vfs_bio.c revision 82127
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 82127 2001-08-22 04:07:27Z dillon $ 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 */ 285static __inline__ 286void 287vfs_buf_test_cache(struct buf *bp, 288 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 289 vm_page_t m) 290{ 291 GIANT_REQUIRED; 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 * Calculating buffer cache scaling values and reserve space for buffer 323 * headers. This is called during low level kernel initialization and 324 * may be called more then once. We CANNOT write to the memory area 325 * being reserved at this time. 326 */ 327caddr_t 328kern_vfs_bio_buffer_alloc(caddr_t v, int physmem_est) 329{ 330 /* 331 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 332 * For the first 64MB of ram nominally allocate sufficient buffers to 333 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 334 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 335 * the buffer cache we limit the eventual kva reservation to 336 * maxbcache bytes. 337 * 338 * factor represents the 1/4 x ram conversion. 339 */ 340 if (nbuf == 0) { 341 int factor = 4 * BKVASIZE / PAGE_SIZE; 342 343 nbuf = 50; 344 if (physmem_est > 1024) 345 nbuf += min((physmem_est - 1024) / factor, 346 16384 / factor); 347 if (physmem_est > 16384) 348 nbuf += (physmem_est - 16384) * 2 / (factor * 5); 349 350 if (maxbcache && nbuf > maxbcache / BKVASIZE) 351 nbuf = maxbcache / BKVASIZE; 352 } 353 354 /* 355 * Do not allow the buffer_map to be more then 1/2 the size of the 356 * kernel_map. 357 */ 358 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 359 (BKVASIZE * 2)) { 360 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 361 (BKVASIZE * 2); 362 printf("Warning: nbufs capped at %d\n", nbuf); 363 } 364 365 /* 366 * swbufs are used as temporary holders for I/O, such as paging I/O. 367 * We have no less then 16 and no more then 256. 368 */ 369 nswbuf = max(min(nbuf/4, 256), 16); 370 371 /* 372 * Reserve space for the buffer cache buffers 373 */ 374 swbuf = (void *)v; 375 v = (caddr_t)(swbuf + nswbuf); 376 buf = (void *)v; 377 v = (caddr_t)(buf + nbuf); 378 379 /* 380 * Calculate the hash table size and reserve space 381 */ 382 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 383 ; 384 bufhashtbl = (void *)v; 385 v = (caddr_t)(bufhashtbl + bufhashmask); 386 --bufhashmask; 387 388 return(v); 389} 390 391void 392bufinit(void) 393{ 394 struct buf *bp; 395 int i; 396 397 GIANT_REQUIRED; 398 399 TAILQ_INIT(&bswlist); 400 LIST_INIT(&invalhash); 401 mtx_init(&buftimelock, "buftime lock", MTX_DEF); 402 403 for (i = 0; i <= bufhashmask; i++) 404 LIST_INIT(&bufhashtbl[i]); 405 406 /* next, make a null set of free lists */ 407 for (i = 0; i < BUFFER_QUEUES; i++) 408 TAILQ_INIT(&bufqueues[i]); 409 410 /* finally, initialize each buffer header and stick on empty q */ 411 for (i = 0; i < nbuf; i++) { 412 bp = &buf[i]; 413 bzero(bp, sizeof *bp); 414 bp->b_flags = B_INVAL; /* we're just an empty header */ 415 bp->b_dev = NODEV; 416 bp->b_rcred = NOCRED; 417 bp->b_wcred = NOCRED; 418 bp->b_qindex = QUEUE_EMPTY; 419 bp->b_xflags = 0; 420 LIST_INIT(&bp->b_dep); 421 BUF_LOCKINIT(bp); 422 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 423 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 424 } 425 426 /* 427 * maxbufspace is the absolute maximum amount of buffer space we are 428 * allowed to reserve in KVM and in real terms. The absolute maximum 429 * is nominally used by buf_daemon. hibufspace is the nominal maximum 430 * used by most other processes. The differential is required to 431 * ensure that buf_daemon is able to run when other processes might 432 * be blocked waiting for buffer space. 433 * 434 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 435 * this may result in KVM fragmentation which is not handled optimally 436 * by the system. 437 */ 438 maxbufspace = nbuf * BKVASIZE; 439 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 440 lobufspace = hibufspace - MAXBSIZE; 441 442 lorunningspace = 512 * 1024; 443 hirunningspace = 1024 * 1024; 444 445/* 446 * Limit the amount of malloc memory since it is wired permanently into 447 * the kernel space. Even though this is accounted for in the buffer 448 * allocation, we don't want the malloced region to grow uncontrolled. 449 * The malloc scheme improves memory utilization significantly on average 450 * (small) directories. 451 */ 452 maxbufmallocspace = hibufspace / 20; 453 454/* 455 * Reduce the chance of a deadlock occuring by limiting the number 456 * of delayed-write dirty buffers we allow to stack up. 457 */ 458 hidirtybuffers = nbuf / 4 + 20; 459 numdirtybuffers = 0; 460/* 461 * To support extreme low-memory systems, make sure hidirtybuffers cannot 462 * eat up all available buffer space. This occurs when our minimum cannot 463 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 464 * BKVASIZE'd (8K) buffers. 465 */ 466 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 467 hidirtybuffers >>= 1; 468 } 469 lodirtybuffers = hidirtybuffers / 2; 470 471/* 472 * Try to keep the number of free buffers in the specified range, 473 * and give special processes (e.g. like buf_daemon) access to an 474 * emergency reserve. 475 */ 476 lofreebuffers = nbuf / 18 + 5; 477 hifreebuffers = 2 * lofreebuffers; 478 numfreebuffers = nbuf; 479 480/* 481 * Maximum number of async ops initiated per buf_daemon loop. This is 482 * somewhat of a hack at the moment, we really need to limit ourselves 483 * based on the number of bytes of I/O in-transit that were initiated 484 * from buf_daemon. 485 */ 486 487 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 488 bogus_page = vm_page_alloc(kernel_object, 489 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 490 VM_ALLOC_NORMAL); 491 cnt.v_wire_count++; 492} 493 494/* 495 * bfreekva() - free the kva allocation for a buffer. 496 * 497 * Must be called at splbio() or higher as this is the only locking for 498 * buffer_map. 499 * 500 * Since this call frees up buffer space, we call bufspacewakeup(). 501 */ 502static void 503bfreekva(struct buf * bp) 504{ 505 GIANT_REQUIRED; 506 507 if (bp->b_kvasize) { 508 ++buffreekvacnt; 509 bufspace -= bp->b_kvasize; 510 vm_map_delete(buffer_map, 511 (vm_offset_t) bp->b_kvabase, 512 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 513 ); 514 bp->b_kvasize = 0; 515 bufspacewakeup(); 516 } 517} 518 519/* 520 * bremfree: 521 * 522 * Remove the buffer from the appropriate free list. 523 */ 524void 525bremfree(struct buf * bp) 526{ 527 int s = splbio(); 528 int old_qindex = bp->b_qindex; 529 530 GIANT_REQUIRED; 531 532 if (bp->b_qindex != QUEUE_NONE) { 533 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 534 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 535 bp->b_qindex = QUEUE_NONE; 536 } else { 537 if (BUF_REFCNT(bp) <= 1) 538 panic("bremfree: removing a buffer not on a queue"); 539 } 540 541 /* 542 * Fixup numfreebuffers count. If the buffer is invalid or not 543 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 544 * the buffer was free and we must decrement numfreebuffers. 545 */ 546 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 547 switch(old_qindex) { 548 case QUEUE_DIRTY: 549 case QUEUE_CLEAN: 550 case QUEUE_EMPTY: 551 case QUEUE_EMPTYKVA: 552 --numfreebuffers; 553 break; 554 default: 555 break; 556 } 557 } 558 splx(s); 559} 560 561 562/* 563 * Get a buffer with the specified data. Look in the cache first. We 564 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 565 * is set, the buffer is valid and we do not have to do anything ( see 566 * getblk() ). This is really just a special case of breadn(). 567 */ 568int 569bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 570 struct buf ** bpp) 571{ 572 573 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 574} 575 576/* 577 * Operates like bread, but also starts asynchronous I/O on 578 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 579 * to initiating I/O . If B_CACHE is set, the buffer is valid 580 * and we do not have to do anything. 581 */ 582int 583breadn(struct vnode * vp, daddr_t blkno, int size, 584 daddr_t * rablkno, int *rabsize, 585 int cnt, struct ucred * cred, struct buf ** bpp) 586{ 587 struct buf *bp, *rabp; 588 int i; 589 int rv = 0, readwait = 0; 590 591 *bpp = bp = getblk(vp, blkno, size, 0, 0); 592 593 /* if not found in cache, do some I/O */ 594 if ((bp->b_flags & B_CACHE) == 0) { 595 if (curproc != PCPU_GET(idleproc)) 596 curproc->p_stats->p_ru.ru_inblock++; 597 bp->b_iocmd = BIO_READ; 598 bp->b_flags &= ~B_INVAL; 599 bp->b_ioflags &= ~BIO_ERROR; 600 if (bp->b_rcred == NOCRED) { 601 if (cred != NOCRED) 602 crhold(cred); 603 bp->b_rcred = cred; 604 } 605 vfs_busy_pages(bp, 0); 606 VOP_STRATEGY(vp, bp); 607 ++readwait; 608 } 609 610 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 611 if (inmem(vp, *rablkno)) 612 continue; 613 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 614 615 if ((rabp->b_flags & B_CACHE) == 0) { 616 if (curproc != PCPU_GET(idleproc)) 617 curproc->p_stats->p_ru.ru_inblock++; 618 rabp->b_flags |= B_ASYNC; 619 rabp->b_flags &= ~B_INVAL; 620 rabp->b_ioflags &= ~BIO_ERROR; 621 rabp->b_iocmd = BIO_READ; 622 if (rabp->b_rcred == NOCRED) { 623 if (cred != NOCRED) 624 crhold(cred); 625 rabp->b_rcred = cred; 626 } 627 vfs_busy_pages(rabp, 0); 628 BUF_KERNPROC(rabp); 629 VOP_STRATEGY(vp, rabp); 630 } else { 631 brelse(rabp); 632 } 633 } 634 635 if (readwait) { 636 rv = bufwait(bp); 637 } 638 return (rv); 639} 640 641/* 642 * Write, release buffer on completion. (Done by iodone 643 * if async). Do not bother writing anything if the buffer 644 * is invalid. 645 * 646 * Note that we set B_CACHE here, indicating that buffer is 647 * fully valid and thus cacheable. This is true even of NFS 648 * now so we set it generally. This could be set either here 649 * or in biodone() since the I/O is synchronous. We put it 650 * here. 651 */ 652 653int dobkgrdwrite = 1; 654SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, ""); 655 656int 657bwrite(struct buf * bp) 658{ 659 int oldflags, s; 660 struct buf *newbp; 661 662 if (bp->b_flags & B_INVAL) { 663 brelse(bp); 664 return (0); 665 } 666 667 oldflags = bp->b_flags; 668 669 if (BUF_REFCNT(bp) == 0) 670 panic("bwrite: buffer is not busy???"); 671 s = splbio(); 672 /* 673 * If a background write is already in progress, delay 674 * writing this block if it is asynchronous. Otherwise 675 * wait for the background write to complete. 676 */ 677 if (bp->b_xflags & BX_BKGRDINPROG) { 678 if (bp->b_flags & B_ASYNC) { 679 splx(s); 680 bdwrite(bp); 681 return (0); 682 } 683 bp->b_xflags |= BX_BKGRDWAIT; 684 tsleep(&bp->b_xflags, PRIBIO, "biord", 0); 685 if (bp->b_xflags & BX_BKGRDINPROG) 686 panic("bwrite: still writing"); 687 } 688 689 /* Mark the buffer clean */ 690 bundirty(bp); 691 692 /* 693 * If this buffer is marked for background writing and we 694 * do not have to wait for it, make a copy and write the 695 * copy so as to leave this buffer ready for further use. 696 * 697 * This optimization eats a lot of memory. If we have a page 698 * or buffer shortfall we can't do it. 699 */ 700 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 701 (bp->b_flags & B_ASYNC) && 702 !vm_page_count_severe() && 703 !buf_dirty_count_severe()) { 704 if (bp->b_iodone != NULL) { 705 printf("bp->b_iodone = %p\n", bp->b_iodone); 706 panic("bwrite: need chained iodone"); 707 } 708 709 /* get a new block */ 710 newbp = geteblk(bp->b_bufsize); 711 712 /* set it to be identical to the old block */ 713 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 714 bgetvp(bp->b_vp, newbp); 715 newbp->b_lblkno = bp->b_lblkno; 716 newbp->b_blkno = bp->b_blkno; 717 newbp->b_offset = bp->b_offset; 718 newbp->b_iodone = vfs_backgroundwritedone; 719 newbp->b_flags |= B_ASYNC; 720 newbp->b_flags &= ~B_INVAL; 721 722 /* move over the dependencies */ 723 if (LIST_FIRST(&bp->b_dep) != NULL) 724 buf_movedeps(bp, newbp); 725 726 /* 727 * Initiate write on the copy, release the original to 728 * the B_LOCKED queue so that it cannot go away until 729 * the background write completes. If not locked it could go 730 * away and then be reconstituted while it was being written. 731 * If the reconstituted buffer were written, we could end up 732 * with two background copies being written at the same time. 733 */ 734 bp->b_xflags |= BX_BKGRDINPROG; 735 bp->b_flags |= B_LOCKED; 736 bqrelse(bp); 737 bp = newbp; 738 } 739 740 bp->b_flags &= ~B_DONE; 741 bp->b_ioflags &= ~BIO_ERROR; 742 bp->b_flags |= B_WRITEINPROG | B_CACHE; 743 bp->b_iocmd = BIO_WRITE; 744 745 bp->b_vp->v_numoutput++; 746 vfs_busy_pages(bp, 1); 747 748 /* 749 * Normal bwrites pipeline writes 750 */ 751 bp->b_runningbufspace = bp->b_bufsize; 752 runningbufspace += bp->b_runningbufspace; 753 754 if (curproc != PCPU_GET(idleproc)) 755 curproc->p_stats->p_ru.ru_oublock++; 756 splx(s); 757 if (oldflags & B_ASYNC) 758 BUF_KERNPROC(bp); 759 BUF_STRATEGY(bp); 760 761 if ((oldflags & B_ASYNC) == 0) { 762 int rtval = bufwait(bp); 763 brelse(bp); 764 return (rtval); 765 } else { 766 /* 767 * don't allow the async write to saturate the I/O 768 * system. There is no chance of deadlock here because 769 * we are blocking on I/O that is already in-progress. 770 */ 771 waitrunningbufspace(); 772 } 773 774 return (0); 775} 776 777/* 778 * Complete a background write started from bwrite. 779 */ 780static void 781vfs_backgroundwritedone(bp) 782 struct buf *bp; 783{ 784 struct buf *origbp; 785 786 /* 787 * Find the original buffer that we are writing. 788 */ 789 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 790 panic("backgroundwritedone: lost buffer"); 791 /* 792 * Process dependencies then return any unfinished ones. 793 */ 794 if (LIST_FIRST(&bp->b_dep) != NULL) 795 buf_complete(bp); 796 if (LIST_FIRST(&bp->b_dep) != NULL) 797 buf_movedeps(bp, origbp); 798 /* 799 * Clear the BX_BKGRDINPROG flag in the original buffer 800 * and awaken it if it is waiting for the write to complete. 801 * If BX_BKGRDINPROG is not set in the original buffer it must 802 * have been released and re-instantiated - which is not legal. 803 */ 804 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), ("backgroundwritedone: lost buffer2")); 805 origbp->b_xflags &= ~BX_BKGRDINPROG; 806 if (origbp->b_xflags & BX_BKGRDWAIT) { 807 origbp->b_xflags &= ~BX_BKGRDWAIT; 808 wakeup(&origbp->b_xflags); 809 } 810 /* 811 * Clear the B_LOCKED flag and remove it from the locked 812 * queue if it currently resides there. 813 */ 814 origbp->b_flags &= ~B_LOCKED; 815 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 816 bremfree(origbp); 817 bqrelse(origbp); 818 } 819 /* 820 * This buffer is marked B_NOCACHE, so when it is released 821 * by biodone, it will be tossed. We mark it with BIO_READ 822 * to avoid biodone doing a second vwakeup. 823 */ 824 bp->b_flags |= B_NOCACHE; 825 bp->b_iocmd = BIO_READ; 826 bp->b_flags &= ~(B_CACHE | B_DONE); 827 bp->b_iodone = 0; 828 bufdone(bp); 829} 830 831/* 832 * Delayed write. (Buffer is marked dirty). Do not bother writing 833 * anything if the buffer is marked invalid. 834 * 835 * Note that since the buffer must be completely valid, we can safely 836 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 837 * biodone() in order to prevent getblk from writing the buffer 838 * out synchronously. 839 */ 840void 841bdwrite(struct buf * bp) 842{ 843 GIANT_REQUIRED; 844 845 if (BUF_REFCNT(bp) == 0) 846 panic("bdwrite: buffer is not busy"); 847 848 if (bp->b_flags & B_INVAL) { 849 brelse(bp); 850 return; 851 } 852 bdirty(bp); 853 854 /* 855 * Set B_CACHE, indicating that the buffer is fully valid. This is 856 * true even of NFS now. 857 */ 858 bp->b_flags |= B_CACHE; 859 860 /* 861 * This bmap keeps the system from needing to do the bmap later, 862 * perhaps when the system is attempting to do a sync. Since it 863 * is likely that the indirect block -- or whatever other datastructure 864 * that the filesystem needs is still in memory now, it is a good 865 * thing to do this. Note also, that if the pageout daemon is 866 * requesting a sync -- there might not be enough memory to do 867 * the bmap then... So, this is important to do. 868 */ 869 if (bp->b_lblkno == bp->b_blkno) { 870 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 871 } 872 873 /* 874 * Set the *dirty* buffer range based upon the VM system dirty pages. 875 */ 876 vfs_setdirty(bp); 877 878 /* 879 * We need to do this here to satisfy the vnode_pager and the 880 * pageout daemon, so that it thinks that the pages have been 881 * "cleaned". Note that since the pages are in a delayed write 882 * buffer -- the VFS layer "will" see that the pages get written 883 * out on the next sync, or perhaps the cluster will be completed. 884 */ 885 vfs_clean_pages(bp); 886 bqrelse(bp); 887 888 /* 889 * Wakeup the buffer flushing daemon if we have a lot of dirty 890 * buffers (midpoint between our recovery point and our stall 891 * point). 892 */ 893 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 894 895 /* 896 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 897 * due to the softdep code. 898 */ 899} 900 901/* 902 * bdirty: 903 * 904 * Turn buffer into delayed write request. We must clear BIO_READ and 905 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 906 * itself to properly update it in the dirty/clean lists. We mark it 907 * B_DONE to ensure that any asynchronization of the buffer properly 908 * clears B_DONE ( else a panic will occur later ). 909 * 910 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 911 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 912 * should only be called if the buffer is known-good. 913 * 914 * Since the buffer is not on a queue, we do not update the numfreebuffers 915 * count. 916 * 917 * Must be called at splbio(). 918 * The buffer must be on QUEUE_NONE. 919 */ 920void 921bdirty(bp) 922 struct buf *bp; 923{ 924 KASSERT(bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 925 bp->b_flags &= ~(B_RELBUF); 926 bp->b_iocmd = BIO_WRITE; 927 928 if ((bp->b_flags & B_DELWRI) == 0) { 929 bp->b_flags |= B_DONE | B_DELWRI; 930 reassignbuf(bp, bp->b_vp); 931 ++numdirtybuffers; 932 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 933 } 934} 935 936/* 937 * bundirty: 938 * 939 * Clear B_DELWRI for buffer. 940 * 941 * Since the buffer is not on a queue, we do not update the numfreebuffers 942 * count. 943 * 944 * Must be called at splbio(). 945 * The buffer must be on QUEUE_NONE. 946 */ 947 948void 949bundirty(bp) 950 struct buf *bp; 951{ 952 KASSERT(bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 953 954 if (bp->b_flags & B_DELWRI) { 955 bp->b_flags &= ~B_DELWRI; 956 reassignbuf(bp, bp->b_vp); 957 --numdirtybuffers; 958 numdirtywakeup(lodirtybuffers); 959 } 960 /* 961 * Since it is now being written, we can clear its deferred write flag. 962 */ 963 bp->b_flags &= ~B_DEFERRED; 964} 965 966/* 967 * bawrite: 968 * 969 * Asynchronous write. Start output on a buffer, but do not wait for 970 * it to complete. The buffer is released when the output completes. 971 * 972 * bwrite() ( or the VOP routine anyway ) is responsible for handling 973 * B_INVAL buffers. Not us. 974 */ 975void 976bawrite(struct buf * bp) 977{ 978 bp->b_flags |= B_ASYNC; 979 (void) BUF_WRITE(bp); 980} 981 982/* 983 * bowrite: 984 * 985 * Ordered write. Start output on a buffer, and flag it so that the 986 * device will write it in the order it was queued. The buffer is 987 * released when the output completes. bwrite() ( or the VOP routine 988 * anyway ) is responsible for handling B_INVAL buffers. 989 */ 990int 991bowrite(struct buf * bp) 992{ 993 bp->b_ioflags |= BIO_ORDERED; 994 bp->b_flags |= B_ASYNC; 995 return (BUF_WRITE(bp)); 996} 997 998/* 999 * bwillwrite: 1000 * 1001 * Called prior to the locking of any vnodes when we are expecting to 1002 * write. We do not want to starve the buffer cache with too many 1003 * dirty buffers so we block here. By blocking prior to the locking 1004 * of any vnodes we attempt to avoid the situation where a locked vnode 1005 * prevents the various system daemons from flushing related buffers. 1006 */ 1007 1008void 1009bwillwrite(void) 1010{ 1011 if (numdirtybuffers >= hidirtybuffers) { 1012 int s; 1013 1014 s = splbio(); 1015 while (numdirtybuffers >= hidirtybuffers) { 1016 bd_wakeup(1); 1017 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1018 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0); 1019 } 1020 splx(s); 1021 } 1022} 1023 1024/* 1025 * Return true if we have too many dirty buffers. 1026 */ 1027int 1028buf_dirty_count_severe(void) 1029{ 1030 return(numdirtybuffers >= hidirtybuffers); 1031} 1032 1033/* 1034 * brelse: 1035 * 1036 * Release a busy buffer and, if requested, free its resources. The 1037 * buffer will be stashed in the appropriate bufqueue[] allowing it 1038 * to be accessed later as a cache entity or reused for other purposes. 1039 */ 1040void 1041brelse(struct buf * bp) 1042{ 1043 int s; 1044 1045 GIANT_REQUIRED; 1046 1047 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1048 1049 s = splbio(); 1050 1051 if (bp->b_flags & B_LOCKED) 1052 bp->b_ioflags &= ~BIO_ERROR; 1053 1054 if (bp->b_iocmd == BIO_WRITE && 1055 (bp->b_ioflags & BIO_ERROR) && 1056 !(bp->b_flags & B_INVAL)) { 1057 /* 1058 * Failed write, redirty. Must clear BIO_ERROR to prevent 1059 * pages from being scrapped. If B_INVAL is set then 1060 * this case is not run and the next case is run to 1061 * destroy the buffer. B_INVAL can occur if the buffer 1062 * is outside the range supported by the underlying device. 1063 */ 1064 bp->b_ioflags &= ~BIO_ERROR; 1065 bdirty(bp); 1066 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1067 (bp->b_ioflags & BIO_ERROR) || 1068 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1069 /* 1070 * Either a failed I/O or we were asked to free or not 1071 * cache the buffer. 1072 */ 1073 bp->b_flags |= B_INVAL; 1074 if (LIST_FIRST(&bp->b_dep) != NULL) 1075 buf_deallocate(bp); 1076 if (bp->b_flags & B_DELWRI) { 1077 --numdirtybuffers; 1078 numdirtywakeup(lodirtybuffers); 1079 } 1080 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1081 if ((bp->b_flags & B_VMIO) == 0) { 1082 if (bp->b_bufsize) 1083 allocbuf(bp, 0); 1084 if (bp->b_vp) 1085 brelvp(bp); 1086 } 1087 } 1088 1089 /* 1090 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1091 * is called with B_DELWRI set, the underlying pages may wind up 1092 * getting freed causing a previous write (bdwrite()) to get 'lost' 1093 * because pages associated with a B_DELWRI bp are marked clean. 1094 * 1095 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1096 * if B_DELWRI is set. 1097 * 1098 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1099 * on pages to return pages to the VM page queues. 1100 */ 1101 if (bp->b_flags & B_DELWRI) 1102 bp->b_flags &= ~B_RELBUF; 1103 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1104 bp->b_flags |= B_RELBUF; 1105 1106 /* 1107 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1108 * constituted, not even NFS buffers now. Two flags effect this. If 1109 * B_INVAL, the struct buf is invalidated but the VM object is kept 1110 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1111 * 1112 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1113 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1114 * buffer is also B_INVAL because it hits the re-dirtying code above. 1115 * 1116 * Normally we can do this whether a buffer is B_DELWRI or not. If 1117 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1118 * the commit state and we cannot afford to lose the buffer. If the 1119 * buffer has a background write in progress, we need to keep it 1120 * around to prevent it from being reconstituted and starting a second 1121 * background write. 1122 */ 1123 if ((bp->b_flags & B_VMIO) 1124 && !(bp->b_vp->v_tag == VT_NFS && 1125 !vn_isdisk(bp->b_vp, NULL) && 1126 (bp->b_flags & B_DELWRI)) 1127 ) { 1128 1129 int i, j, resid; 1130 vm_page_t m; 1131 off_t foff; 1132 vm_pindex_t poff; 1133 vm_object_t obj; 1134 struct vnode *vp; 1135 1136 vp = bp->b_vp; 1137 1138 /* 1139 * Get the base offset and length of the buffer. Note that 1140 * for block sizes that are less then PAGE_SIZE, the b_data 1141 * base of the buffer does not represent exactly b_offset and 1142 * neither b_offset nor b_size are necessarily page aligned. 1143 * Instead, the starting position of b_offset is: 1144 * 1145 * b_data + (b_offset & PAGE_MASK) 1146 * 1147 * block sizes less then DEV_BSIZE (usually 512) are not 1148 * supported due to the page granularity bits (m->valid, 1149 * m->dirty, etc...). 1150 * 1151 * See man buf(9) for more information 1152 */ 1153 resid = bp->b_bufsize; 1154 foff = bp->b_offset; 1155 1156 for (i = 0; i < bp->b_npages; i++) { 1157 int had_bogus = 0; 1158 1159 m = bp->b_pages[i]; 1160 vm_page_flag_clear(m, PG_ZERO); 1161 1162 /* 1163 * If we hit a bogus page, fixup *all* the bogus pages 1164 * now. 1165 */ 1166 if (m == bogus_page) { 1167 VOP_GETVOBJECT(vp, &obj); 1168 poff = OFF_TO_IDX(bp->b_offset); 1169 had_bogus = 1; 1170 1171 for (j = i; j < bp->b_npages; j++) { 1172 vm_page_t mtmp; 1173 mtmp = bp->b_pages[j]; 1174 if (mtmp == bogus_page) { 1175 mtmp = vm_page_lookup(obj, poff + j); 1176 if (!mtmp) { 1177 panic("brelse: page missing\n"); 1178 } 1179 bp->b_pages[j] = mtmp; 1180 } 1181 } 1182 1183 if ((bp->b_flags & B_INVAL) == 0) { 1184 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1185 } 1186 m = bp->b_pages[i]; 1187 } 1188 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1189 int poffset = foff & PAGE_MASK; 1190 int presid = resid > (PAGE_SIZE - poffset) ? 1191 (PAGE_SIZE - poffset) : resid; 1192 1193 KASSERT(presid >= 0, ("brelse: extra page")); 1194 vm_page_set_invalid(m, poffset, presid); 1195 if (had_bogus) 1196 printf("avoided corruption bug in bogus_page/brelse code\n"); 1197 } 1198 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1199 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1200 } 1201 1202 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1203 vfs_vmio_release(bp); 1204 1205 } else if (bp->b_flags & B_VMIO) { 1206 1207 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1208 vfs_vmio_release(bp); 1209 } 1210 1211 } 1212 1213 if (bp->b_qindex != QUEUE_NONE) 1214 panic("brelse: free buffer onto another queue???"); 1215 if (BUF_REFCNT(bp) > 1) { 1216 /* do not release to free list */ 1217 BUF_UNLOCK(bp); 1218 splx(s); 1219 return; 1220 } 1221 1222 /* enqueue */ 1223 1224 /* buffers with no memory */ 1225 if (bp->b_bufsize == 0) { 1226 bp->b_flags |= B_INVAL; 1227 bp->b_xflags &= ~BX_BKGRDWRITE; 1228 if (bp->b_xflags & BX_BKGRDINPROG) 1229 panic("losing buffer 1"); 1230 if (bp->b_kvasize) { 1231 bp->b_qindex = QUEUE_EMPTYKVA; 1232 } else { 1233 bp->b_qindex = QUEUE_EMPTY; 1234 } 1235 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1236 LIST_REMOVE(bp, b_hash); 1237 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1238 bp->b_dev = NODEV; 1239 /* buffers with junk contents */ 1240 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || (bp->b_ioflags & BIO_ERROR)) { 1241 bp->b_flags |= B_INVAL; 1242 bp->b_xflags &= ~BX_BKGRDWRITE; 1243 if (bp->b_xflags & BX_BKGRDINPROG) 1244 panic("losing buffer 2"); 1245 bp->b_qindex = QUEUE_CLEAN; 1246 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1247 LIST_REMOVE(bp, b_hash); 1248 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1249 bp->b_dev = NODEV; 1250 1251 /* buffers that are locked */ 1252 } else if (bp->b_flags & B_LOCKED) { 1253 bp->b_qindex = QUEUE_LOCKED; 1254 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1255 1256 /* remaining buffers */ 1257 } else { 1258 if (bp->b_flags & B_DELWRI) 1259 bp->b_qindex = QUEUE_DIRTY; 1260 else 1261 bp->b_qindex = QUEUE_CLEAN; 1262 if (bp->b_flags & B_AGE) 1263 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1264 else 1265 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1266 } 1267 1268 /* 1269 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1270 * on the correct queue. 1271 */ 1272 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) { 1273 bp->b_flags &= ~B_DELWRI; 1274 --numdirtybuffers; 1275 numdirtywakeup(lodirtybuffers); 1276 } 1277 1278 /* 1279 * Fixup numfreebuffers count. The bp is on an appropriate queue 1280 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1281 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1282 * if B_INVAL is set ). 1283 */ 1284 1285 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1286 bufcountwakeup(); 1287 1288 /* 1289 * Something we can maybe free or reuse 1290 */ 1291 if (bp->b_bufsize || bp->b_kvasize) 1292 bufspacewakeup(); 1293 1294 /* unlock */ 1295 BUF_UNLOCK(bp); 1296 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1297 bp->b_ioflags &= ~BIO_ORDERED; 1298 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1299 panic("brelse: not dirty"); 1300 splx(s); 1301} 1302 1303/* 1304 * Release a buffer back to the appropriate queue but do not try to free 1305 * it. The buffer is expected to be used again soon. 1306 * 1307 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1308 * biodone() to requeue an async I/O on completion. It is also used when 1309 * known good buffers need to be requeued but we think we may need the data 1310 * again soon. 1311 * 1312 * XXX we should be able to leave the B_RELBUF hint set on completion. 1313 */ 1314void 1315bqrelse(struct buf * bp) 1316{ 1317 int s; 1318 1319 s = splbio(); 1320 1321 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1322 1323 if (bp->b_qindex != QUEUE_NONE) 1324 panic("bqrelse: free buffer onto another queue???"); 1325 if (BUF_REFCNT(bp) > 1) { 1326 /* do not release to free list */ 1327 BUF_UNLOCK(bp); 1328 splx(s); 1329 return; 1330 } 1331 if (bp->b_flags & B_LOCKED) { 1332 bp->b_ioflags &= ~BIO_ERROR; 1333 bp->b_qindex = QUEUE_LOCKED; 1334 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1335 /* buffers with stale but valid contents */ 1336 } else if (bp->b_flags & B_DELWRI) { 1337 bp->b_qindex = QUEUE_DIRTY; 1338 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1339 } else if (vm_page_count_severe()) { 1340 /* 1341 * We are too low on memory, we have to try to free the 1342 * buffer (most importantly: the wired pages making up its 1343 * backing store) *now*. 1344 */ 1345 splx(s); 1346 brelse(bp); 1347 return; 1348 } else { 1349 bp->b_qindex = QUEUE_CLEAN; 1350 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1351 } 1352 1353 if ((bp->b_flags & B_LOCKED) == 0 && 1354 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1355 bufcountwakeup(); 1356 } 1357 1358 /* 1359 * Something we can maybe free or reuse. 1360 */ 1361 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1362 bufspacewakeup(); 1363 1364 /* unlock */ 1365 BUF_UNLOCK(bp); 1366 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1367 bp->b_ioflags &= ~BIO_ORDERED; 1368 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1369 panic("bqrelse: not dirty"); 1370 splx(s); 1371} 1372 1373static void 1374vfs_vmio_release(bp) 1375 struct buf *bp; 1376{ 1377 int i; 1378 vm_page_t m; 1379 1380 GIANT_REQUIRED; 1381 1382 for (i = 0; i < bp->b_npages; i++) { 1383 m = bp->b_pages[i]; 1384 bp->b_pages[i] = NULL; 1385 /* 1386 * In order to keep page LRU ordering consistent, put 1387 * everything on the inactive queue. 1388 */ 1389 vm_page_unwire(m, 0); 1390 /* 1391 * We don't mess with busy pages, it is 1392 * the responsibility of the process that 1393 * busied the pages to deal with them. 1394 */ 1395 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1396 continue; 1397 1398 if (m->wire_count == 0) { 1399 vm_page_flag_clear(m, PG_ZERO); 1400 /* 1401 * Might as well free the page if we can and it has 1402 * no valid data. We also free the page if the 1403 * buffer was used for direct I/O 1404 */ 1405 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { 1406 vm_page_busy(m); 1407 vm_page_protect(m, VM_PROT_NONE); 1408 vm_page_free(m); 1409 } else if (bp->b_flags & B_DIRECT) { 1410 vm_page_try_to_free(m); 1411 } else if (vm_page_count_severe()) { 1412 vm_page_try_to_cache(m); 1413 } 1414 } 1415 } 1416 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1417 1418 if (bp->b_bufsize) { 1419 bufspacewakeup(); 1420 bp->b_bufsize = 0; 1421 } 1422 bp->b_npages = 0; 1423 bp->b_flags &= ~B_VMIO; 1424 if (bp->b_vp) 1425 brelvp(bp); 1426} 1427 1428/* 1429 * Check to see if a block is currently memory resident. 1430 */ 1431struct buf * 1432gbincore(struct vnode * vp, daddr_t blkno) 1433{ 1434 struct buf *bp; 1435 struct bufhashhdr *bh; 1436 1437 bh = bufhash(vp, blkno); 1438 1439 /* Search hash chain */ 1440 LIST_FOREACH(bp, bh, b_hash) { 1441 /* hit */ 1442 if (bp->b_vp == vp && bp->b_lblkno == blkno && 1443 (bp->b_flags & B_INVAL) == 0) { 1444 break; 1445 } 1446 } 1447 return (bp); 1448} 1449 1450/* 1451 * vfs_bio_awrite: 1452 * 1453 * Implement clustered async writes for clearing out B_DELWRI buffers. 1454 * This is much better then the old way of writing only one buffer at 1455 * a time. Note that we may not be presented with the buffers in the 1456 * correct order, so we search for the cluster in both directions. 1457 */ 1458int 1459vfs_bio_awrite(struct buf * bp) 1460{ 1461 int i; 1462 int j; 1463 daddr_t lblkno = bp->b_lblkno; 1464 struct vnode *vp = bp->b_vp; 1465 int s; 1466 int ncl; 1467 struct buf *bpa; 1468 int nwritten; 1469 int size; 1470 int maxcl; 1471 1472 s = splbio(); 1473 /* 1474 * right now we support clustered writing only to regular files. If 1475 * we find a clusterable block we could be in the middle of a cluster 1476 * rather then at the beginning. 1477 */ 1478 if ((vp->v_type == VREG) && 1479 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1480 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1481 1482 size = vp->v_mount->mnt_stat.f_iosize; 1483 maxcl = MAXPHYS / size; 1484 1485 for (i = 1; i < maxcl; i++) { 1486 if ((bpa = gbincore(vp, lblkno + i)) && 1487 BUF_REFCNT(bpa) == 0 && 1488 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1489 (B_DELWRI | B_CLUSTEROK)) && 1490 (bpa->b_bufsize == size)) { 1491 if ((bpa->b_blkno == bpa->b_lblkno) || 1492 (bpa->b_blkno != 1493 bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 1494 break; 1495 } else { 1496 break; 1497 } 1498 } 1499 for (j = 1; i + j <= maxcl && j <= lblkno; j++) { 1500 if ((bpa = gbincore(vp, lblkno - j)) && 1501 BUF_REFCNT(bpa) == 0 && 1502 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1503 (B_DELWRI | B_CLUSTEROK)) && 1504 (bpa->b_bufsize == size)) { 1505 if ((bpa->b_blkno == bpa->b_lblkno) || 1506 (bpa->b_blkno != 1507 bp->b_blkno - ((j * size) >> DEV_BSHIFT))) 1508 break; 1509 } else { 1510 break; 1511 } 1512 } 1513 --j; 1514 ncl = i + j; 1515 /* 1516 * this is a possible cluster write 1517 */ 1518 if (ncl != 1) { 1519 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1520 splx(s); 1521 return nwritten; 1522 } 1523 } 1524 1525 BUF_LOCK(bp, LK_EXCLUSIVE); 1526 bremfree(bp); 1527 bp->b_flags |= B_ASYNC; 1528 1529 splx(s); 1530 /* 1531 * default (old) behavior, writing out only one block 1532 * 1533 * XXX returns b_bufsize instead of b_bcount for nwritten? 1534 */ 1535 nwritten = bp->b_bufsize; 1536 (void) BUF_WRITE(bp); 1537 1538 return nwritten; 1539} 1540 1541/* 1542 * getnewbuf: 1543 * 1544 * Find and initialize a new buffer header, freeing up existing buffers 1545 * in the bufqueues as necessary. The new buffer is returned locked. 1546 * 1547 * Important: B_INVAL is not set. If the caller wishes to throw the 1548 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1549 * 1550 * We block if: 1551 * We have insufficient buffer headers 1552 * We have insufficient buffer space 1553 * buffer_map is too fragmented ( space reservation fails ) 1554 * If we have to flush dirty buffers ( but we try to avoid this ) 1555 * 1556 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1557 * Instead we ask the buf daemon to do it for us. We attempt to 1558 * avoid piecemeal wakeups of the pageout daemon. 1559 */ 1560 1561static struct buf * 1562getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1563{ 1564 struct buf *bp; 1565 struct buf *nbp; 1566 int defrag = 0; 1567 int nqindex; 1568 static int flushingbufs; 1569 1570 GIANT_REQUIRED; 1571 1572 /* 1573 * We can't afford to block since we might be holding a vnode lock, 1574 * which may prevent system daemons from running. We deal with 1575 * low-memory situations by proactively returning memory and running 1576 * async I/O rather then sync I/O. 1577 */ 1578 1579 ++getnewbufcalls; 1580 --getnewbufrestarts; 1581restart: 1582 ++getnewbufrestarts; 1583 1584 /* 1585 * Setup for scan. If we do not have enough free buffers, 1586 * we setup a degenerate case that immediately fails. Note 1587 * that if we are specially marked process, we are allowed to 1588 * dip into our reserves. 1589 * 1590 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1591 * 1592 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1593 * However, there are a number of cases (defragging, reusing, ...) 1594 * where we cannot backup. 1595 */ 1596 nqindex = QUEUE_EMPTYKVA; 1597 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1598 1599 if (nbp == NULL) { 1600 /* 1601 * If no EMPTYKVA buffers and we are either 1602 * defragging or reusing, locate a CLEAN buffer 1603 * to free or reuse. If bufspace useage is low 1604 * skip this step so we can allocate a new buffer. 1605 */ 1606 if (defrag || bufspace >= lobufspace) { 1607 nqindex = QUEUE_CLEAN; 1608 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1609 } 1610 1611 /* 1612 * If we could not find or were not allowed to reuse a 1613 * CLEAN buffer, check to see if it is ok to use an EMPTY 1614 * buffer. We can only use an EMPTY buffer if allocating 1615 * its KVA would not otherwise run us out of buffer space. 1616 */ 1617 if (nbp == NULL && defrag == 0 && 1618 bufspace + maxsize < hibufspace) { 1619 nqindex = QUEUE_EMPTY; 1620 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1621 } 1622 } 1623 1624 /* 1625 * Run scan, possibly freeing data and/or kva mappings on the fly 1626 * depending. 1627 */ 1628 1629 while ((bp = nbp) != NULL) { 1630 int qindex = nqindex; 1631 1632 /* 1633 * Calculate next bp ( we can only use it if we do not block 1634 * or do other fancy things ). 1635 */ 1636 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1637 switch(qindex) { 1638 case QUEUE_EMPTY: 1639 nqindex = QUEUE_EMPTYKVA; 1640 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1641 break; 1642 /* fall through */ 1643 case QUEUE_EMPTYKVA: 1644 nqindex = QUEUE_CLEAN; 1645 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1646 break; 1647 /* fall through */ 1648 case QUEUE_CLEAN: 1649 /* 1650 * nbp is NULL. 1651 */ 1652 break; 1653 } 1654 } 1655 1656 /* 1657 * Sanity Checks 1658 */ 1659 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1660 1661 /* 1662 * Note: we no longer distinguish between VMIO and non-VMIO 1663 * buffers. 1664 */ 1665 1666 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1667 1668 /* 1669 * If we are defragging then we need a buffer with 1670 * b_kvasize != 0. XXX this situation should no longer 1671 * occur, if defrag is non-zero the buffer's b_kvasize 1672 * should also be non-zero at this point. XXX 1673 */ 1674 if (defrag && bp->b_kvasize == 0) { 1675 printf("Warning: defrag empty buffer %p\n", bp); 1676 continue; 1677 } 1678 1679 /* 1680 * Start freeing the bp. This is somewhat involved. nbp 1681 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1682 */ 1683 1684 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1685 panic("getnewbuf: locked buf"); 1686 bremfree(bp); 1687 1688 if (qindex == QUEUE_CLEAN) { 1689 if (bp->b_flags & B_VMIO) { 1690 bp->b_flags &= ~B_ASYNC; 1691 vfs_vmio_release(bp); 1692 } 1693 if (bp->b_vp) 1694 brelvp(bp); 1695 } 1696 1697 /* 1698 * NOTE: nbp is now entirely invalid. We can only restart 1699 * the scan from this point on. 1700 * 1701 * Get the rest of the buffer freed up. b_kva* is still 1702 * valid after this operation. 1703 */ 1704 1705 if (bp->b_rcred != NOCRED) { 1706 crfree(bp->b_rcred); 1707 bp->b_rcred = NOCRED; 1708 } 1709 if (bp->b_wcred != NOCRED) { 1710 crfree(bp->b_wcred); 1711 bp->b_wcred = NOCRED; 1712 } 1713 if (LIST_FIRST(&bp->b_dep) != NULL) 1714 buf_deallocate(bp); 1715 if (bp->b_xflags & BX_BKGRDINPROG) 1716 panic("losing buffer 3"); 1717 LIST_REMOVE(bp, b_hash); 1718 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1719 1720 if (bp->b_bufsize) 1721 allocbuf(bp, 0); 1722 1723 bp->b_flags = 0; 1724 bp->b_ioflags = 0; 1725 bp->b_xflags = 0; 1726 bp->b_dev = NODEV; 1727 bp->b_vp = NULL; 1728 bp->b_blkno = bp->b_lblkno = 0; 1729 bp->b_offset = NOOFFSET; 1730 bp->b_iodone = 0; 1731 bp->b_error = 0; 1732 bp->b_resid = 0; 1733 bp->b_bcount = 0; 1734 bp->b_npages = 0; 1735 bp->b_dirtyoff = bp->b_dirtyend = 0; 1736 bp->b_magic = B_MAGIC_BIO; 1737 bp->b_op = &buf_ops_bio; 1738 1739 LIST_INIT(&bp->b_dep); 1740 1741 /* 1742 * If we are defragging then free the buffer. 1743 */ 1744 if (defrag) { 1745 bp->b_flags |= B_INVAL; 1746 bfreekva(bp); 1747 brelse(bp); 1748 defrag = 0; 1749 goto restart; 1750 } 1751 1752 /* 1753 * If we are overcomitted then recover the buffer and its 1754 * KVM space. This occurs in rare situations when multiple 1755 * processes are blocked in getnewbuf() or allocbuf(). 1756 */ 1757 if (bufspace >= hibufspace) 1758 flushingbufs = 1; 1759 if (flushingbufs && bp->b_kvasize != 0) { 1760 bp->b_flags |= B_INVAL; 1761 bfreekva(bp); 1762 brelse(bp); 1763 goto restart; 1764 } 1765 if (bufspace < lobufspace) 1766 flushingbufs = 0; 1767 break; 1768 } 1769 1770 /* 1771 * If we exhausted our list, sleep as appropriate. We may have to 1772 * wakeup various daemons and write out some dirty buffers. 1773 * 1774 * Generally we are sleeping due to insufficient buffer space. 1775 */ 1776 1777 if (bp == NULL) { 1778 int flags; 1779 char *waitmsg; 1780 1781 if (defrag) { 1782 flags = VFS_BIO_NEED_BUFSPACE; 1783 waitmsg = "nbufkv"; 1784 } else if (bufspace >= hibufspace) { 1785 waitmsg = "nbufbs"; 1786 flags = VFS_BIO_NEED_BUFSPACE; 1787 } else { 1788 waitmsg = "newbuf"; 1789 flags = VFS_BIO_NEED_ANY; 1790 } 1791 1792 bd_speedup(); /* heeeelp */ 1793 1794 needsbuffer |= flags; 1795 while (needsbuffer & flags) { 1796 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, 1797 waitmsg, slptimeo)) 1798 return (NULL); 1799 } 1800 } else { 1801 /* 1802 * We finally have a valid bp. We aren't quite out of the 1803 * woods, we still have to reserve kva space. In order 1804 * to keep fragmentation sane we only allocate kva in 1805 * BKVASIZE chunks. 1806 */ 1807 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1808 1809 if (maxsize != bp->b_kvasize) { 1810 vm_offset_t addr = 0; 1811 1812 bfreekva(bp); 1813 1814 if (vm_map_findspace(buffer_map, 1815 vm_map_min(buffer_map), maxsize, &addr)) { 1816 /* 1817 * Uh oh. Buffer map is to fragmented. We 1818 * must defragment the map. 1819 */ 1820 ++bufdefragcnt; 1821 defrag = 1; 1822 bp->b_flags |= B_INVAL; 1823 brelse(bp); 1824 goto restart; 1825 } 1826 if (addr) { 1827 vm_map_insert(buffer_map, NULL, 0, 1828 addr, addr + maxsize, 1829 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1830 1831 bp->b_kvabase = (caddr_t) addr; 1832 bp->b_kvasize = maxsize; 1833 bufspace += bp->b_kvasize; 1834 ++bufreusecnt; 1835 } 1836 } 1837 bp->b_data = bp->b_kvabase; 1838 } 1839 return(bp); 1840} 1841 1842/* 1843 * buf_daemon: 1844 * 1845 * buffer flushing daemon. Buffers are normally flushed by the 1846 * update daemon but if it cannot keep up this process starts to 1847 * take the load in an attempt to prevent getnewbuf() from blocking. 1848 */ 1849 1850static struct proc *bufdaemonproc; 1851 1852static struct kproc_desc buf_kp = { 1853 "bufdaemon", 1854 buf_daemon, 1855 &bufdaemonproc 1856}; 1857SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1858 1859static void 1860buf_daemon() 1861{ 1862 int s; 1863 1864 mtx_lock(&Giant); 1865 1866 /* 1867 * This process needs to be suspended prior to shutdown sync. 1868 */ 1869 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 1870 SHUTDOWN_PRI_LAST); 1871 1872 /* 1873 * This process is allowed to take the buffer cache to the limit 1874 */ 1875 curproc->p_flag |= P_BUFEXHAUST; 1876 s = splbio(); 1877 1878 for (;;) { 1879 kthread_suspend_check(bufdaemonproc); 1880 1881 bd_request = 0; 1882 1883 /* 1884 * Do the flush. Limit the amount of in-transit I/O we 1885 * allow to build up, otherwise we would completely saturate 1886 * the I/O system. Wakeup any waiting processes before we 1887 * normally would so they can run in parallel with our drain. 1888 */ 1889 while (numdirtybuffers > lodirtybuffers) { 1890 if (flushbufqueues() == 0) 1891 break; 1892 waitrunningbufspace(); 1893 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 1894 } 1895 1896 /* 1897 * Only clear bd_request if we have reached our low water 1898 * mark. The buf_daemon normally waits 5 seconds and 1899 * then incrementally flushes any dirty buffers that have 1900 * built up, within reason. 1901 * 1902 * If we were unable to hit our low water mark and couldn't 1903 * find any flushable buffers, we sleep half a second. 1904 * Otherwise we loop immediately. 1905 */ 1906 if (numdirtybuffers <= lodirtybuffers) { 1907 /* 1908 * We reached our low water mark, reset the 1909 * request and sleep until we are needed again. 1910 * The sleep is just so the suspend code works. 1911 */ 1912 bd_request = 0; 1913 tsleep(&bd_request, PVM, "psleep", hz); 1914 } else { 1915 /* 1916 * We couldn't find any flushable dirty buffers but 1917 * still have too many dirty buffers, we 1918 * have to sleep and try again. (rare) 1919 */ 1920 tsleep(&bd_request, PVM, "qsleep", hz / 2); 1921 } 1922 } 1923} 1924 1925/* 1926 * flushbufqueues: 1927 * 1928 * Try to flush a buffer in the dirty queue. We must be careful to 1929 * free up B_INVAL buffers instead of write them, which NFS is 1930 * particularly sensitive to. 1931 */ 1932 1933static int 1934flushbufqueues(void) 1935{ 1936 struct buf *bp; 1937 int r = 0; 1938 1939 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1940 1941 while (bp) { 1942 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); 1943 if ((bp->b_flags & B_DELWRI) != 0 && 1944 (bp->b_xflags & BX_BKGRDINPROG) == 0) { 1945 if (bp->b_flags & B_INVAL) { 1946 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1947 panic("flushbufqueues: locked buf"); 1948 bremfree(bp); 1949 brelse(bp); 1950 ++r; 1951 break; 1952 } 1953 if (LIST_FIRST(&bp->b_dep) != NULL && 1954 (bp->b_flags & B_DEFERRED) == 0 && 1955 buf_countdeps(bp, 0)) { 1956 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], 1957 bp, b_freelist); 1958 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], 1959 bp, b_freelist); 1960 bp->b_flags |= B_DEFERRED; 1961 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1962 continue; 1963 } 1964 vfs_bio_awrite(bp); 1965 ++r; 1966 break; 1967 } 1968 bp = TAILQ_NEXT(bp, b_freelist); 1969 } 1970 return (r); 1971} 1972 1973/* 1974 * Check to see if a block is currently memory resident. 1975 */ 1976struct buf * 1977incore(struct vnode * vp, daddr_t blkno) 1978{ 1979 struct buf *bp; 1980 1981 int s = splbio(); 1982 bp = gbincore(vp, blkno); 1983 splx(s); 1984 return (bp); 1985} 1986 1987/* 1988 * Returns true if no I/O is needed to access the 1989 * associated VM object. This is like incore except 1990 * it also hunts around in the VM system for the data. 1991 */ 1992 1993int 1994inmem(struct vnode * vp, daddr_t blkno) 1995{ 1996 vm_object_t obj; 1997 vm_offset_t toff, tinc, size; 1998 vm_page_t m; 1999 vm_ooffset_t off; 2000 2001 GIANT_REQUIRED; 2002 2003 if (incore(vp, blkno)) 2004 return 1; 2005 if (vp->v_mount == NULL) 2006 return 0; 2007 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_flag & VOBJBUF) == 0) 2008 return 0; 2009 2010 size = PAGE_SIZE; 2011 if (size > vp->v_mount->mnt_stat.f_iosize) 2012 size = vp->v_mount->mnt_stat.f_iosize; 2013 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2014 2015 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2016 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2017 if (!m) 2018 goto notinmem; 2019 tinc = size; 2020 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2021 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2022 if (vm_page_is_valid(m, 2023 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2024 goto notinmem; 2025 } 2026 return 1; 2027 2028notinmem: 2029 return (0); 2030} 2031 2032/* 2033 * vfs_setdirty: 2034 * 2035 * Sets the dirty range for a buffer based on the status of the dirty 2036 * bits in the pages comprising the buffer. 2037 * 2038 * The range is limited to the size of the buffer. 2039 * 2040 * This routine is primarily used by NFS, but is generalized for the 2041 * B_VMIO case. 2042 */ 2043static void 2044vfs_setdirty(struct buf *bp) 2045{ 2046 int i; 2047 vm_object_t object; 2048 2049 GIANT_REQUIRED; 2050 /* 2051 * Degenerate case - empty buffer 2052 */ 2053 2054 if (bp->b_bufsize == 0) 2055 return; 2056 2057 /* 2058 * We qualify the scan for modified pages on whether the 2059 * object has been flushed yet. The OBJ_WRITEABLE flag 2060 * is not cleared simply by protecting pages off. 2061 */ 2062 2063 if ((bp->b_flags & B_VMIO) == 0) 2064 return; 2065 2066 object = bp->b_pages[0]->object; 2067 2068 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2069 printf("Warning: object %p writeable but not mightbedirty\n", object); 2070 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2071 printf("Warning: object %p mightbedirty but not writeable\n", object); 2072 2073 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2074 vm_offset_t boffset; 2075 vm_offset_t eoffset; 2076 2077 /* 2078 * test the pages to see if they have been modified directly 2079 * by users through the VM system. 2080 */ 2081 for (i = 0; i < bp->b_npages; i++) { 2082 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2083 vm_page_test_dirty(bp->b_pages[i]); 2084 } 2085 2086 /* 2087 * Calculate the encompassing dirty range, boffset and eoffset, 2088 * (eoffset - boffset) bytes. 2089 */ 2090 2091 for (i = 0; i < bp->b_npages; i++) { 2092 if (bp->b_pages[i]->dirty) 2093 break; 2094 } 2095 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2096 2097 for (i = bp->b_npages - 1; i >= 0; --i) { 2098 if (bp->b_pages[i]->dirty) { 2099 break; 2100 } 2101 } 2102 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2103 2104 /* 2105 * Fit it to the buffer. 2106 */ 2107 2108 if (eoffset > bp->b_bcount) 2109 eoffset = bp->b_bcount; 2110 2111 /* 2112 * If we have a good dirty range, merge with the existing 2113 * dirty range. 2114 */ 2115 2116 if (boffset < eoffset) { 2117 if (bp->b_dirtyoff > boffset) 2118 bp->b_dirtyoff = boffset; 2119 if (bp->b_dirtyend < eoffset) 2120 bp->b_dirtyend = eoffset; 2121 } 2122 } 2123} 2124 2125/* 2126 * getblk: 2127 * 2128 * Get a block given a specified block and offset into a file/device. 2129 * The buffers B_DONE bit will be cleared on return, making it almost 2130 * ready for an I/O initiation. B_INVAL may or may not be set on 2131 * return. The caller should clear B_INVAL prior to initiating a 2132 * READ. 2133 * 2134 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2135 * an existing buffer. 2136 * 2137 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2138 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2139 * and then cleared based on the backing VM. If the previous buffer is 2140 * non-0-sized but invalid, B_CACHE will be cleared. 2141 * 2142 * If getblk() must create a new buffer, the new buffer is returned with 2143 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2144 * case it is returned with B_INVAL clear and B_CACHE set based on the 2145 * backing VM. 2146 * 2147 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2148 * B_CACHE bit is clear. 2149 * 2150 * What this means, basically, is that the caller should use B_CACHE to 2151 * determine whether the buffer is fully valid or not and should clear 2152 * B_INVAL prior to issuing a read. If the caller intends to validate 2153 * the buffer by loading its data area with something, the caller needs 2154 * to clear B_INVAL. If the caller does this without issuing an I/O, 2155 * the caller should set B_CACHE ( as an optimization ), else the caller 2156 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2157 * a write attempt or if it was a successfull read. If the caller 2158 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2159 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2160 */ 2161struct buf * 2162getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2163{ 2164 struct buf *bp; 2165 int s; 2166 struct bufhashhdr *bh; 2167 2168 if (size > MAXBSIZE) 2169 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2170 2171 s = splbio(); 2172loop: 2173 /* 2174 * Block if we are low on buffers. Certain processes are allowed 2175 * to completely exhaust the buffer cache. 2176 * 2177 * If this check ever becomes a bottleneck it may be better to 2178 * move it into the else, when gbincore() fails. At the moment 2179 * it isn't a problem. 2180 * 2181 * XXX remove if 0 sections (clean this up after its proven) 2182 */ 2183 if (numfreebuffers == 0) { 2184 if (curproc == PCPU_GET(idleproc)) 2185 return NULL; 2186 needsbuffer |= VFS_BIO_NEED_ANY; 2187 } 2188 2189 if ((bp = gbincore(vp, blkno))) { 2190 /* 2191 * Buffer is in-core. If the buffer is not busy, it must 2192 * be on a queue. 2193 */ 2194 2195 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2196 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, 2197 "getblk", slpflag, slptimeo) == ENOLCK) 2198 goto loop; 2199 splx(s); 2200 return (struct buf *) NULL; 2201 } 2202 2203 /* 2204 * The buffer is locked. B_CACHE is cleared if the buffer is 2205 * invalid. Ohterwise, for a non-VMIO buffer, B_CACHE is set 2206 * and for a VMIO buffer B_CACHE is adjusted according to the 2207 * backing VM cache. 2208 */ 2209 if (bp->b_flags & B_INVAL) 2210 bp->b_flags &= ~B_CACHE; 2211 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2212 bp->b_flags |= B_CACHE; 2213 bremfree(bp); 2214 2215 /* 2216 * check for size inconsistancies for non-VMIO case. 2217 */ 2218 2219 if (bp->b_bcount != size) { 2220 if ((bp->b_flags & B_VMIO) == 0 || 2221 (size > bp->b_kvasize)) { 2222 if (bp->b_flags & B_DELWRI) { 2223 bp->b_flags |= B_NOCACHE; 2224 BUF_WRITE(bp); 2225 } else { 2226 if ((bp->b_flags & B_VMIO) && 2227 (LIST_FIRST(&bp->b_dep) == NULL)) { 2228 bp->b_flags |= B_RELBUF; 2229 brelse(bp); 2230 } else { 2231 bp->b_flags |= B_NOCACHE; 2232 BUF_WRITE(bp); 2233 } 2234 } 2235 goto loop; 2236 } 2237 } 2238 2239 /* 2240 * If the size is inconsistant in the VMIO case, we can resize 2241 * the buffer. This might lead to B_CACHE getting set or 2242 * cleared. If the size has not changed, B_CACHE remains 2243 * unchanged from its previous state. 2244 */ 2245 2246 if (bp->b_bcount != size) 2247 allocbuf(bp, size); 2248 2249 KASSERT(bp->b_offset != NOOFFSET, 2250 ("getblk: no buffer offset")); 2251 2252 /* 2253 * A buffer with B_DELWRI set and B_CACHE clear must 2254 * be committed before we can return the buffer in 2255 * order to prevent the caller from issuing a read 2256 * ( due to B_CACHE not being set ) and overwriting 2257 * it. 2258 * 2259 * Most callers, including NFS and FFS, need this to 2260 * operate properly either because they assume they 2261 * can issue a read if B_CACHE is not set, or because 2262 * ( for example ) an uncached B_DELWRI might loop due 2263 * to softupdates re-dirtying the buffer. In the latter 2264 * case, B_CACHE is set after the first write completes, 2265 * preventing further loops. 2266 */ 2267 2268 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2269 BUF_WRITE(bp); 2270 goto loop; 2271 } 2272 2273 splx(s); 2274 bp->b_flags &= ~B_DONE; 2275 } else { 2276 /* 2277 * Buffer is not in-core, create new buffer. The buffer 2278 * returned by getnewbuf() is locked. Note that the returned 2279 * buffer is also considered valid (not marked B_INVAL). 2280 */ 2281 int bsize, maxsize, vmio; 2282 off_t offset; 2283 2284 if (vn_isdisk(vp, NULL)) 2285 bsize = DEV_BSIZE; 2286 else if (vp->v_mountedhere) 2287 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2288 else if (vp->v_mount) 2289 bsize = vp->v_mount->mnt_stat.f_iosize; 2290 else 2291 bsize = size; 2292 2293 offset = (off_t)blkno * bsize; 2294 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && (vp->v_flag & VOBJBUF); 2295 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2296 maxsize = imax(maxsize, bsize); 2297 2298 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2299 if (slpflag || slptimeo) { 2300 splx(s); 2301 return NULL; 2302 } 2303 goto loop; 2304 } 2305 2306 /* 2307 * This code is used to make sure that a buffer is not 2308 * created while the getnewbuf routine is blocked. 2309 * This can be a problem whether the vnode is locked or not. 2310 * If the buffer is created out from under us, we have to 2311 * throw away the one we just created. There is now window 2312 * race because we are safely running at splbio() from the 2313 * point of the duplicate buffer creation through to here, 2314 * and we've locked the buffer. 2315 */ 2316 if (gbincore(vp, blkno)) { 2317 bp->b_flags |= B_INVAL; 2318 brelse(bp); 2319 goto loop; 2320 } 2321 2322 /* 2323 * Insert the buffer into the hash, so that it can 2324 * be found by incore. 2325 */ 2326 bp->b_blkno = bp->b_lblkno = blkno; 2327 bp->b_offset = offset; 2328 2329 bgetvp(vp, bp); 2330 LIST_REMOVE(bp, b_hash); 2331 bh = bufhash(vp, blkno); 2332 LIST_INSERT_HEAD(bh, bp, b_hash); 2333 2334 /* 2335 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2336 * buffer size starts out as 0, B_CACHE will be set by 2337 * allocbuf() for the VMIO case prior to it testing the 2338 * backing store for validity. 2339 */ 2340 2341 if (vmio) { 2342 bp->b_flags |= B_VMIO; 2343#if defined(VFS_BIO_DEBUG) 2344 if (vp->v_type != VREG) 2345 printf("getblk: vmioing file type %d???\n", vp->v_type); 2346#endif 2347 } else { 2348 bp->b_flags &= ~B_VMIO; 2349 } 2350 2351 allocbuf(bp, size); 2352 2353 splx(s); 2354 bp->b_flags &= ~B_DONE; 2355 } 2356 return (bp); 2357} 2358 2359/* 2360 * Get an empty, disassociated buffer of given size. The buffer is initially 2361 * set to B_INVAL. 2362 */ 2363struct buf * 2364geteblk(int size) 2365{ 2366 struct buf *bp; 2367 int s; 2368 int maxsize; 2369 2370 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2371 2372 s = splbio(); 2373 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0); 2374 splx(s); 2375 allocbuf(bp, size); 2376 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2377 return (bp); 2378} 2379 2380 2381/* 2382 * This code constitutes the buffer memory from either anonymous system 2383 * memory (in the case of non-VMIO operations) or from an associated 2384 * VM object (in the case of VMIO operations). This code is able to 2385 * resize a buffer up or down. 2386 * 2387 * Note that this code is tricky, and has many complications to resolve 2388 * deadlock or inconsistant data situations. Tread lightly!!! 2389 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2390 * the caller. Calling this code willy nilly can result in the loss of data. 2391 * 2392 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2393 * B_CACHE for the non-VMIO case. 2394 */ 2395 2396int 2397allocbuf(struct buf *bp, int size) 2398{ 2399 int newbsize, mbsize; 2400 int i; 2401 2402 GIANT_REQUIRED; 2403 2404 if (BUF_REFCNT(bp) == 0) 2405 panic("allocbuf: buffer not busy"); 2406 2407 if (bp->b_kvasize < size) 2408 panic("allocbuf: buffer too small"); 2409 2410 if ((bp->b_flags & B_VMIO) == 0) { 2411 caddr_t origbuf; 2412 int origbufsize; 2413 /* 2414 * Just get anonymous memory from the kernel. Don't 2415 * mess with B_CACHE. 2416 */ 2417 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2418#if !defined(NO_B_MALLOC) 2419 if (bp->b_flags & B_MALLOC) 2420 newbsize = mbsize; 2421 else 2422#endif 2423 newbsize = round_page(size); 2424 2425 if (newbsize < bp->b_bufsize) { 2426#if !defined(NO_B_MALLOC) 2427 /* 2428 * malloced buffers are not shrunk 2429 */ 2430 if (bp->b_flags & B_MALLOC) { 2431 if (newbsize) { 2432 bp->b_bcount = size; 2433 } else { 2434 free(bp->b_data, M_BIOBUF); 2435 if (bp->b_bufsize) { 2436 bufmallocspace -= bp->b_bufsize; 2437 bufspacewakeup(); 2438 bp->b_bufsize = 0; 2439 } 2440 bp->b_data = bp->b_kvabase; 2441 bp->b_bcount = 0; 2442 bp->b_flags &= ~B_MALLOC; 2443 } 2444 return 1; 2445 } 2446#endif 2447 vm_hold_free_pages( 2448 bp, 2449 (vm_offset_t) bp->b_data + newbsize, 2450 (vm_offset_t) bp->b_data + bp->b_bufsize); 2451 } else if (newbsize > bp->b_bufsize) { 2452#if !defined(NO_B_MALLOC) 2453 /* 2454 * We only use malloced memory on the first allocation. 2455 * and revert to page-allocated memory when the buffer 2456 * grows. 2457 */ 2458 if ( (bufmallocspace < maxbufmallocspace) && 2459 (bp->b_bufsize == 0) && 2460 (mbsize <= PAGE_SIZE/2)) { 2461 2462 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2463 bp->b_bufsize = mbsize; 2464 bp->b_bcount = size; 2465 bp->b_flags |= B_MALLOC; 2466 bufmallocspace += mbsize; 2467 return 1; 2468 } 2469#endif 2470 origbuf = NULL; 2471 origbufsize = 0; 2472#if !defined(NO_B_MALLOC) 2473 /* 2474 * If the buffer is growing on its other-than-first allocation, 2475 * then we revert to the page-allocation scheme. 2476 */ 2477 if (bp->b_flags & B_MALLOC) { 2478 origbuf = bp->b_data; 2479 origbufsize = bp->b_bufsize; 2480 bp->b_data = bp->b_kvabase; 2481 if (bp->b_bufsize) { 2482 bufmallocspace -= bp->b_bufsize; 2483 bufspacewakeup(); 2484 bp->b_bufsize = 0; 2485 } 2486 bp->b_flags &= ~B_MALLOC; 2487 newbsize = round_page(newbsize); 2488 } 2489#endif 2490 vm_hold_load_pages( 2491 bp, 2492 (vm_offset_t) bp->b_data + bp->b_bufsize, 2493 (vm_offset_t) bp->b_data + newbsize); 2494#if !defined(NO_B_MALLOC) 2495 if (origbuf) { 2496 bcopy(origbuf, bp->b_data, origbufsize); 2497 free(origbuf, M_BIOBUF); 2498 } 2499#endif 2500 } 2501 } else { 2502 vm_page_t m; 2503 int desiredpages; 2504 2505 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2506 desiredpages = (size == 0) ? 0 : 2507 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2508 2509#if !defined(NO_B_MALLOC) 2510 if (bp->b_flags & B_MALLOC) 2511 panic("allocbuf: VMIO buffer can't be malloced"); 2512#endif 2513 /* 2514 * Set B_CACHE initially if buffer is 0 length or will become 2515 * 0-length. 2516 */ 2517 if (size == 0 || bp->b_bufsize == 0) 2518 bp->b_flags |= B_CACHE; 2519 2520 if (newbsize < bp->b_bufsize) { 2521 /* 2522 * DEV_BSIZE aligned new buffer size is less then the 2523 * DEV_BSIZE aligned existing buffer size. Figure out 2524 * if we have to remove any pages. 2525 */ 2526 if (desiredpages < bp->b_npages) { 2527 for (i = desiredpages; i < bp->b_npages; i++) { 2528 /* 2529 * the page is not freed here -- it 2530 * is the responsibility of 2531 * vnode_pager_setsize 2532 */ 2533 m = bp->b_pages[i]; 2534 KASSERT(m != bogus_page, 2535 ("allocbuf: bogus page found")); 2536 while (vm_page_sleep_busy(m, TRUE, "biodep")) 2537 ; 2538 2539 bp->b_pages[i] = NULL; 2540 vm_page_unwire(m, 0); 2541 } 2542 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2543 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2544 bp->b_npages = desiredpages; 2545 } 2546 } else if (size > bp->b_bcount) { 2547 /* 2548 * We are growing the buffer, possibly in a 2549 * byte-granular fashion. 2550 */ 2551 struct vnode *vp; 2552 vm_object_t obj; 2553 vm_offset_t toff; 2554 vm_offset_t tinc; 2555 2556 /* 2557 * Step 1, bring in the VM pages from the object, 2558 * allocating them if necessary. We must clear 2559 * B_CACHE if these pages are not valid for the 2560 * range covered by the buffer. 2561 */ 2562 2563 vp = bp->b_vp; 2564 VOP_GETVOBJECT(vp, &obj); 2565 2566 while (bp->b_npages < desiredpages) { 2567 vm_page_t m; 2568 vm_pindex_t pi; 2569 2570 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2571 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2572 /* 2573 * note: must allocate system pages 2574 * since blocking here could intefere 2575 * with paging I/O, no matter which 2576 * process we are. 2577 */ 2578 m = vm_page_alloc(obj, pi, VM_ALLOC_SYSTEM); 2579 if (m == NULL) { 2580 VM_WAIT; 2581 vm_pageout_deficit += desiredpages - bp->b_npages; 2582 } else { 2583 vm_page_wire(m); 2584 vm_page_wakeup(m); 2585 bp->b_flags &= ~B_CACHE; 2586 bp->b_pages[bp->b_npages] = m; 2587 ++bp->b_npages; 2588 } 2589 continue; 2590 } 2591 2592 /* 2593 * We found a page. If we have to sleep on it, 2594 * retry because it might have gotten freed out 2595 * from under us. 2596 * 2597 * We can only test PG_BUSY here. Blocking on 2598 * m->busy might lead to a deadlock: 2599 * 2600 * vm_fault->getpages->cluster_read->allocbuf 2601 * 2602 */ 2603 2604 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 2605 continue; 2606 2607 /* 2608 * We have a good page. Should we wakeup the 2609 * page daemon? 2610 */ 2611 if ((curproc != pageproc) && 2612 ((m->queue - m->pc) == PQ_CACHE) && 2613 ((cnt.v_free_count + cnt.v_cache_count) < 2614 (cnt.v_free_min + cnt.v_cache_min))) { 2615 pagedaemon_wakeup(); 2616 } 2617 vm_page_flag_clear(m, PG_ZERO); 2618 vm_page_wire(m); 2619 bp->b_pages[bp->b_npages] = m; 2620 ++bp->b_npages; 2621 } 2622 2623 /* 2624 * Step 2. We've loaded the pages into the buffer, 2625 * we have to figure out if we can still have B_CACHE 2626 * set. Note that B_CACHE is set according to the 2627 * byte-granular range ( bcount and size ), new the 2628 * aligned range ( newbsize ). 2629 * 2630 * The VM test is against m->valid, which is DEV_BSIZE 2631 * aligned. Needless to say, the validity of the data 2632 * needs to also be DEV_BSIZE aligned. Note that this 2633 * fails with NFS if the server or some other client 2634 * extends the file's EOF. If our buffer is resized, 2635 * B_CACHE may remain set! XXX 2636 */ 2637 2638 toff = bp->b_bcount; 2639 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2640 2641 while ((bp->b_flags & B_CACHE) && toff < size) { 2642 vm_pindex_t pi; 2643 2644 if (tinc > (size - toff)) 2645 tinc = size - toff; 2646 2647 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2648 PAGE_SHIFT; 2649 2650 vfs_buf_test_cache( 2651 bp, 2652 bp->b_offset, 2653 toff, 2654 tinc, 2655 bp->b_pages[pi] 2656 ); 2657 toff += tinc; 2658 tinc = PAGE_SIZE; 2659 } 2660 2661 /* 2662 * Step 3, fixup the KVM pmap. Remember that 2663 * bp->b_data is relative to bp->b_offset, but 2664 * bp->b_offset may be offset into the first page. 2665 */ 2666 2667 bp->b_data = (caddr_t) 2668 trunc_page((vm_offset_t)bp->b_data); 2669 pmap_qenter( 2670 (vm_offset_t)bp->b_data, 2671 bp->b_pages, 2672 bp->b_npages 2673 ); 2674 2675 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2676 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2677 } 2678 } 2679 if (newbsize < bp->b_bufsize) 2680 bufspacewakeup(); 2681 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2682 bp->b_bcount = size; /* requested buffer size */ 2683 return 1; 2684} 2685 2686/* 2687 * bufwait: 2688 * 2689 * Wait for buffer I/O completion, returning error status. The buffer 2690 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR 2691 * error and cleared. 2692 */ 2693int 2694bufwait(register struct buf * bp) 2695{ 2696 int s; 2697 2698 s = splbio(); 2699 while ((bp->b_flags & B_DONE) == 0) { 2700 if (bp->b_iocmd == BIO_READ) 2701 tsleep(bp, PRIBIO, "biord", 0); 2702 else 2703 tsleep(bp, PRIBIO, "biowr", 0); 2704 } 2705 splx(s); 2706 if (bp->b_flags & B_EINTR) { 2707 bp->b_flags &= ~B_EINTR; 2708 return (EINTR); 2709 } 2710 if (bp->b_ioflags & BIO_ERROR) { 2711 return (bp->b_error ? bp->b_error : EIO); 2712 } else { 2713 return (0); 2714 } 2715} 2716 2717 /* 2718 * Call back function from struct bio back up to struct buf. 2719 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 2720 */ 2721void 2722bufdonebio(struct bio *bp) 2723{ 2724 bufdone(bp->bio_caller2); 2725} 2726 2727/* 2728 * bufdone: 2729 * 2730 * Finish I/O on a buffer, optionally calling a completion function. 2731 * This is usually called from an interrupt so process blocking is 2732 * not allowed. 2733 * 2734 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2735 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2736 * assuming B_INVAL is clear. 2737 * 2738 * For the VMIO case, we set B_CACHE if the op was a read and no 2739 * read error occured, or if the op was a write. B_CACHE is never 2740 * set if the buffer is invalid or otherwise uncacheable. 2741 * 2742 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2743 * initiator to leave B_INVAL set to brelse the buffer out of existance 2744 * in the biodone routine. 2745 */ 2746void 2747bufdone(struct buf *bp) 2748{ 2749 int s, error; 2750 void (*biodone) __P((struct buf *)); 2751 2752 GIANT_REQUIRED; 2753 2754 s = splbio(); 2755 2756 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 2757 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 2758 2759 bp->b_flags |= B_DONE; 2760 runningbufwakeup(bp); 2761 2762 if (bp->b_iocmd == BIO_DELETE) { 2763 brelse(bp); 2764 splx(s); 2765 return; 2766 } 2767 2768 if (bp->b_iocmd == BIO_WRITE) { 2769 vwakeup(bp); 2770 } 2771 2772 /* call optional completion function if requested */ 2773 if (bp->b_iodone != NULL) { 2774 biodone = bp->b_iodone; 2775 bp->b_iodone = NULL; 2776 (*biodone) (bp); 2777 splx(s); 2778 return; 2779 } 2780 if (LIST_FIRST(&bp->b_dep) != NULL) 2781 buf_complete(bp); 2782 2783 if (bp->b_flags & B_VMIO) { 2784 int i; 2785 vm_ooffset_t foff; 2786 vm_page_t m; 2787 vm_object_t obj; 2788 int iosize; 2789 struct vnode *vp = bp->b_vp; 2790 2791 error = VOP_GETVOBJECT(vp, &obj); 2792 2793#if defined(VFS_BIO_DEBUG) 2794 if (vp->v_usecount == 0) { 2795 panic("biodone: zero vnode ref count"); 2796 } 2797 2798 if (error) { 2799 panic("biodone: missing VM object"); 2800 } 2801 2802 if ((vp->v_flag & VOBJBUF) == 0) { 2803 panic("biodone: vnode is not setup for merged cache"); 2804 } 2805#endif 2806 2807 foff = bp->b_offset; 2808 KASSERT(bp->b_offset != NOOFFSET, 2809 ("biodone: no buffer offset")); 2810 2811 if (error) { 2812 panic("biodone: no object"); 2813 } 2814#if defined(VFS_BIO_DEBUG) 2815 if (obj->paging_in_progress < bp->b_npages) { 2816 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 2817 obj->paging_in_progress, bp->b_npages); 2818 } 2819#endif 2820 2821 /* 2822 * Set B_CACHE if the op was a normal read and no error 2823 * occured. B_CACHE is set for writes in the b*write() 2824 * routines. 2825 */ 2826 iosize = bp->b_bcount - bp->b_resid; 2827 if (bp->b_iocmd == BIO_READ && 2828 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 2829 !(bp->b_ioflags & BIO_ERROR)) { 2830 bp->b_flags |= B_CACHE; 2831 } 2832 2833 for (i = 0; i < bp->b_npages; i++) { 2834 int bogusflag = 0; 2835 int resid; 2836 2837 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2838 if (resid > iosize) 2839 resid = iosize; 2840 2841 /* 2842 * cleanup bogus pages, restoring the originals 2843 */ 2844 m = bp->b_pages[i]; 2845 if (m == bogus_page) { 2846 bogusflag = 1; 2847 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2848 if (m == NULL) 2849 panic("biodone: page disappeared!"); 2850 bp->b_pages[i] = m; 2851 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2852 } 2853#if defined(VFS_BIO_DEBUG) 2854 if (OFF_TO_IDX(foff) != m->pindex) { 2855 printf( 2856"biodone: foff(%lu)/m->pindex(%d) mismatch\n", 2857 (unsigned long)foff, m->pindex); 2858 } 2859#endif 2860 2861 /* 2862 * In the write case, the valid and clean bits are 2863 * already changed correctly ( see bdwrite() ), so we 2864 * only need to do this here in the read case. 2865 */ 2866 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 2867 vfs_page_set_valid(bp, foff, i, m); 2868 } 2869 vm_page_flag_clear(m, PG_ZERO); 2870 2871 /* 2872 * when debugging new filesystems or buffer I/O methods, this 2873 * is the most common error that pops up. if you see this, you 2874 * have not set the page busy flag correctly!!! 2875 */ 2876 if (m->busy == 0) { 2877 printf("biodone: page busy < 0, " 2878 "pindex: %d, foff: 0x(%x,%x), " 2879 "resid: %d, index: %d\n", 2880 (int) m->pindex, (int)(foff >> 32), 2881 (int) foff & 0xffffffff, resid, i); 2882 if (!vn_isdisk(vp, NULL)) 2883 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 2884 bp->b_vp->v_mount->mnt_stat.f_iosize, 2885 (int) bp->b_lblkno, 2886 bp->b_flags, bp->b_npages); 2887 else 2888 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 2889 (int) bp->b_lblkno, 2890 bp->b_flags, bp->b_npages); 2891 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2892 m->valid, m->dirty, m->wire_count); 2893 panic("biodone: page busy < 0\n"); 2894 } 2895 vm_page_io_finish(m); 2896 vm_object_pip_subtract(obj, 1); 2897 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2898 iosize -= resid; 2899 } 2900 if (obj) 2901 vm_object_pip_wakeupn(obj, 0); 2902 } 2903 2904 /* 2905 * For asynchronous completions, release the buffer now. The brelse 2906 * will do a wakeup there if necessary - so no need to do a wakeup 2907 * here in the async case. The sync case always needs to do a wakeup. 2908 */ 2909 2910 if (bp->b_flags & B_ASYNC) { 2911 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 2912 brelse(bp); 2913 else 2914 bqrelse(bp); 2915 } else { 2916 wakeup(bp); 2917 } 2918 splx(s); 2919} 2920 2921/* 2922 * This routine is called in lieu of iodone in the case of 2923 * incomplete I/O. This keeps the busy status for pages 2924 * consistant. 2925 */ 2926void 2927vfs_unbusy_pages(struct buf * bp) 2928{ 2929 int i; 2930 2931 GIANT_REQUIRED; 2932 2933 runningbufwakeup(bp); 2934 if (bp->b_flags & B_VMIO) { 2935 struct vnode *vp = bp->b_vp; 2936 vm_object_t obj; 2937 2938 VOP_GETVOBJECT(vp, &obj); 2939 2940 for (i = 0; i < bp->b_npages; i++) { 2941 vm_page_t m = bp->b_pages[i]; 2942 2943 if (m == bogus_page) { 2944 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 2945 if (!m) { 2946 panic("vfs_unbusy_pages: page missing\n"); 2947 } 2948 bp->b_pages[i] = m; 2949 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2950 } 2951 vm_object_pip_subtract(obj, 1); 2952 vm_page_flag_clear(m, PG_ZERO); 2953 vm_page_io_finish(m); 2954 } 2955 vm_object_pip_wakeupn(obj, 0); 2956 } 2957} 2958 2959/* 2960 * vfs_page_set_valid: 2961 * 2962 * Set the valid bits in a page based on the supplied offset. The 2963 * range is restricted to the buffer's size. 2964 * 2965 * This routine is typically called after a read completes. 2966 */ 2967static void 2968vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 2969{ 2970 vm_ooffset_t soff, eoff; 2971 2972 GIANT_REQUIRED; 2973 /* 2974 * Start and end offsets in buffer. eoff - soff may not cross a 2975 * page boundry or cross the end of the buffer. The end of the 2976 * buffer, in this case, is our file EOF, not the allocation size 2977 * of the buffer. 2978 */ 2979 soff = off; 2980 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2981 if (eoff > bp->b_offset + bp->b_bcount) 2982 eoff = bp->b_offset + bp->b_bcount; 2983 2984 /* 2985 * Set valid range. This is typically the entire buffer and thus the 2986 * entire page. 2987 */ 2988 if (eoff > soff) { 2989 vm_page_set_validclean( 2990 m, 2991 (vm_offset_t) (soff & PAGE_MASK), 2992 (vm_offset_t) (eoff - soff) 2993 ); 2994 } 2995} 2996 2997/* 2998 * This routine is called before a device strategy routine. 2999 * It is used to tell the VM system that paging I/O is in 3000 * progress, and treat the pages associated with the buffer 3001 * almost as being PG_BUSY. Also the object paging_in_progress 3002 * flag is handled to make sure that the object doesn't become 3003 * inconsistant. 3004 * 3005 * Since I/O has not been initiated yet, certain buffer flags 3006 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3007 * and should be ignored. 3008 */ 3009void 3010vfs_busy_pages(struct buf * bp, int clear_modify) 3011{ 3012 int i, bogus; 3013 3014 GIANT_REQUIRED; 3015 3016 if (bp->b_flags & B_VMIO) { 3017 struct vnode *vp = bp->b_vp; 3018 vm_object_t obj; 3019 vm_ooffset_t foff; 3020 3021 VOP_GETVOBJECT(vp, &obj); 3022 foff = bp->b_offset; 3023 KASSERT(bp->b_offset != NOOFFSET, 3024 ("vfs_busy_pages: no buffer offset")); 3025 vfs_setdirty(bp); 3026 3027retry: 3028 for (i = 0; i < bp->b_npages; i++) { 3029 vm_page_t m = bp->b_pages[i]; 3030 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 3031 goto retry; 3032 } 3033 3034 bogus = 0; 3035 for (i = 0; i < bp->b_npages; i++) { 3036 vm_page_t m = bp->b_pages[i]; 3037 3038 vm_page_flag_clear(m, PG_ZERO); 3039 if ((bp->b_flags & B_CLUSTER) == 0) { 3040 vm_object_pip_add(obj, 1); 3041 vm_page_io_start(m); 3042 } 3043 3044 /* 3045 * When readying a buffer for a read ( i.e 3046 * clear_modify == 0 ), it is important to do 3047 * bogus_page replacement for valid pages in 3048 * partially instantiated buffers. Partially 3049 * instantiated buffers can, in turn, occur when 3050 * reconstituting a buffer from its VM backing store 3051 * base. We only have to do this if B_CACHE is 3052 * clear ( which causes the I/O to occur in the 3053 * first place ). The replacement prevents the read 3054 * I/O from overwriting potentially dirty VM-backed 3055 * pages. XXX bogus page replacement is, uh, bogus. 3056 * It may not work properly with small-block devices. 3057 * We need to find a better way. 3058 */ 3059 3060 vm_page_protect(m, VM_PROT_NONE); 3061 if (clear_modify) 3062 vfs_page_set_valid(bp, foff, i, m); 3063 else if (m->valid == VM_PAGE_BITS_ALL && 3064 (bp->b_flags & B_CACHE) == 0) { 3065 bp->b_pages[i] = bogus_page; 3066 bogus++; 3067 } 3068 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3069 } 3070 if (bogus) 3071 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3072 } 3073} 3074 3075/* 3076 * Tell the VM system that the pages associated with this buffer 3077 * are clean. This is used for delayed writes where the data is 3078 * going to go to disk eventually without additional VM intevention. 3079 * 3080 * Note that while we only really need to clean through to b_bcount, we 3081 * just go ahead and clean through to b_bufsize. 3082 */ 3083static void 3084vfs_clean_pages(struct buf * bp) 3085{ 3086 int i; 3087 3088 GIANT_REQUIRED; 3089 3090 if (bp->b_flags & B_VMIO) { 3091 vm_ooffset_t foff; 3092 3093 foff = bp->b_offset; 3094 KASSERT(bp->b_offset != NOOFFSET, 3095 ("vfs_clean_pages: no buffer offset")); 3096 for (i = 0; i < bp->b_npages; i++) { 3097 vm_page_t m = bp->b_pages[i]; 3098 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3099 vm_ooffset_t eoff = noff; 3100 3101 if (eoff > bp->b_offset + bp->b_bufsize) 3102 eoff = bp->b_offset + bp->b_bufsize; 3103 vfs_page_set_valid(bp, foff, i, m); 3104 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3105 foff = noff; 3106 } 3107 } 3108} 3109 3110/* 3111 * vfs_bio_set_validclean: 3112 * 3113 * Set the range within the buffer to valid and clean. The range is 3114 * relative to the beginning of the buffer, b_offset. Note that b_offset 3115 * itself may be offset from the beginning of the first page. 3116 * 3117 */ 3118 3119void 3120vfs_bio_set_validclean(struct buf *bp, int base, int size) 3121{ 3122 if (bp->b_flags & B_VMIO) { 3123 int i; 3124 int n; 3125 3126 /* 3127 * Fixup base to be relative to beginning of first page. 3128 * Set initial n to be the maximum number of bytes in the 3129 * first page that can be validated. 3130 */ 3131 3132 base += (bp->b_offset & PAGE_MASK); 3133 n = PAGE_SIZE - (base & PAGE_MASK); 3134 3135 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3136 vm_page_t m = bp->b_pages[i]; 3137 3138 if (n > size) 3139 n = size; 3140 3141 vm_page_set_validclean(m, base & PAGE_MASK, n); 3142 base += n; 3143 size -= n; 3144 n = PAGE_SIZE; 3145 } 3146 } 3147} 3148 3149/* 3150 * vfs_bio_clrbuf: 3151 * 3152 * clear a buffer. This routine essentially fakes an I/O, so we need 3153 * to clear BIO_ERROR and B_INVAL. 3154 * 3155 * Note that while we only theoretically need to clear through b_bcount, 3156 * we go ahead and clear through b_bufsize. 3157 */ 3158 3159void 3160vfs_bio_clrbuf(struct buf *bp) { 3161 int i, mask = 0; 3162 caddr_t sa, ea; 3163 3164 GIANT_REQUIRED; 3165 3166 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3167 bp->b_flags &= ~B_INVAL; 3168 bp->b_ioflags &= ~BIO_ERROR; 3169 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3170 (bp->b_offset & PAGE_MASK) == 0) { 3171 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3172 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3173 ((bp->b_pages[0]->valid & mask) != mask)) { 3174 bzero(bp->b_data, bp->b_bufsize); 3175 } 3176 bp->b_pages[0]->valid |= mask; 3177 bp->b_resid = 0; 3178 return; 3179 } 3180 ea = sa = bp->b_data; 3181 for(i=0;i<bp->b_npages;i++,sa=ea) { 3182 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3183 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3184 ea = (caddr_t)(vm_offset_t)ulmin( 3185 (u_long)(vm_offset_t)ea, 3186 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3187 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3188 if ((bp->b_pages[i]->valid & mask) == mask) 3189 continue; 3190 if ((bp->b_pages[i]->valid & mask) == 0) { 3191 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3192 bzero(sa, ea - sa); 3193 } 3194 } else { 3195 for (; sa < ea; sa += DEV_BSIZE, j++) { 3196 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3197 (bp->b_pages[i]->valid & (1<<j)) == 0) 3198 bzero(sa, DEV_BSIZE); 3199 } 3200 } 3201 bp->b_pages[i]->valid |= mask; 3202 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3203 } 3204 bp->b_resid = 0; 3205 } else { 3206 clrbuf(bp); 3207 } 3208} 3209 3210/* 3211 * vm_hold_load_pages and vm_hold_free_pages get pages into 3212 * a buffers address space. The pages are anonymous and are 3213 * not associated with a file object. 3214 */ 3215static void 3216vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3217{ 3218 vm_offset_t pg; 3219 vm_page_t p; 3220 int index; 3221 3222 GIANT_REQUIRED; 3223 3224 to = round_page(to); 3225 from = round_page(from); 3226 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3227 3228 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3229tryagain: 3230 /* 3231 * note: must allocate system pages since blocking here 3232 * could intefere with paging I/O, no matter which 3233 * process we are. 3234 */ 3235 p = vm_page_alloc(kernel_object, 3236 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3237 VM_ALLOC_SYSTEM); 3238 if (!p) { 3239 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3240 VM_WAIT; 3241 goto tryagain; 3242 } 3243 vm_page_wire(p); 3244 p->valid = VM_PAGE_BITS_ALL; 3245 vm_page_flag_clear(p, PG_ZERO); 3246 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 3247 bp->b_pages[index] = p; 3248 vm_page_wakeup(p); 3249 } 3250 bp->b_npages = index; 3251} 3252 3253void 3254vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3255{ 3256 vm_offset_t pg; 3257 vm_page_t p; 3258 int index, newnpages; 3259 3260 GIANT_REQUIRED; 3261 3262 from = round_page(from); 3263 to = round_page(to); 3264 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3265 3266 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3267 p = bp->b_pages[index]; 3268 if (p && (index < bp->b_npages)) { 3269 if (p->busy) { 3270 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 3271 bp->b_blkno, bp->b_lblkno); 3272 } 3273 bp->b_pages[index] = NULL; 3274 pmap_kremove(pg); 3275 vm_page_busy(p); 3276 vm_page_unwire(p, 0); 3277 vm_page_free(p); 3278 } 3279 } 3280 bp->b_npages = newnpages; 3281} 3282 3283 3284#include "opt_ddb.h" 3285#ifdef DDB 3286#include <ddb/ddb.h> 3287 3288DB_SHOW_COMMAND(buffer, db_show_buffer) 3289{ 3290 /* get args */ 3291 struct buf *bp = (struct buf *)addr; 3292 3293 if (!have_addr) { 3294 db_printf("usage: show buffer <addr>\n"); 3295 return; 3296 } 3297 3298 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3299 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 3300 "b_resid = %ld\nb_dev = (%d,%d), b_data = %p, " 3301 "b_blkno = %d, b_pblkno = %d\n", 3302 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3303 major(bp->b_dev), minor(bp->b_dev), 3304 bp->b_data, bp->b_blkno, bp->b_pblkno); 3305 if (bp->b_npages) { 3306 int i; 3307 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3308 for (i = 0; i < bp->b_npages; i++) { 3309 vm_page_t m; 3310 m = bp->b_pages[i]; 3311 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3312 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3313 if ((i + 1) < bp->b_npages) 3314 db_printf(","); 3315 } 3316 db_printf("\n"); 3317 } 3318} 3319#endif /* DDB */ 3320