vfs_bio.c revision 144084
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 144084 2005-03-25 00:20:37Z jeff $"); 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/lock.h> 52#include <sys/malloc.h> 53#include <sys/mount.h> 54#include <sys/mutex.h> 55#include <sys/kernel.h> 56#include <sys/kthread.h> 57#include <sys/proc.h> 58#include <sys/resourcevar.h> 59#include <sys/sysctl.h> 60#include <sys/vmmeter.h> 61#include <sys/vnode.h> 62#include <geom/geom.h> 63#include <vm/vm.h> 64#include <vm/vm_param.h> 65#include <vm/vm_kern.h> 66#include <vm/vm_pageout.h> 67#include <vm/vm_page.h> 68#include <vm/vm_object.h> 69#include <vm/vm_extern.h> 70#include <vm/vm_map.h> 71#include "opt_directio.h" 72#include "opt_swap.h" 73 74static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 75 76struct bio_ops bioops; /* I/O operation notification */ 77 78struct buf_ops buf_ops_bio = { 79 .bop_name = "buf_ops_bio", 80 .bop_write = bufwrite, 81 .bop_strategy = bufstrategy, 82 .bop_sync = bufsync, 83}; 84 85/* 86 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 87 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 88 */ 89struct buf *buf; /* buffer header pool */ 90 91static struct proc *bufdaemonproc; 92 93static int inmem(struct vnode *vp, daddr_t blkno); 94static void vm_hold_free_pages(struct buf *bp, vm_offset_t from, 95 vm_offset_t to); 96static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 97 vm_offset_t to); 98static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 99 int pageno, vm_page_t m); 100static void vfs_clean_pages(struct buf *bp); 101static void vfs_setdirty(struct buf *bp); 102static void vfs_vmio_release(struct buf *bp); 103static int vfs_bio_clcheck(struct vnode *vp, int size, 104 daddr_t lblkno, daddr_t blkno); 105static int flushbufqueues(int flushdeps); 106static void buf_daemon(void); 107static void bremfreel(struct buf *bp); 108 109int vmiodirenable = TRUE; 110SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 111 "Use the VM system for directory writes"); 112int runningbufspace; 113SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 114 "Amount of presently outstanding async buffer io"); 115static int bufspace; 116SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 117 "KVA memory used for bufs"); 118static int maxbufspace; 119SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 120 "Maximum allowed value of bufspace (including buf_daemon)"); 121static int bufmallocspace; 122SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 123 "Amount of malloced memory for buffers"); 124static int maxbufmallocspace; 125SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 126 "Maximum amount of malloced memory for buffers"); 127static int lobufspace; 128SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 129 "Minimum amount of buffers we want to have"); 130static int hibufspace; 131SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 132 "Maximum allowed value of bufspace (excluding buf_daemon)"); 133static int bufreusecnt; 134SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 135 "Number of times we have reused a buffer"); 136static int buffreekvacnt; 137SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 138 "Number of times we have freed the KVA space from some buffer"); 139static int bufdefragcnt; 140SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 141 "Number of times we have had to repeat buffer allocation to defragment"); 142static int lorunningspace; 143SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 144 "Minimum preferred space used for in-progress I/O"); 145static int hirunningspace; 146SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 147 "Maximum amount of space to use for in-progress I/O"); 148static int dirtybufferflushes; 149SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 150 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 151static int altbufferflushes; 152SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 153 0, "Number of fsync flushes to limit dirty buffers"); 154static int recursiveflushes; 155SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 156 0, "Number of flushes skipped due to being recursive"); 157static int numdirtybuffers; 158SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 159 "Number of buffers that are dirty (has unwritten changes) at the moment"); 160static int lodirtybuffers; 161SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 162 "How many buffers we want to have free before bufdaemon can sleep"); 163static int hidirtybuffers; 164SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 165 "When the number of dirty buffers is considered severe"); 166static int dirtybufthresh; 167SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 168 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 169static int numfreebuffers; 170SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 171 "Number of free buffers"); 172static int lofreebuffers; 173SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 174 "XXX Unused"); 175static int hifreebuffers; 176SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 177 "XXX Complicatedly unused"); 178static int getnewbufcalls; 179SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 180 "Number of calls to getnewbuf"); 181static int getnewbufrestarts; 182SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 183 "Number of times getnewbuf has had to restart a buffer aquisition"); 184 185/* 186 * Wakeup point for bufdaemon, as well as indicator of whether it is already 187 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 188 * is idling. 189 */ 190static int bd_request; 191 192/* 193 * This lock synchronizes access to bd_request. 194 */ 195static struct mtx bdlock; 196 197/* 198 * bogus page -- for I/O to/from partially complete buffers 199 * this is a temporary solution to the problem, but it is not 200 * really that bad. it would be better to split the buffer 201 * for input in the case of buffers partially already in memory, 202 * but the code is intricate enough already. 203 */ 204vm_page_t bogus_page; 205 206/* 207 * Synchronization (sleep/wakeup) variable for active buffer space requests. 208 * Set when wait starts, cleared prior to wakeup(). 209 * Used in runningbufwakeup() and waitrunningbufspace(). 210 */ 211static int runningbufreq; 212 213/* 214 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 215 * waitrunningbufspace(). 216 */ 217static struct mtx rbreqlock; 218 219/* 220 * Synchronization (sleep/wakeup) variable for buffer requests. 221 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 222 * by and/or. 223 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 224 * getnewbuf(), and getblk(). 225 */ 226static int needsbuffer; 227 228/* 229 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 230 */ 231static struct mtx nblock; 232 233/* 234 * Lock that protects against bwait()/bdone()/B_DONE races. 235 */ 236 237static struct mtx bdonelock; 238 239/* 240 * Definitions for the buffer free lists. 241 */ 242#define BUFFER_QUEUES 5 /* number of free buffer queues */ 243 244#define QUEUE_NONE 0 /* on no queue */ 245#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 246#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 247#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 248#define QUEUE_EMPTY 4 /* empty buffer headers */ 249 250/* Queues for free buffers with various properties */ 251static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 252 253/* Lock for the bufqueues */ 254static struct mtx bqlock; 255 256/* 257 * Single global constant for BUF_WMESG, to avoid getting multiple references. 258 * buf_wmesg is referred from macros. 259 */ 260const char *buf_wmesg = BUF_WMESG; 261 262#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 263#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 264#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 265#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 266 267#ifdef DIRECTIO 268extern void ffs_rawread_setup(void); 269#endif /* DIRECTIO */ 270/* 271 * numdirtywakeup: 272 * 273 * If someone is blocked due to there being too many dirty buffers, 274 * and numdirtybuffers is now reasonable, wake them up. 275 */ 276 277static __inline void 278numdirtywakeup(int level) 279{ 280 281 if (numdirtybuffers <= level) { 282 mtx_lock(&nblock); 283 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 284 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 285 wakeup(&needsbuffer); 286 } 287 mtx_unlock(&nblock); 288 } 289} 290 291/* 292 * bufspacewakeup: 293 * 294 * Called when buffer space is potentially available for recovery. 295 * getnewbuf() will block on this flag when it is unable to free 296 * sufficient buffer space. Buffer space becomes recoverable when 297 * bp's get placed back in the queues. 298 */ 299 300static __inline void 301bufspacewakeup(void) 302{ 303 304 /* 305 * If someone is waiting for BUF space, wake them up. Even 306 * though we haven't freed the kva space yet, the waiting 307 * process will be able to now. 308 */ 309 mtx_lock(&nblock); 310 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 311 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 312 wakeup(&needsbuffer); 313 } 314 mtx_unlock(&nblock); 315} 316 317/* 318 * runningbufwakeup() - in-progress I/O accounting. 319 * 320 */ 321static __inline void 322runningbufwakeup(struct buf *bp) 323{ 324 325 if (bp->b_runningbufspace) { 326 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 327 bp->b_runningbufspace = 0; 328 mtx_lock(&rbreqlock); 329 if (runningbufreq && runningbufspace <= lorunningspace) { 330 runningbufreq = 0; 331 wakeup(&runningbufreq); 332 } 333 mtx_unlock(&rbreqlock); 334 } 335} 336 337/* 338 * bufcountwakeup: 339 * 340 * Called when a buffer has been added to one of the free queues to 341 * account for the buffer and to wakeup anyone waiting for free buffers. 342 * This typically occurs when large amounts of metadata are being handled 343 * by the buffer cache ( else buffer space runs out first, usually ). 344 */ 345 346static __inline void 347bufcountwakeup(void) 348{ 349 350 atomic_add_int(&numfreebuffers, 1); 351 mtx_lock(&nblock); 352 if (needsbuffer) { 353 needsbuffer &= ~VFS_BIO_NEED_ANY; 354 if (numfreebuffers >= hifreebuffers) 355 needsbuffer &= ~VFS_BIO_NEED_FREE; 356 wakeup(&needsbuffer); 357 } 358 mtx_unlock(&nblock); 359} 360 361/* 362 * waitrunningbufspace() 363 * 364 * runningbufspace is a measure of the amount of I/O currently 365 * running. This routine is used in async-write situations to 366 * prevent creating huge backups of pending writes to a device. 367 * Only asynchronous writes are governed by this function. 368 * 369 * Reads will adjust runningbufspace, but will not block based on it. 370 * The read load has a side effect of reducing the allowed write load. 371 * 372 * This does NOT turn an async write into a sync write. It waits 373 * for earlier writes to complete and generally returns before the 374 * caller's write has reached the device. 375 */ 376static __inline void 377waitrunningbufspace(void) 378{ 379 380 mtx_lock(&rbreqlock); 381 while (runningbufspace > hirunningspace) { 382 ++runningbufreq; 383 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 384 } 385 mtx_unlock(&rbreqlock); 386} 387 388 389/* 390 * vfs_buf_test_cache: 391 * 392 * Called when a buffer is extended. This function clears the B_CACHE 393 * bit if the newly extended portion of the buffer does not contain 394 * valid data. 395 */ 396static __inline 397void 398vfs_buf_test_cache(struct buf *bp, 399 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 400 vm_page_t m) 401{ 402 403 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 404 if (bp->b_flags & B_CACHE) { 405 int base = (foff + off) & PAGE_MASK; 406 if (vm_page_is_valid(m, base, size) == 0) 407 bp->b_flags &= ~B_CACHE; 408 } 409} 410 411/* Wake up the buffer deamon if necessary */ 412static __inline 413void 414bd_wakeup(int dirtybuflevel) 415{ 416 417 mtx_lock(&bdlock); 418 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 419 bd_request = 1; 420 wakeup(&bd_request); 421 } 422 mtx_unlock(&bdlock); 423} 424 425/* 426 * bd_speedup - speedup the buffer cache flushing code 427 */ 428 429static __inline 430void 431bd_speedup(void) 432{ 433 434 bd_wakeup(1); 435} 436 437/* 438 * Calculating buffer cache scaling values and reserve space for buffer 439 * headers. This is called during low level kernel initialization and 440 * may be called more then once. We CANNOT write to the memory area 441 * being reserved at this time. 442 */ 443caddr_t 444kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 445{ 446 447 /* 448 * physmem_est is in pages. Convert it to kilobytes (assumes 449 * PAGE_SIZE is >= 1K) 450 */ 451 physmem_est = physmem_est * (PAGE_SIZE / 1024); 452 453 /* 454 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 455 * For the first 64MB of ram nominally allocate sufficient buffers to 456 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 457 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 458 * the buffer cache we limit the eventual kva reservation to 459 * maxbcache bytes. 460 * 461 * factor represents the 1/4 x ram conversion. 462 */ 463 if (nbuf == 0) { 464 int factor = 4 * BKVASIZE / 1024; 465 466 nbuf = 50; 467 if (physmem_est > 4096) 468 nbuf += min((physmem_est - 4096) / factor, 469 65536 / factor); 470 if (physmem_est > 65536) 471 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 472 473 if (maxbcache && nbuf > maxbcache / BKVASIZE) 474 nbuf = maxbcache / BKVASIZE; 475 } 476 477#if 0 478 /* 479 * Do not allow the buffer_map to be more then 1/2 the size of the 480 * kernel_map. 481 */ 482 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 483 (BKVASIZE * 2)) { 484 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 485 (BKVASIZE * 2); 486 printf("Warning: nbufs capped at %d\n", nbuf); 487 } 488#endif 489 490 /* 491 * swbufs are used as temporary holders for I/O, such as paging I/O. 492 * We have no less then 16 and no more then 256. 493 */ 494 nswbuf = max(min(nbuf/4, 256), 16); 495#ifdef NSWBUF_MIN 496 if (nswbuf < NSWBUF_MIN) 497 nswbuf = NSWBUF_MIN; 498#endif 499#ifdef DIRECTIO 500 ffs_rawread_setup(); 501#endif 502 503 /* 504 * Reserve space for the buffer cache buffers 505 */ 506 swbuf = (void *)v; 507 v = (caddr_t)(swbuf + nswbuf); 508 buf = (void *)v; 509 v = (caddr_t)(buf + nbuf); 510 511 return(v); 512} 513 514/* Initialize the buffer subsystem. Called before use of any buffers. */ 515void 516bufinit(void) 517{ 518 struct buf *bp; 519 int i; 520 521 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 522 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 523 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 524 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 525 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 526 527 /* next, make a null set of free lists */ 528 for (i = 0; i < BUFFER_QUEUES; i++) 529 TAILQ_INIT(&bufqueues[i]); 530 531 /* finally, initialize each buffer header and stick on empty q */ 532 for (i = 0; i < nbuf; i++) { 533 bp = &buf[i]; 534 bzero(bp, sizeof *bp); 535 bp->b_flags = B_INVAL; /* we're just an empty header */ 536 bp->b_rcred = NOCRED; 537 bp->b_wcred = NOCRED; 538 bp->b_qindex = QUEUE_EMPTY; 539 bp->b_vflags = 0; 540 bp->b_xflags = 0; 541 LIST_INIT(&bp->b_dep); 542 BUF_LOCKINIT(bp); 543 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 544 } 545 546 /* 547 * maxbufspace is the absolute maximum amount of buffer space we are 548 * allowed to reserve in KVM and in real terms. The absolute maximum 549 * is nominally used by buf_daemon. hibufspace is the nominal maximum 550 * used by most other processes. The differential is required to 551 * ensure that buf_daemon is able to run when other processes might 552 * be blocked waiting for buffer space. 553 * 554 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 555 * this may result in KVM fragmentation which is not handled optimally 556 * by the system. 557 */ 558 maxbufspace = nbuf * BKVASIZE; 559 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 560 lobufspace = hibufspace - MAXBSIZE; 561 562 lorunningspace = 512 * 1024; 563 hirunningspace = 1024 * 1024; 564 565/* 566 * Limit the amount of malloc memory since it is wired permanently into 567 * the kernel space. Even though this is accounted for in the buffer 568 * allocation, we don't want the malloced region to grow uncontrolled. 569 * The malloc scheme improves memory utilization significantly on average 570 * (small) directories. 571 */ 572 maxbufmallocspace = hibufspace / 20; 573 574/* 575 * Reduce the chance of a deadlock occuring by limiting the number 576 * of delayed-write dirty buffers we allow to stack up. 577 */ 578 hidirtybuffers = nbuf / 4 + 20; 579 dirtybufthresh = hidirtybuffers * 9 / 10; 580 numdirtybuffers = 0; 581/* 582 * To support extreme low-memory systems, make sure hidirtybuffers cannot 583 * eat up all available buffer space. This occurs when our minimum cannot 584 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 585 * BKVASIZE'd (8K) buffers. 586 */ 587 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 588 hidirtybuffers >>= 1; 589 } 590 lodirtybuffers = hidirtybuffers / 2; 591 592/* 593 * Try to keep the number of free buffers in the specified range, 594 * and give special processes (e.g. like buf_daemon) access to an 595 * emergency reserve. 596 */ 597 lofreebuffers = nbuf / 18 + 5; 598 hifreebuffers = 2 * lofreebuffers; 599 numfreebuffers = nbuf; 600 601/* 602 * Maximum number of async ops initiated per buf_daemon loop. This is 603 * somewhat of a hack at the moment, we really need to limit ourselves 604 * based on the number of bytes of I/O in-transit that were initiated 605 * from buf_daemon. 606 */ 607 608 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 609 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 610} 611 612/* 613 * bfreekva() - free the kva allocation for a buffer. 614 * 615 * Must be called at splbio() or higher as this is the only locking for 616 * buffer_map. 617 * 618 * Since this call frees up buffer space, we call bufspacewakeup(). 619 */ 620static void 621bfreekva(struct buf *bp) 622{ 623 624 if (bp->b_kvasize) { 625 atomic_add_int(&buffreekvacnt, 1); 626 atomic_subtract_int(&bufspace, bp->b_kvasize); 627 vm_map_lock(buffer_map); 628 vm_map_delete(buffer_map, 629 (vm_offset_t) bp->b_kvabase, 630 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 631 ); 632 vm_map_unlock(buffer_map); 633 bp->b_kvasize = 0; 634 bufspacewakeup(); 635 } 636} 637 638/* 639 * bremfree: 640 * 641 * Mark the buffer for removal from the appropriate free list in brelse. 642 * 643 */ 644void 645bremfree(struct buf *bp) 646{ 647 648 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 649 KASSERT(BUF_REFCNT(bp), ("bremfree: buf must be locked.")); 650 KASSERT((bp->b_flags & B_REMFREE) == 0 && bp->b_qindex != QUEUE_NONE, 651 ("bremfree: buffer %p not on a queue.", bp)); 652 653 bp->b_flags |= B_REMFREE; 654 /* Fixup numfreebuffers count. */ 655 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 656 atomic_subtract_int(&numfreebuffers, 1); 657} 658 659/* 660 * bremfreef: 661 * 662 * Force an immediate removal from a free list. Used only in nfs when 663 * it abuses the b_freelist pointer. 664 */ 665void 666bremfreef(struct buf *bp) 667{ 668 mtx_lock(&bqlock); 669 bremfreel(bp); 670 mtx_unlock(&bqlock); 671} 672 673/* 674 * bremfreel: 675 * 676 * Removes a buffer from the free list, must be called with the 677 * bqlock held. 678 */ 679static void 680bremfreel(struct buf *bp) 681{ 682 int s = splbio(); 683 684 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 685 bp, bp->b_vp, bp->b_flags); 686 KASSERT(BUF_REFCNT(bp), ("bremfreel: buffer %p not locked.", bp)); 687 KASSERT(bp->b_qindex != QUEUE_NONE, 688 ("bremfreel: buffer %p not on a queue.", bp)); 689 mtx_assert(&bqlock, MA_OWNED); 690 691 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 692 bp->b_qindex = QUEUE_NONE; 693 /* 694 * If this was a delayed bremfree() we only need to remove the buffer 695 * from the queue and return the stats are already done. 696 */ 697 if (bp->b_flags & B_REMFREE) { 698 bp->b_flags &= ~B_REMFREE; 699 splx(s); 700 return; 701 } 702 /* 703 * Fixup numfreebuffers count. If the buffer is invalid or not 704 * delayed-write, the buffer was free and we must decrement 705 * numfreebuffers. 706 */ 707 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 708 atomic_subtract_int(&numfreebuffers, 1); 709 splx(s); 710} 711 712 713/* 714 * Get a buffer with the specified data. Look in the cache first. We 715 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 716 * is set, the buffer is valid and we do not have to do anything ( see 717 * getblk() ). This is really just a special case of breadn(). 718 */ 719int 720bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 721 struct buf **bpp) 722{ 723 724 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 725} 726 727/* 728 * Operates like bread, but also starts asynchronous I/O on 729 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 730 * to initiating I/O . If B_CACHE is set, the buffer is valid 731 * and we do not have to do anything. 732 */ 733int 734breadn(struct vnode * vp, daddr_t blkno, int size, 735 daddr_t * rablkno, int *rabsize, 736 int cnt, struct ucred * cred, struct buf **bpp) 737{ 738 struct buf *bp, *rabp; 739 int i; 740 int rv = 0, readwait = 0; 741 742 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 743 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 744 745 /* if not found in cache, do some I/O */ 746 if ((bp->b_flags & B_CACHE) == 0) { 747 if (curthread != PCPU_GET(idlethread)) 748 curthread->td_proc->p_stats->p_ru.ru_inblock++; 749 bp->b_iocmd = BIO_READ; 750 bp->b_flags &= ~B_INVAL; 751 bp->b_ioflags &= ~BIO_ERROR; 752 if (bp->b_rcred == NOCRED && cred != NOCRED) 753 bp->b_rcred = crhold(cred); 754 vfs_busy_pages(bp, 0); 755 bp->b_iooffset = dbtob(bp->b_blkno); 756 bstrategy(bp); 757 ++readwait; 758 } 759 760 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 761 if (inmem(vp, *rablkno)) 762 continue; 763 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 764 765 if ((rabp->b_flags & B_CACHE) == 0) { 766 if (curthread != PCPU_GET(idlethread)) 767 curthread->td_proc->p_stats->p_ru.ru_inblock++; 768 rabp->b_flags |= B_ASYNC; 769 rabp->b_flags &= ~B_INVAL; 770 rabp->b_ioflags &= ~BIO_ERROR; 771 rabp->b_iocmd = BIO_READ; 772 if (rabp->b_rcred == NOCRED && cred != NOCRED) 773 rabp->b_rcred = crhold(cred); 774 vfs_busy_pages(rabp, 0); 775 BUF_KERNPROC(rabp); 776 rabp->b_iooffset = dbtob(rabp->b_blkno); 777 bstrategy(rabp); 778 } else { 779 brelse(rabp); 780 } 781 } 782 783 if (readwait) { 784 rv = bufwait(bp); 785 } 786 return (rv); 787} 788 789/* 790 * Write, release buffer on completion. (Done by iodone 791 * if async). Do not bother writing anything if the buffer 792 * is invalid. 793 * 794 * Note that we set B_CACHE here, indicating that buffer is 795 * fully valid and thus cacheable. This is true even of NFS 796 * now so we set it generally. This could be set either here 797 * or in biodone() since the I/O is synchronous. We put it 798 * here. 799 */ 800int 801bufwrite(struct buf *bp) 802{ 803 int oldflags, s; 804 805 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 806 if (bp->b_flags & B_INVAL) { 807 brelse(bp); 808 return (0); 809 } 810 811 oldflags = bp->b_flags; 812 813 if (BUF_REFCNT(bp) == 0) 814 panic("bufwrite: buffer is not busy???"); 815 s = splbio(); 816 817 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 818 ("FFS background buffer should not get here %p", bp)); 819 820 /* Mark the buffer clean */ 821 bundirty(bp); 822 823 bp->b_flags &= ~B_DONE; 824 bp->b_ioflags &= ~BIO_ERROR; 825 bp->b_flags |= B_CACHE; 826 bp->b_iocmd = BIO_WRITE; 827 828 bufobj_wref(bp->b_bufobj); 829 vfs_busy_pages(bp, 1); 830 831 /* 832 * Normal bwrites pipeline writes 833 */ 834 bp->b_runningbufspace = bp->b_bufsize; 835 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 836 837 if (curthread != PCPU_GET(idlethread)) 838 curthread->td_proc->p_stats->p_ru.ru_oublock++; 839 splx(s); 840 if (oldflags & B_ASYNC) 841 BUF_KERNPROC(bp); 842 bp->b_iooffset = dbtob(bp->b_blkno); 843 bstrategy(bp); 844 845 if ((oldflags & B_ASYNC) == 0) { 846 int rtval = bufwait(bp); 847 brelse(bp); 848 return (rtval); 849 } else { 850 /* 851 * don't allow the async write to saturate the I/O 852 * system. We will not deadlock here because 853 * we are blocking waiting for I/O that is already in-progress 854 * to complete. We do not block here if it is the update 855 * or syncer daemon trying to clean up as that can lead 856 * to deadlock. 857 */ 858 if (curthread->td_proc != bufdaemonproc && 859 curthread->td_proc != updateproc) 860 waitrunningbufspace(); 861 } 862 863 return (0); 864} 865 866/* 867 * Delayed write. (Buffer is marked dirty). Do not bother writing 868 * anything if the buffer is marked invalid. 869 * 870 * Note that since the buffer must be completely valid, we can safely 871 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 872 * biodone() in order to prevent getblk from writing the buffer 873 * out synchronously. 874 */ 875void 876bdwrite(struct buf *bp) 877{ 878 struct thread *td = curthread; 879 struct vnode *vp; 880 struct buf *nbp; 881 struct bufobj *bo; 882 883 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 884 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 885 KASSERT(BUF_REFCNT(bp) != 0, ("bdwrite: buffer is not busy")); 886 887 if (bp->b_flags & B_INVAL) { 888 brelse(bp); 889 return; 890 } 891 892 /* 893 * If we have too many dirty buffers, don't create any more. 894 * If we are wildly over our limit, then force a complete 895 * cleanup. Otherwise, just keep the situation from getting 896 * out of control. Note that we have to avoid a recursive 897 * disaster and not try to clean up after our own cleanup! 898 */ 899 vp = bp->b_vp; 900 bo = bp->b_bufobj; 901 if ((td->td_pflags & TDP_COWINPROGRESS) == 0) { 902 BO_LOCK(bo); 903 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 904 BO_UNLOCK(bo); 905 (void) VOP_FSYNC(vp, MNT_NOWAIT, td); 906 altbufferflushes++; 907 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 908 /* 909 * Try to find a buffer to flush. 910 */ 911 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 912 if ((nbp->b_vflags & BV_BKGRDINPROG) || 913 BUF_LOCK(nbp, 914 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 915 continue; 916 if (bp == nbp) 917 panic("bdwrite: found ourselves"); 918 BO_UNLOCK(bo); 919 /* Don't countdeps with the bo lock held. */ 920 if (buf_countdeps(nbp, 0)) { 921 BO_LOCK(bo); 922 BUF_UNLOCK(nbp); 923 continue; 924 } 925 if (nbp->b_flags & B_CLUSTEROK) { 926 vfs_bio_awrite(nbp); 927 } else { 928 bremfree(nbp); 929 bawrite(nbp); 930 } 931 dirtybufferflushes++; 932 break; 933 } 934 if (nbp == NULL) 935 BO_UNLOCK(bo); 936 } else 937 BO_UNLOCK(bo); 938 } else 939 recursiveflushes++; 940 941 bdirty(bp); 942 /* 943 * Set B_CACHE, indicating that the buffer is fully valid. This is 944 * true even of NFS now. 945 */ 946 bp->b_flags |= B_CACHE; 947 948 /* 949 * This bmap keeps the system from needing to do the bmap later, 950 * perhaps when the system is attempting to do a sync. Since it 951 * is likely that the indirect block -- or whatever other datastructure 952 * that the filesystem needs is still in memory now, it is a good 953 * thing to do this. Note also, that if the pageout daemon is 954 * requesting a sync -- there might not be enough memory to do 955 * the bmap then... So, this is important to do. 956 */ 957 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 958 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 959 } 960 961 /* 962 * Set the *dirty* buffer range based upon the VM system dirty pages. 963 */ 964 vfs_setdirty(bp); 965 966 /* 967 * We need to do this here to satisfy the vnode_pager and the 968 * pageout daemon, so that it thinks that the pages have been 969 * "cleaned". Note that since the pages are in a delayed write 970 * buffer -- the VFS layer "will" see that the pages get written 971 * out on the next sync, or perhaps the cluster will be completed. 972 */ 973 vfs_clean_pages(bp); 974 bqrelse(bp); 975 976 /* 977 * Wakeup the buffer flushing daemon if we have a lot of dirty 978 * buffers (midpoint between our recovery point and our stall 979 * point). 980 */ 981 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 982 983 /* 984 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 985 * due to the softdep code. 986 */ 987} 988 989/* 990 * bdirty: 991 * 992 * Turn buffer into delayed write request. We must clear BIO_READ and 993 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 994 * itself to properly update it in the dirty/clean lists. We mark it 995 * B_DONE to ensure that any asynchronization of the buffer properly 996 * clears B_DONE ( else a panic will occur later ). 997 * 998 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 999 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1000 * should only be called if the buffer is known-good. 1001 * 1002 * Since the buffer is not on a queue, we do not update the numfreebuffers 1003 * count. 1004 * 1005 * Must be called at splbio(). 1006 * The buffer must be on QUEUE_NONE. 1007 */ 1008void 1009bdirty(struct buf *bp) 1010{ 1011 1012 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1013 bp, bp->b_vp, bp->b_flags); 1014 KASSERT(BUF_REFCNT(bp) == 1, ("bdirty: bp %p not locked",bp)); 1015 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1016 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1017 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1018 bp->b_flags &= ~(B_RELBUF); 1019 bp->b_iocmd = BIO_WRITE; 1020 1021 if ((bp->b_flags & B_DELWRI) == 0) { 1022 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1023 reassignbuf(bp); 1024 atomic_add_int(&numdirtybuffers, 1); 1025 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1026 } 1027} 1028 1029/* 1030 * bundirty: 1031 * 1032 * Clear B_DELWRI for buffer. 1033 * 1034 * Since the buffer is not on a queue, we do not update the numfreebuffers 1035 * count. 1036 * 1037 * Must be called at splbio(). 1038 * The buffer must be on QUEUE_NONE. 1039 */ 1040 1041void 1042bundirty(struct buf *bp) 1043{ 1044 1045 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1046 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1047 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1048 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1049 KASSERT(BUF_REFCNT(bp) == 1, ("bundirty: bp %p not locked",bp)); 1050 1051 if (bp->b_flags & B_DELWRI) { 1052 bp->b_flags &= ~B_DELWRI; 1053 reassignbuf(bp); 1054 atomic_subtract_int(&numdirtybuffers, 1); 1055 numdirtywakeup(lodirtybuffers); 1056 } 1057 /* 1058 * Since it is now being written, we can clear its deferred write flag. 1059 */ 1060 bp->b_flags &= ~B_DEFERRED; 1061} 1062 1063/* 1064 * bawrite: 1065 * 1066 * Asynchronous write. Start output on a buffer, but do not wait for 1067 * it to complete. The buffer is released when the output completes. 1068 * 1069 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1070 * B_INVAL buffers. Not us. 1071 */ 1072void 1073bawrite(struct buf *bp) 1074{ 1075 1076 bp->b_flags |= B_ASYNC; 1077 (void) bwrite(bp); 1078} 1079 1080/* 1081 * bwillwrite: 1082 * 1083 * Called prior to the locking of any vnodes when we are expecting to 1084 * write. We do not want to starve the buffer cache with too many 1085 * dirty buffers so we block here. By blocking prior to the locking 1086 * of any vnodes we attempt to avoid the situation where a locked vnode 1087 * prevents the various system daemons from flushing related buffers. 1088 */ 1089 1090void 1091bwillwrite(void) 1092{ 1093 1094 if (numdirtybuffers >= hidirtybuffers) { 1095 int s; 1096 1097 s = splbio(); 1098 mtx_lock(&nblock); 1099 while (numdirtybuffers >= hidirtybuffers) { 1100 bd_wakeup(1); 1101 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1102 msleep(&needsbuffer, &nblock, 1103 (PRIBIO + 4), "flswai", 0); 1104 } 1105 splx(s); 1106 mtx_unlock(&nblock); 1107 } 1108} 1109 1110/* 1111 * Return true if we have too many dirty buffers. 1112 */ 1113int 1114buf_dirty_count_severe(void) 1115{ 1116 1117 return(numdirtybuffers >= hidirtybuffers); 1118} 1119 1120/* 1121 * brelse: 1122 * 1123 * Release a busy buffer and, if requested, free its resources. The 1124 * buffer will be stashed in the appropriate bufqueue[] allowing it 1125 * to be accessed later as a cache entity or reused for other purposes. 1126 */ 1127void 1128brelse(struct buf *bp) 1129{ 1130 int s; 1131 1132 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1133 bp, bp->b_vp, bp->b_flags); 1134 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1135 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1136 1137 s = splbio(); 1138 1139 if (bp->b_iocmd == BIO_WRITE && 1140 (bp->b_ioflags & BIO_ERROR) && 1141 !(bp->b_flags & B_INVAL)) { 1142 /* 1143 * Failed write, redirty. Must clear BIO_ERROR to prevent 1144 * pages from being scrapped. If B_INVAL is set then 1145 * this case is not run and the next case is run to 1146 * destroy the buffer. B_INVAL can occur if the buffer 1147 * is outside the range supported by the underlying device. 1148 */ 1149 bp->b_ioflags &= ~BIO_ERROR; 1150 bdirty(bp); 1151 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1152 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1153 /* 1154 * Either a failed I/O or we were asked to free or not 1155 * cache the buffer. 1156 */ 1157 bp->b_flags |= B_INVAL; 1158 if (LIST_FIRST(&bp->b_dep) != NULL) 1159 buf_deallocate(bp); 1160 if (bp->b_flags & B_DELWRI) { 1161 atomic_subtract_int(&numdirtybuffers, 1); 1162 numdirtywakeup(lodirtybuffers); 1163 } 1164 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1165 if ((bp->b_flags & B_VMIO) == 0) { 1166 if (bp->b_bufsize) 1167 allocbuf(bp, 0); 1168 if (bp->b_vp) 1169 brelvp(bp); 1170 } 1171 } 1172 1173 /* 1174 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1175 * is called with B_DELWRI set, the underlying pages may wind up 1176 * getting freed causing a previous write (bdwrite()) to get 'lost' 1177 * because pages associated with a B_DELWRI bp are marked clean. 1178 * 1179 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1180 * if B_DELWRI is set. 1181 * 1182 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1183 * on pages to return pages to the VM page queues. 1184 */ 1185 if (bp->b_flags & B_DELWRI) 1186 bp->b_flags &= ~B_RELBUF; 1187 else if (vm_page_count_severe()) { 1188 /* 1189 * XXX This lock may not be necessary since BKGRDINPROG 1190 * cannot be set while we hold the buf lock, it can only be 1191 * cleared if it is already pending. 1192 */ 1193 if (bp->b_vp) { 1194 BO_LOCK(bp->b_bufobj); 1195 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1196 bp->b_flags |= B_RELBUF; 1197 BO_UNLOCK(bp->b_bufobj); 1198 } else 1199 bp->b_flags |= B_RELBUF; 1200 } 1201 1202 /* 1203 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1204 * constituted, not even NFS buffers now. Two flags effect this. If 1205 * B_INVAL, the struct buf is invalidated but the VM object is kept 1206 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1207 * 1208 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1209 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1210 * buffer is also B_INVAL because it hits the re-dirtying code above. 1211 * 1212 * Normally we can do this whether a buffer is B_DELWRI or not. If 1213 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1214 * the commit state and we cannot afford to lose the buffer. If the 1215 * buffer has a background write in progress, we need to keep it 1216 * around to prevent it from being reconstituted and starting a second 1217 * background write. 1218 */ 1219 if ((bp->b_flags & B_VMIO) 1220 && !(bp->b_vp->v_mount != NULL && 1221 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1222 !vn_isdisk(bp->b_vp, NULL) && 1223 (bp->b_flags & B_DELWRI)) 1224 ) { 1225 1226 int i, j, resid; 1227 vm_page_t m; 1228 off_t foff; 1229 vm_pindex_t poff; 1230 vm_object_t obj; 1231 1232 obj = bp->b_bufobj->bo_object; 1233 1234 /* 1235 * Get the base offset and length of the buffer. Note that 1236 * in the VMIO case if the buffer block size is not 1237 * page-aligned then b_data pointer may not be page-aligned. 1238 * But our b_pages[] array *IS* page aligned. 1239 * 1240 * block sizes less then DEV_BSIZE (usually 512) are not 1241 * supported due to the page granularity bits (m->valid, 1242 * m->dirty, etc...). 1243 * 1244 * See man buf(9) for more information 1245 */ 1246 resid = bp->b_bufsize; 1247 foff = bp->b_offset; 1248 VM_OBJECT_LOCK(obj); 1249 for (i = 0; i < bp->b_npages; i++) { 1250 int had_bogus = 0; 1251 1252 m = bp->b_pages[i]; 1253 1254 /* 1255 * If we hit a bogus page, fixup *all* the bogus pages 1256 * now. 1257 */ 1258 if (m == bogus_page) { 1259 poff = OFF_TO_IDX(bp->b_offset); 1260 had_bogus = 1; 1261 1262 for (j = i; j < bp->b_npages; j++) { 1263 vm_page_t mtmp; 1264 mtmp = bp->b_pages[j]; 1265 if (mtmp == bogus_page) { 1266 mtmp = vm_page_lookup(obj, poff + j); 1267 if (!mtmp) { 1268 panic("brelse: page missing\n"); 1269 } 1270 bp->b_pages[j] = mtmp; 1271 } 1272 } 1273 1274 if ((bp->b_flags & B_INVAL) == 0) { 1275 pmap_qenter( 1276 trunc_page((vm_offset_t)bp->b_data), 1277 bp->b_pages, bp->b_npages); 1278 } 1279 m = bp->b_pages[i]; 1280 } 1281 if ((bp->b_flags & B_NOCACHE) || 1282 (bp->b_ioflags & BIO_ERROR)) { 1283 int poffset = foff & PAGE_MASK; 1284 int presid = resid > (PAGE_SIZE - poffset) ? 1285 (PAGE_SIZE - poffset) : resid; 1286 1287 KASSERT(presid >= 0, ("brelse: extra page")); 1288 vm_page_lock_queues(); 1289 vm_page_set_invalid(m, poffset, presid); 1290 vm_page_unlock_queues(); 1291 if (had_bogus) 1292 printf("avoided corruption bug in bogus_page/brelse code\n"); 1293 } 1294 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1295 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1296 } 1297 VM_OBJECT_UNLOCK(obj); 1298 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1299 vfs_vmio_release(bp); 1300 1301 } else if (bp->b_flags & B_VMIO) { 1302 1303 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1304 vfs_vmio_release(bp); 1305 } 1306 1307 } 1308 1309 if (BUF_REFCNT(bp) > 1) { 1310 /* do not release to free list */ 1311 BUF_UNLOCK(bp); 1312 splx(s); 1313 return; 1314 } 1315 1316 /* enqueue */ 1317 mtx_lock(&bqlock); 1318 /* Handle delayed bremfree() processing. */ 1319 if (bp->b_flags & B_REMFREE) 1320 bremfreel(bp); 1321 if (bp->b_qindex != QUEUE_NONE) 1322 panic("brelse: free buffer onto another queue???"); 1323 1324 /* buffers with no memory */ 1325 if (bp->b_bufsize == 0) { 1326 bp->b_flags |= B_INVAL; 1327 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1328 if (bp->b_vflags & BV_BKGRDINPROG) 1329 panic("losing buffer 1"); 1330 if (bp->b_kvasize) { 1331 bp->b_qindex = QUEUE_EMPTYKVA; 1332 } else { 1333 bp->b_qindex = QUEUE_EMPTY; 1334 } 1335 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1336 /* buffers with junk contents */ 1337 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1338 (bp->b_ioflags & BIO_ERROR)) { 1339 bp->b_flags |= B_INVAL; 1340 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1341 if (bp->b_vflags & BV_BKGRDINPROG) 1342 panic("losing buffer 2"); 1343 bp->b_qindex = QUEUE_CLEAN; 1344 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1345 /* remaining buffers */ 1346 } else { 1347 if (bp->b_flags & B_DELWRI) 1348 bp->b_qindex = QUEUE_DIRTY; 1349 else 1350 bp->b_qindex = QUEUE_CLEAN; 1351 if (bp->b_flags & B_AGE) 1352 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1353 else 1354 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1355 } 1356 mtx_unlock(&bqlock); 1357 1358 /* 1359 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1360 * placed the buffer on the correct queue. We must also disassociate 1361 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1362 * find it. 1363 */ 1364 if (bp->b_flags & B_INVAL) { 1365 if (bp->b_flags & B_DELWRI) 1366 bundirty(bp); 1367 if (bp->b_vp) 1368 brelvp(bp); 1369 } 1370 1371 /* 1372 * Fixup numfreebuffers count. The bp is on an appropriate queue 1373 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1374 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1375 * if B_INVAL is set ). 1376 */ 1377 1378 if (!(bp->b_flags & B_DELWRI)) 1379 bufcountwakeup(); 1380 1381 /* 1382 * Something we can maybe free or reuse 1383 */ 1384 if (bp->b_bufsize || bp->b_kvasize) 1385 bufspacewakeup(); 1386 1387 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1388 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1389 panic("brelse: not dirty"); 1390 /* unlock */ 1391 BUF_UNLOCK(bp); 1392 splx(s); 1393} 1394 1395/* 1396 * Release a buffer back to the appropriate queue but do not try to free 1397 * it. The buffer is expected to be used again soon. 1398 * 1399 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1400 * biodone() to requeue an async I/O on completion. It is also used when 1401 * known good buffers need to be requeued but we think we may need the data 1402 * again soon. 1403 * 1404 * XXX we should be able to leave the B_RELBUF hint set on completion. 1405 */ 1406void 1407bqrelse(struct buf *bp) 1408{ 1409 int s; 1410 1411 s = splbio(); 1412 1413 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1414 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1415 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1416 1417 if (BUF_REFCNT(bp) > 1) { 1418 /* do not release to free list */ 1419 BUF_UNLOCK(bp); 1420 splx(s); 1421 return; 1422 } 1423 mtx_lock(&bqlock); 1424 /* Handle delayed bremfree() processing. */ 1425 if (bp->b_flags & B_REMFREE) 1426 bremfreel(bp); 1427 if (bp->b_qindex != QUEUE_NONE) 1428 panic("bqrelse: free buffer onto another queue???"); 1429 /* buffers with stale but valid contents */ 1430 if (bp->b_flags & B_DELWRI) { 1431 bp->b_qindex = QUEUE_DIRTY; 1432 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1433 } else { 1434 /* 1435 * XXX This lock may not be necessary since BKGRDINPROG 1436 * cannot be set while we hold the buf lock, it can only be 1437 * cleared if it is already pending. 1438 */ 1439 BO_LOCK(bp->b_bufobj); 1440 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1441 BO_UNLOCK(bp->b_bufobj); 1442 bp->b_qindex = QUEUE_CLEAN; 1443 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1444 b_freelist); 1445 } else { 1446 /* 1447 * We are too low on memory, we have to try to free 1448 * the buffer (most importantly: the wired pages 1449 * making up its backing store) *now*. 1450 */ 1451 BO_UNLOCK(bp->b_bufobj); 1452 mtx_unlock(&bqlock); 1453 splx(s); 1454 brelse(bp); 1455 return; 1456 } 1457 } 1458 mtx_unlock(&bqlock); 1459 1460 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1461 bufcountwakeup(); 1462 1463 /* 1464 * Something we can maybe free or reuse. 1465 */ 1466 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1467 bufspacewakeup(); 1468 1469 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1470 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1471 panic("bqrelse: not dirty"); 1472 /* unlock */ 1473 BUF_UNLOCK(bp); 1474 splx(s); 1475} 1476 1477/* Give pages used by the bp back to the VM system (where possible) */ 1478static void 1479vfs_vmio_release(struct buf *bp) 1480{ 1481 int i; 1482 vm_page_t m; 1483 1484 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1485 vm_page_lock_queues(); 1486 for (i = 0; i < bp->b_npages; i++) { 1487 m = bp->b_pages[i]; 1488 bp->b_pages[i] = NULL; 1489 /* 1490 * In order to keep page LRU ordering consistent, put 1491 * everything on the inactive queue. 1492 */ 1493 vm_page_unwire(m, 0); 1494 /* 1495 * We don't mess with busy pages, it is 1496 * the responsibility of the process that 1497 * busied the pages to deal with them. 1498 */ 1499 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1500 continue; 1501 1502 if (m->wire_count == 0) { 1503 /* 1504 * Might as well free the page if we can and it has 1505 * no valid data. We also free the page if the 1506 * buffer was used for direct I/O 1507 */ 1508 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1509 m->hold_count == 0) { 1510 pmap_remove_all(m); 1511 vm_page_free(m); 1512 } else if (bp->b_flags & B_DIRECT) { 1513 vm_page_try_to_free(m); 1514 } else if (vm_page_count_severe()) { 1515 vm_page_try_to_cache(m); 1516 } 1517 } 1518 } 1519 vm_page_unlock_queues(); 1520 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1521 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1522 1523 if (bp->b_bufsize) { 1524 bufspacewakeup(); 1525 bp->b_bufsize = 0; 1526 } 1527 bp->b_npages = 0; 1528 bp->b_flags &= ~B_VMIO; 1529 if (bp->b_vp) 1530 brelvp(bp); 1531} 1532 1533/* 1534 * Check to see if a block at a particular lbn is available for a clustered 1535 * write. 1536 */ 1537static int 1538vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1539{ 1540 struct buf *bpa; 1541 int match; 1542 1543 match = 0; 1544 1545 /* If the buf isn't in core skip it */ 1546 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1547 return (0); 1548 1549 /* If the buf is busy we don't want to wait for it */ 1550 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1551 return (0); 1552 1553 /* Only cluster with valid clusterable delayed write buffers */ 1554 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1555 (B_DELWRI | B_CLUSTEROK)) 1556 goto done; 1557 1558 if (bpa->b_bufsize != size) 1559 goto done; 1560 1561 /* 1562 * Check to see if it is in the expected place on disk and that the 1563 * block has been mapped. 1564 */ 1565 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1566 match = 1; 1567done: 1568 BUF_UNLOCK(bpa); 1569 return (match); 1570} 1571 1572/* 1573 * vfs_bio_awrite: 1574 * 1575 * Implement clustered async writes for clearing out B_DELWRI buffers. 1576 * This is much better then the old way of writing only one buffer at 1577 * a time. Note that we may not be presented with the buffers in the 1578 * correct order, so we search for the cluster in both directions. 1579 */ 1580int 1581vfs_bio_awrite(struct buf *bp) 1582{ 1583 int i; 1584 int j; 1585 daddr_t lblkno = bp->b_lblkno; 1586 struct vnode *vp = bp->b_vp; 1587 int s; 1588 int ncl; 1589 int nwritten; 1590 int size; 1591 int maxcl; 1592 1593 s = splbio(); 1594 /* 1595 * right now we support clustered writing only to regular files. If 1596 * we find a clusterable block we could be in the middle of a cluster 1597 * rather then at the beginning. 1598 */ 1599 if ((vp->v_type == VREG) && 1600 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1601 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1602 1603 size = vp->v_mount->mnt_stat.f_iosize; 1604 maxcl = MAXPHYS / size; 1605 1606 VI_LOCK(vp); 1607 for (i = 1; i < maxcl; i++) 1608 if (vfs_bio_clcheck(vp, size, lblkno + i, 1609 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1610 break; 1611 1612 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1613 if (vfs_bio_clcheck(vp, size, lblkno - j, 1614 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1615 break; 1616 1617 VI_UNLOCK(vp); 1618 --j; 1619 ncl = i + j; 1620 /* 1621 * this is a possible cluster write 1622 */ 1623 if (ncl != 1) { 1624 BUF_UNLOCK(bp); 1625 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1626 splx(s); 1627 return nwritten; 1628 } 1629 } 1630 1631 bremfree(bp); 1632 bp->b_flags |= B_ASYNC; 1633 1634 splx(s); 1635 /* 1636 * default (old) behavior, writing out only one block 1637 * 1638 * XXX returns b_bufsize instead of b_bcount for nwritten? 1639 */ 1640 nwritten = bp->b_bufsize; 1641 (void) bwrite(bp); 1642 1643 return nwritten; 1644} 1645 1646/* 1647 * getnewbuf: 1648 * 1649 * Find and initialize a new buffer header, freeing up existing buffers 1650 * in the bufqueues as necessary. The new buffer is returned locked. 1651 * 1652 * Important: B_INVAL is not set. If the caller wishes to throw the 1653 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1654 * 1655 * We block if: 1656 * We have insufficient buffer headers 1657 * We have insufficient buffer space 1658 * buffer_map is too fragmented ( space reservation fails ) 1659 * If we have to flush dirty buffers ( but we try to avoid this ) 1660 * 1661 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1662 * Instead we ask the buf daemon to do it for us. We attempt to 1663 * avoid piecemeal wakeups of the pageout daemon. 1664 */ 1665 1666static struct buf * 1667getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1668{ 1669 struct buf *bp; 1670 struct buf *nbp; 1671 int defrag = 0; 1672 int nqindex; 1673 static int flushingbufs; 1674 1675 /* 1676 * We can't afford to block since we might be holding a vnode lock, 1677 * which may prevent system daemons from running. We deal with 1678 * low-memory situations by proactively returning memory and running 1679 * async I/O rather then sync I/O. 1680 */ 1681 1682 atomic_add_int(&getnewbufcalls, 1); 1683 atomic_subtract_int(&getnewbufrestarts, 1); 1684restart: 1685 atomic_add_int(&getnewbufrestarts, 1); 1686 1687 /* 1688 * Setup for scan. If we do not have enough free buffers, 1689 * we setup a degenerate case that immediately fails. Note 1690 * that if we are specially marked process, we are allowed to 1691 * dip into our reserves. 1692 * 1693 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1694 * 1695 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1696 * However, there are a number of cases (defragging, reusing, ...) 1697 * where we cannot backup. 1698 */ 1699 mtx_lock(&bqlock); 1700 nqindex = QUEUE_EMPTYKVA; 1701 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1702 1703 if (nbp == NULL) { 1704 /* 1705 * If no EMPTYKVA buffers and we are either 1706 * defragging or reusing, locate a CLEAN buffer 1707 * to free or reuse. If bufspace useage is low 1708 * skip this step so we can allocate a new buffer. 1709 */ 1710 if (defrag || bufspace >= lobufspace) { 1711 nqindex = QUEUE_CLEAN; 1712 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1713 } 1714 1715 /* 1716 * If we could not find or were not allowed to reuse a 1717 * CLEAN buffer, check to see if it is ok to use an EMPTY 1718 * buffer. We can only use an EMPTY buffer if allocating 1719 * its KVA would not otherwise run us out of buffer space. 1720 */ 1721 if (nbp == NULL && defrag == 0 && 1722 bufspace + maxsize < hibufspace) { 1723 nqindex = QUEUE_EMPTY; 1724 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1725 } 1726 } 1727 1728 /* 1729 * Run scan, possibly freeing data and/or kva mappings on the fly 1730 * depending. 1731 */ 1732 1733 while ((bp = nbp) != NULL) { 1734 int qindex = nqindex; 1735 1736 /* 1737 * Calculate next bp ( we can only use it if we do not block 1738 * or do other fancy things ). 1739 */ 1740 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1741 switch(qindex) { 1742 case QUEUE_EMPTY: 1743 nqindex = QUEUE_EMPTYKVA; 1744 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1745 break; 1746 /* FALLTHROUGH */ 1747 case QUEUE_EMPTYKVA: 1748 nqindex = QUEUE_CLEAN; 1749 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1750 break; 1751 /* FALLTHROUGH */ 1752 case QUEUE_CLEAN: 1753 /* 1754 * nbp is NULL. 1755 */ 1756 break; 1757 } 1758 } 1759 /* 1760 * If we are defragging then we need a buffer with 1761 * b_kvasize != 0. XXX this situation should no longer 1762 * occur, if defrag is non-zero the buffer's b_kvasize 1763 * should also be non-zero at this point. XXX 1764 */ 1765 if (defrag && bp->b_kvasize == 0) { 1766 printf("Warning: defrag empty buffer %p\n", bp); 1767 continue; 1768 } 1769 1770 /* 1771 * Start freeing the bp. This is somewhat involved. nbp 1772 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1773 */ 1774 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1775 continue; 1776 if (bp->b_vp) { 1777 BO_LOCK(bp->b_bufobj); 1778 if (bp->b_vflags & BV_BKGRDINPROG) { 1779 BO_UNLOCK(bp->b_bufobj); 1780 BUF_UNLOCK(bp); 1781 continue; 1782 } 1783 BO_UNLOCK(bp->b_bufobj); 1784 } 1785 CTR6(KTR_BUF, 1786 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1787 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1788 bp->b_kvasize, bp->b_bufsize, qindex); 1789 1790 /* 1791 * Sanity Checks 1792 */ 1793 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1794 1795 /* 1796 * Note: we no longer distinguish between VMIO and non-VMIO 1797 * buffers. 1798 */ 1799 1800 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1801 1802 bremfreel(bp); 1803 mtx_unlock(&bqlock); 1804 1805 if (qindex == QUEUE_CLEAN) { 1806 if (bp->b_flags & B_VMIO) { 1807 bp->b_flags &= ~B_ASYNC; 1808 vfs_vmio_release(bp); 1809 } 1810 if (bp->b_vp) 1811 brelvp(bp); 1812 } 1813 1814 /* 1815 * NOTE: nbp is now entirely invalid. We can only restart 1816 * the scan from this point on. 1817 * 1818 * Get the rest of the buffer freed up. b_kva* is still 1819 * valid after this operation. 1820 */ 1821 1822 if (bp->b_rcred != NOCRED) { 1823 crfree(bp->b_rcred); 1824 bp->b_rcred = NOCRED; 1825 } 1826 if (bp->b_wcred != NOCRED) { 1827 crfree(bp->b_wcred); 1828 bp->b_wcred = NOCRED; 1829 } 1830 if (LIST_FIRST(&bp->b_dep) != NULL) 1831 buf_deallocate(bp); 1832 if (bp->b_vflags & BV_BKGRDINPROG) 1833 panic("losing buffer 3"); 1834 1835 if (bp->b_bufsize) 1836 allocbuf(bp, 0); 1837 1838 bp->b_flags = 0; 1839 bp->b_ioflags = 0; 1840 bp->b_xflags = 0; 1841 bp->b_vflags = 0; 1842 bp->b_vp = NULL; 1843 bp->b_blkno = bp->b_lblkno = 0; 1844 bp->b_offset = NOOFFSET; 1845 bp->b_iodone = 0; 1846 bp->b_error = 0; 1847 bp->b_resid = 0; 1848 bp->b_bcount = 0; 1849 bp->b_npages = 0; 1850 bp->b_dirtyoff = bp->b_dirtyend = 0; 1851 bp->b_bufobj = NULL; 1852 1853 LIST_INIT(&bp->b_dep); 1854 1855 /* 1856 * If we are defragging then free the buffer. 1857 */ 1858 if (defrag) { 1859 bp->b_flags |= B_INVAL; 1860 bfreekva(bp); 1861 brelse(bp); 1862 defrag = 0; 1863 goto restart; 1864 } 1865 1866 /* 1867 * If we are overcomitted then recover the buffer and its 1868 * KVM space. This occurs in rare situations when multiple 1869 * processes are blocked in getnewbuf() or allocbuf(). 1870 */ 1871 if (bufspace >= hibufspace) 1872 flushingbufs = 1; 1873 if (flushingbufs && bp->b_kvasize != 0) { 1874 bp->b_flags |= B_INVAL; 1875 bfreekva(bp); 1876 brelse(bp); 1877 goto restart; 1878 } 1879 if (bufspace < lobufspace) 1880 flushingbufs = 0; 1881 break; 1882 } 1883 1884 /* 1885 * If we exhausted our list, sleep as appropriate. We may have to 1886 * wakeup various daemons and write out some dirty buffers. 1887 * 1888 * Generally we are sleeping due to insufficient buffer space. 1889 */ 1890 1891 if (bp == NULL) { 1892 int flags; 1893 char *waitmsg; 1894 1895 mtx_unlock(&bqlock); 1896 if (defrag) { 1897 flags = VFS_BIO_NEED_BUFSPACE; 1898 waitmsg = "nbufkv"; 1899 } else if (bufspace >= hibufspace) { 1900 waitmsg = "nbufbs"; 1901 flags = VFS_BIO_NEED_BUFSPACE; 1902 } else { 1903 waitmsg = "newbuf"; 1904 flags = VFS_BIO_NEED_ANY; 1905 } 1906 1907 bd_speedup(); /* heeeelp */ 1908 1909 mtx_lock(&nblock); 1910 needsbuffer |= flags; 1911 while (needsbuffer & flags) { 1912 if (msleep(&needsbuffer, &nblock, 1913 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1914 mtx_unlock(&nblock); 1915 return (NULL); 1916 } 1917 } 1918 mtx_unlock(&nblock); 1919 } else { 1920 /* 1921 * We finally have a valid bp. We aren't quite out of the 1922 * woods, we still have to reserve kva space. In order 1923 * to keep fragmentation sane we only allocate kva in 1924 * BKVASIZE chunks. 1925 */ 1926 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1927 1928 if (maxsize != bp->b_kvasize) { 1929 vm_offset_t addr = 0; 1930 1931 bfreekva(bp); 1932 1933 vm_map_lock(buffer_map); 1934 if (vm_map_findspace(buffer_map, 1935 vm_map_min(buffer_map), maxsize, &addr)) { 1936 /* 1937 * Uh oh. Buffer map is to fragmented. We 1938 * must defragment the map. 1939 */ 1940 atomic_add_int(&bufdefragcnt, 1); 1941 vm_map_unlock(buffer_map); 1942 defrag = 1; 1943 bp->b_flags |= B_INVAL; 1944 brelse(bp); 1945 goto restart; 1946 } 1947 if (addr) { 1948 vm_map_insert(buffer_map, NULL, 0, 1949 addr, addr + maxsize, 1950 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1951 1952 bp->b_kvabase = (caddr_t) addr; 1953 bp->b_kvasize = maxsize; 1954 atomic_add_int(&bufspace, bp->b_kvasize); 1955 atomic_add_int(&bufreusecnt, 1); 1956 } 1957 vm_map_unlock(buffer_map); 1958 } 1959 bp->b_saveaddr = bp->b_kvabase; 1960 bp->b_data = bp->b_saveaddr; 1961 } 1962 return(bp); 1963} 1964 1965/* 1966 * buf_daemon: 1967 * 1968 * buffer flushing daemon. Buffers are normally flushed by the 1969 * update daemon but if it cannot keep up this process starts to 1970 * take the load in an attempt to prevent getnewbuf() from blocking. 1971 */ 1972 1973static struct kproc_desc buf_kp = { 1974 "bufdaemon", 1975 buf_daemon, 1976 &bufdaemonproc 1977}; 1978SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1979 1980static void 1981buf_daemon() 1982{ 1983 int s; 1984 1985 mtx_lock(&Giant); 1986 1987 /* 1988 * This process needs to be suspended prior to shutdown sync. 1989 */ 1990 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 1991 SHUTDOWN_PRI_LAST); 1992 1993 /* 1994 * This process is allowed to take the buffer cache to the limit 1995 */ 1996 s = splbio(); 1997 mtx_lock(&bdlock); 1998 1999 for (;;) { 2000 bd_request = 0; 2001 mtx_unlock(&bdlock); 2002 2003 kthread_suspend_check(bufdaemonproc); 2004 2005 /* 2006 * Do the flush. Limit the amount of in-transit I/O we 2007 * allow to build up, otherwise we would completely saturate 2008 * the I/O system. Wakeup any waiting processes before we 2009 * normally would so they can run in parallel with our drain. 2010 */ 2011 while (numdirtybuffers > lodirtybuffers) { 2012 if (flushbufqueues(0) == 0) { 2013 /* 2014 * Could not find any buffers without rollback 2015 * dependencies, so just write the first one 2016 * in the hopes of eventually making progress. 2017 */ 2018 flushbufqueues(1); 2019 break; 2020 } 2021 waitrunningbufspace(); 2022 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2023 } 2024 2025 /* 2026 * Only clear bd_request if we have reached our low water 2027 * mark. The buf_daemon normally waits 1 second and 2028 * then incrementally flushes any dirty buffers that have 2029 * built up, within reason. 2030 * 2031 * If we were unable to hit our low water mark and couldn't 2032 * find any flushable buffers, we sleep half a second. 2033 * Otherwise we loop immediately. 2034 */ 2035 mtx_lock(&bdlock); 2036 if (numdirtybuffers <= lodirtybuffers) { 2037 /* 2038 * We reached our low water mark, reset the 2039 * request and sleep until we are needed again. 2040 * The sleep is just so the suspend code works. 2041 */ 2042 bd_request = 0; 2043 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2044 } else { 2045 /* 2046 * We couldn't find any flushable dirty buffers but 2047 * still have too many dirty buffers, we 2048 * have to sleep and try again. (rare) 2049 */ 2050 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2051 } 2052 } 2053} 2054 2055/* 2056 * flushbufqueues: 2057 * 2058 * Try to flush a buffer in the dirty queue. We must be careful to 2059 * free up B_INVAL buffers instead of write them, which NFS is 2060 * particularly sensitive to. 2061 */ 2062static int flushwithdeps = 0; 2063SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2064 0, "Number of buffers flushed with dependecies that require rollbacks"); 2065 2066static int 2067flushbufqueues(int flushdeps) 2068{ 2069 struct thread *td = curthread; 2070 struct vnode *vp; 2071 struct mount *mp; 2072 struct buf *bp; 2073 int hasdeps; 2074 2075 mtx_lock(&bqlock); 2076 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2077 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2078 continue; 2079 BO_LOCK(bp->b_bufobj); 2080 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2081 (bp->b_flags & B_DELWRI) == 0) { 2082 BO_UNLOCK(bp->b_bufobj); 2083 BUF_UNLOCK(bp); 2084 continue; 2085 } 2086 BO_UNLOCK(bp->b_bufobj); 2087 if (bp->b_flags & B_INVAL) { 2088 bremfreel(bp); 2089 mtx_unlock(&bqlock); 2090 brelse(bp); 2091 return (1); 2092 } 2093 2094 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2095 if (flushdeps == 0) { 2096 BUF_UNLOCK(bp); 2097 continue; 2098 } 2099 hasdeps = 1; 2100 } else 2101 hasdeps = 0; 2102 /* 2103 * We must hold the lock on a vnode before writing 2104 * one of its buffers. Otherwise we may confuse, or 2105 * in the case of a snapshot vnode, deadlock the 2106 * system. 2107 * 2108 * The lock order here is the reverse of the normal 2109 * of vnode followed by buf lock. This is ok because 2110 * the NOWAIT will prevent deadlock. 2111 */ 2112 vp = bp->b_vp; 2113 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2114 BUF_UNLOCK(bp); 2115 continue; 2116 } 2117 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2118 mtx_unlock(&bqlock); 2119 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2120 bp, bp->b_vp, bp->b_flags); 2121 vfs_bio_awrite(bp); 2122 vn_finished_write(mp); 2123 VOP_UNLOCK(vp, 0, td); 2124 flushwithdeps += hasdeps; 2125 return (1); 2126 } 2127 vn_finished_write(mp); 2128 BUF_UNLOCK(bp); 2129 } 2130 mtx_unlock(&bqlock); 2131 return (0); 2132} 2133 2134/* 2135 * Check to see if a block is currently memory resident. 2136 */ 2137struct buf * 2138incore(struct bufobj *bo, daddr_t blkno) 2139{ 2140 struct buf *bp; 2141 2142 int s = splbio(); 2143 BO_LOCK(bo); 2144 bp = gbincore(bo, blkno); 2145 BO_UNLOCK(bo); 2146 splx(s); 2147 return (bp); 2148} 2149 2150/* 2151 * Returns true if no I/O is needed to access the 2152 * associated VM object. This is like incore except 2153 * it also hunts around in the VM system for the data. 2154 */ 2155 2156static int 2157inmem(struct vnode * vp, daddr_t blkno) 2158{ 2159 vm_object_t obj; 2160 vm_offset_t toff, tinc, size; 2161 vm_page_t m; 2162 vm_ooffset_t off; 2163 2164 ASSERT_VOP_LOCKED(vp, "inmem"); 2165 2166 if (incore(&vp->v_bufobj, blkno)) 2167 return 1; 2168 if (vp->v_mount == NULL) 2169 return 0; 2170 obj = vp->v_object; 2171 if (obj == NULL) 2172 return (0); 2173 2174 size = PAGE_SIZE; 2175 if (size > vp->v_mount->mnt_stat.f_iosize) 2176 size = vp->v_mount->mnt_stat.f_iosize; 2177 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2178 2179 VM_OBJECT_LOCK(obj); 2180 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2181 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2182 if (!m) 2183 goto notinmem; 2184 tinc = size; 2185 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2186 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2187 if (vm_page_is_valid(m, 2188 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2189 goto notinmem; 2190 } 2191 VM_OBJECT_UNLOCK(obj); 2192 return 1; 2193 2194notinmem: 2195 VM_OBJECT_UNLOCK(obj); 2196 return (0); 2197} 2198 2199/* 2200 * vfs_setdirty: 2201 * 2202 * Sets the dirty range for a buffer based on the status of the dirty 2203 * bits in the pages comprising the buffer. 2204 * 2205 * The range is limited to the size of the buffer. 2206 * 2207 * This routine is primarily used by NFS, but is generalized for the 2208 * B_VMIO case. 2209 */ 2210static void 2211vfs_setdirty(struct buf *bp) 2212{ 2213 int i; 2214 vm_object_t object; 2215 2216 /* 2217 * Degenerate case - empty buffer 2218 */ 2219 2220 if (bp->b_bufsize == 0) 2221 return; 2222 2223 /* 2224 * We qualify the scan for modified pages on whether the 2225 * object has been flushed yet. The OBJ_WRITEABLE flag 2226 * is not cleared simply by protecting pages off. 2227 */ 2228 2229 if ((bp->b_flags & B_VMIO) == 0) 2230 return; 2231 2232 object = bp->b_pages[0]->object; 2233 VM_OBJECT_LOCK(object); 2234 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2235 printf("Warning: object %p writeable but not mightbedirty\n", object); 2236 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2237 printf("Warning: object %p mightbedirty but not writeable\n", object); 2238 2239 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2240 vm_offset_t boffset; 2241 vm_offset_t eoffset; 2242 2243 vm_page_lock_queues(); 2244 /* 2245 * test the pages to see if they have been modified directly 2246 * by users through the VM system. 2247 */ 2248 for (i = 0; i < bp->b_npages; i++) 2249 vm_page_test_dirty(bp->b_pages[i]); 2250 2251 /* 2252 * Calculate the encompassing dirty range, boffset and eoffset, 2253 * (eoffset - boffset) bytes. 2254 */ 2255 2256 for (i = 0; i < bp->b_npages; i++) { 2257 if (bp->b_pages[i]->dirty) 2258 break; 2259 } 2260 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2261 2262 for (i = bp->b_npages - 1; i >= 0; --i) { 2263 if (bp->b_pages[i]->dirty) { 2264 break; 2265 } 2266 } 2267 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2268 2269 vm_page_unlock_queues(); 2270 /* 2271 * Fit it to the buffer. 2272 */ 2273 2274 if (eoffset > bp->b_bcount) 2275 eoffset = bp->b_bcount; 2276 2277 /* 2278 * If we have a good dirty range, merge with the existing 2279 * dirty range. 2280 */ 2281 2282 if (boffset < eoffset) { 2283 if (bp->b_dirtyoff > boffset) 2284 bp->b_dirtyoff = boffset; 2285 if (bp->b_dirtyend < eoffset) 2286 bp->b_dirtyend = eoffset; 2287 } 2288 } 2289 VM_OBJECT_UNLOCK(object); 2290} 2291 2292/* 2293 * getblk: 2294 * 2295 * Get a block given a specified block and offset into a file/device. 2296 * The buffers B_DONE bit will be cleared on return, making it almost 2297 * ready for an I/O initiation. B_INVAL may or may not be set on 2298 * return. The caller should clear B_INVAL prior to initiating a 2299 * READ. 2300 * 2301 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2302 * an existing buffer. 2303 * 2304 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2305 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2306 * and then cleared based on the backing VM. If the previous buffer is 2307 * non-0-sized but invalid, B_CACHE will be cleared. 2308 * 2309 * If getblk() must create a new buffer, the new buffer is returned with 2310 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2311 * case it is returned with B_INVAL clear and B_CACHE set based on the 2312 * backing VM. 2313 * 2314 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2315 * B_CACHE bit is clear. 2316 * 2317 * What this means, basically, is that the caller should use B_CACHE to 2318 * determine whether the buffer is fully valid or not and should clear 2319 * B_INVAL prior to issuing a read. If the caller intends to validate 2320 * the buffer by loading its data area with something, the caller needs 2321 * to clear B_INVAL. If the caller does this without issuing an I/O, 2322 * the caller should set B_CACHE ( as an optimization ), else the caller 2323 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2324 * a write attempt or if it was a successfull read. If the caller 2325 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2326 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2327 */ 2328struct buf * 2329getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2330 int flags) 2331{ 2332 struct buf *bp; 2333 struct bufobj *bo; 2334 int s; 2335 int error; 2336 2337 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2338 ASSERT_VOP_LOCKED(vp, "getblk"); 2339 if (size > MAXBSIZE) 2340 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2341 2342 bo = &vp->v_bufobj; 2343 s = splbio(); 2344loop: 2345 /* 2346 * Block if we are low on buffers. Certain processes are allowed 2347 * to completely exhaust the buffer cache. 2348 * 2349 * If this check ever becomes a bottleneck it may be better to 2350 * move it into the else, when gbincore() fails. At the moment 2351 * it isn't a problem. 2352 * 2353 * XXX remove if 0 sections (clean this up after its proven) 2354 */ 2355 if (numfreebuffers == 0) { 2356 if (curthread == PCPU_GET(idlethread)) 2357 return NULL; 2358 mtx_lock(&nblock); 2359 needsbuffer |= VFS_BIO_NEED_ANY; 2360 mtx_unlock(&nblock); 2361 } 2362 2363 VI_LOCK(vp); 2364 bp = gbincore(bo, blkno); 2365 if (bp != NULL) { 2366 int lockflags; 2367 /* 2368 * Buffer is in-core. If the buffer is not busy, it must 2369 * be on a queue. 2370 */ 2371 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2372 2373 if (flags & GB_LOCK_NOWAIT) 2374 lockflags |= LK_NOWAIT; 2375 2376 error = BUF_TIMELOCK(bp, lockflags, 2377 VI_MTX(vp), "getblk", slpflag, slptimeo); 2378 2379 /* 2380 * If we slept and got the lock we have to restart in case 2381 * the buffer changed identities. 2382 */ 2383 if (error == ENOLCK) 2384 goto loop; 2385 /* We timed out or were interrupted. */ 2386 else if (error) 2387 return (NULL); 2388 2389 /* 2390 * The buffer is locked. B_CACHE is cleared if the buffer is 2391 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2392 * and for a VMIO buffer B_CACHE is adjusted according to the 2393 * backing VM cache. 2394 */ 2395 if (bp->b_flags & B_INVAL) 2396 bp->b_flags &= ~B_CACHE; 2397 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2398 bp->b_flags |= B_CACHE; 2399 bremfree(bp); 2400 2401 /* 2402 * check for size inconsistancies for non-VMIO case. 2403 */ 2404 2405 if (bp->b_bcount != size) { 2406 if ((bp->b_flags & B_VMIO) == 0 || 2407 (size > bp->b_kvasize)) { 2408 if (bp->b_flags & B_DELWRI) { 2409 bp->b_flags |= B_NOCACHE; 2410 bwrite(bp); 2411 } else { 2412 if ((bp->b_flags & B_VMIO) && 2413 (LIST_FIRST(&bp->b_dep) == NULL)) { 2414 bp->b_flags |= B_RELBUF; 2415 brelse(bp); 2416 } else { 2417 bp->b_flags |= B_NOCACHE; 2418 bwrite(bp); 2419 } 2420 } 2421 goto loop; 2422 } 2423 } 2424 2425 /* 2426 * If the size is inconsistant in the VMIO case, we can resize 2427 * the buffer. This might lead to B_CACHE getting set or 2428 * cleared. If the size has not changed, B_CACHE remains 2429 * unchanged from its previous state. 2430 */ 2431 2432 if (bp->b_bcount != size) 2433 allocbuf(bp, size); 2434 2435 KASSERT(bp->b_offset != NOOFFSET, 2436 ("getblk: no buffer offset")); 2437 2438 /* 2439 * A buffer with B_DELWRI set and B_CACHE clear must 2440 * be committed before we can return the buffer in 2441 * order to prevent the caller from issuing a read 2442 * ( due to B_CACHE not being set ) and overwriting 2443 * it. 2444 * 2445 * Most callers, including NFS and FFS, need this to 2446 * operate properly either because they assume they 2447 * can issue a read if B_CACHE is not set, or because 2448 * ( for example ) an uncached B_DELWRI might loop due 2449 * to softupdates re-dirtying the buffer. In the latter 2450 * case, B_CACHE is set after the first write completes, 2451 * preventing further loops. 2452 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2453 * above while extending the buffer, we cannot allow the 2454 * buffer to remain with B_CACHE set after the write 2455 * completes or it will represent a corrupt state. To 2456 * deal with this we set B_NOCACHE to scrap the buffer 2457 * after the write. 2458 * 2459 * We might be able to do something fancy, like setting 2460 * B_CACHE in bwrite() except if B_DELWRI is already set, 2461 * so the below call doesn't set B_CACHE, but that gets real 2462 * confusing. This is much easier. 2463 */ 2464 2465 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2466 bp->b_flags |= B_NOCACHE; 2467 bwrite(bp); 2468 goto loop; 2469 } 2470 2471 splx(s); 2472 bp->b_flags &= ~B_DONE; 2473 } else { 2474 int bsize, maxsize, vmio; 2475 off_t offset; 2476 2477 /* 2478 * Buffer is not in-core, create new buffer. The buffer 2479 * returned by getnewbuf() is locked. Note that the returned 2480 * buffer is also considered valid (not marked B_INVAL). 2481 */ 2482 VI_UNLOCK(vp); 2483 /* 2484 * If the user does not want us to create the buffer, bail out 2485 * here. 2486 */ 2487 if (flags & GB_NOCREAT) { 2488 splx(s); 2489 return NULL; 2490 } 2491 2492 bsize = bo->bo_bsize; 2493 offset = blkno * bsize; 2494 vmio = vp->v_object != NULL; 2495 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2496 maxsize = imax(maxsize, bsize); 2497 2498 bp = getnewbuf(slpflag, slptimeo, size, maxsize); 2499 if (bp == NULL) { 2500 if (slpflag || slptimeo) { 2501 splx(s); 2502 return NULL; 2503 } 2504 goto loop; 2505 } 2506 2507 /* 2508 * This code is used to make sure that a buffer is not 2509 * created while the getnewbuf routine is blocked. 2510 * This can be a problem whether the vnode is locked or not. 2511 * If the buffer is created out from under us, we have to 2512 * throw away the one we just created. There is now window 2513 * race because we are safely running at splbio() from the 2514 * point of the duplicate buffer creation through to here, 2515 * and we've locked the buffer. 2516 * 2517 * Note: this must occur before we associate the buffer 2518 * with the vp especially considering limitations in 2519 * the splay tree implementation when dealing with duplicate 2520 * lblkno's. 2521 */ 2522 BO_LOCK(bo); 2523 if (gbincore(bo, blkno)) { 2524 BO_UNLOCK(bo); 2525 bp->b_flags |= B_INVAL; 2526 brelse(bp); 2527 goto loop; 2528 } 2529 2530 /* 2531 * Insert the buffer into the hash, so that it can 2532 * be found by incore. 2533 */ 2534 bp->b_blkno = bp->b_lblkno = blkno; 2535 bp->b_offset = offset; 2536 2537 bgetvp(vp, bp); 2538 BO_UNLOCK(bo); 2539 2540 /* 2541 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2542 * buffer size starts out as 0, B_CACHE will be set by 2543 * allocbuf() for the VMIO case prior to it testing the 2544 * backing store for validity. 2545 */ 2546 2547 if (vmio) { 2548 bp->b_flags |= B_VMIO; 2549#if defined(VFS_BIO_DEBUG) 2550 if (vn_canvmio(vp) != TRUE) 2551 printf("getblk: VMIO on vnode type %d\n", 2552 vp->v_type); 2553#endif 2554 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2555 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2556 bp, vp->v_object, bp->b_bufobj->bo_object)); 2557 } else { 2558 bp->b_flags &= ~B_VMIO; 2559 KASSERT(bp->b_bufobj->bo_object == NULL, 2560 ("ARGH! has b_bufobj->bo_object %p %p\n", 2561 bp, bp->b_bufobj->bo_object)); 2562 } 2563 2564 allocbuf(bp, size); 2565 2566 splx(s); 2567 bp->b_flags &= ~B_DONE; 2568 } 2569 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2570 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2571 KASSERT(bp->b_bufobj == bo, 2572 ("wrong b_bufobj %p should be %p", bp->b_bufobj, bo)); 2573 return (bp); 2574} 2575 2576/* 2577 * Get an empty, disassociated buffer of given size. The buffer is initially 2578 * set to B_INVAL. 2579 */ 2580struct buf * 2581geteblk(int size) 2582{ 2583 struct buf *bp; 2584 int s; 2585 int maxsize; 2586 2587 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2588 2589 s = splbio(); 2590 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2591 continue; 2592 splx(s); 2593 allocbuf(bp, size); 2594 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2595 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2596 return (bp); 2597} 2598 2599 2600/* 2601 * This code constitutes the buffer memory from either anonymous system 2602 * memory (in the case of non-VMIO operations) or from an associated 2603 * VM object (in the case of VMIO operations). This code is able to 2604 * resize a buffer up or down. 2605 * 2606 * Note that this code is tricky, and has many complications to resolve 2607 * deadlock or inconsistant data situations. Tread lightly!!! 2608 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2609 * the caller. Calling this code willy nilly can result in the loss of data. 2610 * 2611 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2612 * B_CACHE for the non-VMIO case. 2613 */ 2614 2615int 2616allocbuf(struct buf *bp, int size) 2617{ 2618 int newbsize, mbsize; 2619 int i; 2620 2621 if (BUF_REFCNT(bp) == 0) 2622 panic("allocbuf: buffer not busy"); 2623 2624 if (bp->b_kvasize < size) 2625 panic("allocbuf: buffer too small"); 2626 2627 if ((bp->b_flags & B_VMIO) == 0) { 2628 caddr_t origbuf; 2629 int origbufsize; 2630 /* 2631 * Just get anonymous memory from the kernel. Don't 2632 * mess with B_CACHE. 2633 */ 2634 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2635 if (bp->b_flags & B_MALLOC) 2636 newbsize = mbsize; 2637 else 2638 newbsize = round_page(size); 2639 2640 if (newbsize < bp->b_bufsize) { 2641 /* 2642 * malloced buffers are not shrunk 2643 */ 2644 if (bp->b_flags & B_MALLOC) { 2645 if (newbsize) { 2646 bp->b_bcount = size; 2647 } else { 2648 free(bp->b_data, M_BIOBUF); 2649 if (bp->b_bufsize) { 2650 atomic_subtract_int( 2651 &bufmallocspace, 2652 bp->b_bufsize); 2653 bufspacewakeup(); 2654 bp->b_bufsize = 0; 2655 } 2656 bp->b_saveaddr = bp->b_kvabase; 2657 bp->b_data = bp->b_saveaddr; 2658 bp->b_bcount = 0; 2659 bp->b_flags &= ~B_MALLOC; 2660 } 2661 return 1; 2662 } 2663 vm_hold_free_pages( 2664 bp, 2665 (vm_offset_t) bp->b_data + newbsize, 2666 (vm_offset_t) bp->b_data + bp->b_bufsize); 2667 } else if (newbsize > bp->b_bufsize) { 2668 /* 2669 * We only use malloced memory on the first allocation. 2670 * and revert to page-allocated memory when the buffer 2671 * grows. 2672 */ 2673 /* 2674 * There is a potential smp race here that could lead 2675 * to bufmallocspace slightly passing the max. It 2676 * is probably extremely rare and not worth worrying 2677 * over. 2678 */ 2679 if ( (bufmallocspace < maxbufmallocspace) && 2680 (bp->b_bufsize == 0) && 2681 (mbsize <= PAGE_SIZE/2)) { 2682 2683 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2684 bp->b_bufsize = mbsize; 2685 bp->b_bcount = size; 2686 bp->b_flags |= B_MALLOC; 2687 atomic_add_int(&bufmallocspace, mbsize); 2688 return 1; 2689 } 2690 origbuf = NULL; 2691 origbufsize = 0; 2692 /* 2693 * If the buffer is growing on its other-than-first allocation, 2694 * then we revert to the page-allocation scheme. 2695 */ 2696 if (bp->b_flags & B_MALLOC) { 2697 origbuf = bp->b_data; 2698 origbufsize = bp->b_bufsize; 2699 bp->b_data = bp->b_kvabase; 2700 if (bp->b_bufsize) { 2701 atomic_subtract_int(&bufmallocspace, 2702 bp->b_bufsize); 2703 bufspacewakeup(); 2704 bp->b_bufsize = 0; 2705 } 2706 bp->b_flags &= ~B_MALLOC; 2707 newbsize = round_page(newbsize); 2708 } 2709 vm_hold_load_pages( 2710 bp, 2711 (vm_offset_t) bp->b_data + bp->b_bufsize, 2712 (vm_offset_t) bp->b_data + newbsize); 2713 if (origbuf) { 2714 bcopy(origbuf, bp->b_data, origbufsize); 2715 free(origbuf, M_BIOBUF); 2716 } 2717 } 2718 } else { 2719 int desiredpages; 2720 2721 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2722 desiredpages = (size == 0) ? 0 : 2723 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2724 2725 if (bp->b_flags & B_MALLOC) 2726 panic("allocbuf: VMIO buffer can't be malloced"); 2727 /* 2728 * Set B_CACHE initially if buffer is 0 length or will become 2729 * 0-length. 2730 */ 2731 if (size == 0 || bp->b_bufsize == 0) 2732 bp->b_flags |= B_CACHE; 2733 2734 if (newbsize < bp->b_bufsize) { 2735 /* 2736 * DEV_BSIZE aligned new buffer size is less then the 2737 * DEV_BSIZE aligned existing buffer size. Figure out 2738 * if we have to remove any pages. 2739 */ 2740 if (desiredpages < bp->b_npages) { 2741 vm_page_t m; 2742 2743 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2744 vm_page_lock_queues(); 2745 for (i = desiredpages; i < bp->b_npages; i++) { 2746 /* 2747 * the page is not freed here -- it 2748 * is the responsibility of 2749 * vnode_pager_setsize 2750 */ 2751 m = bp->b_pages[i]; 2752 KASSERT(m != bogus_page, 2753 ("allocbuf: bogus page found")); 2754 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2755 vm_page_lock_queues(); 2756 2757 bp->b_pages[i] = NULL; 2758 vm_page_unwire(m, 0); 2759 } 2760 vm_page_unlock_queues(); 2761 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2762 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2763 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2764 bp->b_npages = desiredpages; 2765 } 2766 } else if (size > bp->b_bcount) { 2767 /* 2768 * We are growing the buffer, possibly in a 2769 * byte-granular fashion. 2770 */ 2771 struct vnode *vp; 2772 vm_object_t obj; 2773 vm_offset_t toff; 2774 vm_offset_t tinc; 2775 2776 /* 2777 * Step 1, bring in the VM pages from the object, 2778 * allocating them if necessary. We must clear 2779 * B_CACHE if these pages are not valid for the 2780 * range covered by the buffer. 2781 */ 2782 2783 vp = bp->b_vp; 2784 obj = bp->b_bufobj->bo_object; 2785 2786 VM_OBJECT_LOCK(obj); 2787 while (bp->b_npages < desiredpages) { 2788 vm_page_t m; 2789 vm_pindex_t pi; 2790 2791 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2792 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2793 /* 2794 * note: must allocate system pages 2795 * since blocking here could intefere 2796 * with paging I/O, no matter which 2797 * process we are. 2798 */ 2799 m = vm_page_alloc(obj, pi, 2800 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2801 VM_ALLOC_WIRED); 2802 if (m == NULL) { 2803 atomic_add_int(&vm_pageout_deficit, 2804 desiredpages - bp->b_npages); 2805 VM_OBJECT_UNLOCK(obj); 2806 VM_WAIT; 2807 VM_OBJECT_LOCK(obj); 2808 } else { 2809 bp->b_flags &= ~B_CACHE; 2810 bp->b_pages[bp->b_npages] = m; 2811 ++bp->b_npages; 2812 } 2813 continue; 2814 } 2815 2816 /* 2817 * We found a page. If we have to sleep on it, 2818 * retry because it might have gotten freed out 2819 * from under us. 2820 * 2821 * We can only test PG_BUSY here. Blocking on 2822 * m->busy might lead to a deadlock: 2823 * 2824 * vm_fault->getpages->cluster_read->allocbuf 2825 * 2826 */ 2827 vm_page_lock_queues(); 2828 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2829 continue; 2830 2831 /* 2832 * We have a good page. Should we wakeup the 2833 * page daemon? 2834 */ 2835 if ((curproc != pageproc) && 2836 ((m->queue - m->pc) == PQ_CACHE) && 2837 ((cnt.v_free_count + cnt.v_cache_count) < 2838 (cnt.v_free_min + cnt.v_cache_min))) { 2839 pagedaemon_wakeup(); 2840 } 2841 vm_page_wire(m); 2842 vm_page_unlock_queues(); 2843 bp->b_pages[bp->b_npages] = m; 2844 ++bp->b_npages; 2845 } 2846 2847 /* 2848 * Step 2. We've loaded the pages into the buffer, 2849 * we have to figure out if we can still have B_CACHE 2850 * set. Note that B_CACHE is set according to the 2851 * byte-granular range ( bcount and size ), new the 2852 * aligned range ( newbsize ). 2853 * 2854 * The VM test is against m->valid, which is DEV_BSIZE 2855 * aligned. Needless to say, the validity of the data 2856 * needs to also be DEV_BSIZE aligned. Note that this 2857 * fails with NFS if the server or some other client 2858 * extends the file's EOF. If our buffer is resized, 2859 * B_CACHE may remain set! XXX 2860 */ 2861 2862 toff = bp->b_bcount; 2863 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2864 2865 while ((bp->b_flags & B_CACHE) && toff < size) { 2866 vm_pindex_t pi; 2867 2868 if (tinc > (size - toff)) 2869 tinc = size - toff; 2870 2871 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2872 PAGE_SHIFT; 2873 2874 vfs_buf_test_cache( 2875 bp, 2876 bp->b_offset, 2877 toff, 2878 tinc, 2879 bp->b_pages[pi] 2880 ); 2881 toff += tinc; 2882 tinc = PAGE_SIZE; 2883 } 2884 VM_OBJECT_UNLOCK(obj); 2885 2886 /* 2887 * Step 3, fixup the KVM pmap. Remember that 2888 * bp->b_data is relative to bp->b_offset, but 2889 * bp->b_offset may be offset into the first page. 2890 */ 2891 2892 bp->b_data = (caddr_t) 2893 trunc_page((vm_offset_t)bp->b_data); 2894 pmap_qenter( 2895 (vm_offset_t)bp->b_data, 2896 bp->b_pages, 2897 bp->b_npages 2898 ); 2899 2900 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2901 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2902 } 2903 } 2904 if (newbsize < bp->b_bufsize) 2905 bufspacewakeup(); 2906 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2907 bp->b_bcount = size; /* requested buffer size */ 2908 return 1; 2909} 2910 2911void 2912biodone(struct bio *bp) 2913{ 2914 2915 mtx_lock(&bdonelock); 2916 bp->bio_flags |= BIO_DONE; 2917 if (bp->bio_done == NULL) 2918 wakeup(bp); 2919 mtx_unlock(&bdonelock); 2920 if (bp->bio_done != NULL) 2921 bp->bio_done(bp); 2922} 2923 2924/* 2925 * Wait for a BIO to finish. 2926 * 2927 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2928 * case is not yet clear. 2929 */ 2930int 2931biowait(struct bio *bp, const char *wchan) 2932{ 2933 2934 mtx_lock(&bdonelock); 2935 while ((bp->bio_flags & BIO_DONE) == 0) 2936 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 2937 mtx_unlock(&bdonelock); 2938 if (bp->bio_error != 0) 2939 return (bp->bio_error); 2940 if (!(bp->bio_flags & BIO_ERROR)) 2941 return (0); 2942 return (EIO); 2943} 2944 2945void 2946biofinish(struct bio *bp, struct devstat *stat, int error) 2947{ 2948 2949 if (error) { 2950 bp->bio_error = error; 2951 bp->bio_flags |= BIO_ERROR; 2952 } 2953 if (stat != NULL) 2954 devstat_end_transaction_bio(stat, bp); 2955 biodone(bp); 2956} 2957 2958/* 2959 * bufwait: 2960 * 2961 * Wait for buffer I/O completion, returning error status. The buffer 2962 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 2963 * error and cleared. 2964 */ 2965int 2966bufwait(struct buf *bp) 2967{ 2968 int s; 2969 2970 s = splbio(); 2971 if (bp->b_iocmd == BIO_READ) 2972 bwait(bp, PRIBIO, "biord"); 2973 else 2974 bwait(bp, PRIBIO, "biowr"); 2975 splx(s); 2976 if (bp->b_flags & B_EINTR) { 2977 bp->b_flags &= ~B_EINTR; 2978 return (EINTR); 2979 } 2980 if (bp->b_ioflags & BIO_ERROR) { 2981 return (bp->b_error ? bp->b_error : EIO); 2982 } else { 2983 return (0); 2984 } 2985} 2986 2987 /* 2988 * Call back function from struct bio back up to struct buf. 2989 */ 2990static void 2991bufdonebio(struct bio *bip) 2992{ 2993 struct buf *bp; 2994 2995 bp = bip->bio_caller2; 2996 bp->b_resid = bp->b_bcount - bip->bio_completed; 2997 bp->b_resid = bip->bio_resid; /* XXX: remove */ 2998 bp->b_ioflags = bip->bio_flags; 2999 bp->b_error = bip->bio_error; 3000 if (bp->b_error) 3001 bp->b_ioflags |= BIO_ERROR; 3002 bufdone(bp); 3003 g_destroy_bio(bip); 3004} 3005 3006void 3007dev_strategy(struct cdev *dev, struct buf *bp) 3008{ 3009 struct cdevsw *csw; 3010 struct bio *bip; 3011 3012 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3013 panic("b_iocmd botch"); 3014 for (;;) { 3015 bip = g_new_bio(); 3016 if (bip != NULL) 3017 break; 3018 /* Try again later */ 3019 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3020 } 3021 bip->bio_cmd = bp->b_iocmd; 3022 bip->bio_offset = bp->b_iooffset; 3023 bip->bio_length = bp->b_bcount; 3024 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3025 bip->bio_data = bp->b_data; 3026 bip->bio_done = bufdonebio; 3027 bip->bio_caller2 = bp; 3028 bip->bio_dev = dev; 3029 KASSERT(dev->si_refcount > 0, 3030 ("dev_strategy on un-referenced struct cdev *(%s)", 3031 devtoname(dev))); 3032 csw = dev_refthread(dev); 3033 if (csw == NULL) { 3034 bp->b_error = ENXIO; 3035 bp->b_ioflags = BIO_ERROR; 3036 bufdone(bp); 3037 return; 3038 } 3039 (*csw->d_strategy)(bip); 3040 dev_relthread(dev); 3041} 3042 3043/* 3044 * bufdone: 3045 * 3046 * Finish I/O on a buffer, optionally calling a completion function. 3047 * This is usually called from an interrupt so process blocking is 3048 * not allowed. 3049 * 3050 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3051 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3052 * assuming B_INVAL is clear. 3053 * 3054 * For the VMIO case, we set B_CACHE if the op was a read and no 3055 * read error occured, or if the op was a write. B_CACHE is never 3056 * set if the buffer is invalid or otherwise uncacheable. 3057 * 3058 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3059 * initiator to leave B_INVAL set to brelse the buffer out of existance 3060 * in the biodone routine. 3061 */ 3062void 3063bufdone(struct buf *bp) 3064{ 3065 struct bufobj *dropobj; 3066 int s; 3067 void (*biodone)(struct buf *); 3068 3069 3070 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3071 s = splbio(); 3072 dropobj = NULL; 3073 3074 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3075 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3076 3077 runningbufwakeup(bp); 3078 if (bp->b_iocmd == BIO_WRITE) 3079 dropobj = bp->b_bufobj; 3080 /* call optional completion function if requested */ 3081 if (bp->b_iodone != NULL) { 3082 biodone = bp->b_iodone; 3083 bp->b_iodone = NULL; 3084 /* 3085 * Device drivers may or may not hold giant, hold it here 3086 * if we're calling into unknown code. 3087 */ 3088 mtx_lock(&Giant); 3089 bp->b_flags |= B_DONE; 3090 (*biodone) (bp); 3091 mtx_unlock(&Giant); 3092 if (dropobj) 3093 bufobj_wdrop(dropobj); 3094 splx(s); 3095 return; 3096 } 3097 if (LIST_FIRST(&bp->b_dep) != NULL) 3098 buf_complete(bp); 3099 3100 if (bp->b_flags & B_VMIO) { 3101 int i; 3102 vm_ooffset_t foff; 3103 vm_page_t m; 3104 vm_object_t obj; 3105 int iosize; 3106 struct vnode *vp = bp->b_vp; 3107 3108 obj = bp->b_bufobj->bo_object; 3109 3110#if defined(VFS_BIO_DEBUG) 3111 mp_fixme("usecount and vflag accessed without locks."); 3112 if (vp->v_usecount == 0) { 3113 panic("biodone: zero vnode ref count"); 3114 } 3115 3116 KASSERT(vp->v_object != NULL, 3117 ("biodone: vnode %p has no vm_object", vp)); 3118#endif 3119 3120 foff = bp->b_offset; 3121 KASSERT(bp->b_offset != NOOFFSET, 3122 ("biodone: no buffer offset")); 3123 3124 VM_OBJECT_LOCK(obj); 3125#if defined(VFS_BIO_DEBUG) 3126 if (obj->paging_in_progress < bp->b_npages) { 3127 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3128 obj->paging_in_progress, bp->b_npages); 3129 } 3130#endif 3131 3132 /* 3133 * Set B_CACHE if the op was a normal read and no error 3134 * occured. B_CACHE is set for writes in the b*write() 3135 * routines. 3136 */ 3137 iosize = bp->b_bcount - bp->b_resid; 3138 if (bp->b_iocmd == BIO_READ && 3139 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3140 !(bp->b_ioflags & BIO_ERROR)) { 3141 bp->b_flags |= B_CACHE; 3142 } 3143 vm_page_lock_queues(); 3144 for (i = 0; i < bp->b_npages; i++) { 3145 int bogusflag = 0; 3146 int resid; 3147 3148 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3149 if (resid > iosize) 3150 resid = iosize; 3151 3152 /* 3153 * cleanup bogus pages, restoring the originals 3154 */ 3155 m = bp->b_pages[i]; 3156 if (m == bogus_page) { 3157 bogusflag = 1; 3158 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3159 if (m == NULL) 3160 panic("biodone: page disappeared!"); 3161 bp->b_pages[i] = m; 3162 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3163 } 3164#if defined(VFS_BIO_DEBUG) 3165 if (OFF_TO_IDX(foff) != m->pindex) { 3166 printf( 3167"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3168 (intmax_t)foff, (uintmax_t)m->pindex); 3169 } 3170#endif 3171 3172 /* 3173 * In the write case, the valid and clean bits are 3174 * already changed correctly ( see bdwrite() ), so we 3175 * only need to do this here in the read case. 3176 */ 3177 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3178 vfs_page_set_valid(bp, foff, i, m); 3179 } 3180 3181 /* 3182 * when debugging new filesystems or buffer I/O methods, this 3183 * is the most common error that pops up. if you see this, you 3184 * have not set the page busy flag correctly!!! 3185 */ 3186 if (m->busy == 0) { 3187 printf("biodone: page busy < 0, " 3188 "pindex: %d, foff: 0x(%x,%x), " 3189 "resid: %d, index: %d\n", 3190 (int) m->pindex, (int)(foff >> 32), 3191 (int) foff & 0xffffffff, resid, i); 3192 if (!vn_isdisk(vp, NULL)) 3193 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3194 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3195 (intmax_t) bp->b_lblkno, 3196 bp->b_flags, bp->b_npages); 3197 else 3198 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3199 (intmax_t) bp->b_lblkno, 3200 bp->b_flags, bp->b_npages); 3201 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3202 (u_long)m->valid, (u_long)m->dirty, 3203 m->wire_count); 3204 panic("biodone: page busy < 0\n"); 3205 } 3206 vm_page_io_finish(m); 3207 vm_object_pip_subtract(obj, 1); 3208 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3209 iosize -= resid; 3210 } 3211 vm_page_unlock_queues(); 3212 vm_object_pip_wakeupn(obj, 0); 3213 VM_OBJECT_UNLOCK(obj); 3214 } 3215 3216 /* 3217 * For asynchronous completions, release the buffer now. The brelse 3218 * will do a wakeup there if necessary - so no need to do a wakeup 3219 * here in the async case. The sync case always needs to do a wakeup. 3220 */ 3221 3222 if (bp->b_flags & B_ASYNC) { 3223 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3224 brelse(bp); 3225 else 3226 bqrelse(bp); 3227 } else 3228 bdone(bp); 3229 if (dropobj) 3230 bufobj_wdrop(dropobj); 3231 splx(s); 3232} 3233 3234/* 3235 * This routine is called in lieu of iodone in the case of 3236 * incomplete I/O. This keeps the busy status for pages 3237 * consistant. 3238 */ 3239void 3240vfs_unbusy_pages(struct buf *bp) 3241{ 3242 int i; 3243 vm_object_t obj; 3244 vm_page_t m; 3245 3246 runningbufwakeup(bp); 3247 if (!(bp->b_flags & B_VMIO)) 3248 return; 3249 3250 obj = bp->b_bufobj->bo_object; 3251 VM_OBJECT_LOCK(obj); 3252 vm_page_lock_queues(); 3253 for (i = 0; i < bp->b_npages; i++) { 3254 m = bp->b_pages[i]; 3255 if (m == bogus_page) { 3256 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3257 if (!m) 3258 panic("vfs_unbusy_pages: page missing\n"); 3259 bp->b_pages[i] = m; 3260 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3261 bp->b_pages, bp->b_npages); 3262 } 3263 vm_object_pip_subtract(obj, 1); 3264 vm_page_io_finish(m); 3265 } 3266 vm_page_unlock_queues(); 3267 vm_object_pip_wakeupn(obj, 0); 3268 VM_OBJECT_UNLOCK(obj); 3269} 3270 3271/* 3272 * vfs_page_set_valid: 3273 * 3274 * Set the valid bits in a page based on the supplied offset. The 3275 * range is restricted to the buffer's size. 3276 * 3277 * This routine is typically called after a read completes. 3278 */ 3279static void 3280vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3281{ 3282 vm_ooffset_t soff, eoff; 3283 3284 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3285 /* 3286 * Start and end offsets in buffer. eoff - soff may not cross a 3287 * page boundry or cross the end of the buffer. The end of the 3288 * buffer, in this case, is our file EOF, not the allocation size 3289 * of the buffer. 3290 */ 3291 soff = off; 3292 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3293 if (eoff > bp->b_offset + bp->b_bcount) 3294 eoff = bp->b_offset + bp->b_bcount; 3295 3296 /* 3297 * Set valid range. This is typically the entire buffer and thus the 3298 * entire page. 3299 */ 3300 if (eoff > soff) { 3301 vm_page_set_validclean( 3302 m, 3303 (vm_offset_t) (soff & PAGE_MASK), 3304 (vm_offset_t) (eoff - soff) 3305 ); 3306 } 3307} 3308 3309/* 3310 * This routine is called before a device strategy routine. 3311 * It is used to tell the VM system that paging I/O is in 3312 * progress, and treat the pages associated with the buffer 3313 * almost as being PG_BUSY. Also the object paging_in_progress 3314 * flag is handled to make sure that the object doesn't become 3315 * inconsistant. 3316 * 3317 * Since I/O has not been initiated yet, certain buffer flags 3318 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3319 * and should be ignored. 3320 */ 3321void 3322vfs_busy_pages(struct buf *bp, int clear_modify) 3323{ 3324 int i, bogus; 3325 vm_object_t obj; 3326 vm_ooffset_t foff; 3327 vm_page_t m; 3328 3329 if (!(bp->b_flags & B_VMIO)) 3330 return; 3331 3332 obj = bp->b_bufobj->bo_object; 3333 foff = bp->b_offset; 3334 KASSERT(bp->b_offset != NOOFFSET, 3335 ("vfs_busy_pages: no buffer offset")); 3336 vfs_setdirty(bp); 3337 VM_OBJECT_LOCK(obj); 3338retry: 3339 vm_page_lock_queues(); 3340 for (i = 0; i < bp->b_npages; i++) { 3341 m = bp->b_pages[i]; 3342 3343 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3344 goto retry; 3345 } 3346 bogus = 0; 3347 for (i = 0; i < bp->b_npages; i++) { 3348 m = bp->b_pages[i]; 3349 3350 if ((bp->b_flags & B_CLUSTER) == 0) { 3351 vm_object_pip_add(obj, 1); 3352 vm_page_io_start(m); 3353 } 3354 /* 3355 * When readying a buffer for a read ( i.e 3356 * clear_modify == 0 ), it is important to do 3357 * bogus_page replacement for valid pages in 3358 * partially instantiated buffers. Partially 3359 * instantiated buffers can, in turn, occur when 3360 * reconstituting a buffer from its VM backing store 3361 * base. We only have to do this if B_CACHE is 3362 * clear ( which causes the I/O to occur in the 3363 * first place ). The replacement prevents the read 3364 * I/O from overwriting potentially dirty VM-backed 3365 * pages. XXX bogus page replacement is, uh, bogus. 3366 * It may not work properly with small-block devices. 3367 * We need to find a better way. 3368 */ 3369 pmap_remove_all(m); 3370 if (clear_modify) 3371 vfs_page_set_valid(bp, foff, i, m); 3372 else if (m->valid == VM_PAGE_BITS_ALL && 3373 (bp->b_flags & B_CACHE) == 0) { 3374 bp->b_pages[i] = bogus_page; 3375 bogus++; 3376 } 3377 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3378 } 3379 vm_page_unlock_queues(); 3380 VM_OBJECT_UNLOCK(obj); 3381 if (bogus) 3382 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3383 bp->b_pages, bp->b_npages); 3384} 3385 3386/* 3387 * Tell the VM system that the pages associated with this buffer 3388 * are clean. This is used for delayed writes where the data is 3389 * going to go to disk eventually without additional VM intevention. 3390 * 3391 * Note that while we only really need to clean through to b_bcount, we 3392 * just go ahead and clean through to b_bufsize. 3393 */ 3394static void 3395vfs_clean_pages(struct buf *bp) 3396{ 3397 int i; 3398 vm_ooffset_t foff, noff, eoff; 3399 vm_page_t m; 3400 3401 if (!(bp->b_flags & B_VMIO)) 3402 return; 3403 3404 foff = bp->b_offset; 3405 KASSERT(bp->b_offset != NOOFFSET, 3406 ("vfs_clean_pages: no buffer offset")); 3407 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3408 vm_page_lock_queues(); 3409 for (i = 0; i < bp->b_npages; i++) { 3410 m = bp->b_pages[i]; 3411 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3412 eoff = noff; 3413 3414 if (eoff > bp->b_offset + bp->b_bufsize) 3415 eoff = bp->b_offset + bp->b_bufsize; 3416 vfs_page_set_valid(bp, foff, i, m); 3417 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3418 foff = noff; 3419 } 3420 vm_page_unlock_queues(); 3421 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3422} 3423 3424/* 3425 * vfs_bio_set_validclean: 3426 * 3427 * Set the range within the buffer to valid and clean. The range is 3428 * relative to the beginning of the buffer, b_offset. Note that b_offset 3429 * itself may be offset from the beginning of the first page. 3430 * 3431 */ 3432 3433void 3434vfs_bio_set_validclean(struct buf *bp, int base, int size) 3435{ 3436 int i, n; 3437 vm_page_t m; 3438 3439 if (!(bp->b_flags & B_VMIO)) 3440 return; 3441 /* 3442 * Fixup base to be relative to beginning of first page. 3443 * Set initial n to be the maximum number of bytes in the 3444 * first page that can be validated. 3445 */ 3446 3447 base += (bp->b_offset & PAGE_MASK); 3448 n = PAGE_SIZE - (base & PAGE_MASK); 3449 3450 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3451 vm_page_lock_queues(); 3452 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3453 m = bp->b_pages[i]; 3454 if (n > size) 3455 n = size; 3456 vm_page_set_validclean(m, base & PAGE_MASK, n); 3457 base += n; 3458 size -= n; 3459 n = PAGE_SIZE; 3460 } 3461 vm_page_unlock_queues(); 3462 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3463} 3464 3465/* 3466 * vfs_bio_clrbuf: 3467 * 3468 * clear a buffer. This routine essentially fakes an I/O, so we need 3469 * to clear BIO_ERROR and B_INVAL. 3470 * 3471 * Note that while we only theoretically need to clear through b_bcount, 3472 * we go ahead and clear through b_bufsize. 3473 */ 3474 3475void 3476vfs_bio_clrbuf(struct buf *bp) 3477{ 3478 int i, j, mask = 0; 3479 caddr_t sa, ea; 3480 3481 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3482 clrbuf(bp); 3483 return; 3484 } 3485 3486 bp->b_flags &= ~B_INVAL; 3487 bp->b_ioflags &= ~BIO_ERROR; 3488 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3489 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3490 (bp->b_offset & PAGE_MASK) == 0) { 3491 if (bp->b_pages[0] == bogus_page) 3492 goto unlock; 3493 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3494 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3495 if ((bp->b_pages[0]->valid & mask) == mask) 3496 goto unlock; 3497 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3498 ((bp->b_pages[0]->valid & mask) == 0)) { 3499 bzero(bp->b_data, bp->b_bufsize); 3500 bp->b_pages[0]->valid |= mask; 3501 goto unlock; 3502 } 3503 } 3504 ea = sa = bp->b_data; 3505 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3506 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3507 ea = (caddr_t)(vm_offset_t)ulmin( 3508 (u_long)(vm_offset_t)ea, 3509 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3510 if (bp->b_pages[i] == bogus_page) 3511 continue; 3512 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3513 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3514 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3515 if ((bp->b_pages[i]->valid & mask) == mask) 3516 continue; 3517 if ((bp->b_pages[i]->valid & mask) == 0) { 3518 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3519 bzero(sa, ea - sa); 3520 } else { 3521 for (; sa < ea; sa += DEV_BSIZE, j++) { 3522 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3523 (bp->b_pages[i]->valid & (1 << j)) == 0) 3524 bzero(sa, DEV_BSIZE); 3525 } 3526 } 3527 bp->b_pages[i]->valid |= mask; 3528 } 3529unlock: 3530 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3531 bp->b_resid = 0; 3532} 3533 3534/* 3535 * vm_hold_load_pages and vm_hold_free_pages get pages into 3536 * a buffers address space. The pages are anonymous and are 3537 * not associated with a file object. 3538 */ 3539static void 3540vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3541{ 3542 vm_offset_t pg; 3543 vm_page_t p; 3544 int index; 3545 3546 to = round_page(to); 3547 from = round_page(from); 3548 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3549 3550 VM_OBJECT_LOCK(kernel_object); 3551 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3552tryagain: 3553 /* 3554 * note: must allocate system pages since blocking here 3555 * could intefere with paging I/O, no matter which 3556 * process we are. 3557 */ 3558 p = vm_page_alloc(kernel_object, 3559 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3560 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3561 if (!p) { 3562 atomic_add_int(&vm_pageout_deficit, 3563 (to - pg) >> PAGE_SHIFT); 3564 VM_OBJECT_UNLOCK(kernel_object); 3565 VM_WAIT; 3566 VM_OBJECT_LOCK(kernel_object); 3567 goto tryagain; 3568 } 3569 p->valid = VM_PAGE_BITS_ALL; 3570 pmap_qenter(pg, &p, 1); 3571 bp->b_pages[index] = p; 3572 } 3573 VM_OBJECT_UNLOCK(kernel_object); 3574 bp->b_npages = index; 3575} 3576 3577/* Return pages associated with this buf to the vm system */ 3578static void 3579vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3580{ 3581 vm_offset_t pg; 3582 vm_page_t p; 3583 int index, newnpages; 3584 3585 from = round_page(from); 3586 to = round_page(to); 3587 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3588 3589 VM_OBJECT_LOCK(kernel_object); 3590 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3591 p = bp->b_pages[index]; 3592 if (p && (index < bp->b_npages)) { 3593 if (p->busy) { 3594 printf( 3595 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3596 (intmax_t)bp->b_blkno, 3597 (intmax_t)bp->b_lblkno); 3598 } 3599 bp->b_pages[index] = NULL; 3600 pmap_qremove(pg, 1); 3601 vm_page_lock_queues(); 3602 vm_page_unwire(p, 0); 3603 vm_page_free(p); 3604 vm_page_unlock_queues(); 3605 } 3606 } 3607 VM_OBJECT_UNLOCK(kernel_object); 3608 bp->b_npages = newnpages; 3609} 3610 3611/* 3612 * Map an IO request into kernel virtual address space. 3613 * 3614 * All requests are (re)mapped into kernel VA space. 3615 * Notice that we use b_bufsize for the size of the buffer 3616 * to be mapped. b_bcount might be modified by the driver. 3617 * 3618 * Note that even if the caller determines that the address space should 3619 * be valid, a race or a smaller-file mapped into a larger space may 3620 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3621 * check the return value. 3622 */ 3623int 3624vmapbuf(struct buf *bp) 3625{ 3626 caddr_t addr, kva; 3627 vm_prot_t prot; 3628 int pidx, i; 3629 struct vm_page *m; 3630 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3631 3632 if (bp->b_bufsize < 0) 3633 return (-1); 3634 prot = VM_PROT_READ; 3635 if (bp->b_iocmd == BIO_READ) 3636 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3637 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3638 addr < bp->b_data + bp->b_bufsize; 3639 addr += PAGE_SIZE, pidx++) { 3640 /* 3641 * Do the vm_fault if needed; do the copy-on-write thing 3642 * when reading stuff off device into memory. 3643 * 3644 * NOTE! Must use pmap_extract() because addr may be in 3645 * the userland address space, and kextract is only guarenteed 3646 * to work for the kernland address space (see: sparc64 port). 3647 */ 3648retry: 3649 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3650 prot) < 0) { 3651 vm_page_lock_queues(); 3652 for (i = 0; i < pidx; ++i) { 3653 vm_page_unhold(bp->b_pages[i]); 3654 bp->b_pages[i] = NULL; 3655 } 3656 vm_page_unlock_queues(); 3657 return(-1); 3658 } 3659 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3660 if (m == NULL) 3661 goto retry; 3662 bp->b_pages[pidx] = m; 3663 } 3664 if (pidx > btoc(MAXPHYS)) 3665 panic("vmapbuf: mapped more than MAXPHYS"); 3666 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3667 3668 kva = bp->b_saveaddr; 3669 bp->b_npages = pidx; 3670 bp->b_saveaddr = bp->b_data; 3671 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3672 return(0); 3673} 3674 3675/* 3676 * Free the io map PTEs associated with this IO operation. 3677 * We also invalidate the TLB entries and restore the original b_addr. 3678 */ 3679void 3680vunmapbuf(struct buf *bp) 3681{ 3682 int pidx; 3683 int npages; 3684 3685 npages = bp->b_npages; 3686 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3687 vm_page_lock_queues(); 3688 for (pidx = 0; pidx < npages; pidx++) 3689 vm_page_unhold(bp->b_pages[pidx]); 3690 vm_page_unlock_queues(); 3691 3692 bp->b_data = bp->b_saveaddr; 3693} 3694 3695void 3696bdone(struct buf *bp) 3697{ 3698 3699 mtx_lock(&bdonelock); 3700 bp->b_flags |= B_DONE; 3701 wakeup(bp); 3702 mtx_unlock(&bdonelock); 3703} 3704 3705void 3706bwait(struct buf *bp, u_char pri, const char *wchan) 3707{ 3708 3709 mtx_lock(&bdonelock); 3710 while ((bp->b_flags & B_DONE) == 0) 3711 msleep(bp, &bdonelock, pri, wchan, 0); 3712 mtx_unlock(&bdonelock); 3713} 3714 3715int 3716bufsync(struct bufobj *bo, int waitfor, struct thread *td) 3717{ 3718 3719 return (VOP_FSYNC(bo->__bo_vnode, waitfor, td)); 3720} 3721 3722void 3723bufstrategy(struct bufobj *bo, struct buf *bp) 3724{ 3725 int i = 0; 3726 struct vnode *vp; 3727 3728 vp = bp->b_vp; 3729 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3730 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3731 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3732 i = VOP_STRATEGY(vp, bp); 3733 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3734} 3735 3736void 3737bufobj_wref(struct bufobj *bo) 3738{ 3739 3740 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3741 BO_LOCK(bo); 3742 bo->bo_numoutput++; 3743 BO_UNLOCK(bo); 3744} 3745 3746void 3747bufobj_wdrop(struct bufobj *bo) 3748{ 3749 3750 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3751 BO_LOCK(bo); 3752 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3753 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3754 bo->bo_flag &= ~BO_WWAIT; 3755 wakeup(&bo->bo_numoutput); 3756 } 3757 BO_UNLOCK(bo); 3758} 3759 3760int 3761bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3762{ 3763 int error; 3764 3765 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3766 ASSERT_BO_LOCKED(bo); 3767 error = 0; 3768 while (bo->bo_numoutput) { 3769 bo->bo_flag |= BO_WWAIT; 3770 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3771 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3772 if (error) 3773 break; 3774 } 3775 return (error); 3776} 3777 3778#include "opt_ddb.h" 3779#ifdef DDB 3780#include <ddb/ddb.h> 3781 3782/* DDB command to show buffer data */ 3783DB_SHOW_COMMAND(buffer, db_show_buffer) 3784{ 3785 /* get args */ 3786 struct buf *bp = (struct buf *)addr; 3787 3788 if (!have_addr) { 3789 db_printf("usage: show buffer <addr>\n"); 3790 return; 3791 } 3792 3793 db_printf("buf at %p\n", bp); 3794 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3795 db_printf( 3796 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3797 "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n", 3798 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3799 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno); 3800 db_printf("lockstatus = %d, excl count = %d, excl owner %p\n", 3801 lockstatus(&bp->b_lock, NULL), bp->b_lock.lk_exclusivecount, 3802 bp->b_lock.lk_lockholder); 3803 if (bp->b_npages) { 3804 int i; 3805 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3806 for (i = 0; i < bp->b_npages; i++) { 3807 vm_page_t m; 3808 m = bp->b_pages[i]; 3809 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3810 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3811 if ((i + 1) < bp->b_npages) 3812 db_printf(","); 3813 } 3814 db_printf("\n"); 3815 } 3816} 3817 3818DB_SHOW_COMMAND(lockedbufs, lockedbufs) 3819{ 3820 struct buf *bp; 3821 int i; 3822 3823 for (i = 0; i < nbuf; i++) { 3824 bp = &buf[i]; 3825 if (lockcount(&bp->b_lock)) { 3826 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 3827 db_printf("\n"); 3828 } 3829 } 3830} 3831#endif /* DDB */ 3832