vfs_bio.c revision 76117
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 76117 2001-04-29 02:45:39Z grog $ 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 if (bp->b_flags & B_CACHE) { 292 int base = (foff + off) & PAGE_MASK; 293 if (vm_page_is_valid(m, base, size) == 0) 294 bp->b_flags &= ~B_CACHE; 295 } 296} 297 298static __inline__ 299void 300bd_wakeup(int dirtybuflevel) 301{ 302 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 303 bd_request = 1; 304 wakeup(&bd_request); 305 } 306} 307 308/* 309 * bd_speedup - speedup the buffer cache flushing code 310 */ 311 312static __inline__ 313void 314bd_speedup(void) 315{ 316 bd_wakeup(1); 317} 318 319/* 320 * Initialize buffer headers and related structures. 321 */ 322 323caddr_t 324bufhashinit(caddr_t vaddr) 325{ 326 /* first, make a null hash table */ 327 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 328 ; 329 bufhashtbl = (void *)vaddr; 330 vaddr = vaddr + sizeof(*bufhashtbl) * bufhashmask; 331 --bufhashmask; 332 return(vaddr); 333} 334 335void 336bufinit(void) 337{ 338 struct buf *bp; 339 int i; 340 341 TAILQ_INIT(&bswlist); 342 LIST_INIT(&invalhash); 343 mtx_init(&buftimelock, "buftime lock", MTX_DEF); 344 345 for (i = 0; i <= bufhashmask; i++) 346 LIST_INIT(&bufhashtbl[i]); 347 348 /* next, make a null set of free lists */ 349 for (i = 0; i < BUFFER_QUEUES; i++) 350 TAILQ_INIT(&bufqueues[i]); 351 352 /* finally, initialize each buffer header and stick on empty q */ 353 for (i = 0; i < nbuf; i++) { 354 bp = &buf[i]; 355 bzero(bp, sizeof *bp); 356 bp->b_flags = B_INVAL; /* we're just an empty header */ 357 bp->b_dev = NODEV; 358 bp->b_rcred = NOCRED; 359 bp->b_wcred = NOCRED; 360 bp->b_qindex = QUEUE_EMPTY; 361 bp->b_xflags = 0; 362 LIST_INIT(&bp->b_dep); 363 BUF_LOCKINIT(bp); 364 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 365 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 366 } 367 368 /* 369 * maxbufspace is the absolute maximum amount of buffer space we are 370 * allowed to reserve in KVM and in real terms. The absolute maximum 371 * is nominally used by buf_daemon. hibufspace is the nominal maximum 372 * used by most other processes. The differential is required to 373 * ensure that buf_daemon is able to run when other processes might 374 * be blocked waiting for buffer space. 375 * 376 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 377 * this may result in KVM fragmentation which is not handled optimally 378 * by the system. 379 */ 380 maxbufspace = nbuf * BKVASIZE; 381 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 382 lobufspace = hibufspace - MAXBSIZE; 383 384 lorunningspace = 512 * 1024; 385 hirunningspace = 1024 * 1024; 386 387/* 388 * Limit the amount of malloc memory since it is wired permanently into 389 * the kernel space. Even though this is accounted for in the buffer 390 * allocation, we don't want the malloced region to grow uncontrolled. 391 * The malloc scheme improves memory utilization significantly on average 392 * (small) directories. 393 */ 394 maxbufmallocspace = hibufspace / 20; 395 396/* 397 * Reduce the chance of a deadlock occuring by limiting the number 398 * of delayed-write dirty buffers we allow to stack up. 399 */ 400 hidirtybuffers = nbuf / 4 + 20; 401 numdirtybuffers = 0; 402/* 403 * To support extreme low-memory systems, make sure hidirtybuffers cannot 404 * eat up all available buffer space. This occurs when our minimum cannot 405 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 406 * BKVASIZE'd (8K) buffers. 407 */ 408 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 409 hidirtybuffers >>= 1; 410 } 411 lodirtybuffers = hidirtybuffers / 2; 412 413/* 414 * Try to keep the number of free buffers in the specified range, 415 * and give special processes (e.g. like buf_daemon) access to an 416 * emergency reserve. 417 */ 418 lofreebuffers = nbuf / 18 + 5; 419 hifreebuffers = 2 * lofreebuffers; 420 numfreebuffers = nbuf; 421 422/* 423 * Maximum number of async ops initiated per buf_daemon loop. This is 424 * somewhat of a hack at the moment, we really need to limit ourselves 425 * based on the number of bytes of I/O in-transit that were initiated 426 * from buf_daemon. 427 */ 428 429 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 430 bogus_page = vm_page_alloc(kernel_object, 431 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 432 VM_ALLOC_NORMAL); 433 cnt.v_wire_count++; 434 435} 436 437/* 438 * bfreekva() - free the kva allocation for a buffer. 439 * 440 * Must be called at splbio() or higher as this is the only locking for 441 * buffer_map. 442 * 443 * Since this call frees up buffer space, we call bufspacewakeup(). 444 */ 445static void 446bfreekva(struct buf * bp) 447{ 448 if (bp->b_kvasize) { 449 ++buffreekvacnt; 450 bufspace -= bp->b_kvasize; 451 vm_map_delete(buffer_map, 452 (vm_offset_t) bp->b_kvabase, 453 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 454 ); 455 bp->b_kvasize = 0; 456 bufspacewakeup(); 457 } 458} 459 460/* 461 * bremfree: 462 * 463 * Remove the buffer from the appropriate free list. 464 */ 465void 466bremfree(struct buf * bp) 467{ 468 int s = splbio(); 469 int old_qindex = bp->b_qindex; 470 471 if (bp->b_qindex != QUEUE_NONE) { 472 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 473 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 474 bp->b_qindex = QUEUE_NONE; 475 } else { 476 if (BUF_REFCNT(bp) <= 1) 477 panic("bremfree: removing a buffer not on a queue"); 478 } 479 480 /* 481 * Fixup numfreebuffers count. If the buffer is invalid or not 482 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 483 * the buffer was free and we must decrement numfreebuffers. 484 */ 485 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 486 switch(old_qindex) { 487 case QUEUE_DIRTY: 488 case QUEUE_CLEAN: 489 case QUEUE_EMPTY: 490 case QUEUE_EMPTYKVA: 491 --numfreebuffers; 492 break; 493 default: 494 break; 495 } 496 } 497 splx(s); 498} 499 500 501/* 502 * Get a buffer with the specified data. Look in the cache first. We 503 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 504 * is set, the buffer is valid and we do not have to do anything ( see 505 * getblk() ). This is really just a special case of breadn(). 506 */ 507int 508bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 509 struct buf ** bpp) 510{ 511 512 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 513} 514 515/* 516 * Operates like bread, but also starts asynchronous I/O on 517 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 518 * to initiating I/O . If B_CACHE is set, the buffer is valid 519 * and we do not have to do anything. 520 */ 521int 522breadn(struct vnode * vp, daddr_t blkno, int size, 523 daddr_t * rablkno, int *rabsize, 524 int cnt, struct ucred * cred, struct buf ** bpp) 525{ 526 struct buf *bp, *rabp; 527 int i; 528 int rv = 0, readwait = 0; 529 530 *bpp = bp = getblk(vp, blkno, size, 0, 0); 531 532 /* if not found in cache, do some I/O */ 533 if ((bp->b_flags & B_CACHE) == 0) { 534 if (curproc != PCPU_GET(idleproc)) 535 curproc->p_stats->p_ru.ru_inblock++; 536 bp->b_iocmd = BIO_READ; 537 bp->b_flags &= ~B_INVAL; 538 bp->b_ioflags &= ~BIO_ERROR; 539 if (bp->b_rcred == NOCRED) { 540 if (cred != NOCRED) 541 crhold(cred); 542 bp->b_rcred = cred; 543 } 544 vfs_busy_pages(bp, 0); 545 VOP_STRATEGY(vp, bp); 546 ++readwait; 547 } 548 549 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 550 if (inmem(vp, *rablkno)) 551 continue; 552 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 553 554 if ((rabp->b_flags & B_CACHE) == 0) { 555 if (curproc != PCPU_GET(idleproc)) 556 curproc->p_stats->p_ru.ru_inblock++; 557 rabp->b_flags |= B_ASYNC; 558 rabp->b_flags &= ~B_INVAL; 559 rabp->b_ioflags &= ~BIO_ERROR; 560 rabp->b_iocmd = BIO_READ; 561 if (rabp->b_rcred == NOCRED) { 562 if (cred != NOCRED) 563 crhold(cred); 564 rabp->b_rcred = cred; 565 } 566 vfs_busy_pages(rabp, 0); 567 BUF_KERNPROC(rabp); 568 VOP_STRATEGY(vp, rabp); 569 } else { 570 brelse(rabp); 571 } 572 } 573 574 if (readwait) { 575 rv = bufwait(bp); 576 } 577 return (rv); 578} 579 580/* 581 * Write, release buffer on completion. (Done by iodone 582 * if async). Do not bother writing anything if the buffer 583 * is invalid. 584 * 585 * Note that we set B_CACHE here, indicating that buffer is 586 * fully valid and thus cacheable. This is true even of NFS 587 * now so we set it generally. This could be set either here 588 * or in biodone() since the I/O is synchronous. We put it 589 * here. 590 */ 591 592int dobkgrdwrite = 1; 593SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, ""); 594 595int 596bwrite(struct buf * bp) 597{ 598 int oldflags, s; 599 struct buf *newbp; 600 601 if (bp->b_flags & B_INVAL) { 602 brelse(bp); 603 return (0); 604 } 605 606 oldflags = bp->b_flags; 607 608 if (BUF_REFCNT(bp) == 0) 609 panic("bwrite: buffer is not busy???"); 610 s = splbio(); 611 /* 612 * If a background write is already in progress, delay 613 * writing this block if it is asynchronous. Otherwise 614 * wait for the background write to complete. 615 */ 616 if (bp->b_xflags & BX_BKGRDINPROG) { 617 if (bp->b_flags & B_ASYNC) { 618 splx(s); 619 bdwrite(bp); 620 return (0); 621 } 622 bp->b_xflags |= BX_BKGRDWAIT; 623 tsleep(&bp->b_xflags, PRIBIO, "biord", 0); 624 if (bp->b_xflags & BX_BKGRDINPROG) 625 panic("bwrite: still writing"); 626 } 627 628 /* Mark the buffer clean */ 629 bundirty(bp); 630 631 /* 632 * If this buffer is marked for background writing and we 633 * do not have to wait for it, make a copy and write the 634 * copy so as to leave this buffer ready for further use. 635 * 636 * This optimization eats a lot of memory. If we have a page 637 * or buffer shortfall we can't do it. 638 */ 639 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 640 (bp->b_flags & B_ASYNC) && 641 !vm_page_count_severe() && 642 !buf_dirty_count_severe()) { 643 if (bp->b_iodone != NULL) { 644 printf("bp->b_iodone = %p\n", bp->b_iodone); 645 panic("bwrite: need chained iodone"); 646 } 647 648 /* get a new block */ 649 newbp = geteblk(bp->b_bufsize); 650 651 /* set it to be identical to the old block */ 652 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 653 bgetvp(bp->b_vp, newbp); 654 newbp->b_lblkno = bp->b_lblkno; 655 newbp->b_blkno = bp->b_blkno; 656 newbp->b_offset = bp->b_offset; 657 newbp->b_iodone = vfs_backgroundwritedone; 658 newbp->b_flags |= B_ASYNC; 659 newbp->b_flags &= ~B_INVAL; 660 661 /* move over the dependencies */ 662 if (LIST_FIRST(&bp->b_dep) != NULL) 663 buf_movedeps(bp, newbp); 664 665 /* 666 * Initiate write on the copy, release the original to 667 * the B_LOCKED queue so that it cannot go away until 668 * the background write completes. If not locked it could go 669 * away and then be reconstituted while it was being written. 670 * If the reconstituted buffer were written, we could end up 671 * with two background copies being written at the same time. 672 */ 673 bp->b_xflags |= BX_BKGRDINPROG; 674 bp->b_flags |= B_LOCKED; 675 bqrelse(bp); 676 bp = newbp; 677 } 678 679 bp->b_flags &= ~B_DONE; 680 bp->b_ioflags &= ~BIO_ERROR; 681 bp->b_flags |= B_WRITEINPROG | B_CACHE; 682 bp->b_iocmd = BIO_WRITE; 683 684 bp->b_vp->v_numoutput++; 685 vfs_busy_pages(bp, 1); 686 687 /* 688 * Normal bwrites pipeline writes 689 */ 690 bp->b_runningbufspace = bp->b_bufsize; 691 runningbufspace += bp->b_runningbufspace; 692 693 if (curproc != PCPU_GET(idleproc)) 694 curproc->p_stats->p_ru.ru_oublock++; 695 splx(s); 696 if (oldflags & B_ASYNC) 697 BUF_KERNPROC(bp); 698 BUF_STRATEGY(bp); 699 700 if ((oldflags & B_ASYNC) == 0) { 701 int rtval = bufwait(bp); 702 brelse(bp); 703 return (rtval); 704 } else { 705 /* 706 * don't allow the async write to saturate the I/O 707 * system. There is no chance of deadlock here because 708 * we are blocking on I/O that is already in-progress. 709 */ 710 waitrunningbufspace(); 711 } 712 713 return (0); 714} 715 716/* 717 * Complete a background write started from bwrite. 718 */ 719static void 720vfs_backgroundwritedone(bp) 721 struct buf *bp; 722{ 723 struct buf *origbp; 724 725 /* 726 * Find the original buffer that we are writing. 727 */ 728 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 729 panic("backgroundwritedone: lost buffer"); 730 /* 731 * Process dependencies then return any unfinished ones. 732 */ 733 if (LIST_FIRST(&bp->b_dep) != NULL) 734 buf_complete(bp); 735 if (LIST_FIRST(&bp->b_dep) != NULL) 736 buf_movedeps(bp, origbp); 737 /* 738 * Clear the BX_BKGRDINPROG flag in the original buffer 739 * and awaken it if it is waiting for the write to complete. 740 * If BX_BKGRDINPROG is not set in the original buffer it must 741 * have been released and re-instantiated - which is not legal. 742 */ 743 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), ("backgroundwritedone: lost buffer2")); 744 origbp->b_xflags &= ~BX_BKGRDINPROG; 745 if (origbp->b_xflags & BX_BKGRDWAIT) { 746 origbp->b_xflags &= ~BX_BKGRDWAIT; 747 wakeup(&origbp->b_xflags); 748 } 749 /* 750 * Clear the B_LOCKED flag and remove it from the locked 751 * queue if it currently resides there. 752 */ 753 origbp->b_flags &= ~B_LOCKED; 754 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 755 bremfree(origbp); 756 bqrelse(origbp); 757 } 758 /* 759 * This buffer is marked B_NOCACHE, so when it is released 760 * by biodone, it will be tossed. We mark it with BIO_READ 761 * to avoid biodone doing a second vwakeup. 762 */ 763 bp->b_flags |= B_NOCACHE; 764 bp->b_iocmd = BIO_READ; 765 bp->b_flags &= ~(B_CACHE | B_DONE); 766 bp->b_iodone = 0; 767 bufdone(bp); 768} 769 770/* 771 * Delayed write. (Buffer is marked dirty). Do not bother writing 772 * anything if the buffer is marked invalid. 773 * 774 * Note that since the buffer must be completely valid, we can safely 775 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 776 * biodone() in order to prevent getblk from writing the buffer 777 * out synchronously. 778 */ 779void 780bdwrite(struct buf * bp) 781{ 782 if (BUF_REFCNT(bp) == 0) 783 panic("bdwrite: buffer is not busy"); 784 785 if (bp->b_flags & B_INVAL) { 786 brelse(bp); 787 return; 788 } 789 bdirty(bp); 790 791 /* 792 * Set B_CACHE, indicating that the buffer is fully valid. This is 793 * true even of NFS now. 794 */ 795 bp->b_flags |= B_CACHE; 796 797 /* 798 * This bmap keeps the system from needing to do the bmap later, 799 * perhaps when the system is attempting to do a sync. Since it 800 * is likely that the indirect block -- or whatever other datastructure 801 * that the filesystem needs is still in memory now, it is a good 802 * thing to do this. Note also, that if the pageout daemon is 803 * requesting a sync -- there might not be enough memory to do 804 * the bmap then... So, this is important to do. 805 */ 806 if (bp->b_lblkno == bp->b_blkno) { 807 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 808 } 809 810 /* 811 * Set the *dirty* buffer range based upon the VM system dirty pages. 812 */ 813 vfs_setdirty(bp); 814 815 /* 816 * We need to do this here to satisfy the vnode_pager and the 817 * pageout daemon, so that it thinks that the pages have been 818 * "cleaned". Note that since the pages are in a delayed write 819 * buffer -- the VFS layer "will" see that the pages get written 820 * out on the next sync, or perhaps the cluster will be completed. 821 */ 822 vfs_clean_pages(bp); 823 bqrelse(bp); 824 825 /* 826 * Wakeup the buffer flushing daemon if we have a lot of dirty 827 * buffers (midpoint between our recovery point and our stall 828 * point). 829 */ 830 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 831 832 /* 833 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 834 * due to the softdep code. 835 */ 836} 837 838/* 839 * bdirty: 840 * 841 * Turn buffer into delayed write request. We must clear BIO_READ and 842 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 843 * itself to properly update it in the dirty/clean lists. We mark it 844 * B_DONE to ensure that any asynchronization of the buffer properly 845 * clears B_DONE ( else a panic will occur later ). 846 * 847 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 848 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 849 * should only be called if the buffer is known-good. 850 * 851 * Since the buffer is not on a queue, we do not update the numfreebuffers 852 * count. 853 * 854 * Must be called at splbio(). 855 * The buffer must be on QUEUE_NONE. 856 */ 857void 858bdirty(bp) 859 struct buf *bp; 860{ 861 KASSERT(bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 862 bp->b_flags &= ~(B_RELBUF); 863 bp->b_iocmd = BIO_WRITE; 864 865 if ((bp->b_flags & B_DELWRI) == 0) { 866 bp->b_flags |= B_DONE | B_DELWRI; 867 reassignbuf(bp, bp->b_vp); 868 ++numdirtybuffers; 869 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 870 } 871} 872 873/* 874 * bundirty: 875 * 876 * Clear B_DELWRI for buffer. 877 * 878 * Since the buffer is not on a queue, we do not update the numfreebuffers 879 * count. 880 * 881 * Must be called at splbio(). 882 * The buffer must be on QUEUE_NONE. 883 */ 884 885void 886bundirty(bp) 887 struct buf *bp; 888{ 889 KASSERT(bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 890 891 if (bp->b_flags & B_DELWRI) { 892 bp->b_flags &= ~B_DELWRI; 893 reassignbuf(bp, bp->b_vp); 894 --numdirtybuffers; 895 numdirtywakeup(lodirtybuffers); 896 } 897 /* 898 * Since it is now being written, we can clear its deferred write flag. 899 */ 900 bp->b_flags &= ~B_DEFERRED; 901} 902 903/* 904 * bawrite: 905 * 906 * Asynchronous write. Start output on a buffer, but do not wait for 907 * it to complete. The buffer is released when the output completes. 908 * 909 * bwrite() ( or the VOP routine anyway ) is responsible for handling 910 * B_INVAL buffers. Not us. 911 */ 912void 913bawrite(struct buf * bp) 914{ 915 bp->b_flags |= B_ASYNC; 916 (void) BUF_WRITE(bp); 917} 918 919/* 920 * bowrite: 921 * 922 * Ordered write. Start output on a buffer, and flag it so that the 923 * device will write it in the order it was queued. The buffer is 924 * released when the output completes. bwrite() ( or the VOP routine 925 * anyway ) is responsible for handling B_INVAL buffers. 926 */ 927int 928bowrite(struct buf * bp) 929{ 930 bp->b_ioflags |= BIO_ORDERED; 931 bp->b_flags |= B_ASYNC; 932 return (BUF_WRITE(bp)); 933} 934 935/* 936 * bwillwrite: 937 * 938 * Called prior to the locking of any vnodes when we are expecting to 939 * write. We do not want to starve the buffer cache with too many 940 * dirty buffers so we block here. By blocking prior to the locking 941 * of any vnodes we attempt to avoid the situation where a locked vnode 942 * prevents the various system daemons from flushing related buffers. 943 */ 944 945void 946bwillwrite(void) 947{ 948 if (numdirtybuffers >= hidirtybuffers) { 949 int s; 950 951 s = splbio(); 952 while (numdirtybuffers >= hidirtybuffers) { 953 bd_wakeup(1); 954 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 955 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0); 956 } 957 splx(s); 958 } 959} 960 961/* 962 * Return true if we have too many dirty buffers. 963 */ 964int 965buf_dirty_count_severe(void) 966{ 967 return(numdirtybuffers >= hidirtybuffers); 968} 969 970/* 971 * brelse: 972 * 973 * Release a busy buffer and, if requested, free its resources. The 974 * buffer will be stashed in the appropriate bufqueue[] allowing it 975 * to be accessed later as a cache entity or reused for other purposes. 976 */ 977void 978brelse(struct buf * bp) 979{ 980 int s; 981 982 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 983 984 s = splbio(); 985 986 if (bp->b_flags & B_LOCKED) 987 bp->b_ioflags &= ~BIO_ERROR; 988 989 if (bp->b_iocmd == BIO_WRITE && 990 (bp->b_ioflags & BIO_ERROR) && 991 !(bp->b_flags & B_INVAL)) { 992 /* 993 * Failed write, redirty. Must clear BIO_ERROR to prevent 994 * pages from being scrapped. If B_INVAL is set then 995 * this case is not run and the next case is run to 996 * destroy the buffer. B_INVAL can occur if the buffer 997 * is outside the range supported by the underlying device. 998 */ 999 bp->b_ioflags &= ~BIO_ERROR; 1000 bdirty(bp); 1001 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1002 (bp->b_ioflags & BIO_ERROR) || 1003 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1004 /* 1005 * Either a failed I/O or we were asked to free or not 1006 * cache the buffer. 1007 */ 1008 bp->b_flags |= B_INVAL; 1009 if (LIST_FIRST(&bp->b_dep) != NULL) 1010 buf_deallocate(bp); 1011 if (bp->b_flags & B_DELWRI) { 1012 --numdirtybuffers; 1013 numdirtywakeup(lodirtybuffers); 1014 } 1015 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1016 if ((bp->b_flags & B_VMIO) == 0) { 1017 if (bp->b_bufsize) 1018 allocbuf(bp, 0); 1019 if (bp->b_vp) 1020 brelvp(bp); 1021 } 1022 } 1023 1024 /* 1025 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1026 * is called with B_DELWRI set, the underlying pages may wind up 1027 * getting freed causing a previous write (bdwrite()) to get 'lost' 1028 * because pages associated with a B_DELWRI bp are marked clean. 1029 * 1030 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1031 * if B_DELWRI is set. 1032 * 1033 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1034 * on pages to return pages to the VM page queues. 1035 */ 1036 if (bp->b_flags & B_DELWRI) 1037 bp->b_flags &= ~B_RELBUF; 1038 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1039 bp->b_flags |= B_RELBUF; 1040 1041 /* 1042 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1043 * constituted, not even NFS buffers now. Two flags effect this. If 1044 * B_INVAL, the struct buf is invalidated but the VM object is kept 1045 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1046 * 1047 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1048 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1049 * buffer is also B_INVAL because it hits the re-dirtying code above. 1050 * 1051 * Normally we can do this whether a buffer is B_DELWRI or not. If 1052 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1053 * the commit state and we cannot afford to lose the buffer. If the 1054 * buffer has a background write in progress, we need to keep it 1055 * around to prevent it from being reconstituted and starting a second 1056 * background write. 1057 */ 1058 if ((bp->b_flags & B_VMIO) 1059 && !(bp->b_vp->v_tag == VT_NFS && 1060 !vn_isdisk(bp->b_vp, NULL) && 1061 (bp->b_flags & B_DELWRI)) 1062 ) { 1063 1064 int i, j, resid; 1065 vm_page_t m; 1066 off_t foff; 1067 vm_pindex_t poff; 1068 vm_object_t obj; 1069 struct vnode *vp; 1070 1071 vp = bp->b_vp; 1072 1073 /* 1074 * Get the base offset and length of the buffer. Note that 1075 * for block sizes that are less then PAGE_SIZE, the b_data 1076 * base of the buffer does not represent exactly b_offset and 1077 * neither b_offset nor b_size are necessarily page aligned. 1078 * Instead, the starting position of b_offset is: 1079 * 1080 * b_data + (b_offset & PAGE_MASK) 1081 * 1082 * block sizes less then DEV_BSIZE (usually 512) are not 1083 * supported due to the page granularity bits (m->valid, 1084 * m->dirty, etc...). 1085 * 1086 * See man buf(9) for more information 1087 */ 1088 resid = bp->b_bufsize; 1089 foff = bp->b_offset; 1090 1091 for (i = 0; i < bp->b_npages; i++) { 1092 int had_bogus = 0; 1093 1094 m = bp->b_pages[i]; 1095 vm_page_flag_clear(m, PG_ZERO); 1096 1097 /* 1098 * If we hit a bogus page, fixup *all* the bogus pages 1099 * now. 1100 */ 1101 if (m == bogus_page) { 1102 VOP_GETVOBJECT(vp, &obj); 1103 poff = OFF_TO_IDX(bp->b_offset); 1104 had_bogus = 1; 1105 1106 for (j = i; j < bp->b_npages; j++) { 1107 vm_page_t mtmp; 1108 mtmp = bp->b_pages[j]; 1109 if (mtmp == bogus_page) { 1110 mtmp = vm_page_lookup(obj, poff + j); 1111 if (!mtmp) { 1112 panic("brelse: page missing\n"); 1113 } 1114 bp->b_pages[j] = mtmp; 1115 } 1116 } 1117 1118 if ((bp->b_flags & B_INVAL) == 0) { 1119 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1120 } 1121 m = bp->b_pages[i]; 1122 } 1123 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1124 int poffset = foff & PAGE_MASK; 1125 int presid = resid > (PAGE_SIZE - poffset) ? 1126 (PAGE_SIZE - poffset) : resid; 1127 1128 KASSERT(presid >= 0, ("brelse: extra page")); 1129 vm_page_set_invalid(m, poffset, presid); 1130 if (had_bogus) 1131 printf("avoided corruption bug in bogus_page/brelse code\n"); 1132 } 1133 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1134 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1135 } 1136 1137 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1138 vfs_vmio_release(bp); 1139 1140 } else if (bp->b_flags & B_VMIO) { 1141 1142 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1143 vfs_vmio_release(bp); 1144 1145 } 1146 1147 if (bp->b_qindex != QUEUE_NONE) 1148 panic("brelse: free buffer onto another queue???"); 1149 if (BUF_REFCNT(bp) > 1) { 1150 /* do not release to free list */ 1151 BUF_UNLOCK(bp); 1152 splx(s); 1153 return; 1154 } 1155 1156 /* enqueue */ 1157 1158 /* buffers with no memory */ 1159 if (bp->b_bufsize == 0) { 1160 bp->b_flags |= B_INVAL; 1161 bp->b_xflags &= ~BX_BKGRDWRITE; 1162 if (bp->b_xflags & BX_BKGRDINPROG) 1163 panic("losing buffer 1"); 1164 if (bp->b_kvasize) { 1165 bp->b_qindex = QUEUE_EMPTYKVA; 1166 } else { 1167 bp->b_qindex = QUEUE_EMPTY; 1168 } 1169 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1170 LIST_REMOVE(bp, b_hash); 1171 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1172 bp->b_dev = NODEV; 1173 /* buffers with junk contents */ 1174 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || (bp->b_ioflags & BIO_ERROR)) { 1175 bp->b_flags |= B_INVAL; 1176 bp->b_xflags &= ~BX_BKGRDWRITE; 1177 if (bp->b_xflags & BX_BKGRDINPROG) 1178 panic("losing buffer 2"); 1179 bp->b_qindex = QUEUE_CLEAN; 1180 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1181 LIST_REMOVE(bp, b_hash); 1182 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1183 bp->b_dev = NODEV; 1184 1185 /* buffers that are locked */ 1186 } else if (bp->b_flags & B_LOCKED) { 1187 bp->b_qindex = QUEUE_LOCKED; 1188 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1189 1190 /* remaining buffers */ 1191 } else { 1192 if (bp->b_flags & B_DELWRI) 1193 bp->b_qindex = QUEUE_DIRTY; 1194 else 1195 bp->b_qindex = QUEUE_CLEAN; 1196 if (bp->b_flags & B_AGE) 1197 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1198 else 1199 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1200 } 1201 1202 /* 1203 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1204 * on the correct queue. 1205 */ 1206 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) { 1207 bp->b_flags &= ~B_DELWRI; 1208 --numdirtybuffers; 1209 numdirtywakeup(lodirtybuffers); 1210 } 1211 1212 /* 1213 * Fixup numfreebuffers count. The bp is on an appropriate queue 1214 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1215 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1216 * if B_INVAL is set ). 1217 */ 1218 1219 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1220 bufcountwakeup(); 1221 1222 /* 1223 * Something we can maybe free or reuse 1224 */ 1225 if (bp->b_bufsize || bp->b_kvasize) 1226 bufspacewakeup(); 1227 1228 /* unlock */ 1229 BUF_UNLOCK(bp); 1230 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1231 bp->b_ioflags &= ~BIO_ORDERED; 1232 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1233 panic("brelse: not dirty"); 1234 splx(s); 1235} 1236 1237/* 1238 * Release a buffer back to the appropriate queue but do not try to free 1239 * it. The buffer is expected to be used again soon. 1240 * 1241 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1242 * biodone() to requeue an async I/O on completion. It is also used when 1243 * known good buffers need to be requeued but we think we may need the data 1244 * again soon. 1245 */ 1246void 1247bqrelse(struct buf * bp) 1248{ 1249 int s; 1250 1251 s = splbio(); 1252 1253 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1254 1255 if (bp->b_qindex != QUEUE_NONE) 1256 panic("bqrelse: free buffer onto another queue???"); 1257 if (BUF_REFCNT(bp) > 1) { 1258 /* do not release to free list */ 1259 BUF_UNLOCK(bp); 1260 splx(s); 1261 return; 1262 } 1263 if (bp->b_flags & B_LOCKED) { 1264 bp->b_ioflags &= ~BIO_ERROR; 1265 bp->b_qindex = QUEUE_LOCKED; 1266 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1267 /* buffers with stale but valid contents */ 1268 } else if (bp->b_flags & B_DELWRI) { 1269 bp->b_qindex = QUEUE_DIRTY; 1270 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1271 } else if (vm_page_count_severe()) { 1272 /* 1273 * We are too low on memory, we have to try to free the 1274 * buffer (most importantly: the wired pages making up its 1275 * backing store) *now*. 1276 */ 1277 splx(s); 1278 brelse(bp); 1279 return; 1280 } else { 1281 bp->b_qindex = QUEUE_CLEAN; 1282 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1283 } 1284 1285 if ((bp->b_flags & B_LOCKED) == 0 && 1286 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1287 bufcountwakeup(); 1288 } 1289 1290 /* 1291 * Something we can maybe free or reuse. 1292 */ 1293 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1294 bufspacewakeup(); 1295 1296 /* unlock */ 1297 BUF_UNLOCK(bp); 1298 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1299 bp->b_ioflags &= ~BIO_ORDERED; 1300 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1301 panic("bqrelse: not dirty"); 1302 splx(s); 1303} 1304 1305static void 1306vfs_vmio_release(bp) 1307 struct buf *bp; 1308{ 1309 int i, s; 1310 vm_page_t m; 1311 1312 s = splvm(); 1313 for (i = 0; i < bp->b_npages; i++) { 1314 m = bp->b_pages[i]; 1315 bp->b_pages[i] = NULL; 1316 /* 1317 * In order to keep page LRU ordering consistent, put 1318 * everything on the inactive queue. 1319 */ 1320 vm_page_unwire(m, 0); 1321 /* 1322 * We don't mess with busy pages, it is 1323 * the responsibility of the process that 1324 * busied the pages to deal with them. 1325 */ 1326 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1327 continue; 1328 1329 if (m->wire_count == 0) { 1330 vm_page_flag_clear(m, PG_ZERO); 1331 /* 1332 * Might as well free the page if we can and it has 1333 * no valid data. 1334 */ 1335 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { 1336 vm_page_busy(m); 1337 vm_page_protect(m, VM_PROT_NONE); 1338 vm_page_free(m); 1339 } else if (vm_page_count_severe()) { 1340 vm_page_try_to_cache(m); 1341 } 1342 } 1343 } 1344 splx(s); 1345 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1346 if (bp->b_bufsize) { 1347 bufspacewakeup(); 1348 bp->b_bufsize = 0; 1349 } 1350 bp->b_npages = 0; 1351 bp->b_flags &= ~B_VMIO; 1352 if (bp->b_vp) 1353 brelvp(bp); 1354} 1355 1356/* 1357 * Check to see if a block is currently memory resident. 1358 */ 1359struct buf * 1360gbincore(struct vnode * vp, daddr_t blkno) 1361{ 1362 struct buf *bp; 1363 struct bufhashhdr *bh; 1364 1365 bh = bufhash(vp, blkno); 1366 1367 /* Search hash chain */ 1368 LIST_FOREACH(bp, bh, b_hash) { 1369 /* hit */ 1370 if (bp->b_vp == vp && bp->b_lblkno == blkno && 1371 (bp->b_flags & B_INVAL) == 0) { 1372 break; 1373 } 1374 } 1375 return (bp); 1376} 1377 1378/* 1379 * vfs_bio_awrite: 1380 * 1381 * Implement clustered async writes for clearing out B_DELWRI buffers. 1382 * This is much better then the old way of writing only one buffer at 1383 * a time. Note that we may not be presented with the buffers in the 1384 * correct order, so we search for the cluster in both directions. 1385 */ 1386int 1387vfs_bio_awrite(struct buf * bp) 1388{ 1389 int i; 1390 int j; 1391 daddr_t lblkno = bp->b_lblkno; 1392 struct vnode *vp = bp->b_vp; 1393 int s; 1394 int ncl; 1395 struct buf *bpa; 1396 int nwritten; 1397 int size; 1398 int maxcl; 1399 1400 s = splbio(); 1401 /* 1402 * right now we support clustered writing only to regular files. If 1403 * we find a clusterable block we could be in the middle of a cluster 1404 * rather then at the beginning. 1405 */ 1406 if ((vp->v_type == VREG) && 1407 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1408 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1409 1410 size = vp->v_mount->mnt_stat.f_iosize; 1411 maxcl = MAXPHYS / size; 1412 1413 for (i = 1; i < maxcl; i++) { 1414 if ((bpa = gbincore(vp, lblkno + i)) && 1415 BUF_REFCNT(bpa) == 0 && 1416 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1417 (B_DELWRI | B_CLUSTEROK)) && 1418 (bpa->b_bufsize == size)) { 1419 if ((bpa->b_blkno == bpa->b_lblkno) || 1420 (bpa->b_blkno != 1421 bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 1422 break; 1423 } else { 1424 break; 1425 } 1426 } 1427 for (j = 1; i + j <= maxcl && j <= lblkno; j++) { 1428 if ((bpa = gbincore(vp, lblkno - j)) && 1429 BUF_REFCNT(bpa) == 0 && 1430 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1431 (B_DELWRI | B_CLUSTEROK)) && 1432 (bpa->b_bufsize == size)) { 1433 if ((bpa->b_blkno == bpa->b_lblkno) || 1434 (bpa->b_blkno != 1435 bp->b_blkno - ((j * size) >> DEV_BSHIFT))) 1436 break; 1437 } else { 1438 break; 1439 } 1440 } 1441 --j; 1442 ncl = i + j; 1443 /* 1444 * this is a possible cluster write 1445 */ 1446 if (ncl != 1) { 1447 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1448 splx(s); 1449 return nwritten; 1450 } 1451 } 1452 1453 BUF_LOCK(bp, LK_EXCLUSIVE); 1454 bremfree(bp); 1455 bp->b_flags |= B_ASYNC; 1456 1457 splx(s); 1458 /* 1459 * default (old) behavior, writing out only one block 1460 * 1461 * XXX returns b_bufsize instead of b_bcount for nwritten? 1462 */ 1463 nwritten = bp->b_bufsize; 1464 (void) BUF_WRITE(bp); 1465 1466 return nwritten; 1467} 1468 1469/* 1470 * getnewbuf: 1471 * 1472 * Find and initialize a new buffer header, freeing up existing buffers 1473 * in the bufqueues as necessary. The new buffer is returned locked. 1474 * 1475 * Important: B_INVAL is not set. If the caller wishes to throw the 1476 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1477 * 1478 * We block if: 1479 * We have insufficient buffer headers 1480 * We have insufficient buffer space 1481 * buffer_map is too fragmented ( space reservation fails ) 1482 * If we have to flush dirty buffers ( but we try to avoid this ) 1483 * 1484 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1485 * Instead we ask the buf daemon to do it for us. We attempt to 1486 * avoid piecemeal wakeups of the pageout daemon. 1487 */ 1488 1489static struct buf * 1490getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1491{ 1492 struct buf *bp; 1493 struct buf *nbp; 1494 int defrag = 0; 1495 int nqindex; 1496 static int flushingbufs; 1497 1498 /* 1499 * We can't afford to block since we might be holding a vnode lock, 1500 * which may prevent system daemons from running. We deal with 1501 * low-memory situations by proactively returning memory and running 1502 * async I/O rather then sync I/O. 1503 */ 1504 1505 ++getnewbufcalls; 1506 --getnewbufrestarts; 1507restart: 1508 ++getnewbufrestarts; 1509 1510 /* 1511 * Setup for scan. If we do not have enough free buffers, 1512 * we setup a degenerate case that immediately fails. Note 1513 * that if we are specially marked process, we are allowed to 1514 * dip into our reserves. 1515 * 1516 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1517 * 1518 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1519 * However, there are a number of cases (defragging, reusing, ...) 1520 * where we cannot backup. 1521 */ 1522 nqindex = QUEUE_EMPTYKVA; 1523 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1524 1525 if (nbp == NULL) { 1526 /* 1527 * If no EMPTYKVA buffers and we are either 1528 * defragging or reusing, locate a CLEAN buffer 1529 * to free or reuse. If bufspace useage is low 1530 * skip this step so we can allocate a new buffer. 1531 */ 1532 if (defrag || bufspace >= lobufspace) { 1533 nqindex = QUEUE_CLEAN; 1534 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1535 } 1536 1537 /* 1538 * If we could not find or were not allowed to reuse a 1539 * CLEAN buffer, check to see if it is ok to use an EMPTY 1540 * buffer. We can only use an EMPTY buffer if allocating 1541 * its KVA would not otherwise run us out of buffer space. 1542 */ 1543 if (nbp == NULL && defrag == 0 && 1544 bufspace + maxsize < hibufspace) { 1545 nqindex = QUEUE_EMPTY; 1546 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1547 } 1548 } 1549 1550 /* 1551 * Run scan, possibly freeing data and/or kva mappings on the fly 1552 * depending. 1553 */ 1554 1555 while ((bp = nbp) != NULL) { 1556 int qindex = nqindex; 1557 1558 /* 1559 * Calculate next bp ( we can only use it if we do not block 1560 * or do other fancy things ). 1561 */ 1562 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1563 switch(qindex) { 1564 case QUEUE_EMPTY: 1565 nqindex = QUEUE_EMPTYKVA; 1566 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1567 break; 1568 /* fall through */ 1569 case QUEUE_EMPTYKVA: 1570 nqindex = QUEUE_CLEAN; 1571 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1572 break; 1573 /* fall through */ 1574 case QUEUE_CLEAN: 1575 /* 1576 * nbp is NULL. 1577 */ 1578 break; 1579 } 1580 } 1581 1582 /* 1583 * Sanity Checks 1584 */ 1585 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1586 1587 /* 1588 * Note: we no longer distinguish between VMIO and non-VMIO 1589 * buffers. 1590 */ 1591 1592 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1593 1594 /* 1595 * If we are defragging then we need a buffer with 1596 * b_kvasize != 0. XXX this situation should no longer 1597 * occur, if defrag is non-zero the buffer's b_kvasize 1598 * should also be non-zero at this point. XXX 1599 */ 1600 if (defrag && bp->b_kvasize == 0) { 1601 printf("Warning: defrag empty buffer %p\n", bp); 1602 continue; 1603 } 1604 1605 /* 1606 * Start freeing the bp. This is somewhat involved. nbp 1607 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1608 */ 1609 1610 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1611 panic("getnewbuf: locked buf"); 1612 bremfree(bp); 1613 1614 if (qindex == QUEUE_CLEAN) { 1615 if (bp->b_flags & B_VMIO) { 1616 bp->b_flags &= ~B_ASYNC; 1617 vfs_vmio_release(bp); 1618 } 1619 if (bp->b_vp) 1620 brelvp(bp); 1621 } 1622 1623 /* 1624 * NOTE: nbp is now entirely invalid. We can only restart 1625 * the scan from this point on. 1626 * 1627 * Get the rest of the buffer freed up. b_kva* is still 1628 * valid after this operation. 1629 */ 1630 1631 if (bp->b_rcred != NOCRED) { 1632 crfree(bp->b_rcred); 1633 bp->b_rcred = NOCRED; 1634 } 1635 if (bp->b_wcred != NOCRED) { 1636 crfree(bp->b_wcred); 1637 bp->b_wcred = NOCRED; 1638 } 1639 if (LIST_FIRST(&bp->b_dep) != NULL) 1640 buf_deallocate(bp); 1641 if (bp->b_xflags & BX_BKGRDINPROG) 1642 panic("losing buffer 3"); 1643 LIST_REMOVE(bp, b_hash); 1644 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1645 1646 if (bp->b_bufsize) 1647 allocbuf(bp, 0); 1648 1649 bp->b_flags = 0; 1650 bp->b_ioflags = 0; 1651 bp->b_xflags = 0; 1652 bp->b_dev = NODEV; 1653 bp->b_vp = NULL; 1654 bp->b_blkno = bp->b_lblkno = 0; 1655 bp->b_offset = NOOFFSET; 1656 bp->b_iodone = 0; 1657 bp->b_error = 0; 1658 bp->b_resid = 0; 1659 bp->b_bcount = 0; 1660 bp->b_npages = 0; 1661 bp->b_dirtyoff = bp->b_dirtyend = 0; 1662 bp->b_magic = B_MAGIC_BIO; 1663 bp->b_op = &buf_ops_bio; 1664 1665 LIST_INIT(&bp->b_dep); 1666 1667 /* 1668 * If we are defragging then free the buffer. 1669 */ 1670 if (defrag) { 1671 bp->b_flags |= B_INVAL; 1672 bfreekva(bp); 1673 brelse(bp); 1674 defrag = 0; 1675 goto restart; 1676 } 1677 1678 /* 1679 * If we are overcomitted then recover the buffer and its 1680 * KVM space. This occurs in rare situations when multiple 1681 * processes are blocked in getnewbuf() or allocbuf(). 1682 */ 1683 if (bufspace >= hibufspace) 1684 flushingbufs = 1; 1685 if (flushingbufs && bp->b_kvasize != 0) { 1686 bp->b_flags |= B_INVAL; 1687 bfreekva(bp); 1688 brelse(bp); 1689 goto restart; 1690 } 1691 if (bufspace < lobufspace) 1692 flushingbufs = 0; 1693 break; 1694 } 1695 1696 /* 1697 * If we exhausted our list, sleep as appropriate. We may have to 1698 * wakeup various daemons and write out some dirty buffers. 1699 * 1700 * Generally we are sleeping due to insufficient buffer space. 1701 */ 1702 1703 if (bp == NULL) { 1704 int flags; 1705 char *waitmsg; 1706 1707 if (defrag) { 1708 flags = VFS_BIO_NEED_BUFSPACE; 1709 waitmsg = "nbufkv"; 1710 } else if (bufspace >= hibufspace) { 1711 waitmsg = "nbufbs"; 1712 flags = VFS_BIO_NEED_BUFSPACE; 1713 } else { 1714 waitmsg = "newbuf"; 1715 flags = VFS_BIO_NEED_ANY; 1716 } 1717 1718 bd_speedup(); /* heeeelp */ 1719 1720 needsbuffer |= flags; 1721 while (needsbuffer & flags) { 1722 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, 1723 waitmsg, slptimeo)) 1724 return (NULL); 1725 } 1726 } else { 1727 /* 1728 * We finally have a valid bp. We aren't quite out of the 1729 * woods, we still have to reserve kva space. In order 1730 * to keep fragmentation sane we only allocate kva in 1731 * BKVASIZE chunks. 1732 */ 1733 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1734 1735 if (maxsize != bp->b_kvasize) { 1736 vm_offset_t addr = 0; 1737 1738 bfreekva(bp); 1739 1740 if (vm_map_findspace(buffer_map, 1741 vm_map_min(buffer_map), maxsize, &addr)) { 1742 /* 1743 * Uh oh. Buffer map is to fragmented. We 1744 * must defragment the map. 1745 */ 1746 ++bufdefragcnt; 1747 defrag = 1; 1748 bp->b_flags |= B_INVAL; 1749 brelse(bp); 1750 goto restart; 1751 } 1752 if (addr) { 1753 vm_map_insert(buffer_map, NULL, 0, 1754 addr, addr + maxsize, 1755 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1756 1757 bp->b_kvabase = (caddr_t) addr; 1758 bp->b_kvasize = maxsize; 1759 bufspace += bp->b_kvasize; 1760 ++bufreusecnt; 1761 } 1762 } 1763 bp->b_data = bp->b_kvabase; 1764 } 1765 return(bp); 1766} 1767 1768/* 1769 * buf_daemon: 1770 * 1771 * buffer flushing daemon. Buffers are normally flushed by the 1772 * update daemon but if it cannot keep up this process starts to 1773 * take the load in an attempt to prevent getnewbuf() from blocking. 1774 */ 1775 1776static struct proc *bufdaemonproc; 1777 1778static struct kproc_desc buf_kp = { 1779 "bufdaemon", 1780 buf_daemon, 1781 &bufdaemonproc 1782}; 1783SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1784 1785static void 1786buf_daemon() 1787{ 1788 int s; 1789 1790 mtx_lock(&Giant); 1791 1792 /* 1793 * This process needs to be suspended prior to shutdown sync. 1794 */ 1795 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 1796 SHUTDOWN_PRI_LAST); 1797 1798 /* 1799 * This process is allowed to take the buffer cache to the limit 1800 */ 1801 curproc->p_flag |= P_BUFEXHAUST; 1802 s = splbio(); 1803 1804 for (;;) { 1805 kthread_suspend_check(bufdaemonproc); 1806 1807 bd_request = 0; 1808 1809 /* 1810 * Do the flush. Limit the amount of in-transit I/O we 1811 * allow to build up, otherwise we would completely saturate 1812 * the I/O system. Wakeup any waiting processes before we 1813 * normally would so they can run in parallel with our drain. 1814 */ 1815 while (numdirtybuffers > lodirtybuffers) { 1816 if (flushbufqueues() == 0) 1817 break; 1818 waitrunningbufspace(); 1819 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 1820 } 1821 1822 /* 1823 * Only clear bd_request if we have reached our low water 1824 * mark. The buf_daemon normally waits 5 seconds and 1825 * then incrementally flushes any dirty buffers that have 1826 * built up, within reason. 1827 * 1828 * If we were unable to hit our low water mark and couldn't 1829 * find any flushable buffers, we sleep half a second. 1830 * Otherwise we loop immediately. 1831 */ 1832 if (numdirtybuffers <= lodirtybuffers) { 1833 /* 1834 * We reached our low water mark, reset the 1835 * request and sleep until we are needed again. 1836 * The sleep is just so the suspend code works. 1837 */ 1838 bd_request = 0; 1839 tsleep(&bd_request, PVM, "psleep", hz); 1840 } else { 1841 /* 1842 * We couldn't find any flushable dirty buffers but 1843 * still have too many dirty buffers, we 1844 * have to sleep and try again. (rare) 1845 */ 1846 tsleep(&bd_request, PVM, "qsleep", hz / 2); 1847 } 1848 } 1849} 1850 1851/* 1852 * flushbufqueues: 1853 * 1854 * Try to flush a buffer in the dirty queue. We must be careful to 1855 * free up B_INVAL buffers instead of write them, which NFS is 1856 * particularly sensitive to. 1857 */ 1858 1859static int 1860flushbufqueues(void) 1861{ 1862 struct buf *bp; 1863 int r = 0; 1864 1865 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1866 1867 while (bp) { 1868 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); 1869 if ((bp->b_flags & B_DELWRI) != 0 && 1870 (bp->b_xflags & BX_BKGRDINPROG) == 0) { 1871 if (bp->b_flags & B_INVAL) { 1872 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1873 panic("flushbufqueues: locked buf"); 1874 bremfree(bp); 1875 brelse(bp); 1876 ++r; 1877 break; 1878 } 1879 if (LIST_FIRST(&bp->b_dep) != NULL && 1880 (bp->b_flags & B_DEFERRED) == 0 && 1881 buf_countdeps(bp, 0)) { 1882 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], 1883 bp, b_freelist); 1884 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], 1885 bp, b_freelist); 1886 bp->b_flags |= B_DEFERRED; 1887 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1888 continue; 1889 } 1890 vfs_bio_awrite(bp); 1891 ++r; 1892 break; 1893 } 1894 bp = TAILQ_NEXT(bp, b_freelist); 1895 } 1896 return (r); 1897} 1898 1899/* 1900 * Check to see if a block is currently memory resident. 1901 */ 1902struct buf * 1903incore(struct vnode * vp, daddr_t blkno) 1904{ 1905 struct buf *bp; 1906 1907 int s = splbio(); 1908 bp = gbincore(vp, blkno); 1909 splx(s); 1910 return (bp); 1911} 1912 1913/* 1914 * Returns true if no I/O is needed to access the 1915 * associated VM object. This is like incore except 1916 * it also hunts around in the VM system for the data. 1917 */ 1918 1919int 1920inmem(struct vnode * vp, daddr_t blkno) 1921{ 1922 vm_object_t obj; 1923 vm_offset_t toff, tinc, size; 1924 vm_page_t m; 1925 vm_ooffset_t off; 1926 1927 if (incore(vp, blkno)) 1928 return 1; 1929 if (vp->v_mount == NULL) 1930 return 0; 1931 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_flag & VOBJBUF) == 0) 1932 return 0; 1933 1934 size = PAGE_SIZE; 1935 if (size > vp->v_mount->mnt_stat.f_iosize) 1936 size = vp->v_mount->mnt_stat.f_iosize; 1937 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 1938 1939 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1940 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 1941 if (!m) 1942 return 0; 1943 tinc = size; 1944 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 1945 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 1946 if (vm_page_is_valid(m, 1947 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 1948 return 0; 1949 } 1950 return 1; 1951} 1952 1953/* 1954 * vfs_setdirty: 1955 * 1956 * Sets the dirty range for a buffer based on the status of the dirty 1957 * bits in the pages comprising the buffer. 1958 * 1959 * The range is limited to the size of the buffer. 1960 * 1961 * This routine is primarily used by NFS, but is generalized for the 1962 * B_VMIO case. 1963 */ 1964static void 1965vfs_setdirty(struct buf *bp) 1966{ 1967 int i; 1968 vm_object_t object; 1969 1970 /* 1971 * Degenerate case - empty buffer 1972 */ 1973 1974 if (bp->b_bufsize == 0) 1975 return; 1976 1977 /* 1978 * We qualify the scan for modified pages on whether the 1979 * object has been flushed yet. The OBJ_WRITEABLE flag 1980 * is not cleared simply by protecting pages off. 1981 */ 1982 1983 if ((bp->b_flags & B_VMIO) == 0) 1984 return; 1985 1986 object = bp->b_pages[0]->object; 1987 1988 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 1989 printf("Warning: object %p writeable but not mightbedirty\n", object); 1990 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 1991 printf("Warning: object %p mightbedirty but not writeable\n", object); 1992 1993 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 1994 vm_offset_t boffset; 1995 vm_offset_t eoffset; 1996 1997 /* 1998 * test the pages to see if they have been modified directly 1999 * by users through the VM system. 2000 */ 2001 for (i = 0; i < bp->b_npages; i++) { 2002 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2003 vm_page_test_dirty(bp->b_pages[i]); 2004 } 2005 2006 /* 2007 * Calculate the encompassing dirty range, boffset and eoffset, 2008 * (eoffset - boffset) bytes. 2009 */ 2010 2011 for (i = 0; i < bp->b_npages; i++) { 2012 if (bp->b_pages[i]->dirty) 2013 break; 2014 } 2015 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2016 2017 for (i = bp->b_npages - 1; i >= 0; --i) { 2018 if (bp->b_pages[i]->dirty) { 2019 break; 2020 } 2021 } 2022 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2023 2024 /* 2025 * Fit it to the buffer. 2026 */ 2027 2028 if (eoffset > bp->b_bcount) 2029 eoffset = bp->b_bcount; 2030 2031 /* 2032 * If we have a good dirty range, merge with the existing 2033 * dirty range. 2034 */ 2035 2036 if (boffset < eoffset) { 2037 if (bp->b_dirtyoff > boffset) 2038 bp->b_dirtyoff = boffset; 2039 if (bp->b_dirtyend < eoffset) 2040 bp->b_dirtyend = eoffset; 2041 } 2042 } 2043} 2044 2045/* 2046 * getblk: 2047 * 2048 * Get a block given a specified block and offset into a file/device. 2049 * The buffers B_DONE bit will be cleared on return, making it almost 2050 * ready for an I/O initiation. B_INVAL may or may not be set on 2051 * return. The caller should clear B_INVAL prior to initiating a 2052 * READ. 2053 * 2054 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2055 * an existing buffer. 2056 * 2057 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2058 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2059 * and then cleared based on the backing VM. If the previous buffer is 2060 * non-0-sized but invalid, B_CACHE will be cleared. 2061 * 2062 * If getblk() must create a new buffer, the new buffer is returned with 2063 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2064 * case it is returned with B_INVAL clear and B_CACHE set based on the 2065 * backing VM. 2066 * 2067 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2068 * B_CACHE bit is clear. 2069 * 2070 * What this means, basically, is that the caller should use B_CACHE to 2071 * determine whether the buffer is fully valid or not and should clear 2072 * B_INVAL prior to issuing a read. If the caller intends to validate 2073 * the buffer by loading its data area with something, the caller needs 2074 * to clear B_INVAL. If the caller does this without issuing an I/O, 2075 * the caller should set B_CACHE ( as an optimization ), else the caller 2076 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2077 * a write attempt or if it was a successfull read. If the caller 2078 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2079 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2080 */ 2081struct buf * 2082getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2083{ 2084 struct buf *bp; 2085 int s; 2086 struct bufhashhdr *bh; 2087 2088 if (size > MAXBSIZE) 2089 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2090 2091 s = splbio(); 2092loop: 2093 /* 2094 * Block if we are low on buffers. Certain processes are allowed 2095 * to completely exhaust the buffer cache. 2096 * 2097 * If this check ever becomes a bottleneck it may be better to 2098 * move it into the else, when gbincore() fails. At the moment 2099 * it isn't a problem. 2100 * 2101 * XXX remove if 0 sections (clean this up after its proven) 2102 */ 2103 if (numfreebuffers == 0) { 2104 if (curproc == PCPU_GET(idleproc)) 2105 return NULL; 2106 needsbuffer |= VFS_BIO_NEED_ANY; 2107 } 2108 2109 if ((bp = gbincore(vp, blkno))) { 2110 /* 2111 * Buffer is in-core. If the buffer is not busy, it must 2112 * be on a queue. 2113 */ 2114 2115 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2116 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, 2117 "getblk", slpflag, slptimeo) == ENOLCK) 2118 goto loop; 2119 splx(s); 2120 return (struct buf *) NULL; 2121 } 2122 2123 /* 2124 * The buffer is locked. B_CACHE is cleared if the buffer is 2125 * invalid. Ohterwise, for a non-VMIO buffer, B_CACHE is set 2126 * and for a VMIO buffer B_CACHE is adjusted according to the 2127 * backing VM cache. 2128 */ 2129 if (bp->b_flags & B_INVAL) 2130 bp->b_flags &= ~B_CACHE; 2131 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2132 bp->b_flags |= B_CACHE; 2133 bremfree(bp); 2134 2135 /* 2136 * check for size inconsistancies for non-VMIO case. 2137 */ 2138 2139 if (bp->b_bcount != size) { 2140 if ((bp->b_flags & B_VMIO) == 0 || 2141 (size > bp->b_kvasize)) { 2142 if (bp->b_flags & B_DELWRI) { 2143 bp->b_flags |= B_NOCACHE; 2144 BUF_WRITE(bp); 2145 } else { 2146 if ((bp->b_flags & B_VMIO) && 2147 (LIST_FIRST(&bp->b_dep) == NULL)) { 2148 bp->b_flags |= B_RELBUF; 2149 brelse(bp); 2150 } else { 2151 bp->b_flags |= B_NOCACHE; 2152 BUF_WRITE(bp); 2153 } 2154 } 2155 goto loop; 2156 } 2157 } 2158 2159 /* 2160 * If the size is inconsistant in the VMIO case, we can resize 2161 * the buffer. This might lead to B_CACHE getting set or 2162 * cleared. If the size has not changed, B_CACHE remains 2163 * unchanged from its previous state. 2164 */ 2165 2166 if (bp->b_bcount != size) 2167 allocbuf(bp, size); 2168 2169 KASSERT(bp->b_offset != NOOFFSET, 2170 ("getblk: no buffer offset")); 2171 2172 /* 2173 * A buffer with B_DELWRI set and B_CACHE clear must 2174 * be committed before we can return the buffer in 2175 * order to prevent the caller from issuing a read 2176 * ( due to B_CACHE not being set ) and overwriting 2177 * it. 2178 * 2179 * Most callers, including NFS and FFS, need this to 2180 * operate properly either because they assume they 2181 * can issue a read if B_CACHE is not set, or because 2182 * ( for example ) an uncached B_DELWRI might loop due 2183 * to softupdates re-dirtying the buffer. In the latter 2184 * case, B_CACHE is set after the first write completes, 2185 * preventing further loops. 2186 */ 2187 2188 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2189 BUF_WRITE(bp); 2190 goto loop; 2191 } 2192 2193 splx(s); 2194 bp->b_flags &= ~B_DONE; 2195 } else { 2196 /* 2197 * Buffer is not in-core, create new buffer. The buffer 2198 * returned by getnewbuf() is locked. Note that the returned 2199 * buffer is also considered valid (not marked B_INVAL). 2200 */ 2201 int bsize, maxsize, vmio; 2202 off_t offset; 2203 2204 if (vn_isdisk(vp, NULL)) 2205 bsize = DEV_BSIZE; 2206 else if (vp->v_mountedhere) 2207 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2208 else if (vp->v_mount) 2209 bsize = vp->v_mount->mnt_stat.f_iosize; 2210 else 2211 bsize = size; 2212 2213 offset = (off_t)blkno * bsize; 2214 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && (vp->v_flag & VOBJBUF); 2215 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2216 maxsize = imax(maxsize, bsize); 2217 2218 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2219 if (slpflag || slptimeo) { 2220 splx(s); 2221 return NULL; 2222 } 2223 goto loop; 2224 } 2225 2226 /* 2227 * This code is used to make sure that a buffer is not 2228 * created while the getnewbuf routine is blocked. 2229 * This can be a problem whether the vnode is locked or not. 2230 * If the buffer is created out from under us, we have to 2231 * throw away the one we just created. There is now window 2232 * race because we are safely running at splbio() from the 2233 * point of the duplicate buffer creation through to here, 2234 * and we've locked the buffer. 2235 */ 2236 if (gbincore(vp, blkno)) { 2237 bp->b_flags |= B_INVAL; 2238 brelse(bp); 2239 goto loop; 2240 } 2241 2242 /* 2243 * Insert the buffer into the hash, so that it can 2244 * be found by incore. 2245 */ 2246 bp->b_blkno = bp->b_lblkno = blkno; 2247 bp->b_offset = offset; 2248 2249 bgetvp(vp, bp); 2250 LIST_REMOVE(bp, b_hash); 2251 bh = bufhash(vp, blkno); 2252 LIST_INSERT_HEAD(bh, bp, b_hash); 2253 2254 /* 2255 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2256 * buffer size starts out as 0, B_CACHE will be set by 2257 * allocbuf() for the VMIO case prior to it testing the 2258 * backing store for validity. 2259 */ 2260 2261 if (vmio) { 2262 bp->b_flags |= B_VMIO; 2263#if defined(VFS_BIO_DEBUG) 2264 if (vp->v_type != VREG) 2265 printf("getblk: vmioing file type %d???\n", vp->v_type); 2266#endif 2267 } else { 2268 bp->b_flags &= ~B_VMIO; 2269 } 2270 2271 allocbuf(bp, size); 2272 2273 splx(s); 2274 bp->b_flags &= ~B_DONE; 2275 } 2276 return (bp); 2277} 2278 2279/* 2280 * Get an empty, disassociated buffer of given size. The buffer is initially 2281 * set to B_INVAL. 2282 */ 2283struct buf * 2284geteblk(int size) 2285{ 2286 struct buf *bp; 2287 int s; 2288 int maxsize; 2289 2290 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2291 2292 s = splbio(); 2293 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0); 2294 splx(s); 2295 allocbuf(bp, size); 2296 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2297 return (bp); 2298} 2299 2300 2301/* 2302 * This code constitutes the buffer memory from either anonymous system 2303 * memory (in the case of non-VMIO operations) or from an associated 2304 * VM object (in the case of VMIO operations). This code is able to 2305 * resize a buffer up or down. 2306 * 2307 * Note that this code is tricky, and has many complications to resolve 2308 * deadlock or inconsistant data situations. Tread lightly!!! 2309 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2310 * the caller. Calling this code willy nilly can result in the loss of data. 2311 * 2312 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2313 * B_CACHE for the non-VMIO case. 2314 */ 2315 2316int 2317allocbuf(struct buf *bp, int size) 2318{ 2319 int newbsize, mbsize; 2320 int i; 2321 2322 if (BUF_REFCNT(bp) == 0) 2323 panic("allocbuf: buffer not busy"); 2324 2325 if (bp->b_kvasize < size) 2326 panic("allocbuf: buffer too small"); 2327 2328 if ((bp->b_flags & B_VMIO) == 0) { 2329 caddr_t origbuf; 2330 int origbufsize; 2331 /* 2332 * Just get anonymous memory from the kernel. Don't 2333 * mess with B_CACHE. 2334 */ 2335 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2336#if !defined(NO_B_MALLOC) 2337 if (bp->b_flags & B_MALLOC) 2338 newbsize = mbsize; 2339 else 2340#endif 2341 newbsize = round_page(size); 2342 2343 if (newbsize < bp->b_bufsize) { 2344#if !defined(NO_B_MALLOC) 2345 /* 2346 * malloced buffers are not shrunk 2347 */ 2348 if (bp->b_flags & B_MALLOC) { 2349 if (newbsize) { 2350 bp->b_bcount = size; 2351 } else { 2352 free(bp->b_data, M_BIOBUF); 2353 if (bp->b_bufsize) { 2354 bufmallocspace -= bp->b_bufsize; 2355 bufspacewakeup(); 2356 bp->b_bufsize = 0; 2357 } 2358 bp->b_data = bp->b_kvabase; 2359 bp->b_bcount = 0; 2360 bp->b_flags &= ~B_MALLOC; 2361 } 2362 return 1; 2363 } 2364#endif 2365 vm_hold_free_pages( 2366 bp, 2367 (vm_offset_t) bp->b_data + newbsize, 2368 (vm_offset_t) bp->b_data + bp->b_bufsize); 2369 } else if (newbsize > bp->b_bufsize) { 2370#if !defined(NO_B_MALLOC) 2371 /* 2372 * We only use malloced memory on the first allocation. 2373 * and revert to page-allocated memory when the buffer 2374 * grows. 2375 */ 2376 if ( (bufmallocspace < maxbufmallocspace) && 2377 (bp->b_bufsize == 0) && 2378 (mbsize <= PAGE_SIZE/2)) { 2379 2380 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2381 bp->b_bufsize = mbsize; 2382 bp->b_bcount = size; 2383 bp->b_flags |= B_MALLOC; 2384 bufmallocspace += mbsize; 2385 return 1; 2386 } 2387#endif 2388 origbuf = NULL; 2389 origbufsize = 0; 2390#if !defined(NO_B_MALLOC) 2391 /* 2392 * If the buffer is growing on its other-than-first allocation, 2393 * then we revert to the page-allocation scheme. 2394 */ 2395 if (bp->b_flags & B_MALLOC) { 2396 origbuf = bp->b_data; 2397 origbufsize = bp->b_bufsize; 2398 bp->b_data = bp->b_kvabase; 2399 if (bp->b_bufsize) { 2400 bufmallocspace -= bp->b_bufsize; 2401 bufspacewakeup(); 2402 bp->b_bufsize = 0; 2403 } 2404 bp->b_flags &= ~B_MALLOC; 2405 newbsize = round_page(newbsize); 2406 } 2407#endif 2408 vm_hold_load_pages( 2409 bp, 2410 (vm_offset_t) bp->b_data + bp->b_bufsize, 2411 (vm_offset_t) bp->b_data + newbsize); 2412#if !defined(NO_B_MALLOC) 2413 if (origbuf) { 2414 bcopy(origbuf, bp->b_data, origbufsize); 2415 free(origbuf, M_BIOBUF); 2416 } 2417#endif 2418 } 2419 } else { 2420 vm_page_t m; 2421 int desiredpages; 2422 2423 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2424 desiredpages = (size == 0) ? 0 : 2425 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2426 2427#if !defined(NO_B_MALLOC) 2428 if (bp->b_flags & B_MALLOC) 2429 panic("allocbuf: VMIO buffer can't be malloced"); 2430#endif 2431 /* 2432 * Set B_CACHE initially if buffer is 0 length or will become 2433 * 0-length. 2434 */ 2435 if (size == 0 || bp->b_bufsize == 0) 2436 bp->b_flags |= B_CACHE; 2437 2438 if (newbsize < bp->b_bufsize) { 2439 /* 2440 * DEV_BSIZE aligned new buffer size is less then the 2441 * DEV_BSIZE aligned existing buffer size. Figure out 2442 * if we have to remove any pages. 2443 */ 2444 if (desiredpages < bp->b_npages) { 2445 for (i = desiredpages; i < bp->b_npages; i++) { 2446 /* 2447 * the page is not freed here -- it 2448 * is the responsibility of 2449 * vnode_pager_setsize 2450 */ 2451 m = bp->b_pages[i]; 2452 KASSERT(m != bogus_page, 2453 ("allocbuf: bogus page found")); 2454 while (vm_page_sleep_busy(m, TRUE, "biodep")) 2455 ; 2456 2457 bp->b_pages[i] = NULL; 2458 vm_page_unwire(m, 0); 2459 } 2460 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2461 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2462 bp->b_npages = desiredpages; 2463 } 2464 } else if (size > bp->b_bcount) { 2465 /* 2466 * We are growing the buffer, possibly in a 2467 * byte-granular fashion. 2468 */ 2469 struct vnode *vp; 2470 vm_object_t obj; 2471 vm_offset_t toff; 2472 vm_offset_t tinc; 2473 2474 /* 2475 * Step 1, bring in the VM pages from the object, 2476 * allocating them if necessary. We must clear 2477 * B_CACHE if these pages are not valid for the 2478 * range covered by the buffer. 2479 */ 2480 2481 vp = bp->b_vp; 2482 VOP_GETVOBJECT(vp, &obj); 2483 2484 while (bp->b_npages < desiredpages) { 2485 vm_page_t m; 2486 vm_pindex_t pi; 2487 2488 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2489 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2490 /* 2491 * note: must allocate system pages 2492 * since blocking here could intefere 2493 * with paging I/O, no matter which 2494 * process we are. 2495 */ 2496 m = vm_page_alloc(obj, pi, VM_ALLOC_SYSTEM); 2497 if (m == NULL) { 2498 VM_WAIT; 2499 vm_pageout_deficit += desiredpages - bp->b_npages; 2500 } else { 2501 vm_page_wire(m); 2502 vm_page_wakeup(m); 2503 bp->b_flags &= ~B_CACHE; 2504 bp->b_pages[bp->b_npages] = m; 2505 ++bp->b_npages; 2506 } 2507 continue; 2508 } 2509 2510 /* 2511 * We found a page. If we have to sleep on it, 2512 * retry because it might have gotten freed out 2513 * from under us. 2514 * 2515 * We can only test PG_BUSY here. Blocking on 2516 * m->busy might lead to a deadlock: 2517 * 2518 * vm_fault->getpages->cluster_read->allocbuf 2519 * 2520 */ 2521 2522 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 2523 continue; 2524 2525 /* 2526 * We have a good page. Should we wakeup the 2527 * page daemon? 2528 */ 2529 if ((curproc != pageproc) && 2530 ((m->queue - m->pc) == PQ_CACHE) && 2531 ((cnt.v_free_count + cnt.v_cache_count) < 2532 (cnt.v_free_min + cnt.v_cache_min))) { 2533 pagedaemon_wakeup(); 2534 } 2535 vm_page_flag_clear(m, PG_ZERO); 2536 vm_page_wire(m); 2537 bp->b_pages[bp->b_npages] = m; 2538 ++bp->b_npages; 2539 } 2540 2541 /* 2542 * Step 2. We've loaded the pages into the buffer, 2543 * we have to figure out if we can still have B_CACHE 2544 * set. Note that B_CACHE is set according to the 2545 * byte-granular range ( bcount and size ), new the 2546 * aligned range ( newbsize ). 2547 * 2548 * The VM test is against m->valid, which is DEV_BSIZE 2549 * aligned. Needless to say, the validity of the data 2550 * needs to also be DEV_BSIZE aligned. Note that this 2551 * fails with NFS if the server or some other client 2552 * extends the file's EOF. If our buffer is resized, 2553 * B_CACHE may remain set! XXX 2554 */ 2555 2556 toff = bp->b_bcount; 2557 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2558 2559 while ((bp->b_flags & B_CACHE) && toff < size) { 2560 vm_pindex_t pi; 2561 2562 if (tinc > (size - toff)) 2563 tinc = size - toff; 2564 2565 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2566 PAGE_SHIFT; 2567 2568 vfs_buf_test_cache( 2569 bp, 2570 bp->b_offset, 2571 toff, 2572 tinc, 2573 bp->b_pages[pi] 2574 ); 2575 toff += tinc; 2576 tinc = PAGE_SIZE; 2577 } 2578 2579 /* 2580 * Step 3, fixup the KVM pmap. Remember that 2581 * bp->b_data is relative to bp->b_offset, but 2582 * bp->b_offset may be offset into the first page. 2583 */ 2584 2585 bp->b_data = (caddr_t) 2586 trunc_page((vm_offset_t)bp->b_data); 2587 pmap_qenter( 2588 (vm_offset_t)bp->b_data, 2589 bp->b_pages, 2590 bp->b_npages 2591 ); 2592 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2593 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2594 } 2595 } 2596 if (newbsize < bp->b_bufsize) 2597 bufspacewakeup(); 2598 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2599 bp->b_bcount = size; /* requested buffer size */ 2600 return 1; 2601} 2602 2603/* 2604 * bufwait: 2605 * 2606 * Wait for buffer I/O completion, returning error status. The buffer 2607 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR 2608 * error and cleared. 2609 */ 2610int 2611bufwait(register struct buf * bp) 2612{ 2613 int s; 2614 2615 s = splbio(); 2616 while ((bp->b_flags & B_DONE) == 0) { 2617 if (bp->b_iocmd == BIO_READ) 2618 tsleep(bp, PRIBIO, "biord", 0); 2619 else 2620 tsleep(bp, PRIBIO, "biowr", 0); 2621 } 2622 splx(s); 2623 if (bp->b_flags & B_EINTR) { 2624 bp->b_flags &= ~B_EINTR; 2625 return (EINTR); 2626 } 2627 if (bp->b_ioflags & BIO_ERROR) { 2628 return (bp->b_error ? bp->b_error : EIO); 2629 } else { 2630 return (0); 2631 } 2632} 2633 2634 /* 2635 * Call back function from struct bio back up to struct buf. 2636 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 2637 */ 2638void 2639bufdonebio(struct bio *bp) 2640{ 2641 bufdone(bp->bio_caller2); 2642} 2643 2644/* 2645 * bufdone: 2646 * 2647 * Finish I/O on a buffer, optionally calling a completion function. 2648 * This is usually called from an interrupt so process blocking is 2649 * not allowed. 2650 * 2651 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2652 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2653 * assuming B_INVAL is clear. 2654 * 2655 * For the VMIO case, we set B_CACHE if the op was a read and no 2656 * read error occured, or if the op was a write. B_CACHE is never 2657 * set if the buffer is invalid or otherwise uncacheable. 2658 * 2659 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2660 * initiator to leave B_INVAL set to brelse the buffer out of existance 2661 * in the biodone routine. 2662 */ 2663void 2664bufdone(struct buf *bp) 2665{ 2666 int s, error; 2667 void (*biodone) __P((struct buf *)); 2668 2669 s = splbio(); 2670 2671 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 2672 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 2673 2674 bp->b_flags |= B_DONE; 2675 runningbufwakeup(bp); 2676 2677 if (bp->b_iocmd == BIO_DELETE) { 2678 brelse(bp); 2679 splx(s); 2680 return; 2681 } 2682 2683 if (bp->b_iocmd == BIO_WRITE) { 2684 vwakeup(bp); 2685 } 2686 2687 /* call optional completion function if requested */ 2688 if (bp->b_iodone != NULL) { 2689 biodone = bp->b_iodone; 2690 bp->b_iodone = NULL; 2691 (*biodone) (bp); 2692 splx(s); 2693 return; 2694 } 2695 if (LIST_FIRST(&bp->b_dep) != NULL) 2696 buf_complete(bp); 2697 2698 if (bp->b_flags & B_VMIO) { 2699 int i; 2700 vm_ooffset_t foff; 2701 vm_page_t m; 2702 vm_object_t obj; 2703 int iosize; 2704 struct vnode *vp = bp->b_vp; 2705 2706 error = VOP_GETVOBJECT(vp, &obj); 2707 2708#if defined(VFS_BIO_DEBUG) 2709 if (vp->v_usecount == 0) { 2710 panic("biodone: zero vnode ref count"); 2711 } 2712 2713 if (error) { 2714 panic("biodone: missing VM object"); 2715 } 2716 2717 if ((vp->v_flag & VOBJBUF) == 0) { 2718 panic("biodone: vnode is not setup for merged cache"); 2719 } 2720#endif 2721 2722 foff = bp->b_offset; 2723 KASSERT(bp->b_offset != NOOFFSET, 2724 ("biodone: no buffer offset")); 2725 2726 if (error) { 2727 panic("biodone: no object"); 2728 } 2729#if defined(VFS_BIO_DEBUG) 2730 if (obj->paging_in_progress < bp->b_npages) { 2731 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 2732 obj->paging_in_progress, bp->b_npages); 2733 } 2734#endif 2735 2736 /* 2737 * Set B_CACHE if the op was a normal read and no error 2738 * occured. B_CACHE is set for writes in the b*write() 2739 * routines. 2740 */ 2741 iosize = bp->b_bcount - bp->b_resid; 2742 if (bp->b_iocmd == BIO_READ && 2743 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 2744 !(bp->b_ioflags & BIO_ERROR)) { 2745 bp->b_flags |= B_CACHE; 2746 } 2747 2748 for (i = 0; i < bp->b_npages; i++) { 2749 int bogusflag = 0; 2750 int resid; 2751 2752 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2753 if (resid > iosize) 2754 resid = iosize; 2755 2756 /* 2757 * cleanup bogus pages, restoring the originals 2758 */ 2759 m = bp->b_pages[i]; 2760 if (m == bogus_page) { 2761 bogusflag = 1; 2762 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2763 if (m == NULL) 2764 panic("biodone: page disappeared!"); 2765 bp->b_pages[i] = m; 2766 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2767 } 2768#if defined(VFS_BIO_DEBUG) 2769 if (OFF_TO_IDX(foff) != m->pindex) { 2770 printf( 2771"biodone: foff(%lu)/m->pindex(%d) mismatch\n", 2772 (unsigned long)foff, m->pindex); 2773 } 2774#endif 2775 2776 /* 2777 * In the write case, the valid and clean bits are 2778 * already changed correctly ( see bdwrite() ), so we 2779 * only need to do this here in the read case. 2780 */ 2781 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 2782 vfs_page_set_valid(bp, foff, i, m); 2783 } 2784 vm_page_flag_clear(m, PG_ZERO); 2785 2786 /* 2787 * when debugging new filesystems or buffer I/O methods, this 2788 * is the most common error that pops up. if you see this, you 2789 * have not set the page busy flag correctly!!! 2790 */ 2791 if (m->busy == 0) { 2792 printf("biodone: page busy < 0, " 2793 "pindex: %d, foff: 0x(%x,%x), " 2794 "resid: %d, index: %d\n", 2795 (int) m->pindex, (int)(foff >> 32), 2796 (int) foff & 0xffffffff, resid, i); 2797 if (!vn_isdisk(vp, NULL)) 2798 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 2799 bp->b_vp->v_mount->mnt_stat.f_iosize, 2800 (int) bp->b_lblkno, 2801 bp->b_flags, bp->b_npages); 2802 else 2803 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 2804 (int) bp->b_lblkno, 2805 bp->b_flags, bp->b_npages); 2806 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2807 m->valid, m->dirty, m->wire_count); 2808 panic("biodone: page busy < 0\n"); 2809 } 2810 vm_page_io_finish(m); 2811 vm_object_pip_subtract(obj, 1); 2812 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2813 iosize -= resid; 2814 } 2815 if (obj) 2816 vm_object_pip_wakeupn(obj, 0); 2817 } 2818 2819 /* 2820 * For asynchronous completions, release the buffer now. The brelse 2821 * will do a wakeup there if necessary - so no need to do a wakeup 2822 * here in the async case. The sync case always needs to do a wakeup. 2823 */ 2824 2825 if (bp->b_flags & B_ASYNC) { 2826 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 2827 brelse(bp); 2828 else 2829 bqrelse(bp); 2830 } else { 2831 wakeup(bp); 2832 } 2833 splx(s); 2834} 2835 2836/* 2837 * This routine is called in lieu of iodone in the case of 2838 * incomplete I/O. This keeps the busy status for pages 2839 * consistant. 2840 */ 2841void 2842vfs_unbusy_pages(struct buf * bp) 2843{ 2844 int i; 2845 2846 runningbufwakeup(bp); 2847 if (bp->b_flags & B_VMIO) { 2848 struct vnode *vp = bp->b_vp; 2849 vm_object_t obj; 2850 2851 VOP_GETVOBJECT(vp, &obj); 2852 2853 for (i = 0; i < bp->b_npages; i++) { 2854 vm_page_t m = bp->b_pages[i]; 2855 2856 if (m == bogus_page) { 2857 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 2858 if (!m) { 2859 panic("vfs_unbusy_pages: page missing\n"); 2860 } 2861 bp->b_pages[i] = m; 2862 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2863 } 2864 vm_object_pip_subtract(obj, 1); 2865 vm_page_flag_clear(m, PG_ZERO); 2866 vm_page_io_finish(m); 2867 } 2868 vm_object_pip_wakeupn(obj, 0); 2869 } 2870} 2871 2872/* 2873 * vfs_page_set_valid: 2874 * 2875 * Set the valid bits in a page based on the supplied offset. The 2876 * range is restricted to the buffer's size. 2877 * 2878 * This routine is typically called after a read completes. 2879 */ 2880static void 2881vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 2882{ 2883 vm_ooffset_t soff, eoff; 2884 2885 /* 2886 * Start and end offsets in buffer. eoff - soff may not cross a 2887 * page boundry or cross the end of the buffer. The end of the 2888 * buffer, in this case, is our file EOF, not the allocation size 2889 * of the buffer. 2890 */ 2891 soff = off; 2892 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2893 if (eoff > bp->b_offset + bp->b_bcount) 2894 eoff = bp->b_offset + bp->b_bcount; 2895 2896 /* 2897 * Set valid range. This is typically the entire buffer and thus the 2898 * entire page. 2899 */ 2900 if (eoff > soff) { 2901 vm_page_set_validclean( 2902 m, 2903 (vm_offset_t) (soff & PAGE_MASK), 2904 (vm_offset_t) (eoff - soff) 2905 ); 2906 } 2907} 2908 2909/* 2910 * This routine is called before a device strategy routine. 2911 * It is used to tell the VM system that paging I/O is in 2912 * progress, and treat the pages associated with the buffer 2913 * almost as being PG_BUSY. Also the object paging_in_progress 2914 * flag is handled to make sure that the object doesn't become 2915 * inconsistant. 2916 * 2917 * Since I/O has not been initiated yet, certain buffer flags 2918 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 2919 * and should be ignored. 2920 */ 2921void 2922vfs_busy_pages(struct buf * bp, int clear_modify) 2923{ 2924 int i, bogus; 2925 2926 if (bp->b_flags & B_VMIO) { 2927 struct vnode *vp = bp->b_vp; 2928 vm_object_t obj; 2929 vm_ooffset_t foff; 2930 2931 VOP_GETVOBJECT(vp, &obj); 2932 foff = bp->b_offset; 2933 KASSERT(bp->b_offset != NOOFFSET, 2934 ("vfs_busy_pages: no buffer offset")); 2935 vfs_setdirty(bp); 2936 2937retry: 2938 for (i = 0; i < bp->b_npages; i++) { 2939 vm_page_t m = bp->b_pages[i]; 2940 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 2941 goto retry; 2942 } 2943 2944 bogus = 0; 2945 for (i = 0; i < bp->b_npages; i++) { 2946 vm_page_t m = bp->b_pages[i]; 2947 2948 vm_page_flag_clear(m, PG_ZERO); 2949 if ((bp->b_flags & B_CLUSTER) == 0) { 2950 vm_object_pip_add(obj, 1); 2951 vm_page_io_start(m); 2952 } 2953 2954 /* 2955 * When readying a buffer for a read ( i.e 2956 * clear_modify == 0 ), it is important to do 2957 * bogus_page replacement for valid pages in 2958 * partially instantiated buffers. Partially 2959 * instantiated buffers can, in turn, occur when 2960 * reconstituting a buffer from its VM backing store 2961 * base. We only have to do this if B_CACHE is 2962 * clear ( which causes the I/O to occur in the 2963 * first place ). The replacement prevents the read 2964 * I/O from overwriting potentially dirty VM-backed 2965 * pages. XXX bogus page replacement is, uh, bogus. 2966 * It may not work properly with small-block devices. 2967 * We need to find a better way. 2968 */ 2969 2970 vm_page_protect(m, VM_PROT_NONE); 2971 if (clear_modify) 2972 vfs_page_set_valid(bp, foff, i, m); 2973 else if (m->valid == VM_PAGE_BITS_ALL && 2974 (bp->b_flags & B_CACHE) == 0) { 2975 bp->b_pages[i] = bogus_page; 2976 bogus++; 2977 } 2978 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2979 } 2980 if (bogus) 2981 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2982 } 2983} 2984 2985/* 2986 * Tell the VM system that the pages associated with this buffer 2987 * are clean. This is used for delayed writes where the data is 2988 * going to go to disk eventually without additional VM intevention. 2989 * 2990 * Note that while we only really need to clean through to b_bcount, we 2991 * just go ahead and clean through to b_bufsize. 2992 */ 2993static void 2994vfs_clean_pages(struct buf * bp) 2995{ 2996 int i; 2997 2998 if (bp->b_flags & B_VMIO) { 2999 vm_ooffset_t foff; 3000 3001 foff = bp->b_offset; 3002 KASSERT(bp->b_offset != NOOFFSET, 3003 ("vfs_clean_pages: no buffer offset")); 3004 for (i = 0; i < bp->b_npages; i++) { 3005 vm_page_t m = bp->b_pages[i]; 3006 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3007 vm_ooffset_t eoff = noff; 3008 3009 if (eoff > bp->b_offset + bp->b_bufsize) 3010 eoff = bp->b_offset + bp->b_bufsize; 3011 vfs_page_set_valid(bp, foff, i, m); 3012 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3013 foff = noff; 3014 } 3015 } 3016} 3017 3018/* 3019 * vfs_bio_set_validclean: 3020 * 3021 * Set the range within the buffer to valid and clean. The range is 3022 * relative to the beginning of the buffer, b_offset. Note that b_offset 3023 * itself may be offset from the beginning of the first page. 3024 */ 3025 3026void 3027vfs_bio_set_validclean(struct buf *bp, int base, int size) 3028{ 3029 if (bp->b_flags & B_VMIO) { 3030 int i; 3031 int n; 3032 3033 /* 3034 * Fixup base to be relative to beginning of first page. 3035 * Set initial n to be the maximum number of bytes in the 3036 * first page that can be validated. 3037 */ 3038 3039 base += (bp->b_offset & PAGE_MASK); 3040 n = PAGE_SIZE - (base & PAGE_MASK); 3041 3042 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3043 vm_page_t m = bp->b_pages[i]; 3044 3045 if (n > size) 3046 n = size; 3047 3048 vm_page_set_validclean(m, base & PAGE_MASK, n); 3049 base += n; 3050 size -= n; 3051 n = PAGE_SIZE; 3052 } 3053 } 3054} 3055 3056/* 3057 * vfs_bio_clrbuf: 3058 * 3059 * clear a buffer. This routine essentially fakes an I/O, so we need 3060 * to clear BIO_ERROR and B_INVAL. 3061 * 3062 * Note that while we only theoretically need to clear through b_bcount, 3063 * we go ahead and clear through b_bufsize. 3064 */ 3065 3066void 3067vfs_bio_clrbuf(struct buf *bp) { 3068 int i, mask = 0; 3069 caddr_t sa, ea; 3070 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3071 bp->b_flags &= ~B_INVAL; 3072 bp->b_ioflags &= ~BIO_ERROR; 3073 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3074 (bp->b_offset & PAGE_MASK) == 0) { 3075 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3076 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3077 ((bp->b_pages[0]->valid & mask) != mask)) { 3078 bzero(bp->b_data, bp->b_bufsize); 3079 } 3080 bp->b_pages[0]->valid |= mask; 3081 bp->b_resid = 0; 3082 return; 3083 } 3084 ea = sa = bp->b_data; 3085 for(i=0;i<bp->b_npages;i++,sa=ea) { 3086 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3087 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3088 ea = (caddr_t)(vm_offset_t)ulmin( 3089 (u_long)(vm_offset_t)ea, 3090 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3091 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3092 if ((bp->b_pages[i]->valid & mask) == mask) 3093 continue; 3094 if ((bp->b_pages[i]->valid & mask) == 0) { 3095 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3096 bzero(sa, ea - sa); 3097 } 3098 } else { 3099 for (; sa < ea; sa += DEV_BSIZE, j++) { 3100 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3101 (bp->b_pages[i]->valid & (1<<j)) == 0) 3102 bzero(sa, DEV_BSIZE); 3103 } 3104 } 3105 bp->b_pages[i]->valid |= mask; 3106 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3107 } 3108 bp->b_resid = 0; 3109 } else { 3110 clrbuf(bp); 3111 } 3112} 3113 3114/* 3115 * vm_hold_load_pages and vm_hold_unload pages get pages into 3116 * a buffers address space. The pages are anonymous and are 3117 * not associated with a file object. 3118 */ 3119void 3120vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3121{ 3122 vm_offset_t pg; 3123 vm_page_t p; 3124 int index; 3125 3126 to = round_page(to); 3127 from = round_page(from); 3128 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3129 3130 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3131 3132tryagain: 3133 3134 /* 3135 * note: must allocate system pages since blocking here 3136 * could intefere with paging I/O, no matter which 3137 * process we are. 3138 */ 3139 p = vm_page_alloc(kernel_object, 3140 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3141 VM_ALLOC_SYSTEM); 3142 if (!p) { 3143 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3144 VM_WAIT; 3145 goto tryagain; 3146 } 3147 vm_page_wire(p); 3148 p->valid = VM_PAGE_BITS_ALL; 3149 vm_page_flag_clear(p, PG_ZERO); 3150 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 3151 bp->b_pages[index] = p; 3152 vm_page_wakeup(p); 3153 } 3154 bp->b_npages = index; 3155} 3156 3157void 3158vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3159{ 3160 vm_offset_t pg; 3161 vm_page_t p; 3162 int index, newnpages; 3163 3164 from = round_page(from); 3165 to = round_page(to); 3166 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3167 3168 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3169 p = bp->b_pages[index]; 3170 if (p && (index < bp->b_npages)) { 3171 if (p->busy) { 3172 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 3173 bp->b_blkno, bp->b_lblkno); 3174 } 3175 bp->b_pages[index] = NULL; 3176 pmap_kremove(pg); 3177 vm_page_busy(p); 3178 vm_page_unwire(p, 0); 3179 vm_page_free(p); 3180 } 3181 } 3182 bp->b_npages = newnpages; 3183} 3184 3185 3186#include "opt_ddb.h" 3187#ifdef DDB 3188#include <ddb/ddb.h> 3189 3190DB_SHOW_COMMAND(buffer, db_show_buffer) 3191{ 3192 /* get args */ 3193 struct buf *bp = (struct buf *)addr; 3194 3195 if (!have_addr) { 3196 db_printf("usage: show buffer <addr>\n"); 3197 return; 3198 } 3199 3200 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3201 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 3202 "b_resid = %ld\nb_dev = (%d,%d), b_data = %p, " 3203 "b_blkno = %d, b_pblkno = %d\n", 3204 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3205 major(bp->b_dev), minor(bp->b_dev), 3206 bp->b_data, bp->b_blkno, bp->b_pblkno); 3207 if (bp->b_npages) { 3208 int i; 3209 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3210 for (i = 0; i < bp->b_npages; i++) { 3211 vm_page_t m; 3212 m = bp->b_pages[i]; 3213 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3214 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3215 if ((i + 1) < bp->b_npages) 3216 db_printf(","); 3217 } 3218 db_printf("\n"); 3219 } 3220} 3221#endif /* DDB */ 3222