vfs_bio.c revision 284944
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 284944 2015-06-30 05:53:15Z 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(struct vnode *vp, int); 117static int flushbufqueues(struct vnode *, 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 CTASSERT(MAXBCACHEBUF >= MAXBSIZE); 778 mtx_init(&bqclean, "bufq clean lock", NULL, MTX_DEF); 779 mtx_init(&bqdirty, "bufq dirty lock", NULL, MTX_DEF); 780 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 781 rw_init(&nblock, "needsbuffer lock"); 782 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 783 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF); 784 785 /* next, make a null set of free lists */ 786 for (i = 0; i < BUFFER_QUEUES; i++) 787 TAILQ_INIT(&bufqueues[i]); 788 789 /* finally, initialize each buffer header and stick on empty q */ 790 for (i = 0; i < nbuf; i++) { 791 bp = &buf[i]; 792 bzero(bp, sizeof *bp); 793 bp->b_flags = B_INVAL | B_INFREECNT; 794 bp->b_rcred = NOCRED; 795 bp->b_wcred = NOCRED; 796 bp->b_qindex = QUEUE_EMPTY; 797 bp->b_xflags = 0; 798 LIST_INIT(&bp->b_dep); 799 BUF_LOCKINIT(bp); 800 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 801#ifdef INVARIANTS 802 bq_len[QUEUE_EMPTY]++; 803#endif 804 } 805 806 /* 807 * maxbufspace is the absolute maximum amount of buffer space we are 808 * allowed to reserve in KVM and in real terms. The absolute maximum 809 * is nominally used by buf_daemon. hibufspace is the nominal maximum 810 * used by most other processes. The differential is required to 811 * ensure that buf_daemon is able to run when other processes might 812 * be blocked waiting for buffer space. 813 * 814 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 815 * this may result in KVM fragmentation which is not handled optimally 816 * by the system. 817 */ 818 maxbufspace = (long)nbuf * BKVASIZE; 819 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBCACHEBUF * 10); 820 lobufspace = hibufspace - MAXBCACHEBUF; 821 822 /* 823 * Note: The 16 MiB upper limit for hirunningspace was chosen 824 * arbitrarily and may need further tuning. It corresponds to 825 * 128 outstanding write IO requests (if IO size is 128 KiB), 826 * which fits with many RAID controllers' tagged queuing limits. 827 * The lower 1 MiB limit is the historical upper limit for 828 * hirunningspace. 829 */ 830 hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBCACHEBUF), 831 16 * 1024 * 1024), 1024 * 1024); 832 lorunningspace = roundup((hirunningspace * 2) / 3, MAXBCACHEBUF); 833 834/* 835 * Limit the amount of malloc memory since it is wired permanently into 836 * the kernel space. Even though this is accounted for in the buffer 837 * allocation, we don't want the malloced region to grow uncontrolled. 838 * The malloc scheme improves memory utilization significantly on average 839 * (small) directories. 840 */ 841 maxbufmallocspace = hibufspace / 20; 842 843/* 844 * Reduce the chance of a deadlock occuring by limiting the number 845 * of delayed-write dirty buffers we allow to stack up. 846 */ 847 hidirtybuffers = nbuf / 4 + 20; 848 dirtybufthresh = hidirtybuffers * 9 / 10; 849 numdirtybuffers = 0; 850/* 851 * To support extreme low-memory systems, make sure hidirtybuffers cannot 852 * eat up all available buffer space. This occurs when our minimum cannot 853 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 854 * BKVASIZE'd buffers. 855 */ 856 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 857 hidirtybuffers >>= 1; 858 } 859 lodirtybuffers = hidirtybuffers / 2; 860 861/* 862 * Try to keep the number of free buffers in the specified range, 863 * and give special processes (e.g. like buf_daemon) access to an 864 * emergency reserve. 865 */ 866 lofreebuffers = nbuf / 18 + 5; 867 hifreebuffers = 2 * lofreebuffers; 868 numfreebuffers = nbuf; 869 870 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 871 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 872 unmapped_buf = (caddr_t)kva_alloc(MAXPHYS); 873} 874 875#ifdef INVARIANTS 876static inline void 877vfs_buf_check_mapped(struct buf *bp) 878{ 879 880 KASSERT((bp->b_flags & B_UNMAPPED) == 0, 881 ("mapped buf %p %x", bp, bp->b_flags)); 882 KASSERT(bp->b_kvabase != unmapped_buf, 883 ("mapped buf: b_kvabase was not updated %p", bp)); 884 KASSERT(bp->b_data != unmapped_buf, 885 ("mapped buf: b_data was not updated %p", bp)); 886} 887 888static inline void 889vfs_buf_check_unmapped(struct buf *bp) 890{ 891 892 KASSERT((bp->b_flags & B_UNMAPPED) == B_UNMAPPED, 893 ("unmapped buf %p %x", bp, bp->b_flags)); 894 KASSERT(bp->b_kvabase == unmapped_buf, 895 ("unmapped buf: corrupted b_kvabase %p", bp)); 896 KASSERT(bp->b_data == unmapped_buf, 897 ("unmapped buf: corrupted b_data %p", bp)); 898} 899 900#define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) 901#define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) 902#else 903#define BUF_CHECK_MAPPED(bp) do {} while (0) 904#define BUF_CHECK_UNMAPPED(bp) do {} while (0) 905#endif 906 907static void 908bpmap_qenter(struct buf *bp) 909{ 910 911 BUF_CHECK_MAPPED(bp); 912 913 /* 914 * bp->b_data is relative to bp->b_offset, but 915 * bp->b_offset may be offset into the first page. 916 */ 917 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 918 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 919 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 920 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 921} 922 923/* 924 * bfreekva() - free the kva allocation for a buffer. 925 * 926 * Since this call frees up buffer space, we call bufspacewakeup(). 927 */ 928static void 929bfreekva(struct buf *bp) 930{ 931 932 if (bp->b_kvasize == 0) 933 return; 934 935 atomic_add_int(&buffreekvacnt, 1); 936 atomic_subtract_long(&bufspace, bp->b_kvasize); 937 if ((bp->b_flags & B_UNMAPPED) == 0) { 938 BUF_CHECK_MAPPED(bp); 939 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, 940 bp->b_kvasize); 941 } else { 942 BUF_CHECK_UNMAPPED(bp); 943 if ((bp->b_flags & B_KVAALLOC) != 0) { 944 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvaalloc, 945 bp->b_kvasize); 946 } 947 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize); 948 bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC); 949 } 950 bp->b_kvasize = 0; 951 bufspacewakeup(); 952} 953 954/* 955 * binsfree: 956 * 957 * Insert the buffer into the appropriate free list. 958 */ 959static void 960binsfree(struct buf *bp, int qindex) 961{ 962 struct mtx *olock, *nlock; 963 964 BUF_ASSERT_XLOCKED(bp); 965 966 nlock = bqlock(qindex); 967 /* Handle delayed bremfree() processing. */ 968 if (bp->b_flags & B_REMFREE) { 969 olock = bqlock(bp->b_qindex); 970 mtx_lock(olock); 971 bremfreel(bp); 972 if (olock != nlock) { 973 mtx_unlock(olock); 974 mtx_lock(nlock); 975 } 976 } else 977 mtx_lock(nlock); 978 979 if (bp->b_qindex != QUEUE_NONE) 980 panic("binsfree: free buffer onto another queue???"); 981 982 bp->b_qindex = qindex; 983 if (bp->b_flags & B_AGE) 984 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 985 else 986 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 987#ifdef INVARIANTS 988 bq_len[bp->b_qindex]++; 989#endif 990 mtx_unlock(nlock); 991 992 /* 993 * Something we can maybe free or reuse. 994 */ 995 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 996 bufspacewakeup(); 997 998 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 999 bufcountadd(bp); 1000} 1001 1002/* 1003 * bremfree: 1004 * 1005 * Mark the buffer for removal from the appropriate free list. 1006 * 1007 */ 1008void 1009bremfree(struct buf *bp) 1010{ 1011 1012 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1013 KASSERT((bp->b_flags & B_REMFREE) == 0, 1014 ("bremfree: buffer %p already marked for delayed removal.", bp)); 1015 KASSERT(bp->b_qindex != QUEUE_NONE, 1016 ("bremfree: buffer %p not on a queue.", bp)); 1017 BUF_ASSERT_XLOCKED(bp); 1018 1019 bp->b_flags |= B_REMFREE; 1020 bufcountsub(bp); 1021} 1022 1023/* 1024 * bremfreef: 1025 * 1026 * Force an immediate removal from a free list. Used only in nfs when 1027 * it abuses the b_freelist pointer. 1028 */ 1029void 1030bremfreef(struct buf *bp) 1031{ 1032 struct mtx *qlock; 1033 1034 qlock = bqlock(bp->b_qindex); 1035 mtx_lock(qlock); 1036 bremfreel(bp); 1037 mtx_unlock(qlock); 1038} 1039 1040/* 1041 * bremfreel: 1042 * 1043 * Removes a buffer from the free list, must be called with the 1044 * correct qlock held. 1045 */ 1046static void 1047bremfreel(struct buf *bp) 1048{ 1049 1050 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 1051 bp, bp->b_vp, bp->b_flags); 1052 KASSERT(bp->b_qindex != QUEUE_NONE, 1053 ("bremfreel: buffer %p not on a queue.", bp)); 1054 BUF_ASSERT_XLOCKED(bp); 1055 mtx_assert(bqlock(bp->b_qindex), MA_OWNED); 1056 1057 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 1058#ifdef INVARIANTS 1059 KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow", 1060 bp->b_qindex)); 1061 bq_len[bp->b_qindex]--; 1062#endif 1063 bp->b_qindex = QUEUE_NONE; 1064 /* 1065 * If this was a delayed bremfree() we only need to remove the buffer 1066 * from the queue and return the stats are already done. 1067 */ 1068 if (bp->b_flags & B_REMFREE) { 1069 bp->b_flags &= ~B_REMFREE; 1070 return; 1071 } 1072 bufcountsub(bp); 1073} 1074 1075/* 1076 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 1077 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 1078 * the buffer is valid and we do not have to do anything. 1079 */ 1080void 1081breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 1082 int cnt, struct ucred * cred) 1083{ 1084 struct buf *rabp; 1085 int i; 1086 1087 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 1088 if (inmem(vp, *rablkno)) 1089 continue; 1090 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 1091 1092 if ((rabp->b_flags & B_CACHE) == 0) { 1093 if (!TD_IS_IDLETHREAD(curthread)) 1094 curthread->td_ru.ru_inblock++; 1095 rabp->b_flags |= B_ASYNC; 1096 rabp->b_flags &= ~B_INVAL; 1097 rabp->b_ioflags &= ~BIO_ERROR; 1098 rabp->b_iocmd = BIO_READ; 1099 if (rabp->b_rcred == NOCRED && cred != NOCRED) 1100 rabp->b_rcred = crhold(cred); 1101 vfs_busy_pages(rabp, 0); 1102 BUF_KERNPROC(rabp); 1103 rabp->b_iooffset = dbtob(rabp->b_blkno); 1104 bstrategy(rabp); 1105 } else { 1106 brelse(rabp); 1107 } 1108 } 1109} 1110 1111/* 1112 * Entry point for bread() and breadn() via #defines in sys/buf.h. 1113 * 1114 * Get a buffer with the specified data. Look in the cache first. We 1115 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 1116 * is set, the buffer is valid and we do not have to do anything, see 1117 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 1118 */ 1119int 1120breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno, 1121 int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp) 1122{ 1123 struct buf *bp; 1124 int rv = 0, readwait = 0; 1125 1126 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 1127 /* 1128 * Can only return NULL if GB_LOCK_NOWAIT flag is specified. 1129 */ 1130 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags); 1131 if (bp == NULL) 1132 return (EBUSY); 1133 1134 /* if not found in cache, do some I/O */ 1135 if ((bp->b_flags & B_CACHE) == 0) { 1136 if (!TD_IS_IDLETHREAD(curthread)) 1137 curthread->td_ru.ru_inblock++; 1138 bp->b_iocmd = BIO_READ; 1139 bp->b_flags &= ~B_INVAL; 1140 bp->b_ioflags &= ~BIO_ERROR; 1141 if (bp->b_rcred == NOCRED && cred != NOCRED) 1142 bp->b_rcred = crhold(cred); 1143 vfs_busy_pages(bp, 0); 1144 bp->b_iooffset = dbtob(bp->b_blkno); 1145 bstrategy(bp); 1146 ++readwait; 1147 } 1148 1149 breada(vp, rablkno, rabsize, cnt, cred); 1150 1151 if (readwait) { 1152 rv = bufwait(bp); 1153 } 1154 return (rv); 1155} 1156 1157/* 1158 * Write, release buffer on completion. (Done by iodone 1159 * if async). Do not bother writing anything if the buffer 1160 * is invalid. 1161 * 1162 * Note that we set B_CACHE here, indicating that buffer is 1163 * fully valid and thus cacheable. This is true even of NFS 1164 * now so we set it generally. This could be set either here 1165 * or in biodone() since the I/O is synchronous. We put it 1166 * here. 1167 */ 1168int 1169bufwrite(struct buf *bp) 1170{ 1171 int oldflags; 1172 struct vnode *vp; 1173 long space; 1174 int vp_md; 1175 1176 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1177 if (bp->b_flags & B_INVAL) { 1178 brelse(bp); 1179 return (0); 1180 } 1181 1182 if (bp->b_flags & B_BARRIER) 1183 barrierwrites++; 1184 1185 oldflags = bp->b_flags; 1186 1187 BUF_ASSERT_HELD(bp); 1188 1189 if (bp->b_pin_count > 0) 1190 bunpin_wait(bp); 1191 1192 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 1193 ("FFS background buffer should not get here %p", bp)); 1194 1195 vp = bp->b_vp; 1196 if (vp) 1197 vp_md = vp->v_vflag & VV_MD; 1198 else 1199 vp_md = 0; 1200 1201 /* 1202 * Mark the buffer clean. Increment the bufobj write count 1203 * before bundirty() call, to prevent other thread from seeing 1204 * empty dirty list and zero counter for writes in progress, 1205 * falsely indicating that the bufobj is clean. 1206 */ 1207 bufobj_wref(bp->b_bufobj); 1208 bundirty(bp); 1209 1210 bp->b_flags &= ~B_DONE; 1211 bp->b_ioflags &= ~BIO_ERROR; 1212 bp->b_flags |= B_CACHE; 1213 bp->b_iocmd = BIO_WRITE; 1214 1215 vfs_busy_pages(bp, 1); 1216 1217 /* 1218 * Normal bwrites pipeline writes 1219 */ 1220 bp->b_runningbufspace = bp->b_bufsize; 1221 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace); 1222 1223 if (!TD_IS_IDLETHREAD(curthread)) 1224 curthread->td_ru.ru_oublock++; 1225 if (oldflags & B_ASYNC) 1226 BUF_KERNPROC(bp); 1227 bp->b_iooffset = dbtob(bp->b_blkno); 1228 bstrategy(bp); 1229 1230 if ((oldflags & B_ASYNC) == 0) { 1231 int rtval = bufwait(bp); 1232 brelse(bp); 1233 return (rtval); 1234 } else if (space > hirunningspace) { 1235 /* 1236 * don't allow the async write to saturate the I/O 1237 * system. We will not deadlock here because 1238 * we are blocking waiting for I/O that is already in-progress 1239 * to complete. We do not block here if it is the update 1240 * or syncer daemon trying to clean up as that can lead 1241 * to deadlock. 1242 */ 1243 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 1244 waitrunningbufspace(); 1245 } 1246 1247 return (0); 1248} 1249 1250void 1251bufbdflush(struct bufobj *bo, struct buf *bp) 1252{ 1253 struct buf *nbp; 1254 1255 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 1256 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 1257 altbufferflushes++; 1258 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 1259 BO_LOCK(bo); 1260 /* 1261 * Try to find a buffer to flush. 1262 */ 1263 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 1264 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1265 BUF_LOCK(nbp, 1266 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1267 continue; 1268 if (bp == nbp) 1269 panic("bdwrite: found ourselves"); 1270 BO_UNLOCK(bo); 1271 /* Don't countdeps with the bo lock held. */ 1272 if (buf_countdeps(nbp, 0)) { 1273 BO_LOCK(bo); 1274 BUF_UNLOCK(nbp); 1275 continue; 1276 } 1277 if (nbp->b_flags & B_CLUSTEROK) { 1278 vfs_bio_awrite(nbp); 1279 } else { 1280 bremfree(nbp); 1281 bawrite(nbp); 1282 } 1283 dirtybufferflushes++; 1284 break; 1285 } 1286 if (nbp == NULL) 1287 BO_UNLOCK(bo); 1288 } 1289} 1290 1291/* 1292 * Delayed write. (Buffer is marked dirty). Do not bother writing 1293 * anything if the buffer is marked invalid. 1294 * 1295 * Note that since the buffer must be completely valid, we can safely 1296 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1297 * biodone() in order to prevent getblk from writing the buffer 1298 * out synchronously. 1299 */ 1300void 1301bdwrite(struct buf *bp) 1302{ 1303 struct thread *td = curthread; 1304 struct vnode *vp; 1305 struct bufobj *bo; 1306 1307 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1308 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1309 KASSERT((bp->b_flags & B_BARRIER) == 0, 1310 ("Barrier request in delayed write %p", bp)); 1311 BUF_ASSERT_HELD(bp); 1312 1313 if (bp->b_flags & B_INVAL) { 1314 brelse(bp); 1315 return; 1316 } 1317 1318 /* 1319 * If we have too many dirty buffers, don't create any more. 1320 * If we are wildly over our limit, then force a complete 1321 * cleanup. Otherwise, just keep the situation from getting 1322 * out of control. Note that we have to avoid a recursive 1323 * disaster and not try to clean up after our own cleanup! 1324 */ 1325 vp = bp->b_vp; 1326 bo = bp->b_bufobj; 1327 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 1328 td->td_pflags |= TDP_INBDFLUSH; 1329 BO_BDFLUSH(bo, bp); 1330 td->td_pflags &= ~TDP_INBDFLUSH; 1331 } else 1332 recursiveflushes++; 1333 1334 bdirty(bp); 1335 /* 1336 * Set B_CACHE, indicating that the buffer is fully valid. This is 1337 * true even of NFS now. 1338 */ 1339 bp->b_flags |= B_CACHE; 1340 1341 /* 1342 * This bmap keeps the system from needing to do the bmap later, 1343 * perhaps when the system is attempting to do a sync. Since it 1344 * is likely that the indirect block -- or whatever other datastructure 1345 * that the filesystem needs is still in memory now, it is a good 1346 * thing to do this. Note also, that if the pageout daemon is 1347 * requesting a sync -- there might not be enough memory to do 1348 * the bmap then... So, this is important to do. 1349 */ 1350 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1351 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1352 } 1353 1354 /* 1355 * Set the *dirty* buffer range based upon the VM system dirty 1356 * pages. 1357 * 1358 * Mark the buffer pages as clean. We need to do this here to 1359 * satisfy the vnode_pager and the pageout daemon, so that it 1360 * thinks that the pages have been "cleaned". Note that since 1361 * the pages are in a delayed write buffer -- the VFS layer 1362 * "will" see that the pages get written out on the next sync, 1363 * or perhaps the cluster will be completed. 1364 */ 1365 vfs_clean_pages_dirty_buf(bp); 1366 bqrelse(bp); 1367 1368 /* 1369 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1370 * due to the softdep code. 1371 */ 1372} 1373 1374/* 1375 * bdirty: 1376 * 1377 * Turn buffer into delayed write request. We must clear BIO_READ and 1378 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1379 * itself to properly update it in the dirty/clean lists. We mark it 1380 * B_DONE to ensure that any asynchronization of the buffer properly 1381 * clears B_DONE ( else a panic will occur later ). 1382 * 1383 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1384 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1385 * should only be called if the buffer is known-good. 1386 * 1387 * Since the buffer is not on a queue, we do not update the numfreebuffers 1388 * count. 1389 * 1390 * The buffer must be on QUEUE_NONE. 1391 */ 1392void 1393bdirty(struct buf *bp) 1394{ 1395 1396 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1397 bp, bp->b_vp, bp->b_flags); 1398 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1399 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1400 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1401 BUF_ASSERT_HELD(bp); 1402 bp->b_flags &= ~(B_RELBUF); 1403 bp->b_iocmd = BIO_WRITE; 1404 1405 if ((bp->b_flags & B_DELWRI) == 0) { 1406 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1407 reassignbuf(bp); 1408 bdirtyadd(); 1409 } 1410} 1411 1412/* 1413 * bundirty: 1414 * 1415 * Clear B_DELWRI for buffer. 1416 * 1417 * Since the buffer is not on a queue, we do not update the numfreebuffers 1418 * count. 1419 * 1420 * The buffer must be on QUEUE_NONE. 1421 */ 1422 1423void 1424bundirty(struct buf *bp) 1425{ 1426 1427 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1428 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1429 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1430 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1431 BUF_ASSERT_HELD(bp); 1432 1433 if (bp->b_flags & B_DELWRI) { 1434 bp->b_flags &= ~B_DELWRI; 1435 reassignbuf(bp); 1436 bdirtysub(); 1437 } 1438 /* 1439 * Since it is now being written, we can clear its deferred write flag. 1440 */ 1441 bp->b_flags &= ~B_DEFERRED; 1442} 1443 1444/* 1445 * bawrite: 1446 * 1447 * Asynchronous write. Start output on a buffer, but do not wait for 1448 * it to complete. The buffer is released when the output completes. 1449 * 1450 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1451 * B_INVAL buffers. Not us. 1452 */ 1453void 1454bawrite(struct buf *bp) 1455{ 1456 1457 bp->b_flags |= B_ASYNC; 1458 (void) bwrite(bp); 1459} 1460 1461/* 1462 * babarrierwrite: 1463 * 1464 * Asynchronous barrier write. Start output on a buffer, but do not 1465 * wait for it to complete. Place a write barrier after this write so 1466 * that this buffer and all buffers written before it are committed to 1467 * the disk before any buffers written after this write are committed 1468 * to the disk. The buffer is released when the output completes. 1469 */ 1470void 1471babarrierwrite(struct buf *bp) 1472{ 1473 1474 bp->b_flags |= B_ASYNC | B_BARRIER; 1475 (void) bwrite(bp); 1476} 1477 1478/* 1479 * bbarrierwrite: 1480 * 1481 * Synchronous barrier write. Start output on a buffer and wait for 1482 * it to complete. Place a write barrier after this write so that 1483 * this buffer and all buffers written before it are committed to 1484 * the disk before any buffers written after this write are committed 1485 * to the disk. The buffer is released when the output completes. 1486 */ 1487int 1488bbarrierwrite(struct buf *bp) 1489{ 1490 1491 bp->b_flags |= B_BARRIER; 1492 return (bwrite(bp)); 1493} 1494 1495/* 1496 * bwillwrite: 1497 * 1498 * Called prior to the locking of any vnodes when we are expecting to 1499 * write. We do not want to starve the buffer cache with too many 1500 * dirty buffers so we block here. By blocking prior to the locking 1501 * of any vnodes we attempt to avoid the situation where a locked vnode 1502 * prevents the various system daemons from flushing related buffers. 1503 */ 1504void 1505bwillwrite(void) 1506{ 1507 1508 if (numdirtybuffers >= hidirtybuffers) { 1509 mtx_lock(&bdirtylock); 1510 while (numdirtybuffers >= hidirtybuffers) { 1511 bdirtywait = 1; 1512 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4), 1513 "flswai", 0); 1514 } 1515 mtx_unlock(&bdirtylock); 1516 } 1517} 1518 1519/* 1520 * Return true if we have too many dirty buffers. 1521 */ 1522int 1523buf_dirty_count_severe(void) 1524{ 1525 1526 return(numdirtybuffers >= hidirtybuffers); 1527} 1528 1529static __noinline int 1530buf_vm_page_count_severe(void) 1531{ 1532 1533 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1); 1534 1535 return vm_page_count_severe(); 1536} 1537 1538/* 1539 * brelse: 1540 * 1541 * Release a busy buffer and, if requested, free its resources. The 1542 * buffer will be stashed in the appropriate bufqueue[] allowing it 1543 * to be accessed later as a cache entity or reused for other purposes. 1544 */ 1545void 1546brelse(struct buf *bp) 1547{ 1548 int qindex; 1549 1550 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1551 bp, bp->b_vp, bp->b_flags); 1552 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1553 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1554 1555 if (BUF_LOCKRECURSED(bp)) { 1556 /* 1557 * Do not process, in particular, do not handle the 1558 * B_INVAL/B_RELBUF and do not release to free list. 1559 */ 1560 BUF_UNLOCK(bp); 1561 return; 1562 } 1563 1564 if (bp->b_flags & B_MANAGED) { 1565 bqrelse(bp); 1566 return; 1567 } 1568 1569 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1570 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1571 /* 1572 * Failed write, redirty. Must clear BIO_ERROR to prevent 1573 * pages from being scrapped. If the error is anything 1574 * other than an I/O error (EIO), assume that retrying 1575 * is futile. 1576 */ 1577 bp->b_ioflags &= ~BIO_ERROR; 1578 bdirty(bp); 1579 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1580 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1581 /* 1582 * Either a failed I/O or we were asked to free or not 1583 * cache the buffer. 1584 */ 1585 bp->b_flags |= B_INVAL; 1586 if (!LIST_EMPTY(&bp->b_dep)) 1587 buf_deallocate(bp); 1588 if (bp->b_flags & B_DELWRI) 1589 bdirtysub(); 1590 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1591 if ((bp->b_flags & B_VMIO) == 0) { 1592 if (bp->b_bufsize) 1593 allocbuf(bp, 0); 1594 if (bp->b_vp) 1595 brelvp(bp); 1596 } 1597 } 1598 1599 /* 1600 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1601 * is called with B_DELWRI set, the underlying pages may wind up 1602 * getting freed causing a previous write (bdwrite()) to get 'lost' 1603 * because pages associated with a B_DELWRI bp are marked clean. 1604 * 1605 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1606 * if B_DELWRI is set. 1607 * 1608 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1609 * on pages to return pages to the VM page queues. 1610 */ 1611 if (bp->b_flags & B_DELWRI) 1612 bp->b_flags &= ~B_RELBUF; 1613 else if (buf_vm_page_count_severe()) { 1614 /* 1615 * BKGRDINPROG can only be set with the buf and bufobj 1616 * locks both held. We tolerate a race to clear it here. 1617 */ 1618 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1619 bp->b_flags |= B_RELBUF; 1620 } 1621 1622 /* 1623 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1624 * constituted, not even NFS buffers now. Two flags effect this. If 1625 * B_INVAL, the struct buf is invalidated but the VM object is kept 1626 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1627 * 1628 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1629 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1630 * buffer is also B_INVAL because it hits the re-dirtying code above. 1631 * 1632 * Normally we can do this whether a buffer is B_DELWRI or not. If 1633 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1634 * the commit state and we cannot afford to lose the buffer. If the 1635 * buffer has a background write in progress, we need to keep it 1636 * around to prevent it from being reconstituted and starting a second 1637 * background write. 1638 */ 1639 if ((bp->b_flags & B_VMIO) 1640 && !(bp->b_vp->v_mount != NULL && 1641 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1642 !vn_isdisk(bp->b_vp, NULL) && 1643 (bp->b_flags & B_DELWRI)) 1644 ) { 1645 1646 int i, j, resid; 1647 vm_page_t m; 1648 off_t foff; 1649 vm_pindex_t poff; 1650 vm_object_t obj; 1651 1652 obj = bp->b_bufobj->bo_object; 1653 1654 /* 1655 * Get the base offset and length of the buffer. Note that 1656 * in the VMIO case if the buffer block size is not 1657 * page-aligned then b_data pointer may not be page-aligned. 1658 * But our b_pages[] array *IS* page aligned. 1659 * 1660 * block sizes less then DEV_BSIZE (usually 512) are not 1661 * supported due to the page granularity bits (m->valid, 1662 * m->dirty, etc...). 1663 * 1664 * See man buf(9) for more information 1665 */ 1666 resid = bp->b_bufsize; 1667 foff = bp->b_offset; 1668 for (i = 0; i < bp->b_npages; i++) { 1669 int had_bogus = 0; 1670 1671 m = bp->b_pages[i]; 1672 1673 /* 1674 * If we hit a bogus page, fixup *all* the bogus pages 1675 * now. 1676 */ 1677 if (m == bogus_page) { 1678 poff = OFF_TO_IDX(bp->b_offset); 1679 had_bogus = 1; 1680 1681 VM_OBJECT_RLOCK(obj); 1682 for (j = i; j < bp->b_npages; j++) { 1683 vm_page_t mtmp; 1684 mtmp = bp->b_pages[j]; 1685 if (mtmp == bogus_page) { 1686 mtmp = vm_page_lookup(obj, poff + j); 1687 if (!mtmp) { 1688 panic("brelse: page missing\n"); 1689 } 1690 bp->b_pages[j] = mtmp; 1691 } 1692 } 1693 VM_OBJECT_RUNLOCK(obj); 1694 1695 if ((bp->b_flags & (B_INVAL | B_UNMAPPED)) == 0) { 1696 BUF_CHECK_MAPPED(bp); 1697 pmap_qenter( 1698 trunc_page((vm_offset_t)bp->b_data), 1699 bp->b_pages, bp->b_npages); 1700 } 1701 m = bp->b_pages[i]; 1702 } 1703 if ((bp->b_flags & B_NOCACHE) || 1704 (bp->b_ioflags & BIO_ERROR && 1705 bp->b_iocmd == BIO_READ)) { 1706 int poffset = foff & PAGE_MASK; 1707 int presid = resid > (PAGE_SIZE - poffset) ? 1708 (PAGE_SIZE - poffset) : resid; 1709 1710 KASSERT(presid >= 0, ("brelse: extra page")); 1711 VM_OBJECT_WLOCK(obj); 1712 while (vm_page_xbusied(m)) { 1713 vm_page_lock(m); 1714 VM_OBJECT_WUNLOCK(obj); 1715 vm_page_busy_sleep(m, "mbncsh"); 1716 VM_OBJECT_WLOCK(obj); 1717 } 1718 if (pmap_page_wired_mappings(m) == 0) 1719 vm_page_set_invalid(m, poffset, presid); 1720 VM_OBJECT_WUNLOCK(obj); 1721 if (had_bogus) 1722 printf("avoided corruption bug in bogus_page/brelse code\n"); 1723 } 1724 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1725 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1726 } 1727 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1728 vfs_vmio_release(bp); 1729 1730 } else if (bp->b_flags & B_VMIO) { 1731 1732 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1733 vfs_vmio_release(bp); 1734 } 1735 1736 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1737 if (bp->b_bufsize != 0) 1738 allocbuf(bp, 0); 1739 if (bp->b_vp != NULL) 1740 brelvp(bp); 1741 } 1742 1743 /* 1744 * If the buffer has junk contents signal it and eventually 1745 * clean up B_DELWRI and diassociate the vnode so that gbincore() 1746 * doesn't find it. 1747 */ 1748 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1749 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1750 bp->b_flags |= B_INVAL; 1751 if (bp->b_flags & B_INVAL) { 1752 if (bp->b_flags & B_DELWRI) 1753 bundirty(bp); 1754 if (bp->b_vp) 1755 brelvp(bp); 1756 } 1757 1758 /* buffers with no memory */ 1759 if (bp->b_bufsize == 0) { 1760 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1761 if (bp->b_vflags & BV_BKGRDINPROG) 1762 panic("losing buffer 1"); 1763 if (bp->b_kvasize) 1764 qindex = QUEUE_EMPTYKVA; 1765 else 1766 qindex = QUEUE_EMPTY; 1767 bp->b_flags |= B_AGE; 1768 /* buffers with junk contents */ 1769 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1770 (bp->b_ioflags & BIO_ERROR)) { 1771 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1772 if (bp->b_vflags & BV_BKGRDINPROG) 1773 panic("losing buffer 2"); 1774 qindex = QUEUE_CLEAN; 1775 bp->b_flags |= B_AGE; 1776 /* remaining buffers */ 1777 } else if (bp->b_flags & B_DELWRI) 1778 qindex = QUEUE_DIRTY; 1779 else 1780 qindex = QUEUE_CLEAN; 1781 1782 binsfree(bp, qindex); 1783 1784 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1785 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1786 panic("brelse: not dirty"); 1787 /* unlock */ 1788 BUF_UNLOCK(bp); 1789} 1790 1791/* 1792 * Release a buffer back to the appropriate queue but do not try to free 1793 * it. The buffer is expected to be used again soon. 1794 * 1795 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1796 * biodone() to requeue an async I/O on completion. It is also used when 1797 * known good buffers need to be requeued but we think we may need the data 1798 * again soon. 1799 * 1800 * XXX we should be able to leave the B_RELBUF hint set on completion. 1801 */ 1802void 1803bqrelse(struct buf *bp) 1804{ 1805 int qindex; 1806 1807 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1808 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1809 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1810 1811 if (BUF_LOCKRECURSED(bp)) { 1812 /* do not release to free list */ 1813 BUF_UNLOCK(bp); 1814 return; 1815 } 1816 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1817 1818 if (bp->b_flags & B_MANAGED) { 1819 if (bp->b_flags & B_REMFREE) 1820 bremfreef(bp); 1821 goto out; 1822 } 1823 1824 /* buffers with stale but valid contents */ 1825 if (bp->b_flags & B_DELWRI) { 1826 qindex = QUEUE_DIRTY; 1827 } else { 1828 if ((bp->b_flags & B_DELWRI) == 0 && 1829 (bp->b_xflags & BX_VNDIRTY)) 1830 panic("bqrelse: not dirty"); 1831 /* 1832 * BKGRDINPROG can only be set with the buf and bufobj 1833 * locks both held. We tolerate a race to clear it here. 1834 */ 1835 if (buf_vm_page_count_severe() && 1836 (bp->b_vflags & BV_BKGRDINPROG) == 0) { 1837 /* 1838 * We are too low on memory, we have to try to free 1839 * the buffer (most importantly: the wired pages 1840 * making up its backing store) *now*. 1841 */ 1842 brelse(bp); 1843 return; 1844 } 1845 qindex = QUEUE_CLEAN; 1846 } 1847 binsfree(bp, qindex); 1848 1849out: 1850 /* unlock */ 1851 BUF_UNLOCK(bp); 1852} 1853 1854/* Give pages used by the bp back to the VM system (where possible) */ 1855static void 1856vfs_vmio_release(struct buf *bp) 1857{ 1858 vm_object_t obj; 1859 vm_page_t m; 1860 int i; 1861 1862 if ((bp->b_flags & B_UNMAPPED) == 0) { 1863 BUF_CHECK_MAPPED(bp); 1864 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 1865 } else 1866 BUF_CHECK_UNMAPPED(bp); 1867 obj = bp->b_bufobj->bo_object; 1868 if (obj != NULL) 1869 VM_OBJECT_WLOCK(obj); 1870 for (i = 0; i < bp->b_npages; i++) { 1871 m = bp->b_pages[i]; 1872 bp->b_pages[i] = NULL; 1873 /* 1874 * In order to keep page LRU ordering consistent, put 1875 * everything on the inactive queue. 1876 */ 1877 vm_page_lock(m); 1878 vm_page_unwire(m, 0); 1879 1880 /* 1881 * Might as well free the page if we can and it has 1882 * no valid data. We also free the page if the 1883 * buffer was used for direct I/O 1884 */ 1885 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) { 1886 if (m->wire_count == 0 && !vm_page_busied(m)) 1887 vm_page_free(m); 1888 } else if (bp->b_flags & B_DIRECT) 1889 vm_page_try_to_free(m); 1890 else if (buf_vm_page_count_severe()) 1891 vm_page_try_to_cache(m); 1892 vm_page_unlock(m); 1893 } 1894 if (obj != NULL) 1895 VM_OBJECT_WUNLOCK(obj); 1896 1897 if (bp->b_bufsize) { 1898 bufspacewakeup(); 1899 bp->b_bufsize = 0; 1900 } 1901 bp->b_npages = 0; 1902 bp->b_flags &= ~B_VMIO; 1903 if (bp->b_vp) 1904 brelvp(bp); 1905} 1906 1907/* 1908 * Check to see if a block at a particular lbn is available for a clustered 1909 * write. 1910 */ 1911static int 1912vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1913{ 1914 struct buf *bpa; 1915 int match; 1916 1917 match = 0; 1918 1919 /* If the buf isn't in core skip it */ 1920 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1921 return (0); 1922 1923 /* If the buf is busy we don't want to wait for it */ 1924 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1925 return (0); 1926 1927 /* Only cluster with valid clusterable delayed write buffers */ 1928 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1929 (B_DELWRI | B_CLUSTEROK)) 1930 goto done; 1931 1932 if (bpa->b_bufsize != size) 1933 goto done; 1934 1935 /* 1936 * Check to see if it is in the expected place on disk and that the 1937 * block has been mapped. 1938 */ 1939 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1940 match = 1; 1941done: 1942 BUF_UNLOCK(bpa); 1943 return (match); 1944} 1945 1946/* 1947 * vfs_bio_awrite: 1948 * 1949 * Implement clustered async writes for clearing out B_DELWRI buffers. 1950 * This is much better then the old way of writing only one buffer at 1951 * a time. Note that we may not be presented with the buffers in the 1952 * correct order, so we search for the cluster in both directions. 1953 */ 1954int 1955vfs_bio_awrite(struct buf *bp) 1956{ 1957 struct bufobj *bo; 1958 int i; 1959 int j; 1960 daddr_t lblkno = bp->b_lblkno; 1961 struct vnode *vp = bp->b_vp; 1962 int ncl; 1963 int nwritten; 1964 int size; 1965 int maxcl; 1966 int gbflags; 1967 1968 bo = &vp->v_bufobj; 1969 gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0; 1970 /* 1971 * right now we support clustered writing only to regular files. If 1972 * we find a clusterable block we could be in the middle of a cluster 1973 * rather then at the beginning. 1974 */ 1975 if ((vp->v_type == VREG) && 1976 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1977 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1978 1979 size = vp->v_mount->mnt_stat.f_iosize; 1980 maxcl = MAXPHYS / size; 1981 1982 BO_RLOCK(bo); 1983 for (i = 1; i < maxcl; i++) 1984 if (vfs_bio_clcheck(vp, size, lblkno + i, 1985 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1986 break; 1987 1988 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1989 if (vfs_bio_clcheck(vp, size, lblkno - j, 1990 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1991 break; 1992 BO_RUNLOCK(bo); 1993 --j; 1994 ncl = i + j; 1995 /* 1996 * this is a possible cluster write 1997 */ 1998 if (ncl != 1) { 1999 BUF_UNLOCK(bp); 2000 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 2001 gbflags); 2002 return (nwritten); 2003 } 2004 } 2005 bremfree(bp); 2006 bp->b_flags |= B_ASYNC; 2007 /* 2008 * default (old) behavior, writing out only one block 2009 * 2010 * XXX returns b_bufsize instead of b_bcount for nwritten? 2011 */ 2012 nwritten = bp->b_bufsize; 2013 (void) bwrite(bp); 2014 2015 return (nwritten); 2016} 2017 2018static void 2019setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags) 2020{ 2021 2022 KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 2023 bp->b_kvasize == 0, ("call bfreekva(%p)", bp)); 2024 if ((gbflags & GB_UNMAPPED) == 0) { 2025 bp->b_kvabase = (caddr_t)addr; 2026 } else if ((gbflags & GB_KVAALLOC) != 0) { 2027 KASSERT((gbflags & GB_UNMAPPED) != 0, 2028 ("GB_KVAALLOC without GB_UNMAPPED")); 2029 bp->b_kvaalloc = (caddr_t)addr; 2030 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 2031 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 2032 } 2033 bp->b_kvasize = maxsize; 2034} 2035 2036/* 2037 * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if 2038 * needed. 2039 */ 2040static int 2041allocbufkva(struct buf *bp, int maxsize, int gbflags) 2042{ 2043 vm_offset_t addr; 2044 2045 bfreekva(bp); 2046 addr = 0; 2047 2048 if (vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr)) { 2049 /* 2050 * Buffer map is too fragmented. Request the caller 2051 * to defragment the map. 2052 */ 2053 atomic_add_int(&bufdefragcnt, 1); 2054 return (1); 2055 } 2056 setbufkva(bp, addr, maxsize, gbflags); 2057 atomic_add_long(&bufspace, bp->b_kvasize); 2058 return (0); 2059} 2060 2061/* 2062 * Ask the bufdaemon for help, or act as bufdaemon itself, when a 2063 * locked vnode is supplied. 2064 */ 2065static void 2066getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo, 2067 int defrag) 2068{ 2069 struct thread *td; 2070 char *waitmsg; 2071 int error, fl, flags, norunbuf; 2072 2073 mtx_assert(&bqclean, MA_OWNED); 2074 2075 if (defrag) { 2076 flags = VFS_BIO_NEED_BUFSPACE; 2077 waitmsg = "nbufkv"; 2078 } else if (bufspace >= hibufspace) { 2079 waitmsg = "nbufbs"; 2080 flags = VFS_BIO_NEED_BUFSPACE; 2081 } else { 2082 waitmsg = "newbuf"; 2083 flags = VFS_BIO_NEED_ANY; 2084 } 2085 atomic_set_int(&needsbuffer, flags); 2086 mtx_unlock(&bqclean); 2087 2088 bd_speedup(); /* heeeelp */ 2089 if ((gbflags & GB_NOWAIT_BD) != 0) 2090 return; 2091 2092 td = curthread; 2093 rw_wlock(&nblock); 2094 while ((needsbuffer & flags) != 0) { 2095 if (vp != NULL && vp->v_type != VCHR && 2096 (td->td_pflags & TDP_BUFNEED) == 0) { 2097 rw_wunlock(&nblock); 2098 /* 2099 * getblk() is called with a vnode locked, and 2100 * some majority of the dirty buffers may as 2101 * well belong to the vnode. Flushing the 2102 * buffers there would make a progress that 2103 * cannot be achieved by the buf_daemon, that 2104 * cannot lock the vnode. 2105 */ 2106 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2107 (td->td_pflags & TDP_NORUNNINGBUF); 2108 2109 /* 2110 * Play bufdaemon. The getnewbuf() function 2111 * may be called while the thread owns lock 2112 * for another dirty buffer for the same 2113 * vnode, which makes it impossible to use 2114 * VOP_FSYNC() there, due to the buffer lock 2115 * recursion. 2116 */ 2117 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2118 fl = buf_flush(vp, flushbufqtarget); 2119 td->td_pflags &= norunbuf; 2120 rw_wlock(&nblock); 2121 if (fl != 0) 2122 continue; 2123 if ((needsbuffer & flags) == 0) 2124 break; 2125 } 2126 error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock, 2127 (PRIBIO + 4) | slpflag, waitmsg, slptimeo); 2128 if (error != 0) 2129 break; 2130 } 2131 rw_wunlock(&nblock); 2132} 2133 2134static void 2135getnewbuf_reuse_bp(struct buf *bp, int qindex) 2136{ 2137 2138 CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 2139 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 2140 bp->b_kvasize, bp->b_bufsize, qindex); 2141 mtx_assert(&bqclean, MA_NOTOWNED); 2142 2143 /* 2144 * Note: we no longer distinguish between VMIO and non-VMIO 2145 * buffers. 2146 */ 2147 KASSERT((bp->b_flags & B_DELWRI) == 0, 2148 ("delwri buffer %p found in queue %d", bp, qindex)); 2149 2150 if (qindex == QUEUE_CLEAN) { 2151 if (bp->b_flags & B_VMIO) { 2152 bp->b_flags &= ~B_ASYNC; 2153 vfs_vmio_release(bp); 2154 } 2155 if (bp->b_vp != NULL) 2156 brelvp(bp); 2157 } 2158 2159 /* 2160 * Get the rest of the buffer freed up. b_kva* is still valid 2161 * after this operation. 2162 */ 2163 2164 if (bp->b_rcred != NOCRED) { 2165 crfree(bp->b_rcred); 2166 bp->b_rcred = NOCRED; 2167 } 2168 if (bp->b_wcred != NOCRED) { 2169 crfree(bp->b_wcred); 2170 bp->b_wcred = NOCRED; 2171 } 2172 if (!LIST_EMPTY(&bp->b_dep)) 2173 buf_deallocate(bp); 2174 if (bp->b_vflags & BV_BKGRDINPROG) 2175 panic("losing buffer 3"); 2176 KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p. qindex: %d", 2177 bp, bp->b_vp, qindex)); 2178 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 2179 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 2180 2181 if (bp->b_bufsize) 2182 allocbuf(bp, 0); 2183 2184 bp->b_flags &= B_UNMAPPED | B_KVAALLOC; 2185 bp->b_ioflags = 0; 2186 bp->b_xflags = 0; 2187 KASSERT((bp->b_flags & B_INFREECNT) == 0, 2188 ("buf %p still counted as free?", bp)); 2189 bp->b_vflags = 0; 2190 bp->b_vp = NULL; 2191 bp->b_blkno = bp->b_lblkno = 0; 2192 bp->b_offset = NOOFFSET; 2193 bp->b_iodone = 0; 2194 bp->b_error = 0; 2195 bp->b_resid = 0; 2196 bp->b_bcount = 0; 2197 bp->b_npages = 0; 2198 bp->b_dirtyoff = bp->b_dirtyend = 0; 2199 bp->b_bufobj = NULL; 2200 bp->b_pin_count = 0; 2201 bp->b_fsprivate1 = NULL; 2202 bp->b_fsprivate2 = NULL; 2203 bp->b_fsprivate3 = NULL; 2204 2205 LIST_INIT(&bp->b_dep); 2206} 2207 2208static int flushingbufs; 2209 2210static struct buf * 2211getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata) 2212{ 2213 struct buf *bp, *nbp; 2214 int nqindex, qindex, pass; 2215 2216 KASSERT(!unmapped || !defrag, ("both unmapped and defrag")); 2217 2218 pass = 1; 2219restart: 2220 atomic_add_int(&getnewbufrestarts, 1); 2221 2222 /* 2223 * Setup for scan. If we do not have enough free buffers, 2224 * we setup a degenerate case that immediately fails. Note 2225 * that if we are specially marked process, we are allowed to 2226 * dip into our reserves. 2227 * 2228 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 2229 * for the allocation of the mapped buffer. For unmapped, the 2230 * easiest is to start with EMPTY outright. 2231 * 2232 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 2233 * However, there are a number of cases (defragging, reusing, ...) 2234 * where we cannot backup. 2235 */ 2236 nbp = NULL; 2237 mtx_lock(&bqclean); 2238 if (!defrag && unmapped) { 2239 nqindex = QUEUE_EMPTY; 2240 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2241 } 2242 if (nbp == NULL) { 2243 nqindex = QUEUE_EMPTYKVA; 2244 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2245 } 2246 2247 /* 2248 * If no EMPTYKVA buffers and we are either defragging or 2249 * reusing, locate a CLEAN buffer to free or reuse. If 2250 * bufspace useage is low skip this step so we can allocate a 2251 * new buffer. 2252 */ 2253 if (nbp == NULL && (defrag || bufspace >= lobufspace)) { 2254 nqindex = QUEUE_CLEAN; 2255 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2256 } 2257 2258 /* 2259 * If we could not find or were not allowed to reuse a CLEAN 2260 * buffer, check to see if it is ok to use an EMPTY buffer. 2261 * We can only use an EMPTY buffer if allocating its KVA would 2262 * not otherwise run us out of buffer space. No KVA is needed 2263 * for the unmapped allocation. 2264 */ 2265 if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace || 2266 metadata)) { 2267 nqindex = QUEUE_EMPTY; 2268 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2269 } 2270 2271 /* 2272 * All available buffers might be clean, retry ignoring the 2273 * lobufspace as the last resort. 2274 */ 2275 if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) { 2276 nqindex = QUEUE_CLEAN; 2277 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2278 } 2279 2280 /* 2281 * Run scan, possibly freeing data and/or kva mappings on the fly 2282 * depending. 2283 */ 2284 while ((bp = nbp) != NULL) { 2285 qindex = nqindex; 2286 2287 /* 2288 * Calculate next bp (we can only use it if we do not 2289 * block or do other fancy things). 2290 */ 2291 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 2292 switch (qindex) { 2293 case QUEUE_EMPTY: 2294 nqindex = QUEUE_EMPTYKVA; 2295 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2296 if (nbp != NULL) 2297 break; 2298 /* FALLTHROUGH */ 2299 case QUEUE_EMPTYKVA: 2300 nqindex = QUEUE_CLEAN; 2301 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2302 if (nbp != NULL) 2303 break; 2304 /* FALLTHROUGH */ 2305 case QUEUE_CLEAN: 2306 if (metadata && pass == 1) { 2307 pass = 2; 2308 nqindex = QUEUE_EMPTY; 2309 nbp = TAILQ_FIRST( 2310 &bufqueues[QUEUE_EMPTY]); 2311 } 2312 /* 2313 * nbp is NULL. 2314 */ 2315 break; 2316 } 2317 } 2318 /* 2319 * If we are defragging then we need a buffer with 2320 * b_kvasize != 0. XXX this situation should no longer 2321 * occur, if defrag is non-zero the buffer's b_kvasize 2322 * should also be non-zero at this point. XXX 2323 */ 2324 if (defrag && bp->b_kvasize == 0) { 2325 printf("Warning: defrag empty buffer %p\n", bp); 2326 continue; 2327 } 2328 2329 /* 2330 * Start freeing the bp. This is somewhat involved. nbp 2331 * remains valid only for QUEUE_EMPTY[KVA] bp's. 2332 */ 2333 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2334 continue; 2335 /* 2336 * BKGRDINPROG can only be set with the buf and bufobj 2337 * locks both held. We tolerate a race to clear it here. 2338 */ 2339 if (bp->b_vflags & BV_BKGRDINPROG) { 2340 BUF_UNLOCK(bp); 2341 continue; 2342 } 2343 2344 KASSERT(bp->b_qindex == qindex, 2345 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2346 2347 bremfreel(bp); 2348 mtx_unlock(&bqclean); 2349 /* 2350 * NOTE: nbp is now entirely invalid. We can only restart 2351 * the scan from this point on. 2352 */ 2353 2354 getnewbuf_reuse_bp(bp, qindex); 2355 mtx_assert(&bqclean, MA_NOTOWNED); 2356 2357 /* 2358 * If we are defragging then free the buffer. 2359 */ 2360 if (defrag) { 2361 bp->b_flags |= B_INVAL; 2362 bfreekva(bp); 2363 brelse(bp); 2364 defrag = 0; 2365 goto restart; 2366 } 2367 2368 /* 2369 * Notify any waiters for the buffer lock about 2370 * identity change by freeing the buffer. 2371 */ 2372 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 2373 bp->b_flags |= B_INVAL; 2374 bfreekva(bp); 2375 brelse(bp); 2376 goto restart; 2377 } 2378 2379 if (metadata) 2380 break; 2381 2382 /* 2383 * If we are overcomitted then recover the buffer and its 2384 * KVM space. This occurs in rare situations when multiple 2385 * processes are blocked in getnewbuf() or allocbuf(). 2386 */ 2387 if (bufspace >= hibufspace) 2388 flushingbufs = 1; 2389 if (flushingbufs && bp->b_kvasize != 0) { 2390 bp->b_flags |= B_INVAL; 2391 bfreekva(bp); 2392 brelse(bp); 2393 goto restart; 2394 } 2395 if (bufspace < lobufspace) 2396 flushingbufs = 0; 2397 break; 2398 } 2399 return (bp); 2400} 2401 2402/* 2403 * getnewbuf: 2404 * 2405 * Find and initialize a new buffer header, freeing up existing buffers 2406 * in the bufqueues as necessary. The new buffer is returned locked. 2407 * 2408 * Important: B_INVAL is not set. If the caller wishes to throw the 2409 * buffer away, the caller must set B_INVAL prior to calling brelse(). 2410 * 2411 * We block if: 2412 * We have insufficient buffer headers 2413 * We have insufficient buffer space 2414 * buffer_arena is too fragmented ( space reservation fails ) 2415 * If we have to flush dirty buffers ( but we try to avoid this ) 2416 */ 2417static struct buf * 2418getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 2419 int gbflags) 2420{ 2421 struct buf *bp; 2422 int defrag, metadata; 2423 2424 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 2425 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 2426 if (!unmapped_buf_allowed) 2427 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 2428 2429 defrag = 0; 2430 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 2431 vp->v_type == VCHR) 2432 metadata = 1; 2433 else 2434 metadata = 0; 2435 /* 2436 * We can't afford to block since we might be holding a vnode lock, 2437 * which may prevent system daemons from running. We deal with 2438 * low-memory situations by proactively returning memory and running 2439 * async I/O rather then sync I/O. 2440 */ 2441 atomic_add_int(&getnewbufcalls, 1); 2442 atomic_subtract_int(&getnewbufrestarts, 1); 2443restart: 2444 bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED | 2445 GB_KVAALLOC)) == GB_UNMAPPED, metadata); 2446 if (bp != NULL) 2447 defrag = 0; 2448 2449 /* 2450 * If we exhausted our list, sleep as appropriate. We may have to 2451 * wakeup various daemons and write out some dirty buffers. 2452 * 2453 * Generally we are sleeping due to insufficient buffer space. 2454 */ 2455 if (bp == NULL) { 2456 mtx_assert(&bqclean, MA_OWNED); 2457 getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag); 2458 mtx_assert(&bqclean, MA_NOTOWNED); 2459 } else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) { 2460 mtx_assert(&bqclean, MA_NOTOWNED); 2461 2462 bfreekva(bp); 2463 bp->b_flags |= B_UNMAPPED; 2464 bp->b_kvabase = bp->b_data = unmapped_buf; 2465 bp->b_kvasize = maxsize; 2466 atomic_add_long(&bufspace, bp->b_kvasize); 2467 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 2468 atomic_add_int(&bufreusecnt, 1); 2469 } else { 2470 mtx_assert(&bqclean, MA_NOTOWNED); 2471 2472 /* 2473 * We finally have a valid bp. We aren't quite out of the 2474 * woods, we still have to reserve kva space. In order 2475 * to keep fragmentation sane we only allocate kva in 2476 * BKVASIZE chunks. 2477 */ 2478 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2479 2480 if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED | 2481 B_KVAALLOC)) == B_UNMAPPED) { 2482 if (allocbufkva(bp, maxsize, gbflags)) { 2483 defrag = 1; 2484 bp->b_flags |= B_INVAL; 2485 brelse(bp); 2486 goto restart; 2487 } 2488 atomic_add_int(&bufreusecnt, 1); 2489 } else if ((bp->b_flags & B_KVAALLOC) != 0 && 2490 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) { 2491 /* 2492 * If the reused buffer has KVA allocated, 2493 * reassign b_kvaalloc to b_kvabase. 2494 */ 2495 bp->b_kvabase = bp->b_kvaalloc; 2496 bp->b_flags &= ~B_KVAALLOC; 2497 atomic_subtract_long(&unmapped_bufspace, 2498 bp->b_kvasize); 2499 atomic_add_int(&bufreusecnt, 1); 2500 } else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 2501 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED | 2502 GB_KVAALLOC)) { 2503 /* 2504 * The case of reused buffer already have KVA 2505 * mapped, but the request is for unmapped 2506 * buffer with KVA allocated. 2507 */ 2508 bp->b_kvaalloc = bp->b_kvabase; 2509 bp->b_data = bp->b_kvabase = unmapped_buf; 2510 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 2511 atomic_add_long(&unmapped_bufspace, 2512 bp->b_kvasize); 2513 atomic_add_int(&bufreusecnt, 1); 2514 } 2515 if ((gbflags & GB_UNMAPPED) == 0) { 2516 bp->b_saveaddr = bp->b_kvabase; 2517 bp->b_data = bp->b_saveaddr; 2518 bp->b_flags &= ~B_UNMAPPED; 2519 BUF_CHECK_MAPPED(bp); 2520 } 2521 } 2522 return (bp); 2523} 2524 2525/* 2526 * buf_daemon: 2527 * 2528 * buffer flushing daemon. Buffers are normally flushed by the 2529 * update daemon but if it cannot keep up this process starts to 2530 * take the load in an attempt to prevent getnewbuf() from blocking. 2531 */ 2532 2533static struct kproc_desc buf_kp = { 2534 "bufdaemon", 2535 buf_daemon, 2536 &bufdaemonproc 2537}; 2538SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2539 2540static int 2541buf_flush(struct vnode *vp, int target) 2542{ 2543 int flushed; 2544 2545 flushed = flushbufqueues(vp, target, 0); 2546 if (flushed == 0) { 2547 /* 2548 * Could not find any buffers without rollback 2549 * dependencies, so just write the first one 2550 * in the hopes of eventually making progress. 2551 */ 2552 if (vp != NULL && target > 2) 2553 target /= 2; 2554 flushbufqueues(vp, target, 1); 2555 } 2556 return (flushed); 2557} 2558 2559static void 2560buf_daemon() 2561{ 2562 int lodirty; 2563 2564 /* 2565 * This process needs to be suspended prior to shutdown sync. 2566 */ 2567 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2568 SHUTDOWN_PRI_LAST); 2569 2570 /* 2571 * This process is allowed to take the buffer cache to the limit 2572 */ 2573 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2574 mtx_lock(&bdlock); 2575 for (;;) { 2576 bd_request = 0; 2577 mtx_unlock(&bdlock); 2578 2579 kproc_suspend_check(bufdaemonproc); 2580 lodirty = lodirtybuffers; 2581 if (bd_speedupreq) { 2582 lodirty = numdirtybuffers / 2; 2583 bd_speedupreq = 0; 2584 } 2585 /* 2586 * Do the flush. Limit the amount of in-transit I/O we 2587 * allow to build up, otherwise we would completely saturate 2588 * the I/O system. 2589 */ 2590 while (numdirtybuffers > lodirty) { 2591 if (buf_flush(NULL, numdirtybuffers - lodirty) == 0) 2592 break; 2593 kern_yield(PRI_USER); 2594 } 2595 2596 /* 2597 * Only clear bd_request if we have reached our low water 2598 * mark. The buf_daemon normally waits 1 second and 2599 * then incrementally flushes any dirty buffers that have 2600 * built up, within reason. 2601 * 2602 * If we were unable to hit our low water mark and couldn't 2603 * find any flushable buffers, we sleep for a short period 2604 * to avoid endless loops on unlockable buffers. 2605 */ 2606 mtx_lock(&bdlock); 2607 if (numdirtybuffers <= lodirtybuffers) { 2608 /* 2609 * We reached our low water mark, reset the 2610 * request and sleep until we are needed again. 2611 * The sleep is just so the suspend code works. 2612 */ 2613 bd_request = 0; 2614 /* 2615 * Do an extra wakeup in case dirty threshold 2616 * changed via sysctl and the explicit transition 2617 * out of shortfall was missed. 2618 */ 2619 bdirtywakeup(); 2620 if (runningbufspace <= lorunningspace) 2621 runningwakeup(); 2622 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2623 } else { 2624 /* 2625 * We couldn't find any flushable dirty buffers but 2626 * still have too many dirty buffers, we 2627 * have to sleep and try again. (rare) 2628 */ 2629 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2630 } 2631 } 2632} 2633 2634/* 2635 * flushbufqueues: 2636 * 2637 * Try to flush a buffer in the dirty queue. We must be careful to 2638 * free up B_INVAL buffers instead of write them, which NFS is 2639 * particularly sensitive to. 2640 */ 2641static int flushwithdeps = 0; 2642SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2643 0, "Number of buffers flushed with dependecies that require rollbacks"); 2644 2645static int 2646flushbufqueues(struct vnode *lvp, int target, int flushdeps) 2647{ 2648 struct buf *sentinel; 2649 struct vnode *vp; 2650 struct mount *mp; 2651 struct buf *bp; 2652 int hasdeps; 2653 int flushed; 2654 int queue; 2655 int error; 2656 bool unlock; 2657 2658 flushed = 0; 2659 queue = QUEUE_DIRTY; 2660 bp = NULL; 2661 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2662 sentinel->b_qindex = QUEUE_SENTINEL; 2663 mtx_lock(&bqdirty); 2664 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2665 mtx_unlock(&bqdirty); 2666 while (flushed != target) { 2667 maybe_yield(); 2668 mtx_lock(&bqdirty); 2669 bp = TAILQ_NEXT(sentinel, b_freelist); 2670 if (bp != NULL) { 2671 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2672 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2673 b_freelist); 2674 } else { 2675 mtx_unlock(&bqdirty); 2676 break; 2677 } 2678 /* 2679 * Skip sentinels inserted by other invocations of the 2680 * flushbufqueues(), taking care to not reorder them. 2681 * 2682 * Only flush the buffers that belong to the 2683 * vnode locked by the curthread. 2684 */ 2685 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL && 2686 bp->b_vp != lvp)) { 2687 mtx_unlock(&bqdirty); 2688 continue; 2689 } 2690 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL); 2691 mtx_unlock(&bqdirty); 2692 if (error != 0) 2693 continue; 2694 if (bp->b_pin_count > 0) { 2695 BUF_UNLOCK(bp); 2696 continue; 2697 } 2698 /* 2699 * BKGRDINPROG can only be set with the buf and bufobj 2700 * locks both held. We tolerate a race to clear it here. 2701 */ 2702 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2703 (bp->b_flags & B_DELWRI) == 0) { 2704 BUF_UNLOCK(bp); 2705 continue; 2706 } 2707 if (bp->b_flags & B_INVAL) { 2708 bremfreef(bp); 2709 brelse(bp); 2710 flushed++; 2711 continue; 2712 } 2713 2714 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2715 if (flushdeps == 0) { 2716 BUF_UNLOCK(bp); 2717 continue; 2718 } 2719 hasdeps = 1; 2720 } else 2721 hasdeps = 0; 2722 /* 2723 * We must hold the lock on a vnode before writing 2724 * one of its buffers. Otherwise we may confuse, or 2725 * in the case of a snapshot vnode, deadlock the 2726 * system. 2727 * 2728 * The lock order here is the reverse of the normal 2729 * of vnode followed by buf lock. This is ok because 2730 * the NOWAIT will prevent deadlock. 2731 */ 2732 vp = bp->b_vp; 2733 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2734 BUF_UNLOCK(bp); 2735 continue; 2736 } 2737 if (lvp == NULL) { 2738 unlock = true; 2739 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); 2740 } else { 2741 ASSERT_VOP_LOCKED(vp, "getbuf"); 2742 unlock = false; 2743 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 : 2744 vn_lock(vp, LK_TRYUPGRADE); 2745 } 2746 if (error == 0) { 2747 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2748 bp, bp->b_vp, bp->b_flags); 2749 if (curproc == bufdaemonproc) { 2750 vfs_bio_awrite(bp); 2751 } else { 2752 bremfree(bp); 2753 bwrite(bp); 2754 notbufdflushes++; 2755 } 2756 vn_finished_write(mp); 2757 if (unlock) 2758 VOP_UNLOCK(vp, 0); 2759 flushwithdeps += hasdeps; 2760 flushed++; 2761 2762 /* 2763 * Sleeping on runningbufspace while holding 2764 * vnode lock leads to deadlock. 2765 */ 2766 if (curproc == bufdaemonproc && 2767 runningbufspace > hirunningspace) 2768 waitrunningbufspace(); 2769 continue; 2770 } 2771 vn_finished_write(mp); 2772 BUF_UNLOCK(bp); 2773 } 2774 mtx_lock(&bqdirty); 2775 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2776 mtx_unlock(&bqdirty); 2777 free(sentinel, M_TEMP); 2778 return (flushed); 2779} 2780 2781/* 2782 * Check to see if a block is currently memory resident. 2783 */ 2784struct buf * 2785incore(struct bufobj *bo, daddr_t blkno) 2786{ 2787 struct buf *bp; 2788 2789 BO_RLOCK(bo); 2790 bp = gbincore(bo, blkno); 2791 BO_RUNLOCK(bo); 2792 return (bp); 2793} 2794 2795/* 2796 * Returns true if no I/O is needed to access the 2797 * associated VM object. This is like incore except 2798 * it also hunts around in the VM system for the data. 2799 */ 2800 2801static int 2802inmem(struct vnode * vp, daddr_t blkno) 2803{ 2804 vm_object_t obj; 2805 vm_offset_t toff, tinc, size; 2806 vm_page_t m; 2807 vm_ooffset_t off; 2808 2809 ASSERT_VOP_LOCKED(vp, "inmem"); 2810 2811 if (incore(&vp->v_bufobj, blkno)) 2812 return 1; 2813 if (vp->v_mount == NULL) 2814 return 0; 2815 obj = vp->v_object; 2816 if (obj == NULL) 2817 return (0); 2818 2819 size = PAGE_SIZE; 2820 if (size > vp->v_mount->mnt_stat.f_iosize) 2821 size = vp->v_mount->mnt_stat.f_iosize; 2822 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2823 2824 VM_OBJECT_RLOCK(obj); 2825 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2826 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2827 if (!m) 2828 goto notinmem; 2829 tinc = size; 2830 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2831 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2832 if (vm_page_is_valid(m, 2833 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2834 goto notinmem; 2835 } 2836 VM_OBJECT_RUNLOCK(obj); 2837 return 1; 2838 2839notinmem: 2840 VM_OBJECT_RUNLOCK(obj); 2841 return (0); 2842} 2843 2844/* 2845 * Set the dirty range for a buffer based on the status of the dirty 2846 * bits in the pages comprising the buffer. The range is limited 2847 * to the size of the buffer. 2848 * 2849 * Tell the VM system that the pages associated with this buffer 2850 * are clean. This is used for delayed writes where the data is 2851 * going to go to disk eventually without additional VM intevention. 2852 * 2853 * Note that while we only really need to clean through to b_bcount, we 2854 * just go ahead and clean through to b_bufsize. 2855 */ 2856static void 2857vfs_clean_pages_dirty_buf(struct buf *bp) 2858{ 2859 vm_ooffset_t foff, noff, eoff; 2860 vm_page_t m; 2861 int i; 2862 2863 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 2864 return; 2865 2866 foff = bp->b_offset; 2867 KASSERT(bp->b_offset != NOOFFSET, 2868 ("vfs_clean_pages_dirty_buf: no buffer offset")); 2869 2870 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 2871 vfs_drain_busy_pages(bp); 2872 vfs_setdirty_locked_object(bp); 2873 for (i = 0; i < bp->b_npages; i++) { 2874 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2875 eoff = noff; 2876 if (eoff > bp->b_offset + bp->b_bufsize) 2877 eoff = bp->b_offset + bp->b_bufsize; 2878 m = bp->b_pages[i]; 2879 vfs_page_set_validclean(bp, foff, m); 2880 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2881 foff = noff; 2882 } 2883 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 2884} 2885 2886static void 2887vfs_setdirty_locked_object(struct buf *bp) 2888{ 2889 vm_object_t object; 2890 int i; 2891 2892 object = bp->b_bufobj->bo_object; 2893 VM_OBJECT_ASSERT_WLOCKED(object); 2894 2895 /* 2896 * We qualify the scan for modified pages on whether the 2897 * object has been flushed yet. 2898 */ 2899 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { 2900 vm_offset_t boffset; 2901 vm_offset_t eoffset; 2902 2903 /* 2904 * test the pages to see if they have been modified directly 2905 * by users through the VM system. 2906 */ 2907 for (i = 0; i < bp->b_npages; i++) 2908 vm_page_test_dirty(bp->b_pages[i]); 2909 2910 /* 2911 * Calculate the encompassing dirty range, boffset and eoffset, 2912 * (eoffset - boffset) bytes. 2913 */ 2914 2915 for (i = 0; i < bp->b_npages; i++) { 2916 if (bp->b_pages[i]->dirty) 2917 break; 2918 } 2919 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2920 2921 for (i = bp->b_npages - 1; i >= 0; --i) { 2922 if (bp->b_pages[i]->dirty) { 2923 break; 2924 } 2925 } 2926 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2927 2928 /* 2929 * Fit it to the buffer. 2930 */ 2931 2932 if (eoffset > bp->b_bcount) 2933 eoffset = bp->b_bcount; 2934 2935 /* 2936 * If we have a good dirty range, merge with the existing 2937 * dirty range. 2938 */ 2939 2940 if (boffset < eoffset) { 2941 if (bp->b_dirtyoff > boffset) 2942 bp->b_dirtyoff = boffset; 2943 if (bp->b_dirtyend < eoffset) 2944 bp->b_dirtyend = eoffset; 2945 } 2946 } 2947} 2948 2949/* 2950 * Allocate the KVA mapping for an existing buffer. It handles the 2951 * cases of both B_UNMAPPED buffer, and buffer with the preallocated 2952 * KVA which is not mapped (B_KVAALLOC). 2953 */ 2954static void 2955bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 2956{ 2957 struct buf *scratch_bp; 2958 int bsize, maxsize, need_mapping, need_kva; 2959 off_t offset; 2960 2961 need_mapping = (bp->b_flags & B_UNMAPPED) != 0 && 2962 (gbflags & GB_UNMAPPED) == 0; 2963 need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED && 2964 (gbflags & GB_KVAALLOC) != 0; 2965 if (!need_mapping && !need_kva) 2966 return; 2967 2968 BUF_CHECK_UNMAPPED(bp); 2969 2970 if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) { 2971 /* 2972 * Buffer is not mapped, but the KVA was already 2973 * reserved at the time of the instantiation. Use the 2974 * allocated space. 2975 */ 2976 bp->b_flags &= ~B_KVAALLOC; 2977 KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0")); 2978 bp->b_kvabase = bp->b_kvaalloc; 2979 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize); 2980 goto has_addr; 2981 } 2982 2983 /* 2984 * Calculate the amount of the address space we would reserve 2985 * if the buffer was mapped. 2986 */ 2987 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 2988 offset = blkno * bsize; 2989 maxsize = size + (offset & PAGE_MASK); 2990 maxsize = imax(maxsize, bsize); 2991 2992mapping_loop: 2993 if (allocbufkva(bp, maxsize, gbflags)) { 2994 /* 2995 * Request defragmentation. getnewbuf() returns us the 2996 * allocated space by the scratch buffer KVA. 2997 */ 2998 scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags | 2999 (GB_UNMAPPED | GB_KVAALLOC)); 3000 if (scratch_bp == NULL) { 3001 if ((gbflags & GB_NOWAIT_BD) != 0) { 3002 /* 3003 * XXXKIB: defragmentation cannot 3004 * succeed, not sure what else to do. 3005 */ 3006 panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp); 3007 } 3008 atomic_add_int(&mappingrestarts, 1); 3009 goto mapping_loop; 3010 } 3011 KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0, 3012 ("scratch bp !B_KVAALLOC %p", scratch_bp)); 3013 setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc, 3014 scratch_bp->b_kvasize, gbflags); 3015 3016 /* Get rid of the scratch buffer. */ 3017 scratch_bp->b_kvasize = 0; 3018 scratch_bp->b_flags |= B_INVAL; 3019 scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC); 3020 brelse(scratch_bp); 3021 } 3022 if (!need_mapping) 3023 return; 3024 3025has_addr: 3026 bp->b_saveaddr = bp->b_kvabase; 3027 bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */ 3028 bp->b_flags &= ~B_UNMAPPED; 3029 BUF_CHECK_MAPPED(bp); 3030 bpmap_qenter(bp); 3031} 3032 3033/* 3034 * getblk: 3035 * 3036 * Get a block given a specified block and offset into a file/device. 3037 * The buffers B_DONE bit will be cleared on return, making it almost 3038 * ready for an I/O initiation. B_INVAL may or may not be set on 3039 * return. The caller should clear B_INVAL prior to initiating a 3040 * READ. 3041 * 3042 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 3043 * an existing buffer. 3044 * 3045 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 3046 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 3047 * and then cleared based on the backing VM. If the previous buffer is 3048 * non-0-sized but invalid, B_CACHE will be cleared. 3049 * 3050 * If getblk() must create a new buffer, the new buffer is returned with 3051 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 3052 * case it is returned with B_INVAL clear and B_CACHE set based on the 3053 * backing VM. 3054 * 3055 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 3056 * B_CACHE bit is clear. 3057 * 3058 * What this means, basically, is that the caller should use B_CACHE to 3059 * determine whether the buffer is fully valid or not and should clear 3060 * B_INVAL prior to issuing a read. If the caller intends to validate 3061 * the buffer by loading its data area with something, the caller needs 3062 * to clear B_INVAL. If the caller does this without issuing an I/O, 3063 * the caller should set B_CACHE ( as an optimization ), else the caller 3064 * should issue the I/O and biodone() will set B_CACHE if the I/O was 3065 * a write attempt or if it was a successfull read. If the caller 3066 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 3067 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 3068 */ 3069struct buf * 3070getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 3071 int flags) 3072{ 3073 struct buf *bp; 3074 struct bufobj *bo; 3075 int bsize, error, maxsize, vmio; 3076 off_t offset; 3077 3078 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 3079 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3080 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3081 ASSERT_VOP_LOCKED(vp, "getblk"); 3082 if (size > MAXBCACHEBUF) 3083 panic("getblk: size(%d) > MAXBCACHEBUF(%d)\n", size, 3084 MAXBCACHEBUF); 3085 if (!unmapped_buf_allowed) 3086 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3087 3088 bo = &vp->v_bufobj; 3089loop: 3090 BO_RLOCK(bo); 3091 bp = gbincore(bo, blkno); 3092 if (bp != NULL) { 3093 int lockflags; 3094 /* 3095 * Buffer is in-core. If the buffer is not busy nor managed, 3096 * it must be on a queue. 3097 */ 3098 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 3099 3100 if (flags & GB_LOCK_NOWAIT) 3101 lockflags |= LK_NOWAIT; 3102 3103 error = BUF_TIMELOCK(bp, lockflags, 3104 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); 3105 3106 /* 3107 * If we slept and got the lock we have to restart in case 3108 * the buffer changed identities. 3109 */ 3110 if (error == ENOLCK) 3111 goto loop; 3112 /* We timed out or were interrupted. */ 3113 else if (error) 3114 return (NULL); 3115 /* If recursed, assume caller knows the rules. */ 3116 else if (BUF_LOCKRECURSED(bp)) 3117 goto end; 3118 3119 /* 3120 * The buffer is locked. B_CACHE is cleared if the buffer is 3121 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 3122 * and for a VMIO buffer B_CACHE is adjusted according to the 3123 * backing VM cache. 3124 */ 3125 if (bp->b_flags & B_INVAL) 3126 bp->b_flags &= ~B_CACHE; 3127 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 3128 bp->b_flags |= B_CACHE; 3129 if (bp->b_flags & B_MANAGED) 3130 MPASS(bp->b_qindex == QUEUE_NONE); 3131 else 3132 bremfree(bp); 3133 3134 /* 3135 * check for size inconsistencies for non-VMIO case. 3136 */ 3137 if (bp->b_bcount != size) { 3138 if ((bp->b_flags & B_VMIO) == 0 || 3139 (size > bp->b_kvasize)) { 3140 if (bp->b_flags & B_DELWRI) { 3141 /* 3142 * If buffer is pinned and caller does 3143 * not want sleep waiting for it to be 3144 * unpinned, bail out 3145 * */ 3146 if (bp->b_pin_count > 0) { 3147 if (flags & GB_LOCK_NOWAIT) { 3148 bqrelse(bp); 3149 return (NULL); 3150 } else { 3151 bunpin_wait(bp); 3152 } 3153 } 3154 bp->b_flags |= B_NOCACHE; 3155 bwrite(bp); 3156 } else { 3157 if (LIST_EMPTY(&bp->b_dep)) { 3158 bp->b_flags |= B_RELBUF; 3159 brelse(bp); 3160 } else { 3161 bp->b_flags |= B_NOCACHE; 3162 bwrite(bp); 3163 } 3164 } 3165 goto loop; 3166 } 3167 } 3168 3169 /* 3170 * Handle the case of unmapped buffer which should 3171 * become mapped, or the buffer for which KVA 3172 * reservation is requested. 3173 */ 3174 bp_unmapped_get_kva(bp, blkno, size, flags); 3175 3176 /* 3177 * If the size is inconsistant in the VMIO case, we can resize 3178 * the buffer. This might lead to B_CACHE getting set or 3179 * cleared. If the size has not changed, B_CACHE remains 3180 * unchanged from its previous state. 3181 */ 3182 if (bp->b_bcount != size) 3183 allocbuf(bp, size); 3184 3185 KASSERT(bp->b_offset != NOOFFSET, 3186 ("getblk: no buffer offset")); 3187 3188 /* 3189 * A buffer with B_DELWRI set and B_CACHE clear must 3190 * be committed before we can return the buffer in 3191 * order to prevent the caller from issuing a read 3192 * ( due to B_CACHE not being set ) and overwriting 3193 * it. 3194 * 3195 * Most callers, including NFS and FFS, need this to 3196 * operate properly either because they assume they 3197 * can issue a read if B_CACHE is not set, or because 3198 * ( for example ) an uncached B_DELWRI might loop due 3199 * to softupdates re-dirtying the buffer. In the latter 3200 * case, B_CACHE is set after the first write completes, 3201 * preventing further loops. 3202 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3203 * above while extending the buffer, we cannot allow the 3204 * buffer to remain with B_CACHE set after the write 3205 * completes or it will represent a corrupt state. To 3206 * deal with this we set B_NOCACHE to scrap the buffer 3207 * after the write. 3208 * 3209 * We might be able to do something fancy, like setting 3210 * B_CACHE in bwrite() except if B_DELWRI is already set, 3211 * so the below call doesn't set B_CACHE, but that gets real 3212 * confusing. This is much easier. 3213 */ 3214 3215 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3216 bp->b_flags |= B_NOCACHE; 3217 bwrite(bp); 3218 goto loop; 3219 } 3220 bp->b_flags &= ~B_DONE; 3221 } else { 3222 /* 3223 * Buffer is not in-core, create new buffer. The buffer 3224 * returned by getnewbuf() is locked. Note that the returned 3225 * buffer is also considered valid (not marked B_INVAL). 3226 */ 3227 BO_RUNLOCK(bo); 3228 /* 3229 * If the user does not want us to create the buffer, bail out 3230 * here. 3231 */ 3232 if (flags & GB_NOCREAT) 3233 return NULL; 3234 if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread)) 3235 return NULL; 3236 3237 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 3238 offset = blkno * bsize; 3239 vmio = vp->v_object != NULL; 3240 if (vmio) { 3241 maxsize = size + (offset & PAGE_MASK); 3242 } else { 3243 maxsize = size; 3244 /* Do not allow non-VMIO notmapped buffers. */ 3245 flags &= ~GB_UNMAPPED; 3246 } 3247 maxsize = imax(maxsize, bsize); 3248 3249 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 3250 if (bp == NULL) { 3251 if (slpflag || slptimeo) 3252 return NULL; 3253 goto loop; 3254 } 3255 3256 /* 3257 * This code is used to make sure that a buffer is not 3258 * created while the getnewbuf routine is blocked. 3259 * This can be a problem whether the vnode is locked or not. 3260 * If the buffer is created out from under us, we have to 3261 * throw away the one we just created. 3262 * 3263 * Note: this must occur before we associate the buffer 3264 * with the vp especially considering limitations in 3265 * the splay tree implementation when dealing with duplicate 3266 * lblkno's. 3267 */ 3268 BO_LOCK(bo); 3269 if (gbincore(bo, blkno)) { 3270 BO_UNLOCK(bo); 3271 bp->b_flags |= B_INVAL; 3272 brelse(bp); 3273 goto loop; 3274 } 3275 3276 /* 3277 * Insert the buffer into the hash, so that it can 3278 * be found by incore. 3279 */ 3280 bp->b_blkno = bp->b_lblkno = blkno; 3281 bp->b_offset = offset; 3282 bgetvp(vp, bp); 3283 BO_UNLOCK(bo); 3284 3285 /* 3286 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 3287 * buffer size starts out as 0, B_CACHE will be set by 3288 * allocbuf() for the VMIO case prior to it testing the 3289 * backing store for validity. 3290 */ 3291 3292 if (vmio) { 3293 bp->b_flags |= B_VMIO; 3294 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 3295 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 3296 bp, vp->v_object, bp->b_bufobj->bo_object)); 3297 } else { 3298 bp->b_flags &= ~B_VMIO; 3299 KASSERT(bp->b_bufobj->bo_object == NULL, 3300 ("ARGH! has b_bufobj->bo_object %p %p\n", 3301 bp, bp->b_bufobj->bo_object)); 3302 BUF_CHECK_MAPPED(bp); 3303 } 3304 3305 allocbuf(bp, size); 3306 bp->b_flags &= ~B_DONE; 3307 } 3308 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 3309 BUF_ASSERT_HELD(bp); 3310end: 3311 KASSERT(bp->b_bufobj == bo, 3312 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3313 return (bp); 3314} 3315 3316/* 3317 * Get an empty, disassociated buffer of given size. The buffer is initially 3318 * set to B_INVAL. 3319 */ 3320struct buf * 3321geteblk(int size, int flags) 3322{ 3323 struct buf *bp; 3324 int maxsize; 3325 3326 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3327 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 3328 if ((flags & GB_NOWAIT_BD) && 3329 (curthread->td_pflags & TDP_BUFNEED) != 0) 3330 return (NULL); 3331 } 3332 allocbuf(bp, size); 3333 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3334 BUF_ASSERT_HELD(bp); 3335 return (bp); 3336} 3337 3338 3339/* 3340 * This code constitutes the buffer memory from either anonymous system 3341 * memory (in the case of non-VMIO operations) or from an associated 3342 * VM object (in the case of VMIO operations). This code is able to 3343 * resize a buffer up or down. 3344 * 3345 * Note that this code is tricky, and has many complications to resolve 3346 * deadlock or inconsistant data situations. Tread lightly!!! 3347 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3348 * the caller. Calling this code willy nilly can result in the loss of data. 3349 * 3350 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3351 * B_CACHE for the non-VMIO case. 3352 */ 3353 3354int 3355allocbuf(struct buf *bp, int size) 3356{ 3357 int newbsize, mbsize; 3358 int i; 3359 3360 BUF_ASSERT_HELD(bp); 3361 3362 if (bp->b_kvasize < size) 3363 panic("allocbuf: buffer too small"); 3364 3365 if ((bp->b_flags & B_VMIO) == 0) { 3366 caddr_t origbuf; 3367 int origbufsize; 3368 /* 3369 * Just get anonymous memory from the kernel. Don't 3370 * mess with B_CACHE. 3371 */ 3372 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3373 if (bp->b_flags & B_MALLOC) 3374 newbsize = mbsize; 3375 else 3376 newbsize = round_page(size); 3377 3378 if (newbsize < bp->b_bufsize) { 3379 /* 3380 * malloced buffers are not shrunk 3381 */ 3382 if (bp->b_flags & B_MALLOC) { 3383 if (newbsize) { 3384 bp->b_bcount = size; 3385 } else { 3386 free(bp->b_data, M_BIOBUF); 3387 if (bp->b_bufsize) { 3388 atomic_subtract_long( 3389 &bufmallocspace, 3390 bp->b_bufsize); 3391 bufspacewakeup(); 3392 bp->b_bufsize = 0; 3393 } 3394 bp->b_saveaddr = bp->b_kvabase; 3395 bp->b_data = bp->b_saveaddr; 3396 bp->b_bcount = 0; 3397 bp->b_flags &= ~B_MALLOC; 3398 } 3399 return 1; 3400 } 3401 vm_hold_free_pages(bp, newbsize); 3402 } else if (newbsize > bp->b_bufsize) { 3403 /* 3404 * We only use malloced memory on the first allocation. 3405 * and revert to page-allocated memory when the buffer 3406 * grows. 3407 */ 3408 /* 3409 * There is a potential smp race here that could lead 3410 * to bufmallocspace slightly passing the max. It 3411 * is probably extremely rare and not worth worrying 3412 * over. 3413 */ 3414 if ( (bufmallocspace < maxbufmallocspace) && 3415 (bp->b_bufsize == 0) && 3416 (mbsize <= PAGE_SIZE/2)) { 3417 3418 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 3419 bp->b_bufsize = mbsize; 3420 bp->b_bcount = size; 3421 bp->b_flags |= B_MALLOC; 3422 atomic_add_long(&bufmallocspace, mbsize); 3423 return 1; 3424 } 3425 origbuf = NULL; 3426 origbufsize = 0; 3427 /* 3428 * If the buffer is growing on its other-than-first allocation, 3429 * then we revert to the page-allocation scheme. 3430 */ 3431 if (bp->b_flags & B_MALLOC) { 3432 origbuf = bp->b_data; 3433 origbufsize = bp->b_bufsize; 3434 bp->b_data = bp->b_kvabase; 3435 if (bp->b_bufsize) { 3436 atomic_subtract_long(&bufmallocspace, 3437 bp->b_bufsize); 3438 bufspacewakeup(); 3439 bp->b_bufsize = 0; 3440 } 3441 bp->b_flags &= ~B_MALLOC; 3442 newbsize = round_page(newbsize); 3443 } 3444 vm_hold_load_pages( 3445 bp, 3446 (vm_offset_t) bp->b_data + bp->b_bufsize, 3447 (vm_offset_t) bp->b_data + newbsize); 3448 if (origbuf) { 3449 bcopy(origbuf, bp->b_data, origbufsize); 3450 free(origbuf, M_BIOBUF); 3451 } 3452 } 3453 } else { 3454 int desiredpages; 3455 3456 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3457 desiredpages = (size == 0) ? 0 : 3458 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 3459 3460 if (bp->b_flags & B_MALLOC) 3461 panic("allocbuf: VMIO buffer can't be malloced"); 3462 /* 3463 * Set B_CACHE initially if buffer is 0 length or will become 3464 * 0-length. 3465 */ 3466 if (size == 0 || bp->b_bufsize == 0) 3467 bp->b_flags |= B_CACHE; 3468 3469 if (newbsize < bp->b_bufsize) { 3470 /* 3471 * DEV_BSIZE aligned new buffer size is less then the 3472 * DEV_BSIZE aligned existing buffer size. Figure out 3473 * if we have to remove any pages. 3474 */ 3475 if (desiredpages < bp->b_npages) { 3476 vm_page_t m; 3477 3478 if ((bp->b_flags & B_UNMAPPED) == 0) { 3479 BUF_CHECK_MAPPED(bp); 3480 pmap_qremove((vm_offset_t)trunc_page( 3481 (vm_offset_t)bp->b_data) + 3482 (desiredpages << PAGE_SHIFT), 3483 (bp->b_npages - desiredpages)); 3484 } else 3485 BUF_CHECK_UNMAPPED(bp); 3486 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3487 for (i = desiredpages; i < bp->b_npages; i++) { 3488 /* 3489 * the page is not freed here -- it 3490 * is the responsibility of 3491 * vnode_pager_setsize 3492 */ 3493 m = bp->b_pages[i]; 3494 KASSERT(m != bogus_page, 3495 ("allocbuf: bogus page found")); 3496 while (vm_page_sleep_if_busy(m, 3497 "biodep")) 3498 continue; 3499 3500 bp->b_pages[i] = NULL; 3501 vm_page_lock(m); 3502 vm_page_unwire(m, 0); 3503 vm_page_unlock(m); 3504 } 3505 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3506 bp->b_npages = desiredpages; 3507 } 3508 } else if (size > bp->b_bcount) { 3509 /* 3510 * We are growing the buffer, possibly in a 3511 * byte-granular fashion. 3512 */ 3513 vm_object_t obj; 3514 vm_offset_t toff; 3515 vm_offset_t tinc; 3516 3517 /* 3518 * Step 1, bring in the VM pages from the object, 3519 * allocating them if necessary. We must clear 3520 * B_CACHE if these pages are not valid for the 3521 * range covered by the buffer. 3522 */ 3523 3524 obj = bp->b_bufobj->bo_object; 3525 3526 VM_OBJECT_WLOCK(obj); 3527 while (bp->b_npages < desiredpages) { 3528 vm_page_t m; 3529 3530 /* 3531 * We must allocate system pages since blocking 3532 * here could interfere with paging I/O, no 3533 * matter which process we are. 3534 * 3535 * Only exclusive busy can be tested here. 3536 * Blocking on shared busy might lead to 3537 * deadlocks once allocbuf() is called after 3538 * pages are vfs_busy_pages(). 3539 */ 3540 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + 3541 bp->b_npages, VM_ALLOC_NOBUSY | 3542 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3543 VM_ALLOC_IGN_SBUSY | 3544 VM_ALLOC_COUNT(desiredpages - bp->b_npages)); 3545 if (m->valid == 0) 3546 bp->b_flags &= ~B_CACHE; 3547 bp->b_pages[bp->b_npages] = m; 3548 ++bp->b_npages; 3549 } 3550 3551 /* 3552 * Step 2. We've loaded the pages into the buffer, 3553 * we have to figure out if we can still have B_CACHE 3554 * set. Note that B_CACHE is set according to the 3555 * byte-granular range ( bcount and size ), new the 3556 * aligned range ( newbsize ). 3557 * 3558 * The VM test is against m->valid, which is DEV_BSIZE 3559 * aligned. Needless to say, the validity of the data 3560 * needs to also be DEV_BSIZE aligned. Note that this 3561 * fails with NFS if the server or some other client 3562 * extends the file's EOF. If our buffer is resized, 3563 * B_CACHE may remain set! XXX 3564 */ 3565 3566 toff = bp->b_bcount; 3567 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3568 3569 while ((bp->b_flags & B_CACHE) && toff < size) { 3570 vm_pindex_t pi; 3571 3572 if (tinc > (size - toff)) 3573 tinc = size - toff; 3574 3575 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3576 PAGE_SHIFT; 3577 3578 vfs_buf_test_cache( 3579 bp, 3580 bp->b_offset, 3581 toff, 3582 tinc, 3583 bp->b_pages[pi] 3584 ); 3585 toff += tinc; 3586 tinc = PAGE_SIZE; 3587 } 3588 VM_OBJECT_WUNLOCK(obj); 3589 3590 /* 3591 * Step 3, fixup the KVM pmap. 3592 */ 3593 if ((bp->b_flags & B_UNMAPPED) == 0) 3594 bpmap_qenter(bp); 3595 else 3596 BUF_CHECK_UNMAPPED(bp); 3597 } 3598 } 3599 if (newbsize < bp->b_bufsize) 3600 bufspacewakeup(); 3601 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3602 bp->b_bcount = size; /* requested buffer size */ 3603 return 1; 3604} 3605 3606extern int inflight_transient_maps; 3607 3608void 3609biodone(struct bio *bp) 3610{ 3611 struct mtx *mtxp; 3612 void (*done)(struct bio *); 3613 vm_offset_t start, end; 3614 3615 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 3616 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING; 3617 bp->bio_flags |= BIO_UNMAPPED; 3618 start = trunc_page((vm_offset_t)bp->bio_data); 3619 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 3620 pmap_qremove(start, OFF_TO_IDX(end - start)); 3621 vmem_free(transient_arena, start, end - start); 3622 atomic_add_int(&inflight_transient_maps, -1); 3623 } 3624 done = bp->bio_done; 3625 if (done == NULL) { 3626 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3627 mtx_lock(mtxp); 3628 bp->bio_flags |= BIO_DONE; 3629 wakeup(bp); 3630 mtx_unlock(mtxp); 3631 } else { 3632 bp->bio_flags |= BIO_DONE; 3633 done(bp); 3634 } 3635} 3636 3637/* 3638 * Wait for a BIO to finish. 3639 */ 3640int 3641biowait(struct bio *bp, const char *wchan) 3642{ 3643 struct mtx *mtxp; 3644 3645 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3646 mtx_lock(mtxp); 3647 while ((bp->bio_flags & BIO_DONE) == 0) 3648 msleep(bp, mtxp, PRIBIO, wchan, 0); 3649 mtx_unlock(mtxp); 3650 if (bp->bio_error != 0) 3651 return (bp->bio_error); 3652 if (!(bp->bio_flags & BIO_ERROR)) 3653 return (0); 3654 return (EIO); 3655} 3656 3657void 3658biofinish(struct bio *bp, struct devstat *stat, int error) 3659{ 3660 3661 if (error) { 3662 bp->bio_error = error; 3663 bp->bio_flags |= BIO_ERROR; 3664 } 3665 if (stat != NULL) 3666 devstat_end_transaction_bio(stat, bp); 3667 biodone(bp); 3668} 3669 3670/* 3671 * bufwait: 3672 * 3673 * Wait for buffer I/O completion, returning error status. The buffer 3674 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3675 * error and cleared. 3676 */ 3677int 3678bufwait(struct buf *bp) 3679{ 3680 if (bp->b_iocmd == BIO_READ) 3681 bwait(bp, PRIBIO, "biord"); 3682 else 3683 bwait(bp, PRIBIO, "biowr"); 3684 if (bp->b_flags & B_EINTR) { 3685 bp->b_flags &= ~B_EINTR; 3686 return (EINTR); 3687 } 3688 if (bp->b_ioflags & BIO_ERROR) { 3689 return (bp->b_error ? bp->b_error : EIO); 3690 } else { 3691 return (0); 3692 } 3693} 3694 3695 /* 3696 * Call back function from struct bio back up to struct buf. 3697 */ 3698static void 3699bufdonebio(struct bio *bip) 3700{ 3701 struct buf *bp; 3702 3703 bp = bip->bio_caller2; 3704 bp->b_resid = bp->b_bcount - bip->bio_completed; 3705 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3706 bp->b_ioflags = bip->bio_flags; 3707 bp->b_error = bip->bio_error; 3708 if (bp->b_error) 3709 bp->b_ioflags |= BIO_ERROR; 3710 bufdone(bp); 3711 g_destroy_bio(bip); 3712} 3713 3714void 3715dev_strategy(struct cdev *dev, struct buf *bp) 3716{ 3717 struct cdevsw *csw; 3718 int ref; 3719 3720 KASSERT(dev->si_refcount > 0, 3721 ("dev_strategy on un-referenced struct cdev *(%s) %p", 3722 devtoname(dev), dev)); 3723 3724 csw = dev_refthread(dev, &ref); 3725 dev_strategy_csw(dev, csw, bp); 3726 dev_relthread(dev, ref); 3727} 3728 3729void 3730dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp) 3731{ 3732 struct bio *bip; 3733 3734 KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE, 3735 ("b_iocmd botch")); 3736 KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) || 3737 dev->si_threadcount > 0, 3738 ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev), 3739 dev)); 3740 if (csw == NULL) { 3741 bp->b_error = ENXIO; 3742 bp->b_ioflags = BIO_ERROR; 3743 bufdone(bp); 3744 return; 3745 } 3746 for (;;) { 3747 bip = g_new_bio(); 3748 if (bip != NULL) 3749 break; 3750 /* Try again later */ 3751 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3752 } 3753 bip->bio_cmd = bp->b_iocmd; 3754 bip->bio_offset = bp->b_iooffset; 3755 bip->bio_length = bp->b_bcount; 3756 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3757 bdata2bio(bp, bip); 3758 bip->bio_done = bufdonebio; 3759 bip->bio_caller2 = bp; 3760 bip->bio_dev = dev; 3761 (*csw->d_strategy)(bip); 3762} 3763 3764/* 3765 * bufdone: 3766 * 3767 * Finish I/O on a buffer, optionally calling a completion function. 3768 * This is usually called from an interrupt so process blocking is 3769 * not allowed. 3770 * 3771 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3772 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3773 * assuming B_INVAL is clear. 3774 * 3775 * For the VMIO case, we set B_CACHE if the op was a read and no 3776 * read error occured, or if the op was a write. B_CACHE is never 3777 * set if the buffer is invalid or otherwise uncacheable. 3778 * 3779 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3780 * initiator to leave B_INVAL set to brelse the buffer out of existance 3781 * in the biodone routine. 3782 */ 3783void 3784bufdone(struct buf *bp) 3785{ 3786 struct bufobj *dropobj; 3787 void (*biodone)(struct buf *); 3788 3789 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3790 dropobj = NULL; 3791 3792 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3793 BUF_ASSERT_HELD(bp); 3794 3795 runningbufwakeup(bp); 3796 if (bp->b_iocmd == BIO_WRITE) 3797 dropobj = bp->b_bufobj; 3798 /* call optional completion function if requested */ 3799 if (bp->b_iodone != NULL) { 3800 biodone = bp->b_iodone; 3801 bp->b_iodone = NULL; 3802 (*biodone) (bp); 3803 if (dropobj) 3804 bufobj_wdrop(dropobj); 3805 return; 3806 } 3807 3808 bufdone_finish(bp); 3809 3810 if (dropobj) 3811 bufobj_wdrop(dropobj); 3812} 3813 3814void 3815bufdone_finish(struct buf *bp) 3816{ 3817 BUF_ASSERT_HELD(bp); 3818 3819 if (!LIST_EMPTY(&bp->b_dep)) 3820 buf_complete(bp); 3821 3822 if (bp->b_flags & B_VMIO) { 3823 vm_ooffset_t foff; 3824 vm_page_t m; 3825 vm_object_t obj; 3826 struct vnode *vp; 3827 int bogus, i, iosize; 3828 3829 obj = bp->b_bufobj->bo_object; 3830 KASSERT(obj->paging_in_progress >= bp->b_npages, 3831 ("biodone_finish: paging in progress(%d) < b_npages(%d)", 3832 obj->paging_in_progress, bp->b_npages)); 3833 3834 vp = bp->b_vp; 3835 KASSERT(vp->v_holdcnt > 0, 3836 ("biodone_finish: vnode %p has zero hold count", vp)); 3837 KASSERT(vp->v_object != NULL, 3838 ("biodone_finish: vnode %p has no vm_object", vp)); 3839 3840 foff = bp->b_offset; 3841 KASSERT(bp->b_offset != NOOFFSET, 3842 ("biodone_finish: bp %p has no buffer offset", bp)); 3843 3844 /* 3845 * Set B_CACHE if the op was a normal read and no error 3846 * occured. B_CACHE is set for writes in the b*write() 3847 * routines. 3848 */ 3849 iosize = bp->b_bcount - bp->b_resid; 3850 if (bp->b_iocmd == BIO_READ && 3851 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3852 !(bp->b_ioflags & BIO_ERROR)) { 3853 bp->b_flags |= B_CACHE; 3854 } 3855 bogus = 0; 3856 VM_OBJECT_WLOCK(obj); 3857 for (i = 0; i < bp->b_npages; i++) { 3858 int bogusflag = 0; 3859 int resid; 3860 3861 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3862 if (resid > iosize) 3863 resid = iosize; 3864 3865 /* 3866 * cleanup bogus pages, restoring the originals 3867 */ 3868 m = bp->b_pages[i]; 3869 if (m == bogus_page) { 3870 bogus = bogusflag = 1; 3871 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3872 if (m == NULL) 3873 panic("biodone: page disappeared!"); 3874 bp->b_pages[i] = m; 3875 } 3876 KASSERT(OFF_TO_IDX(foff) == m->pindex, 3877 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch", 3878 (intmax_t)foff, (uintmax_t)m->pindex)); 3879 3880 /* 3881 * In the write case, the valid and clean bits are 3882 * already changed correctly ( see bdwrite() ), so we 3883 * only need to do this here in the read case. 3884 */ 3885 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3886 KASSERT((m->dirty & vm_page_bits(foff & 3887 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3888 " page %p has unexpected dirty bits", m)); 3889 vfs_page_set_valid(bp, foff, m); 3890 } 3891 3892 vm_page_sunbusy(m); 3893 vm_object_pip_subtract(obj, 1); 3894 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3895 iosize -= resid; 3896 } 3897 vm_object_pip_wakeupn(obj, 0); 3898 VM_OBJECT_WUNLOCK(obj); 3899 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 3900 BUF_CHECK_MAPPED(bp); 3901 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3902 bp->b_pages, bp->b_npages); 3903 } 3904 } 3905 3906 /* 3907 * For asynchronous completions, release the buffer now. The brelse 3908 * will do a wakeup there if necessary - so no need to do a wakeup 3909 * here in the async case. The sync case always needs to do a wakeup. 3910 */ 3911 3912 if (bp->b_flags & B_ASYNC) { 3913 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3914 brelse(bp); 3915 else 3916 bqrelse(bp); 3917 } else 3918 bdone(bp); 3919} 3920 3921/* 3922 * This routine is called in lieu of iodone in the case of 3923 * incomplete I/O. This keeps the busy status for pages 3924 * consistant. 3925 */ 3926void 3927vfs_unbusy_pages(struct buf *bp) 3928{ 3929 int i; 3930 vm_object_t obj; 3931 vm_page_t m; 3932 3933 runningbufwakeup(bp); 3934 if (!(bp->b_flags & B_VMIO)) 3935 return; 3936 3937 obj = bp->b_bufobj->bo_object; 3938 VM_OBJECT_WLOCK(obj); 3939 for (i = 0; i < bp->b_npages; i++) { 3940 m = bp->b_pages[i]; 3941 if (m == bogus_page) { 3942 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3943 if (!m) 3944 panic("vfs_unbusy_pages: page missing\n"); 3945 bp->b_pages[i] = m; 3946 if ((bp->b_flags & B_UNMAPPED) == 0) { 3947 BUF_CHECK_MAPPED(bp); 3948 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3949 bp->b_pages, bp->b_npages); 3950 } else 3951 BUF_CHECK_UNMAPPED(bp); 3952 } 3953 vm_object_pip_subtract(obj, 1); 3954 vm_page_sunbusy(m); 3955 } 3956 vm_object_pip_wakeupn(obj, 0); 3957 VM_OBJECT_WUNLOCK(obj); 3958} 3959 3960/* 3961 * vfs_page_set_valid: 3962 * 3963 * Set the valid bits in a page based on the supplied offset. The 3964 * range is restricted to the buffer's size. 3965 * 3966 * This routine is typically called after a read completes. 3967 */ 3968static void 3969vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3970{ 3971 vm_ooffset_t eoff; 3972 3973 /* 3974 * Compute the end offset, eoff, such that [off, eoff) does not span a 3975 * page boundary and eoff is not greater than the end of the buffer. 3976 * The end of the buffer, in this case, is our file EOF, not the 3977 * allocation size of the buffer. 3978 */ 3979 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3980 if (eoff > bp->b_offset + bp->b_bcount) 3981 eoff = bp->b_offset + bp->b_bcount; 3982 3983 /* 3984 * Set valid range. This is typically the entire buffer and thus the 3985 * entire page. 3986 */ 3987 if (eoff > off) 3988 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 3989} 3990 3991/* 3992 * vfs_page_set_validclean: 3993 * 3994 * Set the valid bits and clear the dirty bits in a page based on the 3995 * supplied offset. The range is restricted to the buffer's size. 3996 */ 3997static void 3998vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3999{ 4000 vm_ooffset_t soff, eoff; 4001 4002 /* 4003 * Start and end offsets in buffer. eoff - soff may not cross a 4004 * page boundry or cross the end of the buffer. The end of the 4005 * buffer, in this case, is our file EOF, not the allocation size 4006 * of the buffer. 4007 */ 4008 soff = off; 4009 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4010 if (eoff > bp->b_offset + bp->b_bcount) 4011 eoff = bp->b_offset + bp->b_bcount; 4012 4013 /* 4014 * Set valid range. This is typically the entire buffer and thus the 4015 * entire page. 4016 */ 4017 if (eoff > soff) { 4018 vm_page_set_validclean( 4019 m, 4020 (vm_offset_t) (soff & PAGE_MASK), 4021 (vm_offset_t) (eoff - soff) 4022 ); 4023 } 4024} 4025 4026/* 4027 * Ensure that all buffer pages are not exclusive busied. If any page is 4028 * exclusive busy, drain it. 4029 */ 4030void 4031vfs_drain_busy_pages(struct buf *bp) 4032{ 4033 vm_page_t m; 4034 int i, last_busied; 4035 4036 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); 4037 last_busied = 0; 4038 for (i = 0; i < bp->b_npages; i++) { 4039 m = bp->b_pages[i]; 4040 if (vm_page_xbusied(m)) { 4041 for (; last_busied < i; last_busied++) 4042 vm_page_sbusy(bp->b_pages[last_busied]); 4043 while (vm_page_xbusied(m)) { 4044 vm_page_lock(m); 4045 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4046 vm_page_busy_sleep(m, "vbpage"); 4047 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4048 } 4049 } 4050 } 4051 for (i = 0; i < last_busied; i++) 4052 vm_page_sunbusy(bp->b_pages[i]); 4053} 4054 4055/* 4056 * This routine is called before a device strategy routine. 4057 * It is used to tell the VM system that paging I/O is in 4058 * progress, and treat the pages associated with the buffer 4059 * almost as being exclusive busy. Also the object paging_in_progress 4060 * flag is handled to make sure that the object doesn't become 4061 * inconsistant. 4062 * 4063 * Since I/O has not been initiated yet, certain buffer flags 4064 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 4065 * and should be ignored. 4066 */ 4067void 4068vfs_busy_pages(struct buf *bp, int clear_modify) 4069{ 4070 int i, bogus; 4071 vm_object_t obj; 4072 vm_ooffset_t foff; 4073 vm_page_t m; 4074 4075 if (!(bp->b_flags & B_VMIO)) 4076 return; 4077 4078 obj = bp->b_bufobj->bo_object; 4079 foff = bp->b_offset; 4080 KASSERT(bp->b_offset != NOOFFSET, 4081 ("vfs_busy_pages: no buffer offset")); 4082 VM_OBJECT_WLOCK(obj); 4083 vfs_drain_busy_pages(bp); 4084 if (bp->b_bufsize != 0) 4085 vfs_setdirty_locked_object(bp); 4086 bogus = 0; 4087 for (i = 0; i < bp->b_npages; i++) { 4088 m = bp->b_pages[i]; 4089 4090 if ((bp->b_flags & B_CLUSTER) == 0) { 4091 vm_object_pip_add(obj, 1); 4092 vm_page_sbusy(m); 4093 } 4094 /* 4095 * When readying a buffer for a read ( i.e 4096 * clear_modify == 0 ), it is important to do 4097 * bogus_page replacement for valid pages in 4098 * partially instantiated buffers. Partially 4099 * instantiated buffers can, in turn, occur when 4100 * reconstituting a buffer from its VM backing store 4101 * base. We only have to do this if B_CACHE is 4102 * clear ( which causes the I/O to occur in the 4103 * first place ). The replacement prevents the read 4104 * I/O from overwriting potentially dirty VM-backed 4105 * pages. XXX bogus page replacement is, uh, bogus. 4106 * It may not work properly with small-block devices. 4107 * We need to find a better way. 4108 */ 4109 if (clear_modify) { 4110 pmap_remove_write(m); 4111 vfs_page_set_validclean(bp, foff, m); 4112 } else if (m->valid == VM_PAGE_BITS_ALL && 4113 (bp->b_flags & B_CACHE) == 0) { 4114 bp->b_pages[i] = bogus_page; 4115 bogus++; 4116 } 4117 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4118 } 4119 VM_OBJECT_WUNLOCK(obj); 4120 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 4121 BUF_CHECK_MAPPED(bp); 4122 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4123 bp->b_pages, bp->b_npages); 4124 } 4125} 4126 4127/* 4128 * vfs_bio_set_valid: 4129 * 4130 * Set the range within the buffer to valid. The range is 4131 * relative to the beginning of the buffer, b_offset. Note that 4132 * b_offset itself may be offset from the beginning of the first 4133 * page. 4134 */ 4135void 4136vfs_bio_set_valid(struct buf *bp, int base, int size) 4137{ 4138 int i, n; 4139 vm_page_t m; 4140 4141 if (!(bp->b_flags & B_VMIO)) 4142 return; 4143 4144 /* 4145 * Fixup base to be relative to beginning of first page. 4146 * Set initial n to be the maximum number of bytes in the 4147 * first page that can be validated. 4148 */ 4149 base += (bp->b_offset & PAGE_MASK); 4150 n = PAGE_SIZE - (base & PAGE_MASK); 4151 4152 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4153 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4154 m = bp->b_pages[i]; 4155 if (n > size) 4156 n = size; 4157 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4158 base += n; 4159 size -= n; 4160 n = PAGE_SIZE; 4161 } 4162 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4163} 4164 4165/* 4166 * vfs_bio_clrbuf: 4167 * 4168 * If the specified buffer is a non-VMIO buffer, clear the entire 4169 * buffer. If the specified buffer is a VMIO buffer, clear and 4170 * validate only the previously invalid portions of the buffer. 4171 * This routine essentially fakes an I/O, so we need to clear 4172 * BIO_ERROR and B_INVAL. 4173 * 4174 * Note that while we only theoretically need to clear through b_bcount, 4175 * we go ahead and clear through b_bufsize. 4176 */ 4177void 4178vfs_bio_clrbuf(struct buf *bp) 4179{ 4180 int i, j, mask, sa, ea, slide; 4181 4182 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4183 clrbuf(bp); 4184 return; 4185 } 4186 bp->b_flags &= ~B_INVAL; 4187 bp->b_ioflags &= ~BIO_ERROR; 4188 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4189 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4190 (bp->b_offset & PAGE_MASK) == 0) { 4191 if (bp->b_pages[0] == bogus_page) 4192 goto unlock; 4193 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4194 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object); 4195 if ((bp->b_pages[0]->valid & mask) == mask) 4196 goto unlock; 4197 if ((bp->b_pages[0]->valid & mask) == 0) { 4198 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize); 4199 bp->b_pages[0]->valid |= mask; 4200 goto unlock; 4201 } 4202 } 4203 sa = bp->b_offset & PAGE_MASK; 4204 slide = 0; 4205 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4206 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4207 ea = slide & PAGE_MASK; 4208 if (ea == 0) 4209 ea = PAGE_SIZE; 4210 if (bp->b_pages[i] == bogus_page) 4211 continue; 4212 j = sa / DEV_BSIZE; 4213 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4214 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object); 4215 if ((bp->b_pages[i]->valid & mask) == mask) 4216 continue; 4217 if ((bp->b_pages[i]->valid & mask) == 0) 4218 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4219 else { 4220 for (; sa < ea; sa += DEV_BSIZE, j++) { 4221 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4222 pmap_zero_page_area(bp->b_pages[i], 4223 sa, DEV_BSIZE); 4224 } 4225 } 4226 } 4227 bp->b_pages[i]->valid |= mask; 4228 } 4229unlock: 4230 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4231 bp->b_resid = 0; 4232} 4233 4234void 4235vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4236{ 4237 vm_page_t m; 4238 int i, n; 4239 4240 if ((bp->b_flags & B_UNMAPPED) == 0) { 4241 BUF_CHECK_MAPPED(bp); 4242 bzero(bp->b_data + base, size); 4243 } else { 4244 BUF_CHECK_UNMAPPED(bp); 4245 n = PAGE_SIZE - (base & PAGE_MASK); 4246 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4247 m = bp->b_pages[i]; 4248 if (n > size) 4249 n = size; 4250 pmap_zero_page_area(m, base & PAGE_MASK, n); 4251 base += n; 4252 size -= n; 4253 n = PAGE_SIZE; 4254 } 4255 } 4256} 4257 4258/* 4259 * vm_hold_load_pages and vm_hold_free_pages get pages into 4260 * a buffers address space. The pages are anonymous and are 4261 * not associated with a file object. 4262 */ 4263static void 4264vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4265{ 4266 vm_offset_t pg; 4267 vm_page_t p; 4268 int index; 4269 4270 BUF_CHECK_MAPPED(bp); 4271 4272 to = round_page(to); 4273 from = round_page(from); 4274 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4275 4276 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4277tryagain: 4278 /* 4279 * note: must allocate system pages since blocking here 4280 * could interfere with paging I/O, no matter which 4281 * process we are. 4282 */ 4283 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | 4284 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT)); 4285 if (p == NULL) { 4286 VM_WAIT; 4287 goto tryagain; 4288 } 4289 pmap_qenter(pg, &p, 1); 4290 bp->b_pages[index] = p; 4291 } 4292 bp->b_npages = index; 4293} 4294 4295/* Return pages associated with this buf to the vm system */ 4296static void 4297vm_hold_free_pages(struct buf *bp, int newbsize) 4298{ 4299 vm_offset_t from; 4300 vm_page_t p; 4301 int index, newnpages; 4302 4303 BUF_CHECK_MAPPED(bp); 4304 4305 from = round_page((vm_offset_t)bp->b_data + newbsize); 4306 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4307 if (bp->b_npages > newnpages) 4308 pmap_qremove(from, bp->b_npages - newnpages); 4309 for (index = newnpages; index < bp->b_npages; index++) { 4310 p = bp->b_pages[index]; 4311 bp->b_pages[index] = NULL; 4312 if (vm_page_sbusied(p)) 4313 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 4314 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); 4315 p->wire_count--; 4316 vm_page_free(p); 4317 atomic_subtract_int(&cnt.v_wire_count, 1); 4318 } 4319 bp->b_npages = newnpages; 4320} 4321 4322/* 4323 * Map an IO request into kernel virtual address space. 4324 * 4325 * All requests are (re)mapped into kernel VA space. 4326 * Notice that we use b_bufsize for the size of the buffer 4327 * to be mapped. b_bcount might be modified by the driver. 4328 * 4329 * Note that even if the caller determines that the address space should 4330 * be valid, a race or a smaller-file mapped into a larger space may 4331 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 4332 * check the return value. 4333 */ 4334int 4335vmapbuf(struct buf *bp, int mapbuf) 4336{ 4337 caddr_t kva; 4338 vm_prot_t prot; 4339 int pidx; 4340 4341 if (bp->b_bufsize < 0) 4342 return (-1); 4343 prot = VM_PROT_READ; 4344 if (bp->b_iocmd == BIO_READ) 4345 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 4346 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 4347 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, 4348 btoc(MAXPHYS))) < 0) 4349 return (-1); 4350 bp->b_npages = pidx; 4351 if (mapbuf || !unmapped_buf_allowed) { 4352 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 4353 kva = bp->b_saveaddr; 4354 bp->b_saveaddr = bp->b_data; 4355 bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK); 4356 bp->b_flags &= ~B_UNMAPPED; 4357 } else { 4358 bp->b_flags |= B_UNMAPPED; 4359 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK; 4360 bp->b_saveaddr = bp->b_data; 4361 bp->b_data = unmapped_buf; 4362 } 4363 return(0); 4364} 4365 4366/* 4367 * Free the io map PTEs associated with this IO operation. 4368 * We also invalidate the TLB entries and restore the original b_addr. 4369 */ 4370void 4371vunmapbuf(struct buf *bp) 4372{ 4373 int npages; 4374 4375 npages = bp->b_npages; 4376 if (bp->b_flags & B_UNMAPPED) 4377 bp->b_flags &= ~B_UNMAPPED; 4378 else 4379 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4380 vm_page_unhold_pages(bp->b_pages, npages); 4381 4382 bp->b_data = bp->b_saveaddr; 4383} 4384 4385void 4386bdone(struct buf *bp) 4387{ 4388 struct mtx *mtxp; 4389 4390 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4391 mtx_lock(mtxp); 4392 bp->b_flags |= B_DONE; 4393 wakeup(bp); 4394 mtx_unlock(mtxp); 4395} 4396 4397void 4398bwait(struct buf *bp, u_char pri, const char *wchan) 4399{ 4400 struct mtx *mtxp; 4401 4402 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4403 mtx_lock(mtxp); 4404 while ((bp->b_flags & B_DONE) == 0) 4405 msleep(bp, mtxp, pri, wchan, 0); 4406 mtx_unlock(mtxp); 4407} 4408 4409int 4410bufsync(struct bufobj *bo, int waitfor) 4411{ 4412 4413 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 4414} 4415 4416void 4417bufstrategy(struct bufobj *bo, struct buf *bp) 4418{ 4419 int i = 0; 4420 struct vnode *vp; 4421 4422 vp = bp->b_vp; 4423 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 4424 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 4425 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 4426 i = VOP_STRATEGY(vp, bp); 4427 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 4428} 4429 4430void 4431bufobj_wrefl(struct bufobj *bo) 4432{ 4433 4434 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4435 ASSERT_BO_WLOCKED(bo); 4436 bo->bo_numoutput++; 4437} 4438 4439void 4440bufobj_wref(struct bufobj *bo) 4441{ 4442 4443 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4444 BO_LOCK(bo); 4445 bo->bo_numoutput++; 4446 BO_UNLOCK(bo); 4447} 4448 4449void 4450bufobj_wdrop(struct bufobj *bo) 4451{ 4452 4453 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 4454 BO_LOCK(bo); 4455 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 4456 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 4457 bo->bo_flag &= ~BO_WWAIT; 4458 wakeup(&bo->bo_numoutput); 4459 } 4460 BO_UNLOCK(bo); 4461} 4462 4463int 4464bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 4465{ 4466 int error; 4467 4468 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 4469 ASSERT_BO_WLOCKED(bo); 4470 error = 0; 4471 while (bo->bo_numoutput) { 4472 bo->bo_flag |= BO_WWAIT; 4473 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), 4474 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 4475 if (error) 4476 break; 4477 } 4478 return (error); 4479} 4480 4481void 4482bpin(struct buf *bp) 4483{ 4484 struct mtx *mtxp; 4485 4486 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4487 mtx_lock(mtxp); 4488 bp->b_pin_count++; 4489 mtx_unlock(mtxp); 4490} 4491 4492void 4493bunpin(struct buf *bp) 4494{ 4495 struct mtx *mtxp; 4496 4497 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4498 mtx_lock(mtxp); 4499 if (--bp->b_pin_count == 0) 4500 wakeup(bp); 4501 mtx_unlock(mtxp); 4502} 4503 4504void 4505bunpin_wait(struct buf *bp) 4506{ 4507 struct mtx *mtxp; 4508 4509 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4510 mtx_lock(mtxp); 4511 while (bp->b_pin_count > 0) 4512 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4513 mtx_unlock(mtxp); 4514} 4515 4516/* 4517 * Set bio_data or bio_ma for struct bio from the struct buf. 4518 */ 4519void 4520bdata2bio(struct buf *bp, struct bio *bip) 4521{ 4522 4523 if ((bp->b_flags & B_UNMAPPED) != 0) { 4524 KASSERT(unmapped_buf_allowed, ("unmapped")); 4525 bip->bio_ma = bp->b_pages; 4526 bip->bio_ma_n = bp->b_npages; 4527 bip->bio_data = unmapped_buf; 4528 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 4529 bip->bio_flags |= BIO_UNMAPPED; 4530 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 4531 PAGE_SIZE == bp->b_npages, 4532 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset, 4533 (long long)bip->bio_length, bip->bio_ma_n)); 4534 } else { 4535 bip->bio_data = bp->b_data; 4536 bip->bio_ma = NULL; 4537 } 4538} 4539 4540#include "opt_ddb.h" 4541#ifdef DDB 4542#include <ddb/ddb.h> 4543 4544/* DDB command to show buffer data */ 4545DB_SHOW_COMMAND(buffer, db_show_buffer) 4546{ 4547 /* get args */ 4548 struct buf *bp = (struct buf *)addr; 4549 4550 if (!have_addr) { 4551 db_printf("usage: show buffer <addr>\n"); 4552 return; 4553 } 4554 4555 db_printf("buf at %p\n", bp); 4556 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n", 4557 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, 4558 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS); 4559 db_printf( 4560 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4561 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, " 4562 "b_dep = %p\n", 4563 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4564 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4565 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first); 4566 if (bp->b_npages) { 4567 int i; 4568 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4569 for (i = 0; i < bp->b_npages; i++) { 4570 vm_page_t m; 4571 m = bp->b_pages[i]; 4572 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4573 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4574 if ((i + 1) < bp->b_npages) 4575 db_printf(","); 4576 } 4577 db_printf("\n"); 4578 } 4579 db_printf(" "); 4580 BUF_LOCKPRINTINFO(bp); 4581} 4582 4583DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4584{ 4585 struct buf *bp; 4586 int i; 4587 4588 for (i = 0; i < nbuf; i++) { 4589 bp = &buf[i]; 4590 if (BUF_ISLOCKED(bp)) { 4591 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4592 db_printf("\n"); 4593 } 4594 } 4595} 4596 4597DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4598{ 4599 struct vnode *vp; 4600 struct buf *bp; 4601 4602 if (!have_addr) { 4603 db_printf("usage: show vnodebufs <addr>\n"); 4604 return; 4605 } 4606 vp = (struct vnode *)addr; 4607 db_printf("Clean buffers:\n"); 4608 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4609 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4610 db_printf("\n"); 4611 } 4612 db_printf("Dirty buffers:\n"); 4613 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4614 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4615 db_printf("\n"); 4616 } 4617} 4618 4619DB_COMMAND(countfreebufs, db_coundfreebufs) 4620{ 4621 struct buf *bp; 4622 int i, used = 0, nfree = 0; 4623 4624 if (have_addr) { 4625 db_printf("usage: countfreebufs\n"); 4626 return; 4627 } 4628 4629 for (i = 0; i < nbuf; i++) { 4630 bp = &buf[i]; 4631 if ((bp->b_flags & B_INFREECNT) != 0) 4632 nfree++; 4633 else 4634 used++; 4635 } 4636 4637 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4638 nfree + used); 4639 db_printf("numfreebuffers is %d\n", numfreebuffers); 4640} 4641#endif /* DDB */ 4642