vfs_bio.c revision 42961
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 1.194 1999/01/21 08:29:05 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_NOCACHE | B_INVAL | B_ERROR | B_FREEBUF)) || 581 (bp->b_bufsize <= 0)) { 582 bp->b_flags |= B_INVAL; 583 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 584 (*bioops.io_deallocate)(bp); 585 if (bp->b_flags & B_DELWRI) 586 --numdirtybuffers; 587 bp->b_flags &= ~(B_DELWRI | B_CACHE | B_FREEBUF); 588 if ((bp->b_flags & B_VMIO) == 0) { 589 if (bp->b_bufsize) 590 allocbuf(bp, 0); 591 if (bp->b_vp) 592 brelvp(bp); 593 } 594 } 595 596 /* 597 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 598 * is called with B_DELWRI set, the underlying pages may wind up 599 * getting freed causing a previous write (bdwrite()) to get 'lost' 600 * because pages associated with a B_DELWRI bp are marked clean. 601 * 602 * We still allow the B_INVAL case to call vfs_vmio_release(), even 603 * if B_DELWRI is set. 604 */ 605 606 if (bp->b_flags & B_DELWRI) 607 bp->b_flags &= ~B_RELBUF; 608 609 /* 610 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 611 * constituted, so the B_INVAL flag is used to *invalidate* the buffer, 612 * but the VM object is kept around. The B_NOCACHE flag is used to 613 * invalidate the pages in the VM object. 614 * 615 * The b_{validoff,validend,dirtyoff,dirtyend} values are relative 616 * to b_offset and currently have byte granularity, whereas the 617 * valid flags in the vm_pages have only DEV_BSIZE resolution. 618 * The byte resolution fields are used to avoid unnecessary re-reads 619 * of the buffer but the code really needs to be genericized so 620 * other filesystem modules can take advantage of these fields. 621 * 622 * XXX this seems to cause performance problems. 623 */ 624 if ((bp->b_flags & B_VMIO) 625 && !(bp->b_vp->v_tag == VT_NFS && 626 bp->b_vp->v_type != VBLK && 627 (bp->b_flags & B_DELWRI) != 0) 628#ifdef notdef 629 && (bp->b_vp->v_tag != VT_NFS 630 || bp->b_vp->v_type == VBLK 631 || (bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) 632 || bp->b_validend == 0 633 || (bp->b_validoff == 0 634 && bp->b_validend == bp->b_bufsize)) 635#endif 636 ) { 637 638 int i, j, resid; 639 vm_page_t m; 640 off_t foff; 641 vm_pindex_t poff; 642 vm_object_t obj; 643 struct vnode *vp; 644 645 vp = bp->b_vp; 646 647 /* 648 * Get the base offset and length of the buffer. Note that 649 * for block sizes that are less then PAGE_SIZE, the b_data 650 * base of the buffer does not represent exactly b_offset and 651 * neither b_offset nor b_size are necessarily page aligned. 652 * Instead, the starting position of b_offset is: 653 * 654 * b_data + (b_offset & PAGE_MASK) 655 * 656 * block sizes less then DEV_BSIZE (usually 512) are not 657 * supported due to the page granularity bits (m->valid, 658 * m->dirty, etc...). 659 * 660 * See man buf(9) for more information 661 */ 662 663 resid = bp->b_bufsize; 664 foff = bp->b_offset; 665 666 for (i = 0; i < bp->b_npages; i++) { 667 m = bp->b_pages[i]; 668 vm_page_flag_clear(m, PG_ZERO); 669 if (m == bogus_page) { 670 671 obj = (vm_object_t) vp->v_object; 672 poff = OFF_TO_IDX(bp->b_offset); 673 674 for (j = i; j < bp->b_npages; j++) { 675 m = bp->b_pages[j]; 676 if (m == bogus_page) { 677 m = vm_page_lookup(obj, poff + j); 678#if !defined(MAX_PERF) 679 if (!m) { 680 panic("brelse: page missing\n"); 681 } 682#endif 683 bp->b_pages[j] = m; 684 } 685 } 686 687 if ((bp->b_flags & B_INVAL) == 0) { 688 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 689 } 690 } 691 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 692 int poffset = foff & PAGE_MASK; 693 int presid = resid > (PAGE_SIZE - poffset) ? 694 (PAGE_SIZE - poffset) : resid; 695 696 KASSERT(presid >= 0, ("brelse: extra page")); 697 vm_page_set_invalid(m, poffset, presid); 698 } 699 resid -= PAGE_SIZE - (foff & PAGE_MASK); 700 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 701 } 702 703 if (bp->b_flags & (B_INVAL | B_RELBUF)) 704 vfs_vmio_release(bp); 705 706 } else if (bp->b_flags & B_VMIO) { 707 708 if (bp->b_flags & (B_INVAL | B_RELBUF)) 709 vfs_vmio_release(bp); 710 711 } 712 713#if !defined(MAX_PERF) 714 if (bp->b_qindex != QUEUE_NONE) 715 panic("brelse: free buffer onto another queue???"); 716#endif 717 718 /* enqueue */ 719 /* buffers with no memory */ 720 if (bp->b_bufsize == 0) { 721 bp->b_flags |= B_INVAL; 722 bp->b_qindex = QUEUE_EMPTY; 723 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 724 LIST_REMOVE(bp, b_hash); 725 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 726 bp->b_dev = NODEV; 727 kvafreespace += bp->b_kvasize; 728 729 /* buffers with junk contents */ 730 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) { 731 bp->b_flags |= B_INVAL; 732 bp->b_qindex = QUEUE_AGE; 733 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_AGE], bp, b_freelist); 734 LIST_REMOVE(bp, b_hash); 735 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 736 bp->b_dev = NODEV; 737 738 /* buffers that are locked */ 739 } else if (bp->b_flags & B_LOCKED) { 740 bp->b_qindex = QUEUE_LOCKED; 741 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 742 743 /* buffers with stale but valid contents */ 744 } else if (bp->b_flags & B_AGE) { 745 bp->b_qindex = QUEUE_AGE; 746 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_AGE], bp, b_freelist); 747 748 /* buffers with valid and quite potentially reuseable contents */ 749 } else { 750 bp->b_qindex = QUEUE_LRU; 751 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 752 } 753 754 if ((bp->b_flags & B_INVAL) || 755 (bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) { 756 if (bp->b_flags & B_DELWRI) { 757 --numdirtybuffers; 758 bp->b_flags &= ~B_DELWRI; 759 } 760 vfs_bio_need_satisfy(); 761 } 762 763 /* unlock */ 764 bp->b_flags &= ~(B_ORDERED | B_WANTED | B_BUSY | 765 B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 766 splx(s); 767} 768 769/* 770 * Release a buffer. 771 */ 772void 773bqrelse(struct buf * bp) 774{ 775 int s; 776 777 s = splbio(); 778 779 /* anyone need this block? */ 780 if (bp->b_flags & B_WANTED) { 781 bp->b_flags &= ~(B_WANTED | B_AGE); 782 wakeup(bp); 783 } 784 785#if !defined(MAX_PERF) 786 if (bp->b_qindex != QUEUE_NONE) 787 panic("bqrelse: free buffer onto another queue???"); 788#endif 789 790 if (bp->b_flags & B_LOCKED) { 791 bp->b_flags &= ~B_ERROR; 792 bp->b_qindex = QUEUE_LOCKED; 793 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 794 /* buffers with stale but valid contents */ 795 } else { 796 bp->b_qindex = QUEUE_LRU; 797 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 798 } 799 800 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) { 801 vfs_bio_need_satisfy(); 802 } 803 804 /* unlock */ 805 bp->b_flags &= ~(B_ORDERED | B_WANTED | B_BUSY | 806 B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 807 splx(s); 808} 809 810static void 811vfs_vmio_release(bp) 812 struct buf *bp; 813{ 814 int i, s; 815 vm_page_t m; 816 817 s = splvm(); 818 for (i = 0; i < bp->b_npages; i++) { 819 m = bp->b_pages[i]; 820 bp->b_pages[i] = NULL; 821 /* 822 * In order to keep page LRU ordering consistent, put 823 * everything on the inactive queue. 824 */ 825 vm_page_unwire(m, 0); 826 /* 827 * We don't mess with busy pages, it is 828 * the responsibility of the process that 829 * busied the pages to deal with them. 830 */ 831 if ((m->flags & PG_BUSY) || (m->busy != 0)) 832 continue; 833 834 if (m->wire_count == 0) { 835 vm_page_flag_clear(m, PG_ZERO); 836 /* 837 * Might as well free the page if we can and it has 838 * no valid data. 839 */ 840 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { 841 vm_page_busy(m); 842 vm_page_protect(m, VM_PROT_NONE); 843 vm_page_free(m); 844 } 845 } 846 } 847 splx(s); 848 bufspace -= bp->b_bufsize; 849 vmiospace -= bp->b_bufsize; 850 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 851 bp->b_npages = 0; 852 bp->b_bufsize = 0; 853 bp->b_flags &= ~B_VMIO; 854 if (bp->b_vp) 855 brelvp(bp); 856} 857 858/* 859 * Check to see if a block is currently memory resident. 860 */ 861struct buf * 862gbincore(struct vnode * vp, daddr_t blkno) 863{ 864 struct buf *bp; 865 struct bufhashhdr *bh; 866 867 bh = BUFHASH(vp, blkno); 868 bp = bh->lh_first; 869 870 /* Search hash chain */ 871 while (bp != NULL) { 872 /* hit */ 873 if (bp->b_vp == vp && bp->b_lblkno == blkno && 874 (bp->b_flags & B_INVAL) == 0) { 875 break; 876 } 877 bp = bp->b_hash.le_next; 878 } 879 return (bp); 880} 881 882/* 883 * this routine implements clustered async writes for 884 * clearing out B_DELWRI buffers... This is much better 885 * than the old way of writing only one buffer at a time. 886 */ 887int 888vfs_bio_awrite(struct buf * bp) 889{ 890 int i; 891 daddr_t lblkno = bp->b_lblkno; 892 struct vnode *vp = bp->b_vp; 893 int s; 894 int ncl; 895 struct buf *bpa; 896 int nwritten; 897 int size; 898 int maxcl; 899 900 s = splbio(); 901 /* 902 * right now we support clustered writing only to regular files 903 */ 904 if ((vp->v_type == VREG) && 905 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 906 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 907 908 size = vp->v_mount->mnt_stat.f_iosize; 909 maxcl = MAXPHYS / size; 910 911 for (i = 1; i < maxcl; i++) { 912 if ((bpa = gbincore(vp, lblkno + i)) && 913 ((bpa->b_flags & (B_BUSY | B_DELWRI | B_CLUSTEROK | B_INVAL)) == 914 (B_DELWRI | B_CLUSTEROK)) && 915 (bpa->b_bufsize == size)) { 916 if ((bpa->b_blkno == bpa->b_lblkno) || 917 (bpa->b_blkno != bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 918 break; 919 } else { 920 break; 921 } 922 } 923 ncl = i; 924 /* 925 * this is a possible cluster write 926 */ 927 if (ncl != 1) { 928 nwritten = cluster_wbuild(vp, size, lblkno, ncl); 929 splx(s); 930 return nwritten; 931 } 932 } 933 934 bremfree(bp); 935 bp->b_flags |= B_BUSY | B_ASYNC; 936 937 splx(s); 938 /* 939 * default (old) behavior, writing out only one block 940 */ 941 nwritten = bp->b_bufsize; 942 (void) VOP_BWRITE(bp); 943 return nwritten; 944} 945 946 947/* 948 * Find a buffer header which is available for use. 949 */ 950static struct buf * 951getnewbuf(struct vnode *vp, daddr_t blkno, 952 int slpflag, int slptimeo, int size, int maxsize) 953{ 954 struct buf *bp, *bp1; 955 int nbyteswritten = 0; 956 vm_offset_t addr; 957 static int writerecursion = 0; 958 959start: 960 if (bufspace >= maxbufspace) 961 goto trytofreespace; 962 963 /* can we constitute a new buffer? */ 964 if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]))) { 965#if !defined(MAX_PERF) 966 if (bp->b_qindex != QUEUE_EMPTY) 967 panic("getnewbuf: inconsistent EMPTY queue, qindex=%d", 968 bp->b_qindex); 969#endif 970 bp->b_flags |= B_BUSY; 971 bremfree(bp); 972 goto fillbuf; 973 } 974trytofreespace: 975 /* 976 * We keep the file I/O from hogging metadata I/O 977 * This is desirable because file data is cached in the 978 * VM/Buffer cache even if a buffer is freed. 979 */ 980 if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]))) { 981#if !defined(MAX_PERF) 982 if (bp->b_qindex != QUEUE_AGE) 983 panic("getnewbuf: inconsistent AGE queue, qindex=%d", 984 bp->b_qindex); 985#endif 986 } else if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]))) { 987#if !defined(MAX_PERF) 988 if (bp->b_qindex != QUEUE_LRU) 989 panic("getnewbuf: inconsistent LRU queue, qindex=%d", 990 bp->b_qindex); 991#endif 992 } 993 if (!bp) { 994 /* wait for a free buffer of any kind */ 995 needsbuffer |= VFS_BIO_NEED_ANY; 996 do 997 tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, "newbuf", 998 slptimeo); 999 while (needsbuffer & VFS_BIO_NEED_ANY); 1000 return (0); 1001 } 1002 KASSERT(!(bp->b_flags & B_BUSY), 1003 ("getnewbuf: busy buffer on free list\n")); 1004 /* 1005 * We are fairly aggressive about freeing VMIO buffers, but since 1006 * the buffering is intact without buffer headers, there is not 1007 * much loss. We gain by maintaining non-VMIOed metadata in buffers. 1008 */ 1009 if ((bp->b_qindex == QUEUE_LRU) && (bp->b_usecount > 0)) { 1010 if ((bp->b_flags & B_VMIO) == 0 || 1011 (vmiospace < maxvmiobufspace)) { 1012 --bp->b_usecount; 1013 TAILQ_REMOVE(&bufqueues[QUEUE_LRU], bp, b_freelist); 1014 if (TAILQ_FIRST(&bufqueues[QUEUE_LRU]) != NULL) { 1015 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1016 goto start; 1017 } 1018 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1019 } 1020 } 1021 1022 1023 /* if we are a delayed write, convert to an async write */ 1024 if ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) { 1025 1026 /* 1027 * If our delayed write is likely to be used soon, then 1028 * recycle back onto the LRU queue. 1029 */ 1030 if (vp && (bp->b_vp == vp) && (bp->b_qindex == QUEUE_LRU) && 1031 (bp->b_lblkno >= blkno) && (maxsize > 0)) { 1032 1033 if (bp->b_usecount > 0) { 1034 if (bp->b_lblkno < blkno + (MAXPHYS / maxsize)) { 1035 1036 TAILQ_REMOVE(&bufqueues[QUEUE_LRU], bp, b_freelist); 1037 1038 if (TAILQ_FIRST(&bufqueues[QUEUE_LRU]) != NULL) { 1039 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1040 bp->b_usecount--; 1041 goto start; 1042 } 1043 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1044 } 1045 } 1046 } 1047 1048 /* 1049 * Certain layered filesystems can recursively re-enter the vfs_bio 1050 * code, due to delayed writes. This helps keep the system from 1051 * deadlocking. 1052 */ 1053 if (writerecursion > 0) { 1054 if (writerecursion > 5) { 1055 bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]); 1056 while (bp) { 1057 if ((bp->b_flags & B_DELWRI) == 0) 1058 break; 1059 bp = TAILQ_NEXT(bp, b_freelist); 1060 } 1061 if (bp == NULL) { 1062 bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]); 1063 while (bp) { 1064 if ((bp->b_flags & B_DELWRI) == 0) 1065 break; 1066 bp = TAILQ_NEXT(bp, b_freelist); 1067 } 1068 } 1069 if (bp == NULL) 1070 panic("getnewbuf: cannot get buffer, infinite recursion failure"); 1071 } else { 1072 bremfree(bp); 1073 bp->b_flags |= B_BUSY | B_AGE | B_ASYNC; 1074 nbyteswritten += bp->b_bufsize; 1075 ++writerecursion; 1076 VOP_BWRITE(bp); 1077 --writerecursion; 1078 if (!slpflag && !slptimeo) { 1079 return (0); 1080 } 1081 goto start; 1082 } 1083 } else { 1084 ++writerecursion; 1085 nbyteswritten += vfs_bio_awrite(bp); 1086 --writerecursion; 1087 if (!slpflag && !slptimeo) { 1088 return (0); 1089 } 1090 goto start; 1091 } 1092 } 1093 1094 if (bp->b_flags & B_WANTED) { 1095 bp->b_flags &= ~B_WANTED; 1096 wakeup(bp); 1097 } 1098 bremfree(bp); 1099 bp->b_flags |= B_BUSY; 1100 1101 if (bp->b_flags & B_VMIO) { 1102 bp->b_flags &= ~B_ASYNC; 1103 vfs_vmio_release(bp); 1104 } 1105 1106 if (bp->b_vp) 1107 brelvp(bp); 1108 1109fillbuf: 1110 1111 /* we are not free, nor do we contain interesting data */ 1112 if (bp->b_rcred != NOCRED) { 1113 crfree(bp->b_rcred); 1114 bp->b_rcred = NOCRED; 1115 } 1116 if (bp->b_wcred != NOCRED) { 1117 crfree(bp->b_wcred); 1118 bp->b_wcred = NOCRED; 1119 } 1120 if (LIST_FIRST(&bp->b_dep) != NULL && 1121 bioops.io_deallocate) 1122 (*bioops.io_deallocate)(bp); 1123 1124 LIST_REMOVE(bp, b_hash); 1125 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1126 if (bp->b_bufsize) { 1127 allocbuf(bp, 0); 1128 } 1129 bp->b_flags = B_BUSY; 1130 bp->b_dev = NODEV; 1131 bp->b_vp = NULL; 1132 bp->b_blkno = bp->b_lblkno = 0; 1133 bp->b_offset = NOOFFSET; 1134 bp->b_iodone = 0; 1135 bp->b_error = 0; 1136 bp->b_resid = 0; 1137 bp->b_bcount = 0; 1138 bp->b_npages = 0; 1139 bp->b_dirtyoff = bp->b_dirtyend = 0; 1140 bp->b_validoff = bp->b_validend = 0; 1141 bp->b_usecount = 5; 1142 /* Here, not kern_physio.c, is where this should be done*/ 1143 LIST_INIT(&bp->b_dep); 1144 1145 maxsize = (maxsize + PAGE_MASK) & ~PAGE_MASK; 1146 1147 /* 1148 * we assume that buffer_map is not at address 0 1149 */ 1150 addr = 0; 1151 if (maxsize != bp->b_kvasize) { 1152 bfreekva(bp); 1153 1154findkvaspace: 1155 /* 1156 * See if we have buffer kva space 1157 */ 1158 if (vm_map_findspace(buffer_map, 1159 vm_map_min(buffer_map), maxsize, &addr)) { 1160 if (kvafreespace > 0) { 1161 int totfree = 0, freed; 1162 do { 1163 freed = 0; 1164 for (bp1 = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1165 bp1 != NULL; bp1 = TAILQ_NEXT(bp1, b_freelist)) { 1166 if (bp1->b_kvasize != 0) { 1167 totfree += bp1->b_kvasize; 1168 freed = bp1->b_kvasize; 1169 bremfree(bp1); 1170 bfreekva(bp1); 1171 brelse(bp1); 1172 break; 1173 } 1174 } 1175 } while (freed); 1176 /* 1177 * if we found free space, then retry with the same buffer. 1178 */ 1179 if (totfree) 1180 goto findkvaspace; 1181 } 1182 bp->b_flags |= B_INVAL; 1183 brelse(bp); 1184 goto trytofreespace; 1185 } 1186 } 1187 1188 /* 1189 * See if we are below are allocated minimum 1190 */ 1191 if (bufspace >= (maxbufspace + nbyteswritten)) { 1192 bp->b_flags |= B_INVAL; 1193 brelse(bp); 1194 goto trytofreespace; 1195 } 1196 1197 /* 1198 * create a map entry for the buffer -- in essence 1199 * reserving the kva space. 1200 */ 1201 if (addr) { 1202 vm_map_insert(buffer_map, NULL, 0, 1203 addr, addr + maxsize, 1204 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1205 1206 bp->b_kvabase = (caddr_t) addr; 1207 bp->b_kvasize = maxsize; 1208 } 1209 bp->b_data = bp->b_kvabase; 1210 1211 return (bp); 1212} 1213 1214static void 1215waitfreebuffers(int slpflag, int slptimeo) { 1216 while (numfreebuffers < hifreebuffers) { 1217 flushdirtybuffers(slpflag, slptimeo); 1218 if (numfreebuffers < hifreebuffers) 1219 break; 1220 needsbuffer |= VFS_BIO_NEED_FREE; 1221 if (tsleep(&needsbuffer, (PRIBIO + 4)|slpflag, "biofre", slptimeo)) 1222 break; 1223 } 1224} 1225 1226static void 1227flushdirtybuffers(int slpflag, int slptimeo) { 1228 int s; 1229 static pid_t flushing = 0; 1230 1231 s = splbio(); 1232 1233 if (flushing) { 1234 if (flushing == curproc->p_pid) { 1235 splx(s); 1236 return; 1237 } 1238 while (flushing) { 1239 if (tsleep(&flushing, (PRIBIO + 4)|slpflag, "biofls", slptimeo)) { 1240 splx(s); 1241 return; 1242 } 1243 } 1244 } 1245 flushing = curproc->p_pid; 1246 1247 while (numdirtybuffers > lodirtybuffers) { 1248 struct buf *bp; 1249 needsbuffer |= VFS_BIO_NEED_LOWLIMIT; 1250 bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]); 1251 if (bp == NULL) 1252 bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]); 1253 1254 while (bp && ((bp->b_flags & B_DELWRI) == 0)) { 1255 bp = TAILQ_NEXT(bp, b_freelist); 1256 } 1257 1258 if (bp) { 1259 vfs_bio_awrite(bp); 1260 continue; 1261 } 1262 break; 1263 } 1264 1265 flushing = 0; 1266 wakeup(&flushing); 1267 splx(s); 1268} 1269 1270/* 1271 * Check to see if a block is currently memory resident. 1272 */ 1273struct buf * 1274incore(struct vnode * vp, daddr_t blkno) 1275{ 1276 struct buf *bp; 1277 1278 int s = splbio(); 1279 bp = gbincore(vp, blkno); 1280 splx(s); 1281 return (bp); 1282} 1283 1284/* 1285 * Returns true if no I/O is needed to access the 1286 * associated VM object. This is like incore except 1287 * it also hunts around in the VM system for the data. 1288 */ 1289 1290int 1291inmem(struct vnode * vp, daddr_t blkno) 1292{ 1293 vm_object_t obj; 1294 vm_offset_t toff, tinc, size; 1295 vm_page_t m; 1296 vm_ooffset_t off; 1297 1298 if (incore(vp, blkno)) 1299 return 1; 1300 if (vp->v_mount == NULL) 1301 return 0; 1302 if ((vp->v_object == NULL) || (vp->v_flag & VOBJBUF) == 0) 1303 return 0; 1304 1305 obj = vp->v_object; 1306 size = PAGE_SIZE; 1307 if (size > vp->v_mount->mnt_stat.f_iosize) 1308 size = vp->v_mount->mnt_stat.f_iosize; 1309 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 1310 1311 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1312 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 1313 if (!m) 1314 return 0; 1315 tinc = size; 1316 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 1317 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 1318 if (vm_page_is_valid(m, 1319 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 1320 return 0; 1321 } 1322 return 1; 1323} 1324 1325/* 1326 * now we set the dirty range for the buffer -- 1327 * for NFS -- if the file is mapped and pages have 1328 * been written to, let it know. We want the 1329 * entire range of the buffer to be marked dirty if 1330 * any of the pages have been written to for consistancy 1331 * with the b_validoff, b_validend set in the nfs write 1332 * code, and used by the nfs read code. 1333 */ 1334static void 1335vfs_setdirty(struct buf *bp) { 1336 int i; 1337 vm_object_t object; 1338 vm_offset_t boffset; 1339#if 0 1340 vm_offset_t offset; 1341#endif 1342 1343 /* 1344 * We qualify the scan for modified pages on whether the 1345 * object has been flushed yet. The OBJ_WRITEABLE flag 1346 * is not cleared simply by protecting pages off. 1347 */ 1348 if ((bp->b_flags & B_VMIO) && 1349 ((object = bp->b_pages[0]->object)->flags & (OBJ_WRITEABLE|OBJ_CLEANING))) { 1350 /* 1351 * test the pages to see if they have been modified directly 1352 * by users through the VM system. 1353 */ 1354 for (i = 0; i < bp->b_npages; i++) { 1355 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 1356 vm_page_test_dirty(bp->b_pages[i]); 1357 } 1358 1359 /* 1360 * scan forwards for the first page modified 1361 */ 1362 for (i = 0; i < bp->b_npages; i++) { 1363 if (bp->b_pages[i]->dirty) { 1364 break; 1365 } 1366 } 1367 1368 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 1369 if (boffset < bp->b_dirtyoff) { 1370 bp->b_dirtyoff = max(boffset, 0); 1371 } 1372 1373 /* 1374 * scan backwards for the last page modified 1375 */ 1376 for (i = bp->b_npages - 1; i >= 0; --i) { 1377 if (bp->b_pages[i]->dirty) { 1378 break; 1379 } 1380 } 1381 boffset = (i + 1); 1382#if 0 1383 offset = boffset + bp->b_pages[0]->pindex; 1384 if (offset >= object->size) 1385 boffset = object->size - bp->b_pages[0]->pindex; 1386#endif 1387 boffset = (boffset << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 1388 if (bp->b_dirtyend < boffset) 1389 bp->b_dirtyend = min(boffset, bp->b_bufsize); 1390 } 1391} 1392 1393/* 1394 * Get a block given a specified block and offset into a file/device. 1395 */ 1396struct buf * 1397getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 1398{ 1399 struct buf *bp; 1400 int i, s; 1401 struct bufhashhdr *bh; 1402 1403#if !defined(MAX_PERF) 1404 if (size > MAXBSIZE) 1405 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 1406#endif 1407 1408 s = splbio(); 1409loop: 1410 if (numfreebuffers < lofreebuffers) { 1411 waitfreebuffers(slpflag, slptimeo); 1412 } 1413 1414 if ((bp = gbincore(vp, blkno))) { 1415 if (bp->b_flags & B_BUSY) { 1416 bp->b_flags |= B_WANTED; 1417 if (bp->b_usecount < BUF_MAXUSE) 1418 ++bp->b_usecount; 1419 1420 if (!tsleep(bp, 1421 (PRIBIO + 4) | slpflag, "getblk", slptimeo)) { 1422 goto loop; 1423 } 1424 1425 splx(s); 1426 return (struct buf *) NULL; 1427 } 1428 bp->b_flags |= B_BUSY | B_CACHE; 1429 bremfree(bp); 1430 1431 /* 1432 * check for size inconsistancies for non-VMIO case. 1433 */ 1434 1435 if (bp->b_bcount != size) { 1436 if ((bp->b_flags & B_VMIO) == 0 || 1437 (size > bp->b_kvasize) 1438 ) { 1439 if (bp->b_flags & B_DELWRI) { 1440 bp->b_flags |= B_NOCACHE; 1441 VOP_BWRITE(bp); 1442 } else { 1443 if ((bp->b_flags & B_VMIO) && 1444 (LIST_FIRST(&bp->b_dep) == NULL)) { 1445 bp->b_flags |= B_RELBUF; 1446 brelse(bp); 1447 } else { 1448 bp->b_flags |= B_NOCACHE; 1449 VOP_BWRITE(bp); 1450 } 1451 } 1452 goto loop; 1453 } 1454 } 1455 1456 /* 1457 * If the size is inconsistant in the VMIO case, we can resize 1458 * the buffer. This might lead to B_CACHE getting cleared. 1459 */ 1460 1461 if (bp->b_bcount != size) 1462 allocbuf(bp, size); 1463 1464 KASSERT(bp->b_offset != NOOFFSET, 1465 ("getblk: no buffer offset")); 1466 1467 /* 1468 * Check that the constituted buffer really deserves for the 1469 * B_CACHE bit to be set. B_VMIO type buffers might not 1470 * contain fully valid pages. Normal (old-style) buffers 1471 * should be fully valid. This might also lead to B_CACHE 1472 * getting clear. 1473 */ 1474 if ((bp->b_flags & B_VMIO|B_CACHE) == (B_VMIO|B_CACHE)) { 1475 int checksize = bp->b_bufsize; 1476 int poffset = bp->b_offset & PAGE_MASK; 1477 int resid; 1478 for (i = 0; i < bp->b_npages; i++) { 1479 resid = (checksize > (PAGE_SIZE - poffset)) ? 1480 (PAGE_SIZE - poffset) : checksize; 1481 if (!vm_page_is_valid(bp->b_pages[i], poffset, resid)) { 1482 bp->b_flags &= ~(B_CACHE | B_DONE); 1483 break; 1484 } 1485 checksize -= resid; 1486 poffset = 0; 1487 } 1488 } 1489 1490 /* 1491 * If B_DELWRI is set and B_CACHE got cleared ( or was 1492 * already clear ), we have to commit the write and 1493 * retry. The NFS code absolutely depends on this, 1494 * and so might the FFS code. In anycase, it formalizes 1495 * the B_CACHE rules. See sys/buf.h. 1496 */ 1497 1498 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 1499 VOP_BWRITE(bp); 1500 goto loop; 1501 } 1502 1503 if (bp->b_usecount < BUF_MAXUSE) 1504 ++bp->b_usecount; 1505 splx(s); 1506 return (bp); 1507 } else { 1508 int bsize, maxsize, vmio; 1509 off_t offset; 1510 1511 if (vp->v_type == VBLK) 1512 bsize = DEV_BSIZE; 1513 else if (vp->v_mountedhere) 1514 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 1515 else if (vp->v_mount) 1516 bsize = vp->v_mount->mnt_stat.f_iosize; 1517 else 1518 bsize = size; 1519 1520 offset = (off_t)blkno * bsize; 1521 vmio = (vp->v_object != 0) && (vp->v_flag & VOBJBUF); 1522 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 1523 maxsize = imax(maxsize, bsize); 1524 1525 if ((bp = getnewbuf(vp, blkno, 1526 slpflag, slptimeo, size, maxsize)) == 0) { 1527 if (slpflag || slptimeo) { 1528 splx(s); 1529 return NULL; 1530 } 1531 goto loop; 1532 } 1533 1534 /* 1535 * This code is used to make sure that a buffer is not 1536 * created while the getnewbuf routine is blocked. 1537 * Normally the vnode is locked so this isn't a problem. 1538 * VBLK type I/O requests, however, don't lock the vnode. 1539 */ 1540 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE && gbincore(vp, blkno)) { 1541 bp->b_flags |= B_INVAL; 1542 brelse(bp); 1543 goto loop; 1544 } 1545 1546 /* 1547 * Insert the buffer into the hash, so that it can 1548 * be found by incore. 1549 */ 1550 bp->b_blkno = bp->b_lblkno = blkno; 1551 bp->b_offset = offset; 1552 1553 bgetvp(vp, bp); 1554 LIST_REMOVE(bp, b_hash); 1555 bh = BUFHASH(vp, blkno); 1556 LIST_INSERT_HEAD(bh, bp, b_hash); 1557 1558 if (vmio) { 1559 bp->b_flags |= (B_VMIO | B_CACHE); 1560#if defined(VFS_BIO_DEBUG) 1561 if (vp->v_type != VREG && vp->v_type != VBLK) 1562 printf("getblk: vmioing file type %d???\n", vp->v_type); 1563#endif 1564 } else { 1565 bp->b_flags &= ~B_VMIO; 1566 } 1567 1568 allocbuf(bp, size); 1569 1570 splx(s); 1571 return (bp); 1572 } 1573} 1574 1575/* 1576 * Get an empty, disassociated buffer of given size. 1577 */ 1578struct buf * 1579geteblk(int size) 1580{ 1581 struct buf *bp; 1582 int s; 1583 1584 s = splbio(); 1585 while ((bp = getnewbuf(0, (daddr_t) 0, 0, 0, size, MAXBSIZE)) == 0); 1586 splx(s); 1587 allocbuf(bp, size); 1588 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 1589 return (bp); 1590} 1591 1592 1593/* 1594 * This code constitutes the buffer memory from either anonymous system 1595 * memory (in the case of non-VMIO operations) or from an associated 1596 * VM object (in the case of VMIO operations). This code is able to 1597 * resize a buffer up or down. 1598 * 1599 * Note that this code is tricky, and has many complications to resolve 1600 * deadlock or inconsistant data situations. Tread lightly!!! 1601 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 1602 * the caller. Calling this code willy nilly can result in the loss of data. 1603 */ 1604 1605int 1606allocbuf(struct buf *bp, int size) 1607{ 1608 int newbsize, mbsize; 1609 int i; 1610 1611#if !defined(MAX_PERF) 1612 if (!(bp->b_flags & B_BUSY)) 1613 panic("allocbuf: buffer not busy"); 1614 1615 if (bp->b_kvasize < size) 1616 panic("allocbuf: buffer too small"); 1617#endif 1618 1619 if ((bp->b_flags & B_VMIO) == 0) { 1620 caddr_t origbuf; 1621 int origbufsize; 1622 /* 1623 * Just get anonymous memory from the kernel 1624 */ 1625 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 1626#if !defined(NO_B_MALLOC) 1627 if (bp->b_flags & B_MALLOC) 1628 newbsize = mbsize; 1629 else 1630#endif 1631 newbsize = round_page(size); 1632 1633 if (newbsize < bp->b_bufsize) { 1634#if !defined(NO_B_MALLOC) 1635 /* 1636 * malloced buffers are not shrunk 1637 */ 1638 if (bp->b_flags & B_MALLOC) { 1639 if (newbsize) { 1640 bp->b_bcount = size; 1641 } else { 1642 free(bp->b_data, M_BIOBUF); 1643 bufspace -= bp->b_bufsize; 1644 bufmallocspace -= bp->b_bufsize; 1645 bp->b_data = bp->b_kvabase; 1646 bp->b_bufsize = 0; 1647 bp->b_bcount = 0; 1648 bp->b_flags &= ~B_MALLOC; 1649 } 1650 return 1; 1651 } 1652#endif 1653 vm_hold_free_pages( 1654 bp, 1655 (vm_offset_t) bp->b_data + newbsize, 1656 (vm_offset_t) bp->b_data + bp->b_bufsize); 1657 } else if (newbsize > bp->b_bufsize) { 1658#if !defined(NO_B_MALLOC) 1659 /* 1660 * We only use malloced memory on the first allocation. 1661 * and revert to page-allocated memory when the buffer grows. 1662 */ 1663 if ( (bufmallocspace < maxbufmallocspace) && 1664 (bp->b_bufsize == 0) && 1665 (mbsize <= PAGE_SIZE/2)) { 1666 1667 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 1668 bp->b_bufsize = mbsize; 1669 bp->b_bcount = size; 1670 bp->b_flags |= B_MALLOC; 1671 bufspace += mbsize; 1672 bufmallocspace += mbsize; 1673 return 1; 1674 } 1675#endif 1676 origbuf = NULL; 1677 origbufsize = 0; 1678#if !defined(NO_B_MALLOC) 1679 /* 1680 * If the buffer is growing on its other-than-first allocation, 1681 * then we revert to the page-allocation scheme. 1682 */ 1683 if (bp->b_flags & B_MALLOC) { 1684 origbuf = bp->b_data; 1685 origbufsize = bp->b_bufsize; 1686 bp->b_data = bp->b_kvabase; 1687 bufspace -= bp->b_bufsize; 1688 bufmallocspace -= bp->b_bufsize; 1689 bp->b_bufsize = 0; 1690 bp->b_flags &= ~B_MALLOC; 1691 newbsize = round_page(newbsize); 1692 } 1693#endif 1694 vm_hold_load_pages( 1695 bp, 1696 (vm_offset_t) bp->b_data + bp->b_bufsize, 1697 (vm_offset_t) bp->b_data + newbsize); 1698#if !defined(NO_B_MALLOC) 1699 if (origbuf) { 1700 bcopy(origbuf, bp->b_data, origbufsize); 1701 free(origbuf, M_BIOBUF); 1702 } 1703#endif 1704 } 1705 } else { 1706 vm_page_t m; 1707 int desiredpages; 1708 1709 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 1710 desiredpages = (size == 0) ? 0 : 1711 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 1712 1713#if !defined(NO_B_MALLOC) 1714 if (bp->b_flags & B_MALLOC) 1715 panic("allocbuf: VMIO buffer can't be malloced"); 1716#endif 1717 1718 if (newbsize < bp->b_bufsize) { 1719 if (desiredpages < bp->b_npages) { 1720 for (i = desiredpages; i < bp->b_npages; i++) { 1721 /* 1722 * the page is not freed here -- it 1723 * is the responsibility of vnode_pager_setsize 1724 */ 1725 m = bp->b_pages[i]; 1726 KASSERT(m != bogus_page, 1727 ("allocbuf: bogus page found")); 1728 while (vm_page_sleep_busy(m, TRUE, "biodep")) 1729 ; 1730 1731 bp->b_pages[i] = NULL; 1732 vm_page_unwire(m, 0); 1733 } 1734 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 1735 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 1736 bp->b_npages = desiredpages; 1737 } 1738 } else if (newbsize > bp->b_bufsize) { 1739 vm_object_t obj; 1740 vm_offset_t tinc, toff; 1741 vm_ooffset_t off; 1742 vm_pindex_t objoff; 1743 int pageindex, curbpnpages; 1744 struct vnode *vp; 1745 int bsize; 1746 int orig_validoff = bp->b_validoff; 1747 int orig_validend = bp->b_validend; 1748 1749 vp = bp->b_vp; 1750 1751 if (vp->v_type == VBLK) 1752 bsize = DEV_BSIZE; 1753 else 1754 bsize = vp->v_mount->mnt_stat.f_iosize; 1755 1756 if (bp->b_npages < desiredpages) { 1757 obj = vp->v_object; 1758 tinc = PAGE_SIZE; 1759 1760 off = bp->b_offset; 1761 KASSERT(bp->b_offset != NOOFFSET, 1762 ("allocbuf: no buffer offset")); 1763 curbpnpages = bp->b_npages; 1764 doretry: 1765 bp->b_validoff = orig_validoff; 1766 bp->b_validend = orig_validend; 1767 bp->b_flags |= B_CACHE; 1768 for (toff = 0; toff < newbsize; toff += tinc) { 1769 objoff = OFF_TO_IDX(off + toff); 1770 pageindex = objoff - OFF_TO_IDX(off); 1771 tinc = PAGE_SIZE - ((off + toff) & PAGE_MASK); 1772 if (pageindex < curbpnpages) { 1773 1774 m = bp->b_pages[pageindex]; 1775#ifdef VFS_BIO_DIAG 1776 if (m->pindex != objoff) 1777 panic("allocbuf: page changed offset?!!!?"); 1778#endif 1779 if (tinc > (newbsize - toff)) 1780 tinc = newbsize - toff; 1781 if (bp->b_flags & B_CACHE) 1782 vfs_buf_set_valid(bp, off, toff, tinc, m); 1783 continue; 1784 } 1785 m = vm_page_lookup(obj, objoff); 1786 if (!m) { 1787 m = vm_page_alloc(obj, objoff, VM_ALLOC_NORMAL); 1788 if (!m) { 1789 VM_WAIT; 1790 vm_pageout_deficit += (desiredpages - curbpnpages); 1791 goto doretry; 1792 } 1793 1794 vm_page_wire(m); 1795 vm_page_wakeup(m); 1796 bp->b_flags &= ~B_CACHE; 1797 1798 } else if (vm_page_sleep_busy(m, FALSE, "pgtblk")) { 1799 /* 1800 * If we had to sleep, retry. 1801 * 1802 * Also note that we only test 1803 * PG_BUSY here, not m->busy. 1804 * 1805 * We cannot sleep on m->busy 1806 * here because a vm_fault -> 1807 * getpages -> cluster-read -> 1808 * ...-> allocbuf sequence 1809 * will convert PG_BUSY to 1810 * m->busy so we have to let 1811 * m->busy through if we do 1812 * not want to deadlock. 1813 */ 1814 goto doretry; 1815 } else { 1816 if ((curproc != pageproc) && 1817 ((m->queue - m->pc) == PQ_CACHE) && 1818 ((cnt.v_free_count + cnt.v_cache_count) < 1819 (cnt.v_free_min + cnt.v_cache_min))) { 1820 pagedaemon_wakeup(); 1821 } 1822 if (tinc > (newbsize - toff)) 1823 tinc = newbsize - toff; 1824 if (bp->b_flags & B_CACHE) 1825 vfs_buf_set_valid(bp, off, toff, tinc, m); 1826 vm_page_flag_clear(m, PG_ZERO); 1827 vm_page_wire(m); 1828 } 1829 bp->b_pages[pageindex] = m; 1830 curbpnpages = pageindex + 1; 1831 } 1832 if (vp->v_tag == VT_NFS && 1833 vp->v_type != VBLK) { 1834 if (bp->b_dirtyend > 0) { 1835 bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff); 1836 bp->b_validend = max(bp->b_validend, bp->b_dirtyend); 1837 } 1838 if (bp->b_validend == 0) 1839 bp->b_flags &= ~B_CACHE; 1840 } 1841 bp->b_data = (caddr_t) trunc_page((vm_offset_t)bp->b_data); 1842 bp->b_npages = curbpnpages; 1843 pmap_qenter((vm_offset_t) bp->b_data, 1844 bp->b_pages, bp->b_npages); 1845 ((vm_offset_t) bp->b_data) |= off & PAGE_MASK; 1846 } 1847 } 1848 } 1849 if (bp->b_flags & B_VMIO) 1850 vmiospace += (newbsize - bp->b_bufsize); 1851 bufspace += (newbsize - bp->b_bufsize); 1852 bp->b_bufsize = newbsize; 1853 bp->b_bcount = size; 1854 return 1; 1855} 1856 1857/* 1858 * Wait for buffer I/O completion, returning error status. 1859 */ 1860int 1861biowait(register struct buf * bp) 1862{ 1863 int s; 1864 1865 s = splbio(); 1866 while ((bp->b_flags & B_DONE) == 0) 1867#if defined(NO_SCHEDULE_MODS) 1868 tsleep(bp, PRIBIO, "biowait", 0); 1869#else 1870 if (bp->b_flags & B_READ) 1871 tsleep(bp, PRIBIO, "biord", 0); 1872 else 1873 tsleep(bp, PRIBIO, "biowr", 0); 1874#endif 1875 splx(s); 1876 if (bp->b_flags & B_EINTR) { 1877 bp->b_flags &= ~B_EINTR; 1878 return (EINTR); 1879 } 1880 if (bp->b_flags & B_ERROR) { 1881 return (bp->b_error ? bp->b_error : EIO); 1882 } else { 1883 return (0); 1884 } 1885} 1886 1887/* 1888 * Finish I/O on a buffer, calling an optional function. 1889 * This is usually called from interrupt level, so process blocking 1890 * is not *a good idea*. 1891 */ 1892void 1893biodone(register struct buf * bp) 1894{ 1895 int s; 1896 1897 s = splbio(); 1898 1899#if !defined(MAX_PERF) 1900 if (!(bp->b_flags & B_BUSY)) 1901 panic("biodone: buffer not busy"); 1902#endif 1903 1904 if (bp->b_flags & B_DONE) { 1905 splx(s); 1906#if !defined(MAX_PERF) 1907 printf("biodone: buffer already done\n"); 1908#endif 1909 return; 1910 } 1911 bp->b_flags |= B_DONE; 1912 1913 if (bp->b_flags & B_FREEBUF) { 1914 brelse(bp); 1915 splx(s); 1916 return; 1917 } 1918 1919 if ((bp->b_flags & B_READ) == 0) { 1920 vwakeup(bp); 1921 } 1922 1923 /* call optional completion function if requested */ 1924 if (bp->b_flags & B_CALL) { 1925 bp->b_flags &= ~B_CALL; 1926 (*bp->b_iodone) (bp); 1927 splx(s); 1928 return; 1929 } 1930 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 1931 (*bioops.io_complete)(bp); 1932 1933 if (bp->b_flags & B_VMIO) { 1934 int i, resid; 1935 vm_ooffset_t foff; 1936 vm_page_t m; 1937 vm_object_t obj; 1938 int iosize; 1939 struct vnode *vp = bp->b_vp; 1940 1941 obj = vp->v_object; 1942 1943#if defined(VFS_BIO_DEBUG) 1944 if (vp->v_usecount == 0) { 1945 panic("biodone: zero vnode ref count"); 1946 } 1947 1948 if (vp->v_object == NULL) { 1949 panic("biodone: missing VM object"); 1950 } 1951 1952 if ((vp->v_flag & VOBJBUF) == 0) { 1953 panic("biodone: vnode is not setup for merged cache"); 1954 } 1955#endif 1956 1957 foff = bp->b_offset; 1958 KASSERT(bp->b_offset != NOOFFSET, 1959 ("biodone: no buffer offset")); 1960 1961#if !defined(MAX_PERF) 1962 if (!obj) { 1963 panic("biodone: no object"); 1964 } 1965#endif 1966#if defined(VFS_BIO_DEBUG) 1967 if (obj->paging_in_progress < bp->b_npages) { 1968 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 1969 obj->paging_in_progress, bp->b_npages); 1970 } 1971#endif 1972 iosize = bp->b_bufsize; 1973 for (i = 0; i < bp->b_npages; i++) { 1974 int bogusflag = 0; 1975 m = bp->b_pages[i]; 1976 if (m == bogus_page) { 1977 bogusflag = 1; 1978 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 1979 if (!m) { 1980#if defined(VFS_BIO_DEBUG) 1981 printf("biodone: page disappeared\n"); 1982#endif 1983 vm_object_pip_subtract(obj, 1); 1984 continue; 1985 } 1986 bp->b_pages[i] = m; 1987 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1988 } 1989#if defined(VFS_BIO_DEBUG) 1990 if (OFF_TO_IDX(foff) != m->pindex) { 1991 printf( 1992"biodone: foff(%lu)/m->pindex(%d) mismatch\n", 1993 (unsigned long)foff, m->pindex); 1994 } 1995#endif 1996 resid = IDX_TO_OFF(m->pindex + 1) - foff; 1997 if (resid > iosize) 1998 resid = iosize; 1999 2000 /* 2001 * In the write case, the valid and clean bits are 2002 * already changed correctly, so we only need to do this 2003 * here in the read case. 2004 */ 2005 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) { 2006 vfs_page_set_valid(bp, foff, i, m); 2007 } 2008 vm_page_flag_clear(m, PG_ZERO); 2009 2010 /* 2011 * when debugging new filesystems or buffer I/O methods, this 2012 * is the most common error that pops up. if you see this, you 2013 * have not set the page busy flag correctly!!! 2014 */ 2015 if (m->busy == 0) { 2016#if !defined(MAX_PERF) 2017 printf("biodone: page busy < 0, " 2018 "pindex: %d, foff: 0x(%x,%x), " 2019 "resid: %d, index: %d\n", 2020 (int) m->pindex, (int)(foff >> 32), 2021 (int) foff & 0xffffffff, resid, i); 2022#endif 2023 if (vp->v_type != VBLK) 2024#if !defined(MAX_PERF) 2025 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 2026 bp->b_vp->v_mount->mnt_stat.f_iosize, 2027 (int) bp->b_lblkno, 2028 bp->b_flags, bp->b_npages); 2029 else 2030 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 2031 (int) bp->b_lblkno, 2032 bp->b_flags, bp->b_npages); 2033 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2034 m->valid, m->dirty, m->wire_count); 2035#endif 2036 panic("biodone: page busy < 0\n"); 2037 } 2038 vm_page_io_finish(m); 2039 vm_object_pip_subtract(obj, 1); 2040 foff += resid; 2041 iosize -= resid; 2042 } 2043 if (obj) 2044 vm_object_pip_wakeupn(obj, 0); 2045 } 2046 /* 2047 * For asynchronous completions, release the buffer now. The brelse 2048 * checks for B_WANTED and will do the wakeup there if necessary - so 2049 * no need to do a wakeup here in the async case. 2050 */ 2051 2052 if (bp->b_flags & B_ASYNC) { 2053 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0) 2054 brelse(bp); 2055 else 2056 bqrelse(bp); 2057 } else { 2058 bp->b_flags &= ~B_WANTED; 2059 wakeup(bp); 2060 } 2061 splx(s); 2062} 2063 2064#if 0 /* not with kirks code */ 2065static int vfs_update_interval = 30; 2066 2067static void 2068vfs_update() 2069{ 2070 while (1) { 2071 tsleep(&vfs_update_wakeup, PUSER, "update", 2072 hz * vfs_update_interval); 2073 vfs_update_wakeup = 0; 2074 sync(curproc, NULL); 2075 } 2076} 2077 2078static int 2079sysctl_kern_updateinterval SYSCTL_HANDLER_ARGS 2080{ 2081 int error = sysctl_handle_int(oidp, 2082 oidp->oid_arg1, oidp->oid_arg2, req); 2083 if (!error) 2084 wakeup(&vfs_update_wakeup); 2085 return error; 2086} 2087 2088SYSCTL_PROC(_kern, KERN_UPDATEINTERVAL, update, CTLTYPE_INT|CTLFLAG_RW, 2089 &vfs_update_interval, 0, sysctl_kern_updateinterval, "I", ""); 2090 2091#endif 2092 2093 2094/* 2095 * This routine is called in lieu of iodone in the case of 2096 * incomplete I/O. This keeps the busy status for pages 2097 * consistant. 2098 */ 2099void 2100vfs_unbusy_pages(struct buf * bp) 2101{ 2102 int i; 2103 2104 if (bp->b_flags & B_VMIO) { 2105 struct vnode *vp = bp->b_vp; 2106 vm_object_t obj = vp->v_object; 2107 2108 for (i = 0; i < bp->b_npages; i++) { 2109 vm_page_t m = bp->b_pages[i]; 2110 2111 if (m == bogus_page) { 2112 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 2113#if !defined(MAX_PERF) 2114 if (!m) { 2115 panic("vfs_unbusy_pages: page missing\n"); 2116 } 2117#endif 2118 bp->b_pages[i] = m; 2119 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2120 } 2121 vm_object_pip_subtract(obj, 1); 2122 vm_page_flag_clear(m, PG_ZERO); 2123 vm_page_io_finish(m); 2124 } 2125 vm_object_pip_wakeupn(obj, 0); 2126 } 2127} 2128 2129/* 2130 * Set NFS' b_validoff and b_validend fields from the valid bits 2131 * of a page. If the consumer is not NFS, and the page is not 2132 * valid for the entire range, clear the B_CACHE flag to force 2133 * the consumer to re-read the page. 2134 * 2135 * B_CACHE interaction is especially tricky. 2136 */ 2137static void 2138vfs_buf_set_valid(struct buf *bp, 2139 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 2140 vm_page_t m) 2141{ 2142 if (bp->b_vp->v_tag == VT_NFS && bp->b_vp->v_type != VBLK) { 2143 vm_offset_t svalid, evalid; 2144 int validbits = m->valid >> (((foff+off)&PAGE_MASK)/DEV_BSIZE); 2145 2146 /* 2147 * This only bothers with the first valid range in the 2148 * page. 2149 */ 2150 svalid = off; 2151 while (validbits && !(validbits & 1)) { 2152 svalid += DEV_BSIZE; 2153 validbits >>= 1; 2154 } 2155 evalid = svalid; 2156 while (validbits & 1) { 2157 evalid += DEV_BSIZE; 2158 validbits >>= 1; 2159 } 2160 evalid = min(evalid, off + size); 2161 /* 2162 * We can only set b_validoff/end if this range is contiguous 2163 * with the range built up already. If we cannot set 2164 * b_validoff/end, we must clear B_CACHE to force an update 2165 * to clean the bp up. 2166 */ 2167 if (svalid == bp->b_validend) { 2168 bp->b_validoff = min(bp->b_validoff, svalid); 2169 bp->b_validend = max(bp->b_validend, evalid); 2170 } else { 2171 bp->b_flags &= ~B_CACHE; 2172 } 2173 } else if (!vm_page_is_valid(m, 2174 (vm_offset_t) ((foff + off) & PAGE_MASK), 2175 size)) { 2176 bp->b_flags &= ~B_CACHE; 2177 } 2178} 2179 2180/* 2181 * Set the valid bits in a page, taking care of the b_validoff, 2182 * b_validend fields which NFS uses to optimise small reads. Off is 2183 * the offset within the file and pageno is the page index within the buf. 2184 * 2185 * XXX we have to set the valid & clean bits for all page fragments 2186 * touched by b_validoff/validend, even if the page fragment goes somewhat 2187 * beyond b_validoff/validend due to alignment. 2188 */ 2189static void 2190vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 2191{ 2192 struct vnode *vp = bp->b_vp; 2193 vm_ooffset_t soff, eoff; 2194 2195 soff = off; 2196 eoff = (off + PAGE_SIZE) & ~PAGE_MASK; 2197 if (eoff > bp->b_offset + bp->b_bufsize) 2198 eoff = bp->b_offset + bp->b_bufsize; 2199 if (vp->v_tag == VT_NFS && vp->v_type != VBLK) { 2200 vm_ooffset_t sv, ev; 2201 vm_page_set_invalid(m, 2202 (vm_offset_t) (soff & PAGE_MASK), 2203 (vm_offset_t) (eoff - soff)); 2204 sv = (bp->b_offset + bp->b_validoff + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2205 ev = (bp->b_offset + bp->b_validend + (DEV_BSIZE - 1)) & 2206 ~(DEV_BSIZE - 1); 2207 soff = qmax(sv, soff); 2208 eoff = qmin(ev, eoff); 2209 } 2210 if (eoff > soff) 2211 vm_page_set_validclean(m, 2212 (vm_offset_t) (soff & PAGE_MASK), 2213 (vm_offset_t) (eoff - soff)); 2214} 2215 2216/* 2217 * This routine is called before a device strategy routine. 2218 * It is used to tell the VM system that paging I/O is in 2219 * progress, and treat the pages associated with the buffer 2220 * almost as being PG_BUSY. Also the object paging_in_progress 2221 * flag is handled to make sure that the object doesn't become 2222 * inconsistant. 2223 */ 2224void 2225vfs_busy_pages(struct buf * bp, int clear_modify) 2226{ 2227 int i, bogus; 2228 2229 if (bp->b_flags & B_VMIO) { 2230 struct vnode *vp = bp->b_vp; 2231 vm_object_t obj = vp->v_object; 2232 vm_ooffset_t foff; 2233 2234 foff = bp->b_offset; 2235 KASSERT(bp->b_offset != NOOFFSET, 2236 ("vfs_busy_pages: no buffer offset")); 2237 vfs_setdirty(bp); 2238 2239retry: 2240 for (i = 0; i < bp->b_npages; i++) { 2241 vm_page_t m = bp->b_pages[i]; 2242 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 2243 goto retry; 2244 } 2245 2246 bogus = 0; 2247 for (i = 0; i < bp->b_npages; i++) { 2248 vm_page_t m = bp->b_pages[i]; 2249 2250 vm_page_flag_clear(m, PG_ZERO); 2251 if ((bp->b_flags & B_CLUSTER) == 0) { 2252 vm_object_pip_add(obj, 1); 2253 vm_page_io_start(m); 2254 } 2255 2256 vm_page_protect(m, VM_PROT_NONE); 2257 if (clear_modify) 2258 vfs_page_set_valid(bp, foff, i, m); 2259 else if (m->valid == VM_PAGE_BITS_ALL && 2260 (bp->b_flags & B_CACHE) == 0) { 2261 bp->b_pages[i] = bogus_page; 2262 bogus++; 2263 } 2264 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 2265 } 2266 if (bogus) 2267 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2268 } 2269} 2270 2271/* 2272 * Tell the VM system that the pages associated with this buffer 2273 * are clean. This is used for delayed writes where the data is 2274 * going to go to disk eventually without additional VM intevention. 2275 */ 2276void 2277vfs_clean_pages(struct buf * bp) 2278{ 2279 int i; 2280 2281 if (bp->b_flags & B_VMIO) { 2282 vm_ooffset_t foff; 2283 foff = bp->b_offset; 2284 KASSERT(bp->b_offset != NOOFFSET, 2285 ("vfs_clean_pages: no buffer offset")); 2286 for (i = 0; i < bp->b_npages; i++) { 2287 vm_page_t m = bp->b_pages[i]; 2288 vfs_page_set_valid(bp, foff, i, m); 2289 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 2290 } 2291 } 2292} 2293 2294void 2295vfs_bio_clrbuf(struct buf *bp) { 2296 int i, mask = 0; 2297 caddr_t sa, ea; 2298 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 2299 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 2300 (bp->b_offset & PAGE_MASK) == 0) { 2301 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 2302 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 2303 ((bp->b_pages[0]->valid & mask) != mask)) { 2304 bzero(bp->b_data, bp->b_bufsize); 2305 } 2306 bp->b_pages[0]->valid |= mask; 2307 bp->b_resid = 0; 2308 return; 2309 } 2310 ea = sa = bp->b_data; 2311 for(i=0;i<bp->b_npages;i++,sa=ea) { 2312 int j = ((u_long)sa & PAGE_MASK) / DEV_BSIZE; 2313 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 2314 ea = (caddr_t)ulmin((u_long)ea, 2315 (u_long)bp->b_data + bp->b_bufsize); 2316 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 2317 if ((bp->b_pages[i]->valid & mask) == mask) 2318 continue; 2319 if ((bp->b_pages[i]->valid & mask) == 0) { 2320 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 2321 bzero(sa, ea - sa); 2322 } 2323 } else { 2324 for (; sa < ea; sa += DEV_BSIZE, j++) { 2325 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 2326 (bp->b_pages[i]->valid & (1<<j)) == 0) 2327 bzero(sa, DEV_BSIZE); 2328 } 2329 } 2330 bp->b_pages[i]->valid |= mask; 2331 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2332 } 2333 bp->b_resid = 0; 2334 } else { 2335 clrbuf(bp); 2336 } 2337} 2338 2339/* 2340 * vm_hold_load_pages and vm_hold_unload pages get pages into 2341 * a buffers address space. The pages are anonymous and are 2342 * not associated with a file object. 2343 */ 2344void 2345vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 2346{ 2347 vm_offset_t pg; 2348 vm_page_t p; 2349 int index; 2350 2351 to = round_page(to); 2352 from = round_page(from); 2353 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 2354 2355 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 2356 2357tryagain: 2358 2359 p = vm_page_alloc(kernel_object, 2360 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 2361 VM_ALLOC_NORMAL); 2362 if (!p) { 2363 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 2364 VM_WAIT; 2365 goto tryagain; 2366 } 2367 vm_page_wire(p); 2368 p->valid = VM_PAGE_BITS_ALL; 2369 vm_page_flag_clear(p, PG_ZERO); 2370 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 2371 bp->b_pages[index] = p; 2372 vm_page_wakeup(p); 2373 } 2374 bp->b_npages = index; 2375} 2376 2377void 2378vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 2379{ 2380 vm_offset_t pg; 2381 vm_page_t p; 2382 int index, newnpages; 2383 2384 from = round_page(from); 2385 to = round_page(to); 2386 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 2387 2388 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 2389 p = bp->b_pages[index]; 2390 if (p && (index < bp->b_npages)) { 2391#if !defined(MAX_PERF) 2392 if (p->busy) { 2393 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 2394 bp->b_blkno, bp->b_lblkno); 2395 } 2396#endif 2397 bp->b_pages[index] = NULL; 2398 pmap_kremove(pg); 2399 vm_page_busy(p); 2400 vm_page_unwire(p, 0); 2401 vm_page_free(p); 2402 } 2403 } 2404 bp->b_npages = newnpages; 2405} 2406 2407 2408#include "opt_ddb.h" 2409#ifdef DDB 2410#include <ddb/ddb.h> 2411 2412DB_SHOW_COMMAND(buffer, db_show_buffer) 2413{ 2414 /* get args */ 2415 struct buf *bp = (struct buf *)addr; 2416 2417 if (!have_addr) { 2418 db_printf("usage: show buffer <addr>\n"); 2419 return; 2420 } 2421 2422 db_printf("b_proc = %p,\nb_flags = 0x%b\n", (void *)bp->b_proc, 2423 (u_int)bp->b_flags, PRINT_BUF_FLAGS); 2424 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 2425 "b_resid = %ld\nb_dev = 0x%x, b_data = %p, " 2426 "b_blkno = %d, b_pblkno = %d\n", 2427 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 2428 bp->b_dev, bp->b_data, bp->b_blkno, bp->b_pblkno); 2429 if (bp->b_npages) { 2430 int i; 2431 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 2432 for (i = 0; i < bp->b_npages; i++) { 2433 vm_page_t m; 2434 m = bp->b_pages[i]; 2435 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 2436 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 2437 if ((i + 1) < bp->b_npages) 2438 db_printf(","); 2439 } 2440 db_printf("\n"); 2441 } 2442} 2443#endif /* DDB */ 2444