vfs_bio.c revision 77085
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 * $FreeBSD: head/sys/kern/vfs_bio.c 77085 2001-05-23 22:24:49Z jhb $ 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#include <sys/param.h> 31#include <sys/systm.h> 32#include <sys/bio.h> 33#include <sys/buf.h> 34#include <sys/eventhandler.h> 35#include <sys/lock.h> 36#include <sys/malloc.h> 37#include <sys/mount.h> 38#include <sys/mutex.h> 39#include <sys/kernel.h> 40#include <sys/kthread.h> 41#include <sys/ktr.h> 42#include <sys/proc.h> 43#include <sys/reboot.h> 44#include <sys/resourcevar.h> 45#include <sys/sysctl.h> 46#include <sys/vmmeter.h> 47#include <sys/vnode.h> 48#include <vm/vm.h> 49#include <vm/vm_param.h> 50#include <vm/vm_kern.h> 51#include <vm/vm_pageout.h> 52#include <vm/vm_page.h> 53#include <vm/vm_object.h> 54#include <vm/vm_extern.h> 55#include <vm/vm_map.h> 56 57static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 58 59struct bio_ops bioops; /* I/O operation notification */ 60 61struct buf_ops buf_ops_bio = { 62 "buf_ops_bio", 63 bwrite 64}; 65 66struct buf *buf; /* buffer header pool */ 67struct swqueue bswlist; 68struct mtx buftimelock; /* Interlock on setting prio and timo */ 69 70static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 71 vm_offset_t to); 72static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 73 vm_offset_t to); 74static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 75 int pageno, vm_page_t m); 76static void vfs_clean_pages(struct buf * bp); 77static void vfs_setdirty(struct buf *bp); 78static void vfs_vmio_release(struct buf *bp); 79static void vfs_backgroundwritedone(struct buf *bp); 80static int flushbufqueues(void); 81 82static int bd_request; 83 84static void buf_daemon __P((void)); 85/* 86 * bogus page -- for I/O to/from partially complete buffers 87 * this is a temporary solution to the problem, but it is not 88 * really that bad. it would be better to split the buffer 89 * for input in the case of buffers partially already in memory, 90 * but the code is intricate enough already. 91 */ 92vm_page_t bogus_page; 93int vmiodirenable = FALSE; 94int runningbufspace; 95static vm_offset_t bogus_offset; 96 97static int bufspace, maxbufspace, 98 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace; 99static int bufreusecnt, bufdefragcnt, buffreekvacnt; 100static int needsbuffer; 101static int lorunningspace, hirunningspace, runningbufreq; 102static int numdirtybuffers, lodirtybuffers, hidirtybuffers; 103static int numfreebuffers, lofreebuffers, hifreebuffers; 104static int getnewbufcalls; 105static int getnewbufrestarts; 106 107SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, 108 &numdirtybuffers, 0, ""); 109SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, 110 &lodirtybuffers, 0, ""); 111SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, 112 &hidirtybuffers, 0, ""); 113SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, 114 &numfreebuffers, 0, ""); 115SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, 116 &lofreebuffers, 0, ""); 117SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, 118 &hifreebuffers, 0, ""); 119SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, 120 &runningbufspace, 0, ""); 121SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, 122 &lorunningspace, 0, ""); 123SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, 124 &hirunningspace, 0, ""); 125SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, 126 &maxbufspace, 0, ""); 127SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, 128 &hibufspace, 0, ""); 129SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, 130 &lobufspace, 0, ""); 131SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, 132 &bufspace, 0, ""); 133SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, 134 &maxbufmallocspace, 0, ""); 135SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, 136 &bufmallocspace, 0, ""); 137SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, 138 &getnewbufcalls, 0, ""); 139SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, 140 &getnewbufrestarts, 0, ""); 141SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, 142 &vmiodirenable, 0, ""); 143SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, 144 &bufdefragcnt, 0, ""); 145SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, 146 &buffreekvacnt, 0, ""); 147SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, 148 &bufreusecnt, 0, ""); 149 150static int bufhashmask; 151static LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash; 152struct bqueues bufqueues[BUFFER_QUEUES] = { { 0 } }; 153char *buf_wmesg = BUF_WMESG; 154 155extern int vm_swap_size; 156 157#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 158#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 159#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 160#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 161 162/* 163 * Buffer hash table code. Note that the logical block scans linearly, which 164 * gives us some L1 cache locality. 165 */ 166 167static __inline 168struct bufhashhdr * 169bufhash(struct vnode *vnp, daddr_t bn) 170{ 171 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]); 172} 173 174/* 175 * numdirtywakeup: 176 * 177 * If someone is blocked due to there being too many dirty buffers, 178 * and numdirtybuffers is now reasonable, wake them up. 179 */ 180 181static __inline void 182numdirtywakeup(int level) 183{ 184 if (numdirtybuffers <= level) { 185 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 186 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 187 wakeup(&needsbuffer); 188 } 189 } 190} 191 192/* 193 * bufspacewakeup: 194 * 195 * Called when buffer space is potentially available for recovery. 196 * getnewbuf() will block on this flag when it is unable to free 197 * sufficient buffer space. Buffer space becomes recoverable when 198 * bp's get placed back in the queues. 199 */ 200 201static __inline void 202bufspacewakeup(void) 203{ 204 /* 205 * If someone is waiting for BUF space, wake them up. Even 206 * though we haven't freed the kva space yet, the waiting 207 * process will be able to now. 208 */ 209 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 210 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 211 wakeup(&needsbuffer); 212 } 213} 214 215/* 216 * runningbufwakeup() - in-progress I/O accounting. 217 * 218 */ 219static __inline void 220runningbufwakeup(struct buf *bp) 221{ 222 if (bp->b_runningbufspace) { 223 runningbufspace -= bp->b_runningbufspace; 224 bp->b_runningbufspace = 0; 225 if (runningbufreq && runningbufspace <= lorunningspace) { 226 runningbufreq = 0; 227 wakeup(&runningbufreq); 228 } 229 } 230} 231 232/* 233 * bufcountwakeup: 234 * 235 * Called when a buffer has been added to one of the free queues to 236 * account for the buffer and to wakeup anyone waiting for free buffers. 237 * This typically occurs when large amounts of metadata are being handled 238 * by the buffer cache ( else buffer space runs out first, usually ). 239 */ 240 241static __inline void 242bufcountwakeup(void) 243{ 244 ++numfreebuffers; 245 if (needsbuffer) { 246 needsbuffer &= ~VFS_BIO_NEED_ANY; 247 if (numfreebuffers >= hifreebuffers) 248 needsbuffer &= ~VFS_BIO_NEED_FREE; 249 wakeup(&needsbuffer); 250 } 251} 252 253/* 254 * waitrunningbufspace() 255 * 256 * runningbufspace is a measure of the amount of I/O currently 257 * running. This routine is used in async-write situations to 258 * prevent creating huge backups of pending writes to a device. 259 * Only asynchronous writes are governed by this function. 260 * 261 * Reads will adjust runningbufspace, but will not block based on it. 262 * The read load has a side effect of reducing the allowed write load. 263 * 264 * This does NOT turn an async write into a sync write. It waits 265 * for earlier writes to complete and generally returns before the 266 * caller's write has reached the device. 267 */ 268static __inline void 269waitrunningbufspace(void) 270{ 271 while (runningbufspace > hirunningspace) { 272 ++runningbufreq; 273 tsleep(&runningbufreq, PVM, "wdrain", 0); 274 } 275} 276 277 278/* 279 * vfs_buf_test_cache: 280 * 281 * Called when a buffer is extended. This function clears the B_CACHE 282 * bit if the newly extended portion of the buffer does not contain 283 * valid data. 284 * 285 * must be called with vm_mtx held 286 */ 287static __inline__ 288void 289vfs_buf_test_cache(struct buf *bp, 290 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 291 vm_page_t m) 292{ 293 if (bp->b_flags & B_CACHE) { 294 int base = (foff + off) & PAGE_MASK; 295 if (vm_page_is_valid(m, base, size) == 0) 296 bp->b_flags &= ~B_CACHE; 297 } 298} 299 300static __inline__ 301void 302bd_wakeup(int dirtybuflevel) 303{ 304 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 305 bd_request = 1; 306 wakeup(&bd_request); 307 } 308} 309 310/* 311 * bd_speedup - speedup the buffer cache flushing code 312 */ 313 314static __inline__ 315void 316bd_speedup(void) 317{ 318 bd_wakeup(1); 319} 320 321/* 322 * Initialize buffer headers and related structures. 323 */ 324 325caddr_t 326bufhashinit(caddr_t vaddr) 327{ 328 /* first, make a null hash table */ 329 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 330 ; 331 bufhashtbl = (void *)vaddr; 332 vaddr = vaddr + sizeof(*bufhashtbl) * bufhashmask; 333 --bufhashmask; 334 return(vaddr); 335} 336 337void 338bufinit(void) 339{ 340 struct buf *bp; 341 int i; 342 343 TAILQ_INIT(&bswlist); 344 LIST_INIT(&invalhash); 345 mtx_init(&buftimelock, "buftime lock", MTX_DEF); 346 347 for (i = 0; i <= bufhashmask; i++) 348 LIST_INIT(&bufhashtbl[i]); 349 350 /* next, make a null set of free lists */ 351 for (i = 0; i < BUFFER_QUEUES; i++) 352 TAILQ_INIT(&bufqueues[i]); 353 354 /* finally, initialize each buffer header and stick on empty q */ 355 for (i = 0; i < nbuf; i++) { 356 bp = &buf[i]; 357 bzero(bp, sizeof *bp); 358 bp->b_flags = B_INVAL; /* we're just an empty header */ 359 bp->b_dev = NODEV; 360 bp->b_rcred = NOCRED; 361 bp->b_wcred = NOCRED; 362 bp->b_qindex = QUEUE_EMPTY; 363 bp->b_xflags = 0; 364 LIST_INIT(&bp->b_dep); 365 BUF_LOCKINIT(bp); 366 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 367 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 368 } 369 370 /* 371 * maxbufspace is the absolute maximum amount of buffer space we are 372 * allowed to reserve in KVM and in real terms. The absolute maximum 373 * is nominally used by buf_daemon. hibufspace is the nominal maximum 374 * used by most other processes. The differential is required to 375 * ensure that buf_daemon is able to run when other processes might 376 * be blocked waiting for buffer space. 377 * 378 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 379 * this may result in KVM fragmentation which is not handled optimally 380 * by the system. 381 */ 382 maxbufspace = nbuf * BKVASIZE; 383 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 384 lobufspace = hibufspace - MAXBSIZE; 385 386 lorunningspace = 512 * 1024; 387 hirunningspace = 1024 * 1024; 388 389/* 390 * Limit the amount of malloc memory since it is wired permanently into 391 * the kernel space. Even though this is accounted for in the buffer 392 * allocation, we don't want the malloced region to grow uncontrolled. 393 * The malloc scheme improves memory utilization significantly on average 394 * (small) directories. 395 */ 396 maxbufmallocspace = hibufspace / 20; 397 398/* 399 * Reduce the chance of a deadlock occuring by limiting the number 400 * of delayed-write dirty buffers we allow to stack up. 401 */ 402 hidirtybuffers = nbuf / 4 + 20; 403 numdirtybuffers = 0; 404/* 405 * To support extreme low-memory systems, make sure hidirtybuffers cannot 406 * eat up all available buffer space. This occurs when our minimum cannot 407 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 408 * BKVASIZE'd (8K) buffers. 409 */ 410 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 411 hidirtybuffers >>= 1; 412 } 413 lodirtybuffers = hidirtybuffers / 2; 414 415/* 416 * Try to keep the number of free buffers in the specified range, 417 * and give special processes (e.g. like buf_daemon) access to an 418 * emergency reserve. 419 */ 420 lofreebuffers = nbuf / 18 + 5; 421 hifreebuffers = 2 * lofreebuffers; 422 numfreebuffers = nbuf; 423 424/* 425 * Maximum number of async ops initiated per buf_daemon loop. This is 426 * somewhat of a hack at the moment, we really need to limit ourselves 427 * based on the number of bytes of I/O in-transit that were initiated 428 * from buf_daemon. 429 */ 430 431 mtx_lock(&vm_mtx); 432 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 433 bogus_page = vm_page_alloc(kernel_object, 434 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 435 VM_ALLOC_NORMAL); 436 cnt.v_wire_count++; 437 mtx_unlock(&vm_mtx); 438 439} 440 441/* 442 * bfreekva() - free the kva allocation for a buffer. 443 * 444 * Must be called at splbio() or higher as this is the only locking for 445 * buffer_map. 446 * 447 * Since this call frees up buffer space, we call bufspacewakeup(). 448 * 449 * Must be called without the vm_mtx. 450 */ 451static void 452bfreekva(struct buf * bp) 453{ 454 455 mtx_assert(&vm_mtx, MA_NOTOWNED); 456 if (bp->b_kvasize) { 457 ++buffreekvacnt; 458 bufspace -= bp->b_kvasize; 459 mtx_lock(&vm_mtx); 460 vm_map_delete(buffer_map, 461 (vm_offset_t) bp->b_kvabase, 462 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 463 ); 464 mtx_unlock(&vm_mtx); 465 bp->b_kvasize = 0; 466 bufspacewakeup(); 467 } 468} 469 470/* 471 * bremfree: 472 * 473 * Remove the buffer from the appropriate free list. 474 */ 475void 476bremfree(struct buf * bp) 477{ 478 int s = splbio(); 479 int old_qindex = bp->b_qindex; 480 481 if (bp->b_qindex != QUEUE_NONE) { 482 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 483 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 484 bp->b_qindex = QUEUE_NONE; 485 } else { 486 if (BUF_REFCNT(bp) <= 1) 487 panic("bremfree: removing a buffer not on a queue"); 488 } 489 490 /* 491 * Fixup numfreebuffers count. If the buffer is invalid or not 492 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 493 * the buffer was free and we must decrement numfreebuffers. 494 */ 495 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 496 switch(old_qindex) { 497 case QUEUE_DIRTY: 498 case QUEUE_CLEAN: 499 case QUEUE_EMPTY: 500 case QUEUE_EMPTYKVA: 501 --numfreebuffers; 502 break; 503 default: 504 break; 505 } 506 } 507 splx(s); 508} 509 510 511/* 512 * Get a buffer with the specified data. Look in the cache first. We 513 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 514 * is set, the buffer is valid and we do not have to do anything ( see 515 * getblk() ). This is really just a special case of breadn(). 516 */ 517int 518bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 519 struct buf ** bpp) 520{ 521 522 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 523} 524 525/* 526 * Operates like bread, but also starts asynchronous I/O on 527 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 528 * to initiating I/O . If B_CACHE is set, the buffer is valid 529 * and we do not have to do anything. 530 */ 531int 532breadn(struct vnode * vp, daddr_t blkno, int size, 533 daddr_t * rablkno, int *rabsize, 534 int cnt, struct ucred * cred, struct buf ** bpp) 535{ 536 struct buf *bp, *rabp; 537 int i; 538 int rv = 0, readwait = 0; 539 540 *bpp = bp = getblk(vp, blkno, size, 0, 0); 541 542 /* if not found in cache, do some I/O */ 543 if ((bp->b_flags & B_CACHE) == 0) { 544 if (curproc != PCPU_GET(idleproc)) 545 curproc->p_stats->p_ru.ru_inblock++; 546 bp->b_iocmd = BIO_READ; 547 bp->b_flags &= ~B_INVAL; 548 bp->b_ioflags &= ~BIO_ERROR; 549 if (bp->b_rcred == NOCRED) { 550 if (cred != NOCRED) 551 crhold(cred); 552 bp->b_rcred = cred; 553 } 554 vfs_busy_pages(bp, 0); 555 VOP_STRATEGY(vp, bp); 556 ++readwait; 557 } 558 559 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 560 if (inmem(vp, *rablkno)) 561 continue; 562 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 563 564 if ((rabp->b_flags & B_CACHE) == 0) { 565 if (curproc != PCPU_GET(idleproc)) 566 curproc->p_stats->p_ru.ru_inblock++; 567 rabp->b_flags |= B_ASYNC; 568 rabp->b_flags &= ~B_INVAL; 569 rabp->b_ioflags &= ~BIO_ERROR; 570 rabp->b_iocmd = BIO_READ; 571 if (rabp->b_rcred == NOCRED) { 572 if (cred != NOCRED) 573 crhold(cred); 574 rabp->b_rcred = cred; 575 } 576 vfs_busy_pages(rabp, 0); 577 BUF_KERNPROC(rabp); 578 VOP_STRATEGY(vp, rabp); 579 } else { 580 brelse(rabp); 581 } 582 } 583 584 if (readwait) { 585 rv = bufwait(bp); 586 } 587 return (rv); 588} 589 590/* 591 * Write, release buffer on completion. (Done by iodone 592 * if async). Do not bother writing anything if the buffer 593 * is invalid. 594 * 595 * Note that we set B_CACHE here, indicating that buffer is 596 * fully valid and thus cacheable. This is true even of NFS 597 * now so we set it generally. This could be set either here 598 * or in biodone() since the I/O is synchronous. We put it 599 * here. 600 */ 601 602int dobkgrdwrite = 1; 603SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, ""); 604 605int 606bwrite(struct buf * bp) 607{ 608 int oldflags, s; 609 struct buf *newbp; 610 611 if (bp->b_flags & B_INVAL) { 612 brelse(bp); 613 return (0); 614 } 615 616 oldflags = bp->b_flags; 617 618 if (BUF_REFCNT(bp) == 0) 619 panic("bwrite: buffer is not busy???"); 620 s = splbio(); 621 /* 622 * If a background write is already in progress, delay 623 * writing this block if it is asynchronous. Otherwise 624 * wait for the background write to complete. 625 */ 626 if (bp->b_xflags & BX_BKGRDINPROG) { 627 if (bp->b_flags & B_ASYNC) { 628 splx(s); 629 bdwrite(bp); 630 return (0); 631 } 632 bp->b_xflags |= BX_BKGRDWAIT; 633 tsleep(&bp->b_xflags, PRIBIO, "biord", 0); 634 if (bp->b_xflags & BX_BKGRDINPROG) 635 panic("bwrite: still writing"); 636 } 637 638 /* Mark the buffer clean */ 639 bundirty(bp); 640 641 /* 642 * If this buffer is marked for background writing and we 643 * do not have to wait for it, make a copy and write the 644 * copy so as to leave this buffer ready for further use. 645 * 646 * This optimization eats a lot of memory. If we have a page 647 * or buffer shortfall we can't do it. 648 */ 649 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 650 (bp->b_flags & B_ASYNC) && 651 !vm_page_count_severe() && 652 !buf_dirty_count_severe()) { 653 if (bp->b_iodone != NULL) { 654 printf("bp->b_iodone = %p\n", bp->b_iodone); 655 panic("bwrite: need chained iodone"); 656 } 657 658 /* get a new block */ 659 newbp = geteblk(bp->b_bufsize); 660 661 /* set it to be identical to the old block */ 662 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 663 bgetvp(bp->b_vp, newbp); 664 newbp->b_lblkno = bp->b_lblkno; 665 newbp->b_blkno = bp->b_blkno; 666 newbp->b_offset = bp->b_offset; 667 newbp->b_iodone = vfs_backgroundwritedone; 668 newbp->b_flags |= B_ASYNC; 669 newbp->b_flags &= ~B_INVAL; 670 671 /* move over the dependencies */ 672 if (LIST_FIRST(&bp->b_dep) != NULL) 673 buf_movedeps(bp, newbp); 674 675 /* 676 * Initiate write on the copy, release the original to 677 * the B_LOCKED queue so that it cannot go away until 678 * the background write completes. If not locked it could go 679 * away and then be reconstituted while it was being written. 680 * If the reconstituted buffer were written, we could end up 681 * with two background copies being written at the same time. 682 */ 683 bp->b_xflags |= BX_BKGRDINPROG; 684 bp->b_flags |= B_LOCKED; 685 bqrelse(bp); 686 bp = newbp; 687 } 688 689 bp->b_flags &= ~B_DONE; 690 bp->b_ioflags &= ~BIO_ERROR; 691 bp->b_flags |= B_WRITEINPROG | B_CACHE; 692 bp->b_iocmd = BIO_WRITE; 693 694 bp->b_vp->v_numoutput++; 695 vfs_busy_pages(bp, 1); 696 697 /* 698 * Normal bwrites pipeline writes 699 */ 700 bp->b_runningbufspace = bp->b_bufsize; 701 runningbufspace += bp->b_runningbufspace; 702 703 if (curproc != PCPU_GET(idleproc)) 704 curproc->p_stats->p_ru.ru_oublock++; 705 splx(s); 706 if (oldflags & B_ASYNC) 707 BUF_KERNPROC(bp); 708 BUF_STRATEGY(bp); 709 710 if ((oldflags & B_ASYNC) == 0) { 711 int rtval = bufwait(bp); 712 brelse(bp); 713 return (rtval); 714 } else { 715 /* 716 * don't allow the async write to saturate the I/O 717 * system. There is no chance of deadlock here because 718 * we are blocking on I/O that is already in-progress. 719 */ 720 waitrunningbufspace(); 721 } 722 723 return (0); 724} 725 726/* 727 * Complete a background write started from bwrite. 728 */ 729static void 730vfs_backgroundwritedone(bp) 731 struct buf *bp; 732{ 733 struct buf *origbp; 734 735 /* 736 * Find the original buffer that we are writing. 737 */ 738 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 739 panic("backgroundwritedone: lost buffer"); 740 /* 741 * Process dependencies then return any unfinished ones. 742 */ 743 if (LIST_FIRST(&bp->b_dep) != NULL) 744 buf_complete(bp); 745 if (LIST_FIRST(&bp->b_dep) != NULL) 746 buf_movedeps(bp, origbp); 747 /* 748 * Clear the BX_BKGRDINPROG flag in the original buffer 749 * and awaken it if it is waiting for the write to complete. 750 * If BX_BKGRDINPROG is not set in the original buffer it must 751 * have been released and re-instantiated - which is not legal. 752 */ 753 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), ("backgroundwritedone: lost buffer2")); 754 origbp->b_xflags &= ~BX_BKGRDINPROG; 755 if (origbp->b_xflags & BX_BKGRDWAIT) { 756 origbp->b_xflags &= ~BX_BKGRDWAIT; 757 wakeup(&origbp->b_xflags); 758 } 759 /* 760 * Clear the B_LOCKED flag and remove it from the locked 761 * queue if it currently resides there. 762 */ 763 origbp->b_flags &= ~B_LOCKED; 764 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 765 bremfree(origbp); 766 bqrelse(origbp); 767 } 768 /* 769 * This buffer is marked B_NOCACHE, so when it is released 770 * by biodone, it will be tossed. We mark it with BIO_READ 771 * to avoid biodone doing a second vwakeup. 772 */ 773 bp->b_flags |= B_NOCACHE; 774 bp->b_iocmd = BIO_READ; 775 bp->b_flags &= ~(B_CACHE | B_DONE); 776 bp->b_iodone = 0; 777 bufdone(bp); 778} 779 780/* 781 * Delayed write. (Buffer is marked dirty). Do not bother writing 782 * anything if the buffer is marked invalid. 783 * 784 * Note that since the buffer must be completely valid, we can safely 785 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 786 * biodone() in order to prevent getblk from writing the buffer 787 * out synchronously. 788 */ 789void 790bdwrite(struct buf * bp) 791{ 792 if (BUF_REFCNT(bp) == 0) 793 panic("bdwrite: buffer is not busy"); 794 795 if (bp->b_flags & B_INVAL) { 796 brelse(bp); 797 return; 798 } 799 bdirty(bp); 800 801 /* 802 * Set B_CACHE, indicating that the buffer is fully valid. This is 803 * true even of NFS now. 804 */ 805 bp->b_flags |= B_CACHE; 806 807 /* 808 * This bmap keeps the system from needing to do the bmap later, 809 * perhaps when the system is attempting to do a sync. Since it 810 * is likely that the indirect block -- or whatever other datastructure 811 * that the filesystem needs is still in memory now, it is a good 812 * thing to do this. Note also, that if the pageout daemon is 813 * requesting a sync -- there might not be enough memory to do 814 * the bmap then... So, this is important to do. 815 */ 816 if (bp->b_lblkno == bp->b_blkno) { 817 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 818 } 819 820 mtx_lock(&vm_mtx); 821 /* 822 * Set the *dirty* buffer range based upon the VM system dirty pages. 823 */ 824 vfs_setdirty(bp); 825 826 /* 827 * We need to do this here to satisfy the vnode_pager and the 828 * pageout daemon, so that it thinks that the pages have been 829 * "cleaned". Note that since the pages are in a delayed write 830 * buffer -- the VFS layer "will" see that the pages get written 831 * out on the next sync, or perhaps the cluster will be completed. 832 */ 833 vfs_clean_pages(bp); 834 mtx_unlock(&vm_mtx); 835 bqrelse(bp); 836 837 /* 838 * Wakeup the buffer flushing daemon if we have a lot of dirty 839 * buffers (midpoint between our recovery point and our stall 840 * point). 841 */ 842 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 843 844 /* 845 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 846 * due to the softdep code. 847 */ 848} 849 850/* 851 * bdirty: 852 * 853 * Turn buffer into delayed write request. We must clear BIO_READ and 854 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 855 * itself to properly update it in the dirty/clean lists. We mark it 856 * B_DONE to ensure that any asynchronization of the buffer properly 857 * clears B_DONE ( else a panic will occur later ). 858 * 859 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 860 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 861 * should only be called if the buffer is known-good. 862 * 863 * Since the buffer is not on a queue, we do not update the numfreebuffers 864 * count. 865 * 866 * Must be called at splbio(). 867 * The buffer must be on QUEUE_NONE. 868 */ 869void 870bdirty(bp) 871 struct buf *bp; 872{ 873 KASSERT(bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 874 bp->b_flags &= ~(B_RELBUF); 875 bp->b_iocmd = BIO_WRITE; 876 877 if ((bp->b_flags & B_DELWRI) == 0) { 878 bp->b_flags |= B_DONE | B_DELWRI; 879 reassignbuf(bp, bp->b_vp); 880 ++numdirtybuffers; 881 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 882 } 883} 884 885/* 886 * bundirty: 887 * 888 * Clear B_DELWRI for buffer. 889 * 890 * Since the buffer is not on a queue, we do not update the numfreebuffers 891 * count. 892 * 893 * Must be called at splbio(). 894 * The buffer must be on QUEUE_NONE. 895 */ 896 897void 898bundirty(bp) 899 struct buf *bp; 900{ 901 KASSERT(bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 902 903 if (bp->b_flags & B_DELWRI) { 904 bp->b_flags &= ~B_DELWRI; 905 reassignbuf(bp, bp->b_vp); 906 --numdirtybuffers; 907 numdirtywakeup(lodirtybuffers); 908 } 909 /* 910 * Since it is now being written, we can clear its deferred write flag. 911 */ 912 bp->b_flags &= ~B_DEFERRED; 913} 914 915/* 916 * bawrite: 917 * 918 * Asynchronous write. Start output on a buffer, but do not wait for 919 * it to complete. The buffer is released when the output completes. 920 * 921 * bwrite() ( or the VOP routine anyway ) is responsible for handling 922 * B_INVAL buffers. Not us. 923 */ 924void 925bawrite(struct buf * bp) 926{ 927 bp->b_flags |= B_ASYNC; 928 (void) BUF_WRITE(bp); 929} 930 931/* 932 * bowrite: 933 * 934 * Ordered write. Start output on a buffer, and flag it so that the 935 * device will write it in the order it was queued. The buffer is 936 * released when the output completes. bwrite() ( or the VOP routine 937 * anyway ) is responsible for handling B_INVAL buffers. 938 */ 939int 940bowrite(struct buf * bp) 941{ 942 bp->b_ioflags |= BIO_ORDERED; 943 bp->b_flags |= B_ASYNC; 944 return (BUF_WRITE(bp)); 945} 946 947/* 948 * bwillwrite: 949 * 950 * Called prior to the locking of any vnodes when we are expecting to 951 * write. We do not want to starve the buffer cache with too many 952 * dirty buffers so we block here. By blocking prior to the locking 953 * of any vnodes we attempt to avoid the situation where a locked vnode 954 * prevents the various system daemons from flushing related buffers. 955 */ 956 957void 958bwillwrite(void) 959{ 960 if (numdirtybuffers >= hidirtybuffers) { 961 int s; 962 963 s = splbio(); 964 while (numdirtybuffers >= hidirtybuffers) { 965 bd_wakeup(1); 966 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 967 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0); 968 } 969 splx(s); 970 } 971} 972 973/* 974 * Return true if we have too many dirty buffers. 975 */ 976int 977buf_dirty_count_severe(void) 978{ 979 return(numdirtybuffers >= hidirtybuffers); 980} 981 982/* 983 * brelse: 984 * 985 * Release a busy buffer and, if requested, free its resources. The 986 * buffer will be stashed in the appropriate bufqueue[] allowing it 987 * to be accessed later as a cache entity or reused for other purposes. 988 * 989 * vm_mtx must be not be held. 990 */ 991void 992brelse(struct buf * bp) 993{ 994 int s; 995 996 mtx_assert(&vm_mtx, MA_NOTOWNED); 997 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 998 999 s = splbio(); 1000 1001 if (bp->b_flags & B_LOCKED) 1002 bp->b_ioflags &= ~BIO_ERROR; 1003 1004 if (bp->b_iocmd == BIO_WRITE && 1005 (bp->b_ioflags & BIO_ERROR) && 1006 !(bp->b_flags & B_INVAL)) { 1007 /* 1008 * Failed write, redirty. Must clear BIO_ERROR to prevent 1009 * pages from being scrapped. If B_INVAL is set then 1010 * this case is not run and the next case is run to 1011 * destroy the buffer. B_INVAL can occur if the buffer 1012 * is outside the range supported by the underlying device. 1013 */ 1014 bp->b_ioflags &= ~BIO_ERROR; 1015 bdirty(bp); 1016 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1017 (bp->b_ioflags & BIO_ERROR) || 1018 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1019 /* 1020 * Either a failed I/O or we were asked to free or not 1021 * cache the buffer. 1022 */ 1023 bp->b_flags |= B_INVAL; 1024 if (LIST_FIRST(&bp->b_dep) != NULL) 1025 buf_deallocate(bp); 1026 if (bp->b_flags & B_DELWRI) { 1027 --numdirtybuffers; 1028 numdirtywakeup(lodirtybuffers); 1029 } 1030 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1031 if ((bp->b_flags & B_VMIO) == 0) { 1032 if (bp->b_bufsize) 1033 allocbuf(bp, 0); 1034 if (bp->b_vp) 1035 brelvp(bp); 1036 } 1037 } 1038 1039 /* 1040 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1041 * is called with B_DELWRI set, the underlying pages may wind up 1042 * getting freed causing a previous write (bdwrite()) to get 'lost' 1043 * because pages associated with a B_DELWRI bp are marked clean. 1044 * 1045 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1046 * if B_DELWRI is set. 1047 * 1048 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1049 * on pages to return pages to the VM page queues. 1050 */ 1051 if (bp->b_flags & B_DELWRI) 1052 bp->b_flags &= ~B_RELBUF; 1053 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1054 bp->b_flags |= B_RELBUF; 1055 1056 /* 1057 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1058 * constituted, not even NFS buffers now. Two flags effect this. If 1059 * B_INVAL, the struct buf is invalidated but the VM object is kept 1060 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1061 * 1062 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1063 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1064 * buffer is also B_INVAL because it hits the re-dirtying code above. 1065 * 1066 * Normally we can do this whether a buffer is B_DELWRI or not. If 1067 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1068 * the commit state and we cannot afford to lose the buffer. If the 1069 * buffer has a background write in progress, we need to keep it 1070 * around to prevent it from being reconstituted and starting a second 1071 * background write. 1072 */ 1073 if ((bp->b_flags & B_VMIO) 1074 && !(bp->b_vp->v_tag == VT_NFS && 1075 !vn_isdisk(bp->b_vp, NULL) && 1076 (bp->b_flags & B_DELWRI)) 1077 ) { 1078 1079 int i, j, resid; 1080 vm_page_t m; 1081 off_t foff; 1082 vm_pindex_t poff; 1083 vm_object_t obj; 1084 struct vnode *vp; 1085 1086 vp = bp->b_vp; 1087 1088 /* 1089 * Get the base offset and length of the buffer. Note that 1090 * for block sizes that are less then PAGE_SIZE, the b_data 1091 * base of the buffer does not represent exactly b_offset and 1092 * neither b_offset nor b_size are necessarily page aligned. 1093 * Instead, the starting position of b_offset is: 1094 * 1095 * b_data + (b_offset & PAGE_MASK) 1096 * 1097 * block sizes less then DEV_BSIZE (usually 512) are not 1098 * supported due to the page granularity bits (m->valid, 1099 * m->dirty, etc...). 1100 * 1101 * See man buf(9) for more information 1102 */ 1103 resid = bp->b_bufsize; 1104 foff = bp->b_offset; 1105 1106 mtx_lock(&vm_mtx); 1107 for (i = 0; i < bp->b_npages; i++) { 1108 int had_bogus = 0; 1109 1110 m = bp->b_pages[i]; 1111 vm_page_flag_clear(m, PG_ZERO); 1112 1113 /* 1114 * If we hit a bogus page, fixup *all* the bogus pages 1115 * now. 1116 */ 1117 if (m == bogus_page) { 1118 mtx_unlock(&vm_mtx); 1119 VOP_GETVOBJECT(vp, &obj); 1120 poff = OFF_TO_IDX(bp->b_offset); 1121 had_bogus = 1; 1122 1123 mtx_lock(&vm_mtx); 1124 for (j = i; j < bp->b_npages; j++) { 1125 vm_page_t mtmp; 1126 mtmp = bp->b_pages[j]; 1127 if (mtmp == bogus_page) { 1128 mtmp = vm_page_lookup(obj, poff + j); 1129 if (!mtmp) { 1130 panic("brelse: page missing\n"); 1131 } 1132 bp->b_pages[j] = mtmp; 1133 } 1134 } 1135 1136 if ((bp->b_flags & B_INVAL) == 0) { 1137 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1138 } 1139 m = bp->b_pages[i]; 1140 } 1141 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1142 int poffset = foff & PAGE_MASK; 1143 int presid = resid > (PAGE_SIZE - poffset) ? 1144 (PAGE_SIZE - poffset) : resid; 1145 1146 KASSERT(presid >= 0, ("brelse: extra page")); 1147 vm_page_set_invalid(m, poffset, presid); 1148 if (had_bogus) 1149 printf("avoided corruption bug in bogus_page/brelse code\n"); 1150 } 1151 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1152 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1153 } 1154 1155 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1156 vfs_vmio_release(bp); 1157 mtx_unlock(&vm_mtx); 1158 1159 } else if (bp->b_flags & B_VMIO) { 1160 1161 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1162 mtx_lock(&vm_mtx); 1163 vfs_vmio_release(bp); 1164 mtx_unlock(&vm_mtx); 1165 } 1166 1167 } 1168 1169 if (bp->b_qindex != QUEUE_NONE) 1170 panic("brelse: free buffer onto another queue???"); 1171 if (BUF_REFCNT(bp) > 1) { 1172 /* do not release to free list */ 1173 BUF_UNLOCK(bp); 1174 splx(s); 1175 return; 1176 } 1177 1178 /* enqueue */ 1179 1180 /* buffers with no memory */ 1181 if (bp->b_bufsize == 0) { 1182 bp->b_flags |= B_INVAL; 1183 bp->b_xflags &= ~BX_BKGRDWRITE; 1184 if (bp->b_xflags & BX_BKGRDINPROG) 1185 panic("losing buffer 1"); 1186 if (bp->b_kvasize) { 1187 bp->b_qindex = QUEUE_EMPTYKVA; 1188 } else { 1189 bp->b_qindex = QUEUE_EMPTY; 1190 } 1191 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1192 LIST_REMOVE(bp, b_hash); 1193 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1194 bp->b_dev = NODEV; 1195 /* buffers with junk contents */ 1196 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || (bp->b_ioflags & BIO_ERROR)) { 1197 bp->b_flags |= B_INVAL; 1198 bp->b_xflags &= ~BX_BKGRDWRITE; 1199 if (bp->b_xflags & BX_BKGRDINPROG) 1200 panic("losing buffer 2"); 1201 bp->b_qindex = QUEUE_CLEAN; 1202 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1203 LIST_REMOVE(bp, b_hash); 1204 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1205 bp->b_dev = NODEV; 1206 1207 /* buffers that are locked */ 1208 } else if (bp->b_flags & B_LOCKED) { 1209 bp->b_qindex = QUEUE_LOCKED; 1210 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1211 1212 /* remaining buffers */ 1213 } else { 1214 if (bp->b_flags & B_DELWRI) 1215 bp->b_qindex = QUEUE_DIRTY; 1216 else 1217 bp->b_qindex = QUEUE_CLEAN; 1218 if (bp->b_flags & B_AGE) 1219 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1220 else 1221 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1222 } 1223 1224 /* 1225 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1226 * on the correct queue. 1227 */ 1228 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) { 1229 bp->b_flags &= ~B_DELWRI; 1230 --numdirtybuffers; 1231 numdirtywakeup(lodirtybuffers); 1232 } 1233 1234 /* 1235 * Fixup numfreebuffers count. The bp is on an appropriate queue 1236 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1237 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1238 * if B_INVAL is set ). 1239 */ 1240 1241 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1242 bufcountwakeup(); 1243 1244 /* 1245 * Something we can maybe free or reuse 1246 */ 1247 if (bp->b_bufsize || bp->b_kvasize) 1248 bufspacewakeup(); 1249 1250 /* unlock */ 1251 BUF_UNLOCK(bp); 1252 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1253 bp->b_ioflags &= ~BIO_ORDERED; 1254 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1255 panic("brelse: not dirty"); 1256 splx(s); 1257} 1258 1259/* 1260 * Release a buffer back to the appropriate queue but do not try to free 1261 * it. The buffer is expected to be used again soon. 1262 * 1263 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1264 * biodone() to requeue an async I/O on completion. It is also used when 1265 * known good buffers need to be requeued but we think we may need the data 1266 * again soon. 1267 */ 1268void 1269bqrelse(struct buf * bp) 1270{ 1271 int s; 1272 1273 s = splbio(); 1274 1275 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1276 1277 if (bp->b_qindex != QUEUE_NONE) 1278 panic("bqrelse: free buffer onto another queue???"); 1279 if (BUF_REFCNT(bp) > 1) { 1280 /* do not release to free list */ 1281 BUF_UNLOCK(bp); 1282 splx(s); 1283 return; 1284 } 1285 if (bp->b_flags & B_LOCKED) { 1286 bp->b_ioflags &= ~BIO_ERROR; 1287 bp->b_qindex = QUEUE_LOCKED; 1288 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1289 /* buffers with stale but valid contents */ 1290 } else if (bp->b_flags & B_DELWRI) { 1291 bp->b_qindex = QUEUE_DIRTY; 1292 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1293 } else if (vm_page_count_severe()) { 1294 /* 1295 * We are too low on memory, we have to try to free the 1296 * buffer (most importantly: the wired pages making up its 1297 * backing store) *now*. 1298 */ 1299 splx(s); 1300 brelse(bp); 1301 return; 1302 } else { 1303 bp->b_qindex = QUEUE_CLEAN; 1304 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1305 } 1306 1307 if ((bp->b_flags & B_LOCKED) == 0 && 1308 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1309 bufcountwakeup(); 1310 } 1311 1312 /* 1313 * Something we can maybe free or reuse. 1314 */ 1315 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1316 bufspacewakeup(); 1317 1318 /* unlock */ 1319 BUF_UNLOCK(bp); 1320 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1321 bp->b_ioflags &= ~BIO_ORDERED; 1322 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1323 panic("bqrelse: not dirty"); 1324 splx(s); 1325} 1326 1327/* 1328 * Must be called with vm_mtx held. 1329 */ 1330static void 1331vfs_vmio_release(bp) 1332 struct buf *bp; 1333{ 1334 int i; 1335 vm_page_t m; 1336 1337 mtx_assert(&vm_mtx, MA_OWNED); 1338 for (i = 0; i < bp->b_npages; i++) { 1339 m = bp->b_pages[i]; 1340 bp->b_pages[i] = NULL; 1341 /* 1342 * In order to keep page LRU ordering consistent, put 1343 * everything on the inactive queue. 1344 */ 1345 vm_page_unwire(m, 0); 1346 /* 1347 * We don't mess with busy pages, it is 1348 * the responsibility of the process that 1349 * busied the pages to deal with them. 1350 */ 1351 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1352 continue; 1353 1354 if (m->wire_count == 0) { 1355 vm_page_flag_clear(m, PG_ZERO); 1356 /* 1357 * Might as well free the page if we can and it has 1358 * no valid data. 1359 */ 1360 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { 1361 vm_page_busy(m); 1362 vm_page_protect(m, VM_PROT_NONE); 1363 vm_page_free(m); 1364 } else if (vm_page_count_severe()) { 1365 vm_page_try_to_cache(m); 1366 } 1367 } 1368 } 1369 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1370 1371 /* could drop vm_mtx here */ 1372 1373 if (bp->b_bufsize) { 1374 bufspacewakeup(); 1375 bp->b_bufsize = 0; 1376 } 1377 bp->b_npages = 0; 1378 bp->b_flags &= ~B_VMIO; 1379 if (bp->b_vp) 1380 brelvp(bp); 1381} 1382 1383/* 1384 * Check to see if a block is currently memory resident. 1385 */ 1386struct buf * 1387gbincore(struct vnode * vp, daddr_t blkno) 1388{ 1389 struct buf *bp; 1390 struct bufhashhdr *bh; 1391 1392 bh = bufhash(vp, blkno); 1393 1394 /* Search hash chain */ 1395 LIST_FOREACH(bp, bh, b_hash) { 1396 /* hit */ 1397 if (bp->b_vp == vp && bp->b_lblkno == blkno && 1398 (bp->b_flags & B_INVAL) == 0) { 1399 break; 1400 } 1401 } 1402 return (bp); 1403} 1404 1405/* 1406 * vfs_bio_awrite: 1407 * 1408 * Implement clustered async writes for clearing out B_DELWRI buffers. 1409 * This is much better then the old way of writing only one buffer at 1410 * a time. Note that we may not be presented with the buffers in the 1411 * correct order, so we search for the cluster in both directions. 1412 */ 1413int 1414vfs_bio_awrite(struct buf * bp) 1415{ 1416 int i; 1417 int j; 1418 daddr_t lblkno = bp->b_lblkno; 1419 struct vnode *vp = bp->b_vp; 1420 int s; 1421 int ncl; 1422 struct buf *bpa; 1423 int nwritten; 1424 int size; 1425 int maxcl; 1426 1427 s = splbio(); 1428 /* 1429 * right now we support clustered writing only to regular files. If 1430 * we find a clusterable block we could be in the middle of a cluster 1431 * rather then at the beginning. 1432 */ 1433 if ((vp->v_type == VREG) && 1434 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1435 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1436 1437 size = vp->v_mount->mnt_stat.f_iosize; 1438 maxcl = MAXPHYS / size; 1439 1440 for (i = 1; i < maxcl; i++) { 1441 if ((bpa = gbincore(vp, lblkno + i)) && 1442 BUF_REFCNT(bpa) == 0 && 1443 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1444 (B_DELWRI | B_CLUSTEROK)) && 1445 (bpa->b_bufsize == size)) { 1446 if ((bpa->b_blkno == bpa->b_lblkno) || 1447 (bpa->b_blkno != 1448 bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 1449 break; 1450 } else { 1451 break; 1452 } 1453 } 1454 for (j = 1; i + j <= maxcl && j <= lblkno; j++) { 1455 if ((bpa = gbincore(vp, lblkno - j)) && 1456 BUF_REFCNT(bpa) == 0 && 1457 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1458 (B_DELWRI | B_CLUSTEROK)) && 1459 (bpa->b_bufsize == size)) { 1460 if ((bpa->b_blkno == bpa->b_lblkno) || 1461 (bpa->b_blkno != 1462 bp->b_blkno - ((j * size) >> DEV_BSHIFT))) 1463 break; 1464 } else { 1465 break; 1466 } 1467 } 1468 --j; 1469 ncl = i + j; 1470 /* 1471 * this is a possible cluster write 1472 */ 1473 if (ncl != 1) { 1474 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1475 splx(s); 1476 return nwritten; 1477 } 1478 } 1479 1480 BUF_LOCK(bp, LK_EXCLUSIVE); 1481 bremfree(bp); 1482 bp->b_flags |= B_ASYNC; 1483 1484 splx(s); 1485 /* 1486 * default (old) behavior, writing out only one block 1487 * 1488 * XXX returns b_bufsize instead of b_bcount for nwritten? 1489 */ 1490 nwritten = bp->b_bufsize; 1491 (void) BUF_WRITE(bp); 1492 1493 return nwritten; 1494} 1495 1496/* 1497 * getnewbuf: 1498 * 1499 * Find and initialize a new buffer header, freeing up existing buffers 1500 * in the bufqueues as necessary. The new buffer is returned locked. 1501 * 1502 * Important: B_INVAL is not set. If the caller wishes to throw the 1503 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1504 * 1505 * We block if: 1506 * We have insufficient buffer headers 1507 * We have insufficient buffer space 1508 * buffer_map is too fragmented ( space reservation fails ) 1509 * If we have to flush dirty buffers ( but we try to avoid this ) 1510 * 1511 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1512 * Instead we ask the buf daemon to do it for us. We attempt to 1513 * avoid piecemeal wakeups of the pageout daemon. 1514 */ 1515 1516static struct buf * 1517getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1518{ 1519 struct buf *bp; 1520 struct buf *nbp; 1521 int defrag = 0; 1522 int nqindex; 1523 static int flushingbufs; 1524 1525 /* 1526 * We can't afford to block since we might be holding a vnode lock, 1527 * which may prevent system daemons from running. We deal with 1528 * low-memory situations by proactively returning memory and running 1529 * async I/O rather then sync I/O. 1530 */ 1531 1532 ++getnewbufcalls; 1533 --getnewbufrestarts; 1534restart: 1535 ++getnewbufrestarts; 1536 1537 /* 1538 * Setup for scan. If we do not have enough free buffers, 1539 * we setup a degenerate case that immediately fails. Note 1540 * that if we are specially marked process, we are allowed to 1541 * dip into our reserves. 1542 * 1543 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1544 * 1545 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1546 * However, there are a number of cases (defragging, reusing, ...) 1547 * where we cannot backup. 1548 */ 1549 nqindex = QUEUE_EMPTYKVA; 1550 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1551 1552 if (nbp == NULL) { 1553 /* 1554 * If no EMPTYKVA buffers and we are either 1555 * defragging or reusing, locate a CLEAN buffer 1556 * to free or reuse. If bufspace useage is low 1557 * skip this step so we can allocate a new buffer. 1558 */ 1559 if (defrag || bufspace >= lobufspace) { 1560 nqindex = QUEUE_CLEAN; 1561 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1562 } 1563 1564 /* 1565 * If we could not find or were not allowed to reuse a 1566 * CLEAN buffer, check to see if it is ok to use an EMPTY 1567 * buffer. We can only use an EMPTY buffer if allocating 1568 * its KVA would not otherwise run us out of buffer space. 1569 */ 1570 if (nbp == NULL && defrag == 0 && 1571 bufspace + maxsize < hibufspace) { 1572 nqindex = QUEUE_EMPTY; 1573 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1574 } 1575 } 1576 1577 /* 1578 * Run scan, possibly freeing data and/or kva mappings on the fly 1579 * depending. 1580 */ 1581 1582 while ((bp = nbp) != NULL) { 1583 int qindex = nqindex; 1584 1585 /* 1586 * Calculate next bp ( we can only use it if we do not block 1587 * or do other fancy things ). 1588 */ 1589 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1590 switch(qindex) { 1591 case QUEUE_EMPTY: 1592 nqindex = QUEUE_EMPTYKVA; 1593 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1594 break; 1595 /* fall through */ 1596 case QUEUE_EMPTYKVA: 1597 nqindex = QUEUE_CLEAN; 1598 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1599 break; 1600 /* fall through */ 1601 case QUEUE_CLEAN: 1602 /* 1603 * nbp is NULL. 1604 */ 1605 break; 1606 } 1607 } 1608 1609 /* 1610 * Sanity Checks 1611 */ 1612 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1613 1614 /* 1615 * Note: we no longer distinguish between VMIO and non-VMIO 1616 * buffers. 1617 */ 1618 1619 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1620 1621 /* 1622 * If we are defragging then we need a buffer with 1623 * b_kvasize != 0. XXX this situation should no longer 1624 * occur, if defrag is non-zero the buffer's b_kvasize 1625 * should also be non-zero at this point. XXX 1626 */ 1627 if (defrag && bp->b_kvasize == 0) { 1628 printf("Warning: defrag empty buffer %p\n", bp); 1629 continue; 1630 } 1631 1632 /* 1633 * Start freeing the bp. This is somewhat involved. nbp 1634 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1635 */ 1636 1637 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1638 panic("getnewbuf: locked buf"); 1639 bremfree(bp); 1640 1641 if (qindex == QUEUE_CLEAN) { 1642 if (bp->b_flags & B_VMIO) { 1643 bp->b_flags &= ~B_ASYNC; 1644 mtx_lock(&vm_mtx); 1645 vfs_vmio_release(bp); 1646 mtx_unlock(&vm_mtx); 1647 } 1648 if (bp->b_vp) 1649 brelvp(bp); 1650 } 1651 1652 /* 1653 * NOTE: nbp is now entirely invalid. We can only restart 1654 * the scan from this point on. 1655 * 1656 * Get the rest of the buffer freed up. b_kva* is still 1657 * valid after this operation. 1658 */ 1659 1660 if (bp->b_rcred != NOCRED) { 1661 crfree(bp->b_rcred); 1662 bp->b_rcred = NOCRED; 1663 } 1664 if (bp->b_wcred != NOCRED) { 1665 crfree(bp->b_wcred); 1666 bp->b_wcred = NOCRED; 1667 } 1668 if (LIST_FIRST(&bp->b_dep) != NULL) 1669 buf_deallocate(bp); 1670 if (bp->b_xflags & BX_BKGRDINPROG) 1671 panic("losing buffer 3"); 1672 LIST_REMOVE(bp, b_hash); 1673 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1674 1675 if (bp->b_bufsize) 1676 allocbuf(bp, 0); 1677 1678 bp->b_flags = 0; 1679 bp->b_ioflags = 0; 1680 bp->b_xflags = 0; 1681 bp->b_dev = NODEV; 1682 bp->b_vp = NULL; 1683 bp->b_blkno = bp->b_lblkno = 0; 1684 bp->b_offset = NOOFFSET; 1685 bp->b_iodone = 0; 1686 bp->b_error = 0; 1687 bp->b_resid = 0; 1688 bp->b_bcount = 0; 1689 bp->b_npages = 0; 1690 bp->b_dirtyoff = bp->b_dirtyend = 0; 1691 bp->b_magic = B_MAGIC_BIO; 1692 bp->b_op = &buf_ops_bio; 1693 1694 LIST_INIT(&bp->b_dep); 1695 1696 /* 1697 * If we are defragging then free the buffer. 1698 */ 1699 if (defrag) { 1700 bp->b_flags |= B_INVAL; 1701 bfreekva(bp); 1702 brelse(bp); 1703 defrag = 0; 1704 goto restart; 1705 } 1706 1707 /* 1708 * If we are overcomitted then recover the buffer and its 1709 * KVM space. This occurs in rare situations when multiple 1710 * processes are blocked in getnewbuf() or allocbuf(). 1711 */ 1712 if (bufspace >= hibufspace) 1713 flushingbufs = 1; 1714 if (flushingbufs && bp->b_kvasize != 0) { 1715 bp->b_flags |= B_INVAL; 1716 bfreekva(bp); 1717 brelse(bp); 1718 goto restart; 1719 } 1720 if (bufspace < lobufspace) 1721 flushingbufs = 0; 1722 break; 1723 } 1724 1725 /* 1726 * If we exhausted our list, sleep as appropriate. We may have to 1727 * wakeup various daemons and write out some dirty buffers. 1728 * 1729 * Generally we are sleeping due to insufficient buffer space. 1730 */ 1731 1732 if (bp == NULL) { 1733 int flags; 1734 char *waitmsg; 1735 1736 if (defrag) { 1737 flags = VFS_BIO_NEED_BUFSPACE; 1738 waitmsg = "nbufkv"; 1739 } else if (bufspace >= hibufspace) { 1740 waitmsg = "nbufbs"; 1741 flags = VFS_BIO_NEED_BUFSPACE; 1742 } else { 1743 waitmsg = "newbuf"; 1744 flags = VFS_BIO_NEED_ANY; 1745 } 1746 1747 bd_speedup(); /* heeeelp */ 1748 1749 needsbuffer |= flags; 1750 while (needsbuffer & flags) { 1751 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, 1752 waitmsg, slptimeo)) 1753 return (NULL); 1754 } 1755 } else { 1756 /* 1757 * We finally have a valid bp. We aren't quite out of the 1758 * woods, we still have to reserve kva space. In order 1759 * to keep fragmentation sane we only allocate kva in 1760 * BKVASIZE chunks. 1761 */ 1762 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1763 1764 if (maxsize != bp->b_kvasize) { 1765 vm_offset_t addr = 0; 1766 1767 bfreekva(bp); 1768 1769 mtx_lock(&vm_mtx); 1770 if (vm_map_findspace(buffer_map, 1771 vm_map_min(buffer_map), maxsize, &addr)) { 1772 /* 1773 * Uh oh. Buffer map is to fragmented. We 1774 * must defragment the map. 1775 */ 1776 mtx_unlock(&vm_mtx); 1777 ++bufdefragcnt; 1778 defrag = 1; 1779 bp->b_flags |= B_INVAL; 1780 brelse(bp); 1781 goto restart; 1782 } 1783 if (addr) { 1784 vm_map_insert(buffer_map, NULL, 0, 1785 addr, addr + maxsize, 1786 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1787 1788 bp->b_kvabase = (caddr_t) addr; 1789 bp->b_kvasize = maxsize; 1790 bufspace += bp->b_kvasize; 1791 ++bufreusecnt; 1792 } 1793 mtx_unlock(&vm_mtx); 1794 } 1795 bp->b_data = bp->b_kvabase; 1796 } 1797 return(bp); 1798} 1799 1800/* 1801 * buf_daemon: 1802 * 1803 * buffer flushing daemon. Buffers are normally flushed by the 1804 * update daemon but if it cannot keep up this process starts to 1805 * take the load in an attempt to prevent getnewbuf() from blocking. 1806 */ 1807 1808static struct proc *bufdaemonproc; 1809 1810static struct kproc_desc buf_kp = { 1811 "bufdaemon", 1812 buf_daemon, 1813 &bufdaemonproc 1814}; 1815SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1816 1817static void 1818buf_daemon() 1819{ 1820 int s; 1821 1822 mtx_lock(&Giant); 1823 1824 /* 1825 * This process needs to be suspended prior to shutdown sync. 1826 */ 1827 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 1828 SHUTDOWN_PRI_LAST); 1829 1830 /* 1831 * This process is allowed to take the buffer cache to the limit 1832 */ 1833 curproc->p_flag |= P_BUFEXHAUST; 1834 s = splbio(); 1835 1836 for (;;) { 1837 kthread_suspend_check(bufdaemonproc); 1838 1839 bd_request = 0; 1840 1841 /* 1842 * Do the flush. Limit the amount of in-transit I/O we 1843 * allow to build up, otherwise we would completely saturate 1844 * the I/O system. Wakeup any waiting processes before we 1845 * normally would so they can run in parallel with our drain. 1846 */ 1847 while (numdirtybuffers > lodirtybuffers) { 1848 if (flushbufqueues() == 0) 1849 break; 1850 waitrunningbufspace(); 1851 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 1852 } 1853 1854 /* 1855 * Only clear bd_request if we have reached our low water 1856 * mark. The buf_daemon normally waits 5 seconds and 1857 * then incrementally flushes any dirty buffers that have 1858 * built up, within reason. 1859 * 1860 * If we were unable to hit our low water mark and couldn't 1861 * find any flushable buffers, we sleep half a second. 1862 * Otherwise we loop immediately. 1863 */ 1864 if (numdirtybuffers <= lodirtybuffers) { 1865 /* 1866 * We reached our low water mark, reset the 1867 * request and sleep until we are needed again. 1868 * The sleep is just so the suspend code works. 1869 */ 1870 bd_request = 0; 1871 tsleep(&bd_request, PVM, "psleep", hz); 1872 } else { 1873 /* 1874 * We couldn't find any flushable dirty buffers but 1875 * still have too many dirty buffers, we 1876 * have to sleep and try again. (rare) 1877 */ 1878 tsleep(&bd_request, PVM, "qsleep", hz / 2); 1879 } 1880 } 1881} 1882 1883/* 1884 * flushbufqueues: 1885 * 1886 * Try to flush a buffer in the dirty queue. We must be careful to 1887 * free up B_INVAL buffers instead of write them, which NFS is 1888 * particularly sensitive to. 1889 */ 1890 1891static int 1892flushbufqueues(void) 1893{ 1894 struct buf *bp; 1895 int r = 0; 1896 1897 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1898 1899 while (bp) { 1900 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); 1901 if ((bp->b_flags & B_DELWRI) != 0 && 1902 (bp->b_xflags & BX_BKGRDINPROG) == 0) { 1903 if (bp->b_flags & B_INVAL) { 1904 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1905 panic("flushbufqueues: locked buf"); 1906 bremfree(bp); 1907 brelse(bp); 1908 ++r; 1909 break; 1910 } 1911 if (LIST_FIRST(&bp->b_dep) != NULL && 1912 (bp->b_flags & B_DEFERRED) == 0 && 1913 buf_countdeps(bp, 0)) { 1914 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], 1915 bp, b_freelist); 1916 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], 1917 bp, b_freelist); 1918 bp->b_flags |= B_DEFERRED; 1919 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1920 continue; 1921 } 1922 vfs_bio_awrite(bp); 1923 ++r; 1924 break; 1925 } 1926 bp = TAILQ_NEXT(bp, b_freelist); 1927 } 1928 return (r); 1929} 1930 1931/* 1932 * Check to see if a block is currently memory resident. 1933 */ 1934struct buf * 1935incore(struct vnode * vp, daddr_t blkno) 1936{ 1937 struct buf *bp; 1938 1939 int s = splbio(); 1940 bp = gbincore(vp, blkno); 1941 splx(s); 1942 return (bp); 1943} 1944 1945/* 1946 * Returns true if no I/O is needed to access the 1947 * associated VM object. This is like incore except 1948 * it also hunts around in the VM system for the data. 1949 */ 1950 1951int 1952inmem(struct vnode * vp, daddr_t blkno) 1953{ 1954 vm_object_t obj; 1955 vm_offset_t toff, tinc, size; 1956 vm_page_t m; 1957 vm_ooffset_t off; 1958 1959 if (incore(vp, blkno)) 1960 return 1; 1961 if (vp->v_mount == NULL) 1962 return 0; 1963 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_flag & VOBJBUF) == 0) 1964 return 0; 1965 1966 size = PAGE_SIZE; 1967 if (size > vp->v_mount->mnt_stat.f_iosize) 1968 size = vp->v_mount->mnt_stat.f_iosize; 1969 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 1970 1971 mtx_lock(&vm_mtx); 1972 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1973 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 1974 if (!m) 1975 goto notinmem; 1976 tinc = size; 1977 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 1978 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 1979 if (vm_page_is_valid(m, 1980 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 1981 goto notinmem; 1982 } 1983 mtx_unlock(&vm_mtx); 1984 return 1; 1985 1986notinmem: 1987 mtx_unlock(&vm_mtx); 1988 return (0); 1989} 1990 1991/* 1992 * vfs_setdirty: 1993 * 1994 * Sets the dirty range for a buffer based on the status of the dirty 1995 * bits in the pages comprising the buffer. 1996 * 1997 * The range is limited to the size of the buffer. 1998 * 1999 * This routine is primarily used by NFS, but is generalized for the 2000 * B_VMIO case. 2001 * 2002 * Must be called with vm_mtx 2003 */ 2004static void 2005vfs_setdirty(struct buf *bp) 2006{ 2007 int i; 2008 vm_object_t object; 2009 2010 mtx_assert(&vm_mtx, MA_OWNED); 2011 /* 2012 * Degenerate case - empty buffer 2013 */ 2014 2015 if (bp->b_bufsize == 0) 2016 return; 2017 2018 /* 2019 * We qualify the scan for modified pages on whether the 2020 * object has been flushed yet. The OBJ_WRITEABLE flag 2021 * is not cleared simply by protecting pages off. 2022 */ 2023 2024 if ((bp->b_flags & B_VMIO) == 0) 2025 return; 2026 2027 object = bp->b_pages[0]->object; 2028 2029 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2030 printf("Warning: object %p writeable but not mightbedirty\n", object); 2031 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2032 printf("Warning: object %p mightbedirty but not writeable\n", object); 2033 2034 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2035 vm_offset_t boffset; 2036 vm_offset_t eoffset; 2037 2038 /* 2039 * test the pages to see if they have been modified directly 2040 * by users through the VM system. 2041 */ 2042 for (i = 0; i < bp->b_npages; i++) { 2043 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2044 vm_page_test_dirty(bp->b_pages[i]); 2045 } 2046 2047 /* 2048 * Calculate the encompassing dirty range, boffset and eoffset, 2049 * (eoffset - boffset) bytes. 2050 */ 2051 2052 for (i = 0; i < bp->b_npages; i++) { 2053 if (bp->b_pages[i]->dirty) 2054 break; 2055 } 2056 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2057 2058 for (i = bp->b_npages - 1; i >= 0; --i) { 2059 if (bp->b_pages[i]->dirty) { 2060 break; 2061 } 2062 } 2063 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2064 2065 /* 2066 * Fit it to the buffer. 2067 */ 2068 2069 if (eoffset > bp->b_bcount) 2070 eoffset = bp->b_bcount; 2071 2072 /* 2073 * If we have a good dirty range, merge with the existing 2074 * dirty range. 2075 */ 2076 2077 if (boffset < eoffset) { 2078 if (bp->b_dirtyoff > boffset) 2079 bp->b_dirtyoff = boffset; 2080 if (bp->b_dirtyend < eoffset) 2081 bp->b_dirtyend = eoffset; 2082 } 2083 } 2084} 2085 2086/* 2087 * getblk: 2088 * 2089 * Get a block given a specified block and offset into a file/device. 2090 * The buffers B_DONE bit will be cleared on return, making it almost 2091 * ready for an I/O initiation. B_INVAL may or may not be set on 2092 * return. The caller should clear B_INVAL prior to initiating a 2093 * READ. 2094 * 2095 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2096 * an existing buffer. 2097 * 2098 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2099 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2100 * and then cleared based on the backing VM. If the previous buffer is 2101 * non-0-sized but invalid, B_CACHE will be cleared. 2102 * 2103 * If getblk() must create a new buffer, the new buffer is returned with 2104 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2105 * case it is returned with B_INVAL clear and B_CACHE set based on the 2106 * backing VM. 2107 * 2108 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2109 * B_CACHE bit is clear. 2110 * 2111 * What this means, basically, is that the caller should use B_CACHE to 2112 * determine whether the buffer is fully valid or not and should clear 2113 * B_INVAL prior to issuing a read. If the caller intends to validate 2114 * the buffer by loading its data area with something, the caller needs 2115 * to clear B_INVAL. If the caller does this without issuing an I/O, 2116 * the caller should set B_CACHE ( as an optimization ), else the caller 2117 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2118 * a write attempt or if it was a successfull read. If the caller 2119 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2120 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2121 */ 2122struct buf * 2123getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2124{ 2125 struct buf *bp; 2126 int s; 2127 struct bufhashhdr *bh; 2128 2129 if (size > MAXBSIZE) 2130 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2131 2132 s = splbio(); 2133loop: 2134 /* 2135 * Block if we are low on buffers. Certain processes are allowed 2136 * to completely exhaust the buffer cache. 2137 * 2138 * If this check ever becomes a bottleneck it may be better to 2139 * move it into the else, when gbincore() fails. At the moment 2140 * it isn't a problem. 2141 * 2142 * XXX remove if 0 sections (clean this up after its proven) 2143 */ 2144 if (numfreebuffers == 0) { 2145 if (curproc == PCPU_GET(idleproc)) 2146 return NULL; 2147 needsbuffer |= VFS_BIO_NEED_ANY; 2148 } 2149 2150 if ((bp = gbincore(vp, blkno))) { 2151 /* 2152 * Buffer is in-core. If the buffer is not busy, it must 2153 * be on a queue. 2154 */ 2155 2156 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2157 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, 2158 "getblk", slpflag, slptimeo) == ENOLCK) 2159 goto loop; 2160 splx(s); 2161 return (struct buf *) NULL; 2162 } 2163 2164 /* 2165 * The buffer is locked. B_CACHE is cleared if the buffer is 2166 * invalid. Ohterwise, for a non-VMIO buffer, B_CACHE is set 2167 * and for a VMIO buffer B_CACHE is adjusted according to the 2168 * backing VM cache. 2169 */ 2170 if (bp->b_flags & B_INVAL) 2171 bp->b_flags &= ~B_CACHE; 2172 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2173 bp->b_flags |= B_CACHE; 2174 bremfree(bp); 2175 2176 /* 2177 * check for size inconsistancies for non-VMIO case. 2178 */ 2179 2180 if (bp->b_bcount != size) { 2181 if ((bp->b_flags & B_VMIO) == 0 || 2182 (size > bp->b_kvasize)) { 2183 if (bp->b_flags & B_DELWRI) { 2184 bp->b_flags |= B_NOCACHE; 2185 BUF_WRITE(bp); 2186 } else { 2187 if ((bp->b_flags & B_VMIO) && 2188 (LIST_FIRST(&bp->b_dep) == NULL)) { 2189 bp->b_flags |= B_RELBUF; 2190 brelse(bp); 2191 } else { 2192 bp->b_flags |= B_NOCACHE; 2193 BUF_WRITE(bp); 2194 } 2195 } 2196 goto loop; 2197 } 2198 } 2199 2200 /* 2201 * If the size is inconsistant in the VMIO case, we can resize 2202 * the buffer. This might lead to B_CACHE getting set or 2203 * cleared. If the size has not changed, B_CACHE remains 2204 * unchanged from its previous state. 2205 */ 2206 2207 if (bp->b_bcount != size) 2208 allocbuf(bp, size); 2209 2210 KASSERT(bp->b_offset != NOOFFSET, 2211 ("getblk: no buffer offset")); 2212 2213 /* 2214 * A buffer with B_DELWRI set and B_CACHE clear must 2215 * be committed before we can return the buffer in 2216 * order to prevent the caller from issuing a read 2217 * ( due to B_CACHE not being set ) and overwriting 2218 * it. 2219 * 2220 * Most callers, including NFS and FFS, need this to 2221 * operate properly either because they assume they 2222 * can issue a read if B_CACHE is not set, or because 2223 * ( for example ) an uncached B_DELWRI might loop due 2224 * to softupdates re-dirtying the buffer. In the latter 2225 * case, B_CACHE is set after the first write completes, 2226 * preventing further loops. 2227 */ 2228 2229 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2230 BUF_WRITE(bp); 2231 goto loop; 2232 } 2233 2234 splx(s); 2235 bp->b_flags &= ~B_DONE; 2236 } else { 2237 /* 2238 * Buffer is not in-core, create new buffer. The buffer 2239 * returned by getnewbuf() is locked. Note that the returned 2240 * buffer is also considered valid (not marked B_INVAL). 2241 */ 2242 int bsize, maxsize, vmio; 2243 off_t offset; 2244 2245 if (vn_isdisk(vp, NULL)) 2246 bsize = DEV_BSIZE; 2247 else if (vp->v_mountedhere) 2248 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2249 else if (vp->v_mount) 2250 bsize = vp->v_mount->mnt_stat.f_iosize; 2251 else 2252 bsize = size; 2253 2254 offset = (off_t)blkno * bsize; 2255 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && (vp->v_flag & VOBJBUF); 2256 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2257 maxsize = imax(maxsize, bsize); 2258 2259 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2260 if (slpflag || slptimeo) { 2261 splx(s); 2262 return NULL; 2263 } 2264 goto loop; 2265 } 2266 2267 /* 2268 * This code is used to make sure that a buffer is not 2269 * created while the getnewbuf routine is blocked. 2270 * This can be a problem whether the vnode is locked or not. 2271 * If the buffer is created out from under us, we have to 2272 * throw away the one we just created. There is now window 2273 * race because we are safely running at splbio() from the 2274 * point of the duplicate buffer creation through to here, 2275 * and we've locked the buffer. 2276 */ 2277 if (gbincore(vp, blkno)) { 2278 bp->b_flags |= B_INVAL; 2279 brelse(bp); 2280 goto loop; 2281 } 2282 2283 /* 2284 * Insert the buffer into the hash, so that it can 2285 * be found by incore. 2286 */ 2287 bp->b_blkno = bp->b_lblkno = blkno; 2288 bp->b_offset = offset; 2289 2290 bgetvp(vp, bp); 2291 LIST_REMOVE(bp, b_hash); 2292 bh = bufhash(vp, blkno); 2293 LIST_INSERT_HEAD(bh, bp, b_hash); 2294 2295 /* 2296 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2297 * buffer size starts out as 0, B_CACHE will be set by 2298 * allocbuf() for the VMIO case prior to it testing the 2299 * backing store for validity. 2300 */ 2301 2302 if (vmio) { 2303 bp->b_flags |= B_VMIO; 2304#if defined(VFS_BIO_DEBUG) 2305 if (vp->v_type != VREG) 2306 printf("getblk: vmioing file type %d???\n", vp->v_type); 2307#endif 2308 } else { 2309 bp->b_flags &= ~B_VMIO; 2310 } 2311 2312 allocbuf(bp, size); 2313 2314 splx(s); 2315 bp->b_flags &= ~B_DONE; 2316 } 2317 return (bp); 2318} 2319 2320/* 2321 * Get an empty, disassociated buffer of given size. The buffer is initially 2322 * set to B_INVAL. 2323 */ 2324struct buf * 2325geteblk(int size) 2326{ 2327 struct buf *bp; 2328 int s; 2329 int maxsize; 2330 2331 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2332 2333 s = splbio(); 2334 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0); 2335 splx(s); 2336 allocbuf(bp, size); 2337 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2338 return (bp); 2339} 2340 2341 2342/* 2343 * This code constitutes the buffer memory from either anonymous system 2344 * memory (in the case of non-VMIO operations) or from an associated 2345 * VM object (in the case of VMIO operations). This code is able to 2346 * resize a buffer up or down. 2347 * 2348 * Note that this code is tricky, and has many complications to resolve 2349 * deadlock or inconsistant data situations. Tread lightly!!! 2350 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2351 * the caller. Calling this code willy nilly can result in the loss of data. 2352 * 2353 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2354 * B_CACHE for the non-VMIO case. 2355 */ 2356 2357int 2358allocbuf(struct buf *bp, int size) 2359{ 2360 int newbsize, mbsize; 2361 int i; 2362 2363 if (BUF_REFCNT(bp) == 0) 2364 panic("allocbuf: buffer not busy"); 2365 2366 if (bp->b_kvasize < size) 2367 panic("allocbuf: buffer too small"); 2368 2369 if ((bp->b_flags & B_VMIO) == 0) { 2370 caddr_t origbuf; 2371 int origbufsize; 2372 /* 2373 * Just get anonymous memory from the kernel. Don't 2374 * mess with B_CACHE. 2375 */ 2376 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2377#if !defined(NO_B_MALLOC) 2378 if (bp->b_flags & B_MALLOC) 2379 newbsize = mbsize; 2380 else 2381#endif 2382 newbsize = round_page(size); 2383 2384 if (newbsize < bp->b_bufsize) { 2385#if !defined(NO_B_MALLOC) 2386 /* 2387 * malloced buffers are not shrunk 2388 */ 2389 if (bp->b_flags & B_MALLOC) { 2390 if (newbsize) { 2391 bp->b_bcount = size; 2392 } else { 2393 free(bp->b_data, M_BIOBUF); 2394 if (bp->b_bufsize) { 2395 bufmallocspace -= bp->b_bufsize; 2396 bufspacewakeup(); 2397 bp->b_bufsize = 0; 2398 } 2399 bp->b_data = bp->b_kvabase; 2400 bp->b_bcount = 0; 2401 bp->b_flags &= ~B_MALLOC; 2402 } 2403 return 1; 2404 } 2405#endif 2406 vm_hold_free_pages( 2407 bp, 2408 (vm_offset_t) bp->b_data + newbsize, 2409 (vm_offset_t) bp->b_data + bp->b_bufsize); 2410 } else if (newbsize > bp->b_bufsize) { 2411#if !defined(NO_B_MALLOC) 2412 /* 2413 * We only use malloced memory on the first allocation. 2414 * and revert to page-allocated memory when the buffer 2415 * grows. 2416 */ 2417 if ( (bufmallocspace < maxbufmallocspace) && 2418 (bp->b_bufsize == 0) && 2419 (mbsize <= PAGE_SIZE/2)) { 2420 2421 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2422 bp->b_bufsize = mbsize; 2423 bp->b_bcount = size; 2424 bp->b_flags |= B_MALLOC; 2425 bufmallocspace += mbsize; 2426 return 1; 2427 } 2428#endif 2429 origbuf = NULL; 2430 origbufsize = 0; 2431#if !defined(NO_B_MALLOC) 2432 /* 2433 * If the buffer is growing on its other-than-first allocation, 2434 * then we revert to the page-allocation scheme. 2435 */ 2436 if (bp->b_flags & B_MALLOC) { 2437 origbuf = bp->b_data; 2438 origbufsize = bp->b_bufsize; 2439 bp->b_data = bp->b_kvabase; 2440 if (bp->b_bufsize) { 2441 bufmallocspace -= bp->b_bufsize; 2442 bufspacewakeup(); 2443 bp->b_bufsize = 0; 2444 } 2445 bp->b_flags &= ~B_MALLOC; 2446 newbsize = round_page(newbsize); 2447 } 2448#endif 2449 vm_hold_load_pages( 2450 bp, 2451 (vm_offset_t) bp->b_data + bp->b_bufsize, 2452 (vm_offset_t) bp->b_data + newbsize); 2453#if !defined(NO_B_MALLOC) 2454 if (origbuf) { 2455 bcopy(origbuf, bp->b_data, origbufsize); 2456 free(origbuf, M_BIOBUF); 2457 } 2458#endif 2459 } 2460 } else { 2461 vm_page_t m; 2462 int desiredpages; 2463 2464 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2465 desiredpages = (size == 0) ? 0 : 2466 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2467 2468#if !defined(NO_B_MALLOC) 2469 if (bp->b_flags & B_MALLOC) 2470 panic("allocbuf: VMIO buffer can't be malloced"); 2471#endif 2472 /* 2473 * Set B_CACHE initially if buffer is 0 length or will become 2474 * 0-length. 2475 */ 2476 if (size == 0 || bp->b_bufsize == 0) 2477 bp->b_flags |= B_CACHE; 2478 2479 if (newbsize < bp->b_bufsize) { 2480 /* 2481 * DEV_BSIZE aligned new buffer size is less then the 2482 * DEV_BSIZE aligned existing buffer size. Figure out 2483 * if we have to remove any pages. 2484 */ 2485 mtx_lock(&vm_mtx); 2486 if (desiredpages < bp->b_npages) { 2487 for (i = desiredpages; i < bp->b_npages; i++) { 2488 /* 2489 * the page is not freed here -- it 2490 * is the responsibility of 2491 * vnode_pager_setsize 2492 */ 2493 m = bp->b_pages[i]; 2494 KASSERT(m != bogus_page, 2495 ("allocbuf: bogus page found")); 2496 while (vm_page_sleep_busy(m, TRUE, "biodep")) 2497 ; 2498 2499 bp->b_pages[i] = NULL; 2500 vm_page_unwire(m, 0); 2501 } 2502 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2503 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2504 bp->b_npages = desiredpages; 2505 } 2506 mtx_unlock(&vm_mtx); 2507 } else if (size > bp->b_bcount) { 2508 /* 2509 * We are growing the buffer, possibly in a 2510 * byte-granular fashion. 2511 */ 2512 struct vnode *vp; 2513 vm_object_t obj; 2514 vm_offset_t toff; 2515 vm_offset_t tinc; 2516 2517 /* 2518 * Step 1, bring in the VM pages from the object, 2519 * allocating them if necessary. We must clear 2520 * B_CACHE if these pages are not valid for the 2521 * range covered by the buffer. 2522 */ 2523 2524 vp = bp->b_vp; 2525 VOP_GETVOBJECT(vp, &obj); 2526 2527 mtx_lock(&vm_mtx); 2528 while (bp->b_npages < desiredpages) { 2529 vm_page_t m; 2530 vm_pindex_t pi; 2531 2532 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2533 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2534 /* 2535 * note: must allocate system pages 2536 * since blocking here could intefere 2537 * with paging I/O, no matter which 2538 * process we are. 2539 */ 2540 m = vm_page_alloc(obj, pi, VM_ALLOC_SYSTEM); 2541 if (m == NULL) { 2542 VM_WAIT; 2543 vm_pageout_deficit += desiredpages - bp->b_npages; 2544 } else { 2545 vm_page_wire(m); 2546 vm_page_wakeup(m); 2547 bp->b_flags &= ~B_CACHE; 2548 bp->b_pages[bp->b_npages] = m; 2549 ++bp->b_npages; 2550 } 2551 continue; 2552 } 2553 2554 /* 2555 * We found a page. If we have to sleep on it, 2556 * retry because it might have gotten freed out 2557 * from under us. 2558 * 2559 * We can only test PG_BUSY here. Blocking on 2560 * m->busy might lead to a deadlock: 2561 * 2562 * vm_fault->getpages->cluster_read->allocbuf 2563 * 2564 */ 2565 2566 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 2567 continue; 2568 2569 /* 2570 * We have a good page. Should we wakeup the 2571 * page daemon? 2572 */ 2573 if ((curproc != pageproc) && 2574 ((m->queue - m->pc) == PQ_CACHE) && 2575 ((cnt.v_free_count + cnt.v_cache_count) < 2576 (cnt.v_free_min + cnt.v_cache_min))) { 2577 pagedaemon_wakeup(); 2578 } 2579 vm_page_flag_clear(m, PG_ZERO); 2580 vm_page_wire(m); 2581 bp->b_pages[bp->b_npages] = m; 2582 ++bp->b_npages; 2583 } 2584 2585 /* 2586 * Step 2. We've loaded the pages into the buffer, 2587 * we have to figure out if we can still have B_CACHE 2588 * set. Note that B_CACHE is set according to the 2589 * byte-granular range ( bcount and size ), new the 2590 * aligned range ( newbsize ). 2591 * 2592 * The VM test is against m->valid, which is DEV_BSIZE 2593 * aligned. Needless to say, the validity of the data 2594 * needs to also be DEV_BSIZE aligned. Note that this 2595 * fails with NFS if the server or some other client 2596 * extends the file's EOF. If our buffer is resized, 2597 * B_CACHE may remain set! XXX 2598 */ 2599 2600 toff = bp->b_bcount; 2601 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2602 2603 while ((bp->b_flags & B_CACHE) && toff < size) { 2604 vm_pindex_t pi; 2605 2606 if (tinc > (size - toff)) 2607 tinc = size - toff; 2608 2609 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2610 PAGE_SHIFT; 2611 2612 vfs_buf_test_cache( 2613 bp, 2614 bp->b_offset, 2615 toff, 2616 tinc, 2617 bp->b_pages[pi] 2618 ); 2619 toff += tinc; 2620 tinc = PAGE_SIZE; 2621 } 2622 2623 /* 2624 * Step 3, fixup the KVM pmap. Remember that 2625 * bp->b_data is relative to bp->b_offset, but 2626 * bp->b_offset may be offset into the first page. 2627 */ 2628 2629 bp->b_data = (caddr_t) 2630 trunc_page((vm_offset_t)bp->b_data); 2631 pmap_qenter( 2632 (vm_offset_t)bp->b_data, 2633 bp->b_pages, 2634 bp->b_npages 2635 ); 2636 2637 mtx_unlock(&vm_mtx); 2638 2639 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2640 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2641 } 2642 } 2643 if (newbsize < bp->b_bufsize) 2644 bufspacewakeup(); 2645 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2646 bp->b_bcount = size; /* requested buffer size */ 2647 return 1; 2648} 2649 2650/* 2651 * bufwait: 2652 * 2653 * Wait for buffer I/O completion, returning error status. The buffer 2654 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR 2655 * error and cleared. 2656 */ 2657int 2658bufwait(register struct buf * bp) 2659{ 2660 int s; 2661 2662 s = splbio(); 2663 while ((bp->b_flags & B_DONE) == 0) { 2664 if (bp->b_iocmd == BIO_READ) 2665 tsleep(bp, PRIBIO, "biord", 0); 2666 else 2667 tsleep(bp, PRIBIO, "biowr", 0); 2668 } 2669 splx(s); 2670 if (bp->b_flags & B_EINTR) { 2671 bp->b_flags &= ~B_EINTR; 2672 return (EINTR); 2673 } 2674 if (bp->b_ioflags & BIO_ERROR) { 2675 return (bp->b_error ? bp->b_error : EIO); 2676 } else { 2677 return (0); 2678 } 2679} 2680 2681 /* 2682 * Call back function from struct bio back up to struct buf. 2683 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 2684 */ 2685void 2686bufdonebio(struct bio *bp) 2687{ 2688 bufdone(bp->bio_caller2); 2689} 2690 2691/* 2692 * bufdone: 2693 * 2694 * Finish I/O on a buffer, optionally calling a completion function. 2695 * This is usually called from an interrupt so process blocking is 2696 * not allowed. 2697 * 2698 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2699 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2700 * assuming B_INVAL is clear. 2701 * 2702 * For the VMIO case, we set B_CACHE if the op was a read and no 2703 * read error occured, or if the op was a write. B_CACHE is never 2704 * set if the buffer is invalid or otherwise uncacheable. 2705 * 2706 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2707 * initiator to leave B_INVAL set to brelse the buffer out of existance 2708 * in the biodone routine. 2709 */ 2710void 2711bufdone(struct buf *bp) 2712{ 2713 int s, error; 2714 void (*biodone) __P((struct buf *)); 2715 2716 s = splbio(); 2717 2718 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 2719 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 2720 2721 bp->b_flags |= B_DONE; 2722 runningbufwakeup(bp); 2723 2724 if (bp->b_iocmd == BIO_DELETE) { 2725 brelse(bp); 2726 splx(s); 2727 return; 2728 } 2729 2730 if (bp->b_iocmd == BIO_WRITE) { 2731 vwakeup(bp); 2732 } 2733 2734 /* call optional completion function if requested */ 2735 if (bp->b_iodone != NULL) { 2736 biodone = bp->b_iodone; 2737 bp->b_iodone = NULL; 2738 (*biodone) (bp); 2739 splx(s); 2740 return; 2741 } 2742 if (LIST_FIRST(&bp->b_dep) != NULL) 2743 buf_complete(bp); 2744 2745 if (bp->b_flags & B_VMIO) { 2746 int i; 2747 vm_ooffset_t foff; 2748 vm_page_t m; 2749 vm_object_t obj; 2750 int iosize; 2751 struct vnode *vp = bp->b_vp; 2752 2753 error = VOP_GETVOBJECT(vp, &obj); 2754 2755#if defined(VFS_BIO_DEBUG) 2756 if (vp->v_usecount == 0) { 2757 panic("biodone: zero vnode ref count"); 2758 } 2759 2760 if (error) { 2761 panic("biodone: missing VM object"); 2762 } 2763 2764 if ((vp->v_flag & VOBJBUF) == 0) { 2765 panic("biodone: vnode is not setup for merged cache"); 2766 } 2767#endif 2768 2769 foff = bp->b_offset; 2770 KASSERT(bp->b_offset != NOOFFSET, 2771 ("biodone: no buffer offset")); 2772 2773 if (error) { 2774 panic("biodone: no object"); 2775 } 2776 mtx_lock(&vm_mtx); 2777#if defined(VFS_BIO_DEBUG) 2778 if (obj->paging_in_progress < bp->b_npages) { 2779 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 2780 obj->paging_in_progress, bp->b_npages); 2781 } 2782#endif 2783 2784 /* 2785 * Set B_CACHE if the op was a normal read and no error 2786 * occured. B_CACHE is set for writes in the b*write() 2787 * routines. 2788 */ 2789 iosize = bp->b_bcount - bp->b_resid; 2790 if (bp->b_iocmd == BIO_READ && 2791 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 2792 !(bp->b_ioflags & BIO_ERROR)) { 2793 bp->b_flags |= B_CACHE; 2794 } 2795 2796 for (i = 0; i < bp->b_npages; i++) { 2797 int bogusflag = 0; 2798 int resid; 2799 2800 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2801 if (resid > iosize) 2802 resid = iosize; 2803 2804 /* 2805 * cleanup bogus pages, restoring the originals 2806 */ 2807 m = bp->b_pages[i]; 2808 if (m == bogus_page) { 2809 bogusflag = 1; 2810 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2811 if (m == NULL) 2812 panic("biodone: page disappeared!"); 2813 bp->b_pages[i] = m; 2814 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2815 } 2816#if defined(VFS_BIO_DEBUG) 2817 if (OFF_TO_IDX(foff) != m->pindex) { 2818 printf( 2819"biodone: foff(%lu)/m->pindex(%d) mismatch\n", 2820 (unsigned long)foff, m->pindex); 2821 } 2822#endif 2823 2824 /* 2825 * In the write case, the valid and clean bits are 2826 * already changed correctly ( see bdwrite() ), so we 2827 * only need to do this here in the read case. 2828 */ 2829 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 2830 vfs_page_set_valid(bp, foff, i, m); 2831 } 2832 vm_page_flag_clear(m, PG_ZERO); 2833 2834 /* 2835 * when debugging new filesystems or buffer I/O methods, this 2836 * is the most common error that pops up. if you see this, you 2837 * have not set the page busy flag correctly!!! 2838 */ 2839 if (m->busy == 0) { 2840 printf("biodone: page busy < 0, " 2841 "pindex: %d, foff: 0x(%x,%x), " 2842 "resid: %d, index: %d\n", 2843 (int) m->pindex, (int)(foff >> 32), 2844 (int) foff & 0xffffffff, resid, i); 2845 if (!vn_isdisk(vp, NULL)) 2846 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 2847 bp->b_vp->v_mount->mnt_stat.f_iosize, 2848 (int) bp->b_lblkno, 2849 bp->b_flags, bp->b_npages); 2850 else 2851 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 2852 (int) bp->b_lblkno, 2853 bp->b_flags, bp->b_npages); 2854 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2855 m->valid, m->dirty, m->wire_count); 2856 panic("biodone: page busy < 0\n"); 2857 } 2858 vm_page_io_finish(m); 2859 vm_object_pip_subtract(obj, 1); 2860 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2861 iosize -= resid; 2862 } 2863 if (obj) 2864 vm_object_pip_wakeupn(obj, 0); 2865 mtx_unlock(&vm_mtx); 2866 } 2867 2868 /* 2869 * For asynchronous completions, release the buffer now. The brelse 2870 * will do a wakeup there if necessary - so no need to do a wakeup 2871 * here in the async case. The sync case always needs to do a wakeup. 2872 */ 2873 2874 if (bp->b_flags & B_ASYNC) { 2875 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 2876 brelse(bp); 2877 else 2878 bqrelse(bp); 2879 } else { 2880 wakeup(bp); 2881 } 2882 splx(s); 2883} 2884 2885/* 2886 * This routine is called in lieu of iodone in the case of 2887 * incomplete I/O. This keeps the busy status for pages 2888 * consistant. 2889 * 2890 * vm_mtx should not be held 2891 */ 2892void 2893vfs_unbusy_pages(struct buf * bp) 2894{ 2895 int i; 2896 2897 mtx_assert(&vm_mtx, MA_NOTOWNED); 2898 runningbufwakeup(bp); 2899 if (bp->b_flags & B_VMIO) { 2900 struct vnode *vp = bp->b_vp; 2901 vm_object_t obj; 2902 2903 VOP_GETVOBJECT(vp, &obj); 2904 2905 mtx_lock(&vm_mtx); 2906 for (i = 0; i < bp->b_npages; i++) { 2907 vm_page_t m = bp->b_pages[i]; 2908 2909 if (m == bogus_page) { 2910 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 2911 if (!m) { 2912 panic("vfs_unbusy_pages: page missing\n"); 2913 } 2914 bp->b_pages[i] = m; 2915 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2916 } 2917 vm_object_pip_subtract(obj, 1); 2918 vm_page_flag_clear(m, PG_ZERO); 2919 vm_page_io_finish(m); 2920 } 2921 vm_object_pip_wakeupn(obj, 0); 2922 mtx_unlock(&vm_mtx); 2923 } 2924} 2925 2926/* 2927 * vfs_page_set_valid: 2928 * 2929 * Set the valid bits in a page based on the supplied offset. The 2930 * range is restricted to the buffer's size. 2931 * 2932 * This routine is typically called after a read completes. 2933 * 2934 * vm_mtx should be held 2935 */ 2936static void 2937vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 2938{ 2939 vm_ooffset_t soff, eoff; 2940 2941 mtx_assert(&vm_mtx, MA_OWNED); 2942 /* 2943 * Start and end offsets in buffer. eoff - soff may not cross a 2944 * page boundry or cross the end of the buffer. The end of the 2945 * buffer, in this case, is our file EOF, not the allocation size 2946 * of the buffer. 2947 */ 2948 soff = off; 2949 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2950 if (eoff > bp->b_offset + bp->b_bcount) 2951 eoff = bp->b_offset + bp->b_bcount; 2952 2953 /* 2954 * Set valid range. This is typically the entire buffer and thus the 2955 * entire page. 2956 */ 2957 if (eoff > soff) { 2958 vm_page_set_validclean( 2959 m, 2960 (vm_offset_t) (soff & PAGE_MASK), 2961 (vm_offset_t) (eoff - soff) 2962 ); 2963 } 2964} 2965 2966/* 2967 * This routine is called before a device strategy routine. 2968 * It is used to tell the VM system that paging I/O is in 2969 * progress, and treat the pages associated with the buffer 2970 * almost as being PG_BUSY. Also the object paging_in_progress 2971 * flag is handled to make sure that the object doesn't become 2972 * inconsistant. 2973 * 2974 * Since I/O has not been initiated yet, certain buffer flags 2975 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 2976 * and should be ignored. 2977 * 2978 * vm_mtx should not be held 2979 */ 2980void 2981vfs_busy_pages(struct buf * bp, int clear_modify) 2982{ 2983 int i, bogus; 2984 2985 mtx_assert(&vm_mtx, MA_NOTOWNED); 2986 if (bp->b_flags & B_VMIO) { 2987 struct vnode *vp = bp->b_vp; 2988 vm_object_t obj; 2989 vm_ooffset_t foff; 2990 2991 VOP_GETVOBJECT(vp, &obj); 2992 foff = bp->b_offset; 2993 KASSERT(bp->b_offset != NOOFFSET, 2994 ("vfs_busy_pages: no buffer offset")); 2995 mtx_lock(&vm_mtx); 2996 vfs_setdirty(bp); 2997 2998retry: 2999 for (i = 0; i < bp->b_npages; i++) { 3000 vm_page_t m = bp->b_pages[i]; 3001 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 3002 goto retry; 3003 } 3004 3005 bogus = 0; 3006 for (i = 0; i < bp->b_npages; i++) { 3007 vm_page_t m = bp->b_pages[i]; 3008 3009 vm_page_flag_clear(m, PG_ZERO); 3010 if ((bp->b_flags & B_CLUSTER) == 0) { 3011 vm_object_pip_add(obj, 1); 3012 vm_page_io_start(m); 3013 } 3014 3015 /* 3016 * When readying a buffer for a read ( i.e 3017 * clear_modify == 0 ), it is important to do 3018 * bogus_page replacement for valid pages in 3019 * partially instantiated buffers. Partially 3020 * instantiated buffers can, in turn, occur when 3021 * reconstituting a buffer from its VM backing store 3022 * base. We only have to do this if B_CACHE is 3023 * clear ( which causes the I/O to occur in the 3024 * first place ). The replacement prevents the read 3025 * I/O from overwriting potentially dirty VM-backed 3026 * pages. XXX bogus page replacement is, uh, bogus. 3027 * It may not work properly with small-block devices. 3028 * We need to find a better way. 3029 */ 3030 3031 vm_page_protect(m, VM_PROT_NONE); 3032 if (clear_modify) 3033 vfs_page_set_valid(bp, foff, i, m); 3034 else if (m->valid == VM_PAGE_BITS_ALL && 3035 (bp->b_flags & B_CACHE) == 0) { 3036 bp->b_pages[i] = bogus_page; 3037 bogus++; 3038 } 3039 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3040 } 3041 if (bogus) 3042 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3043 mtx_unlock(&vm_mtx); 3044 } 3045} 3046 3047/* 3048 * Tell the VM system that the pages associated with this buffer 3049 * are clean. This is used for delayed writes where the data is 3050 * going to go to disk eventually without additional VM intevention. 3051 * 3052 * Note that while we only really need to clean through to b_bcount, we 3053 * just go ahead and clean through to b_bufsize. 3054 * 3055 * should be called with vm_mtx held 3056 */ 3057static void 3058vfs_clean_pages(struct buf * bp) 3059{ 3060 int i; 3061 3062 mtx_assert(&vm_mtx, MA_OWNED); 3063 if (bp->b_flags & B_VMIO) { 3064 vm_ooffset_t foff; 3065 3066 foff = bp->b_offset; 3067 KASSERT(bp->b_offset != NOOFFSET, 3068 ("vfs_clean_pages: no buffer offset")); 3069 for (i = 0; i < bp->b_npages; i++) { 3070 vm_page_t m = bp->b_pages[i]; 3071 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3072 vm_ooffset_t eoff = noff; 3073 3074 if (eoff > bp->b_offset + bp->b_bufsize) 3075 eoff = bp->b_offset + bp->b_bufsize; 3076 vfs_page_set_valid(bp, foff, i, m); 3077 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3078 foff = noff; 3079 } 3080 } 3081} 3082 3083/* 3084 * vfs_bio_set_validclean: 3085 * 3086 * Set the range within the buffer to valid and clean. The range is 3087 * relative to the beginning of the buffer, b_offset. Note that b_offset 3088 * itself may be offset from the beginning of the first page. 3089 * 3090 */ 3091 3092void 3093vfs_bio_set_validclean(struct buf *bp, int base, int size) 3094{ 3095 if (bp->b_flags & B_VMIO) { 3096 int i; 3097 int n; 3098 3099 /* 3100 * Fixup base to be relative to beginning of first page. 3101 * Set initial n to be the maximum number of bytes in the 3102 * first page that can be validated. 3103 */ 3104 3105 base += (bp->b_offset & PAGE_MASK); 3106 n = PAGE_SIZE - (base & PAGE_MASK); 3107 3108 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3109 vm_page_t m = bp->b_pages[i]; 3110 3111 if (n > size) 3112 n = size; 3113 3114 vm_page_set_validclean(m, base & PAGE_MASK, n); 3115 base += n; 3116 size -= n; 3117 n = PAGE_SIZE; 3118 } 3119 } 3120} 3121 3122/* 3123 * vfs_bio_clrbuf: 3124 * 3125 * clear a buffer. This routine essentially fakes an I/O, so we need 3126 * to clear BIO_ERROR and B_INVAL. 3127 * 3128 * Note that while we only theoretically need to clear through b_bcount, 3129 * we go ahead and clear through b_bufsize. 3130 * 3131 * We'll get vm_mtx here for safety if processing a VMIO buffer. 3132 * I don't think vm_mtx is needed, but we're twiddling vm_page flags. 3133 */ 3134 3135void 3136vfs_bio_clrbuf(struct buf *bp) { 3137 int i, mask = 0; 3138 caddr_t sa, ea; 3139 3140 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3141 mtx_lock(&vm_mtx); 3142 bp->b_flags &= ~B_INVAL; 3143 bp->b_ioflags &= ~BIO_ERROR; 3144 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3145 (bp->b_offset & PAGE_MASK) == 0) { 3146 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3147 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3148 ((bp->b_pages[0]->valid & mask) != mask)) { 3149 bzero(bp->b_data, bp->b_bufsize); 3150 } 3151 bp->b_pages[0]->valid |= mask; 3152 bp->b_resid = 0; 3153 mtx_unlock(&vm_mtx); 3154 return; 3155 } 3156 ea = sa = bp->b_data; 3157 for(i=0;i<bp->b_npages;i++,sa=ea) { 3158 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3159 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3160 ea = (caddr_t)(vm_offset_t)ulmin( 3161 (u_long)(vm_offset_t)ea, 3162 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3163 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3164 if ((bp->b_pages[i]->valid & mask) == mask) 3165 continue; 3166 if ((bp->b_pages[i]->valid & mask) == 0) { 3167 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3168 bzero(sa, ea - sa); 3169 } 3170 } else { 3171 for (; sa < ea; sa += DEV_BSIZE, j++) { 3172 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3173 (bp->b_pages[i]->valid & (1<<j)) == 0) 3174 bzero(sa, DEV_BSIZE); 3175 } 3176 } 3177 bp->b_pages[i]->valid |= mask; 3178 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3179 } 3180 bp->b_resid = 0; 3181 mtx_unlock(&vm_mtx); 3182 } else { 3183 clrbuf(bp); 3184 } 3185} 3186 3187/* 3188 * vm_hold_load_pages and vm_hold_free_pages get pages into 3189 * a buffers address space. The pages are anonymous and are 3190 * not associated with a file object. 3191 * 3192 * vm_mtx should not be held 3193 */ 3194static void 3195vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3196{ 3197 vm_offset_t pg; 3198 vm_page_t p; 3199 int index; 3200 3201 mtx_assert(&vm_mtx, MA_NOTOWNED); 3202 to = round_page(to); 3203 from = round_page(from); 3204 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3205 3206 mtx_lock(&vm_mtx); 3207 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3208 3209tryagain: 3210 3211 /* 3212 * note: must allocate system pages since blocking here 3213 * could intefere with paging I/O, no matter which 3214 * process we are. 3215 */ 3216 p = vm_page_alloc(kernel_object, 3217 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3218 VM_ALLOC_SYSTEM); 3219 if (!p) { 3220 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3221 VM_WAIT; 3222 goto tryagain; 3223 } 3224 vm_page_wire(p); 3225 p->valid = VM_PAGE_BITS_ALL; 3226 vm_page_flag_clear(p, PG_ZERO); 3227 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 3228 bp->b_pages[index] = p; 3229 vm_page_wakeup(p); 3230 } 3231 bp->b_npages = index; 3232 mtx_unlock(&vm_mtx); 3233} 3234 3235void 3236vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3237{ 3238 vm_offset_t pg; 3239 vm_page_t p; 3240 int index, newnpages; 3241 3242 mtx_assert(&vm_mtx, MA_NOTOWNED); 3243 from = round_page(from); 3244 to = round_page(to); 3245 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3246 3247 mtx_lock(&vm_mtx); 3248 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3249 p = bp->b_pages[index]; 3250 if (p && (index < bp->b_npages)) { 3251 if (p->busy) { 3252 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 3253 bp->b_blkno, bp->b_lblkno); 3254 } 3255 bp->b_pages[index] = NULL; 3256 pmap_kremove(pg); 3257 vm_page_busy(p); 3258 vm_page_unwire(p, 0); 3259 vm_page_free(p); 3260 } 3261 } 3262 bp->b_npages = newnpages; 3263 mtx_unlock(&vm_mtx); 3264} 3265 3266 3267#include "opt_ddb.h" 3268#ifdef DDB 3269#include <ddb/ddb.h> 3270 3271DB_SHOW_COMMAND(buffer, db_show_buffer) 3272{ 3273 /* get args */ 3274 struct buf *bp = (struct buf *)addr; 3275 3276 if (!have_addr) { 3277 db_printf("usage: show buffer <addr>\n"); 3278 return; 3279 } 3280 3281 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3282 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 3283 "b_resid = %ld\nb_dev = (%d,%d), b_data = %p, " 3284 "b_blkno = %d, b_pblkno = %d\n", 3285 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3286 major(bp->b_dev), minor(bp->b_dev), 3287 bp->b_data, bp->b_blkno, bp->b_pblkno); 3288 if (bp->b_npages) { 3289 int i; 3290 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3291 for (i = 0; i < bp->b_npages; i++) { 3292 vm_page_t m; 3293 m = bp->b_pages[i]; 3294 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3295 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3296 if ((i + 1) < bp->b_npages) 3297 db_printf(","); 3298 } 3299 db_printf("\n"); 3300 } 3301} 3302#endif /* DDB */ 3303