vfs_bio.c revision 44433
1/* 2 * Copyright (c) 1994,1997 John S. Dyson 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice immediately at the beginning of the file, without modification, 10 * this list of conditions, and the following disclaimer. 11 * 2. Absolutely no warranty of function or purpose is made by the author 12 * John S. Dyson. 13 * 14 * $Id: vfs_bio.c,v 1.199 1999/01/27 21:49:58 dillon Exp $ 15 */ 16 17/* 18 * this file contains a new buffer I/O scheme implementing a coherent 19 * VM object and buffer cache scheme. Pains have been taken to make 20 * sure that the performance degradation associated with schemes such 21 * as this is not realized. 22 * 23 * Author: John S. Dyson 24 * Significant help during the development and debugging phases 25 * had been provided by David Greenman, also of the FreeBSD core team. 26 * 27 * see man buf(9) for more info. 28 */ 29 30#define VMIO 31#include <sys/param.h> 32#include <sys/systm.h> 33#include <sys/sysproto.h> 34#include <sys/kernel.h> 35#include <sys/sysctl.h> 36#include <sys/proc.h> 37#include <sys/vnode.h> 38#include <sys/vmmeter.h> 39#include <sys/lock.h> 40#include <miscfs/specfs/specdev.h> 41#include <vm/vm.h> 42#include <vm/vm_param.h> 43#include <vm/vm_prot.h> 44#include <vm/vm_kern.h> 45#include <vm/vm_pageout.h> 46#include <vm/vm_page.h> 47#include <vm/vm_object.h> 48#include <vm/vm_extern.h> 49#include <vm/vm_map.h> 50#include <sys/buf.h> 51#include <sys/mount.h> 52#include <sys/malloc.h> 53#include <sys/resourcevar.h> 54 55static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 56 57struct bio_ops bioops; /* I/O operation notification */ 58 59#if 0 /* replaced bu sched_sync */ 60static void vfs_update __P((void)); 61static struct proc *updateproc; 62static struct kproc_desc up_kp = { 63 "update", 64 vfs_update, 65 &updateproc 66}; 67SYSINIT_KT(update, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 68#endif 69 70struct buf *buf; /* buffer header pool */ 71struct swqueue bswlist; 72 73static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 74 vm_offset_t to); 75static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 76 vm_offset_t to); 77static void vfs_buf_set_valid(struct buf *bp, vm_ooffset_t foff, 78 vm_offset_t off, vm_offset_t size, 79 vm_page_t m); 80static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 81 int pageno, vm_page_t m); 82static void vfs_clean_pages(struct buf * bp); 83static void vfs_setdirty(struct buf *bp); 84static void vfs_vmio_release(struct buf *bp); 85static void flushdirtybuffers(int slpflag, int slptimeo); 86 87int needsbuffer; 88 89/* 90 * Internal update daemon, process 3 91 * The variable vfs_update_wakeup allows for internal syncs. 92 */ 93int vfs_update_wakeup; 94 95 96/* 97 * buffers base kva 98 */ 99 100/* 101 * bogus page -- for I/O to/from partially complete buffers 102 * this is a temporary solution to the problem, but it is not 103 * really that bad. it would be better to split the buffer 104 * for input in the case of buffers partially already in memory, 105 * but the code is intricate enough already. 106 */ 107vm_page_t bogus_page; 108static vm_offset_t bogus_offset; 109 110static int bufspace, maxbufspace, vmiospace, maxvmiobufspace, 111 bufmallocspace, maxbufmallocspace; 112int numdirtybuffers; 113static int lodirtybuffers, hidirtybuffers; 114static int numfreebuffers, lofreebuffers, hifreebuffers; 115static int kvafreespace; 116 117SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, 118 &numdirtybuffers, 0, ""); 119SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, 120 &lodirtybuffers, 0, ""); 121SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, 122 &hidirtybuffers, 0, ""); 123SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, 124 &numfreebuffers, 0, ""); 125SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, 126 &lofreebuffers, 0, ""); 127SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, 128 &hifreebuffers, 0, ""); 129SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RW, 130 &maxbufspace, 0, ""); 131SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, 132 &bufspace, 0, ""); 133SYSCTL_INT(_vfs, OID_AUTO, maxvmiobufspace, CTLFLAG_RW, 134 &maxvmiobufspace, 0, ""); 135SYSCTL_INT(_vfs, OID_AUTO, vmiospace, CTLFLAG_RD, 136 &vmiospace, 0, ""); 137SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, 138 &maxbufmallocspace, 0, ""); 139SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, 140 &bufmallocspace, 0, ""); 141SYSCTL_INT(_vfs, OID_AUTO, kvafreespace, CTLFLAG_RD, 142 &kvafreespace, 0, ""); 143 144static LIST_HEAD(bufhashhdr, buf) bufhashtbl[BUFHSZ], invalhash; 145struct bqueues bufqueues[BUFFER_QUEUES] = { { 0 } }; 146 147extern int vm_swap_size; 148 149#define BUF_MAXUSE 24 150 151#define VFS_BIO_NEED_ANY 1 152#define VFS_BIO_NEED_LOWLIMIT 2 153#define VFS_BIO_NEED_FREE 4 154 155/* 156 * Initialize buffer headers and related structures. 157 */ 158void 159bufinit() 160{ 161 struct buf *bp; 162 int i; 163 164 TAILQ_INIT(&bswlist); 165 LIST_INIT(&invalhash); 166 167 /* first, make a null hash table */ 168 for (i = 0; i < BUFHSZ; i++) 169 LIST_INIT(&bufhashtbl[i]); 170 171 /* next, make a null set of free lists */ 172 for (i = 0; i < BUFFER_QUEUES; i++) 173 TAILQ_INIT(&bufqueues[i]); 174 175 /* finally, initialize each buffer header and stick on empty q */ 176 for (i = 0; i < nbuf; i++) { 177 bp = &buf[i]; 178 bzero(bp, sizeof *bp); 179 bp->b_flags = B_INVAL; /* we're just an empty header */ 180 bp->b_dev = NODEV; 181 bp->b_rcred = NOCRED; 182 bp->b_wcred = NOCRED; 183 bp->b_qindex = QUEUE_EMPTY; 184 bp->b_xflags = 0; 185 LIST_INIT(&bp->b_dep); 186 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 187 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 188 } 189/* 190 * maxbufspace is currently calculated to support all filesystem blocks 191 * to be 8K. If you happen to use a 16K filesystem, the size of the buffer 192 * cache is still the same as it would be for 8K filesystems. This 193 * keeps the size of the buffer cache "in check" for big block filesystems. 194 */ 195 maxbufspace = (nbuf + 8) * DFLTBSIZE; 196/* 197 * reserve 1/3 of the buffers for metadata (VDIR) which might not be VMIO'ed 198 */ 199 maxvmiobufspace = 2 * maxbufspace / 3; 200/* 201 * Limit the amount of malloc memory since it is wired permanently into 202 * the kernel space. Even though this is accounted for in the buffer 203 * allocation, we don't want the malloced region to grow uncontrolled. 204 * The malloc scheme improves memory utilization significantly on average 205 * (small) directories. 206 */ 207 maxbufmallocspace = maxbufspace / 20; 208 209/* 210 * Remove the probability of deadlock conditions by limiting the 211 * number of dirty buffers. 212 */ 213 hidirtybuffers = nbuf / 8 + 20; 214 lodirtybuffers = nbuf / 16 + 10; 215 numdirtybuffers = 0; 216 lofreebuffers = nbuf / 18 + 5; 217 hifreebuffers = 2 * lofreebuffers; 218 numfreebuffers = nbuf; 219 kvafreespace = 0; 220 221 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 222 bogus_page = vm_page_alloc(kernel_object, 223 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 224 VM_ALLOC_NORMAL); 225 226} 227 228/* 229 * Free the kva allocation for a buffer 230 * Must be called only at splbio or higher, 231 * as this is the only locking for buffer_map. 232 */ 233static void 234bfreekva(struct buf * bp) 235{ 236 if (bp->b_kvasize == 0) 237 return; 238 239 vm_map_delete(buffer_map, 240 (vm_offset_t) bp->b_kvabase, 241 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 242 243 bp->b_kvasize = 0; 244 245} 246 247/* 248 * remove the buffer from the appropriate free list 249 */ 250void 251bremfree(struct buf * bp) 252{ 253 int s = splbio(); 254 255 if (bp->b_qindex != QUEUE_NONE) { 256 if (bp->b_qindex == QUEUE_EMPTY) { 257 kvafreespace -= bp->b_kvasize; 258 } 259 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 260 bp->b_qindex = QUEUE_NONE; 261 } else { 262#if !defined(MAX_PERF) 263 panic("bremfree: removing a buffer when not on a queue"); 264#endif 265 } 266 if ((bp->b_flags & B_INVAL) || 267 (bp->b_flags & (B_DELWRI|B_LOCKED)) == 0) 268 --numfreebuffers; 269 splx(s); 270} 271 272 273/* 274 * Get a buffer with the specified data. Look in the cache first. 275 */ 276int 277bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 278 struct buf ** bpp) 279{ 280 struct buf *bp; 281 282 bp = getblk(vp, blkno, size, 0, 0); 283 *bpp = bp; 284 285 /* if not found in cache, do some I/O */ 286 if ((bp->b_flags & B_CACHE) == 0) { 287 if (curproc != NULL) 288 curproc->p_stats->p_ru.ru_inblock++; 289 bp->b_flags |= B_READ; 290 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 291 if (bp->b_rcred == NOCRED) { 292 if (cred != NOCRED) 293 crhold(cred); 294 bp->b_rcred = cred; 295 } 296 vfs_busy_pages(bp, 0); 297 VOP_STRATEGY(vp, bp); 298 return (biowait(bp)); 299 } 300 return (0); 301} 302 303/* 304 * Operates like bread, but also starts asynchronous I/O on 305 * read-ahead blocks. 306 */ 307int 308breadn(struct vnode * vp, daddr_t blkno, int size, 309 daddr_t * rablkno, int *rabsize, 310 int cnt, struct ucred * cred, struct buf ** bpp) 311{ 312 struct buf *bp, *rabp; 313 int i; 314 int rv = 0, readwait = 0; 315 316 *bpp = bp = getblk(vp, blkno, size, 0, 0); 317 318 /* if not found in cache, do some I/O */ 319 if ((bp->b_flags & B_CACHE) == 0) { 320 if (curproc != NULL) 321 curproc->p_stats->p_ru.ru_inblock++; 322 bp->b_flags |= B_READ; 323 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 324 if (bp->b_rcred == NOCRED) { 325 if (cred != NOCRED) 326 crhold(cred); 327 bp->b_rcred = cred; 328 } 329 vfs_busy_pages(bp, 0); 330 VOP_STRATEGY(vp, bp); 331 ++readwait; 332 } 333 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 334 if (inmem(vp, *rablkno)) 335 continue; 336 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 337 338 if ((rabp->b_flags & B_CACHE) == 0) { 339 if (curproc != NULL) 340 curproc->p_stats->p_ru.ru_inblock++; 341 rabp->b_flags |= B_READ | B_ASYNC; 342 rabp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 343 if (rabp->b_rcred == NOCRED) { 344 if (cred != NOCRED) 345 crhold(cred); 346 rabp->b_rcred = cred; 347 } 348 vfs_busy_pages(rabp, 0); 349 VOP_STRATEGY(vp, rabp); 350 } else { 351 brelse(rabp); 352 } 353 } 354 355 if (readwait) { 356 rv = biowait(bp); 357 } 358 return (rv); 359} 360 361/* 362 * Write, release buffer on completion. (Done by iodone 363 * if async.) 364 */ 365int 366bwrite(struct buf * bp) 367{ 368 int oldflags, s; 369 struct vnode *vp; 370 struct mount *mp; 371 372 373 if (bp->b_flags & B_INVAL) { 374 brelse(bp); 375 return (0); 376 } 377 378 oldflags = bp->b_flags; 379 380#if !defined(MAX_PERF) 381 if ((bp->b_flags & B_BUSY) == 0) 382 panic("bwrite: buffer is not busy???"); 383#endif 384 385 bp->b_flags &= ~(B_READ | B_DONE | B_ERROR | B_DELWRI); 386 bp->b_flags |= B_WRITEINPROG; 387 388 s = splbio(); 389 if ((oldflags & B_DELWRI) == B_DELWRI) { 390 --numdirtybuffers; 391 reassignbuf(bp, bp->b_vp); 392 } 393 394 bp->b_vp->v_numoutput++; 395 vfs_busy_pages(bp, 1); 396 if (curproc != NULL) 397 curproc->p_stats->p_ru.ru_oublock++; 398 splx(s); 399 VOP_STRATEGY(bp->b_vp, bp); 400 401 /* 402 * Collect statistics on synchronous and asynchronous writes. 403 * Writes to block devices are charged to their associated 404 * filesystem (if any). 405 */ 406 if ((vp = bp->b_vp) != NULL) { 407 if (vp->v_type == VBLK) 408 mp = vp->v_specmountpoint; 409 else 410 mp = vp->v_mount; 411 if (mp != NULL) 412 if ((oldflags & B_ASYNC) == 0) 413 mp->mnt_stat.f_syncwrites++; 414 else 415 mp->mnt_stat.f_asyncwrites++; 416 } 417 418 if ((oldflags & B_ASYNC) == 0) { 419 int rtval = biowait(bp); 420 brelse(bp); 421 return (rtval); 422 } 423 return (0); 424} 425 426void 427vfs_bio_need_satisfy(void) { 428 ++numfreebuffers; 429 if (!needsbuffer) 430 return; 431 if (numdirtybuffers < lodirtybuffers) { 432 needsbuffer &= ~(VFS_BIO_NEED_ANY | VFS_BIO_NEED_LOWLIMIT); 433 } else { 434 needsbuffer &= ~VFS_BIO_NEED_ANY; 435 } 436 if (numfreebuffers >= hifreebuffers) { 437 needsbuffer &= ~VFS_BIO_NEED_FREE; 438 } 439 wakeup(&needsbuffer); 440} 441 442/* 443 * Delayed write. (Buffer is marked dirty). 444 */ 445void 446bdwrite(struct buf * bp) 447{ 448 struct vnode *vp; 449 450#if !defined(MAX_PERF) 451 if ((bp->b_flags & B_BUSY) == 0) { 452 panic("bdwrite: buffer is not busy"); 453 } 454#endif 455 456 if (bp->b_flags & B_INVAL) { 457 brelse(bp); 458 return; 459 } 460 bp->b_flags &= ~(B_READ|B_RELBUF); 461 if ((bp->b_flags & B_DELWRI) == 0) { 462 bp->b_flags |= B_DONE | B_DELWRI; 463 reassignbuf(bp, bp->b_vp); 464 ++numdirtybuffers; 465 } 466 467 /* 468 * This bmap keeps the system from needing to do the bmap later, 469 * perhaps when the system is attempting to do a sync. Since it 470 * is likely that the indirect block -- or whatever other datastructure 471 * that the filesystem needs is still in memory now, it is a good 472 * thing to do this. Note also, that if the pageout daemon is 473 * requesting a sync -- there might not be enough memory to do 474 * the bmap then... So, this is important to do. 475 */ 476 if (bp->b_lblkno == bp->b_blkno) { 477 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 478 } 479 480 /* 481 * Set the *dirty* buffer range based upon the VM system dirty pages. 482 */ 483 vfs_setdirty(bp); 484 485 /* 486 * We need to do this here to satisfy the vnode_pager and the 487 * pageout daemon, so that it thinks that the pages have been 488 * "cleaned". Note that since the pages are in a delayed write 489 * buffer -- the VFS layer "will" see that the pages get written 490 * out on the next sync, or perhaps the cluster will be completed. 491 */ 492 vfs_clean_pages(bp); 493 bqrelse(bp); 494 495 /* 496 * XXX The soft dependency code is not prepared to 497 * have I/O done when a bdwrite is requested. For 498 * now we just let the write be delayed if it is 499 * requested by the soft dependency code. 500 */ 501 if ((vp = bp->b_vp) && 502 ((vp->v_type == VBLK && vp->v_specmountpoint && 503 (vp->v_specmountpoint->mnt_flag & MNT_SOFTDEP)) || 504 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SOFTDEP)))) 505 return; 506 507 if (numdirtybuffers >= hidirtybuffers) 508 flushdirtybuffers(0, 0); 509 510 return; 511} 512 513 514/* 515 * Same as first half of bdwrite, mark buffer dirty, but do not release it. 516 * Check how this compares with vfs_setdirty(); XXX [JRE] 517 */ 518void 519bdirty(bp) 520 struct buf *bp; 521{ 522 523 bp->b_flags &= ~(B_READ|B_RELBUF); /* XXX ??? check this */ 524 if ((bp->b_flags & B_DELWRI) == 0) { 525 bp->b_flags |= B_DONE | B_DELWRI; /* why done? XXX JRE */ 526 reassignbuf(bp, bp->b_vp); 527 ++numdirtybuffers; 528 } 529} 530 531/* 532 * Asynchronous write. 533 * Start output on a buffer, but do not wait for it to complete. 534 * The buffer is released when the output completes. 535 */ 536void 537bawrite(struct buf * bp) 538{ 539 bp->b_flags |= B_ASYNC; 540 (void) VOP_BWRITE(bp); 541} 542 543/* 544 * Ordered write. 545 * Start output on a buffer, and flag it so that the device will write 546 * it in the order it was queued. The buffer is released when the output 547 * completes. 548 */ 549int 550bowrite(struct buf * bp) 551{ 552 bp->b_flags |= B_ORDERED|B_ASYNC; 553 return (VOP_BWRITE(bp)); 554} 555 556/* 557 * Release a buffer. 558 */ 559void 560brelse(struct buf * bp) 561{ 562 int s; 563 564 if (bp->b_flags & B_CLUSTER) { 565 relpbuf(bp, NULL); 566 return; 567 } 568 569 s = splbio(); 570 571 /* anyone need this block? */ 572 if (bp->b_flags & B_WANTED) { 573 bp->b_flags &= ~(B_WANTED | B_AGE); 574 wakeup(bp); 575 } 576 577 if (bp->b_flags & B_LOCKED) 578 bp->b_flags &= ~B_ERROR; 579 580 if ((bp->b_flags & (B_READ | B_ERROR)) == B_ERROR) { 581 bp->b_flags &= ~B_ERROR; 582 bdirty(bp); 583 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_FREEBUF)) || 584 (bp->b_bufsize <= 0)) { 585 bp->b_flags |= B_INVAL; 586 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 587 (*bioops.io_deallocate)(bp); 588 if (bp->b_flags & B_DELWRI) 589 --numdirtybuffers; 590 bp->b_flags &= ~(B_DELWRI | B_CACHE | B_FREEBUF); 591 if ((bp->b_flags & B_VMIO) == 0) { 592 if (bp->b_bufsize) 593 allocbuf(bp, 0); 594 if (bp->b_vp) 595 brelvp(bp); 596 } 597 } 598 599 /* 600 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 601 * is called with B_DELWRI set, the underlying pages may wind up 602 * getting freed causing a previous write (bdwrite()) to get 'lost' 603 * because pages associated with a B_DELWRI bp are marked clean. 604 * 605 * We still allow the B_INVAL case to call vfs_vmio_release(), even 606 * if B_DELWRI is set. 607 */ 608 609 if (bp->b_flags & B_DELWRI) 610 bp->b_flags &= ~B_RELBUF; 611 612 /* 613 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 614 * constituted, so the B_INVAL flag is used to *invalidate* the buffer, 615 * but the VM object is kept around. The B_NOCACHE flag is used to 616 * invalidate the pages in the VM object. 617 * 618 * The b_{validoff,validend,dirtyoff,dirtyend} values are relative 619 * to b_offset and currently have byte granularity, whereas the 620 * valid flags in the vm_pages have only DEV_BSIZE resolution. 621 * The byte resolution fields are used to avoid unnecessary re-reads 622 * of the buffer but the code really needs to be genericized so 623 * other filesystem modules can take advantage of these fields. 624 * 625 * XXX this seems to cause performance problems. 626 */ 627 if ((bp->b_flags & B_VMIO) 628 && !(bp->b_vp->v_tag == VT_NFS && 629 bp->b_vp->v_type != VBLK && 630 (bp->b_flags & B_DELWRI) != 0) 631#ifdef notdef 632 && (bp->b_vp->v_tag != VT_NFS 633 || bp->b_vp->v_type == VBLK 634 || (bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) 635 || bp->b_validend == 0 636 || (bp->b_validoff == 0 637 && bp->b_validend == bp->b_bufsize)) 638#endif 639 ) { 640 641 int i, j, resid; 642 vm_page_t m; 643 off_t foff; 644 vm_pindex_t poff; 645 vm_object_t obj; 646 struct vnode *vp; 647 648 vp = bp->b_vp; 649 650 /* 651 * Get the base offset and length of the buffer. Note that 652 * for block sizes that are less then PAGE_SIZE, the b_data 653 * base of the buffer does not represent exactly b_offset and 654 * neither b_offset nor b_size are necessarily page aligned. 655 * Instead, the starting position of b_offset is: 656 * 657 * b_data + (b_offset & PAGE_MASK) 658 * 659 * block sizes less then DEV_BSIZE (usually 512) are not 660 * supported due to the page granularity bits (m->valid, 661 * m->dirty, etc...). 662 * 663 * See man buf(9) for more information 664 */ 665 666 resid = bp->b_bufsize; 667 foff = bp->b_offset; 668 669 for (i = 0; i < bp->b_npages; i++) { 670 m = bp->b_pages[i]; 671 vm_page_flag_clear(m, PG_ZERO); 672 if (m == bogus_page) { 673 674 obj = (vm_object_t) vp->v_object; 675 poff = OFF_TO_IDX(bp->b_offset); 676 677 for (j = i; j < bp->b_npages; j++) { 678 m = bp->b_pages[j]; 679 if (m == bogus_page) { 680 m = vm_page_lookup(obj, poff + j); 681#if !defined(MAX_PERF) 682 if (!m) { 683 panic("brelse: page missing\n"); 684 } 685#endif 686 bp->b_pages[j] = m; 687 } 688 } 689 690 if ((bp->b_flags & B_INVAL) == 0) { 691 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 692 } 693 } 694 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 695 int poffset = foff & PAGE_MASK; 696 int presid = resid > (PAGE_SIZE - poffset) ? 697 (PAGE_SIZE - poffset) : resid; 698 699 KASSERT(presid >= 0, ("brelse: extra page")); 700 vm_page_set_invalid(m, poffset, presid); 701 } 702 resid -= PAGE_SIZE - (foff & PAGE_MASK); 703 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 704 } 705 706 if (bp->b_flags & (B_INVAL | B_RELBUF)) 707 vfs_vmio_release(bp); 708 709 } else if (bp->b_flags & B_VMIO) { 710 711 if (bp->b_flags & (B_INVAL | B_RELBUF)) 712 vfs_vmio_release(bp); 713 714 } 715 716#if !defined(MAX_PERF) 717 if (bp->b_qindex != QUEUE_NONE) 718 panic("brelse: free buffer onto another queue???"); 719#endif 720 721 /* enqueue */ 722 /* buffers with no memory */ 723 if (bp->b_bufsize == 0) { 724 bp->b_flags |= B_INVAL; 725 bp->b_qindex = QUEUE_EMPTY; 726 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 727 LIST_REMOVE(bp, b_hash); 728 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 729 bp->b_dev = NODEV; 730 kvafreespace += bp->b_kvasize; 731 732 /* buffers with junk contents */ 733 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) { 734 bp->b_flags |= B_INVAL; 735 bp->b_qindex = QUEUE_AGE; 736 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_AGE], bp, b_freelist); 737 LIST_REMOVE(bp, b_hash); 738 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 739 bp->b_dev = NODEV; 740 741 /* buffers that are locked */ 742 } else if (bp->b_flags & B_LOCKED) { 743 bp->b_qindex = QUEUE_LOCKED; 744 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 745 746 /* buffers with stale but valid contents */ 747 } else if (bp->b_flags & B_AGE) { 748 bp->b_qindex = QUEUE_AGE; 749 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_AGE], bp, b_freelist); 750 751 /* buffers with valid and quite potentially reuseable contents */ 752 } else { 753 bp->b_qindex = QUEUE_LRU; 754 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 755 } 756 757 if ((bp->b_flags & B_INVAL) || 758 (bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) { 759 if (bp->b_flags & B_DELWRI) { 760 --numdirtybuffers; 761 bp->b_flags &= ~B_DELWRI; 762 } 763 vfs_bio_need_satisfy(); 764 } 765 766 /* unlock */ 767 bp->b_flags &= ~(B_ORDERED | B_WANTED | B_BUSY | 768 B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 769 splx(s); 770} 771 772/* 773 * Release a buffer. 774 */ 775void 776bqrelse(struct buf * bp) 777{ 778 int s; 779 780 s = splbio(); 781 782 /* anyone need this block? */ 783 if (bp->b_flags & B_WANTED) { 784 bp->b_flags &= ~(B_WANTED | B_AGE); 785 wakeup(bp); 786 } 787 788#if !defined(MAX_PERF) 789 if (bp->b_qindex != QUEUE_NONE) 790 panic("bqrelse: free buffer onto another queue???"); 791#endif 792 793 if (bp->b_flags & B_LOCKED) { 794 bp->b_flags &= ~B_ERROR; 795 bp->b_qindex = QUEUE_LOCKED; 796 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 797 /* buffers with stale but valid contents */ 798 } else { 799 bp->b_qindex = QUEUE_LRU; 800 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 801 } 802 803 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) { 804 vfs_bio_need_satisfy(); 805 } 806 807 /* unlock */ 808 bp->b_flags &= ~(B_ORDERED | B_WANTED | B_BUSY | 809 B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 810 splx(s); 811} 812 813static void 814vfs_vmio_release(bp) 815 struct buf *bp; 816{ 817 int i, s; 818 vm_page_t m; 819 820 s = splvm(); 821 for (i = 0; i < bp->b_npages; i++) { 822 m = bp->b_pages[i]; 823 bp->b_pages[i] = NULL; 824 /* 825 * In order to keep page LRU ordering consistent, put 826 * everything on the inactive queue. 827 */ 828 vm_page_unwire(m, 0); 829 /* 830 * We don't mess with busy pages, it is 831 * the responsibility of the process that 832 * busied the pages to deal with them. 833 */ 834 if ((m->flags & PG_BUSY) || (m->busy != 0)) 835 continue; 836 837 if (m->wire_count == 0) { 838 vm_page_flag_clear(m, PG_ZERO); 839 /* 840 * Might as well free the page if we can and it has 841 * no valid data. 842 */ 843 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { 844 vm_page_busy(m); 845 vm_page_protect(m, VM_PROT_NONE); 846 vm_page_free(m); 847 } 848 } 849 } 850 splx(s); 851 bufspace -= bp->b_bufsize; 852 vmiospace -= bp->b_bufsize; 853 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 854 bp->b_npages = 0; 855 bp->b_bufsize = 0; 856 bp->b_flags &= ~B_VMIO; 857 if (bp->b_vp) 858 brelvp(bp); 859} 860 861/* 862 * Check to see if a block is currently memory resident. 863 */ 864struct buf * 865gbincore(struct vnode * vp, daddr_t blkno) 866{ 867 struct buf *bp; 868 struct bufhashhdr *bh; 869 870 bh = BUFHASH(vp, blkno); 871 bp = bh->lh_first; 872 873 /* Search hash chain */ 874 while (bp != NULL) { 875 /* hit */ 876 if (bp->b_vp == vp && bp->b_lblkno == blkno && 877 (bp->b_flags & B_INVAL) == 0) { 878 break; 879 } 880 bp = bp->b_hash.le_next; 881 } 882 return (bp); 883} 884 885/* 886 * this routine implements clustered async writes for 887 * clearing out B_DELWRI buffers... This is much better 888 * than the old way of writing only one buffer at a time. 889 */ 890int 891vfs_bio_awrite(struct buf * bp) 892{ 893 int i; 894 daddr_t lblkno = bp->b_lblkno; 895 struct vnode *vp = bp->b_vp; 896 int s; 897 int ncl; 898 struct buf *bpa; 899 int nwritten; 900 int size; 901 int maxcl; 902 903 s = splbio(); 904 /* 905 * right now we support clustered writing only to regular files 906 */ 907 if ((vp->v_type == VREG) && 908 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 909 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 910 911 size = vp->v_mount->mnt_stat.f_iosize; 912 maxcl = MAXPHYS / size; 913 914 for (i = 1; i < maxcl; i++) { 915 if ((bpa = gbincore(vp, lblkno + i)) && 916 ((bpa->b_flags & (B_BUSY | B_DELWRI | B_CLUSTEROK | B_INVAL)) == 917 (B_DELWRI | B_CLUSTEROK)) && 918 (bpa->b_bufsize == size)) { 919 if ((bpa->b_blkno == bpa->b_lblkno) || 920 (bpa->b_blkno != bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 921 break; 922 } else { 923 break; 924 } 925 } 926 ncl = i; 927 /* 928 * this is a possible cluster write 929 */ 930 if (ncl != 1) { 931 nwritten = cluster_wbuild(vp, size, lblkno, ncl); 932 splx(s); 933 return nwritten; 934 } 935 } 936 937 bremfree(bp); 938 bp->b_flags |= B_BUSY | B_ASYNC; 939 940 splx(s); 941 /* 942 * default (old) behavior, writing out only one block 943 */ 944 nwritten = bp->b_bufsize; 945 (void) VOP_BWRITE(bp); 946 return nwritten; 947} 948 949 950/* 951 * Find a buffer header which is available for use. 952 */ 953static struct buf * 954getnewbuf(struct vnode *vp, daddr_t blkno, 955 int slpflag, int slptimeo, int size, int maxsize) 956{ 957 struct buf *bp, *bp1; 958 int nbyteswritten = 0; 959 vm_offset_t addr; 960 static int writerecursion = 0; 961 962start: 963 if (bufspace >= maxbufspace) 964 goto trytofreespace; 965 966 /* can we constitute a new buffer? */ 967 if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]))) { 968#if !defined(MAX_PERF) 969 if (bp->b_qindex != QUEUE_EMPTY) 970 panic("getnewbuf: inconsistent EMPTY queue, qindex=%d", 971 bp->b_qindex); 972#endif 973 bp->b_flags |= B_BUSY; 974 bremfree(bp); 975 goto fillbuf; 976 } 977trytofreespace: 978 /* 979 * We keep the file I/O from hogging metadata I/O 980 * This is desirable because file data is cached in the 981 * VM/Buffer cache even if a buffer is freed. 982 */ 983 if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]))) { 984#if !defined(MAX_PERF) 985 if (bp->b_qindex != QUEUE_AGE) 986 panic("getnewbuf: inconsistent AGE queue, qindex=%d", 987 bp->b_qindex); 988#endif 989 } else if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]))) { 990#if !defined(MAX_PERF) 991 if (bp->b_qindex != QUEUE_LRU) 992 panic("getnewbuf: inconsistent LRU queue, qindex=%d", 993 bp->b_qindex); 994#endif 995 } 996 if (!bp) { 997 /* wait for a free buffer of any kind */ 998 needsbuffer |= VFS_BIO_NEED_ANY; 999 do 1000 tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, "newbuf", 1001 slptimeo); 1002 while (needsbuffer & VFS_BIO_NEED_ANY); 1003 return (0); 1004 } 1005 KASSERT(!(bp->b_flags & B_BUSY), 1006 ("getnewbuf: busy buffer on free list\n")); 1007 /* 1008 * We are fairly aggressive about freeing VMIO buffers, but since 1009 * the buffering is intact without buffer headers, there is not 1010 * much loss. We gain by maintaining non-VMIOed metadata in buffers. 1011 */ 1012 if ((bp->b_qindex == QUEUE_LRU) && (bp->b_usecount > 0)) { 1013 if ((bp->b_flags & B_VMIO) == 0 || 1014 (vmiospace < maxvmiobufspace)) { 1015 --bp->b_usecount; 1016 TAILQ_REMOVE(&bufqueues[QUEUE_LRU], bp, b_freelist); 1017 if (TAILQ_FIRST(&bufqueues[QUEUE_LRU]) != NULL) { 1018 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1019 goto start; 1020 } 1021 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1022 } 1023 } 1024 1025 1026 /* if we are a delayed write, convert to an async write */ 1027 if ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) { 1028 1029 /* 1030 * If our delayed write is likely to be used soon, then 1031 * recycle back onto the LRU queue. 1032 */ 1033 if (vp && (bp->b_vp == vp) && (bp->b_qindex == QUEUE_LRU) && 1034 (bp->b_lblkno >= blkno) && (maxsize > 0)) { 1035 1036 if (bp->b_usecount > 0) { 1037 if (bp->b_lblkno < blkno + (MAXPHYS / maxsize)) { 1038 1039 TAILQ_REMOVE(&bufqueues[QUEUE_LRU], bp, b_freelist); 1040 1041 if (TAILQ_FIRST(&bufqueues[QUEUE_LRU]) != NULL) { 1042 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1043 bp->b_usecount--; 1044 goto start; 1045 } 1046 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1047 } 1048 } 1049 } 1050 1051 /* 1052 * Certain layered filesystems can recursively re-enter the vfs_bio 1053 * code, due to delayed writes. This helps keep the system from 1054 * deadlocking. 1055 */ 1056 if (writerecursion > 0) { 1057 if (writerecursion > 5) { 1058 bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]); 1059 while (bp) { 1060 if ((bp->b_flags & B_DELWRI) == 0) 1061 break; 1062 bp = TAILQ_NEXT(bp, b_freelist); 1063 } 1064 if (bp == NULL) { 1065 bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]); 1066 while (bp) { 1067 if ((bp->b_flags & B_DELWRI) == 0) 1068 break; 1069 bp = TAILQ_NEXT(bp, b_freelist); 1070 } 1071 } 1072 if (bp == NULL) 1073 panic("getnewbuf: cannot get buffer, infinite recursion failure"); 1074 } else { 1075 bremfree(bp); 1076 bp->b_flags |= B_BUSY | B_AGE | B_ASYNC; 1077 nbyteswritten += bp->b_bufsize; 1078 ++writerecursion; 1079 VOP_BWRITE(bp); 1080 --writerecursion; 1081 if (!slpflag && !slptimeo) { 1082 return (0); 1083 } 1084 goto start; 1085 } 1086 } else { 1087 ++writerecursion; 1088 nbyteswritten += vfs_bio_awrite(bp); 1089 --writerecursion; 1090 if (!slpflag && !slptimeo) { 1091 return (0); 1092 } 1093 goto start; 1094 } 1095 } 1096 1097 if (bp->b_flags & B_WANTED) { 1098 bp->b_flags &= ~B_WANTED; 1099 wakeup(bp); 1100 } 1101 bremfree(bp); 1102 bp->b_flags |= B_BUSY; 1103 1104 if (bp->b_flags & B_VMIO) { 1105 bp->b_flags &= ~B_ASYNC; 1106 vfs_vmio_release(bp); 1107 } 1108 1109 if (bp->b_vp) 1110 brelvp(bp); 1111 1112fillbuf: 1113 1114 /* we are not free, nor do we contain interesting data */ 1115 if (bp->b_rcred != NOCRED) { 1116 crfree(bp->b_rcred); 1117 bp->b_rcred = NOCRED; 1118 } 1119 if (bp->b_wcred != NOCRED) { 1120 crfree(bp->b_wcred); 1121 bp->b_wcred = NOCRED; 1122 } 1123 if (LIST_FIRST(&bp->b_dep) != NULL && 1124 bioops.io_deallocate) 1125 (*bioops.io_deallocate)(bp); 1126 1127 LIST_REMOVE(bp, b_hash); 1128 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1129 if (bp->b_bufsize) { 1130 allocbuf(bp, 0); 1131 } 1132 bp->b_flags = B_BUSY; 1133 bp->b_dev = NODEV; 1134 bp->b_vp = NULL; 1135 bp->b_blkno = bp->b_lblkno = 0; 1136 bp->b_offset = NOOFFSET; 1137 bp->b_iodone = 0; 1138 bp->b_error = 0; 1139 bp->b_resid = 0; 1140 bp->b_bcount = 0; 1141 bp->b_npages = 0; 1142 bp->b_dirtyoff = bp->b_dirtyend = 0; 1143 bp->b_validoff = bp->b_validend = 0; 1144 bp->b_usecount = 5; 1145 /* Here, not kern_physio.c, is where this should be done*/ 1146 LIST_INIT(&bp->b_dep); 1147 1148 maxsize = (maxsize + PAGE_MASK) & ~PAGE_MASK; 1149 1150 /* 1151 * we assume that buffer_map is not at address 0 1152 */ 1153 addr = 0; 1154 if (maxsize != bp->b_kvasize) { 1155 bfreekva(bp); 1156 1157findkvaspace: 1158 /* 1159 * See if we have buffer kva space 1160 */ 1161 if (vm_map_findspace(buffer_map, 1162 vm_map_min(buffer_map), maxsize, &addr)) { 1163 if (kvafreespace > 0) { 1164 int totfree = 0, freed; 1165 do { 1166 freed = 0; 1167 for (bp1 = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1168 bp1 != NULL; bp1 = TAILQ_NEXT(bp1, b_freelist)) { 1169 if (bp1->b_kvasize != 0) { 1170 totfree += bp1->b_kvasize; 1171 freed = bp1->b_kvasize; 1172 bremfree(bp1); 1173 bfreekva(bp1); 1174 brelse(bp1); 1175 break; 1176 } 1177 } 1178 } while (freed); 1179 /* 1180 * if we found free space, then retry with the same buffer. 1181 */ 1182 if (totfree) 1183 goto findkvaspace; 1184 } 1185 bp->b_flags |= B_INVAL; 1186 brelse(bp); 1187 goto trytofreespace; 1188 } 1189 } 1190 1191 /* 1192 * See if we are below are allocated minimum 1193 */ 1194 if (bufspace >= (maxbufspace + nbyteswritten)) { 1195 bp->b_flags |= B_INVAL; 1196 brelse(bp); 1197 goto trytofreespace; 1198 } 1199 1200 /* 1201 * create a map entry for the buffer -- in essence 1202 * reserving the kva space. 1203 */ 1204 if (addr) { 1205 vm_map_insert(buffer_map, NULL, 0, 1206 addr, addr + maxsize, 1207 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1208 1209 bp->b_kvabase = (caddr_t) addr; 1210 bp->b_kvasize = maxsize; 1211 } 1212 bp->b_data = bp->b_kvabase; 1213 1214 return (bp); 1215} 1216 1217static void 1218waitfreebuffers(int slpflag, int slptimeo) { 1219 while (numfreebuffers < hifreebuffers) { 1220 flushdirtybuffers(slpflag, slptimeo); 1221 if (numfreebuffers < hifreebuffers) 1222 break; 1223 needsbuffer |= VFS_BIO_NEED_FREE; 1224 if (tsleep(&needsbuffer, (PRIBIO + 4)|slpflag, "biofre", slptimeo)) 1225 break; 1226 } 1227} 1228 1229static void 1230flushdirtybuffers(int slpflag, int slptimeo) { 1231 int s; 1232 static pid_t flushing = 0; 1233 1234 s = splbio(); 1235 1236 if (flushing) { 1237 if (flushing == curproc->p_pid) { 1238 splx(s); 1239 return; 1240 } 1241 while (flushing) { 1242 if (tsleep(&flushing, (PRIBIO + 4)|slpflag, "biofls", slptimeo)) { 1243 splx(s); 1244 return; 1245 } 1246 } 1247 } 1248 flushing = curproc->p_pid; 1249 1250 while (numdirtybuffers > lodirtybuffers) { 1251 struct buf *bp; 1252 needsbuffer |= VFS_BIO_NEED_LOWLIMIT; 1253 bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]); 1254 if (bp == NULL) 1255 bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]); 1256 1257 while (bp && ((bp->b_flags & B_DELWRI) == 0)) { 1258 bp = TAILQ_NEXT(bp, b_freelist); 1259 } 1260 1261 if (bp) { 1262 vfs_bio_awrite(bp); 1263 continue; 1264 } 1265 break; 1266 } 1267 1268 flushing = 0; 1269 wakeup(&flushing); 1270 splx(s); 1271} 1272 1273/* 1274 * Check to see if a block is currently memory resident. 1275 */ 1276struct buf * 1277incore(struct vnode * vp, daddr_t blkno) 1278{ 1279 struct buf *bp; 1280 1281 int s = splbio(); 1282 bp = gbincore(vp, blkno); 1283 splx(s); 1284 return (bp); 1285} 1286 1287/* 1288 * Returns true if no I/O is needed to access the 1289 * associated VM object. This is like incore except 1290 * it also hunts around in the VM system for the data. 1291 */ 1292 1293int 1294inmem(struct vnode * vp, daddr_t blkno) 1295{ 1296 vm_object_t obj; 1297 vm_offset_t toff, tinc, size; 1298 vm_page_t m; 1299 vm_ooffset_t off; 1300 1301 if (incore(vp, blkno)) 1302 return 1; 1303 if (vp->v_mount == NULL) 1304 return 0; 1305 if ((vp->v_object == NULL) || (vp->v_flag & VOBJBUF) == 0) 1306 return 0; 1307 1308 obj = vp->v_object; 1309 size = PAGE_SIZE; 1310 if (size > vp->v_mount->mnt_stat.f_iosize) 1311 size = vp->v_mount->mnt_stat.f_iosize; 1312 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 1313 1314 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1315 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 1316 if (!m) 1317 return 0; 1318 tinc = size; 1319 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 1320 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 1321 if (vm_page_is_valid(m, 1322 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 1323 return 0; 1324 } 1325 return 1; 1326} 1327 1328/* 1329 * now we set the dirty range for the buffer -- 1330 * for NFS -- if the file is mapped and pages have 1331 * been written to, let it know. We want the 1332 * entire range of the buffer to be marked dirty if 1333 * any of the pages have been written to for consistancy 1334 * with the b_validoff, b_validend set in the nfs write 1335 * code, and used by the nfs read code. 1336 */ 1337static void 1338vfs_setdirty(struct buf *bp) { 1339 int i; 1340 vm_object_t object; 1341 vm_offset_t boffset; 1342#if 0 1343 vm_offset_t offset; 1344#endif 1345 1346 /* 1347 * We qualify the scan for modified pages on whether the 1348 * object has been flushed yet. The OBJ_WRITEABLE flag 1349 * is not cleared simply by protecting pages off. 1350 */ 1351 if ((bp->b_flags & B_VMIO) && 1352 ((object = bp->b_pages[0]->object)->flags & (OBJ_WRITEABLE|OBJ_CLEANING))) { 1353 /* 1354 * test the pages to see if they have been modified directly 1355 * by users through the VM system. 1356 */ 1357 for (i = 0; i < bp->b_npages; i++) { 1358 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 1359 vm_page_test_dirty(bp->b_pages[i]); 1360 } 1361 1362 /* 1363 * scan forwards for the first page modified 1364 */ 1365 for (i = 0; i < bp->b_npages; i++) { 1366 if (bp->b_pages[i]->dirty) { 1367 break; 1368 } 1369 } 1370 1371 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 1372 if (boffset < bp->b_dirtyoff) { 1373 bp->b_dirtyoff = max(boffset, 0); 1374 } 1375 1376 /* 1377 * scan backwards for the last page modified 1378 */ 1379 for (i = bp->b_npages - 1; i >= 0; --i) { 1380 if (bp->b_pages[i]->dirty) { 1381 break; 1382 } 1383 } 1384 boffset = (i + 1); 1385#if 0 1386 offset = boffset + bp->b_pages[0]->pindex; 1387 if (offset >= object->size) 1388 boffset = object->size - bp->b_pages[0]->pindex; 1389#endif 1390 boffset = (boffset << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 1391 if (bp->b_dirtyend < boffset) 1392 bp->b_dirtyend = min(boffset, bp->b_bufsize); 1393 } 1394} 1395 1396/* 1397 * Get a block given a specified block and offset into a file/device. 1398 */ 1399struct buf * 1400getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 1401{ 1402 struct buf *bp; 1403 int i, s; 1404 struct bufhashhdr *bh; 1405 1406#if !defined(MAX_PERF) 1407 if (size > MAXBSIZE) 1408 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 1409#endif 1410 1411 s = splbio(); 1412loop: 1413 if (numfreebuffers < lofreebuffers) { 1414 waitfreebuffers(slpflag, slptimeo); 1415 } 1416 1417 if ((bp = gbincore(vp, blkno))) { 1418 if (bp->b_flags & B_BUSY) { 1419 bp->b_flags |= B_WANTED; 1420 if (bp->b_usecount < BUF_MAXUSE) 1421 ++bp->b_usecount; 1422 1423 if (!tsleep(bp, 1424 (PRIBIO + 4) | slpflag, "getblk", slptimeo)) { 1425 goto loop; 1426 } 1427 1428 splx(s); 1429 return (struct buf *) NULL; 1430 } 1431 bp->b_flags |= B_BUSY | B_CACHE; 1432 bremfree(bp); 1433 1434 /* 1435 * check for size inconsistancies for non-VMIO case. 1436 */ 1437 1438 if (bp->b_bcount != size) { 1439 if ((bp->b_flags & B_VMIO) == 0 || 1440 (size > bp->b_kvasize) 1441 ) { 1442 if (bp->b_flags & B_DELWRI) { 1443 bp->b_flags |= B_NOCACHE; 1444 VOP_BWRITE(bp); 1445 } else { 1446 if ((bp->b_flags & B_VMIO) && 1447 (LIST_FIRST(&bp->b_dep) == NULL)) { 1448 bp->b_flags |= B_RELBUF; 1449 brelse(bp); 1450 } else { 1451 bp->b_flags |= B_NOCACHE; 1452 VOP_BWRITE(bp); 1453 } 1454 } 1455 goto loop; 1456 } 1457 } 1458 1459 /* 1460 * If the size is inconsistant in the VMIO case, we can resize 1461 * the buffer. This might lead to B_CACHE getting cleared. 1462 */ 1463 1464 if (bp->b_bcount != size) 1465 allocbuf(bp, size); 1466 1467 KASSERT(bp->b_offset != NOOFFSET, 1468 ("getblk: no buffer offset")); 1469 1470 /* 1471 * Check that the constituted buffer really deserves for the 1472 * B_CACHE bit to be set. B_VMIO type buffers might not 1473 * contain fully valid pages. Normal (old-style) buffers 1474 * should be fully valid. This might also lead to B_CACHE 1475 * getting clear. 1476 * 1477 * If B_CACHE is already clear, don't bother checking to see 1478 * if we have to clear it again. 1479 * 1480 * XXX this code should not be necessary unless the B_CACHE 1481 * handling is broken elsewhere in the kernel. We need to 1482 * check the cases and then turn the clearing part of this 1483 * code into a panic. 1484 */ 1485 if ( 1486 (bp->b_flags & (B_VMIO|B_CACHE)) == (B_VMIO|B_CACHE) && 1487 (bp->b_vp->v_tag != VT_NFS || bp->b_validend <= 0) 1488 ) { 1489 int checksize = bp->b_bufsize; 1490 int poffset = bp->b_offset & PAGE_MASK; 1491 int resid; 1492 for (i = 0; i < bp->b_npages; i++) { 1493 resid = (checksize > (PAGE_SIZE - poffset)) ? 1494 (PAGE_SIZE - poffset) : checksize; 1495 if (!vm_page_is_valid(bp->b_pages[i], poffset, resid)) { 1496 bp->b_flags &= ~(B_CACHE | B_DONE); 1497 break; 1498 } 1499 checksize -= resid; 1500 poffset = 0; 1501 } 1502 } 1503 1504 /* 1505 * If B_DELWRI is set and B_CACHE got cleared ( or was 1506 * already clear ), we have to commit the write and 1507 * retry. The NFS code absolutely depends on this, 1508 * and so might the FFS code. In anycase, it formalizes 1509 * the B_CACHE rules. See sys/buf.h. 1510 */ 1511 1512 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 1513 VOP_BWRITE(bp); 1514 goto loop; 1515 } 1516 1517 if (bp->b_usecount < BUF_MAXUSE) 1518 ++bp->b_usecount; 1519 splx(s); 1520 return (bp); 1521 } else { 1522 int bsize, maxsize, vmio; 1523 off_t offset; 1524 1525 if (vp->v_type == VBLK) 1526 bsize = DEV_BSIZE; 1527 else if (vp->v_mountedhere) 1528 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 1529 else if (vp->v_mount) 1530 bsize = vp->v_mount->mnt_stat.f_iosize; 1531 else 1532 bsize = size; 1533 1534 offset = (off_t)blkno * bsize; 1535 vmio = (vp->v_object != 0) && (vp->v_flag & VOBJBUF); 1536 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 1537 maxsize = imax(maxsize, bsize); 1538 1539 if ((bp = getnewbuf(vp, blkno, 1540 slpflag, slptimeo, size, maxsize)) == 0) { 1541 if (slpflag || slptimeo) { 1542 splx(s); 1543 return NULL; 1544 } 1545 goto loop; 1546 } 1547 1548 /* 1549 * This code is used to make sure that a buffer is not 1550 * created while the getnewbuf routine is blocked. 1551 * Normally the vnode is locked so this isn't a problem. 1552 * VBLK type I/O requests, however, don't lock the vnode. 1553 */ 1554 if (gbincore(vp, blkno)) { 1555 bp->b_flags |= B_INVAL; 1556 brelse(bp); 1557 goto loop; 1558 } 1559 1560 /* 1561 * Insert the buffer into the hash, so that it can 1562 * be found by incore. 1563 */ 1564 bp->b_blkno = bp->b_lblkno = blkno; 1565 bp->b_offset = offset; 1566 1567 bgetvp(vp, bp); 1568 LIST_REMOVE(bp, b_hash); 1569 bh = BUFHASH(vp, blkno); 1570 LIST_INSERT_HEAD(bh, bp, b_hash); 1571 1572 if (vmio) { 1573 bp->b_flags |= (B_VMIO | B_CACHE); 1574#if defined(VFS_BIO_DEBUG) 1575 if (vp->v_type != VREG && vp->v_type != VBLK) 1576 printf("getblk: vmioing file type %d???\n", vp->v_type); 1577#endif 1578 } else { 1579 bp->b_flags &= ~B_VMIO; 1580 } 1581 1582 allocbuf(bp, size); 1583 1584 splx(s); 1585 return (bp); 1586 } 1587} 1588 1589/* 1590 * Get an empty, disassociated buffer of given size. 1591 */ 1592struct buf * 1593geteblk(int size) 1594{ 1595 struct buf *bp; 1596 int s; 1597 1598 s = splbio(); 1599 while ((bp = getnewbuf(0, (daddr_t) 0, 0, 0, size, MAXBSIZE)) == 0); 1600 splx(s); 1601 allocbuf(bp, size); 1602 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 1603 return (bp); 1604} 1605 1606 1607/* 1608 * This code constitutes the buffer memory from either anonymous system 1609 * memory (in the case of non-VMIO operations) or from an associated 1610 * VM object (in the case of VMIO operations). This code is able to 1611 * resize a buffer up or down. 1612 * 1613 * Note that this code is tricky, and has many complications to resolve 1614 * deadlock or inconsistant data situations. Tread lightly!!! 1615 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 1616 * the caller. Calling this code willy nilly can result in the loss of data. 1617 */ 1618 1619int 1620allocbuf(struct buf *bp, int size) 1621{ 1622 int newbsize, mbsize; 1623 int i; 1624 1625#if !defined(MAX_PERF) 1626 if (!(bp->b_flags & B_BUSY)) 1627 panic("allocbuf: buffer not busy"); 1628 1629 if (bp->b_kvasize < size) 1630 panic("allocbuf: buffer too small"); 1631#endif 1632 1633 if ((bp->b_flags & B_VMIO) == 0) { 1634 caddr_t origbuf; 1635 int origbufsize; 1636 /* 1637 * Just get anonymous memory from the kernel 1638 */ 1639 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 1640#if !defined(NO_B_MALLOC) 1641 if (bp->b_flags & B_MALLOC) 1642 newbsize = mbsize; 1643 else 1644#endif 1645 newbsize = round_page(size); 1646 1647 if (newbsize < bp->b_bufsize) { 1648#if !defined(NO_B_MALLOC) 1649 /* 1650 * malloced buffers are not shrunk 1651 */ 1652 if (bp->b_flags & B_MALLOC) { 1653 if (newbsize) { 1654 bp->b_bcount = size; 1655 } else { 1656 free(bp->b_data, M_BIOBUF); 1657 bufspace -= bp->b_bufsize; 1658 bufmallocspace -= bp->b_bufsize; 1659 bp->b_data = bp->b_kvabase; 1660 bp->b_bufsize = 0; 1661 bp->b_bcount = 0; 1662 bp->b_flags &= ~B_MALLOC; 1663 } 1664 return 1; 1665 } 1666#endif 1667 vm_hold_free_pages( 1668 bp, 1669 (vm_offset_t) bp->b_data + newbsize, 1670 (vm_offset_t) bp->b_data + bp->b_bufsize); 1671 } else if (newbsize > bp->b_bufsize) { 1672#if !defined(NO_B_MALLOC) 1673 /* 1674 * We only use malloced memory on the first allocation. 1675 * and revert to page-allocated memory when the buffer grows. 1676 */ 1677 if ( (bufmallocspace < maxbufmallocspace) && 1678 (bp->b_bufsize == 0) && 1679 (mbsize <= PAGE_SIZE/2)) { 1680 1681 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 1682 bp->b_bufsize = mbsize; 1683 bp->b_bcount = size; 1684 bp->b_flags |= B_MALLOC; 1685 bufspace += mbsize; 1686 bufmallocspace += mbsize; 1687 return 1; 1688 } 1689#endif 1690 origbuf = NULL; 1691 origbufsize = 0; 1692#if !defined(NO_B_MALLOC) 1693 /* 1694 * If the buffer is growing on its other-than-first allocation, 1695 * then we revert to the page-allocation scheme. 1696 */ 1697 if (bp->b_flags & B_MALLOC) { 1698 origbuf = bp->b_data; 1699 origbufsize = bp->b_bufsize; 1700 bp->b_data = bp->b_kvabase; 1701 bufspace -= bp->b_bufsize; 1702 bufmallocspace -= bp->b_bufsize; 1703 bp->b_bufsize = 0; 1704 bp->b_flags &= ~B_MALLOC; 1705 newbsize = round_page(newbsize); 1706 } 1707#endif 1708 vm_hold_load_pages( 1709 bp, 1710 (vm_offset_t) bp->b_data + bp->b_bufsize, 1711 (vm_offset_t) bp->b_data + newbsize); 1712#if !defined(NO_B_MALLOC) 1713 if (origbuf) { 1714 bcopy(origbuf, bp->b_data, origbufsize); 1715 free(origbuf, M_BIOBUF); 1716 } 1717#endif 1718 } 1719 } else { 1720 vm_page_t m; 1721 int desiredpages; 1722 1723 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 1724 desiredpages = (size == 0) ? 0 : 1725 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 1726 1727#if !defined(NO_B_MALLOC) 1728 if (bp->b_flags & B_MALLOC) 1729 panic("allocbuf: VMIO buffer can't be malloced"); 1730#endif 1731 1732 if (newbsize < bp->b_bufsize) { 1733 if (desiredpages < bp->b_npages) { 1734 for (i = desiredpages; i < bp->b_npages; i++) { 1735 /* 1736 * the page is not freed here -- it 1737 * is the responsibility of vnode_pager_setsize 1738 */ 1739 m = bp->b_pages[i]; 1740 KASSERT(m != bogus_page, 1741 ("allocbuf: bogus page found")); 1742 while (vm_page_sleep_busy(m, TRUE, "biodep")) 1743 ; 1744 1745 bp->b_pages[i] = NULL; 1746 vm_page_unwire(m, 0); 1747 } 1748 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 1749 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 1750 bp->b_npages = desiredpages; 1751 } 1752 } else if (newbsize > bp->b_bufsize) { 1753 vm_object_t obj; 1754 vm_offset_t tinc, toff; 1755 vm_ooffset_t off; 1756 vm_pindex_t objoff; 1757 int pageindex, curbpnpages; 1758 struct vnode *vp; 1759 int bsize; 1760 int orig_validoff = bp->b_validoff; 1761 int orig_validend = bp->b_validend; 1762 1763 vp = bp->b_vp; 1764 1765 if (vp->v_type == VBLK) 1766 bsize = DEV_BSIZE; 1767 else 1768 bsize = vp->v_mount->mnt_stat.f_iosize; 1769 1770 if (bp->b_npages < desiredpages) { 1771 obj = vp->v_object; 1772 tinc = PAGE_SIZE; 1773 1774 off = bp->b_offset; 1775 KASSERT(bp->b_offset != NOOFFSET, 1776 ("allocbuf: no buffer offset")); 1777 curbpnpages = bp->b_npages; 1778 doretry: 1779 bp->b_validoff = orig_validoff; 1780 bp->b_validend = orig_validend; 1781 bp->b_flags |= B_CACHE; 1782 for (toff = 0; toff < newbsize; toff += tinc) { 1783 objoff = OFF_TO_IDX(off + toff); 1784 pageindex = objoff - OFF_TO_IDX(off); 1785 tinc = PAGE_SIZE - ((off + toff) & PAGE_MASK); 1786 if (pageindex < curbpnpages) { 1787 1788 m = bp->b_pages[pageindex]; 1789#ifdef VFS_BIO_DIAG 1790 if (m->pindex != objoff) 1791 panic("allocbuf: page changed offset?!!!?"); 1792#endif 1793 if (tinc > (newbsize - toff)) 1794 tinc = newbsize - toff; 1795 if (bp->b_flags & B_CACHE) 1796 vfs_buf_set_valid(bp, off, toff, tinc, m); 1797 continue; 1798 } 1799 m = vm_page_lookup(obj, objoff); 1800 if (!m) { 1801 m = vm_page_alloc(obj, objoff, VM_ALLOC_NORMAL); 1802 if (!m) { 1803 VM_WAIT; 1804 vm_pageout_deficit += (desiredpages - curbpnpages); 1805 goto doretry; 1806 } 1807 1808 vm_page_wire(m); 1809 vm_page_wakeup(m); 1810 bp->b_flags &= ~B_CACHE; 1811 1812 } else if (vm_page_sleep_busy(m, FALSE, "pgtblk")) { 1813 /* 1814 * If we had to sleep, retry. 1815 * 1816 * Also note that we only test 1817 * PG_BUSY here, not m->busy. 1818 * 1819 * We cannot sleep on m->busy 1820 * here because a vm_fault -> 1821 * getpages -> cluster-read -> 1822 * ...-> allocbuf sequence 1823 * will convert PG_BUSY to 1824 * m->busy so we have to let 1825 * m->busy through if we do 1826 * not want to deadlock. 1827 */ 1828 goto doretry; 1829 } else { 1830 if ((curproc != pageproc) && 1831 ((m->queue - m->pc) == PQ_CACHE) && 1832 ((cnt.v_free_count + cnt.v_cache_count) < 1833 (cnt.v_free_min + cnt.v_cache_min))) { 1834 pagedaemon_wakeup(); 1835 } 1836 if (tinc > (newbsize - toff)) 1837 tinc = newbsize - toff; 1838 if (bp->b_flags & B_CACHE) 1839 vfs_buf_set_valid(bp, off, toff, tinc, m); 1840 vm_page_flag_clear(m, PG_ZERO); 1841 vm_page_wire(m); 1842 } 1843 bp->b_pages[pageindex] = m; 1844 curbpnpages = pageindex + 1; 1845 } 1846 if (vp->v_tag == VT_NFS && 1847 vp->v_type != VBLK) { 1848 if (bp->b_dirtyend > 0) { 1849 bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff); 1850 bp->b_validend = max(bp->b_validend, bp->b_dirtyend); 1851 } 1852 if (bp->b_validend == 0) 1853 bp->b_flags &= ~B_CACHE; 1854 } 1855 bp->b_data = (caddr_t) trunc_page((vm_offset_t)bp->b_data); 1856 bp->b_npages = curbpnpages; 1857 pmap_qenter((vm_offset_t) bp->b_data, 1858 bp->b_pages, bp->b_npages); 1859 ((vm_offset_t) bp->b_data) |= off & PAGE_MASK; 1860 } 1861 } 1862 } 1863 if (bp->b_flags & B_VMIO) 1864 vmiospace += (newbsize - bp->b_bufsize); 1865 bufspace += (newbsize - bp->b_bufsize); 1866 bp->b_bufsize = newbsize; 1867 bp->b_bcount = size; 1868 return 1; 1869} 1870 1871/* 1872 * Wait for buffer I/O completion, returning error status. 1873 */ 1874int 1875biowait(register struct buf * bp) 1876{ 1877 int s; 1878 1879 s = splbio(); 1880 while ((bp->b_flags & B_DONE) == 0) 1881#if defined(NO_SCHEDULE_MODS) 1882 tsleep(bp, PRIBIO, "biowait", 0); 1883#else 1884 if (bp->b_flags & B_READ) 1885 tsleep(bp, PRIBIO, "biord", 0); 1886 else 1887 tsleep(bp, PRIBIO, "biowr", 0); 1888#endif 1889 splx(s); 1890 if (bp->b_flags & B_EINTR) { 1891 bp->b_flags &= ~B_EINTR; 1892 return (EINTR); 1893 } 1894 if (bp->b_flags & B_ERROR) { 1895 return (bp->b_error ? bp->b_error : EIO); 1896 } else { 1897 return (0); 1898 } 1899} 1900 1901/* 1902 * Finish I/O on a buffer, calling an optional function. 1903 * This is usually called from interrupt level, so process blocking 1904 * is not *a good idea*. 1905 */ 1906void 1907biodone(register struct buf * bp) 1908{ 1909 int s; 1910 1911 s = splbio(); 1912 1913#if !defined(MAX_PERF) 1914 if (!(bp->b_flags & B_BUSY)) 1915 panic("biodone: buffer not busy"); 1916#endif 1917 1918 if (bp->b_flags & B_DONE) { 1919 splx(s); 1920#if !defined(MAX_PERF) 1921 printf("biodone: buffer already done\n"); 1922#endif 1923 return; 1924 } 1925 bp->b_flags |= B_DONE; 1926 1927 if (bp->b_flags & B_FREEBUF) { 1928 brelse(bp); 1929 splx(s); 1930 return; 1931 } 1932 1933 if ((bp->b_flags & B_READ) == 0) { 1934 vwakeup(bp); 1935 } 1936 1937 /* call optional completion function if requested */ 1938 if (bp->b_flags & B_CALL) { 1939 bp->b_flags &= ~B_CALL; 1940 (*bp->b_iodone) (bp); 1941 splx(s); 1942 return; 1943 } 1944 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 1945 (*bioops.io_complete)(bp); 1946 1947 if (bp->b_flags & B_VMIO) { 1948 int i, resid; 1949 vm_ooffset_t foff; 1950 vm_page_t m; 1951 vm_object_t obj; 1952 int iosize; 1953 struct vnode *vp = bp->b_vp; 1954 1955 obj = vp->v_object; 1956 1957#if defined(VFS_BIO_DEBUG) 1958 if (vp->v_usecount == 0) { 1959 panic("biodone: zero vnode ref count"); 1960 } 1961 1962 if (vp->v_object == NULL) { 1963 panic("biodone: missing VM object"); 1964 } 1965 1966 if ((vp->v_flag & VOBJBUF) == 0) { 1967 panic("biodone: vnode is not setup for merged cache"); 1968 } 1969#endif 1970 1971 foff = bp->b_offset; 1972 KASSERT(bp->b_offset != NOOFFSET, 1973 ("biodone: no buffer offset")); 1974 1975#if !defined(MAX_PERF) 1976 if (!obj) { 1977 panic("biodone: no object"); 1978 } 1979#endif 1980#if defined(VFS_BIO_DEBUG) 1981 if (obj->paging_in_progress < bp->b_npages) { 1982 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 1983 obj->paging_in_progress, bp->b_npages); 1984 } 1985#endif 1986 iosize = bp->b_bufsize; 1987 for (i = 0; i < bp->b_npages; i++) { 1988 int bogusflag = 0; 1989 m = bp->b_pages[i]; 1990 if (m == bogus_page) { 1991 bogusflag = 1; 1992 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 1993 if (!m) { 1994#if defined(VFS_BIO_DEBUG) 1995 printf("biodone: page disappeared\n"); 1996#endif 1997 vm_object_pip_subtract(obj, 1); 1998 continue; 1999 } 2000 bp->b_pages[i] = m; 2001 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2002 } 2003#if defined(VFS_BIO_DEBUG) 2004 if (OFF_TO_IDX(foff) != m->pindex) { 2005 printf( 2006"biodone: foff(%lu)/m->pindex(%d) mismatch\n", 2007 (unsigned long)foff, m->pindex); 2008 } 2009#endif 2010 resid = IDX_TO_OFF(m->pindex + 1) - foff; 2011 if (resid > iosize) 2012 resid = iosize; 2013 2014 /* 2015 * In the write case, the valid and clean bits are 2016 * already changed correctly, so we only need to do this 2017 * here in the read case. 2018 */ 2019 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) { 2020 vfs_page_set_valid(bp, foff, i, m); 2021 } 2022 vm_page_flag_clear(m, PG_ZERO); 2023 2024 /* 2025 * when debugging new filesystems or buffer I/O methods, this 2026 * is the most common error that pops up. if you see this, you 2027 * have not set the page busy flag correctly!!! 2028 */ 2029 if (m->busy == 0) { 2030#if !defined(MAX_PERF) 2031 printf("biodone: page busy < 0, " 2032 "pindex: %d, foff: 0x(%x,%x), " 2033 "resid: %d, index: %d\n", 2034 (int) m->pindex, (int)(foff >> 32), 2035 (int) foff & 0xffffffff, resid, i); 2036#endif 2037 if (vp->v_type != VBLK) 2038#if !defined(MAX_PERF) 2039 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 2040 bp->b_vp->v_mount->mnt_stat.f_iosize, 2041 (int) bp->b_lblkno, 2042 bp->b_flags, bp->b_npages); 2043 else 2044 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 2045 (int) bp->b_lblkno, 2046 bp->b_flags, bp->b_npages); 2047 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2048 m->valid, m->dirty, m->wire_count); 2049#endif 2050 panic("biodone: page busy < 0\n"); 2051 } 2052 vm_page_io_finish(m); 2053 vm_object_pip_subtract(obj, 1); 2054 foff += resid; 2055 iosize -= resid; 2056 } 2057 if (obj) 2058 vm_object_pip_wakeupn(obj, 0); 2059 } 2060 /* 2061 * For asynchronous completions, release the buffer now. The brelse 2062 * checks for B_WANTED and will do the wakeup there if necessary - so 2063 * no need to do a wakeup here in the async case. 2064 */ 2065 2066 if (bp->b_flags & B_ASYNC) { 2067 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0) 2068 brelse(bp); 2069 else 2070 bqrelse(bp); 2071 } else { 2072 bp->b_flags &= ~B_WANTED; 2073 wakeup(bp); 2074 } 2075 splx(s); 2076} 2077 2078#if 0 /* not with kirks code */ 2079static int vfs_update_interval = 30; 2080 2081static void 2082vfs_update() 2083{ 2084 while (1) { 2085 tsleep(&vfs_update_wakeup, PUSER, "update", 2086 hz * vfs_update_interval); 2087 vfs_update_wakeup = 0; 2088 sync(curproc, NULL); 2089 } 2090} 2091 2092static int 2093sysctl_kern_updateinterval SYSCTL_HANDLER_ARGS 2094{ 2095 int error = sysctl_handle_int(oidp, 2096 oidp->oid_arg1, oidp->oid_arg2, req); 2097 if (!error) 2098 wakeup(&vfs_update_wakeup); 2099 return error; 2100} 2101 2102SYSCTL_PROC(_kern, KERN_UPDATEINTERVAL, update, CTLTYPE_INT|CTLFLAG_RW, 2103 &vfs_update_interval, 0, sysctl_kern_updateinterval, "I", ""); 2104 2105#endif 2106 2107 2108/* 2109 * This routine is called in lieu of iodone in the case of 2110 * incomplete I/O. This keeps the busy status for pages 2111 * consistant. 2112 */ 2113void 2114vfs_unbusy_pages(struct buf * bp) 2115{ 2116 int i; 2117 2118 if (bp->b_flags & B_VMIO) { 2119 struct vnode *vp = bp->b_vp; 2120 vm_object_t obj = vp->v_object; 2121 2122 for (i = 0; i < bp->b_npages; i++) { 2123 vm_page_t m = bp->b_pages[i]; 2124 2125 if (m == bogus_page) { 2126 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 2127#if !defined(MAX_PERF) 2128 if (!m) { 2129 panic("vfs_unbusy_pages: page missing\n"); 2130 } 2131#endif 2132 bp->b_pages[i] = m; 2133 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2134 } 2135 vm_object_pip_subtract(obj, 1); 2136 vm_page_flag_clear(m, PG_ZERO); 2137 vm_page_io_finish(m); 2138 } 2139 vm_object_pip_wakeupn(obj, 0); 2140 } 2141} 2142 2143/* 2144 * Set NFS' b_validoff and b_validend fields from the valid bits 2145 * of a page. If the consumer is not NFS, and the page is not 2146 * valid for the entire range, clear the B_CACHE flag to force 2147 * the consumer to re-read the page. 2148 * 2149 * B_CACHE interaction is especially tricky. 2150 */ 2151static void 2152vfs_buf_set_valid(struct buf *bp, 2153 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 2154 vm_page_t m) 2155{ 2156 if (bp->b_vp->v_tag == VT_NFS && bp->b_vp->v_type != VBLK) { 2157 vm_offset_t svalid, evalid; 2158 int validbits = m->valid >> (((foff+off)&PAGE_MASK)/DEV_BSIZE); 2159 2160 /* 2161 * This only bothers with the first valid range in the 2162 * page. 2163 */ 2164 svalid = off; 2165 while (validbits && !(validbits & 1)) { 2166 svalid += DEV_BSIZE; 2167 validbits >>= 1; 2168 } 2169 evalid = svalid; 2170 while (validbits & 1) { 2171 evalid += DEV_BSIZE; 2172 validbits >>= 1; 2173 } 2174 evalid = min(evalid, off + size); 2175 /* 2176 * We can only set b_validoff/end if this range is contiguous 2177 * with the range built up already. If we cannot set 2178 * b_validoff/end, we must clear B_CACHE to force an update 2179 * to clean the bp up. 2180 */ 2181 if (svalid == bp->b_validend) { 2182 bp->b_validoff = min(bp->b_validoff, svalid); 2183 bp->b_validend = max(bp->b_validend, evalid); 2184 } else { 2185 bp->b_flags &= ~B_CACHE; 2186 } 2187 } else if (!vm_page_is_valid(m, 2188 (vm_offset_t) ((foff + off) & PAGE_MASK), 2189 size)) { 2190 bp->b_flags &= ~B_CACHE; 2191 } 2192} 2193 2194/* 2195 * Set the valid bits in a page, taking care of the b_validoff, 2196 * b_validend fields which NFS uses to optimise small reads. Off is 2197 * the offset within the file and pageno is the page index within the buf. 2198 * 2199 * XXX we have to set the valid & clean bits for all page fragments 2200 * touched by b_validoff/validend, even if the page fragment goes somewhat 2201 * beyond b_validoff/validend due to alignment. 2202 */ 2203static void 2204vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 2205{ 2206 struct vnode *vp = bp->b_vp; 2207 vm_ooffset_t soff, eoff; 2208 2209 soff = off; 2210 eoff = (off + PAGE_SIZE) & ~PAGE_MASK; 2211 if (eoff > bp->b_offset + bp->b_bufsize) 2212 eoff = bp->b_offset + bp->b_bufsize; 2213 if (vp->v_tag == VT_NFS && vp->v_type != VBLK) { 2214 vm_ooffset_t sv, ev; 2215 vm_page_set_invalid(m, 2216 (vm_offset_t) (soff & PAGE_MASK), 2217 (vm_offset_t) (eoff - soff)); 2218 sv = (bp->b_offset + bp->b_validoff + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2219 ev = (bp->b_offset + bp->b_validend + (DEV_BSIZE - 1)) & 2220 ~(DEV_BSIZE - 1); 2221 soff = qmax(sv, soff); 2222 eoff = qmin(ev, eoff); 2223 } 2224 if (eoff > soff) 2225 vm_page_set_validclean(m, 2226 (vm_offset_t) (soff & PAGE_MASK), 2227 (vm_offset_t) (eoff - soff)); 2228} 2229 2230/* 2231 * This routine is called before a device strategy routine. 2232 * It is used to tell the VM system that paging I/O is in 2233 * progress, and treat the pages associated with the buffer 2234 * almost as being PG_BUSY. Also the object paging_in_progress 2235 * flag is handled to make sure that the object doesn't become 2236 * inconsistant. 2237 */ 2238void 2239vfs_busy_pages(struct buf * bp, int clear_modify) 2240{ 2241 int i, bogus; 2242 2243 if (bp->b_flags & B_VMIO) { 2244 struct vnode *vp = bp->b_vp; 2245 vm_object_t obj = vp->v_object; 2246 vm_ooffset_t foff; 2247 2248 foff = bp->b_offset; 2249 KASSERT(bp->b_offset != NOOFFSET, 2250 ("vfs_busy_pages: no buffer offset")); 2251 vfs_setdirty(bp); 2252 2253retry: 2254 for (i = 0; i < bp->b_npages; i++) { 2255 vm_page_t m = bp->b_pages[i]; 2256 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 2257 goto retry; 2258 } 2259 2260 bogus = 0; 2261 for (i = 0; i < bp->b_npages; i++) { 2262 vm_page_t m = bp->b_pages[i]; 2263 2264 vm_page_flag_clear(m, PG_ZERO); 2265 if ((bp->b_flags & B_CLUSTER) == 0) { 2266 vm_object_pip_add(obj, 1); 2267 vm_page_io_start(m); 2268 } 2269 2270 vm_page_protect(m, VM_PROT_NONE); 2271 if (clear_modify) 2272 vfs_page_set_valid(bp, foff, i, m); 2273 else if (m->valid == VM_PAGE_BITS_ALL && 2274 (bp->b_flags & B_CACHE) == 0) { 2275 bp->b_pages[i] = bogus_page; 2276 bogus++; 2277 } 2278 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 2279 } 2280 if (bogus) 2281 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2282 } 2283} 2284 2285/* 2286 * Tell the VM system that the pages associated with this buffer 2287 * are clean. This is used for delayed writes where the data is 2288 * going to go to disk eventually without additional VM intevention. 2289 */ 2290void 2291vfs_clean_pages(struct buf * bp) 2292{ 2293 int i; 2294 2295 if (bp->b_flags & B_VMIO) { 2296 vm_ooffset_t foff; 2297 foff = bp->b_offset; 2298 KASSERT(bp->b_offset != NOOFFSET, 2299 ("vfs_clean_pages: no buffer offset")); 2300 for (i = 0; i < bp->b_npages; i++) { 2301 vm_page_t m = bp->b_pages[i]; 2302 vfs_page_set_valid(bp, foff, i, m); 2303 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 2304 } 2305 } 2306} 2307 2308void 2309vfs_bio_clrbuf(struct buf *bp) { 2310 int i, mask = 0; 2311 caddr_t sa, ea; 2312 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 2313 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 2314 (bp->b_offset & PAGE_MASK) == 0) { 2315 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 2316 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 2317 ((bp->b_pages[0]->valid & mask) != mask)) { 2318 bzero(bp->b_data, bp->b_bufsize); 2319 } 2320 bp->b_pages[0]->valid |= mask; 2321 bp->b_resid = 0; 2322 return; 2323 } 2324 ea = sa = bp->b_data; 2325 for(i=0;i<bp->b_npages;i++,sa=ea) { 2326 int j = ((u_long)sa & PAGE_MASK) / DEV_BSIZE; 2327 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 2328 ea = (caddr_t)ulmin((u_long)ea, 2329 (u_long)bp->b_data + bp->b_bufsize); 2330 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 2331 if ((bp->b_pages[i]->valid & mask) == mask) 2332 continue; 2333 if ((bp->b_pages[i]->valid & mask) == 0) { 2334 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 2335 bzero(sa, ea - sa); 2336 } 2337 } else { 2338 for (; sa < ea; sa += DEV_BSIZE, j++) { 2339 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 2340 (bp->b_pages[i]->valid & (1<<j)) == 0) 2341 bzero(sa, DEV_BSIZE); 2342 } 2343 } 2344 bp->b_pages[i]->valid |= mask; 2345 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2346 } 2347 bp->b_resid = 0; 2348 } else { 2349 clrbuf(bp); 2350 } 2351} 2352 2353/* 2354 * vm_hold_load_pages and vm_hold_unload pages get pages into 2355 * a buffers address space. The pages are anonymous and are 2356 * not associated with a file object. 2357 */ 2358void 2359vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 2360{ 2361 vm_offset_t pg; 2362 vm_page_t p; 2363 int index; 2364 2365 to = round_page(to); 2366 from = round_page(from); 2367 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 2368 2369 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 2370 2371tryagain: 2372 2373 p = vm_page_alloc(kernel_object, 2374 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 2375 VM_ALLOC_NORMAL); 2376 if (!p) { 2377 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 2378 VM_WAIT; 2379 goto tryagain; 2380 } 2381 vm_page_wire(p); 2382 p->valid = VM_PAGE_BITS_ALL; 2383 vm_page_flag_clear(p, PG_ZERO); 2384 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 2385 bp->b_pages[index] = p; 2386 vm_page_wakeup(p); 2387 } 2388 bp->b_npages = index; 2389} 2390 2391void 2392vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 2393{ 2394 vm_offset_t pg; 2395 vm_page_t p; 2396 int index, newnpages; 2397 2398 from = round_page(from); 2399 to = round_page(to); 2400 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 2401 2402 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 2403 p = bp->b_pages[index]; 2404 if (p && (index < bp->b_npages)) { 2405#if !defined(MAX_PERF) 2406 if (p->busy) { 2407 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 2408 bp->b_blkno, bp->b_lblkno); 2409 } 2410#endif 2411 bp->b_pages[index] = NULL; 2412 pmap_kremove(pg); 2413 vm_page_busy(p); 2414 vm_page_unwire(p, 0); 2415 vm_page_free(p); 2416 } 2417 } 2418 bp->b_npages = newnpages; 2419} 2420 2421 2422#include "opt_ddb.h" 2423#ifdef DDB 2424#include <ddb/ddb.h> 2425 2426DB_SHOW_COMMAND(buffer, db_show_buffer) 2427{ 2428 /* get args */ 2429 struct buf *bp = (struct buf *)addr; 2430 2431 if (!have_addr) { 2432 db_printf("usage: show buffer <addr>\n"); 2433 return; 2434 } 2435 2436 db_printf("b_proc = %p,\nb_flags = 0x%b\n", (void *)bp->b_proc, 2437 (u_int)bp->b_flags, PRINT_BUF_FLAGS); 2438 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 2439 "b_resid = %ld\nb_dev = 0x%x, b_data = %p, " 2440 "b_blkno = %d, b_pblkno = %d\n", 2441 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 2442 bp->b_dev, bp->b_data, bp->b_blkno, bp->b_pblkno); 2443 if (bp->b_npages) { 2444 int i; 2445 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 2446 for (i = 0; i < bp->b_npages; i++) { 2447 vm_page_t m; 2448 m = bp->b_pages[i]; 2449 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 2450 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 2451 if ((i + 1) < bp->b_npages) 2452 db_printf(","); 2453 } 2454 db_printf("\n"); 2455 } 2456} 2457#endif /* DDB */ 2458