vfs_bio.c revision 248084
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 248084 2013-03-09 02:32:23Z attilio $"); 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 return nwritten; 1799 } 1800 } 1801 bremfree(bp); 1802 bp->b_flags |= B_ASYNC; 1803 /* 1804 * default (old) behavior, writing out only one block 1805 * 1806 * XXX returns b_bufsize instead of b_bcount for nwritten? 1807 */ 1808 nwritten = bp->b_bufsize; 1809 (void) bwrite(bp); 1810 1811 return nwritten; 1812} 1813 1814/* 1815 * getnewbuf: 1816 * 1817 * Find and initialize a new buffer header, freeing up existing buffers 1818 * in the bufqueues as necessary. The new buffer is returned locked. 1819 * 1820 * Important: B_INVAL is not set. If the caller wishes to throw the 1821 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1822 * 1823 * We block if: 1824 * We have insufficient buffer headers 1825 * We have insufficient buffer space 1826 * buffer_map is too fragmented ( space reservation fails ) 1827 * If we have to flush dirty buffers ( but we try to avoid this ) 1828 * 1829 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1830 * Instead we ask the buf daemon to do it for us. We attempt to 1831 * avoid piecemeal wakeups of the pageout daemon. 1832 */ 1833 1834static struct buf * 1835getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 1836 int gbflags) 1837{ 1838 struct thread *td; 1839 struct buf *bp; 1840 struct buf *nbp; 1841 int defrag = 0; 1842 int nqindex; 1843 static int flushingbufs; 1844 1845 td = curthread; 1846 /* 1847 * We can't afford to block since we might be holding a vnode lock, 1848 * which may prevent system daemons from running. We deal with 1849 * low-memory situations by proactively returning memory and running 1850 * async I/O rather then sync I/O. 1851 */ 1852 atomic_add_int(&getnewbufcalls, 1); 1853 atomic_subtract_int(&getnewbufrestarts, 1); 1854restart: 1855 atomic_add_int(&getnewbufrestarts, 1); 1856 1857 /* 1858 * Setup for scan. If we do not have enough free buffers, 1859 * we setup a degenerate case that immediately fails. Note 1860 * that if we are specially marked process, we are allowed to 1861 * dip into our reserves. 1862 * 1863 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1864 * 1865 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1866 * However, there are a number of cases (defragging, reusing, ...) 1867 * where we cannot backup. 1868 */ 1869 mtx_lock(&bqlock); 1870 nqindex = QUEUE_EMPTYKVA; 1871 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1872 1873 if (nbp == NULL) { 1874 /* 1875 * If no EMPTYKVA buffers and we are either 1876 * defragging or reusing, locate a CLEAN buffer 1877 * to free or reuse. If bufspace useage is low 1878 * skip this step so we can allocate a new buffer. 1879 */ 1880 if (defrag || bufspace >= lobufspace) { 1881 nqindex = QUEUE_CLEAN; 1882 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1883 } 1884 1885 /* 1886 * If we could not find or were not allowed to reuse a 1887 * CLEAN buffer, check to see if it is ok to use an EMPTY 1888 * buffer. We can only use an EMPTY buffer if allocating 1889 * its KVA would not otherwise run us out of buffer space. 1890 */ 1891 if (nbp == NULL && defrag == 0 && 1892 bufspace + maxsize < hibufspace) { 1893 nqindex = QUEUE_EMPTY; 1894 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1895 } 1896 } 1897 1898 /* 1899 * Run scan, possibly freeing data and/or kva mappings on the fly 1900 * depending. 1901 */ 1902 1903 while ((bp = nbp) != NULL) { 1904 int qindex = nqindex; 1905 1906 /* 1907 * Calculate next bp ( we can only use it if we do not block 1908 * or do other fancy things ). 1909 */ 1910 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1911 switch(qindex) { 1912 case QUEUE_EMPTY: 1913 nqindex = QUEUE_EMPTYKVA; 1914 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1915 break; 1916 /* FALLTHROUGH */ 1917 case QUEUE_EMPTYKVA: 1918 nqindex = QUEUE_CLEAN; 1919 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1920 break; 1921 /* FALLTHROUGH */ 1922 case QUEUE_CLEAN: 1923 /* 1924 * nbp is NULL. 1925 */ 1926 break; 1927 } 1928 } 1929 /* 1930 * If we are defragging then we need a buffer with 1931 * b_kvasize != 0. XXX this situation should no longer 1932 * occur, if defrag is non-zero the buffer's b_kvasize 1933 * should also be non-zero at this point. XXX 1934 */ 1935 if (defrag && bp->b_kvasize == 0) { 1936 printf("Warning: defrag empty buffer %p\n", bp); 1937 continue; 1938 } 1939 1940 /* 1941 * Start freeing the bp. This is somewhat involved. nbp 1942 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1943 */ 1944 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1945 continue; 1946 if (bp->b_vp) { 1947 BO_LOCK(bp->b_bufobj); 1948 if (bp->b_vflags & BV_BKGRDINPROG) { 1949 BO_UNLOCK(bp->b_bufobj); 1950 BUF_UNLOCK(bp); 1951 continue; 1952 } 1953 BO_UNLOCK(bp->b_bufobj); 1954 } 1955 CTR6(KTR_BUF, 1956 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1957 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1958 bp->b_kvasize, bp->b_bufsize, qindex); 1959 1960 /* 1961 * Sanity Checks 1962 */ 1963 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1964 1965 /* 1966 * Note: we no longer distinguish between VMIO and non-VMIO 1967 * buffers. 1968 */ 1969 1970 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1971 1972 if (bp->b_bufobj != NULL) 1973 BO_LOCK(bp->b_bufobj); 1974 bremfreel(bp); 1975 if (bp->b_bufobj != NULL) 1976 BO_UNLOCK(bp->b_bufobj); 1977 mtx_unlock(&bqlock); 1978 1979 if (qindex == QUEUE_CLEAN) { 1980 if (bp->b_flags & B_VMIO) { 1981 bp->b_flags &= ~B_ASYNC; 1982 vfs_vmio_release(bp); 1983 } 1984 if (bp->b_vp) 1985 brelvp(bp); 1986 } 1987 1988 /* 1989 * NOTE: nbp is now entirely invalid. We can only restart 1990 * the scan from this point on. 1991 * 1992 * Get the rest of the buffer freed up. b_kva* is still 1993 * valid after this operation. 1994 */ 1995 1996 if (bp->b_rcred != NOCRED) { 1997 crfree(bp->b_rcred); 1998 bp->b_rcred = NOCRED; 1999 } 2000 if (bp->b_wcred != NOCRED) { 2001 crfree(bp->b_wcred); 2002 bp->b_wcred = NOCRED; 2003 } 2004 if (!LIST_EMPTY(&bp->b_dep)) 2005 buf_deallocate(bp); 2006 if (bp->b_vflags & BV_BKGRDINPROG) 2007 panic("losing buffer 3"); 2008 KASSERT(bp->b_vp == NULL, 2009 ("bp: %p still has vnode %p. qindex: %d", 2010 bp, bp->b_vp, qindex)); 2011 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 2012 ("bp: %p still on a buffer list. xflags %X", 2013 bp, bp->b_xflags)); 2014 2015 if (bp->b_bufsize) 2016 allocbuf(bp, 0); 2017 2018 bp->b_flags = 0; 2019 bp->b_ioflags = 0; 2020 bp->b_xflags = 0; 2021 KASSERT((bp->b_vflags & BV_INFREECNT) == 0, 2022 ("buf %p still counted as free?", bp)); 2023 bp->b_vflags = 0; 2024 bp->b_vp = NULL; 2025 bp->b_blkno = bp->b_lblkno = 0; 2026 bp->b_offset = NOOFFSET; 2027 bp->b_iodone = 0; 2028 bp->b_error = 0; 2029 bp->b_resid = 0; 2030 bp->b_bcount = 0; 2031 bp->b_npages = 0; 2032 bp->b_dirtyoff = bp->b_dirtyend = 0; 2033 bp->b_bufobj = NULL; 2034 bp->b_pin_count = 0; 2035 bp->b_fsprivate1 = NULL; 2036 bp->b_fsprivate2 = NULL; 2037 bp->b_fsprivate3 = NULL; 2038 2039 LIST_INIT(&bp->b_dep); 2040 2041 /* 2042 * If we are defragging then free the buffer. 2043 */ 2044 if (defrag) { 2045 bp->b_flags |= B_INVAL; 2046 bfreekva(bp); 2047 brelse(bp); 2048 defrag = 0; 2049 goto restart; 2050 } 2051 2052 /* 2053 * Notify any waiters for the buffer lock about 2054 * identity change by freeing the buffer. 2055 */ 2056 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 2057 bp->b_flags |= B_INVAL; 2058 bfreekva(bp); 2059 brelse(bp); 2060 goto restart; 2061 } 2062 2063 /* 2064 * If we are overcomitted then recover the buffer and its 2065 * KVM space. This occurs in rare situations when multiple 2066 * processes are blocked in getnewbuf() or allocbuf(). 2067 */ 2068 if (bufspace >= hibufspace) 2069 flushingbufs = 1; 2070 if (flushingbufs && bp->b_kvasize != 0) { 2071 bp->b_flags |= B_INVAL; 2072 bfreekva(bp); 2073 brelse(bp); 2074 goto restart; 2075 } 2076 if (bufspace < lobufspace) 2077 flushingbufs = 0; 2078 break; 2079 } 2080 2081 /* 2082 * If we exhausted our list, sleep as appropriate. We may have to 2083 * wakeup various daemons and write out some dirty buffers. 2084 * 2085 * Generally we are sleeping due to insufficient buffer space. 2086 */ 2087 2088 if (bp == NULL) { 2089 int flags, norunbuf; 2090 char *waitmsg; 2091 int fl; 2092 2093 if (defrag) { 2094 flags = VFS_BIO_NEED_BUFSPACE; 2095 waitmsg = "nbufkv"; 2096 } else if (bufspace >= hibufspace) { 2097 waitmsg = "nbufbs"; 2098 flags = VFS_BIO_NEED_BUFSPACE; 2099 } else { 2100 waitmsg = "newbuf"; 2101 flags = VFS_BIO_NEED_ANY; 2102 } 2103 mtx_lock(&nblock); 2104 needsbuffer |= flags; 2105 mtx_unlock(&nblock); 2106 mtx_unlock(&bqlock); 2107 2108 bd_speedup(); /* heeeelp */ 2109 if (gbflags & GB_NOWAIT_BD) 2110 return (NULL); 2111 2112 mtx_lock(&nblock); 2113 while (needsbuffer & flags) { 2114 if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) { 2115 mtx_unlock(&nblock); 2116 /* 2117 * getblk() is called with a vnode 2118 * locked, and some majority of the 2119 * dirty buffers may as well belong to 2120 * the vnode. Flushing the buffers 2121 * there would make a progress that 2122 * cannot be achieved by the 2123 * buf_daemon, that cannot lock the 2124 * vnode. 2125 */ 2126 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2127 (td->td_pflags & TDP_NORUNNINGBUF); 2128 /* play bufdaemon */ 2129 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2130 fl = buf_do_flush(vp); 2131 td->td_pflags &= norunbuf; 2132 mtx_lock(&nblock); 2133 if (fl != 0) 2134 continue; 2135 if ((needsbuffer & flags) == 0) 2136 break; 2137 } 2138 if (msleep(&needsbuffer, &nblock, 2139 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2140 mtx_unlock(&nblock); 2141 return (NULL); 2142 } 2143 } 2144 mtx_unlock(&nblock); 2145 } else { 2146 /* 2147 * We finally have a valid bp. We aren't quite out of the 2148 * woods, we still have to reserve kva space. In order 2149 * to keep fragmentation sane we only allocate kva in 2150 * BKVASIZE chunks. 2151 */ 2152 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2153 2154 if (maxsize != bp->b_kvasize) { 2155 vm_offset_t addr = 0; 2156 int rv; 2157 2158 bfreekva(bp); 2159 2160 vm_map_lock(buffer_map); 2161 if (vm_map_findspace(buffer_map, 2162 vm_map_min(buffer_map), maxsize, &addr)) { 2163 /* 2164 * Buffer map is too fragmented. 2165 * We must defragment the map. 2166 */ 2167 atomic_add_int(&bufdefragcnt, 1); 2168 vm_map_unlock(buffer_map); 2169 defrag = 1; 2170 bp->b_flags |= B_INVAL; 2171 brelse(bp); 2172 goto restart; 2173 } 2174 rv = vm_map_insert(buffer_map, NULL, 0, addr, 2175 addr + maxsize, VM_PROT_ALL, VM_PROT_ALL, 2176 MAP_NOFAULT); 2177 KASSERT(rv == KERN_SUCCESS, 2178 ("vm_map_insert(buffer_map) rv %d", rv)); 2179 vm_map_unlock(buffer_map); 2180 bp->b_kvabase = (caddr_t)addr; 2181 bp->b_kvasize = maxsize; 2182 atomic_add_long(&bufspace, bp->b_kvasize); 2183 atomic_add_int(&bufreusecnt, 1); 2184 } 2185 bp->b_saveaddr = bp->b_kvabase; 2186 bp->b_data = bp->b_saveaddr; 2187 } 2188 return (bp); 2189} 2190 2191/* 2192 * buf_daemon: 2193 * 2194 * buffer flushing daemon. Buffers are normally flushed by the 2195 * update daemon but if it cannot keep up this process starts to 2196 * take the load in an attempt to prevent getnewbuf() from blocking. 2197 */ 2198 2199static struct kproc_desc buf_kp = { 2200 "bufdaemon", 2201 buf_daemon, 2202 &bufdaemonproc 2203}; 2204SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2205 2206static int 2207buf_do_flush(struct vnode *vp) 2208{ 2209 int flushed; 2210 2211 flushed = flushbufqueues(vp, QUEUE_DIRTY, 0); 2212 if (flushed == 0) { 2213 /* 2214 * Could not find any buffers without rollback 2215 * dependencies, so just write the first one 2216 * in the hopes of eventually making progress. 2217 */ 2218 flushbufqueues(vp, QUEUE_DIRTY, 1); 2219 } 2220 return (flushed); 2221} 2222 2223static void 2224buf_daemon() 2225{ 2226 int lodirtysave; 2227 2228 /* 2229 * This process needs to be suspended prior to shutdown sync. 2230 */ 2231 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2232 SHUTDOWN_PRI_LAST); 2233 2234 /* 2235 * This process is allowed to take the buffer cache to the limit 2236 */ 2237 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2238 mtx_lock(&bdlock); 2239 for (;;) { 2240 bd_request = 0; 2241 mtx_unlock(&bdlock); 2242 2243 kproc_suspend_check(bufdaemonproc); 2244 lodirtysave = lodirtybuffers; 2245 if (bd_speedupreq) { 2246 lodirtybuffers = numdirtybuffers / 2; 2247 bd_speedupreq = 0; 2248 } 2249 /* 2250 * Do the flush. Limit the amount of in-transit I/O we 2251 * allow to build up, otherwise we would completely saturate 2252 * the I/O system. Wakeup any waiting processes before we 2253 * normally would so they can run in parallel with our drain. 2254 */ 2255 while (numdirtybuffers > lodirtybuffers) { 2256 if (buf_do_flush(NULL) == 0) 2257 break; 2258 kern_yield(PRI_USER); 2259 } 2260 lodirtybuffers = lodirtysave; 2261 2262 /* 2263 * Only clear bd_request if we have reached our low water 2264 * mark. The buf_daemon normally waits 1 second and 2265 * then incrementally flushes any dirty buffers that have 2266 * built up, within reason. 2267 * 2268 * If we were unable to hit our low water mark and couldn't 2269 * find any flushable buffers, we sleep half a second. 2270 * Otherwise we loop immediately. 2271 */ 2272 mtx_lock(&bdlock); 2273 if (numdirtybuffers <= lodirtybuffers) { 2274 /* 2275 * We reached our low water mark, reset the 2276 * request and sleep until we are needed again. 2277 * The sleep is just so the suspend code works. 2278 */ 2279 bd_request = 0; 2280 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2281 } else { 2282 /* 2283 * We couldn't find any flushable dirty buffers but 2284 * still have too many dirty buffers, we 2285 * have to sleep and try again. (rare) 2286 */ 2287 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2288 } 2289 } 2290} 2291 2292/* 2293 * flushbufqueues: 2294 * 2295 * Try to flush a buffer in the dirty queue. We must be careful to 2296 * free up B_INVAL buffers instead of write them, which NFS is 2297 * particularly sensitive to. 2298 */ 2299static int flushwithdeps = 0; 2300SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2301 0, "Number of buffers flushed with dependecies that require rollbacks"); 2302 2303static int 2304flushbufqueues(struct vnode *lvp, int queue, int flushdeps) 2305{ 2306 struct buf *sentinel; 2307 struct vnode *vp; 2308 struct mount *mp; 2309 struct buf *bp; 2310 int hasdeps; 2311 int flushed; 2312 int target; 2313 2314 if (lvp == NULL) { 2315 target = numdirtybuffers - lodirtybuffers; 2316 if (flushdeps && target > 2) 2317 target /= 2; 2318 } else 2319 target = flushbufqtarget; 2320 flushed = 0; 2321 bp = NULL; 2322 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2323 sentinel->b_qindex = QUEUE_SENTINEL; 2324 mtx_lock(&bqlock); 2325 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2326 while (flushed != target) { 2327 bp = TAILQ_NEXT(sentinel, b_freelist); 2328 if (bp != NULL) { 2329 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2330 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2331 b_freelist); 2332 } else 2333 break; 2334 /* 2335 * Skip sentinels inserted by other invocations of the 2336 * flushbufqueues(), taking care to not reorder them. 2337 */ 2338 if (bp->b_qindex == QUEUE_SENTINEL) 2339 continue; 2340 /* 2341 * Only flush the buffers that belong to the 2342 * vnode locked by the curthread. 2343 */ 2344 if (lvp != NULL && bp->b_vp != lvp) 2345 continue; 2346 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2347 continue; 2348 if (bp->b_pin_count > 0) { 2349 BUF_UNLOCK(bp); 2350 continue; 2351 } 2352 BO_LOCK(bp->b_bufobj); 2353 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2354 (bp->b_flags & B_DELWRI) == 0) { 2355 BO_UNLOCK(bp->b_bufobj); 2356 BUF_UNLOCK(bp); 2357 continue; 2358 } 2359 BO_UNLOCK(bp->b_bufobj); 2360 if (bp->b_flags & B_INVAL) { 2361 bremfreel(bp); 2362 mtx_unlock(&bqlock); 2363 brelse(bp); 2364 flushed++; 2365 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2366 mtx_lock(&bqlock); 2367 continue; 2368 } 2369 2370 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2371 if (flushdeps == 0) { 2372 BUF_UNLOCK(bp); 2373 continue; 2374 } 2375 hasdeps = 1; 2376 } else 2377 hasdeps = 0; 2378 /* 2379 * We must hold the lock on a vnode before writing 2380 * one of its buffers. Otherwise we may confuse, or 2381 * in the case of a snapshot vnode, deadlock the 2382 * system. 2383 * 2384 * The lock order here is the reverse of the normal 2385 * of vnode followed by buf lock. This is ok because 2386 * the NOWAIT will prevent deadlock. 2387 */ 2388 vp = bp->b_vp; 2389 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2390 BUF_UNLOCK(bp); 2391 continue; 2392 } 2393 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) { 2394 mtx_unlock(&bqlock); 2395 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2396 bp, bp->b_vp, bp->b_flags); 2397 if (curproc == bufdaemonproc) 2398 vfs_bio_awrite(bp); 2399 else { 2400 bremfree(bp); 2401 bwrite(bp); 2402 notbufdflashes++; 2403 } 2404 vn_finished_write(mp); 2405 VOP_UNLOCK(vp, 0); 2406 flushwithdeps += hasdeps; 2407 flushed++; 2408 2409 /* 2410 * Sleeping on runningbufspace while holding 2411 * vnode lock leads to deadlock. 2412 */ 2413 if (curproc == bufdaemonproc) 2414 waitrunningbufspace(); 2415 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2416 mtx_lock(&bqlock); 2417 continue; 2418 } 2419 vn_finished_write(mp); 2420 BUF_UNLOCK(bp); 2421 } 2422 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2423 mtx_unlock(&bqlock); 2424 free(sentinel, M_TEMP); 2425 return (flushed); 2426} 2427 2428/* 2429 * Check to see if a block is currently memory resident. 2430 */ 2431struct buf * 2432incore(struct bufobj *bo, daddr_t blkno) 2433{ 2434 struct buf *bp; 2435 2436 BO_LOCK(bo); 2437 bp = gbincore(bo, blkno); 2438 BO_UNLOCK(bo); 2439 return (bp); 2440} 2441 2442/* 2443 * Returns true if no I/O is needed to access the 2444 * associated VM object. This is like incore except 2445 * it also hunts around in the VM system for the data. 2446 */ 2447 2448static int 2449inmem(struct vnode * vp, daddr_t blkno) 2450{ 2451 vm_object_t obj; 2452 vm_offset_t toff, tinc, size; 2453 vm_page_t m; 2454 vm_ooffset_t off; 2455 2456 ASSERT_VOP_LOCKED(vp, "inmem"); 2457 2458 if (incore(&vp->v_bufobj, blkno)) 2459 return 1; 2460 if (vp->v_mount == NULL) 2461 return 0; 2462 obj = vp->v_object; 2463 if (obj == NULL) 2464 return (0); 2465 2466 size = PAGE_SIZE; 2467 if (size > vp->v_mount->mnt_stat.f_iosize) 2468 size = vp->v_mount->mnt_stat.f_iosize; 2469 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2470 2471 VM_OBJECT_WLOCK(obj); 2472 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2473 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2474 if (!m) 2475 goto notinmem; 2476 tinc = size; 2477 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2478 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2479 if (vm_page_is_valid(m, 2480 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2481 goto notinmem; 2482 } 2483 VM_OBJECT_WUNLOCK(obj); 2484 return 1; 2485 2486notinmem: 2487 VM_OBJECT_WUNLOCK(obj); 2488 return (0); 2489} 2490 2491/* 2492 * Set the dirty range for a buffer based on the status of the dirty 2493 * bits in the pages comprising the buffer. The range is limited 2494 * to the size of the buffer. 2495 * 2496 * Tell the VM system that the pages associated with this buffer 2497 * are clean. This is used for delayed writes where the data is 2498 * going to go to disk eventually without additional VM intevention. 2499 * 2500 * Note that while we only really need to clean through to b_bcount, we 2501 * just go ahead and clean through to b_bufsize. 2502 */ 2503static void 2504vfs_clean_pages_dirty_buf(struct buf *bp) 2505{ 2506 vm_ooffset_t foff, noff, eoff; 2507 vm_page_t m; 2508 int i; 2509 2510 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 2511 return; 2512 2513 foff = bp->b_offset; 2514 KASSERT(bp->b_offset != NOOFFSET, 2515 ("vfs_clean_pages_dirty_buf: no buffer offset")); 2516 2517 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 2518 vfs_drain_busy_pages(bp); 2519 vfs_setdirty_locked_object(bp); 2520 for (i = 0; i < bp->b_npages; i++) { 2521 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2522 eoff = noff; 2523 if (eoff > bp->b_offset + bp->b_bufsize) 2524 eoff = bp->b_offset + bp->b_bufsize; 2525 m = bp->b_pages[i]; 2526 vfs_page_set_validclean(bp, foff, m); 2527 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2528 foff = noff; 2529 } 2530 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 2531} 2532 2533static void 2534vfs_setdirty_locked_object(struct buf *bp) 2535{ 2536 vm_object_t object; 2537 int i; 2538 2539 object = bp->b_bufobj->bo_object; 2540 VM_OBJECT_ASSERT_WLOCKED(object); 2541 2542 /* 2543 * We qualify the scan for modified pages on whether the 2544 * object has been flushed yet. 2545 */ 2546 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { 2547 vm_offset_t boffset; 2548 vm_offset_t eoffset; 2549 2550 /* 2551 * test the pages to see if they have been modified directly 2552 * by users through the VM system. 2553 */ 2554 for (i = 0; i < bp->b_npages; i++) 2555 vm_page_test_dirty(bp->b_pages[i]); 2556 2557 /* 2558 * Calculate the encompassing dirty range, boffset and eoffset, 2559 * (eoffset - boffset) bytes. 2560 */ 2561 2562 for (i = 0; i < bp->b_npages; i++) { 2563 if (bp->b_pages[i]->dirty) 2564 break; 2565 } 2566 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2567 2568 for (i = bp->b_npages - 1; i >= 0; --i) { 2569 if (bp->b_pages[i]->dirty) { 2570 break; 2571 } 2572 } 2573 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2574 2575 /* 2576 * Fit it to the buffer. 2577 */ 2578 2579 if (eoffset > bp->b_bcount) 2580 eoffset = bp->b_bcount; 2581 2582 /* 2583 * If we have a good dirty range, merge with the existing 2584 * dirty range. 2585 */ 2586 2587 if (boffset < eoffset) { 2588 if (bp->b_dirtyoff > boffset) 2589 bp->b_dirtyoff = boffset; 2590 if (bp->b_dirtyend < eoffset) 2591 bp->b_dirtyend = eoffset; 2592 } 2593 } 2594} 2595 2596/* 2597 * getblk: 2598 * 2599 * Get a block given a specified block and offset into a file/device. 2600 * The buffers B_DONE bit will be cleared on return, making it almost 2601 * ready for an I/O initiation. B_INVAL may or may not be set on 2602 * return. The caller should clear B_INVAL prior to initiating a 2603 * READ. 2604 * 2605 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2606 * an existing buffer. 2607 * 2608 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2609 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2610 * and then cleared based on the backing VM. If the previous buffer is 2611 * non-0-sized but invalid, B_CACHE will be cleared. 2612 * 2613 * If getblk() must create a new buffer, the new buffer is returned with 2614 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2615 * case it is returned with B_INVAL clear and B_CACHE set based on the 2616 * backing VM. 2617 * 2618 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2619 * B_CACHE bit is clear. 2620 * 2621 * What this means, basically, is that the caller should use B_CACHE to 2622 * determine whether the buffer is fully valid or not and should clear 2623 * B_INVAL prior to issuing a read. If the caller intends to validate 2624 * the buffer by loading its data area with something, the caller needs 2625 * to clear B_INVAL. If the caller does this without issuing an I/O, 2626 * the caller should set B_CACHE ( as an optimization ), else the caller 2627 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2628 * a write attempt or if it was a successfull read. If the caller 2629 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2630 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2631 */ 2632struct buf * 2633getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2634 int flags) 2635{ 2636 struct buf *bp; 2637 struct bufobj *bo; 2638 int error; 2639 2640 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2641 ASSERT_VOP_LOCKED(vp, "getblk"); 2642 if (size > MAXBSIZE) 2643 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2644 2645 bo = &vp->v_bufobj; 2646loop: 2647 /* 2648 * Block if we are low on buffers. Certain processes are allowed 2649 * to completely exhaust the buffer cache. 2650 * 2651 * If this check ever becomes a bottleneck it may be better to 2652 * move it into the else, when gbincore() fails. At the moment 2653 * it isn't a problem. 2654 */ 2655 if (numfreebuffers == 0) { 2656 if (TD_IS_IDLETHREAD(curthread)) 2657 return NULL; 2658 mtx_lock(&nblock); 2659 needsbuffer |= VFS_BIO_NEED_ANY; 2660 mtx_unlock(&nblock); 2661 } 2662 2663 BO_LOCK(bo); 2664 bp = gbincore(bo, blkno); 2665 if (bp != NULL) { 2666 int lockflags; 2667 /* 2668 * Buffer is in-core. If the buffer is not busy nor managed, 2669 * it must be on a queue. 2670 */ 2671 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2672 2673 if (flags & GB_LOCK_NOWAIT) 2674 lockflags |= LK_NOWAIT; 2675 2676 error = BUF_TIMELOCK(bp, lockflags, 2677 BO_MTX(bo), "getblk", slpflag, slptimeo); 2678 2679 /* 2680 * If we slept and got the lock we have to restart in case 2681 * the buffer changed identities. 2682 */ 2683 if (error == ENOLCK) 2684 goto loop; 2685 /* We timed out or were interrupted. */ 2686 else if (error) 2687 return (NULL); 2688 /* If recursed, assume caller knows the rules. */ 2689 else if (BUF_LOCKRECURSED(bp)) 2690 goto end; 2691 2692 /* 2693 * The buffer is locked. B_CACHE is cleared if the buffer is 2694 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2695 * and for a VMIO buffer B_CACHE is adjusted according to the 2696 * backing VM cache. 2697 */ 2698 if (bp->b_flags & B_INVAL) 2699 bp->b_flags &= ~B_CACHE; 2700 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2701 bp->b_flags |= B_CACHE; 2702 if (bp->b_flags & B_MANAGED) 2703 MPASS(bp->b_qindex == QUEUE_NONE); 2704 else { 2705 BO_LOCK(bo); 2706 bremfree(bp); 2707 BO_UNLOCK(bo); 2708 } 2709 2710 /* 2711 * check for size inconsistancies for non-VMIO case. 2712 */ 2713 2714 if (bp->b_bcount != size) { 2715 if ((bp->b_flags & B_VMIO) == 0 || 2716 (size > bp->b_kvasize)) { 2717 if (bp->b_flags & B_DELWRI) { 2718 /* 2719 * If buffer is pinned and caller does 2720 * not want sleep waiting for it to be 2721 * unpinned, bail out 2722 * */ 2723 if (bp->b_pin_count > 0) { 2724 if (flags & GB_LOCK_NOWAIT) { 2725 bqrelse(bp); 2726 return (NULL); 2727 } else { 2728 bunpin_wait(bp); 2729 } 2730 } 2731 bp->b_flags |= B_NOCACHE; 2732 bwrite(bp); 2733 } else { 2734 if (LIST_EMPTY(&bp->b_dep)) { 2735 bp->b_flags |= B_RELBUF; 2736 brelse(bp); 2737 } else { 2738 bp->b_flags |= B_NOCACHE; 2739 bwrite(bp); 2740 } 2741 } 2742 goto loop; 2743 } 2744 } 2745 2746 /* 2747 * If the size is inconsistant in the VMIO case, we can resize 2748 * the buffer. This might lead to B_CACHE getting set or 2749 * cleared. If the size has not changed, B_CACHE remains 2750 * unchanged from its previous state. 2751 */ 2752 2753 if (bp->b_bcount != size) 2754 allocbuf(bp, size); 2755 2756 KASSERT(bp->b_offset != NOOFFSET, 2757 ("getblk: no buffer offset")); 2758 2759 /* 2760 * A buffer with B_DELWRI set and B_CACHE clear must 2761 * be committed before we can return the buffer in 2762 * order to prevent the caller from issuing a read 2763 * ( due to B_CACHE not being set ) and overwriting 2764 * it. 2765 * 2766 * Most callers, including NFS and FFS, need this to 2767 * operate properly either because they assume they 2768 * can issue a read if B_CACHE is not set, or because 2769 * ( for example ) an uncached B_DELWRI might loop due 2770 * to softupdates re-dirtying the buffer. In the latter 2771 * case, B_CACHE is set after the first write completes, 2772 * preventing further loops. 2773 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2774 * above while extending the buffer, we cannot allow the 2775 * buffer to remain with B_CACHE set after the write 2776 * completes or it will represent a corrupt state. To 2777 * deal with this we set B_NOCACHE to scrap the buffer 2778 * after the write. 2779 * 2780 * We might be able to do something fancy, like setting 2781 * B_CACHE in bwrite() except if B_DELWRI is already set, 2782 * so the below call doesn't set B_CACHE, but that gets real 2783 * confusing. This is much easier. 2784 */ 2785 2786 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2787 bp->b_flags |= B_NOCACHE; 2788 bwrite(bp); 2789 goto loop; 2790 } 2791 bp->b_flags &= ~B_DONE; 2792 } else { 2793 int bsize, maxsize, vmio; 2794 off_t offset; 2795 2796 /* 2797 * Buffer is not in-core, create new buffer. The buffer 2798 * returned by getnewbuf() is locked. Note that the returned 2799 * buffer is also considered valid (not marked B_INVAL). 2800 */ 2801 BO_UNLOCK(bo); 2802 /* 2803 * If the user does not want us to create the buffer, bail out 2804 * here. 2805 */ 2806 if (flags & GB_NOCREAT) 2807 return NULL; 2808 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 2809 offset = blkno * bsize; 2810 vmio = vp->v_object != NULL; 2811 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2812 maxsize = imax(maxsize, bsize); 2813 2814 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 2815 if (bp == NULL) { 2816 if (slpflag || slptimeo) 2817 return NULL; 2818 goto loop; 2819 } 2820 2821 /* 2822 * This code is used to make sure that a buffer is not 2823 * created while the getnewbuf routine is blocked. 2824 * This can be a problem whether the vnode is locked or not. 2825 * If the buffer is created out from under us, we have to 2826 * throw away the one we just created. 2827 * 2828 * Note: this must occur before we associate the buffer 2829 * with the vp especially considering limitations in 2830 * the splay tree implementation when dealing with duplicate 2831 * lblkno's. 2832 */ 2833 BO_LOCK(bo); 2834 if (gbincore(bo, blkno)) { 2835 BO_UNLOCK(bo); 2836 bp->b_flags |= B_INVAL; 2837 brelse(bp); 2838 goto loop; 2839 } 2840 2841 /* 2842 * Insert the buffer into the hash, so that it can 2843 * be found by incore. 2844 */ 2845 bp->b_blkno = bp->b_lblkno = blkno; 2846 bp->b_offset = offset; 2847 bgetvp(vp, bp); 2848 BO_UNLOCK(bo); 2849 2850 /* 2851 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2852 * buffer size starts out as 0, B_CACHE will be set by 2853 * allocbuf() for the VMIO case prior to it testing the 2854 * backing store for validity. 2855 */ 2856 2857 if (vmio) { 2858 bp->b_flags |= B_VMIO; 2859 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2860 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2861 bp, vp->v_object, bp->b_bufobj->bo_object)); 2862 } else { 2863 bp->b_flags &= ~B_VMIO; 2864 KASSERT(bp->b_bufobj->bo_object == NULL, 2865 ("ARGH! has b_bufobj->bo_object %p %p\n", 2866 bp, bp->b_bufobj->bo_object)); 2867 } 2868 2869 allocbuf(bp, size); 2870 bp->b_flags &= ~B_DONE; 2871 } 2872 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2873 BUF_ASSERT_HELD(bp); 2874end: 2875 KASSERT(bp->b_bufobj == bo, 2876 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2877 return (bp); 2878} 2879 2880/* 2881 * Get an empty, disassociated buffer of given size. The buffer is initially 2882 * set to B_INVAL. 2883 */ 2884struct buf * 2885geteblk(int size, int flags) 2886{ 2887 struct buf *bp; 2888 int maxsize; 2889 2890 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2891 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 2892 if ((flags & GB_NOWAIT_BD) && 2893 (curthread->td_pflags & TDP_BUFNEED) != 0) 2894 return (NULL); 2895 } 2896 allocbuf(bp, size); 2897 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2898 BUF_ASSERT_HELD(bp); 2899 return (bp); 2900} 2901 2902 2903/* 2904 * This code constitutes the buffer memory from either anonymous system 2905 * memory (in the case of non-VMIO operations) or from an associated 2906 * VM object (in the case of VMIO operations). This code is able to 2907 * resize a buffer up or down. 2908 * 2909 * Note that this code is tricky, and has many complications to resolve 2910 * deadlock or inconsistant data situations. Tread lightly!!! 2911 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2912 * the caller. Calling this code willy nilly can result in the loss of data. 2913 * 2914 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2915 * B_CACHE for the non-VMIO case. 2916 */ 2917 2918int 2919allocbuf(struct buf *bp, int size) 2920{ 2921 int newbsize, mbsize; 2922 int i; 2923 2924 BUF_ASSERT_HELD(bp); 2925 2926 if (bp->b_kvasize < size) 2927 panic("allocbuf: buffer too small"); 2928 2929 if ((bp->b_flags & B_VMIO) == 0) { 2930 caddr_t origbuf; 2931 int origbufsize; 2932 /* 2933 * Just get anonymous memory from the kernel. Don't 2934 * mess with B_CACHE. 2935 */ 2936 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2937 if (bp->b_flags & B_MALLOC) 2938 newbsize = mbsize; 2939 else 2940 newbsize = round_page(size); 2941 2942 if (newbsize < bp->b_bufsize) { 2943 /* 2944 * malloced buffers are not shrunk 2945 */ 2946 if (bp->b_flags & B_MALLOC) { 2947 if (newbsize) { 2948 bp->b_bcount = size; 2949 } else { 2950 free(bp->b_data, M_BIOBUF); 2951 if (bp->b_bufsize) { 2952 atomic_subtract_long( 2953 &bufmallocspace, 2954 bp->b_bufsize); 2955 bufspacewakeup(); 2956 bp->b_bufsize = 0; 2957 } 2958 bp->b_saveaddr = bp->b_kvabase; 2959 bp->b_data = bp->b_saveaddr; 2960 bp->b_bcount = 0; 2961 bp->b_flags &= ~B_MALLOC; 2962 } 2963 return 1; 2964 } 2965 vm_hold_free_pages(bp, newbsize); 2966 } else if (newbsize > bp->b_bufsize) { 2967 /* 2968 * We only use malloced memory on the first allocation. 2969 * and revert to page-allocated memory when the buffer 2970 * grows. 2971 */ 2972 /* 2973 * There is a potential smp race here that could lead 2974 * to bufmallocspace slightly passing the max. It 2975 * is probably extremely rare and not worth worrying 2976 * over. 2977 */ 2978 if ( (bufmallocspace < maxbufmallocspace) && 2979 (bp->b_bufsize == 0) && 2980 (mbsize <= PAGE_SIZE/2)) { 2981 2982 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2983 bp->b_bufsize = mbsize; 2984 bp->b_bcount = size; 2985 bp->b_flags |= B_MALLOC; 2986 atomic_add_long(&bufmallocspace, mbsize); 2987 return 1; 2988 } 2989 origbuf = NULL; 2990 origbufsize = 0; 2991 /* 2992 * If the buffer is growing on its other-than-first allocation, 2993 * then we revert to the page-allocation scheme. 2994 */ 2995 if (bp->b_flags & B_MALLOC) { 2996 origbuf = bp->b_data; 2997 origbufsize = bp->b_bufsize; 2998 bp->b_data = bp->b_kvabase; 2999 if (bp->b_bufsize) { 3000 atomic_subtract_long(&bufmallocspace, 3001 bp->b_bufsize); 3002 bufspacewakeup(); 3003 bp->b_bufsize = 0; 3004 } 3005 bp->b_flags &= ~B_MALLOC; 3006 newbsize = round_page(newbsize); 3007 } 3008 vm_hold_load_pages( 3009 bp, 3010 (vm_offset_t) bp->b_data + bp->b_bufsize, 3011 (vm_offset_t) bp->b_data + newbsize); 3012 if (origbuf) { 3013 bcopy(origbuf, bp->b_data, origbufsize); 3014 free(origbuf, M_BIOBUF); 3015 } 3016 } 3017 } else { 3018 int desiredpages; 3019 3020 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3021 desiredpages = (size == 0) ? 0 : 3022 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 3023 3024 if (bp->b_flags & B_MALLOC) 3025 panic("allocbuf: VMIO buffer can't be malloced"); 3026 /* 3027 * Set B_CACHE initially if buffer is 0 length or will become 3028 * 0-length. 3029 */ 3030 if (size == 0 || bp->b_bufsize == 0) 3031 bp->b_flags |= B_CACHE; 3032 3033 if (newbsize < bp->b_bufsize) { 3034 /* 3035 * DEV_BSIZE aligned new buffer size is less then the 3036 * DEV_BSIZE aligned existing buffer size. Figure out 3037 * if we have to remove any pages. 3038 */ 3039 if (desiredpages < bp->b_npages) { 3040 vm_page_t m; 3041 3042 pmap_qremove((vm_offset_t)trunc_page( 3043 (vm_offset_t)bp->b_data) + 3044 (desiredpages << PAGE_SHIFT), 3045 (bp->b_npages - desiredpages)); 3046 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3047 for (i = desiredpages; i < bp->b_npages; i++) { 3048 /* 3049 * the page is not freed here -- it 3050 * is the responsibility of 3051 * vnode_pager_setsize 3052 */ 3053 m = bp->b_pages[i]; 3054 KASSERT(m != bogus_page, 3055 ("allocbuf: bogus page found")); 3056 while (vm_page_sleep_if_busy(m, TRUE, 3057 "biodep")) 3058 continue; 3059 3060 bp->b_pages[i] = NULL; 3061 vm_page_lock(m); 3062 vm_page_unwire(m, 0); 3063 vm_page_unlock(m); 3064 } 3065 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3066 bp->b_npages = desiredpages; 3067 } 3068 } else if (size > bp->b_bcount) { 3069 /* 3070 * We are growing the buffer, possibly in a 3071 * byte-granular fashion. 3072 */ 3073 vm_object_t obj; 3074 vm_offset_t toff; 3075 vm_offset_t tinc; 3076 3077 /* 3078 * Step 1, bring in the VM pages from the object, 3079 * allocating them if necessary. We must clear 3080 * B_CACHE if these pages are not valid for the 3081 * range covered by the buffer. 3082 */ 3083 3084 obj = bp->b_bufobj->bo_object; 3085 3086 VM_OBJECT_WLOCK(obj); 3087 while (bp->b_npages < desiredpages) { 3088 vm_page_t m; 3089 3090 /* 3091 * We must allocate system pages since blocking 3092 * here could interfere with paging I/O, no 3093 * matter which process we are. 3094 * 3095 * We can only test VPO_BUSY here. Blocking on 3096 * m->busy might lead to a deadlock: 3097 * vm_fault->getpages->cluster_read->allocbuf 3098 * Thus, we specify VM_ALLOC_IGN_SBUSY. 3099 */ 3100 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + 3101 bp->b_npages, VM_ALLOC_NOBUSY | 3102 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3103 VM_ALLOC_RETRY | VM_ALLOC_IGN_SBUSY | 3104 VM_ALLOC_COUNT(desiredpages - bp->b_npages)); 3105 if (m->valid == 0) 3106 bp->b_flags &= ~B_CACHE; 3107 bp->b_pages[bp->b_npages] = m; 3108 ++bp->b_npages; 3109 } 3110 3111 /* 3112 * Step 2. We've loaded the pages into the buffer, 3113 * we have to figure out if we can still have B_CACHE 3114 * set. Note that B_CACHE is set according to the 3115 * byte-granular range ( bcount and size ), new the 3116 * aligned range ( newbsize ). 3117 * 3118 * The VM test is against m->valid, which is DEV_BSIZE 3119 * aligned. Needless to say, the validity of the data 3120 * needs to also be DEV_BSIZE aligned. Note that this 3121 * fails with NFS if the server or some other client 3122 * extends the file's EOF. If our buffer is resized, 3123 * B_CACHE may remain set! XXX 3124 */ 3125 3126 toff = bp->b_bcount; 3127 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3128 3129 while ((bp->b_flags & B_CACHE) && toff < size) { 3130 vm_pindex_t pi; 3131 3132 if (tinc > (size - toff)) 3133 tinc = size - toff; 3134 3135 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3136 PAGE_SHIFT; 3137 3138 vfs_buf_test_cache( 3139 bp, 3140 bp->b_offset, 3141 toff, 3142 tinc, 3143 bp->b_pages[pi] 3144 ); 3145 toff += tinc; 3146 tinc = PAGE_SIZE; 3147 } 3148 VM_OBJECT_WUNLOCK(obj); 3149 3150 /* 3151 * Step 3, fixup the KVM pmap. Remember that 3152 * bp->b_data is relative to bp->b_offset, but 3153 * bp->b_offset may be offset into the first page. 3154 */ 3155 3156 bp->b_data = (caddr_t) 3157 trunc_page((vm_offset_t)bp->b_data); 3158 pmap_qenter( 3159 (vm_offset_t)bp->b_data, 3160 bp->b_pages, 3161 bp->b_npages 3162 ); 3163 3164 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3165 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 3166 } 3167 } 3168 if (newbsize < bp->b_bufsize) 3169 bufspacewakeup(); 3170 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3171 bp->b_bcount = size; /* requested buffer size */ 3172 return 1; 3173} 3174 3175void 3176biodone(struct bio *bp) 3177{ 3178 struct mtx *mtxp; 3179 void (*done)(struct bio *); 3180 3181 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3182 mtx_lock(mtxp); 3183 bp->bio_flags |= BIO_DONE; 3184 done = bp->bio_done; 3185 if (done == NULL) 3186 wakeup(bp); 3187 mtx_unlock(mtxp); 3188 if (done != NULL) 3189 done(bp); 3190} 3191 3192/* 3193 * Wait for a BIO to finish. 3194 * 3195 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3196 * case is not yet clear. 3197 */ 3198int 3199biowait(struct bio *bp, const char *wchan) 3200{ 3201 struct mtx *mtxp; 3202 3203 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3204 mtx_lock(mtxp); 3205 while ((bp->bio_flags & BIO_DONE) == 0) 3206 msleep(bp, mtxp, PRIBIO, wchan, hz / 10); 3207 mtx_unlock(mtxp); 3208 if (bp->bio_error != 0) 3209 return (bp->bio_error); 3210 if (!(bp->bio_flags & BIO_ERROR)) 3211 return (0); 3212 return (EIO); 3213} 3214 3215void 3216biofinish(struct bio *bp, struct devstat *stat, int error) 3217{ 3218 3219 if (error) { 3220 bp->bio_error = error; 3221 bp->bio_flags |= BIO_ERROR; 3222 } 3223 if (stat != NULL) 3224 devstat_end_transaction_bio(stat, bp); 3225 biodone(bp); 3226} 3227 3228/* 3229 * bufwait: 3230 * 3231 * Wait for buffer I/O completion, returning error status. The buffer 3232 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3233 * error and cleared. 3234 */ 3235int 3236bufwait(struct buf *bp) 3237{ 3238 if (bp->b_iocmd == BIO_READ) 3239 bwait(bp, PRIBIO, "biord"); 3240 else 3241 bwait(bp, PRIBIO, "biowr"); 3242 if (bp->b_flags & B_EINTR) { 3243 bp->b_flags &= ~B_EINTR; 3244 return (EINTR); 3245 } 3246 if (bp->b_ioflags & BIO_ERROR) { 3247 return (bp->b_error ? bp->b_error : EIO); 3248 } else { 3249 return (0); 3250 } 3251} 3252 3253 /* 3254 * Call back function from struct bio back up to struct buf. 3255 */ 3256static void 3257bufdonebio(struct bio *bip) 3258{ 3259 struct buf *bp; 3260 3261 bp = bip->bio_caller2; 3262 bp->b_resid = bp->b_bcount - bip->bio_completed; 3263 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3264 bp->b_ioflags = bip->bio_flags; 3265 bp->b_error = bip->bio_error; 3266 if (bp->b_error) 3267 bp->b_ioflags |= BIO_ERROR; 3268 bufdone(bp); 3269 g_destroy_bio(bip); 3270} 3271 3272void 3273dev_strategy(struct cdev *dev, struct buf *bp) 3274{ 3275 struct cdevsw *csw; 3276 struct bio *bip; 3277 int ref; 3278 3279 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3280 panic("b_iocmd botch"); 3281 for (;;) { 3282 bip = g_new_bio(); 3283 if (bip != NULL) 3284 break; 3285 /* Try again later */ 3286 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3287 } 3288 bip->bio_cmd = bp->b_iocmd; 3289 bip->bio_offset = bp->b_iooffset; 3290 bip->bio_length = bp->b_bcount; 3291 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3292 bip->bio_data = bp->b_data; 3293 bip->bio_done = bufdonebio; 3294 bip->bio_caller2 = bp; 3295 bip->bio_dev = dev; 3296 KASSERT(dev->si_refcount > 0, 3297 ("dev_strategy on un-referenced struct cdev *(%s)", 3298 devtoname(dev))); 3299 csw = dev_refthread(dev, &ref); 3300 if (csw == NULL) { 3301 g_destroy_bio(bip); 3302 bp->b_error = ENXIO; 3303 bp->b_ioflags = BIO_ERROR; 3304 bufdone(bp); 3305 return; 3306 } 3307 (*csw->d_strategy)(bip); 3308 dev_relthread(dev, ref); 3309} 3310 3311/* 3312 * bufdone: 3313 * 3314 * Finish I/O on a buffer, optionally calling a completion function. 3315 * This is usually called from an interrupt so process blocking is 3316 * not allowed. 3317 * 3318 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3319 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3320 * assuming B_INVAL is clear. 3321 * 3322 * For the VMIO case, we set B_CACHE if the op was a read and no 3323 * read error occured, or if the op was a write. B_CACHE is never 3324 * set if the buffer is invalid or otherwise uncacheable. 3325 * 3326 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3327 * initiator to leave B_INVAL set to brelse the buffer out of existance 3328 * in the biodone routine. 3329 */ 3330void 3331bufdone(struct buf *bp) 3332{ 3333 struct bufobj *dropobj; 3334 void (*biodone)(struct buf *); 3335 3336 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3337 dropobj = NULL; 3338 3339 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3340 BUF_ASSERT_HELD(bp); 3341 3342 runningbufwakeup(bp); 3343 if (bp->b_iocmd == BIO_WRITE) 3344 dropobj = bp->b_bufobj; 3345 /* call optional completion function if requested */ 3346 if (bp->b_iodone != NULL) { 3347 biodone = bp->b_iodone; 3348 bp->b_iodone = NULL; 3349 (*biodone) (bp); 3350 if (dropobj) 3351 bufobj_wdrop(dropobj); 3352 return; 3353 } 3354 3355 bufdone_finish(bp); 3356 3357 if (dropobj) 3358 bufobj_wdrop(dropobj); 3359} 3360 3361void 3362bufdone_finish(struct buf *bp) 3363{ 3364 BUF_ASSERT_HELD(bp); 3365 3366 if (!LIST_EMPTY(&bp->b_dep)) 3367 buf_complete(bp); 3368 3369 if (bp->b_flags & B_VMIO) { 3370 vm_ooffset_t foff; 3371 vm_page_t m; 3372 vm_object_t obj; 3373 struct vnode *vp; 3374 int bogus, i, iosize; 3375 3376 obj = bp->b_bufobj->bo_object; 3377 KASSERT(obj->paging_in_progress >= bp->b_npages, 3378 ("biodone_finish: paging in progress(%d) < b_npages(%d)", 3379 obj->paging_in_progress, bp->b_npages)); 3380 3381 vp = bp->b_vp; 3382 KASSERT(vp->v_holdcnt > 0, 3383 ("biodone_finish: vnode %p has zero hold count", vp)); 3384 KASSERT(vp->v_object != NULL, 3385 ("biodone_finish: vnode %p has no vm_object", vp)); 3386 3387 foff = bp->b_offset; 3388 KASSERT(bp->b_offset != NOOFFSET, 3389 ("biodone_finish: bp %p has no buffer offset", bp)); 3390 3391 /* 3392 * Set B_CACHE if the op was a normal read and no error 3393 * occured. B_CACHE is set for writes in the b*write() 3394 * routines. 3395 */ 3396 iosize = bp->b_bcount - bp->b_resid; 3397 if (bp->b_iocmd == BIO_READ && 3398 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3399 !(bp->b_ioflags & BIO_ERROR)) { 3400 bp->b_flags |= B_CACHE; 3401 } 3402 bogus = 0; 3403 VM_OBJECT_WLOCK(obj); 3404 for (i = 0; i < bp->b_npages; i++) { 3405 int bogusflag = 0; 3406 int resid; 3407 3408 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3409 if (resid > iosize) 3410 resid = iosize; 3411 3412 /* 3413 * cleanup bogus pages, restoring the originals 3414 */ 3415 m = bp->b_pages[i]; 3416 if (m == bogus_page) { 3417 bogus = bogusflag = 1; 3418 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3419 if (m == NULL) 3420 panic("biodone: page disappeared!"); 3421 bp->b_pages[i] = m; 3422 } 3423 KASSERT(OFF_TO_IDX(foff) == m->pindex, 3424 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch", 3425 (intmax_t)foff, (uintmax_t)m->pindex)); 3426 3427 /* 3428 * In the write case, the valid and clean bits are 3429 * already changed correctly ( see bdwrite() ), so we 3430 * only need to do this here in the read case. 3431 */ 3432 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3433 KASSERT((m->dirty & vm_page_bits(foff & 3434 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3435 " page %p has unexpected dirty bits", m)); 3436 vfs_page_set_valid(bp, foff, m); 3437 } 3438 3439 vm_page_io_finish(m); 3440 vm_object_pip_subtract(obj, 1); 3441 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3442 iosize -= resid; 3443 } 3444 vm_object_pip_wakeupn(obj, 0); 3445 VM_OBJECT_WUNLOCK(obj); 3446 if (bogus) 3447 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3448 bp->b_pages, bp->b_npages); 3449 } 3450 3451 /* 3452 * For asynchronous completions, release the buffer now. The brelse 3453 * will do a wakeup there if necessary - so no need to do a wakeup 3454 * here in the async case. The sync case always needs to do a wakeup. 3455 */ 3456 3457 if (bp->b_flags & B_ASYNC) { 3458 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3459 brelse(bp); 3460 else 3461 bqrelse(bp); 3462 } else 3463 bdone(bp); 3464} 3465 3466/* 3467 * This routine is called in lieu of iodone in the case of 3468 * incomplete I/O. This keeps the busy status for pages 3469 * consistant. 3470 */ 3471void 3472vfs_unbusy_pages(struct buf *bp) 3473{ 3474 int i; 3475 vm_object_t obj; 3476 vm_page_t m; 3477 3478 runningbufwakeup(bp); 3479 if (!(bp->b_flags & B_VMIO)) 3480 return; 3481 3482 obj = bp->b_bufobj->bo_object; 3483 VM_OBJECT_WLOCK(obj); 3484 for (i = 0; i < bp->b_npages; i++) { 3485 m = bp->b_pages[i]; 3486 if (m == bogus_page) { 3487 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3488 if (!m) 3489 panic("vfs_unbusy_pages: page missing\n"); 3490 bp->b_pages[i] = m; 3491 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3492 bp->b_pages, bp->b_npages); 3493 } 3494 vm_object_pip_subtract(obj, 1); 3495 vm_page_io_finish(m); 3496 } 3497 vm_object_pip_wakeupn(obj, 0); 3498 VM_OBJECT_WUNLOCK(obj); 3499} 3500 3501/* 3502 * vfs_page_set_valid: 3503 * 3504 * Set the valid bits in a page based on the supplied offset. The 3505 * range is restricted to the buffer's size. 3506 * 3507 * This routine is typically called after a read completes. 3508 */ 3509static void 3510vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3511{ 3512 vm_ooffset_t eoff; 3513 3514 /* 3515 * Compute the end offset, eoff, such that [off, eoff) does not span a 3516 * page boundary and eoff is not greater than the end of the buffer. 3517 * The end of the buffer, in this case, is our file EOF, not the 3518 * allocation size of the buffer. 3519 */ 3520 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3521 if (eoff > bp->b_offset + bp->b_bcount) 3522 eoff = bp->b_offset + bp->b_bcount; 3523 3524 /* 3525 * Set valid range. This is typically the entire buffer and thus the 3526 * entire page. 3527 */ 3528 if (eoff > off) 3529 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 3530} 3531 3532/* 3533 * vfs_page_set_validclean: 3534 * 3535 * Set the valid bits and clear the dirty bits in a page based on the 3536 * supplied offset. The range is restricted to the buffer's size. 3537 */ 3538static void 3539vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3540{ 3541 vm_ooffset_t soff, eoff; 3542 3543 /* 3544 * Start and end offsets in buffer. eoff - soff may not cross a 3545 * page boundry or cross the end of the buffer. The end of the 3546 * buffer, in this case, is our file EOF, not the allocation size 3547 * of the buffer. 3548 */ 3549 soff = off; 3550 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3551 if (eoff > bp->b_offset + bp->b_bcount) 3552 eoff = bp->b_offset + bp->b_bcount; 3553 3554 /* 3555 * Set valid range. This is typically the entire buffer and thus the 3556 * entire page. 3557 */ 3558 if (eoff > soff) { 3559 vm_page_set_validclean( 3560 m, 3561 (vm_offset_t) (soff & PAGE_MASK), 3562 (vm_offset_t) (eoff - soff) 3563 ); 3564 } 3565} 3566 3567/* 3568 * Ensure that all buffer pages are not busied by VPO_BUSY flag. If 3569 * any page is busy, drain the flag. 3570 */ 3571static void 3572vfs_drain_busy_pages(struct buf *bp) 3573{ 3574 vm_page_t m; 3575 int i, last_busied; 3576 3577 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); 3578 last_busied = 0; 3579 for (i = 0; i < bp->b_npages; i++) { 3580 m = bp->b_pages[i]; 3581 if ((m->oflags & VPO_BUSY) != 0) { 3582 for (; last_busied < i; last_busied++) 3583 vm_page_busy(bp->b_pages[last_busied]); 3584 while ((m->oflags & VPO_BUSY) != 0) 3585 vm_page_sleep(m, "vbpage"); 3586 } 3587 } 3588 for (i = 0; i < last_busied; i++) 3589 vm_page_wakeup(bp->b_pages[i]); 3590} 3591 3592/* 3593 * This routine is called before a device strategy routine. 3594 * It is used to tell the VM system that paging I/O is in 3595 * progress, and treat the pages associated with the buffer 3596 * almost as being VPO_BUSY. Also the object paging_in_progress 3597 * flag is handled to make sure that the object doesn't become 3598 * inconsistant. 3599 * 3600 * Since I/O has not been initiated yet, certain buffer flags 3601 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3602 * and should be ignored. 3603 */ 3604void 3605vfs_busy_pages(struct buf *bp, int clear_modify) 3606{ 3607 int i, bogus; 3608 vm_object_t obj; 3609 vm_ooffset_t foff; 3610 vm_page_t m; 3611 3612 if (!(bp->b_flags & B_VMIO)) 3613 return; 3614 3615 obj = bp->b_bufobj->bo_object; 3616 foff = bp->b_offset; 3617 KASSERT(bp->b_offset != NOOFFSET, 3618 ("vfs_busy_pages: no buffer offset")); 3619 VM_OBJECT_WLOCK(obj); 3620 vfs_drain_busy_pages(bp); 3621 if (bp->b_bufsize != 0) 3622 vfs_setdirty_locked_object(bp); 3623 bogus = 0; 3624 for (i = 0; i < bp->b_npages; i++) { 3625 m = bp->b_pages[i]; 3626 3627 if ((bp->b_flags & B_CLUSTER) == 0) { 3628 vm_object_pip_add(obj, 1); 3629 vm_page_io_start(m); 3630 } 3631 /* 3632 * When readying a buffer for a read ( i.e 3633 * clear_modify == 0 ), it is important to do 3634 * bogus_page replacement for valid pages in 3635 * partially instantiated buffers. Partially 3636 * instantiated buffers can, in turn, occur when 3637 * reconstituting a buffer from its VM backing store 3638 * base. We only have to do this if B_CACHE is 3639 * clear ( which causes the I/O to occur in the 3640 * first place ). The replacement prevents the read 3641 * I/O from overwriting potentially dirty VM-backed 3642 * pages. XXX bogus page replacement is, uh, bogus. 3643 * It may not work properly with small-block devices. 3644 * We need to find a better way. 3645 */ 3646 if (clear_modify) { 3647 pmap_remove_write(m); 3648 vfs_page_set_validclean(bp, foff, m); 3649 } else if (m->valid == VM_PAGE_BITS_ALL && 3650 (bp->b_flags & B_CACHE) == 0) { 3651 bp->b_pages[i] = bogus_page; 3652 bogus++; 3653 } 3654 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3655 } 3656 VM_OBJECT_WUNLOCK(obj); 3657 if (bogus) 3658 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3659 bp->b_pages, bp->b_npages); 3660} 3661 3662/* 3663 * vfs_bio_set_valid: 3664 * 3665 * Set the range within the buffer to valid. The range is 3666 * relative to the beginning of the buffer, b_offset. Note that 3667 * b_offset itself may be offset from the beginning of the first 3668 * page. 3669 */ 3670void 3671vfs_bio_set_valid(struct buf *bp, int base, int size) 3672{ 3673 int i, n; 3674 vm_page_t m; 3675 3676 if (!(bp->b_flags & B_VMIO)) 3677 return; 3678 3679 /* 3680 * Fixup base to be relative to beginning of first page. 3681 * Set initial n to be the maximum number of bytes in the 3682 * first page that can be validated. 3683 */ 3684 base += (bp->b_offset & PAGE_MASK); 3685 n = PAGE_SIZE - (base & PAGE_MASK); 3686 3687 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3688 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3689 m = bp->b_pages[i]; 3690 if (n > size) 3691 n = size; 3692 vm_page_set_valid_range(m, base & PAGE_MASK, n); 3693 base += n; 3694 size -= n; 3695 n = PAGE_SIZE; 3696 } 3697 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3698} 3699 3700/* 3701 * vfs_bio_clrbuf: 3702 * 3703 * If the specified buffer is a non-VMIO buffer, clear the entire 3704 * buffer. If the specified buffer is a VMIO buffer, clear and 3705 * validate only the previously invalid portions of the buffer. 3706 * This routine essentially fakes an I/O, so we need to clear 3707 * BIO_ERROR and B_INVAL. 3708 * 3709 * Note that while we only theoretically need to clear through b_bcount, 3710 * we go ahead and clear through b_bufsize. 3711 */ 3712void 3713vfs_bio_clrbuf(struct buf *bp) 3714{ 3715 int i, j, mask; 3716 caddr_t sa, ea; 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 bzero(bp->b_data, bp->b_bufsize); 3735 bp->b_pages[0]->valid |= mask; 3736 goto unlock; 3737 } 3738 } 3739 ea = sa = bp->b_data; 3740 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3741 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3742 ea = (caddr_t)(vm_offset_t)ulmin( 3743 (u_long)(vm_offset_t)ea, 3744 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3745 if (bp->b_pages[i] == bogus_page) 3746 continue; 3747 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3748 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3749 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object); 3750 if ((bp->b_pages[i]->valid & mask) == mask) 3751 continue; 3752 if ((bp->b_pages[i]->valid & mask) == 0) 3753 bzero(sa, ea - sa); 3754 else { 3755 for (; sa < ea; sa += DEV_BSIZE, j++) { 3756 if ((bp->b_pages[i]->valid & (1 << j)) == 0) 3757 bzero(sa, DEV_BSIZE); 3758 } 3759 } 3760 bp->b_pages[i]->valid |= mask; 3761 } 3762unlock: 3763 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3764 bp->b_resid = 0; 3765} 3766 3767/* 3768 * vm_hold_load_pages and vm_hold_free_pages get pages into 3769 * a buffers address space. The pages are anonymous and are 3770 * not associated with a file object. 3771 */ 3772static void 3773vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3774{ 3775 vm_offset_t pg; 3776 vm_page_t p; 3777 int index; 3778 3779 to = round_page(to); 3780 from = round_page(from); 3781 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3782 3783 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3784tryagain: 3785 /* 3786 * note: must allocate system pages since blocking here 3787 * could interfere with paging I/O, no matter which 3788 * process we are. 3789 */ 3790 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | 3791 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT)); 3792 if (p == NULL) { 3793 VM_WAIT; 3794 goto tryagain; 3795 } 3796 pmap_qenter(pg, &p, 1); 3797 bp->b_pages[index] = p; 3798 } 3799 bp->b_npages = index; 3800} 3801 3802/* Return pages associated with this buf to the vm system */ 3803static void 3804vm_hold_free_pages(struct buf *bp, int newbsize) 3805{ 3806 vm_offset_t from; 3807 vm_page_t p; 3808 int index, newnpages; 3809 3810 from = round_page((vm_offset_t)bp->b_data + newbsize); 3811 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3812 if (bp->b_npages > newnpages) 3813 pmap_qremove(from, bp->b_npages - newnpages); 3814 for (index = newnpages; index < bp->b_npages; index++) { 3815 p = bp->b_pages[index]; 3816 bp->b_pages[index] = NULL; 3817 if (p->busy != 0) 3818 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3819 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); 3820 p->wire_count--; 3821 vm_page_free(p); 3822 atomic_subtract_int(&cnt.v_wire_count, 1); 3823 } 3824 bp->b_npages = newnpages; 3825} 3826 3827/* 3828 * Map an IO request into kernel virtual address space. 3829 * 3830 * All requests are (re)mapped into kernel VA space. 3831 * Notice that we use b_bufsize for the size of the buffer 3832 * to be mapped. b_bcount might be modified by the driver. 3833 * 3834 * Note that even if the caller determines that the address space should 3835 * be valid, a race or a smaller-file mapped into a larger space may 3836 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3837 * check the return value. 3838 */ 3839int 3840vmapbuf(struct buf *bp) 3841{ 3842 caddr_t kva; 3843 vm_prot_t prot; 3844 int pidx; 3845 3846 if (bp->b_bufsize < 0) 3847 return (-1); 3848 prot = VM_PROT_READ; 3849 if (bp->b_iocmd == BIO_READ) 3850 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3851 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 3852 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, 3853 btoc(MAXPHYS))) < 0) 3854 return (-1); 3855 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3856 3857 kva = bp->b_saveaddr; 3858 bp->b_npages = pidx; 3859 bp->b_saveaddr = bp->b_data; 3860 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3861 return(0); 3862} 3863 3864/* 3865 * Free the io map PTEs associated with this IO operation. 3866 * We also invalidate the TLB entries and restore the original b_addr. 3867 */ 3868void 3869vunmapbuf(struct buf *bp) 3870{ 3871 int npages; 3872 3873 npages = bp->b_npages; 3874 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3875 vm_page_unhold_pages(bp->b_pages, npages); 3876 3877 bp->b_data = bp->b_saveaddr; 3878} 3879 3880void 3881bdone(struct buf *bp) 3882{ 3883 struct mtx *mtxp; 3884 3885 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3886 mtx_lock(mtxp); 3887 bp->b_flags |= B_DONE; 3888 wakeup(bp); 3889 mtx_unlock(mtxp); 3890} 3891 3892void 3893bwait(struct buf *bp, u_char pri, const char *wchan) 3894{ 3895 struct mtx *mtxp; 3896 3897 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3898 mtx_lock(mtxp); 3899 while ((bp->b_flags & B_DONE) == 0) 3900 msleep(bp, mtxp, pri, wchan, 0); 3901 mtx_unlock(mtxp); 3902} 3903 3904int 3905bufsync(struct bufobj *bo, int waitfor) 3906{ 3907 3908 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 3909} 3910 3911void 3912bufstrategy(struct bufobj *bo, struct buf *bp) 3913{ 3914 int i = 0; 3915 struct vnode *vp; 3916 3917 vp = bp->b_vp; 3918 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3919 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3920 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3921 i = VOP_STRATEGY(vp, bp); 3922 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3923} 3924 3925void 3926bufobj_wrefl(struct bufobj *bo) 3927{ 3928 3929 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3930 ASSERT_BO_LOCKED(bo); 3931 bo->bo_numoutput++; 3932} 3933 3934void 3935bufobj_wref(struct bufobj *bo) 3936{ 3937 3938 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3939 BO_LOCK(bo); 3940 bo->bo_numoutput++; 3941 BO_UNLOCK(bo); 3942} 3943 3944void 3945bufobj_wdrop(struct bufobj *bo) 3946{ 3947 3948 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3949 BO_LOCK(bo); 3950 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3951 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3952 bo->bo_flag &= ~BO_WWAIT; 3953 wakeup(&bo->bo_numoutput); 3954 } 3955 BO_UNLOCK(bo); 3956} 3957 3958int 3959bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3960{ 3961 int error; 3962 3963 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3964 ASSERT_BO_LOCKED(bo); 3965 error = 0; 3966 while (bo->bo_numoutput) { 3967 bo->bo_flag |= BO_WWAIT; 3968 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3969 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3970 if (error) 3971 break; 3972 } 3973 return (error); 3974} 3975 3976void 3977bpin(struct buf *bp) 3978{ 3979 struct mtx *mtxp; 3980 3981 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3982 mtx_lock(mtxp); 3983 bp->b_pin_count++; 3984 mtx_unlock(mtxp); 3985} 3986 3987void 3988bunpin(struct buf *bp) 3989{ 3990 struct mtx *mtxp; 3991 3992 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3993 mtx_lock(mtxp); 3994 if (--bp->b_pin_count == 0) 3995 wakeup(bp); 3996 mtx_unlock(mtxp); 3997} 3998 3999void 4000bunpin_wait(struct buf *bp) 4001{ 4002 struct mtx *mtxp; 4003 4004 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4005 mtx_lock(mtxp); 4006 while (bp->b_pin_count > 0) 4007 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4008 mtx_unlock(mtxp); 4009} 4010 4011#include "opt_ddb.h" 4012#ifdef DDB 4013#include <ddb/ddb.h> 4014 4015/* DDB command to show buffer data */ 4016DB_SHOW_COMMAND(buffer, db_show_buffer) 4017{ 4018 /* get args */ 4019 struct buf *bp = (struct buf *)addr; 4020 4021 if (!have_addr) { 4022 db_printf("usage: show buffer <addr>\n"); 4023 return; 4024 } 4025 4026 db_printf("buf at %p\n", bp); 4027 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n", 4028 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, 4029 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS); 4030 db_printf( 4031 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4032 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, " 4033 "b_dep = %p\n", 4034 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4035 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4036 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first); 4037 if (bp->b_npages) { 4038 int i; 4039 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4040 for (i = 0; i < bp->b_npages; i++) { 4041 vm_page_t m; 4042 m = bp->b_pages[i]; 4043 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4044 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4045 if ((i + 1) < bp->b_npages) 4046 db_printf(","); 4047 } 4048 db_printf("\n"); 4049 } 4050 db_printf(" "); 4051 BUF_LOCKPRINTINFO(bp); 4052} 4053 4054DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4055{ 4056 struct buf *bp; 4057 int i; 4058 4059 for (i = 0; i < nbuf; i++) { 4060 bp = &buf[i]; 4061 if (BUF_ISLOCKED(bp)) { 4062 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4063 db_printf("\n"); 4064 } 4065 } 4066} 4067 4068DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4069{ 4070 struct vnode *vp; 4071 struct buf *bp; 4072 4073 if (!have_addr) { 4074 db_printf("usage: show vnodebufs <addr>\n"); 4075 return; 4076 } 4077 vp = (struct vnode *)addr; 4078 db_printf("Clean buffers:\n"); 4079 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4080 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4081 db_printf("\n"); 4082 } 4083 db_printf("Dirty buffers:\n"); 4084 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4085 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4086 db_printf("\n"); 4087 } 4088} 4089 4090DB_COMMAND(countfreebufs, db_coundfreebufs) 4091{ 4092 struct buf *bp; 4093 int i, used = 0, nfree = 0; 4094 4095 if (have_addr) { 4096 db_printf("usage: countfreebufs\n"); 4097 return; 4098 } 4099 4100 for (i = 0; i < nbuf; i++) { 4101 bp = &buf[i]; 4102 if ((bp->b_vflags & BV_INFREECNT) != 0) 4103 nfree++; 4104 else 4105 used++; 4106 } 4107 4108 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4109 nfree + used); 4110 db_printf("numfreebuffers is %d\n", numfreebuffers); 4111} 4112#endif /* DDB */ 4113