js_indirs, freework, fw_segs); 2593 TAILQ_INSERT_HEAD(INDIR_HASH(ump, freework->fw_blkno), freework, 2594 fw_next); 2595 freework->fw_state &= ~DEPCOMPLETE; 2596} 2597 2598static void 2599indirblk_remove(freework) 2600 struct freework *freework; 2601{ 2602 struct ufsmount *ump; 2603 2604 ump = VFSTOUFS(freework->fw_list.wk_mp); 2605 LIST_REMOVE(freework, fw_segs); 2606 TAILQ_REMOVE(INDIR_HASH(ump, freework->fw_blkno), freework, fw_next); 2607 freework->fw_state |= DEPCOMPLETE; 2608 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 2609 WORKITEM_FREE(freework, D_FREEWORK); 2610} 2611 2612/* 2613 * Executed during filesystem system initialization before 2614 * mounting any filesystems. 2615 */ 2616void 2617softdep_initialize() 2618{ 2619 2620 TAILQ_INIT(&softdepmounts); 2621#ifdef __LP64__ 2622 max_softdeps = desiredvnodes * 4; 2623#else 2624 max_softdeps = desiredvnodes * 2; 2625#endif 2626 2627 /* initialise bioops hack */ 2628 bioops.io_start = softdep_disk_io_initiation; 2629 bioops.io_complete = softdep_disk_write_complete; 2630 bioops.io_deallocate = softdep_deallocate_dependencies; 2631 bioops.io_countdeps = softdep_count_dependencies; 2632 softdep_ast_cleanup = softdep_ast_cleanup_proc; 2633 2634 /* Initialize the callout with an mtx. */ 2635 callout_init_mtx(&softdep_callout, &lk, 0); 2636} 2637 2638/* 2639 * Executed after all filesystems have been unmounted during 2640 * filesystem module unload. 2641 */ 2642void 2643softdep_uninitialize() 2644{ 2645 2646 /* clear bioops hack */ 2647 bioops.io_start = NULL; 2648 bioops.io_complete = NULL; 2649 bioops.io_deallocate = NULL; 2650 bioops.io_countdeps = NULL; 2651 softdep_ast_cleanup = NULL; 2652 2653 callout_drain(&softdep_callout); 2654} 2655 2656/* 2657 * Called at mount time to notify the dependency code that a 2658 * filesystem wishes to use it. 2659 */ 2660int 2661softdep_mount(devvp, mp, fs, cred) 2662 struct vnode *devvp; 2663 struct mount *mp; 2664 struct fs *fs; 2665 struct ucred *cred; 2666{ 2667 struct csum_total cstotal; 2668 struct mount_softdeps *sdp; 2669 struct ufsmount *ump; 2670 struct cg *cgp; 2671 struct buf *bp; 2672 u_int cyl, i; 2673 int error; 2674 2675 sdp = malloc(sizeof(struct mount_softdeps), M_MOUNTDATA, 2676 M_WAITOK | M_ZERO); 2677 MNT_ILOCK(mp); 2678 mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP; 2679 if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) { 2680 mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) | 2681 MNTK_SOFTDEP | MNTK_NOASYNC; 2682 } 2683 ump = VFSTOUFS(mp); 2684 ump->um_softdep = sdp; 2685 MNT_IUNLOCK(mp); 2686 rw_init(LOCK_PTR(ump), "per-fs softdep"); 2687 sdp->sd_ump = ump; 2688 LIST_INIT(&ump->softdep_workitem_pending); 2689 LIST_INIT(&ump->softdep_journal_pending); 2690 TAILQ_INIT(&ump->softdep_unlinked); 2691 LIST_INIT(&ump->softdep_dirtycg); 2692 ump->softdep_worklist_tail = NULL; 2693 ump->softdep_on_worklist = 0; 2694 ump->softdep_deps = 0; 2695 LIST_INIT(&ump->softdep_mkdirlisthd); 2696 ump->pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 2697 &ump->pagedep_hash_size); 2698 ump->pagedep_nextclean = 0; 2699 ump->inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, 2700 &ump->inodedep_hash_size); 2701 ump->inodedep_nextclean = 0; 2702 ump->newblk_hashtbl = hashinit(max_softdeps / 2, M_NEWBLK, 2703 &ump->newblk_hash_size); 2704 ump->bmsafemap_hashtbl = hashinit(1024, M_BMSAFEMAP, 2705 &ump->bmsafemap_hash_size); 2706 i = 1 << (ffs(desiredvnodes / 10) - 1); 2707 ump->indir_hashtbl = malloc(i * sizeof(struct indir_hashhead), 2708 M_FREEWORK, M_WAITOK); 2709 ump->indir_hash_size = i - 1; 2710 for (i = 0; i <= ump->indir_hash_size; i++) 2711 TAILQ_INIT(&ump->indir_hashtbl[i]); 2712#ifdef INVARIANTS 2713 for (i = 0; i <= D_LAST; i++) 2714 LIST_INIT(&ump->softdep_alldeps[i]); 2715#endif 2716 ACQUIRE_GBLLOCK(&lk); 2717 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next); 2718 FREE_GBLLOCK(&lk); 2719 if ((fs->fs_flags & FS_SUJ) && 2720 (error = journal_mount(mp, fs, cred)) != 0) { 2721 printf("Failed to start journal: %d\n", error); 2722 softdep_unmount(mp); 2723 return (error); 2724 } 2725 /* 2726 * Start our flushing thread in the bufdaemon process. 2727 */ 2728 ACQUIRE_LOCK(ump); 2729 ump->softdep_flags |= FLUSH_STARTING; 2730 FREE_LOCK(ump); 2731 kproc_kthread_add(&softdep_flush, mp, &bufdaemonproc, 2732 &ump->softdep_flushtd, 0, 0, "softdepflush", "%s worker", 2733 mp->mnt_stat.f_mntonname); 2734 ACQUIRE_LOCK(ump); 2735 while ((ump->softdep_flags & FLUSH_STARTING) != 0) { 2736 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM, "sdstart", 2737 hz / 2); 2738 } 2739 FREE_LOCK(ump); 2740 /* 2741 * When doing soft updates, the counters in the 2742 * superblock may have gotten out of sync. Recomputation 2743 * can take a long time and can be deferred for background 2744 * fsck. However, the old behavior of scanning the cylinder 2745 * groups and recalculating them at mount time is available 2746 * by setting vfs.ffs.compute_summary_at_mount to one. 2747 */ 2748 if (compute_summary_at_mount == 0 || fs->fs_clean != 0) 2749 return (0); 2750 bzero(&cstotal, sizeof cstotal); 2751 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 2752 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 2753 fs->fs_cgsize, cred, &bp)) != 0) { 2754 brelse(bp); 2755 softdep_unmount(mp); 2756 return (error); 2757 } 2758 cgp = (struct cg *)bp->b_data; 2759 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 2760 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 2761 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 2762 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 2763 fs->fs_cs(fs, cyl) = cgp->cg_cs; 2764 brelse(bp); 2765 } 2766#ifdef INVARIANTS 2767 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 2768 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt); 2769#endif 2770 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 2771 return (0); 2772} 2773 2774void 2775softdep_unmount(mp) 2776 struct mount *mp; 2777{ 2778 struct ufsmount *ump; 2779#ifdef INVARIANTS 2780 int i; 2781#endif 2782 2783 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 2784 ("softdep_unmount called on non-softdep filesystem")); 2785 ump = VFSTOUFS(mp); 2786 MNT_ILOCK(mp); 2787 mp->mnt_flag &= ~MNT_SOFTDEP; 2788 if (MOUNTEDSUJ(mp) == 0) { 2789 MNT_IUNLOCK(mp); 2790 } else { 2791 mp->mnt_flag &= ~MNT_SUJ; 2792 MNT_IUNLOCK(mp); 2793 journal_unmount(ump); 2794 } 2795 /* 2796 * Shut down our flushing thread. Check for NULL is if 2797 * softdep_mount errors out before the thread has been created. 2798 */ 2799 if (ump->softdep_flushtd != NULL) { 2800 ACQUIRE_LOCK(ump); 2801 ump->softdep_flags |= FLUSH_EXIT; 2802 wakeup(&ump->softdep_flushtd); 2803 msleep(&ump->softdep_flags, LOCK_PTR(ump), PVM | PDROP, 2804 "sdwait", 0); 2805 KASSERT((ump->softdep_flags & FLUSH_EXIT) == 0, 2806 ("Thread shutdown failed")); 2807 } 2808 /* 2809 * Free up our resources. 2810 */ 2811 ACQUIRE_GBLLOCK(&lk); 2812 TAILQ_REMOVE(&softdepmounts, ump->um_softdep, sd_next); 2813 FREE_GBLLOCK(&lk); 2814 rw_destroy(LOCK_PTR(ump)); 2815 hashdestroy(ump->pagedep_hashtbl, M_PAGEDEP, ump->pagedep_hash_size); 2816 hashdestroy(ump->inodedep_hashtbl, M_INODEDEP, ump->inodedep_hash_size); 2817 hashdestroy(ump->newblk_hashtbl, M_NEWBLK, ump->newblk_hash_size); 2818 hashdestroy(ump->bmsafemap_hashtbl, M_BMSAFEMAP, 2819 ump->bmsafemap_hash_size); 2820 free(ump->indir_hashtbl, M_FREEWORK); 2821#ifdef INVARIANTS 2822 for (i = 0; i <= D_LAST; i++) { 2823 KASSERT(ump->softdep_curdeps[i] == 0, 2824 ("Unmount %s: Dep type %s != 0 (%ld)", ump->um_fs->fs_fsmnt, 2825 TYPENAME(i), ump->softdep_curdeps[i])); 2826 KASSERT(LIST_EMPTY(&ump->softdep_alldeps[i]), 2827 ("Unmount %s: Dep type %s not empty (%p)", ump->um_fs->fs_fsmnt, 2828 TYPENAME(i), LIST_FIRST(&ump->softdep_alldeps[i]))); 2829 } 2830#endif 2831 free(ump->um_softdep, M_MOUNTDATA); 2832} 2833 2834static struct jblocks * 2835jblocks_create(void) 2836{ 2837 struct jblocks *jblocks; 2838 2839 jblocks = malloc(sizeof(*jblocks), M_JBLOCKS, M_WAITOK | M_ZERO); 2840 TAILQ_INIT(&jblocks->jb_segs); 2841 jblocks->jb_avail = 10; 2842 jblocks->jb_extent = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2843 M_JBLOCKS, M_WAITOK | M_ZERO); 2844 2845 return (jblocks); 2846} 2847 2848static ufs2_daddr_t 2849jblocks_alloc(jblocks, bytes, actual) 2850 struct jblocks *jblocks; 2851 int bytes; 2852 int *actual; 2853{ 2854 ufs2_daddr_t daddr; 2855 struct jextent *jext; 2856 int freecnt; 2857 int blocks; 2858 2859 blocks = bytes / DEV_BSIZE; 2860 jext = &jblocks->jb_extent[jblocks->jb_head]; 2861 freecnt = jext->je_blocks - jblocks->jb_off; 2862 if (freecnt == 0) { 2863 jblocks->jb_off = 0; 2864 if (++jblocks->jb_head > jblocks->jb_used) 2865 jblocks->jb_head = 0; 2866 jext = &jblocks->jb_extent[jblocks->jb_head]; 2867 freecnt = jext->je_blocks; 2868 } 2869 if (freecnt > blocks) 2870 freecnt = blocks; 2871 *actual = freecnt * DEV_BSIZE; 2872 daddr = jext->je_daddr + jblocks->jb_off; 2873 jblocks->jb_off += freecnt; 2874 jblocks->jb_free -= freecnt; 2875 2876 return (daddr); 2877} 2878 2879static void 2880jblocks_free(jblocks, mp, bytes) 2881 struct jblocks *jblocks; 2882 struct mount *mp; 2883 int bytes; 2884{ 2885 2886 LOCK_OWNED(VFSTOUFS(mp)); 2887 jblocks->jb_free += bytes / DEV_BSIZE; 2888 if (jblocks->jb_suspended) 2889 worklist_speedup(mp); 2890 wakeup(jblocks); 2891} 2892 2893static void 2894jblocks_destroy(jblocks) 2895 struct jblocks *jblocks; 2896{ 2897 2898 if (jblocks->jb_extent) 2899 free(jblocks->jb_extent, M_JBLOCKS); 2900 free(jblocks, M_JBLOCKS); 2901} 2902 2903static void 2904jblocks_add(jblocks, daddr, blocks) 2905 struct jblocks *jblocks; 2906 ufs2_daddr_t daddr; 2907 int blocks; 2908{ 2909 struct jextent *jext; 2910 2911 jblocks->jb_blocks += blocks; 2912 jblocks->jb_free += blocks; 2913 jext = &jblocks->jb_extent[jblocks->jb_used]; 2914 /* Adding the first block. */ 2915 if (jext->je_daddr == 0) { 2916 jext->je_daddr = daddr; 2917 jext->je_blocks = blocks; 2918 return; 2919 } 2920 /* Extending the last extent. */ 2921 if (jext->je_daddr + jext->je_blocks == daddr) { 2922 jext->je_blocks += blocks; 2923 return; 2924 } 2925 /* Adding a new extent. */ 2926 if (++jblocks->jb_used == jblocks->jb_avail) { 2927 jblocks->jb_avail *= 2; 2928 jext = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2929 M_JBLOCKS, M_WAITOK | M_ZERO); 2930 memcpy(jext, jblocks->jb_extent, 2931 sizeof(struct jextent) * jblocks->jb_used); 2932 free(jblocks->jb_extent, M_JBLOCKS); 2933 jblocks->jb_extent = jext; 2934 } 2935 jext = &jblocks->jb_extent[jblocks->jb_used]; 2936 jext->je_daddr = daddr; 2937 jext->je_blocks = blocks; 2938 return; 2939} 2940 2941int 2942softdep_journal_lookup(mp, vpp) 2943 struct mount *mp; 2944 struct vnode **vpp; 2945{ 2946 struct componentname cnp; 2947 struct vnode *dvp; 2948 ino_t sujournal; 2949 int error; 2950 2951 error = VFS_VGET(mp, UFS_ROOTINO, LK_EXCLUSIVE, &dvp); 2952 if (error) 2953 return (error); 2954 bzero(&cnp, sizeof(cnp)); 2955 cnp.cn_nameiop = LOOKUP; 2956 cnp.cn_flags = ISLASTCN; 2957 cnp.cn_thread = curthread; 2958 cnp.cn_cred = curthread->td_ucred; 2959 cnp.cn_pnbuf = SUJ_FILE; 2960 cnp.cn_nameptr = SUJ_FILE; 2961 cnp.cn_namelen = strlen(SUJ_FILE); 2962 error = ufs_lookup_ino(dvp, NULL, &cnp, &sujournal); 2963 vput(dvp); 2964 if (error != 0) 2965 return (error); 2966 error = VFS_VGET(mp, sujournal, LK_EXCLUSIVE, vpp); 2967 return (error); 2968} 2969 2970/* 2971 * Open and verify the journal file. 2972 */ 2973static int 2974journal_mount(mp, fs, cred) 2975 struct mount *mp; 2976 struct fs *fs; 2977 struct ucred *cred; 2978{ 2979 struct jblocks *jblocks; 2980 struct ufsmount *ump; 2981 struct vnode *vp; 2982 struct inode *ip; 2983 ufs2_daddr_t blkno; 2984 int bcount; 2985 int error; 2986 int i; 2987 2988 ump = VFSTOUFS(mp); 2989 ump->softdep_journal_tail = NULL; 2990 ump->softdep_on_journal = 0; 2991 ump->softdep_accdeps = 0; 2992 ump->softdep_req = 0; 2993 ump->softdep_jblocks = NULL; 2994 error = softdep_journal_lookup(mp, &vp); 2995 if (error != 0) { 2996 printf("Failed to find journal. Use tunefs to create one\n"); 2997 return (error); 2998 } 2999 ip = VTOI(vp); 3000 if (ip->i_size < SUJ_MIN) { 3001 error = ENOSPC; 3002 goto out; 3003 } 3004 bcount = lblkno(fs, ip->i_size); /* Only use whole blocks. */ 3005 jblocks = jblocks_create(); 3006 for (i = 0; i < bcount; i++) { 3007 error = ufs_bmaparray(vp, i, &blkno, NULL, NULL, NULL); 3008 if (error) 3009 break; 3010 jblocks_add(jblocks, blkno, fsbtodb(fs, fs->fs_frag)); 3011 } 3012 if (error) { 3013 jblocks_destroy(jblocks); 3014 goto out; 3015 } 3016 jblocks->jb_low = jblocks->jb_free / 3; /* Reserve 33%. */ 3017 jblocks->jb_min = jblocks->jb_free / 10; /* Suspend at 10%. */ 3018 ump->softdep_jblocks = jblocks; 3019out: 3020 if (error == 0) { 3021 MNT_ILOCK(mp); 3022 mp->mnt_flag |= MNT_SUJ; 3023 mp->mnt_flag &= ~MNT_SOFTDEP; 3024 MNT_IUNLOCK(mp); 3025 /* 3026 * Only validate the journal contents if the 3027 * filesystem is clean, otherwise we write the logs 3028 * but they'll never be used. If the filesystem was 3029 * still dirty when we mounted it the journal is 3030 * invalid and a new journal can only be valid if it 3031 * starts from a clean mount. 3032 */ 3033 if (fs->fs_clean) { 3034 DIP_SET(ip, i_modrev, fs->fs_mtime); 3035 ip->i_flags |= IN_MODIFIED; 3036 ffs_update(vp, 1); 3037 } 3038 } 3039 vput(vp); 3040 return (error); 3041} 3042 3043static void 3044journal_unmount(ump) 3045 struct ufsmount *ump; 3046{ 3047 3048 if (ump->softdep_jblocks) 3049 jblocks_destroy(ump->softdep_jblocks); 3050 ump->softdep_jblocks = NULL; 3051} 3052 3053/* 3054 * Called when a journal record is ready to be written. Space is allocated 3055 * and the journal entry is created when the journal is flushed to stable 3056 * store. 3057 */ 3058static void 3059add_to_journal(wk) 3060 struct worklist *wk; 3061{ 3062 struct ufsmount *ump; 3063 3064 ump = VFSTOUFS(wk->wk_mp); 3065 LOCK_OWNED(ump); 3066 if (wk->wk_state & ONWORKLIST) 3067 panic("add_to_journal: %s(0x%X) already on list", 3068 TYPENAME(wk->wk_type), wk->wk_state); 3069 wk->wk_state |= ONWORKLIST | DEPCOMPLETE; 3070 if (LIST_EMPTY(&ump->softdep_journal_pending)) { 3071 ump->softdep_jblocks->jb_age = ticks; 3072 LIST_INSERT_HEAD(&ump->softdep_journal_pending, wk, wk_list); 3073 } else 3074 LIST_INSERT_AFTER(ump->softdep_journal_tail, wk, wk_list); 3075 ump->softdep_journal_tail = wk; 3076 ump->softdep_on_journal += 1; 3077} 3078 3079/* 3080 * Remove an arbitrary item for the journal worklist maintain the tail 3081 * pointer. This happens when a new operation obviates the need to 3082 * journal an old operation. 3083 */ 3084static void 3085remove_from_journal(wk) 3086 struct worklist *wk; 3087{ 3088 struct ufsmount *ump; 3089 3090 ump = VFSTOUFS(wk->wk_mp); 3091 LOCK_OWNED(ump); 3092#ifdef INVARIANTS 3093 { 3094 struct worklist *wkn; 3095 3096 LIST_FOREACH(wkn, &ump->softdep_journal_pending, wk_list) 3097 if (wkn == wk) 3098 break; 3099 if (wkn == NULL) 3100 panic("remove_from_journal: %p is not in journal", wk); 3101 } 3102#endif 3103 /* 3104 * We emulate a TAILQ to save space in most structures which do not 3105 * require TAILQ semantics. Here we must update the tail position 3106 * when removing the tail which is not the final entry. This works 3107 * only if the worklist linkage are at the beginning of the structure. 3108 */ 3109 if (ump->softdep_journal_tail == wk) 3110 ump->softdep_journal_tail = 3111 (struct worklist *)wk->wk_list.le_prev; 3112 WORKLIST_REMOVE(wk); 3113 ump->softdep_on_journal -= 1; 3114} 3115 3116/* 3117 * Check for journal space as well as dependency limits so the prelink 3118 * code can throttle both journaled and non-journaled filesystems. 3119 * Threshold is 0 for low and 1 for min. 3120 */ 3121static int 3122journal_space(ump, thresh) 3123 struct ufsmount *ump; 3124 int thresh; 3125{ 3126 struct jblocks *jblocks; 3127 int limit, avail; 3128 3129 jblocks = ump->softdep_jblocks; 3130 if (jblocks == NULL) 3131 return (1); 3132 /* 3133 * We use a tighter restriction here to prevent request_cleanup() 3134 * running in threads from running into locks we currently hold. 3135 * We have to be over the limit and our filesystem has to be 3136 * responsible for more than our share of that usage. 3137 */ 3138 limit = (max_softdeps / 10) * 9; 3139 if (dep_current[D_INODEDEP] > limit && 3140 ump->softdep_curdeps[D_INODEDEP] > limit / stat_flush_threads) 3141 return (0); 3142 if (thresh) 3143 thresh = jblocks->jb_min; 3144 else 3145 thresh = jblocks->jb_low; 3146 avail = (ump->softdep_on_journal * JREC_SIZE) / DEV_BSIZE; 3147 avail = jblocks->jb_free - avail; 3148 3149 return (avail > thresh); 3150} 3151 3152static void 3153journal_suspend(ump) 3154 struct ufsmount *ump; 3155{ 3156 struct jblocks *jblocks; 3157 struct mount *mp; 3158 bool set; 3159 3160 mp = UFSTOVFS(ump); 3161 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) 3162 return; 3163 3164 jblocks = ump->softdep_jblocks; 3165 vfs_op_enter(mp); 3166 set = false; 3167 MNT_ILOCK(mp); 3168 if ((mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { 3169 stat_journal_min++; 3170 mp->mnt_kern_flag |= MNTK_SUSPEND; 3171 mp->mnt_susp_owner = ump->softdep_flushtd; 3172 set = true; 3173 } 3174 jblocks->jb_suspended = 1; 3175 MNT_IUNLOCK(mp); 3176 if (!set) 3177 vfs_op_exit(mp); 3178} 3179 3180static int 3181journal_unsuspend(struct ufsmount *ump) 3182{ 3183 struct jblocks *jblocks; 3184 struct mount *mp; 3185 3186 mp = UFSTOVFS(ump); 3187 jblocks = ump->softdep_jblocks; 3188 3189 if (jblocks != NULL && jblocks->jb_suspended && 3190 journal_space(ump, jblocks->jb_min)) { 3191 jblocks->jb_suspended = 0; 3192 FREE_LOCK(ump); 3193 mp->mnt_susp_owner = curthread; 3194 vfs_write_resume(mp, 0); 3195 ACQUIRE_LOCK(ump); 3196 return (1); 3197 } 3198 return (0); 3199} 3200 3201/* 3202 * Called before any allocation function to be certain that there is 3203 * sufficient space in the journal prior to creating any new records. 3204 * Since in the case of block allocation we may have multiple locked 3205 * buffers at the time of the actual allocation we can not block 3206 * when the journal records are created. Doing so would create a deadlock 3207 * if any of these buffers needed to be flushed to reclaim space. Instead 3208 * we require a sufficiently large amount of available space such that 3209 * each thread in the system could have passed this allocation check and 3210 * still have sufficient free space. With 20% of a minimum journal size 3211 * of 1MB we have 6553 records available. 3212 */ 3213int 3214softdep_prealloc(vp, waitok) 3215 struct vnode *vp; 3216 int waitok; 3217{ 3218 struct ufsmount *ump; 3219 3220 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 3221 ("softdep_prealloc called on non-softdep filesystem")); 3222 /* 3223 * Nothing to do if we are not running journaled soft updates. 3224 * If we currently hold the snapshot lock, we must avoid 3225 * handling other resources that could cause deadlock. Do not 3226 * touch quotas vnode since it is typically recursed with 3227 * other vnode locks held. 3228 */ 3229 if (DOINGSUJ(vp) == 0 || IS_SNAPSHOT(VTOI(vp)) || 3230 (vp->v_vflag & VV_SYSTEM) != 0) 3231 return (0); 3232 ump = VFSTOUFS(vp->v_mount); 3233 ACQUIRE_LOCK(ump); 3234 if (journal_space(ump, 0)) { 3235 FREE_LOCK(ump); 3236 return (0); 3237 } 3238 stat_journal_low++; 3239 FREE_LOCK(ump); 3240 if (waitok == MNT_NOWAIT) 3241 return (ENOSPC); 3242 /* 3243 * Attempt to sync this vnode once to flush any journal 3244 * work attached to it. 3245 */ 3246 if ((curthread->td_pflags & TDP_COWINPROGRESS) == 0) 3247 ffs_syncvnode(vp, waitok, 0); 3248 ACQUIRE_LOCK(ump); 3249 process_removes(vp); 3250 process_truncates(vp); 3251 if (journal_space(ump, 0) == 0) { 3252 softdep_speedup(ump); 3253 if (journal_space(ump, 1) == 0) 3254 journal_suspend(ump); 3255 } 3256 FREE_LOCK(ump); 3257 3258 return (0); 3259} 3260 3261/* 3262 * Try hard to sync all data and metadata for the vnode, and workitems 3263 * flushing which might conflict with the vnode lock. This is a 3264 * helper for softdep_prerename(). 3265 */ 3266static int 3267softdep_prerename_vnode(ump, vp) 3268 struct ufsmount *ump; 3269 struct vnode *vp; 3270{ 3271 int error; 3272 3273 ASSERT_VOP_ELOCKED(vp, "prehandle"); 3274 if (vp->v_data == NULL) 3275 return (0); 3276 error = VOP_FSYNC(vp, MNT_WAIT, curthread); 3277 if (error != 0) 3278 return (error); 3279 ACQUIRE_LOCK(ump); 3280 process_removes(vp); 3281 process_truncates(vp); 3282 FREE_LOCK(ump); 3283 return (0); 3284} 3285 3286/* 3287 * Must be called from VOP_RENAME() after all vnodes are locked. 3288 * Ensures that there is enough journal space for rename. It is 3289 * sufficiently different from softdep_prelink() by having to handle 3290 * four vnodes. 3291 */ 3292int 3293softdep_prerename(fdvp, fvp, tdvp, tvp) 3294 struct vnode *fdvp; 3295 struct vnode *fvp; 3296 struct vnode *tdvp; 3297 struct vnode *tvp; 3298{ 3299 struct ufsmount *ump; 3300 int error; 3301 3302 ump = VFSTOUFS(fdvp->v_mount); 3303 3304 if (journal_space(ump, 0)) 3305 return (0); 3306 3307 VOP_UNLOCK(tdvp); 3308 VOP_UNLOCK(fvp); 3309 if (tvp != NULL && tvp != tdvp) 3310 VOP_UNLOCK(tvp); 3311 3312 error = softdep_prerename_vnode(ump, fdvp); 3313 VOP_UNLOCK(fdvp); 3314 if (error != 0) 3315 return (error); 3316 3317 VOP_LOCK(fvp, LK_EXCLUSIVE | LK_RETRY); 3318 error = softdep_prerename_vnode(ump, fvp); 3319 VOP_UNLOCK(fvp); 3320 if (error != 0) 3321 return (error); 3322 3323 if (tdvp != fdvp) { 3324 VOP_LOCK(tdvp, LK_EXCLUSIVE | LK_RETRY); 3325 error = softdep_prerename_vnode(ump, tdvp); 3326 VOP_UNLOCK(tdvp); 3327 if (error != 0) 3328 return (error); 3329 } 3330 3331 if (tvp != fvp && tvp != NULL) { 3332 VOP_LOCK(tvp, LK_EXCLUSIVE | LK_RETRY); 3333 error = softdep_prerename_vnode(ump, tvp); 3334 VOP_UNLOCK(tvp); 3335 if (error != 0) 3336 return (error); 3337 } 3338 3339 ACQUIRE_LOCK(ump); 3340 softdep_speedup(ump); 3341 process_worklist_item(UFSTOVFS(ump), 2, LK_NOWAIT); 3342 if (journal_space(ump, 0) == 0) { 3343 softdep_speedup(ump); 3344 if (journal_space(ump, 1) == 0) 3345 journal_suspend(ump); 3346 } 3347 FREE_LOCK(ump); 3348 return (ERELOOKUP); 3349} 3350 3351/* 3352 * Before adjusting a link count on a vnode verify that we have sufficient 3353 * journal space. If not, process operations that depend on the currently 3354 * locked pair of vnodes to try to flush space as the syncer, buf daemon, 3355 * and softdep flush threads can not acquire these locks to reclaim space. 3356 * 3357 * Returns 0 if all owned locks are still valid and were not dropped 3358 * in the process, in other case it returns either an error from sync, 3359 * or ERELOOKUP if any of the locks were re-acquired. In the later 3360 * case, the state of the vnodes cannot be relied upon and our VFS 3361 * syscall must be restarted at top level from the lookup. 3362 */ 3363int 3364softdep_prelink(dvp, vp) 3365 struct vnode *dvp; 3366 struct vnode *vp; 3367{ 3368 struct ufsmount *ump; 3369 3370 ASSERT_VOP_ELOCKED(dvp, "prelink dvp"); 3371 if (vp != NULL) 3372 ASSERT_VOP_ELOCKED(vp, "prelink vp"); 3373 ump = VFSTOUFS(dvp->v_mount); 3374 3375 /* 3376 * Nothing to do if we have sufficient journal space. We skip 3377 * flushing when vp is a snapshot to avoid deadlock where 3378 * another thread is trying to update the inodeblock for dvp 3379 * and is waiting on snaplk that vp holds. 3380 */ 3381 if (journal_space(ump, 0) || (vp != NULL && IS_SNAPSHOT(VTOI(vp)))) 3382 return (0); 3383 3384 stat_journal_low++; 3385 if (vp != NULL) { 3386 VOP_UNLOCK(dvp); 3387 ffs_syncvnode(vp, MNT_NOWAIT, 0); 3388 vn_lock_pair(dvp, false, vp, true); 3389 if (dvp->v_data == NULL) 3390 return (ERELOOKUP); 3391 } 3392 if (vp != NULL) 3393 VOP_UNLOCK(vp); 3394 ffs_syncvnode(dvp, MNT_WAIT, 0); 3395 VOP_UNLOCK(dvp); 3396 3397 /* Process vp before dvp as it may create .. removes. */ 3398 if (vp != NULL) { 3399 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 3400 if (vp->v_data == NULL) { 3401 vn_lock_pair(dvp, false, vp, true); 3402 return (ERELOOKUP); 3403 } 3404 ACQUIRE_LOCK(ump); 3405 process_removes(vp); 3406 process_truncates(vp); 3407 FREE_LOCK(ump); 3408 VOP_UNLOCK(vp); 3409 } 3410 3411 vn_lock(dvp, LK_EXCLUSIVE | LK_RETRY); 3412 if (dvp->v_data == NULL) { 3413 vn_lock_pair(dvp, true, vp, false); 3414 return (ERELOOKUP); 3415 } 3416 3417 ACQUIRE_LOCK(ump); 3418 process_removes(dvp); 3419 process_truncates(dvp); 3420 VOP_UNLOCK(dvp); 3421 softdep_speedup(ump); 3422 3423 process_worklist_item(UFSTOVFS(ump), 2, LK_NOWAIT); 3424 if (journal_space(ump, 0) == 0) { 3425 softdep_speedup(ump); 3426 if (journal_space(ump, 1) == 0) 3427 journal_suspend(ump); 3428 } 3429 FREE_LOCK(ump); 3430 3431 vn_lock_pair(dvp, false, vp, false); 3432 return (ERELOOKUP); 3433} 3434 3435static void 3436jseg_write(ump, jseg, data) 3437 struct ufsmount *ump; 3438 struct jseg *jseg; 3439 uint8_t *data; 3440{ 3441 struct jsegrec *rec; 3442 3443 rec = (struct jsegrec *)data; 3444 rec->jsr_seq = jseg->js_seq; 3445 rec->jsr_oldest = jseg->js_oldseq; 3446 rec->jsr_cnt = jseg->js_cnt; 3447 rec->jsr_blocks = jseg->js_size / ump->um_devvp->v_bufobj.bo_bsize; 3448 rec->jsr_crc = 0; 3449 rec->jsr_time = ump->um_fs->fs_mtime; 3450} 3451 3452static inline void 3453inoref_write(inoref, jseg, rec) 3454 struct inoref *inoref; 3455 struct jseg *jseg; 3456 struct jrefrec *rec; 3457{ 3458 3459 inoref->if_jsegdep->jd_seg = jseg; 3460 rec->jr_ino = inoref->if_ino; 3461 rec->jr_parent = inoref->if_parent; 3462 rec->jr_nlink = inoref->if_nlink; 3463 rec->jr_mode = inoref->if_mode; 3464 rec->jr_diroff = inoref->if_diroff; 3465} 3466 3467static void 3468jaddref_write(jaddref, jseg, data) 3469 struct jaddref *jaddref; 3470 struct jseg *jseg; 3471 uint8_t *data; 3472{ 3473 struct jrefrec *rec; 3474 3475 rec = (struct jrefrec *)data; 3476 rec->jr_op = JOP_ADDREF; 3477 inoref_write(&jaddref->ja_ref, jseg, rec); 3478} 3479 3480static void 3481jremref_write(jremref, jseg, data) 3482 struct jremref *jremref; 3483 struct jseg *jseg; 3484 uint8_t *data; 3485{ 3486 struct jrefrec *rec; 3487 3488 rec = (struct jrefrec *)data; 3489 rec->jr_op = JOP_REMREF; 3490 inoref_write(&jremref->jr_ref, jseg, rec); 3491} 3492 3493static void 3494jmvref_write(jmvref, jseg, data) 3495 struct jmvref *jmvref; 3496 struct jseg *jseg; 3497 uint8_t *data; 3498{ 3499 struct jmvrec *rec; 3500 3501 rec = (struct jmvrec *)data; 3502 rec->jm_op = JOP_MVREF; 3503 rec->jm_ino = jmvref->jm_ino; 3504 rec->jm_parent = jmvref->jm_parent; 3505 rec->jm_oldoff = jmvref->jm_oldoff; 3506 rec->jm_newoff = jmvref->jm_newoff; 3507} 3508 3509static void 3510jnewblk_write(jnewblk, jseg, data) 3511 struct jnewblk *jnewblk; 3512 struct jseg *jseg; 3513 uint8_t *data; 3514{ 3515 struct jblkrec *rec; 3516 3517 jnewblk->jn_jsegdep->jd_seg = jseg; 3518 rec = (struct jblkrec *)data; 3519 rec->jb_op = JOP_NEWBLK; 3520 rec->jb_ino = jnewblk->jn_ino; 3521 rec->jb_blkno = jnewblk->jn_blkno; 3522 rec->jb_lbn = jnewblk->jn_lbn; 3523 rec->jb_frags = jnewblk->jn_frags; 3524 rec->jb_oldfrags = jnewblk->jn_oldfrags; 3525} 3526 3527static void 3528jfreeblk_write(jfreeblk, jseg, data) 3529 struct jfreeblk *jfreeblk; 3530 struct jseg *jseg; 3531 uint8_t *data; 3532{ 3533 struct jblkrec *rec; 3534 3535 jfreeblk->jf_dep.jb_jsegdep->jd_seg = jseg; 3536 rec = (struct jblkrec *)data; 3537 rec->jb_op = JOP_FREEBLK; 3538 rec->jb_ino = jfreeblk->jf_ino; 3539 rec->jb_blkno = jfreeblk->jf_blkno; 3540 rec->jb_lbn = jfreeblk->jf_lbn; 3541 rec->jb_frags = jfreeblk->jf_frags; 3542 rec->jb_oldfrags = 0; 3543} 3544 3545static void 3546jfreefrag_write(jfreefrag, jseg, data) 3547 struct jfreefrag *jfreefrag; 3548 struct jseg *jseg; 3549 uint8_t *data; 3550{ 3551 struct jblkrec *rec; 3552 3553 jfreefrag->fr_jsegdep->jd_seg = jseg; 3554 rec = (struct jblkrec *)data; 3555 rec->jb_op = JOP_FREEBLK; 3556 rec->jb_ino = jfreefrag->fr_ino; 3557 rec->jb_blkno = jfreefrag->fr_blkno; 3558 rec->jb_lbn = jfreefrag->fr_lbn; 3559 rec->jb_frags = jfreefrag->fr_frags; 3560 rec->jb_oldfrags = 0; 3561} 3562 3563static void 3564jtrunc_write(jtrunc, jseg, data) 3565 struct jtrunc *jtrunc; 3566 struct jseg *jseg; 3567 uint8_t *data; 3568{ 3569 struct jtrncrec *rec; 3570 3571 jtrunc->jt_dep.jb_jsegdep->jd_seg = jseg; 3572 rec = (struct jtrncrec *)data; 3573 rec->jt_op = JOP_TRUNC; 3574 rec->jt_ino = jtrunc->jt_ino; 3575 rec->jt_size = jtrunc->jt_size; 3576 rec->jt_extsize = jtrunc->jt_extsize; 3577} 3578 3579static void 3580jfsync_write(jfsync, jseg, data) 3581 struct jfsync *jfsync; 3582 struct jseg *jseg; 3583 uint8_t *data; 3584{ 3585 struct jtrncrec *rec; 3586 3587 rec = (struct jtrncrec *)data; 3588 rec->jt_op = JOP_SYNC; 3589 rec->jt_ino = jfsync->jfs_ino; 3590 rec->jt_size = jfsync->jfs_size; 3591 rec->jt_extsize = jfsync->jfs_extsize; 3592} 3593 3594static void 3595softdep_flushjournal(mp) 3596 struct mount *mp; 3597{ 3598 struct jblocks *jblocks; 3599 struct ufsmount *ump; 3600 3601 if (MOUNTEDSUJ(mp) == 0) 3602 return; 3603 ump = VFSTOUFS(mp); 3604 jblocks = ump->softdep_jblocks; 3605 ACQUIRE_LOCK(ump); 3606 while (ump->softdep_on_journal) { 3607 jblocks->jb_needseg = 1; 3608 softdep_process_journal(mp, NULL, MNT_WAIT); 3609 } 3610 FREE_LOCK(ump); 3611} 3612 3613static void softdep_synchronize_completed(struct bio *); 3614static void softdep_synchronize(struct bio *, struct ufsmount *, void *); 3615 3616static void 3617softdep_synchronize_completed(bp) 3618 struct bio *bp; 3619{ 3620 struct jseg *oldest; 3621 struct jseg *jseg; 3622 struct ufsmount *ump; 3623 3624 /* 3625 * caller1 marks the last segment written before we issued the 3626 * synchronize cache. 3627 */ 3628 jseg = bp->bio_caller1; 3629 if (jseg == NULL) { 3630 g_destroy_bio(bp); 3631 return; 3632 } 3633 ump = VFSTOUFS(jseg->js_list.wk_mp); 3634 ACQUIRE_LOCK(ump); 3635 oldest = NULL; 3636 /* 3637 * Mark all the journal entries waiting on the synchronize cache 3638 * as completed so they may continue on. 3639 */ 3640 while (jseg != NULL && (jseg->js_state & COMPLETE) == 0) { 3641 jseg->js_state |= COMPLETE; 3642 oldest = jseg; 3643 jseg = TAILQ_PREV(jseg, jseglst, js_next); 3644 } 3645 /* 3646 * Restart deferred journal entry processing from the oldest 3647 * completed jseg. 3648 */ 3649 if (oldest) 3650 complete_jsegs(oldest); 3651 3652 FREE_LOCK(ump); 3653 g_destroy_bio(bp); 3654} 3655 3656/* 3657 * Send BIO_FLUSH/SYNCHRONIZE CACHE to the device to enforce write ordering 3658 * barriers. The journal must be written prior to any blocks that depend 3659 * on it and the journal can not be released until the blocks have be 3660 * written. This code handles both barriers simultaneously. 3661 */ 3662static void 3663softdep_synchronize(bp, ump, caller1) 3664 struct bio *bp; 3665 struct ufsmount *ump; 3666 void *caller1; 3667{ 3668 3669 bp->bio_cmd = BIO_FLUSH; 3670 bp->bio_flags |= BIO_ORDERED; 3671 bp->bio_data = NULL; 3672 bp->bio_offset = ump->um_cp->provider->mediasize; 3673 bp->bio_length = 0; 3674 bp->bio_done = softdep_synchronize_completed; 3675 bp->bio_caller1 = caller1; 3676 g_io_request(bp, ump->um_cp); 3677} 3678 3679/* 3680 * Flush some journal records to disk. 3681 */ 3682static void 3683softdep_process_journal(mp, needwk, flags) 3684 struct mount *mp; 3685 struct worklist *needwk; 3686 int flags; 3687{ 3688 struct jblocks *jblocks; 3689 struct ufsmount *ump; 3690 struct worklist *wk; 3691 struct jseg *jseg; 3692 struct buf *bp; 3693 struct bio *bio; 3694 uint8_t *data; 3695 struct fs *fs; 3696 int shouldflush; 3697 int segwritten; 3698 int jrecmin; /* Minimum records per block. */ 3699 int jrecmax; /* Maximum records per block. */ 3700 int size; 3701 int cnt; 3702 int off; 3703 int devbsize; 3704 3705 if (MOUNTEDSUJ(mp) == 0) 3706 return; 3707 shouldflush = softdep_flushcache; 3708 bio = NULL; 3709 jseg = NULL; 3710 ump = VFSTOUFS(mp); 3711 LOCK_OWNED(ump); 3712 fs = ump->um_fs; 3713 jblocks = ump->softdep_jblocks; 3714 devbsize = ump->um_devvp->v_bufobj.bo_bsize; 3715 /* 3716 * We write anywhere between a disk block and fs block. The upper 3717 * bound is picked to prevent buffer cache fragmentation and limit 3718 * processing time per I/O. 3719 */ 3720 jrecmin = (devbsize / JREC_SIZE) - 1; /* -1 for seg header */ 3721 jrecmax = (fs->fs_bsize / devbsize) * jrecmin; 3722 segwritten = 0; 3723 for (;;) { 3724 cnt = ump->softdep_on_journal; 3725 /* 3726 * Criteria for writing a segment: 3727 * 1) We have a full block. 3728 * 2) We're called from jwait() and haven't found the 3729 * journal item yet. 3730 * 3) Always write if needseg is set. 3731 * 4) If we are called from process_worklist and have 3732 * not yet written anything we write a partial block 3733 * to enforce a 1 second maximum latency on journal 3734 * entries. 3735 */ 3736 if (cnt < (jrecmax - 1) && needwk == NULL && 3737 jblocks->jb_needseg == 0 && (segwritten || cnt == 0)) 3738 break; 3739 cnt++; 3740 /* 3741 * Verify some free journal space. softdep_prealloc() should 3742 * guarantee that we don't run out so this is indicative of 3743 * a problem with the flow control. Try to recover 3744 * gracefully in any event. 3745 */ 3746 while (jblocks->jb_free == 0) { 3747 if (flags != MNT_WAIT) 3748 break; 3749 printf("softdep: Out of journal space!\n"); 3750 softdep_speedup(ump); 3751 msleep(jblocks, LOCK_PTR(ump), PRIBIO, "jblocks", hz); 3752 } 3753 FREE_LOCK(ump); 3754 jseg = malloc(sizeof(*jseg), M_JSEG, M_SOFTDEP_FLAGS); 3755 workitem_alloc(&jseg->js_list, D_JSEG, mp); 3756 LIST_INIT(&jseg->js_entries); 3757 LIST_INIT(&jseg->js_indirs); 3758 jseg->js_state = ATTACHED; 3759 if (shouldflush == 0) 3760 jseg->js_state |= COMPLETE; 3761 else if (bio == NULL) 3762 bio = g_alloc_bio(); 3763 jseg->js_jblocks = jblocks; 3764 bp = geteblk(fs->fs_bsize, 0); 3765 ACQUIRE_LOCK(ump); 3766 /* 3767 * If there was a race while we were allocating the block 3768 * and jseg the entry we care about was likely written. 3769 * We bail out in both the WAIT and NOWAIT case and assume 3770 * the caller will loop if the entry it cares about is 3771 * not written. 3772 */ 3773 cnt = ump->softdep_on_journal; 3774 if (cnt + jblocks->jb_needseg == 0 || jblocks->jb_free == 0) { 3775 bp->b_flags |= B_INVAL | B_NOCACHE; 3776 WORKITEM_FREE(jseg, D_JSEG); 3777 FREE_LOCK(ump); 3778 brelse(bp); 3779 ACQUIRE_LOCK(ump); 3780 break; 3781 } 3782 /* 3783 * Calculate the disk block size required for the available 3784 * records rounded to the min size. 3785 */ 3786 if (cnt == 0) 3787 size = devbsize; 3788 else if (cnt < jrecmax) 3789 size = howmany(cnt, jrecmin) * devbsize; 3790 else 3791 size = fs->fs_bsize; 3792 /* 3793 * Allocate a disk block for this journal data and account 3794 * for truncation of the requested size if enough contiguous 3795 * space was not available. 3796 */ 3797 bp->b_blkno = jblocks_alloc(jblocks, size, &size); 3798 bp->b_lblkno = bp->b_blkno; 3799 bp->b_offset = bp->b_blkno * DEV_BSIZE; 3800 bp->b_bcount = size; 3801 bp->b_flags &= ~B_INVAL; 3802 bp->b_flags |= B_VALIDSUSPWRT | B_NOCOPY; 3803 /* 3804 * Initialize our jseg with cnt records. Assign the next 3805 * sequence number to it and link it in-order. 3806 */ 3807 cnt = MIN(cnt, (size / devbsize) * jrecmin); 3808 jseg->js_buf = bp; 3809 jseg->js_cnt = cnt; 3810 jseg->js_refs = cnt + 1; /* Self ref. */ 3811 jseg->js_size = size; 3812 jseg->js_seq = jblocks->jb_nextseq++; 3813 if (jblocks->jb_oldestseg == NULL) 3814 jblocks->jb_oldestseg = jseg; 3815 jseg->js_oldseq = jblocks->jb_oldestseg->js_seq; 3816 TAILQ_INSERT_TAIL(&jblocks->jb_segs, jseg, js_next); 3817 if (jblocks->jb_writeseg == NULL) 3818 jblocks->jb_writeseg = jseg; 3819 /* 3820 * Start filling in records from the pending list. 3821 */ 3822 data = bp->b_data; 3823 off = 0; 3824 3825 /* 3826 * Always put a header on the first block. 3827 * XXX As with below, there might not be a chance to get 3828 * into the loop. Ensure that something valid is written. 3829 */ 3830 jseg_write(ump, jseg, data); 3831 off += JREC_SIZE; 3832 data = bp->b_data + off; 3833 3834 /* 3835 * XXX Something is wrong here. There's no work to do, 3836 * but we need to perform and I/O and allow it to complete 3837 * anyways. 3838 */ 3839 if (LIST_EMPTY(&ump->softdep_journal_pending)) 3840 stat_emptyjblocks++; 3841 3842 while ((wk = LIST_FIRST(&ump->softdep_journal_pending)) 3843 != NULL) { 3844 if (cnt == 0) 3845 break; 3846 /* Place a segment header on every device block. */ 3847 if ((off % devbsize) == 0) { 3848 jseg_write(ump, jseg, data); 3849 off += JREC_SIZE; 3850 data = bp->b_data + off; 3851 } 3852 if (wk == needwk) 3853 needwk = NULL; 3854 remove_from_journal(wk); 3855 wk->wk_state |= INPROGRESS; 3856 WORKLIST_INSERT(&jseg->js_entries, wk); 3857 switch (wk->wk_type) { 3858 case D_JADDREF: 3859 jaddref_write(WK_JADDREF(wk), jseg, data); 3860 break; 3861 case D_JREMREF: 3862 jremref_write(WK_JREMREF(wk), jseg, data); 3863 break; 3864 case D_JMVREF: 3865 jmvref_write(WK_JMVREF(wk), jseg, data); 3866 break; 3867 case D_JNEWBLK: 3868 jnewblk_write(WK_JNEWBLK(wk), jseg, data); 3869 break; 3870 case D_JFREEBLK: 3871 jfreeblk_write(WK_JFREEBLK(wk), jseg, data); 3872 break; 3873 case D_JFREEFRAG: 3874 jfreefrag_write(WK_JFREEFRAG(wk), jseg, data); 3875 break; 3876 case D_JTRUNC: 3877 jtrunc_write(WK_JTRUNC(wk), jseg, data); 3878 break; 3879 case D_JFSYNC: 3880 jfsync_write(WK_JFSYNC(wk), jseg, data); 3881 break; 3882 default: 3883 panic("process_journal: Unknown type %s", 3884 TYPENAME(wk->wk_type)); 3885 /* NOTREACHED */ 3886 } 3887 off += JREC_SIZE; 3888 data = bp->b_data + off; 3889 cnt--; 3890 } 3891 3892 /* Clear any remaining space so we don't leak kernel data */ 3893 if (size > off) 3894 bzero(data, size - off); 3895 3896 /* 3897 * Write this one buffer and continue. 3898 */ 3899 segwritten = 1; 3900 jblocks->jb_needseg = 0; 3901 WORKLIST_INSERT(&bp->b_dep, &jseg->js_list); 3902 FREE_LOCK(ump); 3903 bp->b_xflags |= BX_CVTENXIO; 3904 pbgetvp(ump->um_devvp, bp); 3905 /* 3906 * We only do the blocking wait once we find the journal 3907 * entry we're looking for. 3908 */ 3909 if (needwk == NULL && flags == MNT_WAIT) 3910 bwrite(bp); 3911 else 3912 bawrite(bp); 3913 ACQUIRE_LOCK(ump); 3914 } 3915 /* 3916 * If we wrote a segment issue a synchronize cache so the journal 3917 * is reflected on disk before the data is written. Since reclaiming 3918 * journal space also requires writing a journal record this 3919 * process also enforces a barrier before reclamation. 3920 */ 3921 if (segwritten && shouldflush) { 3922 softdep_synchronize(bio, ump, 3923 TAILQ_LAST(&jblocks->jb_segs, jseglst)); 3924 } else if (bio) 3925 g_destroy_bio(bio); 3926 /* 3927 * If we've suspended the filesystem because we ran out of journal 3928 * space either try to sync it here to make some progress or 3929 * unsuspend it if we already have. 3930 */ 3931 if (flags == 0 && jblocks->jb_suspended) { 3932 if (journal_unsuspend(ump)) 3933 return; 3934 FREE_LOCK(ump); 3935 VFS_SYNC(mp, MNT_NOWAIT); 3936 ffs_sbupdate(ump, MNT_WAIT, 0); 3937 ACQUIRE_LOCK(ump); 3938 } 3939} 3940 3941/* 3942 * Complete a jseg, allowing all dependencies awaiting journal writes 3943 * to proceed. Each journal dependency also attaches a jsegdep to dependent 3944 * structures so that the journal segment can be freed to reclaim space. 3945 */ 3946static void 3947complete_jseg(jseg) 3948 struct jseg *jseg; 3949{ 3950 struct worklist *wk; 3951 struct jmvref *jmvref; 3952#ifdef INVARIANTS 3953 int i = 0; 3954#endif 3955 3956 while ((wk = LIST_FIRST(&jseg->js_entries)) != NULL) { 3957 WORKLIST_REMOVE(wk); 3958 wk->wk_state &= ~INPROGRESS; 3959 wk->wk_state |= COMPLETE; 3960 KASSERT(i++ < jseg->js_cnt, 3961 ("handle_written_jseg: overflow %d >= %d", 3962 i - 1, jseg->js_cnt)); 3963 switch (wk->wk_type) { 3964 case D_JADDREF: 3965 handle_written_jaddref(WK_JADDREF(wk)); 3966 break; 3967 case D_JREMREF: 3968 handle_written_jremref(WK_JREMREF(wk)); 3969 break; 3970 case D_JMVREF: 3971 rele_jseg(jseg); /* No jsegdep. */ 3972 jmvref = WK_JMVREF(wk); 3973 LIST_REMOVE(jmvref, jm_deps); 3974 if ((jmvref->jm_pagedep->pd_state & ONWORKLIST) == 0) 3975 free_pagedep(jmvref->jm_pagedep); 3976 WORKITEM_FREE(jmvref, D_JMVREF); 3977 break; 3978 case D_JNEWBLK: 3979 handle_written_jnewblk(WK_JNEWBLK(wk)); 3980 break; 3981 case D_JFREEBLK: 3982 handle_written_jblkdep(&WK_JFREEBLK(wk)->jf_dep); 3983 break; 3984 case D_JTRUNC: 3985 handle_written_jblkdep(&WK_JTRUNC(wk)->jt_dep); 3986 break; 3987 case D_JFSYNC: 3988 rele_jseg(jseg); /* No jsegdep. */ 3989 WORKITEM_FREE(wk, D_JFSYNC); 3990 break; 3991 case D_JFREEFRAG: 3992 handle_written_jfreefrag(WK_JFREEFRAG(wk)); 3993 break; 3994 default: 3995 panic("handle_written_jseg: Unknown type %s", 3996 TYPENAME(wk->wk_type)); 3997 /* NOTREACHED */ 3998 } 3999 } 4000 /* Release the self reference so the structure may be freed. */ 4001 rele_jseg(jseg); 4002} 4003 4004/* 4005 * Determine which jsegs are ready for completion processing. Waits for 4006 * synchronize cache to complete as well as forcing in-order completion 4007 * of journal entries. 4008 */ 4009static void 4010complete_jsegs(jseg) 4011 struct jseg *jseg; 4012{ 4013 struct jblocks *jblocks; 4014 struct jseg *jsegn; 4015 4016 jblocks = jseg->js_jblocks; 4017 /* 4018 * Don't allow out of order completions. If this isn't the first 4019 * block wait for it to write before we're done. 4020 */ 4021 if (jseg != jblocks->jb_writeseg) 4022 return; 4023 /* Iterate through available jsegs processing their entries. */ 4024 while (jseg && (jseg->js_state & ALLCOMPLETE) == ALLCOMPLETE) { 4025 jblocks->jb_oldestwrseq = jseg->js_oldseq; 4026 jsegn = TAILQ_NEXT(jseg, js_next); 4027 complete_jseg(jseg); 4028 jseg = jsegn; 4029 } 4030 jblocks->jb_writeseg = jseg; 4031 /* 4032 * Attempt to free jsegs now that oldestwrseq may have advanced. 4033 */ 4034 free_jsegs(jblocks); 4035} 4036 4037/* 4038 * Mark a jseg as DEPCOMPLETE and throw away the buffer. Attempt to handle 4039 * the final completions. 4040 */ 4041static void 4042handle_written_jseg(jseg, bp) 4043 struct jseg *jseg; 4044 struct buf *bp; 4045{ 4046 4047 if (jseg->js_refs == 0) 4048 panic("handle_written_jseg: No self-reference on %p", jseg); 4049 jseg->js_state |= DEPCOMPLETE; 4050 /* 4051 * We'll never need this buffer again, set flags so it will be 4052 * discarded. 4053 */ 4054 bp->b_flags |= B_INVAL | B_NOCACHE; 4055 pbrelvp(bp); 4056 complete_jsegs(jseg); 4057} 4058 4059static inline struct jsegdep * 4060inoref_jseg(inoref) 4061 struct inoref *inoref; 4062{ 4063 struct jsegdep *jsegdep; 4064 4065 jsegdep = inoref->if_jsegdep; 4066 inoref->if_jsegdep = NULL; 4067 4068 return (jsegdep); 4069} 4070 4071/* 4072 * Called once a jremref has made it to stable store. The jremref is marked 4073 * complete and we attempt to free it. Any pagedeps writes sleeping waiting 4074 * for the jremref to complete will be awoken by free_jremref. 4075 */ 4076static void 4077handle_written_jremref(jremref) 4078 struct jremref *jremref; 4079{ 4080 struct inodedep *inodedep; 4081 struct jsegdep *jsegdep; 4082 struct dirrem *dirrem; 4083 4084 /* Grab the jsegdep. */ 4085 jsegdep = inoref_jseg(&jremref->jr_ref); 4086 /* 4087 * Remove us from the inoref list. 4088 */ 4089 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 4090 0, &inodedep) == 0) 4091 panic("handle_written_jremref: Lost inodedep"); 4092 TAILQ_REMOVE(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 4093 /* 4094 * Complete the dirrem. 4095 */ 4096 dirrem = jremref->jr_dirrem; 4097 jremref->jr_dirrem = NULL; 4098 LIST_REMOVE(jremref, jr_deps); 4099 jsegdep->jd_state |= jremref->jr_state & MKDIR_PARENT; 4100 jwork_insert(&dirrem->dm_jwork, jsegdep); 4101 if (LIST_EMPTY(&dirrem->dm_jremrefhd) && 4102 (dirrem->dm_state & COMPLETE) != 0) 4103 add_to_worklist(&dirrem->dm_list, 0); 4104 free_jremref(jremref); 4105} 4106 4107/* 4108 * Called once a jaddref has made it to stable store. The dependency is 4109 * marked complete and any dependent structures are added to the inode 4110 * bufwait list to be completed as soon as it is written. If a bitmap write 4111 * depends on this entry we move the inode into the inodedephd of the 4112 * bmsafemap dependency and attempt to remove the jaddref from the bmsafemap. 4113 */ 4114static void 4115handle_written_jaddref(jaddref) 4116 struct jaddref *jaddref; 4117{ 4118 struct jsegdep *jsegdep; 4119 struct inodedep *inodedep; 4120 struct diradd *diradd; 4121 struct mkdir *mkdir; 4122 4123 /* Grab the jsegdep. */ 4124 jsegdep = inoref_jseg(&jaddref->ja_ref); 4125 mkdir = NULL; 4126 diradd = NULL; 4127 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 4128 0, &inodedep) == 0) 4129 panic("handle_written_jaddref: Lost inodedep."); 4130 if (jaddref->ja_diradd == NULL) 4131 panic("handle_written_jaddref: No dependency"); 4132 if (jaddref->ja_diradd->da_list.wk_type == D_DIRADD) { 4133 diradd = jaddref->ja_diradd; 4134 WORKLIST_INSERT(&inodedep->id_bufwait, &diradd->da_list); 4135 } else if (jaddref->ja_state & MKDIR_PARENT) { 4136 mkdir = jaddref->ja_mkdir; 4137 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir->md_list); 4138 } else if (jaddref->ja_state & MKDIR_BODY) 4139 mkdir = jaddref->ja_mkdir; 4140 else 4141 panic("handle_written_jaddref: Unknown dependency %p", 4142 jaddref->ja_diradd); 4143 jaddref->ja_diradd = NULL; /* also clears ja_mkdir */ 4144 /* 4145 * Remove us from the inode list. 4146 */ 4147 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, if_deps); 4148 /* 4149 * The mkdir may be waiting on the jaddref to clear before freeing. 4150 */ 4151 if (mkdir) { 4152 KASSERT(mkdir->md_list.wk_type == D_MKDIR, 4153 ("handle_written_jaddref: Incorrect type for mkdir %s", 4154 TYPENAME(mkdir->md_list.wk_type))); 4155 mkdir->md_jaddref = NULL; 4156 diradd = mkdir->md_diradd; 4157 mkdir->md_state |= DEPCOMPLETE; 4158 complete_mkdir(mkdir); 4159 } 4160 jwork_insert(&diradd->da_jwork, jsegdep); 4161 if (jaddref->ja_state & NEWBLOCK) { 4162 inodedep->id_state |= ONDEPLIST; 4163 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_inodedephd, 4164 inodedep, id_deps); 4165 } 4166 free_jaddref(jaddref); 4167} 4168 4169/* 4170 * Called once a jnewblk journal is written. The allocdirect or allocindir 4171 * is placed in the bmsafemap to await notification of a written bitmap. If 4172 * the operation was canceled we add the segdep to the appropriate 4173 * dependency to free the journal space once the canceling operation 4174 * completes. 4175 */ 4176static void 4177handle_written_jnewblk(jnewblk) 4178 struct jnewblk *jnewblk; 4179{ 4180 struct bmsafemap *bmsafemap; 4181 struct freefrag *freefrag; 4182 struct freework *freework; 4183 struct jsegdep *jsegdep; 4184 struct newblk *newblk; 4185 4186 /* Grab the jsegdep. */ 4187 jsegdep = jnewblk->jn_jsegdep; 4188 jnewblk->jn_jsegdep = NULL; 4189 if (jnewblk->jn_dep == NULL) 4190 panic("handle_written_jnewblk: No dependency for the segdep."); 4191 switch (jnewblk->jn_dep->wk_type) { 4192 case D_NEWBLK: 4193 case D_ALLOCDIRECT: 4194 case D_ALLOCINDIR: 4195 /* 4196 * Add the written block to the bmsafemap so it can 4197 * be notified when the bitmap is on disk. 4198 */ 4199 newblk = WK_NEWBLK(jnewblk->jn_dep); 4200 newblk->nb_jnewblk = NULL; 4201 if ((newblk->nb_state & GOINGAWAY) == 0) { 4202 bmsafemap = newblk->nb_bmsafemap; 4203 newblk->nb_state |= ONDEPLIST; 4204 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, 4205 nb_deps); 4206 } 4207 jwork_insert(&newblk->nb_jwork, jsegdep); 4208 break; 4209 case D_FREEFRAG: 4210 /* 4211 * A newblock being removed by a freefrag when replaced by 4212 * frag extension. 4213 */ 4214 freefrag = WK_FREEFRAG(jnewblk->jn_dep); 4215 freefrag->ff_jdep = NULL; 4216 jwork_insert(&freefrag->ff_jwork, jsegdep); 4217 break; 4218 case D_FREEWORK: 4219 /* 4220 * A direct block was removed by truncate. 4221 */ 4222 freework = WK_FREEWORK(jnewblk->jn_dep); 4223 freework->fw_jnewblk = NULL; 4224 jwork_insert(&freework->fw_freeblks->fb_jwork, jsegdep); 4225 break; 4226 default: 4227 panic("handle_written_jnewblk: Unknown type %d.", 4228 jnewblk->jn_dep->wk_type); 4229 } 4230 jnewblk->jn_dep = NULL; 4231 free_jnewblk(jnewblk); 4232} 4233 4234/* 4235 * Cancel a jfreefrag that won't be needed, probably due to colliding with 4236 * an in-flight allocation that has not yet been committed. Divorce us 4237 * from the freefrag and mark it DEPCOMPLETE so that it may be added 4238 * to the worklist. 4239 */ 4240static void 4241cancel_jfreefrag(jfreefrag) 4242 struct jfreefrag *jfreefrag; 4243{ 4244 struct freefrag *freefrag; 4245 4246 if (jfreefrag->fr_jsegdep) { 4247 free_jsegdep(jfreefrag->fr_jsegdep); 4248 jfreefrag->fr_jsegdep = NULL; 4249 } 4250 freefrag = jfreefrag->fr_freefrag; 4251 jfreefrag->fr_freefrag = NULL; 4252 free_jfreefrag(jfreefrag); 4253 freefrag->ff_state |= DEPCOMPLETE; 4254 CTR1(KTR_SUJ, "cancel_jfreefrag: blkno %jd", freefrag->ff_blkno); 4255} 4256 4257/* 4258 * Free a jfreefrag when the parent freefrag is rendered obsolete. 4259 */ 4260static void 4261free_jfreefrag(jfreefrag) 4262 struct jfreefrag *jfreefrag; 4263{ 4264 4265 if (jfreefrag->fr_state & INPROGRESS) 4266 WORKLIST_REMOVE(&jfreefrag->fr_list); 4267 else if (jfreefrag->fr_state & ONWORKLIST) 4268 remove_from_journal(&jfreefrag->fr_list); 4269 if (jfreefrag->fr_freefrag != NULL) 4270 panic("free_jfreefrag: Still attached to a freefrag."); 4271 WORKITEM_FREE(jfreefrag, D_JFREEFRAG); 4272} 4273 4274/* 4275 * Called when the journal write for a jfreefrag completes. The parent 4276 * freefrag is added to the worklist if this completes its dependencies. 4277 */ 4278static void 4279handle_written_jfreefrag(jfreefrag) 4280 struct jfreefrag *jfreefrag; 4281{ 4282 struct jsegdep *jsegdep; 4283 struct freefrag *freefrag; 4284 4285 /* Grab the jsegdep. */ 4286 jsegdep = jfreefrag->fr_jsegdep; 4287 jfreefrag->fr_jsegdep = NULL; 4288 freefrag = jfreefrag->fr_freefrag; 4289 if (freefrag == NULL) 4290 panic("handle_written_jfreefrag: No freefrag."); 4291 freefrag->ff_state |= DEPCOMPLETE; 4292 freefrag->ff_jdep = NULL; 4293 jwork_insert(&freefrag->ff_jwork, jsegdep); 4294 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 4295 add_to_worklist(&freefrag->ff_list, 0); 4296 jfreefrag->fr_freefrag = NULL; 4297 free_jfreefrag(jfreefrag); 4298} 4299 4300/* 4301 * Called when the journal write for a jfreeblk completes. The jfreeblk 4302 * is removed from the freeblks list of pending journal writes and the 4303 * jsegdep is moved to the freeblks jwork to be completed when all blocks 4304 * have been reclaimed. 4305 */ 4306static void 4307handle_written_jblkdep(jblkdep) 4308 struct jblkdep *jblkdep; 4309{ 4310 struct freeblks *freeblks; 4311 struct jsegdep *jsegdep; 4312 4313 /* Grab the jsegdep. */ 4314 jsegdep = jblkdep->jb_jsegdep; 4315 jblkdep->jb_jsegdep = NULL; 4316 freeblks = jblkdep->jb_freeblks; 4317 LIST_REMOVE(jblkdep, jb_deps); 4318 jwork_insert(&freeblks->fb_jwork, jsegdep); 4319 /* 4320 * If the freeblks is all journaled, we can add it to the worklist. 4321 */ 4322 if (LIST_EMPTY(&freeblks->fb_jblkdephd) && 4323 (freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 4324 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 4325 4326 free_jblkdep(jblkdep); 4327} 4328 4329static struct jsegdep * 4330newjsegdep(struct worklist *wk) 4331{ 4332 struct jsegdep *jsegdep; 4333 4334 jsegdep = malloc(sizeof(*jsegdep), M_JSEGDEP, M_SOFTDEP_FLAGS); 4335 workitem_alloc(&jsegdep->jd_list, D_JSEGDEP, wk->wk_mp); 4336 jsegdep->jd_seg = NULL; 4337 4338 return (jsegdep); 4339} 4340 4341static struct jmvref * 4342newjmvref(dp, ino, oldoff, newoff) 4343 struct inode *dp; 4344 ino_t ino; 4345 off_t oldoff; 4346 off_t newoff; 4347{ 4348 struct jmvref *jmvref; 4349 4350 jmvref = malloc(sizeof(*jmvref), M_JMVREF, M_SOFTDEP_FLAGS); 4351 workitem_alloc(&jmvref->jm_list, D_JMVREF, ITOVFS(dp)); 4352 jmvref->jm_list.wk_state = ATTACHED | DEPCOMPLETE; 4353 jmvref->jm_parent = dp->i_number; 4354 jmvref->jm_ino = ino; 4355 jmvref->jm_oldoff = oldoff; 4356 jmvref->jm_newoff = newoff; 4357 4358 return (jmvref); 4359} 4360 4361/* 4362 * Allocate a new jremref that tracks the removal of ip from dp with the 4363 * directory entry offset of diroff. Mark the entry as ATTACHED and 4364 * DEPCOMPLETE as we have all the information required for the journal write 4365 * and the directory has already been removed from the buffer. The caller 4366 * is responsible for linking the jremref into the pagedep and adding it 4367 * to the journal to write. The MKDIR_PARENT flag is set if we're doing 4368 * a DOTDOT addition so handle_workitem_remove() can properly assign 4369 * the jsegdep when we're done. 4370 */ 4371static struct jremref * 4372newjremref(struct dirrem *dirrem, struct inode *dp, struct inode *ip, 4373 off_t diroff, nlink_t nlink) 4374{ 4375 struct jremref *jremref; 4376 4377 jremref = malloc(sizeof(*jremref), M_JREMREF, M_SOFTDEP_FLAGS); 4378 workitem_alloc(&jremref->jr_list, D_JREMREF, ITOVFS(dp)); 4379 jremref->jr_state = ATTACHED; 4380 newinoref(&jremref->jr_ref, ip->i_number, dp->i_number, diroff, 4381 nlink, ip->i_mode); 4382 jremref->jr_dirrem = dirrem; 4383 4384 return (jremref); 4385} 4386 4387static inline void 4388newinoref(struct inoref *inoref, ino_t ino, ino_t parent, off_t diroff, 4389 nlink_t nlink, uint16_t mode) 4390{ 4391 4392 inoref->if_jsegdep = newjsegdep(&inoref->if_list); 4393 inoref->if_diroff = diroff; 4394 inoref->if_ino = ino; 4395 inoref->if_parent = parent; 4396 inoref->if_nlink = nlink; 4397 inoref->if_mode = mode; 4398} 4399 4400/* 4401 * Allocate a new jaddref to track the addition of ino to dp at diroff. The 4402 * directory offset may not be known until later. The caller is responsible 4403 * adding the entry to the journal when this information is available. nlink 4404 * should be the link count prior to the addition and mode is only required 4405 * to have the correct FMT. 4406 */ 4407static struct jaddref * 4408newjaddref(struct inode *dp, ino_t ino, off_t diroff, int16_t nlink, 4409 uint16_t mode) 4410{ 4411 struct jaddref *jaddref; 4412 4413 jaddref = malloc(sizeof(*jaddref), M_JADDREF, M_SOFTDEP_FLAGS); 4414 workitem_alloc(&jaddref->ja_list, D_JADDREF, ITOVFS(dp)); 4415 jaddref->ja_state = ATTACHED; 4416 jaddref->ja_mkdir = NULL; 4417 newinoref(&jaddref->ja_ref, ino, dp->i_number, diroff, nlink, mode); 4418 4419 return (jaddref); 4420} 4421 4422/* 4423 * Create a new free dependency for a freework. The caller is responsible 4424 * for adjusting the reference count when it has the lock held. The freedep 4425 * will track an outstanding bitmap write that will ultimately clear the 4426 * freework to continue. 4427 */ 4428static struct freedep * 4429newfreedep(struct freework *freework) 4430{ 4431 struct freedep *freedep; 4432 4433 freedep = malloc(sizeof(*freedep), M_FREEDEP, M_SOFTDEP_FLAGS); 4434 workitem_alloc(&freedep->fd_list, D_FREEDEP, freework->fw_list.wk_mp); 4435 freedep->fd_freework = freework; 4436 4437 return (freedep); 4438} 4439 4440/* 4441 * Free a freedep structure once the buffer it is linked to is written. If 4442 * this is the last reference to the freework schedule it for completion. 4443 */ 4444static void 4445free_freedep(freedep) 4446 struct freedep *freedep; 4447{ 4448 struct freework *freework; 4449 4450 freework = freedep->fd_freework; 4451 freework->fw_freeblks->fb_cgwait--; 4452 if (--freework->fw_ref == 0) 4453 freework_enqueue(freework); 4454 WORKITEM_FREE(freedep, D_FREEDEP); 4455} 4456 4457/* 4458 * Allocate a new freework structure that may be a level in an indirect 4459 * when parent is not NULL or a top level block when it is. The top level 4460 * freework structures are allocated without the per-filesystem lock held 4461 * and before the freeblks is visible outside of softdep_setup_freeblocks(). 4462 */ 4463static struct freework * 4464newfreework(ump, freeblks, parent, lbn, nb, frags, off, journal) 4465 struct ufsmount *ump; 4466 struct freeblks *freeblks; 4467 struct freework *parent; 4468 ufs_lbn_t lbn; 4469 ufs2_daddr_t nb; 4470 int frags; 4471 int off; 4472 int journal; 4473{ 4474 struct freework *freework; 4475 4476 freework = malloc(sizeof(*freework), M_FREEWORK, M_SOFTDEP_FLAGS); 4477 workitem_alloc(&freework->fw_list, D_FREEWORK, freeblks->fb_list.wk_mp); 4478 freework->fw_state = ATTACHED; 4479 freework->fw_jnewblk = NULL; 4480 freework->fw_freeblks = freeblks; 4481 freework->fw_parent = parent; 4482 freework->fw_lbn = lbn; 4483 freework->fw_blkno = nb; 4484 freework->fw_frags = frags; 4485 freework->fw_indir = NULL; 4486 freework->fw_ref = (MOUNTEDSUJ(UFSTOVFS(ump)) == 0 || 4487 lbn >= -UFS_NXADDR) ? 0 : NINDIR(ump->um_fs) + 1; 4488 freework->fw_start = freework->fw_off = off; 4489 if (journal) 4490 newjfreeblk(freeblks, lbn, nb, frags); 4491 if (parent == NULL) { 4492 ACQUIRE_LOCK(ump); 4493 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 4494 freeblks->fb_ref++; 4495 FREE_LOCK(ump); 4496 } 4497 4498 return (freework); 4499} 4500 4501/* 4502 * Eliminate a jfreeblk for a block that does not need journaling. 4503 */ 4504static void 4505cancel_jfreeblk(freeblks, blkno) 4506 struct freeblks *freeblks; 4507 ufs2_daddr_t blkno; 4508{ 4509 struct jfreeblk *jfreeblk; 4510 struct jblkdep *jblkdep; 4511 4512 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) { 4513 if (jblkdep->jb_list.wk_type != D_JFREEBLK) 4514 continue; 4515 jfreeblk = WK_JFREEBLK(&jblkdep->jb_list); 4516 if (jfreeblk->jf_blkno == blkno) 4517 break; 4518 } 4519 if (jblkdep == NULL) 4520 return; 4521 CTR1(KTR_SUJ, "cancel_jfreeblk: blkno %jd", blkno); 4522 free_jsegdep(jblkdep->jb_jsegdep); 4523 LIST_REMOVE(jblkdep, jb_deps); 4524 WORKITEM_FREE(jfreeblk, D_JFREEBLK); 4525} 4526 4527/* 4528 * Allocate a new jfreeblk to journal top level block pointer when truncating 4529 * a file. The caller must add this to the worklist when the per-filesystem 4530 * lock is held. 4531 */ 4532static struct jfreeblk * 4533newjfreeblk(freeblks, lbn, blkno, frags) 4534 struct freeblks *freeblks; 4535 ufs_lbn_t lbn; 4536 ufs2_daddr_t blkno; 4537 int frags; 4538{ 4539 struct jfreeblk *jfreeblk; 4540 4541 jfreeblk = malloc(sizeof(*jfreeblk), M_JFREEBLK, M_SOFTDEP_FLAGS); 4542 workitem_alloc(&jfreeblk->jf_dep.jb_list, D_JFREEBLK, 4543 freeblks->fb_list.wk_mp); 4544 jfreeblk->jf_dep.jb_jsegdep = newjsegdep(&jfreeblk->jf_dep.jb_list); 4545 jfreeblk->jf_dep.jb_freeblks = freeblks; 4546 jfreeblk->jf_ino = freeblks->fb_inum; 4547 jfreeblk->jf_lbn = lbn; 4548 jfreeblk->jf_blkno = blkno; 4549 jfreeblk->jf_frags = frags; 4550 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jfreeblk->jf_dep, jb_deps); 4551 4552 return (jfreeblk); 4553} 4554 4555/* 4556 * The journal is only prepared to handle full-size block numbers, so we 4557 * have to adjust the record to reflect the change to a full-size block. 4558 * For example, suppose we have a block made up of fragments 8-15 and 4559 * want to free its last two fragments. We are given a request that says: 4560 * FREEBLK ino=5, blkno=14, lbn=0, frags=2, oldfrags=0 4561 * where frags are the number of fragments to free and oldfrags are the 4562 * number of fragments to keep. To block align it, we have to change it to 4563 * have a valid full-size blkno, so it becomes: 4564 * FREEBLK ino=5, blkno=8, lbn=0, frags=2, oldfrags=6 4565 */ 4566static void 4567adjust_newfreework(freeblks, frag_offset) 4568 struct freeblks *freeblks; 4569 int frag_offset; 4570{ 4571 struct jfreeblk *jfreeblk; 4572 4573 KASSERT((LIST_FIRST(&freeblks->fb_jblkdephd) != NULL && 4574 LIST_FIRST(&freeblks->fb_jblkdephd)->jb_list.wk_type == D_JFREEBLK), 4575 ("adjust_newfreework: Missing freeblks dependency")); 4576 4577 jfreeblk = WK_JFREEBLK(LIST_FIRST(&freeblks->fb_jblkdephd)); 4578 jfreeblk->jf_blkno -= frag_offset; 4579 jfreeblk->jf_frags += frag_offset; 4580} 4581 4582/* 4583 * Allocate a new jtrunc to track a partial truncation. 4584 */ 4585static struct jtrunc * 4586newjtrunc(freeblks, size, extsize) 4587 struct freeblks *freeblks; 4588 off_t size; 4589 int extsize; 4590{ 4591 struct jtrunc *jtrunc; 4592 4593 jtrunc = malloc(sizeof(*jtrunc), M_JTRUNC, M_SOFTDEP_FLAGS); 4594 workitem_alloc(&jtrunc->jt_dep.jb_list, D_JTRUNC, 4595 freeblks->fb_list.wk_mp); 4596 jtrunc->jt_dep.jb_jsegdep = newjsegdep(&jtrunc->jt_dep.jb_list); 4597 jtrunc->jt_dep.jb_freeblks = freeblks; 4598 jtrunc->jt_ino = freeblks->fb_inum; 4599 jtrunc->jt_size = size; 4600 jtrunc->jt_extsize = extsize; 4601 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jtrunc->jt_dep, jb_deps); 4602 4603 return (jtrunc); 4604} 4605 4606/* 4607 * If we're canceling a new bitmap we have to search for another ref 4608 * to move into the bmsafemap dep. This might be better expressed 4609 * with another structure. 4610 */ 4611static void 4612move_newblock_dep(jaddref, inodedep) 4613 struct jaddref *jaddref; 4614 struct inodedep *inodedep; 4615{ 4616 struct inoref *inoref; 4617 struct jaddref *jaddrefn; 4618 4619 jaddrefn = NULL; 4620 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 4621 inoref = TAILQ_NEXT(inoref, if_deps)) { 4622 if ((jaddref->ja_state & NEWBLOCK) && 4623 inoref->if_list.wk_type == D_JADDREF) { 4624 jaddrefn = (struct jaddref *)inoref; 4625 break; 4626 } 4627 } 4628 if (jaddrefn == NULL) 4629 return; 4630 jaddrefn->ja_state &= ~(ATTACHED | UNDONE); 4631 jaddrefn->ja_state |= jaddref->ja_state & 4632 (ATTACHED | UNDONE | NEWBLOCK); 4633 jaddref->ja_state &= ~(ATTACHED | UNDONE | NEWBLOCK); 4634 jaddref->ja_state |= ATTACHED; 4635 LIST_REMOVE(jaddref, ja_bmdeps); 4636 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_jaddrefhd, jaddrefn, 4637 ja_bmdeps); 4638} 4639 4640/* 4641 * Cancel a jaddref either before it has been written or while it is being 4642 * written. This happens when a link is removed before the add reaches 4643 * the disk. The jaddref dependency is kept linked into the bmsafemap 4644 * and inode to prevent the link count or bitmap from reaching the disk 4645 * until handle_workitem_remove() re-adjusts the counts and bitmaps as 4646 * required. 4647 * 4648 * Returns 1 if the canceled addref requires journaling of the remove and 4649 * 0 otherwise. 4650 */ 4651static int 4652cancel_jaddref(jaddref, inodedep, wkhd) 4653 struct jaddref *jaddref; 4654 struct inodedep *inodedep; 4655 struct workhead *wkhd; 4656{ 4657 struct inoref *inoref; 4658 struct jsegdep *jsegdep; 4659 int needsj; 4660 4661 KASSERT((jaddref->ja_state & COMPLETE) == 0, 4662 ("cancel_jaddref: Canceling complete jaddref")); 4663 if (jaddref->ja_state & (INPROGRESS | COMPLETE)) 4664 needsj = 1; 4665 else 4666 needsj = 0; 4667 if (inodedep == NULL) 4668 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 4669 0, &inodedep) == 0) 4670 panic("cancel_jaddref: Lost inodedep"); 4671 /* 4672 * We must adjust the nlink of any reference operation that follows 4673 * us so that it is consistent with the in-memory reference. This 4674 * ensures that inode nlink rollbacks always have the correct link. 4675 */ 4676 if (needsj == 0) { 4677 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 4678 inoref = TAILQ_NEXT(inoref, if_deps)) { 4679 if (inoref->if_state & GOINGAWAY) 4680 break; 4681 inoref->if_nlink--; 4682 } 4683 } 4684 jsegdep = inoref_jseg(&jaddref->ja_ref); 4685 if (jaddref->ja_state & NEWBLOCK) 4686 move_newblock_dep(jaddref, inodedep); 4687 wake_worklist(&jaddref->ja_list); 4688 jaddref->ja_mkdir = NULL; 4689 if (jaddref->ja_state & INPROGRESS) { 4690 jaddref->ja_state &= ~INPROGRESS; 4691 WORKLIST_REMOVE(&jaddref->ja_list); 4692 jwork_insert(wkhd, jsegdep); 4693 } else { 4694 free_jsegdep(jsegdep); 4695 if (jaddref->ja_state & DEPCOMPLETE) 4696 remove_from_journal(&jaddref->ja_list); 4697 } 4698 jaddref->ja_state |= (GOINGAWAY | DEPCOMPLETE); 4699 /* 4700 * Leave NEWBLOCK jaddrefs on the inodedep so handle_workitem_remove 4701 * can arrange for them to be freed with the bitmap. Otherwise we 4702 * no longer need this addref attached to the inoreflst and it 4703 * will incorrectly adjust nlink if we leave it. 4704 */ 4705 if ((jaddref->ja_state & NEWBLOCK) == 0) { 4706 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 4707 if_deps); 4708 jaddref->ja_state |= COMPLETE; 4709 free_jaddref(jaddref); 4710 return (needsj); 4711 } 4712 /* 4713 * Leave the head of the list for jsegdeps for fast merging. 4714 */ 4715 if (LIST_FIRST(wkhd) != NULL) { 4716 jaddref->ja_state |= ONWORKLIST; 4717 LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jaddref->ja_list, wk_list); 4718 } else 4719 WORKLIST_INSERT(wkhd, &jaddref->ja_list); 4720 4721 return (needsj); 4722} 4723 4724/* 4725 * Attempt to free a jaddref structure when some work completes. This 4726 * should only succeed once the entry is written and all dependencies have 4727 * been notified. 4728 */ 4729static void 4730free_jaddref(jaddref) 4731 struct jaddref *jaddref; 4732{ 4733 4734 if ((jaddref->ja_state & ALLCOMPLETE) != ALLCOMPLETE) 4735 return; 4736 if (jaddref->ja_ref.if_jsegdep) 4737 panic("free_jaddref: segdep attached to jaddref %p(0x%X)\n", 4738 jaddref, jaddref->ja_state); 4739 if (jaddref->ja_state & NEWBLOCK) 4740 LIST_REMOVE(jaddref, ja_bmdeps); 4741 if (jaddref->ja_state & (INPROGRESS | ONWORKLIST)) 4742 panic("free_jaddref: Bad state %p(0x%X)", 4743 jaddref, jaddref->ja_state); 4744 if (jaddref->ja_mkdir != NULL) 4745 panic("free_jaddref: Work pending, 0x%X\n", jaddref->ja_state); 4746 WORKITEM_FREE(jaddref, D_JADDREF); 4747} 4748 4749/* 4750 * Free a jremref structure once it has been written or discarded. 4751 */ 4752static void 4753free_jremref(jremref) 4754 struct jremref *jremref; 4755{ 4756 4757 if (jremref->jr_ref.if_jsegdep) 4758 free_jsegdep(jremref->jr_ref.if_jsegdep); 4759 if (jremref->jr_state & INPROGRESS) 4760 panic("free_jremref: IO still pending"); 4761 WORKITEM_FREE(jremref, D_JREMREF); 4762} 4763 4764/* 4765 * Free a jnewblk structure. 4766 */ 4767static void 4768free_jnewblk(jnewblk) 4769 struct jnewblk *jnewblk; 4770{ 4771 4772 if ((jnewblk->jn_state & ALLCOMPLETE) != ALLCOMPLETE) 4773 return; 4774 LIST_REMOVE(jnewblk, jn_deps); 4775 if (jnewblk->jn_dep != NULL) 4776 panic("free_jnewblk: Dependency still attached."); 4777 WORKITEM_FREE(jnewblk, D_JNEWBLK); 4778} 4779 4780/* 4781 * Cancel a jnewblk which has been been made redundant by frag extension. 4782 */ 4783static void 4784cancel_jnewblk(jnewblk, wkhd) 4785 struct jnewblk *jnewblk; 4786 struct workhead *wkhd; 4787{ 4788 struct jsegdep *jsegdep; 4789 4790 CTR1(KTR_SUJ, "cancel_jnewblk: blkno %jd", jnewblk->jn_blkno); 4791 jsegdep = jnewblk->jn_jsegdep; 4792 if (jnewblk->jn_jsegdep == NULL || jnewblk->jn_dep == NULL) 4793 panic("cancel_jnewblk: Invalid state"); 4794 jnewblk->jn_jsegdep = NULL; 4795 jnewblk->jn_dep = NULL; 4796 jnewblk->jn_state |= GOINGAWAY; 4797 if (jnewblk->jn_state & INPROGRESS) { 4798 jnewblk->jn_state &= ~INPROGRESS; 4799 WORKLIST_REMOVE(&jnewblk->jn_list); 4800 jwork_insert(wkhd, jsegdep); 4801 } else { 4802 free_jsegdep(jsegdep); 4803 remove_from_journal(&jnewblk->jn_list); 4804 } 4805 wake_worklist(&jnewblk->jn_list); 4806 WORKLIST_INSERT(wkhd, &jnewblk->jn_list); 4807} 4808 4809static void 4810free_jblkdep(jblkdep) 4811 struct jblkdep *jblkdep; 4812{ 4813 4814 if (jblkdep->jb_list.wk_type == D_JFREEBLK) 4815 WORKITEM_FREE(jblkdep, D_JFREEBLK); 4816 else if (jblkdep->jb_list.wk_type == D_JTRUNC) 4817 WORKITEM_FREE(jblkdep, D_JTRUNC); 4818 else 4819 panic("free_jblkdep: Unexpected type %s", 4820 TYPENAME(jblkdep->jb_list.wk_type)); 4821} 4822 4823/* 4824 * Free a single jseg once it is no longer referenced in memory or on 4825 * disk. Reclaim journal blocks and dependencies waiting for the segment 4826 * to disappear. 4827 */ 4828static void 4829free_jseg(jseg, jblocks) 4830 struct jseg *jseg; 4831 struct jblocks *jblocks; 4832{ 4833 struct freework *freework; 4834 4835 /* 4836 * Free freework structures that were lingering to indicate freed 4837 * indirect blocks that forced journal write ordering on reallocate. 4838 */ 4839 while ((freework = LIST_FIRST(&jseg->js_indirs)) != NULL) 4840 indirblk_remove(freework); 4841 if (jblocks->jb_oldestseg == jseg) 4842 jblocks->jb_oldestseg = TAILQ_NEXT(jseg, js_next); 4843 TAILQ_REMOVE(&jblocks->jb_segs, jseg, js_next); 4844 jblocks_free(jblocks, jseg->js_list.wk_mp, jseg->js_size); 4845 KASSERT(LIST_EMPTY(&jseg->js_entries), 4846 ("free_jseg: Freed jseg has valid entries.")); 4847 WORKITEM_FREE(jseg, D_JSEG); 4848} 4849 4850/* 4851 * Free all jsegs that meet the criteria for being reclaimed and update 4852 * oldestseg. 4853 */ 4854static void 4855free_jsegs(jblocks) 4856 struct jblocks *jblocks; 4857{ 4858 struct jseg *jseg; 4859 4860 /* 4861 * Free only those jsegs which have none allocated before them to 4862 * preserve the journal space ordering. 4863 */ 4864 while ((jseg = TAILQ_FIRST(&jblocks->jb_segs)) != NULL) { 4865 /* 4866 * Only reclaim space when nothing depends on this journal 4867 * set and another set has written that it is no longer 4868 * valid. 4869 */ 4870 if (jseg->js_refs != 0) { 4871 jblocks->jb_oldestseg = jseg; 4872 return; 4873 } 4874 if ((jseg->js_state & ALLCOMPLETE) != ALLCOMPLETE) 4875 break; 4876 if (jseg->js_seq > jblocks->jb_oldestwrseq) 4877 break; 4878 /* 4879 * We can free jsegs that didn't write entries when 4880 * oldestwrseq == js_seq. 4881 */ 4882 if (jseg->js_seq == jblocks->jb_oldestwrseq && 4883 jseg->js_cnt != 0) 4884 break; 4885 free_jseg(jseg, jblocks); 4886 } 4887 /* 4888 * If we exited the loop above we still must discover the 4889 * oldest valid segment. 4890 */ 4891 if (jseg) 4892 for (jseg = jblocks->jb_oldestseg; jseg != NULL; 4893 jseg = TAILQ_NEXT(jseg, js_next)) 4894 if (jseg->js_refs != 0) 4895 break; 4896 jblocks->jb_oldestseg = jseg; 4897 /* 4898 * The journal has no valid records but some jsegs may still be 4899 * waiting on oldestwrseq to advance. We force a small record 4900 * out to permit these lingering records to be reclaimed. 4901 */ 4902 if (jblocks->jb_oldestseg == NULL && !TAILQ_EMPTY(&jblocks->jb_segs)) 4903 jblocks->jb_needseg = 1; 4904} 4905 4906/* 4907 * Release one reference to a jseg and free it if the count reaches 0. This 4908 * should eventually reclaim journal space as well. 4909 */ 4910static void 4911rele_jseg(jseg) 4912 struct jseg *jseg; 4913{ 4914 4915 KASSERT(jseg->js_refs > 0, 4916 ("free_jseg: Invalid refcnt %d", jseg->js_refs)); 4917 if (--jseg->js_refs != 0) 4918 return; 4919 free_jsegs(jseg->js_jblocks); 4920} 4921 4922/* 4923 * Release a jsegdep and decrement the jseg count. 4924 */ 4925static void 4926free_jsegdep(jsegdep) 4927 struct jsegdep *jsegdep; 4928{ 4929 4930 if (jsegdep->jd_seg) 4931 rele_jseg(jsegdep->jd_seg); 4932 WORKITEM_FREE(jsegdep, D_JSEGDEP); 4933} 4934 4935/* 4936 * Wait for a journal item to make it to disk. Initiate journal processing 4937 * if required. 4938 */ 4939static int 4940jwait(wk, waitfor) 4941 struct worklist *wk; 4942 int waitfor; 4943{ 4944 4945 LOCK_OWNED(VFSTOUFS(wk->wk_mp)); 4946 /* 4947 * Blocking journal waits cause slow synchronous behavior. Record 4948 * stats on the frequency of these blocking operations. 4949 */ 4950 if (waitfor == MNT_WAIT) { 4951 stat_journal_wait++; 4952 switch (wk->wk_type) { 4953 case D_JREMREF: 4954 case D_JMVREF: 4955 stat_jwait_filepage++; 4956 break; 4957 case D_JTRUNC: 4958 case D_JFREEBLK: 4959 stat_jwait_freeblks++; 4960 break; 4961 case D_JNEWBLK: 4962 stat_jwait_newblk++; 4963 break; 4964 case D_JADDREF: 4965 stat_jwait_inode++; 4966 break; 4967 default: 4968 break; 4969 } 4970 } 4971 /* 4972 * If IO has not started we process the journal. We can't mark the 4973 * worklist item as IOWAITING because we drop the lock while 4974 * processing the journal and the worklist entry may be freed after 4975 * this point. The caller may call back in and re-issue the request. 4976 */ 4977 if ((wk->wk_state & INPROGRESS) == 0) { 4978 softdep_process_journal(wk->wk_mp, wk, waitfor); 4979 if (waitfor != MNT_WAIT) 4980 return (EBUSY); 4981 return (0); 4982 } 4983 if (waitfor != MNT_WAIT) 4984 return (EBUSY); 4985 wait_worklist(wk, "jwait"); 4986 return (0); 4987} 4988 4989/* 4990 * Lookup an inodedep based on an inode pointer and set the nlinkdelta as 4991 * appropriate. This is a convenience function to reduce duplicate code 4992 * for the setup and revert functions below. 4993 */ 4994static struct inodedep * 4995inodedep_lookup_ip(ip) 4996 struct inode *ip; 4997{ 4998 struct inodedep *inodedep; 4999 5000 KASSERT(ip->i_nlink >= ip->i_effnlink, 5001 ("inodedep_lookup_ip: bad delta")); 5002 (void) inodedep_lookup(ITOVFS(ip), ip->i_number, DEPALLOC, 5003 &inodedep); 5004 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 5005 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 5006 5007 return (inodedep); 5008} 5009 5010/* 5011 * Called prior to creating a new inode and linking it to a directory. The 5012 * jaddref structure must already be allocated by softdep_setup_inomapdep 5013 * and it is discovered here so we can initialize the mode and update 5014 * nlinkdelta. 5015 */ 5016void 5017softdep_setup_create(dp, ip) 5018 struct inode *dp; 5019 struct inode *ip; 5020{ 5021 struct inodedep *inodedep; 5022 struct jaddref *jaddref; 5023 struct vnode *dvp; 5024 5025 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 5026 ("softdep_setup_create called on non-softdep filesystem")); 5027 KASSERT(ip->i_nlink == 1, 5028 ("softdep_setup_create: Invalid link count.")); 5029 dvp = ITOV(dp); 5030 ACQUIRE_LOCK(ITOUMP(dp)); 5031 inodedep = inodedep_lookup_ip(ip); 5032 if (DOINGSUJ(dvp)) { 5033 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 5034 inoreflst); 5035 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 5036 ("softdep_setup_create: No addref structure present.")); 5037 } 5038 FREE_LOCK(ITOUMP(dp)); 5039} 5040 5041/* 5042 * Create a jaddref structure to track the addition of a DOTDOT link when 5043 * we are reparenting an inode as part of a rename. This jaddref will be 5044 * found by softdep_setup_directory_change. Adjusts nlinkdelta for 5045 * non-journaling softdep. 5046 */ 5047void 5048softdep_setup_dotdot_link(dp, ip) 5049 struct inode *dp; 5050 struct inode *ip; 5051{ 5052 struct inodedep *inodedep; 5053 struct jaddref *jaddref; 5054 struct vnode *dvp; 5055 5056 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 5057 ("softdep_setup_dotdot_link called on non-softdep filesystem")); 5058 dvp = ITOV(dp); 5059 jaddref = NULL; 5060 /* 5061 * We don't set MKDIR_PARENT as this is not tied to a mkdir and 5062 * is used as a normal link would be. 5063 */ 5064 if (DOINGSUJ(dvp)) 5065 jaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 5066 dp->i_effnlink - 1, dp->i_mode); 5067 ACQUIRE_LOCK(ITOUMP(dp)); 5068 inodedep = inodedep_lookup_ip(dp); 5069 if (jaddref) 5070 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 5071 if_deps); 5072 FREE_LOCK(ITOUMP(dp)); 5073} 5074 5075/* 5076 * Create a jaddref structure to track a new link to an inode. The directory 5077 * offset is not known until softdep_setup_directory_add or 5078 * softdep_setup_directory_change. Adjusts nlinkdelta for non-journaling 5079 * softdep. 5080 */ 5081void 5082softdep_setup_link(dp, ip) 5083 struct inode *dp; 5084 struct inode *ip; 5085{ 5086 struct inodedep *inodedep; 5087 struct jaddref *jaddref; 5088 struct vnode *dvp; 5089 5090 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 5091 ("softdep_setup_link called on non-softdep filesystem")); 5092 dvp = ITOV(dp); 5093 jaddref = NULL; 5094 if (DOINGSUJ(dvp)) 5095 jaddref = newjaddref(dp, ip->i_number, 0, ip->i_effnlink - 1, 5096 ip->i_mode); 5097 ACQUIRE_LOCK(ITOUMP(dp)); 5098 inodedep = inodedep_lookup_ip(ip); 5099 if (jaddref) 5100 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 5101 if_deps); 5102 FREE_LOCK(ITOUMP(dp)); 5103} 5104 5105/* 5106 * Called to create the jaddref structures to track . and .. references as 5107 * well as lookup and further initialize the incomplete jaddref created 5108 * by softdep_setup_inomapdep when the inode was allocated. Adjusts 5109 * nlinkdelta for non-journaling softdep. 5110 */ 5111void 5112softdep_setup_mkdir(dp, ip) 5113 struct inode *dp; 5114 struct inode *ip; 5115{ 5116 struct inodedep *inodedep; 5117 struct jaddref *dotdotaddref; 5118 struct jaddref *dotaddref; 5119 struct jaddref *jaddref; 5120 struct vnode *dvp; 5121 5122 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 5123 ("softdep_setup_mkdir called on non-softdep filesystem")); 5124 dvp = ITOV(dp); 5125 dotaddref = dotdotaddref = NULL; 5126 if (DOINGSUJ(dvp)) { 5127 dotaddref = newjaddref(ip, ip->i_number, DOT_OFFSET, 1, 5128 ip->i_mode); 5129 dotaddref->ja_state |= MKDIR_BODY; 5130 dotdotaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 5131 dp->i_effnlink - 1, dp->i_mode); 5132 dotdotaddref->ja_state |= MKDIR_PARENT; 5133 } 5134 ACQUIRE_LOCK(ITOUMP(dp)); 5135 inodedep = inodedep_lookup_ip(ip); 5136 if (DOINGSUJ(dvp)) { 5137 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 5138 inoreflst); 5139 KASSERT(jaddref != NULL, 5140 ("softdep_setup_mkdir: No addref structure present.")); 5141 KASSERT(jaddref->ja_parent == dp->i_number, 5142 ("softdep_setup_mkdir: bad parent %ju", 5143 (uintmax_t)jaddref->ja_parent)); 5144 TAILQ_INSERT_BEFORE(&jaddref->ja_ref, &dotaddref->ja_ref, 5145 if_deps); 5146 } 5147 inodedep = inodedep_lookup_ip(dp); 5148 if (DOINGSUJ(dvp)) 5149 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, 5150 &dotdotaddref->ja_ref, if_deps); 5151 FREE_LOCK(ITOUMP(dp)); 5152} 5153 5154/* 5155 * Called to track nlinkdelta of the inode and parent directories prior to 5156 * unlinking a directory. 5157 */ 5158void 5159softdep_setup_rmdir(dp, ip) 5160 struct inode *dp; 5161 struct inode *ip; 5162{ 5163 struct vnode *dvp; 5164 5165 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 5166 ("softdep_setup_rmdir called on non-softdep filesystem")); 5167 dvp = ITOV(dp); 5168 ACQUIRE_LOCK(ITOUMP(dp)); 5169 (void) inodedep_lookup_ip(ip); 5170 (void) inodedep_lookup_ip(dp); 5171 FREE_LOCK(ITOUMP(dp)); 5172} 5173 5174/* 5175 * Called to track nlinkdelta of the inode and parent directories prior to 5176 * unlink. 5177 */ 5178void 5179softdep_setup_unlink(dp, ip) 5180 struct inode *dp; 5181 struct inode *ip; 5182{ 5183 struct vnode *dvp; 5184 5185 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 5186 ("softdep_setup_unlink called on non-softdep filesystem")); 5187 dvp = ITOV(dp); 5188 ACQUIRE_LOCK(ITOUMP(dp)); 5189 (void) inodedep_lookup_ip(ip); 5190 (void) inodedep_lookup_ip(dp); 5191 FREE_LOCK(ITOUMP(dp)); 5192} 5193 5194/* 5195 * Called to release the journal structures created by a failed non-directory 5196 * creation. Adjusts nlinkdelta for non-journaling softdep. 5197 */ 5198void 5199softdep_revert_create(dp, ip) 5200 struct inode *dp; 5201 struct inode *ip; 5202{ 5203 struct inodedep *inodedep; 5204 struct jaddref *jaddref; 5205 struct vnode *dvp; 5206 5207 KASSERT(MOUNTEDSOFTDEP(ITOVFS((dp))) != 0, 5208 ("softdep_revert_create called on non-softdep filesystem")); 5209 dvp = ITOV(dp); 5210 ACQUIRE_LOCK(ITOUMP(dp)); 5211 inodedep = inodedep_lookup_ip(ip); 5212 if (DOINGSUJ(dvp)) { 5213 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 5214 inoreflst); 5215 KASSERT(jaddref->ja_parent == dp->i_number, 5216 ("softdep_revert_create: addref parent mismatch")); 5217 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 5218 } 5219 FREE_LOCK(ITOUMP(dp)); 5220} 5221 5222/* 5223 * Called to release the journal structures created by a failed link 5224 * addition. Adjusts nlinkdelta for non-journaling softdep. 5225 */ 5226void 5227softdep_revert_link(dp, ip) 5228 struct inode *dp; 5229 struct inode *ip; 5230{ 5231 struct inodedep *inodedep; 5232 struct jaddref *jaddref; 5233 struct vnode *dvp; 5234 5235 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 5236 ("softdep_revert_link called on non-softdep filesystem")); 5237 dvp = ITOV(dp); 5238 ACQUIRE_LOCK(ITOUMP(dp)); 5239 inodedep = inodedep_lookup_ip(ip); 5240 if (DOINGSUJ(dvp)) { 5241 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 5242 inoreflst); 5243 KASSERT(jaddref->ja_parent == dp->i_number, 5244 ("softdep_revert_link: addref parent mismatch")); 5245 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 5246 } 5247 FREE_LOCK(ITOUMP(dp)); 5248} 5249 5250/* 5251 * Called to release the journal structures created by a failed mkdir 5252 * attempt. Adjusts nlinkdelta for non-journaling softdep. 5253 */ 5254void 5255softdep_revert_mkdir(dp, ip) 5256 struct inode *dp; 5257 struct inode *ip; 5258{ 5259 struct inodedep *inodedep; 5260 struct jaddref *jaddref; 5261 struct jaddref *dotaddref; 5262 struct vnode *dvp; 5263 5264 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 5265 ("softdep_revert_mkdir called on non-softdep filesystem")); 5266 dvp = ITOV(dp); 5267 5268 ACQUIRE_LOCK(ITOUMP(dp)); 5269 inodedep = inodedep_lookup_ip(dp); 5270 if (DOINGSUJ(dvp)) { 5271 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 5272 inoreflst); 5273 KASSERT(jaddref->ja_parent == ip->i_number, 5274 ("softdep_revert_mkdir: dotdot addref parent mismatch")); 5275 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 5276 } 5277 inodedep = inodedep_lookup_ip(ip); 5278 if (DOINGSUJ(dvp)) { 5279 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 5280 inoreflst); 5281 KASSERT(jaddref->ja_parent == dp->i_number, 5282 ("softdep_revert_mkdir: addref parent mismatch")); 5283 dotaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 5284 inoreflst, if_deps); 5285 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 5286 KASSERT(dotaddref->ja_parent == ip->i_number, 5287 ("softdep_revert_mkdir: dot addref parent mismatch")); 5288 cancel_jaddref(dotaddref, inodedep, &inodedep->id_inowait); 5289 } 5290 FREE_LOCK(ITOUMP(dp)); 5291} 5292 5293/* 5294 * Called to correct nlinkdelta after a failed rmdir. 5295 */ 5296void 5297softdep_revert_rmdir(dp, ip) 5298 struct inode *dp; 5299 struct inode *ip; 5300{ 5301 5302 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 5303 ("softdep_revert_rmdir called on non-softdep filesystem")); 5304 ACQUIRE_LOCK(ITOUMP(dp)); 5305 (void) inodedep_lookup_ip(ip); 5306 (void) inodedep_lookup_ip(dp); 5307 FREE_LOCK(ITOUMP(dp)); 5308} 5309 5310/* 5311 * Protecting the freemaps (or bitmaps). 5312 * 5313 * To eliminate the need to execute fsck before mounting a filesystem 5314 * after a power failure, one must (conservatively) guarantee that the 5315 * on-disk copy of the bitmaps never indicate that a live inode or block is 5316 * free. So, when a block or inode is allocated, the bitmap should be 5317 * updated (on disk) before any new pointers. When a block or inode is 5318 * freed, the bitmap should not be updated until all pointers have been 5319 * reset. The latter dependency is handled by the delayed de-allocation 5320 * approach described below for block and inode de-allocation. The former 5321 * dependency is handled by calling the following procedure when a block or 5322 * inode is allocated. When an inode is allocated an "inodedep" is created 5323 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 5324 * Each "inodedep" is also inserted into the hash indexing structure so 5325 * that any additional link additions can be made dependent on the inode 5326 * allocation. 5327 * 5328 * The ufs filesystem maintains a number of free block counts (e.g., per 5329 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 5330 * in addition to the bitmaps. These counts are used to improve efficiency 5331 * during allocation and therefore must be consistent with the bitmaps. 5332 * There is no convenient way to guarantee post-crash consistency of these 5333 * counts with simple update ordering, for two main reasons: (1) The counts 5334 * and bitmaps for a single cylinder group block are not in the same disk 5335 * sector. If a disk write is interrupted (e.g., by power failure), one may 5336 * be written and the other not. (2) Some of the counts are located in the 5337 * superblock rather than the cylinder group block. So, we focus our soft 5338 * updates implementation on protecting the bitmaps. When mounting a 5339 * filesystem, we recompute the auxiliary counts from the bitmaps. 5340 */ 5341 5342/* 5343 * Called just after updating the cylinder group block to allocate an inode. 5344 */ 5345void 5346softdep_setup_inomapdep(bp, ip, newinum, mode) 5347 struct buf *bp; /* buffer for cylgroup block with inode map */ 5348 struct inode *ip; /* inode related to allocation */ 5349 ino_t newinum; /* new inode number being allocated */ 5350 int mode; 5351{ 5352 struct inodedep *inodedep; 5353 struct bmsafemap *bmsafemap; 5354 struct jaddref *jaddref; 5355 struct mount *mp; 5356 struct fs *fs; 5357 5358 mp = ITOVFS(ip); 5359 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5360 ("softdep_setup_inomapdep called on non-softdep filesystem")); 5361 fs = VFSTOUFS(mp)->um_fs; 5362 jaddref = NULL; 5363 5364 /* 5365 * Allocate the journal reference add structure so that the bitmap 5366 * can be dependent on it. 5367 */ 5368 if (MOUNTEDSUJ(mp)) { 5369 jaddref = newjaddref(ip, newinum, 0, 0, mode); 5370 jaddref->ja_state |= NEWBLOCK; 5371 } 5372 5373 /* 5374 * Create a dependency for the newly allocated inode. 5375 * Panic if it already exists as something is seriously wrong. 5376 * Otherwise add it to the dependency list for the buffer holding 5377 * the cylinder group map from which it was allocated. 5378 * 5379 * We have to preallocate a bmsafemap entry in case it is needed 5380 * in bmsafemap_lookup since once we allocate the inodedep, we 5381 * have to finish initializing it before we can FREE_LOCK(). 5382 * By preallocating, we avoid FREE_LOCK() while doing a malloc 5383 * in bmsafemap_lookup. We cannot call bmsafemap_lookup before 5384 * creating the inodedep as it can be freed during the time 5385 * that we FREE_LOCK() while allocating the inodedep. We must 5386 * call workitem_alloc() before entering the locked section as 5387 * it also acquires the lock and we must avoid trying doing so 5388 * recursively. 5389 */ 5390 bmsafemap = malloc(sizeof(struct bmsafemap), 5391 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 5392 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 5393 ACQUIRE_LOCK(ITOUMP(ip)); 5394 if ((inodedep_lookup(mp, newinum, DEPALLOC, &inodedep))) 5395 panic("softdep_setup_inomapdep: dependency %p for new" 5396 "inode already exists", inodedep); 5397 bmsafemap = bmsafemap_lookup(mp, bp, ino_to_cg(fs, newinum), bmsafemap); 5398 if (jaddref) { 5399 LIST_INSERT_HEAD(&bmsafemap->sm_jaddrefhd, jaddref, ja_bmdeps); 5400 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 5401 if_deps); 5402 } else { 5403 inodedep->id_state |= ONDEPLIST; 5404 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 5405 } 5406 inodedep->id_bmsafemap = bmsafemap; 5407 inodedep->id_state &= ~DEPCOMPLETE; 5408 FREE_LOCK(ITOUMP(ip)); 5409} 5410 5411/* 5412 * Called just after updating the cylinder group block to 5413 * allocate block or fragment. 5414 */ 5415void 5416softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 5417 struct buf *bp; /* buffer for cylgroup block with block map */ 5418 struct mount *mp; /* filesystem doing allocation */ 5419 ufs2_daddr_t newblkno; /* number of newly allocated block */ 5420 int frags; /* Number of fragments. */ 5421 int oldfrags; /* Previous number of fragments for extend. */ 5422{ 5423 struct newblk *newblk; 5424 struct bmsafemap *bmsafemap; 5425 struct jnewblk *jnewblk; 5426 struct ufsmount *ump; 5427 struct fs *fs; 5428 5429 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5430 ("softdep_setup_blkmapdep called on non-softdep filesystem")); 5431 ump = VFSTOUFS(mp); 5432 fs = ump->um_fs; 5433 jnewblk = NULL; 5434 /* 5435 * Create a dependency for the newly allocated block. 5436 * Add it to the dependency list for the buffer holding 5437 * the cylinder group map from which it was allocated. 5438 */ 5439 if (MOUNTEDSUJ(mp)) { 5440 jnewblk = malloc(sizeof(*jnewblk), M_JNEWBLK, M_SOFTDEP_FLAGS); 5441 workitem_alloc(&jnewblk->jn_list, D_JNEWBLK, mp); 5442 jnewblk->jn_jsegdep = newjsegdep(&jnewblk->jn_list); 5443 jnewblk->jn_state = ATTACHED; 5444 jnewblk->jn_blkno = newblkno; 5445 jnewblk->jn_frags = frags; 5446 jnewblk->jn_oldfrags = oldfrags; 5447#ifdef INVARIANTS 5448 { 5449 struct cg *cgp; 5450 uint8_t *blksfree; 5451 long bno; 5452 int i; 5453 5454 cgp = (struct cg *)bp->b_data; 5455 blksfree = cg_blksfree(cgp); 5456 bno = dtogd(fs, jnewblk->jn_blkno); 5457 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; 5458 i++) { 5459 if (isset(blksfree, bno + i)) 5460 panic("softdep_setup_blkmapdep: " 5461 "free fragment %d from %d-%d " 5462 "state 0x%X dep %p", i, 5463 jnewblk->jn_oldfrags, 5464 jnewblk->jn_frags, 5465 jnewblk->jn_state, 5466 jnewblk->jn_dep); 5467 } 5468 } 5469#endif 5470 } 5471 5472 CTR3(KTR_SUJ, 5473 "softdep_setup_blkmapdep: blkno %jd frags %d oldfrags %d", 5474 newblkno, frags, oldfrags); 5475 ACQUIRE_LOCK(ump); 5476 if (newblk_lookup(mp, newblkno, DEPALLOC, &newblk) != 0) 5477 panic("softdep_setup_blkmapdep: found block"); 5478 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp, 5479 dtog(fs, newblkno), NULL); 5480 if (jnewblk) { 5481 jnewblk->jn_dep = (struct worklist *)newblk; 5482 LIST_INSERT_HEAD(&bmsafemap->sm_jnewblkhd, jnewblk, jn_deps); 5483 } else { 5484 newblk->nb_state |= ONDEPLIST; 5485 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 5486 } 5487 newblk->nb_bmsafemap = bmsafemap; 5488 newblk->nb_jnewblk = jnewblk; 5489 FREE_LOCK(ump); 5490} 5491 5492#define BMSAFEMAP_HASH(ump, cg) \ 5493 (&(ump)->bmsafemap_hashtbl[(cg) & (ump)->bmsafemap_hash_size]) 5494 5495static int 5496bmsafemap_find(bmsafemaphd, cg, bmsafemapp) 5497 struct bmsafemap_hashhead *bmsafemaphd; 5498 int cg; 5499 struct bmsafemap **bmsafemapp; 5500{ 5501 struct bmsafemap *bmsafemap; 5502 5503 LIST_FOREACH(bmsafemap, bmsafemaphd, sm_hash) 5504 if (bmsafemap->sm_cg == cg) 5505 break; 5506 if (bmsafemap) { 5507 *bmsafemapp = bmsafemap; 5508 return (1); 5509 } 5510 *bmsafemapp = NULL; 5511 5512 return (0); 5513} 5514 5515/* 5516 * Find the bmsafemap associated with a cylinder group buffer. 5517 * If none exists, create one. The buffer must be locked when 5518 * this routine is called and this routine must be called with 5519 * the softdep lock held. To avoid giving up the lock while 5520 * allocating a new bmsafemap, a preallocated bmsafemap may be 5521 * provided. If it is provided but not needed, it is freed. 5522 */ 5523static struct bmsafemap * 5524bmsafemap_lookup(mp, bp, cg, newbmsafemap) 5525 struct mount *mp; 5526 struct buf *bp; 5527 int cg; 5528 struct bmsafemap *newbmsafemap; 5529{ 5530 struct bmsafemap_hashhead *bmsafemaphd; 5531 struct bmsafemap *bmsafemap, *collision; 5532 struct worklist *wk; 5533 struct ufsmount *ump; 5534 5535 ump = VFSTOUFS(mp); 5536 LOCK_OWNED(ump); 5537 KASSERT(bp != NULL, ("bmsafemap_lookup: missing buffer")); 5538 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5539 if (wk->wk_type == D_BMSAFEMAP) { 5540 if (newbmsafemap) 5541 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP); 5542 return (WK_BMSAFEMAP(wk)); 5543 } 5544 } 5545 bmsafemaphd = BMSAFEMAP_HASH(ump, cg); 5546 if (bmsafemap_find(bmsafemaphd, cg, &bmsafemap) == 1) { 5547 if (newbmsafemap) 5548 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP); 5549 return (bmsafemap); 5550 } 5551 if (newbmsafemap) { 5552 bmsafemap = newbmsafemap; 5553 } else { 5554 FREE_LOCK(ump); 5555 bmsafemap = malloc(sizeof(struct bmsafemap), 5556 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 5557 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 5558 ACQUIRE_LOCK(ump); 5559 } 5560 bmsafemap->sm_buf = bp; 5561 LIST_INIT(&bmsafemap->sm_inodedephd); 5562 LIST_INIT(&bmsafemap->sm_inodedepwr); 5563 LIST_INIT(&bmsafemap->sm_newblkhd); 5564 LIST_INIT(&bmsafemap->sm_newblkwr); 5565 LIST_INIT(&bmsafemap->sm_jaddrefhd); 5566 LIST_INIT(&bmsafemap->sm_jnewblkhd); 5567 LIST_INIT(&bmsafemap->sm_freehd); 5568 LIST_INIT(&bmsafemap->sm_freewr); 5569 if (bmsafemap_find(bmsafemaphd, cg, &collision) == 1) { 5570 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 5571 return (collision); 5572 } 5573 bmsafemap->sm_cg = cg; 5574 LIST_INSERT_HEAD(bmsafemaphd, bmsafemap, sm_hash); 5575 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next); 5576 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 5577 return (bmsafemap); 5578} 5579 5580/* 5581 * Direct block allocation dependencies. 5582 * 5583 * When a new block is allocated, the corresponding disk locations must be 5584 * initialized (with zeros or new data) before the on-disk inode points to 5585 * them. Also, the freemap from which the block was allocated must be 5586 * updated (on disk) before the inode's pointer. These two dependencies are 5587 * independent of each other and are needed for all file blocks and indirect 5588 * blocks that are pointed to directly by the inode. Just before the 5589 * "in-core" version of the inode is updated with a newly allocated block 5590 * number, a procedure (below) is called to setup allocation dependency 5591 * structures. These structures are removed when the corresponding 5592 * dependencies are satisfied or when the block allocation becomes obsolete 5593 * (i.e., the file is deleted, the block is de-allocated, or the block is a 5594 * fragment that gets upgraded). All of these cases are handled in 5595 * procedures described later. 5596 * 5597 * When a file extension causes a fragment to be upgraded, either to a larger 5598 * fragment or to a full block, the on-disk location may change (if the 5599 * previous fragment could not simply be extended). In this case, the old 5600 * fragment must be de-allocated, but not until after the inode's pointer has 5601 * been updated. In most cases, this is handled by later procedures, which 5602 * will construct a "freefrag" structure to be added to the workitem queue 5603 * when the inode update is complete (or obsolete). The main exception to 5604 * this is when an allocation occurs while a pending allocation dependency 5605 * (for the same block pointer) remains. This case is handled in the main 5606 * allocation dependency setup procedure by immediately freeing the 5607 * unreferenced fragments. 5608 */ 5609void 5610softdep_setup_allocdirect(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 5611 struct inode *ip; /* inode to which block is being added */ 5612 ufs_lbn_t off; /* block pointer within inode */ 5613 ufs2_daddr_t newblkno; /* disk block number being added */ 5614 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */ 5615 long newsize; /* size of new block */ 5616 long oldsize; /* size of new block */ 5617 struct buf *bp; /* bp for allocated block */ 5618{ 5619 struct allocdirect *adp, *oldadp; 5620 struct allocdirectlst *adphead; 5621 struct freefrag *freefrag; 5622 struct inodedep *inodedep; 5623 struct pagedep *pagedep; 5624 struct jnewblk *jnewblk; 5625 struct newblk *newblk; 5626 struct mount *mp; 5627 ufs_lbn_t lbn; 5628 5629 lbn = bp->b_lblkno; 5630 mp = ITOVFS(ip); 5631 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5632 ("softdep_setup_allocdirect called on non-softdep filesystem")); 5633 if (oldblkno && oldblkno != newblkno) 5634 /* 5635 * The usual case is that a smaller fragment that 5636 * was just allocated has been replaced with a bigger 5637 * fragment or a full-size block. If it is marked as 5638 * B_DELWRI, the current contents have not been written 5639 * to disk. It is possible that the block was written 5640 * earlier, but very uncommon. If the block has never 5641 * been written, there is no need to send a BIO_DELETE 5642 * for it when it is freed. The gain from avoiding the 5643 * TRIMs for the common case of unwritten blocks far 5644 * exceeds the cost of the write amplification for the 5645 * uncommon case of failing to send a TRIM for a block 5646 * that had been written. 5647 */ 5648 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn, 5649 (bp->b_flags & B_DELWRI) != 0 ? NOTRIM_KEY : SINGLETON_KEY); 5650 else 5651 freefrag = NULL; 5652 5653 CTR6(KTR_SUJ, 5654 "softdep_setup_allocdirect: ino %d blkno %jd oldblkno %jd " 5655 "off %jd newsize %ld oldsize %d", 5656 ip->i_number, newblkno, oldblkno, off, newsize, oldsize); 5657 ACQUIRE_LOCK(ITOUMP(ip)); 5658 if (off >= UFS_NDADDR) { 5659 if (lbn > 0) 5660 panic("softdep_setup_allocdirect: bad lbn %jd, off %jd", 5661 lbn, off); 5662 /* allocating an indirect block */ 5663 if (oldblkno != 0) 5664 panic("softdep_setup_allocdirect: non-zero indir"); 5665 } else { 5666 if (off != lbn) 5667 panic("softdep_setup_allocdirect: lbn %jd != off %jd", 5668 lbn, off); 5669 /* 5670 * Allocating a direct block. 5671 * 5672 * If we are allocating a directory block, then we must 5673 * allocate an associated pagedep to track additions and 5674 * deletions. 5675 */ 5676 if ((ip->i_mode & IFMT) == IFDIR) 5677 pagedep_lookup(mp, bp, ip->i_number, off, DEPALLOC, 5678 &pagedep); 5679 } 5680 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 5681 panic("softdep_setup_allocdirect: lost block"); 5682 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5683 ("softdep_setup_allocdirect: newblk already initialized")); 5684 /* 5685 * Convert the newblk to an allocdirect. 5686 */ 5687 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT); 5688 adp = (struct allocdirect *)newblk; 5689 newblk->nb_freefrag = freefrag; 5690 adp->ad_offset = off; 5691 adp->ad_oldblkno = oldblkno; 5692 adp->ad_newsize = newsize; 5693 adp->ad_oldsize = oldsize; 5694 5695 /* 5696 * Finish initializing the journal. 5697 */ 5698 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5699 jnewblk->jn_ino = ip->i_number; 5700 jnewblk->jn_lbn = lbn; 5701 add_to_journal(&jnewblk->jn_list); 5702 } 5703 if (freefrag && freefrag->ff_jdep != NULL && 5704 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5705 add_to_journal(freefrag->ff_jdep); 5706 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5707 adp->ad_inodedep = inodedep; 5708 5709 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 5710 /* 5711 * The list of allocdirects must be kept in sorted and ascending 5712 * order so that the rollback routines can quickly determine the 5713 * first uncommitted block (the size of the file stored on disk 5714 * ends at the end of the lowest committed fragment, or if there 5715 * are no fragments, at the end of the highest committed block). 5716 * Since files generally grow, the typical case is that the new 5717 * block is to be added at the end of the list. We speed this 5718 * special case by checking against the last allocdirect in the 5719 * list before laboriously traversing the list looking for the 5720 * insertion point. 5721 */ 5722 adphead = &inodedep->id_newinoupdt; 5723 oldadp = TAILQ_LAST(adphead, allocdirectlst); 5724 if (oldadp == NULL || oldadp->ad_offset <= off) { 5725 /* insert at end of list */ 5726 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 5727 if (oldadp != NULL && oldadp->ad_offset == off) 5728 allocdirect_merge(adphead, adp, oldadp); 5729 FREE_LOCK(ITOUMP(ip)); 5730 return; 5731 } 5732 TAILQ_FOREACH(oldadp, adphead, ad_next) { 5733 if (oldadp->ad_offset >= off) 5734 break; 5735 } 5736 if (oldadp == NULL) 5737 panic("softdep_setup_allocdirect: lost entry"); 5738 /* insert in middle of list */ 5739 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 5740 if (oldadp->ad_offset == off) 5741 allocdirect_merge(adphead, adp, oldadp); 5742 5743 FREE_LOCK(ITOUMP(ip)); 5744} 5745 5746/* 5747 * Merge a newer and older journal record to be stored either in a 5748 * newblock or freefrag. This handles aggregating journal records for 5749 * fragment allocation into a second record as well as replacing a 5750 * journal free with an aborted journal allocation. A segment for the 5751 * oldest record will be placed on wkhd if it has been written. If not 5752 * the segment for the newer record will suffice. 5753 */ 5754static struct worklist * 5755jnewblk_merge(new, old, wkhd) 5756 struct worklist *new; 5757 struct worklist *old; 5758 struct workhead *wkhd; 5759{ 5760 struct jnewblk *njnewblk; 5761 struct jnewblk *jnewblk; 5762 5763 /* Handle NULLs to simplify callers. */ 5764 if (new == NULL) 5765 return (old); 5766 if (old == NULL) 5767 return (new); 5768 /* Replace a jfreefrag with a jnewblk. */ 5769 if (new->wk_type == D_JFREEFRAG) { 5770 if (WK_JNEWBLK(old)->jn_blkno != WK_JFREEFRAG(new)->fr_blkno) 5771 panic("jnewblk_merge: blkno mismatch: %p, %p", 5772 old, new); 5773 cancel_jfreefrag(WK_JFREEFRAG(new)); 5774 return (old); 5775 } 5776 if (old->wk_type != D_JNEWBLK || new->wk_type != D_JNEWBLK) 5777 panic("jnewblk_merge: Bad type: old %d new %d\n", 5778 old->wk_type, new->wk_type); 5779 /* 5780 * Handle merging of two jnewblk records that describe 5781 * different sets of fragments in the same block. 5782 */ 5783 jnewblk = WK_JNEWBLK(old); 5784 njnewblk = WK_JNEWBLK(new); 5785 if (jnewblk->jn_blkno != njnewblk->jn_blkno) 5786 panic("jnewblk_merge: Merging disparate blocks."); 5787 /* 5788 * The record may be rolled back in the cg. 5789 */ 5790 if (jnewblk->jn_state & UNDONE) { 5791 jnewblk->jn_state &= ~UNDONE; 5792 njnewblk->jn_state |= UNDONE; 5793 njnewblk->jn_state &= ~ATTACHED; 5794 } 5795 /* 5796 * We modify the newer addref and free the older so that if neither 5797 * has been written the most up-to-date copy will be on disk. If 5798 * both have been written but rolled back we only temporarily need 5799 * one of them to fix the bits when the cg write completes. 5800 */ 5801 jnewblk->jn_state |= ATTACHED | COMPLETE; 5802 njnewblk->jn_oldfrags = jnewblk->jn_oldfrags; 5803 cancel_jnewblk(jnewblk, wkhd); 5804 WORKLIST_REMOVE(&jnewblk->jn_list); 5805 free_jnewblk(jnewblk); 5806 return (new); 5807} 5808 5809/* 5810 * Replace an old allocdirect dependency with a newer one. 5811 */ 5812static void 5813allocdirect_merge(adphead, newadp, oldadp) 5814 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 5815 struct allocdirect *newadp; /* allocdirect being added */ 5816 struct allocdirect *oldadp; /* existing allocdirect being checked */ 5817{ 5818 struct worklist *wk; 5819 struct freefrag *freefrag; 5820 5821 freefrag = NULL; 5822 LOCK_OWNED(VFSTOUFS(newadp->ad_list.wk_mp)); 5823 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 5824 newadp->ad_oldsize != oldadp->ad_newsize || 5825 newadp->ad_offset >= UFS_NDADDR) 5826 panic("%s %jd != new %jd || old size %ld != new %ld", 5827 "allocdirect_merge: old blkno", 5828 (intmax_t)newadp->ad_oldblkno, 5829 (intmax_t)oldadp->ad_newblkno, 5830 newadp->ad_oldsize, oldadp->ad_newsize); 5831 newadp->ad_oldblkno = oldadp->ad_oldblkno; 5832 newadp->ad_oldsize = oldadp->ad_oldsize; 5833 /* 5834 * If the old dependency had a fragment to free or had never 5835 * previously had a block allocated, then the new dependency 5836 * can immediately post its freefrag and adopt the old freefrag. 5837 * This action is done by swapping the freefrag dependencies. 5838 * The new dependency gains the old one's freefrag, and the 5839 * old one gets the new one and then immediately puts it on 5840 * the worklist when it is freed by free_newblk. It is 5841 * not possible to do this swap when the old dependency had a 5842 * non-zero size but no previous fragment to free. This condition 5843 * arises when the new block is an extension of the old block. 5844 * Here, the first part of the fragment allocated to the new 5845 * dependency is part of the block currently claimed on disk by 5846 * the old dependency, so cannot legitimately be freed until the 5847 * conditions for the new dependency are fulfilled. 5848 */ 5849 freefrag = newadp->ad_freefrag; 5850 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 5851 newadp->ad_freefrag = oldadp->ad_freefrag; 5852 oldadp->ad_freefrag = freefrag; 5853 } 5854 /* 5855 * If we are tracking a new directory-block allocation, 5856 * move it from the old allocdirect to the new allocdirect. 5857 */ 5858 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) { 5859 WORKLIST_REMOVE(wk); 5860 if (!LIST_EMPTY(&oldadp->ad_newdirblk)) 5861 panic("allocdirect_merge: extra newdirblk"); 5862 WORKLIST_INSERT(&newadp->ad_newdirblk, wk); 5863 } 5864 TAILQ_REMOVE(adphead, oldadp, ad_next); 5865 /* 5866 * We need to move any journal dependencies over to the freefrag 5867 * that releases this block if it exists. Otherwise we are 5868 * extending an existing block and we'll wait until that is 5869 * complete to release the journal space and extend the 5870 * new journal to cover this old space as well. 5871 */ 5872 if (freefrag == NULL) { 5873 if (oldadp->ad_newblkno != newadp->ad_newblkno) 5874 panic("allocdirect_merge: %jd != %jd", 5875 oldadp->ad_newblkno, newadp->ad_newblkno); 5876 newadp->ad_block.nb_jnewblk = (struct jnewblk *) 5877 jnewblk_merge(&newadp->ad_block.nb_jnewblk->jn_list, 5878 &oldadp->ad_block.nb_jnewblk->jn_list, 5879 &newadp->ad_block.nb_jwork); 5880 oldadp->ad_block.nb_jnewblk = NULL; 5881 cancel_newblk(&oldadp->ad_block, NULL, 5882 &newadp->ad_block.nb_jwork); 5883 } else { 5884 wk = (struct worklist *) cancel_newblk(&oldadp->ad_block, 5885 &freefrag->ff_list, &freefrag->ff_jwork); 5886 freefrag->ff_jdep = jnewblk_merge(freefrag->ff_jdep, wk, 5887 &freefrag->ff_jwork); 5888 } 5889 free_newblk(&oldadp->ad_block); 5890} 5891 5892/* 5893 * Allocate a jfreefrag structure to journal a single block free. 5894 */ 5895static struct jfreefrag * 5896newjfreefrag(freefrag, ip, blkno, size, lbn) 5897 struct freefrag *freefrag; 5898 struct inode *ip; 5899 ufs2_daddr_t blkno; 5900 long size; 5901 ufs_lbn_t lbn; 5902{ 5903 struct jfreefrag *jfreefrag; 5904 struct fs *fs; 5905 5906 fs = ITOFS(ip); 5907 jfreefrag = malloc(sizeof(struct jfreefrag), M_JFREEFRAG, 5908 M_SOFTDEP_FLAGS); 5909 workitem_alloc(&jfreefrag->fr_list, D_JFREEFRAG, ITOVFS(ip)); 5910 jfreefrag->fr_jsegdep = newjsegdep(&jfreefrag->fr_list); 5911 jfreefrag->fr_state = ATTACHED | DEPCOMPLETE; 5912 jfreefrag->fr_ino = ip->i_number; 5913 jfreefrag->fr_lbn = lbn; 5914 jfreefrag->fr_blkno = blkno; 5915 jfreefrag->fr_frags = numfrags(fs, size); 5916 jfreefrag->fr_freefrag = freefrag; 5917 5918 return (jfreefrag); 5919} 5920 5921/* 5922 * Allocate a new freefrag structure. 5923 */ 5924static struct freefrag * 5925newfreefrag(ip, blkno, size, lbn, key) 5926 struct inode *ip; 5927 ufs2_daddr_t blkno; 5928 long size; 5929 ufs_lbn_t lbn; 5930 u_long key; 5931{ 5932 struct freefrag *freefrag; 5933 struct ufsmount *ump; 5934 struct fs *fs; 5935 5936 CTR4(KTR_SUJ, "newfreefrag: ino %d blkno %jd size %ld lbn %jd", 5937 ip->i_number, blkno, size, lbn); 5938 ump = ITOUMP(ip); 5939 fs = ump->um_fs; 5940 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 5941 panic("newfreefrag: frag size"); 5942 freefrag = malloc(sizeof(struct freefrag), 5943 M_FREEFRAG, M_SOFTDEP_FLAGS); 5944 workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ump)); 5945 freefrag->ff_state = ATTACHED; 5946 LIST_INIT(&freefrag->ff_jwork); 5947 freefrag->ff_inum = ip->i_number; 5948 freefrag->ff_vtype = ITOV(ip)->v_type; 5949 freefrag->ff_blkno = blkno; 5950 freefrag->ff_fragsize = size; 5951 freefrag->ff_key = key; 5952 5953 if (MOUNTEDSUJ(UFSTOVFS(ump))) { 5954 freefrag->ff_jdep = (struct worklist *) 5955 newjfreefrag(freefrag, ip, blkno, size, lbn); 5956 } else { 5957 freefrag->ff_state |= DEPCOMPLETE; 5958 freefrag->ff_jdep = NULL; 5959 } 5960 5961 return (freefrag); 5962} 5963 5964/* 5965 * This workitem de-allocates fragments that were replaced during 5966 * file block allocation. 5967 */ 5968static void 5969handle_workitem_freefrag(freefrag) 5970 struct freefrag *freefrag; 5971{ 5972 struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp); 5973 struct workhead wkhd; 5974 5975 CTR3(KTR_SUJ, 5976 "handle_workitem_freefrag: ino %d blkno %jd size %ld", 5977 freefrag->ff_inum, freefrag->ff_blkno, freefrag->ff_fragsize); 5978 /* 5979 * It would be illegal to add new completion items to the 5980 * freefrag after it was schedule to be done so it must be 5981 * safe to modify the list head here. 5982 */ 5983 LIST_INIT(&wkhd); 5984 ACQUIRE_LOCK(ump); 5985 LIST_SWAP(&freefrag->ff_jwork, &wkhd, worklist, wk_list); 5986 /* 5987 * If the journal has not been written we must cancel it here. 5988 */ 5989 if (freefrag->ff_jdep) { 5990 if (freefrag->ff_jdep->wk_type != D_JNEWBLK) 5991 panic("handle_workitem_freefrag: Unexpected type %d\n", 5992 freefrag->ff_jdep->wk_type); 5993 cancel_jnewblk(WK_JNEWBLK(freefrag->ff_jdep), &wkhd); 5994 } 5995 FREE_LOCK(ump); 5996 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno, 5997 freefrag->ff_fragsize, freefrag->ff_inum, freefrag->ff_vtype, 5998 &wkhd, freefrag->ff_key); 5999 ACQUIRE_LOCK(ump); 6000 WORKITEM_FREE(freefrag, D_FREEFRAG); 6001 FREE_LOCK(ump); 6002} 6003 6004/* 6005 * Set up a dependency structure for an external attributes data block. 6006 * This routine follows much of the structure of softdep_setup_allocdirect. 6007 * See the description of softdep_setup_allocdirect above for details. 6008 */ 6009void 6010softdep_setup_allocext(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 6011 struct inode *ip; 6012 ufs_lbn_t off; 6013 ufs2_daddr_t newblkno; 6014 ufs2_daddr_t oldblkno; 6015 long newsize; 6016 long oldsize; 6017 struct buf *bp; 6018{ 6019 struct allocdirect *adp, *oldadp; 6020 struct allocdirectlst *adphead; 6021 struct freefrag *freefrag; 6022 struct inodedep *inodedep; 6023 struct jnewblk *jnewblk; 6024 struct newblk *newblk; 6025 struct mount *mp; 6026 struct ufsmount *ump; 6027 ufs_lbn_t lbn; 6028 6029 mp = ITOVFS(ip); 6030 ump = VFSTOUFS(mp); 6031 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 6032 ("softdep_setup_allocext called on non-softdep filesystem")); 6033 KASSERT(off < UFS_NXADDR, 6034 ("softdep_setup_allocext: lbn %lld > UFS_NXADDR", (long long)off)); 6035 6036 lbn = bp->b_lblkno; 6037 if (oldblkno && oldblkno != newblkno) 6038 /* 6039 * The usual case is that a smaller fragment that 6040 * was just allocated has been replaced with a bigger 6041 * fragment or a full-size block. If it is marked as 6042 * B_DELWRI, the current contents have not been written 6043 * to disk. It is possible that the block was written 6044 * earlier, but very uncommon. If the block has never 6045 * been written, there is no need to send a BIO_DELETE 6046 * for it when it is freed. The gain from avoiding the 6047 * TRIMs for the common case of unwritten blocks far 6048 * exceeds the cost of the write amplification for the 6049 * uncommon case of failing to send a TRIM for a block 6050 * that had been written. 6051 */ 6052 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn, 6053 (bp->b_flags & B_DELWRI) != 0 ? NOTRIM_KEY : SINGLETON_KEY); 6054 else 6055 freefrag = NULL; 6056 6057 ACQUIRE_LOCK(ump); 6058 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 6059 panic("softdep_setup_allocext: lost block"); 6060 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 6061 ("softdep_setup_allocext: newblk already initialized")); 6062 /* 6063 * Convert the newblk to an allocdirect. 6064 */ 6065 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT); 6066 adp = (struct allocdirect *)newblk; 6067 newblk->nb_freefrag = freefrag; 6068 adp->ad_offset = off; 6069 adp->ad_oldblkno = oldblkno; 6070 adp->ad_newsize = newsize; 6071 adp->ad_oldsize = oldsize; 6072 adp->ad_state |= EXTDATA; 6073 6074 /* 6075 * Finish initializing the journal. 6076 */ 6077 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 6078 jnewblk->jn_ino = ip->i_number; 6079 jnewblk->jn_lbn = lbn; 6080 add_to_journal(&jnewblk->jn_list); 6081 } 6082 if (freefrag && freefrag->ff_jdep != NULL && 6083 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 6084 add_to_journal(freefrag->ff_jdep); 6085 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6086 adp->ad_inodedep = inodedep; 6087 6088 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 6089 /* 6090 * The list of allocdirects must be kept in sorted and ascending 6091 * order so that the rollback routines can quickly determine the 6092 * first uncommitted block (the size of the file stored on disk 6093 * ends at the end of the lowest committed fragment, or if there 6094 * are no fragments, at the end of the highest committed block). 6095 * Since files generally grow, the typical case is that the new 6096 * block is to be added at the end of the list. We speed this 6097 * special case by checking against the last allocdirect in the 6098 * list before laboriously traversing the list looking for the 6099 * insertion point. 6100 */ 6101 adphead = &inodedep->id_newextupdt; 6102 oldadp = TAILQ_LAST(adphead, allocdirectlst); 6103 if (oldadp == NULL || oldadp->ad_offset <= off) { 6104 /* insert at end of list */ 6105 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 6106 if (oldadp != NULL && oldadp->ad_offset == off) 6107 allocdirect_merge(adphead, adp, oldadp); 6108 FREE_LOCK(ump); 6109 return; 6110 } 6111 TAILQ_FOREACH(oldadp, adphead, ad_next) { 6112 if (oldadp->ad_offset >= off) 6113 break; 6114 } 6115 if (oldadp == NULL) 6116 panic("softdep_setup_allocext: lost entry"); 6117 /* insert in middle of list */ 6118 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 6119 if (oldadp->ad_offset == off) 6120 allocdirect_merge(adphead, adp, oldadp); 6121 FREE_LOCK(ump); 6122} 6123 6124/* 6125 * Indirect block allocation dependencies. 6126 * 6127 * The same dependencies that exist for a direct block also exist when 6128 * a new block is allocated and pointed to by an entry in a block of 6129 * indirect pointers. The undo/redo states described above are also 6130 * used here. Because an indirect block contains many pointers that 6131 * may have dependencies, a second copy of the entire in-memory indirect 6132 * block is kept. The buffer cache copy is always completely up-to-date. 6133 * The second copy, which is used only as a source for disk writes, 6134 * contains only the safe pointers (i.e., those that have no remaining 6135 * update dependencies). The second copy is freed when all pointers 6136 * are safe. The cache is not allowed to replace indirect blocks with 6137 * pending update dependencies. If a buffer containing an indirect 6138 * block with dependencies is written, these routines will mark it 6139 * dirty again. It can only be successfully written once all the 6140 * dependencies are removed. The ffs_fsync routine in conjunction with 6141 * softdep_sync_metadata work together to get all the dependencies 6142 * removed so that a file can be successfully written to disk. Three 6143 * procedures are used when setting up indirect block pointer 6144 * dependencies. The division is necessary because of the organization 6145 * of the "balloc" routine and because of the distinction between file 6146 * pages and file metadata blocks. 6147 */ 6148 6149/* 6150 * Allocate a new allocindir structure. 6151 */ 6152static struct allocindir * 6153newallocindir(ip, ptrno, newblkno, oldblkno, lbn) 6154 struct inode *ip; /* inode for file being extended */ 6155 int ptrno; /* offset of pointer in indirect block */ 6156 ufs2_daddr_t newblkno; /* disk block number being added */ 6157 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 6158 ufs_lbn_t lbn; 6159{ 6160 struct newblk *newblk; 6161 struct allocindir *aip; 6162 struct freefrag *freefrag; 6163 struct jnewblk *jnewblk; 6164 6165 if (oldblkno) 6166 freefrag = newfreefrag(ip, oldblkno, ITOFS(ip)->fs_bsize, lbn, 6167 SINGLETON_KEY); 6168 else 6169 freefrag = NULL; 6170 ACQUIRE_LOCK(ITOUMP(ip)); 6171 if (newblk_lookup(ITOVFS(ip), newblkno, 0, &newblk) == 0) 6172 panic("new_allocindir: lost block"); 6173 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 6174 ("newallocindir: newblk already initialized")); 6175 WORKITEM_REASSIGN(newblk, D_ALLOCINDIR); 6176 newblk->nb_freefrag = freefrag; 6177 aip = (struct allocindir *)newblk; 6178 aip->ai_offset = ptrno; 6179 aip->ai_oldblkno = oldblkno; 6180 aip->ai_lbn = lbn; 6181 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 6182 jnewblk->jn_ino = ip->i_number; 6183 jnewblk->jn_lbn = lbn; 6184 add_to_journal(&jnewblk->jn_list); 6185 } 6186 if (freefrag && freefrag->ff_jdep != NULL && 6187 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 6188 add_to_journal(freefrag->ff_jdep); 6189 return (aip); 6190} 6191 6192/* 6193 * Called just before setting an indirect block pointer 6194 * to a newly allocated file page. 6195 */ 6196void 6197softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 6198 struct inode *ip; /* inode for file being extended */ 6199 ufs_lbn_t lbn; /* allocated block number within file */ 6200 struct buf *bp; /* buffer with indirect blk referencing page */ 6201 int ptrno; /* offset of pointer in indirect block */ 6202 ufs2_daddr_t newblkno; /* disk block number being added */ 6203 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 6204 struct buf *nbp; /* buffer holding allocated page */ 6205{ 6206 struct inodedep *inodedep; 6207 struct freefrag *freefrag; 6208 struct allocindir *aip; 6209 struct pagedep *pagedep; 6210 struct mount *mp; 6211 struct ufsmount *ump; 6212 6213 mp = ITOVFS(ip); 6214 ump = VFSTOUFS(mp); 6215 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 6216 ("softdep_setup_allocindir_page called on non-softdep filesystem")); 6217 KASSERT(lbn == nbp->b_lblkno, 6218 ("softdep_setup_allocindir_page: lbn %jd != lblkno %jd", 6219 lbn, bp->b_lblkno)); 6220 CTR4(KTR_SUJ, 6221 "softdep_setup_allocindir_page: ino %d blkno %jd oldblkno %jd " 6222 "lbn %jd", ip->i_number, newblkno, oldblkno, lbn); 6223 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page"); 6224 aip = newallocindir(ip, ptrno, newblkno, oldblkno, lbn); 6225 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6226 /* 6227 * If we are allocating a directory page, then we must 6228 * allocate an associated pagedep to track additions and 6229 * deletions. 6230 */ 6231 if ((ip->i_mode & IFMT) == IFDIR) 6232 pagedep_lookup(mp, nbp, ip->i_number, lbn, DEPALLOC, &pagedep); 6233 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 6234 freefrag = setup_allocindir_phase2(bp, ip, inodedep, aip, lbn); 6235 FREE_LOCK(ump); 6236 if (freefrag) 6237 handle_workitem_freefrag(freefrag); 6238} 6239 6240/* 6241 * Called just before setting an indirect block pointer to a 6242 * newly allocated indirect block. 6243 */ 6244void 6245softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 6246 struct buf *nbp; /* newly allocated indirect block */ 6247 struct inode *ip; /* inode for file being extended */ 6248 struct buf *bp; /* indirect block referencing allocated block */ 6249 int ptrno; /* offset of pointer in indirect block */ 6250 ufs2_daddr_t newblkno; /* disk block number being added */ 6251{ 6252 struct inodedep *inodedep; 6253 struct allocindir *aip; 6254 struct ufsmount *ump; 6255 ufs_lbn_t lbn; 6256 6257 ump = ITOUMP(ip); 6258 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 6259 ("softdep_setup_allocindir_meta called on non-softdep filesystem")); 6260 CTR3(KTR_SUJ, 6261 "softdep_setup_allocindir_meta: ino %d blkno %jd ptrno %d", 6262 ip->i_number, newblkno, ptrno); 6263 lbn = nbp->b_lblkno; 6264 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta"); 6265 aip = newallocindir(ip, ptrno, newblkno, 0, lbn); 6266 inodedep_lookup(UFSTOVFS(ump), ip->i_number, DEPALLOC, &inodedep); 6267 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 6268 if (setup_allocindir_phase2(bp, ip, inodedep, aip, lbn)) 6269 panic("softdep_setup_allocindir_meta: Block already existed"); 6270 FREE_LOCK(ump); 6271} 6272 6273static void 6274indirdep_complete(indirdep) 6275 struct indirdep *indirdep; 6276{ 6277 struct allocindir *aip; 6278 6279 LIST_REMOVE(indirdep, ir_next); 6280 indirdep->ir_state |= DEPCOMPLETE; 6281 6282 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) { 6283 LIST_REMOVE(aip, ai_next); 6284 free_newblk(&aip->ai_block); 6285 } 6286 /* 6287 * If this indirdep is not attached to a buf it was simply waiting 6288 * on completion to clear completehd. free_indirdep() asserts 6289 * that nothing is dangling. 6290 */ 6291 if ((indirdep->ir_state & ONWORKLIST) == 0) 6292 free_indirdep(indirdep); 6293} 6294 6295static struct indirdep * 6296indirdep_lookup(mp, ip, bp) 6297 struct mount *mp; 6298 struct inode *ip; 6299 struct buf *bp; 6300{ 6301 struct indirdep *indirdep, *newindirdep; 6302 struct newblk *newblk; 6303 struct ufsmount *ump; 6304 struct worklist *wk; 6305 struct fs *fs; 6306 ufs2_daddr_t blkno; 6307 6308 ump = VFSTOUFS(mp); 6309 LOCK_OWNED(ump); 6310 indirdep = NULL; 6311 newindirdep = NULL; 6312 fs = ump->um_fs; 6313 for (;;) { 6314 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 6315 if (wk->wk_type != D_INDIRDEP) 6316 continue; 6317 indirdep = WK_INDIRDEP(wk); 6318 break; 6319 } 6320 /* Found on the buffer worklist, no new structure to free. */ 6321 if (indirdep != NULL && newindirdep == NULL) 6322 return (indirdep); 6323 if (indirdep != NULL && newindirdep != NULL) 6324 panic("indirdep_lookup: simultaneous create"); 6325 /* None found on the buffer and a new structure is ready. */ 6326 if (indirdep == NULL && newindirdep != NULL) 6327 break; 6328 /* None found and no new structure available. */ 6329 FREE_LOCK(ump); 6330 newindirdep = malloc(sizeof(struct indirdep), 6331 M_INDIRDEP, M_SOFTDEP_FLAGS); 6332 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, mp); 6333 newindirdep->ir_state = ATTACHED; 6334 if (I_IS_UFS1(ip)) 6335 newindirdep->ir_state |= UFS1FMT; 6336 TAILQ_INIT(&newindirdep->ir_trunc); 6337 newindirdep->ir_saveddata = NULL; 6338 LIST_INIT(&newindirdep->ir_deplisthd); 6339 LIST_INIT(&newindirdep->ir_donehd); 6340 LIST_INIT(&newindirdep->ir_writehd); 6341 LIST_INIT(&newindirdep->ir_completehd); 6342 if (bp->b_blkno == bp->b_lblkno) { 6343 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp, 6344 NULL, NULL); 6345 bp->b_blkno = blkno; 6346 } 6347 newindirdep->ir_freeblks = NULL; 6348 newindirdep->ir_savebp = 6349 getblk(ump->um_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0); 6350 newindirdep->ir_bp = bp; 6351 BUF_KERNPROC(newindirdep->ir_savebp); 6352 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 6353 ACQUIRE_LOCK(ump); 6354 } 6355 indirdep = newindirdep; 6356 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 6357 /* 6358 * If the block is not yet allocated we don't set DEPCOMPLETE so 6359 * that we don't free dependencies until the pointers are valid. 6360 * This could search b_dep for D_ALLOCDIRECT/D_ALLOCINDIR rather 6361 * than using the hash. 6362 */ 6363 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk)) 6364 LIST_INSERT_HEAD(&newblk->nb_indirdeps, indirdep, ir_next); 6365 else 6366 indirdep->ir_state |= DEPCOMPLETE; 6367 return (indirdep); 6368} 6369 6370/* 6371 * Called to finish the allocation of the "aip" allocated 6372 * by one of the two routines above. 6373 */ 6374static struct freefrag * 6375setup_allocindir_phase2(bp, ip, inodedep, aip, lbn) 6376 struct buf *bp; /* in-memory copy of the indirect block */ 6377 struct inode *ip; /* inode for file being extended */ 6378 struct inodedep *inodedep; /* Inodedep for ip */ 6379 struct allocindir *aip; /* allocindir allocated by the above routines */ 6380 ufs_lbn_t lbn; /* Logical block number for this block. */ 6381{ 6382 struct fs *fs; 6383 struct indirdep *indirdep; 6384 struct allocindir *oldaip; 6385 struct freefrag *freefrag; 6386 struct mount *mp; 6387 struct ufsmount *ump; 6388 6389 mp = ITOVFS(ip); 6390 ump = VFSTOUFS(mp); 6391 LOCK_OWNED(ump); 6392 fs = ump->um_fs; 6393 if (bp->b_lblkno >= 0) 6394 panic("setup_allocindir_phase2: not indir blk"); 6395 KASSERT(aip->ai_offset >= 0 && aip->ai_offset < NINDIR(fs), 6396 ("setup_allocindir_phase2: Bad offset %d", aip->ai_offset)); 6397 indirdep = indirdep_lookup(mp, ip, bp); 6398 KASSERT(indirdep->ir_savebp != NULL, 6399 ("setup_allocindir_phase2 NULL ir_savebp")); 6400 aip->ai_indirdep = indirdep; 6401 /* 6402 * Check for an unwritten dependency for this indirect offset. If 6403 * there is, merge the old dependency into the new one. This happens 6404 * as a result of reallocblk only. 6405 */ 6406 freefrag = NULL; 6407 if (aip->ai_oldblkno != 0) { 6408 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) { 6409 if (oldaip->ai_offset == aip->ai_offset) { 6410 freefrag = allocindir_merge(aip, oldaip); 6411 goto done; 6412 } 6413 } 6414 LIST_FOREACH(oldaip, &indirdep->ir_donehd, ai_next) { 6415 if (oldaip->ai_offset == aip->ai_offset) { 6416 freefrag = allocindir_merge(aip, oldaip); 6417 goto done; 6418 } 6419 } 6420 } 6421done: 6422 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 6423 return (freefrag); 6424} 6425 6426/* 6427 * Merge two allocindirs which refer to the same block. Move newblock 6428 * dependencies and setup the freefrags appropriately. 6429 */ 6430static struct freefrag * 6431allocindir_merge(aip, oldaip) 6432 struct allocindir *aip; 6433 struct allocindir *oldaip; 6434{ 6435 struct freefrag *freefrag; 6436 struct worklist *wk; 6437 6438 if (oldaip->ai_newblkno != aip->ai_oldblkno) 6439 panic("allocindir_merge: blkno"); 6440 aip->ai_oldblkno = oldaip->ai_oldblkno; 6441 freefrag = aip->ai_freefrag; 6442 aip->ai_freefrag = oldaip->ai_freefrag; 6443 oldaip->ai_freefrag = NULL; 6444 KASSERT(freefrag != NULL, ("setup_allocindir_phase2: No freefrag")); 6445 /* 6446 * If we are tracking a new directory-block allocation, 6447 * move it from the old allocindir to the new allocindir. 6448 */ 6449 if ((wk = LIST_FIRST(&oldaip->ai_newdirblk)) != NULL) { 6450 WORKLIST_REMOVE(wk); 6451 if (!LIST_EMPTY(&oldaip->ai_newdirblk)) 6452 panic("allocindir_merge: extra newdirblk"); 6453 WORKLIST_INSERT(&aip->ai_newdirblk, wk); 6454 } 6455 /* 6456 * We can skip journaling for this freefrag and just complete 6457 * any pending journal work for the allocindir that is being 6458 * removed after the freefrag completes. 6459 */ 6460 if (freefrag->ff_jdep) 6461 cancel_jfreefrag(WK_JFREEFRAG(freefrag->ff_jdep)); 6462 LIST_REMOVE(oldaip, ai_next); 6463 freefrag->ff_jdep = (struct worklist *)cancel_newblk(&oldaip->ai_block, 6464 &freefrag->ff_list, &freefrag->ff_jwork); 6465 free_newblk(&oldaip->ai_block); 6466 6467 return (freefrag); 6468} 6469 6470static inline void 6471setup_freedirect(freeblks, ip, i, needj) 6472 struct freeblks *freeblks; 6473 struct inode *ip; 6474 int i; 6475 int needj; 6476{ 6477 struct ufsmount *ump; 6478 ufs2_daddr_t blkno; 6479 int frags; 6480 6481 blkno = DIP(ip, i_db[i]); 6482 if (blkno == 0) 6483 return; 6484 DIP_SET(ip, i_db[i], 0); 6485 ump = ITOUMP(ip); 6486 frags = sblksize(ump->um_fs, ip->i_size, i); 6487 frags = numfrags(ump->um_fs, frags); 6488 newfreework(ump, freeblks, NULL, i, blkno, frags, 0, needj); 6489} 6490 6491static inline void 6492setup_freeext(freeblks, ip, i, needj) 6493 struct freeblks *freeblks; 6494 struct inode *ip; 6495 int i; 6496 int needj; 6497{ 6498 struct ufsmount *ump; 6499 ufs2_daddr_t blkno; 6500 int frags; 6501 6502 blkno = ip->i_din2->di_extb[i]; 6503 if (blkno == 0) 6504 return; 6505 ip->i_din2->di_extb[i] = 0; 6506 ump = ITOUMP(ip); 6507 frags = sblksize(ump->um_fs, ip->i_din2->di_extsize, i); 6508 frags = numfrags(ump->um_fs, frags); 6509 newfreework(ump, freeblks, NULL, -1 - i, blkno, frags, 0, needj); 6510} 6511 6512static inline void 6513setup_freeindir(freeblks, ip, i, lbn, needj) 6514 struct freeblks *freeblks; 6515 struct inode *ip; 6516 int i; 6517 ufs_lbn_t lbn; 6518 int needj; 6519{ 6520 struct ufsmount *ump; 6521 ufs2_daddr_t blkno; 6522 6523 blkno = DIP(ip, i_ib[i]); 6524 if (blkno == 0) 6525 return; 6526 DIP_SET(ip, i_ib[i], 0); 6527 ump = ITOUMP(ip); 6528 newfreework(ump, freeblks, NULL, lbn, blkno, ump->um_fs->fs_frag, 6529 0, needj); 6530} 6531 6532static inline struct freeblks * 6533newfreeblks(mp, ip) 6534 struct mount *mp; 6535 struct inode *ip; 6536{ 6537 struct freeblks *freeblks; 6538 6539 freeblks = malloc(sizeof(struct freeblks), 6540 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); 6541 workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp); 6542 LIST_INIT(&freeblks->fb_jblkdephd); 6543 LIST_INIT(&freeblks->fb_jwork); 6544 freeblks->fb_ref = 0; 6545 freeblks->fb_cgwait = 0; 6546 freeblks->fb_state = ATTACHED; 6547 freeblks->fb_uid = ip->i_uid; 6548 freeblks->fb_inum = ip->i_number; 6549 freeblks->fb_vtype = ITOV(ip)->v_type; 6550 freeblks->fb_modrev = DIP(ip, i_modrev); 6551 freeblks->fb_devvp = ITODEVVP(ip); 6552 freeblks->fb_chkcnt = 0; 6553 freeblks->fb_len = 0; 6554 6555 return (freeblks); 6556} 6557 6558static void 6559trunc_indirdep(indirdep, freeblks, bp, off) 6560 struct indirdep *indirdep; 6561 struct freeblks *freeblks; 6562 struct buf *bp; 6563 int off; 6564{ 6565 struct allocindir *aip, *aipn; 6566 6567 /* 6568 * The first set of allocindirs won't be in savedbp. 6569 */ 6570 LIST_FOREACH_SAFE(aip, &indirdep->ir_deplisthd, ai_next, aipn) 6571 if (aip->ai_offset > off) 6572 cancel_allocindir(aip, bp, freeblks, 1); 6573 LIST_FOREACH_SAFE(aip, &indirdep->ir_donehd, ai_next, aipn) 6574 if (aip->ai_offset > off) 6575 cancel_allocindir(aip, bp, freeblks, 1); 6576 /* 6577 * These will exist in savedbp. 6578 */ 6579 LIST_FOREACH_SAFE(aip, &indirdep->ir_writehd, ai_next, aipn) 6580 if (aip->ai_offset > off) 6581 cancel_allocindir(aip, NULL, freeblks, 0); 6582 LIST_FOREACH_SAFE(aip, &indirdep->ir_completehd, ai_next, aipn) 6583 if (aip->ai_offset > off) 6584 cancel_allocindir(aip, NULL, freeblks, 0); 6585} 6586 6587/* 6588 * Follow the chain of indirects down to lastlbn creating a freework 6589 * structure for each. This will be used to start indir_trunc() at 6590 * the right offset and create the journal records for the parrtial 6591 * truncation. A second step will handle the truncated dependencies. 6592 */ 6593static int 6594setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno) 6595 struct freeblks *freeblks; 6596 struct inode *ip; 6597 ufs_lbn_t lbn; 6598 ufs_lbn_t lastlbn; 6599 ufs2_daddr_t blkno; 6600{ 6601 struct indirdep *indirdep; 6602 struct indirdep *indirn; 6603 struct freework *freework; 6604 struct newblk *newblk; 6605 struct mount *mp; 6606 struct ufsmount *ump; 6607 struct buf *bp; 6608 uint8_t *start; 6609 uint8_t *end; 6610 ufs_lbn_t lbnadd; 6611 int level; 6612 int error; 6613 int off; 6614 6615 freework = NULL; 6616 if (blkno == 0) 6617 return (0); 6618 mp = freeblks->fb_list.wk_mp; 6619 ump = VFSTOUFS(mp); 6620 /* 6621 * Here, calls to VOP_BMAP() will fail. However, we already have 6622 * the on-disk address, so we just pass it to bread() instead of 6623 * having bread() attempt to calculate it using VOP_BMAP(). 6624 */ 6625 error = ffs_breadz(ump, ITOV(ip), lbn, blkptrtodb(ump, blkno), 6626 (int)mp->mnt_stat.f_iosize, NULL, NULL, 0, NOCRED, 0, NULL, &bp); 6627 if (error) 6628 return (error); 6629 level = lbn_level(lbn); 6630 lbnadd = lbn_offset(ump->um_fs, level); 6631 /* 6632 * Compute the offset of the last block we want to keep. Store 6633 * in the freework the first block we want to completely free. 6634 */ 6635 off = (lastlbn - -(lbn + level)) / lbnadd; 6636 if (off + 1 == NINDIR(ump->um_fs)) 6637 goto nowork; 6638 freework = newfreework(ump, freeblks, NULL, lbn, blkno, 0, off + 1, 0); 6639 /* 6640 * Link the freework into the indirdep. This will prevent any new 6641 * allocations from proceeding until we are finished with the 6642 * truncate and the block is written. 6643 */ 6644 ACQUIRE_LOCK(ump); 6645 indirdep = indirdep_lookup(mp, ip, bp); 6646 if (indirdep->ir_freeblks) 6647 panic("setup_trunc_indir: indirdep already truncated."); 6648 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, freework, fw_next); 6649 freework->fw_indir = indirdep; 6650 /* 6651 * Cancel any allocindirs that will not make it to disk. 6652 * We have to do this for all copies of the indirdep that 6653 * live on this newblk. 6654 */ 6655 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 6656 if (newblk_lookup(mp, dbtofsb(ump->um_fs, bp->b_blkno), 0, 6657 &newblk) == 0) 6658 panic("setup_trunc_indir: lost block"); 6659 LIST_FOREACH(indirn, &newblk->nb_indirdeps, ir_next) 6660 trunc_indirdep(indirn, freeblks, bp, off); 6661 } else 6662 trunc_indirdep(indirdep, freeblks, bp, off); 6663 FREE_LOCK(ump); 6664 /* 6665 * Creation is protected by the buf lock. The saveddata is only 6666 * needed if a full truncation follows a partial truncation but it 6667 * is difficult to allocate in that case so we fetch it anyway. 6668 */ 6669 if (indirdep->ir_saveddata == NULL) 6670 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 6671 M_SOFTDEP_FLAGS); 6672nowork: 6673 /* Fetch the blkno of the child and the zero start offset. */ 6674 if (I_IS_UFS1(ip)) { 6675 blkno = ((ufs1_daddr_t *)bp->b_data)[off]; 6676 start = (uint8_t *)&((ufs1_daddr_t *)bp->b_data)[off+1]; 6677 } else { 6678 blkno = ((ufs2_daddr_t *)bp->b_data)[off]; 6679 start = (uint8_t *)&((ufs2_daddr_t *)bp->b_data)[off+1]; 6680 } 6681 if (freework) { 6682 /* Zero the truncated pointers. */ 6683 end = bp->b_data + bp->b_bcount; 6684 bzero(start, end - start); 6685 bdwrite(bp); 6686 } else 6687 bqrelse(bp); 6688 if (level == 0) 6689 return (0); 6690 lbn++; /* adjust level */ 6691 lbn -= (off * lbnadd); 6692 return setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno); 6693} 6694 6695/* 6696 * Complete the partial truncation of an indirect block setup by 6697 * setup_trunc_indir(). This zeros the truncated pointers in the saved 6698 * copy and writes them to disk before the freeblks is allowed to complete. 6699 */ 6700static void 6701complete_trunc_indir(freework) 6702 struct freework *freework; 6703{ 6704 struct freework *fwn; 6705 struct indirdep *indirdep; 6706 struct ufsmount *ump; 6707 struct buf *bp; 6708 uintptr_t start; 6709 int count; 6710 6711 ump = VFSTOUFS(freework->fw_list.wk_mp); 6712 LOCK_OWNED(ump); 6713 indirdep = freework->fw_indir; 6714 for (;;) { 6715 bp = indirdep->ir_bp; 6716 /* See if the block was discarded. */ 6717 if (bp == NULL) 6718 break; 6719 /* Inline part of getdirtybuf(). We dont want bremfree. */ 6720 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) 6721 break; 6722 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 6723 LOCK_PTR(ump)) == 0) 6724 BUF_UNLOCK(bp); 6725 ACQUIRE_LOCK(ump); 6726 } 6727 freework->fw_state |= DEPCOMPLETE; 6728 TAILQ_REMOVE(&indirdep->ir_trunc, freework, fw_next); 6729 /* 6730 * Zero the pointers in the saved copy. 6731 */ 6732 if (indirdep->ir_state & UFS1FMT) 6733 start = sizeof(ufs1_daddr_t); 6734 else 6735 start = sizeof(ufs2_daddr_t); 6736 start *= freework->fw_start; 6737 count = indirdep->ir_savebp->b_bcount - start; 6738 start += (uintptr_t)indirdep->ir_savebp->b_data; 6739 bzero((char *)start, count); 6740 /* 6741 * We need to start the next truncation in the list if it has not 6742 * been started yet. 6743 */ 6744 fwn = TAILQ_FIRST(&indirdep->ir_trunc); 6745 if (fwn != NULL) { 6746 if (fwn->fw_freeblks == indirdep->ir_freeblks) 6747 TAILQ_REMOVE(&indirdep->ir_trunc, fwn, fw_next); 6748 if ((fwn->fw_state & ONWORKLIST) == 0) 6749 freework_enqueue(fwn); 6750 } 6751 /* 6752 * If bp is NULL the block was fully truncated, restore 6753 * the saved block list otherwise free it if it is no 6754 * longer needed. 6755 */ 6756 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 6757 if (bp == NULL) 6758 bcopy(indirdep->ir_saveddata, 6759 indirdep->ir_savebp->b_data, 6760 indirdep->ir_savebp->b_bcount); 6761 free(indirdep->ir_saveddata, M_INDIRDEP); 6762 indirdep->ir_saveddata = NULL; 6763 } 6764 /* 6765 * When bp is NULL there is a full truncation pending. We 6766 * must wait for this full truncation to be journaled before 6767 * we can release this freework because the disk pointers will 6768 * never be written as zero. 6769 */ 6770 if (bp == NULL) { 6771 if (LIST_EMPTY(&indirdep->ir_freeblks->fb_jblkdephd)) 6772 handle_written_freework(freework); 6773 else 6774 WORKLIST_INSERT(&indirdep->ir_freeblks->fb_freeworkhd, 6775 &freework->fw_list); 6776 if (fwn == NULL) { 6777 freework->fw_indir = (void *)0x0000deadbeef0000; 6778 bp = indirdep->ir_savebp; 6779 indirdep->ir_savebp = NULL; 6780 free_indirdep(indirdep); 6781 FREE_LOCK(ump); 6782 brelse(bp); 6783 ACQUIRE_LOCK(ump); 6784 } 6785 } else { 6786 /* Complete when the real copy is written. */ 6787 WORKLIST_INSERT(&bp->b_dep, &freework->fw_list); 6788 BUF_UNLOCK(bp); 6789 } 6790} 6791 6792/* 6793 * Calculate the number of blocks we are going to release where datablocks 6794 * is the current total and length is the new file size. 6795 */ 6796static ufs2_daddr_t 6797blkcount(fs, datablocks, length) 6798 struct fs *fs; 6799 ufs2_daddr_t datablocks; 6800 off_t length; 6801{ 6802 off_t totblks, numblks; 6803 6804 totblks = 0; 6805 numblks = howmany(length, fs->fs_bsize); 6806 if (numblks <= UFS_NDADDR) { 6807 totblks = howmany(length, fs->fs_fsize); 6808 goto out; 6809 } 6810 totblks = blkstofrags(fs, numblks); 6811 numblks -= UFS_NDADDR; 6812 /* 6813 * Count all single, then double, then triple indirects required. 6814 * Subtracting one indirects worth of blocks for each pass 6815 * acknowledges one of each pointed to by the inode. 6816 */ 6817 for (;;) { 6818 totblks += blkstofrags(fs, howmany(numblks, NINDIR(fs))); 6819 numblks -= NINDIR(fs); 6820 if (numblks <= 0) 6821 break; 6822 numblks = howmany(numblks, NINDIR(fs)); 6823 } 6824out: 6825 totblks = fsbtodb(fs, totblks); 6826 /* 6827 * Handle sparse files. We can't reclaim more blocks than the inode 6828 * references. We will correct it later in handle_complete_freeblks() 6829 * when we know the real count. 6830 */ 6831 if (totblks > datablocks) 6832 return (0); 6833 return (datablocks - totblks); 6834} 6835 6836/* 6837 * Handle freeblocks for journaled softupdate filesystems. 6838 * 6839 * Contrary to normal softupdates, we must preserve the block pointers in 6840 * indirects until their subordinates are free. This is to avoid journaling 6841 * every block that is freed which may consume more space than the journal 6842 * itself. The recovery program will see the free block journals at the 6843 * base of the truncated area and traverse them to reclaim space. The 6844 * pointers in the inode may be cleared immediately after the journal 6845 * records are written because each direct and indirect pointer in the 6846 * inode is recorded in a journal. This permits full truncation to proceed 6847 * asynchronously. The write order is journal -> inode -> cgs -> indirects. 6848 * 6849 * The algorithm is as follows: 6850 * 1) Traverse the in-memory state and create journal entries to release 6851 * the relevant blocks and full indirect trees. 6852 * 2) Traverse the indirect block chain adding partial truncation freework 6853 * records to indirects in the path to lastlbn. The freework will 6854 * prevent new allocation dependencies from being satisfied in this 6855 * indirect until the truncation completes. 6856 * 3) Read and lock the inode block, performing an update with the new size 6857 * and pointers. This prevents truncated data from becoming valid on 6858 * disk through step 4. 6859 * 4) Reap unsatisfied dependencies that are beyond the truncated area, 6860 * eliminate journal work for those records that do not require it. 6861 * 5) Schedule the journal records to be written followed by the inode block. 6862 * 6) Allocate any necessary frags for the end of file. 6863 * 7) Zero any partially truncated blocks. 6864 * 6865 * From this truncation proceeds asynchronously using the freework and 6866 * indir_trunc machinery. The file will not be extended again into a 6867 * partially truncated indirect block until all work is completed but 6868 * the normal dependency mechanism ensures that it is rolled back/forward 6869 * as appropriate. Further truncation may occur without delay and is 6870 * serialized in indir_trunc(). 6871 */ 6872void 6873softdep_journal_freeblocks(ip, cred, length, flags) 6874 struct inode *ip; /* The inode whose length is to be reduced */ 6875 struct ucred *cred; 6876 off_t length; /* The new length for the file */ 6877 int flags; /* IO_EXT and/or IO_NORMAL */ 6878{ 6879 struct freeblks *freeblks, *fbn; 6880 struct worklist *wk, *wkn; 6881 struct inodedep *inodedep; 6882 struct jblkdep *jblkdep; 6883 struct allocdirect *adp, *adpn; 6884 struct ufsmount *ump; 6885 struct fs *fs; 6886 struct buf *bp; 6887 struct vnode *vp; 6888 struct mount *mp; 6889 daddr_t dbn; 6890 ufs2_daddr_t extblocks, datablocks; 6891 ufs_lbn_t tmpval, lbn, lastlbn; 6892 int frags, lastoff, iboff, allocblock, needj, error, i; 6893 6894 ump = ITOUMP(ip); 6895 mp = UFSTOVFS(ump); 6896 fs = ump->um_fs; 6897 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 6898 ("softdep_journal_freeblocks called on non-softdep filesystem")); 6899 vp = ITOV(ip); 6900 needj = 1; 6901 iboff = -1; 6902 allocblock = 0; 6903 extblocks = 0; 6904 datablocks = 0; 6905 frags = 0; 6906 freeblks = newfreeblks(mp, ip); 6907 ACQUIRE_LOCK(ump); 6908 /* 6909 * If we're truncating a removed file that will never be written 6910 * we don't need to journal the block frees. The canceled journals 6911 * for the allocations will suffice. 6912 */ 6913 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6914 if ((inodedep->id_state & (UNLINKED | DEPCOMPLETE)) == UNLINKED && 6915 length == 0) 6916 needj = 0; 6917 CTR3(KTR_SUJ, "softdep_journal_freeblks: ip %d length %ld needj %d", 6918 ip->i_number, length, needj); 6919 FREE_LOCK(ump); 6920 /* 6921 * Calculate the lbn that we are truncating to. This results in -1 6922 * if we're truncating the 0 bytes. So it is the last lbn we want 6923 * to keep, not the first lbn we want to truncate. 6924 */ 6925 lastlbn = lblkno(fs, length + fs->fs_bsize - 1) - 1; 6926 lastoff = blkoff(fs, length); 6927 /* 6928 * Compute frags we are keeping in lastlbn. 0 means all. 6929 */ 6930 if (lastlbn >= 0 && lastlbn < UFS_NDADDR) { 6931 frags = fragroundup(fs, lastoff); 6932 /* adp offset of last valid allocdirect. */ 6933 iboff = lastlbn; 6934 } else if (lastlbn > 0) 6935 iboff = UFS_NDADDR; 6936 if (fs->fs_magic == FS_UFS2_MAGIC) 6937 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 6938 /* 6939 * Handle normal data blocks and indirects. This section saves 6940 * values used after the inode update to complete frag and indirect 6941 * truncation. 6942 */ 6943 if ((flags & IO_NORMAL) != 0) { 6944 /* 6945 * Handle truncation of whole direct and indirect blocks. 6946 */ 6947 for (i = iboff + 1; i < UFS_NDADDR; i++) 6948 setup_freedirect(freeblks, ip, i, needj); 6949 for (i = 0, tmpval = NINDIR(fs), lbn = UFS_NDADDR; 6950 i < UFS_NIADDR; 6951 i++, lbn += tmpval, tmpval *= NINDIR(fs)) { 6952 /* Release a whole indirect tree. */ 6953 if (lbn > lastlbn) { 6954 setup_freeindir(freeblks, ip, i, -lbn -i, 6955 needj); 6956 continue; 6957 } 6958 iboff = i + UFS_NDADDR; 6959 /* 6960 * Traverse partially truncated indirect tree. 6961 */ 6962 if (lbn <= lastlbn && lbn + tmpval - 1 > lastlbn) 6963 setup_trunc_indir(freeblks, ip, -lbn - i, 6964 lastlbn, DIP(ip, i_ib[i])); 6965 } 6966 /* 6967 * Handle partial truncation to a frag boundary. 6968 */ 6969 if (frags) { 6970 ufs2_daddr_t blkno; 6971 long oldfrags; 6972 6973 oldfrags = blksize(fs, ip, lastlbn); 6974 blkno = DIP(ip, i_db[lastlbn]); 6975 if (blkno && oldfrags != frags) { 6976 oldfrags -= frags; 6977 oldfrags = numfrags(fs, oldfrags); 6978 blkno += numfrags(fs, frags); 6979 newfreework(ump, freeblks, NULL, lastlbn, 6980 blkno, oldfrags, 0, needj); 6981 if (needj) 6982 adjust_newfreework(freeblks, 6983 numfrags(fs, frags)); 6984 } else if (blkno == 0) 6985 allocblock = 1; 6986 } 6987 /* 6988 * Add a journal record for partial truncate if we are 6989 * handling indirect blocks. Non-indirects need no extra 6990 * journaling. 6991 */ 6992 if (length != 0 && lastlbn >= UFS_NDADDR) { 6993 UFS_INODE_SET_FLAG(ip, IN_TRUNCATED); 6994 newjtrunc(freeblks, length, 0); 6995 } 6996 ip->i_size = length; 6997 DIP_SET(ip, i_size, ip->i_size); 6998 UFS_INODE_SET_FLAG(ip, IN_SIZEMOD | IN_CHANGE); 6999 datablocks = DIP(ip, i_blocks) - extblocks; 7000 if (length != 0) 7001 datablocks = blkcount(fs, datablocks, length); 7002 freeblks->fb_len = length; 7003 } 7004 if ((flags & IO_EXT) != 0) { 7005 for (i = 0; i < UFS_NXADDR; i++) 7006 setup_freeext(freeblks, ip, i, needj); 7007 ip->i_din2->di_extsize = 0; 7008 datablocks += extblocks; 7009 UFS_INODE_SET_FLAG(ip, IN_SIZEMOD | IN_CHANGE); 7010 } 7011#ifdef QUOTA 7012 /* Reference the quotas in case the block count is wrong in the end. */ 7013 quotaref(vp, freeblks->fb_quota); 7014 (void) chkdq(ip, -datablocks, NOCRED, FORCE); 7015#endif 7016 freeblks->fb_chkcnt = -datablocks; 7017 UFS_LOCK(ump); 7018 fs->fs_pendingblocks += datablocks; 7019 UFS_UNLOCK(ump); 7020 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 7021 /* 7022 * Handle truncation of incomplete alloc direct dependencies. We 7023 * hold the inode block locked to prevent incomplete dependencies 7024 * from reaching the disk while we are eliminating those that 7025 * have been truncated. This is a partially inlined ffs_update(). 7026 */ 7027 ufs_itimes(vp); 7028 ip->i_flag &= ~(IN_LAZYACCESS | IN_LAZYMOD | IN_MODIFIED); 7029 dbn = fsbtodb(fs, ino_to_fsba(fs, ip->i_number)); 7030 error = ffs_breadz(ump, ump->um_devvp, dbn, dbn, (int)fs->fs_bsize, 7031 NULL, NULL, 0, cred, 0, NULL, &bp); 7032 if (error) { 7033 softdep_error("softdep_journal_freeblocks", error); 7034 return; 7035 } 7036 if (bp->b_bufsize == fs->fs_bsize) 7037 bp->b_flags |= B_CLUSTEROK; 7038 softdep_update_inodeblock(ip, bp, 0); 7039 if (ump->um_fstype == UFS1) { 7040 *((struct ufs1_dinode *)bp->b_data + 7041 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1; 7042 } else { 7043 ffs_update_dinode_ckhash(fs, ip->i_din2); 7044 *((struct ufs2_dinode *)bp->b_data + 7045 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2; 7046 } 7047 ACQUIRE_LOCK(ump); 7048 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 7049 if ((inodedep->id_state & IOSTARTED) != 0) 7050 panic("softdep_setup_freeblocks: inode busy"); 7051 /* 7052 * Add the freeblks structure to the list of operations that 7053 * must await the zero'ed inode being written to disk. If we 7054 * still have a bitmap dependency (needj), then the inode 7055 * has never been written to disk, so we can process the 7056 * freeblks below once we have deleted the dependencies. 7057 */ 7058 if (needj) 7059 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 7060 else 7061 freeblks->fb_state |= COMPLETE; 7062 if ((flags & IO_NORMAL) != 0) { 7063 TAILQ_FOREACH_SAFE(adp, &inodedep->id_inoupdt, ad_next, adpn) { 7064 if (adp->ad_offset > iboff) 7065 cancel_allocdirect(&inodedep->id_inoupdt, adp, 7066 freeblks); 7067 /* 7068 * Truncate the allocdirect. We could eliminate 7069 * or modify journal records as well. 7070 */ 7071 else if (adp->ad_offset == iboff && frags) 7072 adp->ad_newsize = frags; 7073 } 7074 } 7075 if ((flags & IO_EXT) != 0) 7076 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 7077 cancel_allocdirect(&inodedep->id_extupdt, adp, 7078 freeblks); 7079 /* 7080 * Scan the bufwait list for newblock dependencies that will never 7081 * make it to disk. 7082 */ 7083 LIST_FOREACH_SAFE(wk, &inodedep->id_bufwait, wk_list, wkn) { 7084 if (wk->wk_type != D_ALLOCDIRECT) 7085 continue; 7086 adp = WK_ALLOCDIRECT(wk); 7087 if (((flags & IO_NORMAL) != 0 && (adp->ad_offset > iboff)) || 7088 ((flags & IO_EXT) != 0 && (adp->ad_state & EXTDATA))) { 7089 cancel_jfreeblk(freeblks, adp->ad_newblkno); 7090 cancel_newblk(WK_NEWBLK(wk), NULL, &freeblks->fb_jwork); 7091 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk); 7092 } 7093 } 7094 /* 7095 * Add journal work. 7096 */ 7097 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) 7098 add_to_journal(&jblkdep->jb_list); 7099 FREE_LOCK(ump); 7100 bdwrite(bp); 7101 /* 7102 * Truncate dependency structures beyond length. 7103 */ 7104 trunc_dependencies(ip, freeblks, lastlbn, frags, flags); 7105 /* 7106 * This is only set when we need to allocate a fragment because 7107 * none existed at the end of a frag-sized file. It handles only 7108 * allocating a new, zero filled block. 7109 */ 7110 if (allocblock) { 7111 ip->i_size = length - lastoff; 7112 DIP_SET(ip, i_size, ip->i_size); 7113 error = UFS_BALLOC(vp, length - 1, 1, cred, BA_CLRBUF, &bp); 7114 if (error != 0) { 7115 softdep_error("softdep_journal_freeblks", error); 7116 return; 7117 } 7118 ip->i_size = length; 7119 DIP_SET(ip, i_size, length); 7120 UFS_INODE_SET_FLAG(ip, IN_SIZEMOD | IN_CHANGE | IN_UPDATE); 7121 allocbuf(bp, frags); 7122 ffs_update(vp, 0); 7123 bawrite(bp); 7124 } else if (lastoff != 0 && vp->v_type != VDIR) { 7125 int size; 7126 7127 /* 7128 * Zero the end of a truncated frag or block. 7129 */ 7130 size = sblksize(fs, length, lastlbn); 7131 error = bread(vp, lastlbn, size, cred, &bp); 7132 if (error == 0) { 7133 bzero((char *)bp->b_data + lastoff, size - lastoff); 7134 bawrite(bp); 7135 } else if (!ffs_fsfail_cleanup(ump, error)) { 7136 softdep_error("softdep_journal_freeblks", error); 7137 return; 7138 } 7139 } 7140 ACQUIRE_LOCK(ump); 7141 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 7142 TAILQ_INSERT_TAIL(&inodedep->id_freeblklst, freeblks, fb_next); 7143 freeblks->fb_state |= DEPCOMPLETE | ONDEPLIST; 7144 /* 7145 * We zero earlier truncations so they don't erroneously 7146 * update i_blocks. 7147 */ 7148 if (freeblks->fb_len == 0 && (flags & IO_NORMAL) != 0) 7149 TAILQ_FOREACH(fbn, &inodedep->id_freeblklst, fb_next) 7150 fbn->fb_len = 0; 7151 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE && 7152 LIST_EMPTY(&freeblks->fb_jblkdephd)) 7153 freeblks->fb_state |= INPROGRESS; 7154 else 7155 freeblks = NULL; 7156 FREE_LOCK(ump); 7157 if (freeblks) 7158 handle_workitem_freeblocks(freeblks, 0); 7159 trunc_pages(ip, length, extblocks, flags); 7160 7161} 7162 7163/* 7164 * Flush a JOP_SYNC to the journal. 7165 */ 7166void 7167softdep_journal_fsync(ip) 7168 struct inode *ip; 7169{ 7170 struct jfsync *jfsync; 7171 struct ufsmount *ump; 7172 7173 ump = ITOUMP(ip); 7174 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 7175 ("softdep_journal_fsync called on non-softdep filesystem")); 7176 if ((ip->i_flag & IN_TRUNCATED) == 0) 7177 return; 7178 ip->i_flag &= ~IN_TRUNCATED; 7179 jfsync = malloc(sizeof(*jfsync), M_JFSYNC, M_SOFTDEP_FLAGS | M_ZERO); 7180 workitem_alloc(&jfsync->jfs_list, D_JFSYNC, UFSTOVFS(ump)); 7181 jfsync->jfs_size = ip->i_size; 7182 jfsync->jfs_ino = ip->i_number; 7183 ACQUIRE_LOCK(ump); 7184 add_to_journal(&jfsync->jfs_list); 7185 jwait(&jfsync->jfs_list, MNT_WAIT); 7186 FREE_LOCK(ump); 7187} 7188 7189/* 7190 * Block de-allocation dependencies. 7191 * 7192 * When blocks are de-allocated, the on-disk pointers must be nullified before 7193 * the blocks are made available for use by other files. (The true 7194 * requirement is that old pointers must be nullified before new on-disk 7195 * pointers are set. We chose this slightly more stringent requirement to 7196 * reduce complexity.) Our implementation handles this dependency by updating 7197 * the inode (or indirect block) appropriately but delaying the actual block 7198 * de-allocation (i.e., freemap and free space count manipulation) until 7199 * after the updated versions reach stable storage. After the disk is 7200 * updated, the blocks can be safely de-allocated whenever it is convenient. 7201 * This implementation handles only the common case of reducing a file's 7202 * length to zero. Other cases are handled by the conventional synchronous 7203 * write approach. 7204 * 7205 * The ffs implementation with which we worked double-checks 7206 * the state of the block pointers and file size as it reduces 7207 * a file's length. Some of this code is replicated here in our 7208 * soft updates implementation. The freeblks->fb_chkcnt field is 7209 * used to transfer a part of this information to the procedure 7210 * that eventually de-allocates the blocks. 7211 * 7212 * This routine should be called from the routine that shortens 7213 * a file's length, before the inode's size or block pointers 7214 * are modified. It will save the block pointer information for 7215 * later release and zero the inode so that the calling routine 7216 * can release it. 7217 */ 7218void 7219softdep_setup_freeblocks(ip, length, flags) 7220 struct inode *ip; /* The inode whose length is to be reduced */ 7221 off_t length; /* The new length for the file */ 7222 int flags; /* IO_EXT and/or IO_NORMAL */ 7223{ 7224 struct ufs1_dinode *dp1; 7225 struct ufs2_dinode *dp2; 7226 struct freeblks *freeblks; 7227 struct inodedep *inodedep; 7228 struct allocdirect *adp; 7229 struct ufsmount *ump; 7230 struct buf *bp; 7231 struct fs *fs; 7232 ufs2_daddr_t extblocks, datablocks; 7233 struct mount *mp; 7234 int i, delay, error; 7235 ufs_lbn_t tmpval; 7236 ufs_lbn_t lbn; 7237 7238 ump = ITOUMP(ip); 7239 mp = UFSTOVFS(ump); 7240 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 7241 ("softdep_setup_freeblocks called on non-softdep filesystem")); 7242 CTR2(KTR_SUJ, "softdep_setup_freeblks: ip %d length %ld", 7243 ip->i_number, length); 7244 KASSERT(length == 0, ("softdep_setup_freeblocks: non-zero length")); 7245 fs = ump->um_fs; 7246 if ((error = bread(ump->um_devvp, 7247 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 7248 (int)fs->fs_bsize, NOCRED, &bp)) != 0) { 7249 if (!ffs_fsfail_cleanup(ump, error)) 7250 softdep_error("softdep_setup_freeblocks", error); 7251 return; 7252 } 7253 freeblks = newfreeblks(mp, ip); 7254 extblocks = 0; 7255 datablocks = 0; 7256 if (fs->fs_magic == FS_UFS2_MAGIC) 7257 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 7258 if ((flags & IO_NORMAL) != 0) { 7259 for (i = 0; i < UFS_NDADDR; i++) 7260 setup_freedirect(freeblks, ip, i, 0); 7261 for (i = 0, tmpval = NINDIR(fs), lbn = UFS_NDADDR; 7262 i < UFS_NIADDR; 7263 i++, lbn += tmpval, tmpval *= NINDIR(fs)) 7264 setup_freeindir(freeblks, ip, i, -lbn -i, 0); 7265 ip->i_size = 0; 7266 DIP_SET(ip, i_size, 0); 7267 UFS_INODE_SET_FLAG(ip, IN_SIZEMOD | IN_CHANGE); 7268 datablocks = DIP(ip, i_blocks) - extblocks; 7269 } 7270 if ((flags & IO_EXT) != 0) { 7271 for (i = 0; i < UFS_NXADDR; i++) 7272 setup_freeext(freeblks, ip, i, 0); 7273 ip->i_din2->di_extsize = 0; 7274 datablocks += extblocks; 7275 UFS_INODE_SET_FLAG(ip, IN_SIZEMOD | IN_CHANGE); 7276 } 7277#ifdef QUOTA 7278 /* Reference the quotas in case the block count is wrong in the end. */ 7279 quotaref(ITOV(ip), freeblks->fb_quota); 7280 (void) chkdq(ip, -datablocks, NOCRED, FORCE); 7281#endif 7282 freeblks->fb_chkcnt = -datablocks; 7283 UFS_LOCK(ump); 7284 fs->fs_pendingblocks += datablocks; 7285 UFS_UNLOCK(ump); 7286 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 7287 /* 7288 * Push the zero'ed inode to its disk buffer so that we are free 7289 * to delete its dependencies below. Once the dependencies are gone 7290 * the buffer can be safely released. 7291 */ 7292 if (ump->um_fstype == UFS1) { 7293 dp1 = ((struct ufs1_dinode *)bp->b_data + 7294 ino_to_fsbo(fs, ip->i_number)); 7295 ip->i_din1->di_freelink = dp1->di_freelink; 7296 *dp1 = *ip->i_din1; 7297 } else { 7298 dp2 = ((struct ufs2_dinode *)bp->b_data + 7299 ino_to_fsbo(fs, ip->i_number)); 7300 ip->i_din2->di_freelink = dp2->di_freelink; 7301 ffs_update_dinode_ckhash(fs, ip->i_din2); 7302 *dp2 = *ip->i_din2; 7303 } 7304 /* 7305 * Find and eliminate any inode dependencies. 7306 */ 7307 ACQUIRE_LOCK(ump); 7308 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 7309 if ((inodedep->id_state & IOSTARTED) != 0) 7310 panic("softdep_setup_freeblocks: inode busy"); 7311 /* 7312 * Add the freeblks structure to the list of operations that 7313 * must await the zero'ed inode being written to disk. If we 7314 * still have a bitmap dependency (delay == 0), then the inode 7315 * has never been written to disk, so we can process the 7316 * freeblks below once we have deleted the dependencies. 7317 */ 7318 delay = (inodedep->id_state & DEPCOMPLETE); 7319 if (delay) 7320 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 7321 else 7322 freeblks->fb_state |= COMPLETE; 7323 /* 7324 * Because the file length has been truncated to zero, any 7325 * pending block allocation dependency structures associated 7326 * with this inode are obsolete and can simply be de-allocated. 7327 * We must first merge the two dependency lists to get rid of 7328 * any duplicate freefrag structures, then purge the merged list. 7329 * If we still have a bitmap dependency, then the inode has never 7330 * been written to disk, so we can free any fragments without delay. 7331 */ 7332 if (flags & IO_NORMAL) { 7333 merge_inode_lists(&inodedep->id_newinoupdt, 7334 &inodedep->id_inoupdt); 7335 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 7336 cancel_allocdirect(&inodedep->id_inoupdt, adp, 7337 freeblks); 7338 } 7339 if (flags & IO_EXT) { 7340 merge_inode_lists(&inodedep->id_newextupdt, 7341 &inodedep->id_extupdt); 7342 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 7343 cancel_allocdirect(&inodedep->id_extupdt, adp, 7344 freeblks); 7345 } 7346 FREE_LOCK(ump); 7347 bdwrite(bp); 7348 trunc_dependencies(ip, freeblks, -1, 0, flags); 7349 ACQUIRE_LOCK(ump); 7350 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 7351 (void) free_inodedep(inodedep); 7352 freeblks->fb_state |= DEPCOMPLETE; 7353 /* 7354 * If the inode with zeroed block pointers is now on disk 7355 * we can start freeing blocks. 7356 */ 7357 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 7358 freeblks->fb_state |= INPROGRESS; 7359 else 7360 freeblks = NULL; 7361 FREE_LOCK(ump); 7362 if (freeblks) 7363 handle_workitem_freeblocks(freeblks, 0); 7364 trunc_pages(ip, length, extblocks, flags); 7365} 7366 7367/* 7368 * Eliminate pages from the page cache that back parts of this inode and 7369 * adjust the vnode pager's idea of our size. This prevents stale data 7370 * from hanging around in the page cache. 7371 */ 7372static void 7373trunc_pages(ip, length, extblocks, flags) 7374 struct inode *ip; 7375 off_t length; 7376 ufs2_daddr_t extblocks; 7377 int flags; 7378{ 7379 struct vnode *vp; 7380 struct fs *fs; 7381 ufs_lbn_t lbn; 7382 off_t end, extend; 7383 7384 vp = ITOV(ip); 7385 fs = ITOFS(ip); 7386 extend = OFF_TO_IDX(lblktosize(fs, -extblocks)); 7387 if ((flags & IO_EXT) != 0) 7388 vn_pages_remove(vp, extend, 0); 7389 if ((flags & IO_NORMAL) == 0) 7390 return; 7391 BO_LOCK(&vp->v_bufobj); 7392 drain_output(vp); 7393 BO_UNLOCK(&vp->v_bufobj); 7394 /* 7395 * The vnode pager eliminates file pages we eliminate indirects 7396 * below. 7397 */ 7398 vnode_pager_setsize(vp, length); 7399 /* 7400 * Calculate the end based on the last indirect we want to keep. If 7401 * the block extends into indirects we can just use the negative of 7402 * its lbn. Doubles and triples exist at lower numbers so we must 7403 * be careful not to remove those, if they exist. double and triple 7404 * indirect lbns do not overlap with others so it is not important 7405 * to verify how many levels are required. 7406 */ 7407 lbn = lblkno(fs, length); 7408 if (lbn >= UFS_NDADDR) { 7409 /* Calculate the virtual lbn of the triple indirect. */ 7410 lbn = -lbn - (UFS_NIADDR - 1); 7411 end = OFF_TO_IDX(lblktosize(fs, lbn)); 7412 } else 7413 end = extend; 7414 vn_pages_remove(vp, OFF_TO_IDX(OFF_MAX), end); 7415} 7416 7417/* 7418 * See if the buf bp is in the range eliminated by truncation. 7419 */ 7420static int 7421trunc_check_buf(bp, blkoffp, lastlbn, lastoff, flags) 7422 struct buf *bp; 7423 int *blkoffp; 7424 ufs_lbn_t lastlbn; 7425 int lastoff; 7426 int flags; 7427{ 7428 ufs_lbn_t lbn; 7429 7430 *blkoffp = 0; 7431 /* Only match ext/normal blocks as appropriate. */ 7432 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) || 7433 ((flags & IO_NORMAL) == 0 && (bp->b_xflags & BX_ALTDATA) == 0)) 7434 return (0); 7435 /* ALTDATA is always a full truncation. */ 7436 if ((bp->b_xflags & BX_ALTDATA) != 0) 7437 return (1); 7438 /* -1 is full truncation. */ 7439 if (lastlbn == -1) 7440 return (1); 7441 /* 7442 * If this is a partial truncate we only want those 7443 * blocks and indirect blocks that cover the range 7444 * we're after. 7445 */ 7446 lbn = bp->b_lblkno; 7447 if (lbn < 0) 7448 lbn = -(lbn + lbn_level(lbn)); 7449 if (lbn < lastlbn) 7450 return (0); 7451 /* Here we only truncate lblkno if it's partial. */ 7452 if (lbn == lastlbn) { 7453 if (lastoff == 0) 7454 return (0); 7455 *blkoffp = lastoff; 7456 } 7457 return (1); 7458} 7459 7460/* 7461 * Eliminate any dependencies that exist in memory beyond lblkno:off 7462 */ 7463static void 7464trunc_dependencies(ip, freeblks, lastlbn, lastoff, flags) 7465 struct inode *ip; 7466 struct freeblks *freeblks; 7467 ufs_lbn_t lastlbn; 7468 int lastoff; 7469 int flags; 7470{ 7471 struct bufobj *bo; 7472 struct vnode *vp; 7473 struct buf *bp; 7474 int blkoff; 7475 7476 /* 7477 * We must wait for any I/O in progress to finish so that 7478 * all potential buffers on the dirty list will be visible. 7479 * Once they are all there, walk the list and get rid of 7480 * any dependencies. 7481 */ 7482 vp = ITOV(ip); 7483 bo = &vp->v_bufobj; 7484 BO_LOCK(bo); 7485 drain_output(vp); 7486 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) 7487 bp->b_vflags &= ~BV_SCANNED; 7488restart: 7489 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { 7490 if (bp->b_vflags & BV_SCANNED) 7491 continue; 7492 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 7493 bp->b_vflags |= BV_SCANNED; 7494 continue; 7495 } 7496 KASSERT(bp->b_bufobj == bo, ("Wrong object in buffer")); 7497 if ((bp = getdirtybuf(bp, BO_LOCKPTR(bo), MNT_WAIT)) == NULL) 7498 goto restart; 7499 BO_UNLOCK(bo); 7500 if (deallocate_dependencies(bp, freeblks, blkoff)) 7501 bqrelse(bp); 7502 else 7503 brelse(bp); 7504 BO_LOCK(bo); 7505 goto restart; 7506 } 7507 /* 7508 * Now do the work of vtruncbuf while also matching indirect blocks. 7509 */ 7510 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) 7511 bp->b_vflags &= ~BV_SCANNED; 7512cleanrestart: 7513 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) { 7514 if (bp->b_vflags & BV_SCANNED) 7515 continue; 7516 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 7517 bp->b_vflags |= BV_SCANNED; 7518 continue; 7519 } 7520 if (BUF_LOCK(bp, 7521 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 7522 BO_LOCKPTR(bo)) == ENOLCK) { 7523 BO_LOCK(bo); 7524 goto cleanrestart; 7525 } 7526 BO_LOCK(bo); 7527 bp->b_vflags |= BV_SCANNED; 7528 BO_UNLOCK(bo); 7529 bremfree(bp); 7530 if (blkoff != 0) { 7531 allocbuf(bp, blkoff); 7532 bqrelse(bp); 7533 } else { 7534 bp->b_flags |= B_INVAL | B_NOCACHE | B_RELBUF; 7535 brelse(bp); 7536 } 7537 BO_LOCK(bo); 7538 goto cleanrestart; 7539 } 7540 drain_output(vp); 7541 BO_UNLOCK(bo); 7542} 7543 7544static int 7545cancel_pagedep(pagedep, freeblks, blkoff) 7546 struct pagedep *pagedep; 7547 struct freeblks *freeblks; 7548 int blkoff; 7549{ 7550 struct jremref *jremref; 7551 struct jmvref *jmvref; 7552 struct dirrem *dirrem, *tmp; 7553 int i; 7554 7555 /* 7556 * Copy any directory remove dependencies to the list 7557 * to be processed after the freeblks proceeds. If 7558 * directory entry never made it to disk they 7559 * can be dumped directly onto the work list. 7560 */ 7561 LIST_FOREACH_SAFE(dirrem, &pagedep->pd_dirremhd, dm_next, tmp) { 7562 /* Skip this directory removal if it is intended to remain. */ 7563 if (dirrem->dm_offset < blkoff) 7564 continue; 7565 /* 7566 * If there are any dirrems we wait for the journal write 7567 * to complete and then restart the buf scan as the lock 7568 * has been dropped. 7569 */ 7570 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) { 7571 jwait(&jremref->jr_list, MNT_WAIT); 7572 return (ERESTART); 7573 } 7574 LIST_REMOVE(dirrem, dm_next); 7575 dirrem->dm_dirinum = pagedep->pd_ino; 7576 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &dirrem->dm_list); 7577 } 7578 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) { 7579 jwait(&jmvref->jm_list, MNT_WAIT); 7580 return (ERESTART); 7581 } 7582 /* 7583 * When we're partially truncating a pagedep we just want to flush 7584 * journal entries and return. There can not be any adds in the 7585 * truncated portion of the directory and newblk must remain if 7586 * part of the block remains. 7587 */ 7588 if (blkoff != 0) { 7589 struct diradd *dap; 7590 7591 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 7592 if (dap->da_offset > blkoff) 7593 panic("cancel_pagedep: diradd %p off %d > %d", 7594 dap, dap->da_offset, blkoff); 7595 for (i = 0; i < DAHASHSZ; i++) 7596 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) 7597 if (dap->da_offset > blkoff) 7598 panic("cancel_pagedep: diradd %p off %d > %d", 7599 dap, dap->da_offset, blkoff); 7600 return (0); 7601 } 7602 /* 7603 * There should be no directory add dependencies present 7604 * as the directory could not be truncated until all 7605 * children were removed. 7606 */ 7607 KASSERT(LIST_FIRST(&pagedep->pd_pendinghd) == NULL, 7608 ("deallocate_dependencies: pendinghd != NULL")); 7609 for (i = 0; i < DAHASHSZ; i++) 7610 KASSERT(LIST_FIRST(&pagedep->pd_diraddhd[i]) == NULL, 7611 ("deallocate_dependencies: diraddhd != NULL")); 7612 if ((pagedep->pd_state & NEWBLOCK) != 0) 7613 free_newdirblk(pagedep->pd_newdirblk); 7614 if (free_pagedep(pagedep) == 0) 7615 panic("Failed to free pagedep %p", pagedep); 7616 return (0); 7617} 7618 7619/* 7620 * Reclaim any dependency structures from a buffer that is about to 7621 * be reallocated to a new vnode. The buffer must be locked, thus, 7622 * no I/O completion operations can occur while we are manipulating 7623 * its associated dependencies. The mutex is held so that other I/O's 7624 * associated with related dependencies do not occur. 7625 */ 7626static int 7627deallocate_dependencies(bp, freeblks, off) 7628 struct buf *bp; 7629 struct freeblks *freeblks; 7630 int off; 7631{ 7632 struct indirdep *indirdep; 7633 struct pagedep *pagedep; 7634 struct worklist *wk, *wkn; 7635 struct ufsmount *ump; 7636 7637 ump = softdep_bp_to_mp(bp); 7638 if (ump == NULL) 7639 goto done; 7640 ACQUIRE_LOCK(ump); 7641 LIST_FOREACH_SAFE(wk, &bp->b_dep, wk_list, wkn) { 7642 switch (wk->wk_type) { 7643 case D_INDIRDEP: 7644 indirdep = WK_INDIRDEP(wk); 7645 if (bp->b_lblkno >= 0 || 7646 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 7647 panic("deallocate_dependencies: not indir"); 7648 cancel_indirdep(indirdep, bp, freeblks); 7649 continue; 7650 7651 case D_PAGEDEP: 7652 pagedep = WK_PAGEDEP(wk); 7653 if (cancel_pagedep(pagedep, freeblks, off)) { 7654 FREE_LOCK(ump); 7655 return (ERESTART); 7656 } 7657 continue; 7658 7659 case D_ALLOCINDIR: 7660 /* 7661 * Simply remove the allocindir, we'll find it via 7662 * the indirdep where we can clear pointers if 7663 * needed. 7664 */ 7665 WORKLIST_REMOVE(wk); 7666 continue; 7667 7668 case D_FREEWORK: 7669 /* 7670 * A truncation is waiting for the zero'd pointers 7671 * to be written. It can be freed when the freeblks 7672 * is journaled. 7673 */ 7674 WORKLIST_REMOVE(wk); 7675 wk->wk_state |= ONDEPLIST; 7676 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk); 7677 break; 7678 7679 case D_ALLOCDIRECT: 7680 if (off != 0) 7681 continue; 7682 /* FALLTHROUGH */ 7683 default: 7684 panic("deallocate_dependencies: Unexpected type %s", 7685 TYPENAME(wk->wk_type)); 7686 /* NOTREACHED */ 7687 } 7688 } 7689 FREE_LOCK(ump); 7690done: 7691 /* 7692 * Don't throw away this buf, we were partially truncating and 7693 * some deps may always remain. 7694 */ 7695 if (off) { 7696 allocbuf(bp, off); 7697 bp->b_vflags |= BV_SCANNED; 7698 return (EBUSY); 7699 } 7700 bp->b_flags |= B_INVAL | B_NOCACHE; 7701 7702 return (0); 7703} 7704 7705/* 7706 * An allocdirect is being canceled due to a truncate. We must make sure 7707 * the journal entry is released in concert with the blkfree that releases 7708 * the storage. Completed journal entries must not be released until the 7709 * space is no longer pointed to by the inode or in the bitmap. 7710 */ 7711static void 7712cancel_allocdirect(adphead, adp, freeblks) 7713 struct allocdirectlst *adphead; 7714 struct allocdirect *adp; 7715 struct freeblks *freeblks; 7716{ 7717 struct freework *freework; 7718 struct newblk *newblk; 7719 struct worklist *wk; 7720 7721 TAILQ_REMOVE(adphead, adp, ad_next); 7722 newblk = (struct newblk *)adp; 7723 freework = NULL; 7724 /* 7725 * Find the correct freework structure. 7726 */ 7727 LIST_FOREACH(wk, &freeblks->fb_freeworkhd, wk_list) { 7728 if (wk->wk_type != D_FREEWORK) 7729 continue; 7730 freework = WK_FREEWORK(wk); 7731 if (freework->fw_blkno == newblk->nb_newblkno) 7732 break; 7733 } 7734 if (freework == NULL) 7735 panic("cancel_allocdirect: Freework not found"); 7736 /* 7737 * If a newblk exists at all we still have the journal entry that 7738 * initiated the allocation so we do not need to journal the free. 7739 */ 7740 cancel_jfreeblk(freeblks, freework->fw_blkno); 7741 /* 7742 * If the journal hasn't been written the jnewblk must be passed 7743 * to the call to ffs_blkfree that reclaims the space. We accomplish 7744 * this by linking the journal dependency into the freework to be 7745 * freed when freework_freeblock() is called. If the journal has 7746 * been written we can simply reclaim the journal space when the 7747 * freeblks work is complete. 7748 */ 7749 freework->fw_jnewblk = cancel_newblk(newblk, &freework->fw_list, 7750 &freeblks->fb_jwork); 7751 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 7752} 7753 7754/* 7755 * Cancel a new block allocation. May be an indirect or direct block. We 7756 * remove it from various lists and return any journal record that needs to 7757 * be resolved by the caller. 7758 * 7759 * A special consideration is made for indirects which were never pointed 7760 * at on disk and will never be found once this block is released. 7761 */ 7762static struct jnewblk * 7763cancel_newblk(newblk, wk, wkhd) 7764 struct newblk *newblk; 7765 struct worklist *wk; 7766 struct workhead *wkhd; 7767{ 7768 struct jnewblk *jnewblk; 7769 7770 CTR1(KTR_SUJ, "cancel_newblk: blkno %jd", newblk->nb_newblkno); 7771 7772 newblk->nb_state |= GOINGAWAY; 7773 /* 7774 * Previously we traversed the completedhd on each indirdep 7775 * attached to this newblk to cancel them and gather journal 7776 * work. Since we need only the oldest journal segment and 7777 * the lowest point on the tree will always have the oldest 7778 * journal segment we are free to release the segments 7779 * of any subordinates and may leave the indirdep list to 7780 * indirdep_complete() when this newblk is freed. 7781 */ 7782 if (newblk->nb_state & ONDEPLIST) { 7783 newblk->nb_state &= ~ONDEPLIST; 7784 LIST_REMOVE(newblk, nb_deps); 7785 } 7786 if (newblk->nb_state & ONWORKLIST) 7787 WORKLIST_REMOVE(&newblk->nb_list); 7788 /* 7789 * If the journal entry hasn't been written we save a pointer to 7790 * the dependency that frees it until it is written or the 7791 * superseding operation completes. 7792 */ 7793 jnewblk = newblk->nb_jnewblk; 7794 if (jnewblk != NULL && wk != NULL) { 7795 newblk->nb_jnewblk = NULL; 7796 jnewblk->jn_dep = wk; 7797 } 7798 if (!LIST_EMPTY(&newblk->nb_jwork)) 7799 jwork_move(wkhd, &newblk->nb_jwork); 7800 /* 7801 * When truncating we must free the newdirblk early to remove 7802 * the pagedep from the hash before returning. 7803 */ 7804 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7805 free_newdirblk(WK_NEWDIRBLK(wk)); 7806 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7807 panic("cancel_newblk: extra newdirblk"); 7808 7809 return (jnewblk); 7810} 7811 7812/* 7813 * Schedule the freefrag associated with a newblk to be released once 7814 * the pointers are written and the previous block is no longer needed. 7815 */ 7816static void 7817newblk_freefrag(newblk) 7818 struct newblk *newblk; 7819{ 7820 struct freefrag *freefrag; 7821 7822 if (newblk->nb_freefrag == NULL) 7823 return; 7824 freefrag = newblk->nb_freefrag; 7825 newblk->nb_freefrag = NULL; 7826 freefrag->ff_state |= COMPLETE; 7827 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 7828 add_to_worklist(&freefrag->ff_list, 0); 7829} 7830 7831/* 7832 * Free a newblk. Generate a new freefrag work request if appropriate. 7833 * This must be called after the inode pointer and any direct block pointers 7834 * are valid or fully removed via truncate or frag extension. 7835 */ 7836static void 7837free_newblk(newblk) 7838 struct newblk *newblk; 7839{ 7840 struct indirdep *indirdep; 7841 struct worklist *wk; 7842 7843 KASSERT(newblk->nb_jnewblk == NULL, 7844 ("free_newblk: jnewblk %p still attached", newblk->nb_jnewblk)); 7845 KASSERT(newblk->nb_list.wk_type != D_NEWBLK, 7846 ("free_newblk: unclaimed newblk")); 7847 LOCK_OWNED(VFSTOUFS(newblk->nb_list.wk_mp)); 7848 newblk_freefrag(newblk); 7849 if (newblk->nb_state & ONDEPLIST) 7850 LIST_REMOVE(newblk, nb_deps); 7851 if (newblk->nb_state & ONWORKLIST) 7852 WORKLIST_REMOVE(&newblk->nb_list); 7853 LIST_REMOVE(newblk, nb_hash); 7854 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7855 free_newdirblk(WK_NEWDIRBLK(wk)); 7856 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7857 panic("free_newblk: extra newdirblk"); 7858 while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL) 7859 indirdep_complete(indirdep); 7860 handle_jwork(&newblk->nb_jwork); 7861 WORKITEM_FREE(newblk, D_NEWBLK); 7862} 7863 7864/* 7865 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep. 7866 */ 7867static void 7868free_newdirblk(newdirblk) 7869 struct newdirblk *newdirblk; 7870{ 7871 struct pagedep *pagedep; 7872 struct diradd *dap; 7873 struct worklist *wk; 7874 7875 LOCK_OWNED(VFSTOUFS(newdirblk->db_list.wk_mp)); 7876 WORKLIST_REMOVE(&newdirblk->db_list); 7877 /* 7878 * If the pagedep is still linked onto the directory buffer 7879 * dependency chain, then some of the entries on the 7880 * pd_pendinghd list may not be committed to disk yet. In 7881 * this case, we will simply clear the NEWBLOCK flag and 7882 * let the pd_pendinghd list be processed when the pagedep 7883 * is next written. If the pagedep is no longer on the buffer 7884 * dependency chain, then all the entries on the pd_pending 7885 * list are committed to disk and we can free them here. 7886 */ 7887 pagedep = newdirblk->db_pagedep; 7888 pagedep->pd_state &= ~NEWBLOCK; 7889 if ((pagedep->pd_state & ONWORKLIST) == 0) { 7890 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 7891 free_diradd(dap, NULL); 7892 /* 7893 * If no dependencies remain, the pagedep will be freed. 7894 */ 7895 free_pagedep(pagedep); 7896 } 7897 /* Should only ever be one item in the list. */ 7898 while ((wk = LIST_FIRST(&newdirblk->db_mkdir)) != NULL) { 7899 WORKLIST_REMOVE(wk); 7900 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 7901 } 7902 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 7903} 7904 7905/* 7906 * Prepare an inode to be freed. The actual free operation is not 7907 * done until the zero'ed inode has been written to disk. 7908 */ 7909void 7910softdep_freefile(pvp, ino, mode) 7911 struct vnode *pvp; 7912 ino_t ino; 7913 int mode; 7914{ 7915 struct inode *ip = VTOI(pvp); 7916 struct inodedep *inodedep; 7917 struct freefile *freefile; 7918 struct freeblks *freeblks; 7919 struct ufsmount *ump; 7920 7921 ump = ITOUMP(ip); 7922 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 7923 ("softdep_freefile called on non-softdep filesystem")); 7924 /* 7925 * This sets up the inode de-allocation dependency. 7926 */ 7927 freefile = malloc(sizeof(struct freefile), 7928 M_FREEFILE, M_SOFTDEP_FLAGS); 7929 workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount); 7930 freefile->fx_mode = mode; 7931 freefile->fx_oldinum = ino; 7932 freefile->fx_devvp = ump->um_devvp; 7933 LIST_INIT(&freefile->fx_jwork); 7934 UFS_LOCK(ump); 7935 ump->um_fs->fs_pendinginodes += 1; 7936 UFS_UNLOCK(ump); 7937 7938 /* 7939 * If the inodedep does not exist, then the zero'ed inode has 7940 * been written to disk. If the allocated inode has never been 7941 * written to disk, then the on-disk inode is zero'ed. In either 7942 * case we can free the file immediately. If the journal was 7943 * canceled before being written the inode will never make it to 7944 * disk and we must send the canceled journal entrys to 7945 * ffs_freefile() to be cleared in conjunction with the bitmap. 7946 * Any blocks waiting on the inode to write can be safely freed 7947 * here as it will never been written. 7948 */ 7949 ACQUIRE_LOCK(ump); 7950 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7951 if (inodedep) { 7952 /* 7953 * Clear out freeblks that no longer need to reference 7954 * this inode. 7955 */ 7956 while ((freeblks = 7957 TAILQ_FIRST(&inodedep->id_freeblklst)) != NULL) { 7958 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, 7959 fb_next); 7960 freeblks->fb_state &= ~ONDEPLIST; 7961 } 7962 /* 7963 * Remove this inode from the unlinked list. 7964 */ 7965 if (inodedep->id_state & UNLINKED) { 7966 /* 7967 * Save the journal work to be freed with the bitmap 7968 * before we clear UNLINKED. Otherwise it can be lost 7969 * if the inode block is written. 7970 */ 7971 handle_bufwait(inodedep, &freefile->fx_jwork); 7972 clear_unlinked_inodedep(inodedep); 7973 /* 7974 * Re-acquire inodedep as we've dropped the 7975 * per-filesystem lock in clear_unlinked_inodedep(). 7976 */ 7977 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7978 } 7979 } 7980 if (inodedep == NULL || check_inode_unwritten(inodedep)) { 7981 FREE_LOCK(ump); 7982 handle_workitem_freefile(freefile); 7983 return; 7984 } 7985 if ((inodedep->id_state & DEPCOMPLETE) == 0) 7986 inodedep->id_state |= GOINGAWAY; 7987 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 7988 FREE_LOCK(ump); 7989 if (ip->i_number == ino) 7990 UFS_INODE_SET_FLAG(ip, IN_MODIFIED); 7991} 7992 7993/* 7994 * Check to see if an inode has never been written to disk. If 7995 * so free the inodedep and return success, otherwise return failure. 7996 * 7997 * If we still have a bitmap dependency, then the inode has never 7998 * been written to disk. Drop the dependency as it is no longer 7999 * necessary since the inode is being deallocated. We set the 8000 * ALLCOMPLETE flags since the bitmap now properly shows that the 8001 * inode is not allocated. Even if the inode is actively being 8002 * written, it has been rolled back to its zero'ed state, so we 8003 * are ensured that a zero inode is what is on the disk. For short 8004 * lived files, this change will usually result in removing all the 8005 * dependencies from the inode so that it can be freed immediately. 8006 */ 8007static int 8008check_inode_unwritten(inodedep) 8009 struct inodedep *inodedep; 8010{ 8011 8012 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 8013 8014 if ((inodedep->id_state & (DEPCOMPLETE | UNLINKED)) != 0 || 8015 !LIST_EMPTY(&inodedep->id_dirremhd) || 8016 !LIST_EMPTY(&inodedep->id_pendinghd) || 8017 !LIST_EMPTY(&inodedep->id_bufwait) || 8018 !LIST_EMPTY(&inodedep->id_inowait) || 8019 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 8020 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 8021 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 8022 !TAILQ_EMPTY(&inodedep->id_extupdt) || 8023 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 8024 !TAILQ_EMPTY(&inodedep->id_freeblklst) || 8025 inodedep->id_mkdiradd != NULL || 8026 inodedep->id_nlinkdelta != 0) 8027 return (0); 8028 /* 8029 * Another process might be in initiate_write_inodeblock_ufs[12] 8030 * trying to allocate memory without holding "Softdep Lock". 8031 */ 8032 if ((inodedep->id_state & IOSTARTED) != 0 && 8033 inodedep->id_savedino1 == NULL) 8034 return (0); 8035 8036 if (inodedep->id_state & ONDEPLIST) 8037 LIST_REMOVE(inodedep, id_deps); 8038 inodedep->id_state &= ~ONDEPLIST; 8039 inodedep->id_state |= ALLCOMPLETE; 8040 inodedep->id_bmsafemap = NULL; 8041 if (inodedep->id_state & ONWORKLIST) 8042 WORKLIST_REMOVE(&inodedep->id_list); 8043 if (inodedep->id_savedino1 != NULL) { 8044 free(inodedep->id_savedino1, M_SAVEDINO); 8045 inodedep->id_savedino1 = NULL; 8046 } 8047 if (free_inodedep(inodedep) == 0) 8048 panic("check_inode_unwritten: busy inode"); 8049 return (1); 8050} 8051 8052static int 8053check_inodedep_free(inodedep) 8054 struct inodedep *inodedep; 8055{ 8056 8057 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 8058 if ((inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 8059 !LIST_EMPTY(&inodedep->id_dirremhd) || 8060 !LIST_EMPTY(&inodedep->id_pendinghd) || 8061 !LIST_EMPTY(&inodedep->id_bufwait) || 8062 !LIST_EMPTY(&inodedep->id_inowait) || 8063 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 8064 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 8065 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 8066 !TAILQ_EMPTY(&inodedep->id_extupdt) || 8067 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 8068 !TAILQ_EMPTY(&inodedep->id_freeblklst) || 8069 inodedep->id_mkdiradd != NULL || 8070 inodedep->id_nlinkdelta != 0 || 8071 inodedep->id_savedino1 != NULL) 8072 return (0); 8073 return (1); 8074} 8075 8076/* 8077 * Try to free an inodedep structure. Return 1 if it could be freed. 8078 */ 8079static int 8080free_inodedep(inodedep) 8081 struct inodedep *inodedep; 8082{ 8083 8084 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 8085 if ((inodedep->id_state & (ONWORKLIST | UNLINKED)) != 0 || 8086 !check_inodedep_free(inodedep)) 8087 return (0); 8088 if (inodedep->id_state & ONDEPLIST) 8089 LIST_REMOVE(inodedep, id_deps); 8090 LIST_REMOVE(inodedep, id_hash); 8091 WORKITEM_FREE(inodedep, D_INODEDEP); 8092 return (1); 8093} 8094 8095/* 8096 * Free the block referenced by a freework structure. The parent freeblks 8097 * structure is released and completed when the final cg bitmap reaches 8098 * the disk. This routine may be freeing a jnewblk which never made it to 8099 * disk in which case we do not have to wait as the operation is undone 8100 * in memory immediately. 8101 */ 8102static void 8103freework_freeblock(freework, key) 8104 struct freework *freework; 8105 u_long key; 8106{ 8107 struct freeblks *freeblks; 8108 struct jnewblk *jnewblk; 8109 struct ufsmount *ump; 8110 struct workhead wkhd; 8111 struct fs *fs; 8112 int bsize; 8113 int needj; 8114 8115 ump = VFSTOUFS(freework->fw_list.wk_mp); 8116 LOCK_OWNED(ump); 8117 /* 8118 * Handle partial truncate separately. 8119 */ 8120 if (freework->fw_indir) { 8121 complete_trunc_indir(freework); 8122 return; 8123 } 8124 freeblks = freework->fw_freeblks; 8125 fs = ump->um_fs; 8126 needj = MOUNTEDSUJ(freeblks->fb_list.wk_mp) != 0; 8127 bsize = lfragtosize(fs, freework->fw_frags); 8128 LIST_INIT(&wkhd); 8129 /* 8130 * DEPCOMPLETE is cleared in indirblk_insert() if the block lives 8131 * on the indirblk hashtable and prevents premature freeing. 8132 */ 8133 freework->fw_state |= DEPCOMPLETE; 8134 /* 8135 * SUJ needs to wait for the segment referencing freed indirect 8136 * blocks to expire so that we know the checker will not confuse 8137 * a re-allocated indirect block with its old contents. 8138 */ 8139 if (needj && freework->fw_lbn <= -UFS_NDADDR) 8140 indirblk_insert(freework); 8141 /* 8142 * If we are canceling an existing jnewblk pass it to the free 8143 * routine, otherwise pass the freeblk which will ultimately 8144 * release the freeblks. If we're not journaling, we can just 8145 * free the freeblks immediately. 8146 */ 8147 jnewblk = freework->fw_jnewblk; 8148 if (jnewblk != NULL) { 8149 cancel_jnewblk(jnewblk, &wkhd); 8150 needj = 0; 8151 } else if (needj) { 8152 freework->fw_state |= DELAYEDFREE; 8153 freeblks->fb_cgwait++; 8154 WORKLIST_INSERT(&wkhd, &freework->fw_list); 8155 } 8156 FREE_LOCK(ump); 8157 freeblks_free(ump, freeblks, btodb(bsize)); 8158 CTR4(KTR_SUJ, 8159 "freework_freeblock: ino %jd blkno %jd lbn %jd size %d", 8160 freeblks->fb_inum, freework->fw_blkno, freework->fw_lbn, bsize); 8161 ffs_blkfree(ump, fs, freeblks->fb_devvp, freework->fw_blkno, bsize, 8162 freeblks->fb_inum, freeblks->fb_vtype, &wkhd, key); 8163 ACQUIRE_LOCK(ump); 8164 /* 8165 * The jnewblk will be discarded and the bits in the map never 8166 * made it to disk. We can immediately free the freeblk. 8167 */ 8168 if (needj == 0) 8169 handle_written_freework(freework); 8170} 8171 8172/* 8173 * We enqueue freework items that need processing back on the freeblks and 8174 * add the freeblks to the worklist. This makes it easier to find all work 8175 * required to flush a truncation in process_truncates(). 8176 */ 8177static void 8178freework_enqueue(freework) 8179 struct freework *freework; 8180{ 8181 struct freeblks *freeblks; 8182 8183 freeblks = freework->fw_freeblks; 8184 if ((freework->fw_state & INPROGRESS) == 0) 8185 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 8186 if ((freeblks->fb_state & 8187 (ONWORKLIST | INPROGRESS | ALLCOMPLETE)) == ALLCOMPLETE && 8188 LIST_EMPTY(&freeblks->fb_jblkdephd)) 8189 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 8190} 8191 8192/* 8193 * Start, continue, or finish the process of freeing an indirect block tree. 8194 * The free operation may be paused at any point with fw_off containing the 8195 * offset to restart from. This enables us to implement some flow control 8196 * for large truncates which may fan out and generate a huge number of 8197 * dependencies. 8198 */ 8199static void 8200handle_workitem_indirblk(freework) 8201 struct freework *freework; 8202{ 8203 struct freeblks *freeblks; 8204 struct ufsmount *ump; 8205 struct fs *fs; 8206 8207 freeblks = freework->fw_freeblks; 8208 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 8209 fs = ump->um_fs; 8210 if (freework->fw_state & DEPCOMPLETE) { 8211 handle_written_freework(freework); 8212 return; 8213 } 8214 if (freework->fw_off == NINDIR(fs)) { 8215 freework_freeblock(freework, SINGLETON_KEY); 8216 return; 8217 } 8218 freework->fw_state |= INPROGRESS; 8219 FREE_LOCK(ump); 8220 indir_trunc(freework, fsbtodb(fs, freework->fw_blkno), 8221 freework->fw_lbn); 8222 ACQUIRE_LOCK(ump); 8223} 8224 8225/* 8226 * Called when a freework structure attached to a cg buf is written. The 8227 * ref on either the parent or the freeblks structure is released and 8228 * the freeblks is added back to the worklist if there is more work to do. 8229 */ 8230static void 8231handle_written_freework(freework) 8232 struct freework *freework; 8233{ 8234 struct freeblks *freeblks; 8235 struct freework *parent; 8236 8237 freeblks = freework->fw_freeblks; 8238 parent = freework->fw_parent; 8239 if (freework->fw_state & DELAYEDFREE) 8240 freeblks->fb_cgwait--; 8241 freework->fw_state |= COMPLETE; 8242 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 8243 WORKITEM_FREE(freework, D_FREEWORK); 8244 if (parent) { 8245 if (--parent->fw_ref == 0) 8246 freework_enqueue(parent); 8247 return; 8248 } 8249 if (--freeblks->fb_ref != 0) 8250 return; 8251 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST | INPROGRESS)) == 8252 ALLCOMPLETE && LIST_EMPTY(&freeblks->fb_jblkdephd)) 8253 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 8254} 8255 8256/* 8257 * This workitem routine performs the block de-allocation. 8258 * The workitem is added to the pending list after the updated 8259 * inode block has been written to disk. As mentioned above, 8260 * checks regarding the number of blocks de-allocated (compared 8261 * to the number of blocks allocated for the file) are also 8262 * performed in this function. 8263 */ 8264static int 8265handle_workitem_freeblocks(freeblks, flags) 8266 struct freeblks *freeblks; 8267 int flags; 8268{ 8269 struct freework *freework; 8270 struct newblk *newblk; 8271 struct allocindir *aip; 8272 struct ufsmount *ump; 8273 struct worklist *wk; 8274 u_long key; 8275 8276 KASSERT(LIST_EMPTY(&freeblks->fb_jblkdephd), 8277 ("handle_workitem_freeblocks: Journal entries not written.")); 8278 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 8279 key = ffs_blkrelease_start(ump, freeblks->fb_devvp, freeblks->fb_inum); 8280 ACQUIRE_LOCK(ump); 8281 while ((wk = LIST_FIRST(&freeblks->fb_freeworkhd)) != NULL) { 8282 WORKLIST_REMOVE(wk); 8283 switch (wk->wk_type) { 8284 case D_DIRREM: 8285 wk->wk_state |= COMPLETE; 8286 add_to_worklist(wk, 0); 8287 continue; 8288 8289 case D_ALLOCDIRECT: 8290 free_newblk(WK_NEWBLK(wk)); 8291 continue; 8292 8293 case D_ALLOCINDIR: 8294 aip = WK_ALLOCINDIR(wk); 8295 freework = NULL; 8296 if (aip->ai_state & DELAYEDFREE) { 8297 FREE_LOCK(ump); 8298 freework = newfreework(ump, freeblks, NULL, 8299 aip->ai_lbn, aip->ai_newblkno, 8300 ump->um_fs->fs_frag, 0, 0); 8301 ACQUIRE_LOCK(ump); 8302 } 8303 newblk = WK_NEWBLK(wk); 8304 if (newblk->nb_jnewblk) { 8305 freework->fw_jnewblk = newblk->nb_jnewblk; 8306 newblk->nb_jnewblk->jn_dep = &freework->fw_list; 8307 newblk->nb_jnewblk = NULL; 8308 } 8309 free_newblk(newblk); 8310 continue; 8311 8312 case D_FREEWORK: 8313 freework = WK_FREEWORK(wk); 8314 if (freework->fw_lbn <= -UFS_NDADDR) 8315 handle_workitem_indirblk(freework); 8316 else 8317 freework_freeblock(freework, key); 8318 continue; 8319 default: 8320 panic("handle_workitem_freeblocks: Unknown type %s", 8321 TYPENAME(wk->wk_type)); 8322 } 8323 } 8324 if (freeblks->fb_ref != 0) { 8325 freeblks->fb_state &= ~INPROGRESS; 8326 wake_worklist(&freeblks->fb_list); 8327 freeblks = NULL; 8328 } 8329 FREE_LOCK(ump); 8330 ffs_blkrelease_finish(ump, key); 8331 if (freeblks) 8332 return handle_complete_freeblocks(freeblks, flags); 8333 return (0); 8334} 8335 8336/* 8337 * Handle completion of block free via truncate. This allows fs_pending 8338 * to track the actual free block count more closely than if we only updated 8339 * it at the end. We must be careful to handle cases where the block count 8340 * on free was incorrect. 8341 */ 8342static void 8343freeblks_free(ump, freeblks, blocks) 8344 struct ufsmount *ump; 8345 struct freeblks *freeblks; 8346 int blocks; 8347{ 8348 struct fs *fs; 8349 ufs2_daddr_t remain; 8350 8351 UFS_LOCK(ump); 8352 remain = -freeblks->fb_chkcnt; 8353 freeblks->fb_chkcnt += blocks; 8354 if (remain > 0) { 8355 if (remain < blocks) 8356 blocks = remain; 8357 fs = ump->um_fs; 8358 fs->fs_pendingblocks -= blocks; 8359 } 8360 UFS_UNLOCK(ump); 8361} 8362 8363/* 8364 * Once all of the freework workitems are complete we can retire the 8365 * freeblocks dependency and any journal work awaiting completion. This 8366 * can not be called until all other dependencies are stable on disk. 8367 */ 8368static int 8369handle_complete_freeblocks(freeblks, flags) 8370 struct freeblks *freeblks; 8371 int flags; 8372{ 8373 struct inodedep *inodedep; 8374 struct inode *ip; 8375 struct vnode *vp; 8376 struct fs *fs; 8377 struct ufsmount *ump; 8378 ufs2_daddr_t spare; 8379 8380 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 8381 fs = ump->um_fs; 8382 flags = LK_EXCLUSIVE | flags; 8383 spare = freeblks->fb_chkcnt; 8384 8385 /* 8386 * If we did not release the expected number of blocks we may have 8387 * to adjust the inode block count here. Only do so if it wasn't 8388 * a truncation to zero and the modrev still matches. 8389 */ 8390 if (spare && freeblks->fb_len != 0) { 8391 if (ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_inum, 8392 flags, &vp, FFSV_FORCEINSMQ) != 0) 8393 return (EBUSY); 8394 ip = VTOI(vp); 8395 if (ip->i_mode == 0) { 8396 vgone(vp); 8397 } else if (DIP(ip, i_modrev) == freeblks->fb_modrev) { 8398 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - spare); 8399 UFS_INODE_SET_FLAG(ip, IN_CHANGE); 8400 /* 8401 * We must wait so this happens before the 8402 * journal is reclaimed. 8403 */ 8404 ffs_update(vp, 1); 8405 } 8406 vput(vp); 8407 } 8408 if (spare < 0) { 8409 UFS_LOCK(ump); 8410 fs->fs_pendingblocks += spare; 8411 UFS_UNLOCK(ump); 8412 } 8413#ifdef QUOTA 8414 /* Handle spare. */ 8415 if (spare) 8416 quotaadj(freeblks->fb_quota, ump, -spare); 8417 quotarele(freeblks->fb_quota); 8418#endif 8419 ACQUIRE_LOCK(ump); 8420 if (freeblks->fb_state & ONDEPLIST) { 8421 inodedep_lookup(freeblks->fb_list.wk_mp, freeblks->fb_inum, 8422 0, &inodedep); 8423 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, fb_next); 8424 freeblks->fb_state &= ~ONDEPLIST; 8425 if (TAILQ_EMPTY(&inodedep->id_freeblklst)) 8426 free_inodedep(inodedep); 8427 } 8428 /* 8429 * All of the freeblock deps must be complete prior to this call 8430 * so it's now safe to complete earlier outstanding journal entries. 8431 */ 8432 handle_jwork(&freeblks->fb_jwork); 8433 WORKITEM_FREE(freeblks, D_FREEBLKS); 8434 FREE_LOCK(ump); 8435 return (0); 8436} 8437 8438/* 8439 * Release blocks associated with the freeblks and stored in the indirect 8440 * block dbn. If level is greater than SINGLE, the block is an indirect block 8441 * and recursive calls to indirtrunc must be used to cleanse other indirect 8442 * blocks. 8443 * 8444 * This handles partial and complete truncation of blocks. Partial is noted 8445 * with goingaway == 0. In this case the freework is completed after the 8446 * zero'd indirects are written to disk. For full truncation the freework 8447 * is completed after the block is freed. 8448 */ 8449static void 8450indir_trunc(freework, dbn, lbn) 8451 struct freework *freework; 8452 ufs2_daddr_t dbn; 8453 ufs_lbn_t lbn; 8454{ 8455 struct freework *nfreework; 8456 struct workhead wkhd; 8457 struct freeblks *freeblks; 8458 struct buf *bp; 8459 struct fs *fs; 8460 struct indirdep *indirdep; 8461 struct mount *mp; 8462 struct ufsmount *ump; 8463 ufs1_daddr_t *bap1; 8464 ufs2_daddr_t nb, nnb, *bap2; 8465 ufs_lbn_t lbnadd, nlbn; 8466 u_long key; 8467 int nblocks, ufs1fmt, freedblocks; 8468 int goingaway, freedeps, needj, level, cnt, i, error; 8469 8470 freeblks = freework->fw_freeblks; 8471 mp = freeblks->fb_list.wk_mp; 8472 ump = VFSTOUFS(mp); 8473 fs = ump->um_fs; 8474 /* 8475 * Get buffer of block pointers to be freed. There are three cases: 8476 * 8477 * 1) Partial truncate caches the indirdep pointer in the freework 8478 * which provides us a back copy to the save bp which holds the 8479 * pointers we want to clear. When this completes the zero 8480 * pointers are written to the real copy. 8481 * 2) The indirect is being completely truncated, cancel_indirdep() 8482 * eliminated the real copy and placed the indirdep on the saved 8483 * copy. The indirdep and buf are discarded when this completes. 8484 * 3) The indirect was not in memory, we read a copy off of the disk 8485 * using the devvp and drop and invalidate the buffer when we're 8486 * done. 8487 */ 8488 goingaway = 1; 8489 indirdep = NULL; 8490 if (freework->fw_indir != NULL) { 8491 goingaway = 0; 8492 indirdep = freework->fw_indir; 8493 bp = indirdep->ir_savebp; 8494 if (bp == NULL || bp->b_blkno != dbn) 8495 panic("indir_trunc: Bad saved buf %p blkno %jd", 8496 bp, (intmax_t)dbn); 8497 } else if ((bp = incore(&freeblks->fb_devvp->v_bufobj, dbn)) != NULL) { 8498 /* 8499 * The lock prevents the buf dep list from changing and 8500 * indirects on devvp should only ever have one dependency. 8501 */ 8502 indirdep = WK_INDIRDEP(LIST_FIRST(&bp->b_dep)); 8503 if (indirdep == NULL || (indirdep->ir_state & GOINGAWAY) == 0) 8504 panic("indir_trunc: Bad indirdep %p from buf %p", 8505 indirdep, bp); 8506 } else { 8507 error = ffs_breadz(ump, freeblks->fb_devvp, dbn, dbn, 8508 (int)fs->fs_bsize, NULL, NULL, 0, NOCRED, 0, NULL, &bp); 8509 if (error) 8510 return; 8511 } 8512 ACQUIRE_LOCK(ump); 8513 /* Protects against a race with complete_trunc_indir(). */ 8514 freework->fw_state &= ~INPROGRESS; 8515 /* 8516 * If we have an indirdep we need to enforce the truncation order 8517 * and discard it when it is complete. 8518 */ 8519 if (indirdep) { 8520 if (freework != TAILQ_FIRST(&indirdep->ir_trunc) && 8521 !TAILQ_EMPTY(&indirdep->ir_trunc)) { 8522 /* 8523 * Add the complete truncate to the list on the 8524 * indirdep to enforce in-order processing. 8525 */ 8526 if (freework->fw_indir == NULL) 8527 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, 8528 freework, fw_next); 8529 FREE_LOCK(ump); 8530 return; 8531 } 8532 /* 8533 * If we're goingaway, free the indirdep. Otherwise it will 8534 * linger until the write completes. 8535 */ 8536 if (goingaway) { 8537 KASSERT(indirdep->ir_savebp == bp, 8538 ("indir_trunc: losing ir_savebp %p", 8539 indirdep->ir_savebp)); 8540 indirdep->ir_savebp = NULL; 8541 free_indirdep(indirdep); 8542 } 8543 } 8544 FREE_LOCK(ump); 8545 /* Initialize pointers depending on block size. */ 8546 if (ump->um_fstype == UFS1) { 8547 bap1 = (ufs1_daddr_t *)bp->b_data; 8548 nb = bap1[freework->fw_off]; 8549 ufs1fmt = 1; 8550 bap2 = NULL; 8551 } else { 8552 bap2 = (ufs2_daddr_t *)bp->b_data; 8553 nb = bap2[freework->fw_off]; 8554 ufs1fmt = 0; 8555 bap1 = NULL; 8556 } 8557 level = lbn_level(lbn); 8558 needj = MOUNTEDSUJ(UFSTOVFS(ump)) != 0; 8559 lbnadd = lbn_offset(fs, level); 8560 nblocks = btodb(fs->fs_bsize); 8561 nfreework = freework; 8562 freedeps = 0; 8563 cnt = 0; 8564 /* 8565 * Reclaim blocks. Traverses into nested indirect levels and 8566 * arranges for the current level to be freed when subordinates 8567 * are free when journaling. 8568 */ 8569 key = ffs_blkrelease_start(ump, freeblks->fb_devvp, freeblks->fb_inum); 8570 for (i = freework->fw_off; i < NINDIR(fs); i++, nb = nnb) { 8571 if (UFS_CHECK_BLKNO(mp, freeblks->fb_inum, nb, 8572 fs->fs_bsize) != 0) 8573 nb = 0; 8574 if (i != NINDIR(fs) - 1) { 8575 if (ufs1fmt) 8576 nnb = bap1[i+1]; 8577 else 8578 nnb = bap2[i+1]; 8579 } else 8580 nnb = 0; 8581 if (nb == 0) 8582 continue; 8583 cnt++; 8584 if (level != 0) { 8585 nlbn = (lbn + 1) - (i * lbnadd); 8586 if (needj != 0) { 8587 nfreework = newfreework(ump, freeblks, freework, 8588 nlbn, nb, fs->fs_frag, 0, 0); 8589 freedeps++; 8590 } 8591 indir_trunc(nfreework, fsbtodb(fs, nb), nlbn); 8592 } else { 8593 struct freedep *freedep; 8594 8595 /* 8596 * Attempt to aggregate freedep dependencies for 8597 * all blocks being released to the same CG. 8598 */ 8599 LIST_INIT(&wkhd); 8600 if (needj != 0 && 8601 (nnb == 0 || (dtog(fs, nb) != dtog(fs, nnb)))) { 8602 freedep = newfreedep(freework); 8603 WORKLIST_INSERT_UNLOCKED(&wkhd, 8604 &freedep->fd_list); 8605 freedeps++; 8606 } 8607 CTR3(KTR_SUJ, 8608 "indir_trunc: ino %jd blkno %jd size %d", 8609 freeblks->fb_inum, nb, fs->fs_bsize); 8610 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, 8611 fs->fs_bsize, freeblks->fb_inum, 8612 freeblks->fb_vtype, &wkhd, key); 8613 } 8614 } 8615 ffs_blkrelease_finish(ump, key); 8616 if (goingaway) { 8617 bp->b_flags |= B_INVAL | B_NOCACHE; 8618 brelse(bp); 8619 } 8620 freedblocks = 0; 8621 if (level == 0) 8622 freedblocks = (nblocks * cnt); 8623 if (needj == 0) 8624 freedblocks += nblocks; 8625 freeblks_free(ump, freeblks, freedblocks); 8626 /* 8627 * If we are journaling set up the ref counts and offset so this 8628 * indirect can be completed when its children are free. 8629 */ 8630 if (needj) { 8631 ACQUIRE_LOCK(ump); 8632 freework->fw_off = i; 8633 freework->fw_ref += freedeps; 8634 freework->fw_ref -= NINDIR(fs) + 1; 8635 if (level == 0) 8636 freeblks->fb_cgwait += freedeps; 8637 if (freework->fw_ref == 0) 8638 freework_freeblock(freework, SINGLETON_KEY); 8639 FREE_LOCK(ump); 8640 return; 8641 } 8642 /* 8643 * If we're not journaling we can free the indirect now. 8644 */ 8645 dbn = dbtofsb(fs, dbn); 8646 CTR3(KTR_SUJ, 8647 "indir_trunc 2: ino %jd blkno %jd size %d", 8648 freeblks->fb_inum, dbn, fs->fs_bsize); 8649 ffs_blkfree(ump, fs, freeblks->fb_devvp, dbn, fs->fs_bsize, 8650 freeblks->fb_inum, freeblks->fb_vtype, NULL, SINGLETON_KEY); 8651 /* Non SUJ softdep does single-threaded truncations. */ 8652 if (freework->fw_blkno == dbn) { 8653 freework->fw_state |= ALLCOMPLETE; 8654 ACQUIRE_LOCK(ump); 8655 handle_written_freework(freework); 8656 FREE_LOCK(ump); 8657 } 8658 return; 8659} 8660 8661/* 8662 * Cancel an allocindir when it is removed via truncation. When bp is not 8663 * NULL the indirect never appeared on disk and is scheduled to be freed 8664 * independently of the indir so we can more easily track journal work. 8665 */ 8666static void 8667cancel_allocindir(aip, bp, freeblks, trunc) 8668 struct allocindir *aip; 8669 struct buf *bp; 8670 struct freeblks *freeblks; 8671 int trunc; 8672{ 8673 struct indirdep *indirdep; 8674 struct freefrag *freefrag; 8675 struct newblk *newblk; 8676 8677 newblk = (struct newblk *)aip; 8678 LIST_REMOVE(aip, ai_next); 8679 /* 8680 * We must eliminate the pointer in bp if it must be freed on its 8681 * own due to partial truncate or pending journal work. 8682 */ 8683 if (bp && (trunc || newblk->nb_jnewblk)) { 8684 /* 8685 * Clear the pointer and mark the aip to be freed 8686 * directly if it never existed on disk. 8687 */ 8688 aip->ai_state |= DELAYEDFREE; 8689 indirdep = aip->ai_indirdep; 8690 if (indirdep->ir_state & UFS1FMT) 8691 ((ufs1_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 8692 else 8693 ((ufs2_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 8694 } 8695 /* 8696 * When truncating the previous pointer will be freed via 8697 * savedbp. Eliminate the freefrag which would dup free. 8698 */ 8699 if (trunc && (freefrag = newblk->nb_freefrag) != NULL) { 8700 newblk->nb_freefrag = NULL; 8701 if (freefrag->ff_jdep) 8702 cancel_jfreefrag( 8703 WK_JFREEFRAG(freefrag->ff_jdep)); 8704 jwork_move(&freeblks->fb_jwork, &freefrag->ff_jwork); 8705 WORKITEM_FREE(freefrag, D_FREEFRAG); 8706 } 8707 /* 8708 * If the journal hasn't been written the jnewblk must be passed 8709 * to the call to ffs_blkfree that reclaims the space. We accomplish 8710 * this by leaving the journal dependency on the newblk to be freed 8711 * when a freework is created in handle_workitem_freeblocks(). 8712 */ 8713 cancel_newblk(newblk, NULL, &freeblks->fb_jwork); 8714 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 8715} 8716 8717/* 8718 * Create the mkdir dependencies for . and .. in a new directory. Link them 8719 * in to a newdirblk so any subsequent additions are tracked properly. The 8720 * caller is responsible for adding the mkdir1 dependency to the journal 8721 * and updating id_mkdiradd. This function returns with the per-filesystem 8722 * lock held. 8723 */ 8724static struct mkdir * 8725setup_newdir(dap, newinum, dinum, newdirbp, mkdirp) 8726 struct diradd *dap; 8727 ino_t newinum; 8728 ino_t dinum; 8729 struct buf *newdirbp; 8730 struct mkdir **mkdirp; 8731{ 8732 struct newblk *newblk; 8733 struct pagedep *pagedep; 8734 struct inodedep *inodedep; 8735 struct newdirblk *newdirblk; 8736 struct mkdir *mkdir1, *mkdir2; 8737 struct worklist *wk; 8738 struct jaddref *jaddref; 8739 struct ufsmount *ump; 8740 struct mount *mp; 8741 8742 mp = dap->da_list.wk_mp; 8743 ump = VFSTOUFS(mp); 8744 newdirblk = malloc(sizeof(struct newdirblk), M_NEWDIRBLK, 8745 M_SOFTDEP_FLAGS); 8746 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 8747 LIST_INIT(&newdirblk->db_mkdir); 8748 mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 8749 workitem_alloc(&mkdir1->md_list, D_MKDIR, mp); 8750 mkdir1->md_state = ATTACHED | MKDIR_BODY; 8751 mkdir1->md_diradd = dap; 8752 mkdir1->md_jaddref = NULL; 8753 mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 8754 workitem_alloc(&mkdir2->md_list, D_MKDIR, mp); 8755 mkdir2->md_state = ATTACHED | MKDIR_PARENT; 8756 mkdir2->md_diradd = dap; 8757 mkdir2->md_jaddref = NULL; 8758 if (MOUNTEDSUJ(mp) == 0) { 8759 mkdir1->md_state |= DEPCOMPLETE; 8760 mkdir2->md_state |= DEPCOMPLETE; 8761 } 8762 /* 8763 * Dependency on "." and ".." being written to disk. 8764 */ 8765 mkdir1->md_buf = newdirbp; 8766 ACQUIRE_LOCK(VFSTOUFS(mp)); 8767 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir1, md_mkdirs); 8768 /* 8769 * We must link the pagedep, allocdirect, and newdirblk for 8770 * the initial file page so the pointer to the new directory 8771 * is not written until the directory contents are live and 8772 * any subsequent additions are not marked live until the 8773 * block is reachable via the inode. 8774 */ 8775 if (pagedep_lookup(mp, newdirbp, newinum, 0, 0, &pagedep) == 0) 8776 panic("setup_newdir: lost pagedep"); 8777 LIST_FOREACH(wk, &newdirbp->b_dep, wk_list) 8778 if (wk->wk_type == D_ALLOCDIRECT) 8779 break; 8780 if (wk == NULL) 8781 panic("setup_newdir: lost allocdirect"); 8782 if (pagedep->pd_state & NEWBLOCK) 8783 panic("setup_newdir: NEWBLOCK already set"); 8784 newblk = WK_NEWBLK(wk); 8785 pagedep->pd_state |= NEWBLOCK; 8786 pagedep->pd_newdirblk = newdirblk; 8787 newdirblk->db_pagedep = pagedep; 8788 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 8789 WORKLIST_INSERT(&newdirblk->db_mkdir, &mkdir1->md_list); 8790 /* 8791 * Look up the inodedep for the parent directory so that we 8792 * can link mkdir2 into the pending dotdot jaddref or 8793 * the inode write if there is none. If the inode is 8794 * ALLCOMPLETE and no jaddref is present all dependencies have 8795 * been satisfied and mkdir2 can be freed. 8796 */ 8797 inodedep_lookup(mp, dinum, 0, &inodedep); 8798 if (MOUNTEDSUJ(mp)) { 8799 if (inodedep == NULL) 8800 panic("setup_newdir: Lost parent."); 8801 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8802 inoreflst); 8803 KASSERT(jaddref != NULL && jaddref->ja_parent == newinum && 8804 (jaddref->ja_state & MKDIR_PARENT), 8805 ("setup_newdir: bad dotdot jaddref %p", jaddref)); 8806 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs); 8807 mkdir2->md_jaddref = jaddref; 8808 jaddref->ja_mkdir = mkdir2; 8809 } else if (inodedep == NULL || 8810 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 8811 dap->da_state &= ~MKDIR_PARENT; 8812 WORKITEM_FREE(mkdir2, D_MKDIR); 8813 mkdir2 = NULL; 8814 } else { 8815 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs); 8816 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir2->md_list); 8817 } 8818 *mkdirp = mkdir2; 8819 8820 return (mkdir1); 8821} 8822 8823/* 8824 * Directory entry addition dependencies. 8825 * 8826 * When adding a new directory entry, the inode (with its incremented link 8827 * count) must be written to disk before the directory entry's pointer to it. 8828 * Also, if the inode is newly allocated, the corresponding freemap must be 8829 * updated (on disk) before the directory entry's pointer. These requirements 8830 * are met via undo/redo on the directory entry's pointer, which consists 8831 * simply of the inode number. 8832 * 8833 * As directory entries are added and deleted, the free space within a 8834 * directory block can become fragmented. The ufs filesystem will compact 8835 * a fragmented directory block to make space for a new entry. When this 8836 * occurs, the offsets of previously added entries change. Any "diradd" 8837 * dependency structures corresponding to these entries must be updated with 8838 * the new offsets. 8839 */ 8840 8841/* 8842 * This routine is called after the in-memory inode's link 8843 * count has been incremented, but before the directory entry's 8844 * pointer to the inode has been set. 8845 */ 8846int 8847softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 8848 struct buf *bp; /* buffer containing directory block */ 8849 struct inode *dp; /* inode for directory */ 8850 off_t diroffset; /* offset of new entry in directory */ 8851 ino_t newinum; /* inode referenced by new directory entry */ 8852 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 8853 int isnewblk; /* entry is in a newly allocated block */ 8854{ 8855 int offset; /* offset of new entry within directory block */ 8856 ufs_lbn_t lbn; /* block in directory containing new entry */ 8857 struct fs *fs; 8858 struct diradd *dap; 8859 struct newblk *newblk; 8860 struct pagedep *pagedep; 8861 struct inodedep *inodedep; 8862 struct newdirblk *newdirblk; 8863 struct mkdir *mkdir1, *mkdir2; 8864 struct jaddref *jaddref; 8865 struct ufsmount *ump; 8866 struct mount *mp; 8867 int isindir; 8868 8869 mp = ITOVFS(dp); 8870 ump = VFSTOUFS(mp); 8871 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 8872 ("softdep_setup_directory_add called on non-softdep filesystem")); 8873 /* 8874 * Whiteouts have no dependencies. 8875 */ 8876 if (newinum == UFS_WINO) { 8877 if (newdirbp != NULL) 8878 bdwrite(newdirbp); 8879 return (0); 8880 } 8881 jaddref = NULL; 8882 mkdir1 = mkdir2 = NULL; 8883 fs = ump->um_fs; 8884 lbn = lblkno(fs, diroffset); 8885 offset = blkoff(fs, diroffset); 8886 dap = malloc(sizeof(struct diradd), M_DIRADD, 8887 M_SOFTDEP_FLAGS|M_ZERO); 8888 workitem_alloc(&dap->da_list, D_DIRADD, mp); 8889 dap->da_offset = offset; 8890 dap->da_newinum = newinum; 8891 dap->da_state = ATTACHED; 8892 LIST_INIT(&dap->da_jwork); 8893 isindir = bp->b_lblkno >= UFS_NDADDR; 8894 newdirblk = NULL; 8895 if (isnewblk && 8896 (isindir ? blkoff(fs, diroffset) : fragoff(fs, diroffset)) == 0) { 8897 newdirblk = malloc(sizeof(struct newdirblk), 8898 M_NEWDIRBLK, M_SOFTDEP_FLAGS); 8899 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 8900 LIST_INIT(&newdirblk->db_mkdir); 8901 } 8902 /* 8903 * If we're creating a new directory setup the dependencies and set 8904 * the dap state to wait for them. Otherwise it's COMPLETE and 8905 * we can move on. 8906 */ 8907 if (newdirbp == NULL) { 8908 dap->da_state |= DEPCOMPLETE; 8909 ACQUIRE_LOCK(ump); 8910 } else { 8911 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 8912 mkdir1 = setup_newdir(dap, newinum, dp->i_number, newdirbp, 8913 &mkdir2); 8914 } 8915 /* 8916 * Link into parent directory pagedep to await its being written. 8917 */ 8918 pagedep_lookup(mp, bp, dp->i_number, lbn, DEPALLOC, &pagedep); 8919#ifdef INVARIANTS 8920 if (diradd_lookup(pagedep, offset) != NULL) 8921 panic("softdep_setup_directory_add: %p already at off %d\n", 8922 diradd_lookup(pagedep, offset), offset); 8923#endif 8924 dap->da_pagedep = pagedep; 8925 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 8926 da_pdlist); 8927 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); 8928 /* 8929 * If we're journaling, link the diradd into the jaddref so it 8930 * may be completed after the journal entry is written. Otherwise, 8931 * link the diradd into its inodedep. If the inode is not yet 8932 * written place it on the bufwait list, otherwise do the post-inode 8933 * write processing to put it on the id_pendinghd list. 8934 */ 8935 if (MOUNTEDSUJ(mp)) { 8936 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8937 inoreflst); 8938 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 8939 ("softdep_setup_directory_add: bad jaddref %p", jaddref)); 8940 jaddref->ja_diroff = diroffset; 8941 jaddref->ja_diradd = dap; 8942 add_to_journal(&jaddref->ja_list); 8943 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 8944 diradd_inode_written(dap, inodedep); 8945 else 8946 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 8947 /* 8948 * Add the journal entries for . and .. links now that the primary 8949 * link is written. 8950 */ 8951 if (mkdir1 != NULL && MOUNTEDSUJ(mp)) { 8952 jaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 8953 inoreflst, if_deps); 8954 KASSERT(jaddref != NULL && 8955 jaddref->ja_ino == jaddref->ja_parent && 8956 (jaddref->ja_state & MKDIR_BODY), 8957 ("softdep_setup_directory_add: bad dot jaddref %p", 8958 jaddref)); 8959 mkdir1->md_jaddref = jaddref; 8960 jaddref->ja_mkdir = mkdir1; 8961 /* 8962 * It is important that the dotdot journal entry 8963 * is added prior to the dot entry since dot writes 8964 * both the dot and dotdot links. These both must 8965 * be added after the primary link for the journal 8966 * to remain consistent. 8967 */ 8968 add_to_journal(&mkdir2->md_jaddref->ja_list); 8969 add_to_journal(&jaddref->ja_list); 8970 } 8971 /* 8972 * If we are adding a new directory remember this diradd so that if 8973 * we rename it we can keep the dot and dotdot dependencies. If 8974 * we are adding a new name for an inode that has a mkdiradd we 8975 * must be in rename and we have to move the dot and dotdot 8976 * dependencies to this new name. The old name is being orphaned 8977 * soon. 8978 */ 8979 if (mkdir1 != NULL) { 8980 if (inodedep->id_mkdiradd != NULL) 8981 panic("softdep_setup_directory_add: Existing mkdir"); 8982 inodedep->id_mkdiradd = dap; 8983 } else if (inodedep->id_mkdiradd) 8984 merge_diradd(inodedep, dap); 8985 if (newdirblk != NULL) { 8986 /* 8987 * There is nothing to do if we are already tracking 8988 * this block. 8989 */ 8990 if ((pagedep->pd_state & NEWBLOCK) != 0) { 8991 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 8992 FREE_LOCK(ump); 8993 return (0); 8994 } 8995 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk) 8996 == 0) 8997 panic("softdep_setup_directory_add: lost entry"); 8998 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 8999 pagedep->pd_state |= NEWBLOCK; 9000 pagedep->pd_newdirblk = newdirblk; 9001 newdirblk->db_pagedep = pagedep; 9002 FREE_LOCK(ump); 9003 /* 9004 * If we extended into an indirect signal direnter to sync. 9005 */ 9006 if (isindir) 9007 return (1); 9008 return (0); 9009 } 9010 FREE_LOCK(ump); 9011 return (0); 9012} 9013 9014/* 9015 * This procedure is called to change the offset of a directory 9016 * entry when compacting a directory block which must be owned 9017 * exclusively by the caller. Note that the actual entry movement 9018 * must be done in this procedure to ensure that no I/O completions 9019 * occur while the move is in progress. 9020 */ 9021void 9022softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 9023 struct buf *bp; /* Buffer holding directory block. */ 9024 struct inode *dp; /* inode for directory */ 9025 caddr_t base; /* address of dp->i_offset */ 9026 caddr_t oldloc; /* address of old directory location */ 9027 caddr_t newloc; /* address of new directory location */ 9028 int entrysize; /* size of directory entry */ 9029{ 9030 int offset, oldoffset, newoffset; 9031 struct pagedep *pagedep; 9032 struct jmvref *jmvref; 9033 struct diradd *dap; 9034 struct direct *de; 9035 struct mount *mp; 9036 struct ufsmount *ump; 9037 ufs_lbn_t lbn; 9038 int flags; 9039 9040 mp = ITOVFS(dp); 9041 ump = VFSTOUFS(mp); 9042 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 9043 ("softdep_change_directoryentry_offset called on " 9044 "non-softdep filesystem")); 9045 de = (struct direct *)oldloc; 9046 jmvref = NULL; 9047 flags = 0; 9048 /* 9049 * Moves are always journaled as it would be too complex to 9050 * determine if any affected adds or removes are present in the 9051 * journal. 9052 */ 9053 if (MOUNTEDSUJ(mp)) { 9054 flags = DEPALLOC; 9055 jmvref = newjmvref(dp, de->d_ino, 9056 I_OFFSET(dp) + (oldloc - base), 9057 I_OFFSET(dp) + (newloc - base)); 9058 } 9059 lbn = lblkno(ump->um_fs, I_OFFSET(dp)); 9060 offset = blkoff(ump->um_fs, I_OFFSET(dp)); 9061 oldoffset = offset + (oldloc - base); 9062 newoffset = offset + (newloc - base); 9063 ACQUIRE_LOCK(ump); 9064 if (pagedep_lookup(mp, bp, dp->i_number, lbn, flags, &pagedep) == 0) 9065 goto done; 9066 dap = diradd_lookup(pagedep, oldoffset); 9067 if (dap) { 9068 dap->da_offset = newoffset; 9069 newoffset = DIRADDHASH(newoffset); 9070 oldoffset = DIRADDHASH(oldoffset); 9071 if ((dap->da_state & ALLCOMPLETE) != ALLCOMPLETE && 9072 newoffset != oldoffset) { 9073 LIST_REMOVE(dap, da_pdlist); 9074 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[newoffset], 9075 dap, da_pdlist); 9076 } 9077 } 9078done: 9079 if (jmvref) { 9080 jmvref->jm_pagedep = pagedep; 9081 LIST_INSERT_HEAD(&pagedep->pd_jmvrefhd, jmvref, jm_deps); 9082 add_to_journal(&jmvref->jm_list); 9083 } 9084 bcopy(oldloc, newloc, entrysize); 9085 FREE_LOCK(ump); 9086} 9087 9088/* 9089 * Move the mkdir dependencies and journal work from one diradd to another 9090 * when renaming a directory. The new name must depend on the mkdir deps 9091 * completing as the old name did. Directories can only have one valid link 9092 * at a time so one must be canonical. 9093 */ 9094static void 9095merge_diradd(inodedep, newdap) 9096 struct inodedep *inodedep; 9097 struct diradd *newdap; 9098{ 9099 struct diradd *olddap; 9100 struct mkdir *mkdir, *nextmd; 9101 struct ufsmount *ump; 9102 short state; 9103 9104 olddap = inodedep->id_mkdiradd; 9105 inodedep->id_mkdiradd = newdap; 9106 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 9107 newdap->da_state &= ~DEPCOMPLETE; 9108 ump = VFSTOUFS(inodedep->id_list.wk_mp); 9109 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 9110 mkdir = nextmd) { 9111 nextmd = LIST_NEXT(mkdir, md_mkdirs); 9112 if (mkdir->md_diradd != olddap) 9113 continue; 9114 mkdir->md_diradd = newdap; 9115 state = mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY); 9116 newdap->da_state |= state; 9117 olddap->da_state &= ~state; 9118 if ((olddap->da_state & 9119 (MKDIR_PARENT | MKDIR_BODY)) == 0) 9120 break; 9121 } 9122 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 9123 panic("merge_diradd: unfound ref"); 9124 } 9125 /* 9126 * Any mkdir related journal items are not safe to be freed until 9127 * the new name is stable. 9128 */ 9129 jwork_move(&newdap->da_jwork, &olddap->da_jwork); 9130 olddap->da_state |= DEPCOMPLETE; 9131 complete_diradd(olddap); 9132} 9133 9134/* 9135 * Move the diradd to the pending list when all diradd dependencies are 9136 * complete. 9137 */ 9138static void 9139complete_diradd(dap) 9140 struct diradd *dap; 9141{ 9142 struct pagedep *pagedep; 9143 9144 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 9145 if (dap->da_state & DIRCHG) 9146 pagedep = dap->da_previous->dm_pagedep; 9147 else 9148 pagedep = dap->da_pagedep; 9149 LIST_REMOVE(dap, da_pdlist); 9150 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 9151 } 9152} 9153 9154/* 9155 * Cancel a diradd when a dirrem overlaps with it. We must cancel the journal 9156 * add entries and conditonally journal the remove. 9157 */ 9158static void 9159cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref) 9160 struct diradd *dap; 9161 struct dirrem *dirrem; 9162 struct jremref *jremref; 9163 struct jremref *dotremref; 9164 struct jremref *dotdotremref; 9165{ 9166 struct inodedep *inodedep; 9167 struct jaddref *jaddref; 9168 struct inoref *inoref; 9169 struct ufsmount *ump; 9170 struct mkdir *mkdir; 9171 9172 /* 9173 * If no remove references were allocated we're on a non-journaled 9174 * filesystem and can skip the cancel step. 9175 */ 9176 if (jremref == NULL) { 9177 free_diradd(dap, NULL); 9178 return; 9179 } 9180 /* 9181 * Cancel the primary name an free it if it does not require 9182 * journaling. 9183 */ 9184 if (inodedep_lookup(dap->da_list.wk_mp, dap->da_newinum, 9185 0, &inodedep) != 0) { 9186 /* Abort the addref that reference this diradd. */ 9187 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 9188 if (inoref->if_list.wk_type != D_JADDREF) 9189 continue; 9190 jaddref = (struct jaddref *)inoref; 9191 if (jaddref->ja_diradd != dap) 9192 continue; 9193 if (cancel_jaddref(jaddref, inodedep, 9194 &dirrem->dm_jwork) == 0) { 9195 free_jremref(jremref); 9196 jremref = NULL; 9197 } 9198 break; 9199 } 9200 } 9201 /* 9202 * Cancel subordinate names and free them if they do not require 9203 * journaling. 9204 */ 9205 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 9206 ump = VFSTOUFS(dap->da_list.wk_mp); 9207 LIST_FOREACH(mkdir, &ump->softdep_mkdirlisthd, md_mkdirs) { 9208 if (mkdir->md_diradd != dap) 9209 continue; 9210 if ((jaddref = mkdir->md_jaddref) == NULL) 9211 continue; 9212 mkdir->md_jaddref = NULL; 9213 if (mkdir->md_state & MKDIR_PARENT) { 9214 if (cancel_jaddref(jaddref, NULL, 9215 &dirrem->dm_jwork) == 0) { 9216 free_jremref(dotdotremref); 9217 dotdotremref = NULL; 9218 } 9219 } else { 9220 if (cancel_jaddref(jaddref, inodedep, 9221 &dirrem->dm_jwork) == 0) { 9222 free_jremref(dotremref); 9223 dotremref = NULL; 9224 } 9225 } 9226 } 9227 } 9228 9229 if (jremref) 9230 journal_jremref(dirrem, jremref, inodedep); 9231 if (dotremref) 9232 journal_jremref(dirrem, dotremref, inodedep); 9233 if (dotdotremref) 9234 journal_jremref(dirrem, dotdotremref, NULL); 9235 jwork_move(&dirrem->dm_jwork, &dap->da_jwork); 9236 free_diradd(dap, &dirrem->dm_jwork); 9237} 9238 9239/* 9240 * Free a diradd dependency structure. 9241 */ 9242static void 9243free_diradd(dap, wkhd) 9244 struct diradd *dap; 9245 struct workhead *wkhd; 9246{ 9247 struct dirrem *dirrem; 9248 struct pagedep *pagedep; 9249 struct inodedep *inodedep; 9250 struct mkdir *mkdir, *nextmd; 9251 struct ufsmount *ump; 9252 9253 ump = VFSTOUFS(dap->da_list.wk_mp); 9254 LOCK_OWNED(ump); 9255 LIST_REMOVE(dap, da_pdlist); 9256 if (dap->da_state & ONWORKLIST) 9257 WORKLIST_REMOVE(&dap->da_list); 9258 if ((dap->da_state & DIRCHG) == 0) { 9259 pagedep = dap->da_pagedep; 9260 } else { 9261 dirrem = dap->da_previous; 9262 pagedep = dirrem->dm_pagedep; 9263 dirrem->dm_dirinum = pagedep->pd_ino; 9264 dirrem->dm_state |= COMPLETE; 9265 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 9266 add_to_worklist(&dirrem->dm_list, 0); 9267 } 9268 if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum, 9269 0, &inodedep) != 0) 9270 if (inodedep->id_mkdiradd == dap) 9271 inodedep->id_mkdiradd = NULL; 9272 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 9273 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 9274 mkdir = nextmd) { 9275 nextmd = LIST_NEXT(mkdir, md_mkdirs); 9276 if (mkdir->md_diradd != dap) 9277 continue; 9278 dap->da_state &= 9279 ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 9280 LIST_REMOVE(mkdir, md_mkdirs); 9281 if (mkdir->md_state & ONWORKLIST) 9282 WORKLIST_REMOVE(&mkdir->md_list); 9283 if (mkdir->md_jaddref != NULL) 9284 panic("free_diradd: Unexpected jaddref"); 9285 WORKITEM_FREE(mkdir, D_MKDIR); 9286 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 9287 break; 9288 } 9289 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 9290 panic("free_diradd: unfound ref"); 9291 } 9292 if (inodedep) 9293 free_inodedep(inodedep); 9294 /* 9295 * Free any journal segments waiting for the directory write. 9296 */ 9297 handle_jwork(&dap->da_jwork); 9298 WORKITEM_FREE(dap, D_DIRADD); 9299} 9300 9301/* 9302 * Directory entry removal dependencies. 9303 * 9304 * When removing a directory entry, the entry's inode pointer must be 9305 * zero'ed on disk before the corresponding inode's link count is decremented 9306 * (possibly freeing the inode for re-use). This dependency is handled by 9307 * updating the directory entry but delaying the inode count reduction until 9308 * after the directory block has been written to disk. After this point, the 9309 * inode count can be decremented whenever it is convenient. 9310 */ 9311 9312/* 9313 * This routine should be called immediately after removing 9314 * a directory entry. The inode's link count should not be 9315 * decremented by the calling procedure -- the soft updates 9316 * code will do this task when it is safe. 9317 */ 9318void 9319softdep_setup_remove(bp, dp, ip, isrmdir) 9320 struct buf *bp; /* buffer containing directory block */ 9321 struct inode *dp; /* inode for the directory being modified */ 9322 struct inode *ip; /* inode for directory entry being removed */ 9323 int isrmdir; /* indicates if doing RMDIR */ 9324{ 9325 struct dirrem *dirrem, *prevdirrem; 9326 struct inodedep *inodedep; 9327 struct ufsmount *ump; 9328 int direct; 9329 9330 ump = ITOUMP(ip); 9331 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 9332 ("softdep_setup_remove called on non-softdep filesystem")); 9333 /* 9334 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. We want 9335 * newdirrem() to setup the full directory remove which requires 9336 * isrmdir > 1. 9337 */ 9338 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 9339 /* 9340 * Add the dirrem to the inodedep's pending remove list for quick 9341 * discovery later. 9342 */ 9343 if (inodedep_lookup(UFSTOVFS(ump), ip->i_number, 0, &inodedep) == 0) 9344 panic("softdep_setup_remove: Lost inodedep."); 9345 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 9346 dirrem->dm_state |= ONDEPLIST; 9347 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9348 9349 /* 9350 * If the COMPLETE flag is clear, then there were no active 9351 * entries and we want to roll back to a zeroed entry until 9352 * the new inode is committed to disk. If the COMPLETE flag is 9353 * set then we have deleted an entry that never made it to 9354 * disk. If the entry we deleted resulted from a name change, 9355 * then the old name still resides on disk. We cannot delete 9356 * its inode (returned to us in prevdirrem) until the zeroed 9357 * directory entry gets to disk. The new inode has never been 9358 * referenced on the disk, so can be deleted immediately. 9359 */ 9360 if ((dirrem->dm_state & COMPLETE) == 0) { 9361 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 9362 dm_next); 9363 FREE_LOCK(ump); 9364 } else { 9365 if (prevdirrem != NULL) 9366 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 9367 prevdirrem, dm_next); 9368 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 9369 direct = LIST_EMPTY(&dirrem->dm_jremrefhd); 9370 FREE_LOCK(ump); 9371 if (direct) 9372 handle_workitem_remove(dirrem, 0); 9373 } 9374} 9375 9376/* 9377 * Check for an entry matching 'offset' on both the pd_dirraddhd list and the 9378 * pd_pendinghd list of a pagedep. 9379 */ 9380static struct diradd * 9381diradd_lookup(pagedep, offset) 9382 struct pagedep *pagedep; 9383 int offset; 9384{ 9385 struct diradd *dap; 9386 9387 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 9388 if (dap->da_offset == offset) 9389 return (dap); 9390 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 9391 if (dap->da_offset == offset) 9392 return (dap); 9393 return (NULL); 9394} 9395 9396/* 9397 * Search for a .. diradd dependency in a directory that is being removed. 9398 * If the directory was renamed to a new parent we have a diradd rather 9399 * than a mkdir for the .. entry. We need to cancel it now before 9400 * it is found in truncate(). 9401 */ 9402static struct jremref * 9403cancel_diradd_dotdot(ip, dirrem, jremref) 9404 struct inode *ip; 9405 struct dirrem *dirrem; 9406 struct jremref *jremref; 9407{ 9408 struct pagedep *pagedep; 9409 struct diradd *dap; 9410 struct worklist *wk; 9411 9412 if (pagedep_lookup(ITOVFS(ip), NULL, ip->i_number, 0, 0, &pagedep) == 0) 9413 return (jremref); 9414 dap = diradd_lookup(pagedep, DOTDOT_OFFSET); 9415 if (dap == NULL) 9416 return (jremref); 9417 cancel_diradd(dap, dirrem, jremref, NULL, NULL); 9418 /* 9419 * Mark any journal work as belonging to the parent so it is freed 9420 * with the .. reference. 9421 */ 9422 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 9423 wk->wk_state |= MKDIR_PARENT; 9424 return (NULL); 9425} 9426 9427/* 9428 * Cancel the MKDIR_PARENT mkdir component of a diradd when we're going to 9429 * replace it with a dirrem/diradd pair as a result of re-parenting a 9430 * directory. This ensures that we don't simultaneously have a mkdir and 9431 * a diradd for the same .. entry. 9432 */ 9433static struct jremref * 9434cancel_mkdir_dotdot(ip, dirrem, jremref) 9435 struct inode *ip; 9436 struct dirrem *dirrem; 9437 struct jremref *jremref; 9438{ 9439 struct inodedep *inodedep; 9440 struct jaddref *jaddref; 9441 struct ufsmount *ump; 9442 struct mkdir *mkdir; 9443 struct diradd *dap; 9444 struct mount *mp; 9445 9446 mp = ITOVFS(ip); 9447 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 9448 return (jremref); 9449 dap = inodedep->id_mkdiradd; 9450 if (dap == NULL || (dap->da_state & MKDIR_PARENT) == 0) 9451 return (jremref); 9452 ump = VFSTOUFS(inodedep->id_list.wk_mp); 9453 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 9454 mkdir = LIST_NEXT(mkdir, md_mkdirs)) 9455 if (mkdir->md_diradd == dap && mkdir->md_state & MKDIR_PARENT) 9456 break; 9457 if (mkdir == NULL) 9458 panic("cancel_mkdir_dotdot: Unable to find mkdir\n"); 9459 if ((jaddref = mkdir->md_jaddref) != NULL) { 9460 mkdir->md_jaddref = NULL; 9461 jaddref->ja_state &= ~MKDIR_PARENT; 9462 if (inodedep_lookup(mp, jaddref->ja_ino, 0, &inodedep) == 0) 9463 panic("cancel_mkdir_dotdot: Lost parent inodedep"); 9464 if (cancel_jaddref(jaddref, inodedep, &dirrem->dm_jwork)) { 9465 journal_jremref(dirrem, jremref, inodedep); 9466 jremref = NULL; 9467 } 9468 } 9469 if (mkdir->md_state & ONWORKLIST) 9470 WORKLIST_REMOVE(&mkdir->md_list); 9471 mkdir->md_state |= ALLCOMPLETE; 9472 complete_mkdir(mkdir); 9473 return (jremref); 9474} 9475 9476static void 9477journal_jremref(dirrem, jremref, inodedep) 9478 struct dirrem *dirrem; 9479 struct jremref *jremref; 9480 struct inodedep *inodedep; 9481{ 9482 9483 if (inodedep == NULL) 9484 if (inodedep_lookup(jremref->jr_list.wk_mp, 9485 jremref->jr_ref.if_ino, 0, &inodedep) == 0) 9486 panic("journal_jremref: Lost inodedep"); 9487 LIST_INSERT_HEAD(&dirrem->dm_jremrefhd, jremref, jr_deps); 9488 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 9489 add_to_journal(&jremref->jr_list); 9490} 9491 9492static void 9493dirrem_journal(dirrem, jremref, dotremref, dotdotremref) 9494 struct dirrem *dirrem; 9495 struct jremref *jremref; 9496 struct jremref *dotremref; 9497 struct jremref *dotdotremref; 9498{ 9499 struct inodedep *inodedep; 9500 9501 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 0, 9502 &inodedep) == 0) 9503 panic("dirrem_journal: Lost inodedep"); 9504 journal_jremref(dirrem, jremref, inodedep); 9505 if (dotremref) 9506 journal_jremref(dirrem, dotremref, inodedep); 9507 if (dotdotremref) 9508 journal_jremref(dirrem, dotdotremref, NULL); 9509} 9510 9511/* 9512 * Allocate a new dirrem if appropriate and return it along with 9513 * its associated pagedep. Called without a lock, returns with lock. 9514 */ 9515static struct dirrem * 9516newdirrem(bp, dp, ip, isrmdir, prevdirremp) 9517 struct buf *bp; /* buffer containing directory block */ 9518 struct inode *dp; /* inode for the directory being modified */ 9519 struct inode *ip; /* inode for directory entry being removed */ 9520 int isrmdir; /* indicates if doing RMDIR */ 9521 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 9522{ 9523 int offset; 9524 ufs_lbn_t lbn; 9525 struct diradd *dap; 9526 struct dirrem *dirrem; 9527 struct pagedep *pagedep; 9528 struct jremref *jremref; 9529 struct jremref *dotremref; 9530 struct jremref *dotdotremref; 9531 struct vnode *dvp; 9532 struct ufsmount *ump; 9533 9534 /* 9535 * Whiteouts have no deletion dependencies. 9536 */ 9537 if (ip == NULL) 9538 panic("newdirrem: whiteout"); 9539 dvp = ITOV(dp); 9540 ump = ITOUMP(dp); 9541 9542 /* 9543 * If the system is over its limit and our filesystem is 9544 * responsible for more than our share of that usage and 9545 * we are not a snapshot, request some inodedep cleanup. 9546 * Limiting the number of dirrem structures will also limit 9547 * the number of freefile and freeblks structures. 9548 */ 9549 ACQUIRE_LOCK(ump); 9550 if (!IS_SNAPSHOT(ip) && softdep_excess_items(ump, D_DIRREM)) 9551 schedule_cleanup(UFSTOVFS(ump)); 9552 else 9553 FREE_LOCK(ump); 9554 dirrem = malloc(sizeof(struct dirrem), M_DIRREM, M_SOFTDEP_FLAGS | 9555 M_ZERO); 9556 workitem_alloc(&dirrem->dm_list, D_DIRREM, dvp->v_mount); 9557 LIST_INIT(&dirrem->dm_jremrefhd); 9558 LIST_INIT(&dirrem->dm_jwork); 9559 dirrem->dm_state = isrmdir ? RMDIR : 0; 9560 dirrem->dm_oldinum = ip->i_number; 9561 *prevdirremp = NULL; 9562 /* 9563 * Allocate remove reference structures to track journal write 9564 * dependencies. We will always have one for the link and 9565 * when doing directories we will always have one more for dot. 9566 * When renaming a directory we skip the dotdot link change so 9567 * this is not needed. 9568 */ 9569 jremref = dotremref = dotdotremref = NULL; 9570 if (DOINGSUJ(dvp)) { 9571 if (isrmdir) { 9572 jremref = newjremref(dirrem, dp, ip, I_OFFSET(dp), 9573 ip->i_effnlink + 2); 9574 dotremref = newjremref(dirrem, ip, ip, DOT_OFFSET, 9575 ip->i_effnlink + 1); 9576 dotdotremref = newjremref(dirrem, ip, dp, DOTDOT_OFFSET, 9577 dp->i_effnlink + 1); 9578 dotdotremref->jr_state |= MKDIR_PARENT; 9579 } else 9580 jremref = newjremref(dirrem, dp, ip, I_OFFSET(dp), 9581 ip->i_effnlink + 1); 9582 } 9583 ACQUIRE_LOCK(ump); 9584 lbn = lblkno(ump->um_fs, I_OFFSET(dp)); 9585 offset = blkoff(ump->um_fs, I_OFFSET(dp)); 9586 pagedep_lookup(UFSTOVFS(ump), bp, dp->i_number, lbn, DEPALLOC, 9587 &pagedep); 9588 dirrem->dm_pagedep = pagedep; 9589 dirrem->dm_offset = offset; 9590 /* 9591 * If we're renaming a .. link to a new directory, cancel any 9592 * existing MKDIR_PARENT mkdir. If it has already been canceled 9593 * the jremref is preserved for any potential diradd in this 9594 * location. This can not coincide with a rmdir. 9595 */ 9596 if (I_OFFSET(dp) == DOTDOT_OFFSET) { 9597 if (isrmdir) 9598 panic("newdirrem: .. directory change during remove?"); 9599 jremref = cancel_mkdir_dotdot(dp, dirrem, jremref); 9600 } 9601 /* 9602 * If we're removing a directory search for the .. dependency now and 9603 * cancel it. Any pending journal work will be added to the dirrem 9604 * to be completed when the workitem remove completes. 9605 */ 9606 if (isrmdir) 9607 dotdotremref = cancel_diradd_dotdot(ip, dirrem, dotdotremref); 9608 /* 9609 * Check for a diradd dependency for the same directory entry. 9610 * If present, then both dependencies become obsolete and can 9611 * be de-allocated. 9612 */ 9613 dap = diradd_lookup(pagedep, offset); 9614 if (dap == NULL) { 9615 /* 9616 * Link the jremref structures into the dirrem so they are 9617 * written prior to the pagedep. 9618 */ 9619 if (jremref) 9620 dirrem_journal(dirrem, jremref, dotremref, 9621 dotdotremref); 9622 return (dirrem); 9623 } 9624 /* 9625 * Must be ATTACHED at this point. 9626 */ 9627 if ((dap->da_state & ATTACHED) == 0) 9628 panic("newdirrem: not ATTACHED"); 9629 if (dap->da_newinum != ip->i_number) 9630 panic("newdirrem: inum %ju should be %ju", 9631 (uintmax_t)ip->i_number, (uintmax_t)dap->da_newinum); 9632 /* 9633 * If we are deleting a changed name that never made it to disk, 9634 * then return the dirrem describing the previous inode (which 9635 * represents the inode currently referenced from this entry on disk). 9636 */ 9637 if ((dap->da_state & DIRCHG) != 0) { 9638 *prevdirremp = dap->da_previous; 9639 dap->da_state &= ~DIRCHG; 9640 dap->da_pagedep = pagedep; 9641 } 9642 /* 9643 * We are deleting an entry that never made it to disk. 9644 * Mark it COMPLETE so we can delete its inode immediately. 9645 */ 9646 dirrem->dm_state |= COMPLETE; 9647 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref); 9648#ifdef INVARIANTS 9649 if (isrmdir == 0) { 9650 struct worklist *wk; 9651 9652 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 9653 if (wk->wk_state & (MKDIR_BODY | MKDIR_PARENT)) 9654 panic("bad wk %p (0x%X)\n", wk, wk->wk_state); 9655 } 9656#endif 9657 9658 return (dirrem); 9659} 9660 9661/* 9662 * Directory entry change dependencies. 9663 * 9664 * Changing an existing directory entry requires that an add operation 9665 * be completed first followed by a deletion. The semantics for the addition 9666 * are identical to the description of adding a new entry above except 9667 * that the rollback is to the old inode number rather than zero. Once 9668 * the addition dependency is completed, the removal is done as described 9669 * in the removal routine above. 9670 */ 9671 9672/* 9673 * This routine should be called immediately after changing 9674 * a directory entry. The inode's link count should not be 9675 * decremented by the calling procedure -- the soft updates 9676 * code will perform this task when it is safe. 9677 */ 9678void 9679softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 9680 struct buf *bp; /* buffer containing directory block */ 9681 struct inode *dp; /* inode for the directory being modified */ 9682 struct inode *ip; /* inode for directory entry being removed */ 9683 ino_t newinum; /* new inode number for changed entry */ 9684 int isrmdir; /* indicates if doing RMDIR */ 9685{ 9686 int offset; 9687 struct diradd *dap = NULL; 9688 struct dirrem *dirrem, *prevdirrem; 9689 struct pagedep *pagedep; 9690 struct inodedep *inodedep; 9691 struct jaddref *jaddref; 9692 struct mount *mp; 9693 struct ufsmount *ump; 9694 9695 mp = ITOVFS(dp); 9696 ump = VFSTOUFS(mp); 9697 offset = blkoff(ump->um_fs, I_OFFSET(dp)); 9698 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 9699 ("softdep_setup_directory_change called on non-softdep filesystem")); 9700 9701 /* 9702 * Whiteouts do not need diradd dependencies. 9703 */ 9704 if (newinum != UFS_WINO) { 9705 dap = malloc(sizeof(struct diradd), 9706 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); 9707 workitem_alloc(&dap->da_list, D_DIRADD, mp); 9708 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 9709 dap->da_offset = offset; 9710 dap->da_newinum = newinum; 9711 LIST_INIT(&dap->da_jwork); 9712 } 9713 9714 /* 9715 * Allocate a new dirrem and ACQUIRE_LOCK. 9716 */ 9717 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 9718 pagedep = dirrem->dm_pagedep; 9719 /* 9720 * The possible values for isrmdir: 9721 * 0 - non-directory file rename 9722 * 1 - directory rename within same directory 9723 * inum - directory rename to new directory of given inode number 9724 * When renaming to a new directory, we are both deleting and 9725 * creating a new directory entry, so the link count on the new 9726 * directory should not change. Thus we do not need the followup 9727 * dirrem which is usually done in handle_workitem_remove. We set 9728 * the DIRCHG flag to tell handle_workitem_remove to skip the 9729 * followup dirrem. 9730 */ 9731 if (isrmdir > 1) 9732 dirrem->dm_state |= DIRCHG; 9733 9734 /* 9735 * Whiteouts have no additional dependencies, 9736 * so just put the dirrem on the correct list. 9737 */ 9738 if (newinum == UFS_WINO) { 9739 if ((dirrem->dm_state & COMPLETE) == 0) { 9740 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 9741 dm_next); 9742 } else { 9743 dirrem->dm_dirinum = pagedep->pd_ino; 9744 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 9745 add_to_worklist(&dirrem->dm_list, 0); 9746 } 9747 FREE_LOCK(ump); 9748 return; 9749 } 9750 /* 9751 * Add the dirrem to the inodedep's pending remove list for quick 9752 * discovery later. A valid nlinkdelta ensures that this lookup 9753 * will not fail. 9754 */ 9755 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 9756 panic("softdep_setup_directory_change: Lost inodedep."); 9757 dirrem->dm_state |= ONDEPLIST; 9758 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9759 9760 /* 9761 * If the COMPLETE flag is clear, then there were no active 9762 * entries and we want to roll back to the previous inode until 9763 * the new inode is committed to disk. If the COMPLETE flag is 9764 * set, then we have deleted an entry that never made it to disk. 9765 * If the entry we deleted resulted from a name change, then the old 9766 * inode reference still resides on disk. Any rollback that we do 9767 * needs to be to that old inode (returned to us in prevdirrem). If 9768 * the entry we deleted resulted from a create, then there is 9769 * no entry on the disk, so we want to roll back to zero rather 9770 * than the uncommitted inode. In either of the COMPLETE cases we 9771 * want to immediately free the unwritten and unreferenced inode. 9772 */ 9773 if ((dirrem->dm_state & COMPLETE) == 0) { 9774 dap->da_previous = dirrem; 9775 } else { 9776 if (prevdirrem != NULL) { 9777 dap->da_previous = prevdirrem; 9778 } else { 9779 dap->da_state &= ~DIRCHG; 9780 dap->da_pagedep = pagedep; 9781 } 9782 dirrem->dm_dirinum = pagedep->pd_ino; 9783 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 9784 add_to_worklist(&dirrem->dm_list, 0); 9785 } 9786 /* 9787 * Lookup the jaddref for this journal entry. We must finish 9788 * initializing it and make the diradd write dependent on it. 9789 * If we're not journaling, put it on the id_bufwait list if the 9790 * inode is not yet written. If it is written, do the post-inode 9791 * write processing to put it on the id_pendinghd list. 9792 */ 9793 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); 9794 if (MOUNTEDSUJ(mp)) { 9795 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 9796 inoreflst); 9797 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 9798 ("softdep_setup_directory_change: bad jaddref %p", 9799 jaddref)); 9800 jaddref->ja_diroff = I_OFFSET(dp); 9801 jaddref->ja_diradd = dap; 9802 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 9803 dap, da_pdlist); 9804 add_to_journal(&jaddref->ja_list); 9805 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 9806 dap->da_state |= COMPLETE; 9807 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 9808 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 9809 } else { 9810 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 9811 dap, da_pdlist); 9812 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 9813 } 9814 /* 9815 * If we're making a new name for a directory that has not been 9816 * committed when need to move the dot and dotdot references to 9817 * this new name. 9818 */ 9819 if (inodedep->id_mkdiradd && I_OFFSET(dp) != DOTDOT_OFFSET) 9820 merge_diradd(inodedep, dap); 9821 FREE_LOCK(ump); 9822} 9823 9824/* 9825 * Called whenever the link count on an inode is changed. 9826 * It creates an inode dependency so that the new reference(s) 9827 * to the inode cannot be committed to disk until the updated 9828 * inode has been written. 9829 */ 9830void 9831softdep_change_linkcnt(ip) 9832 struct inode *ip; /* the inode with the increased link count */ 9833{ 9834 struct inodedep *inodedep; 9835 struct ufsmount *ump; 9836 9837 ump = ITOUMP(ip); 9838 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 9839 ("softdep_change_linkcnt called on non-softdep filesystem")); 9840 ACQUIRE_LOCK(ump); 9841 inodedep_lookup(UFSTOVFS(ump), ip->i_number, DEPALLOC, &inodedep); 9842 if (ip->i_nlink < ip->i_effnlink) 9843 panic("softdep_change_linkcnt: bad delta"); 9844 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9845 FREE_LOCK(ump); 9846} 9847 9848/* 9849 * Attach a sbdep dependency to the superblock buf so that we can keep 9850 * track of the head of the linked list of referenced but unlinked inodes. 9851 */ 9852void 9853softdep_setup_sbupdate(ump, fs, bp) 9854 struct ufsmount *ump; 9855 struct fs *fs; 9856 struct buf *bp; 9857{ 9858 struct sbdep *sbdep; 9859 struct worklist *wk; 9860 9861 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 9862 ("softdep_setup_sbupdate called on non-softdep filesystem")); 9863 LIST_FOREACH(wk, &bp->b_dep, wk_list) 9864 if (wk->wk_type == D_SBDEP) 9865 break; 9866 if (wk != NULL) 9867 return; 9868 sbdep = malloc(sizeof(struct sbdep), M_SBDEP, M_SOFTDEP_FLAGS); 9869 workitem_alloc(&sbdep->sb_list, D_SBDEP, UFSTOVFS(ump)); 9870 sbdep->sb_fs = fs; 9871 sbdep->sb_ump = ump; 9872 ACQUIRE_LOCK(ump); 9873 WORKLIST_INSERT(&bp->b_dep, &sbdep->sb_list); 9874 FREE_LOCK(ump); 9875} 9876 9877/* 9878 * Return the first unlinked inodedep which is ready to be the head of the 9879 * list. The inodedep and all those after it must have valid next pointers. 9880 */ 9881static struct inodedep * 9882first_unlinked_inodedep(ump) 9883 struct ufsmount *ump; 9884{ 9885 struct inodedep *inodedep; 9886 struct inodedep *idp; 9887 9888 LOCK_OWNED(ump); 9889 for (inodedep = TAILQ_LAST(&ump->softdep_unlinked, inodedeplst); 9890 inodedep; inodedep = idp) { 9891 if ((inodedep->id_state & UNLINKNEXT) == 0) 9892 return (NULL); 9893 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9894 if (idp == NULL || (idp->id_state & UNLINKNEXT) == 0) 9895 break; 9896 if ((inodedep->id_state & UNLINKPREV) == 0) 9897 break; 9898 } 9899 return (inodedep); 9900} 9901 9902/* 9903 * Set the sujfree unlinked head pointer prior to writing a superblock. 9904 */ 9905static void 9906initiate_write_sbdep(sbdep) 9907 struct sbdep *sbdep; 9908{ 9909 struct inodedep *inodedep; 9910 struct fs *bpfs; 9911 struct fs *fs; 9912 9913 bpfs = sbdep->sb_fs; 9914 fs = sbdep->sb_ump->um_fs; 9915 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9916 if (inodedep) { 9917 fs->fs_sujfree = inodedep->id_ino; 9918 inodedep->id_state |= UNLINKPREV; 9919 } else 9920 fs->fs_sujfree = 0; 9921 bpfs->fs_sujfree = fs->fs_sujfree; 9922 /* 9923 * Because we have made changes to the superblock, we need to 9924 * recompute its check-hash. 9925 */ 9926 bpfs->fs_ckhash = ffs_calc_sbhash(bpfs); 9927} 9928 9929/* 9930 * After a superblock is written determine whether it must be written again 9931 * due to a changing unlinked list head. 9932 */ 9933static int 9934handle_written_sbdep(sbdep, bp) 9935 struct sbdep *sbdep; 9936 struct buf *bp; 9937{ 9938 struct inodedep *inodedep; 9939 struct fs *fs; 9940 9941 LOCK_OWNED(sbdep->sb_ump); 9942 fs = sbdep->sb_fs; 9943 /* 9944 * If the superblock doesn't match the in-memory list start over. 9945 */ 9946 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9947 if ((inodedep && fs->fs_sujfree != inodedep->id_ino) || 9948 (inodedep == NULL && fs->fs_sujfree != 0)) { 9949 bdirty(bp); 9950 return (1); 9951 } 9952 WORKITEM_FREE(sbdep, D_SBDEP); 9953 if (fs->fs_sujfree == 0) 9954 return (0); 9955 /* 9956 * Now that we have a record of this inode in stable store allow it 9957 * to be written to free up pending work. Inodes may see a lot of 9958 * write activity after they are unlinked which we must not hold up. 9959 */ 9960 for (; inodedep != NULL; inodedep = TAILQ_NEXT(inodedep, id_unlinked)) { 9961 if ((inodedep->id_state & UNLINKLINKS) != UNLINKLINKS) 9962 panic("handle_written_sbdep: Bad inodedep %p (0x%X)", 9963 inodedep, inodedep->id_state); 9964 if (inodedep->id_state & UNLINKONLIST) 9965 break; 9966 inodedep->id_state |= DEPCOMPLETE | UNLINKONLIST; 9967 } 9968 9969 return (0); 9970} 9971 9972/* 9973 * Mark an inodedep as unlinked and insert it into the in-memory unlinked list. 9974 */ 9975static void 9976unlinked_inodedep(mp, inodedep) 9977 struct mount *mp; 9978 struct inodedep *inodedep; 9979{ 9980 struct ufsmount *ump; 9981 9982 ump = VFSTOUFS(mp); 9983 LOCK_OWNED(ump); 9984 if (MOUNTEDSUJ(mp) == 0) 9985 return; 9986 ump->um_fs->fs_fmod = 1; 9987 if (inodedep->id_state & UNLINKED) 9988 panic("unlinked_inodedep: %p already unlinked\n", inodedep); 9989 inodedep->id_state |= UNLINKED; 9990 TAILQ_INSERT_HEAD(&ump->softdep_unlinked, inodedep, id_unlinked); 9991} 9992 9993/* 9994 * Remove an inodedep from the unlinked inodedep list. This may require 9995 * disk writes if the inode has made it that far. 9996 */ 9997static void 9998clear_unlinked_inodedep(inodedep) 9999 struct inodedep *inodedep; 10000{ 10001 struct ufs2_dinode *dip; 10002 struct ufsmount *ump; 10003 struct inodedep *idp; 10004 struct inodedep *idn; 10005 struct fs *fs, *bpfs; 10006 struct buf *bp; 10007 daddr_t dbn; 10008 ino_t ino; 10009 ino_t nino; 10010 ino_t pino; 10011 int error; 10012 10013 ump = VFSTOUFS(inodedep->id_list.wk_mp); 10014 fs = ump->um_fs; 10015 ino = inodedep->id_ino; 10016 error = 0; 10017 for (;;) { 10018 LOCK_OWNED(ump); 10019 KASSERT((inodedep->id_state & UNLINKED) != 0, 10020 ("clear_unlinked_inodedep: inodedep %p not unlinked", 10021 inodedep)); 10022 /* 10023 * If nothing has yet been written simply remove us from 10024 * the in memory list and return. This is the most common 10025 * case where handle_workitem_remove() loses the final 10026 * reference. 10027 */ 10028 if ((inodedep->id_state & UNLINKLINKS) == 0) 10029 break; 10030 /* 10031 * If we have a NEXT pointer and no PREV pointer we can simply 10032 * clear NEXT's PREV and remove ourselves from the list. Be 10033 * careful not to clear PREV if the superblock points at 10034 * next as well. 10035 */ 10036 idn = TAILQ_NEXT(inodedep, id_unlinked); 10037 if ((inodedep->id_state & UNLINKLINKS) == UNLINKNEXT) { 10038 if (idn && fs->fs_sujfree != idn->id_ino) 10039 idn->id_state &= ~UNLINKPREV; 10040 break; 10041 } 10042 /* 10043 * Here we have an inodedep which is actually linked into 10044 * the list. We must remove it by forcing a write to the 10045 * link before us, whether it be the superblock or an inode. 10046 * Unfortunately the list may change while we're waiting 10047 * on the buf lock for either resource so we must loop until 10048 * we lock the right one. If both the superblock and an 10049 * inode point to this inode we must clear the inode first 10050 * followed by the superblock. 10051 */ 10052 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 10053 pino = 0; 10054 if (idp && (idp->id_state & UNLINKNEXT)) 10055 pino = idp->id_ino; 10056 FREE_LOCK(ump); 10057 if (pino == 0) { 10058 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 10059 (int)fs->fs_sbsize, 0, 0, 0); 10060 } else { 10061 dbn = fsbtodb(fs, ino_to_fsba(fs, pino)); 10062 error = ffs_breadz(ump, ump->um_devvp, dbn, dbn, 10063 (int)fs->fs_bsize, NULL, NULL, 0, NOCRED, 0, NULL, 10064 &bp); 10065 } 10066 ACQUIRE_LOCK(ump); 10067 if (error) 10068 break; 10069 /* If the list has changed restart the loop. */ 10070 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 10071 nino = 0; 10072 if (idp && (idp->id_state & UNLINKNEXT)) 10073 nino = idp->id_ino; 10074 if (nino != pino || 10075 (inodedep->id_state & UNLINKPREV) != UNLINKPREV) { 10076 FREE_LOCK(ump); 10077 brelse(bp); 10078 ACQUIRE_LOCK(ump); 10079 continue; 10080 } 10081 nino = 0; 10082 idn = TAILQ_NEXT(inodedep, id_unlinked); 10083 if (idn) 10084 nino = idn->id_ino; 10085 /* 10086 * Remove us from the in memory list. After this we cannot 10087 * access the inodedep. 10088 */ 10089 KASSERT((inodedep->id_state & UNLINKED) != 0, 10090 ("clear_unlinked_inodedep: inodedep %p not unlinked", 10091 inodedep)); 10092 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 10093 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 10094 FREE_LOCK(ump); 10095 /* 10096 * The predecessor's next pointer is manually updated here 10097 * so that the NEXT flag is never cleared for an element 10098 * that is in the list. 10099 */ 10100 if (pino == 0) { 10101 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 10102 bpfs = (struct fs *)bp->b_data; 10103 ffs_oldfscompat_write(bpfs, ump); 10104 softdep_setup_sbupdate(ump, bpfs, bp); 10105 /* 10106 * Because we may have made changes to the superblock, 10107 * we need to recompute its check-hash. 10108 */ 10109 bpfs->fs_ckhash = ffs_calc_sbhash(bpfs); 10110 } else if (fs->fs_magic == FS_UFS1_MAGIC) { 10111 ((struct ufs1_dinode *)bp->b_data + 10112 ino_to_fsbo(fs, pino))->di_freelink = nino; 10113 } else { 10114 dip = (struct ufs2_dinode *)bp->b_data + 10115 ino_to_fsbo(fs, pino); 10116 dip->di_freelink = nino; 10117 ffs_update_dinode_ckhash(fs, dip); 10118 } 10119 /* 10120 * If the bwrite fails we have no recourse to recover. The 10121 * filesystem is corrupted already. 10122 */ 10123 bwrite(bp); 10124 ACQUIRE_LOCK(ump); 10125 /* 10126 * If the superblock pointer still needs to be cleared force 10127 * a write here. 10128 */ 10129 if (fs->fs_sujfree == ino) { 10130 FREE_LOCK(ump); 10131 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 10132 (int)fs->fs_sbsize, 0, 0, 0); 10133 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 10134 bpfs = (struct fs *)bp->b_data; 10135 ffs_oldfscompat_write(bpfs, ump); 10136 softdep_setup_sbupdate(ump, bpfs, bp); 10137 /* 10138 * Because we may have made changes to the superblock, 10139 * we need to recompute its check-hash. 10140 */ 10141 bpfs->fs_ckhash = ffs_calc_sbhash(bpfs); 10142 bwrite(bp); 10143 ACQUIRE_LOCK(ump); 10144 } 10145 10146 if (fs->fs_sujfree != ino) 10147 return; 10148 panic("clear_unlinked_inodedep: Failed to clear free head"); 10149 } 10150 if (inodedep->id_ino == fs->fs_sujfree) 10151 panic("clear_unlinked_inodedep: Freeing head of free list"); 10152 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 10153 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 10154 return; 10155} 10156 10157/* 10158 * This workitem decrements the inode's link count. 10159 * If the link count reaches zero, the file is removed. 10160 */ 10161static int 10162handle_workitem_remove(dirrem, flags) 10163 struct dirrem *dirrem; 10164 int flags; 10165{ 10166 struct inodedep *inodedep; 10167 struct workhead dotdotwk; 10168 struct worklist *wk; 10169 struct ufsmount *ump; 10170 struct mount *mp; 10171 struct vnode *vp; 10172 struct inode *ip; 10173 ino_t oldinum; 10174 10175 if (dirrem->dm_state & ONWORKLIST) 10176 panic("handle_workitem_remove: dirrem %p still on worklist", 10177 dirrem); 10178 oldinum = dirrem->dm_oldinum; 10179 mp = dirrem->dm_list.wk_mp; 10180 ump = VFSTOUFS(mp); 10181 flags |= LK_EXCLUSIVE; 10182 if (ffs_vgetf(mp, oldinum, flags, &vp, FFSV_FORCEINSMQ) != 0) 10183 return (EBUSY); 10184 ip = VTOI(vp); 10185 MPASS(ip->i_mode != 0); 10186 ACQUIRE_LOCK(ump); 10187 if ((inodedep_lookup(mp, oldinum, 0, &inodedep)) == 0) 10188 panic("handle_workitem_remove: lost inodedep"); 10189 if (dirrem->dm_state & ONDEPLIST) 10190 LIST_REMOVE(dirrem, dm_inonext); 10191 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 10192 ("handle_workitem_remove: Journal entries not written.")); 10193 10194 /* 10195 * Move all dependencies waiting on the remove to complete 10196 * from the dirrem to the inode inowait list to be completed 10197 * after the inode has been updated and written to disk. 10198 * 10199 * Any marked MKDIR_PARENT are saved to be completed when the 10200 * dotdot ref is removed unless DIRCHG is specified. For 10201 * directory change operations there will be no further 10202 * directory writes and the jsegdeps need to be moved along 10203 * with the rest to be completed when the inode is free or 10204 * stable in the inode free list. 10205 */ 10206 LIST_INIT(&dotdotwk); 10207 while ((wk = LIST_FIRST(&dirrem->dm_jwork)) != NULL) { 10208 WORKLIST_REMOVE(wk); 10209 if ((dirrem->dm_state & DIRCHG) == 0 && 10210 wk->wk_state & MKDIR_PARENT) { 10211 wk->wk_state &= ~MKDIR_PARENT; 10212 WORKLIST_INSERT(&dotdotwk, wk); 10213 continue; 10214 } 10215 WORKLIST_INSERT(&inodedep->id_inowait, wk); 10216 } 10217 LIST_SWAP(&dirrem->dm_jwork, &dotdotwk, worklist, wk_list); 10218 /* 10219 * Normal file deletion. 10220 */ 10221 if ((dirrem->dm_state & RMDIR) == 0) { 10222 ip->i_nlink--; 10223 KASSERT(ip->i_nlink >= 0, ("handle_workitem_remove: file ino " 10224 "%ju negative i_nlink %d", (intmax_t)ip->i_number, 10225 ip->i_nlink)); 10226 DIP_SET(ip, i_nlink, ip->i_nlink); 10227 UFS_INODE_SET_FLAG(ip, IN_CHANGE); 10228 if (ip->i_nlink < ip->i_effnlink) 10229 panic("handle_workitem_remove: bad file delta"); 10230 if (ip->i_nlink == 0) 10231 unlinked_inodedep(mp, inodedep); 10232 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 10233 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 10234 ("handle_workitem_remove: worklist not empty. %s", 10235 TYPENAME(LIST_FIRST(&dirrem->dm_jwork)->wk_type))); 10236 WORKITEM_FREE(dirrem, D_DIRREM); 10237 FREE_LOCK(ump); 10238 goto out; 10239 } 10240 /* 10241 * Directory deletion. Decrement reference count for both the 10242 * just deleted parent directory entry and the reference for ".". 10243 * Arrange to have the reference count on the parent decremented 10244 * to account for the loss of "..". 10245 */ 10246 ip->i_nlink -= 2; 10247 KASSERT(ip->i_nlink >= 0, ("handle_workitem_remove: directory ino " 10248 "%ju negative i_nlink %d", (intmax_t)ip->i_number, ip->i_nlink)); 10249 DIP_SET(ip, i_nlink, ip->i_nlink); 10250 UFS_INODE_SET_FLAG(ip, IN_CHANGE); 10251 if (ip->i_nlink < ip->i_effnlink) 10252 panic("handle_workitem_remove: bad dir delta"); 10253 if (ip->i_nlink == 0) 10254 unlinked_inodedep(mp, inodedep); 10255 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 10256 /* 10257 * Rename a directory to a new parent. Since, we are both deleting 10258 * and creating a new directory entry, the link count on the new 10259 * directory should not change. Thus we skip the followup dirrem. 10260 */ 10261 if (dirrem->dm_state & DIRCHG) { 10262 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 10263 ("handle_workitem_remove: DIRCHG and worklist not empty.")); 10264 WORKITEM_FREE(dirrem, D_DIRREM); 10265 FREE_LOCK(ump); 10266 goto out; 10267 } 10268 dirrem->dm_state = ONDEPLIST; 10269 dirrem->dm_oldinum = dirrem->dm_dirinum; 10270 /* 10271 * Place the dirrem on the parent's diremhd list. 10272 */ 10273 if (inodedep_lookup(mp, dirrem->dm_oldinum, 0, &inodedep) == 0) 10274 panic("handle_workitem_remove: lost dir inodedep"); 10275 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 10276 /* 10277 * If the allocated inode has never been written to disk, then 10278 * the on-disk inode is zero'ed and we can remove the file 10279 * immediately. When journaling if the inode has been marked 10280 * unlinked and not DEPCOMPLETE we know it can never be written. 10281 */ 10282 inodedep_lookup(mp, oldinum, 0, &inodedep); 10283 if (inodedep == NULL || 10284 (inodedep->id_state & (DEPCOMPLETE | UNLINKED)) == UNLINKED || 10285 check_inode_unwritten(inodedep)) { 10286 FREE_LOCK(ump); 10287 vput(vp); 10288 return handle_workitem_remove(dirrem, flags); 10289 } 10290 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 10291 FREE_LOCK(ump); 10292 UFS_INODE_SET_FLAG(ip, IN_CHANGE); 10293out: 10294 ffs_update(vp, 0); 10295 vput(vp); 10296 return (0); 10297} 10298 10299/* 10300 * Inode de-allocation dependencies. 10301 * 10302 * When an inode's link count is reduced to zero, it can be de-allocated. We 10303 * found it convenient to postpone de-allocation until after the inode is 10304 * written to disk with its new link count (zero). At this point, all of the 10305 * on-disk inode's block pointers are nullified and, with careful dependency 10306 * list ordering, all dependencies related to the inode will be satisfied and 10307 * the corresponding dependency structures de-allocated. So, if/when the 10308 * inode is reused, there will be no mixing of old dependencies with new 10309 * ones. This artificial dependency is set up by the block de-allocation 10310 * procedure above (softdep_setup_freeblocks) and completed by the 10311 * following procedure. 10312 */ 10313static void 10314handle_workitem_freefile(freefile) 10315 struct freefile *freefile; 10316{ 10317 struct workhead wkhd; 10318 struct fs *fs; 10319 struct ufsmount *ump; 10320 int error; 10321#ifdef INVARIANTS 10322 struct inodedep *idp; 10323#endif 10324 10325 ump = VFSTOUFS(freefile->fx_list.wk_mp); 10326 fs = ump->um_fs; 10327#ifdef INVARIANTS 10328 ACQUIRE_LOCK(ump); 10329 error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp); 10330 FREE_LOCK(ump); 10331 if (error) 10332 panic("handle_workitem_freefile: inodedep %p survived", idp); 10333#endif 10334 UFS_LOCK(ump); 10335 fs->fs_pendinginodes -= 1; 10336 UFS_UNLOCK(ump); 10337 LIST_INIT(&wkhd); 10338 LIST_SWAP(&freefile->fx_jwork, &wkhd, worklist, wk_list); 10339 if ((error = ffs_freefile(ump, fs, freefile->fx_devvp, 10340 freefile->fx_oldinum, freefile->fx_mode, &wkhd)) != 0) 10341 softdep_error("handle_workitem_freefile", error); 10342 ACQUIRE_LOCK(ump); 10343 WORKITEM_FREE(freefile, D_FREEFILE); 10344 FREE_LOCK(ump); 10345} 10346 10347/* 10348 * Helper function which unlinks marker element from work list and returns 10349 * the next element on the list. 10350 */ 10351static __inline struct worklist * 10352markernext(struct worklist *marker) 10353{ 10354 struct worklist *next; 10355 10356 next = LIST_NEXT(marker, wk_list); 10357 LIST_REMOVE(marker, wk_list); 10358 return next; 10359} 10360 10361/* 10362 * Disk writes. 10363 * 10364 * The dependency structures constructed above are most actively used when file 10365 * system blocks are written to disk. No constraints are placed on when a 10366 * block can be written, but unsatisfied update dependencies are made safe by 10367 * modifying (or replacing) the source memory for the duration of the disk 10368 * write. When the disk write completes, the memory block is again brought 10369 * up-to-date. 10370 * 10371 * In-core inode structure reclamation. 10372 * 10373 * Because there are a finite number of "in-core" inode structures, they are 10374 * reused regularly. By transferring all inode-related dependencies to the 10375 * in-memory inode block and indexing them separately (via "inodedep"s), we 10376 * can allow "in-core" inode structures to be reused at any time and avoid 10377 * any increase in contention. 10378 * 10379 * Called just before entering the device driver to initiate a new disk I/O. 10380 * The buffer must be locked, thus, no I/O completion operations can occur 10381 * while we are manipulating its associated dependencies. 10382 */ 10383static void 10384softdep_disk_io_initiation(bp) 10385 struct buf *bp; /* structure describing disk write to occur */ 10386{ 10387 struct worklist *wk; 10388 struct worklist marker; 10389 struct inodedep *inodedep; 10390 struct freeblks *freeblks; 10391 struct jblkdep *jblkdep; 10392 struct newblk *newblk; 10393 struct ufsmount *ump; 10394 10395 /* 10396 * We only care about write operations. There should never 10397 * be dependencies for reads. 10398 */ 10399 if (bp->b_iocmd != BIO_WRITE) 10400 panic("softdep_disk_io_initiation: not write"); 10401 10402 if (bp->b_vflags & BV_BKGRDINPROG) 10403 panic("softdep_disk_io_initiation: Writing buffer with " 10404 "background write in progress: %p", bp); 10405 10406 ump = softdep_bp_to_mp(bp); 10407 if (ump == NULL) 10408 return; 10409 10410 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 10411 PHOLD(curproc); /* Don't swap out kernel stack */ 10412 ACQUIRE_LOCK(ump); 10413 /* 10414 * Do any necessary pre-I/O processing. 10415 */ 10416 for (wk = LIST_FIRST(&bp->b_dep); wk != NULL; 10417 wk = markernext(&marker)) { 10418 LIST_INSERT_AFTER(wk, &marker, wk_list); 10419 switch (wk->wk_type) { 10420 case D_PAGEDEP: 10421 initiate_write_filepage(WK_PAGEDEP(wk), bp); 10422 continue; 10423 10424 case D_INODEDEP: 10425 inodedep = WK_INODEDEP(wk); 10426 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) 10427 initiate_write_inodeblock_ufs1(inodedep, bp); 10428 else 10429 initiate_write_inodeblock_ufs2(inodedep, bp); 10430 continue; 10431 10432 case D_INDIRDEP: 10433 initiate_write_indirdep(WK_INDIRDEP(wk), bp); 10434 continue; 10435 10436 case D_BMSAFEMAP: 10437 initiate_write_bmsafemap(WK_BMSAFEMAP(wk), bp); 10438 continue; 10439 10440 case D_JSEG: 10441 WK_JSEG(wk)->js_buf = NULL; 10442 continue; 10443 10444 case D_FREEBLKS: 10445 freeblks = WK_FREEBLKS(wk); 10446 jblkdep = LIST_FIRST(&freeblks->fb_jblkdephd); 10447 /* 10448 * We have to wait for the freeblks to be journaled 10449 * before we can write an inodeblock with updated 10450 * pointers. Be careful to arrange the marker so 10451 * we revisit the freeblks if it's not removed by 10452 * the first jwait(). 10453 */ 10454 if (jblkdep != NULL) { 10455 LIST_REMOVE(&marker, wk_list); 10456 LIST_INSERT_BEFORE(wk, &marker, wk_list); 10457 jwait(&jblkdep->jb_list, MNT_WAIT); 10458 } 10459 continue; 10460 case D_ALLOCDIRECT: 10461 case D_ALLOCINDIR: 10462 /* 10463 * We have to wait for the jnewblk to be journaled 10464 * before we can write to a block if the contents 10465 * may be confused with an earlier file's indirect 10466 * at recovery time. Handle the marker as described 10467 * above. 10468 */ 10469 newblk = WK_NEWBLK(wk); 10470 if (newblk->nb_jnewblk != NULL && 10471 indirblk_lookup(newblk->nb_list.wk_mp, 10472 newblk->nb_newblkno)) { 10473 LIST_REMOVE(&marker, wk_list); 10474 LIST_INSERT_BEFORE(wk, &marker, wk_list); 10475 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 10476 } 10477 continue; 10478 10479 case D_SBDEP: 10480 initiate_write_sbdep(WK_SBDEP(wk)); 10481 continue; 10482 10483 case D_MKDIR: 10484 case D_FREEWORK: 10485 case D_FREEDEP: 10486 case D_JSEGDEP: 10487 continue; 10488 10489 default: 10490 panic("handle_disk_io_initiation: Unexpected type %s", 10491 TYPENAME(wk->wk_type)); 10492 /* NOTREACHED */ 10493 } 10494 } 10495 FREE_LOCK(ump); 10496 PRELE(curproc); /* Allow swapout of kernel stack */ 10497} 10498 10499/* 10500 * Called from within the procedure above to deal with unsatisfied 10501 * allocation dependencies in a directory. The buffer must be locked, 10502 * thus, no I/O completion operations can occur while we are 10503 * manipulating its associated dependencies. 10504 */ 10505static void 10506initiate_write_filepage(pagedep, bp) 10507 struct pagedep *pagedep; 10508 struct buf *bp; 10509{ 10510 struct jremref *jremref; 10511 struct jmvref *jmvref; 10512 struct dirrem *dirrem; 10513 struct diradd *dap; 10514 struct direct *ep; 10515 int i; 10516 10517 if (pagedep->pd_state & IOSTARTED) { 10518 /* 10519 * This can only happen if there is a driver that does not 10520 * understand chaining. Here biodone will reissue the call 10521 * to strategy for the incomplete buffers. 10522 */ 10523 printf("initiate_write_filepage: already started\n"); 10524 return; 10525 } 10526 pagedep->pd_state |= IOSTARTED; 10527 /* 10528 * Wait for all journal remove dependencies to hit the disk. 10529 * We can not allow any potentially conflicting directory adds 10530 * to be visible before removes and rollback is too difficult. 10531 * The per-filesystem lock may be dropped and re-acquired, however 10532 * we hold the buf locked so the dependency can not go away. 10533 */ 10534 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) 10535 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) 10536 jwait(&jremref->jr_list, MNT_WAIT); 10537 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) 10538 jwait(&jmvref->jm_list, MNT_WAIT); 10539 for (i = 0; i < DAHASHSZ; i++) { 10540 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 10541 ep = (struct direct *) 10542 ((char *)bp->b_data + dap->da_offset); 10543 if (ep->d_ino != dap->da_newinum) 10544 panic("%s: dir inum %ju != new %ju", 10545 "initiate_write_filepage", 10546 (uintmax_t)ep->d_ino, 10547 (uintmax_t)dap->da_newinum); 10548 if (dap->da_state & DIRCHG) 10549 ep->d_ino = dap->da_previous->dm_oldinum; 10550 else 10551 ep->d_ino = 0; 10552 dap->da_state &= ~ATTACHED; 10553 dap->da_state |= UNDONE; 10554 } 10555 } 10556} 10557 10558/* 10559 * Version of initiate_write_inodeblock that handles UFS1 dinodes. 10560 * Note that any bug fixes made to this routine must be done in the 10561 * version found below. 10562 * 10563 * Called from within the procedure above to deal with unsatisfied 10564 * allocation dependencies in an inodeblock. The buffer must be 10565 * locked, thus, no I/O completion operations can occur while we 10566 * are manipulating its associated dependencies. 10567 */ 10568static void 10569initiate_write_inodeblock_ufs1(inodedep, bp) 10570 struct inodedep *inodedep; 10571 struct buf *bp; /* The inode block */ 10572{ 10573 struct allocdirect *adp, *lastadp; 10574 struct ufs1_dinode *dp; 10575 struct ufs1_dinode *sip; 10576 struct inoref *inoref; 10577 struct ufsmount *ump; 10578 struct fs *fs; 10579 ufs_lbn_t i; 10580#ifdef INVARIANTS 10581 ufs_lbn_t prevlbn = 0; 10582#endif 10583 int deplist; 10584 10585 if (inodedep->id_state & IOSTARTED) 10586 panic("initiate_write_inodeblock_ufs1: already started"); 10587 inodedep->id_state |= IOSTARTED; 10588 fs = inodedep->id_fs; 10589 ump = VFSTOUFS(inodedep->id_list.wk_mp); 10590 LOCK_OWNED(ump); 10591 dp = (struct ufs1_dinode *)bp->b_data + 10592 ino_to_fsbo(fs, inodedep->id_ino); 10593 10594 /* 10595 * If we're on the unlinked list but have not yet written our 10596 * next pointer initialize it here. 10597 */ 10598 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 10599 struct inodedep *inon; 10600 10601 inon = TAILQ_NEXT(inodedep, id_unlinked); 10602 dp->di_freelink = inon ? inon->id_ino : 0; 10603 } 10604 /* 10605 * If the bitmap is not yet written, then the allocated 10606 * inode cannot be written to disk. 10607 */ 10608 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 10609 if (inodedep->id_savedino1 != NULL) 10610 panic("initiate_write_inodeblock_ufs1: I/O underway"); 10611 FREE_LOCK(ump); 10612 sip = malloc(sizeof(struct ufs1_dinode), 10613 M_SAVEDINO, M_SOFTDEP_FLAGS); 10614 ACQUIRE_LOCK(ump); 10615 inodedep->id_savedino1 = sip; 10616 *inodedep->id_savedino1 = *dp; 10617 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 10618 dp->di_gen = inodedep->id_savedino1->di_gen; 10619 dp->di_freelink = inodedep->id_savedino1->di_freelink; 10620 return; 10621 } 10622 /* 10623 * If no dependencies, then there is nothing to roll back. 10624 */ 10625 inodedep->id_savedsize = dp->di_size; 10626 inodedep->id_savedextsize = 0; 10627 inodedep->id_savednlink = dp->di_nlink; 10628 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 10629 TAILQ_EMPTY(&inodedep->id_inoreflst)) 10630 return; 10631 /* 10632 * Revert the link count to that of the first unwritten journal entry. 10633 */ 10634 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 10635 if (inoref) 10636 dp->di_nlink = inoref->if_nlink; 10637 /* 10638 * Set the dependencies to busy. 10639 */ 10640 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10641 adp = TAILQ_NEXT(adp, ad_next)) { 10642#ifdef INVARIANTS 10643 if (deplist != 0 && prevlbn >= adp->ad_offset) 10644 panic("softdep_write_inodeblock: lbn order"); 10645 prevlbn = adp->ad_offset; 10646 if (adp->ad_offset < UFS_NDADDR && 10647 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 10648 panic("initiate_write_inodeblock_ufs1: " 10649 "direct pointer #%jd mismatch %d != %jd", 10650 (intmax_t)adp->ad_offset, 10651 dp->di_db[adp->ad_offset], 10652 (intmax_t)adp->ad_newblkno); 10653 if (adp->ad_offset >= UFS_NDADDR && 10654 dp->di_ib[adp->ad_offset - UFS_NDADDR] != adp->ad_newblkno) 10655 panic("initiate_write_inodeblock_ufs1: " 10656 "indirect pointer #%jd mismatch %d != %jd", 10657 (intmax_t)adp->ad_offset - UFS_NDADDR, 10658 dp->di_ib[adp->ad_offset - UFS_NDADDR], 10659 (intmax_t)adp->ad_newblkno); 10660 deplist |= 1 << adp->ad_offset; 10661 if ((adp->ad_state & ATTACHED) == 0) 10662 panic("initiate_write_inodeblock_ufs1: " 10663 "Unknown state 0x%x", adp->ad_state); 10664#endif /* INVARIANTS */ 10665 adp->ad_state &= ~ATTACHED; 10666 adp->ad_state |= UNDONE; 10667 } 10668 /* 10669 * The on-disk inode cannot claim to be any larger than the last 10670 * fragment that has been written. Otherwise, the on-disk inode 10671 * might have fragments that were not the last block in the file 10672 * which would corrupt the filesystem. 10673 */ 10674 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10675 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10676 if (adp->ad_offset >= UFS_NDADDR) 10677 break; 10678 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 10679 /* keep going until hitting a rollback to a frag */ 10680 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10681 continue; 10682 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10683 for (i = adp->ad_offset + 1; i < UFS_NDADDR; i++) { 10684#ifdef INVARIANTS 10685 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 10686 panic("initiate_write_inodeblock_ufs1: " 10687 "lost dep1"); 10688#endif /* INVARIANTS */ 10689 dp->di_db[i] = 0; 10690 } 10691 for (i = 0; i < UFS_NIADDR; i++) { 10692#ifdef INVARIANTS 10693 if (dp->di_ib[i] != 0 && 10694 (deplist & ((1 << UFS_NDADDR) << i)) == 0) 10695 panic("initiate_write_inodeblock_ufs1: " 10696 "lost dep2"); 10697#endif /* INVARIANTS */ 10698 dp->di_ib[i] = 0; 10699 } 10700 return; 10701 } 10702 /* 10703 * If we have zero'ed out the last allocated block of the file, 10704 * roll back the size to the last currently allocated block. 10705 * We know that this last allocated block is a full-sized as 10706 * we already checked for fragments in the loop above. 10707 */ 10708 if (lastadp != NULL && 10709 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10710 for (i = lastadp->ad_offset; i >= 0; i--) 10711 if (dp->di_db[i] != 0) 10712 break; 10713 dp->di_size = (i + 1) * fs->fs_bsize; 10714 } 10715 /* 10716 * The only dependencies are for indirect blocks. 10717 * 10718 * The file size for indirect block additions is not guaranteed. 10719 * Such a guarantee would be non-trivial to achieve. The conventional 10720 * synchronous write implementation also does not make this guarantee. 10721 * Fsck should catch and fix discrepancies. Arguably, the file size 10722 * can be over-estimated without destroying integrity when the file 10723 * moves into the indirect blocks (i.e., is large). If we want to 10724 * postpone fsck, we are stuck with this argument. 10725 */ 10726 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 10727 dp->di_ib[adp->ad_offset - UFS_NDADDR] = 0; 10728} 10729 10730/* 10731 * Version of initiate_write_inodeblock that handles UFS2 dinodes. 10732 * Note that any bug fixes made to this routine must be done in the 10733 * version found above. 10734 * 10735 * Called from within the procedure above to deal with unsatisfied 10736 * allocation dependencies in an inodeblock. The buffer must be 10737 * locked, thus, no I/O completion operations can occur while we 10738 * are manipulating its associated dependencies. 10739 */ 10740static void 10741initiate_write_inodeblock_ufs2(inodedep, bp) 10742 struct inodedep *inodedep; 10743 struct buf *bp; /* The inode block */ 10744{ 10745 struct allocdirect *adp, *lastadp; 10746 struct ufs2_dinode *dp; 10747 struct ufs2_dinode *sip; 10748 struct inoref *inoref; 10749 struct ufsmount *ump; 10750 struct fs *fs; 10751 ufs_lbn_t i; 10752#ifdef INVARIANTS 10753 ufs_lbn_t prevlbn = 0; 10754#endif 10755 int deplist; 10756 10757 if (inodedep->id_state & IOSTARTED) 10758 panic("initiate_write_inodeblock_ufs2: already started"); 10759 inodedep->id_state |= IOSTARTED; 10760 fs = inodedep->id_fs; 10761 ump = VFSTOUFS(inodedep->id_list.wk_mp); 10762 LOCK_OWNED(ump); 10763 dp = (struct ufs2_dinode *)bp->b_data + 10764 ino_to_fsbo(fs, inodedep->id_ino); 10765 10766 /* 10767 * If we're on the unlinked list but have not yet written our 10768 * next pointer initialize it here. 10769 */ 10770 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 10771 struct inodedep *inon; 10772 10773 inon = TAILQ_NEXT(inodedep, id_unlinked); 10774 dp->di_freelink = inon ? inon->id_ino : 0; 10775 ffs_update_dinode_ckhash(fs, dp); 10776 } 10777 /* 10778 * If the bitmap is not yet written, then the allocated 10779 * inode cannot be written to disk. 10780 */ 10781 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 10782 if (inodedep->id_savedino2 != NULL) 10783 panic("initiate_write_inodeblock_ufs2: I/O underway"); 10784 FREE_LOCK(ump); 10785 sip = malloc(sizeof(struct ufs2_dinode), 10786 M_SAVEDINO, M_SOFTDEP_FLAGS); 10787 ACQUIRE_LOCK(ump); 10788 inodedep->id_savedino2 = sip; 10789 *inodedep->id_savedino2 = *dp; 10790 bzero((caddr_t)dp, sizeof(struct ufs2_dinode)); 10791 dp->di_gen = inodedep->id_savedino2->di_gen; 10792 dp->di_freelink = inodedep->id_savedino2->di_freelink; 10793 return; 10794 } 10795 /* 10796 * If no dependencies, then there is nothing to roll back. 10797 */ 10798 inodedep->id_savedsize = dp->di_size; 10799 inodedep->id_savedextsize = dp->di_extsize; 10800 inodedep->id_savednlink = dp->di_nlink; 10801 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 10802 TAILQ_EMPTY(&inodedep->id_extupdt) && 10803 TAILQ_EMPTY(&inodedep->id_inoreflst)) 10804 return; 10805 /* 10806 * Revert the link count to that of the first unwritten journal entry. 10807 */ 10808 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 10809 if (inoref) 10810 dp->di_nlink = inoref->if_nlink; 10811 10812 /* 10813 * Set the ext data dependencies to busy. 10814 */ 10815 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 10816 adp = TAILQ_NEXT(adp, ad_next)) { 10817#ifdef INVARIANTS 10818 if (deplist != 0 && prevlbn >= adp->ad_offset) 10819 panic("initiate_write_inodeblock_ufs2: lbn order"); 10820 prevlbn = adp->ad_offset; 10821 if (dp->di_extb[adp->ad_offset] != adp->ad_newblkno) 10822 panic("initiate_write_inodeblock_ufs2: " 10823 "ext pointer #%jd mismatch %jd != %jd", 10824 (intmax_t)adp->ad_offset, 10825 (intmax_t)dp->di_extb[adp->ad_offset], 10826 (intmax_t)adp->ad_newblkno); 10827 deplist |= 1 << adp->ad_offset; 10828 if ((adp->ad_state & ATTACHED) == 0) 10829 panic("initiate_write_inodeblock_ufs2: Unknown " 10830 "state 0x%x", adp->ad_state); 10831#endif /* INVARIANTS */ 10832 adp->ad_state &= ~ATTACHED; 10833 adp->ad_state |= UNDONE; 10834 } 10835 /* 10836 * The on-disk inode cannot claim to be any larger than the last 10837 * fragment that has been written. Otherwise, the on-disk inode 10838 * might have fragments that were not the last block in the ext 10839 * data which would corrupt the filesystem. 10840 */ 10841 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 10842 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10843 dp->di_extb[adp->ad_offset] = adp->ad_oldblkno; 10844 /* keep going until hitting a rollback to a frag */ 10845 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10846 continue; 10847 dp->di_extsize = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10848 for (i = adp->ad_offset + 1; i < UFS_NXADDR; i++) { 10849#ifdef INVARIANTS 10850 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) 10851 panic("initiate_write_inodeblock_ufs2: " 10852 "lost dep1"); 10853#endif /* INVARIANTS */ 10854 dp->di_extb[i] = 0; 10855 } 10856 lastadp = NULL; 10857 break; 10858 } 10859 /* 10860 * If we have zero'ed out the last allocated block of the ext 10861 * data, roll back the size to the last currently allocated block. 10862 * We know that this last allocated block is a full-sized as 10863 * we already checked for fragments in the loop above. 10864 */ 10865 if (lastadp != NULL && 10866 dp->di_extsize <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10867 for (i = lastadp->ad_offset; i >= 0; i--) 10868 if (dp->di_extb[i] != 0) 10869 break; 10870 dp->di_extsize = (i + 1) * fs->fs_bsize; 10871 } 10872 /* 10873 * Set the file data dependencies to busy. 10874 */ 10875 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10876 adp = TAILQ_NEXT(adp, ad_next)) { 10877#ifdef INVARIANTS 10878 if (deplist != 0 && prevlbn >= adp->ad_offset) 10879 panic("softdep_write_inodeblock: lbn order"); 10880 if ((adp->ad_state & ATTACHED) == 0) 10881 panic("inodedep %p and adp %p not attached", inodedep, adp); 10882 prevlbn = adp->ad_offset; 10883 if (!ffs_fsfail_cleanup(ump, 0) && 10884 adp->ad_offset < UFS_NDADDR && 10885 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 10886 panic("initiate_write_inodeblock_ufs2: " 10887 "direct pointer #%jd mismatch %jd != %jd", 10888 (intmax_t)adp->ad_offset, 10889 (intmax_t)dp->di_db[adp->ad_offset], 10890 (intmax_t)adp->ad_newblkno); 10891 if (!ffs_fsfail_cleanup(ump, 0) && 10892 adp->ad_offset >= UFS_NDADDR && 10893 dp->di_ib[adp->ad_offset - UFS_NDADDR] != adp->ad_newblkno) 10894 panic("initiate_write_inodeblock_ufs2: " 10895 "indirect pointer #%jd mismatch %jd != %jd", 10896 (intmax_t)adp->ad_offset - UFS_NDADDR, 10897 (intmax_t)dp->di_ib[adp->ad_offset - UFS_NDADDR], 10898 (intmax_t)adp->ad_newblkno); 10899 deplist |= 1 << adp->ad_offset; 10900 if ((adp->ad_state & ATTACHED) == 0) 10901 panic("initiate_write_inodeblock_ufs2: Unknown " 10902 "state 0x%x", adp->ad_state); 10903#endif /* INVARIANTS */ 10904 adp->ad_state &= ~ATTACHED; 10905 adp->ad_state |= UNDONE; 10906 } 10907 /* 10908 * The on-disk inode cannot claim to be any larger than the last 10909 * fragment that has been written. Otherwise, the on-disk inode 10910 * might have fragments that were not the last block in the file 10911 * which would corrupt the filesystem. 10912 */ 10913 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10914 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10915 if (adp->ad_offset >= UFS_NDADDR) 10916 break; 10917 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 10918 /* keep going until hitting a rollback to a frag */ 10919 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10920 continue; 10921 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10922 for (i = adp->ad_offset + 1; i < UFS_NDADDR; i++) { 10923#ifdef INVARIANTS 10924 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 10925 panic("initiate_write_inodeblock_ufs2: " 10926 "lost dep2"); 10927#endif /* INVARIANTS */ 10928 dp->di_db[i] = 0; 10929 } 10930 for (i = 0; i < UFS_NIADDR; i++) { 10931#ifdef INVARIANTS 10932 if (dp->di_ib[i] != 0 && 10933 (deplist & ((1 << UFS_NDADDR) << i)) == 0) 10934 panic("initiate_write_inodeblock_ufs2: " 10935 "lost dep3"); 10936#endif /* INVARIANTS */ 10937 dp->di_ib[i] = 0; 10938 } 10939 ffs_update_dinode_ckhash(fs, dp); 10940 return; 10941 } 10942 /* 10943 * If we have zero'ed out the last allocated block of the file, 10944 * roll back the size to the last currently allocated block. 10945 * We know that this last allocated block is a full-sized as 10946 * we already checked for fragments in the loop above. 10947 */ 10948 if (lastadp != NULL && 10949 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10950 for (i = lastadp->ad_offset; i >= 0; i--) 10951 if (dp->di_db[i] != 0) 10952 break; 10953 dp->di_size = (i + 1) * fs->fs_bsize; 10954 } 10955 /* 10956 * The only dependencies are for indirect blocks. 10957 * 10958 * The file size for indirect block additions is not guaranteed. 10959 * Such a guarantee would be non-trivial to achieve. The conventional 10960 * synchronous write implementation also does not make this guarantee. 10961 * Fsck should catch and fix discrepancies. Arguably, the file size 10962 * can be over-estimated without destroying integrity when the file 10963 * moves into the indirect blocks (i.e., is large). If we want to 10964 * postpone fsck, we are stuck with this argument. 10965 */ 10966 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 10967 dp->di_ib[adp->ad_offset - UFS_NDADDR] = 0; 10968 ffs_update_dinode_ckhash(fs, dp); 10969} 10970 10971/* 10972 * Cancel an indirdep as a result of truncation. Release all of the 10973 * children allocindirs and place their journal work on the appropriate 10974 * list. 10975 */ 10976static void 10977cancel_indirdep(indirdep, bp, freeblks) 10978 struct indirdep *indirdep; 10979 struct buf *bp; 10980 struct freeblks *freeblks; 10981{ 10982 struct allocindir *aip; 10983 10984 /* 10985 * None of the indirect pointers will ever be visible, 10986 * so they can simply be tossed. GOINGAWAY ensures 10987 * that allocated pointers will be saved in the buffer 10988 * cache until they are freed. Note that they will 10989 * only be able to be found by their physical address 10990 * since the inode mapping the logical address will 10991 * be gone. The save buffer used for the safe copy 10992 * was allocated in setup_allocindir_phase2 using 10993 * the physical address so it could be used for this 10994 * purpose. Hence we swap the safe copy with the real 10995 * copy, allowing the safe copy to be freed and holding 10996 * on to the real copy for later use in indir_trunc. 10997 */ 10998 if (indirdep->ir_state & GOINGAWAY) 10999 panic("cancel_indirdep: already gone"); 11000 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 11001 indirdep->ir_state |= DEPCOMPLETE; 11002 LIST_REMOVE(indirdep, ir_next); 11003 } 11004 indirdep->ir_state |= GOINGAWAY; 11005 /* 11006 * Pass in bp for blocks still have journal writes 11007 * pending so we can cancel them on their own. 11008 */ 11009 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != NULL) 11010 cancel_allocindir(aip, bp, freeblks, 0); 11011 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL) 11012 cancel_allocindir(aip, NULL, freeblks, 0); 11013 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL) 11014 cancel_allocindir(aip, NULL, freeblks, 0); 11015 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) 11016 cancel_allocindir(aip, NULL, freeblks, 0); 11017 /* 11018 * If there are pending partial truncations we need to keep the 11019 * old block copy around until they complete. This is because 11020 * the current b_data is not a perfect superset of the available 11021 * blocks. 11022 */ 11023 if (TAILQ_EMPTY(&indirdep->ir_trunc)) 11024 bcopy(bp->b_data, indirdep->ir_savebp->b_data, bp->b_bcount); 11025 else 11026 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 11027 WORKLIST_REMOVE(&indirdep->ir_list); 11028 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, &indirdep->ir_list); 11029 indirdep->ir_bp = NULL; 11030 indirdep->ir_freeblks = freeblks; 11031} 11032 11033/* 11034 * Free an indirdep once it no longer has new pointers to track. 11035 */ 11036static void 11037free_indirdep(indirdep) 11038 struct indirdep *indirdep; 11039{ 11040 11041 KASSERT(TAILQ_EMPTY(&indirdep->ir_trunc), 11042 ("free_indirdep: Indir trunc list not empty.")); 11043 KASSERT(LIST_EMPTY(&indirdep->ir_completehd), 11044 ("free_indirdep: Complete head not empty.")); 11045 KASSERT(LIST_EMPTY(&indirdep->ir_writehd), 11046 ("free_indirdep: write head not empty.")); 11047 KASSERT(LIST_EMPTY(&indirdep->ir_donehd), 11048 ("free_indirdep: done head not empty.")); 11049 KASSERT(LIST_EMPTY(&indirdep->ir_deplisthd), 11050 ("free_indirdep: deplist head not empty.")); 11051 KASSERT((indirdep->ir_state & DEPCOMPLETE), 11052 ("free_indirdep: %p still on newblk list.", indirdep)); 11053 KASSERT(indirdep->ir_saveddata == NULL, 11054 ("free_indirdep: %p still has saved data.", indirdep)); 11055 KASSERT(indirdep->ir_savebp == NULL, 11056 ("free_indirdep: %p still has savebp buffer.", indirdep)); 11057 if (indirdep->ir_state & ONWORKLIST) 11058 WORKLIST_REMOVE(&indirdep->ir_list); 11059 WORKITEM_FREE(indirdep, D_INDIRDEP); 11060} 11061 11062/* 11063 * Called before a write to an indirdep. This routine is responsible for 11064 * rolling back pointers to a safe state which includes only those 11065 * allocindirs which have been completed. 11066 */ 11067static void 11068initiate_write_indirdep(indirdep, bp) 11069 struct indirdep *indirdep; 11070 struct buf *bp; 11071{ 11072 struct ufsmount *ump; 11073 11074 indirdep->ir_state |= IOSTARTED; 11075 if (indirdep->ir_state & GOINGAWAY) 11076 panic("disk_io_initiation: indirdep gone"); 11077 /* 11078 * If there are no remaining dependencies, this will be writing 11079 * the real pointers. 11080 */ 11081 if (LIST_EMPTY(&indirdep->ir_deplisthd) && 11082 TAILQ_EMPTY(&indirdep->ir_trunc)) 11083 return; 11084 /* 11085 * Replace up-to-date version with safe version. 11086 */ 11087 if (indirdep->ir_saveddata == NULL) { 11088 ump = VFSTOUFS(indirdep->ir_list.wk_mp); 11089 LOCK_OWNED(ump); 11090 FREE_LOCK(ump); 11091 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 11092 M_SOFTDEP_FLAGS); 11093 ACQUIRE_LOCK(ump); 11094 } 11095 indirdep->ir_state &= ~ATTACHED; 11096 indirdep->ir_state |= UNDONE; 11097 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 11098 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 11099 bp->b_bcount); 11100} 11101 11102/* 11103 * Called when an inode has been cleared in a cg bitmap. This finally 11104 * eliminates any canceled jaddrefs 11105 */ 11106void 11107softdep_setup_inofree(mp, bp, ino, wkhd) 11108 struct mount *mp; 11109 struct buf *bp; 11110 ino_t ino; 11111 struct workhead *wkhd; 11112{ 11113 struct worklist *wk, *wkn; 11114 struct inodedep *inodedep; 11115 struct ufsmount *ump; 11116 uint8_t *inosused; 11117 struct cg *cgp; 11118 struct fs *fs; 11119 11120 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 11121 ("softdep_setup_inofree called on non-softdep filesystem")); 11122 ump = VFSTOUFS(mp); 11123 ACQUIRE_LOCK(ump); 11124 if (!ffs_fsfail_cleanup(ump, 0)) { 11125 fs = ump->um_fs; 11126 cgp = (struct cg *)bp->b_data; 11127 inosused = cg_inosused(cgp); 11128 if (isset(inosused, ino % fs->fs_ipg)) 11129 panic("softdep_setup_inofree: inode %ju not freed.", 11130 (uintmax_t)ino); 11131 } 11132 if (inodedep_lookup(mp, ino, 0, &inodedep)) 11133 panic("softdep_setup_inofree: ino %ju has existing inodedep %p", 11134 (uintmax_t)ino, inodedep); 11135 if (wkhd) { 11136 LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) { 11137 if (wk->wk_type != D_JADDREF) 11138 continue; 11139 WORKLIST_REMOVE(wk); 11140 /* 11141 * We can free immediately even if the jaddref 11142 * isn't attached in a background write as now 11143 * the bitmaps are reconciled. 11144 */ 11145 wk->wk_state |= COMPLETE | ATTACHED; 11146 free_jaddref(WK_JADDREF(wk)); 11147 } 11148 jwork_move(&bp->b_dep, wkhd); 11149 } 11150 FREE_LOCK(ump); 11151} 11152 11153/* 11154 * Called via ffs_blkfree() after a set of frags has been cleared from a cg 11155 * map. Any dependencies waiting for the write to clear are added to the 11156 * buf's list and any jnewblks that are being canceled are discarded 11157 * immediately. 11158 */ 11159void 11160softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 11161 struct mount *mp; 11162 struct buf *bp; 11163 ufs2_daddr_t blkno; 11164 int frags; 11165 struct workhead *wkhd; 11166{ 11167 struct bmsafemap *bmsafemap; 11168 struct jnewblk *jnewblk; 11169 struct ufsmount *ump; 11170 struct worklist *wk; 11171 struct fs *fs; 11172#ifdef INVARIANTS 11173 uint8_t *blksfree; 11174 struct cg *cgp; 11175 ufs2_daddr_t jstart; 11176 ufs2_daddr_t jend; 11177 ufs2_daddr_t end; 11178 long bno; 11179 int i; 11180#endif 11181 11182 CTR3(KTR_SUJ, 11183 "softdep_setup_blkfree: blkno %jd frags %d wk head %p", 11184 blkno, frags, wkhd); 11185 11186 ump = VFSTOUFS(mp); 11187 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 11188 ("softdep_setup_blkfree called on non-softdep filesystem")); 11189 ACQUIRE_LOCK(ump); 11190 /* Lookup the bmsafemap so we track when it is dirty. */ 11191 fs = ump->um_fs; 11192 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 11193 /* 11194 * Detach any jnewblks which have been canceled. They must linger 11195 * until the bitmap is cleared again by ffs_blkfree() to prevent 11196 * an unjournaled allocation from hitting the disk. 11197 */ 11198 if (wkhd) { 11199 while ((wk = LIST_FIRST(wkhd)) != NULL) { 11200 CTR2(KTR_SUJ, 11201 "softdep_setup_blkfree: blkno %jd wk type %d", 11202 blkno, wk->wk_type); 11203 WORKLIST_REMOVE(wk); 11204 if (wk->wk_type != D_JNEWBLK) { 11205 WORKLIST_INSERT(&bmsafemap->sm_freehd, wk); 11206 continue; 11207 } 11208 jnewblk = WK_JNEWBLK(wk); 11209 KASSERT(jnewblk->jn_state & GOINGAWAY, 11210 ("softdep_setup_blkfree: jnewblk not canceled.")); 11211#ifdef INVARIANTS 11212 /* 11213 * Assert that this block is free in the bitmap 11214 * before we discard the jnewblk. 11215 */ 11216 cgp = (struct cg *)bp->b_data; 11217 blksfree = cg_blksfree(cgp); 11218 bno = dtogd(fs, jnewblk->jn_blkno); 11219 for (i = jnewblk->jn_oldfrags; 11220 i < jnewblk->jn_frags; i++) { 11221 if (isset(blksfree, bno + i)) 11222 continue; 11223 panic("softdep_setup_blkfree: not free"); 11224 } 11225#endif 11226 /* 11227 * Even if it's not attached we can free immediately 11228 * as the new bitmap is correct. 11229 */ 11230 wk->wk_state |= COMPLETE | ATTACHED; 11231 free_jnewblk(jnewblk); 11232 } 11233 } 11234 11235#ifdef INVARIANTS 11236 /* 11237 * Assert that we are not freeing a block which has an outstanding 11238 * allocation dependency. 11239 */ 11240 fs = VFSTOUFS(mp)->um_fs; 11241 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 11242 end = blkno + frags; 11243 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 11244 /* 11245 * Don't match against blocks that will be freed when the 11246 * background write is done. 11247 */ 11248 if ((jnewblk->jn_state & (ATTACHED | COMPLETE | DEPCOMPLETE)) == 11249 (COMPLETE | DEPCOMPLETE)) 11250 continue; 11251 jstart = jnewblk->jn_blkno + jnewblk->jn_oldfrags; 11252 jend = jnewblk->jn_blkno + jnewblk->jn_frags; 11253 if ((blkno >= jstart && blkno < jend) || 11254 (end > jstart && end <= jend)) { 11255 printf("state 0x%X %jd - %d %d dep %p\n", 11256 jnewblk->jn_state, jnewblk->jn_blkno, 11257 jnewblk->jn_oldfrags, jnewblk->jn_frags, 11258 jnewblk->jn_dep); 11259 panic("softdep_setup_blkfree: " 11260 "%jd-%jd(%d) overlaps with %jd-%jd", 11261 blkno, end, frags, jstart, jend); 11262 } 11263 } 11264#endif 11265 FREE_LOCK(ump); 11266} 11267 11268/* 11269 * Revert a block allocation when the journal record that describes it 11270 * is not yet written. 11271 */ 11272static int 11273jnewblk_rollback(jnewblk, fs, cgp, blksfree) 11274 struct jnewblk *jnewblk; 11275 struct fs *fs; 11276 struct cg *cgp; 11277 uint8_t *blksfree; 11278{ 11279 ufs1_daddr_t fragno; 11280 long cgbno, bbase; 11281 int frags, blk; 11282 int i; 11283 11284 frags = 0; 11285 cgbno = dtogd(fs, jnewblk->jn_blkno); 11286 /* 11287 * We have to test which frags need to be rolled back. We may 11288 * be operating on a stale copy when doing background writes. 11289 */ 11290 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) 11291 if (isclr(blksfree, cgbno + i)) 11292 frags++; 11293 if (frags == 0) 11294 return (0); 11295 /* 11296 * This is mostly ffs_blkfree() sans some validation and 11297 * superblock updates. 11298 */ 11299 if (frags == fs->fs_frag) { 11300 fragno = fragstoblks(fs, cgbno); 11301 ffs_setblock(fs, blksfree, fragno); 11302 ffs_clusteracct(fs, cgp, fragno, 1); 11303 cgp->cg_cs.cs_nbfree++; 11304 } else { 11305 cgbno += jnewblk->jn_oldfrags; 11306 bbase = cgbno - fragnum(fs, cgbno); 11307 /* Decrement the old frags. */ 11308 blk = blkmap(fs, blksfree, bbase); 11309 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 11310 /* Deallocate the fragment */ 11311 for (i = 0; i < frags; i++) 11312 setbit(blksfree, cgbno + i); 11313 cgp->cg_cs.cs_nffree += frags; 11314 /* Add back in counts associated with the new frags */ 11315 blk = blkmap(fs, blksfree, bbase); 11316 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 11317 /* If a complete block has been reassembled, account for it. */ 11318 fragno = fragstoblks(fs, bbase); 11319 if (ffs_isblock(fs, blksfree, fragno)) { 11320 cgp->cg_cs.cs_nffree -= fs->fs_frag; 11321 ffs_clusteracct(fs, cgp, fragno, 1); 11322 cgp->cg_cs.cs_nbfree++; 11323 } 11324 } 11325 stat_jnewblk++; 11326 jnewblk->jn_state &= ~ATTACHED; 11327 jnewblk->jn_state |= UNDONE; 11328 11329 return (frags); 11330} 11331 11332static void 11333initiate_write_bmsafemap(bmsafemap, bp) 11334 struct bmsafemap *bmsafemap; 11335 struct buf *bp; /* The cg block. */ 11336{ 11337 struct jaddref *jaddref; 11338 struct jnewblk *jnewblk; 11339 uint8_t *inosused; 11340 uint8_t *blksfree; 11341 struct cg *cgp; 11342 struct fs *fs; 11343 ino_t ino; 11344 11345 /* 11346 * If this is a background write, we did this at the time that 11347 * the copy was made, so do not need to do it again. 11348 */ 11349 if (bmsafemap->sm_state & IOSTARTED) 11350 return; 11351 bmsafemap->sm_state |= IOSTARTED; 11352 /* 11353 * Clear any inode allocations which are pending journal writes. 11354 */ 11355 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd) != NULL) { 11356 cgp = (struct cg *)bp->b_data; 11357 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11358 inosused = cg_inosused(cgp); 11359 LIST_FOREACH(jaddref, &bmsafemap->sm_jaddrefhd, ja_bmdeps) { 11360 ino = jaddref->ja_ino % fs->fs_ipg; 11361 if (isset(inosused, ino)) { 11362 if ((jaddref->ja_mode & IFMT) == IFDIR) 11363 cgp->cg_cs.cs_ndir--; 11364 cgp->cg_cs.cs_nifree++; 11365 clrbit(inosused, ino); 11366 jaddref->ja_state &= ~ATTACHED; 11367 jaddref->ja_state |= UNDONE; 11368 stat_jaddref++; 11369 } else 11370 panic("initiate_write_bmsafemap: inode %ju " 11371 "marked free", (uintmax_t)jaddref->ja_ino); 11372 } 11373 } 11374 /* 11375 * Clear any block allocations which are pending journal writes. 11376 */ 11377 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 11378 cgp = (struct cg *)bp->b_data; 11379 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11380 blksfree = cg_blksfree(cgp); 11381 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 11382 if (jnewblk_rollback(jnewblk, fs, cgp, blksfree)) 11383 continue; 11384 panic("initiate_write_bmsafemap: block %jd " 11385 "marked free", jnewblk->jn_blkno); 11386 } 11387 } 11388 /* 11389 * Move allocation lists to the written lists so they can be 11390 * cleared once the block write is complete. 11391 */ 11392 LIST_SWAP(&bmsafemap->sm_inodedephd, &bmsafemap->sm_inodedepwr, 11393 inodedep, id_deps); 11394 LIST_SWAP(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr, 11395 newblk, nb_deps); 11396 LIST_SWAP(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, worklist, 11397 wk_list); 11398} 11399 11400void 11401softdep_handle_error(struct buf *bp) 11402{ 11403 struct ufsmount *ump; 11404 11405 ump = softdep_bp_to_mp(bp); 11406 if (ump == NULL) 11407 return; 11408 11409 if (ffs_fsfail_cleanup(ump, bp->b_error)) { 11410 /* 11411 * No future writes will succeed, so the on-disk image is safe. 11412 * Pretend that this write succeeded so that the softdep state 11413 * will be cleaned up naturally. 11414 */ 11415 bp->b_ioflags &= ~BIO_ERROR; 11416 bp->b_error = 0; 11417 } 11418} 11419 11420/* 11421 * This routine is called during the completion interrupt 11422 * service routine for a disk write (from the procedure called 11423 * by the device driver to inform the filesystem caches of 11424 * a request completion). It should be called early in this 11425 * procedure, before the block is made available to other 11426 * processes or other routines are called. 11427 * 11428 */ 11429static void 11430softdep_disk_write_complete(bp) 11431 struct buf *bp; /* describes the completed disk write */ 11432{ 11433 struct worklist *wk; 11434 struct worklist *owk; 11435 struct ufsmount *ump; 11436 struct workhead reattach; 11437 struct freeblks *freeblks; 11438 struct buf *sbp; 11439 11440 ump = softdep_bp_to_mp(bp); 11441 KASSERT(LIST_EMPTY(&bp->b_dep) || ump != NULL, 11442 ("softdep_disk_write_complete: softdep_bp_to_mp returned NULL " 11443 "with outstanding dependencies for buffer %p", bp)); 11444 if (ump == NULL) 11445 return; 11446 if ((bp->b_ioflags & BIO_ERROR) != 0) 11447 softdep_handle_error(bp); 11448 /* 11449 * If an error occurred while doing the write, then the data 11450 * has not hit the disk and the dependencies cannot be processed. 11451 * But we do have to go through and roll forward any dependencies 11452 * that were rolled back before the disk write. 11453 */ 11454 sbp = NULL; 11455 ACQUIRE_LOCK(ump); 11456 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) { 11457 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 11458 switch (wk->wk_type) { 11459 case D_PAGEDEP: 11460 handle_written_filepage(WK_PAGEDEP(wk), bp, 0); 11461 continue; 11462 11463 case D_INODEDEP: 11464 handle_written_inodeblock(WK_INODEDEP(wk), 11465 bp, 0); 11466 continue; 11467 11468 case D_BMSAFEMAP: 11469 handle_written_bmsafemap(WK_BMSAFEMAP(wk), 11470 bp, 0); 11471 continue; 11472 11473 case D_INDIRDEP: 11474 handle_written_indirdep(WK_INDIRDEP(wk), 11475 bp, &sbp, 0); 11476 continue; 11477 default: 11478 /* nothing to roll forward */ 11479 continue; 11480 } 11481 } 11482 FREE_LOCK(ump); 11483 if (sbp) 11484 brelse(sbp); 11485 return; 11486 } 11487 LIST_INIT(&reattach); 11488 11489 /* 11490 * Ump SU lock must not be released anywhere in this code segment. 11491 */ 11492 owk = NULL; 11493 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 11494 WORKLIST_REMOVE(wk); 11495 atomic_add_long(&dep_write[wk->wk_type], 1); 11496 if (wk == owk) 11497 panic("duplicate worklist: %p\n", wk); 11498 owk = wk; 11499 switch (wk->wk_type) { 11500 case D_PAGEDEP: 11501 if (handle_written_filepage(WK_PAGEDEP(wk), bp, 11502 WRITESUCCEEDED)) 11503 WORKLIST_INSERT(&reattach, wk); 11504 continue; 11505 11506 case D_INODEDEP: 11507 if (handle_written_inodeblock(WK_INODEDEP(wk), bp, 11508 WRITESUCCEEDED)) 11509 WORKLIST_INSERT(&reattach, wk); 11510 continue; 11511 11512 case D_BMSAFEMAP: 11513 if (handle_written_bmsafemap(WK_BMSAFEMAP(wk), bp, 11514 WRITESUCCEEDED)) 11515 WORKLIST_INSERT(&reattach, wk); 11516 continue; 11517 11518 case D_MKDIR: 11519 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 11520 continue; 11521 11522 case D_ALLOCDIRECT: 11523 wk->wk_state |= COMPLETE; 11524 handle_allocdirect_partdone(WK_ALLOCDIRECT(wk), NULL); 11525 continue; 11526 11527 case D_ALLOCINDIR: 11528 wk->wk_state |= COMPLETE; 11529 handle_allocindir_partdone(WK_ALLOCINDIR(wk)); 11530 continue; 11531 11532 case D_INDIRDEP: 11533 if (handle_written_indirdep(WK_INDIRDEP(wk), bp, &sbp, 11534 WRITESUCCEEDED)) 11535 WORKLIST_INSERT(&reattach, wk); 11536 continue; 11537 11538 case D_FREEBLKS: 11539 wk->wk_state |= COMPLETE; 11540 freeblks = WK_FREEBLKS(wk); 11541 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE && 11542 LIST_EMPTY(&freeblks->fb_jblkdephd)) 11543 add_to_worklist(wk, WK_NODELAY); 11544 continue; 11545 11546 case D_FREEWORK: 11547 handle_written_freework(WK_FREEWORK(wk)); 11548 break; 11549 11550 case D_JSEGDEP: 11551 free_jsegdep(WK_JSEGDEP(wk)); 11552 continue; 11553 11554 case D_JSEG: 11555 handle_written_jseg(WK_JSEG(wk), bp); 11556 continue; 11557 11558 case D_SBDEP: 11559 if (handle_written_sbdep(WK_SBDEP(wk), bp)) 11560 WORKLIST_INSERT(&reattach, wk); 11561 continue; 11562 11563 case D_FREEDEP: 11564 free_freedep(WK_FREEDEP(wk)); 11565 continue; 11566 11567 default: 11568 panic("handle_disk_write_complete: Unknown type %s", 11569 TYPENAME(wk->wk_type)); 11570 /* NOTREACHED */ 11571 } 11572 } 11573 /* 11574 * Reattach any requests that must be redone. 11575 */ 11576 while ((wk = LIST_FIRST(&reattach)) != NULL) { 11577 WORKLIST_REMOVE(wk); 11578 WORKLIST_INSERT(&bp->b_dep, wk); 11579 } 11580 FREE_LOCK(ump); 11581 if (sbp) 11582 brelse(sbp); 11583} 11584 11585/* 11586 * Called from within softdep_disk_write_complete above. 11587 */ 11588static void 11589handle_allocdirect_partdone(adp, wkhd) 11590 struct allocdirect *adp; /* the completed allocdirect */ 11591 struct workhead *wkhd; /* Work to do when inode is writtne. */ 11592{ 11593 struct allocdirectlst *listhead; 11594 struct allocdirect *listadp; 11595 struct inodedep *inodedep; 11596 long bsize; 11597 11598 LOCK_OWNED(VFSTOUFS(adp->ad_block.nb_list.wk_mp)); 11599 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 11600 return; 11601 /* 11602 * The on-disk inode cannot claim to be any larger than the last 11603 * fragment that has been written. Otherwise, the on-disk inode 11604 * might have fragments that were not the last block in the file 11605 * which would corrupt the filesystem. Thus, we cannot free any 11606 * allocdirects after one whose ad_oldblkno claims a fragment as 11607 * these blocks must be rolled back to zero before writing the inode. 11608 * We check the currently active set of allocdirects in id_inoupdt 11609 * or id_extupdt as appropriate. 11610 */ 11611 inodedep = adp->ad_inodedep; 11612 bsize = inodedep->id_fs->fs_bsize; 11613 if (adp->ad_state & EXTDATA) 11614 listhead = &inodedep->id_extupdt; 11615 else 11616 listhead = &inodedep->id_inoupdt; 11617 TAILQ_FOREACH(listadp, listhead, ad_next) { 11618 /* found our block */ 11619 if (listadp == adp) 11620 break; 11621 /* continue if ad_oldlbn is not a fragment */ 11622 if (listadp->ad_oldsize == 0 || 11623 listadp->ad_oldsize == bsize) 11624 continue; 11625 /* hit a fragment */ 11626 return; 11627 } 11628 /* 11629 * If we have reached the end of the current list without 11630 * finding the just finished dependency, then it must be 11631 * on the future dependency list. Future dependencies cannot 11632 * be freed until they are moved to the current list. 11633 */ 11634 if (listadp == NULL) { 11635#ifdef INVARIANTS 11636 if (adp->ad_state & EXTDATA) 11637 listhead = &inodedep->id_newextupdt; 11638 else 11639 listhead = &inodedep->id_newinoupdt; 11640 TAILQ_FOREACH(listadp, listhead, ad_next) 11641 /* found our block */ 11642 if (listadp == adp) 11643 break; 11644 if (listadp == NULL) 11645 panic("handle_allocdirect_partdone: lost dep"); 11646#endif /* INVARIANTS */ 11647 return; 11648 } 11649 /* 11650 * If we have found the just finished dependency, then queue 11651 * it along with anything that follows it that is complete. 11652 * Since the pointer has not yet been written in the inode 11653 * as the dependency prevents it, place the allocdirect on the 11654 * bufwait list where it will be freed once the pointer is 11655 * valid. 11656 */ 11657 if (wkhd == NULL) 11658 wkhd = &inodedep->id_bufwait; 11659 for (; adp; adp = listadp) { 11660 listadp = TAILQ_NEXT(adp, ad_next); 11661 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 11662 return; 11663 TAILQ_REMOVE(listhead, adp, ad_next); 11664 WORKLIST_INSERT(wkhd, &adp->ad_block.nb_list); 11665 } 11666} 11667 11668/* 11669 * Called from within softdep_disk_write_complete above. This routine 11670 * completes successfully written allocindirs. 11671 */ 11672static void 11673handle_allocindir_partdone(aip) 11674 struct allocindir *aip; /* the completed allocindir */ 11675{ 11676 struct indirdep *indirdep; 11677 11678 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 11679 return; 11680 indirdep = aip->ai_indirdep; 11681 LIST_REMOVE(aip, ai_next); 11682 /* 11683 * Don't set a pointer while the buffer is undergoing IO or while 11684 * we have active truncations. 11685 */ 11686 if (indirdep->ir_state & UNDONE || !TAILQ_EMPTY(&indirdep->ir_trunc)) { 11687 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 11688 return; 11689 } 11690 if (indirdep->ir_state & UFS1FMT) 11691 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 11692 aip->ai_newblkno; 11693 else 11694 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 11695 aip->ai_newblkno; 11696 /* 11697 * Await the pointer write before freeing the allocindir. 11698 */ 11699 LIST_INSERT_HEAD(&indirdep->ir_writehd, aip, ai_next); 11700} 11701 11702/* 11703 * Release segments held on a jwork list. 11704 */ 11705static void 11706handle_jwork(wkhd) 11707 struct workhead *wkhd; 11708{ 11709 struct worklist *wk; 11710 11711 while ((wk = LIST_FIRST(wkhd)) != NULL) { 11712 WORKLIST_REMOVE(wk); 11713 switch (wk->wk_type) { 11714 case D_JSEGDEP: 11715 free_jsegdep(WK_JSEGDEP(wk)); 11716 continue; 11717 case D_FREEDEP: 11718 free_freedep(WK_FREEDEP(wk)); 11719 continue; 11720 case D_FREEFRAG: 11721 rele_jseg(WK_JSEG(WK_FREEFRAG(wk)->ff_jdep)); 11722 WORKITEM_FREE(wk, D_FREEFRAG); 11723 continue; 11724 case D_FREEWORK: 11725 handle_written_freework(WK_FREEWORK(wk)); 11726 continue; 11727 default: 11728 panic("handle_jwork: Unknown type %s\n", 11729 TYPENAME(wk->wk_type)); 11730 } 11731 } 11732} 11733 11734/* 11735 * Handle the bufwait list on an inode when it is safe to release items 11736 * held there. This normally happens after an inode block is written but 11737 * may be delayed and handled later if there are pending journal items that 11738 * are not yet safe to be released. 11739 */ 11740static struct freefile * 11741handle_bufwait(inodedep, refhd) 11742 struct inodedep *inodedep; 11743 struct workhead *refhd; 11744{ 11745 struct jaddref *jaddref; 11746 struct freefile *freefile; 11747 struct worklist *wk; 11748 11749 freefile = NULL; 11750 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 11751 WORKLIST_REMOVE(wk); 11752 switch (wk->wk_type) { 11753 case D_FREEFILE: 11754 /* 11755 * We defer adding freefile to the worklist 11756 * until all other additions have been made to 11757 * ensure that it will be done after all the 11758 * old blocks have been freed. 11759 */ 11760 if (freefile != NULL) 11761 panic("handle_bufwait: freefile"); 11762 freefile = WK_FREEFILE(wk); 11763 continue; 11764 11765 case D_MKDIR: 11766 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 11767 continue; 11768 11769 case D_DIRADD: 11770 diradd_inode_written(WK_DIRADD(wk), inodedep); 11771 continue; 11772 11773 case D_FREEFRAG: 11774 wk->wk_state |= COMPLETE; 11775 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE) 11776 add_to_worklist(wk, 0); 11777 continue; 11778 11779 case D_DIRREM: 11780 wk->wk_state |= COMPLETE; 11781 add_to_worklist(wk, 0); 11782 continue; 11783 11784 case D_ALLOCDIRECT: 11785 case D_ALLOCINDIR: 11786 free_newblk(WK_NEWBLK(wk)); 11787 continue; 11788 11789 case D_JNEWBLK: 11790 wk->wk_state |= COMPLETE; 11791 free_jnewblk(WK_JNEWBLK(wk)); 11792 continue; 11793 11794 /* 11795 * Save freed journal segments and add references on 11796 * the supplied list which will delay their release 11797 * until the cg bitmap is cleared on disk. 11798 */ 11799 case D_JSEGDEP: 11800 if (refhd == NULL) 11801 free_jsegdep(WK_JSEGDEP(wk)); 11802 else 11803 WORKLIST_INSERT(refhd, wk); 11804 continue; 11805 11806 case D_JADDREF: 11807 jaddref = WK_JADDREF(wk); 11808 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 11809 if_deps); 11810 /* 11811 * Transfer any jaddrefs to the list to be freed with 11812 * the bitmap if we're handling a removed file. 11813 */ 11814 if (refhd == NULL) { 11815 wk->wk_state |= COMPLETE; 11816 free_jaddref(jaddref); 11817 } else 11818 WORKLIST_INSERT(refhd, wk); 11819 continue; 11820 11821 default: 11822 panic("handle_bufwait: Unknown type %p(%s)", 11823 wk, TYPENAME(wk->wk_type)); 11824 /* NOTREACHED */ 11825 } 11826 } 11827 return (freefile); 11828} 11829/* 11830 * Called from within softdep_disk_write_complete above to restore 11831 * in-memory inode block contents to their most up-to-date state. Note 11832 * that this routine is always called from interrupt level with further 11833 * interrupts from this device blocked. 11834 * 11835 * If the write did not succeed, we will do all the roll-forward 11836 * operations, but we will not take the actions that will allow its 11837 * dependencies to be processed. 11838 */ 11839static int 11840handle_written_inodeblock(inodedep, bp, flags) 11841 struct inodedep *inodedep; 11842 struct buf *bp; /* buffer containing the inode block */ 11843 int flags; 11844{ 11845 struct freefile *freefile; 11846 struct allocdirect *adp, *nextadp; 11847 struct ufs1_dinode *dp1 = NULL; 11848 struct ufs2_dinode *dp2 = NULL; 11849 struct workhead wkhd; 11850 int hadchanges, fstype; 11851 ino_t freelink; 11852 11853 LIST_INIT(&wkhd); 11854 hadchanges = 0; 11855 freefile = NULL; 11856 if ((inodedep->id_state & IOSTARTED) == 0) 11857 panic("handle_written_inodeblock: not started"); 11858 inodedep->id_state &= ~IOSTARTED; 11859 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) { 11860 fstype = UFS1; 11861 dp1 = (struct ufs1_dinode *)bp->b_data + 11862 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 11863 freelink = dp1->di_freelink; 11864 } else { 11865 fstype = UFS2; 11866 dp2 = (struct ufs2_dinode *)bp->b_data + 11867 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 11868 freelink = dp2->di_freelink; 11869 } 11870 /* 11871 * Leave this inodeblock dirty until it's in the list. 11872 */ 11873 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) == UNLINKED && 11874 (flags & WRITESUCCEEDED)) { 11875 struct inodedep *inon; 11876 11877 inon = TAILQ_NEXT(inodedep, id_unlinked); 11878 if ((inon == NULL && freelink == 0) || 11879 (inon && inon->id_ino == freelink)) { 11880 if (inon) 11881 inon->id_state |= UNLINKPREV; 11882 inodedep->id_state |= UNLINKNEXT; 11883 } 11884 hadchanges = 1; 11885 } 11886 /* 11887 * If we had to rollback the inode allocation because of 11888 * bitmaps being incomplete, then simply restore it. 11889 * Keep the block dirty so that it will not be reclaimed until 11890 * all associated dependencies have been cleared and the 11891 * corresponding updates written to disk. 11892 */ 11893 if (inodedep->id_savedino1 != NULL) { 11894 hadchanges = 1; 11895 if (fstype == UFS1) 11896 *dp1 = *inodedep->id_savedino1; 11897 else 11898 *dp2 = *inodedep->id_savedino2; 11899 free(inodedep->id_savedino1, M_SAVEDINO); 11900 inodedep->id_savedino1 = NULL; 11901 if ((bp->b_flags & B_DELWRI) == 0) 11902 stat_inode_bitmap++; 11903 bdirty(bp); 11904 /* 11905 * If the inode is clear here and GOINGAWAY it will never 11906 * be written. Process the bufwait and clear any pending 11907 * work which may include the freefile. 11908 */ 11909 if (inodedep->id_state & GOINGAWAY) 11910 goto bufwait; 11911 return (1); 11912 } 11913 if (flags & WRITESUCCEEDED) 11914 inodedep->id_state |= COMPLETE; 11915 /* 11916 * Roll forward anything that had to be rolled back before 11917 * the inode could be updated. 11918 */ 11919 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 11920 nextadp = TAILQ_NEXT(adp, ad_next); 11921 if (adp->ad_state & ATTACHED) 11922 panic("handle_written_inodeblock: new entry"); 11923 if (fstype == UFS1) { 11924 if (adp->ad_offset < UFS_NDADDR) { 11925 if (dp1->di_db[adp->ad_offset]!=adp->ad_oldblkno) 11926 panic("%s %s #%jd mismatch %d != %jd", 11927 "handle_written_inodeblock:", 11928 "direct pointer", 11929 (intmax_t)adp->ad_offset, 11930 dp1->di_db[adp->ad_offset], 11931 (intmax_t)adp->ad_oldblkno); 11932 dp1->di_db[adp->ad_offset] = adp->ad_newblkno; 11933 } else { 11934 if (dp1->di_ib[adp->ad_offset - UFS_NDADDR] != 11935 0) 11936 panic("%s: %s #%jd allocated as %d", 11937 "handle_written_inodeblock", 11938 "indirect pointer", 11939 (intmax_t)adp->ad_offset - 11940 UFS_NDADDR, 11941 dp1->di_ib[adp->ad_offset - 11942 UFS_NDADDR]); 11943 dp1->di_ib[adp->ad_offset - UFS_NDADDR] = 11944 adp->ad_newblkno; 11945 } 11946 } else { 11947 if (adp->ad_offset < UFS_NDADDR) { 11948 if (dp2->di_db[adp->ad_offset]!=adp->ad_oldblkno) 11949 panic("%s: %s #%jd %s %jd != %jd", 11950 "handle_written_inodeblock", 11951 "direct pointer", 11952 (intmax_t)adp->ad_offset, "mismatch", 11953 (intmax_t)dp2->di_db[adp->ad_offset], 11954 (intmax_t)adp->ad_oldblkno); 11955 dp2->di_db[adp->ad_offset] = adp->ad_newblkno; 11956 } else { 11957 if (dp2->di_ib[adp->ad_offset - UFS_NDADDR] != 11958 0) 11959 panic("%s: %s #%jd allocated as %jd", 11960 "handle_written_inodeblock", 11961 "indirect pointer", 11962 (intmax_t)adp->ad_offset - 11963 UFS_NDADDR, 11964 (intmax_t) 11965 dp2->di_ib[adp->ad_offset - 11966 UFS_NDADDR]); 11967 dp2->di_ib[adp->ad_offset - UFS_NDADDR] = 11968 adp->ad_newblkno; 11969 } 11970 } 11971 adp->ad_state &= ~UNDONE; 11972 adp->ad_state |= ATTACHED; 11973 hadchanges = 1; 11974 } 11975 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) { 11976 nextadp = TAILQ_NEXT(adp, ad_next); 11977 if (adp->ad_state & ATTACHED) 11978 panic("handle_written_inodeblock: new entry"); 11979 if (dp2->di_extb[adp->ad_offset] != adp->ad_oldblkno) 11980 panic("%s: direct pointers #%jd %s %jd != %jd", 11981 "handle_written_inodeblock", 11982 (intmax_t)adp->ad_offset, "mismatch", 11983 (intmax_t)dp2->di_extb[adp->ad_offset], 11984 (intmax_t)adp->ad_oldblkno); 11985 dp2->di_extb[adp->ad_offset] = adp->ad_newblkno; 11986 adp->ad_state &= ~UNDONE; 11987 adp->ad_state |= ATTACHED; 11988 hadchanges = 1; 11989 } 11990 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 11991 stat_direct_blk_ptrs++; 11992 /* 11993 * Reset the file size to its most up-to-date value. 11994 */ 11995 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) 11996 panic("handle_written_inodeblock: bad size"); 11997 if (inodedep->id_savednlink > UFS_LINK_MAX) 11998 panic("handle_written_inodeblock: Invalid link count " 11999 "%jd for inodedep %p", (uintmax_t)inodedep->id_savednlink, 12000 inodedep); 12001 if (fstype == UFS1) { 12002 if (dp1->di_nlink != inodedep->id_savednlink) { 12003 dp1->di_nlink = inodedep->id_savednlink; 12004 hadchanges = 1; 12005 } 12006 if (dp1->di_size != inodedep->id_savedsize) { 12007 dp1->di_size = inodedep->id_savedsize; 12008 hadchanges = 1; 12009 } 12010 } else { 12011 if (dp2->di_nlink != inodedep->id_savednlink) { 12012 dp2->di_nlink = inodedep->id_savednlink; 12013 hadchanges = 1; 12014 } 12015 if (dp2->di_size != inodedep->id_savedsize) { 12016 dp2->di_size = inodedep->id_savedsize; 12017 hadchanges = 1; 12018 } 12019 if (dp2->di_extsize != inodedep->id_savedextsize) { 12020 dp2->di_extsize = inodedep->id_savedextsize; 12021 hadchanges = 1; 12022 } 12023 } 12024 inodedep->id_savedsize = -1; 12025 inodedep->id_savedextsize = -1; 12026 inodedep->id_savednlink = -1; 12027 /* 12028 * If there were any rollbacks in the inode block, then it must be 12029 * marked dirty so that its will eventually get written back in 12030 * its correct form. 12031 */ 12032 if (hadchanges) { 12033 if (fstype == UFS2) 12034 ffs_update_dinode_ckhash(inodedep->id_fs, dp2); 12035 bdirty(bp); 12036 } 12037bufwait: 12038 /* 12039 * If the write did not succeed, we have done all the roll-forward 12040 * operations, but we cannot take the actions that will allow its 12041 * dependencies to be processed. 12042 */ 12043 if ((flags & WRITESUCCEEDED) == 0) 12044 return (hadchanges); 12045 /* 12046 * Process any allocdirects that completed during the update. 12047 */ 12048 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 12049 handle_allocdirect_partdone(adp, &wkhd); 12050 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 12051 handle_allocdirect_partdone(adp, &wkhd); 12052 /* 12053 * Process deallocations that were held pending until the 12054 * inode had been written to disk. Freeing of the inode 12055 * is delayed until after all blocks have been freed to 12056 * avoid creation of new <vfsid, inum, lbn> triples 12057 * before the old ones have been deleted. Completely 12058 * unlinked inodes are not processed until the unlinked 12059 * inode list is written or the last reference is removed. 12060 */ 12061 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) != UNLINKED) { 12062 freefile = handle_bufwait(inodedep, NULL); 12063 if (freefile && !LIST_EMPTY(&wkhd)) { 12064 WORKLIST_INSERT(&wkhd, &freefile->fx_list); 12065 freefile = NULL; 12066 } 12067 } 12068 /* 12069 * Move rolled forward dependency completions to the bufwait list 12070 * now that those that were already written have been processed. 12071 */ 12072 if (!LIST_EMPTY(&wkhd) && hadchanges == 0) 12073 panic("handle_written_inodeblock: bufwait but no changes"); 12074 jwork_move(&inodedep->id_bufwait, &wkhd); 12075 12076 if (freefile != NULL) { 12077 /* 12078 * If the inode is goingaway it was never written. Fake up 12079 * the state here so free_inodedep() can succeed. 12080 */ 12081 if (inodedep->id_state & GOINGAWAY) 12082 inodedep->id_state |= COMPLETE | DEPCOMPLETE; 12083 if (free_inodedep(inodedep) == 0) 12084 panic("handle_written_inodeblock: live inodedep %p", 12085 inodedep); 12086 add_to_worklist(&freefile->fx_list, 0); 12087 return (0); 12088 } 12089 12090 /* 12091 * If no outstanding dependencies, free it. 12092 */ 12093 if (free_inodedep(inodedep) || 12094 (TAILQ_FIRST(&inodedep->id_inoreflst) == 0 && 12095 TAILQ_FIRST(&inodedep->id_inoupdt) == 0 && 12096 TAILQ_FIRST(&inodedep->id_extupdt) == 0 && 12097 LIST_FIRST(&inodedep->id_bufwait) == 0)) 12098 return (0); 12099 return (hadchanges); 12100} 12101 12102/* 12103 * Perform needed roll-forwards and kick off any dependencies that 12104 * can now be processed. 12105 * 12106 * If the write did not succeed, we will do all the roll-forward 12107 * operations, but we will not take the actions that will allow its 12108 * dependencies to be processed. 12109 */ 12110static int 12111handle_written_indirdep(indirdep, bp, bpp, flags) 12112 struct indirdep *indirdep; 12113 struct buf *bp; 12114 struct buf **bpp; 12115 int flags; 12116{ 12117 struct allocindir *aip; 12118 struct buf *sbp; 12119 int chgs; 12120 12121 if (indirdep->ir_state & GOINGAWAY) 12122 panic("handle_written_indirdep: indirdep gone"); 12123 if ((indirdep->ir_state & IOSTARTED) == 0) 12124 panic("handle_written_indirdep: IO not started"); 12125 chgs = 0; 12126 /* 12127 * If there were rollbacks revert them here. 12128 */ 12129 if (indirdep->ir_saveddata) { 12130 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 12131 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 12132 free(indirdep->ir_saveddata, M_INDIRDEP); 12133 indirdep->ir_saveddata = NULL; 12134 } 12135 chgs = 1; 12136 } 12137 indirdep->ir_state &= ~(UNDONE | IOSTARTED); 12138 indirdep->ir_state |= ATTACHED; 12139 /* 12140 * If the write did not succeed, we have done all the roll-forward 12141 * operations, but we cannot take the actions that will allow its 12142 * dependencies to be processed. 12143 */ 12144 if ((flags & WRITESUCCEEDED) == 0) { 12145 stat_indir_blk_ptrs++; 12146 bdirty(bp); 12147 return (1); 12148 } 12149 /* 12150 * Move allocindirs with written pointers to the completehd if 12151 * the indirdep's pointer is not yet written. Otherwise 12152 * free them here. 12153 */ 12154 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL) { 12155 LIST_REMOVE(aip, ai_next); 12156 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 12157 LIST_INSERT_HEAD(&indirdep->ir_completehd, aip, 12158 ai_next); 12159 newblk_freefrag(&aip->ai_block); 12160 continue; 12161 } 12162 free_newblk(&aip->ai_block); 12163 } 12164 /* 12165 * Move allocindirs that have finished dependency processing from 12166 * the done list to the write list after updating the pointers. 12167 */ 12168 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 12169 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL) { 12170 handle_allocindir_partdone(aip); 12171 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 12172 panic("disk_write_complete: not gone"); 12173 chgs = 1; 12174 } 12175 } 12176 /* 12177 * Preserve the indirdep if there were any changes or if it is not 12178 * yet valid on disk. 12179 */ 12180 if (chgs) { 12181 stat_indir_blk_ptrs++; 12182 bdirty(bp); 12183 return (1); 12184 } 12185 /* 12186 * If there were no changes we can discard the savedbp and detach 12187 * ourselves from the buf. We are only carrying completed pointers 12188 * in this case. 12189 */ 12190 sbp = indirdep->ir_savebp; 12191 sbp->b_flags |= B_INVAL | B_NOCACHE; 12192 indirdep->ir_savebp = NULL; 12193 indirdep->ir_bp = NULL; 12194 if (*bpp != NULL) 12195 panic("handle_written_indirdep: bp already exists."); 12196 *bpp = sbp; 12197 /* 12198 * The indirdep may not be freed until its parent points at it. 12199 */ 12200 if (indirdep->ir_state & DEPCOMPLETE) 12201 free_indirdep(indirdep); 12202 12203 return (0); 12204} 12205 12206/* 12207 * Process a diradd entry after its dependent inode has been written. 12208 */ 12209static void 12210diradd_inode_written(dap, inodedep) 12211 struct diradd *dap; 12212 struct inodedep *inodedep; 12213{ 12214 12215 LOCK_OWNED(VFSTOUFS(dap->da_list.wk_mp)); 12216 dap->da_state |= COMPLETE; 12217 complete_diradd(dap); 12218 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 12219} 12220 12221/* 12222 * Returns true if the bmsafemap will have rollbacks when written. Must only 12223 * be called with the per-filesystem lock and the buf lock on the cg held. 12224 */ 12225static int 12226bmsafemap_backgroundwrite(bmsafemap, bp) 12227 struct bmsafemap *bmsafemap; 12228 struct buf *bp; 12229{ 12230 int dirty; 12231 12232 LOCK_OWNED(VFSTOUFS(bmsafemap->sm_list.wk_mp)); 12233 dirty = !LIST_EMPTY(&bmsafemap->sm_jaddrefhd) | 12234 !LIST_EMPTY(&bmsafemap->sm_jnewblkhd); 12235 /* 12236 * If we're initiating a background write we need to process the 12237 * rollbacks as they exist now, not as they exist when IO starts. 12238 * No other consumers will look at the contents of the shadowed 12239 * buf so this is safe to do here. 12240 */ 12241 if (bp->b_xflags & BX_BKGRDMARKER) 12242 initiate_write_bmsafemap(bmsafemap, bp); 12243 12244 return (dirty); 12245} 12246 12247/* 12248 * Re-apply an allocation when a cg write is complete. 12249 */ 12250static int 12251jnewblk_rollforward(jnewblk, fs, cgp, blksfree) 12252 struct jnewblk *jnewblk; 12253 struct fs *fs; 12254 struct cg *cgp; 12255 uint8_t *blksfree; 12256{ 12257 ufs1_daddr_t fragno; 12258 ufs2_daddr_t blkno; 12259 long cgbno, bbase; 12260 int frags, blk; 12261 int i; 12262 12263 frags = 0; 12264 cgbno = dtogd(fs, jnewblk->jn_blkno); 12265 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) { 12266 if (isclr(blksfree, cgbno + i)) 12267 panic("jnewblk_rollforward: re-allocated fragment"); 12268 frags++; 12269 } 12270 if (frags == fs->fs_frag) { 12271 blkno = fragstoblks(fs, cgbno); 12272 ffs_clrblock(fs, blksfree, (long)blkno); 12273 ffs_clusteracct(fs, cgp, blkno, -1); 12274 cgp->cg_cs.cs_nbfree--; 12275 } else { 12276 bbase = cgbno - fragnum(fs, cgbno); 12277 cgbno += jnewblk->jn_oldfrags; 12278 /* If a complete block had been reassembled, account for it. */ 12279 fragno = fragstoblks(fs, bbase); 12280 if (ffs_isblock(fs, blksfree, fragno)) { 12281 cgp->cg_cs.cs_nffree += fs->fs_frag; 12282 ffs_clusteracct(fs, cgp, fragno, -1); 12283 cgp->cg_cs.cs_nbfree--; 12284 } 12285 /* Decrement the old frags. */ 12286 blk = blkmap(fs, blksfree, bbase); 12287 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 12288 /* Allocate the fragment */ 12289 for (i = 0; i < frags; i++) 12290 clrbit(blksfree, cgbno + i); 12291 cgp->cg_cs.cs_nffree -= frags; 12292 /* Add back in counts associated with the new frags */ 12293 blk = blkmap(fs, blksfree, bbase); 12294 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 12295 } 12296 return (frags); 12297} 12298 12299/* 12300 * Complete a write to a bmsafemap structure. Roll forward any bitmap 12301 * changes if it's not a background write. Set all written dependencies 12302 * to DEPCOMPLETE and free the structure if possible. 12303 * 12304 * If the write did not succeed, we will do all the roll-forward 12305 * operations, but we will not take the actions that will allow its 12306 * dependencies to be processed. 12307 */ 12308static int 12309handle_written_bmsafemap(bmsafemap, bp, flags) 12310 struct bmsafemap *bmsafemap; 12311 struct buf *bp; 12312 int flags; 12313{ 12314 struct newblk *newblk; 12315 struct inodedep *inodedep; 12316 struct jaddref *jaddref, *jatmp; 12317 struct jnewblk *jnewblk, *jntmp; 12318 struct ufsmount *ump; 12319 uint8_t *inosused; 12320 uint8_t *blksfree; 12321 struct cg *cgp; 12322 struct fs *fs; 12323 ino_t ino; 12324 int foreground; 12325 int chgs; 12326 12327 if ((bmsafemap->sm_state & IOSTARTED) == 0) 12328 panic("handle_written_bmsafemap: Not started\n"); 12329 ump = VFSTOUFS(bmsafemap->sm_list.wk_mp); 12330 chgs = 0; 12331 bmsafemap->sm_state &= ~IOSTARTED; 12332 foreground = (bp->b_xflags & BX_BKGRDMARKER) == 0; 12333 /* 12334 * If write was successful, release journal work that was waiting 12335 * on the write. Otherwise move the work back. 12336 */ 12337 if (flags & WRITESUCCEEDED) 12338 handle_jwork(&bmsafemap->sm_freewr); 12339 else 12340 LIST_CONCAT(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, 12341 worklist, wk_list); 12342 12343 /* 12344 * Restore unwritten inode allocation pending jaddref writes. 12345 */ 12346 if (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd)) { 12347 cgp = (struct cg *)bp->b_data; 12348 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 12349 inosused = cg_inosused(cgp); 12350 LIST_FOREACH_SAFE(jaddref, &bmsafemap->sm_jaddrefhd, 12351 ja_bmdeps, jatmp) { 12352 if ((jaddref->ja_state & UNDONE) == 0) 12353 continue; 12354 ino = jaddref->ja_ino % fs->fs_ipg; 12355 if (isset(inosused, ino)) 12356 panic("handle_written_bmsafemap: " 12357 "re-allocated inode"); 12358 /* Do the roll-forward only if it's a real copy. */ 12359 if (foreground) { 12360 if ((jaddref->ja_mode & IFMT) == IFDIR) 12361 cgp->cg_cs.cs_ndir++; 12362 cgp->cg_cs.cs_nifree--; 12363 setbit(inosused, ino); 12364 chgs = 1; 12365 } 12366 jaddref->ja_state &= ~UNDONE; 12367 jaddref->ja_state |= ATTACHED; 12368 free_jaddref(jaddref); 12369 } 12370 } 12371 /* 12372 * Restore any block allocations which are pending journal writes. 12373 */ 12374 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 12375 cgp = (struct cg *)bp->b_data; 12376 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 12377 blksfree = cg_blksfree(cgp); 12378 LIST_FOREACH_SAFE(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps, 12379 jntmp) { 12380 if ((jnewblk->jn_state & UNDONE) == 0) 12381 continue; 12382 /* Do the roll-forward only if it's a real copy. */ 12383 if (foreground && 12384 jnewblk_rollforward(jnewblk, fs, cgp, blksfree)) 12385 chgs = 1; 12386 jnewblk->jn_state &= ~(UNDONE | NEWBLOCK); 12387 jnewblk->jn_state |= ATTACHED; 12388 free_jnewblk(jnewblk); 12389 } 12390 } 12391 /* 12392 * If the write did not succeed, we have done all the roll-forward 12393 * operations, but we cannot take the actions that will allow its 12394 * dependencies to be processed. 12395 */ 12396 if ((flags & WRITESUCCEEDED) == 0) { 12397 LIST_CONCAT(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr, 12398 newblk, nb_deps); 12399 LIST_CONCAT(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, 12400 worklist, wk_list); 12401 if (foreground) 12402 bdirty(bp); 12403 return (1); 12404 } 12405 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkwr))) { 12406 newblk->nb_state |= DEPCOMPLETE; 12407 newblk->nb_state &= ~ONDEPLIST; 12408 newblk->nb_bmsafemap = NULL; 12409 LIST_REMOVE(newblk, nb_deps); 12410 if (newblk->nb_list.wk_type == D_ALLOCDIRECT) 12411 handle_allocdirect_partdone( 12412 WK_ALLOCDIRECT(&newblk->nb_list), NULL); 12413 else if (newblk->nb_list.wk_type == D_ALLOCINDIR) 12414 handle_allocindir_partdone( 12415 WK_ALLOCINDIR(&newblk->nb_list)); 12416 else if (newblk->nb_list.wk_type != D_NEWBLK) 12417 panic("handle_written_bmsafemap: Unexpected type: %s", 12418 TYPENAME(newblk->nb_list.wk_type)); 12419 } 12420 while ((inodedep = LIST_FIRST(&bmsafemap->sm_inodedepwr)) != NULL) { 12421 inodedep->id_state |= DEPCOMPLETE; 12422 inodedep->id_state &= ~ONDEPLIST; 12423 LIST_REMOVE(inodedep, id_deps); 12424 inodedep->id_bmsafemap = NULL; 12425 } 12426 LIST_REMOVE(bmsafemap, sm_next); 12427 if (chgs == 0 && LIST_EMPTY(&bmsafemap->sm_jaddrefhd) && 12428 LIST_EMPTY(&bmsafemap->sm_jnewblkhd) && 12429 LIST_EMPTY(&bmsafemap->sm_newblkhd) && 12430 LIST_EMPTY(&bmsafemap->sm_inodedephd) && 12431 LIST_EMPTY(&bmsafemap->sm_freehd)) { 12432 LIST_REMOVE(bmsafemap, sm_hash); 12433 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 12434 return (0); 12435 } 12436 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next); 12437 if (foreground) 12438 bdirty(bp); 12439 return (1); 12440} 12441 12442/* 12443 * Try to free a mkdir dependency. 12444 */ 12445static void 12446complete_mkdir(mkdir) 12447 struct mkdir *mkdir; 12448{ 12449 struct diradd *dap; 12450 12451 if ((mkdir->md_state & ALLCOMPLETE) != ALLCOMPLETE) 12452 return; 12453 LIST_REMOVE(mkdir, md_mkdirs); 12454 dap = mkdir->md_diradd; 12455 dap->da_state &= ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 12456 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) { 12457 dap->da_state |= DEPCOMPLETE; 12458 complete_diradd(dap); 12459 } 12460 WORKITEM_FREE(mkdir, D_MKDIR); 12461} 12462 12463/* 12464 * Handle the completion of a mkdir dependency. 12465 */ 12466static void 12467handle_written_mkdir(mkdir, type) 12468 struct mkdir *mkdir; 12469 int type; 12470{ 12471 12472 if ((mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)) != type) 12473 panic("handle_written_mkdir: bad type"); 12474 mkdir->md_state |= COMPLETE; 12475 complete_mkdir(mkdir); 12476} 12477 12478static int 12479free_pagedep(pagedep) 12480 struct pagedep *pagedep; 12481{ 12482 int i; 12483 12484 if (pagedep->pd_state & NEWBLOCK) 12485 return (0); 12486 if (!LIST_EMPTY(&pagedep->pd_dirremhd)) 12487 return (0); 12488 for (i = 0; i < DAHASHSZ; i++) 12489 if (!LIST_EMPTY(&pagedep->pd_diraddhd[i])) 12490 return (0); 12491 if (!LIST_EMPTY(&pagedep->pd_pendinghd)) 12492 return (0); 12493 if (!LIST_EMPTY(&pagedep->pd_jmvrefhd)) 12494 return (0); 12495 if (pagedep->pd_state & ONWORKLIST) 12496 WORKLIST_REMOVE(&pagedep->pd_list); 12497 LIST_REMOVE(pagedep, pd_hash); 12498 WORKITEM_FREE(pagedep, D_PAGEDEP); 12499 12500 return (1); 12501} 12502 12503/* 12504 * Called from within softdep_disk_write_complete above. 12505 * A write operation was just completed. Removed inodes can 12506 * now be freed and associated block pointers may be committed. 12507 * Note that this routine is always called from interrupt level 12508 * with further interrupts from this device blocked. 12509 * 12510 * If the write did not succeed, we will do all the roll-forward 12511 * operations, but we will not take the actions that will allow its 12512 * dependencies to be processed. 12513 */ 12514static int 12515handle_written_filepage(pagedep, bp, flags) 12516 struct pagedep *pagedep; 12517 struct buf *bp; /* buffer containing the written page */ 12518 int flags; 12519{ 12520 struct dirrem *dirrem; 12521 struct diradd *dap, *nextdap; 12522 struct direct *ep; 12523 int i, chgs; 12524 12525 if ((pagedep->pd_state & IOSTARTED) == 0) 12526 panic("handle_written_filepage: not started"); 12527 pagedep->pd_state &= ~IOSTARTED; 12528 if ((flags & WRITESUCCEEDED) == 0) 12529 goto rollforward; 12530 /* 12531 * Process any directory removals that have been committed. 12532 */ 12533 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 12534 LIST_REMOVE(dirrem, dm_next); 12535 dirrem->dm_state |= COMPLETE; 12536 dirrem->dm_dirinum = pagedep->pd_ino; 12537 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 12538 ("handle_written_filepage: Journal entries not written.")); 12539 add_to_worklist(&dirrem->dm_list, 0); 12540 } 12541 /* 12542 * Free any directory additions that have been committed. 12543 * If it is a newly allocated block, we have to wait until 12544 * the on-disk directory inode claims the new block. 12545 */ 12546 if ((pagedep->pd_state & NEWBLOCK) == 0) 12547 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 12548 free_diradd(dap, NULL); 12549rollforward: 12550 /* 12551 * Uncommitted directory entries must be restored. 12552 */ 12553 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 12554 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 12555 dap = nextdap) { 12556 nextdap = LIST_NEXT(dap, da_pdlist); 12557 if (dap->da_state & ATTACHED) 12558 panic("handle_written_filepage: attached"); 12559 ep = (struct direct *) 12560 ((char *)bp->b_data + dap->da_offset); 12561 ep->d_ino = dap->da_newinum; 12562 dap->da_state &= ~UNDONE; 12563 dap->da_state |= ATTACHED; 12564 chgs = 1; 12565 /* 12566 * If the inode referenced by the directory has 12567 * been written out, then the dependency can be 12568 * moved to the pending list. 12569 */ 12570 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 12571 LIST_REMOVE(dap, da_pdlist); 12572 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 12573 da_pdlist); 12574 } 12575 } 12576 } 12577 /* 12578 * If there were any rollbacks in the directory, then it must be 12579 * marked dirty so that its will eventually get written back in 12580 * its correct form. 12581 */ 12582 if (chgs || (flags & WRITESUCCEEDED) == 0) { 12583 if ((bp->b_flags & B_DELWRI) == 0) 12584 stat_dir_entry++; 12585 bdirty(bp); 12586 return (1); 12587 } 12588 /* 12589 * If we are not waiting for a new directory block to be 12590 * claimed by its inode, then the pagedep will be freed. 12591 * Otherwise it will remain to track any new entries on 12592 * the page in case they are fsync'ed. 12593 */ 12594 free_pagedep(pagedep); 12595 return (0); 12596} 12597 12598/* 12599 * Writing back in-core inode structures. 12600 * 12601 * The filesystem only accesses an inode's contents when it occupies an 12602 * "in-core" inode structure. These "in-core" structures are separate from 12603 * the page frames used to cache inode blocks. Only the latter are 12604 * transferred to/from the disk. So, when the updated contents of the 12605 * "in-core" inode structure are copied to the corresponding in-memory inode 12606 * block, the dependencies are also transferred. The following procedure is 12607 * called when copying a dirty "in-core" inode to a cached inode block. 12608 */ 12609 12610/* 12611 * Called when an inode is loaded from disk. If the effective link count 12612 * differed from the actual link count when it was last flushed, then we 12613 * need to ensure that the correct effective link count is put back. 12614 */ 12615void 12616softdep_load_inodeblock(ip) 12617 struct inode *ip; /* the "in_core" copy of the inode */ 12618{ 12619 struct inodedep *inodedep; 12620 struct ufsmount *ump; 12621 12622 ump = ITOUMP(ip); 12623 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 12624 ("softdep_load_inodeblock called on non-softdep filesystem")); 12625 /* 12626 * Check for alternate nlink count. 12627 */ 12628 ip->i_effnlink = ip->i_nlink; 12629 ACQUIRE_LOCK(ump); 12630 if (inodedep_lookup(UFSTOVFS(ump), ip->i_number, 0, &inodedep) == 0) { 12631 FREE_LOCK(ump); 12632 return; 12633 } 12634 if (ip->i_nlink != inodedep->id_nlinkwrote && 12635 inodedep->id_nlinkwrote != -1) { 12636 KASSERT(ip->i_nlink == 0 && 12637 (ump->um_flags & UM_FSFAIL_CLEANUP) != 0, 12638 ("read bad i_nlink value")); 12639 ip->i_effnlink = ip->i_nlink = inodedep->id_nlinkwrote; 12640 } 12641 ip->i_effnlink -= inodedep->id_nlinkdelta; 12642 KASSERT(ip->i_effnlink >= 0, 12643 ("softdep_load_inodeblock: negative i_effnlink")); 12644 FREE_LOCK(ump); 12645} 12646 12647/* 12648 * This routine is called just before the "in-core" inode 12649 * information is to be copied to the in-memory inode block. 12650 * Recall that an inode block contains several inodes. If 12651 * the force flag is set, then the dependencies will be 12652 * cleared so that the update can always be made. Note that 12653 * the buffer is locked when this routine is called, so we 12654 * will never be in the middle of writing the inode block 12655 * to disk. 12656 */ 12657void 12658softdep_update_inodeblock(ip, bp, waitfor) 12659 struct inode *ip; /* the "in_core" copy of the inode */ 12660 struct buf *bp; /* the buffer containing the inode block */ 12661 int waitfor; /* nonzero => update must be allowed */ 12662{ 12663 struct inodedep *inodedep; 12664 struct inoref *inoref; 12665 struct ufsmount *ump; 12666 struct worklist *wk; 12667 struct mount *mp; 12668 struct buf *ibp; 12669 struct fs *fs; 12670 int error; 12671 12672 ump = ITOUMP(ip); 12673 mp = UFSTOVFS(ump); 12674 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 12675 ("softdep_update_inodeblock called on non-softdep filesystem")); 12676 fs = ump->um_fs; 12677 /* 12678 * Preserve the freelink that is on disk. clear_unlinked_inodedep() 12679 * does not have access to the in-core ip so must write directly into 12680 * the inode block buffer when setting freelink. 12681 */ 12682 if (fs->fs_magic == FS_UFS1_MAGIC) 12683 DIP_SET(ip, i_freelink, ((struct ufs1_dinode *)bp->b_data + 12684 ino_to_fsbo(fs, ip->i_number))->di_freelink); 12685 else 12686 DIP_SET(ip, i_freelink, ((struct ufs2_dinode *)bp->b_data + 12687 ino_to_fsbo(fs, ip->i_number))->di_freelink); 12688 /* 12689 * If the effective link count is not equal to the actual link 12690 * count, then we must track the difference in an inodedep while 12691 * the inode is (potentially) tossed out of the cache. Otherwise, 12692 * if there is no existing inodedep, then there are no dependencies 12693 * to track. 12694 */ 12695 ACQUIRE_LOCK(ump); 12696again: 12697 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 12698 FREE_LOCK(ump); 12699 if (ip->i_effnlink != ip->i_nlink) 12700 panic("softdep_update_inodeblock: bad link count"); 12701 return; 12702 } 12703 KASSERT(ip->i_nlink >= inodedep->id_nlinkdelta, 12704 ("softdep_update_inodeblock inconsistent ip %p i_nlink %d " 12705 "inodedep %p id_nlinkdelta %jd", 12706 ip, ip->i_nlink, inodedep, (intmax_t)inodedep->id_nlinkdelta)); 12707 inodedep->id_nlinkwrote = ip->i_nlink; 12708 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) 12709 panic("softdep_update_inodeblock: bad delta"); 12710 /* 12711 * If we're flushing all dependencies we must also move any waiting 12712 * for journal writes onto the bufwait list prior to I/O. 12713 */ 12714 if (waitfor) { 12715 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12716 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12717 == DEPCOMPLETE) { 12718 jwait(&inoref->if_list, MNT_WAIT); 12719 goto again; 12720 } 12721 } 12722 } 12723 /* 12724 * Changes have been initiated. Anything depending on these 12725 * changes cannot occur until this inode has been written. 12726 */ 12727 inodedep->id_state &= ~COMPLETE; 12728 if ((inodedep->id_state & ONWORKLIST) == 0) 12729 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 12730 /* 12731 * Any new dependencies associated with the incore inode must 12732 * now be moved to the list associated with the buffer holding 12733 * the in-memory copy of the inode. Once merged process any 12734 * allocdirects that are completed by the merger. 12735 */ 12736 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt); 12737 if (!TAILQ_EMPTY(&inodedep->id_inoupdt)) 12738 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt), 12739 NULL); 12740 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt); 12741 if (!TAILQ_EMPTY(&inodedep->id_extupdt)) 12742 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt), 12743 NULL); 12744 /* 12745 * Now that the inode has been pushed into the buffer, the 12746 * operations dependent on the inode being written to disk 12747 * can be moved to the id_bufwait so that they will be 12748 * processed when the buffer I/O completes. 12749 */ 12750 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 12751 WORKLIST_REMOVE(wk); 12752 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 12753 } 12754 /* 12755 * Newly allocated inodes cannot be written until the bitmap 12756 * that allocates them have been written (indicated by 12757 * DEPCOMPLETE being set in id_state). If we are doing a 12758 * forced sync (e.g., an fsync on a file), we force the bitmap 12759 * to be written so that the update can be done. 12760 */ 12761 if (waitfor == 0) { 12762 FREE_LOCK(ump); 12763 return; 12764 } 12765retry: 12766 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) != 0) { 12767 FREE_LOCK(ump); 12768 return; 12769 } 12770 ibp = inodedep->id_bmsafemap->sm_buf; 12771 ibp = getdirtybuf(ibp, LOCK_PTR(ump), MNT_WAIT); 12772 if (ibp == NULL) { 12773 /* 12774 * If ibp came back as NULL, the dependency could have been 12775 * freed while we slept. Look it up again, and check to see 12776 * that it has completed. 12777 */ 12778 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 12779 goto retry; 12780 FREE_LOCK(ump); 12781 return; 12782 } 12783 FREE_LOCK(ump); 12784 if ((error = bwrite(ibp)) != 0) 12785 softdep_error("softdep_update_inodeblock: bwrite", error); 12786} 12787 12788/* 12789 * Merge the a new inode dependency list (such as id_newinoupdt) into an 12790 * old inode dependency list (such as id_inoupdt). 12791 */ 12792static void 12793merge_inode_lists(newlisthead, oldlisthead) 12794 struct allocdirectlst *newlisthead; 12795 struct allocdirectlst *oldlisthead; 12796{ 12797 struct allocdirect *listadp, *newadp; 12798 12799 newadp = TAILQ_FIRST(newlisthead); 12800 if (newadp != NULL) 12801 LOCK_OWNED(VFSTOUFS(newadp->ad_block.nb_list.wk_mp)); 12802 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) { 12803 if (listadp->ad_offset < newadp->ad_offset) { 12804 listadp = TAILQ_NEXT(listadp, ad_next); 12805 continue; 12806 } 12807 TAILQ_REMOVE(newlisthead, newadp, ad_next); 12808 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 12809 if (listadp->ad_offset == newadp->ad_offset) { 12810 allocdirect_merge(oldlisthead, newadp, 12811 listadp); 12812 listadp = newadp; 12813 } 12814 newadp = TAILQ_FIRST(newlisthead); 12815 } 12816 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) { 12817 TAILQ_REMOVE(newlisthead, newadp, ad_next); 12818 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next); 12819 } 12820} 12821 12822/* 12823 * If we are doing an fsync, then we must ensure that any directory 12824 * entries for the inode have been written after the inode gets to disk. 12825 */ 12826int 12827softdep_fsync(vp) 12828 struct vnode *vp; /* the "in_core" copy of the inode */ 12829{ 12830 struct inodedep *inodedep; 12831 struct pagedep *pagedep; 12832 struct inoref *inoref; 12833 struct ufsmount *ump; 12834 struct worklist *wk; 12835 struct diradd *dap; 12836 struct mount *mp; 12837 struct vnode *pvp; 12838 struct inode *ip; 12839 struct buf *bp; 12840 struct fs *fs; 12841 struct thread *td = curthread; 12842 int error, flushparent, pagedep_new_block; 12843 ino_t parentino; 12844 ufs_lbn_t lbn; 12845 12846 ip = VTOI(vp); 12847 mp = vp->v_mount; 12848 ump = VFSTOUFS(mp); 12849 fs = ump->um_fs; 12850 if (MOUNTEDSOFTDEP(mp) == 0) 12851 return (0); 12852 ACQUIRE_LOCK(ump); 12853restart: 12854 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 12855 FREE_LOCK(ump); 12856 return (0); 12857 } 12858 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12859 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12860 == DEPCOMPLETE) { 12861 jwait(&inoref->if_list, MNT_WAIT); 12862 goto restart; 12863 } 12864 } 12865 if (!LIST_EMPTY(&inodedep->id_inowait) || 12866 !TAILQ_EMPTY(&inodedep->id_extupdt) || 12867 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 12868 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 12869 !TAILQ_EMPTY(&inodedep->id_newinoupdt)) 12870 panic("softdep_fsync: pending ops %p", inodedep); 12871 for (error = 0, flushparent = 0; ; ) { 12872 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 12873 break; 12874 if (wk->wk_type != D_DIRADD) 12875 panic("softdep_fsync: Unexpected type %s", 12876 TYPENAME(wk->wk_type)); 12877 dap = WK_DIRADD(wk); 12878 /* 12879 * Flush our parent if this directory entry has a MKDIR_PARENT 12880 * dependency or is contained in a newly allocated block. 12881 */ 12882 if (dap->da_state & DIRCHG) 12883 pagedep = dap->da_previous->dm_pagedep; 12884 else 12885 pagedep = dap->da_pagedep; 12886 parentino = pagedep->pd_ino; 12887 lbn = pagedep->pd_lbn; 12888 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 12889 panic("softdep_fsync: dirty"); 12890 if ((dap->da_state & MKDIR_PARENT) || 12891 (pagedep->pd_state & NEWBLOCK)) 12892 flushparent = 1; 12893 else 12894 flushparent = 0; 12895 /* 12896 * If we are being fsync'ed as part of vgone'ing this vnode, 12897 * then we will not be able to release and recover the 12898 * vnode below, so we just have to give up on writing its 12899 * directory entry out. It will eventually be written, just 12900 * not now, but then the user was not asking to have it 12901 * written, so we are not breaking any promises. 12902 */ 12903 if (VN_IS_DOOMED(vp)) 12904 break; 12905 /* 12906 * We prevent deadlock by always fetching inodes from the 12907 * root, moving down the directory tree. Thus, when fetching 12908 * our parent directory, we first try to get the lock. If 12909 * that fails, we must unlock ourselves before requesting 12910 * the lock on our parent. See the comment in ufs_lookup 12911 * for details on possible races. 12912 */ 12913 FREE_LOCK(ump); 12914 error = get_parent_vp(vp, mp, parentino, NULL, NULL, NULL, 12915 &pvp); 12916 if (error == ERELOOKUP) 12917 error = 0; 12918 if (error != 0) 12919 return (error); 12920 /* 12921 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps 12922 * that are contained in direct blocks will be resolved by 12923 * doing a ffs_update. Pagedeps contained in indirect blocks 12924 * may require a complete sync'ing of the directory. So, we 12925 * try the cheap and fast ffs_update first, and if that fails, 12926 * then we do the slower ffs_syncvnode of the directory. 12927 */ 12928 if (flushparent) { 12929 int locked; 12930 12931 if ((error = ffs_update(pvp, 1)) != 0) { 12932 vput(pvp); 12933 return (error); 12934 } 12935 ACQUIRE_LOCK(ump); 12936 locked = 1; 12937 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) { 12938 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) { 12939 if (wk->wk_type != D_DIRADD) 12940 panic("softdep_fsync: Unexpected type %s", 12941 TYPENAME(wk->wk_type)); 12942 dap = WK_DIRADD(wk); 12943 if (dap->da_state & DIRCHG) 12944 pagedep = dap->da_previous->dm_pagedep; 12945 else 12946 pagedep = dap->da_pagedep; 12947 pagedep_new_block = pagedep->pd_state & NEWBLOCK; 12948 FREE_LOCK(ump); 12949 locked = 0; 12950 if (pagedep_new_block && (error = 12951 ffs_syncvnode(pvp, MNT_WAIT, 0))) { 12952 vput(pvp); 12953 return (error); 12954 } 12955 } 12956 } 12957 if (locked) 12958 FREE_LOCK(ump); 12959 } 12960 /* 12961 * Flush directory page containing the inode's name. 12962 */ 12963 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred, 12964 &bp); 12965 if (error == 0) 12966 error = bwrite(bp); 12967 else 12968 brelse(bp); 12969 vput(pvp); 12970 if (!ffs_fsfail_cleanup(ump, error)) 12971 return (error); 12972 ACQUIRE_LOCK(ump); 12973 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 12974 break; 12975 } 12976 FREE_LOCK(ump); 12977 return (0); 12978} 12979 12980/* 12981 * Flush all the dirty bitmaps associated with the block device 12982 * before flushing the rest of the dirty blocks so as to reduce 12983 * the number of dependencies that will have to be rolled back. 12984 * 12985 * XXX Unused? 12986 */ 12987void 12988softdep_fsync_mountdev(vp) 12989 struct vnode *vp; 12990{ 12991 struct buf *bp, *nbp; 12992 struct worklist *wk; 12993 struct bufobj *bo; 12994 12995 if (!vn_isdisk(vp)) 12996 panic("softdep_fsync_mountdev: vnode not a disk"); 12997 bo = &vp->v_bufobj; 12998restart: 12999 BO_LOCK(bo); 13000 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 13001 /* 13002 * If it is already scheduled, skip to the next buffer. 13003 */ 13004 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 13005 continue; 13006 13007 if ((bp->b_flags & B_DELWRI) == 0) 13008 panic("softdep_fsync_mountdev: not dirty"); 13009 /* 13010 * We are only interested in bitmaps with outstanding 13011 * dependencies. 13012 */ 13013 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 13014 wk->wk_type != D_BMSAFEMAP || 13015 (bp->b_vflags & BV_BKGRDINPROG)) { 13016 BUF_UNLOCK(bp); 13017 continue; 13018 } 13019 BO_UNLOCK(bo); 13020 bremfree(bp); 13021 (void) bawrite(bp); 13022 goto restart; 13023 } 13024 drain_output(vp); 13025 BO_UNLOCK(bo); 13026} 13027 13028/* 13029 * Sync all cylinder groups that were dirty at the time this function is 13030 * called. Newly dirtied cgs will be inserted before the sentinel. This 13031 * is used to flush freedep activity that may be holding up writes to a 13032 * indirect block. 13033 */ 13034static int 13035sync_cgs(mp, waitfor) 13036 struct mount *mp; 13037 int waitfor; 13038{ 13039 struct bmsafemap *bmsafemap; 13040 struct bmsafemap *sentinel; 13041 struct ufsmount *ump; 13042 struct buf *bp; 13043 int error; 13044 13045 sentinel = malloc(sizeof(*sentinel), M_BMSAFEMAP, M_ZERO | M_WAITOK); 13046 sentinel->sm_cg = -1; 13047 ump = VFSTOUFS(mp); 13048 error = 0; 13049 ACQUIRE_LOCK(ump); 13050 LIST_INSERT_HEAD(&ump->softdep_dirtycg, sentinel, sm_next); 13051 for (bmsafemap = LIST_NEXT(sentinel, sm_next); bmsafemap != NULL; 13052 bmsafemap = LIST_NEXT(sentinel, sm_next)) { 13053 /* Skip sentinels and cgs with no work to release. */ 13054 if (bmsafemap->sm_cg == -1 || 13055 (LIST_EMPTY(&bmsafemap->sm_freehd) && 13056 LIST_EMPTY(&bmsafemap->sm_freewr))) { 13057 LIST_REMOVE(sentinel, sm_next); 13058 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next); 13059 continue; 13060 } 13061 /* 13062 * If we don't get the lock and we're waiting try again, if 13063 * not move on to the next buf and try to sync it. 13064 */ 13065 bp = getdirtybuf(bmsafemap->sm_buf, LOCK_PTR(ump), waitfor); 13066 if (bp == NULL && waitfor == MNT_WAIT) 13067 continue; 13068 LIST_REMOVE(sentinel, sm_next); 13069 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next); 13070 if (bp == NULL) 13071 continue; 13072 FREE_LOCK(ump); 13073 if (waitfor == MNT_NOWAIT) 13074 bawrite(bp); 13075 else 13076 error = bwrite(bp); 13077 ACQUIRE_LOCK(ump); 13078 if (error) 13079 break; 13080 } 13081 LIST_REMOVE(sentinel, sm_next); 13082 FREE_LOCK(ump); 13083 free(sentinel, M_BMSAFEMAP); 13084 return (error); 13085} 13086 13087/* 13088 * This routine is called when we are trying to synchronously flush a 13089 * file. This routine must eliminate any filesystem metadata dependencies 13090 * so that the syncing routine can succeed. 13091 */ 13092int 13093softdep_sync_metadata(struct vnode *vp) 13094{ 13095 struct inode *ip; 13096 int error; 13097 13098 ip = VTOI(vp); 13099 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 13100 ("softdep_sync_metadata called on non-softdep filesystem")); 13101 /* 13102 * Ensure that any direct block dependencies have been cleared, 13103 * truncations are started, and inode references are journaled. 13104 */ 13105 ACQUIRE_LOCK(VFSTOUFS(vp->v_mount)); 13106 /* 13107 * Write all journal records to prevent rollbacks on devvp. 13108 */ 13109 if (vp->v_type == VCHR) 13110 softdep_flushjournal(vp->v_mount); 13111 error = flush_inodedep_deps(vp, vp->v_mount, ip->i_number); 13112 /* 13113 * Ensure that all truncates are written so we won't find deps on 13114 * indirect blocks. 13115 */ 13116 process_truncates(vp); 13117 FREE_LOCK(VFSTOUFS(vp->v_mount)); 13118 13119 return (error); 13120} 13121 13122/* 13123 * This routine is called when we are attempting to sync a buf with 13124 * dependencies. If waitfor is MNT_NOWAIT it attempts to schedule any 13125 * other IO it can but returns EBUSY if the buffer is not yet able to 13126 * be written. Dependencies which will not cause rollbacks will always 13127 * return 0. 13128 */ 13129int 13130softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 13131{ 13132 struct indirdep *indirdep; 13133 struct pagedep *pagedep; 13134 struct allocindir *aip; 13135 struct newblk *newblk; 13136 struct ufsmount *ump; 13137 struct buf *nbp; 13138 struct worklist *wk; 13139 int i, error; 13140 13141 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 13142 ("softdep_sync_buf called on non-softdep filesystem")); 13143 /* 13144 * For VCHR we just don't want to force flush any dependencies that 13145 * will cause rollbacks. 13146 */ 13147 if (vp->v_type == VCHR) { 13148 if (waitfor == MNT_NOWAIT && softdep_count_dependencies(bp, 0)) 13149 return (EBUSY); 13150 return (0); 13151 } 13152 ump = VFSTOUFS(vp->v_mount); 13153 ACQUIRE_LOCK(ump); 13154 /* 13155 * As we hold the buffer locked, none of its dependencies 13156 * will disappear. 13157 */ 13158 error = 0; 13159top: 13160 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 13161 switch (wk->wk_type) { 13162 case D_ALLOCDIRECT: 13163 case D_ALLOCINDIR: 13164 newblk = WK_NEWBLK(wk); 13165 if (newblk->nb_jnewblk != NULL) { 13166 if (waitfor == MNT_NOWAIT) { 13167 error = EBUSY; 13168 goto out_unlock; 13169 } 13170 jwait(&newblk->nb_jnewblk->jn_list, waitfor); 13171 goto top; 13172 } 13173 if (newblk->nb_state & DEPCOMPLETE || 13174 waitfor == MNT_NOWAIT) 13175 continue; 13176 nbp = newblk->nb_bmsafemap->sm_buf; 13177 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor); 13178 if (nbp == NULL) 13179 goto top; 13180 FREE_LOCK(ump); 13181 if ((error = bwrite(nbp)) != 0) 13182 goto out; 13183 ACQUIRE_LOCK(ump); 13184 continue; 13185 13186 case D_INDIRDEP: 13187 indirdep = WK_INDIRDEP(wk); 13188 if (waitfor == MNT_NOWAIT) { 13189 if (!TAILQ_EMPTY(&indirdep->ir_trunc) || 13190 !LIST_EMPTY(&indirdep->ir_deplisthd)) { 13191 error = EBUSY; 13192 goto out_unlock; 13193 } 13194 } 13195 if (!TAILQ_EMPTY(&indirdep->ir_trunc)) 13196 panic("softdep_sync_buf: truncation pending."); 13197 restart: 13198 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 13199 newblk = (struct newblk *)aip; 13200 if (newblk->nb_jnewblk != NULL) { 13201 jwait(&newblk->nb_jnewblk->jn_list, 13202 waitfor); 13203 goto restart; 13204 } 13205 if (newblk->nb_state & DEPCOMPLETE) 13206 continue; 13207 nbp = newblk->nb_bmsafemap->sm_buf; 13208 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor); 13209 if (nbp == NULL) 13210 goto restart; 13211 FREE_LOCK(ump); 13212 if ((error = bwrite(nbp)) != 0) 13213 goto out; 13214 ACQUIRE_LOCK(ump); 13215 goto restart; 13216 } 13217 continue; 13218 13219 case D_PAGEDEP: 13220 /* 13221 * Only flush directory entries in synchronous passes. 13222 */ 13223 if (waitfor != MNT_WAIT) { 13224 error = EBUSY; 13225 goto out_unlock; 13226 } 13227 /* 13228 * While syncing snapshots, we must allow recursive 13229 * lookups. 13230 */ 13231 BUF_AREC(bp); 13232 /* 13233 * We are trying to sync a directory that may 13234 * have dependencies on both its own metadata 13235 * and/or dependencies on the inodes of any 13236 * recently allocated files. We walk its diradd 13237 * lists pushing out the associated inode. 13238 */ 13239 pagedep = WK_PAGEDEP(wk); 13240 for (i = 0; i < DAHASHSZ; i++) { 13241 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 13242 continue; 13243 error = flush_pagedep_deps(vp, wk->wk_mp, 13244 &pagedep->pd_diraddhd[i], bp); 13245 if (error != 0) { 13246 if (error != ERELOOKUP) 13247 BUF_NOREC(bp); 13248 goto out_unlock; 13249 } 13250 } 13251 BUF_NOREC(bp); 13252 continue; 13253 13254 case D_FREEWORK: 13255 case D_FREEDEP: 13256 case D_JSEGDEP: 13257 case D_JNEWBLK: 13258 continue; 13259 13260 default: 13261 panic("softdep_sync_buf: Unknown type %s", 13262 TYPENAME(wk->wk_type)); 13263 /* NOTREACHED */ 13264 } 13265 } 13266out_unlock: 13267 FREE_LOCK(ump); 13268out: 13269 return (error); 13270} 13271 13272/* 13273 * Flush the dependencies associated with an inodedep. 13274 */ 13275static int 13276flush_inodedep_deps(vp, mp, ino) 13277 struct vnode *vp; 13278 struct mount *mp; 13279 ino_t ino; 13280{ 13281 struct inodedep *inodedep; 13282 struct inoref *inoref; 13283 struct ufsmount *ump; 13284 int error, waitfor; 13285 13286 /* 13287 * This work is done in two passes. The first pass grabs most 13288 * of the buffers and begins asynchronously writing them. The 13289 * only way to wait for these asynchronous writes is to sleep 13290 * on the filesystem vnode which may stay busy for a long time 13291 * if the filesystem is active. So, instead, we make a second 13292 * pass over the dependencies blocking on each write. In the 13293 * usual case we will be blocking against a write that we 13294 * initiated, so when it is done the dependency will have been 13295 * resolved. Thus the second pass is expected to end quickly. 13296 * We give a brief window at the top of the loop to allow 13297 * any pending I/O to complete. 13298 */ 13299 ump = VFSTOUFS(mp); 13300 LOCK_OWNED(ump); 13301 for (error = 0, waitfor = MNT_NOWAIT; ; ) { 13302 if (error) 13303 return (error); 13304 FREE_LOCK(ump); 13305 ACQUIRE_LOCK(ump); 13306restart: 13307 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 13308 return (0); 13309 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 13310 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 13311 == DEPCOMPLETE) { 13312 jwait(&inoref->if_list, MNT_WAIT); 13313 goto restart; 13314 } 13315 } 13316 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) || 13317 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) || 13318 flush_deplist(&inodedep->id_extupdt, waitfor, &error) || 13319 flush_deplist(&inodedep->id_newextupdt, waitfor, &error)) 13320 continue; 13321 /* 13322 * If pass2, we are done, otherwise do pass 2. 13323 */ 13324 if (waitfor == MNT_WAIT) 13325 break; 13326 waitfor = MNT_WAIT; 13327 } 13328 /* 13329 * Try freeing inodedep in case all dependencies have been removed. 13330 */ 13331 if (inodedep_lookup(mp, ino, 0, &inodedep) != 0) 13332 (void) free_inodedep(inodedep); 13333 return (0); 13334} 13335 13336/* 13337 * Flush an inode dependency list. 13338 */ 13339static int 13340flush_deplist(listhead, waitfor, errorp) 13341 struct allocdirectlst *listhead; 13342 int waitfor; 13343 int *errorp; 13344{ 13345 struct allocdirect *adp; 13346 struct newblk *newblk; 13347 struct ufsmount *ump; 13348 struct buf *bp; 13349 13350 if ((adp = TAILQ_FIRST(listhead)) == NULL) 13351 return (0); 13352 ump = VFSTOUFS(adp->ad_list.wk_mp); 13353 LOCK_OWNED(ump); 13354 TAILQ_FOREACH(adp, listhead, ad_next) { 13355 newblk = (struct newblk *)adp; 13356 if (newblk->nb_jnewblk != NULL) { 13357 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 13358 return (1); 13359 } 13360 if (newblk->nb_state & DEPCOMPLETE) 13361 continue; 13362 bp = newblk->nb_bmsafemap->sm_buf; 13363 bp = getdirtybuf(bp, LOCK_PTR(ump), waitfor); 13364 if (bp == NULL) { 13365 if (waitfor == MNT_NOWAIT) 13366 continue; 13367 return (1); 13368 } 13369 FREE_LOCK(ump); 13370 if (waitfor == MNT_NOWAIT) 13371 bawrite(bp); 13372 else 13373 *errorp = bwrite(bp); 13374 ACQUIRE_LOCK(ump); 13375 return (1); 13376 } 13377 return (0); 13378} 13379 13380/* 13381 * Flush dependencies associated with an allocdirect block. 13382 */ 13383static int 13384flush_newblk_dep(vp, mp, lbn) 13385 struct vnode *vp; 13386 struct mount *mp; 13387 ufs_lbn_t lbn; 13388{ 13389 struct newblk *newblk; 13390 struct ufsmount *ump; 13391 struct bufobj *bo; 13392 struct inode *ip; 13393 struct buf *bp; 13394 ufs2_daddr_t blkno; 13395 int error; 13396 13397 error = 0; 13398 bo = &vp->v_bufobj; 13399 ip = VTOI(vp); 13400 blkno = DIP(ip, i_db[lbn]); 13401 if (blkno == 0) 13402 panic("flush_newblk_dep: Missing block"); 13403 ump = VFSTOUFS(mp); 13404 ACQUIRE_LOCK(ump); 13405 /* 13406 * Loop until all dependencies related to this block are satisfied. 13407 * We must be careful to restart after each sleep in case a write 13408 * completes some part of this process for us. 13409 */ 13410 for (;;) { 13411 if (newblk_lookup(mp, blkno, 0, &newblk) == 0) { 13412 FREE_LOCK(ump); 13413 break; 13414 } 13415 if (newblk->nb_list.wk_type != D_ALLOCDIRECT) 13416 panic("flush_newblk_dep: Bad newblk %p", newblk); 13417 /* 13418 * Flush the journal. 13419 */ 13420 if (newblk->nb_jnewblk != NULL) { 13421 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 13422 continue; 13423 } 13424 /* 13425 * Write the bitmap dependency. 13426 */ 13427 if ((newblk->nb_state & DEPCOMPLETE) == 0) { 13428 bp = newblk->nb_bmsafemap->sm_buf; 13429 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT); 13430 if (bp == NULL) 13431 continue; 13432 FREE_LOCK(ump); 13433 error = bwrite(bp); 13434 if (error) 13435 break; 13436 ACQUIRE_LOCK(ump); 13437 continue; 13438 } 13439 /* 13440 * Write the buffer. 13441 */ 13442 FREE_LOCK(ump); 13443 BO_LOCK(bo); 13444 bp = gbincore(bo, lbn); 13445 if (bp != NULL) { 13446 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 13447 LK_INTERLOCK, BO_LOCKPTR(bo)); 13448 if (error == ENOLCK) { 13449 ACQUIRE_LOCK(ump); 13450 error = 0; 13451 continue; /* Slept, retry */ 13452 } 13453 if (error != 0) 13454 break; /* Failed */ 13455 if (bp->b_flags & B_DELWRI) { 13456 bremfree(bp); 13457 error = bwrite(bp); 13458 if (error) 13459 break; 13460 } else 13461 BUF_UNLOCK(bp); 13462 } else 13463 BO_UNLOCK(bo); 13464 /* 13465 * We have to wait for the direct pointers to 13466 * point at the newdirblk before the dependency 13467 * will go away. 13468 */ 13469 error = ffs_update(vp, 1); 13470 if (error) 13471 break; 13472 ACQUIRE_LOCK(ump); 13473 } 13474 return (error); 13475} 13476 13477/* 13478 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 13479 */ 13480static int 13481flush_pagedep_deps(pvp, mp, diraddhdp, locked_bp) 13482 struct vnode *pvp; 13483 struct mount *mp; 13484 struct diraddhd *diraddhdp; 13485 struct buf *locked_bp; 13486{ 13487 struct inodedep *inodedep; 13488 struct inoref *inoref; 13489 struct ufsmount *ump; 13490 struct diradd *dap; 13491 struct vnode *vp; 13492 int error = 0; 13493 struct buf *bp; 13494 ino_t inum; 13495 struct diraddhd unfinished; 13496 13497 LIST_INIT(&unfinished); 13498 ump = VFSTOUFS(mp); 13499 LOCK_OWNED(ump); 13500restart: 13501 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 13502 /* 13503 * Flush ourselves if this directory entry 13504 * has a MKDIR_PARENT dependency. 13505 */ 13506 if (dap->da_state & MKDIR_PARENT) { 13507 FREE_LOCK(ump); 13508 if ((error = ffs_update(pvp, 1)) != 0) 13509 break; 13510 ACQUIRE_LOCK(ump); 13511 /* 13512 * If that cleared dependencies, go on to next. 13513 */ 13514 if (dap != LIST_FIRST(diraddhdp)) 13515 continue; 13516 /* 13517 * All MKDIR_PARENT dependencies and all the 13518 * NEWBLOCK pagedeps that are contained in direct 13519 * blocks were resolved by doing above ffs_update. 13520 * Pagedeps contained in indirect blocks may 13521 * require a complete sync'ing of the directory. 13522 * We are in the midst of doing a complete sync, 13523 * so if they are not resolved in this pass we 13524 * defer them for now as they will be sync'ed by 13525 * our caller shortly. 13526 */ 13527 LIST_REMOVE(dap, da_pdlist); 13528 LIST_INSERT_HEAD(&unfinished, dap, da_pdlist); 13529 continue; 13530 } 13531 /* 13532 * A newly allocated directory must have its "." and 13533 * ".." entries written out before its name can be 13534 * committed in its parent. 13535 */ 13536 inum = dap->da_newinum; 13537 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 13538 panic("flush_pagedep_deps: lost inode1"); 13539 /* 13540 * Wait for any pending journal adds to complete so we don't 13541 * cause rollbacks while syncing. 13542 */ 13543 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 13544 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 13545 == DEPCOMPLETE) { 13546 jwait(&inoref->if_list, MNT_WAIT); 13547 goto restart; 13548 } 13549 } 13550 if (dap->da_state & MKDIR_BODY) { 13551 FREE_LOCK(ump); 13552 error = get_parent_vp(pvp, mp, inum, locked_bp, 13553 diraddhdp, &unfinished, &vp); 13554 if (error != 0) 13555 break; 13556 error = flush_newblk_dep(vp, mp, 0); 13557 /* 13558 * If we still have the dependency we might need to 13559 * update the vnode to sync the new link count to 13560 * disk. 13561 */ 13562 if (error == 0 && dap == LIST_FIRST(diraddhdp)) 13563 error = ffs_update(vp, 1); 13564 vput(vp); 13565 if (error != 0) 13566 break; 13567 ACQUIRE_LOCK(ump); 13568 /* 13569 * If that cleared dependencies, go on to next. 13570 */ 13571 if (dap != LIST_FIRST(diraddhdp)) 13572 continue; 13573 if (dap->da_state & MKDIR_BODY) { 13574 inodedep_lookup(UFSTOVFS(ump), inum, 0, 13575 &inodedep); 13576 panic("flush_pagedep_deps: MKDIR_BODY " 13577 "inodedep %p dap %p vp %p", 13578 inodedep, dap, vp); 13579 } 13580 } 13581 /* 13582 * Flush the inode on which the directory entry depends. 13583 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 13584 * the only remaining dependency is that the updated inode 13585 * count must get pushed to disk. The inode has already 13586 * been pushed into its inode buffer (via VOP_UPDATE) at 13587 * the time of the reference count change. So we need only 13588 * locate that buffer, ensure that there will be no rollback 13589 * caused by a bitmap dependency, then write the inode buffer. 13590 */ 13591retry: 13592 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 13593 panic("flush_pagedep_deps: lost inode"); 13594 /* 13595 * If the inode still has bitmap dependencies, 13596 * push them to disk. 13597 */ 13598 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) == 0) { 13599 bp = inodedep->id_bmsafemap->sm_buf; 13600 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT); 13601 if (bp == NULL) 13602 goto retry; 13603 FREE_LOCK(ump); 13604 if ((error = bwrite(bp)) != 0) 13605 break; 13606 ACQUIRE_LOCK(ump); 13607 if (dap != LIST_FIRST(diraddhdp)) 13608 continue; 13609 } 13610 /* 13611 * If the inode is still sitting in a buffer waiting 13612 * to be written or waiting for the link count to be 13613 * adjusted update it here to flush it to disk. 13614 */ 13615 if (dap == LIST_FIRST(diraddhdp)) { 13616 FREE_LOCK(ump); 13617 error = get_parent_vp(pvp, mp, inum, locked_bp, 13618 diraddhdp, &unfinished, &vp); 13619 if (error != 0) 13620 break; 13621 error = ffs_update(vp, 1); 13622 vput(vp); 13623 if (error) 13624 break; 13625 ACQUIRE_LOCK(ump); 13626 } 13627 /* 13628 * If we have failed to get rid of all the dependencies 13629 * then something is seriously wrong. 13630 */ 13631 if (dap == LIST_FIRST(diraddhdp)) { 13632 inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep); 13633 panic("flush_pagedep_deps: failed to flush " 13634 "inodedep %p ino %ju dap %p", 13635 inodedep, (uintmax_t)inum, dap); 13636 } 13637 } 13638 if (error) 13639 ACQUIRE_LOCK(ump); 13640 while ((dap = LIST_FIRST(&unfinished)) != NULL) { 13641 LIST_REMOVE(dap, da_pdlist); 13642 LIST_INSERT_HEAD(diraddhdp, dap, da_pdlist); 13643 } 13644 return (error); 13645} 13646 13647/* 13648 * A large burst of file addition or deletion activity can drive the 13649 * memory load excessively high. First attempt to slow things down 13650 * using the techniques below. If that fails, this routine requests 13651 * the offending operations to fall back to running synchronously 13652 * until the memory load returns to a reasonable level. 13653 */ 13654int 13655softdep_slowdown(vp) 13656 struct vnode *vp; 13657{ 13658 struct ufsmount *ump; 13659 int jlow; 13660 int max_softdeps_hard; 13661 13662 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 13663 ("softdep_slowdown called on non-softdep filesystem")); 13664 ump = VFSTOUFS(vp->v_mount); 13665 ACQUIRE_LOCK(ump); 13666 jlow = 0; 13667 /* 13668 * Check for journal space if needed. 13669 */ 13670 if (DOINGSUJ(vp)) { 13671 if (journal_space(ump, 0) == 0) 13672 jlow = 1; 13673 } 13674 /* 13675 * If the system is under its limits and our filesystem is 13676 * not responsible for more than our share of the usage and 13677 * we are not low on journal space, then no need to slow down. 13678 */ 13679 max_softdeps_hard = max_softdeps * 11 / 10; 13680 if (dep_current[D_DIRREM] < max_softdeps_hard / 2 && 13681 dep_current[D_INODEDEP] < max_softdeps_hard && 13682 dep_current[D_INDIRDEP] < max_softdeps_hard / 1000 && 13683 dep_current[D_FREEBLKS] < max_softdeps_hard && jlow == 0 && 13684 ump->softdep_curdeps[D_DIRREM] < 13685 (max_softdeps_hard / 2) / stat_flush_threads && 13686 ump->softdep_curdeps[D_INODEDEP] < 13687 max_softdeps_hard / stat_flush_threads && 13688 ump->softdep_curdeps[D_INDIRDEP] < 13689 (max_softdeps_hard / 1000) / stat_flush_threads && 13690 ump->softdep_curdeps[D_FREEBLKS] < 13691 max_softdeps_hard / stat_flush_threads) { 13692 FREE_LOCK(ump); 13693 return (0); 13694 } 13695 /* 13696 * If the journal is low or our filesystem is over its limit 13697 * then speedup the cleanup. 13698 */ 13699 if (ump->softdep_curdeps[D_INDIRDEP] < 13700 (max_softdeps_hard / 1000) / stat_flush_threads || jlow) 13701 softdep_speedup(ump); 13702 stat_sync_limit_hit += 1; 13703 FREE_LOCK(ump); 13704 /* 13705 * We only slow down the rate at which new dependencies are 13706 * generated if we are not using journaling. With journaling, 13707 * the cleanup should always be sufficient to keep things 13708 * under control. 13709 */ 13710 if (DOINGSUJ(vp)) 13711 return (0); 13712 return (1); 13713} 13714 13715static int 13716softdep_request_cleanup_filter(struct vnode *vp, void *arg __unused) 13717{ 13718 return ((vp->v_iflag & VI_OWEINACT) != 0 && vp->v_usecount == 0 && 13719 ((vp->v_vflag & VV_NOSYNC) != 0 || VTOI(vp)->i_effnlink == 0)); 13720} 13721 13722static void 13723softdep_request_cleanup_inactivate(struct mount *mp) 13724{ 13725 struct vnode *vp, *mvp; 13726 int error; 13727 13728 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, softdep_request_cleanup_filter, 13729 NULL) { 13730 vholdl(vp); 13731 vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK | LK_RETRY); 13732 VI_LOCK(vp); 13733 if (vp->v_data != NULL && vp->v_usecount == 0) { 13734 while ((vp->v_iflag & VI_OWEINACT) != 0) { 13735 error = vinactive(vp); 13736 if (error != 0 && error != ERELOOKUP) 13737 break; 13738 } 13739 atomic_add_int(&stat_delayed_inact, 1); 13740 } 13741 VOP_UNLOCK(vp); 13742 vdropl(vp); 13743 } 13744} 13745 13746/* 13747 * Called by the allocation routines when they are about to fail 13748 * in the hope that we can free up the requested resource (inodes 13749 * or disk space). 13750 * 13751 * First check to see if the work list has anything on it. If it has, 13752 * clean up entries until we successfully free the requested resource. 13753 * Because this process holds inodes locked, we cannot handle any remove 13754 * requests that might block on a locked inode as that could lead to 13755 * deadlock. If the worklist yields none of the requested resource, 13756 * start syncing out vnodes to free up the needed space. 13757 */ 13758int 13759softdep_request_cleanup(fs, vp, cred, resource) 13760 struct fs *fs; 13761 struct vnode *vp; 13762 struct ucred *cred; 13763 int resource; 13764{ 13765 struct ufsmount *ump; 13766 struct mount *mp; 13767 long starttime; 13768 ufs2_daddr_t needed; 13769 int error, failed_vnode; 13770 13771 /* 13772 * If we are being called because of a process doing a 13773 * copy-on-write, then it is not safe to process any 13774 * worklist items as we will recurse into the copyonwrite 13775 * routine. This will result in an incoherent snapshot. 13776 * If the vnode that we hold is a snapshot, we must avoid 13777 * handling other resources that could cause deadlock. 13778 */ 13779 if ((curthread->td_pflags & TDP_COWINPROGRESS) || IS_SNAPSHOT(VTOI(vp))) 13780 return (0); 13781 13782 if (resource == FLUSH_BLOCKS_WAIT) 13783 stat_cleanup_blkrequests += 1; 13784 else 13785 stat_cleanup_inorequests += 1; 13786 13787 mp = vp->v_mount; 13788 ump = VFSTOUFS(mp); 13789 mtx_assert(UFS_MTX(ump), MA_OWNED); 13790 UFS_UNLOCK(ump); 13791 error = ffs_update(vp, 1); 13792 if (error != 0 || MOUNTEDSOFTDEP(mp) == 0) { 13793 UFS_LOCK(ump); 13794 return (0); 13795 } 13796 /* 13797 * If we are in need of resources, start by cleaning up 13798 * any block removals associated with our inode. 13799 */ 13800 ACQUIRE_LOCK(ump); 13801 process_removes(vp); 13802 process_truncates(vp); 13803 FREE_LOCK(ump); 13804 /* 13805 * Now clean up at least as many resources as we will need. 13806 * 13807 * When requested to clean up inodes, the number that are needed 13808 * is set by the number of simultaneous writers (mnt_writeopcount) 13809 * plus a bit of slop (2) in case some more writers show up while 13810 * we are cleaning. 13811 * 13812 * When requested to free up space, the amount of space that 13813 * we need is enough blocks to allocate a full-sized segment 13814 * (fs_contigsumsize). The number of such segments that will 13815 * be needed is set by the number of simultaneous writers 13816 * (mnt_writeopcount) plus a bit of slop (2) in case some more 13817 * writers show up while we are cleaning. 13818 * 13819 * Additionally, if we are unpriviledged and allocating space, 13820 * we need to ensure that we clean up enough blocks to get the 13821 * needed number of blocks over the threshold of the minimum 13822 * number of blocks required to be kept free by the filesystem 13823 * (fs_minfree). 13824 */ 13825 if (resource == FLUSH_INODES_WAIT) { 13826 needed = vfs_mount_fetch_counter(vp->v_mount, 13827 MNT_COUNT_WRITEOPCOUNT) + 2; 13828 } else if (resource == FLUSH_BLOCKS_WAIT) { 13829 needed = (vfs_mount_fetch_counter(vp->v_mount, 13830 MNT_COUNT_WRITEOPCOUNT) + 2) * fs->fs_contigsumsize; 13831 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE)) 13832 needed += fragstoblks(fs, 13833 roundup((fs->fs_dsize * fs->fs_minfree / 100) - 13834 fs->fs_cstotal.cs_nffree, fs->fs_frag)); 13835 } else { 13836 printf("softdep_request_cleanup: Unknown resource type %d\n", 13837 resource); 13838 UFS_LOCK(ump); 13839 return (0); 13840 } 13841 starttime = time_second; 13842retry: 13843 if (resource == FLUSH_BLOCKS_WAIT && 13844 fs->fs_cstotal.cs_nbfree <= needed) 13845 softdep_send_speedup(ump, needed * fs->fs_bsize, 13846 BIO_SPEEDUP_TRIM); 13847 if ((resource == FLUSH_BLOCKS_WAIT && ump->softdep_on_worklist > 0 && 13848 fs->fs_cstotal.cs_nbfree <= needed) || 13849 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 13850 fs->fs_cstotal.cs_nifree <= needed)) { 13851 ACQUIRE_LOCK(ump); 13852 if (ump->softdep_on_worklist > 0 && 13853 process_worklist_item(UFSTOVFS(ump), 13854 ump->softdep_on_worklist, LK_NOWAIT) != 0) 13855 stat_worklist_push += 1; 13856 FREE_LOCK(ump); 13857 } 13858 13859 /* 13860 * Check that there are vnodes pending inactivation. As they 13861 * have been unlinked, inactivating them will free up their 13862 * inodes. 13863 */ 13864 ACQUIRE_LOCK(ump); 13865 if (resource == FLUSH_INODES_WAIT && 13866 fs->fs_cstotal.cs_nifree <= needed && 13867 fs->fs_pendinginodes <= needed) { 13868 if ((ump->um_softdep->sd_flags & FLUSH_DI_ACTIVE) == 0) { 13869 ump->um_softdep->sd_flags |= FLUSH_DI_ACTIVE; 13870 FREE_LOCK(ump); 13871 softdep_request_cleanup_inactivate(mp); 13872 ACQUIRE_LOCK(ump); 13873 ump->um_softdep->sd_flags &= ~FLUSH_DI_ACTIVE; 13874 wakeup(&ump->um_softdep->sd_flags); 13875 } else { 13876 while ((ump->um_softdep->sd_flags & 13877 FLUSH_DI_ACTIVE) != 0) { 13878 msleep(&ump->um_softdep->sd_flags, 13879 LOCK_PTR(ump), PVM, "ffsvina", hz); 13880 } 13881 } 13882 } 13883 FREE_LOCK(ump); 13884 13885 /* 13886 * If we still need resources and there are no more worklist 13887 * entries to process to obtain them, we have to start flushing 13888 * the dirty vnodes to force the release of additional requests 13889 * to the worklist that we can then process to reap addition 13890 * resources. We walk the vnodes associated with the mount point 13891 * until we get the needed worklist requests that we can reap. 13892 * 13893 * If there are several threads all needing to clean the same 13894 * mount point, only one is allowed to walk the mount list. 13895 * When several threads all try to walk the same mount list, 13896 * they end up competing with each other and often end up in 13897 * livelock. This approach ensures that forward progress is 13898 * made at the cost of occational ENOSPC errors being returned 13899 * that might otherwise have been avoided. 13900 */ 13901 error = 1; 13902 if ((resource == FLUSH_BLOCKS_WAIT && 13903 fs->fs_cstotal.cs_nbfree <= needed) || 13904 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 13905 fs->fs_cstotal.cs_nifree <= needed)) { 13906 ACQUIRE_LOCK(ump); 13907 if ((ump->um_softdep->sd_flags & FLUSH_RC_ACTIVE) == 0) { 13908 ump->um_softdep->sd_flags |= FLUSH_RC_ACTIVE; 13909 FREE_LOCK(ump); 13910 failed_vnode = softdep_request_cleanup_flush(mp, ump); 13911 ACQUIRE_LOCK(ump); 13912 ump->um_softdep->sd_flags &= ~FLUSH_RC_ACTIVE; 13913 wakeup(&ump->um_softdep->sd_flags); 13914 FREE_LOCK(ump); 13915 if (ump->softdep_on_worklist > 0) { 13916 stat_cleanup_retries += 1; 13917 if (!failed_vnode) 13918 goto retry; 13919 } 13920 } else { 13921 while ((ump->um_softdep->sd_flags & 13922 FLUSH_RC_ACTIVE) != 0) { 13923 msleep(&ump->um_softdep->sd_flags, 13924 LOCK_PTR(ump), PVM, "ffsrca", hz); 13925 } 13926 FREE_LOCK(ump); 13927 error = 0; 13928 } 13929 stat_cleanup_failures += 1; 13930 } 13931 if (time_second - starttime > stat_cleanup_high_delay) 13932 stat_cleanup_high_delay = time_second - starttime; 13933 UFS_LOCK(ump); 13934 return (error); 13935} 13936 13937/* 13938 * Scan the vnodes for the specified mount point flushing out any 13939 * vnodes that can be locked without waiting. Finally, try to flush 13940 * the device associated with the mount point if it can be locked 13941 * without waiting. 13942 * 13943 * We return 0 if we were able to lock every vnode in our scan. 13944 * If we had to skip one or more vnodes, we return 1. 13945 */ 13946static int 13947softdep_request_cleanup_flush(mp, ump) 13948 struct mount *mp; 13949 struct ufsmount *ump; 13950{ 13951 struct thread *td; 13952 struct vnode *lvp, *mvp; 13953 int failed_vnode; 13954 13955 failed_vnode = 0; 13956 td = curthread; 13957 MNT_VNODE_FOREACH_ALL(lvp, mp, mvp) { 13958 if (TAILQ_FIRST(&lvp->v_bufobj.bo_dirty.bv_hd) == 0) { 13959 VI_UNLOCK(lvp); 13960 continue; 13961 } 13962 if (vget(lvp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT) != 0) { 13963 failed_vnode = 1; 13964 continue; 13965 } 13966 if (lvp->v_vflag & VV_NOSYNC) { /* unlinked */ 13967 vput(lvp); 13968 continue; 13969 } 13970 (void) ffs_syncvnode(lvp, MNT_NOWAIT, 0); 13971 vput(lvp); 13972 } 13973 lvp = ump->um_devvp; 13974 if (vn_lock(lvp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 13975 VOP_FSYNC(lvp, MNT_NOWAIT, td); 13976 VOP_UNLOCK(lvp); 13977 } 13978 return (failed_vnode); 13979} 13980 13981static bool 13982softdep_excess_items(struct ufsmount *ump, int item) 13983{ 13984 13985 KASSERT(item >= 0 && item < D_LAST, ("item %d", item)); 13986 return (dep_current[item] > max_softdeps && 13987 ump->softdep_curdeps[item] > max_softdeps / 13988 stat_flush_threads); 13989} 13990 13991static void 13992schedule_cleanup(struct mount *mp) 13993{ 13994 struct ufsmount *ump; 13995 struct thread *td; 13996 13997 ump = VFSTOUFS(mp); 13998 LOCK_OWNED(ump); 13999 FREE_LOCK(ump); 14000 td = curthread; 14001 if ((td->td_pflags & TDP_KTHREAD) != 0 && 14002 (td->td_proc->p_flag2 & P2_AST_SU) == 0) { 14003 /* 14004 * No ast is delivered to kernel threads, so nobody 14005 * would deref the mp. Some kernel threads 14006 * explicitely check for AST, e.g. NFS daemon does 14007 * this in the serving loop. 14008 */ 14009 return; 14010 } 14011 if (td->td_su != NULL) 14012 vfs_rel(td->td_su); 14013 vfs_ref(mp); 14014 td->td_su = mp; 14015 thread_lock(td); 14016 td->td_flags |= TDF_ASTPENDING; 14017 thread_unlock(td); 14018} 14019 14020static void 14021softdep_ast_cleanup_proc(struct thread *td) 14022{ 14023 struct mount *mp; 14024 struct ufsmount *ump; 14025 int error; 14026 bool req; 14027 14028 while ((mp = td->td_su) != NULL) { 14029 td->td_su = NULL; 14030 error = vfs_busy(mp, MBF_NOWAIT); 14031 vfs_rel(mp); 14032 if (error != 0) 14033 return; 14034 if (ffs_own_mount(mp) && MOUNTEDSOFTDEP(mp)) { 14035 ump = VFSTOUFS(mp); 14036 for (;;) { 14037 req = false; 14038 ACQUIRE_LOCK(ump); 14039 if (softdep_excess_items(ump, D_INODEDEP)) { 14040 req = true; 14041 request_cleanup(mp, FLUSH_INODES); 14042 } 14043 if (softdep_excess_items(ump, D_DIRREM)) { 14044 req = true; 14045 request_cleanup(mp, FLUSH_BLOCKS); 14046 } 14047 FREE_LOCK(ump); 14048 if (softdep_excess_items(ump, D_NEWBLK) || 14049 softdep_excess_items(ump, D_ALLOCDIRECT) || 14050 softdep_excess_items(ump, D_ALLOCINDIR)) { 14051 error = vn_start_write(NULL, &mp, 14052 V_WAIT); 14053 if (error == 0) { 14054 req = true; 14055 VFS_SYNC(mp, MNT_WAIT); 14056 vn_finished_write(mp); 14057 } 14058 } 14059 if ((td->td_pflags & TDP_KTHREAD) != 0 || !req) 14060 break; 14061 } 14062 } 14063 vfs_unbusy(mp); 14064 } 14065 if ((mp = td->td_su) != NULL) { 14066 td->td_su = NULL; 14067 vfs_rel(mp); 14068 } 14069} 14070 14071/* 14072 * If memory utilization has gotten too high, deliberately slow things 14073 * down and speed up the I/O processing. 14074 */ 14075static int 14076request_cleanup(mp, resource) 14077 struct mount *mp; 14078 int resource; 14079{ 14080 struct thread *td = curthread; 14081 struct ufsmount *ump; 14082 14083 ump = VFSTOUFS(mp); 14084 LOCK_OWNED(ump); 14085 /* 14086 * We never hold up the filesystem syncer or buf daemon. 14087 */ 14088 if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF)) 14089 return (0); 14090 /* 14091 * First check to see if the work list has gotten backlogged. 14092 * If it has, co-opt this process to help clean up two entries. 14093 * Because this process may hold inodes locked, we cannot 14094 * handle any remove requests that might block on a locked 14095 * inode as that could lead to deadlock. We set TDP_SOFTDEP 14096 * to avoid recursively processing the worklist. 14097 */ 14098 if (ump->softdep_on_worklist > max_softdeps / 10) { 14099 td->td_pflags |= TDP_SOFTDEP; 14100 process_worklist_item(mp, 2, LK_NOWAIT); 14101 td->td_pflags &= ~TDP_SOFTDEP; 14102 stat_worklist_push += 2; 14103 return(1); 14104 } 14105 /* 14106 * Next, we attempt to speed up the syncer process. If that 14107 * is successful, then we allow the process to continue. 14108 */ 14109 if (softdep_speedup(ump) && 14110 resource != FLUSH_BLOCKS_WAIT && 14111 resource != FLUSH_INODES_WAIT) 14112 return(0); 14113 /* 14114 * If we are resource constrained on inode dependencies, try 14115 * flushing some dirty inodes. Otherwise, we are constrained 14116 * by file deletions, so try accelerating flushes of directories 14117 * with removal dependencies. We would like to do the cleanup 14118 * here, but we probably hold an inode locked at this point and 14119 * that might deadlock against one that we try to clean. So, 14120 * the best that we can do is request the syncer daemon to do 14121 * the cleanup for us. 14122 */ 14123 switch (resource) { 14124 case FLUSH_INODES: 14125 case FLUSH_INODES_WAIT: 14126 ACQUIRE_GBLLOCK(&lk); 14127 stat_ino_limit_push += 1; 14128 req_clear_inodedeps += 1; 14129 FREE_GBLLOCK(&lk); 14130 stat_countp = &stat_ino_limit_hit; 14131 break; 14132 14133 case FLUSH_BLOCKS: 14134 case FLUSH_BLOCKS_WAIT: 14135 ACQUIRE_GBLLOCK(&lk); 14136 stat_blk_limit_push += 1; 14137 req_clear_remove += 1; 14138 FREE_GBLLOCK(&lk); 14139 stat_countp = &stat_blk_limit_hit; 14140 break; 14141 14142 default: 14143 panic("request_cleanup: unknown type"); 14144 } 14145 /* 14146 * Hopefully the syncer daemon will catch up and awaken us. 14147 * We wait at most tickdelay before proceeding in any case. 14148 */ 14149 ACQUIRE_GBLLOCK(&lk); 14150 FREE_LOCK(ump); 14151 proc_waiting += 1; 14152 if (callout_pending(&softdep_callout) == FALSE) 14153 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2, 14154 pause_timer, 0); 14155 14156 if ((td->td_pflags & TDP_KTHREAD) == 0) 14157 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0); 14158 proc_waiting -= 1; 14159 FREE_GBLLOCK(&lk); 14160 ACQUIRE_LOCK(ump); 14161 return (1); 14162} 14163 14164/* 14165 * Awaken processes pausing in request_cleanup and clear proc_waiting 14166 * to indicate that there is no longer a timer running. Pause_timer 14167 * will be called with the global softdep mutex (&lk) locked. 14168 */ 14169static void 14170pause_timer(arg) 14171 void *arg; 14172{ 14173 14174 GBLLOCK_OWNED(&lk); 14175 /* 14176 * The callout_ API has acquired mtx and will hold it around this 14177 * function call. 14178 */ 14179 *stat_countp += proc_waiting; 14180 wakeup(&proc_waiting); 14181} 14182 14183/* 14184 * If requested, try removing inode or removal dependencies. 14185 */ 14186static void 14187check_clear_deps(mp) 14188 struct mount *mp; 14189{ 14190 struct ufsmount *ump; 14191 bool suj_susp; 14192 14193 /* 14194 * Tell the lower layers that any TRIM or WRITE transactions that have 14195 * been delayed for performance reasons should proceed to help alleviate 14196 * the shortage faster. The race between checking req_* and the softdep 14197 * mutex (lk) is fine since this is an advisory operation that at most 14198 * causes deferred work to be done sooner. 14199 */ 14200 ump = VFSTOUFS(mp); 14201 suj_susp = MOUNTEDSUJ(mp) && ump->softdep_jblocks->jb_suspended; 14202 if (req_clear_remove || req_clear_inodedeps || suj_susp) { 14203 FREE_LOCK(ump); 14204 softdep_send_speedup(ump, 0, BIO_SPEEDUP_TRIM | BIO_SPEEDUP_WRITE); 14205 ACQUIRE_LOCK(ump); 14206 } 14207 14208 /* 14209 * If we are suspended, it may be because of our using 14210 * too many inodedeps, so help clear them out. 14211 */ 14212 if (suj_susp) 14213 clear_inodedeps(mp); 14214 14215 /* 14216 * General requests for cleanup of backed up dependencies 14217 */ 14218 ACQUIRE_GBLLOCK(&lk); 14219 if (req_clear_inodedeps) { 14220 req_clear_inodedeps -= 1; 14221 FREE_GBLLOCK(&lk); 14222 clear_inodedeps(mp); 14223 ACQUIRE_GBLLOCK(&lk); 14224 wakeup(&proc_waiting); 14225 } 14226 if (req_clear_remove) { 14227 req_clear_remove -= 1; 14228 FREE_GBLLOCK(&lk); 14229 clear_remove(mp); 14230 ACQUIRE_GBLLOCK(&lk); 14231 wakeup(&proc_waiting); 14232 } 14233 FREE_GBLLOCK(&lk); 14234} 14235 14236/* 14237 * Flush out a directory with at least one removal dependency in an effort to 14238 * reduce the number of dirrem, freefile, and freeblks dependency structures. 14239 */ 14240static void 14241clear_remove(mp) 14242 struct mount *mp; 14243{ 14244 struct pagedep_hashhead *pagedephd; 14245 struct pagedep *pagedep; 14246 struct ufsmount *ump; 14247 struct vnode *vp; 14248 struct bufobj *bo; 14249 int error, cnt; 14250 ino_t ino; 14251 14252 ump = VFSTOUFS(mp); 14253 LOCK_OWNED(ump); 14254 14255 for (cnt = 0; cnt <= ump->pagedep_hash_size; cnt++) { 14256 pagedephd = &ump->pagedep_hashtbl[ump->pagedep_nextclean++]; 14257 if (ump->pagedep_nextclean > ump->pagedep_hash_size) 14258 ump->pagedep_nextclean = 0; 14259 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 14260 if (LIST_EMPTY(&pagedep->pd_dirremhd)) 14261 continue; 14262 ino = pagedep->pd_ino; 14263 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 14264 continue; 14265 FREE_LOCK(ump); 14266 14267 /* 14268 * Let unmount clear deps 14269 */ 14270 error = vfs_busy(mp, MBF_NOWAIT); 14271 if (error != 0) 14272 goto finish_write; 14273 error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 14274 FFSV_FORCEINSMQ); 14275 vfs_unbusy(mp); 14276 if (error != 0) { 14277 softdep_error("clear_remove: vget", error); 14278 goto finish_write; 14279 } 14280 MPASS(VTOI(vp)->i_mode != 0); 14281 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 14282 softdep_error("clear_remove: fsync", error); 14283 bo = &vp->v_bufobj; 14284 BO_LOCK(bo); 14285 drain_output(vp); 14286 BO_UNLOCK(bo); 14287 vput(vp); 14288 finish_write: 14289 vn_finished_write(mp); 14290 ACQUIRE_LOCK(ump); 14291 return; 14292 } 14293 } 14294} 14295 14296/* 14297 * Clear out a block of dirty inodes in an effort to reduce 14298 * the number of inodedep dependency structures. 14299 */ 14300static void 14301clear_inodedeps(mp) 14302 struct mount *mp; 14303{ 14304 struct inodedep_hashhead *inodedephd; 14305 struct inodedep *inodedep; 14306 struct ufsmount *ump; 14307 struct vnode *vp; 14308 struct fs *fs; 14309 int error, cnt; 14310 ino_t firstino, lastino, ino; 14311 14312 ump = VFSTOUFS(mp); 14313 fs = ump->um_fs; 14314 LOCK_OWNED(ump); 14315 /* 14316 * Pick a random inode dependency to be cleared. 14317 * We will then gather up all the inodes in its block 14318 * that have dependencies and flush them out. 14319 */ 14320 for (cnt = 0; cnt <= ump->inodedep_hash_size; cnt++) { 14321 inodedephd = &ump->inodedep_hashtbl[ump->inodedep_nextclean++]; 14322 if (ump->inodedep_nextclean > ump->inodedep_hash_size) 14323 ump->inodedep_nextclean = 0; 14324 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 14325 break; 14326 } 14327 if (inodedep == NULL) 14328 return; 14329 /* 14330 * Find the last inode in the block with dependencies. 14331 */ 14332 firstino = rounddown2(inodedep->id_ino, INOPB(fs)); 14333 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 14334 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0) 14335 break; 14336 /* 14337 * Asynchronously push all but the last inode with dependencies. 14338 * Synchronously push the last inode with dependencies to ensure 14339 * that the inode block gets written to free up the inodedeps. 14340 */ 14341 for (ino = firstino; ino <= lastino; ino++) { 14342 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 14343 continue; 14344 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 14345 continue; 14346 FREE_LOCK(ump); 14347 error = vfs_busy(mp, MBF_NOWAIT); /* Let unmount clear deps */ 14348 if (error != 0) { 14349 vn_finished_write(mp); 14350 ACQUIRE_LOCK(ump); 14351 return; 14352 } 14353 if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 14354 FFSV_FORCEINSMQ)) != 0) { 14355 softdep_error("clear_inodedeps: vget", error); 14356 vfs_unbusy(mp); 14357 vn_finished_write(mp); 14358 ACQUIRE_LOCK(ump); 14359 return; 14360 } 14361 vfs_unbusy(mp); 14362 if (VTOI(vp)->i_mode == 0) { 14363 vgone(vp); 14364 } else if (ino == lastino) { 14365 do { 14366 error = ffs_syncvnode(vp, MNT_WAIT, 0); 14367 } while (error == ERELOOKUP); 14368 if (error != 0) 14369 softdep_error("clear_inodedeps: fsync1", error); 14370 } else { 14371 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 14372 softdep_error("clear_inodedeps: fsync2", error); 14373 BO_LOCK(&vp->v_bufobj); 14374 drain_output(vp); 14375 BO_UNLOCK(&vp->v_bufobj); 14376 } 14377 vput(vp); 14378 vn_finished_write(mp); 14379 ACQUIRE_LOCK(ump); 14380 } 14381} 14382 14383void 14384softdep_buf_append(bp, wkhd) 14385 struct buf *bp; 14386 struct workhead *wkhd; 14387{ 14388 struct worklist *wk; 14389 struct ufsmount *ump; 14390 14391 if ((wk = LIST_FIRST(wkhd)) == NULL) 14392 return; 14393 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 14394 ("softdep_buf_append called on non-softdep filesystem")); 14395 ump = VFSTOUFS(wk->wk_mp); 14396 ACQUIRE_LOCK(ump); 14397 while ((wk = LIST_FIRST(wkhd)) != NULL) { 14398 WORKLIST_REMOVE(wk); 14399 WORKLIST_INSERT(&bp->b_dep, wk); 14400 } 14401 FREE_LOCK(ump); 14402 14403} 14404 14405void 14406softdep_inode_append(ip, cred, wkhd) 14407 struct inode *ip; 14408 struct ucred *cred; 14409 struct workhead *wkhd; 14410{ 14411 struct buf *bp; 14412 struct fs *fs; 14413 struct ufsmount *ump; 14414 int error; 14415 14416 ump = ITOUMP(ip); 14417 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 14418 ("softdep_inode_append called on non-softdep filesystem")); 14419 fs = ump->um_fs; 14420 error = bread(ump->um_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 14421 (int)fs->fs_bsize, cred, &bp); 14422 if (error) { 14423 bqrelse(bp); 14424 softdep_freework(wkhd); 14425 return; 14426 } 14427 softdep_buf_append(bp, wkhd); 14428 bqrelse(bp); 14429} 14430 14431void 14432softdep_freework(wkhd) 14433 struct workhead *wkhd; 14434{ 14435 struct worklist *wk; 14436 struct ufsmount *ump; 14437 14438 if ((wk = LIST_FIRST(wkhd)) == NULL) 14439 return; 14440 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 14441 ("softdep_freework called on non-softdep filesystem")); 14442 ump = VFSTOUFS(wk->wk_mp); 14443 ACQUIRE_LOCK(ump); 14444 handle_jwork(wkhd); 14445 FREE_LOCK(ump); 14446} 14447 14448static struct ufsmount * 14449softdep_bp_to_mp(bp) 14450 struct buf *bp; 14451{ 14452 struct mount *mp; 14453 struct vnode *vp; 14454 14455 if (LIST_EMPTY(&bp->b_dep)) 14456 return (NULL); 14457 vp = bp->b_vp; 14458 KASSERT(vp != NULL, 14459 ("%s, buffer with dependencies lacks vnode", __func__)); 14460 14461 /* 14462 * The ump mount point is stable after we get a correct 14463 * pointer, since bp is locked and this prevents unmount from 14464 * proceeding. But to get to it, we cannot dereference bp->b_dep 14465 * head wk_mp, because we do not yet own SU ump lock and 14466 * workitem might be freed while dereferenced. 14467 */ 14468retry: 14469 switch (vp->v_type) { 14470 case VCHR: 14471 VI_LOCK(vp); 14472 mp = vp->v_type == VCHR ? vp->v_rdev->si_mountpt : NULL; 14473 VI_UNLOCK(vp); 14474 if (mp == NULL) 14475 goto retry; 14476 break; 14477 case VREG: 14478 case VDIR: 14479 case VLNK: 14480 case VFIFO: 14481 case VSOCK: 14482 mp = vp->v_mount; 14483 break; 14484 case VBLK: 14485 vn_printf(vp, "softdep_bp_to_mp: unexpected block device\n"); 14486 /* FALLTHROUGH */ 14487 case VNON: 14488 case VBAD: 14489 case VMARKER: 14490 mp = NULL; 14491 break; 14492 default: 14493 vn_printf(vp, "unknown vnode type"); 14494 mp = NULL; 14495 break; 14496 } 14497 return (VFSTOUFS(mp)); 14498} 14499 14500/* 14501 * Function to determine if the buffer has outstanding dependencies 14502 * that will cause a roll-back if the buffer is written. If wantcount 14503 * is set, return number of dependencies, otherwise just yes or no. 14504 */ 14505static int 14506softdep_count_dependencies(bp, wantcount) 14507 struct buf *bp; 14508 int wantcount; 14509{ 14510 struct worklist *wk; 14511 struct ufsmount *ump; 14512 struct bmsafemap *bmsafemap; 14513 struct freework *freework; 14514 struct inodedep *inodedep; 14515 struct indirdep *indirdep; 14516 struct freeblks *freeblks; 14517 struct allocindir *aip; 14518 struct pagedep *pagedep; 14519 struct dirrem *dirrem; 14520 struct newblk *newblk; 14521 struct mkdir *mkdir; 14522 struct diradd *dap; 14523 int i, retval; 14524 14525 ump = softdep_bp_to_mp(bp); 14526 if (ump == NULL) 14527 return (0); 14528 retval = 0; 14529 ACQUIRE_LOCK(ump); 14530 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 14531 switch (wk->wk_type) { 14532 case D_INODEDEP: 14533 inodedep = WK_INODEDEP(wk); 14534 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 14535 /* bitmap allocation dependency */ 14536 retval += 1; 14537 if (!wantcount) 14538 goto out; 14539 } 14540 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 14541 /* direct block pointer dependency */ 14542 retval += 1; 14543 if (!wantcount) 14544 goto out; 14545 } 14546 if (TAILQ_FIRST(&inodedep->id_extupdt)) { 14547 /* direct block pointer dependency */ 14548 retval += 1; 14549 if (!wantcount) 14550 goto out; 14551 } 14552 if (TAILQ_FIRST(&inodedep->id_inoreflst)) { 14553 /* Add reference dependency. */ 14554 retval += 1; 14555 if (!wantcount) 14556 goto out; 14557 } 14558 continue; 14559 14560 case D_INDIRDEP: 14561 indirdep = WK_INDIRDEP(wk); 14562 14563 TAILQ_FOREACH(freework, &indirdep->ir_trunc, fw_next) { 14564 /* indirect truncation dependency */ 14565 retval += 1; 14566 if (!wantcount) 14567 goto out; 14568 } 14569 14570 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 14571 /* indirect block pointer dependency */ 14572 retval += 1; 14573 if (!wantcount) 14574 goto out; 14575 } 14576 continue; 14577 14578 case D_PAGEDEP: 14579 pagedep = WK_PAGEDEP(wk); 14580 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 14581 if (LIST_FIRST(&dirrem->dm_jremrefhd)) { 14582 /* Journal remove ref dependency. */ 14583 retval += 1; 14584 if (!wantcount) 14585 goto out; 14586 } 14587 } 14588 for (i = 0; i < DAHASHSZ; i++) { 14589 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 14590 /* directory entry dependency */ 14591 retval += 1; 14592 if (!wantcount) 14593 goto out; 14594 } 14595 } 14596 continue; 14597 14598 case D_BMSAFEMAP: 14599 bmsafemap = WK_BMSAFEMAP(wk); 14600 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd)) { 14601 /* Add reference dependency. */ 14602 retval += 1; 14603 if (!wantcount) 14604 goto out; 14605 } 14606 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd)) { 14607 /* Allocate block dependency. */ 14608 retval += 1; 14609 if (!wantcount) 14610 goto out; 14611 } 14612 continue; 14613 14614 case D_FREEBLKS: 14615 freeblks = WK_FREEBLKS(wk); 14616 if (LIST_FIRST(&freeblks->fb_jblkdephd)) { 14617 /* Freeblk journal dependency. */ 14618 retval += 1; 14619 if (!wantcount) 14620 goto out; 14621 } 14622 continue; 14623 14624 case D_ALLOCDIRECT: 14625 case D_ALLOCINDIR: 14626 newblk = WK_NEWBLK(wk); 14627 if (newblk->nb_jnewblk) { 14628 /* Journal allocate dependency. */ 14629 retval += 1; 14630 if (!wantcount) 14631 goto out; 14632 } 14633 continue; 14634 14635 case D_MKDIR: 14636 mkdir = WK_MKDIR(wk); 14637 if (mkdir->md_jaddref) { 14638 /* Journal reference dependency. */ 14639 retval += 1; 14640 if (!wantcount) 14641 goto out; 14642 } 14643 continue; 14644 14645 case D_FREEWORK: 14646 case D_FREEDEP: 14647 case D_JSEGDEP: 14648 case D_JSEG: 14649 case D_SBDEP: 14650 /* never a dependency on these blocks */ 14651 continue; 14652 14653 default: 14654 panic("softdep_count_dependencies: Unexpected type %s", 14655 TYPENAME(wk->wk_type)); 14656 /* NOTREACHED */ 14657 } 14658 } 14659out: 14660 FREE_LOCK(ump); 14661 return (retval); 14662} 14663 14664/* 14665 * Acquire exclusive access to a buffer. 14666 * Must be called with a locked mtx parameter. 14667 * Return acquired buffer or NULL on failure. 14668 */ 14669static struct buf * 14670getdirtybuf(bp, lock, waitfor) 14671 struct buf *bp; 14672 struct rwlock *lock; 14673 int waitfor; 14674{ 14675 int error; 14676 14677 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) { 14678 if (waitfor != MNT_WAIT) 14679 return (NULL); 14680 error = BUF_LOCK(bp, 14681 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, lock); 14682 /* 14683 * Even if we successfully acquire bp here, we have dropped 14684 * lock, which may violates our guarantee. 14685 */ 14686 if (error == 0) 14687 BUF_UNLOCK(bp); 14688 else if (error != ENOLCK) 14689 panic("getdirtybuf: inconsistent lock: %d", error); 14690 rw_wlock(lock); 14691 return (NULL); 14692 } 14693 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 14694 if (lock != BO_LOCKPTR(bp->b_bufobj) && waitfor == MNT_WAIT) { 14695 rw_wunlock(lock); 14696 BO_LOCK(bp->b_bufobj); 14697 BUF_UNLOCK(bp); 14698 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 14699 bp->b_vflags |= BV_BKGRDWAIT; 14700 msleep(&bp->b_xflags, BO_LOCKPTR(bp->b_bufobj), 14701 PRIBIO | PDROP, "getbuf", 0); 14702 } else 14703 BO_UNLOCK(bp->b_bufobj); 14704 rw_wlock(lock); 14705 return (NULL); 14706 } 14707 BUF_UNLOCK(bp); 14708 if (waitfor != MNT_WAIT) 14709 return (NULL); 14710#ifdef DEBUG_VFS_LOCKS 14711 if (bp->b_vp->v_type != VCHR) 14712 ASSERT_BO_WLOCKED(bp->b_bufobj); 14713#endif 14714 bp->b_vflags |= BV_BKGRDWAIT; 14715 rw_sleep(&bp->b_xflags, lock, PRIBIO, "getbuf", 0); 14716 return (NULL); 14717 } 14718 if ((bp->b_flags & B_DELWRI) == 0) { 14719 BUF_UNLOCK(bp); 14720 return (NULL); 14721 } 14722 bremfree(bp); 14723 return (bp); 14724} 14725 14726/* 14727 * Check if it is safe to suspend the file system now. On entry, 14728 * the vnode interlock for devvp should be held. Return 0 with 14729 * the mount interlock held if the file system can be suspended now, 14730 * otherwise return EAGAIN with the mount interlock held. 14731 */ 14732int 14733softdep_check_suspend(struct mount *mp, 14734 struct vnode *devvp, 14735 int softdep_depcnt, 14736 int softdep_accdepcnt, 14737 int secondary_writes, 14738 int secondary_accwrites) 14739{ 14740 struct bufobj *bo; 14741 struct ufsmount *ump; 14742 struct inodedep *inodedep; 14743 int error, unlinked; 14744 14745 bo = &devvp->v_bufobj; 14746 ASSERT_BO_WLOCKED(bo); 14747 14748 /* 14749 * If we are not running with soft updates, then we need only 14750 * deal with secondary writes as we try to suspend. 14751 */ 14752 if (MOUNTEDSOFTDEP(mp) == 0) { 14753 MNT_ILOCK(mp); 14754 while (mp->mnt_secondary_writes != 0) { 14755 BO_UNLOCK(bo); 14756 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), 14757 (PUSER - 1) | PDROP, "secwr", 0); 14758 BO_LOCK(bo); 14759 MNT_ILOCK(mp); 14760 } 14761 14762 /* 14763 * Reasons for needing more work before suspend: 14764 * - Dirty buffers on devvp. 14765 * - Secondary writes occurred after start of vnode sync loop 14766 */ 14767 error = 0; 14768 if (bo->bo_numoutput > 0 || 14769 bo->bo_dirty.bv_cnt > 0 || 14770 secondary_writes != 0 || 14771 mp->mnt_secondary_writes != 0 || 14772 secondary_accwrites != mp->mnt_secondary_accwrites) 14773 error = EAGAIN; 14774 BO_UNLOCK(bo); 14775 return (error); 14776 } 14777 14778 /* 14779 * If we are running with soft updates, then we need to coordinate 14780 * with them as we try to suspend. 14781 */ 14782 ump = VFSTOUFS(mp); 14783 for (;;) { 14784 if (!TRY_ACQUIRE_LOCK(ump)) { 14785 BO_UNLOCK(bo); 14786 ACQUIRE_LOCK(ump); 14787 FREE_LOCK(ump); 14788 BO_LOCK(bo); 14789 continue; 14790 } 14791 MNT_ILOCK(mp); 14792 if (mp->mnt_secondary_writes != 0) { 14793 FREE_LOCK(ump); 14794 BO_UNLOCK(bo); 14795 msleep(&mp->mnt_secondary_writes, 14796 MNT_MTX(mp), 14797 (PUSER - 1) | PDROP, "secwr", 0); 14798 BO_LOCK(bo); 14799 continue; 14800 } 14801 break; 14802 } 14803 14804 unlinked = 0; 14805 if (MOUNTEDSUJ(mp)) { 14806 for (inodedep = TAILQ_FIRST(&ump->softdep_unlinked); 14807 inodedep != NULL; 14808 inodedep = TAILQ_NEXT(inodedep, id_unlinked)) { 14809 if ((inodedep->id_state & (UNLINKED | UNLINKLINKS | 14810 UNLINKONLIST)) != (UNLINKED | UNLINKLINKS | 14811 UNLINKONLIST) || 14812 !check_inodedep_free(inodedep)) 14813 continue; 14814 unlinked++; 14815 } 14816 } 14817 14818 /* 14819 * Reasons for needing more work before suspend: 14820 * - Dirty buffers on devvp. 14821 * - Softdep activity occurred after start of vnode sync loop 14822 * - Secondary writes occurred after start of vnode sync loop 14823 */ 14824 error = 0; 14825 if (bo->bo_numoutput > 0 || 14826 bo->bo_dirty.bv_cnt > 0 || 14827 softdep_depcnt != unlinked || 14828 ump->softdep_deps != unlinked || 14829 softdep_accdepcnt != ump->softdep_accdeps || 14830 secondary_writes != 0 || 14831 mp->mnt_secondary_writes != 0 || 14832 secondary_accwrites != mp->mnt_secondary_accwrites) 14833 error = EAGAIN; 14834 FREE_LOCK(ump); 14835 BO_UNLOCK(bo); 14836 return (error); 14837} 14838 14839/* 14840 * Get the number of dependency structures for the file system, both 14841 * the current number and the total number allocated. These will 14842 * later be used to detect that softdep processing has occurred. 14843 */ 14844void 14845softdep_get_depcounts(struct mount *mp, 14846 int *softdep_depsp, 14847 int *softdep_accdepsp) 14848{ 14849 struct ufsmount *ump; 14850 14851 if (MOUNTEDSOFTDEP(mp) == 0) { 14852 *softdep_depsp = 0; 14853 *softdep_accdepsp = 0; 14854 return; 14855 } 14856 ump = VFSTOUFS(mp); 14857 ACQUIRE_LOCK(ump); 14858 *softdep_depsp = ump->softdep_deps; 14859 *softdep_accdepsp = ump->softdep_accdeps; 14860 FREE_LOCK(ump); 14861} 14862 14863/* 14864 * Wait for pending output on a vnode to complete. 14865 */ 14866static void 14867drain_output(vp) 14868 struct vnode *vp; 14869{ 14870 14871 ASSERT_VOP_LOCKED(vp, "drain_output"); 14872 (void)bufobj_wwait(&vp->v_bufobj, 0, 0); 14873} 14874 14875/* 14876 * Called whenever a buffer that is being invalidated or reallocated 14877 * contains dependencies. This should only happen if an I/O error has 14878 * occurred. The routine is called with the buffer locked. 14879 */ 14880static void 14881softdep_deallocate_dependencies(bp) 14882 struct buf *bp; 14883{ 14884 14885 if ((bp->b_ioflags & BIO_ERROR) == 0) 14886 panic("softdep_deallocate_dependencies: dangling deps"); 14887 if (bp->b_vp != NULL && bp->b_vp->v_mount != NULL) 14888 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 14889 else 14890 printf("softdep_deallocate_dependencies: " 14891 "got error %d while accessing filesystem\n", bp->b_error); 14892 if (bp->b_error != ENXIO) 14893 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 14894} 14895 14896/* 14897 * Function to handle asynchronous write errors in the filesystem. 14898 */ 14899static void 14900softdep_error(func, error) 14901 char *func; 14902 int error; 14903{ 14904 14905 /* XXX should do something better! */ 14906 printf("%s: got error %d while accessing filesystem\n", func, error); 14907} 14908 14909#ifdef DDB 14910 14911/* exported to ffs_vfsops.c */ 14912extern void db_print_ffs(struct ufsmount *ump); 14913void 14914db_print_ffs(struct ufsmount *ump) 14915{ 14916 db_printf("mp %p (%s) devvp %p\n", ump->um_mountp, 14917 ump->um_mountp->mnt_stat.f_mntonname, ump->um_devvp); 14918 db_printf(" fs %p su_wl %d su_deps %d su_req %d\n", 14919 ump->um_fs, ump->softdep_on_worklist, 14920 ump->softdep_deps, ump->softdep_req); 14921} 14922 14923static void 14924worklist_print(struct worklist *wk, int verbose) 14925{ 14926 14927 if (!verbose) { 14928 db_printf("%s: %p state 0x%b\n", TYPENAME(wk->wk_type), wk, 14929 (u_int)wk->wk_state, PRINT_SOFTDEP_FLAGS); 14930 return; 14931 } 14932 db_printf("worklist: %p type %s state 0x%b next %p\n ", wk, 14933 TYPENAME(wk->wk_type), (u_int)wk->wk_state, PRINT_SOFTDEP_FLAGS, 14934 LIST_NEXT(wk, wk_list)); 14935 db_print_ffs(VFSTOUFS(wk->wk_mp)); 14936} 14937 14938static void 14939inodedep_print(struct inodedep *inodedep, int verbose) 14940{ 14941 14942 worklist_print(&inodedep->id_list, 0); 14943 db_printf(" fs %p ino %jd inoblk %jd delta %jd nlink %jd\n", 14944 inodedep->id_fs, 14945 (intmax_t)inodedep->id_ino, 14946 (intmax_t)fsbtodb(inodedep->id_fs, 14947 ino_to_fsba(inodedep->id_fs, inodedep->id_ino)), 14948 (intmax_t)inodedep->id_nlinkdelta, 14949 (intmax_t)inodedep->id_savednlink); 14950 14951 if (verbose == 0) 14952 return; 14953 14954 db_printf(" bmsafemap %p, mkdiradd %p, inoreflst %p\n", 14955 inodedep->id_bmsafemap, 14956 inodedep->id_mkdiradd, 14957 TAILQ_FIRST(&inodedep->id_inoreflst)); 14958 db_printf(" dirremhd %p, pendinghd %p, bufwait %p\n", 14959 LIST_FIRST(&inodedep->id_dirremhd), 14960 LIST_FIRST(&inodedep->id_pendinghd), 14961 LIST_FIRST(&inodedep->id_bufwait)); 14962 db_printf(" inowait %p, inoupdt %p, newinoupdt %p\n", 14963 LIST_FIRST(&inodedep->id_inowait), 14964 TAILQ_FIRST(&inodedep->id_inoupdt), 14965 TAILQ_FIRST(&inodedep->id_newinoupdt)); 14966 db_printf(" extupdt %p, newextupdt %p, freeblklst %p\n", 14967 TAILQ_FIRST(&inodedep->id_extupdt), 14968 TAILQ_FIRST(&inodedep->id_newextupdt), 14969 TAILQ_FIRST(&inodedep->id_freeblklst)); 14970 db_printf(" saveino %p, savedsize %jd, savedextsize %jd\n", 14971 inodedep->id_savedino1, 14972 (intmax_t)inodedep->id_savedsize, 14973 (intmax_t)inodedep->id_savedextsize); 14974} 14975 14976static void 14977newblk_print(struct newblk *nbp) 14978{ 14979 14980 worklist_print(&nbp->nb_list, 0); 14981 db_printf(" newblkno %jd\n", (intmax_t)nbp->nb_newblkno); 14982 db_printf(" jnewblk %p, bmsafemap %p, freefrag %p\n", 14983 &nbp->nb_jnewblk, 14984 &nbp->nb_bmsafemap, 14985 &nbp->nb_freefrag); 14986 db_printf(" indirdeps %p, newdirblk %p, jwork %p\n", 14987 LIST_FIRST(&nbp->nb_indirdeps), 14988 LIST_FIRST(&nbp->nb_newdirblk), 14989 LIST_FIRST(&nbp->nb_jwork)); 14990} 14991 14992static void 14993allocdirect_print(struct allocdirect *adp) 14994{ 14995 14996 newblk_print(&adp->ad_block); 14997 db_printf(" oldblkno %jd, oldsize %ld, newsize %ld\n", 14998 adp->ad_oldblkno, adp->ad_oldsize, adp->ad_newsize); 14999 db_printf(" offset %d, inodedep %p\n", 15000 adp->ad_offset, adp->ad_inodedep); 15001} 15002 15003static void 15004allocindir_print(struct allocindir *aip) 15005{ 15006 15007 newblk_print(&aip->ai_block); 15008 db_printf(" oldblkno %jd, lbn %jd\n", 15009 (intmax_t)aip->ai_oldblkno, (intmax_t)aip->ai_lbn); 15010 db_printf(" offset %d, indirdep %p\n", 15011 aip->ai_offset, aip->ai_indirdep); 15012} 15013 15014static void 15015mkdir_print(struct mkdir *mkdir) 15016{ 15017 15018 worklist_print(&mkdir->md_list, 0); 15019 db_printf(" diradd %p, jaddref %p, buf %p\n", 15020 mkdir->md_diradd, mkdir->md_jaddref, mkdir->md_buf); 15021} 15022 15023DB_SHOW_COMMAND(sd_inodedep, db_show_sd_inodedep) 15024{ 15025 15026 if (have_addr == 0) { 15027 db_printf("inodedep address required\n"); 15028 return; 15029 } 15030 inodedep_print((struct inodedep*)addr, 1); 15031} 15032 15033DB_SHOW_COMMAND(sd_allinodedeps, db_show_sd_allinodedeps) 15034{ 15035 struct inodedep_hashhead *inodedephd; 15036 struct inodedep *inodedep; 15037 struct ufsmount *ump; 15038 int cnt; 15039 15040 if (have_addr == 0) { 15041 db_printf("ufsmount address required\n"); 15042 return; 15043 } 15044 ump = (struct ufsmount *)addr; 15045 for (cnt = 0; cnt < ump->inodedep_hash_size; cnt++) { 15046 inodedephd = &ump->inodedep_hashtbl[cnt]; 15047 LIST_FOREACH(inodedep, inodedephd, id_hash) { 15048 inodedep_print(inodedep, 0); 15049 } 15050 } 15051} 15052 15053DB_SHOW_COMMAND(sd_worklist, db_show_sd_worklist) 15054{ 15055 15056 if (have_addr == 0) { 15057 db_printf("worklist address required\n"); 15058 return; 15059 } 15060 worklist_print((struct worklist *)addr, 1); 15061} 15062 15063DB_SHOW_COMMAND(sd_workhead, db_show_sd_workhead) 15064{ 15065 struct worklist *wk; 15066 struct workhead *wkhd; 15067 15068 if (have_addr == 0) { 15069 db_printf("worklist address required " 15070 "(for example value in bp->b_dep)\n"); 15071 return; 15072 } 15073 /* 15074 * We often do not have the address of the worklist head but 15075 * instead a pointer to its first entry (e.g., we have the 15076 * contents of bp->b_dep rather than &bp->b_dep). But the back 15077 * pointer of bp->b_dep will point at the head of the list, so 15078 * we cheat and use that instead. If we are in the middle of 15079 * a list we will still get the same result, so nothing 15080 * unexpected will result. 15081 */ 15082 wk = (struct worklist *)addr; 15083 if (wk == NULL) 15084 return; 15085 wkhd = (struct workhead *)wk->wk_list.le_prev; 15086 LIST_FOREACH(wk, wkhd, wk_list) { 15087 switch(wk->wk_type) { 15088 case D_INODEDEP: 15089 inodedep_print(WK_INODEDEP(wk), 0); 15090 continue; 15091 case D_ALLOCDIRECT: 15092 allocdirect_print(WK_ALLOCDIRECT(wk)); 15093 continue; 15094 case D_ALLOCINDIR: 15095 allocindir_print(WK_ALLOCINDIR(wk)); 15096 continue; 15097 case D_MKDIR: 15098 mkdir_print(WK_MKDIR(wk)); 15099 continue; 15100 default: 15101 worklist_print(wk, 0); 15102 continue; 15103 } 15104 } 15105} 15106 15107DB_SHOW_COMMAND(sd_mkdir, db_show_sd_mkdir) 15108{ 15109 if (have_addr == 0) { 15110 db_printf("mkdir address required\n"); 15111 return; 15112 } 15113 mkdir_print((struct mkdir *)addr); 15114} 15115 15116DB_SHOW_COMMAND(sd_mkdir_list, db_show_sd_mkdir_list) 15117{ 15118 struct mkdirlist *mkdirlisthd; 15119 struct mkdir *mkdir; 15120 15121 if (have_addr == 0) { 15122 db_printf("mkdir listhead address required\n"); 15123 return; 15124 } 15125 mkdirlisthd = (struct mkdirlist *)addr; 15126 LIST_FOREACH(mkdir, mkdirlisthd, md_mkdirs) { 15127 mkdir_print(mkdir); 15128 if (mkdir->md_diradd != NULL) { 15129 db_printf(" "); 15130 worklist_print(&mkdir->md_diradd->da_list, 0); 15131 } 15132 if (mkdir->md_jaddref != NULL) { 15133 db_printf(" "); 15134 worklist_print(&mkdir->md_jaddref->ja_list, 0); 15135 } 15136 } 15137} 15138 15139DB_SHOW_COMMAND(sd_allocdirect, db_show_sd_allocdirect) 15140{ 15141 if (have_addr == 0) { 15142 db_printf("allocdirect address required\n"); 15143 return; 15144 } 15145 allocdirect_print((struct allocdirect *)addr); 15146} 15147 15148DB_SHOW_COMMAND(sd_allocindir, db_show_sd_allocindir) 15149{ 15150 if (have_addr == 0) { 15151 db_printf("allocindir address required\n"); 15152 return; 15153 } 15154 allocindir_print((struct allocindir *)addr); 15155} 15156 15157#endif /* DDB */ 15158 15159#endif /* SOFTUPDATES */ 15160