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: src/sys/kern/vfs_bio.c,v 1.242.2.20 2003/05/28 18:38:10 alc Exp $ 15 * $DragonFly: src/sys/kern/vfs_bio.c,v 1.58 2006/03/05 18:38:34 dillon Exp $ 16 */ 17 18 /* 19 * this file contains a new buffer I/O scheme implementing a coherent 20 * VM object and buffer cache scheme. Pains have been taken to make 21 * sure that the performance degradation associated with schemes such 22 * as this is not realized. 23 * 24 * Author: John S. Dyson 25 * Significant help during the development and debugging phases 26 * had been provided by David Greenman, also of the FreeBSD core team. 27 * 28 * see man buf(9) for more info. 29 */ 30 31 #include <sys/param.h> 32 #include <sys/systm.h> 33 #include <sys/buf.h> 34 #include <sys/conf.h> 35 #include <sys/eventhandler.h> 36 #include <sys/lock.h> 37 #include <sys/malloc.h> 38 #include <sys/mount.h> 39 #include <sys/kernel.h> 40 #include <sys/kthread.h> 41 #include <sys/proc.h> 42 #include <sys/reboot.h> 43 #include <sys/resourcevar.h> 44 #include <sys/sysctl.h> 45 #include <sys/vmmeter.h> 46 #include <sys/vnode.h> 47 #include <sys/proc.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 #include <sys/buf2.h> 58 #include <sys/thread2.h> 59 #include <vm/vm_page2.h> 60 61 /* 62 * Buffer queues. 63 */ 64 #define BUFFER_QUEUES 6 65 enum bufq_type { 66 BQUEUE_NONE, /* not on any queue */ 67 BQUEUE_LOCKED, /* locked buffers */ 68 BQUEUE_CLEAN, /* non-B_DELWRI buffers */ 69 BQUEUE_DIRTY, /* B_DELWRI buffers */ 70 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */ 71 BQUEUE_EMPTY /* empty buffer headers */ 72 }; 73 TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES]; 74 75 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 76 77 struct bio_ops bioops; /* I/O operation notification */ 78 79 struct buf *buf; /* buffer header pool */ 80 81 static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 82 vm_offset_t to); 83 static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 84 vm_offset_t to); 85 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 86 int pageno, vm_page_t m); 87 static void vfs_clean_pages(struct buf * bp); 88 static void vfs_setdirty(struct buf *bp); 89 static void vfs_vmio_release(struct buf *bp); 90 static int flushbufqueues(void); 91 92 static int bd_request; 93 94 static void buf_daemon (void); 95 /* 96 * bogus page -- for I/O to/from partially complete buffers 97 * this is a temporary solution to the problem, but it is not 98 * really that bad. it would be better to split the buffer 99 * for input in the case of buffers partially already in memory, 100 * but the code is intricate enough already. 101 */ 102 vm_page_t bogus_page; 103 int vmiodirenable = TRUE; 104 int runningbufspace; 105 106 static int bufspace, maxbufspace, 107 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace; 108 static int bufreusecnt, bufdefragcnt, buffreekvacnt; 109 static int needsbuffer; 110 static int lorunningspace, hirunningspace, runningbufreq; 111 static int numdirtybuffers, lodirtybuffers, hidirtybuffers; 112 static int numfreebuffers, lofreebuffers, hifreebuffers; 113 static int getnewbufcalls; 114 static int getnewbufrestarts; 115 116 /* 117 * Sysctls for operational control of the buffer cache. 118 */ 119 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 120 "Number of dirty buffers to flush before bufdaemon becomes inactive"); 121 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 122 "High watermark used to trigger explicit flushing of dirty buffers"); 123 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 124 "Low watermark for special reserve in low-memory situations"); 125 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 126 "High watermark for special reserve in low-memory situations"); 127 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 128 "Minimum amount of buffer space required for active I/O"); 129 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 130 "Maximum amount of buffer space to usable for active I/O"); 131 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 132 "Use the VM system for performing directory writes"); 133 /* 134 * Sysctls determining current state of the buffer cache. 135 */ 136 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 137 "Pending number of dirty buffers"); 138 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 139 "Number of free buffers on the buffer cache free list"); 140 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 141 "I/O bytes currently in progress due to asynchronous writes"); 142 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 143 "Hard limit on maximum amount of memory usable for buffer space"); 144 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 145 "Soft limit on maximum amount of memory usable for buffer space"); 146 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 147 "Minimum amount of memory to reserve for system buffer space"); 148 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 149 "Amount of memory available for buffers"); 150 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace, 151 0, "Maximum amount of memory reserved for buffers using malloc"); 152 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 153 "Amount of memory left for buffers using malloc-scheme"); 154 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0, 155 "New buffer header acquisition requests"); 156 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts, 157 0, "New buffer header acquisition restarts"); 158 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0, 159 "Buffer acquisition restarts due to fragmented buffer map"); 160 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0, 161 "Amount of time KVA space was deallocated in an arbitrary buffer"); 162 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0, 163 "Amount of time buffer re-use operations were successful"); 164 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf), 165 "sizeof(struct buf)"); 166 167 char *buf_wmesg = BUF_WMESG; 168 169 extern int vm_swap_size; 170 171 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 172 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 173 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 174 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 175 176 /* 177 * numdirtywakeup: 178 * 179 * If someone is blocked due to there being too many dirty buffers, 180 * and numdirtybuffers is now reasonable, wake them up. 181 */ 182 183 static __inline void 184 numdirtywakeup(int level) 185 { 186 if (numdirtybuffers <= level) { 187 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 188 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 189 wakeup(&needsbuffer); 190 } 191 } 192 } 193 194 /* 195 * bufspacewakeup: 196 * 197 * Called when buffer space is potentially available for recovery. 198 * getnewbuf() will block on this flag when it is unable to free 199 * sufficient buffer space. Buffer space becomes recoverable when 200 * bp's get placed back in the queues. 201 */ 202 203 static __inline void 204 bufspacewakeup(void) 205 { 206 /* 207 * If someone is waiting for BUF space, wake them up. Even 208 * though we haven't freed the kva space yet, the waiting 209 * process will be able to now. 210 */ 211 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 212 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 213 wakeup(&needsbuffer); 214 } 215 } 216 217 /* 218 * runningbufwakeup: 219 * 220 * Accounting for I/O in progress. 221 * 222 */ 223 static __inline void 224 runningbufwakeup(struct buf *bp) 225 { 226 if (bp->b_runningbufspace) { 227 runningbufspace -= bp->b_runningbufspace; 228 bp->b_runningbufspace = 0; 229 if (runningbufreq && runningbufspace <= lorunningspace) { 230 runningbufreq = 0; 231 wakeup(&runningbufreq); 232 } 233 } 234 } 235 236 /* 237 * bufcountwakeup: 238 * 239 * Called when a buffer has been added to one of the free queues to 240 * account for the buffer and to wakeup anyone waiting for free buffers. 241 * This typically occurs when large amounts of metadata are being handled 242 * by the buffer cache ( else buffer space runs out first, usually ). 243 */ 244 245 static __inline void 246 bufcountwakeup(void) 247 { 248 ++numfreebuffers; 249 if (needsbuffer) { 250 needsbuffer &= ~VFS_BIO_NEED_ANY; 251 if (numfreebuffers >= hifreebuffers) 252 needsbuffer &= ~VFS_BIO_NEED_FREE; 253 wakeup(&needsbuffer); 254 } 255 } 256 257 /* 258 * waitrunningbufspace() 259 * 260 * runningbufspace is a measure of the amount of I/O currently 261 * running. This routine is used in async-write situations to 262 * prevent creating huge backups of pending writes to a device. 263 * Only asynchronous writes are governed by this function. 264 * 265 * Reads will adjust runningbufspace, but will not block based on it. 266 * The read load has a side effect of reducing the allowed write load. 267 * 268 * This does NOT turn an async write into a sync write. It waits 269 * for earlier writes to complete and generally returns before the 270 * caller's write has reached the device. 271 */ 272 static __inline void 273 waitrunningbufspace(void) 274 { 275 if (runningbufspace > hirunningspace) { 276 crit_enter(); 277 while (runningbufspace > hirunningspace) { 278 ++runningbufreq; 279 tsleep(&runningbufreq, 0, "wdrain", 0); 280 } 281 crit_exit(); 282 } 283 } 284 285 /* 286 * vfs_buf_test_cache: 287 * 288 * Called when a buffer is extended. This function clears the B_CACHE 289 * bit if the newly extended portion of the buffer does not contain 290 * valid data. 291 */ 292 static __inline__ 293 void 294 vfs_buf_test_cache(struct buf *bp, 295 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 296 vm_page_t m) 297 { 298 if (bp->b_flags & B_CACHE) { 299 int base = (foff + off) & PAGE_MASK; 300 if (vm_page_is_valid(m, base, size) == 0) 301 bp->b_flags &= ~B_CACHE; 302 } 303 } 304 305 /* 306 * bd_wakeup: 307 * 308 * Wake up the buffer daemon if the number of outstanding dirty buffers 309 * is above specified threshold 'dirtybuflevel'. 310 * 311 * The buffer daemon is explicitly woken up when (a) the pending number 312 * of dirty buffers exceeds the recovery and stall mid-point value, 313 * (b) during bwillwrite() or (c) buf freelist was exhausted. 314 */ 315 static __inline__ 316 void 317 bd_wakeup(int dirtybuflevel) 318 { 319 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 320 bd_request = 1; 321 wakeup(&bd_request); 322 } 323 } 324 325 /* 326 * bd_speedup: 327 * 328 * Speed up the buffer cache flushing process. 329 */ 330 331 static __inline__ 332 void 333 bd_speedup(void) 334 { 335 bd_wakeup(1); 336 } 337 338 /* 339 * bufinit: 340 * 341 * Load time initialisation of the buffer cache, called from machine 342 * dependant initialization code. 343 */ 344 void 345 bufinit(void) 346 { 347 struct buf *bp; 348 vm_offset_t bogus_offset; 349 int i; 350 351 /* next, make a null set of free lists */ 352 for (i = 0; i < BUFFER_QUEUES; i++) 353 TAILQ_INIT(&bufqueues[i]); 354 355 /* finally, initialize each buffer header and stick on empty q */ 356 for (i = 0; i < nbuf; i++) { 357 bp = &buf[i]; 358 bzero(bp, sizeof *bp); 359 bp->b_flags = B_INVAL; /* we're just an empty header */ 360 bp->b_qindex = BQUEUE_EMPTY; 361 initbufbio(bp); 362 xio_init(&bp->b_xio); 363 LIST_INIT(&bp->b_dep); 364 BUF_LOCKINIT(bp); 365 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist); 366 } 367 368 /* 369 * maxbufspace is the absolute maximum amount of buffer space we are 370 * allowed to reserve in KVM and in real terms. The absolute maximum 371 * is nominally used by buf_daemon. hibufspace is the nominal maximum 372 * used by most other processes. The differential is required to 373 * ensure that buf_daemon is able to run when other processes might 374 * be blocked waiting for buffer space. 375 * 376 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 377 * this may result in KVM fragmentation which is not handled optimally 378 * by the system. 379 */ 380 maxbufspace = nbuf * BKVASIZE; 381 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 382 lobufspace = hibufspace - MAXBSIZE; 383 384 lorunningspace = 512 * 1024; 385 hirunningspace = 1024 * 1024; 386 387 /* 388 * Limit the amount of malloc memory since it is wired permanently into 389 * the kernel space. Even though this is accounted for in the buffer 390 * allocation, we don't want the malloced region to grow uncontrolled. 391 * The malloc scheme improves memory utilization significantly on average 392 * (small) directories. 393 */ 394 maxbufmallocspace = hibufspace / 20; 395 396 /* 397 * Reduce the chance of a deadlock occuring by limiting the number 398 * of delayed-write dirty buffers we allow to stack up. 399 */ 400 hidirtybuffers = nbuf / 4 + 20; 401 numdirtybuffers = 0; 402 /* 403 * To support extreme low-memory systems, make sure hidirtybuffers cannot 404 * eat up all available buffer space. This occurs when our minimum cannot 405 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 406 * BKVASIZE'd (8K) buffers. 407 */ 408 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 409 hidirtybuffers >>= 1; 410 } 411 lodirtybuffers = hidirtybuffers / 2; 412 413 /* 414 * Try to keep the number of free buffers in the specified range, 415 * and give special processes (e.g. like buf_daemon) access to an 416 * emergency reserve. 417 */ 418 lofreebuffers = nbuf / 18 + 5; 419 hifreebuffers = 2 * lofreebuffers; 420 numfreebuffers = nbuf; 421 422 /* 423 * Maximum number of async ops initiated per buf_daemon loop. This is 424 * somewhat of a hack at the moment, we really need to limit ourselves 425 * based on the number of bytes of I/O in-transit that were initiated 426 * from buf_daemon. 427 */ 428 429 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 430 bogus_page = vm_page_alloc(kernel_object, 431 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 432 VM_ALLOC_NORMAL); 433 vmstats.v_wire_count++; 434 435 } 436 437 /* 438 * Initialize the embedded bio structures 439 */ 440 void 441 initbufbio(struct buf *bp) 442 { 443 bp->b_bio1.bio_buf = bp; 444 bp->b_bio1.bio_prev = NULL; 445 bp->b_bio1.bio_blkno = (daddr_t)-1; 446 bp->b_bio1.bio_offset = NOOFFSET; 447 bp->b_bio1.bio_next = &bp->b_bio2; 448 bp->b_bio1.bio_done = NULL; 449 450 bp->b_bio2.bio_buf = bp; 451 bp->b_bio2.bio_prev = &bp->b_bio1; 452 bp->b_bio2.bio_blkno = (daddr_t)-1; 453 bp->b_bio2.bio_offset = NOOFFSET; 454 bp->b_bio2.bio_next = NULL; 455 bp->b_bio2.bio_done = NULL; 456 } 457 458 /* 459 * Reinitialize the embedded bio structures as well as any additional 460 * translation cache layers. 461 */ 462 void 463 reinitbufbio(struct buf *bp) 464 { 465 struct bio *bio; 466 467 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) { 468 bio->bio_done = NULL; 469 bio->bio_blkno = (daddr_t)-1; 470 bio->bio_offset = NOOFFSET; 471 } 472 } 473 474 /* 475 * Push another BIO layer onto an existing BIO and return it. The new 476 * BIO layer may already exist, holding cached translation data. 477 */ 478 struct bio * 479 push_bio(struct bio *bio) 480 { 481 struct bio *nbio; 482 483 if ((nbio = bio->bio_next) == NULL) { 484 int index = bio - &bio->bio_buf->b_bio_array[0]; 485 if (index >= NBUF_BIO) { 486 panic("push_bio: too many layers bp %p\n", 487 bio->bio_buf); 488 } 489 nbio = &bio->bio_buf->b_bio_array[index + 1]; 490 bio->bio_next = nbio; 491 nbio->bio_prev = bio; 492 nbio->bio_buf = bio->bio_buf; 493 nbio->bio_blkno = (daddr_t)-1; 494 nbio->bio_offset = NOOFFSET; 495 nbio->bio_done = NULL; 496 nbio->bio_next = NULL; 497 } 498 KKASSERT(nbio->bio_done == NULL); 499 return(nbio); 500 } 501 502 void 503 pop_bio(struct bio *bio) 504 { 505 /* NOP */ 506 } 507 508 void 509 clearbiocache(struct bio *bio) 510 { 511 while (bio) { 512 bio->bio_blkno = (daddr_t)-1; 513 bio->bio_offset = NOOFFSET; 514 bio = bio->bio_next; 515 } 516 } 517 518 /* 519 * bfreekva: 520 * 521 * Free the KVA allocation for buffer 'bp'. 522 * 523 * Must be called from a critical section as this is the only locking for 524 * buffer_map. 525 * 526 * Since this call frees up buffer space, we call bufspacewakeup(). 527 */ 528 static void 529 bfreekva(struct buf * bp) 530 { 531 int count; 532 533 if (bp->b_kvasize) { 534 ++buffreekvacnt; 535 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 536 vm_map_lock(buffer_map); 537 bufspace -= bp->b_kvasize; 538 vm_map_delete(buffer_map, 539 (vm_offset_t) bp->b_kvabase, 540 (vm_offset_t) bp->b_kvabase + bp->b_kvasize, 541 &count 542 ); 543 vm_map_unlock(buffer_map); 544 vm_map_entry_release(count); 545 bp->b_kvasize = 0; 546 bufspacewakeup(); 547 } 548 } 549 550 /* 551 * bremfree: 552 * 553 * Remove the buffer from the appropriate free list. 554 */ 555 void 556 bremfree(struct buf * bp) 557 { 558 int old_qindex; 559 560 crit_enter(); 561 old_qindex = bp->b_qindex; 562 563 if (bp->b_qindex != BQUEUE_NONE) { 564 KASSERT(BUF_REFCNTNB(bp) == 1, 565 ("bremfree: bp %p not locked",bp)); 566 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 567 bp->b_qindex = BQUEUE_NONE; 568 } else { 569 if (BUF_REFCNTNB(bp) <= 1) 570 panic("bremfree: removing a buffer not on a queue"); 571 } 572 573 /* 574 * Fixup numfreebuffers count. If the buffer is invalid or not 575 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 576 * the buffer was free and we must decrement numfreebuffers. 577 */ 578 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 579 switch(old_qindex) { 580 case BQUEUE_DIRTY: 581 case BQUEUE_CLEAN: 582 case BQUEUE_EMPTY: 583 case BQUEUE_EMPTYKVA: 584 --numfreebuffers; 585 break; 586 default: 587 break; 588 } 589 } 590 crit_exit(); 591 } 592 593 594 /* 595 * bread: 596 * 597 * Get a buffer with the specified data. Look in the cache first. We 598 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 599 * is set, the buffer is valid and we do not have to do anything ( see 600 * getblk() ). 601 */ 602 int 603 bread(struct vnode * vp, daddr_t blkno, int size, struct buf ** bpp) 604 { 605 struct buf *bp; 606 607 bp = getblk(vp, blkno, size, 0, 0); 608 *bpp = bp; 609 610 /* if not found in cache, do some I/O */ 611 if ((bp->b_flags & B_CACHE) == 0) { 612 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp)); 613 bp->b_flags |= B_READ; 614 bp->b_flags &= ~(B_ERROR | B_INVAL); 615 vfs_busy_pages(bp, 0); 616 vn_strategy(vp, &bp->b_bio1); 617 return (biowait(bp)); 618 } 619 return (0); 620 } 621 622 /* 623 * breadn: 624 * 625 * Operates like bread, but also starts asynchronous I/O on 626 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior 627 * to initiating I/O . If B_CACHE is set, the buffer is valid 628 * and we do not have to do anything. 629 */ 630 int 631 breadn(struct vnode * vp, daddr_t blkno, int size, daddr_t * rablkno, 632 int *rabsize, int cnt, struct buf ** bpp) 633 { 634 struct buf *bp, *rabp; 635 int i; 636 int rv = 0, readwait = 0; 637 638 *bpp = bp = getblk(vp, blkno, size, 0, 0); 639 640 /* if not found in cache, do some I/O */ 641 if ((bp->b_flags & B_CACHE) == 0) { 642 bp->b_flags |= B_READ; 643 bp->b_flags &= ~(B_ERROR | B_INVAL); 644 vfs_busy_pages(bp, 0); 645 vn_strategy(vp, &bp->b_bio1); 646 ++readwait; 647 } 648 649 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 650 if (inmem(vp, *rablkno)) 651 continue; 652 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 653 654 if ((rabp->b_flags & B_CACHE) == 0) { 655 rabp->b_flags |= B_READ | B_ASYNC; 656 rabp->b_flags &= ~(B_ERROR | B_INVAL); 657 vfs_busy_pages(rabp, 0); 658 BUF_KERNPROC(rabp); 659 vn_strategy(vp, &rabp->b_bio1); 660 } else { 661 brelse(rabp); 662 } 663 } 664 665 if (readwait) { 666 rv = biowait(bp); 667 } 668 return (rv); 669 } 670 671 /* 672 * bwrite: 673 * 674 * Write, release buffer on completion. (Done by iodone 675 * if async). Do not bother writing anything if the buffer 676 * is invalid. 677 * 678 * Note that we set B_CACHE here, indicating that buffer is 679 * fully valid and thus cacheable. This is true even of NFS 680 * now so we set it generally. This could be set either here 681 * or in biodone() since the I/O is synchronous. We put it 682 * here. 683 */ 684 int 685 bwrite(struct buf * bp) 686 { 687 int oldflags; 688 689 if (bp->b_flags & B_INVAL) { 690 brelse(bp); 691 return (0); 692 } 693 694 oldflags = bp->b_flags; 695 696 if (BUF_REFCNTNB(bp) == 0) 697 panic("bwrite: buffer is not busy???"); 698 crit_enter(); 699 /* 700 * If a background write is already in progress, delay 701 * writing this block if it is asynchronous. Otherwise 702 * wait for the background write to complete. 703 */ 704 if (bp->b_xflags & BX_BKGRDINPROG) { 705 if (bp->b_flags & B_ASYNC) { 706 crit_exit(); 707 bdwrite(bp); 708 return (0); 709 } 710 bp->b_xflags |= BX_BKGRDWAIT; 711 tsleep(&bp->b_xflags, 0, "biord", 0); 712 if (bp->b_xflags & BX_BKGRDINPROG) 713 panic("bwrite: still writing"); 714 } 715 716 /* Mark the buffer clean */ 717 bundirty(bp); 718 719 bp->b_flags &= ~(B_READ | B_DONE | B_ERROR); 720 bp->b_flags |= B_CACHE; 721 722 vfs_busy_pages(bp, 1); 723 724 /* 725 * Normal bwrites pipeline writes 726 */ 727 bp->b_runningbufspace = bp->b_bufsize; 728 runningbufspace += bp->b_runningbufspace; 729 730 crit_exit(); 731 if (oldflags & B_ASYNC) 732 BUF_KERNPROC(bp); 733 vn_strategy(bp->b_vp, &bp->b_bio1); 734 735 if ((oldflags & B_ASYNC) == 0) { 736 int rtval = biowait(bp); 737 brelse(bp); 738 return (rtval); 739 } else if ((oldflags & B_NOWDRAIN) == 0) { 740 /* 741 * don't allow the async write to saturate the I/O 742 * system. Deadlocks can occur only if a device strategy 743 * routine (like in VN) turns around and issues another 744 * high-level write, in which case B_NOWDRAIN is expected 745 * to be set. Otherwise we will not deadlock here because 746 * we are blocking waiting for I/O that is already in-progress 747 * to complete. 748 */ 749 waitrunningbufspace(); 750 } 751 752 return (0); 753 } 754 755 /* 756 * bdwrite: 757 * 758 * Delayed write. (Buffer is marked dirty). Do not bother writing 759 * anything if the buffer is marked invalid. 760 * 761 * Note that since the buffer must be completely valid, we can safely 762 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 763 * biodone() in order to prevent getblk from writing the buffer 764 * out synchronously. 765 */ 766 void 767 bdwrite(struct buf *bp) 768 { 769 if (BUF_REFCNTNB(bp) == 0) 770 panic("bdwrite: buffer is not busy"); 771 772 if (bp->b_flags & B_INVAL) { 773 brelse(bp); 774 return; 775 } 776 bdirty(bp); 777 778 /* 779 * Set B_CACHE, indicating that the buffer is fully valid. This is 780 * true even of NFS now. 781 */ 782 bp->b_flags |= B_CACHE; 783 784 /* 785 * This bmap keeps the system from needing to do the bmap later, 786 * perhaps when the system is attempting to do a sync. Since it 787 * is likely that the indirect block -- or whatever other datastructure 788 * that the filesystem needs is still in memory now, it is a good 789 * thing to do this. Note also, that if the pageout daemon is 790 * requesting a sync -- there might not be enough memory to do 791 * the bmap then... So, this is important to do. 792 */ 793 if (bp->b_bio2.bio_blkno == (daddr_t)-1) { 794 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_bio2.bio_blkno, 795 NULL, NULL); 796 } 797 798 /* 799 * Set the *dirty* buffer range based upon the VM system dirty pages. 800 */ 801 vfs_setdirty(bp); 802 803 /* 804 * We need to do this here to satisfy the vnode_pager and the 805 * pageout daemon, so that it thinks that the pages have been 806 * "cleaned". Note that since the pages are in a delayed write 807 * buffer -- the VFS layer "will" see that the pages get written 808 * out on the next sync, or perhaps the cluster will be completed. 809 */ 810 vfs_clean_pages(bp); 811 bqrelse(bp); 812 813 /* 814 * Wakeup the buffer flushing daemon if we have a lot of dirty 815 * buffers (midpoint between our recovery point and our stall 816 * point). 817 */ 818 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 819 820 /* 821 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 822 * due to the softdep code. 823 */ 824 } 825 826 /* 827 * bdirty: 828 * 829 * Turn buffer into delayed write request. We must clear B_READ and 830 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 831 * itself to properly update it in the dirty/clean lists. We mark it 832 * B_DONE to ensure that any asynchronization of the buffer properly 833 * clears B_DONE ( else a panic will occur later ). 834 * 835 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 836 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 837 * should only be called if the buffer is known-good. 838 * 839 * Since the buffer is not on a queue, we do not update the numfreebuffers 840 * count. 841 * 842 * Must be called from a critical section. 843 * The buffer must be on BQUEUE_NONE. 844 */ 845 void 846 bdirty(struct buf *bp) 847 { 848 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 849 bp->b_flags &= ~(B_READ|B_RELBUF); 850 851 if ((bp->b_flags & B_DELWRI) == 0) { 852 bp->b_flags |= B_DONE | B_DELWRI; 853 reassignbuf(bp); 854 ++numdirtybuffers; 855 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 856 } 857 } 858 859 /* 860 * bundirty: 861 * 862 * Clear B_DELWRI for buffer. 863 * 864 * Since the buffer is not on a queue, we do not update the numfreebuffers 865 * count. 866 * 867 * Must be called from a critical section. 868 * 869 * The buffer is typically on BQUEUE_NONE but there is one case in 870 * brelse() that calls this function after placing the buffer on 871 * a different queue. 872 */ 873 874 void 875 bundirty(struct buf *bp) 876 { 877 if (bp->b_flags & B_DELWRI) { 878 bp->b_flags &= ~B_DELWRI; 879 reassignbuf(bp); 880 --numdirtybuffers; 881 numdirtywakeup(lodirtybuffers); 882 } 883 /* 884 * Since it is now being written, we can clear its deferred write flag. 885 */ 886 bp->b_flags &= ~B_DEFERRED; 887 } 888 889 /* 890 * bawrite: 891 * 892 * Asynchronous write. Start output on a buffer, but do not wait for 893 * it to complete. The buffer is released when the output completes. 894 * 895 * bwrite() ( or the VOP routine anyway ) is responsible for handling 896 * B_INVAL buffers. Not us. 897 */ 898 void 899 bawrite(struct buf * bp) 900 { 901 bp->b_flags |= B_ASYNC; 902 (void) VOP_BWRITE(bp->b_vp, bp); 903 } 904 905 /* 906 * bowrite: 907 * 908 * Ordered write. Start output on a buffer, and flag it so that the 909 * device will write it in the order it was queued. The buffer is 910 * released when the output completes. bwrite() ( or the VOP routine 911 * anyway ) is responsible for handling B_INVAL buffers. 912 */ 913 int 914 bowrite(struct buf * bp) 915 { 916 bp->b_flags |= B_ORDERED | B_ASYNC; 917 return (VOP_BWRITE(bp->b_vp, bp)); 918 } 919 920 /* 921 * bwillwrite: 922 * 923 * Called prior to the locking of any vnodes when we are expecting to 924 * write. We do not want to starve the buffer cache with too many 925 * dirty buffers so we block here. By blocking prior to the locking 926 * of any vnodes we attempt to avoid the situation where a locked vnode 927 * prevents the various system daemons from flushing related buffers. 928 */ 929 930 void 931 bwillwrite(void) 932 { 933 if (numdirtybuffers >= hidirtybuffers) { 934 crit_enter(); 935 while (numdirtybuffers >= hidirtybuffers) { 936 bd_wakeup(1); 937 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 938 tsleep(&needsbuffer, 0, "flswai", 0); 939 } 940 crit_exit(); 941 } 942 } 943 944 /* 945 * buf_dirty_count_severe: 946 * 947 * Return true if we have too many dirty buffers. 948 */ 949 int 950 buf_dirty_count_severe(void) 951 { 952 return(numdirtybuffers >= hidirtybuffers); 953 } 954 955 /* 956 * brelse: 957 * 958 * Release a busy buffer and, if requested, free its resources. The 959 * buffer will be stashed in the appropriate bufqueue[] allowing it 960 * to be accessed later as a cache entity or reused for other purposes. 961 */ 962 void 963 brelse(struct buf * bp) 964 { 965 #ifdef INVARIANTS 966 int saved_flags = bp->b_flags; 967 #endif 968 969 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 970 971 crit_enter(); 972 973 if (bp->b_flags & B_LOCKED) 974 bp->b_flags &= ~B_ERROR; 975 976 if ((bp->b_flags & (B_READ | B_ERROR | B_INVAL)) == B_ERROR) { 977 /* 978 * Failed write, redirty. Must clear B_ERROR to prevent 979 * pages from being scrapped. If B_INVAL is set then 980 * this case is not run and the next case is run to 981 * destroy the buffer. B_INVAL can occur if the buffer 982 * is outside the range supported by the underlying device. 983 */ 984 bp->b_flags &= ~B_ERROR; 985 bdirty(bp); 986 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_FREEBUF)) || 987 (bp->b_bufsize <= 0)) { 988 /* 989 * Either a failed I/O or we were asked to free or not 990 * cache the buffer. 991 */ 992 bp->b_flags |= B_INVAL; 993 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 994 (*bioops.io_deallocate)(bp); 995 if (bp->b_flags & B_DELWRI) { 996 --numdirtybuffers; 997 numdirtywakeup(lodirtybuffers); 998 } 999 bp->b_flags &= ~(B_DELWRI | B_CACHE | B_FREEBUF); 1000 } 1001 1002 /* 1003 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1004 * is called with B_DELWRI set, the underlying pages may wind up 1005 * getting freed causing a previous write (bdwrite()) to get 'lost' 1006 * because pages associated with a B_DELWRI bp are marked clean. 1007 * 1008 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1009 * if B_DELWRI is set. 1010 * 1011 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1012 * on pages to return pages to the VM page queues. 1013 */ 1014 if (bp->b_flags & B_DELWRI) 1015 bp->b_flags &= ~B_RELBUF; 1016 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1017 bp->b_flags |= B_RELBUF; 1018 1019 /* 1020 * At this point destroying the buffer is governed by the B_INVAL 1021 * or B_RELBUF flags. 1022 */ 1023 1024 /* 1025 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1026 * constituted, not even NFS buffers now. Two flags effect this. If 1027 * B_INVAL, the struct buf is invalidated but the VM object is kept 1028 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1029 * 1030 * If B_ERROR or B_NOCACHE is set, pages in the VM object will be 1031 * invalidated. B_ERROR cannot be set for a failed write unless the 1032 * buffer is also B_INVAL because it hits the re-dirtying code above. 1033 * 1034 * Normally we can do this whether a buffer is B_DELWRI or not. If 1035 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1036 * the commit state and we cannot afford to lose the buffer. If the 1037 * buffer has a background write in progress, we need to keep it 1038 * around to prevent it from being reconstituted and starting a second 1039 * background write. 1040 */ 1041 if ((bp->b_flags & B_VMIO) 1042 && !(bp->b_vp->v_tag == VT_NFS && 1043 !vn_isdisk(bp->b_vp, NULL) && 1044 (bp->b_flags & B_DELWRI)) 1045 ) { 1046 /* 1047 * Rundown for VMIO buffers which are not dirty NFS buffers. 1048 */ 1049 int i, j, resid; 1050 vm_page_t m; 1051 off_t foff; 1052 vm_pindex_t poff; 1053 vm_object_t obj; 1054 struct vnode *vp; 1055 1056 vp = bp->b_vp; 1057 1058 /* 1059 * Get the base offset and length of the buffer. Note that 1060 * in the VMIO case if the buffer block size is not 1061 * page-aligned then b_data pointer may not be page-aligned. 1062 * But our b_xio.xio_pages array *IS* page aligned. 1063 * 1064 * block sizes less then DEV_BSIZE (usually 512) are not 1065 * supported due to the page granularity bits (m->valid, 1066 * m->dirty, etc...). 1067 * 1068 * See man buf(9) for more information 1069 */ 1070 1071 resid = bp->b_bufsize; 1072 foff = bp->b_loffset; 1073 1074 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1075 m = bp->b_xio.xio_pages[i]; 1076 vm_page_flag_clear(m, PG_ZERO); 1077 /* 1078 * If we hit a bogus page, fixup *all* of them 1079 * now. Note that we left these pages wired 1080 * when we removed them so they had better exist, 1081 * and they cannot be ripped out from under us so 1082 * no critical section protection is necessary. 1083 */ 1084 if (m == bogus_page) { 1085 VOP_GETVOBJECT(vp, &obj); 1086 poff = OFF_TO_IDX(bp->b_loffset); 1087 1088 for (j = i; j < bp->b_xio.xio_npages; j++) { 1089 vm_page_t mtmp; 1090 1091 mtmp = bp->b_xio.xio_pages[j]; 1092 if (mtmp == bogus_page) { 1093 mtmp = vm_page_lookup(obj, poff + j); 1094 if (!mtmp) { 1095 panic("brelse: page missing"); 1096 } 1097 bp->b_xio.xio_pages[j] = mtmp; 1098 } 1099 } 1100 1101 if ((bp->b_flags & B_INVAL) == 0) { 1102 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 1103 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 1104 } 1105 m = bp->b_xio.xio_pages[i]; 1106 } 1107 1108 /* 1109 * Invalidate the backing store if B_NOCACHE is set 1110 * (e.g. used with vinvalbuf()). If this is NFS 1111 * we impose a requirement that the block size be 1112 * a multiple of PAGE_SIZE and create a temporary 1113 * hack to basically invalidate the whole page. The 1114 * problem is that NFS uses really odd buffer sizes 1115 * especially when tracking piecemeal writes and 1116 * it also vinvalbuf()'s a lot, which would result 1117 * in only partial page validation and invalidation 1118 * here. If the file page is mmap()'d, however, 1119 * all the valid bits get set so after we invalidate 1120 * here we would end up with weird m->valid values 1121 * like 0xfc. nfs_getpages() can't handle this so 1122 * we clear all the valid bits for the NFS case 1123 * instead of just some of them. 1124 * 1125 * The real bug is the VM system having to set m->valid 1126 * to VM_PAGE_BITS_ALL for faulted-in pages, which 1127 * itself is an artifact of the whole 512-byte 1128 * granular mess that exists to support odd block 1129 * sizes and UFS meta-data block sizes (e.g. 6144). 1130 * A complete rewrite is required. 1131 */ 1132 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 1133 int poffset = foff & PAGE_MASK; 1134 int presid; 1135 1136 presid = PAGE_SIZE - poffset; 1137 if (bp->b_vp->v_tag == VT_NFS && 1138 bp->b_vp->v_type == VREG) { 1139 ; /* entire page */ 1140 } else if (presid > resid) { 1141 presid = resid; 1142 } 1143 KASSERT(presid >= 0, ("brelse: extra page")); 1144 vm_page_set_invalid(m, poffset, presid); 1145 } 1146 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1147 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1148 } 1149 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1150 vfs_vmio_release(bp); 1151 } else if (bp->b_flags & B_VMIO) { 1152 /* 1153 * Rundown for VMIO buffers which are dirty NFS buffers. Such 1154 * buffers contain tracking ranges for NFS and cannot normally 1155 * be released. Due to the dirty check above this series of 1156 * conditionals, B_RELBUF probably will never be set in this 1157 * codepath. 1158 */ 1159 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1160 vfs_vmio_release(bp); 1161 } else { 1162 /* 1163 * Rundown for non-VMIO buffers. 1164 */ 1165 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1166 #if 0 1167 if (bp->b_vp) 1168 printf("brelse bp %p %08x/%08lx: Warning, caught and fixed brelvp bug\n", bp, saved_flags, bp->b_flags); 1169 #endif 1170 if (bp->b_bufsize) 1171 allocbuf(bp, 0); 1172 if (bp->b_vp) 1173 brelvp(bp); 1174 } 1175 } 1176 1177 if (bp->b_qindex != BQUEUE_NONE) 1178 panic("brelse: free buffer onto another queue???"); 1179 if (BUF_REFCNTNB(bp) > 1) { 1180 /* Temporary panic to verify exclusive locking */ 1181 /* This panic goes away when we allow shared refs */ 1182 panic("brelse: multiple refs"); 1183 /* do not release to free list */ 1184 BUF_UNLOCK(bp); 1185 crit_exit(); 1186 return; 1187 } 1188 1189 /* 1190 * Figure out the correct queue to place the cleaned up buffer on. 1191 * Buffers placed in the EMPTY or EMPTYKVA had better already be 1192 * disassociated from their vnode. 1193 */ 1194 1195 if (bp->b_bufsize == 0) { 1196 /* 1197 * Buffers with no memory. Due to conditionals near the top 1198 * of brelse() such buffers should probably already be 1199 * marked B_INVAL and disassociated from their vnode. 1200 */ 1201 bp->b_flags |= B_INVAL; 1202 bp->b_xflags &= ~BX_BKGRDWRITE; 1203 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08lx vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp)); 1204 KKASSERT((bp->b_flags & B_HASHED) == 0); 1205 if (bp->b_xflags & BX_BKGRDINPROG) 1206 panic("losing buffer 1"); 1207 if (bp->b_kvasize) { 1208 bp->b_qindex = BQUEUE_EMPTYKVA; 1209 } else { 1210 bp->b_qindex = BQUEUE_EMPTY; 1211 } 1212 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1213 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) { 1214 /* 1215 * Buffers with junk contents. Again these buffers had better 1216 * already be disassociated from their vnode. 1217 */ 1218 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08lx vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp)); 1219 KKASSERT((bp->b_flags & B_HASHED) == 0); 1220 bp->b_flags |= B_INVAL; 1221 bp->b_xflags &= ~BX_BKGRDWRITE; 1222 if (bp->b_xflags & BX_BKGRDINPROG) 1223 panic("losing buffer 2"); 1224 bp->b_qindex = BQUEUE_CLEAN; 1225 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1226 } else if (bp->b_flags & B_LOCKED) { 1227 /* 1228 * Buffers that are locked. 1229 */ 1230 bp->b_qindex = BQUEUE_LOCKED; 1231 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist); 1232 } else { 1233 /* 1234 * Remaining buffers. These buffers are still associated with 1235 * their vnode. 1236 */ 1237 switch(bp->b_flags & (B_DELWRI|B_AGE)) { 1238 case B_DELWRI | B_AGE: 1239 bp->b_qindex = BQUEUE_DIRTY; 1240 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_DIRTY], bp, b_freelist); 1241 break; 1242 case B_DELWRI: 1243 bp->b_qindex = BQUEUE_DIRTY; 1244 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist); 1245 break; 1246 case B_AGE: 1247 bp->b_qindex = BQUEUE_CLEAN; 1248 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1249 break; 1250 default: 1251 bp->b_qindex = BQUEUE_CLEAN; 1252 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1253 break; 1254 } 1255 } 1256 1257 /* 1258 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1259 * on the correct queue. 1260 */ 1261 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) 1262 bundirty(bp); 1263 1264 /* 1265 * Fixup numfreebuffers count. The bp is on an appropriate queue 1266 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1267 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1268 * if B_INVAL is set ). 1269 */ 1270 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1271 bufcountwakeup(); 1272 1273 /* 1274 * Something we can maybe free or reuse 1275 */ 1276 if (bp->b_bufsize || bp->b_kvasize) 1277 bufspacewakeup(); 1278 1279 /* unlock */ 1280 BUF_UNLOCK(bp); 1281 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | 1282 B_DIRECT | B_NOWDRAIN); 1283 crit_exit(); 1284 } 1285 1286 /* 1287 * bqrelse: 1288 * 1289 * Release a buffer back to the appropriate queue but do not try to free 1290 * it. The buffer is expected to be used again soon. 1291 * 1292 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1293 * biodone() to requeue an async I/O on completion. It is also used when 1294 * known good buffers need to be requeued but we think we may need the data 1295 * again soon. 1296 * 1297 * XXX we should be able to leave the B_RELBUF hint set on completion. 1298 */ 1299 void 1300 bqrelse(struct buf * bp) 1301 { 1302 crit_enter(); 1303 1304 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1305 1306 if (bp->b_qindex != BQUEUE_NONE) 1307 panic("bqrelse: free buffer onto another queue???"); 1308 if (BUF_REFCNTNB(bp) > 1) { 1309 /* do not release to free list */ 1310 panic("bqrelse: multiple refs"); 1311 BUF_UNLOCK(bp); 1312 crit_exit(); 1313 return; 1314 } 1315 if (bp->b_flags & B_LOCKED) { 1316 bp->b_flags &= ~B_ERROR; 1317 bp->b_qindex = BQUEUE_LOCKED; 1318 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist); 1319 /* buffers with stale but valid contents */ 1320 } else if (bp->b_flags & B_DELWRI) { 1321 bp->b_qindex = BQUEUE_DIRTY; 1322 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist); 1323 } else if (vm_page_count_severe()) { 1324 /* 1325 * We are too low on memory, we have to try to free the 1326 * buffer (most importantly: the wired pages making up its 1327 * backing store) *now*. 1328 */ 1329 crit_exit(); 1330 brelse(bp); 1331 return; 1332 } else { 1333 bp->b_qindex = BQUEUE_CLEAN; 1334 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1335 } 1336 1337 if ((bp->b_flags & B_LOCKED) == 0 && 1338 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1339 bufcountwakeup(); 1340 } 1341 1342 /* 1343 * Something we can maybe free or reuse. 1344 */ 1345 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1346 bufspacewakeup(); 1347 1348 /* 1349 * Final cleanup and unlock. Clear bits that are only used while a 1350 * buffer is actively locked. 1351 */ 1352 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1353 BUF_UNLOCK(bp); 1354 crit_exit(); 1355 } 1356 1357 /* 1358 * vfs_vmio_release: 1359 * 1360 * Return backing pages held by the buffer 'bp' back to the VM system 1361 * if possible. The pages are freed if they are no longer valid or 1362 * attempt to free if it was used for direct I/O otherwise they are 1363 * sent to the page cache. 1364 * 1365 * Pages that were marked busy are left alone and skipped. 1366 * 1367 * The KVA mapping (b_data) for the underlying pages is removed by 1368 * this function. 1369 */ 1370 static void 1371 vfs_vmio_release(struct buf *bp) 1372 { 1373 int i; 1374 vm_page_t m; 1375 1376 crit_enter(); 1377 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1378 m = bp->b_xio.xio_pages[i]; 1379 bp->b_xio.xio_pages[i] = NULL; 1380 /* 1381 * In order to keep page LRU ordering consistent, put 1382 * everything on the inactive queue. 1383 */ 1384 vm_page_unwire(m, 0); 1385 /* 1386 * We don't mess with busy pages, it is 1387 * the responsibility of the process that 1388 * busied the pages to deal with them. 1389 */ 1390 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1391 continue; 1392 1393 if (m->wire_count == 0) { 1394 vm_page_flag_clear(m, PG_ZERO); 1395 /* 1396 * Might as well free the page if we can and it has 1397 * no valid data. We also free the page if the 1398 * buffer was used for direct I/O. 1399 */ 1400 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1401 m->hold_count == 0) { 1402 vm_page_busy(m); 1403 vm_page_protect(m, VM_PROT_NONE); 1404 vm_page_free(m); 1405 } else if (bp->b_flags & B_DIRECT) { 1406 vm_page_try_to_free(m); 1407 } else if (vm_page_count_severe()) { 1408 vm_page_try_to_cache(m); 1409 } 1410 } 1411 } 1412 crit_exit(); 1413 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages); 1414 if (bp->b_bufsize) { 1415 bufspacewakeup(); 1416 bp->b_bufsize = 0; 1417 } 1418 bp->b_xio.xio_npages = 0; 1419 bp->b_flags &= ~B_VMIO; 1420 if (bp->b_vp) 1421 brelvp(bp); 1422 } 1423 1424 /* 1425 * vfs_bio_awrite: 1426 * 1427 * Implement clustered async writes for clearing out B_DELWRI buffers. 1428 * This is much better then the old way of writing only one buffer at 1429 * a time. Note that we may not be presented with the buffers in the 1430 * correct order, so we search for the cluster in both directions. 1431 * 1432 * The buffer is locked on call. 1433 */ 1434 int 1435 vfs_bio_awrite(struct buf *bp) 1436 { 1437 int i; 1438 int j; 1439 daddr_t lblkno = bp->b_lblkno; 1440 struct vnode *vp = bp->b_vp; 1441 int ncl; 1442 struct buf *bpa; 1443 int nwritten; 1444 int size; 1445 int maxcl; 1446 1447 crit_enter(); 1448 /* 1449 * right now we support clustered writing only to regular files. If 1450 * we find a clusterable block we could be in the middle of a cluster 1451 * rather then at the beginning. 1452 * 1453 * NOTE: b_bio1 contains the logical loffset/lblkno and is aliased 1454 * to b_lblkno and b_loffset. b_bio2 contains the translated block 1455 * number. 1456 */ 1457 if ((vp->v_type == VREG) && 1458 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1459 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1460 1461 size = vp->v_mount->mnt_stat.f_iosize; 1462 maxcl = MAXPHYS / size; 1463 1464 for (i = 1; i < maxcl; i++) { 1465 if ((bpa = findblk(vp, lblkno + i)) && 1466 BUF_REFCNT(bpa) == 0 && 1467 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1468 (B_DELWRI | B_CLUSTEROK)) && 1469 (bpa->b_bufsize == size)) { 1470 if ((bpa->b_bio2.bio_blkno == (daddr_t)-1) || 1471 (bpa->b_bio2.bio_blkno != 1472 bp->b_bio2.bio_blkno + ((i * size) >> DEV_BSHIFT))) 1473 break; 1474 } else { 1475 break; 1476 } 1477 } 1478 for (j = 1; i + j <= maxcl && j <= lblkno; j++) { 1479 if ((bpa = findblk(vp, lblkno - j)) && 1480 BUF_REFCNT(bpa) == 0 && 1481 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1482 (B_DELWRI | B_CLUSTEROK)) && 1483 (bpa->b_bufsize == size)) { 1484 if ((bpa->b_bio2.bio_blkno == (daddr_t)-1) || 1485 (bpa->b_bio2.bio_blkno != 1486 bp->b_bio2.bio_blkno - ((j * size) >> DEV_BSHIFT))) 1487 break; 1488 } else { 1489 break; 1490 } 1491 } 1492 --j; 1493 ncl = i + j; 1494 /* 1495 * this is a possible cluster write 1496 */ 1497 if (ncl != 1) { 1498 BUF_UNLOCK(bp); 1499 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1500 crit_exit(); 1501 return nwritten; 1502 } 1503 } 1504 1505 bremfree(bp); 1506 bp->b_flags |= B_ASYNC; 1507 1508 crit_exit(); 1509 /* 1510 * default (old) behavior, writing out only one block 1511 * 1512 * XXX returns b_bufsize instead of b_bcount for nwritten? 1513 */ 1514 nwritten = bp->b_bufsize; 1515 (void) VOP_BWRITE(bp->b_vp, bp); 1516 1517 return nwritten; 1518 } 1519 1520 /* 1521 * getnewbuf: 1522 * 1523 * Find and initialize a new buffer header, freeing up existing buffers 1524 * in the bufqueues as necessary. The new buffer is returned locked. 1525 * 1526 * Important: B_INVAL is not set. If the caller wishes to throw the 1527 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1528 * 1529 * We block if: 1530 * We have insufficient buffer headers 1531 * We have insufficient buffer space 1532 * buffer_map is too fragmented ( space reservation fails ) 1533 * If we have to flush dirty buffers ( but we try to avoid this ) 1534 * 1535 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1536 * Instead we ask the buf daemon to do it for us. We attempt to 1537 * avoid piecemeal wakeups of the pageout daemon. 1538 */ 1539 1540 static struct buf * 1541 getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1542 { 1543 struct buf *bp; 1544 struct buf *nbp; 1545 int defrag = 0; 1546 int nqindex; 1547 static int flushingbufs; 1548 1549 /* 1550 * We can't afford to block since we might be holding a vnode lock, 1551 * which may prevent system daemons from running. We deal with 1552 * low-memory situations by proactively returning memory and running 1553 * async I/O rather then sync I/O. 1554 */ 1555 1556 ++getnewbufcalls; 1557 --getnewbufrestarts; 1558 restart: 1559 ++getnewbufrestarts; 1560 1561 /* 1562 * Setup for scan. If we do not have enough free buffers, 1563 * we setup a degenerate case that immediately fails. Note 1564 * that if we are specially marked process, we are allowed to 1565 * dip into our reserves. 1566 * 1567 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1568 * 1569 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1570 * However, there are a number of cases (defragging, reusing, ...) 1571 * where we cannot backup. 1572 */ 1573 nqindex = BQUEUE_EMPTYKVA; 1574 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]); 1575 1576 if (nbp == NULL) { 1577 /* 1578 * If no EMPTYKVA buffers and we are either 1579 * defragging or reusing, locate a CLEAN buffer 1580 * to free or reuse. If bufspace useage is low 1581 * skip this step so we can allocate a new buffer. 1582 */ 1583 if (defrag || bufspace >= lobufspace) { 1584 nqindex = BQUEUE_CLEAN; 1585 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]); 1586 } 1587 1588 /* 1589 * If we could not find or were not allowed to reuse a 1590 * CLEAN buffer, check to see if it is ok to use an EMPTY 1591 * buffer. We can only use an EMPTY buffer if allocating 1592 * its KVA would not otherwise run us out of buffer space. 1593 */ 1594 if (nbp == NULL && defrag == 0 && 1595 bufspace + maxsize < hibufspace) { 1596 nqindex = BQUEUE_EMPTY; 1597 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]); 1598 } 1599 } 1600 1601 /* 1602 * Run scan, possibly freeing data and/or kva mappings on the fly 1603 * depending. 1604 */ 1605 1606 while ((bp = nbp) != NULL) { 1607 int qindex = nqindex; 1608 1609 /* 1610 * Calculate next bp ( we can only use it if we do not block 1611 * or do other fancy things ). 1612 */ 1613 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1614 switch(qindex) { 1615 case BQUEUE_EMPTY: 1616 nqindex = BQUEUE_EMPTYKVA; 1617 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]))) 1618 break; 1619 /* fall through */ 1620 case BQUEUE_EMPTYKVA: 1621 nqindex = BQUEUE_CLEAN; 1622 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]))) 1623 break; 1624 /* fall through */ 1625 case BQUEUE_CLEAN: 1626 /* 1627 * nbp is NULL. 1628 */ 1629 break; 1630 } 1631 } 1632 1633 /* 1634 * Sanity Checks 1635 */ 1636 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1637 1638 /* 1639 * Note: we no longer distinguish between VMIO and non-VMIO 1640 * buffers. 1641 */ 1642 1643 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1644 1645 /* 1646 * If we are defragging then we need a buffer with 1647 * b_kvasize != 0. XXX this situation should no longer 1648 * occur, if defrag is non-zero the buffer's b_kvasize 1649 * should also be non-zero at this point. XXX 1650 */ 1651 if (defrag && bp->b_kvasize == 0) { 1652 printf("Warning: defrag empty buffer %p\n", bp); 1653 continue; 1654 } 1655 1656 /* 1657 * Start freeing the bp. This is somewhat involved. nbp 1658 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers 1659 * on the clean list must be disassociated from their 1660 * current vnode. Buffers on the empty[kva] lists have 1661 * already been disassociated. 1662 */ 1663 1664 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 1665 printf("getnewbuf: warning, locked buf %p, race corrected\n", bp); 1666 tsleep(&bd_request, 0, "gnbxxx", hz / 100); 1667 goto restart; 1668 } 1669 if (bp->b_qindex != qindex) { 1670 printf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex); 1671 BUF_UNLOCK(bp); 1672 goto restart; 1673 } 1674 bremfree(bp); 1675 1676 if (qindex == BQUEUE_CLEAN) { 1677 if (bp->b_flags & B_VMIO) { 1678 bp->b_flags &= ~B_ASYNC; 1679 vfs_vmio_release(bp); 1680 } 1681 if (bp->b_vp) 1682 brelvp(bp); 1683 } 1684 1685 /* 1686 * NOTE: nbp is now entirely invalid. We can only restart 1687 * the scan from this point on. 1688 * 1689 * Get the rest of the buffer freed up. b_kva* is still 1690 * valid after this operation. 1691 */ 1692 1693 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08lx vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex)); 1694 KKASSERT((bp->b_flags & B_HASHED) == 0); 1695 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 1696 (*bioops.io_deallocate)(bp); 1697 if (bp->b_xflags & BX_BKGRDINPROG) 1698 panic("losing buffer 3"); 1699 1700 /* 1701 * critical section protection is not required when 1702 * scrapping a buffer's contents because it is already 1703 * wired. 1704 */ 1705 if (bp->b_bufsize) 1706 allocbuf(bp, 0); 1707 1708 bp->b_flags = 0; 1709 bp->b_xflags = 0; 1710 bp->b_vp = NULL; 1711 bp->b_error = 0; 1712 bp->b_resid = 0; 1713 bp->b_bcount = 0; 1714 bp->b_xio.xio_npages = 0; 1715 bp->b_dirtyoff = bp->b_dirtyend = 0; 1716 reinitbufbio(bp); 1717 1718 LIST_INIT(&bp->b_dep); 1719 1720 /* 1721 * If we are defragging then free the buffer. 1722 */ 1723 if (defrag) { 1724 bp->b_flags |= B_INVAL; 1725 bfreekva(bp); 1726 brelse(bp); 1727 defrag = 0; 1728 goto restart; 1729 } 1730 1731 /* 1732 * If we are overcomitted then recover the buffer and its 1733 * KVM space. This occurs in rare situations when multiple 1734 * processes are blocked in getnewbuf() or allocbuf(). 1735 */ 1736 if (bufspace >= hibufspace) 1737 flushingbufs = 1; 1738 if (flushingbufs && bp->b_kvasize != 0) { 1739 bp->b_flags |= B_INVAL; 1740 bfreekva(bp); 1741 brelse(bp); 1742 goto restart; 1743 } 1744 if (bufspace < lobufspace) 1745 flushingbufs = 0; 1746 break; 1747 } 1748 1749 /* 1750 * If we exhausted our list, sleep as appropriate. We may have to 1751 * wakeup various daemons and write out some dirty buffers. 1752 * 1753 * Generally we are sleeping due to insufficient buffer space. 1754 */ 1755 1756 if (bp == NULL) { 1757 int flags; 1758 char *waitmsg; 1759 1760 if (defrag) { 1761 flags = VFS_BIO_NEED_BUFSPACE; 1762 waitmsg = "nbufkv"; 1763 } else if (bufspace >= hibufspace) { 1764 waitmsg = "nbufbs"; 1765 flags = VFS_BIO_NEED_BUFSPACE; 1766 } else { 1767 waitmsg = "newbuf"; 1768 flags = VFS_BIO_NEED_ANY; 1769 } 1770 1771 bd_speedup(); /* heeeelp */ 1772 1773 needsbuffer |= flags; 1774 while (needsbuffer & flags) { 1775 if (tsleep(&needsbuffer, slpflag, waitmsg, slptimeo)) 1776 return (NULL); 1777 } 1778 } else { 1779 /* 1780 * We finally have a valid bp. We aren't quite out of the 1781 * woods, we still have to reserve kva space. In order 1782 * to keep fragmentation sane we only allocate kva in 1783 * BKVASIZE chunks. 1784 */ 1785 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1786 1787 if (maxsize != bp->b_kvasize) { 1788 vm_offset_t addr = 0; 1789 int count; 1790 1791 bfreekva(bp); 1792 1793 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1794 vm_map_lock(buffer_map); 1795 1796 if (vm_map_findspace(buffer_map, 1797 vm_map_min(buffer_map), maxsize, 1798 maxsize, &addr)) { 1799 /* 1800 * Uh oh. Buffer map is too fragmented. We 1801 * must defragment the map. 1802 */ 1803 vm_map_unlock(buffer_map); 1804 vm_map_entry_release(count); 1805 ++bufdefragcnt; 1806 defrag = 1; 1807 bp->b_flags |= B_INVAL; 1808 brelse(bp); 1809 goto restart; 1810 } 1811 if (addr) { 1812 vm_map_insert(buffer_map, &count, 1813 NULL, 0, 1814 addr, addr + maxsize, 1815 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1816 1817 bp->b_kvabase = (caddr_t) addr; 1818 bp->b_kvasize = maxsize; 1819 bufspace += bp->b_kvasize; 1820 ++bufreusecnt; 1821 } 1822 vm_map_unlock(buffer_map); 1823 vm_map_entry_release(count); 1824 } 1825 bp->b_data = bp->b_kvabase; 1826 } 1827 return(bp); 1828 } 1829 1830 /* 1831 * buf_daemon: 1832 * 1833 * Buffer flushing daemon. Buffers are normally flushed by the 1834 * update daemon but if it cannot keep up this process starts to 1835 * take the load in an attempt to prevent getnewbuf() from blocking. 1836 */ 1837 1838 static struct thread *bufdaemonthread; 1839 1840 static struct kproc_desc buf_kp = { 1841 "bufdaemon", 1842 buf_daemon, 1843 &bufdaemonthread 1844 }; 1845 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1846 1847 static void 1848 buf_daemon() 1849 { 1850 /* 1851 * This process needs to be suspended prior to shutdown sync. 1852 */ 1853 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 1854 bufdaemonthread, SHUTDOWN_PRI_LAST); 1855 1856 /* 1857 * This process is allowed to take the buffer cache to the limit 1858 */ 1859 crit_enter(); 1860 1861 for (;;) { 1862 kproc_suspend_loop(); 1863 1864 /* 1865 * Do the flush. Limit the amount of in-transit I/O we 1866 * allow to build up, otherwise we would completely saturate 1867 * the I/O system. Wakeup any waiting processes before we 1868 * normally would so they can run in parallel with our drain. 1869 */ 1870 while (numdirtybuffers > lodirtybuffers) { 1871 if (flushbufqueues() == 0) 1872 break; 1873 waitrunningbufspace(); 1874 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 1875 } 1876 1877 /* 1878 * Only clear bd_request if we have reached our low water 1879 * mark. The buf_daemon normally waits 5 seconds and 1880 * then incrementally flushes any dirty buffers that have 1881 * built up, within reason. 1882 * 1883 * If we were unable to hit our low water mark and couldn't 1884 * find any flushable buffers, we sleep half a second. 1885 * Otherwise we loop immediately. 1886 */ 1887 if (numdirtybuffers <= lodirtybuffers) { 1888 /* 1889 * We reached our low water mark, reset the 1890 * request and sleep until we are needed again. 1891 * The sleep is just so the suspend code works. 1892 */ 1893 bd_request = 0; 1894 tsleep(&bd_request, 0, "psleep", hz); 1895 } else { 1896 /* 1897 * We couldn't find any flushable dirty buffers but 1898 * still have too many dirty buffers, we 1899 * have to sleep and try again. (rare) 1900 */ 1901 tsleep(&bd_request, 0, "qsleep", hz / 2); 1902 } 1903 } 1904 } 1905 1906 /* 1907 * flushbufqueues: 1908 * 1909 * Try to flush a buffer in the dirty queue. We must be careful to 1910 * free up B_INVAL buffers instead of write them, which NFS is 1911 * particularly sensitive to. 1912 */ 1913 1914 static int 1915 flushbufqueues(void) 1916 { 1917 struct buf *bp; 1918 int r = 0; 1919 1920 bp = TAILQ_FIRST(&bufqueues[BQUEUE_DIRTY]); 1921 1922 while (bp) { 1923 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); 1924 if ((bp->b_flags & B_DELWRI) != 0 && 1925 (bp->b_xflags & BX_BKGRDINPROG) == 0) { 1926 if (bp->b_flags & B_INVAL) { 1927 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1928 panic("flushbufqueues: locked buf"); 1929 bremfree(bp); 1930 brelse(bp); 1931 ++r; 1932 break; 1933 } 1934 if (LIST_FIRST(&bp->b_dep) != NULL && 1935 bioops.io_countdeps && 1936 (bp->b_flags & B_DEFERRED) == 0 && 1937 (*bioops.io_countdeps)(bp, 0)) { 1938 TAILQ_REMOVE(&bufqueues[BQUEUE_DIRTY], 1939 bp, b_freelist); 1940 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], 1941 bp, b_freelist); 1942 bp->b_flags |= B_DEFERRED; 1943 bp = TAILQ_FIRST(&bufqueues[BQUEUE_DIRTY]); 1944 continue; 1945 } 1946 1947 /* 1948 * Only write it out if we can successfully lock 1949 * it. 1950 */ 1951 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 1952 vfs_bio_awrite(bp); 1953 ++r; 1954 break; 1955 } 1956 } 1957 bp = TAILQ_NEXT(bp, b_freelist); 1958 } 1959 return (r); 1960 } 1961 1962 /* 1963 * inmem: 1964 * 1965 * Returns true if no I/O is needed to access the associated VM object. 1966 * This is like findblk except it also hunts around in the VM system for 1967 * the data. 1968 * 1969 * Note that we ignore vm_page_free() races from interrupts against our 1970 * lookup, since if the caller is not protected our return value will not 1971 * be any more valid then otherwise once we exit the critical section. 1972 */ 1973 int 1974 inmem(struct vnode * vp, daddr_t blkno) 1975 { 1976 vm_object_t obj; 1977 vm_offset_t toff, tinc, size; 1978 vm_page_t m; 1979 vm_ooffset_t off; 1980 1981 if (findblk(vp, blkno)) 1982 return 1; 1983 if (vp->v_mount == NULL) 1984 return 0; 1985 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_flag & VOBJBUF) == 0) 1986 return 0; 1987 1988 size = PAGE_SIZE; 1989 if (size > vp->v_mount->mnt_stat.f_iosize) 1990 size = vp->v_mount->mnt_stat.f_iosize; 1991 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 1992 1993 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1994 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 1995 if (!m) 1996 return 0; 1997 tinc = size; 1998 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 1999 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2000 if (vm_page_is_valid(m, 2001 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2002 return 0; 2003 } 2004 return 1; 2005 } 2006 2007 /* 2008 * vfs_setdirty: 2009 * 2010 * Sets the dirty range for a buffer based on the status of the dirty 2011 * bits in the pages comprising the buffer. 2012 * 2013 * The range is limited to the size of the buffer. 2014 * 2015 * This routine is primarily used by NFS, but is generalized for the 2016 * B_VMIO case. 2017 */ 2018 static void 2019 vfs_setdirty(struct buf *bp) 2020 { 2021 int i; 2022 vm_object_t object; 2023 2024 /* 2025 * Degenerate case - empty buffer 2026 */ 2027 2028 if (bp->b_bufsize == 0) 2029 return; 2030 2031 /* 2032 * We qualify the scan for modified pages on whether the 2033 * object has been flushed yet. The OBJ_WRITEABLE flag 2034 * is not cleared simply by protecting pages off. 2035 */ 2036 2037 if ((bp->b_flags & B_VMIO) == 0) 2038 return; 2039 2040 object = bp->b_xio.xio_pages[0]->object; 2041 2042 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2043 printf("Warning: object %p writeable but not mightbedirty\n", object); 2044 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2045 printf("Warning: object %p mightbedirty but not writeable\n", object); 2046 2047 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2048 vm_offset_t boffset; 2049 vm_offset_t eoffset; 2050 2051 /* 2052 * test the pages to see if they have been modified directly 2053 * by users through the VM system. 2054 */ 2055 for (i = 0; i < bp->b_xio.xio_npages; i++) { 2056 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO); 2057 vm_page_test_dirty(bp->b_xio.xio_pages[i]); 2058 } 2059 2060 /* 2061 * Calculate the encompassing dirty range, boffset and eoffset, 2062 * (eoffset - boffset) bytes. 2063 */ 2064 2065 for (i = 0; i < bp->b_xio.xio_npages; i++) { 2066 if (bp->b_xio.xio_pages[i]->dirty) 2067 break; 2068 } 2069 boffset = (i << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK); 2070 2071 for (i = bp->b_xio.xio_npages - 1; i >= 0; --i) { 2072 if (bp->b_xio.xio_pages[i]->dirty) { 2073 break; 2074 } 2075 } 2076 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK); 2077 2078 /* 2079 * Fit it to the buffer. 2080 */ 2081 2082 if (eoffset > bp->b_bcount) 2083 eoffset = bp->b_bcount; 2084 2085 /* 2086 * If we have a good dirty range, merge with the existing 2087 * dirty range. 2088 */ 2089 2090 if (boffset < eoffset) { 2091 if (bp->b_dirtyoff > boffset) 2092 bp->b_dirtyoff = boffset; 2093 if (bp->b_dirtyend < eoffset) 2094 bp->b_dirtyend = eoffset; 2095 } 2096 } 2097 } 2098 2099 /* 2100 * findblk: 2101 * 2102 * Locate and return the specified buffer, or NULL if the buffer does 2103 * not exist. Do not attempt to lock the buffer or manipulate it in 2104 * any way. The caller must validate that the correct buffer has been 2105 * obtain after locking it. 2106 */ 2107 struct buf * 2108 findblk(struct vnode *vp, daddr_t blkno) 2109 { 2110 struct buf *bp; 2111 2112 crit_enter(); 2113 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, blkno); 2114 crit_exit(); 2115 return(bp); 2116 } 2117 2118 /* 2119 * getblk: 2120 * 2121 * Get a block given a specified block and offset into a file/device. 2122 * The buffers B_DONE bit will be cleared on return, making it almost 2123 * ready for an I/O initiation. B_INVAL may or may not be set on 2124 * return. The caller should clear B_INVAL prior to initiating a 2125 * READ. 2126 * 2127 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE 2128 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ, 2129 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer 2130 * without doing any of those things the system will likely believe 2131 * the buffer to be valid (especially if it is not B_VMIO), and the 2132 * next getblk() will return the buffer with B_CACHE set. 2133 * 2134 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2135 * an existing buffer. 2136 * 2137 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2138 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2139 * and then cleared based on the backing VM. If the previous buffer is 2140 * non-0-sized but invalid, B_CACHE will be cleared. 2141 * 2142 * If getblk() must create a new buffer, the new buffer is returned with 2143 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2144 * case it is returned with B_INVAL clear and B_CACHE set based on the 2145 * backing VM. 2146 * 2147 * getblk() also forces a VOP_BWRITE() for any B_DELWRI buffer whos 2148 * B_CACHE bit is clear. 2149 * 2150 * What this means, basically, is that the caller should use B_CACHE to 2151 * determine whether the buffer is fully valid or not and should clear 2152 * B_INVAL prior to issuing a read. If the caller intends to validate 2153 * the buffer by loading its data area with something, the caller needs 2154 * to clear B_INVAL. If the caller does this without issuing an I/O, 2155 * the caller should set B_CACHE ( as an optimization ), else the caller 2156 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2157 * a write attempt or if it was a successfull read. If the caller 2158 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR 2159 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2160 */ 2161 struct buf * 2162 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2163 { 2164 struct buf *bp; 2165 2166 if (size > MAXBSIZE) 2167 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE); 2168 2169 crit_enter(); 2170 loop: 2171 /* 2172 * Block if we are low on buffers. Certain processes are allowed 2173 * to completely exhaust the buffer cache. 2174 * 2175 * If this check ever becomes a bottleneck it may be better to 2176 * move it into the else, when findblk() fails. At the moment 2177 * it isn't a problem. 2178 * 2179 * XXX remove, we cannot afford to block anywhere if holding a vnode 2180 * lock in low-memory situation, so take it to the max. 2181 */ 2182 if (numfreebuffers == 0) { 2183 if (!curproc) 2184 return NULL; 2185 needsbuffer |= VFS_BIO_NEED_ANY; 2186 tsleep(&needsbuffer, slpflag, "newbuf", slptimeo); 2187 } 2188 2189 if ((bp = findblk(vp, blkno))) { 2190 /* 2191 * The buffer was found in the cache, but we need to lock it. 2192 * Even with LK_NOWAIT the lockmgr may break our critical 2193 * section, so double-check the validity of the buffer 2194 * once the lock has been obtained. 2195 */ 2196 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2197 int lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL; 2198 if (slpflag & PCATCH) 2199 lkflags |= LK_PCATCH; 2200 if (BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo) == 2201 ENOLCK) { 2202 goto loop; 2203 } 2204 crit_exit(); 2205 return (NULL); 2206 } 2207 2208 /* 2209 * Once the buffer has been locked, make sure we didn't race 2210 * a buffer recyclement. Buffers that are no longer hashed 2211 * will have b_vp == NULL, so this takes care of that check 2212 * as well. 2213 */ 2214 if (bp->b_vp != vp || bp->b_lblkno != blkno) { 2215 printf("Warning buffer %p (vp %p lblkno %d) was recycled\n", bp, vp, (int)blkno); 2216 BUF_UNLOCK(bp); 2217 goto loop; 2218 } 2219 2220 /* 2221 * Make sure that B_INVAL buffers do not have a cached 2222 * block number translation. 2223 */ 2224 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_blkno != (daddr_t)-1)) { 2225 printf("Warning invalid buffer %p (vp %p lblkno %d) did not have cleared bio_blkno cache\n", bp, vp, (int)blkno); 2226 clearbiocache(&bp->b_bio2); 2227 } 2228 2229 /* 2230 * The buffer is locked. B_CACHE is cleared if the buffer is 2231 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2232 * and for a VMIO buffer B_CACHE is adjusted according to the 2233 * backing VM cache. 2234 */ 2235 if (bp->b_flags & B_INVAL) 2236 bp->b_flags &= ~B_CACHE; 2237 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2238 bp->b_flags |= B_CACHE; 2239 bremfree(bp); 2240 2241 /* 2242 * check for size inconsistancies for non-VMIO case. 2243 */ 2244 2245 if (bp->b_bcount != size) { 2246 if ((bp->b_flags & B_VMIO) == 0 || 2247 (size > bp->b_kvasize)) { 2248 if (bp->b_flags & B_DELWRI) { 2249 bp->b_flags |= B_NOCACHE; 2250 VOP_BWRITE(bp->b_vp, bp); 2251 } else { 2252 if ((bp->b_flags & B_VMIO) && 2253 (LIST_FIRST(&bp->b_dep) == NULL)) { 2254 bp->b_flags |= B_RELBUF; 2255 brelse(bp); 2256 } else { 2257 bp->b_flags |= B_NOCACHE; 2258 VOP_BWRITE(bp->b_vp, bp); 2259 } 2260 } 2261 goto loop; 2262 } 2263 } 2264 2265 /* 2266 * If the size is inconsistant in the VMIO case, we can resize 2267 * the buffer. This might lead to B_CACHE getting set or 2268 * cleared. If the size has not changed, B_CACHE remains 2269 * unchanged from its previous state. 2270 */ 2271 2272 if (bp->b_bcount != size) 2273 allocbuf(bp, size); 2274 2275 KASSERT(bp->b_loffset != NOOFFSET, 2276 ("getblk: no buffer offset")); 2277 2278 /* 2279 * A buffer with B_DELWRI set and B_CACHE clear must 2280 * be committed before we can return the buffer in 2281 * order to prevent the caller from issuing a read 2282 * ( due to B_CACHE not being set ) and overwriting 2283 * it. 2284 * 2285 * Most callers, including NFS and FFS, need this to 2286 * operate properly either because they assume they 2287 * can issue a read if B_CACHE is not set, or because 2288 * ( for example ) an uncached B_DELWRI might loop due 2289 * to softupdates re-dirtying the buffer. In the latter 2290 * case, B_CACHE is set after the first write completes, 2291 * preventing further loops. 2292 * 2293 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2294 * above while extending the buffer, we cannot allow the 2295 * buffer to remain with B_CACHE set after the write 2296 * completes or it will represent a corrupt state. To 2297 * deal with this we set B_NOCACHE to scrap the buffer 2298 * after the write. 2299 * 2300 * We might be able to do something fancy, like setting 2301 * B_CACHE in bwrite() except if B_DELWRI is already set, 2302 * so the below call doesn't set B_CACHE, but that gets real 2303 * confusing. This is much easier. 2304 */ 2305 2306 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2307 bp->b_flags |= B_NOCACHE; 2308 VOP_BWRITE(bp->b_vp, bp); 2309 goto loop; 2310 } 2311 2312 crit_exit(); 2313 bp->b_flags &= ~B_DONE; 2314 } else { 2315 /* 2316 * Buffer is not in-core, create new buffer. The buffer 2317 * returned by getnewbuf() is locked. Note that the returned 2318 * buffer is also considered valid (not marked B_INVAL). 2319 * 2320 * Calculating the offset for the I/O requires figuring out 2321 * the block size. We use DEV_BSIZE for VBLK or VCHR and 2322 * the mount's f_iosize otherwise. If the vnode does not 2323 * have an associated mount we assume that the passed size is 2324 * the block size. 2325 * 2326 * Note that vn_isdisk() cannot be used here since it may 2327 * return a failure for numerous reasons. Note that the 2328 * buffer size may be larger then the block size (the caller 2329 * will use block numbers with the proper multiple). Beware 2330 * of using any v_* fields which are part of unions. In 2331 * particular, in DragonFly the mount point overloading 2332 * mechanism is such that the underlying directory (with a 2333 * non-NULL v_mountedhere) is not a special case. 2334 */ 2335 int bsize, maxsize, vmio; 2336 off_t offset; 2337 2338 if (vp->v_type == VBLK || vp->v_type == VCHR) 2339 bsize = DEV_BSIZE; 2340 else if (vp->v_mount) 2341 bsize = vp->v_mount->mnt_stat.f_iosize; 2342 else 2343 bsize = size; 2344 2345 offset = (off_t)blkno * bsize; 2346 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && (vp->v_flag & VOBJBUF); 2347 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2348 maxsize = imax(maxsize, bsize); 2349 2350 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2351 if (slpflag || slptimeo) { 2352 crit_exit(); 2353 return NULL; 2354 } 2355 goto loop; 2356 } 2357 2358 /* 2359 * This code is used to make sure that a buffer is not 2360 * created while the getnewbuf routine is blocked. 2361 * This can be a problem whether the vnode is locked or not. 2362 * If the buffer is created out from under us, we have to 2363 * throw away the one we just created. There is now window 2364 * race because we are safely running in a critical section 2365 * from the point of the duplicate buffer creation through 2366 * to here, and we've locked the buffer. 2367 */ 2368 if (findblk(vp, blkno)) { 2369 bp->b_flags |= B_INVAL; 2370 brelse(bp); 2371 goto loop; 2372 } 2373 2374 /* 2375 * Insert the buffer into the hash, so that it can 2376 * be found by findblk(). 2377 * 2378 * Make sure the translation layer has been cleared. 2379 */ 2380 bp->b_lblkno = blkno; 2381 bp->b_loffset = offset; 2382 bp->b_bio2.bio_blkno = (daddr_t)-1; 2383 /* bp->b_bio2.bio_next = NULL; */ 2384 2385 bgetvp(vp, bp); 2386 2387 /* 2388 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2389 * buffer size starts out as 0, B_CACHE will be set by 2390 * allocbuf() for the VMIO case prior to it testing the 2391 * backing store for validity. 2392 */ 2393 2394 if (vmio) { 2395 bp->b_flags |= B_VMIO; 2396 #if defined(VFS_BIO_DEBUG) 2397 if (vn_canvmio(vp) != TRUE) 2398 printf("getblk: vmioing file type %d???\n", vp->v_type); 2399 #endif 2400 } else { 2401 bp->b_flags &= ~B_VMIO; 2402 } 2403 2404 allocbuf(bp, size); 2405 2406 crit_exit(); 2407 bp->b_flags &= ~B_DONE; 2408 } 2409 return (bp); 2410 } 2411 2412 /* 2413 * geteblk: 2414 * 2415 * Get an empty, disassociated buffer of given size. The buffer is 2416 * initially set to B_INVAL. 2417 * 2418 * critical section protection is not required for the allocbuf() 2419 * call because races are impossible here. 2420 */ 2421 struct buf * 2422 geteblk(int size) 2423 { 2424 struct buf *bp; 2425 int maxsize; 2426 2427 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2428 2429 crit_enter(); 2430 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2431 ; 2432 crit_exit(); 2433 allocbuf(bp, size); 2434 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2435 return (bp); 2436 } 2437 2438 2439 /* 2440 * allocbuf: 2441 * 2442 * This code constitutes the buffer memory from either anonymous system 2443 * memory (in the case of non-VMIO operations) or from an associated 2444 * VM object (in the case of VMIO operations). This code is able to 2445 * resize a buffer up or down. 2446 * 2447 * Note that this code is tricky, and has many complications to resolve 2448 * deadlock or inconsistant data situations. Tread lightly!!! 2449 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2450 * the caller. Calling this code willy nilly can result in the loss of data. 2451 * 2452 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2453 * B_CACHE for the non-VMIO case. 2454 * 2455 * This routine does not need to be called from a critical section but you 2456 * must own the buffer. 2457 */ 2458 int 2459 allocbuf(struct buf *bp, int size) 2460 { 2461 int newbsize, mbsize; 2462 int i; 2463 2464 if (BUF_REFCNT(bp) == 0) 2465 panic("allocbuf: buffer not busy"); 2466 2467 if (bp->b_kvasize < size) 2468 panic("allocbuf: buffer too small"); 2469 2470 if ((bp->b_flags & B_VMIO) == 0) { 2471 caddr_t origbuf; 2472 int origbufsize; 2473 /* 2474 * Just get anonymous memory from the kernel. Don't 2475 * mess with B_CACHE. 2476 */ 2477 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2478 if (bp->b_flags & B_MALLOC) 2479 newbsize = mbsize; 2480 else 2481 newbsize = round_page(size); 2482 2483 if (newbsize < bp->b_bufsize) { 2484 /* 2485 * malloced buffers are not shrunk 2486 */ 2487 if (bp->b_flags & B_MALLOC) { 2488 if (newbsize) { 2489 bp->b_bcount = size; 2490 } else { 2491 free(bp->b_data, M_BIOBUF); 2492 if (bp->b_bufsize) { 2493 bufmallocspace -= bp->b_bufsize; 2494 bufspacewakeup(); 2495 bp->b_bufsize = 0; 2496 } 2497 bp->b_data = bp->b_kvabase; 2498 bp->b_bcount = 0; 2499 bp->b_flags &= ~B_MALLOC; 2500 } 2501 return 1; 2502 } 2503 vm_hold_free_pages( 2504 bp, 2505 (vm_offset_t) bp->b_data + newbsize, 2506 (vm_offset_t) bp->b_data + bp->b_bufsize); 2507 } else if (newbsize > bp->b_bufsize) { 2508 /* 2509 * We only use malloced memory on the first allocation. 2510 * and revert to page-allocated memory when the buffer 2511 * grows. 2512 */ 2513 if ( (bufmallocspace < maxbufmallocspace) && 2514 (bp->b_bufsize == 0) && 2515 (mbsize <= PAGE_SIZE/2)) { 2516 2517 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2518 bp->b_bufsize = mbsize; 2519 bp->b_bcount = size; 2520 bp->b_flags |= B_MALLOC; 2521 bufmallocspace += mbsize; 2522 return 1; 2523 } 2524 origbuf = NULL; 2525 origbufsize = 0; 2526 /* 2527 * If the buffer is growing on its other-than-first allocation, 2528 * then we revert to the page-allocation scheme. 2529 */ 2530 if (bp->b_flags & B_MALLOC) { 2531 origbuf = bp->b_data; 2532 origbufsize = bp->b_bufsize; 2533 bp->b_data = bp->b_kvabase; 2534 if (bp->b_bufsize) { 2535 bufmallocspace -= bp->b_bufsize; 2536 bufspacewakeup(); 2537 bp->b_bufsize = 0; 2538 } 2539 bp->b_flags &= ~B_MALLOC; 2540 newbsize = round_page(newbsize); 2541 } 2542 vm_hold_load_pages( 2543 bp, 2544 (vm_offset_t) bp->b_data + bp->b_bufsize, 2545 (vm_offset_t) bp->b_data + newbsize); 2546 if (origbuf) { 2547 bcopy(origbuf, bp->b_data, origbufsize); 2548 free(origbuf, M_BIOBUF); 2549 } 2550 } 2551 } else { 2552 vm_page_t m; 2553 int desiredpages; 2554 2555 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2556 desiredpages = (size == 0) ? 0 : 2557 num_pages((bp->b_loffset & PAGE_MASK) + newbsize); 2558 2559 if (bp->b_flags & B_MALLOC) 2560 panic("allocbuf: VMIO buffer can't be malloced"); 2561 /* 2562 * Set B_CACHE initially if buffer is 0 length or will become 2563 * 0-length. 2564 */ 2565 if (size == 0 || bp->b_bufsize == 0) 2566 bp->b_flags |= B_CACHE; 2567 2568 if (newbsize < bp->b_bufsize) { 2569 /* 2570 * DEV_BSIZE aligned new buffer size is less then the 2571 * DEV_BSIZE aligned existing buffer size. Figure out 2572 * if we have to remove any pages. 2573 */ 2574 if (desiredpages < bp->b_xio.xio_npages) { 2575 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) { 2576 /* 2577 * the page is not freed here -- it 2578 * is the responsibility of 2579 * vnode_pager_setsize 2580 */ 2581 m = bp->b_xio.xio_pages[i]; 2582 KASSERT(m != bogus_page, 2583 ("allocbuf: bogus page found")); 2584 while (vm_page_sleep_busy(m, TRUE, "biodep")) 2585 ; 2586 2587 bp->b_xio.xio_pages[i] = NULL; 2588 vm_page_unwire(m, 0); 2589 } 2590 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2591 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages)); 2592 bp->b_xio.xio_npages = desiredpages; 2593 } 2594 } else if (size > bp->b_bcount) { 2595 /* 2596 * We are growing the buffer, possibly in a 2597 * byte-granular fashion. 2598 */ 2599 struct vnode *vp; 2600 vm_object_t obj; 2601 vm_offset_t toff; 2602 vm_offset_t tinc; 2603 2604 /* 2605 * Step 1, bring in the VM pages from the object, 2606 * allocating them if necessary. We must clear 2607 * B_CACHE if these pages are not valid for the 2608 * range covered by the buffer. 2609 * 2610 * critical section protection is required to protect 2611 * against interrupts unbusying and freeing pages 2612 * between our vm_page_lookup() and our 2613 * busycheck/wiring call. 2614 */ 2615 vp = bp->b_vp; 2616 VOP_GETVOBJECT(vp, &obj); 2617 2618 crit_enter(); 2619 while (bp->b_xio.xio_npages < desiredpages) { 2620 vm_page_t m; 2621 vm_pindex_t pi; 2622 2623 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages; 2624 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2625 /* 2626 * note: must allocate system pages 2627 * since blocking here could intefere 2628 * with paging I/O, no matter which 2629 * process we are. 2630 */ 2631 m = vm_page_alloc(obj, pi, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM); 2632 if (m == NULL) { 2633 vm_wait(); 2634 vm_pageout_deficit += desiredpages - 2635 bp->b_xio.xio_npages; 2636 } else { 2637 vm_page_wire(m); 2638 vm_page_wakeup(m); 2639 bp->b_flags &= ~B_CACHE; 2640 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 2641 ++bp->b_xio.xio_npages; 2642 } 2643 continue; 2644 } 2645 2646 /* 2647 * We found a page. If we have to sleep on it, 2648 * retry because it might have gotten freed out 2649 * from under us. 2650 * 2651 * We can only test PG_BUSY here. Blocking on 2652 * m->busy might lead to a deadlock: 2653 * 2654 * vm_fault->getpages->cluster_read->allocbuf 2655 * 2656 */ 2657 2658 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 2659 continue; 2660 2661 /* 2662 * We have a good page. Should we wakeup the 2663 * page daemon? 2664 */ 2665 if ((curthread != pagethread) && 2666 ((m->queue - m->pc) == PQ_CACHE) && 2667 ((vmstats.v_free_count + vmstats.v_cache_count) < 2668 (vmstats.v_free_min + vmstats.v_cache_min))) { 2669 pagedaemon_wakeup(); 2670 } 2671 vm_page_flag_clear(m, PG_ZERO); 2672 vm_page_wire(m); 2673 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 2674 ++bp->b_xio.xio_npages; 2675 } 2676 crit_exit(); 2677 2678 /* 2679 * Step 2. We've loaded the pages into the buffer, 2680 * we have to figure out if we can still have B_CACHE 2681 * set. Note that B_CACHE is set according to the 2682 * byte-granular range ( bcount and size ), not the 2683 * aligned range ( newbsize ). 2684 * 2685 * The VM test is against m->valid, which is DEV_BSIZE 2686 * aligned. Needless to say, the validity of the data 2687 * needs to also be DEV_BSIZE aligned. Note that this 2688 * fails with NFS if the server or some other client 2689 * extends the file's EOF. If our buffer is resized, 2690 * B_CACHE may remain set! XXX 2691 */ 2692 2693 toff = bp->b_bcount; 2694 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK); 2695 2696 while ((bp->b_flags & B_CACHE) && toff < size) { 2697 vm_pindex_t pi; 2698 2699 if (tinc > (size - toff)) 2700 tinc = size - toff; 2701 2702 pi = ((bp->b_loffset & PAGE_MASK) + toff) >> 2703 PAGE_SHIFT; 2704 2705 vfs_buf_test_cache( 2706 bp, 2707 bp->b_loffset, 2708 toff, 2709 tinc, 2710 bp->b_xio.xio_pages[pi] 2711 ); 2712 toff += tinc; 2713 tinc = PAGE_SIZE; 2714 } 2715 2716 /* 2717 * Step 3, fixup the KVM pmap. Remember that 2718 * bp->b_data is relative to bp->b_loffset, but 2719 * bp->b_loffset may be offset into the first page. 2720 */ 2721 2722 bp->b_data = (caddr_t) 2723 trunc_page((vm_offset_t)bp->b_data); 2724 pmap_qenter( 2725 (vm_offset_t)bp->b_data, 2726 bp->b_xio.xio_pages, 2727 bp->b_xio.xio_npages 2728 ); 2729 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2730 (vm_offset_t)(bp->b_loffset & PAGE_MASK)); 2731 } 2732 } 2733 if (newbsize < bp->b_bufsize) 2734 bufspacewakeup(); 2735 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2736 bp->b_bcount = size; /* requested buffer size */ 2737 return 1; 2738 } 2739 2740 /* 2741 * biowait: 2742 * 2743 * Wait for buffer I/O completion, returning error status. The buffer 2744 * is left locked and B_DONE on return. B_EINTR is converted into an 2745 * EINTR error and cleared. 2746 */ 2747 int 2748 biowait(struct buf * bp) 2749 { 2750 crit_enter(); 2751 while ((bp->b_flags & B_DONE) == 0) { 2752 if (bp->b_flags & B_READ) 2753 tsleep(bp, 0, "biord", 0); 2754 else 2755 tsleep(bp, 0, "biowr", 0); 2756 } 2757 crit_exit(); 2758 if (bp->b_flags & B_EINTR) { 2759 bp->b_flags &= ~B_EINTR; 2760 return (EINTR); 2761 } 2762 if (bp->b_flags & B_ERROR) { 2763 return (bp->b_error ? bp->b_error : EIO); 2764 } else { 2765 return (0); 2766 } 2767 } 2768 2769 /* 2770 * This associates a tracking count with an I/O. vn_strategy() and 2771 * dev_dstrategy() do this automatically but there are a few cases 2772 * where a vnode or device layer is bypassed when a block translation 2773 * is cached. In such cases bio_start_transaction() may be called on 2774 * the bypassed layers so the system gets an I/O in progress indication 2775 * for those higher layers. 2776 */ 2777 void 2778 bio_start_transaction(struct bio *bio, struct bio_track *track) 2779 { 2780 bio->bio_track = track; 2781 atomic_add_int(&track->bk_active, 1); 2782 } 2783 2784 /* 2785 * Initiate I/O on a vnode. 2786 */ 2787 void 2788 vn_strategy(struct vnode *vp, struct bio *bio) 2789 { 2790 struct bio_track *track; 2791 2792 if (bio->bio_buf->b_flags & B_READ) 2793 track = &vp->v_track_read; 2794 else 2795 track = &vp->v_track_write; 2796 bio->bio_track = track; 2797 atomic_add_int(&track->bk_active, 1); 2798 vop_strategy(*vp->v_ops, vp, bio); 2799 } 2800 2801 2802 /* 2803 * biodone: 2804 * 2805 * Finish I/O on a buffer, optionally calling a completion function. 2806 * This is usually called from an interrupt so process blocking is 2807 * not allowed. 2808 * 2809 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2810 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2811 * assuming B_INVAL is clear. 2812 * 2813 * For the VMIO case, we set B_CACHE if the op was a read and no 2814 * read error occured, or if the op was a write. B_CACHE is never 2815 * set if the buffer is invalid or otherwise uncacheable. 2816 * 2817 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2818 * initiator to leave B_INVAL set to brelse the buffer out of existance 2819 * in the biodone routine. 2820 */ 2821 void 2822 biodone(struct bio *bio) 2823 { 2824 struct buf *bp = bio->bio_buf; 2825 int error; 2826 2827 crit_enter(); 2828 2829 KASSERT(BUF_REFCNTNB(bp) > 0, 2830 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp))); 2831 KASSERT(!(bp->b_flags & B_DONE), 2832 ("biodone: bp %p already done", bp)); 2833 2834 bp->b_flags |= B_DONE; 2835 runningbufwakeup(bp); 2836 2837 /* 2838 * Run up the chain of BIO's. 2839 */ 2840 while (bio) { 2841 biodone_t *done_func; 2842 struct bio_track *track; 2843 2844 /* 2845 * BIO tracking. Most but not all BIOs are tracked. 2846 */ 2847 if ((track = bio->bio_track) != NULL) { 2848 atomic_subtract_int(&track->bk_active, 1); 2849 if (track->bk_active < 0) { 2850 panic("biodone: bad active count bio %p\n", 2851 bio); 2852 } 2853 if (track->bk_waitflag) { 2854 track->bk_waitflag = 0; 2855 wakeup(track); 2856 } 2857 bio->bio_track = NULL; 2858 } 2859 2860 /* 2861 * A bio_done function terminates the loop. The function 2862 * will be responsible for any further chaining and/or 2863 * buffer management. 2864 */ 2865 if ((done_func = bio->bio_done) != NULL) { 2866 bio->bio_done = NULL; 2867 done_func(bio); 2868 crit_exit(); 2869 return; 2870 } 2871 bio = bio->bio_prev; 2872 } 2873 2874 /* 2875 * Special case (XXX) - not a read or write. 2876 */ 2877 if (bp->b_flags & B_FREEBUF) { 2878 brelse(bp); 2879 crit_exit(); 2880 return; 2881 } 2882 2883 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 2884 (*bioops.io_complete)(bp); 2885 2886 if (bp->b_flags & B_VMIO) { 2887 int i; 2888 vm_ooffset_t foff; 2889 vm_page_t m; 2890 vm_object_t obj; 2891 int iosize; 2892 struct vnode *vp = bp->b_vp; 2893 2894 error = VOP_GETVOBJECT(vp, &obj); 2895 2896 #if defined(VFS_BIO_DEBUG) 2897 if (vp->v_holdcnt == 0) { 2898 panic("biodone: zero vnode hold count"); 2899 } 2900 2901 if (error) { 2902 panic("biodone: missing VM object"); 2903 } 2904 2905 if ((vp->v_flag & VOBJBUF) == 0) { 2906 panic("biodone: vnode is not setup for merged cache"); 2907 } 2908 #endif 2909 2910 foff = bp->b_loffset; 2911 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset")); 2912 2913 if (error) { 2914 panic("biodone: no object"); 2915 } 2916 #if defined(VFS_BIO_DEBUG) 2917 if (obj->paging_in_progress < bp->b_xio.xio_npages) { 2918 printf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n", 2919 obj->paging_in_progress, bp->b_xio.xio_npages); 2920 } 2921 #endif 2922 2923 /* 2924 * Set B_CACHE if the op was a normal read and no error 2925 * occured. B_CACHE is set for writes in the b*write() 2926 * routines. 2927 */ 2928 iosize = bp->b_bcount - bp->b_resid; 2929 if ((bp->b_flags & (B_READ|B_FREEBUF|B_INVAL|B_NOCACHE|B_ERROR)) == B_READ) { 2930 bp->b_flags |= B_CACHE; 2931 } 2932 2933 for (i = 0; i < bp->b_xio.xio_npages; i++) { 2934 int bogusflag = 0; 2935 int resid; 2936 2937 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2938 if (resid > iosize) 2939 resid = iosize; 2940 2941 /* 2942 * cleanup bogus pages, restoring the originals. Since 2943 * the originals should still be wired, we don't have 2944 * to worry about interrupt/freeing races destroying 2945 * the VM object association. 2946 */ 2947 m = bp->b_xio.xio_pages[i]; 2948 if (m == bogus_page) { 2949 bogusflag = 1; 2950 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2951 if (m == NULL) 2952 panic("biodone: page disappeared"); 2953 bp->b_xio.xio_pages[i] = m; 2954 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 2955 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 2956 } 2957 #if defined(VFS_BIO_DEBUG) 2958 if (OFF_TO_IDX(foff) != m->pindex) { 2959 printf( 2960 "biodone: foff(%lu)/m->pindex(%d) mismatch\n", 2961 (unsigned long)foff, m->pindex); 2962 } 2963 #endif 2964 2965 /* 2966 * In the write case, the valid and clean bits are 2967 * already changed correctly ( see bdwrite() ), so we 2968 * only need to do this here in the read case. 2969 */ 2970 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) { 2971 vfs_page_set_valid(bp, foff, i, m); 2972 } 2973 vm_page_flag_clear(m, PG_ZERO); 2974 2975 /* 2976 * when debugging new filesystems or buffer I/O methods, this 2977 * is the most common error that pops up. if you see this, you 2978 * have not set the page busy flag correctly!!! 2979 */ 2980 if (m->busy == 0) { 2981 printf("biodone: page busy < 0, " 2982 "pindex: %d, foff: 0x(%x,%x), " 2983 "resid: %d, index: %d\n", 2984 (int) m->pindex, (int)(foff >> 32), 2985 (int) foff & 0xffffffff, resid, i); 2986 if (!vn_isdisk(vp, NULL)) 2987 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 2988 bp->b_vp->v_mount->mnt_stat.f_iosize, 2989 (int) bp->b_lblkno, 2990 bp->b_flags, bp->b_xio.xio_npages); 2991 else 2992 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 2993 (int) bp->b_lblkno, 2994 bp->b_flags, bp->b_xio.xio_npages); 2995 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2996 m->valid, m->dirty, m->wire_count); 2997 panic("biodone: page busy < 0"); 2998 } 2999 vm_page_io_finish(m); 3000 vm_object_pip_subtract(obj, 1); 3001 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3002 iosize -= resid; 3003 } 3004 if (obj) 3005 vm_object_pip_wakeupn(obj, 0); 3006 } 3007 3008 /* 3009 * For asynchronous completions, release the buffer now. The brelse 3010 * will do a wakeup there if necessary - so no need to do a wakeup 3011 * here in the async case. The sync case always needs to do a wakeup. 3012 */ 3013 3014 if (bp->b_flags & B_ASYNC) { 3015 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0) 3016 brelse(bp); 3017 else 3018 bqrelse(bp); 3019 } else { 3020 wakeup(bp); 3021 } 3022 crit_exit(); 3023 } 3024 3025 /* 3026 * vfs_unbusy_pages: 3027 * 3028 * This routine is called in lieu of iodone in the case of 3029 * incomplete I/O. This keeps the busy status for pages 3030 * consistant. 3031 */ 3032 void 3033 vfs_unbusy_pages(struct buf *bp) 3034 { 3035 int i; 3036 3037 runningbufwakeup(bp); 3038 if (bp->b_flags & B_VMIO) { 3039 struct vnode *vp = bp->b_vp; 3040 vm_object_t obj; 3041 3042 VOP_GETVOBJECT(vp, &obj); 3043 3044 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3045 vm_page_t m = bp->b_xio.xio_pages[i]; 3046 3047 /* 3048 * When restoring bogus changes the original pages 3049 * should still be wired, so we are in no danger of 3050 * losing the object association and do not need 3051 * critical section protection particularly. 3052 */ 3053 if (m == bogus_page) { 3054 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i); 3055 if (!m) { 3056 panic("vfs_unbusy_pages: page missing"); 3057 } 3058 bp->b_xio.xio_pages[i] = m; 3059 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3060 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3061 } 3062 vm_object_pip_subtract(obj, 1); 3063 vm_page_flag_clear(m, PG_ZERO); 3064 vm_page_io_finish(m); 3065 } 3066 vm_object_pip_wakeupn(obj, 0); 3067 } 3068 } 3069 3070 /* 3071 * vfs_page_set_valid: 3072 * 3073 * Set the valid bits in a page based on the supplied offset. The 3074 * range is restricted to the buffer's size. 3075 * 3076 * This routine is typically called after a read completes. 3077 */ 3078 static void 3079 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3080 { 3081 vm_ooffset_t soff, eoff; 3082 3083 /* 3084 * Start and end offsets in buffer. eoff - soff may not cross a 3085 * page boundry or cross the end of the buffer. The end of the 3086 * buffer, in this case, is our file EOF, not the allocation size 3087 * of the buffer. 3088 */ 3089 soff = off; 3090 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3091 if (eoff > bp->b_loffset + bp->b_bcount) 3092 eoff = bp->b_loffset + bp->b_bcount; 3093 3094 /* 3095 * Set valid range. This is typically the entire buffer and thus the 3096 * entire page. 3097 */ 3098 if (eoff > soff) { 3099 vm_page_set_validclean( 3100 m, 3101 (vm_offset_t) (soff & PAGE_MASK), 3102 (vm_offset_t) (eoff - soff) 3103 ); 3104 } 3105 } 3106 3107 /* 3108 * vfs_busy_pages: 3109 * 3110 * This routine is called before a device strategy routine. 3111 * It is used to tell the VM system that paging I/O is in 3112 * progress, and treat the pages associated with the buffer 3113 * almost as being PG_BUSY. Also the object 'paging_in_progress' 3114 * flag is handled to make sure that the object doesn't become 3115 * inconsistant. 3116 * 3117 * Since I/O has not been initiated yet, certain buffer flags 3118 * such as B_ERROR or B_INVAL may be in an inconsistant state 3119 * and should be ignored. 3120 */ 3121 void 3122 vfs_busy_pages(struct buf *bp, int clear_modify) 3123 { 3124 int i, bogus; 3125 struct proc *p = curthread->td_proc; 3126 3127 if (bp->b_flags & B_VMIO) { 3128 struct vnode *vp = bp->b_vp; 3129 vm_object_t obj; 3130 vm_ooffset_t foff; 3131 3132 VOP_GETVOBJECT(vp, &obj); 3133 foff = bp->b_loffset; 3134 KASSERT(bp->b_loffset != NOOFFSET, 3135 ("vfs_busy_pages: no buffer offset")); 3136 vfs_setdirty(bp); 3137 3138 retry: 3139 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3140 vm_page_t m = bp->b_xio.xio_pages[i]; 3141 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 3142 goto retry; 3143 } 3144 3145 bogus = 0; 3146 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3147 vm_page_t m = bp->b_xio.xio_pages[i]; 3148 3149 vm_page_flag_clear(m, PG_ZERO); 3150 if ((bp->b_flags & B_CLUSTER) == 0) { 3151 vm_object_pip_add(obj, 1); 3152 vm_page_io_start(m); 3153 } 3154 3155 /* 3156 * When readying a buffer for a read ( i.e 3157 * clear_modify == 0 ), it is important to do 3158 * bogus_page replacement for valid pages in 3159 * partially instantiated buffers. Partially 3160 * instantiated buffers can, in turn, occur when 3161 * reconstituting a buffer from its VM backing store 3162 * base. We only have to do this if B_CACHE is 3163 * clear ( which causes the I/O to occur in the 3164 * first place ). The replacement prevents the read 3165 * I/O from overwriting potentially dirty VM-backed 3166 * pages. XXX bogus page replacement is, uh, bogus. 3167 * It may not work properly with small-block devices. 3168 * We need to find a better way. 3169 */ 3170 3171 vm_page_protect(m, VM_PROT_NONE); 3172 if (clear_modify) 3173 vfs_page_set_valid(bp, foff, i, m); 3174 else if (m->valid == VM_PAGE_BITS_ALL && 3175 (bp->b_flags & B_CACHE) == 0) { 3176 bp->b_xio.xio_pages[i] = bogus_page; 3177 bogus++; 3178 } 3179 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3180 } 3181 if (bogus) 3182 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3183 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3184 } 3185 3186 /* 3187 * This is the easiest place to put the process accounting for the I/O 3188 * for now. 3189 */ 3190 if (p != NULL) { 3191 if (bp->b_flags & B_READ) 3192 p->p_stats->p_ru.ru_inblock++; 3193 else 3194 p->p_stats->p_ru.ru_oublock++; 3195 } 3196 } 3197 3198 /* 3199 * vfs_clean_pages: 3200 * 3201 * Tell the VM system that the pages associated with this buffer 3202 * are clean. This is used for delayed writes where the data is 3203 * going to go to disk eventually without additional VM intevention. 3204 * 3205 * Note that while we only really need to clean through to b_bcount, we 3206 * just go ahead and clean through to b_bufsize. 3207 */ 3208 static void 3209 vfs_clean_pages(struct buf *bp) 3210 { 3211 int i; 3212 3213 if (bp->b_flags & B_VMIO) { 3214 vm_ooffset_t foff; 3215 3216 foff = bp->b_loffset; 3217 KASSERT(foff != NOOFFSET, ("vfs_clean_pages: no buffer offset")); 3218 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3219 vm_page_t m = bp->b_xio.xio_pages[i]; 3220 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3221 vm_ooffset_t eoff = noff; 3222 3223 if (eoff > bp->b_loffset + bp->b_bufsize) 3224 eoff = bp->b_loffset + bp->b_bufsize; 3225 vfs_page_set_valid(bp, foff, i, m); 3226 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3227 foff = noff; 3228 } 3229 } 3230 } 3231 3232 /* 3233 * vfs_bio_set_validclean: 3234 * 3235 * Set the range within the buffer to valid and clean. The range is 3236 * relative to the beginning of the buffer, b_loffset. Note that 3237 * b_loffset itself may be offset from the beginning of the first page. 3238 */ 3239 3240 void 3241 vfs_bio_set_validclean(struct buf *bp, int base, int size) 3242 { 3243 if (bp->b_flags & B_VMIO) { 3244 int i; 3245 int n; 3246 3247 /* 3248 * Fixup base to be relative to beginning of first page. 3249 * Set initial n to be the maximum number of bytes in the 3250 * first page that can be validated. 3251 */ 3252 3253 base += (bp->b_loffset & PAGE_MASK); 3254 n = PAGE_SIZE - (base & PAGE_MASK); 3255 3256 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_xio.xio_npages; ++i) { 3257 vm_page_t m = bp->b_xio.xio_pages[i]; 3258 3259 if (n > size) 3260 n = size; 3261 3262 vm_page_set_validclean(m, base & PAGE_MASK, n); 3263 base += n; 3264 size -= n; 3265 n = PAGE_SIZE; 3266 } 3267 } 3268 } 3269 3270 /* 3271 * vfs_bio_clrbuf: 3272 * 3273 * Clear a buffer. This routine essentially fakes an I/O, so we need 3274 * to clear B_ERROR and B_INVAL. 3275 * 3276 * Note that while we only theoretically need to clear through b_bcount, 3277 * we go ahead and clear through b_bufsize. 3278 */ 3279 3280 void 3281 vfs_bio_clrbuf(struct buf *bp) 3282 { 3283 int i, mask = 0; 3284 caddr_t sa, ea; 3285 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3286 bp->b_flags &= ~(B_INVAL|B_ERROR); 3287 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3288 (bp->b_loffset & PAGE_MASK) == 0) { 3289 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3290 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) { 3291 bp->b_resid = 0; 3292 return; 3293 } 3294 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) && 3295 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) { 3296 bzero(bp->b_data, bp->b_bufsize); 3297 bp->b_xio.xio_pages[0]->valid |= mask; 3298 bp->b_resid = 0; 3299 return; 3300 } 3301 } 3302 ea = sa = bp->b_data; 3303 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) { 3304 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3305 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3306 ea = (caddr_t)(vm_offset_t)ulmin( 3307 (u_long)(vm_offset_t)ea, 3308 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3309 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3310 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask) 3311 continue; 3312 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) { 3313 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) { 3314 bzero(sa, ea - sa); 3315 } 3316 } else { 3317 for (; sa < ea; sa += DEV_BSIZE, j++) { 3318 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) && 3319 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0) 3320 bzero(sa, DEV_BSIZE); 3321 } 3322 } 3323 bp->b_xio.xio_pages[i]->valid |= mask; 3324 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO); 3325 } 3326 bp->b_resid = 0; 3327 } else { 3328 clrbuf(bp); 3329 } 3330 } 3331 3332 /* 3333 * vm_hold_load_pages: 3334 * 3335 * Load pages into the buffer's address space. The pages are 3336 * allocated from the kernel object in order to reduce interference 3337 * with the any VM paging I/O activity. The range of loaded 3338 * pages will be wired. 3339 * 3340 * If a page cannot be allocated, the 'pagedaemon' is woken up to 3341 * retrieve the full range (to - from) of pages. 3342 * 3343 */ 3344 void 3345 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3346 { 3347 vm_offset_t pg; 3348 vm_page_t p; 3349 int index; 3350 3351 to = round_page(to); 3352 from = round_page(from); 3353 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3354 3355 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3356 3357 tryagain: 3358 3359 /* 3360 * Note: must allocate system pages since blocking here 3361 * could intefere with paging I/O, no matter which 3362 * process we are. 3363 */ 3364 p = vm_page_alloc(kernel_object, 3365 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3366 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM); 3367 if (!p) { 3368 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3369 vm_wait(); 3370 goto tryagain; 3371 } 3372 vm_page_wire(p); 3373 p->valid = VM_PAGE_BITS_ALL; 3374 vm_page_flag_clear(p, PG_ZERO); 3375 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 3376 bp->b_xio.xio_pages[index] = p; 3377 vm_page_wakeup(p); 3378 } 3379 bp->b_xio.xio_npages = index; 3380 } 3381 3382 /* 3383 * vm_hold_free_pages: 3384 * 3385 * Return pages associated with the buffer back to the VM system. 3386 * 3387 * The range of pages underlying the buffer's address space will 3388 * be unmapped and un-wired. 3389 */ 3390 void 3391 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3392 { 3393 vm_offset_t pg; 3394 vm_page_t p; 3395 int index, newnpages; 3396 3397 from = round_page(from); 3398 to = round_page(to); 3399 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3400 3401 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3402 p = bp->b_xio.xio_pages[index]; 3403 if (p && (index < bp->b_xio.xio_npages)) { 3404 if (p->busy) { 3405 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 3406 bp->b_bio2.bio_blkno, bp->b_lblkno); 3407 } 3408 bp->b_xio.xio_pages[index] = NULL; 3409 pmap_kremove(pg); 3410 vm_page_busy(p); 3411 vm_page_unwire(p, 0); 3412 vm_page_free(p); 3413 } 3414 } 3415 bp->b_xio.xio_npages = newnpages; 3416 } 3417 3418 /* 3419 * vmapbuf: 3420 * 3421 * Map an IO request into kernel virtual address space. 3422 * 3423 * All requests are (re)mapped into kernel VA space. 3424 * Notice that we use b_bufsize for the size of the buffer 3425 * to be mapped. b_bcount might be modified by the driver. 3426 */ 3427 int 3428 vmapbuf(struct buf *bp) 3429 { 3430 caddr_t addr, v, kva; 3431 vm_paddr_t pa; 3432 int pidx; 3433 int i; 3434 struct vm_page *m; 3435 3436 if ((bp->b_flags & B_PHYS) == 0) 3437 panic("vmapbuf"); 3438 if (bp->b_bufsize < 0) 3439 return (-1); 3440 for (v = bp->b_saveaddr, 3441 addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), 3442 pidx = 0; 3443 addr < bp->b_data + bp->b_bufsize; 3444 addr += PAGE_SIZE, v += PAGE_SIZE, pidx++) { 3445 /* 3446 * Do the vm_fault if needed; do the copy-on-write thing 3447 * when reading stuff off device into memory. 3448 */ 3449 retry: 3450 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data, 3451 (bp->b_flags&B_READ)?(VM_PROT_READ|VM_PROT_WRITE):VM_PROT_READ); 3452 if (i < 0) { 3453 for (i = 0; i < pidx; ++i) { 3454 vm_page_unhold(bp->b_xio.xio_pages[i]); 3455 bp->b_xio.xio_pages[i] = NULL; 3456 } 3457 return(-1); 3458 } 3459 3460 /* 3461 * WARNING! If sparc support is MFCd in the future this will 3462 * have to be changed from pmap_kextract() to pmap_extract() 3463 * ala -current. 3464 */ 3465 #ifdef __sparc64__ 3466 #error "If MFCing sparc support use pmap_extract" 3467 #endif 3468 pa = pmap_kextract((vm_offset_t)addr); 3469 if (pa == 0) { 3470 printf("vmapbuf: warning, race against user address during I/O"); 3471 goto retry; 3472 } 3473 m = PHYS_TO_VM_PAGE(pa); 3474 vm_page_hold(m); 3475 bp->b_xio.xio_pages[pidx] = m; 3476 } 3477 if (pidx > btoc(MAXPHYS)) 3478 panic("vmapbuf: mapped more than MAXPHYS"); 3479 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_xio.xio_pages, pidx); 3480 3481 kva = bp->b_saveaddr; 3482 bp->b_xio.xio_npages = pidx; 3483 bp->b_saveaddr = bp->b_data; 3484 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3485 return(0); 3486 } 3487 3488 /* 3489 * vunmapbuf: 3490 * 3491 * Free the io map PTEs associated with this IO operation. 3492 * We also invalidate the TLB entries and restore the original b_addr. 3493 */ 3494 void 3495 vunmapbuf(struct buf *bp) 3496 { 3497 int pidx; 3498 int npages; 3499 vm_page_t *m; 3500 3501 if ((bp->b_flags & B_PHYS) == 0) 3502 panic("vunmapbuf"); 3503 3504 npages = bp->b_xio.xio_npages; 3505 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3506 npages); 3507 m = bp->b_xio.xio_pages; 3508 for (pidx = 0; pidx < npages; pidx++) 3509 vm_page_unhold(*m++); 3510 3511 bp->b_data = bp->b_saveaddr; 3512 } 3513 3514 /* 3515 * print out statistics from the current status of the buffer pool 3516 * this can be toggeled by the system control option debug.syncprt 3517 */ 3518 #ifdef DEBUG 3519 void 3520 vfs_bufstats(void) 3521 { 3522 int i, j, count; 3523 struct buf *bp; 3524 struct bqueues *dp; 3525 int counts[(MAXBSIZE / PAGE_SIZE) + 1]; 3526 static char *bname[3] = { "LOCKED", "LRU", "AGE" }; 3527 3528 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) { 3529 count = 0; 3530 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++) 3531 counts[j] = 0; 3532 crit_enter(); 3533 TAILQ_FOREACH(bp, dp, b_freelist) { 3534 counts[bp->b_bufsize/PAGE_SIZE]++; 3535 count++; 3536 } 3537 crit_exit(); 3538 printf("%s: total-%d", bname[i], count); 3539 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++) 3540 if (counts[j] != 0) 3541 printf(", %d-%d", j * PAGE_SIZE, counts[j]); 3542 printf("\n"); 3543 } 3544 } 3545 #endif 3546 3547 #include "opt_ddb.h" 3548 #ifdef DDB 3549 #include <ddb/ddb.h> 3550 3551 DB_SHOW_COMMAND(buffer, db_show_buffer) 3552 { 3553 /* get args */ 3554 struct buf *bp = (struct buf *)addr; 3555 3556 if (!have_addr) { 3557 db_printf("usage: show buffer <addr>\n"); 3558 return; 3559 } 3560 3561 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3562 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 3563 "b_resid = %ld\n, b_data = %p, " 3564 "bio_blkno(disk) = %d, bio_blkno(phys) = %d\n", 3565 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3566 bp->b_data, bp->b_bio2.bio_blkno, (bp->b_bio2.bio_next ? bp->b_bio2.bio_next->bio_blkno : (daddr_t)-1)); 3567 if (bp->b_xio.xio_npages) { 3568 int i; 3569 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ", 3570 bp->b_xio.xio_npages); 3571 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3572 vm_page_t m; 3573 m = bp->b_xio.xio_pages[i]; 3574 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3575 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3576 if ((i + 1) < bp->b_xio.xio_npages) 3577 db_printf(","); 3578 } 3579 db_printf("\n"); 3580 } 3581 } 3582 #endif /* DDB */ 3583