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 * $FreeBSD$ 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 57 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 58 59 struct bio_ops bioops; /* I/O operation notification */ 60 61 struct buf *buf; /* buffer header pool */ 62 struct swqueue bswlist; 63 struct mtx buftimelock; /* Interlock on setting prio and timo */ 64 65 static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 66 vm_offset_t to); 67 static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 68 vm_offset_t to); 69 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 70 int pageno, vm_page_t m); 71 static void vfs_clean_pages(struct buf * bp); 72 static void vfs_setdirty(struct buf *bp); 73 static void vfs_vmio_release(struct buf *bp); 74 static void vfs_backgroundwritedone(struct buf *bp); 75 static int flushbufqueues(void); 76 77 static int bd_request; 78 79 static 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 */ 87 vm_page_t bogus_page; 88 int vmiodirenable = FALSE; 89 int runningbufspace; 90 static vm_offset_t bogus_offset; 91 92 static int bufspace, maxbufspace, 93 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace; 94 static int bufreusecnt, bufdefragcnt, buffreekvacnt; 95 static int needsbuffer; 96 static int lorunningspace, hirunningspace, runningbufreq; 97 static int numdirtybuffers, lodirtybuffers, hidirtybuffers; 98 static int numfreebuffers, lofreebuffers, hifreebuffers; 99 static int getnewbufcalls; 100 static int getnewbufrestarts; 101 102 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, 103 &numdirtybuffers, 0, ""); 104 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, 105 &lodirtybuffers, 0, ""); 106 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, 107 &hidirtybuffers, 0, ""); 108 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, 109 &numfreebuffers, 0, ""); 110 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, 111 &lofreebuffers, 0, ""); 112 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, 113 &hifreebuffers, 0, ""); 114 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, 115 &runningbufspace, 0, ""); 116 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, 117 &lorunningspace, 0, ""); 118 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, 119 &hirunningspace, 0, ""); 120 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, 121 &maxbufspace, 0, ""); 122 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, 123 &hibufspace, 0, ""); 124 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, 125 &lobufspace, 0, ""); 126 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, 127 &bufspace, 0, ""); 128 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, 129 &maxbufmallocspace, 0, ""); 130 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, 131 &bufmallocspace, 0, ""); 132 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, 133 &getnewbufcalls, 0, ""); 134 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, 135 &getnewbufrestarts, 0, ""); 136 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, 137 &vmiodirenable, 0, ""); 138 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, 139 &bufdefragcnt, 0, ""); 140 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, 141 &buffreekvacnt, 0, ""); 142 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, 143 &bufreusecnt, 0, ""); 144 145 static int bufhashmask; 146 static LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash; 147 struct bqueues bufqueues[BUFFER_QUEUES] = { { 0 } }; 148 char *buf_wmesg = BUF_WMESG; 149 150 extern int vm_swap_size; 151 152 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 153 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 154 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 155 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 156 157 /* 158 * Buffer hash table code. Note that the logical block scans linearly, which 159 * gives us some L1 cache locality. 160 */ 161 162 static __inline 163 struct bufhashhdr * 164 bufhash(struct vnode *vnp, daddr_t bn) 165 { 166 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]); 167 } 168 169 /* 170 * numdirtywakeup: 171 * 172 * If someone is blocked due to there being too many dirty buffers, 173 * and numdirtybuffers is now reasonable, wake them up. 174 */ 175 176 static __inline void 177 numdirtywakeup(int level) 178 { 179 if (numdirtybuffers <= level) { 180 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 181 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 182 wakeup(&needsbuffer); 183 } 184 } 185 } 186 187 /* 188 * bufspacewakeup: 189 * 190 * Called when buffer space is potentially available for recovery. 191 * getnewbuf() will block on this flag when it is unable to free 192 * sufficient buffer space. Buffer space becomes recoverable when 193 * bp's get placed back in the queues. 194 */ 195 196 static __inline void 197 bufspacewakeup(void) 198 { 199 /* 200 * If someone is waiting for BUF space, wake them up. Even 201 * though we haven't freed the kva space yet, the waiting 202 * process will be able to now. 203 */ 204 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 205 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 206 wakeup(&needsbuffer); 207 } 208 } 209 210 /* 211 * runningbufwakeup() - in-progress I/O accounting. 212 * 213 */ 214 static __inline void 215 runningbufwakeup(struct buf *bp) 216 { 217 if (bp->b_runningbufspace) { 218 runningbufspace -= bp->b_runningbufspace; 219 bp->b_runningbufspace = 0; 220 if (runningbufreq && runningbufspace <= lorunningspace) { 221 runningbufreq = 0; 222 wakeup(&runningbufreq); 223 } 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 */ 235 236 static __inline void 237 bufcountwakeup(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 * waitrunningbufspace() 250 * 251 * runningbufspace is a measure of the amount of I/O currently 252 * running. This routine is used in async-write situations to 253 * prevent creating huge backups of pending writes to a device. 254 * Only asynchronous writes are governed by this function. 255 * 256 * Reads will adjust runningbufspace, but will not block based on it. 257 * The read load has a side effect of reducing the allowed write load. 258 * 259 * This does NOT turn an async write into a sync write. It waits 260 * for earlier writes to complete and generally returns before the 261 * caller's write has reached the device. 262 */ 263 static __inline void 264 waitrunningbufspace(void) 265 { 266 while (runningbufspace > hirunningspace) { 267 ++runningbufreq; 268 tsleep(&runningbufreq, PVM, "wdrain", 0); 269 } 270 } 271 272 273 /* 274 * vfs_buf_test_cache: 275 * 276 * Called when a buffer is extended. This function clears the B_CACHE 277 * bit if the newly extended portion of the buffer does not contain 278 * valid data. 279 */ 280 static __inline__ 281 void 282 vfs_buf_test_cache(struct buf *bp, 283 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 284 vm_page_t m) 285 { 286 if (bp->b_flags & B_CACHE) { 287 int base = (foff + off) & PAGE_MASK; 288 if (vm_page_is_valid(m, base, size) == 0) 289 bp->b_flags &= ~B_CACHE; 290 } 291 } 292 293 static __inline__ 294 void 295 bd_wakeup(int dirtybuflevel) 296 { 297 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 298 bd_request = 1; 299 wakeup(&bd_request); 300 } 301 } 302 303 /* 304 * bd_speedup - speedup the buffer cache flushing code 305 */ 306 307 static __inline__ 308 void 309 bd_speedup(void) 310 { 311 bd_wakeup(1); 312 } 313 314 /* 315 * Initialize buffer headers and related structures. 316 */ 317 318 caddr_t 319 bufhashinit(caddr_t vaddr) 320 { 321 /* first, make a null hash table */ 322 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 323 ; 324 bufhashtbl = (void *)vaddr; 325 vaddr = vaddr + sizeof(*bufhashtbl) * bufhashmask; 326 --bufhashmask; 327 return(vaddr); 328 } 329 330 void 331 bufinit(void) 332 { 333 struct buf *bp; 334 int i; 335 336 TAILQ_INIT(&bswlist); 337 LIST_INIT(&invalhash); 338 mtx_init(&buftimelock, "buftime lock", MTX_DEF); 339 340 for (i = 0; i <= bufhashmask; i++) 341 LIST_INIT(&bufhashtbl[i]); 342 343 /* next, make a null set of free lists */ 344 for (i = 0; i < BUFFER_QUEUES; i++) 345 TAILQ_INIT(&bufqueues[i]); 346 347 /* finally, initialize each buffer header and stick on empty q */ 348 for (i = 0; i < nbuf; i++) { 349 bp = &buf[i]; 350 bzero(bp, sizeof *bp); 351 bp->b_flags = B_INVAL; /* we're just an empty header */ 352 bp->b_dev = NODEV; 353 bp->b_rcred = NOCRED; 354 bp->b_wcred = NOCRED; 355 bp->b_qindex = QUEUE_EMPTY; 356 bp->b_xflags = 0; 357 LIST_INIT(&bp->b_dep); 358 BUF_LOCKINIT(bp); 359 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 360 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 361 } 362 363 /* 364 * maxbufspace is the absolute maximum amount of buffer space we are 365 * allowed to reserve in KVM and in real terms. The absolute maximum 366 * is nominally used by buf_daemon. hibufspace is the nominal maximum 367 * used by most other processes. The differential is required to 368 * ensure that buf_daemon is able to run when other processes might 369 * be blocked waiting for buffer space. 370 * 371 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 372 * this may result in KVM fragmentation which is not handled optimally 373 * by the system. 374 */ 375 maxbufspace = nbuf * BKVASIZE; 376 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 377 lobufspace = hibufspace - MAXBSIZE; 378 379 lorunningspace = 512 * 1024; 380 hirunningspace = 1024 * 1024; 381 382 /* 383 * Limit the amount of malloc memory since it is wired permanently into 384 * the kernel space. Even though this is accounted for in the buffer 385 * allocation, we don't want the malloced region to grow uncontrolled. 386 * The malloc scheme improves memory utilization significantly on average 387 * (small) directories. 388 */ 389 maxbufmallocspace = hibufspace / 20; 390 391 /* 392 * Reduce the chance of a deadlock occuring by limiting the number 393 * of delayed-write dirty buffers we allow to stack up. 394 */ 395 hidirtybuffers = nbuf / 4 + 20; 396 numdirtybuffers = 0; 397 /* 398 * To support extreme low-memory systems, make sure hidirtybuffers cannot 399 * eat up all available buffer space. This occurs when our minimum cannot 400 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 401 * BKVASIZE'd (8K) buffers. 402 */ 403 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 404 hidirtybuffers >>= 1; 405 } 406 lodirtybuffers = hidirtybuffers / 2; 407 408 /* 409 * Try to keep the number of free buffers in the specified range, 410 * and give special processes (e.g. like buf_daemon) access to an 411 * emergency reserve. 412 */ 413 lofreebuffers = nbuf / 18 + 5; 414 hifreebuffers = 2 * lofreebuffers; 415 numfreebuffers = nbuf; 416 417 /* 418 * Maximum number of async ops initiated per buf_daemon loop. This is 419 * somewhat of a hack at the moment, we really need to limit ourselves 420 * based on the number of bytes of I/O in-transit that were initiated 421 * from buf_daemon. 422 */ 423 424 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 425 bogus_page = vm_page_alloc(kernel_object, 426 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 427 VM_ALLOC_NORMAL); 428 cnt.v_wire_count++; 429 430 } 431 432 /* 433 * bfreekva() - free the kva allocation for a buffer. 434 * 435 * Must be called at splbio() or higher as this is the only locking for 436 * buffer_map. 437 * 438 * Since this call frees up buffer space, we call bufspacewakeup(). 439 */ 440 static void 441 bfreekva(struct buf * bp) 442 { 443 if (bp->b_kvasize) { 444 ++buffreekvacnt; 445 bufspace -= bp->b_kvasize; 446 vm_map_delete(buffer_map, 447 (vm_offset_t) bp->b_kvabase, 448 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 449 ); 450 bp->b_kvasize = 0; 451 bufspacewakeup(); 452 } 453 } 454 455 /* 456 * bremfree: 457 * 458 * Remove the buffer from the appropriate free list. 459 */ 460 void 461 bremfree(struct buf * bp) 462 { 463 int s = splbio(); 464 int old_qindex = bp->b_qindex; 465 466 if (bp->b_qindex != QUEUE_NONE) { 467 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 468 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 469 bp->b_qindex = QUEUE_NONE; 470 } else { 471 if (BUF_REFCNT(bp) <= 1) 472 panic("bremfree: removing a buffer not on a queue"); 473 } 474 475 /* 476 * Fixup numfreebuffers count. If the buffer is invalid or not 477 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 478 * the buffer was free and we must decrement numfreebuffers. 479 */ 480 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 481 switch(old_qindex) { 482 case QUEUE_DIRTY: 483 case QUEUE_CLEAN: 484 case QUEUE_EMPTY: 485 case QUEUE_EMPTYKVA: 486 --numfreebuffers; 487 break; 488 default: 489 break; 490 } 491 } 492 splx(s); 493 } 494 495 496 /* 497 * Get a buffer with the specified data. Look in the cache first. We 498 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 499 * is set, the buffer is valid and we do not have to do anything ( see 500 * getblk() ). 501 */ 502 int 503 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 504 struct buf ** bpp) 505 { 506 struct buf *bp; 507 508 bp = getblk(vp, blkno, size, 0, 0); 509 *bpp = bp; 510 511 /* if not found in cache, do some I/O */ 512 if ((bp->b_flags & B_CACHE) == 0) { 513 if (curproc != PCPU_GET(idleproc)) 514 curproc->p_stats->p_ru.ru_inblock++; 515 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp)); 516 bp->b_iocmd = BIO_READ; 517 bp->b_flags &= ~B_INVAL; 518 bp->b_ioflags &= ~BIO_ERROR; 519 if (bp->b_rcred == NOCRED) { 520 if (cred != NOCRED) 521 crhold(cred); 522 bp->b_rcred = cred; 523 } 524 vfs_busy_pages(bp, 0); 525 VOP_STRATEGY(vp, bp); 526 return (bufwait(bp)); 527 } 528 return (0); 529 } 530 531 /* 532 * Operates like bread, but also starts asynchronous I/O on 533 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 534 * to initiating I/O . If B_CACHE is set, the buffer is valid 535 * and we do not have to do anything. 536 */ 537 int 538 breadn(struct vnode * vp, daddr_t blkno, int size, 539 daddr_t * rablkno, int *rabsize, 540 int cnt, struct ucred * cred, struct buf ** bpp) 541 { 542 struct buf *bp, *rabp; 543 int i; 544 int rv = 0, readwait = 0; 545 546 *bpp = bp = getblk(vp, blkno, size, 0, 0); 547 548 /* if not found in cache, do some I/O */ 549 if ((bp->b_flags & B_CACHE) == 0) { 550 if (curproc != PCPU_GET(idleproc)) 551 curproc->p_stats->p_ru.ru_inblock++; 552 bp->b_iocmd = BIO_READ; 553 bp->b_flags &= ~B_INVAL; 554 bp->b_ioflags &= ~BIO_ERROR; 555 if (bp->b_rcred == NOCRED) { 556 if (cred != NOCRED) 557 crhold(cred); 558 bp->b_rcred = cred; 559 } 560 vfs_busy_pages(bp, 0); 561 VOP_STRATEGY(vp, bp); 562 ++readwait; 563 } 564 565 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 566 if (inmem(vp, *rablkno)) 567 continue; 568 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 569 570 if ((rabp->b_flags & B_CACHE) == 0) { 571 if (curproc != PCPU_GET(idleproc)) 572 curproc->p_stats->p_ru.ru_inblock++; 573 rabp->b_flags |= B_ASYNC; 574 rabp->b_flags &= ~B_INVAL; 575 rabp->b_ioflags &= ~BIO_ERROR; 576 rabp->b_iocmd = BIO_READ; 577 if (rabp->b_rcred == NOCRED) { 578 if (cred != NOCRED) 579 crhold(cred); 580 rabp->b_rcred = cred; 581 } 582 vfs_busy_pages(rabp, 0); 583 BUF_KERNPROC(rabp); 584 VOP_STRATEGY(vp, rabp); 585 } else { 586 brelse(rabp); 587 } 588 } 589 590 if (readwait) { 591 rv = bufwait(bp); 592 } 593 return (rv); 594 } 595 596 /* 597 * Write, release buffer on completion. (Done by iodone 598 * if async). Do not bother writing anything if the buffer 599 * is invalid. 600 * 601 * Note that we set B_CACHE here, indicating that buffer is 602 * fully valid and thus cacheable. This is true even of NFS 603 * now so we set it generally. This could be set either here 604 * or in biodone() since the I/O is synchronous. We put it 605 * here. 606 */ 607 int 608 bwrite(struct buf * bp) 609 { 610 int oldflags, s; 611 struct buf *newbp; 612 613 if (bp->b_flags & B_INVAL) { 614 brelse(bp); 615 return (0); 616 } 617 618 oldflags = bp->b_flags; 619 620 if (BUF_REFCNT(bp) == 0) 621 panic("bwrite: buffer is not busy???"); 622 s = splbio(); 623 /* 624 * If a background write is already in progress, delay 625 * writing this block if it is asynchronous. Otherwise 626 * wait for the background write to complete. 627 */ 628 if (bp->b_xflags & BX_BKGRDINPROG) { 629 if (bp->b_flags & B_ASYNC) { 630 splx(s); 631 bdwrite(bp); 632 return (0); 633 } 634 bp->b_xflags |= BX_BKGRDWAIT; 635 tsleep(&bp->b_xflags, PRIBIO, "biord", 0); 636 if (bp->b_xflags & BX_BKGRDINPROG) 637 panic("bwrite: still writing"); 638 } 639 640 /* Mark the buffer clean */ 641 bundirty(bp); 642 643 /* 644 * If this buffer is marked for background writing and we 645 * do not have to wait for it, make a copy and write the 646 * copy so as to leave this buffer ready for further use. 647 * 648 * This optimization eats a lot of memory. If we have a page 649 * or buffer shortfall we can't do it. 650 */ 651 if ((bp->b_xflags & BX_BKGRDWRITE) && 652 (bp->b_flags & B_ASYNC) && 653 !vm_page_count_severe() && 654 !buf_dirty_count_severe()) { 655 if (bp->b_iodone != NULL) { 656 printf("bp->b_iodone = %p\n", bp->b_iodone); 657 panic("bwrite: need chained iodone"); 658 } 659 660 /* get a new block */ 661 newbp = geteblk(bp->b_bufsize); 662 663 /* set it to be identical to the old block */ 664 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 665 bgetvp(bp->b_vp, newbp); 666 newbp->b_lblkno = bp->b_lblkno; 667 newbp->b_blkno = bp->b_blkno; 668 newbp->b_offset = bp->b_offset; 669 newbp->b_iodone = vfs_backgroundwritedone; 670 newbp->b_flags |= B_ASYNC; 671 newbp->b_flags &= ~B_INVAL; 672 673 /* move over the dependencies */ 674 if (LIST_FIRST(&bp->b_dep) != NULL) 675 buf_movedeps(bp, newbp); 676 677 /* 678 * Initiate write on the copy, release the original to 679 * the B_LOCKED queue so that it cannot go away until 680 * the background write completes. If not locked it could go 681 * away and then be reconstituted while it was being written. 682 * If the reconstituted buffer were written, we could end up 683 * with two background copies being written at the same time. 684 */ 685 bp->b_xflags |= BX_BKGRDINPROG; 686 bp->b_flags |= B_LOCKED; 687 bqrelse(bp); 688 bp = newbp; 689 } 690 691 bp->b_flags &= ~B_DONE; 692 bp->b_ioflags &= ~BIO_ERROR; 693 bp->b_flags |= B_WRITEINPROG | B_CACHE; 694 bp->b_iocmd = BIO_WRITE; 695 696 bp->b_vp->v_numoutput++; 697 vfs_busy_pages(bp, 1); 698 if (curproc != PCPU_GET(idleproc)) 699 curproc->p_stats->p_ru.ru_oublock++; 700 splx(s); 701 if (oldflags & B_ASYNC) 702 BUF_KERNPROC(bp); 703 BUF_STRATEGY(bp); 704 705 if ((oldflags & B_ASYNC) == 0) { 706 int rtval = bufwait(bp); 707 brelse(bp); 708 return (rtval); 709 } else { 710 /* 711 * don't allow the async write to saturate the I/O 712 * system. There is no chance of deadlock here because 713 * we are blocking on I/O that is already in-progress. 714 */ 715 waitrunningbufspace(); 716 } 717 718 return (0); 719 } 720 721 /* 722 * Complete a background write started from bwrite. 723 */ 724 static void 725 vfs_backgroundwritedone(bp) 726 struct buf *bp; 727 { 728 struct buf *origbp; 729 730 /* 731 * Find the original buffer that we are writing. 732 */ 733 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 734 panic("backgroundwritedone: lost buffer"); 735 /* 736 * Process dependencies then return any unfinished ones. 737 */ 738 if (LIST_FIRST(&bp->b_dep) != NULL) 739 buf_complete(bp); 740 if (LIST_FIRST(&bp->b_dep) != NULL) 741 buf_movedeps(bp, origbp); 742 /* 743 * Clear the BX_BKGRDINPROG flag in the original buffer 744 * and awaken it if it is waiting for the write to complete. 745 * If BX_BKGRDINPROG is not set in the original buffer it must 746 * have been released and re-instantiated - which is not legal. 747 */ 748 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), ("backgroundwritedone: lost buffer2")); 749 origbp->b_xflags &= ~BX_BKGRDINPROG; 750 if (origbp->b_xflags & BX_BKGRDWAIT) { 751 origbp->b_xflags &= ~BX_BKGRDWAIT; 752 wakeup(&origbp->b_xflags); 753 } 754 /* 755 * Clear the B_LOCKED flag and remove it from the locked 756 * queue if it currently resides there. 757 */ 758 origbp->b_flags &= ~B_LOCKED; 759 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 760 bremfree(origbp); 761 bqrelse(origbp); 762 } 763 /* 764 * This buffer is marked B_NOCACHE, so when it is released 765 * by biodone, it will be tossed. We mark it with BIO_READ 766 * to avoid biodone doing a second vwakeup. 767 */ 768 bp->b_flags |= B_NOCACHE; 769 bp->b_iocmd = BIO_READ; 770 bp->b_flags &= ~(B_CACHE | B_DONE); 771 bp->b_iodone = 0; 772 bufdone(bp); 773 } 774 775 /* 776 * Delayed write. (Buffer is marked dirty). Do not bother writing 777 * anything if the buffer is marked invalid. 778 * 779 * Note that since the buffer must be completely valid, we can safely 780 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 781 * biodone() in order to prevent getblk from writing the buffer 782 * out synchronously. 783 */ 784 void 785 bdwrite(struct buf * bp) 786 { 787 if (BUF_REFCNT(bp) == 0) 788 panic("bdwrite: buffer is not busy"); 789 790 if (bp->b_flags & B_INVAL) { 791 brelse(bp); 792 return; 793 } 794 bdirty(bp); 795 796 /* 797 * Set B_CACHE, indicating that the buffer is fully valid. This is 798 * true even of NFS now. 799 */ 800 bp->b_flags |= B_CACHE; 801 802 /* 803 * This bmap keeps the system from needing to do the bmap later, 804 * perhaps when the system is attempting to do a sync. Since it 805 * is likely that the indirect block -- or whatever other datastructure 806 * that the filesystem needs is still in memory now, it is a good 807 * thing to do this. Note also, that if the pageout daemon is 808 * requesting a sync -- there might not be enough memory to do 809 * the bmap then... So, this is important to do. 810 */ 811 if (bp->b_lblkno == bp->b_blkno) { 812 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 813 } 814 815 /* 816 * Set the *dirty* buffer range based upon the VM system dirty pages. 817 */ 818 vfs_setdirty(bp); 819 820 /* 821 * We need to do this here to satisfy the vnode_pager and the 822 * pageout daemon, so that it thinks that the pages have been 823 * "cleaned". Note that since the pages are in a delayed write 824 * buffer -- the VFS layer "will" see that the pages get written 825 * out on the next sync, or perhaps the cluster will be completed. 826 */ 827 vfs_clean_pages(bp); 828 bqrelse(bp); 829 830 /* 831 * Wakeup the buffer flushing daemon if we have a lot of dirty 832 * buffers (midpoint between our recovery point and our stall 833 * point). 834 */ 835 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 836 837 /* 838 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 839 * due to the softdep code. 840 */ 841 } 842 843 /* 844 * bdirty: 845 * 846 * Turn buffer into delayed write request. We must clear BIO_READ and 847 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 848 * itself to properly update it in the dirty/clean lists. We mark it 849 * B_DONE to ensure that any asynchronization of the buffer properly 850 * clears B_DONE ( else a panic will occur later ). 851 * 852 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 853 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 854 * should only be called if the buffer is known-good. 855 * 856 * Since the buffer is not on a queue, we do not update the numfreebuffers 857 * count. 858 * 859 * Must be called at splbio(). 860 * The buffer must be on QUEUE_NONE. 861 */ 862 void 863 bdirty(bp) 864 struct buf *bp; 865 { 866 KASSERT(bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 867 bp->b_flags &= ~(B_RELBUF); 868 bp->b_iocmd = BIO_WRITE; 869 870 if ((bp->b_flags & B_DELWRI) == 0) { 871 bp->b_flags |= B_DONE | B_DELWRI; 872 reassignbuf(bp, bp->b_vp); 873 ++numdirtybuffers; 874 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 875 } 876 } 877 878 /* 879 * bundirty: 880 * 881 * Clear B_DELWRI for buffer. 882 * 883 * Since the buffer is not on a queue, we do not update the numfreebuffers 884 * count. 885 * 886 * Must be called at splbio(). 887 * The buffer must be on QUEUE_NONE. 888 */ 889 890 void 891 bundirty(bp) 892 struct buf *bp; 893 { 894 KASSERT(bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 895 896 if (bp->b_flags & B_DELWRI) { 897 bp->b_flags &= ~B_DELWRI; 898 reassignbuf(bp, bp->b_vp); 899 --numdirtybuffers; 900 numdirtywakeup(lodirtybuffers); 901 } 902 /* 903 * Since it is now being written, we can clear its deferred write flag. 904 */ 905 bp->b_flags &= ~B_DEFERRED; 906 } 907 908 /* 909 * bawrite: 910 * 911 * Asynchronous write. Start output on a buffer, but do not wait for 912 * it to complete. The buffer is released when the output completes. 913 * 914 * bwrite() ( or the VOP routine anyway ) is responsible for handling 915 * B_INVAL buffers. Not us. 916 */ 917 void 918 bawrite(struct buf * bp) 919 { 920 bp->b_flags |= B_ASYNC; 921 (void) BUF_WRITE(bp); 922 } 923 924 /* 925 * bowrite: 926 * 927 * Ordered write. Start output on a buffer, and flag it so that the 928 * device will write it in the order it was queued. The buffer is 929 * released when the output completes. bwrite() ( or the VOP routine 930 * anyway ) is responsible for handling B_INVAL buffers. 931 */ 932 int 933 bowrite(struct buf * bp) 934 { 935 bp->b_ioflags |= BIO_ORDERED; 936 bp->b_flags |= B_ASYNC; 937 return (BUF_WRITE(bp)); 938 } 939 940 /* 941 * bwillwrite: 942 * 943 * Called prior to the locking of any vnodes when we are expecting to 944 * write. We do not want to starve the buffer cache with too many 945 * dirty buffers so we block here. By blocking prior to the locking 946 * of any vnodes we attempt to avoid the situation where a locked vnode 947 * prevents the various system daemons from flushing related buffers. 948 */ 949 950 void 951 bwillwrite(void) 952 { 953 if (numdirtybuffers >= hidirtybuffers) { 954 int s; 955 956 s = splbio(); 957 while (numdirtybuffers >= hidirtybuffers) { 958 bd_wakeup(1); 959 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 960 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0); 961 } 962 splx(s); 963 } 964 } 965 966 /* 967 * Return true if we have too many dirty buffers. 968 */ 969 int 970 buf_dirty_count_severe(void) 971 { 972 return(numdirtybuffers >= hidirtybuffers); 973 } 974 975 /* 976 * brelse: 977 * 978 * Release a busy buffer and, if requested, free its resources. The 979 * buffer will be stashed in the appropriate bufqueue[] allowing it 980 * to be accessed later as a cache entity or reused for other purposes. 981 */ 982 void 983 brelse(struct buf * bp) 984 { 985 int s; 986 987 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 988 989 s = splbio(); 990 991 if (bp->b_flags & B_LOCKED) 992 bp->b_ioflags &= ~BIO_ERROR; 993 994 if (bp->b_iocmd == BIO_WRITE && 995 (bp->b_ioflags & BIO_ERROR) && 996 !(bp->b_flags & B_INVAL)) { 997 /* 998 * Failed write, redirty. Must clear BIO_ERROR to prevent 999 * pages from being scrapped. If B_INVAL is set then 1000 * this case is not run and the next case is run to 1001 * destroy the buffer. B_INVAL can occur if the buffer 1002 * is outside the range supported by the underlying device. 1003 */ 1004 bp->b_ioflags &= ~BIO_ERROR; 1005 bdirty(bp); 1006 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1007 (bp->b_ioflags & BIO_ERROR) || 1008 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1009 /* 1010 * Either a failed I/O or we were asked to free or not 1011 * cache the buffer. 1012 */ 1013 bp->b_flags |= B_INVAL; 1014 if (LIST_FIRST(&bp->b_dep) != NULL) 1015 buf_deallocate(bp); 1016 if (bp->b_flags & B_DELWRI) { 1017 --numdirtybuffers; 1018 numdirtywakeup(lodirtybuffers); 1019 } 1020 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1021 if ((bp->b_flags & B_VMIO) == 0) { 1022 if (bp->b_bufsize) 1023 allocbuf(bp, 0); 1024 if (bp->b_vp) 1025 brelvp(bp); 1026 } 1027 } 1028 1029 /* 1030 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1031 * is called with B_DELWRI set, the underlying pages may wind up 1032 * getting freed causing a previous write (bdwrite()) to get 'lost' 1033 * because pages associated with a B_DELWRI bp are marked clean. 1034 * 1035 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1036 * if B_DELWRI is set. 1037 * 1038 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1039 * on pages to return pages to the VM page queues. 1040 */ 1041 if (bp->b_flags & B_DELWRI) 1042 bp->b_flags &= ~B_RELBUF; 1043 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1044 bp->b_flags |= B_RELBUF; 1045 1046 /* 1047 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1048 * constituted, not even NFS buffers now. Two flags effect this. If 1049 * B_INVAL, the struct buf is invalidated but the VM object is kept 1050 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1051 * 1052 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1053 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1054 * buffer is also B_INVAL because it hits the re-dirtying code above. 1055 * 1056 * Normally we can do this whether a buffer is B_DELWRI or not. If 1057 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1058 * the commit state and we cannot afford to lose the buffer. If the 1059 * buffer has a background write in progress, we need to keep it 1060 * around to prevent it from being reconstituted and starting a second 1061 * background write. 1062 */ 1063 if ((bp->b_flags & B_VMIO) 1064 && !(bp->b_vp->v_tag == VT_NFS && 1065 !vn_isdisk(bp->b_vp, NULL) && 1066 (bp->b_flags & B_DELWRI)) 1067 ) { 1068 1069 int i, j, resid; 1070 vm_page_t m; 1071 off_t foff; 1072 vm_pindex_t poff; 1073 vm_object_t obj; 1074 struct vnode *vp; 1075 1076 vp = bp->b_vp; 1077 1078 /* 1079 * Get the base offset and length of the buffer. Note that 1080 * for block sizes that are less then PAGE_SIZE, the b_data 1081 * base of the buffer does not represent exactly b_offset and 1082 * neither b_offset nor b_size are necessarily page aligned. 1083 * Instead, the starting position of b_offset is: 1084 * 1085 * b_data + (b_offset & PAGE_MASK) 1086 * 1087 * block sizes less then DEV_BSIZE (usually 512) are not 1088 * supported due to the page granularity bits (m->valid, 1089 * m->dirty, etc...). 1090 * 1091 * See man buf(9) for more information 1092 */ 1093 resid = bp->b_bufsize; 1094 foff = bp->b_offset; 1095 1096 for (i = 0; i < bp->b_npages; i++) { 1097 int had_bogus = 0; 1098 1099 m = bp->b_pages[i]; 1100 vm_page_flag_clear(m, PG_ZERO); 1101 1102 /* 1103 * If we hit a bogus page, fixup *all* the bogus pages 1104 * now. 1105 */ 1106 if (m == bogus_page) { 1107 VOP_GETVOBJECT(vp, &obj); 1108 poff = OFF_TO_IDX(bp->b_offset); 1109 had_bogus = 1; 1110 1111 for (j = i; j < bp->b_npages; j++) { 1112 vm_page_t mtmp; 1113 mtmp = bp->b_pages[j]; 1114 if (mtmp == bogus_page) { 1115 mtmp = vm_page_lookup(obj, poff + j); 1116 if (!mtmp) { 1117 panic("brelse: page missing\n"); 1118 } 1119 bp->b_pages[j] = mtmp; 1120 } 1121 } 1122 1123 if ((bp->b_flags & B_INVAL) == 0) { 1124 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1125 } 1126 m = bp->b_pages[i]; 1127 } 1128 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1129 int poffset = foff & PAGE_MASK; 1130 int presid = resid > (PAGE_SIZE - poffset) ? 1131 (PAGE_SIZE - poffset) : resid; 1132 1133 KASSERT(presid >= 0, ("brelse: extra page")); 1134 vm_page_set_invalid(m, poffset, presid); 1135 if (had_bogus) 1136 printf("avoided corruption bug in bogus_page/brelse code\n"); 1137 } 1138 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1139 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1140 } 1141 1142 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1143 vfs_vmio_release(bp); 1144 1145 } else if (bp->b_flags & B_VMIO) { 1146 1147 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1148 vfs_vmio_release(bp); 1149 1150 } 1151 1152 if (bp->b_qindex != QUEUE_NONE) 1153 panic("brelse: free buffer onto another queue???"); 1154 if (BUF_REFCNT(bp) > 1) { 1155 /* do not release to free list */ 1156 BUF_UNLOCK(bp); 1157 splx(s); 1158 return; 1159 } 1160 1161 /* enqueue */ 1162 1163 /* buffers with no memory */ 1164 if (bp->b_bufsize == 0) { 1165 bp->b_flags |= B_INVAL; 1166 bp->b_xflags &= ~BX_BKGRDWRITE; 1167 if (bp->b_xflags & BX_BKGRDINPROG) 1168 panic("losing buffer 1"); 1169 if (bp->b_kvasize) { 1170 bp->b_qindex = QUEUE_EMPTYKVA; 1171 } else { 1172 bp->b_qindex = QUEUE_EMPTY; 1173 } 1174 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1175 LIST_REMOVE(bp, b_hash); 1176 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1177 bp->b_dev = NODEV; 1178 /* buffers with junk contents */ 1179 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || (bp->b_ioflags & BIO_ERROR)) { 1180 bp->b_flags |= B_INVAL; 1181 bp->b_xflags &= ~BX_BKGRDWRITE; 1182 if (bp->b_xflags & BX_BKGRDINPROG) 1183 panic("losing buffer 2"); 1184 bp->b_qindex = QUEUE_CLEAN; 1185 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1186 LIST_REMOVE(bp, b_hash); 1187 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1188 bp->b_dev = NODEV; 1189 1190 /* buffers that are locked */ 1191 } else if (bp->b_flags & B_LOCKED) { 1192 bp->b_qindex = QUEUE_LOCKED; 1193 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1194 1195 /* remaining buffers */ 1196 } else { 1197 switch(bp->b_flags & (B_DELWRI|B_AGE)) { 1198 case B_DELWRI | B_AGE: 1199 bp->b_qindex = QUEUE_DIRTY; 1200 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1201 break; 1202 case B_DELWRI: 1203 bp->b_qindex = QUEUE_DIRTY; 1204 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1205 break; 1206 case B_AGE: 1207 bp->b_qindex = QUEUE_CLEAN; 1208 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1209 break; 1210 default: 1211 bp->b_qindex = QUEUE_CLEAN; 1212 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1213 break; 1214 } 1215 } 1216 1217 /* 1218 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1219 * on the correct queue. 1220 */ 1221 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) { 1222 bp->b_flags &= ~B_DELWRI; 1223 --numdirtybuffers; 1224 numdirtywakeup(lodirtybuffers); 1225 } 1226 1227 /* 1228 * Fixup numfreebuffers count. The bp is on an appropriate queue 1229 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1230 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1231 * if B_INVAL is set ). 1232 */ 1233 1234 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1235 bufcountwakeup(); 1236 1237 /* 1238 * Something we can maybe free or reuse 1239 */ 1240 if (bp->b_bufsize || bp->b_kvasize) 1241 bufspacewakeup(); 1242 1243 /* unlock */ 1244 BUF_UNLOCK(bp); 1245 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1246 bp->b_ioflags &= ~BIO_ORDERED; 1247 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1248 panic("brelse: not dirty"); 1249 splx(s); 1250 } 1251 1252 /* 1253 * Release a buffer back to the appropriate queue but do not try to free 1254 * it. The buffer is expected to be used again soon. 1255 * 1256 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1257 * biodone() to requeue an async I/O on completion. It is also used when 1258 * known good buffers need to be requeued but we think we may need the data 1259 * again soon. 1260 */ 1261 void 1262 bqrelse(struct buf * bp) 1263 { 1264 int s; 1265 1266 s = splbio(); 1267 1268 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1269 1270 if (bp->b_qindex != QUEUE_NONE) 1271 panic("bqrelse: free buffer onto another queue???"); 1272 if (BUF_REFCNT(bp) > 1) { 1273 /* do not release to free list */ 1274 BUF_UNLOCK(bp); 1275 splx(s); 1276 return; 1277 } 1278 if (bp->b_flags & B_LOCKED) { 1279 bp->b_ioflags &= ~BIO_ERROR; 1280 bp->b_qindex = QUEUE_LOCKED; 1281 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1282 /* buffers with stale but valid contents */ 1283 } else if (bp->b_flags & B_DELWRI) { 1284 bp->b_qindex = QUEUE_DIRTY; 1285 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1286 } else if (vm_page_count_severe()) { 1287 /* 1288 * We are too low on memory, we have to try to free the 1289 * buffer (most importantly: the wired pages making up its 1290 * backing store) *now*. 1291 */ 1292 splx(s); 1293 brelse(bp); 1294 return; 1295 } else { 1296 bp->b_qindex = QUEUE_CLEAN; 1297 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1298 } 1299 1300 if ((bp->b_flags & B_LOCKED) == 0 && 1301 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1302 bufcountwakeup(); 1303 } 1304 1305 /* 1306 * Something we can maybe free or reuse. 1307 */ 1308 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1309 bufspacewakeup(); 1310 1311 /* unlock */ 1312 BUF_UNLOCK(bp); 1313 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1314 bp->b_ioflags &= ~BIO_ORDERED; 1315 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1316 panic("bqrelse: not dirty"); 1317 splx(s); 1318 } 1319 1320 static void 1321 vfs_vmio_release(bp) 1322 struct buf *bp; 1323 { 1324 int i, s; 1325 vm_page_t m; 1326 1327 s = splvm(); 1328 for (i = 0; i < bp->b_npages; i++) { 1329 m = bp->b_pages[i]; 1330 bp->b_pages[i] = NULL; 1331 /* 1332 * In order to keep page LRU ordering consistent, put 1333 * everything on the inactive queue. 1334 */ 1335 vm_page_unwire(m, 0); 1336 /* 1337 * We don't mess with busy pages, it is 1338 * the responsibility of the process that 1339 * busied the pages to deal with them. 1340 */ 1341 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1342 continue; 1343 1344 if (m->wire_count == 0) { 1345 vm_page_flag_clear(m, PG_ZERO); 1346 /* 1347 * Might as well free the page if we can and it has 1348 * no valid data. 1349 */ 1350 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { 1351 vm_page_busy(m); 1352 vm_page_protect(m, VM_PROT_NONE); 1353 vm_page_free(m); 1354 } else if (vm_page_count_severe()) { 1355 vm_page_try_to_cache(m); 1356 } 1357 } 1358 } 1359 splx(s); 1360 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1361 if (bp->b_bufsize) { 1362 bufspacewakeup(); 1363 bp->b_bufsize = 0; 1364 } 1365 bp->b_npages = 0; 1366 bp->b_flags &= ~B_VMIO; 1367 if (bp->b_vp) 1368 brelvp(bp); 1369 } 1370 1371 /* 1372 * Check to see if a block is currently memory resident. 1373 */ 1374 struct buf * 1375 gbincore(struct vnode * vp, daddr_t blkno) 1376 { 1377 struct buf *bp; 1378 struct bufhashhdr *bh; 1379 1380 bh = bufhash(vp, blkno); 1381 1382 /* Search hash chain */ 1383 LIST_FOREACH(bp, bh, b_hash) { 1384 /* hit */ 1385 if (bp->b_vp == vp && bp->b_lblkno == blkno && 1386 (bp->b_flags & B_INVAL) == 0) { 1387 break; 1388 } 1389 } 1390 return (bp); 1391 } 1392 1393 /* 1394 * vfs_bio_awrite: 1395 * 1396 * Implement clustered async writes for clearing out B_DELWRI buffers. 1397 * This is much better then the old way of writing only one buffer at 1398 * a time. Note that we may not be presented with the buffers in the 1399 * correct order, so we search for the cluster in both directions. 1400 */ 1401 int 1402 vfs_bio_awrite(struct buf * bp) 1403 { 1404 int i; 1405 int j; 1406 daddr_t lblkno = bp->b_lblkno; 1407 struct vnode *vp = bp->b_vp; 1408 int s; 1409 int ncl; 1410 struct buf *bpa; 1411 int nwritten; 1412 int size; 1413 int maxcl; 1414 1415 s = splbio(); 1416 /* 1417 * right now we support clustered writing only to regular files. If 1418 * we find a clusterable block we could be in the middle of a cluster 1419 * rather then at the beginning. 1420 */ 1421 if ((vp->v_type == VREG) && 1422 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1423 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1424 1425 size = vp->v_mount->mnt_stat.f_iosize; 1426 maxcl = MAXPHYS / size; 1427 1428 for (i = 1; i < maxcl; i++) { 1429 if ((bpa = gbincore(vp, lblkno + i)) && 1430 BUF_REFCNT(bpa) == 0 && 1431 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1432 (B_DELWRI | B_CLUSTEROK)) && 1433 (bpa->b_bufsize == size)) { 1434 if ((bpa->b_blkno == bpa->b_lblkno) || 1435 (bpa->b_blkno != 1436 bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 1437 break; 1438 } else { 1439 break; 1440 } 1441 } 1442 for (j = 1; i + j <= maxcl && j <= lblkno; j++) { 1443 if ((bpa = gbincore(vp, lblkno - j)) && 1444 BUF_REFCNT(bpa) == 0 && 1445 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1446 (B_DELWRI | B_CLUSTEROK)) && 1447 (bpa->b_bufsize == size)) { 1448 if ((bpa->b_blkno == bpa->b_lblkno) || 1449 (bpa->b_blkno != 1450 bp->b_blkno - ((j * size) >> DEV_BSHIFT))) 1451 break; 1452 } else { 1453 break; 1454 } 1455 } 1456 --j; 1457 ncl = i + j; 1458 /* 1459 * this is a possible cluster write 1460 */ 1461 if (ncl != 1) { 1462 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1463 splx(s); 1464 return nwritten; 1465 } 1466 } 1467 1468 BUF_LOCK(bp, LK_EXCLUSIVE); 1469 bremfree(bp); 1470 bp->b_flags |= B_ASYNC; 1471 1472 splx(s); 1473 /* 1474 * default (old) behavior, writing out only one block 1475 * 1476 * XXX returns b_bufsize instead of b_bcount for nwritten? 1477 */ 1478 nwritten = bp->b_bufsize; 1479 (void) BUF_WRITE(bp); 1480 1481 return nwritten; 1482 } 1483 1484 /* 1485 * getnewbuf: 1486 * 1487 * Find and initialize a new buffer header, freeing up existing buffers 1488 * in the bufqueues as necessary. The new buffer is returned locked. 1489 * 1490 * Important: B_INVAL is not set. If the caller wishes to throw the 1491 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1492 * 1493 * We block if: 1494 * We have insufficient buffer headers 1495 * We have insufficient buffer space 1496 * buffer_map is too fragmented ( space reservation fails ) 1497 * If we have to flush dirty buffers ( but we try to avoid this ) 1498 * 1499 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1500 * Instead we ask the buf daemon to do it for us. We attempt to 1501 * avoid piecemeal wakeups of the pageout daemon. 1502 */ 1503 1504 static struct buf * 1505 getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1506 { 1507 struct buf *bp; 1508 struct buf *nbp; 1509 int defrag = 0; 1510 int nqindex; 1511 static int flushingbufs; 1512 1513 /* 1514 * We can't afford to block since we might be holding a vnode lock, 1515 * which may prevent system daemons from running. We deal with 1516 * low-memory situations by proactively returning memory and running 1517 * async I/O rather then sync I/O. 1518 */ 1519 1520 ++getnewbufcalls; 1521 --getnewbufrestarts; 1522 restart: 1523 ++getnewbufrestarts; 1524 1525 /* 1526 * Setup for scan. If we do not have enough free buffers, 1527 * we setup a degenerate case that immediately fails. Note 1528 * that if we are specially marked process, we are allowed to 1529 * dip into our reserves. 1530 * 1531 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1532 * 1533 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1534 * However, there are a number of cases (defragging, reusing, ...) 1535 * where we cannot backup. 1536 */ 1537 nqindex = QUEUE_EMPTYKVA; 1538 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1539 1540 if (nbp == NULL) { 1541 /* 1542 * If no EMPTYKVA buffers and we are either 1543 * defragging or reusing, locate a CLEAN buffer 1544 * to free or reuse. If bufspace useage is low 1545 * skip this step so we can allocate a new buffer. 1546 */ 1547 if (defrag || bufspace >= lobufspace) { 1548 nqindex = QUEUE_CLEAN; 1549 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1550 } 1551 1552 /* 1553 * If we could not find or were not allowed to reuse a 1554 * CLEAN buffer, check to see if it is ok to use an EMPTY 1555 * buffer. We can only use an EMPTY buffer if allocating 1556 * its KVA would not otherwise run us out of buffer space. 1557 */ 1558 if (nbp == NULL && defrag == 0 && 1559 bufspace + maxsize < hibufspace) { 1560 nqindex = QUEUE_EMPTY; 1561 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1562 } 1563 } 1564 1565 /* 1566 * Run scan, possibly freeing data and/or kva mappings on the fly 1567 * depending. 1568 */ 1569 1570 while ((bp = nbp) != NULL) { 1571 int qindex = nqindex; 1572 1573 /* 1574 * Calculate next bp ( we can only use it if we do not block 1575 * or do other fancy things ). 1576 */ 1577 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1578 switch(qindex) { 1579 case QUEUE_EMPTY: 1580 nqindex = QUEUE_EMPTYKVA; 1581 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1582 break; 1583 /* fall through */ 1584 case QUEUE_EMPTYKVA: 1585 nqindex = QUEUE_CLEAN; 1586 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1587 break; 1588 /* fall through */ 1589 case QUEUE_CLEAN: 1590 /* 1591 * nbp is NULL. 1592 */ 1593 break; 1594 } 1595 } 1596 1597 /* 1598 * Sanity Checks 1599 */ 1600 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1601 1602 /* 1603 * Note: we no longer distinguish between VMIO and non-VMIO 1604 * buffers. 1605 */ 1606 1607 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1608 1609 /* 1610 * If we are defragging then we need a buffer with 1611 * b_kvasize != 0. XXX this situation should no longer 1612 * occur, if defrag is non-zero the buffer's b_kvasize 1613 * should also be non-zero at this point. XXX 1614 */ 1615 if (defrag && bp->b_kvasize == 0) { 1616 printf("Warning: defrag empty buffer %p\n", bp); 1617 continue; 1618 } 1619 1620 /* 1621 * Start freeing the bp. This is somewhat involved. nbp 1622 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1623 */ 1624 1625 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1626 panic("getnewbuf: locked buf"); 1627 bremfree(bp); 1628 1629 if (qindex == QUEUE_CLEAN) { 1630 if (bp->b_flags & B_VMIO) { 1631 bp->b_flags &= ~B_ASYNC; 1632 vfs_vmio_release(bp); 1633 } 1634 if (bp->b_vp) 1635 brelvp(bp); 1636 } 1637 1638 /* 1639 * NOTE: nbp is now entirely invalid. We can only restart 1640 * the scan from this point on. 1641 * 1642 * Get the rest of the buffer freed up. b_kva* is still 1643 * valid after this operation. 1644 */ 1645 1646 if (bp->b_rcred != NOCRED) { 1647 crfree(bp->b_rcred); 1648 bp->b_rcred = NOCRED; 1649 } 1650 if (bp->b_wcred != NOCRED) { 1651 crfree(bp->b_wcred); 1652 bp->b_wcred = NOCRED; 1653 } 1654 if (LIST_FIRST(&bp->b_dep) != NULL) 1655 buf_deallocate(bp); 1656 if (bp->b_xflags & BX_BKGRDINPROG) 1657 panic("losing buffer 3"); 1658 LIST_REMOVE(bp, b_hash); 1659 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1660 1661 if (bp->b_bufsize) 1662 allocbuf(bp, 0); 1663 1664 bp->b_flags = 0; 1665 bp->b_ioflags = 0; 1666 bp->b_xflags = 0; 1667 bp->b_dev = NODEV; 1668 bp->b_vp = NULL; 1669 bp->b_blkno = bp->b_lblkno = 0; 1670 bp->b_offset = NOOFFSET; 1671 bp->b_iodone = 0; 1672 bp->b_error = 0; 1673 bp->b_resid = 0; 1674 bp->b_bcount = 0; 1675 bp->b_npages = 0; 1676 bp->b_dirtyoff = bp->b_dirtyend = 0; 1677 1678 LIST_INIT(&bp->b_dep); 1679 1680 /* 1681 * If we are defragging then free the buffer. 1682 */ 1683 if (defrag) { 1684 bp->b_flags |= B_INVAL; 1685 bfreekva(bp); 1686 brelse(bp); 1687 defrag = 0; 1688 goto restart; 1689 } 1690 1691 /* 1692 * If we are overcomitted then recover the buffer and its 1693 * KVM space. This occurs in rare situations when multiple 1694 * processes are blocked in getnewbuf() or allocbuf(). 1695 */ 1696 if (bufspace >= hibufspace) 1697 flushingbufs = 1; 1698 if (flushingbufs && bp->b_kvasize != 0) { 1699 bp->b_flags |= B_INVAL; 1700 bfreekva(bp); 1701 brelse(bp); 1702 goto restart; 1703 } 1704 if (bufspace < lobufspace) 1705 flushingbufs = 0; 1706 break; 1707 } 1708 1709 /* 1710 * If we exhausted our list, sleep as appropriate. We may have to 1711 * wakeup various daemons and write out some dirty buffers. 1712 * 1713 * Generally we are sleeping due to insufficient buffer space. 1714 */ 1715 1716 if (bp == NULL) { 1717 int flags; 1718 char *waitmsg; 1719 1720 if (defrag) { 1721 flags = VFS_BIO_NEED_BUFSPACE; 1722 waitmsg = "nbufkv"; 1723 } else if (bufspace >= hibufspace) { 1724 waitmsg = "nbufbs"; 1725 flags = VFS_BIO_NEED_BUFSPACE; 1726 } else { 1727 waitmsg = "newbuf"; 1728 flags = VFS_BIO_NEED_ANY; 1729 } 1730 1731 bd_speedup(); /* heeeelp */ 1732 1733 needsbuffer |= flags; 1734 while (needsbuffer & flags) { 1735 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, 1736 waitmsg, slptimeo)) 1737 return (NULL); 1738 } 1739 } else { 1740 /* 1741 * We finally have a valid bp. We aren't quite out of the 1742 * woods, we still have to reserve kva space. In order 1743 * to keep fragmentation sane we only allocate kva in 1744 * BKVASIZE chunks. 1745 */ 1746 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1747 1748 if (maxsize != bp->b_kvasize) { 1749 vm_offset_t addr = 0; 1750 1751 bfreekva(bp); 1752 1753 if (vm_map_findspace(buffer_map, 1754 vm_map_min(buffer_map), maxsize, &addr)) { 1755 /* 1756 * Uh oh. Buffer map is to fragmented. We 1757 * must defragment the map. 1758 */ 1759 ++bufdefragcnt; 1760 defrag = 1; 1761 bp->b_flags |= B_INVAL; 1762 brelse(bp); 1763 goto restart; 1764 } 1765 if (addr) { 1766 vm_map_insert(buffer_map, NULL, 0, 1767 addr, addr + maxsize, 1768 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1769 1770 bp->b_kvabase = (caddr_t) addr; 1771 bp->b_kvasize = maxsize; 1772 bufspace += bp->b_kvasize; 1773 ++bufreusecnt; 1774 } 1775 } 1776 bp->b_data = bp->b_kvabase; 1777 } 1778 return(bp); 1779 } 1780 1781 /* 1782 * buf_daemon: 1783 * 1784 * buffer flushing daemon. Buffers are normally flushed by the 1785 * update daemon but if it cannot keep up this process starts to 1786 * take the load in an attempt to prevent getnewbuf() from blocking. 1787 */ 1788 1789 static struct proc *bufdaemonproc; 1790 1791 static struct kproc_desc buf_kp = { 1792 "bufdaemon", 1793 buf_daemon, 1794 &bufdaemonproc 1795 }; 1796 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1797 1798 static void 1799 buf_daemon() 1800 { 1801 int s; 1802 1803 mtx_lock(&Giant); 1804 1805 /* 1806 * This process needs to be suspended prior to shutdown sync. 1807 */ 1808 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 1809 SHUTDOWN_PRI_LAST); 1810 1811 /* 1812 * This process is allowed to take the buffer cache to the limit 1813 */ 1814 curproc->p_flag |= P_BUFEXHAUST; 1815 s = splbio(); 1816 1817 for (;;) { 1818 kthread_suspend_check(bufdaemonproc); 1819 1820 bd_request = 0; 1821 1822 /* 1823 * Do the flush. Limit the amount of in-transit I/O we 1824 * allow to build up, otherwise we would completely saturate 1825 * the I/O system. Wakeup any waiting processes before we 1826 * normally would so they can run in parallel with our drain. 1827 */ 1828 while (numdirtybuffers > lodirtybuffers) { 1829 if (flushbufqueues() == 0) 1830 break; 1831 waitrunningbufspace(); 1832 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 1833 } 1834 1835 /* 1836 * Only clear bd_request if we have reached our low water 1837 * mark. The buf_daemon normally waits 5 seconds and 1838 * then incrementally flushes any dirty buffers that have 1839 * built up, within reason. 1840 * 1841 * If we were unable to hit our low water mark and couldn't 1842 * find any flushable buffers, we sleep half a second. 1843 * Otherwise we loop immediately. 1844 */ 1845 if (numdirtybuffers <= lodirtybuffers) { 1846 /* 1847 * We reached our low water mark, reset the 1848 * request and sleep until we are needed again. 1849 * The sleep is just so the suspend code works. 1850 */ 1851 bd_request = 0; 1852 tsleep(&bd_request, PVM, "psleep", hz); 1853 } else { 1854 /* 1855 * We couldn't find any flushable dirty buffers but 1856 * still have too many dirty buffers, we 1857 * have to sleep and try again. (rare) 1858 */ 1859 tsleep(&bd_request, PVM, "qsleep", hz / 2); 1860 } 1861 } 1862 } 1863 1864 /* 1865 * flushbufqueues: 1866 * 1867 * Try to flush a buffer in the dirty queue. We must be careful to 1868 * free up B_INVAL buffers instead of write them, which NFS is 1869 * particularly sensitive to. 1870 */ 1871 1872 static int 1873 flushbufqueues(void) 1874 { 1875 struct buf *bp; 1876 int r = 0; 1877 1878 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1879 1880 while (bp) { 1881 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); 1882 if ((bp->b_flags & B_DELWRI) != 0 && 1883 (bp->b_xflags & BX_BKGRDINPROG) == 0) { 1884 if (bp->b_flags & B_INVAL) { 1885 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1886 panic("flushbufqueues: locked buf"); 1887 bremfree(bp); 1888 brelse(bp); 1889 ++r; 1890 break; 1891 } 1892 if (LIST_FIRST(&bp->b_dep) != NULL && 1893 (bp->b_flags & B_DEFERRED) == 0 && 1894 buf_countdeps(bp, 0)) { 1895 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], 1896 bp, b_freelist); 1897 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], 1898 bp, b_freelist); 1899 bp->b_flags |= B_DEFERRED; 1900 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1901 continue; 1902 } 1903 vfs_bio_awrite(bp); 1904 ++r; 1905 break; 1906 } 1907 bp = TAILQ_NEXT(bp, b_freelist); 1908 } 1909 return (r); 1910 } 1911 1912 /* 1913 * Check to see if a block is currently memory resident. 1914 */ 1915 struct buf * 1916 incore(struct vnode * vp, daddr_t blkno) 1917 { 1918 struct buf *bp; 1919 1920 int s = splbio(); 1921 bp = gbincore(vp, blkno); 1922 splx(s); 1923 return (bp); 1924 } 1925 1926 /* 1927 * Returns true if no I/O is needed to access the 1928 * associated VM object. This is like incore except 1929 * it also hunts around in the VM system for the data. 1930 */ 1931 1932 int 1933 inmem(struct vnode * vp, daddr_t blkno) 1934 { 1935 vm_object_t obj; 1936 vm_offset_t toff, tinc, size; 1937 vm_page_t m; 1938 vm_ooffset_t off; 1939 1940 if (incore(vp, blkno)) 1941 return 1; 1942 if (vp->v_mount == NULL) 1943 return 0; 1944 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_flag & VOBJBUF) == 0) 1945 return 0; 1946 1947 size = PAGE_SIZE; 1948 if (size > vp->v_mount->mnt_stat.f_iosize) 1949 size = vp->v_mount->mnt_stat.f_iosize; 1950 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 1951 1952 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1953 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 1954 if (!m) 1955 return 0; 1956 tinc = size; 1957 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 1958 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 1959 if (vm_page_is_valid(m, 1960 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 1961 return 0; 1962 } 1963 return 1; 1964 } 1965 1966 /* 1967 * vfs_setdirty: 1968 * 1969 * Sets the dirty range for a buffer based on the status of the dirty 1970 * bits in the pages comprising the buffer. 1971 * 1972 * The range is limited to the size of the buffer. 1973 * 1974 * This routine is primarily used by NFS, but is generalized for the 1975 * B_VMIO case. 1976 */ 1977 static void 1978 vfs_setdirty(struct buf *bp) 1979 { 1980 int i; 1981 vm_object_t object; 1982 1983 /* 1984 * Degenerate case - empty buffer 1985 */ 1986 1987 if (bp->b_bufsize == 0) 1988 return; 1989 1990 /* 1991 * We qualify the scan for modified pages on whether the 1992 * object has been flushed yet. The OBJ_WRITEABLE flag 1993 * is not cleared simply by protecting pages off. 1994 */ 1995 1996 if ((bp->b_flags & B_VMIO) == 0) 1997 return; 1998 1999 object = bp->b_pages[0]->object; 2000 2001 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2002 printf("Warning: object %p writeable but not mightbedirty\n", object); 2003 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2004 printf("Warning: object %p mightbedirty but not writeable\n", object); 2005 2006 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2007 vm_offset_t boffset; 2008 vm_offset_t eoffset; 2009 2010 /* 2011 * test the pages to see if they have been modified directly 2012 * by users through the VM system. 2013 */ 2014 for (i = 0; i < bp->b_npages; i++) { 2015 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2016 vm_page_test_dirty(bp->b_pages[i]); 2017 } 2018 2019 /* 2020 * Calculate the encompassing dirty range, boffset and eoffset, 2021 * (eoffset - boffset) bytes. 2022 */ 2023 2024 for (i = 0; i < bp->b_npages; i++) { 2025 if (bp->b_pages[i]->dirty) 2026 break; 2027 } 2028 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2029 2030 for (i = bp->b_npages - 1; i >= 0; --i) { 2031 if (bp->b_pages[i]->dirty) { 2032 break; 2033 } 2034 } 2035 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2036 2037 /* 2038 * Fit it to the buffer. 2039 */ 2040 2041 if (eoffset > bp->b_bcount) 2042 eoffset = bp->b_bcount; 2043 2044 /* 2045 * If we have a good dirty range, merge with the existing 2046 * dirty range. 2047 */ 2048 2049 if (boffset < eoffset) { 2050 if (bp->b_dirtyoff > boffset) 2051 bp->b_dirtyoff = boffset; 2052 if (bp->b_dirtyend < eoffset) 2053 bp->b_dirtyend = eoffset; 2054 } 2055 } 2056 } 2057 2058 /* 2059 * getblk: 2060 * 2061 * Get a block given a specified block and offset into a file/device. 2062 * The buffers B_DONE bit will be cleared on return, making it almost 2063 * ready for an I/O initiation. B_INVAL may or may not be set on 2064 * return. The caller should clear B_INVAL prior to initiating a 2065 * READ. 2066 * 2067 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2068 * an existing buffer. 2069 * 2070 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2071 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2072 * and then cleared based on the backing VM. If the previous buffer is 2073 * non-0-sized but invalid, B_CACHE will be cleared. 2074 * 2075 * If getblk() must create a new buffer, the new buffer is returned with 2076 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2077 * case it is returned with B_INVAL clear and B_CACHE set based on the 2078 * backing VM. 2079 * 2080 * getblk() also forces a VOP_BWRITE() for any B_DELWRI buffer whos 2081 * B_CACHE bit is clear. 2082 * 2083 * What this means, basically, is that the caller should use B_CACHE to 2084 * determine whether the buffer is fully valid or not and should clear 2085 * B_INVAL prior to issuing a read. If the caller intends to validate 2086 * the buffer by loading its data area with something, the caller needs 2087 * to clear B_INVAL. If the caller does this without issuing an I/O, 2088 * the caller should set B_CACHE ( as an optimization ), else the caller 2089 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2090 * a write attempt or if it was a successfull read. If the caller 2091 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2092 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2093 */ 2094 struct buf * 2095 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2096 { 2097 struct buf *bp; 2098 int s; 2099 struct bufhashhdr *bh; 2100 2101 if (size > MAXBSIZE) 2102 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2103 2104 s = splbio(); 2105 loop: 2106 /* 2107 * Block if we are low on buffers. Certain processes are allowed 2108 * to completely exhaust the buffer cache. 2109 * 2110 * If this check ever becomes a bottleneck it may be better to 2111 * move it into the else, when gbincore() fails. At the moment 2112 * it isn't a problem. 2113 * 2114 * XXX remove if 0 sections (clean this up after its proven) 2115 */ 2116 if (numfreebuffers == 0) { 2117 if (curproc == PCPU_GET(idleproc)) 2118 return NULL; 2119 needsbuffer |= VFS_BIO_NEED_ANY; 2120 } 2121 2122 if ((bp = gbincore(vp, blkno))) { 2123 /* 2124 * Buffer is in-core. If the buffer is not busy, it must 2125 * be on a queue. 2126 */ 2127 2128 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2129 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, 2130 "getblk", slpflag, slptimeo) == ENOLCK) 2131 goto loop; 2132 splx(s); 2133 return (struct buf *) NULL; 2134 } 2135 2136 /* 2137 * The buffer is locked. B_CACHE is cleared if the buffer is 2138 * invalid. Ohterwise, for a non-VMIO buffer, B_CACHE is set 2139 * and for a VMIO buffer B_CACHE is adjusted according to the 2140 * backing VM cache. 2141 */ 2142 if (bp->b_flags & B_INVAL) 2143 bp->b_flags &= ~B_CACHE; 2144 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2145 bp->b_flags |= B_CACHE; 2146 bremfree(bp); 2147 2148 /* 2149 * check for size inconsistancies for non-VMIO case. 2150 */ 2151 2152 if (bp->b_bcount != size) { 2153 if ((bp->b_flags & B_VMIO) == 0 || 2154 (size > bp->b_kvasize)) { 2155 if (bp->b_flags & B_DELWRI) { 2156 bp->b_flags |= B_NOCACHE; 2157 BUF_WRITE(bp); 2158 } else { 2159 if ((bp->b_flags & B_VMIO) && 2160 (LIST_FIRST(&bp->b_dep) == NULL)) { 2161 bp->b_flags |= B_RELBUF; 2162 brelse(bp); 2163 } else { 2164 bp->b_flags |= B_NOCACHE; 2165 BUF_WRITE(bp); 2166 } 2167 } 2168 goto loop; 2169 } 2170 } 2171 2172 /* 2173 * If the size is inconsistant in the VMIO case, we can resize 2174 * the buffer. This might lead to B_CACHE getting set or 2175 * cleared. If the size has not changed, B_CACHE remains 2176 * unchanged from its previous state. 2177 */ 2178 2179 if (bp->b_bcount != size) 2180 allocbuf(bp, size); 2181 2182 KASSERT(bp->b_offset != NOOFFSET, 2183 ("getblk: no buffer offset")); 2184 2185 /* 2186 * A buffer with B_DELWRI set and B_CACHE clear must 2187 * be committed before we can return the buffer in 2188 * order to prevent the caller from issuing a read 2189 * ( due to B_CACHE not being set ) and overwriting 2190 * it. 2191 * 2192 * Most callers, including NFS and FFS, need this to 2193 * operate properly either because they assume they 2194 * can issue a read if B_CACHE is not set, or because 2195 * ( for example ) an uncached B_DELWRI might loop due 2196 * to softupdates re-dirtying the buffer. In the latter 2197 * case, B_CACHE is set after the first write completes, 2198 * preventing further loops. 2199 */ 2200 2201 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2202 BUF_WRITE(bp); 2203 goto loop; 2204 } 2205 2206 splx(s); 2207 bp->b_flags &= ~B_DONE; 2208 } else { 2209 /* 2210 * Buffer is not in-core, create new buffer. The buffer 2211 * returned by getnewbuf() is locked. Note that the returned 2212 * buffer is also considered valid (not marked B_INVAL). 2213 */ 2214 int bsize, maxsize, vmio; 2215 off_t offset; 2216 2217 if (vn_isdisk(vp, NULL)) 2218 bsize = DEV_BSIZE; 2219 else if (vp->v_mountedhere) 2220 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2221 else if (vp->v_mount) 2222 bsize = vp->v_mount->mnt_stat.f_iosize; 2223 else 2224 bsize = size; 2225 2226 offset = (off_t)blkno * bsize; 2227 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && (vp->v_flag & VOBJBUF); 2228 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2229 maxsize = imax(maxsize, bsize); 2230 2231 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2232 if (slpflag || slptimeo) { 2233 splx(s); 2234 return NULL; 2235 } 2236 goto loop; 2237 } 2238 2239 /* 2240 * This code is used to make sure that a buffer is not 2241 * created while the getnewbuf routine is blocked. 2242 * This can be a problem whether the vnode is locked or not. 2243 * If the buffer is created out from under us, we have to 2244 * throw away the one we just created. There is now window 2245 * race because we are safely running at splbio() from the 2246 * point of the duplicate buffer creation through to here, 2247 * and we've locked the buffer. 2248 */ 2249 if (gbincore(vp, blkno)) { 2250 bp->b_flags |= B_INVAL; 2251 brelse(bp); 2252 goto loop; 2253 } 2254 2255 /* 2256 * Insert the buffer into the hash, so that it can 2257 * be found by incore. 2258 */ 2259 bp->b_blkno = bp->b_lblkno = blkno; 2260 bp->b_offset = offset; 2261 2262 bgetvp(vp, bp); 2263 LIST_REMOVE(bp, b_hash); 2264 bh = bufhash(vp, blkno); 2265 LIST_INSERT_HEAD(bh, bp, b_hash); 2266 2267 /* 2268 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2269 * buffer size starts out as 0, B_CACHE will be set by 2270 * allocbuf() for the VMIO case prior to it testing the 2271 * backing store for validity. 2272 */ 2273 2274 if (vmio) { 2275 bp->b_flags |= B_VMIO; 2276 #if defined(VFS_BIO_DEBUG) 2277 if (vp->v_type != VREG) 2278 printf("getblk: vmioing file type %d???\n", vp->v_type); 2279 #endif 2280 } else { 2281 bp->b_flags &= ~B_VMIO; 2282 } 2283 2284 allocbuf(bp, size); 2285 2286 splx(s); 2287 bp->b_flags &= ~B_DONE; 2288 } 2289 return (bp); 2290 } 2291 2292 /* 2293 * Get an empty, disassociated buffer of given size. The buffer is initially 2294 * set to B_INVAL. 2295 */ 2296 struct buf * 2297 geteblk(int size) 2298 { 2299 struct buf *bp; 2300 int s; 2301 int maxsize; 2302 2303 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2304 2305 s = splbio(); 2306 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0); 2307 splx(s); 2308 allocbuf(bp, size); 2309 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2310 return (bp); 2311 } 2312 2313 2314 /* 2315 * This code constitutes the buffer memory from either anonymous system 2316 * memory (in the case of non-VMIO operations) or from an associated 2317 * VM object (in the case of VMIO operations). This code is able to 2318 * resize a buffer up or down. 2319 * 2320 * Note that this code is tricky, and has many complications to resolve 2321 * deadlock or inconsistant data situations. Tread lightly!!! 2322 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2323 * the caller. Calling this code willy nilly can result in the loss of data. 2324 * 2325 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2326 * B_CACHE for the non-VMIO case. 2327 */ 2328 2329 int 2330 allocbuf(struct buf *bp, int size) 2331 { 2332 int newbsize, mbsize; 2333 int i; 2334 2335 if (BUF_REFCNT(bp) == 0) 2336 panic("allocbuf: buffer not busy"); 2337 2338 if (bp->b_kvasize < size) 2339 panic("allocbuf: buffer too small"); 2340 2341 if ((bp->b_flags & B_VMIO) == 0) { 2342 caddr_t origbuf; 2343 int origbufsize; 2344 /* 2345 * Just get anonymous memory from the kernel. Don't 2346 * mess with B_CACHE. 2347 */ 2348 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2349 #if !defined(NO_B_MALLOC) 2350 if (bp->b_flags & B_MALLOC) 2351 newbsize = mbsize; 2352 else 2353 #endif 2354 newbsize = round_page(size); 2355 2356 if (newbsize < bp->b_bufsize) { 2357 #if !defined(NO_B_MALLOC) 2358 /* 2359 * malloced buffers are not shrunk 2360 */ 2361 if (bp->b_flags & B_MALLOC) { 2362 if (newbsize) { 2363 bp->b_bcount = size; 2364 } else { 2365 free(bp->b_data, M_BIOBUF); 2366 if (bp->b_bufsize) { 2367 bufmallocspace -= bp->b_bufsize; 2368 bufspacewakeup(); 2369 bp->b_bufsize = 0; 2370 } 2371 bp->b_data = bp->b_kvabase; 2372 bp->b_bcount = 0; 2373 bp->b_flags &= ~B_MALLOC; 2374 } 2375 return 1; 2376 } 2377 #endif 2378 vm_hold_free_pages( 2379 bp, 2380 (vm_offset_t) bp->b_data + newbsize, 2381 (vm_offset_t) bp->b_data + bp->b_bufsize); 2382 } else if (newbsize > bp->b_bufsize) { 2383 #if !defined(NO_B_MALLOC) 2384 /* 2385 * We only use malloced memory on the first allocation. 2386 * and revert to page-allocated memory when the buffer 2387 * grows. 2388 */ 2389 if ( (bufmallocspace < maxbufmallocspace) && 2390 (bp->b_bufsize == 0) && 2391 (mbsize <= PAGE_SIZE/2)) { 2392 2393 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2394 bp->b_bufsize = mbsize; 2395 bp->b_bcount = size; 2396 bp->b_flags |= B_MALLOC; 2397 bufmallocspace += mbsize; 2398 return 1; 2399 } 2400 #endif 2401 origbuf = NULL; 2402 origbufsize = 0; 2403 #if !defined(NO_B_MALLOC) 2404 /* 2405 * If the buffer is growing on its other-than-first allocation, 2406 * then we revert to the page-allocation scheme. 2407 */ 2408 if (bp->b_flags & B_MALLOC) { 2409 origbuf = bp->b_data; 2410 origbufsize = bp->b_bufsize; 2411 bp->b_data = bp->b_kvabase; 2412 if (bp->b_bufsize) { 2413 bufmallocspace -= bp->b_bufsize; 2414 bufspacewakeup(); 2415 bp->b_bufsize = 0; 2416 } 2417 bp->b_flags &= ~B_MALLOC; 2418 newbsize = round_page(newbsize); 2419 } 2420 #endif 2421 vm_hold_load_pages( 2422 bp, 2423 (vm_offset_t) bp->b_data + bp->b_bufsize, 2424 (vm_offset_t) bp->b_data + newbsize); 2425 #if !defined(NO_B_MALLOC) 2426 if (origbuf) { 2427 bcopy(origbuf, bp->b_data, origbufsize); 2428 free(origbuf, M_BIOBUF); 2429 } 2430 #endif 2431 } 2432 } else { 2433 vm_page_t m; 2434 int desiredpages; 2435 2436 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2437 desiredpages = (size == 0) ? 0 : 2438 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2439 2440 #if !defined(NO_B_MALLOC) 2441 if (bp->b_flags & B_MALLOC) 2442 panic("allocbuf: VMIO buffer can't be malloced"); 2443 #endif 2444 /* 2445 * Set B_CACHE initially if buffer is 0 length or will become 2446 * 0-length. 2447 */ 2448 if (size == 0 || bp->b_bufsize == 0) 2449 bp->b_flags |= B_CACHE; 2450 2451 if (newbsize < bp->b_bufsize) { 2452 /* 2453 * DEV_BSIZE aligned new buffer size is less then the 2454 * DEV_BSIZE aligned existing buffer size. Figure out 2455 * if we have to remove any pages. 2456 */ 2457 if (desiredpages < bp->b_npages) { 2458 for (i = desiredpages; i < bp->b_npages; i++) { 2459 /* 2460 * the page is not freed here -- it 2461 * is the responsibility of 2462 * vnode_pager_setsize 2463 */ 2464 m = bp->b_pages[i]; 2465 KASSERT(m != bogus_page, 2466 ("allocbuf: bogus page found")); 2467 while (vm_page_sleep_busy(m, TRUE, "biodep")) 2468 ; 2469 2470 bp->b_pages[i] = NULL; 2471 vm_page_unwire(m, 0); 2472 } 2473 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2474 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2475 bp->b_npages = desiredpages; 2476 } 2477 } else if (size > bp->b_bcount) { 2478 /* 2479 * We are growing the buffer, possibly in a 2480 * byte-granular fashion. 2481 */ 2482 struct vnode *vp; 2483 vm_object_t obj; 2484 vm_offset_t toff; 2485 vm_offset_t tinc; 2486 2487 /* 2488 * Step 1, bring in the VM pages from the object, 2489 * allocating them if necessary. We must clear 2490 * B_CACHE if these pages are not valid for the 2491 * range covered by the buffer. 2492 */ 2493 2494 vp = bp->b_vp; 2495 VOP_GETVOBJECT(vp, &obj); 2496 2497 while (bp->b_npages < desiredpages) { 2498 vm_page_t m; 2499 vm_pindex_t pi; 2500 2501 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2502 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2503 /* 2504 * note: must allocate system pages 2505 * since blocking here could intefere 2506 * with paging I/O, no matter which 2507 * process we are. 2508 */ 2509 m = vm_page_alloc(obj, pi, VM_ALLOC_SYSTEM); 2510 if (m == NULL) { 2511 VM_WAIT; 2512 vm_pageout_deficit += desiredpages - bp->b_npages; 2513 } else { 2514 vm_page_wire(m); 2515 vm_page_wakeup(m); 2516 bp->b_flags &= ~B_CACHE; 2517 bp->b_pages[bp->b_npages] = m; 2518 ++bp->b_npages; 2519 } 2520 continue; 2521 } 2522 2523 /* 2524 * We found a page. If we have to sleep on it, 2525 * retry because it might have gotten freed out 2526 * from under us. 2527 * 2528 * We can only test PG_BUSY here. Blocking on 2529 * m->busy might lead to a deadlock: 2530 * 2531 * vm_fault->getpages->cluster_read->allocbuf 2532 * 2533 */ 2534 2535 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 2536 continue; 2537 2538 /* 2539 * We have a good page. Should we wakeup the 2540 * page daemon? 2541 */ 2542 if ((curproc != pageproc) && 2543 ((m->queue - m->pc) == PQ_CACHE) && 2544 ((cnt.v_free_count + cnt.v_cache_count) < 2545 (cnt.v_free_min + cnt.v_cache_min))) { 2546 pagedaemon_wakeup(); 2547 } 2548 vm_page_flag_clear(m, PG_ZERO); 2549 vm_page_wire(m); 2550 bp->b_pages[bp->b_npages] = m; 2551 ++bp->b_npages; 2552 } 2553 2554 /* 2555 * Step 2. We've loaded the pages into the buffer, 2556 * we have to figure out if we can still have B_CACHE 2557 * set. Note that B_CACHE is set according to the 2558 * byte-granular range ( bcount and size ), new the 2559 * aligned range ( newbsize ). 2560 * 2561 * The VM test is against m->valid, which is DEV_BSIZE 2562 * aligned. Needless to say, the validity of the data 2563 * needs to also be DEV_BSIZE aligned. Note that this 2564 * fails with NFS if the server or some other client 2565 * extends the file's EOF. If our buffer is resized, 2566 * B_CACHE may remain set! XXX 2567 */ 2568 2569 toff = bp->b_bcount; 2570 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2571 2572 while ((bp->b_flags & B_CACHE) && toff < size) { 2573 vm_pindex_t pi; 2574 2575 if (tinc > (size - toff)) 2576 tinc = size - toff; 2577 2578 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2579 PAGE_SHIFT; 2580 2581 vfs_buf_test_cache( 2582 bp, 2583 bp->b_offset, 2584 toff, 2585 tinc, 2586 bp->b_pages[pi] 2587 ); 2588 toff += tinc; 2589 tinc = PAGE_SIZE; 2590 } 2591 2592 /* 2593 * Step 3, fixup the KVM pmap. Remember that 2594 * bp->b_data is relative to bp->b_offset, but 2595 * bp->b_offset may be offset into the first page. 2596 */ 2597 2598 bp->b_data = (caddr_t) 2599 trunc_page((vm_offset_t)bp->b_data); 2600 pmap_qenter( 2601 (vm_offset_t)bp->b_data, 2602 bp->b_pages, 2603 bp->b_npages 2604 ); 2605 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2606 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2607 } 2608 } 2609 if (newbsize < bp->b_bufsize) 2610 bufspacewakeup(); 2611 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2612 bp->b_bcount = size; /* requested buffer size */ 2613 return 1; 2614 } 2615 2616 /* 2617 * bufwait: 2618 * 2619 * Wait for buffer I/O completion, returning error status. The buffer 2620 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR 2621 * error and cleared. 2622 */ 2623 int 2624 bufwait(register struct buf * bp) 2625 { 2626 int s; 2627 2628 s = splbio(); 2629 while ((bp->b_flags & B_DONE) == 0) { 2630 if (bp->b_iocmd == BIO_READ) 2631 tsleep(bp, PRIBIO, "biord", 0); 2632 else 2633 tsleep(bp, PRIBIO, "biowr", 0); 2634 } 2635 splx(s); 2636 if (bp->b_flags & B_EINTR) { 2637 bp->b_flags &= ~B_EINTR; 2638 return (EINTR); 2639 } 2640 if (bp->b_ioflags & BIO_ERROR) { 2641 return (bp->b_error ? bp->b_error : EIO); 2642 } else { 2643 return (0); 2644 } 2645 } 2646 2647 /* 2648 * Call back function from struct bio back up to struct buf. 2649 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 2650 */ 2651 void 2652 bufdonebio(struct bio *bp) 2653 { 2654 bufdone(bp->bio_caller2); 2655 } 2656 2657 /* 2658 * bufdone: 2659 * 2660 * Finish I/O on a buffer, optionally calling a completion function. 2661 * This is usually called from an interrupt so process blocking is 2662 * not allowed. 2663 * 2664 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2665 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2666 * assuming B_INVAL is clear. 2667 * 2668 * For the VMIO case, we set B_CACHE if the op was a read and no 2669 * read error occured, or if the op was a write. B_CACHE is never 2670 * set if the buffer is invalid or otherwise uncacheable. 2671 * 2672 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2673 * initiator to leave B_INVAL set to brelse the buffer out of existance 2674 * in the biodone routine. 2675 */ 2676 void 2677 bufdone(struct buf *bp) 2678 { 2679 int s, error; 2680 void (*biodone) __P((struct buf *)); 2681 2682 s = splbio(); 2683 2684 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 2685 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 2686 2687 bp->b_flags |= B_DONE; 2688 runningbufwakeup(bp); 2689 2690 if (bp->b_iocmd == BIO_DELETE) { 2691 brelse(bp); 2692 splx(s); 2693 return; 2694 } 2695 2696 if (bp->b_iocmd == BIO_WRITE) { 2697 vwakeup(bp); 2698 } 2699 2700 /* call optional completion function if requested */ 2701 if (bp->b_iodone != NULL) { 2702 biodone = bp->b_iodone; 2703 bp->b_iodone = NULL; 2704 (*biodone) (bp); 2705 splx(s); 2706 return; 2707 } 2708 if (LIST_FIRST(&bp->b_dep) != NULL) 2709 buf_complete(bp); 2710 2711 if (bp->b_flags & B_VMIO) { 2712 int i; 2713 vm_ooffset_t foff; 2714 vm_page_t m; 2715 vm_object_t obj; 2716 int iosize; 2717 struct vnode *vp = bp->b_vp; 2718 2719 error = VOP_GETVOBJECT(vp, &obj); 2720 2721 #if defined(VFS_BIO_DEBUG) 2722 if (vp->v_usecount == 0) { 2723 panic("biodone: zero vnode ref count"); 2724 } 2725 2726 if (error) { 2727 panic("biodone: missing VM object"); 2728 } 2729 2730 if ((vp->v_flag & VOBJBUF) == 0) { 2731 panic("biodone: vnode is not setup for merged cache"); 2732 } 2733 #endif 2734 2735 foff = bp->b_offset; 2736 KASSERT(bp->b_offset != NOOFFSET, 2737 ("biodone: no buffer offset")); 2738 2739 if (error) { 2740 panic("biodone: no object"); 2741 } 2742 #if defined(VFS_BIO_DEBUG) 2743 if (obj->paging_in_progress < bp->b_npages) { 2744 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 2745 obj->paging_in_progress, bp->b_npages); 2746 } 2747 #endif 2748 2749 /* 2750 * Set B_CACHE if the op was a normal read and no error 2751 * occured. B_CACHE is set for writes in the b*write() 2752 * routines. 2753 */ 2754 iosize = bp->b_bcount - bp->b_resid; 2755 if (bp->b_iocmd == BIO_READ && 2756 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 2757 !(bp->b_ioflags & BIO_ERROR)) { 2758 bp->b_flags |= B_CACHE; 2759 } 2760 2761 for (i = 0; i < bp->b_npages; i++) { 2762 int bogusflag = 0; 2763 int resid; 2764 2765 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2766 if (resid > iosize) 2767 resid = iosize; 2768 2769 /* 2770 * cleanup bogus pages, restoring the originals 2771 */ 2772 m = bp->b_pages[i]; 2773 if (m == bogus_page) { 2774 bogusflag = 1; 2775 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2776 if (m == NULL) 2777 panic("biodone: page disappeared!"); 2778 bp->b_pages[i] = m; 2779 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2780 } 2781 #if defined(VFS_BIO_DEBUG) 2782 if (OFF_TO_IDX(foff) != m->pindex) { 2783 printf( 2784 "biodone: foff(%lu)/m->pindex(%d) mismatch\n", 2785 (unsigned long)foff, m->pindex); 2786 } 2787 #endif 2788 2789 /* 2790 * In the write case, the valid and clean bits are 2791 * already changed correctly ( see bdwrite() ), so we 2792 * only need to do this here in the read case. 2793 */ 2794 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 2795 vfs_page_set_valid(bp, foff, i, m); 2796 } 2797 vm_page_flag_clear(m, PG_ZERO); 2798 2799 /* 2800 * when debugging new filesystems or buffer I/O methods, this 2801 * is the most common error that pops up. if you see this, you 2802 * have not set the page busy flag correctly!!! 2803 */ 2804 if (m->busy == 0) { 2805 printf("biodone: page busy < 0, " 2806 "pindex: %d, foff: 0x(%x,%x), " 2807 "resid: %d, index: %d\n", 2808 (int) m->pindex, (int)(foff >> 32), 2809 (int) foff & 0xffffffff, resid, i); 2810 if (!vn_isdisk(vp, NULL)) 2811 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 2812 bp->b_vp->v_mount->mnt_stat.f_iosize, 2813 (int) bp->b_lblkno, 2814 bp->b_flags, bp->b_npages); 2815 else 2816 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 2817 (int) bp->b_lblkno, 2818 bp->b_flags, bp->b_npages); 2819 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2820 m->valid, m->dirty, m->wire_count); 2821 panic("biodone: page busy < 0\n"); 2822 } 2823 vm_page_io_finish(m); 2824 vm_object_pip_subtract(obj, 1); 2825 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2826 iosize -= resid; 2827 } 2828 if (obj) 2829 vm_object_pip_wakeupn(obj, 0); 2830 } 2831 2832 /* 2833 * For asynchronous completions, release the buffer now. The brelse 2834 * will do a wakeup there if necessary - so no need to do a wakeup 2835 * here in the async case. The sync case always needs to do a wakeup. 2836 */ 2837 2838 if (bp->b_flags & B_ASYNC) { 2839 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 2840 brelse(bp); 2841 else 2842 bqrelse(bp); 2843 } else { 2844 wakeup(bp); 2845 } 2846 splx(s); 2847 } 2848 2849 /* 2850 * This routine is called in lieu of iodone in the case of 2851 * incomplete I/O. This keeps the busy status for pages 2852 * consistant. 2853 */ 2854 void 2855 vfs_unbusy_pages(struct buf * bp) 2856 { 2857 int i; 2858 2859 runningbufwakeup(bp); 2860 if (bp->b_flags & B_VMIO) { 2861 struct vnode *vp = bp->b_vp; 2862 vm_object_t obj; 2863 2864 VOP_GETVOBJECT(vp, &obj); 2865 2866 for (i = 0; i < bp->b_npages; i++) { 2867 vm_page_t m = bp->b_pages[i]; 2868 2869 if (m == bogus_page) { 2870 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 2871 if (!m) { 2872 panic("vfs_unbusy_pages: page missing\n"); 2873 } 2874 bp->b_pages[i] = m; 2875 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2876 } 2877 vm_object_pip_subtract(obj, 1); 2878 vm_page_flag_clear(m, PG_ZERO); 2879 vm_page_io_finish(m); 2880 } 2881 vm_object_pip_wakeupn(obj, 0); 2882 } 2883 } 2884 2885 /* 2886 * vfs_page_set_valid: 2887 * 2888 * Set the valid bits in a page based on the supplied offset. The 2889 * range is restricted to the buffer's size. 2890 * 2891 * This routine is typically called after a read completes. 2892 */ 2893 static void 2894 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 2895 { 2896 vm_ooffset_t soff, eoff; 2897 2898 /* 2899 * Start and end offsets in buffer. eoff - soff may not cross a 2900 * page boundry or cross the end of the buffer. The end of the 2901 * buffer, in this case, is our file EOF, not the allocation size 2902 * of the buffer. 2903 */ 2904 soff = off; 2905 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2906 if (eoff > bp->b_offset + bp->b_bcount) 2907 eoff = bp->b_offset + bp->b_bcount; 2908 2909 /* 2910 * Set valid range. This is typically the entire buffer and thus the 2911 * entire page. 2912 */ 2913 if (eoff > soff) { 2914 vm_page_set_validclean( 2915 m, 2916 (vm_offset_t) (soff & PAGE_MASK), 2917 (vm_offset_t) (eoff - soff) 2918 ); 2919 } 2920 } 2921 2922 /* 2923 * This routine is called before a device strategy routine. 2924 * It is used to tell the VM system that paging I/O is in 2925 * progress, and treat the pages associated with the buffer 2926 * almost as being PG_BUSY. Also the object paging_in_progress 2927 * flag is handled to make sure that the object doesn't become 2928 * inconsistant. 2929 * 2930 * Since I/O has not been initiated yet, certain buffer flags 2931 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 2932 * and should be ignored. 2933 */ 2934 void 2935 vfs_busy_pages(struct buf * bp, int clear_modify) 2936 { 2937 int i, bogus; 2938 2939 bp->b_runningbufspace = bp->b_bufsize; 2940 runningbufspace += bp->b_runningbufspace; 2941 2942 if (bp->b_flags & B_VMIO) { 2943 struct vnode *vp = bp->b_vp; 2944 vm_object_t obj; 2945 vm_ooffset_t foff; 2946 2947 VOP_GETVOBJECT(vp, &obj); 2948 foff = bp->b_offset; 2949 KASSERT(bp->b_offset != NOOFFSET, 2950 ("vfs_busy_pages: no buffer offset")); 2951 vfs_setdirty(bp); 2952 2953 retry: 2954 for (i = 0; i < bp->b_npages; i++) { 2955 vm_page_t m = bp->b_pages[i]; 2956 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 2957 goto retry; 2958 } 2959 2960 bogus = 0; 2961 for (i = 0; i < bp->b_npages; i++) { 2962 vm_page_t m = bp->b_pages[i]; 2963 2964 vm_page_flag_clear(m, PG_ZERO); 2965 if ((bp->b_flags & B_CLUSTER) == 0) { 2966 vm_object_pip_add(obj, 1); 2967 vm_page_io_start(m); 2968 } 2969 2970 /* 2971 * When readying a buffer for a read ( i.e 2972 * clear_modify == 0 ), it is important to do 2973 * bogus_page replacement for valid pages in 2974 * partially instantiated buffers. Partially 2975 * instantiated buffers can, in turn, occur when 2976 * reconstituting a buffer from its VM backing store 2977 * base. We only have to do this if B_CACHE is 2978 * clear ( which causes the I/O to occur in the 2979 * first place ). The replacement prevents the read 2980 * I/O from overwriting potentially dirty VM-backed 2981 * pages. XXX bogus page replacement is, uh, bogus. 2982 * It may not work properly with small-block devices. 2983 * We need to find a better way. 2984 */ 2985 2986 vm_page_protect(m, VM_PROT_NONE); 2987 if (clear_modify) 2988 vfs_page_set_valid(bp, foff, i, m); 2989 else if (m->valid == VM_PAGE_BITS_ALL && 2990 (bp->b_flags & B_CACHE) == 0) { 2991 bp->b_pages[i] = bogus_page; 2992 bogus++; 2993 } 2994 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2995 } 2996 if (bogus) 2997 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2998 } 2999 } 3000 3001 /* 3002 * Tell the VM system that the pages associated with this buffer 3003 * are clean. This is used for delayed writes where the data is 3004 * going to go to disk eventually without additional VM intevention. 3005 * 3006 * Note that while we only really need to clean through to b_bcount, we 3007 * just go ahead and clean through to b_bufsize. 3008 */ 3009 static void 3010 vfs_clean_pages(struct buf * bp) 3011 { 3012 int i; 3013 3014 if (bp->b_flags & B_VMIO) { 3015 vm_ooffset_t foff; 3016 3017 foff = bp->b_offset; 3018 KASSERT(bp->b_offset != NOOFFSET, 3019 ("vfs_clean_pages: no buffer offset")); 3020 for (i = 0; i < bp->b_npages; i++) { 3021 vm_page_t m = bp->b_pages[i]; 3022 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3023 vm_ooffset_t eoff = noff; 3024 3025 if (eoff > bp->b_offset + bp->b_bufsize) 3026 eoff = bp->b_offset + bp->b_bufsize; 3027 vfs_page_set_valid(bp, foff, i, m); 3028 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3029 foff = noff; 3030 } 3031 } 3032 } 3033 3034 /* 3035 * vfs_bio_set_validclean: 3036 * 3037 * Set the range within the buffer to valid and clean. The range is 3038 * relative to the beginning of the buffer, b_offset. Note that b_offset 3039 * itself may be offset from the beginning of the first page. 3040 */ 3041 3042 void 3043 vfs_bio_set_validclean(struct buf *bp, int base, int size) 3044 { 3045 if (bp->b_flags & B_VMIO) { 3046 int i; 3047 int n; 3048 3049 /* 3050 * Fixup base to be relative to beginning of first page. 3051 * Set initial n to be the maximum number of bytes in the 3052 * first page that can be validated. 3053 */ 3054 3055 base += (bp->b_offset & PAGE_MASK); 3056 n = PAGE_SIZE - (base & PAGE_MASK); 3057 3058 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3059 vm_page_t m = bp->b_pages[i]; 3060 3061 if (n > size) 3062 n = size; 3063 3064 vm_page_set_validclean(m, base & PAGE_MASK, n); 3065 base += n; 3066 size -= n; 3067 n = PAGE_SIZE; 3068 } 3069 } 3070 } 3071 3072 /* 3073 * vfs_bio_clrbuf: 3074 * 3075 * clear a buffer. This routine essentially fakes an I/O, so we need 3076 * to clear BIO_ERROR and B_INVAL. 3077 * 3078 * Note that while we only theoretically need to clear through b_bcount, 3079 * we go ahead and clear through b_bufsize. 3080 */ 3081 3082 void 3083 vfs_bio_clrbuf(struct buf *bp) { 3084 int i, mask = 0; 3085 caddr_t sa, ea; 3086 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3087 bp->b_flags &= ~B_INVAL; 3088 bp->b_ioflags &= ~BIO_ERROR; 3089 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3090 (bp->b_offset & PAGE_MASK) == 0) { 3091 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3092 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3093 ((bp->b_pages[0]->valid & mask) != mask)) { 3094 bzero(bp->b_data, bp->b_bufsize); 3095 } 3096 bp->b_pages[0]->valid |= mask; 3097 bp->b_resid = 0; 3098 return; 3099 } 3100 ea = sa = bp->b_data; 3101 for(i=0;i<bp->b_npages;i++,sa=ea) { 3102 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3103 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3104 ea = (caddr_t)(vm_offset_t)ulmin( 3105 (u_long)(vm_offset_t)ea, 3106 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3107 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3108 if ((bp->b_pages[i]->valid & mask) == mask) 3109 continue; 3110 if ((bp->b_pages[i]->valid & mask) == 0) { 3111 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3112 bzero(sa, ea - sa); 3113 } 3114 } else { 3115 for (; sa < ea; sa += DEV_BSIZE, j++) { 3116 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3117 (bp->b_pages[i]->valid & (1<<j)) == 0) 3118 bzero(sa, DEV_BSIZE); 3119 } 3120 } 3121 bp->b_pages[i]->valid |= mask; 3122 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3123 } 3124 bp->b_resid = 0; 3125 } else { 3126 clrbuf(bp); 3127 } 3128 } 3129 3130 /* 3131 * vm_hold_load_pages and vm_hold_unload pages get pages into 3132 * a buffers address space. The pages are anonymous and are 3133 * not associated with a file object. 3134 */ 3135 void 3136 vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3137 { 3138 vm_offset_t pg; 3139 vm_page_t p; 3140 int index; 3141 3142 to = round_page(to); 3143 from = round_page(from); 3144 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3145 3146 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3147 3148 tryagain: 3149 3150 /* 3151 * note: must allocate system pages since blocking here 3152 * could intefere with paging I/O, no matter which 3153 * process we are. 3154 */ 3155 p = vm_page_alloc(kernel_object, 3156 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3157 VM_ALLOC_SYSTEM); 3158 if (!p) { 3159 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3160 VM_WAIT; 3161 goto tryagain; 3162 } 3163 vm_page_wire(p); 3164 p->valid = VM_PAGE_BITS_ALL; 3165 vm_page_flag_clear(p, PG_ZERO); 3166 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 3167 bp->b_pages[index] = p; 3168 vm_page_wakeup(p); 3169 } 3170 bp->b_npages = index; 3171 } 3172 3173 void 3174 vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3175 { 3176 vm_offset_t pg; 3177 vm_page_t p; 3178 int index, newnpages; 3179 3180 from = round_page(from); 3181 to = round_page(to); 3182 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3183 3184 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3185 p = bp->b_pages[index]; 3186 if (p && (index < bp->b_npages)) { 3187 if (p->busy) { 3188 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 3189 bp->b_blkno, bp->b_lblkno); 3190 } 3191 bp->b_pages[index] = NULL; 3192 pmap_kremove(pg); 3193 vm_page_busy(p); 3194 vm_page_unwire(p, 0); 3195 vm_page_free(p); 3196 } 3197 } 3198 bp->b_npages = newnpages; 3199 } 3200 3201 3202 #include "opt_ddb.h" 3203 #ifdef DDB 3204 #include <ddb/ddb.h> 3205 3206 DB_SHOW_COMMAND(buffer, db_show_buffer) 3207 { 3208 /* get args */ 3209 struct buf *bp = (struct buf *)addr; 3210 3211 if (!have_addr) { 3212 db_printf("usage: show buffer <addr>\n"); 3213 return; 3214 } 3215 3216 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3217 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 3218 "b_resid = %ld\nb_dev = (%d,%d), b_data = %p, " 3219 "b_blkno = %d, b_pblkno = %d\n", 3220 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3221 major(bp->b_dev), minor(bp->b_dev), 3222 bp->b_data, bp->b_blkno, bp->b_pblkno); 3223 if (bp->b_npages) { 3224 int i; 3225 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3226 for (i = 0; i < bp->b_npages; i++) { 3227 vm_page_t m; 3228 m = bp->b_pages[i]; 3229 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3230 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3231 if ((i + 1) < bp->b_npages) 3232 db_printf(","); 3233 } 3234 db_printf("\n"); 3235 } 3236 } 3237 #endif /* DDB */ 3238