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 1.77 2006/05/27 20:17:16 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 <sys/spinlock2.h> 60 #include <vm/vm_page2.h> 61 62 #include "opt_ddb.h" 63 #ifdef DDB 64 #include <ddb/ddb.h> 65 #endif 66 67 /* 68 * Buffer queues. 69 */ 70 #define BUFFER_QUEUES 6 71 enum bufq_type { 72 BQUEUE_NONE, /* not on any queue */ 73 BQUEUE_LOCKED, /* locked buffers */ 74 BQUEUE_CLEAN, /* non-B_DELWRI buffers */ 75 BQUEUE_DIRTY, /* B_DELWRI buffers */ 76 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */ 77 BQUEUE_EMPTY /* empty buffer headers */ 78 }; 79 TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES]; 80 81 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 82 83 struct bio_ops bioops; /* I/O operation notification */ 84 85 struct buf *buf; /* buffer header pool */ 86 87 static void vm_hold_free_pages(struct buf *bp, vm_offset_t from, 88 vm_offset_t to); 89 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 90 vm_offset_t to); 91 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 92 int pageno, vm_page_t m); 93 static void vfs_clean_pages(struct buf *bp); 94 static void vfs_setdirty(struct buf *bp); 95 static void vfs_vmio_release(struct buf *bp); 96 static int flushbufqueues(void); 97 98 static void buf_daemon (void); 99 /* 100 * bogus page -- for I/O to/from partially complete buffers 101 * this is a temporary solution to the problem, but it is not 102 * really that bad. it would be better to split the buffer 103 * for input in the case of buffers partially already in memory, 104 * but the code is intricate enough already. 105 */ 106 vm_page_t bogus_page; 107 int runningbufspace; 108 109 static int bufspace, maxbufspace, 110 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace; 111 static int bufreusecnt, bufdefragcnt, buffreekvacnt; 112 static int lorunningspace, hirunningspace, runningbufreq; 113 static int numdirtybuffers, lodirtybuffers, hidirtybuffers; 114 static int numfreebuffers, lofreebuffers, hifreebuffers; 115 static int getnewbufcalls; 116 static int getnewbufrestarts; 117 118 static int needsbuffer; /* locked by needsbuffer_spin */ 119 static int bd_request; /* locked by needsbuffer_spin */ 120 static struct spinlock needsbuffer_spin; 121 122 /* 123 * Sysctls for operational control of the buffer cache. 124 */ 125 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 126 "Number of dirty buffers to flush before bufdaemon becomes inactive"); 127 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 128 "High watermark used to trigger explicit flushing of dirty buffers"); 129 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 130 "Low watermark for special reserve in low-memory situations"); 131 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 132 "High watermark for special reserve in low-memory situations"); 133 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 134 "Minimum amount of buffer space required for active I/O"); 135 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 136 "Maximum amount of buffer space to usable for active I/O"); 137 /* 138 * Sysctls determining current state of the buffer cache. 139 */ 140 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 141 "Pending number of dirty buffers"); 142 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 143 "Number of free buffers on the buffer cache free list"); 144 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 145 "I/O bytes currently in progress due to asynchronous writes"); 146 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 147 "Hard limit on maximum amount of memory usable for buffer space"); 148 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 149 "Soft limit on maximum amount of memory usable for buffer space"); 150 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 151 "Minimum amount of memory to reserve for system buffer space"); 152 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 153 "Amount of memory available for buffers"); 154 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace, 155 0, "Maximum amount of memory reserved for buffers using malloc"); 156 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 157 "Amount of memory left for buffers using malloc-scheme"); 158 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0, 159 "New buffer header acquisition requests"); 160 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts, 161 0, "New buffer header acquisition restarts"); 162 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0, 163 "Buffer acquisition restarts due to fragmented buffer map"); 164 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0, 165 "Amount of time KVA space was deallocated in an arbitrary buffer"); 166 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0, 167 "Amount of time buffer re-use operations were successful"); 168 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf), 169 "sizeof(struct buf)"); 170 171 char *buf_wmesg = BUF_WMESG; 172 173 extern int vm_swap_size; 174 175 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 176 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 177 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 178 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 179 180 /* 181 * numdirtywakeup: 182 * 183 * If someone is blocked due to there being too many dirty buffers, 184 * and numdirtybuffers is now reasonable, wake them up. 185 */ 186 187 static __inline void 188 numdirtywakeup(int level) 189 { 190 if (numdirtybuffers <= level) { 191 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 192 spin_lock_wr(&needsbuffer_spin); 193 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 194 spin_unlock_wr(&needsbuffer_spin); 195 wakeup(&needsbuffer); 196 } 197 } 198 } 199 200 /* 201 * bufspacewakeup: 202 * 203 * Called when buffer space is potentially available for recovery. 204 * getnewbuf() will block on this flag when it is unable to free 205 * sufficient buffer space. Buffer space becomes recoverable when 206 * bp's get placed back in the queues. 207 */ 208 209 static __inline void 210 bufspacewakeup(void) 211 { 212 /* 213 * If someone is waiting for BUF space, wake them up. Even 214 * though we haven't freed the kva space yet, the waiting 215 * process will be able to now. 216 */ 217 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 218 spin_lock_wr(&needsbuffer_spin); 219 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 220 spin_unlock_wr(&needsbuffer_spin); 221 wakeup(&needsbuffer); 222 } 223 } 224 225 /* 226 * runningbufwakeup: 227 * 228 * Accounting for I/O in progress. 229 * 230 */ 231 static __inline void 232 runningbufwakeup(struct buf *bp) 233 { 234 if (bp->b_runningbufspace) { 235 runningbufspace -= bp->b_runningbufspace; 236 bp->b_runningbufspace = 0; 237 if (runningbufreq && runningbufspace <= lorunningspace) { 238 runningbufreq = 0; 239 wakeup(&runningbufreq); 240 } 241 } 242 } 243 244 /* 245 * bufcountwakeup: 246 * 247 * Called when a buffer has been added to one of the free queues to 248 * account for the buffer and to wakeup anyone waiting for free buffers. 249 * This typically occurs when large amounts of metadata are being handled 250 * by the buffer cache ( else buffer space runs out first, usually ). 251 */ 252 253 static __inline void 254 bufcountwakeup(void) 255 { 256 ++numfreebuffers; 257 if (needsbuffer) { 258 spin_lock_wr(&needsbuffer_spin); 259 needsbuffer &= ~VFS_BIO_NEED_ANY; 260 if (numfreebuffers >= hifreebuffers) 261 needsbuffer &= ~VFS_BIO_NEED_FREE; 262 spin_unlock_wr(&needsbuffer_spin); 263 wakeup(&needsbuffer); 264 } 265 } 266 267 /* 268 * waitrunningbufspace() 269 * 270 * runningbufspace is a measure of the amount of I/O currently 271 * running. This routine is used in async-write situations to 272 * prevent creating huge backups of pending writes to a device. 273 * Only asynchronous writes are governed by this function. 274 * 275 * Reads will adjust runningbufspace, but will not block based on it. 276 * The read load has a side effect of reducing the allowed write load. 277 * 278 * This does NOT turn an async write into a sync write. It waits 279 * for earlier writes to complete and generally returns before the 280 * caller's write has reached the device. 281 */ 282 static __inline void 283 waitrunningbufspace(void) 284 { 285 if (runningbufspace > hirunningspace) { 286 crit_enter(); 287 while (runningbufspace > hirunningspace) { 288 ++runningbufreq; 289 tsleep(&runningbufreq, 0, "wdrain", 0); 290 } 291 crit_exit(); 292 } 293 } 294 295 /* 296 * vfs_buf_test_cache: 297 * 298 * Called when a buffer is extended. This function clears the B_CACHE 299 * bit if the newly extended portion of the buffer does not contain 300 * valid data. 301 */ 302 static __inline__ 303 void 304 vfs_buf_test_cache(struct buf *bp, 305 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 306 vm_page_t m) 307 { 308 if (bp->b_flags & B_CACHE) { 309 int base = (foff + off) & PAGE_MASK; 310 if (vm_page_is_valid(m, base, size) == 0) 311 bp->b_flags &= ~B_CACHE; 312 } 313 } 314 315 /* 316 * bd_wakeup: 317 * 318 * Wake up the buffer daemon if the number of outstanding dirty buffers 319 * is above specified threshold 'dirtybuflevel'. 320 * 321 * The buffer daemon is explicitly woken up when (a) the pending number 322 * of dirty buffers exceeds the recovery and stall mid-point value, 323 * (b) during bwillwrite() or (c) buf freelist was exhausted. 324 */ 325 static __inline__ 326 void 327 bd_wakeup(int dirtybuflevel) 328 { 329 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 330 spin_lock_wr(&needsbuffer_spin); 331 bd_request = 1; 332 spin_unlock_wr(&needsbuffer_spin); 333 wakeup(&bd_request); 334 } 335 } 336 337 /* 338 * bd_speedup: 339 * 340 * Speed up the buffer cache flushing process. 341 */ 342 343 static __inline__ 344 void 345 bd_speedup(void) 346 { 347 bd_wakeup(1); 348 } 349 350 /* 351 * bufinit: 352 * 353 * Load time initialisation of the buffer cache, called from machine 354 * dependant initialization code. 355 */ 356 void 357 bufinit(void) 358 { 359 struct buf *bp; 360 vm_offset_t bogus_offset; 361 int i; 362 363 spin_init(&needsbuffer_spin); 364 365 /* next, make a null set of free lists */ 366 for (i = 0; i < BUFFER_QUEUES; i++) 367 TAILQ_INIT(&bufqueues[i]); 368 369 /* finally, initialize each buffer header and stick on empty q */ 370 for (i = 0; i < nbuf; i++) { 371 bp = &buf[i]; 372 bzero(bp, sizeof *bp); 373 bp->b_flags = B_INVAL; /* we're just an empty header */ 374 bp->b_cmd = BUF_CMD_DONE; 375 bp->b_qindex = BQUEUE_EMPTY; 376 initbufbio(bp); 377 xio_init(&bp->b_xio); 378 LIST_INIT(&bp->b_dep); 379 BUF_LOCKINIT(bp); 380 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist); 381 } 382 383 /* 384 * maxbufspace is the absolute maximum amount of buffer space we are 385 * allowed to reserve in KVM and in real terms. The absolute maximum 386 * is nominally used by buf_daemon. hibufspace is the nominal maximum 387 * used by most other processes. The differential is required to 388 * ensure that buf_daemon is able to run when other processes might 389 * be blocked waiting for buffer space. 390 * 391 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 392 * this may result in KVM fragmentation which is not handled optimally 393 * by the system. 394 */ 395 maxbufspace = nbuf * BKVASIZE; 396 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 397 lobufspace = hibufspace - MAXBSIZE; 398 399 lorunningspace = 512 * 1024; 400 hirunningspace = 1024 * 1024; 401 402 /* 403 * Limit the amount of malloc memory since it is wired permanently into 404 * the kernel space. Even though this is accounted for in the buffer 405 * allocation, we don't want the malloced region to grow uncontrolled. 406 * The malloc scheme improves memory utilization significantly on average 407 * (small) directories. 408 */ 409 maxbufmallocspace = hibufspace / 20; 410 411 /* 412 * Reduce the chance of a deadlock occuring by limiting the number 413 * of delayed-write dirty buffers we allow to stack up. 414 */ 415 hidirtybuffers = nbuf / 4 + 20; 416 numdirtybuffers = 0; 417 /* 418 * To support extreme low-memory systems, make sure hidirtybuffers cannot 419 * eat up all available buffer space. This occurs when our minimum cannot 420 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 421 * BKVASIZE'd (8K) buffers. 422 */ 423 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 424 hidirtybuffers >>= 1; 425 } 426 lodirtybuffers = hidirtybuffers / 2; 427 428 /* 429 * Try to keep the number of free buffers in the specified range, 430 * and give special processes (e.g. like buf_daemon) access to an 431 * emergency reserve. 432 */ 433 lofreebuffers = nbuf / 18 + 5; 434 hifreebuffers = 2 * lofreebuffers; 435 numfreebuffers = nbuf; 436 437 /* 438 * Maximum number of async ops initiated per buf_daemon loop. This is 439 * somewhat of a hack at the moment, we really need to limit ourselves 440 * based on the number of bytes of I/O in-transit that were initiated 441 * from buf_daemon. 442 */ 443 444 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 445 bogus_page = vm_page_alloc(kernel_object, 446 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 447 VM_ALLOC_NORMAL); 448 vmstats.v_wire_count++; 449 450 } 451 452 /* 453 * Initialize the embedded bio structures 454 */ 455 void 456 initbufbio(struct buf *bp) 457 { 458 bp->b_bio1.bio_buf = bp; 459 bp->b_bio1.bio_prev = NULL; 460 bp->b_bio1.bio_offset = NOOFFSET; 461 bp->b_bio1.bio_next = &bp->b_bio2; 462 bp->b_bio1.bio_done = NULL; 463 464 bp->b_bio2.bio_buf = bp; 465 bp->b_bio2.bio_prev = &bp->b_bio1; 466 bp->b_bio2.bio_offset = NOOFFSET; 467 bp->b_bio2.bio_next = NULL; 468 bp->b_bio2.bio_done = NULL; 469 } 470 471 /* 472 * Reinitialize the embedded bio structures as well as any additional 473 * translation cache layers. 474 */ 475 void 476 reinitbufbio(struct buf *bp) 477 { 478 struct bio *bio; 479 480 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) { 481 bio->bio_done = NULL; 482 bio->bio_offset = NOOFFSET; 483 } 484 } 485 486 /* 487 * Push another BIO layer onto an existing BIO and return it. The new 488 * BIO layer may already exist, holding cached translation data. 489 */ 490 struct bio * 491 push_bio(struct bio *bio) 492 { 493 struct bio *nbio; 494 495 if ((nbio = bio->bio_next) == NULL) { 496 int index = bio - &bio->bio_buf->b_bio_array[0]; 497 if (index >= NBUF_BIO) { 498 panic("push_bio: too many layers bp %p\n", 499 bio->bio_buf); 500 } 501 nbio = &bio->bio_buf->b_bio_array[index + 1]; 502 bio->bio_next = nbio; 503 nbio->bio_prev = bio; 504 nbio->bio_buf = bio->bio_buf; 505 nbio->bio_offset = NOOFFSET; 506 nbio->bio_done = NULL; 507 nbio->bio_next = NULL; 508 } 509 KKASSERT(nbio->bio_done == NULL); 510 return(nbio); 511 } 512 513 void 514 pop_bio(struct bio *bio) 515 { 516 /* NOP */ 517 } 518 519 void 520 clearbiocache(struct bio *bio) 521 { 522 while (bio) { 523 bio->bio_offset = NOOFFSET; 524 bio = bio->bio_next; 525 } 526 } 527 528 /* 529 * bfreekva: 530 * 531 * Free the KVA allocation for buffer 'bp'. 532 * 533 * Must be called from a critical section as this is the only locking for 534 * buffer_map. 535 * 536 * Since this call frees up buffer space, we call bufspacewakeup(). 537 */ 538 static void 539 bfreekva(struct buf *bp) 540 { 541 int count; 542 543 if (bp->b_kvasize) { 544 ++buffreekvacnt; 545 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 546 vm_map_lock(buffer_map); 547 bufspace -= bp->b_kvasize; 548 vm_map_delete(buffer_map, 549 (vm_offset_t) bp->b_kvabase, 550 (vm_offset_t) bp->b_kvabase + bp->b_kvasize, 551 &count 552 ); 553 vm_map_unlock(buffer_map); 554 vm_map_entry_release(count); 555 bp->b_kvasize = 0; 556 bufspacewakeup(); 557 } 558 } 559 560 /* 561 * bremfree: 562 * 563 * Remove the buffer from the appropriate free list. 564 */ 565 void 566 bremfree(struct buf *bp) 567 { 568 int old_qindex; 569 570 crit_enter(); 571 old_qindex = bp->b_qindex; 572 573 if (bp->b_qindex != BQUEUE_NONE) { 574 KASSERT(BUF_REFCNTNB(bp) == 1, 575 ("bremfree: bp %p not locked",bp)); 576 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 577 bp->b_qindex = BQUEUE_NONE; 578 } else { 579 if (BUF_REFCNTNB(bp) <= 1) 580 panic("bremfree: removing a buffer not on a queue"); 581 } 582 583 /* 584 * Fixup numfreebuffers count. If the buffer is invalid or not 585 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 586 * the buffer was free and we must decrement numfreebuffers. 587 */ 588 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 589 switch(old_qindex) { 590 case BQUEUE_DIRTY: 591 case BQUEUE_CLEAN: 592 case BQUEUE_EMPTY: 593 case BQUEUE_EMPTYKVA: 594 --numfreebuffers; 595 break; 596 default: 597 break; 598 } 599 } 600 crit_exit(); 601 } 602 603 604 /* 605 * bread: 606 * 607 * Get a buffer with the specified data. Look in the cache first. We 608 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 609 * is set, the buffer is valid and we do not have to do anything ( see 610 * getblk() ). 611 */ 612 int 613 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp) 614 { 615 struct buf *bp; 616 617 bp = getblk(vp, loffset, size, 0, 0); 618 *bpp = bp; 619 620 /* if not found in cache, do some I/O */ 621 if ((bp->b_flags & B_CACHE) == 0) { 622 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp)); 623 bp->b_flags &= ~(B_ERROR | B_INVAL); 624 bp->b_cmd = BUF_CMD_READ; 625 vfs_busy_pages(vp, bp); 626 vn_strategy(vp, &bp->b_bio1); 627 return (biowait(bp)); 628 } 629 return (0); 630 } 631 632 /* 633 * breadn: 634 * 635 * Operates like bread, but also starts asynchronous I/O on 636 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior 637 * to initiating I/O . If B_CACHE is set, the buffer is valid 638 * and we do not have to do anything. 639 */ 640 int 641 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset, 642 int *rabsize, int cnt, struct buf **bpp) 643 { 644 struct buf *bp, *rabp; 645 int i; 646 int rv = 0, readwait = 0; 647 648 *bpp = bp = getblk(vp, loffset, size, 0, 0); 649 650 /* if not found in cache, do some I/O */ 651 if ((bp->b_flags & B_CACHE) == 0) { 652 bp->b_flags &= ~(B_ERROR | B_INVAL); 653 bp->b_cmd = BUF_CMD_READ; 654 vfs_busy_pages(vp, bp); 655 vn_strategy(vp, &bp->b_bio1); 656 ++readwait; 657 } 658 659 for (i = 0; i < cnt; i++, raoffset++, rabsize++) { 660 if (inmem(vp, *raoffset)) 661 continue; 662 rabp = getblk(vp, *raoffset, *rabsize, 0, 0); 663 664 if ((rabp->b_flags & B_CACHE) == 0) { 665 rabp->b_flags |= B_ASYNC; 666 rabp->b_flags &= ~(B_ERROR | B_INVAL); 667 rabp->b_cmd = BUF_CMD_READ; 668 vfs_busy_pages(vp, rabp); 669 BUF_KERNPROC(rabp); 670 vn_strategy(vp, &rabp->b_bio1); 671 } else { 672 brelse(rabp); 673 } 674 } 675 676 if (readwait) { 677 rv = biowait(bp); 678 } 679 return (rv); 680 } 681 682 /* 683 * bwrite: 684 * 685 * Write, release buffer on completion. (Done by iodone 686 * if async). Do not bother writing anything if the buffer 687 * is invalid. 688 * 689 * Note that we set B_CACHE here, indicating that buffer is 690 * fully valid and thus cacheable. This is true even of NFS 691 * now so we set it generally. This could be set either here 692 * or in biodone() since the I/O is synchronous. We put it 693 * here. 694 */ 695 int 696 bwrite(struct buf *bp) 697 { 698 int oldflags; 699 700 if (bp->b_flags & B_INVAL) { 701 brelse(bp); 702 return (0); 703 } 704 705 oldflags = bp->b_flags; 706 707 if (BUF_REFCNTNB(bp) == 0) 708 panic("bwrite: buffer is not busy???"); 709 crit_enter(); 710 711 /* Mark the buffer clean */ 712 bundirty(bp); 713 714 bp->b_flags &= ~B_ERROR; 715 bp->b_flags |= B_CACHE; 716 bp->b_cmd = BUF_CMD_WRITE; 717 vfs_busy_pages(bp->b_vp, bp); 718 719 /* 720 * Normal bwrites pipeline writes. NOTE: b_bufsize is only 721 * valid for vnode-backed buffers. 722 */ 723 bp->b_runningbufspace = bp->b_bufsize; 724 runningbufspace += bp->b_runningbufspace; 725 726 crit_exit(); 727 if (oldflags & B_ASYNC) 728 BUF_KERNPROC(bp); 729 vn_strategy(bp->b_vp, &bp->b_bio1); 730 731 if ((oldflags & B_ASYNC) == 0) { 732 int rtval = biowait(bp); 733 brelse(bp); 734 return (rtval); 735 } else if ((oldflags & B_NOWDRAIN) == 0) { 736 /* 737 * don't allow the async write to saturate the I/O 738 * system. Deadlocks can occur only if a device strategy 739 * routine (like in VN) turns around and issues another 740 * high-level write, in which case B_NOWDRAIN is expected 741 * to be set. Otherwise we will not deadlock here because 742 * we are blocking waiting for I/O that is already in-progress 743 * to complete. 744 */ 745 waitrunningbufspace(); 746 } 747 748 return (0); 749 } 750 751 /* 752 * bdwrite: 753 * 754 * Delayed write. (Buffer is marked dirty). Do not bother writing 755 * anything if the buffer is marked invalid. 756 * 757 * Note that since the buffer must be completely valid, we can safely 758 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 759 * biodone() in order to prevent getblk from writing the buffer 760 * out synchronously. 761 */ 762 void 763 bdwrite(struct buf *bp) 764 { 765 if (BUF_REFCNTNB(bp) == 0) 766 panic("bdwrite: buffer is not busy"); 767 768 if (bp->b_flags & B_INVAL) { 769 brelse(bp); 770 return; 771 } 772 bdirty(bp); 773 774 /* 775 * Set B_CACHE, indicating that the buffer is fully valid. This is 776 * true even of NFS now. 777 */ 778 bp->b_flags |= B_CACHE; 779 780 /* 781 * This bmap keeps the system from needing to do the bmap later, 782 * perhaps when the system is attempting to do a sync. Since it 783 * is likely that the indirect block -- or whatever other datastructure 784 * that the filesystem needs is still in memory now, it is a good 785 * thing to do this. Note also, that if the pageout daemon is 786 * requesting a sync -- there might not be enough memory to do 787 * the bmap then... So, this is important to do. 788 */ 789 if (bp->b_bio2.bio_offset == NOOFFSET) { 790 VOP_BMAP(bp->b_vp, bp->b_loffset, NULL, &bp->b_bio2.bio_offset, 791 NULL, NULL); 792 } 793 794 /* 795 * Set the *dirty* buffer range based upon the VM system dirty pages. 796 */ 797 vfs_setdirty(bp); 798 799 /* 800 * We need to do this here to satisfy the vnode_pager and the 801 * pageout daemon, so that it thinks that the pages have been 802 * "cleaned". Note that since the pages are in a delayed write 803 * buffer -- the VFS layer "will" see that the pages get written 804 * out on the next sync, or perhaps the cluster will be completed. 805 */ 806 vfs_clean_pages(bp); 807 bqrelse(bp); 808 809 /* 810 * Wakeup the buffer flushing daemon if we have a lot of dirty 811 * buffers (midpoint between our recovery point and our stall 812 * point). 813 */ 814 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 815 816 /* 817 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 818 * due to the softdep code. 819 */ 820 } 821 822 /* 823 * bdirty: 824 * 825 * Turn buffer into delayed write request by marking it B_DELWRI. 826 * B_RELBUF and B_NOCACHE must be cleared. 827 * 828 * We reassign the buffer to itself to properly update it in the 829 * dirty/clean lists. 830 * 831 * Since the buffer is not on a queue, we do not update the 832 * numfreebuffers count. 833 * 834 * Must be called from a critical section. 835 * The buffer must be on BQUEUE_NONE. 836 */ 837 void 838 bdirty(struct buf *bp) 839 { 840 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 841 if (bp->b_flags & B_NOCACHE) { 842 printf("bdirty: clearing B_NOCACHE on buf %p\n", bp); 843 bp->b_flags &= ~B_NOCACHE; 844 } 845 if (bp->b_flags & B_INVAL) { 846 printf("bdirty: warning, dirtying invalid buffer %p\n", bp); 847 } 848 bp->b_flags &= ~B_RELBUF; 849 850 if ((bp->b_flags & B_DELWRI) == 0) { 851 bp->b_flags |= B_DELWRI; 852 reassignbuf(bp); 853 ++numdirtybuffers; 854 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 855 } 856 } 857 858 /* 859 * bundirty: 860 * 861 * Clear B_DELWRI for buffer. 862 * 863 * Since the buffer is not on a queue, we do not update the numfreebuffers 864 * count. 865 * 866 * Must be called from a critical section. 867 * 868 * The buffer is typically on BQUEUE_NONE but there is one case in 869 * brelse() that calls this function after placing the buffer on 870 * a different queue. 871 */ 872 873 void 874 bundirty(struct buf *bp) 875 { 876 if (bp->b_flags & B_DELWRI) { 877 bp->b_flags &= ~B_DELWRI; 878 reassignbuf(bp); 879 --numdirtybuffers; 880 numdirtywakeup(lodirtybuffers); 881 } 882 /* 883 * Since it is now being written, we can clear its deferred write flag. 884 */ 885 bp->b_flags &= ~B_DEFERRED; 886 } 887 888 /* 889 * bawrite: 890 * 891 * Asynchronous write. Start output on a buffer, but do not wait for 892 * it to complete. The buffer is released when the output completes. 893 * 894 * bwrite() ( or the VOP routine anyway ) is responsible for handling 895 * B_INVAL buffers. Not us. 896 */ 897 void 898 bawrite(struct buf *bp) 899 { 900 bp->b_flags |= B_ASYNC; 901 bwrite(bp); 902 } 903 904 /* 905 * bowrite: 906 * 907 * Ordered write. Start output on a buffer, and flag it so that the 908 * device will write it in the order it was queued. The buffer is 909 * released when the output completes. bwrite() ( or the VOP routine 910 * anyway ) is responsible for handling B_INVAL buffers. 911 */ 912 int 913 bowrite(struct buf *bp) 914 { 915 bp->b_flags |= B_ORDERED | B_ASYNC; 916 return (bwrite(bp)); 917 } 918 919 /* 920 * bwillwrite: 921 * 922 * Called prior to the locking of any vnodes when we are expecting to 923 * write. We do not want to starve the buffer cache with too many 924 * dirty buffers so we block here. By blocking prior to the locking 925 * of any vnodes we attempt to avoid the situation where a locked vnode 926 * prevents the various system daemons from flushing related buffers. 927 */ 928 929 void 930 bwillwrite(void) 931 { 932 if (numdirtybuffers >= hidirtybuffers) { 933 while (numdirtybuffers >= hidirtybuffers) { 934 bd_wakeup(1); 935 spin_lock_wr(&needsbuffer_spin); 936 if (numdirtybuffers >= hidirtybuffers) { 937 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 938 msleep(&needsbuffer, &needsbuffer_spin, 0, 939 "flswai", 0); 940 } 941 spin_unlock_wr(&needsbuffer_spin); 942 } 943 } 944 } 945 946 /* 947 * buf_dirty_count_severe: 948 * 949 * Return true if we have too many dirty buffers. 950 */ 951 int 952 buf_dirty_count_severe(void) 953 { 954 return(numdirtybuffers >= hidirtybuffers); 955 } 956 957 /* 958 * brelse: 959 * 960 * Release a busy buffer and, if requested, free its resources. The 961 * buffer will be stashed in the appropriate bufqueue[] allowing it 962 * to be accessed later as a cache entity or reused for other purposes. 963 */ 964 void 965 brelse(struct buf *bp) 966 { 967 #ifdef INVARIANTS 968 int saved_flags = bp->b_flags; 969 #endif 970 971 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 972 973 crit_enter(); 974 975 /* 976 * If B_NOCACHE is set we are being asked to destroy the buffer and 977 * its backing store. Clear B_DELWRI. 978 * 979 * B_NOCACHE is set in two cases: (1) when the caller really wants 980 * to destroy the buffer and backing store and (2) when the caller 981 * wants to destroy the buffer and backing store after a write 982 * completes. 983 */ 984 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) { 985 bundirty(bp); 986 } 987 988 if (bp->b_flags & B_LOCKED) 989 bp->b_flags &= ~B_ERROR; 990 991 /* 992 * If a write error occurs and the caller does not want to throw 993 * away the buffer, redirty the buffer. This will also clear 994 * B_NOCACHE. 995 */ 996 if (bp->b_cmd == BUF_CMD_WRITE && 997 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) { 998 /* 999 * Failed write, redirty. Must clear B_ERROR to prevent 1000 * pages from being scrapped. If B_INVAL is set then 1001 * this case is not run and the next case is run to 1002 * destroy the buffer. B_INVAL can occur if the buffer 1003 * is outside the range supported by the underlying device. 1004 */ 1005 bp->b_flags &= ~B_ERROR; 1006 bdirty(bp); 1007 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) || 1008 (bp->b_bufsize <= 0) || bp->b_cmd == BUF_CMD_FREEBLKS) { 1009 /* 1010 * Either a failed I/O or we were asked to free or not 1011 * cache the buffer. 1012 */ 1013 bp->b_flags |= B_INVAL; 1014 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 1015 (*bioops.io_deallocate)(bp); 1016 if (bp->b_flags & B_DELWRI) { 1017 --numdirtybuffers; 1018 numdirtywakeup(lodirtybuffers); 1019 } 1020 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1021 } 1022 1023 /* 1024 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1025 * is called with B_DELWRI set, the underlying pages may wind up 1026 * getting freed causing a previous write (bdwrite()) to get 'lost' 1027 * because pages associated with a B_DELWRI bp are marked clean. 1028 * 1029 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1030 * if B_DELWRI is set. 1031 * 1032 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1033 * on pages to return pages to the VM page queues. 1034 */ 1035 if (bp->b_flags & B_DELWRI) 1036 bp->b_flags &= ~B_RELBUF; 1037 else if (vm_page_count_severe()) 1038 bp->b_flags |= B_RELBUF; 1039 1040 /* 1041 * At this point destroying the buffer is governed by the B_INVAL 1042 * or B_RELBUF flags. 1043 */ 1044 bp->b_cmd = BUF_CMD_DONE; 1045 1046 /* 1047 * VMIO buffer rundown. Make sure the VM page array is restored 1048 * after an I/O may have replaces some of the pages with bogus pages 1049 * in order to not destroy dirty pages in a fill-in read. 1050 * 1051 * Note that due to the code above, if a buffer is marked B_DELWRI 1052 * then the B_RELBUF and B_NOCACHE bits will always be clear. 1053 * B_INVAL may still be set, however. 1054 * 1055 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer 1056 * but not the backing store. B_NOCACHE will destroy the backing 1057 * store. 1058 * 1059 * Note that dirty NFS buffers contain byte-granular write ranges 1060 * and should not be destroyed w/ B_INVAL even if the backing store 1061 * is left intact. 1062 */ 1063 if (bp->b_flags & B_VMIO) { 1064 /* 1065 * Rundown for VMIO buffers which are not dirty NFS buffers. 1066 */ 1067 int i, j, resid; 1068 vm_page_t m; 1069 off_t foff; 1070 vm_pindex_t poff; 1071 vm_object_t obj; 1072 struct vnode *vp; 1073 1074 vp = bp->b_vp; 1075 1076 /* 1077 * Get the base offset and length of the buffer. Note that 1078 * in the VMIO case if the buffer block size is not 1079 * page-aligned then b_data pointer may not be page-aligned. 1080 * But our b_xio.xio_pages array *IS* page aligned. 1081 * 1082 * block sizes less then DEV_BSIZE (usually 512) are not 1083 * supported due to the page granularity bits (m->valid, 1084 * m->dirty, etc...). 1085 * 1086 * See man buf(9) for more information 1087 */ 1088 1089 resid = bp->b_bufsize; 1090 foff = bp->b_loffset; 1091 1092 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1093 m = bp->b_xio.xio_pages[i]; 1094 vm_page_flag_clear(m, PG_ZERO); 1095 /* 1096 * If we hit a bogus page, fixup *all* of them 1097 * now. Note that we left these pages wired 1098 * when we removed them so they had better exist, 1099 * and they cannot be ripped out from under us so 1100 * no critical section protection is necessary. 1101 */ 1102 if (m == bogus_page) { 1103 obj = vp->v_object; 1104 poff = OFF_TO_IDX(bp->b_loffset); 1105 1106 for (j = i; j < bp->b_xio.xio_npages; j++) { 1107 vm_page_t mtmp; 1108 1109 mtmp = bp->b_xio.xio_pages[j]; 1110 if (mtmp == bogus_page) { 1111 mtmp = vm_page_lookup(obj, poff + j); 1112 if (!mtmp) { 1113 panic("brelse: page missing"); 1114 } 1115 bp->b_xio.xio_pages[j] = mtmp; 1116 } 1117 } 1118 1119 if ((bp->b_flags & B_INVAL) == 0) { 1120 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 1121 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 1122 } 1123 m = bp->b_xio.xio_pages[i]; 1124 } 1125 1126 /* 1127 * Invalidate the backing store if B_NOCACHE is set 1128 * (e.g. used with vinvalbuf()). If this is NFS 1129 * we impose a requirement that the block size be 1130 * a multiple of PAGE_SIZE and create a temporary 1131 * hack to basically invalidate the whole page. The 1132 * problem is that NFS uses really odd buffer sizes 1133 * especially when tracking piecemeal writes and 1134 * it also vinvalbuf()'s a lot, which would result 1135 * in only partial page validation and invalidation 1136 * here. If the file page is mmap()'d, however, 1137 * all the valid bits get set so after we invalidate 1138 * here we would end up with weird m->valid values 1139 * like 0xfc. nfs_getpages() can't handle this so 1140 * we clear all the valid bits for the NFS case 1141 * instead of just some of them. 1142 * 1143 * The real bug is the VM system having to set m->valid 1144 * to VM_PAGE_BITS_ALL for faulted-in pages, which 1145 * itself is an artifact of the whole 512-byte 1146 * granular mess that exists to support odd block 1147 * sizes and UFS meta-data block sizes (e.g. 6144). 1148 * A complete rewrite is required. 1149 */ 1150 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 1151 int poffset = foff & PAGE_MASK; 1152 int presid; 1153 1154 presid = PAGE_SIZE - poffset; 1155 if (bp->b_vp->v_tag == VT_NFS && 1156 bp->b_vp->v_type == VREG) { 1157 ; /* entire page */ 1158 } else if (presid > resid) { 1159 presid = resid; 1160 } 1161 KASSERT(presid >= 0, ("brelse: extra page")); 1162 vm_page_set_invalid(m, poffset, presid); 1163 } 1164 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1165 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1166 } 1167 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1168 vfs_vmio_release(bp); 1169 } else { 1170 /* 1171 * Rundown for non-VMIO buffers. 1172 */ 1173 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1174 #if 0 1175 if (bp->b_vp) 1176 printf("brelse bp %p %08x/%08x: Warning, caught and fixed brelvp bug\n", bp, saved_flags, bp->b_flags); 1177 #endif 1178 if (bp->b_bufsize) 1179 allocbuf(bp, 0); 1180 if (bp->b_vp) 1181 brelvp(bp); 1182 } 1183 } 1184 1185 if (bp->b_qindex != BQUEUE_NONE) 1186 panic("brelse: free buffer onto another queue???"); 1187 if (BUF_REFCNTNB(bp) > 1) { 1188 /* Temporary panic to verify exclusive locking */ 1189 /* This panic goes away when we allow shared refs */ 1190 panic("brelse: multiple refs"); 1191 /* do not release to free list */ 1192 BUF_UNLOCK(bp); 1193 crit_exit(); 1194 return; 1195 } 1196 1197 /* 1198 * Figure out the correct queue to place the cleaned up buffer on. 1199 * Buffers placed in the EMPTY or EMPTYKVA had better already be 1200 * disassociated from their vnode. 1201 */ 1202 1203 if (bp->b_bufsize == 0) { 1204 /* 1205 * Buffers with no memory. Due to conditionals near the top 1206 * of brelse() such buffers should probably already be 1207 * marked B_INVAL and disassociated from their vnode. 1208 */ 1209 bp->b_flags |= B_INVAL; 1210 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp)); 1211 KKASSERT((bp->b_flags & B_HASHED) == 0); 1212 if (bp->b_kvasize) { 1213 bp->b_qindex = BQUEUE_EMPTYKVA; 1214 } else { 1215 bp->b_qindex = BQUEUE_EMPTY; 1216 } 1217 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1218 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) { 1219 /* 1220 * Buffers with junk contents. Again these buffers had better 1221 * already be disassociated from their vnode. 1222 */ 1223 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp)); 1224 KKASSERT((bp->b_flags & B_HASHED) == 0); 1225 bp->b_flags |= B_INVAL; 1226 bp->b_qindex = BQUEUE_CLEAN; 1227 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1228 } else if (bp->b_flags & B_LOCKED) { 1229 /* 1230 * Buffers that are locked. 1231 */ 1232 bp->b_qindex = BQUEUE_LOCKED; 1233 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist); 1234 } else { 1235 /* 1236 * Remaining buffers. These buffers are still associated with 1237 * their vnode. 1238 */ 1239 switch(bp->b_flags & (B_DELWRI|B_AGE)) { 1240 case B_DELWRI | B_AGE: 1241 bp->b_qindex = BQUEUE_DIRTY; 1242 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_DIRTY], bp, b_freelist); 1243 break; 1244 case B_DELWRI: 1245 bp->b_qindex = BQUEUE_DIRTY; 1246 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist); 1247 break; 1248 case B_AGE: 1249 bp->b_qindex = BQUEUE_CLEAN; 1250 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1251 break; 1252 default: 1253 bp->b_qindex = BQUEUE_CLEAN; 1254 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1255 break; 1256 } 1257 } 1258 1259 /* 1260 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1261 * on the correct queue. 1262 */ 1263 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) 1264 bundirty(bp); 1265 1266 /* 1267 * Fixup numfreebuffers count. The bp is on an appropriate queue 1268 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1269 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1270 * if B_INVAL is set ). 1271 */ 1272 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1273 bufcountwakeup(); 1274 1275 /* 1276 * Something we can maybe free or reuse 1277 */ 1278 if (bp->b_bufsize || bp->b_kvasize) 1279 bufspacewakeup(); 1280 1281 /* 1282 * Clean up temporary flags and unlock the buffer. 1283 */ 1284 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | 1285 B_DIRECT | B_NOWDRAIN); 1286 BUF_UNLOCK(bp); 1287 crit_exit(); 1288 } 1289 1290 /* 1291 * bqrelse: 1292 * 1293 * Release a buffer back to the appropriate queue but do not try to free 1294 * it. The buffer is expected to be used again soon. 1295 * 1296 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1297 * biodone() to requeue an async I/O on completion. It is also used when 1298 * known good buffers need to be requeued but we think we may need the data 1299 * again soon. 1300 * 1301 * XXX we should be able to leave the B_RELBUF hint set on completion. 1302 */ 1303 void 1304 bqrelse(struct buf *bp) 1305 { 1306 crit_enter(); 1307 1308 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1309 1310 if (bp->b_qindex != BQUEUE_NONE) 1311 panic("bqrelse: free buffer onto another queue???"); 1312 if (BUF_REFCNTNB(bp) > 1) { 1313 /* do not release to free list */ 1314 panic("bqrelse: multiple refs"); 1315 BUF_UNLOCK(bp); 1316 crit_exit(); 1317 return; 1318 } 1319 if (bp->b_flags & B_LOCKED) { 1320 bp->b_flags &= ~B_ERROR; 1321 bp->b_qindex = BQUEUE_LOCKED; 1322 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist); 1323 /* buffers with stale but valid contents */ 1324 } else if (bp->b_flags & B_DELWRI) { 1325 bp->b_qindex = BQUEUE_DIRTY; 1326 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist); 1327 } else if (vm_page_count_severe()) { 1328 /* 1329 * We are too low on memory, we have to try to free the 1330 * buffer (most importantly: the wired pages making up its 1331 * backing store) *now*. 1332 */ 1333 crit_exit(); 1334 brelse(bp); 1335 return; 1336 } else { 1337 bp->b_qindex = BQUEUE_CLEAN; 1338 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1339 } 1340 1341 if ((bp->b_flags & B_LOCKED) == 0 && 1342 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1343 bufcountwakeup(); 1344 } 1345 1346 /* 1347 * Something we can maybe free or reuse. 1348 */ 1349 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1350 bufspacewakeup(); 1351 1352 /* 1353 * Final cleanup and unlock. Clear bits that are only used while a 1354 * buffer is actively locked. 1355 */ 1356 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1357 BUF_UNLOCK(bp); 1358 crit_exit(); 1359 } 1360 1361 /* 1362 * vfs_vmio_release: 1363 * 1364 * Return backing pages held by the buffer 'bp' back to the VM system 1365 * if possible. The pages are freed if they are no longer valid or 1366 * attempt to free if it was used for direct I/O otherwise they are 1367 * sent to the page cache. 1368 * 1369 * Pages that were marked busy are left alone and skipped. 1370 * 1371 * The KVA mapping (b_data) for the underlying pages is removed by 1372 * this function. 1373 */ 1374 static void 1375 vfs_vmio_release(struct buf *bp) 1376 { 1377 int i; 1378 vm_page_t m; 1379 1380 crit_enter(); 1381 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1382 m = bp->b_xio.xio_pages[i]; 1383 bp->b_xio.xio_pages[i] = NULL; 1384 /* 1385 * In order to keep page LRU ordering consistent, put 1386 * everything on the inactive queue. 1387 */ 1388 vm_page_unwire(m, 0); 1389 /* 1390 * We don't mess with busy pages, it is 1391 * the responsibility of the process that 1392 * busied the pages to deal with them. 1393 */ 1394 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1395 continue; 1396 1397 if (m->wire_count == 0) { 1398 vm_page_flag_clear(m, PG_ZERO); 1399 /* 1400 * Might as well free the page if we can and it has 1401 * no valid data. We also free the page if the 1402 * buffer was used for direct I/O. 1403 */ 1404 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1405 m->hold_count == 0) { 1406 vm_page_busy(m); 1407 vm_page_protect(m, VM_PROT_NONE); 1408 vm_page_free(m); 1409 } else if (bp->b_flags & B_DIRECT) { 1410 vm_page_try_to_free(m); 1411 } else if (vm_page_count_severe()) { 1412 vm_page_try_to_cache(m); 1413 } 1414 } 1415 } 1416 crit_exit(); 1417 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages); 1418 if (bp->b_bufsize) { 1419 bufspacewakeup(); 1420 bp->b_bufsize = 0; 1421 } 1422 bp->b_xio.xio_npages = 0; 1423 bp->b_flags &= ~B_VMIO; 1424 if (bp->b_vp) 1425 brelvp(bp); 1426 } 1427 1428 /* 1429 * vfs_bio_awrite: 1430 * 1431 * Implement clustered async writes for clearing out B_DELWRI buffers. 1432 * This is much better then the old way of writing only one buffer at 1433 * a time. Note that we may not be presented with the buffers in the 1434 * correct order, so we search for the cluster in both directions. 1435 * 1436 * The buffer is locked on call. 1437 */ 1438 int 1439 vfs_bio_awrite(struct buf *bp) 1440 { 1441 int i; 1442 int j; 1443 off_t loffset = bp->b_loffset; 1444 struct vnode *vp = bp->b_vp; 1445 int nbytes; 1446 struct buf *bpa; 1447 int nwritten; 1448 int size; 1449 1450 crit_enter(); 1451 /* 1452 * right now we support clustered writing only to regular files. If 1453 * we find a clusterable block we could be in the middle of a cluster 1454 * rather then at the beginning. 1455 * 1456 * NOTE: b_bio1 contains the logical loffset and is aliased 1457 * to b_loffset. b_bio2 contains the translated block number. 1458 */ 1459 if ((vp->v_type == VREG) && 1460 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1461 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1462 1463 size = vp->v_mount->mnt_stat.f_iosize; 1464 1465 for (i = size; i < MAXPHYS; i += size) { 1466 if ((bpa = findblk(vp, loffset + i)) && 1467 BUF_REFCNT(bpa) == 0 && 1468 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1469 (B_DELWRI | B_CLUSTEROK)) && 1470 (bpa->b_bufsize == size)) { 1471 if ((bpa->b_bio2.bio_offset == NOOFFSET) || 1472 (bpa->b_bio2.bio_offset != 1473 bp->b_bio2.bio_offset + i)) 1474 break; 1475 } else { 1476 break; 1477 } 1478 } 1479 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) { 1480 if ((bpa = findblk(vp, loffset - j)) && 1481 BUF_REFCNT(bpa) == 0 && 1482 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1483 (B_DELWRI | B_CLUSTEROK)) && 1484 (bpa->b_bufsize == size)) { 1485 if ((bpa->b_bio2.bio_offset == NOOFFSET) || 1486 (bpa->b_bio2.bio_offset != 1487 bp->b_bio2.bio_offset - j)) 1488 break; 1489 } else { 1490 break; 1491 } 1492 } 1493 j -= size; 1494 nbytes = (i + j); 1495 /* 1496 * this is a possible cluster write 1497 */ 1498 if (nbytes != size) { 1499 BUF_UNLOCK(bp); 1500 nwritten = cluster_wbuild(vp, size, 1501 loffset - j, nbytes); 1502 crit_exit(); 1503 return nwritten; 1504 } 1505 } 1506 1507 bremfree(bp); 1508 bp->b_flags |= B_ASYNC; 1509 1510 crit_exit(); 1511 /* 1512 * default (old) behavior, writing out only one block 1513 * 1514 * XXX returns b_bufsize instead of b_bcount for nwritten? 1515 */ 1516 nwritten = bp->b_bufsize; 1517 bwrite(bp); 1518 1519 return nwritten; 1520 } 1521 1522 /* 1523 * getnewbuf: 1524 * 1525 * Find and initialize a new buffer header, freeing up existing buffers 1526 * in the bufqueues as necessary. The new buffer is returned locked. 1527 * 1528 * Important: B_INVAL is not set. If the caller wishes to throw the 1529 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1530 * 1531 * We block if: 1532 * We have insufficient buffer headers 1533 * We have insufficient buffer space 1534 * buffer_map is too fragmented ( space reservation fails ) 1535 * If we have to flush dirty buffers ( but we try to avoid this ) 1536 * 1537 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1538 * Instead we ask the buf daemon to do it for us. We attempt to 1539 * avoid piecemeal wakeups of the pageout daemon. 1540 */ 1541 1542 static struct buf * 1543 getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1544 { 1545 struct buf *bp; 1546 struct buf *nbp; 1547 int defrag = 0; 1548 int nqindex; 1549 static int flushingbufs; 1550 1551 /* 1552 * We can't afford to block since we might be holding a vnode lock, 1553 * which may prevent system daemons from running. We deal with 1554 * low-memory situations by proactively returning memory and running 1555 * async I/O rather then sync I/O. 1556 */ 1557 1558 ++getnewbufcalls; 1559 --getnewbufrestarts; 1560 restart: 1561 ++getnewbufrestarts; 1562 1563 /* 1564 * Setup for scan. If we do not have enough free buffers, 1565 * we setup a degenerate case that immediately fails. Note 1566 * that if we are specially marked process, we are allowed to 1567 * dip into our reserves. 1568 * 1569 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1570 * 1571 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1572 * However, there are a number of cases (defragging, reusing, ...) 1573 * where we cannot backup. 1574 */ 1575 nqindex = BQUEUE_EMPTYKVA; 1576 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]); 1577 1578 if (nbp == NULL) { 1579 /* 1580 * If no EMPTYKVA buffers and we are either 1581 * defragging or reusing, locate a CLEAN buffer 1582 * to free or reuse. If bufspace useage is low 1583 * skip this step so we can allocate a new buffer. 1584 */ 1585 if (defrag || bufspace >= lobufspace) { 1586 nqindex = BQUEUE_CLEAN; 1587 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]); 1588 } 1589 1590 /* 1591 * If we could not find or were not allowed to reuse a 1592 * CLEAN buffer, check to see if it is ok to use an EMPTY 1593 * buffer. We can only use an EMPTY buffer if allocating 1594 * its KVA would not otherwise run us out of buffer space. 1595 */ 1596 if (nbp == NULL && defrag == 0 && 1597 bufspace + maxsize < hibufspace) { 1598 nqindex = BQUEUE_EMPTY; 1599 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]); 1600 } 1601 } 1602 1603 /* 1604 * Run scan, possibly freeing data and/or kva mappings on the fly 1605 * depending. 1606 */ 1607 1608 while ((bp = nbp) != NULL) { 1609 int qindex = nqindex; 1610 1611 /* 1612 * Calculate next bp ( we can only use it if we do not block 1613 * or do other fancy things ). 1614 */ 1615 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1616 switch(qindex) { 1617 case BQUEUE_EMPTY: 1618 nqindex = BQUEUE_EMPTYKVA; 1619 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]))) 1620 break; 1621 /* fall through */ 1622 case BQUEUE_EMPTYKVA: 1623 nqindex = BQUEUE_CLEAN; 1624 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]))) 1625 break; 1626 /* fall through */ 1627 case BQUEUE_CLEAN: 1628 /* 1629 * nbp is NULL. 1630 */ 1631 break; 1632 } 1633 } 1634 1635 /* 1636 * Sanity Checks 1637 */ 1638 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1639 1640 /* 1641 * Note: we no longer distinguish between VMIO and non-VMIO 1642 * buffers. 1643 */ 1644 1645 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1646 1647 /* 1648 * If we are defragging then we need a buffer with 1649 * b_kvasize != 0. XXX this situation should no longer 1650 * occur, if defrag is non-zero the buffer's b_kvasize 1651 * should also be non-zero at this point. XXX 1652 */ 1653 if (defrag && bp->b_kvasize == 0) { 1654 printf("Warning: defrag empty buffer %p\n", bp); 1655 continue; 1656 } 1657 1658 /* 1659 * Start freeing the bp. This is somewhat involved. nbp 1660 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers 1661 * on the clean list must be disassociated from their 1662 * current vnode. Buffers on the empty[kva] lists have 1663 * already been disassociated. 1664 */ 1665 1666 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 1667 printf("getnewbuf: warning, locked buf %p, race corrected\n", bp); 1668 tsleep(&bd_request, 0, "gnbxxx", hz / 100); 1669 goto restart; 1670 } 1671 if (bp->b_qindex != qindex) { 1672 printf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex); 1673 BUF_UNLOCK(bp); 1674 goto restart; 1675 } 1676 bremfree(bp); 1677 1678 if (qindex == BQUEUE_CLEAN) { 1679 if (bp->b_flags & B_VMIO) { 1680 bp->b_flags &= ~B_ASYNC; 1681 vfs_vmio_release(bp); 1682 } 1683 if (bp->b_vp) 1684 brelvp(bp); 1685 } 1686 1687 /* 1688 * NOTE: nbp is now entirely invalid. We can only restart 1689 * the scan from this point on. 1690 * 1691 * Get the rest of the buffer freed up. b_kva* is still 1692 * valid after this operation. 1693 */ 1694 1695 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex)); 1696 KKASSERT((bp->b_flags & B_HASHED) == 0); 1697 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 1698 (*bioops.io_deallocate)(bp); 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 = B_BNOCLIP; 1709 bp->b_cmd = BUF_CMD_DONE; 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 spin_lock_wr(&needsbuffer_spin); 1894 bd_request = 0; 1895 msleep(&bd_request, &needsbuffer_spin, 0, "psleep", hz); 1896 spin_unlock_wr(&needsbuffer_spin); 1897 } else { 1898 /* 1899 * We couldn't find any flushable dirty buffers but 1900 * still have too many dirty buffers, we 1901 * have to sleep and try again. (rare) 1902 */ 1903 tsleep(&bd_request, 0, "qsleep", hz / 2); 1904 } 1905 } 1906 } 1907 1908 /* 1909 * flushbufqueues: 1910 * 1911 * Try to flush a buffer in the dirty queue. We must be careful to 1912 * free up B_INVAL buffers instead of write them, which NFS is 1913 * particularly sensitive to. 1914 */ 1915 1916 static int 1917 flushbufqueues(void) 1918 { 1919 struct buf *bp; 1920 int r = 0; 1921 1922 bp = TAILQ_FIRST(&bufqueues[BQUEUE_DIRTY]); 1923 1924 while (bp) { 1925 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); 1926 if (bp->b_flags & B_DELWRI) { 1927 if (bp->b_flags & B_INVAL) { 1928 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1929 panic("flushbufqueues: locked buf"); 1930 bremfree(bp); 1931 brelse(bp); 1932 ++r; 1933 break; 1934 } 1935 if (LIST_FIRST(&bp->b_dep) != NULL && 1936 bioops.io_countdeps && 1937 (bp->b_flags & B_DEFERRED) == 0 && 1938 (*bioops.io_countdeps)(bp, 0)) { 1939 TAILQ_REMOVE(&bufqueues[BQUEUE_DIRTY], 1940 bp, b_freelist); 1941 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], 1942 bp, b_freelist); 1943 bp->b_flags |= B_DEFERRED; 1944 bp = TAILQ_FIRST(&bufqueues[BQUEUE_DIRTY]); 1945 continue; 1946 } 1947 1948 /* 1949 * Only write it out if we can successfully lock 1950 * it. 1951 */ 1952 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 1953 vfs_bio_awrite(bp); 1954 ++r; 1955 break; 1956 } 1957 } 1958 bp = TAILQ_NEXT(bp, b_freelist); 1959 } 1960 return (r); 1961 } 1962 1963 /* 1964 * inmem: 1965 * 1966 * Returns true if no I/O is needed to access the associated VM object. 1967 * This is like findblk except it also hunts around in the VM system for 1968 * the data. 1969 * 1970 * Note that we ignore vm_page_free() races from interrupts against our 1971 * lookup, since if the caller is not protected our return value will not 1972 * be any more valid then otherwise once we exit the critical section. 1973 */ 1974 int 1975 inmem(struct vnode *vp, off_t loffset) 1976 { 1977 vm_object_t obj; 1978 vm_offset_t toff, tinc, size; 1979 vm_page_t m; 1980 1981 if (findblk(vp, loffset)) 1982 return 1; 1983 if (vp->v_mount == NULL) 1984 return 0; 1985 if ((obj = vp->v_object) == NULL) 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 1992 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1993 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff)); 1994 if (m == NULL) 1995 return 0; 1996 tinc = size; 1997 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK)) 1998 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK); 1999 if (vm_page_is_valid(m, 2000 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) 2001 return 0; 2002 } 2003 return 1; 2004 } 2005 2006 /* 2007 * vfs_setdirty: 2008 * 2009 * Sets the dirty range for a buffer based on the status of the dirty 2010 * bits in the pages comprising the buffer. 2011 * 2012 * The range is limited to the size of the buffer. 2013 * 2014 * This routine is primarily used by NFS, but is generalized for the 2015 * B_VMIO case. 2016 */ 2017 static void 2018 vfs_setdirty(struct buf *bp) 2019 { 2020 int i; 2021 vm_object_t object; 2022 2023 /* 2024 * Degenerate case - empty buffer 2025 */ 2026 2027 if (bp->b_bufsize == 0) 2028 return; 2029 2030 /* 2031 * We qualify the scan for modified pages on whether the 2032 * object has been flushed yet. The OBJ_WRITEABLE flag 2033 * is not cleared simply by protecting pages off. 2034 */ 2035 2036 if ((bp->b_flags & B_VMIO) == 0) 2037 return; 2038 2039 object = bp->b_xio.xio_pages[0]->object; 2040 2041 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2042 printf("Warning: object %p writeable but not mightbedirty\n", object); 2043 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2044 printf("Warning: object %p mightbedirty but not writeable\n", object); 2045 2046 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2047 vm_offset_t boffset; 2048 vm_offset_t eoffset; 2049 2050 /* 2051 * test the pages to see if they have been modified directly 2052 * by users through the VM system. 2053 */ 2054 for (i = 0; i < bp->b_xio.xio_npages; i++) { 2055 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO); 2056 vm_page_test_dirty(bp->b_xio.xio_pages[i]); 2057 } 2058 2059 /* 2060 * Calculate the encompassing dirty range, boffset and eoffset, 2061 * (eoffset - boffset) bytes. 2062 */ 2063 2064 for (i = 0; i < bp->b_xio.xio_npages; i++) { 2065 if (bp->b_xio.xio_pages[i]->dirty) 2066 break; 2067 } 2068 boffset = (i << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK); 2069 2070 for (i = bp->b_xio.xio_npages - 1; i >= 0; --i) { 2071 if (bp->b_xio.xio_pages[i]->dirty) { 2072 break; 2073 } 2074 } 2075 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK); 2076 2077 /* 2078 * Fit it to the buffer. 2079 */ 2080 2081 if (eoffset > bp->b_bcount) 2082 eoffset = bp->b_bcount; 2083 2084 /* 2085 * If we have a good dirty range, merge with the existing 2086 * dirty range. 2087 */ 2088 2089 if (boffset < eoffset) { 2090 if (bp->b_dirtyoff > boffset) 2091 bp->b_dirtyoff = boffset; 2092 if (bp->b_dirtyend < eoffset) 2093 bp->b_dirtyend = eoffset; 2094 } 2095 } 2096 } 2097 2098 /* 2099 * findblk: 2100 * 2101 * Locate and return the specified buffer, or NULL if the buffer does 2102 * not exist. Do not attempt to lock the buffer or manipulate it in 2103 * any way. The caller must validate that the correct buffer has been 2104 * obtain after locking it. 2105 */ 2106 struct buf * 2107 findblk(struct vnode *vp, off_t loffset) 2108 { 2109 struct buf *bp; 2110 2111 crit_enter(); 2112 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset); 2113 crit_exit(); 2114 return(bp); 2115 } 2116 2117 /* 2118 * getblk: 2119 * 2120 * Get a block given a specified block and offset into a file/device. 2121 * B_INVAL may or may not be set on return. The caller should clear 2122 * B_INVAL prior to initiating a READ. 2123 * 2124 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE 2125 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ, 2126 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer 2127 * without doing any of those things the system will likely believe 2128 * the buffer to be valid (especially if it is not B_VMIO), and the 2129 * next getblk() will return the buffer with B_CACHE set. 2130 * 2131 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2132 * an existing buffer. 2133 * 2134 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2135 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2136 * and then cleared based on the backing VM. If the previous buffer is 2137 * non-0-sized but invalid, B_CACHE will be cleared. 2138 * 2139 * If getblk() must create a new buffer, the new buffer is returned with 2140 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2141 * case it is returned with B_INVAL clear and B_CACHE set based on the 2142 * backing VM. 2143 * 2144 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2145 * B_CACHE bit is clear. 2146 * 2147 * What this means, basically, is that the caller should use B_CACHE to 2148 * determine whether the buffer is fully valid or not and should clear 2149 * B_INVAL prior to issuing a read. If the caller intends to validate 2150 * the buffer by loading its data area with something, the caller needs 2151 * to clear B_INVAL. If the caller does this without issuing an I/O, 2152 * the caller should set B_CACHE ( as an optimization ), else the caller 2153 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2154 * a write attempt or if it was a successfull read. If the caller 2155 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR 2156 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2157 */ 2158 struct buf * 2159 getblk(struct vnode *vp, off_t loffset, int size, int slpflag, int slptimeo) 2160 { 2161 struct buf *bp; 2162 2163 if (size > MAXBSIZE) 2164 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE); 2165 if (vp->v_object == NULL) 2166 panic("getblk: vnode %p has no object!", vp); 2167 2168 crit_enter(); 2169 loop: 2170 /* 2171 * Block if we are low on buffers. Certain processes are allowed 2172 * to completely exhaust the buffer cache. 2173 * 2174 * If this check ever becomes a bottleneck it may be better to 2175 * move it into the else, when findblk() fails. At the moment 2176 * it isn't a problem. 2177 * 2178 * XXX remove, we cannot afford to block anywhere if holding a vnode 2179 * lock in low-memory situation, so take it to the max. 2180 */ 2181 if (numfreebuffers == 0) { 2182 if (!curproc) 2183 return NULL; 2184 needsbuffer |= VFS_BIO_NEED_ANY; 2185 tsleep(&needsbuffer, slpflag, "newbuf", slptimeo); 2186 } 2187 2188 if ((bp = findblk(vp, loffset))) { 2189 /* 2190 * The buffer was found in the cache, but we need to lock it. 2191 * Even with LK_NOWAIT the lockmgr may break our critical 2192 * section, so double-check the validity of the buffer 2193 * once the lock has been obtained. 2194 */ 2195 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2196 int lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL; 2197 if (slpflag & PCATCH) 2198 lkflags |= LK_PCATCH; 2199 if (BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo) == 2200 ENOLCK) { 2201 goto loop; 2202 } 2203 crit_exit(); 2204 return (NULL); 2205 } 2206 2207 /* 2208 * Once the buffer has been locked, make sure we didn't race 2209 * a buffer recyclement. Buffers that are no longer hashed 2210 * will have b_vp == NULL, so this takes care of that check 2211 * as well. 2212 */ 2213 if (bp->b_vp != vp || bp->b_loffset != loffset) { 2214 printf("Warning buffer %p (vp %p loffset %lld) was recycled\n", bp, vp, loffset); 2215 BUF_UNLOCK(bp); 2216 goto loop; 2217 } 2218 2219 /* 2220 * All vnode-based buffers must be backed by a VM object. 2221 */ 2222 KKASSERT(bp->b_flags & B_VMIO); 2223 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 2224 2225 /* 2226 * Make sure that B_INVAL buffers do not have a cached 2227 * block number translation. 2228 */ 2229 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) { 2230 printf("Warning invalid buffer %p (vp %p loffset %lld) did not have cleared bio_offset cache\n", bp, vp, loffset); 2231 clearbiocache(&bp->b_bio2); 2232 } 2233 2234 /* 2235 * The buffer is locked. B_CACHE is cleared if the buffer is 2236 * invalid. 2237 */ 2238 if (bp->b_flags & B_INVAL) 2239 bp->b_flags &= ~B_CACHE; 2240 bremfree(bp); 2241 2242 /* 2243 * Any size inconsistancy with a dirty buffer or a buffer 2244 * with a softupdates dependancy must be resolved. Resizing 2245 * the buffer in such circumstances can lead to problems. 2246 */ 2247 if (size != bp->b_bcount) { 2248 if (bp->b_flags & B_DELWRI) { 2249 bp->b_flags |= B_NOCACHE; 2250 bwrite(bp); 2251 } else if (LIST_FIRST(&bp->b_dep)) { 2252 bp->b_flags |= B_NOCACHE; 2253 bwrite(bp); 2254 } else { 2255 bp->b_flags |= B_RELBUF; 2256 brelse(bp); 2257 } 2258 goto loop; 2259 } 2260 KKASSERT(size <= bp->b_kvasize); 2261 KASSERT(bp->b_loffset != NOOFFSET, 2262 ("getblk: no buffer offset")); 2263 2264 /* 2265 * A buffer with B_DELWRI set and B_CACHE clear must 2266 * be committed before we can return the buffer in 2267 * order to prevent the caller from issuing a read 2268 * ( due to B_CACHE not being set ) and overwriting 2269 * it. 2270 * 2271 * Most callers, including NFS and FFS, need this to 2272 * operate properly either because they assume they 2273 * can issue a read if B_CACHE is not set, or because 2274 * ( for example ) an uncached B_DELWRI might loop due 2275 * to softupdates re-dirtying the buffer. In the latter 2276 * case, B_CACHE is set after the first write completes, 2277 * preventing further loops. 2278 * 2279 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2280 * above while extending the buffer, we cannot allow the 2281 * buffer to remain with B_CACHE set after the write 2282 * completes or it will represent a corrupt state. To 2283 * deal with this we set B_NOCACHE to scrap the buffer 2284 * after the write. 2285 * 2286 * We might be able to do something fancy, like setting 2287 * B_CACHE in bwrite() except if B_DELWRI is already set, 2288 * so the below call doesn't set B_CACHE, but that gets real 2289 * confusing. This is much easier. 2290 */ 2291 2292 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2293 bp->b_flags |= B_NOCACHE; 2294 bwrite(bp); 2295 goto loop; 2296 } 2297 crit_exit(); 2298 } else { 2299 /* 2300 * Buffer is not in-core, create new buffer. The buffer 2301 * returned by getnewbuf() is locked. Note that the returned 2302 * buffer is also considered valid (not marked B_INVAL). 2303 * 2304 * Calculating the offset for the I/O requires figuring out 2305 * the block size. We use DEV_BSIZE for VBLK or VCHR and 2306 * the mount's f_iosize otherwise. If the vnode does not 2307 * have an associated mount we assume that the passed size is 2308 * the block size. 2309 * 2310 * Note that vn_isdisk() cannot be used here since it may 2311 * return a failure for numerous reasons. Note that the 2312 * buffer size may be larger then the block size (the caller 2313 * will use block numbers with the proper multiple). Beware 2314 * of using any v_* fields which are part of unions. In 2315 * particular, in DragonFly the mount point overloading 2316 * mechanism is such that the underlying directory (with a 2317 * non-NULL v_mountedhere) is not a special case. 2318 */ 2319 int bsize, maxsize; 2320 2321 if (vp->v_type == VBLK || vp->v_type == VCHR) 2322 bsize = DEV_BSIZE; 2323 else if (vp->v_mount) 2324 bsize = vp->v_mount->mnt_stat.f_iosize; 2325 else 2326 bsize = size; 2327 2328 maxsize = size + (loffset & PAGE_MASK); 2329 maxsize = imax(maxsize, bsize); 2330 2331 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2332 if (slpflag || slptimeo) { 2333 crit_exit(); 2334 return NULL; 2335 } 2336 goto loop; 2337 } 2338 2339 /* 2340 * This code is used to make sure that a buffer is not 2341 * created while the getnewbuf routine is blocked. 2342 * This can be a problem whether the vnode is locked or not. 2343 * If the buffer is created out from under us, we have to 2344 * throw away the one we just created. There is now window 2345 * race because we are safely running in a critical section 2346 * from the point of the duplicate buffer creation through 2347 * to here, and we've locked the buffer. 2348 */ 2349 if (findblk(vp, loffset)) { 2350 bp->b_flags |= B_INVAL; 2351 brelse(bp); 2352 goto loop; 2353 } 2354 2355 /* 2356 * Insert the buffer into the hash, so that it can 2357 * be found by findblk(). 2358 * 2359 * Make sure the translation layer has been cleared. 2360 */ 2361 bp->b_loffset = loffset; 2362 bp->b_bio2.bio_offset = NOOFFSET; 2363 /* bp->b_bio2.bio_next = NULL; */ 2364 2365 bgetvp(vp, bp); 2366 2367 /* 2368 * All vnode-based buffers must be backed by a VM object. 2369 */ 2370 KKASSERT(vp->v_object != NULL); 2371 bp->b_flags |= B_VMIO; 2372 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 2373 2374 allocbuf(bp, size); 2375 2376 crit_exit(); 2377 } 2378 return (bp); 2379 } 2380 2381 /* 2382 * geteblk: 2383 * 2384 * Get an empty, disassociated buffer of given size. The buffer is 2385 * initially set to B_INVAL. 2386 * 2387 * critical section protection is not required for the allocbuf() 2388 * call because races are impossible here. 2389 */ 2390 struct buf * 2391 geteblk(int size) 2392 { 2393 struct buf *bp; 2394 int maxsize; 2395 2396 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2397 2398 crit_enter(); 2399 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2400 ; 2401 crit_exit(); 2402 allocbuf(bp, size); 2403 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2404 return (bp); 2405 } 2406 2407 2408 /* 2409 * allocbuf: 2410 * 2411 * This code constitutes the buffer memory from either anonymous system 2412 * memory (in the case of non-VMIO operations) or from an associated 2413 * VM object (in the case of VMIO operations). This code is able to 2414 * resize a buffer up or down. 2415 * 2416 * Note that this code is tricky, and has many complications to resolve 2417 * deadlock or inconsistant data situations. Tread lightly!!! 2418 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2419 * the caller. Calling this code willy nilly can result in the loss of data. 2420 * 2421 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2422 * B_CACHE for the non-VMIO case. 2423 * 2424 * This routine does not need to be called from a critical section but you 2425 * must own the buffer. 2426 */ 2427 int 2428 allocbuf(struct buf *bp, int size) 2429 { 2430 int newbsize, mbsize; 2431 int i; 2432 2433 if (BUF_REFCNT(bp) == 0) 2434 panic("allocbuf: buffer not busy"); 2435 2436 if (bp->b_kvasize < size) 2437 panic("allocbuf: buffer too small"); 2438 2439 if ((bp->b_flags & B_VMIO) == 0) { 2440 caddr_t origbuf; 2441 int origbufsize; 2442 /* 2443 * Just get anonymous memory from the kernel. Don't 2444 * mess with B_CACHE. 2445 */ 2446 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2447 if (bp->b_flags & B_MALLOC) 2448 newbsize = mbsize; 2449 else 2450 newbsize = round_page(size); 2451 2452 if (newbsize < bp->b_bufsize) { 2453 /* 2454 * Malloced buffers are not shrunk 2455 */ 2456 if (bp->b_flags & B_MALLOC) { 2457 if (newbsize) { 2458 bp->b_bcount = size; 2459 } else { 2460 free(bp->b_data, M_BIOBUF); 2461 if (bp->b_bufsize) { 2462 bufmallocspace -= bp->b_bufsize; 2463 bufspacewakeup(); 2464 bp->b_bufsize = 0; 2465 } 2466 bp->b_data = bp->b_kvabase; 2467 bp->b_bcount = 0; 2468 bp->b_flags &= ~B_MALLOC; 2469 } 2470 return 1; 2471 } 2472 vm_hold_free_pages( 2473 bp, 2474 (vm_offset_t) bp->b_data + newbsize, 2475 (vm_offset_t) bp->b_data + bp->b_bufsize); 2476 } else if (newbsize > bp->b_bufsize) { 2477 /* 2478 * We only use malloced memory on the first allocation. 2479 * and revert to page-allocated memory when the buffer 2480 * grows. 2481 */ 2482 if ((bufmallocspace < maxbufmallocspace) && 2483 (bp->b_bufsize == 0) && 2484 (mbsize <= PAGE_SIZE/2)) { 2485 2486 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2487 bp->b_bufsize = mbsize; 2488 bp->b_bcount = size; 2489 bp->b_flags |= B_MALLOC; 2490 bufmallocspace += mbsize; 2491 return 1; 2492 } 2493 origbuf = NULL; 2494 origbufsize = 0; 2495 /* 2496 * If the buffer is growing on its other-than-first 2497 * allocation, then we revert to the page-allocation 2498 * scheme. 2499 */ 2500 if (bp->b_flags & B_MALLOC) { 2501 origbuf = bp->b_data; 2502 origbufsize = bp->b_bufsize; 2503 bp->b_data = bp->b_kvabase; 2504 if (bp->b_bufsize) { 2505 bufmallocspace -= bp->b_bufsize; 2506 bufspacewakeup(); 2507 bp->b_bufsize = 0; 2508 } 2509 bp->b_flags &= ~B_MALLOC; 2510 newbsize = round_page(newbsize); 2511 } 2512 vm_hold_load_pages( 2513 bp, 2514 (vm_offset_t) bp->b_data + bp->b_bufsize, 2515 (vm_offset_t) bp->b_data + newbsize); 2516 if (origbuf) { 2517 bcopy(origbuf, bp->b_data, origbufsize); 2518 free(origbuf, M_BIOBUF); 2519 } 2520 } 2521 } else { 2522 vm_page_t m; 2523 int desiredpages; 2524 2525 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2526 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) + 2527 newbsize + PAGE_MASK) >> PAGE_SHIFT; 2528 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES); 2529 2530 if (bp->b_flags & B_MALLOC) 2531 panic("allocbuf: VMIO buffer can't be malloced"); 2532 /* 2533 * Set B_CACHE initially if buffer is 0 length or will become 2534 * 0-length. 2535 */ 2536 if (size == 0 || bp->b_bufsize == 0) 2537 bp->b_flags |= B_CACHE; 2538 2539 if (newbsize < bp->b_bufsize) { 2540 /* 2541 * DEV_BSIZE aligned new buffer size is less then the 2542 * DEV_BSIZE aligned existing buffer size. Figure out 2543 * if we have to remove any pages. 2544 */ 2545 if (desiredpages < bp->b_xio.xio_npages) { 2546 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) { 2547 /* 2548 * the page is not freed here -- it 2549 * is the responsibility of 2550 * vnode_pager_setsize 2551 */ 2552 m = bp->b_xio.xio_pages[i]; 2553 KASSERT(m != bogus_page, 2554 ("allocbuf: bogus page found")); 2555 while (vm_page_sleep_busy(m, TRUE, "biodep")) 2556 ; 2557 2558 bp->b_xio.xio_pages[i] = NULL; 2559 vm_page_unwire(m, 0); 2560 } 2561 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2562 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages)); 2563 bp->b_xio.xio_npages = desiredpages; 2564 } 2565 } else if (size > bp->b_bcount) { 2566 /* 2567 * We are growing the buffer, possibly in a 2568 * byte-granular fashion. 2569 */ 2570 struct vnode *vp; 2571 vm_object_t obj; 2572 vm_offset_t toff; 2573 vm_offset_t tinc; 2574 2575 /* 2576 * Step 1, bring in the VM pages from the object, 2577 * allocating them if necessary. We must clear 2578 * B_CACHE if these pages are not valid for the 2579 * range covered by the buffer. 2580 * 2581 * critical section protection is required to protect 2582 * against interrupts unbusying and freeing pages 2583 * between our vm_page_lookup() and our 2584 * busycheck/wiring call. 2585 */ 2586 vp = bp->b_vp; 2587 obj = vp->v_object; 2588 2589 crit_enter(); 2590 while (bp->b_xio.xio_npages < desiredpages) { 2591 vm_page_t m; 2592 vm_pindex_t pi; 2593 2594 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages; 2595 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2596 /* 2597 * note: must allocate system pages 2598 * since blocking here could intefere 2599 * with paging I/O, no matter which 2600 * process we are. 2601 */ 2602 m = vm_page_alloc(obj, pi, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM); 2603 if (m == NULL) { 2604 vm_wait(); 2605 vm_pageout_deficit += desiredpages - 2606 bp->b_xio.xio_npages; 2607 } else { 2608 vm_page_wire(m); 2609 vm_page_wakeup(m); 2610 bp->b_flags &= ~B_CACHE; 2611 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 2612 ++bp->b_xio.xio_npages; 2613 } 2614 continue; 2615 } 2616 2617 /* 2618 * We found a page. If we have to sleep on it, 2619 * retry because it might have gotten freed out 2620 * from under us. 2621 * 2622 * We can only test PG_BUSY here. Blocking on 2623 * m->busy might lead to a deadlock: 2624 * 2625 * vm_fault->getpages->cluster_read->allocbuf 2626 * 2627 */ 2628 2629 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 2630 continue; 2631 2632 /* 2633 * We have a good page. Should we wakeup the 2634 * page daemon? 2635 */ 2636 if ((curthread != pagethread) && 2637 ((m->queue - m->pc) == PQ_CACHE) && 2638 ((vmstats.v_free_count + vmstats.v_cache_count) < 2639 (vmstats.v_free_min + vmstats.v_cache_min))) { 2640 pagedaemon_wakeup(); 2641 } 2642 vm_page_flag_clear(m, PG_ZERO); 2643 vm_page_wire(m); 2644 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 2645 ++bp->b_xio.xio_npages; 2646 } 2647 crit_exit(); 2648 2649 /* 2650 * Step 2. We've loaded the pages into the buffer, 2651 * we have to figure out if we can still have B_CACHE 2652 * set. Note that B_CACHE is set according to the 2653 * byte-granular range ( bcount and size ), not the 2654 * aligned range ( newbsize ). 2655 * 2656 * The VM test is against m->valid, which is DEV_BSIZE 2657 * aligned. Needless to say, the validity of the data 2658 * needs to also be DEV_BSIZE aligned. Note that this 2659 * fails with NFS if the server or some other client 2660 * extends the file's EOF. If our buffer is resized, 2661 * B_CACHE may remain set! XXX 2662 */ 2663 2664 toff = bp->b_bcount; 2665 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK); 2666 2667 while ((bp->b_flags & B_CACHE) && toff < size) { 2668 vm_pindex_t pi; 2669 2670 if (tinc > (size - toff)) 2671 tinc = size - toff; 2672 2673 pi = ((bp->b_loffset & PAGE_MASK) + toff) >> 2674 PAGE_SHIFT; 2675 2676 vfs_buf_test_cache( 2677 bp, 2678 bp->b_loffset, 2679 toff, 2680 tinc, 2681 bp->b_xio.xio_pages[pi] 2682 ); 2683 toff += tinc; 2684 tinc = PAGE_SIZE; 2685 } 2686 2687 /* 2688 * Step 3, fixup the KVM pmap. Remember that 2689 * bp->b_data is relative to bp->b_loffset, but 2690 * bp->b_loffset may be offset into the first page. 2691 */ 2692 2693 bp->b_data = (caddr_t) 2694 trunc_page((vm_offset_t)bp->b_data); 2695 pmap_qenter( 2696 (vm_offset_t)bp->b_data, 2697 bp->b_xio.xio_pages, 2698 bp->b_xio.xio_npages 2699 ); 2700 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2701 (vm_offset_t)(bp->b_loffset & PAGE_MASK)); 2702 } 2703 } 2704 if (newbsize < bp->b_bufsize) 2705 bufspacewakeup(); 2706 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2707 bp->b_bcount = size; /* requested buffer size */ 2708 return 1; 2709 } 2710 2711 /* 2712 * biowait: 2713 * 2714 * Wait for buffer I/O completion, returning error status. The buffer 2715 * is left locked on return. B_EINTR is converted into an EINTR error 2716 * and cleared. 2717 * 2718 * NOTE! The original b_cmd is lost on return, since b_cmd will be 2719 * set to BUF_CMD_DONE. 2720 */ 2721 int 2722 biowait(struct buf *bp) 2723 { 2724 crit_enter(); 2725 while (bp->b_cmd != BUF_CMD_DONE) { 2726 if (bp->b_cmd == BUF_CMD_READ) 2727 tsleep(bp, 0, "biord", 0); 2728 else 2729 tsleep(bp, 0, "biowr", 0); 2730 } 2731 crit_exit(); 2732 if (bp->b_flags & B_EINTR) { 2733 bp->b_flags &= ~B_EINTR; 2734 return (EINTR); 2735 } 2736 if (bp->b_flags & B_ERROR) { 2737 return (bp->b_error ? bp->b_error : EIO); 2738 } else { 2739 return (0); 2740 } 2741 } 2742 2743 /* 2744 * This associates a tracking count with an I/O. vn_strategy() and 2745 * dev_dstrategy() do this automatically but there are a few cases 2746 * where a vnode or device layer is bypassed when a block translation 2747 * is cached. In such cases bio_start_transaction() may be called on 2748 * the bypassed layers so the system gets an I/O in progress indication 2749 * for those higher layers. 2750 */ 2751 void 2752 bio_start_transaction(struct bio *bio, struct bio_track *track) 2753 { 2754 bio->bio_track = track; 2755 atomic_add_int(&track->bk_active, 1); 2756 } 2757 2758 /* 2759 * Initiate I/O on a vnode. 2760 */ 2761 void 2762 vn_strategy(struct vnode *vp, struct bio *bio) 2763 { 2764 struct bio_track *track; 2765 2766 KKASSERT(bio->bio_buf->b_cmd != BUF_CMD_DONE); 2767 if (bio->bio_buf->b_cmd == BUF_CMD_READ) 2768 track = &vp->v_track_read; 2769 else 2770 track = &vp->v_track_write; 2771 bio->bio_track = track; 2772 atomic_add_int(&track->bk_active, 1); 2773 vop_strategy(*vp->v_ops, vp, bio); 2774 } 2775 2776 2777 /* 2778 * biodone: 2779 * 2780 * Finish I/O on a buffer, optionally calling a completion function. 2781 * This is usually called from an interrupt so process blocking is 2782 * not allowed. 2783 * 2784 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2785 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2786 * assuming B_INVAL is clear. 2787 * 2788 * For the VMIO case, we set B_CACHE if the op was a read and no 2789 * read error occured, or if the op was a write. B_CACHE is never 2790 * set if the buffer is invalid or otherwise uncacheable. 2791 * 2792 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2793 * initiator to leave B_INVAL set to brelse the buffer out of existance 2794 * in the biodone routine. 2795 */ 2796 void 2797 biodone(struct bio *bio) 2798 { 2799 struct buf *bp = bio->bio_buf; 2800 buf_cmd_t cmd; 2801 2802 crit_enter(); 2803 2804 KASSERT(BUF_REFCNTNB(bp) > 0, 2805 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp))); 2806 KASSERT(bp->b_cmd != BUF_CMD_DONE, 2807 ("biodone: bp %p already done!", bp)); 2808 2809 runningbufwakeup(bp); 2810 2811 /* 2812 * Run up the chain of BIO's. Leave b_cmd intact for the duration. 2813 */ 2814 while (bio) { 2815 biodone_t *done_func; 2816 struct bio_track *track; 2817 2818 /* 2819 * BIO tracking. Most but not all BIOs are tracked. 2820 */ 2821 if ((track = bio->bio_track) != NULL) { 2822 atomic_subtract_int(&track->bk_active, 1); 2823 if (track->bk_active < 0) { 2824 panic("biodone: bad active count bio %p\n", 2825 bio); 2826 } 2827 if (track->bk_waitflag) { 2828 track->bk_waitflag = 0; 2829 wakeup(track); 2830 } 2831 bio->bio_track = NULL; 2832 } 2833 2834 /* 2835 * A bio_done function terminates the loop. The function 2836 * will be responsible for any further chaining and/or 2837 * buffer management. 2838 * 2839 * WARNING! The done function can deallocate the buffer! 2840 */ 2841 if ((done_func = bio->bio_done) != NULL) { 2842 bio->bio_done = NULL; 2843 done_func(bio); 2844 crit_exit(); 2845 return; 2846 } 2847 bio = bio->bio_prev; 2848 } 2849 2850 cmd = bp->b_cmd; 2851 bp->b_cmd = BUF_CMD_DONE; 2852 2853 /* 2854 * Only reads and writes are processed past this point. 2855 */ 2856 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) { 2857 brelse(bp); 2858 crit_exit(); 2859 return; 2860 } 2861 2862 /* 2863 * Warning: softupdates may re-dirty the buffer. 2864 */ 2865 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 2866 (*bioops.io_complete)(bp); 2867 2868 if (bp->b_flags & B_VMIO) { 2869 int i; 2870 vm_ooffset_t foff; 2871 vm_page_t m; 2872 vm_object_t obj; 2873 int iosize; 2874 struct vnode *vp = bp->b_vp; 2875 2876 obj = vp->v_object; 2877 2878 #if defined(VFS_BIO_DEBUG) 2879 if (vp->v_holdcnt == 0) 2880 panic("biodone: zero vnode hold count"); 2881 if ((vp->v_flag & VOBJBUF) == 0) 2882 panic("biodone: vnode is not setup for merged cache"); 2883 #endif 2884 2885 foff = bp->b_loffset; 2886 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset")); 2887 KASSERT(obj != NULL, ("biodone: missing VM object")); 2888 2889 #if defined(VFS_BIO_DEBUG) 2890 if (obj->paging_in_progress < bp->b_xio.xio_npages) { 2891 printf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n", 2892 obj->paging_in_progress, bp->b_xio.xio_npages); 2893 } 2894 #endif 2895 2896 /* 2897 * Set B_CACHE if the op was a normal read and no error 2898 * occured. B_CACHE is set for writes in the b*write() 2899 * routines. 2900 */ 2901 iosize = bp->b_bcount - bp->b_resid; 2902 if (cmd == BUF_CMD_READ && (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) { 2903 bp->b_flags |= B_CACHE; 2904 } 2905 2906 for (i = 0; i < bp->b_xio.xio_npages; i++) { 2907 int bogusflag = 0; 2908 int resid; 2909 2910 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2911 if (resid > iosize) 2912 resid = iosize; 2913 2914 /* 2915 * cleanup bogus pages, restoring the originals. Since 2916 * the originals should still be wired, we don't have 2917 * to worry about interrupt/freeing races destroying 2918 * the VM object association. 2919 */ 2920 m = bp->b_xio.xio_pages[i]; 2921 if (m == bogus_page) { 2922 bogusflag = 1; 2923 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2924 if (m == NULL) 2925 panic("biodone: page disappeared"); 2926 bp->b_xio.xio_pages[i] = m; 2927 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 2928 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 2929 } 2930 #if defined(VFS_BIO_DEBUG) 2931 if (OFF_TO_IDX(foff) != m->pindex) { 2932 printf( 2933 "biodone: foff(%lu)/m->pindex(%d) mismatch\n", 2934 (unsigned long)foff, m->pindex); 2935 } 2936 #endif 2937 2938 /* 2939 * In the write case, the valid and clean bits are 2940 * already changed correctly ( see bdwrite() ), so we 2941 * only need to do this here in the read case. 2942 */ 2943 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) { 2944 vfs_page_set_valid(bp, foff, i, m); 2945 } 2946 vm_page_flag_clear(m, PG_ZERO); 2947 2948 /* 2949 * when debugging new filesystems or buffer I/O methods, this 2950 * is the most common error that pops up. if you see this, you 2951 * have not set the page busy flag correctly!!! 2952 */ 2953 if (m->busy == 0) { 2954 printf("biodone: page busy < 0, " 2955 "pindex: %d, foff: 0x(%x,%x), " 2956 "resid: %d, index: %d\n", 2957 (int) m->pindex, (int)(foff >> 32), 2958 (int) foff & 0xffffffff, resid, i); 2959 if (!vn_isdisk(vp, NULL)) 2960 printf(" iosize: %ld, loffset: %lld, flags: 0x%08x, npages: %d\n", 2961 bp->b_vp->v_mount->mnt_stat.f_iosize, 2962 bp->b_loffset, 2963 bp->b_flags, bp->b_xio.xio_npages); 2964 else 2965 printf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n", 2966 bp->b_loffset, 2967 bp->b_flags, bp->b_xio.xio_npages); 2968 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2969 m->valid, m->dirty, m->wire_count); 2970 panic("biodone: page busy < 0"); 2971 } 2972 vm_page_io_finish(m); 2973 vm_object_pip_subtract(obj, 1); 2974 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2975 iosize -= resid; 2976 } 2977 if (obj) 2978 vm_object_pip_wakeupn(obj, 0); 2979 } 2980 2981 /* 2982 * For asynchronous completions, release the buffer now. The brelse 2983 * will do a wakeup there if necessary - so no need to do a wakeup 2984 * here in the async case. The sync case always needs to do a wakeup. 2985 */ 2986 2987 if (bp->b_flags & B_ASYNC) { 2988 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0) 2989 brelse(bp); 2990 else 2991 bqrelse(bp); 2992 } else { 2993 wakeup(bp); 2994 } 2995 crit_exit(); 2996 } 2997 2998 /* 2999 * vfs_unbusy_pages: 3000 * 3001 * This routine is called in lieu of iodone in the case of 3002 * incomplete I/O. This keeps the busy status for pages 3003 * consistant. 3004 */ 3005 void 3006 vfs_unbusy_pages(struct buf *bp) 3007 { 3008 int i; 3009 3010 runningbufwakeup(bp); 3011 if (bp->b_flags & B_VMIO) { 3012 struct vnode *vp = bp->b_vp; 3013 vm_object_t obj; 3014 3015 obj = vp->v_object; 3016 3017 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3018 vm_page_t m = bp->b_xio.xio_pages[i]; 3019 3020 /* 3021 * When restoring bogus changes the original pages 3022 * should still be wired, so we are in no danger of 3023 * losing the object association and do not need 3024 * critical section protection particularly. 3025 */ 3026 if (m == bogus_page) { 3027 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i); 3028 if (!m) { 3029 panic("vfs_unbusy_pages: page missing"); 3030 } 3031 bp->b_xio.xio_pages[i] = m; 3032 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3033 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3034 } 3035 vm_object_pip_subtract(obj, 1); 3036 vm_page_flag_clear(m, PG_ZERO); 3037 vm_page_io_finish(m); 3038 } 3039 vm_object_pip_wakeupn(obj, 0); 3040 } 3041 } 3042 3043 /* 3044 * vfs_page_set_valid: 3045 * 3046 * Set the valid bits in a page based on the supplied offset. The 3047 * range is restricted to the buffer's size. 3048 * 3049 * This routine is typically called after a read completes. 3050 */ 3051 static void 3052 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3053 { 3054 vm_ooffset_t soff, eoff; 3055 3056 /* 3057 * Start and end offsets in buffer. eoff - soff may not cross a 3058 * page boundry or cross the end of the buffer. The end of the 3059 * buffer, in this case, is our file EOF, not the allocation size 3060 * of the buffer. 3061 */ 3062 soff = off; 3063 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3064 if (eoff > bp->b_loffset + bp->b_bcount) 3065 eoff = bp->b_loffset + bp->b_bcount; 3066 3067 /* 3068 * Set valid range. This is typically the entire buffer and thus the 3069 * entire page. 3070 */ 3071 if (eoff > soff) { 3072 vm_page_set_validclean( 3073 m, 3074 (vm_offset_t) (soff & PAGE_MASK), 3075 (vm_offset_t) (eoff - soff) 3076 ); 3077 } 3078 } 3079 3080 /* 3081 * vfs_busy_pages: 3082 * 3083 * This routine is called before a device strategy routine. 3084 * It is used to tell the VM system that paging I/O is in 3085 * progress, and treat the pages associated with the buffer 3086 * almost as being PG_BUSY. Also the object 'paging_in_progress' 3087 * flag is handled to make sure that the object doesn't become 3088 * inconsistant. 3089 * 3090 * Since I/O has not been initiated yet, certain buffer flags 3091 * such as B_ERROR or B_INVAL may be in an inconsistant state 3092 * and should be ignored. 3093 */ 3094 void 3095 vfs_busy_pages(struct vnode *vp, struct buf *bp) 3096 { 3097 int i, bogus; 3098 struct proc *p = curthread->td_proc; 3099 3100 /* 3101 * The buffer's I/O command must already be set. If reading, 3102 * B_CACHE must be 0 (double check against callers only doing 3103 * I/O when B_CACHE is 0). 3104 */ 3105 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 3106 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0); 3107 3108 if (bp->b_flags & B_VMIO) { 3109 vm_object_t obj; 3110 vm_ooffset_t foff; 3111 3112 obj = vp->v_object; 3113 foff = bp->b_loffset; 3114 KASSERT(bp->b_loffset != NOOFFSET, 3115 ("vfs_busy_pages: no buffer offset")); 3116 vfs_setdirty(bp); 3117 3118 retry: 3119 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3120 vm_page_t m = bp->b_xio.xio_pages[i]; 3121 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 3122 goto retry; 3123 } 3124 3125 bogus = 0; 3126 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3127 vm_page_t m = bp->b_xio.xio_pages[i]; 3128 3129 vm_page_flag_clear(m, PG_ZERO); 3130 if ((bp->b_flags & B_CLUSTER) == 0) { 3131 vm_object_pip_add(obj, 1); 3132 vm_page_io_start(m); 3133 } 3134 3135 /* 3136 * When readying a vnode-backed buffer for a write 3137 * we must zero-fill any invalid portions of the 3138 * backing VM pages. 3139 * 3140 * When readying a vnode-backed buffer for a read 3141 * we must replace any dirty pages with a bogus 3142 * page so we do not destroy dirty data when 3143 * filling in gaps. Dirty pages might not 3144 * necessarily be marked dirty yet, so use m->valid 3145 * as a reasonable test. 3146 * 3147 * Bogus page replacement is, uh, bogus. We need 3148 * to find a better way. 3149 */ 3150 vm_page_protect(m, VM_PROT_NONE); 3151 if (bp->b_cmd == BUF_CMD_WRITE) { 3152 vfs_page_set_valid(bp, foff, i, m); 3153 } else if (m->valid == VM_PAGE_BITS_ALL) { 3154 bp->b_xio.xio_pages[i] = bogus_page; 3155 bogus++; 3156 } 3157 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3158 } 3159 if (bogus) 3160 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3161 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3162 } 3163 3164 /* 3165 * This is the easiest place to put the process accounting for the I/O 3166 * for now. 3167 */ 3168 if (p != NULL) { 3169 if (bp->b_cmd == BUF_CMD_READ) 3170 p->p_stats->p_ru.ru_inblock++; 3171 else 3172 p->p_stats->p_ru.ru_oublock++; 3173 } 3174 } 3175 3176 /* 3177 * vfs_clean_pages: 3178 * 3179 * Tell the VM system that the pages associated with this buffer 3180 * are clean. This is used for delayed writes where the data is 3181 * going to go to disk eventually without additional VM intevention. 3182 * 3183 * Note that while we only really need to clean through to b_bcount, we 3184 * just go ahead and clean through to b_bufsize. 3185 */ 3186 static void 3187 vfs_clean_pages(struct buf *bp) 3188 { 3189 int i; 3190 3191 if (bp->b_flags & B_VMIO) { 3192 vm_ooffset_t foff; 3193 3194 foff = bp->b_loffset; 3195 KASSERT(foff != NOOFFSET, ("vfs_clean_pages: no buffer offset")); 3196 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3197 vm_page_t m = bp->b_xio.xio_pages[i]; 3198 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3199 vm_ooffset_t eoff = noff; 3200 3201 if (eoff > bp->b_loffset + bp->b_bufsize) 3202 eoff = bp->b_loffset + bp->b_bufsize; 3203 vfs_page_set_valid(bp, foff, i, m); 3204 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3205 foff = noff; 3206 } 3207 } 3208 } 3209 3210 /* 3211 * vfs_bio_set_validclean: 3212 * 3213 * Set the range within the buffer to valid and clean. The range is 3214 * relative to the beginning of the buffer, b_loffset. Note that 3215 * b_loffset itself may be offset from the beginning of the first page. 3216 */ 3217 3218 void 3219 vfs_bio_set_validclean(struct buf *bp, int base, int size) 3220 { 3221 if (bp->b_flags & B_VMIO) { 3222 int i; 3223 int n; 3224 3225 /* 3226 * Fixup base to be relative to beginning of first page. 3227 * Set initial n to be the maximum number of bytes in the 3228 * first page that can be validated. 3229 */ 3230 3231 base += (bp->b_loffset & PAGE_MASK); 3232 n = PAGE_SIZE - (base & PAGE_MASK); 3233 3234 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_xio.xio_npages; ++i) { 3235 vm_page_t m = bp->b_xio.xio_pages[i]; 3236 3237 if (n > size) 3238 n = size; 3239 3240 vm_page_set_validclean(m, base & PAGE_MASK, n); 3241 base += n; 3242 size -= n; 3243 n = PAGE_SIZE; 3244 } 3245 } 3246 } 3247 3248 /* 3249 * vfs_bio_clrbuf: 3250 * 3251 * Clear a buffer. This routine essentially fakes an I/O, so we need 3252 * to clear B_ERROR and B_INVAL. 3253 * 3254 * Note that while we only theoretically need to clear through b_bcount, 3255 * we go ahead and clear through b_bufsize. 3256 */ 3257 3258 void 3259 vfs_bio_clrbuf(struct buf *bp) 3260 { 3261 int i, mask = 0; 3262 caddr_t sa, ea; 3263 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3264 bp->b_flags &= ~(B_INVAL|B_ERROR); 3265 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3266 (bp->b_loffset & PAGE_MASK) == 0) { 3267 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3268 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) { 3269 bp->b_resid = 0; 3270 return; 3271 } 3272 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) && 3273 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) { 3274 bzero(bp->b_data, bp->b_bufsize); 3275 bp->b_xio.xio_pages[0]->valid |= mask; 3276 bp->b_resid = 0; 3277 return; 3278 } 3279 } 3280 ea = sa = bp->b_data; 3281 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) { 3282 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3283 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3284 ea = (caddr_t)(vm_offset_t)ulmin( 3285 (u_long)(vm_offset_t)ea, 3286 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3287 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3288 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask) 3289 continue; 3290 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) { 3291 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) { 3292 bzero(sa, ea - sa); 3293 } 3294 } else { 3295 for (; sa < ea; sa += DEV_BSIZE, j++) { 3296 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) && 3297 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0) 3298 bzero(sa, DEV_BSIZE); 3299 } 3300 } 3301 bp->b_xio.xio_pages[i]->valid |= mask; 3302 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO); 3303 } 3304 bp->b_resid = 0; 3305 } else { 3306 clrbuf(bp); 3307 } 3308 } 3309 3310 /* 3311 * vm_hold_load_pages: 3312 * 3313 * Load pages into the buffer's address space. The pages are 3314 * allocated from the kernel object in order to reduce interference 3315 * with the any VM paging I/O activity. The range of loaded 3316 * pages will be wired. 3317 * 3318 * If a page cannot be allocated, the 'pagedaemon' is woken up to 3319 * retrieve the full range (to - from) of pages. 3320 * 3321 */ 3322 void 3323 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3324 { 3325 vm_offset_t pg; 3326 vm_page_t p; 3327 int index; 3328 3329 to = round_page(to); 3330 from = round_page(from); 3331 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3332 3333 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3334 3335 tryagain: 3336 3337 /* 3338 * Note: must allocate system pages since blocking here 3339 * could intefere with paging I/O, no matter which 3340 * process we are. 3341 */ 3342 p = vm_page_alloc(kernel_object, 3343 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3344 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM); 3345 if (!p) { 3346 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3347 vm_wait(); 3348 goto tryagain; 3349 } 3350 vm_page_wire(p); 3351 p->valid = VM_PAGE_BITS_ALL; 3352 vm_page_flag_clear(p, PG_ZERO); 3353 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 3354 bp->b_xio.xio_pages[index] = p; 3355 vm_page_wakeup(p); 3356 } 3357 bp->b_xio.xio_npages = index; 3358 } 3359 3360 /* 3361 * vm_hold_free_pages: 3362 * 3363 * Return pages associated with the buffer back to the VM system. 3364 * 3365 * The range of pages underlying the buffer's address space will 3366 * be unmapped and un-wired. 3367 */ 3368 void 3369 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3370 { 3371 vm_offset_t pg; 3372 vm_page_t p; 3373 int index, newnpages; 3374 3375 from = round_page(from); 3376 to = round_page(to); 3377 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3378 3379 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3380 p = bp->b_xio.xio_pages[index]; 3381 if (p && (index < bp->b_xio.xio_npages)) { 3382 if (p->busy) { 3383 printf("vm_hold_free_pages: doffset: %lld, loffset: %lld\n", 3384 bp->b_bio2.bio_offset, bp->b_loffset); 3385 } 3386 bp->b_xio.xio_pages[index] = NULL; 3387 pmap_kremove(pg); 3388 vm_page_busy(p); 3389 vm_page_unwire(p, 0); 3390 vm_page_free(p); 3391 } 3392 } 3393 bp->b_xio.xio_npages = newnpages; 3394 } 3395 3396 /* 3397 * vmapbuf: 3398 * 3399 * Map a user buffer into KVM via a pbuf. On return the buffer's 3400 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array 3401 * initialized. 3402 */ 3403 int 3404 vmapbuf(struct buf *bp, caddr_t udata, int bytes) 3405 { 3406 caddr_t addr; 3407 vm_paddr_t pa; 3408 int pidx; 3409 int i; 3410 int vmprot; 3411 struct vm_page *m; 3412 3413 /* 3414 * bp had better have a command and it better be a pbuf. 3415 */ 3416 KKASSERT(bp->b_cmd != BUF_CMD_DONE); 3417 KKASSERT(bp->b_flags & B_PAGING); 3418 3419 if (bytes < 0) 3420 return (-1); 3421 3422 /* 3423 * Map the user data into KVM. Mappings have to be page-aligned. 3424 */ 3425 addr = (caddr_t)trunc_page((vm_offset_t)udata); 3426 pidx = 0; 3427 3428 vmprot = VM_PROT_READ; 3429 if (bp->b_cmd == BUF_CMD_READ) 3430 vmprot |= VM_PROT_WRITE; 3431 3432 while (addr < udata + bytes) { 3433 /* 3434 * Do the vm_fault if needed; do the copy-on-write thing 3435 * when reading stuff off device into memory. 3436 */ 3437 retry: 3438 if (addr >= udata) 3439 i = vm_fault_quick(addr, vmprot); 3440 else 3441 i = vm_fault_quick(udata, vmprot); 3442 if (i < 0) { 3443 for (i = 0; i < pidx; ++i) { 3444 vm_page_unhold(bp->b_xio.xio_pages[i]); 3445 bp->b_xio.xio_pages[i] = NULL; 3446 } 3447 return(-1); 3448 } 3449 3450 /* 3451 * Extract from current process's address map. Since the 3452 * fault succeeded, an empty page indicates a race. 3453 */ 3454 pa = pmap_extract(&curproc->p_vmspace->vm_pmap, (vm_offset_t)addr); 3455 if (pa == 0) { 3456 printf("vmapbuf: warning, race against user address during I/O"); 3457 goto retry; 3458 } 3459 m = PHYS_TO_VM_PAGE(pa); 3460 vm_page_hold(m); 3461 bp->b_xio.xio_pages[pidx] = m; 3462 addr += PAGE_SIZE; 3463 ++pidx; 3464 } 3465 3466 /* 3467 * Map the page array and set the buffer fields to point to 3468 * the mapped data buffer. 3469 */ 3470 if (pidx > btoc(MAXPHYS)) 3471 panic("vmapbuf: mapped more than MAXPHYS"); 3472 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx); 3473 3474 bp->b_xio.xio_npages = pidx; 3475 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK); 3476 bp->b_bcount = bytes; 3477 bp->b_bufsize = bytes; 3478 return(0); 3479 } 3480 3481 /* 3482 * vunmapbuf: 3483 * 3484 * Free the io map PTEs associated with this IO operation. 3485 * We also invalidate the TLB entries and restore the original b_addr. 3486 */ 3487 void 3488 vunmapbuf(struct buf *bp) 3489 { 3490 int pidx; 3491 int npages; 3492 3493 KKASSERT(bp->b_flags & B_PAGING); 3494 3495 npages = bp->b_xio.xio_npages; 3496 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3497 for (pidx = 0; pidx < npages; ++pidx) { 3498 vm_page_unhold(bp->b_xio.xio_pages[pidx]); 3499 bp->b_xio.xio_pages[pidx] = NULL; 3500 } 3501 bp->b_xio.xio_npages = 0; 3502 bp->b_data = bp->b_kvabase; 3503 } 3504 3505 /* 3506 * Scan all buffers in the system and issue the callback. 3507 */ 3508 int 3509 scan_all_buffers(int (*callback)(struct buf *, void *), void *info) 3510 { 3511 int count = 0; 3512 int error; 3513 int n; 3514 3515 for (n = 0; n < nbuf; ++n) { 3516 if ((error = callback(&buf[n], info)) < 0) { 3517 count = error; 3518 break; 3519 } 3520 count += error; 3521 } 3522 return (count); 3523 } 3524 3525 /* 3526 * print out statistics from the current status of the buffer pool 3527 * this can be toggeled by the system control option debug.syncprt 3528 */ 3529 #ifdef DEBUG 3530 void 3531 vfs_bufstats(void) 3532 { 3533 int i, j, count; 3534 struct buf *bp; 3535 struct bqueues *dp; 3536 int counts[(MAXBSIZE / PAGE_SIZE) + 1]; 3537 static char *bname[3] = { "LOCKED", "LRU", "AGE" }; 3538 3539 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) { 3540 count = 0; 3541 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++) 3542 counts[j] = 0; 3543 crit_enter(); 3544 TAILQ_FOREACH(bp, dp, b_freelist) { 3545 counts[bp->b_bufsize/PAGE_SIZE]++; 3546 count++; 3547 } 3548 crit_exit(); 3549 printf("%s: total-%d", bname[i], count); 3550 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++) 3551 if (counts[j] != 0) 3552 printf(", %d-%d", j * PAGE_SIZE, counts[j]); 3553 printf("\n"); 3554 } 3555 } 3556 #endif 3557 3558 #ifdef DDB 3559 3560 DB_SHOW_COMMAND(buffer, db_show_buffer) 3561 { 3562 /* get args */ 3563 struct buf *bp = (struct buf *)addr; 3564 3565 if (!have_addr) { 3566 db_printf("usage: show buffer <addr>\n"); 3567 return; 3568 } 3569 3570 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3571 db_printf("b_cmd = %d\n", bp->b_cmd); 3572 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, " 3573 "b_resid = %d\n, b_data = %p, " 3574 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n", 3575 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3576 bp->b_data, bp->b_bio2.bio_offset, (bp->b_bio2.bio_next ? bp->b_bio2.bio_next->bio_offset : (off_t)-1)); 3577 if (bp->b_xio.xio_npages) { 3578 int i; 3579 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ", 3580 bp->b_xio.xio_npages); 3581 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3582 vm_page_t m; 3583 m = bp->b_xio.xio_pages[i]; 3584 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3585 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3586 if ((i + 1) < bp->b_xio.xio_npages) 3587 db_printf(","); 3588 } 3589 db_printf("\n"); 3590 } 3591 } 3592 #endif /* DDB */ 3593