vfs_bio.c revision 248515
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 248515 2013-03-19 14:43:57Z 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_vflags & BV_INFREECNT) == 0, 422 ("buf %p already counted as free", bp)); 423 if (bp->b_bufobj != NULL) 424 mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED); 425 bp->b_vflags |= BV_INFREECNT; 426 old = atomic_fetchadd_int(&numfreebuffers, 1); 427 KASSERT(old >= 0 && old < nbuf, 428 ("numfreebuffers climbed to %d", old + 1)); 429 mtx_lock(&nblock); 430 if (needsbuffer) { 431 needsbuffer &= ~VFS_BIO_NEED_ANY; 432 if (numfreebuffers >= hifreebuffers) 433 needsbuffer &= ~VFS_BIO_NEED_FREE; 434 wakeup(&needsbuffer); 435 } 436 mtx_unlock(&nblock); 437} 438 439/* 440 * waitrunningbufspace() 441 * 442 * runningbufspace is a measure of the amount of I/O currently 443 * running. This routine is used in async-write situations to 444 * prevent creating huge backups of pending writes to a device. 445 * Only asynchronous writes are governed by this function. 446 * 447 * Reads will adjust runningbufspace, but will not block based on it. 448 * The read load has a side effect of reducing the allowed write load. 449 * 450 * This does NOT turn an async write into a sync write. It waits 451 * for earlier writes to complete and generally returns before the 452 * caller's write has reached the device. 453 */ 454void 455waitrunningbufspace(void) 456{ 457 458 mtx_lock(&rbreqlock); 459 while (runningbufspace > hirunningspace) { 460 ++runningbufreq; 461 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 462 } 463 mtx_unlock(&rbreqlock); 464} 465 466 467/* 468 * vfs_buf_test_cache: 469 * 470 * Called when a buffer is extended. This function clears the B_CACHE 471 * bit if the newly extended portion of the buffer does not contain 472 * valid data. 473 */ 474static __inline 475void 476vfs_buf_test_cache(struct buf *bp, 477 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 478 vm_page_t m) 479{ 480 481 VM_OBJECT_ASSERT_WLOCKED(m->object); 482 if (bp->b_flags & B_CACHE) { 483 int base = (foff + off) & PAGE_MASK; 484 if (vm_page_is_valid(m, base, size) == 0) 485 bp->b_flags &= ~B_CACHE; 486 } 487} 488 489/* Wake up the buffer daemon if necessary */ 490static __inline 491void 492bd_wakeup(int dirtybuflevel) 493{ 494 495 mtx_lock(&bdlock); 496 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 497 bd_request = 1; 498 wakeup(&bd_request); 499 } 500 mtx_unlock(&bdlock); 501} 502 503/* 504 * bd_speedup - speedup the buffer cache flushing code 505 */ 506 507void 508bd_speedup(void) 509{ 510 int needwake; 511 512 mtx_lock(&bdlock); 513 needwake = 0; 514 if (bd_speedupreq == 0 || bd_request == 0) 515 needwake = 1; 516 bd_speedupreq = 1; 517 bd_request = 1; 518 if (needwake) 519 wakeup(&bd_request); 520 mtx_unlock(&bdlock); 521} 522 523/* 524 * Calculating buffer cache scaling values and reserve space for buffer 525 * headers. This is called during low level kernel initialization and 526 * may be called more then once. We CANNOT write to the memory area 527 * being reserved at this time. 528 */ 529caddr_t 530kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 531{ 532 int tuned_nbuf; 533 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz; 534 535 /* 536 * physmem_est is in pages. Convert it to kilobytes (assumes 537 * PAGE_SIZE is >= 1K) 538 */ 539 physmem_est = physmem_est * (PAGE_SIZE / 1024); 540 541 /* 542 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 543 * For the first 64MB of ram nominally allocate sufficient buffers to 544 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 545 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 546 * the buffer cache we limit the eventual kva reservation to 547 * maxbcache bytes. 548 * 549 * factor represents the 1/4 x ram conversion. 550 */ 551 if (nbuf == 0) { 552 int factor = 4 * BKVASIZE / 1024; 553 554 nbuf = 50; 555 if (physmem_est > 4096) 556 nbuf += min((physmem_est - 4096) / factor, 557 65536 / factor); 558 if (physmem_est > 65536) 559 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 560 561 if (maxbcache && nbuf > maxbcache / BKVASIZE) 562 nbuf = maxbcache / BKVASIZE; 563 tuned_nbuf = 1; 564 } else 565 tuned_nbuf = 0; 566 567 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 568 maxbuf = (LONG_MAX / 3) / BKVASIZE; 569 if (nbuf > maxbuf) { 570 if (!tuned_nbuf) 571 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 572 maxbuf); 573 nbuf = maxbuf; 574 } 575 576 /* 577 * Ideal allocation size for the transient bio submap if 10% 578 * of the maximal space buffer map. This roughly corresponds 579 * to the amount of the buffer mapped for typical UFS load. 580 * 581 * Clip the buffer map to reserve space for the transient 582 * BIOs, if its extent is bigger than 90% of the maximum 583 * buffer map extent on the platform. 584 * 585 * The fall-back to the maxbuf in case of maxbcache unset, 586 * allows to not trim the buffer KVA for the architectures 587 * with ample KVA space. 588 */ 589 if (bio_transient_maxcnt == 0) { 590 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE; 591 buf_sz = (long)nbuf * BKVASIZE; 592 if (buf_sz < maxbuf_sz / 10 * 9) { 593 /* 594 * There is more KVA than memory. Do not 595 * adjust buffer map size, and assign the rest 596 * of maxbuf to transient map. 597 */ 598 biotmap_sz = maxbuf_sz - buf_sz; 599 } else { 600 /* 601 * Buffer map spans all KVA we could afford on 602 * this platform. Give 10% of the buffer map 603 * to the transient bio map. 604 */ 605 biotmap_sz = buf_sz / 10; 606 buf_sz -= biotmap_sz; 607 } 608 if (biotmap_sz / INT_MAX > MAXPHYS) 609 bio_transient_maxcnt = INT_MAX; 610 else 611 bio_transient_maxcnt = biotmap_sz / MAXPHYS; 612 /* 613 * Artifically limit to 1024 simultaneous in-flight I/Os 614 * using the transient mapping. 615 */ 616 if (bio_transient_maxcnt > 1024) 617 bio_transient_maxcnt = 1024; 618 if (tuned_nbuf) 619 nbuf = buf_sz / BKVASIZE; 620 } 621 622 /* 623 * swbufs are used as temporary holders for I/O, such as paging I/O. 624 * We have no less then 16 and no more then 256. 625 */ 626 nswbuf = max(min(nbuf/4, 256), 16); 627#ifdef NSWBUF_MIN 628 if (nswbuf < NSWBUF_MIN) 629 nswbuf = NSWBUF_MIN; 630#endif 631#ifdef DIRECTIO 632 ffs_rawread_setup(); 633#endif 634 635 /* 636 * Reserve space for the buffer cache buffers 637 */ 638 swbuf = (void *)v; 639 v = (caddr_t)(swbuf + nswbuf); 640 buf = (void *)v; 641 v = (caddr_t)(buf + nbuf); 642 643 return(v); 644} 645 646/* Initialize the buffer subsystem. Called before use of any buffers. */ 647void 648bufinit(void) 649{ 650 struct buf *bp; 651 int i; 652 653 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 654 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 655 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 656 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 657 658 /* next, make a null set of free lists */ 659 for (i = 0; i < BUFFER_QUEUES; i++) 660 TAILQ_INIT(&bufqueues[i]); 661 662 /* finally, initialize each buffer header and stick on empty q */ 663 for (i = 0; i < nbuf; i++) { 664 bp = &buf[i]; 665 bzero(bp, sizeof *bp); 666 bp->b_flags = B_INVAL; /* we're just an empty header */ 667 bp->b_rcred = NOCRED; 668 bp->b_wcred = NOCRED; 669 bp->b_qindex = QUEUE_EMPTY; 670 bp->b_vflags = BV_INFREECNT; /* buf is counted as free */ 671 bp->b_xflags = 0; 672 LIST_INIT(&bp->b_dep); 673 BUF_LOCKINIT(bp); 674 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 675#ifdef INVARIANTS 676 bq_len[QUEUE_EMPTY]++; 677#endif 678 } 679 680 /* 681 * maxbufspace is the absolute maximum amount of buffer space we are 682 * allowed to reserve in KVM and in real terms. The absolute maximum 683 * is nominally used by buf_daemon. hibufspace is the nominal maximum 684 * used by most other processes. The differential is required to 685 * ensure that buf_daemon is able to run when other processes might 686 * be blocked waiting for buffer space. 687 * 688 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 689 * this may result in KVM fragmentation which is not handled optimally 690 * by the system. 691 */ 692 maxbufspace = (long)nbuf * BKVASIZE; 693 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 694 lobufspace = hibufspace - MAXBSIZE; 695 696 /* 697 * Note: The 16 MiB upper limit for hirunningspace was chosen 698 * arbitrarily and may need further tuning. It corresponds to 699 * 128 outstanding write IO requests (if IO size is 128 KiB), 700 * which fits with many RAID controllers' tagged queuing limits. 701 * The lower 1 MiB limit is the historical upper limit for 702 * hirunningspace. 703 */ 704 hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE), 705 16 * 1024 * 1024), 1024 * 1024); 706 lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE); 707 708/* 709 * Limit the amount of malloc memory since it is wired permanently into 710 * the kernel space. Even though this is accounted for in the buffer 711 * allocation, we don't want the malloced region to grow uncontrolled. 712 * The malloc scheme improves memory utilization significantly on average 713 * (small) directories. 714 */ 715 maxbufmallocspace = hibufspace / 20; 716 717/* 718 * Reduce the chance of a deadlock occuring by limiting the number 719 * of delayed-write dirty buffers we allow to stack up. 720 */ 721 hidirtybuffers = nbuf / 4 + 20; 722 dirtybufthresh = hidirtybuffers * 9 / 10; 723 numdirtybuffers = 0; 724/* 725 * To support extreme low-memory systems, make sure hidirtybuffers cannot 726 * eat up all available buffer space. This occurs when our minimum cannot 727 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 728 * BKVASIZE'd buffers. 729 */ 730 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 731 hidirtybuffers >>= 1; 732 } 733 lodirtybuffers = hidirtybuffers / 2; 734 735/* 736 * Try to keep the number of free buffers in the specified range, 737 * and give special processes (e.g. like buf_daemon) access to an 738 * emergency reserve. 739 */ 740 lofreebuffers = nbuf / 18 + 5; 741 hifreebuffers = 2 * lofreebuffers; 742 numfreebuffers = nbuf; 743 744 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 745 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 746 unmapped_buf = (caddr_t)kmem_alloc_nofault(kernel_map, MAXPHYS); 747} 748 749#ifdef INVARIANTS 750static inline void 751vfs_buf_check_mapped(struct buf *bp) 752{ 753 754 KASSERT((bp->b_flags & B_UNMAPPED) == 0, 755 ("mapped buf %p %x", bp, bp->b_flags)); 756 KASSERT(bp->b_kvabase != unmapped_buf, 757 ("mapped buf: b_kvabase was not updated %p", bp)); 758 KASSERT(bp->b_data != unmapped_buf, 759 ("mapped buf: b_data was not updated %p", bp)); 760} 761 762static inline void 763vfs_buf_check_unmapped(struct buf *bp) 764{ 765 766 KASSERT((bp->b_flags & B_UNMAPPED) == B_UNMAPPED, 767 ("unmapped buf %p %x", bp, bp->b_flags)); 768 KASSERT(bp->b_kvabase == unmapped_buf, 769 ("unmapped buf: corrupted b_kvabase %p", bp)); 770 KASSERT(bp->b_data == unmapped_buf, 771 ("unmapped buf: corrupted b_data %p", bp)); 772} 773 774#define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) 775#define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) 776#else 777#define BUF_CHECK_MAPPED(bp) do {} while (0) 778#define BUF_CHECK_UNMAPPED(bp) do {} while (0) 779#endif 780 781static void 782bpmap_qenter(struct buf *bp) 783{ 784 785 BUF_CHECK_MAPPED(bp); 786 787 /* 788 * bp->b_data is relative to bp->b_offset, but 789 * bp->b_offset may be offset into the first page. 790 */ 791 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 792 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 793 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 794 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 795} 796 797/* 798 * bfreekva() - free the kva allocation for a buffer. 799 * 800 * Since this call frees up buffer space, we call bufspacewakeup(). 801 */ 802static void 803bfreekva(struct buf *bp) 804{ 805 806 if (bp->b_kvasize == 0) 807 return; 808 809 atomic_add_int(&buffreekvacnt, 1); 810 atomic_subtract_long(&bufspace, bp->b_kvasize); 811 if ((bp->b_flags & B_UNMAPPED) == 0) { 812 BUF_CHECK_MAPPED(bp); 813 vm_map_remove(buffer_map, (vm_offset_t)bp->b_kvabase, 814 (vm_offset_t)bp->b_kvabase + bp->b_kvasize); 815 } else { 816 BUF_CHECK_UNMAPPED(bp); 817 if ((bp->b_flags & B_KVAALLOC) != 0) { 818 vm_map_remove(buffer_map, (vm_offset_t)bp->b_kvaalloc, 819 (vm_offset_t)bp->b_kvaalloc + bp->b_kvasize); 820 } 821 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize); 822 bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC); 823 } 824 bp->b_kvasize = 0; 825 bufspacewakeup(); 826} 827 828/* 829 * bremfree: 830 * 831 * Mark the buffer for removal from the appropriate free list in brelse. 832 * 833 */ 834void 835bremfree(struct buf *bp) 836{ 837 int old; 838 839 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 840 KASSERT((bp->b_flags & B_REMFREE) == 0, 841 ("bremfree: buffer %p already marked for delayed removal.", bp)); 842 KASSERT(bp->b_qindex != QUEUE_NONE, 843 ("bremfree: buffer %p not on a queue.", bp)); 844 BUF_ASSERT_HELD(bp); 845 846 bp->b_flags |= B_REMFREE; 847 /* Fixup numfreebuffers count. */ 848 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 849 KASSERT((bp->b_vflags & BV_INFREECNT) != 0, 850 ("buf %p not counted in numfreebuffers", bp)); 851 if (bp->b_bufobj != NULL) 852 mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED); 853 bp->b_vflags &= ~BV_INFREECNT; 854 old = atomic_fetchadd_int(&numfreebuffers, -1); 855 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 856 } 857} 858 859/* 860 * bremfreef: 861 * 862 * Force an immediate removal from a free list. Used only in nfs when 863 * it abuses the b_freelist pointer. 864 */ 865void 866bremfreef(struct buf *bp) 867{ 868 mtx_lock(&bqlock); 869 bremfreel(bp); 870 mtx_unlock(&bqlock); 871} 872 873/* 874 * bremfreel: 875 * 876 * Removes a buffer from the free list, must be called with the 877 * bqlock held. 878 */ 879static void 880bremfreel(struct buf *bp) 881{ 882 int old; 883 884 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 885 bp, bp->b_vp, bp->b_flags); 886 KASSERT(bp->b_qindex != QUEUE_NONE, 887 ("bremfreel: buffer %p not on a queue.", bp)); 888 BUF_ASSERT_HELD(bp); 889 mtx_assert(&bqlock, MA_OWNED); 890 891 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 892#ifdef INVARIANTS 893 KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow", 894 bp->b_qindex)); 895 bq_len[bp->b_qindex]--; 896#endif 897 bp->b_qindex = QUEUE_NONE; 898 /* 899 * If this was a delayed bremfree() we only need to remove the buffer 900 * from the queue and return the stats are already done. 901 */ 902 if (bp->b_flags & B_REMFREE) { 903 bp->b_flags &= ~B_REMFREE; 904 return; 905 } 906 /* 907 * Fixup numfreebuffers count. If the buffer is invalid or not 908 * delayed-write, the buffer was free and we must decrement 909 * numfreebuffers. 910 */ 911 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 912 KASSERT((bp->b_vflags & BV_INFREECNT) != 0, 913 ("buf %p not counted in numfreebuffers", bp)); 914 if (bp->b_bufobj != NULL) 915 mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED); 916 bp->b_vflags &= ~BV_INFREECNT; 917 old = atomic_fetchadd_int(&numfreebuffers, -1); 918 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 919 } 920} 921 922/* 923 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 924 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 925 * the buffer is valid and we do not have to do anything. 926 */ 927void 928breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 929 int cnt, struct ucred * cred) 930{ 931 struct buf *rabp; 932 int i; 933 934 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 935 if (inmem(vp, *rablkno)) 936 continue; 937 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 938 939 if ((rabp->b_flags & B_CACHE) == 0) { 940 if (!TD_IS_IDLETHREAD(curthread)) 941 curthread->td_ru.ru_inblock++; 942 rabp->b_flags |= B_ASYNC; 943 rabp->b_flags &= ~B_INVAL; 944 rabp->b_ioflags &= ~BIO_ERROR; 945 rabp->b_iocmd = BIO_READ; 946 if (rabp->b_rcred == NOCRED && cred != NOCRED) 947 rabp->b_rcred = crhold(cred); 948 vfs_busy_pages(rabp, 0); 949 BUF_KERNPROC(rabp); 950 rabp->b_iooffset = dbtob(rabp->b_blkno); 951 bstrategy(rabp); 952 } else { 953 brelse(rabp); 954 } 955 } 956} 957 958/* 959 * Entry point for bread() and breadn() via #defines in sys/buf.h. 960 * 961 * Get a buffer with the specified data. Look in the cache first. We 962 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 963 * is set, the buffer is valid and we do not have to do anything, see 964 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 965 */ 966int 967breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno, 968 int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp) 969{ 970 struct buf *bp; 971 int rv = 0, readwait = 0; 972 973 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 974 /* 975 * Can only return NULL if GB_LOCK_NOWAIT flag is specified. 976 */ 977 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags); 978 if (bp == NULL) 979 return (EBUSY); 980 981 /* if not found in cache, do some I/O */ 982 if ((bp->b_flags & B_CACHE) == 0) { 983 if (!TD_IS_IDLETHREAD(curthread)) 984 curthread->td_ru.ru_inblock++; 985 bp->b_iocmd = BIO_READ; 986 bp->b_flags &= ~B_INVAL; 987 bp->b_ioflags &= ~BIO_ERROR; 988 if (bp->b_rcred == NOCRED && cred != NOCRED) 989 bp->b_rcred = crhold(cred); 990 vfs_busy_pages(bp, 0); 991 bp->b_iooffset = dbtob(bp->b_blkno); 992 bstrategy(bp); 993 ++readwait; 994 } 995 996 breada(vp, rablkno, rabsize, cnt, cred); 997 998 if (readwait) { 999 rv = bufwait(bp); 1000 } 1001 return (rv); 1002} 1003 1004/* 1005 * Write, release buffer on completion. (Done by iodone 1006 * if async). Do not bother writing anything if the buffer 1007 * is invalid. 1008 * 1009 * Note that we set B_CACHE here, indicating that buffer is 1010 * fully valid and thus cacheable. This is true even of NFS 1011 * now so we set it generally. This could be set either here 1012 * or in biodone() since the I/O is synchronous. We put it 1013 * here. 1014 */ 1015int 1016bufwrite(struct buf *bp) 1017{ 1018 int oldflags; 1019 struct vnode *vp; 1020 int vp_md; 1021 1022 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1023 if (bp->b_flags & B_INVAL) { 1024 brelse(bp); 1025 return (0); 1026 } 1027 1028 if (bp->b_flags & B_BARRIER) 1029 barrierwrites++; 1030 1031 oldflags = bp->b_flags; 1032 1033 BUF_ASSERT_HELD(bp); 1034 1035 if (bp->b_pin_count > 0) 1036 bunpin_wait(bp); 1037 1038 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 1039 ("FFS background buffer should not get here %p", bp)); 1040 1041 vp = bp->b_vp; 1042 if (vp) 1043 vp_md = vp->v_vflag & VV_MD; 1044 else 1045 vp_md = 0; 1046 1047 /* Mark the buffer clean */ 1048 bundirty(bp); 1049 1050 bp->b_flags &= ~B_DONE; 1051 bp->b_ioflags &= ~BIO_ERROR; 1052 bp->b_flags |= B_CACHE; 1053 bp->b_iocmd = BIO_WRITE; 1054 1055 bufobj_wref(bp->b_bufobj); 1056 vfs_busy_pages(bp, 1); 1057 1058 /* 1059 * Normal bwrites pipeline writes 1060 */ 1061 bp->b_runningbufspace = bp->b_bufsize; 1062 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 1063 1064 if (!TD_IS_IDLETHREAD(curthread)) 1065 curthread->td_ru.ru_oublock++; 1066 if (oldflags & B_ASYNC) 1067 BUF_KERNPROC(bp); 1068 bp->b_iooffset = dbtob(bp->b_blkno); 1069 bstrategy(bp); 1070 1071 if ((oldflags & B_ASYNC) == 0) { 1072 int rtval = bufwait(bp); 1073 brelse(bp); 1074 return (rtval); 1075 } else { 1076 /* 1077 * don't allow the async write to saturate the I/O 1078 * system. We will not deadlock here because 1079 * we are blocking waiting for I/O that is already in-progress 1080 * to complete. We do not block here if it is the update 1081 * or syncer daemon trying to clean up as that can lead 1082 * to deadlock. 1083 */ 1084 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 1085 waitrunningbufspace(); 1086 } 1087 1088 return (0); 1089} 1090 1091void 1092bufbdflush(struct bufobj *bo, struct buf *bp) 1093{ 1094 struct buf *nbp; 1095 1096 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 1097 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 1098 altbufferflushes++; 1099 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 1100 BO_LOCK(bo); 1101 /* 1102 * Try to find a buffer to flush. 1103 */ 1104 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 1105 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1106 BUF_LOCK(nbp, 1107 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1108 continue; 1109 if (bp == nbp) 1110 panic("bdwrite: found ourselves"); 1111 BO_UNLOCK(bo); 1112 /* Don't countdeps with the bo lock held. */ 1113 if (buf_countdeps(nbp, 0)) { 1114 BO_LOCK(bo); 1115 BUF_UNLOCK(nbp); 1116 continue; 1117 } 1118 if (nbp->b_flags & B_CLUSTEROK) { 1119 vfs_bio_awrite(nbp); 1120 } else { 1121 bremfree(nbp); 1122 bawrite(nbp); 1123 } 1124 dirtybufferflushes++; 1125 break; 1126 } 1127 if (nbp == NULL) 1128 BO_UNLOCK(bo); 1129 } 1130} 1131 1132/* 1133 * Delayed write. (Buffer is marked dirty). Do not bother writing 1134 * anything if the buffer is marked invalid. 1135 * 1136 * Note that since the buffer must be completely valid, we can safely 1137 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1138 * biodone() in order to prevent getblk from writing the buffer 1139 * out synchronously. 1140 */ 1141void 1142bdwrite(struct buf *bp) 1143{ 1144 struct thread *td = curthread; 1145 struct vnode *vp; 1146 struct bufobj *bo; 1147 1148 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1149 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1150 KASSERT((bp->b_flags & B_BARRIER) == 0, 1151 ("Barrier request in delayed write %p", bp)); 1152 BUF_ASSERT_HELD(bp); 1153 1154 if (bp->b_flags & B_INVAL) { 1155 brelse(bp); 1156 return; 1157 } 1158 1159 /* 1160 * If we have too many dirty buffers, don't create any more. 1161 * If we are wildly over our limit, then force a complete 1162 * cleanup. Otherwise, just keep the situation from getting 1163 * out of control. Note that we have to avoid a recursive 1164 * disaster and not try to clean up after our own cleanup! 1165 */ 1166 vp = bp->b_vp; 1167 bo = bp->b_bufobj; 1168 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 1169 td->td_pflags |= TDP_INBDFLUSH; 1170 BO_BDFLUSH(bo, bp); 1171 td->td_pflags &= ~TDP_INBDFLUSH; 1172 } else 1173 recursiveflushes++; 1174 1175 bdirty(bp); 1176 /* 1177 * Set B_CACHE, indicating that the buffer is fully valid. This is 1178 * true even of NFS now. 1179 */ 1180 bp->b_flags |= B_CACHE; 1181 1182 /* 1183 * This bmap keeps the system from needing to do the bmap later, 1184 * perhaps when the system is attempting to do a sync. Since it 1185 * is likely that the indirect block -- or whatever other datastructure 1186 * that the filesystem needs is still in memory now, it is a good 1187 * thing to do this. Note also, that if the pageout daemon is 1188 * requesting a sync -- there might not be enough memory to do 1189 * the bmap then... So, this is important to do. 1190 */ 1191 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1192 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1193 } 1194 1195 /* 1196 * Set the *dirty* buffer range based upon the VM system dirty 1197 * pages. 1198 * 1199 * Mark the buffer pages as clean. We need to do this here to 1200 * satisfy the vnode_pager and the pageout daemon, so that it 1201 * thinks that the pages have been "cleaned". Note that since 1202 * the pages are in a delayed write buffer -- the VFS layer 1203 * "will" see that the pages get written out on the next sync, 1204 * or perhaps the cluster will be completed. 1205 */ 1206 vfs_clean_pages_dirty_buf(bp); 1207 bqrelse(bp); 1208 1209 /* 1210 * Wakeup the buffer flushing daemon if we have a lot of dirty 1211 * buffers (midpoint between our recovery point and our stall 1212 * point). 1213 */ 1214 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1215 1216 /* 1217 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1218 * due to the softdep code. 1219 */ 1220} 1221 1222/* 1223 * bdirty: 1224 * 1225 * Turn buffer into delayed write request. We must clear BIO_READ and 1226 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1227 * itself to properly update it in the dirty/clean lists. We mark it 1228 * B_DONE to ensure that any asynchronization of the buffer properly 1229 * clears B_DONE ( else a panic will occur later ). 1230 * 1231 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1232 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1233 * should only be called if the buffer is known-good. 1234 * 1235 * Since the buffer is not on a queue, we do not update the numfreebuffers 1236 * count. 1237 * 1238 * The buffer must be on QUEUE_NONE. 1239 */ 1240void 1241bdirty(struct buf *bp) 1242{ 1243 1244 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1245 bp, bp->b_vp, bp->b_flags); 1246 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1247 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1248 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1249 BUF_ASSERT_HELD(bp); 1250 bp->b_flags &= ~(B_RELBUF); 1251 bp->b_iocmd = BIO_WRITE; 1252 1253 if ((bp->b_flags & B_DELWRI) == 0) { 1254 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1255 reassignbuf(bp); 1256 atomic_add_int(&numdirtybuffers, 1); 1257 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1258 } 1259} 1260 1261/* 1262 * bundirty: 1263 * 1264 * Clear B_DELWRI for buffer. 1265 * 1266 * Since the buffer is not on a queue, we do not update the numfreebuffers 1267 * count. 1268 * 1269 * The buffer must be on QUEUE_NONE. 1270 */ 1271 1272void 1273bundirty(struct buf *bp) 1274{ 1275 1276 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1277 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1278 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1279 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1280 BUF_ASSERT_HELD(bp); 1281 1282 if (bp->b_flags & B_DELWRI) { 1283 bp->b_flags &= ~B_DELWRI; 1284 reassignbuf(bp); 1285 atomic_subtract_int(&numdirtybuffers, 1); 1286 numdirtywakeup(lodirtybuffers); 1287 } 1288 /* 1289 * Since it is now being written, we can clear its deferred write flag. 1290 */ 1291 bp->b_flags &= ~B_DEFERRED; 1292} 1293 1294/* 1295 * bawrite: 1296 * 1297 * Asynchronous write. Start output on a buffer, but do not wait for 1298 * it to complete. The buffer is released when the output completes. 1299 * 1300 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1301 * B_INVAL buffers. Not us. 1302 */ 1303void 1304bawrite(struct buf *bp) 1305{ 1306 1307 bp->b_flags |= B_ASYNC; 1308 (void) bwrite(bp); 1309} 1310 1311/* 1312 * babarrierwrite: 1313 * 1314 * Asynchronous barrier write. Start output on a buffer, but do not 1315 * wait for it to complete. Place a write barrier after this write so 1316 * that this buffer and all buffers written before it are committed to 1317 * the disk before any buffers written after this write are committed 1318 * to the disk. The buffer is released when the output completes. 1319 */ 1320void 1321babarrierwrite(struct buf *bp) 1322{ 1323 1324 bp->b_flags |= B_ASYNC | B_BARRIER; 1325 (void) bwrite(bp); 1326} 1327 1328/* 1329 * bbarrierwrite: 1330 * 1331 * Synchronous barrier write. Start output on a buffer and wait for 1332 * it to complete. Place a write barrier after this write so that 1333 * this buffer and all buffers written before it are committed to 1334 * the disk before any buffers written after this write are committed 1335 * to the disk. The buffer is released when the output completes. 1336 */ 1337int 1338bbarrierwrite(struct buf *bp) 1339{ 1340 1341 bp->b_flags |= B_BARRIER; 1342 return (bwrite(bp)); 1343} 1344 1345/* 1346 * bwillwrite: 1347 * 1348 * Called prior to the locking of any vnodes when we are expecting to 1349 * write. We do not want to starve the buffer cache with too many 1350 * dirty buffers so we block here. By blocking prior to the locking 1351 * of any vnodes we attempt to avoid the situation where a locked vnode 1352 * prevents the various system daemons from flushing related buffers. 1353 */ 1354 1355void 1356bwillwrite(void) 1357{ 1358 1359 if (numdirtybuffers >= hidirtybuffers) { 1360 mtx_lock(&nblock); 1361 while (numdirtybuffers >= hidirtybuffers) { 1362 bd_wakeup(1); 1363 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1364 msleep(&needsbuffer, &nblock, 1365 (PRIBIO + 4), "flswai", 0); 1366 } 1367 mtx_unlock(&nblock); 1368 } 1369} 1370 1371/* 1372 * Return true if we have too many dirty buffers. 1373 */ 1374int 1375buf_dirty_count_severe(void) 1376{ 1377 1378 return(numdirtybuffers >= hidirtybuffers); 1379} 1380 1381static __noinline int 1382buf_vm_page_count_severe(void) 1383{ 1384 1385 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1); 1386 1387 return vm_page_count_severe(); 1388} 1389 1390/* 1391 * brelse: 1392 * 1393 * Release a busy buffer and, if requested, free its resources. The 1394 * buffer will be stashed in the appropriate bufqueue[] allowing it 1395 * to be accessed later as a cache entity or reused for other purposes. 1396 */ 1397void 1398brelse(struct buf *bp) 1399{ 1400 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1401 bp, bp->b_vp, bp->b_flags); 1402 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1403 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1404 1405 if (BUF_LOCKRECURSED(bp)) { 1406 /* 1407 * Do not process, in particular, do not handle the 1408 * B_INVAL/B_RELBUF and do not release to free list. 1409 */ 1410 BUF_UNLOCK(bp); 1411 return; 1412 } 1413 1414 if (bp->b_flags & B_MANAGED) { 1415 bqrelse(bp); 1416 return; 1417 } 1418 1419 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1420 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1421 /* 1422 * Failed write, redirty. Must clear BIO_ERROR to prevent 1423 * pages from being scrapped. If the error is anything 1424 * other than an I/O error (EIO), assume that retrying 1425 * is futile. 1426 */ 1427 bp->b_ioflags &= ~BIO_ERROR; 1428 bdirty(bp); 1429 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1430 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1431 /* 1432 * Either a failed I/O or we were asked to free or not 1433 * cache the buffer. 1434 */ 1435 bp->b_flags |= B_INVAL; 1436 if (!LIST_EMPTY(&bp->b_dep)) 1437 buf_deallocate(bp); 1438 if (bp->b_flags & B_DELWRI) { 1439 atomic_subtract_int(&numdirtybuffers, 1); 1440 numdirtywakeup(lodirtybuffers); 1441 } 1442 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1443 if ((bp->b_flags & B_VMIO) == 0) { 1444 if (bp->b_bufsize) 1445 allocbuf(bp, 0); 1446 if (bp->b_vp) 1447 brelvp(bp); 1448 } 1449 } 1450 1451 /* 1452 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1453 * is called with B_DELWRI set, the underlying pages may wind up 1454 * getting freed causing a previous write (bdwrite()) to get 'lost' 1455 * because pages associated with a B_DELWRI bp are marked clean. 1456 * 1457 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1458 * if B_DELWRI is set. 1459 * 1460 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1461 * on pages to return pages to the VM page queues. 1462 */ 1463 if (bp->b_flags & B_DELWRI) 1464 bp->b_flags &= ~B_RELBUF; 1465 else if (buf_vm_page_count_severe()) { 1466 /* 1467 * The locking of the BO_LOCK is not necessary since 1468 * BKGRDINPROG cannot be set while we hold the buf 1469 * lock, it can only be cleared if it is already 1470 * pending. 1471 */ 1472 if (bp->b_vp) { 1473 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1474 bp->b_flags |= B_RELBUF; 1475 } else 1476 bp->b_flags |= B_RELBUF; 1477 } 1478 1479 /* 1480 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1481 * constituted, not even NFS buffers now. Two flags effect this. If 1482 * B_INVAL, the struct buf is invalidated but the VM object is kept 1483 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1484 * 1485 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1486 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1487 * buffer is also B_INVAL because it hits the re-dirtying code above. 1488 * 1489 * Normally we can do this whether a buffer is B_DELWRI or not. If 1490 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1491 * the commit state and we cannot afford to lose the buffer. If the 1492 * buffer has a background write in progress, we need to keep it 1493 * around to prevent it from being reconstituted and starting a second 1494 * background write. 1495 */ 1496 if ((bp->b_flags & B_VMIO) 1497 && !(bp->b_vp->v_mount != NULL && 1498 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1499 !vn_isdisk(bp->b_vp, NULL) && 1500 (bp->b_flags & B_DELWRI)) 1501 ) { 1502 1503 int i, j, resid; 1504 vm_page_t m; 1505 off_t foff; 1506 vm_pindex_t poff; 1507 vm_object_t obj; 1508 1509 obj = bp->b_bufobj->bo_object; 1510 1511 /* 1512 * Get the base offset and length of the buffer. Note that 1513 * in the VMIO case if the buffer block size is not 1514 * page-aligned then b_data pointer may not be page-aligned. 1515 * But our b_pages[] array *IS* page aligned. 1516 * 1517 * block sizes less then DEV_BSIZE (usually 512) are not 1518 * supported due to the page granularity bits (m->valid, 1519 * m->dirty, etc...). 1520 * 1521 * See man buf(9) for more information 1522 */ 1523 resid = bp->b_bufsize; 1524 foff = bp->b_offset; 1525 VM_OBJECT_WLOCK(obj); 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 for (j = i; j < bp->b_npages; j++) { 1540 vm_page_t mtmp; 1541 mtmp = bp->b_pages[j]; 1542 if (mtmp == bogus_page) { 1543 mtmp = vm_page_lookup(obj, poff + j); 1544 if (!mtmp) { 1545 panic("brelse: page missing\n"); 1546 } 1547 bp->b_pages[j] = mtmp; 1548 } 1549 } 1550 1551 if ((bp->b_flags & (B_INVAL | B_UNMAPPED)) == 0) { 1552 BUF_CHECK_MAPPED(bp); 1553 pmap_qenter( 1554 trunc_page((vm_offset_t)bp->b_data), 1555 bp->b_pages, bp->b_npages); 1556 } 1557 m = bp->b_pages[i]; 1558 } 1559 if ((bp->b_flags & B_NOCACHE) || 1560 (bp->b_ioflags & BIO_ERROR && 1561 bp->b_iocmd == BIO_READ)) { 1562 int poffset = foff & PAGE_MASK; 1563 int presid = resid > (PAGE_SIZE - poffset) ? 1564 (PAGE_SIZE - poffset) : resid; 1565 1566 KASSERT(presid >= 0, ("brelse: extra page")); 1567 vm_page_set_invalid(m, poffset, presid); 1568 if (had_bogus) 1569 printf("avoided corruption bug in bogus_page/brelse code\n"); 1570 } 1571 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1572 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1573 } 1574 VM_OBJECT_WUNLOCK(obj); 1575 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1576 vfs_vmio_release(bp); 1577 1578 } else if (bp->b_flags & B_VMIO) { 1579 1580 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1581 vfs_vmio_release(bp); 1582 } 1583 1584 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1585 if (bp->b_bufsize != 0) 1586 allocbuf(bp, 0); 1587 if (bp->b_vp != NULL) 1588 brelvp(bp); 1589 } 1590 1591 /* enqueue */ 1592 mtx_lock(&bqlock); 1593 /* Handle delayed bremfree() processing. */ 1594 if (bp->b_flags & B_REMFREE) { 1595 struct bufobj *bo; 1596 1597 bo = bp->b_bufobj; 1598 if (bo != NULL) 1599 BO_LOCK(bo); 1600 bremfreel(bp); 1601 if (bo != NULL) 1602 BO_UNLOCK(bo); 1603 } 1604 if (bp->b_qindex != QUEUE_NONE) 1605 panic("brelse: free buffer onto another queue???"); 1606 1607 /* 1608 * If the buffer has junk contents signal it and eventually 1609 * clean up B_DELWRI and diassociate the vnode so that gbincore() 1610 * doesn't find it. 1611 */ 1612 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1613 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1614 bp->b_flags |= B_INVAL; 1615 if (bp->b_flags & B_INVAL) { 1616 if (bp->b_flags & B_DELWRI) 1617 bundirty(bp); 1618 if (bp->b_vp) 1619 brelvp(bp); 1620 } 1621 1622 /* buffers with no memory */ 1623 if (bp->b_bufsize == 0) { 1624 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1625 if (bp->b_vflags & BV_BKGRDINPROG) 1626 panic("losing buffer 1"); 1627 if (bp->b_kvasize) { 1628 bp->b_qindex = QUEUE_EMPTYKVA; 1629 } else { 1630 bp->b_qindex = QUEUE_EMPTY; 1631 } 1632 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1633 /* buffers with junk contents */ 1634 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1635 (bp->b_ioflags & BIO_ERROR)) { 1636 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1637 if (bp->b_vflags & BV_BKGRDINPROG) 1638 panic("losing buffer 2"); 1639 bp->b_qindex = QUEUE_CLEAN; 1640 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1641 /* remaining buffers */ 1642 } else { 1643 if (bp->b_flags & B_DELWRI) 1644 bp->b_qindex = QUEUE_DIRTY; 1645 else 1646 bp->b_qindex = QUEUE_CLEAN; 1647 if (bp->b_flags & B_AGE) { 1648 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, 1649 b_freelist); 1650 } else { 1651 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, 1652 b_freelist); 1653 } 1654 } 1655#ifdef INVARIANTS 1656 bq_len[bp->b_qindex]++; 1657#endif 1658 mtx_unlock(&bqlock); 1659 1660 /* 1661 * Fixup numfreebuffers count. The bp is on an appropriate queue 1662 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1663 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1664 * if B_INVAL is set ). 1665 */ 1666 1667 if (!(bp->b_flags & B_DELWRI)) { 1668 struct bufobj *bo; 1669 1670 bo = bp->b_bufobj; 1671 if (bo != NULL) 1672 BO_LOCK(bo); 1673 bufcountwakeup(bp); 1674 if (bo != NULL) 1675 BO_UNLOCK(bo); 1676 } 1677 1678 /* 1679 * Something we can maybe free or reuse 1680 */ 1681 if (bp->b_bufsize || bp->b_kvasize) 1682 bufspacewakeup(); 1683 1684 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1685 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1686 panic("brelse: not dirty"); 1687 /* unlock */ 1688 BUF_UNLOCK(bp); 1689} 1690 1691/* 1692 * Release a buffer back to the appropriate queue but do not try to free 1693 * it. The buffer is expected to be used again soon. 1694 * 1695 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1696 * biodone() to requeue an async I/O on completion. It is also used when 1697 * known good buffers need to be requeued but we think we may need the data 1698 * again soon. 1699 * 1700 * XXX we should be able to leave the B_RELBUF hint set on completion. 1701 */ 1702void 1703bqrelse(struct buf *bp) 1704{ 1705 struct bufobj *bo; 1706 1707 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1708 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1709 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1710 1711 if (BUF_LOCKRECURSED(bp)) { 1712 /* do not release to free list */ 1713 BUF_UNLOCK(bp); 1714 return; 1715 } 1716 1717 bo = bp->b_bufobj; 1718 if (bp->b_flags & B_MANAGED) { 1719 if (bp->b_flags & B_REMFREE) { 1720 mtx_lock(&bqlock); 1721 if (bo != NULL) 1722 BO_LOCK(bo); 1723 bremfreel(bp); 1724 if (bo != NULL) 1725 BO_UNLOCK(bo); 1726 mtx_unlock(&bqlock); 1727 } 1728 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1729 BUF_UNLOCK(bp); 1730 return; 1731 } 1732 1733 mtx_lock(&bqlock); 1734 /* Handle delayed bremfree() processing. */ 1735 if (bp->b_flags & B_REMFREE) { 1736 if (bo != NULL) 1737 BO_LOCK(bo); 1738 bremfreel(bp); 1739 if (bo != NULL) 1740 BO_UNLOCK(bo); 1741 } 1742 if (bp->b_qindex != QUEUE_NONE) 1743 panic("bqrelse: free buffer onto another queue???"); 1744 /* buffers with stale but valid contents */ 1745 if (bp->b_flags & B_DELWRI) { 1746 bp->b_qindex = QUEUE_DIRTY; 1747 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1748#ifdef INVARIANTS 1749 bq_len[bp->b_qindex]++; 1750#endif 1751 } else { 1752 /* 1753 * The locking of the BO_LOCK for checking of the 1754 * BV_BKGRDINPROG is not necessary since the 1755 * BV_BKGRDINPROG cannot be set while we hold the buf 1756 * lock, it can only be cleared if it is already 1757 * pending. 1758 */ 1759 if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) { 1760 bp->b_qindex = QUEUE_CLEAN; 1761 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1762 b_freelist); 1763#ifdef INVARIANTS 1764 bq_len[QUEUE_CLEAN]++; 1765#endif 1766 } else { 1767 /* 1768 * We are too low on memory, we have to try to free 1769 * the buffer (most importantly: the wired pages 1770 * making up its backing store) *now*. 1771 */ 1772 mtx_unlock(&bqlock); 1773 brelse(bp); 1774 return; 1775 } 1776 } 1777 mtx_unlock(&bqlock); 1778 1779 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) { 1780 if (bo != NULL) 1781 BO_LOCK(bo); 1782 bufcountwakeup(bp); 1783 if (bo != NULL) 1784 BO_UNLOCK(bo); 1785 } 1786 1787 /* 1788 * Something we can maybe free or reuse. 1789 */ 1790 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1791 bufspacewakeup(); 1792 1793 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1794 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1795 panic("bqrelse: not dirty"); 1796 /* unlock */ 1797 BUF_UNLOCK(bp); 1798} 1799 1800/* Give pages used by the bp back to the VM system (where possible) */ 1801static void 1802vfs_vmio_release(struct buf *bp) 1803{ 1804 int i; 1805 vm_page_t m; 1806 1807 if ((bp->b_flags & B_UNMAPPED) == 0) { 1808 BUF_CHECK_MAPPED(bp); 1809 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 1810 } else 1811 BUF_CHECK_UNMAPPED(bp); 1812 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 1813 for (i = 0; i < bp->b_npages; i++) { 1814 m = bp->b_pages[i]; 1815 bp->b_pages[i] = NULL; 1816 /* 1817 * In order to keep page LRU ordering consistent, put 1818 * everything on the inactive queue. 1819 */ 1820 vm_page_lock(m); 1821 vm_page_unwire(m, 0); 1822 /* 1823 * We don't mess with busy pages, it is 1824 * the responsibility of the process that 1825 * busied the pages to deal with them. 1826 */ 1827 if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 && 1828 m->wire_count == 0) { 1829 /* 1830 * Might as well free the page if we can and it has 1831 * no valid data. We also free the page if the 1832 * buffer was used for direct I/O 1833 */ 1834 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) { 1835 vm_page_free(m); 1836 } else if (bp->b_flags & B_DIRECT) { 1837 vm_page_try_to_free(m); 1838 } else if (buf_vm_page_count_severe()) { 1839 vm_page_try_to_cache(m); 1840 } 1841 } 1842 vm_page_unlock(m); 1843 } 1844 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 1845 1846 if (bp->b_bufsize) { 1847 bufspacewakeup(); 1848 bp->b_bufsize = 0; 1849 } 1850 bp->b_npages = 0; 1851 bp->b_flags &= ~B_VMIO; 1852 if (bp->b_vp) 1853 brelvp(bp); 1854} 1855 1856/* 1857 * Check to see if a block at a particular lbn is available for a clustered 1858 * write. 1859 */ 1860static int 1861vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1862{ 1863 struct buf *bpa; 1864 int match; 1865 1866 match = 0; 1867 1868 /* If the buf isn't in core skip it */ 1869 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1870 return (0); 1871 1872 /* If the buf is busy we don't want to wait for it */ 1873 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1874 return (0); 1875 1876 /* Only cluster with valid clusterable delayed write buffers */ 1877 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1878 (B_DELWRI | B_CLUSTEROK)) 1879 goto done; 1880 1881 if (bpa->b_bufsize != size) 1882 goto done; 1883 1884 /* 1885 * Check to see if it is in the expected place on disk and that the 1886 * block has been mapped. 1887 */ 1888 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1889 match = 1; 1890done: 1891 BUF_UNLOCK(bpa); 1892 return (match); 1893} 1894 1895/* 1896 * vfs_bio_awrite: 1897 * 1898 * Implement clustered async writes for clearing out B_DELWRI buffers. 1899 * This is much better then the old way of writing only one buffer at 1900 * a time. Note that we may not be presented with the buffers in the 1901 * correct order, so we search for the cluster in both directions. 1902 */ 1903int 1904vfs_bio_awrite(struct buf *bp) 1905{ 1906 struct bufobj *bo; 1907 int i; 1908 int j; 1909 daddr_t lblkno = bp->b_lblkno; 1910 struct vnode *vp = bp->b_vp; 1911 int ncl; 1912 int nwritten; 1913 int size; 1914 int maxcl; 1915 int gbflags; 1916 1917 bo = &vp->v_bufobj; 1918 gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0; 1919 /* 1920 * right now we support clustered writing only to regular files. If 1921 * we find a clusterable block we could be in the middle of a cluster 1922 * rather then at the beginning. 1923 */ 1924 if ((vp->v_type == VREG) && 1925 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1926 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1927 1928 size = vp->v_mount->mnt_stat.f_iosize; 1929 maxcl = MAXPHYS / size; 1930 1931 BO_LOCK(bo); 1932 for (i = 1; i < maxcl; i++) 1933 if (vfs_bio_clcheck(vp, size, lblkno + i, 1934 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1935 break; 1936 1937 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1938 if (vfs_bio_clcheck(vp, size, lblkno - j, 1939 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1940 break; 1941 BO_UNLOCK(bo); 1942 --j; 1943 ncl = i + j; 1944 /* 1945 * this is a possible cluster write 1946 */ 1947 if (ncl != 1) { 1948 BUF_UNLOCK(bp); 1949 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 1950 gbflags); 1951 return (nwritten); 1952 } 1953 } 1954 bremfree(bp); 1955 bp->b_flags |= B_ASYNC; 1956 /* 1957 * default (old) behavior, writing out only one block 1958 * 1959 * XXX returns b_bufsize instead of b_bcount for nwritten? 1960 */ 1961 nwritten = bp->b_bufsize; 1962 (void) bwrite(bp); 1963 1964 return (nwritten); 1965} 1966 1967static void 1968setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags) 1969{ 1970 1971 KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 1972 bp->b_kvasize == 0, ("call bfreekva(%p)", bp)); 1973 if ((gbflags & GB_UNMAPPED) == 0) { 1974 bp->b_kvabase = (caddr_t)addr; 1975 } else if ((gbflags & GB_KVAALLOC) != 0) { 1976 KASSERT((gbflags & GB_UNMAPPED) != 0, 1977 ("GB_KVAALLOC without GB_UNMAPPED")); 1978 bp->b_kvaalloc = (caddr_t)addr; 1979 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 1980 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 1981 } 1982 bp->b_kvasize = maxsize; 1983} 1984 1985/* 1986 * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if 1987 * needed. 1988 */ 1989static int 1990allocbufkva(struct buf *bp, int maxsize, int gbflags) 1991{ 1992 vm_offset_t addr; 1993 int rv; 1994 1995 bfreekva(bp); 1996 addr = 0; 1997 1998 vm_map_lock(buffer_map); 1999 if (vm_map_findspace(buffer_map, vm_map_min(buffer_map), maxsize, 2000 &addr)) { 2001 vm_map_unlock(buffer_map); 2002 /* 2003 * Buffer map is too fragmented. Request the caller 2004 * to defragment the map. 2005 */ 2006 atomic_add_int(&bufdefragcnt, 1); 2007 return (1); 2008 } 2009 rv = vm_map_insert(buffer_map, NULL, 0, addr, addr + maxsize, 2010 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); 2011 KASSERT(rv == KERN_SUCCESS, ("vm_map_insert(buffer_map) rv %d", rv)); 2012 vm_map_unlock(buffer_map); 2013 setbufkva(bp, addr, maxsize, gbflags); 2014 atomic_add_long(&bufspace, bp->b_kvasize); 2015 return (0); 2016} 2017 2018/* 2019 * Ask the bufdaemon for help, or act as bufdaemon itself, when a 2020 * locked vnode is supplied. 2021 */ 2022static void 2023getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo, 2024 int defrag) 2025{ 2026 struct thread *td; 2027 char *waitmsg; 2028 int fl, flags, norunbuf; 2029 2030 mtx_assert(&bqlock, MA_OWNED); 2031 2032 if (defrag) { 2033 flags = VFS_BIO_NEED_BUFSPACE; 2034 waitmsg = "nbufkv"; 2035 } else if (bufspace >= hibufspace) { 2036 waitmsg = "nbufbs"; 2037 flags = VFS_BIO_NEED_BUFSPACE; 2038 } else { 2039 waitmsg = "newbuf"; 2040 flags = VFS_BIO_NEED_ANY; 2041 } 2042 mtx_lock(&nblock); 2043 needsbuffer |= flags; 2044 mtx_unlock(&nblock); 2045 mtx_unlock(&bqlock); 2046 2047 bd_speedup(); /* heeeelp */ 2048 if ((gbflags & GB_NOWAIT_BD) != 0) 2049 return; 2050 2051 td = curthread; 2052 mtx_lock(&nblock); 2053 while (needsbuffer & flags) { 2054 if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) { 2055 mtx_unlock(&nblock); 2056 /* 2057 * getblk() is called with a vnode locked, and 2058 * some majority of the dirty buffers may as 2059 * well belong to the vnode. Flushing the 2060 * buffers there would make a progress that 2061 * cannot be achieved by the buf_daemon, that 2062 * cannot lock the vnode. 2063 */ 2064 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2065 (td->td_pflags & TDP_NORUNNINGBUF); 2066 /* play bufdaemon */ 2067 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2068 fl = buf_do_flush(vp); 2069 td->td_pflags &= norunbuf; 2070 mtx_lock(&nblock); 2071 if (fl != 0) 2072 continue; 2073 if ((needsbuffer & flags) == 0) 2074 break; 2075 } 2076 if (msleep(&needsbuffer, &nblock, (PRIBIO + 4) | slpflag, 2077 waitmsg, slptimeo)) 2078 break; 2079 } 2080 mtx_unlock(&nblock); 2081} 2082 2083static void 2084getnewbuf_reuse_bp(struct buf *bp, int qindex) 2085{ 2086 2087 CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 2088 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 2089 bp->b_kvasize, bp->b_bufsize, qindex); 2090 mtx_assert(&bqlock, MA_NOTOWNED); 2091 2092 /* 2093 * Note: we no longer distinguish between VMIO and non-VMIO 2094 * buffers. 2095 */ 2096 KASSERT((bp->b_flags & B_DELWRI) == 0, 2097 ("delwri buffer %p found in queue %d", bp, qindex)); 2098 2099 if (qindex == QUEUE_CLEAN) { 2100 if (bp->b_flags & B_VMIO) { 2101 bp->b_flags &= ~B_ASYNC; 2102 vfs_vmio_release(bp); 2103 } 2104 if (bp->b_vp != NULL) 2105 brelvp(bp); 2106 } 2107 2108 /* 2109 * Get the rest of the buffer freed up. b_kva* is still valid 2110 * after this operation. 2111 */ 2112 2113 if (bp->b_rcred != NOCRED) { 2114 crfree(bp->b_rcred); 2115 bp->b_rcred = NOCRED; 2116 } 2117 if (bp->b_wcred != NOCRED) { 2118 crfree(bp->b_wcred); 2119 bp->b_wcred = NOCRED; 2120 } 2121 if (!LIST_EMPTY(&bp->b_dep)) 2122 buf_deallocate(bp); 2123 if (bp->b_vflags & BV_BKGRDINPROG) 2124 panic("losing buffer 3"); 2125 KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p. qindex: %d", 2126 bp, bp->b_vp, qindex)); 2127 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 2128 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 2129 2130 if (bp->b_bufsize) 2131 allocbuf(bp, 0); 2132 2133 bp->b_flags &= B_UNMAPPED | B_KVAALLOC; 2134 bp->b_ioflags = 0; 2135 bp->b_xflags = 0; 2136 KASSERT((bp->b_vflags & BV_INFREECNT) == 0, 2137 ("buf %p still counted as free?", bp)); 2138 bp->b_vflags = 0; 2139 bp->b_vp = NULL; 2140 bp->b_blkno = bp->b_lblkno = 0; 2141 bp->b_offset = NOOFFSET; 2142 bp->b_iodone = 0; 2143 bp->b_error = 0; 2144 bp->b_resid = 0; 2145 bp->b_bcount = 0; 2146 bp->b_npages = 0; 2147 bp->b_dirtyoff = bp->b_dirtyend = 0; 2148 bp->b_bufobj = NULL; 2149 bp->b_pin_count = 0; 2150 bp->b_fsprivate1 = NULL; 2151 bp->b_fsprivate2 = NULL; 2152 bp->b_fsprivate3 = NULL; 2153 2154 LIST_INIT(&bp->b_dep); 2155} 2156 2157static int flushingbufs; 2158 2159static struct buf * 2160getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata) 2161{ 2162 struct buf *bp, *nbp; 2163 int nqindex, qindex, pass; 2164 2165 KASSERT(!unmapped || !defrag, ("both unmapped and defrag")); 2166 2167 pass = 1; 2168restart: 2169 atomic_add_int(&getnewbufrestarts, 1); 2170 2171 /* 2172 * Setup for scan. If we do not have enough free buffers, 2173 * we setup a degenerate case that immediately fails. Note 2174 * that if we are specially marked process, we are allowed to 2175 * dip into our reserves. 2176 * 2177 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 2178 * for the allocation of the mapped buffer. For unmapped, the 2179 * easiest is to start with EMPTY outright. 2180 * 2181 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 2182 * However, there are a number of cases (defragging, reusing, ...) 2183 * where we cannot backup. 2184 */ 2185 nbp = NULL; 2186 mtx_lock(&bqlock); 2187 if (!defrag && unmapped) { 2188 nqindex = QUEUE_EMPTY; 2189 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2190 } 2191 if (nbp == NULL) { 2192 nqindex = QUEUE_EMPTYKVA; 2193 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2194 } 2195 2196 /* 2197 * If no EMPTYKVA buffers and we are either defragging or 2198 * reusing, locate a CLEAN buffer to free or reuse. If 2199 * bufspace useage is low skip this step so we can allocate a 2200 * new buffer. 2201 */ 2202 if (nbp == NULL && (defrag || bufspace >= lobufspace)) { 2203 nqindex = QUEUE_CLEAN; 2204 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2205 } 2206 2207 /* 2208 * If we could not find or were not allowed to reuse a CLEAN 2209 * buffer, check to see if it is ok to use an EMPTY buffer. 2210 * We can only use an EMPTY buffer if allocating its KVA would 2211 * not otherwise run us out of buffer space. No KVA is needed 2212 * for the unmapped allocation. 2213 */ 2214 if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace || 2215 metadata)) { 2216 nqindex = QUEUE_EMPTY; 2217 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2218 } 2219 2220 /* 2221 * All available buffers might be clean, retry ignoring the 2222 * lobufspace as the last resort. 2223 */ 2224 if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) { 2225 nqindex = QUEUE_CLEAN; 2226 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2227 } 2228 2229 /* 2230 * Run scan, possibly freeing data and/or kva mappings on the fly 2231 * depending. 2232 */ 2233 while ((bp = nbp) != NULL) { 2234 qindex = nqindex; 2235 2236 /* 2237 * Calculate next bp (we can only use it if we do not 2238 * block or do other fancy things). 2239 */ 2240 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 2241 switch (qindex) { 2242 case QUEUE_EMPTY: 2243 nqindex = QUEUE_EMPTYKVA; 2244 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2245 if (nbp != NULL) 2246 break; 2247 /* FALLTHROUGH */ 2248 case QUEUE_EMPTYKVA: 2249 nqindex = QUEUE_CLEAN; 2250 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2251 if (nbp != NULL) 2252 break; 2253 /* FALLTHROUGH */ 2254 case QUEUE_CLEAN: 2255 if (metadata && pass == 1) { 2256 pass = 2; 2257 nqindex = QUEUE_EMPTY; 2258 nbp = TAILQ_FIRST( 2259 &bufqueues[QUEUE_EMPTY]); 2260 } 2261 /* 2262 * nbp is NULL. 2263 */ 2264 break; 2265 } 2266 } 2267 /* 2268 * If we are defragging then we need a buffer with 2269 * b_kvasize != 0. XXX this situation should no longer 2270 * occur, if defrag is non-zero the buffer's b_kvasize 2271 * should also be non-zero at this point. XXX 2272 */ 2273 if (defrag && bp->b_kvasize == 0) { 2274 printf("Warning: defrag empty buffer %p\n", bp); 2275 continue; 2276 } 2277 2278 /* 2279 * Start freeing the bp. This is somewhat involved. nbp 2280 * remains valid only for QUEUE_EMPTY[KVA] bp's. 2281 */ 2282 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2283 continue; 2284 if (bp->b_vp) { 2285 BO_LOCK(bp->b_bufobj); 2286 if (bp->b_vflags & BV_BKGRDINPROG) { 2287 BO_UNLOCK(bp->b_bufobj); 2288 BUF_UNLOCK(bp); 2289 continue; 2290 } 2291 BO_UNLOCK(bp->b_bufobj); 2292 } 2293 2294 KASSERT(bp->b_qindex == qindex, 2295 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2296 2297 if (bp->b_bufobj != NULL) 2298 BO_LOCK(bp->b_bufobj); 2299 bremfreel(bp); 2300 if (bp->b_bufobj != NULL) 2301 BO_UNLOCK(bp->b_bufobj); 2302 mtx_unlock(&bqlock); 2303 /* 2304 * NOTE: nbp is now entirely invalid. We can only restart 2305 * the scan from this point on. 2306 */ 2307 2308 getnewbuf_reuse_bp(bp, qindex); 2309 mtx_assert(&bqlock, MA_NOTOWNED); 2310 2311 /* 2312 * If we are defragging then free the buffer. 2313 */ 2314 if (defrag) { 2315 bp->b_flags |= B_INVAL; 2316 bfreekva(bp); 2317 brelse(bp); 2318 defrag = 0; 2319 goto restart; 2320 } 2321 2322 /* 2323 * Notify any waiters for the buffer lock about 2324 * identity change by freeing the buffer. 2325 */ 2326 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 2327 bp->b_flags |= B_INVAL; 2328 bfreekva(bp); 2329 brelse(bp); 2330 goto restart; 2331 } 2332 2333 if (metadata) 2334 break; 2335 2336 /* 2337 * If we are overcomitted then recover the buffer and its 2338 * KVM space. This occurs in rare situations when multiple 2339 * processes are blocked in getnewbuf() or allocbuf(). 2340 */ 2341 if (bufspace >= hibufspace) 2342 flushingbufs = 1; 2343 if (flushingbufs && bp->b_kvasize != 0) { 2344 bp->b_flags |= B_INVAL; 2345 bfreekva(bp); 2346 brelse(bp); 2347 goto restart; 2348 } 2349 if (bufspace < lobufspace) 2350 flushingbufs = 0; 2351 break; 2352 } 2353 return (bp); 2354} 2355 2356/* 2357 * getnewbuf: 2358 * 2359 * Find and initialize a new buffer header, freeing up existing buffers 2360 * in the bufqueues as necessary. The new buffer is returned locked. 2361 * 2362 * Important: B_INVAL is not set. If the caller wishes to throw the 2363 * buffer away, the caller must set B_INVAL prior to calling brelse(). 2364 * 2365 * We block if: 2366 * We have insufficient buffer headers 2367 * We have insufficient buffer space 2368 * buffer_map is too fragmented ( space reservation fails ) 2369 * If we have to flush dirty buffers ( but we try to avoid this ) 2370 * 2371 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 2372 * Instead we ask the buf daemon to do it for us. We attempt to 2373 * avoid piecemeal wakeups of the pageout daemon. 2374 */ 2375static struct buf * 2376getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 2377 int gbflags) 2378{ 2379 struct buf *bp; 2380 int defrag, metadata; 2381 2382 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 2383 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 2384 if (!unmapped_buf_allowed) 2385 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 2386 2387 defrag = 0; 2388 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 2389 vp->v_type == VCHR) 2390 metadata = 1; 2391 else 2392 metadata = 0; 2393 /* 2394 * We can't afford to block since we might be holding a vnode lock, 2395 * which may prevent system daemons from running. We deal with 2396 * low-memory situations by proactively returning memory and running 2397 * async I/O rather then sync I/O. 2398 */ 2399 atomic_add_int(&getnewbufcalls, 1); 2400 atomic_subtract_int(&getnewbufrestarts, 1); 2401restart: 2402 bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED | 2403 GB_KVAALLOC)) == GB_UNMAPPED, metadata); 2404 if (bp != NULL) 2405 defrag = 0; 2406 2407 /* 2408 * If we exhausted our list, sleep as appropriate. We may have to 2409 * wakeup various daemons and write out some dirty buffers. 2410 * 2411 * Generally we are sleeping due to insufficient buffer space. 2412 */ 2413 if (bp == NULL) { 2414 mtx_assert(&bqlock, MA_OWNED); 2415 getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag); 2416 mtx_assert(&bqlock, MA_NOTOWNED); 2417 } else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) { 2418 mtx_assert(&bqlock, MA_NOTOWNED); 2419 2420 bfreekva(bp); 2421 bp->b_flags |= B_UNMAPPED; 2422 bp->b_kvabase = bp->b_data = unmapped_buf; 2423 bp->b_kvasize = maxsize; 2424 atomic_add_long(&bufspace, bp->b_kvasize); 2425 atomic_add_long(&unmapped_bufspace, bp->b_kvasize); 2426 atomic_add_int(&bufreusecnt, 1); 2427 } else { 2428 mtx_assert(&bqlock, MA_NOTOWNED); 2429 2430 /* 2431 * We finally have a valid bp. We aren't quite out of the 2432 * woods, we still have to reserve kva space. In order 2433 * to keep fragmentation sane we only allocate kva in 2434 * BKVASIZE chunks. 2435 */ 2436 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2437 2438 if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED | 2439 B_KVAALLOC)) == B_UNMAPPED) { 2440 if (allocbufkva(bp, maxsize, gbflags)) { 2441 defrag = 1; 2442 bp->b_flags |= B_INVAL; 2443 brelse(bp); 2444 goto restart; 2445 } 2446 atomic_add_int(&bufreusecnt, 1); 2447 } else if ((bp->b_flags & B_KVAALLOC) != 0 && 2448 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) { 2449 /* 2450 * If the reused buffer has KVA allocated, 2451 * reassign b_kvaalloc to b_kvabase. 2452 */ 2453 bp->b_kvabase = bp->b_kvaalloc; 2454 bp->b_flags &= ~B_KVAALLOC; 2455 atomic_subtract_long(&unmapped_bufspace, 2456 bp->b_kvasize); 2457 atomic_add_int(&bufreusecnt, 1); 2458 } else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 && 2459 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED | 2460 GB_KVAALLOC)) { 2461 /* 2462 * The case of reused buffer already have KVA 2463 * mapped, but the request is for unmapped 2464 * buffer with KVA allocated. 2465 */ 2466 bp->b_kvaalloc = bp->b_kvabase; 2467 bp->b_data = bp->b_kvabase = unmapped_buf; 2468 bp->b_flags |= B_UNMAPPED | B_KVAALLOC; 2469 atomic_add_long(&unmapped_bufspace, 2470 bp->b_kvasize); 2471 atomic_add_int(&bufreusecnt, 1); 2472 } 2473 if ((gbflags & GB_UNMAPPED) == 0) { 2474 bp->b_saveaddr = bp->b_kvabase; 2475 bp->b_data = bp->b_saveaddr; 2476 bp->b_flags &= ~B_UNMAPPED; 2477 BUF_CHECK_MAPPED(bp); 2478 } 2479 } 2480 return (bp); 2481} 2482 2483/* 2484 * buf_daemon: 2485 * 2486 * buffer flushing daemon. Buffers are normally flushed by the 2487 * update daemon but if it cannot keep up this process starts to 2488 * take the load in an attempt to prevent getnewbuf() from blocking. 2489 */ 2490 2491static struct kproc_desc buf_kp = { 2492 "bufdaemon", 2493 buf_daemon, 2494 &bufdaemonproc 2495}; 2496SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2497 2498static int 2499buf_do_flush(struct vnode *vp) 2500{ 2501 int flushed; 2502 2503 flushed = flushbufqueues(vp, QUEUE_DIRTY, 0); 2504 if (flushed == 0) { 2505 /* 2506 * Could not find any buffers without rollback 2507 * dependencies, so just write the first one 2508 * in the hopes of eventually making progress. 2509 */ 2510 flushbufqueues(vp, QUEUE_DIRTY, 1); 2511 } 2512 return (flushed); 2513} 2514 2515static void 2516buf_daemon() 2517{ 2518 int lodirtysave; 2519 2520 /* 2521 * This process needs to be suspended prior to shutdown sync. 2522 */ 2523 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2524 SHUTDOWN_PRI_LAST); 2525 2526 /* 2527 * This process is allowed to take the buffer cache to the limit 2528 */ 2529 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2530 mtx_lock(&bdlock); 2531 for (;;) { 2532 bd_request = 0; 2533 mtx_unlock(&bdlock); 2534 2535 kproc_suspend_check(bufdaemonproc); 2536 lodirtysave = lodirtybuffers; 2537 if (bd_speedupreq) { 2538 lodirtybuffers = numdirtybuffers / 2; 2539 bd_speedupreq = 0; 2540 } 2541 /* 2542 * Do the flush. Limit the amount of in-transit I/O we 2543 * allow to build up, otherwise we would completely saturate 2544 * the I/O system. Wakeup any waiting processes before we 2545 * normally would so they can run in parallel with our drain. 2546 */ 2547 while (numdirtybuffers > lodirtybuffers) { 2548 if (buf_do_flush(NULL) == 0) 2549 break; 2550 kern_yield(PRI_USER); 2551 } 2552 lodirtybuffers = lodirtysave; 2553 2554 /* 2555 * Only clear bd_request if we have reached our low water 2556 * mark. The buf_daemon normally waits 1 second and 2557 * then incrementally flushes any dirty buffers that have 2558 * built up, within reason. 2559 * 2560 * If we were unable to hit our low water mark and couldn't 2561 * find any flushable buffers, we sleep half a second. 2562 * Otherwise we loop immediately. 2563 */ 2564 mtx_lock(&bdlock); 2565 if (numdirtybuffers <= lodirtybuffers) { 2566 /* 2567 * We reached our low water mark, reset the 2568 * request and sleep until we are needed again. 2569 * The sleep is just so the suspend code works. 2570 */ 2571 bd_request = 0; 2572 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2573 } else { 2574 /* 2575 * We couldn't find any flushable dirty buffers but 2576 * still have too many dirty buffers, we 2577 * have to sleep and try again. (rare) 2578 */ 2579 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2580 } 2581 } 2582} 2583 2584/* 2585 * flushbufqueues: 2586 * 2587 * Try to flush a buffer in the dirty queue. We must be careful to 2588 * free up B_INVAL buffers instead of write them, which NFS is 2589 * particularly sensitive to. 2590 */ 2591static int flushwithdeps = 0; 2592SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2593 0, "Number of buffers flushed with dependecies that require rollbacks"); 2594 2595static int 2596flushbufqueues(struct vnode *lvp, int queue, int flushdeps) 2597{ 2598 struct buf *sentinel; 2599 struct vnode *vp; 2600 struct mount *mp; 2601 struct buf *bp; 2602 int hasdeps; 2603 int flushed; 2604 int target; 2605 2606 if (lvp == NULL) { 2607 target = numdirtybuffers - lodirtybuffers; 2608 if (flushdeps && target > 2) 2609 target /= 2; 2610 } else 2611 target = flushbufqtarget; 2612 flushed = 0; 2613 bp = NULL; 2614 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2615 sentinel->b_qindex = QUEUE_SENTINEL; 2616 mtx_lock(&bqlock); 2617 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2618 while (flushed != target) { 2619 bp = TAILQ_NEXT(sentinel, b_freelist); 2620 if (bp != NULL) { 2621 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2622 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2623 b_freelist); 2624 } else 2625 break; 2626 /* 2627 * Skip sentinels inserted by other invocations of the 2628 * flushbufqueues(), taking care to not reorder them. 2629 */ 2630 if (bp->b_qindex == QUEUE_SENTINEL) 2631 continue; 2632 /* 2633 * Only flush the buffers that belong to the 2634 * vnode locked by the curthread. 2635 */ 2636 if (lvp != NULL && bp->b_vp != lvp) 2637 continue; 2638 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2639 continue; 2640 if (bp->b_pin_count > 0) { 2641 BUF_UNLOCK(bp); 2642 continue; 2643 } 2644 BO_LOCK(bp->b_bufobj); 2645 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2646 (bp->b_flags & B_DELWRI) == 0) { 2647 BO_UNLOCK(bp->b_bufobj); 2648 BUF_UNLOCK(bp); 2649 continue; 2650 } 2651 BO_UNLOCK(bp->b_bufobj); 2652 if (bp->b_flags & B_INVAL) { 2653 bremfreel(bp); 2654 mtx_unlock(&bqlock); 2655 brelse(bp); 2656 flushed++; 2657 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2658 mtx_lock(&bqlock); 2659 continue; 2660 } 2661 2662 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2663 if (flushdeps == 0) { 2664 BUF_UNLOCK(bp); 2665 continue; 2666 } 2667 hasdeps = 1; 2668 } else 2669 hasdeps = 0; 2670 /* 2671 * We must hold the lock on a vnode before writing 2672 * one of its buffers. Otherwise we may confuse, or 2673 * in the case of a snapshot vnode, deadlock the 2674 * system. 2675 * 2676 * The lock order here is the reverse of the normal 2677 * of vnode followed by buf lock. This is ok because 2678 * the NOWAIT will prevent deadlock. 2679 */ 2680 vp = bp->b_vp; 2681 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2682 BUF_UNLOCK(bp); 2683 continue; 2684 } 2685 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) { 2686 mtx_unlock(&bqlock); 2687 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2688 bp, bp->b_vp, bp->b_flags); 2689 if (curproc == bufdaemonproc) 2690 vfs_bio_awrite(bp); 2691 else { 2692 bremfree(bp); 2693 bwrite(bp); 2694 notbufdflashes++; 2695 } 2696 vn_finished_write(mp); 2697 VOP_UNLOCK(vp, 0); 2698 flushwithdeps += hasdeps; 2699 flushed++; 2700 2701 /* 2702 * Sleeping on runningbufspace while holding 2703 * vnode lock leads to deadlock. 2704 */ 2705 if (curproc == bufdaemonproc) 2706 waitrunningbufspace(); 2707 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2708 mtx_lock(&bqlock); 2709 continue; 2710 } 2711 vn_finished_write(mp); 2712 BUF_UNLOCK(bp); 2713 } 2714 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2715 mtx_unlock(&bqlock); 2716 free(sentinel, M_TEMP); 2717 return (flushed); 2718} 2719 2720/* 2721 * Check to see if a block is currently memory resident. 2722 */ 2723struct buf * 2724incore(struct bufobj *bo, daddr_t blkno) 2725{ 2726 struct buf *bp; 2727 2728 BO_LOCK(bo); 2729 bp = gbincore(bo, blkno); 2730 BO_UNLOCK(bo); 2731 return (bp); 2732} 2733 2734/* 2735 * Returns true if no I/O is needed to access the 2736 * associated VM object. This is like incore except 2737 * it also hunts around in the VM system for the data. 2738 */ 2739 2740static int 2741inmem(struct vnode * vp, daddr_t blkno) 2742{ 2743 vm_object_t obj; 2744 vm_offset_t toff, tinc, size; 2745 vm_page_t m; 2746 vm_ooffset_t off; 2747 2748 ASSERT_VOP_LOCKED(vp, "inmem"); 2749 2750 if (incore(&vp->v_bufobj, blkno)) 2751 return 1; 2752 if (vp->v_mount == NULL) 2753 return 0; 2754 obj = vp->v_object; 2755 if (obj == NULL) 2756 return (0); 2757 2758 size = PAGE_SIZE; 2759 if (size > vp->v_mount->mnt_stat.f_iosize) 2760 size = vp->v_mount->mnt_stat.f_iosize; 2761 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2762 2763 VM_OBJECT_WLOCK(obj); 2764 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2765 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2766 if (!m) 2767 goto notinmem; 2768 tinc = size; 2769 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2770 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2771 if (vm_page_is_valid(m, 2772 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2773 goto notinmem; 2774 } 2775 VM_OBJECT_WUNLOCK(obj); 2776 return 1; 2777 2778notinmem: 2779 VM_OBJECT_WUNLOCK(obj); 2780 return (0); 2781} 2782 2783/* 2784 * Set the dirty range for a buffer based on the status of the dirty 2785 * bits in the pages comprising the buffer. The range is limited 2786 * to the size of the buffer. 2787 * 2788 * Tell the VM system that the pages associated with this buffer 2789 * are clean. This is used for delayed writes where the data is 2790 * going to go to disk eventually without additional VM intevention. 2791 * 2792 * Note that while we only really need to clean through to b_bcount, we 2793 * just go ahead and clean through to b_bufsize. 2794 */ 2795static void 2796vfs_clean_pages_dirty_buf(struct buf *bp) 2797{ 2798 vm_ooffset_t foff, noff, eoff; 2799 vm_page_t m; 2800 int i; 2801 2802 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 2803 return; 2804 2805 foff = bp->b_offset; 2806 KASSERT(bp->b_offset != NOOFFSET, 2807 ("vfs_clean_pages_dirty_buf: no buffer offset")); 2808 2809 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 2810 vfs_drain_busy_pages(bp); 2811 vfs_setdirty_locked_object(bp); 2812 for (i = 0; i < bp->b_npages; i++) { 2813 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2814 eoff = noff; 2815 if (eoff > bp->b_offset + bp->b_bufsize) 2816 eoff = bp->b_offset + bp->b_bufsize; 2817 m = bp->b_pages[i]; 2818 vfs_page_set_validclean(bp, foff, m); 2819 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2820 foff = noff; 2821 } 2822 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 2823} 2824 2825static void 2826vfs_setdirty_locked_object(struct buf *bp) 2827{ 2828 vm_object_t object; 2829 int i; 2830 2831 object = bp->b_bufobj->bo_object; 2832 VM_OBJECT_ASSERT_WLOCKED(object); 2833 2834 /* 2835 * We qualify the scan for modified pages on whether the 2836 * object has been flushed yet. 2837 */ 2838 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { 2839 vm_offset_t boffset; 2840 vm_offset_t eoffset; 2841 2842 /* 2843 * test the pages to see if they have been modified directly 2844 * by users through the VM system. 2845 */ 2846 for (i = 0; i < bp->b_npages; i++) 2847 vm_page_test_dirty(bp->b_pages[i]); 2848 2849 /* 2850 * Calculate the encompassing dirty range, boffset and eoffset, 2851 * (eoffset - boffset) bytes. 2852 */ 2853 2854 for (i = 0; i < bp->b_npages; i++) { 2855 if (bp->b_pages[i]->dirty) 2856 break; 2857 } 2858 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2859 2860 for (i = bp->b_npages - 1; i >= 0; --i) { 2861 if (bp->b_pages[i]->dirty) { 2862 break; 2863 } 2864 } 2865 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2866 2867 /* 2868 * Fit it to the buffer. 2869 */ 2870 2871 if (eoffset > bp->b_bcount) 2872 eoffset = bp->b_bcount; 2873 2874 /* 2875 * If we have a good dirty range, merge with the existing 2876 * dirty range. 2877 */ 2878 2879 if (boffset < eoffset) { 2880 if (bp->b_dirtyoff > boffset) 2881 bp->b_dirtyoff = boffset; 2882 if (bp->b_dirtyend < eoffset) 2883 bp->b_dirtyend = eoffset; 2884 } 2885 } 2886} 2887 2888/* 2889 * Allocate the KVA mapping for an existing buffer. It handles the 2890 * cases of both B_UNMAPPED buffer, and buffer with the preallocated 2891 * KVA which is not mapped (B_KVAALLOC). 2892 */ 2893static void 2894bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 2895{ 2896 struct buf *scratch_bp; 2897 int bsize, maxsize, need_mapping, need_kva; 2898 off_t offset; 2899 2900 need_mapping = (bp->b_flags & B_UNMAPPED) != 0 && 2901 (gbflags & GB_UNMAPPED) == 0; 2902 need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED && 2903 (gbflags & GB_KVAALLOC) != 0; 2904 if (!need_mapping && !need_kva) 2905 return; 2906 2907 BUF_CHECK_UNMAPPED(bp); 2908 2909 if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) { 2910 /* 2911 * Buffer is not mapped, but the KVA was already 2912 * reserved at the time of the instantiation. Use the 2913 * allocated space. 2914 */ 2915 bp->b_flags &= ~B_KVAALLOC; 2916 KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0")); 2917 bp->b_kvabase = bp->b_kvaalloc; 2918 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize); 2919 goto has_addr; 2920 } 2921 2922 /* 2923 * Calculate the amount of the address space we would reserve 2924 * if the buffer was mapped. 2925 */ 2926 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 2927 offset = blkno * bsize; 2928 maxsize = size + (offset & PAGE_MASK); 2929 maxsize = imax(maxsize, bsize); 2930 2931mapping_loop: 2932 if (allocbufkva(bp, maxsize, gbflags)) { 2933 /* 2934 * Request defragmentation. getnewbuf() returns us the 2935 * allocated space by the scratch buffer KVA. 2936 */ 2937 scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags | 2938 (GB_UNMAPPED | GB_KVAALLOC)); 2939 if (scratch_bp == NULL) { 2940 if ((gbflags & GB_NOWAIT_BD) != 0) { 2941 /* 2942 * XXXKIB: defragmentation cannot 2943 * succeed, not sure what else to do. 2944 */ 2945 panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp); 2946 } 2947 atomic_add_int(&mappingrestarts, 1); 2948 goto mapping_loop; 2949 } 2950 KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0, 2951 ("scratch bp !B_KVAALLOC %p", scratch_bp)); 2952 setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc, 2953 scratch_bp->b_kvasize, gbflags); 2954 2955 /* Get rid of the scratch buffer. */ 2956 scratch_bp->b_kvasize = 0; 2957 scratch_bp->b_flags |= B_INVAL; 2958 scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC); 2959 brelse(scratch_bp); 2960 } 2961 if (!need_mapping) 2962 return; 2963 2964has_addr: 2965 bp->b_saveaddr = bp->b_kvabase; 2966 bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */ 2967 bp->b_flags &= ~B_UNMAPPED; 2968 BUF_CHECK_MAPPED(bp); 2969 bpmap_qenter(bp); 2970} 2971 2972/* 2973 * getblk: 2974 * 2975 * Get a block given a specified block and offset into a file/device. 2976 * The buffers B_DONE bit will be cleared on return, making it almost 2977 * ready for an I/O initiation. B_INVAL may or may not be set on 2978 * return. The caller should clear B_INVAL prior to initiating a 2979 * READ. 2980 * 2981 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2982 * an existing buffer. 2983 * 2984 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2985 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2986 * and then cleared based on the backing VM. If the previous buffer is 2987 * non-0-sized but invalid, B_CACHE will be cleared. 2988 * 2989 * If getblk() must create a new buffer, the new buffer is returned with 2990 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2991 * case it is returned with B_INVAL clear and B_CACHE set based on the 2992 * backing VM. 2993 * 2994 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2995 * B_CACHE bit is clear. 2996 * 2997 * What this means, basically, is that the caller should use B_CACHE to 2998 * determine whether the buffer is fully valid or not and should clear 2999 * B_INVAL prior to issuing a read. If the caller intends to validate 3000 * the buffer by loading its data area with something, the caller needs 3001 * to clear B_INVAL. If the caller does this without issuing an I/O, 3002 * the caller should set B_CACHE ( as an optimization ), else the caller 3003 * should issue the I/O and biodone() will set B_CACHE if the I/O was 3004 * a write attempt or if it was a successfull read. If the caller 3005 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 3006 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 3007 */ 3008struct buf * 3009getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 3010 int flags) 3011{ 3012 struct buf *bp; 3013 struct bufobj *bo; 3014 int bsize, error, maxsize, vmio; 3015 off_t offset; 3016 3017 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 3018 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3019 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3020 ASSERT_VOP_LOCKED(vp, "getblk"); 3021 if (size > MAXBSIZE) 3022 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 3023 if (!unmapped_buf_allowed) 3024 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3025 3026 bo = &vp->v_bufobj; 3027loop: 3028 /* 3029 * Block if we are low on buffers. Certain processes are allowed 3030 * to completely exhaust the buffer cache. 3031 * 3032 * If this check ever becomes a bottleneck it may be better to 3033 * move it into the else, when gbincore() fails. At the moment 3034 * it isn't a problem. 3035 */ 3036 if (numfreebuffers == 0) { 3037 if (TD_IS_IDLETHREAD(curthread)) 3038 return NULL; 3039 mtx_lock(&nblock); 3040 needsbuffer |= VFS_BIO_NEED_ANY; 3041 mtx_unlock(&nblock); 3042 } 3043 3044 BO_LOCK(bo); 3045 bp = gbincore(bo, blkno); 3046 if (bp != NULL) { 3047 int lockflags; 3048 /* 3049 * Buffer is in-core. If the buffer is not busy nor managed, 3050 * it must be on a queue. 3051 */ 3052 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 3053 3054 if (flags & GB_LOCK_NOWAIT) 3055 lockflags |= LK_NOWAIT; 3056 3057 error = BUF_TIMELOCK(bp, lockflags, 3058 BO_MTX(bo), "getblk", slpflag, slptimeo); 3059 3060 /* 3061 * If we slept and got the lock we have to restart in case 3062 * the buffer changed identities. 3063 */ 3064 if (error == ENOLCK) 3065 goto loop; 3066 /* We timed out or were interrupted. */ 3067 else if (error) 3068 return (NULL); 3069 /* If recursed, assume caller knows the rules. */ 3070 else if (BUF_LOCKRECURSED(bp)) 3071 goto end; 3072 3073 /* 3074 * The buffer is locked. B_CACHE is cleared if the buffer is 3075 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 3076 * and for a VMIO buffer B_CACHE is adjusted according to the 3077 * backing VM cache. 3078 */ 3079 if (bp->b_flags & B_INVAL) 3080 bp->b_flags &= ~B_CACHE; 3081 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 3082 bp->b_flags |= B_CACHE; 3083 if (bp->b_flags & B_MANAGED) 3084 MPASS(bp->b_qindex == QUEUE_NONE); 3085 else { 3086 BO_LOCK(bo); 3087 bremfree(bp); 3088 BO_UNLOCK(bo); 3089 } 3090 3091 /* 3092 * check for size inconsistencies for non-VMIO case. 3093 */ 3094 if (bp->b_bcount != size) { 3095 if ((bp->b_flags & B_VMIO) == 0 || 3096 (size > bp->b_kvasize)) { 3097 if (bp->b_flags & B_DELWRI) { 3098 /* 3099 * If buffer is pinned and caller does 3100 * not want sleep waiting for it to be 3101 * unpinned, bail out 3102 * */ 3103 if (bp->b_pin_count > 0) { 3104 if (flags & GB_LOCK_NOWAIT) { 3105 bqrelse(bp); 3106 return (NULL); 3107 } else { 3108 bunpin_wait(bp); 3109 } 3110 } 3111 bp->b_flags |= B_NOCACHE; 3112 bwrite(bp); 3113 } else { 3114 if (LIST_EMPTY(&bp->b_dep)) { 3115 bp->b_flags |= B_RELBUF; 3116 brelse(bp); 3117 } else { 3118 bp->b_flags |= B_NOCACHE; 3119 bwrite(bp); 3120 } 3121 } 3122 goto loop; 3123 } 3124 } 3125 3126 /* 3127 * Handle the case of unmapped buffer which should 3128 * become mapped, or the buffer for which KVA 3129 * reservation is requested. 3130 */ 3131 bp_unmapped_get_kva(bp, blkno, size, flags); 3132 3133 /* 3134 * If the size is inconsistant in the VMIO case, we can resize 3135 * the buffer. This might lead to B_CACHE getting set or 3136 * cleared. If the size has not changed, B_CACHE remains 3137 * unchanged from its previous state. 3138 */ 3139 if (bp->b_bcount != size) 3140 allocbuf(bp, size); 3141 3142 KASSERT(bp->b_offset != NOOFFSET, 3143 ("getblk: no buffer offset")); 3144 3145 /* 3146 * A buffer with B_DELWRI set and B_CACHE clear must 3147 * be committed before we can return the buffer in 3148 * order to prevent the caller from issuing a read 3149 * ( due to B_CACHE not being set ) and overwriting 3150 * it. 3151 * 3152 * Most callers, including NFS and FFS, need this to 3153 * operate properly either because they assume they 3154 * can issue a read if B_CACHE is not set, or because 3155 * ( for example ) an uncached B_DELWRI might loop due 3156 * to softupdates re-dirtying the buffer. In the latter 3157 * case, B_CACHE is set after the first write completes, 3158 * preventing further loops. 3159 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3160 * above while extending the buffer, we cannot allow the 3161 * buffer to remain with B_CACHE set after the write 3162 * completes or it will represent a corrupt state. To 3163 * deal with this we set B_NOCACHE to scrap the buffer 3164 * after the write. 3165 * 3166 * We might be able to do something fancy, like setting 3167 * B_CACHE in bwrite() except if B_DELWRI is already set, 3168 * so the below call doesn't set B_CACHE, but that gets real 3169 * confusing. This is much easier. 3170 */ 3171 3172 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3173 bp->b_flags |= B_NOCACHE; 3174 bwrite(bp); 3175 goto loop; 3176 } 3177 bp->b_flags &= ~B_DONE; 3178 } else { 3179 /* 3180 * Buffer is not in-core, create new buffer. The buffer 3181 * returned by getnewbuf() is locked. Note that the returned 3182 * buffer is also considered valid (not marked B_INVAL). 3183 */ 3184 BO_UNLOCK(bo); 3185 /* 3186 * If the user does not want us to create the buffer, bail out 3187 * here. 3188 */ 3189 if (flags & GB_NOCREAT) 3190 return NULL; 3191 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 3192 offset = blkno * bsize; 3193 vmio = vp->v_object != NULL; 3194 if (vmio) { 3195 maxsize = size + (offset & PAGE_MASK); 3196 } else { 3197 maxsize = size; 3198 /* Do not allow non-VMIO notmapped buffers. */ 3199 flags &= ~GB_UNMAPPED; 3200 } 3201 maxsize = imax(maxsize, bsize); 3202 3203 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 3204 if (bp == NULL) { 3205 if (slpflag || slptimeo) 3206 return NULL; 3207 goto loop; 3208 } 3209 3210 /* 3211 * This code is used to make sure that a buffer is not 3212 * created while the getnewbuf routine is blocked. 3213 * This can be a problem whether the vnode is locked or not. 3214 * If the buffer is created out from under us, we have to 3215 * throw away the one we just created. 3216 * 3217 * Note: this must occur before we associate the buffer 3218 * with the vp especially considering limitations in 3219 * the splay tree implementation when dealing with duplicate 3220 * lblkno's. 3221 */ 3222 BO_LOCK(bo); 3223 if (gbincore(bo, blkno)) { 3224 BO_UNLOCK(bo); 3225 bp->b_flags |= B_INVAL; 3226 brelse(bp); 3227 goto loop; 3228 } 3229 3230 /* 3231 * Insert the buffer into the hash, so that it can 3232 * be found by incore. 3233 */ 3234 bp->b_blkno = bp->b_lblkno = blkno; 3235 bp->b_offset = offset; 3236 bgetvp(vp, bp); 3237 BO_UNLOCK(bo); 3238 3239 /* 3240 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 3241 * buffer size starts out as 0, B_CACHE will be set by 3242 * allocbuf() for the VMIO case prior to it testing the 3243 * backing store for validity. 3244 */ 3245 3246 if (vmio) { 3247 bp->b_flags |= B_VMIO; 3248 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 3249 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 3250 bp, vp->v_object, bp->b_bufobj->bo_object)); 3251 } else { 3252 bp->b_flags &= ~B_VMIO; 3253 KASSERT(bp->b_bufobj->bo_object == NULL, 3254 ("ARGH! has b_bufobj->bo_object %p %p\n", 3255 bp, bp->b_bufobj->bo_object)); 3256 BUF_CHECK_MAPPED(bp); 3257 } 3258 3259 allocbuf(bp, size); 3260 bp->b_flags &= ~B_DONE; 3261 } 3262 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 3263 BUF_ASSERT_HELD(bp); 3264end: 3265 KASSERT(bp->b_bufobj == bo, 3266 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3267 return (bp); 3268} 3269 3270/* 3271 * Get an empty, disassociated buffer of given size. The buffer is initially 3272 * set to B_INVAL. 3273 */ 3274struct buf * 3275geteblk(int size, int flags) 3276{ 3277 struct buf *bp; 3278 int maxsize; 3279 3280 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3281 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 3282 if ((flags & GB_NOWAIT_BD) && 3283 (curthread->td_pflags & TDP_BUFNEED) != 0) 3284 return (NULL); 3285 } 3286 allocbuf(bp, size); 3287 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3288 BUF_ASSERT_HELD(bp); 3289 return (bp); 3290} 3291 3292 3293/* 3294 * This code constitutes the buffer memory from either anonymous system 3295 * memory (in the case of non-VMIO operations) or from an associated 3296 * VM object (in the case of VMIO operations). This code is able to 3297 * resize a buffer up or down. 3298 * 3299 * Note that this code is tricky, and has many complications to resolve 3300 * deadlock or inconsistant data situations. Tread lightly!!! 3301 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3302 * the caller. Calling this code willy nilly can result in the loss of data. 3303 * 3304 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3305 * B_CACHE for the non-VMIO case. 3306 */ 3307 3308int 3309allocbuf(struct buf *bp, int size) 3310{ 3311 int newbsize, mbsize; 3312 int i; 3313 3314 BUF_ASSERT_HELD(bp); 3315 3316 if (bp->b_kvasize < size) 3317 panic("allocbuf: buffer too small"); 3318 3319 if ((bp->b_flags & B_VMIO) == 0) { 3320 caddr_t origbuf; 3321 int origbufsize; 3322 /* 3323 * Just get anonymous memory from the kernel. Don't 3324 * mess with B_CACHE. 3325 */ 3326 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3327 if (bp->b_flags & B_MALLOC) 3328 newbsize = mbsize; 3329 else 3330 newbsize = round_page(size); 3331 3332 if (newbsize < bp->b_bufsize) { 3333 /* 3334 * malloced buffers are not shrunk 3335 */ 3336 if (bp->b_flags & B_MALLOC) { 3337 if (newbsize) { 3338 bp->b_bcount = size; 3339 } else { 3340 free(bp->b_data, M_BIOBUF); 3341 if (bp->b_bufsize) { 3342 atomic_subtract_long( 3343 &bufmallocspace, 3344 bp->b_bufsize); 3345 bufspacewakeup(); 3346 bp->b_bufsize = 0; 3347 } 3348 bp->b_saveaddr = bp->b_kvabase; 3349 bp->b_data = bp->b_saveaddr; 3350 bp->b_bcount = 0; 3351 bp->b_flags &= ~B_MALLOC; 3352 } 3353 return 1; 3354 } 3355 vm_hold_free_pages(bp, newbsize); 3356 } else if (newbsize > bp->b_bufsize) { 3357 /* 3358 * We only use malloced memory on the first allocation. 3359 * and revert to page-allocated memory when the buffer 3360 * grows. 3361 */ 3362 /* 3363 * There is a potential smp race here that could lead 3364 * to bufmallocspace slightly passing the max. It 3365 * is probably extremely rare and not worth worrying 3366 * over. 3367 */ 3368 if ( (bufmallocspace < maxbufmallocspace) && 3369 (bp->b_bufsize == 0) && 3370 (mbsize <= PAGE_SIZE/2)) { 3371 3372 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 3373 bp->b_bufsize = mbsize; 3374 bp->b_bcount = size; 3375 bp->b_flags |= B_MALLOC; 3376 atomic_add_long(&bufmallocspace, mbsize); 3377 return 1; 3378 } 3379 origbuf = NULL; 3380 origbufsize = 0; 3381 /* 3382 * If the buffer is growing on its other-than-first allocation, 3383 * then we revert to the page-allocation scheme. 3384 */ 3385 if (bp->b_flags & B_MALLOC) { 3386 origbuf = bp->b_data; 3387 origbufsize = bp->b_bufsize; 3388 bp->b_data = bp->b_kvabase; 3389 if (bp->b_bufsize) { 3390 atomic_subtract_long(&bufmallocspace, 3391 bp->b_bufsize); 3392 bufspacewakeup(); 3393 bp->b_bufsize = 0; 3394 } 3395 bp->b_flags &= ~B_MALLOC; 3396 newbsize = round_page(newbsize); 3397 } 3398 vm_hold_load_pages( 3399 bp, 3400 (vm_offset_t) bp->b_data + bp->b_bufsize, 3401 (vm_offset_t) bp->b_data + newbsize); 3402 if (origbuf) { 3403 bcopy(origbuf, bp->b_data, origbufsize); 3404 free(origbuf, M_BIOBUF); 3405 } 3406 } 3407 } else { 3408 int desiredpages; 3409 3410 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3411 desiredpages = (size == 0) ? 0 : 3412 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 3413 3414 if (bp->b_flags & B_MALLOC) 3415 panic("allocbuf: VMIO buffer can't be malloced"); 3416 /* 3417 * Set B_CACHE initially if buffer is 0 length or will become 3418 * 0-length. 3419 */ 3420 if (size == 0 || bp->b_bufsize == 0) 3421 bp->b_flags |= B_CACHE; 3422 3423 if (newbsize < bp->b_bufsize) { 3424 /* 3425 * DEV_BSIZE aligned new buffer size is less then the 3426 * DEV_BSIZE aligned existing buffer size. Figure out 3427 * if we have to remove any pages. 3428 */ 3429 if (desiredpages < bp->b_npages) { 3430 vm_page_t m; 3431 3432 if ((bp->b_flags & B_UNMAPPED) == 0) { 3433 BUF_CHECK_MAPPED(bp); 3434 pmap_qremove((vm_offset_t)trunc_page( 3435 (vm_offset_t)bp->b_data) + 3436 (desiredpages << PAGE_SHIFT), 3437 (bp->b_npages - desiredpages)); 3438 } else 3439 BUF_CHECK_UNMAPPED(bp); 3440 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3441 for (i = desiredpages; i < bp->b_npages; i++) { 3442 /* 3443 * the page is not freed here -- it 3444 * is the responsibility of 3445 * vnode_pager_setsize 3446 */ 3447 m = bp->b_pages[i]; 3448 KASSERT(m != bogus_page, 3449 ("allocbuf: bogus page found")); 3450 while (vm_page_sleep_if_busy(m, TRUE, 3451 "biodep")) 3452 continue; 3453 3454 bp->b_pages[i] = NULL; 3455 vm_page_lock(m); 3456 vm_page_unwire(m, 0); 3457 vm_page_unlock(m); 3458 } 3459 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3460 bp->b_npages = desiredpages; 3461 } 3462 } else if (size > bp->b_bcount) { 3463 /* 3464 * We are growing the buffer, possibly in a 3465 * byte-granular fashion. 3466 */ 3467 vm_object_t obj; 3468 vm_offset_t toff; 3469 vm_offset_t tinc; 3470 3471 /* 3472 * Step 1, bring in the VM pages from the object, 3473 * allocating them if necessary. We must clear 3474 * B_CACHE if these pages are not valid for the 3475 * range covered by the buffer. 3476 */ 3477 3478 obj = bp->b_bufobj->bo_object; 3479 3480 VM_OBJECT_WLOCK(obj); 3481 while (bp->b_npages < desiredpages) { 3482 vm_page_t m; 3483 3484 /* 3485 * We must allocate system pages since blocking 3486 * here could interfere with paging I/O, no 3487 * matter which process we are. 3488 * 3489 * We can only test VPO_BUSY here. Blocking on 3490 * m->busy might lead to a deadlock: 3491 * vm_fault->getpages->cluster_read->allocbuf 3492 * Thus, we specify VM_ALLOC_IGN_SBUSY. 3493 */ 3494 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + 3495 bp->b_npages, VM_ALLOC_NOBUSY | 3496 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3497 VM_ALLOC_RETRY | VM_ALLOC_IGN_SBUSY | 3498 VM_ALLOC_COUNT(desiredpages - bp->b_npages)); 3499 if (m->valid == 0) 3500 bp->b_flags &= ~B_CACHE; 3501 bp->b_pages[bp->b_npages] = m; 3502 ++bp->b_npages; 3503 } 3504 3505 /* 3506 * Step 2. We've loaded the pages into the buffer, 3507 * we have to figure out if we can still have B_CACHE 3508 * set. Note that B_CACHE is set according to the 3509 * byte-granular range ( bcount and size ), new the 3510 * aligned range ( newbsize ). 3511 * 3512 * The VM test is against m->valid, which is DEV_BSIZE 3513 * aligned. Needless to say, the validity of the data 3514 * needs to also be DEV_BSIZE aligned. Note that this 3515 * fails with NFS if the server or some other client 3516 * extends the file's EOF. If our buffer is resized, 3517 * B_CACHE may remain set! XXX 3518 */ 3519 3520 toff = bp->b_bcount; 3521 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3522 3523 while ((bp->b_flags & B_CACHE) && toff < size) { 3524 vm_pindex_t pi; 3525 3526 if (tinc > (size - toff)) 3527 tinc = size - toff; 3528 3529 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3530 PAGE_SHIFT; 3531 3532 vfs_buf_test_cache( 3533 bp, 3534 bp->b_offset, 3535 toff, 3536 tinc, 3537 bp->b_pages[pi] 3538 ); 3539 toff += tinc; 3540 tinc = PAGE_SIZE; 3541 } 3542 VM_OBJECT_WUNLOCK(obj); 3543 3544 /* 3545 * Step 3, fixup the KVM pmap. 3546 */ 3547 if ((bp->b_flags & B_UNMAPPED) == 0) 3548 bpmap_qenter(bp); 3549 else 3550 BUF_CHECK_UNMAPPED(bp); 3551 } 3552 } 3553 if (newbsize < bp->b_bufsize) 3554 bufspacewakeup(); 3555 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3556 bp->b_bcount = size; /* requested buffer size */ 3557 return 1; 3558} 3559 3560extern int inflight_transient_maps; 3561 3562void 3563biodone(struct bio *bp) 3564{ 3565 struct mtx *mtxp; 3566 void (*done)(struct bio *); 3567 vm_offset_t start, end; 3568 int transient; 3569 3570 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3571 mtx_lock(mtxp); 3572 bp->bio_flags |= BIO_DONE; 3573 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 3574 start = trunc_page((vm_offset_t)bp->bio_data); 3575 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 3576 transient = 1; 3577 } else { 3578 transient = 0; 3579 start = end = 0; 3580 } 3581 done = bp->bio_done; 3582 if (done == NULL) 3583 wakeup(bp); 3584 mtx_unlock(mtxp); 3585 if (done != NULL) 3586 done(bp); 3587 if (transient) { 3588 pmap_qremove(start, OFF_TO_IDX(end - start)); 3589 vm_map_remove(bio_transient_map, start, end); 3590 atomic_add_int(&inflight_transient_maps, -1); 3591 } 3592} 3593 3594/* 3595 * Wait for a BIO to finish. 3596 * 3597 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3598 * case is not yet clear. 3599 */ 3600int 3601biowait(struct bio *bp, const char *wchan) 3602{ 3603 struct mtx *mtxp; 3604 3605 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3606 mtx_lock(mtxp); 3607 while ((bp->bio_flags & BIO_DONE) == 0) 3608 msleep(bp, mtxp, PRIBIO, wchan, hz / 10); 3609 mtx_unlock(mtxp); 3610 if (bp->bio_error != 0) 3611 return (bp->bio_error); 3612 if (!(bp->bio_flags & BIO_ERROR)) 3613 return (0); 3614 return (EIO); 3615} 3616 3617void 3618biofinish(struct bio *bp, struct devstat *stat, int error) 3619{ 3620 3621 if (error) { 3622 bp->bio_error = error; 3623 bp->bio_flags |= BIO_ERROR; 3624 } 3625 if (stat != NULL) 3626 devstat_end_transaction_bio(stat, bp); 3627 biodone(bp); 3628} 3629 3630/* 3631 * bufwait: 3632 * 3633 * Wait for buffer I/O completion, returning error status. The buffer 3634 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3635 * error and cleared. 3636 */ 3637int 3638bufwait(struct buf *bp) 3639{ 3640 if (bp->b_iocmd == BIO_READ) 3641 bwait(bp, PRIBIO, "biord"); 3642 else 3643 bwait(bp, PRIBIO, "biowr"); 3644 if (bp->b_flags & B_EINTR) { 3645 bp->b_flags &= ~B_EINTR; 3646 return (EINTR); 3647 } 3648 if (bp->b_ioflags & BIO_ERROR) { 3649 return (bp->b_error ? bp->b_error : EIO); 3650 } else { 3651 return (0); 3652 } 3653} 3654 3655 /* 3656 * Call back function from struct bio back up to struct buf. 3657 */ 3658static void 3659bufdonebio(struct bio *bip) 3660{ 3661 struct buf *bp; 3662 3663 bp = bip->bio_caller2; 3664 bp->b_resid = bp->b_bcount - bip->bio_completed; 3665 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3666 bp->b_ioflags = bip->bio_flags; 3667 bp->b_error = bip->bio_error; 3668 if (bp->b_error) 3669 bp->b_ioflags |= BIO_ERROR; 3670 bufdone(bp); 3671 g_destroy_bio(bip); 3672} 3673 3674void 3675dev_strategy(struct cdev *dev, struct buf *bp) 3676{ 3677 struct cdevsw *csw; 3678 struct bio *bip; 3679 int ref; 3680 3681 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3682 panic("b_iocmd botch"); 3683 for (;;) { 3684 bip = g_new_bio(); 3685 if (bip != NULL) 3686 break; 3687 /* Try again later */ 3688 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3689 } 3690 bip->bio_cmd = bp->b_iocmd; 3691 bip->bio_offset = bp->b_iooffset; 3692 bip->bio_length = bp->b_bcount; 3693 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3694 bdata2bio(bp, bip); 3695 bip->bio_done = bufdonebio; 3696 bip->bio_caller2 = bp; 3697 bip->bio_dev = dev; 3698 KASSERT(dev->si_refcount > 0, 3699 ("dev_strategy on un-referenced struct cdev *(%s)", 3700 devtoname(dev))); 3701 csw = dev_refthread(dev, &ref); 3702 if (csw == NULL) { 3703 g_destroy_bio(bip); 3704 bp->b_error = ENXIO; 3705 bp->b_ioflags = BIO_ERROR; 3706 bufdone(bp); 3707 return; 3708 } 3709 (*csw->d_strategy)(bip); 3710 dev_relthread(dev, ref); 3711} 3712 3713/* 3714 * bufdone: 3715 * 3716 * Finish I/O on a buffer, optionally calling a completion function. 3717 * This is usually called from an interrupt so process blocking is 3718 * not allowed. 3719 * 3720 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3721 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3722 * assuming B_INVAL is clear. 3723 * 3724 * For the VMIO case, we set B_CACHE if the op was a read and no 3725 * read error occured, or if the op was a write. B_CACHE is never 3726 * set if the buffer is invalid or otherwise uncacheable. 3727 * 3728 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3729 * initiator to leave B_INVAL set to brelse the buffer out of existance 3730 * in the biodone routine. 3731 */ 3732void 3733bufdone(struct buf *bp) 3734{ 3735 struct bufobj *dropobj; 3736 void (*biodone)(struct buf *); 3737 3738 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3739 dropobj = NULL; 3740 3741 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3742 BUF_ASSERT_HELD(bp); 3743 3744 runningbufwakeup(bp); 3745 if (bp->b_iocmd == BIO_WRITE) 3746 dropobj = bp->b_bufobj; 3747 /* call optional completion function if requested */ 3748 if (bp->b_iodone != NULL) { 3749 biodone = bp->b_iodone; 3750 bp->b_iodone = NULL; 3751 (*biodone) (bp); 3752 if (dropobj) 3753 bufobj_wdrop(dropobj); 3754 return; 3755 } 3756 3757 bufdone_finish(bp); 3758 3759 if (dropobj) 3760 bufobj_wdrop(dropobj); 3761} 3762 3763void 3764bufdone_finish(struct buf *bp) 3765{ 3766 BUF_ASSERT_HELD(bp); 3767 3768 if (!LIST_EMPTY(&bp->b_dep)) 3769 buf_complete(bp); 3770 3771 if (bp->b_flags & B_VMIO) { 3772 vm_ooffset_t foff; 3773 vm_page_t m; 3774 vm_object_t obj; 3775 struct vnode *vp; 3776 int bogus, i, iosize; 3777 3778 obj = bp->b_bufobj->bo_object; 3779 KASSERT(obj->paging_in_progress >= bp->b_npages, 3780 ("biodone_finish: paging in progress(%d) < b_npages(%d)", 3781 obj->paging_in_progress, bp->b_npages)); 3782 3783 vp = bp->b_vp; 3784 KASSERT(vp->v_holdcnt > 0, 3785 ("biodone_finish: vnode %p has zero hold count", vp)); 3786 KASSERT(vp->v_object != NULL, 3787 ("biodone_finish: vnode %p has no vm_object", vp)); 3788 3789 foff = bp->b_offset; 3790 KASSERT(bp->b_offset != NOOFFSET, 3791 ("biodone_finish: bp %p has no buffer offset", bp)); 3792 3793 /* 3794 * Set B_CACHE if the op was a normal read and no error 3795 * occured. B_CACHE is set for writes in the b*write() 3796 * routines. 3797 */ 3798 iosize = bp->b_bcount - bp->b_resid; 3799 if (bp->b_iocmd == BIO_READ && 3800 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3801 !(bp->b_ioflags & BIO_ERROR)) { 3802 bp->b_flags |= B_CACHE; 3803 } 3804 bogus = 0; 3805 VM_OBJECT_WLOCK(obj); 3806 for (i = 0; i < bp->b_npages; i++) { 3807 int bogusflag = 0; 3808 int resid; 3809 3810 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3811 if (resid > iosize) 3812 resid = iosize; 3813 3814 /* 3815 * cleanup bogus pages, restoring the originals 3816 */ 3817 m = bp->b_pages[i]; 3818 if (m == bogus_page) { 3819 bogus = bogusflag = 1; 3820 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3821 if (m == NULL) 3822 panic("biodone: page disappeared!"); 3823 bp->b_pages[i] = m; 3824 } 3825 KASSERT(OFF_TO_IDX(foff) == m->pindex, 3826 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch", 3827 (intmax_t)foff, (uintmax_t)m->pindex)); 3828 3829 /* 3830 * In the write case, the valid and clean bits are 3831 * already changed correctly ( see bdwrite() ), so we 3832 * only need to do this here in the read case. 3833 */ 3834 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3835 KASSERT((m->dirty & vm_page_bits(foff & 3836 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3837 " page %p has unexpected dirty bits", m)); 3838 vfs_page_set_valid(bp, foff, m); 3839 } 3840 3841 vm_page_io_finish(m); 3842 vm_object_pip_subtract(obj, 1); 3843 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3844 iosize -= resid; 3845 } 3846 vm_object_pip_wakeupn(obj, 0); 3847 VM_OBJECT_WUNLOCK(obj); 3848 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 3849 BUF_CHECK_MAPPED(bp); 3850 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3851 bp->b_pages, bp->b_npages); 3852 } 3853 } 3854 3855 /* 3856 * For asynchronous completions, release the buffer now. The brelse 3857 * will do a wakeup there if necessary - so no need to do a wakeup 3858 * here in the async case. The sync case always needs to do a wakeup. 3859 */ 3860 3861 if (bp->b_flags & B_ASYNC) { 3862 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3863 brelse(bp); 3864 else 3865 bqrelse(bp); 3866 } else 3867 bdone(bp); 3868} 3869 3870/* 3871 * This routine is called in lieu of iodone in the case of 3872 * incomplete I/O. This keeps the busy status for pages 3873 * consistant. 3874 */ 3875void 3876vfs_unbusy_pages(struct buf *bp) 3877{ 3878 int i; 3879 vm_object_t obj; 3880 vm_page_t m; 3881 3882 runningbufwakeup(bp); 3883 if (!(bp->b_flags & B_VMIO)) 3884 return; 3885 3886 obj = bp->b_bufobj->bo_object; 3887 VM_OBJECT_WLOCK(obj); 3888 for (i = 0; i < bp->b_npages; i++) { 3889 m = bp->b_pages[i]; 3890 if (m == bogus_page) { 3891 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3892 if (!m) 3893 panic("vfs_unbusy_pages: page missing\n"); 3894 bp->b_pages[i] = m; 3895 if ((bp->b_flags & B_UNMAPPED) == 0) { 3896 BUF_CHECK_MAPPED(bp); 3897 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3898 bp->b_pages, bp->b_npages); 3899 } else 3900 BUF_CHECK_UNMAPPED(bp); 3901 } 3902 vm_object_pip_subtract(obj, 1); 3903 vm_page_io_finish(m); 3904 } 3905 vm_object_pip_wakeupn(obj, 0); 3906 VM_OBJECT_WUNLOCK(obj); 3907} 3908 3909/* 3910 * vfs_page_set_valid: 3911 * 3912 * Set the valid bits in a page based on the supplied offset. The 3913 * range is restricted to the buffer's size. 3914 * 3915 * This routine is typically called after a read completes. 3916 */ 3917static void 3918vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3919{ 3920 vm_ooffset_t eoff; 3921 3922 /* 3923 * Compute the end offset, eoff, such that [off, eoff) does not span a 3924 * page boundary and eoff is not greater than the end of the buffer. 3925 * The end of the buffer, in this case, is our file EOF, not the 3926 * allocation size of the buffer. 3927 */ 3928 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3929 if (eoff > bp->b_offset + bp->b_bcount) 3930 eoff = bp->b_offset + bp->b_bcount; 3931 3932 /* 3933 * Set valid range. This is typically the entire buffer and thus the 3934 * entire page. 3935 */ 3936 if (eoff > off) 3937 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 3938} 3939 3940/* 3941 * vfs_page_set_validclean: 3942 * 3943 * Set the valid bits and clear the dirty bits in a page based on the 3944 * supplied offset. The range is restricted to the buffer's size. 3945 */ 3946static void 3947vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3948{ 3949 vm_ooffset_t soff, eoff; 3950 3951 /* 3952 * Start and end offsets in buffer. eoff - soff may not cross a 3953 * page boundry or cross the end of the buffer. The end of the 3954 * buffer, in this case, is our file EOF, not the allocation size 3955 * of the buffer. 3956 */ 3957 soff = off; 3958 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3959 if (eoff > bp->b_offset + bp->b_bcount) 3960 eoff = bp->b_offset + bp->b_bcount; 3961 3962 /* 3963 * Set valid range. This is typically the entire buffer and thus the 3964 * entire page. 3965 */ 3966 if (eoff > soff) { 3967 vm_page_set_validclean( 3968 m, 3969 (vm_offset_t) (soff & PAGE_MASK), 3970 (vm_offset_t) (eoff - soff) 3971 ); 3972 } 3973} 3974 3975/* 3976 * Ensure that all buffer pages are not busied by VPO_BUSY flag. If 3977 * any page is busy, drain the flag. 3978 */ 3979static void 3980vfs_drain_busy_pages(struct buf *bp) 3981{ 3982 vm_page_t m; 3983 int i, last_busied; 3984 3985 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); 3986 last_busied = 0; 3987 for (i = 0; i < bp->b_npages; i++) { 3988 m = bp->b_pages[i]; 3989 if ((m->oflags & VPO_BUSY) != 0) { 3990 for (; last_busied < i; last_busied++) 3991 vm_page_busy(bp->b_pages[last_busied]); 3992 while ((m->oflags & VPO_BUSY) != 0) 3993 vm_page_sleep(m, "vbpage"); 3994 } 3995 } 3996 for (i = 0; i < last_busied; i++) 3997 vm_page_wakeup(bp->b_pages[i]); 3998} 3999 4000/* 4001 * This routine is called before a device strategy routine. 4002 * It is used to tell the VM system that paging I/O is in 4003 * progress, and treat the pages associated with the buffer 4004 * almost as being VPO_BUSY. Also the object paging_in_progress 4005 * flag is handled to make sure that the object doesn't become 4006 * inconsistant. 4007 * 4008 * Since I/O has not been initiated yet, certain buffer flags 4009 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 4010 * and should be ignored. 4011 */ 4012void 4013vfs_busy_pages(struct buf *bp, int clear_modify) 4014{ 4015 int i, bogus; 4016 vm_object_t obj; 4017 vm_ooffset_t foff; 4018 vm_page_t m; 4019 4020 if (!(bp->b_flags & B_VMIO)) 4021 return; 4022 4023 obj = bp->b_bufobj->bo_object; 4024 foff = bp->b_offset; 4025 KASSERT(bp->b_offset != NOOFFSET, 4026 ("vfs_busy_pages: no buffer offset")); 4027 VM_OBJECT_WLOCK(obj); 4028 vfs_drain_busy_pages(bp); 4029 if (bp->b_bufsize != 0) 4030 vfs_setdirty_locked_object(bp); 4031 bogus = 0; 4032 for (i = 0; i < bp->b_npages; i++) { 4033 m = bp->b_pages[i]; 4034 4035 if ((bp->b_flags & B_CLUSTER) == 0) { 4036 vm_object_pip_add(obj, 1); 4037 vm_page_io_start(m); 4038 } 4039 /* 4040 * When readying a buffer for a read ( i.e 4041 * clear_modify == 0 ), it is important to do 4042 * bogus_page replacement for valid pages in 4043 * partially instantiated buffers. Partially 4044 * instantiated buffers can, in turn, occur when 4045 * reconstituting a buffer from its VM backing store 4046 * base. We only have to do this if B_CACHE is 4047 * clear ( which causes the I/O to occur in the 4048 * first place ). The replacement prevents the read 4049 * I/O from overwriting potentially dirty VM-backed 4050 * pages. XXX bogus page replacement is, uh, bogus. 4051 * It may not work properly with small-block devices. 4052 * We need to find a better way. 4053 */ 4054 if (clear_modify) { 4055 pmap_remove_write(m); 4056 vfs_page_set_validclean(bp, foff, m); 4057 } else if (m->valid == VM_PAGE_BITS_ALL && 4058 (bp->b_flags & B_CACHE) == 0) { 4059 bp->b_pages[i] = bogus_page; 4060 bogus++; 4061 } 4062 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4063 } 4064 VM_OBJECT_WUNLOCK(obj); 4065 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) { 4066 BUF_CHECK_MAPPED(bp); 4067 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4068 bp->b_pages, bp->b_npages); 4069 } 4070} 4071 4072/* 4073 * vfs_bio_set_valid: 4074 * 4075 * Set the range within the buffer to valid. The range is 4076 * relative to the beginning of the buffer, b_offset. Note that 4077 * b_offset itself may be offset from the beginning of the first 4078 * page. 4079 */ 4080void 4081vfs_bio_set_valid(struct buf *bp, int base, int size) 4082{ 4083 int i, n; 4084 vm_page_t m; 4085 4086 if (!(bp->b_flags & B_VMIO)) 4087 return; 4088 4089 /* 4090 * Fixup base to be relative to beginning of first page. 4091 * Set initial n to be the maximum number of bytes in the 4092 * first page that can be validated. 4093 */ 4094 base += (bp->b_offset & PAGE_MASK); 4095 n = PAGE_SIZE - (base & PAGE_MASK); 4096 4097 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4098 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4099 m = bp->b_pages[i]; 4100 if (n > size) 4101 n = size; 4102 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4103 base += n; 4104 size -= n; 4105 n = PAGE_SIZE; 4106 } 4107 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4108} 4109 4110/* 4111 * vfs_bio_clrbuf: 4112 * 4113 * If the specified buffer is a non-VMIO buffer, clear the entire 4114 * buffer. If the specified buffer is a VMIO buffer, clear and 4115 * validate only the previously invalid portions of the buffer. 4116 * This routine essentially fakes an I/O, so we need to clear 4117 * BIO_ERROR and B_INVAL. 4118 * 4119 * Note that while we only theoretically need to clear through b_bcount, 4120 * we go ahead and clear through b_bufsize. 4121 */ 4122void 4123vfs_bio_clrbuf(struct buf *bp) 4124{ 4125 int i, j, mask, sa, ea, slide; 4126 4127 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4128 clrbuf(bp); 4129 return; 4130 } 4131 bp->b_flags &= ~B_INVAL; 4132 bp->b_ioflags &= ~BIO_ERROR; 4133 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4134 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4135 (bp->b_offset & PAGE_MASK) == 0) { 4136 if (bp->b_pages[0] == bogus_page) 4137 goto unlock; 4138 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4139 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object); 4140 if ((bp->b_pages[0]->valid & mask) == mask) 4141 goto unlock; 4142 if ((bp->b_pages[0]->valid & mask) == 0) { 4143 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize); 4144 bp->b_pages[0]->valid |= mask; 4145 goto unlock; 4146 } 4147 } 4148 sa = bp->b_offset & PAGE_MASK; 4149 slide = 0; 4150 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4151 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4152 ea = slide & PAGE_MASK; 4153 if (ea == 0) 4154 ea = PAGE_SIZE; 4155 if (bp->b_pages[i] == bogus_page) 4156 continue; 4157 j = sa / DEV_BSIZE; 4158 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4159 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object); 4160 if ((bp->b_pages[i]->valid & mask) == mask) 4161 continue; 4162 if ((bp->b_pages[i]->valid & mask) == 0) 4163 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4164 else { 4165 for (; sa < ea; sa += DEV_BSIZE, j++) { 4166 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4167 pmap_zero_page_area(bp->b_pages[i], 4168 sa, DEV_BSIZE); 4169 } 4170 } 4171 } 4172 bp->b_pages[i]->valid |= mask; 4173 } 4174unlock: 4175 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4176 bp->b_resid = 0; 4177} 4178 4179void 4180vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4181{ 4182 vm_page_t m; 4183 int i, n; 4184 4185 if ((bp->b_flags & B_UNMAPPED) == 0) { 4186 BUF_CHECK_MAPPED(bp); 4187 bzero(bp->b_data + base, size); 4188 } else { 4189 BUF_CHECK_UNMAPPED(bp); 4190 n = PAGE_SIZE - (base & PAGE_MASK); 4191 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4192 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4193 m = bp->b_pages[i]; 4194 if (n > size) 4195 n = size; 4196 pmap_zero_page_area(m, base & PAGE_MASK, n); 4197 base += n; 4198 size -= n; 4199 n = PAGE_SIZE; 4200 } 4201 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4202 } 4203} 4204 4205/* 4206 * vm_hold_load_pages and vm_hold_free_pages get pages into 4207 * a buffers address space. The pages are anonymous and are 4208 * not associated with a file object. 4209 */ 4210static void 4211vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4212{ 4213 vm_offset_t pg; 4214 vm_page_t p; 4215 int index; 4216 4217 BUF_CHECK_MAPPED(bp); 4218 4219 to = round_page(to); 4220 from = round_page(from); 4221 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4222 4223 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4224tryagain: 4225 /* 4226 * note: must allocate system pages since blocking here 4227 * could interfere with paging I/O, no matter which 4228 * process we are. 4229 */ 4230 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | 4231 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT)); 4232 if (p == NULL) { 4233 VM_WAIT; 4234 goto tryagain; 4235 } 4236 pmap_qenter(pg, &p, 1); 4237 bp->b_pages[index] = p; 4238 } 4239 bp->b_npages = index; 4240} 4241 4242/* Return pages associated with this buf to the vm system */ 4243static void 4244vm_hold_free_pages(struct buf *bp, int newbsize) 4245{ 4246 vm_offset_t from; 4247 vm_page_t p; 4248 int index, newnpages; 4249 4250 BUF_CHECK_MAPPED(bp); 4251 4252 from = round_page((vm_offset_t)bp->b_data + newbsize); 4253 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4254 if (bp->b_npages > newnpages) 4255 pmap_qremove(from, bp->b_npages - newnpages); 4256 for (index = newnpages; index < bp->b_npages; index++) { 4257 p = bp->b_pages[index]; 4258 bp->b_pages[index] = NULL; 4259 if (p->busy != 0) 4260 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 4261 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); 4262 p->wire_count--; 4263 vm_page_free(p); 4264 atomic_subtract_int(&cnt.v_wire_count, 1); 4265 } 4266 bp->b_npages = newnpages; 4267} 4268 4269/* 4270 * Map an IO request into kernel virtual address space. 4271 * 4272 * All requests are (re)mapped into kernel VA space. 4273 * Notice that we use b_bufsize for the size of the buffer 4274 * to be mapped. b_bcount might be modified by the driver. 4275 * 4276 * Note that even if the caller determines that the address space should 4277 * be valid, a race or a smaller-file mapped into a larger space may 4278 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 4279 * check the return value. 4280 */ 4281int 4282vmapbuf(struct buf *bp, int mapbuf) 4283{ 4284 caddr_t kva; 4285 vm_prot_t prot; 4286 int pidx; 4287 4288 if (bp->b_bufsize < 0) 4289 return (-1); 4290 prot = VM_PROT_READ; 4291 if (bp->b_iocmd == BIO_READ) 4292 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 4293 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 4294 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, 4295 btoc(MAXPHYS))) < 0) 4296 return (-1); 4297 bp->b_npages = pidx; 4298 if (mapbuf || !unmapped_buf_allowed) { 4299 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 4300 kva = bp->b_saveaddr; 4301 bp->b_saveaddr = bp->b_data; 4302 bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK); 4303 bp->b_flags &= ~B_UNMAPPED; 4304 } else { 4305 bp->b_flags |= B_UNMAPPED; 4306 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK; 4307 bp->b_saveaddr = bp->b_data; 4308 bp->b_data = unmapped_buf; 4309 } 4310 return(0); 4311} 4312 4313/* 4314 * Free the io map PTEs associated with this IO operation. 4315 * We also invalidate the TLB entries and restore the original b_addr. 4316 */ 4317void 4318vunmapbuf(struct buf *bp) 4319{ 4320 int npages; 4321 4322 npages = bp->b_npages; 4323 if (bp->b_flags & B_UNMAPPED) 4324 bp->b_flags &= ~B_UNMAPPED; 4325 else 4326 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4327 vm_page_unhold_pages(bp->b_pages, npages); 4328 4329 bp->b_data = bp->b_saveaddr; 4330} 4331 4332void 4333bdone(struct buf *bp) 4334{ 4335 struct mtx *mtxp; 4336 4337 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4338 mtx_lock(mtxp); 4339 bp->b_flags |= B_DONE; 4340 wakeup(bp); 4341 mtx_unlock(mtxp); 4342} 4343 4344void 4345bwait(struct buf *bp, u_char pri, const char *wchan) 4346{ 4347 struct mtx *mtxp; 4348 4349 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4350 mtx_lock(mtxp); 4351 while ((bp->b_flags & B_DONE) == 0) 4352 msleep(bp, mtxp, pri, wchan, 0); 4353 mtx_unlock(mtxp); 4354} 4355 4356int 4357bufsync(struct bufobj *bo, int waitfor) 4358{ 4359 4360 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 4361} 4362 4363void 4364bufstrategy(struct bufobj *bo, struct buf *bp) 4365{ 4366 int i = 0; 4367 struct vnode *vp; 4368 4369 vp = bp->b_vp; 4370 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 4371 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 4372 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 4373 i = VOP_STRATEGY(vp, bp); 4374 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 4375} 4376 4377void 4378bufobj_wrefl(struct bufobj *bo) 4379{ 4380 4381 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4382 ASSERT_BO_LOCKED(bo); 4383 bo->bo_numoutput++; 4384} 4385 4386void 4387bufobj_wref(struct bufobj *bo) 4388{ 4389 4390 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4391 BO_LOCK(bo); 4392 bo->bo_numoutput++; 4393 BO_UNLOCK(bo); 4394} 4395 4396void 4397bufobj_wdrop(struct bufobj *bo) 4398{ 4399 4400 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 4401 BO_LOCK(bo); 4402 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 4403 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 4404 bo->bo_flag &= ~BO_WWAIT; 4405 wakeup(&bo->bo_numoutput); 4406 } 4407 BO_UNLOCK(bo); 4408} 4409 4410int 4411bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 4412{ 4413 int error; 4414 4415 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 4416 ASSERT_BO_LOCKED(bo); 4417 error = 0; 4418 while (bo->bo_numoutput) { 4419 bo->bo_flag |= BO_WWAIT; 4420 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 4421 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 4422 if (error) 4423 break; 4424 } 4425 return (error); 4426} 4427 4428void 4429bpin(struct buf *bp) 4430{ 4431 struct mtx *mtxp; 4432 4433 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4434 mtx_lock(mtxp); 4435 bp->b_pin_count++; 4436 mtx_unlock(mtxp); 4437} 4438 4439void 4440bunpin(struct buf *bp) 4441{ 4442 struct mtx *mtxp; 4443 4444 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4445 mtx_lock(mtxp); 4446 if (--bp->b_pin_count == 0) 4447 wakeup(bp); 4448 mtx_unlock(mtxp); 4449} 4450 4451void 4452bunpin_wait(struct buf *bp) 4453{ 4454 struct mtx *mtxp; 4455 4456 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4457 mtx_lock(mtxp); 4458 while (bp->b_pin_count > 0) 4459 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4460 mtx_unlock(mtxp); 4461} 4462 4463/* 4464 * Set bio_data or bio_ma for struct bio from the struct buf. 4465 */ 4466void 4467bdata2bio(struct buf *bp, struct bio *bip) 4468{ 4469 4470 if ((bp->b_flags & B_UNMAPPED) != 0) { 4471 KASSERT(unmapped_buf_allowed, ("unmapped")); 4472 bip->bio_ma = bp->b_pages; 4473 bip->bio_ma_n = bp->b_npages; 4474 bip->bio_data = unmapped_buf; 4475 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 4476 bip->bio_flags |= BIO_UNMAPPED; 4477 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 4478 PAGE_SIZE == bp->b_npages, 4479 ("Buffer %p too short: %d %d %d", bp, bip->bio_ma_offset, 4480 bip->bio_length, bip->bio_ma_n)); 4481 } else { 4482 bip->bio_data = bp->b_data; 4483 bip->bio_ma = NULL; 4484 } 4485} 4486 4487#include "opt_ddb.h" 4488#ifdef DDB 4489#include <ddb/ddb.h> 4490 4491/* DDB command to show buffer data */ 4492DB_SHOW_COMMAND(buffer, db_show_buffer) 4493{ 4494 /* get args */ 4495 struct buf *bp = (struct buf *)addr; 4496 4497 if (!have_addr) { 4498 db_printf("usage: show buffer <addr>\n"); 4499 return; 4500 } 4501 4502 db_printf("buf at %p\n", bp); 4503 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n", 4504 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, 4505 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS); 4506 db_printf( 4507 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4508 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, " 4509 "b_dep = %p\n", 4510 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4511 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4512 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first); 4513 if (bp->b_npages) { 4514 int i; 4515 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4516 for (i = 0; i < bp->b_npages; i++) { 4517 vm_page_t m; 4518 m = bp->b_pages[i]; 4519 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4520 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4521 if ((i + 1) < bp->b_npages) 4522 db_printf(","); 4523 } 4524 db_printf("\n"); 4525 } 4526 db_printf(" "); 4527 BUF_LOCKPRINTINFO(bp); 4528} 4529 4530DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4531{ 4532 struct buf *bp; 4533 int i; 4534 4535 for (i = 0; i < nbuf; i++) { 4536 bp = &buf[i]; 4537 if (BUF_ISLOCKED(bp)) { 4538 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4539 db_printf("\n"); 4540 } 4541 } 4542} 4543 4544DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4545{ 4546 struct vnode *vp; 4547 struct buf *bp; 4548 4549 if (!have_addr) { 4550 db_printf("usage: show vnodebufs <addr>\n"); 4551 return; 4552 } 4553 vp = (struct vnode *)addr; 4554 db_printf("Clean buffers:\n"); 4555 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4556 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4557 db_printf("\n"); 4558 } 4559 db_printf("Dirty buffers:\n"); 4560 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4561 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4562 db_printf("\n"); 4563 } 4564} 4565 4566DB_COMMAND(countfreebufs, db_coundfreebufs) 4567{ 4568 struct buf *bp; 4569 int i, used = 0, nfree = 0; 4570 4571 if (have_addr) { 4572 db_printf("usage: countfreebufs\n"); 4573 return; 4574 } 4575 4576 for (i = 0; i < nbuf; i++) { 4577 bp = &buf[i]; 4578 if ((bp->b_vflags & BV_INFREECNT) != 0) 4579 nfree++; 4580 else 4581 used++; 4582 } 4583 4584 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4585 nfree + used); 4586 db_printf("numfreebuffers is %d\n", numfreebuffers); 4587} 4588#endif /* DDB */ 4589