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/kernel.h> 39 #include <sys/kthread.h> 40 #include <sys/proc.h> 41 #include <sys/reboot.h> 42 #include <sys/resourcevar.h> 43 #include <sys/sysctl.h> 44 #include <sys/vmmeter.h> 45 #include <sys/vnode.h> 46 #include <vm/vm.h> 47 #include <vm/vm_param.h> 48 #include <vm/vm_kern.h> 49 #include <vm/vm_pageout.h> 50 #include <vm/vm_page.h> 51 #include <vm/vm_object.h> 52 #include <vm/vm_extern.h> 53 #include <vm/vm_map.h> 54 55 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 56 57 struct bio_ops bioops; /* I/O operation notification */ 58 59 struct buf *buf; /* buffer header pool */ 60 struct swqueue bswlist; 61 struct simplelock buftimelock; /* Interlock on setting prio and timo */ 62 63 static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 64 vm_offset_t to); 65 static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 66 vm_offset_t to); 67 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 68 int pageno, vm_page_t m); 69 static void vfs_clean_pages(struct buf * bp); 70 static void vfs_setdirty(struct buf *bp); 71 static void vfs_vmio_release(struct buf *bp); 72 static void vfs_backgroundwritedone(struct buf *bp); 73 static int flushbufqueues(void); 74 75 static int bd_request; 76 77 static void buf_daemon __P((void)); 78 /* 79 * bogus page -- for I/O to/from partially complete buffers 80 * this is a temporary solution to the problem, but it is not 81 * really that bad. it would be better to split the buffer 82 * for input in the case of buffers partially already in memory, 83 * but the code is intricate enough already. 84 */ 85 vm_page_t bogus_page; 86 int runningbufspace; 87 int vmiodirenable = FALSE; 88 static vm_offset_t bogus_offset; 89 90 static int bufspace, maxbufspace, 91 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace; 92 static int bufreusecnt, bufdefragcnt, buffreekvacnt; 93 static int maxbdrun; 94 static int needsbuffer; 95 static int numdirtybuffers, hidirtybuffers; 96 static int numfreebuffers, lofreebuffers, hifreebuffers; 97 static int getnewbufcalls; 98 static int getnewbufrestarts; 99 100 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, 101 &numdirtybuffers, 0, ""); 102 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, 103 &hidirtybuffers, 0, ""); 104 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, 105 &numfreebuffers, 0, ""); 106 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, 107 &lofreebuffers, 0, ""); 108 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, 109 &hifreebuffers, 0, ""); 110 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, 111 &runningbufspace, 0, ""); 112 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, 113 &maxbufspace, 0, ""); 114 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, 115 &hibufspace, 0, ""); 116 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, 117 &lobufspace, 0, ""); 118 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, 119 &bufspace, 0, ""); 120 SYSCTL_INT(_vfs, OID_AUTO, maxbdrun, CTLFLAG_RW, 121 &maxbdrun, 0, ""); 122 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, 123 &maxbufmallocspace, 0, ""); 124 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, 125 &bufmallocspace, 0, ""); 126 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, 127 &getnewbufcalls, 0, ""); 128 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, 129 &getnewbufrestarts, 0, ""); 130 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, 131 &vmiodirenable, 0, ""); 132 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, 133 &bufdefragcnt, 0, ""); 134 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, 135 &buffreekvacnt, 0, ""); 136 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, 137 &bufreusecnt, 0, ""); 138 139 static int bufhashmask; 140 static LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash; 141 struct bqueues bufqueues[BUFFER_QUEUES] = { { 0 } }; 142 char *buf_wmesg = BUF_WMESG; 143 144 extern int vm_swap_size; 145 146 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 147 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 148 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 149 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 150 151 /* 152 * Buffer hash table code. Note that the logical block scans linearly, which 153 * gives us some L1 cache locality. 154 */ 155 156 static __inline 157 struct bufhashhdr * 158 bufhash(struct vnode *vnp, daddr_t bn) 159 { 160 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]); 161 } 162 163 /* 164 * numdirtywakeup: 165 * 166 * If someone is blocked due to there being too many dirty buffers, 167 * and numdirtybuffers is now reasonable, wake them up. 168 */ 169 170 static __inline void 171 numdirtywakeup(void) 172 { 173 if (numdirtybuffers < hidirtybuffers) { 174 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 175 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 176 wakeup(&needsbuffer); 177 } 178 } 179 } 180 181 /* 182 * bufspacewakeup: 183 * 184 * Called when buffer space is potentially available for recovery. 185 * getnewbuf() will block on this flag when it is unable to free 186 * sufficient buffer space. Buffer space becomes recoverable when 187 * bp's get placed back in the queues. 188 */ 189 190 static __inline void 191 bufspacewakeup(void) 192 { 193 /* 194 * If someone is waiting for BUF space, wake them up. Even 195 * though we haven't freed the kva space yet, the waiting 196 * process will be able to now. 197 */ 198 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 199 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 200 wakeup(&needsbuffer); 201 } 202 } 203 204 /* 205 * bufcountwakeup: 206 * 207 * Called when a buffer has been added to one of the free queues to 208 * account for the buffer and to wakeup anyone waiting for free buffers. 209 * This typically occurs when large amounts of metadata are being handled 210 * by the buffer cache ( else buffer space runs out first, usually ). 211 */ 212 213 static __inline void 214 bufcountwakeup(void) 215 { 216 ++numfreebuffers; 217 if (needsbuffer) { 218 needsbuffer &= ~VFS_BIO_NEED_ANY; 219 if (numfreebuffers >= hifreebuffers) 220 needsbuffer &= ~VFS_BIO_NEED_FREE; 221 wakeup(&needsbuffer); 222 } 223 } 224 225 /* 226 * vfs_buf_test_cache: 227 * 228 * Called when a buffer is extended. This function clears the B_CACHE 229 * bit if the newly extended portion of the buffer does not contain 230 * valid data. 231 */ 232 static __inline__ 233 void 234 vfs_buf_test_cache(struct buf *bp, 235 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 236 vm_page_t m) 237 { 238 if (bp->b_flags & B_CACHE) { 239 int base = (foff + off) & PAGE_MASK; 240 if (vm_page_is_valid(m, base, size) == 0) 241 bp->b_flags &= ~B_CACHE; 242 } 243 } 244 245 static __inline__ 246 void 247 bd_wakeup(int dirtybuflevel) 248 { 249 if (numdirtybuffers >= dirtybuflevel && bd_request == 0) { 250 bd_request = 1; 251 wakeup(&bd_request); 252 } 253 } 254 255 /* 256 * bd_speedup - speedup the buffer cache flushing code 257 */ 258 259 static __inline__ 260 void 261 bd_speedup(void) 262 { 263 bd_wakeup(1); 264 } 265 266 /* 267 * Initialize buffer headers and related structures. 268 */ 269 270 caddr_t 271 bufhashinit(caddr_t vaddr) 272 { 273 /* first, make a null hash table */ 274 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 275 ; 276 bufhashtbl = (void *)vaddr; 277 vaddr = vaddr + sizeof(*bufhashtbl) * bufhashmask; 278 --bufhashmask; 279 return(vaddr); 280 } 281 282 void 283 bufinit(void) 284 { 285 struct buf *bp; 286 int i; 287 288 TAILQ_INIT(&bswlist); 289 LIST_INIT(&invalhash); 290 simple_lock_init(&buftimelock); 291 292 for (i = 0; i <= bufhashmask; i++) 293 LIST_INIT(&bufhashtbl[i]); 294 295 /* next, make a null set of free lists */ 296 for (i = 0; i < BUFFER_QUEUES; i++) 297 TAILQ_INIT(&bufqueues[i]); 298 299 /* finally, initialize each buffer header and stick on empty q */ 300 for (i = 0; i < nbuf; i++) { 301 bp = &buf[i]; 302 bzero(bp, sizeof *bp); 303 bp->b_flags = B_INVAL; /* we're just an empty header */ 304 bp->b_dev = NODEV; 305 bp->b_rcred = NOCRED; 306 bp->b_wcred = NOCRED; 307 bp->b_qindex = QUEUE_EMPTY; 308 bp->b_xflags = 0; 309 LIST_INIT(&bp->b_dep); 310 BUF_LOCKINIT(bp); 311 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 312 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 313 } 314 315 /* 316 * maxbufspace is the absolute maximum amount of buffer space we are 317 * allowed to reserve in KVM and in real terms. The absolute maximum 318 * is nominally used by buf_daemon. hibufspace is the nominal maximum 319 * used by most other processes. The differential is required to 320 * ensure that buf_daemon is able to run when other processes might 321 * be blocked waiting for buffer space. 322 * 323 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 324 * this may result in KVM fragmentation which is not handled optimally 325 * by the system. 326 */ 327 maxbufspace = nbuf * BKVASIZE; 328 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 329 lobufspace = hibufspace - MAXBSIZE; 330 331 /* 332 * Limit the amount of malloc memory since it is wired permanently into 333 * the kernel space. Even though this is accounted for in the buffer 334 * allocation, we don't want the malloced region to grow uncontrolled. 335 * The malloc scheme improves memory utilization significantly on average 336 * (small) directories. 337 */ 338 maxbufmallocspace = hibufspace / 20; 339 340 /* 341 * Reduce the chance of a deadlock occuring by limiting the number 342 * of delayed-write dirty buffers we allow to stack up. 343 */ 344 hidirtybuffers = nbuf / 4 + 20; 345 numdirtybuffers = 0; 346 /* 347 * To support extreme low-memory systems, make sure hidirtybuffers cannot 348 * eat up all available buffer space. This occurs when our minimum cannot 349 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 350 * BKVASIZE'd (8K) buffers. 351 */ 352 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 353 hidirtybuffers >>= 1; 354 } 355 356 /* 357 * Try to keep the number of free buffers in the specified range, 358 * and give special processes (e.g. like buf_daemon) access to an 359 * emergency reserve. 360 */ 361 lofreebuffers = nbuf / 18 + 5; 362 hifreebuffers = 2 * lofreebuffers; 363 numfreebuffers = nbuf; 364 365 /* 366 * Maximum number of async ops initiated per buf_daemon loop. This is 367 * somewhat of a hack at the moment, we really need to limit ourselves 368 * based on the number of bytes of I/O in-transit that were initiated 369 * from buf_daemon. 370 */ 371 if ((maxbdrun = nswbuf / 4) < 4) 372 maxbdrun = 4; 373 374 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 375 bogus_page = vm_page_alloc(kernel_object, 376 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 377 VM_ALLOC_NORMAL); 378 cnt.v_wire_count++; 379 380 } 381 382 /* 383 * bfreekva() - free the kva allocation for a buffer. 384 * 385 * Must be called at splbio() or higher as this is the only locking for 386 * buffer_map. 387 * 388 * Since this call frees up buffer space, we call bufspacewakeup(). 389 */ 390 static void 391 bfreekva(struct buf * bp) 392 { 393 if (bp->b_kvasize) { 394 ++buffreekvacnt; 395 bufspace -= bp->b_kvasize; 396 vm_map_delete(buffer_map, 397 (vm_offset_t) bp->b_kvabase, 398 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 399 ); 400 bp->b_kvasize = 0; 401 bufspacewakeup(); 402 } 403 } 404 405 /* 406 * bremfree: 407 * 408 * Remove the buffer from the appropriate free list. 409 */ 410 void 411 bremfree(struct buf * bp) 412 { 413 int s = splbio(); 414 int old_qindex = bp->b_qindex; 415 416 if (bp->b_qindex != QUEUE_NONE) { 417 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 418 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 419 bp->b_qindex = QUEUE_NONE; 420 runningbufspace += bp->b_bufsize; 421 } else { 422 if (BUF_REFCNT(bp) <= 1) 423 panic("bremfree: removing a buffer not on a queue"); 424 } 425 426 /* 427 * Fixup numfreebuffers count. If the buffer is invalid or not 428 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 429 * the buffer was free and we must decrement numfreebuffers. 430 */ 431 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 432 switch(old_qindex) { 433 case QUEUE_DIRTY: 434 case QUEUE_CLEAN: 435 case QUEUE_EMPTY: 436 case QUEUE_EMPTYKVA: 437 --numfreebuffers; 438 break; 439 default: 440 break; 441 } 442 } 443 splx(s); 444 } 445 446 447 /* 448 * Get a buffer with the specified data. Look in the cache first. We 449 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 450 * is set, the buffer is valid and we do not have to do anything ( see 451 * getblk() ). 452 */ 453 int 454 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 455 struct buf ** bpp) 456 { 457 struct buf *bp; 458 459 bp = getblk(vp, blkno, size, 0, 0); 460 *bpp = bp; 461 462 /* if not found in cache, do some I/O */ 463 if ((bp->b_flags & B_CACHE) == 0) { 464 if (curproc != NULL) 465 curproc->p_stats->p_ru.ru_inblock++; 466 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp)); 467 bp->b_iocmd = BIO_READ; 468 bp->b_flags &= ~B_INVAL; 469 bp->b_ioflags &= ~BIO_ERROR; 470 if (bp->b_rcred == NOCRED) { 471 if (cred != NOCRED) 472 crhold(cred); 473 bp->b_rcred = cred; 474 } 475 vfs_busy_pages(bp, 0); 476 VOP_STRATEGY(vp, bp); 477 return (bufwait(bp)); 478 } 479 return (0); 480 } 481 482 /* 483 * Operates like bread, but also starts asynchronous I/O on 484 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 485 * to initiating I/O . If B_CACHE is set, the buffer is valid 486 * and we do not have to do anything. 487 */ 488 int 489 breadn(struct vnode * vp, daddr_t blkno, int size, 490 daddr_t * rablkno, int *rabsize, 491 int cnt, struct ucred * cred, struct buf ** bpp) 492 { 493 struct buf *bp, *rabp; 494 int i; 495 int rv = 0, readwait = 0; 496 497 *bpp = bp = getblk(vp, blkno, size, 0, 0); 498 499 /* if not found in cache, do some I/O */ 500 if ((bp->b_flags & B_CACHE) == 0) { 501 if (curproc != NULL) 502 curproc->p_stats->p_ru.ru_inblock++; 503 bp->b_iocmd = BIO_READ; 504 bp->b_flags &= ~B_INVAL; 505 bp->b_ioflags &= ~BIO_ERROR; 506 if (bp->b_rcred == NOCRED) { 507 if (cred != NOCRED) 508 crhold(cred); 509 bp->b_rcred = cred; 510 } 511 vfs_busy_pages(bp, 0); 512 VOP_STRATEGY(vp, bp); 513 ++readwait; 514 } 515 516 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 517 if (inmem(vp, *rablkno)) 518 continue; 519 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 520 521 if ((rabp->b_flags & B_CACHE) == 0) { 522 if (curproc != NULL) 523 curproc->p_stats->p_ru.ru_inblock++; 524 rabp->b_flags |= B_ASYNC; 525 rabp->b_flags &= ~B_INVAL; 526 rabp->b_ioflags &= ~BIO_ERROR; 527 rabp->b_iocmd = BIO_READ; 528 if (rabp->b_rcred == NOCRED) { 529 if (cred != NOCRED) 530 crhold(cred); 531 rabp->b_rcred = cred; 532 } 533 vfs_busy_pages(rabp, 0); 534 BUF_KERNPROC(rabp); 535 VOP_STRATEGY(vp, rabp); 536 } else { 537 brelse(rabp); 538 } 539 } 540 541 if (readwait) { 542 rv = bufwait(bp); 543 } 544 return (rv); 545 } 546 547 /* 548 * Write, release buffer on completion. (Done by iodone 549 * if async). Do not bother writing anything if the buffer 550 * is invalid. 551 * 552 * Note that we set B_CACHE here, indicating that buffer is 553 * fully valid and thus cacheable. This is true even of NFS 554 * now so we set it generally. This could be set either here 555 * or in biodone() since the I/O is synchronous. We put it 556 * here. 557 */ 558 int 559 bwrite(struct buf * bp) 560 { 561 int oldflags, s; 562 struct buf *newbp; 563 564 if (bp->b_flags & B_INVAL) { 565 brelse(bp); 566 return (0); 567 } 568 569 oldflags = bp->b_flags; 570 571 if (BUF_REFCNT(bp) == 0) 572 panic("bwrite: buffer is not busy???"); 573 s = splbio(); 574 /* 575 * If a background write is already in progress, delay 576 * writing this block if it is asynchronous. Otherwise 577 * wait for the background write to complete. 578 */ 579 if (bp->b_xflags & BX_BKGRDINPROG) { 580 if (bp->b_flags & B_ASYNC) { 581 splx(s); 582 bdwrite(bp); 583 return (0); 584 } 585 bp->b_xflags |= BX_BKGRDWAIT; 586 tsleep(&bp->b_xflags, PRIBIO, "biord", 0); 587 if (bp->b_xflags & BX_BKGRDINPROG) 588 panic("bwrite: still writing"); 589 } 590 591 /* Mark the buffer clean */ 592 bundirty(bp); 593 594 /* 595 * If this buffer is marked for background writing and we 596 * do not have to wait for it, make a copy and write the 597 * copy so as to leave this buffer ready for further use. 598 */ 599 if ((bp->b_xflags & BX_BKGRDWRITE) && (bp->b_flags & B_ASYNC)) { 600 if (bp->b_iodone != NULL) { 601 printf("bp->b_iodone = %p\n", bp->b_iodone); 602 panic("bwrite: need chained iodone"); 603 } 604 605 /* get a new block */ 606 newbp = geteblk(bp->b_bufsize); 607 608 /* set it to be identical to the old block */ 609 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 610 bgetvp(bp->b_vp, newbp); 611 newbp->b_lblkno = bp->b_lblkno; 612 newbp->b_blkno = bp->b_blkno; 613 newbp->b_offset = bp->b_offset; 614 newbp->b_iodone = vfs_backgroundwritedone; 615 newbp->b_flags |= B_ASYNC; 616 newbp->b_flags &= ~B_INVAL; 617 618 /* move over the dependencies */ 619 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps) 620 (*bioops.io_movedeps)(bp, newbp); 621 622 /* 623 * Initiate write on the copy, release the original to 624 * the B_LOCKED queue so that it cannot go away until 625 * the background write completes. If not locked it could go 626 * away and then be reconstituted while it was being written. 627 * If the reconstituted buffer were written, we could end up 628 * with two background copies being written at the same time. 629 */ 630 bp->b_xflags |= BX_BKGRDINPROG; 631 bp->b_flags |= B_LOCKED; 632 bqrelse(bp); 633 bp = newbp; 634 } 635 636 bp->b_flags &= ~B_DONE; 637 bp->b_ioflags &= ~BIO_ERROR; 638 bp->b_flags |= B_WRITEINPROG | B_CACHE; 639 bp->b_iocmd = BIO_WRITE; 640 641 bp->b_vp->v_numoutput++; 642 vfs_busy_pages(bp, 1); 643 if (curproc != NULL) 644 curproc->p_stats->p_ru.ru_oublock++; 645 splx(s); 646 if (oldflags & B_ASYNC) 647 BUF_KERNPROC(bp); 648 BUF_STRATEGY(bp); 649 650 if ((oldflags & B_ASYNC) == 0) { 651 int rtval = bufwait(bp); 652 brelse(bp); 653 return (rtval); 654 } 655 656 return (0); 657 } 658 659 /* 660 * Complete a background write started from bwrite. 661 */ 662 static void 663 vfs_backgroundwritedone(bp) 664 struct buf *bp; 665 { 666 struct buf *origbp; 667 668 /* 669 * Find the original buffer that we are writing. 670 */ 671 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 672 panic("backgroundwritedone: lost buffer"); 673 /* 674 * Process dependencies then return any unfinished ones. 675 */ 676 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 677 (*bioops.io_complete)(bp); 678 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps) 679 (*bioops.io_movedeps)(bp, origbp); 680 /* 681 * Clear the BX_BKGRDINPROG flag in the original buffer 682 * and awaken it if it is waiting for the write to complete. 683 */ 684 origbp->b_xflags &= ~BX_BKGRDINPROG; 685 if (origbp->b_xflags & BX_BKGRDWAIT) { 686 origbp->b_xflags &= ~BX_BKGRDWAIT; 687 wakeup(&origbp->b_xflags); 688 } 689 /* 690 * Clear the B_LOCKED flag and remove it from the locked 691 * queue if it currently resides there. 692 */ 693 origbp->b_flags &= ~B_LOCKED; 694 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 695 bremfree(origbp); 696 bqrelse(origbp); 697 } 698 /* 699 * This buffer is marked B_NOCACHE, so when it is released 700 * by biodone, it will be tossed. We mark it with BIO_READ 701 * to avoid biodone doing a second vwakeup. 702 */ 703 bp->b_flags |= B_NOCACHE; 704 bp->b_iocmd = BIO_READ; 705 bp->b_flags &= ~(B_CACHE | B_DONE); 706 bp->b_iodone = 0; 707 bufdone(bp); 708 } 709 710 /* 711 * Delayed write. (Buffer is marked dirty). Do not bother writing 712 * anything if the buffer is marked invalid. 713 * 714 * Note that since the buffer must be completely valid, we can safely 715 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 716 * biodone() in order to prevent getblk from writing the buffer 717 * out synchronously. 718 */ 719 void 720 bdwrite(struct buf * bp) 721 { 722 if (BUF_REFCNT(bp) == 0) 723 panic("bdwrite: buffer is not busy"); 724 725 if (bp->b_flags & B_INVAL) { 726 brelse(bp); 727 return; 728 } 729 bdirty(bp); 730 731 /* 732 * Set B_CACHE, indicating that the buffer is fully valid. This is 733 * true even of NFS now. 734 */ 735 bp->b_flags |= B_CACHE; 736 737 /* 738 * This bmap keeps the system from needing to do the bmap later, 739 * perhaps when the system is attempting to do a sync. Since it 740 * is likely that the indirect block -- or whatever other datastructure 741 * that the filesystem needs is still in memory now, it is a good 742 * thing to do this. Note also, that if the pageout daemon is 743 * requesting a sync -- there might not be enough memory to do 744 * the bmap then... So, this is important to do. 745 */ 746 if (bp->b_lblkno == bp->b_blkno) { 747 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 748 } 749 750 /* 751 * Set the *dirty* buffer range based upon the VM system dirty pages. 752 */ 753 vfs_setdirty(bp); 754 755 /* 756 * We need to do this here to satisfy the vnode_pager and the 757 * pageout daemon, so that it thinks that the pages have been 758 * "cleaned". Note that since the pages are in a delayed write 759 * buffer -- the VFS layer "will" see that the pages get written 760 * out on the next sync, or perhaps the cluster will be completed. 761 */ 762 vfs_clean_pages(bp); 763 bqrelse(bp); 764 765 /* 766 * Wakeup the buffer flushing daemon if we have saturated the 767 * buffer cache. 768 */ 769 770 bd_wakeup(hidirtybuffers); 771 772 /* 773 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 774 * due to the softdep code. 775 */ 776 } 777 778 /* 779 * bdirty: 780 * 781 * Turn buffer into delayed write request. We must clear BIO_READ and 782 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 783 * itself to properly update it in the dirty/clean lists. We mark it 784 * B_DONE to ensure that any asynchronization of the buffer properly 785 * clears B_DONE ( else a panic will occur later ). 786 * 787 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 788 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 789 * should only be called if the buffer is known-good. 790 * 791 * Since the buffer is not on a queue, we do not update the numfreebuffers 792 * count. 793 * 794 * Must be called at splbio(). 795 * The buffer must be on QUEUE_NONE. 796 */ 797 void 798 bdirty(bp) 799 struct buf *bp; 800 { 801 KASSERT(bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 802 bp->b_flags &= ~(B_RELBUF); 803 bp->b_iocmd = BIO_WRITE; 804 805 if ((bp->b_flags & B_DELWRI) == 0) { 806 bp->b_flags |= B_DONE | B_DELWRI; 807 reassignbuf(bp, bp->b_vp); 808 ++numdirtybuffers; 809 bd_wakeup(hidirtybuffers); 810 } 811 } 812 813 /* 814 * bundirty: 815 * 816 * Clear B_DELWRI for buffer. 817 * 818 * Since the buffer is not on a queue, we do not update the numfreebuffers 819 * count. 820 * 821 * Must be called at splbio(). 822 * The buffer must be on QUEUE_NONE. 823 */ 824 825 void 826 bundirty(bp) 827 struct buf *bp; 828 { 829 KASSERT(bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 830 831 if (bp->b_flags & B_DELWRI) { 832 bp->b_flags &= ~B_DELWRI; 833 reassignbuf(bp, bp->b_vp); 834 --numdirtybuffers; 835 numdirtywakeup(); 836 } 837 /* 838 * Since it is now being written, we can clear its deferred write flag. 839 */ 840 bp->b_flags &= ~B_DEFERRED; 841 } 842 843 /* 844 * bawrite: 845 * 846 * Asynchronous write. Start output on a buffer, but do not wait for 847 * it to complete. The buffer is released when the output completes. 848 * 849 * bwrite() ( or the VOP routine anyway ) is responsible for handling 850 * B_INVAL buffers. Not us. 851 */ 852 void 853 bawrite(struct buf * bp) 854 { 855 bp->b_flags |= B_ASYNC; 856 (void) BUF_WRITE(bp); 857 } 858 859 /* 860 * bowrite: 861 * 862 * Ordered write. Start output on a buffer, and flag it so that the 863 * device will write it in the order it was queued. The buffer is 864 * released when the output completes. bwrite() ( or the VOP routine 865 * anyway ) is responsible for handling B_INVAL buffers. 866 */ 867 int 868 bowrite(struct buf * bp) 869 { 870 bp->b_ioflags |= BIO_ORDERED; 871 bp->b_flags |= B_ASYNC; 872 return (BUF_WRITE(bp)); 873 } 874 875 /* 876 * bwillwrite: 877 * 878 * Called prior to the locking of any vnodes when we are expecting to 879 * write. We do not want to starve the buffer cache with too many 880 * dirty buffers so we block here. By blocking prior to the locking 881 * of any vnodes we attempt to avoid the situation where a locked vnode 882 * prevents the various system daemons from flushing related buffers. 883 */ 884 885 void 886 bwillwrite(void) 887 { 888 int slop = hidirtybuffers / 10; 889 890 if (numdirtybuffers > hidirtybuffers + slop) { 891 int s; 892 893 s = splbio(); 894 while (numdirtybuffers > hidirtybuffers) { 895 bd_wakeup(hidirtybuffers); 896 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 897 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0); 898 } 899 splx(s); 900 } 901 } 902 903 /* 904 * brelse: 905 * 906 * Release a busy buffer and, if requested, free its resources. The 907 * buffer will be stashed in the appropriate bufqueue[] allowing it 908 * to be accessed later as a cache entity or reused for other purposes. 909 */ 910 void 911 brelse(struct buf * bp) 912 { 913 int s; 914 915 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 916 917 s = splbio(); 918 919 if (bp->b_flags & B_LOCKED) 920 bp->b_ioflags &= ~BIO_ERROR; 921 922 if (bp->b_iocmd == BIO_WRITE && 923 (bp->b_ioflags & BIO_ERROR) && 924 !(bp->b_flags & B_INVAL)) { 925 /* 926 * Failed write, redirty. Must clear BIO_ERROR to prevent 927 * pages from being scrapped. If B_INVAL is set then 928 * this case is not run and the next case is run to 929 * destroy the buffer. B_INVAL can occur if the buffer 930 * is outside the range supported by the underlying device. 931 */ 932 bp->b_ioflags &= ~BIO_ERROR; 933 bdirty(bp); 934 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 935 (bp->b_ioflags & BIO_ERROR) || 936 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 937 /* 938 * Either a failed I/O or we were asked to free or not 939 * cache the buffer. 940 */ 941 bp->b_flags |= B_INVAL; 942 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 943 (*bioops.io_deallocate)(bp); 944 if (bp->b_flags & B_DELWRI) { 945 --numdirtybuffers; 946 numdirtywakeup(); 947 } 948 bp->b_flags &= ~(B_DELWRI | B_CACHE); 949 if ((bp->b_flags & B_VMIO) == 0) { 950 if (bp->b_bufsize) 951 allocbuf(bp, 0); 952 if (bp->b_vp) 953 brelvp(bp); 954 } 955 } 956 957 /* 958 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 959 * is called with B_DELWRI set, the underlying pages may wind up 960 * getting freed causing a previous write (bdwrite()) to get 'lost' 961 * because pages associated with a B_DELWRI bp are marked clean. 962 * 963 * We still allow the B_INVAL case to call vfs_vmio_release(), even 964 * if B_DELWRI is set. 965 */ 966 967 if (bp->b_flags & B_DELWRI) 968 bp->b_flags &= ~B_RELBUF; 969 970 /* 971 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 972 * constituted, not even NFS buffers now. Two flags effect this. If 973 * B_INVAL, the struct buf is invalidated but the VM object is kept 974 * around ( i.e. so it is trivial to reconstitute the buffer later ). 975 * 976 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 977 * invalidated. BIO_ERROR cannot be set for a failed write unless the 978 * buffer is also B_INVAL because it hits the re-dirtying code above. 979 * 980 * Normally we can do this whether a buffer is B_DELWRI or not. If 981 * the buffer is an NFS buffer, it is tracking piecemeal writes or 982 * the commit state and we cannot afford to lose the buffer. If the 983 * buffer has a background write in progress, we need to keep it 984 * around to prevent it from being reconstituted and starting a second 985 * background write. 986 */ 987 if ((bp->b_flags & B_VMIO) 988 && !(bp->b_vp->v_tag == VT_NFS && 989 !vn_isdisk(bp->b_vp, NULL) && 990 (bp->b_flags & B_DELWRI) && 991 (bp->b_xflags & BX_BKGRDINPROG)) 992 ) { 993 994 int i, j, resid; 995 vm_page_t m; 996 off_t foff; 997 vm_pindex_t poff; 998 vm_object_t obj; 999 struct vnode *vp; 1000 1001 vp = bp->b_vp; 1002 1003 /* 1004 * Get the base offset and length of the buffer. Note that 1005 * for block sizes that are less then PAGE_SIZE, the b_data 1006 * base of the buffer does not represent exactly b_offset and 1007 * neither b_offset nor b_size are necessarily page aligned. 1008 * Instead, the starting position of b_offset is: 1009 * 1010 * b_data + (b_offset & PAGE_MASK) 1011 * 1012 * block sizes less then DEV_BSIZE (usually 512) are not 1013 * supported due to the page granularity bits (m->valid, 1014 * m->dirty, etc...). 1015 * 1016 * See man buf(9) for more information 1017 */ 1018 1019 resid = bp->b_bufsize; 1020 foff = bp->b_offset; 1021 1022 for (i = 0; i < bp->b_npages; i++) { 1023 m = bp->b_pages[i]; 1024 vm_page_flag_clear(m, PG_ZERO); 1025 if (m == bogus_page) { 1026 1027 obj = (vm_object_t) vp->v_object; 1028 poff = OFF_TO_IDX(bp->b_offset); 1029 1030 for (j = i; j < bp->b_npages; j++) { 1031 m = bp->b_pages[j]; 1032 if (m == bogus_page) { 1033 m = vm_page_lookup(obj, poff + j); 1034 if (!m) { 1035 panic("brelse: page missing\n"); 1036 } 1037 bp->b_pages[j] = m; 1038 } 1039 } 1040 1041 if ((bp->b_flags & B_INVAL) == 0) { 1042 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1043 } 1044 } 1045 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1046 int poffset = foff & PAGE_MASK; 1047 int presid = resid > (PAGE_SIZE - poffset) ? 1048 (PAGE_SIZE - poffset) : resid; 1049 1050 KASSERT(presid >= 0, ("brelse: extra page")); 1051 vm_page_set_invalid(m, poffset, presid); 1052 } 1053 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1054 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 1055 } 1056 1057 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1058 vfs_vmio_release(bp); 1059 1060 } else if (bp->b_flags & B_VMIO) { 1061 1062 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1063 vfs_vmio_release(bp); 1064 1065 } 1066 1067 if (bp->b_qindex != QUEUE_NONE) 1068 panic("brelse: free buffer onto another queue???"); 1069 if (BUF_REFCNT(bp) > 1) { 1070 /* Temporary panic to verify exclusive locking */ 1071 /* This panic goes away when we allow shared refs */ 1072 panic("brelse: multiple refs"); 1073 /* do not release to free list */ 1074 BUF_UNLOCK(bp); 1075 splx(s); 1076 return; 1077 } 1078 1079 /* enqueue */ 1080 1081 /* buffers with no memory */ 1082 if (bp->b_bufsize == 0) { 1083 bp->b_flags |= B_INVAL; 1084 bp->b_xflags &= ~BX_BKGRDWRITE; 1085 if (bp->b_xflags & BX_BKGRDINPROG) 1086 panic("losing buffer 1"); 1087 if (bp->b_kvasize) { 1088 bp->b_qindex = QUEUE_EMPTYKVA; 1089 } else { 1090 bp->b_qindex = QUEUE_EMPTY; 1091 } 1092 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1093 LIST_REMOVE(bp, b_hash); 1094 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1095 bp->b_dev = NODEV; 1096 /* buffers with junk contents */ 1097 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || (bp->b_ioflags & BIO_ERROR)) { 1098 bp->b_flags |= B_INVAL; 1099 bp->b_xflags &= ~BX_BKGRDWRITE; 1100 if (bp->b_xflags & BX_BKGRDINPROG) 1101 panic("losing buffer 2"); 1102 bp->b_qindex = QUEUE_CLEAN; 1103 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1104 LIST_REMOVE(bp, b_hash); 1105 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1106 bp->b_dev = NODEV; 1107 1108 /* buffers that are locked */ 1109 } else if (bp->b_flags & B_LOCKED) { 1110 bp->b_qindex = QUEUE_LOCKED; 1111 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1112 1113 /* remaining buffers */ 1114 } else { 1115 switch(bp->b_flags & (B_DELWRI|B_AGE)) { 1116 case B_DELWRI | B_AGE: 1117 bp->b_qindex = QUEUE_DIRTY; 1118 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1119 break; 1120 case B_DELWRI: 1121 bp->b_qindex = QUEUE_DIRTY; 1122 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1123 break; 1124 case B_AGE: 1125 bp->b_qindex = QUEUE_CLEAN; 1126 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1127 break; 1128 default: 1129 bp->b_qindex = QUEUE_CLEAN; 1130 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1131 break; 1132 } 1133 } 1134 1135 /* 1136 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1137 * on the correct queue. 1138 */ 1139 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) { 1140 bp->b_flags &= ~B_DELWRI; 1141 --numdirtybuffers; 1142 numdirtywakeup(); 1143 } 1144 1145 runningbufspace -= bp->b_bufsize; 1146 1147 /* 1148 * Fixup numfreebuffers count. The bp is on an appropriate queue 1149 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1150 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1151 * if B_INVAL is set ). 1152 */ 1153 1154 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1155 bufcountwakeup(); 1156 1157 /* 1158 * Something we can maybe free. 1159 */ 1160 1161 if (bp->b_bufsize || bp->b_kvasize) 1162 bufspacewakeup(); 1163 1164 /* unlock */ 1165 BUF_UNLOCK(bp); 1166 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1167 bp->b_ioflags &= ~BIO_ORDERED; 1168 splx(s); 1169 } 1170 1171 /* 1172 * Release a buffer back to the appropriate queue but do not try to free 1173 * it. 1174 * 1175 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1176 * biodone() to requeue an async I/O on completion. It is also used when 1177 * known good buffers need to be requeued but we think we may need the data 1178 * again soon. 1179 */ 1180 void 1181 bqrelse(struct buf * bp) 1182 { 1183 int s; 1184 1185 s = splbio(); 1186 1187 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1188 1189 if (bp->b_qindex != QUEUE_NONE) 1190 panic("bqrelse: free buffer onto another queue???"); 1191 if (BUF_REFCNT(bp) > 1) { 1192 /* do not release to free list */ 1193 panic("bqrelse: multiple refs"); 1194 BUF_UNLOCK(bp); 1195 splx(s); 1196 return; 1197 } 1198 if (bp->b_flags & B_LOCKED) { 1199 bp->b_ioflags &= ~BIO_ERROR; 1200 bp->b_qindex = QUEUE_LOCKED; 1201 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1202 /* buffers with stale but valid contents */ 1203 } else if (bp->b_flags & B_DELWRI) { 1204 bp->b_qindex = QUEUE_DIRTY; 1205 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1206 } else { 1207 bp->b_qindex = QUEUE_CLEAN; 1208 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1209 } 1210 1211 runningbufspace -= bp->b_bufsize; 1212 1213 if ((bp->b_flags & B_LOCKED) == 0 && 1214 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1215 bufcountwakeup(); 1216 } 1217 1218 /* 1219 * Something we can maybe wakeup 1220 */ 1221 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1222 bufspacewakeup(); 1223 1224 /* unlock */ 1225 BUF_UNLOCK(bp); 1226 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1227 bp->b_ioflags &= ~BIO_ORDERED; 1228 splx(s); 1229 } 1230 1231 static void 1232 vfs_vmio_release(bp) 1233 struct buf *bp; 1234 { 1235 int i, s; 1236 vm_page_t m; 1237 1238 s = splvm(); 1239 for (i = 0; i < bp->b_npages; i++) { 1240 m = bp->b_pages[i]; 1241 bp->b_pages[i] = NULL; 1242 /* 1243 * In order to keep page LRU ordering consistent, put 1244 * everything on the inactive queue. 1245 */ 1246 vm_page_unwire(m, 0); 1247 /* 1248 * We don't mess with busy pages, it is 1249 * the responsibility of the process that 1250 * busied the pages to deal with them. 1251 */ 1252 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1253 continue; 1254 1255 if (m->wire_count == 0) { 1256 vm_page_flag_clear(m, PG_ZERO); 1257 /* 1258 * Might as well free the page if we can and it has 1259 * no valid data. 1260 */ 1261 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { 1262 vm_page_busy(m); 1263 vm_page_protect(m, VM_PROT_NONE); 1264 vm_page_free(m); 1265 } 1266 } 1267 } 1268 runningbufspace -= bp->b_bufsize; 1269 splx(s); 1270 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1271 if (bp->b_bufsize) 1272 bufspacewakeup(); 1273 bp->b_npages = 0; 1274 bp->b_bufsize = 0; 1275 bp->b_flags &= ~B_VMIO; 1276 if (bp->b_vp) 1277 brelvp(bp); 1278 } 1279 1280 /* 1281 * Check to see if a block is currently memory resident. 1282 */ 1283 struct buf * 1284 gbincore(struct vnode * vp, daddr_t blkno) 1285 { 1286 struct buf *bp; 1287 struct bufhashhdr *bh; 1288 1289 bh = bufhash(vp, blkno); 1290 1291 /* Search hash chain */ 1292 LIST_FOREACH(bp, bh, b_hash) { 1293 /* hit */ 1294 if (bp->b_vp == vp && bp->b_lblkno == blkno && 1295 (bp->b_flags & B_INVAL) == 0) { 1296 break; 1297 } 1298 } 1299 return (bp); 1300 } 1301 1302 /* 1303 * vfs_bio_awrite: 1304 * 1305 * Implement clustered async writes for clearing out B_DELWRI buffers. 1306 * This is much better then the old way of writing only one buffer at 1307 * a time. Note that we may not be presented with the buffers in the 1308 * correct order, so we search for the cluster in both directions. 1309 */ 1310 int 1311 vfs_bio_awrite(struct buf * bp) 1312 { 1313 int i; 1314 int j; 1315 daddr_t lblkno = bp->b_lblkno; 1316 struct vnode *vp = bp->b_vp; 1317 int s; 1318 int ncl; 1319 struct buf *bpa; 1320 int nwritten; 1321 int size; 1322 int maxcl; 1323 1324 s = splbio(); 1325 /* 1326 * right now we support clustered writing only to regular files. If 1327 * we find a clusterable block we could be in the middle of a cluster 1328 * rather then at the beginning. 1329 */ 1330 if ((vp->v_type == VREG) && 1331 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1332 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1333 1334 size = vp->v_mount->mnt_stat.f_iosize; 1335 maxcl = MAXPHYS / size; 1336 1337 for (i = 1; i < maxcl; i++) { 1338 if ((bpa = gbincore(vp, lblkno + i)) && 1339 BUF_REFCNT(bpa) == 0 && 1340 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1341 (B_DELWRI | B_CLUSTEROK)) && 1342 (bpa->b_bufsize == size)) { 1343 if ((bpa->b_blkno == bpa->b_lblkno) || 1344 (bpa->b_blkno != 1345 bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 1346 break; 1347 } else { 1348 break; 1349 } 1350 } 1351 for (j = 1; i + j <= maxcl && j <= lblkno; j++) { 1352 if ((bpa = gbincore(vp, lblkno - j)) && 1353 BUF_REFCNT(bpa) == 0 && 1354 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1355 (B_DELWRI | B_CLUSTEROK)) && 1356 (bpa->b_bufsize == size)) { 1357 if ((bpa->b_blkno == bpa->b_lblkno) || 1358 (bpa->b_blkno != 1359 bp->b_blkno - ((j * size) >> DEV_BSHIFT))) 1360 break; 1361 } else { 1362 break; 1363 } 1364 } 1365 --j; 1366 ncl = i + j; 1367 /* 1368 * this is a possible cluster write 1369 */ 1370 if (ncl != 1) { 1371 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1372 splx(s); 1373 return nwritten; 1374 } 1375 } 1376 1377 BUF_LOCK(bp, LK_EXCLUSIVE); 1378 bremfree(bp); 1379 bp->b_flags |= B_ASYNC; 1380 1381 splx(s); 1382 /* 1383 * default (old) behavior, writing out only one block 1384 * 1385 * XXX returns b_bufsize instead of b_bcount for nwritten? 1386 */ 1387 nwritten = bp->b_bufsize; 1388 (void) BUF_WRITE(bp); 1389 1390 return nwritten; 1391 } 1392 1393 /* 1394 * getnewbuf: 1395 * 1396 * Find and initialize a new buffer header, freeing up existing buffers 1397 * in the bufqueues as necessary. The new buffer is returned locked. 1398 * 1399 * Important: B_INVAL is not set. If the caller wishes to throw the 1400 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1401 * 1402 * We block if: 1403 * We have insufficient buffer headers 1404 * We have insufficient buffer space 1405 * buffer_map is too fragmented ( space reservation fails ) 1406 * If we have to flush dirty buffers ( but we try to avoid this ) 1407 * 1408 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1409 * Instead we ask the buf daemon to do it for us. We attempt to 1410 * avoid piecemeal wakeups of the pageout daemon. 1411 */ 1412 1413 static struct buf * 1414 getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1415 { 1416 struct buf *bp; 1417 struct buf *nbp; 1418 int defrag = 0; 1419 int nqindex; 1420 int isspecial; 1421 static int flushingbufs; 1422 1423 if (curproc && (curproc->p_flag & P_BUFEXHAUST) == 0) 1424 isspecial = 0; 1425 else 1426 isspecial = 1; 1427 1428 ++getnewbufcalls; 1429 --getnewbufrestarts; 1430 restart: 1431 ++getnewbufrestarts; 1432 1433 /* 1434 * Setup for scan. If we do not have enough free buffers, 1435 * we setup a degenerate case that immediately fails. Note 1436 * that if we are specially marked process, we are allowed to 1437 * dip into our reserves. 1438 * 1439 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1440 * 1441 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1442 * However, there are a number of cases (defragging, reusing, ...) 1443 * where we cannot backup. 1444 */ 1445 1446 if (isspecial == 0 && numfreebuffers < lofreebuffers) { 1447 /* 1448 * This will cause an immediate failure 1449 */ 1450 nqindex = QUEUE_CLEAN; 1451 nbp = NULL; 1452 } else { 1453 /* 1454 * Locate a buffer which already has KVA assigned. First 1455 * try EMPTYKVA buffers. 1456 */ 1457 nqindex = QUEUE_EMPTYKVA; 1458 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1459 1460 if (nbp == NULL) { 1461 /* 1462 * If no EMPTYKVA buffers and we are either 1463 * defragging or reusing, locate a CLEAN buffer 1464 * to free or reuse. If bufspace useage is low 1465 * skip this step so we can allocate a new buffer. 1466 */ 1467 if (defrag || bufspace >= lobufspace) { 1468 nqindex = QUEUE_CLEAN; 1469 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1470 } 1471 1472 /* 1473 * Nada. If we are allowed to allocate an EMPTY 1474 * buffer, go get one. 1475 */ 1476 if (nbp == NULL && defrag == 0 && 1477 (isspecial || bufspace < hibufspace)) { 1478 nqindex = QUEUE_EMPTY; 1479 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1480 } 1481 } 1482 } 1483 1484 /* 1485 * Run scan, possibly freeing data and/or kva mappings on the fly 1486 * depending. 1487 */ 1488 1489 while ((bp = nbp) != NULL) { 1490 int qindex = nqindex; 1491 1492 /* 1493 * Calculate next bp ( we can only use it if we do not block 1494 * or do other fancy things ). 1495 */ 1496 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1497 switch(qindex) { 1498 case QUEUE_EMPTY: 1499 nqindex = QUEUE_EMPTYKVA; 1500 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1501 break; 1502 /* fall through */ 1503 case QUEUE_EMPTYKVA: 1504 nqindex = QUEUE_CLEAN; 1505 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1506 break; 1507 /* fall through */ 1508 case QUEUE_CLEAN: 1509 /* 1510 * nbp is NULL. 1511 */ 1512 break; 1513 } 1514 } 1515 1516 /* 1517 * Sanity Checks 1518 */ 1519 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1520 1521 /* 1522 * Note: we no longer distinguish between VMIO and non-VMIO 1523 * buffers. 1524 */ 1525 1526 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1527 1528 /* 1529 * If we are defragging then we need a buffer with 1530 * b_kvasize != 0. XXX this situation should no longer 1531 * occur, if defrag is non-zero the buffer's b_kvasize 1532 * should also be non-zero at this point. XXX 1533 */ 1534 if (defrag && bp->b_kvasize == 0) { 1535 printf("Warning: defrag empty buffer %p\n", bp); 1536 continue; 1537 } 1538 1539 /* 1540 * Start freeing the bp. This is somewhat involved. nbp 1541 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1542 */ 1543 1544 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1545 panic("getnewbuf: locked buf"); 1546 bremfree(bp); 1547 1548 if (qindex == QUEUE_CLEAN) { 1549 if (bp->b_flags & B_VMIO) { 1550 bp->b_flags &= ~B_ASYNC; 1551 vfs_vmio_release(bp); 1552 } 1553 if (bp->b_vp) 1554 brelvp(bp); 1555 } 1556 1557 /* 1558 * NOTE: nbp is now entirely invalid. We can only restart 1559 * the scan from this point on. 1560 * 1561 * Get the rest of the buffer freed up. b_kva* is still 1562 * valid after this operation. 1563 */ 1564 1565 if (bp->b_rcred != NOCRED) { 1566 crfree(bp->b_rcred); 1567 bp->b_rcred = NOCRED; 1568 } 1569 if (bp->b_wcred != NOCRED) { 1570 crfree(bp->b_wcred); 1571 bp->b_wcred = NOCRED; 1572 } 1573 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 1574 (*bioops.io_deallocate)(bp); 1575 if (bp->b_xflags & BX_BKGRDINPROG) 1576 panic("losing buffer 3"); 1577 LIST_REMOVE(bp, b_hash); 1578 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1579 1580 if (bp->b_bufsize) 1581 allocbuf(bp, 0); 1582 1583 bp->b_flags = 0; 1584 bp->b_ioflags = 0; 1585 bp->b_xflags = 0; 1586 bp->b_dev = NODEV; 1587 bp->b_vp = NULL; 1588 bp->b_blkno = bp->b_lblkno = 0; 1589 bp->b_offset = NOOFFSET; 1590 bp->b_iodone = 0; 1591 bp->b_error = 0; 1592 bp->b_resid = 0; 1593 bp->b_bcount = 0; 1594 bp->b_npages = 0; 1595 bp->b_dirtyoff = bp->b_dirtyend = 0; 1596 1597 LIST_INIT(&bp->b_dep); 1598 1599 /* 1600 * If we are defragging then free the buffer. 1601 */ 1602 if (defrag) { 1603 bp->b_flags |= B_INVAL; 1604 bfreekva(bp); 1605 brelse(bp); 1606 defrag = 0; 1607 goto restart; 1608 } 1609 1610 /* 1611 * If we are a normal process then deal with bufspace 1612 * hysteresis. A normal process tries to keep bufspace 1613 * between lobufspace and hibufspace. Note: if we encounter 1614 * a buffer with b_kvasize == 0 then it means we started 1615 * our scan on the EMPTY list and should allocate a new 1616 * buffer. 1617 */ 1618 if (isspecial == 0) { 1619 if (bufspace > hibufspace) 1620 flushingbufs = 1; 1621 if (flushingbufs && bp->b_kvasize != 0) { 1622 bp->b_flags |= B_INVAL; 1623 bfreekva(bp); 1624 brelse(bp); 1625 goto restart; 1626 } 1627 if (bufspace < lobufspace) 1628 flushingbufs = 0; 1629 } 1630 break; 1631 } 1632 1633 /* 1634 * If we exhausted our list, sleep as appropriate. We may have to 1635 * wakeup various daemons and write out some dirty buffers. 1636 * 1637 * Generally we are sleeping due to insufficient buffer space. 1638 */ 1639 1640 if (bp == NULL) { 1641 int flags; 1642 char *waitmsg; 1643 1644 if (defrag) { 1645 flags = VFS_BIO_NEED_BUFSPACE; 1646 waitmsg = "nbufkv"; 1647 } else if (bufspace >= hibufspace) { 1648 waitmsg = "nbufbs"; 1649 flags = VFS_BIO_NEED_BUFSPACE; 1650 } else { 1651 waitmsg = "newbuf"; 1652 flags = VFS_BIO_NEED_ANY; 1653 } 1654 1655 bd_speedup(); /* heeeelp */ 1656 1657 needsbuffer |= flags; 1658 while (needsbuffer & flags) { 1659 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, 1660 waitmsg, slptimeo)) 1661 return (NULL); 1662 } 1663 } else { 1664 /* 1665 * We finally have a valid bp. We aren't quite out of the 1666 * woods, we still have to reserve kva space. In order 1667 * to keep fragmentation sane we only allocate kva in 1668 * BKVASIZE chunks. 1669 */ 1670 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1671 1672 if (maxsize != bp->b_kvasize) { 1673 vm_offset_t addr = 0; 1674 1675 bfreekva(bp); 1676 1677 if (vm_map_findspace(buffer_map, 1678 vm_map_min(buffer_map), maxsize, &addr)) { 1679 /* 1680 * Uh oh. Buffer map is to fragmented. We 1681 * must defragment the map. 1682 */ 1683 ++bufdefragcnt; 1684 defrag = 1; 1685 bp->b_flags |= B_INVAL; 1686 brelse(bp); 1687 goto restart; 1688 } 1689 if (addr) { 1690 vm_map_insert(buffer_map, NULL, 0, 1691 addr, addr + maxsize, 1692 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1693 1694 bp->b_kvabase = (caddr_t) addr; 1695 bp->b_kvasize = maxsize; 1696 bufspace += bp->b_kvasize; 1697 ++bufreusecnt; 1698 } 1699 } 1700 bp->b_data = bp->b_kvabase; 1701 } 1702 return(bp); 1703 } 1704 1705 /* 1706 * waitfreebuffers: 1707 * 1708 * Wait for sufficient free buffers. Only called from normal processes. 1709 */ 1710 1711 static void 1712 waitfreebuffers(int slpflag, int slptimeo) 1713 { 1714 while (numfreebuffers < hifreebuffers) { 1715 if (numfreebuffers >= hifreebuffers) 1716 break; 1717 needsbuffer |= VFS_BIO_NEED_FREE; 1718 if (tsleep(&needsbuffer, (PRIBIO + 4)|slpflag, "biofre", slptimeo)) 1719 break; 1720 } 1721 } 1722 1723 /* 1724 * buf_daemon: 1725 * 1726 * buffer flushing daemon. Buffers are normally flushed by the 1727 * update daemon but if it cannot keep up this process starts to 1728 * take the load in an attempt to prevent getnewbuf() from blocking. 1729 */ 1730 1731 static struct proc *bufdaemonproc; 1732 static int bd_interval; 1733 static int bd_flushto; 1734 static int bd_flushinc; 1735 1736 static struct kproc_desc buf_kp = { 1737 "bufdaemon", 1738 buf_daemon, 1739 &bufdaemonproc 1740 }; 1741 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1742 1743 static void 1744 buf_daemon() 1745 { 1746 int s; 1747 1748 /* 1749 * This process needs to be suspended prior to shutdown sync. 1750 */ 1751 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, bufdaemonproc, 1752 SHUTDOWN_PRI_LAST); 1753 1754 /* 1755 * This process is allowed to take the buffer cache to the limit 1756 */ 1757 curproc->p_flag |= P_BUFEXHAUST; 1758 s = splbio(); 1759 1760 bd_interval = 5 * hz; /* dynamically adjusted */ 1761 bd_flushto = hidirtybuffers; /* dynamically adjusted */ 1762 bd_flushinc = 1; 1763 1764 for (;;) { 1765 kproc_suspend_loop(bufdaemonproc); 1766 1767 bd_request = 0; 1768 1769 /* 1770 * Do the flush. Limit the number of buffers we flush in one 1771 * go. The failure condition occurs when processes are writing 1772 * buffers faster then we can dispose of them. In this case 1773 * we may be flushing so often that the previous set of flushes 1774 * have not had time to complete, causing us to run out of 1775 * physical buffers and block. 1776 */ 1777 { 1778 int runcount = maxbdrun; 1779 1780 while (numdirtybuffers > bd_flushto && runcount) { 1781 --runcount; 1782 if (flushbufqueues() == 0) 1783 break; 1784 } 1785 } 1786 1787 if (bd_request || 1788 tsleep(&bd_request, PVM, "psleep", bd_interval) == 0) { 1789 /* 1790 * Another request is pending or we were woken up 1791 * without timing out. Flush more. 1792 */ 1793 --bd_flushto; 1794 if (bd_flushto >= numdirtybuffers - 5) { 1795 bd_flushto = numdirtybuffers - 10; 1796 bd_flushinc = 1; 1797 } 1798 if (bd_flushto < 2) 1799 bd_flushto = 2; 1800 } else { 1801 /* 1802 * We slept and timed out, we can slow down. 1803 */ 1804 bd_flushto += bd_flushinc; 1805 if (bd_flushto > hidirtybuffers) 1806 bd_flushto = hidirtybuffers; 1807 ++bd_flushinc; 1808 if (bd_flushinc > hidirtybuffers / 20 + 1) 1809 bd_flushinc = hidirtybuffers / 20 + 1; 1810 } 1811 1812 /* 1813 * Set the interval on a linear scale based on hidirtybuffers 1814 * with a maximum frequency of 1/10 second. 1815 */ 1816 bd_interval = bd_flushto * 5 * hz / hidirtybuffers; 1817 if (bd_interval < hz / 10) 1818 bd_interval = hz / 10; 1819 } 1820 } 1821 1822 /* 1823 * flushbufqueues: 1824 * 1825 * Try to flush a buffer in the dirty queue. We must be careful to 1826 * free up B_INVAL buffers instead of write them, which NFS is 1827 * particularly sensitive to. 1828 */ 1829 1830 static int 1831 flushbufqueues(void) 1832 { 1833 struct buf *bp; 1834 int r = 0; 1835 1836 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1837 1838 while (bp) { 1839 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); 1840 if ((bp->b_flags & B_DELWRI) != 0 && 1841 (bp->b_xflags & BX_BKGRDINPROG) == 0) { 1842 if (bp->b_flags & B_INVAL) { 1843 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1844 panic("flushbufqueues: locked buf"); 1845 bremfree(bp); 1846 brelse(bp); 1847 ++r; 1848 break; 1849 } 1850 if (LIST_FIRST(&bp->b_dep) != NULL && 1851 bioops.io_countdeps && 1852 (bp->b_flags & B_DEFERRED) == 0 && 1853 (*bioops.io_countdeps)(bp, 0)) { 1854 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], 1855 bp, b_freelist); 1856 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], 1857 bp, b_freelist); 1858 bp->b_flags |= B_DEFERRED; 1859 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1860 continue; 1861 } 1862 vfs_bio_awrite(bp); 1863 ++r; 1864 break; 1865 } 1866 bp = TAILQ_NEXT(bp, b_freelist); 1867 } 1868 return (r); 1869 } 1870 1871 /* 1872 * Check to see if a block is currently memory resident. 1873 */ 1874 struct buf * 1875 incore(struct vnode * vp, daddr_t blkno) 1876 { 1877 struct buf *bp; 1878 1879 int s = splbio(); 1880 bp = gbincore(vp, blkno); 1881 splx(s); 1882 return (bp); 1883 } 1884 1885 /* 1886 * Returns true if no I/O is needed to access the 1887 * associated VM object. This is like incore except 1888 * it also hunts around in the VM system for the data. 1889 */ 1890 1891 int 1892 inmem(struct vnode * vp, daddr_t blkno) 1893 { 1894 vm_object_t obj; 1895 vm_offset_t toff, tinc, size; 1896 vm_page_t m; 1897 vm_ooffset_t off; 1898 1899 if (incore(vp, blkno)) 1900 return 1; 1901 if (vp->v_mount == NULL) 1902 return 0; 1903 if ((vp->v_object == NULL) || (vp->v_flag & VOBJBUF) == 0) 1904 return 0; 1905 1906 obj = vp->v_object; 1907 size = PAGE_SIZE; 1908 if (size > vp->v_mount->mnt_stat.f_iosize) 1909 size = vp->v_mount->mnt_stat.f_iosize; 1910 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 1911 1912 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1913 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 1914 if (!m) 1915 return 0; 1916 tinc = size; 1917 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 1918 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 1919 if (vm_page_is_valid(m, 1920 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 1921 return 0; 1922 } 1923 return 1; 1924 } 1925 1926 /* 1927 * vfs_setdirty: 1928 * 1929 * Sets the dirty range for a buffer based on the status of the dirty 1930 * bits in the pages comprising the buffer. 1931 * 1932 * The range is limited to the size of the buffer. 1933 * 1934 * This routine is primarily used by NFS, but is generalized for the 1935 * B_VMIO case. 1936 */ 1937 static void 1938 vfs_setdirty(struct buf *bp) 1939 { 1940 int i; 1941 vm_object_t object; 1942 1943 /* 1944 * Degenerate case - empty buffer 1945 */ 1946 1947 if (bp->b_bufsize == 0) 1948 return; 1949 1950 /* 1951 * We qualify the scan for modified pages on whether the 1952 * object has been flushed yet. The OBJ_WRITEABLE flag 1953 * is not cleared simply by protecting pages off. 1954 */ 1955 1956 if ((bp->b_flags & B_VMIO) == 0) 1957 return; 1958 1959 object = bp->b_pages[0]->object; 1960 1961 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 1962 printf("Warning: object %p writeable but not mightbedirty\n", object); 1963 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 1964 printf("Warning: object %p mightbedirty but not writeable\n", object); 1965 1966 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 1967 vm_offset_t boffset; 1968 vm_offset_t eoffset; 1969 1970 /* 1971 * test the pages to see if they have been modified directly 1972 * by users through the VM system. 1973 */ 1974 for (i = 0; i < bp->b_npages; i++) { 1975 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 1976 vm_page_test_dirty(bp->b_pages[i]); 1977 } 1978 1979 /* 1980 * Calculate the encompassing dirty range, boffset and eoffset, 1981 * (eoffset - boffset) bytes. 1982 */ 1983 1984 for (i = 0; i < bp->b_npages; i++) { 1985 if (bp->b_pages[i]->dirty) 1986 break; 1987 } 1988 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 1989 1990 for (i = bp->b_npages - 1; i >= 0; --i) { 1991 if (bp->b_pages[i]->dirty) { 1992 break; 1993 } 1994 } 1995 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 1996 1997 /* 1998 * Fit it to the buffer. 1999 */ 2000 2001 if (eoffset > bp->b_bcount) 2002 eoffset = bp->b_bcount; 2003 2004 /* 2005 * If we have a good dirty range, merge with the existing 2006 * dirty range. 2007 */ 2008 2009 if (boffset < eoffset) { 2010 if (bp->b_dirtyoff > boffset) 2011 bp->b_dirtyoff = boffset; 2012 if (bp->b_dirtyend < eoffset) 2013 bp->b_dirtyend = eoffset; 2014 } 2015 } 2016 } 2017 2018 /* 2019 * getblk: 2020 * 2021 * Get a block given a specified block and offset into a file/device. 2022 * The buffers B_DONE bit will be cleared on return, making it almost 2023 * ready for an I/O initiation. B_INVAL may or may not be set on 2024 * return. The caller should clear B_INVAL prior to initiating a 2025 * READ. 2026 * 2027 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2028 * an existing buffer. 2029 * 2030 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2031 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2032 * and then cleared based on the backing VM. If the previous buffer is 2033 * non-0-sized but invalid, B_CACHE will be cleared. 2034 * 2035 * If getblk() must create a new buffer, the new buffer is returned with 2036 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2037 * case it is returned with B_INVAL clear and B_CACHE set based on the 2038 * backing VM. 2039 * 2040 * getblk() also forces a VOP_BWRITE() for any B_DELWRI buffer whos 2041 * B_CACHE bit is clear. 2042 * 2043 * What this means, basically, is that the caller should use B_CACHE to 2044 * determine whether the buffer is fully valid or not and should clear 2045 * B_INVAL prior to issuing a read. If the caller intends to validate 2046 * the buffer by loading its data area with something, the caller needs 2047 * to clear B_INVAL. If the caller does this without issuing an I/O, 2048 * the caller should set B_CACHE ( as an optimization ), else the caller 2049 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2050 * a write attempt or if it was a successfull read. If the caller 2051 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2052 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2053 */ 2054 struct buf * 2055 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2056 { 2057 struct buf *bp; 2058 int s; 2059 struct bufhashhdr *bh; 2060 2061 if (size > MAXBSIZE) 2062 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2063 2064 s = splbio(); 2065 loop: 2066 /* 2067 * Block if we are low on buffers. Certain processes are allowed 2068 * to completely exhaust the buffer cache. 2069 * 2070 * If this check ever becomes a bottleneck it may be better to 2071 * move it into the else, when gbincore() fails. At the moment 2072 * it isn't a problem. 2073 */ 2074 if (!curproc || (curproc->p_flag & P_BUFEXHAUST)) { 2075 if (numfreebuffers == 0) { 2076 if (!curproc) 2077 return NULL; 2078 needsbuffer |= VFS_BIO_NEED_ANY; 2079 tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, "newbuf", 2080 slptimeo); 2081 } 2082 } else if (numfreebuffers < lofreebuffers) { 2083 waitfreebuffers(slpflag, slptimeo); 2084 } 2085 2086 if ((bp = gbincore(vp, blkno))) { 2087 /* 2088 * Buffer is in-core. If the buffer is not busy, it must 2089 * be on a queue. 2090 */ 2091 2092 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2093 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, 2094 "getblk", slpflag, slptimeo) == ENOLCK) 2095 goto loop; 2096 splx(s); 2097 return (struct buf *) NULL; 2098 } 2099 2100 /* 2101 * The buffer is locked. B_CACHE is cleared if the buffer is 2102 * invalid. Ohterwise, for a non-VMIO buffer, B_CACHE is set 2103 * and for a VMIO buffer B_CACHE is adjusted according to the 2104 * backing VM cache. 2105 */ 2106 if (bp->b_flags & B_INVAL) 2107 bp->b_flags &= ~B_CACHE; 2108 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2109 bp->b_flags |= B_CACHE; 2110 bremfree(bp); 2111 2112 /* 2113 * check for size inconsistancies for non-VMIO case. 2114 */ 2115 2116 if (bp->b_bcount != size) { 2117 if ((bp->b_flags & B_VMIO) == 0 || 2118 (size > bp->b_kvasize)) { 2119 if (bp->b_flags & B_DELWRI) { 2120 bp->b_flags |= B_NOCACHE; 2121 BUF_WRITE(bp); 2122 } else { 2123 if ((bp->b_flags & B_VMIO) && 2124 (LIST_FIRST(&bp->b_dep) == NULL)) { 2125 bp->b_flags |= B_RELBUF; 2126 brelse(bp); 2127 } else { 2128 bp->b_flags |= B_NOCACHE; 2129 BUF_WRITE(bp); 2130 } 2131 } 2132 goto loop; 2133 } 2134 } 2135 2136 /* 2137 * If the size is inconsistant in the VMIO case, we can resize 2138 * the buffer. This might lead to B_CACHE getting set or 2139 * cleared. If the size has not changed, B_CACHE remains 2140 * unchanged from its previous state. 2141 */ 2142 2143 if (bp->b_bcount != size) 2144 allocbuf(bp, size); 2145 2146 KASSERT(bp->b_offset != NOOFFSET, 2147 ("getblk: no buffer offset")); 2148 2149 /* 2150 * A buffer with B_DELWRI set and B_CACHE clear must 2151 * be committed before we can return the buffer in 2152 * order to prevent the caller from issuing a read 2153 * ( due to B_CACHE not being set ) and overwriting 2154 * it. 2155 * 2156 * Most callers, including NFS and FFS, need this to 2157 * operate properly either because they assume they 2158 * can issue a read if B_CACHE is not set, or because 2159 * ( for example ) an uncached B_DELWRI might loop due 2160 * to softupdates re-dirtying the buffer. In the latter 2161 * case, B_CACHE is set after the first write completes, 2162 * preventing further loops. 2163 */ 2164 2165 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2166 BUF_WRITE(bp); 2167 goto loop; 2168 } 2169 2170 splx(s); 2171 bp->b_flags &= ~B_DONE; 2172 } else { 2173 /* 2174 * Buffer is not in-core, create new buffer. The buffer 2175 * returned by getnewbuf() is locked. Note that the returned 2176 * buffer is also considered valid (not marked B_INVAL). 2177 */ 2178 int bsize, maxsize, vmio; 2179 off_t offset; 2180 2181 if (vn_isdisk(vp, NULL)) 2182 bsize = DEV_BSIZE; 2183 else if (vp->v_mountedhere) 2184 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2185 else if (vp->v_mount) 2186 bsize = vp->v_mount->mnt_stat.f_iosize; 2187 else 2188 bsize = size; 2189 2190 offset = (off_t)blkno * bsize; 2191 vmio = (vp->v_object != 0) && (vp->v_flag & VOBJBUF); 2192 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2193 maxsize = imax(maxsize, bsize); 2194 2195 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2196 if (slpflag || slptimeo) { 2197 splx(s); 2198 return NULL; 2199 } 2200 goto loop; 2201 } 2202 2203 /* 2204 * This code is used to make sure that a buffer is not 2205 * created while the getnewbuf routine is blocked. 2206 * This can be a problem whether the vnode is locked or not. 2207 * If the buffer is created out from under us, we have to 2208 * throw away the one we just created. There is now window 2209 * race because we are safely running at splbio() from the 2210 * point of the duplicate buffer creation through to here, 2211 * and we've locked the buffer. 2212 */ 2213 if (gbincore(vp, blkno)) { 2214 bp->b_flags |= B_INVAL; 2215 brelse(bp); 2216 goto loop; 2217 } 2218 2219 /* 2220 * Insert the buffer into the hash, so that it can 2221 * be found by incore. 2222 */ 2223 bp->b_blkno = bp->b_lblkno = blkno; 2224 bp->b_offset = offset; 2225 2226 bgetvp(vp, bp); 2227 LIST_REMOVE(bp, b_hash); 2228 bh = bufhash(vp, blkno); 2229 LIST_INSERT_HEAD(bh, bp, b_hash); 2230 2231 /* 2232 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2233 * buffer size starts out as 0, B_CACHE will be set by 2234 * allocbuf() for the VMIO case prior to it testing the 2235 * backing store for validity. 2236 */ 2237 2238 if (vmio) { 2239 bp->b_flags |= B_VMIO; 2240 #if defined(VFS_BIO_DEBUG) 2241 if (vp->v_type != VREG && vp->v_type != VBLK) 2242 printf("getblk: vmioing file type %d???\n", vp->v_type); 2243 #endif 2244 } else { 2245 bp->b_flags &= ~B_VMIO; 2246 } 2247 2248 allocbuf(bp, size); 2249 2250 splx(s); 2251 bp->b_flags &= ~B_DONE; 2252 } 2253 return (bp); 2254 } 2255 2256 /* 2257 * Get an empty, disassociated buffer of given size. The buffer is initially 2258 * set to B_INVAL. 2259 */ 2260 struct buf * 2261 geteblk(int size) 2262 { 2263 struct buf *bp; 2264 int s; 2265 int maxsize; 2266 2267 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2268 2269 s = splbio(); 2270 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0); 2271 splx(s); 2272 allocbuf(bp, size); 2273 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2274 return (bp); 2275 } 2276 2277 2278 /* 2279 * This code constitutes the buffer memory from either anonymous system 2280 * memory (in the case of non-VMIO operations) or from an associated 2281 * VM object (in the case of VMIO operations). This code is able to 2282 * resize a buffer up or down. 2283 * 2284 * Note that this code is tricky, and has many complications to resolve 2285 * deadlock or inconsistant data situations. Tread lightly!!! 2286 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2287 * the caller. Calling this code willy nilly can result in the loss of data. 2288 * 2289 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2290 * B_CACHE for the non-VMIO case. 2291 */ 2292 2293 int 2294 allocbuf(struct buf *bp, int size) 2295 { 2296 int newbsize, mbsize; 2297 int i; 2298 2299 if (BUF_REFCNT(bp) == 0) 2300 panic("allocbuf: buffer not busy"); 2301 2302 if (bp->b_kvasize < size) 2303 panic("allocbuf: buffer too small"); 2304 2305 if ((bp->b_flags & B_VMIO) == 0) { 2306 caddr_t origbuf; 2307 int origbufsize; 2308 /* 2309 * Just get anonymous memory from the kernel. Don't 2310 * mess with B_CACHE. 2311 */ 2312 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2313 #if !defined(NO_B_MALLOC) 2314 if (bp->b_flags & B_MALLOC) 2315 newbsize = mbsize; 2316 else 2317 #endif 2318 newbsize = round_page(size); 2319 2320 if (newbsize < bp->b_bufsize) { 2321 #if !defined(NO_B_MALLOC) 2322 /* 2323 * malloced buffers are not shrunk 2324 */ 2325 if (bp->b_flags & B_MALLOC) { 2326 if (newbsize) { 2327 bp->b_bcount = size; 2328 } else { 2329 free(bp->b_data, M_BIOBUF); 2330 bufmallocspace -= bp->b_bufsize; 2331 runningbufspace -= bp->b_bufsize; 2332 if (bp->b_bufsize) 2333 bufspacewakeup(); 2334 bp->b_data = bp->b_kvabase; 2335 bp->b_bufsize = 0; 2336 bp->b_bcount = 0; 2337 bp->b_flags &= ~B_MALLOC; 2338 } 2339 return 1; 2340 } 2341 #endif 2342 vm_hold_free_pages( 2343 bp, 2344 (vm_offset_t) bp->b_data + newbsize, 2345 (vm_offset_t) bp->b_data + bp->b_bufsize); 2346 } else if (newbsize > bp->b_bufsize) { 2347 #if !defined(NO_B_MALLOC) 2348 /* 2349 * We only use malloced memory on the first allocation. 2350 * and revert to page-allocated memory when the buffer 2351 * grows. 2352 */ 2353 if ( (bufmallocspace < maxbufmallocspace) && 2354 (bp->b_bufsize == 0) && 2355 (mbsize <= PAGE_SIZE/2)) { 2356 2357 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2358 bp->b_bufsize = mbsize; 2359 bp->b_bcount = size; 2360 bp->b_flags |= B_MALLOC; 2361 bufmallocspace += mbsize; 2362 runningbufspace += bp->b_bufsize; 2363 return 1; 2364 } 2365 #endif 2366 origbuf = NULL; 2367 origbufsize = 0; 2368 #if !defined(NO_B_MALLOC) 2369 /* 2370 * If the buffer is growing on its other-than-first allocation, 2371 * then we revert to the page-allocation scheme. 2372 */ 2373 if (bp->b_flags & B_MALLOC) { 2374 origbuf = bp->b_data; 2375 origbufsize = bp->b_bufsize; 2376 bp->b_data = bp->b_kvabase; 2377 bufmallocspace -= bp->b_bufsize; 2378 runningbufspace -= bp->b_bufsize; 2379 if (bp->b_bufsize) 2380 bufspacewakeup(); 2381 bp->b_bufsize = 0; 2382 bp->b_flags &= ~B_MALLOC; 2383 newbsize = round_page(newbsize); 2384 } 2385 #endif 2386 vm_hold_load_pages( 2387 bp, 2388 (vm_offset_t) bp->b_data + bp->b_bufsize, 2389 (vm_offset_t) bp->b_data + newbsize); 2390 #if !defined(NO_B_MALLOC) 2391 if (origbuf) { 2392 bcopy(origbuf, bp->b_data, origbufsize); 2393 free(origbuf, M_BIOBUF); 2394 } 2395 #endif 2396 } 2397 } else { 2398 vm_page_t m; 2399 int desiredpages; 2400 2401 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2402 desiredpages = (size == 0) ? 0 : 2403 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2404 2405 #if !defined(NO_B_MALLOC) 2406 if (bp->b_flags & B_MALLOC) 2407 panic("allocbuf: VMIO buffer can't be malloced"); 2408 #endif 2409 /* 2410 * Set B_CACHE initially if buffer is 0 length or will become 2411 * 0-length. 2412 */ 2413 if (size == 0 || bp->b_bufsize == 0) 2414 bp->b_flags |= B_CACHE; 2415 2416 if (newbsize < bp->b_bufsize) { 2417 /* 2418 * DEV_BSIZE aligned new buffer size is less then the 2419 * DEV_BSIZE aligned existing buffer size. Figure out 2420 * if we have to remove any pages. 2421 */ 2422 if (desiredpages < bp->b_npages) { 2423 for (i = desiredpages; i < bp->b_npages; i++) { 2424 /* 2425 * the page is not freed here -- it 2426 * is the responsibility of 2427 * vnode_pager_setsize 2428 */ 2429 m = bp->b_pages[i]; 2430 KASSERT(m != bogus_page, 2431 ("allocbuf: bogus page found")); 2432 while (vm_page_sleep_busy(m, TRUE, "biodep")) 2433 ; 2434 2435 bp->b_pages[i] = NULL; 2436 vm_page_unwire(m, 0); 2437 } 2438 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2439 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2440 bp->b_npages = desiredpages; 2441 } 2442 } else if (size > bp->b_bcount) { 2443 /* 2444 * We are growing the buffer, possibly in a 2445 * byte-granular fashion. 2446 */ 2447 struct vnode *vp; 2448 vm_object_t obj; 2449 vm_offset_t toff; 2450 vm_offset_t tinc; 2451 2452 /* 2453 * Step 1, bring in the VM pages from the object, 2454 * allocating them if necessary. We must clear 2455 * B_CACHE if these pages are not valid for the 2456 * range covered by the buffer. 2457 */ 2458 2459 vp = bp->b_vp; 2460 obj = vp->v_object; 2461 2462 while (bp->b_npages < desiredpages) { 2463 vm_page_t m; 2464 vm_pindex_t pi; 2465 2466 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2467 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2468 m = vm_page_alloc(obj, pi, VM_ALLOC_NORMAL); 2469 if (m == NULL) { 2470 VM_WAIT; 2471 vm_pageout_deficit += desiredpages - bp->b_npages; 2472 } else { 2473 vm_page_wire(m); 2474 vm_page_wakeup(m); 2475 bp->b_flags &= ~B_CACHE; 2476 bp->b_pages[bp->b_npages] = m; 2477 ++bp->b_npages; 2478 } 2479 continue; 2480 } 2481 2482 /* 2483 * We found a page. If we have to sleep on it, 2484 * retry because it might have gotten freed out 2485 * from under us. 2486 * 2487 * We can only test PG_BUSY here. Blocking on 2488 * m->busy might lead to a deadlock: 2489 * 2490 * vm_fault->getpages->cluster_read->allocbuf 2491 * 2492 */ 2493 2494 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 2495 continue; 2496 2497 /* 2498 * We have a good page. Should we wakeup the 2499 * page daemon? 2500 */ 2501 if ((curproc != pageproc) && 2502 ((m->queue - m->pc) == PQ_CACHE) && 2503 ((cnt.v_free_count + cnt.v_cache_count) < 2504 (cnt.v_free_min + cnt.v_cache_min))) { 2505 pagedaemon_wakeup(); 2506 } 2507 vm_page_flag_clear(m, PG_ZERO); 2508 vm_page_wire(m); 2509 bp->b_pages[bp->b_npages] = m; 2510 ++bp->b_npages; 2511 } 2512 2513 /* 2514 * Step 2. We've loaded the pages into the buffer, 2515 * we have to figure out if we can still have B_CACHE 2516 * set. Note that B_CACHE is set according to the 2517 * byte-granular range ( bcount and size ), new the 2518 * aligned range ( newbsize ). 2519 * 2520 * The VM test is against m->valid, which is DEV_BSIZE 2521 * aligned. Needless to say, the validity of the data 2522 * needs to also be DEV_BSIZE aligned. Note that this 2523 * fails with NFS if the server or some other client 2524 * extends the file's EOF. If our buffer is resized, 2525 * B_CACHE may remain set! XXX 2526 */ 2527 2528 toff = bp->b_bcount; 2529 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2530 2531 while ((bp->b_flags & B_CACHE) && toff < size) { 2532 vm_pindex_t pi; 2533 2534 if (tinc > (size - toff)) 2535 tinc = size - toff; 2536 2537 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2538 PAGE_SHIFT; 2539 2540 vfs_buf_test_cache( 2541 bp, 2542 bp->b_offset, 2543 toff, 2544 tinc, 2545 bp->b_pages[pi] 2546 ); 2547 toff += tinc; 2548 tinc = PAGE_SIZE; 2549 } 2550 2551 /* 2552 * Step 3, fixup the KVM pmap. Remember that 2553 * bp->b_data is relative to bp->b_offset, but 2554 * bp->b_offset may be offset into the first page. 2555 */ 2556 2557 bp->b_data = (caddr_t) 2558 trunc_page((vm_offset_t)bp->b_data); 2559 pmap_qenter( 2560 (vm_offset_t)bp->b_data, 2561 bp->b_pages, 2562 bp->b_npages 2563 ); 2564 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2565 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2566 } 2567 } 2568 runningbufspace += (newbsize - bp->b_bufsize); 2569 if (newbsize < bp->b_bufsize) 2570 bufspacewakeup(); 2571 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2572 bp->b_bcount = size; /* requested buffer size */ 2573 return 1; 2574 } 2575 2576 /* 2577 * bufwait: 2578 * 2579 * Wait for buffer I/O completion, returning error status. The buffer 2580 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR 2581 * error and cleared. 2582 */ 2583 int 2584 bufwait(register struct buf * bp) 2585 { 2586 int s; 2587 2588 s = splbio(); 2589 while ((bp->b_flags & B_DONE) == 0) { 2590 if (bp->b_iocmd == BIO_READ) 2591 tsleep(bp, PRIBIO, "biord", 0); 2592 else 2593 tsleep(bp, PRIBIO, "biowr", 0); 2594 } 2595 splx(s); 2596 if (bp->b_flags & B_EINTR) { 2597 bp->b_flags &= ~B_EINTR; 2598 return (EINTR); 2599 } 2600 if (bp->b_ioflags & BIO_ERROR) { 2601 return (bp->b_error ? bp->b_error : EIO); 2602 } else { 2603 return (0); 2604 } 2605 } 2606 2607 /* 2608 * Call back function from struct bio back up to struct buf. 2609 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 2610 */ 2611 void 2612 bufdonebio(struct bio *bp) 2613 { 2614 bufdone(bp->bio_caller2); 2615 } 2616 2617 /* 2618 * bufdone: 2619 * 2620 * Finish I/O on a buffer, optionally calling a completion function. 2621 * This is usually called from an interrupt so process blocking is 2622 * not allowed. 2623 * 2624 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2625 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2626 * assuming B_INVAL is clear. 2627 * 2628 * For the VMIO case, we set B_CACHE if the op was a read and no 2629 * read error occured, or if the op was a write. B_CACHE is never 2630 * set if the buffer is invalid or otherwise uncacheable. 2631 * 2632 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2633 * initiator to leave B_INVAL set to brelse the buffer out of existance 2634 * in the biodone routine. 2635 */ 2636 void 2637 bufdone(struct buf *bp) 2638 { 2639 int s; 2640 void (*biodone) __P((struct buf *)); 2641 2642 s = splbio(); 2643 2644 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 2645 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 2646 2647 bp->b_flags |= B_DONE; 2648 2649 if (bp->b_iocmd == BIO_DELETE) { 2650 brelse(bp); 2651 splx(s); 2652 return; 2653 } 2654 2655 if (bp->b_iocmd == BIO_WRITE) { 2656 vwakeup(bp); 2657 } 2658 2659 /* call optional completion function if requested */ 2660 if (bp->b_iodone != NULL) { 2661 biodone = bp->b_iodone; 2662 bp->b_iodone = NULL; 2663 (*biodone) (bp); 2664 splx(s); 2665 return; 2666 } 2667 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 2668 (*bioops.io_complete)(bp); 2669 2670 if (bp->b_flags & B_VMIO) { 2671 int i, resid; 2672 vm_ooffset_t foff; 2673 vm_page_t m; 2674 vm_object_t obj; 2675 int iosize; 2676 struct vnode *vp = bp->b_vp; 2677 2678 obj = vp->v_object; 2679 2680 #if defined(VFS_BIO_DEBUG) 2681 if (vp->v_usecount == 0) { 2682 panic("biodone: zero vnode ref count"); 2683 } 2684 2685 if (vp->v_object == NULL) { 2686 panic("biodone: missing VM object"); 2687 } 2688 2689 if ((vp->v_flag & VOBJBUF) == 0) { 2690 panic("biodone: vnode is not setup for merged cache"); 2691 } 2692 #endif 2693 2694 foff = bp->b_offset; 2695 KASSERT(bp->b_offset != NOOFFSET, 2696 ("biodone: no buffer offset")); 2697 2698 if (!obj) { 2699 panic("biodone: no object"); 2700 } 2701 #if defined(VFS_BIO_DEBUG) 2702 if (obj->paging_in_progress < bp->b_npages) { 2703 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 2704 obj->paging_in_progress, bp->b_npages); 2705 } 2706 #endif 2707 2708 /* 2709 * Set B_CACHE if the op was a normal read and no error 2710 * occured. B_CACHE is set for writes in the b*write() 2711 * routines. 2712 */ 2713 iosize = bp->b_bcount - bp->b_resid; 2714 if (bp->b_iocmd == BIO_READ && 2715 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 2716 !(bp->b_ioflags & BIO_ERROR)) { 2717 bp->b_flags |= B_CACHE; 2718 } 2719 2720 for (i = 0; i < bp->b_npages; i++) { 2721 int bogusflag = 0; 2722 m = bp->b_pages[i]; 2723 if (m == bogus_page) { 2724 bogusflag = 1; 2725 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2726 if (!m) { 2727 #if defined(VFS_BIO_DEBUG) 2728 printf("biodone: page disappeared\n"); 2729 #endif 2730 vm_object_pip_subtract(obj, 1); 2731 bp->b_flags &= ~B_CACHE; 2732 continue; 2733 } 2734 bp->b_pages[i] = m; 2735 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2736 } 2737 #if defined(VFS_BIO_DEBUG) 2738 if (OFF_TO_IDX(foff) != m->pindex) { 2739 printf( 2740 "biodone: foff(%lu)/m->pindex(%d) mismatch\n", 2741 (unsigned long)foff, m->pindex); 2742 } 2743 #endif 2744 resid = IDX_TO_OFF(m->pindex + 1) - foff; 2745 if (resid > iosize) 2746 resid = iosize; 2747 2748 /* 2749 * In the write case, the valid and clean bits are 2750 * already changed correctly ( see bdwrite() ), so we 2751 * only need to do this here in the read case. 2752 */ 2753 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 2754 vfs_page_set_valid(bp, foff, i, m); 2755 } 2756 vm_page_flag_clear(m, PG_ZERO); 2757 2758 /* 2759 * when debugging new filesystems or buffer I/O methods, this 2760 * is the most common error that pops up. if you see this, you 2761 * have not set the page busy flag correctly!!! 2762 */ 2763 if (m->busy == 0) { 2764 printf("biodone: page busy < 0, " 2765 "pindex: %d, foff: 0x(%x,%x), " 2766 "resid: %d, index: %d\n", 2767 (int) m->pindex, (int)(foff >> 32), 2768 (int) foff & 0xffffffff, resid, i); 2769 if (!vn_isdisk(vp, NULL)) 2770 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 2771 bp->b_vp->v_mount->mnt_stat.f_iosize, 2772 (int) bp->b_lblkno, 2773 bp->b_flags, bp->b_npages); 2774 else 2775 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 2776 (int) bp->b_lblkno, 2777 bp->b_flags, bp->b_npages); 2778 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2779 m->valid, m->dirty, m->wire_count); 2780 panic("biodone: page busy < 0\n"); 2781 } 2782 vm_page_io_finish(m); 2783 vm_object_pip_subtract(obj, 1); 2784 foff += resid; 2785 iosize -= resid; 2786 } 2787 if (obj) 2788 vm_object_pip_wakeupn(obj, 0); 2789 } 2790 /* 2791 * For asynchronous completions, release the buffer now. The brelse 2792 * will do a wakeup there if necessary - so no need to do a wakeup 2793 * here in the async case. The sync case always needs to do a wakeup. 2794 */ 2795 2796 if (bp->b_flags & B_ASYNC) { 2797 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 2798 brelse(bp); 2799 else 2800 bqrelse(bp); 2801 } else { 2802 wakeup(bp); 2803 } 2804 splx(s); 2805 } 2806 2807 /* 2808 * This routine is called in lieu of iodone in the case of 2809 * incomplete I/O. This keeps the busy status for pages 2810 * consistant. 2811 */ 2812 void 2813 vfs_unbusy_pages(struct buf * bp) 2814 { 2815 int i; 2816 2817 if (bp->b_flags & B_VMIO) { 2818 struct vnode *vp = bp->b_vp; 2819 vm_object_t obj = vp->v_object; 2820 2821 for (i = 0; i < bp->b_npages; i++) { 2822 vm_page_t m = bp->b_pages[i]; 2823 2824 if (m == bogus_page) { 2825 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 2826 if (!m) { 2827 panic("vfs_unbusy_pages: page missing\n"); 2828 } 2829 bp->b_pages[i] = m; 2830 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2831 } 2832 vm_object_pip_subtract(obj, 1); 2833 vm_page_flag_clear(m, PG_ZERO); 2834 vm_page_io_finish(m); 2835 } 2836 vm_object_pip_wakeupn(obj, 0); 2837 } 2838 } 2839 2840 /* 2841 * vfs_page_set_valid: 2842 * 2843 * Set the valid bits in a page based on the supplied offset. The 2844 * range is restricted to the buffer's size. 2845 * 2846 * This routine is typically called after a read completes. 2847 */ 2848 static void 2849 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 2850 { 2851 vm_ooffset_t soff, eoff; 2852 2853 /* 2854 * Start and end offsets in buffer. eoff - soff may not cross a 2855 * page boundry or cross the end of the buffer. The end of the 2856 * buffer, in this case, is our file EOF, not the allocation size 2857 * of the buffer. 2858 */ 2859 soff = off; 2860 eoff = (off + PAGE_SIZE) & ~PAGE_MASK; 2861 if (eoff > bp->b_offset + bp->b_bcount) 2862 eoff = bp->b_offset + bp->b_bcount; 2863 2864 /* 2865 * Set valid range. This is typically the entire buffer and thus the 2866 * entire page. 2867 */ 2868 if (eoff > soff) { 2869 vm_page_set_validclean( 2870 m, 2871 (vm_offset_t) (soff & PAGE_MASK), 2872 (vm_offset_t) (eoff - soff) 2873 ); 2874 } 2875 } 2876 2877 /* 2878 * This routine is called before a device strategy routine. 2879 * It is used to tell the VM system that paging I/O is in 2880 * progress, and treat the pages associated with the buffer 2881 * almost as being PG_BUSY. Also the object paging_in_progress 2882 * flag is handled to make sure that the object doesn't become 2883 * inconsistant. 2884 * 2885 * Since I/O has not been initiated yet, certain buffer flags 2886 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 2887 * and should be ignored. 2888 */ 2889 void 2890 vfs_busy_pages(struct buf * bp, int clear_modify) 2891 { 2892 int i, bogus; 2893 2894 if (bp->b_flags & B_VMIO) { 2895 struct vnode *vp = bp->b_vp; 2896 vm_object_t obj = vp->v_object; 2897 vm_ooffset_t foff; 2898 2899 foff = bp->b_offset; 2900 KASSERT(bp->b_offset != NOOFFSET, 2901 ("vfs_busy_pages: no buffer offset")); 2902 vfs_setdirty(bp); 2903 2904 retry: 2905 for (i = 0; i < bp->b_npages; i++) { 2906 vm_page_t m = bp->b_pages[i]; 2907 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 2908 goto retry; 2909 } 2910 2911 bogus = 0; 2912 for (i = 0; i < bp->b_npages; i++) { 2913 vm_page_t m = bp->b_pages[i]; 2914 2915 vm_page_flag_clear(m, PG_ZERO); 2916 if ((bp->b_flags & B_CLUSTER) == 0) { 2917 vm_object_pip_add(obj, 1); 2918 vm_page_io_start(m); 2919 } 2920 2921 /* 2922 * When readying a buffer for a read ( i.e 2923 * clear_modify == 0 ), it is important to do 2924 * bogus_page replacement for valid pages in 2925 * partially instantiated buffers. Partially 2926 * instantiated buffers can, in turn, occur when 2927 * reconstituting a buffer from its VM backing store 2928 * base. We only have to do this if B_CACHE is 2929 * clear ( which causes the I/O to occur in the 2930 * first place ). The replacement prevents the read 2931 * I/O from overwriting potentially dirty VM-backed 2932 * pages. XXX bogus page replacement is, uh, bogus. 2933 * It may not work properly with small-block devices. 2934 * We need to find a better way. 2935 */ 2936 2937 vm_page_protect(m, VM_PROT_NONE); 2938 if (clear_modify) 2939 vfs_page_set_valid(bp, foff, i, m); 2940 else if (m->valid == VM_PAGE_BITS_ALL && 2941 (bp->b_flags & B_CACHE) == 0) { 2942 bp->b_pages[i] = bogus_page; 2943 bogus++; 2944 } 2945 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 2946 } 2947 if (bogus) 2948 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2949 } 2950 } 2951 2952 /* 2953 * Tell the VM system that the pages associated with this buffer 2954 * are clean. This is used for delayed writes where the data is 2955 * going to go to disk eventually without additional VM intevention. 2956 * 2957 * Note that while we only really need to clean through to b_bcount, we 2958 * just go ahead and clean through to b_bufsize. 2959 */ 2960 static void 2961 vfs_clean_pages(struct buf * bp) 2962 { 2963 int i; 2964 2965 if (bp->b_flags & B_VMIO) { 2966 vm_ooffset_t foff; 2967 2968 foff = bp->b_offset; 2969 KASSERT(bp->b_offset != NOOFFSET, 2970 ("vfs_clean_pages: no buffer offset")); 2971 for (i = 0; i < bp->b_npages; i++) { 2972 vm_page_t m = bp->b_pages[i]; 2973 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~PAGE_MASK; 2974 vm_ooffset_t eoff = noff; 2975 2976 if (eoff > bp->b_offset + bp->b_bufsize) 2977 eoff = bp->b_offset + bp->b_bufsize; 2978 vfs_page_set_valid(bp, foff, i, m); 2979 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2980 foff = noff; 2981 } 2982 } 2983 } 2984 2985 /* 2986 * vfs_bio_set_validclean: 2987 * 2988 * Set the range within the buffer to valid and clean. The range is 2989 * relative to the beginning of the buffer, b_offset. Note that b_offset 2990 * itself may be offset from the beginning of the first page. 2991 */ 2992 2993 void 2994 vfs_bio_set_validclean(struct buf *bp, int base, int size) 2995 { 2996 if (bp->b_flags & B_VMIO) { 2997 int i; 2998 int n; 2999 3000 /* 3001 * Fixup base to be relative to beginning of first page. 3002 * Set initial n to be the maximum number of bytes in the 3003 * first page that can be validated. 3004 */ 3005 3006 base += (bp->b_offset & PAGE_MASK); 3007 n = PAGE_SIZE - (base & PAGE_MASK); 3008 3009 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3010 vm_page_t m = bp->b_pages[i]; 3011 3012 if (n > size) 3013 n = size; 3014 3015 vm_page_set_validclean(m, base & PAGE_MASK, n); 3016 base += n; 3017 size -= n; 3018 n = PAGE_SIZE; 3019 } 3020 } 3021 } 3022 3023 /* 3024 * vfs_bio_clrbuf: 3025 * 3026 * clear a buffer. This routine essentially fakes an I/O, so we need 3027 * to clear BIO_ERROR and B_INVAL. 3028 * 3029 * Note that while we only theoretically need to clear through b_bcount, 3030 * we go ahead and clear through b_bufsize. 3031 */ 3032 3033 void 3034 vfs_bio_clrbuf(struct buf *bp) { 3035 int i, mask = 0; 3036 caddr_t sa, ea; 3037 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3038 bp->b_flags &= ~B_INVAL; 3039 bp->b_ioflags &= ~BIO_ERROR; 3040 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3041 (bp->b_offset & PAGE_MASK) == 0) { 3042 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3043 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3044 ((bp->b_pages[0]->valid & mask) != mask)) { 3045 bzero(bp->b_data, bp->b_bufsize); 3046 } 3047 bp->b_pages[0]->valid |= mask; 3048 bp->b_resid = 0; 3049 return; 3050 } 3051 ea = sa = bp->b_data; 3052 for(i=0;i<bp->b_npages;i++,sa=ea) { 3053 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3054 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3055 ea = (caddr_t)(vm_offset_t)ulmin( 3056 (u_long)(vm_offset_t)ea, 3057 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3058 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3059 if ((bp->b_pages[i]->valid & mask) == mask) 3060 continue; 3061 if ((bp->b_pages[i]->valid & mask) == 0) { 3062 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3063 bzero(sa, ea - sa); 3064 } 3065 } else { 3066 for (; sa < ea; sa += DEV_BSIZE, j++) { 3067 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3068 (bp->b_pages[i]->valid & (1<<j)) == 0) 3069 bzero(sa, DEV_BSIZE); 3070 } 3071 } 3072 bp->b_pages[i]->valid |= mask; 3073 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3074 } 3075 bp->b_resid = 0; 3076 } else { 3077 clrbuf(bp); 3078 } 3079 } 3080 3081 /* 3082 * vm_hold_load_pages and vm_hold_unload pages get pages into 3083 * a buffers address space. The pages are anonymous and are 3084 * not associated with a file object. 3085 */ 3086 void 3087 vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3088 { 3089 vm_offset_t pg; 3090 vm_page_t p; 3091 int index; 3092 3093 to = round_page(to); 3094 from = round_page(from); 3095 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3096 3097 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3098 3099 tryagain: 3100 3101 p = vm_page_alloc(kernel_object, 3102 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3103 VM_ALLOC_NORMAL); 3104 if (!p) { 3105 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3106 VM_WAIT; 3107 goto tryagain; 3108 } 3109 vm_page_wire(p); 3110 p->valid = VM_PAGE_BITS_ALL; 3111 vm_page_flag_clear(p, PG_ZERO); 3112 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 3113 bp->b_pages[index] = p; 3114 vm_page_wakeup(p); 3115 } 3116 bp->b_npages = index; 3117 } 3118 3119 void 3120 vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3121 { 3122 vm_offset_t pg; 3123 vm_page_t p; 3124 int index, newnpages; 3125 3126 from = round_page(from); 3127 to = round_page(to); 3128 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3129 3130 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3131 p = bp->b_pages[index]; 3132 if (p && (index < bp->b_npages)) { 3133 if (p->busy) { 3134 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 3135 bp->b_blkno, bp->b_lblkno); 3136 } 3137 bp->b_pages[index] = NULL; 3138 pmap_kremove(pg); 3139 vm_page_busy(p); 3140 vm_page_unwire(p, 0); 3141 vm_page_free(p); 3142 } 3143 } 3144 bp->b_npages = newnpages; 3145 } 3146 3147 3148 #include "opt_ddb.h" 3149 #ifdef DDB 3150 #include <ddb/ddb.h> 3151 3152 DB_SHOW_COMMAND(buffer, db_show_buffer) 3153 { 3154 /* get args */ 3155 struct buf *bp = (struct buf *)addr; 3156 3157 if (!have_addr) { 3158 db_printf("usage: show buffer <addr>\n"); 3159 return; 3160 } 3161 3162 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3163 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 3164 "b_resid = %ld\nb_dev = (%d,%d), b_data = %p, " 3165 "b_blkno = %d, b_pblkno = %d\n", 3166 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3167 major(bp->b_dev), minor(bp->b_dev), 3168 bp->b_data, bp->b_blkno, bp->b_pblkno); 3169 if (bp->b_npages) { 3170 int i; 3171 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3172 for (i = 0; i < bp->b_npages; i++) { 3173 vm_page_t m; 3174 m = bp->b_pages[i]; 3175 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3176 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3177 if ((i + 1) < bp->b_npages) 3178 db_printf(","); 3179 } 3180 db_printf("\n"); 3181 } 3182 } 3183 #endif /* DDB */ 3184