vfs_bio.c revision 53212
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 53212 1999-11-16 10:56:05Z 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/kernel.h> 33#include <sys/sysctl.h> 34#include <sys/proc.h> 35#include <sys/kthread.h> 36#include <sys/vnode.h> 37#include <sys/vmmeter.h> 38#include <sys/lock.h> 39#include <vm/vm.h> 40#include <vm/vm_param.h> 41#include <vm/vm_kern.h> 42#include <vm/vm_pageout.h> 43#include <vm/vm_page.h> 44#include <vm/vm_object.h> 45#include <vm/vm_extern.h> 46#include <vm/vm_map.h> 47#include <sys/buf.h> 48#include <sys/mount.h> 49#include <sys/malloc.h> 50#include <sys/resourcevar.h> 51#include <sys/conf.h> 52 53static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 54 55struct bio_ops bioops; /* I/O operation notification */ 56 57struct buf *buf; /* buffer header pool */ 58struct swqueue bswlist; 59 60static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 61 vm_offset_t to); 62static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 63 vm_offset_t to); 64static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 65 int pageno, vm_page_t m); 66static void vfs_clean_pages(struct buf * bp); 67static void vfs_setdirty(struct buf *bp); 68static void vfs_vmio_release(struct buf *bp); 69static int flushbufqueues(void); 70 71static int bd_request; 72 73static void buf_daemon __P((void)); 74/* 75 * bogus page -- for I/O to/from partially complete buffers 76 * this is a temporary solution to the problem, but it is not 77 * really that bad. it would be better to split the buffer 78 * for input in the case of buffers partially already in memory, 79 * but the code is intricate enough already. 80 */ 81vm_page_t bogus_page; 82int runningbufspace; 83int vmiodirenable = FALSE; 84int buf_maxio = DFLTPHYS; 85static vm_offset_t bogus_offset; 86 87static int bufspace, maxbufspace, vmiospace, 88 bufmallocspace, maxbufmallocspace, hibufspace; 89static int maxbdrun; 90static int needsbuffer; 91static int numdirtybuffers, lodirtybuffers, hidirtybuffers; 92static int numfreebuffers, lofreebuffers, hifreebuffers; 93static int getnewbufcalls; 94static int getnewbufrestarts; 95static int kvafreespace; 96 97SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, 98 &numdirtybuffers, 0, ""); 99SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, 100 &lodirtybuffers, 0, ""); 101SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, 102 &hidirtybuffers, 0, ""); 103SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, 104 &numfreebuffers, 0, ""); 105SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, 106 &lofreebuffers, 0, ""); 107SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, 108 &hifreebuffers, 0, ""); 109SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, 110 &runningbufspace, 0, ""); 111SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RW, 112 &maxbufspace, 0, ""); 113SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, 114 &hibufspace, 0, ""); 115SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, 116 &bufspace, 0, ""); 117SYSCTL_INT(_vfs, OID_AUTO, maxbdrun, CTLFLAG_RW, 118 &maxbdrun, 0, ""); 119SYSCTL_INT(_vfs, OID_AUTO, vmiospace, CTLFLAG_RD, 120 &vmiospace, 0, ""); 121SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, 122 &maxbufmallocspace, 0, ""); 123SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, 124 &bufmallocspace, 0, ""); 125SYSCTL_INT(_vfs, OID_AUTO, kvafreespace, CTLFLAG_RD, 126 &kvafreespace, 0, ""); 127SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, 128 &getnewbufcalls, 0, ""); 129SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, 130 &getnewbufrestarts, 0, ""); 131SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, 132 &vmiodirenable, 0, ""); 133 134 135static int bufhashmask; 136static LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash; 137struct bqueues bufqueues[BUFFER_QUEUES] = { { 0 } }; 138char *buf_wmesg = BUF_WMESG; 139 140extern int vm_swap_size; 141 142#define BUF_MAXUSE 24 143 144#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 145#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 146#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 147#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 148#define VFS_BIO_NEED_KVASPACE 0x10 /* wait for buffer_map space, emerg */ 149 150/* 151 * Buffer hash table code. Note that the logical block scans linearly, which 152 * gives us some L1 cache locality. 153 */ 154 155static __inline 156struct bufhashhdr * 157bufhash(struct vnode *vnp, daddr_t bn) 158{ 159 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]); 160} 161 162/* 163 * kvaspacewakeup: 164 * 165 * Called when kva space is potential available for recovery or when 166 * kva space is recovered in the buffer_map. This function wakes up 167 * anyone waiting for buffer_map kva space. Even though the buffer_map 168 * is larger then maxbufspace, this situation will typically occur 169 * when the buffer_map gets fragmented. 170 */ 171 172static __inline void 173kvaspacewakeup(void) 174{ 175 /* 176 * If someone is waiting for KVA space, wake them up. Even 177 * though we haven't freed the kva space yet, the waiting 178 * process will be able to now. 179 */ 180 if (needsbuffer & VFS_BIO_NEED_KVASPACE) { 181 needsbuffer &= ~VFS_BIO_NEED_KVASPACE; 182 wakeup(&needsbuffer); 183 } 184} 185 186/* 187 * numdirtywakeup: 188 * 189 * If someone is blocked due to there being too many dirty buffers, 190 * and numdirtybuffers is now reasonable, wake them up. 191 */ 192 193static __inline void 194numdirtywakeup(void) 195{ 196 if (numdirtybuffers < hidirtybuffers) { 197 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 198 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 199 wakeup(&needsbuffer); 200 } 201 } 202} 203 204/* 205 * bufspacewakeup: 206 * 207 * Called when buffer space is potentially available for recovery or when 208 * buffer space is recovered. getnewbuf() will block on this flag when 209 * it is unable to free sufficient buffer space. Buffer space becomes 210 * recoverable when bp's get placed back in the queues. 211 */ 212 213static __inline void 214bufspacewakeup(void) 215{ 216 /* 217 * If someone is waiting for BUF space, wake them up. Even 218 * though we haven't freed the kva space yet, the waiting 219 * process will be able to now. 220 */ 221 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 222 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 223 wakeup(&needsbuffer); 224 } 225} 226 227/* 228 * bufcountwakeup: 229 * 230 * Called when a buffer has been added to one of the free queues to 231 * account for the buffer and to wakeup anyone waiting for free buffers. 232 * This typically occurs when large amounts of metadata are being handled 233 * by the buffer cache ( else buffer space runs out first, usually ). 234 */ 235 236static __inline void 237bufcountwakeup(void) 238{ 239 ++numfreebuffers; 240 if (needsbuffer) { 241 needsbuffer &= ~VFS_BIO_NEED_ANY; 242 if (numfreebuffers >= hifreebuffers) 243 needsbuffer &= ~VFS_BIO_NEED_FREE; 244 wakeup(&needsbuffer); 245 } 246} 247 248/* 249 * vfs_buf_test_cache: 250 * 251 * Called when a buffer is extended. This function clears the B_CACHE 252 * bit if the newly extended portion of the buffer does not contain 253 * valid data. 254 */ 255static __inline__ 256void 257vfs_buf_test_cache(struct buf *bp, 258 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 259 vm_page_t m) 260{ 261 if (bp->b_flags & B_CACHE) { 262 int base = (foff + off) & PAGE_MASK; 263 if (vm_page_is_valid(m, base, size) == 0) 264 bp->b_flags &= ~B_CACHE; 265 } 266} 267 268static __inline__ 269void 270bd_wakeup(int dirtybuflevel) 271{ 272 if (numdirtybuffers >= dirtybuflevel && bd_request == 0) { 273 bd_request = 1; 274 wakeup(&bd_request); 275 } 276} 277 278 279/* 280 * Initialize buffer headers and related structures. 281 */ 282 283caddr_t 284bufhashinit(caddr_t vaddr) 285{ 286 /* first, make a null hash table */ 287 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 288 ; 289 bufhashtbl = (void *)vaddr; 290 vaddr = vaddr + sizeof(*bufhashtbl) * bufhashmask; 291 --bufhashmask; 292 return(vaddr); 293} 294 295void 296bufinit(void) 297{ 298 struct buf *bp; 299 int i; 300 301 TAILQ_INIT(&bswlist); 302 LIST_INIT(&invalhash); 303 simple_lock_init(&buftimelock); 304 305 for (i = 0; i <= bufhashmask; i++) 306 LIST_INIT(&bufhashtbl[i]); 307 308 /* next, make a null set of free lists */ 309 for (i = 0; i < BUFFER_QUEUES; i++) 310 TAILQ_INIT(&bufqueues[i]); 311 312 /* finally, initialize each buffer header and stick on empty q */ 313 for (i = 0; i < nbuf; i++) { 314 bp = &buf[i]; 315 bzero(bp, sizeof *bp); 316 bp->b_flags = B_INVAL; /* we're just an empty header */ 317 bp->b_dev = NODEV; 318 bp->b_rcred = NOCRED; 319 bp->b_wcred = NOCRED; 320 bp->b_qindex = QUEUE_EMPTY; 321 bp->b_xflags = 0; 322 LIST_INIT(&bp->b_dep); 323 BUF_LOCKINIT(bp); 324 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 325 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 326 } 327 328 /* 329 * maxbufspace is currently calculated to be maximally efficient 330 * when the filesystem block size is DFLTBSIZE or DFLTBSIZE*2 331 * (4K or 8K). To reduce the number of stall points our calculation 332 * is based on DFLTBSIZE which should reduce the chances of actually 333 * running out of buffer headers. The maxbufspace calculation is also 334 * based on DFLTBSIZE (4K) instead of BKVASIZE (8K) in order to 335 * reduce the chance that a KVA allocation will fail due to 336 * fragmentation. While this does not usually create a stall, 337 * the KVA map allocation/free functions are O(N) rather then O(1) 338 * so running them constantly would result in inefficient O(N*M) 339 * buffer cache operation. 340 */ 341 maxbufspace = (nbuf + 8) * DFLTBSIZE; 342 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 5); 343/* 344 * Limit the amount of malloc memory since it is wired permanently into 345 * the kernel space. Even though this is accounted for in the buffer 346 * allocation, we don't want the malloced region to grow uncontrolled. 347 * The malloc scheme improves memory utilization significantly on average 348 * (small) directories. 349 */ 350 maxbufmallocspace = hibufspace / 20; 351 352/* 353 * Reduce the chance of a deadlock occuring by limiting the number 354 * of delayed-write dirty buffers we allow to stack up. 355 */ 356 lodirtybuffers = nbuf / 7 + 10; 357 hidirtybuffers = nbuf / 4 + 20; 358 numdirtybuffers = 0; 359/* 360 * To support extreme low-memory systems, make sure hidirtybuffers cannot 361 * eat up all available buffer space. This occurs when our minimum cannot 362 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 363 * BKVASIZE'd (8K) buffers. We also reduce buf_maxio in this case (used 364 * by the clustering code) in an attempt to further reduce the load on 365 * the buffer cache. 366 */ 367 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 368 lodirtybuffers >>= 1; 369 hidirtybuffers >>= 1; 370 buf_maxio >>= 1; 371 } 372 if (lodirtybuffers < 2) { 373 lodirtybuffers = 2; 374 hidirtybuffers = 4; 375 } 376 377 /* 378 * Temporary, BKVASIZE may be manipulated soon, make sure we don't 379 * do something illegal. XXX 380 */ 381#if BKVASIZE < MAXBSIZE 382 if (buf_maxio < BKVASIZE * 2) 383 buf_maxio = BKVASIZE * 2; 384#else 385 if (buf_maxio < MAXBSIZE) 386 buf_maxio = MAXBSIZE; 387#endif 388 389/* 390 * Try to keep the number of free buffers in the specified range, 391 * and give the syncer access to an emergency reserve. 392 */ 393 lofreebuffers = nbuf / 18 + 5; 394 hifreebuffers = 2 * lofreebuffers; 395 numfreebuffers = nbuf; 396 397/* 398 * Maximum number of async ops initiated per buf_daemon loop. This is 399 * somewhat of a hack at the moment, we really need to limit ourselves 400 * based on the number of bytes of I/O in-transit that were initiated 401 * from buf_daemon. 402 */ 403 if ((maxbdrun = nswbuf / 4) < 4) 404 maxbdrun = 4; 405 406 kvafreespace = 0; 407 408 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 409 bogus_page = vm_page_alloc(kernel_object, 410 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 411 VM_ALLOC_NORMAL); 412 cnt.v_wire_count++; 413 414} 415 416/* 417 * Free the kva allocation for a buffer 418 * Must be called only at splbio or higher, 419 * as this is the only locking for buffer_map. 420 */ 421static void 422bfreekva(struct buf * bp) 423{ 424 if (bp->b_kvasize) { 425 vm_map_delete(buffer_map, 426 (vm_offset_t) bp->b_kvabase, 427 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 428 ); 429 bp->b_kvasize = 0; 430 kvaspacewakeup(); 431 } 432} 433 434/* 435 * bremfree: 436 * 437 * Remove the buffer from the appropriate free list. 438 */ 439void 440bremfree(struct buf * bp) 441{ 442 int s = splbio(); 443 int old_qindex = bp->b_qindex; 444 445 if (bp->b_qindex != QUEUE_NONE) { 446 if (bp->b_qindex == QUEUE_EMPTYKVA) { 447 kvafreespace -= bp->b_kvasize; 448 } 449 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 450 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 451 bp->b_qindex = QUEUE_NONE; 452 runningbufspace += bp->b_bufsize; 453 } else { 454#if !defined(MAX_PERF) 455 if (BUF_REFCNT(bp) <= 1) 456 panic("bremfree: removing a buffer not on a queue"); 457#endif 458 } 459 460 /* 461 * Fixup numfreebuffers count. If the buffer is invalid or not 462 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 463 * the buffer was free and we must decrement numfreebuffers. 464 */ 465 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 466 switch(old_qindex) { 467 case QUEUE_DIRTY: 468 case QUEUE_CLEAN: 469 case QUEUE_EMPTY: 470 case QUEUE_EMPTYKVA: 471 --numfreebuffers; 472 break; 473 default: 474 break; 475 } 476 } 477 splx(s); 478} 479 480 481/* 482 * Get a buffer with the specified data. Look in the cache first. We 483 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 484 * is set, the buffer is valid and we do not have to do anything ( see 485 * getblk() ). 486 */ 487int 488bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 489 struct buf ** bpp) 490{ 491 struct buf *bp; 492 493 bp = getblk(vp, blkno, size, 0, 0); 494 *bpp = bp; 495 496 /* if not found in cache, do some I/O */ 497 if ((bp->b_flags & B_CACHE) == 0) { 498 if (curproc != NULL) 499 curproc->p_stats->p_ru.ru_inblock++; 500 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp)); 501 bp->b_flags |= B_READ; 502 bp->b_flags &= ~(B_ERROR | B_INVAL); 503 if (bp->b_rcred == NOCRED) { 504 if (cred != NOCRED) 505 crhold(cred); 506 bp->b_rcred = cred; 507 } 508 vfs_busy_pages(bp, 0); 509 VOP_STRATEGY(vp, bp); 510 return (biowait(bp)); 511 } 512 return (0); 513} 514 515/* 516 * Operates like bread, but also starts asynchronous I/O on 517 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior 518 * to initiating I/O . If B_CACHE is set, the buffer is valid 519 * and we do not have to do anything. 520 */ 521int 522breadn(struct vnode * vp, daddr_t blkno, int size, 523 daddr_t * rablkno, int *rabsize, 524 int cnt, struct ucred * cred, struct buf ** bpp) 525{ 526 struct buf *bp, *rabp; 527 int i; 528 int rv = 0, readwait = 0; 529 530 *bpp = bp = getblk(vp, blkno, size, 0, 0); 531 532 /* if not found in cache, do some I/O */ 533 if ((bp->b_flags & B_CACHE) == 0) { 534 if (curproc != NULL) 535 curproc->p_stats->p_ru.ru_inblock++; 536 bp->b_flags |= B_READ; 537 bp->b_flags &= ~(B_ERROR | B_INVAL); 538 if (bp->b_rcred == NOCRED) { 539 if (cred != NOCRED) 540 crhold(cred); 541 bp->b_rcred = cred; 542 } 543 vfs_busy_pages(bp, 0); 544 VOP_STRATEGY(vp, bp); 545 ++readwait; 546 } 547 548 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 549 if (inmem(vp, *rablkno)) 550 continue; 551 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 552 553 if ((rabp->b_flags & B_CACHE) == 0) { 554 if (curproc != NULL) 555 curproc->p_stats->p_ru.ru_inblock++; 556 rabp->b_flags |= B_READ | B_ASYNC; 557 rabp->b_flags &= ~(B_ERROR | B_INVAL); 558 if (rabp->b_rcred == NOCRED) { 559 if (cred != NOCRED) 560 crhold(cred); 561 rabp->b_rcred = cred; 562 } 563 vfs_busy_pages(rabp, 0); 564 BUF_KERNPROC(rabp); 565 VOP_STRATEGY(vp, rabp); 566 } else { 567 brelse(rabp); 568 } 569 } 570 571 if (readwait) { 572 rv = biowait(bp); 573 } 574 return (rv); 575} 576 577/* 578 * Write, release buffer on completion. (Done by iodone 579 * if async). Do not bother writing anything if the buffer 580 * is invalid. 581 * 582 * Note that we set B_CACHE here, indicating that buffer is 583 * fully valid and thus cacheable. This is true even of NFS 584 * now so we set it generally. This could be set either here 585 * or in biodone() since the I/O is synchronous. We put it 586 * here. 587 */ 588int 589bwrite(struct buf * bp) 590{ 591 int oldflags, s; 592 struct vnode *vp; 593 struct mount *mp; 594 595 if (bp->b_flags & B_INVAL) { 596 brelse(bp); 597 return (0); 598 } 599 600 oldflags = bp->b_flags; 601 602#if !defined(MAX_PERF) 603 if (BUF_REFCNT(bp) == 0) 604 panic("bwrite: buffer is not busy???"); 605#endif 606 s = splbio(); 607 bundirty(bp); 608 609 bp->b_flags &= ~(B_READ | B_DONE | B_ERROR); 610 bp->b_flags |= B_WRITEINPROG | B_CACHE; 611 612 bp->b_vp->v_numoutput++; 613 vfs_busy_pages(bp, 1); 614 if (curproc != NULL) 615 curproc->p_stats->p_ru.ru_oublock++; 616 splx(s); 617 if (oldflags & B_ASYNC) 618 BUF_KERNPROC(bp); 619 VOP_STRATEGY(bp->b_vp, bp); 620 621 /* 622 * Collect statistics on synchronous and asynchronous writes. 623 * Writes to block devices are charged to their associated 624 * filesystem (if any). 625 */ 626 if ((vp = bp->b_vp) != NULL) { 627 if (vp->v_type == VBLK) 628 mp = vp->v_specmountpoint; 629 else 630 mp = vp->v_mount; 631 if (mp != NULL) { 632 if ((oldflags & B_ASYNC) == 0) 633 mp->mnt_stat.f_syncwrites++; 634 else 635 mp->mnt_stat.f_asyncwrites++; 636 } 637 } 638 639 if ((oldflags & B_ASYNC) == 0) { 640 int rtval = biowait(bp); 641 brelse(bp); 642 return (rtval); 643 } 644 645 return (0); 646} 647 648/* 649 * Delayed write. (Buffer is marked dirty). Do not bother writing 650 * anything if the buffer is marked invalid. 651 * 652 * Note that since the buffer must be completely valid, we can safely 653 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 654 * biodone() in order to prevent getblk from writing the buffer 655 * out synchronously. 656 */ 657void 658bdwrite(struct buf * bp) 659{ 660#if !defined(MAX_PERF) 661 if (BUF_REFCNT(bp) == 0) 662 panic("bdwrite: buffer is not busy"); 663#endif 664 665 if (bp->b_flags & B_INVAL) { 666 brelse(bp); 667 return; 668 } 669 bdirty(bp); 670 671 /* 672 * Set B_CACHE, indicating that the buffer is fully valid. This is 673 * true even of NFS now. 674 */ 675 bp->b_flags |= B_CACHE; 676 677 /* 678 * This bmap keeps the system from needing to do the bmap later, 679 * perhaps when the system is attempting to do a sync. Since it 680 * is likely that the indirect block -- or whatever other datastructure 681 * that the filesystem needs is still in memory now, it is a good 682 * thing to do this. Note also, that if the pageout daemon is 683 * requesting a sync -- there might not be enough memory to do 684 * the bmap then... So, this is important to do. 685 */ 686 if (bp->b_lblkno == bp->b_blkno) { 687 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 688 } 689 690 /* 691 * Set the *dirty* buffer range based upon the VM system dirty pages. 692 */ 693 vfs_setdirty(bp); 694 695 /* 696 * We need to do this here to satisfy the vnode_pager and the 697 * pageout daemon, so that it thinks that the pages have been 698 * "cleaned". Note that since the pages are in a delayed write 699 * buffer -- the VFS layer "will" see that the pages get written 700 * out on the next sync, or perhaps the cluster will be completed. 701 */ 702 vfs_clean_pages(bp); 703 bqrelse(bp); 704 705 /* 706 * Wakeup the buffer flushing daemon if we have saturated the 707 * buffer cache. 708 */ 709 710 bd_wakeup(hidirtybuffers); 711 712 /* 713 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 714 * due to the softdep code. 715 */ 716} 717 718/* 719 * bdirty: 720 * 721 * Turn buffer into delayed write request. We must clear B_READ and 722 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 723 * itself to properly update it in the dirty/clean lists. We mark it 724 * B_DONE to ensure that any asynchronization of the buffer properly 725 * clears B_DONE ( else a panic will occur later ). 726 * 727 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 728 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 729 * should only be called if the buffer is known-good. 730 * 731 * Since the buffer is not on a queue, we do not update the numfreebuffers 732 * count. 733 * 734 * Must be called at splbio(). 735 * The buffer must be on QUEUE_NONE. 736 */ 737void 738bdirty(bp) 739 struct buf *bp; 740{ 741 KASSERT(bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 742 bp->b_flags &= ~(B_READ|B_RELBUF); 743 744 if ((bp->b_flags & B_DELWRI) == 0) { 745 bp->b_flags |= B_DONE | B_DELWRI; 746 reassignbuf(bp, bp->b_vp); 747 ++numdirtybuffers; 748 bd_wakeup(hidirtybuffers); 749 } 750} 751 752/* 753 * bundirty: 754 * 755 * Clear B_DELWRI for buffer. 756 * 757 * Since the buffer is not on a queue, we do not update the numfreebuffers 758 * count. 759 * 760 * Must be called at splbio(). 761 * The buffer must be on QUEUE_NONE. 762 */ 763 764void 765bundirty(bp) 766 struct buf *bp; 767{ 768 KASSERT(bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 769 770 if (bp->b_flags & B_DELWRI) { 771 bp->b_flags &= ~B_DELWRI; 772 reassignbuf(bp, bp->b_vp); 773 --numdirtybuffers; 774 numdirtywakeup(); 775 } 776} 777 778/* 779 * bawrite: 780 * 781 * Asynchronous write. Start output on a buffer, but do not wait for 782 * it to complete. The buffer is released when the output completes. 783 * 784 * bwrite() ( or the VOP routine anyway ) is responsible for handling 785 * B_INVAL buffers. Not us. 786 */ 787void 788bawrite(struct buf * bp) 789{ 790 bp->b_flags |= B_ASYNC; 791 (void) VOP_BWRITE(bp->b_vp, bp); 792} 793 794/* 795 * bowrite: 796 * 797 * Ordered write. Start output on a buffer, and flag it so that the 798 * device will write it in the order it was queued. The buffer is 799 * released when the output completes. bwrite() ( or the VOP routine 800 * anyway ) is responsible for handling B_INVAL buffers. 801 */ 802int 803bowrite(struct buf * bp) 804{ 805 bp->b_flags |= B_ORDERED | B_ASYNC; 806 return (VOP_BWRITE(bp->b_vp, bp)); 807} 808 809/* 810 * bwillwrite: 811 * 812 * Called prior to the locking of any vnodes when we are expecting to 813 * write. We do not want to starve the buffer cache with too many 814 * dirty buffers so we block here. By blocking prior to the locking 815 * of any vnodes we attempt to avoid the situation where a locked vnode 816 * prevents the various system daemons from flushing related buffers. 817 */ 818 819void 820bwillwrite(void) 821{ 822 int twenty = (hidirtybuffers - lodirtybuffers) / 5; 823 824 if (numdirtybuffers > hidirtybuffers + twenty) { 825 int s; 826 827 s = splbio(); 828 while (numdirtybuffers > hidirtybuffers) { 829 bd_wakeup(hidirtybuffers); 830 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 831 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0); 832 } 833 splx(s); 834 } 835} 836 837/* 838 * brelse: 839 * 840 * Release a busy buffer and, if requested, free its resources. The 841 * buffer will be stashed in the appropriate bufqueue[] allowing it 842 * to be accessed later as a cache entity or reused for other purposes. 843 */ 844void 845brelse(struct buf * bp) 846{ 847 int s; 848 int kvawakeup = 0; 849 850 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 851 852 s = splbio(); 853 854 if (bp->b_flags & B_LOCKED) 855 bp->b_flags &= ~B_ERROR; 856 857 if ((bp->b_flags & (B_READ | B_ERROR | B_INVAL)) == B_ERROR) { 858 /* 859 * Failed write, redirty. Must clear B_ERROR to prevent 860 * pages from being scrapped. If B_INVAL is set then 861 * this case is not run and the next case is run to 862 * destroy the buffer. B_INVAL can occur if the buffer 863 * is outside the range supported by the underlying device. 864 */ 865 bp->b_flags &= ~B_ERROR; 866 bdirty(bp); 867 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_FREEBUF)) || 868 (bp->b_bufsize <= 0)) { 869 /* 870 * Either a failed I/O or we were asked to free or not 871 * cache the buffer. 872 */ 873 bp->b_flags |= B_INVAL; 874 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 875 (*bioops.io_deallocate)(bp); 876 if (bp->b_flags & B_DELWRI) { 877 --numdirtybuffers; 878 numdirtywakeup(); 879 } 880 bp->b_flags &= ~(B_DELWRI | B_CACHE | B_FREEBUF); 881 if ((bp->b_flags & B_VMIO) == 0) { 882 if (bp->b_bufsize) 883 allocbuf(bp, 0); 884 if (bp->b_vp) 885 brelvp(bp); 886 } 887 } 888 889 /* 890 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 891 * is called with B_DELWRI set, the underlying pages may wind up 892 * getting freed causing a previous write (bdwrite()) to get 'lost' 893 * because pages associated with a B_DELWRI bp are marked clean. 894 * 895 * We still allow the B_INVAL case to call vfs_vmio_release(), even 896 * if B_DELWRI is set. 897 */ 898 899 if (bp->b_flags & B_DELWRI) 900 bp->b_flags &= ~B_RELBUF; 901 902 /* 903 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 904 * constituted, not even NFS buffers now. Two flags effect this. If 905 * B_INVAL, the struct buf is invalidated but the VM object is kept 906 * around ( i.e. so it is trivial to reconstitute the buffer later ). 907 * 908 * If B_ERROR or B_NOCACHE is set, pages in the VM object will be 909 * invalidated. B_ERROR cannot be set for a failed write unless the 910 * buffer is also B_INVAL because it hits the re-dirtying code above. 911 * 912 * Normally we can do this whether a buffer is B_DELWRI or not. If 913 * the buffer is an NFS buffer, it is tracking piecemeal writes or 914 * the commit state and we cannot afford to lose the buffer. 915 */ 916 if ((bp->b_flags & B_VMIO) 917 && !(bp->b_vp->v_tag == VT_NFS && 918 bp->b_vp->v_type != VBLK && 919 (bp->b_flags & B_DELWRI)) 920 ) { 921 922 int i, j, resid; 923 vm_page_t m; 924 off_t foff; 925 vm_pindex_t poff; 926 vm_object_t obj; 927 struct vnode *vp; 928 929 vp = bp->b_vp; 930 931 /* 932 * Get the base offset and length of the buffer. Note that 933 * for block sizes that are less then PAGE_SIZE, the b_data 934 * base of the buffer does not represent exactly b_offset and 935 * neither b_offset nor b_size are necessarily page aligned. 936 * Instead, the starting position of b_offset is: 937 * 938 * b_data + (b_offset & PAGE_MASK) 939 * 940 * block sizes less then DEV_BSIZE (usually 512) are not 941 * supported due to the page granularity bits (m->valid, 942 * m->dirty, etc...). 943 * 944 * See man buf(9) for more information 945 */ 946 947 resid = bp->b_bufsize; 948 foff = bp->b_offset; 949 950 for (i = 0; i < bp->b_npages; i++) { 951 m = bp->b_pages[i]; 952 vm_page_flag_clear(m, PG_ZERO); 953 if (m == bogus_page) { 954 955 obj = (vm_object_t) vp->v_object; 956 poff = OFF_TO_IDX(bp->b_offset); 957 958 for (j = i; j < bp->b_npages; j++) { 959 m = bp->b_pages[j]; 960 if (m == bogus_page) { 961 m = vm_page_lookup(obj, poff + j); 962#if !defined(MAX_PERF) 963 if (!m) { 964 panic("brelse: page missing\n"); 965 } 966#endif 967 bp->b_pages[j] = m; 968 } 969 } 970 971 if ((bp->b_flags & B_INVAL) == 0) { 972 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 973 } 974 } 975 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 976 int poffset = foff & PAGE_MASK; 977 int presid = resid > (PAGE_SIZE - poffset) ? 978 (PAGE_SIZE - poffset) : resid; 979 980 KASSERT(presid >= 0, ("brelse: extra page")); 981 vm_page_set_invalid(m, poffset, presid); 982 } 983 resid -= PAGE_SIZE - (foff & PAGE_MASK); 984 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 985 } 986 987 if (bp->b_flags & (B_INVAL | B_RELBUF)) 988 vfs_vmio_release(bp); 989 990 } else if (bp->b_flags & B_VMIO) { 991 992 if (bp->b_flags & (B_INVAL | B_RELBUF)) 993 vfs_vmio_release(bp); 994 995 } 996 997#if !defined(MAX_PERF) 998 if (bp->b_qindex != QUEUE_NONE) 999 panic("brelse: free buffer onto another queue???"); 1000#endif 1001 if (BUF_REFCNT(bp) > 1) { 1002 /* Temporary panic to verify exclusive locking */ 1003 /* This panic goes away when we allow shared refs */ 1004 panic("brelse: multiple refs"); 1005 /* do not release to free list */ 1006 BUF_UNLOCK(bp); 1007 splx(s); 1008 return; 1009 } 1010 1011 /* enqueue */ 1012 1013 /* buffers with no memory */ 1014 if (bp->b_bufsize == 0) { 1015 bp->b_flags |= B_INVAL; 1016 if (bp->b_kvasize) { 1017 bp->b_qindex = QUEUE_EMPTYKVA; 1018 kvawakeup = 1; 1019 } else { 1020 bp->b_qindex = QUEUE_EMPTY; 1021 } 1022 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1023 LIST_REMOVE(bp, b_hash); 1024 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1025 bp->b_dev = NODEV; 1026 kvafreespace += bp->b_kvasize; 1027 /* buffers with junk contents */ 1028 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) { 1029 bp->b_flags |= B_INVAL; 1030 bp->b_qindex = QUEUE_CLEAN; 1031 if (bp->b_kvasize) 1032 kvawakeup = 1; 1033 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1034 LIST_REMOVE(bp, b_hash); 1035 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1036 bp->b_dev = NODEV; 1037 1038 /* buffers that are locked */ 1039 } else if (bp->b_flags & B_LOCKED) { 1040 bp->b_qindex = QUEUE_LOCKED; 1041 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1042 1043 /* remaining buffers */ 1044 } else { 1045 switch(bp->b_flags & (B_DELWRI|B_AGE)) { 1046 case B_DELWRI | B_AGE: 1047 bp->b_qindex = QUEUE_DIRTY; 1048 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1049 break; 1050 case B_DELWRI: 1051 bp->b_qindex = QUEUE_DIRTY; 1052 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1053 break; 1054 case B_AGE: 1055 bp->b_qindex = QUEUE_CLEAN; 1056 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1057 if (bp->b_kvasize) 1058 kvawakeup = 1; 1059 break; 1060 default: 1061 bp->b_qindex = QUEUE_CLEAN; 1062 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1063 if (bp->b_kvasize) 1064 kvawakeup = 1; 1065 break; 1066 } 1067 } 1068 1069 /* 1070 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1071 * on the correct queue. 1072 */ 1073 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) { 1074 bp->b_flags &= ~B_DELWRI; 1075 --numdirtybuffers; 1076 numdirtywakeup(); 1077 } 1078 1079 runningbufspace -= bp->b_bufsize; 1080 1081 /* 1082 * Fixup numfreebuffers count. The bp is on an appropriate queue 1083 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1084 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1085 * if B_INVAL is set ). 1086 */ 1087 1088 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1089 bufcountwakeup(); 1090 1091 /* 1092 * Something we can maybe free. 1093 */ 1094 1095 if (bp->b_bufsize) 1096 bufspacewakeup(); 1097 if (kvawakeup) 1098 kvaspacewakeup(); 1099 1100 /* unlock */ 1101 BUF_UNLOCK(bp); 1102 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1103 splx(s); 1104} 1105 1106/* 1107 * Release a buffer back to the appropriate queue but do not try to free 1108 * it. 1109 * 1110 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1111 * biodone() to requeue an async I/O on completion. It is also used when 1112 * known good buffers need to be requeued but we think we may need the data 1113 * again soon. 1114 */ 1115void 1116bqrelse(struct buf * bp) 1117{ 1118 int s; 1119 1120 s = splbio(); 1121 1122 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1123 1124#if !defined(MAX_PERF) 1125 if (bp->b_qindex != QUEUE_NONE) 1126 panic("bqrelse: free buffer onto another queue???"); 1127#endif 1128 if (BUF_REFCNT(bp) > 1) { 1129 /* do not release to free list */ 1130 panic("bqrelse: multiple refs"); 1131 BUF_UNLOCK(bp); 1132 splx(s); 1133 return; 1134 } 1135 if (bp->b_flags & B_LOCKED) { 1136 bp->b_flags &= ~B_ERROR; 1137 bp->b_qindex = QUEUE_LOCKED; 1138 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1139 /* buffers with stale but valid contents */ 1140 } else if (bp->b_flags & B_DELWRI) { 1141 bp->b_qindex = QUEUE_DIRTY; 1142 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1143 } else { 1144 bp->b_qindex = QUEUE_CLEAN; 1145 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1146 } 1147 1148 runningbufspace -= bp->b_bufsize; 1149 1150 if ((bp->b_flags & B_LOCKED) == 0 && 1151 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1152 bufcountwakeup(); 1153 } 1154 1155 /* 1156 * Something we can maybe wakeup 1157 */ 1158 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1159 bufspacewakeup(); 1160 1161 /* unlock */ 1162 BUF_UNLOCK(bp); 1163 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1164 splx(s); 1165} 1166 1167static void 1168vfs_vmio_release(bp) 1169 struct buf *bp; 1170{ 1171 int i, s; 1172 vm_page_t m; 1173 1174 s = splvm(); 1175 for (i = 0; i < bp->b_npages; i++) { 1176 m = bp->b_pages[i]; 1177 bp->b_pages[i] = NULL; 1178 /* 1179 * In order to keep page LRU ordering consistent, put 1180 * everything on the inactive queue. 1181 */ 1182 vm_page_unwire(m, 0); 1183 /* 1184 * We don't mess with busy pages, it is 1185 * the responsibility of the process that 1186 * busied the pages to deal with them. 1187 */ 1188 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1189 continue; 1190 1191 if (m->wire_count == 0) { 1192 vm_page_flag_clear(m, PG_ZERO); 1193 /* 1194 * Might as well free the page if we can and it has 1195 * no valid data. 1196 */ 1197 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { 1198 vm_page_busy(m); 1199 vm_page_protect(m, VM_PROT_NONE); 1200 vm_page_free(m); 1201 } 1202 } 1203 } 1204 bufspace -= bp->b_bufsize; 1205 vmiospace -= bp->b_bufsize; 1206 runningbufspace -= bp->b_bufsize; 1207 splx(s); 1208 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1209 if (bp->b_bufsize) 1210 bufspacewakeup(); 1211 bp->b_npages = 0; 1212 bp->b_bufsize = 0; 1213 bp->b_flags &= ~B_VMIO; 1214 if (bp->b_vp) 1215 brelvp(bp); 1216} 1217 1218/* 1219 * Check to see if a block is currently memory resident. 1220 */ 1221struct buf * 1222gbincore(struct vnode * vp, daddr_t blkno) 1223{ 1224 struct buf *bp; 1225 struct bufhashhdr *bh; 1226 1227 bh = bufhash(vp, blkno); 1228 1229 /* Search hash chain */ 1230 LIST_FOREACH(bp, bh, b_hash) { 1231 /* hit */ 1232 if (bp->b_vp == vp && bp->b_lblkno == blkno && 1233 (bp->b_flags & B_INVAL) == 0) { 1234 break; 1235 } 1236 } 1237 return (bp); 1238} 1239 1240/* 1241 * vfs_bio_awrite: 1242 * 1243 * Implement clustered async writes for clearing out B_DELWRI buffers. 1244 * This is much better then the old way of writing only one buffer at 1245 * a time. Note that we may not be presented with the buffers in the 1246 * correct order, so we search for the cluster in both directions. 1247 */ 1248int 1249vfs_bio_awrite(struct buf * bp) 1250{ 1251 int i; 1252 int j; 1253 daddr_t lblkno = bp->b_lblkno; 1254 struct vnode *vp = bp->b_vp; 1255 int s; 1256 int ncl; 1257 struct buf *bpa; 1258 int nwritten; 1259 int size; 1260 int maxcl; 1261 1262 s = splbio(); 1263 /* 1264 * right now we support clustered writing only to regular files. If 1265 * we find a clusterable block we could be in the middle of a cluster 1266 * rather then at the beginning. 1267 */ 1268 if ((vp->v_type == VREG) && 1269 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1270 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1271 1272 size = vp->v_mount->mnt_stat.f_iosize; 1273 maxcl = MAXPHYS / size; 1274 1275 for (i = 1; i < maxcl; i++) { 1276 if ((bpa = gbincore(vp, lblkno + i)) && 1277 BUF_REFCNT(bpa) == 0 && 1278 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1279 (B_DELWRI | B_CLUSTEROK)) && 1280 (bpa->b_bufsize == size)) { 1281 if ((bpa->b_blkno == bpa->b_lblkno) || 1282 (bpa->b_blkno != 1283 bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 1284 break; 1285 } else { 1286 break; 1287 } 1288 } 1289 for (j = 1; i + j <= maxcl && j <= lblkno; j++) { 1290 if ((bpa = gbincore(vp, lblkno - j)) && 1291 BUF_REFCNT(bpa) == 0 && 1292 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1293 (B_DELWRI | B_CLUSTEROK)) && 1294 (bpa->b_bufsize == size)) { 1295 if ((bpa->b_blkno == bpa->b_lblkno) || 1296 (bpa->b_blkno != 1297 bp->b_blkno - ((j * size) >> DEV_BSHIFT))) 1298 break; 1299 } else { 1300 break; 1301 } 1302 } 1303 --j; 1304 ncl = i + j; 1305 /* 1306 * this is a possible cluster write 1307 */ 1308 if (ncl != 1) { 1309 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1310 splx(s); 1311 return nwritten; 1312 } 1313 } 1314 1315 BUF_LOCK(bp, LK_EXCLUSIVE); 1316 bremfree(bp); 1317 bp->b_flags |= B_ASYNC; 1318 1319 splx(s); 1320 /* 1321 * default (old) behavior, writing out only one block 1322 * 1323 * XXX returns b_bufsize instead of b_bcount for nwritten? 1324 */ 1325 nwritten = bp->b_bufsize; 1326 (void) VOP_BWRITE(bp->b_vp, bp); 1327 1328 return nwritten; 1329} 1330 1331/* 1332 * getnewbuf: 1333 * 1334 * Find and initialize a new buffer header, freeing up existing buffers 1335 * in the bufqueues as necessary. The new buffer is returned locked. 1336 * 1337 * Important: B_INVAL is not set. If the caller wishes to throw the 1338 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1339 * 1340 * We block if: 1341 * We have insufficient buffer headers 1342 * We have insufficient buffer space 1343 * buffer_map is too fragmented ( space reservation fails ) 1344 * If we have to flush dirty buffers ( but we try to avoid this ) 1345 * 1346 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1347 * Instead we ask the buf daemon to do it for us. We attempt to 1348 * avoid piecemeal wakeups of the pageout daemon. 1349 */ 1350 1351static struct buf * 1352getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1353{ 1354 struct buf *bp; 1355 struct buf *nbp; 1356 struct buf *dbp; 1357 int outofspace; 1358 int nqindex; 1359 int defrag = 0; 1360 1361 ++getnewbufcalls; 1362 --getnewbufrestarts; 1363restart: 1364 ++getnewbufrestarts; 1365 1366 /* 1367 * Calculate whether we are out of buffer space. This state is 1368 * recalculated on every restart. If we are out of space, we 1369 * have to turn off defragmentation. Setting defrag to -1 when 1370 * outofspace is positive means "defrag while freeing buffers". 1371 * The looping conditional will be muffed up if defrag is left 1372 * positive when outofspace is positive. 1373 */ 1374 1375 dbp = NULL; 1376 outofspace = 0; 1377 if (bufspace >= hibufspace) { 1378 if ((curproc && (curproc->p_flag & P_BUFEXHAUST) == 0) || 1379 bufspace >= maxbufspace) { 1380 outofspace = 1; 1381 if (defrag > 0) 1382 defrag = -1; 1383 } 1384 } 1385 1386 /* 1387 * defrag state is semi-persistant. 1 means we are flagged for 1388 * defragging. -1 means we actually defragged something. 1389 */ 1390 /* nop */ 1391 1392 /* 1393 * Setup for scan. If we do not have enough free buffers, 1394 * we setup a degenerate case that immediately fails. Note 1395 * that if we are specially marked process, we are allowed to 1396 * dip into our reserves. 1397 * 1398 * Normally we want to find an EMPTYKVA buffer. That is, a 1399 * buffer with kva already allocated. If there are no EMPTYKVA 1400 * buffers we back up to the truely EMPTY buffers. When defragging 1401 * we do not bother backing up since we have to locate buffers with 1402 * kva to defrag. If we are out of space we skip both EMPTY and 1403 * EMPTYKVA and dig right into the CLEAN queue. 1404 * 1405 * In this manner we avoid scanning unnecessary buffers. It is very 1406 * important for us to do this because the buffer cache is almost 1407 * constantly out of space or in need of defragmentation. 1408 */ 1409 1410 if (curproc && (curproc->p_flag & P_BUFEXHAUST) == 0 && 1411 numfreebuffers < lofreebuffers) { 1412 nqindex = QUEUE_CLEAN; 1413 nbp = NULL; 1414 } else { 1415 nqindex = QUEUE_EMPTYKVA; 1416 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1417 if (nbp == NULL) { 1418 if (defrag <= 0) { 1419 nqindex = QUEUE_EMPTY; 1420 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1421 } 1422 } 1423 if (outofspace || nbp == NULL) { 1424 nqindex = QUEUE_CLEAN; 1425 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1426 } 1427 } 1428 1429 /* 1430 * Run scan, possibly freeing data and/or kva mappings on the fly 1431 * depending. 1432 */ 1433 1434 while ((bp = nbp) != NULL) { 1435 int qindex = nqindex; 1436 1437 /* 1438 * Calculate next bp ( we can only use it if we do not block 1439 * or do other fancy things ). 1440 */ 1441 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1442 switch(qindex) { 1443 case QUEUE_EMPTY: 1444 nqindex = QUEUE_EMPTYKVA; 1445 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1446 break; 1447 /* fall through */ 1448 case QUEUE_EMPTYKVA: 1449 nqindex = QUEUE_CLEAN; 1450 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1451 break; 1452 /* fall through */ 1453 case QUEUE_CLEAN: 1454 /* 1455 * nbp is NULL. 1456 */ 1457 break; 1458 } 1459 } 1460 1461 /* 1462 * Sanity Checks 1463 */ 1464 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1465 1466 /* 1467 * Note: we no longer distinguish between VMIO and non-VMIO 1468 * buffers. 1469 */ 1470 1471 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1472 1473 /* 1474 * If we are defragging and the buffer isn't useful for fixing 1475 * that problem we continue. If we are out of space and the 1476 * buffer isn't useful for fixing that problem we continue. 1477 */ 1478 1479 if (defrag > 0 && bp->b_kvasize == 0) 1480 continue; 1481 if (outofspace > 0 && bp->b_bufsize == 0) 1482 continue; 1483 1484 /* 1485 * Start freeing the bp. This is somewhat involved. nbp 1486 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1487 */ 1488 1489 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1490 panic("getnewbuf: locked buf"); 1491 bremfree(bp); 1492 1493 if (qindex == QUEUE_CLEAN) { 1494 if (bp->b_flags & B_VMIO) { 1495 bp->b_flags &= ~B_ASYNC; 1496 vfs_vmio_release(bp); 1497 } 1498 if (bp->b_vp) 1499 brelvp(bp); 1500 } 1501 1502 /* 1503 * NOTE: nbp is now entirely invalid. We can only restart 1504 * the scan from this point on. 1505 * 1506 * Get the rest of the buffer freed up. b_kva* is still 1507 * valid after this operation. 1508 */ 1509 1510 if (bp->b_rcred != NOCRED) { 1511 crfree(bp->b_rcred); 1512 bp->b_rcred = NOCRED; 1513 } 1514 if (bp->b_wcred != NOCRED) { 1515 crfree(bp->b_wcred); 1516 bp->b_wcred = NOCRED; 1517 } 1518 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 1519 (*bioops.io_deallocate)(bp); 1520 LIST_REMOVE(bp, b_hash); 1521 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1522 1523 if (bp->b_bufsize) 1524 allocbuf(bp, 0); 1525 1526 bp->b_flags = 0; 1527 bp->b_dev = NODEV; 1528 bp->b_vp = NULL; 1529 bp->b_blkno = bp->b_lblkno = 0; 1530 bp->b_offset = NOOFFSET; 1531 bp->b_iodone = 0; 1532 bp->b_error = 0; 1533 bp->b_resid = 0; 1534 bp->b_bcount = 0; 1535 bp->b_npages = 0; 1536 bp->b_dirtyoff = bp->b_dirtyend = 0; 1537 1538 LIST_INIT(&bp->b_dep); 1539 1540 /* 1541 * Ok, now that we have a free buffer, if we are defragging 1542 * we have to recover the kvaspace. If we are out of space 1543 * we have to free the buffer (which we just did), but we 1544 * do not have to recover kva space unless we hit a defrag 1545 * hicup. Being able to avoid freeing the kva space leads 1546 * to a significant reduction in overhead. 1547 */ 1548 1549 if (defrag > 0) { 1550 defrag = -1; 1551 bp->b_flags |= B_INVAL; 1552 bfreekva(bp); 1553 brelse(bp); 1554 goto restart; 1555 } 1556 1557 if (outofspace > 0) { 1558 outofspace = -1; 1559 bp->b_flags |= B_INVAL; 1560 if (defrag < 0) 1561 bfreekva(bp); 1562 brelse(bp); 1563 goto restart; 1564 } 1565 1566 /* 1567 * We are done 1568 */ 1569 break; 1570 } 1571 1572 /* 1573 * If we exhausted our list, sleep as appropriate. We may have to 1574 * wakeup various daemons and write out some dirty buffers. 1575 * 1576 * Generally we are sleeping due to insufficient buffer space. 1577 */ 1578 1579 if (bp == NULL) { 1580 int flags; 1581 char *waitmsg; 1582 1583 if (defrag > 0) { 1584 flags = VFS_BIO_NEED_KVASPACE; 1585 waitmsg = "nbufkv"; 1586 } else if (outofspace > 0) { 1587 waitmsg = "nbufbs"; 1588 flags = VFS_BIO_NEED_BUFSPACE; 1589 } else { 1590 waitmsg = "newbuf"; 1591 flags = VFS_BIO_NEED_ANY; 1592 } 1593 1594 /* XXX */ 1595 1596 (void) speedup_syncer(); 1597 needsbuffer |= flags; 1598 while (needsbuffer & flags) { 1599 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, 1600 waitmsg, slptimeo)) 1601 return (NULL); 1602 } 1603 } else { 1604 /* 1605 * We finally have a valid bp. We aren't quite out of the 1606 * woods, we still have to reserve kva space. 1607 */ 1608 vm_offset_t addr = 0; 1609 1610 maxsize = (maxsize + PAGE_MASK) & ~PAGE_MASK; 1611 1612 if (maxsize != bp->b_kvasize) { 1613 bfreekva(bp); 1614 1615 if (vm_map_findspace(buffer_map, 1616 vm_map_min(buffer_map), maxsize, &addr)) { 1617 /* 1618 * Uh oh. Buffer map is to fragmented. Try 1619 * to defragment. 1620 */ 1621 if (defrag <= 0) { 1622 defrag = 1; 1623 bp->b_flags |= B_INVAL; 1624 brelse(bp); 1625 goto restart; 1626 } 1627 /* 1628 * Uh oh. We couldn't seem to defragment 1629 */ 1630 panic("getnewbuf: unreachable code reached"); 1631 } 1632 } 1633 if (addr) { 1634 vm_map_insert(buffer_map, NULL, 0, 1635 addr, addr + maxsize, 1636 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1637 1638 bp->b_kvabase = (caddr_t) addr; 1639 bp->b_kvasize = maxsize; 1640 } 1641 bp->b_data = bp->b_kvabase; 1642 } 1643 return(bp); 1644} 1645 1646/* 1647 * waitfreebuffers: 1648 * 1649 * Wait for sufficient free buffers. Only called from normal processes. 1650 */ 1651 1652static void 1653waitfreebuffers(int slpflag, int slptimeo) 1654{ 1655 while (numfreebuffers < hifreebuffers) { 1656 if (numfreebuffers >= hifreebuffers) 1657 break; 1658 needsbuffer |= VFS_BIO_NEED_FREE; 1659 if (tsleep(&needsbuffer, (PRIBIO + 4)|slpflag, "biofre", slptimeo)) 1660 break; 1661 } 1662} 1663 1664/* 1665 * buf_daemon: 1666 * 1667 * buffer flushing daemon. Buffers are normally flushed by the 1668 * update daemon but if it cannot keep up this process starts to 1669 * take the load in an attempt to prevent getnewbuf() from blocking. 1670 */ 1671 1672static struct proc *bufdaemonproc; 1673static int bd_interval; 1674static int bd_flushto; 1675 1676static struct kproc_desc buf_kp = { 1677 "bufdaemon", 1678 buf_daemon, 1679 &bufdaemonproc 1680}; 1681SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1682 1683static void 1684buf_daemon() 1685{ 1686 int s; 1687 /* 1688 * This process is allowed to take the buffer cache to the limit 1689 */ 1690 curproc->p_flag |= P_BUFEXHAUST; 1691 s = splbio(); 1692 1693 bd_interval = 5 * hz; /* dynamically adjusted */ 1694 bd_flushto = hidirtybuffers; /* dynamically adjusted */ 1695 1696 while (TRUE) { 1697 bd_request = 0; 1698 1699 /* 1700 * Do the flush. Limit the number of buffers we flush in one 1701 * go. The failure condition occurs when processes are writing 1702 * buffers faster then we can dispose of them. In this case 1703 * we may be flushing so often that the previous set of flushes 1704 * have not had time to complete, causing us to run out of 1705 * physical buffers and block. 1706 */ 1707 { 1708 int runcount = maxbdrun; 1709 1710 while (numdirtybuffers > bd_flushto && runcount) { 1711 --runcount; 1712 if (flushbufqueues() == 0) 1713 break; 1714 } 1715 } 1716 1717 /* 1718 * If nobody is requesting anything we sleep 1719 */ 1720 if (bd_request == 0) 1721 tsleep(&bd_request, PVM, "psleep", bd_interval); 1722 1723 /* 1724 * We calculate how much to add or subtract from bd_flushto 1725 * and bd_interval based on how far off we are from the 1726 * optimal number of dirty buffers, which is 20% below the 1727 * hidirtybuffers mark. We cannot use hidirtybuffers straight 1728 * because being right on the mark will cause getnewbuf() 1729 * to oscillate our wakeup. 1730 * 1731 * The larger the error in either direction, the more we adjust 1732 * bd_flushto and bd_interval. The time interval is adjusted 1733 * by 2 seconds per whole-buffer-range of error. This is an 1734 * exponential convergence algorithm, with large errors 1735 * producing large changes and small errors producing small 1736 * changes. 1737 */ 1738 1739 { 1740 int brange = hidirtybuffers - lodirtybuffers; 1741 int middb = hidirtybuffers - brange / 5; 1742 int deltabuf = middb - numdirtybuffers; 1743 1744 bd_flushto += deltabuf / 20; 1745 bd_interval += deltabuf * (2 * hz) / (brange * 1); 1746 } 1747 if (bd_flushto < lodirtybuffers) 1748 bd_flushto = lodirtybuffers; 1749 if (bd_flushto > hidirtybuffers) 1750 bd_flushto = hidirtybuffers; 1751 if (bd_interval < hz / 10) 1752 bd_interval = hz / 10; 1753 if (bd_interval > 5 * hz) 1754 bd_interval = 5 * hz; 1755 } 1756} 1757 1758/* 1759 * flushbufqueues: 1760 * 1761 * Try to flush a buffer in the dirty queue. We must be careful to 1762 * free up B_INVAL buffers instead of write them, which NFS is 1763 * particularly sensitive to. 1764 */ 1765 1766static int 1767flushbufqueues(void) 1768{ 1769 struct buf *bp; 1770 int r = 0; 1771 1772 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1773 1774 while (bp) { 1775 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); 1776 if ((bp->b_flags & B_DELWRI) != 0) { 1777 if (bp->b_flags & B_INVAL) { 1778 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1779 panic("flushbufqueues: locked buf"); 1780 bremfree(bp); 1781 brelse(bp); 1782 ++r; 1783 break; 1784 } 1785 vfs_bio_awrite(bp); 1786 ++r; 1787 break; 1788 } 1789 bp = TAILQ_NEXT(bp, b_freelist); 1790 } 1791 return(r); 1792} 1793 1794/* 1795 * Check to see if a block is currently memory resident. 1796 */ 1797struct buf * 1798incore(struct vnode * vp, daddr_t blkno) 1799{ 1800 struct buf *bp; 1801 1802 int s = splbio(); 1803 bp = gbincore(vp, blkno); 1804 splx(s); 1805 return (bp); 1806} 1807 1808/* 1809 * Returns true if no I/O is needed to access the 1810 * associated VM object. This is like incore except 1811 * it also hunts around in the VM system for the data. 1812 */ 1813 1814int 1815inmem(struct vnode * vp, daddr_t blkno) 1816{ 1817 vm_object_t obj; 1818 vm_offset_t toff, tinc, size; 1819 vm_page_t m; 1820 vm_ooffset_t off; 1821 1822 if (incore(vp, blkno)) 1823 return 1; 1824 if (vp->v_mount == NULL) 1825 return 0; 1826 if ((vp->v_object == NULL) || (vp->v_flag & VOBJBUF) == 0) 1827 return 0; 1828 1829 obj = vp->v_object; 1830 size = PAGE_SIZE; 1831 if (size > vp->v_mount->mnt_stat.f_iosize) 1832 size = vp->v_mount->mnt_stat.f_iosize; 1833 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 1834 1835 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1836 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 1837 if (!m) 1838 return 0; 1839 tinc = size; 1840 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 1841 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 1842 if (vm_page_is_valid(m, 1843 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 1844 return 0; 1845 } 1846 return 1; 1847} 1848 1849/* 1850 * vfs_setdirty: 1851 * 1852 * Sets the dirty range for a buffer based on the status of the dirty 1853 * bits in the pages comprising the buffer. 1854 * 1855 * The range is limited to the size of the buffer. 1856 * 1857 * This routine is primarily used by NFS, but is generalized for the 1858 * B_VMIO case. 1859 */ 1860static void 1861vfs_setdirty(struct buf *bp) 1862{ 1863 int i; 1864 vm_object_t object; 1865 1866 /* 1867 * Degenerate case - empty buffer 1868 */ 1869 1870 if (bp->b_bufsize == 0) 1871 return; 1872 1873 /* 1874 * We qualify the scan for modified pages on whether the 1875 * object has been flushed yet. The OBJ_WRITEABLE flag 1876 * is not cleared simply by protecting pages off. 1877 */ 1878 1879 if ((bp->b_flags & B_VMIO) == 0) 1880 return; 1881 1882 object = bp->b_pages[0]->object; 1883 1884 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 1885 printf("Warning: object %p writeable but not mightbedirty\n", object); 1886 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 1887 printf("Warning: object %p mightbedirty but not writeable\n", object); 1888 1889 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 1890 vm_offset_t boffset; 1891 vm_offset_t eoffset; 1892 1893 /* 1894 * test the pages to see if they have been modified directly 1895 * by users through the VM system. 1896 */ 1897 for (i = 0; i < bp->b_npages; i++) { 1898 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 1899 vm_page_test_dirty(bp->b_pages[i]); 1900 } 1901 1902 /* 1903 * Calculate the encompassing dirty range, boffset and eoffset, 1904 * (eoffset - boffset) bytes. 1905 */ 1906 1907 for (i = 0; i < bp->b_npages; i++) { 1908 if (bp->b_pages[i]->dirty) 1909 break; 1910 } 1911 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 1912 1913 for (i = bp->b_npages - 1; i >= 0; --i) { 1914 if (bp->b_pages[i]->dirty) { 1915 break; 1916 } 1917 } 1918 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 1919 1920 /* 1921 * Fit it to the buffer. 1922 */ 1923 1924 if (eoffset > bp->b_bcount) 1925 eoffset = bp->b_bcount; 1926 1927 /* 1928 * If we have a good dirty range, merge with the existing 1929 * dirty range. 1930 */ 1931 1932 if (boffset < eoffset) { 1933 if (bp->b_dirtyoff > boffset) 1934 bp->b_dirtyoff = boffset; 1935 if (bp->b_dirtyend < eoffset) 1936 bp->b_dirtyend = eoffset; 1937 } 1938 } 1939} 1940 1941/* 1942 * getblk: 1943 * 1944 * Get a block given a specified block and offset into a file/device. 1945 * The buffers B_DONE bit will be cleared on return, making it almost 1946 * ready for an I/O initiation. B_INVAL may or may not be set on 1947 * return. The caller should clear B_INVAL prior to initiating a 1948 * READ. 1949 * 1950 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 1951 * an existing buffer. 1952 * 1953 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 1954 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 1955 * and then cleared based on the backing VM. If the previous buffer is 1956 * non-0-sized but invalid, B_CACHE will be cleared. 1957 * 1958 * If getblk() must create a new buffer, the new buffer is returned with 1959 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 1960 * case it is returned with B_INVAL clear and B_CACHE set based on the 1961 * backing VM. 1962 * 1963 * getblk() also forces a VOP_BWRITE() for any B_DELWRI buffer whos 1964 * B_CACHE bit is clear. 1965 * 1966 * What this means, basically, is that the caller should use B_CACHE to 1967 * determine whether the buffer is fully valid or not and should clear 1968 * B_INVAL prior to issuing a read. If the caller intends to validate 1969 * the buffer by loading its data area with something, the caller needs 1970 * to clear B_INVAL. If the caller does this without issuing an I/O, 1971 * the caller should set B_CACHE ( as an optimization ), else the caller 1972 * should issue the I/O and biodone() will set B_CACHE if the I/O was 1973 * a write attempt or if it was a successfull read. If the caller 1974 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR 1975 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 1976 */ 1977struct buf * 1978getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 1979{ 1980 struct buf *bp; 1981 int s; 1982 struct bufhashhdr *bh; 1983 1984#if !defined(MAX_PERF) 1985 if (size > MAXBSIZE) 1986 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 1987#endif 1988 1989 s = splbio(); 1990loop: 1991 /* 1992 * Block if we are low on buffers. Certain processes are allowed 1993 * to completely exhaust the buffer cache. 1994 * 1995 * If this check ever becomes a bottleneck it may be better to 1996 * move it into the else, when gbincore() fails. At the moment 1997 * it isn't a problem. 1998 */ 1999 if (!curproc || (curproc->p_flag & P_BUFEXHAUST)) { 2000 if (numfreebuffers == 0) { 2001 if (!curproc) 2002 return NULL; 2003 needsbuffer |= VFS_BIO_NEED_ANY; 2004 tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, "newbuf", 2005 slptimeo); 2006 } 2007 } else if (numfreebuffers < lofreebuffers) { 2008 waitfreebuffers(slpflag, slptimeo); 2009 } 2010 2011 if ((bp = gbincore(vp, blkno))) { 2012 /* 2013 * Buffer is in-core. If the buffer is not busy, it must 2014 * be on a queue. 2015 */ 2016 2017 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2018 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, 2019 "getblk", slpflag, slptimeo) == ENOLCK) 2020 goto loop; 2021 splx(s); 2022 return (struct buf *) NULL; 2023 } 2024 2025 /* 2026 * The buffer is locked. B_CACHE is cleared if the buffer is 2027 * invalid. Ohterwise, for a non-VMIO buffer, B_CACHE is set 2028 * and for a VMIO buffer B_CACHE is adjusted according to the 2029 * backing VM cache. 2030 */ 2031 if (bp->b_flags & B_INVAL) 2032 bp->b_flags &= ~B_CACHE; 2033 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2034 bp->b_flags |= B_CACHE; 2035 bremfree(bp); 2036 2037 /* 2038 * check for size inconsistancies for non-VMIO case. 2039 */ 2040 2041 if (bp->b_bcount != size) { 2042 if ((bp->b_flags & B_VMIO) == 0 || 2043 (size > bp->b_kvasize)) { 2044 if (bp->b_flags & B_DELWRI) { 2045 bp->b_flags |= B_NOCACHE; 2046 VOP_BWRITE(bp->b_vp, bp); 2047 } else { 2048 if ((bp->b_flags & B_VMIO) && 2049 (LIST_FIRST(&bp->b_dep) == NULL)) { 2050 bp->b_flags |= B_RELBUF; 2051 brelse(bp); 2052 } else { 2053 bp->b_flags |= B_NOCACHE; 2054 VOP_BWRITE(bp->b_vp, bp); 2055 } 2056 } 2057 goto loop; 2058 } 2059 } 2060 2061 /* 2062 * If the size is inconsistant in the VMIO case, we can resize 2063 * the buffer. This might lead to B_CACHE getting set or 2064 * cleared. If the size has not changed, B_CACHE remains 2065 * unchanged from its previous state. 2066 */ 2067 2068 if (bp->b_bcount != size) 2069 allocbuf(bp, size); 2070 2071 KASSERT(bp->b_offset != NOOFFSET, 2072 ("getblk: no buffer offset")); 2073 2074 /* 2075 * A buffer with B_DELWRI set and B_CACHE clear must 2076 * be committed before we can return the buffer in 2077 * order to prevent the caller from issuing a read 2078 * ( due to B_CACHE not being set ) and overwriting 2079 * it. 2080 * 2081 * Most callers, including NFS and FFS, need this to 2082 * operate properly either because they assume they 2083 * can issue a read if B_CACHE is not set, or because 2084 * ( for example ) an uncached B_DELWRI might loop due 2085 * to softupdates re-dirtying the buffer. In the latter 2086 * case, B_CACHE is set after the first write completes, 2087 * preventing further loops. 2088 */ 2089 2090 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2091 VOP_BWRITE(bp->b_vp, bp); 2092 goto loop; 2093 } 2094 2095 splx(s); 2096 bp->b_flags &= ~B_DONE; 2097 } else { 2098 /* 2099 * Buffer is not in-core, create new buffer. The buffer 2100 * returned by getnewbuf() is locked. Note that the returned 2101 * buffer is also considered valid (not marked B_INVAL). 2102 */ 2103 int bsize, maxsize, vmio; 2104 off_t offset; 2105 2106 if (vp->v_type == VBLK) 2107 bsize = DEV_BSIZE; 2108 else if (vp->v_mountedhere) 2109 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2110 else if (vp->v_mount) 2111 bsize = vp->v_mount->mnt_stat.f_iosize; 2112 else 2113 bsize = size; 2114 2115 offset = (off_t)blkno * bsize; 2116 vmio = (vp->v_object != 0) && (vp->v_flag & VOBJBUF); 2117 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2118 maxsize = imax(maxsize, bsize); 2119 2120 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2121 if (slpflag || slptimeo) { 2122 splx(s); 2123 return NULL; 2124 } 2125 goto loop; 2126 } 2127 2128 /* 2129 * This code is used to make sure that a buffer is not 2130 * created while the getnewbuf routine is blocked. 2131 * This can be a problem whether the vnode is locked or not. 2132 * If the buffer is created out from under us, we have to 2133 * throw away the one we just created. There is now window 2134 * race because we are safely running at splbio() from the 2135 * point of the duplicate buffer creation through to here, 2136 * and we've locked the buffer. 2137 */ 2138 if (gbincore(vp, blkno)) { 2139 bp->b_flags |= B_INVAL; 2140 brelse(bp); 2141 goto loop; 2142 } 2143 2144 /* 2145 * Insert the buffer into the hash, so that it can 2146 * be found by incore. 2147 */ 2148 bp->b_blkno = bp->b_lblkno = blkno; 2149 bp->b_offset = offset; 2150 2151 bgetvp(vp, bp); 2152 LIST_REMOVE(bp, b_hash); 2153 bh = bufhash(vp, blkno); 2154 LIST_INSERT_HEAD(bh, bp, b_hash); 2155 2156 /* 2157 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2158 * buffer size starts out as 0, B_CACHE will be set by 2159 * allocbuf() for the VMIO case prior to it testing the 2160 * backing store for validity. 2161 */ 2162 2163 if (vmio) { 2164 bp->b_flags |= B_VMIO; 2165#if defined(VFS_BIO_DEBUG) 2166 if (vp->v_type != VREG && vp->v_type != VBLK) 2167 printf("getblk: vmioing file type %d???\n", vp->v_type); 2168#endif 2169 } else { 2170 bp->b_flags &= ~B_VMIO; 2171 } 2172 2173 allocbuf(bp, size); 2174 2175 splx(s); 2176 bp->b_flags &= ~B_DONE; 2177 } 2178 return (bp); 2179} 2180 2181/* 2182 * Get an empty, disassociated buffer of given size. The buffer is initially 2183 * set to B_INVAL. 2184 */ 2185struct buf * 2186geteblk(int size) 2187{ 2188 struct buf *bp; 2189 int s; 2190 2191 s = splbio(); 2192 while ((bp = getnewbuf(0, 0, size, MAXBSIZE)) == 0); 2193 splx(s); 2194 allocbuf(bp, size); 2195 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2196 return (bp); 2197} 2198 2199 2200/* 2201 * This code constitutes the buffer memory from either anonymous system 2202 * memory (in the case of non-VMIO operations) or from an associated 2203 * VM object (in the case of VMIO operations). This code is able to 2204 * resize a buffer up or down. 2205 * 2206 * Note that this code is tricky, and has many complications to resolve 2207 * deadlock or inconsistant data situations. Tread lightly!!! 2208 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2209 * the caller. Calling this code willy nilly can result in the loss of data. 2210 * 2211 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2212 * B_CACHE for the non-VMIO case. 2213 */ 2214 2215int 2216allocbuf(struct buf *bp, int size) 2217{ 2218 int newbsize, mbsize; 2219 int i; 2220 2221#if !defined(MAX_PERF) 2222 if (BUF_REFCNT(bp) == 0) 2223 panic("allocbuf: buffer not busy"); 2224 2225 if (bp->b_kvasize < size) 2226 panic("allocbuf: buffer too small"); 2227#endif 2228 2229 if ((bp->b_flags & B_VMIO) == 0) { 2230 caddr_t origbuf; 2231 int origbufsize; 2232 /* 2233 * Just get anonymous memory from the kernel. Don't 2234 * mess with B_CACHE. 2235 */ 2236 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2237#if !defined(NO_B_MALLOC) 2238 if (bp->b_flags & B_MALLOC) 2239 newbsize = mbsize; 2240 else 2241#endif 2242 newbsize = round_page(size); 2243 2244 if (newbsize < bp->b_bufsize) { 2245#if !defined(NO_B_MALLOC) 2246 /* 2247 * malloced buffers are not shrunk 2248 */ 2249 if (bp->b_flags & B_MALLOC) { 2250 if (newbsize) { 2251 bp->b_bcount = size; 2252 } else { 2253 free(bp->b_data, M_BIOBUF); 2254 bufspace -= bp->b_bufsize; 2255 bufmallocspace -= bp->b_bufsize; 2256 runningbufspace -= bp->b_bufsize; 2257 if (bp->b_bufsize) 2258 bufspacewakeup(); 2259 bp->b_data = bp->b_kvabase; 2260 bp->b_bufsize = 0; 2261 bp->b_bcount = 0; 2262 bp->b_flags &= ~B_MALLOC; 2263 } 2264 return 1; 2265 } 2266#endif 2267 vm_hold_free_pages( 2268 bp, 2269 (vm_offset_t) bp->b_data + newbsize, 2270 (vm_offset_t) bp->b_data + bp->b_bufsize); 2271 } else if (newbsize > bp->b_bufsize) { 2272#if !defined(NO_B_MALLOC) 2273 /* 2274 * We only use malloced memory on the first allocation. 2275 * and revert to page-allocated memory when the buffer 2276 * grows. 2277 */ 2278 if ( (bufmallocspace < maxbufmallocspace) && 2279 (bp->b_bufsize == 0) && 2280 (mbsize <= PAGE_SIZE/2)) { 2281 2282 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2283 bp->b_bufsize = mbsize; 2284 bp->b_bcount = size; 2285 bp->b_flags |= B_MALLOC; 2286 bufspace += mbsize; 2287 bufmallocspace += mbsize; 2288 runningbufspace += bp->b_bufsize; 2289 return 1; 2290 } 2291#endif 2292 origbuf = NULL; 2293 origbufsize = 0; 2294#if !defined(NO_B_MALLOC) 2295 /* 2296 * If the buffer is growing on its other-than-first allocation, 2297 * then we revert to the page-allocation scheme. 2298 */ 2299 if (bp->b_flags & B_MALLOC) { 2300 origbuf = bp->b_data; 2301 origbufsize = bp->b_bufsize; 2302 bp->b_data = bp->b_kvabase; 2303 bufspace -= bp->b_bufsize; 2304 bufmallocspace -= bp->b_bufsize; 2305 runningbufspace -= bp->b_bufsize; 2306 if (bp->b_bufsize) 2307 bufspacewakeup(); 2308 bp->b_bufsize = 0; 2309 bp->b_flags &= ~B_MALLOC; 2310 newbsize = round_page(newbsize); 2311 } 2312#endif 2313 vm_hold_load_pages( 2314 bp, 2315 (vm_offset_t) bp->b_data + bp->b_bufsize, 2316 (vm_offset_t) bp->b_data + newbsize); 2317#if !defined(NO_B_MALLOC) 2318 if (origbuf) { 2319 bcopy(origbuf, bp->b_data, origbufsize); 2320 free(origbuf, M_BIOBUF); 2321 } 2322#endif 2323 } 2324 } else { 2325 vm_page_t m; 2326 int desiredpages; 2327 2328 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2329 desiredpages = (size == 0) ? 0 : 2330 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2331 2332#if !defined(NO_B_MALLOC) 2333 if (bp->b_flags & B_MALLOC) 2334 panic("allocbuf: VMIO buffer can't be malloced"); 2335#endif 2336 /* 2337 * Set B_CACHE initially if buffer is 0 length or will become 2338 * 0-length. 2339 */ 2340 if (size == 0 || bp->b_bufsize == 0) 2341 bp->b_flags |= B_CACHE; 2342 2343 if (newbsize < bp->b_bufsize) { 2344 /* 2345 * DEV_BSIZE aligned new buffer size is less then the 2346 * DEV_BSIZE aligned existing buffer size. Figure out 2347 * if we have to remove any pages. 2348 */ 2349 if (desiredpages < bp->b_npages) { 2350 for (i = desiredpages; i < bp->b_npages; i++) { 2351 /* 2352 * the page is not freed here -- it 2353 * is the responsibility of 2354 * vnode_pager_setsize 2355 */ 2356 m = bp->b_pages[i]; 2357 KASSERT(m != bogus_page, 2358 ("allocbuf: bogus page found")); 2359 while (vm_page_sleep_busy(m, TRUE, "biodep")) 2360 ; 2361 2362 bp->b_pages[i] = NULL; 2363 vm_page_unwire(m, 0); 2364 } 2365 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2366 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2367 bp->b_npages = desiredpages; 2368 } 2369 } else if (size > bp->b_bcount) { 2370 /* 2371 * We are growing the buffer, possibly in a 2372 * byte-granular fashion. 2373 */ 2374 struct vnode *vp; 2375 vm_object_t obj; 2376 vm_offset_t toff; 2377 vm_offset_t tinc; 2378 2379 /* 2380 * Step 1, bring in the VM pages from the object, 2381 * allocating them if necessary. We must clear 2382 * B_CACHE if these pages are not valid for the 2383 * range covered by the buffer. 2384 */ 2385 2386 vp = bp->b_vp; 2387 obj = vp->v_object; 2388 2389 while (bp->b_npages < desiredpages) { 2390 vm_page_t m; 2391 vm_pindex_t pi; 2392 2393 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2394 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2395 m = vm_page_alloc(obj, pi, VM_ALLOC_NORMAL); 2396 if (m == NULL) { 2397 VM_WAIT; 2398 vm_pageout_deficit += desiredpages - bp->b_npages; 2399 } else { 2400 vm_page_wire(m); 2401 vm_page_wakeup(m); 2402 bp->b_flags &= ~B_CACHE; 2403 bp->b_pages[bp->b_npages] = m; 2404 ++bp->b_npages; 2405 } 2406 continue; 2407 } 2408 2409 /* 2410 * We found a page. If we have to sleep on it, 2411 * retry because it might have gotten freed out 2412 * from under us. 2413 * 2414 * We can only test PG_BUSY here. Blocking on 2415 * m->busy might lead to a deadlock: 2416 * 2417 * vm_fault->getpages->cluster_read->allocbuf 2418 * 2419 */ 2420 2421 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 2422 continue; 2423 2424 /* 2425 * We have a good page. Should we wakeup the 2426 * page daemon? 2427 */ 2428 if ((curproc != pageproc) && 2429 ((m->queue - m->pc) == PQ_CACHE) && 2430 ((cnt.v_free_count + cnt.v_cache_count) < 2431 (cnt.v_free_min + cnt.v_cache_min))) { 2432 pagedaemon_wakeup(); 2433 } 2434 vm_page_flag_clear(m, PG_ZERO); 2435 vm_page_wire(m); 2436 bp->b_pages[bp->b_npages] = m; 2437 ++bp->b_npages; 2438 } 2439 2440 /* 2441 * Step 2. We've loaded the pages into the buffer, 2442 * we have to figure out if we can still have B_CACHE 2443 * set. Note that B_CACHE is set according to the 2444 * byte-granular range ( bcount and size ), new the 2445 * aligned range ( newbsize ). 2446 * 2447 * The VM test is against m->valid, which is DEV_BSIZE 2448 * aligned. Needless to say, the validity of the data 2449 * needs to also be DEV_BSIZE aligned. Note that this 2450 * fails with NFS if the server or some other client 2451 * extends the file's EOF. If our buffer is resized, 2452 * B_CACHE may remain set! XXX 2453 */ 2454 2455 toff = bp->b_bcount; 2456 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2457 2458 while ((bp->b_flags & B_CACHE) && toff < size) { 2459 vm_pindex_t pi; 2460 2461 if (tinc > (size - toff)) 2462 tinc = size - toff; 2463 2464 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2465 PAGE_SHIFT; 2466 2467 vfs_buf_test_cache( 2468 bp, 2469 bp->b_offset, 2470 toff, 2471 tinc, 2472 bp->b_pages[pi] 2473 ); 2474 toff += tinc; 2475 tinc = PAGE_SIZE; 2476 } 2477 2478 /* 2479 * Step 3, fixup the KVM pmap. Remember that 2480 * bp->b_data is relative to bp->b_offset, but 2481 * bp->b_offset may be offset into the first page. 2482 */ 2483 2484 bp->b_data = (caddr_t) 2485 trunc_page((vm_offset_t)bp->b_data); 2486 pmap_qenter( 2487 (vm_offset_t)bp->b_data, 2488 bp->b_pages, 2489 bp->b_npages 2490 ); 2491 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2492 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2493 } 2494 } 2495 if (bp->b_flags & B_VMIO) 2496 vmiospace += (newbsize - bp->b_bufsize); 2497 bufspace += (newbsize - bp->b_bufsize); 2498 runningbufspace += (newbsize - bp->b_bufsize); 2499 if (newbsize < bp->b_bufsize) 2500 bufspacewakeup(); 2501 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2502 bp->b_bcount = size; /* requested buffer size */ 2503 return 1; 2504} 2505 2506/* 2507 * biowait: 2508 * 2509 * Wait for buffer I/O completion, returning error status. The buffer 2510 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR 2511 * error and cleared. 2512 */ 2513int 2514biowait(register struct buf * bp) 2515{ 2516 int s; 2517 2518 s = splbio(); 2519 while ((bp->b_flags & B_DONE) == 0) { 2520#if defined(NO_SCHEDULE_MODS) 2521 tsleep(bp, PRIBIO, "biowait", 0); 2522#else 2523 if (bp->b_flags & B_READ) 2524 tsleep(bp, PRIBIO, "biord", 0); 2525 else 2526 tsleep(bp, PRIBIO, "biowr", 0); 2527#endif 2528 } 2529 splx(s); 2530 if (bp->b_flags & B_EINTR) { 2531 bp->b_flags &= ~B_EINTR; 2532 return (EINTR); 2533 } 2534 if (bp->b_flags & B_ERROR) { 2535 return (bp->b_error ? bp->b_error : EIO); 2536 } else { 2537 return (0); 2538 } 2539} 2540 2541/* 2542 * biodone: 2543 * 2544 * Finish I/O on a buffer, optionally calling a completion function. 2545 * This is usually called from an interrupt so process blocking is 2546 * not allowed. 2547 * 2548 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2549 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2550 * assuming B_INVAL is clear. 2551 * 2552 * For the VMIO case, we set B_CACHE if the op was a read and no 2553 * read error occured, or if the op was a write. B_CACHE is never 2554 * set if the buffer is invalid or otherwise uncacheable. 2555 * 2556 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2557 * initiator to leave B_INVAL set to brelse the buffer out of existance 2558 * in the biodone routine. 2559 */ 2560void 2561biodone(register struct buf * bp) 2562{ 2563 int s; 2564 2565 s = splbio(); 2566 2567 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 2568 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 2569 2570 bp->b_flags |= B_DONE; 2571 2572 if (bp->b_flags & B_FREEBUF) { 2573 brelse(bp); 2574 splx(s); 2575 return; 2576 } 2577 2578 if ((bp->b_flags & B_READ) == 0) { 2579 vwakeup(bp); 2580 } 2581 2582 /* call optional completion function if requested */ 2583 if (bp->b_flags & B_CALL) { 2584 bp->b_flags &= ~B_CALL; 2585 (*bp->b_iodone) (bp); 2586 splx(s); 2587 return; 2588 } 2589 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 2590 (*bioops.io_complete)(bp); 2591 2592 if (bp->b_flags & B_VMIO) { 2593 int i, resid; 2594 vm_ooffset_t foff; 2595 vm_page_t m; 2596 vm_object_t obj; 2597 int iosize; 2598 struct vnode *vp = bp->b_vp; 2599 2600 obj = vp->v_object; 2601 2602#if defined(VFS_BIO_DEBUG) 2603 if (vp->v_usecount == 0) { 2604 panic("biodone: zero vnode ref count"); 2605 } 2606 2607 if (vp->v_object == NULL) { 2608 panic("biodone: missing VM object"); 2609 } 2610 2611 if ((vp->v_flag & VOBJBUF) == 0) { 2612 panic("biodone: vnode is not setup for merged cache"); 2613 } 2614#endif 2615 2616 foff = bp->b_offset; 2617 KASSERT(bp->b_offset != NOOFFSET, 2618 ("biodone: no buffer offset")); 2619 2620#if !defined(MAX_PERF) 2621 if (!obj) { 2622 panic("biodone: no object"); 2623 } 2624#endif 2625#if defined(VFS_BIO_DEBUG) 2626 if (obj->paging_in_progress < bp->b_npages) { 2627 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 2628 obj->paging_in_progress, bp->b_npages); 2629 } 2630#endif 2631 2632 /* 2633 * Set B_CACHE if the op was a normal read and no error 2634 * occured. B_CACHE is set for writes in the b*write() 2635 * routines. 2636 */ 2637 iosize = bp->b_bcount - bp->b_resid; 2638 if ((bp->b_flags & (B_READ|B_FREEBUF|B_INVAL|B_NOCACHE|B_ERROR)) == B_READ) { 2639 bp->b_flags |= B_CACHE; 2640 } 2641 2642 for (i = 0; i < bp->b_npages; i++) { 2643 int bogusflag = 0; 2644 m = bp->b_pages[i]; 2645 if (m == bogus_page) { 2646 bogusflag = 1; 2647 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2648 if (!m) { 2649#if defined(VFS_BIO_DEBUG) 2650 printf("biodone: page disappeared\n"); 2651#endif 2652 vm_object_pip_subtract(obj, 1); 2653 bp->b_flags &= ~B_CACHE; 2654 continue; 2655 } 2656 bp->b_pages[i] = m; 2657 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2658 } 2659#if defined(VFS_BIO_DEBUG) 2660 if (OFF_TO_IDX(foff) != m->pindex) { 2661 printf( 2662"biodone: foff(%lu)/m->pindex(%d) mismatch\n", 2663 (unsigned long)foff, m->pindex); 2664 } 2665#endif 2666 resid = IDX_TO_OFF(m->pindex + 1) - foff; 2667 if (resid > iosize) 2668 resid = iosize; 2669 2670 /* 2671 * In the write case, the valid and clean bits are 2672 * already changed correctly ( see bdwrite() ), so we 2673 * only need to do this here in the read case. 2674 */ 2675 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) { 2676 vfs_page_set_valid(bp, foff, i, m); 2677 } 2678 vm_page_flag_clear(m, PG_ZERO); 2679 2680 /* 2681 * when debugging new filesystems or buffer I/O methods, this 2682 * is the most common error that pops up. if you see this, you 2683 * have not set the page busy flag correctly!!! 2684 */ 2685 if (m->busy == 0) { 2686#if !defined(MAX_PERF) 2687 printf("biodone: page busy < 0, " 2688 "pindex: %d, foff: 0x(%x,%x), " 2689 "resid: %d, index: %d\n", 2690 (int) m->pindex, (int)(foff >> 32), 2691 (int) foff & 0xffffffff, resid, i); 2692#endif 2693 if (vp->v_type != VBLK) 2694#if !defined(MAX_PERF) 2695 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 2696 bp->b_vp->v_mount->mnt_stat.f_iosize, 2697 (int) bp->b_lblkno, 2698 bp->b_flags, bp->b_npages); 2699 else 2700 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 2701 (int) bp->b_lblkno, 2702 bp->b_flags, bp->b_npages); 2703 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2704 m->valid, m->dirty, m->wire_count); 2705#endif 2706 panic("biodone: page busy < 0\n"); 2707 } 2708 vm_page_io_finish(m); 2709 vm_object_pip_subtract(obj, 1); 2710 foff += resid; 2711 iosize -= resid; 2712 } 2713 if (obj) 2714 vm_object_pip_wakeupn(obj, 0); 2715 } 2716 /* 2717 * For asynchronous completions, release the buffer now. The brelse 2718 * will do a wakeup there if necessary - so no need to do a wakeup 2719 * here in the async case. The sync case always needs to do a wakeup. 2720 */ 2721 2722 if (bp->b_flags & B_ASYNC) { 2723 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0) 2724 brelse(bp); 2725 else 2726 bqrelse(bp); 2727 } else { 2728 wakeup(bp); 2729 } 2730 splx(s); 2731} 2732 2733/* 2734 * This routine is called in lieu of iodone in the case of 2735 * incomplete I/O. This keeps the busy status for pages 2736 * consistant. 2737 */ 2738void 2739vfs_unbusy_pages(struct buf * bp) 2740{ 2741 int i; 2742 2743 if (bp->b_flags & B_VMIO) { 2744 struct vnode *vp = bp->b_vp; 2745 vm_object_t obj = vp->v_object; 2746 2747 for (i = 0; i < bp->b_npages; i++) { 2748 vm_page_t m = bp->b_pages[i]; 2749 2750 if (m == bogus_page) { 2751 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 2752#if !defined(MAX_PERF) 2753 if (!m) { 2754 panic("vfs_unbusy_pages: page missing\n"); 2755 } 2756#endif 2757 bp->b_pages[i] = m; 2758 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2759 } 2760 vm_object_pip_subtract(obj, 1); 2761 vm_page_flag_clear(m, PG_ZERO); 2762 vm_page_io_finish(m); 2763 } 2764 vm_object_pip_wakeupn(obj, 0); 2765 } 2766} 2767 2768/* 2769 * vfs_page_set_valid: 2770 * 2771 * Set the valid bits in a page based on the supplied offset. The 2772 * range is restricted to the buffer's size. 2773 * 2774 * This routine is typically called after a read completes. 2775 */ 2776static void 2777vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 2778{ 2779 vm_ooffset_t soff, eoff; 2780 2781 /* 2782 * Start and end offsets in buffer. eoff - soff may not cross a 2783 * page boundry or cross the end of the buffer. The end of the 2784 * buffer, in this case, is our file EOF, not the allocation size 2785 * of the buffer. 2786 */ 2787 soff = off; 2788 eoff = (off + PAGE_SIZE) & ~PAGE_MASK; 2789 if (eoff > bp->b_offset + bp->b_bcount) 2790 eoff = bp->b_offset + bp->b_bcount; 2791 2792 /* 2793 * Set valid range. This is typically the entire buffer and thus the 2794 * entire page. 2795 */ 2796 if (eoff > soff) { 2797 vm_page_set_validclean( 2798 m, 2799 (vm_offset_t) (soff & PAGE_MASK), 2800 (vm_offset_t) (eoff - soff) 2801 ); 2802 } 2803} 2804 2805/* 2806 * This routine is called before a device strategy routine. 2807 * It is used to tell the VM system that paging I/O is in 2808 * progress, and treat the pages associated with the buffer 2809 * almost as being PG_BUSY. Also the object paging_in_progress 2810 * flag is handled to make sure that the object doesn't become 2811 * inconsistant. 2812 * 2813 * Since I/O has not been initiated yet, certain buffer flags 2814 * such as B_ERROR or B_INVAL may be in an inconsistant state 2815 * and should be ignored. 2816 */ 2817void 2818vfs_busy_pages(struct buf * bp, int clear_modify) 2819{ 2820 int i, bogus; 2821 2822 if (bp->b_flags & B_VMIO) { 2823 struct vnode *vp = bp->b_vp; 2824 vm_object_t obj = vp->v_object; 2825 vm_ooffset_t foff; 2826 2827 foff = bp->b_offset; 2828 KASSERT(bp->b_offset != NOOFFSET, 2829 ("vfs_busy_pages: no buffer offset")); 2830 vfs_setdirty(bp); 2831 2832retry: 2833 for (i = 0; i < bp->b_npages; i++) { 2834 vm_page_t m = bp->b_pages[i]; 2835 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 2836 goto retry; 2837 } 2838 2839 bogus = 0; 2840 for (i = 0; i < bp->b_npages; i++) { 2841 vm_page_t m = bp->b_pages[i]; 2842 2843 vm_page_flag_clear(m, PG_ZERO); 2844 if ((bp->b_flags & B_CLUSTER) == 0) { 2845 vm_object_pip_add(obj, 1); 2846 vm_page_io_start(m); 2847 } 2848 2849 /* 2850 * When readying a buffer for a read ( i.e 2851 * clear_modify == 0 ), it is important to do 2852 * bogus_page replacement for valid pages in 2853 * partially instantiated buffers. Partially 2854 * instantiated buffers can, in turn, occur when 2855 * reconstituting a buffer from its VM backing store 2856 * base. We only have to do this if B_CACHE is 2857 * clear ( which causes the I/O to occur in the 2858 * first place ). The replacement prevents the read 2859 * I/O from overwriting potentially dirty VM-backed 2860 * pages. XXX bogus page replacement is, uh, bogus. 2861 * It may not work properly with small-block devices. 2862 * We need to find a better way. 2863 */ 2864 2865 vm_page_protect(m, VM_PROT_NONE); 2866 if (clear_modify) 2867 vfs_page_set_valid(bp, foff, i, m); 2868 else if (m->valid == VM_PAGE_BITS_ALL && 2869 (bp->b_flags & B_CACHE) == 0) { 2870 bp->b_pages[i] = bogus_page; 2871 bogus++; 2872 } 2873 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 2874 } 2875 if (bogus) 2876 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2877 } 2878} 2879 2880/* 2881 * Tell the VM system that the pages associated with this buffer 2882 * are clean. This is used for delayed writes where the data is 2883 * going to go to disk eventually without additional VM intevention. 2884 * 2885 * Note that while we only really need to clean through to b_bcount, we 2886 * just go ahead and clean through to b_bufsize. 2887 */ 2888static void 2889vfs_clean_pages(struct buf * bp) 2890{ 2891 int i; 2892 2893 if (bp->b_flags & B_VMIO) { 2894 vm_ooffset_t foff; 2895 2896 foff = bp->b_offset; 2897 KASSERT(bp->b_offset != NOOFFSET, 2898 ("vfs_clean_pages: no buffer offset")); 2899 for (i = 0; i < bp->b_npages; i++) { 2900 vm_page_t m = bp->b_pages[i]; 2901 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~PAGE_MASK; 2902 vm_ooffset_t eoff = noff; 2903 2904 if (eoff > bp->b_offset + bp->b_bufsize) 2905 eoff = bp->b_offset + bp->b_bufsize; 2906 vfs_page_set_valid(bp, foff, i, m); 2907 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2908 foff = noff; 2909 } 2910 } 2911} 2912 2913/* 2914 * vfs_bio_set_validclean: 2915 * 2916 * Set the range within the buffer to valid and clean. The range is 2917 * relative to the beginning of the buffer, b_offset. Note that b_offset 2918 * itself may be offset from the beginning of the first page. 2919 */ 2920 2921void 2922vfs_bio_set_validclean(struct buf *bp, int base, int size) 2923{ 2924 if (bp->b_flags & B_VMIO) { 2925 int i; 2926 int n; 2927 2928 /* 2929 * Fixup base to be relative to beginning of first page. 2930 * Set initial n to be the maximum number of bytes in the 2931 * first page that can be validated. 2932 */ 2933 2934 base += (bp->b_offset & PAGE_MASK); 2935 n = PAGE_SIZE - (base & PAGE_MASK); 2936 2937 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 2938 vm_page_t m = bp->b_pages[i]; 2939 2940 if (n > size) 2941 n = size; 2942 2943 vm_page_set_validclean(m, base & PAGE_MASK, n); 2944 base += n; 2945 size -= n; 2946 n = PAGE_SIZE; 2947 } 2948 } 2949} 2950 2951/* 2952 * vfs_bio_clrbuf: 2953 * 2954 * clear a buffer. This routine essentially fakes an I/O, so we need 2955 * to clear B_ERROR and B_INVAL. 2956 * 2957 * Note that while we only theoretically need to clear through b_bcount, 2958 * we go ahead and clear through b_bufsize. 2959 */ 2960 2961void 2962vfs_bio_clrbuf(struct buf *bp) { 2963 int i, mask = 0; 2964 caddr_t sa, ea; 2965 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 2966 bp->b_flags &= ~(B_INVAL|B_ERROR); 2967 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 2968 (bp->b_offset & PAGE_MASK) == 0) { 2969 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 2970 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 2971 ((bp->b_pages[0]->valid & mask) != mask)) { 2972 bzero(bp->b_data, bp->b_bufsize); 2973 } 2974 bp->b_pages[0]->valid |= mask; 2975 bp->b_resid = 0; 2976 return; 2977 } 2978 ea = sa = bp->b_data; 2979 for(i=0;i<bp->b_npages;i++,sa=ea) { 2980 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 2981 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 2982 ea = (caddr_t)(vm_offset_t)ulmin( 2983 (u_long)(vm_offset_t)ea, 2984 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 2985 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 2986 if ((bp->b_pages[i]->valid & mask) == mask) 2987 continue; 2988 if ((bp->b_pages[i]->valid & mask) == 0) { 2989 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 2990 bzero(sa, ea - sa); 2991 } 2992 } else { 2993 for (; sa < ea; sa += DEV_BSIZE, j++) { 2994 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 2995 (bp->b_pages[i]->valid & (1<<j)) == 0) 2996 bzero(sa, DEV_BSIZE); 2997 } 2998 } 2999 bp->b_pages[i]->valid |= mask; 3000 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3001 } 3002 bp->b_resid = 0; 3003 } else { 3004 clrbuf(bp); 3005 } 3006} 3007 3008/* 3009 * vm_hold_load_pages and vm_hold_unload pages get pages into 3010 * a buffers address space. The pages are anonymous and are 3011 * not associated with a file object. 3012 */ 3013void 3014vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3015{ 3016 vm_offset_t pg; 3017 vm_page_t p; 3018 int index; 3019 3020 to = round_page(to); 3021 from = round_page(from); 3022 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3023 3024 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3025 3026tryagain: 3027 3028 p = vm_page_alloc(kernel_object, 3029 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3030 VM_ALLOC_NORMAL); 3031 if (!p) { 3032 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3033 VM_WAIT; 3034 goto tryagain; 3035 } 3036 vm_page_wire(p); 3037 p->valid = VM_PAGE_BITS_ALL; 3038 vm_page_flag_clear(p, PG_ZERO); 3039 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 3040 bp->b_pages[index] = p; 3041 vm_page_wakeup(p); 3042 } 3043 bp->b_npages = index; 3044} 3045 3046void 3047vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3048{ 3049 vm_offset_t pg; 3050 vm_page_t p; 3051 int index, newnpages; 3052 3053 from = round_page(from); 3054 to = round_page(to); 3055 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3056 3057 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3058 p = bp->b_pages[index]; 3059 if (p && (index < bp->b_npages)) { 3060#if !defined(MAX_PERF) 3061 if (p->busy) { 3062 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 3063 bp->b_blkno, bp->b_lblkno); 3064 } 3065#endif 3066 bp->b_pages[index] = NULL; 3067 pmap_kremove(pg); 3068 vm_page_busy(p); 3069 vm_page_unwire(p, 0); 3070 vm_page_free(p); 3071 } 3072 } 3073 bp->b_npages = newnpages; 3074} 3075 3076 3077#include "opt_ddb.h" 3078#ifdef DDB 3079#include <ddb/ddb.h> 3080 3081DB_SHOW_COMMAND(buffer, db_show_buffer) 3082{ 3083 /* get args */ 3084 struct buf *bp = (struct buf *)addr; 3085 3086 if (!have_addr) { 3087 db_printf("usage: show buffer <addr>\n"); 3088 return; 3089 } 3090 3091 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3092 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 3093 "b_resid = %ld\nb_dev = (%d,%d), b_data = %p, " 3094 "b_blkno = %d, b_pblkno = %d\n", 3095 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3096 major(bp->b_dev), minor(bp->b_dev), 3097 bp->b_data, bp->b_blkno, bp->b_pblkno); 3098 if (bp->b_npages) { 3099 int i; 3100 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3101 for (i = 0; i < bp->b_npages; i++) { 3102 vm_page_t m; 3103 m = bp->b_pages[i]; 3104 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3105 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3106 if ((i + 1) < bp->b_npages) 3107 db_printf(","); 3108 } 3109 db_printf("\n"); 3110 } 3111} 3112#endif /* DDB */ 3113