vfs_bio.c revision 275797
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 275797 2014-12-15 11:00:56Z kib $"); 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 vm_object_t obj; 1856 vm_page_t m; 1857 int i; 1858 1859 if ((bp->b_flags & B_UNMAPPED) == 0) { 1860 BUF_CHECK_MAPPED(bp); 1861 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 1862 } else 1863 BUF_CHECK_UNMAPPED(bp); 1864 obj = bp->b_bufobj->bo_object; 1865 if (obj != NULL) 1866 VM_OBJECT_WLOCK(obj); 1867 for (i = 0; i < bp->b_npages; i++) { 1868 m = bp->b_pages[i]; 1869 bp->b_pages[i] = NULL; 1870 /* 1871 * In order to keep page LRU ordering consistent, put 1872 * everything on the inactive queue. 1873 */ 1874 vm_page_lock(m); 1875 vm_page_unwire(m, 0); 1876 1877 /* 1878 * Might as well free the page if we can and it has 1879 * no valid data. We also free the page if the 1880 * buffer was used for direct I/O 1881 */ 1882 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) { 1883 if (m->wire_count == 0 && !vm_page_busied(m)) 1884 vm_page_free(m); 1885 } else if (bp->b_flags & B_DIRECT) 1886 vm_page_try_to_free(m); 1887 else if (buf_vm_page_count_severe()) 1888 vm_page_try_to_cache(m); 1889 vm_page_unlock(m); 1890 } 1891 if (obj != NULL) 1892 VM_OBJECT_WUNLOCK(obj); 1893 1894 if (bp->b_bufsize) { 1895 bufspacewakeup(); 1896 bp->b_bufsize = 0; 1897 } 1898 bp->b_npages = 0; 1899 bp->b_flags &= ~B_VMIO; 1900 if (bp->b_vp) 1901 brelvp(bp); 1902} 1903 1904/* 1905 * Check to see if a block at a particular lbn is available for a clustered 1906 * write. 1907 */ 1908static int 1909vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1910{ 1911 struct buf *bpa; 1912 int match; 1913 1914 match = 0; 1915 1916 /* If the buf isn't in core skip it */ 1917 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1918 return (0); 1919 1920 /* If the buf is busy we don't want to wait for it */ 1921 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1922 return (0); 1923 1924 /* Only cluster with valid clusterable delayed write buffers */ 1925 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1926 (B_DELWRI | B_CLUSTEROK)) 1927 goto done; 1928 1929 if (bpa->b_bufsize != size) 1930 goto done; 1931 1932 /* 1933 * Check to see if it is in the expected place on disk and that the 1934 * block has been mapped. 1935 */ 1936 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1937 match = 1; 1938done: 1939 BUF_UNLOCK(bpa); 1940 return (match); 1941} 1942 1943/* 1944 * vfs_bio_awrite: 1945 * 1946 * Implement clustered async writes for clearing out B_DELWRI buffers. 1947 * This is much better then the old way of writing only one buffer at 1948 * a time. Note that we may not be presented with the buffers in the 1949 * correct order, so we search for the cluster in both directions. 1950 */ 1951int 1952vfs_bio_awrite(struct buf *bp) 1953{ 1954 struct bufobj *bo; 1955 int i; 1956 int j; 1957 daddr_t lblkno = bp->b_lblkno; 1958 struct vnode *vp = bp->b_vp; 1959 int ncl; 1960 int nwritten; 1961 int size; 1962 int maxcl; 1963 int gbflags; 1964 1965 bo = &vp->v_bufobj; 1966 gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0; 1967 /* 1968 * right now we support clustered writing only to regular files. If 1969 * we find a clusterable block we could be in the middle of a cluster 1970 * rather then at the beginning. 1971 */ 1972 if ((vp->v_type == VREG) && 1973 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1974 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1975 1976 size = vp->v_mount->mnt_stat.f_iosize; 1977 maxcl = MAXPHYS / size; 1978 1979 BO_RLOCK(bo); 1980 for (i = 1; i < maxcl; i++) 1981 if (vfs_bio_clcheck(vp, size, lblkno + i, 1982 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1983 break; 1984 1985 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1986 if (vfs_bio_clcheck(vp, size, lblkno - j, 1987 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1988 break; 1989 BO_RUNLOCK(bo); 1990 --j; 1991 ncl = i + j; 1992 /* 1993 * this is a possible cluster write 1994 */ 1995 if (ncl != 1) { 1996 BUF_UNLOCK(bp); 1997 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 1998 gbflags); 1999 return (nwritten); 2000 } 2001 } 2002 bremfree(bp); 2003 bp->b_flags |= B_ASYNC; 2004 /* 2005 * default (old) behavior, writing out only one block 2006 * 2007 * XXX returns b_bufsize instead of b_bcount for nwritten? 2008 */ 2009 nwritten = bp->b_bufsize; 2010 (void) bwrite(bp); 2011 2012 return (nwritten); 2013} 2014 2015static void 2016setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags) 2017{ 2018 2019 KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 2020 bp->b_kvasize == 0, ("call bfreekva(%p)", bp)); 2021 if ((gbflags & GB_UNMAPPED) == 0) { 2022 bp->b_kvabase = (caddr_t)addr; 2023 } else if ((gbflags & GB_KVAALLOC) != 0) { 2024 KASSERT((gbflags & GB_UNMAPPED) != 0, 2025 ("GB_KVAALLOC without GB_UNMAPPED")); 2026 bp->b_kvaalloc = (caddr_t)addr; 2027 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 2028 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 2029 } 2030 bp->b_kvasize = maxsize; 2031} 2032 2033/* 2034 * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if 2035 * needed. 2036 */ 2037static int 2038allocbufkva(struct buf *bp, int maxsize, int gbflags) 2039{ 2040 vm_offset_t addr; 2041 2042 bfreekva(bp); 2043 addr = 0; 2044 2045 if (vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr)) { 2046 /* 2047 * Buffer map is too fragmented. Request the caller 2048 * to defragment the map. 2049 */ 2050 atomic_add_int(&bufdefragcnt, 1); 2051 return (1); 2052 } 2053 setbufkva(bp, addr, maxsize, gbflags); 2054 atomic_add_long(&bufspace, bp->b_kvasize); 2055 return (0); 2056} 2057 2058/* 2059 * Ask the bufdaemon for help, or act as bufdaemon itself, when a 2060 * locked vnode is supplied. 2061 */ 2062static void 2063getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo, 2064 int defrag) 2065{ 2066 struct thread *td; 2067 char *waitmsg; 2068 int cnt, error, flags, norunbuf, wait; 2069 2070 mtx_assert(&bqclean, MA_OWNED); 2071 2072 if (defrag) { 2073 flags = VFS_BIO_NEED_BUFSPACE; 2074 waitmsg = "nbufkv"; 2075 } else if (bufspace >= hibufspace) { 2076 waitmsg = "nbufbs"; 2077 flags = VFS_BIO_NEED_BUFSPACE; 2078 } else { 2079 waitmsg = "newbuf"; 2080 flags = VFS_BIO_NEED_ANY; 2081 } 2082 atomic_set_int(&needsbuffer, flags); 2083 mtx_unlock(&bqclean); 2084 2085 bd_speedup(); /* heeeelp */ 2086 if ((gbflags & GB_NOWAIT_BD) != 0) 2087 return; 2088 2089 td = curthread; 2090 cnt = 0; 2091 wait = MNT_NOWAIT; 2092 rw_wlock(&nblock); 2093 while ((needsbuffer & flags) != 0) { 2094 if (vp != NULL && vp->v_type != VCHR && 2095 (td->td_pflags & TDP_BUFNEED) == 0) { 2096 rw_wunlock(&nblock); 2097 /* 2098 * getblk() is called with a vnode locked, and 2099 * some majority of the dirty buffers may as 2100 * well belong to the vnode. Flushing the 2101 * buffers there would make a progress that 2102 * cannot be achieved by the buf_daemon, that 2103 * cannot lock the vnode. 2104 */ 2105 if (cnt++ > 2) 2106 wait = MNT_WAIT; 2107 ASSERT_VOP_LOCKED(vp, "bufd_helper"); 2108 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 : 2109 vn_lock(vp, LK_TRYUPGRADE); 2110 if (error == 0) { 2111 /* play bufdaemon */ 2112 norunbuf = curthread_pflags_set(TDP_BUFNEED | 2113 TDP_NORUNNINGBUF); 2114 VOP_FSYNC(vp, wait, td); 2115 atomic_add_long(¬bufdflushes, 1); 2116 curthread_pflags_restore(norunbuf); 2117 } 2118 rw_wlock(&nblock); 2119 if ((needsbuffer & flags) == 0) 2120 break; 2121 } 2122 error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock, 2123 (PRIBIO + 4) | slpflag, waitmsg, slptimeo); 2124 if (error != 0) 2125 break; 2126 } 2127 rw_wunlock(&nblock); 2128} 2129 2130static void 2131getnewbuf_reuse_bp(struct buf *bp, int qindex) 2132{ 2133 2134 CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 2135 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 2136 bp->b_kvasize, bp->b_bufsize, qindex); 2137 mtx_assert(&bqclean, MA_NOTOWNED); 2138 2139 /* 2140 * Note: we no longer distinguish between VMIO and non-VMIO 2141 * buffers. 2142 */ 2143 KASSERT((bp->b_flags & B_DELWRI) == 0, 2144 ("delwri buffer %p found in queue %d", bp, qindex)); 2145 2146 if (qindex == QUEUE_CLEAN) { 2147 if (bp->b_flags & B_VMIO) { 2148 bp->b_flags &= ~B_ASYNC; 2149 vfs_vmio_release(bp); 2150 } 2151 if (bp->b_vp != NULL) 2152 brelvp(bp); 2153 } 2154 2155 /* 2156 * Get the rest of the buffer freed up. b_kva* is still valid 2157 * after this operation. 2158 */ 2159 2160 if (bp->b_rcred != NOCRED) { 2161 crfree(bp->b_rcred); 2162 bp->b_rcred = NOCRED; 2163 } 2164 if (bp->b_wcred != NOCRED) { 2165 crfree(bp->b_wcred); 2166 bp->b_wcred = NOCRED; 2167 } 2168 if (!LIST_EMPTY(&bp->b_dep)) 2169 buf_deallocate(bp); 2170 if (bp->b_vflags & BV_BKGRDINPROG) 2171 panic("losing buffer 3"); 2172 KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p. qindex: %d", 2173 bp, bp->b_vp, qindex)); 2174 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 2175 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 2176 2177 if (bp->b_bufsize) 2178 allocbuf(bp, 0); 2179 2180 bp->b_flags &= B_UNMAPPED | B_KVAALLOC; 2181 bp->b_ioflags = 0; 2182 bp->b_xflags = 0; 2183 KASSERT((bp->b_flags & B_INFREECNT) == 0, 2184 ("buf %p still counted as free?", bp)); 2185 bp->b_vflags = 0; 2186 bp->b_vp = NULL; 2187 bp->b_blkno = bp->b_lblkno = 0; 2188 bp->b_offset = NOOFFSET; 2189 bp->b_iodone = 0; 2190 bp->b_error = 0; 2191 bp->b_resid = 0; 2192 bp->b_bcount = 0; 2193 bp->b_npages = 0; 2194 bp->b_dirtyoff = bp->b_dirtyend = 0; 2195 bp->b_bufobj = NULL; 2196 bp->b_pin_count = 0; 2197 bp->b_fsprivate1 = NULL; 2198 bp->b_fsprivate2 = NULL; 2199 bp->b_fsprivate3 = NULL; 2200 2201 LIST_INIT(&bp->b_dep); 2202} 2203 2204static int flushingbufs; 2205 2206static struct buf * 2207getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata) 2208{ 2209 struct buf *bp, *nbp; 2210 int nqindex, qindex, pass; 2211 2212 KASSERT(!unmapped || !defrag, ("both unmapped and defrag")); 2213 2214 pass = 1; 2215restart: 2216 atomic_add_int(&getnewbufrestarts, 1); 2217 2218 /* 2219 * Setup for scan. If we do not have enough free buffers, 2220 * we setup a degenerate case that immediately fails. Note 2221 * that if we are specially marked process, we are allowed to 2222 * dip into our reserves. 2223 * 2224 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 2225 * for the allocation of the mapped buffer. For unmapped, the 2226 * easiest is to start with EMPTY outright. 2227 * 2228 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 2229 * However, there are a number of cases (defragging, reusing, ...) 2230 * where we cannot backup. 2231 */ 2232 nbp = NULL; 2233 mtx_lock(&bqclean); 2234 if (!defrag && unmapped) { 2235 nqindex = QUEUE_EMPTY; 2236 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2237 } 2238 if (nbp == NULL) { 2239 nqindex = QUEUE_EMPTYKVA; 2240 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2241 } 2242 2243 /* 2244 * If no EMPTYKVA buffers and we are either defragging or 2245 * reusing, locate a CLEAN buffer to free or reuse. If 2246 * bufspace useage is low skip this step so we can allocate a 2247 * new buffer. 2248 */ 2249 if (nbp == NULL && (defrag || bufspace >= lobufspace)) { 2250 nqindex = QUEUE_CLEAN; 2251 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2252 } 2253 2254 /* 2255 * If we could not find or were not allowed to reuse a CLEAN 2256 * buffer, check to see if it is ok to use an EMPTY buffer. 2257 * We can only use an EMPTY buffer if allocating its KVA would 2258 * not otherwise run us out of buffer space. No KVA is needed 2259 * for the unmapped allocation. 2260 */ 2261 if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace || 2262 metadata)) { 2263 nqindex = QUEUE_EMPTY; 2264 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2265 } 2266 2267 /* 2268 * All available buffers might be clean, retry ignoring the 2269 * lobufspace as the last resort. 2270 */ 2271 if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) { 2272 nqindex = QUEUE_CLEAN; 2273 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2274 } 2275 2276 /* 2277 * Run scan, possibly freeing data and/or kva mappings on the fly 2278 * depending. 2279 */ 2280 while ((bp = nbp) != NULL) { 2281 qindex = nqindex; 2282 2283 /* 2284 * Calculate next bp (we can only use it if we do not 2285 * block or do other fancy things). 2286 */ 2287 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 2288 switch (qindex) { 2289 case QUEUE_EMPTY: 2290 nqindex = QUEUE_EMPTYKVA; 2291 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2292 if (nbp != NULL) 2293 break; 2294 /* FALLTHROUGH */ 2295 case QUEUE_EMPTYKVA: 2296 nqindex = QUEUE_CLEAN; 2297 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2298 if (nbp != NULL) 2299 break; 2300 /* FALLTHROUGH */ 2301 case QUEUE_CLEAN: 2302 if (metadata && pass == 1) { 2303 pass = 2; 2304 nqindex = QUEUE_EMPTY; 2305 nbp = TAILQ_FIRST( 2306 &bufqueues[QUEUE_EMPTY]); 2307 } 2308 /* 2309 * nbp is NULL. 2310 */ 2311 break; 2312 } 2313 } 2314 /* 2315 * If we are defragging then we need a buffer with 2316 * b_kvasize != 0. XXX this situation should no longer 2317 * occur, if defrag is non-zero the buffer's b_kvasize 2318 * should also be non-zero at this point. XXX 2319 */ 2320 if (defrag && bp->b_kvasize == 0) { 2321 printf("Warning: defrag empty buffer %p\n", bp); 2322 continue; 2323 } 2324 2325 /* 2326 * Start freeing the bp. This is somewhat involved. nbp 2327 * remains valid only for QUEUE_EMPTY[KVA] bp's. 2328 */ 2329 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2330 continue; 2331 /* 2332 * BKGRDINPROG can only be set with the buf and bufobj 2333 * locks both held. We tolerate a race to clear it here. 2334 */ 2335 if (bp->b_vflags & BV_BKGRDINPROG) { 2336 BUF_UNLOCK(bp); 2337 continue; 2338 } 2339 2340 KASSERT(bp->b_qindex == qindex, 2341 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2342 2343 bremfreel(bp); 2344 mtx_unlock(&bqclean); 2345 /* 2346 * NOTE: nbp is now entirely invalid. We can only restart 2347 * the scan from this point on. 2348 */ 2349 2350 getnewbuf_reuse_bp(bp, qindex); 2351 mtx_assert(&bqclean, MA_NOTOWNED); 2352 2353 /* 2354 * If we are defragging then free the buffer. 2355 */ 2356 if (defrag) { 2357 bp->b_flags |= B_INVAL; 2358 bfreekva(bp); 2359 brelse(bp); 2360 defrag = 0; 2361 goto restart; 2362 } 2363 2364 /* 2365 * Notify any waiters for the buffer lock about 2366 * identity change by freeing the buffer. 2367 */ 2368 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 2369 bp->b_flags |= B_INVAL; 2370 bfreekva(bp); 2371 brelse(bp); 2372 goto restart; 2373 } 2374 2375 if (metadata) 2376 break; 2377 2378 /* 2379 * If we are overcomitted then recover the buffer and its 2380 * KVM space. This occurs in rare situations when multiple 2381 * processes are blocked in getnewbuf() or allocbuf(). 2382 */ 2383 if (bufspace >= hibufspace) 2384 flushingbufs = 1; 2385 if (flushingbufs && bp->b_kvasize != 0) { 2386 bp->b_flags |= B_INVAL; 2387 bfreekva(bp); 2388 brelse(bp); 2389 goto restart; 2390 } 2391 if (bufspace < lobufspace) 2392 flushingbufs = 0; 2393 break; 2394 } 2395 return (bp); 2396} 2397 2398/* 2399 * getnewbuf: 2400 * 2401 * Find and initialize a new buffer header, freeing up existing buffers 2402 * in the bufqueues as necessary. The new buffer is returned locked. 2403 * 2404 * Important: B_INVAL is not set. If the caller wishes to throw the 2405 * buffer away, the caller must set B_INVAL prior to calling brelse(). 2406 * 2407 * We block if: 2408 * We have insufficient buffer headers 2409 * We have insufficient buffer space 2410 * buffer_arena is too fragmented ( space reservation fails ) 2411 * If we have to flush dirty buffers ( but we try to avoid this ) 2412 */ 2413static struct buf * 2414getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 2415 int gbflags) 2416{ 2417 struct buf *bp; 2418 int defrag, metadata; 2419 2420 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 2421 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 2422 if (!unmapped_buf_allowed) 2423 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 2424 2425 defrag = 0; 2426 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 2427 vp->v_type == VCHR) 2428 metadata = 1; 2429 else 2430 metadata = 0; 2431 /* 2432 * We can't afford to block since we might be holding a vnode lock, 2433 * which may prevent system daemons from running. We deal with 2434 * low-memory situations by proactively returning memory and running 2435 * async I/O rather then sync I/O. 2436 */ 2437 atomic_add_int(&getnewbufcalls, 1); 2438 atomic_subtract_int(&getnewbufrestarts, 1); 2439restart: 2440 bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED | 2441 GB_KVAALLOC)) == GB_UNMAPPED, metadata); 2442 if (bp != NULL) 2443 defrag = 0; 2444 2445 /* 2446 * If we exhausted our list, sleep as appropriate. We may have to 2447 * wakeup various daemons and write out some dirty buffers. 2448 * 2449 * Generally we are sleeping due to insufficient buffer space. 2450 */ 2451 if (bp == NULL) { 2452 mtx_assert(&bqclean, MA_OWNED); 2453 getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag); 2454 mtx_assert(&bqclean, MA_NOTOWNED); 2455 } else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) { 2456 mtx_assert(&bqclean, MA_NOTOWNED); 2457 2458 bfreekva(bp); 2459 bp->b_flags |= B_UNMAPPED; 2460 bp->b_kvabase = bp->b_data = unmapped_buf; 2461 bp->b_kvasize = maxsize; 2462 atomic_add_long(&bufspace, bp->b_kvasize); 2463 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 2464 atomic_add_int(&bufreusecnt, 1); 2465 } else { 2466 mtx_assert(&bqclean, MA_NOTOWNED); 2467 2468 /* 2469 * We finally have a valid bp. We aren't quite out of the 2470 * woods, we still have to reserve kva space. In order 2471 * to keep fragmentation sane we only allocate kva in 2472 * BKVASIZE chunks. 2473 */ 2474 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2475 2476 if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED | 2477 B_KVAALLOC)) == B_UNMAPPED) { 2478 if (allocbufkva(bp, maxsize, gbflags)) { 2479 defrag = 1; 2480 bp->b_flags |= B_INVAL; 2481 brelse(bp); 2482 goto restart; 2483 } 2484 atomic_add_int(&bufreusecnt, 1); 2485 } else if ((bp->b_flags & B_KVAALLOC) != 0 && 2486 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) { 2487 /* 2488 * If the reused buffer has KVA allocated, 2489 * reassign b_kvaalloc to b_kvabase. 2490 */ 2491 bp->b_kvabase = bp->b_kvaalloc; 2492 bp->b_flags &= ~B_KVAALLOC; 2493 atomic_subtract_long(&unmapped_bufspace, 2494 bp->b_kvasize); 2495 atomic_add_int(&bufreusecnt, 1); 2496 } else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 2497 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED | 2498 GB_KVAALLOC)) { 2499 /* 2500 * The case of reused buffer already have KVA 2501 * mapped, but the request is for unmapped 2502 * buffer with KVA allocated. 2503 */ 2504 bp->b_kvaalloc = bp->b_kvabase; 2505 bp->b_data = bp->b_kvabase = unmapped_buf; 2506 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 2507 atomic_add_long(&unmapped_bufspace, 2508 bp->b_kvasize); 2509 atomic_add_int(&bufreusecnt, 1); 2510 } 2511 if ((gbflags & GB_UNMAPPED) == 0) { 2512 bp->b_saveaddr = bp->b_kvabase; 2513 bp->b_data = bp->b_saveaddr; 2514 bp->b_flags &= ~B_UNMAPPED; 2515 BUF_CHECK_MAPPED(bp); 2516 } 2517 } 2518 return (bp); 2519} 2520 2521/* 2522 * buf_daemon: 2523 * 2524 * buffer flushing daemon. Buffers are normally flushed by the 2525 * update daemon but if it cannot keep up this process starts to 2526 * take the load in an attempt to prevent getnewbuf() from blocking. 2527 */ 2528 2529static struct kproc_desc buf_kp = { 2530 "bufdaemon", 2531 buf_daemon, 2532 &bufdaemonproc 2533}; 2534SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2535 2536static int 2537buf_flush(int target) 2538{ 2539 int flushed; 2540 2541 flushed = flushbufqueues(target, 0); 2542 if (flushed == 0) { 2543 /* 2544 * Could not find any buffers without rollback 2545 * dependencies, so just write the first one 2546 * in the hopes of eventually making progress. 2547 */ 2548 flushed = flushbufqueues(target, 1); 2549 } 2550 return (flushed); 2551} 2552 2553static void 2554buf_daemon() 2555{ 2556 int lodirty; 2557 2558 /* 2559 * This process needs to be suspended prior to shutdown sync. 2560 */ 2561 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2562 SHUTDOWN_PRI_LAST); 2563 2564 /* 2565 * This process is allowed to take the buffer cache to the limit 2566 */ 2567 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2568 mtx_lock(&bdlock); 2569 for (;;) { 2570 bd_request = 0; 2571 mtx_unlock(&bdlock); 2572 2573 kproc_suspend_check(bufdaemonproc); 2574 lodirty = lodirtybuffers; 2575 if (bd_speedupreq) { 2576 lodirty = numdirtybuffers / 2; 2577 bd_speedupreq = 0; 2578 } 2579 /* 2580 * Do the flush. Limit the amount of in-transit I/O we 2581 * allow to build up, otherwise we would completely saturate 2582 * the I/O system. 2583 */ 2584 while (numdirtybuffers > lodirty) { 2585 if (buf_flush(numdirtybuffers - lodirty) == 0) 2586 break; 2587 kern_yield(PRI_USER); 2588 } 2589 2590 /* 2591 * Only clear bd_request if we have reached our low water 2592 * mark. The buf_daemon normally waits 1 second and 2593 * then incrementally flushes any dirty buffers that have 2594 * built up, within reason. 2595 * 2596 * If we were unable to hit our low water mark and couldn't 2597 * find any flushable buffers, we sleep for a short period 2598 * to avoid endless loops on unlockable buffers. 2599 */ 2600 mtx_lock(&bdlock); 2601 if (numdirtybuffers <= lodirtybuffers) { 2602 /* 2603 * We reached our low water mark, reset the 2604 * request and sleep until we are needed again. 2605 * The sleep is just so the suspend code works. 2606 */ 2607 bd_request = 0; 2608 /* 2609 * Do an extra wakeup in case dirty threshold 2610 * changed via sysctl and the explicit transition 2611 * out of shortfall was missed. 2612 */ 2613 bdirtywakeup(); 2614 if (runningbufspace <= lorunningspace) 2615 runningwakeup(); 2616 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2617 } else { 2618 /* 2619 * We couldn't find any flushable dirty buffers but 2620 * still have too many dirty buffers, we 2621 * have to sleep and try again. (rare) 2622 */ 2623 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2624 } 2625 } 2626} 2627 2628/* 2629 * flushbufqueues: 2630 * 2631 * Try to flush a buffer in the dirty queue. We must be careful to 2632 * free up B_INVAL buffers instead of write them, which NFS is 2633 * particularly sensitive to. 2634 */ 2635static int flushwithdeps = 0; 2636SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2637 0, "Number of buffers flushed with dependecies that require rollbacks"); 2638 2639static int 2640flushbufqueues(int target, int flushdeps) 2641{ 2642 struct buf *sentinel; 2643 struct vnode *vp; 2644 struct mount *mp; 2645 struct buf *bp; 2646 int hasdeps; 2647 int flushed; 2648 int queue; 2649 int error; 2650 2651 flushed = 0; 2652 queue = QUEUE_DIRTY; 2653 bp = NULL; 2654 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2655 sentinel->b_qindex = QUEUE_SENTINEL; 2656 mtx_lock(&bqdirty); 2657 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2658 mtx_unlock(&bqdirty); 2659 while (flushed != target) { 2660 maybe_yield(); 2661 mtx_lock(&bqdirty); 2662 bp = TAILQ_NEXT(sentinel, b_freelist); 2663 if (bp != NULL) { 2664 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2665 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2666 b_freelist); 2667 } else { 2668 mtx_unlock(&bqdirty); 2669 break; 2670 } 2671 KASSERT(bp->b_qindex != QUEUE_SENTINEL, 2672 ("parallel calls to flushbufqueues() bp %p", bp)); 2673 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL); 2674 mtx_unlock(&bqdirty); 2675 if (error != 0) 2676 continue; 2677 if (bp->b_pin_count > 0) { 2678 BUF_UNLOCK(bp); 2679 continue; 2680 } 2681 /* 2682 * BKGRDINPROG can only be set with the buf and bufobj 2683 * locks both held. We tolerate a race to clear it here. 2684 */ 2685 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2686 (bp->b_flags & B_DELWRI) == 0) { 2687 BUF_UNLOCK(bp); 2688 continue; 2689 } 2690 if (bp->b_flags & B_INVAL) { 2691 bremfreef(bp); 2692 brelse(bp); 2693 flushed++; 2694 continue; 2695 } 2696 2697 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2698 if (flushdeps == 0) { 2699 BUF_UNLOCK(bp); 2700 continue; 2701 } 2702 hasdeps = 1; 2703 } else 2704 hasdeps = 0; 2705 /* 2706 * We must hold the lock on a vnode before writing 2707 * one of its buffers. Otherwise we may confuse, or 2708 * in the case of a snapshot vnode, deadlock the 2709 * system. 2710 * 2711 * The lock order here is the reverse of the normal 2712 * of vnode followed by buf lock. This is ok because 2713 * the NOWAIT will prevent deadlock. 2714 */ 2715 vp = bp->b_vp; 2716 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2717 BUF_UNLOCK(bp); 2718 continue; 2719 } 2720 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); 2721 if (error == 0) { 2722 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2723 bp, bp->b_vp, bp->b_flags); 2724 vfs_bio_awrite(bp); 2725 vn_finished_write(mp); 2726 VOP_UNLOCK(vp, 0); 2727 flushwithdeps += hasdeps; 2728 flushed++; 2729 if (runningbufspace > hirunningspace) 2730 waitrunningbufspace(); 2731 continue; 2732 } 2733 vn_finished_write(mp); 2734 BUF_UNLOCK(bp); 2735 } 2736 mtx_lock(&bqdirty); 2737 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2738 mtx_unlock(&bqdirty); 2739 free(sentinel, M_TEMP); 2740 return (flushed); 2741} 2742 2743/* 2744 * Check to see if a block is currently memory resident. 2745 */ 2746struct buf * 2747incore(struct bufobj *bo, daddr_t blkno) 2748{ 2749 struct buf *bp; 2750 2751 BO_RLOCK(bo); 2752 bp = gbincore(bo, blkno); 2753 BO_RUNLOCK(bo); 2754 return (bp); 2755} 2756 2757/* 2758 * Returns true if no I/O is needed to access the 2759 * associated VM object. This is like incore except 2760 * it also hunts around in the VM system for the data. 2761 */ 2762 2763static int 2764inmem(struct vnode * vp, daddr_t blkno) 2765{ 2766 vm_object_t obj; 2767 vm_offset_t toff, tinc, size; 2768 vm_page_t m; 2769 vm_ooffset_t off; 2770 2771 ASSERT_VOP_LOCKED(vp, "inmem"); 2772 2773 if (incore(&vp->v_bufobj, blkno)) 2774 return 1; 2775 if (vp->v_mount == NULL) 2776 return 0; 2777 obj = vp->v_object; 2778 if (obj == NULL) 2779 return (0); 2780 2781 size = PAGE_SIZE; 2782 if (size > vp->v_mount->mnt_stat.f_iosize) 2783 size = vp->v_mount->mnt_stat.f_iosize; 2784 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2785 2786 VM_OBJECT_RLOCK(obj); 2787 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2788 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2789 if (!m) 2790 goto notinmem; 2791 tinc = size; 2792 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2793 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2794 if (vm_page_is_valid(m, 2795 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2796 goto notinmem; 2797 } 2798 VM_OBJECT_RUNLOCK(obj); 2799 return 1; 2800 2801notinmem: 2802 VM_OBJECT_RUNLOCK(obj); 2803 return (0); 2804} 2805 2806/* 2807 * Set the dirty range for a buffer based on the status of the dirty 2808 * bits in the pages comprising the buffer. The range is limited 2809 * to the size of the buffer. 2810 * 2811 * Tell the VM system that the pages associated with this buffer 2812 * are clean. This is used for delayed writes where the data is 2813 * going to go to disk eventually without additional VM intevention. 2814 * 2815 * Note that while we only really need to clean through to b_bcount, we 2816 * just go ahead and clean through to b_bufsize. 2817 */ 2818static void 2819vfs_clean_pages_dirty_buf(struct buf *bp) 2820{ 2821 vm_ooffset_t foff, noff, eoff; 2822 vm_page_t m; 2823 int i; 2824 2825 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 2826 return; 2827 2828 foff = bp->b_offset; 2829 KASSERT(bp->b_offset != NOOFFSET, 2830 ("vfs_clean_pages_dirty_buf: no buffer offset")); 2831 2832 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 2833 vfs_drain_busy_pages(bp); 2834 vfs_setdirty_locked_object(bp); 2835 for (i = 0; i < bp->b_npages; i++) { 2836 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2837 eoff = noff; 2838 if (eoff > bp->b_offset + bp->b_bufsize) 2839 eoff = bp->b_offset + bp->b_bufsize; 2840 m = bp->b_pages[i]; 2841 vfs_page_set_validclean(bp, foff, m); 2842 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2843 foff = noff; 2844 } 2845 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 2846} 2847 2848static void 2849vfs_setdirty_locked_object(struct buf *bp) 2850{ 2851 vm_object_t object; 2852 int i; 2853 2854 object = bp->b_bufobj->bo_object; 2855 VM_OBJECT_ASSERT_WLOCKED(object); 2856 2857 /* 2858 * We qualify the scan for modified pages on whether the 2859 * object has been flushed yet. 2860 */ 2861 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { 2862 vm_offset_t boffset; 2863 vm_offset_t eoffset; 2864 2865 /* 2866 * test the pages to see if they have been modified directly 2867 * by users through the VM system. 2868 */ 2869 for (i = 0; i < bp->b_npages; i++) 2870 vm_page_test_dirty(bp->b_pages[i]); 2871 2872 /* 2873 * Calculate the encompassing dirty range, boffset and eoffset, 2874 * (eoffset - boffset) bytes. 2875 */ 2876 2877 for (i = 0; i < bp->b_npages; i++) { 2878 if (bp->b_pages[i]->dirty) 2879 break; 2880 } 2881 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2882 2883 for (i = bp->b_npages - 1; i >= 0; --i) { 2884 if (bp->b_pages[i]->dirty) { 2885 break; 2886 } 2887 } 2888 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2889 2890 /* 2891 * Fit it to the buffer. 2892 */ 2893 2894 if (eoffset > bp->b_bcount) 2895 eoffset = bp->b_bcount; 2896 2897 /* 2898 * If we have a good dirty range, merge with the existing 2899 * dirty range. 2900 */ 2901 2902 if (boffset < eoffset) { 2903 if (bp->b_dirtyoff > boffset) 2904 bp->b_dirtyoff = boffset; 2905 if (bp->b_dirtyend < eoffset) 2906 bp->b_dirtyend = eoffset; 2907 } 2908 } 2909} 2910 2911/* 2912 * Allocate the KVA mapping for an existing buffer. It handles the 2913 * cases of both B_UNMAPPED buffer, and buffer with the preallocated 2914 * KVA which is not mapped (B_KVAALLOC). 2915 */ 2916static void 2917bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 2918{ 2919 struct buf *scratch_bp; 2920 int bsize, maxsize, need_mapping, need_kva; 2921 off_t offset; 2922 2923 need_mapping = (bp->b_flags & B_UNMAPPED) != 0 && 2924 (gbflags & GB_UNMAPPED) == 0; 2925 need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED && 2926 (gbflags & GB_KVAALLOC) != 0; 2927 if (!need_mapping && !need_kva) 2928 return; 2929 2930 BUF_CHECK_UNMAPPED(bp); 2931 2932 if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) { 2933 /* 2934 * Buffer is not mapped, but the KVA was already 2935 * reserved at the time of the instantiation. Use the 2936 * allocated space. 2937 */ 2938 bp->b_flags &= ~B_KVAALLOC; 2939 KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0")); 2940 bp->b_kvabase = bp->b_kvaalloc; 2941 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize); 2942 goto has_addr; 2943 } 2944 2945 /* 2946 * Calculate the amount of the address space we would reserve 2947 * if the buffer was mapped. 2948 */ 2949 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 2950 offset = blkno * bsize; 2951 maxsize = size + (offset & PAGE_MASK); 2952 maxsize = imax(maxsize, bsize); 2953 2954mapping_loop: 2955 if (allocbufkva(bp, maxsize, gbflags)) { 2956 /* 2957 * Request defragmentation. getnewbuf() returns us the 2958 * allocated space by the scratch buffer KVA. 2959 */ 2960 scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags | 2961 (GB_UNMAPPED | GB_KVAALLOC)); 2962 if (scratch_bp == NULL) { 2963 if ((gbflags & GB_NOWAIT_BD) != 0) { 2964 /* 2965 * XXXKIB: defragmentation cannot 2966 * succeed, not sure what else to do. 2967 */ 2968 panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp); 2969 } 2970 atomic_add_int(&mappingrestarts, 1); 2971 goto mapping_loop; 2972 } 2973 KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0, 2974 ("scratch bp !B_KVAALLOC %p", scratch_bp)); 2975 setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc, 2976 scratch_bp->b_kvasize, gbflags); 2977 2978 /* Get rid of the scratch buffer. */ 2979 scratch_bp->b_kvasize = 0; 2980 scratch_bp->b_flags |= B_INVAL; 2981 scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC); 2982 brelse(scratch_bp); 2983 } 2984 if (!need_mapping) 2985 return; 2986 2987has_addr: 2988 bp->b_saveaddr = bp->b_kvabase; 2989 bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */ 2990 bp->b_flags &= ~B_UNMAPPED; 2991 BUF_CHECK_MAPPED(bp); 2992 bpmap_qenter(bp); 2993} 2994 2995/* 2996 * getblk: 2997 * 2998 * Get a block given a specified block and offset into a file/device. 2999 * The buffers B_DONE bit will be cleared on return, making it almost 3000 * ready for an I/O initiation. B_INVAL may or may not be set on 3001 * return. The caller should clear B_INVAL prior to initiating a 3002 * READ. 3003 * 3004 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 3005 * an existing buffer. 3006 * 3007 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 3008 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 3009 * and then cleared based on the backing VM. If the previous buffer is 3010 * non-0-sized but invalid, B_CACHE will be cleared. 3011 * 3012 * If getblk() must create a new buffer, the new buffer is returned with 3013 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 3014 * case it is returned with B_INVAL clear and B_CACHE set based on the 3015 * backing VM. 3016 * 3017 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 3018 * B_CACHE bit is clear. 3019 * 3020 * What this means, basically, is that the caller should use B_CACHE to 3021 * determine whether the buffer is fully valid or not and should clear 3022 * B_INVAL prior to issuing a read. If the caller intends to validate 3023 * the buffer by loading its data area with something, the caller needs 3024 * to clear B_INVAL. If the caller does this without issuing an I/O, 3025 * the caller should set B_CACHE ( as an optimization ), else the caller 3026 * should issue the I/O and biodone() will set B_CACHE if the I/O was 3027 * a write attempt or if it was a successfull read. If the caller 3028 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 3029 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 3030 */ 3031struct buf * 3032getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 3033 int flags) 3034{ 3035 struct buf *bp; 3036 struct bufobj *bo; 3037 int bsize, error, maxsize, vmio; 3038 off_t offset; 3039 3040 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 3041 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3042 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3043 ASSERT_VOP_LOCKED(vp, "getblk"); 3044 if (size > MAXBSIZE) 3045 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 3046 if (!unmapped_buf_allowed) 3047 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3048 3049 bo = &vp->v_bufobj; 3050loop: 3051 BO_RLOCK(bo); 3052 bp = gbincore(bo, blkno); 3053 if (bp != NULL) { 3054 int lockflags; 3055 /* 3056 * Buffer is in-core. If the buffer is not busy nor managed, 3057 * it must be on a queue. 3058 */ 3059 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 3060 3061 if (flags & GB_LOCK_NOWAIT) 3062 lockflags |= LK_NOWAIT; 3063 3064 error = BUF_TIMELOCK(bp, lockflags, 3065 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); 3066 3067 /* 3068 * If we slept and got the lock we have to restart in case 3069 * the buffer changed identities. 3070 */ 3071 if (error == ENOLCK) 3072 goto loop; 3073 /* We timed out or were interrupted. */ 3074 else if (error) 3075 return (NULL); 3076 /* If recursed, assume caller knows the rules. */ 3077 else if (BUF_LOCKRECURSED(bp)) 3078 goto end; 3079 3080 /* 3081 * The buffer is locked. B_CACHE is cleared if the buffer is 3082 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 3083 * and for a VMIO buffer B_CACHE is adjusted according to the 3084 * backing VM cache. 3085 */ 3086 if (bp->b_flags & B_INVAL) 3087 bp->b_flags &= ~B_CACHE; 3088 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 3089 bp->b_flags |= B_CACHE; 3090 if (bp->b_flags & B_MANAGED) 3091 MPASS(bp->b_qindex == QUEUE_NONE); 3092 else 3093 bremfree(bp); 3094 3095 /* 3096 * check for size inconsistencies for non-VMIO case. 3097 */ 3098 if (bp->b_bcount != size) { 3099 if ((bp->b_flags & B_VMIO) == 0 || 3100 (size > bp->b_kvasize)) { 3101 if (bp->b_flags & B_DELWRI) { 3102 /* 3103 * If buffer is pinned and caller does 3104 * not want sleep waiting for it to be 3105 * unpinned, bail out 3106 * */ 3107 if (bp->b_pin_count > 0) { 3108 if (flags & GB_LOCK_NOWAIT) { 3109 bqrelse(bp); 3110 return (NULL); 3111 } else { 3112 bunpin_wait(bp); 3113 } 3114 } 3115 bp->b_flags |= B_NOCACHE; 3116 bwrite(bp); 3117 } else { 3118 if (LIST_EMPTY(&bp->b_dep)) { 3119 bp->b_flags |= B_RELBUF; 3120 brelse(bp); 3121 } else { 3122 bp->b_flags |= B_NOCACHE; 3123 bwrite(bp); 3124 } 3125 } 3126 goto loop; 3127 } 3128 } 3129 3130 /* 3131 * Handle the case of unmapped buffer which should 3132 * become mapped, or the buffer for which KVA 3133 * reservation is requested. 3134 */ 3135 bp_unmapped_get_kva(bp, blkno, size, flags); 3136 3137 /* 3138 * If the size is inconsistant in the VMIO case, we can resize 3139 * the buffer. This might lead to B_CACHE getting set or 3140 * cleared. If the size has not changed, B_CACHE remains 3141 * unchanged from its previous state. 3142 */ 3143 if (bp->b_bcount != size) 3144 allocbuf(bp, size); 3145 3146 KASSERT(bp->b_offset != NOOFFSET, 3147 ("getblk: no buffer offset")); 3148 3149 /* 3150 * A buffer with B_DELWRI set and B_CACHE clear must 3151 * be committed before we can return the buffer in 3152 * order to prevent the caller from issuing a read 3153 * ( due to B_CACHE not being set ) and overwriting 3154 * it. 3155 * 3156 * Most callers, including NFS and FFS, need this to 3157 * operate properly either because they assume they 3158 * can issue a read if B_CACHE is not set, or because 3159 * ( for example ) an uncached B_DELWRI might loop due 3160 * to softupdates re-dirtying the buffer. In the latter 3161 * case, B_CACHE is set after the first write completes, 3162 * preventing further loops. 3163 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3164 * above while extending the buffer, we cannot allow the 3165 * buffer to remain with B_CACHE set after the write 3166 * completes or it will represent a corrupt state. To 3167 * deal with this we set B_NOCACHE to scrap the buffer 3168 * after the write. 3169 * 3170 * We might be able to do something fancy, like setting 3171 * B_CACHE in bwrite() except if B_DELWRI is already set, 3172 * so the below call doesn't set B_CACHE, but that gets real 3173 * confusing. This is much easier. 3174 */ 3175 3176 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3177 bp->b_flags |= B_NOCACHE; 3178 bwrite(bp); 3179 goto loop; 3180 } 3181 bp->b_flags &= ~B_DONE; 3182 } else { 3183 /* 3184 * Buffer is not in-core, create new buffer. The buffer 3185 * returned by getnewbuf() is locked. Note that the returned 3186 * buffer is also considered valid (not marked B_INVAL). 3187 */ 3188 BO_RUNLOCK(bo); 3189 /* 3190 * If the user does not want us to create the buffer, bail out 3191 * here. 3192 */ 3193 if (flags & GB_NOCREAT) 3194 return NULL; 3195 if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread)) 3196 return NULL; 3197 3198 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 3199 offset = blkno * bsize; 3200 vmio = vp->v_object != NULL; 3201 if (vmio) { 3202 maxsize = size + (offset & PAGE_MASK); 3203 } else { 3204 maxsize = size; 3205 /* Do not allow non-VMIO notmapped buffers. */ 3206 flags &= ~GB_UNMAPPED; 3207 } 3208 maxsize = imax(maxsize, bsize); 3209 3210 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 3211 if (bp == NULL) { 3212 if (slpflag || slptimeo) 3213 return NULL; 3214 goto loop; 3215 } 3216 3217 /* 3218 * This code is used to make sure that a buffer is not 3219 * created while the getnewbuf routine is blocked. 3220 * This can be a problem whether the vnode is locked or not. 3221 * If the buffer is created out from under us, we have to 3222 * throw away the one we just created. 3223 * 3224 * Note: this must occur before we associate the buffer 3225 * with the vp especially considering limitations in 3226 * the splay tree implementation when dealing with duplicate 3227 * lblkno's. 3228 */ 3229 BO_LOCK(bo); 3230 if (gbincore(bo, blkno)) { 3231 BO_UNLOCK(bo); 3232 bp->b_flags |= B_INVAL; 3233 brelse(bp); 3234 goto loop; 3235 } 3236 3237 /* 3238 * Insert the buffer into the hash, so that it can 3239 * be found by incore. 3240 */ 3241 bp->b_blkno = bp->b_lblkno = blkno; 3242 bp->b_offset = offset; 3243 bgetvp(vp, bp); 3244 BO_UNLOCK(bo); 3245 3246 /* 3247 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 3248 * buffer size starts out as 0, B_CACHE will be set by 3249 * allocbuf() for the VMIO case prior to it testing the 3250 * backing store for validity. 3251 */ 3252 3253 if (vmio) { 3254 bp->b_flags |= B_VMIO; 3255 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 3256 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 3257 bp, vp->v_object, bp->b_bufobj->bo_object)); 3258 } else { 3259 bp->b_flags &= ~B_VMIO; 3260 KASSERT(bp->b_bufobj->bo_object == NULL, 3261 ("ARGH! has b_bufobj->bo_object %p %p\n", 3262 bp, bp->b_bufobj->bo_object)); 3263 BUF_CHECK_MAPPED(bp); 3264 } 3265 3266 allocbuf(bp, size); 3267 bp->b_flags &= ~B_DONE; 3268 } 3269 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 3270 BUF_ASSERT_HELD(bp); 3271end: 3272 KASSERT(bp->b_bufobj == bo, 3273 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3274 return (bp); 3275} 3276 3277/* 3278 * Get an empty, disassociated buffer of given size. The buffer is initially 3279 * set to B_INVAL. 3280 */ 3281struct buf * 3282geteblk(int size, int flags) 3283{ 3284 struct buf *bp; 3285 int maxsize; 3286 3287 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3288 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 3289 if ((flags & GB_NOWAIT_BD) && 3290 (curthread->td_pflags & TDP_BUFNEED) != 0) 3291 return (NULL); 3292 } 3293 allocbuf(bp, size); 3294 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3295 BUF_ASSERT_HELD(bp); 3296 return (bp); 3297} 3298 3299 3300/* 3301 * This code constitutes the buffer memory from either anonymous system 3302 * memory (in the case of non-VMIO operations) or from an associated 3303 * VM object (in the case of VMIO operations). This code is able to 3304 * resize a buffer up or down. 3305 * 3306 * Note that this code is tricky, and has many complications to resolve 3307 * deadlock or inconsistant data situations. Tread lightly!!! 3308 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3309 * the caller. Calling this code willy nilly can result in the loss of data. 3310 * 3311 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3312 * B_CACHE for the non-VMIO case. 3313 */ 3314 3315int 3316allocbuf(struct buf *bp, int size) 3317{ 3318 int newbsize, mbsize; 3319 int i; 3320 3321 BUF_ASSERT_HELD(bp); 3322 3323 if (bp->b_kvasize < size) 3324 panic("allocbuf: buffer too small"); 3325 3326 if ((bp->b_flags & B_VMIO) == 0) { 3327 caddr_t origbuf; 3328 int origbufsize; 3329 /* 3330 * Just get anonymous memory from the kernel. Don't 3331 * mess with B_CACHE. 3332 */ 3333 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3334 if (bp->b_flags & B_MALLOC) 3335 newbsize = mbsize; 3336 else 3337 newbsize = round_page(size); 3338 3339 if (newbsize < bp->b_bufsize) { 3340 /* 3341 * malloced buffers are not shrunk 3342 */ 3343 if (bp->b_flags & B_MALLOC) { 3344 if (newbsize) { 3345 bp->b_bcount = size; 3346 } else { 3347 free(bp->b_data, M_BIOBUF); 3348 if (bp->b_bufsize) { 3349 atomic_subtract_long( 3350 &bufmallocspace, 3351 bp->b_bufsize); 3352 bufspacewakeup(); 3353 bp->b_bufsize = 0; 3354 } 3355 bp->b_saveaddr = bp->b_kvabase; 3356 bp->b_data = bp->b_saveaddr; 3357 bp->b_bcount = 0; 3358 bp->b_flags &= ~B_MALLOC; 3359 } 3360 return 1; 3361 } 3362 vm_hold_free_pages(bp, newbsize); 3363 } else if (newbsize > bp->b_bufsize) { 3364 /* 3365 * We only use malloced memory on the first allocation. 3366 * and revert to page-allocated memory when the buffer 3367 * grows. 3368 */ 3369 /* 3370 * There is a potential smp race here that could lead 3371 * to bufmallocspace slightly passing the max. It 3372 * is probably extremely rare and not worth worrying 3373 * over. 3374 */ 3375 if ( (bufmallocspace < maxbufmallocspace) && 3376 (bp->b_bufsize == 0) && 3377 (mbsize <= PAGE_SIZE/2)) { 3378 3379 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 3380 bp->b_bufsize = mbsize; 3381 bp->b_bcount = size; 3382 bp->b_flags |= B_MALLOC; 3383 atomic_add_long(&bufmallocspace, mbsize); 3384 return 1; 3385 } 3386 origbuf = NULL; 3387 origbufsize = 0; 3388 /* 3389 * If the buffer is growing on its other-than-first allocation, 3390 * then we revert to the page-allocation scheme. 3391 */ 3392 if (bp->b_flags & B_MALLOC) { 3393 origbuf = bp->b_data; 3394 origbufsize = bp->b_bufsize; 3395 bp->b_data = bp->b_kvabase; 3396 if (bp->b_bufsize) { 3397 atomic_subtract_long(&bufmallocspace, 3398 bp->b_bufsize); 3399 bufspacewakeup(); 3400 bp->b_bufsize = 0; 3401 } 3402 bp->b_flags &= ~B_MALLOC; 3403 newbsize = round_page(newbsize); 3404 } 3405 vm_hold_load_pages( 3406 bp, 3407 (vm_offset_t) bp->b_data + bp->b_bufsize, 3408 (vm_offset_t) bp->b_data + newbsize); 3409 if (origbuf) { 3410 bcopy(origbuf, bp->b_data, origbufsize); 3411 free(origbuf, M_BIOBUF); 3412 } 3413 } 3414 } else { 3415 int desiredpages; 3416 3417 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3418 desiredpages = (size == 0) ? 0 : 3419 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 3420 3421 if (bp->b_flags & B_MALLOC) 3422 panic("allocbuf: VMIO buffer can't be malloced"); 3423 /* 3424 * Set B_CACHE initially if buffer is 0 length or will become 3425 * 0-length. 3426 */ 3427 if (size == 0 || bp->b_bufsize == 0) 3428 bp->b_flags |= B_CACHE; 3429 3430 if (newbsize < bp->b_bufsize) { 3431 /* 3432 * DEV_BSIZE aligned new buffer size is less then the 3433 * DEV_BSIZE aligned existing buffer size. Figure out 3434 * if we have to remove any pages. 3435 */ 3436 if (desiredpages < bp->b_npages) { 3437 vm_page_t m; 3438 3439 if ((bp->b_flags & B_UNMAPPED) == 0) { 3440 BUF_CHECK_MAPPED(bp); 3441 pmap_qremove((vm_offset_t)trunc_page( 3442 (vm_offset_t)bp->b_data) + 3443 (desiredpages << PAGE_SHIFT), 3444 (bp->b_npages - desiredpages)); 3445 } else 3446 BUF_CHECK_UNMAPPED(bp); 3447 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3448 for (i = desiredpages; i < bp->b_npages; i++) { 3449 /* 3450 * the page is not freed here -- it 3451 * is the responsibility of 3452 * vnode_pager_setsize 3453 */ 3454 m = bp->b_pages[i]; 3455 KASSERT(m != bogus_page, 3456 ("allocbuf: bogus page found")); 3457 while (vm_page_sleep_if_busy(m, 3458 "biodep")) 3459 continue; 3460 3461 bp->b_pages[i] = NULL; 3462 vm_page_lock(m); 3463 vm_page_unwire(m, 0); 3464 vm_page_unlock(m); 3465 } 3466 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3467 bp->b_npages = desiredpages; 3468 } 3469 } else if (size > bp->b_bcount) { 3470 /* 3471 * We are growing the buffer, possibly in a 3472 * byte-granular fashion. 3473 */ 3474 vm_object_t obj; 3475 vm_offset_t toff; 3476 vm_offset_t tinc; 3477 3478 /* 3479 * Step 1, bring in the VM pages from the object, 3480 * allocating them if necessary. We must clear 3481 * B_CACHE if these pages are not valid for the 3482 * range covered by the buffer. 3483 */ 3484 3485 obj = bp->b_bufobj->bo_object; 3486 3487 VM_OBJECT_WLOCK(obj); 3488 while (bp->b_npages < desiredpages) { 3489 vm_page_t m; 3490 3491 /* 3492 * We must allocate system pages since blocking 3493 * here could interfere with paging I/O, no 3494 * matter which process we are. 3495 * 3496 * Only exclusive busy can be tested here. 3497 * Blocking on shared busy might lead to 3498 * deadlocks once allocbuf() is called after 3499 * pages are vfs_busy_pages(). 3500 */ 3501 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + 3502 bp->b_npages, VM_ALLOC_NOBUSY | 3503 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3504 VM_ALLOC_IGN_SBUSY | 3505 VM_ALLOC_COUNT(desiredpages - bp->b_npages)); 3506 if (m->valid == 0) 3507 bp->b_flags &= ~B_CACHE; 3508 bp->b_pages[bp->b_npages] = m; 3509 ++bp->b_npages; 3510 } 3511 3512 /* 3513 * Step 2. We've loaded the pages into the buffer, 3514 * we have to figure out if we can still have B_CACHE 3515 * set. Note that B_CACHE is set according to the 3516 * byte-granular range ( bcount and size ), new the 3517 * aligned range ( newbsize ). 3518 * 3519 * The VM test is against m->valid, which is DEV_BSIZE 3520 * aligned. Needless to say, the validity of the data 3521 * needs to also be DEV_BSIZE aligned. Note that this 3522 * fails with NFS if the server or some other client 3523 * extends the file's EOF. If our buffer is resized, 3524 * B_CACHE may remain set! XXX 3525 */ 3526 3527 toff = bp->b_bcount; 3528 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3529 3530 while ((bp->b_flags & B_CACHE) && toff < size) { 3531 vm_pindex_t pi; 3532 3533 if (tinc > (size - toff)) 3534 tinc = size - toff; 3535 3536 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3537 PAGE_SHIFT; 3538 3539 vfs_buf_test_cache( 3540 bp, 3541 bp->b_offset, 3542 toff, 3543 tinc, 3544 bp->b_pages[pi] 3545 ); 3546 toff += tinc; 3547 tinc = PAGE_SIZE; 3548 } 3549 VM_OBJECT_WUNLOCK(obj); 3550 3551 /* 3552 * Step 3, fixup the KVM pmap. 3553 */ 3554 if ((bp->b_flags & B_UNMAPPED) == 0) 3555 bpmap_qenter(bp); 3556 else 3557 BUF_CHECK_UNMAPPED(bp); 3558 } 3559 } 3560 if (newbsize < bp->b_bufsize) 3561 bufspacewakeup(); 3562 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3563 bp->b_bcount = size; /* requested buffer size */ 3564 return 1; 3565} 3566 3567extern int inflight_transient_maps; 3568 3569void 3570biodone(struct bio *bp) 3571{ 3572 struct mtx *mtxp; 3573 void (*done)(struct bio *); 3574 vm_offset_t start, end; 3575 3576 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 3577 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING; 3578 bp->bio_flags |= BIO_UNMAPPED; 3579 start = trunc_page((vm_offset_t)bp->bio_data); 3580 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 3581 pmap_qremove(start, OFF_TO_IDX(end - start)); 3582 vmem_free(transient_arena, start, end - start); 3583 atomic_add_int(&inflight_transient_maps, -1); 3584 } 3585 done = bp->bio_done; 3586 if (done == NULL) { 3587 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3588 mtx_lock(mtxp); 3589 bp->bio_flags |= BIO_DONE; 3590 wakeup(bp); 3591 mtx_unlock(mtxp); 3592 } else { 3593 bp->bio_flags |= BIO_DONE; 3594 done(bp); 3595 } 3596} 3597 3598/* 3599 * Wait for a BIO to finish. 3600 */ 3601int 3602biowait(struct bio *bp, const char *wchan) 3603{ 3604 struct mtx *mtxp; 3605 3606 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3607 mtx_lock(mtxp); 3608 while ((bp->bio_flags & BIO_DONE) == 0) 3609 msleep(bp, mtxp, PRIBIO, wchan, 0); 3610 mtx_unlock(mtxp); 3611 if (bp->bio_error != 0) 3612 return (bp->bio_error); 3613 if (!(bp->bio_flags & BIO_ERROR)) 3614 return (0); 3615 return (EIO); 3616} 3617 3618void 3619biofinish(struct bio *bp, struct devstat *stat, int error) 3620{ 3621 3622 if (error) { 3623 bp->bio_error = error; 3624 bp->bio_flags |= BIO_ERROR; 3625 } 3626 if (stat != NULL) 3627 devstat_end_transaction_bio(stat, bp); 3628 biodone(bp); 3629} 3630 3631/* 3632 * bufwait: 3633 * 3634 * Wait for buffer I/O completion, returning error status. The buffer 3635 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3636 * error and cleared. 3637 */ 3638int 3639bufwait(struct buf *bp) 3640{ 3641 if (bp->b_iocmd == BIO_READ) 3642 bwait(bp, PRIBIO, "biord"); 3643 else 3644 bwait(bp, PRIBIO, "biowr"); 3645 if (bp->b_flags & B_EINTR) { 3646 bp->b_flags &= ~B_EINTR; 3647 return (EINTR); 3648 } 3649 if (bp->b_ioflags & BIO_ERROR) { 3650 return (bp->b_error ? bp->b_error : EIO); 3651 } else { 3652 return (0); 3653 } 3654} 3655 3656 /* 3657 * Call back function from struct bio back up to struct buf. 3658 */ 3659static void 3660bufdonebio(struct bio *bip) 3661{ 3662 struct buf *bp; 3663 3664 bp = bip->bio_caller2; 3665 bp->b_resid = bp->b_bcount - bip->bio_completed; 3666 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3667 bp->b_ioflags = bip->bio_flags; 3668 bp->b_error = bip->bio_error; 3669 if (bp->b_error) 3670 bp->b_ioflags |= BIO_ERROR; 3671 bufdone(bp); 3672 g_destroy_bio(bip); 3673} 3674 3675void 3676dev_strategy(struct cdev *dev, struct buf *bp) 3677{ 3678 struct cdevsw *csw; 3679 int ref; 3680 3681 KASSERT(dev->si_refcount > 0, 3682 ("dev_strategy on un-referenced struct cdev *(%s) %p", 3683 devtoname(dev), dev)); 3684 3685 csw = dev_refthread(dev, &ref); 3686 dev_strategy_csw(dev, csw, bp); 3687 dev_relthread(dev, ref); 3688} 3689 3690void 3691dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp) 3692{ 3693 struct bio *bip; 3694 3695 KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE, 3696 ("b_iocmd botch")); 3697 KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) || 3698 dev->si_threadcount > 0, 3699 ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev), 3700 dev)); 3701 if (csw == NULL) { 3702 bp->b_error = ENXIO; 3703 bp->b_ioflags = BIO_ERROR; 3704 bufdone(bp); 3705 return; 3706 } 3707 for (;;) { 3708 bip = g_new_bio(); 3709 if (bip != NULL) 3710 break; 3711 /* Try again later */ 3712 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3713 } 3714 bip->bio_cmd = bp->b_iocmd; 3715 bip->bio_offset = bp->b_iooffset; 3716 bip->bio_length = bp->b_bcount; 3717 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3718 bdata2bio(bp, bip); 3719 bip->bio_done = bufdonebio; 3720 bip->bio_caller2 = bp; 3721 bip->bio_dev = dev; 3722 (*csw->d_strategy)(bip); 3723} 3724 3725/* 3726 * bufdone: 3727 * 3728 * Finish I/O on a buffer, optionally calling a completion function. 3729 * This is usually called from an interrupt so process blocking is 3730 * not allowed. 3731 * 3732 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3733 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3734 * assuming B_INVAL is clear. 3735 * 3736 * For the VMIO case, we set B_CACHE if the op was a read and no 3737 * read error occured, or if the op was a write. B_CACHE is never 3738 * set if the buffer is invalid or otherwise uncacheable. 3739 * 3740 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3741 * initiator to leave B_INVAL set to brelse the buffer out of existance 3742 * in the biodone routine. 3743 */ 3744void 3745bufdone(struct buf *bp) 3746{ 3747 struct bufobj *dropobj; 3748 void (*biodone)(struct buf *); 3749 3750 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3751 dropobj = NULL; 3752 3753 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3754 BUF_ASSERT_HELD(bp); 3755 3756 runningbufwakeup(bp); 3757 if (bp->b_iocmd == BIO_WRITE) 3758 dropobj = bp->b_bufobj; 3759 /* call optional completion function if requested */ 3760 if (bp->b_iodone != NULL) { 3761 biodone = bp->b_iodone; 3762 bp->b_iodone = NULL; 3763 (*biodone) (bp); 3764 if (dropobj) 3765 bufobj_wdrop(dropobj); 3766 return; 3767 } 3768 3769 bufdone_finish(bp); 3770 3771 if (dropobj) 3772 bufobj_wdrop(dropobj); 3773} 3774 3775void 3776bufdone_finish(struct buf *bp) 3777{ 3778 BUF_ASSERT_HELD(bp); 3779 3780 if (!LIST_EMPTY(&bp->b_dep)) 3781 buf_complete(bp); 3782 3783 if (bp->b_flags & B_VMIO) { 3784 vm_ooffset_t foff; 3785 vm_page_t m; 3786 vm_object_t obj; 3787 struct vnode *vp; 3788 int bogus, i, iosize; 3789 3790 obj = bp->b_bufobj->bo_object; 3791 KASSERT(obj->paging_in_progress >= bp->b_npages, 3792 ("biodone_finish: paging in progress(%d) < b_npages(%d)", 3793 obj->paging_in_progress, bp->b_npages)); 3794 3795 vp = bp->b_vp; 3796 KASSERT(vp->v_holdcnt > 0, 3797 ("biodone_finish: vnode %p has zero hold count", vp)); 3798 KASSERT(vp->v_object != NULL, 3799 ("biodone_finish: vnode %p has no vm_object", vp)); 3800 3801 foff = bp->b_offset; 3802 KASSERT(bp->b_offset != NOOFFSET, 3803 ("biodone_finish: bp %p has no buffer offset", bp)); 3804 3805 /* 3806 * Set B_CACHE if the op was a normal read and no error 3807 * occured. B_CACHE is set for writes in the b*write() 3808 * routines. 3809 */ 3810 iosize = bp->b_bcount - bp->b_resid; 3811 if (bp->b_iocmd == BIO_READ && 3812 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3813 !(bp->b_ioflags & BIO_ERROR)) { 3814 bp->b_flags |= B_CACHE; 3815 } 3816 bogus = 0; 3817 VM_OBJECT_WLOCK(obj); 3818 for (i = 0; i < bp->b_npages; i++) { 3819 int bogusflag = 0; 3820 int resid; 3821 3822 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3823 if (resid > iosize) 3824 resid = iosize; 3825 3826 /* 3827 * cleanup bogus pages, restoring the originals 3828 */ 3829 m = bp->b_pages[i]; 3830 if (m == bogus_page) { 3831 bogus = bogusflag = 1; 3832 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3833 if (m == NULL) 3834 panic("biodone: page disappeared!"); 3835 bp->b_pages[i] = m; 3836 } 3837 KASSERT(OFF_TO_IDX(foff) == m->pindex, 3838 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch", 3839 (intmax_t)foff, (uintmax_t)m->pindex)); 3840 3841 /* 3842 * In the write case, the valid and clean bits are 3843 * already changed correctly ( see bdwrite() ), so we 3844 * only need to do this here in the read case. 3845 */ 3846 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3847 KASSERT((m->dirty & vm_page_bits(foff & 3848 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3849 " page %p has unexpected dirty bits", m)); 3850 vfs_page_set_valid(bp, foff, m); 3851 } 3852 3853 vm_page_sunbusy(m); 3854 vm_object_pip_subtract(obj, 1); 3855 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3856 iosize -= resid; 3857 } 3858 vm_object_pip_wakeupn(obj, 0); 3859 VM_OBJECT_WUNLOCK(obj); 3860 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 3861 BUF_CHECK_MAPPED(bp); 3862 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3863 bp->b_pages, bp->b_npages); 3864 } 3865 } 3866 3867 /* 3868 * For asynchronous completions, release the buffer now. The brelse 3869 * will do a wakeup there if necessary - so no need to do a wakeup 3870 * here in the async case. The sync case always needs to do a wakeup. 3871 */ 3872 3873 if (bp->b_flags & B_ASYNC) { 3874 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3875 brelse(bp); 3876 else 3877 bqrelse(bp); 3878 } else 3879 bdone(bp); 3880} 3881 3882/* 3883 * This routine is called in lieu of iodone in the case of 3884 * incomplete I/O. This keeps the busy status for pages 3885 * consistant. 3886 */ 3887void 3888vfs_unbusy_pages(struct buf *bp) 3889{ 3890 int i; 3891 vm_object_t obj; 3892 vm_page_t m; 3893 3894 runningbufwakeup(bp); 3895 if (!(bp->b_flags & B_VMIO)) 3896 return; 3897 3898 obj = bp->b_bufobj->bo_object; 3899 VM_OBJECT_WLOCK(obj); 3900 for (i = 0; i < bp->b_npages; i++) { 3901 m = bp->b_pages[i]; 3902 if (m == bogus_page) { 3903 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3904 if (!m) 3905 panic("vfs_unbusy_pages: page missing\n"); 3906 bp->b_pages[i] = m; 3907 if ((bp->b_flags & B_UNMAPPED) == 0) { 3908 BUF_CHECK_MAPPED(bp); 3909 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3910 bp->b_pages, bp->b_npages); 3911 } else 3912 BUF_CHECK_UNMAPPED(bp); 3913 } 3914 vm_object_pip_subtract(obj, 1); 3915 vm_page_sunbusy(m); 3916 } 3917 vm_object_pip_wakeupn(obj, 0); 3918 VM_OBJECT_WUNLOCK(obj); 3919} 3920 3921/* 3922 * vfs_page_set_valid: 3923 * 3924 * Set the valid bits in a page based on the supplied offset. The 3925 * range is restricted to the buffer's size. 3926 * 3927 * This routine is typically called after a read completes. 3928 */ 3929static void 3930vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3931{ 3932 vm_ooffset_t eoff; 3933 3934 /* 3935 * Compute the end offset, eoff, such that [off, eoff) does not span a 3936 * page boundary and eoff is not greater than the end of the buffer. 3937 * The end of the buffer, in this case, is our file EOF, not the 3938 * allocation size of the buffer. 3939 */ 3940 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3941 if (eoff > bp->b_offset + bp->b_bcount) 3942 eoff = bp->b_offset + bp->b_bcount; 3943 3944 /* 3945 * Set valid range. This is typically the entire buffer and thus the 3946 * entire page. 3947 */ 3948 if (eoff > off) 3949 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 3950} 3951 3952/* 3953 * vfs_page_set_validclean: 3954 * 3955 * Set the valid bits and clear the dirty bits in a page based on the 3956 * supplied offset. The range is restricted to the buffer's size. 3957 */ 3958static void 3959vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3960{ 3961 vm_ooffset_t soff, eoff; 3962 3963 /* 3964 * Start and end offsets in buffer. eoff - soff may not cross a 3965 * page boundry or cross the end of the buffer. The end of the 3966 * buffer, in this case, is our file EOF, not the allocation size 3967 * of the buffer. 3968 */ 3969 soff = off; 3970 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3971 if (eoff > bp->b_offset + bp->b_bcount) 3972 eoff = bp->b_offset + bp->b_bcount; 3973 3974 /* 3975 * Set valid range. This is typically the entire buffer and thus the 3976 * entire page. 3977 */ 3978 if (eoff > soff) { 3979 vm_page_set_validclean( 3980 m, 3981 (vm_offset_t) (soff & PAGE_MASK), 3982 (vm_offset_t) (eoff - soff) 3983 ); 3984 } 3985} 3986 3987/* 3988 * Ensure that all buffer pages are not exclusive busied. If any page is 3989 * exclusive busy, drain it. 3990 */ 3991void 3992vfs_drain_busy_pages(struct buf *bp) 3993{ 3994 vm_page_t m; 3995 int i, last_busied; 3996 3997 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); 3998 last_busied = 0; 3999 for (i = 0; i < bp->b_npages; i++) { 4000 m = bp->b_pages[i]; 4001 if (vm_page_xbusied(m)) { 4002 for (; last_busied < i; last_busied++) 4003 vm_page_sbusy(bp->b_pages[last_busied]); 4004 while (vm_page_xbusied(m)) { 4005 vm_page_lock(m); 4006 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4007 vm_page_busy_sleep(m, "vbpage"); 4008 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4009 } 4010 } 4011 } 4012 for (i = 0; i < last_busied; i++) 4013 vm_page_sunbusy(bp->b_pages[i]); 4014} 4015 4016/* 4017 * This routine is called before a device strategy routine. 4018 * It is used to tell the VM system that paging I/O is in 4019 * progress, and treat the pages associated with the buffer 4020 * almost as being exclusive busy. Also the object paging_in_progress 4021 * flag is handled to make sure that the object doesn't become 4022 * inconsistant. 4023 * 4024 * Since I/O has not been initiated yet, certain buffer flags 4025 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 4026 * and should be ignored. 4027 */ 4028void 4029vfs_busy_pages(struct buf *bp, int clear_modify) 4030{ 4031 int i, bogus; 4032 vm_object_t obj; 4033 vm_ooffset_t foff; 4034 vm_page_t m; 4035 4036 if (!(bp->b_flags & B_VMIO)) 4037 return; 4038 4039 obj = bp->b_bufobj->bo_object; 4040 foff = bp->b_offset; 4041 KASSERT(bp->b_offset != NOOFFSET, 4042 ("vfs_busy_pages: no buffer offset")); 4043 VM_OBJECT_WLOCK(obj); 4044 vfs_drain_busy_pages(bp); 4045 if (bp->b_bufsize != 0) 4046 vfs_setdirty_locked_object(bp); 4047 bogus = 0; 4048 for (i = 0; i < bp->b_npages; i++) { 4049 m = bp->b_pages[i]; 4050 4051 if ((bp->b_flags & B_CLUSTER) == 0) { 4052 vm_object_pip_add(obj, 1); 4053 vm_page_sbusy(m); 4054 } 4055 /* 4056 * When readying a buffer for a read ( i.e 4057 * clear_modify == 0 ), it is important to do 4058 * bogus_page replacement for valid pages in 4059 * partially instantiated buffers. Partially 4060 * instantiated buffers can, in turn, occur when 4061 * reconstituting a buffer from its VM backing store 4062 * base. We only have to do this if B_CACHE is 4063 * clear ( which causes the I/O to occur in the 4064 * first place ). The replacement prevents the read 4065 * I/O from overwriting potentially dirty VM-backed 4066 * pages. XXX bogus page replacement is, uh, bogus. 4067 * It may not work properly with small-block devices. 4068 * We need to find a better way. 4069 */ 4070 if (clear_modify) { 4071 pmap_remove_write(m); 4072 vfs_page_set_validclean(bp, foff, m); 4073 } else if (m->valid == VM_PAGE_BITS_ALL && 4074 (bp->b_flags & B_CACHE) == 0) { 4075 bp->b_pages[i] = bogus_page; 4076 bogus++; 4077 } 4078 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4079 } 4080 VM_OBJECT_WUNLOCK(obj); 4081 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 4082 BUF_CHECK_MAPPED(bp); 4083 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4084 bp->b_pages, bp->b_npages); 4085 } 4086} 4087 4088/* 4089 * vfs_bio_set_valid: 4090 * 4091 * Set the range within the buffer to valid. The range is 4092 * relative to the beginning of the buffer, b_offset. Note that 4093 * b_offset itself may be offset from the beginning of the first 4094 * page. 4095 */ 4096void 4097vfs_bio_set_valid(struct buf *bp, int base, int size) 4098{ 4099 int i, n; 4100 vm_page_t m; 4101 4102 if (!(bp->b_flags & B_VMIO)) 4103 return; 4104 4105 /* 4106 * Fixup base to be relative to beginning of first page. 4107 * Set initial n to be the maximum number of bytes in the 4108 * first page that can be validated. 4109 */ 4110 base += (bp->b_offset & PAGE_MASK); 4111 n = PAGE_SIZE - (base & PAGE_MASK); 4112 4113 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4114 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4115 m = bp->b_pages[i]; 4116 if (n > size) 4117 n = size; 4118 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4119 base += n; 4120 size -= n; 4121 n = PAGE_SIZE; 4122 } 4123 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4124} 4125 4126/* 4127 * vfs_bio_clrbuf: 4128 * 4129 * If the specified buffer is a non-VMIO buffer, clear the entire 4130 * buffer. If the specified buffer is a VMIO buffer, clear and 4131 * validate only the previously invalid portions of the buffer. 4132 * This routine essentially fakes an I/O, so we need to clear 4133 * BIO_ERROR and B_INVAL. 4134 * 4135 * Note that while we only theoretically need to clear through b_bcount, 4136 * we go ahead and clear through b_bufsize. 4137 */ 4138void 4139vfs_bio_clrbuf(struct buf *bp) 4140{ 4141 int i, j, mask, sa, ea, slide; 4142 4143 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4144 clrbuf(bp); 4145 return; 4146 } 4147 bp->b_flags &= ~B_INVAL; 4148 bp->b_ioflags &= ~BIO_ERROR; 4149 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4150 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4151 (bp->b_offset & PAGE_MASK) == 0) { 4152 if (bp->b_pages[0] == bogus_page) 4153 goto unlock; 4154 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4155 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object); 4156 if ((bp->b_pages[0]->valid & mask) == mask) 4157 goto unlock; 4158 if ((bp->b_pages[0]->valid & mask) == 0) { 4159 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize); 4160 bp->b_pages[0]->valid |= mask; 4161 goto unlock; 4162 } 4163 } 4164 sa = bp->b_offset & PAGE_MASK; 4165 slide = 0; 4166 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4167 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4168 ea = slide & PAGE_MASK; 4169 if (ea == 0) 4170 ea = PAGE_SIZE; 4171 if (bp->b_pages[i] == bogus_page) 4172 continue; 4173 j = sa / DEV_BSIZE; 4174 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4175 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object); 4176 if ((bp->b_pages[i]->valid & mask) == mask) 4177 continue; 4178 if ((bp->b_pages[i]->valid & mask) == 0) 4179 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4180 else { 4181 for (; sa < ea; sa += DEV_BSIZE, j++) { 4182 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4183 pmap_zero_page_area(bp->b_pages[i], 4184 sa, DEV_BSIZE); 4185 } 4186 } 4187 } 4188 bp->b_pages[i]->valid |= mask; 4189 } 4190unlock: 4191 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4192 bp->b_resid = 0; 4193} 4194 4195void 4196vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4197{ 4198 vm_page_t m; 4199 int i, n; 4200 4201 if ((bp->b_flags & B_UNMAPPED) == 0) { 4202 BUF_CHECK_MAPPED(bp); 4203 bzero(bp->b_data + base, size); 4204 } else { 4205 BUF_CHECK_UNMAPPED(bp); 4206 n = PAGE_SIZE - (base & PAGE_MASK); 4207 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4208 m = bp->b_pages[i]; 4209 if (n > size) 4210 n = size; 4211 pmap_zero_page_area(m, base & PAGE_MASK, n); 4212 base += n; 4213 size -= n; 4214 n = PAGE_SIZE; 4215 } 4216 } 4217} 4218 4219/* 4220 * vm_hold_load_pages and vm_hold_free_pages get pages into 4221 * a buffers address space. The pages are anonymous and are 4222 * not associated with a file object. 4223 */ 4224static void 4225vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4226{ 4227 vm_offset_t pg; 4228 vm_page_t p; 4229 int index; 4230 4231 BUF_CHECK_MAPPED(bp); 4232 4233 to = round_page(to); 4234 from = round_page(from); 4235 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4236 4237 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4238tryagain: 4239 /* 4240 * note: must allocate system pages since blocking here 4241 * could interfere with paging I/O, no matter which 4242 * process we are. 4243 */ 4244 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | 4245 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT)); 4246 if (p == NULL) { 4247 VM_WAIT; 4248 goto tryagain; 4249 } 4250 pmap_qenter(pg, &p, 1); 4251 bp->b_pages[index] = p; 4252 } 4253 bp->b_npages = index; 4254} 4255 4256/* Return pages associated with this buf to the vm system */ 4257static void 4258vm_hold_free_pages(struct buf *bp, int newbsize) 4259{ 4260 vm_offset_t from; 4261 vm_page_t p; 4262 int index, newnpages; 4263 4264 BUF_CHECK_MAPPED(bp); 4265 4266 from = round_page((vm_offset_t)bp->b_data + newbsize); 4267 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4268 if (bp->b_npages > newnpages) 4269 pmap_qremove(from, bp->b_npages - newnpages); 4270 for (index = newnpages; index < bp->b_npages; index++) { 4271 p = bp->b_pages[index]; 4272 bp->b_pages[index] = NULL; 4273 if (vm_page_sbusied(p)) 4274 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 4275 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); 4276 p->wire_count--; 4277 vm_page_free(p); 4278 atomic_subtract_int(&cnt.v_wire_count, 1); 4279 } 4280 bp->b_npages = newnpages; 4281} 4282 4283/* 4284 * Map an IO request into kernel virtual address space. 4285 * 4286 * All requests are (re)mapped into kernel VA space. 4287 * Notice that we use b_bufsize for the size of the buffer 4288 * to be mapped. b_bcount might be modified by the driver. 4289 * 4290 * Note that even if the caller determines that the address space should 4291 * be valid, a race or a smaller-file mapped into a larger space may 4292 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 4293 * check the return value. 4294 */ 4295int 4296vmapbuf(struct buf *bp, int mapbuf) 4297{ 4298 caddr_t kva; 4299 vm_prot_t prot; 4300 int pidx; 4301 4302 if (bp->b_bufsize < 0) 4303 return (-1); 4304 prot = VM_PROT_READ; 4305 if (bp->b_iocmd == BIO_READ) 4306 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 4307 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 4308 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, 4309 btoc(MAXPHYS))) < 0) 4310 return (-1); 4311 bp->b_npages = pidx; 4312 if (mapbuf || !unmapped_buf_allowed) { 4313 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 4314 kva = bp->b_saveaddr; 4315 bp->b_saveaddr = bp->b_data; 4316 bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK); 4317 bp->b_flags &= ~B_UNMAPPED; 4318 } else { 4319 bp->b_flags |= B_UNMAPPED; 4320 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK; 4321 bp->b_saveaddr = bp->b_data; 4322 bp->b_data = unmapped_buf; 4323 } 4324 return(0); 4325} 4326 4327/* 4328 * Free the io map PTEs associated with this IO operation. 4329 * We also invalidate the TLB entries and restore the original b_addr. 4330 */ 4331void 4332vunmapbuf(struct buf *bp) 4333{ 4334 int npages; 4335 4336 npages = bp->b_npages; 4337 if (bp->b_flags & B_UNMAPPED) 4338 bp->b_flags &= ~B_UNMAPPED; 4339 else 4340 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4341 vm_page_unhold_pages(bp->b_pages, npages); 4342 4343 bp->b_data = bp->b_saveaddr; 4344} 4345 4346void 4347bdone(struct buf *bp) 4348{ 4349 struct mtx *mtxp; 4350 4351 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4352 mtx_lock(mtxp); 4353 bp->b_flags |= B_DONE; 4354 wakeup(bp); 4355 mtx_unlock(mtxp); 4356} 4357 4358void 4359bwait(struct buf *bp, u_char pri, const char *wchan) 4360{ 4361 struct mtx *mtxp; 4362 4363 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4364 mtx_lock(mtxp); 4365 while ((bp->b_flags & B_DONE) == 0) 4366 msleep(bp, mtxp, pri, wchan, 0); 4367 mtx_unlock(mtxp); 4368} 4369 4370int 4371bufsync(struct bufobj *bo, int waitfor) 4372{ 4373 4374 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 4375} 4376 4377void 4378bufstrategy(struct bufobj *bo, struct buf *bp) 4379{ 4380 int i = 0; 4381 struct vnode *vp; 4382 4383 vp = bp->b_vp; 4384 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 4385 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 4386 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 4387 i = VOP_STRATEGY(vp, bp); 4388 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 4389} 4390 4391void 4392bufobj_wrefl(struct bufobj *bo) 4393{ 4394 4395 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4396 ASSERT_BO_WLOCKED(bo); 4397 bo->bo_numoutput++; 4398} 4399 4400void 4401bufobj_wref(struct bufobj *bo) 4402{ 4403 4404 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4405 BO_LOCK(bo); 4406 bo->bo_numoutput++; 4407 BO_UNLOCK(bo); 4408} 4409 4410void 4411bufobj_wdrop(struct bufobj *bo) 4412{ 4413 4414 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 4415 BO_LOCK(bo); 4416 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 4417 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 4418 bo->bo_flag &= ~BO_WWAIT; 4419 wakeup(&bo->bo_numoutput); 4420 } 4421 BO_UNLOCK(bo); 4422} 4423 4424int 4425bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 4426{ 4427 int error; 4428 4429 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 4430 ASSERT_BO_WLOCKED(bo); 4431 error = 0; 4432 while (bo->bo_numoutput) { 4433 bo->bo_flag |= BO_WWAIT; 4434 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), 4435 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 4436 if (error) 4437 break; 4438 } 4439 return (error); 4440} 4441 4442void 4443bpin(struct buf *bp) 4444{ 4445 struct mtx *mtxp; 4446 4447 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4448 mtx_lock(mtxp); 4449 bp->b_pin_count++; 4450 mtx_unlock(mtxp); 4451} 4452 4453void 4454bunpin(struct buf *bp) 4455{ 4456 struct mtx *mtxp; 4457 4458 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4459 mtx_lock(mtxp); 4460 if (--bp->b_pin_count == 0) 4461 wakeup(bp); 4462 mtx_unlock(mtxp); 4463} 4464 4465void 4466bunpin_wait(struct buf *bp) 4467{ 4468 struct mtx *mtxp; 4469 4470 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4471 mtx_lock(mtxp); 4472 while (bp->b_pin_count > 0) 4473 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4474 mtx_unlock(mtxp); 4475} 4476 4477/* 4478 * Set bio_data or bio_ma for struct bio from the struct buf. 4479 */ 4480void 4481bdata2bio(struct buf *bp, struct bio *bip) 4482{ 4483 4484 if ((bp->b_flags & B_UNMAPPED) != 0) { 4485 KASSERT(unmapped_buf_allowed, ("unmapped")); 4486 bip->bio_ma = bp->b_pages; 4487 bip->bio_ma_n = bp->b_npages; 4488 bip->bio_data = unmapped_buf; 4489 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 4490 bip->bio_flags |= BIO_UNMAPPED; 4491 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 4492 PAGE_SIZE == bp->b_npages, 4493 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset, 4494 (long long)bip->bio_length, bip->bio_ma_n)); 4495 } else { 4496 bip->bio_data = bp->b_data; 4497 bip->bio_ma = NULL; 4498 } 4499} 4500 4501#include "opt_ddb.h" 4502#ifdef DDB 4503#include <ddb/ddb.h> 4504 4505/* DDB command to show buffer data */ 4506DB_SHOW_COMMAND(buffer, db_show_buffer) 4507{ 4508 /* get args */ 4509 struct buf *bp = (struct buf *)addr; 4510 4511 if (!have_addr) { 4512 db_printf("usage: show buffer <addr>\n"); 4513 return; 4514 } 4515 4516 db_printf("buf at %p\n", bp); 4517 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n", 4518 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, 4519 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS); 4520 db_printf( 4521 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4522 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, " 4523 "b_dep = %p\n", 4524 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4525 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4526 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first); 4527 if (bp->b_npages) { 4528 int i; 4529 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4530 for (i = 0; i < bp->b_npages; i++) { 4531 vm_page_t m; 4532 m = bp->b_pages[i]; 4533 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4534 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4535 if ((i + 1) < bp->b_npages) 4536 db_printf(","); 4537 } 4538 db_printf("\n"); 4539 } 4540 db_printf(" "); 4541 BUF_LOCKPRINTINFO(bp); 4542} 4543 4544DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4545{ 4546 struct buf *bp; 4547 int i; 4548 4549 for (i = 0; i < nbuf; i++) { 4550 bp = &buf[i]; 4551 if (BUF_ISLOCKED(bp)) { 4552 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4553 db_printf("\n"); 4554 } 4555 } 4556} 4557 4558DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4559{ 4560 struct vnode *vp; 4561 struct buf *bp; 4562 4563 if (!have_addr) { 4564 db_printf("usage: show vnodebufs <addr>\n"); 4565 return; 4566 } 4567 vp = (struct vnode *)addr; 4568 db_printf("Clean buffers:\n"); 4569 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4570 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4571 db_printf("\n"); 4572 } 4573 db_printf("Dirty buffers:\n"); 4574 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4575 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4576 db_printf("\n"); 4577 } 4578} 4579 4580DB_COMMAND(countfreebufs, db_coundfreebufs) 4581{ 4582 struct buf *bp; 4583 int i, used = 0, nfree = 0; 4584 4585 if (have_addr) { 4586 db_printf("usage: countfreebufs\n"); 4587 return; 4588 } 4589 4590 for (i = 0; i < nbuf; i++) { 4591 bp = &buf[i]; 4592 if ((bp->b_flags & B_INFREECNT) != 0) 4593 nfree++; 4594 else 4595 used++; 4596 } 4597 4598 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4599 nfree + used); 4600 db_printf("numfreebuffers is %d\n", numfreebuffers); 4601} 4602#endif /* DDB */ 4603