vfs_bio.c revision 112181
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 112181 2003-03-13 07:19:23Z 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 * Definitions for the buffer free lists. 217 */ 218#define BUFFER_QUEUES 6 /* number of free buffer queues */ 219 220#define QUEUE_NONE 0 /* on no queue */ 221#define QUEUE_LOCKED 1 /* locked buffers */ 222#define QUEUE_CLEAN 2 /* non-B_DELWRI buffers */ 223#define QUEUE_DIRTY 3 /* B_DELWRI buffers */ 224#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 225#define QUEUE_EMPTY 5 /* empty buffer headers */ 226 227/* Queues for free buffers with various properties */ 228static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 229 230/* Lock for the bufqueues */ 231static struct mtx bqlock; 232 233/* 234 * Single global constant for BUF_WMESG, to avoid getting multiple references. 235 * buf_wmesg is referred from macros. 236 */ 237const char *buf_wmesg = BUF_WMESG; 238 239#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 240#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 241#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 242#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 243 244/* 245 * numdirtywakeup: 246 * 247 * If someone is blocked due to there being too many dirty buffers, 248 * and numdirtybuffers is now reasonable, wake them up. 249 */ 250 251static __inline void 252numdirtywakeup(int level) 253{ 254 if (numdirtybuffers <= level) { 255 mtx_lock(&nblock); 256 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 257 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 258 wakeup(&needsbuffer); 259 } 260 mtx_unlock(&nblock); 261 } 262} 263 264/* 265 * bufspacewakeup: 266 * 267 * Called when buffer space is potentially available for recovery. 268 * getnewbuf() will block on this flag when it is unable to free 269 * sufficient buffer space. Buffer space becomes recoverable when 270 * bp's get placed back in the queues. 271 */ 272 273static __inline void 274bufspacewakeup(void) 275{ 276 /* 277 * If someone is waiting for BUF space, wake them up. Even 278 * though we haven't freed the kva space yet, the waiting 279 * process will be able to now. 280 */ 281 mtx_lock(&nblock); 282 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 283 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 284 wakeup(&needsbuffer); 285 } 286 mtx_unlock(&nblock); 287} 288 289/* 290 * runningbufwakeup() - in-progress I/O accounting. 291 * 292 */ 293static __inline void 294runningbufwakeup(struct buf *bp) 295{ 296 if (bp->b_runningbufspace) { 297 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 298 bp->b_runningbufspace = 0; 299 mtx_lock(&rbreqlock); 300 if (runningbufreq && runningbufspace <= lorunningspace) { 301 runningbufreq = 0; 302 wakeup(&runningbufreq); 303 } 304 mtx_unlock(&rbreqlock); 305 } 306} 307 308/* 309 * bufcountwakeup: 310 * 311 * Called when a buffer has been added to one of the free queues to 312 * account for the buffer and to wakeup anyone waiting for free buffers. 313 * This typically occurs when large amounts of metadata are being handled 314 * by the buffer cache ( else buffer space runs out first, usually ). 315 */ 316 317static __inline void 318bufcountwakeup(void) 319{ 320 atomic_add_int(&numfreebuffers, 1); 321 mtx_lock(&nblock); 322 if (needsbuffer) { 323 needsbuffer &= ~VFS_BIO_NEED_ANY; 324 if (numfreebuffers >= hifreebuffers) 325 needsbuffer &= ~VFS_BIO_NEED_FREE; 326 wakeup(&needsbuffer); 327 } 328 mtx_unlock(&nblock); 329} 330 331/* 332 * waitrunningbufspace() 333 * 334 * runningbufspace is a measure of the amount of I/O currently 335 * running. This routine is used in async-write situations to 336 * prevent creating huge backups of pending writes to a device. 337 * Only asynchronous writes are governed by this function. 338 * 339 * Reads will adjust runningbufspace, but will not block based on it. 340 * The read load has a side effect of reducing the allowed write load. 341 * 342 * This does NOT turn an async write into a sync write. It waits 343 * for earlier writes to complete and generally returns before the 344 * caller's write has reached the device. 345 */ 346static __inline void 347waitrunningbufspace(void) 348{ 349 mtx_lock(&rbreqlock); 350 while (runningbufspace > hirunningspace) { 351 ++runningbufreq; 352 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 353 } 354 mtx_unlock(&rbreqlock); 355} 356 357 358/* 359 * vfs_buf_test_cache: 360 * 361 * Called when a buffer is extended. This function clears the B_CACHE 362 * bit if the newly extended portion of the buffer does not contain 363 * valid data. 364 */ 365static __inline__ 366void 367vfs_buf_test_cache(struct buf *bp, 368 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 369 vm_page_t m) 370{ 371 GIANT_REQUIRED; 372 373 if (bp->b_flags & B_CACHE) { 374 int base = (foff + off) & PAGE_MASK; 375 if (vm_page_is_valid(m, base, size) == 0) 376 bp->b_flags &= ~B_CACHE; 377 } 378} 379 380/* Wake up the buffer deamon if necessary */ 381static __inline__ 382void 383bd_wakeup(int dirtybuflevel) 384{ 385 mtx_lock(&bdlock); 386 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 387 bd_request = 1; 388 wakeup(&bd_request); 389 } 390 mtx_unlock(&bdlock); 391} 392 393/* 394 * bd_speedup - speedup the buffer cache flushing code 395 */ 396 397static __inline__ 398void 399bd_speedup(void) 400{ 401 bd_wakeup(1); 402} 403 404/* 405 * Calculating buffer cache scaling values and reserve space for buffer 406 * headers. This is called during low level kernel initialization and 407 * may be called more then once. We CANNOT write to the memory area 408 * being reserved at this time. 409 */ 410caddr_t 411kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 412{ 413 /* 414 * physmem_est is in pages. Convert it to kilobytes (assumes 415 * PAGE_SIZE is >= 1K) 416 */ 417 physmem_est = physmem_est * (PAGE_SIZE / 1024); 418 419 /* 420 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 421 * For the first 64MB of ram nominally allocate sufficient buffers to 422 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 423 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 424 * the buffer cache we limit the eventual kva reservation to 425 * maxbcache bytes. 426 * 427 * factor represents the 1/4 x ram conversion. 428 */ 429 if (nbuf == 0) { 430 int factor = 4 * BKVASIZE / 1024; 431 432 nbuf = 50; 433 if (physmem_est > 4096) 434 nbuf += min((physmem_est - 4096) / factor, 435 65536 / factor); 436 if (physmem_est > 65536) 437 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 438 439 if (maxbcache && nbuf > maxbcache / BKVASIZE) 440 nbuf = maxbcache / BKVASIZE; 441 } 442 443#if 0 444 /* 445 * Do not allow the buffer_map to be more then 1/2 the size of the 446 * kernel_map. 447 */ 448 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 449 (BKVASIZE * 2)) { 450 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 451 (BKVASIZE * 2); 452 printf("Warning: nbufs capped at %d\n", nbuf); 453 } 454#endif 455 456 /* 457 * swbufs are used as temporary holders for I/O, such as paging I/O. 458 * We have no less then 16 and no more then 256. 459 */ 460 nswbuf = max(min(nbuf/4, 256), 16); 461 462 /* 463 * Reserve space for the buffer cache buffers 464 */ 465 swbuf = (void *)v; 466 v = (caddr_t)(swbuf + nswbuf); 467 buf = (void *)v; 468 v = (caddr_t)(buf + nbuf); 469 470 return(v); 471} 472 473/* Initialize the buffer subsystem. Called before use of any buffers. */ 474void 475bufinit(void) 476{ 477 struct buf *bp; 478 vm_offset_t bogus_offset; 479 int i; 480 481 GIANT_REQUIRED; 482 483 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 484 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 485 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 486 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 487 488 /* next, make a null set of free lists */ 489 for (i = 0; i < BUFFER_QUEUES; i++) 490 TAILQ_INIT(&bufqueues[i]); 491 492 /* finally, initialize each buffer header and stick on empty q */ 493 for (i = 0; i < nbuf; i++) { 494 bp = &buf[i]; 495 bzero(bp, sizeof *bp); 496 bp->b_flags = B_INVAL; /* we're just an empty header */ 497 bp->b_dev = NODEV; 498 bp->b_rcred = NOCRED; 499 bp->b_wcred = NOCRED; 500 bp->b_qindex = QUEUE_EMPTY; 501 bp->b_vflags = 0; 502 bp->b_xflags = 0; 503 LIST_INIT(&bp->b_dep); 504 BUF_LOCKINIT(bp); 505 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 506 } 507 508 /* 509 * maxbufspace is the absolute maximum amount of buffer space we are 510 * allowed to reserve in KVM and in real terms. The absolute maximum 511 * is nominally used by buf_daemon. hibufspace is the nominal maximum 512 * used by most other processes. The differential is required to 513 * ensure that buf_daemon is able to run when other processes might 514 * be blocked waiting for buffer space. 515 * 516 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 517 * this may result in KVM fragmentation which is not handled optimally 518 * by the system. 519 */ 520 maxbufspace = nbuf * BKVASIZE; 521 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 522 lobufspace = hibufspace - MAXBSIZE; 523 524 lorunningspace = 512 * 1024; 525 hirunningspace = 1024 * 1024; 526 527/* 528 * Limit the amount of malloc memory since it is wired permanently into 529 * the kernel space. Even though this is accounted for in the buffer 530 * allocation, we don't want the malloced region to grow uncontrolled. 531 * The malloc scheme improves memory utilization significantly on average 532 * (small) directories. 533 */ 534 maxbufmallocspace = hibufspace / 20; 535 536/* 537 * Reduce the chance of a deadlock occuring by limiting the number 538 * of delayed-write dirty buffers we allow to stack up. 539 */ 540 hidirtybuffers = nbuf / 4 + 20; 541 dirtybufthresh = hidirtybuffers * 9 / 10; 542 numdirtybuffers = 0; 543/* 544 * To support extreme low-memory systems, make sure hidirtybuffers cannot 545 * eat up all available buffer space. This occurs when our minimum cannot 546 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 547 * BKVASIZE'd (8K) buffers. 548 */ 549 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 550 hidirtybuffers >>= 1; 551 } 552 lodirtybuffers = hidirtybuffers / 2; 553 554/* 555 * Try to keep the number of free buffers in the specified range, 556 * and give special processes (e.g. like buf_daemon) access to an 557 * emergency reserve. 558 */ 559 lofreebuffers = nbuf / 18 + 5; 560 hifreebuffers = 2 * lofreebuffers; 561 numfreebuffers = nbuf; 562 563/* 564 * Maximum number of async ops initiated per buf_daemon loop. This is 565 * somewhat of a hack at the moment, we really need to limit ourselves 566 * based on the number of bytes of I/O in-transit that were initiated 567 * from buf_daemon. 568 */ 569 570 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 571 vm_object_lock(kernel_object); 572 bogus_page = vm_page_alloc(kernel_object, 573 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 574 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 575 vm_object_unlock(kernel_object); 576} 577 578/* 579 * bfreekva() - free the kva allocation for a buffer. 580 * 581 * Must be called at splbio() or higher as this is the only locking for 582 * buffer_map. 583 * 584 * Since this call frees up buffer space, we call bufspacewakeup(). 585 */ 586static void 587bfreekva(struct buf * bp) 588{ 589 GIANT_REQUIRED; 590 591 if (bp->b_kvasize) { 592 atomic_add_int(&buffreekvacnt, 1); 593 atomic_subtract_int(&bufspace, bp->b_kvasize); 594 vm_map_delete(buffer_map, 595 (vm_offset_t) bp->b_kvabase, 596 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 597 ); 598 bp->b_kvasize = 0; 599 bufspacewakeup(); 600 } 601} 602 603/* 604 * bremfree: 605 * 606 * Remove the buffer from the appropriate free list. 607 */ 608void 609bremfree(struct buf * bp) 610{ 611 mtx_lock(&bqlock); 612 bremfreel(bp); 613 mtx_unlock(&bqlock); 614} 615 616void 617bremfreel(struct buf * bp) 618{ 619 int s = splbio(); 620 int old_qindex = bp->b_qindex; 621 622 GIANT_REQUIRED; 623 624 if (bp->b_qindex != QUEUE_NONE) { 625 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 626 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 627 bp->b_qindex = QUEUE_NONE; 628 } else { 629 if (BUF_REFCNT(bp) <= 1) 630 panic("bremfree: removing a buffer not on a queue"); 631 } 632 633 /* 634 * Fixup numfreebuffers count. If the buffer is invalid or not 635 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 636 * the buffer was free and we must decrement numfreebuffers. 637 */ 638 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 639 switch(old_qindex) { 640 case QUEUE_DIRTY: 641 case QUEUE_CLEAN: 642 case QUEUE_EMPTY: 643 case QUEUE_EMPTYKVA: 644 atomic_subtract_int(&numfreebuffers, 1); 645 break; 646 default: 647 break; 648 } 649 } 650 splx(s); 651} 652 653 654/* 655 * Get a buffer with the specified data. Look in the cache first. We 656 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 657 * is set, the buffer is valid and we do not have to do anything ( see 658 * getblk() ). This is really just a special case of breadn(). 659 */ 660int 661bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 662 struct buf ** bpp) 663{ 664 665 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 666} 667 668/* 669 * Operates like bread, but also starts asynchronous I/O on 670 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 671 * to initiating I/O . If B_CACHE is set, the buffer is valid 672 * and we do not have to do anything. 673 */ 674int 675breadn(struct vnode * vp, daddr_t blkno, int size, 676 daddr_t * rablkno, int *rabsize, 677 int cnt, struct ucred * cred, struct buf ** bpp) 678{ 679 struct buf *bp, *rabp; 680 int i; 681 int rv = 0, readwait = 0; 682 683 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 684 685 /* if not found in cache, do some I/O */ 686 if ((bp->b_flags & B_CACHE) == 0) { 687 if (curthread != PCPU_GET(idlethread)) 688 curthread->td_proc->p_stats->p_ru.ru_inblock++; 689 bp->b_iocmd = BIO_READ; 690 bp->b_flags &= ~B_INVAL; 691 bp->b_ioflags &= ~BIO_ERROR; 692 if (bp->b_rcred == NOCRED && cred != NOCRED) 693 bp->b_rcred = crhold(cred); 694 vfs_busy_pages(bp, 0); 695 if (vp->v_type == VCHR) 696 VOP_SPECSTRATEGY(vp, bp); 697 else 698 VOP_STRATEGY(vp, bp); 699 ++readwait; 700 } 701 702 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 703 if (inmem(vp, *rablkno)) 704 continue; 705 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 706 707 if ((rabp->b_flags & B_CACHE) == 0) { 708 if (curthread != PCPU_GET(idlethread)) 709 curthread->td_proc->p_stats->p_ru.ru_inblock++; 710 rabp->b_flags |= B_ASYNC; 711 rabp->b_flags &= ~B_INVAL; 712 rabp->b_ioflags &= ~BIO_ERROR; 713 rabp->b_iocmd = BIO_READ; 714 if (rabp->b_rcred == NOCRED && cred != NOCRED) 715 rabp->b_rcred = crhold(cred); 716 vfs_busy_pages(rabp, 0); 717 BUF_KERNPROC(rabp); 718 if (vp->v_type == VCHR) 719 VOP_SPECSTRATEGY(vp, rabp); 720 else 721 VOP_STRATEGY(vp, rabp); 722 } else { 723 brelse(rabp); 724 } 725 } 726 727 if (readwait) { 728 rv = bufwait(bp); 729 } 730 return (rv); 731} 732 733/* 734 * Write, release buffer on completion. (Done by iodone 735 * if async). Do not bother writing anything if the buffer 736 * is invalid. 737 * 738 * Note that we set B_CACHE here, indicating that buffer is 739 * fully valid and thus cacheable. This is true even of NFS 740 * now so we set it generally. This could be set either here 741 * or in biodone() since the I/O is synchronous. We put it 742 * here. 743 */ 744 745int 746bwrite(struct buf * bp) 747{ 748 int oldflags, s; 749 struct buf *newbp; 750 751 if (bp->b_flags & B_INVAL) { 752 brelse(bp); 753 return (0); 754 } 755 756 oldflags = bp->b_flags; 757 758 if (BUF_REFCNT(bp) == 0) 759 panic("bwrite: buffer is not busy???"); 760 s = splbio(); 761 /* 762 * If a background write is already in progress, delay 763 * writing this block if it is asynchronous. Otherwise 764 * wait for the background write to complete. 765 */ 766 if (bp->b_xflags & BX_BKGRDINPROG) { 767 if (bp->b_flags & B_ASYNC) { 768 splx(s); 769 bdwrite(bp); 770 return (0); 771 } 772 bp->b_xflags |= BX_BKGRDWAIT; 773 tsleep(&bp->b_xflags, PRIBIO, "bwrbg", 0); 774 if (bp->b_xflags & BX_BKGRDINPROG) 775 panic("bwrite: still writing"); 776 } 777 778 /* Mark the buffer clean */ 779 bundirty(bp); 780 781 /* 782 * If this buffer is marked for background writing and we 783 * do not have to wait for it, make a copy and write the 784 * copy so as to leave this buffer ready for further use. 785 * 786 * This optimization eats a lot of memory. If we have a page 787 * or buffer shortfall we can't do it. 788 */ 789 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 790 (bp->b_flags & B_ASYNC) && 791 !vm_page_count_severe() && 792 !buf_dirty_count_severe()) { 793 if (bp->b_iodone != NULL) { 794 printf("bp->b_iodone = %p\n", bp->b_iodone); 795 panic("bwrite: need chained iodone"); 796 } 797 798 /* get a new block */ 799 newbp = geteblk(bp->b_bufsize); 800 801 /* 802 * set it to be identical to the old block. We have to 803 * set b_lblkno and BKGRDMARKER before calling bgetvp() 804 * to avoid confusing the splay tree and gbincore(). 805 */ 806 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 807 newbp->b_lblkno = bp->b_lblkno; 808 newbp->b_xflags |= BX_BKGRDMARKER; 809 /* XXX The BX_ flags need to be protected as well */ 810 VI_LOCK(bp->b_vp); 811 bgetvp(bp->b_vp, newbp); 812 VI_UNLOCK(bp->b_vp); 813 newbp->b_blkno = bp->b_blkno; 814 newbp->b_offset = bp->b_offset; 815 newbp->b_iodone = vfs_backgroundwritedone; 816 newbp->b_flags |= B_ASYNC; 817 newbp->b_flags &= ~B_INVAL; 818 819 /* move over the dependencies */ 820 if (LIST_FIRST(&bp->b_dep) != NULL) 821 buf_movedeps(bp, newbp); 822 823 /* 824 * Initiate write on the copy, release the original to 825 * the B_LOCKED queue so that it cannot go away until 826 * the background write completes. If not locked it could go 827 * away and then be reconstituted while it was being written. 828 * If the reconstituted buffer were written, we could end up 829 * with two background copies being written at the same time. 830 */ 831 bp->b_xflags |= BX_BKGRDINPROG; 832 bp->b_flags |= B_LOCKED; 833 bqrelse(bp); 834 bp = newbp; 835 } 836 837 bp->b_flags &= ~B_DONE; 838 bp->b_ioflags &= ~BIO_ERROR; 839 bp->b_flags |= B_WRITEINPROG | B_CACHE; 840 bp->b_iocmd = BIO_WRITE; 841 842 VI_LOCK(bp->b_vp); 843 bp->b_vp->v_numoutput++; 844 VI_UNLOCK(bp->b_vp); 845 vfs_busy_pages(bp, 1); 846 847 /* 848 * Normal bwrites pipeline writes 849 */ 850 bp->b_runningbufspace = bp->b_bufsize; 851 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 852 853 if (curthread != PCPU_GET(idlethread)) 854 curthread->td_proc->p_stats->p_ru.ru_oublock++; 855 splx(s); 856 if (oldflags & B_ASYNC) 857 BUF_KERNPROC(bp); 858 if (bp->b_vp->v_type == VCHR) 859 VOP_SPECSTRATEGY(bp->b_vp, bp); 860 else 861 VOP_STRATEGY(bp->b_vp, bp); 862 863 if ((oldflags & B_ASYNC) == 0) { 864 int rtval = bufwait(bp); 865 brelse(bp); 866 return (rtval); 867 } else if ((oldflags & B_NOWDRAIN) == 0) { 868 /* 869 * don't allow the async write to saturate the I/O 870 * system. Deadlocks can occur only if a device strategy 871 * routine (like in MD) turns around and issues another 872 * high-level write, in which case B_NOWDRAIN is expected 873 * to be set. Otherwise we will not deadlock here because 874 * we are blocking waiting for I/O that is already in-progress 875 * to complete. 876 */ 877 waitrunningbufspace(); 878 } 879 880 return (0); 881} 882 883/* 884 * Complete a background write started from bwrite. 885 */ 886static void 887vfs_backgroundwritedone(bp) 888 struct buf *bp; 889{ 890 struct buf *origbp; 891 892 /* 893 * Find the original buffer that we are writing. 894 */ 895 VI_LOCK(bp->b_vp); 896 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 897 panic("backgroundwritedone: lost buffer"); 898 VI_UNLOCK(bp->b_vp); 899 /* 900 * Process dependencies then return any unfinished ones. 901 */ 902 if (LIST_FIRST(&bp->b_dep) != NULL) 903 buf_complete(bp); 904 if (LIST_FIRST(&bp->b_dep) != NULL) 905 buf_movedeps(bp, origbp); 906 907 /* XXX Find out if origbp can disappear or get inconsistent */ 908 /* 909 * Clear the BX_BKGRDINPROG flag in the original buffer 910 * and awaken it if it is waiting for the write to complete. 911 * If BX_BKGRDINPROG is not set in the original buffer it must 912 * have been released and re-instantiated - which is not legal. 913 */ 914 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), 915 ("backgroundwritedone: lost buffer2")); 916 origbp->b_xflags &= ~BX_BKGRDINPROG; 917 if (origbp->b_xflags & BX_BKGRDWAIT) { 918 origbp->b_xflags &= ~BX_BKGRDWAIT; 919 wakeup(&origbp->b_xflags); 920 } 921 /* 922 * Clear the B_LOCKED flag and remove it from the locked 923 * queue if it currently resides there. 924 */ 925 origbp->b_flags &= ~B_LOCKED; 926 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) { 927 bremfree(origbp); 928 bqrelse(origbp); 929 } 930 /* 931 * This buffer is marked B_NOCACHE, so when it is released 932 * by biodone, it will be tossed. We mark it with BIO_READ 933 * to avoid biodone doing a second vwakeup. 934 */ 935 bp->b_flags |= B_NOCACHE; 936 bp->b_iocmd = BIO_READ; 937 bp->b_flags &= ~(B_CACHE | B_DONE); 938 bp->b_iodone = 0; 939 bufdone(bp); 940} 941 942/* 943 * Delayed write. (Buffer is marked dirty). Do not bother writing 944 * anything if the buffer is marked invalid. 945 * 946 * Note that since the buffer must be completely valid, we can safely 947 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 948 * biodone() in order to prevent getblk from writing the buffer 949 * out synchronously. 950 */ 951void 952bdwrite(struct buf * bp) 953{ 954 struct thread *td = curthread; 955 struct vnode *vp; 956 struct buf *nbp; 957 958 GIANT_REQUIRED; 959 960 if (BUF_REFCNT(bp) == 0) 961 panic("bdwrite: buffer is not busy"); 962 963 if (bp->b_flags & B_INVAL) { 964 brelse(bp); 965 return; 966 } 967 968 /* 969 * If we have too many dirty buffers, don't create any more. 970 * If we are wildly over our limit, then force a complete 971 * cleanup. Otherwise, just keep the situation from getting 972 * out of control. Note that we have to avoid a recursive 973 * disaster and not try to clean up after our own cleanup! 974 */ 975 vp = bp->b_vp; 976 VI_LOCK(vp); 977 if (td->td_proc->p_flag & P_COWINPROGRESS) { 978 recursiveflushes++; 979 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) { 980 VI_UNLOCK(vp); 981 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 982 VI_LOCK(vp); 983 altbufferflushes++; 984 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) { 985 /* 986 * Try to find a buffer to flush. 987 */ 988 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) { 989 if ((nbp->b_xflags & BX_BKGRDINPROG) || 990 buf_countdeps(nbp, 0) || 991 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 992 continue; 993 if (bp == nbp) 994 panic("bdwrite: found ourselves"); 995 VI_UNLOCK(vp); 996 if (nbp->b_flags & B_CLUSTEROK) { 997 vfs_bio_awrite(nbp); 998 } else { 999 bremfree(nbp); 1000 bawrite(nbp); 1001 } 1002 VI_LOCK(vp); 1003 dirtybufferflushes++; 1004 break; 1005 } 1006 } 1007 VI_UNLOCK(vp); 1008 1009 bdirty(bp); 1010 /* 1011 * Set B_CACHE, indicating that the buffer is fully valid. This is 1012 * true even of NFS now. 1013 */ 1014 bp->b_flags |= B_CACHE; 1015 1016 /* 1017 * This bmap keeps the system from needing to do the bmap later, 1018 * perhaps when the system is attempting to do a sync. Since it 1019 * is likely that the indirect block -- or whatever other datastructure 1020 * that the filesystem needs is still in memory now, it is a good 1021 * thing to do this. Note also, that if the pageout daemon is 1022 * requesting a sync -- there might not be enough memory to do 1023 * the bmap then... So, this is important to do. 1024 */ 1025 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1026 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1027 } 1028 1029 /* 1030 * Set the *dirty* buffer range based upon the VM system dirty pages. 1031 */ 1032 vfs_setdirty(bp); 1033 1034 /* 1035 * We need to do this here to satisfy the vnode_pager and the 1036 * pageout daemon, so that it thinks that the pages have been 1037 * "cleaned". Note that since the pages are in a delayed write 1038 * buffer -- the VFS layer "will" see that the pages get written 1039 * out on the next sync, or perhaps the cluster will be completed. 1040 */ 1041 vfs_clean_pages(bp); 1042 bqrelse(bp); 1043 1044 /* 1045 * Wakeup the buffer flushing daemon if we have a lot of dirty 1046 * buffers (midpoint between our recovery point and our stall 1047 * point). 1048 */ 1049 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1050 1051 /* 1052 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1053 * due to the softdep code. 1054 */ 1055} 1056 1057/* 1058 * bdirty: 1059 * 1060 * Turn buffer into delayed write request. We must clear BIO_READ and 1061 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1062 * itself to properly update it in the dirty/clean lists. We mark it 1063 * B_DONE to ensure that any asynchronization of the buffer properly 1064 * clears B_DONE ( else a panic will occur later ). 1065 * 1066 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1067 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1068 * should only be called if the buffer is known-good. 1069 * 1070 * Since the buffer is not on a queue, we do not update the numfreebuffers 1071 * count. 1072 * 1073 * Must be called at splbio(). 1074 * The buffer must be on QUEUE_NONE. 1075 */ 1076void 1077bdirty(bp) 1078 struct buf *bp; 1079{ 1080 KASSERT(bp->b_qindex == QUEUE_NONE, 1081 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1082 bp->b_flags &= ~(B_RELBUF); 1083 bp->b_iocmd = BIO_WRITE; 1084 1085 if ((bp->b_flags & B_DELWRI) == 0) { 1086 bp->b_flags |= B_DONE | B_DELWRI; 1087 reassignbuf(bp, bp->b_vp); 1088 atomic_add_int(&numdirtybuffers, 1); 1089 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1090 } 1091} 1092 1093/* 1094 * bundirty: 1095 * 1096 * Clear B_DELWRI for buffer. 1097 * 1098 * Since the buffer is not on a queue, we do not update the numfreebuffers 1099 * count. 1100 * 1101 * Must be called at splbio(). 1102 * The buffer must be on QUEUE_NONE. 1103 */ 1104 1105void 1106bundirty(bp) 1107 struct buf *bp; 1108{ 1109 KASSERT(bp->b_qindex == QUEUE_NONE, 1110 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1111 1112 if (bp->b_flags & B_DELWRI) { 1113 bp->b_flags &= ~B_DELWRI; 1114 reassignbuf(bp, bp->b_vp); 1115 atomic_subtract_int(&numdirtybuffers, 1); 1116 numdirtywakeup(lodirtybuffers); 1117 } 1118 /* 1119 * Since it is now being written, we can clear its deferred write flag. 1120 */ 1121 bp->b_flags &= ~B_DEFERRED; 1122} 1123 1124/* 1125 * bawrite: 1126 * 1127 * Asynchronous write. Start output on a buffer, but do not wait for 1128 * it to complete. The buffer is released when the output completes. 1129 * 1130 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1131 * B_INVAL buffers. Not us. 1132 */ 1133void 1134bawrite(struct buf * bp) 1135{ 1136 bp->b_flags |= B_ASYNC; 1137 (void) BUF_WRITE(bp); 1138} 1139 1140/* 1141 * bwillwrite: 1142 * 1143 * Called prior to the locking of any vnodes when we are expecting to 1144 * write. We do not want to starve the buffer cache with too many 1145 * dirty buffers so we block here. By blocking prior to the locking 1146 * of any vnodes we attempt to avoid the situation where a locked vnode 1147 * prevents the various system daemons from flushing related buffers. 1148 */ 1149 1150void 1151bwillwrite(void) 1152{ 1153 if (numdirtybuffers >= hidirtybuffers) { 1154 int s; 1155 1156 mtx_lock(&Giant); 1157 s = splbio(); 1158 mtx_lock(&nblock); 1159 while (numdirtybuffers >= hidirtybuffers) { 1160 bd_wakeup(1); 1161 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1162 msleep(&needsbuffer, &nblock, 1163 (PRIBIO + 4), "flswai", 0); 1164 } 1165 splx(s); 1166 mtx_unlock(&nblock); 1167 mtx_unlock(&Giant); 1168 } 1169} 1170 1171/* 1172 * Return true if we have too many dirty buffers. 1173 */ 1174int 1175buf_dirty_count_severe(void) 1176{ 1177 return(numdirtybuffers >= hidirtybuffers); 1178} 1179 1180/* 1181 * brelse: 1182 * 1183 * Release a busy buffer and, if requested, free its resources. The 1184 * buffer will be stashed in the appropriate bufqueue[] allowing it 1185 * to be accessed later as a cache entity or reused for other purposes. 1186 */ 1187void 1188brelse(struct buf * bp) 1189{ 1190 int s; 1191 1192 GIANT_REQUIRED; 1193 1194 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1195 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1196 1197 s = splbio(); 1198 1199 if (bp->b_flags & B_LOCKED) 1200 bp->b_ioflags &= ~BIO_ERROR; 1201 1202 if (bp->b_iocmd == BIO_WRITE && 1203 (bp->b_ioflags & BIO_ERROR) && 1204 !(bp->b_flags & B_INVAL)) { 1205 /* 1206 * Failed write, redirty. Must clear BIO_ERROR to prevent 1207 * pages from being scrapped. If B_INVAL is set then 1208 * this case is not run and the next case is run to 1209 * destroy the buffer. B_INVAL can occur if the buffer 1210 * is outside the range supported by the underlying device. 1211 */ 1212 bp->b_ioflags &= ~BIO_ERROR; 1213 bdirty(bp); 1214 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1215 (bp->b_ioflags & BIO_ERROR) || 1216 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1217 /* 1218 * Either a failed I/O or we were asked to free or not 1219 * cache the buffer. 1220 */ 1221 bp->b_flags |= B_INVAL; 1222 if (LIST_FIRST(&bp->b_dep) != NULL) 1223 buf_deallocate(bp); 1224 if (bp->b_flags & B_DELWRI) { 1225 atomic_subtract_int(&numdirtybuffers, 1); 1226 numdirtywakeup(lodirtybuffers); 1227 } 1228 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1229 if ((bp->b_flags & B_VMIO) == 0) { 1230 if (bp->b_bufsize) 1231 allocbuf(bp, 0); 1232 if (bp->b_vp) 1233 brelvp(bp); 1234 } 1235 } 1236 1237 /* 1238 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1239 * is called with B_DELWRI set, the underlying pages may wind up 1240 * getting freed causing a previous write (bdwrite()) to get 'lost' 1241 * because pages associated with a B_DELWRI bp are marked clean. 1242 * 1243 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1244 * if B_DELWRI is set. 1245 * 1246 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1247 * on pages to return pages to the VM page queues. 1248 */ 1249 if (bp->b_flags & B_DELWRI) 1250 bp->b_flags &= ~B_RELBUF; 1251 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1252 bp->b_flags |= B_RELBUF; 1253 1254 /* 1255 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1256 * constituted, not even NFS buffers now. Two flags effect this. If 1257 * B_INVAL, the struct buf is invalidated but the VM object is kept 1258 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1259 * 1260 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1261 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1262 * buffer is also B_INVAL because it hits the re-dirtying code above. 1263 * 1264 * Normally we can do this whether a buffer is B_DELWRI or not. If 1265 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1266 * the commit state and we cannot afford to lose the buffer. If the 1267 * buffer has a background write in progress, we need to keep it 1268 * around to prevent it from being reconstituted and starting a second 1269 * background write. 1270 */ 1271 if ((bp->b_flags & B_VMIO) 1272 && !(bp->b_vp->v_mount != NULL && 1273 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1274 !vn_isdisk(bp->b_vp, NULL) && 1275 (bp->b_flags & B_DELWRI)) 1276 ) { 1277 1278 int i, j, resid; 1279 vm_page_t m; 1280 off_t foff; 1281 vm_pindex_t poff; 1282 vm_object_t obj; 1283 struct vnode *vp; 1284 1285 vp = bp->b_vp; 1286 obj = bp->b_object; 1287 1288 /* 1289 * Get the base offset and length of the buffer. Note that 1290 * in the VMIO case if the buffer block size is not 1291 * page-aligned then b_data pointer may not be page-aligned. 1292 * But our b_pages[] array *IS* page aligned. 1293 * 1294 * block sizes less then DEV_BSIZE (usually 512) are not 1295 * supported due to the page granularity bits (m->valid, 1296 * m->dirty, etc...). 1297 * 1298 * See man buf(9) for more information 1299 */ 1300 resid = bp->b_bufsize; 1301 foff = bp->b_offset; 1302 1303 for (i = 0; i < bp->b_npages; i++) { 1304 int had_bogus = 0; 1305 1306 m = bp->b_pages[i]; 1307 vm_page_lock_queues(); 1308 vm_page_flag_clear(m, PG_ZERO); 1309 vm_page_unlock_queues(); 1310 1311 /* 1312 * If we hit a bogus page, fixup *all* the bogus pages 1313 * now. 1314 */ 1315 if (m == bogus_page) { 1316 poff = OFF_TO_IDX(bp->b_offset); 1317 had_bogus = 1; 1318 1319 for (j = i; j < bp->b_npages; j++) { 1320 vm_page_t mtmp; 1321 mtmp = bp->b_pages[j]; 1322 if (mtmp == bogus_page) { 1323 mtmp = vm_page_lookup(obj, poff + j); 1324 if (!mtmp) { 1325 panic("brelse: page missing\n"); 1326 } 1327 bp->b_pages[j] = mtmp; 1328 } 1329 } 1330 1331 if ((bp->b_flags & B_INVAL) == 0) { 1332 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1333 } 1334 m = bp->b_pages[i]; 1335 } 1336 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1337 int poffset = foff & PAGE_MASK; 1338 int presid = resid > (PAGE_SIZE - poffset) ? 1339 (PAGE_SIZE - poffset) : resid; 1340 1341 KASSERT(presid >= 0, ("brelse: extra page")); 1342 vm_page_set_invalid(m, poffset, presid); 1343 if (had_bogus) 1344 printf("avoided corruption bug in bogus_page/brelse code\n"); 1345 } 1346 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1347 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1348 } 1349 1350 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1351 vfs_vmio_release(bp); 1352 1353 } else if (bp->b_flags & B_VMIO) { 1354 1355 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1356 vfs_vmio_release(bp); 1357 } 1358 1359 } 1360 1361 if (bp->b_qindex != QUEUE_NONE) 1362 panic("brelse: free buffer onto another queue???"); 1363 if (BUF_REFCNT(bp) > 1) { 1364 /* do not release to free list */ 1365 BUF_UNLOCK(bp); 1366 splx(s); 1367 return; 1368 } 1369 1370 /* enqueue */ 1371 mtx_lock(&bqlock); 1372 1373 /* buffers with no memory */ 1374 if (bp->b_bufsize == 0) { 1375 bp->b_flags |= B_INVAL; 1376 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1377 if (bp->b_xflags & BX_BKGRDINPROG) 1378 panic("losing buffer 1"); 1379 if (bp->b_kvasize) { 1380 bp->b_qindex = QUEUE_EMPTYKVA; 1381 } else { 1382 bp->b_qindex = QUEUE_EMPTY; 1383 } 1384 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1385 bp->b_dev = NODEV; 1386 /* buffers with junk contents */ 1387 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1388 (bp->b_ioflags & BIO_ERROR)) { 1389 bp->b_flags |= B_INVAL; 1390 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1391 if (bp->b_xflags & BX_BKGRDINPROG) 1392 panic("losing buffer 2"); 1393 bp->b_qindex = QUEUE_CLEAN; 1394 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1395 bp->b_dev = NODEV; 1396 1397 /* buffers that are locked */ 1398 } else if (bp->b_flags & B_LOCKED) { 1399 bp->b_qindex = QUEUE_LOCKED; 1400 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1401 1402 /* remaining buffers */ 1403 } else { 1404 if (bp->b_flags & B_DELWRI) 1405 bp->b_qindex = QUEUE_DIRTY; 1406 else 1407 bp->b_qindex = QUEUE_CLEAN; 1408 if (bp->b_flags & B_AGE) 1409 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1410 else 1411 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1412 } 1413 mtx_unlock(&bqlock); 1414 1415 /* 1416 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1417 * placed the buffer on the correct queue. We must also disassociate 1418 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1419 * find it. 1420 */ 1421 if (bp->b_flags & B_INVAL) { 1422 if (bp->b_flags & B_DELWRI) 1423 bundirty(bp); 1424 if (bp->b_vp) 1425 brelvp(bp); 1426 } 1427 1428 /* 1429 * Fixup numfreebuffers count. The bp is on an appropriate queue 1430 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1431 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1432 * if B_INVAL is set ). 1433 */ 1434 1435 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1436 bufcountwakeup(); 1437 1438 /* 1439 * Something we can maybe free or reuse 1440 */ 1441 if (bp->b_bufsize || bp->b_kvasize) 1442 bufspacewakeup(); 1443 1444 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | 1445 B_DIRECT | B_NOWDRAIN); 1446 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1447 panic("brelse: not dirty"); 1448 /* unlock */ 1449 BUF_UNLOCK(bp); 1450 splx(s); 1451} 1452 1453/* 1454 * Release a buffer back to the appropriate queue but do not try to free 1455 * it. The buffer is expected to be used again soon. 1456 * 1457 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1458 * biodone() to requeue an async I/O on completion. It is also used when 1459 * known good buffers need to be requeued but we think we may need the data 1460 * again soon. 1461 * 1462 * XXX we should be able to leave the B_RELBUF hint set on completion. 1463 */ 1464void 1465bqrelse(struct buf * bp) 1466{ 1467 int s; 1468 1469 s = splbio(); 1470 1471 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1472 1473 if (bp->b_qindex != QUEUE_NONE) 1474 panic("bqrelse: free buffer onto another queue???"); 1475 if (BUF_REFCNT(bp) > 1) { 1476 /* do not release to free list */ 1477 BUF_UNLOCK(bp); 1478 splx(s); 1479 return; 1480 } 1481 mtx_lock(&bqlock); 1482 if (bp->b_flags & B_LOCKED) { 1483 bp->b_ioflags &= ~BIO_ERROR; 1484 bp->b_qindex = QUEUE_LOCKED; 1485 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1486 /* buffers with stale but valid contents */ 1487 } else if (bp->b_flags & B_DELWRI) { 1488 bp->b_qindex = QUEUE_DIRTY; 1489 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1490 } else if (vm_page_count_severe()) { 1491 /* 1492 * We are too low on memory, we have to try to free the 1493 * buffer (most importantly: the wired pages making up its 1494 * backing store) *now*. 1495 */ 1496 mtx_unlock(&bqlock); 1497 splx(s); 1498 brelse(bp); 1499 return; 1500 } else { 1501 bp->b_qindex = QUEUE_CLEAN; 1502 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1503 } 1504 mtx_unlock(&bqlock); 1505 1506 if ((bp->b_flags & B_LOCKED) == 0 && 1507 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1508 bufcountwakeup(); 1509 } 1510 1511 /* 1512 * Something we can maybe free or reuse. 1513 */ 1514 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1515 bufspacewakeup(); 1516 1517 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1518 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1519 panic("bqrelse: not dirty"); 1520 /* unlock */ 1521 BUF_UNLOCK(bp); 1522 splx(s); 1523} 1524 1525/* Give pages used by the bp back to the VM system (where possible) */ 1526static void 1527vfs_vmio_release(bp) 1528 struct buf *bp; 1529{ 1530 int i; 1531 vm_page_t m; 1532 1533 GIANT_REQUIRED; 1534 vm_page_lock_queues(); 1535 for (i = 0; i < bp->b_npages; i++) { 1536 m = bp->b_pages[i]; 1537 bp->b_pages[i] = NULL; 1538 /* 1539 * In order to keep page LRU ordering consistent, put 1540 * everything on the inactive queue. 1541 */ 1542 vm_page_unwire(m, 0); 1543 /* 1544 * We don't mess with busy pages, it is 1545 * the responsibility of the process that 1546 * busied the pages to deal with them. 1547 */ 1548 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1549 continue; 1550 1551 if (m->wire_count == 0) { 1552 vm_page_flag_clear(m, PG_ZERO); 1553 /* 1554 * Might as well free the page if we can and it has 1555 * no valid data. We also free the page if the 1556 * buffer was used for direct I/O 1557 */ 1558 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1559 m->hold_count == 0) { 1560 vm_page_busy(m); 1561 pmap_remove_all(m); 1562 vm_page_free(m); 1563 } else if (bp->b_flags & B_DIRECT) { 1564 vm_page_try_to_free(m); 1565 } else if (vm_page_count_severe()) { 1566 vm_page_try_to_cache(m); 1567 } 1568 } 1569 } 1570 vm_page_unlock_queues(); 1571 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1572 1573 if (bp->b_bufsize) { 1574 bufspacewakeup(); 1575 bp->b_bufsize = 0; 1576 } 1577 bp->b_npages = 0; 1578 bp->b_flags &= ~B_VMIO; 1579 if (bp->b_vp) 1580 brelvp(bp); 1581} 1582 1583/* 1584 * Check to see if a block at a particular lbn is available for a clustered 1585 * write. 1586 */ 1587static int 1588vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1589{ 1590 struct buf *bpa; 1591 int match; 1592 1593 match = 0; 1594 1595 /* If the buf isn't in core skip it */ 1596 if ((bpa = gbincore(vp, lblkno)) == NULL) 1597 return (0); 1598 1599 /* If the buf is busy we don't want to wait for it */ 1600 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1601 return (0); 1602 1603 /* Only cluster with valid clusterable delayed write buffers */ 1604 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1605 (B_DELWRI | B_CLUSTEROK)) 1606 goto done; 1607 1608 if (bpa->b_bufsize != size) 1609 goto done; 1610 1611 /* 1612 * Check to see if it is in the expected place on disk and that the 1613 * block has been mapped. 1614 */ 1615 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1616 match = 1; 1617done: 1618 BUF_UNLOCK(bpa); 1619 return (match); 1620} 1621 1622/* 1623 * vfs_bio_awrite: 1624 * 1625 * Implement clustered async writes for clearing out B_DELWRI buffers. 1626 * This is much better then the old way of writing only one buffer at 1627 * a time. Note that we may not be presented with the buffers in the 1628 * correct order, so we search for the cluster in both directions. 1629 */ 1630int 1631vfs_bio_awrite(struct buf * bp) 1632{ 1633 int i; 1634 int j; 1635 daddr_t lblkno = bp->b_lblkno; 1636 struct vnode *vp = bp->b_vp; 1637 int s; 1638 int ncl; 1639 int nwritten; 1640 int size; 1641 int maxcl; 1642 1643 s = splbio(); 1644 /* 1645 * right now we support clustered writing only to regular files. If 1646 * we find a clusterable block we could be in the middle of a cluster 1647 * rather then at the beginning. 1648 */ 1649 if ((vp->v_type == VREG) && 1650 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1651 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1652 1653 size = vp->v_mount->mnt_stat.f_iosize; 1654 maxcl = MAXPHYS / size; 1655 1656 VI_LOCK(vp); 1657 for (i = 1; i < maxcl; i++) 1658 if (vfs_bio_clcheck(vp, size, lblkno + i, 1659 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1660 break; 1661 1662 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1663 if (vfs_bio_clcheck(vp, size, lblkno - j, 1664 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1665 break; 1666 1667 VI_UNLOCK(vp); 1668 --j; 1669 ncl = i + j; 1670 /* 1671 * this is a possible cluster write 1672 */ 1673 if (ncl != 1) { 1674 BUF_UNLOCK(bp); 1675 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1676 splx(s); 1677 return nwritten; 1678 } 1679 } 1680 1681 bremfree(bp); 1682 bp->b_flags |= B_ASYNC; 1683 1684 splx(s); 1685 /* 1686 * default (old) behavior, writing out only one block 1687 * 1688 * XXX returns b_bufsize instead of b_bcount for nwritten? 1689 */ 1690 nwritten = bp->b_bufsize; 1691 (void) BUF_WRITE(bp); 1692 1693 return nwritten; 1694} 1695 1696/* 1697 * getnewbuf: 1698 * 1699 * Find and initialize a new buffer header, freeing up existing buffers 1700 * in the bufqueues as necessary. The new buffer is returned locked. 1701 * 1702 * Important: B_INVAL is not set. If the caller wishes to throw the 1703 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1704 * 1705 * We block if: 1706 * We have insufficient buffer headers 1707 * We have insufficient buffer space 1708 * buffer_map is too fragmented ( space reservation fails ) 1709 * If we have to flush dirty buffers ( but we try to avoid this ) 1710 * 1711 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1712 * Instead we ask the buf daemon to do it for us. We attempt to 1713 * avoid piecemeal wakeups of the pageout daemon. 1714 */ 1715 1716static struct buf * 1717getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1718{ 1719 struct buf *bp; 1720 struct buf *nbp; 1721 int defrag = 0; 1722 int nqindex; 1723 static int flushingbufs; 1724 1725 GIANT_REQUIRED; 1726 1727 /* 1728 * We can't afford to block since we might be holding a vnode lock, 1729 * which may prevent system daemons from running. We deal with 1730 * low-memory situations by proactively returning memory and running 1731 * async I/O rather then sync I/O. 1732 */ 1733 1734 atomic_add_int(&getnewbufcalls, 1); 1735 atomic_subtract_int(&getnewbufrestarts, 1); 1736restart: 1737 atomic_add_int(&getnewbufrestarts, 1); 1738 1739 /* 1740 * Setup for scan. If we do not have enough free buffers, 1741 * we setup a degenerate case that immediately fails. Note 1742 * that if we are specially marked process, we are allowed to 1743 * dip into our reserves. 1744 * 1745 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1746 * 1747 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1748 * However, there are a number of cases (defragging, reusing, ...) 1749 * where we cannot backup. 1750 */ 1751 mtx_lock(&bqlock); 1752 nqindex = QUEUE_EMPTYKVA; 1753 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1754 1755 if (nbp == NULL) { 1756 /* 1757 * If no EMPTYKVA buffers and we are either 1758 * defragging or reusing, locate a CLEAN buffer 1759 * to free or reuse. If bufspace useage is low 1760 * skip this step so we can allocate a new buffer. 1761 */ 1762 if (defrag || bufspace >= lobufspace) { 1763 nqindex = QUEUE_CLEAN; 1764 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1765 } 1766 1767 /* 1768 * If we could not find or were not allowed to reuse a 1769 * CLEAN buffer, check to see if it is ok to use an EMPTY 1770 * buffer. We can only use an EMPTY buffer if allocating 1771 * its KVA would not otherwise run us out of buffer space. 1772 */ 1773 if (nbp == NULL && defrag == 0 && 1774 bufspace + maxsize < hibufspace) { 1775 nqindex = QUEUE_EMPTY; 1776 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1777 } 1778 } 1779 1780 /* 1781 * Run scan, possibly freeing data and/or kva mappings on the fly 1782 * depending. 1783 */ 1784 1785 while ((bp = nbp) != NULL) { 1786 int qindex = nqindex; 1787 1788 /* 1789 * Calculate next bp ( we can only use it if we do not block 1790 * or do other fancy things ). 1791 */ 1792 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1793 switch(qindex) { 1794 case QUEUE_EMPTY: 1795 nqindex = QUEUE_EMPTYKVA; 1796 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1797 break; 1798 /* FALLTHROUGH */ 1799 case QUEUE_EMPTYKVA: 1800 nqindex = QUEUE_CLEAN; 1801 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1802 break; 1803 /* FALLTHROUGH */ 1804 case QUEUE_CLEAN: 1805 /* 1806 * nbp is NULL. 1807 */ 1808 break; 1809 } 1810 } 1811 1812 /* 1813 * Sanity Checks 1814 */ 1815 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1816 1817 /* 1818 * Note: we no longer distinguish between VMIO and non-VMIO 1819 * buffers. 1820 */ 1821 1822 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1823 1824 /* 1825 * If we are defragging then we need a buffer with 1826 * b_kvasize != 0. XXX this situation should no longer 1827 * occur, if defrag is non-zero the buffer's b_kvasize 1828 * should also be non-zero at this point. XXX 1829 */ 1830 if (defrag && bp->b_kvasize == 0) { 1831 printf("Warning: defrag empty buffer %p\n", bp); 1832 continue; 1833 } 1834 1835 /* 1836 * Start freeing the bp. This is somewhat involved. nbp 1837 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1838 */ 1839 1840 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1841 panic("getnewbuf: locked buf"); 1842 bremfreel(bp); 1843 mtx_unlock(&bqlock); 1844 1845 if (qindex == QUEUE_CLEAN) { 1846 if (bp->b_flags & B_VMIO) { 1847 bp->b_flags &= ~B_ASYNC; 1848 vfs_vmio_release(bp); 1849 } 1850 if (bp->b_vp) 1851 brelvp(bp); 1852 } 1853 1854 /* 1855 * NOTE: nbp is now entirely invalid. We can only restart 1856 * the scan from this point on. 1857 * 1858 * Get the rest of the buffer freed up. b_kva* is still 1859 * valid after this operation. 1860 */ 1861 1862 if (bp->b_rcred != NOCRED) { 1863 crfree(bp->b_rcred); 1864 bp->b_rcred = NOCRED; 1865 } 1866 if (bp->b_wcred != NOCRED) { 1867 crfree(bp->b_wcred); 1868 bp->b_wcred = NOCRED; 1869 } 1870 if (LIST_FIRST(&bp->b_dep) != NULL) 1871 buf_deallocate(bp); 1872 if (bp->b_xflags & BX_BKGRDINPROG) 1873 panic("losing buffer 3"); 1874 1875 if (bp->b_bufsize) 1876 allocbuf(bp, 0); 1877 1878 bp->b_flags = 0; 1879 bp->b_ioflags = 0; 1880 bp->b_xflags = 0; 1881 bp->b_vflags = 0; 1882 bp->b_dev = NODEV; 1883 bp->b_vp = NULL; 1884 bp->b_blkno = bp->b_lblkno = 0; 1885 bp->b_offset = NOOFFSET; 1886 bp->b_iodone = 0; 1887 bp->b_error = 0; 1888 bp->b_resid = 0; 1889 bp->b_bcount = 0; 1890 bp->b_npages = 0; 1891 bp->b_dirtyoff = bp->b_dirtyend = 0; 1892 bp->b_magic = B_MAGIC_BIO; 1893 bp->b_op = &buf_ops_bio; 1894 bp->b_object = NULL; 1895 1896 LIST_INIT(&bp->b_dep); 1897 1898 /* 1899 * If we are defragging then free the buffer. 1900 */ 1901 if (defrag) { 1902 bp->b_flags |= B_INVAL; 1903 bfreekva(bp); 1904 brelse(bp); 1905 defrag = 0; 1906 goto restart; 1907 } 1908 1909 /* 1910 * If we are overcomitted then recover the buffer and its 1911 * KVM space. This occurs in rare situations when multiple 1912 * processes are blocked in getnewbuf() or allocbuf(). 1913 */ 1914 if (bufspace >= hibufspace) 1915 flushingbufs = 1; 1916 if (flushingbufs && bp->b_kvasize != 0) { 1917 bp->b_flags |= B_INVAL; 1918 bfreekva(bp); 1919 brelse(bp); 1920 goto restart; 1921 } 1922 if (bufspace < lobufspace) 1923 flushingbufs = 0; 1924 break; 1925 } 1926 1927 /* 1928 * If we exhausted our list, sleep as appropriate. We may have to 1929 * wakeup various daemons and write out some dirty buffers. 1930 * 1931 * Generally we are sleeping due to insufficient buffer space. 1932 */ 1933 1934 if (bp == NULL) { 1935 int flags; 1936 char *waitmsg; 1937 1938 mtx_unlock(&bqlock); 1939 if (defrag) { 1940 flags = VFS_BIO_NEED_BUFSPACE; 1941 waitmsg = "nbufkv"; 1942 } else if (bufspace >= hibufspace) { 1943 waitmsg = "nbufbs"; 1944 flags = VFS_BIO_NEED_BUFSPACE; 1945 } else { 1946 waitmsg = "newbuf"; 1947 flags = VFS_BIO_NEED_ANY; 1948 } 1949 1950 bd_speedup(); /* heeeelp */ 1951 1952 mtx_lock(&nblock); 1953 needsbuffer |= flags; 1954 while (needsbuffer & flags) { 1955 if (msleep(&needsbuffer, &nblock, 1956 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1957 mtx_unlock(&nblock); 1958 return (NULL); 1959 } 1960 } 1961 mtx_unlock(&nblock); 1962 } else { 1963 /* 1964 * We finally have a valid bp. We aren't quite out of the 1965 * woods, we still have to reserve kva space. In order 1966 * to keep fragmentation sane we only allocate kva in 1967 * BKVASIZE chunks. 1968 */ 1969 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1970 1971 if (maxsize != bp->b_kvasize) { 1972 vm_offset_t addr = 0; 1973 1974 bfreekva(bp); 1975 1976 if (vm_map_findspace(buffer_map, 1977 vm_map_min(buffer_map), maxsize, &addr)) { 1978 /* 1979 * Uh oh. Buffer map is to fragmented. We 1980 * must defragment the map. 1981 */ 1982 atomic_add_int(&bufdefragcnt, 1); 1983 defrag = 1; 1984 bp->b_flags |= B_INVAL; 1985 brelse(bp); 1986 goto restart; 1987 } 1988 if (addr) { 1989 vm_map_insert(buffer_map, NULL, 0, 1990 addr, addr + maxsize, 1991 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1992 1993 bp->b_kvabase = (caddr_t) addr; 1994 bp->b_kvasize = maxsize; 1995 atomic_add_int(&bufspace, bp->b_kvasize); 1996 atomic_add_int(&bufreusecnt, 1); 1997 } 1998 } 1999 bp->b_data = bp->b_kvabase; 2000 } 2001 return(bp); 2002} 2003 2004/* 2005 * buf_daemon: 2006 * 2007 * buffer flushing daemon. Buffers are normally flushed by the 2008 * update daemon but if it cannot keep up this process starts to 2009 * take the load in an attempt to prevent getnewbuf() from blocking. 2010 */ 2011 2012static struct proc *bufdaemonproc; 2013 2014static struct kproc_desc buf_kp = { 2015 "bufdaemon", 2016 buf_daemon, 2017 &bufdaemonproc 2018}; 2019SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2020 2021static void 2022buf_daemon() 2023{ 2024 int s; 2025 2026 mtx_lock(&Giant); 2027 2028 /* 2029 * This process needs to be suspended prior to shutdown sync. 2030 */ 2031 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2032 SHUTDOWN_PRI_LAST); 2033 2034 /* 2035 * This process is allowed to take the buffer cache to the limit 2036 */ 2037 s = splbio(); 2038 mtx_lock(&bdlock); 2039 2040 for (;;) { 2041 bd_request = 0; 2042 mtx_unlock(&bdlock); 2043 2044 kthread_suspend_check(bufdaemonproc); 2045 2046 /* 2047 * Do the flush. Limit the amount of in-transit I/O we 2048 * allow to build up, otherwise we would completely saturate 2049 * the I/O system. Wakeup any waiting processes before we 2050 * normally would so they can run in parallel with our drain. 2051 */ 2052 while (numdirtybuffers > lodirtybuffers) { 2053 if (flushbufqueues(0) == 0) { 2054 /* 2055 * Could not find any buffers without rollback 2056 * dependencies, so just write the first one 2057 * in the hopes of eventually making progress. 2058 */ 2059 flushbufqueues(1); 2060 break; 2061 } 2062 waitrunningbufspace(); 2063 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2064 } 2065 2066 /* 2067 * Only clear bd_request if we have reached our low water 2068 * mark. The buf_daemon normally waits 1 second and 2069 * then incrementally flushes any dirty buffers that have 2070 * built up, within reason. 2071 * 2072 * If we were unable to hit our low water mark and couldn't 2073 * find any flushable buffers, we sleep half a second. 2074 * Otherwise we loop immediately. 2075 */ 2076 mtx_lock(&bdlock); 2077 if (numdirtybuffers <= lodirtybuffers) { 2078 /* 2079 * We reached our low water mark, reset the 2080 * request and sleep until we are needed again. 2081 * The sleep is just so the suspend code works. 2082 */ 2083 bd_request = 0; 2084 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2085 } else { 2086 /* 2087 * We couldn't find any flushable dirty buffers but 2088 * still have too many dirty buffers, we 2089 * have to sleep and try again. (rare) 2090 */ 2091 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2092 } 2093 } 2094} 2095 2096/* 2097 * flushbufqueues: 2098 * 2099 * Try to flush a buffer in the dirty queue. We must be careful to 2100 * free up B_INVAL buffers instead of write them, which NFS is 2101 * particularly sensitive to. 2102 */ 2103int flushwithdeps = 0; 2104SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2105 0, "Number of buffers flushed with dependecies that require rollbacks"); 2106static int 2107flushbufqueues(int flushdeps) 2108{ 2109 struct thread *td = curthread; 2110 struct vnode *vp; 2111 struct buf *bp; 2112 int hasdeps; 2113 2114 mtx_lock(&bqlock); 2115 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2116 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2117 continue; 2118 KASSERT((bp->b_flags & B_DELWRI), 2119 ("unexpected clean buffer %p", bp)); 2120 if ((bp->b_xflags & BX_BKGRDINPROG) != 0) { 2121 BUF_UNLOCK(bp); 2122 continue; 2123 } 2124 if (bp->b_flags & B_INVAL) { 2125 bremfreel(bp); 2126 mtx_unlock(&bqlock); 2127 brelse(bp); 2128 return (1); 2129 } 2130 2131 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2132 if (flushdeps == 0) { 2133 BUF_UNLOCK(bp); 2134 continue; 2135 } 2136 hasdeps = 1; 2137 } else 2138 hasdeps = 0; 2139 /* 2140 * We must hold the lock on a vnode before writing 2141 * one of its buffers. Otherwise we may confuse, or 2142 * in the case of a snapshot vnode, deadlock the 2143 * system. 2144 * 2145 * The lock order here is the reverse of the normal 2146 * of vnode followed by buf lock. This is ok because 2147 * the NOWAIT will prevent deadlock. 2148 */ 2149 if ((vp = bp->b_vp) == NULL || 2150 vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2151 mtx_unlock(&bqlock); 2152 vfs_bio_awrite(bp); 2153 if (vp != NULL) 2154 VOP_UNLOCK(vp, 0, td); 2155 flushwithdeps += hasdeps; 2156 return (1); 2157 } 2158 BUF_UNLOCK(bp); 2159 } 2160 mtx_unlock(&bqlock); 2161 return (0); 2162} 2163 2164/* 2165 * Check to see if a block is currently memory resident. 2166 */ 2167struct buf * 2168incore(struct vnode * vp, daddr_t blkno) 2169{ 2170 struct buf *bp; 2171 2172 int s = splbio(); 2173 VI_LOCK(vp); 2174 bp = gbincore(vp, blkno); 2175 VI_UNLOCK(vp); 2176 splx(s); 2177 return (bp); 2178} 2179 2180/* 2181 * Returns true if no I/O is needed to access the 2182 * associated VM object. This is like incore except 2183 * it also hunts around in the VM system for the data. 2184 */ 2185 2186int 2187inmem(struct vnode * vp, daddr_t blkno) 2188{ 2189 vm_object_t obj; 2190 vm_offset_t toff, tinc, size; 2191 vm_page_t m; 2192 vm_ooffset_t off; 2193 2194 GIANT_REQUIRED; 2195 ASSERT_VOP_LOCKED(vp, "inmem"); 2196 2197 if (incore(vp, blkno)) 2198 return 1; 2199 if (vp->v_mount == NULL) 2200 return 0; 2201 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2202 return 0; 2203 2204 size = PAGE_SIZE; 2205 if (size > vp->v_mount->mnt_stat.f_iosize) 2206 size = vp->v_mount->mnt_stat.f_iosize; 2207 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2208 2209 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2210 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2211 if (!m) 2212 goto notinmem; 2213 tinc = size; 2214 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2215 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2216 if (vm_page_is_valid(m, 2217 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2218 goto notinmem; 2219 } 2220 return 1; 2221 2222notinmem: 2223 return (0); 2224} 2225 2226/* 2227 * vfs_setdirty: 2228 * 2229 * Sets the dirty range for a buffer based on the status of the dirty 2230 * bits in the pages comprising the buffer. 2231 * 2232 * The range is limited to the size of the buffer. 2233 * 2234 * This routine is primarily used by NFS, but is generalized for the 2235 * B_VMIO case. 2236 */ 2237static void 2238vfs_setdirty(struct buf *bp) 2239{ 2240 int i; 2241 vm_object_t object; 2242 2243 GIANT_REQUIRED; 2244 /* 2245 * Degenerate case - empty buffer 2246 */ 2247 2248 if (bp->b_bufsize == 0) 2249 return; 2250 2251 /* 2252 * We qualify the scan for modified pages on whether the 2253 * object has been flushed yet. The OBJ_WRITEABLE flag 2254 * is not cleared simply by protecting pages off. 2255 */ 2256 2257 if ((bp->b_flags & B_VMIO) == 0) 2258 return; 2259 2260 object = bp->b_pages[0]->object; 2261 2262 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2263 printf("Warning: object %p writeable but not mightbedirty\n", object); 2264 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2265 printf("Warning: object %p mightbedirty but not writeable\n", object); 2266 2267 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2268 vm_offset_t boffset; 2269 vm_offset_t eoffset; 2270 2271 vm_page_lock_queues(); 2272 /* 2273 * test the pages to see if they have been modified directly 2274 * by users through the VM system. 2275 */ 2276 for (i = 0; i < bp->b_npages; i++) { 2277 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2278 vm_page_test_dirty(bp->b_pages[i]); 2279 } 2280 2281 /* 2282 * Calculate the encompassing dirty range, boffset and eoffset, 2283 * (eoffset - boffset) bytes. 2284 */ 2285 2286 for (i = 0; i < bp->b_npages; i++) { 2287 if (bp->b_pages[i]->dirty) 2288 break; 2289 } 2290 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2291 2292 for (i = bp->b_npages - 1; i >= 0; --i) { 2293 if (bp->b_pages[i]->dirty) { 2294 break; 2295 } 2296 } 2297 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2298 2299 vm_page_unlock_queues(); 2300 /* 2301 * Fit it to the buffer. 2302 */ 2303 2304 if (eoffset > bp->b_bcount) 2305 eoffset = bp->b_bcount; 2306 2307 /* 2308 * If we have a good dirty range, merge with the existing 2309 * dirty range. 2310 */ 2311 2312 if (boffset < eoffset) { 2313 if (bp->b_dirtyoff > boffset) 2314 bp->b_dirtyoff = boffset; 2315 if (bp->b_dirtyend < eoffset) 2316 bp->b_dirtyend = eoffset; 2317 } 2318 } 2319} 2320 2321/* 2322 * getblk: 2323 * 2324 * Get a block given a specified block and offset into a file/device. 2325 * The buffers B_DONE bit will be cleared on return, making it almost 2326 * ready for an I/O initiation. B_INVAL may or may not be set on 2327 * return. The caller should clear B_INVAL prior to initiating a 2328 * READ. 2329 * 2330 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2331 * an existing buffer. 2332 * 2333 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2334 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2335 * and then cleared based on the backing VM. If the previous buffer is 2336 * non-0-sized but invalid, B_CACHE will be cleared. 2337 * 2338 * If getblk() must create a new buffer, the new buffer is returned with 2339 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2340 * case it is returned with B_INVAL clear and B_CACHE set based on the 2341 * backing VM. 2342 * 2343 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2344 * B_CACHE bit is clear. 2345 * 2346 * What this means, basically, is that the caller should use B_CACHE to 2347 * determine whether the buffer is fully valid or not and should clear 2348 * B_INVAL prior to issuing a read. If the caller intends to validate 2349 * the buffer by loading its data area with something, the caller needs 2350 * to clear B_INVAL. If the caller does this without issuing an I/O, 2351 * the caller should set B_CACHE ( as an optimization ), else the caller 2352 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2353 * a write attempt or if it was a successfull read. If the caller 2354 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2355 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2356 */ 2357struct buf * 2358getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2359 int flags) 2360{ 2361 struct buf *bp; 2362 int s; 2363 int error; 2364 ASSERT_VOP_LOCKED(vp, "getblk"); 2365 2366 if (size > MAXBSIZE) 2367 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2368 2369 s = splbio(); 2370loop: 2371 /* 2372 * Block if we are low on buffers. Certain processes are allowed 2373 * to completely exhaust the buffer cache. 2374 * 2375 * If this check ever becomes a bottleneck it may be better to 2376 * move it into the else, when gbincore() fails. At the moment 2377 * it isn't a problem. 2378 * 2379 * XXX remove if 0 sections (clean this up after its proven) 2380 */ 2381 if (numfreebuffers == 0) { 2382 if (curthread == PCPU_GET(idlethread)) 2383 return NULL; 2384 mtx_lock(&nblock); 2385 needsbuffer |= VFS_BIO_NEED_ANY; 2386 mtx_unlock(&nblock); 2387 } 2388 2389 VI_LOCK(vp); 2390 if ((bp = gbincore(vp, blkno))) { 2391 int lockflags; 2392 /* 2393 * Buffer is in-core. If the buffer is not busy, it must 2394 * be on a queue. 2395 */ 2396 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2397 2398 if (flags & GB_LOCK_NOWAIT) 2399 lockflags |= LK_NOWAIT; 2400 2401 error = BUF_TIMELOCK(bp, lockflags, 2402 VI_MTX(vp), "getblk", slpflag, slptimeo); 2403 2404 /* 2405 * If we slept and got the lock we have to restart in case 2406 * the buffer changed identities. 2407 */ 2408 if (error == ENOLCK) 2409 goto loop; 2410 /* We timed out or were interrupted. */ 2411 else if (error) 2412 return (NULL); 2413 2414 /* 2415 * The buffer is locked. B_CACHE is cleared if the buffer is 2416 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2417 * and for a VMIO buffer B_CACHE is adjusted according to the 2418 * backing VM cache. 2419 */ 2420 if (bp->b_flags & B_INVAL) 2421 bp->b_flags &= ~B_CACHE; 2422 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2423 bp->b_flags |= B_CACHE; 2424 bremfree(bp); 2425 2426 /* 2427 * check for size inconsistancies for non-VMIO case. 2428 */ 2429 2430 if (bp->b_bcount != size) { 2431 if ((bp->b_flags & B_VMIO) == 0 || 2432 (size > bp->b_kvasize)) { 2433 if (bp->b_flags & B_DELWRI) { 2434 bp->b_flags |= B_NOCACHE; 2435 BUF_WRITE(bp); 2436 } else { 2437 if ((bp->b_flags & B_VMIO) && 2438 (LIST_FIRST(&bp->b_dep) == NULL)) { 2439 bp->b_flags |= B_RELBUF; 2440 brelse(bp); 2441 } else { 2442 bp->b_flags |= B_NOCACHE; 2443 BUF_WRITE(bp); 2444 } 2445 } 2446 goto loop; 2447 } 2448 } 2449 2450 /* 2451 * If the size is inconsistant in the VMIO case, we can resize 2452 * the buffer. This might lead to B_CACHE getting set or 2453 * cleared. If the size has not changed, B_CACHE remains 2454 * unchanged from its previous state. 2455 */ 2456 2457 if (bp->b_bcount != size) 2458 allocbuf(bp, size); 2459 2460 KASSERT(bp->b_offset != NOOFFSET, 2461 ("getblk: no buffer offset")); 2462 2463 /* 2464 * A buffer with B_DELWRI set and B_CACHE clear must 2465 * be committed before we can return the buffer in 2466 * order to prevent the caller from issuing a read 2467 * ( due to B_CACHE not being set ) and overwriting 2468 * it. 2469 * 2470 * Most callers, including NFS and FFS, need this to 2471 * operate properly either because they assume they 2472 * can issue a read if B_CACHE is not set, or because 2473 * ( for example ) an uncached B_DELWRI might loop due 2474 * to softupdates re-dirtying the buffer. In the latter 2475 * case, B_CACHE is set after the first write completes, 2476 * preventing further loops. 2477 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2478 * above while extending the buffer, we cannot allow the 2479 * buffer to remain with B_CACHE set after the write 2480 * completes or it will represent a corrupt state. To 2481 * deal with this we set B_NOCACHE to scrap the buffer 2482 * after the write. 2483 * 2484 * We might be able to do something fancy, like setting 2485 * B_CACHE in bwrite() except if B_DELWRI is already set, 2486 * so the below call doesn't set B_CACHE, but that gets real 2487 * confusing. This is much easier. 2488 */ 2489 2490 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2491 bp->b_flags |= B_NOCACHE; 2492 BUF_WRITE(bp); 2493 goto loop; 2494 } 2495 2496 splx(s); 2497 bp->b_flags &= ~B_DONE; 2498 } else { 2499 int bsize, maxsize, vmio; 2500 off_t offset; 2501 2502 /* 2503 * Buffer is not in-core, create new buffer. The buffer 2504 * returned by getnewbuf() is locked. Note that the returned 2505 * buffer is also considered valid (not marked B_INVAL). 2506 */ 2507 VI_UNLOCK(vp); 2508 if (vn_isdisk(vp, NULL)) 2509 bsize = DEV_BSIZE; 2510 else if (vp->v_mountedhere) 2511 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2512 else if (vp->v_mount) 2513 bsize = vp->v_mount->mnt_stat.f_iosize; 2514 else 2515 bsize = size; 2516 2517 offset = blkno * bsize; 2518 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2519 (vp->v_vflag & VV_OBJBUF); 2520 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2521 maxsize = imax(maxsize, bsize); 2522 2523 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2524 if (slpflag || slptimeo) { 2525 splx(s); 2526 return NULL; 2527 } 2528 goto loop; 2529 } 2530 2531 /* 2532 * This code is used to make sure that a buffer is not 2533 * created while the getnewbuf routine is blocked. 2534 * This can be a problem whether the vnode is locked or not. 2535 * If the buffer is created out from under us, we have to 2536 * throw away the one we just created. There is now window 2537 * race because we are safely running at splbio() from the 2538 * point of the duplicate buffer creation through to here, 2539 * and we've locked the buffer. 2540 * 2541 * Note: this must occur before we associate the buffer 2542 * with the vp especially considering limitations in 2543 * the splay tree implementation when dealing with duplicate 2544 * lblkno's. 2545 */ 2546 VI_LOCK(vp); 2547 if (gbincore(vp, blkno)) { 2548 VI_UNLOCK(vp); 2549 bp->b_flags |= B_INVAL; 2550 brelse(bp); 2551 goto loop; 2552 } 2553 2554 /* 2555 * Insert the buffer into the hash, so that it can 2556 * be found by incore. 2557 */ 2558 bp->b_blkno = bp->b_lblkno = blkno; 2559 bp->b_offset = offset; 2560 2561 bgetvp(vp, bp); 2562 VI_UNLOCK(vp); 2563 2564 /* 2565 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2566 * buffer size starts out as 0, B_CACHE will be set by 2567 * allocbuf() for the VMIO case prior to it testing the 2568 * backing store for validity. 2569 */ 2570 2571 if (vmio) { 2572 bp->b_flags |= B_VMIO; 2573#if defined(VFS_BIO_DEBUG) 2574 if (vp->v_type != VREG) 2575 printf("getblk: vmioing file type %d???\n", vp->v_type); 2576#endif 2577 VOP_GETVOBJECT(vp, &bp->b_object); 2578 } else { 2579 bp->b_flags &= ~B_VMIO; 2580 bp->b_object = NULL; 2581 } 2582 2583 allocbuf(bp, size); 2584 2585 splx(s); 2586 bp->b_flags &= ~B_DONE; 2587 } 2588 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2589 return (bp); 2590} 2591 2592/* 2593 * Get an empty, disassociated buffer of given size. The buffer is initially 2594 * set to B_INVAL. 2595 */ 2596struct buf * 2597geteblk(int size) 2598{ 2599 struct buf *bp; 2600 int s; 2601 int maxsize; 2602 2603 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2604 2605 s = splbio(); 2606 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2607 continue; 2608 splx(s); 2609 allocbuf(bp, size); 2610 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2611 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2612 return (bp); 2613} 2614 2615 2616/* 2617 * This code constitutes the buffer memory from either anonymous system 2618 * memory (in the case of non-VMIO operations) or from an associated 2619 * VM object (in the case of VMIO operations). This code is able to 2620 * resize a buffer up or down. 2621 * 2622 * Note that this code is tricky, and has many complications to resolve 2623 * deadlock or inconsistant data situations. Tread lightly!!! 2624 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2625 * the caller. Calling this code willy nilly can result in the loss of data. 2626 * 2627 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2628 * B_CACHE for the non-VMIO case. 2629 */ 2630 2631int 2632allocbuf(struct buf *bp, int size) 2633{ 2634 int newbsize, mbsize; 2635 int i; 2636 2637 GIANT_REQUIRED; 2638 2639 if (BUF_REFCNT(bp) == 0) 2640 panic("allocbuf: buffer not busy"); 2641 2642 if (bp->b_kvasize < size) 2643 panic("allocbuf: buffer too small"); 2644 2645 if ((bp->b_flags & B_VMIO) == 0) { 2646 caddr_t origbuf; 2647 int origbufsize; 2648 /* 2649 * Just get anonymous memory from the kernel. Don't 2650 * mess with B_CACHE. 2651 */ 2652 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2653 if (bp->b_flags & B_MALLOC) 2654 newbsize = mbsize; 2655 else 2656 newbsize = round_page(size); 2657 2658 if (newbsize < bp->b_bufsize) { 2659 /* 2660 * malloced buffers are not shrunk 2661 */ 2662 if (bp->b_flags & B_MALLOC) { 2663 if (newbsize) { 2664 bp->b_bcount = size; 2665 } else { 2666 free(bp->b_data, M_BIOBUF); 2667 if (bp->b_bufsize) { 2668 atomic_subtract_int( 2669 &bufmallocspace, 2670 bp->b_bufsize); 2671 bufspacewakeup(); 2672 bp->b_bufsize = 0; 2673 } 2674 bp->b_data = bp->b_kvabase; 2675 bp->b_bcount = 0; 2676 bp->b_flags &= ~B_MALLOC; 2677 } 2678 return 1; 2679 } 2680 vm_hold_free_pages( 2681 bp, 2682 (vm_offset_t) bp->b_data + newbsize, 2683 (vm_offset_t) bp->b_data + bp->b_bufsize); 2684 } else if (newbsize > bp->b_bufsize) { 2685 /* 2686 * We only use malloced memory on the first allocation. 2687 * and revert to page-allocated memory when the buffer 2688 * grows. 2689 */ 2690 /* 2691 * There is a potential smp race here that could lead 2692 * to bufmallocspace slightly passing the max. It 2693 * is probably extremely rare and not worth worrying 2694 * over. 2695 */ 2696 if ( (bufmallocspace < maxbufmallocspace) && 2697 (bp->b_bufsize == 0) && 2698 (mbsize <= PAGE_SIZE/2)) { 2699 2700 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2701 bp->b_bufsize = mbsize; 2702 bp->b_bcount = size; 2703 bp->b_flags |= B_MALLOC; 2704 atomic_add_int(&bufmallocspace, mbsize); 2705 return 1; 2706 } 2707 origbuf = NULL; 2708 origbufsize = 0; 2709 /* 2710 * If the buffer is growing on its other-than-first allocation, 2711 * then we revert to the page-allocation scheme. 2712 */ 2713 if (bp->b_flags & B_MALLOC) { 2714 origbuf = bp->b_data; 2715 origbufsize = bp->b_bufsize; 2716 bp->b_data = bp->b_kvabase; 2717 if (bp->b_bufsize) { 2718 atomic_subtract_int(&bufmallocspace, 2719 bp->b_bufsize); 2720 bufspacewakeup(); 2721 bp->b_bufsize = 0; 2722 } 2723 bp->b_flags &= ~B_MALLOC; 2724 newbsize = round_page(newbsize); 2725 } 2726 vm_hold_load_pages( 2727 bp, 2728 (vm_offset_t) bp->b_data + bp->b_bufsize, 2729 (vm_offset_t) bp->b_data + newbsize); 2730 if (origbuf) { 2731 bcopy(origbuf, bp->b_data, origbufsize); 2732 free(origbuf, M_BIOBUF); 2733 } 2734 } 2735 } else { 2736 int desiredpages; 2737 2738 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2739 desiredpages = (size == 0) ? 0 : 2740 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2741 2742 if (bp->b_flags & B_MALLOC) 2743 panic("allocbuf: VMIO buffer can't be malloced"); 2744 /* 2745 * Set B_CACHE initially if buffer is 0 length or will become 2746 * 0-length. 2747 */ 2748 if (size == 0 || bp->b_bufsize == 0) 2749 bp->b_flags |= B_CACHE; 2750 2751 if (newbsize < bp->b_bufsize) { 2752 /* 2753 * DEV_BSIZE aligned new buffer size is less then the 2754 * DEV_BSIZE aligned existing buffer size. Figure out 2755 * if we have to remove any pages. 2756 */ 2757 if (desiredpages < bp->b_npages) { 2758 vm_page_t m; 2759 2760 vm_page_lock_queues(); 2761 for (i = desiredpages; i < bp->b_npages; i++) { 2762 /* 2763 * the page is not freed here -- it 2764 * is the responsibility of 2765 * vnode_pager_setsize 2766 */ 2767 m = bp->b_pages[i]; 2768 KASSERT(m != bogus_page, 2769 ("allocbuf: bogus page found")); 2770 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2771 vm_page_lock_queues(); 2772 2773 bp->b_pages[i] = NULL; 2774 vm_page_unwire(m, 0); 2775 } 2776 vm_page_unlock_queues(); 2777 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2778 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2779 bp->b_npages = desiredpages; 2780 } 2781 } else if (size > bp->b_bcount) { 2782 /* 2783 * We are growing the buffer, possibly in a 2784 * byte-granular fashion. 2785 */ 2786 struct vnode *vp; 2787 vm_object_t obj; 2788 vm_offset_t toff; 2789 vm_offset_t tinc; 2790 2791 /* 2792 * Step 1, bring in the VM pages from the object, 2793 * allocating them if necessary. We must clear 2794 * B_CACHE if these pages are not valid for the 2795 * range covered by the buffer. 2796 */ 2797 2798 vp = bp->b_vp; 2799 obj = bp->b_object; 2800 2801 while (bp->b_npages < desiredpages) { 2802 vm_page_t m; 2803 vm_pindex_t pi; 2804 2805 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2806 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2807 /* 2808 * note: must allocate system pages 2809 * since blocking here could intefere 2810 * with paging I/O, no matter which 2811 * process we are. 2812 */ 2813 m = vm_page_alloc(obj, pi, 2814 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2815 if (m == NULL) { 2816 atomic_add_int(&vm_pageout_deficit, 2817 desiredpages - bp->b_npages); 2818 VM_WAIT; 2819 } else { 2820 vm_page_lock_queues(); 2821 vm_page_wakeup(m); 2822 vm_page_unlock_queues(); 2823 bp->b_flags &= ~B_CACHE; 2824 bp->b_pages[bp->b_npages] = m; 2825 ++bp->b_npages; 2826 } 2827 continue; 2828 } 2829 2830 /* 2831 * We found a page. If we have to sleep on it, 2832 * retry because it might have gotten freed out 2833 * from under us. 2834 * 2835 * We can only test PG_BUSY here. Blocking on 2836 * m->busy might lead to a deadlock: 2837 * 2838 * vm_fault->getpages->cluster_read->allocbuf 2839 * 2840 */ 2841 vm_page_lock_queues(); 2842 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2843 continue; 2844 2845 /* 2846 * We have a good page. Should we wakeup the 2847 * page daemon? 2848 */ 2849 if ((curproc != pageproc) && 2850 ((m->queue - m->pc) == PQ_CACHE) && 2851 ((cnt.v_free_count + cnt.v_cache_count) < 2852 (cnt.v_free_min + cnt.v_cache_min))) { 2853 pagedaemon_wakeup(); 2854 } 2855 vm_page_flag_clear(m, PG_ZERO); 2856 vm_page_wire(m); 2857 vm_page_unlock_queues(); 2858 bp->b_pages[bp->b_npages] = m; 2859 ++bp->b_npages; 2860 } 2861 2862 /* 2863 * Step 2. We've loaded the pages into the buffer, 2864 * we have to figure out if we can still have B_CACHE 2865 * set. Note that B_CACHE is set according to the 2866 * byte-granular range ( bcount and size ), new the 2867 * aligned range ( newbsize ). 2868 * 2869 * The VM test is against m->valid, which is DEV_BSIZE 2870 * aligned. Needless to say, the validity of the data 2871 * needs to also be DEV_BSIZE aligned. Note that this 2872 * fails with NFS if the server or some other client 2873 * extends the file's EOF. If our buffer is resized, 2874 * B_CACHE may remain set! XXX 2875 */ 2876 2877 toff = bp->b_bcount; 2878 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2879 2880 while ((bp->b_flags & B_CACHE) && toff < size) { 2881 vm_pindex_t pi; 2882 2883 if (tinc > (size - toff)) 2884 tinc = size - toff; 2885 2886 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2887 PAGE_SHIFT; 2888 2889 vfs_buf_test_cache( 2890 bp, 2891 bp->b_offset, 2892 toff, 2893 tinc, 2894 bp->b_pages[pi] 2895 ); 2896 toff += tinc; 2897 tinc = PAGE_SIZE; 2898 } 2899 2900 /* 2901 * Step 3, fixup the KVM pmap. Remember that 2902 * bp->b_data is relative to bp->b_offset, but 2903 * bp->b_offset may be offset into the first page. 2904 */ 2905 2906 bp->b_data = (caddr_t) 2907 trunc_page((vm_offset_t)bp->b_data); 2908 pmap_qenter( 2909 (vm_offset_t)bp->b_data, 2910 bp->b_pages, 2911 bp->b_npages 2912 ); 2913 2914 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2915 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2916 } 2917 } 2918 if (newbsize < bp->b_bufsize) 2919 bufspacewakeup(); 2920 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2921 bp->b_bcount = size; /* requested buffer size */ 2922 return 1; 2923} 2924 2925void 2926biodone(struct bio *bp) 2927{ 2928 bp->bio_flags |= BIO_DONE; 2929 if (bp->bio_done != NULL) 2930 bp->bio_done(bp); 2931 else 2932 wakeup(bp); 2933} 2934 2935/* 2936 * Wait for a BIO to finish. 2937 * 2938 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2939 * case is not yet clear. 2940 */ 2941int 2942biowait(struct bio *bp, const char *wchan) 2943{ 2944 2945 while ((bp->bio_flags & BIO_DONE) == 0) 2946 msleep(bp, NULL, PRIBIO, wchan, hz / 10); 2947 if (bp->bio_error != 0) 2948 return (bp->bio_error); 2949 if (!(bp->bio_flags & BIO_ERROR)) 2950 return (0); 2951 return (EIO); 2952} 2953 2954void 2955biofinish(struct bio *bp, struct devstat *stat, int error) 2956{ 2957 2958 if (error) { 2959 bp->bio_error = error; 2960 bp->bio_flags |= BIO_ERROR; 2961 } 2962 if (stat != NULL) 2963 devstat_end_transaction_bio(stat, bp); 2964 biodone(bp); 2965} 2966 2967void 2968bioq_init(struct bio_queue_head *head) 2969{ 2970 TAILQ_INIT(&head->queue); 2971 head->last_pblkno = 0; 2972 head->insert_point = NULL; 2973 head->switch_point = NULL; 2974} 2975 2976void 2977bioq_remove(struct bio_queue_head *head, struct bio *bp) 2978{ 2979 if (bp == head->switch_point) 2980 head->switch_point = TAILQ_NEXT(bp, bio_queue); 2981 if (bp == head->insert_point) { 2982 head->insert_point = TAILQ_PREV(bp, bio_queue, bio_queue); 2983 if (head->insert_point == NULL) 2984 head->last_pblkno = 0; 2985 } else if (bp == TAILQ_FIRST(&head->queue)) 2986 head->last_pblkno = bp->bio_pblkno; 2987 TAILQ_REMOVE(&head->queue, bp, bio_queue); 2988 if (TAILQ_FIRST(&head->queue) == head->switch_point) 2989 head->switch_point = NULL; 2990} 2991 2992/* 2993 * bufwait: 2994 * 2995 * Wait for buffer I/O completion, returning error status. The buffer 2996 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 2997 * error and cleared. 2998 */ 2999int 3000bufwait(register struct buf * bp) 3001{ 3002 int s; 3003 3004 s = splbio(); 3005 while ((bp->b_flags & B_DONE) == 0) { 3006 if (bp->b_iocmd == BIO_READ) 3007 tsleep(bp, PRIBIO, "biord", 0); 3008 else 3009 tsleep(bp, PRIBIO, "biowr", 0); 3010 } 3011 splx(s); 3012 if (bp->b_flags & B_EINTR) { 3013 bp->b_flags &= ~B_EINTR; 3014 return (EINTR); 3015 } 3016 if (bp->b_ioflags & BIO_ERROR) { 3017 return (bp->b_error ? bp->b_error : EIO); 3018 } else { 3019 return (0); 3020 } 3021} 3022 3023 /* 3024 * Call back function from struct bio back up to struct buf. 3025 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 3026 */ 3027void 3028bufdonebio(struct bio *bp) 3029{ 3030 bufdone(bp->bio_caller2); 3031} 3032 3033/* 3034 * bufdone: 3035 * 3036 * Finish I/O on a buffer, optionally calling a completion function. 3037 * This is usually called from an interrupt so process blocking is 3038 * not allowed. 3039 * 3040 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3041 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3042 * assuming B_INVAL is clear. 3043 * 3044 * For the VMIO case, we set B_CACHE if the op was a read and no 3045 * read error occured, or if the op was a write. B_CACHE is never 3046 * set if the buffer is invalid or otherwise uncacheable. 3047 * 3048 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3049 * initiator to leave B_INVAL set to brelse the buffer out of existance 3050 * in the biodone routine. 3051 */ 3052void 3053bufdone(struct buf *bp) 3054{ 3055 int s; 3056 void (*biodone)(struct buf *); 3057 3058 GIANT_REQUIRED; 3059 3060 s = splbio(); 3061 3062 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3063 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3064 3065 bp->b_flags |= B_DONE; 3066 runningbufwakeup(bp); 3067 3068 if (bp->b_iocmd == BIO_DELETE) { 3069 brelse(bp); 3070 splx(s); 3071 return; 3072 } 3073 3074 if (bp->b_iocmd == BIO_WRITE) { 3075 vwakeup(bp); 3076 } 3077 3078 /* call optional completion function if requested */ 3079 if (bp->b_iodone != NULL) { 3080 biodone = bp->b_iodone; 3081 bp->b_iodone = NULL; 3082 (*biodone) (bp); 3083 splx(s); 3084 return; 3085 } 3086 if (LIST_FIRST(&bp->b_dep) != NULL) 3087 buf_complete(bp); 3088 3089 if (bp->b_flags & B_VMIO) { 3090 int i; 3091 vm_ooffset_t foff; 3092 vm_page_t m; 3093 vm_object_t obj; 3094 int iosize; 3095 struct vnode *vp = bp->b_vp; 3096 3097 obj = bp->b_object; 3098 3099#if defined(VFS_BIO_DEBUG) 3100 mp_fixme("usecount and vflag accessed without locks."); 3101 if (vp->v_usecount == 0) { 3102 panic("biodone: zero vnode ref count"); 3103 } 3104 3105 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3106 panic("biodone: vnode is not setup for merged cache"); 3107 } 3108#endif 3109 3110 foff = bp->b_offset; 3111 KASSERT(bp->b_offset != NOOFFSET, 3112 ("biodone: no buffer offset")); 3113 3114#if defined(VFS_BIO_DEBUG) 3115 if (obj->paging_in_progress < bp->b_npages) { 3116 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3117 obj->paging_in_progress, bp->b_npages); 3118 } 3119#endif 3120 3121 /* 3122 * Set B_CACHE if the op was a normal read and no error 3123 * occured. B_CACHE is set for writes in the b*write() 3124 * routines. 3125 */ 3126 iosize = bp->b_bcount - bp->b_resid; 3127 if (bp->b_iocmd == BIO_READ && 3128 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3129 !(bp->b_ioflags & BIO_ERROR)) { 3130 bp->b_flags |= B_CACHE; 3131 } 3132 vm_page_lock_queues(); 3133 for (i = 0; i < bp->b_npages; i++) { 3134 int bogusflag = 0; 3135 int resid; 3136 3137 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3138 if (resid > iosize) 3139 resid = iosize; 3140 3141 /* 3142 * cleanup bogus pages, restoring the originals 3143 */ 3144 m = bp->b_pages[i]; 3145 if (m == bogus_page) { 3146 bogusflag = 1; 3147 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3148 if (m == NULL) 3149 panic("biodone: page disappeared!"); 3150 bp->b_pages[i] = m; 3151 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3152 } 3153#if defined(VFS_BIO_DEBUG) 3154 if (OFF_TO_IDX(foff) != m->pindex) { 3155 printf( 3156"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3157 (intmax_t)foff, (uintmax_t)m->pindex); 3158 } 3159#endif 3160 3161 /* 3162 * In the write case, the valid and clean bits are 3163 * already changed correctly ( see bdwrite() ), so we 3164 * only need to do this here in the read case. 3165 */ 3166 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3167 vfs_page_set_valid(bp, foff, i, m); 3168 } 3169 vm_page_flag_clear(m, PG_ZERO); 3170 3171 /* 3172 * when debugging new filesystems or buffer I/O methods, this 3173 * is the most common error that pops up. if you see this, you 3174 * have not set the page busy flag correctly!!! 3175 */ 3176 if (m->busy == 0) { 3177 printf("biodone: page busy < 0, " 3178 "pindex: %d, foff: 0x(%x,%x), " 3179 "resid: %d, index: %d\n", 3180 (int) m->pindex, (int)(foff >> 32), 3181 (int) foff & 0xffffffff, resid, i); 3182 if (!vn_isdisk(vp, NULL)) 3183 printf(" iosize: %ld, lblkno: %jd, flags: 0x%x, npages: %d\n", 3184 bp->b_vp->v_mount->mnt_stat.f_iosize, 3185 (intmax_t) bp->b_lblkno, 3186 bp->b_flags, bp->b_npages); 3187 else 3188 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3189 (intmax_t) bp->b_lblkno, 3190 bp->b_flags, bp->b_npages); 3191 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 3192 m->valid, m->dirty, m->wire_count); 3193 panic("biodone: page busy < 0\n"); 3194 } 3195 vm_page_io_finish(m); 3196 vm_object_pip_subtract(obj, 1); 3197 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3198 iosize -= resid; 3199 } 3200 vm_page_unlock_queues(); 3201 if (obj) 3202 vm_object_pip_wakeupn(obj, 0); 3203 } 3204 3205 /* 3206 * For asynchronous completions, release the buffer now. The brelse 3207 * will do a wakeup there if necessary - so no need to do a wakeup 3208 * here in the async case. The sync case always needs to do a wakeup. 3209 */ 3210 3211 if (bp->b_flags & B_ASYNC) { 3212 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3213 brelse(bp); 3214 else 3215 bqrelse(bp); 3216 } else { 3217 wakeup(bp); 3218 } 3219 splx(s); 3220} 3221 3222/* 3223 * This routine is called in lieu of iodone in the case of 3224 * incomplete I/O. This keeps the busy status for pages 3225 * consistant. 3226 */ 3227void 3228vfs_unbusy_pages(struct buf * bp) 3229{ 3230 int i; 3231 3232 GIANT_REQUIRED; 3233 3234 runningbufwakeup(bp); 3235 if (bp->b_flags & B_VMIO) { 3236 vm_object_t obj; 3237 3238 obj = bp->b_object; 3239 vm_page_lock_queues(); 3240 for (i = 0; i < bp->b_npages; i++) { 3241 vm_page_t m = bp->b_pages[i]; 3242 3243 if (m == bogus_page) { 3244 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3245 if (!m) { 3246 panic("vfs_unbusy_pages: page missing\n"); 3247 } 3248 bp->b_pages[i] = m; 3249 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3250 } 3251 vm_object_pip_subtract(obj, 1); 3252 vm_page_flag_clear(m, PG_ZERO); 3253 vm_page_io_finish(m); 3254 } 3255 vm_page_unlock_queues(); 3256 vm_object_pip_wakeupn(obj, 0); 3257 } 3258} 3259 3260/* 3261 * vfs_page_set_valid: 3262 * 3263 * Set the valid bits in a page based on the supplied offset. The 3264 * range is restricted to the buffer's size. 3265 * 3266 * This routine is typically called after a read completes. 3267 */ 3268static void 3269vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3270{ 3271 vm_ooffset_t soff, eoff; 3272 3273 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3274 /* 3275 * Start and end offsets in buffer. eoff - soff may not cross a 3276 * page boundry or cross the end of the buffer. The end of the 3277 * buffer, in this case, is our file EOF, not the allocation size 3278 * of the buffer. 3279 */ 3280 soff = off; 3281 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3282 if (eoff > bp->b_offset + bp->b_bcount) 3283 eoff = bp->b_offset + bp->b_bcount; 3284 3285 /* 3286 * Set valid range. This is typically the entire buffer and thus the 3287 * entire page. 3288 */ 3289 if (eoff > soff) { 3290 vm_page_set_validclean( 3291 m, 3292 (vm_offset_t) (soff & PAGE_MASK), 3293 (vm_offset_t) (eoff - soff) 3294 ); 3295 } 3296} 3297 3298/* 3299 * This routine is called before a device strategy routine. 3300 * It is used to tell the VM system that paging I/O is in 3301 * progress, and treat the pages associated with the buffer 3302 * almost as being PG_BUSY. Also the object paging_in_progress 3303 * flag is handled to make sure that the object doesn't become 3304 * inconsistant. 3305 * 3306 * Since I/O has not been initiated yet, certain buffer flags 3307 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3308 * and should be ignored. 3309 */ 3310void 3311vfs_busy_pages(struct buf * bp, int clear_modify) 3312{ 3313 int i, bogus; 3314 3315 if (bp->b_flags & B_VMIO) { 3316 vm_object_t obj; 3317 vm_ooffset_t foff; 3318 3319 obj = bp->b_object; 3320 foff = bp->b_offset; 3321 KASSERT(bp->b_offset != NOOFFSET, 3322 ("vfs_busy_pages: no buffer offset")); 3323 vfs_setdirty(bp); 3324retry: 3325 vm_page_lock_queues(); 3326 for (i = 0; i < bp->b_npages; i++) { 3327 vm_page_t m = bp->b_pages[i]; 3328 3329 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3330 goto retry; 3331 } 3332 bogus = 0; 3333 for (i = 0; i < bp->b_npages; i++) { 3334 vm_page_t m = bp->b_pages[i]; 3335 3336 vm_page_flag_clear(m, PG_ZERO); 3337 if ((bp->b_flags & B_CLUSTER) == 0) { 3338 vm_object_pip_add(obj, 1); 3339 vm_page_io_start(m); 3340 } 3341 /* 3342 * When readying a buffer for a read ( i.e 3343 * clear_modify == 0 ), it is important to do 3344 * bogus_page replacement for valid pages in 3345 * partially instantiated buffers. Partially 3346 * instantiated buffers can, in turn, occur when 3347 * reconstituting a buffer from its VM backing store 3348 * base. We only have to do this if B_CACHE is 3349 * clear ( which causes the I/O to occur in the 3350 * first place ). The replacement prevents the read 3351 * I/O from overwriting potentially dirty VM-backed 3352 * pages. XXX bogus page replacement is, uh, bogus. 3353 * It may not work properly with small-block devices. 3354 * We need to find a better way. 3355 */ 3356 pmap_remove_all(m); 3357 if (clear_modify) 3358 vfs_page_set_valid(bp, foff, i, m); 3359 else if (m->valid == VM_PAGE_BITS_ALL && 3360 (bp->b_flags & B_CACHE) == 0) { 3361 bp->b_pages[i] = bogus_page; 3362 bogus++; 3363 } 3364 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3365 } 3366 vm_page_unlock_queues(); 3367 if (bogus) 3368 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3369 } 3370} 3371 3372/* 3373 * Tell the VM system that the pages associated with this buffer 3374 * are clean. This is used for delayed writes where the data is 3375 * going to go to disk eventually without additional VM intevention. 3376 * 3377 * Note that while we only really need to clean through to b_bcount, we 3378 * just go ahead and clean through to b_bufsize. 3379 */ 3380static void 3381vfs_clean_pages(struct buf * bp) 3382{ 3383 int i; 3384 3385 GIANT_REQUIRED; 3386 3387 if (bp->b_flags & B_VMIO) { 3388 vm_ooffset_t foff; 3389 3390 foff = bp->b_offset; 3391 KASSERT(bp->b_offset != NOOFFSET, 3392 ("vfs_clean_pages: no buffer offset")); 3393 vm_page_lock_queues(); 3394 for (i = 0; i < bp->b_npages; i++) { 3395 vm_page_t m = bp->b_pages[i]; 3396 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3397 vm_ooffset_t eoff = noff; 3398 3399 if (eoff > bp->b_offset + bp->b_bufsize) 3400 eoff = bp->b_offset + bp->b_bufsize; 3401 vfs_page_set_valid(bp, foff, i, m); 3402 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3403 foff = noff; 3404 } 3405 vm_page_unlock_queues(); 3406 } 3407} 3408 3409/* 3410 * vfs_bio_set_validclean: 3411 * 3412 * Set the range within the buffer to valid and clean. The range is 3413 * relative to the beginning of the buffer, b_offset. Note that b_offset 3414 * itself may be offset from the beginning of the first page. 3415 * 3416 */ 3417 3418void 3419vfs_bio_set_validclean(struct buf *bp, int base, int size) 3420{ 3421 if (bp->b_flags & B_VMIO) { 3422 int i; 3423 int n; 3424 3425 /* 3426 * Fixup base to be relative to beginning of first page. 3427 * Set initial n to be the maximum number of bytes in the 3428 * first page that can be validated. 3429 */ 3430 3431 base += (bp->b_offset & PAGE_MASK); 3432 n = PAGE_SIZE - (base & PAGE_MASK); 3433 3434 vm_page_lock_queues(); 3435 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3436 vm_page_t m = bp->b_pages[i]; 3437 3438 if (n > size) 3439 n = size; 3440 3441 vm_page_set_validclean(m, base & PAGE_MASK, n); 3442 base += n; 3443 size -= n; 3444 n = PAGE_SIZE; 3445 } 3446 vm_page_unlock_queues(); 3447 } 3448} 3449 3450/* 3451 * vfs_bio_clrbuf: 3452 * 3453 * clear a buffer. This routine essentially fakes an I/O, so we need 3454 * to clear BIO_ERROR and B_INVAL. 3455 * 3456 * Note that while we only theoretically need to clear through b_bcount, 3457 * we go ahead and clear through b_bufsize. 3458 */ 3459 3460void 3461vfs_bio_clrbuf(struct buf *bp) 3462{ 3463 int i, mask = 0; 3464 caddr_t sa, ea; 3465 3466 GIANT_REQUIRED; 3467 3468 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3469 bp->b_flags &= ~B_INVAL; 3470 bp->b_ioflags &= ~BIO_ERROR; 3471 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3472 (bp->b_offset & PAGE_MASK) == 0) { 3473 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3474 if ((bp->b_pages[0]->valid & mask) == mask) { 3475 bp->b_resid = 0; 3476 return; 3477 } 3478 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3479 ((bp->b_pages[0]->valid & mask) == 0)) { 3480 bzero(bp->b_data, bp->b_bufsize); 3481 bp->b_pages[0]->valid |= mask; 3482 bp->b_resid = 0; 3483 return; 3484 } 3485 } 3486 ea = sa = bp->b_data; 3487 for(i=0;i<bp->b_npages;i++,sa=ea) { 3488 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3489 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3490 ea = (caddr_t)(vm_offset_t)ulmin( 3491 (u_long)(vm_offset_t)ea, 3492 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3493 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3494 if ((bp->b_pages[i]->valid & mask) == mask) 3495 continue; 3496 if ((bp->b_pages[i]->valid & mask) == 0) { 3497 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3498 bzero(sa, ea - sa); 3499 } 3500 } else { 3501 for (; sa < ea; sa += DEV_BSIZE, j++) { 3502 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3503 (bp->b_pages[i]->valid & (1<<j)) == 0) 3504 bzero(sa, DEV_BSIZE); 3505 } 3506 } 3507 bp->b_pages[i]->valid |= mask; 3508 vm_page_lock_queues(); 3509 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3510 vm_page_unlock_queues(); 3511 } 3512 bp->b_resid = 0; 3513 } else { 3514 clrbuf(bp); 3515 } 3516} 3517 3518/* 3519 * vm_hold_load_pages and vm_hold_free_pages get pages into 3520 * a buffers address space. The pages are anonymous and are 3521 * not associated with a file object. 3522 */ 3523static void 3524vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3525{ 3526 vm_offset_t pg; 3527 vm_page_t p; 3528 int index; 3529 3530 GIANT_REQUIRED; 3531 3532 to = round_page(to); 3533 from = round_page(from); 3534 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3535 3536 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3537tryagain: 3538 /* 3539 * note: must allocate system pages since blocking here 3540 * could intefere with paging I/O, no matter which 3541 * process we are. 3542 */ 3543 vm_object_lock(kernel_object); 3544 p = vm_page_alloc(kernel_object, 3545 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3546 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3547 vm_object_unlock(kernel_object); 3548 if (!p) { 3549 atomic_add_int(&vm_pageout_deficit, 3550 (to - pg) >> PAGE_SHIFT); 3551 VM_WAIT; 3552 goto tryagain; 3553 } 3554 vm_page_lock_queues(); 3555 p->valid = VM_PAGE_BITS_ALL; 3556 vm_page_unlock_queues(); 3557 pmap_qenter(pg, &p, 1); 3558 bp->b_pages[index] = p; 3559 vm_page_lock_queues(); 3560 vm_page_wakeup(p); 3561 vm_page_unlock_queues(); 3562 } 3563 bp->b_npages = index; 3564} 3565 3566/* Return pages associated with this buf to the vm system */ 3567static void 3568vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3569{ 3570 vm_offset_t pg; 3571 vm_page_t p; 3572 int index, newnpages; 3573 3574 GIANT_REQUIRED; 3575 3576 from = round_page(from); 3577 to = round_page(to); 3578 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3579 3580 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3581 p = bp->b_pages[index]; 3582 if (p && (index < bp->b_npages)) { 3583 if (p->busy) { 3584 printf( 3585 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3586 (intmax_t)bp->b_blkno, 3587 (intmax_t)bp->b_lblkno); 3588 } 3589 bp->b_pages[index] = NULL; 3590 pmap_qremove(pg, 1); 3591 vm_page_lock_queues(); 3592 vm_page_busy(p); 3593 vm_page_unwire(p, 0); 3594 vm_page_free(p); 3595 vm_page_unlock_queues(); 3596 } 3597 } 3598 bp->b_npages = newnpages; 3599} 3600 3601/* 3602 * Map an IO request into kernel virtual address space. 3603 * 3604 * All requests are (re)mapped into kernel VA space. 3605 * Notice that we use b_bufsize for the size of the buffer 3606 * to be mapped. b_bcount might be modified by the driver. 3607 * 3608 * Note that even if the caller determines that the address space should 3609 * be valid, a race or a smaller-file mapped into a larger space may 3610 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3611 * check the return value. 3612 */ 3613int 3614vmapbuf(struct buf *bp) 3615{ 3616 caddr_t addr, kva; 3617 vm_offset_t pa; 3618 int pidx, i; 3619 struct vm_page *m; 3620 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3621 3622 GIANT_REQUIRED; 3623 3624 if ((bp->b_flags & B_PHYS) == 0) 3625 panic("vmapbuf"); 3626 3627 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3628 addr < bp->b_data + bp->b_bufsize; 3629 addr += PAGE_SIZE, pidx++) { 3630 /* 3631 * Do the vm_fault if needed; do the copy-on-write thing 3632 * when reading stuff off device into memory. 3633 * 3634 * NOTE! Must use pmap_extract() because addr may be in 3635 * the userland address space, and kextract is only guarenteed 3636 * to work for the kernland address space (see: sparc64 port). 3637 */ 3638retry: 3639 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data, 3640 (bp->b_iocmd == BIO_READ) ? 3641 (VM_PROT_READ|VM_PROT_WRITE) : VM_PROT_READ); 3642 if (i < 0) { 3643 printf("vmapbuf: warning, bad user address during I/O\n"); 3644 vm_page_lock_queues(); 3645 for (i = 0; i < pidx; ++i) { 3646 vm_page_unhold(bp->b_pages[i]); 3647 bp->b_pages[i] = NULL; 3648 } 3649 vm_page_unlock_queues(); 3650 return(-1); 3651 } 3652 pa = trunc_page(pmap_extract(pmap, (vm_offset_t) addr)); 3653 if (pa == 0) { 3654 printf("vmapbuf: warning, race against user address during I/O"); 3655 goto retry; 3656 } 3657 m = PHYS_TO_VM_PAGE(pa); 3658 vm_page_lock_queues(); 3659 vm_page_hold(m); 3660 vm_page_unlock_queues(); 3661 bp->b_pages[pidx] = m; 3662 } 3663 if (pidx > btoc(MAXPHYS)) 3664 panic("vmapbuf: mapped more than MAXPHYS"); 3665 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3666 3667 kva = bp->b_saveaddr; 3668 bp->b_npages = pidx; 3669 bp->b_saveaddr = bp->b_data; 3670 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3671 return(0); 3672} 3673 3674/* 3675 * Free the io map PTEs associated with this IO operation. 3676 * We also invalidate the TLB entries and restore the original b_addr. 3677 */ 3678void 3679vunmapbuf(struct buf *bp) 3680{ 3681 int pidx; 3682 int npages; 3683 3684 GIANT_REQUIRED; 3685 3686 if ((bp->b_flags & B_PHYS) == 0) 3687 panic("vunmapbuf"); 3688 3689 npages = bp->b_npages; 3690 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3691 npages); 3692 vm_page_lock_queues(); 3693 for (pidx = 0; pidx < npages; pidx++) 3694 vm_page_unhold(bp->b_pages[pidx]); 3695 vm_page_unlock_queues(); 3696 3697 bp->b_data = bp->b_saveaddr; 3698} 3699 3700#include "opt_ddb.h" 3701#ifdef DDB 3702#include <ddb/ddb.h> 3703 3704/* DDB command to show buffer data */ 3705DB_SHOW_COMMAND(buffer, db_show_buffer) 3706{ 3707 /* get args */ 3708 struct buf *bp = (struct buf *)addr; 3709 3710 if (!have_addr) { 3711 db_printf("usage: show buffer <addr>\n"); 3712 return; 3713 } 3714 3715 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3716 db_printf( 3717 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3718 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd, b_pblkno = %jd\n", 3719 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3720 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3721 (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno); 3722 if (bp->b_npages) { 3723 int i; 3724 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3725 for (i = 0; i < bp->b_npages; i++) { 3726 vm_page_t m; 3727 m = bp->b_pages[i]; 3728 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3729 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3730 if ((i + 1) < bp->b_npages) 3731 db_printf(","); 3732 } 3733 db_printf("\n"); 3734 } 3735} 3736#endif /* DDB */ 3737