vfs_bio.c revision 121296
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 121296 2003-10-21 06:53:10Z 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 bp->b_iooffset = dbtob(bp->b_blkno); 711 if (vp->v_type == VCHR) 712 VOP_SPECSTRATEGY(vp, bp); 713 else 714 VOP_STRATEGY(vp, bp); 715 ++readwait; 716 } 717 718 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 719 if (inmem(vp, *rablkno)) 720 continue; 721 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 722 723 if ((rabp->b_flags & B_CACHE) == 0) { 724 if (curthread != PCPU_GET(idlethread)) 725 curthread->td_proc->p_stats->p_ru.ru_inblock++; 726 rabp->b_flags |= B_ASYNC; 727 rabp->b_flags &= ~B_INVAL; 728 rabp->b_ioflags &= ~BIO_ERROR; 729 rabp->b_iocmd = BIO_READ; 730 if (rabp->b_rcred == NOCRED && cred != NOCRED) 731 rabp->b_rcred = crhold(cred); 732 vfs_busy_pages(rabp, 0); 733 BUF_KERNPROC(rabp); 734 rabp->b_iooffset = dbtob(rabp->b_blkno); 735 if (vp->v_type == VCHR) 736 VOP_SPECSTRATEGY(vp, rabp); 737 else 738 VOP_STRATEGY(vp, rabp); 739 } else { 740 brelse(rabp); 741 } 742 } 743 744 if (readwait) { 745 rv = bufwait(bp); 746 } 747 return (rv); 748} 749 750/* 751 * Write, release buffer on completion. (Done by iodone 752 * if async). Do not bother writing anything if the buffer 753 * is invalid. 754 * 755 * Note that we set B_CACHE here, indicating that buffer is 756 * fully valid and thus cacheable. This is true even of NFS 757 * now so we set it generally. This could be set either here 758 * or in biodone() since the I/O is synchronous. We put it 759 * here. 760 */ 761 762int 763bwrite(struct buf * bp) 764{ 765 int oldflags, s; 766 struct buf *newbp; 767 768 if (bp->b_flags & B_INVAL) { 769 brelse(bp); 770 return (0); 771 } 772 773 oldflags = bp->b_flags; 774 775 if (BUF_REFCNT(bp) == 0) 776 panic("bwrite: buffer is not busy???"); 777 s = splbio(); 778 /* 779 * If a background write is already in progress, delay 780 * writing this block if it is asynchronous. Otherwise 781 * wait for the background write to complete. 782 */ 783 VI_LOCK(bp->b_vp); 784 if (bp->b_vflags & BV_BKGRDINPROG) { 785 if (bp->b_flags & B_ASYNC) { 786 VI_UNLOCK(bp->b_vp); 787 splx(s); 788 bdwrite(bp); 789 return (0); 790 } 791 bp->b_vflags |= BV_BKGRDWAIT; 792 msleep(&bp->b_xflags, VI_MTX(bp->b_vp), PRIBIO, "bwrbg", 0); 793 if (bp->b_vflags & BV_BKGRDINPROG) 794 panic("bwrite: still writing"); 795 } 796 VI_UNLOCK(bp->b_vp); 797 798 /* Mark the buffer clean */ 799 bundirty(bp); 800 801 /* 802 * If this buffer is marked for background writing and we 803 * do not have to wait for it, make a copy and write the 804 * copy so as to leave this buffer ready for further use. 805 * 806 * This optimization eats a lot of memory. If we have a page 807 * or buffer shortfall we can't do it. 808 */ 809 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 810 (bp->b_flags & B_ASYNC) && 811 !vm_page_count_severe() && 812 !buf_dirty_count_severe()) { 813 if (bp->b_iodone != NULL) { 814 printf("bp->b_iodone = %p\n", bp->b_iodone); 815 panic("bwrite: need chained iodone"); 816 } 817 818 /* get a new block */ 819 newbp = geteblk(bp->b_bufsize); 820 821 /* 822 * set it to be identical to the old block. We have to 823 * set b_lblkno and BKGRDMARKER before calling bgetvp() 824 * to avoid confusing the splay tree and gbincore(). 825 */ 826 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 827 newbp->b_lblkno = bp->b_lblkno; 828 newbp->b_xflags |= BX_BKGRDMARKER; 829 VI_LOCK(bp->b_vp); 830 bp->b_vflags |= BV_BKGRDINPROG; 831 bgetvp(bp->b_vp, newbp); 832 VI_UNLOCK(bp->b_vp); 833 newbp->b_blkno = bp->b_blkno; 834 newbp->b_offset = bp->b_offset; 835 newbp->b_iodone = vfs_backgroundwritedone; 836 newbp->b_flags |= B_ASYNC; 837 newbp->b_flags &= ~B_INVAL; 838 839 /* move over the dependencies */ 840 if (LIST_FIRST(&bp->b_dep) != NULL) 841 buf_movedeps(bp, newbp); 842 843 /* 844 * Initiate write on the copy, release the original to 845 * the B_LOCKED queue so that it cannot go away until 846 * the background write completes. If not locked it could go 847 * away and then be reconstituted while it was being written. 848 * If the reconstituted buffer were written, we could end up 849 * with two background copies being written at the same time. 850 */ 851 bqrelse(bp); 852 bp = newbp; 853 } 854 855 bp->b_flags &= ~B_DONE; 856 bp->b_ioflags &= ~BIO_ERROR; 857 bp->b_flags |= B_WRITEINPROG | B_CACHE; 858 bp->b_iocmd = BIO_WRITE; 859 860 VI_LOCK(bp->b_vp); 861 bp->b_vp->v_numoutput++; 862 VI_UNLOCK(bp->b_vp); 863 vfs_busy_pages(bp, 1); 864 865 /* 866 * Normal bwrites pipeline writes 867 */ 868 bp->b_runningbufspace = bp->b_bufsize; 869 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 870 871 if (curthread != PCPU_GET(idlethread)) 872 curthread->td_proc->p_stats->p_ru.ru_oublock++; 873 splx(s); 874 if (oldflags & B_ASYNC) 875 BUF_KERNPROC(bp); 876 bp->b_iooffset = dbtob(bp->b_blkno); 877 if (bp->b_vp->v_type == VCHR) 878 VOP_SPECSTRATEGY(bp->b_vp, bp); 879 else 880 VOP_STRATEGY(bp->b_vp, bp); 881 882 if ((oldflags & B_ASYNC) == 0) { 883 int rtval = bufwait(bp); 884 brelse(bp); 885 return (rtval); 886 } else { 887 /* 888 * don't allow the async write to saturate the I/O 889 * system. We will not deadlock here because 890 * we are blocking waiting for I/O that is already in-progress 891 * to complete. 892 */ 893 waitrunningbufspace(); 894 } 895 896 return (0); 897} 898 899/* 900 * Complete a background write started from bwrite. 901 */ 902static void 903vfs_backgroundwritedone(bp) 904 struct buf *bp; 905{ 906 struct buf *origbp; 907 908 /* 909 * Find the original buffer that we are writing. 910 */ 911 VI_LOCK(bp->b_vp); 912 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 913 panic("backgroundwritedone: lost buffer"); 914 915 /* 916 * Clear the BV_BKGRDINPROG flag in the original buffer 917 * and awaken it if it is waiting for the write to complete. 918 * If BV_BKGRDINPROG is not set in the original buffer it must 919 * have been released and re-instantiated - which is not legal. 920 */ 921 KASSERT((origbp->b_vflags & BV_BKGRDINPROG), 922 ("backgroundwritedone: lost buffer2")); 923 origbp->b_vflags &= ~BV_BKGRDINPROG; 924 if (origbp->b_vflags & BV_BKGRDWAIT) { 925 origbp->b_vflags &= ~BV_BKGRDWAIT; 926 wakeup(&origbp->b_xflags); 927 } 928 VI_UNLOCK(bp->b_vp); 929 /* 930 * Process dependencies then return any unfinished ones. 931 */ 932 if (LIST_FIRST(&bp->b_dep) != NULL) 933 buf_complete(bp); 934 if (LIST_FIRST(&bp->b_dep) != NULL) 935 buf_movedeps(bp, origbp); 936 937 /* 938 * This buffer is marked B_NOCACHE, so when it is released 939 * by biodone, it will be tossed. We mark it with BIO_READ 940 * to avoid biodone doing a second vwakeup. 941 */ 942 bp->b_flags |= B_NOCACHE; 943 bp->b_iocmd = BIO_READ; 944 bp->b_flags &= ~(B_CACHE | B_DONE); 945 bp->b_iodone = 0; 946 bufdone(bp); 947} 948 949/* 950 * Delayed write. (Buffer is marked dirty). Do not bother writing 951 * anything if the buffer is marked invalid. 952 * 953 * Note that since the buffer must be completely valid, we can safely 954 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 955 * biodone() in order to prevent getblk from writing the buffer 956 * out synchronously. 957 */ 958void 959bdwrite(struct buf * bp) 960{ 961 struct thread *td = curthread; 962 struct vnode *vp; 963 struct buf *nbp; 964 965 GIANT_REQUIRED; 966 967 if (BUF_REFCNT(bp) == 0) 968 panic("bdwrite: buffer is not busy"); 969 970 if (bp->b_flags & B_INVAL) { 971 brelse(bp); 972 return; 973 } 974 975 /* 976 * If we have too many dirty buffers, don't create any more. 977 * If we are wildly over our limit, then force a complete 978 * cleanup. Otherwise, just keep the situation from getting 979 * out of control. Note that we have to avoid a recursive 980 * disaster and not try to clean up after our own cleanup! 981 */ 982 vp = bp->b_vp; 983 VI_LOCK(vp); 984 if (td->td_proc->p_flag & P_COWINPROGRESS) { 985 recursiveflushes++; 986 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) { 987 VI_UNLOCK(vp); 988 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 989 VI_LOCK(vp); 990 altbufferflushes++; 991 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) { 992 /* 993 * Try to find a buffer to flush. 994 */ 995 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) { 996 if ((nbp->b_vflags & BV_BKGRDINPROG) || 997 buf_countdeps(nbp, 0) || 998 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 999 continue; 1000 if (bp == nbp) 1001 panic("bdwrite: found ourselves"); 1002 VI_UNLOCK(vp); 1003 if (nbp->b_flags & B_CLUSTEROK) { 1004 vfs_bio_awrite(nbp); 1005 } else { 1006 bremfree(nbp); 1007 bawrite(nbp); 1008 } 1009 VI_LOCK(vp); 1010 dirtybufferflushes++; 1011 break; 1012 } 1013 } 1014 VI_UNLOCK(vp); 1015 1016 bdirty(bp); 1017 /* 1018 * Set B_CACHE, indicating that the buffer is fully valid. This is 1019 * true even of NFS now. 1020 */ 1021 bp->b_flags |= B_CACHE; 1022 1023 /* 1024 * This bmap keeps the system from needing to do the bmap later, 1025 * perhaps when the system is attempting to do a sync. Since it 1026 * is likely that the indirect block -- or whatever other datastructure 1027 * that the filesystem needs is still in memory now, it is a good 1028 * thing to do this. Note also, that if the pageout daemon is 1029 * requesting a sync -- there might not be enough memory to do 1030 * the bmap then... So, this is important to do. 1031 */ 1032 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1033 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1034 } 1035 1036 /* 1037 * Set the *dirty* buffer range based upon the VM system dirty pages. 1038 */ 1039 vfs_setdirty(bp); 1040 1041 /* 1042 * We need to do this here to satisfy the vnode_pager and the 1043 * pageout daemon, so that it thinks that the pages have been 1044 * "cleaned". Note that since the pages are in a delayed write 1045 * buffer -- the VFS layer "will" see that the pages get written 1046 * out on the next sync, or perhaps the cluster will be completed. 1047 */ 1048 vfs_clean_pages(bp); 1049 bqrelse(bp); 1050 1051 /* 1052 * Wakeup the buffer flushing daemon if we have a lot of dirty 1053 * buffers (midpoint between our recovery point and our stall 1054 * point). 1055 */ 1056 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1057 1058 /* 1059 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1060 * due to the softdep code. 1061 */ 1062} 1063 1064/* 1065 * bdirty: 1066 * 1067 * Turn buffer into delayed write request. We must clear BIO_READ and 1068 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1069 * itself to properly update it in the dirty/clean lists. We mark it 1070 * B_DONE to ensure that any asynchronization of the buffer properly 1071 * clears B_DONE ( else a panic will occur later ). 1072 * 1073 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1074 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1075 * should only be called if the buffer is known-good. 1076 * 1077 * Since the buffer is not on a queue, we do not update the numfreebuffers 1078 * count. 1079 * 1080 * Must be called at splbio(). 1081 * The buffer must be on QUEUE_NONE. 1082 */ 1083void 1084bdirty(bp) 1085 struct buf *bp; 1086{ 1087 KASSERT(bp->b_qindex == QUEUE_NONE, 1088 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1089 bp->b_flags &= ~(B_RELBUF); 1090 bp->b_iocmd = BIO_WRITE; 1091 1092 if ((bp->b_flags & B_DELWRI) == 0) { 1093 bp->b_flags |= B_DONE | B_DELWRI; 1094 reassignbuf(bp, bp->b_vp); 1095 atomic_add_int(&numdirtybuffers, 1); 1096 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1097 } 1098} 1099 1100/* 1101 * bundirty: 1102 * 1103 * Clear B_DELWRI for buffer. 1104 * 1105 * Since the buffer is not on a queue, we do not update the numfreebuffers 1106 * count. 1107 * 1108 * Must be called at splbio(). 1109 * The buffer must be on QUEUE_NONE. 1110 */ 1111 1112void 1113bundirty(bp) 1114 struct buf *bp; 1115{ 1116 KASSERT(bp->b_qindex == QUEUE_NONE, 1117 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1118 1119 if (bp->b_flags & B_DELWRI) { 1120 bp->b_flags &= ~B_DELWRI; 1121 reassignbuf(bp, bp->b_vp); 1122 atomic_subtract_int(&numdirtybuffers, 1); 1123 numdirtywakeup(lodirtybuffers); 1124 } 1125 /* 1126 * Since it is now being written, we can clear its deferred write flag. 1127 */ 1128 bp->b_flags &= ~B_DEFERRED; 1129} 1130 1131/* 1132 * bawrite: 1133 * 1134 * Asynchronous write. Start output on a buffer, but do not wait for 1135 * it to complete. The buffer is released when the output completes. 1136 * 1137 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1138 * B_INVAL buffers. Not us. 1139 */ 1140void 1141bawrite(struct buf * bp) 1142{ 1143 bp->b_flags |= B_ASYNC; 1144 (void) BUF_WRITE(bp); 1145} 1146 1147/* 1148 * bwillwrite: 1149 * 1150 * Called prior to the locking of any vnodes when we are expecting to 1151 * write. We do not want to starve the buffer cache with too many 1152 * dirty buffers so we block here. By blocking prior to the locking 1153 * of any vnodes we attempt to avoid the situation where a locked vnode 1154 * prevents the various system daemons from flushing related buffers. 1155 */ 1156 1157void 1158bwillwrite(void) 1159{ 1160 if (numdirtybuffers >= hidirtybuffers) { 1161 int s; 1162 1163 mtx_lock(&Giant); 1164 s = splbio(); 1165 mtx_lock(&nblock); 1166 while (numdirtybuffers >= hidirtybuffers) { 1167 bd_wakeup(1); 1168 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1169 msleep(&needsbuffer, &nblock, 1170 (PRIBIO + 4), "flswai", 0); 1171 } 1172 splx(s); 1173 mtx_unlock(&nblock); 1174 mtx_unlock(&Giant); 1175 } 1176} 1177 1178/* 1179 * Return true if we have too many dirty buffers. 1180 */ 1181int 1182buf_dirty_count_severe(void) 1183{ 1184 return(numdirtybuffers >= hidirtybuffers); 1185} 1186 1187/* 1188 * brelse: 1189 * 1190 * Release a busy buffer and, if requested, free its resources. The 1191 * buffer will be stashed in the appropriate bufqueue[] allowing it 1192 * to be accessed later as a cache entity or reused for other purposes. 1193 */ 1194void 1195brelse(struct buf * bp) 1196{ 1197 int s; 1198 1199 GIANT_REQUIRED; 1200 1201 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1202 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1203 1204 s = splbio(); 1205 1206 if (bp->b_iocmd == BIO_WRITE && 1207 (bp->b_ioflags & BIO_ERROR) && 1208 !(bp->b_flags & B_INVAL)) { 1209 /* 1210 * Failed write, redirty. Must clear BIO_ERROR to prevent 1211 * pages from being scrapped. If B_INVAL is set then 1212 * this case is not run and the next case is run to 1213 * destroy the buffer. B_INVAL can occur if the buffer 1214 * is outside the range supported by the underlying device. 1215 */ 1216 bp->b_ioflags &= ~BIO_ERROR; 1217 bdirty(bp); 1218 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1219 (bp->b_ioflags & BIO_ERROR) || 1220 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1221 /* 1222 * Either a failed I/O or we were asked to free or not 1223 * cache the buffer. 1224 */ 1225 bp->b_flags |= B_INVAL; 1226 if (LIST_FIRST(&bp->b_dep) != NULL) 1227 buf_deallocate(bp); 1228 if (bp->b_flags & B_DELWRI) { 1229 atomic_subtract_int(&numdirtybuffers, 1); 1230 numdirtywakeup(lodirtybuffers); 1231 } 1232 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1233 if ((bp->b_flags & B_VMIO) == 0) { 1234 if (bp->b_bufsize) 1235 allocbuf(bp, 0); 1236 if (bp->b_vp) 1237 brelvp(bp); 1238 } 1239 } 1240 1241 /* 1242 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1243 * is called with B_DELWRI set, the underlying pages may wind up 1244 * getting freed causing a previous write (bdwrite()) to get 'lost' 1245 * because pages associated with a B_DELWRI bp are marked clean. 1246 * 1247 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1248 * if B_DELWRI is set. 1249 * 1250 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1251 * on pages to return pages to the VM page queues. 1252 */ 1253 if (bp->b_flags & B_DELWRI) 1254 bp->b_flags &= ~B_RELBUF; 1255 else if (vm_page_count_severe()) { 1256 /* 1257 * XXX This lock may not be necessary since BKGRDINPROG 1258 * cannot be set while we hold the buf lock, it can only be 1259 * cleared if it is already pending. 1260 */ 1261 if (bp->b_vp) { 1262 VI_LOCK(bp->b_vp); 1263 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1264 bp->b_flags |= B_RELBUF; 1265 VI_UNLOCK(bp->b_vp); 1266 } else 1267 bp->b_flags |= B_RELBUF; 1268 } 1269 1270 /* 1271 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1272 * constituted, not even NFS buffers now. Two flags effect this. If 1273 * B_INVAL, the struct buf is invalidated but the VM object is kept 1274 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1275 * 1276 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1277 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1278 * buffer is also B_INVAL because it hits the re-dirtying code above. 1279 * 1280 * Normally we can do this whether a buffer is B_DELWRI or not. If 1281 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1282 * the commit state and we cannot afford to lose the buffer. If the 1283 * buffer has a background write in progress, we need to keep it 1284 * around to prevent it from being reconstituted and starting a second 1285 * background write. 1286 */ 1287 if ((bp->b_flags & B_VMIO) 1288 && !(bp->b_vp->v_mount != NULL && 1289 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1290 !vn_isdisk(bp->b_vp, NULL) && 1291 (bp->b_flags & B_DELWRI)) 1292 ) { 1293 1294 int i, j, resid; 1295 vm_page_t m; 1296 off_t foff; 1297 vm_pindex_t poff; 1298 vm_object_t obj; 1299 struct vnode *vp; 1300 1301 vp = bp->b_vp; 1302 obj = bp->b_object; 1303 1304 /* 1305 * Get the base offset and length of the buffer. Note that 1306 * in the VMIO case if the buffer block size is not 1307 * page-aligned then b_data pointer may not be page-aligned. 1308 * But our b_pages[] array *IS* page aligned. 1309 * 1310 * block sizes less then DEV_BSIZE (usually 512) are not 1311 * supported due to the page granularity bits (m->valid, 1312 * m->dirty, etc...). 1313 * 1314 * See man buf(9) for more information 1315 */ 1316 resid = bp->b_bufsize; 1317 foff = bp->b_offset; 1318 if (obj != NULL) 1319 VM_OBJECT_LOCK(obj); 1320 for (i = 0; i < bp->b_npages; i++) { 1321 int had_bogus = 0; 1322 1323 m = bp->b_pages[i]; 1324 vm_page_lock_queues(); 1325 vm_page_flag_clear(m, PG_ZERO); 1326 vm_page_unlock_queues(); 1327 1328 /* 1329 * If we hit a bogus page, fixup *all* the bogus pages 1330 * now. 1331 */ 1332 if (m == bogus_page) { 1333 poff = OFF_TO_IDX(bp->b_offset); 1334 had_bogus = 1; 1335 1336 for (j = i; j < bp->b_npages; j++) { 1337 vm_page_t mtmp; 1338 mtmp = bp->b_pages[j]; 1339 if (mtmp == bogus_page) { 1340 mtmp = vm_page_lookup(obj, poff + j); 1341 if (!mtmp) { 1342 panic("brelse: page missing\n"); 1343 } 1344 bp->b_pages[j] = mtmp; 1345 } 1346 } 1347 1348 if ((bp->b_flags & B_INVAL) == 0) { 1349 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1350 } 1351 m = bp->b_pages[i]; 1352 } 1353 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1354 int poffset = foff & PAGE_MASK; 1355 int presid = resid > (PAGE_SIZE - poffset) ? 1356 (PAGE_SIZE - poffset) : resid; 1357 1358 KASSERT(presid >= 0, ("brelse: extra page")); 1359 vm_page_lock_queues(); 1360 vm_page_set_invalid(m, poffset, presid); 1361 vm_page_unlock_queues(); 1362 if (had_bogus) 1363 printf("avoided corruption bug in bogus_page/brelse code\n"); 1364 } 1365 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1366 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1367 } 1368 if (obj != NULL) 1369 VM_OBJECT_UNLOCK(obj); 1370 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1371 vfs_vmio_release(bp); 1372 1373 } else if (bp->b_flags & B_VMIO) { 1374 1375 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1376 vfs_vmio_release(bp); 1377 } 1378 1379 } 1380 1381 if (bp->b_qindex != QUEUE_NONE) 1382 panic("brelse: free buffer onto another queue???"); 1383 if (BUF_REFCNT(bp) > 1) { 1384 /* do not release to free list */ 1385 BUF_UNLOCK(bp); 1386 splx(s); 1387 return; 1388 } 1389 1390 /* enqueue */ 1391 mtx_lock(&bqlock); 1392 1393 /* buffers with no memory */ 1394 if (bp->b_bufsize == 0) { 1395 bp->b_flags |= B_INVAL; 1396 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1397 if (bp->b_vflags & BV_BKGRDINPROG) 1398 panic("losing buffer 1"); 1399 if (bp->b_kvasize) { 1400 bp->b_qindex = QUEUE_EMPTYKVA; 1401 } else { 1402 bp->b_qindex = QUEUE_EMPTY; 1403 } 1404 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1405 bp->b_dev = NODEV; 1406 /* buffers with junk contents */ 1407 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1408 (bp->b_ioflags & BIO_ERROR)) { 1409 bp->b_flags |= B_INVAL; 1410 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1411 if (bp->b_vflags & BV_BKGRDINPROG) 1412 panic("losing buffer 2"); 1413 bp->b_qindex = QUEUE_CLEAN; 1414 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1415 bp->b_dev = NODEV; 1416 /* remaining buffers */ 1417 } else { 1418 if (bp->b_flags & B_DELWRI) 1419 bp->b_qindex = QUEUE_DIRTY; 1420 else 1421 bp->b_qindex = QUEUE_CLEAN; 1422 if (bp->b_flags & B_AGE) 1423 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1424 else 1425 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1426 } 1427 mtx_unlock(&bqlock); 1428 1429 /* 1430 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1431 * placed the buffer on the correct queue. We must also disassociate 1432 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1433 * find it. 1434 */ 1435 if (bp->b_flags & B_INVAL) { 1436 if (bp->b_flags & B_DELWRI) 1437 bundirty(bp); 1438 if (bp->b_vp) 1439 brelvp(bp); 1440 } 1441 1442 /* 1443 * Fixup numfreebuffers count. The bp is on an appropriate queue 1444 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1445 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1446 * if B_INVAL is set ). 1447 */ 1448 1449 if (!(bp->b_flags & B_DELWRI)) 1450 bufcountwakeup(); 1451 1452 /* 1453 * Something we can maybe free or reuse 1454 */ 1455 if (bp->b_bufsize || bp->b_kvasize) 1456 bufspacewakeup(); 1457 1458 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1459 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1460 panic("brelse: not dirty"); 1461 /* unlock */ 1462 BUF_UNLOCK(bp); 1463 splx(s); 1464} 1465 1466/* 1467 * Release a buffer back to the appropriate queue but do not try to free 1468 * it. The buffer is expected to be used again soon. 1469 * 1470 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1471 * biodone() to requeue an async I/O on completion. It is also used when 1472 * known good buffers need to be requeued but we think we may need the data 1473 * again soon. 1474 * 1475 * XXX we should be able to leave the B_RELBUF hint set on completion. 1476 */ 1477void 1478bqrelse(struct buf * bp) 1479{ 1480 int s; 1481 1482 s = splbio(); 1483 1484 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1485 1486 if (bp->b_qindex != QUEUE_NONE) 1487 panic("bqrelse: free buffer onto another queue???"); 1488 if (BUF_REFCNT(bp) > 1) { 1489 /* do not release to free list */ 1490 BUF_UNLOCK(bp); 1491 splx(s); 1492 return; 1493 } 1494 mtx_lock(&bqlock); 1495 /* buffers with stale but valid contents */ 1496 if (bp->b_flags & B_DELWRI) { 1497 bp->b_qindex = QUEUE_DIRTY; 1498 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1499 } else { 1500 /* 1501 * XXX This lock may not be necessary since BKGRDINPROG 1502 * cannot be set while we hold the buf lock, it can only be 1503 * cleared if it is already pending. 1504 */ 1505 VI_LOCK(bp->b_vp); 1506 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1507 VI_UNLOCK(bp->b_vp); 1508 bp->b_qindex = QUEUE_CLEAN; 1509 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1510 b_freelist); 1511 } else { 1512 /* 1513 * We are too low on memory, we have to try to free 1514 * the buffer (most importantly: the wired pages 1515 * making up its backing store) *now*. 1516 */ 1517 VI_UNLOCK(bp->b_vp); 1518 mtx_unlock(&bqlock); 1519 splx(s); 1520 brelse(bp); 1521 return; 1522 } 1523 } 1524 mtx_unlock(&bqlock); 1525 1526 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1527 bufcountwakeup(); 1528 1529 /* 1530 * Something we can maybe free or reuse. 1531 */ 1532 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1533 bufspacewakeup(); 1534 1535 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1536 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1537 panic("bqrelse: not dirty"); 1538 /* unlock */ 1539 BUF_UNLOCK(bp); 1540 splx(s); 1541} 1542 1543/* Give pages used by the bp back to the VM system (where possible) */ 1544static void 1545vfs_vmio_release(bp) 1546 struct buf *bp; 1547{ 1548 int i; 1549 vm_page_t m; 1550 1551 GIANT_REQUIRED; 1552 if (bp->b_object != NULL) 1553 VM_OBJECT_LOCK(bp->b_object); 1554 vm_page_lock_queues(); 1555 for (i = 0; i < bp->b_npages; i++) { 1556 m = bp->b_pages[i]; 1557 bp->b_pages[i] = NULL; 1558 /* 1559 * In order to keep page LRU ordering consistent, put 1560 * everything on the inactive queue. 1561 */ 1562 vm_page_unwire(m, 0); 1563 /* 1564 * We don't mess with busy pages, it is 1565 * the responsibility of the process that 1566 * busied the pages to deal with them. 1567 */ 1568 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1569 continue; 1570 1571 if (m->wire_count == 0) { 1572 vm_page_flag_clear(m, PG_ZERO); 1573 /* 1574 * Might as well free the page if we can and it has 1575 * no valid data. We also free the page if the 1576 * buffer was used for direct I/O 1577 */ 1578 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1579 m->hold_count == 0) { 1580 vm_page_busy(m); 1581 pmap_remove_all(m); 1582 vm_page_free(m); 1583 } else if (bp->b_flags & B_DIRECT) { 1584 vm_page_try_to_free(m); 1585 } else if (vm_page_count_severe()) { 1586 vm_page_try_to_cache(m); 1587 } 1588 } 1589 } 1590 vm_page_unlock_queues(); 1591 if (bp->b_object != NULL) 1592 VM_OBJECT_UNLOCK(bp->b_object); 1593 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1594 1595 if (bp->b_bufsize) { 1596 bufspacewakeup(); 1597 bp->b_bufsize = 0; 1598 } 1599 bp->b_npages = 0; 1600 bp->b_flags &= ~B_VMIO; 1601 if (bp->b_vp) 1602 brelvp(bp); 1603} 1604 1605/* 1606 * Check to see if a block at a particular lbn is available for a clustered 1607 * write. 1608 */ 1609static int 1610vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1611{ 1612 struct buf *bpa; 1613 int match; 1614 1615 match = 0; 1616 1617 /* If the buf isn't in core skip it */ 1618 if ((bpa = gbincore(vp, lblkno)) == NULL) 1619 return (0); 1620 1621 /* If the buf is busy we don't want to wait for it */ 1622 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1623 return (0); 1624 1625 /* Only cluster with valid clusterable delayed write buffers */ 1626 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1627 (B_DELWRI | B_CLUSTEROK)) 1628 goto done; 1629 1630 if (bpa->b_bufsize != size) 1631 goto done; 1632 1633 /* 1634 * Check to see if it is in the expected place on disk and that the 1635 * block has been mapped. 1636 */ 1637 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1638 match = 1; 1639done: 1640 BUF_UNLOCK(bpa); 1641 return (match); 1642} 1643 1644/* 1645 * vfs_bio_awrite: 1646 * 1647 * Implement clustered async writes for clearing out B_DELWRI buffers. 1648 * This is much better then the old way of writing only one buffer at 1649 * a time. Note that we may not be presented with the buffers in the 1650 * correct order, so we search for the cluster in both directions. 1651 */ 1652int 1653vfs_bio_awrite(struct buf * bp) 1654{ 1655 int i; 1656 int j; 1657 daddr_t lblkno = bp->b_lblkno; 1658 struct vnode *vp = bp->b_vp; 1659 int s; 1660 int ncl; 1661 int nwritten; 1662 int size; 1663 int maxcl; 1664 1665 s = splbio(); 1666 /* 1667 * right now we support clustered writing only to regular files. If 1668 * we find a clusterable block we could be in the middle of a cluster 1669 * rather then at the beginning. 1670 */ 1671 if ((vp->v_type == VREG) && 1672 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1673 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1674 1675 size = vp->v_mount->mnt_stat.f_iosize; 1676 maxcl = MAXPHYS / size; 1677 1678 VI_LOCK(vp); 1679 for (i = 1; i < maxcl; i++) 1680 if (vfs_bio_clcheck(vp, size, lblkno + i, 1681 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1682 break; 1683 1684 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1685 if (vfs_bio_clcheck(vp, size, lblkno - j, 1686 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1687 break; 1688 1689 VI_UNLOCK(vp); 1690 --j; 1691 ncl = i + j; 1692 /* 1693 * this is a possible cluster write 1694 */ 1695 if (ncl != 1) { 1696 BUF_UNLOCK(bp); 1697 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1698 splx(s); 1699 return nwritten; 1700 } 1701 } 1702 1703 bremfree(bp); 1704 bp->b_flags |= B_ASYNC; 1705 1706 splx(s); 1707 /* 1708 * default (old) behavior, writing out only one block 1709 * 1710 * XXX returns b_bufsize instead of b_bcount for nwritten? 1711 */ 1712 nwritten = bp->b_bufsize; 1713 (void) BUF_WRITE(bp); 1714 1715 return nwritten; 1716} 1717 1718/* 1719 * getnewbuf: 1720 * 1721 * Find and initialize a new buffer header, freeing up existing buffers 1722 * in the bufqueues as necessary. The new buffer is returned locked. 1723 * 1724 * Important: B_INVAL is not set. If the caller wishes to throw the 1725 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1726 * 1727 * We block if: 1728 * We have insufficient buffer headers 1729 * We have insufficient buffer space 1730 * buffer_map is too fragmented ( space reservation fails ) 1731 * If we have to flush dirty buffers ( but we try to avoid this ) 1732 * 1733 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1734 * Instead we ask the buf daemon to do it for us. We attempt to 1735 * avoid piecemeal wakeups of the pageout daemon. 1736 */ 1737 1738static struct buf * 1739getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1740{ 1741 struct buf *bp; 1742 struct buf *nbp; 1743 int defrag = 0; 1744 int nqindex; 1745 static int flushingbufs; 1746 1747 GIANT_REQUIRED; 1748 1749 /* 1750 * We can't afford to block since we might be holding a vnode lock, 1751 * which may prevent system daemons from running. We deal with 1752 * low-memory situations by proactively returning memory and running 1753 * async I/O rather then sync I/O. 1754 */ 1755 1756 atomic_add_int(&getnewbufcalls, 1); 1757 atomic_subtract_int(&getnewbufrestarts, 1); 1758restart: 1759 atomic_add_int(&getnewbufrestarts, 1); 1760 1761 /* 1762 * Setup for scan. If we do not have enough free buffers, 1763 * we setup a degenerate case that immediately fails. Note 1764 * that if we are specially marked process, we are allowed to 1765 * dip into our reserves. 1766 * 1767 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1768 * 1769 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1770 * However, there are a number of cases (defragging, reusing, ...) 1771 * where we cannot backup. 1772 */ 1773 mtx_lock(&bqlock); 1774 nqindex = QUEUE_EMPTYKVA; 1775 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1776 1777 if (nbp == NULL) { 1778 /* 1779 * If no EMPTYKVA buffers and we are either 1780 * defragging or reusing, locate a CLEAN buffer 1781 * to free or reuse. If bufspace useage is low 1782 * skip this step so we can allocate a new buffer. 1783 */ 1784 if (defrag || bufspace >= lobufspace) { 1785 nqindex = QUEUE_CLEAN; 1786 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1787 } 1788 1789 /* 1790 * If we could not find or were not allowed to reuse a 1791 * CLEAN buffer, check to see if it is ok to use an EMPTY 1792 * buffer. We can only use an EMPTY buffer if allocating 1793 * its KVA would not otherwise run us out of buffer space. 1794 */ 1795 if (nbp == NULL && defrag == 0 && 1796 bufspace + maxsize < hibufspace) { 1797 nqindex = QUEUE_EMPTY; 1798 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1799 } 1800 } 1801 1802 /* 1803 * Run scan, possibly freeing data and/or kva mappings on the fly 1804 * depending. 1805 */ 1806 1807 while ((bp = nbp) != NULL) { 1808 int qindex = nqindex; 1809 1810 /* 1811 * Calculate next bp ( we can only use it if we do not block 1812 * or do other fancy things ). 1813 */ 1814 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1815 switch(qindex) { 1816 case QUEUE_EMPTY: 1817 nqindex = QUEUE_EMPTYKVA; 1818 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1819 break; 1820 /* FALLTHROUGH */ 1821 case QUEUE_EMPTYKVA: 1822 nqindex = QUEUE_CLEAN; 1823 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1824 break; 1825 /* FALLTHROUGH */ 1826 case QUEUE_CLEAN: 1827 /* 1828 * nbp is NULL. 1829 */ 1830 break; 1831 } 1832 } 1833 if (bp->b_vp) { 1834 VI_LOCK(bp->b_vp); 1835 if (bp->b_vflags & BV_BKGRDINPROG) { 1836 VI_UNLOCK(bp->b_vp); 1837 continue; 1838 } 1839 VI_UNLOCK(bp->b_vp); 1840 } 1841 1842 /* 1843 * Sanity Checks 1844 */ 1845 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1846 1847 /* 1848 * Note: we no longer distinguish between VMIO and non-VMIO 1849 * buffers. 1850 */ 1851 1852 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1853 1854 /* 1855 * If we are defragging then we need a buffer with 1856 * b_kvasize != 0. XXX this situation should no longer 1857 * occur, if defrag is non-zero the buffer's b_kvasize 1858 * should also be non-zero at this point. XXX 1859 */ 1860 if (defrag && bp->b_kvasize == 0) { 1861 printf("Warning: defrag empty buffer %p\n", bp); 1862 continue; 1863 } 1864 1865 /* 1866 * Start freeing the bp. This is somewhat involved. nbp 1867 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1868 */ 1869 1870 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1871 panic("getnewbuf: locked buf"); 1872 bremfreel(bp); 1873 mtx_unlock(&bqlock); 1874 1875 if (qindex == QUEUE_CLEAN) { 1876 if (bp->b_flags & B_VMIO) { 1877 bp->b_flags &= ~B_ASYNC; 1878 vfs_vmio_release(bp); 1879 } 1880 if (bp->b_vp) 1881 brelvp(bp); 1882 } 1883 1884 /* 1885 * NOTE: nbp is now entirely invalid. We can only restart 1886 * the scan from this point on. 1887 * 1888 * Get the rest of the buffer freed up. b_kva* is still 1889 * valid after this operation. 1890 */ 1891 1892 if (bp->b_rcred != NOCRED) { 1893 crfree(bp->b_rcred); 1894 bp->b_rcred = NOCRED; 1895 } 1896 if (bp->b_wcred != NOCRED) { 1897 crfree(bp->b_wcred); 1898 bp->b_wcred = NOCRED; 1899 } 1900 if (LIST_FIRST(&bp->b_dep) != NULL) 1901 buf_deallocate(bp); 1902 if (bp->b_vflags & BV_BKGRDINPROG) 1903 panic("losing buffer 3"); 1904 1905 if (bp->b_bufsize) 1906 allocbuf(bp, 0); 1907 1908 bp->b_flags = 0; 1909 bp->b_ioflags = 0; 1910 bp->b_xflags = 0; 1911 bp->b_vflags = 0; 1912 bp->b_dev = NODEV; 1913 bp->b_vp = NULL; 1914 bp->b_blkno = bp->b_lblkno = 0; 1915 bp->b_offset = NOOFFSET; 1916 bp->b_iodone = 0; 1917 bp->b_error = 0; 1918 bp->b_resid = 0; 1919 bp->b_bcount = 0; 1920 bp->b_npages = 0; 1921 bp->b_dirtyoff = bp->b_dirtyend = 0; 1922 bp->b_magic = B_MAGIC_BIO; 1923 bp->b_op = &buf_ops_bio; 1924 bp->b_object = NULL; 1925 1926 LIST_INIT(&bp->b_dep); 1927 1928 /* 1929 * If we are defragging then free the buffer. 1930 */ 1931 if (defrag) { 1932 bp->b_flags |= B_INVAL; 1933 bfreekva(bp); 1934 brelse(bp); 1935 defrag = 0; 1936 goto restart; 1937 } 1938 1939 /* 1940 * If we are overcomitted then recover the buffer and its 1941 * KVM space. This occurs in rare situations when multiple 1942 * processes are blocked in getnewbuf() or allocbuf(). 1943 */ 1944 if (bufspace >= hibufspace) 1945 flushingbufs = 1; 1946 if (flushingbufs && bp->b_kvasize != 0) { 1947 bp->b_flags |= B_INVAL; 1948 bfreekva(bp); 1949 brelse(bp); 1950 goto restart; 1951 } 1952 if (bufspace < lobufspace) 1953 flushingbufs = 0; 1954 break; 1955 } 1956 1957 /* 1958 * If we exhausted our list, sleep as appropriate. We may have to 1959 * wakeup various daemons and write out some dirty buffers. 1960 * 1961 * Generally we are sleeping due to insufficient buffer space. 1962 */ 1963 1964 if (bp == NULL) { 1965 int flags; 1966 char *waitmsg; 1967 1968 mtx_unlock(&bqlock); 1969 if (defrag) { 1970 flags = VFS_BIO_NEED_BUFSPACE; 1971 waitmsg = "nbufkv"; 1972 } else if (bufspace >= hibufspace) { 1973 waitmsg = "nbufbs"; 1974 flags = VFS_BIO_NEED_BUFSPACE; 1975 } else { 1976 waitmsg = "newbuf"; 1977 flags = VFS_BIO_NEED_ANY; 1978 } 1979 1980 bd_speedup(); /* heeeelp */ 1981 1982 mtx_lock(&nblock); 1983 needsbuffer |= flags; 1984 while (needsbuffer & flags) { 1985 if (msleep(&needsbuffer, &nblock, 1986 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1987 mtx_unlock(&nblock); 1988 return (NULL); 1989 } 1990 } 1991 mtx_unlock(&nblock); 1992 } else { 1993 /* 1994 * We finally have a valid bp. We aren't quite out of the 1995 * woods, we still have to reserve kva space. In order 1996 * to keep fragmentation sane we only allocate kva in 1997 * BKVASIZE chunks. 1998 */ 1999 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2000 2001 if (maxsize != bp->b_kvasize) { 2002 vm_offset_t addr = 0; 2003 2004 bfreekva(bp); 2005 2006 if (vm_map_findspace(buffer_map, 2007 vm_map_min(buffer_map), maxsize, &addr)) { 2008 /* 2009 * Uh oh. Buffer map is to fragmented. We 2010 * must defragment the map. 2011 */ 2012 atomic_add_int(&bufdefragcnt, 1); 2013 defrag = 1; 2014 bp->b_flags |= B_INVAL; 2015 brelse(bp); 2016 goto restart; 2017 } 2018 if (addr) { 2019 vm_map_insert(buffer_map, NULL, 0, 2020 addr, addr + maxsize, 2021 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2022 2023 bp->b_kvabase = (caddr_t) addr; 2024 bp->b_kvasize = maxsize; 2025 atomic_add_int(&bufspace, bp->b_kvasize); 2026 atomic_add_int(&bufreusecnt, 1); 2027 } 2028 } 2029 bp->b_saveaddr = bp->b_kvabase; 2030 bp->b_data = bp->b_saveaddr; 2031 } 2032 return(bp); 2033} 2034 2035/* 2036 * buf_daemon: 2037 * 2038 * buffer flushing daemon. Buffers are normally flushed by the 2039 * update daemon but if it cannot keep up this process starts to 2040 * take the load in an attempt to prevent getnewbuf() from blocking. 2041 */ 2042 2043static struct proc *bufdaemonproc; 2044 2045static struct kproc_desc buf_kp = { 2046 "bufdaemon", 2047 buf_daemon, 2048 &bufdaemonproc 2049}; 2050SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2051 2052static void 2053buf_daemon() 2054{ 2055 int s; 2056 2057 mtx_lock(&Giant); 2058 2059 /* 2060 * This process needs to be suspended prior to shutdown sync. 2061 */ 2062 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2063 SHUTDOWN_PRI_LAST); 2064 2065 /* 2066 * This process is allowed to take the buffer cache to the limit 2067 */ 2068 s = splbio(); 2069 mtx_lock(&bdlock); 2070 2071 for (;;) { 2072 bd_request = 0; 2073 mtx_unlock(&bdlock); 2074 2075 kthread_suspend_check(bufdaemonproc); 2076 2077 /* 2078 * Do the flush. Limit the amount of in-transit I/O we 2079 * allow to build up, otherwise we would completely saturate 2080 * the I/O system. Wakeup any waiting processes before we 2081 * normally would so they can run in parallel with our drain. 2082 */ 2083 while (numdirtybuffers > lodirtybuffers) { 2084 if (flushbufqueues(0) == 0) { 2085 /* 2086 * Could not find any buffers without rollback 2087 * dependencies, so just write the first one 2088 * in the hopes of eventually making progress. 2089 */ 2090 flushbufqueues(1); 2091 break; 2092 } 2093 waitrunningbufspace(); 2094 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2095 } 2096 2097 /* 2098 * Only clear bd_request if we have reached our low water 2099 * mark. The buf_daemon normally waits 1 second and 2100 * then incrementally flushes any dirty buffers that have 2101 * built up, within reason. 2102 * 2103 * If we were unable to hit our low water mark and couldn't 2104 * find any flushable buffers, we sleep half a second. 2105 * Otherwise we loop immediately. 2106 */ 2107 mtx_lock(&bdlock); 2108 if (numdirtybuffers <= lodirtybuffers) { 2109 /* 2110 * We reached our low water mark, reset the 2111 * request and sleep until we are needed again. 2112 * The sleep is just so the suspend code works. 2113 */ 2114 bd_request = 0; 2115 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2116 } else { 2117 /* 2118 * We couldn't find any flushable dirty buffers but 2119 * still have too many dirty buffers, we 2120 * have to sleep and try again. (rare) 2121 */ 2122 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2123 } 2124 } 2125} 2126 2127/* 2128 * flushbufqueues: 2129 * 2130 * Try to flush a buffer in the dirty queue. We must be careful to 2131 * free up B_INVAL buffers instead of write them, which NFS is 2132 * particularly sensitive to. 2133 */ 2134int flushwithdeps = 0; 2135SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2136 0, "Number of buffers flushed with dependecies that require rollbacks"); 2137static int 2138flushbufqueues(int flushdeps) 2139{ 2140 struct thread *td = curthread; 2141 struct vnode *vp; 2142 struct mount *mp; 2143 struct buf *bp; 2144 int hasdeps; 2145 2146 mtx_lock(&bqlock); 2147 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2148 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2149 continue; 2150 KASSERT((bp->b_flags & B_DELWRI), 2151 ("unexpected clean buffer %p", bp)); 2152 VI_LOCK(bp->b_vp); 2153 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 2154 VI_UNLOCK(bp->b_vp); 2155 BUF_UNLOCK(bp); 2156 continue; 2157 } 2158 VI_UNLOCK(bp->b_vp); 2159 if (bp->b_flags & B_INVAL) { 2160 bremfreel(bp); 2161 mtx_unlock(&bqlock); 2162 brelse(bp); 2163 return (1); 2164 } 2165 2166 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2167 if (flushdeps == 0) { 2168 BUF_UNLOCK(bp); 2169 continue; 2170 } 2171 hasdeps = 1; 2172 } else 2173 hasdeps = 0; 2174 /* 2175 * We must hold the lock on a vnode before writing 2176 * one of its buffers. Otherwise we may confuse, or 2177 * in the case of a snapshot vnode, deadlock the 2178 * system. 2179 * 2180 * The lock order here is the reverse of the normal 2181 * of vnode followed by buf lock. This is ok because 2182 * the NOWAIT will prevent deadlock. 2183 */ 2184 vp = bp->b_vp; 2185 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2186 BUF_UNLOCK(bp); 2187 continue; 2188 } 2189 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2190 mtx_unlock(&bqlock); 2191 vfs_bio_awrite(bp); 2192 vn_finished_write(mp); 2193 VOP_UNLOCK(vp, 0, td); 2194 flushwithdeps += hasdeps; 2195 return (1); 2196 } 2197 vn_finished_write(mp); 2198 BUF_UNLOCK(bp); 2199 } 2200 mtx_unlock(&bqlock); 2201 return (0); 2202} 2203 2204/* 2205 * Check to see if a block is currently memory resident. 2206 */ 2207struct buf * 2208incore(struct vnode * vp, daddr_t blkno) 2209{ 2210 struct buf *bp; 2211 2212 int s = splbio(); 2213 VI_LOCK(vp); 2214 bp = gbincore(vp, blkno); 2215 VI_UNLOCK(vp); 2216 splx(s); 2217 return (bp); 2218} 2219 2220/* 2221 * Returns true if no I/O is needed to access the 2222 * associated VM object. This is like incore except 2223 * it also hunts around in the VM system for the data. 2224 */ 2225 2226int 2227inmem(struct vnode * vp, daddr_t blkno) 2228{ 2229 vm_object_t obj; 2230 vm_offset_t toff, tinc, size; 2231 vm_page_t m; 2232 vm_ooffset_t off; 2233 2234 GIANT_REQUIRED; 2235 ASSERT_VOP_LOCKED(vp, "inmem"); 2236 2237 if (incore(vp, blkno)) 2238 return 1; 2239 if (vp->v_mount == NULL) 2240 return 0; 2241 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2242 return 0; 2243 2244 size = PAGE_SIZE; 2245 if (size > vp->v_mount->mnt_stat.f_iosize) 2246 size = vp->v_mount->mnt_stat.f_iosize; 2247 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2248 2249 VM_OBJECT_LOCK(obj); 2250 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2251 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2252 if (!m) 2253 goto notinmem; 2254 tinc = size; 2255 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2256 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2257 if (vm_page_is_valid(m, 2258 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2259 goto notinmem; 2260 } 2261 VM_OBJECT_UNLOCK(obj); 2262 return 1; 2263 2264notinmem: 2265 VM_OBJECT_UNLOCK(obj); 2266 return (0); 2267} 2268 2269/* 2270 * vfs_setdirty: 2271 * 2272 * Sets the dirty range for a buffer based on the status of the dirty 2273 * bits in the pages comprising the buffer. 2274 * 2275 * The range is limited to the size of the buffer. 2276 * 2277 * This routine is primarily used by NFS, but is generalized for the 2278 * B_VMIO case. 2279 */ 2280static void 2281vfs_setdirty(struct buf *bp) 2282{ 2283 int i; 2284 vm_object_t object; 2285 2286 GIANT_REQUIRED; 2287 /* 2288 * Degenerate case - empty buffer 2289 */ 2290 2291 if (bp->b_bufsize == 0) 2292 return; 2293 2294 /* 2295 * We qualify the scan for modified pages on whether the 2296 * object has been flushed yet. The OBJ_WRITEABLE flag 2297 * is not cleared simply by protecting pages off. 2298 */ 2299 2300 if ((bp->b_flags & B_VMIO) == 0) 2301 return; 2302 2303 object = bp->b_pages[0]->object; 2304 VM_OBJECT_LOCK(object); 2305 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2306 printf("Warning: object %p writeable but not mightbedirty\n", object); 2307 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2308 printf("Warning: object %p mightbedirty but not writeable\n", object); 2309 2310 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2311 vm_offset_t boffset; 2312 vm_offset_t eoffset; 2313 2314 vm_page_lock_queues(); 2315 /* 2316 * test the pages to see if they have been modified directly 2317 * by users through the VM system. 2318 */ 2319 for (i = 0; i < bp->b_npages; i++) { 2320 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2321 vm_page_test_dirty(bp->b_pages[i]); 2322 } 2323 2324 /* 2325 * Calculate the encompassing dirty range, boffset and eoffset, 2326 * (eoffset - boffset) bytes. 2327 */ 2328 2329 for (i = 0; i < bp->b_npages; i++) { 2330 if (bp->b_pages[i]->dirty) 2331 break; 2332 } 2333 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2334 2335 for (i = bp->b_npages - 1; i >= 0; --i) { 2336 if (bp->b_pages[i]->dirty) { 2337 break; 2338 } 2339 } 2340 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2341 2342 vm_page_unlock_queues(); 2343 /* 2344 * Fit it to the buffer. 2345 */ 2346 2347 if (eoffset > bp->b_bcount) 2348 eoffset = bp->b_bcount; 2349 2350 /* 2351 * If we have a good dirty range, merge with the existing 2352 * dirty range. 2353 */ 2354 2355 if (boffset < eoffset) { 2356 if (bp->b_dirtyoff > boffset) 2357 bp->b_dirtyoff = boffset; 2358 if (bp->b_dirtyend < eoffset) 2359 bp->b_dirtyend = eoffset; 2360 } 2361 } 2362 VM_OBJECT_UNLOCK(object); 2363} 2364 2365/* 2366 * getblk: 2367 * 2368 * Get a block given a specified block and offset into a file/device. 2369 * The buffers B_DONE bit will be cleared on return, making it almost 2370 * ready for an I/O initiation. B_INVAL may or may not be set on 2371 * return. The caller should clear B_INVAL prior to initiating a 2372 * READ. 2373 * 2374 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2375 * an existing buffer. 2376 * 2377 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2378 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2379 * and then cleared based on the backing VM. If the previous buffer is 2380 * non-0-sized but invalid, B_CACHE will be cleared. 2381 * 2382 * If getblk() must create a new buffer, the new buffer is returned with 2383 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2384 * case it is returned with B_INVAL clear and B_CACHE set based on the 2385 * backing VM. 2386 * 2387 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2388 * B_CACHE bit is clear. 2389 * 2390 * What this means, basically, is that the caller should use B_CACHE to 2391 * determine whether the buffer is fully valid or not and should clear 2392 * B_INVAL prior to issuing a read. If the caller intends to validate 2393 * the buffer by loading its data area with something, the caller needs 2394 * to clear B_INVAL. If the caller does this without issuing an I/O, 2395 * the caller should set B_CACHE ( as an optimization ), else the caller 2396 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2397 * a write attempt or if it was a successfull read. If the caller 2398 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2399 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2400 */ 2401struct buf * 2402getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2403 int flags) 2404{ 2405 struct buf *bp; 2406 int s; 2407 int error; 2408 ASSERT_VOP_LOCKED(vp, "getblk"); 2409 2410 if (size > MAXBSIZE) 2411 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2412 2413 s = splbio(); 2414loop: 2415 /* 2416 * Block if we are low on buffers. Certain processes are allowed 2417 * to completely exhaust the buffer cache. 2418 * 2419 * If this check ever becomes a bottleneck it may be better to 2420 * move it into the else, when gbincore() fails. At the moment 2421 * it isn't a problem. 2422 * 2423 * XXX remove if 0 sections (clean this up after its proven) 2424 */ 2425 if (numfreebuffers == 0) { 2426 if (curthread == PCPU_GET(idlethread)) 2427 return NULL; 2428 mtx_lock(&nblock); 2429 needsbuffer |= VFS_BIO_NEED_ANY; 2430 mtx_unlock(&nblock); 2431 } 2432 2433 VI_LOCK(vp); 2434 if ((bp = gbincore(vp, blkno))) { 2435 int lockflags; 2436 /* 2437 * Buffer is in-core. If the buffer is not busy, it must 2438 * be on a queue. 2439 */ 2440 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2441 2442 if (flags & GB_LOCK_NOWAIT) 2443 lockflags |= LK_NOWAIT; 2444 2445 error = BUF_TIMELOCK(bp, lockflags, 2446 VI_MTX(vp), "getblk", slpflag, slptimeo); 2447 2448 /* 2449 * If we slept and got the lock we have to restart in case 2450 * the buffer changed identities. 2451 */ 2452 if (error == ENOLCK) 2453 goto loop; 2454 /* We timed out or were interrupted. */ 2455 else if (error) 2456 return (NULL); 2457 2458 /* 2459 * The buffer is locked. B_CACHE is cleared if the buffer is 2460 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2461 * and for a VMIO buffer B_CACHE is adjusted according to the 2462 * backing VM cache. 2463 */ 2464 if (bp->b_flags & B_INVAL) 2465 bp->b_flags &= ~B_CACHE; 2466 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2467 bp->b_flags |= B_CACHE; 2468 bremfree(bp); 2469 2470 /* 2471 * check for size inconsistancies for non-VMIO case. 2472 */ 2473 2474 if (bp->b_bcount != size) { 2475 if ((bp->b_flags & B_VMIO) == 0 || 2476 (size > bp->b_kvasize)) { 2477 if (bp->b_flags & B_DELWRI) { 2478 bp->b_flags |= B_NOCACHE; 2479 BUF_WRITE(bp); 2480 } else { 2481 if ((bp->b_flags & B_VMIO) && 2482 (LIST_FIRST(&bp->b_dep) == NULL)) { 2483 bp->b_flags |= B_RELBUF; 2484 brelse(bp); 2485 } else { 2486 bp->b_flags |= B_NOCACHE; 2487 BUF_WRITE(bp); 2488 } 2489 } 2490 goto loop; 2491 } 2492 } 2493 2494 /* 2495 * If the size is inconsistant in the VMIO case, we can resize 2496 * the buffer. This might lead to B_CACHE getting set or 2497 * cleared. If the size has not changed, B_CACHE remains 2498 * unchanged from its previous state. 2499 */ 2500 2501 if (bp->b_bcount != size) 2502 allocbuf(bp, size); 2503 2504 KASSERT(bp->b_offset != NOOFFSET, 2505 ("getblk: no buffer offset")); 2506 2507 /* 2508 * A buffer with B_DELWRI set and B_CACHE clear must 2509 * be committed before we can return the buffer in 2510 * order to prevent the caller from issuing a read 2511 * ( due to B_CACHE not being set ) and overwriting 2512 * it. 2513 * 2514 * Most callers, including NFS and FFS, need this to 2515 * operate properly either because they assume they 2516 * can issue a read if B_CACHE is not set, or because 2517 * ( for example ) an uncached B_DELWRI might loop due 2518 * to softupdates re-dirtying the buffer. In the latter 2519 * case, B_CACHE is set after the first write completes, 2520 * preventing further loops. 2521 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2522 * above while extending the buffer, we cannot allow the 2523 * buffer to remain with B_CACHE set after the write 2524 * completes or it will represent a corrupt state. To 2525 * deal with this we set B_NOCACHE to scrap the buffer 2526 * after the write. 2527 * 2528 * We might be able to do something fancy, like setting 2529 * B_CACHE in bwrite() except if B_DELWRI is already set, 2530 * so the below call doesn't set B_CACHE, but that gets real 2531 * confusing. This is much easier. 2532 */ 2533 2534 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2535 bp->b_flags |= B_NOCACHE; 2536 BUF_WRITE(bp); 2537 goto loop; 2538 } 2539 2540 splx(s); 2541 bp->b_flags &= ~B_DONE; 2542 } else { 2543 int bsize, maxsize, vmio; 2544 off_t offset; 2545 2546 /* 2547 * Buffer is not in-core, create new buffer. The buffer 2548 * returned by getnewbuf() is locked. Note that the returned 2549 * buffer is also considered valid (not marked B_INVAL). 2550 */ 2551 VI_UNLOCK(vp); 2552 /* 2553 * If the user does not want us to create the buffer, bail out 2554 * here. 2555 */ 2556 if (flags & GB_NOCREAT) { 2557 splx(s); 2558 return NULL; 2559 } 2560 if (vn_isdisk(vp, NULL)) 2561 bsize = DEV_BSIZE; 2562 else if (vp->v_mountedhere) 2563 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2564 else if (vp->v_mount) 2565 bsize = vp->v_mount->mnt_stat.f_iosize; 2566 else 2567 bsize = size; 2568 2569 offset = blkno * bsize; 2570 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2571 (vp->v_vflag & VV_OBJBUF); 2572 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2573 maxsize = imax(maxsize, bsize); 2574 2575 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2576 if (slpflag || slptimeo) { 2577 splx(s); 2578 return NULL; 2579 } 2580 goto loop; 2581 } 2582 2583 /* 2584 * This code is used to make sure that a buffer is not 2585 * created while the getnewbuf routine is blocked. 2586 * This can be a problem whether the vnode is locked or not. 2587 * If the buffer is created out from under us, we have to 2588 * throw away the one we just created. There is now window 2589 * race because we are safely running at splbio() from the 2590 * point of the duplicate buffer creation through to here, 2591 * and we've locked the buffer. 2592 * 2593 * Note: this must occur before we associate the buffer 2594 * with the vp especially considering limitations in 2595 * the splay tree implementation when dealing with duplicate 2596 * lblkno's. 2597 */ 2598 VI_LOCK(vp); 2599 if (gbincore(vp, blkno)) { 2600 VI_UNLOCK(vp); 2601 bp->b_flags |= B_INVAL; 2602 brelse(bp); 2603 goto loop; 2604 } 2605 2606 /* 2607 * Insert the buffer into the hash, so that it can 2608 * be found by incore. 2609 */ 2610 bp->b_blkno = bp->b_lblkno = blkno; 2611 bp->b_offset = offset; 2612 2613 bgetvp(vp, bp); 2614 VI_UNLOCK(vp); 2615 2616 /* 2617 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2618 * buffer size starts out as 0, B_CACHE will be set by 2619 * allocbuf() for the VMIO case prior to it testing the 2620 * backing store for validity. 2621 */ 2622 2623 if (vmio) { 2624 bp->b_flags |= B_VMIO; 2625#if defined(VFS_BIO_DEBUG) 2626 if (vp->v_type != VREG) 2627 printf("getblk: vmioing file type %d???\n", vp->v_type); 2628#endif 2629 VOP_GETVOBJECT(vp, &bp->b_object); 2630 } else { 2631 bp->b_flags &= ~B_VMIO; 2632 bp->b_object = NULL; 2633 } 2634 2635 allocbuf(bp, size); 2636 2637 splx(s); 2638 bp->b_flags &= ~B_DONE; 2639 } 2640 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2641 return (bp); 2642} 2643 2644/* 2645 * Get an empty, disassociated buffer of given size. The buffer is initially 2646 * set to B_INVAL. 2647 */ 2648struct buf * 2649geteblk(int size) 2650{ 2651 struct buf *bp; 2652 int s; 2653 int maxsize; 2654 2655 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2656 2657 s = splbio(); 2658 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2659 continue; 2660 splx(s); 2661 allocbuf(bp, size); 2662 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2663 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2664 return (bp); 2665} 2666 2667 2668/* 2669 * This code constitutes the buffer memory from either anonymous system 2670 * memory (in the case of non-VMIO operations) or from an associated 2671 * VM object (in the case of VMIO operations). This code is able to 2672 * resize a buffer up or down. 2673 * 2674 * Note that this code is tricky, and has many complications to resolve 2675 * deadlock or inconsistant data situations. Tread lightly!!! 2676 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2677 * the caller. Calling this code willy nilly can result in the loss of data. 2678 * 2679 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2680 * B_CACHE for the non-VMIO case. 2681 */ 2682 2683int 2684allocbuf(struct buf *bp, int size) 2685{ 2686 int newbsize, mbsize; 2687 int i; 2688 2689 GIANT_REQUIRED; 2690 2691 if (BUF_REFCNT(bp) == 0) 2692 panic("allocbuf: buffer not busy"); 2693 2694 if (bp->b_kvasize < size) 2695 panic("allocbuf: buffer too small"); 2696 2697 if ((bp->b_flags & B_VMIO) == 0) { 2698 caddr_t origbuf; 2699 int origbufsize; 2700 /* 2701 * Just get anonymous memory from the kernel. Don't 2702 * mess with B_CACHE. 2703 */ 2704 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2705 if (bp->b_flags & B_MALLOC) 2706 newbsize = mbsize; 2707 else 2708 newbsize = round_page(size); 2709 2710 if (newbsize < bp->b_bufsize) { 2711 /* 2712 * malloced buffers are not shrunk 2713 */ 2714 if (bp->b_flags & B_MALLOC) { 2715 if (newbsize) { 2716 bp->b_bcount = size; 2717 } else { 2718 free(bp->b_data, M_BIOBUF); 2719 if (bp->b_bufsize) { 2720 atomic_subtract_int( 2721 &bufmallocspace, 2722 bp->b_bufsize); 2723 bufspacewakeup(); 2724 bp->b_bufsize = 0; 2725 } 2726 bp->b_saveaddr = bp->b_kvabase; 2727 bp->b_data = bp->b_saveaddr; 2728 bp->b_bcount = 0; 2729 bp->b_flags &= ~B_MALLOC; 2730 } 2731 return 1; 2732 } 2733 vm_hold_free_pages( 2734 bp, 2735 (vm_offset_t) bp->b_data + newbsize, 2736 (vm_offset_t) bp->b_data + bp->b_bufsize); 2737 } else if (newbsize > bp->b_bufsize) { 2738 /* 2739 * We only use malloced memory on the first allocation. 2740 * and revert to page-allocated memory when the buffer 2741 * grows. 2742 */ 2743 /* 2744 * There is a potential smp race here that could lead 2745 * to bufmallocspace slightly passing the max. It 2746 * is probably extremely rare and not worth worrying 2747 * over. 2748 */ 2749 if ( (bufmallocspace < maxbufmallocspace) && 2750 (bp->b_bufsize == 0) && 2751 (mbsize <= PAGE_SIZE/2)) { 2752 2753 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2754 bp->b_bufsize = mbsize; 2755 bp->b_bcount = size; 2756 bp->b_flags |= B_MALLOC; 2757 atomic_add_int(&bufmallocspace, mbsize); 2758 return 1; 2759 } 2760 origbuf = NULL; 2761 origbufsize = 0; 2762 /* 2763 * If the buffer is growing on its other-than-first allocation, 2764 * then we revert to the page-allocation scheme. 2765 */ 2766 if (bp->b_flags & B_MALLOC) { 2767 origbuf = bp->b_data; 2768 origbufsize = bp->b_bufsize; 2769 bp->b_data = bp->b_kvabase; 2770 if (bp->b_bufsize) { 2771 atomic_subtract_int(&bufmallocspace, 2772 bp->b_bufsize); 2773 bufspacewakeup(); 2774 bp->b_bufsize = 0; 2775 } 2776 bp->b_flags &= ~B_MALLOC; 2777 newbsize = round_page(newbsize); 2778 } 2779 vm_hold_load_pages( 2780 bp, 2781 (vm_offset_t) bp->b_data + bp->b_bufsize, 2782 (vm_offset_t) bp->b_data + newbsize); 2783 if (origbuf) { 2784 bcopy(origbuf, bp->b_data, origbufsize); 2785 free(origbuf, M_BIOBUF); 2786 } 2787 } 2788 } else { 2789 int desiredpages; 2790 2791 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2792 desiredpages = (size == 0) ? 0 : 2793 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2794 2795 if (bp->b_flags & B_MALLOC) 2796 panic("allocbuf: VMIO buffer can't be malloced"); 2797 /* 2798 * Set B_CACHE initially if buffer is 0 length or will become 2799 * 0-length. 2800 */ 2801 if (size == 0 || bp->b_bufsize == 0) 2802 bp->b_flags |= B_CACHE; 2803 2804 if (newbsize < bp->b_bufsize) { 2805 /* 2806 * DEV_BSIZE aligned new buffer size is less then the 2807 * DEV_BSIZE aligned existing buffer size. Figure out 2808 * if we have to remove any pages. 2809 */ 2810 if (desiredpages < bp->b_npages) { 2811 vm_page_t m; 2812 2813 vm_page_lock_queues(); 2814 for (i = desiredpages; i < bp->b_npages; i++) { 2815 /* 2816 * the page is not freed here -- it 2817 * is the responsibility of 2818 * vnode_pager_setsize 2819 */ 2820 m = bp->b_pages[i]; 2821 KASSERT(m != bogus_page, 2822 ("allocbuf: bogus page found")); 2823 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2824 vm_page_lock_queues(); 2825 2826 bp->b_pages[i] = NULL; 2827 vm_page_unwire(m, 0); 2828 } 2829 vm_page_unlock_queues(); 2830 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2831 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2832 bp->b_npages = desiredpages; 2833 } 2834 } else if (size > bp->b_bcount) { 2835 /* 2836 * We are growing the buffer, possibly in a 2837 * byte-granular fashion. 2838 */ 2839 struct vnode *vp; 2840 vm_object_t obj; 2841 vm_offset_t toff; 2842 vm_offset_t tinc; 2843 2844 /* 2845 * Step 1, bring in the VM pages from the object, 2846 * allocating them if necessary. We must clear 2847 * B_CACHE if these pages are not valid for the 2848 * range covered by the buffer. 2849 */ 2850 2851 vp = bp->b_vp; 2852 obj = bp->b_object; 2853 2854 VM_OBJECT_LOCK(obj); 2855 while (bp->b_npages < desiredpages) { 2856 vm_page_t m; 2857 vm_pindex_t pi; 2858 2859 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2860 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2861 /* 2862 * note: must allocate system pages 2863 * since blocking here could intefere 2864 * with paging I/O, no matter which 2865 * process we are. 2866 */ 2867 m = vm_page_alloc(obj, pi, 2868 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2869 if (m == NULL) { 2870 atomic_add_int(&vm_pageout_deficit, 2871 desiredpages - bp->b_npages); 2872 VM_OBJECT_UNLOCK(obj); 2873 VM_WAIT; 2874 VM_OBJECT_LOCK(obj); 2875 } else { 2876 vm_page_lock_queues(); 2877 vm_page_wakeup(m); 2878 vm_page_unlock_queues(); 2879 bp->b_flags &= ~B_CACHE; 2880 bp->b_pages[bp->b_npages] = m; 2881 ++bp->b_npages; 2882 } 2883 continue; 2884 } 2885 2886 /* 2887 * We found a page. If we have to sleep on it, 2888 * retry because it might have gotten freed out 2889 * from under us. 2890 * 2891 * We can only test PG_BUSY here. Blocking on 2892 * m->busy might lead to a deadlock: 2893 * 2894 * vm_fault->getpages->cluster_read->allocbuf 2895 * 2896 */ 2897 vm_page_lock_queues(); 2898 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2899 continue; 2900 2901 /* 2902 * We have a good page. Should we wakeup the 2903 * page daemon? 2904 */ 2905 if ((curproc != pageproc) && 2906 ((m->queue - m->pc) == PQ_CACHE) && 2907 ((cnt.v_free_count + cnt.v_cache_count) < 2908 (cnt.v_free_min + cnt.v_cache_min))) { 2909 pagedaemon_wakeup(); 2910 } 2911 vm_page_flag_clear(m, PG_ZERO); 2912 vm_page_wire(m); 2913 vm_page_unlock_queues(); 2914 bp->b_pages[bp->b_npages] = m; 2915 ++bp->b_npages; 2916 } 2917 2918 /* 2919 * Step 2. We've loaded the pages into the buffer, 2920 * we have to figure out if we can still have B_CACHE 2921 * set. Note that B_CACHE is set according to the 2922 * byte-granular range ( bcount and size ), new the 2923 * aligned range ( newbsize ). 2924 * 2925 * The VM test is against m->valid, which is DEV_BSIZE 2926 * aligned. Needless to say, the validity of the data 2927 * needs to also be DEV_BSIZE aligned. Note that this 2928 * fails with NFS if the server or some other client 2929 * extends the file's EOF. If our buffer is resized, 2930 * B_CACHE may remain set! XXX 2931 */ 2932 2933 toff = bp->b_bcount; 2934 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2935 2936 while ((bp->b_flags & B_CACHE) && toff < size) { 2937 vm_pindex_t pi; 2938 2939 if (tinc > (size - toff)) 2940 tinc = size - toff; 2941 2942 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2943 PAGE_SHIFT; 2944 2945 vfs_buf_test_cache( 2946 bp, 2947 bp->b_offset, 2948 toff, 2949 tinc, 2950 bp->b_pages[pi] 2951 ); 2952 toff += tinc; 2953 tinc = PAGE_SIZE; 2954 } 2955 VM_OBJECT_UNLOCK(obj); 2956 2957 /* 2958 * Step 3, fixup the KVM pmap. Remember that 2959 * bp->b_data is relative to bp->b_offset, but 2960 * bp->b_offset may be offset into the first page. 2961 */ 2962 2963 bp->b_data = (caddr_t) 2964 trunc_page((vm_offset_t)bp->b_data); 2965 pmap_qenter( 2966 (vm_offset_t)bp->b_data, 2967 bp->b_pages, 2968 bp->b_npages 2969 ); 2970 2971 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2972 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2973 } 2974 } 2975 if (newbsize < bp->b_bufsize) 2976 bufspacewakeup(); 2977 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2978 bp->b_bcount = size; /* requested buffer size */ 2979 return 1; 2980} 2981 2982void 2983biodone(struct bio *bp) 2984{ 2985 mtx_lock(&bdonelock); 2986 bp->bio_flags |= BIO_DONE; 2987 if (bp->bio_done == NULL) 2988 wakeup(bp); 2989 mtx_unlock(&bdonelock); 2990 if (bp->bio_done != NULL) 2991 bp->bio_done(bp); 2992} 2993 2994/* 2995 * Wait for a BIO to finish. 2996 * 2997 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2998 * case is not yet clear. 2999 */ 3000int 3001biowait(struct bio *bp, const char *wchan) 3002{ 3003 3004 mtx_lock(&bdonelock); 3005 while ((bp->bio_flags & BIO_DONE) == 0) 3006 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3007 mtx_unlock(&bdonelock); 3008 if (bp->bio_error != 0) 3009 return (bp->bio_error); 3010 if (!(bp->bio_flags & BIO_ERROR)) 3011 return (0); 3012 return (EIO); 3013} 3014 3015void 3016biofinish(struct bio *bp, struct devstat *stat, int error) 3017{ 3018 3019 if (error) { 3020 bp->bio_error = error; 3021 bp->bio_flags |= BIO_ERROR; 3022 } 3023 if (stat != NULL) 3024 devstat_end_transaction_bio(stat, bp); 3025 biodone(bp); 3026} 3027 3028/* 3029 * bufwait: 3030 * 3031 * Wait for buffer I/O completion, returning error status. The buffer 3032 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3033 * error and cleared. 3034 */ 3035int 3036bufwait(register struct buf * bp) 3037{ 3038 int s; 3039 3040 s = splbio(); 3041 if (bp->b_iocmd == BIO_READ) 3042 bwait(bp, PRIBIO, "biord"); 3043 else 3044 bwait(bp, PRIBIO, "biowr"); 3045 splx(s); 3046 if (bp->b_flags & B_EINTR) { 3047 bp->b_flags &= ~B_EINTR; 3048 return (EINTR); 3049 } 3050 if (bp->b_ioflags & BIO_ERROR) { 3051 return (bp->b_error ? bp->b_error : EIO); 3052 } else { 3053 return (0); 3054 } 3055} 3056 3057 /* 3058 * Call back function from struct bio back up to struct buf. 3059 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 3060 */ 3061static void 3062bufdonebio(struct bio *bp) 3063{ 3064 3065 /* Device drivers may or may not hold giant, hold it here. */ 3066 mtx_lock(&Giant); 3067 bufdone(bp->bio_caller2); 3068 mtx_unlock(&Giant); 3069} 3070 3071void 3072dev_strategy(struct buf *bp) 3073{ 3074 3075 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3076 panic("b_iocmd botch"); 3077 if (bp->b_flags & B_PHYS) 3078 bp->b_io.bio_offset = bp->b_offset; 3079 else 3080 bp->b_io.bio_offset = dbtob(bp->b_blkno); 3081 bp->b_io.bio_done = bufdonebio; 3082 bp->b_io.bio_caller2 = bp; 3083 (*devsw(bp->b_io.bio_dev)->d_strategy)(&bp->b_io); 3084} 3085 3086/* 3087 * bufdone: 3088 * 3089 * Finish I/O on a buffer, optionally calling a completion function. 3090 * This is usually called from an interrupt so process blocking is 3091 * not allowed. 3092 * 3093 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3094 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3095 * assuming B_INVAL is clear. 3096 * 3097 * For the VMIO case, we set B_CACHE if the op was a read and no 3098 * read error occured, or if the op was a write. B_CACHE is never 3099 * set if the buffer is invalid or otherwise uncacheable. 3100 * 3101 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3102 * initiator to leave B_INVAL set to brelse the buffer out of existance 3103 * in the biodone routine. 3104 */ 3105void 3106bufdone(struct buf *bp) 3107{ 3108 int s; 3109 void (*biodone)(struct buf *); 3110 3111 GIANT_REQUIRED; 3112 3113 s = splbio(); 3114 3115 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3116 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3117 3118 bp->b_flags |= B_DONE; 3119 runningbufwakeup(bp); 3120 3121 if (bp->b_iocmd == BIO_DELETE) { 3122 brelse(bp); 3123 splx(s); 3124 return; 3125 } 3126 3127 if (bp->b_iocmd == BIO_WRITE) { 3128 vwakeup(bp); 3129 } 3130 3131 /* call optional completion function if requested */ 3132 if (bp->b_iodone != NULL) { 3133 biodone = bp->b_iodone; 3134 bp->b_iodone = NULL; 3135 (*biodone) (bp); 3136 splx(s); 3137 return; 3138 } 3139 if (LIST_FIRST(&bp->b_dep) != NULL) 3140 buf_complete(bp); 3141 3142 if (bp->b_flags & B_VMIO) { 3143 int i; 3144 vm_ooffset_t foff; 3145 vm_page_t m; 3146 vm_object_t obj; 3147 int iosize; 3148 struct vnode *vp = bp->b_vp; 3149 3150 obj = bp->b_object; 3151 3152#if defined(VFS_BIO_DEBUG) 3153 mp_fixme("usecount and vflag accessed without locks."); 3154 if (vp->v_usecount == 0) { 3155 panic("biodone: zero vnode ref count"); 3156 } 3157 3158 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3159 panic("biodone: vnode is not setup for merged cache"); 3160 } 3161#endif 3162 3163 foff = bp->b_offset; 3164 KASSERT(bp->b_offset != NOOFFSET, 3165 ("biodone: no buffer offset")); 3166 3167 if (obj != NULL) 3168 VM_OBJECT_LOCK(obj); 3169#if defined(VFS_BIO_DEBUG) 3170 if (obj->paging_in_progress < bp->b_npages) { 3171 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3172 obj->paging_in_progress, bp->b_npages); 3173 } 3174#endif 3175 3176 /* 3177 * Set B_CACHE if the op was a normal read and no error 3178 * occured. B_CACHE is set for writes in the b*write() 3179 * routines. 3180 */ 3181 iosize = bp->b_bcount - bp->b_resid; 3182 if (bp->b_iocmd == BIO_READ && 3183 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3184 !(bp->b_ioflags & BIO_ERROR)) { 3185 bp->b_flags |= B_CACHE; 3186 } 3187 vm_page_lock_queues(); 3188 for (i = 0; i < bp->b_npages; i++) { 3189 int bogusflag = 0; 3190 int resid; 3191 3192 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3193 if (resid > iosize) 3194 resid = iosize; 3195 3196 /* 3197 * cleanup bogus pages, restoring the originals 3198 */ 3199 m = bp->b_pages[i]; 3200 if (m == bogus_page) { 3201 bogusflag = 1; 3202 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3203 if (m == NULL) 3204 panic("biodone: page disappeared!"); 3205 bp->b_pages[i] = m; 3206 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3207 } 3208#if defined(VFS_BIO_DEBUG) 3209 if (OFF_TO_IDX(foff) != m->pindex) { 3210 printf( 3211"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3212 (intmax_t)foff, (uintmax_t)m->pindex); 3213 } 3214#endif 3215 3216 /* 3217 * In the write case, the valid and clean bits are 3218 * already changed correctly ( see bdwrite() ), so we 3219 * only need to do this here in the read case. 3220 */ 3221 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3222 vfs_page_set_valid(bp, foff, i, m); 3223 } 3224 vm_page_flag_clear(m, PG_ZERO); 3225 3226 /* 3227 * when debugging new filesystems or buffer I/O methods, this 3228 * is the most common error that pops up. if you see this, you 3229 * have not set the page busy flag correctly!!! 3230 */ 3231 if (m->busy == 0) { 3232 printf("biodone: page busy < 0, " 3233 "pindex: %d, foff: 0x(%x,%x), " 3234 "resid: %d, index: %d\n", 3235 (int) m->pindex, (int)(foff >> 32), 3236 (int) foff & 0xffffffff, resid, i); 3237 if (!vn_isdisk(vp, NULL)) 3238 printf(" iosize: %ld, lblkno: %jd, flags: 0x%x, npages: %d\n", 3239 bp->b_vp->v_mount->mnt_stat.f_iosize, 3240 (intmax_t) bp->b_lblkno, 3241 bp->b_flags, bp->b_npages); 3242 else 3243 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3244 (intmax_t) bp->b_lblkno, 3245 bp->b_flags, bp->b_npages); 3246 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3247 (u_long)m->valid, (u_long)m->dirty, 3248 m->wire_count); 3249 panic("biodone: page busy < 0\n"); 3250 } 3251 vm_page_io_finish(m); 3252 vm_object_pip_subtract(obj, 1); 3253 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3254 iosize -= resid; 3255 } 3256 vm_page_unlock_queues(); 3257 if (obj != NULL) { 3258 vm_object_pip_wakeupn(obj, 0); 3259 VM_OBJECT_UNLOCK(obj); 3260 } 3261 } 3262 3263 /* 3264 * For asynchronous completions, release the buffer now. The brelse 3265 * will do a wakeup there if necessary - so no need to do a wakeup 3266 * here in the async case. The sync case always needs to do a wakeup. 3267 */ 3268 3269 if (bp->b_flags & B_ASYNC) { 3270 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3271 brelse(bp); 3272 else 3273 bqrelse(bp); 3274 } else { 3275 bdone(bp); 3276 } 3277 splx(s); 3278} 3279 3280/* 3281 * This routine is called in lieu of iodone in the case of 3282 * incomplete I/O. This keeps the busy status for pages 3283 * consistant. 3284 */ 3285void 3286vfs_unbusy_pages(struct buf * bp) 3287{ 3288 int i; 3289 3290 GIANT_REQUIRED; 3291 3292 runningbufwakeup(bp); 3293 if (bp->b_flags & B_VMIO) { 3294 vm_object_t obj; 3295 3296 obj = bp->b_object; 3297 VM_OBJECT_LOCK(obj); 3298 vm_page_lock_queues(); 3299 for (i = 0; i < bp->b_npages; i++) { 3300 vm_page_t m = bp->b_pages[i]; 3301 3302 if (m == bogus_page) { 3303 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3304 if (!m) { 3305 panic("vfs_unbusy_pages: page missing\n"); 3306 } 3307 bp->b_pages[i] = m; 3308 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3309 } 3310 vm_object_pip_subtract(obj, 1); 3311 vm_page_flag_clear(m, PG_ZERO); 3312 vm_page_io_finish(m); 3313 } 3314 vm_page_unlock_queues(); 3315 vm_object_pip_wakeupn(obj, 0); 3316 VM_OBJECT_UNLOCK(obj); 3317 } 3318} 3319 3320/* 3321 * vfs_page_set_valid: 3322 * 3323 * Set the valid bits in a page based on the supplied offset. The 3324 * range is restricted to the buffer's size. 3325 * 3326 * This routine is typically called after a read completes. 3327 */ 3328static void 3329vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3330{ 3331 vm_ooffset_t soff, eoff; 3332 3333 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3334 /* 3335 * Start and end offsets in buffer. eoff - soff may not cross a 3336 * page boundry or cross the end of the buffer. The end of the 3337 * buffer, in this case, is our file EOF, not the allocation size 3338 * of the buffer. 3339 */ 3340 soff = off; 3341 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3342 if (eoff > bp->b_offset + bp->b_bcount) 3343 eoff = bp->b_offset + bp->b_bcount; 3344 3345 /* 3346 * Set valid range. This is typically the entire buffer and thus the 3347 * entire page. 3348 */ 3349 if (eoff > soff) { 3350 vm_page_set_validclean( 3351 m, 3352 (vm_offset_t) (soff & PAGE_MASK), 3353 (vm_offset_t) (eoff - soff) 3354 ); 3355 } 3356} 3357 3358/* 3359 * This routine is called before a device strategy routine. 3360 * It is used to tell the VM system that paging I/O is in 3361 * progress, and treat the pages associated with the buffer 3362 * almost as being PG_BUSY. Also the object paging_in_progress 3363 * flag is handled to make sure that the object doesn't become 3364 * inconsistant. 3365 * 3366 * Since I/O has not been initiated yet, certain buffer flags 3367 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3368 * and should be ignored. 3369 */ 3370void 3371vfs_busy_pages(struct buf * bp, int clear_modify) 3372{ 3373 int i, bogus; 3374 3375 if (bp->b_flags & B_VMIO) { 3376 vm_object_t obj; 3377 vm_ooffset_t foff; 3378 3379 obj = bp->b_object; 3380 foff = bp->b_offset; 3381 KASSERT(bp->b_offset != NOOFFSET, 3382 ("vfs_busy_pages: no buffer offset")); 3383 vfs_setdirty(bp); 3384 if (obj != NULL) 3385 VM_OBJECT_LOCK(obj); 3386retry: 3387 vm_page_lock_queues(); 3388 for (i = 0; i < bp->b_npages; i++) { 3389 vm_page_t m = bp->b_pages[i]; 3390 3391 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3392 goto retry; 3393 } 3394 bogus = 0; 3395 for (i = 0; i < bp->b_npages; i++) { 3396 vm_page_t m = bp->b_pages[i]; 3397 3398 vm_page_flag_clear(m, PG_ZERO); 3399 if ((bp->b_flags & B_CLUSTER) == 0) { 3400 vm_object_pip_add(obj, 1); 3401 vm_page_io_start(m); 3402 } 3403 /* 3404 * When readying a buffer for a read ( i.e 3405 * clear_modify == 0 ), it is important to do 3406 * bogus_page replacement for valid pages in 3407 * partially instantiated buffers. Partially 3408 * instantiated buffers can, in turn, occur when 3409 * reconstituting a buffer from its VM backing store 3410 * base. We only have to do this if B_CACHE is 3411 * clear ( which causes the I/O to occur in the 3412 * first place ). The replacement prevents the read 3413 * I/O from overwriting potentially dirty VM-backed 3414 * pages. XXX bogus page replacement is, uh, bogus. 3415 * It may not work properly with small-block devices. 3416 * We need to find a better way. 3417 */ 3418 pmap_remove_all(m); 3419 if (clear_modify) 3420 vfs_page_set_valid(bp, foff, i, m); 3421 else if (m->valid == VM_PAGE_BITS_ALL && 3422 (bp->b_flags & B_CACHE) == 0) { 3423 bp->b_pages[i] = bogus_page; 3424 bogus++; 3425 } 3426 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3427 } 3428 vm_page_unlock_queues(); 3429 if (obj != NULL) 3430 VM_OBJECT_UNLOCK(obj); 3431 if (bogus) 3432 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3433 } 3434} 3435 3436/* 3437 * Tell the VM system that the pages associated with this buffer 3438 * are clean. This is used for delayed writes where the data is 3439 * going to go to disk eventually without additional VM intevention. 3440 * 3441 * Note that while we only really need to clean through to b_bcount, we 3442 * just go ahead and clean through to b_bufsize. 3443 */ 3444static void 3445vfs_clean_pages(struct buf * bp) 3446{ 3447 int i; 3448 3449 if (bp->b_flags & B_VMIO) { 3450 vm_ooffset_t foff; 3451 3452 foff = bp->b_offset; 3453 KASSERT(bp->b_offset != NOOFFSET, 3454 ("vfs_clean_pages: no buffer offset")); 3455 VM_OBJECT_LOCK(bp->b_object); 3456 vm_page_lock_queues(); 3457 for (i = 0; i < bp->b_npages; i++) { 3458 vm_page_t m = bp->b_pages[i]; 3459 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3460 vm_ooffset_t eoff = noff; 3461 3462 if (eoff > bp->b_offset + bp->b_bufsize) 3463 eoff = bp->b_offset + bp->b_bufsize; 3464 vfs_page_set_valid(bp, foff, i, m); 3465 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3466 foff = noff; 3467 } 3468 vm_page_unlock_queues(); 3469 VM_OBJECT_UNLOCK(bp->b_object); 3470 } 3471} 3472 3473/* 3474 * vfs_bio_set_validclean: 3475 * 3476 * Set the range within the buffer to valid and clean. The range is 3477 * relative to the beginning of the buffer, b_offset. Note that b_offset 3478 * itself may be offset from the beginning of the first page. 3479 * 3480 */ 3481 3482void 3483vfs_bio_set_validclean(struct buf *bp, int base, int size) 3484{ 3485 if (bp->b_flags & B_VMIO) { 3486 int i; 3487 int n; 3488 3489 /* 3490 * Fixup base to be relative to beginning of first page. 3491 * Set initial n to be the maximum number of bytes in the 3492 * first page that can be validated. 3493 */ 3494 3495 base += (bp->b_offset & PAGE_MASK); 3496 n = PAGE_SIZE - (base & PAGE_MASK); 3497 3498 VM_OBJECT_LOCK(bp->b_object); 3499 vm_page_lock_queues(); 3500 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3501 vm_page_t m = bp->b_pages[i]; 3502 3503 if (n > size) 3504 n = size; 3505 3506 vm_page_set_validclean(m, base & PAGE_MASK, n); 3507 base += n; 3508 size -= n; 3509 n = PAGE_SIZE; 3510 } 3511 vm_page_unlock_queues(); 3512 VM_OBJECT_UNLOCK(bp->b_object); 3513 } 3514} 3515 3516/* 3517 * vfs_bio_clrbuf: 3518 * 3519 * clear a buffer. This routine essentially fakes an I/O, so we need 3520 * to clear BIO_ERROR and B_INVAL. 3521 * 3522 * Note that while we only theoretically need to clear through b_bcount, 3523 * we go ahead and clear through b_bufsize. 3524 */ 3525 3526void 3527vfs_bio_clrbuf(struct buf *bp) 3528{ 3529 int i, mask = 0; 3530 caddr_t sa, ea; 3531 3532 GIANT_REQUIRED; 3533 3534 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3535 bp->b_flags &= ~B_INVAL; 3536 bp->b_ioflags &= ~BIO_ERROR; 3537 if (bp->b_object != NULL) 3538 VM_OBJECT_LOCK(bp->b_object); 3539 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3540 (bp->b_offset & PAGE_MASK) == 0) { 3541 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3542 if (bp->b_pages[0] != bogus_page) 3543 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3544 if ((bp->b_pages[0]->valid & mask) == mask) 3545 goto unlock; 3546 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3547 ((bp->b_pages[0]->valid & mask) == 0)) { 3548 bzero(bp->b_data, bp->b_bufsize); 3549 bp->b_pages[0]->valid |= mask; 3550 goto unlock; 3551 } 3552 } 3553 ea = sa = bp->b_data; 3554 for(i=0;i<bp->b_npages;i++,sa=ea) { 3555 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3556 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3557 ea = (caddr_t)(vm_offset_t)ulmin( 3558 (u_long)(vm_offset_t)ea, 3559 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3560 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3561 if (bp->b_pages[i] != bogus_page) 3562 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3563 if ((bp->b_pages[i]->valid & mask) == mask) 3564 continue; 3565 if ((bp->b_pages[i]->valid & mask) == 0) { 3566 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3567 bzero(sa, ea - sa); 3568 } 3569 } else { 3570 for (; sa < ea; sa += DEV_BSIZE, j++) { 3571 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3572 (bp->b_pages[i]->valid & (1<<j)) == 0) 3573 bzero(sa, DEV_BSIZE); 3574 } 3575 } 3576 bp->b_pages[i]->valid |= mask; 3577 vm_page_lock_queues(); 3578 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3579 vm_page_unlock_queues(); 3580 } 3581unlock: 3582 if (bp->b_object != NULL) 3583 VM_OBJECT_UNLOCK(bp->b_object); 3584 bp->b_resid = 0; 3585 } else { 3586 clrbuf(bp); 3587 } 3588} 3589 3590/* 3591 * vm_hold_load_pages and vm_hold_free_pages get pages into 3592 * a buffers address space. The pages are anonymous and are 3593 * not associated with a file object. 3594 */ 3595static void 3596vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3597{ 3598 vm_offset_t pg; 3599 vm_page_t p; 3600 int index; 3601 3602 to = round_page(to); 3603 from = round_page(from); 3604 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3605 3606 VM_OBJECT_LOCK(kernel_object); 3607 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3608tryagain: 3609 /* 3610 * note: must allocate system pages since blocking here 3611 * could intefere with paging I/O, no matter which 3612 * process we are. 3613 */ 3614 p = vm_page_alloc(kernel_object, 3615 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3616 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3617 if (!p) { 3618 atomic_add_int(&vm_pageout_deficit, 3619 (to - pg) >> PAGE_SHIFT); 3620 VM_OBJECT_UNLOCK(kernel_object); 3621 VM_WAIT; 3622 VM_OBJECT_LOCK(kernel_object); 3623 goto tryagain; 3624 } 3625 p->valid = VM_PAGE_BITS_ALL; 3626 pmap_qenter(pg, &p, 1); 3627 bp->b_pages[index] = p; 3628 vm_page_lock_queues(); 3629 vm_page_wakeup(p); 3630 vm_page_unlock_queues(); 3631 } 3632 VM_OBJECT_UNLOCK(kernel_object); 3633 bp->b_npages = index; 3634} 3635 3636/* Return pages associated with this buf to the vm system */ 3637static void 3638vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3639{ 3640 vm_offset_t pg; 3641 vm_page_t p; 3642 int index, newnpages; 3643 3644 GIANT_REQUIRED; 3645 3646 from = round_page(from); 3647 to = round_page(to); 3648 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3649 3650 VM_OBJECT_LOCK(kernel_object); 3651 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3652 p = bp->b_pages[index]; 3653 if (p && (index < bp->b_npages)) { 3654 if (p->busy) { 3655 printf( 3656 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3657 (intmax_t)bp->b_blkno, 3658 (intmax_t)bp->b_lblkno); 3659 } 3660 bp->b_pages[index] = NULL; 3661 pmap_qremove(pg, 1); 3662 vm_page_lock_queues(); 3663 vm_page_busy(p); 3664 vm_page_unwire(p, 0); 3665 vm_page_free(p); 3666 vm_page_unlock_queues(); 3667 } 3668 } 3669 VM_OBJECT_UNLOCK(kernel_object); 3670 bp->b_npages = newnpages; 3671} 3672 3673/* 3674 * Map an IO request into kernel virtual address space. 3675 * 3676 * All requests are (re)mapped into kernel VA space. 3677 * Notice that we use b_bufsize for the size of the buffer 3678 * to be mapped. b_bcount might be modified by the driver. 3679 * 3680 * Note that even if the caller determines that the address space should 3681 * be valid, a race or a smaller-file mapped into a larger space may 3682 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3683 * check the return value. 3684 */ 3685int 3686vmapbuf(struct buf *bp) 3687{ 3688 caddr_t addr, kva; 3689 vm_prot_t prot; 3690 int pidx, i; 3691 struct vm_page *m; 3692 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3693 3694 GIANT_REQUIRED; 3695 3696 if (bp->b_bufsize < 0) 3697 return (-1); 3698 prot = (bp->b_iocmd == BIO_READ) ? VM_PROT_READ | VM_PROT_WRITE : 3699 VM_PROT_READ; 3700 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3701 addr < bp->b_data + bp->b_bufsize; 3702 addr += PAGE_SIZE, pidx++) { 3703 /* 3704 * Do the vm_fault if needed; do the copy-on-write thing 3705 * when reading stuff off device into memory. 3706 * 3707 * NOTE! Must use pmap_extract() because addr may be in 3708 * the userland address space, and kextract is only guarenteed 3709 * to work for the kernland address space (see: sparc64 port). 3710 */ 3711retry: 3712 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3713 prot) < 0) { 3714 vm_page_lock_queues(); 3715 for (i = 0; i < pidx; ++i) { 3716 vm_page_unhold(bp->b_pages[i]); 3717 bp->b_pages[i] = NULL; 3718 } 3719 vm_page_unlock_queues(); 3720 return(-1); 3721 } 3722 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3723 if (m == NULL) 3724 goto retry; 3725 bp->b_pages[pidx] = m; 3726 } 3727 if (pidx > btoc(MAXPHYS)) 3728 panic("vmapbuf: mapped more than MAXPHYS"); 3729 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3730 3731 kva = bp->b_saveaddr; 3732 bp->b_npages = pidx; 3733 bp->b_saveaddr = bp->b_data; 3734 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3735 return(0); 3736} 3737 3738/* 3739 * Free the io map PTEs associated with this IO operation. 3740 * We also invalidate the TLB entries and restore the original b_addr. 3741 */ 3742void 3743vunmapbuf(struct buf *bp) 3744{ 3745 int pidx; 3746 int npages; 3747 3748 GIANT_REQUIRED; 3749 3750 npages = bp->b_npages; 3751 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3752 npages); 3753 vm_page_lock_queues(); 3754 for (pidx = 0; pidx < npages; pidx++) 3755 vm_page_unhold(bp->b_pages[pidx]); 3756 vm_page_unlock_queues(); 3757 3758 bp->b_data = bp->b_saveaddr; 3759} 3760 3761void 3762bdone(struct buf *bp) 3763{ 3764 mtx_lock(&bdonelock); 3765 bp->b_flags |= B_DONE; 3766 wakeup(bp); 3767 mtx_unlock(&bdonelock); 3768} 3769 3770void 3771bwait(struct buf *bp, u_char pri, const char *wchan) 3772{ 3773 mtx_lock(&bdonelock); 3774 while ((bp->b_flags & B_DONE) == 0) 3775 msleep(bp, &bdonelock, pri, wchan, 0); 3776 mtx_unlock(&bdonelock); 3777} 3778 3779#include "opt_ddb.h" 3780#ifdef DDB 3781#include <ddb/ddb.h> 3782 3783/* DDB command to show buffer data */ 3784DB_SHOW_COMMAND(buffer, db_show_buffer) 3785{ 3786 /* get args */ 3787 struct buf *bp = (struct buf *)addr; 3788 3789 if (!have_addr) { 3790 db_printf("usage: show buffer <addr>\n"); 3791 return; 3792 } 3793 3794 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3795 db_printf( 3796 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3797 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd\n", 3798 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3799 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3800 (intmax_t)bp->b_blkno); 3801 if (bp->b_npages) { 3802 int i; 3803 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3804 for (i = 0; i < bp->b_npages; i++) { 3805 vm_page_t m; 3806 m = bp->b_pages[i]; 3807 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3808 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3809 if ((i + 1) < bp->b_npages) 3810 db_printf(","); 3811 } 3812 db_printf("\n"); 3813 } 3814} 3815#endif /* DDB */ 3816