vfs_bio.c revision 251282
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: head/sys/kern/vfs_bio.c 251282 2013-06-03 04:16:48Z 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/vmmeter.h> 68#include <sys/vnode.h> 69#include <geom/geom.h> 70#include <vm/vm.h> 71#include <vm/vm_param.h> 72#include <vm/vm_kern.h> 73#include <vm/vm_pageout.h> 74#include <vm/vm_page.h> 75#include <vm/vm_object.h> 76#include <vm/vm_extern.h> 77#include <vm/vm_map.h> 78#include "opt_compat.h" 79#include "opt_directio.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 101static struct proc *bufdaemonproc; 102 103static int inmem(struct vnode *vp, daddr_t blkno); 104static void vm_hold_free_pages(struct buf *bp, int newbsize); 105static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 106 vm_offset_t to); 107static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 108static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 109 vm_page_t m); 110static void vfs_drain_busy_pages(struct buf *bp); 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_do_flush(struct vnode *vp); 117static int flushbufqueues(struct vnode *, int, int); 118static void buf_daemon(void); 119static void bremfreel(struct buf *bp); 120#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 121 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 122static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 123#endif 124 125int vmiodirenable = TRUE; 126SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 127 "Use the VM system for directory writes"); 128long runningbufspace; 129SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 130 "Amount of presently outstanding async buffer io"); 131static long bufspace; 132#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 133 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 134SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 135 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers"); 136#else 137SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 138 "Virtual memory used for buffers"); 139#endif 140static long unmapped_bufspace; 141SYSCTL_LONG(_vfs, OID_AUTO, unmapped_bufspace, CTLFLAG_RD, 142 &unmapped_bufspace, 0, 143 "Amount of unmapped buffers, inclusive in the bufspace"); 144static long maxbufspace; 145SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 146 "Maximum allowed value of bufspace (including buf_daemon)"); 147static long bufmallocspace; 148SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 149 "Amount of malloced memory for buffers"); 150static long maxbufmallocspace; 151SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 152 "Maximum amount of malloced memory for buffers"); 153static long lobufspace; 154SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 155 "Minimum amount of buffers we want to have"); 156long hibufspace; 157SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 158 "Maximum allowed value of bufspace (excluding buf_daemon)"); 159static int bufreusecnt; 160SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 161 "Number of times we have reused a buffer"); 162static int buffreekvacnt; 163SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 164 "Number of times we have freed the KVA space from some buffer"); 165static int bufdefragcnt; 166SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 167 "Number of times we have had to repeat buffer allocation to defragment"); 168static long lorunningspace; 169SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 170 "Minimum preferred space used for in-progress I/O"); 171static long hirunningspace; 172SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 173 "Maximum amount of space to use for in-progress I/O"); 174int dirtybufferflushes; 175SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 176 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 177int bdwriteskip; 178SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 179 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 180int altbufferflushes; 181SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 182 0, "Number of fsync flushes to limit dirty buffers"); 183static int recursiveflushes; 184SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 185 0, "Number of flushes skipped due to being recursive"); 186static int numdirtybuffers; 187SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 188 "Number of buffers that are dirty (has unwritten changes) at the moment"); 189static int lodirtybuffers; 190SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 191 "How many buffers we want to have free before bufdaemon can sleep"); 192static int hidirtybuffers; 193SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 194 "When the number of dirty buffers is considered severe"); 195int dirtybufthresh; 196SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 197 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 198static int numfreebuffers; 199SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 200 "Number of free buffers"); 201static int lofreebuffers; 202SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 203 "XXX Unused"); 204static int hifreebuffers; 205SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 206 "XXX Complicatedly unused"); 207static int getnewbufcalls; 208SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 209 "Number of calls to getnewbuf"); 210static int getnewbufrestarts; 211SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 212 "Number of times getnewbuf has had to restart a buffer aquisition"); 213static int mappingrestarts; 214SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0, 215 "Number of times getblk has had to restart a buffer mapping for " 216 "unmapped buffer"); 217static int flushbufqtarget = 100; 218SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 219 "Amount of work to do in flushbufqueues when helping bufdaemon"); 220static long notbufdflashes; 221SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, ¬bufdflashes, 0, 222 "Number of dirty buffer flushes done by the bufdaemon helpers"); 223static long barrierwrites; 224SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0, 225 "Number of barrier writes"); 226SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD, 227 &unmapped_buf_allowed, 0, 228 "Permit the use of the unmapped i/o"); 229 230/* 231 * Wakeup point for bufdaemon, as well as indicator of whether it is already 232 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 233 * is idling. 234 */ 235static int bd_request; 236 237/* 238 * Request for the buf daemon to write more buffers than is indicated by 239 * lodirtybuf. This may be necessary to push out excess dependencies or 240 * defragment the address space where a simple count of the number of dirty 241 * buffers is insufficient to characterize the demand for flushing them. 242 */ 243static int bd_speedupreq; 244 245/* 246 * This lock synchronizes access to bd_request. 247 */ 248static struct mtx bdlock; 249 250/* 251 * bogus page -- for I/O to/from partially complete buffers 252 * this is a temporary solution to the problem, but it is not 253 * really that bad. it would be better to split the buffer 254 * for input in the case of buffers partially already in memory, 255 * but the code is intricate enough already. 256 */ 257vm_page_t bogus_page; 258 259/* 260 * Synchronization (sleep/wakeup) variable for active buffer space requests. 261 * Set when wait starts, cleared prior to wakeup(). 262 * Used in runningbufwakeup() and waitrunningbufspace(). 263 */ 264static int runningbufreq; 265 266/* 267 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 268 * waitrunningbufspace(). 269 */ 270static struct mtx rbreqlock; 271 272/* 273 * Synchronization (sleep/wakeup) variable for buffer requests. 274 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 275 * by and/or. 276 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 277 * getnewbuf(), and getblk(). 278 */ 279static int needsbuffer; 280 281/* 282 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 283 */ 284static struct mtx nblock; 285 286/* 287 * Definitions for the buffer free lists. 288 */ 289#define BUFFER_QUEUES 5 /* number of free buffer queues */ 290 291#define QUEUE_NONE 0 /* on no queue */ 292#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 293#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 294#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 295#define QUEUE_EMPTY 4 /* empty buffer headers */ 296#define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */ 297 298/* Queues for free buffers with various properties */ 299static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 300#ifdef INVARIANTS 301static int bq_len[BUFFER_QUEUES]; 302#endif 303 304/* Lock for the bufqueues */ 305static struct mtx bqlock; 306 307/* 308 * Single global constant for BUF_WMESG, to avoid getting multiple references. 309 * buf_wmesg is referred from macros. 310 */ 311const char *buf_wmesg = BUF_WMESG; 312 313#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 314#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 315#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 316#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 317 318#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 319 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 320static int 321sysctl_bufspace(SYSCTL_HANDLER_ARGS) 322{ 323 long lvalue; 324 int ivalue; 325 326 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 327 return (sysctl_handle_long(oidp, arg1, arg2, req)); 328 lvalue = *(long *)arg1; 329 if (lvalue > INT_MAX) 330 /* On overflow, still write out a long to trigger ENOMEM. */ 331 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 332 ivalue = lvalue; 333 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 334} 335#endif 336 337#ifdef DIRECTIO 338extern void ffs_rawread_setup(void); 339#endif /* DIRECTIO */ 340/* 341 * numdirtywakeup: 342 * 343 * If someone is blocked due to there being too many dirty buffers, 344 * and numdirtybuffers is now reasonable, wake them up. 345 */ 346 347static __inline void 348numdirtywakeup(int level) 349{ 350 351 if (numdirtybuffers <= level) { 352 mtx_lock(&nblock); 353 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 354 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 355 wakeup(&needsbuffer); 356 } 357 mtx_unlock(&nblock); 358 } 359} 360 361/* 362 * bufspacewakeup: 363 * 364 * Called when buffer space is potentially available for recovery. 365 * getnewbuf() will block on this flag when it is unable to free 366 * sufficient buffer space. Buffer space becomes recoverable when 367 * bp's get placed back in the queues. 368 */ 369 370static __inline void 371bufspacewakeup(void) 372{ 373 374 /* 375 * If someone is waiting for BUF space, wake them up. Even 376 * though we haven't freed the kva space yet, the waiting 377 * process will be able to now. 378 */ 379 mtx_lock(&nblock); 380 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 381 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 382 wakeup(&needsbuffer); 383 } 384 mtx_unlock(&nblock); 385} 386 387/* 388 * runningbufwakeup() - in-progress I/O accounting. 389 * 390 */ 391void 392runningbufwakeup(struct buf *bp) 393{ 394 395 if (bp->b_runningbufspace) { 396 atomic_subtract_long(&runningbufspace, bp->b_runningbufspace); 397 bp->b_runningbufspace = 0; 398 mtx_lock(&rbreqlock); 399 if (runningbufreq && runningbufspace <= lorunningspace) { 400 runningbufreq = 0; 401 wakeup(&runningbufreq); 402 } 403 mtx_unlock(&rbreqlock); 404 } 405} 406 407/* 408 * bufcountwakeup: 409 * 410 * Called when a buffer has been added to one of the free queues to 411 * account for the buffer and to wakeup anyone waiting for free buffers. 412 * This typically occurs when large amounts of metadata are being handled 413 * by the buffer cache ( else buffer space runs out first, usually ). 414 */ 415 416static __inline void 417bufcountwakeup(struct buf *bp) 418{ 419 int old; 420 421 KASSERT((bp->b_flags & B_INFREECNT) == 0, 422 ("buf %p already counted as free", bp)); 423 bp->b_flags |= B_INFREECNT; 424 old = atomic_fetchadd_int(&numfreebuffers, 1); 425 KASSERT(old >= 0 && old < nbuf, 426 ("numfreebuffers climbed to %d", old + 1)); 427 mtx_lock(&nblock); 428 if (needsbuffer) { 429 needsbuffer &= ~VFS_BIO_NEED_ANY; 430 if (numfreebuffers >= hifreebuffers) 431 needsbuffer &= ~VFS_BIO_NEED_FREE; 432 wakeup(&needsbuffer); 433 } 434 mtx_unlock(&nblock); 435} 436 437/* 438 * waitrunningbufspace() 439 * 440 * runningbufspace is a measure of the amount of I/O currently 441 * running. This routine is used in async-write situations to 442 * prevent creating huge backups of pending writes to a device. 443 * Only asynchronous writes are governed by this function. 444 * 445 * Reads will adjust runningbufspace, but will not block based on it. 446 * The read load has a side effect of reducing the allowed write load. 447 * 448 * This does NOT turn an async write into a sync write. It waits 449 * for earlier writes to complete and generally returns before the 450 * caller's write has reached the device. 451 */ 452void 453waitrunningbufspace(void) 454{ 455 456 mtx_lock(&rbreqlock); 457 while (runningbufspace > hirunningspace) { 458 ++runningbufreq; 459 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 460 } 461 mtx_unlock(&rbreqlock); 462} 463 464 465/* 466 * vfs_buf_test_cache: 467 * 468 * Called when a buffer is extended. This function clears the B_CACHE 469 * bit if the newly extended portion of the buffer does not contain 470 * valid data. 471 */ 472static __inline 473void 474vfs_buf_test_cache(struct buf *bp, 475 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 476 vm_page_t m) 477{ 478 479 VM_OBJECT_ASSERT_WLOCKED(m->object); 480 if (bp->b_flags & B_CACHE) { 481 int base = (foff + off) & PAGE_MASK; 482 if (vm_page_is_valid(m, base, size) == 0) 483 bp->b_flags &= ~B_CACHE; 484 } 485} 486 487/* Wake up the buffer daemon if necessary */ 488static __inline 489void 490bd_wakeup(int dirtybuflevel) 491{ 492 493 mtx_lock(&bdlock); 494 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 495 bd_request = 1; 496 wakeup(&bd_request); 497 } 498 mtx_unlock(&bdlock); 499} 500 501/* 502 * bd_speedup - speedup the buffer cache flushing code 503 */ 504 505void 506bd_speedup(void) 507{ 508 int needwake; 509 510 mtx_lock(&bdlock); 511 needwake = 0; 512 if (bd_speedupreq == 0 || bd_request == 0) 513 needwake = 1; 514 bd_speedupreq = 1; 515 bd_request = 1; 516 if (needwake) 517 wakeup(&bd_request); 518 mtx_unlock(&bdlock); 519} 520 521#ifdef __i386__ 522#define TRANSIENT_DENOM 5 523#else 524#define TRANSIENT_DENOM 10 525#endif 526 527/* 528 * Calculating buffer cache scaling values and reserve space for buffer 529 * headers. This is called during low level kernel initialization and 530 * may be called more then once. We CANNOT write to the memory area 531 * being reserved at this time. 532 */ 533caddr_t 534kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 535{ 536 int tuned_nbuf; 537 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz; 538 539 /* 540 * physmem_est is in pages. Convert it to kilobytes (assumes 541 * PAGE_SIZE is >= 1K) 542 */ 543 physmem_est = physmem_est * (PAGE_SIZE / 1024); 544 545 /* 546 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 547 * For the first 64MB of ram nominally allocate sufficient buffers to 548 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 549 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 550 * the buffer cache we limit the eventual kva reservation to 551 * maxbcache bytes. 552 * 553 * factor represents the 1/4 x ram conversion. 554 */ 555 if (nbuf == 0) { 556 int factor = 4 * BKVASIZE / 1024; 557 558 nbuf = 50; 559 if (physmem_est > 4096) 560 nbuf += min((physmem_est - 4096) / factor, 561 65536 / factor); 562 if (physmem_est > 65536) 563 nbuf += min((physmem_est - 65536) * 2 / (factor * 5), 564 32 * 1024 * 1024 / (factor * 5)); 565 566 if (maxbcache && nbuf > maxbcache / BKVASIZE) 567 nbuf = maxbcache / BKVASIZE; 568 tuned_nbuf = 1; 569 } else 570 tuned_nbuf = 0; 571 572 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 573 maxbuf = (LONG_MAX / 3) / BKVASIZE; 574 if (nbuf > maxbuf) { 575 if (!tuned_nbuf) 576 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 577 maxbuf); 578 nbuf = maxbuf; 579 } 580 581 /* 582 * Ideal allocation size for the transient bio submap if 10% 583 * of the maximal space buffer map. This roughly corresponds 584 * to the amount of the buffer mapped for typical UFS load. 585 * 586 * Clip the buffer map to reserve space for the transient 587 * BIOs, if its extent is bigger than 90% (80% on i386) of the 588 * maximum buffer map extent on the platform. 589 * 590 * The fall-back to the maxbuf in case of maxbcache unset, 591 * allows to not trim the buffer KVA for the architectures 592 * with ample KVA space. 593 */ 594 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) { 595 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE; 596 buf_sz = (long)nbuf * BKVASIZE; 597 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM * 598 (TRANSIENT_DENOM - 1)) { 599 /* 600 * There is more KVA than memory. Do not 601 * adjust buffer map size, and assign the rest 602 * of maxbuf to transient map. 603 */ 604 biotmap_sz = maxbuf_sz - buf_sz; 605 } else { 606 /* 607 * Buffer map spans all KVA we could afford on 608 * this platform. Give 10% (20% on i386) of 609 * the buffer map to the transient bio map. 610 */ 611 biotmap_sz = buf_sz / TRANSIENT_DENOM; 612 buf_sz -= biotmap_sz; 613 } 614 if (biotmap_sz / INT_MAX > MAXPHYS) 615 bio_transient_maxcnt = INT_MAX; 616 else 617 bio_transient_maxcnt = biotmap_sz / MAXPHYS; 618 /* 619 * Artifically limit to 1024 simultaneous in-flight I/Os 620 * using the transient mapping. 621 */ 622 if (bio_transient_maxcnt > 1024) 623 bio_transient_maxcnt = 1024; 624 if (tuned_nbuf) 625 nbuf = buf_sz / BKVASIZE; 626 } 627 628 /* 629 * swbufs are used as temporary holders for I/O, such as paging I/O. 630 * We have no less then 16 and no more then 256. 631 */ 632 nswbuf = max(min(nbuf/4, 256), 16); 633#ifdef NSWBUF_MIN 634 if (nswbuf < NSWBUF_MIN) 635 nswbuf = NSWBUF_MIN; 636#endif 637#ifdef DIRECTIO 638 ffs_rawread_setup(); 639#endif 640 641 /* 642 * Reserve space for the buffer cache buffers 643 */ 644 swbuf = (void *)v; 645 v = (caddr_t)(swbuf + nswbuf); 646 buf = (void *)v; 647 v = (caddr_t)(buf + nbuf); 648 649 return(v); 650} 651 652/* Initialize the buffer subsystem. Called before use of any buffers. */ 653void 654bufinit(void) 655{ 656 struct buf *bp; 657 int i; 658 659 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 660 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 661 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 662 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 663 664 /* next, make a null set of free lists */ 665 for (i = 0; i < BUFFER_QUEUES; i++) 666 TAILQ_INIT(&bufqueues[i]); 667 668 /* finally, initialize each buffer header and stick on empty q */ 669 for (i = 0; i < nbuf; i++) { 670 bp = &buf[i]; 671 bzero(bp, sizeof *bp); 672 bp->b_flags = B_INVAL | B_INFREECNT; 673 bp->b_rcred = NOCRED; 674 bp->b_wcred = NOCRED; 675 bp->b_qindex = QUEUE_EMPTY; 676 bp->b_xflags = 0; 677 LIST_INIT(&bp->b_dep); 678 BUF_LOCKINIT(bp); 679 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 680#ifdef INVARIANTS 681 bq_len[QUEUE_EMPTY]++; 682#endif 683 } 684 685 /* 686 * maxbufspace is the absolute maximum amount of buffer space we are 687 * allowed to reserve in KVM and in real terms. The absolute maximum 688 * is nominally used by buf_daemon. hibufspace is the nominal maximum 689 * used by most other processes. The differential is required to 690 * ensure that buf_daemon is able to run when other processes might 691 * be blocked waiting for buffer space. 692 * 693 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 694 * this may result in KVM fragmentation which is not handled optimally 695 * by the system. 696 */ 697 maxbufspace = (long)nbuf * BKVASIZE; 698 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 699 lobufspace = hibufspace - MAXBSIZE; 700 701 /* 702 * Note: The 16 MiB upper limit for hirunningspace was chosen 703 * arbitrarily and may need further tuning. It corresponds to 704 * 128 outstanding write IO requests (if IO size is 128 KiB), 705 * which fits with many RAID controllers' tagged queuing limits. 706 * The lower 1 MiB limit is the historical upper limit for 707 * hirunningspace. 708 */ 709 hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE), 710 16 * 1024 * 1024), 1024 * 1024); 711 lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE); 712 713/* 714 * Limit the amount of malloc memory since it is wired permanently into 715 * the kernel space. Even though this is accounted for in the buffer 716 * allocation, we don't want the malloced region to grow uncontrolled. 717 * The malloc scheme improves memory utilization significantly on average 718 * (small) directories. 719 */ 720 maxbufmallocspace = hibufspace / 20; 721 722/* 723 * Reduce the chance of a deadlock occuring by limiting the number 724 * of delayed-write dirty buffers we allow to stack up. 725 */ 726 hidirtybuffers = nbuf / 4 + 20; 727 dirtybufthresh = hidirtybuffers * 9 / 10; 728 numdirtybuffers = 0; 729/* 730 * To support extreme low-memory systems, make sure hidirtybuffers cannot 731 * eat up all available buffer space. This occurs when our minimum cannot 732 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 733 * BKVASIZE'd buffers. 734 */ 735 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 736 hidirtybuffers >>= 1; 737 } 738 lodirtybuffers = hidirtybuffers / 2; 739 740/* 741 * Try to keep the number of free buffers in the specified range, 742 * and give special processes (e.g. like buf_daemon) access to an 743 * emergency reserve. 744 */ 745 lofreebuffers = nbuf / 18 + 5; 746 hifreebuffers = 2 * lofreebuffers; 747 numfreebuffers = nbuf; 748 749 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 750 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 751 unmapped_buf = (caddr_t)kmem_alloc_nofault(kernel_map, MAXPHYS); 752} 753 754#ifdef INVARIANTS 755static inline void 756vfs_buf_check_mapped(struct buf *bp) 757{ 758 759 KASSERT((bp->b_flags & B_UNMAPPED) == 0, 760 ("mapped buf %p %x", bp, bp->b_flags)); 761 KASSERT(bp->b_kvabase != unmapped_buf, 762 ("mapped buf: b_kvabase was not updated %p", bp)); 763 KASSERT(bp->b_data != unmapped_buf, 764 ("mapped buf: b_data was not updated %p", bp)); 765} 766 767static inline void 768vfs_buf_check_unmapped(struct buf *bp) 769{ 770 771 KASSERT((bp->b_flags & B_UNMAPPED) == B_UNMAPPED, 772 ("unmapped buf %p %x", bp, bp->b_flags)); 773 KASSERT(bp->b_kvabase == unmapped_buf, 774 ("unmapped buf: corrupted b_kvabase %p", bp)); 775 KASSERT(bp->b_data == unmapped_buf, 776 ("unmapped buf: corrupted b_data %p", bp)); 777} 778 779#define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) 780#define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) 781#else 782#define BUF_CHECK_MAPPED(bp) do {} while (0) 783#define BUF_CHECK_UNMAPPED(bp) do {} while (0) 784#endif 785 786static void 787bpmap_qenter(struct buf *bp) 788{ 789 790 BUF_CHECK_MAPPED(bp); 791 792 /* 793 * bp->b_data is relative to bp->b_offset, but 794 * bp->b_offset may be offset into the first page. 795 */ 796 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 797 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 798 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 799 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 800} 801 802/* 803 * bfreekva() - free the kva allocation for a buffer. 804 * 805 * Since this call frees up buffer space, we call bufspacewakeup(). 806 */ 807static void 808bfreekva(struct buf *bp) 809{ 810 811 if (bp->b_kvasize == 0) 812 return; 813 814 atomic_add_int(&buffreekvacnt, 1); 815 atomic_subtract_long(&bufspace, bp->b_kvasize); 816 if ((bp->b_flags & B_UNMAPPED) == 0) { 817 BUF_CHECK_MAPPED(bp); 818 vm_map_remove(buffer_map, (vm_offset_t)bp->b_kvabase, 819 (vm_offset_t)bp->b_kvabase + bp->b_kvasize); 820 } else { 821 BUF_CHECK_UNMAPPED(bp); 822 if ((bp->b_flags & B_KVAALLOC) != 0) { 823 vm_map_remove(buffer_map, (vm_offset_t)bp->b_kvaalloc, 824 (vm_offset_t)bp->b_kvaalloc + bp->b_kvasize); 825 } 826 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize); 827 bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC); 828 } 829 bp->b_kvasize = 0; 830 bufspacewakeup(); 831} 832 833/* 834 * bremfree: 835 * 836 * Mark the buffer for removal from the appropriate free list in brelse. 837 * 838 */ 839void 840bremfree(struct buf *bp) 841{ 842 int old; 843 844 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 845 KASSERT((bp->b_flags & B_REMFREE) == 0, 846 ("bremfree: buffer %p already marked for delayed removal.", bp)); 847 KASSERT(bp->b_qindex != QUEUE_NONE, 848 ("bremfree: buffer %p not on a queue.", bp)); 849 BUF_ASSERT_XLOCKED(bp); 850 851 bp->b_flags |= B_REMFREE; 852 /* Fixup numfreebuffers count. */ 853 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 854 KASSERT((bp->b_flags & B_INFREECNT) != 0, 855 ("buf %p not counted in numfreebuffers", bp)); 856 bp->b_flags &= ~B_INFREECNT; 857 old = atomic_fetchadd_int(&numfreebuffers, -1); 858 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 859 } 860} 861 862/* 863 * bremfreef: 864 * 865 * Force an immediate removal from a free list. Used only in nfs when 866 * it abuses the b_freelist pointer. 867 */ 868void 869bremfreef(struct buf *bp) 870{ 871 mtx_lock(&bqlock); 872 bremfreel(bp); 873 mtx_unlock(&bqlock); 874} 875 876/* 877 * bremfreel: 878 * 879 * Removes a buffer from the free list, must be called with the 880 * bqlock held. 881 */ 882static void 883bremfreel(struct buf *bp) 884{ 885 int old; 886 887 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 888 bp, bp->b_vp, bp->b_flags); 889 KASSERT(bp->b_qindex != QUEUE_NONE, 890 ("bremfreel: buffer %p not on a queue.", bp)); 891 BUF_ASSERT_XLOCKED(bp); 892 mtx_assert(&bqlock, MA_OWNED); 893 894 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 895#ifdef INVARIANTS 896 KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow", 897 bp->b_qindex)); 898 bq_len[bp->b_qindex]--; 899#endif 900 bp->b_qindex = QUEUE_NONE; 901 /* 902 * If this was a delayed bremfree() we only need to remove the buffer 903 * from the queue and return the stats are already done. 904 */ 905 if (bp->b_flags & B_REMFREE) { 906 bp->b_flags &= ~B_REMFREE; 907 return; 908 } 909 /* 910 * Fixup numfreebuffers count. If the buffer is invalid or not 911 * delayed-write, the buffer was free and we must decrement 912 * numfreebuffers. 913 */ 914 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 915 KASSERT((bp->b_flags & B_INFREECNT) != 0, 916 ("buf %p not counted in numfreebuffers", bp)); 917 bp->b_flags &= ~B_INFREECNT; 918 old = atomic_fetchadd_int(&numfreebuffers, -1); 919 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 920 } 921} 922 923/* 924 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 925 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 926 * the buffer is valid and we do not have to do anything. 927 */ 928void 929breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 930 int cnt, struct ucred * cred) 931{ 932 struct buf *rabp; 933 int i; 934 935 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 936 if (inmem(vp, *rablkno)) 937 continue; 938 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 939 940 if ((rabp->b_flags & B_CACHE) == 0) { 941 if (!TD_IS_IDLETHREAD(curthread)) 942 curthread->td_ru.ru_inblock++; 943 rabp->b_flags |= B_ASYNC; 944 rabp->b_flags &= ~B_INVAL; 945 rabp->b_ioflags &= ~BIO_ERROR; 946 rabp->b_iocmd = BIO_READ; 947 if (rabp->b_rcred == NOCRED && cred != NOCRED) 948 rabp->b_rcred = crhold(cred); 949 vfs_busy_pages(rabp, 0); 950 BUF_KERNPROC(rabp); 951 rabp->b_iooffset = dbtob(rabp->b_blkno); 952 bstrategy(rabp); 953 } else { 954 brelse(rabp); 955 } 956 } 957} 958 959/* 960 * Entry point for bread() and breadn() via #defines in sys/buf.h. 961 * 962 * Get a buffer with the specified data. Look in the cache first. We 963 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 964 * is set, the buffer is valid and we do not have to do anything, see 965 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 966 */ 967int 968breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno, 969 int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp) 970{ 971 struct buf *bp; 972 int rv = 0, readwait = 0; 973 974 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 975 /* 976 * Can only return NULL if GB_LOCK_NOWAIT flag is specified. 977 */ 978 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags); 979 if (bp == NULL) 980 return (EBUSY); 981 982 /* if not found in cache, do some I/O */ 983 if ((bp->b_flags & B_CACHE) == 0) { 984 if (!TD_IS_IDLETHREAD(curthread)) 985 curthread->td_ru.ru_inblock++; 986 bp->b_iocmd = BIO_READ; 987 bp->b_flags &= ~B_INVAL; 988 bp->b_ioflags &= ~BIO_ERROR; 989 if (bp->b_rcred == NOCRED && cred != NOCRED) 990 bp->b_rcred = crhold(cred); 991 vfs_busy_pages(bp, 0); 992 bp->b_iooffset = dbtob(bp->b_blkno); 993 bstrategy(bp); 994 ++readwait; 995 } 996 997 breada(vp, rablkno, rabsize, cnt, cred); 998 999 if (readwait) { 1000 rv = bufwait(bp); 1001 } 1002 return (rv); 1003} 1004 1005/* 1006 * Write, release buffer on completion. (Done by iodone 1007 * if async). Do not bother writing anything if the buffer 1008 * is invalid. 1009 * 1010 * Note that we set B_CACHE here, indicating that buffer is 1011 * fully valid and thus cacheable. This is true even of NFS 1012 * now so we set it generally. This could be set either here 1013 * or in biodone() since the I/O is synchronous. We put it 1014 * here. 1015 */ 1016int 1017bufwrite(struct buf *bp) 1018{ 1019 int oldflags; 1020 struct vnode *vp; 1021 int vp_md; 1022 1023 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1024 if (bp->b_flags & B_INVAL) { 1025 brelse(bp); 1026 return (0); 1027 } 1028 1029 if (bp->b_flags & B_BARRIER) 1030 barrierwrites++; 1031 1032 oldflags = bp->b_flags; 1033 1034 BUF_ASSERT_HELD(bp); 1035 1036 if (bp->b_pin_count > 0) 1037 bunpin_wait(bp); 1038 1039 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 1040 ("FFS background buffer should not get here %p", bp)); 1041 1042 vp = bp->b_vp; 1043 if (vp) 1044 vp_md = vp->v_vflag & VV_MD; 1045 else 1046 vp_md = 0; 1047 1048 /* 1049 * Mark the buffer clean. Increment the bufobj write count 1050 * before bundirty() call, to prevent other thread from seeing 1051 * empty dirty list and zero counter for writes in progress, 1052 * falsely indicating that the bufobj is clean. 1053 */ 1054 bufobj_wref(bp->b_bufobj); 1055 bundirty(bp); 1056 1057 bp->b_flags &= ~B_DONE; 1058 bp->b_ioflags &= ~BIO_ERROR; 1059 bp->b_flags |= B_CACHE; 1060 bp->b_iocmd = BIO_WRITE; 1061 1062 vfs_busy_pages(bp, 1); 1063 1064 /* 1065 * Normal bwrites pipeline writes 1066 */ 1067 bp->b_runningbufspace = bp->b_bufsize; 1068 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 1069 1070 if (!TD_IS_IDLETHREAD(curthread)) 1071 curthread->td_ru.ru_oublock++; 1072 if (oldflags & B_ASYNC) 1073 BUF_KERNPROC(bp); 1074 bp->b_iooffset = dbtob(bp->b_blkno); 1075 bstrategy(bp); 1076 1077 if ((oldflags & B_ASYNC) == 0) { 1078 int rtval = bufwait(bp); 1079 brelse(bp); 1080 return (rtval); 1081 } else { 1082 /* 1083 * don't allow the async write to saturate the I/O 1084 * system. We will not deadlock here because 1085 * we are blocking waiting for I/O that is already in-progress 1086 * to complete. We do not block here if it is the update 1087 * or syncer daemon trying to clean up as that can lead 1088 * to deadlock. 1089 */ 1090 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 1091 waitrunningbufspace(); 1092 } 1093 1094 return (0); 1095} 1096 1097void 1098bufbdflush(struct bufobj *bo, struct buf *bp) 1099{ 1100 struct buf *nbp; 1101 1102 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 1103 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 1104 altbufferflushes++; 1105 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 1106 BO_LOCK(bo); 1107 /* 1108 * Try to find a buffer to flush. 1109 */ 1110 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 1111 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1112 BUF_LOCK(nbp, 1113 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1114 continue; 1115 if (bp == nbp) 1116 panic("bdwrite: found ourselves"); 1117 BO_UNLOCK(bo); 1118 /* Don't countdeps with the bo lock held. */ 1119 if (buf_countdeps(nbp, 0)) { 1120 BO_LOCK(bo); 1121 BUF_UNLOCK(nbp); 1122 continue; 1123 } 1124 if (nbp->b_flags & B_CLUSTEROK) { 1125 vfs_bio_awrite(nbp); 1126 } else { 1127 bremfree(nbp); 1128 bawrite(nbp); 1129 } 1130 dirtybufferflushes++; 1131 break; 1132 } 1133 if (nbp == NULL) 1134 BO_UNLOCK(bo); 1135 } 1136} 1137 1138/* 1139 * Delayed write. (Buffer is marked dirty). Do not bother writing 1140 * anything if the buffer is marked invalid. 1141 * 1142 * Note that since the buffer must be completely valid, we can safely 1143 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1144 * biodone() in order to prevent getblk from writing the buffer 1145 * out synchronously. 1146 */ 1147void 1148bdwrite(struct buf *bp) 1149{ 1150 struct thread *td = curthread; 1151 struct vnode *vp; 1152 struct bufobj *bo; 1153 1154 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1155 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1156 KASSERT((bp->b_flags & B_BARRIER) == 0, 1157 ("Barrier request in delayed write %p", bp)); 1158 BUF_ASSERT_HELD(bp); 1159 1160 if (bp->b_flags & B_INVAL) { 1161 brelse(bp); 1162 return; 1163 } 1164 1165 /* 1166 * If we have too many dirty buffers, don't create any more. 1167 * If we are wildly over our limit, then force a complete 1168 * cleanup. Otherwise, just keep the situation from getting 1169 * out of control. Note that we have to avoid a recursive 1170 * disaster and not try to clean up after our own cleanup! 1171 */ 1172 vp = bp->b_vp; 1173 bo = bp->b_bufobj; 1174 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 1175 td->td_pflags |= TDP_INBDFLUSH; 1176 BO_BDFLUSH(bo, bp); 1177 td->td_pflags &= ~TDP_INBDFLUSH; 1178 } else 1179 recursiveflushes++; 1180 1181 bdirty(bp); 1182 /* 1183 * Set B_CACHE, indicating that the buffer is fully valid. This is 1184 * true even of NFS now. 1185 */ 1186 bp->b_flags |= B_CACHE; 1187 1188 /* 1189 * This bmap keeps the system from needing to do the bmap later, 1190 * perhaps when the system is attempting to do a sync. Since it 1191 * is likely that the indirect block -- or whatever other datastructure 1192 * that the filesystem needs is still in memory now, it is a good 1193 * thing to do this. Note also, that if the pageout daemon is 1194 * requesting a sync -- there might not be enough memory to do 1195 * the bmap then... So, this is important to do. 1196 */ 1197 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1198 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1199 } 1200 1201 /* 1202 * Set the *dirty* buffer range based upon the VM system dirty 1203 * pages. 1204 * 1205 * Mark the buffer pages as clean. We need to do this here to 1206 * satisfy the vnode_pager and the pageout daemon, so that it 1207 * thinks that the pages have been "cleaned". Note that since 1208 * the pages are in a delayed write buffer -- the VFS layer 1209 * "will" see that the pages get written out on the next sync, 1210 * or perhaps the cluster will be completed. 1211 */ 1212 vfs_clean_pages_dirty_buf(bp); 1213 bqrelse(bp); 1214 1215 /* 1216 * Wakeup the buffer flushing daemon if we have a lot of dirty 1217 * buffers (midpoint between our recovery point and our stall 1218 * point). 1219 */ 1220 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1221 1222 /* 1223 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1224 * due to the softdep code. 1225 */ 1226} 1227 1228/* 1229 * bdirty: 1230 * 1231 * Turn buffer into delayed write request. We must clear BIO_READ and 1232 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1233 * itself to properly update it in the dirty/clean lists. We mark it 1234 * B_DONE to ensure that any asynchronization of the buffer properly 1235 * clears B_DONE ( else a panic will occur later ). 1236 * 1237 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1238 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1239 * should only be called if the buffer is known-good. 1240 * 1241 * Since the buffer is not on a queue, we do not update the numfreebuffers 1242 * count. 1243 * 1244 * The buffer must be on QUEUE_NONE. 1245 */ 1246void 1247bdirty(struct buf *bp) 1248{ 1249 1250 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1251 bp, bp->b_vp, bp->b_flags); 1252 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1253 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1254 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1255 BUF_ASSERT_HELD(bp); 1256 bp->b_flags &= ~(B_RELBUF); 1257 bp->b_iocmd = BIO_WRITE; 1258 1259 if ((bp->b_flags & B_DELWRI) == 0) { 1260 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1261 reassignbuf(bp); 1262 atomic_add_int(&numdirtybuffers, 1); 1263 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1264 } 1265} 1266 1267/* 1268 * bundirty: 1269 * 1270 * Clear B_DELWRI for buffer. 1271 * 1272 * Since the buffer is not on a queue, we do not update the numfreebuffers 1273 * count. 1274 * 1275 * The buffer must be on QUEUE_NONE. 1276 */ 1277 1278void 1279bundirty(struct buf *bp) 1280{ 1281 1282 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1283 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1284 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1285 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1286 BUF_ASSERT_HELD(bp); 1287 1288 if (bp->b_flags & B_DELWRI) { 1289 bp->b_flags &= ~B_DELWRI; 1290 reassignbuf(bp); 1291 atomic_subtract_int(&numdirtybuffers, 1); 1292 numdirtywakeup(lodirtybuffers); 1293 } 1294 /* 1295 * Since it is now being written, we can clear its deferred write flag. 1296 */ 1297 bp->b_flags &= ~B_DEFERRED; 1298} 1299 1300/* 1301 * bawrite: 1302 * 1303 * Asynchronous write. Start output on a buffer, but do not wait for 1304 * it to complete. The buffer is released when the output completes. 1305 * 1306 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1307 * B_INVAL buffers. Not us. 1308 */ 1309void 1310bawrite(struct buf *bp) 1311{ 1312 1313 bp->b_flags |= B_ASYNC; 1314 (void) bwrite(bp); 1315} 1316 1317/* 1318 * babarrierwrite: 1319 * 1320 * Asynchronous barrier write. Start output on a buffer, but do not 1321 * wait for it to complete. Place a write barrier after this write so 1322 * that this buffer and all buffers written before it are committed to 1323 * the disk before any buffers written after this write are committed 1324 * to the disk. The buffer is released when the output completes. 1325 */ 1326void 1327babarrierwrite(struct buf *bp) 1328{ 1329 1330 bp->b_flags |= B_ASYNC | B_BARRIER; 1331 (void) bwrite(bp); 1332} 1333 1334/* 1335 * bbarrierwrite: 1336 * 1337 * Synchronous barrier write. Start output on a buffer and wait for 1338 * it to complete. Place a write barrier after this write so that 1339 * this buffer and all buffers written before it are committed to 1340 * the disk before any buffers written after this write are committed 1341 * to the disk. The buffer is released when the output completes. 1342 */ 1343int 1344bbarrierwrite(struct buf *bp) 1345{ 1346 1347 bp->b_flags |= B_BARRIER; 1348 return (bwrite(bp)); 1349} 1350 1351/* 1352 * bwillwrite: 1353 * 1354 * Called prior to the locking of any vnodes when we are expecting to 1355 * write. We do not want to starve the buffer cache with too many 1356 * dirty buffers so we block here. By blocking prior to the locking 1357 * of any vnodes we attempt to avoid the situation where a locked vnode 1358 * prevents the various system daemons from flushing related buffers. 1359 */ 1360 1361void 1362bwillwrite(void) 1363{ 1364 1365 if (numdirtybuffers >= hidirtybuffers) { 1366 mtx_lock(&nblock); 1367 while (numdirtybuffers >= hidirtybuffers) { 1368 bd_wakeup(1); 1369 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1370 msleep(&needsbuffer, &nblock, 1371 (PRIBIO + 4), "flswai", 0); 1372 } 1373 mtx_unlock(&nblock); 1374 } 1375} 1376 1377/* 1378 * Return true if we have too many dirty buffers. 1379 */ 1380int 1381buf_dirty_count_severe(void) 1382{ 1383 1384 return(numdirtybuffers >= hidirtybuffers); 1385} 1386 1387static __noinline int 1388buf_vm_page_count_severe(void) 1389{ 1390 1391 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1); 1392 1393 return vm_page_count_severe(); 1394} 1395 1396/* 1397 * brelse: 1398 * 1399 * Release a busy buffer and, if requested, free its resources. The 1400 * buffer will be stashed in the appropriate bufqueue[] allowing it 1401 * to be accessed later as a cache entity or reused for other purposes. 1402 */ 1403void 1404brelse(struct buf *bp) 1405{ 1406 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1407 bp, bp->b_vp, bp->b_flags); 1408 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1409 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1410 1411 if (BUF_LOCKRECURSED(bp)) { 1412 /* 1413 * Do not process, in particular, do not handle the 1414 * B_INVAL/B_RELBUF and do not release to free list. 1415 */ 1416 BUF_UNLOCK(bp); 1417 return; 1418 } 1419 1420 if (bp->b_flags & B_MANAGED) { 1421 bqrelse(bp); 1422 return; 1423 } 1424 1425 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1426 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1427 /* 1428 * Failed write, redirty. Must clear BIO_ERROR to prevent 1429 * pages from being scrapped. If the error is anything 1430 * other than an I/O error (EIO), assume that retrying 1431 * is futile. 1432 */ 1433 bp->b_ioflags &= ~BIO_ERROR; 1434 bdirty(bp); 1435 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1436 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1437 /* 1438 * Either a failed I/O or we were asked to free or not 1439 * cache the buffer. 1440 */ 1441 bp->b_flags |= B_INVAL; 1442 if (!LIST_EMPTY(&bp->b_dep)) 1443 buf_deallocate(bp); 1444 if (bp->b_flags & B_DELWRI) { 1445 atomic_subtract_int(&numdirtybuffers, 1); 1446 numdirtywakeup(lodirtybuffers); 1447 } 1448 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1449 if ((bp->b_flags & B_VMIO) == 0) { 1450 if (bp->b_bufsize) 1451 allocbuf(bp, 0); 1452 if (bp->b_vp) 1453 brelvp(bp); 1454 } 1455 } 1456 1457 /* 1458 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1459 * is called with B_DELWRI set, the underlying pages may wind up 1460 * getting freed causing a previous write (bdwrite()) to get 'lost' 1461 * because pages associated with a B_DELWRI bp are marked clean. 1462 * 1463 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1464 * if B_DELWRI is set. 1465 * 1466 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1467 * on pages to return pages to the VM page queues. 1468 */ 1469 if (bp->b_flags & B_DELWRI) 1470 bp->b_flags &= ~B_RELBUF; 1471 else if (buf_vm_page_count_severe()) { 1472 /* 1473 * BKGRDINPROG can only be set with the buf and bufobj 1474 * locks both held. We tolerate a race to clear it here. 1475 */ 1476 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1477 bp->b_flags |= B_RELBUF; 1478 } 1479 1480 /* 1481 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1482 * constituted, not even NFS buffers now. Two flags effect this. If 1483 * B_INVAL, the struct buf is invalidated but the VM object is kept 1484 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1485 * 1486 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1487 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1488 * buffer is also B_INVAL because it hits the re-dirtying code above. 1489 * 1490 * Normally we can do this whether a buffer is B_DELWRI or not. If 1491 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1492 * the commit state and we cannot afford to lose the buffer. If the 1493 * buffer has a background write in progress, we need to keep it 1494 * around to prevent it from being reconstituted and starting a second 1495 * background write. 1496 */ 1497 if ((bp->b_flags & B_VMIO) 1498 && !(bp->b_vp->v_mount != NULL && 1499 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1500 !vn_isdisk(bp->b_vp, NULL) && 1501 (bp->b_flags & B_DELWRI)) 1502 ) { 1503 1504 int i, j, resid; 1505 vm_page_t m; 1506 off_t foff; 1507 vm_pindex_t poff; 1508 vm_object_t obj; 1509 1510 obj = bp->b_bufobj->bo_object; 1511 1512 /* 1513 * Get the base offset and length of the buffer. Note that 1514 * in the VMIO case if the buffer block size is not 1515 * page-aligned then b_data pointer may not be page-aligned. 1516 * But our b_pages[] array *IS* page aligned. 1517 * 1518 * block sizes less then DEV_BSIZE (usually 512) are not 1519 * supported due to the page granularity bits (m->valid, 1520 * m->dirty, etc...). 1521 * 1522 * See man buf(9) for more information 1523 */ 1524 resid = bp->b_bufsize; 1525 foff = bp->b_offset; 1526 for (i = 0; i < bp->b_npages; i++) { 1527 int had_bogus = 0; 1528 1529 m = bp->b_pages[i]; 1530 1531 /* 1532 * If we hit a bogus page, fixup *all* the bogus pages 1533 * now. 1534 */ 1535 if (m == bogus_page) { 1536 poff = OFF_TO_IDX(bp->b_offset); 1537 had_bogus = 1; 1538 1539 VM_OBJECT_RLOCK(obj); 1540 for (j = i; j < bp->b_npages; j++) { 1541 vm_page_t mtmp; 1542 mtmp = bp->b_pages[j]; 1543 if (mtmp == bogus_page) { 1544 mtmp = vm_page_lookup(obj, poff + j); 1545 if (!mtmp) { 1546 panic("brelse: page missing\n"); 1547 } 1548 bp->b_pages[j] = mtmp; 1549 } 1550 } 1551 VM_OBJECT_RUNLOCK(obj); 1552 1553 if ((bp->b_flags & (B_INVAL | B_UNMAPPED)) == 0) { 1554 BUF_CHECK_MAPPED(bp); 1555 pmap_qenter( 1556 trunc_page((vm_offset_t)bp->b_data), 1557 bp->b_pages, bp->b_npages); 1558 } 1559 m = bp->b_pages[i]; 1560 } 1561 if ((bp->b_flags & B_NOCACHE) || 1562 (bp->b_ioflags & BIO_ERROR && 1563 bp->b_iocmd == BIO_READ)) { 1564 int poffset = foff & PAGE_MASK; 1565 int presid = resid > (PAGE_SIZE - poffset) ? 1566 (PAGE_SIZE - poffset) : resid; 1567 1568 KASSERT(presid >= 0, ("brelse: extra page")); 1569 VM_OBJECT_WLOCK(obj); 1570 vm_page_set_invalid(m, poffset, presid); 1571 VM_OBJECT_WUNLOCK(obj); 1572 if (had_bogus) 1573 printf("avoided corruption bug in bogus_page/brelse code\n"); 1574 } 1575 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1576 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1577 } 1578 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1579 vfs_vmio_release(bp); 1580 1581 } else if (bp->b_flags & B_VMIO) { 1582 1583 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1584 vfs_vmio_release(bp); 1585 } 1586 1587 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1588 if (bp->b_bufsize != 0) 1589 allocbuf(bp, 0); 1590 if (bp->b_vp != NULL) 1591 brelvp(bp); 1592 } 1593 1594 /* enqueue */ 1595 mtx_lock(&bqlock); 1596 /* Handle delayed bremfree() processing. */ 1597 if (bp->b_flags & B_REMFREE) 1598 bremfreel(bp); 1599 1600 if (bp->b_qindex != QUEUE_NONE) 1601 panic("brelse: free buffer onto another queue???"); 1602 1603 /* 1604 * If the buffer has junk contents signal it and eventually 1605 * clean up B_DELWRI and diassociate the vnode so that gbincore() 1606 * doesn't find it. 1607 */ 1608 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1609 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1610 bp->b_flags |= B_INVAL; 1611 if (bp->b_flags & B_INVAL) { 1612 if (bp->b_flags & B_DELWRI) 1613 bundirty(bp); 1614 if (bp->b_vp) 1615 brelvp(bp); 1616 } 1617 1618 /* buffers with no memory */ 1619 if (bp->b_bufsize == 0) { 1620 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1621 if (bp->b_vflags & BV_BKGRDINPROG) 1622 panic("losing buffer 1"); 1623 if (bp->b_kvasize) { 1624 bp->b_qindex = QUEUE_EMPTYKVA; 1625 } else { 1626 bp->b_qindex = QUEUE_EMPTY; 1627 } 1628 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1629 /* buffers with junk contents */ 1630 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1631 (bp->b_ioflags & BIO_ERROR)) { 1632 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1633 if (bp->b_vflags & BV_BKGRDINPROG) 1634 panic("losing buffer 2"); 1635 bp->b_qindex = QUEUE_CLEAN; 1636 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1637 /* remaining buffers */ 1638 } else { 1639 if (bp->b_flags & B_DELWRI) 1640 bp->b_qindex = QUEUE_DIRTY; 1641 else 1642 bp->b_qindex = QUEUE_CLEAN; 1643 if (bp->b_flags & B_AGE) { 1644 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, 1645 b_freelist); 1646 } else { 1647 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, 1648 b_freelist); 1649 } 1650 } 1651#ifdef INVARIANTS 1652 bq_len[bp->b_qindex]++; 1653#endif 1654 mtx_unlock(&bqlock); 1655 1656 /* 1657 * Fixup numfreebuffers count. The bp is on an appropriate queue 1658 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1659 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1660 * if B_INVAL is set ). 1661 */ 1662 1663 if (!(bp->b_flags & B_DELWRI)) 1664 bufcountwakeup(bp); 1665 1666 /* 1667 * Something we can maybe free or reuse 1668 */ 1669 if (bp->b_bufsize || bp->b_kvasize) 1670 bufspacewakeup(); 1671 1672 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1673 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1674 panic("brelse: not dirty"); 1675 /* unlock */ 1676 BUF_UNLOCK(bp); 1677} 1678 1679/* 1680 * Release a buffer back to the appropriate queue but do not try to free 1681 * it. The buffer is expected to be used again soon. 1682 * 1683 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1684 * biodone() to requeue an async I/O on completion. It is also used when 1685 * known good buffers need to be requeued but we think we may need the data 1686 * again soon. 1687 * 1688 * XXX we should be able to leave the B_RELBUF hint set on completion. 1689 */ 1690void 1691bqrelse(struct buf *bp) 1692{ 1693 struct bufobj *bo; 1694 1695 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1696 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1697 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1698 1699 if (BUF_LOCKRECURSED(bp)) { 1700 /* do not release to free list */ 1701 BUF_UNLOCK(bp); 1702 return; 1703 } 1704 1705 bo = bp->b_bufobj; 1706 if (bp->b_flags & B_MANAGED) { 1707 if (bp->b_flags & B_REMFREE) { 1708 mtx_lock(&bqlock); 1709 bremfreel(bp); 1710 mtx_unlock(&bqlock); 1711 } 1712 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1713 BUF_UNLOCK(bp); 1714 return; 1715 } 1716 1717 mtx_lock(&bqlock); 1718 /* Handle delayed bremfree() processing. */ 1719 if (bp->b_flags & B_REMFREE) 1720 bremfreel(bp); 1721 1722 if (bp->b_qindex != QUEUE_NONE) 1723 panic("bqrelse: free buffer onto another queue???"); 1724 /* buffers with stale but valid contents */ 1725 if (bp->b_flags & B_DELWRI) { 1726 bp->b_qindex = QUEUE_DIRTY; 1727 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1728#ifdef INVARIANTS 1729 bq_len[bp->b_qindex]++; 1730#endif 1731 } else { 1732 /* 1733 * BKGRDINPROG can only be set with the buf and bufobj 1734 * locks both held. We tolerate a race to clear it here. 1735 */ 1736 if (!buf_vm_page_count_severe() || 1737 (bp->b_vflags & BV_BKGRDINPROG)) { 1738 bp->b_qindex = QUEUE_CLEAN; 1739 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1740 b_freelist); 1741#ifdef INVARIANTS 1742 bq_len[QUEUE_CLEAN]++; 1743#endif 1744 } else { 1745 /* 1746 * We are too low on memory, we have to try to free 1747 * the buffer (most importantly: the wired pages 1748 * making up its backing store) *now*. 1749 */ 1750 mtx_unlock(&bqlock); 1751 brelse(bp); 1752 return; 1753 } 1754 } 1755 mtx_unlock(&bqlock); 1756 1757 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1758 bufcountwakeup(bp); 1759 1760 /* 1761 * Something we can maybe free or reuse. 1762 */ 1763 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1764 bufspacewakeup(); 1765 1766 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1767 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1768 panic("bqrelse: not dirty"); 1769 /* unlock */ 1770 BUF_UNLOCK(bp); 1771} 1772 1773/* Give pages used by the bp back to the VM system (where possible) */ 1774static void 1775vfs_vmio_release(struct buf *bp) 1776{ 1777 int i; 1778 vm_page_t m; 1779 1780 if ((bp->b_flags & B_UNMAPPED) == 0) { 1781 BUF_CHECK_MAPPED(bp); 1782 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 1783 } else 1784 BUF_CHECK_UNMAPPED(bp); 1785 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 1786 for (i = 0; i < bp->b_npages; i++) { 1787 m = bp->b_pages[i]; 1788 bp->b_pages[i] = NULL; 1789 /* 1790 * In order to keep page LRU ordering consistent, put 1791 * everything on the inactive queue. 1792 */ 1793 vm_page_lock(m); 1794 vm_page_unwire(m, 0); 1795 /* 1796 * We don't mess with busy pages, it is 1797 * the responsibility of the process that 1798 * busied the pages to deal with them. 1799 */ 1800 if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 && 1801 m->wire_count == 0) { 1802 /* 1803 * Might as well free the page if we can and it has 1804 * no valid data. We also free the page if the 1805 * buffer was used for direct I/O 1806 */ 1807 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) { 1808 vm_page_free(m); 1809 } else if (bp->b_flags & B_DIRECT) { 1810 vm_page_try_to_free(m); 1811 } else if (buf_vm_page_count_severe()) { 1812 vm_page_try_to_cache(m); 1813 } 1814 } 1815 vm_page_unlock(m); 1816 } 1817 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 1818 1819 if (bp->b_bufsize) { 1820 bufspacewakeup(); 1821 bp->b_bufsize = 0; 1822 } 1823 bp->b_npages = 0; 1824 bp->b_flags &= ~B_VMIO; 1825 if (bp->b_vp) 1826 brelvp(bp); 1827} 1828 1829/* 1830 * Check to see if a block at a particular lbn is available for a clustered 1831 * write. 1832 */ 1833static int 1834vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1835{ 1836 struct buf *bpa; 1837 int match; 1838 1839 match = 0; 1840 1841 /* If the buf isn't in core skip it */ 1842 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1843 return (0); 1844 1845 /* If the buf is busy we don't want to wait for it */ 1846 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1847 return (0); 1848 1849 /* Only cluster with valid clusterable delayed write buffers */ 1850 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1851 (B_DELWRI | B_CLUSTEROK)) 1852 goto done; 1853 1854 if (bpa->b_bufsize != size) 1855 goto done; 1856 1857 /* 1858 * Check to see if it is in the expected place on disk and that the 1859 * block has been mapped. 1860 */ 1861 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1862 match = 1; 1863done: 1864 BUF_UNLOCK(bpa); 1865 return (match); 1866} 1867 1868/* 1869 * vfs_bio_awrite: 1870 * 1871 * Implement clustered async writes for clearing out B_DELWRI buffers. 1872 * This is much better then the old way of writing only one buffer at 1873 * a time. Note that we may not be presented with the buffers in the 1874 * correct order, so we search for the cluster in both directions. 1875 */ 1876int 1877vfs_bio_awrite(struct buf *bp) 1878{ 1879 struct bufobj *bo; 1880 int i; 1881 int j; 1882 daddr_t lblkno = bp->b_lblkno; 1883 struct vnode *vp = bp->b_vp; 1884 int ncl; 1885 int nwritten; 1886 int size; 1887 int maxcl; 1888 int gbflags; 1889 1890 bo = &vp->v_bufobj; 1891 gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0; 1892 /* 1893 * right now we support clustered writing only to regular files. If 1894 * we find a clusterable block we could be in the middle of a cluster 1895 * rather then at the beginning. 1896 */ 1897 if ((vp->v_type == VREG) && 1898 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1899 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1900 1901 size = vp->v_mount->mnt_stat.f_iosize; 1902 maxcl = MAXPHYS / size; 1903 1904 BO_RLOCK(bo); 1905 for (i = 1; i < maxcl; i++) 1906 if (vfs_bio_clcheck(vp, size, lblkno + i, 1907 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1908 break; 1909 1910 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1911 if (vfs_bio_clcheck(vp, size, lblkno - j, 1912 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1913 break; 1914 BO_RUNLOCK(bo); 1915 --j; 1916 ncl = i + j; 1917 /* 1918 * this is a possible cluster write 1919 */ 1920 if (ncl != 1) { 1921 BUF_UNLOCK(bp); 1922 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 1923 gbflags); 1924 return (nwritten); 1925 } 1926 } 1927 bremfree(bp); 1928 bp->b_flags |= B_ASYNC; 1929 /* 1930 * default (old) behavior, writing out only one block 1931 * 1932 * XXX returns b_bufsize instead of b_bcount for nwritten? 1933 */ 1934 nwritten = bp->b_bufsize; 1935 (void) bwrite(bp); 1936 1937 return (nwritten); 1938} 1939 1940static void 1941setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags) 1942{ 1943 1944 KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 1945 bp->b_kvasize == 0, ("call bfreekva(%p)", bp)); 1946 if ((gbflags & GB_UNMAPPED) == 0) { 1947 bp->b_kvabase = (caddr_t)addr; 1948 } else if ((gbflags & GB_KVAALLOC) != 0) { 1949 KASSERT((gbflags & GB_UNMAPPED) != 0, 1950 ("GB_KVAALLOC without GB_UNMAPPED")); 1951 bp->b_kvaalloc = (caddr_t)addr; 1952 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 1953 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 1954 } 1955 bp->b_kvasize = maxsize; 1956} 1957 1958/* 1959 * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if 1960 * needed. 1961 */ 1962static int 1963allocbufkva(struct buf *bp, int maxsize, int gbflags) 1964{ 1965 vm_offset_t addr; 1966 int rv; 1967 1968 bfreekva(bp); 1969 addr = 0; 1970 1971 vm_map_lock(buffer_map); 1972 if (vm_map_findspace(buffer_map, vm_map_min(buffer_map), maxsize, 1973 &addr)) { 1974 vm_map_unlock(buffer_map); 1975 /* 1976 * Buffer map is too fragmented. Request the caller 1977 * to defragment the map. 1978 */ 1979 atomic_add_int(&bufdefragcnt, 1); 1980 return (1); 1981 } 1982 rv = vm_map_insert(buffer_map, NULL, 0, addr, addr + maxsize, 1983 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); 1984 KASSERT(rv == KERN_SUCCESS, ("vm_map_insert(buffer_map) rv %d", rv)); 1985 vm_map_unlock(buffer_map); 1986 setbufkva(bp, addr, maxsize, gbflags); 1987 atomic_add_long(&bufspace, bp->b_kvasize); 1988 return (0); 1989} 1990 1991/* 1992 * Ask the bufdaemon for help, or act as bufdaemon itself, when a 1993 * locked vnode is supplied. 1994 */ 1995static void 1996getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo, 1997 int defrag) 1998{ 1999 struct thread *td; 2000 char *waitmsg; 2001 int fl, flags, norunbuf; 2002 2003 mtx_assert(&bqlock, MA_OWNED); 2004 2005 if (defrag) { 2006 flags = VFS_BIO_NEED_BUFSPACE; 2007 waitmsg = "nbufkv"; 2008 } else if (bufspace >= hibufspace) { 2009 waitmsg = "nbufbs"; 2010 flags = VFS_BIO_NEED_BUFSPACE; 2011 } else { 2012 waitmsg = "newbuf"; 2013 flags = VFS_BIO_NEED_ANY; 2014 } 2015 mtx_lock(&nblock); 2016 needsbuffer |= flags; 2017 mtx_unlock(&nblock); 2018 mtx_unlock(&bqlock); 2019 2020 bd_speedup(); /* heeeelp */ 2021 if ((gbflags & GB_NOWAIT_BD) != 0) 2022 return; 2023 2024 td = curthread; 2025 mtx_lock(&nblock); 2026 while (needsbuffer & flags) { 2027 if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) { 2028 mtx_unlock(&nblock); 2029 /* 2030 * getblk() is called with a vnode locked, and 2031 * some majority of the dirty buffers may as 2032 * well belong to the vnode. Flushing the 2033 * buffers there would make a progress that 2034 * cannot be achieved by the buf_daemon, that 2035 * cannot lock the vnode. 2036 */ 2037 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2038 (td->td_pflags & TDP_NORUNNINGBUF); 2039 /* play bufdaemon */ 2040 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2041 fl = buf_do_flush(vp); 2042 td->td_pflags &= norunbuf; 2043 mtx_lock(&nblock); 2044 if (fl != 0) 2045 continue; 2046 if ((needsbuffer & flags) == 0) 2047 break; 2048 } 2049 if (msleep(&needsbuffer, &nblock, (PRIBIO + 4) | slpflag, 2050 waitmsg, slptimeo)) 2051 break; 2052 } 2053 mtx_unlock(&nblock); 2054} 2055 2056static void 2057getnewbuf_reuse_bp(struct buf *bp, int qindex) 2058{ 2059 2060 CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 2061 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 2062 bp->b_kvasize, bp->b_bufsize, qindex); 2063 mtx_assert(&bqlock, MA_NOTOWNED); 2064 2065 /* 2066 * Note: we no longer distinguish between VMIO and non-VMIO 2067 * buffers. 2068 */ 2069 KASSERT((bp->b_flags & B_DELWRI) == 0, 2070 ("delwri buffer %p found in queue %d", bp, qindex)); 2071 2072 if (qindex == QUEUE_CLEAN) { 2073 if (bp->b_flags & B_VMIO) { 2074 bp->b_flags &= ~B_ASYNC; 2075 vfs_vmio_release(bp); 2076 } 2077 if (bp->b_vp != NULL) 2078 brelvp(bp); 2079 } 2080 2081 /* 2082 * Get the rest of the buffer freed up. b_kva* is still valid 2083 * after this operation. 2084 */ 2085 2086 if (bp->b_rcred != NOCRED) { 2087 crfree(bp->b_rcred); 2088 bp->b_rcred = NOCRED; 2089 } 2090 if (bp->b_wcred != NOCRED) { 2091 crfree(bp->b_wcred); 2092 bp->b_wcred = NOCRED; 2093 } 2094 if (!LIST_EMPTY(&bp->b_dep)) 2095 buf_deallocate(bp); 2096 if (bp->b_vflags & BV_BKGRDINPROG) 2097 panic("losing buffer 3"); 2098 KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p. qindex: %d", 2099 bp, bp->b_vp, qindex)); 2100 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 2101 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 2102 2103 if (bp->b_bufsize) 2104 allocbuf(bp, 0); 2105 2106 bp->b_flags &= B_UNMAPPED | B_KVAALLOC; 2107 bp->b_ioflags = 0; 2108 bp->b_xflags = 0; 2109 KASSERT((bp->b_flags & B_INFREECNT) == 0, 2110 ("buf %p still counted as free?", bp)); 2111 bp->b_vflags = 0; 2112 bp->b_vp = NULL; 2113 bp->b_blkno = bp->b_lblkno = 0; 2114 bp->b_offset = NOOFFSET; 2115 bp->b_iodone = 0; 2116 bp->b_error = 0; 2117 bp->b_resid = 0; 2118 bp->b_bcount = 0; 2119 bp->b_npages = 0; 2120 bp->b_dirtyoff = bp->b_dirtyend = 0; 2121 bp->b_bufobj = NULL; 2122 bp->b_pin_count = 0; 2123 bp->b_fsprivate1 = NULL; 2124 bp->b_fsprivate2 = NULL; 2125 bp->b_fsprivate3 = NULL; 2126 2127 LIST_INIT(&bp->b_dep); 2128} 2129 2130static int flushingbufs; 2131 2132static struct buf * 2133getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata) 2134{ 2135 struct buf *bp, *nbp; 2136 int nqindex, qindex, pass; 2137 2138 KASSERT(!unmapped || !defrag, ("both unmapped and defrag")); 2139 2140 pass = 1; 2141restart: 2142 atomic_add_int(&getnewbufrestarts, 1); 2143 2144 /* 2145 * Setup for scan. If we do not have enough free buffers, 2146 * we setup a degenerate case that immediately fails. Note 2147 * that if we are specially marked process, we are allowed to 2148 * dip into our reserves. 2149 * 2150 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 2151 * for the allocation of the mapped buffer. For unmapped, the 2152 * easiest is to start with EMPTY outright. 2153 * 2154 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 2155 * However, there are a number of cases (defragging, reusing, ...) 2156 * where we cannot backup. 2157 */ 2158 nbp = NULL; 2159 mtx_lock(&bqlock); 2160 if (!defrag && unmapped) { 2161 nqindex = QUEUE_EMPTY; 2162 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2163 } 2164 if (nbp == NULL) { 2165 nqindex = QUEUE_EMPTYKVA; 2166 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2167 } 2168 2169 /* 2170 * If no EMPTYKVA buffers and we are either defragging or 2171 * reusing, locate a CLEAN buffer to free or reuse. If 2172 * bufspace useage is low skip this step so we can allocate a 2173 * new buffer. 2174 */ 2175 if (nbp == NULL && (defrag || bufspace >= lobufspace)) { 2176 nqindex = QUEUE_CLEAN; 2177 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2178 } 2179 2180 /* 2181 * If we could not find or were not allowed to reuse a CLEAN 2182 * buffer, check to see if it is ok to use an EMPTY buffer. 2183 * We can only use an EMPTY buffer if allocating its KVA would 2184 * not otherwise run us out of buffer space. No KVA is needed 2185 * for the unmapped allocation. 2186 */ 2187 if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace || 2188 metadata)) { 2189 nqindex = QUEUE_EMPTY; 2190 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2191 } 2192 2193 /* 2194 * All available buffers might be clean, retry ignoring the 2195 * lobufspace as the last resort. 2196 */ 2197 if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) { 2198 nqindex = QUEUE_CLEAN; 2199 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2200 } 2201 2202 /* 2203 * Run scan, possibly freeing data and/or kva mappings on the fly 2204 * depending. 2205 */ 2206 while ((bp = nbp) != NULL) { 2207 qindex = nqindex; 2208 2209 /* 2210 * Calculate next bp (we can only use it if we do not 2211 * block or do other fancy things). 2212 */ 2213 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 2214 switch (qindex) { 2215 case QUEUE_EMPTY: 2216 nqindex = QUEUE_EMPTYKVA; 2217 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2218 if (nbp != NULL) 2219 break; 2220 /* FALLTHROUGH */ 2221 case QUEUE_EMPTYKVA: 2222 nqindex = QUEUE_CLEAN; 2223 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2224 if (nbp != NULL) 2225 break; 2226 /* FALLTHROUGH */ 2227 case QUEUE_CLEAN: 2228 if (metadata && pass == 1) { 2229 pass = 2; 2230 nqindex = QUEUE_EMPTY; 2231 nbp = TAILQ_FIRST( 2232 &bufqueues[QUEUE_EMPTY]); 2233 } 2234 /* 2235 * nbp is NULL. 2236 */ 2237 break; 2238 } 2239 } 2240 /* 2241 * If we are defragging then we need a buffer with 2242 * b_kvasize != 0. XXX this situation should no longer 2243 * occur, if defrag is non-zero the buffer's b_kvasize 2244 * should also be non-zero at this point. XXX 2245 */ 2246 if (defrag && bp->b_kvasize == 0) { 2247 printf("Warning: defrag empty buffer %p\n", bp); 2248 continue; 2249 } 2250 2251 /* 2252 * Start freeing the bp. This is somewhat involved. nbp 2253 * remains valid only for QUEUE_EMPTY[KVA] bp's. 2254 */ 2255 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2256 continue; 2257 /* 2258 * BKGRDINPROG can only be set with the buf and bufobj 2259 * locks both held. We tolerate a race to clear it here. 2260 */ 2261 if (bp->b_vflags & BV_BKGRDINPROG) { 2262 BUF_UNLOCK(bp); 2263 continue; 2264 } 2265 2266 KASSERT(bp->b_qindex == qindex, 2267 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2268 2269 bremfreel(bp); 2270 mtx_unlock(&bqlock); 2271 /* 2272 * NOTE: nbp is now entirely invalid. We can only restart 2273 * the scan from this point on. 2274 */ 2275 2276 getnewbuf_reuse_bp(bp, qindex); 2277 mtx_assert(&bqlock, MA_NOTOWNED); 2278 2279 /* 2280 * If we are defragging then free the buffer. 2281 */ 2282 if (defrag) { 2283 bp->b_flags |= B_INVAL; 2284 bfreekva(bp); 2285 brelse(bp); 2286 defrag = 0; 2287 goto restart; 2288 } 2289 2290 /* 2291 * Notify any waiters for the buffer lock about 2292 * identity change by freeing the buffer. 2293 */ 2294 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 2295 bp->b_flags |= B_INVAL; 2296 bfreekva(bp); 2297 brelse(bp); 2298 goto restart; 2299 } 2300 2301 if (metadata) 2302 break; 2303 2304 /* 2305 * If we are overcomitted then recover the buffer and its 2306 * KVM space. This occurs in rare situations when multiple 2307 * processes are blocked in getnewbuf() or allocbuf(). 2308 */ 2309 if (bufspace >= hibufspace) 2310 flushingbufs = 1; 2311 if (flushingbufs && bp->b_kvasize != 0) { 2312 bp->b_flags |= B_INVAL; 2313 bfreekva(bp); 2314 brelse(bp); 2315 goto restart; 2316 } 2317 if (bufspace < lobufspace) 2318 flushingbufs = 0; 2319 break; 2320 } 2321 return (bp); 2322} 2323 2324/* 2325 * getnewbuf: 2326 * 2327 * Find and initialize a new buffer header, freeing up existing buffers 2328 * in the bufqueues as necessary. The new buffer is returned locked. 2329 * 2330 * Important: B_INVAL is not set. If the caller wishes to throw the 2331 * buffer away, the caller must set B_INVAL prior to calling brelse(). 2332 * 2333 * We block if: 2334 * We have insufficient buffer headers 2335 * We have insufficient buffer space 2336 * buffer_map is too fragmented ( space reservation fails ) 2337 * If we have to flush dirty buffers ( but we try to avoid this ) 2338 * 2339 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 2340 * Instead we ask the buf daemon to do it for us. We attempt to 2341 * avoid piecemeal wakeups of the pageout daemon. 2342 */ 2343static struct buf * 2344getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 2345 int gbflags) 2346{ 2347 struct buf *bp; 2348 int defrag, metadata; 2349 2350 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 2351 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 2352 if (!unmapped_buf_allowed) 2353 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 2354 2355 defrag = 0; 2356 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 2357 vp->v_type == VCHR) 2358 metadata = 1; 2359 else 2360 metadata = 0; 2361 /* 2362 * We can't afford to block since we might be holding a vnode lock, 2363 * which may prevent system daemons from running. We deal with 2364 * low-memory situations by proactively returning memory and running 2365 * async I/O rather then sync I/O. 2366 */ 2367 atomic_add_int(&getnewbufcalls, 1); 2368 atomic_subtract_int(&getnewbufrestarts, 1); 2369restart: 2370 bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED | 2371 GB_KVAALLOC)) == GB_UNMAPPED, metadata); 2372 if (bp != NULL) 2373 defrag = 0; 2374 2375 /* 2376 * If we exhausted our list, sleep as appropriate. We may have to 2377 * wakeup various daemons and write out some dirty buffers. 2378 * 2379 * Generally we are sleeping due to insufficient buffer space. 2380 */ 2381 if (bp == NULL) { 2382 mtx_assert(&bqlock, MA_OWNED); 2383 getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag); 2384 mtx_assert(&bqlock, MA_NOTOWNED); 2385 } else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) { 2386 mtx_assert(&bqlock, MA_NOTOWNED); 2387 2388 bfreekva(bp); 2389 bp->b_flags |= B_UNMAPPED; 2390 bp->b_kvabase = bp->b_data = unmapped_buf; 2391 bp->b_kvasize = maxsize; 2392 atomic_add_long(&bufspace, bp->b_kvasize); 2393 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 2394 atomic_add_int(&bufreusecnt, 1); 2395 } else { 2396 mtx_assert(&bqlock, MA_NOTOWNED); 2397 2398 /* 2399 * We finally have a valid bp. We aren't quite out of the 2400 * woods, we still have to reserve kva space. In order 2401 * to keep fragmentation sane we only allocate kva in 2402 * BKVASIZE chunks. 2403 */ 2404 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2405 2406 if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED | 2407 B_KVAALLOC)) == B_UNMAPPED) { 2408 if (allocbufkva(bp, maxsize, gbflags)) { 2409 defrag = 1; 2410 bp->b_flags |= B_INVAL; 2411 brelse(bp); 2412 goto restart; 2413 } 2414 atomic_add_int(&bufreusecnt, 1); 2415 } else if ((bp->b_flags & B_KVAALLOC) != 0 && 2416 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) { 2417 /* 2418 * If the reused buffer has KVA allocated, 2419 * reassign b_kvaalloc to b_kvabase. 2420 */ 2421 bp->b_kvabase = bp->b_kvaalloc; 2422 bp->b_flags &= ~B_KVAALLOC; 2423 atomic_subtract_long(&unmapped_bufspace, 2424 bp->b_kvasize); 2425 atomic_add_int(&bufreusecnt, 1); 2426 } else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 2427 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED | 2428 GB_KVAALLOC)) { 2429 /* 2430 * The case of reused buffer already have KVA 2431 * mapped, but the request is for unmapped 2432 * buffer with KVA allocated. 2433 */ 2434 bp->b_kvaalloc = bp->b_kvabase; 2435 bp->b_data = bp->b_kvabase = unmapped_buf; 2436 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 2437 atomic_add_long(&unmapped_bufspace, 2438 bp->b_kvasize); 2439 atomic_add_int(&bufreusecnt, 1); 2440 } 2441 if ((gbflags & GB_UNMAPPED) == 0) { 2442 bp->b_saveaddr = bp->b_kvabase; 2443 bp->b_data = bp->b_saveaddr; 2444 bp->b_flags &= ~B_UNMAPPED; 2445 BUF_CHECK_MAPPED(bp); 2446 } 2447 } 2448 return (bp); 2449} 2450 2451/* 2452 * buf_daemon: 2453 * 2454 * buffer flushing daemon. Buffers are normally flushed by the 2455 * update daemon but if it cannot keep up this process starts to 2456 * take the load in an attempt to prevent getnewbuf() from blocking. 2457 */ 2458 2459static struct kproc_desc buf_kp = { 2460 "bufdaemon", 2461 buf_daemon, 2462 &bufdaemonproc 2463}; 2464SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2465 2466static int 2467buf_do_flush(struct vnode *vp) 2468{ 2469 int flushed; 2470 2471 flushed = flushbufqueues(vp, QUEUE_DIRTY, 0); 2472 if (flushed == 0) { 2473 /* 2474 * Could not find any buffers without rollback 2475 * dependencies, so just write the first one 2476 * in the hopes of eventually making progress. 2477 */ 2478 flushbufqueues(vp, QUEUE_DIRTY, 1); 2479 } 2480 return (flushed); 2481} 2482 2483static void 2484buf_daemon() 2485{ 2486 int lodirtysave; 2487 2488 /* 2489 * This process needs to be suspended prior to shutdown sync. 2490 */ 2491 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2492 SHUTDOWN_PRI_LAST); 2493 2494 /* 2495 * This process is allowed to take the buffer cache to the limit 2496 */ 2497 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2498 mtx_lock(&bdlock); 2499 for (;;) { 2500 bd_request = 0; 2501 mtx_unlock(&bdlock); 2502 2503 kproc_suspend_check(bufdaemonproc); 2504 lodirtysave = lodirtybuffers; 2505 if (bd_speedupreq) { 2506 lodirtybuffers = numdirtybuffers / 2; 2507 bd_speedupreq = 0; 2508 } 2509 /* 2510 * Do the flush. Limit the amount of in-transit I/O we 2511 * allow to build up, otherwise we would completely saturate 2512 * the I/O system. Wakeup any waiting processes before we 2513 * normally would so they can run in parallel with our drain. 2514 */ 2515 while (numdirtybuffers > lodirtybuffers) { 2516 if (buf_do_flush(NULL) == 0) 2517 break; 2518 kern_yield(PRI_USER); 2519 } 2520 lodirtybuffers = lodirtysave; 2521 2522 /* 2523 * Only clear bd_request if we have reached our low water 2524 * mark. The buf_daemon normally waits 1 second and 2525 * then incrementally flushes any dirty buffers that have 2526 * built up, within reason. 2527 * 2528 * If we were unable to hit our low water mark and couldn't 2529 * find any flushable buffers, we sleep half a second. 2530 * Otherwise we loop immediately. 2531 */ 2532 mtx_lock(&bdlock); 2533 if (numdirtybuffers <= lodirtybuffers) { 2534 /* 2535 * We reached our low water mark, reset the 2536 * request and sleep until we are needed again. 2537 * The sleep is just so the suspend code works. 2538 */ 2539 bd_request = 0; 2540 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2541 } else { 2542 /* 2543 * We couldn't find any flushable dirty buffers but 2544 * still have too many dirty buffers, we 2545 * have to sleep and try again. (rare) 2546 */ 2547 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2548 } 2549 } 2550} 2551 2552/* 2553 * flushbufqueues: 2554 * 2555 * Try to flush a buffer in the dirty queue. We must be careful to 2556 * free up B_INVAL buffers instead of write them, which NFS is 2557 * particularly sensitive to. 2558 */ 2559static int flushwithdeps = 0; 2560SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2561 0, "Number of buffers flushed with dependecies that require rollbacks"); 2562 2563static int 2564flushbufqueues(struct vnode *lvp, int queue, int flushdeps) 2565{ 2566 struct buf *sentinel; 2567 struct vnode *vp; 2568 struct mount *mp; 2569 struct buf *bp; 2570 int hasdeps; 2571 int flushed; 2572 int target; 2573 2574 if (lvp == NULL) { 2575 target = numdirtybuffers - lodirtybuffers; 2576 if (flushdeps && target > 2) 2577 target /= 2; 2578 } else 2579 target = flushbufqtarget; 2580 flushed = 0; 2581 bp = NULL; 2582 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2583 sentinel->b_qindex = QUEUE_SENTINEL; 2584 mtx_lock(&bqlock); 2585 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2586 while (flushed != target) { 2587 bp = TAILQ_NEXT(sentinel, b_freelist); 2588 if (bp != NULL) { 2589 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2590 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2591 b_freelist); 2592 } else 2593 break; 2594 /* 2595 * Skip sentinels inserted by other invocations of the 2596 * flushbufqueues(), taking care to not reorder them. 2597 */ 2598 if (bp->b_qindex == QUEUE_SENTINEL) 2599 continue; 2600 /* 2601 * Only flush the buffers that belong to the 2602 * vnode locked by the curthread. 2603 */ 2604 if (lvp != NULL && bp->b_vp != lvp) 2605 continue; 2606 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2607 continue; 2608 if (bp->b_pin_count > 0) { 2609 BUF_UNLOCK(bp); 2610 continue; 2611 } 2612 /* 2613 * BKGRDINPROG can only be set with the buf and bufobj 2614 * locks both held. We tolerate a race to clear it here. 2615 */ 2616 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2617 (bp->b_flags & B_DELWRI) == 0) { 2618 BUF_UNLOCK(bp); 2619 continue; 2620 } 2621 if (bp->b_flags & B_INVAL) { 2622 bremfreel(bp); 2623 mtx_unlock(&bqlock); 2624 brelse(bp); 2625 flushed++; 2626 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2627 mtx_lock(&bqlock); 2628 continue; 2629 } 2630 2631 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2632 if (flushdeps == 0) { 2633 BUF_UNLOCK(bp); 2634 continue; 2635 } 2636 hasdeps = 1; 2637 } else 2638 hasdeps = 0; 2639 /* 2640 * We must hold the lock on a vnode before writing 2641 * one of its buffers. Otherwise we may confuse, or 2642 * in the case of a snapshot vnode, deadlock the 2643 * system. 2644 * 2645 * The lock order here is the reverse of the normal 2646 * of vnode followed by buf lock. This is ok because 2647 * the NOWAIT will prevent deadlock. 2648 */ 2649 vp = bp->b_vp; 2650 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2651 BUF_UNLOCK(bp); 2652 continue; 2653 } 2654 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) { 2655 mtx_unlock(&bqlock); 2656 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2657 bp, bp->b_vp, bp->b_flags); 2658 if (curproc == bufdaemonproc) 2659 vfs_bio_awrite(bp); 2660 else { 2661 bremfree(bp); 2662 bwrite(bp); 2663 notbufdflashes++; 2664 } 2665 vn_finished_write(mp); 2666 VOP_UNLOCK(vp, 0); 2667 flushwithdeps += hasdeps; 2668 flushed++; 2669 2670 /* 2671 * Sleeping on runningbufspace while holding 2672 * vnode lock leads to deadlock. 2673 */ 2674 if (curproc == bufdaemonproc) 2675 waitrunningbufspace(); 2676 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2677 mtx_lock(&bqlock); 2678 continue; 2679 } 2680 vn_finished_write(mp); 2681 BUF_UNLOCK(bp); 2682 } 2683 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2684 mtx_unlock(&bqlock); 2685 free(sentinel, M_TEMP); 2686 return (flushed); 2687} 2688 2689/* 2690 * Check to see if a block is currently memory resident. 2691 */ 2692struct buf * 2693incore(struct bufobj *bo, daddr_t blkno) 2694{ 2695 struct buf *bp; 2696 2697 BO_RLOCK(bo); 2698 bp = gbincore(bo, blkno); 2699 BO_RUNLOCK(bo); 2700 return (bp); 2701} 2702 2703/* 2704 * Returns true if no I/O is needed to access the 2705 * associated VM object. This is like incore except 2706 * it also hunts around in the VM system for the data. 2707 */ 2708 2709static int 2710inmem(struct vnode * vp, daddr_t blkno) 2711{ 2712 vm_object_t obj; 2713 vm_offset_t toff, tinc, size; 2714 vm_page_t m; 2715 vm_ooffset_t off; 2716 2717 ASSERT_VOP_LOCKED(vp, "inmem"); 2718 2719 if (incore(&vp->v_bufobj, blkno)) 2720 return 1; 2721 if (vp->v_mount == NULL) 2722 return 0; 2723 obj = vp->v_object; 2724 if (obj == NULL) 2725 return (0); 2726 2727 size = PAGE_SIZE; 2728 if (size > vp->v_mount->mnt_stat.f_iosize) 2729 size = vp->v_mount->mnt_stat.f_iosize; 2730 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2731 2732 VM_OBJECT_RLOCK(obj); 2733 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2734 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2735 if (!m) 2736 goto notinmem; 2737 tinc = size; 2738 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2739 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2740 if (vm_page_is_valid(m, 2741 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2742 goto notinmem; 2743 } 2744 VM_OBJECT_RUNLOCK(obj); 2745 return 1; 2746 2747notinmem: 2748 VM_OBJECT_RUNLOCK(obj); 2749 return (0); 2750} 2751 2752/* 2753 * Set the dirty range for a buffer based on the status of the dirty 2754 * bits in the pages comprising the buffer. The range is limited 2755 * to the size of the buffer. 2756 * 2757 * Tell the VM system that the pages associated with this buffer 2758 * are clean. This is used for delayed writes where the data is 2759 * going to go to disk eventually without additional VM intevention. 2760 * 2761 * Note that while we only really need to clean through to b_bcount, we 2762 * just go ahead and clean through to b_bufsize. 2763 */ 2764static void 2765vfs_clean_pages_dirty_buf(struct buf *bp) 2766{ 2767 vm_ooffset_t foff, noff, eoff; 2768 vm_page_t m; 2769 int i; 2770 2771 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 2772 return; 2773 2774 foff = bp->b_offset; 2775 KASSERT(bp->b_offset != NOOFFSET, 2776 ("vfs_clean_pages_dirty_buf: no buffer offset")); 2777 2778 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 2779 vfs_drain_busy_pages(bp); 2780 vfs_setdirty_locked_object(bp); 2781 for (i = 0; i < bp->b_npages; i++) { 2782 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2783 eoff = noff; 2784 if (eoff > bp->b_offset + bp->b_bufsize) 2785 eoff = bp->b_offset + bp->b_bufsize; 2786 m = bp->b_pages[i]; 2787 vfs_page_set_validclean(bp, foff, m); 2788 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2789 foff = noff; 2790 } 2791 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 2792} 2793 2794static void 2795vfs_setdirty_locked_object(struct buf *bp) 2796{ 2797 vm_object_t object; 2798 int i; 2799 2800 object = bp->b_bufobj->bo_object; 2801 VM_OBJECT_ASSERT_WLOCKED(object); 2802 2803 /* 2804 * We qualify the scan for modified pages on whether the 2805 * object has been flushed yet. 2806 */ 2807 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { 2808 vm_offset_t boffset; 2809 vm_offset_t eoffset; 2810 2811 /* 2812 * test the pages to see if they have been modified directly 2813 * by users through the VM system. 2814 */ 2815 for (i = 0; i < bp->b_npages; i++) 2816 vm_page_test_dirty(bp->b_pages[i]); 2817 2818 /* 2819 * Calculate the encompassing dirty range, boffset and eoffset, 2820 * (eoffset - boffset) bytes. 2821 */ 2822 2823 for (i = 0; i < bp->b_npages; i++) { 2824 if (bp->b_pages[i]->dirty) 2825 break; 2826 } 2827 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2828 2829 for (i = bp->b_npages - 1; i >= 0; --i) { 2830 if (bp->b_pages[i]->dirty) { 2831 break; 2832 } 2833 } 2834 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2835 2836 /* 2837 * Fit it to the buffer. 2838 */ 2839 2840 if (eoffset > bp->b_bcount) 2841 eoffset = bp->b_bcount; 2842 2843 /* 2844 * If we have a good dirty range, merge with the existing 2845 * dirty range. 2846 */ 2847 2848 if (boffset < eoffset) { 2849 if (bp->b_dirtyoff > boffset) 2850 bp->b_dirtyoff = boffset; 2851 if (bp->b_dirtyend < eoffset) 2852 bp->b_dirtyend = eoffset; 2853 } 2854 } 2855} 2856 2857/* 2858 * Allocate the KVA mapping for an existing buffer. It handles the 2859 * cases of both B_UNMAPPED buffer, and buffer with the preallocated 2860 * KVA which is not mapped (B_KVAALLOC). 2861 */ 2862static void 2863bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 2864{ 2865 struct buf *scratch_bp; 2866 int bsize, maxsize, need_mapping, need_kva; 2867 off_t offset; 2868 2869 need_mapping = (bp->b_flags & B_UNMAPPED) != 0 && 2870 (gbflags & GB_UNMAPPED) == 0; 2871 need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED && 2872 (gbflags & GB_KVAALLOC) != 0; 2873 if (!need_mapping && !need_kva) 2874 return; 2875 2876 BUF_CHECK_UNMAPPED(bp); 2877 2878 if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) { 2879 /* 2880 * Buffer is not mapped, but the KVA was already 2881 * reserved at the time of the instantiation. Use the 2882 * allocated space. 2883 */ 2884 bp->b_flags &= ~B_KVAALLOC; 2885 KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0")); 2886 bp->b_kvabase = bp->b_kvaalloc; 2887 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize); 2888 goto has_addr; 2889 } 2890 2891 /* 2892 * Calculate the amount of the address space we would reserve 2893 * if the buffer was mapped. 2894 */ 2895 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 2896 offset = blkno * bsize; 2897 maxsize = size + (offset & PAGE_MASK); 2898 maxsize = imax(maxsize, bsize); 2899 2900mapping_loop: 2901 if (allocbufkva(bp, maxsize, gbflags)) { 2902 /* 2903 * Request defragmentation. getnewbuf() returns us the 2904 * allocated space by the scratch buffer KVA. 2905 */ 2906 scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags | 2907 (GB_UNMAPPED | GB_KVAALLOC)); 2908 if (scratch_bp == NULL) { 2909 if ((gbflags & GB_NOWAIT_BD) != 0) { 2910 /* 2911 * XXXKIB: defragmentation cannot 2912 * succeed, not sure what else to do. 2913 */ 2914 panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp); 2915 } 2916 atomic_add_int(&mappingrestarts, 1); 2917 goto mapping_loop; 2918 } 2919 KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0, 2920 ("scratch bp !B_KVAALLOC %p", scratch_bp)); 2921 setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc, 2922 scratch_bp->b_kvasize, gbflags); 2923 2924 /* Get rid of the scratch buffer. */ 2925 scratch_bp->b_kvasize = 0; 2926 scratch_bp->b_flags |= B_INVAL; 2927 scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC); 2928 brelse(scratch_bp); 2929 } 2930 if (!need_mapping) 2931 return; 2932 2933has_addr: 2934 bp->b_saveaddr = bp->b_kvabase; 2935 bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */ 2936 bp->b_flags &= ~B_UNMAPPED; 2937 BUF_CHECK_MAPPED(bp); 2938 bpmap_qenter(bp); 2939} 2940 2941/* 2942 * getblk: 2943 * 2944 * Get a block given a specified block and offset into a file/device. 2945 * The buffers B_DONE bit will be cleared on return, making it almost 2946 * ready for an I/O initiation. B_INVAL may or may not be set on 2947 * return. The caller should clear B_INVAL prior to initiating a 2948 * READ. 2949 * 2950 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2951 * an existing buffer. 2952 * 2953 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2954 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2955 * and then cleared based on the backing VM. If the previous buffer is 2956 * non-0-sized but invalid, B_CACHE will be cleared. 2957 * 2958 * If getblk() must create a new buffer, the new buffer is returned with 2959 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2960 * case it is returned with B_INVAL clear and B_CACHE set based on the 2961 * backing VM. 2962 * 2963 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2964 * B_CACHE bit is clear. 2965 * 2966 * What this means, basically, is that the caller should use B_CACHE to 2967 * determine whether the buffer is fully valid or not and should clear 2968 * B_INVAL prior to issuing a read. If the caller intends to validate 2969 * the buffer by loading its data area with something, the caller needs 2970 * to clear B_INVAL. If the caller does this without issuing an I/O, 2971 * the caller should set B_CACHE ( as an optimization ), else the caller 2972 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2973 * a write attempt or if it was a successfull read. If the caller 2974 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2975 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2976 */ 2977struct buf * 2978getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2979 int flags) 2980{ 2981 struct buf *bp; 2982 struct bufobj *bo; 2983 int bsize, error, maxsize, vmio; 2984 off_t offset; 2985 2986 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2987 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 2988 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 2989 ASSERT_VOP_LOCKED(vp, "getblk"); 2990 if (size > MAXBSIZE) 2991 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2992 if (!unmapped_buf_allowed) 2993 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 2994 2995 bo = &vp->v_bufobj; 2996loop: 2997 /* 2998 * Block if we are low on buffers. Certain processes are allowed 2999 * to completely exhaust the buffer cache. 3000 * 3001 * If this check ever becomes a bottleneck it may be better to 3002 * move it into the else, when gbincore() fails. At the moment 3003 * it isn't a problem. 3004 */ 3005 if (numfreebuffers == 0) { 3006 if (TD_IS_IDLETHREAD(curthread)) 3007 return NULL; 3008 mtx_lock(&nblock); 3009 needsbuffer |= VFS_BIO_NEED_ANY; 3010 mtx_unlock(&nblock); 3011 } 3012 3013 BO_RLOCK(bo); 3014 bp = gbincore(bo, blkno); 3015 if (bp != NULL) { 3016 int lockflags; 3017 /* 3018 * Buffer is in-core. If the buffer is not busy nor managed, 3019 * it must be on a queue. 3020 */ 3021 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 3022 3023 if (flags & GB_LOCK_NOWAIT) 3024 lockflags |= LK_NOWAIT; 3025 3026 error = BUF_TIMELOCK(bp, lockflags, 3027 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); 3028 3029 /* 3030 * If we slept and got the lock we have to restart in case 3031 * the buffer changed identities. 3032 */ 3033 if (error == ENOLCK) 3034 goto loop; 3035 /* We timed out or were interrupted. */ 3036 else if (error) 3037 return (NULL); 3038 /* If recursed, assume caller knows the rules. */ 3039 else if (BUF_LOCKRECURSED(bp)) 3040 goto end; 3041 3042 /* 3043 * The buffer is locked. B_CACHE is cleared if the buffer is 3044 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 3045 * and for a VMIO buffer B_CACHE is adjusted according to the 3046 * backing VM cache. 3047 */ 3048 if (bp->b_flags & B_INVAL) 3049 bp->b_flags &= ~B_CACHE; 3050 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 3051 bp->b_flags |= B_CACHE; 3052 if (bp->b_flags & B_MANAGED) 3053 MPASS(bp->b_qindex == QUEUE_NONE); 3054 else 3055 bremfree(bp); 3056 3057 /* 3058 * check for size inconsistencies for non-VMIO case. 3059 */ 3060 if (bp->b_bcount != size) { 3061 if ((bp->b_flags & B_VMIO) == 0 || 3062 (size > bp->b_kvasize)) { 3063 if (bp->b_flags & B_DELWRI) { 3064 /* 3065 * If buffer is pinned and caller does 3066 * not want sleep waiting for it to be 3067 * unpinned, bail out 3068 * */ 3069 if (bp->b_pin_count > 0) { 3070 if (flags & GB_LOCK_NOWAIT) { 3071 bqrelse(bp); 3072 return (NULL); 3073 } else { 3074 bunpin_wait(bp); 3075 } 3076 } 3077 bp->b_flags |= B_NOCACHE; 3078 bwrite(bp); 3079 } else { 3080 if (LIST_EMPTY(&bp->b_dep)) { 3081 bp->b_flags |= B_RELBUF; 3082 brelse(bp); 3083 } else { 3084 bp->b_flags |= B_NOCACHE; 3085 bwrite(bp); 3086 } 3087 } 3088 goto loop; 3089 } 3090 } 3091 3092 /* 3093 * Handle the case of unmapped buffer which should 3094 * become mapped, or the buffer for which KVA 3095 * reservation is requested. 3096 */ 3097 bp_unmapped_get_kva(bp, blkno, size, flags); 3098 3099 /* 3100 * If the size is inconsistant in the VMIO case, we can resize 3101 * the buffer. This might lead to B_CACHE getting set or 3102 * cleared. If the size has not changed, B_CACHE remains 3103 * unchanged from its previous state. 3104 */ 3105 if (bp->b_bcount != size) 3106 allocbuf(bp, size); 3107 3108 KASSERT(bp->b_offset != NOOFFSET, 3109 ("getblk: no buffer offset")); 3110 3111 /* 3112 * A buffer with B_DELWRI set and B_CACHE clear must 3113 * be committed before we can return the buffer in 3114 * order to prevent the caller from issuing a read 3115 * ( due to B_CACHE not being set ) and overwriting 3116 * it. 3117 * 3118 * Most callers, including NFS and FFS, need this to 3119 * operate properly either because they assume they 3120 * can issue a read if B_CACHE is not set, or because 3121 * ( for example ) an uncached B_DELWRI might loop due 3122 * to softupdates re-dirtying the buffer. In the latter 3123 * case, B_CACHE is set after the first write completes, 3124 * preventing further loops. 3125 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3126 * above while extending the buffer, we cannot allow the 3127 * buffer to remain with B_CACHE set after the write 3128 * completes or it will represent a corrupt state. To 3129 * deal with this we set B_NOCACHE to scrap the buffer 3130 * after the write. 3131 * 3132 * We might be able to do something fancy, like setting 3133 * B_CACHE in bwrite() except if B_DELWRI is already set, 3134 * so the below call doesn't set B_CACHE, but that gets real 3135 * confusing. This is much easier. 3136 */ 3137 3138 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3139 bp->b_flags |= B_NOCACHE; 3140 bwrite(bp); 3141 goto loop; 3142 } 3143 bp->b_flags &= ~B_DONE; 3144 } else { 3145 /* 3146 * Buffer is not in-core, create new buffer. The buffer 3147 * returned by getnewbuf() is locked. Note that the returned 3148 * buffer is also considered valid (not marked B_INVAL). 3149 */ 3150 BO_RUNLOCK(bo); 3151 /* 3152 * If the user does not want us to create the buffer, bail out 3153 * here. 3154 */ 3155 if (flags & GB_NOCREAT) 3156 return NULL; 3157 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 3158 offset = blkno * bsize; 3159 vmio = vp->v_object != NULL; 3160 if (vmio) { 3161 maxsize = size + (offset & PAGE_MASK); 3162 } else { 3163 maxsize = size; 3164 /* Do not allow non-VMIO notmapped buffers. */ 3165 flags &= ~GB_UNMAPPED; 3166 } 3167 maxsize = imax(maxsize, bsize); 3168 3169 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 3170 if (bp == NULL) { 3171 if (slpflag || slptimeo) 3172 return NULL; 3173 goto loop; 3174 } 3175 3176 /* 3177 * This code is used to make sure that a buffer is not 3178 * created while the getnewbuf routine is blocked. 3179 * This can be a problem whether the vnode is locked or not. 3180 * If the buffer is created out from under us, we have to 3181 * throw away the one we just created. 3182 * 3183 * Note: this must occur before we associate the buffer 3184 * with the vp especially considering limitations in 3185 * the splay tree implementation when dealing with duplicate 3186 * lblkno's. 3187 */ 3188 BO_LOCK(bo); 3189 if (gbincore(bo, blkno)) { 3190 BO_UNLOCK(bo); 3191 bp->b_flags |= B_INVAL; 3192 brelse(bp); 3193 goto loop; 3194 } 3195 3196 /* 3197 * Insert the buffer into the hash, so that it can 3198 * be found by incore. 3199 */ 3200 bp->b_blkno = bp->b_lblkno = blkno; 3201 bp->b_offset = offset; 3202 bgetvp(vp, bp); 3203 BO_UNLOCK(bo); 3204 3205 /* 3206 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 3207 * buffer size starts out as 0, B_CACHE will be set by 3208 * allocbuf() for the VMIO case prior to it testing the 3209 * backing store for validity. 3210 */ 3211 3212 if (vmio) { 3213 bp->b_flags |= B_VMIO; 3214 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 3215 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 3216 bp, vp->v_object, bp->b_bufobj->bo_object)); 3217 } else { 3218 bp->b_flags &= ~B_VMIO; 3219 KASSERT(bp->b_bufobj->bo_object == NULL, 3220 ("ARGH! has b_bufobj->bo_object %p %p\n", 3221 bp, bp->b_bufobj->bo_object)); 3222 BUF_CHECK_MAPPED(bp); 3223 } 3224 3225 allocbuf(bp, size); 3226 bp->b_flags &= ~B_DONE; 3227 } 3228 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 3229 BUF_ASSERT_HELD(bp); 3230end: 3231 KASSERT(bp->b_bufobj == bo, 3232 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3233 return (bp); 3234} 3235 3236/* 3237 * Get an empty, disassociated buffer of given size. The buffer is initially 3238 * set to B_INVAL. 3239 */ 3240struct buf * 3241geteblk(int size, int flags) 3242{ 3243 struct buf *bp; 3244 int maxsize; 3245 3246 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3247 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 3248 if ((flags & GB_NOWAIT_BD) && 3249 (curthread->td_pflags & TDP_BUFNEED) != 0) 3250 return (NULL); 3251 } 3252 allocbuf(bp, size); 3253 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3254 BUF_ASSERT_HELD(bp); 3255 return (bp); 3256} 3257 3258 3259/* 3260 * This code constitutes the buffer memory from either anonymous system 3261 * memory (in the case of non-VMIO operations) or from an associated 3262 * VM object (in the case of VMIO operations). This code is able to 3263 * resize a buffer up or down. 3264 * 3265 * Note that this code is tricky, and has many complications to resolve 3266 * deadlock or inconsistant data situations. Tread lightly!!! 3267 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3268 * the caller. Calling this code willy nilly can result in the loss of data. 3269 * 3270 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3271 * B_CACHE for the non-VMIO case. 3272 */ 3273 3274int 3275allocbuf(struct buf *bp, int size) 3276{ 3277 int newbsize, mbsize; 3278 int i; 3279 3280 BUF_ASSERT_HELD(bp); 3281 3282 if (bp->b_kvasize < size) 3283 panic("allocbuf: buffer too small"); 3284 3285 if ((bp->b_flags & B_VMIO) == 0) { 3286 caddr_t origbuf; 3287 int origbufsize; 3288 /* 3289 * Just get anonymous memory from the kernel. Don't 3290 * mess with B_CACHE. 3291 */ 3292 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3293 if (bp->b_flags & B_MALLOC) 3294 newbsize = mbsize; 3295 else 3296 newbsize = round_page(size); 3297 3298 if (newbsize < bp->b_bufsize) { 3299 /* 3300 * malloced buffers are not shrunk 3301 */ 3302 if (bp->b_flags & B_MALLOC) { 3303 if (newbsize) { 3304 bp->b_bcount = size; 3305 } else { 3306 free(bp->b_data, M_BIOBUF); 3307 if (bp->b_bufsize) { 3308 atomic_subtract_long( 3309 &bufmallocspace, 3310 bp->b_bufsize); 3311 bufspacewakeup(); 3312 bp->b_bufsize = 0; 3313 } 3314 bp->b_saveaddr = bp->b_kvabase; 3315 bp->b_data = bp->b_saveaddr; 3316 bp->b_bcount = 0; 3317 bp->b_flags &= ~B_MALLOC; 3318 } 3319 return 1; 3320 } 3321 vm_hold_free_pages(bp, newbsize); 3322 } else if (newbsize > bp->b_bufsize) { 3323 /* 3324 * We only use malloced memory on the first allocation. 3325 * and revert to page-allocated memory when the buffer 3326 * grows. 3327 */ 3328 /* 3329 * There is a potential smp race here that could lead 3330 * to bufmallocspace slightly passing the max. It 3331 * is probably extremely rare and not worth worrying 3332 * over. 3333 */ 3334 if ( (bufmallocspace < maxbufmallocspace) && 3335 (bp->b_bufsize == 0) && 3336 (mbsize <= PAGE_SIZE/2)) { 3337 3338 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 3339 bp->b_bufsize = mbsize; 3340 bp->b_bcount = size; 3341 bp->b_flags |= B_MALLOC; 3342 atomic_add_long(&bufmallocspace, mbsize); 3343 return 1; 3344 } 3345 origbuf = NULL; 3346 origbufsize = 0; 3347 /* 3348 * If the buffer is growing on its other-than-first allocation, 3349 * then we revert to the page-allocation scheme. 3350 */ 3351 if (bp->b_flags & B_MALLOC) { 3352 origbuf = bp->b_data; 3353 origbufsize = bp->b_bufsize; 3354 bp->b_data = bp->b_kvabase; 3355 if (bp->b_bufsize) { 3356 atomic_subtract_long(&bufmallocspace, 3357 bp->b_bufsize); 3358 bufspacewakeup(); 3359 bp->b_bufsize = 0; 3360 } 3361 bp->b_flags &= ~B_MALLOC; 3362 newbsize = round_page(newbsize); 3363 } 3364 vm_hold_load_pages( 3365 bp, 3366 (vm_offset_t) bp->b_data + bp->b_bufsize, 3367 (vm_offset_t) bp->b_data + newbsize); 3368 if (origbuf) { 3369 bcopy(origbuf, bp->b_data, origbufsize); 3370 free(origbuf, M_BIOBUF); 3371 } 3372 } 3373 } else { 3374 int desiredpages; 3375 3376 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3377 desiredpages = (size == 0) ? 0 : 3378 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 3379 3380 if (bp->b_flags & B_MALLOC) 3381 panic("allocbuf: VMIO buffer can't be malloced"); 3382 /* 3383 * Set B_CACHE initially if buffer is 0 length or will become 3384 * 0-length. 3385 */ 3386 if (size == 0 || bp->b_bufsize == 0) 3387 bp->b_flags |= B_CACHE; 3388 3389 if (newbsize < bp->b_bufsize) { 3390 /* 3391 * DEV_BSIZE aligned new buffer size is less then the 3392 * DEV_BSIZE aligned existing buffer size. Figure out 3393 * if we have to remove any pages. 3394 */ 3395 if (desiredpages < bp->b_npages) { 3396 vm_page_t m; 3397 3398 if ((bp->b_flags & B_UNMAPPED) == 0) { 3399 BUF_CHECK_MAPPED(bp); 3400 pmap_qremove((vm_offset_t)trunc_page( 3401 (vm_offset_t)bp->b_data) + 3402 (desiredpages << PAGE_SHIFT), 3403 (bp->b_npages - desiredpages)); 3404 } else 3405 BUF_CHECK_UNMAPPED(bp); 3406 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3407 for (i = desiredpages; i < bp->b_npages; i++) { 3408 /* 3409 * the page is not freed here -- it 3410 * is the responsibility of 3411 * vnode_pager_setsize 3412 */ 3413 m = bp->b_pages[i]; 3414 KASSERT(m != bogus_page, 3415 ("allocbuf: bogus page found")); 3416 while (vm_page_sleep_if_busy(m, TRUE, 3417 "biodep")) 3418 continue; 3419 3420 bp->b_pages[i] = NULL; 3421 vm_page_lock(m); 3422 vm_page_unwire(m, 0); 3423 vm_page_unlock(m); 3424 } 3425 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3426 bp->b_npages = desiredpages; 3427 } 3428 } else if (size > bp->b_bcount) { 3429 /* 3430 * We are growing the buffer, possibly in a 3431 * byte-granular fashion. 3432 */ 3433 vm_object_t obj; 3434 vm_offset_t toff; 3435 vm_offset_t tinc; 3436 3437 /* 3438 * Step 1, bring in the VM pages from the object, 3439 * allocating them if necessary. We must clear 3440 * B_CACHE if these pages are not valid for the 3441 * range covered by the buffer. 3442 */ 3443 3444 obj = bp->b_bufobj->bo_object; 3445 3446 VM_OBJECT_WLOCK(obj); 3447 while (bp->b_npages < desiredpages) { 3448 vm_page_t m; 3449 3450 /* 3451 * We must allocate system pages since blocking 3452 * here could interfere with paging I/O, no 3453 * matter which process we are. 3454 * 3455 * We can only test VPO_BUSY here. Blocking on 3456 * m->busy might lead to a deadlock: 3457 * vm_fault->getpages->cluster_read->allocbuf 3458 * Thus, we specify VM_ALLOC_IGN_SBUSY. 3459 */ 3460 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + 3461 bp->b_npages, VM_ALLOC_NOBUSY | 3462 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3463 VM_ALLOC_RETRY | VM_ALLOC_IGN_SBUSY | 3464 VM_ALLOC_COUNT(desiredpages - bp->b_npages)); 3465 if (m->valid == 0) 3466 bp->b_flags &= ~B_CACHE; 3467 bp->b_pages[bp->b_npages] = m; 3468 ++bp->b_npages; 3469 } 3470 3471 /* 3472 * Step 2. We've loaded the pages into the buffer, 3473 * we have to figure out if we can still have B_CACHE 3474 * set. Note that B_CACHE is set according to the 3475 * byte-granular range ( bcount and size ), new the 3476 * aligned range ( newbsize ). 3477 * 3478 * The VM test is against m->valid, which is DEV_BSIZE 3479 * aligned. Needless to say, the validity of the data 3480 * needs to also be DEV_BSIZE aligned. Note that this 3481 * fails with NFS if the server or some other client 3482 * extends the file's EOF. If our buffer is resized, 3483 * B_CACHE may remain set! XXX 3484 */ 3485 3486 toff = bp->b_bcount; 3487 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3488 3489 while ((bp->b_flags & B_CACHE) && toff < size) { 3490 vm_pindex_t pi; 3491 3492 if (tinc > (size - toff)) 3493 tinc = size - toff; 3494 3495 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3496 PAGE_SHIFT; 3497 3498 vfs_buf_test_cache( 3499 bp, 3500 bp->b_offset, 3501 toff, 3502 tinc, 3503 bp->b_pages[pi] 3504 ); 3505 toff += tinc; 3506 tinc = PAGE_SIZE; 3507 } 3508 VM_OBJECT_WUNLOCK(obj); 3509 3510 /* 3511 * Step 3, fixup the KVM pmap. 3512 */ 3513 if ((bp->b_flags & B_UNMAPPED) == 0) 3514 bpmap_qenter(bp); 3515 else 3516 BUF_CHECK_UNMAPPED(bp); 3517 } 3518 } 3519 if (newbsize < bp->b_bufsize) 3520 bufspacewakeup(); 3521 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3522 bp->b_bcount = size; /* requested buffer size */ 3523 return 1; 3524} 3525 3526extern int inflight_transient_maps; 3527 3528void 3529biodone(struct bio *bp) 3530{ 3531 struct mtx *mtxp; 3532 void (*done)(struct bio *); 3533 vm_offset_t start, end; 3534 int transient; 3535 3536 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3537 mtx_lock(mtxp); 3538 bp->bio_flags |= BIO_DONE; 3539 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 3540 start = trunc_page((vm_offset_t)bp->bio_data); 3541 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 3542 transient = 1; 3543 } else { 3544 transient = 0; 3545 start = end = 0; 3546 } 3547 done = bp->bio_done; 3548 if (done == NULL) 3549 wakeup(bp); 3550 mtx_unlock(mtxp); 3551 if (done != NULL) 3552 done(bp); 3553 if (transient) { 3554 pmap_qremove(start, OFF_TO_IDX(end - start)); 3555 vm_map_remove(bio_transient_map, start, end); 3556 atomic_add_int(&inflight_transient_maps, -1); 3557 } 3558} 3559 3560/* 3561 * Wait for a BIO to finish. 3562 * 3563 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3564 * case is not yet clear. 3565 */ 3566int 3567biowait(struct bio *bp, const char *wchan) 3568{ 3569 struct mtx *mtxp; 3570 3571 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3572 mtx_lock(mtxp); 3573 while ((bp->bio_flags & BIO_DONE) == 0) 3574 msleep(bp, mtxp, PRIBIO, wchan, hz / 10); 3575 mtx_unlock(mtxp); 3576 if (bp->bio_error != 0) 3577 return (bp->bio_error); 3578 if (!(bp->bio_flags & BIO_ERROR)) 3579 return (0); 3580 return (EIO); 3581} 3582 3583void 3584biofinish(struct bio *bp, struct devstat *stat, int error) 3585{ 3586 3587 if (error) { 3588 bp->bio_error = error; 3589 bp->bio_flags |= BIO_ERROR; 3590 } 3591 if (stat != NULL) 3592 devstat_end_transaction_bio(stat, bp); 3593 biodone(bp); 3594} 3595 3596/* 3597 * bufwait: 3598 * 3599 * Wait for buffer I/O completion, returning error status. The buffer 3600 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3601 * error and cleared. 3602 */ 3603int 3604bufwait(struct buf *bp) 3605{ 3606 if (bp->b_iocmd == BIO_READ) 3607 bwait(bp, PRIBIO, "biord"); 3608 else 3609 bwait(bp, PRIBIO, "biowr"); 3610 if (bp->b_flags & B_EINTR) { 3611 bp->b_flags &= ~B_EINTR; 3612 return (EINTR); 3613 } 3614 if (bp->b_ioflags & BIO_ERROR) { 3615 return (bp->b_error ? bp->b_error : EIO); 3616 } else { 3617 return (0); 3618 } 3619} 3620 3621 /* 3622 * Call back function from struct bio back up to struct buf. 3623 */ 3624static void 3625bufdonebio(struct bio *bip) 3626{ 3627 struct buf *bp; 3628 3629 bp = bip->bio_caller2; 3630 bp->b_resid = bp->b_bcount - bip->bio_completed; 3631 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3632 bp->b_ioflags = bip->bio_flags; 3633 bp->b_error = bip->bio_error; 3634 if (bp->b_error) 3635 bp->b_ioflags |= BIO_ERROR; 3636 bufdone(bp); 3637 g_destroy_bio(bip); 3638} 3639 3640void 3641dev_strategy(struct cdev *dev, struct buf *bp) 3642{ 3643 struct cdevsw *csw; 3644 int ref; 3645 3646 KASSERT(dev->si_refcount > 0, 3647 ("dev_strategy on un-referenced struct cdev *(%s) %p", 3648 devtoname(dev), dev)); 3649 3650 csw = dev_refthread(dev, &ref); 3651 dev_strategy_csw(dev, csw, bp); 3652 dev_relthread(dev, ref); 3653} 3654 3655void 3656dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp) 3657{ 3658 struct bio *bip; 3659 3660 KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE, 3661 ("b_iocmd botch")); 3662 KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) || 3663 dev->si_threadcount > 0, 3664 ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev), 3665 dev)); 3666 if (csw == NULL) { 3667 bp->b_error = ENXIO; 3668 bp->b_ioflags = BIO_ERROR; 3669 bufdone(bp); 3670 return; 3671 } 3672 for (;;) { 3673 bip = g_new_bio(); 3674 if (bip != NULL) 3675 break; 3676 /* Try again later */ 3677 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3678 } 3679 bip->bio_cmd = bp->b_iocmd; 3680 bip->bio_offset = bp->b_iooffset; 3681 bip->bio_length = bp->b_bcount; 3682 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3683 bdata2bio(bp, bip); 3684 bip->bio_done = bufdonebio; 3685 bip->bio_caller2 = bp; 3686 bip->bio_dev = dev; 3687 (*csw->d_strategy)(bip); 3688} 3689 3690/* 3691 * bufdone: 3692 * 3693 * Finish I/O on a buffer, optionally calling a completion function. 3694 * This is usually called from an interrupt so process blocking is 3695 * not allowed. 3696 * 3697 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3698 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3699 * assuming B_INVAL is clear. 3700 * 3701 * For the VMIO case, we set B_CACHE if the op was a read and no 3702 * read error occured, or if the op was a write. B_CACHE is never 3703 * set if the buffer is invalid or otherwise uncacheable. 3704 * 3705 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3706 * initiator to leave B_INVAL set to brelse the buffer out of existance 3707 * in the biodone routine. 3708 */ 3709void 3710bufdone(struct buf *bp) 3711{ 3712 struct bufobj *dropobj; 3713 void (*biodone)(struct buf *); 3714 3715 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3716 dropobj = NULL; 3717 3718 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3719 BUF_ASSERT_HELD(bp); 3720 3721 runningbufwakeup(bp); 3722 if (bp->b_iocmd == BIO_WRITE) 3723 dropobj = bp->b_bufobj; 3724 /* call optional completion function if requested */ 3725 if (bp->b_iodone != NULL) { 3726 biodone = bp->b_iodone; 3727 bp->b_iodone = NULL; 3728 (*biodone) (bp); 3729 if (dropobj) 3730 bufobj_wdrop(dropobj); 3731 return; 3732 } 3733 3734 bufdone_finish(bp); 3735 3736 if (dropobj) 3737 bufobj_wdrop(dropobj); 3738} 3739 3740void 3741bufdone_finish(struct buf *bp) 3742{ 3743 BUF_ASSERT_HELD(bp); 3744 3745 if (!LIST_EMPTY(&bp->b_dep)) 3746 buf_complete(bp); 3747 3748 if (bp->b_flags & B_VMIO) { 3749 vm_ooffset_t foff; 3750 vm_page_t m; 3751 vm_object_t obj; 3752 struct vnode *vp; 3753 int bogus, i, iosize; 3754 3755 obj = bp->b_bufobj->bo_object; 3756 KASSERT(obj->paging_in_progress >= bp->b_npages, 3757 ("biodone_finish: paging in progress(%d) < b_npages(%d)", 3758 obj->paging_in_progress, bp->b_npages)); 3759 3760 vp = bp->b_vp; 3761 KASSERT(vp->v_holdcnt > 0, 3762 ("biodone_finish: vnode %p has zero hold count", vp)); 3763 KASSERT(vp->v_object != NULL, 3764 ("biodone_finish: vnode %p has no vm_object", vp)); 3765 3766 foff = bp->b_offset; 3767 KASSERT(bp->b_offset != NOOFFSET, 3768 ("biodone_finish: bp %p has no buffer offset", bp)); 3769 3770 /* 3771 * Set B_CACHE if the op was a normal read and no error 3772 * occured. B_CACHE is set for writes in the b*write() 3773 * routines. 3774 */ 3775 iosize = bp->b_bcount - bp->b_resid; 3776 if (bp->b_iocmd == BIO_READ && 3777 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3778 !(bp->b_ioflags & BIO_ERROR)) { 3779 bp->b_flags |= B_CACHE; 3780 } 3781 bogus = 0; 3782 VM_OBJECT_WLOCK(obj); 3783 for (i = 0; i < bp->b_npages; i++) { 3784 int bogusflag = 0; 3785 int resid; 3786 3787 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3788 if (resid > iosize) 3789 resid = iosize; 3790 3791 /* 3792 * cleanup bogus pages, restoring the originals 3793 */ 3794 m = bp->b_pages[i]; 3795 if (m == bogus_page) { 3796 bogus = bogusflag = 1; 3797 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3798 if (m == NULL) 3799 panic("biodone: page disappeared!"); 3800 bp->b_pages[i] = m; 3801 } 3802 KASSERT(OFF_TO_IDX(foff) == m->pindex, 3803 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch", 3804 (intmax_t)foff, (uintmax_t)m->pindex)); 3805 3806 /* 3807 * In the write case, the valid and clean bits are 3808 * already changed correctly ( see bdwrite() ), so we 3809 * only need to do this here in the read case. 3810 */ 3811 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3812 KASSERT((m->dirty & vm_page_bits(foff & 3813 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3814 " page %p has unexpected dirty bits", m)); 3815 vfs_page_set_valid(bp, foff, m); 3816 } 3817 3818 vm_page_io_finish(m); 3819 vm_object_pip_subtract(obj, 1); 3820 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3821 iosize -= resid; 3822 } 3823 vm_object_pip_wakeupn(obj, 0); 3824 VM_OBJECT_WUNLOCK(obj); 3825 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 3826 BUF_CHECK_MAPPED(bp); 3827 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3828 bp->b_pages, bp->b_npages); 3829 } 3830 } 3831 3832 /* 3833 * For asynchronous completions, release the buffer now. The brelse 3834 * will do a wakeup there if necessary - so no need to do a wakeup 3835 * here in the async case. The sync case always needs to do a wakeup. 3836 */ 3837 3838 if (bp->b_flags & B_ASYNC) { 3839 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3840 brelse(bp); 3841 else 3842 bqrelse(bp); 3843 } else 3844 bdone(bp); 3845} 3846 3847/* 3848 * This routine is called in lieu of iodone in the case of 3849 * incomplete I/O. This keeps the busy status for pages 3850 * consistant. 3851 */ 3852void 3853vfs_unbusy_pages(struct buf *bp) 3854{ 3855 int i; 3856 vm_object_t obj; 3857 vm_page_t m; 3858 3859 runningbufwakeup(bp); 3860 if (!(bp->b_flags & B_VMIO)) 3861 return; 3862 3863 obj = bp->b_bufobj->bo_object; 3864 VM_OBJECT_WLOCK(obj); 3865 for (i = 0; i < bp->b_npages; i++) { 3866 m = bp->b_pages[i]; 3867 if (m == bogus_page) { 3868 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3869 if (!m) 3870 panic("vfs_unbusy_pages: page missing\n"); 3871 bp->b_pages[i] = m; 3872 if ((bp->b_flags & B_UNMAPPED) == 0) { 3873 BUF_CHECK_MAPPED(bp); 3874 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3875 bp->b_pages, bp->b_npages); 3876 } else 3877 BUF_CHECK_UNMAPPED(bp); 3878 } 3879 vm_object_pip_subtract(obj, 1); 3880 vm_page_io_finish(m); 3881 } 3882 vm_object_pip_wakeupn(obj, 0); 3883 VM_OBJECT_WUNLOCK(obj); 3884} 3885 3886/* 3887 * vfs_page_set_valid: 3888 * 3889 * Set the valid bits in a page based on the supplied offset. The 3890 * range is restricted to the buffer's size. 3891 * 3892 * This routine is typically called after a read completes. 3893 */ 3894static void 3895vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3896{ 3897 vm_ooffset_t eoff; 3898 3899 /* 3900 * Compute the end offset, eoff, such that [off, eoff) does not span a 3901 * page boundary and eoff is not greater than the end of the buffer. 3902 * The end of the buffer, in this case, is our file EOF, not the 3903 * allocation size of the buffer. 3904 */ 3905 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3906 if (eoff > bp->b_offset + bp->b_bcount) 3907 eoff = bp->b_offset + bp->b_bcount; 3908 3909 /* 3910 * Set valid range. This is typically the entire buffer and thus the 3911 * entire page. 3912 */ 3913 if (eoff > off) 3914 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 3915} 3916 3917/* 3918 * vfs_page_set_validclean: 3919 * 3920 * Set the valid bits and clear the dirty bits in a page based on the 3921 * supplied offset. The range is restricted to the buffer's size. 3922 */ 3923static void 3924vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3925{ 3926 vm_ooffset_t soff, eoff; 3927 3928 /* 3929 * Start and end offsets in buffer. eoff - soff may not cross a 3930 * page boundry or cross the end of the buffer. The end of the 3931 * buffer, in this case, is our file EOF, not the allocation size 3932 * of the buffer. 3933 */ 3934 soff = off; 3935 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3936 if (eoff > bp->b_offset + bp->b_bcount) 3937 eoff = bp->b_offset + bp->b_bcount; 3938 3939 /* 3940 * Set valid range. This is typically the entire buffer and thus the 3941 * entire page. 3942 */ 3943 if (eoff > soff) { 3944 vm_page_set_validclean( 3945 m, 3946 (vm_offset_t) (soff & PAGE_MASK), 3947 (vm_offset_t) (eoff - soff) 3948 ); 3949 } 3950} 3951 3952/* 3953 * Ensure that all buffer pages are not busied by VPO_BUSY flag. If 3954 * any page is busy, drain the flag. 3955 */ 3956static void 3957vfs_drain_busy_pages(struct buf *bp) 3958{ 3959 vm_page_t m; 3960 int i, last_busied; 3961 3962 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); 3963 last_busied = 0; 3964 for (i = 0; i < bp->b_npages; i++) { 3965 m = bp->b_pages[i]; 3966 if ((m->oflags & VPO_BUSY) != 0) { 3967 for (; last_busied < i; last_busied++) 3968 vm_page_busy(bp->b_pages[last_busied]); 3969 while ((m->oflags & VPO_BUSY) != 0) 3970 vm_page_sleep(m, "vbpage"); 3971 } 3972 } 3973 for (i = 0; i < last_busied; i++) 3974 vm_page_wakeup(bp->b_pages[i]); 3975} 3976 3977/* 3978 * This routine is called before a device strategy routine. 3979 * It is used to tell the VM system that paging I/O is in 3980 * progress, and treat the pages associated with the buffer 3981 * almost as being VPO_BUSY. Also the object paging_in_progress 3982 * flag is handled to make sure that the object doesn't become 3983 * inconsistant. 3984 * 3985 * Since I/O has not been initiated yet, certain buffer flags 3986 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3987 * and should be ignored. 3988 */ 3989void 3990vfs_busy_pages(struct buf *bp, int clear_modify) 3991{ 3992 int i, bogus; 3993 vm_object_t obj; 3994 vm_ooffset_t foff; 3995 vm_page_t m; 3996 3997 if (!(bp->b_flags & B_VMIO)) 3998 return; 3999 4000 obj = bp->b_bufobj->bo_object; 4001 foff = bp->b_offset; 4002 KASSERT(bp->b_offset != NOOFFSET, 4003 ("vfs_busy_pages: no buffer offset")); 4004 VM_OBJECT_WLOCK(obj); 4005 vfs_drain_busy_pages(bp); 4006 if (bp->b_bufsize != 0) 4007 vfs_setdirty_locked_object(bp); 4008 bogus = 0; 4009 for (i = 0; i < bp->b_npages; i++) { 4010 m = bp->b_pages[i]; 4011 4012 if ((bp->b_flags & B_CLUSTER) == 0) { 4013 vm_object_pip_add(obj, 1); 4014 vm_page_io_start(m); 4015 } 4016 /* 4017 * When readying a buffer for a read ( i.e 4018 * clear_modify == 0 ), it is important to do 4019 * bogus_page replacement for valid pages in 4020 * partially instantiated buffers. Partially 4021 * instantiated buffers can, in turn, occur when 4022 * reconstituting a buffer from its VM backing store 4023 * base. We only have to do this if B_CACHE is 4024 * clear ( which causes the I/O to occur in the 4025 * first place ). The replacement prevents the read 4026 * I/O from overwriting potentially dirty VM-backed 4027 * pages. XXX bogus page replacement is, uh, bogus. 4028 * It may not work properly with small-block devices. 4029 * We need to find a better way. 4030 */ 4031 if (clear_modify) { 4032 pmap_remove_write(m); 4033 vfs_page_set_validclean(bp, foff, m); 4034 } else if (m->valid == VM_PAGE_BITS_ALL && 4035 (bp->b_flags & B_CACHE) == 0) { 4036 bp->b_pages[i] = bogus_page; 4037 bogus++; 4038 } 4039 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4040 } 4041 VM_OBJECT_WUNLOCK(obj); 4042 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 4043 BUF_CHECK_MAPPED(bp); 4044 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4045 bp->b_pages, bp->b_npages); 4046 } 4047} 4048 4049/* 4050 * vfs_bio_set_valid: 4051 * 4052 * Set the range within the buffer to valid. The range is 4053 * relative to the beginning of the buffer, b_offset. Note that 4054 * b_offset itself may be offset from the beginning of the first 4055 * page. 4056 */ 4057void 4058vfs_bio_set_valid(struct buf *bp, int base, int size) 4059{ 4060 int i, n; 4061 vm_page_t m; 4062 4063 if (!(bp->b_flags & B_VMIO)) 4064 return; 4065 4066 /* 4067 * Fixup base to be relative to beginning of first page. 4068 * Set initial n to be the maximum number of bytes in the 4069 * first page that can be validated. 4070 */ 4071 base += (bp->b_offset & PAGE_MASK); 4072 n = PAGE_SIZE - (base & PAGE_MASK); 4073 4074 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4075 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4076 m = bp->b_pages[i]; 4077 if (n > size) 4078 n = size; 4079 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4080 base += n; 4081 size -= n; 4082 n = PAGE_SIZE; 4083 } 4084 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4085} 4086 4087/* 4088 * vfs_bio_clrbuf: 4089 * 4090 * If the specified buffer is a non-VMIO buffer, clear the entire 4091 * buffer. If the specified buffer is a VMIO buffer, clear and 4092 * validate only the previously invalid portions of the buffer. 4093 * This routine essentially fakes an I/O, so we need to clear 4094 * BIO_ERROR and B_INVAL. 4095 * 4096 * Note that while we only theoretically need to clear through b_bcount, 4097 * we go ahead and clear through b_bufsize. 4098 */ 4099void 4100vfs_bio_clrbuf(struct buf *bp) 4101{ 4102 int i, j, mask, sa, ea, slide; 4103 4104 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4105 clrbuf(bp); 4106 return; 4107 } 4108 bp->b_flags &= ~B_INVAL; 4109 bp->b_ioflags &= ~BIO_ERROR; 4110 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4111 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4112 (bp->b_offset & PAGE_MASK) == 0) { 4113 if (bp->b_pages[0] == bogus_page) 4114 goto unlock; 4115 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4116 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object); 4117 if ((bp->b_pages[0]->valid & mask) == mask) 4118 goto unlock; 4119 if ((bp->b_pages[0]->valid & mask) == 0) { 4120 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize); 4121 bp->b_pages[0]->valid |= mask; 4122 goto unlock; 4123 } 4124 } 4125 sa = bp->b_offset & PAGE_MASK; 4126 slide = 0; 4127 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4128 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4129 ea = slide & PAGE_MASK; 4130 if (ea == 0) 4131 ea = PAGE_SIZE; 4132 if (bp->b_pages[i] == bogus_page) 4133 continue; 4134 j = sa / DEV_BSIZE; 4135 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4136 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object); 4137 if ((bp->b_pages[i]->valid & mask) == mask) 4138 continue; 4139 if ((bp->b_pages[i]->valid & mask) == 0) 4140 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4141 else { 4142 for (; sa < ea; sa += DEV_BSIZE, j++) { 4143 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4144 pmap_zero_page_area(bp->b_pages[i], 4145 sa, DEV_BSIZE); 4146 } 4147 } 4148 } 4149 bp->b_pages[i]->valid |= mask; 4150 } 4151unlock: 4152 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4153 bp->b_resid = 0; 4154} 4155 4156void 4157vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4158{ 4159 vm_page_t m; 4160 int i, n; 4161 4162 if ((bp->b_flags & B_UNMAPPED) == 0) { 4163 BUF_CHECK_MAPPED(bp); 4164 bzero(bp->b_data + base, size); 4165 } else { 4166 BUF_CHECK_UNMAPPED(bp); 4167 n = PAGE_SIZE - (base & PAGE_MASK); 4168 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4169 m = bp->b_pages[i]; 4170 if (n > size) 4171 n = size; 4172 pmap_zero_page_area(m, base & PAGE_MASK, n); 4173 base += n; 4174 size -= n; 4175 n = PAGE_SIZE; 4176 } 4177 } 4178} 4179 4180/* 4181 * vm_hold_load_pages and vm_hold_free_pages get pages into 4182 * a buffers address space. The pages are anonymous and are 4183 * not associated with a file object. 4184 */ 4185static void 4186vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4187{ 4188 vm_offset_t pg; 4189 vm_page_t p; 4190 int index; 4191 4192 BUF_CHECK_MAPPED(bp); 4193 4194 to = round_page(to); 4195 from = round_page(from); 4196 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4197 4198 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4199tryagain: 4200 /* 4201 * note: must allocate system pages since blocking here 4202 * could interfere with paging I/O, no matter which 4203 * process we are. 4204 */ 4205 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | 4206 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT)); 4207 if (p == NULL) { 4208 VM_WAIT; 4209 goto tryagain; 4210 } 4211 pmap_qenter(pg, &p, 1); 4212 bp->b_pages[index] = p; 4213 } 4214 bp->b_npages = index; 4215} 4216 4217/* Return pages associated with this buf to the vm system */ 4218static void 4219vm_hold_free_pages(struct buf *bp, int newbsize) 4220{ 4221 vm_offset_t from; 4222 vm_page_t p; 4223 int index, newnpages; 4224 4225 BUF_CHECK_MAPPED(bp); 4226 4227 from = round_page((vm_offset_t)bp->b_data + newbsize); 4228 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4229 if (bp->b_npages > newnpages) 4230 pmap_qremove(from, bp->b_npages - newnpages); 4231 for (index = newnpages; index < bp->b_npages; index++) { 4232 p = bp->b_pages[index]; 4233 bp->b_pages[index] = NULL; 4234 if (p->busy != 0) 4235 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 4236 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); 4237 p->wire_count--; 4238 vm_page_free(p); 4239 atomic_subtract_int(&cnt.v_wire_count, 1); 4240 } 4241 bp->b_npages = newnpages; 4242} 4243 4244/* 4245 * Map an IO request into kernel virtual address space. 4246 * 4247 * All requests are (re)mapped into kernel VA space. 4248 * Notice that we use b_bufsize for the size of the buffer 4249 * to be mapped. b_bcount might be modified by the driver. 4250 * 4251 * Note that even if the caller determines that the address space should 4252 * be valid, a race or a smaller-file mapped into a larger space may 4253 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 4254 * check the return value. 4255 */ 4256int 4257vmapbuf(struct buf *bp, int mapbuf) 4258{ 4259 caddr_t kva; 4260 vm_prot_t prot; 4261 int pidx; 4262 4263 if (bp->b_bufsize < 0) 4264 return (-1); 4265 prot = VM_PROT_READ; 4266 if (bp->b_iocmd == BIO_READ) 4267 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 4268 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 4269 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, 4270 btoc(MAXPHYS))) < 0) 4271 return (-1); 4272 bp->b_npages = pidx; 4273 if (mapbuf || !unmapped_buf_allowed) { 4274 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 4275 kva = bp->b_saveaddr; 4276 bp->b_saveaddr = bp->b_data; 4277 bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK); 4278 bp->b_flags &= ~B_UNMAPPED; 4279 } else { 4280 bp->b_flags |= B_UNMAPPED; 4281 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK; 4282 bp->b_saveaddr = bp->b_data; 4283 bp->b_data = unmapped_buf; 4284 } 4285 return(0); 4286} 4287 4288/* 4289 * Free the io map PTEs associated with this IO operation. 4290 * We also invalidate the TLB entries and restore the original b_addr. 4291 */ 4292void 4293vunmapbuf(struct buf *bp) 4294{ 4295 int npages; 4296 4297 npages = bp->b_npages; 4298 if (bp->b_flags & B_UNMAPPED) 4299 bp->b_flags &= ~B_UNMAPPED; 4300 else 4301 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4302 vm_page_unhold_pages(bp->b_pages, npages); 4303 4304 bp->b_data = bp->b_saveaddr; 4305} 4306 4307void 4308bdone(struct buf *bp) 4309{ 4310 struct mtx *mtxp; 4311 4312 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4313 mtx_lock(mtxp); 4314 bp->b_flags |= B_DONE; 4315 wakeup(bp); 4316 mtx_unlock(mtxp); 4317} 4318 4319void 4320bwait(struct buf *bp, u_char pri, const char *wchan) 4321{ 4322 struct mtx *mtxp; 4323 4324 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4325 mtx_lock(mtxp); 4326 while ((bp->b_flags & B_DONE) == 0) 4327 msleep(bp, mtxp, pri, wchan, 0); 4328 mtx_unlock(mtxp); 4329} 4330 4331int 4332bufsync(struct bufobj *bo, int waitfor) 4333{ 4334 4335 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 4336} 4337 4338void 4339bufstrategy(struct bufobj *bo, struct buf *bp) 4340{ 4341 int i = 0; 4342 struct vnode *vp; 4343 4344 vp = bp->b_vp; 4345 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 4346 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 4347 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 4348 i = VOP_STRATEGY(vp, bp); 4349 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 4350} 4351 4352void 4353bufobj_wrefl(struct bufobj *bo) 4354{ 4355 4356 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4357 ASSERT_BO_WLOCKED(bo); 4358 bo->bo_numoutput++; 4359} 4360 4361void 4362bufobj_wref(struct bufobj *bo) 4363{ 4364 4365 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4366 BO_LOCK(bo); 4367 bo->bo_numoutput++; 4368 BO_UNLOCK(bo); 4369} 4370 4371void 4372bufobj_wdrop(struct bufobj *bo) 4373{ 4374 4375 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 4376 BO_LOCK(bo); 4377 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 4378 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 4379 bo->bo_flag &= ~BO_WWAIT; 4380 wakeup(&bo->bo_numoutput); 4381 } 4382 BO_UNLOCK(bo); 4383} 4384 4385int 4386bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 4387{ 4388 int error; 4389 4390 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 4391 ASSERT_BO_WLOCKED(bo); 4392 error = 0; 4393 while (bo->bo_numoutput) { 4394 bo->bo_flag |= BO_WWAIT; 4395 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), 4396 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 4397 if (error) 4398 break; 4399 } 4400 return (error); 4401} 4402 4403void 4404bpin(struct buf *bp) 4405{ 4406 struct mtx *mtxp; 4407 4408 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4409 mtx_lock(mtxp); 4410 bp->b_pin_count++; 4411 mtx_unlock(mtxp); 4412} 4413 4414void 4415bunpin(struct buf *bp) 4416{ 4417 struct mtx *mtxp; 4418 4419 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4420 mtx_lock(mtxp); 4421 if (--bp->b_pin_count == 0) 4422 wakeup(bp); 4423 mtx_unlock(mtxp); 4424} 4425 4426void 4427bunpin_wait(struct buf *bp) 4428{ 4429 struct mtx *mtxp; 4430 4431 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4432 mtx_lock(mtxp); 4433 while (bp->b_pin_count > 0) 4434 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4435 mtx_unlock(mtxp); 4436} 4437 4438/* 4439 * Set bio_data or bio_ma for struct bio from the struct buf. 4440 */ 4441void 4442bdata2bio(struct buf *bp, struct bio *bip) 4443{ 4444 4445 if ((bp->b_flags & B_UNMAPPED) != 0) { 4446 KASSERT(unmapped_buf_allowed, ("unmapped")); 4447 bip->bio_ma = bp->b_pages; 4448 bip->bio_ma_n = bp->b_npages; 4449 bip->bio_data = unmapped_buf; 4450 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 4451 bip->bio_flags |= BIO_UNMAPPED; 4452 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 4453 PAGE_SIZE == bp->b_npages, 4454 ("Buffer %p too short: %d %d %d", bp, bip->bio_ma_offset, 4455 bip->bio_length, bip->bio_ma_n)); 4456 } else { 4457 bip->bio_data = bp->b_data; 4458 bip->bio_ma = NULL; 4459 } 4460} 4461 4462#include "opt_ddb.h" 4463#ifdef DDB 4464#include <ddb/ddb.h> 4465 4466/* DDB command to show buffer data */ 4467DB_SHOW_COMMAND(buffer, db_show_buffer) 4468{ 4469 /* get args */ 4470 struct buf *bp = (struct buf *)addr; 4471 4472 if (!have_addr) { 4473 db_printf("usage: show buffer <addr>\n"); 4474 return; 4475 } 4476 4477 db_printf("buf at %p\n", bp); 4478 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n", 4479 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, 4480 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS); 4481 db_printf( 4482 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4483 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, " 4484 "b_dep = %p\n", 4485 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4486 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4487 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first); 4488 if (bp->b_npages) { 4489 int i; 4490 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4491 for (i = 0; i < bp->b_npages; i++) { 4492 vm_page_t m; 4493 m = bp->b_pages[i]; 4494 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4495 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4496 if ((i + 1) < bp->b_npages) 4497 db_printf(","); 4498 } 4499 db_printf("\n"); 4500 } 4501 db_printf(" "); 4502 BUF_LOCKPRINTINFO(bp); 4503} 4504 4505DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4506{ 4507 struct buf *bp; 4508 int i; 4509 4510 for (i = 0; i < nbuf; i++) { 4511 bp = &buf[i]; 4512 if (BUF_ISLOCKED(bp)) { 4513 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4514 db_printf("\n"); 4515 } 4516 } 4517} 4518 4519DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4520{ 4521 struct vnode *vp; 4522 struct buf *bp; 4523 4524 if (!have_addr) { 4525 db_printf("usage: show vnodebufs <addr>\n"); 4526 return; 4527 } 4528 vp = (struct vnode *)addr; 4529 db_printf("Clean buffers:\n"); 4530 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4531 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4532 db_printf("\n"); 4533 } 4534 db_printf("Dirty buffers:\n"); 4535 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4536 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4537 db_printf("\n"); 4538 } 4539} 4540 4541DB_COMMAND(countfreebufs, db_coundfreebufs) 4542{ 4543 struct buf *bp; 4544 int i, used = 0, nfree = 0; 4545 4546 if (have_addr) { 4547 db_printf("usage: countfreebufs\n"); 4548 return; 4549 } 4550 4551 for (i = 0; i < nbuf; i++) { 4552 bp = &buf[i]; 4553 if ((bp->b_flags & B_INFREECNT) != 0) 4554 nfree++; 4555 else 4556 used++; 4557 } 4558 4559 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4560 nfree + used); 4561 db_printf("numfreebuffers is %d\n", numfreebuffers); 4562} 4563#endif /* DDB */ 4564