vfs_bio.c revision 58132
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 58132 2000-03-16 08:51:55Z phk $ 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/buf.h> 33#include <sys/conf.h> 34#include <sys/eventhandler.h> 35#include <sys/lock.h> 36#include <sys/malloc.h> 37#include <sys/mount.h> 38#include <sys/kernel.h> 39#include <sys/kthread.h> 40#include <sys/proc.h> 41#include <sys/reboot.h> 42#include <sys/resourcevar.h> 43#include <sys/sysctl.h> 44#include <sys/vmmeter.h> 45#include <sys/vnode.h> 46#include <vm/vm.h> 47#include <vm/vm_param.h> 48#include <vm/vm_kern.h> 49#include <vm/vm_pageout.h> 50#include <vm/vm_page.h> 51#include <vm/vm_object.h> 52#include <vm/vm_extern.h> 53#include <vm/vm_map.h> 54 55static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 56 57struct bio_ops bioops; /* I/O operation notification */ 58 59struct buf *buf; /* buffer header pool */ 60struct swqueue bswlist; 61 62static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 63 vm_offset_t to); 64static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 65 vm_offset_t to); 66static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 67 int pageno, vm_page_t m); 68static void vfs_clean_pages(struct buf * bp); 69static void vfs_setdirty(struct buf *bp); 70static void vfs_vmio_release(struct buf *bp); 71static void vfs_backgroundwritedone(struct buf *bp); 72static int flushbufqueues(void); 73 74static int bd_request; 75 76static void buf_daemon __P((void)); 77/* 78 * bogus page -- for I/O to/from partially complete buffers 79 * this is a temporary solution to the problem, but it is not 80 * really that bad. it would be better to split the buffer 81 * for input in the case of buffers partially already in memory, 82 * but the code is intricate enough already. 83 */ 84vm_page_t bogus_page; 85int runningbufspace; 86int vmiodirenable = FALSE; 87int buf_maxio = DFLTPHYS; 88static vm_offset_t bogus_offset; 89 90static int bufspace, maxbufspace, vmiospace, 91 bufmallocspace, maxbufmallocspace, hibufspace; 92static int maxbdrun; 93static int needsbuffer; 94static int numdirtybuffers, hidirtybuffers; 95static int numfreebuffers, lofreebuffers, hifreebuffers; 96static int getnewbufcalls; 97static int getnewbufrestarts; 98static int kvafreespace; 99 100SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, 101 &numdirtybuffers, 0, ""); 102SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, 103 &hidirtybuffers, 0, ""); 104SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, 105 &numfreebuffers, 0, ""); 106SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, 107 &lofreebuffers, 0, ""); 108SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, 109 &hifreebuffers, 0, ""); 110SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, 111 &runningbufspace, 0, ""); 112SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RW, 113 &maxbufspace, 0, ""); 114SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, 115 &hibufspace, 0, ""); 116SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, 117 &bufspace, 0, ""); 118SYSCTL_INT(_vfs, OID_AUTO, maxbdrun, CTLFLAG_RW, 119 &maxbdrun, 0, ""); 120SYSCTL_INT(_vfs, OID_AUTO, vmiospace, CTLFLAG_RD, 121 &vmiospace, 0, ""); 122SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, 123 &maxbufmallocspace, 0, ""); 124SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, 125 &bufmallocspace, 0, ""); 126SYSCTL_INT(_vfs, OID_AUTO, kvafreespace, CTLFLAG_RD, 127 &kvafreespace, 0, ""); 128SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, 129 &getnewbufcalls, 0, ""); 130SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, 131 &getnewbufrestarts, 0, ""); 132SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, 133 &vmiodirenable, 0, ""); 134 135 136static int bufhashmask; 137static LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash; 138struct bqueues bufqueues[BUFFER_QUEUES] = { { 0 } }; 139char *buf_wmesg = BUF_WMESG; 140 141extern int vm_swap_size; 142 143#define BUF_MAXUSE 24 144 145#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 146#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 147#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 148#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 149#define VFS_BIO_NEED_KVASPACE 0x10 /* wait for buffer_map space, emerg */ 150 151/* 152 * Buffer hash table code. Note that the logical block scans linearly, which 153 * gives us some L1 cache locality. 154 */ 155 156static __inline 157struct bufhashhdr * 158bufhash(struct vnode *vnp, daddr_t bn) 159{ 160 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]); 161} 162 163/* 164 * kvaspacewakeup: 165 * 166 * Called when kva space is potential available for recovery or when 167 * kva space is recovered in the buffer_map. This function wakes up 168 * anyone waiting for buffer_map kva space. Even though the buffer_map 169 * is larger then maxbufspace, this situation will typically occur 170 * when the buffer_map gets fragmented. 171 */ 172 173static __inline void 174kvaspacewakeup(void) 175{ 176 /* 177 * If someone is waiting for KVA space, wake them up. Even 178 * though we haven't freed the kva space yet, the waiting 179 * process will be able to now. 180 */ 181 if (needsbuffer & VFS_BIO_NEED_KVASPACE) { 182 needsbuffer &= ~VFS_BIO_NEED_KVASPACE; 183 wakeup(&needsbuffer); 184 } 185} 186 187/* 188 * numdirtywakeup: 189 * 190 * If someone is blocked due to there being too many dirty buffers, 191 * and numdirtybuffers is now reasonable, wake them up. 192 */ 193 194static __inline void 195numdirtywakeup(void) 196{ 197 if (numdirtybuffers < hidirtybuffers) { 198 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 199 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 200 wakeup(&needsbuffer); 201 } 202 } 203} 204 205/* 206 * bufspacewakeup: 207 * 208 * Called when buffer space is potentially available for recovery or when 209 * buffer space is recovered. getnewbuf() will block on this flag when 210 * it is unable to free sufficient buffer space. Buffer space becomes 211 * recoverable when bp's get placed back in the queues. 212 */ 213 214static __inline void 215bufspacewakeup(void) 216{ 217 /* 218 * If someone is waiting for BUF space, wake them up. Even 219 * though we haven't freed the kva space yet, the waiting 220 * process will be able to now. 221 */ 222 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 223 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 224 wakeup(&needsbuffer); 225 } 226} 227 228/* 229 * bufcountwakeup: 230 * 231 * Called when a buffer has been added to one of the free queues to 232 * account for the buffer and to wakeup anyone waiting for free buffers. 233 * This typically occurs when large amounts of metadata are being handled 234 * by the buffer cache ( else buffer space runs out first, usually ). 235 */ 236 237static __inline void 238bufcountwakeup(void) 239{ 240 ++numfreebuffers; 241 if (needsbuffer) { 242 needsbuffer &= ~VFS_BIO_NEED_ANY; 243 if (numfreebuffers >= hifreebuffers) 244 needsbuffer &= ~VFS_BIO_NEED_FREE; 245 wakeup(&needsbuffer); 246 } 247} 248 249/* 250 * vfs_buf_test_cache: 251 * 252 * Called when a buffer is extended. This function clears the B_CACHE 253 * bit if the newly extended portion of the buffer does not contain 254 * valid data. 255 */ 256static __inline__ 257void 258vfs_buf_test_cache(struct buf *bp, 259 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 260 vm_page_t m) 261{ 262 if (bp->b_flags & B_CACHE) { 263 int base = (foff + off) & PAGE_MASK; 264 if (vm_page_is_valid(m, base, size) == 0) 265 bp->b_flags &= ~B_CACHE; 266 } 267} 268 269static __inline__ 270void 271bd_wakeup(int dirtybuflevel) 272{ 273 if (numdirtybuffers >= dirtybuflevel && bd_request == 0) { 274 bd_request = 1; 275 wakeup(&bd_request); 276 } 277} 278 279/* 280 * bd_speedup - speedup the buffer cache flushing code 281 */ 282 283static __inline__ 284void 285bd_speedup(void) 286{ 287 bd_wakeup(1); 288} 289 290/* 291 * Initialize buffer headers and related structures. 292 */ 293 294caddr_t 295bufhashinit(caddr_t vaddr) 296{ 297 /* first, make a null hash table */ 298 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 299 ; 300 bufhashtbl = (void *)vaddr; 301 vaddr = vaddr + sizeof(*bufhashtbl) * bufhashmask; 302 --bufhashmask; 303 return(vaddr); 304} 305 306void 307bufinit(void) 308{ 309 struct buf *bp; 310 int i; 311 312 TAILQ_INIT(&bswlist); 313 LIST_INIT(&invalhash); 314 simple_lock_init(&buftimelock); 315 316 for (i = 0; i <= bufhashmask; i++) 317 LIST_INIT(&bufhashtbl[i]); 318 319 /* next, make a null set of free lists */ 320 for (i = 0; i < BUFFER_QUEUES; i++) 321 TAILQ_INIT(&bufqueues[i]); 322 323 /* finally, initialize each buffer header and stick on empty q */ 324 for (i = 0; i < nbuf; i++) { 325 bp = &buf[i]; 326 bzero(bp, sizeof *bp); 327 bp->b_flags = B_INVAL; /* we're just an empty header */ 328 bp->b_dev = NODEV; 329 bp->b_rcred = NOCRED; 330 bp->b_wcred = NOCRED; 331 bp->b_qindex = QUEUE_EMPTY; 332 bp->b_xflags = 0; 333 LIST_INIT(&bp->b_dep); 334 BUF_LOCKINIT(bp); 335 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 336 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 337 } 338 339 /* 340 * maxbufspace is currently calculated to be maximally efficient 341 * when the filesystem block size is DFLTBSIZE or DFLTBSIZE*2 342 * (4K or 8K). To reduce the number of stall points our calculation 343 * is based on DFLTBSIZE which should reduce the chances of actually 344 * running out of buffer headers. The maxbufspace calculation is also 345 * based on DFLTBSIZE (4K) instead of BKVASIZE (8K) in order to 346 * reduce the chance that a KVA allocation will fail due to 347 * fragmentation. While this does not usually create a stall, 348 * the KVA map allocation/free functions are O(N) rather then O(1) 349 * so running them constantly would result in inefficient O(N*M) 350 * buffer cache operation. 351 */ 352 maxbufspace = (nbuf + 8) * DFLTBSIZE; 353 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 354/* 355 * Limit the amount of malloc memory since it is wired permanently into 356 * the kernel space. Even though this is accounted for in the buffer 357 * allocation, we don't want the malloced region to grow uncontrolled. 358 * The malloc scheme improves memory utilization significantly on average 359 * (small) directories. 360 */ 361 maxbufmallocspace = hibufspace / 20; 362 363/* 364 * Reduce the chance of a deadlock occuring by limiting the number 365 * of delayed-write dirty buffers we allow to stack up. 366 */ 367 hidirtybuffers = nbuf / 4 + 20; 368 numdirtybuffers = 0; 369/* 370 * To support extreme low-memory systems, make sure hidirtybuffers cannot 371 * eat up all available buffer space. This occurs when our minimum cannot 372 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 373 * BKVASIZE'd (8K) buffers. We also reduce buf_maxio in this case (used 374 * by the clustering code) in an attempt to further reduce the load on 375 * the buffer cache. 376 */ 377 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 378 hidirtybuffers >>= 1; 379 buf_maxio >>= 1; 380 } 381 382 /* 383 * Temporary, BKVASIZE may be manipulated soon, make sure we don't 384 * do something illegal. XXX 385 */ 386#if BKVASIZE < MAXBSIZE 387 if (buf_maxio < BKVASIZE * 2) 388 buf_maxio = BKVASIZE * 2; 389#else 390 if (buf_maxio < MAXBSIZE) 391 buf_maxio = MAXBSIZE; 392#endif 393 394/* 395 * Try to keep the number of free buffers in the specified range, 396 * and give the syncer access to an emergency reserve. 397 */ 398 lofreebuffers = nbuf / 18 + 5; 399 hifreebuffers = 2 * lofreebuffers; 400 numfreebuffers = nbuf; 401 402/* 403 * Maximum number of async ops initiated per buf_daemon loop. This is 404 * somewhat of a hack at the moment, we really need to limit ourselves 405 * based on the number of bytes of I/O in-transit that were initiated 406 * from buf_daemon. 407 */ 408 if ((maxbdrun = nswbuf / 4) < 4) 409 maxbdrun = 4; 410 411 kvafreespace = 0; 412 413 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 414 bogus_page = vm_page_alloc(kernel_object, 415 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 416 VM_ALLOC_NORMAL); 417 cnt.v_wire_count++; 418 419} 420 421/* 422 * Free the kva allocation for a buffer 423 * Must be called only at splbio or higher, 424 * as this is the only locking for buffer_map. 425 */ 426static void 427bfreekva(struct buf * bp) 428{ 429 if (bp->b_kvasize) { 430 vm_map_delete(buffer_map, 431 (vm_offset_t) bp->b_kvabase, 432 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 433 ); 434 bp->b_kvasize = 0; 435 kvaspacewakeup(); 436 } 437} 438 439/* 440 * bremfree: 441 * 442 * Remove the buffer from the appropriate free list. 443 */ 444void 445bremfree(struct buf * bp) 446{ 447 int s = splbio(); 448 int old_qindex = bp->b_qindex; 449 450 if (bp->b_qindex != QUEUE_NONE) { 451 if (bp->b_qindex == QUEUE_EMPTYKVA) { 452 kvafreespace -= bp->b_kvasize; 453 } 454 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 455 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 456 bp->b_qindex = QUEUE_NONE; 457 runningbufspace += bp->b_bufsize; 458 } else { 459 if (BUF_REFCNT(bp) <= 1) 460 panic("bremfree: removing a buffer not on a queue"); 461 } 462 463 /* 464 * Fixup numfreebuffers count. If the buffer is invalid or not 465 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 466 * the buffer was free and we must decrement numfreebuffers. 467 */ 468 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 469 switch(old_qindex) { 470 case QUEUE_DIRTY: 471 case QUEUE_CLEAN: 472 case QUEUE_EMPTY: 473 case QUEUE_EMPTYKVA: 474 --numfreebuffers; 475 break; 476 default: 477 break; 478 } 479 } 480 splx(s); 481} 482 483 484/* 485 * Get a buffer with the specified data. Look in the cache first. We 486 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 487 * is set, the buffer is valid and we do not have to do anything ( see 488 * getblk() ). 489 */ 490int 491bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 492 struct buf ** bpp) 493{ 494 struct buf *bp; 495 496 bp = getblk(vp, blkno, size, 0, 0); 497 *bpp = bp; 498 499 /* if not found in cache, do some I/O */ 500 if ((bp->b_flags & B_CACHE) == 0) { 501 if (curproc != NULL) 502 curproc->p_stats->p_ru.ru_inblock++; 503 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp)); 504 bp->b_flags |= B_READ; 505 bp->b_flags &= ~(B_ERROR | B_INVAL); 506 if (bp->b_rcred == NOCRED) { 507 if (cred != NOCRED) 508 crhold(cred); 509 bp->b_rcred = cred; 510 } 511 vfs_busy_pages(bp, 0); 512 VOP_STRATEGY(vp, bp); 513 return (biowait(bp)); 514 } 515 return (0); 516} 517 518/* 519 * Operates like bread, but also starts asynchronous I/O on 520 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior 521 * to initiating I/O . If B_CACHE is set, the buffer is valid 522 * and we do not have to do anything. 523 */ 524int 525breadn(struct vnode * vp, daddr_t blkno, int size, 526 daddr_t * rablkno, int *rabsize, 527 int cnt, struct ucred * cred, struct buf ** bpp) 528{ 529 struct buf *bp, *rabp; 530 int i; 531 int rv = 0, readwait = 0; 532 533 *bpp = bp = getblk(vp, blkno, size, 0, 0); 534 535 /* if not found in cache, do some I/O */ 536 if ((bp->b_flags & B_CACHE) == 0) { 537 if (curproc != NULL) 538 curproc->p_stats->p_ru.ru_inblock++; 539 bp->b_flags |= B_READ; 540 bp->b_flags &= ~(B_ERROR | B_INVAL); 541 if (bp->b_rcred == NOCRED) { 542 if (cred != NOCRED) 543 crhold(cred); 544 bp->b_rcred = cred; 545 } 546 vfs_busy_pages(bp, 0); 547 VOP_STRATEGY(vp, bp); 548 ++readwait; 549 } 550 551 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 552 if (inmem(vp, *rablkno)) 553 continue; 554 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 555 556 if ((rabp->b_flags & B_CACHE) == 0) { 557 if (curproc != NULL) 558 curproc->p_stats->p_ru.ru_inblock++; 559 rabp->b_flags |= B_READ | B_ASYNC; 560 rabp->b_flags &= ~(B_ERROR | B_INVAL); 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 = biowait(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 */ 591int 592bwrite(struct buf * bp) 593{ 594 int oldflags, s; 595 struct buf *newbp; 596 597 if (bp->b_flags & B_INVAL) { 598 brelse(bp); 599 return (0); 600 } 601 602 oldflags = bp->b_flags; 603 604 if (BUF_REFCNT(bp) == 0) 605 panic("bwrite: buffer is not busy???"); 606 s = splbio(); 607 /* 608 * If a background write is already in progress, delay 609 * writing this block if it is asynchronous. Otherwise 610 * wait for the background write to complete. 611 */ 612 if (bp->b_xflags & BX_BKGRDINPROG) { 613 if (bp->b_flags & B_ASYNC) { 614 splx(s); 615 bdwrite(bp); 616 return (0); 617 } 618 bp->b_xflags |= BX_BKGRDWAIT; 619 tsleep(&bp->b_xflags, PRIBIO, "biord", 0); 620 if (bp->b_xflags & BX_BKGRDINPROG) 621 panic("bwrite: still writing"); 622 } 623 624 /* Mark the buffer clean */ 625 bundirty(bp); 626 627 /* 628 * If this buffer is marked for background writing and we 629 * do not have to wait for it, make a copy and write the 630 * copy so as to leave this buffer ready for further use. 631 */ 632 if ((bp->b_xflags & BX_BKGRDWRITE) && (bp->b_flags & B_ASYNC)) { 633 if (bp->b_flags & B_CALL) 634 panic("bwrite: need chained iodone"); 635 636 /* get a new block */ 637 newbp = geteblk(bp->b_bufsize); 638 639 /* set it to be identical to the old block */ 640 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 641 bgetvp(bp->b_vp, newbp); 642 newbp->b_lblkno = bp->b_lblkno; 643 newbp->b_blkno = bp->b_blkno; 644 newbp->b_offset = bp->b_offset; 645 newbp->b_iodone = vfs_backgroundwritedone; 646 newbp->b_flags |= B_ASYNC | B_CALL; 647 newbp->b_flags &= ~B_INVAL; 648 649 /* move over the dependencies */ 650 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps) 651 (*bioops.io_movedeps)(bp, newbp); 652 653 /* 654 * Initiate write on the copy, release the original to 655 * the B_LOCKED queue so that it cannot go away until 656 * the background write completes. If not locked it could go 657 * away and then be reconstituted while it was being written. 658 * If the reconstituted buffer were written, we could end up 659 * with two background copies being written at the same time. 660 */ 661 bp->b_xflags |= BX_BKGRDINPROG; 662 bp->b_flags |= B_LOCKED; 663 bqrelse(bp); 664 bp = newbp; 665 } 666 667 bp->b_flags &= ~(B_READ | B_DONE | B_ERROR); 668 bp->b_flags |= B_WRITEINPROG | B_CACHE; 669 670 bp->b_vp->v_numoutput++; 671 vfs_busy_pages(bp, 1); 672 if (curproc != NULL) 673 curproc->p_stats->p_ru.ru_oublock++; 674 splx(s); 675 if (oldflags & B_ASYNC) 676 BUF_KERNPROC(bp); 677 VOP_STRATEGY(bp->b_vp, bp); 678 679 if ((oldflags & B_ASYNC) == 0) { 680 int rtval = biowait(bp); 681 brelse(bp); 682 return (rtval); 683 } 684 685 return (0); 686} 687 688/* 689 * Complete a background write started from bwrite. 690 */ 691static void 692vfs_backgroundwritedone(bp) 693 struct buf *bp; 694{ 695 struct buf *origbp; 696 697 /* 698 * Find the original buffer that we are writing. 699 */ 700 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 701 panic("backgroundwritedone: lost buffer"); 702 /* 703 * Process dependencies then return any unfinished ones. 704 */ 705 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 706 (*bioops.io_complete)(bp); 707 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps) 708 (*bioops.io_movedeps)(bp, origbp); 709 /* 710 * Clear the BX_BKGRDINPROG flag in the original buffer 711 * and awaken it if it is waiting for the write to complete. 712 */ 713 origbp->b_xflags &= ~BX_BKGRDINPROG; 714 if (origbp->b_xflags & BX_BKGRDWAIT) { 715 origbp->b_xflags &= ~BX_BKGRDWAIT; 716 wakeup(&origbp->b_xflags); 717 } 718 /* 719 * Clear the B_LOCKED flag and remove it from the locked 720 * queue if it currently resides there. 721 */ 722 origbp->b_flags &= ~B_LOCKED; 723 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 724 bremfree(origbp); 725 bqrelse(origbp); 726 } 727 /* 728 * This buffer is marked B_NOCACHE, so when it is released 729 * by biodone, it will be tossed. We mark it with B_READ 730 * to avoid biodone doing a second vwakeup. 731 */ 732 bp->b_flags |= B_NOCACHE | B_READ; 733 bp->b_flags &= ~(B_CACHE | B_CALL | B_DONE); 734 bp->b_iodone = 0; 735 biodone(bp); 736} 737 738/* 739 * Delayed write. (Buffer is marked dirty). Do not bother writing 740 * anything if the buffer is marked invalid. 741 * 742 * Note that since the buffer must be completely valid, we can safely 743 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 744 * biodone() in order to prevent getblk from writing the buffer 745 * out synchronously. 746 */ 747void 748bdwrite(struct buf * bp) 749{ 750 if (BUF_REFCNT(bp) == 0) 751 panic("bdwrite: buffer is not busy"); 752 753 if (bp->b_flags & B_INVAL) { 754 brelse(bp); 755 return; 756 } 757 bdirty(bp); 758 759 /* 760 * Set B_CACHE, indicating that the buffer is fully valid. This is 761 * true even of NFS now. 762 */ 763 bp->b_flags |= B_CACHE; 764 765 /* 766 * This bmap keeps the system from needing to do the bmap later, 767 * perhaps when the system is attempting to do a sync. Since it 768 * is likely that the indirect block -- or whatever other datastructure 769 * that the filesystem needs is still in memory now, it is a good 770 * thing to do this. Note also, that if the pageout daemon is 771 * requesting a sync -- there might not be enough memory to do 772 * the bmap then... So, this is important to do. 773 */ 774 if (bp->b_lblkno == bp->b_blkno) { 775 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 776 } 777 778 /* 779 * Set the *dirty* buffer range based upon the VM system dirty pages. 780 */ 781 vfs_setdirty(bp); 782 783 /* 784 * We need to do this here to satisfy the vnode_pager and the 785 * pageout daemon, so that it thinks that the pages have been 786 * "cleaned". Note that since the pages are in a delayed write 787 * buffer -- the VFS layer "will" see that the pages get written 788 * out on the next sync, or perhaps the cluster will be completed. 789 */ 790 vfs_clean_pages(bp); 791 bqrelse(bp); 792 793 /* 794 * Wakeup the buffer flushing daemon if we have saturated the 795 * buffer cache. 796 */ 797 798 bd_wakeup(hidirtybuffers); 799 800 /* 801 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 802 * due to the softdep code. 803 */ 804} 805 806/* 807 * bdirty: 808 * 809 * Turn buffer into delayed write request. We must clear B_READ and 810 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 811 * itself to properly update it in the dirty/clean lists. We mark it 812 * B_DONE to ensure that any asynchronization of the buffer properly 813 * clears B_DONE ( else a panic will occur later ). 814 * 815 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 816 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 817 * should only be called if the buffer is known-good. 818 * 819 * Since the buffer is not on a queue, we do not update the numfreebuffers 820 * count. 821 * 822 * Must be called at splbio(). 823 * The buffer must be on QUEUE_NONE. 824 */ 825void 826bdirty(bp) 827 struct buf *bp; 828{ 829 KASSERT(bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 830 bp->b_flags &= ~(B_READ|B_RELBUF); 831 832 if ((bp->b_flags & B_DELWRI) == 0) { 833 bp->b_flags |= B_DONE | B_DELWRI; 834 reassignbuf(bp, bp->b_vp); 835 ++numdirtybuffers; 836 bd_wakeup(hidirtybuffers); 837 } 838} 839 840/* 841 * bundirty: 842 * 843 * Clear B_DELWRI for buffer. 844 * 845 * Since the buffer is not on a queue, we do not update the numfreebuffers 846 * count. 847 * 848 * Must be called at splbio(). 849 * The buffer must be on QUEUE_NONE. 850 */ 851 852void 853bundirty(bp) 854 struct buf *bp; 855{ 856 KASSERT(bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 857 858 if (bp->b_flags & B_DELWRI) { 859 bp->b_flags &= ~B_DELWRI; 860 reassignbuf(bp, bp->b_vp); 861 --numdirtybuffers; 862 numdirtywakeup(); 863 } 864 /* 865 * Since it is now being written, we can clear its deferred write flag. 866 */ 867 bp->b_flags &= ~B_DEFERRED; 868} 869 870/* 871 * bawrite: 872 * 873 * Asynchronous write. Start output on a buffer, but do not wait for 874 * it to complete. The buffer is released when the output completes. 875 * 876 * bwrite() ( or the VOP routine anyway ) is responsible for handling 877 * B_INVAL buffers. Not us. 878 */ 879void 880bawrite(struct buf * bp) 881{ 882 bp->b_flags |= B_ASYNC; 883 (void) VOP_BWRITE(bp->b_vp, bp); 884} 885 886/* 887 * bowrite: 888 * 889 * Ordered write. Start output on a buffer, and flag it so that the 890 * device will write it in the order it was queued. The buffer is 891 * released when the output completes. bwrite() ( or the VOP routine 892 * anyway ) is responsible for handling B_INVAL buffers. 893 */ 894int 895bowrite(struct buf * bp) 896{ 897 bp->b_flags |= B_ORDERED | B_ASYNC; 898 return (VOP_BWRITE(bp->b_vp, bp)); 899} 900 901/* 902 * bwillwrite: 903 * 904 * Called prior to the locking of any vnodes when we are expecting to 905 * write. We do not want to starve the buffer cache with too many 906 * dirty buffers so we block here. By blocking prior to the locking 907 * of any vnodes we attempt to avoid the situation where a locked vnode 908 * prevents the various system daemons from flushing related buffers. 909 */ 910 911void 912bwillwrite(void) 913{ 914 int slop = hidirtybuffers / 10; 915 916 if (numdirtybuffers > hidirtybuffers + slop) { 917 int s; 918 919 s = splbio(); 920 while (numdirtybuffers > hidirtybuffers) { 921 bd_wakeup(hidirtybuffers); 922 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 923 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0); 924 } 925 splx(s); 926 } 927} 928 929/* 930 * brelse: 931 * 932 * Release a busy buffer and, if requested, free its resources. The 933 * buffer will be stashed in the appropriate bufqueue[] allowing it 934 * to be accessed later as a cache entity or reused for other purposes. 935 */ 936void 937brelse(struct buf * bp) 938{ 939 int s; 940 int kvawakeup = 0; 941 942 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 943 944 s = splbio(); 945 946 if (bp->b_flags & B_LOCKED) 947 bp->b_flags &= ~B_ERROR; 948 949 if ((bp->b_flags & (B_READ | B_ERROR | B_INVAL)) == B_ERROR) { 950 /* 951 * Failed write, redirty. Must clear B_ERROR to prevent 952 * pages from being scrapped. If B_INVAL is set then 953 * this case is not run and the next case is run to 954 * destroy the buffer. B_INVAL can occur if the buffer 955 * is outside the range supported by the underlying device. 956 */ 957 bp->b_flags &= ~B_ERROR; 958 bdirty(bp); 959 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_FREEBUF)) || 960 (bp->b_bufsize <= 0)) { 961 /* 962 * Either a failed I/O or we were asked to free or not 963 * cache the buffer. 964 */ 965 bp->b_flags |= B_INVAL; 966 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 967 (*bioops.io_deallocate)(bp); 968 if (bp->b_flags & B_DELWRI) { 969 --numdirtybuffers; 970 numdirtywakeup(); 971 } 972 bp->b_flags &= ~(B_DELWRI | B_CACHE | B_FREEBUF); 973 if ((bp->b_flags & B_VMIO) == 0) { 974 if (bp->b_bufsize) 975 allocbuf(bp, 0); 976 if (bp->b_vp) 977 brelvp(bp); 978 } 979 } 980 981 /* 982 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 983 * is called with B_DELWRI set, the underlying pages may wind up 984 * getting freed causing a previous write (bdwrite()) to get 'lost' 985 * because pages associated with a B_DELWRI bp are marked clean. 986 * 987 * We still allow the B_INVAL case to call vfs_vmio_release(), even 988 * if B_DELWRI is set. 989 */ 990 991 if (bp->b_flags & B_DELWRI) 992 bp->b_flags &= ~B_RELBUF; 993 994 /* 995 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 996 * constituted, not even NFS buffers now. Two flags effect this. If 997 * B_INVAL, the struct buf is invalidated but the VM object is kept 998 * around ( i.e. so it is trivial to reconstitute the buffer later ). 999 * 1000 * If B_ERROR or B_NOCACHE is set, pages in the VM object will be 1001 * invalidated. B_ERROR cannot be set for a failed write unless the 1002 * buffer is also B_INVAL because it hits the re-dirtying code above. 1003 * 1004 * Normally we can do this whether a buffer is B_DELWRI or not. If 1005 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1006 * the commit state and we cannot afford to lose the buffer. If the 1007 * buffer has a background write in progress, we need to keep it 1008 * around to prevent it from being reconstituted and starting a second 1009 * background write. 1010 */ 1011 if ((bp->b_flags & B_VMIO) 1012 && !(bp->b_vp->v_tag == VT_NFS && 1013 !vn_isdisk(bp->b_vp, NULL) && 1014 (bp->b_flags & B_DELWRI) && 1015 (bp->b_xflags & BX_BKGRDINPROG)) 1016 ) { 1017 1018 int i, j, resid; 1019 vm_page_t m; 1020 off_t foff; 1021 vm_pindex_t poff; 1022 vm_object_t obj; 1023 struct vnode *vp; 1024 1025 vp = bp->b_vp; 1026 1027 /* 1028 * Get the base offset and length of the buffer. Note that 1029 * for block sizes that are less then PAGE_SIZE, the b_data 1030 * base of the buffer does not represent exactly b_offset and 1031 * neither b_offset nor b_size are necessarily page aligned. 1032 * Instead, the starting position of b_offset is: 1033 * 1034 * b_data + (b_offset & PAGE_MASK) 1035 * 1036 * block sizes less then DEV_BSIZE (usually 512) are not 1037 * supported due to the page granularity bits (m->valid, 1038 * m->dirty, etc...). 1039 * 1040 * See man buf(9) for more information 1041 */ 1042 1043 resid = bp->b_bufsize; 1044 foff = bp->b_offset; 1045 1046 for (i = 0; i < bp->b_npages; i++) { 1047 m = bp->b_pages[i]; 1048 vm_page_flag_clear(m, PG_ZERO); 1049 if (m == bogus_page) { 1050 1051 obj = (vm_object_t) vp->v_object; 1052 poff = OFF_TO_IDX(bp->b_offset); 1053 1054 for (j = i; j < bp->b_npages; j++) { 1055 m = bp->b_pages[j]; 1056 if (m == bogus_page) { 1057 m = vm_page_lookup(obj, poff + j); 1058 if (!m) { 1059 panic("brelse: page missing\n"); 1060 } 1061 bp->b_pages[j] = m; 1062 } 1063 } 1064 1065 if ((bp->b_flags & B_INVAL) == 0) { 1066 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1067 } 1068 } 1069 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 1070 int poffset = foff & PAGE_MASK; 1071 int presid = resid > (PAGE_SIZE - poffset) ? 1072 (PAGE_SIZE - poffset) : resid; 1073 1074 KASSERT(presid >= 0, ("brelse: extra page")); 1075 vm_page_set_invalid(m, poffset, presid); 1076 } 1077 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1078 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 1079 } 1080 1081 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1082 vfs_vmio_release(bp); 1083 1084 } else if (bp->b_flags & B_VMIO) { 1085 1086 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1087 vfs_vmio_release(bp); 1088 1089 } 1090 1091 if (bp->b_qindex != QUEUE_NONE) 1092 panic("brelse: free buffer onto another queue???"); 1093 if (BUF_REFCNT(bp) > 1) { 1094 /* Temporary panic to verify exclusive locking */ 1095 /* This panic goes away when we allow shared refs */ 1096 panic("brelse: multiple refs"); 1097 /* do not release to free list */ 1098 BUF_UNLOCK(bp); 1099 splx(s); 1100 return; 1101 } 1102 1103 /* enqueue */ 1104 1105 /* buffers with no memory */ 1106 if (bp->b_bufsize == 0) { 1107 bp->b_flags |= B_INVAL; 1108 bp->b_xflags &= ~BX_BKGRDWRITE; 1109 if (bp->b_xflags & BX_BKGRDINPROG) 1110 panic("losing buffer 1"); 1111 if (bp->b_kvasize) { 1112 bp->b_qindex = QUEUE_EMPTYKVA; 1113 kvawakeup = 1; 1114 } else { 1115 bp->b_qindex = QUEUE_EMPTY; 1116 } 1117 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1118 LIST_REMOVE(bp, b_hash); 1119 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1120 bp->b_dev = NODEV; 1121 kvafreespace += bp->b_kvasize; 1122 /* buffers with junk contents */ 1123 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) { 1124 bp->b_flags |= B_INVAL; 1125 bp->b_xflags &= ~BX_BKGRDWRITE; 1126 if (bp->b_xflags & BX_BKGRDINPROG) 1127 panic("losing buffer 2"); 1128 bp->b_qindex = QUEUE_CLEAN; 1129 if (bp->b_kvasize) 1130 kvawakeup = 1; 1131 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1132 LIST_REMOVE(bp, b_hash); 1133 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1134 bp->b_dev = NODEV; 1135 1136 /* buffers that are locked */ 1137 } else if (bp->b_flags & B_LOCKED) { 1138 bp->b_qindex = QUEUE_LOCKED; 1139 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1140 1141 /* remaining buffers */ 1142 } else { 1143 switch(bp->b_flags & (B_DELWRI|B_AGE)) { 1144 case B_DELWRI | B_AGE: 1145 bp->b_qindex = QUEUE_DIRTY; 1146 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1147 break; 1148 case B_DELWRI: 1149 bp->b_qindex = QUEUE_DIRTY; 1150 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1151 break; 1152 case B_AGE: 1153 bp->b_qindex = QUEUE_CLEAN; 1154 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1155 if (bp->b_kvasize) 1156 kvawakeup = 1; 1157 break; 1158 default: 1159 bp->b_qindex = QUEUE_CLEAN; 1160 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1161 if (bp->b_kvasize) 1162 kvawakeup = 1; 1163 break; 1164 } 1165 } 1166 1167 /* 1168 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1169 * on the correct queue. 1170 */ 1171 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) { 1172 bp->b_flags &= ~B_DELWRI; 1173 --numdirtybuffers; 1174 numdirtywakeup(); 1175 } 1176 1177 runningbufspace -= bp->b_bufsize; 1178 1179 /* 1180 * Fixup numfreebuffers count. The bp is on an appropriate queue 1181 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1182 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1183 * if B_INVAL is set ). 1184 */ 1185 1186 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1187 bufcountwakeup(); 1188 1189 /* 1190 * Something we can maybe free. 1191 */ 1192 1193 if (bp->b_bufsize) 1194 bufspacewakeup(); 1195 if (kvawakeup) 1196 kvaspacewakeup(); 1197 1198 /* unlock */ 1199 BUF_UNLOCK(bp); 1200 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1201 splx(s); 1202} 1203 1204/* 1205 * Release a buffer back to the appropriate queue but do not try to free 1206 * it. 1207 * 1208 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1209 * biodone() to requeue an async I/O on completion. It is also used when 1210 * known good buffers need to be requeued but we think we may need the data 1211 * again soon. 1212 */ 1213void 1214bqrelse(struct buf * bp) 1215{ 1216 int s; 1217 1218 s = splbio(); 1219 1220 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1221 1222 if (bp->b_qindex != QUEUE_NONE) 1223 panic("bqrelse: free buffer onto another queue???"); 1224 if (BUF_REFCNT(bp) > 1) { 1225 /* do not release to free list */ 1226 panic("bqrelse: multiple refs"); 1227 BUF_UNLOCK(bp); 1228 splx(s); 1229 return; 1230 } 1231 if (bp->b_flags & B_LOCKED) { 1232 bp->b_flags &= ~B_ERROR; 1233 bp->b_qindex = QUEUE_LOCKED; 1234 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1235 /* buffers with stale but valid contents */ 1236 } else if (bp->b_flags & B_DELWRI) { 1237 bp->b_qindex = QUEUE_DIRTY; 1238 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1239 } else { 1240 bp->b_qindex = QUEUE_CLEAN; 1241 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1242 } 1243 1244 runningbufspace -= bp->b_bufsize; 1245 1246 if ((bp->b_flags & B_LOCKED) == 0 && 1247 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1248 bufcountwakeup(); 1249 } 1250 1251 /* 1252 * Something we can maybe wakeup 1253 */ 1254 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1255 bufspacewakeup(); 1256 1257 /* unlock */ 1258 BUF_UNLOCK(bp); 1259 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1260 splx(s); 1261} 1262 1263static void 1264vfs_vmio_release(bp) 1265 struct buf *bp; 1266{ 1267 int i, s; 1268 vm_page_t m; 1269 1270 s = splvm(); 1271 for (i = 0; i < bp->b_npages; i++) { 1272 m = bp->b_pages[i]; 1273 bp->b_pages[i] = NULL; 1274 /* 1275 * In order to keep page LRU ordering consistent, put 1276 * everything on the inactive queue. 1277 */ 1278 vm_page_unwire(m, 0); 1279 /* 1280 * We don't mess with busy pages, it is 1281 * the responsibility of the process that 1282 * busied the pages to deal with them. 1283 */ 1284 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1285 continue; 1286 1287 if (m->wire_count == 0) { 1288 vm_page_flag_clear(m, PG_ZERO); 1289 /* 1290 * Might as well free the page if we can and it has 1291 * no valid data. 1292 */ 1293 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { 1294 vm_page_busy(m); 1295 vm_page_protect(m, VM_PROT_NONE); 1296 vm_page_free(m); 1297 } 1298 } 1299 } 1300 bufspace -= bp->b_bufsize; 1301 vmiospace -= bp->b_bufsize; 1302 runningbufspace -= bp->b_bufsize; 1303 splx(s); 1304 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1305 if (bp->b_bufsize) 1306 bufspacewakeup(); 1307 bp->b_npages = 0; 1308 bp->b_bufsize = 0; 1309 bp->b_flags &= ~B_VMIO; 1310 if (bp->b_vp) 1311 brelvp(bp); 1312} 1313 1314/* 1315 * Check to see if a block is currently memory resident. 1316 */ 1317struct buf * 1318gbincore(struct vnode * vp, daddr_t blkno) 1319{ 1320 struct buf *bp; 1321 struct bufhashhdr *bh; 1322 1323 bh = bufhash(vp, blkno); 1324 1325 /* Search hash chain */ 1326 LIST_FOREACH(bp, bh, b_hash) { 1327 /* hit */ 1328 if (bp->b_vp == vp && bp->b_lblkno == blkno && 1329 (bp->b_flags & B_INVAL) == 0) { 1330 break; 1331 } 1332 } 1333 return (bp); 1334} 1335 1336/* 1337 * vfs_bio_awrite: 1338 * 1339 * Implement clustered async writes for clearing out B_DELWRI buffers. 1340 * This is much better then the old way of writing only one buffer at 1341 * a time. Note that we may not be presented with the buffers in the 1342 * correct order, so we search for the cluster in both directions. 1343 */ 1344int 1345vfs_bio_awrite(struct buf * bp) 1346{ 1347 int i; 1348 int j; 1349 daddr_t lblkno = bp->b_lblkno; 1350 struct vnode *vp = bp->b_vp; 1351 int s; 1352 int ncl; 1353 struct buf *bpa; 1354 int nwritten; 1355 int size; 1356 int maxcl; 1357 1358 s = splbio(); 1359 /* 1360 * right now we support clustered writing only to regular files. If 1361 * we find a clusterable block we could be in the middle of a cluster 1362 * rather then at the beginning. 1363 */ 1364 if ((vp->v_type == VREG) && 1365 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1366 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1367 1368 size = vp->v_mount->mnt_stat.f_iosize; 1369 maxcl = MAXPHYS / size; 1370 1371 for (i = 1; i < maxcl; i++) { 1372 if ((bpa = gbincore(vp, lblkno + i)) && 1373 BUF_REFCNT(bpa) == 0 && 1374 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1375 (B_DELWRI | B_CLUSTEROK)) && 1376 (bpa->b_bufsize == size)) { 1377 if ((bpa->b_blkno == bpa->b_lblkno) || 1378 (bpa->b_blkno != 1379 bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 1380 break; 1381 } else { 1382 break; 1383 } 1384 } 1385 for (j = 1; i + j <= maxcl && j <= lblkno; j++) { 1386 if ((bpa = gbincore(vp, lblkno - j)) && 1387 BUF_REFCNT(bpa) == 0 && 1388 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1389 (B_DELWRI | B_CLUSTEROK)) && 1390 (bpa->b_bufsize == size)) { 1391 if ((bpa->b_blkno == bpa->b_lblkno) || 1392 (bpa->b_blkno != 1393 bp->b_blkno - ((j * size) >> DEV_BSHIFT))) 1394 break; 1395 } else { 1396 break; 1397 } 1398 } 1399 --j; 1400 ncl = i + j; 1401 /* 1402 * this is a possible cluster write 1403 */ 1404 if (ncl != 1) { 1405 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1406 splx(s); 1407 return nwritten; 1408 } 1409 } 1410 1411 BUF_LOCK(bp, LK_EXCLUSIVE); 1412 bremfree(bp); 1413 bp->b_flags |= B_ASYNC; 1414 1415 splx(s); 1416 /* 1417 * default (old) behavior, writing out only one block 1418 * 1419 * XXX returns b_bufsize instead of b_bcount for nwritten? 1420 */ 1421 nwritten = bp->b_bufsize; 1422 (void) VOP_BWRITE(bp->b_vp, bp); 1423 1424 return nwritten; 1425} 1426 1427/* 1428 * getnewbuf: 1429 * 1430 * Find and initialize a new buffer header, freeing up existing buffers 1431 * in the bufqueues as necessary. The new buffer is returned locked. 1432 * 1433 * Important: B_INVAL is not set. If the caller wishes to throw the 1434 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1435 * 1436 * We block if: 1437 * We have insufficient buffer headers 1438 * We have insufficient buffer space 1439 * buffer_map is too fragmented ( space reservation fails ) 1440 * If we have to flush dirty buffers ( but we try to avoid this ) 1441 * 1442 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1443 * Instead we ask the buf daemon to do it for us. We attempt to 1444 * avoid piecemeal wakeups of the pageout daemon. 1445 */ 1446 1447static struct buf * 1448getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1449{ 1450 struct buf *bp; 1451 struct buf *nbp; 1452 struct buf *dbp; 1453 int outofspace; 1454 int nqindex; 1455 int defrag = 0; 1456 1457 ++getnewbufcalls; 1458 --getnewbufrestarts; 1459restart: 1460 ++getnewbufrestarts; 1461 1462 /* 1463 * Calculate whether we are out of buffer space. This state is 1464 * recalculated on every restart. If we are out of space, we 1465 * have to turn off defragmentation. Setting defrag to -1 when 1466 * outofspace is positive means "defrag while freeing buffers". 1467 * The looping conditional will be muffed up if defrag is left 1468 * positive when outofspace is positive. 1469 */ 1470 1471 dbp = NULL; 1472 outofspace = 0; 1473 if (bufspace >= hibufspace) { 1474 if ((curproc && (curproc->p_flag & P_BUFEXHAUST) == 0) || 1475 bufspace >= maxbufspace) { 1476 outofspace = 1; 1477 if (defrag > 0) 1478 defrag = -1; 1479 } 1480 } 1481 1482 /* 1483 * defrag state is semi-persistant. 1 means we are flagged for 1484 * defragging. -1 means we actually defragged something. 1485 */ 1486 /* nop */ 1487 1488 /* 1489 * Setup for scan. If we do not have enough free buffers, 1490 * we setup a degenerate case that immediately fails. Note 1491 * that if we are specially marked process, we are allowed to 1492 * dip into our reserves. 1493 * 1494 * Normally we want to find an EMPTYKVA buffer. That is, a 1495 * buffer with kva already allocated. If there are no EMPTYKVA 1496 * buffers we back up to the truely EMPTY buffers. When defragging 1497 * we do not bother backing up since we have to locate buffers with 1498 * kva to defrag. If we are out of space we skip both EMPTY and 1499 * EMPTYKVA and dig right into the CLEAN queue. 1500 * 1501 * In this manner we avoid scanning unnecessary buffers. It is very 1502 * important for us to do this because the buffer cache is almost 1503 * constantly out of space or in need of defragmentation. 1504 */ 1505 1506 if (curproc && (curproc->p_flag & P_BUFEXHAUST) == 0 && 1507 numfreebuffers < lofreebuffers) { 1508 nqindex = QUEUE_CLEAN; 1509 nbp = NULL; 1510 } else { 1511 nqindex = QUEUE_EMPTYKVA; 1512 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1513 if (nbp == NULL) { 1514 if (defrag <= 0) { 1515 nqindex = QUEUE_EMPTY; 1516 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1517 } 1518 } 1519 if (outofspace || nbp == NULL) { 1520 nqindex = QUEUE_CLEAN; 1521 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1522 } 1523 } 1524 1525 /* 1526 * Run scan, possibly freeing data and/or kva mappings on the fly 1527 * depending. 1528 */ 1529 1530 while ((bp = nbp) != NULL) { 1531 int qindex = nqindex; 1532 1533 /* 1534 * Calculate next bp ( we can only use it if we do not block 1535 * or do other fancy things ). 1536 */ 1537 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1538 switch(qindex) { 1539 case QUEUE_EMPTY: 1540 nqindex = QUEUE_EMPTYKVA; 1541 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1542 break; 1543 /* fall through */ 1544 case QUEUE_EMPTYKVA: 1545 nqindex = QUEUE_CLEAN; 1546 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1547 break; 1548 /* fall through */ 1549 case QUEUE_CLEAN: 1550 /* 1551 * nbp is NULL. 1552 */ 1553 break; 1554 } 1555 } 1556 1557 /* 1558 * Sanity Checks 1559 */ 1560 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1561 1562 /* 1563 * Note: we no longer distinguish between VMIO and non-VMIO 1564 * buffers. 1565 */ 1566 1567 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1568 1569 /* 1570 * If we are defragging and the buffer isn't useful for fixing 1571 * that problem we continue. If we are out of space and the 1572 * buffer isn't useful for fixing that problem we continue. 1573 */ 1574 1575 if (defrag > 0 && bp->b_kvasize == 0) 1576 continue; 1577 if (outofspace > 0 && bp->b_bufsize == 0) 1578 continue; 1579 1580 /* 1581 * Start freeing the bp. This is somewhat involved. nbp 1582 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1583 */ 1584 1585 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1586 panic("getnewbuf: locked buf"); 1587 bremfree(bp); 1588 1589 if (qindex == QUEUE_CLEAN) { 1590 if (bp->b_flags & B_VMIO) { 1591 bp->b_flags &= ~B_ASYNC; 1592 vfs_vmio_release(bp); 1593 } 1594 if (bp->b_vp) 1595 brelvp(bp); 1596 } 1597 1598 /* 1599 * NOTE: nbp is now entirely invalid. We can only restart 1600 * the scan from this point on. 1601 * 1602 * Get the rest of the buffer freed up. b_kva* is still 1603 * valid after this operation. 1604 */ 1605 1606 if (bp->b_rcred != NOCRED) { 1607 crfree(bp->b_rcred); 1608 bp->b_rcred = NOCRED; 1609 } 1610 if (bp->b_wcred != NOCRED) { 1611 crfree(bp->b_wcred); 1612 bp->b_wcred = NOCRED; 1613 } 1614 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 1615 (*bioops.io_deallocate)(bp); 1616 if (bp->b_xflags & BX_BKGRDINPROG) 1617 panic("losing buffer 3"); 1618 LIST_REMOVE(bp, b_hash); 1619 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1620 1621 if (bp->b_bufsize) 1622 allocbuf(bp, 0); 1623 1624 bp->b_flags = 0; 1625 bp->b_xflags = 0; 1626 bp->b_dev = NODEV; 1627 bp->b_vp = NULL; 1628 bp->b_blkno = bp->b_lblkno = 0; 1629 bp->b_offset = NOOFFSET; 1630 bp->b_iodone = 0; 1631 bp->b_error = 0; 1632 bp->b_resid = 0; 1633 bp->b_bcount = 0; 1634 bp->b_npages = 0; 1635 bp->b_dirtyoff = bp->b_dirtyend = 0; 1636 1637 LIST_INIT(&bp->b_dep); 1638 1639 /* 1640 * Ok, now that we have a free buffer, if we are defragging 1641 * we have to recover the kvaspace. If we are out of space 1642 * we have to free the buffer (which we just did), but we 1643 * do not have to recover kva space unless we hit a defrag 1644 * hicup. Being able to avoid freeing the kva space leads 1645 * to a significant reduction in overhead. 1646 */ 1647 1648 if (defrag > 0) { 1649 defrag = -1; 1650 bp->b_flags |= B_INVAL; 1651 bfreekva(bp); 1652 brelse(bp); 1653 goto restart; 1654 } 1655 1656 if (outofspace > 0) { 1657 outofspace = -1; 1658 bp->b_flags |= B_INVAL; 1659 if (defrag < 0) 1660 bfreekva(bp); 1661 brelse(bp); 1662 goto restart; 1663 } 1664 1665 /* 1666 * We are done 1667 */ 1668 break; 1669 } 1670 1671 /* 1672 * If we exhausted our list, sleep as appropriate. We may have to 1673 * wakeup various daemons and write out some dirty buffers. 1674 * 1675 * Generally we are sleeping due to insufficient buffer space. 1676 */ 1677 1678 if (bp == NULL) { 1679 int flags; 1680 char *waitmsg; 1681 1682 if (defrag > 0) { 1683 flags = VFS_BIO_NEED_KVASPACE; 1684 waitmsg = "nbufkv"; 1685 } else if (outofspace > 0) { 1686 waitmsg = "nbufbs"; 1687 flags = VFS_BIO_NEED_BUFSPACE; 1688 } else { 1689 waitmsg = "newbuf"; 1690 flags = VFS_BIO_NEED_ANY; 1691 } 1692 1693 bd_speedup(); /* heeeelp */ 1694 1695 needsbuffer |= flags; 1696 while (needsbuffer & flags) { 1697 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, 1698 waitmsg, slptimeo)) 1699 return (NULL); 1700 } 1701 } else { 1702 /* 1703 * We finally have a valid bp. We aren't quite out of the 1704 * woods, we still have to reserve kva space. 1705 */ 1706 vm_offset_t addr = 0; 1707 1708 maxsize = (maxsize + PAGE_MASK) & ~PAGE_MASK; 1709 1710 if (maxsize != bp->b_kvasize) { 1711 bfreekva(bp); 1712 1713 if (vm_map_findspace(buffer_map, 1714 vm_map_min(buffer_map), maxsize, &addr)) { 1715 /* 1716 * Uh oh. Buffer map is to fragmented. Try 1717 * to defragment. 1718 */ 1719 if (defrag <= 0) { 1720 defrag = 1; 1721 bp->b_flags |= B_INVAL; 1722 brelse(bp); 1723 goto restart; 1724 } 1725 /* 1726 * Uh oh. We couldn't seem to defragment 1727 */ 1728 panic("getnewbuf: unreachable code reached"); 1729 } 1730 } 1731 if (addr) { 1732 vm_map_insert(buffer_map, NULL, 0, 1733 addr, addr + maxsize, 1734 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1735 1736 bp->b_kvabase = (caddr_t) addr; 1737 bp->b_kvasize = maxsize; 1738 } 1739 bp->b_data = bp->b_kvabase; 1740 } 1741 return(bp); 1742} 1743 1744/* 1745 * waitfreebuffers: 1746 * 1747 * Wait for sufficient free buffers. Only called from normal processes. 1748 */ 1749 1750static void 1751waitfreebuffers(int slpflag, int slptimeo) 1752{ 1753 while (numfreebuffers < hifreebuffers) { 1754 if (numfreebuffers >= hifreebuffers) 1755 break; 1756 needsbuffer |= VFS_BIO_NEED_FREE; 1757 if (tsleep(&needsbuffer, (PRIBIO + 4)|slpflag, "biofre", slptimeo)) 1758 break; 1759 } 1760} 1761 1762/* 1763 * buf_daemon: 1764 * 1765 * buffer flushing daemon. Buffers are normally flushed by the 1766 * update daemon but if it cannot keep up this process starts to 1767 * take the load in an attempt to prevent getnewbuf() from blocking. 1768 */ 1769 1770static struct proc *bufdaemonproc; 1771static int bd_interval; 1772static int bd_flushto; 1773static int bd_flushinc; 1774 1775static struct kproc_desc buf_kp = { 1776 "bufdaemon", 1777 buf_daemon, 1778 &bufdaemonproc 1779}; 1780SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1781 1782static void 1783buf_daemon() 1784{ 1785 int s; 1786 1787 /* 1788 * This process needs to be suspended prior to shutdown sync. 1789 */ 1790 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, bufdaemonproc, 1791 SHUTDOWN_PRI_LAST); 1792 1793 /* 1794 * This process is allowed to take the buffer cache to the limit 1795 */ 1796 curproc->p_flag |= P_BUFEXHAUST; 1797 s = splbio(); 1798 1799 bd_interval = 5 * hz; /* dynamically adjusted */ 1800 bd_flushto = hidirtybuffers; /* dynamically adjusted */ 1801 bd_flushinc = 1; 1802 1803 for (;;) { 1804 kproc_suspend_loop(bufdaemonproc); 1805 1806 bd_request = 0; 1807 1808 /* 1809 * Do the flush. Limit the number of buffers we flush in one 1810 * go. The failure condition occurs when processes are writing 1811 * buffers faster then we can dispose of them. In this case 1812 * we may be flushing so often that the previous set of flushes 1813 * have not had time to complete, causing us to run out of 1814 * physical buffers and block. 1815 */ 1816 { 1817 int runcount = maxbdrun; 1818 1819 while (numdirtybuffers > bd_flushto && runcount) { 1820 --runcount; 1821 if (flushbufqueues() == 0) 1822 break; 1823 } 1824 } 1825 1826 if (bd_request || 1827 tsleep(&bd_request, PVM, "psleep", bd_interval) == 0) { 1828 /* 1829 * Another request is pending or we were woken up 1830 * without timing out. Flush more. 1831 */ 1832 --bd_flushto; 1833 if (bd_flushto >= numdirtybuffers - 5) { 1834 bd_flushto = numdirtybuffers - 10; 1835 bd_flushinc = 1; 1836 } 1837 if (bd_flushto < 2) 1838 bd_flushto = 2; 1839 } else { 1840 /* 1841 * We slept and timed out, we can slow down. 1842 */ 1843 bd_flushto += bd_flushinc; 1844 if (bd_flushto > hidirtybuffers) 1845 bd_flushto = hidirtybuffers; 1846 ++bd_flushinc; 1847 if (bd_flushinc > hidirtybuffers / 20 + 1) 1848 bd_flushinc = hidirtybuffers / 20 + 1; 1849 } 1850 1851 /* 1852 * Set the interval on a linear scale based on hidirtybuffers 1853 * with a maximum frequency of 1/10 second. 1854 */ 1855 bd_interval = bd_flushto * 5 * hz / hidirtybuffers; 1856 if (bd_interval < hz / 10) 1857 bd_interval = hz / 10; 1858 } 1859} 1860 1861/* 1862 * flushbufqueues: 1863 * 1864 * Try to flush a buffer in the dirty queue. We must be careful to 1865 * free up B_INVAL buffers instead of write them, which NFS is 1866 * particularly sensitive to. 1867 */ 1868 1869static int 1870flushbufqueues(void) 1871{ 1872 struct buf *bp; 1873 int r = 0; 1874 1875 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1876 1877 while (bp) { 1878 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); 1879 if ((bp->b_flags & B_DELWRI) != 0 && 1880 (bp->b_xflags & BX_BKGRDINPROG) == 0) { 1881 if (bp->b_flags & B_INVAL) { 1882 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1883 panic("flushbufqueues: locked buf"); 1884 bremfree(bp); 1885 brelse(bp); 1886 ++r; 1887 break; 1888 } 1889 if (LIST_FIRST(&bp->b_dep) != NULL && 1890 bioops.io_countdeps && 1891 (bp->b_flags & B_DEFERRED) == 0 && 1892 (*bioops.io_countdeps)(bp, 0)) { 1893 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], 1894 bp, b_freelist); 1895 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], 1896 bp, b_freelist); 1897 bp->b_flags |= B_DEFERRED; 1898 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1899 continue; 1900 } 1901 vfs_bio_awrite(bp); 1902 ++r; 1903 break; 1904 } 1905 bp = TAILQ_NEXT(bp, b_freelist); 1906 } 1907 return (r); 1908} 1909 1910/* 1911 * Check to see if a block is currently memory resident. 1912 */ 1913struct buf * 1914incore(struct vnode * vp, daddr_t blkno) 1915{ 1916 struct buf *bp; 1917 1918 int s = splbio(); 1919 bp = gbincore(vp, blkno); 1920 splx(s); 1921 return (bp); 1922} 1923 1924/* 1925 * Returns true if no I/O is needed to access the 1926 * associated VM object. This is like incore except 1927 * it also hunts around in the VM system for the data. 1928 */ 1929 1930int 1931inmem(struct vnode * vp, daddr_t blkno) 1932{ 1933 vm_object_t obj; 1934 vm_offset_t toff, tinc, size; 1935 vm_page_t m; 1936 vm_ooffset_t off; 1937 1938 if (incore(vp, blkno)) 1939 return 1; 1940 if (vp->v_mount == NULL) 1941 return 0; 1942 if ((vp->v_object == NULL) || (vp->v_flag & VOBJBUF) == 0) 1943 return 0; 1944 1945 obj = vp->v_object; 1946 size = PAGE_SIZE; 1947 if (size > vp->v_mount->mnt_stat.f_iosize) 1948 size = vp->v_mount->mnt_stat.f_iosize; 1949 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 1950 1951 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1952 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 1953 if (!m) 1954 return 0; 1955 tinc = size; 1956 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 1957 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 1958 if (vm_page_is_valid(m, 1959 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 1960 return 0; 1961 } 1962 return 1; 1963} 1964 1965/* 1966 * vfs_setdirty: 1967 * 1968 * Sets the dirty range for a buffer based on the status of the dirty 1969 * bits in the pages comprising the buffer. 1970 * 1971 * The range is limited to the size of the buffer. 1972 * 1973 * This routine is primarily used by NFS, but is generalized for the 1974 * B_VMIO case. 1975 */ 1976static void 1977vfs_setdirty(struct buf *bp) 1978{ 1979 int i; 1980 vm_object_t object; 1981 1982 /* 1983 * Degenerate case - empty buffer 1984 */ 1985 1986 if (bp->b_bufsize == 0) 1987 return; 1988 1989 /* 1990 * We qualify the scan for modified pages on whether the 1991 * object has been flushed yet. The OBJ_WRITEABLE flag 1992 * is not cleared simply by protecting pages off. 1993 */ 1994 1995 if ((bp->b_flags & B_VMIO) == 0) 1996 return; 1997 1998 object = bp->b_pages[0]->object; 1999 2000 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2001 printf("Warning: object %p writeable but not mightbedirty\n", object); 2002 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2003 printf("Warning: object %p mightbedirty but not writeable\n", object); 2004 2005 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2006 vm_offset_t boffset; 2007 vm_offset_t eoffset; 2008 2009 /* 2010 * test the pages to see if they have been modified directly 2011 * by users through the VM system. 2012 */ 2013 for (i = 0; i < bp->b_npages; i++) { 2014 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2015 vm_page_test_dirty(bp->b_pages[i]); 2016 } 2017 2018 /* 2019 * Calculate the encompassing dirty range, boffset and eoffset, 2020 * (eoffset - boffset) bytes. 2021 */ 2022 2023 for (i = 0; i < bp->b_npages; i++) { 2024 if (bp->b_pages[i]->dirty) 2025 break; 2026 } 2027 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2028 2029 for (i = bp->b_npages - 1; i >= 0; --i) { 2030 if (bp->b_pages[i]->dirty) { 2031 break; 2032 } 2033 } 2034 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2035 2036 /* 2037 * Fit it to the buffer. 2038 */ 2039 2040 if (eoffset > bp->b_bcount) 2041 eoffset = bp->b_bcount; 2042 2043 /* 2044 * If we have a good dirty range, merge with the existing 2045 * dirty range. 2046 */ 2047 2048 if (boffset < eoffset) { 2049 if (bp->b_dirtyoff > boffset) 2050 bp->b_dirtyoff = boffset; 2051 if (bp->b_dirtyend < eoffset) 2052 bp->b_dirtyend = eoffset; 2053 } 2054 } 2055} 2056 2057/* 2058 * getblk: 2059 * 2060 * Get a block given a specified block and offset into a file/device. 2061 * The buffers B_DONE bit will be cleared on return, making it almost 2062 * ready for an I/O initiation. B_INVAL may or may not be set on 2063 * return. The caller should clear B_INVAL prior to initiating a 2064 * READ. 2065 * 2066 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2067 * an existing buffer. 2068 * 2069 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2070 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2071 * and then cleared based on the backing VM. If the previous buffer is 2072 * non-0-sized but invalid, B_CACHE will be cleared. 2073 * 2074 * If getblk() must create a new buffer, the new buffer is returned with 2075 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2076 * case it is returned with B_INVAL clear and B_CACHE set based on the 2077 * backing VM. 2078 * 2079 * getblk() also forces a VOP_BWRITE() for any B_DELWRI buffer whos 2080 * B_CACHE bit is clear. 2081 * 2082 * What this means, basically, is that the caller should use B_CACHE to 2083 * determine whether the buffer is fully valid or not and should clear 2084 * B_INVAL prior to issuing a read. If the caller intends to validate 2085 * the buffer by loading its data area with something, the caller needs 2086 * to clear B_INVAL. If the caller does this without issuing an I/O, 2087 * the caller should set B_CACHE ( as an optimization ), else the caller 2088 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2089 * a write attempt or if it was a successfull read. If the caller 2090 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR 2091 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2092 */ 2093struct buf * 2094getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2095{ 2096 struct buf *bp; 2097 int s; 2098 struct bufhashhdr *bh; 2099 2100 if (size > MAXBSIZE) 2101 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2102 2103 s = splbio(); 2104loop: 2105 /* 2106 * Block if we are low on buffers. Certain processes are allowed 2107 * to completely exhaust the buffer cache. 2108 * 2109 * If this check ever becomes a bottleneck it may be better to 2110 * move it into the else, when gbincore() fails. At the moment 2111 * it isn't a problem. 2112 */ 2113 if (!curproc || (curproc->p_flag & P_BUFEXHAUST)) { 2114 if (numfreebuffers == 0) { 2115 if (!curproc) 2116 return NULL; 2117 needsbuffer |= VFS_BIO_NEED_ANY; 2118 tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, "newbuf", 2119 slptimeo); 2120 } 2121 } else if (numfreebuffers < lofreebuffers) { 2122 waitfreebuffers(slpflag, slptimeo); 2123 } 2124 2125 if ((bp = gbincore(vp, blkno))) { 2126 /* 2127 * Buffer is in-core. If the buffer is not busy, it must 2128 * be on a queue. 2129 */ 2130 2131 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2132 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, 2133 "getblk", slpflag, slptimeo) == ENOLCK) 2134 goto loop; 2135 splx(s); 2136 return (struct buf *) NULL; 2137 } 2138 2139 /* 2140 * The buffer is locked. B_CACHE is cleared if the buffer is 2141 * invalid. Ohterwise, for a non-VMIO buffer, B_CACHE is set 2142 * and for a VMIO buffer B_CACHE is adjusted according to the 2143 * backing VM cache. 2144 */ 2145 if (bp->b_flags & B_INVAL) 2146 bp->b_flags &= ~B_CACHE; 2147 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2148 bp->b_flags |= B_CACHE; 2149 bremfree(bp); 2150 2151 /* 2152 * check for size inconsistancies for non-VMIO case. 2153 */ 2154 2155 if (bp->b_bcount != size) { 2156 if ((bp->b_flags & B_VMIO) == 0 || 2157 (size > bp->b_kvasize)) { 2158 if (bp->b_flags & B_DELWRI) { 2159 bp->b_flags |= B_NOCACHE; 2160 VOP_BWRITE(bp->b_vp, bp); 2161 } else { 2162 if ((bp->b_flags & B_VMIO) && 2163 (LIST_FIRST(&bp->b_dep) == NULL)) { 2164 bp->b_flags |= B_RELBUF; 2165 brelse(bp); 2166 } else { 2167 bp->b_flags |= B_NOCACHE; 2168 VOP_BWRITE(bp->b_vp, bp); 2169 } 2170 } 2171 goto loop; 2172 } 2173 } 2174 2175 /* 2176 * If the size is inconsistant in the VMIO case, we can resize 2177 * the buffer. This might lead to B_CACHE getting set or 2178 * cleared. If the size has not changed, B_CACHE remains 2179 * unchanged from its previous state. 2180 */ 2181 2182 if (bp->b_bcount != size) 2183 allocbuf(bp, size); 2184 2185 KASSERT(bp->b_offset != NOOFFSET, 2186 ("getblk: no buffer offset")); 2187 2188 /* 2189 * A buffer with B_DELWRI set and B_CACHE clear must 2190 * be committed before we can return the buffer in 2191 * order to prevent the caller from issuing a read 2192 * ( due to B_CACHE not being set ) and overwriting 2193 * it. 2194 * 2195 * Most callers, including NFS and FFS, need this to 2196 * operate properly either because they assume they 2197 * can issue a read if B_CACHE is not set, or because 2198 * ( for example ) an uncached B_DELWRI might loop due 2199 * to softupdates re-dirtying the buffer. In the latter 2200 * case, B_CACHE is set after the first write completes, 2201 * preventing further loops. 2202 */ 2203 2204 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2205 VOP_BWRITE(bp->b_vp, bp); 2206 goto loop; 2207 } 2208 2209 splx(s); 2210 bp->b_flags &= ~B_DONE; 2211 } else { 2212 /* 2213 * Buffer is not in-core, create new buffer. The buffer 2214 * returned by getnewbuf() is locked. Note that the returned 2215 * buffer is also considered valid (not marked B_INVAL). 2216 */ 2217 int bsize, maxsize, vmio; 2218 off_t offset; 2219 2220 if (vn_isdisk(vp, NULL)) 2221 bsize = DEV_BSIZE; 2222 else if (vp->v_mountedhere) 2223 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2224 else if (vp->v_mount) 2225 bsize = vp->v_mount->mnt_stat.f_iosize; 2226 else 2227 bsize = size; 2228 2229 offset = (off_t)blkno * bsize; 2230 vmio = (vp->v_object != 0) && (vp->v_flag & VOBJBUF); 2231 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2232 maxsize = imax(maxsize, bsize); 2233 2234 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2235 if (slpflag || slptimeo) { 2236 splx(s); 2237 return NULL; 2238 } 2239 goto loop; 2240 } 2241 2242 /* 2243 * This code is used to make sure that a buffer is not 2244 * created while the getnewbuf routine is blocked. 2245 * This can be a problem whether the vnode is locked or not. 2246 * If the buffer is created out from under us, we have to 2247 * throw away the one we just created. There is now window 2248 * race because we are safely running at splbio() from the 2249 * point of the duplicate buffer creation through to here, 2250 * and we've locked the buffer. 2251 */ 2252 if (gbincore(vp, blkno)) { 2253 bp->b_flags |= B_INVAL; 2254 brelse(bp); 2255 goto loop; 2256 } 2257 2258 /* 2259 * Insert the buffer into the hash, so that it can 2260 * be found by incore. 2261 */ 2262 bp->b_blkno = bp->b_lblkno = blkno; 2263 bp->b_offset = offset; 2264 2265 bgetvp(vp, bp); 2266 LIST_REMOVE(bp, b_hash); 2267 bh = bufhash(vp, blkno); 2268 LIST_INSERT_HEAD(bh, bp, b_hash); 2269 2270 /* 2271 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2272 * buffer size starts out as 0, B_CACHE will be set by 2273 * allocbuf() for the VMIO case prior to it testing the 2274 * backing store for validity. 2275 */ 2276 2277 if (vmio) { 2278 bp->b_flags |= B_VMIO; 2279#if defined(VFS_BIO_DEBUG) 2280 if (vp->v_type != VREG && vp->v_type != VBLK) 2281 printf("getblk: vmioing file type %d???\n", vp->v_type); 2282#endif 2283 } else { 2284 bp->b_flags &= ~B_VMIO; 2285 } 2286 2287 allocbuf(bp, size); 2288 2289 splx(s); 2290 bp->b_flags &= ~B_DONE; 2291 } 2292 return (bp); 2293} 2294 2295/* 2296 * Get an empty, disassociated buffer of given size. The buffer is initially 2297 * set to B_INVAL. 2298 */ 2299struct buf * 2300geteblk(int size) 2301{ 2302 struct buf *bp; 2303 int s; 2304 2305 s = splbio(); 2306 while ((bp = getnewbuf(0, 0, size, MAXBSIZE)) == 0); 2307 splx(s); 2308 allocbuf(bp, size); 2309 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2310 return (bp); 2311} 2312 2313 2314/* 2315 * This code constitutes the buffer memory from either anonymous system 2316 * memory (in the case of non-VMIO operations) or from an associated 2317 * VM object (in the case of VMIO operations). This code is able to 2318 * resize a buffer up or down. 2319 * 2320 * Note that this code is tricky, and has many complications to resolve 2321 * deadlock or inconsistant data situations. Tread lightly!!! 2322 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2323 * the caller. Calling this code willy nilly can result in the loss of data. 2324 * 2325 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2326 * B_CACHE for the non-VMIO case. 2327 */ 2328 2329int 2330allocbuf(struct buf *bp, int size) 2331{ 2332 int newbsize, mbsize; 2333 int i; 2334 2335 if (BUF_REFCNT(bp) == 0) 2336 panic("allocbuf: buffer not busy"); 2337 2338 if (bp->b_kvasize < size) 2339 panic("allocbuf: buffer too small"); 2340 2341 if ((bp->b_flags & B_VMIO) == 0) { 2342 caddr_t origbuf; 2343 int origbufsize; 2344 /* 2345 * Just get anonymous memory from the kernel. Don't 2346 * mess with B_CACHE. 2347 */ 2348 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2349#if !defined(NO_B_MALLOC) 2350 if (bp->b_flags & B_MALLOC) 2351 newbsize = mbsize; 2352 else 2353#endif 2354 newbsize = round_page(size); 2355 2356 if (newbsize < bp->b_bufsize) { 2357#if !defined(NO_B_MALLOC) 2358 /* 2359 * malloced buffers are not shrunk 2360 */ 2361 if (bp->b_flags & B_MALLOC) { 2362 if (newbsize) { 2363 bp->b_bcount = size; 2364 } else { 2365 free(bp->b_data, M_BIOBUF); 2366 bufspace -= bp->b_bufsize; 2367 bufmallocspace -= bp->b_bufsize; 2368 runningbufspace -= bp->b_bufsize; 2369 if (bp->b_bufsize) 2370 bufspacewakeup(); 2371 bp->b_data = bp->b_kvabase; 2372 bp->b_bufsize = 0; 2373 bp->b_bcount = 0; 2374 bp->b_flags &= ~B_MALLOC; 2375 } 2376 return 1; 2377 } 2378#endif 2379 vm_hold_free_pages( 2380 bp, 2381 (vm_offset_t) bp->b_data + newbsize, 2382 (vm_offset_t) bp->b_data + bp->b_bufsize); 2383 } else if (newbsize > bp->b_bufsize) { 2384#if !defined(NO_B_MALLOC) 2385 /* 2386 * We only use malloced memory on the first allocation. 2387 * and revert to page-allocated memory when the buffer 2388 * grows. 2389 */ 2390 if ( (bufmallocspace < maxbufmallocspace) && 2391 (bp->b_bufsize == 0) && 2392 (mbsize <= PAGE_SIZE/2)) { 2393 2394 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2395 bp->b_bufsize = mbsize; 2396 bp->b_bcount = size; 2397 bp->b_flags |= B_MALLOC; 2398 bufspace += mbsize; 2399 bufmallocspace += mbsize; 2400 runningbufspace += bp->b_bufsize; 2401 return 1; 2402 } 2403#endif 2404 origbuf = NULL; 2405 origbufsize = 0; 2406#if !defined(NO_B_MALLOC) 2407 /* 2408 * If the buffer is growing on its other-than-first allocation, 2409 * then we revert to the page-allocation scheme. 2410 */ 2411 if (bp->b_flags & B_MALLOC) { 2412 origbuf = bp->b_data; 2413 origbufsize = bp->b_bufsize; 2414 bp->b_data = bp->b_kvabase; 2415 bufspace -= bp->b_bufsize; 2416 bufmallocspace -= bp->b_bufsize; 2417 runningbufspace -= bp->b_bufsize; 2418 if (bp->b_bufsize) 2419 bufspacewakeup(); 2420 bp->b_bufsize = 0; 2421 bp->b_flags &= ~B_MALLOC; 2422 newbsize = round_page(newbsize); 2423 } 2424#endif 2425 vm_hold_load_pages( 2426 bp, 2427 (vm_offset_t) bp->b_data + bp->b_bufsize, 2428 (vm_offset_t) bp->b_data + newbsize); 2429#if !defined(NO_B_MALLOC) 2430 if (origbuf) { 2431 bcopy(origbuf, bp->b_data, origbufsize); 2432 free(origbuf, M_BIOBUF); 2433 } 2434#endif 2435 } 2436 } else { 2437 vm_page_t m; 2438 int desiredpages; 2439 2440 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2441 desiredpages = (size == 0) ? 0 : 2442 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2443 2444#if !defined(NO_B_MALLOC) 2445 if (bp->b_flags & B_MALLOC) 2446 panic("allocbuf: VMIO buffer can't be malloced"); 2447#endif 2448 /* 2449 * Set B_CACHE initially if buffer is 0 length or will become 2450 * 0-length. 2451 */ 2452 if (size == 0 || bp->b_bufsize == 0) 2453 bp->b_flags |= B_CACHE; 2454 2455 if (newbsize < bp->b_bufsize) { 2456 /* 2457 * DEV_BSIZE aligned new buffer size is less then the 2458 * DEV_BSIZE aligned existing buffer size. Figure out 2459 * if we have to remove any pages. 2460 */ 2461 if (desiredpages < bp->b_npages) { 2462 for (i = desiredpages; i < bp->b_npages; i++) { 2463 /* 2464 * the page is not freed here -- it 2465 * is the responsibility of 2466 * vnode_pager_setsize 2467 */ 2468 m = bp->b_pages[i]; 2469 KASSERT(m != bogus_page, 2470 ("allocbuf: bogus page found")); 2471 while (vm_page_sleep_busy(m, TRUE, "biodep")) 2472 ; 2473 2474 bp->b_pages[i] = NULL; 2475 vm_page_unwire(m, 0); 2476 } 2477 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2478 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2479 bp->b_npages = desiredpages; 2480 } 2481 } else if (size > bp->b_bcount) { 2482 /* 2483 * We are growing the buffer, possibly in a 2484 * byte-granular fashion. 2485 */ 2486 struct vnode *vp; 2487 vm_object_t obj; 2488 vm_offset_t toff; 2489 vm_offset_t tinc; 2490 2491 /* 2492 * Step 1, bring in the VM pages from the object, 2493 * allocating them if necessary. We must clear 2494 * B_CACHE if these pages are not valid for the 2495 * range covered by the buffer. 2496 */ 2497 2498 vp = bp->b_vp; 2499 obj = vp->v_object; 2500 2501 while (bp->b_npages < desiredpages) { 2502 vm_page_t m; 2503 vm_pindex_t pi; 2504 2505 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2506 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2507 m = vm_page_alloc(obj, pi, VM_ALLOC_NORMAL); 2508 if (m == NULL) { 2509 VM_WAIT; 2510 vm_pageout_deficit += desiredpages - bp->b_npages; 2511 } else { 2512 vm_page_wire(m); 2513 vm_page_wakeup(m); 2514 bp->b_flags &= ~B_CACHE; 2515 bp->b_pages[bp->b_npages] = m; 2516 ++bp->b_npages; 2517 } 2518 continue; 2519 } 2520 2521 /* 2522 * We found a page. If we have to sleep on it, 2523 * retry because it might have gotten freed out 2524 * from under us. 2525 * 2526 * We can only test PG_BUSY here. Blocking on 2527 * m->busy might lead to a deadlock: 2528 * 2529 * vm_fault->getpages->cluster_read->allocbuf 2530 * 2531 */ 2532 2533 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 2534 continue; 2535 2536 /* 2537 * We have a good page. Should we wakeup the 2538 * page daemon? 2539 */ 2540 if ((curproc != pageproc) && 2541 ((m->queue - m->pc) == PQ_CACHE) && 2542 ((cnt.v_free_count + cnt.v_cache_count) < 2543 (cnt.v_free_min + cnt.v_cache_min))) { 2544 pagedaemon_wakeup(); 2545 } 2546 vm_page_flag_clear(m, PG_ZERO); 2547 vm_page_wire(m); 2548 bp->b_pages[bp->b_npages] = m; 2549 ++bp->b_npages; 2550 } 2551 2552 /* 2553 * Step 2. We've loaded the pages into the buffer, 2554 * we have to figure out if we can still have B_CACHE 2555 * set. Note that B_CACHE is set according to the 2556 * byte-granular range ( bcount and size ), new the 2557 * aligned range ( newbsize ). 2558 * 2559 * The VM test is against m->valid, which is DEV_BSIZE 2560 * aligned. Needless to say, the validity of the data 2561 * needs to also be DEV_BSIZE aligned. Note that this 2562 * fails with NFS if the server or some other client 2563 * extends the file's EOF. If our buffer is resized, 2564 * B_CACHE may remain set! XXX 2565 */ 2566 2567 toff = bp->b_bcount; 2568 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2569 2570 while ((bp->b_flags & B_CACHE) && toff < size) { 2571 vm_pindex_t pi; 2572 2573 if (tinc > (size - toff)) 2574 tinc = size - toff; 2575 2576 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2577 PAGE_SHIFT; 2578 2579 vfs_buf_test_cache( 2580 bp, 2581 bp->b_offset, 2582 toff, 2583 tinc, 2584 bp->b_pages[pi] 2585 ); 2586 toff += tinc; 2587 tinc = PAGE_SIZE; 2588 } 2589 2590 /* 2591 * Step 3, fixup the KVM pmap. Remember that 2592 * bp->b_data is relative to bp->b_offset, but 2593 * bp->b_offset may be offset into the first page. 2594 */ 2595 2596 bp->b_data = (caddr_t) 2597 trunc_page((vm_offset_t)bp->b_data); 2598 pmap_qenter( 2599 (vm_offset_t)bp->b_data, 2600 bp->b_pages, 2601 bp->b_npages 2602 ); 2603 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2604 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2605 } 2606 } 2607 if (bp->b_flags & B_VMIO) 2608 vmiospace += (newbsize - bp->b_bufsize); 2609 bufspace += (newbsize - bp->b_bufsize); 2610 runningbufspace += (newbsize - bp->b_bufsize); 2611 if (newbsize < bp->b_bufsize) 2612 bufspacewakeup(); 2613 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2614 bp->b_bcount = size; /* requested buffer size */ 2615 return 1; 2616} 2617 2618/* 2619 * biowait: 2620 * 2621 * Wait for buffer I/O completion, returning error status. The buffer 2622 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR 2623 * error and cleared. 2624 */ 2625int 2626biowait(register struct buf * bp) 2627{ 2628 int s; 2629 2630 s = splbio(); 2631 while ((bp->b_flags & B_DONE) == 0) { 2632#if defined(NO_SCHEDULE_MODS) 2633 tsleep(bp, PRIBIO, "biowait", 0); 2634#else 2635 if (bp->b_flags & B_READ) 2636 tsleep(bp, PRIBIO, "biord", 0); 2637 else 2638 tsleep(bp, PRIBIO, "biowr", 0); 2639#endif 2640 } 2641 splx(s); 2642 if (bp->b_flags & B_EINTR) { 2643 bp->b_flags &= ~B_EINTR; 2644 return (EINTR); 2645 } 2646 if (bp->b_flags & B_ERROR) { 2647 return (bp->b_error ? bp->b_error : EIO); 2648 } else { 2649 return (0); 2650 } 2651} 2652 2653/* 2654 * biodone: 2655 * 2656 * Finish I/O on a buffer, optionally calling a completion function. 2657 * This is usually called from an interrupt so process blocking is 2658 * not allowed. 2659 * 2660 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2661 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2662 * assuming B_INVAL is clear. 2663 * 2664 * For the VMIO case, we set B_CACHE if the op was a read and no 2665 * read error occured, or if the op was a write. B_CACHE is never 2666 * set if the buffer is invalid or otherwise uncacheable. 2667 * 2668 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2669 * initiator to leave B_INVAL set to brelse the buffer out of existance 2670 * in the biodone routine. 2671 */ 2672void 2673biodone(register struct buf * bp) 2674{ 2675 int s; 2676 2677 s = splbio(); 2678 2679 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 2680 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 2681 2682 bp->b_flags |= B_DONE; 2683 2684 if (bp->b_flags & B_FREEBUF) { 2685 brelse(bp); 2686 splx(s); 2687 return; 2688 } 2689 2690 if ((bp->b_flags & B_READ) == 0) { 2691 vwakeup(bp); 2692 } 2693 2694 /* call optional completion function if requested */ 2695 if (bp->b_flags & B_CALL) { 2696 bp->b_flags &= ~B_CALL; 2697 (*bp->b_iodone) (bp); 2698 splx(s); 2699 return; 2700 } 2701 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 2702 (*bioops.io_complete)(bp); 2703 2704 if (bp->b_flags & B_VMIO) { 2705 int i, resid; 2706 vm_ooffset_t foff; 2707 vm_page_t m; 2708 vm_object_t obj; 2709 int iosize; 2710 struct vnode *vp = bp->b_vp; 2711 2712 obj = vp->v_object; 2713 2714#if defined(VFS_BIO_DEBUG) 2715 if (vp->v_usecount == 0) { 2716 panic("biodone: zero vnode ref count"); 2717 } 2718 2719 if (vp->v_object == NULL) { 2720 panic("biodone: missing VM object"); 2721 } 2722 2723 if ((vp->v_flag & VOBJBUF) == 0) { 2724 panic("biodone: vnode is not setup for merged cache"); 2725 } 2726#endif 2727 2728 foff = bp->b_offset; 2729 KASSERT(bp->b_offset != NOOFFSET, 2730 ("biodone: no buffer offset")); 2731 2732 if (!obj) { 2733 panic("biodone: no object"); 2734 } 2735#if defined(VFS_BIO_DEBUG) 2736 if (obj->paging_in_progress < bp->b_npages) { 2737 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 2738 obj->paging_in_progress, bp->b_npages); 2739 } 2740#endif 2741 2742 /* 2743 * Set B_CACHE if the op was a normal read and no error 2744 * occured. B_CACHE is set for writes in the b*write() 2745 * routines. 2746 */ 2747 iosize = bp->b_bcount - bp->b_resid; 2748 if ((bp->b_flags & (B_READ|B_FREEBUF|B_INVAL|B_NOCACHE|B_ERROR)) == B_READ) { 2749 bp->b_flags |= B_CACHE; 2750 } 2751 2752 for (i = 0; i < bp->b_npages; i++) { 2753 int bogusflag = 0; 2754 m = bp->b_pages[i]; 2755 if (m == bogus_page) { 2756 bogusflag = 1; 2757 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2758 if (!m) { 2759#if defined(VFS_BIO_DEBUG) 2760 printf("biodone: page disappeared\n"); 2761#endif 2762 vm_object_pip_subtract(obj, 1); 2763 bp->b_flags &= ~B_CACHE; 2764 continue; 2765 } 2766 bp->b_pages[i] = m; 2767 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2768 } 2769#if defined(VFS_BIO_DEBUG) 2770 if (OFF_TO_IDX(foff) != m->pindex) { 2771 printf( 2772"biodone: foff(%lu)/m->pindex(%d) mismatch\n", 2773 (unsigned long)foff, m->pindex); 2774 } 2775#endif 2776 resid = IDX_TO_OFF(m->pindex + 1) - foff; 2777 if (resid > iosize) 2778 resid = iosize; 2779 2780 /* 2781 * In the write case, the valid and clean bits are 2782 * already changed correctly ( see bdwrite() ), so we 2783 * only need to do this here in the read case. 2784 */ 2785 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) { 2786 vfs_page_set_valid(bp, foff, i, m); 2787 } 2788 vm_page_flag_clear(m, PG_ZERO); 2789 2790 /* 2791 * when debugging new filesystems or buffer I/O methods, this 2792 * is the most common error that pops up. if you see this, you 2793 * have not set the page busy flag correctly!!! 2794 */ 2795 if (m->busy == 0) { 2796 printf("biodone: page busy < 0, " 2797 "pindex: %d, foff: 0x(%x,%x), " 2798 "resid: %d, index: %d\n", 2799 (int) m->pindex, (int)(foff >> 32), 2800 (int) foff & 0xffffffff, resid, i); 2801 if (!vn_isdisk(vp, NULL)) 2802 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 2803 bp->b_vp->v_mount->mnt_stat.f_iosize, 2804 (int) bp->b_lblkno, 2805 bp->b_flags, bp->b_npages); 2806 else 2807 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 2808 (int) bp->b_lblkno, 2809 bp->b_flags, bp->b_npages); 2810 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2811 m->valid, m->dirty, m->wire_count); 2812 panic("biodone: page busy < 0\n"); 2813 } 2814 vm_page_io_finish(m); 2815 vm_object_pip_subtract(obj, 1); 2816 foff += resid; 2817 iosize -= resid; 2818 } 2819 if (obj) 2820 vm_object_pip_wakeupn(obj, 0); 2821 } 2822 /* 2823 * For asynchronous completions, release the buffer now. The brelse 2824 * will do a wakeup there if necessary - so no need to do a wakeup 2825 * here in the async case. The sync case always needs to do a wakeup. 2826 */ 2827 2828 if (bp->b_flags & B_ASYNC) { 2829 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0) 2830 brelse(bp); 2831 else 2832 bqrelse(bp); 2833 } else { 2834 wakeup(bp); 2835 } 2836 splx(s); 2837} 2838 2839/* 2840 * This routine is called in lieu of iodone in the case of 2841 * incomplete I/O. This keeps the busy status for pages 2842 * consistant. 2843 */ 2844void 2845vfs_unbusy_pages(struct buf * bp) 2846{ 2847 int i; 2848 2849 if (bp->b_flags & B_VMIO) { 2850 struct vnode *vp = bp->b_vp; 2851 vm_object_t obj = vp->v_object; 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) & ~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 B_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 = vp->v_object; 2929 vm_ooffset_t foff; 2930 2931 foff = bp->b_offset; 2932 KASSERT(bp->b_offset != NOOFFSET, 2933 ("vfs_busy_pages: no buffer offset")); 2934 vfs_setdirty(bp); 2935 2936retry: 2937 for (i = 0; i < bp->b_npages; i++) { 2938 vm_page_t m = bp->b_pages[i]; 2939 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 2940 goto retry; 2941 } 2942 2943 bogus = 0; 2944 for (i = 0; i < bp->b_npages; i++) { 2945 vm_page_t m = bp->b_pages[i]; 2946 2947 vm_page_flag_clear(m, PG_ZERO); 2948 if ((bp->b_flags & B_CLUSTER) == 0) { 2949 vm_object_pip_add(obj, 1); 2950 vm_page_io_start(m); 2951 } 2952 2953 /* 2954 * When readying a buffer for a read ( i.e 2955 * clear_modify == 0 ), it is important to do 2956 * bogus_page replacement for valid pages in 2957 * partially instantiated buffers. Partially 2958 * instantiated buffers can, in turn, occur when 2959 * reconstituting a buffer from its VM backing store 2960 * base. We only have to do this if B_CACHE is 2961 * clear ( which causes the I/O to occur in the 2962 * first place ). The replacement prevents the read 2963 * I/O from overwriting potentially dirty VM-backed 2964 * pages. XXX bogus page replacement is, uh, bogus. 2965 * It may not work properly with small-block devices. 2966 * We need to find a better way. 2967 */ 2968 2969 vm_page_protect(m, VM_PROT_NONE); 2970 if (clear_modify) 2971 vfs_page_set_valid(bp, foff, i, m); 2972 else if (m->valid == VM_PAGE_BITS_ALL && 2973 (bp->b_flags & B_CACHE) == 0) { 2974 bp->b_pages[i] = bogus_page; 2975 bogus++; 2976 } 2977 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 2978 } 2979 if (bogus) 2980 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2981 } 2982} 2983 2984/* 2985 * Tell the VM system that the pages associated with this buffer 2986 * are clean. This is used for delayed writes where the data is 2987 * going to go to disk eventually without additional VM intevention. 2988 * 2989 * Note that while we only really need to clean through to b_bcount, we 2990 * just go ahead and clean through to b_bufsize. 2991 */ 2992static void 2993vfs_clean_pages(struct buf * bp) 2994{ 2995 int i; 2996 2997 if (bp->b_flags & B_VMIO) { 2998 vm_ooffset_t foff; 2999 3000 foff = bp->b_offset; 3001 KASSERT(bp->b_offset != NOOFFSET, 3002 ("vfs_clean_pages: no buffer offset")); 3003 for (i = 0; i < bp->b_npages; i++) { 3004 vm_page_t m = bp->b_pages[i]; 3005 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~PAGE_MASK; 3006 vm_ooffset_t eoff = noff; 3007 3008 if (eoff > bp->b_offset + bp->b_bufsize) 3009 eoff = bp->b_offset + bp->b_bufsize; 3010 vfs_page_set_valid(bp, foff, i, m); 3011 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3012 foff = noff; 3013 } 3014 } 3015} 3016 3017/* 3018 * vfs_bio_set_validclean: 3019 * 3020 * Set the range within the buffer to valid and clean. The range is 3021 * relative to the beginning of the buffer, b_offset. Note that b_offset 3022 * itself may be offset from the beginning of the first page. 3023 */ 3024 3025void 3026vfs_bio_set_validclean(struct buf *bp, int base, int size) 3027{ 3028 if (bp->b_flags & B_VMIO) { 3029 int i; 3030 int n; 3031 3032 /* 3033 * Fixup base to be relative to beginning of first page. 3034 * Set initial n to be the maximum number of bytes in the 3035 * first page that can be validated. 3036 */ 3037 3038 base += (bp->b_offset & PAGE_MASK); 3039 n = PAGE_SIZE - (base & PAGE_MASK); 3040 3041 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3042 vm_page_t m = bp->b_pages[i]; 3043 3044 if (n > size) 3045 n = size; 3046 3047 vm_page_set_validclean(m, base & PAGE_MASK, n); 3048 base += n; 3049 size -= n; 3050 n = PAGE_SIZE; 3051 } 3052 } 3053} 3054 3055/* 3056 * vfs_bio_clrbuf: 3057 * 3058 * clear a buffer. This routine essentially fakes an I/O, so we need 3059 * to clear B_ERROR and B_INVAL. 3060 * 3061 * Note that while we only theoretically need to clear through b_bcount, 3062 * we go ahead and clear through b_bufsize. 3063 */ 3064 3065void 3066vfs_bio_clrbuf(struct buf *bp) { 3067 int i, mask = 0; 3068 caddr_t sa, ea; 3069 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3070 bp->b_flags &= ~(B_INVAL|B_ERROR); 3071 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3072 (bp->b_offset & PAGE_MASK) == 0) { 3073 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3074 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3075 ((bp->b_pages[0]->valid & mask) != mask)) { 3076 bzero(bp->b_data, bp->b_bufsize); 3077 } 3078 bp->b_pages[0]->valid |= mask; 3079 bp->b_resid = 0; 3080 return; 3081 } 3082 ea = sa = bp->b_data; 3083 for(i=0;i<bp->b_npages;i++,sa=ea) { 3084 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3085 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3086 ea = (caddr_t)(vm_offset_t)ulmin( 3087 (u_long)(vm_offset_t)ea, 3088 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3089 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3090 if ((bp->b_pages[i]->valid & mask) == mask) 3091 continue; 3092 if ((bp->b_pages[i]->valid & mask) == 0) { 3093 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3094 bzero(sa, ea - sa); 3095 } 3096 } else { 3097 for (; sa < ea; sa += DEV_BSIZE, j++) { 3098 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3099 (bp->b_pages[i]->valid & (1<<j)) == 0) 3100 bzero(sa, DEV_BSIZE); 3101 } 3102 } 3103 bp->b_pages[i]->valid |= mask; 3104 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3105 } 3106 bp->b_resid = 0; 3107 } else { 3108 clrbuf(bp); 3109 } 3110} 3111 3112/* 3113 * vm_hold_load_pages and vm_hold_unload pages get pages into 3114 * a buffers address space. The pages are anonymous and are 3115 * not associated with a file object. 3116 */ 3117void 3118vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3119{ 3120 vm_offset_t pg; 3121 vm_page_t p; 3122 int index; 3123 3124 to = round_page(to); 3125 from = round_page(from); 3126 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3127 3128 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3129 3130tryagain: 3131 3132 p = vm_page_alloc(kernel_object, 3133 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3134 VM_ALLOC_NORMAL); 3135 if (!p) { 3136 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3137 VM_WAIT; 3138 goto tryagain; 3139 } 3140 vm_page_wire(p); 3141 p->valid = VM_PAGE_BITS_ALL; 3142 vm_page_flag_clear(p, PG_ZERO); 3143 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 3144 bp->b_pages[index] = p; 3145 vm_page_wakeup(p); 3146 } 3147 bp->b_npages = index; 3148} 3149 3150void 3151vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3152{ 3153 vm_offset_t pg; 3154 vm_page_t p; 3155 int index, newnpages; 3156 3157 from = round_page(from); 3158 to = round_page(to); 3159 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3160 3161 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3162 p = bp->b_pages[index]; 3163 if (p && (index < bp->b_npages)) { 3164 if (p->busy) { 3165 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 3166 bp->b_blkno, bp->b_lblkno); 3167 } 3168 bp->b_pages[index] = NULL; 3169 pmap_kremove(pg); 3170 vm_page_busy(p); 3171 vm_page_unwire(p, 0); 3172 vm_page_free(p); 3173 } 3174 } 3175 bp->b_npages = newnpages; 3176} 3177 3178 3179#include "opt_ddb.h" 3180#ifdef DDB 3181#include <ddb/ddb.h> 3182 3183DB_SHOW_COMMAND(buffer, db_show_buffer) 3184{ 3185 /* get args */ 3186 struct buf *bp = (struct buf *)addr; 3187 3188 if (!have_addr) { 3189 db_printf("usage: show buffer <addr>\n"); 3190 return; 3191 } 3192 3193 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3194 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 3195 "b_resid = %ld\nb_dev = (%d,%d), b_data = %p, " 3196 "b_blkno = %d, b_pblkno = %d\n", 3197 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3198 major(bp->b_dev), minor(bp->b_dev), 3199 bp->b_data, bp->b_blkno, bp->b_pblkno); 3200 if (bp->b_npages) { 3201 int i; 3202 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3203 for (i = 0; i < bp->b_npages; i++) { 3204 vm_page_t m; 3205 m = bp->b_pages[i]; 3206 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3207 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3208 if ((i + 1) < bp->b_npages) 3209 db_printf(","); 3210 } 3211 db_printf("\n"); 3212 } 3213} 3214#endif /* DDB */ 3215