vfs_bio.c revision 126704
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 126704 2004-03-07 00:37:18Z alc $"); 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 struct proc *bufdaemonproc; 76 77static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 78 vm_offset_t to); 79static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 80 vm_offset_t to); 81static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 82 int pageno, vm_page_t m); 83static void vfs_clean_pages(struct buf * bp); 84static void vfs_setdirty(struct buf *bp); 85static void vfs_vmio_release(struct buf *bp); 86static void vfs_backgroundwritedone(struct buf *bp); 87static int vfs_bio_clcheck(struct vnode *vp, int size, 88 daddr_t lblkno, daddr_t blkno); 89static int flushbufqueues(int flushdeps); 90static void buf_daemon(void); 91void bremfreel(struct buf * bp); 92 93int vmiodirenable = TRUE; 94SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 95 "Use the VM system for directory writes"); 96int runningbufspace; 97SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 98 "Amount of presently outstanding async buffer io"); 99static int bufspace; 100SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 101 "KVA memory used for bufs"); 102static int maxbufspace; 103SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 104 "Maximum allowed value of bufspace (including buf_daemon)"); 105static int bufmallocspace; 106SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 107 "Amount of malloced memory for buffers"); 108static int maxbufmallocspace; 109SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 110 "Maximum amount of malloced memory for buffers"); 111static int lobufspace; 112SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 113 "Minimum amount of buffers we want to have"); 114static int hibufspace; 115SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 116 "Maximum allowed value of bufspace (excluding buf_daemon)"); 117static int bufreusecnt; 118SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 119 "Number of times we have reused a buffer"); 120static int buffreekvacnt; 121SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 122 "Number of times we have freed the KVA space from some buffer"); 123static int bufdefragcnt; 124SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 125 "Number of times we have had to repeat buffer allocation to defragment"); 126static int lorunningspace; 127SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 128 "Minimum preferred space used for in-progress I/O"); 129static int hirunningspace; 130SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 131 "Maximum amount of space to use for in-progress I/O"); 132static int dirtybufferflushes; 133SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 134 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 135static int altbufferflushes; 136SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 137 0, "Number of fsync flushes to limit dirty buffers"); 138static int recursiveflushes; 139SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 140 0, "Number of flushes skipped due to being recursive"); 141static int numdirtybuffers; 142SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 143 "Number of buffers that are dirty (has unwritten changes) at the moment"); 144static int lodirtybuffers; 145SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 146 "How many buffers we want to have free before bufdaemon can sleep"); 147static int hidirtybuffers; 148SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 149 "When the number of dirty buffers is considered severe"); 150static int dirtybufthresh; 151SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 152 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 153static int numfreebuffers; 154SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 155 "Number of free buffers"); 156static int lofreebuffers; 157SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 158 "XXX Unused"); 159static int hifreebuffers; 160SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 161 "XXX Complicatedly unused"); 162static int getnewbufcalls; 163SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 164 "Number of calls to getnewbuf"); 165static int getnewbufrestarts; 166SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 167 "Number of times getnewbuf has had to restart a buffer aquisition"); 168static int dobkgrdwrite = 1; 169SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 170 "Do background writes (honoring the BV_BKGRDWRITE flag)?"); 171 172/* 173 * Wakeup point for bufdaemon, as well as indicator of whether it is already 174 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 175 * is idling. 176 */ 177static int bd_request; 178 179/* 180 * This lock synchronizes access to bd_request. 181 */ 182static struct mtx bdlock; 183 184/* 185 * bogus page -- for I/O to/from partially complete buffers 186 * this is a temporary solution to the problem, but it is not 187 * really that bad. it would be better to split the buffer 188 * for input in the case of buffers partially already in memory, 189 * but the code is intricate enough already. 190 */ 191vm_page_t bogus_page; 192 193/* 194 * Synchronization (sleep/wakeup) variable for active buffer space requests. 195 * Set when wait starts, cleared prior to wakeup(). 196 * Used in runningbufwakeup() and waitrunningbufspace(). 197 */ 198static int runningbufreq; 199 200/* 201 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 202 * waitrunningbufspace(). 203 */ 204static struct mtx rbreqlock; 205 206/* 207 * Synchronization (sleep/wakeup) variable for buffer requests. 208 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 209 * by and/or. 210 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 211 * getnewbuf(), and getblk(). 212 */ 213static int needsbuffer; 214 215/* 216 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 217 */ 218static struct mtx nblock; 219 220/* 221 * Lock that protects against bwait()/bdone()/B_DONE races. 222 */ 223 224static struct mtx bdonelock; 225 226/* 227 * Definitions for the buffer free lists. 228 */ 229#define BUFFER_QUEUES 5 /* number of free buffer queues */ 230 231#define QUEUE_NONE 0 /* on no queue */ 232#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 233#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 234#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 235#define QUEUE_EMPTY 4 /* empty buffer headers */ 236 237/* Queues for free buffers with various properties */ 238static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 239 240/* Lock for the bufqueues */ 241static struct mtx bqlock; 242 243/* 244 * Single global constant for BUF_WMESG, to avoid getting multiple references. 245 * buf_wmesg is referred from macros. 246 */ 247const char *buf_wmesg = BUF_WMESG; 248 249#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 250#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 251#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 252#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 253 254#ifdef DIRECTIO 255extern void ffs_rawread_setup(void); 256#endif /* DIRECTIO */ 257/* 258 * numdirtywakeup: 259 * 260 * If someone is blocked due to there being too many dirty buffers, 261 * and numdirtybuffers is now reasonable, wake them up. 262 */ 263 264static __inline void 265numdirtywakeup(int level) 266{ 267 if (numdirtybuffers <= level) { 268 mtx_lock(&nblock); 269 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 270 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 271 wakeup(&needsbuffer); 272 } 273 mtx_unlock(&nblock); 274 } 275} 276 277/* 278 * bufspacewakeup: 279 * 280 * Called when buffer space is potentially available for recovery. 281 * getnewbuf() will block on this flag when it is unable to free 282 * sufficient buffer space. Buffer space becomes recoverable when 283 * bp's get placed back in the queues. 284 */ 285 286static __inline void 287bufspacewakeup(void) 288{ 289 /* 290 * If someone is waiting for BUF space, wake them up. Even 291 * though we haven't freed the kva space yet, the waiting 292 * process will be able to now. 293 */ 294 mtx_lock(&nblock); 295 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 296 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 297 wakeup(&needsbuffer); 298 } 299 mtx_unlock(&nblock); 300} 301 302/* 303 * runningbufwakeup() - in-progress I/O accounting. 304 * 305 */ 306static __inline void 307runningbufwakeup(struct buf *bp) 308{ 309 if (bp->b_runningbufspace) { 310 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 311 bp->b_runningbufspace = 0; 312 mtx_lock(&rbreqlock); 313 if (runningbufreq && runningbufspace <= lorunningspace) { 314 runningbufreq = 0; 315 wakeup(&runningbufreq); 316 } 317 mtx_unlock(&rbreqlock); 318 } 319} 320 321/* 322 * bufcountwakeup: 323 * 324 * Called when a buffer has been added to one of the free queues to 325 * account for the buffer and to wakeup anyone waiting for free buffers. 326 * This typically occurs when large amounts of metadata are being handled 327 * by the buffer cache ( else buffer space runs out first, usually ). 328 */ 329 330static __inline void 331bufcountwakeup(void) 332{ 333 atomic_add_int(&numfreebuffers, 1); 334 mtx_lock(&nblock); 335 if (needsbuffer) { 336 needsbuffer &= ~VFS_BIO_NEED_ANY; 337 if (numfreebuffers >= hifreebuffers) 338 needsbuffer &= ~VFS_BIO_NEED_FREE; 339 wakeup(&needsbuffer); 340 } 341 mtx_unlock(&nblock); 342} 343 344/* 345 * waitrunningbufspace() 346 * 347 * runningbufspace is a measure of the amount of I/O currently 348 * running. This routine is used in async-write situations to 349 * prevent creating huge backups of pending writes to a device. 350 * Only asynchronous writes are governed by this function. 351 * 352 * Reads will adjust runningbufspace, but will not block based on it. 353 * The read load has a side effect of reducing the allowed write load. 354 * 355 * This does NOT turn an async write into a sync write. It waits 356 * for earlier writes to complete and generally returns before the 357 * caller's write has reached the device. 358 */ 359static __inline void 360waitrunningbufspace(void) 361{ 362 mtx_lock(&rbreqlock); 363 while (runningbufspace > hirunningspace) { 364 ++runningbufreq; 365 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 366 } 367 mtx_unlock(&rbreqlock); 368} 369 370 371/* 372 * vfs_buf_test_cache: 373 * 374 * Called when a buffer is extended. This function clears the B_CACHE 375 * bit if the newly extended portion of the buffer does not contain 376 * valid data. 377 */ 378static __inline__ 379void 380vfs_buf_test_cache(struct buf *bp, 381 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 382 vm_page_t m) 383{ 384 GIANT_REQUIRED; 385 386 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 387 if (bp->b_flags & B_CACHE) { 388 int base = (foff + off) & PAGE_MASK; 389 if (vm_page_is_valid(m, base, size) == 0) 390 bp->b_flags &= ~B_CACHE; 391 } 392} 393 394/* Wake up the buffer deamon if necessary */ 395static __inline__ 396void 397bd_wakeup(int dirtybuflevel) 398{ 399 mtx_lock(&bdlock); 400 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 401 bd_request = 1; 402 wakeup(&bd_request); 403 } 404 mtx_unlock(&bdlock); 405} 406 407/* 408 * bd_speedup - speedup the buffer cache flushing code 409 */ 410 411static __inline__ 412void 413bd_speedup(void) 414{ 415 bd_wakeup(1); 416} 417 418/* 419 * Calculating buffer cache scaling values and reserve space for buffer 420 * headers. This is called during low level kernel initialization and 421 * may be called more then once. We CANNOT write to the memory area 422 * being reserved at this time. 423 */ 424caddr_t 425kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 426{ 427 /* 428 * physmem_est is in pages. Convert it to kilobytes (assumes 429 * PAGE_SIZE is >= 1K) 430 */ 431 physmem_est = physmem_est * (PAGE_SIZE / 1024); 432 433 /* 434 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 435 * For the first 64MB of ram nominally allocate sufficient buffers to 436 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 437 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 438 * the buffer cache we limit the eventual kva reservation to 439 * maxbcache bytes. 440 * 441 * factor represents the 1/4 x ram conversion. 442 */ 443 if (nbuf == 0) { 444 int factor = 4 * BKVASIZE / 1024; 445 446 nbuf = 50; 447 if (physmem_est > 4096) 448 nbuf += min((physmem_est - 4096) / factor, 449 65536 / factor); 450 if (physmem_est > 65536) 451 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 452 453 if (maxbcache && nbuf > maxbcache / BKVASIZE) 454 nbuf = maxbcache / BKVASIZE; 455 } 456 457#if 0 458 /* 459 * Do not allow the buffer_map to be more then 1/2 the size of the 460 * kernel_map. 461 */ 462 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 463 (BKVASIZE * 2)) { 464 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 465 (BKVASIZE * 2); 466 printf("Warning: nbufs capped at %d\n", nbuf); 467 } 468#endif 469 470 /* 471 * swbufs are used as temporary holders for I/O, such as paging I/O. 472 * We have no less then 16 and no more then 256. 473 */ 474 nswbuf = max(min(nbuf/4, 256), 16); 475#ifdef NSWBUF_MIN 476 if (nswbuf < NSWBUF_MIN) 477 nswbuf = NSWBUF_MIN; 478#endif 479#ifdef DIRECTIO 480 ffs_rawread_setup(); 481#endif 482 483 /* 484 * Reserve space for the buffer cache buffers 485 */ 486 swbuf = (void *)v; 487 v = (caddr_t)(swbuf + nswbuf); 488 buf = (void *)v; 489 v = (caddr_t)(buf + nbuf); 490 491 return(v); 492} 493 494/* Initialize the buffer subsystem. Called before use of any buffers. */ 495void 496bufinit(void) 497{ 498 struct buf *bp; 499 int i; 500 501 GIANT_REQUIRED; 502 503 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 504 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 505 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 506 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 507 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 508 509 /* next, make a null set of free lists */ 510 for (i = 0; i < BUFFER_QUEUES; i++) 511 TAILQ_INIT(&bufqueues[i]); 512 513 /* finally, initialize each buffer header and stick on empty q */ 514 for (i = 0; i < nbuf; i++) { 515 bp = &buf[i]; 516 bzero(bp, sizeof *bp); 517 bp->b_flags = B_INVAL; /* we're just an empty header */ 518 bp->b_dev = NODEV; 519 bp->b_rcred = NOCRED; 520 bp->b_wcred = NOCRED; 521 bp->b_qindex = QUEUE_EMPTY; 522 bp->b_vflags = 0; 523 bp->b_xflags = 0; 524 LIST_INIT(&bp->b_dep); 525 BUF_LOCKINIT(bp); 526 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 527 } 528 529 /* 530 * maxbufspace is the absolute maximum amount of buffer space we are 531 * allowed to reserve in KVM and in real terms. The absolute maximum 532 * is nominally used by buf_daemon. hibufspace is the nominal maximum 533 * used by most other processes. The differential is required to 534 * ensure that buf_daemon is able to run when other processes might 535 * be blocked waiting for buffer space. 536 * 537 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 538 * this may result in KVM fragmentation which is not handled optimally 539 * by the system. 540 */ 541 maxbufspace = nbuf * BKVASIZE; 542 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 543 lobufspace = hibufspace - MAXBSIZE; 544 545 lorunningspace = 512 * 1024; 546 hirunningspace = 1024 * 1024; 547 548/* 549 * Limit the amount of malloc memory since it is wired permanently into 550 * the kernel space. Even though this is accounted for in the buffer 551 * allocation, we don't want the malloced region to grow uncontrolled. 552 * The malloc scheme improves memory utilization significantly on average 553 * (small) directories. 554 */ 555 maxbufmallocspace = hibufspace / 20; 556 557/* 558 * Reduce the chance of a deadlock occuring by limiting the number 559 * of delayed-write dirty buffers we allow to stack up. 560 */ 561 hidirtybuffers = nbuf / 4 + 20; 562 dirtybufthresh = hidirtybuffers * 9 / 10; 563 numdirtybuffers = 0; 564/* 565 * To support extreme low-memory systems, make sure hidirtybuffers cannot 566 * eat up all available buffer space. This occurs when our minimum cannot 567 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 568 * BKVASIZE'd (8K) buffers. 569 */ 570 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 571 hidirtybuffers >>= 1; 572 } 573 lodirtybuffers = hidirtybuffers / 2; 574 575/* 576 * Try to keep the number of free buffers in the specified range, 577 * and give special processes (e.g. like buf_daemon) access to an 578 * emergency reserve. 579 */ 580 lofreebuffers = nbuf / 18 + 5; 581 hifreebuffers = 2 * lofreebuffers; 582 numfreebuffers = nbuf; 583 584/* 585 * Maximum number of async ops initiated per buf_daemon loop. This is 586 * somewhat of a hack at the moment, we really need to limit ourselves 587 * based on the number of bytes of I/O in-transit that were initiated 588 * from buf_daemon. 589 */ 590 591 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 592 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 593} 594 595/* 596 * bfreekva() - free the kva allocation for a buffer. 597 * 598 * Must be called at splbio() or higher as this is the only locking for 599 * buffer_map. 600 * 601 * Since this call frees up buffer space, we call bufspacewakeup(). 602 */ 603static void 604bfreekva(struct buf * bp) 605{ 606 GIANT_REQUIRED; 607 608 if (bp->b_kvasize) { 609 atomic_add_int(&buffreekvacnt, 1); 610 atomic_subtract_int(&bufspace, bp->b_kvasize); 611 vm_map_delete(buffer_map, 612 (vm_offset_t) bp->b_kvabase, 613 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 614 ); 615 bp->b_kvasize = 0; 616 bufspacewakeup(); 617 } 618} 619 620/* 621 * bremfree: 622 * 623 * Remove the buffer from the appropriate free list. 624 */ 625void 626bremfree(struct buf * bp) 627{ 628 mtx_lock(&bqlock); 629 bremfreel(bp); 630 mtx_unlock(&bqlock); 631} 632 633void 634bremfreel(struct buf * bp) 635{ 636 int s = splbio(); 637 int old_qindex = bp->b_qindex; 638 639 GIANT_REQUIRED; 640 641 if (bp->b_qindex != QUEUE_NONE) { 642 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 643 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 644 bp->b_qindex = QUEUE_NONE; 645 } else { 646 if (BUF_REFCNT(bp) <= 1) 647 panic("bremfree: removing a buffer not on a queue"); 648 } 649 650 /* 651 * Fixup numfreebuffers count. If the buffer is invalid or not 652 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 653 * the buffer was free and we must decrement numfreebuffers. 654 */ 655 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 656 switch(old_qindex) { 657 case QUEUE_DIRTY: 658 case QUEUE_CLEAN: 659 case QUEUE_EMPTY: 660 case QUEUE_EMPTYKVA: 661 atomic_subtract_int(&numfreebuffers, 1); 662 break; 663 default: 664 break; 665 } 666 } 667 splx(s); 668} 669 670 671/* 672 * Get a buffer with the specified data. Look in the cache first. We 673 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 674 * is set, the buffer is valid and we do not have to do anything ( see 675 * getblk() ). This is really just a special case of breadn(). 676 */ 677int 678bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 679 struct buf ** bpp) 680{ 681 682 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 683} 684 685/* 686 * Operates like bread, but also starts asynchronous I/O on 687 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 688 * to initiating I/O . If B_CACHE is set, the buffer is valid 689 * and we do not have to do anything. 690 */ 691int 692breadn(struct vnode * vp, daddr_t blkno, int size, 693 daddr_t * rablkno, int *rabsize, 694 int cnt, struct ucred * cred, struct buf ** bpp) 695{ 696 struct buf *bp, *rabp; 697 int i; 698 int rv = 0, readwait = 0; 699 700 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 701 702 /* if not found in cache, do some I/O */ 703 if ((bp->b_flags & B_CACHE) == 0) { 704 if (curthread != PCPU_GET(idlethread)) 705 curthread->td_proc->p_stats->p_ru.ru_inblock++; 706 bp->b_iocmd = BIO_READ; 707 bp->b_flags &= ~B_INVAL; 708 bp->b_ioflags &= ~BIO_ERROR; 709 if (bp->b_rcred == NOCRED && cred != NOCRED) 710 bp->b_rcred = crhold(cred); 711 vfs_busy_pages(bp, 0); 712 bp->b_iooffset = dbtob(bp->b_blkno); 713 if (vp->v_type == VCHR) 714 VOP_SPECSTRATEGY(vp, bp); 715 else 716 VOP_STRATEGY(vp, bp); 717 ++readwait; 718 } 719 720 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 721 if (inmem(vp, *rablkno)) 722 continue; 723 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 724 725 if ((rabp->b_flags & B_CACHE) == 0) { 726 if (curthread != PCPU_GET(idlethread)) 727 curthread->td_proc->p_stats->p_ru.ru_inblock++; 728 rabp->b_flags |= B_ASYNC; 729 rabp->b_flags &= ~B_INVAL; 730 rabp->b_ioflags &= ~BIO_ERROR; 731 rabp->b_iocmd = BIO_READ; 732 if (rabp->b_rcred == NOCRED && cred != NOCRED) 733 rabp->b_rcred = crhold(cred); 734 vfs_busy_pages(rabp, 0); 735 BUF_KERNPROC(rabp); 736 rabp->b_iooffset = dbtob(rabp->b_blkno); 737 if (vp->v_type == VCHR) 738 VOP_SPECSTRATEGY(vp, rabp); 739 else 740 VOP_STRATEGY(vp, rabp); 741 } else { 742 brelse(rabp); 743 } 744 } 745 746 if (readwait) { 747 rv = bufwait(bp); 748 } 749 return (rv); 750} 751 752/* 753 * Write, release buffer on completion. (Done by iodone 754 * if async). Do not bother writing anything if the buffer 755 * is invalid. 756 * 757 * Note that we set B_CACHE here, indicating that buffer is 758 * fully valid and thus cacheable. This is true even of NFS 759 * now so we set it generally. This could be set either here 760 * or in biodone() since the I/O is synchronous. We put it 761 * here. 762 */ 763 764int 765bwrite(struct buf * bp) 766{ 767 int oldflags, s; 768 struct buf *newbp; 769 770 if (bp->b_flags & B_INVAL) { 771 brelse(bp); 772 return (0); 773 } 774 775 oldflags = bp->b_flags; 776 777 if (BUF_REFCNT(bp) == 0) 778 panic("bwrite: buffer is not busy???"); 779 s = splbio(); 780 /* 781 * If a background write is already in progress, delay 782 * writing this block if it is asynchronous. Otherwise 783 * wait for the background write to complete. 784 */ 785 VI_LOCK(bp->b_vp); 786 if (bp->b_vflags & BV_BKGRDINPROG) { 787 if (bp->b_flags & B_ASYNC) { 788 VI_UNLOCK(bp->b_vp); 789 splx(s); 790 bdwrite(bp); 791 return (0); 792 } 793 bp->b_vflags |= BV_BKGRDWAIT; 794 msleep(&bp->b_xflags, VI_MTX(bp->b_vp), PRIBIO, "bwrbg", 0); 795 if (bp->b_vflags & BV_BKGRDINPROG) 796 panic("bwrite: still writing"); 797 } 798 VI_UNLOCK(bp->b_vp); 799 800 /* Mark the buffer clean */ 801 bundirty(bp); 802 803 /* 804 * If this buffer is marked for background writing and we 805 * do not have to wait for it, make a copy and write the 806 * copy so as to leave this buffer ready for further use. 807 * 808 * This optimization eats a lot of memory. If we have a page 809 * or buffer shortfall we can't do it. 810 */ 811 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 812 (bp->b_flags & B_ASYNC) && 813 !vm_page_count_severe() && 814 !buf_dirty_count_severe()) { 815 if (bp->b_iodone != NULL) { 816 printf("bp->b_iodone = %p\n", bp->b_iodone); 817 panic("bwrite: need chained iodone"); 818 } 819 820 /* get a new block */ 821 newbp = geteblk(bp->b_bufsize); 822 823 /* 824 * set it to be identical to the old block. We have to 825 * set b_lblkno and BKGRDMARKER before calling bgetvp() 826 * to avoid confusing the splay tree and gbincore(). 827 */ 828 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 829 newbp->b_lblkno = bp->b_lblkno; 830 newbp->b_xflags |= BX_BKGRDMARKER; 831 VI_LOCK(bp->b_vp); 832 bp->b_vflags |= BV_BKGRDINPROG; 833 bgetvp(bp->b_vp, newbp); 834 VI_UNLOCK(bp->b_vp); 835 newbp->b_blkno = bp->b_blkno; 836 newbp->b_offset = bp->b_offset; 837 newbp->b_iodone = vfs_backgroundwritedone; 838 newbp->b_flags |= B_ASYNC; 839 newbp->b_flags &= ~B_INVAL; 840 841 /* move over the dependencies */ 842 if (LIST_FIRST(&bp->b_dep) != NULL) 843 buf_movedeps(bp, newbp); 844 845 /* 846 * Initiate write on the copy, release the original to 847 * the B_LOCKED queue so that it cannot go away until 848 * the background write completes. If not locked it could go 849 * away and then be reconstituted while it was being written. 850 * If the reconstituted buffer were written, we could end up 851 * with two background copies being written at the same time. 852 */ 853 bqrelse(bp); 854 bp = newbp; 855 } 856 857 bp->b_flags &= ~B_DONE; 858 bp->b_ioflags &= ~BIO_ERROR; 859 bp->b_flags |= B_WRITEINPROG | B_CACHE; 860 bp->b_iocmd = BIO_WRITE; 861 862 VI_LOCK(bp->b_vp); 863 bp->b_vp->v_numoutput++; 864 VI_UNLOCK(bp->b_vp); 865 vfs_busy_pages(bp, 1); 866 867 /* 868 * Normal bwrites pipeline writes 869 */ 870 bp->b_runningbufspace = bp->b_bufsize; 871 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 872 873 if (curthread != PCPU_GET(idlethread)) 874 curthread->td_proc->p_stats->p_ru.ru_oublock++; 875 splx(s); 876 if (oldflags & B_ASYNC) 877 BUF_KERNPROC(bp); 878 bp->b_iooffset = dbtob(bp->b_blkno); 879 if (bp->b_vp->v_type == VCHR) 880 VOP_SPECSTRATEGY(bp->b_vp, bp); 881 else 882 VOP_STRATEGY(bp->b_vp, bp); 883 884 if ((oldflags & B_ASYNC) == 0) { 885 int rtval = bufwait(bp); 886 brelse(bp); 887 return (rtval); 888 } else { 889 /* 890 * don't allow the async write to saturate the I/O 891 * system. We will not deadlock here because 892 * we are blocking waiting for I/O that is already in-progress 893 * to complete. We do not block here if it is the update 894 * or syncer daemon trying to clean up as that can lead 895 * to deadlock. 896 */ 897 if (curthread->td_proc != bufdaemonproc && 898 curthread->td_proc != updateproc) 899 waitrunningbufspace(); 900 } 901 902 return (0); 903} 904 905/* 906 * Complete a background write started from bwrite. 907 */ 908static void 909vfs_backgroundwritedone(bp) 910 struct buf *bp; 911{ 912 struct buf *origbp; 913 914 /* 915 * Find the original buffer that we are writing. 916 */ 917 VI_LOCK(bp->b_vp); 918 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 919 panic("backgroundwritedone: lost buffer"); 920 921 /* 922 * Clear the BV_BKGRDINPROG flag in the original buffer 923 * and awaken it if it is waiting for the write to complete. 924 * If BV_BKGRDINPROG is not set in the original buffer it must 925 * have been released and re-instantiated - which is not legal. 926 */ 927 KASSERT((origbp->b_vflags & BV_BKGRDINPROG), 928 ("backgroundwritedone: lost buffer2")); 929 origbp->b_vflags &= ~BV_BKGRDINPROG; 930 if (origbp->b_vflags & BV_BKGRDWAIT) { 931 origbp->b_vflags &= ~BV_BKGRDWAIT; 932 wakeup(&origbp->b_xflags); 933 } 934 VI_UNLOCK(bp->b_vp); 935 /* 936 * Process dependencies then return any unfinished ones. 937 */ 938 if (LIST_FIRST(&bp->b_dep) != NULL) 939 buf_complete(bp); 940 if (LIST_FIRST(&bp->b_dep) != NULL) 941 buf_movedeps(bp, origbp); 942 943 /* 944 * This buffer is marked B_NOCACHE, so when it is released 945 * by biodone, it will be tossed. We mark it with BIO_READ 946 * to avoid biodone doing a second vwakeup. 947 */ 948 bp->b_flags |= B_NOCACHE; 949 bp->b_iocmd = BIO_READ; 950 bp->b_flags &= ~(B_CACHE | B_DONE); 951 bp->b_iodone = 0; 952 bufdone(bp); 953} 954 955/* 956 * Delayed write. (Buffer is marked dirty). Do not bother writing 957 * anything if the buffer is marked invalid. 958 * 959 * Note that since the buffer must be completely valid, we can safely 960 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 961 * biodone() in order to prevent getblk from writing the buffer 962 * out synchronously. 963 */ 964void 965bdwrite(struct buf * bp) 966{ 967 struct thread *td = curthread; 968 struct vnode *vp; 969 struct buf *nbp; 970 971 GIANT_REQUIRED; 972 973 if (BUF_REFCNT(bp) == 0) 974 panic("bdwrite: buffer is not busy"); 975 976 if (bp->b_flags & B_INVAL) { 977 brelse(bp); 978 return; 979 } 980 981 /* 982 * If we have too many dirty buffers, don't create any more. 983 * If we are wildly over our limit, then force a complete 984 * cleanup. Otherwise, just keep the situation from getting 985 * out of control. Note that we have to avoid a recursive 986 * disaster and not try to clean up after our own cleanup! 987 */ 988 vp = bp->b_vp; 989 VI_LOCK(vp); 990 if (td->td_pflags & TDP_COWINPROGRESS) { 991 recursiveflushes++; 992 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) { 993 VI_UNLOCK(vp); 994 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 995 VI_LOCK(vp); 996 altbufferflushes++; 997 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) { 998 /* 999 * Try to find a buffer to flush. 1000 */ 1001 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) { 1002 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1003 buf_countdeps(nbp, 0) || 1004 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1005 continue; 1006 if (bp == nbp) 1007 panic("bdwrite: found ourselves"); 1008 VI_UNLOCK(vp); 1009 if (nbp->b_flags & B_CLUSTEROK) { 1010 vfs_bio_awrite(nbp); 1011 } else { 1012 bremfree(nbp); 1013 bawrite(nbp); 1014 } 1015 VI_LOCK(vp); 1016 dirtybufferflushes++; 1017 break; 1018 } 1019 } 1020 VI_UNLOCK(vp); 1021 1022 bdirty(bp); 1023 /* 1024 * Set B_CACHE, indicating that the buffer is fully valid. This is 1025 * true even of NFS now. 1026 */ 1027 bp->b_flags |= B_CACHE; 1028 1029 /* 1030 * This bmap keeps the system from needing to do the bmap later, 1031 * perhaps when the system is attempting to do a sync. Since it 1032 * is likely that the indirect block -- or whatever other datastructure 1033 * that the filesystem needs is still in memory now, it is a good 1034 * thing to do this. Note also, that if the pageout daemon is 1035 * requesting a sync -- there might not be enough memory to do 1036 * the bmap then... So, this is important to do. 1037 */ 1038 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1039 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1040 } 1041 1042 /* 1043 * Set the *dirty* buffer range based upon the VM system dirty pages. 1044 */ 1045 vfs_setdirty(bp); 1046 1047 /* 1048 * We need to do this here to satisfy the vnode_pager and the 1049 * pageout daemon, so that it thinks that the pages have been 1050 * "cleaned". Note that since the pages are in a delayed write 1051 * buffer -- the VFS layer "will" see that the pages get written 1052 * out on the next sync, or perhaps the cluster will be completed. 1053 */ 1054 vfs_clean_pages(bp); 1055 bqrelse(bp); 1056 1057 /* 1058 * Wakeup the buffer flushing daemon if we have a lot of dirty 1059 * buffers (midpoint between our recovery point and our stall 1060 * point). 1061 */ 1062 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1063 1064 /* 1065 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1066 * due to the softdep code. 1067 */ 1068} 1069 1070/* 1071 * bdirty: 1072 * 1073 * Turn buffer into delayed write request. We must clear BIO_READ and 1074 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1075 * itself to properly update it in the dirty/clean lists. We mark it 1076 * B_DONE to ensure that any asynchronization of the buffer properly 1077 * clears B_DONE ( else a panic will occur later ). 1078 * 1079 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1080 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1081 * should only be called if the buffer is known-good. 1082 * 1083 * Since the buffer is not on a queue, we do not update the numfreebuffers 1084 * count. 1085 * 1086 * Must be called at splbio(). 1087 * The buffer must be on QUEUE_NONE. 1088 */ 1089void 1090bdirty(bp) 1091 struct buf *bp; 1092{ 1093 KASSERT(bp->b_qindex == QUEUE_NONE, 1094 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1095 bp->b_flags &= ~(B_RELBUF); 1096 bp->b_iocmd = BIO_WRITE; 1097 1098 if ((bp->b_flags & B_DELWRI) == 0) { 1099 bp->b_flags |= B_DONE | B_DELWRI; 1100 reassignbuf(bp, bp->b_vp); 1101 atomic_add_int(&numdirtybuffers, 1); 1102 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1103 } 1104} 1105 1106/* 1107 * bundirty: 1108 * 1109 * Clear B_DELWRI for buffer. 1110 * 1111 * Since the buffer is not on a queue, we do not update the numfreebuffers 1112 * count. 1113 * 1114 * Must be called at splbio(). 1115 * The buffer must be on QUEUE_NONE. 1116 */ 1117 1118void 1119bundirty(bp) 1120 struct buf *bp; 1121{ 1122 KASSERT(bp->b_qindex == QUEUE_NONE, 1123 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1124 1125 if (bp->b_flags & B_DELWRI) { 1126 bp->b_flags &= ~B_DELWRI; 1127 reassignbuf(bp, bp->b_vp); 1128 atomic_subtract_int(&numdirtybuffers, 1); 1129 numdirtywakeup(lodirtybuffers); 1130 } 1131 /* 1132 * Since it is now being written, we can clear its deferred write flag. 1133 */ 1134 bp->b_flags &= ~B_DEFERRED; 1135} 1136 1137/* 1138 * bawrite: 1139 * 1140 * Asynchronous write. Start output on a buffer, but do not wait for 1141 * it to complete. The buffer is released when the output completes. 1142 * 1143 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1144 * B_INVAL buffers. Not us. 1145 */ 1146void 1147bawrite(struct buf * bp) 1148{ 1149 bp->b_flags |= B_ASYNC; 1150 (void) BUF_WRITE(bp); 1151} 1152 1153/* 1154 * bwillwrite: 1155 * 1156 * Called prior to the locking of any vnodes when we are expecting to 1157 * write. We do not want to starve the buffer cache with too many 1158 * dirty buffers so we block here. By blocking prior to the locking 1159 * of any vnodes we attempt to avoid the situation where a locked vnode 1160 * prevents the various system daemons from flushing related buffers. 1161 */ 1162 1163void 1164bwillwrite(void) 1165{ 1166 if (numdirtybuffers >= hidirtybuffers) { 1167 int s; 1168 1169 mtx_lock(&Giant); 1170 s = splbio(); 1171 mtx_lock(&nblock); 1172 while (numdirtybuffers >= hidirtybuffers) { 1173 bd_wakeup(1); 1174 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1175 msleep(&needsbuffer, &nblock, 1176 (PRIBIO + 4), "flswai", 0); 1177 } 1178 splx(s); 1179 mtx_unlock(&nblock); 1180 mtx_unlock(&Giant); 1181 } 1182} 1183 1184/* 1185 * Return true if we have too many dirty buffers. 1186 */ 1187int 1188buf_dirty_count_severe(void) 1189{ 1190 return(numdirtybuffers >= hidirtybuffers); 1191} 1192 1193/* 1194 * brelse: 1195 * 1196 * Release a busy buffer and, if requested, free its resources. The 1197 * buffer will be stashed in the appropriate bufqueue[] allowing it 1198 * to be accessed later as a cache entity or reused for other purposes. 1199 */ 1200void 1201brelse(struct buf * bp) 1202{ 1203 int s; 1204 1205 GIANT_REQUIRED; 1206 1207 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1208 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1209 1210 s = splbio(); 1211 1212 if (bp->b_iocmd == BIO_WRITE && 1213 (bp->b_ioflags & BIO_ERROR) && 1214 !(bp->b_flags & B_INVAL)) { 1215 /* 1216 * Failed write, redirty. Must clear BIO_ERROR to prevent 1217 * pages from being scrapped. If B_INVAL is set then 1218 * this case is not run and the next case is run to 1219 * destroy the buffer. B_INVAL can occur if the buffer 1220 * is outside the range supported by the underlying device. 1221 */ 1222 bp->b_ioflags &= ~BIO_ERROR; 1223 bdirty(bp); 1224 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1225 (bp->b_ioflags & BIO_ERROR) || 1226 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1227 /* 1228 * Either a failed I/O or we were asked to free or not 1229 * cache the buffer. 1230 */ 1231 bp->b_flags |= B_INVAL; 1232 if (LIST_FIRST(&bp->b_dep) != NULL) 1233 buf_deallocate(bp); 1234 if (bp->b_flags & B_DELWRI) { 1235 atomic_subtract_int(&numdirtybuffers, 1); 1236 numdirtywakeup(lodirtybuffers); 1237 } 1238 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1239 if ((bp->b_flags & B_VMIO) == 0) { 1240 if (bp->b_bufsize) 1241 allocbuf(bp, 0); 1242 if (bp->b_vp) 1243 brelvp(bp); 1244 } 1245 } 1246 1247 /* 1248 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1249 * is called with B_DELWRI set, the underlying pages may wind up 1250 * getting freed causing a previous write (bdwrite()) to get 'lost' 1251 * because pages associated with a B_DELWRI bp are marked clean. 1252 * 1253 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1254 * if B_DELWRI is set. 1255 * 1256 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1257 * on pages to return pages to the VM page queues. 1258 */ 1259 if (bp->b_flags & B_DELWRI) 1260 bp->b_flags &= ~B_RELBUF; 1261 else if (vm_page_count_severe()) { 1262 /* 1263 * XXX This lock may not be necessary since BKGRDINPROG 1264 * cannot be set while we hold the buf lock, it can only be 1265 * cleared if it is already pending. 1266 */ 1267 if (bp->b_vp) { 1268 VI_LOCK(bp->b_vp); 1269 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1270 bp->b_flags |= B_RELBUF; 1271 VI_UNLOCK(bp->b_vp); 1272 } else 1273 bp->b_flags |= B_RELBUF; 1274 } 1275 1276 /* 1277 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1278 * constituted, not even NFS buffers now. Two flags effect this. If 1279 * B_INVAL, the struct buf is invalidated but the VM object is kept 1280 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1281 * 1282 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1283 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1284 * buffer is also B_INVAL because it hits the re-dirtying code above. 1285 * 1286 * Normally we can do this whether a buffer is B_DELWRI or not. If 1287 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1288 * the commit state and we cannot afford to lose the buffer. If the 1289 * buffer has a background write in progress, we need to keep it 1290 * around to prevent it from being reconstituted and starting a second 1291 * background write. 1292 */ 1293 if ((bp->b_flags & B_VMIO) 1294 && !(bp->b_vp->v_mount != NULL && 1295 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1296 !vn_isdisk(bp->b_vp, NULL) && 1297 (bp->b_flags & B_DELWRI)) 1298 ) { 1299 1300 int i, j, resid; 1301 vm_page_t m; 1302 off_t foff; 1303 vm_pindex_t poff; 1304 vm_object_t obj; 1305 1306 obj = bp->b_object; 1307 1308 /* 1309 * Get the base offset and length of the buffer. Note that 1310 * in the VMIO case if the buffer block size is not 1311 * page-aligned then b_data pointer may not be page-aligned. 1312 * But our b_pages[] array *IS* page aligned. 1313 * 1314 * block sizes less then DEV_BSIZE (usually 512) are not 1315 * supported due to the page granularity bits (m->valid, 1316 * m->dirty, etc...). 1317 * 1318 * See man buf(9) for more information 1319 */ 1320 resid = bp->b_bufsize; 1321 foff = bp->b_offset; 1322 VM_OBJECT_LOCK(obj); 1323 for (i = 0; i < bp->b_npages; i++) { 1324 int had_bogus = 0; 1325 1326 m = bp->b_pages[i]; 1327 vm_page_lock_queues(); 1328 vm_page_flag_clear(m, PG_ZERO); 1329 vm_page_unlock_queues(); 1330 1331 /* 1332 * If we hit a bogus page, fixup *all* the bogus pages 1333 * now. 1334 */ 1335 if (m == bogus_page) { 1336 poff = OFF_TO_IDX(bp->b_offset); 1337 had_bogus = 1; 1338 1339 for (j = i; j < bp->b_npages; j++) { 1340 vm_page_t mtmp; 1341 mtmp = bp->b_pages[j]; 1342 if (mtmp == bogus_page) { 1343 mtmp = vm_page_lookup(obj, poff + j); 1344 if (!mtmp) { 1345 panic("brelse: page missing\n"); 1346 } 1347 bp->b_pages[j] = mtmp; 1348 } 1349 } 1350 1351 if ((bp->b_flags & B_INVAL) == 0) { 1352 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1353 } 1354 m = bp->b_pages[i]; 1355 } 1356 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1357 int poffset = foff & PAGE_MASK; 1358 int presid = resid > (PAGE_SIZE - poffset) ? 1359 (PAGE_SIZE - poffset) : resid; 1360 1361 KASSERT(presid >= 0, ("brelse: extra page")); 1362 vm_page_lock_queues(); 1363 vm_page_set_invalid(m, poffset, presid); 1364 vm_page_unlock_queues(); 1365 if (had_bogus) 1366 printf("avoided corruption bug in bogus_page/brelse code\n"); 1367 } 1368 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1369 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1370 } 1371 VM_OBJECT_UNLOCK(obj); 1372 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1373 vfs_vmio_release(bp); 1374 1375 } else if (bp->b_flags & B_VMIO) { 1376 1377 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1378 vfs_vmio_release(bp); 1379 } 1380 1381 } 1382 1383 if (bp->b_qindex != QUEUE_NONE) 1384 panic("brelse: free buffer onto another queue???"); 1385 if (BUF_REFCNT(bp) > 1) { 1386 /* do not release to free list */ 1387 BUF_UNLOCK(bp); 1388 splx(s); 1389 return; 1390 } 1391 1392 /* enqueue */ 1393 mtx_lock(&bqlock); 1394 1395 /* buffers with no memory */ 1396 if (bp->b_bufsize == 0) { 1397 bp->b_flags |= B_INVAL; 1398 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1399 if (bp->b_vflags & BV_BKGRDINPROG) 1400 panic("losing buffer 1"); 1401 if (bp->b_kvasize) { 1402 bp->b_qindex = QUEUE_EMPTYKVA; 1403 } else { 1404 bp->b_qindex = QUEUE_EMPTY; 1405 } 1406 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1407 bp->b_dev = NODEV; 1408 /* buffers with junk contents */ 1409 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1410 (bp->b_ioflags & BIO_ERROR)) { 1411 bp->b_flags |= B_INVAL; 1412 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1413 if (bp->b_vflags & BV_BKGRDINPROG) 1414 panic("losing buffer 2"); 1415 bp->b_qindex = QUEUE_CLEAN; 1416 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1417 bp->b_dev = NODEV; 1418 /* remaining buffers */ 1419 } else { 1420 if (bp->b_flags & B_DELWRI) 1421 bp->b_qindex = QUEUE_DIRTY; 1422 else 1423 bp->b_qindex = QUEUE_CLEAN; 1424 if (bp->b_flags & B_AGE) 1425 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1426 else 1427 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1428 } 1429 mtx_unlock(&bqlock); 1430 1431 /* 1432 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1433 * placed the buffer on the correct queue. We must also disassociate 1434 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1435 * find it. 1436 */ 1437 if (bp->b_flags & B_INVAL) { 1438 if (bp->b_flags & B_DELWRI) 1439 bundirty(bp); 1440 if (bp->b_vp) 1441 brelvp(bp); 1442 } 1443 1444 /* 1445 * Fixup numfreebuffers count. The bp is on an appropriate queue 1446 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1447 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1448 * if B_INVAL is set ). 1449 */ 1450 1451 if (!(bp->b_flags & B_DELWRI)) 1452 bufcountwakeup(); 1453 1454 /* 1455 * Something we can maybe free or reuse 1456 */ 1457 if (bp->b_bufsize || bp->b_kvasize) 1458 bufspacewakeup(); 1459 1460 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1461 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1462 panic("brelse: not dirty"); 1463 /* unlock */ 1464 BUF_UNLOCK(bp); 1465 splx(s); 1466} 1467 1468/* 1469 * Release a buffer back to the appropriate queue but do not try to free 1470 * it. The buffer is expected to be used again soon. 1471 * 1472 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1473 * biodone() to requeue an async I/O on completion. It is also used when 1474 * known good buffers need to be requeued but we think we may need the data 1475 * again soon. 1476 * 1477 * XXX we should be able to leave the B_RELBUF hint set on completion. 1478 */ 1479void 1480bqrelse(struct buf * bp) 1481{ 1482 int s; 1483 1484 s = splbio(); 1485 1486 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1487 1488 if (bp->b_qindex != QUEUE_NONE) 1489 panic("bqrelse: free buffer onto another queue???"); 1490 if (BUF_REFCNT(bp) > 1) { 1491 /* do not release to free list */ 1492 BUF_UNLOCK(bp); 1493 splx(s); 1494 return; 1495 } 1496 mtx_lock(&bqlock); 1497 /* buffers with stale but valid contents */ 1498 if (bp->b_flags & B_DELWRI) { 1499 bp->b_qindex = QUEUE_DIRTY; 1500 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1501 } else { 1502 /* 1503 * XXX This lock may not be necessary since BKGRDINPROG 1504 * cannot be set while we hold the buf lock, it can only be 1505 * cleared if it is already pending. 1506 */ 1507 VI_LOCK(bp->b_vp); 1508 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1509 VI_UNLOCK(bp->b_vp); 1510 bp->b_qindex = QUEUE_CLEAN; 1511 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1512 b_freelist); 1513 } else { 1514 /* 1515 * We are too low on memory, we have to try to free 1516 * the buffer (most importantly: the wired pages 1517 * making up its backing store) *now*. 1518 */ 1519 VI_UNLOCK(bp->b_vp); 1520 mtx_unlock(&bqlock); 1521 splx(s); 1522 brelse(bp); 1523 return; 1524 } 1525 } 1526 mtx_unlock(&bqlock); 1527 1528 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1529 bufcountwakeup(); 1530 1531 /* 1532 * Something we can maybe free or reuse. 1533 */ 1534 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1535 bufspacewakeup(); 1536 1537 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1538 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1539 panic("bqrelse: not dirty"); 1540 /* unlock */ 1541 BUF_UNLOCK(bp); 1542 splx(s); 1543} 1544 1545/* Give pages used by the bp back to the VM system (where possible) */ 1546static void 1547vfs_vmio_release(bp) 1548 struct buf *bp; 1549{ 1550 int i; 1551 vm_page_t m; 1552 1553 GIANT_REQUIRED; 1554 VM_OBJECT_LOCK(bp->b_object); 1555 vm_page_lock_queues(); 1556 for (i = 0; i < bp->b_npages; i++) { 1557 m = bp->b_pages[i]; 1558 bp->b_pages[i] = NULL; 1559 /* 1560 * In order to keep page LRU ordering consistent, put 1561 * everything on the inactive queue. 1562 */ 1563 vm_page_unwire(m, 0); 1564 /* 1565 * We don't mess with busy pages, it is 1566 * the responsibility of the process that 1567 * busied the pages to deal with them. 1568 */ 1569 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1570 continue; 1571 1572 if (m->wire_count == 0) { 1573 vm_page_flag_clear(m, PG_ZERO); 1574 /* 1575 * Might as well free the page if we can and it has 1576 * no valid data. We also free the page if the 1577 * buffer was used for direct I/O 1578 */ 1579 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1580 m->hold_count == 0) { 1581 vm_page_busy(m); 1582 pmap_remove_all(m); 1583 vm_page_free(m); 1584 } else if (bp->b_flags & B_DIRECT) { 1585 vm_page_try_to_free(m); 1586 } else if (vm_page_count_severe()) { 1587 vm_page_try_to_cache(m); 1588 } 1589 } 1590 } 1591 vm_page_unlock_queues(); 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 kproc_desc buf_kp = { 2044 "bufdaemon", 2045 buf_daemon, 2046 &bufdaemonproc 2047}; 2048SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2049 2050static void 2051buf_daemon() 2052{ 2053 int s; 2054 2055 mtx_lock(&Giant); 2056 2057 /* 2058 * This process needs to be suspended prior to shutdown sync. 2059 */ 2060 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2061 SHUTDOWN_PRI_LAST); 2062 2063 /* 2064 * This process is allowed to take the buffer cache to the limit 2065 */ 2066 s = splbio(); 2067 mtx_lock(&bdlock); 2068 2069 for (;;) { 2070 bd_request = 0; 2071 mtx_unlock(&bdlock); 2072 2073 kthread_suspend_check(bufdaemonproc); 2074 2075 /* 2076 * Do the flush. Limit the amount of in-transit I/O we 2077 * allow to build up, otherwise we would completely saturate 2078 * the I/O system. Wakeup any waiting processes before we 2079 * normally would so they can run in parallel with our drain. 2080 */ 2081 while (numdirtybuffers > lodirtybuffers) { 2082 if (flushbufqueues(0) == 0) { 2083 /* 2084 * Could not find any buffers without rollback 2085 * dependencies, so just write the first one 2086 * in the hopes of eventually making progress. 2087 */ 2088 flushbufqueues(1); 2089 break; 2090 } 2091 waitrunningbufspace(); 2092 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2093 } 2094 2095 /* 2096 * Only clear bd_request if we have reached our low water 2097 * mark. The buf_daemon normally waits 1 second and 2098 * then incrementally flushes any dirty buffers that have 2099 * built up, within reason. 2100 * 2101 * If we were unable to hit our low water mark and couldn't 2102 * find any flushable buffers, we sleep half a second. 2103 * Otherwise we loop immediately. 2104 */ 2105 mtx_lock(&bdlock); 2106 if (numdirtybuffers <= lodirtybuffers) { 2107 /* 2108 * We reached our low water mark, reset the 2109 * request and sleep until we are needed again. 2110 * The sleep is just so the suspend code works. 2111 */ 2112 bd_request = 0; 2113 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2114 } else { 2115 /* 2116 * We couldn't find any flushable dirty buffers but 2117 * still have too many dirty buffers, we 2118 * have to sleep and try again. (rare) 2119 */ 2120 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2121 } 2122 } 2123} 2124 2125/* 2126 * flushbufqueues: 2127 * 2128 * Try to flush a buffer in the dirty queue. We must be careful to 2129 * free up B_INVAL buffers instead of write them, which NFS is 2130 * particularly sensitive to. 2131 */ 2132int flushwithdeps = 0; 2133SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2134 0, "Number of buffers flushed with dependecies that require rollbacks"); 2135static int 2136flushbufqueues(int flushdeps) 2137{ 2138 struct thread *td = curthread; 2139 struct vnode *vp; 2140 struct mount *mp; 2141 struct buf *bp; 2142 int hasdeps; 2143 2144 mtx_lock(&bqlock); 2145 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2146 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2147 continue; 2148 KASSERT((bp->b_flags & B_DELWRI), 2149 ("unexpected clean buffer %p", bp)); 2150 VI_LOCK(bp->b_vp); 2151 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 2152 VI_UNLOCK(bp->b_vp); 2153 BUF_UNLOCK(bp); 2154 continue; 2155 } 2156 VI_UNLOCK(bp->b_vp); 2157 if (bp->b_flags & B_INVAL) { 2158 bremfreel(bp); 2159 mtx_unlock(&bqlock); 2160 brelse(bp); 2161 return (1); 2162 } 2163 2164 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2165 if (flushdeps == 0) { 2166 BUF_UNLOCK(bp); 2167 continue; 2168 } 2169 hasdeps = 1; 2170 } else 2171 hasdeps = 0; 2172 /* 2173 * We must hold the lock on a vnode before writing 2174 * one of its buffers. Otherwise we may confuse, or 2175 * in the case of a snapshot vnode, deadlock the 2176 * system. 2177 * 2178 * The lock order here is the reverse of the normal 2179 * of vnode followed by buf lock. This is ok because 2180 * the NOWAIT will prevent deadlock. 2181 */ 2182 vp = bp->b_vp; 2183 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2184 BUF_UNLOCK(bp); 2185 continue; 2186 } 2187 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2188 mtx_unlock(&bqlock); 2189 vfs_bio_awrite(bp); 2190 vn_finished_write(mp); 2191 VOP_UNLOCK(vp, 0, td); 2192 flushwithdeps += hasdeps; 2193 return (1); 2194 } 2195 vn_finished_write(mp); 2196 BUF_UNLOCK(bp); 2197 } 2198 mtx_unlock(&bqlock); 2199 return (0); 2200} 2201 2202/* 2203 * Check to see if a block is currently memory resident. 2204 */ 2205struct buf * 2206incore(struct vnode * vp, daddr_t blkno) 2207{ 2208 struct buf *bp; 2209 2210 int s = splbio(); 2211 VI_LOCK(vp); 2212 bp = gbincore(vp, blkno); 2213 VI_UNLOCK(vp); 2214 splx(s); 2215 return (bp); 2216} 2217 2218/* 2219 * Returns true if no I/O is needed to access the 2220 * associated VM object. This is like incore except 2221 * it also hunts around in the VM system for the data. 2222 */ 2223 2224int 2225inmem(struct vnode * vp, daddr_t blkno) 2226{ 2227 vm_object_t obj; 2228 vm_offset_t toff, tinc, size; 2229 vm_page_t m; 2230 vm_ooffset_t off; 2231 2232 GIANT_REQUIRED; 2233 ASSERT_VOP_LOCKED(vp, "inmem"); 2234 2235 if (incore(vp, blkno)) 2236 return 1; 2237 if (vp->v_mount == NULL) 2238 return 0; 2239 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2240 return 0; 2241 2242 size = PAGE_SIZE; 2243 if (size > vp->v_mount->mnt_stat.f_iosize) 2244 size = vp->v_mount->mnt_stat.f_iosize; 2245 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2246 2247 VM_OBJECT_LOCK(obj); 2248 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2249 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2250 if (!m) 2251 goto notinmem; 2252 tinc = size; 2253 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2254 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2255 if (vm_page_is_valid(m, 2256 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2257 goto notinmem; 2258 } 2259 VM_OBJECT_UNLOCK(obj); 2260 return 1; 2261 2262notinmem: 2263 VM_OBJECT_UNLOCK(obj); 2264 return (0); 2265} 2266 2267/* 2268 * vfs_setdirty: 2269 * 2270 * Sets the dirty range for a buffer based on the status of the dirty 2271 * bits in the pages comprising the buffer. 2272 * 2273 * The range is limited to the size of the buffer. 2274 * 2275 * This routine is primarily used by NFS, but is generalized for the 2276 * B_VMIO case. 2277 */ 2278static void 2279vfs_setdirty(struct buf *bp) 2280{ 2281 int i; 2282 vm_object_t object; 2283 2284 GIANT_REQUIRED; 2285 /* 2286 * Degenerate case - empty buffer 2287 */ 2288 2289 if (bp->b_bufsize == 0) 2290 return; 2291 2292 /* 2293 * We qualify the scan for modified pages on whether the 2294 * object has been flushed yet. The OBJ_WRITEABLE flag 2295 * is not cleared simply by protecting pages off. 2296 */ 2297 2298 if ((bp->b_flags & B_VMIO) == 0) 2299 return; 2300 2301 object = bp->b_pages[0]->object; 2302 VM_OBJECT_LOCK(object); 2303 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2304 printf("Warning: object %p writeable but not mightbedirty\n", object); 2305 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2306 printf("Warning: object %p mightbedirty but not writeable\n", object); 2307 2308 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2309 vm_offset_t boffset; 2310 vm_offset_t eoffset; 2311 2312 vm_page_lock_queues(); 2313 /* 2314 * test the pages to see if they have been modified directly 2315 * by users through the VM system. 2316 */ 2317 for (i = 0; i < bp->b_npages; i++) { 2318 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2319 vm_page_test_dirty(bp->b_pages[i]); 2320 } 2321 2322 /* 2323 * Calculate the encompassing dirty range, boffset and eoffset, 2324 * (eoffset - boffset) bytes. 2325 */ 2326 2327 for (i = 0; i < bp->b_npages; i++) { 2328 if (bp->b_pages[i]->dirty) 2329 break; 2330 } 2331 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2332 2333 for (i = bp->b_npages - 1; i >= 0; --i) { 2334 if (bp->b_pages[i]->dirty) { 2335 break; 2336 } 2337 } 2338 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2339 2340 vm_page_unlock_queues(); 2341 /* 2342 * Fit it to the buffer. 2343 */ 2344 2345 if (eoffset > bp->b_bcount) 2346 eoffset = bp->b_bcount; 2347 2348 /* 2349 * If we have a good dirty range, merge with the existing 2350 * dirty range. 2351 */ 2352 2353 if (boffset < eoffset) { 2354 if (bp->b_dirtyoff > boffset) 2355 bp->b_dirtyoff = boffset; 2356 if (bp->b_dirtyend < eoffset) 2357 bp->b_dirtyend = eoffset; 2358 } 2359 } 2360 VM_OBJECT_UNLOCK(object); 2361} 2362 2363/* 2364 * getblk: 2365 * 2366 * Get a block given a specified block and offset into a file/device. 2367 * The buffers B_DONE bit will be cleared on return, making it almost 2368 * ready for an I/O initiation. B_INVAL may or may not be set on 2369 * return. The caller should clear B_INVAL prior to initiating a 2370 * READ. 2371 * 2372 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2373 * an existing buffer. 2374 * 2375 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2376 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2377 * and then cleared based on the backing VM. If the previous buffer is 2378 * non-0-sized but invalid, B_CACHE will be cleared. 2379 * 2380 * If getblk() must create a new buffer, the new buffer is returned with 2381 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2382 * case it is returned with B_INVAL clear and B_CACHE set based on the 2383 * backing VM. 2384 * 2385 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2386 * B_CACHE bit is clear. 2387 * 2388 * What this means, basically, is that the caller should use B_CACHE to 2389 * determine whether the buffer is fully valid or not and should clear 2390 * B_INVAL prior to issuing a read. If the caller intends to validate 2391 * the buffer by loading its data area with something, the caller needs 2392 * to clear B_INVAL. If the caller does this without issuing an I/O, 2393 * the caller should set B_CACHE ( as an optimization ), else the caller 2394 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2395 * a write attempt or if it was a successfull read. If the caller 2396 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2397 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2398 */ 2399struct buf * 2400getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2401 int flags) 2402{ 2403 struct buf *bp; 2404 int s; 2405 int error; 2406 ASSERT_VOP_LOCKED(vp, "getblk"); 2407 2408 if (size > MAXBSIZE) 2409 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2410 2411 s = splbio(); 2412loop: 2413 /* 2414 * Block if we are low on buffers. Certain processes are allowed 2415 * to completely exhaust the buffer cache. 2416 * 2417 * If this check ever becomes a bottleneck it may be better to 2418 * move it into the else, when gbincore() fails. At the moment 2419 * it isn't a problem. 2420 * 2421 * XXX remove if 0 sections (clean this up after its proven) 2422 */ 2423 if (numfreebuffers == 0) { 2424 if (curthread == PCPU_GET(idlethread)) 2425 return NULL; 2426 mtx_lock(&nblock); 2427 needsbuffer |= VFS_BIO_NEED_ANY; 2428 mtx_unlock(&nblock); 2429 } 2430 2431 VI_LOCK(vp); 2432 if ((bp = gbincore(vp, blkno))) { 2433 int lockflags; 2434 /* 2435 * Buffer is in-core. If the buffer is not busy, it must 2436 * be on a queue. 2437 */ 2438 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2439 2440 if (flags & GB_LOCK_NOWAIT) 2441 lockflags |= LK_NOWAIT; 2442 2443 error = BUF_TIMELOCK(bp, lockflags, 2444 VI_MTX(vp), "getblk", slpflag, slptimeo); 2445 2446 /* 2447 * If we slept and got the lock we have to restart in case 2448 * the buffer changed identities. 2449 */ 2450 if (error == ENOLCK) 2451 goto loop; 2452 /* We timed out or were interrupted. */ 2453 else if (error) 2454 return (NULL); 2455 2456 /* 2457 * The buffer is locked. B_CACHE is cleared if the buffer is 2458 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2459 * and for a VMIO buffer B_CACHE is adjusted according to the 2460 * backing VM cache. 2461 */ 2462 if (bp->b_flags & B_INVAL) 2463 bp->b_flags &= ~B_CACHE; 2464 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2465 bp->b_flags |= B_CACHE; 2466 bremfree(bp); 2467 2468 /* 2469 * check for size inconsistancies for non-VMIO case. 2470 */ 2471 2472 if (bp->b_bcount != size) { 2473 if ((bp->b_flags & B_VMIO) == 0 || 2474 (size > bp->b_kvasize)) { 2475 if (bp->b_flags & B_DELWRI) { 2476 bp->b_flags |= B_NOCACHE; 2477 BUF_WRITE(bp); 2478 } else { 2479 if ((bp->b_flags & B_VMIO) && 2480 (LIST_FIRST(&bp->b_dep) == NULL)) { 2481 bp->b_flags |= B_RELBUF; 2482 brelse(bp); 2483 } else { 2484 bp->b_flags |= B_NOCACHE; 2485 BUF_WRITE(bp); 2486 } 2487 } 2488 goto loop; 2489 } 2490 } 2491 2492 /* 2493 * If the size is inconsistant in the VMIO case, we can resize 2494 * the buffer. This might lead to B_CACHE getting set or 2495 * cleared. If the size has not changed, B_CACHE remains 2496 * unchanged from its previous state. 2497 */ 2498 2499 if (bp->b_bcount != size) 2500 allocbuf(bp, size); 2501 2502 KASSERT(bp->b_offset != NOOFFSET, 2503 ("getblk: no buffer offset")); 2504 2505 /* 2506 * A buffer with B_DELWRI set and B_CACHE clear must 2507 * be committed before we can return the buffer in 2508 * order to prevent the caller from issuing a read 2509 * ( due to B_CACHE not being set ) and overwriting 2510 * it. 2511 * 2512 * Most callers, including NFS and FFS, need this to 2513 * operate properly either because they assume they 2514 * can issue a read if B_CACHE is not set, or because 2515 * ( for example ) an uncached B_DELWRI might loop due 2516 * to softupdates re-dirtying the buffer. In the latter 2517 * case, B_CACHE is set after the first write completes, 2518 * preventing further loops. 2519 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2520 * above while extending the buffer, we cannot allow the 2521 * buffer to remain with B_CACHE set after the write 2522 * completes or it will represent a corrupt state. To 2523 * deal with this we set B_NOCACHE to scrap the buffer 2524 * after the write. 2525 * 2526 * We might be able to do something fancy, like setting 2527 * B_CACHE in bwrite() except if B_DELWRI is already set, 2528 * so the below call doesn't set B_CACHE, but that gets real 2529 * confusing. This is much easier. 2530 */ 2531 2532 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2533 bp->b_flags |= B_NOCACHE; 2534 BUF_WRITE(bp); 2535 goto loop; 2536 } 2537 2538 splx(s); 2539 bp->b_flags &= ~B_DONE; 2540 } else { 2541 int bsize, maxsize, vmio; 2542 off_t offset; 2543 2544 /* 2545 * Buffer is not in-core, create new buffer. The buffer 2546 * returned by getnewbuf() is locked. Note that the returned 2547 * buffer is also considered valid (not marked B_INVAL). 2548 */ 2549 VI_UNLOCK(vp); 2550 /* 2551 * If the user does not want us to create the buffer, bail out 2552 * here. 2553 */ 2554 if (flags & GB_NOCREAT) { 2555 splx(s); 2556 return NULL; 2557 } 2558 if (vn_isdisk(vp, NULL)) 2559 bsize = DEV_BSIZE; 2560 else if (vp->v_mountedhere) 2561 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2562 else if (vp->v_mount) 2563 bsize = vp->v_mount->mnt_stat.f_iosize; 2564 else 2565 bsize = size; 2566 2567 offset = blkno * bsize; 2568 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2569 (vp->v_vflag & VV_OBJBUF); 2570 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2571 maxsize = imax(maxsize, bsize); 2572 2573 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2574 if (slpflag || slptimeo) { 2575 splx(s); 2576 return NULL; 2577 } 2578 goto loop; 2579 } 2580 2581 /* 2582 * This code is used to make sure that a buffer is not 2583 * created while the getnewbuf routine is blocked. 2584 * This can be a problem whether the vnode is locked or not. 2585 * If the buffer is created out from under us, we have to 2586 * throw away the one we just created. There is now window 2587 * race because we are safely running at splbio() from the 2588 * point of the duplicate buffer creation through to here, 2589 * and we've locked the buffer. 2590 * 2591 * Note: this must occur before we associate the buffer 2592 * with the vp especially considering limitations in 2593 * the splay tree implementation when dealing with duplicate 2594 * lblkno's. 2595 */ 2596 VI_LOCK(vp); 2597 if (gbincore(vp, blkno)) { 2598 VI_UNLOCK(vp); 2599 bp->b_flags |= B_INVAL; 2600 brelse(bp); 2601 goto loop; 2602 } 2603 2604 /* 2605 * Insert the buffer into the hash, so that it can 2606 * be found by incore. 2607 */ 2608 bp->b_blkno = bp->b_lblkno = blkno; 2609 bp->b_offset = offset; 2610 2611 bgetvp(vp, bp); 2612 VI_UNLOCK(vp); 2613 2614 /* 2615 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2616 * buffer size starts out as 0, B_CACHE will be set by 2617 * allocbuf() for the VMIO case prior to it testing the 2618 * backing store for validity. 2619 */ 2620 2621 if (vmio) { 2622 bp->b_flags |= B_VMIO; 2623#if defined(VFS_BIO_DEBUG) 2624 if (vp->v_type != VREG) 2625 printf("getblk: vmioing file type %d???\n", vp->v_type); 2626#endif 2627 VOP_GETVOBJECT(vp, &bp->b_object); 2628 } else { 2629 bp->b_flags &= ~B_VMIO; 2630 bp->b_object = NULL; 2631 } 2632 2633 allocbuf(bp, size); 2634 2635 splx(s); 2636 bp->b_flags &= ~B_DONE; 2637 } 2638 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2639 return (bp); 2640} 2641 2642/* 2643 * Get an empty, disassociated buffer of given size. The buffer is initially 2644 * set to B_INVAL. 2645 */ 2646struct buf * 2647geteblk(int size) 2648{ 2649 struct buf *bp; 2650 int s; 2651 int maxsize; 2652 2653 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2654 2655 s = splbio(); 2656 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2657 continue; 2658 splx(s); 2659 allocbuf(bp, size); 2660 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2661 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2662 return (bp); 2663} 2664 2665 2666/* 2667 * This code constitutes the buffer memory from either anonymous system 2668 * memory (in the case of non-VMIO operations) or from an associated 2669 * VM object (in the case of VMIO operations). This code is able to 2670 * resize a buffer up or down. 2671 * 2672 * Note that this code is tricky, and has many complications to resolve 2673 * deadlock or inconsistant data situations. Tread lightly!!! 2674 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2675 * the caller. Calling this code willy nilly can result in the loss of data. 2676 * 2677 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2678 * B_CACHE for the non-VMIO case. 2679 */ 2680 2681int 2682allocbuf(struct buf *bp, int size) 2683{ 2684 int newbsize, mbsize; 2685 int i; 2686 2687 GIANT_REQUIRED; 2688 2689 if (BUF_REFCNT(bp) == 0) 2690 panic("allocbuf: buffer not busy"); 2691 2692 if (bp->b_kvasize < size) 2693 panic("allocbuf: buffer too small"); 2694 2695 if ((bp->b_flags & B_VMIO) == 0) { 2696 caddr_t origbuf; 2697 int origbufsize; 2698 /* 2699 * Just get anonymous memory from the kernel. Don't 2700 * mess with B_CACHE. 2701 */ 2702 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2703 if (bp->b_flags & B_MALLOC) 2704 newbsize = mbsize; 2705 else 2706 newbsize = round_page(size); 2707 2708 if (newbsize < bp->b_bufsize) { 2709 /* 2710 * malloced buffers are not shrunk 2711 */ 2712 if (bp->b_flags & B_MALLOC) { 2713 if (newbsize) { 2714 bp->b_bcount = size; 2715 } else { 2716 free(bp->b_data, M_BIOBUF); 2717 if (bp->b_bufsize) { 2718 atomic_subtract_int( 2719 &bufmallocspace, 2720 bp->b_bufsize); 2721 bufspacewakeup(); 2722 bp->b_bufsize = 0; 2723 } 2724 bp->b_saveaddr = bp->b_kvabase; 2725 bp->b_data = bp->b_saveaddr; 2726 bp->b_bcount = 0; 2727 bp->b_flags &= ~B_MALLOC; 2728 } 2729 return 1; 2730 } 2731 vm_hold_free_pages( 2732 bp, 2733 (vm_offset_t) bp->b_data + newbsize, 2734 (vm_offset_t) bp->b_data + bp->b_bufsize); 2735 } else if (newbsize > bp->b_bufsize) { 2736 /* 2737 * We only use malloced memory on the first allocation. 2738 * and revert to page-allocated memory when the buffer 2739 * grows. 2740 */ 2741 /* 2742 * There is a potential smp race here that could lead 2743 * to bufmallocspace slightly passing the max. It 2744 * is probably extremely rare and not worth worrying 2745 * over. 2746 */ 2747 if ( (bufmallocspace < maxbufmallocspace) && 2748 (bp->b_bufsize == 0) && 2749 (mbsize <= PAGE_SIZE/2)) { 2750 2751 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2752 bp->b_bufsize = mbsize; 2753 bp->b_bcount = size; 2754 bp->b_flags |= B_MALLOC; 2755 atomic_add_int(&bufmallocspace, mbsize); 2756 return 1; 2757 } 2758 origbuf = NULL; 2759 origbufsize = 0; 2760 /* 2761 * If the buffer is growing on its other-than-first allocation, 2762 * then we revert to the page-allocation scheme. 2763 */ 2764 if (bp->b_flags & B_MALLOC) { 2765 origbuf = bp->b_data; 2766 origbufsize = bp->b_bufsize; 2767 bp->b_data = bp->b_kvabase; 2768 if (bp->b_bufsize) { 2769 atomic_subtract_int(&bufmallocspace, 2770 bp->b_bufsize); 2771 bufspacewakeup(); 2772 bp->b_bufsize = 0; 2773 } 2774 bp->b_flags &= ~B_MALLOC; 2775 newbsize = round_page(newbsize); 2776 } 2777 vm_hold_load_pages( 2778 bp, 2779 (vm_offset_t) bp->b_data + bp->b_bufsize, 2780 (vm_offset_t) bp->b_data + newbsize); 2781 if (origbuf) { 2782 bcopy(origbuf, bp->b_data, origbufsize); 2783 free(origbuf, M_BIOBUF); 2784 } 2785 } 2786 } else { 2787 int desiredpages; 2788 2789 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2790 desiredpages = (size == 0) ? 0 : 2791 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2792 2793 if (bp->b_flags & B_MALLOC) 2794 panic("allocbuf: VMIO buffer can't be malloced"); 2795 /* 2796 * Set B_CACHE initially if buffer is 0 length or will become 2797 * 0-length. 2798 */ 2799 if (size == 0 || bp->b_bufsize == 0) 2800 bp->b_flags |= B_CACHE; 2801 2802 if (newbsize < bp->b_bufsize) { 2803 /* 2804 * DEV_BSIZE aligned new buffer size is less then the 2805 * DEV_BSIZE aligned existing buffer size. Figure out 2806 * if we have to remove any pages. 2807 */ 2808 if (desiredpages < bp->b_npages) { 2809 vm_page_t m; 2810 2811 vm_page_lock_queues(); 2812 for (i = desiredpages; i < bp->b_npages; i++) { 2813 /* 2814 * the page is not freed here -- it 2815 * is the responsibility of 2816 * vnode_pager_setsize 2817 */ 2818 m = bp->b_pages[i]; 2819 KASSERT(m != bogus_page, 2820 ("allocbuf: bogus page found")); 2821 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2822 vm_page_lock_queues(); 2823 2824 bp->b_pages[i] = NULL; 2825 vm_page_unwire(m, 0); 2826 } 2827 vm_page_unlock_queues(); 2828 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2829 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2830 bp->b_npages = desiredpages; 2831 } 2832 } else if (size > bp->b_bcount) { 2833 /* 2834 * We are growing the buffer, possibly in a 2835 * byte-granular fashion. 2836 */ 2837 struct vnode *vp; 2838 vm_object_t obj; 2839 vm_offset_t toff; 2840 vm_offset_t tinc; 2841 2842 /* 2843 * Step 1, bring in the VM pages from the object, 2844 * allocating them if necessary. We must clear 2845 * B_CACHE if these pages are not valid for the 2846 * range covered by the buffer. 2847 */ 2848 2849 vp = bp->b_vp; 2850 obj = bp->b_object; 2851 2852 VM_OBJECT_LOCK(obj); 2853 while (bp->b_npages < desiredpages) { 2854 vm_page_t m; 2855 vm_pindex_t pi; 2856 2857 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2858 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2859 /* 2860 * note: must allocate system pages 2861 * since blocking here could intefere 2862 * with paging I/O, no matter which 2863 * process we are. 2864 */ 2865 m = vm_page_alloc(obj, pi, 2866 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2867 if (m == NULL) { 2868 atomic_add_int(&vm_pageout_deficit, 2869 desiredpages - bp->b_npages); 2870 VM_OBJECT_UNLOCK(obj); 2871 VM_WAIT; 2872 VM_OBJECT_LOCK(obj); 2873 } else { 2874 vm_page_lock_queues(); 2875 vm_page_wakeup(m); 2876 vm_page_unlock_queues(); 2877 bp->b_flags &= ~B_CACHE; 2878 bp->b_pages[bp->b_npages] = m; 2879 ++bp->b_npages; 2880 } 2881 continue; 2882 } 2883 2884 /* 2885 * We found a page. If we have to sleep on it, 2886 * retry because it might have gotten freed out 2887 * from under us. 2888 * 2889 * We can only test PG_BUSY here. Blocking on 2890 * m->busy might lead to a deadlock: 2891 * 2892 * vm_fault->getpages->cluster_read->allocbuf 2893 * 2894 */ 2895 vm_page_lock_queues(); 2896 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2897 continue; 2898 2899 /* 2900 * We have a good page. Should we wakeup the 2901 * page daemon? 2902 */ 2903 if ((curproc != pageproc) && 2904 ((m->queue - m->pc) == PQ_CACHE) && 2905 ((cnt.v_free_count + cnt.v_cache_count) < 2906 (cnt.v_free_min + cnt.v_cache_min))) { 2907 pagedaemon_wakeup(); 2908 } 2909 vm_page_flag_clear(m, PG_ZERO); 2910 vm_page_wire(m); 2911 vm_page_unlock_queues(); 2912 bp->b_pages[bp->b_npages] = m; 2913 ++bp->b_npages; 2914 } 2915 2916 /* 2917 * Step 2. We've loaded the pages into the buffer, 2918 * we have to figure out if we can still have B_CACHE 2919 * set. Note that B_CACHE is set according to the 2920 * byte-granular range ( bcount and size ), new the 2921 * aligned range ( newbsize ). 2922 * 2923 * The VM test is against m->valid, which is DEV_BSIZE 2924 * aligned. Needless to say, the validity of the data 2925 * needs to also be DEV_BSIZE aligned. Note that this 2926 * fails with NFS if the server or some other client 2927 * extends the file's EOF. If our buffer is resized, 2928 * B_CACHE may remain set! XXX 2929 */ 2930 2931 toff = bp->b_bcount; 2932 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2933 2934 while ((bp->b_flags & B_CACHE) && toff < size) { 2935 vm_pindex_t pi; 2936 2937 if (tinc > (size - toff)) 2938 tinc = size - toff; 2939 2940 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2941 PAGE_SHIFT; 2942 2943 vfs_buf_test_cache( 2944 bp, 2945 bp->b_offset, 2946 toff, 2947 tinc, 2948 bp->b_pages[pi] 2949 ); 2950 toff += tinc; 2951 tinc = PAGE_SIZE; 2952 } 2953 VM_OBJECT_UNLOCK(obj); 2954 2955 /* 2956 * Step 3, fixup the KVM pmap. Remember that 2957 * bp->b_data is relative to bp->b_offset, but 2958 * bp->b_offset may be offset into the first page. 2959 */ 2960 2961 bp->b_data = (caddr_t) 2962 trunc_page((vm_offset_t)bp->b_data); 2963 pmap_qenter( 2964 (vm_offset_t)bp->b_data, 2965 bp->b_pages, 2966 bp->b_npages 2967 ); 2968 2969 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2970 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2971 } 2972 } 2973 if (newbsize < bp->b_bufsize) 2974 bufspacewakeup(); 2975 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2976 bp->b_bcount = size; /* requested buffer size */ 2977 return 1; 2978} 2979 2980void 2981biodone(struct bio *bp) 2982{ 2983 mtx_lock(&bdonelock); 2984 bp->bio_flags |= BIO_DONE; 2985 if (bp->bio_done == NULL) 2986 wakeup(bp); 2987 mtx_unlock(&bdonelock); 2988 if (bp->bio_done != NULL) 2989 bp->bio_done(bp); 2990} 2991 2992/* 2993 * Wait for a BIO to finish. 2994 * 2995 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2996 * case is not yet clear. 2997 */ 2998int 2999biowait(struct bio *bp, const char *wchan) 3000{ 3001 3002 mtx_lock(&bdonelock); 3003 while ((bp->bio_flags & BIO_DONE) == 0) 3004 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3005 mtx_unlock(&bdonelock); 3006 if (bp->bio_error != 0) 3007 return (bp->bio_error); 3008 if (!(bp->bio_flags & BIO_ERROR)) 3009 return (0); 3010 return (EIO); 3011} 3012 3013void 3014biofinish(struct bio *bp, struct devstat *stat, int error) 3015{ 3016 3017 if (error) { 3018 bp->bio_error = error; 3019 bp->bio_flags |= BIO_ERROR; 3020 } 3021 if (stat != NULL) 3022 devstat_end_transaction_bio(stat, bp); 3023 biodone(bp); 3024} 3025 3026/* 3027 * bufwait: 3028 * 3029 * Wait for buffer I/O completion, returning error status. The buffer 3030 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3031 * error and cleared. 3032 */ 3033int 3034bufwait(register struct buf * bp) 3035{ 3036 int s; 3037 3038 s = splbio(); 3039 if (bp->b_iocmd == BIO_READ) 3040 bwait(bp, PRIBIO, "biord"); 3041 else 3042 bwait(bp, PRIBIO, "biowr"); 3043 splx(s); 3044 if (bp->b_flags & B_EINTR) { 3045 bp->b_flags &= ~B_EINTR; 3046 return (EINTR); 3047 } 3048 if (bp->b_ioflags & BIO_ERROR) { 3049 return (bp->b_error ? bp->b_error : EIO); 3050 } else { 3051 return (0); 3052 } 3053} 3054 3055 /* 3056 * Call back function from struct bio back up to struct buf. 3057 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 3058 */ 3059static void 3060bufdonebio(struct bio *bp) 3061{ 3062 3063 /* Device drivers may or may not hold giant, hold it here. */ 3064 mtx_lock(&Giant); 3065 bufdone(bp->bio_caller2); 3066 mtx_unlock(&Giant); 3067} 3068 3069void 3070dev_strategy(struct buf *bp) 3071{ 3072 struct cdevsw *csw; 3073 3074 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3075 panic("b_iocmd botch"); 3076 bp->b_io.bio_done = bufdonebio; 3077 bp->b_io.bio_caller2 = bp; 3078 csw = devsw(bp->b_io.bio_dev); 3079 KASSERT(bp->b_io.bio_dev->si_refcount > 0, 3080 ("dev_strategy on un-referenced dev_t (%s)", 3081 devtoname(bp->b_io.bio_dev))); 3082 cdevsw_ref(csw); 3083 (*devsw(bp->b_io.bio_dev)->d_strategy)(&bp->b_io); 3084 cdevsw_rel(csw); 3085} 3086 3087/* 3088 * bufdone: 3089 * 3090 * Finish I/O on a buffer, optionally calling a completion function. 3091 * This is usually called from an interrupt so process blocking is 3092 * not allowed. 3093 * 3094 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3095 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3096 * assuming B_INVAL is clear. 3097 * 3098 * For the VMIO case, we set B_CACHE if the op was a read and no 3099 * read error occured, or if the op was a write. B_CACHE is never 3100 * set if the buffer is invalid or otherwise uncacheable. 3101 * 3102 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3103 * initiator to leave B_INVAL set to brelse the buffer out of existance 3104 * in the biodone routine. 3105 */ 3106void 3107bufdone(struct buf *bp) 3108{ 3109 int s; 3110 void (*biodone)(struct buf *); 3111 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 VM_OBJECT_LOCK(obj); 3168#if defined(VFS_BIO_DEBUG) 3169 if (obj->paging_in_progress < bp->b_npages) { 3170 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3171 obj->paging_in_progress, bp->b_npages); 3172 } 3173#endif 3174 3175 /* 3176 * Set B_CACHE if the op was a normal read and no error 3177 * occured. B_CACHE is set for writes in the b*write() 3178 * routines. 3179 */ 3180 iosize = bp->b_bcount - bp->b_resid; 3181 if (bp->b_iocmd == BIO_READ && 3182 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3183 !(bp->b_ioflags & BIO_ERROR)) { 3184 bp->b_flags |= B_CACHE; 3185 } 3186 vm_page_lock_queues(); 3187 for (i = 0; i < bp->b_npages; i++) { 3188 int bogusflag = 0; 3189 int resid; 3190 3191 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3192 if (resid > iosize) 3193 resid = iosize; 3194 3195 /* 3196 * cleanup bogus pages, restoring the originals 3197 */ 3198 m = bp->b_pages[i]; 3199 if (m == bogus_page) { 3200 bogusflag = 1; 3201 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3202 if (m == NULL) 3203 panic("biodone: page disappeared!"); 3204 bp->b_pages[i] = m; 3205 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3206 } 3207#if defined(VFS_BIO_DEBUG) 3208 if (OFF_TO_IDX(foff) != m->pindex) { 3209 printf( 3210"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3211 (intmax_t)foff, (uintmax_t)m->pindex); 3212 } 3213#endif 3214 3215 /* 3216 * In the write case, the valid and clean bits are 3217 * already changed correctly ( see bdwrite() ), so we 3218 * only need to do this here in the read case. 3219 */ 3220 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3221 vfs_page_set_valid(bp, foff, i, m); 3222 } 3223 vm_page_flag_clear(m, PG_ZERO); 3224 3225 /* 3226 * when debugging new filesystems or buffer I/O methods, this 3227 * is the most common error that pops up. if you see this, you 3228 * have not set the page busy flag correctly!!! 3229 */ 3230 if (m->busy == 0) { 3231 printf("biodone: page busy < 0, " 3232 "pindex: %d, foff: 0x(%x,%x), " 3233 "resid: %d, index: %d\n", 3234 (int) m->pindex, (int)(foff >> 32), 3235 (int) foff & 0xffffffff, resid, i); 3236 if (!vn_isdisk(vp, NULL)) 3237 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3238 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3239 (intmax_t) bp->b_lblkno, 3240 bp->b_flags, bp->b_npages); 3241 else 3242 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3243 (intmax_t) bp->b_lblkno, 3244 bp->b_flags, bp->b_npages); 3245 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3246 (u_long)m->valid, (u_long)m->dirty, 3247 m->wire_count); 3248 panic("biodone: page busy < 0\n"); 3249 } 3250 vm_page_io_finish(m); 3251 vm_object_pip_subtract(obj, 1); 3252 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3253 iosize -= resid; 3254 } 3255 vm_page_unlock_queues(); 3256 vm_object_pip_wakeupn(obj, 0); 3257 VM_OBJECT_UNLOCK(obj); 3258 } 3259 3260 /* 3261 * For asynchronous completions, release the buffer now. The brelse 3262 * will do a wakeup there if necessary - so no need to do a wakeup 3263 * here in the async case. The sync case always needs to do a wakeup. 3264 */ 3265 3266 if (bp->b_flags & B_ASYNC) { 3267 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3268 brelse(bp); 3269 else 3270 bqrelse(bp); 3271 } else { 3272 bdone(bp); 3273 } 3274 splx(s); 3275} 3276 3277/* 3278 * This routine is called in lieu of iodone in the case of 3279 * incomplete I/O. This keeps the busy status for pages 3280 * consistant. 3281 */ 3282void 3283vfs_unbusy_pages(struct buf * bp) 3284{ 3285 int i; 3286 3287 runningbufwakeup(bp); 3288 if (bp->b_flags & B_VMIO) { 3289 vm_object_t obj; 3290 3291 obj = bp->b_object; 3292 VM_OBJECT_LOCK(obj); 3293 vm_page_lock_queues(); 3294 for (i = 0; i < bp->b_npages; i++) { 3295 vm_page_t m = bp->b_pages[i]; 3296 3297 if (m == bogus_page) { 3298 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3299 if (!m) { 3300 panic("vfs_unbusy_pages: page missing\n"); 3301 } 3302 bp->b_pages[i] = m; 3303 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3304 } 3305 vm_object_pip_subtract(obj, 1); 3306 vm_page_flag_clear(m, PG_ZERO); 3307 vm_page_io_finish(m); 3308 } 3309 vm_page_unlock_queues(); 3310 vm_object_pip_wakeupn(obj, 0); 3311 VM_OBJECT_UNLOCK(obj); 3312 } 3313} 3314 3315/* 3316 * vfs_page_set_valid: 3317 * 3318 * Set the valid bits in a page based on the supplied offset. The 3319 * range is restricted to the buffer's size. 3320 * 3321 * This routine is typically called after a read completes. 3322 */ 3323static void 3324vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3325{ 3326 vm_ooffset_t soff, eoff; 3327 3328 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3329 /* 3330 * Start and end offsets in buffer. eoff - soff may not cross a 3331 * page boundry or cross the end of the buffer. The end of the 3332 * buffer, in this case, is our file EOF, not the allocation size 3333 * of the buffer. 3334 */ 3335 soff = off; 3336 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3337 if (eoff > bp->b_offset + bp->b_bcount) 3338 eoff = bp->b_offset + bp->b_bcount; 3339 3340 /* 3341 * Set valid range. This is typically the entire buffer and thus the 3342 * entire page. 3343 */ 3344 if (eoff > soff) { 3345 vm_page_set_validclean( 3346 m, 3347 (vm_offset_t) (soff & PAGE_MASK), 3348 (vm_offset_t) (eoff - soff) 3349 ); 3350 } 3351} 3352 3353/* 3354 * This routine is called before a device strategy routine. 3355 * It is used to tell the VM system that paging I/O is in 3356 * progress, and treat the pages associated with the buffer 3357 * almost as being PG_BUSY. Also the object paging_in_progress 3358 * flag is handled to make sure that the object doesn't become 3359 * inconsistant. 3360 * 3361 * Since I/O has not been initiated yet, certain buffer flags 3362 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3363 * and should be ignored. 3364 */ 3365void 3366vfs_busy_pages(struct buf * bp, int clear_modify) 3367{ 3368 int i, bogus; 3369 3370 if (bp->b_flags & B_VMIO) { 3371 vm_object_t obj; 3372 vm_ooffset_t foff; 3373 3374 obj = bp->b_object; 3375 foff = bp->b_offset; 3376 KASSERT(bp->b_offset != NOOFFSET, 3377 ("vfs_busy_pages: no buffer offset")); 3378 vfs_setdirty(bp); 3379 VM_OBJECT_LOCK(obj); 3380retry: 3381 vm_page_lock_queues(); 3382 for (i = 0; i < bp->b_npages; i++) { 3383 vm_page_t m = bp->b_pages[i]; 3384 3385 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3386 goto retry; 3387 } 3388 bogus = 0; 3389 for (i = 0; i < bp->b_npages; i++) { 3390 vm_page_t m = bp->b_pages[i]; 3391 3392 vm_page_flag_clear(m, PG_ZERO); 3393 if ((bp->b_flags & B_CLUSTER) == 0) { 3394 vm_object_pip_add(obj, 1); 3395 vm_page_io_start(m); 3396 } 3397 /* 3398 * When readying a buffer for a read ( i.e 3399 * clear_modify == 0 ), it is important to do 3400 * bogus_page replacement for valid pages in 3401 * partially instantiated buffers. Partially 3402 * instantiated buffers can, in turn, occur when 3403 * reconstituting a buffer from its VM backing store 3404 * base. We only have to do this if B_CACHE is 3405 * clear ( which causes the I/O to occur in the 3406 * first place ). The replacement prevents the read 3407 * I/O from overwriting potentially dirty VM-backed 3408 * pages. XXX bogus page replacement is, uh, bogus. 3409 * It may not work properly with small-block devices. 3410 * We need to find a better way. 3411 */ 3412 pmap_remove_all(m); 3413 if (clear_modify) 3414 vfs_page_set_valid(bp, foff, i, m); 3415 else if (m->valid == VM_PAGE_BITS_ALL && 3416 (bp->b_flags & B_CACHE) == 0) { 3417 bp->b_pages[i] = bogus_page; 3418 bogus++; 3419 } 3420 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3421 } 3422 vm_page_unlock_queues(); 3423 VM_OBJECT_UNLOCK(obj); 3424 if (bogus) 3425 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3426 } 3427} 3428 3429/* 3430 * Tell the VM system that the pages associated with this buffer 3431 * are clean. This is used for delayed writes where the data is 3432 * going to go to disk eventually without additional VM intevention. 3433 * 3434 * Note that while we only really need to clean through to b_bcount, we 3435 * just go ahead and clean through to b_bufsize. 3436 */ 3437static void 3438vfs_clean_pages(struct buf * bp) 3439{ 3440 int i; 3441 3442 if (bp->b_flags & B_VMIO) { 3443 vm_ooffset_t foff; 3444 3445 foff = bp->b_offset; 3446 KASSERT(bp->b_offset != NOOFFSET, 3447 ("vfs_clean_pages: no buffer offset")); 3448 VM_OBJECT_LOCK(bp->b_object); 3449 vm_page_lock_queues(); 3450 for (i = 0; i < bp->b_npages; i++) { 3451 vm_page_t m = bp->b_pages[i]; 3452 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3453 vm_ooffset_t eoff = noff; 3454 3455 if (eoff > bp->b_offset + bp->b_bufsize) 3456 eoff = bp->b_offset + bp->b_bufsize; 3457 vfs_page_set_valid(bp, foff, i, m); 3458 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3459 foff = noff; 3460 } 3461 vm_page_unlock_queues(); 3462 VM_OBJECT_UNLOCK(bp->b_object); 3463 } 3464} 3465 3466/* 3467 * vfs_bio_set_validclean: 3468 * 3469 * Set the range within the buffer to valid and clean. The range is 3470 * relative to the beginning of the buffer, b_offset. Note that b_offset 3471 * itself may be offset from the beginning of the first page. 3472 * 3473 */ 3474 3475void 3476vfs_bio_set_validclean(struct buf *bp, int base, int size) 3477{ 3478 if (bp->b_flags & B_VMIO) { 3479 int i; 3480 int n; 3481 3482 /* 3483 * Fixup base to be relative to beginning of first page. 3484 * Set initial n to be the maximum number of bytes in the 3485 * first page that can be validated. 3486 */ 3487 3488 base += (bp->b_offset & PAGE_MASK); 3489 n = PAGE_SIZE - (base & PAGE_MASK); 3490 3491 VM_OBJECT_LOCK(bp->b_object); 3492 vm_page_lock_queues(); 3493 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3494 vm_page_t m = bp->b_pages[i]; 3495 3496 if (n > size) 3497 n = size; 3498 3499 vm_page_set_validclean(m, base & PAGE_MASK, n); 3500 base += n; 3501 size -= n; 3502 n = PAGE_SIZE; 3503 } 3504 vm_page_unlock_queues(); 3505 VM_OBJECT_UNLOCK(bp->b_object); 3506 } 3507} 3508 3509/* 3510 * vfs_bio_clrbuf: 3511 * 3512 * clear a buffer. This routine essentially fakes an I/O, so we need 3513 * to clear BIO_ERROR and B_INVAL. 3514 * 3515 * Note that while we only theoretically need to clear through b_bcount, 3516 * we go ahead and clear through b_bufsize. 3517 */ 3518 3519void 3520vfs_bio_clrbuf(struct buf *bp) 3521{ 3522 int i, mask = 0; 3523 caddr_t sa, ea; 3524 3525 GIANT_REQUIRED; 3526 3527 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3528 bp->b_flags &= ~B_INVAL; 3529 bp->b_ioflags &= ~BIO_ERROR; 3530 VM_OBJECT_LOCK(bp->b_object); 3531 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3532 (bp->b_offset & PAGE_MASK) == 0) { 3533 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3534 if (bp->b_pages[0] != bogus_page) 3535 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3536 if ((bp->b_pages[0]->valid & mask) == mask) 3537 goto unlock; 3538 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3539 ((bp->b_pages[0]->valid & mask) == 0)) { 3540 bzero(bp->b_data, bp->b_bufsize); 3541 bp->b_pages[0]->valid |= mask; 3542 goto unlock; 3543 } 3544 } 3545 ea = sa = bp->b_data; 3546 for(i=0;i<bp->b_npages;i++,sa=ea) { 3547 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3548 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3549 ea = (caddr_t)(vm_offset_t)ulmin( 3550 (u_long)(vm_offset_t)ea, 3551 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3552 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3553 if (bp->b_pages[i] != bogus_page) 3554 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3555 if ((bp->b_pages[i]->valid & mask) == mask) 3556 continue; 3557 if ((bp->b_pages[i]->valid & mask) == 0) { 3558 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3559 bzero(sa, ea - sa); 3560 } 3561 } else { 3562 for (; sa < ea; sa += DEV_BSIZE, j++) { 3563 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3564 (bp->b_pages[i]->valid & (1<<j)) == 0) 3565 bzero(sa, DEV_BSIZE); 3566 } 3567 } 3568 bp->b_pages[i]->valid |= mask; 3569 vm_page_lock_queues(); 3570 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3571 vm_page_unlock_queues(); 3572 } 3573unlock: 3574 VM_OBJECT_UNLOCK(bp->b_object); 3575 bp->b_resid = 0; 3576 } else { 3577 clrbuf(bp); 3578 } 3579} 3580 3581/* 3582 * vm_hold_load_pages and vm_hold_free_pages get pages into 3583 * a buffers address space. The pages are anonymous and are 3584 * not associated with a file object. 3585 */ 3586static void 3587vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3588{ 3589 vm_offset_t pg; 3590 vm_page_t p; 3591 int index; 3592 3593 to = round_page(to); 3594 from = round_page(from); 3595 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3596 3597 VM_OBJECT_LOCK(kernel_object); 3598 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3599tryagain: 3600 /* 3601 * note: must allocate system pages since blocking here 3602 * could intefere with paging I/O, no matter which 3603 * process we are. 3604 */ 3605 p = vm_page_alloc(kernel_object, 3606 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3607 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3608 if (!p) { 3609 atomic_add_int(&vm_pageout_deficit, 3610 (to - pg) >> PAGE_SHIFT); 3611 VM_OBJECT_UNLOCK(kernel_object); 3612 VM_WAIT; 3613 VM_OBJECT_LOCK(kernel_object); 3614 goto tryagain; 3615 } 3616 p->valid = VM_PAGE_BITS_ALL; 3617 pmap_qenter(pg, &p, 1); 3618 bp->b_pages[index] = p; 3619 vm_page_lock_queues(); 3620 vm_page_wakeup(p); 3621 vm_page_unlock_queues(); 3622 } 3623 VM_OBJECT_UNLOCK(kernel_object); 3624 bp->b_npages = index; 3625} 3626 3627/* Return pages associated with this buf to the vm system */ 3628static void 3629vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3630{ 3631 vm_offset_t pg; 3632 vm_page_t p; 3633 int index, newnpages; 3634 3635 GIANT_REQUIRED; 3636 3637 from = round_page(from); 3638 to = round_page(to); 3639 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3640 3641 VM_OBJECT_LOCK(kernel_object); 3642 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3643 p = bp->b_pages[index]; 3644 if (p && (index < bp->b_npages)) { 3645 if (p->busy) { 3646 printf( 3647 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3648 (intmax_t)bp->b_blkno, 3649 (intmax_t)bp->b_lblkno); 3650 } 3651 bp->b_pages[index] = NULL; 3652 pmap_qremove(pg, 1); 3653 vm_page_lock_queues(); 3654 vm_page_busy(p); 3655 vm_page_unwire(p, 0); 3656 vm_page_free(p); 3657 vm_page_unlock_queues(); 3658 } 3659 } 3660 VM_OBJECT_UNLOCK(kernel_object); 3661 bp->b_npages = newnpages; 3662} 3663 3664/* 3665 * Map an IO request into kernel virtual address space. 3666 * 3667 * All requests are (re)mapped into kernel VA space. 3668 * Notice that we use b_bufsize for the size of the buffer 3669 * to be mapped. b_bcount might be modified by the driver. 3670 * 3671 * Note that even if the caller determines that the address space should 3672 * be valid, a race or a smaller-file mapped into a larger space may 3673 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3674 * check the return value. 3675 */ 3676int 3677vmapbuf(struct buf *bp) 3678{ 3679 caddr_t addr, kva; 3680 vm_prot_t prot; 3681 int pidx, i; 3682 struct vm_page *m; 3683 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3684 3685 GIANT_REQUIRED; 3686 3687 if (bp->b_bufsize < 0) 3688 return (-1); 3689 prot = (bp->b_iocmd == BIO_READ) ? VM_PROT_READ | VM_PROT_WRITE : 3690 VM_PROT_READ; 3691 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3692 addr < bp->b_data + bp->b_bufsize; 3693 addr += PAGE_SIZE, pidx++) { 3694 /* 3695 * Do the vm_fault if needed; do the copy-on-write thing 3696 * when reading stuff off device into memory. 3697 * 3698 * NOTE! Must use pmap_extract() because addr may be in 3699 * the userland address space, and kextract is only guarenteed 3700 * to work for the kernland address space (see: sparc64 port). 3701 */ 3702retry: 3703 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3704 prot) < 0) { 3705 vm_page_lock_queues(); 3706 for (i = 0; i < pidx; ++i) { 3707 vm_page_unhold(bp->b_pages[i]); 3708 bp->b_pages[i] = NULL; 3709 } 3710 vm_page_unlock_queues(); 3711 return(-1); 3712 } 3713 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3714 if (m == NULL) 3715 goto retry; 3716 bp->b_pages[pidx] = m; 3717 } 3718 if (pidx > btoc(MAXPHYS)) 3719 panic("vmapbuf: mapped more than MAXPHYS"); 3720 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3721 3722 kva = bp->b_saveaddr; 3723 bp->b_npages = pidx; 3724 bp->b_saveaddr = bp->b_data; 3725 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3726 return(0); 3727} 3728 3729/* 3730 * Free the io map PTEs associated with this IO operation. 3731 * We also invalidate the TLB entries and restore the original b_addr. 3732 */ 3733void 3734vunmapbuf(struct buf *bp) 3735{ 3736 int pidx; 3737 int npages; 3738 3739 npages = bp->b_npages; 3740 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3741 npages); 3742 vm_page_lock_queues(); 3743 for (pidx = 0; pidx < npages; pidx++) 3744 vm_page_unhold(bp->b_pages[pidx]); 3745 vm_page_unlock_queues(); 3746 3747 bp->b_data = bp->b_saveaddr; 3748} 3749 3750void 3751bdone(struct buf *bp) 3752{ 3753 mtx_lock(&bdonelock); 3754 bp->b_flags |= B_DONE; 3755 wakeup(bp); 3756 mtx_unlock(&bdonelock); 3757} 3758 3759void 3760bwait(struct buf *bp, u_char pri, const char *wchan) 3761{ 3762 mtx_lock(&bdonelock); 3763 while ((bp->b_flags & B_DONE) == 0) 3764 msleep(bp, &bdonelock, pri, wchan, 0); 3765 mtx_unlock(&bdonelock); 3766} 3767 3768#include "opt_ddb.h" 3769#ifdef DDB 3770#include <ddb/ddb.h> 3771 3772/* DDB command to show buffer data */ 3773DB_SHOW_COMMAND(buffer, db_show_buffer) 3774{ 3775 /* get args */ 3776 struct buf *bp = (struct buf *)addr; 3777 3778 if (!have_addr) { 3779 db_printf("usage: show buffer <addr>\n"); 3780 return; 3781 } 3782 3783 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3784 db_printf( 3785 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3786 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd\n", 3787 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3788 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3789 (intmax_t)bp->b_blkno); 3790 if (bp->b_npages) { 3791 int i; 3792 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3793 for (i = 0; i < bp->b_npages; i++) { 3794 vm_page_t m; 3795 m = bp->b_pages[i]; 3796 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3797 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3798 if ((i + 1) < bp->b_npages) 3799 db_printf(","); 3800 } 3801 db_printf("\n"); 3802 } 3803} 3804#endif /* DDB */ 3805