vfs_bio.c revision 112183
1/* 2 * Copyright (c) 1994,1997 John S. Dyson 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice immediately at the beginning of the file, without modification, 10 * this list of conditions, and the following disclaimer. 11 * 2. Absolutely no warranty of function or purpose is made by the author 12 * John S. Dyson. 13 * 14 * $FreeBSD: head/sys/kern/vfs_bio.c 112183 2003-03-13 07:31:45Z jeff $ 15 */ 16 17/* 18 * this file contains a new buffer I/O scheme implementing a coherent 19 * VM object and buffer cache scheme. Pains have been taken to make 20 * sure that the performance degradation associated with schemes such 21 * as this is not realized. 22 * 23 * Author: John S. Dyson 24 * Significant help during the development and debugging phases 25 * had been provided by David Greenman, also of the FreeBSD core team. 26 * 27 * see man buf(9) for more info. 28 */ 29 30#include <sys/param.h> 31#include <sys/systm.h> 32#include <sys/stdint.h> 33#include <sys/bio.h> 34#include <sys/buf.h> 35#include <sys/devicestat.h> 36#include <sys/eventhandler.h> 37#include <sys/lock.h> 38#include <sys/malloc.h> 39#include <sys/mount.h> 40#include <sys/mutex.h> 41#include <sys/kernel.h> 42#include <sys/kthread.h> 43#include <sys/proc.h> 44#include <sys/resourcevar.h> 45#include <sys/sysctl.h> 46#include <sys/vmmeter.h> 47#include <sys/vnode.h> 48#include <vm/vm.h> 49#include <vm/vm_param.h> 50#include <vm/vm_kern.h> 51#include <vm/vm_pageout.h> 52#include <vm/vm_page.h> 53#include <vm/vm_object.h> 54#include <vm/vm_extern.h> 55#include <vm/vm_map.h> 56 57static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 58 59struct bio_ops bioops; /* I/O operation notification */ 60 61struct buf_ops buf_ops_bio = { 62 "buf_ops_bio", 63 bwrite 64}; 65 66/* 67 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 68 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 69 */ 70struct buf *buf; /* buffer header pool */ 71 72static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 73 vm_offset_t to); 74static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 75 vm_offset_t to); 76static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 77 int pageno, vm_page_t m); 78static void vfs_clean_pages(struct buf * bp); 79static void vfs_setdirty(struct buf *bp); 80static void vfs_vmio_release(struct buf *bp); 81static void vfs_backgroundwritedone(struct buf *bp); 82static int vfs_bio_clcheck(struct vnode *vp, int size, 83 daddr_t lblkno, daddr_t blkno); 84static int flushbufqueues(int flushdeps); 85static void buf_daemon(void); 86void bremfreel(struct buf * bp); 87 88int vmiodirenable = TRUE; 89SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 90 "Use the VM system for directory writes"); 91int runningbufspace; 92SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 93 "Amount of presently outstanding async buffer io"); 94static int bufspace; 95SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 96 "KVA memory used for bufs"); 97static int maxbufspace; 98SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 99 "Maximum allowed value of bufspace (including buf_daemon)"); 100static int bufmallocspace; 101SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 102 "Amount of malloced memory for buffers"); 103static int maxbufmallocspace; 104SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 105 "Maximum amount of malloced memory for buffers"); 106static int lobufspace; 107SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 108 "Minimum amount of buffers we want to have"); 109static int hibufspace; 110SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 111 "Maximum allowed value of bufspace (excluding buf_daemon)"); 112static int bufreusecnt; 113SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 114 "Number of times we have reused a buffer"); 115static int buffreekvacnt; 116SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 117 "Number of times we have freed the KVA space from some buffer"); 118static int bufdefragcnt; 119SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 120 "Number of times we have had to repeat buffer allocation to defragment"); 121static int lorunningspace; 122SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 123 "Minimum preferred space used for in-progress I/O"); 124static int hirunningspace; 125SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 126 "Maximum amount of space to use for in-progress I/O"); 127static int dirtybufferflushes; 128SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 129 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 130static int altbufferflushes; 131SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 132 0, "Number of fsync flushes to limit dirty buffers"); 133static int recursiveflushes; 134SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 135 0, "Number of flushes skipped due to being recursive"); 136static int numdirtybuffers; 137SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 138 "Number of buffers that are dirty (has unwritten changes) at the moment"); 139static int lodirtybuffers; 140SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 141 "How many buffers we want to have free before bufdaemon can sleep"); 142static int hidirtybuffers; 143SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 144 "When the number of dirty buffers is considered severe"); 145static int dirtybufthresh; 146SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 147 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 148static int numfreebuffers; 149SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 150 "Number of free buffers"); 151static int lofreebuffers; 152SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 153 "XXX Unused"); 154static int hifreebuffers; 155SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 156 "XXX Complicatedly unused"); 157static int getnewbufcalls; 158SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 159 "Number of calls to getnewbuf"); 160static int getnewbufrestarts; 161SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 162 "Number of times getnewbuf has had to restart a buffer aquisition"); 163static int dobkgrdwrite = 1; 164SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 165 "Do background writes (honoring the BX_BKGRDWRITE flag)?"); 166 167/* 168 * Wakeup point for bufdaemon, as well as indicator of whether it is already 169 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 170 * is idling. 171 */ 172static int bd_request; 173 174/* 175 * This lock synchronizes access to bd_request. 176 */ 177static struct mtx bdlock; 178 179/* 180 * bogus page -- for I/O to/from partially complete buffers 181 * this is a temporary solution to the problem, but it is not 182 * really that bad. it would be better to split the buffer 183 * for input in the case of buffers partially already in memory, 184 * but the code is intricate enough already. 185 */ 186vm_page_t bogus_page; 187 188/* 189 * Synchronization (sleep/wakeup) variable for active buffer space requests. 190 * Set when wait starts, cleared prior to wakeup(). 191 * Used in runningbufwakeup() and waitrunningbufspace(). 192 */ 193static int runningbufreq; 194 195/* 196 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 197 * waitrunningbufspace(). 198 */ 199static struct mtx rbreqlock; 200 201/* 202 * Synchronization (sleep/wakeup) variable for buffer requests. 203 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 204 * by and/or. 205 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 206 * getnewbuf(), and getblk(). 207 */ 208static int needsbuffer; 209 210/* 211 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 212 */ 213static struct mtx nblock; 214 215/* 216 * Lock that protects against bwait()/bdone()/B_DONE races. 217 */ 218 219static struct mtx bdonelock; 220 221/* 222 * Definitions for the buffer free lists. 223 */ 224#define BUFFER_QUEUES 6 /* number of free buffer queues */ 225 226#define QUEUE_NONE 0 /* on no queue */ 227#define QUEUE_LOCKED 1 /* locked buffers */ 228#define QUEUE_CLEAN 2 /* non-B_DELWRI buffers */ 229#define QUEUE_DIRTY 3 /* B_DELWRI buffers */ 230#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 231#define QUEUE_EMPTY 5 /* empty buffer headers */ 232 233/* Queues for free buffers with various properties */ 234static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 235 236/* Lock for the bufqueues */ 237static struct mtx bqlock; 238 239/* 240 * Single global constant for BUF_WMESG, to avoid getting multiple references. 241 * buf_wmesg is referred from macros. 242 */ 243const char *buf_wmesg = BUF_WMESG; 244 245#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 246#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 247#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 248#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 249 250/* 251 * numdirtywakeup: 252 * 253 * If someone is blocked due to there being too many dirty buffers, 254 * and numdirtybuffers is now reasonable, wake them up. 255 */ 256 257static __inline void 258numdirtywakeup(int level) 259{ 260 if (numdirtybuffers <= level) { 261 mtx_lock(&nblock); 262 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 263 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 264 wakeup(&needsbuffer); 265 } 266 mtx_unlock(&nblock); 267 } 268} 269 270/* 271 * bufspacewakeup: 272 * 273 * Called when buffer space is potentially available for recovery. 274 * getnewbuf() will block on this flag when it is unable to free 275 * sufficient buffer space. Buffer space becomes recoverable when 276 * bp's get placed back in the queues. 277 */ 278 279static __inline void 280bufspacewakeup(void) 281{ 282 /* 283 * If someone is waiting for BUF space, wake them up. Even 284 * though we haven't freed the kva space yet, the waiting 285 * process will be able to now. 286 */ 287 mtx_lock(&nblock); 288 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 289 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 290 wakeup(&needsbuffer); 291 } 292 mtx_unlock(&nblock); 293} 294 295/* 296 * runningbufwakeup() - in-progress I/O accounting. 297 * 298 */ 299static __inline void 300runningbufwakeup(struct buf *bp) 301{ 302 if (bp->b_runningbufspace) { 303 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 304 bp->b_runningbufspace = 0; 305 mtx_lock(&rbreqlock); 306 if (runningbufreq && runningbufspace <= lorunningspace) { 307 runningbufreq = 0; 308 wakeup(&runningbufreq); 309 } 310 mtx_unlock(&rbreqlock); 311 } 312} 313 314/* 315 * bufcountwakeup: 316 * 317 * Called when a buffer has been added to one of the free queues to 318 * account for the buffer and to wakeup anyone waiting for free buffers. 319 * This typically occurs when large amounts of metadata are being handled 320 * by the buffer cache ( else buffer space runs out first, usually ). 321 */ 322 323static __inline void 324bufcountwakeup(void) 325{ 326 atomic_add_int(&numfreebuffers, 1); 327 mtx_lock(&nblock); 328 if (needsbuffer) { 329 needsbuffer &= ~VFS_BIO_NEED_ANY; 330 if (numfreebuffers >= hifreebuffers) 331 needsbuffer &= ~VFS_BIO_NEED_FREE; 332 wakeup(&needsbuffer); 333 } 334 mtx_unlock(&nblock); 335} 336 337/* 338 * waitrunningbufspace() 339 * 340 * runningbufspace is a measure of the amount of I/O currently 341 * running. This routine is used in async-write situations to 342 * prevent creating huge backups of pending writes to a device. 343 * Only asynchronous writes are governed by this function. 344 * 345 * Reads will adjust runningbufspace, but will not block based on it. 346 * The read load has a side effect of reducing the allowed write load. 347 * 348 * This does NOT turn an async write into a sync write. It waits 349 * for earlier writes to complete and generally returns before the 350 * caller's write has reached the device. 351 */ 352static __inline void 353waitrunningbufspace(void) 354{ 355 mtx_lock(&rbreqlock); 356 while (runningbufspace > hirunningspace) { 357 ++runningbufreq; 358 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 359 } 360 mtx_unlock(&rbreqlock); 361} 362 363 364/* 365 * vfs_buf_test_cache: 366 * 367 * Called when a buffer is extended. This function clears the B_CACHE 368 * bit if the newly extended portion of the buffer does not contain 369 * valid data. 370 */ 371static __inline__ 372void 373vfs_buf_test_cache(struct buf *bp, 374 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 375 vm_page_t m) 376{ 377 GIANT_REQUIRED; 378 379 if (bp->b_flags & B_CACHE) { 380 int base = (foff + off) & PAGE_MASK; 381 if (vm_page_is_valid(m, base, size) == 0) 382 bp->b_flags &= ~B_CACHE; 383 } 384} 385 386/* Wake up the buffer deamon if necessary */ 387static __inline__ 388void 389bd_wakeup(int dirtybuflevel) 390{ 391 mtx_lock(&bdlock); 392 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 393 bd_request = 1; 394 wakeup(&bd_request); 395 } 396 mtx_unlock(&bdlock); 397} 398 399/* 400 * bd_speedup - speedup the buffer cache flushing code 401 */ 402 403static __inline__ 404void 405bd_speedup(void) 406{ 407 bd_wakeup(1); 408} 409 410/* 411 * Calculating buffer cache scaling values and reserve space for buffer 412 * headers. This is called during low level kernel initialization and 413 * may be called more then once. We CANNOT write to the memory area 414 * being reserved at this time. 415 */ 416caddr_t 417kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 418{ 419 /* 420 * physmem_est is in pages. Convert it to kilobytes (assumes 421 * PAGE_SIZE is >= 1K) 422 */ 423 physmem_est = physmem_est * (PAGE_SIZE / 1024); 424 425 /* 426 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 427 * For the first 64MB of ram nominally allocate sufficient buffers to 428 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 429 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 430 * the buffer cache we limit the eventual kva reservation to 431 * maxbcache bytes. 432 * 433 * factor represents the 1/4 x ram conversion. 434 */ 435 if (nbuf == 0) { 436 int factor = 4 * BKVASIZE / 1024; 437 438 nbuf = 50; 439 if (physmem_est > 4096) 440 nbuf += min((physmem_est - 4096) / factor, 441 65536 / factor); 442 if (physmem_est > 65536) 443 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 444 445 if (maxbcache && nbuf > maxbcache / BKVASIZE) 446 nbuf = maxbcache / BKVASIZE; 447 } 448 449#if 0 450 /* 451 * Do not allow the buffer_map to be more then 1/2 the size of the 452 * kernel_map. 453 */ 454 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 455 (BKVASIZE * 2)) { 456 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 457 (BKVASIZE * 2); 458 printf("Warning: nbufs capped at %d\n", nbuf); 459 } 460#endif 461 462 /* 463 * swbufs are used as temporary holders for I/O, such as paging I/O. 464 * We have no less then 16 and no more then 256. 465 */ 466 nswbuf = max(min(nbuf/4, 256), 16); 467 468 /* 469 * Reserve space for the buffer cache buffers 470 */ 471 swbuf = (void *)v; 472 v = (caddr_t)(swbuf + nswbuf); 473 buf = (void *)v; 474 v = (caddr_t)(buf + nbuf); 475 476 return(v); 477} 478 479/* Initialize the buffer subsystem. Called before use of any buffers. */ 480void 481bufinit(void) 482{ 483 struct buf *bp; 484 vm_offset_t bogus_offset; 485 int i; 486 487 GIANT_REQUIRED; 488 489 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 490 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 491 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 492 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 493 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 494 495 /* next, make a null set of free lists */ 496 for (i = 0; i < BUFFER_QUEUES; i++) 497 TAILQ_INIT(&bufqueues[i]); 498 499 /* finally, initialize each buffer header and stick on empty q */ 500 for (i = 0; i < nbuf; i++) { 501 bp = &buf[i]; 502 bzero(bp, sizeof *bp); 503 bp->b_flags = B_INVAL; /* we're just an empty header */ 504 bp->b_dev = NODEV; 505 bp->b_rcred = NOCRED; 506 bp->b_wcred = NOCRED; 507 bp->b_qindex = QUEUE_EMPTY; 508 bp->b_vflags = 0; 509 bp->b_xflags = 0; 510 LIST_INIT(&bp->b_dep); 511 BUF_LOCKINIT(bp); 512 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 513 } 514 515 /* 516 * maxbufspace is the absolute maximum amount of buffer space we are 517 * allowed to reserve in KVM and in real terms. The absolute maximum 518 * is nominally used by buf_daemon. hibufspace is the nominal maximum 519 * used by most other processes. The differential is required to 520 * ensure that buf_daemon is able to run when other processes might 521 * be blocked waiting for buffer space. 522 * 523 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 524 * this may result in KVM fragmentation which is not handled optimally 525 * by the system. 526 */ 527 maxbufspace = nbuf * BKVASIZE; 528 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 529 lobufspace = hibufspace - MAXBSIZE; 530 531 lorunningspace = 512 * 1024; 532 hirunningspace = 1024 * 1024; 533 534/* 535 * Limit the amount of malloc memory since it is wired permanently into 536 * the kernel space. Even though this is accounted for in the buffer 537 * allocation, we don't want the malloced region to grow uncontrolled. 538 * The malloc scheme improves memory utilization significantly on average 539 * (small) directories. 540 */ 541 maxbufmallocspace = hibufspace / 20; 542 543/* 544 * Reduce the chance of a deadlock occuring by limiting the number 545 * of delayed-write dirty buffers we allow to stack up. 546 */ 547 hidirtybuffers = nbuf / 4 + 20; 548 dirtybufthresh = hidirtybuffers * 9 / 10; 549 numdirtybuffers = 0; 550/* 551 * To support extreme low-memory systems, make sure hidirtybuffers cannot 552 * eat up all available buffer space. This occurs when our minimum cannot 553 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 554 * BKVASIZE'd (8K) buffers. 555 */ 556 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 557 hidirtybuffers >>= 1; 558 } 559 lodirtybuffers = hidirtybuffers / 2; 560 561/* 562 * Try to keep the number of free buffers in the specified range, 563 * and give special processes (e.g. like buf_daemon) access to an 564 * emergency reserve. 565 */ 566 lofreebuffers = nbuf / 18 + 5; 567 hifreebuffers = 2 * lofreebuffers; 568 numfreebuffers = nbuf; 569 570/* 571 * Maximum number of async ops initiated per buf_daemon loop. This is 572 * somewhat of a hack at the moment, we really need to limit ourselves 573 * based on the number of bytes of I/O in-transit that were initiated 574 * from buf_daemon. 575 */ 576 577 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 578 vm_object_lock(kernel_object); 579 bogus_page = vm_page_alloc(kernel_object, 580 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 581 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 582 vm_object_unlock(kernel_object); 583} 584 585/* 586 * bfreekva() - free the kva allocation for a buffer. 587 * 588 * Must be called at splbio() or higher as this is the only locking for 589 * buffer_map. 590 * 591 * Since this call frees up buffer space, we call bufspacewakeup(). 592 */ 593static void 594bfreekva(struct buf * bp) 595{ 596 GIANT_REQUIRED; 597 598 if (bp->b_kvasize) { 599 atomic_add_int(&buffreekvacnt, 1); 600 atomic_subtract_int(&bufspace, bp->b_kvasize); 601 vm_map_delete(buffer_map, 602 (vm_offset_t) bp->b_kvabase, 603 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 604 ); 605 bp->b_kvasize = 0; 606 bufspacewakeup(); 607 } 608} 609 610/* 611 * bremfree: 612 * 613 * Remove the buffer from the appropriate free list. 614 */ 615void 616bremfree(struct buf * bp) 617{ 618 mtx_lock(&bqlock); 619 bremfreel(bp); 620 mtx_unlock(&bqlock); 621} 622 623void 624bremfreel(struct buf * bp) 625{ 626 int s = splbio(); 627 int old_qindex = bp->b_qindex; 628 629 GIANT_REQUIRED; 630 631 if (bp->b_qindex != QUEUE_NONE) { 632 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 633 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 634 bp->b_qindex = QUEUE_NONE; 635 } else { 636 if (BUF_REFCNT(bp) <= 1) 637 panic("bremfree: removing a buffer not on a queue"); 638 } 639 640 /* 641 * Fixup numfreebuffers count. If the buffer is invalid or not 642 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 643 * the buffer was free and we must decrement numfreebuffers. 644 */ 645 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 646 switch(old_qindex) { 647 case QUEUE_DIRTY: 648 case QUEUE_CLEAN: 649 case QUEUE_EMPTY: 650 case QUEUE_EMPTYKVA: 651 atomic_subtract_int(&numfreebuffers, 1); 652 break; 653 default: 654 break; 655 } 656 } 657 splx(s); 658} 659 660 661/* 662 * Get a buffer with the specified data. Look in the cache first. We 663 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 664 * is set, the buffer is valid and we do not have to do anything ( see 665 * getblk() ). This is really just a special case of breadn(). 666 */ 667int 668bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 669 struct buf ** bpp) 670{ 671 672 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 673} 674 675/* 676 * Operates like bread, but also starts asynchronous I/O on 677 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 678 * to initiating I/O . If B_CACHE is set, the buffer is valid 679 * and we do not have to do anything. 680 */ 681int 682breadn(struct vnode * vp, daddr_t blkno, int size, 683 daddr_t * rablkno, int *rabsize, 684 int cnt, struct ucred * cred, struct buf ** bpp) 685{ 686 struct buf *bp, *rabp; 687 int i; 688 int rv = 0, readwait = 0; 689 690 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 691 692 /* if not found in cache, do some I/O */ 693 if ((bp->b_flags & B_CACHE) == 0) { 694 if (curthread != PCPU_GET(idlethread)) 695 curthread->td_proc->p_stats->p_ru.ru_inblock++; 696 bp->b_iocmd = BIO_READ; 697 bp->b_flags &= ~B_INVAL; 698 bp->b_ioflags &= ~BIO_ERROR; 699 if (bp->b_rcred == NOCRED && cred != NOCRED) 700 bp->b_rcred = crhold(cred); 701 vfs_busy_pages(bp, 0); 702 if (vp->v_type == VCHR) 703 VOP_SPECSTRATEGY(vp, bp); 704 else 705 VOP_STRATEGY(vp, bp); 706 ++readwait; 707 } 708 709 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 710 if (inmem(vp, *rablkno)) 711 continue; 712 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 713 714 if ((rabp->b_flags & B_CACHE) == 0) { 715 if (curthread != PCPU_GET(idlethread)) 716 curthread->td_proc->p_stats->p_ru.ru_inblock++; 717 rabp->b_flags |= B_ASYNC; 718 rabp->b_flags &= ~B_INVAL; 719 rabp->b_ioflags &= ~BIO_ERROR; 720 rabp->b_iocmd = BIO_READ; 721 if (rabp->b_rcred == NOCRED && cred != NOCRED) 722 rabp->b_rcred = crhold(cred); 723 vfs_busy_pages(rabp, 0); 724 BUF_KERNPROC(rabp); 725 if (vp->v_type == VCHR) 726 VOP_SPECSTRATEGY(vp, rabp); 727 else 728 VOP_STRATEGY(vp, rabp); 729 } else { 730 brelse(rabp); 731 } 732 } 733 734 if (readwait) { 735 rv = bufwait(bp); 736 } 737 return (rv); 738} 739 740/* 741 * Write, release buffer on completion. (Done by iodone 742 * if async). Do not bother writing anything if the buffer 743 * is invalid. 744 * 745 * Note that we set B_CACHE here, indicating that buffer is 746 * fully valid and thus cacheable. This is true even of NFS 747 * now so we set it generally. This could be set either here 748 * or in biodone() since the I/O is synchronous. We put it 749 * here. 750 */ 751 752int 753bwrite(struct buf * bp) 754{ 755 int oldflags, s; 756 struct buf *newbp; 757 758 if (bp->b_flags & B_INVAL) { 759 brelse(bp); 760 return (0); 761 } 762 763 oldflags = bp->b_flags; 764 765 if (BUF_REFCNT(bp) == 0) 766 panic("bwrite: buffer is not busy???"); 767 s = splbio(); 768 /* 769 * If a background write is already in progress, delay 770 * writing this block if it is asynchronous. Otherwise 771 * wait for the background write to complete. 772 */ 773 if (bp->b_xflags & BX_BKGRDINPROG) { 774 if (bp->b_flags & B_ASYNC) { 775 splx(s); 776 bdwrite(bp); 777 return (0); 778 } 779 bp->b_xflags |= BX_BKGRDWAIT; 780 tsleep(&bp->b_xflags, PRIBIO, "bwrbg", 0); 781 if (bp->b_xflags & BX_BKGRDINPROG) 782 panic("bwrite: still writing"); 783 } 784 785 /* Mark the buffer clean */ 786 bundirty(bp); 787 788 /* 789 * If this buffer is marked for background writing and we 790 * do not have to wait for it, make a copy and write the 791 * copy so as to leave this buffer ready for further use. 792 * 793 * This optimization eats a lot of memory. If we have a page 794 * or buffer shortfall we can't do it. 795 */ 796 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 797 (bp->b_flags & B_ASYNC) && 798 !vm_page_count_severe() && 799 !buf_dirty_count_severe()) { 800 if (bp->b_iodone != NULL) { 801 printf("bp->b_iodone = %p\n", bp->b_iodone); 802 panic("bwrite: need chained iodone"); 803 } 804 805 /* get a new block */ 806 newbp = geteblk(bp->b_bufsize); 807 808 /* 809 * set it to be identical to the old block. We have to 810 * set b_lblkno and BKGRDMARKER before calling bgetvp() 811 * to avoid confusing the splay tree and gbincore(). 812 */ 813 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 814 newbp->b_lblkno = bp->b_lblkno; 815 newbp->b_xflags |= BX_BKGRDMARKER; 816 /* XXX The BX_ flags need to be protected as well */ 817 VI_LOCK(bp->b_vp); 818 bgetvp(bp->b_vp, newbp); 819 VI_UNLOCK(bp->b_vp); 820 newbp->b_blkno = bp->b_blkno; 821 newbp->b_offset = bp->b_offset; 822 newbp->b_iodone = vfs_backgroundwritedone; 823 newbp->b_flags |= B_ASYNC; 824 newbp->b_flags &= ~B_INVAL; 825 826 /* move over the dependencies */ 827 if (LIST_FIRST(&bp->b_dep) != NULL) 828 buf_movedeps(bp, newbp); 829 830 /* 831 * Initiate write on the copy, release the original to 832 * the B_LOCKED queue so that it cannot go away until 833 * the background write completes. If not locked it could go 834 * away and then be reconstituted while it was being written. 835 * If the reconstituted buffer were written, we could end up 836 * with two background copies being written at the same time. 837 */ 838 bp->b_xflags |= BX_BKGRDINPROG; 839 bp->b_flags |= B_LOCKED; 840 bqrelse(bp); 841 bp = newbp; 842 } 843 844 bp->b_flags &= ~B_DONE; 845 bp->b_ioflags &= ~BIO_ERROR; 846 bp->b_flags |= B_WRITEINPROG | B_CACHE; 847 bp->b_iocmd = BIO_WRITE; 848 849 VI_LOCK(bp->b_vp); 850 bp->b_vp->v_numoutput++; 851 VI_UNLOCK(bp->b_vp); 852 vfs_busy_pages(bp, 1); 853 854 /* 855 * Normal bwrites pipeline writes 856 */ 857 bp->b_runningbufspace = bp->b_bufsize; 858 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 859 860 if (curthread != PCPU_GET(idlethread)) 861 curthread->td_proc->p_stats->p_ru.ru_oublock++; 862 splx(s); 863 if (oldflags & B_ASYNC) 864 BUF_KERNPROC(bp); 865 if (bp->b_vp->v_type == VCHR) 866 VOP_SPECSTRATEGY(bp->b_vp, bp); 867 else 868 VOP_STRATEGY(bp->b_vp, bp); 869 870 if ((oldflags & B_ASYNC) == 0) { 871 int rtval = bufwait(bp); 872 brelse(bp); 873 return (rtval); 874 } else if ((oldflags & B_NOWDRAIN) == 0) { 875 /* 876 * don't allow the async write to saturate the I/O 877 * system. Deadlocks can occur only if a device strategy 878 * routine (like in MD) turns around and issues another 879 * high-level write, in which case B_NOWDRAIN is expected 880 * to be set. Otherwise we will not deadlock here because 881 * we are blocking waiting for I/O that is already in-progress 882 * to complete. 883 */ 884 waitrunningbufspace(); 885 } 886 887 return (0); 888} 889 890/* 891 * Complete a background write started from bwrite. 892 */ 893static void 894vfs_backgroundwritedone(bp) 895 struct buf *bp; 896{ 897 struct buf *origbp; 898 899 /* 900 * Find the original buffer that we are writing. 901 */ 902 VI_LOCK(bp->b_vp); 903 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 904 panic("backgroundwritedone: lost buffer"); 905 VI_UNLOCK(bp->b_vp); 906 /* 907 * Process dependencies then return any unfinished ones. 908 */ 909 if (LIST_FIRST(&bp->b_dep) != NULL) 910 buf_complete(bp); 911 if (LIST_FIRST(&bp->b_dep) != NULL) 912 buf_movedeps(bp, origbp); 913 914 /* XXX Find out if origbp can disappear or get inconsistent */ 915 /* 916 * Clear the BX_BKGRDINPROG flag in the original buffer 917 * and awaken it if it is waiting for the write to complete. 918 * If BX_BKGRDINPROG is not set in the original buffer it must 919 * have been released and re-instantiated - which is not legal. 920 */ 921 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), 922 ("backgroundwritedone: lost buffer2")); 923 origbp->b_xflags &= ~BX_BKGRDINPROG; 924 if (origbp->b_xflags & BX_BKGRDWAIT) { 925 origbp->b_xflags &= ~BX_BKGRDWAIT; 926 wakeup(&origbp->b_xflags); 927 } 928 /* 929 * Clear the B_LOCKED flag and remove it from the locked 930 * queue if it currently resides there. 931 */ 932 origbp->b_flags &= ~B_LOCKED; 933 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) { 934 bremfree(origbp); 935 bqrelse(origbp); 936 } 937 /* 938 * This buffer is marked B_NOCACHE, so when it is released 939 * by biodone, it will be tossed. We mark it with BIO_READ 940 * to avoid biodone doing a second vwakeup. 941 */ 942 bp->b_flags |= B_NOCACHE; 943 bp->b_iocmd = BIO_READ; 944 bp->b_flags &= ~(B_CACHE | B_DONE); 945 bp->b_iodone = 0; 946 bufdone(bp); 947} 948 949/* 950 * Delayed write. (Buffer is marked dirty). Do not bother writing 951 * anything if the buffer is marked invalid. 952 * 953 * Note that since the buffer must be completely valid, we can safely 954 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 955 * biodone() in order to prevent getblk from writing the buffer 956 * out synchronously. 957 */ 958void 959bdwrite(struct buf * bp) 960{ 961 struct thread *td = curthread; 962 struct vnode *vp; 963 struct buf *nbp; 964 965 GIANT_REQUIRED; 966 967 if (BUF_REFCNT(bp) == 0) 968 panic("bdwrite: buffer is not busy"); 969 970 if (bp->b_flags & B_INVAL) { 971 brelse(bp); 972 return; 973 } 974 975 /* 976 * If we have too many dirty buffers, don't create any more. 977 * If we are wildly over our limit, then force a complete 978 * cleanup. Otherwise, just keep the situation from getting 979 * out of control. Note that we have to avoid a recursive 980 * disaster and not try to clean up after our own cleanup! 981 */ 982 vp = bp->b_vp; 983 VI_LOCK(vp); 984 if (td->td_proc->p_flag & P_COWINPROGRESS) { 985 recursiveflushes++; 986 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) { 987 VI_UNLOCK(vp); 988 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 989 VI_LOCK(vp); 990 altbufferflushes++; 991 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) { 992 /* 993 * Try to find a buffer to flush. 994 */ 995 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) { 996 if ((nbp->b_xflags & BX_BKGRDINPROG) || 997 buf_countdeps(nbp, 0) || 998 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 999 continue; 1000 if (bp == nbp) 1001 panic("bdwrite: found ourselves"); 1002 VI_UNLOCK(vp); 1003 if (nbp->b_flags & B_CLUSTEROK) { 1004 vfs_bio_awrite(nbp); 1005 } else { 1006 bremfree(nbp); 1007 bawrite(nbp); 1008 } 1009 VI_LOCK(vp); 1010 dirtybufferflushes++; 1011 break; 1012 } 1013 } 1014 VI_UNLOCK(vp); 1015 1016 bdirty(bp); 1017 /* 1018 * Set B_CACHE, indicating that the buffer is fully valid. This is 1019 * true even of NFS now. 1020 */ 1021 bp->b_flags |= B_CACHE; 1022 1023 /* 1024 * This bmap keeps the system from needing to do the bmap later, 1025 * perhaps when the system is attempting to do a sync. Since it 1026 * is likely that the indirect block -- or whatever other datastructure 1027 * that the filesystem needs is still in memory now, it is a good 1028 * thing to do this. Note also, that if the pageout daemon is 1029 * requesting a sync -- there might not be enough memory to do 1030 * the bmap then... So, this is important to do. 1031 */ 1032 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1033 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1034 } 1035 1036 /* 1037 * Set the *dirty* buffer range based upon the VM system dirty pages. 1038 */ 1039 vfs_setdirty(bp); 1040 1041 /* 1042 * We need to do this here to satisfy the vnode_pager and the 1043 * pageout daemon, so that it thinks that the pages have been 1044 * "cleaned". Note that since the pages are in a delayed write 1045 * buffer -- the VFS layer "will" see that the pages get written 1046 * out on the next sync, or perhaps the cluster will be completed. 1047 */ 1048 vfs_clean_pages(bp); 1049 bqrelse(bp); 1050 1051 /* 1052 * Wakeup the buffer flushing daemon if we have a lot of dirty 1053 * buffers (midpoint between our recovery point and our stall 1054 * point). 1055 */ 1056 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1057 1058 /* 1059 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1060 * due to the softdep code. 1061 */ 1062} 1063 1064/* 1065 * bdirty: 1066 * 1067 * Turn buffer into delayed write request. We must clear BIO_READ and 1068 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1069 * itself to properly update it in the dirty/clean lists. We mark it 1070 * B_DONE to ensure that any asynchronization of the buffer properly 1071 * clears B_DONE ( else a panic will occur later ). 1072 * 1073 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1074 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1075 * should only be called if the buffer is known-good. 1076 * 1077 * Since the buffer is not on a queue, we do not update the numfreebuffers 1078 * count. 1079 * 1080 * Must be called at splbio(). 1081 * The buffer must be on QUEUE_NONE. 1082 */ 1083void 1084bdirty(bp) 1085 struct buf *bp; 1086{ 1087 KASSERT(bp->b_qindex == QUEUE_NONE, 1088 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1089 bp->b_flags &= ~(B_RELBUF); 1090 bp->b_iocmd = BIO_WRITE; 1091 1092 if ((bp->b_flags & B_DELWRI) == 0) { 1093 bp->b_flags |= B_DONE | B_DELWRI; 1094 reassignbuf(bp, bp->b_vp); 1095 atomic_add_int(&numdirtybuffers, 1); 1096 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1097 } 1098} 1099 1100/* 1101 * bundirty: 1102 * 1103 * Clear B_DELWRI for buffer. 1104 * 1105 * Since the buffer is not on a queue, we do not update the numfreebuffers 1106 * count. 1107 * 1108 * Must be called at splbio(). 1109 * The buffer must be on QUEUE_NONE. 1110 */ 1111 1112void 1113bundirty(bp) 1114 struct buf *bp; 1115{ 1116 KASSERT(bp->b_qindex == QUEUE_NONE, 1117 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1118 1119 if (bp->b_flags & B_DELWRI) { 1120 bp->b_flags &= ~B_DELWRI; 1121 reassignbuf(bp, bp->b_vp); 1122 atomic_subtract_int(&numdirtybuffers, 1); 1123 numdirtywakeup(lodirtybuffers); 1124 } 1125 /* 1126 * Since it is now being written, we can clear its deferred write flag. 1127 */ 1128 bp->b_flags &= ~B_DEFERRED; 1129} 1130 1131/* 1132 * bawrite: 1133 * 1134 * Asynchronous write. Start output on a buffer, but do not wait for 1135 * it to complete. The buffer is released when the output completes. 1136 * 1137 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1138 * B_INVAL buffers. Not us. 1139 */ 1140void 1141bawrite(struct buf * bp) 1142{ 1143 bp->b_flags |= B_ASYNC; 1144 (void) BUF_WRITE(bp); 1145} 1146 1147/* 1148 * bwillwrite: 1149 * 1150 * Called prior to the locking of any vnodes when we are expecting to 1151 * write. We do not want to starve the buffer cache with too many 1152 * dirty buffers so we block here. By blocking prior to the locking 1153 * of any vnodes we attempt to avoid the situation where a locked vnode 1154 * prevents the various system daemons from flushing related buffers. 1155 */ 1156 1157void 1158bwillwrite(void) 1159{ 1160 if (numdirtybuffers >= hidirtybuffers) { 1161 int s; 1162 1163 mtx_lock(&Giant); 1164 s = splbio(); 1165 mtx_lock(&nblock); 1166 while (numdirtybuffers >= hidirtybuffers) { 1167 bd_wakeup(1); 1168 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1169 msleep(&needsbuffer, &nblock, 1170 (PRIBIO + 4), "flswai", 0); 1171 } 1172 splx(s); 1173 mtx_unlock(&nblock); 1174 mtx_unlock(&Giant); 1175 } 1176} 1177 1178/* 1179 * Return true if we have too many dirty buffers. 1180 */ 1181int 1182buf_dirty_count_severe(void) 1183{ 1184 return(numdirtybuffers >= hidirtybuffers); 1185} 1186 1187/* 1188 * brelse: 1189 * 1190 * Release a busy buffer and, if requested, free its resources. The 1191 * buffer will be stashed in the appropriate bufqueue[] allowing it 1192 * to be accessed later as a cache entity or reused for other purposes. 1193 */ 1194void 1195brelse(struct buf * bp) 1196{ 1197 int s; 1198 1199 GIANT_REQUIRED; 1200 1201 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1202 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1203 1204 s = splbio(); 1205 1206 if (bp->b_flags & B_LOCKED) 1207 bp->b_ioflags &= ~BIO_ERROR; 1208 1209 if (bp->b_iocmd == BIO_WRITE && 1210 (bp->b_ioflags & BIO_ERROR) && 1211 !(bp->b_flags & B_INVAL)) { 1212 /* 1213 * Failed write, redirty. Must clear BIO_ERROR to prevent 1214 * pages from being scrapped. If B_INVAL is set then 1215 * this case is not run and the next case is run to 1216 * destroy the buffer. B_INVAL can occur if the buffer 1217 * is outside the range supported by the underlying device. 1218 */ 1219 bp->b_ioflags &= ~BIO_ERROR; 1220 bdirty(bp); 1221 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1222 (bp->b_ioflags & BIO_ERROR) || 1223 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1224 /* 1225 * Either a failed I/O or we were asked to free or not 1226 * cache the buffer. 1227 */ 1228 bp->b_flags |= B_INVAL; 1229 if (LIST_FIRST(&bp->b_dep) != NULL) 1230 buf_deallocate(bp); 1231 if (bp->b_flags & B_DELWRI) { 1232 atomic_subtract_int(&numdirtybuffers, 1); 1233 numdirtywakeup(lodirtybuffers); 1234 } 1235 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1236 if ((bp->b_flags & B_VMIO) == 0) { 1237 if (bp->b_bufsize) 1238 allocbuf(bp, 0); 1239 if (bp->b_vp) 1240 brelvp(bp); 1241 } 1242 } 1243 1244 /* 1245 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1246 * is called with B_DELWRI set, the underlying pages may wind up 1247 * getting freed causing a previous write (bdwrite()) to get 'lost' 1248 * because pages associated with a B_DELWRI bp are marked clean. 1249 * 1250 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1251 * if B_DELWRI is set. 1252 * 1253 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1254 * on pages to return pages to the VM page queues. 1255 */ 1256 if (bp->b_flags & B_DELWRI) 1257 bp->b_flags &= ~B_RELBUF; 1258 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1259 bp->b_flags |= B_RELBUF; 1260 1261 /* 1262 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1263 * constituted, not even NFS buffers now. Two flags effect this. If 1264 * B_INVAL, the struct buf is invalidated but the VM object is kept 1265 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1266 * 1267 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1268 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1269 * buffer is also B_INVAL because it hits the re-dirtying code above. 1270 * 1271 * Normally we can do this whether a buffer is B_DELWRI or not. If 1272 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1273 * the commit state and we cannot afford to lose the buffer. If the 1274 * buffer has a background write in progress, we need to keep it 1275 * around to prevent it from being reconstituted and starting a second 1276 * background write. 1277 */ 1278 if ((bp->b_flags & B_VMIO) 1279 && !(bp->b_vp->v_mount != NULL && 1280 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1281 !vn_isdisk(bp->b_vp, NULL) && 1282 (bp->b_flags & B_DELWRI)) 1283 ) { 1284 1285 int i, j, resid; 1286 vm_page_t m; 1287 off_t foff; 1288 vm_pindex_t poff; 1289 vm_object_t obj; 1290 struct vnode *vp; 1291 1292 vp = bp->b_vp; 1293 obj = bp->b_object; 1294 1295 /* 1296 * Get the base offset and length of the buffer. Note that 1297 * in the VMIO case if the buffer block size is not 1298 * page-aligned then b_data pointer may not be page-aligned. 1299 * But our b_pages[] array *IS* page aligned. 1300 * 1301 * block sizes less then DEV_BSIZE (usually 512) are not 1302 * supported due to the page granularity bits (m->valid, 1303 * m->dirty, etc...). 1304 * 1305 * See man buf(9) for more information 1306 */ 1307 resid = bp->b_bufsize; 1308 foff = bp->b_offset; 1309 1310 for (i = 0; i < bp->b_npages; i++) { 1311 int had_bogus = 0; 1312 1313 m = bp->b_pages[i]; 1314 vm_page_lock_queues(); 1315 vm_page_flag_clear(m, PG_ZERO); 1316 vm_page_unlock_queues(); 1317 1318 /* 1319 * If we hit a bogus page, fixup *all* the bogus pages 1320 * now. 1321 */ 1322 if (m == bogus_page) { 1323 poff = OFF_TO_IDX(bp->b_offset); 1324 had_bogus = 1; 1325 1326 for (j = i; j < bp->b_npages; j++) { 1327 vm_page_t mtmp; 1328 mtmp = bp->b_pages[j]; 1329 if (mtmp == bogus_page) { 1330 mtmp = vm_page_lookup(obj, poff + j); 1331 if (!mtmp) { 1332 panic("brelse: page missing\n"); 1333 } 1334 bp->b_pages[j] = mtmp; 1335 } 1336 } 1337 1338 if ((bp->b_flags & B_INVAL) == 0) { 1339 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1340 } 1341 m = bp->b_pages[i]; 1342 } 1343 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1344 int poffset = foff & PAGE_MASK; 1345 int presid = resid > (PAGE_SIZE - poffset) ? 1346 (PAGE_SIZE - poffset) : resid; 1347 1348 KASSERT(presid >= 0, ("brelse: extra page")); 1349 vm_page_set_invalid(m, poffset, presid); 1350 if (had_bogus) 1351 printf("avoided corruption bug in bogus_page/brelse code\n"); 1352 } 1353 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1354 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1355 } 1356 1357 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1358 vfs_vmio_release(bp); 1359 1360 } else if (bp->b_flags & B_VMIO) { 1361 1362 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1363 vfs_vmio_release(bp); 1364 } 1365 1366 } 1367 1368 if (bp->b_qindex != QUEUE_NONE) 1369 panic("brelse: free buffer onto another queue???"); 1370 if (BUF_REFCNT(bp) > 1) { 1371 /* do not release to free list */ 1372 BUF_UNLOCK(bp); 1373 splx(s); 1374 return; 1375 } 1376 1377 /* enqueue */ 1378 mtx_lock(&bqlock); 1379 1380 /* buffers with no memory */ 1381 if (bp->b_bufsize == 0) { 1382 bp->b_flags |= B_INVAL; 1383 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1384 if (bp->b_xflags & BX_BKGRDINPROG) 1385 panic("losing buffer 1"); 1386 if (bp->b_kvasize) { 1387 bp->b_qindex = QUEUE_EMPTYKVA; 1388 } else { 1389 bp->b_qindex = QUEUE_EMPTY; 1390 } 1391 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1392 bp->b_dev = NODEV; 1393 /* buffers with junk contents */ 1394 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1395 (bp->b_ioflags & BIO_ERROR)) { 1396 bp->b_flags |= B_INVAL; 1397 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1398 if (bp->b_xflags & BX_BKGRDINPROG) 1399 panic("losing buffer 2"); 1400 bp->b_qindex = QUEUE_CLEAN; 1401 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1402 bp->b_dev = NODEV; 1403 1404 /* buffers that are locked */ 1405 } else if (bp->b_flags & B_LOCKED) { 1406 bp->b_qindex = QUEUE_LOCKED; 1407 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1408 1409 /* remaining buffers */ 1410 } else { 1411 if (bp->b_flags & B_DELWRI) 1412 bp->b_qindex = QUEUE_DIRTY; 1413 else 1414 bp->b_qindex = QUEUE_CLEAN; 1415 if (bp->b_flags & B_AGE) 1416 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1417 else 1418 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1419 } 1420 mtx_unlock(&bqlock); 1421 1422 /* 1423 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1424 * placed the buffer on the correct queue. We must also disassociate 1425 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1426 * find it. 1427 */ 1428 if (bp->b_flags & B_INVAL) { 1429 if (bp->b_flags & B_DELWRI) 1430 bundirty(bp); 1431 if (bp->b_vp) 1432 brelvp(bp); 1433 } 1434 1435 /* 1436 * Fixup numfreebuffers count. The bp is on an appropriate queue 1437 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1438 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1439 * if B_INVAL is set ). 1440 */ 1441 1442 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1443 bufcountwakeup(); 1444 1445 /* 1446 * Something we can maybe free or reuse 1447 */ 1448 if (bp->b_bufsize || bp->b_kvasize) 1449 bufspacewakeup(); 1450 1451 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | 1452 B_DIRECT | B_NOWDRAIN); 1453 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1454 panic("brelse: not dirty"); 1455 /* unlock */ 1456 BUF_UNLOCK(bp); 1457 splx(s); 1458} 1459 1460/* 1461 * Release a buffer back to the appropriate queue but do not try to free 1462 * it. The buffer is expected to be used again soon. 1463 * 1464 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1465 * biodone() to requeue an async I/O on completion. It is also used when 1466 * known good buffers need to be requeued but we think we may need the data 1467 * again soon. 1468 * 1469 * XXX we should be able to leave the B_RELBUF hint set on completion. 1470 */ 1471void 1472bqrelse(struct buf * bp) 1473{ 1474 int s; 1475 1476 s = splbio(); 1477 1478 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1479 1480 if (bp->b_qindex != QUEUE_NONE) 1481 panic("bqrelse: free buffer onto another queue???"); 1482 if (BUF_REFCNT(bp) > 1) { 1483 /* do not release to free list */ 1484 BUF_UNLOCK(bp); 1485 splx(s); 1486 return; 1487 } 1488 mtx_lock(&bqlock); 1489 if (bp->b_flags & B_LOCKED) { 1490 bp->b_ioflags &= ~BIO_ERROR; 1491 bp->b_qindex = QUEUE_LOCKED; 1492 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1493 /* buffers with stale but valid contents */ 1494 } else if (bp->b_flags & B_DELWRI) { 1495 bp->b_qindex = QUEUE_DIRTY; 1496 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1497 } else if (vm_page_count_severe()) { 1498 /* 1499 * We are too low on memory, we have to try to free the 1500 * buffer (most importantly: the wired pages making up its 1501 * backing store) *now*. 1502 */ 1503 mtx_unlock(&bqlock); 1504 splx(s); 1505 brelse(bp); 1506 return; 1507 } else { 1508 bp->b_qindex = QUEUE_CLEAN; 1509 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1510 } 1511 mtx_unlock(&bqlock); 1512 1513 if ((bp->b_flags & B_LOCKED) == 0 && 1514 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1515 bufcountwakeup(); 1516 } 1517 1518 /* 1519 * Something we can maybe free or reuse. 1520 */ 1521 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1522 bufspacewakeup(); 1523 1524 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1525 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1526 panic("bqrelse: not dirty"); 1527 /* unlock */ 1528 BUF_UNLOCK(bp); 1529 splx(s); 1530} 1531 1532/* Give pages used by the bp back to the VM system (where possible) */ 1533static void 1534vfs_vmio_release(bp) 1535 struct buf *bp; 1536{ 1537 int i; 1538 vm_page_t m; 1539 1540 GIANT_REQUIRED; 1541 vm_page_lock_queues(); 1542 for (i = 0; i < bp->b_npages; i++) { 1543 m = bp->b_pages[i]; 1544 bp->b_pages[i] = NULL; 1545 /* 1546 * In order to keep page LRU ordering consistent, put 1547 * everything on the inactive queue. 1548 */ 1549 vm_page_unwire(m, 0); 1550 /* 1551 * We don't mess with busy pages, it is 1552 * the responsibility of the process that 1553 * busied the pages to deal with them. 1554 */ 1555 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1556 continue; 1557 1558 if (m->wire_count == 0) { 1559 vm_page_flag_clear(m, PG_ZERO); 1560 /* 1561 * Might as well free the page if we can and it has 1562 * no valid data. We also free the page if the 1563 * buffer was used for direct I/O 1564 */ 1565 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1566 m->hold_count == 0) { 1567 vm_page_busy(m); 1568 pmap_remove_all(m); 1569 vm_page_free(m); 1570 } else if (bp->b_flags & B_DIRECT) { 1571 vm_page_try_to_free(m); 1572 } else if (vm_page_count_severe()) { 1573 vm_page_try_to_cache(m); 1574 } 1575 } 1576 } 1577 vm_page_unlock_queues(); 1578 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1579 1580 if (bp->b_bufsize) { 1581 bufspacewakeup(); 1582 bp->b_bufsize = 0; 1583 } 1584 bp->b_npages = 0; 1585 bp->b_flags &= ~B_VMIO; 1586 if (bp->b_vp) 1587 brelvp(bp); 1588} 1589 1590/* 1591 * Check to see if a block at a particular lbn is available for a clustered 1592 * write. 1593 */ 1594static int 1595vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1596{ 1597 struct buf *bpa; 1598 int match; 1599 1600 match = 0; 1601 1602 /* If the buf isn't in core skip it */ 1603 if ((bpa = gbincore(vp, lblkno)) == NULL) 1604 return (0); 1605 1606 /* If the buf is busy we don't want to wait for it */ 1607 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1608 return (0); 1609 1610 /* Only cluster with valid clusterable delayed write buffers */ 1611 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1612 (B_DELWRI | B_CLUSTEROK)) 1613 goto done; 1614 1615 if (bpa->b_bufsize != size) 1616 goto done; 1617 1618 /* 1619 * Check to see if it is in the expected place on disk and that the 1620 * block has been mapped. 1621 */ 1622 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1623 match = 1; 1624done: 1625 BUF_UNLOCK(bpa); 1626 return (match); 1627} 1628 1629/* 1630 * vfs_bio_awrite: 1631 * 1632 * Implement clustered async writes for clearing out B_DELWRI buffers. 1633 * This is much better then the old way of writing only one buffer at 1634 * a time. Note that we may not be presented with the buffers in the 1635 * correct order, so we search for the cluster in both directions. 1636 */ 1637int 1638vfs_bio_awrite(struct buf * bp) 1639{ 1640 int i; 1641 int j; 1642 daddr_t lblkno = bp->b_lblkno; 1643 struct vnode *vp = bp->b_vp; 1644 int s; 1645 int ncl; 1646 int nwritten; 1647 int size; 1648 int maxcl; 1649 1650 s = splbio(); 1651 /* 1652 * right now we support clustered writing only to regular files. If 1653 * we find a clusterable block we could be in the middle of a cluster 1654 * rather then at the beginning. 1655 */ 1656 if ((vp->v_type == VREG) && 1657 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1658 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1659 1660 size = vp->v_mount->mnt_stat.f_iosize; 1661 maxcl = MAXPHYS / size; 1662 1663 VI_LOCK(vp); 1664 for (i = 1; i < maxcl; i++) 1665 if (vfs_bio_clcheck(vp, size, lblkno + i, 1666 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1667 break; 1668 1669 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1670 if (vfs_bio_clcheck(vp, size, lblkno - j, 1671 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1672 break; 1673 1674 VI_UNLOCK(vp); 1675 --j; 1676 ncl = i + j; 1677 /* 1678 * this is a possible cluster write 1679 */ 1680 if (ncl != 1) { 1681 BUF_UNLOCK(bp); 1682 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1683 splx(s); 1684 return nwritten; 1685 } 1686 } 1687 1688 bremfree(bp); 1689 bp->b_flags |= B_ASYNC; 1690 1691 splx(s); 1692 /* 1693 * default (old) behavior, writing out only one block 1694 * 1695 * XXX returns b_bufsize instead of b_bcount for nwritten? 1696 */ 1697 nwritten = bp->b_bufsize; 1698 (void) BUF_WRITE(bp); 1699 1700 return nwritten; 1701} 1702 1703/* 1704 * getnewbuf: 1705 * 1706 * Find and initialize a new buffer header, freeing up existing buffers 1707 * in the bufqueues as necessary. The new buffer is returned locked. 1708 * 1709 * Important: B_INVAL is not set. If the caller wishes to throw the 1710 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1711 * 1712 * We block if: 1713 * We have insufficient buffer headers 1714 * We have insufficient buffer space 1715 * buffer_map is too fragmented ( space reservation fails ) 1716 * If we have to flush dirty buffers ( but we try to avoid this ) 1717 * 1718 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1719 * Instead we ask the buf daemon to do it for us. We attempt to 1720 * avoid piecemeal wakeups of the pageout daemon. 1721 */ 1722 1723static struct buf * 1724getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1725{ 1726 struct buf *bp; 1727 struct buf *nbp; 1728 int defrag = 0; 1729 int nqindex; 1730 static int flushingbufs; 1731 1732 GIANT_REQUIRED; 1733 1734 /* 1735 * We can't afford to block since we might be holding a vnode lock, 1736 * which may prevent system daemons from running. We deal with 1737 * low-memory situations by proactively returning memory and running 1738 * async I/O rather then sync I/O. 1739 */ 1740 1741 atomic_add_int(&getnewbufcalls, 1); 1742 atomic_subtract_int(&getnewbufrestarts, 1); 1743restart: 1744 atomic_add_int(&getnewbufrestarts, 1); 1745 1746 /* 1747 * Setup for scan. If we do not have enough free buffers, 1748 * we setup a degenerate case that immediately fails. Note 1749 * that if we are specially marked process, we are allowed to 1750 * dip into our reserves. 1751 * 1752 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1753 * 1754 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1755 * However, there are a number of cases (defragging, reusing, ...) 1756 * where we cannot backup. 1757 */ 1758 mtx_lock(&bqlock); 1759 nqindex = QUEUE_EMPTYKVA; 1760 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1761 1762 if (nbp == NULL) { 1763 /* 1764 * If no EMPTYKVA buffers and we are either 1765 * defragging or reusing, locate a CLEAN buffer 1766 * to free or reuse. If bufspace useage is low 1767 * skip this step so we can allocate a new buffer. 1768 */ 1769 if (defrag || bufspace >= lobufspace) { 1770 nqindex = QUEUE_CLEAN; 1771 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1772 } 1773 1774 /* 1775 * If we could not find or were not allowed to reuse a 1776 * CLEAN buffer, check to see if it is ok to use an EMPTY 1777 * buffer. We can only use an EMPTY buffer if allocating 1778 * its KVA would not otherwise run us out of buffer space. 1779 */ 1780 if (nbp == NULL && defrag == 0 && 1781 bufspace + maxsize < hibufspace) { 1782 nqindex = QUEUE_EMPTY; 1783 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1784 } 1785 } 1786 1787 /* 1788 * Run scan, possibly freeing data and/or kva mappings on the fly 1789 * depending. 1790 */ 1791 1792 while ((bp = nbp) != NULL) { 1793 int qindex = nqindex; 1794 1795 /* 1796 * Calculate next bp ( we can only use it if we do not block 1797 * or do other fancy things ). 1798 */ 1799 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1800 switch(qindex) { 1801 case QUEUE_EMPTY: 1802 nqindex = QUEUE_EMPTYKVA; 1803 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1804 break; 1805 /* FALLTHROUGH */ 1806 case QUEUE_EMPTYKVA: 1807 nqindex = QUEUE_CLEAN; 1808 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1809 break; 1810 /* FALLTHROUGH */ 1811 case QUEUE_CLEAN: 1812 /* 1813 * nbp is NULL. 1814 */ 1815 break; 1816 } 1817 } 1818 1819 /* 1820 * Sanity Checks 1821 */ 1822 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1823 1824 /* 1825 * Note: we no longer distinguish between VMIO and non-VMIO 1826 * buffers. 1827 */ 1828 1829 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1830 1831 /* 1832 * If we are defragging then we need a buffer with 1833 * b_kvasize != 0. XXX this situation should no longer 1834 * occur, if defrag is non-zero the buffer's b_kvasize 1835 * should also be non-zero at this point. XXX 1836 */ 1837 if (defrag && bp->b_kvasize == 0) { 1838 printf("Warning: defrag empty buffer %p\n", bp); 1839 continue; 1840 } 1841 1842 /* 1843 * Start freeing the bp. This is somewhat involved. nbp 1844 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1845 */ 1846 1847 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1848 panic("getnewbuf: locked buf"); 1849 bremfreel(bp); 1850 mtx_unlock(&bqlock); 1851 1852 if (qindex == QUEUE_CLEAN) { 1853 if (bp->b_flags & B_VMIO) { 1854 bp->b_flags &= ~B_ASYNC; 1855 vfs_vmio_release(bp); 1856 } 1857 if (bp->b_vp) 1858 brelvp(bp); 1859 } 1860 1861 /* 1862 * NOTE: nbp is now entirely invalid. We can only restart 1863 * the scan from this point on. 1864 * 1865 * Get the rest of the buffer freed up. b_kva* is still 1866 * valid after this operation. 1867 */ 1868 1869 if (bp->b_rcred != NOCRED) { 1870 crfree(bp->b_rcred); 1871 bp->b_rcred = NOCRED; 1872 } 1873 if (bp->b_wcred != NOCRED) { 1874 crfree(bp->b_wcred); 1875 bp->b_wcred = NOCRED; 1876 } 1877 if (LIST_FIRST(&bp->b_dep) != NULL) 1878 buf_deallocate(bp); 1879 if (bp->b_xflags & BX_BKGRDINPROG) 1880 panic("losing buffer 3"); 1881 1882 if (bp->b_bufsize) 1883 allocbuf(bp, 0); 1884 1885 bp->b_flags = 0; 1886 bp->b_ioflags = 0; 1887 bp->b_xflags = 0; 1888 bp->b_vflags = 0; 1889 bp->b_dev = NODEV; 1890 bp->b_vp = NULL; 1891 bp->b_blkno = bp->b_lblkno = 0; 1892 bp->b_offset = NOOFFSET; 1893 bp->b_iodone = 0; 1894 bp->b_error = 0; 1895 bp->b_resid = 0; 1896 bp->b_bcount = 0; 1897 bp->b_npages = 0; 1898 bp->b_dirtyoff = bp->b_dirtyend = 0; 1899 bp->b_magic = B_MAGIC_BIO; 1900 bp->b_op = &buf_ops_bio; 1901 bp->b_object = NULL; 1902 1903 LIST_INIT(&bp->b_dep); 1904 1905 /* 1906 * If we are defragging then free the buffer. 1907 */ 1908 if (defrag) { 1909 bp->b_flags |= B_INVAL; 1910 bfreekva(bp); 1911 brelse(bp); 1912 defrag = 0; 1913 goto restart; 1914 } 1915 1916 /* 1917 * If we are overcomitted then recover the buffer and its 1918 * KVM space. This occurs in rare situations when multiple 1919 * processes are blocked in getnewbuf() or allocbuf(). 1920 */ 1921 if (bufspace >= hibufspace) 1922 flushingbufs = 1; 1923 if (flushingbufs && bp->b_kvasize != 0) { 1924 bp->b_flags |= B_INVAL; 1925 bfreekva(bp); 1926 brelse(bp); 1927 goto restart; 1928 } 1929 if (bufspace < lobufspace) 1930 flushingbufs = 0; 1931 break; 1932 } 1933 1934 /* 1935 * If we exhausted our list, sleep as appropriate. We may have to 1936 * wakeup various daemons and write out some dirty buffers. 1937 * 1938 * Generally we are sleeping due to insufficient buffer space. 1939 */ 1940 1941 if (bp == NULL) { 1942 int flags; 1943 char *waitmsg; 1944 1945 mtx_unlock(&bqlock); 1946 if (defrag) { 1947 flags = VFS_BIO_NEED_BUFSPACE; 1948 waitmsg = "nbufkv"; 1949 } else if (bufspace >= hibufspace) { 1950 waitmsg = "nbufbs"; 1951 flags = VFS_BIO_NEED_BUFSPACE; 1952 } else { 1953 waitmsg = "newbuf"; 1954 flags = VFS_BIO_NEED_ANY; 1955 } 1956 1957 bd_speedup(); /* heeeelp */ 1958 1959 mtx_lock(&nblock); 1960 needsbuffer |= flags; 1961 while (needsbuffer & flags) { 1962 if (msleep(&needsbuffer, &nblock, 1963 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1964 mtx_unlock(&nblock); 1965 return (NULL); 1966 } 1967 } 1968 mtx_unlock(&nblock); 1969 } else { 1970 /* 1971 * We finally have a valid bp. We aren't quite out of the 1972 * woods, we still have to reserve kva space. In order 1973 * to keep fragmentation sane we only allocate kva in 1974 * BKVASIZE chunks. 1975 */ 1976 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1977 1978 if (maxsize != bp->b_kvasize) { 1979 vm_offset_t addr = 0; 1980 1981 bfreekva(bp); 1982 1983 if (vm_map_findspace(buffer_map, 1984 vm_map_min(buffer_map), maxsize, &addr)) { 1985 /* 1986 * Uh oh. Buffer map is to fragmented. We 1987 * must defragment the map. 1988 */ 1989 atomic_add_int(&bufdefragcnt, 1); 1990 defrag = 1; 1991 bp->b_flags |= B_INVAL; 1992 brelse(bp); 1993 goto restart; 1994 } 1995 if (addr) { 1996 vm_map_insert(buffer_map, NULL, 0, 1997 addr, addr + maxsize, 1998 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1999 2000 bp->b_kvabase = (caddr_t) addr; 2001 bp->b_kvasize = maxsize; 2002 atomic_add_int(&bufspace, bp->b_kvasize); 2003 atomic_add_int(&bufreusecnt, 1); 2004 } 2005 } 2006 bp->b_data = bp->b_kvabase; 2007 } 2008 return(bp); 2009} 2010 2011/* 2012 * buf_daemon: 2013 * 2014 * buffer flushing daemon. Buffers are normally flushed by the 2015 * update daemon but if it cannot keep up this process starts to 2016 * take the load in an attempt to prevent getnewbuf() from blocking. 2017 */ 2018 2019static struct proc *bufdaemonproc; 2020 2021static struct kproc_desc buf_kp = { 2022 "bufdaemon", 2023 buf_daemon, 2024 &bufdaemonproc 2025}; 2026SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2027 2028static void 2029buf_daemon() 2030{ 2031 int s; 2032 2033 mtx_lock(&Giant); 2034 2035 /* 2036 * This process needs to be suspended prior to shutdown sync. 2037 */ 2038 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2039 SHUTDOWN_PRI_LAST); 2040 2041 /* 2042 * This process is allowed to take the buffer cache to the limit 2043 */ 2044 s = splbio(); 2045 mtx_lock(&bdlock); 2046 2047 for (;;) { 2048 bd_request = 0; 2049 mtx_unlock(&bdlock); 2050 2051 kthread_suspend_check(bufdaemonproc); 2052 2053 /* 2054 * Do the flush. Limit the amount of in-transit I/O we 2055 * allow to build up, otherwise we would completely saturate 2056 * the I/O system. Wakeup any waiting processes before we 2057 * normally would so they can run in parallel with our drain. 2058 */ 2059 while (numdirtybuffers > lodirtybuffers) { 2060 if (flushbufqueues(0) == 0) { 2061 /* 2062 * Could not find any buffers without rollback 2063 * dependencies, so just write the first one 2064 * in the hopes of eventually making progress. 2065 */ 2066 flushbufqueues(1); 2067 break; 2068 } 2069 waitrunningbufspace(); 2070 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2071 } 2072 2073 /* 2074 * Only clear bd_request if we have reached our low water 2075 * mark. The buf_daemon normally waits 1 second and 2076 * then incrementally flushes any dirty buffers that have 2077 * built up, within reason. 2078 * 2079 * If we were unable to hit our low water mark and couldn't 2080 * find any flushable buffers, we sleep half a second. 2081 * Otherwise we loop immediately. 2082 */ 2083 mtx_lock(&bdlock); 2084 if (numdirtybuffers <= lodirtybuffers) { 2085 /* 2086 * We reached our low water mark, reset the 2087 * request and sleep until we are needed again. 2088 * The sleep is just so the suspend code works. 2089 */ 2090 bd_request = 0; 2091 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2092 } else { 2093 /* 2094 * We couldn't find any flushable dirty buffers but 2095 * still have too many dirty buffers, we 2096 * have to sleep and try again. (rare) 2097 */ 2098 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2099 } 2100 } 2101} 2102 2103/* 2104 * flushbufqueues: 2105 * 2106 * Try to flush a buffer in the dirty queue. We must be careful to 2107 * free up B_INVAL buffers instead of write them, which NFS is 2108 * particularly sensitive to. 2109 */ 2110int flushwithdeps = 0; 2111SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2112 0, "Number of buffers flushed with dependecies that require rollbacks"); 2113static int 2114flushbufqueues(int flushdeps) 2115{ 2116 struct thread *td = curthread; 2117 struct vnode *vp; 2118 struct buf *bp; 2119 int hasdeps; 2120 2121 mtx_lock(&bqlock); 2122 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2123 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2124 continue; 2125 KASSERT((bp->b_flags & B_DELWRI), 2126 ("unexpected clean buffer %p", bp)); 2127 if ((bp->b_xflags & BX_BKGRDINPROG) != 0) { 2128 BUF_UNLOCK(bp); 2129 continue; 2130 } 2131 if (bp->b_flags & B_INVAL) { 2132 bremfreel(bp); 2133 mtx_unlock(&bqlock); 2134 brelse(bp); 2135 return (1); 2136 } 2137 2138 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2139 if (flushdeps == 0) { 2140 BUF_UNLOCK(bp); 2141 continue; 2142 } 2143 hasdeps = 1; 2144 } else 2145 hasdeps = 0; 2146 /* 2147 * We must hold the lock on a vnode before writing 2148 * one of its buffers. Otherwise we may confuse, or 2149 * in the case of a snapshot vnode, deadlock the 2150 * system. 2151 * 2152 * The lock order here is the reverse of the normal 2153 * of vnode followed by buf lock. This is ok because 2154 * the NOWAIT will prevent deadlock. 2155 */ 2156 if ((vp = bp->b_vp) == NULL || 2157 vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2158 mtx_unlock(&bqlock); 2159 vfs_bio_awrite(bp); 2160 if (vp != NULL) 2161 VOP_UNLOCK(vp, 0, td); 2162 flushwithdeps += hasdeps; 2163 return (1); 2164 } 2165 BUF_UNLOCK(bp); 2166 } 2167 mtx_unlock(&bqlock); 2168 return (0); 2169} 2170 2171/* 2172 * Check to see if a block is currently memory resident. 2173 */ 2174struct buf * 2175incore(struct vnode * vp, daddr_t blkno) 2176{ 2177 struct buf *bp; 2178 2179 int s = splbio(); 2180 VI_LOCK(vp); 2181 bp = gbincore(vp, blkno); 2182 VI_UNLOCK(vp); 2183 splx(s); 2184 return (bp); 2185} 2186 2187/* 2188 * Returns true if no I/O is needed to access the 2189 * associated VM object. This is like incore except 2190 * it also hunts around in the VM system for the data. 2191 */ 2192 2193int 2194inmem(struct vnode * vp, daddr_t blkno) 2195{ 2196 vm_object_t obj; 2197 vm_offset_t toff, tinc, size; 2198 vm_page_t m; 2199 vm_ooffset_t off; 2200 2201 GIANT_REQUIRED; 2202 ASSERT_VOP_LOCKED(vp, "inmem"); 2203 2204 if (incore(vp, blkno)) 2205 return 1; 2206 if (vp->v_mount == NULL) 2207 return 0; 2208 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2209 return 0; 2210 2211 size = PAGE_SIZE; 2212 if (size > vp->v_mount->mnt_stat.f_iosize) 2213 size = vp->v_mount->mnt_stat.f_iosize; 2214 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2215 2216 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2217 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2218 if (!m) 2219 goto notinmem; 2220 tinc = size; 2221 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2222 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2223 if (vm_page_is_valid(m, 2224 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2225 goto notinmem; 2226 } 2227 return 1; 2228 2229notinmem: 2230 return (0); 2231} 2232 2233/* 2234 * vfs_setdirty: 2235 * 2236 * Sets the dirty range for a buffer based on the status of the dirty 2237 * bits in the pages comprising the buffer. 2238 * 2239 * The range is limited to the size of the buffer. 2240 * 2241 * This routine is primarily used by NFS, but is generalized for the 2242 * B_VMIO case. 2243 */ 2244static void 2245vfs_setdirty(struct buf *bp) 2246{ 2247 int i; 2248 vm_object_t object; 2249 2250 GIANT_REQUIRED; 2251 /* 2252 * Degenerate case - empty buffer 2253 */ 2254 2255 if (bp->b_bufsize == 0) 2256 return; 2257 2258 /* 2259 * We qualify the scan for modified pages on whether the 2260 * object has been flushed yet. The OBJ_WRITEABLE flag 2261 * is not cleared simply by protecting pages off. 2262 */ 2263 2264 if ((bp->b_flags & B_VMIO) == 0) 2265 return; 2266 2267 object = bp->b_pages[0]->object; 2268 2269 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2270 printf("Warning: object %p writeable but not mightbedirty\n", object); 2271 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2272 printf("Warning: object %p mightbedirty but not writeable\n", object); 2273 2274 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2275 vm_offset_t boffset; 2276 vm_offset_t eoffset; 2277 2278 vm_page_lock_queues(); 2279 /* 2280 * test the pages to see if they have been modified directly 2281 * by users through the VM system. 2282 */ 2283 for (i = 0; i < bp->b_npages; i++) { 2284 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2285 vm_page_test_dirty(bp->b_pages[i]); 2286 } 2287 2288 /* 2289 * Calculate the encompassing dirty range, boffset and eoffset, 2290 * (eoffset - boffset) bytes. 2291 */ 2292 2293 for (i = 0; i < bp->b_npages; i++) { 2294 if (bp->b_pages[i]->dirty) 2295 break; 2296 } 2297 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2298 2299 for (i = bp->b_npages - 1; i >= 0; --i) { 2300 if (bp->b_pages[i]->dirty) { 2301 break; 2302 } 2303 } 2304 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2305 2306 vm_page_unlock_queues(); 2307 /* 2308 * Fit it to the buffer. 2309 */ 2310 2311 if (eoffset > bp->b_bcount) 2312 eoffset = bp->b_bcount; 2313 2314 /* 2315 * If we have a good dirty range, merge with the existing 2316 * dirty range. 2317 */ 2318 2319 if (boffset < eoffset) { 2320 if (bp->b_dirtyoff > boffset) 2321 bp->b_dirtyoff = boffset; 2322 if (bp->b_dirtyend < eoffset) 2323 bp->b_dirtyend = eoffset; 2324 } 2325 } 2326} 2327 2328/* 2329 * getblk: 2330 * 2331 * Get a block given a specified block and offset into a file/device. 2332 * The buffers B_DONE bit will be cleared on return, making it almost 2333 * ready for an I/O initiation. B_INVAL may or may not be set on 2334 * return. The caller should clear B_INVAL prior to initiating a 2335 * READ. 2336 * 2337 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2338 * an existing buffer. 2339 * 2340 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2341 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2342 * and then cleared based on the backing VM. If the previous buffer is 2343 * non-0-sized but invalid, B_CACHE will be cleared. 2344 * 2345 * If getblk() must create a new buffer, the new buffer is returned with 2346 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2347 * case it is returned with B_INVAL clear and B_CACHE set based on the 2348 * backing VM. 2349 * 2350 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2351 * B_CACHE bit is clear. 2352 * 2353 * What this means, basically, is that the caller should use B_CACHE to 2354 * determine whether the buffer is fully valid or not and should clear 2355 * B_INVAL prior to issuing a read. If the caller intends to validate 2356 * the buffer by loading its data area with something, the caller needs 2357 * to clear B_INVAL. If the caller does this without issuing an I/O, 2358 * the caller should set B_CACHE ( as an optimization ), else the caller 2359 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2360 * a write attempt or if it was a successfull read. If the caller 2361 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2362 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2363 */ 2364struct buf * 2365getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2366 int flags) 2367{ 2368 struct buf *bp; 2369 int s; 2370 int error; 2371 ASSERT_VOP_LOCKED(vp, "getblk"); 2372 2373 if (size > MAXBSIZE) 2374 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2375 2376 s = splbio(); 2377loop: 2378 /* 2379 * Block if we are low on buffers. Certain processes are allowed 2380 * to completely exhaust the buffer cache. 2381 * 2382 * If this check ever becomes a bottleneck it may be better to 2383 * move it into the else, when gbincore() fails. At the moment 2384 * it isn't a problem. 2385 * 2386 * XXX remove if 0 sections (clean this up after its proven) 2387 */ 2388 if (numfreebuffers == 0) { 2389 if (curthread == PCPU_GET(idlethread)) 2390 return NULL; 2391 mtx_lock(&nblock); 2392 needsbuffer |= VFS_BIO_NEED_ANY; 2393 mtx_unlock(&nblock); 2394 } 2395 2396 VI_LOCK(vp); 2397 if ((bp = gbincore(vp, blkno))) { 2398 int lockflags; 2399 /* 2400 * Buffer is in-core. If the buffer is not busy, it must 2401 * be on a queue. 2402 */ 2403 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2404 2405 if (flags & GB_LOCK_NOWAIT) 2406 lockflags |= LK_NOWAIT; 2407 2408 error = BUF_TIMELOCK(bp, lockflags, 2409 VI_MTX(vp), "getblk", slpflag, slptimeo); 2410 2411 /* 2412 * If we slept and got the lock we have to restart in case 2413 * the buffer changed identities. 2414 */ 2415 if (error == ENOLCK) 2416 goto loop; 2417 /* We timed out or were interrupted. */ 2418 else if (error) 2419 return (NULL); 2420 2421 /* 2422 * The buffer is locked. B_CACHE is cleared if the buffer is 2423 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2424 * and for a VMIO buffer B_CACHE is adjusted according to the 2425 * backing VM cache. 2426 */ 2427 if (bp->b_flags & B_INVAL) 2428 bp->b_flags &= ~B_CACHE; 2429 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2430 bp->b_flags |= B_CACHE; 2431 bremfree(bp); 2432 2433 /* 2434 * check for size inconsistancies for non-VMIO case. 2435 */ 2436 2437 if (bp->b_bcount != size) { 2438 if ((bp->b_flags & B_VMIO) == 0 || 2439 (size > bp->b_kvasize)) { 2440 if (bp->b_flags & B_DELWRI) { 2441 bp->b_flags |= B_NOCACHE; 2442 BUF_WRITE(bp); 2443 } else { 2444 if ((bp->b_flags & B_VMIO) && 2445 (LIST_FIRST(&bp->b_dep) == NULL)) { 2446 bp->b_flags |= B_RELBUF; 2447 brelse(bp); 2448 } else { 2449 bp->b_flags |= B_NOCACHE; 2450 BUF_WRITE(bp); 2451 } 2452 } 2453 goto loop; 2454 } 2455 } 2456 2457 /* 2458 * If the size is inconsistant in the VMIO case, we can resize 2459 * the buffer. This might lead to B_CACHE getting set or 2460 * cleared. If the size has not changed, B_CACHE remains 2461 * unchanged from its previous state. 2462 */ 2463 2464 if (bp->b_bcount != size) 2465 allocbuf(bp, size); 2466 2467 KASSERT(bp->b_offset != NOOFFSET, 2468 ("getblk: no buffer offset")); 2469 2470 /* 2471 * A buffer with B_DELWRI set and B_CACHE clear must 2472 * be committed before we can return the buffer in 2473 * order to prevent the caller from issuing a read 2474 * ( due to B_CACHE not being set ) and overwriting 2475 * it. 2476 * 2477 * Most callers, including NFS and FFS, need this to 2478 * operate properly either because they assume they 2479 * can issue a read if B_CACHE is not set, or because 2480 * ( for example ) an uncached B_DELWRI might loop due 2481 * to softupdates re-dirtying the buffer. In the latter 2482 * case, B_CACHE is set after the first write completes, 2483 * preventing further loops. 2484 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2485 * above while extending the buffer, we cannot allow the 2486 * buffer to remain with B_CACHE set after the write 2487 * completes or it will represent a corrupt state. To 2488 * deal with this we set B_NOCACHE to scrap the buffer 2489 * after the write. 2490 * 2491 * We might be able to do something fancy, like setting 2492 * B_CACHE in bwrite() except if B_DELWRI is already set, 2493 * so the below call doesn't set B_CACHE, but that gets real 2494 * confusing. This is much easier. 2495 */ 2496 2497 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2498 bp->b_flags |= B_NOCACHE; 2499 BUF_WRITE(bp); 2500 goto loop; 2501 } 2502 2503 splx(s); 2504 bp->b_flags &= ~B_DONE; 2505 } else { 2506 int bsize, maxsize, vmio; 2507 off_t offset; 2508 2509 /* 2510 * Buffer is not in-core, create new buffer. The buffer 2511 * returned by getnewbuf() is locked. Note that the returned 2512 * buffer is also considered valid (not marked B_INVAL). 2513 */ 2514 VI_UNLOCK(vp); 2515 if (vn_isdisk(vp, NULL)) 2516 bsize = DEV_BSIZE; 2517 else if (vp->v_mountedhere) 2518 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2519 else if (vp->v_mount) 2520 bsize = vp->v_mount->mnt_stat.f_iosize; 2521 else 2522 bsize = size; 2523 2524 offset = blkno * bsize; 2525 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2526 (vp->v_vflag & VV_OBJBUF); 2527 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2528 maxsize = imax(maxsize, bsize); 2529 2530 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2531 if (slpflag || slptimeo) { 2532 splx(s); 2533 return NULL; 2534 } 2535 goto loop; 2536 } 2537 2538 /* 2539 * This code is used to make sure that a buffer is not 2540 * created while the getnewbuf routine is blocked. 2541 * This can be a problem whether the vnode is locked or not. 2542 * If the buffer is created out from under us, we have to 2543 * throw away the one we just created. There is now window 2544 * race because we are safely running at splbio() from the 2545 * point of the duplicate buffer creation through to here, 2546 * and we've locked the buffer. 2547 * 2548 * Note: this must occur before we associate the buffer 2549 * with the vp especially considering limitations in 2550 * the splay tree implementation when dealing with duplicate 2551 * lblkno's. 2552 */ 2553 VI_LOCK(vp); 2554 if (gbincore(vp, blkno)) { 2555 VI_UNLOCK(vp); 2556 bp->b_flags |= B_INVAL; 2557 brelse(bp); 2558 goto loop; 2559 } 2560 2561 /* 2562 * Insert the buffer into the hash, so that it can 2563 * be found by incore. 2564 */ 2565 bp->b_blkno = bp->b_lblkno = blkno; 2566 bp->b_offset = offset; 2567 2568 bgetvp(vp, bp); 2569 VI_UNLOCK(vp); 2570 2571 /* 2572 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2573 * buffer size starts out as 0, B_CACHE will be set by 2574 * allocbuf() for the VMIO case prior to it testing the 2575 * backing store for validity. 2576 */ 2577 2578 if (vmio) { 2579 bp->b_flags |= B_VMIO; 2580#if defined(VFS_BIO_DEBUG) 2581 if (vp->v_type != VREG) 2582 printf("getblk: vmioing file type %d???\n", vp->v_type); 2583#endif 2584 VOP_GETVOBJECT(vp, &bp->b_object); 2585 } else { 2586 bp->b_flags &= ~B_VMIO; 2587 bp->b_object = NULL; 2588 } 2589 2590 allocbuf(bp, size); 2591 2592 splx(s); 2593 bp->b_flags &= ~B_DONE; 2594 } 2595 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2596 return (bp); 2597} 2598 2599/* 2600 * Get an empty, disassociated buffer of given size. The buffer is initially 2601 * set to B_INVAL. 2602 */ 2603struct buf * 2604geteblk(int size) 2605{ 2606 struct buf *bp; 2607 int s; 2608 int maxsize; 2609 2610 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2611 2612 s = splbio(); 2613 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2614 continue; 2615 splx(s); 2616 allocbuf(bp, size); 2617 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2618 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2619 return (bp); 2620} 2621 2622 2623/* 2624 * This code constitutes the buffer memory from either anonymous system 2625 * memory (in the case of non-VMIO operations) or from an associated 2626 * VM object (in the case of VMIO operations). This code is able to 2627 * resize a buffer up or down. 2628 * 2629 * Note that this code is tricky, and has many complications to resolve 2630 * deadlock or inconsistant data situations. Tread lightly!!! 2631 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2632 * the caller. Calling this code willy nilly can result in the loss of data. 2633 * 2634 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2635 * B_CACHE for the non-VMIO case. 2636 */ 2637 2638int 2639allocbuf(struct buf *bp, int size) 2640{ 2641 int newbsize, mbsize; 2642 int i; 2643 2644 GIANT_REQUIRED; 2645 2646 if (BUF_REFCNT(bp) == 0) 2647 panic("allocbuf: buffer not busy"); 2648 2649 if (bp->b_kvasize < size) 2650 panic("allocbuf: buffer too small"); 2651 2652 if ((bp->b_flags & B_VMIO) == 0) { 2653 caddr_t origbuf; 2654 int origbufsize; 2655 /* 2656 * Just get anonymous memory from the kernel. Don't 2657 * mess with B_CACHE. 2658 */ 2659 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2660 if (bp->b_flags & B_MALLOC) 2661 newbsize = mbsize; 2662 else 2663 newbsize = round_page(size); 2664 2665 if (newbsize < bp->b_bufsize) { 2666 /* 2667 * malloced buffers are not shrunk 2668 */ 2669 if (bp->b_flags & B_MALLOC) { 2670 if (newbsize) { 2671 bp->b_bcount = size; 2672 } else { 2673 free(bp->b_data, M_BIOBUF); 2674 if (bp->b_bufsize) { 2675 atomic_subtract_int( 2676 &bufmallocspace, 2677 bp->b_bufsize); 2678 bufspacewakeup(); 2679 bp->b_bufsize = 0; 2680 } 2681 bp->b_data = bp->b_kvabase; 2682 bp->b_bcount = 0; 2683 bp->b_flags &= ~B_MALLOC; 2684 } 2685 return 1; 2686 } 2687 vm_hold_free_pages( 2688 bp, 2689 (vm_offset_t) bp->b_data + newbsize, 2690 (vm_offset_t) bp->b_data + bp->b_bufsize); 2691 } else if (newbsize > bp->b_bufsize) { 2692 /* 2693 * We only use malloced memory on the first allocation. 2694 * and revert to page-allocated memory when the buffer 2695 * grows. 2696 */ 2697 /* 2698 * There is a potential smp race here that could lead 2699 * to bufmallocspace slightly passing the max. It 2700 * is probably extremely rare and not worth worrying 2701 * over. 2702 */ 2703 if ( (bufmallocspace < maxbufmallocspace) && 2704 (bp->b_bufsize == 0) && 2705 (mbsize <= PAGE_SIZE/2)) { 2706 2707 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2708 bp->b_bufsize = mbsize; 2709 bp->b_bcount = size; 2710 bp->b_flags |= B_MALLOC; 2711 atomic_add_int(&bufmallocspace, mbsize); 2712 return 1; 2713 } 2714 origbuf = NULL; 2715 origbufsize = 0; 2716 /* 2717 * If the buffer is growing on its other-than-first allocation, 2718 * then we revert to the page-allocation scheme. 2719 */ 2720 if (bp->b_flags & B_MALLOC) { 2721 origbuf = bp->b_data; 2722 origbufsize = bp->b_bufsize; 2723 bp->b_data = bp->b_kvabase; 2724 if (bp->b_bufsize) { 2725 atomic_subtract_int(&bufmallocspace, 2726 bp->b_bufsize); 2727 bufspacewakeup(); 2728 bp->b_bufsize = 0; 2729 } 2730 bp->b_flags &= ~B_MALLOC; 2731 newbsize = round_page(newbsize); 2732 } 2733 vm_hold_load_pages( 2734 bp, 2735 (vm_offset_t) bp->b_data + bp->b_bufsize, 2736 (vm_offset_t) bp->b_data + newbsize); 2737 if (origbuf) { 2738 bcopy(origbuf, bp->b_data, origbufsize); 2739 free(origbuf, M_BIOBUF); 2740 } 2741 } 2742 } else { 2743 int desiredpages; 2744 2745 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2746 desiredpages = (size == 0) ? 0 : 2747 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2748 2749 if (bp->b_flags & B_MALLOC) 2750 panic("allocbuf: VMIO buffer can't be malloced"); 2751 /* 2752 * Set B_CACHE initially if buffer is 0 length or will become 2753 * 0-length. 2754 */ 2755 if (size == 0 || bp->b_bufsize == 0) 2756 bp->b_flags |= B_CACHE; 2757 2758 if (newbsize < bp->b_bufsize) { 2759 /* 2760 * DEV_BSIZE aligned new buffer size is less then the 2761 * DEV_BSIZE aligned existing buffer size. Figure out 2762 * if we have to remove any pages. 2763 */ 2764 if (desiredpages < bp->b_npages) { 2765 vm_page_t m; 2766 2767 vm_page_lock_queues(); 2768 for (i = desiredpages; i < bp->b_npages; i++) { 2769 /* 2770 * the page is not freed here -- it 2771 * is the responsibility of 2772 * vnode_pager_setsize 2773 */ 2774 m = bp->b_pages[i]; 2775 KASSERT(m != bogus_page, 2776 ("allocbuf: bogus page found")); 2777 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2778 vm_page_lock_queues(); 2779 2780 bp->b_pages[i] = NULL; 2781 vm_page_unwire(m, 0); 2782 } 2783 vm_page_unlock_queues(); 2784 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2785 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2786 bp->b_npages = desiredpages; 2787 } 2788 } else if (size > bp->b_bcount) { 2789 /* 2790 * We are growing the buffer, possibly in a 2791 * byte-granular fashion. 2792 */ 2793 struct vnode *vp; 2794 vm_object_t obj; 2795 vm_offset_t toff; 2796 vm_offset_t tinc; 2797 2798 /* 2799 * Step 1, bring in the VM pages from the object, 2800 * allocating them if necessary. We must clear 2801 * B_CACHE if these pages are not valid for the 2802 * range covered by the buffer. 2803 */ 2804 2805 vp = bp->b_vp; 2806 obj = bp->b_object; 2807 2808 while (bp->b_npages < desiredpages) { 2809 vm_page_t m; 2810 vm_pindex_t pi; 2811 2812 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2813 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2814 /* 2815 * note: must allocate system pages 2816 * since blocking here could intefere 2817 * with paging I/O, no matter which 2818 * process we are. 2819 */ 2820 m = vm_page_alloc(obj, pi, 2821 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2822 if (m == NULL) { 2823 atomic_add_int(&vm_pageout_deficit, 2824 desiredpages - bp->b_npages); 2825 VM_WAIT; 2826 } else { 2827 vm_page_lock_queues(); 2828 vm_page_wakeup(m); 2829 vm_page_unlock_queues(); 2830 bp->b_flags &= ~B_CACHE; 2831 bp->b_pages[bp->b_npages] = m; 2832 ++bp->b_npages; 2833 } 2834 continue; 2835 } 2836 2837 /* 2838 * We found a page. If we have to sleep on it, 2839 * retry because it might have gotten freed out 2840 * from under us. 2841 * 2842 * We can only test PG_BUSY here. Blocking on 2843 * m->busy might lead to a deadlock: 2844 * 2845 * vm_fault->getpages->cluster_read->allocbuf 2846 * 2847 */ 2848 vm_page_lock_queues(); 2849 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2850 continue; 2851 2852 /* 2853 * We have a good page. Should we wakeup the 2854 * page daemon? 2855 */ 2856 if ((curproc != pageproc) && 2857 ((m->queue - m->pc) == PQ_CACHE) && 2858 ((cnt.v_free_count + cnt.v_cache_count) < 2859 (cnt.v_free_min + cnt.v_cache_min))) { 2860 pagedaemon_wakeup(); 2861 } 2862 vm_page_flag_clear(m, PG_ZERO); 2863 vm_page_wire(m); 2864 vm_page_unlock_queues(); 2865 bp->b_pages[bp->b_npages] = m; 2866 ++bp->b_npages; 2867 } 2868 2869 /* 2870 * Step 2. We've loaded the pages into the buffer, 2871 * we have to figure out if we can still have B_CACHE 2872 * set. Note that B_CACHE is set according to the 2873 * byte-granular range ( bcount and size ), new the 2874 * aligned range ( newbsize ). 2875 * 2876 * The VM test is against m->valid, which is DEV_BSIZE 2877 * aligned. Needless to say, the validity of the data 2878 * needs to also be DEV_BSIZE aligned. Note that this 2879 * fails with NFS if the server or some other client 2880 * extends the file's EOF. If our buffer is resized, 2881 * B_CACHE may remain set! XXX 2882 */ 2883 2884 toff = bp->b_bcount; 2885 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2886 2887 while ((bp->b_flags & B_CACHE) && toff < size) { 2888 vm_pindex_t pi; 2889 2890 if (tinc > (size - toff)) 2891 tinc = size - toff; 2892 2893 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2894 PAGE_SHIFT; 2895 2896 vfs_buf_test_cache( 2897 bp, 2898 bp->b_offset, 2899 toff, 2900 tinc, 2901 bp->b_pages[pi] 2902 ); 2903 toff += tinc; 2904 tinc = PAGE_SIZE; 2905 } 2906 2907 /* 2908 * Step 3, fixup the KVM pmap. Remember that 2909 * bp->b_data is relative to bp->b_offset, but 2910 * bp->b_offset may be offset into the first page. 2911 */ 2912 2913 bp->b_data = (caddr_t) 2914 trunc_page((vm_offset_t)bp->b_data); 2915 pmap_qenter( 2916 (vm_offset_t)bp->b_data, 2917 bp->b_pages, 2918 bp->b_npages 2919 ); 2920 2921 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2922 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2923 } 2924 } 2925 if (newbsize < bp->b_bufsize) 2926 bufspacewakeup(); 2927 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2928 bp->b_bcount = size; /* requested buffer size */ 2929 return 1; 2930} 2931 2932void 2933biodone(struct bio *bp) 2934{ 2935 mtx_lock(&bdonelock); 2936 bp->bio_flags |= BIO_DONE; 2937 if (bp->bio_done == NULL) 2938 wakeup(bp); 2939 mtx_unlock(&bdonelock); 2940 if (bp->bio_done != NULL) 2941 bp->bio_done(bp); 2942} 2943 2944/* 2945 * Wait for a BIO to finish. 2946 * 2947 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2948 * case is not yet clear. 2949 */ 2950int 2951biowait(struct bio *bp, const char *wchan) 2952{ 2953 2954 mtx_lock(&bdonelock); 2955 while ((bp->bio_flags & BIO_DONE) == 0) 2956 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 2957 mtx_unlock(&bdonelock); 2958 if (bp->bio_error != 0) 2959 return (bp->bio_error); 2960 if (!(bp->bio_flags & BIO_ERROR)) 2961 return (0); 2962 return (EIO); 2963} 2964 2965void 2966biofinish(struct bio *bp, struct devstat *stat, int error) 2967{ 2968 2969 if (error) { 2970 bp->bio_error = error; 2971 bp->bio_flags |= BIO_ERROR; 2972 } 2973 if (stat != NULL) 2974 devstat_end_transaction_bio(stat, bp); 2975 biodone(bp); 2976} 2977 2978void 2979bioq_init(struct bio_queue_head *head) 2980{ 2981 TAILQ_INIT(&head->queue); 2982 head->last_pblkno = 0; 2983 head->insert_point = NULL; 2984 head->switch_point = NULL; 2985} 2986 2987void 2988bioq_remove(struct bio_queue_head *head, struct bio *bp) 2989{ 2990 if (bp == head->switch_point) 2991 head->switch_point = TAILQ_NEXT(bp, bio_queue); 2992 if (bp == head->insert_point) { 2993 head->insert_point = TAILQ_PREV(bp, bio_queue, bio_queue); 2994 if (head->insert_point == NULL) 2995 head->last_pblkno = 0; 2996 } else if (bp == TAILQ_FIRST(&head->queue)) 2997 head->last_pblkno = bp->bio_pblkno; 2998 TAILQ_REMOVE(&head->queue, bp, bio_queue); 2999 if (TAILQ_FIRST(&head->queue) == head->switch_point) 3000 head->switch_point = NULL; 3001} 3002 3003/* 3004 * bufwait: 3005 * 3006 * Wait for buffer I/O completion, returning error status. The buffer 3007 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3008 * error and cleared. 3009 */ 3010int 3011bufwait(register struct buf * bp) 3012{ 3013 int s; 3014 3015 s = splbio(); 3016 if (bp->b_iocmd == BIO_READ) 3017 bwait(bp, PRIBIO, "biord"); 3018 else 3019 bwait(bp, PRIBIO, "biowr"); 3020 splx(s); 3021 if (bp->b_flags & B_EINTR) { 3022 bp->b_flags &= ~B_EINTR; 3023 return (EINTR); 3024 } 3025 if (bp->b_ioflags & BIO_ERROR) { 3026 return (bp->b_error ? bp->b_error : EIO); 3027 } else { 3028 return (0); 3029 } 3030} 3031 3032 /* 3033 * Call back function from struct bio back up to struct buf. 3034 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 3035 */ 3036void 3037bufdonebio(struct bio *bp) 3038{ 3039 bufdone(bp->bio_caller2); 3040} 3041 3042/* 3043 * bufdone: 3044 * 3045 * Finish I/O on a buffer, optionally calling a completion function. 3046 * This is usually called from an interrupt so process blocking is 3047 * not allowed. 3048 * 3049 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3050 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3051 * assuming B_INVAL is clear. 3052 * 3053 * For the VMIO case, we set B_CACHE if the op was a read and no 3054 * read error occured, or if the op was a write. B_CACHE is never 3055 * set if the buffer is invalid or otherwise uncacheable. 3056 * 3057 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3058 * initiator to leave B_INVAL set to brelse the buffer out of existance 3059 * in the biodone routine. 3060 */ 3061void 3062bufdone(struct buf *bp) 3063{ 3064 int s; 3065 void (*biodone)(struct buf *); 3066 3067 GIANT_REQUIRED; 3068 3069 s = splbio(); 3070 3071 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3072 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3073 3074 bp->b_flags |= B_DONE; 3075 runningbufwakeup(bp); 3076 3077 if (bp->b_iocmd == BIO_DELETE) { 3078 brelse(bp); 3079 splx(s); 3080 return; 3081 } 3082 3083 if (bp->b_iocmd == BIO_WRITE) { 3084 vwakeup(bp); 3085 } 3086 3087 /* call optional completion function if requested */ 3088 if (bp->b_iodone != NULL) { 3089 biodone = bp->b_iodone; 3090 bp->b_iodone = NULL; 3091 (*biodone) (bp); 3092 splx(s); 3093 return; 3094 } 3095 if (LIST_FIRST(&bp->b_dep) != NULL) 3096 buf_complete(bp); 3097 3098 if (bp->b_flags & B_VMIO) { 3099 int i; 3100 vm_ooffset_t foff; 3101 vm_page_t m; 3102 vm_object_t obj; 3103 int iosize; 3104 struct vnode *vp = bp->b_vp; 3105 3106 obj = bp->b_object; 3107 3108#if defined(VFS_BIO_DEBUG) 3109 mp_fixme("usecount and vflag accessed without locks."); 3110 if (vp->v_usecount == 0) { 3111 panic("biodone: zero vnode ref count"); 3112 } 3113 3114 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3115 panic("biodone: vnode is not setup for merged cache"); 3116 } 3117#endif 3118 3119 foff = bp->b_offset; 3120 KASSERT(bp->b_offset != NOOFFSET, 3121 ("biodone: no buffer offset")); 3122 3123#if defined(VFS_BIO_DEBUG) 3124 if (obj->paging_in_progress < bp->b_npages) { 3125 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3126 obj->paging_in_progress, bp->b_npages); 3127 } 3128#endif 3129 3130 /* 3131 * Set B_CACHE if the op was a normal read and no error 3132 * occured. B_CACHE is set for writes in the b*write() 3133 * routines. 3134 */ 3135 iosize = bp->b_bcount - bp->b_resid; 3136 if (bp->b_iocmd == BIO_READ && 3137 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3138 !(bp->b_ioflags & BIO_ERROR)) { 3139 bp->b_flags |= B_CACHE; 3140 } 3141 vm_page_lock_queues(); 3142 for (i = 0; i < bp->b_npages; i++) { 3143 int bogusflag = 0; 3144 int resid; 3145 3146 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3147 if (resid > iosize) 3148 resid = iosize; 3149 3150 /* 3151 * cleanup bogus pages, restoring the originals 3152 */ 3153 m = bp->b_pages[i]; 3154 if (m == bogus_page) { 3155 bogusflag = 1; 3156 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3157 if (m == NULL) 3158 panic("biodone: page disappeared!"); 3159 bp->b_pages[i] = m; 3160 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3161 } 3162#if defined(VFS_BIO_DEBUG) 3163 if (OFF_TO_IDX(foff) != m->pindex) { 3164 printf( 3165"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3166 (intmax_t)foff, (uintmax_t)m->pindex); 3167 } 3168#endif 3169 3170 /* 3171 * In the write case, the valid and clean bits are 3172 * already changed correctly ( see bdwrite() ), so we 3173 * only need to do this here in the read case. 3174 */ 3175 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3176 vfs_page_set_valid(bp, foff, i, m); 3177 } 3178 vm_page_flag_clear(m, PG_ZERO); 3179 3180 /* 3181 * when debugging new filesystems or buffer I/O methods, this 3182 * is the most common error that pops up. if you see this, you 3183 * have not set the page busy flag correctly!!! 3184 */ 3185 if (m->busy == 0) { 3186 printf("biodone: page busy < 0, " 3187 "pindex: %d, foff: 0x(%x,%x), " 3188 "resid: %d, index: %d\n", 3189 (int) m->pindex, (int)(foff >> 32), 3190 (int) foff & 0xffffffff, resid, i); 3191 if (!vn_isdisk(vp, NULL)) 3192 printf(" iosize: %ld, lblkno: %jd, flags: 0x%x, npages: %d\n", 3193 bp->b_vp->v_mount->mnt_stat.f_iosize, 3194 (intmax_t) bp->b_lblkno, 3195 bp->b_flags, bp->b_npages); 3196 else 3197 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3198 (intmax_t) bp->b_lblkno, 3199 bp->b_flags, bp->b_npages); 3200 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 3201 m->valid, m->dirty, m->wire_count); 3202 panic("biodone: page busy < 0\n"); 3203 } 3204 vm_page_io_finish(m); 3205 vm_object_pip_subtract(obj, 1); 3206 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3207 iosize -= resid; 3208 } 3209 vm_page_unlock_queues(); 3210 if (obj) 3211 vm_object_pip_wakeupn(obj, 0); 3212 } 3213 3214 /* 3215 * For asynchronous completions, release the buffer now. The brelse 3216 * will do a wakeup there if necessary - so no need to do a wakeup 3217 * here in the async case. The sync case always needs to do a wakeup. 3218 */ 3219 3220 if (bp->b_flags & B_ASYNC) { 3221 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3222 brelse(bp); 3223 else 3224 bqrelse(bp); 3225 } else { 3226 bdone(bp); 3227 } 3228 splx(s); 3229} 3230 3231/* 3232 * This routine is called in lieu of iodone in the case of 3233 * incomplete I/O. This keeps the busy status for pages 3234 * consistant. 3235 */ 3236void 3237vfs_unbusy_pages(struct buf * bp) 3238{ 3239 int i; 3240 3241 GIANT_REQUIRED; 3242 3243 runningbufwakeup(bp); 3244 if (bp->b_flags & B_VMIO) { 3245 vm_object_t obj; 3246 3247 obj = bp->b_object; 3248 vm_page_lock_queues(); 3249 for (i = 0; i < bp->b_npages; i++) { 3250 vm_page_t m = bp->b_pages[i]; 3251 3252 if (m == bogus_page) { 3253 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3254 if (!m) { 3255 panic("vfs_unbusy_pages: page missing\n"); 3256 } 3257 bp->b_pages[i] = m; 3258 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3259 } 3260 vm_object_pip_subtract(obj, 1); 3261 vm_page_flag_clear(m, PG_ZERO); 3262 vm_page_io_finish(m); 3263 } 3264 vm_page_unlock_queues(); 3265 vm_object_pip_wakeupn(obj, 0); 3266 } 3267} 3268 3269/* 3270 * vfs_page_set_valid: 3271 * 3272 * Set the valid bits in a page based on the supplied offset. The 3273 * range is restricted to the buffer's size. 3274 * 3275 * This routine is typically called after a read completes. 3276 */ 3277static void 3278vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3279{ 3280 vm_ooffset_t soff, eoff; 3281 3282 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3283 /* 3284 * Start and end offsets in buffer. eoff - soff may not cross a 3285 * page boundry or cross the end of the buffer. The end of the 3286 * buffer, in this case, is our file EOF, not the allocation size 3287 * of the buffer. 3288 */ 3289 soff = off; 3290 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3291 if (eoff > bp->b_offset + bp->b_bcount) 3292 eoff = bp->b_offset + bp->b_bcount; 3293 3294 /* 3295 * Set valid range. This is typically the entire buffer and thus the 3296 * entire page. 3297 */ 3298 if (eoff > soff) { 3299 vm_page_set_validclean( 3300 m, 3301 (vm_offset_t) (soff & PAGE_MASK), 3302 (vm_offset_t) (eoff - soff) 3303 ); 3304 } 3305} 3306 3307/* 3308 * This routine is called before a device strategy routine. 3309 * It is used to tell the VM system that paging I/O is in 3310 * progress, and treat the pages associated with the buffer 3311 * almost as being PG_BUSY. Also the object paging_in_progress 3312 * flag is handled to make sure that the object doesn't become 3313 * inconsistant. 3314 * 3315 * Since I/O has not been initiated yet, certain buffer flags 3316 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3317 * and should be ignored. 3318 */ 3319void 3320vfs_busy_pages(struct buf * bp, int clear_modify) 3321{ 3322 int i, bogus; 3323 3324 if (bp->b_flags & B_VMIO) { 3325 vm_object_t obj; 3326 vm_ooffset_t foff; 3327 3328 obj = bp->b_object; 3329 foff = bp->b_offset; 3330 KASSERT(bp->b_offset != NOOFFSET, 3331 ("vfs_busy_pages: no buffer offset")); 3332 vfs_setdirty(bp); 3333retry: 3334 vm_page_lock_queues(); 3335 for (i = 0; i < bp->b_npages; i++) { 3336 vm_page_t m = bp->b_pages[i]; 3337 3338 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3339 goto retry; 3340 } 3341 bogus = 0; 3342 for (i = 0; i < bp->b_npages; i++) { 3343 vm_page_t m = bp->b_pages[i]; 3344 3345 vm_page_flag_clear(m, PG_ZERO); 3346 if ((bp->b_flags & B_CLUSTER) == 0) { 3347 vm_object_pip_add(obj, 1); 3348 vm_page_io_start(m); 3349 } 3350 /* 3351 * When readying a buffer for a read ( i.e 3352 * clear_modify == 0 ), it is important to do 3353 * bogus_page replacement for valid pages in 3354 * partially instantiated buffers. Partially 3355 * instantiated buffers can, in turn, occur when 3356 * reconstituting a buffer from its VM backing store 3357 * base. We only have to do this if B_CACHE is 3358 * clear ( which causes the I/O to occur in the 3359 * first place ). The replacement prevents the read 3360 * I/O from overwriting potentially dirty VM-backed 3361 * pages. XXX bogus page replacement is, uh, bogus. 3362 * It may not work properly with small-block devices. 3363 * We need to find a better way. 3364 */ 3365 pmap_remove_all(m); 3366 if (clear_modify) 3367 vfs_page_set_valid(bp, foff, i, m); 3368 else if (m->valid == VM_PAGE_BITS_ALL && 3369 (bp->b_flags & B_CACHE) == 0) { 3370 bp->b_pages[i] = bogus_page; 3371 bogus++; 3372 } 3373 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3374 } 3375 vm_page_unlock_queues(); 3376 if (bogus) 3377 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3378 } 3379} 3380 3381/* 3382 * Tell the VM system that the pages associated with this buffer 3383 * are clean. This is used for delayed writes where the data is 3384 * going to go to disk eventually without additional VM intevention. 3385 * 3386 * Note that while we only really need to clean through to b_bcount, we 3387 * just go ahead and clean through to b_bufsize. 3388 */ 3389static void 3390vfs_clean_pages(struct buf * bp) 3391{ 3392 int i; 3393 3394 GIANT_REQUIRED; 3395 3396 if (bp->b_flags & B_VMIO) { 3397 vm_ooffset_t foff; 3398 3399 foff = bp->b_offset; 3400 KASSERT(bp->b_offset != NOOFFSET, 3401 ("vfs_clean_pages: no buffer offset")); 3402 vm_page_lock_queues(); 3403 for (i = 0; i < bp->b_npages; i++) { 3404 vm_page_t m = bp->b_pages[i]; 3405 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3406 vm_ooffset_t eoff = noff; 3407 3408 if (eoff > bp->b_offset + bp->b_bufsize) 3409 eoff = bp->b_offset + bp->b_bufsize; 3410 vfs_page_set_valid(bp, foff, i, m); 3411 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3412 foff = noff; 3413 } 3414 vm_page_unlock_queues(); 3415 } 3416} 3417 3418/* 3419 * vfs_bio_set_validclean: 3420 * 3421 * Set the range within the buffer to valid and clean. The range is 3422 * relative to the beginning of the buffer, b_offset. Note that b_offset 3423 * itself may be offset from the beginning of the first page. 3424 * 3425 */ 3426 3427void 3428vfs_bio_set_validclean(struct buf *bp, int base, int size) 3429{ 3430 if (bp->b_flags & B_VMIO) { 3431 int i; 3432 int n; 3433 3434 /* 3435 * Fixup base to be relative to beginning of first page. 3436 * Set initial n to be the maximum number of bytes in the 3437 * first page that can be validated. 3438 */ 3439 3440 base += (bp->b_offset & PAGE_MASK); 3441 n = PAGE_SIZE - (base & PAGE_MASK); 3442 3443 vm_page_lock_queues(); 3444 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3445 vm_page_t m = bp->b_pages[i]; 3446 3447 if (n > size) 3448 n = size; 3449 3450 vm_page_set_validclean(m, base & PAGE_MASK, n); 3451 base += n; 3452 size -= n; 3453 n = PAGE_SIZE; 3454 } 3455 vm_page_unlock_queues(); 3456 } 3457} 3458 3459/* 3460 * vfs_bio_clrbuf: 3461 * 3462 * clear a buffer. This routine essentially fakes an I/O, so we need 3463 * to clear BIO_ERROR and B_INVAL. 3464 * 3465 * Note that while we only theoretically need to clear through b_bcount, 3466 * we go ahead and clear through b_bufsize. 3467 */ 3468 3469void 3470vfs_bio_clrbuf(struct buf *bp) 3471{ 3472 int i, mask = 0; 3473 caddr_t sa, ea; 3474 3475 GIANT_REQUIRED; 3476 3477 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3478 bp->b_flags &= ~B_INVAL; 3479 bp->b_ioflags &= ~BIO_ERROR; 3480 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3481 (bp->b_offset & PAGE_MASK) == 0) { 3482 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3483 if ((bp->b_pages[0]->valid & mask) == mask) { 3484 bp->b_resid = 0; 3485 return; 3486 } 3487 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3488 ((bp->b_pages[0]->valid & mask) == 0)) { 3489 bzero(bp->b_data, bp->b_bufsize); 3490 bp->b_pages[0]->valid |= mask; 3491 bp->b_resid = 0; 3492 return; 3493 } 3494 } 3495 ea = sa = bp->b_data; 3496 for(i=0;i<bp->b_npages;i++,sa=ea) { 3497 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3498 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3499 ea = (caddr_t)(vm_offset_t)ulmin( 3500 (u_long)(vm_offset_t)ea, 3501 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3502 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3503 if ((bp->b_pages[i]->valid & mask) == mask) 3504 continue; 3505 if ((bp->b_pages[i]->valid & mask) == 0) { 3506 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3507 bzero(sa, ea - sa); 3508 } 3509 } else { 3510 for (; sa < ea; sa += DEV_BSIZE, j++) { 3511 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3512 (bp->b_pages[i]->valid & (1<<j)) == 0) 3513 bzero(sa, DEV_BSIZE); 3514 } 3515 } 3516 bp->b_pages[i]->valid |= mask; 3517 vm_page_lock_queues(); 3518 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3519 vm_page_unlock_queues(); 3520 } 3521 bp->b_resid = 0; 3522 } else { 3523 clrbuf(bp); 3524 } 3525} 3526 3527/* 3528 * vm_hold_load_pages and vm_hold_free_pages get pages into 3529 * a buffers address space. The pages are anonymous and are 3530 * not associated with a file object. 3531 */ 3532static void 3533vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3534{ 3535 vm_offset_t pg; 3536 vm_page_t p; 3537 int index; 3538 3539 GIANT_REQUIRED; 3540 3541 to = round_page(to); 3542 from = round_page(from); 3543 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3544 3545 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3546tryagain: 3547 /* 3548 * note: must allocate system pages since blocking here 3549 * could intefere with paging I/O, no matter which 3550 * process we are. 3551 */ 3552 vm_object_lock(kernel_object); 3553 p = vm_page_alloc(kernel_object, 3554 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3555 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3556 vm_object_unlock(kernel_object); 3557 if (!p) { 3558 atomic_add_int(&vm_pageout_deficit, 3559 (to - pg) >> PAGE_SHIFT); 3560 VM_WAIT; 3561 goto tryagain; 3562 } 3563 vm_page_lock_queues(); 3564 p->valid = VM_PAGE_BITS_ALL; 3565 vm_page_unlock_queues(); 3566 pmap_qenter(pg, &p, 1); 3567 bp->b_pages[index] = p; 3568 vm_page_lock_queues(); 3569 vm_page_wakeup(p); 3570 vm_page_unlock_queues(); 3571 } 3572 bp->b_npages = index; 3573} 3574 3575/* Return pages associated with this buf to the vm system */ 3576static void 3577vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3578{ 3579 vm_offset_t pg; 3580 vm_page_t p; 3581 int index, newnpages; 3582 3583 GIANT_REQUIRED; 3584 3585 from = round_page(from); 3586 to = round_page(to); 3587 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3588 3589 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3590 p = bp->b_pages[index]; 3591 if (p && (index < bp->b_npages)) { 3592 if (p->busy) { 3593 printf( 3594 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3595 (intmax_t)bp->b_blkno, 3596 (intmax_t)bp->b_lblkno); 3597 } 3598 bp->b_pages[index] = NULL; 3599 pmap_qremove(pg, 1); 3600 vm_page_lock_queues(); 3601 vm_page_busy(p); 3602 vm_page_unwire(p, 0); 3603 vm_page_free(p); 3604 vm_page_unlock_queues(); 3605 } 3606 } 3607 bp->b_npages = newnpages; 3608} 3609 3610/* 3611 * Map an IO request into kernel virtual address space. 3612 * 3613 * All requests are (re)mapped into kernel VA space. 3614 * Notice that we use b_bufsize for the size of the buffer 3615 * to be mapped. b_bcount might be modified by the driver. 3616 * 3617 * Note that even if the caller determines that the address space should 3618 * be valid, a race or a smaller-file mapped into a larger space may 3619 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3620 * check the return value. 3621 */ 3622int 3623vmapbuf(struct buf *bp) 3624{ 3625 caddr_t addr, kva; 3626 vm_offset_t pa; 3627 int pidx, i; 3628 struct vm_page *m; 3629 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3630 3631 GIANT_REQUIRED; 3632 3633 if ((bp->b_flags & B_PHYS) == 0) 3634 panic("vmapbuf"); 3635 3636 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3637 addr < bp->b_data + bp->b_bufsize; 3638 addr += PAGE_SIZE, pidx++) { 3639 /* 3640 * Do the vm_fault if needed; do the copy-on-write thing 3641 * when reading stuff off device into memory. 3642 * 3643 * NOTE! Must use pmap_extract() because addr may be in 3644 * the userland address space, and kextract is only guarenteed 3645 * to work for the kernland address space (see: sparc64 port). 3646 */ 3647retry: 3648 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data, 3649 (bp->b_iocmd == BIO_READ) ? 3650 (VM_PROT_READ|VM_PROT_WRITE) : VM_PROT_READ); 3651 if (i < 0) { 3652 printf("vmapbuf: warning, bad user address during I/O\n"); 3653 vm_page_lock_queues(); 3654 for (i = 0; i < pidx; ++i) { 3655 vm_page_unhold(bp->b_pages[i]); 3656 bp->b_pages[i] = NULL; 3657 } 3658 vm_page_unlock_queues(); 3659 return(-1); 3660 } 3661 pa = trunc_page(pmap_extract(pmap, (vm_offset_t) addr)); 3662 if (pa == 0) { 3663 printf("vmapbuf: warning, race against user address during I/O"); 3664 goto retry; 3665 } 3666 m = PHYS_TO_VM_PAGE(pa); 3667 vm_page_lock_queues(); 3668 vm_page_hold(m); 3669 vm_page_unlock_queues(); 3670 bp->b_pages[pidx] = m; 3671 } 3672 if (pidx > btoc(MAXPHYS)) 3673 panic("vmapbuf: mapped more than MAXPHYS"); 3674 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3675 3676 kva = bp->b_saveaddr; 3677 bp->b_npages = pidx; 3678 bp->b_saveaddr = bp->b_data; 3679 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3680 return(0); 3681} 3682 3683/* 3684 * Free the io map PTEs associated with this IO operation. 3685 * We also invalidate the TLB entries and restore the original b_addr. 3686 */ 3687void 3688vunmapbuf(struct buf *bp) 3689{ 3690 int pidx; 3691 int npages; 3692 3693 GIANT_REQUIRED; 3694 3695 if ((bp->b_flags & B_PHYS) == 0) 3696 panic("vunmapbuf"); 3697 3698 npages = bp->b_npages; 3699 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3700 npages); 3701 vm_page_lock_queues(); 3702 for (pidx = 0; pidx < npages; pidx++) 3703 vm_page_unhold(bp->b_pages[pidx]); 3704 vm_page_unlock_queues(); 3705 3706 bp->b_data = bp->b_saveaddr; 3707} 3708 3709void 3710bdone(struct buf *bp) 3711{ 3712 mtx_lock(&bdonelock); 3713 bp->b_flags |= B_DONE; 3714 wakeup(bp); 3715 mtx_unlock(&bdonelock); 3716} 3717 3718void 3719bwait(struct buf *bp, u_char pri, const char *wchan) 3720{ 3721 mtx_lock(&bdonelock); 3722 while ((bp->b_flags & B_DONE) == 0) 3723 msleep(bp, &bdonelock, pri, wchan, 0); 3724 mtx_unlock(&bdonelock); 3725} 3726 3727#include "opt_ddb.h" 3728#ifdef DDB 3729#include <ddb/ddb.h> 3730 3731/* DDB command to show buffer data */ 3732DB_SHOW_COMMAND(buffer, db_show_buffer) 3733{ 3734 /* get args */ 3735 struct buf *bp = (struct buf *)addr; 3736 3737 if (!have_addr) { 3738 db_printf("usage: show buffer <addr>\n"); 3739 return; 3740 } 3741 3742 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3743 db_printf( 3744 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3745 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd, b_pblkno = %jd\n", 3746 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3747 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3748 (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno); 3749 if (bp->b_npages) { 3750 int i; 3751 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3752 for (i = 0; i < bp->b_npages; i++) { 3753 vm_page_t m; 3754 m = bp->b_pages[i]; 3755 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3756 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3757 if ((i + 1) < bp->b_npages) 3758 db_printf(","); 3759 } 3760 db_printf("\n"); 3761 } 3762} 3763#endif /* DDB */ 3764