vfs_bio.c revision 273150
1/*- 2 * Copyright (c) 2004 Poul-Henning Kamp 3 * Copyright (c) 1994,1997 John S. Dyson 4 * Copyright (c) 2013 The FreeBSD Foundation 5 * All rights reserved. 6 * 7 * Portions of this software were developed by Konstantin Belousov 8 * under sponsorship from the FreeBSD Foundation. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 22 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 29 * SUCH DAMAGE. 30 */ 31 32/* 33 * this file contains a new buffer I/O scheme implementing a coherent 34 * VM object and buffer cache scheme. Pains have been taken to make 35 * sure that the performance degradation associated with schemes such 36 * as this is not realized. 37 * 38 * Author: John S. Dyson 39 * Significant help during the development and debugging phases 40 * had been provided by David Greenman, also of the FreeBSD core team. 41 * 42 * see man buf(9) for more info. 43 */ 44 45#include <sys/cdefs.h> 46__FBSDID("$FreeBSD: stable/10/sys/kern/vfs_bio.c 273150 2014-10-15 20:04:21Z jkim $"); 47 48#include <sys/param.h> 49#include <sys/systm.h> 50#include <sys/bio.h> 51#include <sys/conf.h> 52#include <sys/buf.h> 53#include <sys/devicestat.h> 54#include <sys/eventhandler.h> 55#include <sys/fail.h> 56#include <sys/limits.h> 57#include <sys/lock.h> 58#include <sys/malloc.h> 59#include <sys/mount.h> 60#include <sys/mutex.h> 61#include <sys/kernel.h> 62#include <sys/kthread.h> 63#include <sys/proc.h> 64#include <sys/resourcevar.h> 65#include <sys/rwlock.h> 66#include <sys/sysctl.h> 67#include <sys/vmem.h> 68#include <sys/vmmeter.h> 69#include <sys/vnode.h> 70#include <geom/geom.h> 71#include <vm/vm.h> 72#include <vm/vm_param.h> 73#include <vm/vm_kern.h> 74#include <vm/vm_pageout.h> 75#include <vm/vm_page.h> 76#include <vm/vm_object.h> 77#include <vm/vm_extern.h> 78#include <vm/vm_map.h> 79#include "opt_compat.h" 80#include "opt_swap.h" 81 82static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 83 84struct bio_ops bioops; /* I/O operation notification */ 85 86struct buf_ops buf_ops_bio = { 87 .bop_name = "buf_ops_bio", 88 .bop_write = bufwrite, 89 .bop_strategy = bufstrategy, 90 .bop_sync = bufsync, 91 .bop_bdflush = bufbdflush, 92}; 93 94/* 95 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 96 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 97 */ 98struct buf *buf; /* buffer header pool */ 99caddr_t unmapped_buf; 100 101/* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */ 102struct proc *bufdaemonproc; 103 104static int inmem(struct vnode *vp, daddr_t blkno); 105static void vm_hold_free_pages(struct buf *bp, int newbsize); 106static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 107 vm_offset_t to); 108static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 109static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 110 vm_page_t m); 111static void vfs_clean_pages_dirty_buf(struct buf *bp); 112static void vfs_setdirty_locked_object(struct buf *bp); 113static void vfs_vmio_release(struct buf *bp); 114static int vfs_bio_clcheck(struct vnode *vp, int size, 115 daddr_t lblkno, daddr_t blkno); 116static int buf_flush(int); 117static int flushbufqueues(int, int); 118static void buf_daemon(void); 119static void bremfreel(struct buf *bp); 120static __inline void bd_wakeup(void); 121#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 122 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 123static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 124#endif 125 126int vmiodirenable = TRUE; 127SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 128 "Use the VM system for directory writes"); 129long runningbufspace; 130SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 131 "Amount of presently outstanding async buffer io"); 132static long bufspace; 133#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 134 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 135SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 136 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers"); 137#else 138SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 139 "Virtual memory used for buffers"); 140#endif 141static long unmapped_bufspace; 142SYSCTL_LONG(_vfs, OID_AUTO, unmapped_bufspace, CTLFLAG_RD, 143 &unmapped_bufspace, 0, 144 "Amount of unmapped buffers, inclusive in the bufspace"); 145static long maxbufspace; 146SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 147 "Maximum allowed value of bufspace (including buf_daemon)"); 148static long bufmallocspace; 149SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 150 "Amount of malloced memory for buffers"); 151static long maxbufmallocspace; 152SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 153 "Maximum amount of malloced memory for buffers"); 154static long lobufspace; 155SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 156 "Minimum amount of buffers we want to have"); 157long hibufspace; 158SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 159 "Maximum allowed value of bufspace (excluding buf_daemon)"); 160static int bufreusecnt; 161SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 162 "Number of times we have reused a buffer"); 163static int buffreekvacnt; 164SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 165 "Number of times we have freed the KVA space from some buffer"); 166static int bufdefragcnt; 167SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 168 "Number of times we have had to repeat buffer allocation to defragment"); 169static long lorunningspace; 170SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 171 "Minimum preferred space used for in-progress I/O"); 172static long hirunningspace; 173SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 174 "Maximum amount of space to use for in-progress I/O"); 175int dirtybufferflushes; 176SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 177 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 178int bdwriteskip; 179SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 180 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 181int altbufferflushes; 182SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 183 0, "Number of fsync flushes to limit dirty buffers"); 184static int recursiveflushes; 185SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 186 0, "Number of flushes skipped due to being recursive"); 187static int numdirtybuffers; 188SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 189 "Number of buffers that are dirty (has unwritten changes) at the moment"); 190static int lodirtybuffers; 191SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 192 "How many buffers we want to have free before bufdaemon can sleep"); 193static int hidirtybuffers; 194SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 195 "When the number of dirty buffers is considered severe"); 196int dirtybufthresh; 197SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 198 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 199static int numfreebuffers; 200SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 201 "Number of free buffers"); 202static int lofreebuffers; 203SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 204 "XXX Unused"); 205static int hifreebuffers; 206SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 207 "XXX Complicatedly unused"); 208static int getnewbufcalls; 209SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 210 "Number of calls to getnewbuf"); 211static int getnewbufrestarts; 212SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 213 "Number of times getnewbuf has had to restart a buffer aquisition"); 214static int mappingrestarts; 215SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0, 216 "Number of times getblk has had to restart a buffer mapping for " 217 "unmapped buffer"); 218static int flushbufqtarget = 100; 219SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 220 "Amount of work to do in flushbufqueues when helping bufdaemon"); 221static long notbufdflushes; 222SYSCTL_LONG(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes, 0, 223 "Number of dirty buffer flushes done by the bufdaemon helpers"); 224static long barrierwrites; 225SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0, 226 "Number of barrier writes"); 227SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD, 228 &unmapped_buf_allowed, 0, 229 "Permit the use of the unmapped i/o"); 230 231/* 232 * Lock for the non-dirty bufqueues 233 */ 234static struct mtx_padalign bqclean; 235 236/* 237 * Lock for the dirty queue. 238 */ 239static struct mtx_padalign bqdirty; 240 241/* 242 * This lock synchronizes access to bd_request. 243 */ 244static struct mtx_padalign bdlock; 245 246/* 247 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 248 * waitrunningbufspace(). 249 */ 250static struct mtx_padalign rbreqlock; 251 252/* 253 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 254 */ 255static struct rwlock_padalign nblock; 256 257/* 258 * Lock that protects bdirtywait. 259 */ 260static struct mtx_padalign bdirtylock; 261 262/* 263 * Wakeup point for bufdaemon, as well as indicator of whether it is already 264 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 265 * is idling. 266 */ 267static int bd_request; 268 269/* 270 * Request for the buf daemon to write more buffers than is indicated by 271 * lodirtybuf. This may be necessary to push out excess dependencies or 272 * defragment the address space where a simple count of the number of dirty 273 * buffers is insufficient to characterize the demand for flushing them. 274 */ 275static int bd_speedupreq; 276 277/* 278 * bogus page -- for I/O to/from partially complete buffers 279 * this is a temporary solution to the problem, but it is not 280 * really that bad. it would be better to split the buffer 281 * for input in the case of buffers partially already in memory, 282 * but the code is intricate enough already. 283 */ 284vm_page_t bogus_page; 285 286/* 287 * Synchronization (sleep/wakeup) variable for active buffer space requests. 288 * Set when wait starts, cleared prior to wakeup(). 289 * Used in runningbufwakeup() and waitrunningbufspace(). 290 */ 291static int runningbufreq; 292 293/* 294 * Synchronization (sleep/wakeup) variable for buffer requests. 295 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 296 * by and/or. 297 * Used in numdirtywakeup(), bufspacewakeup(), bufcountadd(), bwillwrite(), 298 * getnewbuf(), and getblk(). 299 */ 300static volatile int needsbuffer; 301 302/* 303 * Synchronization for bwillwrite() waiters. 304 */ 305static int bdirtywait; 306 307/* 308 * Definitions for the buffer free lists. 309 */ 310#define BUFFER_QUEUES 5 /* number of free buffer queues */ 311 312#define QUEUE_NONE 0 /* on no queue */ 313#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 314#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 315#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 316#define QUEUE_EMPTY 4 /* empty buffer headers */ 317#define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */ 318 319/* Queues for free buffers with various properties */ 320static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 321#ifdef INVARIANTS 322static int bq_len[BUFFER_QUEUES]; 323#endif 324 325/* 326 * Single global constant for BUF_WMESG, to avoid getting multiple references. 327 * buf_wmesg is referred from macros. 328 */ 329const char *buf_wmesg = BUF_WMESG; 330 331#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 332#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 333#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 334 335#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 336 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 337static int 338sysctl_bufspace(SYSCTL_HANDLER_ARGS) 339{ 340 long lvalue; 341 int ivalue; 342 343 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 344 return (sysctl_handle_long(oidp, arg1, arg2, req)); 345 lvalue = *(long *)arg1; 346 if (lvalue > INT_MAX) 347 /* On overflow, still write out a long to trigger ENOMEM. */ 348 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 349 ivalue = lvalue; 350 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 351} 352#endif 353 354/* 355 * bqlock: 356 * 357 * Return the appropriate queue lock based on the index. 358 */ 359static inline struct mtx * 360bqlock(int qindex) 361{ 362 363 if (qindex == QUEUE_DIRTY) 364 return (struct mtx *)(&bqdirty); 365 return (struct mtx *)(&bqclean); 366} 367 368/* 369 * bdirtywakeup: 370 * 371 * Wakeup any bwillwrite() waiters. 372 */ 373static void 374bdirtywakeup(void) 375{ 376 mtx_lock(&bdirtylock); 377 if (bdirtywait) { 378 bdirtywait = 0; 379 wakeup(&bdirtywait); 380 } 381 mtx_unlock(&bdirtylock); 382} 383 384/* 385 * bdirtysub: 386 * 387 * Decrement the numdirtybuffers count by one and wakeup any 388 * threads blocked in bwillwrite(). 389 */ 390static void 391bdirtysub(void) 392{ 393 394 if (atomic_fetchadd_int(&numdirtybuffers, -1) == 395 (lodirtybuffers + hidirtybuffers) / 2) 396 bdirtywakeup(); 397} 398 399/* 400 * bdirtyadd: 401 * 402 * Increment the numdirtybuffers count by one and wakeup the buf 403 * daemon if needed. 404 */ 405static void 406bdirtyadd(void) 407{ 408 409 /* 410 * Only do the wakeup once as we cross the boundary. The 411 * buf daemon will keep running until the condition clears. 412 */ 413 if (atomic_fetchadd_int(&numdirtybuffers, 1) == 414 (lodirtybuffers + hidirtybuffers) / 2) 415 bd_wakeup(); 416} 417 418/* 419 * bufspacewakeup: 420 * 421 * Called when buffer space is potentially available for recovery. 422 * getnewbuf() will block on this flag when it is unable to free 423 * sufficient buffer space. Buffer space becomes recoverable when 424 * bp's get placed back in the queues. 425 */ 426 427static __inline void 428bufspacewakeup(void) 429{ 430 int need_wakeup, on; 431 432 /* 433 * If someone is waiting for BUF space, wake them up. Even 434 * though we haven't freed the kva space yet, the waiting 435 * process will be able to now. 436 */ 437 rw_rlock(&nblock); 438 for (;;) { 439 need_wakeup = 0; 440 on = needsbuffer; 441 if ((on & VFS_BIO_NEED_BUFSPACE) == 0) 442 break; 443 need_wakeup = 1; 444 if (atomic_cmpset_rel_int(&needsbuffer, on, 445 on & ~VFS_BIO_NEED_BUFSPACE)) 446 break; 447 } 448 if (need_wakeup) 449 wakeup(__DEVOLATILE(void *, &needsbuffer)); 450 rw_runlock(&nblock); 451} 452 453/* 454 * runningwakeup: 455 * 456 * Wake up processes that are waiting on asynchronous writes to fall 457 * below lorunningspace. 458 */ 459static void 460runningwakeup(void) 461{ 462 463 mtx_lock(&rbreqlock); 464 if (runningbufreq) { 465 runningbufreq = 0; 466 wakeup(&runningbufreq); 467 } 468 mtx_unlock(&rbreqlock); 469} 470 471/* 472 * runningbufwakeup: 473 * 474 * Decrement the outstanding write count according. 475 */ 476void 477runningbufwakeup(struct buf *bp) 478{ 479 long space, bspace; 480 481 bspace = bp->b_runningbufspace; 482 if (bspace == 0) 483 return; 484 space = atomic_fetchadd_long(&runningbufspace, -bspace); 485 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld", 486 space, bspace)); 487 bp->b_runningbufspace = 0; 488 /* 489 * Only acquire the lock and wakeup on the transition from exceeding 490 * the threshold to falling below it. 491 */ 492 if (space < lorunningspace) 493 return; 494 if (space - bspace > lorunningspace) 495 return; 496 runningwakeup(); 497} 498 499/* 500 * bufcountadd: 501 * 502 * Called when a buffer has been added to one of the free queues to 503 * account for the buffer and to wakeup anyone waiting for free buffers. 504 * This typically occurs when large amounts of metadata are being handled 505 * by the buffer cache ( else buffer space runs out first, usually ). 506 */ 507static __inline void 508bufcountadd(struct buf *bp) 509{ 510 int mask, need_wakeup, old, on; 511 512 KASSERT((bp->b_flags & B_INFREECNT) == 0, 513 ("buf %p already counted as free", bp)); 514 bp->b_flags |= B_INFREECNT; 515 old = atomic_fetchadd_int(&numfreebuffers, 1); 516 KASSERT(old >= 0 && old < nbuf, 517 ("numfreebuffers climbed to %d", old + 1)); 518 mask = VFS_BIO_NEED_ANY; 519 if (numfreebuffers >= hifreebuffers) 520 mask |= VFS_BIO_NEED_FREE; 521 rw_rlock(&nblock); 522 for (;;) { 523 need_wakeup = 0; 524 on = needsbuffer; 525 if (on == 0) 526 break; 527 need_wakeup = 1; 528 if (atomic_cmpset_rel_int(&needsbuffer, on, on & ~mask)) 529 break; 530 } 531 if (need_wakeup) 532 wakeup(__DEVOLATILE(void *, &needsbuffer)); 533 rw_runlock(&nblock); 534} 535 536/* 537 * bufcountsub: 538 * 539 * Decrement the numfreebuffers count as needed. 540 */ 541static void 542bufcountsub(struct buf *bp) 543{ 544 int old; 545 546 /* 547 * Fixup numfreebuffers count. If the buffer is invalid or not 548 * delayed-write, the buffer was free and we must decrement 549 * numfreebuffers. 550 */ 551 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 552 KASSERT((bp->b_flags & B_INFREECNT) != 0, 553 ("buf %p not counted in numfreebuffers", bp)); 554 bp->b_flags &= ~B_INFREECNT; 555 old = atomic_fetchadd_int(&numfreebuffers, -1); 556 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 557 } 558} 559 560/* 561 * waitrunningbufspace() 562 * 563 * runningbufspace is a measure of the amount of I/O currently 564 * running. This routine is used in async-write situations to 565 * prevent creating huge backups of pending writes to a device. 566 * Only asynchronous writes are governed by this function. 567 * 568 * This does NOT turn an async write into a sync write. It waits 569 * for earlier writes to complete and generally returns before the 570 * caller's write has reached the device. 571 */ 572void 573waitrunningbufspace(void) 574{ 575 576 mtx_lock(&rbreqlock); 577 while (runningbufspace > hirunningspace) { 578 runningbufreq = 1; 579 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 580 } 581 mtx_unlock(&rbreqlock); 582} 583 584 585/* 586 * vfs_buf_test_cache: 587 * 588 * Called when a buffer is extended. This function clears the B_CACHE 589 * bit if the newly extended portion of the buffer does not contain 590 * valid data. 591 */ 592static __inline 593void 594vfs_buf_test_cache(struct buf *bp, 595 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 596 vm_page_t m) 597{ 598 599 VM_OBJECT_ASSERT_LOCKED(m->object); 600 if (bp->b_flags & B_CACHE) { 601 int base = (foff + off) & PAGE_MASK; 602 if (vm_page_is_valid(m, base, size) == 0) 603 bp->b_flags &= ~B_CACHE; 604 } 605} 606 607/* Wake up the buffer daemon if necessary */ 608static __inline void 609bd_wakeup(void) 610{ 611 612 mtx_lock(&bdlock); 613 if (bd_request == 0) { 614 bd_request = 1; 615 wakeup(&bd_request); 616 } 617 mtx_unlock(&bdlock); 618} 619 620/* 621 * bd_speedup - speedup the buffer cache flushing code 622 */ 623void 624bd_speedup(void) 625{ 626 int needwake; 627 628 mtx_lock(&bdlock); 629 needwake = 0; 630 if (bd_speedupreq == 0 || bd_request == 0) 631 needwake = 1; 632 bd_speedupreq = 1; 633 bd_request = 1; 634 if (needwake) 635 wakeup(&bd_request); 636 mtx_unlock(&bdlock); 637} 638 639#ifndef NSWBUF_MIN 640#define NSWBUF_MIN 16 641#endif 642 643#ifdef __i386__ 644#define TRANSIENT_DENOM 5 645#else 646#define TRANSIENT_DENOM 10 647#endif 648 649/* 650 * Calculating buffer cache scaling values and reserve space for buffer 651 * headers. This is called during low level kernel initialization and 652 * may be called more then once. We CANNOT write to the memory area 653 * being reserved at this time. 654 */ 655caddr_t 656kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 657{ 658 int tuned_nbuf; 659 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz; 660 661 /* 662 * physmem_est is in pages. Convert it to kilobytes (assumes 663 * PAGE_SIZE is >= 1K) 664 */ 665 physmem_est = physmem_est * (PAGE_SIZE / 1024); 666 667 /* 668 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 669 * For the first 64MB of ram nominally allocate sufficient buffers to 670 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 671 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 672 * the buffer cache we limit the eventual kva reservation to 673 * maxbcache bytes. 674 * 675 * factor represents the 1/4 x ram conversion. 676 */ 677 if (nbuf == 0) { 678 int factor = 4 * BKVASIZE / 1024; 679 680 nbuf = 50; 681 if (physmem_est > 4096) 682 nbuf += min((physmem_est - 4096) / factor, 683 65536 / factor); 684 if (physmem_est > 65536) 685 nbuf += min((physmem_est - 65536) * 2 / (factor * 5), 686 32 * 1024 * 1024 / (factor * 5)); 687 688 if (maxbcache && nbuf > maxbcache / BKVASIZE) 689 nbuf = maxbcache / BKVASIZE; 690 tuned_nbuf = 1; 691 } else 692 tuned_nbuf = 0; 693 694 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 695 maxbuf = (LONG_MAX / 3) / BKVASIZE; 696 if (nbuf > maxbuf) { 697 if (!tuned_nbuf) 698 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 699 maxbuf); 700 nbuf = maxbuf; 701 } 702 703 /* 704 * Ideal allocation size for the transient bio submap if 10% 705 * of the maximal space buffer map. This roughly corresponds 706 * to the amount of the buffer mapped for typical UFS load. 707 * 708 * Clip the buffer map to reserve space for the transient 709 * BIOs, if its extent is bigger than 90% (80% on i386) of the 710 * maximum buffer map extent on the platform. 711 * 712 * The fall-back to the maxbuf in case of maxbcache unset, 713 * allows to not trim the buffer KVA for the architectures 714 * with ample KVA space. 715 */ 716 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) { 717 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE; 718 buf_sz = (long)nbuf * BKVASIZE; 719 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM * 720 (TRANSIENT_DENOM - 1)) { 721 /* 722 * There is more KVA than memory. Do not 723 * adjust buffer map size, and assign the rest 724 * of maxbuf to transient map. 725 */ 726 biotmap_sz = maxbuf_sz - buf_sz; 727 } else { 728 /* 729 * Buffer map spans all KVA we could afford on 730 * this platform. Give 10% (20% on i386) of 731 * the buffer map to the transient bio map. 732 */ 733 biotmap_sz = buf_sz / TRANSIENT_DENOM; 734 buf_sz -= biotmap_sz; 735 } 736 if (biotmap_sz / INT_MAX > MAXPHYS) 737 bio_transient_maxcnt = INT_MAX; 738 else 739 bio_transient_maxcnt = biotmap_sz / MAXPHYS; 740 /* 741 * Artifically limit to 1024 simultaneous in-flight I/Os 742 * using the transient mapping. 743 */ 744 if (bio_transient_maxcnt > 1024) 745 bio_transient_maxcnt = 1024; 746 if (tuned_nbuf) 747 nbuf = buf_sz / BKVASIZE; 748 } 749 750 /* 751 * swbufs are used as temporary holders for I/O, such as paging I/O. 752 * We have no less then 16 and no more then 256. 753 */ 754 nswbuf = min(nbuf / 4, 256); 755 TUNABLE_INT_FETCH("kern.nswbuf", &nswbuf); 756 if (nswbuf < NSWBUF_MIN) 757 nswbuf = NSWBUF_MIN; 758 759 /* 760 * Reserve space for the buffer cache buffers 761 */ 762 swbuf = (void *)v; 763 v = (caddr_t)(swbuf + nswbuf); 764 buf = (void *)v; 765 v = (caddr_t)(buf + nbuf); 766 767 return(v); 768} 769 770/* Initialize the buffer subsystem. Called before use of any buffers. */ 771void 772bufinit(void) 773{ 774 struct buf *bp; 775 int i; 776 777 mtx_init(&bqclean, "bufq clean lock", NULL, MTX_DEF); 778 mtx_init(&bqdirty, "bufq dirty lock", NULL, MTX_DEF); 779 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 780 rw_init(&nblock, "needsbuffer lock"); 781 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 782 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF); 783 784 /* next, make a null set of free lists */ 785 for (i = 0; i < BUFFER_QUEUES; i++) 786 TAILQ_INIT(&bufqueues[i]); 787 788 /* finally, initialize each buffer header and stick on empty q */ 789 for (i = 0; i < nbuf; i++) { 790 bp = &buf[i]; 791 bzero(bp, sizeof *bp); 792 bp->b_flags = B_INVAL | B_INFREECNT; 793 bp->b_rcred = NOCRED; 794 bp->b_wcred = NOCRED; 795 bp->b_qindex = QUEUE_EMPTY; 796 bp->b_xflags = 0; 797 LIST_INIT(&bp->b_dep); 798 BUF_LOCKINIT(bp); 799 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 800#ifdef INVARIANTS 801 bq_len[QUEUE_EMPTY]++; 802#endif 803 } 804 805 /* 806 * maxbufspace is the absolute maximum amount of buffer space we are 807 * allowed to reserve in KVM and in real terms. The absolute maximum 808 * is nominally used by buf_daemon. hibufspace is the nominal maximum 809 * used by most other processes. The differential is required to 810 * ensure that buf_daemon is able to run when other processes might 811 * be blocked waiting for buffer space. 812 * 813 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 814 * this may result in KVM fragmentation which is not handled optimally 815 * by the system. 816 */ 817 maxbufspace = (long)nbuf * BKVASIZE; 818 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 819 lobufspace = hibufspace - MAXBSIZE; 820 821 /* 822 * Note: The 16 MiB upper limit for hirunningspace was chosen 823 * arbitrarily and may need further tuning. It corresponds to 824 * 128 outstanding write IO requests (if IO size is 128 KiB), 825 * which fits with many RAID controllers' tagged queuing limits. 826 * The lower 1 MiB limit is the historical upper limit for 827 * hirunningspace. 828 */ 829 hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE), 830 16 * 1024 * 1024), 1024 * 1024); 831 lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE); 832 833/* 834 * Limit the amount of malloc memory since it is wired permanently into 835 * the kernel space. Even though this is accounted for in the buffer 836 * allocation, we don't want the malloced region to grow uncontrolled. 837 * The malloc scheme improves memory utilization significantly on average 838 * (small) directories. 839 */ 840 maxbufmallocspace = hibufspace / 20; 841 842/* 843 * Reduce the chance of a deadlock occuring by limiting the number 844 * of delayed-write dirty buffers we allow to stack up. 845 */ 846 hidirtybuffers = nbuf / 4 + 20; 847 dirtybufthresh = hidirtybuffers * 9 / 10; 848 numdirtybuffers = 0; 849/* 850 * To support extreme low-memory systems, make sure hidirtybuffers cannot 851 * eat up all available buffer space. This occurs when our minimum cannot 852 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 853 * BKVASIZE'd buffers. 854 */ 855 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 856 hidirtybuffers >>= 1; 857 } 858 lodirtybuffers = hidirtybuffers / 2; 859 860/* 861 * Try to keep the number of free buffers in the specified range, 862 * and give special processes (e.g. like buf_daemon) access to an 863 * emergency reserve. 864 */ 865 lofreebuffers = nbuf / 18 + 5; 866 hifreebuffers = 2 * lofreebuffers; 867 numfreebuffers = nbuf; 868 869 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 870 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 871 unmapped_buf = (caddr_t)kva_alloc(MAXPHYS); 872} 873 874#ifdef INVARIANTS 875static inline void 876vfs_buf_check_mapped(struct buf *bp) 877{ 878 879 KASSERT((bp->b_flags & B_UNMAPPED) == 0, 880 ("mapped buf %p %x", bp, bp->b_flags)); 881 KASSERT(bp->b_kvabase != unmapped_buf, 882 ("mapped buf: b_kvabase was not updated %p", bp)); 883 KASSERT(bp->b_data != unmapped_buf, 884 ("mapped buf: b_data was not updated %p", bp)); 885} 886 887static inline void 888vfs_buf_check_unmapped(struct buf *bp) 889{ 890 891 KASSERT((bp->b_flags & B_UNMAPPED) == B_UNMAPPED, 892 ("unmapped buf %p %x", bp, bp->b_flags)); 893 KASSERT(bp->b_kvabase == unmapped_buf, 894 ("unmapped buf: corrupted b_kvabase %p", bp)); 895 KASSERT(bp->b_data == unmapped_buf, 896 ("unmapped buf: corrupted b_data %p", bp)); 897} 898 899#define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) 900#define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) 901#else 902#define BUF_CHECK_MAPPED(bp) do {} while (0) 903#define BUF_CHECK_UNMAPPED(bp) do {} while (0) 904#endif 905 906static void 907bpmap_qenter(struct buf *bp) 908{ 909 910 BUF_CHECK_MAPPED(bp); 911 912 /* 913 * bp->b_data is relative to bp->b_offset, but 914 * bp->b_offset may be offset into the first page. 915 */ 916 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 917 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 918 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 919 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 920} 921 922/* 923 * bfreekva() - free the kva allocation for a buffer. 924 * 925 * Since this call frees up buffer space, we call bufspacewakeup(). 926 */ 927static void 928bfreekva(struct buf *bp) 929{ 930 931 if (bp->b_kvasize == 0) 932 return; 933 934 atomic_add_int(&buffreekvacnt, 1); 935 atomic_subtract_long(&bufspace, bp->b_kvasize); 936 if ((bp->b_flags & B_UNMAPPED) == 0) { 937 BUF_CHECK_MAPPED(bp); 938 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, 939 bp->b_kvasize); 940 } else { 941 BUF_CHECK_UNMAPPED(bp); 942 if ((bp->b_flags & B_KVAALLOC) != 0) { 943 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvaalloc, 944 bp->b_kvasize); 945 } 946 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize); 947 bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC); 948 } 949 bp->b_kvasize = 0; 950 bufspacewakeup(); 951} 952 953/* 954 * binsfree: 955 * 956 * Insert the buffer into the appropriate free list. 957 */ 958static void 959binsfree(struct buf *bp, int qindex) 960{ 961 struct mtx *olock, *nlock; 962 963 BUF_ASSERT_XLOCKED(bp); 964 965 olock = bqlock(bp->b_qindex); 966 nlock = bqlock(qindex); 967 mtx_lock(olock); 968 /* Handle delayed bremfree() processing. */ 969 if (bp->b_flags & B_REMFREE) 970 bremfreel(bp); 971 972 if (bp->b_qindex != QUEUE_NONE) 973 panic("binsfree: free buffer onto another queue???"); 974 975 bp->b_qindex = qindex; 976 if (olock != nlock) { 977 mtx_unlock(olock); 978 mtx_lock(nlock); 979 } 980 if (bp->b_flags & B_AGE) 981 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 982 else 983 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 984#ifdef INVARIANTS 985 bq_len[bp->b_qindex]++; 986#endif 987 mtx_unlock(nlock); 988 989 /* 990 * Something we can maybe free or reuse. 991 */ 992 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 993 bufspacewakeup(); 994 995 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 996 bufcountadd(bp); 997} 998 999/* 1000 * bremfree: 1001 * 1002 * Mark the buffer for removal from the appropriate free list. 1003 * 1004 */ 1005void 1006bremfree(struct buf *bp) 1007{ 1008 1009 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1010 KASSERT((bp->b_flags & B_REMFREE) == 0, 1011 ("bremfree: buffer %p already marked for delayed removal.", bp)); 1012 KASSERT(bp->b_qindex != QUEUE_NONE, 1013 ("bremfree: buffer %p not on a queue.", bp)); 1014 BUF_ASSERT_XLOCKED(bp); 1015 1016 bp->b_flags |= B_REMFREE; 1017 bufcountsub(bp); 1018} 1019 1020/* 1021 * bremfreef: 1022 * 1023 * Force an immediate removal from a free list. Used only in nfs when 1024 * it abuses the b_freelist pointer. 1025 */ 1026void 1027bremfreef(struct buf *bp) 1028{ 1029 struct mtx *qlock; 1030 1031 qlock = bqlock(bp->b_qindex); 1032 mtx_lock(qlock); 1033 bremfreel(bp); 1034 mtx_unlock(qlock); 1035} 1036 1037/* 1038 * bremfreel: 1039 * 1040 * Removes a buffer from the free list, must be called with the 1041 * correct qlock held. 1042 */ 1043static void 1044bremfreel(struct buf *bp) 1045{ 1046 1047 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 1048 bp, bp->b_vp, bp->b_flags); 1049 KASSERT(bp->b_qindex != QUEUE_NONE, 1050 ("bremfreel: buffer %p not on a queue.", bp)); 1051 BUF_ASSERT_XLOCKED(bp); 1052 mtx_assert(bqlock(bp->b_qindex), MA_OWNED); 1053 1054 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 1055#ifdef INVARIANTS 1056 KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow", 1057 bp->b_qindex)); 1058 bq_len[bp->b_qindex]--; 1059#endif 1060 bp->b_qindex = QUEUE_NONE; 1061 /* 1062 * If this was a delayed bremfree() we only need to remove the buffer 1063 * from the queue and return the stats are already done. 1064 */ 1065 if (bp->b_flags & B_REMFREE) { 1066 bp->b_flags &= ~B_REMFREE; 1067 return; 1068 } 1069 bufcountsub(bp); 1070} 1071 1072/* 1073 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 1074 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 1075 * the buffer is valid and we do not have to do anything. 1076 */ 1077void 1078breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 1079 int cnt, struct ucred * cred) 1080{ 1081 struct buf *rabp; 1082 int i; 1083 1084 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 1085 if (inmem(vp, *rablkno)) 1086 continue; 1087 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 1088 1089 if ((rabp->b_flags & B_CACHE) == 0) { 1090 if (!TD_IS_IDLETHREAD(curthread)) 1091 curthread->td_ru.ru_inblock++; 1092 rabp->b_flags |= B_ASYNC; 1093 rabp->b_flags &= ~B_INVAL; 1094 rabp->b_ioflags &= ~BIO_ERROR; 1095 rabp->b_iocmd = BIO_READ; 1096 if (rabp->b_rcred == NOCRED && cred != NOCRED) 1097 rabp->b_rcred = crhold(cred); 1098 vfs_busy_pages(rabp, 0); 1099 BUF_KERNPROC(rabp); 1100 rabp->b_iooffset = dbtob(rabp->b_blkno); 1101 bstrategy(rabp); 1102 } else { 1103 brelse(rabp); 1104 } 1105 } 1106} 1107 1108/* 1109 * Entry point for bread() and breadn() via #defines in sys/buf.h. 1110 * 1111 * Get a buffer with the specified data. Look in the cache first. We 1112 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 1113 * is set, the buffer is valid and we do not have to do anything, see 1114 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 1115 */ 1116int 1117breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno, 1118 int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp) 1119{ 1120 struct buf *bp; 1121 int rv = 0, readwait = 0; 1122 1123 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 1124 /* 1125 * Can only return NULL if GB_LOCK_NOWAIT flag is specified. 1126 */ 1127 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags); 1128 if (bp == NULL) 1129 return (EBUSY); 1130 1131 /* if not found in cache, do some I/O */ 1132 if ((bp->b_flags & B_CACHE) == 0) { 1133 if (!TD_IS_IDLETHREAD(curthread)) 1134 curthread->td_ru.ru_inblock++; 1135 bp->b_iocmd = BIO_READ; 1136 bp->b_flags &= ~B_INVAL; 1137 bp->b_ioflags &= ~BIO_ERROR; 1138 if (bp->b_rcred == NOCRED && cred != NOCRED) 1139 bp->b_rcred = crhold(cred); 1140 vfs_busy_pages(bp, 0); 1141 bp->b_iooffset = dbtob(bp->b_blkno); 1142 bstrategy(bp); 1143 ++readwait; 1144 } 1145 1146 breada(vp, rablkno, rabsize, cnt, cred); 1147 1148 if (readwait) { 1149 rv = bufwait(bp); 1150 } 1151 return (rv); 1152} 1153 1154/* 1155 * Write, release buffer on completion. (Done by iodone 1156 * if async). Do not bother writing anything if the buffer 1157 * is invalid. 1158 * 1159 * Note that we set B_CACHE here, indicating that buffer is 1160 * fully valid and thus cacheable. This is true even of NFS 1161 * now so we set it generally. This could be set either here 1162 * or in biodone() since the I/O is synchronous. We put it 1163 * here. 1164 */ 1165int 1166bufwrite(struct buf *bp) 1167{ 1168 int oldflags; 1169 struct vnode *vp; 1170 long space; 1171 int vp_md; 1172 1173 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1174 if (bp->b_flags & B_INVAL) { 1175 brelse(bp); 1176 return (0); 1177 } 1178 1179 if (bp->b_flags & B_BARRIER) 1180 barrierwrites++; 1181 1182 oldflags = bp->b_flags; 1183 1184 BUF_ASSERT_HELD(bp); 1185 1186 if (bp->b_pin_count > 0) 1187 bunpin_wait(bp); 1188 1189 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 1190 ("FFS background buffer should not get here %p", bp)); 1191 1192 vp = bp->b_vp; 1193 if (vp) 1194 vp_md = vp->v_vflag & VV_MD; 1195 else 1196 vp_md = 0; 1197 1198 /* 1199 * Mark the buffer clean. Increment the bufobj write count 1200 * before bundirty() call, to prevent other thread from seeing 1201 * empty dirty list and zero counter for writes in progress, 1202 * falsely indicating that the bufobj is clean. 1203 */ 1204 bufobj_wref(bp->b_bufobj); 1205 bundirty(bp); 1206 1207 bp->b_flags &= ~B_DONE; 1208 bp->b_ioflags &= ~BIO_ERROR; 1209 bp->b_flags |= B_CACHE; 1210 bp->b_iocmd = BIO_WRITE; 1211 1212 vfs_busy_pages(bp, 1); 1213 1214 /* 1215 * Normal bwrites pipeline writes 1216 */ 1217 bp->b_runningbufspace = bp->b_bufsize; 1218 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace); 1219 1220 if (!TD_IS_IDLETHREAD(curthread)) 1221 curthread->td_ru.ru_oublock++; 1222 if (oldflags & B_ASYNC) 1223 BUF_KERNPROC(bp); 1224 bp->b_iooffset = dbtob(bp->b_blkno); 1225 bstrategy(bp); 1226 1227 if ((oldflags & B_ASYNC) == 0) { 1228 int rtval = bufwait(bp); 1229 brelse(bp); 1230 return (rtval); 1231 } else if (space > hirunningspace) { 1232 /* 1233 * don't allow the async write to saturate the I/O 1234 * system. We will not deadlock here because 1235 * we are blocking waiting for I/O that is already in-progress 1236 * to complete. We do not block here if it is the update 1237 * or syncer daemon trying to clean up as that can lead 1238 * to deadlock. 1239 */ 1240 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 1241 waitrunningbufspace(); 1242 } 1243 1244 return (0); 1245} 1246 1247void 1248bufbdflush(struct bufobj *bo, struct buf *bp) 1249{ 1250 struct buf *nbp; 1251 1252 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 1253 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 1254 altbufferflushes++; 1255 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 1256 BO_LOCK(bo); 1257 /* 1258 * Try to find a buffer to flush. 1259 */ 1260 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 1261 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1262 BUF_LOCK(nbp, 1263 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1264 continue; 1265 if (bp == nbp) 1266 panic("bdwrite: found ourselves"); 1267 BO_UNLOCK(bo); 1268 /* Don't countdeps with the bo lock held. */ 1269 if (buf_countdeps(nbp, 0)) { 1270 BO_LOCK(bo); 1271 BUF_UNLOCK(nbp); 1272 continue; 1273 } 1274 if (nbp->b_flags & B_CLUSTEROK) { 1275 vfs_bio_awrite(nbp); 1276 } else { 1277 bremfree(nbp); 1278 bawrite(nbp); 1279 } 1280 dirtybufferflushes++; 1281 break; 1282 } 1283 if (nbp == NULL) 1284 BO_UNLOCK(bo); 1285 } 1286} 1287 1288/* 1289 * Delayed write. (Buffer is marked dirty). Do not bother writing 1290 * anything if the buffer is marked invalid. 1291 * 1292 * Note that since the buffer must be completely valid, we can safely 1293 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1294 * biodone() in order to prevent getblk from writing the buffer 1295 * out synchronously. 1296 */ 1297void 1298bdwrite(struct buf *bp) 1299{ 1300 struct thread *td = curthread; 1301 struct vnode *vp; 1302 struct bufobj *bo; 1303 1304 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1305 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1306 KASSERT((bp->b_flags & B_BARRIER) == 0, 1307 ("Barrier request in delayed write %p", bp)); 1308 BUF_ASSERT_HELD(bp); 1309 1310 if (bp->b_flags & B_INVAL) { 1311 brelse(bp); 1312 return; 1313 } 1314 1315 /* 1316 * If we have too many dirty buffers, don't create any more. 1317 * If we are wildly over our limit, then force a complete 1318 * cleanup. Otherwise, just keep the situation from getting 1319 * out of control. Note that we have to avoid a recursive 1320 * disaster and not try to clean up after our own cleanup! 1321 */ 1322 vp = bp->b_vp; 1323 bo = bp->b_bufobj; 1324 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 1325 td->td_pflags |= TDP_INBDFLUSH; 1326 BO_BDFLUSH(bo, bp); 1327 td->td_pflags &= ~TDP_INBDFLUSH; 1328 } else 1329 recursiveflushes++; 1330 1331 bdirty(bp); 1332 /* 1333 * Set B_CACHE, indicating that the buffer is fully valid. This is 1334 * true even of NFS now. 1335 */ 1336 bp->b_flags |= B_CACHE; 1337 1338 /* 1339 * This bmap keeps the system from needing to do the bmap later, 1340 * perhaps when the system is attempting to do a sync. Since it 1341 * is likely that the indirect block -- or whatever other datastructure 1342 * that the filesystem needs is still in memory now, it is a good 1343 * thing to do this. Note also, that if the pageout daemon is 1344 * requesting a sync -- there might not be enough memory to do 1345 * the bmap then... So, this is important to do. 1346 */ 1347 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1348 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1349 } 1350 1351 /* 1352 * Set the *dirty* buffer range based upon the VM system dirty 1353 * pages. 1354 * 1355 * Mark the buffer pages as clean. We need to do this here to 1356 * satisfy the vnode_pager and the pageout daemon, so that it 1357 * thinks that the pages have been "cleaned". Note that since 1358 * the pages are in a delayed write buffer -- the VFS layer 1359 * "will" see that the pages get written out on the next sync, 1360 * or perhaps the cluster will be completed. 1361 */ 1362 vfs_clean_pages_dirty_buf(bp); 1363 bqrelse(bp); 1364 1365 /* 1366 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1367 * due to the softdep code. 1368 */ 1369} 1370 1371/* 1372 * bdirty: 1373 * 1374 * Turn buffer into delayed write request. We must clear BIO_READ and 1375 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1376 * itself to properly update it in the dirty/clean lists. We mark it 1377 * B_DONE to ensure that any asynchronization of the buffer properly 1378 * clears B_DONE ( else a panic will occur later ). 1379 * 1380 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1381 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1382 * should only be called if the buffer is known-good. 1383 * 1384 * Since the buffer is not on a queue, we do not update the numfreebuffers 1385 * count. 1386 * 1387 * The buffer must be on QUEUE_NONE. 1388 */ 1389void 1390bdirty(struct buf *bp) 1391{ 1392 1393 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1394 bp, bp->b_vp, bp->b_flags); 1395 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1396 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1397 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1398 BUF_ASSERT_HELD(bp); 1399 bp->b_flags &= ~(B_RELBUF); 1400 bp->b_iocmd = BIO_WRITE; 1401 1402 if ((bp->b_flags & B_DELWRI) == 0) { 1403 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1404 reassignbuf(bp); 1405 bdirtyadd(); 1406 } 1407} 1408 1409/* 1410 * bundirty: 1411 * 1412 * Clear B_DELWRI for buffer. 1413 * 1414 * Since the buffer is not on a queue, we do not update the numfreebuffers 1415 * count. 1416 * 1417 * The buffer must be on QUEUE_NONE. 1418 */ 1419 1420void 1421bundirty(struct buf *bp) 1422{ 1423 1424 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1425 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1426 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1427 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1428 BUF_ASSERT_HELD(bp); 1429 1430 if (bp->b_flags & B_DELWRI) { 1431 bp->b_flags &= ~B_DELWRI; 1432 reassignbuf(bp); 1433 bdirtysub(); 1434 } 1435 /* 1436 * Since it is now being written, we can clear its deferred write flag. 1437 */ 1438 bp->b_flags &= ~B_DEFERRED; 1439} 1440 1441/* 1442 * bawrite: 1443 * 1444 * Asynchronous write. Start output on a buffer, but do not wait for 1445 * it to complete. The buffer is released when the output completes. 1446 * 1447 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1448 * B_INVAL buffers. Not us. 1449 */ 1450void 1451bawrite(struct buf *bp) 1452{ 1453 1454 bp->b_flags |= B_ASYNC; 1455 (void) bwrite(bp); 1456} 1457 1458/* 1459 * babarrierwrite: 1460 * 1461 * Asynchronous barrier write. Start output on a buffer, but do not 1462 * wait for it to complete. Place a write barrier after this write so 1463 * that this buffer and all buffers written before it are committed to 1464 * the disk before any buffers written after this write are committed 1465 * to the disk. The buffer is released when the output completes. 1466 */ 1467void 1468babarrierwrite(struct buf *bp) 1469{ 1470 1471 bp->b_flags |= B_ASYNC | B_BARRIER; 1472 (void) bwrite(bp); 1473} 1474 1475/* 1476 * bbarrierwrite: 1477 * 1478 * Synchronous barrier write. Start output on a buffer and wait for 1479 * it to complete. Place a write barrier after this write so that 1480 * this buffer and all buffers written before it are committed to 1481 * the disk before any buffers written after this write are committed 1482 * to the disk. The buffer is released when the output completes. 1483 */ 1484int 1485bbarrierwrite(struct buf *bp) 1486{ 1487 1488 bp->b_flags |= B_BARRIER; 1489 return (bwrite(bp)); 1490} 1491 1492/* 1493 * bwillwrite: 1494 * 1495 * Called prior to the locking of any vnodes when we are expecting to 1496 * write. We do not want to starve the buffer cache with too many 1497 * dirty buffers so we block here. By blocking prior to the locking 1498 * of any vnodes we attempt to avoid the situation where a locked vnode 1499 * prevents the various system daemons from flushing related buffers. 1500 */ 1501void 1502bwillwrite(void) 1503{ 1504 1505 if (numdirtybuffers >= hidirtybuffers) { 1506 mtx_lock(&bdirtylock); 1507 while (numdirtybuffers >= hidirtybuffers) { 1508 bdirtywait = 1; 1509 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4), 1510 "flswai", 0); 1511 } 1512 mtx_unlock(&bdirtylock); 1513 } 1514} 1515 1516/* 1517 * Return true if we have too many dirty buffers. 1518 */ 1519int 1520buf_dirty_count_severe(void) 1521{ 1522 1523 return(numdirtybuffers >= hidirtybuffers); 1524} 1525 1526static __noinline int 1527buf_vm_page_count_severe(void) 1528{ 1529 1530 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1); 1531 1532 return vm_page_count_severe(); 1533} 1534 1535/* 1536 * brelse: 1537 * 1538 * Release a busy buffer and, if requested, free its resources. The 1539 * buffer will be stashed in the appropriate bufqueue[] allowing it 1540 * to be accessed later as a cache entity or reused for other purposes. 1541 */ 1542void 1543brelse(struct buf *bp) 1544{ 1545 int qindex; 1546 1547 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1548 bp, bp->b_vp, bp->b_flags); 1549 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1550 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1551 1552 if (BUF_LOCKRECURSED(bp)) { 1553 /* 1554 * Do not process, in particular, do not handle the 1555 * B_INVAL/B_RELBUF and do not release to free list. 1556 */ 1557 BUF_UNLOCK(bp); 1558 return; 1559 } 1560 1561 if (bp->b_flags & B_MANAGED) { 1562 bqrelse(bp); 1563 return; 1564 } 1565 1566 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1567 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1568 /* 1569 * Failed write, redirty. Must clear BIO_ERROR to prevent 1570 * pages from being scrapped. If the error is anything 1571 * other than an I/O error (EIO), assume that retrying 1572 * is futile. 1573 */ 1574 bp->b_ioflags &= ~BIO_ERROR; 1575 bdirty(bp); 1576 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1577 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1578 /* 1579 * Either a failed I/O or we were asked to free or not 1580 * cache the buffer. 1581 */ 1582 bp->b_flags |= B_INVAL; 1583 if (!LIST_EMPTY(&bp->b_dep)) 1584 buf_deallocate(bp); 1585 if (bp->b_flags & B_DELWRI) 1586 bdirtysub(); 1587 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1588 if ((bp->b_flags & B_VMIO) == 0) { 1589 if (bp->b_bufsize) 1590 allocbuf(bp, 0); 1591 if (bp->b_vp) 1592 brelvp(bp); 1593 } 1594 } 1595 1596 /* 1597 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1598 * is called with B_DELWRI set, the underlying pages may wind up 1599 * getting freed causing a previous write (bdwrite()) to get 'lost' 1600 * because pages associated with a B_DELWRI bp are marked clean. 1601 * 1602 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1603 * if B_DELWRI is set. 1604 * 1605 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1606 * on pages to return pages to the VM page queues. 1607 */ 1608 if (bp->b_flags & B_DELWRI) 1609 bp->b_flags &= ~B_RELBUF; 1610 else if (buf_vm_page_count_severe()) { 1611 /* 1612 * BKGRDINPROG can only be set with the buf and bufobj 1613 * locks both held. We tolerate a race to clear it here. 1614 */ 1615 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1616 bp->b_flags |= B_RELBUF; 1617 } 1618 1619 /* 1620 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1621 * constituted, not even NFS buffers now. Two flags effect this. If 1622 * B_INVAL, the struct buf is invalidated but the VM object is kept 1623 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1624 * 1625 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1626 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1627 * buffer is also B_INVAL because it hits the re-dirtying code above. 1628 * 1629 * Normally we can do this whether a buffer is B_DELWRI or not. If 1630 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1631 * the commit state and we cannot afford to lose the buffer. If the 1632 * buffer has a background write in progress, we need to keep it 1633 * around to prevent it from being reconstituted and starting a second 1634 * background write. 1635 */ 1636 if ((bp->b_flags & B_VMIO) 1637 && !(bp->b_vp->v_mount != NULL && 1638 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1639 !vn_isdisk(bp->b_vp, NULL) && 1640 (bp->b_flags & B_DELWRI)) 1641 ) { 1642 1643 int i, j, resid; 1644 vm_page_t m; 1645 off_t foff; 1646 vm_pindex_t poff; 1647 vm_object_t obj; 1648 1649 obj = bp->b_bufobj->bo_object; 1650 1651 /* 1652 * Get the base offset and length of the buffer. Note that 1653 * in the VMIO case if the buffer block size is not 1654 * page-aligned then b_data pointer may not be page-aligned. 1655 * But our b_pages[] array *IS* page aligned. 1656 * 1657 * block sizes less then DEV_BSIZE (usually 512) are not 1658 * supported due to the page granularity bits (m->valid, 1659 * m->dirty, etc...). 1660 * 1661 * See man buf(9) for more information 1662 */ 1663 resid = bp->b_bufsize; 1664 foff = bp->b_offset; 1665 for (i = 0; i < bp->b_npages; i++) { 1666 int had_bogus = 0; 1667 1668 m = bp->b_pages[i]; 1669 1670 /* 1671 * If we hit a bogus page, fixup *all* the bogus pages 1672 * now. 1673 */ 1674 if (m == bogus_page) { 1675 poff = OFF_TO_IDX(bp->b_offset); 1676 had_bogus = 1; 1677 1678 VM_OBJECT_RLOCK(obj); 1679 for (j = i; j < bp->b_npages; j++) { 1680 vm_page_t mtmp; 1681 mtmp = bp->b_pages[j]; 1682 if (mtmp == bogus_page) { 1683 mtmp = vm_page_lookup(obj, poff + j); 1684 if (!mtmp) { 1685 panic("brelse: page missing\n"); 1686 } 1687 bp->b_pages[j] = mtmp; 1688 } 1689 } 1690 VM_OBJECT_RUNLOCK(obj); 1691 1692 if ((bp->b_flags & (B_INVAL | B_UNMAPPED)) == 0) { 1693 BUF_CHECK_MAPPED(bp); 1694 pmap_qenter( 1695 trunc_page((vm_offset_t)bp->b_data), 1696 bp->b_pages, bp->b_npages); 1697 } 1698 m = bp->b_pages[i]; 1699 } 1700 if ((bp->b_flags & B_NOCACHE) || 1701 (bp->b_ioflags & BIO_ERROR && 1702 bp->b_iocmd == BIO_READ)) { 1703 int poffset = foff & PAGE_MASK; 1704 int presid = resid > (PAGE_SIZE - poffset) ? 1705 (PAGE_SIZE - poffset) : resid; 1706 1707 KASSERT(presid >= 0, ("brelse: extra page")); 1708 VM_OBJECT_WLOCK(obj); 1709 while (vm_page_xbusied(m)) { 1710 vm_page_lock(m); 1711 VM_OBJECT_WUNLOCK(obj); 1712 vm_page_busy_sleep(m, "mbncsh"); 1713 VM_OBJECT_WLOCK(obj); 1714 } 1715 if (pmap_page_wired_mappings(m) == 0) 1716 vm_page_set_invalid(m, poffset, presid); 1717 VM_OBJECT_WUNLOCK(obj); 1718 if (had_bogus) 1719 printf("avoided corruption bug in bogus_page/brelse code\n"); 1720 } 1721 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1722 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1723 } 1724 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1725 vfs_vmio_release(bp); 1726 1727 } else if (bp->b_flags & B_VMIO) { 1728 1729 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1730 vfs_vmio_release(bp); 1731 } 1732 1733 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1734 if (bp->b_bufsize != 0) 1735 allocbuf(bp, 0); 1736 if (bp->b_vp != NULL) 1737 brelvp(bp); 1738 } 1739 1740 /* 1741 * If the buffer has junk contents signal it and eventually 1742 * clean up B_DELWRI and diassociate the vnode so that gbincore() 1743 * doesn't find it. 1744 */ 1745 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1746 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1747 bp->b_flags |= B_INVAL; 1748 if (bp->b_flags & B_INVAL) { 1749 if (bp->b_flags & B_DELWRI) 1750 bundirty(bp); 1751 if (bp->b_vp) 1752 brelvp(bp); 1753 } 1754 1755 /* buffers with no memory */ 1756 if (bp->b_bufsize == 0) { 1757 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1758 if (bp->b_vflags & BV_BKGRDINPROG) 1759 panic("losing buffer 1"); 1760 if (bp->b_kvasize) 1761 qindex = QUEUE_EMPTYKVA; 1762 else 1763 qindex = QUEUE_EMPTY; 1764 bp->b_flags |= B_AGE; 1765 /* buffers with junk contents */ 1766 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1767 (bp->b_ioflags & BIO_ERROR)) { 1768 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1769 if (bp->b_vflags & BV_BKGRDINPROG) 1770 panic("losing buffer 2"); 1771 qindex = QUEUE_CLEAN; 1772 bp->b_flags |= B_AGE; 1773 /* remaining buffers */ 1774 } else if (bp->b_flags & B_DELWRI) 1775 qindex = QUEUE_DIRTY; 1776 else 1777 qindex = QUEUE_CLEAN; 1778 1779 binsfree(bp, qindex); 1780 1781 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1782 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1783 panic("brelse: not dirty"); 1784 /* unlock */ 1785 BUF_UNLOCK(bp); 1786} 1787 1788/* 1789 * Release a buffer back to the appropriate queue but do not try to free 1790 * it. The buffer is expected to be used again soon. 1791 * 1792 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1793 * biodone() to requeue an async I/O on completion. It is also used when 1794 * known good buffers need to be requeued but we think we may need the data 1795 * again soon. 1796 * 1797 * XXX we should be able to leave the B_RELBUF hint set on completion. 1798 */ 1799void 1800bqrelse(struct buf *bp) 1801{ 1802 int qindex; 1803 1804 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1805 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1806 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1807 1808 if (BUF_LOCKRECURSED(bp)) { 1809 /* do not release to free list */ 1810 BUF_UNLOCK(bp); 1811 return; 1812 } 1813 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1814 1815 if (bp->b_flags & B_MANAGED) { 1816 if (bp->b_flags & B_REMFREE) 1817 bremfreef(bp); 1818 goto out; 1819 } 1820 1821 /* buffers with stale but valid contents */ 1822 if (bp->b_flags & B_DELWRI) { 1823 qindex = QUEUE_DIRTY; 1824 } else { 1825 if ((bp->b_flags & B_DELWRI) == 0 && 1826 (bp->b_xflags & BX_VNDIRTY)) 1827 panic("bqrelse: not dirty"); 1828 /* 1829 * BKGRDINPROG can only be set with the buf and bufobj 1830 * locks both held. We tolerate a race to clear it here. 1831 */ 1832 if (buf_vm_page_count_severe() && 1833 (bp->b_vflags & BV_BKGRDINPROG) == 0) { 1834 /* 1835 * We are too low on memory, we have to try to free 1836 * the buffer (most importantly: the wired pages 1837 * making up its backing store) *now*. 1838 */ 1839 brelse(bp); 1840 return; 1841 } 1842 qindex = QUEUE_CLEAN; 1843 } 1844 binsfree(bp, qindex); 1845 1846out: 1847 /* unlock */ 1848 BUF_UNLOCK(bp); 1849} 1850 1851/* Give pages used by the bp back to the VM system (where possible) */ 1852static void 1853vfs_vmio_release(struct buf *bp) 1854{ 1855 int i; 1856 vm_page_t m; 1857 1858 if ((bp->b_flags & B_UNMAPPED) == 0) { 1859 BUF_CHECK_MAPPED(bp); 1860 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 1861 } else 1862 BUF_CHECK_UNMAPPED(bp); 1863 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 1864 for (i = 0; i < bp->b_npages; i++) { 1865 m = bp->b_pages[i]; 1866 bp->b_pages[i] = NULL; 1867 /* 1868 * In order to keep page LRU ordering consistent, put 1869 * everything on the inactive queue. 1870 */ 1871 vm_page_lock(m); 1872 vm_page_unwire(m, 0); 1873 1874 /* 1875 * Might as well free the page if we can and it has 1876 * no valid data. We also free the page if the 1877 * buffer was used for direct I/O 1878 */ 1879 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) { 1880 if (m->wire_count == 0 && !vm_page_busied(m)) 1881 vm_page_free(m); 1882 } else if (bp->b_flags & B_DIRECT) 1883 vm_page_try_to_free(m); 1884 else if (buf_vm_page_count_severe()) 1885 vm_page_try_to_cache(m); 1886 vm_page_unlock(m); 1887 } 1888 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 1889 1890 if (bp->b_bufsize) { 1891 bufspacewakeup(); 1892 bp->b_bufsize = 0; 1893 } 1894 bp->b_npages = 0; 1895 bp->b_flags &= ~B_VMIO; 1896 if (bp->b_vp) 1897 brelvp(bp); 1898} 1899 1900/* 1901 * Check to see if a block at a particular lbn is available for a clustered 1902 * write. 1903 */ 1904static int 1905vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1906{ 1907 struct buf *bpa; 1908 int match; 1909 1910 match = 0; 1911 1912 /* If the buf isn't in core skip it */ 1913 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1914 return (0); 1915 1916 /* If the buf is busy we don't want to wait for it */ 1917 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1918 return (0); 1919 1920 /* Only cluster with valid clusterable delayed write buffers */ 1921 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1922 (B_DELWRI | B_CLUSTEROK)) 1923 goto done; 1924 1925 if (bpa->b_bufsize != size) 1926 goto done; 1927 1928 /* 1929 * Check to see if it is in the expected place on disk and that the 1930 * block has been mapped. 1931 */ 1932 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1933 match = 1; 1934done: 1935 BUF_UNLOCK(bpa); 1936 return (match); 1937} 1938 1939/* 1940 * vfs_bio_awrite: 1941 * 1942 * Implement clustered async writes for clearing out B_DELWRI buffers. 1943 * This is much better then the old way of writing only one buffer at 1944 * a time. Note that we may not be presented with the buffers in the 1945 * correct order, so we search for the cluster in both directions. 1946 */ 1947int 1948vfs_bio_awrite(struct buf *bp) 1949{ 1950 struct bufobj *bo; 1951 int i; 1952 int j; 1953 daddr_t lblkno = bp->b_lblkno; 1954 struct vnode *vp = bp->b_vp; 1955 int ncl; 1956 int nwritten; 1957 int size; 1958 int maxcl; 1959 int gbflags; 1960 1961 bo = &vp->v_bufobj; 1962 gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0; 1963 /* 1964 * right now we support clustered writing only to regular files. If 1965 * we find a clusterable block we could be in the middle of a cluster 1966 * rather then at the beginning. 1967 */ 1968 if ((vp->v_type == VREG) && 1969 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1970 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1971 1972 size = vp->v_mount->mnt_stat.f_iosize; 1973 maxcl = MAXPHYS / size; 1974 1975 BO_RLOCK(bo); 1976 for (i = 1; i < maxcl; i++) 1977 if (vfs_bio_clcheck(vp, size, lblkno + i, 1978 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1979 break; 1980 1981 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1982 if (vfs_bio_clcheck(vp, size, lblkno - j, 1983 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1984 break; 1985 BO_RUNLOCK(bo); 1986 --j; 1987 ncl = i + j; 1988 /* 1989 * this is a possible cluster write 1990 */ 1991 if (ncl != 1) { 1992 BUF_UNLOCK(bp); 1993 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 1994 gbflags); 1995 return (nwritten); 1996 } 1997 } 1998 bremfree(bp); 1999 bp->b_flags |= B_ASYNC; 2000 /* 2001 * default (old) behavior, writing out only one block 2002 * 2003 * XXX returns b_bufsize instead of b_bcount for nwritten? 2004 */ 2005 nwritten = bp->b_bufsize; 2006 (void) bwrite(bp); 2007 2008 return (nwritten); 2009} 2010 2011static void 2012setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags) 2013{ 2014 2015 KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 2016 bp->b_kvasize == 0, ("call bfreekva(%p)", bp)); 2017 if ((gbflags & GB_UNMAPPED) == 0) { 2018 bp->b_kvabase = (caddr_t)addr; 2019 } else if ((gbflags & GB_KVAALLOC) != 0) { 2020 KASSERT((gbflags & GB_UNMAPPED) != 0, 2021 ("GB_KVAALLOC without GB_UNMAPPED")); 2022 bp->b_kvaalloc = (caddr_t)addr; 2023 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 2024 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 2025 } 2026 bp->b_kvasize = maxsize; 2027} 2028 2029/* 2030 * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if 2031 * needed. 2032 */ 2033static int 2034allocbufkva(struct buf *bp, int maxsize, int gbflags) 2035{ 2036 vm_offset_t addr; 2037 2038 bfreekva(bp); 2039 addr = 0; 2040 2041 if (vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr)) { 2042 /* 2043 * Buffer map is too fragmented. Request the caller 2044 * to defragment the map. 2045 */ 2046 atomic_add_int(&bufdefragcnt, 1); 2047 return (1); 2048 } 2049 setbufkva(bp, addr, maxsize, gbflags); 2050 atomic_add_long(&bufspace, bp->b_kvasize); 2051 return (0); 2052} 2053 2054/* 2055 * Ask the bufdaemon for help, or act as bufdaemon itself, when a 2056 * locked vnode is supplied. 2057 */ 2058static void 2059getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo, 2060 int defrag) 2061{ 2062 struct thread *td; 2063 char *waitmsg; 2064 int cnt, error, flags, norunbuf, wait; 2065 2066 mtx_assert(&bqclean, MA_OWNED); 2067 2068 if (defrag) { 2069 flags = VFS_BIO_NEED_BUFSPACE; 2070 waitmsg = "nbufkv"; 2071 } else if (bufspace >= hibufspace) { 2072 waitmsg = "nbufbs"; 2073 flags = VFS_BIO_NEED_BUFSPACE; 2074 } else { 2075 waitmsg = "newbuf"; 2076 flags = VFS_BIO_NEED_ANY; 2077 } 2078 atomic_set_int(&needsbuffer, flags); 2079 mtx_unlock(&bqclean); 2080 2081 bd_speedup(); /* heeeelp */ 2082 if ((gbflags & GB_NOWAIT_BD) != 0) 2083 return; 2084 2085 td = curthread; 2086 cnt = 0; 2087 wait = MNT_NOWAIT; 2088 rw_wlock(&nblock); 2089 while ((needsbuffer & flags) != 0) { 2090 if (vp != NULL && vp->v_type != VCHR && 2091 (td->td_pflags & TDP_BUFNEED) == 0) { 2092 rw_wunlock(&nblock); 2093 /* 2094 * getblk() is called with a vnode locked, and 2095 * some majority of the dirty buffers may as 2096 * well belong to the vnode. Flushing the 2097 * buffers there would make a progress that 2098 * cannot be achieved by the buf_daemon, that 2099 * cannot lock the vnode. 2100 */ 2101 if (cnt++ > 2) 2102 wait = MNT_WAIT; 2103 ASSERT_VOP_LOCKED(vp, "bufd_helper"); 2104 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 : 2105 vn_lock(vp, LK_TRYUPGRADE); 2106 if (error == 0) { 2107 /* play bufdaemon */ 2108 norunbuf = curthread_pflags_set(TDP_BUFNEED | 2109 TDP_NORUNNINGBUF); 2110 VOP_FSYNC(vp, wait, td); 2111 atomic_add_long(¬bufdflushes, 1); 2112 curthread_pflags_restore(norunbuf); 2113 } 2114 rw_wlock(&nblock); 2115 if ((needsbuffer & flags) == 0) 2116 break; 2117 } 2118 error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock, 2119 (PRIBIO + 4) | slpflag, waitmsg, slptimeo); 2120 if (error != 0) 2121 break; 2122 } 2123 rw_wunlock(&nblock); 2124} 2125 2126static void 2127getnewbuf_reuse_bp(struct buf *bp, int qindex) 2128{ 2129 2130 CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 2131 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 2132 bp->b_kvasize, bp->b_bufsize, qindex); 2133 mtx_assert(&bqclean, MA_NOTOWNED); 2134 2135 /* 2136 * Note: we no longer distinguish between VMIO and non-VMIO 2137 * buffers. 2138 */ 2139 KASSERT((bp->b_flags & B_DELWRI) == 0, 2140 ("delwri buffer %p found in queue %d", bp, qindex)); 2141 2142 if (qindex == QUEUE_CLEAN) { 2143 if (bp->b_flags & B_VMIO) { 2144 bp->b_flags &= ~B_ASYNC; 2145 vfs_vmio_release(bp); 2146 } 2147 if (bp->b_vp != NULL) 2148 brelvp(bp); 2149 } 2150 2151 /* 2152 * Get the rest of the buffer freed up. b_kva* is still valid 2153 * after this operation. 2154 */ 2155 2156 if (bp->b_rcred != NOCRED) { 2157 crfree(bp->b_rcred); 2158 bp->b_rcred = NOCRED; 2159 } 2160 if (bp->b_wcred != NOCRED) { 2161 crfree(bp->b_wcred); 2162 bp->b_wcred = NOCRED; 2163 } 2164 if (!LIST_EMPTY(&bp->b_dep)) 2165 buf_deallocate(bp); 2166 if (bp->b_vflags & BV_BKGRDINPROG) 2167 panic("losing buffer 3"); 2168 KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p. qindex: %d", 2169 bp, bp->b_vp, qindex)); 2170 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 2171 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 2172 2173 if (bp->b_bufsize) 2174 allocbuf(bp, 0); 2175 2176 bp->b_flags &= B_UNMAPPED | B_KVAALLOC; 2177 bp->b_ioflags = 0; 2178 bp->b_xflags = 0; 2179 KASSERT((bp->b_flags & B_INFREECNT) == 0, 2180 ("buf %p still counted as free?", bp)); 2181 bp->b_vflags = 0; 2182 bp->b_vp = NULL; 2183 bp->b_blkno = bp->b_lblkno = 0; 2184 bp->b_offset = NOOFFSET; 2185 bp->b_iodone = 0; 2186 bp->b_error = 0; 2187 bp->b_resid = 0; 2188 bp->b_bcount = 0; 2189 bp->b_npages = 0; 2190 bp->b_dirtyoff = bp->b_dirtyend = 0; 2191 bp->b_bufobj = NULL; 2192 bp->b_pin_count = 0; 2193 bp->b_fsprivate1 = NULL; 2194 bp->b_fsprivate2 = NULL; 2195 bp->b_fsprivate3 = NULL; 2196 2197 LIST_INIT(&bp->b_dep); 2198} 2199 2200static int flushingbufs; 2201 2202static struct buf * 2203getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata) 2204{ 2205 struct buf *bp, *nbp; 2206 int nqindex, qindex, pass; 2207 2208 KASSERT(!unmapped || !defrag, ("both unmapped and defrag")); 2209 2210 pass = 1; 2211restart: 2212 atomic_add_int(&getnewbufrestarts, 1); 2213 2214 /* 2215 * Setup for scan. If we do not have enough free buffers, 2216 * we setup a degenerate case that immediately fails. Note 2217 * that if we are specially marked process, we are allowed to 2218 * dip into our reserves. 2219 * 2220 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 2221 * for the allocation of the mapped buffer. For unmapped, the 2222 * easiest is to start with EMPTY outright. 2223 * 2224 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 2225 * However, there are a number of cases (defragging, reusing, ...) 2226 * where we cannot backup. 2227 */ 2228 nbp = NULL; 2229 mtx_lock(&bqclean); 2230 if (!defrag && unmapped) { 2231 nqindex = QUEUE_EMPTY; 2232 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2233 } 2234 if (nbp == NULL) { 2235 nqindex = QUEUE_EMPTYKVA; 2236 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2237 } 2238 2239 /* 2240 * If no EMPTYKVA buffers and we are either defragging or 2241 * reusing, locate a CLEAN buffer to free or reuse. If 2242 * bufspace useage is low skip this step so we can allocate a 2243 * new buffer. 2244 */ 2245 if (nbp == NULL && (defrag || bufspace >= lobufspace)) { 2246 nqindex = QUEUE_CLEAN; 2247 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2248 } 2249 2250 /* 2251 * If we could not find or were not allowed to reuse a CLEAN 2252 * buffer, check to see if it is ok to use an EMPTY buffer. 2253 * We can only use an EMPTY buffer if allocating its KVA would 2254 * not otherwise run us out of buffer space. No KVA is needed 2255 * for the unmapped allocation. 2256 */ 2257 if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace || 2258 metadata)) { 2259 nqindex = QUEUE_EMPTY; 2260 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2261 } 2262 2263 /* 2264 * All available buffers might be clean, retry ignoring the 2265 * lobufspace as the last resort. 2266 */ 2267 if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) { 2268 nqindex = QUEUE_CLEAN; 2269 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2270 } 2271 2272 /* 2273 * Run scan, possibly freeing data and/or kva mappings on the fly 2274 * depending. 2275 */ 2276 while ((bp = nbp) != NULL) { 2277 qindex = nqindex; 2278 2279 /* 2280 * Calculate next bp (we can only use it if we do not 2281 * block or do other fancy things). 2282 */ 2283 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 2284 switch (qindex) { 2285 case QUEUE_EMPTY: 2286 nqindex = QUEUE_EMPTYKVA; 2287 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2288 if (nbp != NULL) 2289 break; 2290 /* FALLTHROUGH */ 2291 case QUEUE_EMPTYKVA: 2292 nqindex = QUEUE_CLEAN; 2293 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2294 if (nbp != NULL) 2295 break; 2296 /* FALLTHROUGH */ 2297 case QUEUE_CLEAN: 2298 if (metadata && pass == 1) { 2299 pass = 2; 2300 nqindex = QUEUE_EMPTY; 2301 nbp = TAILQ_FIRST( 2302 &bufqueues[QUEUE_EMPTY]); 2303 } 2304 /* 2305 * nbp is NULL. 2306 */ 2307 break; 2308 } 2309 } 2310 /* 2311 * If we are defragging then we need a buffer with 2312 * b_kvasize != 0. XXX this situation should no longer 2313 * occur, if defrag is non-zero the buffer's b_kvasize 2314 * should also be non-zero at this point. XXX 2315 */ 2316 if (defrag && bp->b_kvasize == 0) { 2317 printf("Warning: defrag empty buffer %p\n", bp); 2318 continue; 2319 } 2320 2321 /* 2322 * Start freeing the bp. This is somewhat involved. nbp 2323 * remains valid only for QUEUE_EMPTY[KVA] bp's. 2324 */ 2325 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2326 continue; 2327 /* 2328 * BKGRDINPROG can only be set with the buf and bufobj 2329 * locks both held. We tolerate a race to clear it here. 2330 */ 2331 if (bp->b_vflags & BV_BKGRDINPROG) { 2332 BUF_UNLOCK(bp); 2333 continue; 2334 } 2335 2336 KASSERT(bp->b_qindex == qindex, 2337 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2338 2339 bremfreel(bp); 2340 mtx_unlock(&bqclean); 2341 /* 2342 * NOTE: nbp is now entirely invalid. We can only restart 2343 * the scan from this point on. 2344 */ 2345 2346 getnewbuf_reuse_bp(bp, qindex); 2347 mtx_assert(&bqclean, MA_NOTOWNED); 2348 2349 /* 2350 * If we are defragging then free the buffer. 2351 */ 2352 if (defrag) { 2353 bp->b_flags |= B_INVAL; 2354 bfreekva(bp); 2355 brelse(bp); 2356 defrag = 0; 2357 goto restart; 2358 } 2359 2360 /* 2361 * Notify any waiters for the buffer lock about 2362 * identity change by freeing the buffer. 2363 */ 2364 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 2365 bp->b_flags |= B_INVAL; 2366 bfreekva(bp); 2367 brelse(bp); 2368 goto restart; 2369 } 2370 2371 if (metadata) 2372 break; 2373 2374 /* 2375 * If we are overcomitted then recover the buffer and its 2376 * KVM space. This occurs in rare situations when multiple 2377 * processes are blocked in getnewbuf() or allocbuf(). 2378 */ 2379 if (bufspace >= hibufspace) 2380 flushingbufs = 1; 2381 if (flushingbufs && bp->b_kvasize != 0) { 2382 bp->b_flags |= B_INVAL; 2383 bfreekva(bp); 2384 brelse(bp); 2385 goto restart; 2386 } 2387 if (bufspace < lobufspace) 2388 flushingbufs = 0; 2389 break; 2390 } 2391 return (bp); 2392} 2393 2394/* 2395 * getnewbuf: 2396 * 2397 * Find and initialize a new buffer header, freeing up existing buffers 2398 * in the bufqueues as necessary. The new buffer is returned locked. 2399 * 2400 * Important: B_INVAL is not set. If the caller wishes to throw the 2401 * buffer away, the caller must set B_INVAL prior to calling brelse(). 2402 * 2403 * We block if: 2404 * We have insufficient buffer headers 2405 * We have insufficient buffer space 2406 * buffer_arena is too fragmented ( space reservation fails ) 2407 * If we have to flush dirty buffers ( but we try to avoid this ) 2408 */ 2409static struct buf * 2410getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 2411 int gbflags) 2412{ 2413 struct buf *bp; 2414 int defrag, metadata; 2415 2416 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 2417 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 2418 if (!unmapped_buf_allowed) 2419 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 2420 2421 defrag = 0; 2422 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 2423 vp->v_type == VCHR) 2424 metadata = 1; 2425 else 2426 metadata = 0; 2427 /* 2428 * We can't afford to block since we might be holding a vnode lock, 2429 * which may prevent system daemons from running. We deal with 2430 * low-memory situations by proactively returning memory and running 2431 * async I/O rather then sync I/O. 2432 */ 2433 atomic_add_int(&getnewbufcalls, 1); 2434 atomic_subtract_int(&getnewbufrestarts, 1); 2435restart: 2436 bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED | 2437 GB_KVAALLOC)) == GB_UNMAPPED, metadata); 2438 if (bp != NULL) 2439 defrag = 0; 2440 2441 /* 2442 * If we exhausted our list, sleep as appropriate. We may have to 2443 * wakeup various daemons and write out some dirty buffers. 2444 * 2445 * Generally we are sleeping due to insufficient buffer space. 2446 */ 2447 if (bp == NULL) { 2448 mtx_assert(&bqclean, MA_OWNED); 2449 getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag); 2450 mtx_assert(&bqclean, MA_NOTOWNED); 2451 } else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) { 2452 mtx_assert(&bqclean, MA_NOTOWNED); 2453 2454 bfreekva(bp); 2455 bp->b_flags |= B_UNMAPPED; 2456 bp->b_kvabase = bp->b_data = unmapped_buf; 2457 bp->b_kvasize = maxsize; 2458 atomic_add_long(&bufspace, bp->b_kvasize); 2459 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 2460 atomic_add_int(&bufreusecnt, 1); 2461 } else { 2462 mtx_assert(&bqclean, MA_NOTOWNED); 2463 2464 /* 2465 * We finally have a valid bp. We aren't quite out of the 2466 * woods, we still have to reserve kva space. In order 2467 * to keep fragmentation sane we only allocate kva in 2468 * BKVASIZE chunks. 2469 */ 2470 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2471 2472 if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED | 2473 B_KVAALLOC)) == B_UNMAPPED) { 2474 if (allocbufkva(bp, maxsize, gbflags)) { 2475 defrag = 1; 2476 bp->b_flags |= B_INVAL; 2477 brelse(bp); 2478 goto restart; 2479 } 2480 atomic_add_int(&bufreusecnt, 1); 2481 } else if ((bp->b_flags & B_KVAALLOC) != 0 && 2482 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) { 2483 /* 2484 * If the reused buffer has KVA allocated, 2485 * reassign b_kvaalloc to b_kvabase. 2486 */ 2487 bp->b_kvabase = bp->b_kvaalloc; 2488 bp->b_flags &= ~B_KVAALLOC; 2489 atomic_subtract_long(&unmapped_bufspace, 2490 bp->b_kvasize); 2491 atomic_add_int(&bufreusecnt, 1); 2492 } else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 2493 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED | 2494 GB_KVAALLOC)) { 2495 /* 2496 * The case of reused buffer already have KVA 2497 * mapped, but the request is for unmapped 2498 * buffer with KVA allocated. 2499 */ 2500 bp->b_kvaalloc = bp->b_kvabase; 2501 bp->b_data = bp->b_kvabase = unmapped_buf; 2502 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 2503 atomic_add_long(&unmapped_bufspace, 2504 bp->b_kvasize); 2505 atomic_add_int(&bufreusecnt, 1); 2506 } 2507 if ((gbflags & GB_UNMAPPED) == 0) { 2508 bp->b_saveaddr = bp->b_kvabase; 2509 bp->b_data = bp->b_saveaddr; 2510 bp->b_flags &= ~B_UNMAPPED; 2511 BUF_CHECK_MAPPED(bp); 2512 } 2513 } 2514 return (bp); 2515} 2516 2517/* 2518 * buf_daemon: 2519 * 2520 * buffer flushing daemon. Buffers are normally flushed by the 2521 * update daemon but if it cannot keep up this process starts to 2522 * take the load in an attempt to prevent getnewbuf() from blocking. 2523 */ 2524 2525static struct kproc_desc buf_kp = { 2526 "bufdaemon", 2527 buf_daemon, 2528 &bufdaemonproc 2529}; 2530SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2531 2532static int 2533buf_flush(int target) 2534{ 2535 int flushed; 2536 2537 flushed = flushbufqueues(target, 0); 2538 if (flushed == 0) { 2539 /* 2540 * Could not find any buffers without rollback 2541 * dependencies, so just write the first one 2542 * in the hopes of eventually making progress. 2543 */ 2544 flushed = flushbufqueues(target, 1); 2545 } 2546 return (flushed); 2547} 2548 2549static void 2550buf_daemon() 2551{ 2552 int lodirty; 2553 2554 /* 2555 * This process needs to be suspended prior to shutdown sync. 2556 */ 2557 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2558 SHUTDOWN_PRI_LAST); 2559 2560 /* 2561 * This process is allowed to take the buffer cache to the limit 2562 */ 2563 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2564 mtx_lock(&bdlock); 2565 for (;;) { 2566 bd_request = 0; 2567 mtx_unlock(&bdlock); 2568 2569 kproc_suspend_check(bufdaemonproc); 2570 lodirty = lodirtybuffers; 2571 if (bd_speedupreq) { 2572 lodirty = numdirtybuffers / 2; 2573 bd_speedupreq = 0; 2574 } 2575 /* 2576 * Do the flush. Limit the amount of in-transit I/O we 2577 * allow to build up, otherwise we would completely saturate 2578 * the I/O system. 2579 */ 2580 while (numdirtybuffers > lodirty) { 2581 if (buf_flush(numdirtybuffers - lodirty) == 0) 2582 break; 2583 kern_yield(PRI_USER); 2584 } 2585 2586 /* 2587 * Only clear bd_request if we have reached our low water 2588 * mark. The buf_daemon normally waits 1 second and 2589 * then incrementally flushes any dirty buffers that have 2590 * built up, within reason. 2591 * 2592 * If we were unable to hit our low water mark and couldn't 2593 * find any flushable buffers, we sleep for a short period 2594 * to avoid endless loops on unlockable buffers. 2595 */ 2596 mtx_lock(&bdlock); 2597 if (numdirtybuffers <= lodirtybuffers) { 2598 /* 2599 * We reached our low water mark, reset the 2600 * request and sleep until we are needed again. 2601 * The sleep is just so the suspend code works. 2602 */ 2603 bd_request = 0; 2604 /* 2605 * Do an extra wakeup in case dirty threshold 2606 * changed via sysctl and the explicit transition 2607 * out of shortfall was missed. 2608 */ 2609 bdirtywakeup(); 2610 if (runningbufspace <= lorunningspace) 2611 runningwakeup(); 2612 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2613 } else { 2614 /* 2615 * We couldn't find any flushable dirty buffers but 2616 * still have too many dirty buffers, we 2617 * have to sleep and try again. (rare) 2618 */ 2619 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2620 } 2621 } 2622} 2623 2624/* 2625 * flushbufqueues: 2626 * 2627 * Try to flush a buffer in the dirty queue. We must be careful to 2628 * free up B_INVAL buffers instead of write them, which NFS is 2629 * particularly sensitive to. 2630 */ 2631static int flushwithdeps = 0; 2632SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2633 0, "Number of buffers flushed with dependecies that require rollbacks"); 2634 2635static int 2636flushbufqueues(int target, int flushdeps) 2637{ 2638 struct buf *sentinel; 2639 struct vnode *vp; 2640 struct mount *mp; 2641 struct buf *bp; 2642 int hasdeps; 2643 int flushed; 2644 int queue; 2645 int error; 2646 2647 flushed = 0; 2648 queue = QUEUE_DIRTY; 2649 bp = NULL; 2650 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2651 sentinel->b_qindex = QUEUE_SENTINEL; 2652 mtx_lock(&bqdirty); 2653 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2654 mtx_unlock(&bqdirty); 2655 while (flushed != target) { 2656 maybe_yield(); 2657 mtx_lock(&bqdirty); 2658 bp = TAILQ_NEXT(sentinel, b_freelist); 2659 if (bp != NULL) { 2660 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2661 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2662 b_freelist); 2663 } else { 2664 mtx_unlock(&bqdirty); 2665 break; 2666 } 2667 KASSERT(bp->b_qindex != QUEUE_SENTINEL, 2668 ("parallel calls to flushbufqueues() bp %p", bp)); 2669 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL); 2670 mtx_unlock(&bqdirty); 2671 if (error != 0) 2672 continue; 2673 if (bp->b_pin_count > 0) { 2674 BUF_UNLOCK(bp); 2675 continue; 2676 } 2677 /* 2678 * BKGRDINPROG can only be set with the buf and bufobj 2679 * locks both held. We tolerate a race to clear it here. 2680 */ 2681 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2682 (bp->b_flags & B_DELWRI) == 0) { 2683 BUF_UNLOCK(bp); 2684 continue; 2685 } 2686 if (bp->b_flags & B_INVAL) { 2687 bremfreef(bp); 2688 brelse(bp); 2689 flushed++; 2690 continue; 2691 } 2692 2693 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2694 if (flushdeps == 0) { 2695 BUF_UNLOCK(bp); 2696 continue; 2697 } 2698 hasdeps = 1; 2699 } else 2700 hasdeps = 0; 2701 /* 2702 * We must hold the lock on a vnode before writing 2703 * one of its buffers. Otherwise we may confuse, or 2704 * in the case of a snapshot vnode, deadlock the 2705 * system. 2706 * 2707 * The lock order here is the reverse of the normal 2708 * of vnode followed by buf lock. This is ok because 2709 * the NOWAIT will prevent deadlock. 2710 */ 2711 vp = bp->b_vp; 2712 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2713 BUF_UNLOCK(bp); 2714 continue; 2715 } 2716 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); 2717 if (error == 0) { 2718 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2719 bp, bp->b_vp, bp->b_flags); 2720 vfs_bio_awrite(bp); 2721 vn_finished_write(mp); 2722 VOP_UNLOCK(vp, 0); 2723 flushwithdeps += hasdeps; 2724 flushed++; 2725 if (runningbufspace > hirunningspace) 2726 waitrunningbufspace(); 2727 continue; 2728 } 2729 vn_finished_write(mp); 2730 BUF_UNLOCK(bp); 2731 } 2732 mtx_lock(&bqdirty); 2733 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2734 mtx_unlock(&bqdirty); 2735 free(sentinel, M_TEMP); 2736 return (flushed); 2737} 2738 2739/* 2740 * Check to see if a block is currently memory resident. 2741 */ 2742struct buf * 2743incore(struct bufobj *bo, daddr_t blkno) 2744{ 2745 struct buf *bp; 2746 2747 BO_RLOCK(bo); 2748 bp = gbincore(bo, blkno); 2749 BO_RUNLOCK(bo); 2750 return (bp); 2751} 2752 2753/* 2754 * Returns true if no I/O is needed to access the 2755 * associated VM object. This is like incore except 2756 * it also hunts around in the VM system for the data. 2757 */ 2758 2759static int 2760inmem(struct vnode * vp, daddr_t blkno) 2761{ 2762 vm_object_t obj; 2763 vm_offset_t toff, tinc, size; 2764 vm_page_t m; 2765 vm_ooffset_t off; 2766 2767 ASSERT_VOP_LOCKED(vp, "inmem"); 2768 2769 if (incore(&vp->v_bufobj, blkno)) 2770 return 1; 2771 if (vp->v_mount == NULL) 2772 return 0; 2773 obj = vp->v_object; 2774 if (obj == NULL) 2775 return (0); 2776 2777 size = PAGE_SIZE; 2778 if (size > vp->v_mount->mnt_stat.f_iosize) 2779 size = vp->v_mount->mnt_stat.f_iosize; 2780 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2781 2782 VM_OBJECT_RLOCK(obj); 2783 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2784 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2785 if (!m) 2786 goto notinmem; 2787 tinc = size; 2788 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2789 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2790 if (vm_page_is_valid(m, 2791 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2792 goto notinmem; 2793 } 2794 VM_OBJECT_RUNLOCK(obj); 2795 return 1; 2796 2797notinmem: 2798 VM_OBJECT_RUNLOCK(obj); 2799 return (0); 2800} 2801 2802/* 2803 * Set the dirty range for a buffer based on the status of the dirty 2804 * bits in the pages comprising the buffer. The range is limited 2805 * to the size of the buffer. 2806 * 2807 * Tell the VM system that the pages associated with this buffer 2808 * are clean. This is used for delayed writes where the data is 2809 * going to go to disk eventually without additional VM intevention. 2810 * 2811 * Note that while we only really need to clean through to b_bcount, we 2812 * just go ahead and clean through to b_bufsize. 2813 */ 2814static void 2815vfs_clean_pages_dirty_buf(struct buf *bp) 2816{ 2817 vm_ooffset_t foff, noff, eoff; 2818 vm_page_t m; 2819 int i; 2820 2821 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 2822 return; 2823 2824 foff = bp->b_offset; 2825 KASSERT(bp->b_offset != NOOFFSET, 2826 ("vfs_clean_pages_dirty_buf: no buffer offset")); 2827 2828 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 2829 vfs_drain_busy_pages(bp); 2830 vfs_setdirty_locked_object(bp); 2831 for (i = 0; i < bp->b_npages; i++) { 2832 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2833 eoff = noff; 2834 if (eoff > bp->b_offset + bp->b_bufsize) 2835 eoff = bp->b_offset + bp->b_bufsize; 2836 m = bp->b_pages[i]; 2837 vfs_page_set_validclean(bp, foff, m); 2838 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2839 foff = noff; 2840 } 2841 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 2842} 2843 2844static void 2845vfs_setdirty_locked_object(struct buf *bp) 2846{ 2847 vm_object_t object; 2848 int i; 2849 2850 object = bp->b_bufobj->bo_object; 2851 VM_OBJECT_ASSERT_WLOCKED(object); 2852 2853 /* 2854 * We qualify the scan for modified pages on whether the 2855 * object has been flushed yet. 2856 */ 2857 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { 2858 vm_offset_t boffset; 2859 vm_offset_t eoffset; 2860 2861 /* 2862 * test the pages to see if they have been modified directly 2863 * by users through the VM system. 2864 */ 2865 for (i = 0; i < bp->b_npages; i++) 2866 vm_page_test_dirty(bp->b_pages[i]); 2867 2868 /* 2869 * Calculate the encompassing dirty range, boffset and eoffset, 2870 * (eoffset - boffset) bytes. 2871 */ 2872 2873 for (i = 0; i < bp->b_npages; i++) { 2874 if (bp->b_pages[i]->dirty) 2875 break; 2876 } 2877 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2878 2879 for (i = bp->b_npages - 1; i >= 0; --i) { 2880 if (bp->b_pages[i]->dirty) { 2881 break; 2882 } 2883 } 2884 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2885 2886 /* 2887 * Fit it to the buffer. 2888 */ 2889 2890 if (eoffset > bp->b_bcount) 2891 eoffset = bp->b_bcount; 2892 2893 /* 2894 * If we have a good dirty range, merge with the existing 2895 * dirty range. 2896 */ 2897 2898 if (boffset < eoffset) { 2899 if (bp->b_dirtyoff > boffset) 2900 bp->b_dirtyoff = boffset; 2901 if (bp->b_dirtyend < eoffset) 2902 bp->b_dirtyend = eoffset; 2903 } 2904 } 2905} 2906 2907/* 2908 * Allocate the KVA mapping for an existing buffer. It handles the 2909 * cases of both B_UNMAPPED buffer, and buffer with the preallocated 2910 * KVA which is not mapped (B_KVAALLOC). 2911 */ 2912static void 2913bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 2914{ 2915 struct buf *scratch_bp; 2916 int bsize, maxsize, need_mapping, need_kva; 2917 off_t offset; 2918 2919 need_mapping = (bp->b_flags & B_UNMAPPED) != 0 && 2920 (gbflags & GB_UNMAPPED) == 0; 2921 need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED && 2922 (gbflags & GB_KVAALLOC) != 0; 2923 if (!need_mapping && !need_kva) 2924 return; 2925 2926 BUF_CHECK_UNMAPPED(bp); 2927 2928 if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) { 2929 /* 2930 * Buffer is not mapped, but the KVA was already 2931 * reserved at the time of the instantiation. Use the 2932 * allocated space. 2933 */ 2934 bp->b_flags &= ~B_KVAALLOC; 2935 KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0")); 2936 bp->b_kvabase = bp->b_kvaalloc; 2937 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize); 2938 goto has_addr; 2939 } 2940 2941 /* 2942 * Calculate the amount of the address space we would reserve 2943 * if the buffer was mapped. 2944 */ 2945 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 2946 offset = blkno * bsize; 2947 maxsize = size + (offset & PAGE_MASK); 2948 maxsize = imax(maxsize, bsize); 2949 2950mapping_loop: 2951 if (allocbufkva(bp, maxsize, gbflags)) { 2952 /* 2953 * Request defragmentation. getnewbuf() returns us the 2954 * allocated space by the scratch buffer KVA. 2955 */ 2956 scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags | 2957 (GB_UNMAPPED | GB_KVAALLOC)); 2958 if (scratch_bp == NULL) { 2959 if ((gbflags & GB_NOWAIT_BD) != 0) { 2960 /* 2961 * XXXKIB: defragmentation cannot 2962 * succeed, not sure what else to do. 2963 */ 2964 panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp); 2965 } 2966 atomic_add_int(&mappingrestarts, 1); 2967 goto mapping_loop; 2968 } 2969 KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0, 2970 ("scratch bp !B_KVAALLOC %p", scratch_bp)); 2971 setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc, 2972 scratch_bp->b_kvasize, gbflags); 2973 2974 /* Get rid of the scratch buffer. */ 2975 scratch_bp->b_kvasize = 0; 2976 scratch_bp->b_flags |= B_INVAL; 2977 scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC); 2978 brelse(scratch_bp); 2979 } 2980 if (!need_mapping) 2981 return; 2982 2983has_addr: 2984 bp->b_saveaddr = bp->b_kvabase; 2985 bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */ 2986 bp->b_flags &= ~B_UNMAPPED; 2987 BUF_CHECK_MAPPED(bp); 2988 bpmap_qenter(bp); 2989} 2990 2991/* 2992 * getblk: 2993 * 2994 * Get a block given a specified block and offset into a file/device. 2995 * The buffers B_DONE bit will be cleared on return, making it almost 2996 * ready for an I/O initiation. B_INVAL may or may not be set on 2997 * return. The caller should clear B_INVAL prior to initiating a 2998 * READ. 2999 * 3000 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 3001 * an existing buffer. 3002 * 3003 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 3004 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 3005 * and then cleared based on the backing VM. If the previous buffer is 3006 * non-0-sized but invalid, B_CACHE will be cleared. 3007 * 3008 * If getblk() must create a new buffer, the new buffer is returned with 3009 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 3010 * case it is returned with B_INVAL clear and B_CACHE set based on the 3011 * backing VM. 3012 * 3013 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 3014 * B_CACHE bit is clear. 3015 * 3016 * What this means, basically, is that the caller should use B_CACHE to 3017 * determine whether the buffer is fully valid or not and should clear 3018 * B_INVAL prior to issuing a read. If the caller intends to validate 3019 * the buffer by loading its data area with something, the caller needs 3020 * to clear B_INVAL. If the caller does this without issuing an I/O, 3021 * the caller should set B_CACHE ( as an optimization ), else the caller 3022 * should issue the I/O and biodone() will set B_CACHE if the I/O was 3023 * a write attempt or if it was a successfull read. If the caller 3024 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 3025 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 3026 */ 3027struct buf * 3028getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 3029 int flags) 3030{ 3031 struct buf *bp; 3032 struct bufobj *bo; 3033 int bsize, error, maxsize, vmio; 3034 off_t offset; 3035 3036 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 3037 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3038 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3039 ASSERT_VOP_LOCKED(vp, "getblk"); 3040 if (size > MAXBSIZE) 3041 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 3042 if (!unmapped_buf_allowed) 3043 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3044 3045 bo = &vp->v_bufobj; 3046loop: 3047 BO_RLOCK(bo); 3048 bp = gbincore(bo, blkno); 3049 if (bp != NULL) { 3050 int lockflags; 3051 /* 3052 * Buffer is in-core. If the buffer is not busy nor managed, 3053 * it must be on a queue. 3054 */ 3055 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 3056 3057 if (flags & GB_LOCK_NOWAIT) 3058 lockflags |= LK_NOWAIT; 3059 3060 error = BUF_TIMELOCK(bp, lockflags, 3061 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); 3062 3063 /* 3064 * If we slept and got the lock we have to restart in case 3065 * the buffer changed identities. 3066 */ 3067 if (error == ENOLCK) 3068 goto loop; 3069 /* We timed out or were interrupted. */ 3070 else if (error) 3071 return (NULL); 3072 /* If recursed, assume caller knows the rules. */ 3073 else if (BUF_LOCKRECURSED(bp)) 3074 goto end; 3075 3076 /* 3077 * The buffer is locked. B_CACHE is cleared if the buffer is 3078 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 3079 * and for a VMIO buffer B_CACHE is adjusted according to the 3080 * backing VM cache. 3081 */ 3082 if (bp->b_flags & B_INVAL) 3083 bp->b_flags &= ~B_CACHE; 3084 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 3085 bp->b_flags |= B_CACHE; 3086 if (bp->b_flags & B_MANAGED) 3087 MPASS(bp->b_qindex == QUEUE_NONE); 3088 else 3089 bremfree(bp); 3090 3091 /* 3092 * check for size inconsistencies for non-VMIO case. 3093 */ 3094 if (bp->b_bcount != size) { 3095 if ((bp->b_flags & B_VMIO) == 0 || 3096 (size > bp->b_kvasize)) { 3097 if (bp->b_flags & B_DELWRI) { 3098 /* 3099 * If buffer is pinned and caller does 3100 * not want sleep waiting for it to be 3101 * unpinned, bail out 3102 * */ 3103 if (bp->b_pin_count > 0) { 3104 if (flags & GB_LOCK_NOWAIT) { 3105 bqrelse(bp); 3106 return (NULL); 3107 } else { 3108 bunpin_wait(bp); 3109 } 3110 } 3111 bp->b_flags |= B_NOCACHE; 3112 bwrite(bp); 3113 } else { 3114 if (LIST_EMPTY(&bp->b_dep)) { 3115 bp->b_flags |= B_RELBUF; 3116 brelse(bp); 3117 } else { 3118 bp->b_flags |= B_NOCACHE; 3119 bwrite(bp); 3120 } 3121 } 3122 goto loop; 3123 } 3124 } 3125 3126 /* 3127 * Handle the case of unmapped buffer which should 3128 * become mapped, or the buffer for which KVA 3129 * reservation is requested. 3130 */ 3131 bp_unmapped_get_kva(bp, blkno, size, flags); 3132 3133 /* 3134 * If the size is inconsistant in the VMIO case, we can resize 3135 * the buffer. This might lead to B_CACHE getting set or 3136 * cleared. If the size has not changed, B_CACHE remains 3137 * unchanged from its previous state. 3138 */ 3139 if (bp->b_bcount != size) 3140 allocbuf(bp, size); 3141 3142 KASSERT(bp->b_offset != NOOFFSET, 3143 ("getblk: no buffer offset")); 3144 3145 /* 3146 * A buffer with B_DELWRI set and B_CACHE clear must 3147 * be committed before we can return the buffer in 3148 * order to prevent the caller from issuing a read 3149 * ( due to B_CACHE not being set ) and overwriting 3150 * it. 3151 * 3152 * Most callers, including NFS and FFS, need this to 3153 * operate properly either because they assume they 3154 * can issue a read if B_CACHE is not set, or because 3155 * ( for example ) an uncached B_DELWRI might loop due 3156 * to softupdates re-dirtying the buffer. In the latter 3157 * case, B_CACHE is set after the first write completes, 3158 * preventing further loops. 3159 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3160 * above while extending the buffer, we cannot allow the 3161 * buffer to remain with B_CACHE set after the write 3162 * completes or it will represent a corrupt state. To 3163 * deal with this we set B_NOCACHE to scrap the buffer 3164 * after the write. 3165 * 3166 * We might be able to do something fancy, like setting 3167 * B_CACHE in bwrite() except if B_DELWRI is already set, 3168 * so the below call doesn't set B_CACHE, but that gets real 3169 * confusing. This is much easier. 3170 */ 3171 3172 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3173 bp->b_flags |= B_NOCACHE; 3174 bwrite(bp); 3175 goto loop; 3176 } 3177 bp->b_flags &= ~B_DONE; 3178 } else { 3179 /* 3180 * Buffer is not in-core, create new buffer. The buffer 3181 * returned by getnewbuf() is locked. Note that the returned 3182 * buffer is also considered valid (not marked B_INVAL). 3183 */ 3184 BO_RUNLOCK(bo); 3185 /* 3186 * If the user does not want us to create the buffer, bail out 3187 * here. 3188 */ 3189 if (flags & GB_NOCREAT) 3190 return NULL; 3191 if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread)) 3192 return NULL; 3193 3194 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 3195 offset = blkno * bsize; 3196 vmio = vp->v_object != NULL; 3197 if (vmio) { 3198 maxsize = size + (offset & PAGE_MASK); 3199 } else { 3200 maxsize = size; 3201 /* Do not allow non-VMIO notmapped buffers. */ 3202 flags &= ~GB_UNMAPPED; 3203 } 3204 maxsize = imax(maxsize, bsize); 3205 3206 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 3207 if (bp == NULL) { 3208 if (slpflag || slptimeo) 3209 return NULL; 3210 goto loop; 3211 } 3212 3213 /* 3214 * This code is used to make sure that a buffer is not 3215 * created while the getnewbuf routine is blocked. 3216 * This can be a problem whether the vnode is locked or not. 3217 * If the buffer is created out from under us, we have to 3218 * throw away the one we just created. 3219 * 3220 * Note: this must occur before we associate the buffer 3221 * with the vp especially considering limitations in 3222 * the splay tree implementation when dealing with duplicate 3223 * lblkno's. 3224 */ 3225 BO_LOCK(bo); 3226 if (gbincore(bo, blkno)) { 3227 BO_UNLOCK(bo); 3228 bp->b_flags |= B_INVAL; 3229 brelse(bp); 3230 goto loop; 3231 } 3232 3233 /* 3234 * Insert the buffer into the hash, so that it can 3235 * be found by incore. 3236 */ 3237 bp->b_blkno = bp->b_lblkno = blkno; 3238 bp->b_offset = offset; 3239 bgetvp(vp, bp); 3240 BO_UNLOCK(bo); 3241 3242 /* 3243 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 3244 * buffer size starts out as 0, B_CACHE will be set by 3245 * allocbuf() for the VMIO case prior to it testing the 3246 * backing store for validity. 3247 */ 3248 3249 if (vmio) { 3250 bp->b_flags |= B_VMIO; 3251 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 3252 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 3253 bp, vp->v_object, bp->b_bufobj->bo_object)); 3254 } else { 3255 bp->b_flags &= ~B_VMIO; 3256 KASSERT(bp->b_bufobj->bo_object == NULL, 3257 ("ARGH! has b_bufobj->bo_object %p %p\n", 3258 bp, bp->b_bufobj->bo_object)); 3259 BUF_CHECK_MAPPED(bp); 3260 } 3261 3262 allocbuf(bp, size); 3263 bp->b_flags &= ~B_DONE; 3264 } 3265 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 3266 BUF_ASSERT_HELD(bp); 3267end: 3268 KASSERT(bp->b_bufobj == bo, 3269 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3270 return (bp); 3271} 3272 3273/* 3274 * Get an empty, disassociated buffer of given size. The buffer is initially 3275 * set to B_INVAL. 3276 */ 3277struct buf * 3278geteblk(int size, int flags) 3279{ 3280 struct buf *bp; 3281 int maxsize; 3282 3283 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3284 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 3285 if ((flags & GB_NOWAIT_BD) && 3286 (curthread->td_pflags & TDP_BUFNEED) != 0) 3287 return (NULL); 3288 } 3289 allocbuf(bp, size); 3290 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3291 BUF_ASSERT_HELD(bp); 3292 return (bp); 3293} 3294 3295 3296/* 3297 * This code constitutes the buffer memory from either anonymous system 3298 * memory (in the case of non-VMIO operations) or from an associated 3299 * VM object (in the case of VMIO operations). This code is able to 3300 * resize a buffer up or down. 3301 * 3302 * Note that this code is tricky, and has many complications to resolve 3303 * deadlock or inconsistant data situations. Tread lightly!!! 3304 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3305 * the caller. Calling this code willy nilly can result in the loss of data. 3306 * 3307 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3308 * B_CACHE for the non-VMIO case. 3309 */ 3310 3311int 3312allocbuf(struct buf *bp, int size) 3313{ 3314 int newbsize, mbsize; 3315 int i; 3316 3317 BUF_ASSERT_HELD(bp); 3318 3319 if (bp->b_kvasize < size) 3320 panic("allocbuf: buffer too small"); 3321 3322 if ((bp->b_flags & B_VMIO) == 0) { 3323 caddr_t origbuf; 3324 int origbufsize; 3325 /* 3326 * Just get anonymous memory from the kernel. Don't 3327 * mess with B_CACHE. 3328 */ 3329 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3330 if (bp->b_flags & B_MALLOC) 3331 newbsize = mbsize; 3332 else 3333 newbsize = round_page(size); 3334 3335 if (newbsize < bp->b_bufsize) { 3336 /* 3337 * malloced buffers are not shrunk 3338 */ 3339 if (bp->b_flags & B_MALLOC) { 3340 if (newbsize) { 3341 bp->b_bcount = size; 3342 } else { 3343 free(bp->b_data, M_BIOBUF); 3344 if (bp->b_bufsize) { 3345 atomic_subtract_long( 3346 &bufmallocspace, 3347 bp->b_bufsize); 3348 bufspacewakeup(); 3349 bp->b_bufsize = 0; 3350 } 3351 bp->b_saveaddr = bp->b_kvabase; 3352 bp->b_data = bp->b_saveaddr; 3353 bp->b_bcount = 0; 3354 bp->b_flags &= ~B_MALLOC; 3355 } 3356 return 1; 3357 } 3358 vm_hold_free_pages(bp, newbsize); 3359 } else if (newbsize > bp->b_bufsize) { 3360 /* 3361 * We only use malloced memory on the first allocation. 3362 * and revert to page-allocated memory when the buffer 3363 * grows. 3364 */ 3365 /* 3366 * There is a potential smp race here that could lead 3367 * to bufmallocspace slightly passing the max. It 3368 * is probably extremely rare and not worth worrying 3369 * over. 3370 */ 3371 if ( (bufmallocspace < maxbufmallocspace) && 3372 (bp->b_bufsize == 0) && 3373 (mbsize <= PAGE_SIZE/2)) { 3374 3375 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 3376 bp->b_bufsize = mbsize; 3377 bp->b_bcount = size; 3378 bp->b_flags |= B_MALLOC; 3379 atomic_add_long(&bufmallocspace, mbsize); 3380 return 1; 3381 } 3382 origbuf = NULL; 3383 origbufsize = 0; 3384 /* 3385 * If the buffer is growing on its other-than-first allocation, 3386 * then we revert to the page-allocation scheme. 3387 */ 3388 if (bp->b_flags & B_MALLOC) { 3389 origbuf = bp->b_data; 3390 origbufsize = bp->b_bufsize; 3391 bp->b_data = bp->b_kvabase; 3392 if (bp->b_bufsize) { 3393 atomic_subtract_long(&bufmallocspace, 3394 bp->b_bufsize); 3395 bufspacewakeup(); 3396 bp->b_bufsize = 0; 3397 } 3398 bp->b_flags &= ~B_MALLOC; 3399 newbsize = round_page(newbsize); 3400 } 3401 vm_hold_load_pages( 3402 bp, 3403 (vm_offset_t) bp->b_data + bp->b_bufsize, 3404 (vm_offset_t) bp->b_data + newbsize); 3405 if (origbuf) { 3406 bcopy(origbuf, bp->b_data, origbufsize); 3407 free(origbuf, M_BIOBUF); 3408 } 3409 } 3410 } else { 3411 int desiredpages; 3412 3413 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3414 desiredpages = (size == 0) ? 0 : 3415 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 3416 3417 if (bp->b_flags & B_MALLOC) 3418 panic("allocbuf: VMIO buffer can't be malloced"); 3419 /* 3420 * Set B_CACHE initially if buffer is 0 length or will become 3421 * 0-length. 3422 */ 3423 if (size == 0 || bp->b_bufsize == 0) 3424 bp->b_flags |= B_CACHE; 3425 3426 if (newbsize < bp->b_bufsize) { 3427 /* 3428 * DEV_BSIZE aligned new buffer size is less then the 3429 * DEV_BSIZE aligned existing buffer size. Figure out 3430 * if we have to remove any pages. 3431 */ 3432 if (desiredpages < bp->b_npages) { 3433 vm_page_t m; 3434 3435 if ((bp->b_flags & B_UNMAPPED) == 0) { 3436 BUF_CHECK_MAPPED(bp); 3437 pmap_qremove((vm_offset_t)trunc_page( 3438 (vm_offset_t)bp->b_data) + 3439 (desiredpages << PAGE_SHIFT), 3440 (bp->b_npages - desiredpages)); 3441 } else 3442 BUF_CHECK_UNMAPPED(bp); 3443 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3444 for (i = desiredpages; i < bp->b_npages; i++) { 3445 /* 3446 * the page is not freed here -- it 3447 * is the responsibility of 3448 * vnode_pager_setsize 3449 */ 3450 m = bp->b_pages[i]; 3451 KASSERT(m != bogus_page, 3452 ("allocbuf: bogus page found")); 3453 while (vm_page_sleep_if_busy(m, 3454 "biodep")) 3455 continue; 3456 3457 bp->b_pages[i] = NULL; 3458 vm_page_lock(m); 3459 vm_page_unwire(m, 0); 3460 vm_page_unlock(m); 3461 } 3462 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3463 bp->b_npages = desiredpages; 3464 } 3465 } else if (size > bp->b_bcount) { 3466 /* 3467 * We are growing the buffer, possibly in a 3468 * byte-granular fashion. 3469 */ 3470 vm_object_t obj; 3471 vm_offset_t toff; 3472 vm_offset_t tinc; 3473 3474 /* 3475 * Step 1, bring in the VM pages from the object, 3476 * allocating them if necessary. We must clear 3477 * B_CACHE if these pages are not valid for the 3478 * range covered by the buffer. 3479 */ 3480 3481 obj = bp->b_bufobj->bo_object; 3482 3483 VM_OBJECT_WLOCK(obj); 3484 while (bp->b_npages < desiredpages) { 3485 vm_page_t m; 3486 3487 /* 3488 * We must allocate system pages since blocking 3489 * here could interfere with paging I/O, no 3490 * matter which process we are. 3491 * 3492 * Only exclusive busy can be tested here. 3493 * Blocking on shared busy might lead to 3494 * deadlocks once allocbuf() is called after 3495 * pages are vfs_busy_pages(). 3496 */ 3497 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + 3498 bp->b_npages, VM_ALLOC_NOBUSY | 3499 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3500 VM_ALLOC_IGN_SBUSY | 3501 VM_ALLOC_COUNT(desiredpages - bp->b_npages)); 3502 if (m->valid == 0) 3503 bp->b_flags &= ~B_CACHE; 3504 bp->b_pages[bp->b_npages] = m; 3505 ++bp->b_npages; 3506 } 3507 3508 /* 3509 * Step 2. We've loaded the pages into the buffer, 3510 * we have to figure out if we can still have B_CACHE 3511 * set. Note that B_CACHE is set according to the 3512 * byte-granular range ( bcount and size ), new the 3513 * aligned range ( newbsize ). 3514 * 3515 * The VM test is against m->valid, which is DEV_BSIZE 3516 * aligned. Needless to say, the validity of the data 3517 * needs to also be DEV_BSIZE aligned. Note that this 3518 * fails with NFS if the server or some other client 3519 * extends the file's EOF. If our buffer is resized, 3520 * B_CACHE may remain set! XXX 3521 */ 3522 3523 toff = bp->b_bcount; 3524 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3525 3526 while ((bp->b_flags & B_CACHE) && toff < size) { 3527 vm_pindex_t pi; 3528 3529 if (tinc > (size - toff)) 3530 tinc = size - toff; 3531 3532 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3533 PAGE_SHIFT; 3534 3535 vfs_buf_test_cache( 3536 bp, 3537 bp->b_offset, 3538 toff, 3539 tinc, 3540 bp->b_pages[pi] 3541 ); 3542 toff += tinc; 3543 tinc = PAGE_SIZE; 3544 } 3545 VM_OBJECT_WUNLOCK(obj); 3546 3547 /* 3548 * Step 3, fixup the KVM pmap. 3549 */ 3550 if ((bp->b_flags & B_UNMAPPED) == 0) 3551 bpmap_qenter(bp); 3552 else 3553 BUF_CHECK_UNMAPPED(bp); 3554 } 3555 } 3556 if (newbsize < bp->b_bufsize) 3557 bufspacewakeup(); 3558 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3559 bp->b_bcount = size; /* requested buffer size */ 3560 return 1; 3561} 3562 3563extern int inflight_transient_maps; 3564 3565void 3566biodone(struct bio *bp) 3567{ 3568 struct mtx *mtxp; 3569 void (*done)(struct bio *); 3570 vm_offset_t start, end; 3571 3572 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 3573 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING; 3574 bp->bio_flags |= BIO_UNMAPPED; 3575 start = trunc_page((vm_offset_t)bp->bio_data); 3576 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 3577 pmap_qremove(start, OFF_TO_IDX(end - start)); 3578 vmem_free(transient_arena, start, end - start); 3579 atomic_add_int(&inflight_transient_maps, -1); 3580 } 3581 done = bp->bio_done; 3582 if (done == NULL) { 3583 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3584 mtx_lock(mtxp); 3585 bp->bio_flags |= BIO_DONE; 3586 wakeup(bp); 3587 mtx_unlock(mtxp); 3588 } else { 3589 bp->bio_flags |= BIO_DONE; 3590 done(bp); 3591 } 3592} 3593 3594/* 3595 * Wait for a BIO to finish. 3596 */ 3597int 3598biowait(struct bio *bp, const char *wchan) 3599{ 3600 struct mtx *mtxp; 3601 3602 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3603 mtx_lock(mtxp); 3604 while ((bp->bio_flags & BIO_DONE) == 0) 3605 msleep(bp, mtxp, PRIBIO, wchan, 0); 3606 mtx_unlock(mtxp); 3607 if (bp->bio_error != 0) 3608 return (bp->bio_error); 3609 if (!(bp->bio_flags & BIO_ERROR)) 3610 return (0); 3611 return (EIO); 3612} 3613 3614void 3615biofinish(struct bio *bp, struct devstat *stat, int error) 3616{ 3617 3618 if (error) { 3619 bp->bio_error = error; 3620 bp->bio_flags |= BIO_ERROR; 3621 } 3622 if (stat != NULL) 3623 devstat_end_transaction_bio(stat, bp); 3624 biodone(bp); 3625} 3626 3627/* 3628 * bufwait: 3629 * 3630 * Wait for buffer I/O completion, returning error status. The buffer 3631 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3632 * error and cleared. 3633 */ 3634int 3635bufwait(struct buf *bp) 3636{ 3637 if (bp->b_iocmd == BIO_READ) 3638 bwait(bp, PRIBIO, "biord"); 3639 else 3640 bwait(bp, PRIBIO, "biowr"); 3641 if (bp->b_flags & B_EINTR) { 3642 bp->b_flags &= ~B_EINTR; 3643 return (EINTR); 3644 } 3645 if (bp->b_ioflags & BIO_ERROR) { 3646 return (bp->b_error ? bp->b_error : EIO); 3647 } else { 3648 return (0); 3649 } 3650} 3651 3652 /* 3653 * Call back function from struct bio back up to struct buf. 3654 */ 3655static void 3656bufdonebio(struct bio *bip) 3657{ 3658 struct buf *bp; 3659 3660 bp = bip->bio_caller2; 3661 bp->b_resid = bp->b_bcount - bip->bio_completed; 3662 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3663 bp->b_ioflags = bip->bio_flags; 3664 bp->b_error = bip->bio_error; 3665 if (bp->b_error) 3666 bp->b_ioflags |= BIO_ERROR; 3667 bufdone(bp); 3668 g_destroy_bio(bip); 3669} 3670 3671void 3672dev_strategy(struct cdev *dev, struct buf *bp) 3673{ 3674 struct cdevsw *csw; 3675 int ref; 3676 3677 KASSERT(dev->si_refcount > 0, 3678 ("dev_strategy on un-referenced struct cdev *(%s) %p", 3679 devtoname(dev), dev)); 3680 3681 csw = dev_refthread(dev, &ref); 3682 dev_strategy_csw(dev, csw, bp); 3683 dev_relthread(dev, ref); 3684} 3685 3686void 3687dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp) 3688{ 3689 struct bio *bip; 3690 3691 KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE, 3692 ("b_iocmd botch")); 3693 KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) || 3694 dev->si_threadcount > 0, 3695 ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev), 3696 dev)); 3697 if (csw == NULL) { 3698 bp->b_error = ENXIO; 3699 bp->b_ioflags = BIO_ERROR; 3700 bufdone(bp); 3701 return; 3702 } 3703 for (;;) { 3704 bip = g_new_bio(); 3705 if (bip != NULL) 3706 break; 3707 /* Try again later */ 3708 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3709 } 3710 bip->bio_cmd = bp->b_iocmd; 3711 bip->bio_offset = bp->b_iooffset; 3712 bip->bio_length = bp->b_bcount; 3713 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3714 bdata2bio(bp, bip); 3715 bip->bio_done = bufdonebio; 3716 bip->bio_caller2 = bp; 3717 bip->bio_dev = dev; 3718 (*csw->d_strategy)(bip); 3719} 3720 3721/* 3722 * bufdone: 3723 * 3724 * Finish I/O on a buffer, optionally calling a completion function. 3725 * This is usually called from an interrupt so process blocking is 3726 * not allowed. 3727 * 3728 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3729 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3730 * assuming B_INVAL is clear. 3731 * 3732 * For the VMIO case, we set B_CACHE if the op was a read and no 3733 * read error occured, or if the op was a write. B_CACHE is never 3734 * set if the buffer is invalid or otherwise uncacheable. 3735 * 3736 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3737 * initiator to leave B_INVAL set to brelse the buffer out of existance 3738 * in the biodone routine. 3739 */ 3740void 3741bufdone(struct buf *bp) 3742{ 3743 struct bufobj *dropobj; 3744 void (*biodone)(struct buf *); 3745 3746 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3747 dropobj = NULL; 3748 3749 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3750 BUF_ASSERT_HELD(bp); 3751 3752 runningbufwakeup(bp); 3753 if (bp->b_iocmd == BIO_WRITE) 3754 dropobj = bp->b_bufobj; 3755 /* call optional completion function if requested */ 3756 if (bp->b_iodone != NULL) { 3757 biodone = bp->b_iodone; 3758 bp->b_iodone = NULL; 3759 (*biodone) (bp); 3760 if (dropobj) 3761 bufobj_wdrop(dropobj); 3762 return; 3763 } 3764 3765 bufdone_finish(bp); 3766 3767 if (dropobj) 3768 bufobj_wdrop(dropobj); 3769} 3770 3771void 3772bufdone_finish(struct buf *bp) 3773{ 3774 BUF_ASSERT_HELD(bp); 3775 3776 if (!LIST_EMPTY(&bp->b_dep)) 3777 buf_complete(bp); 3778 3779 if (bp->b_flags & B_VMIO) { 3780 vm_ooffset_t foff; 3781 vm_page_t m; 3782 vm_object_t obj; 3783 struct vnode *vp; 3784 int bogus, i, iosize; 3785 3786 obj = bp->b_bufobj->bo_object; 3787 KASSERT(obj->paging_in_progress >= bp->b_npages, 3788 ("biodone_finish: paging in progress(%d) < b_npages(%d)", 3789 obj->paging_in_progress, bp->b_npages)); 3790 3791 vp = bp->b_vp; 3792 KASSERT(vp->v_holdcnt > 0, 3793 ("biodone_finish: vnode %p has zero hold count", vp)); 3794 KASSERT(vp->v_object != NULL, 3795 ("biodone_finish: vnode %p has no vm_object", vp)); 3796 3797 foff = bp->b_offset; 3798 KASSERT(bp->b_offset != NOOFFSET, 3799 ("biodone_finish: bp %p has no buffer offset", bp)); 3800 3801 /* 3802 * Set B_CACHE if the op was a normal read and no error 3803 * occured. B_CACHE is set for writes in the b*write() 3804 * routines. 3805 */ 3806 iosize = bp->b_bcount - bp->b_resid; 3807 if (bp->b_iocmd == BIO_READ && 3808 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3809 !(bp->b_ioflags & BIO_ERROR)) { 3810 bp->b_flags |= B_CACHE; 3811 } 3812 bogus = 0; 3813 VM_OBJECT_WLOCK(obj); 3814 for (i = 0; i < bp->b_npages; i++) { 3815 int bogusflag = 0; 3816 int resid; 3817 3818 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3819 if (resid > iosize) 3820 resid = iosize; 3821 3822 /* 3823 * cleanup bogus pages, restoring the originals 3824 */ 3825 m = bp->b_pages[i]; 3826 if (m == bogus_page) { 3827 bogus = bogusflag = 1; 3828 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3829 if (m == NULL) 3830 panic("biodone: page disappeared!"); 3831 bp->b_pages[i] = m; 3832 } 3833 KASSERT(OFF_TO_IDX(foff) == m->pindex, 3834 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch", 3835 (intmax_t)foff, (uintmax_t)m->pindex)); 3836 3837 /* 3838 * In the write case, the valid and clean bits are 3839 * already changed correctly ( see bdwrite() ), so we 3840 * only need to do this here in the read case. 3841 */ 3842 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3843 KASSERT((m->dirty & vm_page_bits(foff & 3844 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3845 " page %p has unexpected dirty bits", m)); 3846 vfs_page_set_valid(bp, foff, m); 3847 } 3848 3849 vm_page_sunbusy(m); 3850 vm_object_pip_subtract(obj, 1); 3851 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3852 iosize -= resid; 3853 } 3854 vm_object_pip_wakeupn(obj, 0); 3855 VM_OBJECT_WUNLOCK(obj); 3856 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 3857 BUF_CHECK_MAPPED(bp); 3858 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3859 bp->b_pages, bp->b_npages); 3860 } 3861 } 3862 3863 /* 3864 * For asynchronous completions, release the buffer now. The brelse 3865 * will do a wakeup there if necessary - so no need to do a wakeup 3866 * here in the async case. The sync case always needs to do a wakeup. 3867 */ 3868 3869 if (bp->b_flags & B_ASYNC) { 3870 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3871 brelse(bp); 3872 else 3873 bqrelse(bp); 3874 } else 3875 bdone(bp); 3876} 3877 3878/* 3879 * This routine is called in lieu of iodone in the case of 3880 * incomplete I/O. This keeps the busy status for pages 3881 * consistant. 3882 */ 3883void 3884vfs_unbusy_pages(struct buf *bp) 3885{ 3886 int i; 3887 vm_object_t obj; 3888 vm_page_t m; 3889 3890 runningbufwakeup(bp); 3891 if (!(bp->b_flags & B_VMIO)) 3892 return; 3893 3894 obj = bp->b_bufobj->bo_object; 3895 VM_OBJECT_WLOCK(obj); 3896 for (i = 0; i < bp->b_npages; i++) { 3897 m = bp->b_pages[i]; 3898 if (m == bogus_page) { 3899 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3900 if (!m) 3901 panic("vfs_unbusy_pages: page missing\n"); 3902 bp->b_pages[i] = m; 3903 if ((bp->b_flags & B_UNMAPPED) == 0) { 3904 BUF_CHECK_MAPPED(bp); 3905 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3906 bp->b_pages, bp->b_npages); 3907 } else 3908 BUF_CHECK_UNMAPPED(bp); 3909 } 3910 vm_object_pip_subtract(obj, 1); 3911 vm_page_sunbusy(m); 3912 } 3913 vm_object_pip_wakeupn(obj, 0); 3914 VM_OBJECT_WUNLOCK(obj); 3915} 3916 3917/* 3918 * vfs_page_set_valid: 3919 * 3920 * Set the valid bits in a page based on the supplied offset. The 3921 * range is restricted to the buffer's size. 3922 * 3923 * This routine is typically called after a read completes. 3924 */ 3925static void 3926vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3927{ 3928 vm_ooffset_t eoff; 3929 3930 /* 3931 * Compute the end offset, eoff, such that [off, eoff) does not span a 3932 * page boundary and eoff is not greater than the end of the buffer. 3933 * The end of the buffer, in this case, is our file EOF, not the 3934 * allocation size of the buffer. 3935 */ 3936 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3937 if (eoff > bp->b_offset + bp->b_bcount) 3938 eoff = bp->b_offset + bp->b_bcount; 3939 3940 /* 3941 * Set valid range. This is typically the entire buffer and thus the 3942 * entire page. 3943 */ 3944 if (eoff > off) 3945 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 3946} 3947 3948/* 3949 * vfs_page_set_validclean: 3950 * 3951 * Set the valid bits and clear the dirty bits in a page based on the 3952 * supplied offset. The range is restricted to the buffer's size. 3953 */ 3954static void 3955vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3956{ 3957 vm_ooffset_t soff, eoff; 3958 3959 /* 3960 * Start and end offsets in buffer. eoff - soff may not cross a 3961 * page boundry or cross the end of the buffer. The end of the 3962 * buffer, in this case, is our file EOF, not the allocation size 3963 * of the buffer. 3964 */ 3965 soff = off; 3966 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3967 if (eoff > bp->b_offset + bp->b_bcount) 3968 eoff = bp->b_offset + bp->b_bcount; 3969 3970 /* 3971 * Set valid range. This is typically the entire buffer and thus the 3972 * entire page. 3973 */ 3974 if (eoff > soff) { 3975 vm_page_set_validclean( 3976 m, 3977 (vm_offset_t) (soff & PAGE_MASK), 3978 (vm_offset_t) (eoff - soff) 3979 ); 3980 } 3981} 3982 3983/* 3984 * Ensure that all buffer pages are not exclusive busied. If any page is 3985 * exclusive busy, drain it. 3986 */ 3987void 3988vfs_drain_busy_pages(struct buf *bp) 3989{ 3990 vm_page_t m; 3991 int i, last_busied; 3992 3993 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); 3994 last_busied = 0; 3995 for (i = 0; i < bp->b_npages; i++) { 3996 m = bp->b_pages[i]; 3997 if (vm_page_xbusied(m)) { 3998 for (; last_busied < i; last_busied++) 3999 vm_page_sbusy(bp->b_pages[last_busied]); 4000 while (vm_page_xbusied(m)) { 4001 vm_page_lock(m); 4002 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4003 vm_page_busy_sleep(m, "vbpage"); 4004 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4005 } 4006 } 4007 } 4008 for (i = 0; i < last_busied; i++) 4009 vm_page_sunbusy(bp->b_pages[i]); 4010} 4011 4012/* 4013 * This routine is called before a device strategy routine. 4014 * It is used to tell the VM system that paging I/O is in 4015 * progress, and treat the pages associated with the buffer 4016 * almost as being exclusive busy. Also the object paging_in_progress 4017 * flag is handled to make sure that the object doesn't become 4018 * inconsistant. 4019 * 4020 * Since I/O has not been initiated yet, certain buffer flags 4021 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 4022 * and should be ignored. 4023 */ 4024void 4025vfs_busy_pages(struct buf *bp, int clear_modify) 4026{ 4027 int i, bogus; 4028 vm_object_t obj; 4029 vm_ooffset_t foff; 4030 vm_page_t m; 4031 4032 if (!(bp->b_flags & B_VMIO)) 4033 return; 4034 4035 obj = bp->b_bufobj->bo_object; 4036 foff = bp->b_offset; 4037 KASSERT(bp->b_offset != NOOFFSET, 4038 ("vfs_busy_pages: no buffer offset")); 4039 VM_OBJECT_WLOCK(obj); 4040 vfs_drain_busy_pages(bp); 4041 if (bp->b_bufsize != 0) 4042 vfs_setdirty_locked_object(bp); 4043 bogus = 0; 4044 for (i = 0; i < bp->b_npages; i++) { 4045 m = bp->b_pages[i]; 4046 4047 if ((bp->b_flags & B_CLUSTER) == 0) { 4048 vm_object_pip_add(obj, 1); 4049 vm_page_sbusy(m); 4050 } 4051 /* 4052 * When readying a buffer for a read ( i.e 4053 * clear_modify == 0 ), it is important to do 4054 * bogus_page replacement for valid pages in 4055 * partially instantiated buffers. Partially 4056 * instantiated buffers can, in turn, occur when 4057 * reconstituting a buffer from its VM backing store 4058 * base. We only have to do this if B_CACHE is 4059 * clear ( which causes the I/O to occur in the 4060 * first place ). The replacement prevents the read 4061 * I/O from overwriting potentially dirty VM-backed 4062 * pages. XXX bogus page replacement is, uh, bogus. 4063 * It may not work properly with small-block devices. 4064 * We need to find a better way. 4065 */ 4066 if (clear_modify) { 4067 pmap_remove_write(m); 4068 vfs_page_set_validclean(bp, foff, m); 4069 } else if (m->valid == VM_PAGE_BITS_ALL && 4070 (bp->b_flags & B_CACHE) == 0) { 4071 bp->b_pages[i] = bogus_page; 4072 bogus++; 4073 } 4074 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4075 } 4076 VM_OBJECT_WUNLOCK(obj); 4077 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 4078 BUF_CHECK_MAPPED(bp); 4079 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4080 bp->b_pages, bp->b_npages); 4081 } 4082} 4083 4084/* 4085 * vfs_bio_set_valid: 4086 * 4087 * Set the range within the buffer to valid. The range is 4088 * relative to the beginning of the buffer, b_offset. Note that 4089 * b_offset itself may be offset from the beginning of the first 4090 * page. 4091 */ 4092void 4093vfs_bio_set_valid(struct buf *bp, int base, int size) 4094{ 4095 int i, n; 4096 vm_page_t m; 4097 4098 if (!(bp->b_flags & B_VMIO)) 4099 return; 4100 4101 /* 4102 * Fixup base to be relative to beginning of first page. 4103 * Set initial n to be the maximum number of bytes in the 4104 * first page that can be validated. 4105 */ 4106 base += (bp->b_offset & PAGE_MASK); 4107 n = PAGE_SIZE - (base & PAGE_MASK); 4108 4109 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4110 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4111 m = bp->b_pages[i]; 4112 if (n > size) 4113 n = size; 4114 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4115 base += n; 4116 size -= n; 4117 n = PAGE_SIZE; 4118 } 4119 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4120} 4121 4122/* 4123 * vfs_bio_clrbuf: 4124 * 4125 * If the specified buffer is a non-VMIO buffer, clear the entire 4126 * buffer. If the specified buffer is a VMIO buffer, clear and 4127 * validate only the previously invalid portions of the buffer. 4128 * This routine essentially fakes an I/O, so we need to clear 4129 * BIO_ERROR and B_INVAL. 4130 * 4131 * Note that while we only theoretically need to clear through b_bcount, 4132 * we go ahead and clear through b_bufsize. 4133 */ 4134void 4135vfs_bio_clrbuf(struct buf *bp) 4136{ 4137 int i, j, mask, sa, ea, slide; 4138 4139 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4140 clrbuf(bp); 4141 return; 4142 } 4143 bp->b_flags &= ~B_INVAL; 4144 bp->b_ioflags &= ~BIO_ERROR; 4145 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4146 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4147 (bp->b_offset & PAGE_MASK) == 0) { 4148 if (bp->b_pages[0] == bogus_page) 4149 goto unlock; 4150 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4151 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object); 4152 if ((bp->b_pages[0]->valid & mask) == mask) 4153 goto unlock; 4154 if ((bp->b_pages[0]->valid & mask) == 0) { 4155 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize); 4156 bp->b_pages[0]->valid |= mask; 4157 goto unlock; 4158 } 4159 } 4160 sa = bp->b_offset & PAGE_MASK; 4161 slide = 0; 4162 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4163 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4164 ea = slide & PAGE_MASK; 4165 if (ea == 0) 4166 ea = PAGE_SIZE; 4167 if (bp->b_pages[i] == bogus_page) 4168 continue; 4169 j = sa / DEV_BSIZE; 4170 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4171 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object); 4172 if ((bp->b_pages[i]->valid & mask) == mask) 4173 continue; 4174 if ((bp->b_pages[i]->valid & mask) == 0) 4175 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4176 else { 4177 for (; sa < ea; sa += DEV_BSIZE, j++) { 4178 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4179 pmap_zero_page_area(bp->b_pages[i], 4180 sa, DEV_BSIZE); 4181 } 4182 } 4183 } 4184 bp->b_pages[i]->valid |= mask; 4185 } 4186unlock: 4187 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4188 bp->b_resid = 0; 4189} 4190 4191void 4192vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4193{ 4194 vm_page_t m; 4195 int i, n; 4196 4197 if ((bp->b_flags & B_UNMAPPED) == 0) { 4198 BUF_CHECK_MAPPED(bp); 4199 bzero(bp->b_data + base, size); 4200 } else { 4201 BUF_CHECK_UNMAPPED(bp); 4202 n = PAGE_SIZE - (base & PAGE_MASK); 4203 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4204 m = bp->b_pages[i]; 4205 if (n > size) 4206 n = size; 4207 pmap_zero_page_area(m, base & PAGE_MASK, n); 4208 base += n; 4209 size -= n; 4210 n = PAGE_SIZE; 4211 } 4212 } 4213} 4214 4215/* 4216 * vm_hold_load_pages and vm_hold_free_pages get pages into 4217 * a buffers address space. The pages are anonymous and are 4218 * not associated with a file object. 4219 */ 4220static void 4221vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4222{ 4223 vm_offset_t pg; 4224 vm_page_t p; 4225 int index; 4226 4227 BUF_CHECK_MAPPED(bp); 4228 4229 to = round_page(to); 4230 from = round_page(from); 4231 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4232 4233 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4234tryagain: 4235 /* 4236 * note: must allocate system pages since blocking here 4237 * could interfere with paging I/O, no matter which 4238 * process we are. 4239 */ 4240 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | 4241 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT)); 4242 if (p == NULL) { 4243 VM_WAIT; 4244 goto tryagain; 4245 } 4246 pmap_qenter(pg, &p, 1); 4247 bp->b_pages[index] = p; 4248 } 4249 bp->b_npages = index; 4250} 4251 4252/* Return pages associated with this buf to the vm system */ 4253static void 4254vm_hold_free_pages(struct buf *bp, int newbsize) 4255{ 4256 vm_offset_t from; 4257 vm_page_t p; 4258 int index, newnpages; 4259 4260 BUF_CHECK_MAPPED(bp); 4261 4262 from = round_page((vm_offset_t)bp->b_data + newbsize); 4263 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4264 if (bp->b_npages > newnpages) 4265 pmap_qremove(from, bp->b_npages - newnpages); 4266 for (index = newnpages; index < bp->b_npages; index++) { 4267 p = bp->b_pages[index]; 4268 bp->b_pages[index] = NULL; 4269 if (vm_page_sbusied(p)) 4270 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 4271 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); 4272 p->wire_count--; 4273 vm_page_free(p); 4274 atomic_subtract_int(&cnt.v_wire_count, 1); 4275 } 4276 bp->b_npages = newnpages; 4277} 4278 4279/* 4280 * Map an IO request into kernel virtual address space. 4281 * 4282 * All requests are (re)mapped into kernel VA space. 4283 * Notice that we use b_bufsize for the size of the buffer 4284 * to be mapped. b_bcount might be modified by the driver. 4285 * 4286 * Note that even if the caller determines that the address space should 4287 * be valid, a race or a smaller-file mapped into a larger space may 4288 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 4289 * check the return value. 4290 */ 4291int 4292vmapbuf(struct buf *bp, int mapbuf) 4293{ 4294 caddr_t kva; 4295 vm_prot_t prot; 4296 int pidx; 4297 4298 if (bp->b_bufsize < 0) 4299 return (-1); 4300 prot = VM_PROT_READ; 4301 if (bp->b_iocmd == BIO_READ) 4302 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 4303 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 4304 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, 4305 btoc(MAXPHYS))) < 0) 4306 return (-1); 4307 bp->b_npages = pidx; 4308 if (mapbuf || !unmapped_buf_allowed) { 4309 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 4310 kva = bp->b_saveaddr; 4311 bp->b_saveaddr = bp->b_data; 4312 bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK); 4313 bp->b_flags &= ~B_UNMAPPED; 4314 } else { 4315 bp->b_flags |= B_UNMAPPED; 4316 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK; 4317 bp->b_saveaddr = bp->b_data; 4318 bp->b_data = unmapped_buf; 4319 } 4320 return(0); 4321} 4322 4323/* 4324 * Free the io map PTEs associated with this IO operation. 4325 * We also invalidate the TLB entries and restore the original b_addr. 4326 */ 4327void 4328vunmapbuf(struct buf *bp) 4329{ 4330 int npages; 4331 4332 npages = bp->b_npages; 4333 if (bp->b_flags & B_UNMAPPED) 4334 bp->b_flags &= ~B_UNMAPPED; 4335 else 4336 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4337 vm_page_unhold_pages(bp->b_pages, npages); 4338 4339 bp->b_data = bp->b_saveaddr; 4340} 4341 4342void 4343bdone(struct buf *bp) 4344{ 4345 struct mtx *mtxp; 4346 4347 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4348 mtx_lock(mtxp); 4349 bp->b_flags |= B_DONE; 4350 wakeup(bp); 4351 mtx_unlock(mtxp); 4352} 4353 4354void 4355bwait(struct buf *bp, u_char pri, const char *wchan) 4356{ 4357 struct mtx *mtxp; 4358 4359 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4360 mtx_lock(mtxp); 4361 while ((bp->b_flags & B_DONE) == 0) 4362 msleep(bp, mtxp, pri, wchan, 0); 4363 mtx_unlock(mtxp); 4364} 4365 4366int 4367bufsync(struct bufobj *bo, int waitfor) 4368{ 4369 4370 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 4371} 4372 4373void 4374bufstrategy(struct bufobj *bo, struct buf *bp) 4375{ 4376 int i = 0; 4377 struct vnode *vp; 4378 4379 vp = bp->b_vp; 4380 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 4381 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 4382 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 4383 i = VOP_STRATEGY(vp, bp); 4384 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 4385} 4386 4387void 4388bufobj_wrefl(struct bufobj *bo) 4389{ 4390 4391 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4392 ASSERT_BO_WLOCKED(bo); 4393 bo->bo_numoutput++; 4394} 4395 4396void 4397bufobj_wref(struct bufobj *bo) 4398{ 4399 4400 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4401 BO_LOCK(bo); 4402 bo->bo_numoutput++; 4403 BO_UNLOCK(bo); 4404} 4405 4406void 4407bufobj_wdrop(struct bufobj *bo) 4408{ 4409 4410 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 4411 BO_LOCK(bo); 4412 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 4413 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 4414 bo->bo_flag &= ~BO_WWAIT; 4415 wakeup(&bo->bo_numoutput); 4416 } 4417 BO_UNLOCK(bo); 4418} 4419 4420int 4421bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 4422{ 4423 int error; 4424 4425 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 4426 ASSERT_BO_WLOCKED(bo); 4427 error = 0; 4428 while (bo->bo_numoutput) { 4429 bo->bo_flag |= BO_WWAIT; 4430 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), 4431 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 4432 if (error) 4433 break; 4434 } 4435 return (error); 4436} 4437 4438void 4439bpin(struct buf *bp) 4440{ 4441 struct mtx *mtxp; 4442 4443 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4444 mtx_lock(mtxp); 4445 bp->b_pin_count++; 4446 mtx_unlock(mtxp); 4447} 4448 4449void 4450bunpin(struct buf *bp) 4451{ 4452 struct mtx *mtxp; 4453 4454 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4455 mtx_lock(mtxp); 4456 if (--bp->b_pin_count == 0) 4457 wakeup(bp); 4458 mtx_unlock(mtxp); 4459} 4460 4461void 4462bunpin_wait(struct buf *bp) 4463{ 4464 struct mtx *mtxp; 4465 4466 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4467 mtx_lock(mtxp); 4468 while (bp->b_pin_count > 0) 4469 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4470 mtx_unlock(mtxp); 4471} 4472 4473/* 4474 * Set bio_data or bio_ma for struct bio from the struct buf. 4475 */ 4476void 4477bdata2bio(struct buf *bp, struct bio *bip) 4478{ 4479 4480 if ((bp->b_flags & B_UNMAPPED) != 0) { 4481 KASSERT(unmapped_buf_allowed, ("unmapped")); 4482 bip->bio_ma = bp->b_pages; 4483 bip->bio_ma_n = bp->b_npages; 4484 bip->bio_data = unmapped_buf; 4485 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 4486 bip->bio_flags |= BIO_UNMAPPED; 4487 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 4488 PAGE_SIZE == bp->b_npages, 4489 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset, 4490 (long long)bip->bio_length, bip->bio_ma_n)); 4491 } else { 4492 bip->bio_data = bp->b_data; 4493 bip->bio_ma = NULL; 4494 } 4495} 4496 4497#include "opt_ddb.h" 4498#ifdef DDB 4499#include <ddb/ddb.h> 4500 4501/* DDB command to show buffer data */ 4502DB_SHOW_COMMAND(buffer, db_show_buffer) 4503{ 4504 /* get args */ 4505 struct buf *bp = (struct buf *)addr; 4506 4507 if (!have_addr) { 4508 db_printf("usage: show buffer <addr>\n"); 4509 return; 4510 } 4511 4512 db_printf("buf at %p\n", bp); 4513 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n", 4514 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, 4515 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS); 4516 db_printf( 4517 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4518 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, " 4519 "b_dep = %p\n", 4520 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4521 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4522 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first); 4523 if (bp->b_npages) { 4524 int i; 4525 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4526 for (i = 0; i < bp->b_npages; i++) { 4527 vm_page_t m; 4528 m = bp->b_pages[i]; 4529 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4530 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4531 if ((i + 1) < bp->b_npages) 4532 db_printf(","); 4533 } 4534 db_printf("\n"); 4535 } 4536 db_printf(" "); 4537 BUF_LOCKPRINTINFO(bp); 4538} 4539 4540DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4541{ 4542 struct buf *bp; 4543 int i; 4544 4545 for (i = 0; i < nbuf; i++) { 4546 bp = &buf[i]; 4547 if (BUF_ISLOCKED(bp)) { 4548 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4549 db_printf("\n"); 4550 } 4551 } 4552} 4553 4554DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4555{ 4556 struct vnode *vp; 4557 struct buf *bp; 4558 4559 if (!have_addr) { 4560 db_printf("usage: show vnodebufs <addr>\n"); 4561 return; 4562 } 4563 vp = (struct vnode *)addr; 4564 db_printf("Clean buffers:\n"); 4565 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4566 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4567 db_printf("\n"); 4568 } 4569 db_printf("Dirty buffers:\n"); 4570 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4571 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4572 db_printf("\n"); 4573 } 4574} 4575 4576DB_COMMAND(countfreebufs, db_coundfreebufs) 4577{ 4578 struct buf *bp; 4579 int i, used = 0, nfree = 0; 4580 4581 if (have_addr) { 4582 db_printf("usage: countfreebufs\n"); 4583 return; 4584 } 4585 4586 for (i = 0; i < nbuf; i++) { 4587 bp = &buf[i]; 4588 if ((bp->b_flags & B_INFREECNT) != 0) 4589 nfree++; 4590 else 4591 used++; 4592 } 4593 4594 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4595 nfree + used); 4596 db_printf("numfreebuffers is %d\n", numfreebuffers); 4597} 4598#endif /* DDB */ 4599