vfs_bio.c revision 248282
1/*- 2 * Copyright (c) 2004 Poul-Henning Kamp 3 * Copyright (c) 1994,1997 John S. Dyson 4 * All rights reserved. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28/* 29 * this file contains a new buffer I/O scheme implementing a coherent 30 * VM object and buffer cache scheme. Pains have been taken to make 31 * sure that the performance degradation associated with schemes such 32 * as this is not realized. 33 * 34 * Author: John S. Dyson 35 * Significant help during the development and debugging phases 36 * had been provided by David Greenman, also of the FreeBSD core team. 37 * 38 * see man buf(9) for more info. 39 */ 40 41#include <sys/cdefs.h> 42__FBSDID("$FreeBSD: head/sys/kern/vfs_bio.c 248282 2013-03-14 20:28:26Z kib $"); 43 44#include <sys/param.h> 45#include <sys/systm.h> 46#include <sys/bio.h> 47#include <sys/conf.h> 48#include <sys/buf.h> 49#include <sys/devicestat.h> 50#include <sys/eventhandler.h> 51#include <sys/fail.h> 52#include <sys/limits.h> 53#include <sys/lock.h> 54#include <sys/malloc.h> 55#include <sys/mount.h> 56#include <sys/mutex.h> 57#include <sys/kernel.h> 58#include <sys/kthread.h> 59#include <sys/proc.h> 60#include <sys/resourcevar.h> 61#include <sys/rwlock.h> 62#include <sys/sysctl.h> 63#include <sys/vmmeter.h> 64#include <sys/vnode.h> 65#include <geom/geom.h> 66#include <vm/vm.h> 67#include <vm/vm_param.h> 68#include <vm/vm_kern.h> 69#include <vm/vm_pageout.h> 70#include <vm/vm_page.h> 71#include <vm/vm_object.h> 72#include <vm/vm_extern.h> 73#include <vm/vm_map.h> 74#include "opt_compat.h" 75#include "opt_directio.h" 76#include "opt_swap.h" 77 78static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 79 80struct bio_ops bioops; /* I/O operation notification */ 81 82struct buf_ops buf_ops_bio = { 83 .bop_name = "buf_ops_bio", 84 .bop_write = bufwrite, 85 .bop_strategy = bufstrategy, 86 .bop_sync = bufsync, 87 .bop_bdflush = bufbdflush, 88}; 89 90/* 91 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 92 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 93 */ 94struct buf *buf; /* buffer header pool */ 95 96static struct proc *bufdaemonproc; 97 98static int inmem(struct vnode *vp, daddr_t blkno); 99static void vm_hold_free_pages(struct buf *bp, int newbsize); 100static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 101 vm_offset_t to); 102static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 103static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 104 vm_page_t m); 105static void vfs_drain_busy_pages(struct buf *bp); 106static void vfs_clean_pages_dirty_buf(struct buf *bp); 107static void vfs_setdirty_locked_object(struct buf *bp); 108static void vfs_vmio_release(struct buf *bp); 109static int vfs_bio_clcheck(struct vnode *vp, int size, 110 daddr_t lblkno, daddr_t blkno); 111static int buf_do_flush(struct vnode *vp); 112static int flushbufqueues(struct vnode *, int, int); 113static void buf_daemon(void); 114static void bremfreel(struct buf *bp); 115#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 116 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 117static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 118#endif 119 120int vmiodirenable = TRUE; 121SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 122 "Use the VM system for directory writes"); 123long runningbufspace; 124SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 125 "Amount of presently outstanding async buffer io"); 126static long bufspace; 127#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 128 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 129SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 130 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers"); 131#else 132SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 133 "Virtual memory used for buffers"); 134#endif 135static long maxbufspace; 136SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 137 "Maximum allowed value of bufspace (including buf_daemon)"); 138static long bufmallocspace; 139SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 140 "Amount of malloced memory for buffers"); 141static long maxbufmallocspace; 142SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 143 "Maximum amount of malloced memory for buffers"); 144static long lobufspace; 145SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 146 "Minimum amount of buffers we want to have"); 147long hibufspace; 148SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 149 "Maximum allowed value of bufspace (excluding buf_daemon)"); 150static int bufreusecnt; 151SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 152 "Number of times we have reused a buffer"); 153static int buffreekvacnt; 154SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 155 "Number of times we have freed the KVA space from some buffer"); 156static int bufdefragcnt; 157SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 158 "Number of times we have had to repeat buffer allocation to defragment"); 159static long lorunningspace; 160SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 161 "Minimum preferred space used for in-progress I/O"); 162static long hirunningspace; 163SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 164 "Maximum amount of space to use for in-progress I/O"); 165int dirtybufferflushes; 166SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 167 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 168int bdwriteskip; 169SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 170 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 171int altbufferflushes; 172SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 173 0, "Number of fsync flushes to limit dirty buffers"); 174static int recursiveflushes; 175SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 176 0, "Number of flushes skipped due to being recursive"); 177static int numdirtybuffers; 178SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 179 "Number of buffers that are dirty (has unwritten changes) at the moment"); 180static int lodirtybuffers; 181SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 182 "How many buffers we want to have free before bufdaemon can sleep"); 183static int hidirtybuffers; 184SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 185 "When the number of dirty buffers is considered severe"); 186int dirtybufthresh; 187SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 188 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 189static int numfreebuffers; 190SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 191 "Number of free buffers"); 192static int lofreebuffers; 193SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 194 "XXX Unused"); 195static int hifreebuffers; 196SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 197 "XXX Complicatedly unused"); 198static int getnewbufcalls; 199SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 200 "Number of calls to getnewbuf"); 201static int getnewbufrestarts; 202SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 203 "Number of times getnewbuf has had to restart a buffer aquisition"); 204static int flushbufqtarget = 100; 205SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 206 "Amount of work to do in flushbufqueues when helping bufdaemon"); 207static long notbufdflashes; 208SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, ¬bufdflashes, 0, 209 "Number of dirty buffer flushes done by the bufdaemon helpers"); 210static long barrierwrites; 211SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0, 212 "Number of barrier writes"); 213 214/* 215 * Wakeup point for bufdaemon, as well as indicator of whether it is already 216 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 217 * is idling. 218 */ 219static int bd_request; 220 221/* 222 * Request for the buf daemon to write more buffers than is indicated by 223 * lodirtybuf. This may be necessary to push out excess dependencies or 224 * defragment the address space where a simple count of the number of dirty 225 * buffers is insufficient to characterize the demand for flushing them. 226 */ 227static int bd_speedupreq; 228 229/* 230 * This lock synchronizes access to bd_request. 231 */ 232static struct mtx bdlock; 233 234/* 235 * bogus page -- for I/O to/from partially complete buffers 236 * this is a temporary solution to the problem, but it is not 237 * really that bad. it would be better to split the buffer 238 * for input in the case of buffers partially already in memory, 239 * but the code is intricate enough already. 240 */ 241vm_page_t bogus_page; 242 243/* 244 * Synchronization (sleep/wakeup) variable for active buffer space requests. 245 * Set when wait starts, cleared prior to wakeup(). 246 * Used in runningbufwakeup() and waitrunningbufspace(). 247 */ 248static int runningbufreq; 249 250/* 251 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 252 * waitrunningbufspace(). 253 */ 254static struct mtx rbreqlock; 255 256/* 257 * Synchronization (sleep/wakeup) variable for buffer requests. 258 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 259 * by and/or. 260 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 261 * getnewbuf(), and getblk(). 262 */ 263static int needsbuffer; 264 265/* 266 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 267 */ 268static struct mtx nblock; 269 270/* 271 * Definitions for the buffer free lists. 272 */ 273#define BUFFER_QUEUES 5 /* number of free buffer queues */ 274 275#define QUEUE_NONE 0 /* on no queue */ 276#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 277#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 278#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 279#define QUEUE_EMPTY 4 /* empty buffer headers */ 280#define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */ 281 282/* Queues for free buffers with various properties */ 283static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 284 285/* Lock for the bufqueues */ 286static struct mtx bqlock; 287 288/* 289 * Single global constant for BUF_WMESG, to avoid getting multiple references. 290 * buf_wmesg is referred from macros. 291 */ 292const char *buf_wmesg = BUF_WMESG; 293 294#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 295#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 296#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 297#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 298 299#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 300 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 301static int 302sysctl_bufspace(SYSCTL_HANDLER_ARGS) 303{ 304 long lvalue; 305 int ivalue; 306 307 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 308 return (sysctl_handle_long(oidp, arg1, arg2, req)); 309 lvalue = *(long *)arg1; 310 if (lvalue > INT_MAX) 311 /* On overflow, still write out a long to trigger ENOMEM. */ 312 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 313 ivalue = lvalue; 314 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 315} 316#endif 317 318#ifdef DIRECTIO 319extern void ffs_rawread_setup(void); 320#endif /* DIRECTIO */ 321/* 322 * numdirtywakeup: 323 * 324 * If someone is blocked due to there being too many dirty buffers, 325 * and numdirtybuffers is now reasonable, wake them up. 326 */ 327 328static __inline void 329numdirtywakeup(int level) 330{ 331 332 if (numdirtybuffers <= level) { 333 mtx_lock(&nblock); 334 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 335 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 336 wakeup(&needsbuffer); 337 } 338 mtx_unlock(&nblock); 339 } 340} 341 342/* 343 * bufspacewakeup: 344 * 345 * Called when buffer space is potentially available for recovery. 346 * getnewbuf() will block on this flag when it is unable to free 347 * sufficient buffer space. Buffer space becomes recoverable when 348 * bp's get placed back in the queues. 349 */ 350 351static __inline void 352bufspacewakeup(void) 353{ 354 355 /* 356 * If someone is waiting for BUF space, wake them up. Even 357 * though we haven't freed the kva space yet, the waiting 358 * process will be able to now. 359 */ 360 mtx_lock(&nblock); 361 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 362 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 363 wakeup(&needsbuffer); 364 } 365 mtx_unlock(&nblock); 366} 367 368/* 369 * runningbufwakeup() - in-progress I/O accounting. 370 * 371 */ 372void 373runningbufwakeup(struct buf *bp) 374{ 375 376 if (bp->b_runningbufspace) { 377 atomic_subtract_long(&runningbufspace, bp->b_runningbufspace); 378 bp->b_runningbufspace = 0; 379 mtx_lock(&rbreqlock); 380 if (runningbufreq && runningbufspace <= lorunningspace) { 381 runningbufreq = 0; 382 wakeup(&runningbufreq); 383 } 384 mtx_unlock(&rbreqlock); 385 } 386} 387 388/* 389 * bufcountwakeup: 390 * 391 * Called when a buffer has been added to one of the free queues to 392 * account for the buffer and to wakeup anyone waiting for free buffers. 393 * This typically occurs when large amounts of metadata are being handled 394 * by the buffer cache ( else buffer space runs out first, usually ). 395 */ 396 397static __inline void 398bufcountwakeup(struct buf *bp) 399{ 400 int old; 401 402 KASSERT((bp->b_vflags & BV_INFREECNT) == 0, 403 ("buf %p already counted as free", bp)); 404 if (bp->b_bufobj != NULL) 405 mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED); 406 bp->b_vflags |= BV_INFREECNT; 407 old = atomic_fetchadd_int(&numfreebuffers, 1); 408 KASSERT(old >= 0 && old < nbuf, 409 ("numfreebuffers climbed to %d", old + 1)); 410 mtx_lock(&nblock); 411 if (needsbuffer) { 412 needsbuffer &= ~VFS_BIO_NEED_ANY; 413 if (numfreebuffers >= hifreebuffers) 414 needsbuffer &= ~VFS_BIO_NEED_FREE; 415 wakeup(&needsbuffer); 416 } 417 mtx_unlock(&nblock); 418} 419 420/* 421 * waitrunningbufspace() 422 * 423 * runningbufspace is a measure of the amount of I/O currently 424 * running. This routine is used in async-write situations to 425 * prevent creating huge backups of pending writes to a device. 426 * Only asynchronous writes are governed by this function. 427 * 428 * Reads will adjust runningbufspace, but will not block based on it. 429 * The read load has a side effect of reducing the allowed write load. 430 * 431 * This does NOT turn an async write into a sync write. It waits 432 * for earlier writes to complete and generally returns before the 433 * caller's write has reached the device. 434 */ 435void 436waitrunningbufspace(void) 437{ 438 439 mtx_lock(&rbreqlock); 440 while (runningbufspace > hirunningspace) { 441 ++runningbufreq; 442 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 443 } 444 mtx_unlock(&rbreqlock); 445} 446 447 448/* 449 * vfs_buf_test_cache: 450 * 451 * Called when a buffer is extended. This function clears the B_CACHE 452 * bit if the newly extended portion of the buffer does not contain 453 * valid data. 454 */ 455static __inline 456void 457vfs_buf_test_cache(struct buf *bp, 458 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 459 vm_page_t m) 460{ 461 462 VM_OBJECT_ASSERT_WLOCKED(m->object); 463 if (bp->b_flags & B_CACHE) { 464 int base = (foff + off) & PAGE_MASK; 465 if (vm_page_is_valid(m, base, size) == 0) 466 bp->b_flags &= ~B_CACHE; 467 } 468} 469 470/* Wake up the buffer daemon if necessary */ 471static __inline 472void 473bd_wakeup(int dirtybuflevel) 474{ 475 476 mtx_lock(&bdlock); 477 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 478 bd_request = 1; 479 wakeup(&bd_request); 480 } 481 mtx_unlock(&bdlock); 482} 483 484/* 485 * bd_speedup - speedup the buffer cache flushing code 486 */ 487 488void 489bd_speedup(void) 490{ 491 int needwake; 492 493 mtx_lock(&bdlock); 494 needwake = 0; 495 if (bd_speedupreq == 0 || bd_request == 0) 496 needwake = 1; 497 bd_speedupreq = 1; 498 bd_request = 1; 499 if (needwake) 500 wakeup(&bd_request); 501 mtx_unlock(&bdlock); 502} 503 504/* 505 * Calculating buffer cache scaling values and reserve space for buffer 506 * headers. This is called during low level kernel initialization and 507 * may be called more then once. We CANNOT write to the memory area 508 * being reserved at this time. 509 */ 510caddr_t 511kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 512{ 513 int tuned_nbuf; 514 long maxbuf; 515 516 /* 517 * physmem_est is in pages. Convert it to kilobytes (assumes 518 * PAGE_SIZE is >= 1K) 519 */ 520 physmem_est = physmem_est * (PAGE_SIZE / 1024); 521 522 /* 523 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 524 * For the first 64MB of ram nominally allocate sufficient buffers to 525 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 526 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 527 * the buffer cache we limit the eventual kva reservation to 528 * maxbcache bytes. 529 * 530 * factor represents the 1/4 x ram conversion. 531 */ 532 if (nbuf == 0) { 533 int factor = 4 * BKVASIZE / 1024; 534 535 nbuf = 50; 536 if (physmem_est > 4096) 537 nbuf += min((physmem_est - 4096) / factor, 538 65536 / factor); 539 if (physmem_est > 65536) 540 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 541 542 if (maxbcache && nbuf > maxbcache / BKVASIZE) 543 nbuf = maxbcache / BKVASIZE; 544 tuned_nbuf = 1; 545 } else 546 tuned_nbuf = 0; 547 548 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 549 maxbuf = (LONG_MAX / 3) / BKVASIZE; 550 if (nbuf > maxbuf) { 551 if (!tuned_nbuf) 552 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 553 maxbuf); 554 nbuf = maxbuf; 555 } 556 557 /* 558 * swbufs are used as temporary holders for I/O, such as paging I/O. 559 * We have no less then 16 and no more then 256. 560 */ 561 nswbuf = max(min(nbuf/4, 256), 16); 562#ifdef NSWBUF_MIN 563 if (nswbuf < NSWBUF_MIN) 564 nswbuf = NSWBUF_MIN; 565#endif 566#ifdef DIRECTIO 567 ffs_rawread_setup(); 568#endif 569 570 /* 571 * Reserve space for the buffer cache buffers 572 */ 573 swbuf = (void *)v; 574 v = (caddr_t)(swbuf + nswbuf); 575 buf = (void *)v; 576 v = (caddr_t)(buf + nbuf); 577 578 return(v); 579} 580 581/* Initialize the buffer subsystem. Called before use of any buffers. */ 582void 583bufinit(void) 584{ 585 struct buf *bp; 586 int i; 587 588 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 589 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 590 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 591 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 592 593 /* next, make a null set of free lists */ 594 for (i = 0; i < BUFFER_QUEUES; i++) 595 TAILQ_INIT(&bufqueues[i]); 596 597 /* finally, initialize each buffer header and stick on empty q */ 598 for (i = 0; i < nbuf; i++) { 599 bp = &buf[i]; 600 bzero(bp, sizeof *bp); 601 bp->b_flags = B_INVAL; /* we're just an empty header */ 602 bp->b_rcred = NOCRED; 603 bp->b_wcred = NOCRED; 604 bp->b_qindex = QUEUE_EMPTY; 605 bp->b_vflags = BV_INFREECNT; /* buf is counted as free */ 606 bp->b_xflags = 0; 607 LIST_INIT(&bp->b_dep); 608 BUF_LOCKINIT(bp); 609 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 610 } 611 612 /* 613 * maxbufspace is the absolute maximum amount of buffer space we are 614 * allowed to reserve in KVM and in real terms. The absolute maximum 615 * is nominally used by buf_daemon. hibufspace is the nominal maximum 616 * used by most other processes. The differential is required to 617 * ensure that buf_daemon is able to run when other processes might 618 * be blocked waiting for buffer space. 619 * 620 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 621 * this may result in KVM fragmentation which is not handled optimally 622 * by the system. 623 */ 624 maxbufspace = (long)nbuf * BKVASIZE; 625 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 626 lobufspace = hibufspace - MAXBSIZE; 627 628 /* 629 * Note: The 16 MiB upper limit for hirunningspace was chosen 630 * arbitrarily and may need further tuning. It corresponds to 631 * 128 outstanding write IO requests (if IO size is 128 KiB), 632 * which fits with many RAID controllers' tagged queuing limits. 633 * The lower 1 MiB limit is the historical upper limit for 634 * hirunningspace. 635 */ 636 hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE), 637 16 * 1024 * 1024), 1024 * 1024); 638 lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE); 639 640/* 641 * Limit the amount of malloc memory since it is wired permanently into 642 * the kernel space. Even though this is accounted for in the buffer 643 * allocation, we don't want the malloced region to grow uncontrolled. 644 * The malloc scheme improves memory utilization significantly on average 645 * (small) directories. 646 */ 647 maxbufmallocspace = hibufspace / 20; 648 649/* 650 * Reduce the chance of a deadlock occuring by limiting the number 651 * of delayed-write dirty buffers we allow to stack up. 652 */ 653 hidirtybuffers = nbuf / 4 + 20; 654 dirtybufthresh = hidirtybuffers * 9 / 10; 655 numdirtybuffers = 0; 656/* 657 * To support extreme low-memory systems, make sure hidirtybuffers cannot 658 * eat up all available buffer space. This occurs when our minimum cannot 659 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 660 * BKVASIZE'd buffers. 661 */ 662 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 663 hidirtybuffers >>= 1; 664 } 665 lodirtybuffers = hidirtybuffers / 2; 666 667/* 668 * Try to keep the number of free buffers in the specified range, 669 * and give special processes (e.g. like buf_daemon) access to an 670 * emergency reserve. 671 */ 672 lofreebuffers = nbuf / 18 + 5; 673 hifreebuffers = 2 * lofreebuffers; 674 numfreebuffers = nbuf; 675 676 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 677 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 678} 679 680/* 681 * bfreekva() - free the kva allocation for a buffer. 682 * 683 * Since this call frees up buffer space, we call bufspacewakeup(). 684 */ 685static void 686bfreekva(struct buf *bp) 687{ 688 689 if (bp->b_kvasize) { 690 atomic_add_int(&buffreekvacnt, 1); 691 atomic_subtract_long(&bufspace, bp->b_kvasize); 692 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase, 693 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 694 bp->b_kvasize = 0; 695 bufspacewakeup(); 696 } 697} 698 699/* 700 * bremfree: 701 * 702 * Mark the buffer for removal from the appropriate free list in brelse. 703 * 704 */ 705void 706bremfree(struct buf *bp) 707{ 708 int old; 709 710 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 711 KASSERT((bp->b_flags & B_REMFREE) == 0, 712 ("bremfree: buffer %p already marked for delayed removal.", bp)); 713 KASSERT(bp->b_qindex != QUEUE_NONE, 714 ("bremfree: buffer %p not on a queue.", bp)); 715 BUF_ASSERT_HELD(bp); 716 717 bp->b_flags |= B_REMFREE; 718 /* Fixup numfreebuffers count. */ 719 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 720 KASSERT((bp->b_vflags & BV_INFREECNT) != 0, 721 ("buf %p not counted in numfreebuffers", bp)); 722 if (bp->b_bufobj != NULL) 723 mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED); 724 bp->b_vflags &= ~BV_INFREECNT; 725 old = atomic_fetchadd_int(&numfreebuffers, -1); 726 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 727 } 728} 729 730/* 731 * bremfreef: 732 * 733 * Force an immediate removal from a free list. Used only in nfs when 734 * it abuses the b_freelist pointer. 735 */ 736void 737bremfreef(struct buf *bp) 738{ 739 mtx_lock(&bqlock); 740 bremfreel(bp); 741 mtx_unlock(&bqlock); 742} 743 744/* 745 * bremfreel: 746 * 747 * Removes a buffer from the free list, must be called with the 748 * bqlock held. 749 */ 750static void 751bremfreel(struct buf *bp) 752{ 753 int old; 754 755 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 756 bp, bp->b_vp, bp->b_flags); 757 KASSERT(bp->b_qindex != QUEUE_NONE, 758 ("bremfreel: buffer %p not on a queue.", bp)); 759 BUF_ASSERT_HELD(bp); 760 mtx_assert(&bqlock, MA_OWNED); 761 762 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 763 bp->b_qindex = QUEUE_NONE; 764 /* 765 * If this was a delayed bremfree() we only need to remove the buffer 766 * from the queue and return the stats are already done. 767 */ 768 if (bp->b_flags & B_REMFREE) { 769 bp->b_flags &= ~B_REMFREE; 770 return; 771 } 772 /* 773 * Fixup numfreebuffers count. If the buffer is invalid or not 774 * delayed-write, the buffer was free and we must decrement 775 * numfreebuffers. 776 */ 777 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 778 KASSERT((bp->b_vflags & BV_INFREECNT) != 0, 779 ("buf %p not counted in numfreebuffers", bp)); 780 if (bp->b_bufobj != NULL) 781 mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED); 782 bp->b_vflags &= ~BV_INFREECNT; 783 old = atomic_fetchadd_int(&numfreebuffers, -1); 784 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 785 } 786} 787 788/* 789 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 790 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 791 * the buffer is valid and we do not have to do anything. 792 */ 793void 794breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 795 int cnt, struct ucred * cred) 796{ 797 struct buf *rabp; 798 int i; 799 800 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 801 if (inmem(vp, *rablkno)) 802 continue; 803 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 804 805 if ((rabp->b_flags & B_CACHE) == 0) { 806 if (!TD_IS_IDLETHREAD(curthread)) 807 curthread->td_ru.ru_inblock++; 808 rabp->b_flags |= B_ASYNC; 809 rabp->b_flags &= ~B_INVAL; 810 rabp->b_ioflags &= ~BIO_ERROR; 811 rabp->b_iocmd = BIO_READ; 812 if (rabp->b_rcred == NOCRED && cred != NOCRED) 813 rabp->b_rcred = crhold(cred); 814 vfs_busy_pages(rabp, 0); 815 BUF_KERNPROC(rabp); 816 rabp->b_iooffset = dbtob(rabp->b_blkno); 817 bstrategy(rabp); 818 } else { 819 brelse(rabp); 820 } 821 } 822} 823 824/* 825 * Entry point for bread() and breadn() via #defines in sys/buf.h. 826 * 827 * Get a buffer with the specified data. Look in the cache first. We 828 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 829 * is set, the buffer is valid and we do not have to do anything, see 830 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 831 */ 832int 833breadn_flags(struct vnode * vp, daddr_t blkno, int size, 834 daddr_t * rablkno, int *rabsize, int cnt, 835 struct ucred * cred, int flags, struct buf **bpp) 836{ 837 struct buf *bp; 838 int rv = 0, readwait = 0; 839 840 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 841 /* 842 * Can only return NULL if GB_LOCK_NOWAIT flag is specified. 843 */ 844 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags); 845 if (bp == NULL) 846 return (EBUSY); 847 848 /* if not found in cache, do some I/O */ 849 if ((bp->b_flags & B_CACHE) == 0) { 850 if (!TD_IS_IDLETHREAD(curthread)) 851 curthread->td_ru.ru_inblock++; 852 bp->b_iocmd = BIO_READ; 853 bp->b_flags &= ~B_INVAL; 854 bp->b_ioflags &= ~BIO_ERROR; 855 if (bp->b_rcred == NOCRED && cred != NOCRED) 856 bp->b_rcred = crhold(cred); 857 vfs_busy_pages(bp, 0); 858 bp->b_iooffset = dbtob(bp->b_blkno); 859 bstrategy(bp); 860 ++readwait; 861 } 862 863 breada(vp, rablkno, rabsize, cnt, cred); 864 865 if (readwait) { 866 rv = bufwait(bp); 867 } 868 return (rv); 869} 870 871/* 872 * Write, release buffer on completion. (Done by iodone 873 * if async). Do not bother writing anything if the buffer 874 * is invalid. 875 * 876 * Note that we set B_CACHE here, indicating that buffer is 877 * fully valid and thus cacheable. This is true even of NFS 878 * now so we set it generally. This could be set either here 879 * or in biodone() since the I/O is synchronous. We put it 880 * here. 881 */ 882int 883bufwrite(struct buf *bp) 884{ 885 int oldflags; 886 struct vnode *vp; 887 int vp_md; 888 889 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 890 if (bp->b_flags & B_INVAL) { 891 brelse(bp); 892 return (0); 893 } 894 895 if (bp->b_flags & B_BARRIER) 896 barrierwrites++; 897 898 oldflags = bp->b_flags; 899 900 BUF_ASSERT_HELD(bp); 901 902 if (bp->b_pin_count > 0) 903 bunpin_wait(bp); 904 905 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 906 ("FFS background buffer should not get here %p", bp)); 907 908 vp = bp->b_vp; 909 if (vp) 910 vp_md = vp->v_vflag & VV_MD; 911 else 912 vp_md = 0; 913 914 /* Mark the buffer clean */ 915 bundirty(bp); 916 917 bp->b_flags &= ~B_DONE; 918 bp->b_ioflags &= ~BIO_ERROR; 919 bp->b_flags |= B_CACHE; 920 bp->b_iocmd = BIO_WRITE; 921 922 bufobj_wref(bp->b_bufobj); 923 vfs_busy_pages(bp, 1); 924 925 /* 926 * Normal bwrites pipeline writes 927 */ 928 bp->b_runningbufspace = bp->b_bufsize; 929 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 930 931 if (!TD_IS_IDLETHREAD(curthread)) 932 curthread->td_ru.ru_oublock++; 933 if (oldflags & B_ASYNC) 934 BUF_KERNPROC(bp); 935 bp->b_iooffset = dbtob(bp->b_blkno); 936 bstrategy(bp); 937 938 if ((oldflags & B_ASYNC) == 0) { 939 int rtval = bufwait(bp); 940 brelse(bp); 941 return (rtval); 942 } else { 943 /* 944 * don't allow the async write to saturate the I/O 945 * system. We will not deadlock here because 946 * we are blocking waiting for I/O that is already in-progress 947 * to complete. We do not block here if it is the update 948 * or syncer daemon trying to clean up as that can lead 949 * to deadlock. 950 */ 951 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 952 waitrunningbufspace(); 953 } 954 955 return (0); 956} 957 958void 959bufbdflush(struct bufobj *bo, struct buf *bp) 960{ 961 struct buf *nbp; 962 963 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 964 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 965 altbufferflushes++; 966 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 967 BO_LOCK(bo); 968 /* 969 * Try to find a buffer to flush. 970 */ 971 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 972 if ((nbp->b_vflags & BV_BKGRDINPROG) || 973 BUF_LOCK(nbp, 974 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 975 continue; 976 if (bp == nbp) 977 panic("bdwrite: found ourselves"); 978 BO_UNLOCK(bo); 979 /* Don't countdeps with the bo lock held. */ 980 if (buf_countdeps(nbp, 0)) { 981 BO_LOCK(bo); 982 BUF_UNLOCK(nbp); 983 continue; 984 } 985 if (nbp->b_flags & B_CLUSTEROK) { 986 vfs_bio_awrite(nbp); 987 } else { 988 bremfree(nbp); 989 bawrite(nbp); 990 } 991 dirtybufferflushes++; 992 break; 993 } 994 if (nbp == NULL) 995 BO_UNLOCK(bo); 996 } 997} 998 999/* 1000 * Delayed write. (Buffer is marked dirty). Do not bother writing 1001 * anything if the buffer is marked invalid. 1002 * 1003 * Note that since the buffer must be completely valid, we can safely 1004 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1005 * biodone() in order to prevent getblk from writing the buffer 1006 * out synchronously. 1007 */ 1008void 1009bdwrite(struct buf *bp) 1010{ 1011 struct thread *td = curthread; 1012 struct vnode *vp; 1013 struct bufobj *bo; 1014 1015 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1016 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1017 KASSERT((bp->b_flags & B_BARRIER) == 0, 1018 ("Barrier request in delayed write %p", bp)); 1019 BUF_ASSERT_HELD(bp); 1020 1021 if (bp->b_flags & B_INVAL) { 1022 brelse(bp); 1023 return; 1024 } 1025 1026 /* 1027 * If we have too many dirty buffers, don't create any more. 1028 * If we are wildly over our limit, then force a complete 1029 * cleanup. Otherwise, just keep the situation from getting 1030 * out of control. Note that we have to avoid a recursive 1031 * disaster and not try to clean up after our own cleanup! 1032 */ 1033 vp = bp->b_vp; 1034 bo = bp->b_bufobj; 1035 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 1036 td->td_pflags |= TDP_INBDFLUSH; 1037 BO_BDFLUSH(bo, bp); 1038 td->td_pflags &= ~TDP_INBDFLUSH; 1039 } else 1040 recursiveflushes++; 1041 1042 bdirty(bp); 1043 /* 1044 * Set B_CACHE, indicating that the buffer is fully valid. This is 1045 * true even of NFS now. 1046 */ 1047 bp->b_flags |= B_CACHE; 1048 1049 /* 1050 * This bmap keeps the system from needing to do the bmap later, 1051 * perhaps when the system is attempting to do a sync. Since it 1052 * is likely that the indirect block -- or whatever other datastructure 1053 * that the filesystem needs is still in memory now, it is a good 1054 * thing to do this. Note also, that if the pageout daemon is 1055 * requesting a sync -- there might not be enough memory to do 1056 * the bmap then... So, this is important to do. 1057 */ 1058 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1059 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1060 } 1061 1062 /* 1063 * Set the *dirty* buffer range based upon the VM system dirty 1064 * pages. 1065 * 1066 * Mark the buffer pages as clean. We need to do this here to 1067 * satisfy the vnode_pager and the pageout daemon, so that it 1068 * thinks that the pages have been "cleaned". Note that since 1069 * the pages are in a delayed write buffer -- the VFS layer 1070 * "will" see that the pages get written out on the next sync, 1071 * or perhaps the cluster will be completed. 1072 */ 1073 vfs_clean_pages_dirty_buf(bp); 1074 bqrelse(bp); 1075 1076 /* 1077 * Wakeup the buffer flushing daemon if we have a lot of dirty 1078 * buffers (midpoint between our recovery point and our stall 1079 * point). 1080 */ 1081 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1082 1083 /* 1084 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1085 * due to the softdep code. 1086 */ 1087} 1088 1089/* 1090 * bdirty: 1091 * 1092 * Turn buffer into delayed write request. We must clear BIO_READ and 1093 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1094 * itself to properly update it in the dirty/clean lists. We mark it 1095 * B_DONE to ensure that any asynchronization of the buffer properly 1096 * clears B_DONE ( else a panic will occur later ). 1097 * 1098 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1099 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1100 * should only be called if the buffer is known-good. 1101 * 1102 * Since the buffer is not on a queue, we do not update the numfreebuffers 1103 * count. 1104 * 1105 * The buffer must be on QUEUE_NONE. 1106 */ 1107void 1108bdirty(struct buf *bp) 1109{ 1110 1111 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1112 bp, bp->b_vp, bp->b_flags); 1113 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1114 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1115 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1116 BUF_ASSERT_HELD(bp); 1117 bp->b_flags &= ~(B_RELBUF); 1118 bp->b_iocmd = BIO_WRITE; 1119 1120 if ((bp->b_flags & B_DELWRI) == 0) { 1121 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1122 reassignbuf(bp); 1123 atomic_add_int(&numdirtybuffers, 1); 1124 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1125 } 1126} 1127 1128/* 1129 * bundirty: 1130 * 1131 * Clear B_DELWRI for buffer. 1132 * 1133 * Since the buffer is not on a queue, we do not update the numfreebuffers 1134 * count. 1135 * 1136 * The buffer must be on QUEUE_NONE. 1137 */ 1138 1139void 1140bundirty(struct buf *bp) 1141{ 1142 1143 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1144 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1145 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1146 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1147 BUF_ASSERT_HELD(bp); 1148 1149 if (bp->b_flags & B_DELWRI) { 1150 bp->b_flags &= ~B_DELWRI; 1151 reassignbuf(bp); 1152 atomic_subtract_int(&numdirtybuffers, 1); 1153 numdirtywakeup(lodirtybuffers); 1154 } 1155 /* 1156 * Since it is now being written, we can clear its deferred write flag. 1157 */ 1158 bp->b_flags &= ~B_DEFERRED; 1159} 1160 1161/* 1162 * bawrite: 1163 * 1164 * Asynchronous write. Start output on a buffer, but do not wait for 1165 * it to complete. The buffer is released when the output completes. 1166 * 1167 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1168 * B_INVAL buffers. Not us. 1169 */ 1170void 1171bawrite(struct buf *bp) 1172{ 1173 1174 bp->b_flags |= B_ASYNC; 1175 (void) bwrite(bp); 1176} 1177 1178/* 1179 * babarrierwrite: 1180 * 1181 * Asynchronous barrier write. Start output on a buffer, but do not 1182 * wait for it to complete. Place a write barrier after this write so 1183 * that this buffer and all buffers written before it are committed to 1184 * the disk before any buffers written after this write are committed 1185 * to the disk. The buffer is released when the output completes. 1186 */ 1187void 1188babarrierwrite(struct buf *bp) 1189{ 1190 1191 bp->b_flags |= B_ASYNC | B_BARRIER; 1192 (void) bwrite(bp); 1193} 1194 1195/* 1196 * bbarrierwrite: 1197 * 1198 * Synchronous barrier write. Start output on a buffer and wait for 1199 * it to complete. Place a write barrier after this write so that 1200 * this buffer and all buffers written before it are committed to 1201 * the disk before any buffers written after this write are committed 1202 * to the disk. The buffer is released when the output completes. 1203 */ 1204int 1205bbarrierwrite(struct buf *bp) 1206{ 1207 1208 bp->b_flags |= B_BARRIER; 1209 return (bwrite(bp)); 1210} 1211 1212/* 1213 * bwillwrite: 1214 * 1215 * Called prior to the locking of any vnodes when we are expecting to 1216 * write. We do not want to starve the buffer cache with too many 1217 * dirty buffers so we block here. By blocking prior to the locking 1218 * of any vnodes we attempt to avoid the situation where a locked vnode 1219 * prevents the various system daemons from flushing related buffers. 1220 */ 1221 1222void 1223bwillwrite(void) 1224{ 1225 1226 if (numdirtybuffers >= hidirtybuffers) { 1227 mtx_lock(&nblock); 1228 while (numdirtybuffers >= hidirtybuffers) { 1229 bd_wakeup(1); 1230 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1231 msleep(&needsbuffer, &nblock, 1232 (PRIBIO + 4), "flswai", 0); 1233 } 1234 mtx_unlock(&nblock); 1235 } 1236} 1237 1238/* 1239 * Return true if we have too many dirty buffers. 1240 */ 1241int 1242buf_dirty_count_severe(void) 1243{ 1244 1245 return(numdirtybuffers >= hidirtybuffers); 1246} 1247 1248static __noinline int 1249buf_vm_page_count_severe(void) 1250{ 1251 1252 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1); 1253 1254 return vm_page_count_severe(); 1255} 1256 1257/* 1258 * brelse: 1259 * 1260 * Release a busy buffer and, if requested, free its resources. The 1261 * buffer will be stashed in the appropriate bufqueue[] allowing it 1262 * to be accessed later as a cache entity or reused for other purposes. 1263 */ 1264void 1265brelse(struct buf *bp) 1266{ 1267 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1268 bp, bp->b_vp, bp->b_flags); 1269 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1270 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1271 1272 if (BUF_LOCKRECURSED(bp)) { 1273 /* 1274 * Do not process, in particular, do not handle the 1275 * B_INVAL/B_RELBUF and do not release to free list. 1276 */ 1277 BUF_UNLOCK(bp); 1278 return; 1279 } 1280 1281 if (bp->b_flags & B_MANAGED) { 1282 bqrelse(bp); 1283 return; 1284 } 1285 1286 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1287 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1288 /* 1289 * Failed write, redirty. Must clear BIO_ERROR to prevent 1290 * pages from being scrapped. If the error is anything 1291 * other than an I/O error (EIO), assume that retrying 1292 * is futile. 1293 */ 1294 bp->b_ioflags &= ~BIO_ERROR; 1295 bdirty(bp); 1296 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1297 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1298 /* 1299 * Either a failed I/O or we were asked to free or not 1300 * cache the buffer. 1301 */ 1302 bp->b_flags |= B_INVAL; 1303 if (!LIST_EMPTY(&bp->b_dep)) 1304 buf_deallocate(bp); 1305 if (bp->b_flags & B_DELWRI) { 1306 atomic_subtract_int(&numdirtybuffers, 1); 1307 numdirtywakeup(lodirtybuffers); 1308 } 1309 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1310 if ((bp->b_flags & B_VMIO) == 0) { 1311 if (bp->b_bufsize) 1312 allocbuf(bp, 0); 1313 if (bp->b_vp) 1314 brelvp(bp); 1315 } 1316 } 1317 1318 /* 1319 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1320 * is called with B_DELWRI set, the underlying pages may wind up 1321 * getting freed causing a previous write (bdwrite()) to get 'lost' 1322 * because pages associated with a B_DELWRI bp are marked clean. 1323 * 1324 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1325 * if B_DELWRI is set. 1326 * 1327 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1328 * on pages to return pages to the VM page queues. 1329 */ 1330 if (bp->b_flags & B_DELWRI) 1331 bp->b_flags &= ~B_RELBUF; 1332 else if (buf_vm_page_count_severe()) { 1333 /* 1334 * The locking of the BO_LOCK is not necessary since 1335 * BKGRDINPROG cannot be set while we hold the buf 1336 * lock, it can only be cleared if it is already 1337 * pending. 1338 */ 1339 if (bp->b_vp) { 1340 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1341 bp->b_flags |= B_RELBUF; 1342 } else 1343 bp->b_flags |= B_RELBUF; 1344 } 1345 1346 /* 1347 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1348 * constituted, not even NFS buffers now. Two flags effect this. If 1349 * B_INVAL, the struct buf is invalidated but the VM object is kept 1350 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1351 * 1352 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1353 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1354 * buffer is also B_INVAL because it hits the re-dirtying code above. 1355 * 1356 * Normally we can do this whether a buffer is B_DELWRI or not. If 1357 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1358 * the commit state and we cannot afford to lose the buffer. If the 1359 * buffer has a background write in progress, we need to keep it 1360 * around to prevent it from being reconstituted and starting a second 1361 * background write. 1362 */ 1363 if ((bp->b_flags & B_VMIO) 1364 && !(bp->b_vp->v_mount != NULL && 1365 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1366 !vn_isdisk(bp->b_vp, NULL) && 1367 (bp->b_flags & B_DELWRI)) 1368 ) { 1369 1370 int i, j, resid; 1371 vm_page_t m; 1372 off_t foff; 1373 vm_pindex_t poff; 1374 vm_object_t obj; 1375 1376 obj = bp->b_bufobj->bo_object; 1377 1378 /* 1379 * Get the base offset and length of the buffer. Note that 1380 * in the VMIO case if the buffer block size is not 1381 * page-aligned then b_data pointer may not be page-aligned. 1382 * But our b_pages[] array *IS* page aligned. 1383 * 1384 * block sizes less then DEV_BSIZE (usually 512) are not 1385 * supported due to the page granularity bits (m->valid, 1386 * m->dirty, etc...). 1387 * 1388 * See man buf(9) for more information 1389 */ 1390 resid = bp->b_bufsize; 1391 foff = bp->b_offset; 1392 VM_OBJECT_WLOCK(obj); 1393 for (i = 0; i < bp->b_npages; i++) { 1394 int had_bogus = 0; 1395 1396 m = bp->b_pages[i]; 1397 1398 /* 1399 * If we hit a bogus page, fixup *all* the bogus pages 1400 * now. 1401 */ 1402 if (m == bogus_page) { 1403 poff = OFF_TO_IDX(bp->b_offset); 1404 had_bogus = 1; 1405 1406 for (j = i; j < bp->b_npages; j++) { 1407 vm_page_t mtmp; 1408 mtmp = bp->b_pages[j]; 1409 if (mtmp == bogus_page) { 1410 mtmp = vm_page_lookup(obj, poff + j); 1411 if (!mtmp) { 1412 panic("brelse: page missing\n"); 1413 } 1414 bp->b_pages[j] = mtmp; 1415 } 1416 } 1417 1418 if ((bp->b_flags & B_INVAL) == 0) { 1419 pmap_qenter( 1420 trunc_page((vm_offset_t)bp->b_data), 1421 bp->b_pages, bp->b_npages); 1422 } 1423 m = bp->b_pages[i]; 1424 } 1425 if ((bp->b_flags & B_NOCACHE) || 1426 (bp->b_ioflags & BIO_ERROR && 1427 bp->b_iocmd == BIO_READ)) { 1428 int poffset = foff & PAGE_MASK; 1429 int presid = resid > (PAGE_SIZE - poffset) ? 1430 (PAGE_SIZE - poffset) : resid; 1431 1432 KASSERT(presid >= 0, ("brelse: extra page")); 1433 vm_page_set_invalid(m, poffset, presid); 1434 if (had_bogus) 1435 printf("avoided corruption bug in bogus_page/brelse code\n"); 1436 } 1437 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1438 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1439 } 1440 VM_OBJECT_WUNLOCK(obj); 1441 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1442 vfs_vmio_release(bp); 1443 1444 } else if (bp->b_flags & B_VMIO) { 1445 1446 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1447 vfs_vmio_release(bp); 1448 } 1449 1450 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1451 if (bp->b_bufsize != 0) 1452 allocbuf(bp, 0); 1453 if (bp->b_vp != NULL) 1454 brelvp(bp); 1455 } 1456 1457 /* enqueue */ 1458 mtx_lock(&bqlock); 1459 /* Handle delayed bremfree() processing. */ 1460 if (bp->b_flags & B_REMFREE) { 1461 struct bufobj *bo; 1462 1463 bo = bp->b_bufobj; 1464 if (bo != NULL) 1465 BO_LOCK(bo); 1466 bremfreel(bp); 1467 if (bo != NULL) 1468 BO_UNLOCK(bo); 1469 } 1470 if (bp->b_qindex != QUEUE_NONE) 1471 panic("brelse: free buffer onto another queue???"); 1472 1473 /* 1474 * If the buffer has junk contents signal it and eventually 1475 * clean up B_DELWRI and diassociate the vnode so that gbincore() 1476 * doesn't find it. 1477 */ 1478 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1479 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1480 bp->b_flags |= B_INVAL; 1481 if (bp->b_flags & B_INVAL) { 1482 if (bp->b_flags & B_DELWRI) 1483 bundirty(bp); 1484 if (bp->b_vp) 1485 brelvp(bp); 1486 } 1487 1488 /* buffers with no memory */ 1489 if (bp->b_bufsize == 0) { 1490 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1491 if (bp->b_vflags & BV_BKGRDINPROG) 1492 panic("losing buffer 1"); 1493 if (bp->b_kvasize) { 1494 bp->b_qindex = QUEUE_EMPTYKVA; 1495 } else { 1496 bp->b_qindex = QUEUE_EMPTY; 1497 } 1498 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1499 /* buffers with junk contents */ 1500 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1501 (bp->b_ioflags & BIO_ERROR)) { 1502 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1503 if (bp->b_vflags & BV_BKGRDINPROG) 1504 panic("losing buffer 2"); 1505 bp->b_qindex = QUEUE_CLEAN; 1506 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1507 /* remaining buffers */ 1508 } else { 1509 if (bp->b_flags & B_DELWRI) 1510 bp->b_qindex = QUEUE_DIRTY; 1511 else 1512 bp->b_qindex = QUEUE_CLEAN; 1513 if (bp->b_flags & B_AGE) 1514 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1515 else 1516 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1517 } 1518 mtx_unlock(&bqlock); 1519 1520 /* 1521 * Fixup numfreebuffers count. The bp is on an appropriate queue 1522 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1523 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1524 * if B_INVAL is set ). 1525 */ 1526 1527 if (!(bp->b_flags & B_DELWRI)) { 1528 struct bufobj *bo; 1529 1530 bo = bp->b_bufobj; 1531 if (bo != NULL) 1532 BO_LOCK(bo); 1533 bufcountwakeup(bp); 1534 if (bo != NULL) 1535 BO_UNLOCK(bo); 1536 } 1537 1538 /* 1539 * Something we can maybe free or reuse 1540 */ 1541 if (bp->b_bufsize || bp->b_kvasize) 1542 bufspacewakeup(); 1543 1544 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1545 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1546 panic("brelse: not dirty"); 1547 /* unlock */ 1548 BUF_UNLOCK(bp); 1549} 1550 1551/* 1552 * Release a buffer back to the appropriate queue but do not try to free 1553 * it. The buffer is expected to be used again soon. 1554 * 1555 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1556 * biodone() to requeue an async I/O on completion. It is also used when 1557 * known good buffers need to be requeued but we think we may need the data 1558 * again soon. 1559 * 1560 * XXX we should be able to leave the B_RELBUF hint set on completion. 1561 */ 1562void 1563bqrelse(struct buf *bp) 1564{ 1565 struct bufobj *bo; 1566 1567 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1568 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1569 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1570 1571 if (BUF_LOCKRECURSED(bp)) { 1572 /* do not release to free list */ 1573 BUF_UNLOCK(bp); 1574 return; 1575 } 1576 1577 bo = bp->b_bufobj; 1578 if (bp->b_flags & B_MANAGED) { 1579 if (bp->b_flags & B_REMFREE) { 1580 mtx_lock(&bqlock); 1581 if (bo != NULL) 1582 BO_LOCK(bo); 1583 bremfreel(bp); 1584 if (bo != NULL) 1585 BO_UNLOCK(bo); 1586 mtx_unlock(&bqlock); 1587 } 1588 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1589 BUF_UNLOCK(bp); 1590 return; 1591 } 1592 1593 mtx_lock(&bqlock); 1594 /* Handle delayed bremfree() processing. */ 1595 if (bp->b_flags & B_REMFREE) { 1596 if (bo != NULL) 1597 BO_LOCK(bo); 1598 bremfreel(bp); 1599 if (bo != NULL) 1600 BO_UNLOCK(bo); 1601 } 1602 if (bp->b_qindex != QUEUE_NONE) 1603 panic("bqrelse: free buffer onto another queue???"); 1604 /* buffers with stale but valid contents */ 1605 if (bp->b_flags & B_DELWRI) { 1606 bp->b_qindex = QUEUE_DIRTY; 1607 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1608 } else { 1609 /* 1610 * The locking of the BO_LOCK for checking of the 1611 * BV_BKGRDINPROG is not necessary since the 1612 * BV_BKGRDINPROG cannot be set while we hold the buf 1613 * lock, it can only be cleared if it is already 1614 * pending. 1615 */ 1616 if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) { 1617 bp->b_qindex = QUEUE_CLEAN; 1618 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1619 b_freelist); 1620 } else { 1621 /* 1622 * We are too low on memory, we have to try to free 1623 * the buffer (most importantly: the wired pages 1624 * making up its backing store) *now*. 1625 */ 1626 mtx_unlock(&bqlock); 1627 brelse(bp); 1628 return; 1629 } 1630 } 1631 mtx_unlock(&bqlock); 1632 1633 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) { 1634 if (bo != NULL) 1635 BO_LOCK(bo); 1636 bufcountwakeup(bp); 1637 if (bo != NULL) 1638 BO_UNLOCK(bo); 1639 } 1640 1641 /* 1642 * Something we can maybe free or reuse. 1643 */ 1644 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1645 bufspacewakeup(); 1646 1647 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1648 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1649 panic("bqrelse: not dirty"); 1650 /* unlock */ 1651 BUF_UNLOCK(bp); 1652} 1653 1654/* Give pages used by the bp back to the VM system (where possible) */ 1655static void 1656vfs_vmio_release(struct buf *bp) 1657{ 1658 int i; 1659 vm_page_t m; 1660 1661 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 1662 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 1663 for (i = 0; i < bp->b_npages; i++) { 1664 m = bp->b_pages[i]; 1665 bp->b_pages[i] = NULL; 1666 /* 1667 * In order to keep page LRU ordering consistent, put 1668 * everything on the inactive queue. 1669 */ 1670 vm_page_lock(m); 1671 vm_page_unwire(m, 0); 1672 /* 1673 * We don't mess with busy pages, it is 1674 * the responsibility of the process that 1675 * busied the pages to deal with them. 1676 */ 1677 if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 && 1678 m->wire_count == 0) { 1679 /* 1680 * Might as well free the page if we can and it has 1681 * no valid data. We also free the page if the 1682 * buffer was used for direct I/O 1683 */ 1684 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) { 1685 vm_page_free(m); 1686 } else if (bp->b_flags & B_DIRECT) { 1687 vm_page_try_to_free(m); 1688 } else if (buf_vm_page_count_severe()) { 1689 vm_page_try_to_cache(m); 1690 } 1691 } 1692 vm_page_unlock(m); 1693 } 1694 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 1695 1696 if (bp->b_bufsize) { 1697 bufspacewakeup(); 1698 bp->b_bufsize = 0; 1699 } 1700 bp->b_npages = 0; 1701 bp->b_flags &= ~B_VMIO; 1702 if (bp->b_vp) 1703 brelvp(bp); 1704} 1705 1706/* 1707 * Check to see if a block at a particular lbn is available for a clustered 1708 * write. 1709 */ 1710static int 1711vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1712{ 1713 struct buf *bpa; 1714 int match; 1715 1716 match = 0; 1717 1718 /* If the buf isn't in core skip it */ 1719 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1720 return (0); 1721 1722 /* If the buf is busy we don't want to wait for it */ 1723 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1724 return (0); 1725 1726 /* Only cluster with valid clusterable delayed write buffers */ 1727 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1728 (B_DELWRI | B_CLUSTEROK)) 1729 goto done; 1730 1731 if (bpa->b_bufsize != size) 1732 goto done; 1733 1734 /* 1735 * Check to see if it is in the expected place on disk and that the 1736 * block has been mapped. 1737 */ 1738 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1739 match = 1; 1740done: 1741 BUF_UNLOCK(bpa); 1742 return (match); 1743} 1744 1745/* 1746 * vfs_bio_awrite: 1747 * 1748 * Implement clustered async writes for clearing out B_DELWRI buffers. 1749 * This is much better then the old way of writing only one buffer at 1750 * a time. Note that we may not be presented with the buffers in the 1751 * correct order, so we search for the cluster in both directions. 1752 */ 1753int 1754vfs_bio_awrite(struct buf *bp) 1755{ 1756 struct bufobj *bo; 1757 int i; 1758 int j; 1759 daddr_t lblkno = bp->b_lblkno; 1760 struct vnode *vp = bp->b_vp; 1761 int ncl; 1762 int nwritten; 1763 int size; 1764 int maxcl; 1765 1766 bo = &vp->v_bufobj; 1767 /* 1768 * right now we support clustered writing only to regular files. If 1769 * we find a clusterable block we could be in the middle of a cluster 1770 * rather then at the beginning. 1771 */ 1772 if ((vp->v_type == VREG) && 1773 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1774 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1775 1776 size = vp->v_mount->mnt_stat.f_iosize; 1777 maxcl = MAXPHYS / size; 1778 1779 BO_LOCK(bo); 1780 for (i = 1; i < maxcl; i++) 1781 if (vfs_bio_clcheck(vp, size, lblkno + i, 1782 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1783 break; 1784 1785 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1786 if (vfs_bio_clcheck(vp, size, lblkno - j, 1787 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1788 break; 1789 BO_UNLOCK(bo); 1790 --j; 1791 ncl = i + j; 1792 /* 1793 * this is a possible cluster write 1794 */ 1795 if (ncl != 1) { 1796 BUF_UNLOCK(bp); 1797 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 1798 0); 1799 return (nwritten); 1800 } 1801 } 1802 bremfree(bp); 1803 bp->b_flags |= B_ASYNC; 1804 /* 1805 * default (old) behavior, writing out only one block 1806 * 1807 * XXX returns b_bufsize instead of b_bcount for nwritten? 1808 */ 1809 nwritten = bp->b_bufsize; 1810 (void) bwrite(bp); 1811 1812 return nwritten; 1813} 1814 1815/* 1816 * getnewbuf: 1817 * 1818 * Find and initialize a new buffer header, freeing up existing buffers 1819 * in the bufqueues as necessary. The new buffer is returned locked. 1820 * 1821 * Important: B_INVAL is not set. If the caller wishes to throw the 1822 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1823 * 1824 * We block if: 1825 * We have insufficient buffer headers 1826 * We have insufficient buffer space 1827 * buffer_map is too fragmented ( space reservation fails ) 1828 * If we have to flush dirty buffers ( but we try to avoid this ) 1829 * 1830 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1831 * Instead we ask the buf daemon to do it for us. We attempt to 1832 * avoid piecemeal wakeups of the pageout daemon. 1833 */ 1834 1835static struct buf * 1836getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 1837 int gbflags) 1838{ 1839 struct thread *td; 1840 struct buf *bp; 1841 struct buf *nbp; 1842 int defrag = 0; 1843 int nqindex; 1844 static int flushingbufs; 1845 1846 td = curthread; 1847 /* 1848 * We can't afford to block since we might be holding a vnode lock, 1849 * which may prevent system daemons from running. We deal with 1850 * low-memory situations by proactively returning memory and running 1851 * async I/O rather then sync I/O. 1852 */ 1853 atomic_add_int(&getnewbufcalls, 1); 1854 atomic_subtract_int(&getnewbufrestarts, 1); 1855restart: 1856 atomic_add_int(&getnewbufrestarts, 1); 1857 1858 /* 1859 * Setup for scan. If we do not have enough free buffers, 1860 * we setup a degenerate case that immediately fails. Note 1861 * that if we are specially marked process, we are allowed to 1862 * dip into our reserves. 1863 * 1864 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1865 * 1866 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1867 * However, there are a number of cases (defragging, reusing, ...) 1868 * where we cannot backup. 1869 */ 1870 mtx_lock(&bqlock); 1871 nqindex = QUEUE_EMPTYKVA; 1872 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1873 1874 if (nbp == NULL) { 1875 /* 1876 * If no EMPTYKVA buffers and we are either 1877 * defragging or reusing, locate a CLEAN buffer 1878 * to free or reuse. If bufspace useage is low 1879 * skip this step so we can allocate a new buffer. 1880 */ 1881 if (defrag || bufspace >= lobufspace) { 1882 nqindex = QUEUE_CLEAN; 1883 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1884 } 1885 1886 /* 1887 * If we could not find or were not allowed to reuse a 1888 * CLEAN buffer, check to see if it is ok to use an EMPTY 1889 * buffer. We can only use an EMPTY buffer if allocating 1890 * its KVA would not otherwise run us out of buffer space. 1891 */ 1892 if (nbp == NULL && defrag == 0 && 1893 bufspace + maxsize < hibufspace) { 1894 nqindex = QUEUE_EMPTY; 1895 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1896 } 1897 } 1898 1899 /* 1900 * Run scan, possibly freeing data and/or kva mappings on the fly 1901 * depending. 1902 */ 1903 1904 while ((bp = nbp) != NULL) { 1905 int qindex = nqindex; 1906 1907 /* 1908 * Calculate next bp ( we can only use it if we do not block 1909 * or do other fancy things ). 1910 */ 1911 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1912 switch(qindex) { 1913 case QUEUE_EMPTY: 1914 nqindex = QUEUE_EMPTYKVA; 1915 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1916 break; 1917 /* FALLTHROUGH */ 1918 case QUEUE_EMPTYKVA: 1919 nqindex = QUEUE_CLEAN; 1920 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1921 break; 1922 /* FALLTHROUGH */ 1923 case QUEUE_CLEAN: 1924 /* 1925 * nbp is NULL. 1926 */ 1927 break; 1928 } 1929 } 1930 /* 1931 * If we are defragging then we need a buffer with 1932 * b_kvasize != 0. XXX this situation should no longer 1933 * occur, if defrag is non-zero the buffer's b_kvasize 1934 * should also be non-zero at this point. XXX 1935 */ 1936 if (defrag && bp->b_kvasize == 0) { 1937 printf("Warning: defrag empty buffer %p\n", bp); 1938 continue; 1939 } 1940 1941 /* 1942 * Start freeing the bp. This is somewhat involved. nbp 1943 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1944 */ 1945 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1946 continue; 1947 if (bp->b_vp) { 1948 BO_LOCK(bp->b_bufobj); 1949 if (bp->b_vflags & BV_BKGRDINPROG) { 1950 BO_UNLOCK(bp->b_bufobj); 1951 BUF_UNLOCK(bp); 1952 continue; 1953 } 1954 BO_UNLOCK(bp->b_bufobj); 1955 } 1956 CTR6(KTR_BUF, 1957 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1958 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1959 bp->b_kvasize, bp->b_bufsize, qindex); 1960 1961 /* 1962 * Sanity Checks 1963 */ 1964 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1965 1966 /* 1967 * Note: we no longer distinguish between VMIO and non-VMIO 1968 * buffers. 1969 */ 1970 1971 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1972 1973 if (bp->b_bufobj != NULL) 1974 BO_LOCK(bp->b_bufobj); 1975 bremfreel(bp); 1976 if (bp->b_bufobj != NULL) 1977 BO_UNLOCK(bp->b_bufobj); 1978 mtx_unlock(&bqlock); 1979 1980 if (qindex == QUEUE_CLEAN) { 1981 if (bp->b_flags & B_VMIO) { 1982 bp->b_flags &= ~B_ASYNC; 1983 vfs_vmio_release(bp); 1984 } 1985 if (bp->b_vp) 1986 brelvp(bp); 1987 } 1988 1989 /* 1990 * NOTE: nbp is now entirely invalid. We can only restart 1991 * the scan from this point on. 1992 * 1993 * Get the rest of the buffer freed up. b_kva* is still 1994 * valid after this operation. 1995 */ 1996 1997 if (bp->b_rcred != NOCRED) { 1998 crfree(bp->b_rcred); 1999 bp->b_rcred = NOCRED; 2000 } 2001 if (bp->b_wcred != NOCRED) { 2002 crfree(bp->b_wcred); 2003 bp->b_wcred = NOCRED; 2004 } 2005 if (!LIST_EMPTY(&bp->b_dep)) 2006 buf_deallocate(bp); 2007 if (bp->b_vflags & BV_BKGRDINPROG) 2008 panic("losing buffer 3"); 2009 KASSERT(bp->b_vp == NULL, 2010 ("bp: %p still has vnode %p. qindex: %d", 2011 bp, bp->b_vp, qindex)); 2012 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 2013 ("bp: %p still on a buffer list. xflags %X", 2014 bp, bp->b_xflags)); 2015 2016 if (bp->b_bufsize) 2017 allocbuf(bp, 0); 2018 2019 bp->b_flags = 0; 2020 bp->b_ioflags = 0; 2021 bp->b_xflags = 0; 2022 KASSERT((bp->b_vflags & BV_INFREECNT) == 0, 2023 ("buf %p still counted as free?", bp)); 2024 bp->b_vflags = 0; 2025 bp->b_vp = NULL; 2026 bp->b_blkno = bp->b_lblkno = 0; 2027 bp->b_offset = NOOFFSET; 2028 bp->b_iodone = 0; 2029 bp->b_error = 0; 2030 bp->b_resid = 0; 2031 bp->b_bcount = 0; 2032 bp->b_npages = 0; 2033 bp->b_dirtyoff = bp->b_dirtyend = 0; 2034 bp->b_bufobj = NULL; 2035 bp->b_pin_count = 0; 2036 bp->b_fsprivate1 = NULL; 2037 bp->b_fsprivate2 = NULL; 2038 bp->b_fsprivate3 = NULL; 2039 2040 LIST_INIT(&bp->b_dep); 2041 2042 /* 2043 * If we are defragging then free the buffer. 2044 */ 2045 if (defrag) { 2046 bp->b_flags |= B_INVAL; 2047 bfreekva(bp); 2048 brelse(bp); 2049 defrag = 0; 2050 goto restart; 2051 } 2052 2053 /* 2054 * Notify any waiters for the buffer lock about 2055 * identity change by freeing the buffer. 2056 */ 2057 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 2058 bp->b_flags |= B_INVAL; 2059 bfreekva(bp); 2060 brelse(bp); 2061 goto restart; 2062 } 2063 2064 /* 2065 * If we are overcomitted then recover the buffer and its 2066 * KVM space. This occurs in rare situations when multiple 2067 * processes are blocked in getnewbuf() or allocbuf(). 2068 */ 2069 if (bufspace >= hibufspace) 2070 flushingbufs = 1; 2071 if (flushingbufs && bp->b_kvasize != 0) { 2072 bp->b_flags |= B_INVAL; 2073 bfreekva(bp); 2074 brelse(bp); 2075 goto restart; 2076 } 2077 if (bufspace < lobufspace) 2078 flushingbufs = 0; 2079 break; 2080 } 2081 2082 /* 2083 * If we exhausted our list, sleep as appropriate. We may have to 2084 * wakeup various daemons and write out some dirty buffers. 2085 * 2086 * Generally we are sleeping due to insufficient buffer space. 2087 */ 2088 2089 if (bp == NULL) { 2090 int flags, norunbuf; 2091 char *waitmsg; 2092 int fl; 2093 2094 if (defrag) { 2095 flags = VFS_BIO_NEED_BUFSPACE; 2096 waitmsg = "nbufkv"; 2097 } else if (bufspace >= hibufspace) { 2098 waitmsg = "nbufbs"; 2099 flags = VFS_BIO_NEED_BUFSPACE; 2100 } else { 2101 waitmsg = "newbuf"; 2102 flags = VFS_BIO_NEED_ANY; 2103 } 2104 mtx_lock(&nblock); 2105 needsbuffer |= flags; 2106 mtx_unlock(&nblock); 2107 mtx_unlock(&bqlock); 2108 2109 bd_speedup(); /* heeeelp */ 2110 if (gbflags & GB_NOWAIT_BD) 2111 return (NULL); 2112 2113 mtx_lock(&nblock); 2114 while (needsbuffer & flags) { 2115 if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) { 2116 mtx_unlock(&nblock); 2117 /* 2118 * getblk() is called with a vnode 2119 * locked, and some majority of the 2120 * dirty buffers may as well belong to 2121 * the vnode. Flushing the buffers 2122 * there would make a progress that 2123 * cannot be achieved by the 2124 * buf_daemon, that cannot lock the 2125 * vnode. 2126 */ 2127 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2128 (td->td_pflags & TDP_NORUNNINGBUF); 2129 /* play bufdaemon */ 2130 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2131 fl = buf_do_flush(vp); 2132 td->td_pflags &= norunbuf; 2133 mtx_lock(&nblock); 2134 if (fl != 0) 2135 continue; 2136 if ((needsbuffer & flags) == 0) 2137 break; 2138 } 2139 if (msleep(&needsbuffer, &nblock, 2140 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2141 mtx_unlock(&nblock); 2142 return (NULL); 2143 } 2144 } 2145 mtx_unlock(&nblock); 2146 } else { 2147 /* 2148 * We finally have a valid bp. We aren't quite out of the 2149 * woods, we still have to reserve kva space. In order 2150 * to keep fragmentation sane we only allocate kva in 2151 * BKVASIZE chunks. 2152 */ 2153 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2154 2155 if (maxsize != bp->b_kvasize) { 2156 vm_offset_t addr = 0; 2157 int rv; 2158 2159 bfreekva(bp); 2160 2161 vm_map_lock(buffer_map); 2162 if (vm_map_findspace(buffer_map, 2163 vm_map_min(buffer_map), maxsize, &addr)) { 2164 /* 2165 * Buffer map is too fragmented. 2166 * We must defragment the map. 2167 */ 2168 atomic_add_int(&bufdefragcnt, 1); 2169 vm_map_unlock(buffer_map); 2170 defrag = 1; 2171 bp->b_flags |= B_INVAL; 2172 brelse(bp); 2173 goto restart; 2174 } 2175 rv = vm_map_insert(buffer_map, NULL, 0, addr, 2176 addr + maxsize, VM_PROT_ALL, VM_PROT_ALL, 2177 MAP_NOFAULT); 2178 KASSERT(rv == KERN_SUCCESS, 2179 ("vm_map_insert(buffer_map) rv %d", rv)); 2180 vm_map_unlock(buffer_map); 2181 bp->b_kvabase = (caddr_t)addr; 2182 bp->b_kvasize = maxsize; 2183 atomic_add_long(&bufspace, bp->b_kvasize); 2184 atomic_add_int(&bufreusecnt, 1); 2185 } 2186 bp->b_saveaddr = bp->b_kvabase; 2187 bp->b_data = bp->b_saveaddr; 2188 } 2189 return (bp); 2190} 2191 2192/* 2193 * buf_daemon: 2194 * 2195 * buffer flushing daemon. Buffers are normally flushed by the 2196 * update daemon but if it cannot keep up this process starts to 2197 * take the load in an attempt to prevent getnewbuf() from blocking. 2198 */ 2199 2200static struct kproc_desc buf_kp = { 2201 "bufdaemon", 2202 buf_daemon, 2203 &bufdaemonproc 2204}; 2205SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2206 2207static int 2208buf_do_flush(struct vnode *vp) 2209{ 2210 int flushed; 2211 2212 flushed = flushbufqueues(vp, QUEUE_DIRTY, 0); 2213 if (flushed == 0) { 2214 /* 2215 * Could not find any buffers without rollback 2216 * dependencies, so just write the first one 2217 * in the hopes of eventually making progress. 2218 */ 2219 flushbufqueues(vp, QUEUE_DIRTY, 1); 2220 } 2221 return (flushed); 2222} 2223 2224static void 2225buf_daemon() 2226{ 2227 int lodirtysave; 2228 2229 /* 2230 * This process needs to be suspended prior to shutdown sync. 2231 */ 2232 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2233 SHUTDOWN_PRI_LAST); 2234 2235 /* 2236 * This process is allowed to take the buffer cache to the limit 2237 */ 2238 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2239 mtx_lock(&bdlock); 2240 for (;;) { 2241 bd_request = 0; 2242 mtx_unlock(&bdlock); 2243 2244 kproc_suspend_check(bufdaemonproc); 2245 lodirtysave = lodirtybuffers; 2246 if (bd_speedupreq) { 2247 lodirtybuffers = numdirtybuffers / 2; 2248 bd_speedupreq = 0; 2249 } 2250 /* 2251 * Do the flush. Limit the amount of in-transit I/O we 2252 * allow to build up, otherwise we would completely saturate 2253 * the I/O system. Wakeup any waiting processes before we 2254 * normally would so they can run in parallel with our drain. 2255 */ 2256 while (numdirtybuffers > lodirtybuffers) { 2257 if (buf_do_flush(NULL) == 0) 2258 break; 2259 kern_yield(PRI_USER); 2260 } 2261 lodirtybuffers = lodirtysave; 2262 2263 /* 2264 * Only clear bd_request if we have reached our low water 2265 * mark. The buf_daemon normally waits 1 second and 2266 * then incrementally flushes any dirty buffers that have 2267 * built up, within reason. 2268 * 2269 * If we were unable to hit our low water mark and couldn't 2270 * find any flushable buffers, we sleep half a second. 2271 * Otherwise we loop immediately. 2272 */ 2273 mtx_lock(&bdlock); 2274 if (numdirtybuffers <= lodirtybuffers) { 2275 /* 2276 * We reached our low water mark, reset the 2277 * request and sleep until we are needed again. 2278 * The sleep is just so the suspend code works. 2279 */ 2280 bd_request = 0; 2281 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2282 } else { 2283 /* 2284 * We couldn't find any flushable dirty buffers but 2285 * still have too many dirty buffers, we 2286 * have to sleep and try again. (rare) 2287 */ 2288 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2289 } 2290 } 2291} 2292 2293/* 2294 * flushbufqueues: 2295 * 2296 * Try to flush a buffer in the dirty queue. We must be careful to 2297 * free up B_INVAL buffers instead of write them, which NFS is 2298 * particularly sensitive to. 2299 */ 2300static int flushwithdeps = 0; 2301SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2302 0, "Number of buffers flushed with dependecies that require rollbacks"); 2303 2304static int 2305flushbufqueues(struct vnode *lvp, int queue, int flushdeps) 2306{ 2307 struct buf *sentinel; 2308 struct vnode *vp; 2309 struct mount *mp; 2310 struct buf *bp; 2311 int hasdeps; 2312 int flushed; 2313 int target; 2314 2315 if (lvp == NULL) { 2316 target = numdirtybuffers - lodirtybuffers; 2317 if (flushdeps && target > 2) 2318 target /= 2; 2319 } else 2320 target = flushbufqtarget; 2321 flushed = 0; 2322 bp = NULL; 2323 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2324 sentinel->b_qindex = QUEUE_SENTINEL; 2325 mtx_lock(&bqlock); 2326 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2327 while (flushed != target) { 2328 bp = TAILQ_NEXT(sentinel, b_freelist); 2329 if (bp != NULL) { 2330 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2331 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2332 b_freelist); 2333 } else 2334 break; 2335 /* 2336 * Skip sentinels inserted by other invocations of the 2337 * flushbufqueues(), taking care to not reorder them. 2338 */ 2339 if (bp->b_qindex == QUEUE_SENTINEL) 2340 continue; 2341 /* 2342 * Only flush the buffers that belong to the 2343 * vnode locked by the curthread. 2344 */ 2345 if (lvp != NULL && bp->b_vp != lvp) 2346 continue; 2347 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2348 continue; 2349 if (bp->b_pin_count > 0) { 2350 BUF_UNLOCK(bp); 2351 continue; 2352 } 2353 BO_LOCK(bp->b_bufobj); 2354 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2355 (bp->b_flags & B_DELWRI) == 0) { 2356 BO_UNLOCK(bp->b_bufobj); 2357 BUF_UNLOCK(bp); 2358 continue; 2359 } 2360 BO_UNLOCK(bp->b_bufobj); 2361 if (bp->b_flags & B_INVAL) { 2362 bremfreel(bp); 2363 mtx_unlock(&bqlock); 2364 brelse(bp); 2365 flushed++; 2366 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2367 mtx_lock(&bqlock); 2368 continue; 2369 } 2370 2371 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2372 if (flushdeps == 0) { 2373 BUF_UNLOCK(bp); 2374 continue; 2375 } 2376 hasdeps = 1; 2377 } else 2378 hasdeps = 0; 2379 /* 2380 * We must hold the lock on a vnode before writing 2381 * one of its buffers. Otherwise we may confuse, or 2382 * in the case of a snapshot vnode, deadlock the 2383 * system. 2384 * 2385 * The lock order here is the reverse of the normal 2386 * of vnode followed by buf lock. This is ok because 2387 * the NOWAIT will prevent deadlock. 2388 */ 2389 vp = bp->b_vp; 2390 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2391 BUF_UNLOCK(bp); 2392 continue; 2393 } 2394 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) { 2395 mtx_unlock(&bqlock); 2396 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2397 bp, bp->b_vp, bp->b_flags); 2398 if (curproc == bufdaemonproc) 2399 vfs_bio_awrite(bp); 2400 else { 2401 bremfree(bp); 2402 bwrite(bp); 2403 notbufdflashes++; 2404 } 2405 vn_finished_write(mp); 2406 VOP_UNLOCK(vp, 0); 2407 flushwithdeps += hasdeps; 2408 flushed++; 2409 2410 /* 2411 * Sleeping on runningbufspace while holding 2412 * vnode lock leads to deadlock. 2413 */ 2414 if (curproc == bufdaemonproc) 2415 waitrunningbufspace(); 2416 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2417 mtx_lock(&bqlock); 2418 continue; 2419 } 2420 vn_finished_write(mp); 2421 BUF_UNLOCK(bp); 2422 } 2423 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2424 mtx_unlock(&bqlock); 2425 free(sentinel, M_TEMP); 2426 return (flushed); 2427} 2428 2429/* 2430 * Check to see if a block is currently memory resident. 2431 */ 2432struct buf * 2433incore(struct bufobj *bo, daddr_t blkno) 2434{ 2435 struct buf *bp; 2436 2437 BO_LOCK(bo); 2438 bp = gbincore(bo, blkno); 2439 BO_UNLOCK(bo); 2440 return (bp); 2441} 2442 2443/* 2444 * Returns true if no I/O is needed to access the 2445 * associated VM object. This is like incore except 2446 * it also hunts around in the VM system for the data. 2447 */ 2448 2449static int 2450inmem(struct vnode * vp, daddr_t blkno) 2451{ 2452 vm_object_t obj; 2453 vm_offset_t toff, tinc, size; 2454 vm_page_t m; 2455 vm_ooffset_t off; 2456 2457 ASSERT_VOP_LOCKED(vp, "inmem"); 2458 2459 if (incore(&vp->v_bufobj, blkno)) 2460 return 1; 2461 if (vp->v_mount == NULL) 2462 return 0; 2463 obj = vp->v_object; 2464 if (obj == NULL) 2465 return (0); 2466 2467 size = PAGE_SIZE; 2468 if (size > vp->v_mount->mnt_stat.f_iosize) 2469 size = vp->v_mount->mnt_stat.f_iosize; 2470 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2471 2472 VM_OBJECT_WLOCK(obj); 2473 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2474 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2475 if (!m) 2476 goto notinmem; 2477 tinc = size; 2478 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2479 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2480 if (vm_page_is_valid(m, 2481 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2482 goto notinmem; 2483 } 2484 VM_OBJECT_WUNLOCK(obj); 2485 return 1; 2486 2487notinmem: 2488 VM_OBJECT_WUNLOCK(obj); 2489 return (0); 2490} 2491 2492/* 2493 * Set the dirty range for a buffer based on the status of the dirty 2494 * bits in the pages comprising the buffer. The range is limited 2495 * to the size of the buffer. 2496 * 2497 * Tell the VM system that the pages associated with this buffer 2498 * are clean. This is used for delayed writes where the data is 2499 * going to go to disk eventually without additional VM intevention. 2500 * 2501 * Note that while we only really need to clean through to b_bcount, we 2502 * just go ahead and clean through to b_bufsize. 2503 */ 2504static void 2505vfs_clean_pages_dirty_buf(struct buf *bp) 2506{ 2507 vm_ooffset_t foff, noff, eoff; 2508 vm_page_t m; 2509 int i; 2510 2511 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 2512 return; 2513 2514 foff = bp->b_offset; 2515 KASSERT(bp->b_offset != NOOFFSET, 2516 ("vfs_clean_pages_dirty_buf: no buffer offset")); 2517 2518 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 2519 vfs_drain_busy_pages(bp); 2520 vfs_setdirty_locked_object(bp); 2521 for (i = 0; i < bp->b_npages; i++) { 2522 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2523 eoff = noff; 2524 if (eoff > bp->b_offset + bp->b_bufsize) 2525 eoff = bp->b_offset + bp->b_bufsize; 2526 m = bp->b_pages[i]; 2527 vfs_page_set_validclean(bp, foff, m); 2528 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2529 foff = noff; 2530 } 2531 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 2532} 2533 2534static void 2535vfs_setdirty_locked_object(struct buf *bp) 2536{ 2537 vm_object_t object; 2538 int i; 2539 2540 object = bp->b_bufobj->bo_object; 2541 VM_OBJECT_ASSERT_WLOCKED(object); 2542 2543 /* 2544 * We qualify the scan for modified pages on whether the 2545 * object has been flushed yet. 2546 */ 2547 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { 2548 vm_offset_t boffset; 2549 vm_offset_t eoffset; 2550 2551 /* 2552 * test the pages to see if they have been modified directly 2553 * by users through the VM system. 2554 */ 2555 for (i = 0; i < bp->b_npages; i++) 2556 vm_page_test_dirty(bp->b_pages[i]); 2557 2558 /* 2559 * Calculate the encompassing dirty range, boffset and eoffset, 2560 * (eoffset - boffset) bytes. 2561 */ 2562 2563 for (i = 0; i < bp->b_npages; i++) { 2564 if (bp->b_pages[i]->dirty) 2565 break; 2566 } 2567 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2568 2569 for (i = bp->b_npages - 1; i >= 0; --i) { 2570 if (bp->b_pages[i]->dirty) { 2571 break; 2572 } 2573 } 2574 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2575 2576 /* 2577 * Fit it to the buffer. 2578 */ 2579 2580 if (eoffset > bp->b_bcount) 2581 eoffset = bp->b_bcount; 2582 2583 /* 2584 * If we have a good dirty range, merge with the existing 2585 * dirty range. 2586 */ 2587 2588 if (boffset < eoffset) { 2589 if (bp->b_dirtyoff > boffset) 2590 bp->b_dirtyoff = boffset; 2591 if (bp->b_dirtyend < eoffset) 2592 bp->b_dirtyend = eoffset; 2593 } 2594 } 2595} 2596 2597/* 2598 * getblk: 2599 * 2600 * Get a block given a specified block and offset into a file/device. 2601 * The buffers B_DONE bit will be cleared on return, making it almost 2602 * ready for an I/O initiation. B_INVAL may or may not be set on 2603 * return. The caller should clear B_INVAL prior to initiating a 2604 * READ. 2605 * 2606 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2607 * an existing buffer. 2608 * 2609 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2610 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2611 * and then cleared based on the backing VM. If the previous buffer is 2612 * non-0-sized but invalid, B_CACHE will be cleared. 2613 * 2614 * If getblk() must create a new buffer, the new buffer is returned with 2615 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2616 * case it is returned with B_INVAL clear and B_CACHE set based on the 2617 * backing VM. 2618 * 2619 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2620 * B_CACHE bit is clear. 2621 * 2622 * What this means, basically, is that the caller should use B_CACHE to 2623 * determine whether the buffer is fully valid or not and should clear 2624 * B_INVAL prior to issuing a read. If the caller intends to validate 2625 * the buffer by loading its data area with something, the caller needs 2626 * to clear B_INVAL. If the caller does this without issuing an I/O, 2627 * the caller should set B_CACHE ( as an optimization ), else the caller 2628 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2629 * a write attempt or if it was a successfull read. If the caller 2630 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2631 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2632 */ 2633struct buf * 2634getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2635 int flags) 2636{ 2637 struct buf *bp; 2638 struct bufobj *bo; 2639 int error; 2640 2641 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2642 ASSERT_VOP_LOCKED(vp, "getblk"); 2643 if (size > MAXBSIZE) 2644 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2645 2646 bo = &vp->v_bufobj; 2647loop: 2648 /* 2649 * Block if we are low on buffers. Certain processes are allowed 2650 * to completely exhaust the buffer cache. 2651 * 2652 * If this check ever becomes a bottleneck it may be better to 2653 * move it into the else, when gbincore() fails. At the moment 2654 * it isn't a problem. 2655 */ 2656 if (numfreebuffers == 0) { 2657 if (TD_IS_IDLETHREAD(curthread)) 2658 return NULL; 2659 mtx_lock(&nblock); 2660 needsbuffer |= VFS_BIO_NEED_ANY; 2661 mtx_unlock(&nblock); 2662 } 2663 2664 BO_LOCK(bo); 2665 bp = gbincore(bo, blkno); 2666 if (bp != NULL) { 2667 int lockflags; 2668 /* 2669 * Buffer is in-core. If the buffer is not busy nor managed, 2670 * it must be on a queue. 2671 */ 2672 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2673 2674 if (flags & GB_LOCK_NOWAIT) 2675 lockflags |= LK_NOWAIT; 2676 2677 error = BUF_TIMELOCK(bp, lockflags, 2678 BO_MTX(bo), "getblk", slpflag, slptimeo); 2679 2680 /* 2681 * If we slept and got the lock we have to restart in case 2682 * the buffer changed identities. 2683 */ 2684 if (error == ENOLCK) 2685 goto loop; 2686 /* We timed out or were interrupted. */ 2687 else if (error) 2688 return (NULL); 2689 /* If recursed, assume caller knows the rules. */ 2690 else if (BUF_LOCKRECURSED(bp)) 2691 goto end; 2692 2693 /* 2694 * The buffer is locked. B_CACHE is cleared if the buffer is 2695 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2696 * and for a VMIO buffer B_CACHE is adjusted according to the 2697 * backing VM cache. 2698 */ 2699 if (bp->b_flags & B_INVAL) 2700 bp->b_flags &= ~B_CACHE; 2701 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2702 bp->b_flags |= B_CACHE; 2703 if (bp->b_flags & B_MANAGED) 2704 MPASS(bp->b_qindex == QUEUE_NONE); 2705 else { 2706 BO_LOCK(bo); 2707 bremfree(bp); 2708 BO_UNLOCK(bo); 2709 } 2710 2711 /* 2712 * check for size inconsistancies for non-VMIO case. 2713 */ 2714 2715 if (bp->b_bcount != size) { 2716 if ((bp->b_flags & B_VMIO) == 0 || 2717 (size > bp->b_kvasize)) { 2718 if (bp->b_flags & B_DELWRI) { 2719 /* 2720 * If buffer is pinned and caller does 2721 * not want sleep waiting for it to be 2722 * unpinned, bail out 2723 * */ 2724 if (bp->b_pin_count > 0) { 2725 if (flags & GB_LOCK_NOWAIT) { 2726 bqrelse(bp); 2727 return (NULL); 2728 } else { 2729 bunpin_wait(bp); 2730 } 2731 } 2732 bp->b_flags |= B_NOCACHE; 2733 bwrite(bp); 2734 } else { 2735 if (LIST_EMPTY(&bp->b_dep)) { 2736 bp->b_flags |= B_RELBUF; 2737 brelse(bp); 2738 } else { 2739 bp->b_flags |= B_NOCACHE; 2740 bwrite(bp); 2741 } 2742 } 2743 goto loop; 2744 } 2745 } 2746 2747 /* 2748 * If the size is inconsistant in the VMIO case, we can resize 2749 * the buffer. This might lead to B_CACHE getting set or 2750 * cleared. If the size has not changed, B_CACHE remains 2751 * unchanged from its previous state. 2752 */ 2753 2754 if (bp->b_bcount != size) 2755 allocbuf(bp, size); 2756 2757 KASSERT(bp->b_offset != NOOFFSET, 2758 ("getblk: no buffer offset")); 2759 2760 /* 2761 * A buffer with B_DELWRI set and B_CACHE clear must 2762 * be committed before we can return the buffer in 2763 * order to prevent the caller from issuing a read 2764 * ( due to B_CACHE not being set ) and overwriting 2765 * it. 2766 * 2767 * Most callers, including NFS and FFS, need this to 2768 * operate properly either because they assume they 2769 * can issue a read if B_CACHE is not set, or because 2770 * ( for example ) an uncached B_DELWRI might loop due 2771 * to softupdates re-dirtying the buffer. In the latter 2772 * case, B_CACHE is set after the first write completes, 2773 * preventing further loops. 2774 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2775 * above while extending the buffer, we cannot allow the 2776 * buffer to remain with B_CACHE set after the write 2777 * completes or it will represent a corrupt state. To 2778 * deal with this we set B_NOCACHE to scrap the buffer 2779 * after the write. 2780 * 2781 * We might be able to do something fancy, like setting 2782 * B_CACHE in bwrite() except if B_DELWRI is already set, 2783 * so the below call doesn't set B_CACHE, but that gets real 2784 * confusing. This is much easier. 2785 */ 2786 2787 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2788 bp->b_flags |= B_NOCACHE; 2789 bwrite(bp); 2790 goto loop; 2791 } 2792 bp->b_flags &= ~B_DONE; 2793 } else { 2794 int bsize, maxsize, vmio; 2795 off_t offset; 2796 2797 /* 2798 * Buffer is not in-core, create new buffer. The buffer 2799 * returned by getnewbuf() is locked. Note that the returned 2800 * buffer is also considered valid (not marked B_INVAL). 2801 */ 2802 BO_UNLOCK(bo); 2803 /* 2804 * If the user does not want us to create the buffer, bail out 2805 * here. 2806 */ 2807 if (flags & GB_NOCREAT) 2808 return NULL; 2809 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 2810 offset = blkno * bsize; 2811 vmio = vp->v_object != NULL; 2812 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2813 maxsize = imax(maxsize, bsize); 2814 2815 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 2816 if (bp == NULL) { 2817 if (slpflag || slptimeo) 2818 return NULL; 2819 goto loop; 2820 } 2821 2822 /* 2823 * This code is used to make sure that a buffer is not 2824 * created while the getnewbuf routine is blocked. 2825 * This can be a problem whether the vnode is locked or not. 2826 * If the buffer is created out from under us, we have to 2827 * throw away the one we just created. 2828 * 2829 * Note: this must occur before we associate the buffer 2830 * with the vp especially considering limitations in 2831 * the splay tree implementation when dealing with duplicate 2832 * lblkno's. 2833 */ 2834 BO_LOCK(bo); 2835 if (gbincore(bo, blkno)) { 2836 BO_UNLOCK(bo); 2837 bp->b_flags |= B_INVAL; 2838 brelse(bp); 2839 goto loop; 2840 } 2841 2842 /* 2843 * Insert the buffer into the hash, so that it can 2844 * be found by incore. 2845 */ 2846 bp->b_blkno = bp->b_lblkno = blkno; 2847 bp->b_offset = offset; 2848 bgetvp(vp, bp); 2849 BO_UNLOCK(bo); 2850 2851 /* 2852 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2853 * buffer size starts out as 0, B_CACHE will be set by 2854 * allocbuf() for the VMIO case prior to it testing the 2855 * backing store for validity. 2856 */ 2857 2858 if (vmio) { 2859 bp->b_flags |= B_VMIO; 2860 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2861 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2862 bp, vp->v_object, bp->b_bufobj->bo_object)); 2863 } else { 2864 bp->b_flags &= ~B_VMIO; 2865 KASSERT(bp->b_bufobj->bo_object == NULL, 2866 ("ARGH! has b_bufobj->bo_object %p %p\n", 2867 bp, bp->b_bufobj->bo_object)); 2868 } 2869 2870 allocbuf(bp, size); 2871 bp->b_flags &= ~B_DONE; 2872 } 2873 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2874 BUF_ASSERT_HELD(bp); 2875end: 2876 KASSERT(bp->b_bufobj == bo, 2877 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2878 return (bp); 2879} 2880 2881/* 2882 * Get an empty, disassociated buffer of given size. The buffer is initially 2883 * set to B_INVAL. 2884 */ 2885struct buf * 2886geteblk(int size, int flags) 2887{ 2888 struct buf *bp; 2889 int maxsize; 2890 2891 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2892 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 2893 if ((flags & GB_NOWAIT_BD) && 2894 (curthread->td_pflags & TDP_BUFNEED) != 0) 2895 return (NULL); 2896 } 2897 allocbuf(bp, size); 2898 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2899 BUF_ASSERT_HELD(bp); 2900 return (bp); 2901} 2902 2903 2904/* 2905 * This code constitutes the buffer memory from either anonymous system 2906 * memory (in the case of non-VMIO operations) or from an associated 2907 * VM object (in the case of VMIO operations). This code is able to 2908 * resize a buffer up or down. 2909 * 2910 * Note that this code is tricky, and has many complications to resolve 2911 * deadlock or inconsistant data situations. Tread lightly!!! 2912 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2913 * the caller. Calling this code willy nilly can result in the loss of data. 2914 * 2915 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2916 * B_CACHE for the non-VMIO case. 2917 */ 2918 2919int 2920allocbuf(struct buf *bp, int size) 2921{ 2922 int newbsize, mbsize; 2923 int i; 2924 2925 BUF_ASSERT_HELD(bp); 2926 2927 if (bp->b_kvasize < size) 2928 panic("allocbuf: buffer too small"); 2929 2930 if ((bp->b_flags & B_VMIO) == 0) { 2931 caddr_t origbuf; 2932 int origbufsize; 2933 /* 2934 * Just get anonymous memory from the kernel. Don't 2935 * mess with B_CACHE. 2936 */ 2937 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2938 if (bp->b_flags & B_MALLOC) 2939 newbsize = mbsize; 2940 else 2941 newbsize = round_page(size); 2942 2943 if (newbsize < bp->b_bufsize) { 2944 /* 2945 * malloced buffers are not shrunk 2946 */ 2947 if (bp->b_flags & B_MALLOC) { 2948 if (newbsize) { 2949 bp->b_bcount = size; 2950 } else { 2951 free(bp->b_data, M_BIOBUF); 2952 if (bp->b_bufsize) { 2953 atomic_subtract_long( 2954 &bufmallocspace, 2955 bp->b_bufsize); 2956 bufspacewakeup(); 2957 bp->b_bufsize = 0; 2958 } 2959 bp->b_saveaddr = bp->b_kvabase; 2960 bp->b_data = bp->b_saveaddr; 2961 bp->b_bcount = 0; 2962 bp->b_flags &= ~B_MALLOC; 2963 } 2964 return 1; 2965 } 2966 vm_hold_free_pages(bp, newbsize); 2967 } else if (newbsize > bp->b_bufsize) { 2968 /* 2969 * We only use malloced memory on the first allocation. 2970 * and revert to page-allocated memory when the buffer 2971 * grows. 2972 */ 2973 /* 2974 * There is a potential smp race here that could lead 2975 * to bufmallocspace slightly passing the max. It 2976 * is probably extremely rare and not worth worrying 2977 * over. 2978 */ 2979 if ( (bufmallocspace < maxbufmallocspace) && 2980 (bp->b_bufsize == 0) && 2981 (mbsize <= PAGE_SIZE/2)) { 2982 2983 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2984 bp->b_bufsize = mbsize; 2985 bp->b_bcount = size; 2986 bp->b_flags |= B_MALLOC; 2987 atomic_add_long(&bufmallocspace, mbsize); 2988 return 1; 2989 } 2990 origbuf = NULL; 2991 origbufsize = 0; 2992 /* 2993 * If the buffer is growing on its other-than-first allocation, 2994 * then we revert to the page-allocation scheme. 2995 */ 2996 if (bp->b_flags & B_MALLOC) { 2997 origbuf = bp->b_data; 2998 origbufsize = bp->b_bufsize; 2999 bp->b_data = bp->b_kvabase; 3000 if (bp->b_bufsize) { 3001 atomic_subtract_long(&bufmallocspace, 3002 bp->b_bufsize); 3003 bufspacewakeup(); 3004 bp->b_bufsize = 0; 3005 } 3006 bp->b_flags &= ~B_MALLOC; 3007 newbsize = round_page(newbsize); 3008 } 3009 vm_hold_load_pages( 3010 bp, 3011 (vm_offset_t) bp->b_data + bp->b_bufsize, 3012 (vm_offset_t) bp->b_data + newbsize); 3013 if (origbuf) { 3014 bcopy(origbuf, bp->b_data, origbufsize); 3015 free(origbuf, M_BIOBUF); 3016 } 3017 } 3018 } else { 3019 int desiredpages; 3020 3021 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3022 desiredpages = (size == 0) ? 0 : 3023 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 3024 3025 if (bp->b_flags & B_MALLOC) 3026 panic("allocbuf: VMIO buffer can't be malloced"); 3027 /* 3028 * Set B_CACHE initially if buffer is 0 length or will become 3029 * 0-length. 3030 */ 3031 if (size == 0 || bp->b_bufsize == 0) 3032 bp->b_flags |= B_CACHE; 3033 3034 if (newbsize < bp->b_bufsize) { 3035 /* 3036 * DEV_BSIZE aligned new buffer size is less then the 3037 * DEV_BSIZE aligned existing buffer size. Figure out 3038 * if we have to remove any pages. 3039 */ 3040 if (desiredpages < bp->b_npages) { 3041 vm_page_t m; 3042 3043 pmap_qremove((vm_offset_t)trunc_page( 3044 (vm_offset_t)bp->b_data) + 3045 (desiredpages << PAGE_SHIFT), 3046 (bp->b_npages - desiredpages)); 3047 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3048 for (i = desiredpages; i < bp->b_npages; i++) { 3049 /* 3050 * the page is not freed here -- it 3051 * is the responsibility of 3052 * vnode_pager_setsize 3053 */ 3054 m = bp->b_pages[i]; 3055 KASSERT(m != bogus_page, 3056 ("allocbuf: bogus page found")); 3057 while (vm_page_sleep_if_busy(m, TRUE, 3058 "biodep")) 3059 continue; 3060 3061 bp->b_pages[i] = NULL; 3062 vm_page_lock(m); 3063 vm_page_unwire(m, 0); 3064 vm_page_unlock(m); 3065 } 3066 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3067 bp->b_npages = desiredpages; 3068 } 3069 } else if (size > bp->b_bcount) { 3070 /* 3071 * We are growing the buffer, possibly in a 3072 * byte-granular fashion. 3073 */ 3074 vm_object_t obj; 3075 vm_offset_t toff; 3076 vm_offset_t tinc; 3077 3078 /* 3079 * Step 1, bring in the VM pages from the object, 3080 * allocating them if necessary. We must clear 3081 * B_CACHE if these pages are not valid for the 3082 * range covered by the buffer. 3083 */ 3084 3085 obj = bp->b_bufobj->bo_object; 3086 3087 VM_OBJECT_WLOCK(obj); 3088 while (bp->b_npages < desiredpages) { 3089 vm_page_t m; 3090 3091 /* 3092 * We must allocate system pages since blocking 3093 * here could interfere with paging I/O, no 3094 * matter which process we are. 3095 * 3096 * We can only test VPO_BUSY here. Blocking on 3097 * m->busy might lead to a deadlock: 3098 * vm_fault->getpages->cluster_read->allocbuf 3099 * Thus, we specify VM_ALLOC_IGN_SBUSY. 3100 */ 3101 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + 3102 bp->b_npages, VM_ALLOC_NOBUSY | 3103 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3104 VM_ALLOC_RETRY | VM_ALLOC_IGN_SBUSY | 3105 VM_ALLOC_COUNT(desiredpages - bp->b_npages)); 3106 if (m->valid == 0) 3107 bp->b_flags &= ~B_CACHE; 3108 bp->b_pages[bp->b_npages] = m; 3109 ++bp->b_npages; 3110 } 3111 3112 /* 3113 * Step 2. We've loaded the pages into the buffer, 3114 * we have to figure out if we can still have B_CACHE 3115 * set. Note that B_CACHE is set according to the 3116 * byte-granular range ( bcount and size ), new the 3117 * aligned range ( newbsize ). 3118 * 3119 * The VM test is against m->valid, which is DEV_BSIZE 3120 * aligned. Needless to say, the validity of the data 3121 * needs to also be DEV_BSIZE aligned. Note that this 3122 * fails with NFS if the server or some other client 3123 * extends the file's EOF. If our buffer is resized, 3124 * B_CACHE may remain set! XXX 3125 */ 3126 3127 toff = bp->b_bcount; 3128 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3129 3130 while ((bp->b_flags & B_CACHE) && toff < size) { 3131 vm_pindex_t pi; 3132 3133 if (tinc > (size - toff)) 3134 tinc = size - toff; 3135 3136 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3137 PAGE_SHIFT; 3138 3139 vfs_buf_test_cache( 3140 bp, 3141 bp->b_offset, 3142 toff, 3143 tinc, 3144 bp->b_pages[pi] 3145 ); 3146 toff += tinc; 3147 tinc = PAGE_SIZE; 3148 } 3149 VM_OBJECT_WUNLOCK(obj); 3150 3151 /* 3152 * Step 3, fixup the KVM pmap. Remember that 3153 * bp->b_data is relative to bp->b_offset, but 3154 * bp->b_offset may be offset into the first page. 3155 */ 3156 3157 bp->b_data = (caddr_t) 3158 trunc_page((vm_offset_t)bp->b_data); 3159 pmap_qenter( 3160 (vm_offset_t)bp->b_data, 3161 bp->b_pages, 3162 bp->b_npages 3163 ); 3164 3165 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3166 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 3167 } 3168 } 3169 if (newbsize < bp->b_bufsize) 3170 bufspacewakeup(); 3171 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3172 bp->b_bcount = size; /* requested buffer size */ 3173 return 1; 3174} 3175 3176void 3177biodone(struct bio *bp) 3178{ 3179 struct mtx *mtxp; 3180 void (*done)(struct bio *); 3181 3182 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3183 mtx_lock(mtxp); 3184 bp->bio_flags |= BIO_DONE; 3185 done = bp->bio_done; 3186 if (done == NULL) 3187 wakeup(bp); 3188 mtx_unlock(mtxp); 3189 if (done != NULL) 3190 done(bp); 3191} 3192 3193/* 3194 * Wait for a BIO to finish. 3195 * 3196 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3197 * case is not yet clear. 3198 */ 3199int 3200biowait(struct bio *bp, const char *wchan) 3201{ 3202 struct mtx *mtxp; 3203 3204 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3205 mtx_lock(mtxp); 3206 while ((bp->bio_flags & BIO_DONE) == 0) 3207 msleep(bp, mtxp, PRIBIO, wchan, hz / 10); 3208 mtx_unlock(mtxp); 3209 if (bp->bio_error != 0) 3210 return (bp->bio_error); 3211 if (!(bp->bio_flags & BIO_ERROR)) 3212 return (0); 3213 return (EIO); 3214} 3215 3216void 3217biofinish(struct bio *bp, struct devstat *stat, int error) 3218{ 3219 3220 if (error) { 3221 bp->bio_error = error; 3222 bp->bio_flags |= BIO_ERROR; 3223 } 3224 if (stat != NULL) 3225 devstat_end_transaction_bio(stat, bp); 3226 biodone(bp); 3227} 3228 3229/* 3230 * bufwait: 3231 * 3232 * Wait for buffer I/O completion, returning error status. The buffer 3233 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3234 * error and cleared. 3235 */ 3236int 3237bufwait(struct buf *bp) 3238{ 3239 if (bp->b_iocmd == BIO_READ) 3240 bwait(bp, PRIBIO, "biord"); 3241 else 3242 bwait(bp, PRIBIO, "biowr"); 3243 if (bp->b_flags & B_EINTR) { 3244 bp->b_flags &= ~B_EINTR; 3245 return (EINTR); 3246 } 3247 if (bp->b_ioflags & BIO_ERROR) { 3248 return (bp->b_error ? bp->b_error : EIO); 3249 } else { 3250 return (0); 3251 } 3252} 3253 3254 /* 3255 * Call back function from struct bio back up to struct buf. 3256 */ 3257static void 3258bufdonebio(struct bio *bip) 3259{ 3260 struct buf *bp; 3261 3262 bp = bip->bio_caller2; 3263 bp->b_resid = bp->b_bcount - bip->bio_completed; 3264 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3265 bp->b_ioflags = bip->bio_flags; 3266 bp->b_error = bip->bio_error; 3267 if (bp->b_error) 3268 bp->b_ioflags |= BIO_ERROR; 3269 bufdone(bp); 3270 g_destroy_bio(bip); 3271} 3272 3273void 3274dev_strategy(struct cdev *dev, struct buf *bp) 3275{ 3276 struct cdevsw *csw; 3277 struct bio *bip; 3278 int ref; 3279 3280 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3281 panic("b_iocmd botch"); 3282 for (;;) { 3283 bip = g_new_bio(); 3284 if (bip != NULL) 3285 break; 3286 /* Try again later */ 3287 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3288 } 3289 bip->bio_cmd = bp->b_iocmd; 3290 bip->bio_offset = bp->b_iooffset; 3291 bip->bio_length = bp->b_bcount; 3292 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3293 bip->bio_data = bp->b_data; 3294 bip->bio_done = bufdonebio; 3295 bip->bio_caller2 = bp; 3296 bip->bio_dev = dev; 3297 KASSERT(dev->si_refcount > 0, 3298 ("dev_strategy on un-referenced struct cdev *(%s)", 3299 devtoname(dev))); 3300 csw = dev_refthread(dev, &ref); 3301 if (csw == NULL) { 3302 g_destroy_bio(bip); 3303 bp->b_error = ENXIO; 3304 bp->b_ioflags = BIO_ERROR; 3305 bufdone(bp); 3306 return; 3307 } 3308 (*csw->d_strategy)(bip); 3309 dev_relthread(dev, ref); 3310} 3311 3312/* 3313 * bufdone: 3314 * 3315 * Finish I/O on a buffer, optionally calling a completion function. 3316 * This is usually called from an interrupt so process blocking is 3317 * not allowed. 3318 * 3319 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3320 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3321 * assuming B_INVAL is clear. 3322 * 3323 * For the VMIO case, we set B_CACHE if the op was a read and no 3324 * read error occured, or if the op was a write. B_CACHE is never 3325 * set if the buffer is invalid or otherwise uncacheable. 3326 * 3327 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3328 * initiator to leave B_INVAL set to brelse the buffer out of existance 3329 * in the biodone routine. 3330 */ 3331void 3332bufdone(struct buf *bp) 3333{ 3334 struct bufobj *dropobj; 3335 void (*biodone)(struct buf *); 3336 3337 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3338 dropobj = NULL; 3339 3340 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3341 BUF_ASSERT_HELD(bp); 3342 3343 runningbufwakeup(bp); 3344 if (bp->b_iocmd == BIO_WRITE) 3345 dropobj = bp->b_bufobj; 3346 /* call optional completion function if requested */ 3347 if (bp->b_iodone != NULL) { 3348 biodone = bp->b_iodone; 3349 bp->b_iodone = NULL; 3350 (*biodone) (bp); 3351 if (dropobj) 3352 bufobj_wdrop(dropobj); 3353 return; 3354 } 3355 3356 bufdone_finish(bp); 3357 3358 if (dropobj) 3359 bufobj_wdrop(dropobj); 3360} 3361 3362void 3363bufdone_finish(struct buf *bp) 3364{ 3365 BUF_ASSERT_HELD(bp); 3366 3367 if (!LIST_EMPTY(&bp->b_dep)) 3368 buf_complete(bp); 3369 3370 if (bp->b_flags & B_VMIO) { 3371 vm_ooffset_t foff; 3372 vm_page_t m; 3373 vm_object_t obj; 3374 struct vnode *vp; 3375 int bogus, i, iosize; 3376 3377 obj = bp->b_bufobj->bo_object; 3378 KASSERT(obj->paging_in_progress >= bp->b_npages, 3379 ("biodone_finish: paging in progress(%d) < b_npages(%d)", 3380 obj->paging_in_progress, bp->b_npages)); 3381 3382 vp = bp->b_vp; 3383 KASSERT(vp->v_holdcnt > 0, 3384 ("biodone_finish: vnode %p has zero hold count", vp)); 3385 KASSERT(vp->v_object != NULL, 3386 ("biodone_finish: vnode %p has no vm_object", vp)); 3387 3388 foff = bp->b_offset; 3389 KASSERT(bp->b_offset != NOOFFSET, 3390 ("biodone_finish: bp %p has no buffer offset", bp)); 3391 3392 /* 3393 * Set B_CACHE if the op was a normal read and no error 3394 * occured. B_CACHE is set for writes in the b*write() 3395 * routines. 3396 */ 3397 iosize = bp->b_bcount - bp->b_resid; 3398 if (bp->b_iocmd == BIO_READ && 3399 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3400 !(bp->b_ioflags & BIO_ERROR)) { 3401 bp->b_flags |= B_CACHE; 3402 } 3403 bogus = 0; 3404 VM_OBJECT_WLOCK(obj); 3405 for (i = 0; i < bp->b_npages; i++) { 3406 int bogusflag = 0; 3407 int resid; 3408 3409 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3410 if (resid > iosize) 3411 resid = iosize; 3412 3413 /* 3414 * cleanup bogus pages, restoring the originals 3415 */ 3416 m = bp->b_pages[i]; 3417 if (m == bogus_page) { 3418 bogus = bogusflag = 1; 3419 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3420 if (m == NULL) 3421 panic("biodone: page disappeared!"); 3422 bp->b_pages[i] = m; 3423 } 3424 KASSERT(OFF_TO_IDX(foff) == m->pindex, 3425 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch", 3426 (intmax_t)foff, (uintmax_t)m->pindex)); 3427 3428 /* 3429 * In the write case, the valid and clean bits are 3430 * already changed correctly ( see bdwrite() ), so we 3431 * only need to do this here in the read case. 3432 */ 3433 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3434 KASSERT((m->dirty & vm_page_bits(foff & 3435 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3436 " page %p has unexpected dirty bits", m)); 3437 vfs_page_set_valid(bp, foff, m); 3438 } 3439 3440 vm_page_io_finish(m); 3441 vm_object_pip_subtract(obj, 1); 3442 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3443 iosize -= resid; 3444 } 3445 vm_object_pip_wakeupn(obj, 0); 3446 VM_OBJECT_WUNLOCK(obj); 3447 if (bogus) 3448 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3449 bp->b_pages, bp->b_npages); 3450 } 3451 3452 /* 3453 * For asynchronous completions, release the buffer now. The brelse 3454 * will do a wakeup there if necessary - so no need to do a wakeup 3455 * here in the async case. The sync case always needs to do a wakeup. 3456 */ 3457 3458 if (bp->b_flags & B_ASYNC) { 3459 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3460 brelse(bp); 3461 else 3462 bqrelse(bp); 3463 } else 3464 bdone(bp); 3465} 3466 3467/* 3468 * This routine is called in lieu of iodone in the case of 3469 * incomplete I/O. This keeps the busy status for pages 3470 * consistant. 3471 */ 3472void 3473vfs_unbusy_pages(struct buf *bp) 3474{ 3475 int i; 3476 vm_object_t obj; 3477 vm_page_t m; 3478 3479 runningbufwakeup(bp); 3480 if (!(bp->b_flags & B_VMIO)) 3481 return; 3482 3483 obj = bp->b_bufobj->bo_object; 3484 VM_OBJECT_WLOCK(obj); 3485 for (i = 0; i < bp->b_npages; i++) { 3486 m = bp->b_pages[i]; 3487 if (m == bogus_page) { 3488 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3489 if (!m) 3490 panic("vfs_unbusy_pages: page missing\n"); 3491 bp->b_pages[i] = m; 3492 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3493 bp->b_pages, bp->b_npages); 3494 } 3495 vm_object_pip_subtract(obj, 1); 3496 vm_page_io_finish(m); 3497 } 3498 vm_object_pip_wakeupn(obj, 0); 3499 VM_OBJECT_WUNLOCK(obj); 3500} 3501 3502/* 3503 * vfs_page_set_valid: 3504 * 3505 * Set the valid bits in a page based on the supplied offset. The 3506 * range is restricted to the buffer's size. 3507 * 3508 * This routine is typically called after a read completes. 3509 */ 3510static void 3511vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3512{ 3513 vm_ooffset_t eoff; 3514 3515 /* 3516 * Compute the end offset, eoff, such that [off, eoff) does not span a 3517 * page boundary and eoff is not greater than the end of the buffer. 3518 * The end of the buffer, in this case, is our file EOF, not the 3519 * allocation size of the buffer. 3520 */ 3521 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3522 if (eoff > bp->b_offset + bp->b_bcount) 3523 eoff = bp->b_offset + bp->b_bcount; 3524 3525 /* 3526 * Set valid range. This is typically the entire buffer and thus the 3527 * entire page. 3528 */ 3529 if (eoff > off) 3530 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 3531} 3532 3533/* 3534 * vfs_page_set_validclean: 3535 * 3536 * Set the valid bits and clear the dirty bits in a page based on the 3537 * supplied offset. The range is restricted to the buffer's size. 3538 */ 3539static void 3540vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3541{ 3542 vm_ooffset_t soff, eoff; 3543 3544 /* 3545 * Start and end offsets in buffer. eoff - soff may not cross a 3546 * page boundry or cross the end of the buffer. The end of the 3547 * buffer, in this case, is our file EOF, not the allocation size 3548 * of the buffer. 3549 */ 3550 soff = off; 3551 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3552 if (eoff > bp->b_offset + bp->b_bcount) 3553 eoff = bp->b_offset + bp->b_bcount; 3554 3555 /* 3556 * Set valid range. This is typically the entire buffer and thus the 3557 * entire page. 3558 */ 3559 if (eoff > soff) { 3560 vm_page_set_validclean( 3561 m, 3562 (vm_offset_t) (soff & PAGE_MASK), 3563 (vm_offset_t) (eoff - soff) 3564 ); 3565 } 3566} 3567 3568/* 3569 * Ensure that all buffer pages are not busied by VPO_BUSY flag. If 3570 * any page is busy, drain the flag. 3571 */ 3572static void 3573vfs_drain_busy_pages(struct buf *bp) 3574{ 3575 vm_page_t m; 3576 int i, last_busied; 3577 3578 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); 3579 last_busied = 0; 3580 for (i = 0; i < bp->b_npages; i++) { 3581 m = bp->b_pages[i]; 3582 if ((m->oflags & VPO_BUSY) != 0) { 3583 for (; last_busied < i; last_busied++) 3584 vm_page_busy(bp->b_pages[last_busied]); 3585 while ((m->oflags & VPO_BUSY) != 0) 3586 vm_page_sleep(m, "vbpage"); 3587 } 3588 } 3589 for (i = 0; i < last_busied; i++) 3590 vm_page_wakeup(bp->b_pages[i]); 3591} 3592 3593/* 3594 * This routine is called before a device strategy routine. 3595 * It is used to tell the VM system that paging I/O is in 3596 * progress, and treat the pages associated with the buffer 3597 * almost as being VPO_BUSY. Also the object paging_in_progress 3598 * flag is handled to make sure that the object doesn't become 3599 * inconsistant. 3600 * 3601 * Since I/O has not been initiated yet, certain buffer flags 3602 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3603 * and should be ignored. 3604 */ 3605void 3606vfs_busy_pages(struct buf *bp, int clear_modify) 3607{ 3608 int i, bogus; 3609 vm_object_t obj; 3610 vm_ooffset_t foff; 3611 vm_page_t m; 3612 3613 if (!(bp->b_flags & B_VMIO)) 3614 return; 3615 3616 obj = bp->b_bufobj->bo_object; 3617 foff = bp->b_offset; 3618 KASSERT(bp->b_offset != NOOFFSET, 3619 ("vfs_busy_pages: no buffer offset")); 3620 VM_OBJECT_WLOCK(obj); 3621 vfs_drain_busy_pages(bp); 3622 if (bp->b_bufsize != 0) 3623 vfs_setdirty_locked_object(bp); 3624 bogus = 0; 3625 for (i = 0; i < bp->b_npages; i++) { 3626 m = bp->b_pages[i]; 3627 3628 if ((bp->b_flags & B_CLUSTER) == 0) { 3629 vm_object_pip_add(obj, 1); 3630 vm_page_io_start(m); 3631 } 3632 /* 3633 * When readying a buffer for a read ( i.e 3634 * clear_modify == 0 ), it is important to do 3635 * bogus_page replacement for valid pages in 3636 * partially instantiated buffers. Partially 3637 * instantiated buffers can, in turn, occur when 3638 * reconstituting a buffer from its VM backing store 3639 * base. We only have to do this if B_CACHE is 3640 * clear ( which causes the I/O to occur in the 3641 * first place ). The replacement prevents the read 3642 * I/O from overwriting potentially dirty VM-backed 3643 * pages. XXX bogus page replacement is, uh, bogus. 3644 * It may not work properly with small-block devices. 3645 * We need to find a better way. 3646 */ 3647 if (clear_modify) { 3648 pmap_remove_write(m); 3649 vfs_page_set_validclean(bp, foff, m); 3650 } else if (m->valid == VM_PAGE_BITS_ALL && 3651 (bp->b_flags & B_CACHE) == 0) { 3652 bp->b_pages[i] = bogus_page; 3653 bogus++; 3654 } 3655 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3656 } 3657 VM_OBJECT_WUNLOCK(obj); 3658 if (bogus) 3659 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3660 bp->b_pages, bp->b_npages); 3661} 3662 3663/* 3664 * vfs_bio_set_valid: 3665 * 3666 * Set the range within the buffer to valid. The range is 3667 * relative to the beginning of the buffer, b_offset. Note that 3668 * b_offset itself may be offset from the beginning of the first 3669 * page. 3670 */ 3671void 3672vfs_bio_set_valid(struct buf *bp, int base, int size) 3673{ 3674 int i, n; 3675 vm_page_t m; 3676 3677 if (!(bp->b_flags & B_VMIO)) 3678 return; 3679 3680 /* 3681 * Fixup base to be relative to beginning of first page. 3682 * Set initial n to be the maximum number of bytes in the 3683 * first page that can be validated. 3684 */ 3685 base += (bp->b_offset & PAGE_MASK); 3686 n = PAGE_SIZE - (base & PAGE_MASK); 3687 3688 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3689 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3690 m = bp->b_pages[i]; 3691 if (n > size) 3692 n = size; 3693 vm_page_set_valid_range(m, base & PAGE_MASK, n); 3694 base += n; 3695 size -= n; 3696 n = PAGE_SIZE; 3697 } 3698 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3699} 3700 3701/* 3702 * vfs_bio_clrbuf: 3703 * 3704 * If the specified buffer is a non-VMIO buffer, clear the entire 3705 * buffer. If the specified buffer is a VMIO buffer, clear and 3706 * validate only the previously invalid portions of the buffer. 3707 * This routine essentially fakes an I/O, so we need to clear 3708 * BIO_ERROR and B_INVAL. 3709 * 3710 * Note that while we only theoretically need to clear through b_bcount, 3711 * we go ahead and clear through b_bufsize. 3712 */ 3713void 3714vfs_bio_clrbuf(struct buf *bp) 3715{ 3716 int i, j, mask, sa, ea, slide; 3717 3718 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3719 clrbuf(bp); 3720 return; 3721 } 3722 bp->b_flags &= ~B_INVAL; 3723 bp->b_ioflags &= ~BIO_ERROR; 3724 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3725 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3726 (bp->b_offset & PAGE_MASK) == 0) { 3727 if (bp->b_pages[0] == bogus_page) 3728 goto unlock; 3729 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3730 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object); 3731 if ((bp->b_pages[0]->valid & mask) == mask) 3732 goto unlock; 3733 if ((bp->b_pages[0]->valid & mask) == 0) { 3734 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize); 3735 bp->b_pages[0]->valid |= mask; 3736 goto unlock; 3737 } 3738 } 3739 sa = bp->b_offset & PAGE_MASK; 3740 slide = 0; 3741 for (i = 0; i < bp->b_npages; i++, sa = 0) { 3742 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 3743 ea = slide & PAGE_MASK; 3744 if (ea == 0) 3745 ea = PAGE_SIZE; 3746 if (bp->b_pages[i] == bogus_page) 3747 continue; 3748 j = sa / DEV_BSIZE; 3749 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3750 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object); 3751 if ((bp->b_pages[i]->valid & mask) == mask) 3752 continue; 3753 if ((bp->b_pages[i]->valid & mask) == 0) 3754 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 3755 else { 3756 for (; sa < ea; sa += DEV_BSIZE, j++) { 3757 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 3758 pmap_zero_page_area(bp->b_pages[i], 3759 sa, DEV_BSIZE); 3760 } 3761 } 3762 } 3763 bp->b_pages[i]->valid |= mask; 3764 } 3765unlock: 3766 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3767 bp->b_resid = 0; 3768} 3769 3770/* 3771 * vm_hold_load_pages and vm_hold_free_pages get pages into 3772 * a buffers address space. The pages are anonymous and are 3773 * not associated with a file object. 3774 */ 3775static void 3776vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3777{ 3778 vm_offset_t pg; 3779 vm_page_t p; 3780 int index; 3781 3782 to = round_page(to); 3783 from = round_page(from); 3784 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3785 3786 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3787tryagain: 3788 /* 3789 * note: must allocate system pages since blocking here 3790 * could interfere with paging I/O, no matter which 3791 * process we are. 3792 */ 3793 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | 3794 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT)); 3795 if (p == NULL) { 3796 VM_WAIT; 3797 goto tryagain; 3798 } 3799 pmap_qenter(pg, &p, 1); 3800 bp->b_pages[index] = p; 3801 } 3802 bp->b_npages = index; 3803} 3804 3805/* Return pages associated with this buf to the vm system */ 3806static void 3807vm_hold_free_pages(struct buf *bp, int newbsize) 3808{ 3809 vm_offset_t from; 3810 vm_page_t p; 3811 int index, newnpages; 3812 3813 from = round_page((vm_offset_t)bp->b_data + newbsize); 3814 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3815 if (bp->b_npages > newnpages) 3816 pmap_qremove(from, bp->b_npages - newnpages); 3817 for (index = newnpages; index < bp->b_npages; index++) { 3818 p = bp->b_pages[index]; 3819 bp->b_pages[index] = NULL; 3820 if (p->busy != 0) 3821 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3822 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); 3823 p->wire_count--; 3824 vm_page_free(p); 3825 atomic_subtract_int(&cnt.v_wire_count, 1); 3826 } 3827 bp->b_npages = newnpages; 3828} 3829 3830/* 3831 * Map an IO request into kernel virtual address space. 3832 * 3833 * All requests are (re)mapped into kernel VA space. 3834 * Notice that we use b_bufsize for the size of the buffer 3835 * to be mapped. b_bcount might be modified by the driver. 3836 * 3837 * Note that even if the caller determines that the address space should 3838 * be valid, a race or a smaller-file mapped into a larger space may 3839 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3840 * check the return value. 3841 */ 3842int 3843vmapbuf(struct buf *bp) 3844{ 3845 caddr_t kva; 3846 vm_prot_t prot; 3847 int pidx; 3848 3849 if (bp->b_bufsize < 0) 3850 return (-1); 3851 prot = VM_PROT_READ; 3852 if (bp->b_iocmd == BIO_READ) 3853 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3854 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 3855 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, 3856 btoc(MAXPHYS))) < 0) 3857 return (-1); 3858 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3859 3860 kva = bp->b_saveaddr; 3861 bp->b_npages = pidx; 3862 bp->b_saveaddr = bp->b_data; 3863 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3864 return(0); 3865} 3866 3867/* 3868 * Free the io map PTEs associated with this IO operation. 3869 * We also invalidate the TLB entries and restore the original b_addr. 3870 */ 3871void 3872vunmapbuf(struct buf *bp) 3873{ 3874 int npages; 3875 3876 npages = bp->b_npages; 3877 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3878 vm_page_unhold_pages(bp->b_pages, npages); 3879 3880 bp->b_data = bp->b_saveaddr; 3881} 3882 3883void 3884bdone(struct buf *bp) 3885{ 3886 struct mtx *mtxp; 3887 3888 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3889 mtx_lock(mtxp); 3890 bp->b_flags |= B_DONE; 3891 wakeup(bp); 3892 mtx_unlock(mtxp); 3893} 3894 3895void 3896bwait(struct buf *bp, u_char pri, const char *wchan) 3897{ 3898 struct mtx *mtxp; 3899 3900 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3901 mtx_lock(mtxp); 3902 while ((bp->b_flags & B_DONE) == 0) 3903 msleep(bp, mtxp, pri, wchan, 0); 3904 mtx_unlock(mtxp); 3905} 3906 3907int 3908bufsync(struct bufobj *bo, int waitfor) 3909{ 3910 3911 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 3912} 3913 3914void 3915bufstrategy(struct bufobj *bo, struct buf *bp) 3916{ 3917 int i = 0; 3918 struct vnode *vp; 3919 3920 vp = bp->b_vp; 3921 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3922 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3923 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3924 i = VOP_STRATEGY(vp, bp); 3925 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3926} 3927 3928void 3929bufobj_wrefl(struct bufobj *bo) 3930{ 3931 3932 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3933 ASSERT_BO_LOCKED(bo); 3934 bo->bo_numoutput++; 3935} 3936 3937void 3938bufobj_wref(struct bufobj *bo) 3939{ 3940 3941 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3942 BO_LOCK(bo); 3943 bo->bo_numoutput++; 3944 BO_UNLOCK(bo); 3945} 3946 3947void 3948bufobj_wdrop(struct bufobj *bo) 3949{ 3950 3951 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3952 BO_LOCK(bo); 3953 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3954 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3955 bo->bo_flag &= ~BO_WWAIT; 3956 wakeup(&bo->bo_numoutput); 3957 } 3958 BO_UNLOCK(bo); 3959} 3960 3961int 3962bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3963{ 3964 int error; 3965 3966 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3967 ASSERT_BO_LOCKED(bo); 3968 error = 0; 3969 while (bo->bo_numoutput) { 3970 bo->bo_flag |= BO_WWAIT; 3971 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3972 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3973 if (error) 3974 break; 3975 } 3976 return (error); 3977} 3978 3979void 3980bpin(struct buf *bp) 3981{ 3982 struct mtx *mtxp; 3983 3984 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3985 mtx_lock(mtxp); 3986 bp->b_pin_count++; 3987 mtx_unlock(mtxp); 3988} 3989 3990void 3991bunpin(struct buf *bp) 3992{ 3993 struct mtx *mtxp; 3994 3995 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3996 mtx_lock(mtxp); 3997 if (--bp->b_pin_count == 0) 3998 wakeup(bp); 3999 mtx_unlock(mtxp); 4000} 4001 4002void 4003bunpin_wait(struct buf *bp) 4004{ 4005 struct mtx *mtxp; 4006 4007 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4008 mtx_lock(mtxp); 4009 while (bp->b_pin_count > 0) 4010 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4011 mtx_unlock(mtxp); 4012} 4013 4014#include "opt_ddb.h" 4015#ifdef DDB 4016#include <ddb/ddb.h> 4017 4018/* DDB command to show buffer data */ 4019DB_SHOW_COMMAND(buffer, db_show_buffer) 4020{ 4021 /* get args */ 4022 struct buf *bp = (struct buf *)addr; 4023 4024 if (!have_addr) { 4025 db_printf("usage: show buffer <addr>\n"); 4026 return; 4027 } 4028 4029 db_printf("buf at %p\n", bp); 4030 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n", 4031 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, 4032 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS); 4033 db_printf( 4034 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4035 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, " 4036 "b_dep = %p\n", 4037 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4038 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4039 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first); 4040 if (bp->b_npages) { 4041 int i; 4042 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4043 for (i = 0; i < bp->b_npages; i++) { 4044 vm_page_t m; 4045 m = bp->b_pages[i]; 4046 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4047 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4048 if ((i + 1) < bp->b_npages) 4049 db_printf(","); 4050 } 4051 db_printf("\n"); 4052 } 4053 db_printf(" "); 4054 BUF_LOCKPRINTINFO(bp); 4055} 4056 4057DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4058{ 4059 struct buf *bp; 4060 int i; 4061 4062 for (i = 0; i < nbuf; i++) { 4063 bp = &buf[i]; 4064 if (BUF_ISLOCKED(bp)) { 4065 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4066 db_printf("\n"); 4067 } 4068 } 4069} 4070 4071DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4072{ 4073 struct vnode *vp; 4074 struct buf *bp; 4075 4076 if (!have_addr) { 4077 db_printf("usage: show vnodebufs <addr>\n"); 4078 return; 4079 } 4080 vp = (struct vnode *)addr; 4081 db_printf("Clean buffers:\n"); 4082 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4083 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4084 db_printf("\n"); 4085 } 4086 db_printf("Dirty buffers:\n"); 4087 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4088 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4089 db_printf("\n"); 4090 } 4091} 4092 4093DB_COMMAND(countfreebufs, db_coundfreebufs) 4094{ 4095 struct buf *bp; 4096 int i, used = 0, nfree = 0; 4097 4098 if (have_addr) { 4099 db_printf("usage: countfreebufs\n"); 4100 return; 4101 } 4102 4103 for (i = 0; i < nbuf; i++) { 4104 bp = &buf[i]; 4105 if ((bp->b_vflags & BV_INFREECNT) != 0) 4106 nfree++; 4107 else 4108 used++; 4109 } 4110 4111 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4112 nfree + used); 4113 db_printf("numfreebuffers is %d\n", numfreebuffers); 4114} 4115#endif /* DDB */ 4116