vfs_bio.c revision 67365
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 67365 2000-10-20 07:58:15Z jhb $ 15 */ 16 17/* 18 * this file contains a new buffer I/O scheme implementing a coherent 19 * VM object and buffer cache scheme. Pains have been taken to make 20 * sure that the performance degradation associated with schemes such 21 * as this is not realized. 22 * 23 * Author: John S. Dyson 24 * Significant help during the development and debugging phases 25 * had been provided by David Greenman, also of the FreeBSD core team. 26 * 27 * see man buf(9) for more info. 28 */ 29 30#include <sys/param.h> 31#include <sys/systm.h> 32#include <sys/bio.h> 33#include <sys/buf.h> 34#include <sys/eventhandler.h> 35#include <sys/lock.h> 36#include <sys/malloc.h> 37#include <sys/mount.h> 38#include <sys/mutex.h> 39#include <sys/kernel.h> 40#include <sys/kthread.h> 41#include <sys/ktr.h> 42#include <sys/proc.h> 43#include <sys/reboot.h> 44#include <sys/resourcevar.h> 45#include <sys/sysctl.h> 46#include <sys/vmmeter.h> 47#include <sys/vnode.h> 48#include <vm/vm.h> 49#include <vm/vm_param.h> 50#include <vm/vm_kern.h> 51#include <vm/vm_pageout.h> 52#include <vm/vm_page.h> 53#include <vm/vm_object.h> 54#include <vm/vm_extern.h> 55#include <vm/vm_map.h> 56 57static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 58 59struct bio_ops bioops; /* I/O operation notification */ 60 61struct buf *buf; /* buffer header pool */ 62struct swqueue bswlist; 63struct mtx buftimelock; /* Interlock on setting prio and timo */ 64 65static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 66 vm_offset_t to); 67static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 68 vm_offset_t to); 69static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 70 int pageno, vm_page_t m); 71static void vfs_clean_pages(struct buf * bp); 72static void vfs_setdirty(struct buf *bp); 73static void vfs_vmio_release(struct buf *bp); 74static void vfs_backgroundwritedone(struct buf *bp); 75static int flushbufqueues(void); 76 77static int bd_request; 78 79static void buf_daemon __P((void)); 80/* 81 * bogus page -- for I/O to/from partially complete buffers 82 * this is a temporary solution to the problem, but it is not 83 * really that bad. it would be better to split the buffer 84 * for input in the case of buffers partially already in memory, 85 * but the code is intricate enough already. 86 */ 87vm_page_t bogus_page; 88int runningbufspace; 89int vmiodirenable = FALSE; 90static vm_offset_t bogus_offset; 91 92static int bufspace, maxbufspace, 93 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace; 94static int bufreusecnt, bufdefragcnt, buffreekvacnt; 95static int maxbdrun; 96static int needsbuffer; 97static int numdirtybuffers, hidirtybuffers; 98static int numfreebuffers, lofreebuffers, hifreebuffers; 99static int getnewbufcalls; 100static int getnewbufrestarts; 101 102SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, 103 &numdirtybuffers, 0, ""); 104SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, 105 &hidirtybuffers, 0, ""); 106SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, 107 &numfreebuffers, 0, ""); 108SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, 109 &lofreebuffers, 0, ""); 110SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, 111 &hifreebuffers, 0, ""); 112SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, 113 &runningbufspace, 0, ""); 114SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, 115 &maxbufspace, 0, ""); 116SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, 117 &hibufspace, 0, ""); 118SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, 119 &lobufspace, 0, ""); 120SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, 121 &bufspace, 0, ""); 122SYSCTL_INT(_vfs, OID_AUTO, maxbdrun, CTLFLAG_RW, 123 &maxbdrun, 0, ""); 124SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, 125 &maxbufmallocspace, 0, ""); 126SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, 127 &bufmallocspace, 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, ""); 134SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, 135 &bufdefragcnt, 0, ""); 136SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, 137 &buffreekvacnt, 0, ""); 138SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, 139 &bufreusecnt, 0, ""); 140 141static int bufhashmask; 142static LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash; 143struct bqueues bufqueues[BUFFER_QUEUES] = { { 0 } }; 144char *buf_wmesg = BUF_WMESG; 145 146extern int vm_swap_size; 147 148#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 149#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 150#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 151#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 152 153/* 154 * Buffer hash table code. Note that the logical block scans linearly, which 155 * gives us some L1 cache locality. 156 */ 157 158static __inline 159struct bufhashhdr * 160bufhash(struct vnode *vnp, daddr_t bn) 161{ 162 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]); 163} 164 165/* 166 * numdirtywakeup: 167 * 168 * If someone is blocked due to there being too many dirty buffers, 169 * and numdirtybuffers is now reasonable, wake them up. 170 */ 171 172static __inline void 173numdirtywakeup(void) 174{ 175 if (numdirtybuffers < hidirtybuffers) { 176 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 177 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 178 wakeup(&needsbuffer); 179 } 180 } 181} 182 183/* 184 * bufspacewakeup: 185 * 186 * Called when buffer space is potentially available for recovery. 187 * getnewbuf() will block on this flag when it is unable to free 188 * sufficient buffer space. Buffer space becomes recoverable when 189 * bp's get placed back in the queues. 190 */ 191 192static __inline void 193bufspacewakeup(void) 194{ 195 /* 196 * If someone is waiting for BUF space, wake them up. Even 197 * though we haven't freed the kva space yet, the waiting 198 * process will be able to now. 199 */ 200 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 201 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 202 wakeup(&needsbuffer); 203 } 204} 205 206/* 207 * bufcountwakeup: 208 * 209 * Called when a buffer has been added to one of the free queues to 210 * account for the buffer and to wakeup anyone waiting for free buffers. 211 * This typically occurs when large amounts of metadata are being handled 212 * by the buffer cache ( else buffer space runs out first, usually ). 213 */ 214 215static __inline void 216bufcountwakeup(void) 217{ 218 ++numfreebuffers; 219 if (needsbuffer) { 220 needsbuffer &= ~VFS_BIO_NEED_ANY; 221 if (numfreebuffers >= hifreebuffers) 222 needsbuffer &= ~VFS_BIO_NEED_FREE; 223 wakeup(&needsbuffer); 224 } 225} 226 227/* 228 * vfs_buf_test_cache: 229 * 230 * Called when a buffer is extended. This function clears the B_CACHE 231 * bit if the newly extended portion of the buffer does not contain 232 * valid data. 233 */ 234static __inline__ 235void 236vfs_buf_test_cache(struct buf *bp, 237 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 238 vm_page_t m) 239{ 240 if (bp->b_flags & B_CACHE) { 241 int base = (foff + off) & PAGE_MASK; 242 if (vm_page_is_valid(m, base, size) == 0) 243 bp->b_flags &= ~B_CACHE; 244 } 245} 246 247static __inline__ 248void 249bd_wakeup(int dirtybuflevel) 250{ 251 if (numdirtybuffers >= dirtybuflevel && bd_request == 0) { 252 bd_request = 1; 253 wakeup(&bd_request); 254 } 255} 256 257/* 258 * bd_speedup - speedup the buffer cache flushing code 259 */ 260 261static __inline__ 262void 263bd_speedup(void) 264{ 265 bd_wakeup(1); 266} 267 268/* 269 * Initialize buffer headers and related structures. 270 */ 271 272caddr_t 273bufhashinit(caddr_t vaddr) 274{ 275 /* first, make a null hash table */ 276 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 277 ; 278 bufhashtbl = (void *)vaddr; 279 vaddr = vaddr + sizeof(*bufhashtbl) * bufhashmask; 280 --bufhashmask; 281 return(vaddr); 282} 283 284void 285bufinit(void) 286{ 287 struct buf *bp; 288 int i; 289 290 TAILQ_INIT(&bswlist); 291 LIST_INIT(&invalhash); 292 mtx_init(&buftimelock, "buftime lock", MTX_DEF); 293 294 for (i = 0; i <= bufhashmask; i++) 295 LIST_INIT(&bufhashtbl[i]); 296 297 /* next, make a null set of free lists */ 298 for (i = 0; i < BUFFER_QUEUES; i++) 299 TAILQ_INIT(&bufqueues[i]); 300 301 /* finally, initialize each buffer header and stick on empty q */ 302 for (i = 0; i < nbuf; i++) { 303 bp = &buf[i]; 304 bzero(bp, sizeof *bp); 305 bp->b_flags = B_INVAL; /* we're just an empty header */ 306 bp->b_dev = NODEV; 307 bp->b_rcred = NOCRED; 308 bp->b_wcred = NOCRED; 309 bp->b_qindex = QUEUE_EMPTY; 310 bp->b_xflags = 0; 311 LIST_INIT(&bp->b_dep); 312 BUF_LOCKINIT(bp); 313 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 314 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 315 } 316 317 /* 318 * maxbufspace is the absolute maximum amount of buffer space we are 319 * allowed to reserve in KVM and in real terms. The absolute maximum 320 * is nominally used by buf_daemon. hibufspace is the nominal maximum 321 * used by most other processes. The differential is required to 322 * ensure that buf_daemon is able to run when other processes might 323 * be blocked waiting for buffer space. 324 * 325 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 326 * this may result in KVM fragmentation which is not handled optimally 327 * by the system. 328 */ 329 maxbufspace = nbuf * BKVASIZE; 330 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 331 lobufspace = hibufspace - MAXBSIZE; 332 333/* 334 * Limit the amount of malloc memory since it is wired permanently into 335 * the kernel space. Even though this is accounted for in the buffer 336 * allocation, we don't want the malloced region to grow uncontrolled. 337 * The malloc scheme improves memory utilization significantly on average 338 * (small) directories. 339 */ 340 maxbufmallocspace = hibufspace / 20; 341 342/* 343 * Reduce the chance of a deadlock occuring by limiting the number 344 * of delayed-write dirty buffers we allow to stack up. 345 */ 346 hidirtybuffers = nbuf / 4 + 20; 347 numdirtybuffers = 0; 348/* 349 * To support extreme low-memory systems, make sure hidirtybuffers cannot 350 * eat up all available buffer space. This occurs when our minimum cannot 351 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 352 * BKVASIZE'd (8K) buffers. 353 */ 354 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 355 hidirtybuffers >>= 1; 356 } 357 358/* 359 * Try to keep the number of free buffers in the specified range, 360 * and give special processes (e.g. like buf_daemon) access to an 361 * emergency reserve. 362 */ 363 lofreebuffers = nbuf / 18 + 5; 364 hifreebuffers = 2 * lofreebuffers; 365 numfreebuffers = nbuf; 366 367/* 368 * Maximum number of async ops initiated per buf_daemon loop. This is 369 * somewhat of a hack at the moment, we really need to limit ourselves 370 * based on the number of bytes of I/O in-transit that were initiated 371 * from buf_daemon. 372 */ 373 if ((maxbdrun = nswbuf / 4) < 4) 374 maxbdrun = 4; 375 376 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 377 bogus_page = vm_page_alloc(kernel_object, 378 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 379 VM_ALLOC_NORMAL); 380 cnt.v_wire_count++; 381 382} 383 384/* 385 * bfreekva() - free the kva allocation for a buffer. 386 * 387 * Must be called at splbio() or higher as this is the only locking for 388 * buffer_map. 389 * 390 * Since this call frees up buffer space, we call bufspacewakeup(). 391 */ 392static void 393bfreekva(struct buf * bp) 394{ 395 if (bp->b_kvasize) { 396 ++buffreekvacnt; 397 bufspace -= bp->b_kvasize; 398 vm_map_delete(buffer_map, 399 (vm_offset_t) bp->b_kvabase, 400 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 401 ); 402 bp->b_kvasize = 0; 403 bufspacewakeup(); 404 } 405} 406 407/* 408 * bremfree: 409 * 410 * Remove the buffer from the appropriate free list. 411 */ 412void 413bremfree(struct buf * bp) 414{ 415 int s = splbio(); 416 int old_qindex = bp->b_qindex; 417 418 if (bp->b_qindex != QUEUE_NONE) { 419 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 420 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 421 bp->b_qindex = QUEUE_NONE; 422 runningbufspace += bp->b_bufsize; 423 } else { 424 if (BUF_REFCNT(bp) <= 1) 425 panic("bremfree: removing a buffer not on a queue"); 426 } 427 428 /* 429 * Fixup numfreebuffers count. If the buffer is invalid or not 430 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 431 * the buffer was free and we must decrement numfreebuffers. 432 */ 433 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 434 switch(old_qindex) { 435 case QUEUE_DIRTY: 436 case QUEUE_CLEAN: 437 case QUEUE_EMPTY: 438 case QUEUE_EMPTYKVA: 439 --numfreebuffers; 440 break; 441 default: 442 break; 443 } 444 } 445 splx(s); 446} 447 448 449/* 450 * Get a buffer with the specified data. Look in the cache first. We 451 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 452 * is set, the buffer is valid and we do not have to do anything ( see 453 * getblk() ). 454 */ 455int 456bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 457 struct buf ** bpp) 458{ 459 struct buf *bp; 460 461 bp = getblk(vp, blkno, size, 0, 0); 462 *bpp = bp; 463 464 /* if not found in cache, do some I/O */ 465 if ((bp->b_flags & B_CACHE) == 0) { 466 if (curproc != idleproc) 467 curproc->p_stats->p_ru.ru_inblock++; 468 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp)); 469 bp->b_iocmd = BIO_READ; 470 bp->b_flags &= ~B_INVAL; 471 bp->b_ioflags &= ~BIO_ERROR; 472 if (bp->b_rcred == NOCRED) { 473 if (cred != NOCRED) 474 crhold(cred); 475 bp->b_rcred = cred; 476 } 477 vfs_busy_pages(bp, 0); 478 VOP_STRATEGY(vp, bp); 479 return (bufwait(bp)); 480 } 481 return (0); 482} 483 484/* 485 * Operates like bread, but also starts asynchronous I/O on 486 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 487 * to initiating I/O . If B_CACHE is set, the buffer is valid 488 * and we do not have to do anything. 489 */ 490int 491breadn(struct vnode * vp, daddr_t blkno, int size, 492 daddr_t * rablkno, int *rabsize, 493 int cnt, struct ucred * cred, struct buf ** bpp) 494{ 495 struct buf *bp, *rabp; 496 int i; 497 int rv = 0, readwait = 0; 498 499 *bpp = bp = getblk(vp, blkno, size, 0, 0); 500 501 /* if not found in cache, do some I/O */ 502 if ((bp->b_flags & B_CACHE) == 0) { 503 if (curproc != idleproc) 504 curproc->p_stats->p_ru.ru_inblock++; 505 bp->b_iocmd = BIO_READ; 506 bp->b_flags &= ~B_INVAL; 507 bp->b_ioflags &= ~BIO_ERROR; 508 if (bp->b_rcred == NOCRED) { 509 if (cred != NOCRED) 510 crhold(cred); 511 bp->b_rcred = cred; 512 } 513 vfs_busy_pages(bp, 0); 514 VOP_STRATEGY(vp, bp); 515 ++readwait; 516 } 517 518 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 519 if (inmem(vp, *rablkno)) 520 continue; 521 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 522 523 if ((rabp->b_flags & B_CACHE) == 0) { 524 if (curproc != idleproc) 525 curproc->p_stats->p_ru.ru_inblock++; 526 rabp->b_flags |= B_ASYNC; 527 rabp->b_flags &= ~B_INVAL; 528 rabp->b_ioflags &= ~BIO_ERROR; 529 rabp->b_iocmd = BIO_READ; 530 if (rabp->b_rcred == NOCRED) { 531 if (cred != NOCRED) 532 crhold(cred); 533 rabp->b_rcred = cred; 534 } 535 vfs_busy_pages(rabp, 0); 536 BUF_KERNPROC(rabp); 537 VOP_STRATEGY(vp, rabp); 538 } else { 539 brelse(rabp); 540 } 541 } 542 543 if (readwait) { 544 rv = bufwait(bp); 545 } 546 return (rv); 547} 548 549/* 550 * Write, release buffer on completion. (Done by iodone 551 * if async). Do not bother writing anything if the buffer 552 * is invalid. 553 * 554 * Note that we set B_CACHE here, indicating that buffer is 555 * fully valid and thus cacheable. This is true even of NFS 556 * now so we set it generally. This could be set either here 557 * or in biodone() since the I/O is synchronous. We put it 558 * here. 559 */ 560int 561bwrite(struct buf * bp) 562{ 563 int oldflags, s; 564 struct buf *newbp; 565 566 if (bp->b_flags & B_INVAL) { 567 brelse(bp); 568 return (0); 569 } 570 571 oldflags = bp->b_flags; 572 573 if (BUF_REFCNT(bp) == 0) 574 panic("bwrite: buffer is not busy???"); 575 s = splbio(); 576 /* 577 * If a background write is already in progress, delay 578 * writing this block if it is asynchronous. Otherwise 579 * wait for the background write to complete. 580 */ 581 if (bp->b_xflags & BX_BKGRDINPROG) { 582 if (bp->b_flags & B_ASYNC) { 583 splx(s); 584 bdwrite(bp); 585 return (0); 586 } 587 bp->b_xflags |= BX_BKGRDWAIT; 588 tsleep(&bp->b_xflags, PRIBIO, "biord", 0); 589 if (bp->b_xflags & BX_BKGRDINPROG) 590 panic("bwrite: still writing"); 591 } 592 593 /* Mark the buffer clean */ 594 bundirty(bp); 595 596 /* 597 * If this buffer is marked for background writing and we 598 * do not have to wait for it, make a copy and write the 599 * copy so as to leave this buffer ready for further use. 600 */ 601 if ((bp->b_xflags & BX_BKGRDWRITE) && (bp->b_flags & B_ASYNC)) { 602 if (bp->b_iodone != NULL) { 603 printf("bp->b_iodone = %p\n", bp->b_iodone); 604 panic("bwrite: need chained iodone"); 605 } 606 607 /* get a new block */ 608 newbp = geteblk(bp->b_bufsize); 609 610 /* set it to be identical to the old block */ 611 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 612 bgetvp(bp->b_vp, newbp); 613 newbp->b_lblkno = bp->b_lblkno; 614 newbp->b_blkno = bp->b_blkno; 615 newbp->b_offset = bp->b_offset; 616 newbp->b_iodone = vfs_backgroundwritedone; 617 newbp->b_flags |= B_ASYNC; 618 newbp->b_flags &= ~B_INVAL; 619 620 /* move over the dependencies */ 621 if (LIST_FIRST(&bp->b_dep) != NULL) 622 buf_movedeps(bp, newbp); 623 624 /* 625 * Initiate write on the copy, release the original to 626 * the B_LOCKED queue so that it cannot go away until 627 * the background write completes. If not locked it could go 628 * away and then be reconstituted while it was being written. 629 * If the reconstituted buffer were written, we could end up 630 * with two background copies being written at the same time. 631 */ 632 bp->b_xflags |= BX_BKGRDINPROG; 633 bp->b_flags |= B_LOCKED; 634 bqrelse(bp); 635 bp = newbp; 636 } 637 638 bp->b_flags &= ~B_DONE; 639 bp->b_ioflags &= ~BIO_ERROR; 640 bp->b_flags |= B_WRITEINPROG | B_CACHE; 641 bp->b_iocmd = BIO_WRITE; 642 643 bp->b_vp->v_numoutput++; 644 vfs_busy_pages(bp, 1); 645 if (curproc != idleproc) 646 curproc->p_stats->p_ru.ru_oublock++; 647 splx(s); 648 if (oldflags & B_ASYNC) 649 BUF_KERNPROC(bp); 650 BUF_STRATEGY(bp); 651 652 if ((oldflags & B_ASYNC) == 0) { 653 int rtval = bufwait(bp); 654 brelse(bp); 655 return (rtval); 656 } 657 658 return (0); 659} 660 661/* 662 * Complete a background write started from bwrite. 663 */ 664static void 665vfs_backgroundwritedone(bp) 666 struct buf *bp; 667{ 668 struct buf *origbp; 669 670 /* 671 * Find the original buffer that we are writing. 672 */ 673 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 674 panic("backgroundwritedone: lost buffer"); 675 /* 676 * Process dependencies then return any unfinished ones. 677 */ 678 if (LIST_FIRST(&bp->b_dep) != NULL) 679 buf_complete(bp); 680 if (LIST_FIRST(&bp->b_dep) != NULL) 681 buf_movedeps(bp, origbp); 682 /* 683 * Clear the BX_BKGRDINPROG flag in the original buffer 684 * and awaken it if it is waiting for the write to complete. 685 */ 686 origbp->b_xflags &= ~BX_BKGRDINPROG; 687 if (origbp->b_xflags & BX_BKGRDWAIT) { 688 origbp->b_xflags &= ~BX_BKGRDWAIT; 689 wakeup(&origbp->b_xflags); 690 } 691 /* 692 * Clear the B_LOCKED flag and remove it from the locked 693 * queue if it currently resides there. 694 */ 695 origbp->b_flags &= ~B_LOCKED; 696 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 697 bremfree(origbp); 698 bqrelse(origbp); 699 } 700 /* 701 * This buffer is marked B_NOCACHE, so when it is released 702 * by biodone, it will be tossed. We mark it with BIO_READ 703 * to avoid biodone doing a second vwakeup. 704 */ 705 bp->b_flags |= B_NOCACHE; 706 bp->b_iocmd = BIO_READ; 707 bp->b_flags &= ~(B_CACHE | B_DONE); 708 bp->b_iodone = 0; 709 bufdone(bp); 710} 711 712/* 713 * Delayed write. (Buffer is marked dirty). Do not bother writing 714 * anything if the buffer is marked invalid. 715 * 716 * Note that since the buffer must be completely valid, we can safely 717 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 718 * biodone() in order to prevent getblk from writing the buffer 719 * out synchronously. 720 */ 721void 722bdwrite(struct buf * bp) 723{ 724 if (BUF_REFCNT(bp) == 0) 725 panic("bdwrite: buffer is not busy"); 726 727 if (bp->b_flags & B_INVAL) { 728 brelse(bp); 729 return; 730 } 731 bdirty(bp); 732 733 /* 734 * Set B_CACHE, indicating that the buffer is fully valid. This is 735 * true even of NFS now. 736 */ 737 bp->b_flags |= B_CACHE; 738 739 /* 740 * This bmap keeps the system from needing to do the bmap later, 741 * perhaps when the system is attempting to do a sync. Since it 742 * is likely that the indirect block -- or whatever other datastructure 743 * that the filesystem needs is still in memory now, it is a good 744 * thing to do this. Note also, that if the pageout daemon is 745 * requesting a sync -- there might not be enough memory to do 746 * the bmap then... So, this is important to do. 747 */ 748 if (bp->b_lblkno == bp->b_blkno) { 749 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 750 } 751 752 /* 753 * Set the *dirty* buffer range based upon the VM system dirty pages. 754 */ 755 vfs_setdirty(bp); 756 757 /* 758 * We need to do this here to satisfy the vnode_pager and the 759 * pageout daemon, so that it thinks that the pages have been 760 * "cleaned". Note that since the pages are in a delayed write 761 * buffer -- the VFS layer "will" see that the pages get written 762 * out on the next sync, or perhaps the cluster will be completed. 763 */ 764 vfs_clean_pages(bp); 765 bqrelse(bp); 766 767 /* 768 * Wakeup the buffer flushing daemon if we have saturated the 769 * buffer cache. 770 */ 771 772 bd_wakeup(hidirtybuffers); 773 774 /* 775 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 776 * due to the softdep code. 777 */ 778} 779 780/* 781 * bdirty: 782 * 783 * Turn buffer into delayed write request. We must clear BIO_READ and 784 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 785 * itself to properly update it in the dirty/clean lists. We mark it 786 * B_DONE to ensure that any asynchronization of the buffer properly 787 * clears B_DONE ( else a panic will occur later ). 788 * 789 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 790 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 791 * should only be called if the buffer is known-good. 792 * 793 * Since the buffer is not on a queue, we do not update the numfreebuffers 794 * count. 795 * 796 * Must be called at splbio(). 797 * The buffer must be on QUEUE_NONE. 798 */ 799void 800bdirty(bp) 801 struct buf *bp; 802{ 803 KASSERT(bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 804 bp->b_flags &= ~(B_RELBUF); 805 bp->b_iocmd = BIO_WRITE; 806 807 if ((bp->b_flags & B_DELWRI) == 0) { 808 bp->b_flags |= B_DONE | B_DELWRI; 809 reassignbuf(bp, bp->b_vp); 810 ++numdirtybuffers; 811 bd_wakeup(hidirtybuffers); 812 } 813} 814 815/* 816 * bundirty: 817 * 818 * Clear B_DELWRI for buffer. 819 * 820 * Since the buffer is not on a queue, we do not update the numfreebuffers 821 * count. 822 * 823 * Must be called at splbio(). 824 * The buffer must be on QUEUE_NONE. 825 */ 826 827void 828bundirty(bp) 829 struct buf *bp; 830{ 831 KASSERT(bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 832 833 if (bp->b_flags & B_DELWRI) { 834 bp->b_flags &= ~B_DELWRI; 835 reassignbuf(bp, bp->b_vp); 836 --numdirtybuffers; 837 numdirtywakeup(); 838 } 839 /* 840 * Since it is now being written, we can clear its deferred write flag. 841 */ 842 bp->b_flags &= ~B_DEFERRED; 843} 844 845/* 846 * bawrite: 847 * 848 * Asynchronous write. Start output on a buffer, but do not wait for 849 * it to complete. The buffer is released when the output completes. 850 * 851 * bwrite() ( or the VOP routine anyway ) is responsible for handling 852 * B_INVAL buffers. Not us. 853 */ 854void 855bawrite(struct buf * bp) 856{ 857 bp->b_flags |= B_ASYNC; 858 (void) BUF_WRITE(bp); 859} 860 861/* 862 * bowrite: 863 * 864 * Ordered write. Start output on a buffer, and flag it so that the 865 * device will write it in the order it was queued. The buffer is 866 * released when the output completes. bwrite() ( or the VOP routine 867 * anyway ) is responsible for handling B_INVAL buffers. 868 */ 869int 870bowrite(struct buf * bp) 871{ 872 bp->b_ioflags |= BIO_ORDERED; 873 bp->b_flags |= B_ASYNC; 874 return (BUF_WRITE(bp)); 875} 876 877/* 878 * bwillwrite: 879 * 880 * Called prior to the locking of any vnodes when we are expecting to 881 * write. We do not want to starve the buffer cache with too many 882 * dirty buffers so we block here. By blocking prior to the locking 883 * of any vnodes we attempt to avoid the situation where a locked vnode 884 * prevents the various system daemons from flushing related buffers. 885 */ 886 887void 888bwillwrite(void) 889{ 890 int slop = hidirtybuffers / 10; 891 892 if (numdirtybuffers > hidirtybuffers + slop) { 893 int s; 894 895 s = splbio(); 896 while (numdirtybuffers > hidirtybuffers) { 897 bd_wakeup(hidirtybuffers); 898 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 899 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0); 900 } 901 splx(s); 902 } 903} 904 905/* 906 * brelse: 907 * 908 * Release a busy buffer and, if requested, free its resources. The 909 * buffer will be stashed in the appropriate bufqueue[] allowing it 910 * to be accessed later as a cache entity or reused for other purposes. 911 */ 912void 913brelse(struct buf * bp) 914{ 915 int s; 916 917 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 918 919 s = splbio(); 920 921 if (bp->b_flags & B_LOCKED) 922 bp->b_ioflags &= ~BIO_ERROR; 923 924 if (bp->b_iocmd == BIO_WRITE && 925 (bp->b_ioflags & BIO_ERROR) && 926 !(bp->b_flags & B_INVAL)) { 927 /* 928 * Failed write, redirty. Must clear BIO_ERROR to prevent 929 * pages from being scrapped. If B_INVAL is set then 930 * this case is not run and the next case is run to 931 * destroy the buffer. B_INVAL can occur if the buffer 932 * is outside the range supported by the underlying device. 933 */ 934 bp->b_ioflags &= ~BIO_ERROR; 935 bdirty(bp); 936 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 937 (bp->b_ioflags & BIO_ERROR) || 938 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 939 /* 940 * Either a failed I/O or we were asked to free or not 941 * cache the buffer. 942 */ 943 bp->b_flags |= B_INVAL; 944 if (LIST_FIRST(&bp->b_dep) != NULL) 945 buf_deallocate(bp); 946 if (bp->b_flags & B_DELWRI) { 947 --numdirtybuffers; 948 numdirtywakeup(); 949 } 950 bp->b_flags &= ~(B_DELWRI | B_CACHE); 951 if ((bp->b_flags & B_VMIO) == 0) { 952 if (bp->b_bufsize) 953 allocbuf(bp, 0); 954 if (bp->b_vp) 955 brelvp(bp); 956 } 957 } 958 959 /* 960 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 961 * is called with B_DELWRI set, the underlying pages may wind up 962 * getting freed causing a previous write (bdwrite()) to get 'lost' 963 * because pages associated with a B_DELWRI bp are marked clean. 964 * 965 * We still allow the B_INVAL case to call vfs_vmio_release(), even 966 * if B_DELWRI is set. 967 */ 968 969 if (bp->b_flags & B_DELWRI) 970 bp->b_flags &= ~B_RELBUF; 971 972 /* 973 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 974 * constituted, not even NFS buffers now. Two flags effect this. If 975 * B_INVAL, the struct buf is invalidated but the VM object is kept 976 * around ( i.e. so it is trivial to reconstitute the buffer later ). 977 * 978 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 979 * invalidated. BIO_ERROR cannot be set for a failed write unless the 980 * buffer is also B_INVAL because it hits the re-dirtying code above. 981 * 982 * Normally we can do this whether a buffer is B_DELWRI or not. If 983 * the buffer is an NFS buffer, it is tracking piecemeal writes or 984 * the commit state and we cannot afford to lose the buffer. If the 985 * buffer has a background write in progress, we need to keep it 986 * around to prevent it from being reconstituted and starting a second 987 * background write. 988 */ 989 if ((bp->b_flags & B_VMIO) 990 && !(bp->b_vp->v_tag == VT_NFS && 991 !vn_isdisk(bp->b_vp, NULL) && 992 (bp->b_flags & B_DELWRI) && 993 (bp->b_xflags & BX_BKGRDINPROG)) 994 ) { 995 996 int i, j, resid; 997 vm_page_t m; 998 off_t foff; 999 vm_pindex_t poff; 1000 vm_object_t obj; 1001 struct vnode *vp; 1002 1003 vp = bp->b_vp; 1004 1005 /* 1006 * Get the base offset and length of the buffer. Note that 1007 * for block sizes that are less then PAGE_SIZE, the b_data 1008 * base of the buffer does not represent exactly b_offset and 1009 * neither b_offset nor b_size are necessarily page aligned. 1010 * Instead, the starting position of b_offset is: 1011 * 1012 * b_data + (b_offset & PAGE_MASK) 1013 * 1014 * block sizes less then DEV_BSIZE (usually 512) are not 1015 * supported due to the page granularity bits (m->valid, 1016 * m->dirty, etc...). 1017 * 1018 * See man buf(9) for more information 1019 */ 1020 1021 resid = bp->b_bufsize; 1022 foff = bp->b_offset; 1023 1024 for (i = 0; i < bp->b_npages; i++) { 1025 m = bp->b_pages[i]; 1026 vm_page_flag_clear(m, PG_ZERO); 1027 if (m == bogus_page) { 1028 1029 VOP_GETVOBJECT(vp, &obj); 1030 poff = OFF_TO_IDX(bp->b_offset); 1031 1032 for (j = i; j < bp->b_npages; j++) { 1033 m = bp->b_pages[j]; 1034 if (m == bogus_page) { 1035 m = vm_page_lookup(obj, poff + j); 1036 if (!m) { 1037 panic("brelse: page missing\n"); 1038 } 1039 bp->b_pages[j] = m; 1040 } 1041 } 1042 1043 if ((bp->b_flags & B_INVAL) == 0) { 1044 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1045 } 1046 } 1047 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1048 int poffset = foff & PAGE_MASK; 1049 int presid = resid > (PAGE_SIZE - poffset) ? 1050 (PAGE_SIZE - poffset) : resid; 1051 1052 KASSERT(presid >= 0, ("brelse: extra page")); 1053 vm_page_set_invalid(m, poffset, presid); 1054 } 1055 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1056 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 1057 } 1058 1059 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1060 vfs_vmio_release(bp); 1061 1062 } else if (bp->b_flags & B_VMIO) { 1063 1064 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1065 vfs_vmio_release(bp); 1066 1067 } 1068 1069 if (bp->b_qindex != QUEUE_NONE) 1070 panic("brelse: free buffer onto another queue???"); 1071 if (BUF_REFCNT(bp) > 1) { 1072 /* do not release to free list */ 1073 BUF_UNLOCK(bp); 1074 splx(s); 1075 return; 1076 } 1077 1078 /* enqueue */ 1079 1080 /* buffers with no memory */ 1081 if (bp->b_bufsize == 0) { 1082 bp->b_flags |= B_INVAL; 1083 bp->b_xflags &= ~BX_BKGRDWRITE; 1084 if (bp->b_xflags & BX_BKGRDINPROG) 1085 panic("losing buffer 1"); 1086 if (bp->b_kvasize) { 1087 bp->b_qindex = QUEUE_EMPTYKVA; 1088 } else { 1089 bp->b_qindex = QUEUE_EMPTY; 1090 } 1091 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1092 LIST_REMOVE(bp, b_hash); 1093 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1094 bp->b_dev = NODEV; 1095 /* buffers with junk contents */ 1096 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || (bp->b_ioflags & BIO_ERROR)) { 1097 bp->b_flags |= B_INVAL; 1098 bp->b_xflags &= ~BX_BKGRDWRITE; 1099 if (bp->b_xflags & BX_BKGRDINPROG) 1100 panic("losing buffer 2"); 1101 bp->b_qindex = QUEUE_CLEAN; 1102 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1103 LIST_REMOVE(bp, b_hash); 1104 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1105 bp->b_dev = NODEV; 1106 1107 /* buffers that are locked */ 1108 } else if (bp->b_flags & B_LOCKED) { 1109 bp->b_qindex = QUEUE_LOCKED; 1110 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1111 1112 /* remaining buffers */ 1113 } else { 1114 switch(bp->b_flags & (B_DELWRI|B_AGE)) { 1115 case B_DELWRI | B_AGE: 1116 bp->b_qindex = QUEUE_DIRTY; 1117 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1118 break; 1119 case B_DELWRI: 1120 bp->b_qindex = QUEUE_DIRTY; 1121 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1122 break; 1123 case B_AGE: 1124 bp->b_qindex = QUEUE_CLEAN; 1125 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1126 break; 1127 default: 1128 bp->b_qindex = QUEUE_CLEAN; 1129 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1130 break; 1131 } 1132 } 1133 1134 /* 1135 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1136 * on the correct queue. 1137 */ 1138 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) { 1139 bp->b_flags &= ~B_DELWRI; 1140 --numdirtybuffers; 1141 numdirtywakeup(); 1142 } 1143 1144 runningbufspace -= bp->b_bufsize; 1145 1146 /* 1147 * Fixup numfreebuffers count. The bp is on an appropriate queue 1148 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1149 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1150 * if B_INVAL is set ). 1151 */ 1152 1153 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1154 bufcountwakeup(); 1155 1156 /* 1157 * Something we can maybe free. 1158 */ 1159 1160 if (bp->b_bufsize || bp->b_kvasize) 1161 bufspacewakeup(); 1162 1163 /* unlock */ 1164 BUF_UNLOCK(bp); 1165 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1166 bp->b_ioflags &= ~BIO_ORDERED; 1167 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1168 panic("brelse: not dirty"); 1169 splx(s); 1170} 1171 1172/* 1173 * Release a buffer back to the appropriate queue but do not try to free 1174 * it. 1175 * 1176 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1177 * biodone() to requeue an async I/O on completion. It is also used when 1178 * known good buffers need to be requeued but we think we may need the data 1179 * again soon. 1180 */ 1181void 1182bqrelse(struct buf * bp) 1183{ 1184 int s; 1185 1186 s = splbio(); 1187 1188 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1189 1190 if (bp->b_qindex != QUEUE_NONE) 1191 panic("bqrelse: free buffer onto another queue???"); 1192 if (BUF_REFCNT(bp) > 1) { 1193 /* do not release to free list */ 1194 BUF_UNLOCK(bp); 1195 splx(s); 1196 return; 1197 } 1198 if (bp->b_flags & B_LOCKED) { 1199 bp->b_ioflags &= ~BIO_ERROR; 1200 bp->b_qindex = QUEUE_LOCKED; 1201 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1202 /* buffers with stale but valid contents */ 1203 } else if (bp->b_flags & B_DELWRI) { 1204 bp->b_qindex = QUEUE_DIRTY; 1205 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1206 } else { 1207 bp->b_qindex = QUEUE_CLEAN; 1208 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1209 } 1210 1211 runningbufspace -= bp->b_bufsize; 1212 1213 if ((bp->b_flags & B_LOCKED) == 0 && 1214 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1215 bufcountwakeup(); 1216 } 1217 1218 /* 1219 * Something we can maybe wakeup 1220 */ 1221 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1222 bufspacewakeup(); 1223 1224 /* unlock */ 1225 BUF_UNLOCK(bp); 1226 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1227 bp->b_ioflags &= ~BIO_ORDERED; 1228 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1229 panic("bqrelse: not dirty"); 1230 splx(s); 1231} 1232 1233static void 1234vfs_vmio_release(bp) 1235 struct buf *bp; 1236{ 1237 int i, s; 1238 vm_page_t m; 1239 1240 s = splvm(); 1241 for (i = 0; i < bp->b_npages; i++) { 1242 m = bp->b_pages[i]; 1243 bp->b_pages[i] = NULL; 1244 /* 1245 * In order to keep page LRU ordering consistent, put 1246 * everything on the inactive queue. 1247 */ 1248 vm_page_unwire(m, 0); 1249 /* 1250 * We don't mess with busy pages, it is 1251 * the responsibility of the process that 1252 * busied the pages to deal with them. 1253 */ 1254 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1255 continue; 1256 1257 if (m->wire_count == 0) { 1258 vm_page_flag_clear(m, PG_ZERO); 1259 /* 1260 * Might as well free the page if we can and it has 1261 * no valid data. 1262 */ 1263 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { 1264 vm_page_busy(m); 1265 vm_page_protect(m, VM_PROT_NONE); 1266 vm_page_free(m); 1267 } 1268 } 1269 } 1270 runningbufspace -= bp->b_bufsize; 1271 splx(s); 1272 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1273 if (bp->b_bufsize) 1274 bufspacewakeup(); 1275 bp->b_npages = 0; 1276 bp->b_bufsize = 0; 1277 bp->b_flags &= ~B_VMIO; 1278 if (bp->b_vp) 1279 brelvp(bp); 1280} 1281 1282/* 1283 * Check to see if a block is currently memory resident. 1284 */ 1285struct buf * 1286gbincore(struct vnode * vp, daddr_t blkno) 1287{ 1288 struct buf *bp; 1289 struct bufhashhdr *bh; 1290 1291 bh = bufhash(vp, blkno); 1292 1293 /* Search hash chain */ 1294 LIST_FOREACH(bp, bh, b_hash) { 1295 /* hit */ 1296 if (bp->b_vp == vp && bp->b_lblkno == blkno && 1297 (bp->b_flags & B_INVAL) == 0) { 1298 break; 1299 } 1300 } 1301 return (bp); 1302} 1303 1304/* 1305 * vfs_bio_awrite: 1306 * 1307 * Implement clustered async writes for clearing out B_DELWRI buffers. 1308 * This is much better then the old way of writing only one buffer at 1309 * a time. Note that we may not be presented with the buffers in the 1310 * correct order, so we search for the cluster in both directions. 1311 */ 1312int 1313vfs_bio_awrite(struct buf * bp) 1314{ 1315 int i; 1316 int j; 1317 daddr_t lblkno = bp->b_lblkno; 1318 struct vnode *vp = bp->b_vp; 1319 int s; 1320 int ncl; 1321 struct buf *bpa; 1322 int nwritten; 1323 int size; 1324 int maxcl; 1325 1326 s = splbio(); 1327 /* 1328 * right now we support clustered writing only to regular files. If 1329 * we find a clusterable block we could be in the middle of a cluster 1330 * rather then at the beginning. 1331 */ 1332 if ((vp->v_type == VREG) && 1333 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1334 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1335 1336 size = vp->v_mount->mnt_stat.f_iosize; 1337 maxcl = MAXPHYS / size; 1338 1339 for (i = 1; i < maxcl; i++) { 1340 if ((bpa = gbincore(vp, lblkno + i)) && 1341 BUF_REFCNT(bpa) == 0 && 1342 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1343 (B_DELWRI | B_CLUSTEROK)) && 1344 (bpa->b_bufsize == size)) { 1345 if ((bpa->b_blkno == bpa->b_lblkno) || 1346 (bpa->b_blkno != 1347 bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 1348 break; 1349 } else { 1350 break; 1351 } 1352 } 1353 for (j = 1; i + j <= maxcl && j <= lblkno; j++) { 1354 if ((bpa = gbincore(vp, lblkno - j)) && 1355 BUF_REFCNT(bpa) == 0 && 1356 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1357 (B_DELWRI | B_CLUSTEROK)) && 1358 (bpa->b_bufsize == size)) { 1359 if ((bpa->b_blkno == bpa->b_lblkno) || 1360 (bpa->b_blkno != 1361 bp->b_blkno - ((j * size) >> DEV_BSHIFT))) 1362 break; 1363 } else { 1364 break; 1365 } 1366 } 1367 --j; 1368 ncl = i + j; 1369 /* 1370 * this is a possible cluster write 1371 */ 1372 if (ncl != 1) { 1373 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1374 splx(s); 1375 return nwritten; 1376 } 1377 } 1378 1379 BUF_LOCK(bp, LK_EXCLUSIVE); 1380 bremfree(bp); 1381 bp->b_flags |= B_ASYNC; 1382 1383 splx(s); 1384 /* 1385 * default (old) behavior, writing out only one block 1386 * 1387 * XXX returns b_bufsize instead of b_bcount for nwritten? 1388 */ 1389 nwritten = bp->b_bufsize; 1390 (void) BUF_WRITE(bp); 1391 1392 return nwritten; 1393} 1394 1395/* 1396 * getnewbuf: 1397 * 1398 * Find and initialize a new buffer header, freeing up existing buffers 1399 * in the bufqueues as necessary. The new buffer is returned locked. 1400 * 1401 * Important: B_INVAL is not set. If the caller wishes to throw the 1402 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1403 * 1404 * We block if: 1405 * We have insufficient buffer headers 1406 * We have insufficient buffer space 1407 * buffer_map is too fragmented ( space reservation fails ) 1408 * If we have to flush dirty buffers ( but we try to avoid this ) 1409 * 1410 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1411 * Instead we ask the buf daemon to do it for us. We attempt to 1412 * avoid piecemeal wakeups of the pageout daemon. 1413 */ 1414 1415static struct buf * 1416getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1417{ 1418 struct buf *bp; 1419 struct buf *nbp; 1420 int defrag = 0; 1421 int nqindex; 1422 int isspecial; 1423 static int flushingbufs; 1424 1425 if (curproc != idleproc && 1426 (curproc->p_flag & (P_COWINPROGRESS|P_BUFEXHAUST)) == 0) 1427 isspecial = 0; 1428 else 1429 isspecial = 1; 1430 1431 ++getnewbufcalls; 1432 --getnewbufrestarts; 1433restart: 1434 ++getnewbufrestarts; 1435 1436 /* 1437 * Setup for scan. If we do not have enough free buffers, 1438 * we setup a degenerate case that immediately fails. Note 1439 * that if we are specially marked process, we are allowed to 1440 * dip into our reserves. 1441 * 1442 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1443 * 1444 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1445 * However, there are a number of cases (defragging, reusing, ...) 1446 * where we cannot backup. 1447 */ 1448 1449 if (isspecial == 0 && numfreebuffers < lofreebuffers) { 1450 /* 1451 * This will cause an immediate failure 1452 */ 1453 nqindex = QUEUE_CLEAN; 1454 nbp = NULL; 1455 } else { 1456 /* 1457 * Locate a buffer which already has KVA assigned. First 1458 * try EMPTYKVA buffers. 1459 */ 1460 nqindex = QUEUE_EMPTYKVA; 1461 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1462 1463 if (nbp == NULL) { 1464 /* 1465 * If no EMPTYKVA buffers and we are either 1466 * defragging or reusing, locate a CLEAN buffer 1467 * to free or reuse. If bufspace useage is low 1468 * skip this step so we can allocate a new buffer. 1469 */ 1470 if (defrag || bufspace >= lobufspace) { 1471 nqindex = QUEUE_CLEAN; 1472 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1473 } 1474 1475 /* 1476 * Nada. If we are allowed to allocate an EMPTY 1477 * buffer, go get one. 1478 */ 1479 if (nbp == NULL && defrag == 0 && 1480 (isspecial || bufspace < hibufspace)) { 1481 nqindex = QUEUE_EMPTY; 1482 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1483 } 1484 } 1485 } 1486 1487 /* 1488 * Run scan, possibly freeing data and/or kva mappings on the fly 1489 * depending. 1490 */ 1491 1492 while ((bp = nbp) != NULL) { 1493 int qindex = nqindex; 1494 1495 /* 1496 * Calculate next bp ( we can only use it if we do not block 1497 * or do other fancy things ). 1498 */ 1499 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1500 switch(qindex) { 1501 case QUEUE_EMPTY: 1502 nqindex = QUEUE_EMPTYKVA; 1503 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1504 break; 1505 /* fall through */ 1506 case QUEUE_EMPTYKVA: 1507 nqindex = QUEUE_CLEAN; 1508 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1509 break; 1510 /* fall through */ 1511 case QUEUE_CLEAN: 1512 /* 1513 * nbp is NULL. 1514 */ 1515 break; 1516 } 1517 } 1518 1519 /* 1520 * Sanity Checks 1521 */ 1522 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1523 1524 /* 1525 * Note: we no longer distinguish between VMIO and non-VMIO 1526 * buffers. 1527 */ 1528 1529 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1530 1531 /* 1532 * If we are defragging then we need a buffer with 1533 * b_kvasize != 0. XXX this situation should no longer 1534 * occur, if defrag is non-zero the buffer's b_kvasize 1535 * should also be non-zero at this point. XXX 1536 */ 1537 if (defrag && bp->b_kvasize == 0) { 1538 printf("Warning: defrag empty buffer %p\n", bp); 1539 continue; 1540 } 1541 1542 /* 1543 * Start freeing the bp. This is somewhat involved. nbp 1544 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1545 */ 1546 1547 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1548 panic("getnewbuf: locked buf"); 1549 bremfree(bp); 1550 1551 if (qindex == QUEUE_CLEAN) { 1552 if (bp->b_flags & B_VMIO) { 1553 bp->b_flags &= ~B_ASYNC; 1554 vfs_vmio_release(bp); 1555 } 1556 if (bp->b_vp) 1557 brelvp(bp); 1558 } 1559 1560 /* 1561 * NOTE: nbp is now entirely invalid. We can only restart 1562 * the scan from this point on. 1563 * 1564 * Get the rest of the buffer freed up. b_kva* is still 1565 * valid after this operation. 1566 */ 1567 1568 if (bp->b_rcred != NOCRED) { 1569 crfree(bp->b_rcred); 1570 bp->b_rcred = NOCRED; 1571 } 1572 if (bp->b_wcred != NOCRED) { 1573 crfree(bp->b_wcred); 1574 bp->b_wcred = NOCRED; 1575 } 1576 if (LIST_FIRST(&bp->b_dep) != NULL) 1577 buf_deallocate(bp); 1578 if (bp->b_xflags & BX_BKGRDINPROG) 1579 panic("losing buffer 3"); 1580 LIST_REMOVE(bp, b_hash); 1581 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1582 1583 if (bp->b_bufsize) 1584 allocbuf(bp, 0); 1585 1586 bp->b_flags = 0; 1587 bp->b_ioflags = 0; 1588 bp->b_xflags = 0; 1589 bp->b_dev = NODEV; 1590 bp->b_vp = NULL; 1591 bp->b_blkno = bp->b_lblkno = 0; 1592 bp->b_offset = NOOFFSET; 1593 bp->b_iodone = 0; 1594 bp->b_error = 0; 1595 bp->b_resid = 0; 1596 bp->b_bcount = 0; 1597 bp->b_npages = 0; 1598 bp->b_dirtyoff = bp->b_dirtyend = 0; 1599 1600 LIST_INIT(&bp->b_dep); 1601 1602 /* 1603 * If we are defragging then free the buffer. 1604 */ 1605 if (defrag) { 1606 bp->b_flags |= B_INVAL; 1607 bfreekva(bp); 1608 brelse(bp); 1609 defrag = 0; 1610 goto restart; 1611 } 1612 1613 /* 1614 * If we are a normal process then deal with bufspace 1615 * hysteresis. A normal process tries to keep bufspace 1616 * between lobufspace and hibufspace. Note: if we encounter 1617 * a buffer with b_kvasize == 0 then it means we started 1618 * our scan on the EMPTY list and should allocate a new 1619 * buffer. 1620 */ 1621 if (isspecial == 0) { 1622 if (bufspace > hibufspace) 1623 flushingbufs = 1; 1624 if (flushingbufs && bp->b_kvasize != 0) { 1625 bp->b_flags |= B_INVAL; 1626 bfreekva(bp); 1627 brelse(bp); 1628 goto restart; 1629 } 1630 if (bufspace < lobufspace) 1631 flushingbufs = 0; 1632 } 1633 break; 1634 } 1635 1636 /* 1637 * If we exhausted our list, sleep as appropriate. We may have to 1638 * wakeup various daemons and write out some dirty buffers. 1639 * 1640 * Generally we are sleeping due to insufficient buffer space. 1641 */ 1642 1643 if (bp == NULL) { 1644 int flags; 1645 char *waitmsg; 1646 1647 if (defrag) { 1648 flags = VFS_BIO_NEED_BUFSPACE; 1649 waitmsg = "nbufkv"; 1650 } else if (bufspace >= hibufspace) { 1651 waitmsg = "nbufbs"; 1652 flags = VFS_BIO_NEED_BUFSPACE; 1653 } else { 1654 waitmsg = "newbuf"; 1655 flags = VFS_BIO_NEED_ANY; 1656 } 1657 1658 bd_speedup(); /* heeeelp */ 1659 1660 needsbuffer |= flags; 1661 while (needsbuffer & flags) { 1662 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, 1663 waitmsg, slptimeo)) 1664 return (NULL); 1665 } 1666 } else { 1667 /* 1668 * We finally have a valid bp. We aren't quite out of the 1669 * woods, we still have to reserve kva space. In order 1670 * to keep fragmentation sane we only allocate kva in 1671 * BKVASIZE chunks. 1672 */ 1673 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1674 1675 if (maxsize != bp->b_kvasize) { 1676 vm_offset_t addr = 0; 1677 1678 bfreekva(bp); 1679 1680 if (vm_map_findspace(buffer_map, 1681 vm_map_min(buffer_map), maxsize, &addr)) { 1682 /* 1683 * Uh oh. Buffer map is to fragmented. We 1684 * must defragment the map. 1685 */ 1686 ++bufdefragcnt; 1687 defrag = 1; 1688 bp->b_flags |= B_INVAL; 1689 brelse(bp); 1690 goto restart; 1691 } 1692 if (addr) { 1693 vm_map_insert(buffer_map, NULL, 0, 1694 addr, addr + maxsize, 1695 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1696 1697 bp->b_kvabase = (caddr_t) addr; 1698 bp->b_kvasize = maxsize; 1699 bufspace += bp->b_kvasize; 1700 ++bufreusecnt; 1701 } 1702 } 1703 bp->b_data = bp->b_kvabase; 1704 } 1705 return(bp); 1706} 1707 1708/* 1709 * waitfreebuffers: 1710 * 1711 * Wait for sufficient free buffers. Only called from normal processes. 1712 */ 1713 1714static void 1715waitfreebuffers(int slpflag, int slptimeo) 1716{ 1717 while (numfreebuffers < hifreebuffers) { 1718 if (numfreebuffers >= hifreebuffers) 1719 break; 1720 needsbuffer |= VFS_BIO_NEED_FREE; 1721 if (tsleep(&needsbuffer, (PRIBIO + 4)|slpflag, "biofre", slptimeo)) 1722 break; 1723 } 1724} 1725 1726/* 1727 * buf_daemon: 1728 * 1729 * buffer flushing daemon. Buffers are normally flushed by the 1730 * update daemon but if it cannot keep up this process starts to 1731 * take the load in an attempt to prevent getnewbuf() from blocking. 1732 */ 1733 1734static struct proc *bufdaemonproc; 1735static int bd_interval; 1736static int bd_flushto; 1737static int bd_flushinc; 1738 1739static struct kproc_desc buf_kp = { 1740 "bufdaemon", 1741 buf_daemon, 1742 &bufdaemonproc 1743}; 1744SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1745 1746static void 1747buf_daemon() 1748{ 1749 int s; 1750 1751 mtx_enter(&Giant, MTX_DEF); 1752 1753 /* 1754 * This process needs to be suspended prior to shutdown sync. 1755 */ 1756 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, bufdaemonproc, 1757 SHUTDOWN_PRI_LAST); 1758 1759 /* 1760 * This process is allowed to take the buffer cache to the limit 1761 */ 1762 curproc->p_flag |= P_BUFEXHAUST; 1763 s = splbio(); 1764 1765 bd_interval = 5 * hz; /* dynamically adjusted */ 1766 bd_flushto = hidirtybuffers; /* dynamically adjusted */ 1767 bd_flushinc = 1; 1768 1769 for (;;) { 1770 kproc_suspend_loop(bufdaemonproc); 1771 1772 bd_request = 0; 1773 1774 /* 1775 * Do the flush. Limit the number of buffers we flush in one 1776 * go. The failure condition occurs when processes are writing 1777 * buffers faster then we can dispose of them. In this case 1778 * we may be flushing so often that the previous set of flushes 1779 * have not had time to complete, causing us to run out of 1780 * physical buffers and block. 1781 */ 1782 { 1783 int runcount = maxbdrun; 1784 1785 while (numdirtybuffers > bd_flushto && runcount) { 1786 --runcount; 1787 if (flushbufqueues() == 0) 1788 break; 1789 } 1790 } 1791 1792 if (bd_request || 1793 tsleep(&bd_request, PVM, "psleep", bd_interval) == 0) { 1794 /* 1795 * Another request is pending or we were woken up 1796 * without timing out. Flush more. 1797 */ 1798 --bd_flushto; 1799 if (bd_flushto >= numdirtybuffers - 5) { 1800 bd_flushto = numdirtybuffers - 10; 1801 bd_flushinc = 1; 1802 } 1803 if (bd_flushto < 2) 1804 bd_flushto = 2; 1805 } else { 1806 /* 1807 * We slept and timed out, we can slow down. 1808 */ 1809 bd_flushto += bd_flushinc; 1810 if (bd_flushto > hidirtybuffers) 1811 bd_flushto = hidirtybuffers; 1812 ++bd_flushinc; 1813 if (bd_flushinc > hidirtybuffers / 20 + 1) 1814 bd_flushinc = hidirtybuffers / 20 + 1; 1815 } 1816 1817 /* 1818 * Set the interval on a linear scale based on hidirtybuffers 1819 * with a maximum frequency of 1/10 second. 1820 */ 1821 bd_interval = bd_flushto * 5 * hz / hidirtybuffers; 1822 if (bd_interval < hz / 10) 1823 bd_interval = hz / 10; 1824 } 1825} 1826 1827/* 1828 * flushbufqueues: 1829 * 1830 * Try to flush a buffer in the dirty queue. We must be careful to 1831 * free up B_INVAL buffers instead of write them, which NFS is 1832 * particularly sensitive to. 1833 */ 1834 1835static int 1836flushbufqueues(void) 1837{ 1838 struct buf *bp; 1839 int r = 0; 1840 1841 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1842 1843 while (bp) { 1844 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); 1845 if ((bp->b_flags & B_DELWRI) != 0 && 1846 (bp->b_xflags & BX_BKGRDINPROG) == 0) { 1847 if (bp->b_flags & B_INVAL) { 1848 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1849 panic("flushbufqueues: locked buf"); 1850 bremfree(bp); 1851 brelse(bp); 1852 ++r; 1853 break; 1854 } 1855 if (LIST_FIRST(&bp->b_dep) != NULL && 1856 (bp->b_flags & B_DEFERRED) == 0 && 1857 buf_countdeps(bp, 0)) { 1858 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], 1859 bp, b_freelist); 1860 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], 1861 bp, b_freelist); 1862 bp->b_flags |= B_DEFERRED; 1863 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1864 continue; 1865 } 1866 vfs_bio_awrite(bp); 1867 ++r; 1868 break; 1869 } 1870 bp = TAILQ_NEXT(bp, b_freelist); 1871 } 1872 return (r); 1873} 1874 1875/* 1876 * Check to see if a block is currently memory resident. 1877 */ 1878struct buf * 1879incore(struct vnode * vp, daddr_t blkno) 1880{ 1881 struct buf *bp; 1882 1883 int s = splbio(); 1884 bp = gbincore(vp, blkno); 1885 splx(s); 1886 return (bp); 1887} 1888 1889/* 1890 * Returns true if no I/O is needed to access the 1891 * associated VM object. This is like incore except 1892 * it also hunts around in the VM system for the data. 1893 */ 1894 1895int 1896inmem(struct vnode * vp, daddr_t blkno) 1897{ 1898 vm_object_t obj; 1899 vm_offset_t toff, tinc, size; 1900 vm_page_t m; 1901 vm_ooffset_t off; 1902 1903 if (incore(vp, blkno)) 1904 return 1; 1905 if (vp->v_mount == NULL) 1906 return 0; 1907 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_flag & VOBJBUF) == 0) 1908 return 0; 1909 1910 size = PAGE_SIZE; 1911 if (size > vp->v_mount->mnt_stat.f_iosize) 1912 size = vp->v_mount->mnt_stat.f_iosize; 1913 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 1914 1915 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1916 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 1917 if (!m) 1918 return 0; 1919 tinc = size; 1920 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 1921 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 1922 if (vm_page_is_valid(m, 1923 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 1924 return 0; 1925 } 1926 return 1; 1927} 1928 1929/* 1930 * vfs_setdirty: 1931 * 1932 * Sets the dirty range for a buffer based on the status of the dirty 1933 * bits in the pages comprising the buffer. 1934 * 1935 * The range is limited to the size of the buffer. 1936 * 1937 * This routine is primarily used by NFS, but is generalized for the 1938 * B_VMIO case. 1939 */ 1940static void 1941vfs_setdirty(struct buf *bp) 1942{ 1943 int i; 1944 vm_object_t object; 1945 1946 /* 1947 * Degenerate case - empty buffer 1948 */ 1949 1950 if (bp->b_bufsize == 0) 1951 return; 1952 1953 /* 1954 * We qualify the scan for modified pages on whether the 1955 * object has been flushed yet. The OBJ_WRITEABLE flag 1956 * is not cleared simply by protecting pages off. 1957 */ 1958 1959 if ((bp->b_flags & B_VMIO) == 0) 1960 return; 1961 1962 object = bp->b_pages[0]->object; 1963 1964 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 1965 printf("Warning: object %p writeable but not mightbedirty\n", object); 1966 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 1967 printf("Warning: object %p mightbedirty but not writeable\n", object); 1968 1969 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 1970 vm_offset_t boffset; 1971 vm_offset_t eoffset; 1972 1973 /* 1974 * test the pages to see if they have been modified directly 1975 * by users through the VM system. 1976 */ 1977 for (i = 0; i < bp->b_npages; i++) { 1978 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 1979 vm_page_test_dirty(bp->b_pages[i]); 1980 } 1981 1982 /* 1983 * Calculate the encompassing dirty range, boffset and eoffset, 1984 * (eoffset - boffset) bytes. 1985 */ 1986 1987 for (i = 0; i < bp->b_npages; i++) { 1988 if (bp->b_pages[i]->dirty) 1989 break; 1990 } 1991 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 1992 1993 for (i = bp->b_npages - 1; i >= 0; --i) { 1994 if (bp->b_pages[i]->dirty) { 1995 break; 1996 } 1997 } 1998 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 1999 2000 /* 2001 * Fit it to the buffer. 2002 */ 2003 2004 if (eoffset > bp->b_bcount) 2005 eoffset = bp->b_bcount; 2006 2007 /* 2008 * If we have a good dirty range, merge with the existing 2009 * dirty range. 2010 */ 2011 2012 if (boffset < eoffset) { 2013 if (bp->b_dirtyoff > boffset) 2014 bp->b_dirtyoff = boffset; 2015 if (bp->b_dirtyend < eoffset) 2016 bp->b_dirtyend = eoffset; 2017 } 2018 } 2019} 2020 2021/* 2022 * getblk: 2023 * 2024 * Get a block given a specified block and offset into a file/device. 2025 * The buffers B_DONE bit will be cleared on return, making it almost 2026 * ready for an I/O initiation. B_INVAL may or may not be set on 2027 * return. The caller should clear B_INVAL prior to initiating a 2028 * READ. 2029 * 2030 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2031 * an existing buffer. 2032 * 2033 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2034 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2035 * and then cleared based on the backing VM. If the previous buffer is 2036 * non-0-sized but invalid, B_CACHE will be cleared. 2037 * 2038 * If getblk() must create a new buffer, the new buffer is returned with 2039 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2040 * case it is returned with B_INVAL clear and B_CACHE set based on the 2041 * backing VM. 2042 * 2043 * getblk() also forces a VOP_BWRITE() for any B_DELWRI buffer whos 2044 * B_CACHE bit is clear. 2045 * 2046 * What this means, basically, is that the caller should use B_CACHE to 2047 * determine whether the buffer is fully valid or not and should clear 2048 * B_INVAL prior to issuing a read. If the caller intends to validate 2049 * the buffer by loading its data area with something, the caller needs 2050 * to clear B_INVAL. If the caller does this without issuing an I/O, 2051 * the caller should set B_CACHE ( as an optimization ), else the caller 2052 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2053 * a write attempt or if it was a successfull read. If the caller 2054 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2055 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2056 */ 2057struct buf * 2058getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2059{ 2060 struct buf *bp; 2061 int s; 2062 struct bufhashhdr *bh; 2063 2064 if (size > MAXBSIZE) 2065 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2066 2067 s = splbio(); 2068loop: 2069 /* 2070 * Block if we are low on buffers. Certain processes are allowed 2071 * to completely exhaust the buffer cache. 2072 * 2073 * If this check ever becomes a bottleneck it may be better to 2074 * move it into the else, when gbincore() fails. At the moment 2075 * it isn't a problem. 2076 */ 2077 if (curproc == idleproc || (curproc->p_flag & P_BUFEXHAUST)) { 2078 if (numfreebuffers == 0) { 2079 if (curproc == idleproc) 2080 return NULL; 2081 needsbuffer |= VFS_BIO_NEED_ANY; 2082 tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, "newbuf", 2083 slptimeo); 2084 } 2085 } else if (numfreebuffers < lofreebuffers) { 2086 waitfreebuffers(slpflag, slptimeo); 2087 } 2088 2089 if ((bp = gbincore(vp, blkno))) { 2090 /* 2091 * Buffer is in-core. If the buffer is not busy, it must 2092 * be on a queue. 2093 */ 2094 2095 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2096 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, 2097 "getblk", slpflag, slptimeo) == ENOLCK) 2098 goto loop; 2099 splx(s); 2100 return (struct buf *) NULL; 2101 } 2102 2103 /* 2104 * The buffer is locked. B_CACHE is cleared if the buffer is 2105 * invalid. Ohterwise, for a non-VMIO buffer, B_CACHE is set 2106 * and for a VMIO buffer B_CACHE is adjusted according to the 2107 * backing VM cache. 2108 */ 2109 if (bp->b_flags & B_INVAL) 2110 bp->b_flags &= ~B_CACHE; 2111 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2112 bp->b_flags |= B_CACHE; 2113 bremfree(bp); 2114 2115 /* 2116 * check for size inconsistancies for non-VMIO case. 2117 */ 2118 2119 if (bp->b_bcount != size) { 2120 if ((bp->b_flags & B_VMIO) == 0 || 2121 (size > bp->b_kvasize)) { 2122 if (bp->b_flags & B_DELWRI) { 2123 bp->b_flags |= B_NOCACHE; 2124 BUF_WRITE(bp); 2125 } else { 2126 if ((bp->b_flags & B_VMIO) && 2127 (LIST_FIRST(&bp->b_dep) == NULL)) { 2128 bp->b_flags |= B_RELBUF; 2129 brelse(bp); 2130 } else { 2131 bp->b_flags |= B_NOCACHE; 2132 BUF_WRITE(bp); 2133 } 2134 } 2135 goto loop; 2136 } 2137 } 2138 2139 /* 2140 * If the size is inconsistant in the VMIO case, we can resize 2141 * the buffer. This might lead to B_CACHE getting set or 2142 * cleared. If the size has not changed, B_CACHE remains 2143 * unchanged from its previous state. 2144 */ 2145 2146 if (bp->b_bcount != size) 2147 allocbuf(bp, size); 2148 2149 KASSERT(bp->b_offset != NOOFFSET, 2150 ("getblk: no buffer offset")); 2151 2152 /* 2153 * A buffer with B_DELWRI set and B_CACHE clear must 2154 * be committed before we can return the buffer in 2155 * order to prevent the caller from issuing a read 2156 * ( due to B_CACHE not being set ) and overwriting 2157 * it. 2158 * 2159 * Most callers, including NFS and FFS, need this to 2160 * operate properly either because they assume they 2161 * can issue a read if B_CACHE is not set, or because 2162 * ( for example ) an uncached B_DELWRI might loop due 2163 * to softupdates re-dirtying the buffer. In the latter 2164 * case, B_CACHE is set after the first write completes, 2165 * preventing further loops. 2166 */ 2167 2168 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2169 BUF_WRITE(bp); 2170 goto loop; 2171 } 2172 2173 splx(s); 2174 bp->b_flags &= ~B_DONE; 2175 } else { 2176 /* 2177 * Buffer is not in-core, create new buffer. The buffer 2178 * returned by getnewbuf() is locked. Note that the returned 2179 * buffer is also considered valid (not marked B_INVAL). 2180 */ 2181 int bsize, maxsize, vmio; 2182 off_t offset; 2183 2184 if (vn_isdisk(vp, NULL)) 2185 bsize = DEV_BSIZE; 2186 else if (vp->v_mountedhere) 2187 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2188 else if (vp->v_mount) 2189 bsize = vp->v_mount->mnt_stat.f_iosize; 2190 else 2191 bsize = size; 2192 2193 offset = (off_t)blkno * bsize; 2194 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && (vp->v_flag & VOBJBUF); 2195 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2196 maxsize = imax(maxsize, bsize); 2197 2198 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2199 if (slpflag || slptimeo) { 2200 splx(s); 2201 return NULL; 2202 } 2203 goto loop; 2204 } 2205 2206 /* 2207 * This code is used to make sure that a buffer is not 2208 * created while the getnewbuf routine is blocked. 2209 * This can be a problem whether the vnode is locked or not. 2210 * If the buffer is created out from under us, we have to 2211 * throw away the one we just created. There is now window 2212 * race because we are safely running at splbio() from the 2213 * point of the duplicate buffer creation through to here, 2214 * and we've locked the buffer. 2215 */ 2216 if (gbincore(vp, blkno)) { 2217 bp->b_flags |= B_INVAL; 2218 brelse(bp); 2219 goto loop; 2220 } 2221 2222 /* 2223 * Insert the buffer into the hash, so that it can 2224 * be found by incore. 2225 */ 2226 bp->b_blkno = bp->b_lblkno = blkno; 2227 bp->b_offset = offset; 2228 2229 bgetvp(vp, bp); 2230 LIST_REMOVE(bp, b_hash); 2231 bh = bufhash(vp, blkno); 2232 LIST_INSERT_HEAD(bh, bp, b_hash); 2233 2234 /* 2235 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2236 * buffer size starts out as 0, B_CACHE will be set by 2237 * allocbuf() for the VMIO case prior to it testing the 2238 * backing store for validity. 2239 */ 2240 2241 if (vmio) { 2242 bp->b_flags |= B_VMIO; 2243#if defined(VFS_BIO_DEBUG) 2244 if (vp->v_type != VREG && vp->v_type != VBLK) 2245 printf("getblk: vmioing file type %d???\n", vp->v_type); 2246#endif 2247 } else { 2248 bp->b_flags &= ~B_VMIO; 2249 } 2250 2251 allocbuf(bp, size); 2252 2253 splx(s); 2254 bp->b_flags &= ~B_DONE; 2255 } 2256 return (bp); 2257} 2258 2259/* 2260 * Get an empty, disassociated buffer of given size. The buffer is initially 2261 * set to B_INVAL. 2262 */ 2263struct buf * 2264geteblk(int size) 2265{ 2266 struct buf *bp; 2267 int s; 2268 int maxsize; 2269 2270 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2271 2272 s = splbio(); 2273 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0); 2274 splx(s); 2275 allocbuf(bp, size); 2276 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2277 return (bp); 2278} 2279 2280 2281/* 2282 * This code constitutes the buffer memory from either anonymous system 2283 * memory (in the case of non-VMIO operations) or from an associated 2284 * VM object (in the case of VMIO operations). This code is able to 2285 * resize a buffer up or down. 2286 * 2287 * Note that this code is tricky, and has many complications to resolve 2288 * deadlock or inconsistant data situations. Tread lightly!!! 2289 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2290 * the caller. Calling this code willy nilly can result in the loss of data. 2291 * 2292 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2293 * B_CACHE for the non-VMIO case. 2294 */ 2295 2296int 2297allocbuf(struct buf *bp, int size) 2298{ 2299 int newbsize, mbsize; 2300 int i; 2301 2302 if (BUF_REFCNT(bp) == 0) 2303 panic("allocbuf: buffer not busy"); 2304 2305 if (bp->b_kvasize < size) 2306 panic("allocbuf: buffer too small"); 2307 2308 if ((bp->b_flags & B_VMIO) == 0) { 2309 caddr_t origbuf; 2310 int origbufsize; 2311 /* 2312 * Just get anonymous memory from the kernel. Don't 2313 * mess with B_CACHE. 2314 */ 2315 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2316#if !defined(NO_B_MALLOC) 2317 if (bp->b_flags & B_MALLOC) 2318 newbsize = mbsize; 2319 else 2320#endif 2321 newbsize = round_page(size); 2322 2323 if (newbsize < bp->b_bufsize) { 2324#if !defined(NO_B_MALLOC) 2325 /* 2326 * malloced buffers are not shrunk 2327 */ 2328 if (bp->b_flags & B_MALLOC) { 2329 if (newbsize) { 2330 bp->b_bcount = size; 2331 } else { 2332 free(bp->b_data, M_BIOBUF); 2333 bufmallocspace -= bp->b_bufsize; 2334 runningbufspace -= bp->b_bufsize; 2335 if (bp->b_bufsize) 2336 bufspacewakeup(); 2337 bp->b_data = bp->b_kvabase; 2338 bp->b_bufsize = 0; 2339 bp->b_bcount = 0; 2340 bp->b_flags &= ~B_MALLOC; 2341 } 2342 return 1; 2343 } 2344#endif 2345 vm_hold_free_pages( 2346 bp, 2347 (vm_offset_t) bp->b_data + newbsize, 2348 (vm_offset_t) bp->b_data + bp->b_bufsize); 2349 } else if (newbsize > bp->b_bufsize) { 2350#if !defined(NO_B_MALLOC) 2351 /* 2352 * We only use malloced memory on the first allocation. 2353 * and revert to page-allocated memory when the buffer 2354 * grows. 2355 */ 2356 if ( (bufmallocspace < maxbufmallocspace) && 2357 (bp->b_bufsize == 0) && 2358 (mbsize <= PAGE_SIZE/2)) { 2359 2360 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2361 bp->b_bufsize = mbsize; 2362 bp->b_bcount = size; 2363 bp->b_flags |= B_MALLOC; 2364 bufmallocspace += mbsize; 2365 runningbufspace += bp->b_bufsize; 2366 return 1; 2367 } 2368#endif 2369 origbuf = NULL; 2370 origbufsize = 0; 2371#if !defined(NO_B_MALLOC) 2372 /* 2373 * If the buffer is growing on its other-than-first allocation, 2374 * then we revert to the page-allocation scheme. 2375 */ 2376 if (bp->b_flags & B_MALLOC) { 2377 origbuf = bp->b_data; 2378 origbufsize = bp->b_bufsize; 2379 bp->b_data = bp->b_kvabase; 2380 bufmallocspace -= bp->b_bufsize; 2381 runningbufspace -= bp->b_bufsize; 2382 if (bp->b_bufsize) 2383 bufspacewakeup(); 2384 bp->b_bufsize = 0; 2385 bp->b_flags &= ~B_MALLOC; 2386 newbsize = round_page(newbsize); 2387 } 2388#endif 2389 vm_hold_load_pages( 2390 bp, 2391 (vm_offset_t) bp->b_data + bp->b_bufsize, 2392 (vm_offset_t) bp->b_data + newbsize); 2393#if !defined(NO_B_MALLOC) 2394 if (origbuf) { 2395 bcopy(origbuf, bp->b_data, origbufsize); 2396 free(origbuf, M_BIOBUF); 2397 } 2398#endif 2399 } 2400 } else { 2401 vm_page_t m; 2402 int desiredpages; 2403 2404 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2405 desiredpages = (size == 0) ? 0 : 2406 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2407 2408#if !defined(NO_B_MALLOC) 2409 if (bp->b_flags & B_MALLOC) 2410 panic("allocbuf: VMIO buffer can't be malloced"); 2411#endif 2412 /* 2413 * Set B_CACHE initially if buffer is 0 length or will become 2414 * 0-length. 2415 */ 2416 if (size == 0 || bp->b_bufsize == 0) 2417 bp->b_flags |= B_CACHE; 2418 2419 if (newbsize < bp->b_bufsize) { 2420 /* 2421 * DEV_BSIZE aligned new buffer size is less then the 2422 * DEV_BSIZE aligned existing buffer size. Figure out 2423 * if we have to remove any pages. 2424 */ 2425 if (desiredpages < bp->b_npages) { 2426 for (i = desiredpages; i < bp->b_npages; i++) { 2427 /* 2428 * the page is not freed here -- it 2429 * is the responsibility of 2430 * vnode_pager_setsize 2431 */ 2432 m = bp->b_pages[i]; 2433 KASSERT(m != bogus_page, 2434 ("allocbuf: bogus page found")); 2435 while (vm_page_sleep_busy(m, TRUE, "biodep")) 2436 ; 2437 2438 bp->b_pages[i] = NULL; 2439 vm_page_unwire(m, 0); 2440 } 2441 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2442 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2443 bp->b_npages = desiredpages; 2444 } 2445 } else if (size > bp->b_bcount) { 2446 /* 2447 * We are growing the buffer, possibly in a 2448 * byte-granular fashion. 2449 */ 2450 struct vnode *vp; 2451 vm_object_t obj; 2452 vm_offset_t toff; 2453 vm_offset_t tinc; 2454 2455 /* 2456 * Step 1, bring in the VM pages from the object, 2457 * allocating them if necessary. We must clear 2458 * B_CACHE if these pages are not valid for the 2459 * range covered by the buffer. 2460 */ 2461 2462 vp = bp->b_vp; 2463 VOP_GETVOBJECT(vp, &obj); 2464 2465 while (bp->b_npages < desiredpages) { 2466 vm_page_t m; 2467 vm_pindex_t pi; 2468 2469 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2470 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2471 m = vm_page_alloc(obj, pi, VM_ALLOC_NORMAL); 2472 if (m == NULL) { 2473 VM_WAIT; 2474 vm_pageout_deficit += desiredpages - bp->b_npages; 2475 } else { 2476 vm_page_wire(m); 2477 vm_page_wakeup(m); 2478 bp->b_flags &= ~B_CACHE; 2479 bp->b_pages[bp->b_npages] = m; 2480 ++bp->b_npages; 2481 } 2482 continue; 2483 } 2484 2485 /* 2486 * We found a page. If we have to sleep on it, 2487 * retry because it might have gotten freed out 2488 * from under us. 2489 * 2490 * We can only test PG_BUSY here. Blocking on 2491 * m->busy might lead to a deadlock: 2492 * 2493 * vm_fault->getpages->cluster_read->allocbuf 2494 * 2495 */ 2496 2497 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 2498 continue; 2499 2500 /* 2501 * We have a good page. Should we wakeup the 2502 * page daemon? 2503 */ 2504 if ((curproc != pageproc) && 2505 ((m->queue - m->pc) == PQ_CACHE) && 2506 ((cnt.v_free_count + cnt.v_cache_count) < 2507 (cnt.v_free_min + cnt.v_cache_min))) { 2508 pagedaemon_wakeup(); 2509 } 2510 vm_page_flag_clear(m, PG_ZERO); 2511 vm_page_wire(m); 2512 bp->b_pages[bp->b_npages] = m; 2513 ++bp->b_npages; 2514 } 2515 2516 /* 2517 * Step 2. We've loaded the pages into the buffer, 2518 * we have to figure out if we can still have B_CACHE 2519 * set. Note that B_CACHE is set according to the 2520 * byte-granular range ( bcount and size ), new the 2521 * aligned range ( newbsize ). 2522 * 2523 * The VM test is against m->valid, which is DEV_BSIZE 2524 * aligned. Needless to say, the validity of the data 2525 * needs to also be DEV_BSIZE aligned. Note that this 2526 * fails with NFS if the server or some other client 2527 * extends the file's EOF. If our buffer is resized, 2528 * B_CACHE may remain set! XXX 2529 */ 2530 2531 toff = bp->b_bcount; 2532 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2533 2534 while ((bp->b_flags & B_CACHE) && toff < size) { 2535 vm_pindex_t pi; 2536 2537 if (tinc > (size - toff)) 2538 tinc = size - toff; 2539 2540 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2541 PAGE_SHIFT; 2542 2543 vfs_buf_test_cache( 2544 bp, 2545 bp->b_offset, 2546 toff, 2547 tinc, 2548 bp->b_pages[pi] 2549 ); 2550 toff += tinc; 2551 tinc = PAGE_SIZE; 2552 } 2553 2554 /* 2555 * Step 3, fixup the KVM pmap. Remember that 2556 * bp->b_data is relative to bp->b_offset, but 2557 * bp->b_offset may be offset into the first page. 2558 */ 2559 2560 bp->b_data = (caddr_t) 2561 trunc_page((vm_offset_t)bp->b_data); 2562 pmap_qenter( 2563 (vm_offset_t)bp->b_data, 2564 bp->b_pages, 2565 bp->b_npages 2566 ); 2567 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2568 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2569 } 2570 } 2571 runningbufspace += (newbsize - bp->b_bufsize); 2572 if (newbsize < bp->b_bufsize) 2573 bufspacewakeup(); 2574 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2575 bp->b_bcount = size; /* requested buffer size */ 2576 return 1; 2577} 2578 2579/* 2580 * bufwait: 2581 * 2582 * Wait for buffer I/O completion, returning error status. The buffer 2583 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR 2584 * error and cleared. 2585 */ 2586int 2587bufwait(register struct buf * bp) 2588{ 2589 int s; 2590 2591 s = splbio(); 2592 while ((bp->b_flags & B_DONE) == 0) { 2593 if (bp->b_iocmd == BIO_READ) 2594 tsleep(bp, PRIBIO, "biord", 0); 2595 else 2596 tsleep(bp, PRIBIO, "biowr", 0); 2597 } 2598 splx(s); 2599 if (bp->b_flags & B_EINTR) { 2600 bp->b_flags &= ~B_EINTR; 2601 return (EINTR); 2602 } 2603 if (bp->b_ioflags & BIO_ERROR) { 2604 return (bp->b_error ? bp->b_error : EIO); 2605 } else { 2606 return (0); 2607 } 2608} 2609 2610 /* 2611 * Call back function from struct bio back up to struct buf. 2612 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 2613 */ 2614void 2615bufdonebio(struct bio *bp) 2616{ 2617 bufdone(bp->bio_caller2); 2618} 2619 2620/* 2621 * bufdone: 2622 * 2623 * Finish I/O on a buffer, optionally calling a completion function. 2624 * This is usually called from an interrupt so process blocking is 2625 * not allowed. 2626 * 2627 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2628 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2629 * assuming B_INVAL is clear. 2630 * 2631 * For the VMIO case, we set B_CACHE if the op was a read and no 2632 * read error occured, or if the op was a write. B_CACHE is never 2633 * set if the buffer is invalid or otherwise uncacheable. 2634 * 2635 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2636 * initiator to leave B_INVAL set to brelse the buffer out of existance 2637 * in the biodone routine. 2638 */ 2639void 2640bufdone(struct buf *bp) 2641{ 2642 int s, error; 2643 void (*biodone) __P((struct buf *)); 2644 2645 s = splbio(); 2646 2647 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 2648 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 2649 2650 bp->b_flags |= B_DONE; 2651 2652 if (bp->b_iocmd == BIO_DELETE) { 2653 brelse(bp); 2654 splx(s); 2655 return; 2656 } 2657 2658 if (bp->b_iocmd == BIO_WRITE) { 2659 vwakeup(bp); 2660 } 2661 2662 /* call optional completion function if requested */ 2663 if (bp->b_iodone != NULL) { 2664 biodone = bp->b_iodone; 2665 bp->b_iodone = NULL; 2666 (*biodone) (bp); 2667 splx(s); 2668 return; 2669 } 2670 if (LIST_FIRST(&bp->b_dep) != NULL) 2671 buf_complete(bp); 2672 2673 if (bp->b_flags & B_VMIO) { 2674 int i, resid; 2675 vm_ooffset_t foff; 2676 vm_page_t m; 2677 vm_object_t obj; 2678 int iosize; 2679 struct vnode *vp = bp->b_vp; 2680 2681 error = VOP_GETVOBJECT(vp, &obj); 2682 2683#if defined(VFS_BIO_DEBUG) 2684 if (vp->v_usecount == 0) { 2685 panic("biodone: zero vnode ref count"); 2686 } 2687 2688 if (error) { 2689 panic("biodone: missing VM object"); 2690 } 2691 2692 if ((vp->v_flag & VOBJBUF) == 0) { 2693 panic("biodone: vnode is not setup for merged cache"); 2694 } 2695#endif 2696 2697 foff = bp->b_offset; 2698 KASSERT(bp->b_offset != NOOFFSET, 2699 ("biodone: no buffer offset")); 2700 2701 if (error) { 2702 panic("biodone: no object"); 2703 } 2704#if defined(VFS_BIO_DEBUG) 2705 if (obj->paging_in_progress < bp->b_npages) { 2706 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 2707 obj->paging_in_progress, bp->b_npages); 2708 } 2709#endif 2710 2711 /* 2712 * Set B_CACHE if the op was a normal read and no error 2713 * occured. B_CACHE is set for writes in the b*write() 2714 * routines. 2715 */ 2716 iosize = bp->b_bcount - bp->b_resid; 2717 if (bp->b_iocmd == BIO_READ && 2718 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 2719 !(bp->b_ioflags & BIO_ERROR)) { 2720 bp->b_flags |= B_CACHE; 2721 } 2722 2723 for (i = 0; i < bp->b_npages; i++) { 2724 int bogusflag = 0; 2725 m = bp->b_pages[i]; 2726 if (m == bogus_page) { 2727 bogusflag = 1; 2728 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2729 if (!m) { 2730#if defined(VFS_BIO_DEBUG) 2731 printf("biodone: page disappeared\n"); 2732#endif 2733 vm_object_pip_subtract(obj, 1); 2734 bp->b_flags &= ~B_CACHE; 2735 continue; 2736 } 2737 bp->b_pages[i] = m; 2738 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2739 } 2740#if defined(VFS_BIO_DEBUG) 2741 if (OFF_TO_IDX(foff) != m->pindex) { 2742 printf( 2743"biodone: foff(%lu)/m->pindex(%d) mismatch\n", 2744 (unsigned long)foff, m->pindex); 2745 } 2746#endif 2747 resid = IDX_TO_OFF(m->pindex + 1) - foff; 2748 if (resid > iosize) 2749 resid = iosize; 2750 2751 /* 2752 * In the write case, the valid and clean bits are 2753 * already changed correctly ( see bdwrite() ), so we 2754 * only need to do this here in the read case. 2755 */ 2756 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 2757 vfs_page_set_valid(bp, foff, i, m); 2758 } 2759 vm_page_flag_clear(m, PG_ZERO); 2760 2761 /* 2762 * when debugging new filesystems or buffer I/O methods, this 2763 * is the most common error that pops up. if you see this, you 2764 * have not set the page busy flag correctly!!! 2765 */ 2766 if (m->busy == 0) { 2767 printf("biodone: page busy < 0, " 2768 "pindex: %d, foff: 0x(%x,%x), " 2769 "resid: %d, index: %d\n", 2770 (int) m->pindex, (int)(foff >> 32), 2771 (int) foff & 0xffffffff, resid, i); 2772 if (!vn_isdisk(vp, NULL)) 2773 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 2774 bp->b_vp->v_mount->mnt_stat.f_iosize, 2775 (int) bp->b_lblkno, 2776 bp->b_flags, bp->b_npages); 2777 else 2778 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 2779 (int) bp->b_lblkno, 2780 bp->b_flags, bp->b_npages); 2781 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2782 m->valid, m->dirty, m->wire_count); 2783 panic("biodone: page busy < 0\n"); 2784 } 2785 vm_page_io_finish(m); 2786 vm_object_pip_subtract(obj, 1); 2787 foff += resid; 2788 iosize -= resid; 2789 } 2790 if (obj) 2791 vm_object_pip_wakeupn(obj, 0); 2792 } 2793 /* 2794 * For asynchronous completions, release the buffer now. The brelse 2795 * will do a wakeup there if necessary - so no need to do a wakeup 2796 * here in the async case. The sync case always needs to do a wakeup. 2797 */ 2798 2799 if (bp->b_flags & B_ASYNC) { 2800 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 2801 brelse(bp); 2802 else 2803 bqrelse(bp); 2804 } else { 2805 wakeup(bp); 2806 } 2807 splx(s); 2808} 2809 2810/* 2811 * This routine is called in lieu of iodone in the case of 2812 * incomplete I/O. This keeps the busy status for pages 2813 * consistant. 2814 */ 2815void 2816vfs_unbusy_pages(struct buf * bp) 2817{ 2818 int i; 2819 2820 if (bp->b_flags & B_VMIO) { 2821 struct vnode *vp = bp->b_vp; 2822 vm_object_t obj; 2823 2824 VOP_GETVOBJECT(vp, &obj); 2825 2826 for (i = 0; i < bp->b_npages; i++) { 2827 vm_page_t m = bp->b_pages[i]; 2828 2829 if (m == bogus_page) { 2830 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 2831 if (!m) { 2832 panic("vfs_unbusy_pages: page missing\n"); 2833 } 2834 bp->b_pages[i] = m; 2835 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2836 } 2837 vm_object_pip_subtract(obj, 1); 2838 vm_page_flag_clear(m, PG_ZERO); 2839 vm_page_io_finish(m); 2840 } 2841 vm_object_pip_wakeupn(obj, 0); 2842 } 2843} 2844 2845/* 2846 * vfs_page_set_valid: 2847 * 2848 * Set the valid bits in a page based on the supplied offset. The 2849 * range is restricted to the buffer's size. 2850 * 2851 * This routine is typically called after a read completes. 2852 */ 2853static void 2854vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 2855{ 2856 vm_ooffset_t soff, eoff; 2857 2858 /* 2859 * Start and end offsets in buffer. eoff - soff may not cross a 2860 * page boundry or cross the end of the buffer. The end of the 2861 * buffer, in this case, is our file EOF, not the allocation size 2862 * of the buffer. 2863 */ 2864 soff = off; 2865 eoff = (off + PAGE_SIZE) & ~PAGE_MASK; 2866 if (eoff > bp->b_offset + bp->b_bcount) 2867 eoff = bp->b_offset + bp->b_bcount; 2868 2869 /* 2870 * Set valid range. This is typically the entire buffer and thus the 2871 * entire page. 2872 */ 2873 if (eoff > soff) { 2874 vm_page_set_validclean( 2875 m, 2876 (vm_offset_t) (soff & PAGE_MASK), 2877 (vm_offset_t) (eoff - soff) 2878 ); 2879 } 2880} 2881 2882/* 2883 * This routine is called before a device strategy routine. 2884 * It is used to tell the VM system that paging I/O is in 2885 * progress, and treat the pages associated with the buffer 2886 * almost as being PG_BUSY. Also the object paging_in_progress 2887 * flag is handled to make sure that the object doesn't become 2888 * inconsistant. 2889 * 2890 * Since I/O has not been initiated yet, certain buffer flags 2891 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 2892 * and should be ignored. 2893 */ 2894void 2895vfs_busy_pages(struct buf * bp, int clear_modify) 2896{ 2897 int i, bogus; 2898 2899 if (bp->b_flags & B_VMIO) { 2900 struct vnode *vp = bp->b_vp; 2901 vm_object_t obj; 2902 vm_ooffset_t foff; 2903 2904 VOP_GETVOBJECT(vp, &obj); 2905 foff = bp->b_offset; 2906 KASSERT(bp->b_offset != NOOFFSET, 2907 ("vfs_busy_pages: no buffer offset")); 2908 vfs_setdirty(bp); 2909 2910retry: 2911 for (i = 0; i < bp->b_npages; i++) { 2912 vm_page_t m = bp->b_pages[i]; 2913 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 2914 goto retry; 2915 } 2916 2917 bogus = 0; 2918 for (i = 0; i < bp->b_npages; i++) { 2919 vm_page_t m = bp->b_pages[i]; 2920 2921 vm_page_flag_clear(m, PG_ZERO); 2922 if ((bp->b_flags & B_CLUSTER) == 0) { 2923 vm_object_pip_add(obj, 1); 2924 vm_page_io_start(m); 2925 } 2926 2927 /* 2928 * When readying a buffer for a read ( i.e 2929 * clear_modify == 0 ), it is important to do 2930 * bogus_page replacement for valid pages in 2931 * partially instantiated buffers. Partially 2932 * instantiated buffers can, in turn, occur when 2933 * reconstituting a buffer from its VM backing store 2934 * base. We only have to do this if B_CACHE is 2935 * clear ( which causes the I/O to occur in the 2936 * first place ). The replacement prevents the read 2937 * I/O from overwriting potentially dirty VM-backed 2938 * pages. XXX bogus page replacement is, uh, bogus. 2939 * It may not work properly with small-block devices. 2940 * We need to find a better way. 2941 */ 2942 2943 vm_page_protect(m, VM_PROT_NONE); 2944 if (clear_modify) 2945 vfs_page_set_valid(bp, foff, i, m); 2946 else if (m->valid == VM_PAGE_BITS_ALL && 2947 (bp->b_flags & B_CACHE) == 0) { 2948 bp->b_pages[i] = bogus_page; 2949 bogus++; 2950 } 2951 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 2952 } 2953 if (bogus) 2954 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2955 } 2956} 2957 2958/* 2959 * Tell the VM system that the pages associated with this buffer 2960 * are clean. This is used for delayed writes where the data is 2961 * going to go to disk eventually without additional VM intevention. 2962 * 2963 * Note that while we only really need to clean through to b_bcount, we 2964 * just go ahead and clean through to b_bufsize. 2965 */ 2966static void 2967vfs_clean_pages(struct buf * bp) 2968{ 2969 int i; 2970 2971 if (bp->b_flags & B_VMIO) { 2972 vm_ooffset_t foff; 2973 2974 foff = bp->b_offset; 2975 KASSERT(bp->b_offset != NOOFFSET, 2976 ("vfs_clean_pages: no buffer offset")); 2977 for (i = 0; i < bp->b_npages; i++) { 2978 vm_page_t m = bp->b_pages[i]; 2979 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~PAGE_MASK; 2980 vm_ooffset_t eoff = noff; 2981 2982 if (eoff > bp->b_offset + bp->b_bufsize) 2983 eoff = bp->b_offset + bp->b_bufsize; 2984 vfs_page_set_valid(bp, foff, i, m); 2985 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2986 foff = noff; 2987 } 2988 } 2989} 2990 2991/* 2992 * vfs_bio_set_validclean: 2993 * 2994 * Set the range within the buffer to valid and clean. The range is 2995 * relative to the beginning of the buffer, b_offset. Note that b_offset 2996 * itself may be offset from the beginning of the first page. 2997 */ 2998 2999void 3000vfs_bio_set_validclean(struct buf *bp, int base, int size) 3001{ 3002 if (bp->b_flags & B_VMIO) { 3003 int i; 3004 int n; 3005 3006 /* 3007 * Fixup base to be relative to beginning of first page. 3008 * Set initial n to be the maximum number of bytes in the 3009 * first page that can be validated. 3010 */ 3011 3012 base += (bp->b_offset & PAGE_MASK); 3013 n = PAGE_SIZE - (base & PAGE_MASK); 3014 3015 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3016 vm_page_t m = bp->b_pages[i]; 3017 3018 if (n > size) 3019 n = size; 3020 3021 vm_page_set_validclean(m, base & PAGE_MASK, n); 3022 base += n; 3023 size -= n; 3024 n = PAGE_SIZE; 3025 } 3026 } 3027} 3028 3029/* 3030 * vfs_bio_clrbuf: 3031 * 3032 * clear a buffer. This routine essentially fakes an I/O, so we need 3033 * to clear BIO_ERROR and B_INVAL. 3034 * 3035 * Note that while we only theoretically need to clear through b_bcount, 3036 * we go ahead and clear through b_bufsize. 3037 */ 3038 3039void 3040vfs_bio_clrbuf(struct buf *bp) { 3041 int i, mask = 0; 3042 caddr_t sa, ea; 3043 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3044 bp->b_flags &= ~B_INVAL; 3045 bp->b_ioflags &= ~BIO_ERROR; 3046 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3047 (bp->b_offset & PAGE_MASK) == 0) { 3048 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3049 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3050 ((bp->b_pages[0]->valid & mask) != mask)) { 3051 bzero(bp->b_data, bp->b_bufsize); 3052 } 3053 bp->b_pages[0]->valid |= mask; 3054 bp->b_resid = 0; 3055 return; 3056 } 3057 ea = sa = bp->b_data; 3058 for(i=0;i<bp->b_npages;i++,sa=ea) { 3059 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3060 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3061 ea = (caddr_t)(vm_offset_t)ulmin( 3062 (u_long)(vm_offset_t)ea, 3063 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3064 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3065 if ((bp->b_pages[i]->valid & mask) == mask) 3066 continue; 3067 if ((bp->b_pages[i]->valid & mask) == 0) { 3068 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3069 bzero(sa, ea - sa); 3070 } 3071 } else { 3072 for (; sa < ea; sa += DEV_BSIZE, j++) { 3073 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3074 (bp->b_pages[i]->valid & (1<<j)) == 0) 3075 bzero(sa, DEV_BSIZE); 3076 } 3077 } 3078 bp->b_pages[i]->valid |= mask; 3079 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3080 } 3081 bp->b_resid = 0; 3082 } else { 3083 clrbuf(bp); 3084 } 3085} 3086 3087/* 3088 * vm_hold_load_pages and vm_hold_unload pages get pages into 3089 * a buffers address space. The pages are anonymous and are 3090 * not associated with a file object. 3091 */ 3092void 3093vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3094{ 3095 vm_offset_t pg; 3096 vm_page_t p; 3097 int index; 3098 3099 to = round_page(to); 3100 from = round_page(from); 3101 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3102 3103 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3104 3105tryagain: 3106 3107 p = vm_page_alloc(kernel_object, 3108 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3109 VM_ALLOC_NORMAL); 3110 if (!p) { 3111 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3112 VM_WAIT; 3113 goto tryagain; 3114 } 3115 vm_page_wire(p); 3116 p->valid = VM_PAGE_BITS_ALL; 3117 vm_page_flag_clear(p, PG_ZERO); 3118 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 3119 bp->b_pages[index] = p; 3120 vm_page_wakeup(p); 3121 } 3122 bp->b_npages = index; 3123} 3124 3125void 3126vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3127{ 3128 vm_offset_t pg; 3129 vm_page_t p; 3130 int index, newnpages; 3131 3132 from = round_page(from); 3133 to = round_page(to); 3134 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3135 3136 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3137 p = bp->b_pages[index]; 3138 if (p && (index < bp->b_npages)) { 3139 if (p->busy) { 3140 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 3141 bp->b_blkno, bp->b_lblkno); 3142 } 3143 bp->b_pages[index] = NULL; 3144 pmap_kremove(pg); 3145 vm_page_busy(p); 3146 vm_page_unwire(p, 0); 3147 vm_page_free(p); 3148 } 3149 } 3150 bp->b_npages = newnpages; 3151} 3152 3153 3154#include "opt_ddb.h" 3155#ifdef DDB 3156#include <ddb/ddb.h> 3157 3158DB_SHOW_COMMAND(buffer, db_show_buffer) 3159{ 3160 /* get args */ 3161 struct buf *bp = (struct buf *)addr; 3162 3163 if (!have_addr) { 3164 db_printf("usage: show buffer <addr>\n"); 3165 return; 3166 } 3167 3168 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3169 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 3170 "b_resid = %ld\nb_dev = (%d,%d), b_data = %p, " 3171 "b_blkno = %d, b_pblkno = %d\n", 3172 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3173 major(bp->b_dev), minor(bp->b_dev), 3174 bp->b_data, bp->b_blkno, bp->b_pblkno); 3175 if (bp->b_npages) { 3176 int i; 3177 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3178 for (i = 0; i < bp->b_npages; i++) { 3179 vm_page_t m; 3180 m = bp->b_pages[i]; 3181 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3182 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3183 if ((i + 1) < bp->b_npages) 3184 db_printf(","); 3185 } 3186 db_printf("\n"); 3187 } 3188} 3189#endif /* DDB */ 3190