vfs_bio.c revision 122537
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 122537 2003-11-12 08:01:40Z mckusick $"); 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 struct vnode *vp; 1306 1307 vp = bp->b_vp; 1308 obj = bp->b_object; 1309 1310 /* 1311 * Get the base offset and length of the buffer. Note that 1312 * in the VMIO case if the buffer block size is not 1313 * page-aligned then b_data pointer may not be page-aligned. 1314 * But our b_pages[] array *IS* page aligned. 1315 * 1316 * block sizes less then DEV_BSIZE (usually 512) are not 1317 * supported due to the page granularity bits (m->valid, 1318 * m->dirty, etc...). 1319 * 1320 * See man buf(9) for more information 1321 */ 1322 resid = bp->b_bufsize; 1323 foff = bp->b_offset; 1324 VM_OBJECT_LOCK(obj); 1325 for (i = 0; i < bp->b_npages; i++) { 1326 int had_bogus = 0; 1327 1328 m = bp->b_pages[i]; 1329 vm_page_lock_queues(); 1330 vm_page_flag_clear(m, PG_ZERO); 1331 vm_page_unlock_queues(); 1332 1333 /* 1334 * If we hit a bogus page, fixup *all* the bogus pages 1335 * now. 1336 */ 1337 if (m == bogus_page) { 1338 poff = OFF_TO_IDX(bp->b_offset); 1339 had_bogus = 1; 1340 1341 for (j = i; j < bp->b_npages; j++) { 1342 vm_page_t mtmp; 1343 mtmp = bp->b_pages[j]; 1344 if (mtmp == bogus_page) { 1345 mtmp = vm_page_lookup(obj, poff + j); 1346 if (!mtmp) { 1347 panic("brelse: page missing\n"); 1348 } 1349 bp->b_pages[j] = mtmp; 1350 } 1351 } 1352 1353 if ((bp->b_flags & B_INVAL) == 0) { 1354 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1355 } 1356 m = bp->b_pages[i]; 1357 } 1358 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1359 int poffset = foff & PAGE_MASK; 1360 int presid = resid > (PAGE_SIZE - poffset) ? 1361 (PAGE_SIZE - poffset) : resid; 1362 1363 KASSERT(presid >= 0, ("brelse: extra page")); 1364 vm_page_lock_queues(); 1365 vm_page_set_invalid(m, poffset, presid); 1366 vm_page_unlock_queues(); 1367 if (had_bogus) 1368 printf("avoided corruption bug in bogus_page/brelse code\n"); 1369 } 1370 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1371 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1372 } 1373 VM_OBJECT_UNLOCK(obj); 1374 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1375 vfs_vmio_release(bp); 1376 1377 } else if (bp->b_flags & B_VMIO) { 1378 1379 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1380 vfs_vmio_release(bp); 1381 } 1382 1383 } 1384 1385 if (bp->b_qindex != QUEUE_NONE) 1386 panic("brelse: free buffer onto another queue???"); 1387 if (BUF_REFCNT(bp) > 1) { 1388 /* do not release to free list */ 1389 BUF_UNLOCK(bp); 1390 splx(s); 1391 return; 1392 } 1393 1394 /* enqueue */ 1395 mtx_lock(&bqlock); 1396 1397 /* buffers with no memory */ 1398 if (bp->b_bufsize == 0) { 1399 bp->b_flags |= B_INVAL; 1400 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1401 if (bp->b_vflags & BV_BKGRDINPROG) 1402 panic("losing buffer 1"); 1403 if (bp->b_kvasize) { 1404 bp->b_qindex = QUEUE_EMPTYKVA; 1405 } else { 1406 bp->b_qindex = QUEUE_EMPTY; 1407 } 1408 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1409 bp->b_dev = NODEV; 1410 /* buffers with junk contents */ 1411 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1412 (bp->b_ioflags & BIO_ERROR)) { 1413 bp->b_flags |= B_INVAL; 1414 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1415 if (bp->b_vflags & BV_BKGRDINPROG) 1416 panic("losing buffer 2"); 1417 bp->b_qindex = QUEUE_CLEAN; 1418 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1419 bp->b_dev = NODEV; 1420 /* remaining buffers */ 1421 } else { 1422 if (bp->b_flags & B_DELWRI) 1423 bp->b_qindex = QUEUE_DIRTY; 1424 else 1425 bp->b_qindex = QUEUE_CLEAN; 1426 if (bp->b_flags & B_AGE) 1427 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1428 else 1429 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1430 } 1431 mtx_unlock(&bqlock); 1432 1433 /* 1434 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1435 * placed the buffer on the correct queue. We must also disassociate 1436 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1437 * find it. 1438 */ 1439 if (bp->b_flags & B_INVAL) { 1440 if (bp->b_flags & B_DELWRI) 1441 bundirty(bp); 1442 if (bp->b_vp) 1443 brelvp(bp); 1444 } 1445 1446 /* 1447 * Fixup numfreebuffers count. The bp is on an appropriate queue 1448 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1449 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1450 * if B_INVAL is set ). 1451 */ 1452 1453 if (!(bp->b_flags & B_DELWRI)) 1454 bufcountwakeup(); 1455 1456 /* 1457 * Something we can maybe free or reuse 1458 */ 1459 if (bp->b_bufsize || bp->b_kvasize) 1460 bufspacewakeup(); 1461 1462 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1463 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1464 panic("brelse: not dirty"); 1465 /* unlock */ 1466 BUF_UNLOCK(bp); 1467 splx(s); 1468} 1469 1470/* 1471 * Release a buffer back to the appropriate queue but do not try to free 1472 * it. The buffer is expected to be used again soon. 1473 * 1474 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1475 * biodone() to requeue an async I/O on completion. It is also used when 1476 * known good buffers need to be requeued but we think we may need the data 1477 * again soon. 1478 * 1479 * XXX we should be able to leave the B_RELBUF hint set on completion. 1480 */ 1481void 1482bqrelse(struct buf * bp) 1483{ 1484 int s; 1485 1486 s = splbio(); 1487 1488 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1489 1490 if (bp->b_qindex != QUEUE_NONE) 1491 panic("bqrelse: free buffer onto another queue???"); 1492 if (BUF_REFCNT(bp) > 1) { 1493 /* do not release to free list */ 1494 BUF_UNLOCK(bp); 1495 splx(s); 1496 return; 1497 } 1498 mtx_lock(&bqlock); 1499 /* buffers with stale but valid contents */ 1500 if (bp->b_flags & B_DELWRI) { 1501 bp->b_qindex = QUEUE_DIRTY; 1502 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1503 } else { 1504 /* 1505 * XXX This lock may not be necessary since BKGRDINPROG 1506 * cannot be set while we hold the buf lock, it can only be 1507 * cleared if it is already pending. 1508 */ 1509 VI_LOCK(bp->b_vp); 1510 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1511 VI_UNLOCK(bp->b_vp); 1512 bp->b_qindex = QUEUE_CLEAN; 1513 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1514 b_freelist); 1515 } else { 1516 /* 1517 * We are too low on memory, we have to try to free 1518 * the buffer (most importantly: the wired pages 1519 * making up its backing store) *now*. 1520 */ 1521 VI_UNLOCK(bp->b_vp); 1522 mtx_unlock(&bqlock); 1523 splx(s); 1524 brelse(bp); 1525 return; 1526 } 1527 } 1528 mtx_unlock(&bqlock); 1529 1530 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1531 bufcountwakeup(); 1532 1533 /* 1534 * Something we can maybe free or reuse. 1535 */ 1536 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1537 bufspacewakeup(); 1538 1539 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1540 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1541 panic("bqrelse: not dirty"); 1542 /* unlock */ 1543 BUF_UNLOCK(bp); 1544 splx(s); 1545} 1546 1547/* Give pages used by the bp back to the VM system (where possible) */ 1548static void 1549vfs_vmio_release(bp) 1550 struct buf *bp; 1551{ 1552 int i; 1553 vm_page_t m; 1554 1555 GIANT_REQUIRED; 1556 if (bp->b_object != NULL) 1557 VM_OBJECT_LOCK(bp->b_object); 1558 vm_page_lock_queues(); 1559 for (i = 0; i < bp->b_npages; i++) { 1560 m = bp->b_pages[i]; 1561 bp->b_pages[i] = NULL; 1562 /* 1563 * In order to keep page LRU ordering consistent, put 1564 * everything on the inactive queue. 1565 */ 1566 vm_page_unwire(m, 0); 1567 /* 1568 * We don't mess with busy pages, it is 1569 * the responsibility of the process that 1570 * busied the pages to deal with them. 1571 */ 1572 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1573 continue; 1574 1575 if (m->wire_count == 0) { 1576 vm_page_flag_clear(m, PG_ZERO); 1577 /* 1578 * Might as well free the page if we can and it has 1579 * no valid data. We also free the page if the 1580 * buffer was used for direct I/O 1581 */ 1582 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1583 m->hold_count == 0) { 1584 vm_page_busy(m); 1585 pmap_remove_all(m); 1586 vm_page_free(m); 1587 } else if (bp->b_flags & B_DIRECT) { 1588 vm_page_try_to_free(m); 1589 } else if (vm_page_count_severe()) { 1590 vm_page_try_to_cache(m); 1591 } 1592 } 1593 } 1594 vm_page_unlock_queues(); 1595 if (bp->b_object != NULL) 1596 VM_OBJECT_UNLOCK(bp->b_object); 1597 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1598 1599 if (bp->b_bufsize) { 1600 bufspacewakeup(); 1601 bp->b_bufsize = 0; 1602 } 1603 bp->b_npages = 0; 1604 bp->b_flags &= ~B_VMIO; 1605 if (bp->b_vp) 1606 brelvp(bp); 1607} 1608 1609/* 1610 * Check to see if a block at a particular lbn is available for a clustered 1611 * write. 1612 */ 1613static int 1614vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1615{ 1616 struct buf *bpa; 1617 int match; 1618 1619 match = 0; 1620 1621 /* If the buf isn't in core skip it */ 1622 if ((bpa = gbincore(vp, lblkno)) == NULL) 1623 return (0); 1624 1625 /* If the buf is busy we don't want to wait for it */ 1626 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1627 return (0); 1628 1629 /* Only cluster with valid clusterable delayed write buffers */ 1630 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1631 (B_DELWRI | B_CLUSTEROK)) 1632 goto done; 1633 1634 if (bpa->b_bufsize != size) 1635 goto done; 1636 1637 /* 1638 * Check to see if it is in the expected place on disk and that the 1639 * block has been mapped. 1640 */ 1641 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1642 match = 1; 1643done: 1644 BUF_UNLOCK(bpa); 1645 return (match); 1646} 1647 1648/* 1649 * vfs_bio_awrite: 1650 * 1651 * Implement clustered async writes for clearing out B_DELWRI buffers. 1652 * This is much better then the old way of writing only one buffer at 1653 * a time. Note that we may not be presented with the buffers in the 1654 * correct order, so we search for the cluster in both directions. 1655 */ 1656int 1657vfs_bio_awrite(struct buf * bp) 1658{ 1659 int i; 1660 int j; 1661 daddr_t lblkno = bp->b_lblkno; 1662 struct vnode *vp = bp->b_vp; 1663 int s; 1664 int ncl; 1665 int nwritten; 1666 int size; 1667 int maxcl; 1668 1669 s = splbio(); 1670 /* 1671 * right now we support clustered writing only to regular files. If 1672 * we find a clusterable block we could be in the middle of a cluster 1673 * rather then at the beginning. 1674 */ 1675 if ((vp->v_type == VREG) && 1676 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1677 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1678 1679 size = vp->v_mount->mnt_stat.f_iosize; 1680 maxcl = MAXPHYS / size; 1681 1682 VI_LOCK(vp); 1683 for (i = 1; i < maxcl; i++) 1684 if (vfs_bio_clcheck(vp, size, lblkno + i, 1685 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1686 break; 1687 1688 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1689 if (vfs_bio_clcheck(vp, size, lblkno - j, 1690 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1691 break; 1692 1693 VI_UNLOCK(vp); 1694 --j; 1695 ncl = i + j; 1696 /* 1697 * this is a possible cluster write 1698 */ 1699 if (ncl != 1) { 1700 BUF_UNLOCK(bp); 1701 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1702 splx(s); 1703 return nwritten; 1704 } 1705 } 1706 1707 bremfree(bp); 1708 bp->b_flags |= B_ASYNC; 1709 1710 splx(s); 1711 /* 1712 * default (old) behavior, writing out only one block 1713 * 1714 * XXX returns b_bufsize instead of b_bcount for nwritten? 1715 */ 1716 nwritten = bp->b_bufsize; 1717 (void) BUF_WRITE(bp); 1718 1719 return nwritten; 1720} 1721 1722/* 1723 * getnewbuf: 1724 * 1725 * Find and initialize a new buffer header, freeing up existing buffers 1726 * in the bufqueues as necessary. The new buffer is returned locked. 1727 * 1728 * Important: B_INVAL is not set. If the caller wishes to throw the 1729 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1730 * 1731 * We block if: 1732 * We have insufficient buffer headers 1733 * We have insufficient buffer space 1734 * buffer_map is too fragmented ( space reservation fails ) 1735 * If we have to flush dirty buffers ( but we try to avoid this ) 1736 * 1737 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1738 * Instead we ask the buf daemon to do it for us. We attempt to 1739 * avoid piecemeal wakeups of the pageout daemon. 1740 */ 1741 1742static struct buf * 1743getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1744{ 1745 struct buf *bp; 1746 struct buf *nbp; 1747 int defrag = 0; 1748 int nqindex; 1749 static int flushingbufs; 1750 1751 GIANT_REQUIRED; 1752 1753 /* 1754 * We can't afford to block since we might be holding a vnode lock, 1755 * which may prevent system daemons from running. We deal with 1756 * low-memory situations by proactively returning memory and running 1757 * async I/O rather then sync I/O. 1758 */ 1759 1760 atomic_add_int(&getnewbufcalls, 1); 1761 atomic_subtract_int(&getnewbufrestarts, 1); 1762restart: 1763 atomic_add_int(&getnewbufrestarts, 1); 1764 1765 /* 1766 * Setup for scan. If we do not have enough free buffers, 1767 * we setup a degenerate case that immediately fails. Note 1768 * that if we are specially marked process, we are allowed to 1769 * dip into our reserves. 1770 * 1771 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1772 * 1773 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1774 * However, there are a number of cases (defragging, reusing, ...) 1775 * where we cannot backup. 1776 */ 1777 mtx_lock(&bqlock); 1778 nqindex = QUEUE_EMPTYKVA; 1779 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1780 1781 if (nbp == NULL) { 1782 /* 1783 * If no EMPTYKVA buffers and we are either 1784 * defragging or reusing, locate a CLEAN buffer 1785 * to free or reuse. If bufspace useage is low 1786 * skip this step so we can allocate a new buffer. 1787 */ 1788 if (defrag || bufspace >= lobufspace) { 1789 nqindex = QUEUE_CLEAN; 1790 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1791 } 1792 1793 /* 1794 * If we could not find or were not allowed to reuse a 1795 * CLEAN buffer, check to see if it is ok to use an EMPTY 1796 * buffer. We can only use an EMPTY buffer if allocating 1797 * its KVA would not otherwise run us out of buffer space. 1798 */ 1799 if (nbp == NULL && defrag == 0 && 1800 bufspace + maxsize < hibufspace) { 1801 nqindex = QUEUE_EMPTY; 1802 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1803 } 1804 } 1805 1806 /* 1807 * Run scan, possibly freeing data and/or kva mappings on the fly 1808 * depending. 1809 */ 1810 1811 while ((bp = nbp) != NULL) { 1812 int qindex = nqindex; 1813 1814 /* 1815 * Calculate next bp ( we can only use it if we do not block 1816 * or do other fancy things ). 1817 */ 1818 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1819 switch(qindex) { 1820 case QUEUE_EMPTY: 1821 nqindex = QUEUE_EMPTYKVA; 1822 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1823 break; 1824 /* FALLTHROUGH */ 1825 case QUEUE_EMPTYKVA: 1826 nqindex = QUEUE_CLEAN; 1827 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1828 break; 1829 /* FALLTHROUGH */ 1830 case QUEUE_CLEAN: 1831 /* 1832 * nbp is NULL. 1833 */ 1834 break; 1835 } 1836 } 1837 if (bp->b_vp) { 1838 VI_LOCK(bp->b_vp); 1839 if (bp->b_vflags & BV_BKGRDINPROG) { 1840 VI_UNLOCK(bp->b_vp); 1841 continue; 1842 } 1843 VI_UNLOCK(bp->b_vp); 1844 } 1845 1846 /* 1847 * Sanity Checks 1848 */ 1849 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1850 1851 /* 1852 * Note: we no longer distinguish between VMIO and non-VMIO 1853 * buffers. 1854 */ 1855 1856 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1857 1858 /* 1859 * If we are defragging then we need a buffer with 1860 * b_kvasize != 0. XXX this situation should no longer 1861 * occur, if defrag is non-zero the buffer's b_kvasize 1862 * should also be non-zero at this point. XXX 1863 */ 1864 if (defrag && bp->b_kvasize == 0) { 1865 printf("Warning: defrag empty buffer %p\n", bp); 1866 continue; 1867 } 1868 1869 /* 1870 * Start freeing the bp. This is somewhat involved. nbp 1871 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1872 */ 1873 1874 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1875 panic("getnewbuf: locked buf"); 1876 bremfreel(bp); 1877 mtx_unlock(&bqlock); 1878 1879 if (qindex == QUEUE_CLEAN) { 1880 if (bp->b_flags & B_VMIO) { 1881 bp->b_flags &= ~B_ASYNC; 1882 vfs_vmio_release(bp); 1883 } 1884 if (bp->b_vp) 1885 brelvp(bp); 1886 } 1887 1888 /* 1889 * NOTE: nbp is now entirely invalid. We can only restart 1890 * the scan from this point on. 1891 * 1892 * Get the rest of the buffer freed up. b_kva* is still 1893 * valid after this operation. 1894 */ 1895 1896 if (bp->b_rcred != NOCRED) { 1897 crfree(bp->b_rcred); 1898 bp->b_rcred = NOCRED; 1899 } 1900 if (bp->b_wcred != NOCRED) { 1901 crfree(bp->b_wcred); 1902 bp->b_wcred = NOCRED; 1903 } 1904 if (LIST_FIRST(&bp->b_dep) != NULL) 1905 buf_deallocate(bp); 1906 if (bp->b_vflags & BV_BKGRDINPROG) 1907 panic("losing buffer 3"); 1908 1909 if (bp->b_bufsize) 1910 allocbuf(bp, 0); 1911 1912 bp->b_flags = 0; 1913 bp->b_ioflags = 0; 1914 bp->b_xflags = 0; 1915 bp->b_vflags = 0; 1916 bp->b_dev = NODEV; 1917 bp->b_vp = NULL; 1918 bp->b_blkno = bp->b_lblkno = 0; 1919 bp->b_offset = NOOFFSET; 1920 bp->b_iodone = 0; 1921 bp->b_error = 0; 1922 bp->b_resid = 0; 1923 bp->b_bcount = 0; 1924 bp->b_npages = 0; 1925 bp->b_dirtyoff = bp->b_dirtyend = 0; 1926 bp->b_magic = B_MAGIC_BIO; 1927 bp->b_op = &buf_ops_bio; 1928 bp->b_object = NULL; 1929 1930 LIST_INIT(&bp->b_dep); 1931 1932 /* 1933 * If we are defragging then free the buffer. 1934 */ 1935 if (defrag) { 1936 bp->b_flags |= B_INVAL; 1937 bfreekva(bp); 1938 brelse(bp); 1939 defrag = 0; 1940 goto restart; 1941 } 1942 1943 /* 1944 * If we are overcomitted then recover the buffer and its 1945 * KVM space. This occurs in rare situations when multiple 1946 * processes are blocked in getnewbuf() or allocbuf(). 1947 */ 1948 if (bufspace >= hibufspace) 1949 flushingbufs = 1; 1950 if (flushingbufs && bp->b_kvasize != 0) { 1951 bp->b_flags |= B_INVAL; 1952 bfreekva(bp); 1953 brelse(bp); 1954 goto restart; 1955 } 1956 if (bufspace < lobufspace) 1957 flushingbufs = 0; 1958 break; 1959 } 1960 1961 /* 1962 * If we exhausted our list, sleep as appropriate. We may have to 1963 * wakeup various daemons and write out some dirty buffers. 1964 * 1965 * Generally we are sleeping due to insufficient buffer space. 1966 */ 1967 1968 if (bp == NULL) { 1969 int flags; 1970 char *waitmsg; 1971 1972 mtx_unlock(&bqlock); 1973 if (defrag) { 1974 flags = VFS_BIO_NEED_BUFSPACE; 1975 waitmsg = "nbufkv"; 1976 } else if (bufspace >= hibufspace) { 1977 waitmsg = "nbufbs"; 1978 flags = VFS_BIO_NEED_BUFSPACE; 1979 } else { 1980 waitmsg = "newbuf"; 1981 flags = VFS_BIO_NEED_ANY; 1982 } 1983 1984 bd_speedup(); /* heeeelp */ 1985 1986 mtx_lock(&nblock); 1987 needsbuffer |= flags; 1988 while (needsbuffer & flags) { 1989 if (msleep(&needsbuffer, &nblock, 1990 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1991 mtx_unlock(&nblock); 1992 return (NULL); 1993 } 1994 } 1995 mtx_unlock(&nblock); 1996 } else { 1997 /* 1998 * We finally have a valid bp. We aren't quite out of the 1999 * woods, we still have to reserve kva space. In order 2000 * to keep fragmentation sane we only allocate kva in 2001 * BKVASIZE chunks. 2002 */ 2003 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2004 2005 if (maxsize != bp->b_kvasize) { 2006 vm_offset_t addr = 0; 2007 2008 bfreekva(bp); 2009 2010 if (vm_map_findspace(buffer_map, 2011 vm_map_min(buffer_map), maxsize, &addr)) { 2012 /* 2013 * Uh oh. Buffer map is to fragmented. We 2014 * must defragment the map. 2015 */ 2016 atomic_add_int(&bufdefragcnt, 1); 2017 defrag = 1; 2018 bp->b_flags |= B_INVAL; 2019 brelse(bp); 2020 goto restart; 2021 } 2022 if (addr) { 2023 vm_map_insert(buffer_map, NULL, 0, 2024 addr, addr + maxsize, 2025 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2026 2027 bp->b_kvabase = (caddr_t) addr; 2028 bp->b_kvasize = maxsize; 2029 atomic_add_int(&bufspace, bp->b_kvasize); 2030 atomic_add_int(&bufreusecnt, 1); 2031 } 2032 } 2033 bp->b_saveaddr = bp->b_kvabase; 2034 bp->b_data = bp->b_saveaddr; 2035 } 2036 return(bp); 2037} 2038 2039/* 2040 * buf_daemon: 2041 * 2042 * buffer flushing daemon. Buffers are normally flushed by the 2043 * update daemon but if it cannot keep up this process starts to 2044 * take the load in an attempt to prevent getnewbuf() from blocking. 2045 */ 2046 2047static struct kproc_desc buf_kp = { 2048 "bufdaemon", 2049 buf_daemon, 2050 &bufdaemonproc 2051}; 2052SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2053 2054static void 2055buf_daemon() 2056{ 2057 int s; 2058 2059 mtx_lock(&Giant); 2060 2061 /* 2062 * This process needs to be suspended prior to shutdown sync. 2063 */ 2064 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2065 SHUTDOWN_PRI_LAST); 2066 2067 /* 2068 * This process is allowed to take the buffer cache to the limit 2069 */ 2070 s = splbio(); 2071 mtx_lock(&bdlock); 2072 2073 for (;;) { 2074 bd_request = 0; 2075 mtx_unlock(&bdlock); 2076 2077 kthread_suspend_check(bufdaemonproc); 2078 2079 /* 2080 * Do the flush. Limit the amount of in-transit I/O we 2081 * allow to build up, otherwise we would completely saturate 2082 * the I/O system. Wakeup any waiting processes before we 2083 * normally would so they can run in parallel with our drain. 2084 */ 2085 while (numdirtybuffers > lodirtybuffers) { 2086 if (flushbufqueues(0) == 0) { 2087 /* 2088 * Could not find any buffers without rollback 2089 * dependencies, so just write the first one 2090 * in the hopes of eventually making progress. 2091 */ 2092 flushbufqueues(1); 2093 break; 2094 } 2095 waitrunningbufspace(); 2096 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2097 } 2098 2099 /* 2100 * Only clear bd_request if we have reached our low water 2101 * mark. The buf_daemon normally waits 1 second and 2102 * then incrementally flushes any dirty buffers that have 2103 * built up, within reason. 2104 * 2105 * If we were unable to hit our low water mark and couldn't 2106 * find any flushable buffers, we sleep half a second. 2107 * Otherwise we loop immediately. 2108 */ 2109 mtx_lock(&bdlock); 2110 if (numdirtybuffers <= lodirtybuffers) { 2111 /* 2112 * We reached our low water mark, reset the 2113 * request and sleep until we are needed again. 2114 * The sleep is just so the suspend code works. 2115 */ 2116 bd_request = 0; 2117 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2118 } else { 2119 /* 2120 * We couldn't find any flushable dirty buffers but 2121 * still have too many dirty buffers, we 2122 * have to sleep and try again. (rare) 2123 */ 2124 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2125 } 2126 } 2127} 2128 2129/* 2130 * flushbufqueues: 2131 * 2132 * Try to flush a buffer in the dirty queue. We must be careful to 2133 * free up B_INVAL buffers instead of write them, which NFS is 2134 * particularly sensitive to. 2135 */ 2136int flushwithdeps = 0; 2137SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2138 0, "Number of buffers flushed with dependecies that require rollbacks"); 2139static int 2140flushbufqueues(int flushdeps) 2141{ 2142 struct thread *td = curthread; 2143 struct vnode *vp; 2144 struct mount *mp; 2145 struct buf *bp; 2146 int hasdeps; 2147 2148 mtx_lock(&bqlock); 2149 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2150 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2151 continue; 2152 KASSERT((bp->b_flags & B_DELWRI), 2153 ("unexpected clean buffer %p", bp)); 2154 VI_LOCK(bp->b_vp); 2155 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 2156 VI_UNLOCK(bp->b_vp); 2157 BUF_UNLOCK(bp); 2158 continue; 2159 } 2160 VI_UNLOCK(bp->b_vp); 2161 if (bp->b_flags & B_INVAL) { 2162 bremfreel(bp); 2163 mtx_unlock(&bqlock); 2164 brelse(bp); 2165 return (1); 2166 } 2167 2168 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2169 if (flushdeps == 0) { 2170 BUF_UNLOCK(bp); 2171 continue; 2172 } 2173 hasdeps = 1; 2174 } else 2175 hasdeps = 0; 2176 /* 2177 * We must hold the lock on a vnode before writing 2178 * one of its buffers. Otherwise we may confuse, or 2179 * in the case of a snapshot vnode, deadlock the 2180 * system. 2181 * 2182 * The lock order here is the reverse of the normal 2183 * of vnode followed by buf lock. This is ok because 2184 * the NOWAIT will prevent deadlock. 2185 */ 2186 vp = bp->b_vp; 2187 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2188 BUF_UNLOCK(bp); 2189 continue; 2190 } 2191 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2192 mtx_unlock(&bqlock); 2193 vfs_bio_awrite(bp); 2194 vn_finished_write(mp); 2195 VOP_UNLOCK(vp, 0, td); 2196 flushwithdeps += hasdeps; 2197 return (1); 2198 } 2199 vn_finished_write(mp); 2200 BUF_UNLOCK(bp); 2201 } 2202 mtx_unlock(&bqlock); 2203 return (0); 2204} 2205 2206/* 2207 * Check to see if a block is currently memory resident. 2208 */ 2209struct buf * 2210incore(struct vnode * vp, daddr_t blkno) 2211{ 2212 struct buf *bp; 2213 2214 int s = splbio(); 2215 VI_LOCK(vp); 2216 bp = gbincore(vp, blkno); 2217 VI_UNLOCK(vp); 2218 splx(s); 2219 return (bp); 2220} 2221 2222/* 2223 * Returns true if no I/O is needed to access the 2224 * associated VM object. This is like incore except 2225 * it also hunts around in the VM system for the data. 2226 */ 2227 2228int 2229inmem(struct vnode * vp, daddr_t blkno) 2230{ 2231 vm_object_t obj; 2232 vm_offset_t toff, tinc, size; 2233 vm_page_t m; 2234 vm_ooffset_t off; 2235 2236 GIANT_REQUIRED; 2237 ASSERT_VOP_LOCKED(vp, "inmem"); 2238 2239 if (incore(vp, blkno)) 2240 return 1; 2241 if (vp->v_mount == NULL) 2242 return 0; 2243 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2244 return 0; 2245 2246 size = PAGE_SIZE; 2247 if (size > vp->v_mount->mnt_stat.f_iosize) 2248 size = vp->v_mount->mnt_stat.f_iosize; 2249 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2250 2251 VM_OBJECT_LOCK(obj); 2252 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2253 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2254 if (!m) 2255 goto notinmem; 2256 tinc = size; 2257 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2258 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2259 if (vm_page_is_valid(m, 2260 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2261 goto notinmem; 2262 } 2263 VM_OBJECT_UNLOCK(obj); 2264 return 1; 2265 2266notinmem: 2267 VM_OBJECT_UNLOCK(obj); 2268 return (0); 2269} 2270 2271/* 2272 * vfs_setdirty: 2273 * 2274 * Sets the dirty range for a buffer based on the status of the dirty 2275 * bits in the pages comprising the buffer. 2276 * 2277 * The range is limited to the size of the buffer. 2278 * 2279 * This routine is primarily used by NFS, but is generalized for the 2280 * B_VMIO case. 2281 */ 2282static void 2283vfs_setdirty(struct buf *bp) 2284{ 2285 int i; 2286 vm_object_t object; 2287 2288 GIANT_REQUIRED; 2289 /* 2290 * Degenerate case - empty buffer 2291 */ 2292 2293 if (bp->b_bufsize == 0) 2294 return; 2295 2296 /* 2297 * We qualify the scan for modified pages on whether the 2298 * object has been flushed yet. The OBJ_WRITEABLE flag 2299 * is not cleared simply by protecting pages off. 2300 */ 2301 2302 if ((bp->b_flags & B_VMIO) == 0) 2303 return; 2304 2305 object = bp->b_pages[0]->object; 2306 VM_OBJECT_LOCK(object); 2307 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2308 printf("Warning: object %p writeable but not mightbedirty\n", object); 2309 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2310 printf("Warning: object %p mightbedirty but not writeable\n", object); 2311 2312 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2313 vm_offset_t boffset; 2314 vm_offset_t eoffset; 2315 2316 vm_page_lock_queues(); 2317 /* 2318 * test the pages to see if they have been modified directly 2319 * by users through the VM system. 2320 */ 2321 for (i = 0; i < bp->b_npages; i++) { 2322 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2323 vm_page_test_dirty(bp->b_pages[i]); 2324 } 2325 2326 /* 2327 * Calculate the encompassing dirty range, boffset and eoffset, 2328 * (eoffset - boffset) bytes. 2329 */ 2330 2331 for (i = 0; i < bp->b_npages; i++) { 2332 if (bp->b_pages[i]->dirty) 2333 break; 2334 } 2335 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2336 2337 for (i = bp->b_npages - 1; i >= 0; --i) { 2338 if (bp->b_pages[i]->dirty) { 2339 break; 2340 } 2341 } 2342 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2343 2344 vm_page_unlock_queues(); 2345 /* 2346 * Fit it to the buffer. 2347 */ 2348 2349 if (eoffset > bp->b_bcount) 2350 eoffset = bp->b_bcount; 2351 2352 /* 2353 * If we have a good dirty range, merge with the existing 2354 * dirty range. 2355 */ 2356 2357 if (boffset < eoffset) { 2358 if (bp->b_dirtyoff > boffset) 2359 bp->b_dirtyoff = boffset; 2360 if (bp->b_dirtyend < eoffset) 2361 bp->b_dirtyend = eoffset; 2362 } 2363 } 2364 VM_OBJECT_UNLOCK(object); 2365} 2366 2367/* 2368 * getblk: 2369 * 2370 * Get a block given a specified block and offset into a file/device. 2371 * The buffers B_DONE bit will be cleared on return, making it almost 2372 * ready for an I/O initiation. B_INVAL may or may not be set on 2373 * return. The caller should clear B_INVAL prior to initiating a 2374 * READ. 2375 * 2376 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2377 * an existing buffer. 2378 * 2379 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2380 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2381 * and then cleared based on the backing VM. If the previous buffer is 2382 * non-0-sized but invalid, B_CACHE will be cleared. 2383 * 2384 * If getblk() must create a new buffer, the new buffer is returned with 2385 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2386 * case it is returned with B_INVAL clear and B_CACHE set based on the 2387 * backing VM. 2388 * 2389 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2390 * B_CACHE bit is clear. 2391 * 2392 * What this means, basically, is that the caller should use B_CACHE to 2393 * determine whether the buffer is fully valid or not and should clear 2394 * B_INVAL prior to issuing a read. If the caller intends to validate 2395 * the buffer by loading its data area with something, the caller needs 2396 * to clear B_INVAL. If the caller does this without issuing an I/O, 2397 * the caller should set B_CACHE ( as an optimization ), else the caller 2398 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2399 * a write attempt or if it was a successfull read. If the caller 2400 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2401 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2402 */ 2403struct buf * 2404getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2405 int flags) 2406{ 2407 struct buf *bp; 2408 int s; 2409 int error; 2410 ASSERT_VOP_LOCKED(vp, "getblk"); 2411 2412 if (size > MAXBSIZE) 2413 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2414 2415 s = splbio(); 2416loop: 2417 /* 2418 * Block if we are low on buffers. Certain processes are allowed 2419 * to completely exhaust the buffer cache. 2420 * 2421 * If this check ever becomes a bottleneck it may be better to 2422 * move it into the else, when gbincore() fails. At the moment 2423 * it isn't a problem. 2424 * 2425 * XXX remove if 0 sections (clean this up after its proven) 2426 */ 2427 if (numfreebuffers == 0) { 2428 if (curthread == PCPU_GET(idlethread)) 2429 return NULL; 2430 mtx_lock(&nblock); 2431 needsbuffer |= VFS_BIO_NEED_ANY; 2432 mtx_unlock(&nblock); 2433 } 2434 2435 VI_LOCK(vp); 2436 if ((bp = gbincore(vp, blkno))) { 2437 int lockflags; 2438 /* 2439 * Buffer is in-core. If the buffer is not busy, it must 2440 * be on a queue. 2441 */ 2442 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2443 2444 if (flags & GB_LOCK_NOWAIT) 2445 lockflags |= LK_NOWAIT; 2446 2447 error = BUF_TIMELOCK(bp, lockflags, 2448 VI_MTX(vp), "getblk", slpflag, slptimeo); 2449 2450 /* 2451 * If we slept and got the lock we have to restart in case 2452 * the buffer changed identities. 2453 */ 2454 if (error == ENOLCK) 2455 goto loop; 2456 /* We timed out or were interrupted. */ 2457 else if (error) 2458 return (NULL); 2459 2460 /* 2461 * The buffer is locked. B_CACHE is cleared if the buffer is 2462 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2463 * and for a VMIO buffer B_CACHE is adjusted according to the 2464 * backing VM cache. 2465 */ 2466 if (bp->b_flags & B_INVAL) 2467 bp->b_flags &= ~B_CACHE; 2468 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2469 bp->b_flags |= B_CACHE; 2470 bremfree(bp); 2471 2472 /* 2473 * check for size inconsistancies for non-VMIO case. 2474 */ 2475 2476 if (bp->b_bcount != size) { 2477 if ((bp->b_flags & B_VMIO) == 0 || 2478 (size > bp->b_kvasize)) { 2479 if (bp->b_flags & B_DELWRI) { 2480 bp->b_flags |= B_NOCACHE; 2481 BUF_WRITE(bp); 2482 } else { 2483 if ((bp->b_flags & B_VMIO) && 2484 (LIST_FIRST(&bp->b_dep) == NULL)) { 2485 bp->b_flags |= B_RELBUF; 2486 brelse(bp); 2487 } else { 2488 bp->b_flags |= B_NOCACHE; 2489 BUF_WRITE(bp); 2490 } 2491 } 2492 goto loop; 2493 } 2494 } 2495 2496 /* 2497 * If the size is inconsistant in the VMIO case, we can resize 2498 * the buffer. This might lead to B_CACHE getting set or 2499 * cleared. If the size has not changed, B_CACHE remains 2500 * unchanged from its previous state. 2501 */ 2502 2503 if (bp->b_bcount != size) 2504 allocbuf(bp, size); 2505 2506 KASSERT(bp->b_offset != NOOFFSET, 2507 ("getblk: no buffer offset")); 2508 2509 /* 2510 * A buffer with B_DELWRI set and B_CACHE clear must 2511 * be committed before we can return the buffer in 2512 * order to prevent the caller from issuing a read 2513 * ( due to B_CACHE not being set ) and overwriting 2514 * it. 2515 * 2516 * Most callers, including NFS and FFS, need this to 2517 * operate properly either because they assume they 2518 * can issue a read if B_CACHE is not set, or because 2519 * ( for example ) an uncached B_DELWRI might loop due 2520 * to softupdates re-dirtying the buffer. In the latter 2521 * case, B_CACHE is set after the first write completes, 2522 * preventing further loops. 2523 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2524 * above while extending the buffer, we cannot allow the 2525 * buffer to remain with B_CACHE set after the write 2526 * completes or it will represent a corrupt state. To 2527 * deal with this we set B_NOCACHE to scrap the buffer 2528 * after the write. 2529 * 2530 * We might be able to do something fancy, like setting 2531 * B_CACHE in bwrite() except if B_DELWRI is already set, 2532 * so the below call doesn't set B_CACHE, but that gets real 2533 * confusing. This is much easier. 2534 */ 2535 2536 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2537 bp->b_flags |= B_NOCACHE; 2538 BUF_WRITE(bp); 2539 goto loop; 2540 } 2541 2542 splx(s); 2543 bp->b_flags &= ~B_DONE; 2544 } else { 2545 int bsize, maxsize, vmio; 2546 off_t offset; 2547 2548 /* 2549 * Buffer is not in-core, create new buffer. The buffer 2550 * returned by getnewbuf() is locked. Note that the returned 2551 * buffer is also considered valid (not marked B_INVAL). 2552 */ 2553 VI_UNLOCK(vp); 2554 /* 2555 * If the user does not want us to create the buffer, bail out 2556 * here. 2557 */ 2558 if (flags & GB_NOCREAT) { 2559 splx(s); 2560 return NULL; 2561 } 2562 if (vn_isdisk(vp, NULL)) 2563 bsize = DEV_BSIZE; 2564 else if (vp->v_mountedhere) 2565 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2566 else if (vp->v_mount) 2567 bsize = vp->v_mount->mnt_stat.f_iosize; 2568 else 2569 bsize = size; 2570 2571 offset = blkno * bsize; 2572 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2573 (vp->v_vflag & VV_OBJBUF); 2574 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2575 maxsize = imax(maxsize, bsize); 2576 2577 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2578 if (slpflag || slptimeo) { 2579 splx(s); 2580 return NULL; 2581 } 2582 goto loop; 2583 } 2584 2585 /* 2586 * This code is used to make sure that a buffer is not 2587 * created while the getnewbuf routine is blocked. 2588 * This can be a problem whether the vnode is locked or not. 2589 * If the buffer is created out from under us, we have to 2590 * throw away the one we just created. There is now window 2591 * race because we are safely running at splbio() from the 2592 * point of the duplicate buffer creation through to here, 2593 * and we've locked the buffer. 2594 * 2595 * Note: this must occur before we associate the buffer 2596 * with the vp especially considering limitations in 2597 * the splay tree implementation when dealing with duplicate 2598 * lblkno's. 2599 */ 2600 VI_LOCK(vp); 2601 if (gbincore(vp, blkno)) { 2602 VI_UNLOCK(vp); 2603 bp->b_flags |= B_INVAL; 2604 brelse(bp); 2605 goto loop; 2606 } 2607 2608 /* 2609 * Insert the buffer into the hash, so that it can 2610 * be found by incore. 2611 */ 2612 bp->b_blkno = bp->b_lblkno = blkno; 2613 bp->b_offset = offset; 2614 2615 bgetvp(vp, bp); 2616 VI_UNLOCK(vp); 2617 2618 /* 2619 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2620 * buffer size starts out as 0, B_CACHE will be set by 2621 * allocbuf() for the VMIO case prior to it testing the 2622 * backing store for validity. 2623 */ 2624 2625 if (vmio) { 2626 bp->b_flags |= B_VMIO; 2627#if defined(VFS_BIO_DEBUG) 2628 if (vp->v_type != VREG) 2629 printf("getblk: vmioing file type %d???\n", vp->v_type); 2630#endif 2631 VOP_GETVOBJECT(vp, &bp->b_object); 2632 } else { 2633 bp->b_flags &= ~B_VMIO; 2634 bp->b_object = NULL; 2635 } 2636 2637 allocbuf(bp, size); 2638 2639 splx(s); 2640 bp->b_flags &= ~B_DONE; 2641 } 2642 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2643 return (bp); 2644} 2645 2646/* 2647 * Get an empty, disassociated buffer of given size. The buffer is initially 2648 * set to B_INVAL. 2649 */ 2650struct buf * 2651geteblk(int size) 2652{ 2653 struct buf *bp; 2654 int s; 2655 int maxsize; 2656 2657 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2658 2659 s = splbio(); 2660 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2661 continue; 2662 splx(s); 2663 allocbuf(bp, size); 2664 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2665 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2666 return (bp); 2667} 2668 2669 2670/* 2671 * This code constitutes the buffer memory from either anonymous system 2672 * memory (in the case of non-VMIO operations) or from an associated 2673 * VM object (in the case of VMIO operations). This code is able to 2674 * resize a buffer up or down. 2675 * 2676 * Note that this code is tricky, and has many complications to resolve 2677 * deadlock or inconsistant data situations. Tread lightly!!! 2678 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2679 * the caller. Calling this code willy nilly can result in the loss of data. 2680 * 2681 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2682 * B_CACHE for the non-VMIO case. 2683 */ 2684 2685int 2686allocbuf(struct buf *bp, int size) 2687{ 2688 int newbsize, mbsize; 2689 int i; 2690 2691 GIANT_REQUIRED; 2692 2693 if (BUF_REFCNT(bp) == 0) 2694 panic("allocbuf: buffer not busy"); 2695 2696 if (bp->b_kvasize < size) 2697 panic("allocbuf: buffer too small"); 2698 2699 if ((bp->b_flags & B_VMIO) == 0) { 2700 caddr_t origbuf; 2701 int origbufsize; 2702 /* 2703 * Just get anonymous memory from the kernel. Don't 2704 * mess with B_CACHE. 2705 */ 2706 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2707 if (bp->b_flags & B_MALLOC) 2708 newbsize = mbsize; 2709 else 2710 newbsize = round_page(size); 2711 2712 if (newbsize < bp->b_bufsize) { 2713 /* 2714 * malloced buffers are not shrunk 2715 */ 2716 if (bp->b_flags & B_MALLOC) { 2717 if (newbsize) { 2718 bp->b_bcount = size; 2719 } else { 2720 free(bp->b_data, M_BIOBUF); 2721 if (bp->b_bufsize) { 2722 atomic_subtract_int( 2723 &bufmallocspace, 2724 bp->b_bufsize); 2725 bufspacewakeup(); 2726 bp->b_bufsize = 0; 2727 } 2728 bp->b_saveaddr = bp->b_kvabase; 2729 bp->b_data = bp->b_saveaddr; 2730 bp->b_bcount = 0; 2731 bp->b_flags &= ~B_MALLOC; 2732 } 2733 return 1; 2734 } 2735 vm_hold_free_pages( 2736 bp, 2737 (vm_offset_t) bp->b_data + newbsize, 2738 (vm_offset_t) bp->b_data + bp->b_bufsize); 2739 } else if (newbsize > bp->b_bufsize) { 2740 /* 2741 * We only use malloced memory on the first allocation. 2742 * and revert to page-allocated memory when the buffer 2743 * grows. 2744 */ 2745 /* 2746 * There is a potential smp race here that could lead 2747 * to bufmallocspace slightly passing the max. It 2748 * is probably extremely rare and not worth worrying 2749 * over. 2750 */ 2751 if ( (bufmallocspace < maxbufmallocspace) && 2752 (bp->b_bufsize == 0) && 2753 (mbsize <= PAGE_SIZE/2)) { 2754 2755 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2756 bp->b_bufsize = mbsize; 2757 bp->b_bcount = size; 2758 bp->b_flags |= B_MALLOC; 2759 atomic_add_int(&bufmallocspace, mbsize); 2760 return 1; 2761 } 2762 origbuf = NULL; 2763 origbufsize = 0; 2764 /* 2765 * If the buffer is growing on its other-than-first allocation, 2766 * then we revert to the page-allocation scheme. 2767 */ 2768 if (bp->b_flags & B_MALLOC) { 2769 origbuf = bp->b_data; 2770 origbufsize = bp->b_bufsize; 2771 bp->b_data = bp->b_kvabase; 2772 if (bp->b_bufsize) { 2773 atomic_subtract_int(&bufmallocspace, 2774 bp->b_bufsize); 2775 bufspacewakeup(); 2776 bp->b_bufsize = 0; 2777 } 2778 bp->b_flags &= ~B_MALLOC; 2779 newbsize = round_page(newbsize); 2780 } 2781 vm_hold_load_pages( 2782 bp, 2783 (vm_offset_t) bp->b_data + bp->b_bufsize, 2784 (vm_offset_t) bp->b_data + newbsize); 2785 if (origbuf) { 2786 bcopy(origbuf, bp->b_data, origbufsize); 2787 free(origbuf, M_BIOBUF); 2788 } 2789 } 2790 } else { 2791 int desiredpages; 2792 2793 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2794 desiredpages = (size == 0) ? 0 : 2795 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2796 2797 if (bp->b_flags & B_MALLOC) 2798 panic("allocbuf: VMIO buffer can't be malloced"); 2799 /* 2800 * Set B_CACHE initially if buffer is 0 length or will become 2801 * 0-length. 2802 */ 2803 if (size == 0 || bp->b_bufsize == 0) 2804 bp->b_flags |= B_CACHE; 2805 2806 if (newbsize < bp->b_bufsize) { 2807 /* 2808 * DEV_BSIZE aligned new buffer size is less then the 2809 * DEV_BSIZE aligned existing buffer size. Figure out 2810 * if we have to remove any pages. 2811 */ 2812 if (desiredpages < bp->b_npages) { 2813 vm_page_t m; 2814 2815 vm_page_lock_queues(); 2816 for (i = desiredpages; i < bp->b_npages; i++) { 2817 /* 2818 * the page is not freed here -- it 2819 * is the responsibility of 2820 * vnode_pager_setsize 2821 */ 2822 m = bp->b_pages[i]; 2823 KASSERT(m != bogus_page, 2824 ("allocbuf: bogus page found")); 2825 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2826 vm_page_lock_queues(); 2827 2828 bp->b_pages[i] = NULL; 2829 vm_page_unwire(m, 0); 2830 } 2831 vm_page_unlock_queues(); 2832 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2833 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2834 bp->b_npages = desiredpages; 2835 } 2836 } else if (size > bp->b_bcount) { 2837 /* 2838 * We are growing the buffer, possibly in a 2839 * byte-granular fashion. 2840 */ 2841 struct vnode *vp; 2842 vm_object_t obj; 2843 vm_offset_t toff; 2844 vm_offset_t tinc; 2845 2846 /* 2847 * Step 1, bring in the VM pages from the object, 2848 * allocating them if necessary. We must clear 2849 * B_CACHE if these pages are not valid for the 2850 * range covered by the buffer. 2851 */ 2852 2853 vp = bp->b_vp; 2854 obj = bp->b_object; 2855 2856 VM_OBJECT_LOCK(obj); 2857 while (bp->b_npages < desiredpages) { 2858 vm_page_t m; 2859 vm_pindex_t pi; 2860 2861 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2862 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2863 /* 2864 * note: must allocate system pages 2865 * since blocking here could intefere 2866 * with paging I/O, no matter which 2867 * process we are. 2868 */ 2869 m = vm_page_alloc(obj, pi, 2870 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2871 if (m == NULL) { 2872 atomic_add_int(&vm_pageout_deficit, 2873 desiredpages - bp->b_npages); 2874 VM_OBJECT_UNLOCK(obj); 2875 VM_WAIT; 2876 VM_OBJECT_LOCK(obj); 2877 } else { 2878 vm_page_lock_queues(); 2879 vm_page_wakeup(m); 2880 vm_page_unlock_queues(); 2881 bp->b_flags &= ~B_CACHE; 2882 bp->b_pages[bp->b_npages] = m; 2883 ++bp->b_npages; 2884 } 2885 continue; 2886 } 2887 2888 /* 2889 * We found a page. If we have to sleep on it, 2890 * retry because it might have gotten freed out 2891 * from under us. 2892 * 2893 * We can only test PG_BUSY here. Blocking on 2894 * m->busy might lead to a deadlock: 2895 * 2896 * vm_fault->getpages->cluster_read->allocbuf 2897 * 2898 */ 2899 vm_page_lock_queues(); 2900 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2901 continue; 2902 2903 /* 2904 * We have a good page. Should we wakeup the 2905 * page daemon? 2906 */ 2907 if ((curproc != pageproc) && 2908 ((m->queue - m->pc) == PQ_CACHE) && 2909 ((cnt.v_free_count + cnt.v_cache_count) < 2910 (cnt.v_free_min + cnt.v_cache_min))) { 2911 pagedaemon_wakeup(); 2912 } 2913 vm_page_flag_clear(m, PG_ZERO); 2914 vm_page_wire(m); 2915 vm_page_unlock_queues(); 2916 bp->b_pages[bp->b_npages] = m; 2917 ++bp->b_npages; 2918 } 2919 2920 /* 2921 * Step 2. We've loaded the pages into the buffer, 2922 * we have to figure out if we can still have B_CACHE 2923 * set. Note that B_CACHE is set according to the 2924 * byte-granular range ( bcount and size ), new the 2925 * aligned range ( newbsize ). 2926 * 2927 * The VM test is against m->valid, which is DEV_BSIZE 2928 * aligned. Needless to say, the validity of the data 2929 * needs to also be DEV_BSIZE aligned. Note that this 2930 * fails with NFS if the server or some other client 2931 * extends the file's EOF. If our buffer is resized, 2932 * B_CACHE may remain set! XXX 2933 */ 2934 2935 toff = bp->b_bcount; 2936 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2937 2938 while ((bp->b_flags & B_CACHE) && toff < size) { 2939 vm_pindex_t pi; 2940 2941 if (tinc > (size - toff)) 2942 tinc = size - toff; 2943 2944 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2945 PAGE_SHIFT; 2946 2947 vfs_buf_test_cache( 2948 bp, 2949 bp->b_offset, 2950 toff, 2951 tinc, 2952 bp->b_pages[pi] 2953 ); 2954 toff += tinc; 2955 tinc = PAGE_SIZE; 2956 } 2957 VM_OBJECT_UNLOCK(obj); 2958 2959 /* 2960 * Step 3, fixup the KVM pmap. Remember that 2961 * bp->b_data is relative to bp->b_offset, but 2962 * bp->b_offset may be offset into the first page. 2963 */ 2964 2965 bp->b_data = (caddr_t) 2966 trunc_page((vm_offset_t)bp->b_data); 2967 pmap_qenter( 2968 (vm_offset_t)bp->b_data, 2969 bp->b_pages, 2970 bp->b_npages 2971 ); 2972 2973 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2974 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2975 } 2976 } 2977 if (newbsize < bp->b_bufsize) 2978 bufspacewakeup(); 2979 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2980 bp->b_bcount = size; /* requested buffer size */ 2981 return 1; 2982} 2983 2984void 2985biodone(struct bio *bp) 2986{ 2987 mtx_lock(&bdonelock); 2988 bp->bio_flags |= BIO_DONE; 2989 if (bp->bio_done == NULL) 2990 wakeup(bp); 2991 mtx_unlock(&bdonelock); 2992 if (bp->bio_done != NULL) 2993 bp->bio_done(bp); 2994} 2995 2996/* 2997 * Wait for a BIO to finish. 2998 * 2999 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3000 * case is not yet clear. 3001 */ 3002int 3003biowait(struct bio *bp, const char *wchan) 3004{ 3005 3006 mtx_lock(&bdonelock); 3007 while ((bp->bio_flags & BIO_DONE) == 0) 3008 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3009 mtx_unlock(&bdonelock); 3010 if (bp->bio_error != 0) 3011 return (bp->bio_error); 3012 if (!(bp->bio_flags & BIO_ERROR)) 3013 return (0); 3014 return (EIO); 3015} 3016 3017void 3018biofinish(struct bio *bp, struct devstat *stat, int error) 3019{ 3020 3021 if (error) { 3022 bp->bio_error = error; 3023 bp->bio_flags |= BIO_ERROR; 3024 } 3025 if (stat != NULL) 3026 devstat_end_transaction_bio(stat, bp); 3027 biodone(bp); 3028} 3029 3030/* 3031 * bufwait: 3032 * 3033 * Wait for buffer I/O completion, returning error status. The buffer 3034 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3035 * error and cleared. 3036 */ 3037int 3038bufwait(register struct buf * bp) 3039{ 3040 int s; 3041 3042 s = splbio(); 3043 if (bp->b_iocmd == BIO_READ) 3044 bwait(bp, PRIBIO, "biord"); 3045 else 3046 bwait(bp, PRIBIO, "biowr"); 3047 splx(s); 3048 if (bp->b_flags & B_EINTR) { 3049 bp->b_flags &= ~B_EINTR; 3050 return (EINTR); 3051 } 3052 if (bp->b_ioflags & BIO_ERROR) { 3053 return (bp->b_error ? bp->b_error : EIO); 3054 } else { 3055 return (0); 3056 } 3057} 3058 3059 /* 3060 * Call back function from struct bio back up to struct buf. 3061 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 3062 */ 3063static void 3064bufdonebio(struct bio *bp) 3065{ 3066 3067 /* Device drivers may or may not hold giant, hold it here. */ 3068 mtx_lock(&Giant); 3069 bufdone(bp->bio_caller2); 3070 mtx_unlock(&Giant); 3071} 3072 3073void 3074dev_strategy(struct buf *bp) 3075{ 3076 3077 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3078 panic("b_iocmd botch"); 3079 if (bp->b_flags & B_PHYS) 3080 bp->b_io.bio_offset = bp->b_offset; 3081 else 3082 bp->b_io.bio_offset = dbtob(bp->b_blkno); 3083 bp->b_io.bio_done = bufdonebio; 3084 bp->b_io.bio_caller2 = bp; 3085 (*devsw(bp->b_io.bio_dev)->d_strategy)(&bp->b_io); 3086} 3087 3088/* 3089 * bufdone: 3090 * 3091 * Finish I/O on a buffer, optionally calling a completion function. 3092 * This is usually called from an interrupt so process blocking is 3093 * not allowed. 3094 * 3095 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3096 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3097 * assuming B_INVAL is clear. 3098 * 3099 * For the VMIO case, we set B_CACHE if the op was a read and no 3100 * read error occured, or if the op was a write. B_CACHE is never 3101 * set if the buffer is invalid or otherwise uncacheable. 3102 * 3103 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3104 * initiator to leave B_INVAL set to brelse the buffer out of existance 3105 * in the biodone routine. 3106 */ 3107void 3108bufdone(struct buf *bp) 3109{ 3110 int s; 3111 void (*biodone)(struct buf *); 3112 3113 GIANT_REQUIRED; 3114 3115 s = splbio(); 3116 3117 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3118 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3119 3120 bp->b_flags |= B_DONE; 3121 runningbufwakeup(bp); 3122 3123 if (bp->b_iocmd == BIO_DELETE) { 3124 brelse(bp); 3125 splx(s); 3126 return; 3127 } 3128 3129 if (bp->b_iocmd == BIO_WRITE) { 3130 vwakeup(bp); 3131 } 3132 3133 /* call optional completion function if requested */ 3134 if (bp->b_iodone != NULL) { 3135 biodone = bp->b_iodone; 3136 bp->b_iodone = NULL; 3137 (*biodone) (bp); 3138 splx(s); 3139 return; 3140 } 3141 if (LIST_FIRST(&bp->b_dep) != NULL) 3142 buf_complete(bp); 3143 3144 if (bp->b_flags & B_VMIO) { 3145 int i; 3146 vm_ooffset_t foff; 3147 vm_page_t m; 3148 vm_object_t obj; 3149 int iosize; 3150 struct vnode *vp = bp->b_vp; 3151 3152 obj = bp->b_object; 3153 3154#if defined(VFS_BIO_DEBUG) 3155 mp_fixme("usecount and vflag accessed without locks."); 3156 if (vp->v_usecount == 0) { 3157 panic("biodone: zero vnode ref count"); 3158 } 3159 3160 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3161 panic("biodone: vnode is not setup for merged cache"); 3162 } 3163#endif 3164 3165 foff = bp->b_offset; 3166 KASSERT(bp->b_offset != NOOFFSET, 3167 ("biodone: no buffer offset")); 3168 3169 VM_OBJECT_LOCK(obj); 3170#if defined(VFS_BIO_DEBUG) 3171 if (obj->paging_in_progress < bp->b_npages) { 3172 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3173 obj->paging_in_progress, bp->b_npages); 3174 } 3175#endif 3176 3177 /* 3178 * Set B_CACHE if the op was a normal read and no error 3179 * occured. B_CACHE is set for writes in the b*write() 3180 * routines. 3181 */ 3182 iosize = bp->b_bcount - bp->b_resid; 3183 if (bp->b_iocmd == BIO_READ && 3184 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3185 !(bp->b_ioflags & BIO_ERROR)) { 3186 bp->b_flags |= B_CACHE; 3187 } 3188 vm_page_lock_queues(); 3189 for (i = 0; i < bp->b_npages; i++) { 3190 int bogusflag = 0; 3191 int resid; 3192 3193 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3194 if (resid > iosize) 3195 resid = iosize; 3196 3197 /* 3198 * cleanup bogus pages, restoring the originals 3199 */ 3200 m = bp->b_pages[i]; 3201 if (m == bogus_page) { 3202 bogusflag = 1; 3203 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3204 if (m == NULL) 3205 panic("biodone: page disappeared!"); 3206 bp->b_pages[i] = m; 3207 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3208 } 3209#if defined(VFS_BIO_DEBUG) 3210 if (OFF_TO_IDX(foff) != m->pindex) { 3211 printf( 3212"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3213 (intmax_t)foff, (uintmax_t)m->pindex); 3214 } 3215#endif 3216 3217 /* 3218 * In the write case, the valid and clean bits are 3219 * already changed correctly ( see bdwrite() ), so we 3220 * only need to do this here in the read case. 3221 */ 3222 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3223 vfs_page_set_valid(bp, foff, i, m); 3224 } 3225 vm_page_flag_clear(m, PG_ZERO); 3226 3227 /* 3228 * when debugging new filesystems or buffer I/O methods, this 3229 * is the most common error that pops up. if you see this, you 3230 * have not set the page busy flag correctly!!! 3231 */ 3232 if (m->busy == 0) { 3233 printf("biodone: page busy < 0, " 3234 "pindex: %d, foff: 0x(%x,%x), " 3235 "resid: %d, index: %d\n", 3236 (int) m->pindex, (int)(foff >> 32), 3237 (int) foff & 0xffffffff, resid, i); 3238 if (!vn_isdisk(vp, NULL)) 3239 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3240 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3241 (intmax_t) bp->b_lblkno, 3242 bp->b_flags, bp->b_npages); 3243 else 3244 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3245 (intmax_t) bp->b_lblkno, 3246 bp->b_flags, bp->b_npages); 3247 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3248 (u_long)m->valid, (u_long)m->dirty, 3249 m->wire_count); 3250 panic("biodone: page busy < 0\n"); 3251 } 3252 vm_page_io_finish(m); 3253 vm_object_pip_subtract(obj, 1); 3254 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3255 iosize -= resid; 3256 } 3257 vm_page_unlock_queues(); 3258 vm_object_pip_wakeupn(obj, 0); 3259 VM_OBJECT_UNLOCK(obj); 3260 } 3261 3262 /* 3263 * For asynchronous completions, release the buffer now. The brelse 3264 * will do a wakeup there if necessary - so no need to do a wakeup 3265 * here in the async case. The sync case always needs to do a wakeup. 3266 */ 3267 3268 if (bp->b_flags & B_ASYNC) { 3269 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3270 brelse(bp); 3271 else 3272 bqrelse(bp); 3273 } else { 3274 bdone(bp); 3275 } 3276 splx(s); 3277} 3278 3279/* 3280 * This routine is called in lieu of iodone in the case of 3281 * incomplete I/O. This keeps the busy status for pages 3282 * consistant. 3283 */ 3284void 3285vfs_unbusy_pages(struct buf * bp) 3286{ 3287 int i; 3288 3289 GIANT_REQUIRED; 3290 3291 runningbufwakeup(bp); 3292 if (bp->b_flags & B_VMIO) { 3293 vm_object_t obj; 3294 3295 obj = bp->b_object; 3296 VM_OBJECT_LOCK(obj); 3297 vm_page_lock_queues(); 3298 for (i = 0; i < bp->b_npages; i++) { 3299 vm_page_t m = bp->b_pages[i]; 3300 3301 if (m == bogus_page) { 3302 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3303 if (!m) { 3304 panic("vfs_unbusy_pages: page missing\n"); 3305 } 3306 bp->b_pages[i] = m; 3307 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3308 } 3309 vm_object_pip_subtract(obj, 1); 3310 vm_page_flag_clear(m, PG_ZERO); 3311 vm_page_io_finish(m); 3312 } 3313 vm_page_unlock_queues(); 3314 vm_object_pip_wakeupn(obj, 0); 3315 VM_OBJECT_UNLOCK(obj); 3316 } 3317} 3318 3319/* 3320 * vfs_page_set_valid: 3321 * 3322 * Set the valid bits in a page based on the supplied offset. The 3323 * range is restricted to the buffer's size. 3324 * 3325 * This routine is typically called after a read completes. 3326 */ 3327static void 3328vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3329{ 3330 vm_ooffset_t soff, eoff; 3331 3332 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3333 /* 3334 * Start and end offsets in buffer. eoff - soff may not cross a 3335 * page boundry or cross the end of the buffer. The end of the 3336 * buffer, in this case, is our file EOF, not the allocation size 3337 * of the buffer. 3338 */ 3339 soff = off; 3340 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3341 if (eoff > bp->b_offset + bp->b_bcount) 3342 eoff = bp->b_offset + bp->b_bcount; 3343 3344 /* 3345 * Set valid range. This is typically the entire buffer and thus the 3346 * entire page. 3347 */ 3348 if (eoff > soff) { 3349 vm_page_set_validclean( 3350 m, 3351 (vm_offset_t) (soff & PAGE_MASK), 3352 (vm_offset_t) (eoff - soff) 3353 ); 3354 } 3355} 3356 3357/* 3358 * This routine is called before a device strategy routine. 3359 * It is used to tell the VM system that paging I/O is in 3360 * progress, and treat the pages associated with the buffer 3361 * almost as being PG_BUSY. Also the object paging_in_progress 3362 * flag is handled to make sure that the object doesn't become 3363 * inconsistant. 3364 * 3365 * Since I/O has not been initiated yet, certain buffer flags 3366 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3367 * and should be ignored. 3368 */ 3369void 3370vfs_busy_pages(struct buf * bp, int clear_modify) 3371{ 3372 int i, bogus; 3373 3374 if (bp->b_flags & B_VMIO) { 3375 vm_object_t obj; 3376 vm_ooffset_t foff; 3377 3378 obj = bp->b_object; 3379 foff = bp->b_offset; 3380 KASSERT(bp->b_offset != NOOFFSET, 3381 ("vfs_busy_pages: no buffer offset")); 3382 vfs_setdirty(bp); 3383 VM_OBJECT_LOCK(obj); 3384retry: 3385 vm_page_lock_queues(); 3386 for (i = 0; i < bp->b_npages; i++) { 3387 vm_page_t m = bp->b_pages[i]; 3388 3389 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3390 goto retry; 3391 } 3392 bogus = 0; 3393 for (i = 0; i < bp->b_npages; i++) { 3394 vm_page_t m = bp->b_pages[i]; 3395 3396 vm_page_flag_clear(m, PG_ZERO); 3397 if ((bp->b_flags & B_CLUSTER) == 0) { 3398 vm_object_pip_add(obj, 1); 3399 vm_page_io_start(m); 3400 } 3401 /* 3402 * When readying a buffer for a read ( i.e 3403 * clear_modify == 0 ), it is important to do 3404 * bogus_page replacement for valid pages in 3405 * partially instantiated buffers. Partially 3406 * instantiated buffers can, in turn, occur when 3407 * reconstituting a buffer from its VM backing store 3408 * base. We only have to do this if B_CACHE is 3409 * clear ( which causes the I/O to occur in the 3410 * first place ). The replacement prevents the read 3411 * I/O from overwriting potentially dirty VM-backed 3412 * pages. XXX bogus page replacement is, uh, bogus. 3413 * It may not work properly with small-block devices. 3414 * We need to find a better way. 3415 */ 3416 pmap_remove_all(m); 3417 if (clear_modify) 3418 vfs_page_set_valid(bp, foff, i, m); 3419 else if (m->valid == VM_PAGE_BITS_ALL && 3420 (bp->b_flags & B_CACHE) == 0) { 3421 bp->b_pages[i] = bogus_page; 3422 bogus++; 3423 } 3424 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3425 } 3426 vm_page_unlock_queues(); 3427 VM_OBJECT_UNLOCK(obj); 3428 if (bogus) 3429 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3430 } 3431} 3432 3433/* 3434 * Tell the VM system that the pages associated with this buffer 3435 * are clean. This is used for delayed writes where the data is 3436 * going to go to disk eventually without additional VM intevention. 3437 * 3438 * Note that while we only really need to clean through to b_bcount, we 3439 * just go ahead and clean through to b_bufsize. 3440 */ 3441static void 3442vfs_clean_pages(struct buf * bp) 3443{ 3444 int i; 3445 3446 if (bp->b_flags & B_VMIO) { 3447 vm_ooffset_t foff; 3448 3449 foff = bp->b_offset; 3450 KASSERT(bp->b_offset != NOOFFSET, 3451 ("vfs_clean_pages: no buffer offset")); 3452 VM_OBJECT_LOCK(bp->b_object); 3453 vm_page_lock_queues(); 3454 for (i = 0; i < bp->b_npages; i++) { 3455 vm_page_t m = bp->b_pages[i]; 3456 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3457 vm_ooffset_t eoff = noff; 3458 3459 if (eoff > bp->b_offset + bp->b_bufsize) 3460 eoff = bp->b_offset + bp->b_bufsize; 3461 vfs_page_set_valid(bp, foff, i, m); 3462 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3463 foff = noff; 3464 } 3465 vm_page_unlock_queues(); 3466 VM_OBJECT_UNLOCK(bp->b_object); 3467 } 3468} 3469 3470/* 3471 * vfs_bio_set_validclean: 3472 * 3473 * Set the range within the buffer to valid and clean. The range is 3474 * relative to the beginning of the buffer, b_offset. Note that b_offset 3475 * itself may be offset from the beginning of the first page. 3476 * 3477 */ 3478 3479void 3480vfs_bio_set_validclean(struct buf *bp, int base, int size) 3481{ 3482 if (bp->b_flags & B_VMIO) { 3483 int i; 3484 int n; 3485 3486 /* 3487 * Fixup base to be relative to beginning of first page. 3488 * Set initial n to be the maximum number of bytes in the 3489 * first page that can be validated. 3490 */ 3491 3492 base += (bp->b_offset & PAGE_MASK); 3493 n = PAGE_SIZE - (base & PAGE_MASK); 3494 3495 VM_OBJECT_LOCK(bp->b_object); 3496 vm_page_lock_queues(); 3497 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3498 vm_page_t m = bp->b_pages[i]; 3499 3500 if (n > size) 3501 n = size; 3502 3503 vm_page_set_validclean(m, base & PAGE_MASK, n); 3504 base += n; 3505 size -= n; 3506 n = PAGE_SIZE; 3507 } 3508 vm_page_unlock_queues(); 3509 VM_OBJECT_UNLOCK(bp->b_object); 3510 } 3511} 3512 3513/* 3514 * vfs_bio_clrbuf: 3515 * 3516 * clear a buffer. This routine essentially fakes an I/O, so we need 3517 * to clear BIO_ERROR and B_INVAL. 3518 * 3519 * Note that while we only theoretically need to clear through b_bcount, 3520 * we go ahead and clear through b_bufsize. 3521 */ 3522 3523void 3524vfs_bio_clrbuf(struct buf *bp) 3525{ 3526 int i, mask = 0; 3527 caddr_t sa, ea; 3528 3529 GIANT_REQUIRED; 3530 3531 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3532 bp->b_flags &= ~B_INVAL; 3533 bp->b_ioflags &= ~BIO_ERROR; 3534 if (bp->b_object != NULL) 3535 VM_OBJECT_LOCK(bp->b_object); 3536 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3537 (bp->b_offset & PAGE_MASK) == 0) { 3538 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3539 if (bp->b_pages[0] != bogus_page) 3540 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3541 if ((bp->b_pages[0]->valid & mask) == mask) 3542 goto unlock; 3543 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3544 ((bp->b_pages[0]->valid & mask) == 0)) { 3545 bzero(bp->b_data, bp->b_bufsize); 3546 bp->b_pages[0]->valid |= mask; 3547 goto unlock; 3548 } 3549 } 3550 ea = sa = bp->b_data; 3551 for(i=0;i<bp->b_npages;i++,sa=ea) { 3552 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3553 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3554 ea = (caddr_t)(vm_offset_t)ulmin( 3555 (u_long)(vm_offset_t)ea, 3556 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3557 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3558 if (bp->b_pages[i] != bogus_page) 3559 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3560 if ((bp->b_pages[i]->valid & mask) == mask) 3561 continue; 3562 if ((bp->b_pages[i]->valid & mask) == 0) { 3563 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3564 bzero(sa, ea - sa); 3565 } 3566 } else { 3567 for (; sa < ea; sa += DEV_BSIZE, j++) { 3568 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3569 (bp->b_pages[i]->valid & (1<<j)) == 0) 3570 bzero(sa, DEV_BSIZE); 3571 } 3572 } 3573 bp->b_pages[i]->valid |= mask; 3574 vm_page_lock_queues(); 3575 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3576 vm_page_unlock_queues(); 3577 } 3578unlock: 3579 if (bp->b_object != NULL) 3580 VM_OBJECT_UNLOCK(bp->b_object); 3581 bp->b_resid = 0; 3582 } else { 3583 clrbuf(bp); 3584 } 3585} 3586 3587/* 3588 * vm_hold_load_pages and vm_hold_free_pages get pages into 3589 * a buffers address space. The pages are anonymous and are 3590 * not associated with a file object. 3591 */ 3592static void 3593vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3594{ 3595 vm_offset_t pg; 3596 vm_page_t p; 3597 int index; 3598 3599 to = round_page(to); 3600 from = round_page(from); 3601 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3602 3603 VM_OBJECT_LOCK(kernel_object); 3604 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3605tryagain: 3606 /* 3607 * note: must allocate system pages since blocking here 3608 * could intefere with paging I/O, no matter which 3609 * process we are. 3610 */ 3611 p = vm_page_alloc(kernel_object, 3612 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3613 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3614 if (!p) { 3615 atomic_add_int(&vm_pageout_deficit, 3616 (to - pg) >> PAGE_SHIFT); 3617 VM_OBJECT_UNLOCK(kernel_object); 3618 VM_WAIT; 3619 VM_OBJECT_LOCK(kernel_object); 3620 goto tryagain; 3621 } 3622 p->valid = VM_PAGE_BITS_ALL; 3623 pmap_qenter(pg, &p, 1); 3624 bp->b_pages[index] = p; 3625 vm_page_lock_queues(); 3626 vm_page_wakeup(p); 3627 vm_page_unlock_queues(); 3628 } 3629 VM_OBJECT_UNLOCK(kernel_object); 3630 bp->b_npages = index; 3631} 3632 3633/* Return pages associated with this buf to the vm system */ 3634static void 3635vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3636{ 3637 vm_offset_t pg; 3638 vm_page_t p; 3639 int index, newnpages; 3640 3641 GIANT_REQUIRED; 3642 3643 from = round_page(from); 3644 to = round_page(to); 3645 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3646 3647 VM_OBJECT_LOCK(kernel_object); 3648 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3649 p = bp->b_pages[index]; 3650 if (p && (index < bp->b_npages)) { 3651 if (p->busy) { 3652 printf( 3653 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3654 (intmax_t)bp->b_blkno, 3655 (intmax_t)bp->b_lblkno); 3656 } 3657 bp->b_pages[index] = NULL; 3658 pmap_qremove(pg, 1); 3659 vm_page_lock_queues(); 3660 vm_page_busy(p); 3661 vm_page_unwire(p, 0); 3662 vm_page_free(p); 3663 vm_page_unlock_queues(); 3664 } 3665 } 3666 VM_OBJECT_UNLOCK(kernel_object); 3667 bp->b_npages = newnpages; 3668} 3669 3670/* 3671 * Map an IO request into kernel virtual address space. 3672 * 3673 * All requests are (re)mapped into kernel VA space. 3674 * Notice that we use b_bufsize for the size of the buffer 3675 * to be mapped. b_bcount might be modified by the driver. 3676 * 3677 * Note that even if the caller determines that the address space should 3678 * be valid, a race or a smaller-file mapped into a larger space may 3679 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3680 * check the return value. 3681 */ 3682int 3683vmapbuf(struct buf *bp) 3684{ 3685 caddr_t addr, kva; 3686 vm_prot_t prot; 3687 int pidx, i; 3688 struct vm_page *m; 3689 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3690 3691 GIANT_REQUIRED; 3692 3693 if (bp->b_bufsize < 0) 3694 return (-1); 3695 prot = (bp->b_iocmd == BIO_READ) ? VM_PROT_READ | VM_PROT_WRITE : 3696 VM_PROT_READ; 3697 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3698 addr < bp->b_data + bp->b_bufsize; 3699 addr += PAGE_SIZE, pidx++) { 3700 /* 3701 * Do the vm_fault if needed; do the copy-on-write thing 3702 * when reading stuff off device into memory. 3703 * 3704 * NOTE! Must use pmap_extract() because addr may be in 3705 * the userland address space, and kextract is only guarenteed 3706 * to work for the kernland address space (see: sparc64 port). 3707 */ 3708retry: 3709 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3710 prot) < 0) { 3711 vm_page_lock_queues(); 3712 for (i = 0; i < pidx; ++i) { 3713 vm_page_unhold(bp->b_pages[i]); 3714 bp->b_pages[i] = NULL; 3715 } 3716 vm_page_unlock_queues(); 3717 return(-1); 3718 } 3719 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3720 if (m == NULL) 3721 goto retry; 3722 bp->b_pages[pidx] = m; 3723 } 3724 if (pidx > btoc(MAXPHYS)) 3725 panic("vmapbuf: mapped more than MAXPHYS"); 3726 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3727 3728 kva = bp->b_saveaddr; 3729 bp->b_npages = pidx; 3730 bp->b_saveaddr = bp->b_data; 3731 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3732 return(0); 3733} 3734 3735/* 3736 * Free the io map PTEs associated with this IO operation. 3737 * We also invalidate the TLB entries and restore the original b_addr. 3738 */ 3739void 3740vunmapbuf(struct buf *bp) 3741{ 3742 int pidx; 3743 int npages; 3744 3745 GIANT_REQUIRED; 3746 3747 npages = bp->b_npages; 3748 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3749 npages); 3750 vm_page_lock_queues(); 3751 for (pidx = 0; pidx < npages; pidx++) 3752 vm_page_unhold(bp->b_pages[pidx]); 3753 vm_page_unlock_queues(); 3754 3755 bp->b_data = bp->b_saveaddr; 3756} 3757 3758void 3759bdone(struct buf *bp) 3760{ 3761 mtx_lock(&bdonelock); 3762 bp->b_flags |= B_DONE; 3763 wakeup(bp); 3764 mtx_unlock(&bdonelock); 3765} 3766 3767void 3768bwait(struct buf *bp, u_char pri, const char *wchan) 3769{ 3770 mtx_lock(&bdonelock); 3771 while ((bp->b_flags & B_DONE) == 0) 3772 msleep(bp, &bdonelock, pri, wchan, 0); 3773 mtx_unlock(&bdonelock); 3774} 3775 3776#include "opt_ddb.h" 3777#ifdef DDB 3778#include <ddb/ddb.h> 3779 3780/* DDB command to show buffer data */ 3781DB_SHOW_COMMAND(buffer, db_show_buffer) 3782{ 3783 /* get args */ 3784 struct buf *bp = (struct buf *)addr; 3785 3786 if (!have_addr) { 3787 db_printf("usage: show buffer <addr>\n"); 3788 return; 3789 } 3790 3791 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3792 db_printf( 3793 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3794 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd\n", 3795 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3796 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3797 (intmax_t)bp->b_blkno); 3798 if (bp->b_npages) { 3799 int i; 3800 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3801 for (i = 0; i < bp->b_npages; i++) { 3802 vm_page_t m; 3803 m = bp->b_pages[i]; 3804 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3805 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3806 if ((i + 1) < bp->b_npages) 3807 db_printf(","); 3808 } 3809 db_printf("\n"); 3810 } 3811} 3812#endif /* DDB */ 3813