vfs_bio.c revision 121218
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 15/* 16 * this file contains a new buffer I/O scheme implementing a coherent 17 * VM object and buffer cache scheme. Pains have been taken to make 18 * sure that the performance degradation associated with schemes such 19 * as this is not realized. 20 * 21 * Author: John S. Dyson 22 * Significant help during the development and debugging phases 23 * had been provided by David Greenman, also of the FreeBSD core team. 24 * 25 * see man buf(9) for more info. 26 */ 27 28#include <sys/cdefs.h> 29__FBSDID("$FreeBSD: head/sys/kern/vfs_bio.c 121218 2003-10-18 17:59:02Z phk $"); 30 31#include <sys/param.h> 32#include <sys/systm.h> 33#include <sys/bio.h> 34#include <sys/conf.h> 35#include <sys/buf.h> 36#include <sys/devicestat.h> 37#include <sys/eventhandler.h> 38#include <sys/lock.h> 39#include <sys/malloc.h> 40#include <sys/mount.h> 41#include <sys/mutex.h> 42#include <sys/kernel.h> 43#include <sys/kthread.h> 44#include <sys/proc.h> 45#include <sys/resourcevar.h> 46#include <sys/sysctl.h> 47#include <sys/vmmeter.h> 48#include <sys/vnode.h> 49#include <vm/vm.h> 50#include <vm/vm_param.h> 51#include <vm/vm_kern.h> 52#include <vm/vm_pageout.h> 53#include <vm/vm_page.h> 54#include <vm/vm_object.h> 55#include <vm/vm_extern.h> 56#include <vm/vm_map.h> 57#include "opt_directio.h" 58#include "opt_swap.h" 59 60static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 61 62struct bio_ops bioops; /* I/O operation notification */ 63 64struct buf_ops buf_ops_bio = { 65 "buf_ops_bio", 66 bwrite 67}; 68 69/* 70 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 71 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 72 */ 73struct buf *buf; /* buffer header pool */ 74 75static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 76 vm_offset_t to); 77static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 78 vm_offset_t to); 79static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 80 int pageno, vm_page_t m); 81static void vfs_clean_pages(struct buf * bp); 82static void vfs_setdirty(struct buf *bp); 83static void vfs_vmio_release(struct buf *bp); 84static void vfs_backgroundwritedone(struct buf *bp); 85static int vfs_bio_clcheck(struct vnode *vp, int size, 86 daddr_t lblkno, daddr_t blkno); 87static int flushbufqueues(int flushdeps); 88static void buf_daemon(void); 89void bremfreel(struct buf * bp); 90 91int vmiodirenable = TRUE; 92SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 93 "Use the VM system for directory writes"); 94int runningbufspace; 95SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 96 "Amount of presently outstanding async buffer io"); 97static int bufspace; 98SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 99 "KVA memory used for bufs"); 100static int maxbufspace; 101SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 102 "Maximum allowed value of bufspace (including buf_daemon)"); 103static int bufmallocspace; 104SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 105 "Amount of malloced memory for buffers"); 106static int maxbufmallocspace; 107SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 108 "Maximum amount of malloced memory for buffers"); 109static int lobufspace; 110SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 111 "Minimum amount of buffers we want to have"); 112static int hibufspace; 113SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 114 "Maximum allowed value of bufspace (excluding buf_daemon)"); 115static int bufreusecnt; 116SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 117 "Number of times we have reused a buffer"); 118static int buffreekvacnt; 119SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 120 "Number of times we have freed the KVA space from some buffer"); 121static int bufdefragcnt; 122SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 123 "Number of times we have had to repeat buffer allocation to defragment"); 124static int lorunningspace; 125SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 126 "Minimum preferred space used for in-progress I/O"); 127static int hirunningspace; 128SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 129 "Maximum amount of space to use for in-progress I/O"); 130static int dirtybufferflushes; 131SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 132 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 133static int altbufferflushes; 134SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 135 0, "Number of fsync flushes to limit dirty buffers"); 136static int recursiveflushes; 137SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 138 0, "Number of flushes skipped due to being recursive"); 139static int numdirtybuffers; 140SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 141 "Number of buffers that are dirty (has unwritten changes) at the moment"); 142static int lodirtybuffers; 143SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 144 "How many buffers we want to have free before bufdaemon can sleep"); 145static int hidirtybuffers; 146SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 147 "When the number of dirty buffers is considered severe"); 148static int dirtybufthresh; 149SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 150 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 151static int numfreebuffers; 152SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 153 "Number of free buffers"); 154static int lofreebuffers; 155SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 156 "XXX Unused"); 157static int hifreebuffers; 158SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 159 "XXX Complicatedly unused"); 160static int getnewbufcalls; 161SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 162 "Number of calls to getnewbuf"); 163static int getnewbufrestarts; 164SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 165 "Number of times getnewbuf has had to restart a buffer aquisition"); 166static int dobkgrdwrite = 1; 167SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 168 "Do background writes (honoring the BV_BKGRDWRITE flag)?"); 169 170/* 171 * Wakeup point for bufdaemon, as well as indicator of whether it is already 172 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 173 * is idling. 174 */ 175static int bd_request; 176 177/* 178 * This lock synchronizes access to bd_request. 179 */ 180static struct mtx bdlock; 181 182/* 183 * bogus page -- for I/O to/from partially complete buffers 184 * this is a temporary solution to the problem, but it is not 185 * really that bad. it would be better to split the buffer 186 * for input in the case of buffers partially already in memory, 187 * but the code is intricate enough already. 188 */ 189vm_page_t bogus_page; 190 191/* 192 * Synchronization (sleep/wakeup) variable for active buffer space requests. 193 * Set when wait starts, cleared prior to wakeup(). 194 * Used in runningbufwakeup() and waitrunningbufspace(). 195 */ 196static int runningbufreq; 197 198/* 199 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 200 * waitrunningbufspace(). 201 */ 202static struct mtx rbreqlock; 203 204/* 205 * Synchronization (sleep/wakeup) variable for buffer requests. 206 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 207 * by and/or. 208 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 209 * getnewbuf(), and getblk(). 210 */ 211static int needsbuffer; 212 213/* 214 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 215 */ 216static struct mtx nblock; 217 218/* 219 * Lock that protects against bwait()/bdone()/B_DONE races. 220 */ 221 222static struct mtx bdonelock; 223 224/* 225 * Definitions for the buffer free lists. 226 */ 227#define BUFFER_QUEUES 5 /* number of free buffer queues */ 228 229#define QUEUE_NONE 0 /* on no queue */ 230#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 231#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 232#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 233#define QUEUE_EMPTY 4 /* empty buffer headers */ 234 235/* Queues for free buffers with various properties */ 236static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 237 238/* Lock for the bufqueues */ 239static struct mtx bqlock; 240 241/* 242 * Single global constant for BUF_WMESG, to avoid getting multiple references. 243 * buf_wmesg is referred from macros. 244 */ 245const char *buf_wmesg = BUF_WMESG; 246 247#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 248#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 249#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 250#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 251 252#ifdef DIRECTIO 253extern void ffs_rawread_setup(void); 254#endif /* DIRECTIO */ 255/* 256 * numdirtywakeup: 257 * 258 * If someone is blocked due to there being too many dirty buffers, 259 * and numdirtybuffers is now reasonable, wake them up. 260 */ 261 262static __inline void 263numdirtywakeup(int level) 264{ 265 if (numdirtybuffers <= level) { 266 mtx_lock(&nblock); 267 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 268 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 269 wakeup(&needsbuffer); 270 } 271 mtx_unlock(&nblock); 272 } 273} 274 275/* 276 * bufspacewakeup: 277 * 278 * Called when buffer space is potentially available for recovery. 279 * getnewbuf() will block on this flag when it is unable to free 280 * sufficient buffer space. Buffer space becomes recoverable when 281 * bp's get placed back in the queues. 282 */ 283 284static __inline void 285bufspacewakeup(void) 286{ 287 /* 288 * If someone is waiting for BUF space, wake them up. Even 289 * though we haven't freed the kva space yet, the waiting 290 * process will be able to now. 291 */ 292 mtx_lock(&nblock); 293 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 294 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 295 wakeup(&needsbuffer); 296 } 297 mtx_unlock(&nblock); 298} 299 300/* 301 * runningbufwakeup() - in-progress I/O accounting. 302 * 303 */ 304static __inline void 305runningbufwakeup(struct buf *bp) 306{ 307 if (bp->b_runningbufspace) { 308 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 309 bp->b_runningbufspace = 0; 310 mtx_lock(&rbreqlock); 311 if (runningbufreq && runningbufspace <= lorunningspace) { 312 runningbufreq = 0; 313 wakeup(&runningbufreq); 314 } 315 mtx_unlock(&rbreqlock); 316 } 317} 318 319/* 320 * bufcountwakeup: 321 * 322 * Called when a buffer has been added to one of the free queues to 323 * account for the buffer and to wakeup anyone waiting for free buffers. 324 * This typically occurs when large amounts of metadata are being handled 325 * by the buffer cache ( else buffer space runs out first, usually ). 326 */ 327 328static __inline void 329bufcountwakeup(void) 330{ 331 atomic_add_int(&numfreebuffers, 1); 332 mtx_lock(&nblock); 333 if (needsbuffer) { 334 needsbuffer &= ~VFS_BIO_NEED_ANY; 335 if (numfreebuffers >= hifreebuffers) 336 needsbuffer &= ~VFS_BIO_NEED_FREE; 337 wakeup(&needsbuffer); 338 } 339 mtx_unlock(&nblock); 340} 341 342/* 343 * waitrunningbufspace() 344 * 345 * runningbufspace is a measure of the amount of I/O currently 346 * running. This routine is used in async-write situations to 347 * prevent creating huge backups of pending writes to a device. 348 * Only asynchronous writes are governed by this function. 349 * 350 * Reads will adjust runningbufspace, but will not block based on it. 351 * The read load has a side effect of reducing the allowed write load. 352 * 353 * This does NOT turn an async write into a sync write. It waits 354 * for earlier writes to complete and generally returns before the 355 * caller's write has reached the device. 356 */ 357static __inline void 358waitrunningbufspace(void) 359{ 360 mtx_lock(&rbreqlock); 361 while (runningbufspace > hirunningspace) { 362 ++runningbufreq; 363 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 364 } 365 mtx_unlock(&rbreqlock); 366} 367 368 369/* 370 * vfs_buf_test_cache: 371 * 372 * Called when a buffer is extended. This function clears the B_CACHE 373 * bit if the newly extended portion of the buffer does not contain 374 * valid data. 375 */ 376static __inline__ 377void 378vfs_buf_test_cache(struct buf *bp, 379 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 380 vm_page_t m) 381{ 382 GIANT_REQUIRED; 383 384 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 385 if (bp->b_flags & B_CACHE) { 386 int base = (foff + off) & PAGE_MASK; 387 if (vm_page_is_valid(m, base, size) == 0) 388 bp->b_flags &= ~B_CACHE; 389 } 390} 391 392/* Wake up the buffer deamon if necessary */ 393static __inline__ 394void 395bd_wakeup(int dirtybuflevel) 396{ 397 mtx_lock(&bdlock); 398 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 399 bd_request = 1; 400 wakeup(&bd_request); 401 } 402 mtx_unlock(&bdlock); 403} 404 405/* 406 * bd_speedup - speedup the buffer cache flushing code 407 */ 408 409static __inline__ 410void 411bd_speedup(void) 412{ 413 bd_wakeup(1); 414} 415 416/* 417 * Calculating buffer cache scaling values and reserve space for buffer 418 * headers. This is called during low level kernel initialization and 419 * may be called more then once. We CANNOT write to the memory area 420 * being reserved at this time. 421 */ 422caddr_t 423kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 424{ 425 /* 426 * physmem_est is in pages. Convert it to kilobytes (assumes 427 * PAGE_SIZE is >= 1K) 428 */ 429 physmem_est = physmem_est * (PAGE_SIZE / 1024); 430 431 /* 432 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 433 * For the first 64MB of ram nominally allocate sufficient buffers to 434 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 435 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 436 * the buffer cache we limit the eventual kva reservation to 437 * maxbcache bytes. 438 * 439 * factor represents the 1/4 x ram conversion. 440 */ 441 if (nbuf == 0) { 442 int factor = 4 * BKVASIZE / 1024; 443 444 nbuf = 50; 445 if (physmem_est > 4096) 446 nbuf += min((physmem_est - 4096) / factor, 447 65536 / factor); 448 if (physmem_est > 65536) 449 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 450 451 if (maxbcache && nbuf > maxbcache / BKVASIZE) 452 nbuf = maxbcache / BKVASIZE; 453 } 454 455#if 0 456 /* 457 * Do not allow the buffer_map to be more then 1/2 the size of the 458 * kernel_map. 459 */ 460 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 461 (BKVASIZE * 2)) { 462 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 463 (BKVASIZE * 2); 464 printf("Warning: nbufs capped at %d\n", nbuf); 465 } 466#endif 467 468 /* 469 * swbufs are used as temporary holders for I/O, such as paging I/O. 470 * We have no less then 16 and no more then 256. 471 */ 472 nswbuf = max(min(nbuf/4, 256), 16); 473#ifdef NSWBUF_MIN 474 if (nswbuf < NSWBUF_MIN) 475 nswbuf = NSWBUF_MIN; 476#endif 477#ifdef DIRECTIO 478 ffs_rawread_setup(); 479#endif 480 481 /* 482 * Reserve space for the buffer cache buffers 483 */ 484 swbuf = (void *)v; 485 v = (caddr_t)(swbuf + nswbuf); 486 buf = (void *)v; 487 v = (caddr_t)(buf + nbuf); 488 489 return(v); 490} 491 492/* Initialize the buffer subsystem. Called before use of any buffers. */ 493void 494bufinit(void) 495{ 496 struct buf *bp; 497 int i; 498 499 GIANT_REQUIRED; 500 501 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 502 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 503 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 504 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 505 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 506 507 /* next, make a null set of free lists */ 508 for (i = 0; i < BUFFER_QUEUES; i++) 509 TAILQ_INIT(&bufqueues[i]); 510 511 /* finally, initialize each buffer header and stick on empty q */ 512 for (i = 0; i < nbuf; i++) { 513 bp = &buf[i]; 514 bzero(bp, sizeof *bp); 515 bp->b_flags = B_INVAL; /* we're just an empty header */ 516 bp->b_dev = NODEV; 517 bp->b_rcred = NOCRED; 518 bp->b_wcred = NOCRED; 519 bp->b_qindex = QUEUE_EMPTY; 520 bp->b_vflags = 0; 521 bp->b_xflags = 0; 522 LIST_INIT(&bp->b_dep); 523 BUF_LOCKINIT(bp); 524 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 525 } 526 527 /* 528 * maxbufspace is the absolute maximum amount of buffer space we are 529 * allowed to reserve in KVM and in real terms. The absolute maximum 530 * is nominally used by buf_daemon. hibufspace is the nominal maximum 531 * used by most other processes. The differential is required to 532 * ensure that buf_daemon is able to run when other processes might 533 * be blocked waiting for buffer space. 534 * 535 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 536 * this may result in KVM fragmentation which is not handled optimally 537 * by the system. 538 */ 539 maxbufspace = nbuf * BKVASIZE; 540 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 541 lobufspace = hibufspace - MAXBSIZE; 542 543 lorunningspace = 512 * 1024; 544 hirunningspace = 1024 * 1024; 545 546/* 547 * Limit the amount of malloc memory since it is wired permanently into 548 * the kernel space. Even though this is accounted for in the buffer 549 * allocation, we don't want the malloced region to grow uncontrolled. 550 * The malloc scheme improves memory utilization significantly on average 551 * (small) directories. 552 */ 553 maxbufmallocspace = hibufspace / 20; 554 555/* 556 * Reduce the chance of a deadlock occuring by limiting the number 557 * of delayed-write dirty buffers we allow to stack up. 558 */ 559 hidirtybuffers = nbuf / 4 + 20; 560 dirtybufthresh = hidirtybuffers * 9 / 10; 561 numdirtybuffers = 0; 562/* 563 * To support extreme low-memory systems, make sure hidirtybuffers cannot 564 * eat up all available buffer space. This occurs when our minimum cannot 565 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 566 * BKVASIZE'd (8K) buffers. 567 */ 568 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 569 hidirtybuffers >>= 1; 570 } 571 lodirtybuffers = hidirtybuffers / 2; 572 573/* 574 * Try to keep the number of free buffers in the specified range, 575 * and give special processes (e.g. like buf_daemon) access to an 576 * emergency reserve. 577 */ 578 lofreebuffers = nbuf / 18 + 5; 579 hifreebuffers = 2 * lofreebuffers; 580 numfreebuffers = nbuf; 581 582/* 583 * Maximum number of async ops initiated per buf_daemon loop. This is 584 * somewhat of a hack at the moment, we really need to limit ourselves 585 * based on the number of bytes of I/O in-transit that were initiated 586 * from buf_daemon. 587 */ 588 589 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 590 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 591} 592 593/* 594 * bfreekva() - free the kva allocation for a buffer. 595 * 596 * Must be called at splbio() or higher as this is the only locking for 597 * buffer_map. 598 * 599 * Since this call frees up buffer space, we call bufspacewakeup(). 600 */ 601static void 602bfreekva(struct buf * bp) 603{ 604 GIANT_REQUIRED; 605 606 if (bp->b_kvasize) { 607 atomic_add_int(&buffreekvacnt, 1); 608 atomic_subtract_int(&bufspace, bp->b_kvasize); 609 vm_map_delete(buffer_map, 610 (vm_offset_t) bp->b_kvabase, 611 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 612 ); 613 bp->b_kvasize = 0; 614 bufspacewakeup(); 615 } 616} 617 618/* 619 * bremfree: 620 * 621 * Remove the buffer from the appropriate free list. 622 */ 623void 624bremfree(struct buf * bp) 625{ 626 mtx_lock(&bqlock); 627 bremfreel(bp); 628 mtx_unlock(&bqlock); 629} 630 631void 632bremfreel(struct buf * bp) 633{ 634 int s = splbio(); 635 int old_qindex = bp->b_qindex; 636 637 GIANT_REQUIRED; 638 639 if (bp->b_qindex != QUEUE_NONE) { 640 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 641 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 642 bp->b_qindex = QUEUE_NONE; 643 } else { 644 if (BUF_REFCNT(bp) <= 1) 645 panic("bremfree: removing a buffer not on a queue"); 646 } 647 648 /* 649 * Fixup numfreebuffers count. If the buffer is invalid or not 650 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 651 * the buffer was free and we must decrement numfreebuffers. 652 */ 653 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 654 switch(old_qindex) { 655 case QUEUE_DIRTY: 656 case QUEUE_CLEAN: 657 case QUEUE_EMPTY: 658 case QUEUE_EMPTYKVA: 659 atomic_subtract_int(&numfreebuffers, 1); 660 break; 661 default: 662 break; 663 } 664 } 665 splx(s); 666} 667 668 669/* 670 * Get a buffer with the specified data. Look in the cache first. We 671 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 672 * is set, the buffer is valid and we do not have to do anything ( see 673 * getblk() ). This is really just a special case of breadn(). 674 */ 675int 676bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 677 struct buf ** bpp) 678{ 679 680 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 681} 682 683/* 684 * Operates like bread, but also starts asynchronous I/O on 685 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 686 * to initiating I/O . If B_CACHE is set, the buffer is valid 687 * and we do not have to do anything. 688 */ 689int 690breadn(struct vnode * vp, daddr_t blkno, int size, 691 daddr_t * rablkno, int *rabsize, 692 int cnt, struct ucred * cred, struct buf ** bpp) 693{ 694 struct buf *bp, *rabp; 695 int i; 696 int rv = 0, readwait = 0; 697 698 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 699 700 /* if not found in cache, do some I/O */ 701 if ((bp->b_flags & B_CACHE) == 0) { 702 if (curthread != PCPU_GET(idlethread)) 703 curthread->td_proc->p_stats->p_ru.ru_inblock++; 704 bp->b_iocmd = BIO_READ; 705 bp->b_flags &= ~B_INVAL; 706 bp->b_ioflags &= ~BIO_ERROR; 707 if (bp->b_rcred == NOCRED && cred != NOCRED) 708 bp->b_rcred = crhold(cred); 709 vfs_busy_pages(bp, 0); 710 if (vp->v_type == VCHR) 711 VOP_SPECSTRATEGY(vp, bp); 712 else 713 VOP_STRATEGY(vp, bp); 714 ++readwait; 715 } 716 717 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 718 if (inmem(vp, *rablkno)) 719 continue; 720 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 721 722 if ((rabp->b_flags & B_CACHE) == 0) { 723 if (curthread != PCPU_GET(idlethread)) 724 curthread->td_proc->p_stats->p_ru.ru_inblock++; 725 rabp->b_flags |= B_ASYNC; 726 rabp->b_flags &= ~B_INVAL; 727 rabp->b_ioflags &= ~BIO_ERROR; 728 rabp->b_iocmd = BIO_READ; 729 if (rabp->b_rcred == NOCRED && cred != NOCRED) 730 rabp->b_rcred = crhold(cred); 731 vfs_busy_pages(rabp, 0); 732 BUF_KERNPROC(rabp); 733 if (vp->v_type == VCHR) 734 VOP_SPECSTRATEGY(vp, rabp); 735 else 736 VOP_STRATEGY(vp, rabp); 737 } else { 738 brelse(rabp); 739 } 740 } 741 742 if (readwait) { 743 rv = bufwait(bp); 744 } 745 return (rv); 746} 747 748/* 749 * Write, release buffer on completion. (Done by iodone 750 * if async). Do not bother writing anything if the buffer 751 * is invalid. 752 * 753 * Note that we set B_CACHE here, indicating that buffer is 754 * fully valid and thus cacheable. This is true even of NFS 755 * now so we set it generally. This could be set either here 756 * or in biodone() since the I/O is synchronous. We put it 757 * here. 758 */ 759 760int 761bwrite(struct buf * bp) 762{ 763 int oldflags, s; 764 struct buf *newbp; 765 766 if (bp->b_flags & B_INVAL) { 767 brelse(bp); 768 return (0); 769 } 770 771 oldflags = bp->b_flags; 772 773 if (BUF_REFCNT(bp) == 0) 774 panic("bwrite: buffer is not busy???"); 775 s = splbio(); 776 /* 777 * If a background write is already in progress, delay 778 * writing this block if it is asynchronous. Otherwise 779 * wait for the background write to complete. 780 */ 781 VI_LOCK(bp->b_vp); 782 if (bp->b_vflags & BV_BKGRDINPROG) { 783 if (bp->b_flags & B_ASYNC) { 784 VI_UNLOCK(bp->b_vp); 785 splx(s); 786 bdwrite(bp); 787 return (0); 788 } 789 bp->b_vflags |= BV_BKGRDWAIT; 790 msleep(&bp->b_xflags, VI_MTX(bp->b_vp), PRIBIO, "bwrbg", 0); 791 if (bp->b_vflags & BV_BKGRDINPROG) 792 panic("bwrite: still writing"); 793 } 794 VI_UNLOCK(bp->b_vp); 795 796 /* Mark the buffer clean */ 797 bundirty(bp); 798 799 /* 800 * If this buffer is marked for background writing and we 801 * do not have to wait for it, make a copy and write the 802 * copy so as to leave this buffer ready for further use. 803 * 804 * This optimization eats a lot of memory. If we have a page 805 * or buffer shortfall we can't do it. 806 */ 807 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 808 (bp->b_flags & B_ASYNC) && 809 !vm_page_count_severe() && 810 !buf_dirty_count_severe()) { 811 if (bp->b_iodone != NULL) { 812 printf("bp->b_iodone = %p\n", bp->b_iodone); 813 panic("bwrite: need chained iodone"); 814 } 815 816 /* get a new block */ 817 newbp = geteblk(bp->b_bufsize); 818 819 /* 820 * set it to be identical to the old block. We have to 821 * set b_lblkno and BKGRDMARKER before calling bgetvp() 822 * to avoid confusing the splay tree and gbincore(). 823 */ 824 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 825 newbp->b_lblkno = bp->b_lblkno; 826 newbp->b_xflags |= BX_BKGRDMARKER; 827 VI_LOCK(bp->b_vp); 828 bp->b_vflags |= BV_BKGRDINPROG; 829 bgetvp(bp->b_vp, newbp); 830 VI_UNLOCK(bp->b_vp); 831 newbp->b_blkno = bp->b_blkno; 832 newbp->b_offset = bp->b_offset; 833 newbp->b_iodone = vfs_backgroundwritedone; 834 newbp->b_flags |= B_ASYNC; 835 newbp->b_flags &= ~B_INVAL; 836 837 /* move over the dependencies */ 838 if (LIST_FIRST(&bp->b_dep) != NULL) 839 buf_movedeps(bp, newbp); 840 841 /* 842 * Initiate write on the copy, release the original to 843 * the B_LOCKED queue so that it cannot go away until 844 * the background write completes. If not locked it could go 845 * away and then be reconstituted while it was being written. 846 * If the reconstituted buffer were written, we could end up 847 * with two background copies being written at the same time. 848 */ 849 bqrelse(bp); 850 bp = newbp; 851 } 852 853 bp->b_flags &= ~B_DONE; 854 bp->b_ioflags &= ~BIO_ERROR; 855 bp->b_flags |= B_WRITEINPROG | B_CACHE; 856 bp->b_iocmd = BIO_WRITE; 857 858 VI_LOCK(bp->b_vp); 859 bp->b_vp->v_numoutput++; 860 VI_UNLOCK(bp->b_vp); 861 vfs_busy_pages(bp, 1); 862 863 /* 864 * Normal bwrites pipeline writes 865 */ 866 bp->b_runningbufspace = bp->b_bufsize; 867 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 868 869 if (curthread != PCPU_GET(idlethread)) 870 curthread->td_proc->p_stats->p_ru.ru_oublock++; 871 splx(s); 872 if (oldflags & B_ASYNC) 873 BUF_KERNPROC(bp); 874 if (bp->b_vp->v_type == VCHR) 875 VOP_SPECSTRATEGY(bp->b_vp, bp); 876 else 877 VOP_STRATEGY(bp->b_vp, bp); 878 879 if ((oldflags & B_ASYNC) == 0) { 880 int rtval = bufwait(bp); 881 brelse(bp); 882 return (rtval); 883 } else { 884 /* 885 * don't allow the async write to saturate the I/O 886 * system. We will not deadlock here because 887 * we are blocking waiting for I/O that is already in-progress 888 * to complete. 889 */ 890 waitrunningbufspace(); 891 } 892 893 return (0); 894} 895 896/* 897 * Complete a background write started from bwrite. 898 */ 899static void 900vfs_backgroundwritedone(bp) 901 struct buf *bp; 902{ 903 struct buf *origbp; 904 905 /* 906 * Find the original buffer that we are writing. 907 */ 908 VI_LOCK(bp->b_vp); 909 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 910 panic("backgroundwritedone: lost buffer"); 911 912 /* 913 * Clear the BV_BKGRDINPROG flag in the original buffer 914 * and awaken it if it is waiting for the write to complete. 915 * If BV_BKGRDINPROG is not set in the original buffer it must 916 * have been released and re-instantiated - which is not legal. 917 */ 918 KASSERT((origbp->b_vflags & BV_BKGRDINPROG), 919 ("backgroundwritedone: lost buffer2")); 920 origbp->b_vflags &= ~BV_BKGRDINPROG; 921 if (origbp->b_vflags & BV_BKGRDWAIT) { 922 origbp->b_vflags &= ~BV_BKGRDWAIT; 923 wakeup(&origbp->b_xflags); 924 } 925 VI_UNLOCK(bp->b_vp); 926 /* 927 * Process dependencies then return any unfinished ones. 928 */ 929 if (LIST_FIRST(&bp->b_dep) != NULL) 930 buf_complete(bp); 931 if (LIST_FIRST(&bp->b_dep) != NULL) 932 buf_movedeps(bp, origbp); 933 934 /* 935 * This buffer is marked B_NOCACHE, so when it is released 936 * by biodone, it will be tossed. We mark it with BIO_READ 937 * to avoid biodone doing a second vwakeup. 938 */ 939 bp->b_flags |= B_NOCACHE; 940 bp->b_iocmd = BIO_READ; 941 bp->b_flags &= ~(B_CACHE | B_DONE); 942 bp->b_iodone = 0; 943 bufdone(bp); 944} 945 946/* 947 * Delayed write. (Buffer is marked dirty). Do not bother writing 948 * anything if the buffer is marked invalid. 949 * 950 * Note that since the buffer must be completely valid, we can safely 951 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 952 * biodone() in order to prevent getblk from writing the buffer 953 * out synchronously. 954 */ 955void 956bdwrite(struct buf * bp) 957{ 958 struct thread *td = curthread; 959 struct vnode *vp; 960 struct buf *nbp; 961 962 GIANT_REQUIRED; 963 964 if (BUF_REFCNT(bp) == 0) 965 panic("bdwrite: buffer is not busy"); 966 967 if (bp->b_flags & B_INVAL) { 968 brelse(bp); 969 return; 970 } 971 972 /* 973 * If we have too many dirty buffers, don't create any more. 974 * If we are wildly over our limit, then force a complete 975 * cleanup. Otherwise, just keep the situation from getting 976 * out of control. Note that we have to avoid a recursive 977 * disaster and not try to clean up after our own cleanup! 978 */ 979 vp = bp->b_vp; 980 VI_LOCK(vp); 981 if (td->td_proc->p_flag & P_COWINPROGRESS) { 982 recursiveflushes++; 983 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) { 984 VI_UNLOCK(vp); 985 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 986 VI_LOCK(vp); 987 altbufferflushes++; 988 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) { 989 /* 990 * Try to find a buffer to flush. 991 */ 992 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) { 993 if ((nbp->b_vflags & BV_BKGRDINPROG) || 994 buf_countdeps(nbp, 0) || 995 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 996 continue; 997 if (bp == nbp) 998 panic("bdwrite: found ourselves"); 999 VI_UNLOCK(vp); 1000 if (nbp->b_flags & B_CLUSTEROK) { 1001 vfs_bio_awrite(nbp); 1002 } else { 1003 bremfree(nbp); 1004 bawrite(nbp); 1005 } 1006 VI_LOCK(vp); 1007 dirtybufferflushes++; 1008 break; 1009 } 1010 } 1011 VI_UNLOCK(vp); 1012 1013 bdirty(bp); 1014 /* 1015 * Set B_CACHE, indicating that the buffer is fully valid. This is 1016 * true even of NFS now. 1017 */ 1018 bp->b_flags |= B_CACHE; 1019 1020 /* 1021 * This bmap keeps the system from needing to do the bmap later, 1022 * perhaps when the system is attempting to do a sync. Since it 1023 * is likely that the indirect block -- or whatever other datastructure 1024 * that the filesystem needs is still in memory now, it is a good 1025 * thing to do this. Note also, that if the pageout daemon is 1026 * requesting a sync -- there might not be enough memory to do 1027 * the bmap then... So, this is important to do. 1028 */ 1029 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1030 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1031 } 1032 1033 /* 1034 * Set the *dirty* buffer range based upon the VM system dirty pages. 1035 */ 1036 vfs_setdirty(bp); 1037 1038 /* 1039 * We need to do this here to satisfy the vnode_pager and the 1040 * pageout daemon, so that it thinks that the pages have been 1041 * "cleaned". Note that since the pages are in a delayed write 1042 * buffer -- the VFS layer "will" see that the pages get written 1043 * out on the next sync, or perhaps the cluster will be completed. 1044 */ 1045 vfs_clean_pages(bp); 1046 bqrelse(bp); 1047 1048 /* 1049 * Wakeup the buffer flushing daemon if we have a lot of dirty 1050 * buffers (midpoint between our recovery point and our stall 1051 * point). 1052 */ 1053 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1054 1055 /* 1056 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1057 * due to the softdep code. 1058 */ 1059} 1060 1061/* 1062 * bdirty: 1063 * 1064 * Turn buffer into delayed write request. We must clear BIO_READ and 1065 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1066 * itself to properly update it in the dirty/clean lists. We mark it 1067 * B_DONE to ensure that any asynchronization of the buffer properly 1068 * clears B_DONE ( else a panic will occur later ). 1069 * 1070 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1071 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1072 * should only be called if the buffer is known-good. 1073 * 1074 * Since the buffer is not on a queue, we do not update the numfreebuffers 1075 * count. 1076 * 1077 * Must be called at splbio(). 1078 * The buffer must be on QUEUE_NONE. 1079 */ 1080void 1081bdirty(bp) 1082 struct buf *bp; 1083{ 1084 KASSERT(bp->b_qindex == QUEUE_NONE, 1085 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1086 bp->b_flags &= ~(B_RELBUF); 1087 bp->b_iocmd = BIO_WRITE; 1088 1089 if ((bp->b_flags & B_DELWRI) == 0) { 1090 bp->b_flags |= B_DONE | B_DELWRI; 1091 reassignbuf(bp, bp->b_vp); 1092 atomic_add_int(&numdirtybuffers, 1); 1093 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1094 } 1095} 1096 1097/* 1098 * bundirty: 1099 * 1100 * Clear B_DELWRI for buffer. 1101 * 1102 * Since the buffer is not on a queue, we do not update the numfreebuffers 1103 * count. 1104 * 1105 * Must be called at splbio(). 1106 * The buffer must be on QUEUE_NONE. 1107 */ 1108 1109void 1110bundirty(bp) 1111 struct buf *bp; 1112{ 1113 KASSERT(bp->b_qindex == QUEUE_NONE, 1114 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1115 1116 if (bp->b_flags & B_DELWRI) { 1117 bp->b_flags &= ~B_DELWRI; 1118 reassignbuf(bp, bp->b_vp); 1119 atomic_subtract_int(&numdirtybuffers, 1); 1120 numdirtywakeup(lodirtybuffers); 1121 } 1122 /* 1123 * Since it is now being written, we can clear its deferred write flag. 1124 */ 1125 bp->b_flags &= ~B_DEFERRED; 1126} 1127 1128/* 1129 * bawrite: 1130 * 1131 * Asynchronous write. Start output on a buffer, but do not wait for 1132 * it to complete. The buffer is released when the output completes. 1133 * 1134 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1135 * B_INVAL buffers. Not us. 1136 */ 1137void 1138bawrite(struct buf * bp) 1139{ 1140 bp->b_flags |= B_ASYNC; 1141 (void) BUF_WRITE(bp); 1142} 1143 1144/* 1145 * bwillwrite: 1146 * 1147 * Called prior to the locking of any vnodes when we are expecting to 1148 * write. We do not want to starve the buffer cache with too many 1149 * dirty buffers so we block here. By blocking prior to the locking 1150 * of any vnodes we attempt to avoid the situation where a locked vnode 1151 * prevents the various system daemons from flushing related buffers. 1152 */ 1153 1154void 1155bwillwrite(void) 1156{ 1157 if (numdirtybuffers >= hidirtybuffers) { 1158 int s; 1159 1160 mtx_lock(&Giant); 1161 s = splbio(); 1162 mtx_lock(&nblock); 1163 while (numdirtybuffers >= hidirtybuffers) { 1164 bd_wakeup(1); 1165 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1166 msleep(&needsbuffer, &nblock, 1167 (PRIBIO + 4), "flswai", 0); 1168 } 1169 splx(s); 1170 mtx_unlock(&nblock); 1171 mtx_unlock(&Giant); 1172 } 1173} 1174 1175/* 1176 * Return true if we have too many dirty buffers. 1177 */ 1178int 1179buf_dirty_count_severe(void) 1180{ 1181 return(numdirtybuffers >= hidirtybuffers); 1182} 1183 1184/* 1185 * brelse: 1186 * 1187 * Release a busy buffer and, if requested, free its resources. The 1188 * buffer will be stashed in the appropriate bufqueue[] allowing it 1189 * to be accessed later as a cache entity or reused for other purposes. 1190 */ 1191void 1192brelse(struct buf * bp) 1193{ 1194 int s; 1195 1196 GIANT_REQUIRED; 1197 1198 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1199 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1200 1201 s = splbio(); 1202 1203 if (bp->b_iocmd == BIO_WRITE && 1204 (bp->b_ioflags & BIO_ERROR) && 1205 !(bp->b_flags & B_INVAL)) { 1206 /* 1207 * Failed write, redirty. Must clear BIO_ERROR to prevent 1208 * pages from being scrapped. If B_INVAL is set then 1209 * this case is not run and the next case is run to 1210 * destroy the buffer. B_INVAL can occur if the buffer 1211 * is outside the range supported by the underlying device. 1212 */ 1213 bp->b_ioflags &= ~BIO_ERROR; 1214 bdirty(bp); 1215 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1216 (bp->b_ioflags & BIO_ERROR) || 1217 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1218 /* 1219 * Either a failed I/O or we were asked to free or not 1220 * cache the buffer. 1221 */ 1222 bp->b_flags |= B_INVAL; 1223 if (LIST_FIRST(&bp->b_dep) != NULL) 1224 buf_deallocate(bp); 1225 if (bp->b_flags & B_DELWRI) { 1226 atomic_subtract_int(&numdirtybuffers, 1); 1227 numdirtywakeup(lodirtybuffers); 1228 } 1229 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1230 if ((bp->b_flags & B_VMIO) == 0) { 1231 if (bp->b_bufsize) 1232 allocbuf(bp, 0); 1233 if (bp->b_vp) 1234 brelvp(bp); 1235 } 1236 } 1237 1238 /* 1239 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1240 * is called with B_DELWRI set, the underlying pages may wind up 1241 * getting freed causing a previous write (bdwrite()) to get 'lost' 1242 * because pages associated with a B_DELWRI bp are marked clean. 1243 * 1244 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1245 * if B_DELWRI is set. 1246 * 1247 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1248 * on pages to return pages to the VM page queues. 1249 */ 1250 if (bp->b_flags & B_DELWRI) 1251 bp->b_flags &= ~B_RELBUF; 1252 else if (vm_page_count_severe()) { 1253 /* 1254 * XXX This lock may not be necessary since BKGRDINPROG 1255 * cannot be set while we hold the buf lock, it can only be 1256 * cleared if it is already pending. 1257 */ 1258 if (bp->b_vp) { 1259 VI_LOCK(bp->b_vp); 1260 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1261 bp->b_flags |= B_RELBUF; 1262 VI_UNLOCK(bp->b_vp); 1263 } else 1264 bp->b_flags |= B_RELBUF; 1265 } 1266 1267 /* 1268 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1269 * constituted, not even NFS buffers now. Two flags effect this. If 1270 * B_INVAL, the struct buf is invalidated but the VM object is kept 1271 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1272 * 1273 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1274 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1275 * buffer is also B_INVAL because it hits the re-dirtying code above. 1276 * 1277 * Normally we can do this whether a buffer is B_DELWRI or not. If 1278 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1279 * the commit state and we cannot afford to lose the buffer. If the 1280 * buffer has a background write in progress, we need to keep it 1281 * around to prevent it from being reconstituted and starting a second 1282 * background write. 1283 */ 1284 if ((bp->b_flags & B_VMIO) 1285 && !(bp->b_vp->v_mount != NULL && 1286 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1287 !vn_isdisk(bp->b_vp, NULL) && 1288 (bp->b_flags & B_DELWRI)) 1289 ) { 1290 1291 int i, j, resid; 1292 vm_page_t m; 1293 off_t foff; 1294 vm_pindex_t poff; 1295 vm_object_t obj; 1296 struct vnode *vp; 1297 1298 vp = bp->b_vp; 1299 obj = bp->b_object; 1300 1301 /* 1302 * Get the base offset and length of the buffer. Note that 1303 * in the VMIO case if the buffer block size is not 1304 * page-aligned then b_data pointer may not be page-aligned. 1305 * But our b_pages[] array *IS* page aligned. 1306 * 1307 * block sizes less then DEV_BSIZE (usually 512) are not 1308 * supported due to the page granularity bits (m->valid, 1309 * m->dirty, etc...). 1310 * 1311 * See man buf(9) for more information 1312 */ 1313 resid = bp->b_bufsize; 1314 foff = bp->b_offset; 1315 if (obj != NULL) 1316 VM_OBJECT_LOCK(obj); 1317 for (i = 0; i < bp->b_npages; i++) { 1318 int had_bogus = 0; 1319 1320 m = bp->b_pages[i]; 1321 vm_page_lock_queues(); 1322 vm_page_flag_clear(m, PG_ZERO); 1323 vm_page_unlock_queues(); 1324 1325 /* 1326 * If we hit a bogus page, fixup *all* the bogus pages 1327 * now. 1328 */ 1329 if (m == bogus_page) { 1330 poff = OFF_TO_IDX(bp->b_offset); 1331 had_bogus = 1; 1332 1333 for (j = i; j < bp->b_npages; j++) { 1334 vm_page_t mtmp; 1335 mtmp = bp->b_pages[j]; 1336 if (mtmp == bogus_page) { 1337 mtmp = vm_page_lookup(obj, poff + j); 1338 if (!mtmp) { 1339 panic("brelse: page missing\n"); 1340 } 1341 bp->b_pages[j] = mtmp; 1342 } 1343 } 1344 1345 if ((bp->b_flags & B_INVAL) == 0) { 1346 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1347 } 1348 m = bp->b_pages[i]; 1349 } 1350 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1351 int poffset = foff & PAGE_MASK; 1352 int presid = resid > (PAGE_SIZE - poffset) ? 1353 (PAGE_SIZE - poffset) : resid; 1354 1355 KASSERT(presid >= 0, ("brelse: extra page")); 1356 vm_page_lock_queues(); 1357 vm_page_set_invalid(m, poffset, presid); 1358 vm_page_unlock_queues(); 1359 if (had_bogus) 1360 printf("avoided corruption bug in bogus_page/brelse code\n"); 1361 } 1362 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1363 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1364 } 1365 if (obj != NULL) 1366 VM_OBJECT_UNLOCK(obj); 1367 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1368 vfs_vmio_release(bp); 1369 1370 } else if (bp->b_flags & B_VMIO) { 1371 1372 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1373 vfs_vmio_release(bp); 1374 } 1375 1376 } 1377 1378 if (bp->b_qindex != QUEUE_NONE) 1379 panic("brelse: free buffer onto another queue???"); 1380 if (BUF_REFCNT(bp) > 1) { 1381 /* do not release to free list */ 1382 BUF_UNLOCK(bp); 1383 splx(s); 1384 return; 1385 } 1386 1387 /* enqueue */ 1388 mtx_lock(&bqlock); 1389 1390 /* buffers with no memory */ 1391 if (bp->b_bufsize == 0) { 1392 bp->b_flags |= B_INVAL; 1393 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1394 if (bp->b_vflags & BV_BKGRDINPROG) 1395 panic("losing buffer 1"); 1396 if (bp->b_kvasize) { 1397 bp->b_qindex = QUEUE_EMPTYKVA; 1398 } else { 1399 bp->b_qindex = QUEUE_EMPTY; 1400 } 1401 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1402 bp->b_dev = NODEV; 1403 /* buffers with junk contents */ 1404 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1405 (bp->b_ioflags & BIO_ERROR)) { 1406 bp->b_flags |= B_INVAL; 1407 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1408 if (bp->b_vflags & BV_BKGRDINPROG) 1409 panic("losing buffer 2"); 1410 bp->b_qindex = QUEUE_CLEAN; 1411 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1412 bp->b_dev = NODEV; 1413 /* remaining buffers */ 1414 } else { 1415 if (bp->b_flags & B_DELWRI) 1416 bp->b_qindex = QUEUE_DIRTY; 1417 else 1418 bp->b_qindex = QUEUE_CLEAN; 1419 if (bp->b_flags & B_AGE) 1420 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1421 else 1422 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1423 } 1424 mtx_unlock(&bqlock); 1425 1426 /* 1427 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1428 * placed the buffer on the correct queue. We must also disassociate 1429 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1430 * find it. 1431 */ 1432 if (bp->b_flags & B_INVAL) { 1433 if (bp->b_flags & B_DELWRI) 1434 bundirty(bp); 1435 if (bp->b_vp) 1436 brelvp(bp); 1437 } 1438 1439 /* 1440 * Fixup numfreebuffers count. The bp is on an appropriate queue 1441 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1442 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1443 * if B_INVAL is set ). 1444 */ 1445 1446 if (!(bp->b_flags & B_DELWRI)) 1447 bufcountwakeup(); 1448 1449 /* 1450 * Something we can maybe free or reuse 1451 */ 1452 if (bp->b_bufsize || bp->b_kvasize) 1453 bufspacewakeup(); 1454 1455 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1456 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1457 panic("brelse: not dirty"); 1458 /* unlock */ 1459 BUF_UNLOCK(bp); 1460 splx(s); 1461} 1462 1463/* 1464 * Release a buffer back to the appropriate queue but do not try to free 1465 * it. The buffer is expected to be used again soon. 1466 * 1467 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1468 * biodone() to requeue an async I/O on completion. It is also used when 1469 * known good buffers need to be requeued but we think we may need the data 1470 * again soon. 1471 * 1472 * XXX we should be able to leave the B_RELBUF hint set on completion. 1473 */ 1474void 1475bqrelse(struct buf * bp) 1476{ 1477 int s; 1478 1479 s = splbio(); 1480 1481 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1482 1483 if (bp->b_qindex != QUEUE_NONE) 1484 panic("bqrelse: free buffer onto another queue???"); 1485 if (BUF_REFCNT(bp) > 1) { 1486 /* do not release to free list */ 1487 BUF_UNLOCK(bp); 1488 splx(s); 1489 return; 1490 } 1491 mtx_lock(&bqlock); 1492 /* buffers with stale but valid contents */ 1493 if (bp->b_flags & B_DELWRI) { 1494 bp->b_qindex = QUEUE_DIRTY; 1495 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1496 } else { 1497 /* 1498 * XXX This lock may not be necessary since BKGRDINPROG 1499 * cannot be set while we hold the buf lock, it can only be 1500 * cleared if it is already pending. 1501 */ 1502 VI_LOCK(bp->b_vp); 1503 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1504 VI_UNLOCK(bp->b_vp); 1505 bp->b_qindex = QUEUE_CLEAN; 1506 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1507 b_freelist); 1508 } else { 1509 /* 1510 * We are too low on memory, we have to try to free 1511 * the buffer (most importantly: the wired pages 1512 * making up its backing store) *now*. 1513 */ 1514 VI_UNLOCK(bp->b_vp); 1515 mtx_unlock(&bqlock); 1516 splx(s); 1517 brelse(bp); 1518 return; 1519 } 1520 } 1521 mtx_unlock(&bqlock); 1522 1523 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1524 bufcountwakeup(); 1525 1526 /* 1527 * Something we can maybe free or reuse. 1528 */ 1529 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1530 bufspacewakeup(); 1531 1532 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1533 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1534 panic("bqrelse: not dirty"); 1535 /* unlock */ 1536 BUF_UNLOCK(bp); 1537 splx(s); 1538} 1539 1540/* Give pages used by the bp back to the VM system (where possible) */ 1541static void 1542vfs_vmio_release(bp) 1543 struct buf *bp; 1544{ 1545 int i; 1546 vm_page_t m; 1547 1548 GIANT_REQUIRED; 1549 if (bp->b_object != NULL) 1550 VM_OBJECT_LOCK(bp->b_object); 1551 vm_page_lock_queues(); 1552 for (i = 0; i < bp->b_npages; i++) { 1553 m = bp->b_pages[i]; 1554 bp->b_pages[i] = NULL; 1555 /* 1556 * In order to keep page LRU ordering consistent, put 1557 * everything on the inactive queue. 1558 */ 1559 vm_page_unwire(m, 0); 1560 /* 1561 * We don't mess with busy pages, it is 1562 * the responsibility of the process that 1563 * busied the pages to deal with them. 1564 */ 1565 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1566 continue; 1567 1568 if (m->wire_count == 0) { 1569 vm_page_flag_clear(m, PG_ZERO); 1570 /* 1571 * Might as well free the page if we can and it has 1572 * no valid data. We also free the page if the 1573 * buffer was used for direct I/O 1574 */ 1575 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1576 m->hold_count == 0) { 1577 vm_page_busy(m); 1578 pmap_remove_all(m); 1579 vm_page_free(m); 1580 } else if (bp->b_flags & B_DIRECT) { 1581 vm_page_try_to_free(m); 1582 } else if (vm_page_count_severe()) { 1583 vm_page_try_to_cache(m); 1584 } 1585 } 1586 } 1587 vm_page_unlock_queues(); 1588 if (bp->b_object != NULL) 1589 VM_OBJECT_UNLOCK(bp->b_object); 1590 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1591 1592 if (bp->b_bufsize) { 1593 bufspacewakeup(); 1594 bp->b_bufsize = 0; 1595 } 1596 bp->b_npages = 0; 1597 bp->b_flags &= ~B_VMIO; 1598 if (bp->b_vp) 1599 brelvp(bp); 1600} 1601 1602/* 1603 * Check to see if a block at a particular lbn is available for a clustered 1604 * write. 1605 */ 1606static int 1607vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1608{ 1609 struct buf *bpa; 1610 int match; 1611 1612 match = 0; 1613 1614 /* If the buf isn't in core skip it */ 1615 if ((bpa = gbincore(vp, lblkno)) == NULL) 1616 return (0); 1617 1618 /* If the buf is busy we don't want to wait for it */ 1619 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1620 return (0); 1621 1622 /* Only cluster with valid clusterable delayed write buffers */ 1623 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1624 (B_DELWRI | B_CLUSTEROK)) 1625 goto done; 1626 1627 if (bpa->b_bufsize != size) 1628 goto done; 1629 1630 /* 1631 * Check to see if it is in the expected place on disk and that the 1632 * block has been mapped. 1633 */ 1634 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1635 match = 1; 1636done: 1637 BUF_UNLOCK(bpa); 1638 return (match); 1639} 1640 1641/* 1642 * vfs_bio_awrite: 1643 * 1644 * Implement clustered async writes for clearing out B_DELWRI buffers. 1645 * This is much better then the old way of writing only one buffer at 1646 * a time. Note that we may not be presented with the buffers in the 1647 * correct order, so we search for the cluster in both directions. 1648 */ 1649int 1650vfs_bio_awrite(struct buf * bp) 1651{ 1652 int i; 1653 int j; 1654 daddr_t lblkno = bp->b_lblkno; 1655 struct vnode *vp = bp->b_vp; 1656 int s; 1657 int ncl; 1658 int nwritten; 1659 int size; 1660 int maxcl; 1661 1662 s = splbio(); 1663 /* 1664 * right now we support clustered writing only to regular files. If 1665 * we find a clusterable block we could be in the middle of a cluster 1666 * rather then at the beginning. 1667 */ 1668 if ((vp->v_type == VREG) && 1669 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1670 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1671 1672 size = vp->v_mount->mnt_stat.f_iosize; 1673 maxcl = MAXPHYS / size; 1674 1675 VI_LOCK(vp); 1676 for (i = 1; i < maxcl; i++) 1677 if (vfs_bio_clcheck(vp, size, lblkno + i, 1678 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1679 break; 1680 1681 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1682 if (vfs_bio_clcheck(vp, size, lblkno - j, 1683 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1684 break; 1685 1686 VI_UNLOCK(vp); 1687 --j; 1688 ncl = i + j; 1689 /* 1690 * this is a possible cluster write 1691 */ 1692 if (ncl != 1) { 1693 BUF_UNLOCK(bp); 1694 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1695 splx(s); 1696 return nwritten; 1697 } 1698 } 1699 1700 bremfree(bp); 1701 bp->b_flags |= B_ASYNC; 1702 1703 splx(s); 1704 /* 1705 * default (old) behavior, writing out only one block 1706 * 1707 * XXX returns b_bufsize instead of b_bcount for nwritten? 1708 */ 1709 nwritten = bp->b_bufsize; 1710 (void) BUF_WRITE(bp); 1711 1712 return nwritten; 1713} 1714 1715/* 1716 * getnewbuf: 1717 * 1718 * Find and initialize a new buffer header, freeing up existing buffers 1719 * in the bufqueues as necessary. The new buffer is returned locked. 1720 * 1721 * Important: B_INVAL is not set. If the caller wishes to throw the 1722 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1723 * 1724 * We block if: 1725 * We have insufficient buffer headers 1726 * We have insufficient buffer space 1727 * buffer_map is too fragmented ( space reservation fails ) 1728 * If we have to flush dirty buffers ( but we try to avoid this ) 1729 * 1730 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1731 * Instead we ask the buf daemon to do it for us. We attempt to 1732 * avoid piecemeal wakeups of the pageout daemon. 1733 */ 1734 1735static struct buf * 1736getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1737{ 1738 struct buf *bp; 1739 struct buf *nbp; 1740 int defrag = 0; 1741 int nqindex; 1742 static int flushingbufs; 1743 1744 GIANT_REQUIRED; 1745 1746 /* 1747 * We can't afford to block since we might be holding a vnode lock, 1748 * which may prevent system daemons from running. We deal with 1749 * low-memory situations by proactively returning memory and running 1750 * async I/O rather then sync I/O. 1751 */ 1752 1753 atomic_add_int(&getnewbufcalls, 1); 1754 atomic_subtract_int(&getnewbufrestarts, 1); 1755restart: 1756 atomic_add_int(&getnewbufrestarts, 1); 1757 1758 /* 1759 * Setup for scan. If we do not have enough free buffers, 1760 * we setup a degenerate case that immediately fails. Note 1761 * that if we are specially marked process, we are allowed to 1762 * dip into our reserves. 1763 * 1764 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1765 * 1766 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1767 * However, there are a number of cases (defragging, reusing, ...) 1768 * where we cannot backup. 1769 */ 1770 mtx_lock(&bqlock); 1771 nqindex = QUEUE_EMPTYKVA; 1772 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1773 1774 if (nbp == NULL) { 1775 /* 1776 * If no EMPTYKVA buffers and we are either 1777 * defragging or reusing, locate a CLEAN buffer 1778 * to free or reuse. If bufspace useage is low 1779 * skip this step so we can allocate a new buffer. 1780 */ 1781 if (defrag || bufspace >= lobufspace) { 1782 nqindex = QUEUE_CLEAN; 1783 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1784 } 1785 1786 /* 1787 * If we could not find or were not allowed to reuse a 1788 * CLEAN buffer, check to see if it is ok to use an EMPTY 1789 * buffer. We can only use an EMPTY buffer if allocating 1790 * its KVA would not otherwise run us out of buffer space. 1791 */ 1792 if (nbp == NULL && defrag == 0 && 1793 bufspace + maxsize < hibufspace) { 1794 nqindex = QUEUE_EMPTY; 1795 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1796 } 1797 } 1798 1799 /* 1800 * Run scan, possibly freeing data and/or kva mappings on the fly 1801 * depending. 1802 */ 1803 1804 while ((bp = nbp) != NULL) { 1805 int qindex = nqindex; 1806 1807 /* 1808 * Calculate next bp ( we can only use it if we do not block 1809 * or do other fancy things ). 1810 */ 1811 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1812 switch(qindex) { 1813 case QUEUE_EMPTY: 1814 nqindex = QUEUE_EMPTYKVA; 1815 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1816 break; 1817 /* FALLTHROUGH */ 1818 case QUEUE_EMPTYKVA: 1819 nqindex = QUEUE_CLEAN; 1820 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1821 break; 1822 /* FALLTHROUGH */ 1823 case QUEUE_CLEAN: 1824 /* 1825 * nbp is NULL. 1826 */ 1827 break; 1828 } 1829 } 1830 if (bp->b_vp) { 1831 VI_LOCK(bp->b_vp); 1832 if (bp->b_vflags & BV_BKGRDINPROG) { 1833 VI_UNLOCK(bp->b_vp); 1834 continue; 1835 } 1836 VI_UNLOCK(bp->b_vp); 1837 } 1838 1839 /* 1840 * Sanity Checks 1841 */ 1842 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1843 1844 /* 1845 * Note: we no longer distinguish between VMIO and non-VMIO 1846 * buffers. 1847 */ 1848 1849 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1850 1851 /* 1852 * If we are defragging then we need a buffer with 1853 * b_kvasize != 0. XXX this situation should no longer 1854 * occur, if defrag is non-zero the buffer's b_kvasize 1855 * should also be non-zero at this point. XXX 1856 */ 1857 if (defrag && bp->b_kvasize == 0) { 1858 printf("Warning: defrag empty buffer %p\n", bp); 1859 continue; 1860 } 1861 1862 /* 1863 * Start freeing the bp. This is somewhat involved. nbp 1864 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1865 */ 1866 1867 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1868 panic("getnewbuf: locked buf"); 1869 bremfreel(bp); 1870 mtx_unlock(&bqlock); 1871 1872 if (qindex == QUEUE_CLEAN) { 1873 if (bp->b_flags & B_VMIO) { 1874 bp->b_flags &= ~B_ASYNC; 1875 vfs_vmio_release(bp); 1876 } 1877 if (bp->b_vp) 1878 brelvp(bp); 1879 } 1880 1881 /* 1882 * NOTE: nbp is now entirely invalid. We can only restart 1883 * the scan from this point on. 1884 * 1885 * Get the rest of the buffer freed up. b_kva* is still 1886 * valid after this operation. 1887 */ 1888 1889 if (bp->b_rcred != NOCRED) { 1890 crfree(bp->b_rcred); 1891 bp->b_rcred = NOCRED; 1892 } 1893 if (bp->b_wcred != NOCRED) { 1894 crfree(bp->b_wcred); 1895 bp->b_wcred = NOCRED; 1896 } 1897 if (LIST_FIRST(&bp->b_dep) != NULL) 1898 buf_deallocate(bp); 1899 if (bp->b_vflags & BV_BKGRDINPROG) 1900 panic("losing buffer 3"); 1901 1902 if (bp->b_bufsize) 1903 allocbuf(bp, 0); 1904 1905 bp->b_flags = 0; 1906 bp->b_ioflags = 0; 1907 bp->b_xflags = 0; 1908 bp->b_vflags = 0; 1909 bp->b_dev = NODEV; 1910 bp->b_vp = NULL; 1911 bp->b_blkno = bp->b_lblkno = 0; 1912 bp->b_offset = NOOFFSET; 1913 bp->b_iodone = 0; 1914 bp->b_error = 0; 1915 bp->b_resid = 0; 1916 bp->b_bcount = 0; 1917 bp->b_npages = 0; 1918 bp->b_dirtyoff = bp->b_dirtyend = 0; 1919 bp->b_magic = B_MAGIC_BIO; 1920 bp->b_op = &buf_ops_bio; 1921 bp->b_object = NULL; 1922 1923 LIST_INIT(&bp->b_dep); 1924 1925 /* 1926 * If we are defragging then free the buffer. 1927 */ 1928 if (defrag) { 1929 bp->b_flags |= B_INVAL; 1930 bfreekva(bp); 1931 brelse(bp); 1932 defrag = 0; 1933 goto restart; 1934 } 1935 1936 /* 1937 * If we are overcomitted then recover the buffer and its 1938 * KVM space. This occurs in rare situations when multiple 1939 * processes are blocked in getnewbuf() or allocbuf(). 1940 */ 1941 if (bufspace >= hibufspace) 1942 flushingbufs = 1; 1943 if (flushingbufs && bp->b_kvasize != 0) { 1944 bp->b_flags |= B_INVAL; 1945 bfreekva(bp); 1946 brelse(bp); 1947 goto restart; 1948 } 1949 if (bufspace < lobufspace) 1950 flushingbufs = 0; 1951 break; 1952 } 1953 1954 /* 1955 * If we exhausted our list, sleep as appropriate. We may have to 1956 * wakeup various daemons and write out some dirty buffers. 1957 * 1958 * Generally we are sleeping due to insufficient buffer space. 1959 */ 1960 1961 if (bp == NULL) { 1962 int flags; 1963 char *waitmsg; 1964 1965 mtx_unlock(&bqlock); 1966 if (defrag) { 1967 flags = VFS_BIO_NEED_BUFSPACE; 1968 waitmsg = "nbufkv"; 1969 } else if (bufspace >= hibufspace) { 1970 waitmsg = "nbufbs"; 1971 flags = VFS_BIO_NEED_BUFSPACE; 1972 } else { 1973 waitmsg = "newbuf"; 1974 flags = VFS_BIO_NEED_ANY; 1975 } 1976 1977 bd_speedup(); /* heeeelp */ 1978 1979 mtx_lock(&nblock); 1980 needsbuffer |= flags; 1981 while (needsbuffer & flags) { 1982 if (msleep(&needsbuffer, &nblock, 1983 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1984 mtx_unlock(&nblock); 1985 return (NULL); 1986 } 1987 } 1988 mtx_unlock(&nblock); 1989 } else { 1990 /* 1991 * We finally have a valid bp. We aren't quite out of the 1992 * woods, we still have to reserve kva space. In order 1993 * to keep fragmentation sane we only allocate kva in 1994 * BKVASIZE chunks. 1995 */ 1996 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1997 1998 if (maxsize != bp->b_kvasize) { 1999 vm_offset_t addr = 0; 2000 2001 bfreekva(bp); 2002 2003 if (vm_map_findspace(buffer_map, 2004 vm_map_min(buffer_map), maxsize, &addr)) { 2005 /* 2006 * Uh oh. Buffer map is to fragmented. We 2007 * must defragment the map. 2008 */ 2009 atomic_add_int(&bufdefragcnt, 1); 2010 defrag = 1; 2011 bp->b_flags |= B_INVAL; 2012 brelse(bp); 2013 goto restart; 2014 } 2015 if (addr) { 2016 vm_map_insert(buffer_map, NULL, 0, 2017 addr, addr + maxsize, 2018 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2019 2020 bp->b_kvabase = (caddr_t) addr; 2021 bp->b_kvasize = maxsize; 2022 atomic_add_int(&bufspace, bp->b_kvasize); 2023 atomic_add_int(&bufreusecnt, 1); 2024 } 2025 } 2026 bp->b_saveaddr = bp->b_kvabase; 2027 bp->b_data = bp->b_saveaddr; 2028 } 2029 return(bp); 2030} 2031 2032/* 2033 * buf_daemon: 2034 * 2035 * buffer flushing daemon. Buffers are normally flushed by the 2036 * update daemon but if it cannot keep up this process starts to 2037 * take the load in an attempt to prevent getnewbuf() from blocking. 2038 */ 2039 2040static struct proc *bufdaemonproc; 2041 2042static struct kproc_desc buf_kp = { 2043 "bufdaemon", 2044 buf_daemon, 2045 &bufdaemonproc 2046}; 2047SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2048 2049static void 2050buf_daemon() 2051{ 2052 int s; 2053 2054 mtx_lock(&Giant); 2055 2056 /* 2057 * This process needs to be suspended prior to shutdown sync. 2058 */ 2059 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2060 SHUTDOWN_PRI_LAST); 2061 2062 /* 2063 * This process is allowed to take the buffer cache to the limit 2064 */ 2065 s = splbio(); 2066 mtx_lock(&bdlock); 2067 2068 for (;;) { 2069 bd_request = 0; 2070 mtx_unlock(&bdlock); 2071 2072 kthread_suspend_check(bufdaemonproc); 2073 2074 /* 2075 * Do the flush. Limit the amount of in-transit I/O we 2076 * allow to build up, otherwise we would completely saturate 2077 * the I/O system. Wakeup any waiting processes before we 2078 * normally would so they can run in parallel with our drain. 2079 */ 2080 while (numdirtybuffers > lodirtybuffers) { 2081 if (flushbufqueues(0) == 0) { 2082 /* 2083 * Could not find any buffers without rollback 2084 * dependencies, so just write the first one 2085 * in the hopes of eventually making progress. 2086 */ 2087 flushbufqueues(1); 2088 break; 2089 } 2090 waitrunningbufspace(); 2091 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2092 } 2093 2094 /* 2095 * Only clear bd_request if we have reached our low water 2096 * mark. The buf_daemon normally waits 1 second and 2097 * then incrementally flushes any dirty buffers that have 2098 * built up, within reason. 2099 * 2100 * If we were unable to hit our low water mark and couldn't 2101 * find any flushable buffers, we sleep half a second. 2102 * Otherwise we loop immediately. 2103 */ 2104 mtx_lock(&bdlock); 2105 if (numdirtybuffers <= lodirtybuffers) { 2106 /* 2107 * We reached our low water mark, reset the 2108 * request and sleep until we are needed again. 2109 * The sleep is just so the suspend code works. 2110 */ 2111 bd_request = 0; 2112 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2113 } else { 2114 /* 2115 * We couldn't find any flushable dirty buffers but 2116 * still have too many dirty buffers, we 2117 * have to sleep and try again. (rare) 2118 */ 2119 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2120 } 2121 } 2122} 2123 2124/* 2125 * flushbufqueues: 2126 * 2127 * Try to flush a buffer in the dirty queue. We must be careful to 2128 * free up B_INVAL buffers instead of write them, which NFS is 2129 * particularly sensitive to. 2130 */ 2131int flushwithdeps = 0; 2132SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2133 0, "Number of buffers flushed with dependecies that require rollbacks"); 2134static int 2135flushbufqueues(int flushdeps) 2136{ 2137 struct thread *td = curthread; 2138 struct vnode *vp; 2139 struct mount *mp; 2140 struct buf *bp; 2141 int hasdeps; 2142 2143 mtx_lock(&bqlock); 2144 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2145 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2146 continue; 2147 KASSERT((bp->b_flags & B_DELWRI), 2148 ("unexpected clean buffer %p", bp)); 2149 VI_LOCK(bp->b_vp); 2150 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 2151 VI_UNLOCK(bp->b_vp); 2152 BUF_UNLOCK(bp); 2153 continue; 2154 } 2155 VI_UNLOCK(bp->b_vp); 2156 if (bp->b_flags & B_INVAL) { 2157 bremfreel(bp); 2158 mtx_unlock(&bqlock); 2159 brelse(bp); 2160 return (1); 2161 } 2162 2163 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2164 if (flushdeps == 0) { 2165 BUF_UNLOCK(bp); 2166 continue; 2167 } 2168 hasdeps = 1; 2169 } else 2170 hasdeps = 0; 2171 /* 2172 * We must hold the lock on a vnode before writing 2173 * one of its buffers. Otherwise we may confuse, or 2174 * in the case of a snapshot vnode, deadlock the 2175 * system. 2176 * 2177 * The lock order here is the reverse of the normal 2178 * of vnode followed by buf lock. This is ok because 2179 * the NOWAIT will prevent deadlock. 2180 */ 2181 vp = bp->b_vp; 2182 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2183 BUF_UNLOCK(bp); 2184 continue; 2185 } 2186 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2187 mtx_unlock(&bqlock); 2188 vfs_bio_awrite(bp); 2189 vn_finished_write(mp); 2190 VOP_UNLOCK(vp, 0, td); 2191 flushwithdeps += hasdeps; 2192 return (1); 2193 } 2194 vn_finished_write(mp); 2195 BUF_UNLOCK(bp); 2196 } 2197 mtx_unlock(&bqlock); 2198 return (0); 2199} 2200 2201/* 2202 * Check to see if a block is currently memory resident. 2203 */ 2204struct buf * 2205incore(struct vnode * vp, daddr_t blkno) 2206{ 2207 struct buf *bp; 2208 2209 int s = splbio(); 2210 VI_LOCK(vp); 2211 bp = gbincore(vp, blkno); 2212 VI_UNLOCK(vp); 2213 splx(s); 2214 return (bp); 2215} 2216 2217/* 2218 * Returns true if no I/O is needed to access the 2219 * associated VM object. This is like incore except 2220 * it also hunts around in the VM system for the data. 2221 */ 2222 2223int 2224inmem(struct vnode * vp, daddr_t blkno) 2225{ 2226 vm_object_t obj; 2227 vm_offset_t toff, tinc, size; 2228 vm_page_t m; 2229 vm_ooffset_t off; 2230 2231 GIANT_REQUIRED; 2232 ASSERT_VOP_LOCKED(vp, "inmem"); 2233 2234 if (incore(vp, blkno)) 2235 return 1; 2236 if (vp->v_mount == NULL) 2237 return 0; 2238 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2239 return 0; 2240 2241 size = PAGE_SIZE; 2242 if (size > vp->v_mount->mnt_stat.f_iosize) 2243 size = vp->v_mount->mnt_stat.f_iosize; 2244 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2245 2246 VM_OBJECT_LOCK(obj); 2247 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2248 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2249 if (!m) 2250 goto notinmem; 2251 tinc = size; 2252 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2253 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2254 if (vm_page_is_valid(m, 2255 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2256 goto notinmem; 2257 } 2258 VM_OBJECT_UNLOCK(obj); 2259 return 1; 2260 2261notinmem: 2262 VM_OBJECT_UNLOCK(obj); 2263 return (0); 2264} 2265 2266/* 2267 * vfs_setdirty: 2268 * 2269 * Sets the dirty range for a buffer based on the status of the dirty 2270 * bits in the pages comprising the buffer. 2271 * 2272 * The range is limited to the size of the buffer. 2273 * 2274 * This routine is primarily used by NFS, but is generalized for the 2275 * B_VMIO case. 2276 */ 2277static void 2278vfs_setdirty(struct buf *bp) 2279{ 2280 int i; 2281 vm_object_t object; 2282 2283 GIANT_REQUIRED; 2284 /* 2285 * Degenerate case - empty buffer 2286 */ 2287 2288 if (bp->b_bufsize == 0) 2289 return; 2290 2291 /* 2292 * We qualify the scan for modified pages on whether the 2293 * object has been flushed yet. The OBJ_WRITEABLE flag 2294 * is not cleared simply by protecting pages off. 2295 */ 2296 2297 if ((bp->b_flags & B_VMIO) == 0) 2298 return; 2299 2300 object = bp->b_pages[0]->object; 2301 VM_OBJECT_LOCK(object); 2302 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2303 printf("Warning: object %p writeable but not mightbedirty\n", object); 2304 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2305 printf("Warning: object %p mightbedirty but not writeable\n", object); 2306 2307 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2308 vm_offset_t boffset; 2309 vm_offset_t eoffset; 2310 2311 vm_page_lock_queues(); 2312 /* 2313 * test the pages to see if they have been modified directly 2314 * by users through the VM system. 2315 */ 2316 for (i = 0; i < bp->b_npages; i++) { 2317 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2318 vm_page_test_dirty(bp->b_pages[i]); 2319 } 2320 2321 /* 2322 * Calculate the encompassing dirty range, boffset and eoffset, 2323 * (eoffset - boffset) bytes. 2324 */ 2325 2326 for (i = 0; i < bp->b_npages; i++) { 2327 if (bp->b_pages[i]->dirty) 2328 break; 2329 } 2330 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2331 2332 for (i = bp->b_npages - 1; i >= 0; --i) { 2333 if (bp->b_pages[i]->dirty) { 2334 break; 2335 } 2336 } 2337 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2338 2339 vm_page_unlock_queues(); 2340 /* 2341 * Fit it to the buffer. 2342 */ 2343 2344 if (eoffset > bp->b_bcount) 2345 eoffset = bp->b_bcount; 2346 2347 /* 2348 * If we have a good dirty range, merge with the existing 2349 * dirty range. 2350 */ 2351 2352 if (boffset < eoffset) { 2353 if (bp->b_dirtyoff > boffset) 2354 bp->b_dirtyoff = boffset; 2355 if (bp->b_dirtyend < eoffset) 2356 bp->b_dirtyend = eoffset; 2357 } 2358 } 2359 VM_OBJECT_UNLOCK(object); 2360} 2361 2362/* 2363 * getblk: 2364 * 2365 * Get a block given a specified block and offset into a file/device. 2366 * The buffers B_DONE bit will be cleared on return, making it almost 2367 * ready for an I/O initiation. B_INVAL may or may not be set on 2368 * return. The caller should clear B_INVAL prior to initiating a 2369 * READ. 2370 * 2371 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2372 * an existing buffer. 2373 * 2374 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2375 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2376 * and then cleared based on the backing VM. If the previous buffer is 2377 * non-0-sized but invalid, B_CACHE will be cleared. 2378 * 2379 * If getblk() must create a new buffer, the new buffer is returned with 2380 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2381 * case it is returned with B_INVAL clear and B_CACHE set based on the 2382 * backing VM. 2383 * 2384 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2385 * B_CACHE bit is clear. 2386 * 2387 * What this means, basically, is that the caller should use B_CACHE to 2388 * determine whether the buffer is fully valid or not and should clear 2389 * B_INVAL prior to issuing a read. If the caller intends to validate 2390 * the buffer by loading its data area with something, the caller needs 2391 * to clear B_INVAL. If the caller does this without issuing an I/O, 2392 * the caller should set B_CACHE ( as an optimization ), else the caller 2393 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2394 * a write attempt or if it was a successfull read. If the caller 2395 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2396 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2397 */ 2398struct buf * 2399getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2400 int flags) 2401{ 2402 struct buf *bp; 2403 int s; 2404 int error; 2405 ASSERT_VOP_LOCKED(vp, "getblk"); 2406 2407 if (size > MAXBSIZE) 2408 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2409 2410 s = splbio(); 2411loop: 2412 /* 2413 * Block if we are low on buffers. Certain processes are allowed 2414 * to completely exhaust the buffer cache. 2415 * 2416 * If this check ever becomes a bottleneck it may be better to 2417 * move it into the else, when gbincore() fails. At the moment 2418 * it isn't a problem. 2419 * 2420 * XXX remove if 0 sections (clean this up after its proven) 2421 */ 2422 if (numfreebuffers == 0) { 2423 if (curthread == PCPU_GET(idlethread)) 2424 return NULL; 2425 mtx_lock(&nblock); 2426 needsbuffer |= VFS_BIO_NEED_ANY; 2427 mtx_unlock(&nblock); 2428 } 2429 2430 VI_LOCK(vp); 2431 if ((bp = gbincore(vp, blkno))) { 2432 int lockflags; 2433 /* 2434 * Buffer is in-core. If the buffer is not busy, it must 2435 * be on a queue. 2436 */ 2437 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2438 2439 if (flags & GB_LOCK_NOWAIT) 2440 lockflags |= LK_NOWAIT; 2441 2442 error = BUF_TIMELOCK(bp, lockflags, 2443 VI_MTX(vp), "getblk", slpflag, slptimeo); 2444 2445 /* 2446 * If we slept and got the lock we have to restart in case 2447 * the buffer changed identities. 2448 */ 2449 if (error == ENOLCK) 2450 goto loop; 2451 /* We timed out or were interrupted. */ 2452 else if (error) 2453 return (NULL); 2454 2455 /* 2456 * The buffer is locked. B_CACHE is cleared if the buffer is 2457 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2458 * and for a VMIO buffer B_CACHE is adjusted according to the 2459 * backing VM cache. 2460 */ 2461 if (bp->b_flags & B_INVAL) 2462 bp->b_flags &= ~B_CACHE; 2463 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2464 bp->b_flags |= B_CACHE; 2465 bremfree(bp); 2466 2467 /* 2468 * check for size inconsistancies for non-VMIO case. 2469 */ 2470 2471 if (bp->b_bcount != size) { 2472 if ((bp->b_flags & B_VMIO) == 0 || 2473 (size > bp->b_kvasize)) { 2474 if (bp->b_flags & B_DELWRI) { 2475 bp->b_flags |= B_NOCACHE; 2476 BUF_WRITE(bp); 2477 } else { 2478 if ((bp->b_flags & B_VMIO) && 2479 (LIST_FIRST(&bp->b_dep) == NULL)) { 2480 bp->b_flags |= B_RELBUF; 2481 brelse(bp); 2482 } else { 2483 bp->b_flags |= B_NOCACHE; 2484 BUF_WRITE(bp); 2485 } 2486 } 2487 goto loop; 2488 } 2489 } 2490 2491 /* 2492 * If the size is inconsistant in the VMIO case, we can resize 2493 * the buffer. This might lead to B_CACHE getting set or 2494 * cleared. If the size has not changed, B_CACHE remains 2495 * unchanged from its previous state. 2496 */ 2497 2498 if (bp->b_bcount != size) 2499 allocbuf(bp, size); 2500 2501 KASSERT(bp->b_offset != NOOFFSET, 2502 ("getblk: no buffer offset")); 2503 2504 /* 2505 * A buffer with B_DELWRI set and B_CACHE clear must 2506 * be committed before we can return the buffer in 2507 * order to prevent the caller from issuing a read 2508 * ( due to B_CACHE not being set ) and overwriting 2509 * it. 2510 * 2511 * Most callers, including NFS and FFS, need this to 2512 * operate properly either because they assume they 2513 * can issue a read if B_CACHE is not set, or because 2514 * ( for example ) an uncached B_DELWRI might loop due 2515 * to softupdates re-dirtying the buffer. In the latter 2516 * case, B_CACHE is set after the first write completes, 2517 * preventing further loops. 2518 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2519 * above while extending the buffer, we cannot allow the 2520 * buffer to remain with B_CACHE set after the write 2521 * completes or it will represent a corrupt state. To 2522 * deal with this we set B_NOCACHE to scrap the buffer 2523 * after the write. 2524 * 2525 * We might be able to do something fancy, like setting 2526 * B_CACHE in bwrite() except if B_DELWRI is already set, 2527 * so the below call doesn't set B_CACHE, but that gets real 2528 * confusing. This is much easier. 2529 */ 2530 2531 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2532 bp->b_flags |= B_NOCACHE; 2533 BUF_WRITE(bp); 2534 goto loop; 2535 } 2536 2537 splx(s); 2538 bp->b_flags &= ~B_DONE; 2539 } else { 2540 int bsize, maxsize, vmio; 2541 off_t offset; 2542 2543 /* 2544 * Buffer is not in-core, create new buffer. The buffer 2545 * returned by getnewbuf() is locked. Note that the returned 2546 * buffer is also considered valid (not marked B_INVAL). 2547 */ 2548 VI_UNLOCK(vp); 2549 /* 2550 * If the user does not want us to create the buffer, bail out 2551 * here. 2552 */ 2553 if (flags & GB_NOCREAT) { 2554 splx(s); 2555 return NULL; 2556 } 2557 if (vn_isdisk(vp, NULL)) 2558 bsize = DEV_BSIZE; 2559 else if (vp->v_mountedhere) 2560 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2561 else if (vp->v_mount) 2562 bsize = vp->v_mount->mnt_stat.f_iosize; 2563 else 2564 bsize = size; 2565 2566 offset = blkno * bsize; 2567 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2568 (vp->v_vflag & VV_OBJBUF); 2569 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2570 maxsize = imax(maxsize, bsize); 2571 2572 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2573 if (slpflag || slptimeo) { 2574 splx(s); 2575 return NULL; 2576 } 2577 goto loop; 2578 } 2579 2580 /* 2581 * This code is used to make sure that a buffer is not 2582 * created while the getnewbuf routine is blocked. 2583 * This can be a problem whether the vnode is locked or not. 2584 * If the buffer is created out from under us, we have to 2585 * throw away the one we just created. There is now window 2586 * race because we are safely running at splbio() from the 2587 * point of the duplicate buffer creation through to here, 2588 * and we've locked the buffer. 2589 * 2590 * Note: this must occur before we associate the buffer 2591 * with the vp especially considering limitations in 2592 * the splay tree implementation when dealing with duplicate 2593 * lblkno's. 2594 */ 2595 VI_LOCK(vp); 2596 if (gbincore(vp, blkno)) { 2597 VI_UNLOCK(vp); 2598 bp->b_flags |= B_INVAL; 2599 brelse(bp); 2600 goto loop; 2601 } 2602 2603 /* 2604 * Insert the buffer into the hash, so that it can 2605 * be found by incore. 2606 */ 2607 bp->b_blkno = bp->b_lblkno = blkno; 2608 bp->b_offset = offset; 2609 2610 bgetvp(vp, bp); 2611 VI_UNLOCK(vp); 2612 2613 /* 2614 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2615 * buffer size starts out as 0, B_CACHE will be set by 2616 * allocbuf() for the VMIO case prior to it testing the 2617 * backing store for validity. 2618 */ 2619 2620 if (vmio) { 2621 bp->b_flags |= B_VMIO; 2622#if defined(VFS_BIO_DEBUG) 2623 if (vp->v_type != VREG) 2624 printf("getblk: vmioing file type %d???\n", vp->v_type); 2625#endif 2626 VOP_GETVOBJECT(vp, &bp->b_object); 2627 } else { 2628 bp->b_flags &= ~B_VMIO; 2629 bp->b_object = NULL; 2630 } 2631 2632 allocbuf(bp, size); 2633 2634 splx(s); 2635 bp->b_flags &= ~B_DONE; 2636 } 2637 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2638 return (bp); 2639} 2640 2641/* 2642 * Get an empty, disassociated buffer of given size. The buffer is initially 2643 * set to B_INVAL. 2644 */ 2645struct buf * 2646geteblk(int size) 2647{ 2648 struct buf *bp; 2649 int s; 2650 int maxsize; 2651 2652 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2653 2654 s = splbio(); 2655 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2656 continue; 2657 splx(s); 2658 allocbuf(bp, size); 2659 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2660 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2661 return (bp); 2662} 2663 2664 2665/* 2666 * This code constitutes the buffer memory from either anonymous system 2667 * memory (in the case of non-VMIO operations) or from an associated 2668 * VM object (in the case of VMIO operations). This code is able to 2669 * resize a buffer up or down. 2670 * 2671 * Note that this code is tricky, and has many complications to resolve 2672 * deadlock or inconsistant data situations. Tread lightly!!! 2673 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2674 * the caller. Calling this code willy nilly can result in the loss of data. 2675 * 2676 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2677 * B_CACHE for the non-VMIO case. 2678 */ 2679 2680int 2681allocbuf(struct buf *bp, int size) 2682{ 2683 int newbsize, mbsize; 2684 int i; 2685 2686 GIANT_REQUIRED; 2687 2688 if (BUF_REFCNT(bp) == 0) 2689 panic("allocbuf: buffer not busy"); 2690 2691 if (bp->b_kvasize < size) 2692 panic("allocbuf: buffer too small"); 2693 2694 if ((bp->b_flags & B_VMIO) == 0) { 2695 caddr_t origbuf; 2696 int origbufsize; 2697 /* 2698 * Just get anonymous memory from the kernel. Don't 2699 * mess with B_CACHE. 2700 */ 2701 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2702 if (bp->b_flags & B_MALLOC) 2703 newbsize = mbsize; 2704 else 2705 newbsize = round_page(size); 2706 2707 if (newbsize < bp->b_bufsize) { 2708 /* 2709 * malloced buffers are not shrunk 2710 */ 2711 if (bp->b_flags & B_MALLOC) { 2712 if (newbsize) { 2713 bp->b_bcount = size; 2714 } else { 2715 free(bp->b_data, M_BIOBUF); 2716 if (bp->b_bufsize) { 2717 atomic_subtract_int( 2718 &bufmallocspace, 2719 bp->b_bufsize); 2720 bufspacewakeup(); 2721 bp->b_bufsize = 0; 2722 } 2723 bp->b_saveaddr = bp->b_kvabase; 2724 bp->b_data = bp->b_saveaddr; 2725 bp->b_bcount = 0; 2726 bp->b_flags &= ~B_MALLOC; 2727 } 2728 return 1; 2729 } 2730 vm_hold_free_pages( 2731 bp, 2732 (vm_offset_t) bp->b_data + newbsize, 2733 (vm_offset_t) bp->b_data + bp->b_bufsize); 2734 } else if (newbsize > bp->b_bufsize) { 2735 /* 2736 * We only use malloced memory on the first allocation. 2737 * and revert to page-allocated memory when the buffer 2738 * grows. 2739 */ 2740 /* 2741 * There is a potential smp race here that could lead 2742 * to bufmallocspace slightly passing the max. It 2743 * is probably extremely rare and not worth worrying 2744 * over. 2745 */ 2746 if ( (bufmallocspace < maxbufmallocspace) && 2747 (bp->b_bufsize == 0) && 2748 (mbsize <= PAGE_SIZE/2)) { 2749 2750 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2751 bp->b_bufsize = mbsize; 2752 bp->b_bcount = size; 2753 bp->b_flags |= B_MALLOC; 2754 atomic_add_int(&bufmallocspace, mbsize); 2755 return 1; 2756 } 2757 origbuf = NULL; 2758 origbufsize = 0; 2759 /* 2760 * If the buffer is growing on its other-than-first allocation, 2761 * then we revert to the page-allocation scheme. 2762 */ 2763 if (bp->b_flags & B_MALLOC) { 2764 origbuf = bp->b_data; 2765 origbufsize = bp->b_bufsize; 2766 bp->b_data = bp->b_kvabase; 2767 if (bp->b_bufsize) { 2768 atomic_subtract_int(&bufmallocspace, 2769 bp->b_bufsize); 2770 bufspacewakeup(); 2771 bp->b_bufsize = 0; 2772 } 2773 bp->b_flags &= ~B_MALLOC; 2774 newbsize = round_page(newbsize); 2775 } 2776 vm_hold_load_pages( 2777 bp, 2778 (vm_offset_t) bp->b_data + bp->b_bufsize, 2779 (vm_offset_t) bp->b_data + newbsize); 2780 if (origbuf) { 2781 bcopy(origbuf, bp->b_data, origbufsize); 2782 free(origbuf, M_BIOBUF); 2783 } 2784 } 2785 } else { 2786 int desiredpages; 2787 2788 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2789 desiredpages = (size == 0) ? 0 : 2790 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2791 2792 if (bp->b_flags & B_MALLOC) 2793 panic("allocbuf: VMIO buffer can't be malloced"); 2794 /* 2795 * Set B_CACHE initially if buffer is 0 length or will become 2796 * 0-length. 2797 */ 2798 if (size == 0 || bp->b_bufsize == 0) 2799 bp->b_flags |= B_CACHE; 2800 2801 if (newbsize < bp->b_bufsize) { 2802 /* 2803 * DEV_BSIZE aligned new buffer size is less then the 2804 * DEV_BSIZE aligned existing buffer size. Figure out 2805 * if we have to remove any pages. 2806 */ 2807 if (desiredpages < bp->b_npages) { 2808 vm_page_t m; 2809 2810 vm_page_lock_queues(); 2811 for (i = desiredpages; i < bp->b_npages; i++) { 2812 /* 2813 * the page is not freed here -- it 2814 * is the responsibility of 2815 * vnode_pager_setsize 2816 */ 2817 m = bp->b_pages[i]; 2818 KASSERT(m != bogus_page, 2819 ("allocbuf: bogus page found")); 2820 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2821 vm_page_lock_queues(); 2822 2823 bp->b_pages[i] = NULL; 2824 vm_page_unwire(m, 0); 2825 } 2826 vm_page_unlock_queues(); 2827 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2828 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2829 bp->b_npages = desiredpages; 2830 } 2831 } else if (size > bp->b_bcount) { 2832 /* 2833 * We are growing the buffer, possibly in a 2834 * byte-granular fashion. 2835 */ 2836 struct vnode *vp; 2837 vm_object_t obj; 2838 vm_offset_t toff; 2839 vm_offset_t tinc; 2840 2841 /* 2842 * Step 1, bring in the VM pages from the object, 2843 * allocating them if necessary. We must clear 2844 * B_CACHE if these pages are not valid for the 2845 * range covered by the buffer. 2846 */ 2847 2848 vp = bp->b_vp; 2849 obj = bp->b_object; 2850 2851 VM_OBJECT_LOCK(obj); 2852 while (bp->b_npages < desiredpages) { 2853 vm_page_t m; 2854 vm_pindex_t pi; 2855 2856 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2857 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2858 /* 2859 * note: must allocate system pages 2860 * since blocking here could intefere 2861 * with paging I/O, no matter which 2862 * process we are. 2863 */ 2864 m = vm_page_alloc(obj, pi, 2865 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2866 if (m == NULL) { 2867 atomic_add_int(&vm_pageout_deficit, 2868 desiredpages - bp->b_npages); 2869 VM_OBJECT_UNLOCK(obj); 2870 VM_WAIT; 2871 VM_OBJECT_LOCK(obj); 2872 } else { 2873 vm_page_lock_queues(); 2874 vm_page_wakeup(m); 2875 vm_page_unlock_queues(); 2876 bp->b_flags &= ~B_CACHE; 2877 bp->b_pages[bp->b_npages] = m; 2878 ++bp->b_npages; 2879 } 2880 continue; 2881 } 2882 2883 /* 2884 * We found a page. If we have to sleep on it, 2885 * retry because it might have gotten freed out 2886 * from under us. 2887 * 2888 * We can only test PG_BUSY here. Blocking on 2889 * m->busy might lead to a deadlock: 2890 * 2891 * vm_fault->getpages->cluster_read->allocbuf 2892 * 2893 */ 2894 vm_page_lock_queues(); 2895 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2896 continue; 2897 2898 /* 2899 * We have a good page. Should we wakeup the 2900 * page daemon? 2901 */ 2902 if ((curproc != pageproc) && 2903 ((m->queue - m->pc) == PQ_CACHE) && 2904 ((cnt.v_free_count + cnt.v_cache_count) < 2905 (cnt.v_free_min + cnt.v_cache_min))) { 2906 pagedaemon_wakeup(); 2907 } 2908 vm_page_flag_clear(m, PG_ZERO); 2909 vm_page_wire(m); 2910 vm_page_unlock_queues(); 2911 bp->b_pages[bp->b_npages] = m; 2912 ++bp->b_npages; 2913 } 2914 2915 /* 2916 * Step 2. We've loaded the pages into the buffer, 2917 * we have to figure out if we can still have B_CACHE 2918 * set. Note that B_CACHE is set according to the 2919 * byte-granular range ( bcount and size ), new the 2920 * aligned range ( newbsize ). 2921 * 2922 * The VM test is against m->valid, which is DEV_BSIZE 2923 * aligned. Needless to say, the validity of the data 2924 * needs to also be DEV_BSIZE aligned. Note that this 2925 * fails with NFS if the server or some other client 2926 * extends the file's EOF. If our buffer is resized, 2927 * B_CACHE may remain set! XXX 2928 */ 2929 2930 toff = bp->b_bcount; 2931 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2932 2933 while ((bp->b_flags & B_CACHE) && toff < size) { 2934 vm_pindex_t pi; 2935 2936 if (tinc > (size - toff)) 2937 tinc = size - toff; 2938 2939 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2940 PAGE_SHIFT; 2941 2942 vfs_buf_test_cache( 2943 bp, 2944 bp->b_offset, 2945 toff, 2946 tinc, 2947 bp->b_pages[pi] 2948 ); 2949 toff += tinc; 2950 tinc = PAGE_SIZE; 2951 } 2952 VM_OBJECT_UNLOCK(obj); 2953 2954 /* 2955 * Step 3, fixup the KVM pmap. Remember that 2956 * bp->b_data is relative to bp->b_offset, but 2957 * bp->b_offset may be offset into the first page. 2958 */ 2959 2960 bp->b_data = (caddr_t) 2961 trunc_page((vm_offset_t)bp->b_data); 2962 pmap_qenter( 2963 (vm_offset_t)bp->b_data, 2964 bp->b_pages, 2965 bp->b_npages 2966 ); 2967 2968 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2969 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2970 } 2971 } 2972 if (newbsize < bp->b_bufsize) 2973 bufspacewakeup(); 2974 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2975 bp->b_bcount = size; /* requested buffer size */ 2976 return 1; 2977} 2978 2979void 2980biodone(struct bio *bp) 2981{ 2982 mtx_lock(&bdonelock); 2983 bp->bio_flags |= BIO_DONE; 2984 if (bp->bio_done == NULL) 2985 wakeup(bp); 2986 mtx_unlock(&bdonelock); 2987 if (bp->bio_done != NULL) 2988 bp->bio_done(bp); 2989} 2990 2991/* 2992 * Wait for a BIO to finish. 2993 * 2994 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2995 * case is not yet clear. 2996 */ 2997int 2998biowait(struct bio *bp, const char *wchan) 2999{ 3000 3001 mtx_lock(&bdonelock); 3002 while ((bp->bio_flags & BIO_DONE) == 0) 3003 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3004 mtx_unlock(&bdonelock); 3005 if (bp->bio_error != 0) 3006 return (bp->bio_error); 3007 if (!(bp->bio_flags & BIO_ERROR)) 3008 return (0); 3009 return (EIO); 3010} 3011 3012void 3013biofinish(struct bio *bp, struct devstat *stat, int error) 3014{ 3015 3016 if (error) { 3017 bp->bio_error = error; 3018 bp->bio_flags |= BIO_ERROR; 3019 } 3020 if (stat != NULL) 3021 devstat_end_transaction_bio(stat, bp); 3022 biodone(bp); 3023} 3024 3025/* 3026 * bufwait: 3027 * 3028 * Wait for buffer I/O completion, returning error status. The buffer 3029 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3030 * error and cleared. 3031 */ 3032int 3033bufwait(register struct buf * bp) 3034{ 3035 int s; 3036 3037 s = splbio(); 3038 if (bp->b_iocmd == BIO_READ) 3039 bwait(bp, PRIBIO, "biord"); 3040 else 3041 bwait(bp, PRIBIO, "biowr"); 3042 splx(s); 3043 if (bp->b_flags & B_EINTR) { 3044 bp->b_flags &= ~B_EINTR; 3045 return (EINTR); 3046 } 3047 if (bp->b_ioflags & BIO_ERROR) { 3048 return (bp->b_error ? bp->b_error : EIO); 3049 } else { 3050 return (0); 3051 } 3052} 3053 3054 /* 3055 * Call back function from struct bio back up to struct buf. 3056 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 3057 */ 3058static void 3059bufdonebio(struct bio *bp) 3060{ 3061 3062 /* Device drivers may or may not hold giant, hold it here. */ 3063 mtx_lock(&Giant); 3064 bufdone(bp->bio_caller2); 3065 mtx_unlock(&Giant); 3066} 3067 3068void 3069dev_strategy(struct buf *bp) 3070{ 3071 3072 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3073 panic("b_iocmd botch"); 3074 if (bp->b_flags & B_PHYS) 3075 bp->b_io.bio_offset = bp->b_offset; 3076 else 3077 bp->b_io.bio_offset = dbtob(bp->b_blkno); 3078 bp->b_io.bio_done = bufdonebio; 3079 bp->b_io.bio_caller2 = bp; 3080 (*devsw(bp->b_io.bio_dev)->d_strategy)(&bp->b_io); 3081} 3082 3083/* 3084 * bufdone: 3085 * 3086 * Finish I/O on a buffer, optionally calling a completion function. 3087 * This is usually called from an interrupt so process blocking is 3088 * not allowed. 3089 * 3090 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3091 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3092 * assuming B_INVAL is clear. 3093 * 3094 * For the VMIO case, we set B_CACHE if the op was a read and no 3095 * read error occured, or if the op was a write. B_CACHE is never 3096 * set if the buffer is invalid or otherwise uncacheable. 3097 * 3098 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3099 * initiator to leave B_INVAL set to brelse the buffer out of existance 3100 * in the biodone routine. 3101 */ 3102void 3103bufdone(struct buf *bp) 3104{ 3105 int s; 3106 void (*biodone)(struct buf *); 3107 3108 GIANT_REQUIRED; 3109 3110 s = splbio(); 3111 3112 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3113 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3114 3115 bp->b_flags |= B_DONE; 3116 runningbufwakeup(bp); 3117 3118 if (bp->b_iocmd == BIO_DELETE) { 3119 brelse(bp); 3120 splx(s); 3121 return; 3122 } 3123 3124 if (bp->b_iocmd == BIO_WRITE) { 3125 vwakeup(bp); 3126 } 3127 3128 /* call optional completion function if requested */ 3129 if (bp->b_iodone != NULL) { 3130 biodone = bp->b_iodone; 3131 bp->b_iodone = NULL; 3132 (*biodone) (bp); 3133 splx(s); 3134 return; 3135 } 3136 if (LIST_FIRST(&bp->b_dep) != NULL) 3137 buf_complete(bp); 3138 3139 if (bp->b_flags & B_VMIO) { 3140 int i; 3141 vm_ooffset_t foff; 3142 vm_page_t m; 3143 vm_object_t obj; 3144 int iosize; 3145 struct vnode *vp = bp->b_vp; 3146 3147 obj = bp->b_object; 3148 3149#if defined(VFS_BIO_DEBUG) 3150 mp_fixme("usecount and vflag accessed without locks."); 3151 if (vp->v_usecount == 0) { 3152 panic("biodone: zero vnode ref count"); 3153 } 3154 3155 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3156 panic("biodone: vnode is not setup for merged cache"); 3157 } 3158#endif 3159 3160 foff = bp->b_offset; 3161 KASSERT(bp->b_offset != NOOFFSET, 3162 ("biodone: no buffer offset")); 3163 3164 if (obj != NULL) 3165 VM_OBJECT_LOCK(obj); 3166#if defined(VFS_BIO_DEBUG) 3167 if (obj->paging_in_progress < bp->b_npages) { 3168 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3169 obj->paging_in_progress, bp->b_npages); 3170 } 3171#endif 3172 3173 /* 3174 * Set B_CACHE if the op was a normal read and no error 3175 * occured. B_CACHE is set for writes in the b*write() 3176 * routines. 3177 */ 3178 iosize = bp->b_bcount - bp->b_resid; 3179 if (bp->b_iocmd == BIO_READ && 3180 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3181 !(bp->b_ioflags & BIO_ERROR)) { 3182 bp->b_flags |= B_CACHE; 3183 } 3184 vm_page_lock_queues(); 3185 for (i = 0; i < bp->b_npages; i++) { 3186 int bogusflag = 0; 3187 int resid; 3188 3189 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3190 if (resid > iosize) 3191 resid = iosize; 3192 3193 /* 3194 * cleanup bogus pages, restoring the originals 3195 */ 3196 m = bp->b_pages[i]; 3197 if (m == bogus_page) { 3198 bogusflag = 1; 3199 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3200 if (m == NULL) 3201 panic("biodone: page disappeared!"); 3202 bp->b_pages[i] = m; 3203 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3204 } 3205#if defined(VFS_BIO_DEBUG) 3206 if (OFF_TO_IDX(foff) != m->pindex) { 3207 printf( 3208"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3209 (intmax_t)foff, (uintmax_t)m->pindex); 3210 } 3211#endif 3212 3213 /* 3214 * In the write case, the valid and clean bits are 3215 * already changed correctly ( see bdwrite() ), so we 3216 * only need to do this here in the read case. 3217 */ 3218 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3219 vfs_page_set_valid(bp, foff, i, m); 3220 } 3221 vm_page_flag_clear(m, PG_ZERO); 3222 3223 /* 3224 * when debugging new filesystems or buffer I/O methods, this 3225 * is the most common error that pops up. if you see this, you 3226 * have not set the page busy flag correctly!!! 3227 */ 3228 if (m->busy == 0) { 3229 printf("biodone: page busy < 0, " 3230 "pindex: %d, foff: 0x(%x,%x), " 3231 "resid: %d, index: %d\n", 3232 (int) m->pindex, (int)(foff >> 32), 3233 (int) foff & 0xffffffff, resid, i); 3234 if (!vn_isdisk(vp, NULL)) 3235 printf(" iosize: %ld, lblkno: %jd, flags: 0x%x, npages: %d\n", 3236 bp->b_vp->v_mount->mnt_stat.f_iosize, 3237 (intmax_t) bp->b_lblkno, 3238 bp->b_flags, bp->b_npages); 3239 else 3240 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3241 (intmax_t) bp->b_lblkno, 3242 bp->b_flags, bp->b_npages); 3243 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3244 (u_long)m->valid, (u_long)m->dirty, 3245 m->wire_count); 3246 panic("biodone: page busy < 0\n"); 3247 } 3248 vm_page_io_finish(m); 3249 vm_object_pip_subtract(obj, 1); 3250 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3251 iosize -= resid; 3252 } 3253 vm_page_unlock_queues(); 3254 if (obj != NULL) { 3255 vm_object_pip_wakeupn(obj, 0); 3256 VM_OBJECT_UNLOCK(obj); 3257 } 3258 } 3259 3260 /* 3261 * For asynchronous completions, release the buffer now. The brelse 3262 * will do a wakeup there if necessary - so no need to do a wakeup 3263 * here in the async case. The sync case always needs to do a wakeup. 3264 */ 3265 3266 if (bp->b_flags & B_ASYNC) { 3267 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3268 brelse(bp); 3269 else 3270 bqrelse(bp); 3271 } else { 3272 bdone(bp); 3273 } 3274 splx(s); 3275} 3276 3277/* 3278 * This routine is called in lieu of iodone in the case of 3279 * incomplete I/O. This keeps the busy status for pages 3280 * consistant. 3281 */ 3282void 3283vfs_unbusy_pages(struct buf * bp) 3284{ 3285 int i; 3286 3287 GIANT_REQUIRED; 3288 3289 runningbufwakeup(bp); 3290 if (bp->b_flags & B_VMIO) { 3291 vm_object_t obj; 3292 3293 obj = bp->b_object; 3294 VM_OBJECT_LOCK(obj); 3295 vm_page_lock_queues(); 3296 for (i = 0; i < bp->b_npages; i++) { 3297 vm_page_t m = bp->b_pages[i]; 3298 3299 if (m == bogus_page) { 3300 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3301 if (!m) { 3302 panic("vfs_unbusy_pages: page missing\n"); 3303 } 3304 bp->b_pages[i] = m; 3305 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3306 } 3307 vm_object_pip_subtract(obj, 1); 3308 vm_page_flag_clear(m, PG_ZERO); 3309 vm_page_io_finish(m); 3310 } 3311 vm_page_unlock_queues(); 3312 vm_object_pip_wakeupn(obj, 0); 3313 VM_OBJECT_UNLOCK(obj); 3314 } 3315} 3316 3317/* 3318 * vfs_page_set_valid: 3319 * 3320 * Set the valid bits in a page based on the supplied offset. The 3321 * range is restricted to the buffer's size. 3322 * 3323 * This routine is typically called after a read completes. 3324 */ 3325static void 3326vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3327{ 3328 vm_ooffset_t soff, eoff; 3329 3330 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3331 /* 3332 * Start and end offsets in buffer. eoff - soff may not cross a 3333 * page boundry or cross the end of the buffer. The end of the 3334 * buffer, in this case, is our file EOF, not the allocation size 3335 * of the buffer. 3336 */ 3337 soff = off; 3338 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3339 if (eoff > bp->b_offset + bp->b_bcount) 3340 eoff = bp->b_offset + bp->b_bcount; 3341 3342 /* 3343 * Set valid range. This is typically the entire buffer and thus the 3344 * entire page. 3345 */ 3346 if (eoff > soff) { 3347 vm_page_set_validclean( 3348 m, 3349 (vm_offset_t) (soff & PAGE_MASK), 3350 (vm_offset_t) (eoff - soff) 3351 ); 3352 } 3353} 3354 3355/* 3356 * This routine is called before a device strategy routine. 3357 * It is used to tell the VM system that paging I/O is in 3358 * progress, and treat the pages associated with the buffer 3359 * almost as being PG_BUSY. Also the object paging_in_progress 3360 * flag is handled to make sure that the object doesn't become 3361 * inconsistant. 3362 * 3363 * Since I/O has not been initiated yet, certain buffer flags 3364 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3365 * and should be ignored. 3366 */ 3367void 3368vfs_busy_pages(struct buf * bp, int clear_modify) 3369{ 3370 int i, bogus; 3371 3372 if (bp->b_flags & B_VMIO) { 3373 vm_object_t obj; 3374 vm_ooffset_t foff; 3375 3376 obj = bp->b_object; 3377 foff = bp->b_offset; 3378 KASSERT(bp->b_offset != NOOFFSET, 3379 ("vfs_busy_pages: no buffer offset")); 3380 vfs_setdirty(bp); 3381 if (obj != NULL) 3382 VM_OBJECT_LOCK(obj); 3383retry: 3384 vm_page_lock_queues(); 3385 for (i = 0; i < bp->b_npages; i++) { 3386 vm_page_t m = bp->b_pages[i]; 3387 3388 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3389 goto retry; 3390 } 3391 bogus = 0; 3392 for (i = 0; i < bp->b_npages; i++) { 3393 vm_page_t m = bp->b_pages[i]; 3394 3395 vm_page_flag_clear(m, PG_ZERO); 3396 if ((bp->b_flags & B_CLUSTER) == 0) { 3397 vm_object_pip_add(obj, 1); 3398 vm_page_io_start(m); 3399 } 3400 /* 3401 * When readying a buffer for a read ( i.e 3402 * clear_modify == 0 ), it is important to do 3403 * bogus_page replacement for valid pages in 3404 * partially instantiated buffers. Partially 3405 * instantiated buffers can, in turn, occur when 3406 * reconstituting a buffer from its VM backing store 3407 * base. We only have to do this if B_CACHE is 3408 * clear ( which causes the I/O to occur in the 3409 * first place ). The replacement prevents the read 3410 * I/O from overwriting potentially dirty VM-backed 3411 * pages. XXX bogus page replacement is, uh, bogus. 3412 * It may not work properly with small-block devices. 3413 * We need to find a better way. 3414 */ 3415 pmap_remove_all(m); 3416 if (clear_modify) 3417 vfs_page_set_valid(bp, foff, i, m); 3418 else if (m->valid == VM_PAGE_BITS_ALL && 3419 (bp->b_flags & B_CACHE) == 0) { 3420 bp->b_pages[i] = bogus_page; 3421 bogus++; 3422 } 3423 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3424 } 3425 vm_page_unlock_queues(); 3426 if (obj != NULL) 3427 VM_OBJECT_UNLOCK(obj); 3428 if (bogus) 3429 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3430 } 3431} 3432 3433/* 3434 * Tell the VM system that the pages associated with this buffer 3435 * are clean. This is used for delayed writes where the data is 3436 * going to go to disk eventually without additional VM intevention. 3437 * 3438 * Note that while we only really need to clean through to b_bcount, we 3439 * just go ahead and clean through to b_bufsize. 3440 */ 3441static void 3442vfs_clean_pages(struct buf * bp) 3443{ 3444 int i; 3445 3446 GIANT_REQUIRED; 3447 3448 if (bp->b_flags & B_VMIO) { 3449 vm_ooffset_t foff; 3450 3451 foff = bp->b_offset; 3452 KASSERT(bp->b_offset != NOOFFSET, 3453 ("vfs_clean_pages: no buffer offset")); 3454 vm_page_lock_queues(); 3455 for (i = 0; i < bp->b_npages; i++) { 3456 vm_page_t m = bp->b_pages[i]; 3457 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3458 vm_ooffset_t eoff = noff; 3459 3460 if (eoff > bp->b_offset + bp->b_bufsize) 3461 eoff = bp->b_offset + bp->b_bufsize; 3462 vfs_page_set_valid(bp, foff, i, m); 3463 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3464 foff = noff; 3465 } 3466 vm_page_unlock_queues(); 3467 } 3468} 3469 3470/* 3471 * vfs_bio_set_validclean: 3472 * 3473 * Set the range within the buffer to valid and clean. The range is 3474 * relative to the beginning of the buffer, b_offset. Note that b_offset 3475 * itself may be offset from the beginning of the first page. 3476 * 3477 */ 3478 3479void 3480vfs_bio_set_validclean(struct buf *bp, int base, int size) 3481{ 3482 if (bp->b_flags & B_VMIO) { 3483 int i; 3484 int n; 3485 3486 /* 3487 * Fixup base to be relative to beginning of first page. 3488 * Set initial n to be the maximum number of bytes in the 3489 * first page that can be validated. 3490 */ 3491 3492 base += (bp->b_offset & PAGE_MASK); 3493 n = PAGE_SIZE - (base & PAGE_MASK); 3494 3495 vm_page_lock_queues(); 3496 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3497 vm_page_t m = bp->b_pages[i]; 3498 3499 if (n > size) 3500 n = size; 3501 3502 vm_page_set_validclean(m, base & PAGE_MASK, n); 3503 base += n; 3504 size -= n; 3505 n = PAGE_SIZE; 3506 } 3507 vm_page_unlock_queues(); 3508 } 3509} 3510 3511/* 3512 * vfs_bio_clrbuf: 3513 * 3514 * clear a buffer. This routine essentially fakes an I/O, so we need 3515 * to clear BIO_ERROR and B_INVAL. 3516 * 3517 * Note that while we only theoretically need to clear through b_bcount, 3518 * we go ahead and clear through b_bufsize. 3519 */ 3520 3521void 3522vfs_bio_clrbuf(struct buf *bp) 3523{ 3524 int i, mask = 0; 3525 caddr_t sa, ea; 3526 3527 GIANT_REQUIRED; 3528 3529 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3530 bp->b_flags &= ~B_INVAL; 3531 bp->b_ioflags &= ~BIO_ERROR; 3532 if (bp->b_object != NULL) 3533 VM_OBJECT_LOCK(bp->b_object); 3534 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3535 (bp->b_offset & PAGE_MASK) == 0) { 3536 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3537 if (bp->b_pages[0] != bogus_page) 3538 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3539 if ((bp->b_pages[0]->valid & mask) == mask) 3540 goto unlock; 3541 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3542 ((bp->b_pages[0]->valid & mask) == 0)) { 3543 bzero(bp->b_data, bp->b_bufsize); 3544 bp->b_pages[0]->valid |= mask; 3545 goto unlock; 3546 } 3547 } 3548 ea = sa = bp->b_data; 3549 for(i=0;i<bp->b_npages;i++,sa=ea) { 3550 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3551 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3552 ea = (caddr_t)(vm_offset_t)ulmin( 3553 (u_long)(vm_offset_t)ea, 3554 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3555 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3556 if (bp->b_pages[i] != bogus_page) 3557 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3558 if ((bp->b_pages[i]->valid & mask) == mask) 3559 continue; 3560 if ((bp->b_pages[i]->valid & mask) == 0) { 3561 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3562 bzero(sa, ea - sa); 3563 } 3564 } else { 3565 for (; sa < ea; sa += DEV_BSIZE, j++) { 3566 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3567 (bp->b_pages[i]->valid & (1<<j)) == 0) 3568 bzero(sa, DEV_BSIZE); 3569 } 3570 } 3571 bp->b_pages[i]->valid |= mask; 3572 vm_page_lock_queues(); 3573 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3574 vm_page_unlock_queues(); 3575 } 3576unlock: 3577 if (bp->b_object != NULL) 3578 VM_OBJECT_UNLOCK(bp->b_object); 3579 bp->b_resid = 0; 3580 } else { 3581 clrbuf(bp); 3582 } 3583} 3584 3585/* 3586 * vm_hold_load_pages and vm_hold_free_pages get pages into 3587 * a buffers address space. The pages are anonymous and are 3588 * not associated with a file object. 3589 */ 3590static void 3591vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3592{ 3593 vm_offset_t pg; 3594 vm_page_t p; 3595 int index; 3596 3597 to = round_page(to); 3598 from = round_page(from); 3599 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3600 3601 VM_OBJECT_LOCK(kernel_object); 3602 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3603tryagain: 3604 /* 3605 * note: must allocate system pages since blocking here 3606 * could intefere with paging I/O, no matter which 3607 * process we are. 3608 */ 3609 p = vm_page_alloc(kernel_object, 3610 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3611 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3612 if (!p) { 3613 atomic_add_int(&vm_pageout_deficit, 3614 (to - pg) >> PAGE_SHIFT); 3615 VM_OBJECT_UNLOCK(kernel_object); 3616 VM_WAIT; 3617 VM_OBJECT_LOCK(kernel_object); 3618 goto tryagain; 3619 } 3620 p->valid = VM_PAGE_BITS_ALL; 3621 pmap_qenter(pg, &p, 1); 3622 bp->b_pages[index] = p; 3623 vm_page_lock_queues(); 3624 vm_page_wakeup(p); 3625 vm_page_unlock_queues(); 3626 } 3627 VM_OBJECT_UNLOCK(kernel_object); 3628 bp->b_npages = index; 3629} 3630 3631/* Return pages associated with this buf to the vm system */ 3632static void 3633vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3634{ 3635 vm_offset_t pg; 3636 vm_page_t p; 3637 int index, newnpages; 3638 3639 GIANT_REQUIRED; 3640 3641 from = round_page(from); 3642 to = round_page(to); 3643 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3644 3645 VM_OBJECT_LOCK(kernel_object); 3646 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3647 p = bp->b_pages[index]; 3648 if (p && (index < bp->b_npages)) { 3649 if (p->busy) { 3650 printf( 3651 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3652 (intmax_t)bp->b_blkno, 3653 (intmax_t)bp->b_lblkno); 3654 } 3655 bp->b_pages[index] = NULL; 3656 pmap_qremove(pg, 1); 3657 vm_page_lock_queues(); 3658 vm_page_busy(p); 3659 vm_page_unwire(p, 0); 3660 vm_page_free(p); 3661 vm_page_unlock_queues(); 3662 } 3663 } 3664 VM_OBJECT_UNLOCK(kernel_object); 3665 bp->b_npages = newnpages; 3666} 3667 3668/* 3669 * Map an IO request into kernel virtual address space. 3670 * 3671 * All requests are (re)mapped into kernel VA space. 3672 * Notice that we use b_bufsize for the size of the buffer 3673 * to be mapped. b_bcount might be modified by the driver. 3674 * 3675 * Note that even if the caller determines that the address space should 3676 * be valid, a race or a smaller-file mapped into a larger space may 3677 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3678 * check the return value. 3679 */ 3680int 3681vmapbuf(struct buf *bp) 3682{ 3683 caddr_t addr, kva; 3684 vm_prot_t prot; 3685 int pidx, i; 3686 struct vm_page *m; 3687 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3688 3689 GIANT_REQUIRED; 3690 3691 KASSERT(bp->b_flags & B_PHYS, ("vmapbuf")); 3692 if (bp->b_bufsize < 0) 3693 return (-1); 3694 prot = (bp->b_iocmd == BIO_READ) ? VM_PROT_READ | VM_PROT_WRITE : 3695 VM_PROT_READ; 3696 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3697 addr < bp->b_data + bp->b_bufsize; 3698 addr += PAGE_SIZE, pidx++) { 3699 /* 3700 * Do the vm_fault if needed; do the copy-on-write thing 3701 * when reading stuff off device into memory. 3702 * 3703 * NOTE! Must use pmap_extract() because addr may be in 3704 * the userland address space, and kextract is only guarenteed 3705 * to work for the kernland address space (see: sparc64 port). 3706 */ 3707retry: 3708 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3709 prot) < 0) { 3710 vm_page_lock_queues(); 3711 for (i = 0; i < pidx; ++i) { 3712 vm_page_unhold(bp->b_pages[i]); 3713 bp->b_pages[i] = NULL; 3714 } 3715 vm_page_unlock_queues(); 3716 return(-1); 3717 } 3718 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3719 if (m == NULL) 3720 goto retry; 3721 bp->b_pages[pidx] = m; 3722 } 3723 if (pidx > btoc(MAXPHYS)) 3724 panic("vmapbuf: mapped more than MAXPHYS"); 3725 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3726 3727 kva = bp->b_saveaddr; 3728 bp->b_npages = pidx; 3729 bp->b_saveaddr = bp->b_data; 3730 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3731 return(0); 3732} 3733 3734/* 3735 * Free the io map PTEs associated with this IO operation. 3736 * We also invalidate the TLB entries and restore the original b_addr. 3737 */ 3738void 3739vunmapbuf(struct buf *bp) 3740{ 3741 int pidx; 3742 int npages; 3743 3744 GIANT_REQUIRED; 3745 3746 KASSERT(bp->b_flags & B_PHYS, ("vunmapbuf")); 3747 3748 npages = bp->b_npages; 3749 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3750 npages); 3751 vm_page_lock_queues(); 3752 for (pidx = 0; pidx < npages; pidx++) 3753 vm_page_unhold(bp->b_pages[pidx]); 3754 vm_page_unlock_queues(); 3755 3756 bp->b_data = bp->b_saveaddr; 3757} 3758 3759void 3760bdone(struct buf *bp) 3761{ 3762 mtx_lock(&bdonelock); 3763 bp->b_flags |= B_DONE; 3764 wakeup(bp); 3765 mtx_unlock(&bdonelock); 3766} 3767 3768void 3769bwait(struct buf *bp, u_char pri, const char *wchan) 3770{ 3771 mtx_lock(&bdonelock); 3772 while ((bp->b_flags & B_DONE) == 0) 3773 msleep(bp, &bdonelock, pri, wchan, 0); 3774 mtx_unlock(&bdonelock); 3775} 3776 3777#include "opt_ddb.h" 3778#ifdef DDB 3779#include <ddb/ddb.h> 3780 3781/* DDB command to show buffer data */ 3782DB_SHOW_COMMAND(buffer, db_show_buffer) 3783{ 3784 /* get args */ 3785 struct buf *bp = (struct buf *)addr; 3786 3787 if (!have_addr) { 3788 db_printf("usage: show buffer <addr>\n"); 3789 return; 3790 } 3791 3792 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3793 db_printf( 3794 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3795 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd\n", 3796 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3797 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3798 (intmax_t)bp->b_blkno); 3799 if (bp->b_npages) { 3800 int i; 3801 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3802 for (i = 0; i < bp->b_npages; i++) { 3803 vm_page_t m; 3804 m = bp->b_pages[i]; 3805 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3806 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3807 if ((i + 1) < bp->b_npages) 3808 db_printf(","); 3809 } 3810 db_printf("\n"); 3811 } 3812} 3813#endif /* DDB */ 3814