vfs_bio.c revision 154695
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 154695 2006-01-23 00:00:45Z alc $"); 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, "biobuf", "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"); 130int 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 * Lock that protects against bwait()/bdone()/B_DONE races. 241 */ 242static struct mtx bpinlock; 243 244/* 245 * Definitions for the buffer free lists. 246 */ 247#define BUFFER_QUEUES 5 /* number of free buffer queues */ 248 249#define QUEUE_NONE 0 /* on no queue */ 250#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 251#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 252#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 253#define QUEUE_EMPTY 4 /* empty buffer headers */ 254 255/* Queues for free buffers with various properties */ 256static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 257 258/* Lock for the bufqueues */ 259static struct mtx bqlock; 260 261/* 262 * Single global constant for BUF_WMESG, to avoid getting multiple references. 263 * buf_wmesg is referred from macros. 264 */ 265const char *buf_wmesg = BUF_WMESG; 266 267#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 268#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 269#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 270#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 271 272#ifdef DIRECTIO 273extern void ffs_rawread_setup(void); 274#endif /* DIRECTIO */ 275/* 276 * numdirtywakeup: 277 * 278 * If someone is blocked due to there being too many dirty buffers, 279 * and numdirtybuffers is now reasonable, wake them up. 280 */ 281 282static __inline void 283numdirtywakeup(int level) 284{ 285 286 if (numdirtybuffers <= level) { 287 mtx_lock(&nblock); 288 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 289 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 290 wakeup(&needsbuffer); 291 } 292 mtx_unlock(&nblock); 293 } 294} 295 296/* 297 * bufspacewakeup: 298 * 299 * Called when buffer space is potentially available for recovery. 300 * getnewbuf() will block on this flag when it is unable to free 301 * sufficient buffer space. Buffer space becomes recoverable when 302 * bp's get placed back in the queues. 303 */ 304 305static __inline void 306bufspacewakeup(void) 307{ 308 309 /* 310 * If someone is waiting for BUF space, wake them up. Even 311 * though we haven't freed the kva space yet, the waiting 312 * process will be able to now. 313 */ 314 mtx_lock(&nblock); 315 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 316 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 317 wakeup(&needsbuffer); 318 } 319 mtx_unlock(&nblock); 320} 321 322/* 323 * runningbufwakeup() - in-progress I/O accounting. 324 * 325 */ 326void 327runningbufwakeup(struct buf *bp) 328{ 329 330 if (bp->b_runningbufspace) { 331 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 332 bp->b_runningbufspace = 0; 333 mtx_lock(&rbreqlock); 334 if (runningbufreq && runningbufspace <= lorunningspace) { 335 runningbufreq = 0; 336 wakeup(&runningbufreq); 337 } 338 mtx_unlock(&rbreqlock); 339 } 340} 341 342/* 343 * bufcountwakeup: 344 * 345 * Called when a buffer has been added to one of the free queues to 346 * account for the buffer and to wakeup anyone waiting for free buffers. 347 * This typically occurs when large amounts of metadata are being handled 348 * by the buffer cache ( else buffer space runs out first, usually ). 349 */ 350 351static __inline void 352bufcountwakeup(void) 353{ 354 355 atomic_add_int(&numfreebuffers, 1); 356 mtx_lock(&nblock); 357 if (needsbuffer) { 358 needsbuffer &= ~VFS_BIO_NEED_ANY; 359 if (numfreebuffers >= hifreebuffers) 360 needsbuffer &= ~VFS_BIO_NEED_FREE; 361 wakeup(&needsbuffer); 362 } 363 mtx_unlock(&nblock); 364} 365 366/* 367 * waitrunningbufspace() 368 * 369 * runningbufspace is a measure of the amount of I/O currently 370 * running. This routine is used in async-write situations to 371 * prevent creating huge backups of pending writes to a device. 372 * Only asynchronous writes are governed by this function. 373 * 374 * Reads will adjust runningbufspace, but will not block based on it. 375 * The read load has a side effect of reducing the allowed write load. 376 * 377 * This does NOT turn an async write into a sync write. It waits 378 * for earlier writes to complete and generally returns before the 379 * caller's write has reached the device. 380 */ 381void 382waitrunningbufspace(void) 383{ 384 385 mtx_lock(&rbreqlock); 386 while (runningbufspace > hirunningspace) { 387 ++runningbufreq; 388 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 389 } 390 mtx_unlock(&rbreqlock); 391} 392 393 394/* 395 * vfs_buf_test_cache: 396 * 397 * Called when a buffer is extended. This function clears the B_CACHE 398 * bit if the newly extended portion of the buffer does not contain 399 * valid data. 400 */ 401static __inline 402void 403vfs_buf_test_cache(struct buf *bp, 404 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 405 vm_page_t m) 406{ 407 408 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 409 if (bp->b_flags & B_CACHE) { 410 int base = (foff + off) & PAGE_MASK; 411 if (vm_page_is_valid(m, base, size) == 0) 412 bp->b_flags &= ~B_CACHE; 413 } 414} 415 416/* Wake up the buffer deamon if necessary */ 417static __inline 418void 419bd_wakeup(int dirtybuflevel) 420{ 421 422 mtx_lock(&bdlock); 423 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 424 bd_request = 1; 425 wakeup(&bd_request); 426 } 427 mtx_unlock(&bdlock); 428} 429 430/* 431 * bd_speedup - speedup the buffer cache flushing code 432 */ 433 434static __inline 435void 436bd_speedup(void) 437{ 438 439 bd_wakeup(1); 440} 441 442/* 443 * Calculating buffer cache scaling values and reserve space for buffer 444 * headers. This is called during low level kernel initialization and 445 * may be called more then once. We CANNOT write to the memory area 446 * being reserved at this time. 447 */ 448caddr_t 449kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 450{ 451 452 /* 453 * physmem_est is in pages. Convert it to kilobytes (assumes 454 * PAGE_SIZE is >= 1K) 455 */ 456 physmem_est = physmem_est * (PAGE_SIZE / 1024); 457 458 /* 459 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 460 * For the first 64MB of ram nominally allocate sufficient buffers to 461 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 462 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 463 * the buffer cache we limit the eventual kva reservation to 464 * maxbcache bytes. 465 * 466 * factor represents the 1/4 x ram conversion. 467 */ 468 if (nbuf == 0) { 469 int factor = 4 * BKVASIZE / 1024; 470 471 nbuf = 50; 472 if (physmem_est > 4096) 473 nbuf += min((physmem_est - 4096) / factor, 474 65536 / factor); 475 if (physmem_est > 65536) 476 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 477 478 if (maxbcache && nbuf > maxbcache / BKVASIZE) 479 nbuf = maxbcache / BKVASIZE; 480 } 481 482#if 0 483 /* 484 * Do not allow the buffer_map to be more then 1/2 the size of the 485 * kernel_map. 486 */ 487 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 488 (BKVASIZE * 2)) { 489 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 490 (BKVASIZE * 2); 491 printf("Warning: nbufs capped at %d\n", nbuf); 492 } 493#endif 494 495 /* 496 * swbufs are used as temporary holders for I/O, such as paging I/O. 497 * We have no less then 16 and no more then 256. 498 */ 499 nswbuf = max(min(nbuf/4, 256), 16); 500#ifdef NSWBUF_MIN 501 if (nswbuf < NSWBUF_MIN) 502 nswbuf = NSWBUF_MIN; 503#endif 504#ifdef DIRECTIO 505 ffs_rawread_setup(); 506#endif 507 508 /* 509 * Reserve space for the buffer cache buffers 510 */ 511 swbuf = (void *)v; 512 v = (caddr_t)(swbuf + nswbuf); 513 buf = (void *)v; 514 v = (caddr_t)(buf + nbuf); 515 516 return(v); 517} 518 519/* Initialize the buffer subsystem. Called before use of any buffers. */ 520void 521bufinit(void) 522{ 523 struct buf *bp; 524 int i; 525 526 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 527 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 528 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 529 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 530 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 531 mtx_init(&bpinlock, "bpin lock", NULL, MTX_DEF); 532 533 /* next, make a null set of free lists */ 534 for (i = 0; i < BUFFER_QUEUES; i++) 535 TAILQ_INIT(&bufqueues[i]); 536 537 /* finally, initialize each buffer header and stick on empty q */ 538 for (i = 0; i < nbuf; i++) { 539 bp = &buf[i]; 540 bzero(bp, sizeof *bp); 541 bp->b_flags = B_INVAL; /* we're just an empty header */ 542 bp->b_rcred = NOCRED; 543 bp->b_wcred = NOCRED; 544 bp->b_qindex = QUEUE_EMPTY; 545 bp->b_vflags = 0; 546 bp->b_xflags = 0; 547 LIST_INIT(&bp->b_dep); 548 BUF_LOCKINIT(bp); 549 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 550 } 551 552 /* 553 * maxbufspace is the absolute maximum amount of buffer space we are 554 * allowed to reserve in KVM and in real terms. The absolute maximum 555 * is nominally used by buf_daemon. hibufspace is the nominal maximum 556 * used by most other processes. The differential is required to 557 * ensure that buf_daemon is able to run when other processes might 558 * be blocked waiting for buffer space. 559 * 560 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 561 * this may result in KVM fragmentation which is not handled optimally 562 * by the system. 563 */ 564 maxbufspace = nbuf * BKVASIZE; 565 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 566 lobufspace = hibufspace - MAXBSIZE; 567 568 lorunningspace = 512 * 1024; 569 hirunningspace = 1024 * 1024; 570 571/* 572 * Limit the amount of malloc memory since it is wired permanently into 573 * the kernel space. Even though this is accounted for in the buffer 574 * allocation, we don't want the malloced region to grow uncontrolled. 575 * The malloc scheme improves memory utilization significantly on average 576 * (small) directories. 577 */ 578 maxbufmallocspace = hibufspace / 20; 579 580/* 581 * Reduce the chance of a deadlock occuring by limiting the number 582 * of delayed-write dirty buffers we allow to stack up. 583 */ 584 hidirtybuffers = nbuf / 4 + 20; 585 dirtybufthresh = hidirtybuffers * 9 / 10; 586 numdirtybuffers = 0; 587/* 588 * To support extreme low-memory systems, make sure hidirtybuffers cannot 589 * eat up all available buffer space. This occurs when our minimum cannot 590 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 591 * BKVASIZE'd (8K) buffers. 592 */ 593 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 594 hidirtybuffers >>= 1; 595 } 596 lodirtybuffers = hidirtybuffers / 2; 597 598/* 599 * Try to keep the number of free buffers in the specified range, 600 * and give special processes (e.g. like buf_daemon) access to an 601 * emergency reserve. 602 */ 603 lofreebuffers = nbuf / 18 + 5; 604 hifreebuffers = 2 * lofreebuffers; 605 numfreebuffers = nbuf; 606 607/* 608 * Maximum number of async ops initiated per buf_daemon loop. This is 609 * somewhat of a hack at the moment, we really need to limit ourselves 610 * based on the number of bytes of I/O in-transit that were initiated 611 * from buf_daemon. 612 */ 613 614 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 615 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 616} 617 618/* 619 * bfreekva() - free the kva allocation for a buffer. 620 * 621 * Since this call frees up buffer space, we call bufspacewakeup(). 622 */ 623static void 624bfreekva(struct buf *bp) 625{ 626 627 if (bp->b_kvasize) { 628 atomic_add_int(&buffreekvacnt, 1); 629 atomic_subtract_int(&bufspace, bp->b_kvasize); 630 vm_map_lock(buffer_map); 631 vm_map_delete(buffer_map, 632 (vm_offset_t) bp->b_kvabase, 633 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 634 ); 635 vm_map_unlock(buffer_map); 636 bp->b_kvasize = 0; 637 bufspacewakeup(); 638 } 639} 640 641/* 642 * bremfree: 643 * 644 * Mark the buffer for removal from the appropriate free list in brelse. 645 * 646 */ 647void 648bremfree(struct buf *bp) 649{ 650 651 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 652 KASSERT(BUF_REFCNT(bp), ("bremfree: buf must be locked.")); 653 KASSERT((bp->b_flags & B_REMFREE) == 0, 654 ("bremfree: buffer %p already marked for delayed removal.", bp)); 655 KASSERT(bp->b_qindex != QUEUE_NONE, 656 ("bremfree: buffer %p not on a queue.", bp)); 657 658 bp->b_flags |= B_REMFREE; 659 /* Fixup numfreebuffers count. */ 660 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 661 atomic_subtract_int(&numfreebuffers, 1); 662} 663 664/* 665 * bremfreef: 666 * 667 * Force an immediate removal from a free list. Used only in nfs when 668 * it abuses the b_freelist pointer. 669 */ 670void 671bremfreef(struct buf *bp) 672{ 673 mtx_lock(&bqlock); 674 bremfreel(bp); 675 mtx_unlock(&bqlock); 676} 677 678/* 679 * bremfreel: 680 * 681 * Removes a buffer from the free list, must be called with the 682 * bqlock held. 683 */ 684static void 685bremfreel(struct buf *bp) 686{ 687 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 688 bp, bp->b_vp, bp->b_flags); 689 KASSERT(BUF_REFCNT(bp), ("bremfreel: buffer %p not locked.", bp)); 690 KASSERT(bp->b_qindex != QUEUE_NONE, 691 ("bremfreel: buffer %p not on a queue.", bp)); 692 mtx_assert(&bqlock, MA_OWNED); 693 694 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 695 bp->b_qindex = QUEUE_NONE; 696 /* 697 * If this was a delayed bremfree() we only need to remove the buffer 698 * from the queue and return the stats are already done. 699 */ 700 if (bp->b_flags & B_REMFREE) { 701 bp->b_flags &= ~B_REMFREE; 702 return; 703 } 704 /* 705 * Fixup numfreebuffers count. If the buffer is invalid or not 706 * delayed-write, the buffer was free and we must decrement 707 * numfreebuffers. 708 */ 709 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 710 atomic_subtract_int(&numfreebuffers, 1); 711} 712 713 714/* 715 * Get a buffer with the specified data. Look in the cache first. We 716 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 717 * is set, the buffer is valid and we do not have to do anything ( see 718 * getblk() ). This is really just a special case of breadn(). 719 */ 720int 721bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 722 struct buf **bpp) 723{ 724 725 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 726} 727 728/* 729 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 730 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 731 * the buffer is valid and we do not have to do anything. 732 */ 733void 734breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 735 int cnt, struct ucred * cred) 736{ 737 struct buf *rabp; 738 int i; 739 740 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 741 if (inmem(vp, *rablkno)) 742 continue; 743 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 744 745 if ((rabp->b_flags & B_CACHE) == 0) { 746 if (curthread != PCPU_GET(idlethread)) 747 curthread->td_proc->p_stats->p_ru.ru_inblock++; 748 rabp->b_flags |= B_ASYNC; 749 rabp->b_flags &= ~B_INVAL; 750 rabp->b_ioflags &= ~BIO_ERROR; 751 rabp->b_iocmd = BIO_READ; 752 if (rabp->b_rcred == NOCRED && cred != NOCRED) 753 rabp->b_rcred = crhold(cred); 754 vfs_busy_pages(rabp, 0); 755 BUF_KERNPROC(rabp); 756 rabp->b_iooffset = dbtob(rabp->b_blkno); 757 bstrategy(rabp); 758 } else { 759 brelse(rabp); 760 } 761 } 762} 763 764/* 765 * Operates like bread, but also starts asynchronous I/O on 766 * read-ahead blocks. 767 */ 768int 769breadn(struct vnode * vp, daddr_t blkno, int size, 770 daddr_t * rablkno, int *rabsize, 771 int cnt, struct ucred * cred, struct buf **bpp) 772{ 773 struct buf *bp; 774 int rv = 0, readwait = 0; 775 776 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 777 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 778 779 /* if not found in cache, do some I/O */ 780 if ((bp->b_flags & B_CACHE) == 0) { 781 if (curthread != PCPU_GET(idlethread)) 782 curthread->td_proc->p_stats->p_ru.ru_inblock++; 783 bp->b_iocmd = BIO_READ; 784 bp->b_flags &= ~B_INVAL; 785 bp->b_ioflags &= ~BIO_ERROR; 786 if (bp->b_rcred == NOCRED && cred != NOCRED) 787 bp->b_rcred = crhold(cred); 788 vfs_busy_pages(bp, 0); 789 bp->b_iooffset = dbtob(bp->b_blkno); 790 bstrategy(bp); 791 ++readwait; 792 } 793 794 breada(vp, rablkno, rabsize, cnt, cred); 795 796 if (readwait) { 797 rv = bufwait(bp); 798 } 799 return (rv); 800} 801 802/* 803 * Write, release buffer on completion. (Done by iodone 804 * if async). Do not bother writing anything if the buffer 805 * is invalid. 806 * 807 * Note that we set B_CACHE here, indicating that buffer is 808 * fully valid and thus cacheable. This is true even of NFS 809 * now so we set it generally. This could be set either here 810 * or in biodone() since the I/O is synchronous. We put it 811 * here. 812 */ 813int 814bufwrite(struct buf *bp) 815{ 816 int oldflags; 817 818 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 819 if (bp->b_flags & B_INVAL) { 820 brelse(bp); 821 return (0); 822 } 823 824 oldflags = bp->b_flags; 825 826 if (BUF_REFCNT(bp) == 0) 827 panic("bufwrite: buffer is not busy???"); 828 829 if (bp->b_pin_count > 0) 830 bunpin_wait(bp); 831 832 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 833 ("FFS background buffer should not get here %p", bp)); 834 835 /* Mark the buffer clean */ 836 bundirty(bp); 837 838 bp->b_flags &= ~B_DONE; 839 bp->b_ioflags &= ~BIO_ERROR; 840 bp->b_flags |= B_CACHE; 841 bp->b_iocmd = BIO_WRITE; 842 843 bufobj_wref(bp->b_bufobj); 844 vfs_busy_pages(bp, 1); 845 846 /* 847 * Normal bwrites pipeline writes 848 */ 849 bp->b_runningbufspace = bp->b_bufsize; 850 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 851 852 if (curthread != PCPU_GET(idlethread)) 853 curthread->td_proc->p_stats->p_ru.ru_oublock++; 854 if (oldflags & B_ASYNC) 855 BUF_KERNPROC(bp); 856 bp->b_iooffset = dbtob(bp->b_blkno); 857 bstrategy(bp); 858 859 if ((oldflags & B_ASYNC) == 0) { 860 int rtval = bufwait(bp); 861 brelse(bp); 862 return (rtval); 863 } else { 864 /* 865 * don't allow the async write to saturate the I/O 866 * system. We will not deadlock here because 867 * we are blocking waiting for I/O that is already in-progress 868 * to complete. We do not block here if it is the update 869 * or syncer daemon trying to clean up as that can lead 870 * to deadlock. 871 */ 872 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0) 873 waitrunningbufspace(); 874 } 875 876 return (0); 877} 878 879/* 880 * Delayed write. (Buffer is marked dirty). Do not bother writing 881 * anything if the buffer is marked invalid. 882 * 883 * Note that since the buffer must be completely valid, we can safely 884 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 885 * biodone() in order to prevent getblk from writing the buffer 886 * out synchronously. 887 */ 888void 889bdwrite(struct buf *bp) 890{ 891 struct thread *td = curthread; 892 struct vnode *vp; 893 struct buf *nbp; 894 struct bufobj *bo; 895 896 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 897 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 898 KASSERT(BUF_REFCNT(bp) != 0, ("bdwrite: buffer is not busy")); 899 900 if (bp->b_flags & B_INVAL) { 901 brelse(bp); 902 return; 903 } 904 905 /* 906 * If we have too many dirty buffers, don't create any more. 907 * If we are wildly over our limit, then force a complete 908 * cleanup. Otherwise, just keep the situation from getting 909 * out of control. Note that we have to avoid a recursive 910 * disaster and not try to clean up after our own cleanup! 911 */ 912 vp = bp->b_vp; 913 bo = bp->b_bufobj; 914 if ((td->td_pflags & TDP_COWINPROGRESS) == 0) { 915 BO_LOCK(bo); 916 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 917 BO_UNLOCK(bo); 918 (void) VOP_FSYNC(vp, MNT_NOWAIT, td); 919 altbufferflushes++; 920 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 921 /* 922 * Try to find a buffer to flush. 923 */ 924 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 925 if ((nbp->b_vflags & BV_BKGRDINPROG) || 926 BUF_LOCK(nbp, 927 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 928 continue; 929 if (bp == nbp) 930 panic("bdwrite: found ourselves"); 931 BO_UNLOCK(bo); 932 /* Don't countdeps with the bo lock held. */ 933 if (buf_countdeps(nbp, 0)) { 934 BO_LOCK(bo); 935 BUF_UNLOCK(nbp); 936 continue; 937 } 938 if (nbp->b_flags & B_CLUSTEROK) { 939 vfs_bio_awrite(nbp); 940 } else { 941 bremfree(nbp); 942 bawrite(nbp); 943 } 944 dirtybufferflushes++; 945 break; 946 } 947 if (nbp == NULL) 948 BO_UNLOCK(bo); 949 } else 950 BO_UNLOCK(bo); 951 } else 952 recursiveflushes++; 953 954 bdirty(bp); 955 /* 956 * Set B_CACHE, indicating that the buffer is fully valid. This is 957 * true even of NFS now. 958 */ 959 bp->b_flags |= B_CACHE; 960 961 /* 962 * This bmap keeps the system from needing to do the bmap later, 963 * perhaps when the system is attempting to do a sync. Since it 964 * is likely that the indirect block -- or whatever other datastructure 965 * that the filesystem needs is still in memory now, it is a good 966 * thing to do this. Note also, that if the pageout daemon is 967 * requesting a sync -- there might not be enough memory to do 968 * the bmap then... So, this is important to do. 969 */ 970 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 971 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 972 } 973 974 /* 975 * Set the *dirty* buffer range based upon the VM system dirty pages. 976 */ 977 vfs_setdirty(bp); 978 979 /* 980 * We need to do this here to satisfy the vnode_pager and the 981 * pageout daemon, so that it thinks that the pages have been 982 * "cleaned". Note that since the pages are in a delayed write 983 * buffer -- the VFS layer "will" see that the pages get written 984 * out on the next sync, or perhaps the cluster will be completed. 985 */ 986 vfs_clean_pages(bp); 987 bqrelse(bp); 988 989 /* 990 * Wakeup the buffer flushing daemon if we have a lot of dirty 991 * buffers (midpoint between our recovery point and our stall 992 * point). 993 */ 994 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 995 996 /* 997 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 998 * due to the softdep code. 999 */ 1000} 1001 1002/* 1003 * bdirty: 1004 * 1005 * Turn buffer into delayed write request. We must clear BIO_READ and 1006 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1007 * itself to properly update it in the dirty/clean lists. We mark it 1008 * B_DONE to ensure that any asynchronization of the buffer properly 1009 * clears B_DONE ( else a panic will occur later ). 1010 * 1011 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1012 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1013 * should only be called if the buffer is known-good. 1014 * 1015 * Since the buffer is not on a queue, we do not update the numfreebuffers 1016 * count. 1017 * 1018 * The buffer must be on QUEUE_NONE. 1019 */ 1020void 1021bdirty(struct buf *bp) 1022{ 1023 1024 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1025 bp, bp->b_vp, bp->b_flags); 1026 KASSERT(BUF_REFCNT(bp) == 1, ("bdirty: bp %p not locked",bp)); 1027 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1028 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1029 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1030 bp->b_flags &= ~(B_RELBUF); 1031 bp->b_iocmd = BIO_WRITE; 1032 1033 if ((bp->b_flags & B_DELWRI) == 0) { 1034 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1035 reassignbuf(bp); 1036 atomic_add_int(&numdirtybuffers, 1); 1037 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1038 } 1039} 1040 1041/* 1042 * bundirty: 1043 * 1044 * Clear B_DELWRI for buffer. 1045 * 1046 * Since the buffer is not on a queue, we do not update the numfreebuffers 1047 * count. 1048 * 1049 * The buffer must be on QUEUE_NONE. 1050 */ 1051 1052void 1053bundirty(struct buf *bp) 1054{ 1055 1056 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1057 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1058 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1059 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1060 KASSERT(BUF_REFCNT(bp) == 1, ("bundirty: bp %p not locked",bp)); 1061 1062 if (bp->b_flags & B_DELWRI) { 1063 bp->b_flags &= ~B_DELWRI; 1064 reassignbuf(bp); 1065 atomic_subtract_int(&numdirtybuffers, 1); 1066 numdirtywakeup(lodirtybuffers); 1067 } 1068 /* 1069 * Since it is now being written, we can clear its deferred write flag. 1070 */ 1071 bp->b_flags &= ~B_DEFERRED; 1072} 1073 1074/* 1075 * bawrite: 1076 * 1077 * Asynchronous write. Start output on a buffer, but do not wait for 1078 * it to complete. The buffer is released when the output completes. 1079 * 1080 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1081 * B_INVAL buffers. Not us. 1082 */ 1083void 1084bawrite(struct buf *bp) 1085{ 1086 1087 bp->b_flags |= B_ASYNC; 1088 (void) bwrite(bp); 1089} 1090 1091/* 1092 * bwillwrite: 1093 * 1094 * Called prior to the locking of any vnodes when we are expecting to 1095 * write. We do not want to starve the buffer cache with too many 1096 * dirty buffers so we block here. By blocking prior to the locking 1097 * of any vnodes we attempt to avoid the situation where a locked vnode 1098 * prevents the various system daemons from flushing related buffers. 1099 */ 1100 1101void 1102bwillwrite(void) 1103{ 1104 1105 if (numdirtybuffers >= hidirtybuffers) { 1106 mtx_lock(&nblock); 1107 while (numdirtybuffers >= hidirtybuffers) { 1108 bd_wakeup(1); 1109 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1110 msleep(&needsbuffer, &nblock, 1111 (PRIBIO + 4), "flswai", 0); 1112 } 1113 mtx_unlock(&nblock); 1114 } 1115} 1116 1117/* 1118 * Return true if we have too many dirty buffers. 1119 */ 1120int 1121buf_dirty_count_severe(void) 1122{ 1123 1124 return(numdirtybuffers >= hidirtybuffers); 1125} 1126 1127/* 1128 * brelse: 1129 * 1130 * Release a busy buffer and, if requested, free its resources. The 1131 * buffer will be stashed in the appropriate bufqueue[] allowing it 1132 * to be accessed later as a cache entity or reused for other purposes. 1133 */ 1134void 1135brelse(struct buf *bp) 1136{ 1137 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1138 bp, bp->b_vp, bp->b_flags); 1139 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1140 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1141 1142 if (bp->b_flags & B_MANAGED) { 1143 bqrelse(bp); 1144 return; 1145 } 1146 1147 if (bp->b_iocmd == BIO_WRITE && 1148 (bp->b_ioflags & BIO_ERROR) && 1149 !(bp->b_flags & B_INVAL)) { 1150 /* 1151 * Failed write, redirty. Must clear BIO_ERROR to prevent 1152 * pages from being scrapped. If B_INVAL is set then 1153 * this case is not run and the next case is run to 1154 * destroy the buffer. B_INVAL can occur if the buffer 1155 * is outside the range supported by the underlying device. 1156 */ 1157 bp->b_ioflags &= ~BIO_ERROR; 1158 bdirty(bp); 1159 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1160 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1161 /* 1162 * Either a failed I/O or we were asked to free or not 1163 * cache the buffer. 1164 */ 1165 bp->b_flags |= B_INVAL; 1166 if (LIST_FIRST(&bp->b_dep) != NULL) 1167 buf_deallocate(bp); 1168 if (bp->b_flags & B_DELWRI) { 1169 atomic_subtract_int(&numdirtybuffers, 1); 1170 numdirtywakeup(lodirtybuffers); 1171 } 1172 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1173 if ((bp->b_flags & B_VMIO) == 0) { 1174 if (bp->b_bufsize) 1175 allocbuf(bp, 0); 1176 if (bp->b_vp) 1177 brelvp(bp); 1178 } 1179 } 1180 1181 /* 1182 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1183 * is called with B_DELWRI set, the underlying pages may wind up 1184 * getting freed causing a previous write (bdwrite()) to get 'lost' 1185 * because pages associated with a B_DELWRI bp are marked clean. 1186 * 1187 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1188 * if B_DELWRI is set. 1189 * 1190 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1191 * on pages to return pages to the VM page queues. 1192 */ 1193 if (bp->b_flags & B_DELWRI) 1194 bp->b_flags &= ~B_RELBUF; 1195 else if (vm_page_count_severe()) { 1196 /* 1197 * XXX This lock may not be necessary since BKGRDINPROG 1198 * cannot be set while we hold the buf lock, it can only be 1199 * cleared if it is already pending. 1200 */ 1201 if (bp->b_vp) { 1202 BO_LOCK(bp->b_bufobj); 1203 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1204 bp->b_flags |= B_RELBUF; 1205 BO_UNLOCK(bp->b_bufobj); 1206 } else 1207 bp->b_flags |= B_RELBUF; 1208 } 1209 1210 /* 1211 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1212 * constituted, not even NFS buffers now. Two flags effect this. If 1213 * B_INVAL, the struct buf is invalidated but the VM object is kept 1214 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1215 * 1216 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1217 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1218 * buffer is also B_INVAL because it hits the re-dirtying code above. 1219 * 1220 * Normally we can do this whether a buffer is B_DELWRI or not. If 1221 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1222 * the commit state and we cannot afford to lose the buffer. If the 1223 * buffer has a background write in progress, we need to keep it 1224 * around to prevent it from being reconstituted and starting a second 1225 * background write. 1226 */ 1227 if ((bp->b_flags & B_VMIO) 1228 && !(bp->b_vp->v_mount != NULL && 1229 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1230 !vn_isdisk(bp->b_vp, NULL) && 1231 (bp->b_flags & B_DELWRI)) 1232 ) { 1233 1234 int i, j, resid; 1235 vm_page_t m; 1236 off_t foff; 1237 vm_pindex_t poff; 1238 vm_object_t obj; 1239 1240 obj = bp->b_bufobj->bo_object; 1241 1242 /* 1243 * Get the base offset and length of the buffer. Note that 1244 * in the VMIO case if the buffer block size is not 1245 * page-aligned then b_data pointer may not be page-aligned. 1246 * But our b_pages[] array *IS* page aligned. 1247 * 1248 * block sizes less then DEV_BSIZE (usually 512) are not 1249 * supported due to the page granularity bits (m->valid, 1250 * m->dirty, etc...). 1251 * 1252 * See man buf(9) for more information 1253 */ 1254 resid = bp->b_bufsize; 1255 foff = bp->b_offset; 1256 VM_OBJECT_LOCK(obj); 1257 for (i = 0; i < bp->b_npages; i++) { 1258 int had_bogus = 0; 1259 1260 m = bp->b_pages[i]; 1261 1262 /* 1263 * If we hit a bogus page, fixup *all* the bogus pages 1264 * now. 1265 */ 1266 if (m == bogus_page) { 1267 poff = OFF_TO_IDX(bp->b_offset); 1268 had_bogus = 1; 1269 1270 for (j = i; j < bp->b_npages; j++) { 1271 vm_page_t mtmp; 1272 mtmp = bp->b_pages[j]; 1273 if (mtmp == bogus_page) { 1274 mtmp = vm_page_lookup(obj, poff + j); 1275 if (!mtmp) { 1276 panic("brelse: page missing\n"); 1277 } 1278 bp->b_pages[j] = mtmp; 1279 } 1280 } 1281 1282 if ((bp->b_flags & B_INVAL) == 0) { 1283 pmap_qenter( 1284 trunc_page((vm_offset_t)bp->b_data), 1285 bp->b_pages, bp->b_npages); 1286 } 1287 m = bp->b_pages[i]; 1288 } 1289 if ((bp->b_flags & B_NOCACHE) || 1290 (bp->b_ioflags & BIO_ERROR)) { 1291 int poffset = foff & PAGE_MASK; 1292 int presid = resid > (PAGE_SIZE - poffset) ? 1293 (PAGE_SIZE - poffset) : resid; 1294 1295 KASSERT(presid >= 0, ("brelse: extra page")); 1296 vm_page_lock_queues(); 1297 vm_page_set_invalid(m, poffset, presid); 1298 vm_page_unlock_queues(); 1299 if (had_bogus) 1300 printf("avoided corruption bug in bogus_page/brelse code\n"); 1301 } 1302 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1303 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1304 } 1305 VM_OBJECT_UNLOCK(obj); 1306 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1307 vfs_vmio_release(bp); 1308 1309 } else if (bp->b_flags & B_VMIO) { 1310 1311 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1312 vfs_vmio_release(bp); 1313 } 1314 1315 } 1316 1317 if (BUF_REFCNT(bp) > 1) { 1318 /* do not release to free list */ 1319 BUF_UNLOCK(bp); 1320 return; 1321 } 1322 1323 /* enqueue */ 1324 mtx_lock(&bqlock); 1325 /* Handle delayed bremfree() processing. */ 1326 if (bp->b_flags & B_REMFREE) 1327 bremfreel(bp); 1328 if (bp->b_qindex != QUEUE_NONE) 1329 panic("brelse: free buffer onto another queue???"); 1330 1331 /* buffers with no memory */ 1332 if (bp->b_bufsize == 0) { 1333 bp->b_flags |= B_INVAL; 1334 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1335 if (bp->b_vflags & BV_BKGRDINPROG) 1336 panic("losing buffer 1"); 1337 if (bp->b_kvasize) { 1338 bp->b_qindex = QUEUE_EMPTYKVA; 1339 } else { 1340 bp->b_qindex = QUEUE_EMPTY; 1341 } 1342 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1343 /* buffers with junk contents */ 1344 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1345 (bp->b_ioflags & BIO_ERROR)) { 1346 bp->b_flags |= B_INVAL; 1347 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1348 if (bp->b_vflags & BV_BKGRDINPROG) 1349 panic("losing buffer 2"); 1350 bp->b_qindex = QUEUE_CLEAN; 1351 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1352 /* remaining buffers */ 1353 } else { 1354 if (bp->b_flags & B_DELWRI) 1355 bp->b_qindex = QUEUE_DIRTY; 1356 else 1357 bp->b_qindex = QUEUE_CLEAN; 1358 if (bp->b_flags & B_AGE) 1359 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1360 else 1361 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1362 } 1363 mtx_unlock(&bqlock); 1364 1365 /* 1366 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1367 * placed the buffer on the correct queue. We must also disassociate 1368 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1369 * find it. 1370 */ 1371 if (bp->b_flags & B_INVAL) { 1372 if (bp->b_flags & B_DELWRI) 1373 bundirty(bp); 1374 if (bp->b_vp) 1375 brelvp(bp); 1376 } 1377 1378 /* 1379 * Fixup numfreebuffers count. The bp is on an appropriate queue 1380 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1381 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1382 * if B_INVAL is set ). 1383 */ 1384 1385 if (!(bp->b_flags & B_DELWRI)) 1386 bufcountwakeup(); 1387 1388 /* 1389 * Something we can maybe free or reuse 1390 */ 1391 if (bp->b_bufsize || bp->b_kvasize) 1392 bufspacewakeup(); 1393 1394 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1395 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1396 panic("brelse: not dirty"); 1397 /* unlock */ 1398 BUF_UNLOCK(bp); 1399} 1400 1401/* 1402 * Release a buffer back to the appropriate queue but do not try to free 1403 * it. The buffer is expected to be used again soon. 1404 * 1405 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1406 * biodone() to requeue an async I/O on completion. It is also used when 1407 * known good buffers need to be requeued but we think we may need the data 1408 * again soon. 1409 * 1410 * XXX we should be able to leave the B_RELBUF hint set on completion. 1411 */ 1412void 1413bqrelse(struct buf *bp) 1414{ 1415 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1416 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1417 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1418 1419 if (BUF_REFCNT(bp) > 1) { 1420 /* do not release to free list */ 1421 BUF_UNLOCK(bp); 1422 return; 1423 } 1424 1425 if (bp->b_flags & B_MANAGED) { 1426 if (bp->b_flags & B_REMFREE) { 1427 mtx_lock(&bqlock); 1428 bremfreel(bp); 1429 mtx_unlock(&bqlock); 1430 } 1431 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1432 BUF_UNLOCK(bp); 1433 return; 1434 } 1435 1436 mtx_lock(&bqlock); 1437 /* Handle delayed bremfree() processing. */ 1438 if (bp->b_flags & B_REMFREE) 1439 bremfreel(bp); 1440 if (bp->b_qindex != QUEUE_NONE) 1441 panic("bqrelse: free buffer onto another queue???"); 1442 /* buffers with stale but valid contents */ 1443 if (bp->b_flags & B_DELWRI) { 1444 bp->b_qindex = QUEUE_DIRTY; 1445 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1446 } else { 1447 /* 1448 * XXX This lock may not be necessary since BKGRDINPROG 1449 * cannot be set while we hold the buf lock, it can only be 1450 * cleared if it is already pending. 1451 */ 1452 BO_LOCK(bp->b_bufobj); 1453 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1454 BO_UNLOCK(bp->b_bufobj); 1455 bp->b_qindex = QUEUE_CLEAN; 1456 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1457 b_freelist); 1458 } else { 1459 /* 1460 * We are too low on memory, we have to try to free 1461 * the buffer (most importantly: the wired pages 1462 * making up its backing store) *now*. 1463 */ 1464 BO_UNLOCK(bp->b_bufobj); 1465 mtx_unlock(&bqlock); 1466 brelse(bp); 1467 return; 1468 } 1469 } 1470 mtx_unlock(&bqlock); 1471 1472 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1473 bufcountwakeup(); 1474 1475 /* 1476 * Something we can maybe free or reuse. 1477 */ 1478 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1479 bufspacewakeup(); 1480 1481 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1482 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1483 panic("bqrelse: not dirty"); 1484 /* unlock */ 1485 BUF_UNLOCK(bp); 1486} 1487 1488/* Give pages used by the bp back to the VM system (where possible) */ 1489static void 1490vfs_vmio_release(struct buf *bp) 1491{ 1492 int i; 1493 vm_page_t m; 1494 1495 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1496 vm_page_lock_queues(); 1497 for (i = 0; i < bp->b_npages; i++) { 1498 m = bp->b_pages[i]; 1499 bp->b_pages[i] = NULL; 1500 /* 1501 * In order to keep page LRU ordering consistent, put 1502 * everything on the inactive queue. 1503 */ 1504 vm_page_unwire(m, 0); 1505 /* 1506 * We don't mess with busy pages, it is 1507 * the responsibility of the process that 1508 * busied the pages to deal with them. 1509 */ 1510 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1511 continue; 1512 1513 if (m->wire_count == 0) { 1514 /* 1515 * Might as well free the page if we can and it has 1516 * no valid data. We also free the page if the 1517 * buffer was used for direct I/O 1518 */ 1519 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1520 m->hold_count == 0) { 1521 vm_page_free(m); 1522 } else if (bp->b_flags & B_DIRECT) { 1523 vm_page_try_to_free(m); 1524 } else if (vm_page_count_severe()) { 1525 vm_page_try_to_cache(m); 1526 } 1527 } 1528 } 1529 vm_page_unlock_queues(); 1530 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1531 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1532 1533 if (bp->b_bufsize) { 1534 bufspacewakeup(); 1535 bp->b_bufsize = 0; 1536 } 1537 bp->b_npages = 0; 1538 bp->b_flags &= ~B_VMIO; 1539 if (bp->b_vp) 1540 brelvp(bp); 1541} 1542 1543/* 1544 * Check to see if a block at a particular lbn is available for a clustered 1545 * write. 1546 */ 1547static int 1548vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1549{ 1550 struct buf *bpa; 1551 int match; 1552 1553 match = 0; 1554 1555 /* If the buf isn't in core skip it */ 1556 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1557 return (0); 1558 1559 /* If the buf is busy we don't want to wait for it */ 1560 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1561 return (0); 1562 1563 /* Only cluster with valid clusterable delayed write buffers */ 1564 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1565 (B_DELWRI | B_CLUSTEROK)) 1566 goto done; 1567 1568 if (bpa->b_bufsize != size) 1569 goto done; 1570 1571 /* 1572 * Check to see if it is in the expected place on disk and that the 1573 * block has been mapped. 1574 */ 1575 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1576 match = 1; 1577done: 1578 BUF_UNLOCK(bpa); 1579 return (match); 1580} 1581 1582/* 1583 * vfs_bio_awrite: 1584 * 1585 * Implement clustered async writes for clearing out B_DELWRI buffers. 1586 * This is much better then the old way of writing only one buffer at 1587 * a time. Note that we may not be presented with the buffers in the 1588 * correct order, so we search for the cluster in both directions. 1589 */ 1590int 1591vfs_bio_awrite(struct buf *bp) 1592{ 1593 int i; 1594 int j; 1595 daddr_t lblkno = bp->b_lblkno; 1596 struct vnode *vp = bp->b_vp; 1597 int ncl; 1598 int nwritten; 1599 int size; 1600 int maxcl; 1601 1602 /* 1603 * right now we support clustered writing only to regular files. If 1604 * we find a clusterable block we could be in the middle of a cluster 1605 * rather then at the beginning. 1606 */ 1607 if ((vp->v_type == VREG) && 1608 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1609 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1610 1611 size = vp->v_mount->mnt_stat.f_iosize; 1612 maxcl = MAXPHYS / size; 1613 1614 VI_LOCK(vp); 1615 for (i = 1; i < maxcl; i++) 1616 if (vfs_bio_clcheck(vp, size, lblkno + i, 1617 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1618 break; 1619 1620 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1621 if (vfs_bio_clcheck(vp, size, lblkno - j, 1622 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1623 break; 1624 1625 VI_UNLOCK(vp); 1626 --j; 1627 ncl = i + j; 1628 /* 1629 * this is a possible cluster write 1630 */ 1631 if (ncl != 1) { 1632 BUF_UNLOCK(bp); 1633 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1634 return nwritten; 1635 } 1636 } 1637 bremfree(bp); 1638 bp->b_flags |= B_ASYNC; 1639 /* 1640 * default (old) behavior, writing out only one block 1641 * 1642 * XXX returns b_bufsize instead of b_bcount for nwritten? 1643 */ 1644 nwritten = bp->b_bufsize; 1645 (void) bwrite(bp); 1646 1647 return nwritten; 1648} 1649 1650/* 1651 * getnewbuf: 1652 * 1653 * Find and initialize a new buffer header, freeing up existing buffers 1654 * in the bufqueues as necessary. The new buffer is returned locked. 1655 * 1656 * Important: B_INVAL is not set. If the caller wishes to throw the 1657 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1658 * 1659 * We block if: 1660 * We have insufficient buffer headers 1661 * We have insufficient buffer space 1662 * buffer_map is too fragmented ( space reservation fails ) 1663 * If we have to flush dirty buffers ( but we try to avoid this ) 1664 * 1665 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1666 * Instead we ask the buf daemon to do it for us. We attempt to 1667 * avoid piecemeal wakeups of the pageout daemon. 1668 */ 1669 1670static struct buf * 1671getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1672{ 1673 struct buf *bp; 1674 struct buf *nbp; 1675 int defrag = 0; 1676 int nqindex; 1677 static int flushingbufs; 1678 1679 /* 1680 * We can't afford to block since we might be holding a vnode lock, 1681 * which may prevent system daemons from running. We deal with 1682 * low-memory situations by proactively returning memory and running 1683 * async I/O rather then sync I/O. 1684 */ 1685 1686 atomic_add_int(&getnewbufcalls, 1); 1687 atomic_subtract_int(&getnewbufrestarts, 1); 1688restart: 1689 atomic_add_int(&getnewbufrestarts, 1); 1690 1691 /* 1692 * Setup for scan. If we do not have enough free buffers, 1693 * we setup a degenerate case that immediately fails. Note 1694 * that if we are specially marked process, we are allowed to 1695 * dip into our reserves. 1696 * 1697 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1698 * 1699 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1700 * However, there are a number of cases (defragging, reusing, ...) 1701 * where we cannot backup. 1702 */ 1703 mtx_lock(&bqlock); 1704 nqindex = QUEUE_EMPTYKVA; 1705 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1706 1707 if (nbp == NULL) { 1708 /* 1709 * If no EMPTYKVA buffers and we are either 1710 * defragging or reusing, locate a CLEAN buffer 1711 * to free or reuse. If bufspace useage is low 1712 * skip this step so we can allocate a new buffer. 1713 */ 1714 if (defrag || bufspace >= lobufspace) { 1715 nqindex = QUEUE_CLEAN; 1716 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1717 } 1718 1719 /* 1720 * If we could not find or were not allowed to reuse a 1721 * CLEAN buffer, check to see if it is ok to use an EMPTY 1722 * buffer. We can only use an EMPTY buffer if allocating 1723 * its KVA would not otherwise run us out of buffer space. 1724 */ 1725 if (nbp == NULL && defrag == 0 && 1726 bufspace + maxsize < hibufspace) { 1727 nqindex = QUEUE_EMPTY; 1728 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1729 } 1730 } 1731 1732 /* 1733 * Run scan, possibly freeing data and/or kva mappings on the fly 1734 * depending. 1735 */ 1736 1737 while ((bp = nbp) != NULL) { 1738 int qindex = nqindex; 1739 1740 /* 1741 * Calculate next bp ( we can only use it if we do not block 1742 * or do other fancy things ). 1743 */ 1744 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1745 switch(qindex) { 1746 case QUEUE_EMPTY: 1747 nqindex = QUEUE_EMPTYKVA; 1748 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1749 break; 1750 /* FALLTHROUGH */ 1751 case QUEUE_EMPTYKVA: 1752 nqindex = QUEUE_CLEAN; 1753 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1754 break; 1755 /* FALLTHROUGH */ 1756 case QUEUE_CLEAN: 1757 /* 1758 * nbp is NULL. 1759 */ 1760 break; 1761 } 1762 } 1763 /* 1764 * If we are defragging then we need a buffer with 1765 * b_kvasize != 0. XXX this situation should no longer 1766 * occur, if defrag is non-zero the buffer's b_kvasize 1767 * should also be non-zero at this point. XXX 1768 */ 1769 if (defrag && bp->b_kvasize == 0) { 1770 printf("Warning: defrag empty buffer %p\n", bp); 1771 continue; 1772 } 1773 1774 /* 1775 * Start freeing the bp. This is somewhat involved. nbp 1776 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1777 */ 1778 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1779 continue; 1780 if (bp->b_vp) { 1781 BO_LOCK(bp->b_bufobj); 1782 if (bp->b_vflags & BV_BKGRDINPROG) { 1783 BO_UNLOCK(bp->b_bufobj); 1784 BUF_UNLOCK(bp); 1785 continue; 1786 } 1787 BO_UNLOCK(bp->b_bufobj); 1788 } 1789 CTR6(KTR_BUF, 1790 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1791 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1792 bp->b_kvasize, bp->b_bufsize, qindex); 1793 1794 /* 1795 * Sanity Checks 1796 */ 1797 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1798 1799 /* 1800 * Note: we no longer distinguish between VMIO and non-VMIO 1801 * buffers. 1802 */ 1803 1804 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1805 1806 bremfreel(bp); 1807 mtx_unlock(&bqlock); 1808 1809 if (qindex == QUEUE_CLEAN) { 1810 if (bp->b_flags & B_VMIO) { 1811 bp->b_flags &= ~B_ASYNC; 1812 vfs_vmio_release(bp); 1813 } 1814 if (bp->b_vp) 1815 brelvp(bp); 1816 } 1817 1818 /* 1819 * NOTE: nbp is now entirely invalid. We can only restart 1820 * the scan from this point on. 1821 * 1822 * Get the rest of the buffer freed up. b_kva* is still 1823 * valid after this operation. 1824 */ 1825 1826 if (bp->b_rcred != NOCRED) { 1827 crfree(bp->b_rcred); 1828 bp->b_rcred = NOCRED; 1829 } 1830 if (bp->b_wcred != NOCRED) { 1831 crfree(bp->b_wcred); 1832 bp->b_wcred = NOCRED; 1833 } 1834 if (LIST_FIRST(&bp->b_dep) != NULL) 1835 buf_deallocate(bp); 1836 if (bp->b_vflags & BV_BKGRDINPROG) 1837 panic("losing buffer 3"); 1838 KASSERT(bp->b_vp == NULL, 1839 ("bp: %p still has vnode %p. qindex: %d", 1840 bp, bp->b_vp, qindex)); 1841 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1842 ("bp: %p still on a buffer list. xflags %X", 1843 bp, bp->b_xflags)); 1844 1845 if (bp->b_bufsize) 1846 allocbuf(bp, 0); 1847 1848 bp->b_flags = 0; 1849 bp->b_ioflags = 0; 1850 bp->b_xflags = 0; 1851 bp->b_vflags = 0; 1852 bp->b_vp = NULL; 1853 bp->b_blkno = bp->b_lblkno = 0; 1854 bp->b_offset = NOOFFSET; 1855 bp->b_iodone = 0; 1856 bp->b_error = 0; 1857 bp->b_resid = 0; 1858 bp->b_bcount = 0; 1859 bp->b_npages = 0; 1860 bp->b_dirtyoff = bp->b_dirtyend = 0; 1861 bp->b_bufobj = NULL; 1862 bp->b_pin_count = 0; 1863 bp->b_fsprivate1 = NULL; 1864 bp->b_fsprivate2 = NULL; 1865 bp->b_fsprivate3 = NULL; 1866 1867 LIST_INIT(&bp->b_dep); 1868 1869 /* 1870 * If we are defragging then free the buffer. 1871 */ 1872 if (defrag) { 1873 bp->b_flags |= B_INVAL; 1874 bfreekva(bp); 1875 brelse(bp); 1876 defrag = 0; 1877 goto restart; 1878 } 1879 1880 /* 1881 * If we are overcomitted then recover the buffer and its 1882 * KVM space. This occurs in rare situations when multiple 1883 * processes are blocked in getnewbuf() or allocbuf(). 1884 */ 1885 if (bufspace >= hibufspace) 1886 flushingbufs = 1; 1887 if (flushingbufs && bp->b_kvasize != 0) { 1888 bp->b_flags |= B_INVAL; 1889 bfreekva(bp); 1890 brelse(bp); 1891 goto restart; 1892 } 1893 if (bufspace < lobufspace) 1894 flushingbufs = 0; 1895 break; 1896 } 1897 1898 /* 1899 * If we exhausted our list, sleep as appropriate. We may have to 1900 * wakeup various daemons and write out some dirty buffers. 1901 * 1902 * Generally we are sleeping due to insufficient buffer space. 1903 */ 1904 1905 if (bp == NULL) { 1906 int flags; 1907 char *waitmsg; 1908 1909 if (defrag) { 1910 flags = VFS_BIO_NEED_BUFSPACE; 1911 waitmsg = "nbufkv"; 1912 } else if (bufspace >= hibufspace) { 1913 waitmsg = "nbufbs"; 1914 flags = VFS_BIO_NEED_BUFSPACE; 1915 } else { 1916 waitmsg = "newbuf"; 1917 flags = VFS_BIO_NEED_ANY; 1918 } 1919 mtx_lock(&nblock); 1920 needsbuffer |= flags; 1921 mtx_unlock(&nblock); 1922 mtx_unlock(&bqlock); 1923 1924 bd_speedup(); /* heeeelp */ 1925 1926 mtx_lock(&nblock); 1927 while (needsbuffer & flags) { 1928 if (msleep(&needsbuffer, &nblock, 1929 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1930 mtx_unlock(&nblock); 1931 return (NULL); 1932 } 1933 } 1934 mtx_unlock(&nblock); 1935 } else { 1936 /* 1937 * We finally have a valid bp. We aren't quite out of the 1938 * woods, we still have to reserve kva space. In order 1939 * to keep fragmentation sane we only allocate kva in 1940 * BKVASIZE chunks. 1941 */ 1942 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1943 1944 if (maxsize != bp->b_kvasize) { 1945 vm_offset_t addr = 0; 1946 1947 bfreekva(bp); 1948 1949 vm_map_lock(buffer_map); 1950 if (vm_map_findspace(buffer_map, 1951 vm_map_min(buffer_map), maxsize, &addr)) { 1952 /* 1953 * Uh oh. Buffer map is to fragmented. We 1954 * must defragment the map. 1955 */ 1956 atomic_add_int(&bufdefragcnt, 1); 1957 vm_map_unlock(buffer_map); 1958 defrag = 1; 1959 bp->b_flags |= B_INVAL; 1960 brelse(bp); 1961 goto restart; 1962 } 1963 if (addr) { 1964 vm_map_insert(buffer_map, NULL, 0, 1965 addr, addr + maxsize, 1966 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1967 1968 bp->b_kvabase = (caddr_t) addr; 1969 bp->b_kvasize = maxsize; 1970 atomic_add_int(&bufspace, bp->b_kvasize); 1971 atomic_add_int(&bufreusecnt, 1); 1972 } 1973 vm_map_unlock(buffer_map); 1974 } 1975 bp->b_saveaddr = bp->b_kvabase; 1976 bp->b_data = bp->b_saveaddr; 1977 } 1978 return(bp); 1979} 1980 1981/* 1982 * buf_daemon: 1983 * 1984 * buffer flushing daemon. Buffers are normally flushed by the 1985 * update daemon but if it cannot keep up this process starts to 1986 * take the load in an attempt to prevent getnewbuf() from blocking. 1987 */ 1988 1989static struct kproc_desc buf_kp = { 1990 "bufdaemon", 1991 buf_daemon, 1992 &bufdaemonproc 1993}; 1994SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1995 1996static void 1997buf_daemon() 1998{ 1999 mtx_lock(&Giant); 2000 2001 /* 2002 * This process needs to be suspended prior to shutdown sync. 2003 */ 2004 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2005 SHUTDOWN_PRI_LAST); 2006 2007 /* 2008 * This process is allowed to take the buffer cache to the limit 2009 */ 2010 curthread->td_pflags |= TDP_NORUNNINGBUF; 2011 mtx_lock(&bdlock); 2012 for (;;) { 2013 bd_request = 0; 2014 mtx_unlock(&bdlock); 2015 2016 kthread_suspend_check(bufdaemonproc); 2017 2018 /* 2019 * Do the flush. Limit the amount of in-transit I/O we 2020 * allow to build up, otherwise we would completely saturate 2021 * the I/O system. Wakeup any waiting processes before we 2022 * normally would so they can run in parallel with our drain. 2023 */ 2024 while (numdirtybuffers > lodirtybuffers) { 2025 if (flushbufqueues(0) == 0) { 2026 /* 2027 * Could not find any buffers without rollback 2028 * dependencies, so just write the first one 2029 * in the hopes of eventually making progress. 2030 */ 2031 flushbufqueues(1); 2032 break; 2033 } 2034 uio_yield(); 2035 } 2036 2037 /* 2038 * Only clear bd_request if we have reached our low water 2039 * mark. The buf_daemon normally waits 1 second and 2040 * then incrementally flushes any dirty buffers that have 2041 * built up, within reason. 2042 * 2043 * If we were unable to hit our low water mark and couldn't 2044 * find any flushable buffers, we sleep half a second. 2045 * Otherwise we loop immediately. 2046 */ 2047 mtx_lock(&bdlock); 2048 if (numdirtybuffers <= lodirtybuffers) { 2049 /* 2050 * We reached our low water mark, reset the 2051 * request and sleep until we are needed again. 2052 * The sleep is just so the suspend code works. 2053 */ 2054 bd_request = 0; 2055 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2056 } else { 2057 /* 2058 * We couldn't find any flushable dirty buffers but 2059 * still have too many dirty buffers, we 2060 * have to sleep and try again. (rare) 2061 */ 2062 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2063 } 2064 } 2065} 2066 2067/* 2068 * flushbufqueues: 2069 * 2070 * Try to flush a buffer in the dirty queue. We must be careful to 2071 * free up B_INVAL buffers instead of write them, which NFS is 2072 * particularly sensitive to. 2073 */ 2074static int flushwithdeps = 0; 2075SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2076 0, "Number of buffers flushed with dependecies that require rollbacks"); 2077 2078static int 2079flushbufqueues(int flushdeps) 2080{ 2081 struct thread *td = curthread; 2082 struct buf sentinel; 2083 struct vnode *vp; 2084 struct mount *mp; 2085 struct buf *bp; 2086 int hasdeps; 2087 int flushed; 2088 int target; 2089 2090 target = numdirtybuffers - lodirtybuffers; 2091 if (flushdeps && target > 2) 2092 target /= 2; 2093 flushed = 0; 2094 bp = NULL; 2095 mtx_lock(&bqlock); 2096 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], &sentinel, b_freelist); 2097 while (flushed != target) { 2098 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 2099 if (bp == &sentinel) 2100 break; 2101 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 2102 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 2103 2104 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2105 continue; 2106 if (bp->b_pin_count > 0) { 2107 BUF_UNLOCK(bp); 2108 continue; 2109 } 2110 BO_LOCK(bp->b_bufobj); 2111 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2112 (bp->b_flags & B_DELWRI) == 0) { 2113 BO_UNLOCK(bp->b_bufobj); 2114 BUF_UNLOCK(bp); 2115 continue; 2116 } 2117 BO_UNLOCK(bp->b_bufobj); 2118 if (bp->b_flags & B_INVAL) { 2119 bremfreel(bp); 2120 mtx_unlock(&bqlock); 2121 brelse(bp); 2122 flushed++; 2123 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2124 mtx_lock(&bqlock); 2125 continue; 2126 } 2127 2128 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2129 if (flushdeps == 0) { 2130 BUF_UNLOCK(bp); 2131 continue; 2132 } 2133 hasdeps = 1; 2134 } else 2135 hasdeps = 0; 2136 /* 2137 * We must hold the lock on a vnode before writing 2138 * one of its buffers. Otherwise we may confuse, or 2139 * in the case of a snapshot vnode, deadlock the 2140 * system. 2141 * 2142 * The lock order here is the reverse of the normal 2143 * of vnode followed by buf lock. This is ok because 2144 * the NOWAIT will prevent deadlock. 2145 */ 2146 vp = bp->b_vp; 2147 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2148 BUF_UNLOCK(bp); 2149 continue; 2150 } 2151 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2152 mtx_unlock(&bqlock); 2153 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2154 bp, bp->b_vp, bp->b_flags); 2155 vfs_bio_awrite(bp); 2156 vn_finished_write(mp); 2157 VOP_UNLOCK(vp, 0, td); 2158 flushwithdeps += hasdeps; 2159 flushed++; 2160 waitrunningbufspace(); 2161 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2162 mtx_lock(&bqlock); 2163 continue; 2164 } 2165 vn_finished_write(mp); 2166 BUF_UNLOCK(bp); 2167 } 2168 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], &sentinel, b_freelist); 2169 mtx_unlock(&bqlock); 2170 return (flushed); 2171} 2172 2173/* 2174 * Check to see if a block is currently memory resident. 2175 */ 2176struct buf * 2177incore(struct bufobj *bo, daddr_t blkno) 2178{ 2179 struct buf *bp; 2180 2181 BO_LOCK(bo); 2182 bp = gbincore(bo, blkno); 2183 BO_UNLOCK(bo); 2184 return (bp); 2185} 2186 2187/* 2188 * Returns true if no I/O is needed to access the 2189 * associated VM object. This is like incore except 2190 * it also hunts around in the VM system for the data. 2191 */ 2192 2193static int 2194inmem(struct vnode * vp, daddr_t blkno) 2195{ 2196 vm_object_t obj; 2197 vm_offset_t toff, tinc, size; 2198 vm_page_t m; 2199 vm_ooffset_t off; 2200 2201 ASSERT_VOP_LOCKED(vp, "inmem"); 2202 2203 if (incore(&vp->v_bufobj, blkno)) 2204 return 1; 2205 if (vp->v_mount == NULL) 2206 return 0; 2207 obj = vp->v_object; 2208 if (obj == NULL) 2209 return (0); 2210 2211 size = PAGE_SIZE; 2212 if (size > vp->v_mount->mnt_stat.f_iosize) 2213 size = vp->v_mount->mnt_stat.f_iosize; 2214 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2215 2216 VM_OBJECT_LOCK(obj); 2217 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2218 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2219 if (!m) 2220 goto notinmem; 2221 tinc = size; 2222 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2223 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2224 if (vm_page_is_valid(m, 2225 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2226 goto notinmem; 2227 } 2228 VM_OBJECT_UNLOCK(obj); 2229 return 1; 2230 2231notinmem: 2232 VM_OBJECT_UNLOCK(obj); 2233 return (0); 2234} 2235 2236/* 2237 * vfs_setdirty: 2238 * 2239 * Sets the dirty range for a buffer based on the status of the dirty 2240 * bits in the pages comprising the buffer. 2241 * 2242 * The range is limited to the size of the buffer. 2243 * 2244 * This routine is primarily used by NFS, but is generalized for the 2245 * B_VMIO case. 2246 */ 2247static void 2248vfs_setdirty(struct buf *bp) 2249{ 2250 int i; 2251 vm_object_t object; 2252 2253 /* 2254 * Degenerate case - empty buffer 2255 */ 2256 2257 if (bp->b_bufsize == 0) 2258 return; 2259 2260 /* 2261 * We qualify the scan for modified pages on whether the 2262 * object has been flushed yet. The OBJ_WRITEABLE flag 2263 * is not cleared simply by protecting pages off. 2264 */ 2265 2266 if ((bp->b_flags & B_VMIO) == 0) 2267 return; 2268 2269 object = bp->b_pages[0]->object; 2270 VM_OBJECT_LOCK(object); 2271 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2272 printf("Warning: object %p writeable but not mightbedirty\n", object); 2273 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2274 printf("Warning: object %p mightbedirty but not writeable\n", object); 2275 2276 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2277 vm_offset_t boffset; 2278 vm_offset_t eoffset; 2279 2280 vm_page_lock_queues(); 2281 /* 2282 * test the pages to see if they have been modified directly 2283 * by users through the VM system. 2284 */ 2285 for (i = 0; i < bp->b_npages; i++) 2286 vm_page_test_dirty(bp->b_pages[i]); 2287 2288 /* 2289 * Calculate the encompassing dirty range, boffset and eoffset, 2290 * (eoffset - boffset) bytes. 2291 */ 2292 2293 for (i = 0; i < bp->b_npages; i++) { 2294 if (bp->b_pages[i]->dirty) 2295 break; 2296 } 2297 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2298 2299 for (i = bp->b_npages - 1; i >= 0; --i) { 2300 if (bp->b_pages[i]->dirty) { 2301 break; 2302 } 2303 } 2304 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2305 2306 vm_page_unlock_queues(); 2307 /* 2308 * Fit it to the buffer. 2309 */ 2310 2311 if (eoffset > bp->b_bcount) 2312 eoffset = bp->b_bcount; 2313 2314 /* 2315 * If we have a good dirty range, merge with the existing 2316 * dirty range. 2317 */ 2318 2319 if (boffset < eoffset) { 2320 if (bp->b_dirtyoff > boffset) 2321 bp->b_dirtyoff = boffset; 2322 if (bp->b_dirtyend < eoffset) 2323 bp->b_dirtyend = eoffset; 2324 } 2325 } 2326 VM_OBJECT_UNLOCK(object); 2327} 2328 2329/* 2330 * getblk: 2331 * 2332 * Get a block given a specified block and offset into a file/device. 2333 * The buffers B_DONE bit will be cleared on return, making it almost 2334 * ready for an I/O initiation. B_INVAL may or may not be set on 2335 * return. The caller should clear B_INVAL prior to initiating a 2336 * READ. 2337 * 2338 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2339 * an existing buffer. 2340 * 2341 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2342 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2343 * and then cleared based on the backing VM. If the previous buffer is 2344 * non-0-sized but invalid, B_CACHE will be cleared. 2345 * 2346 * If getblk() must create a new buffer, the new buffer is returned with 2347 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2348 * case it is returned with B_INVAL clear and B_CACHE set based on the 2349 * backing VM. 2350 * 2351 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2352 * B_CACHE bit is clear. 2353 * 2354 * What this means, basically, is that the caller should use B_CACHE to 2355 * determine whether the buffer is fully valid or not and should clear 2356 * B_INVAL prior to issuing a read. If the caller intends to validate 2357 * the buffer by loading its data area with something, the caller needs 2358 * to clear B_INVAL. If the caller does this without issuing an I/O, 2359 * the caller should set B_CACHE ( as an optimization ), else the caller 2360 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2361 * a write attempt or if it was a successfull read. If the caller 2362 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2363 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2364 */ 2365struct buf * 2366getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2367 int flags) 2368{ 2369 struct buf *bp; 2370 struct bufobj *bo; 2371 int error; 2372 2373 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2374 ASSERT_VOP_LOCKED(vp, "getblk"); 2375 if (size > MAXBSIZE) 2376 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2377 2378 bo = &vp->v_bufobj; 2379loop: 2380 /* 2381 * Block if we are low on buffers. Certain processes are allowed 2382 * to completely exhaust the buffer cache. 2383 * 2384 * If this check ever becomes a bottleneck it may be better to 2385 * move it into the else, when gbincore() fails. At the moment 2386 * it isn't a problem. 2387 * 2388 * XXX remove if 0 sections (clean this up after its proven) 2389 */ 2390 if (numfreebuffers == 0) { 2391 if (curthread == PCPU_GET(idlethread)) 2392 return NULL; 2393 mtx_lock(&nblock); 2394 needsbuffer |= VFS_BIO_NEED_ANY; 2395 mtx_unlock(&nblock); 2396 } 2397 2398 VI_LOCK(vp); 2399 bp = gbincore(bo, blkno); 2400 if (bp != NULL) { 2401 int lockflags; 2402 /* 2403 * Buffer is in-core. If the buffer is not busy, it must 2404 * be on a queue. 2405 */ 2406 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2407 2408 if (flags & GB_LOCK_NOWAIT) 2409 lockflags |= LK_NOWAIT; 2410 2411 error = BUF_TIMELOCK(bp, lockflags, 2412 VI_MTX(vp), "getblk", slpflag, slptimeo); 2413 2414 /* 2415 * If we slept and got the lock we have to restart in case 2416 * the buffer changed identities. 2417 */ 2418 if (error == ENOLCK) 2419 goto loop; 2420 /* We timed out or were interrupted. */ 2421 else if (error) 2422 return (NULL); 2423 2424 /* 2425 * The buffer is locked. B_CACHE is cleared if the buffer is 2426 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2427 * and for a VMIO buffer B_CACHE is adjusted according to the 2428 * backing VM cache. 2429 */ 2430 if (bp->b_flags & B_INVAL) 2431 bp->b_flags &= ~B_CACHE; 2432 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2433 bp->b_flags |= B_CACHE; 2434 bremfree(bp); 2435 2436 /* 2437 * check for size inconsistancies for non-VMIO case. 2438 */ 2439 2440 if (bp->b_bcount != size) { 2441 if ((bp->b_flags & B_VMIO) == 0 || 2442 (size > bp->b_kvasize)) { 2443 if (bp->b_flags & B_DELWRI) { 2444 /* 2445 * If buffer is pinned and caller does 2446 * not want sleep waiting for it to be 2447 * unpinned, bail out 2448 * */ 2449 if (bp->b_pin_count > 0) { 2450 if (flags & GB_LOCK_NOWAIT) { 2451 bqrelse(bp); 2452 return (NULL); 2453 } else { 2454 bunpin_wait(bp); 2455 } 2456 } 2457 bp->b_flags |= B_NOCACHE; 2458 bwrite(bp); 2459 } else { 2460 if (LIST_FIRST(&bp->b_dep) == NULL) { 2461 bp->b_flags |= B_RELBUF; 2462 brelse(bp); 2463 } else { 2464 bp->b_flags |= B_NOCACHE; 2465 bwrite(bp); 2466 } 2467 } 2468 goto loop; 2469 } 2470 } 2471 2472 /* 2473 * If the size is inconsistant in the VMIO case, we can resize 2474 * the buffer. This might lead to B_CACHE getting set or 2475 * cleared. If the size has not changed, B_CACHE remains 2476 * unchanged from its previous state. 2477 */ 2478 2479 if (bp->b_bcount != size) 2480 allocbuf(bp, size); 2481 2482 KASSERT(bp->b_offset != NOOFFSET, 2483 ("getblk: no buffer offset")); 2484 2485 /* 2486 * A buffer with B_DELWRI set and B_CACHE clear must 2487 * be committed before we can return the buffer in 2488 * order to prevent the caller from issuing a read 2489 * ( due to B_CACHE not being set ) and overwriting 2490 * it. 2491 * 2492 * Most callers, including NFS and FFS, need this to 2493 * operate properly either because they assume they 2494 * can issue a read if B_CACHE is not set, or because 2495 * ( for example ) an uncached B_DELWRI might loop due 2496 * to softupdates re-dirtying the buffer. In the latter 2497 * case, B_CACHE is set after the first write completes, 2498 * preventing further loops. 2499 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2500 * above while extending the buffer, we cannot allow the 2501 * buffer to remain with B_CACHE set after the write 2502 * completes or it will represent a corrupt state. To 2503 * deal with this we set B_NOCACHE to scrap the buffer 2504 * after the write. 2505 * 2506 * We might be able to do something fancy, like setting 2507 * B_CACHE in bwrite() except if B_DELWRI is already set, 2508 * so the below call doesn't set B_CACHE, but that gets real 2509 * confusing. This is much easier. 2510 */ 2511 2512 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2513 bp->b_flags |= B_NOCACHE; 2514 bwrite(bp); 2515 goto loop; 2516 } 2517 bp->b_flags &= ~B_DONE; 2518 } else { 2519 int bsize, maxsize, vmio; 2520 off_t offset; 2521 2522 /* 2523 * Buffer is not in-core, create new buffer. The buffer 2524 * returned by getnewbuf() is locked. Note that the returned 2525 * buffer is also considered valid (not marked B_INVAL). 2526 */ 2527 VI_UNLOCK(vp); 2528 /* 2529 * If the user does not want us to create the buffer, bail out 2530 * here. 2531 */ 2532 if (flags & GB_NOCREAT) 2533 return NULL; 2534 bsize = bo->bo_bsize; 2535 offset = blkno * bsize; 2536 vmio = vp->v_object != NULL; 2537 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2538 maxsize = imax(maxsize, bsize); 2539 2540 bp = getnewbuf(slpflag, slptimeo, size, maxsize); 2541 if (bp == NULL) { 2542 if (slpflag || slptimeo) 2543 return NULL; 2544 goto loop; 2545 } 2546 2547 /* 2548 * This code is used to make sure that a buffer is not 2549 * created while the getnewbuf routine is blocked. 2550 * This can be a problem whether the vnode is locked or not. 2551 * If the buffer is created out from under us, we have to 2552 * throw away the one we just created. 2553 * 2554 * Note: this must occur before we associate the buffer 2555 * with the vp especially considering limitations in 2556 * the splay tree implementation when dealing with duplicate 2557 * lblkno's. 2558 */ 2559 BO_LOCK(bo); 2560 if (gbincore(bo, blkno)) { 2561 BO_UNLOCK(bo); 2562 bp->b_flags |= B_INVAL; 2563 brelse(bp); 2564 goto loop; 2565 } 2566 2567 /* 2568 * Insert the buffer into the hash, so that it can 2569 * be found by incore. 2570 */ 2571 bp->b_blkno = bp->b_lblkno = blkno; 2572 bp->b_offset = offset; 2573 2574 bgetvp(vp, bp); 2575 BO_UNLOCK(bo); 2576 2577 /* 2578 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2579 * buffer size starts out as 0, B_CACHE will be set by 2580 * allocbuf() for the VMIO case prior to it testing the 2581 * backing store for validity. 2582 */ 2583 2584 if (vmio) { 2585 bp->b_flags |= B_VMIO; 2586#if defined(VFS_BIO_DEBUG) 2587 if (vn_canvmio(vp) != TRUE) 2588 printf("getblk: VMIO on vnode type %d\n", 2589 vp->v_type); 2590#endif 2591 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2592 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2593 bp, vp->v_object, bp->b_bufobj->bo_object)); 2594 } else { 2595 bp->b_flags &= ~B_VMIO; 2596 KASSERT(bp->b_bufobj->bo_object == NULL, 2597 ("ARGH! has b_bufobj->bo_object %p %p\n", 2598 bp, bp->b_bufobj->bo_object)); 2599 } 2600 2601 allocbuf(bp, size); 2602 bp->b_flags &= ~B_DONE; 2603 } 2604 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2605 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2606 KASSERT(bp->b_bufobj == bo, 2607 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2608 return (bp); 2609} 2610 2611/* 2612 * Get an empty, disassociated buffer of given size. The buffer is initially 2613 * set to B_INVAL. 2614 */ 2615struct buf * 2616geteblk(int size) 2617{ 2618 struct buf *bp; 2619 int maxsize; 2620 2621 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2622 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2623 continue; 2624 allocbuf(bp, size); 2625 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2626 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2627 return (bp); 2628} 2629 2630 2631/* 2632 * This code constitutes the buffer memory from either anonymous system 2633 * memory (in the case of non-VMIO operations) or from an associated 2634 * VM object (in the case of VMIO operations). This code is able to 2635 * resize a buffer up or down. 2636 * 2637 * Note that this code is tricky, and has many complications to resolve 2638 * deadlock or inconsistant data situations. Tread lightly!!! 2639 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2640 * the caller. Calling this code willy nilly can result in the loss of data. 2641 * 2642 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2643 * B_CACHE for the non-VMIO case. 2644 */ 2645 2646int 2647allocbuf(struct buf *bp, int size) 2648{ 2649 int newbsize, mbsize; 2650 int i; 2651 2652 if (BUF_REFCNT(bp) == 0) 2653 panic("allocbuf: buffer not busy"); 2654 2655 if (bp->b_kvasize < size) 2656 panic("allocbuf: buffer too small"); 2657 2658 if ((bp->b_flags & B_VMIO) == 0) { 2659 caddr_t origbuf; 2660 int origbufsize; 2661 /* 2662 * Just get anonymous memory from the kernel. Don't 2663 * mess with B_CACHE. 2664 */ 2665 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2666 if (bp->b_flags & B_MALLOC) 2667 newbsize = mbsize; 2668 else 2669 newbsize = round_page(size); 2670 2671 if (newbsize < bp->b_bufsize) { 2672 /* 2673 * malloced buffers are not shrunk 2674 */ 2675 if (bp->b_flags & B_MALLOC) { 2676 if (newbsize) { 2677 bp->b_bcount = size; 2678 } else { 2679 free(bp->b_data, M_BIOBUF); 2680 if (bp->b_bufsize) { 2681 atomic_subtract_int( 2682 &bufmallocspace, 2683 bp->b_bufsize); 2684 bufspacewakeup(); 2685 bp->b_bufsize = 0; 2686 } 2687 bp->b_saveaddr = bp->b_kvabase; 2688 bp->b_data = bp->b_saveaddr; 2689 bp->b_bcount = 0; 2690 bp->b_flags &= ~B_MALLOC; 2691 } 2692 return 1; 2693 } 2694 vm_hold_free_pages( 2695 bp, 2696 (vm_offset_t) bp->b_data + newbsize, 2697 (vm_offset_t) bp->b_data + bp->b_bufsize); 2698 } else if (newbsize > bp->b_bufsize) { 2699 /* 2700 * We only use malloced memory on the first allocation. 2701 * and revert to page-allocated memory when the buffer 2702 * grows. 2703 */ 2704 /* 2705 * There is a potential smp race here that could lead 2706 * to bufmallocspace slightly passing the max. It 2707 * is probably extremely rare and not worth worrying 2708 * over. 2709 */ 2710 if ( (bufmallocspace < maxbufmallocspace) && 2711 (bp->b_bufsize == 0) && 2712 (mbsize <= PAGE_SIZE/2)) { 2713 2714 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2715 bp->b_bufsize = mbsize; 2716 bp->b_bcount = size; 2717 bp->b_flags |= B_MALLOC; 2718 atomic_add_int(&bufmallocspace, mbsize); 2719 return 1; 2720 } 2721 origbuf = NULL; 2722 origbufsize = 0; 2723 /* 2724 * If the buffer is growing on its other-than-first allocation, 2725 * then we revert to the page-allocation scheme. 2726 */ 2727 if (bp->b_flags & B_MALLOC) { 2728 origbuf = bp->b_data; 2729 origbufsize = bp->b_bufsize; 2730 bp->b_data = bp->b_kvabase; 2731 if (bp->b_bufsize) { 2732 atomic_subtract_int(&bufmallocspace, 2733 bp->b_bufsize); 2734 bufspacewakeup(); 2735 bp->b_bufsize = 0; 2736 } 2737 bp->b_flags &= ~B_MALLOC; 2738 newbsize = round_page(newbsize); 2739 } 2740 vm_hold_load_pages( 2741 bp, 2742 (vm_offset_t) bp->b_data + bp->b_bufsize, 2743 (vm_offset_t) bp->b_data + newbsize); 2744 if (origbuf) { 2745 bcopy(origbuf, bp->b_data, origbufsize); 2746 free(origbuf, M_BIOBUF); 2747 } 2748 } 2749 } else { 2750 int desiredpages; 2751 2752 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2753 desiredpages = (size == 0) ? 0 : 2754 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2755 2756 if (bp->b_flags & B_MALLOC) 2757 panic("allocbuf: VMIO buffer can't be malloced"); 2758 /* 2759 * Set B_CACHE initially if buffer is 0 length or will become 2760 * 0-length. 2761 */ 2762 if (size == 0 || bp->b_bufsize == 0) 2763 bp->b_flags |= B_CACHE; 2764 2765 if (newbsize < bp->b_bufsize) { 2766 /* 2767 * DEV_BSIZE aligned new buffer size is less then the 2768 * DEV_BSIZE aligned existing buffer size. Figure out 2769 * if we have to remove any pages. 2770 */ 2771 if (desiredpages < bp->b_npages) { 2772 vm_page_t m; 2773 2774 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2775 vm_page_lock_queues(); 2776 for (i = desiredpages; i < bp->b_npages; i++) { 2777 /* 2778 * the page is not freed here -- it 2779 * is the responsibility of 2780 * vnode_pager_setsize 2781 */ 2782 m = bp->b_pages[i]; 2783 KASSERT(m != bogus_page, 2784 ("allocbuf: bogus page found")); 2785 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2786 vm_page_lock_queues(); 2787 2788 bp->b_pages[i] = NULL; 2789 vm_page_unwire(m, 0); 2790 } 2791 vm_page_unlock_queues(); 2792 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2793 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2794 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2795 bp->b_npages = desiredpages; 2796 } 2797 } else if (size > bp->b_bcount) { 2798 /* 2799 * We are growing the buffer, possibly in a 2800 * byte-granular fashion. 2801 */ 2802 struct vnode *vp; 2803 vm_object_t obj; 2804 vm_offset_t toff; 2805 vm_offset_t tinc; 2806 2807 /* 2808 * Step 1, bring in the VM pages from the object, 2809 * allocating them if necessary. We must clear 2810 * B_CACHE if these pages are not valid for the 2811 * range covered by the buffer. 2812 */ 2813 2814 vp = bp->b_vp; 2815 obj = bp->b_bufobj->bo_object; 2816 2817 VM_OBJECT_LOCK(obj); 2818 while (bp->b_npages < desiredpages) { 2819 vm_page_t m; 2820 vm_pindex_t pi; 2821 2822 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2823 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2824 /* 2825 * note: must allocate system pages 2826 * since blocking here could intefere 2827 * with paging I/O, no matter which 2828 * process we are. 2829 */ 2830 m = vm_page_alloc(obj, pi, 2831 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2832 VM_ALLOC_WIRED); 2833 if (m == NULL) { 2834 atomic_add_int(&vm_pageout_deficit, 2835 desiredpages - bp->b_npages); 2836 VM_OBJECT_UNLOCK(obj); 2837 VM_WAIT; 2838 VM_OBJECT_LOCK(obj); 2839 } else { 2840 bp->b_flags &= ~B_CACHE; 2841 bp->b_pages[bp->b_npages] = m; 2842 ++bp->b_npages; 2843 } 2844 continue; 2845 } 2846 2847 /* 2848 * We found a page. If we have to sleep on it, 2849 * retry because it might have gotten freed out 2850 * from under us. 2851 * 2852 * We can only test PG_BUSY here. Blocking on 2853 * m->busy might lead to a deadlock: 2854 * 2855 * vm_fault->getpages->cluster_read->allocbuf 2856 * 2857 */ 2858 vm_page_lock_queues(); 2859 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2860 continue; 2861 2862 /* 2863 * We have a good page. Should we wakeup the 2864 * page daemon? 2865 */ 2866 if ((curproc != pageproc) && 2867 (VM_PAGE_INQUEUE1(m, PQ_CACHE)) && 2868 ((cnt.v_free_count + cnt.v_cache_count) < 2869 (cnt.v_free_min + cnt.v_cache_min))) { 2870 pagedaemon_wakeup(); 2871 } 2872 vm_page_wire(m); 2873 vm_page_unlock_queues(); 2874 bp->b_pages[bp->b_npages] = m; 2875 ++bp->b_npages; 2876 } 2877 2878 /* 2879 * Step 2. We've loaded the pages into the buffer, 2880 * we have to figure out if we can still have B_CACHE 2881 * set. Note that B_CACHE is set according to the 2882 * byte-granular range ( bcount and size ), new the 2883 * aligned range ( newbsize ). 2884 * 2885 * The VM test is against m->valid, which is DEV_BSIZE 2886 * aligned. Needless to say, the validity of the data 2887 * needs to also be DEV_BSIZE aligned. Note that this 2888 * fails with NFS if the server or some other client 2889 * extends the file's EOF. If our buffer is resized, 2890 * B_CACHE may remain set! XXX 2891 */ 2892 2893 toff = bp->b_bcount; 2894 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2895 2896 while ((bp->b_flags & B_CACHE) && toff < size) { 2897 vm_pindex_t pi; 2898 2899 if (tinc > (size - toff)) 2900 tinc = size - toff; 2901 2902 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2903 PAGE_SHIFT; 2904 2905 vfs_buf_test_cache( 2906 bp, 2907 bp->b_offset, 2908 toff, 2909 tinc, 2910 bp->b_pages[pi] 2911 ); 2912 toff += tinc; 2913 tinc = PAGE_SIZE; 2914 } 2915 VM_OBJECT_UNLOCK(obj); 2916 2917 /* 2918 * Step 3, fixup the KVM pmap. Remember that 2919 * bp->b_data is relative to bp->b_offset, but 2920 * bp->b_offset may be offset into the first page. 2921 */ 2922 2923 bp->b_data = (caddr_t) 2924 trunc_page((vm_offset_t)bp->b_data); 2925 pmap_qenter( 2926 (vm_offset_t)bp->b_data, 2927 bp->b_pages, 2928 bp->b_npages 2929 ); 2930 2931 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2932 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2933 } 2934 } 2935 if (newbsize < bp->b_bufsize) 2936 bufspacewakeup(); 2937 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2938 bp->b_bcount = size; /* requested buffer size */ 2939 return 1; 2940} 2941 2942void 2943biodone(struct bio *bp) 2944{ 2945 void (*done)(struct bio *); 2946 2947 mtx_lock(&bdonelock); 2948 bp->bio_flags |= BIO_DONE; 2949 done = bp->bio_done; 2950 if (done == NULL) 2951 wakeup(bp); 2952 mtx_unlock(&bdonelock); 2953 if (done != NULL) 2954 done(bp); 2955} 2956 2957/* 2958 * Wait for a BIO to finish. 2959 * 2960 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2961 * case is not yet clear. 2962 */ 2963int 2964biowait(struct bio *bp, const char *wchan) 2965{ 2966 2967 mtx_lock(&bdonelock); 2968 while ((bp->bio_flags & BIO_DONE) == 0) 2969 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 2970 mtx_unlock(&bdonelock); 2971 if (bp->bio_error != 0) 2972 return (bp->bio_error); 2973 if (!(bp->bio_flags & BIO_ERROR)) 2974 return (0); 2975 return (EIO); 2976} 2977 2978void 2979biofinish(struct bio *bp, struct devstat *stat, int error) 2980{ 2981 2982 if (error) { 2983 bp->bio_error = error; 2984 bp->bio_flags |= BIO_ERROR; 2985 } 2986 if (stat != NULL) 2987 devstat_end_transaction_bio(stat, bp); 2988 biodone(bp); 2989} 2990 2991/* 2992 * bufwait: 2993 * 2994 * Wait for buffer I/O completion, returning error status. The buffer 2995 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 2996 * error and cleared. 2997 */ 2998int 2999bufwait(struct buf *bp) 3000{ 3001 if (bp->b_iocmd == BIO_READ) 3002 bwait(bp, PRIBIO, "biord"); 3003 else 3004 bwait(bp, PRIBIO, "biowr"); 3005 if (bp->b_flags & B_EINTR) { 3006 bp->b_flags &= ~B_EINTR; 3007 return (EINTR); 3008 } 3009 if (bp->b_ioflags & BIO_ERROR) { 3010 return (bp->b_error ? bp->b_error : EIO); 3011 } else { 3012 return (0); 3013 } 3014} 3015 3016 /* 3017 * Call back function from struct bio back up to struct buf. 3018 */ 3019static void 3020bufdonebio(struct bio *bip) 3021{ 3022 struct buf *bp; 3023 3024 bp = bip->bio_caller2; 3025 bp->b_resid = bp->b_bcount - bip->bio_completed; 3026 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3027 bp->b_ioflags = bip->bio_flags; 3028 bp->b_error = bip->bio_error; 3029 if (bp->b_error) 3030 bp->b_ioflags |= BIO_ERROR; 3031 bufdone(bp); 3032 g_destroy_bio(bip); 3033} 3034 3035void 3036dev_strategy(struct cdev *dev, struct buf *bp) 3037{ 3038 struct cdevsw *csw; 3039 struct bio *bip; 3040 3041 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3042 panic("b_iocmd botch"); 3043 for (;;) { 3044 bip = g_new_bio(); 3045 if (bip != NULL) 3046 break; 3047 /* Try again later */ 3048 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3049 } 3050 bip->bio_cmd = bp->b_iocmd; 3051 bip->bio_offset = bp->b_iooffset; 3052 bip->bio_length = bp->b_bcount; 3053 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3054 bip->bio_data = bp->b_data; 3055 bip->bio_done = bufdonebio; 3056 bip->bio_caller2 = bp; 3057 bip->bio_dev = dev; 3058 KASSERT(dev->si_refcount > 0, 3059 ("dev_strategy on un-referenced struct cdev *(%s)", 3060 devtoname(dev))); 3061 csw = dev_refthread(dev); 3062 if (csw == NULL) { 3063 bp->b_error = ENXIO; 3064 bp->b_ioflags = BIO_ERROR; 3065 bufdone(bp); 3066 return; 3067 } 3068 (*csw->d_strategy)(bip); 3069 dev_relthread(dev); 3070} 3071 3072/* 3073 * bufdone: 3074 * 3075 * Finish I/O on a buffer, optionally calling a completion function. 3076 * This is usually called from an interrupt so process blocking is 3077 * not allowed. 3078 * 3079 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3080 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3081 * assuming B_INVAL is clear. 3082 * 3083 * For the VMIO case, we set B_CACHE if the op was a read and no 3084 * read error occured, or if the op was a write. B_CACHE is never 3085 * set if the buffer is invalid or otherwise uncacheable. 3086 * 3087 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3088 * initiator to leave B_INVAL set to brelse the buffer out of existance 3089 * in the biodone routine. 3090 */ 3091void 3092bufdone(struct buf *bp) 3093{ 3094 struct bufobj *dropobj; 3095 void (*biodone)(struct buf *); 3096 3097 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3098 dropobj = NULL; 3099 3100 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, 3101 BUF_REFCNT(bp))); 3102 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3103 3104 runningbufwakeup(bp); 3105 if (bp->b_iocmd == BIO_WRITE) 3106 dropobj = bp->b_bufobj; 3107 /* call optional completion function if requested */ 3108 if (bp->b_iodone != NULL) { 3109 biodone = bp->b_iodone; 3110 bp->b_iodone = NULL; 3111 (*biodone) (bp); 3112 if (dropobj) 3113 bufobj_wdrop(dropobj); 3114 return; 3115 } 3116 3117 bufdone_finish(bp); 3118 3119 if (dropobj) 3120 bufobj_wdrop(dropobj); 3121} 3122 3123void 3124bufdone_finish(struct buf *bp) 3125{ 3126 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, 3127 BUF_REFCNT(bp))); 3128 3129 if (LIST_FIRST(&bp->b_dep) != NULL) 3130 buf_complete(bp); 3131 3132 if (bp->b_flags & B_VMIO) { 3133 int i; 3134 vm_ooffset_t foff; 3135 vm_page_t m; 3136 vm_object_t obj; 3137 int iosize; 3138 struct vnode *vp = bp->b_vp; 3139 3140 obj = bp->b_bufobj->bo_object; 3141 3142#if defined(VFS_BIO_DEBUG) 3143 mp_fixme("usecount and vflag accessed without locks."); 3144 if (vp->v_usecount == 0) { 3145 panic("biodone: zero vnode ref count"); 3146 } 3147 3148 KASSERT(vp->v_object != NULL, 3149 ("biodone: vnode %p has no vm_object", vp)); 3150#endif 3151 3152 foff = bp->b_offset; 3153 KASSERT(bp->b_offset != NOOFFSET, 3154 ("biodone: no buffer offset")); 3155 3156 VM_OBJECT_LOCK(obj); 3157#if defined(VFS_BIO_DEBUG) 3158 if (obj->paging_in_progress < bp->b_npages) { 3159 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3160 obj->paging_in_progress, bp->b_npages); 3161 } 3162#endif 3163 3164 /* 3165 * Set B_CACHE if the op was a normal read and no error 3166 * occured. B_CACHE is set for writes in the b*write() 3167 * routines. 3168 */ 3169 iosize = bp->b_bcount - bp->b_resid; 3170 if (bp->b_iocmd == BIO_READ && 3171 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3172 !(bp->b_ioflags & BIO_ERROR)) { 3173 bp->b_flags |= B_CACHE; 3174 } 3175 vm_page_lock_queues(); 3176 for (i = 0; i < bp->b_npages; i++) { 3177 int bogusflag = 0; 3178 int resid; 3179 3180 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3181 if (resid > iosize) 3182 resid = iosize; 3183 3184 /* 3185 * cleanup bogus pages, restoring the originals 3186 */ 3187 m = bp->b_pages[i]; 3188 if (m == bogus_page) { 3189 bogusflag = 1; 3190 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3191 if (m == NULL) 3192 panic("biodone: page disappeared!"); 3193 bp->b_pages[i] = m; 3194 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3195 bp->b_pages, bp->b_npages); 3196 } 3197#if defined(VFS_BIO_DEBUG) 3198 if (OFF_TO_IDX(foff) != m->pindex) { 3199 printf( 3200"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3201 (intmax_t)foff, (uintmax_t)m->pindex); 3202 } 3203#endif 3204 3205 /* 3206 * In the write case, the valid and clean bits are 3207 * already changed correctly ( see bdwrite() ), so we 3208 * only need to do this here in the read case. 3209 */ 3210 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3211 vfs_page_set_valid(bp, foff, i, m); 3212 } 3213 3214 /* 3215 * when debugging new filesystems or buffer I/O methods, this 3216 * is the most common error that pops up. if you see this, you 3217 * have not set the page busy flag correctly!!! 3218 */ 3219 if (m->busy == 0) { 3220 printf("biodone: page busy < 0, " 3221 "pindex: %d, foff: 0x(%x,%x), " 3222 "resid: %d, index: %d\n", 3223 (int) m->pindex, (int)(foff >> 32), 3224 (int) foff & 0xffffffff, resid, i); 3225 if (!vn_isdisk(vp, NULL)) 3226 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3227 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3228 (intmax_t) bp->b_lblkno, 3229 bp->b_flags, bp->b_npages); 3230 else 3231 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3232 (intmax_t) bp->b_lblkno, 3233 bp->b_flags, bp->b_npages); 3234 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3235 (u_long)m->valid, (u_long)m->dirty, 3236 m->wire_count); 3237 panic("biodone: page busy < 0\n"); 3238 } 3239 vm_page_io_finish(m); 3240 vm_object_pip_subtract(obj, 1); 3241 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3242 iosize -= resid; 3243 } 3244 vm_page_unlock_queues(); 3245 vm_object_pip_wakeupn(obj, 0); 3246 VM_OBJECT_UNLOCK(obj); 3247 } 3248 3249 /* 3250 * For asynchronous completions, release the buffer now. The brelse 3251 * will do a wakeup there if necessary - so no need to do a wakeup 3252 * here in the async case. The sync case always needs to do a wakeup. 3253 */ 3254 3255 if (bp->b_flags & B_ASYNC) { 3256 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3257 brelse(bp); 3258 else 3259 bqrelse(bp); 3260 } else 3261 bdone(bp); 3262} 3263 3264/* 3265 * This routine is called in lieu of iodone in the case of 3266 * incomplete I/O. This keeps the busy status for pages 3267 * consistant. 3268 */ 3269void 3270vfs_unbusy_pages(struct buf *bp) 3271{ 3272 int i; 3273 vm_object_t obj; 3274 vm_page_t m; 3275 3276 runningbufwakeup(bp); 3277 if (!(bp->b_flags & B_VMIO)) 3278 return; 3279 3280 obj = bp->b_bufobj->bo_object; 3281 VM_OBJECT_LOCK(obj); 3282 vm_page_lock_queues(); 3283 for (i = 0; i < bp->b_npages; i++) { 3284 m = bp->b_pages[i]; 3285 if (m == bogus_page) { 3286 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3287 if (!m) 3288 panic("vfs_unbusy_pages: page missing\n"); 3289 bp->b_pages[i] = m; 3290 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3291 bp->b_pages, bp->b_npages); 3292 } 3293 vm_object_pip_subtract(obj, 1); 3294 vm_page_io_finish(m); 3295 } 3296 vm_page_unlock_queues(); 3297 vm_object_pip_wakeupn(obj, 0); 3298 VM_OBJECT_UNLOCK(obj); 3299} 3300 3301/* 3302 * vfs_page_set_valid: 3303 * 3304 * Set the valid bits in a page based on the supplied offset. The 3305 * range is restricted to the buffer's size. 3306 * 3307 * This routine is typically called after a read completes. 3308 */ 3309static void 3310vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3311{ 3312 vm_ooffset_t soff, eoff; 3313 3314 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3315 /* 3316 * Start and end offsets in buffer. eoff - soff may not cross a 3317 * page boundry or cross the end of the buffer. The end of the 3318 * buffer, in this case, is our file EOF, not the allocation size 3319 * of the buffer. 3320 */ 3321 soff = off; 3322 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3323 if (eoff > bp->b_offset + bp->b_bcount) 3324 eoff = bp->b_offset + bp->b_bcount; 3325 3326 /* 3327 * Set valid range. This is typically the entire buffer and thus the 3328 * entire page. 3329 */ 3330 if (eoff > soff) { 3331 vm_page_set_validclean( 3332 m, 3333 (vm_offset_t) (soff & PAGE_MASK), 3334 (vm_offset_t) (eoff - soff) 3335 ); 3336 } 3337} 3338 3339/* 3340 * This routine is called before a device strategy routine. 3341 * It is used to tell the VM system that paging I/O is in 3342 * progress, and treat the pages associated with the buffer 3343 * almost as being PG_BUSY. Also the object paging_in_progress 3344 * flag is handled to make sure that the object doesn't become 3345 * inconsistant. 3346 * 3347 * Since I/O has not been initiated yet, certain buffer flags 3348 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3349 * and should be ignored. 3350 */ 3351void 3352vfs_busy_pages(struct buf *bp, int clear_modify) 3353{ 3354 int i, bogus; 3355 vm_object_t obj; 3356 vm_ooffset_t foff; 3357 vm_page_t m; 3358 3359 if (!(bp->b_flags & B_VMIO)) 3360 return; 3361 3362 obj = bp->b_bufobj->bo_object; 3363 foff = bp->b_offset; 3364 KASSERT(bp->b_offset != NOOFFSET, 3365 ("vfs_busy_pages: no buffer offset")); 3366 vfs_setdirty(bp); 3367 VM_OBJECT_LOCK(obj); 3368retry: 3369 vm_page_lock_queues(); 3370 for (i = 0; i < bp->b_npages; i++) { 3371 m = bp->b_pages[i]; 3372 3373 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3374 goto retry; 3375 } 3376 bogus = 0; 3377 for (i = 0; i < bp->b_npages; i++) { 3378 m = bp->b_pages[i]; 3379 3380 if ((bp->b_flags & B_CLUSTER) == 0) { 3381 vm_object_pip_add(obj, 1); 3382 vm_page_io_start(m); 3383 } 3384 /* 3385 * When readying a buffer for a read ( i.e 3386 * clear_modify == 0 ), it is important to do 3387 * bogus_page replacement for valid pages in 3388 * partially instantiated buffers. Partially 3389 * instantiated buffers can, in turn, occur when 3390 * reconstituting a buffer from its VM backing store 3391 * base. We only have to do this if B_CACHE is 3392 * clear ( which causes the I/O to occur in the 3393 * first place ). The replacement prevents the read 3394 * I/O from overwriting potentially dirty VM-backed 3395 * pages. XXX bogus page replacement is, uh, bogus. 3396 * It may not work properly with small-block devices. 3397 * We need to find a better way. 3398 */ 3399 pmap_remove_all(m); 3400 if (clear_modify) 3401 vfs_page_set_valid(bp, foff, i, m); 3402 else if (m->valid == VM_PAGE_BITS_ALL && 3403 (bp->b_flags & B_CACHE) == 0) { 3404 bp->b_pages[i] = bogus_page; 3405 bogus++; 3406 } 3407 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3408 } 3409 vm_page_unlock_queues(); 3410 VM_OBJECT_UNLOCK(obj); 3411 if (bogus) 3412 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3413 bp->b_pages, bp->b_npages); 3414} 3415 3416/* 3417 * Tell the VM system that the pages associated with this buffer 3418 * are clean. This is used for delayed writes where the data is 3419 * going to go to disk eventually without additional VM intevention. 3420 * 3421 * Note that while we only really need to clean through to b_bcount, we 3422 * just go ahead and clean through to b_bufsize. 3423 */ 3424static void 3425vfs_clean_pages(struct buf *bp) 3426{ 3427 int i; 3428 vm_ooffset_t foff, noff, eoff; 3429 vm_page_t m; 3430 3431 if (!(bp->b_flags & B_VMIO)) 3432 return; 3433 3434 foff = bp->b_offset; 3435 KASSERT(bp->b_offset != NOOFFSET, 3436 ("vfs_clean_pages: no buffer offset")); 3437 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3438 vm_page_lock_queues(); 3439 for (i = 0; i < bp->b_npages; i++) { 3440 m = bp->b_pages[i]; 3441 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3442 eoff = noff; 3443 3444 if (eoff > bp->b_offset + bp->b_bufsize) 3445 eoff = bp->b_offset + bp->b_bufsize; 3446 vfs_page_set_valid(bp, foff, i, m); 3447 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3448 foff = noff; 3449 } 3450 vm_page_unlock_queues(); 3451 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3452} 3453 3454/* 3455 * vfs_bio_set_validclean: 3456 * 3457 * Set the range within the buffer to valid and clean. The range is 3458 * relative to the beginning of the buffer, b_offset. Note that b_offset 3459 * itself may be offset from the beginning of the first page. 3460 * 3461 */ 3462 3463void 3464vfs_bio_set_validclean(struct buf *bp, int base, int size) 3465{ 3466 int i, n; 3467 vm_page_t m; 3468 3469 if (!(bp->b_flags & B_VMIO)) 3470 return; 3471 /* 3472 * Fixup base to be relative to beginning of first page. 3473 * Set initial n to be the maximum number of bytes in the 3474 * first page that can be validated. 3475 */ 3476 3477 base += (bp->b_offset & PAGE_MASK); 3478 n = PAGE_SIZE - (base & PAGE_MASK); 3479 3480 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3481 vm_page_lock_queues(); 3482 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3483 m = bp->b_pages[i]; 3484 if (n > size) 3485 n = size; 3486 vm_page_set_validclean(m, base & PAGE_MASK, n); 3487 base += n; 3488 size -= n; 3489 n = PAGE_SIZE; 3490 } 3491 vm_page_unlock_queues(); 3492 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3493} 3494 3495/* 3496 * vfs_bio_clrbuf: 3497 * 3498 * clear a buffer. This routine essentially fakes an I/O, so we need 3499 * to clear BIO_ERROR and B_INVAL. 3500 * 3501 * Note that while we only theoretically need to clear through b_bcount, 3502 * we go ahead and clear through b_bufsize. 3503 */ 3504 3505void 3506vfs_bio_clrbuf(struct buf *bp) 3507{ 3508 int i, j, mask = 0; 3509 caddr_t sa, ea; 3510 3511 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3512 clrbuf(bp); 3513 return; 3514 } 3515 3516 bp->b_flags &= ~B_INVAL; 3517 bp->b_ioflags &= ~BIO_ERROR; 3518 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3519 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3520 (bp->b_offset & PAGE_MASK) == 0) { 3521 if (bp->b_pages[0] == bogus_page) 3522 goto unlock; 3523 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3524 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3525 if ((bp->b_pages[0]->valid & mask) == mask) 3526 goto unlock; 3527 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3528 ((bp->b_pages[0]->valid & mask) == 0)) { 3529 bzero(bp->b_data, bp->b_bufsize); 3530 bp->b_pages[0]->valid |= mask; 3531 goto unlock; 3532 } 3533 } 3534 ea = sa = bp->b_data; 3535 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3536 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3537 ea = (caddr_t)(vm_offset_t)ulmin( 3538 (u_long)(vm_offset_t)ea, 3539 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3540 if (bp->b_pages[i] == bogus_page) 3541 continue; 3542 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3543 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3544 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3545 if ((bp->b_pages[i]->valid & mask) == mask) 3546 continue; 3547 if ((bp->b_pages[i]->valid & mask) == 0) { 3548 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3549 bzero(sa, ea - sa); 3550 } else { 3551 for (; sa < ea; sa += DEV_BSIZE, j++) { 3552 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3553 (bp->b_pages[i]->valid & (1 << j)) == 0) 3554 bzero(sa, DEV_BSIZE); 3555 } 3556 } 3557 bp->b_pages[i]->valid |= mask; 3558 } 3559unlock: 3560 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3561 bp->b_resid = 0; 3562} 3563 3564/* 3565 * vm_hold_load_pages and vm_hold_free_pages get pages into 3566 * a buffers address space. The pages are anonymous and are 3567 * not associated with a file object. 3568 */ 3569static void 3570vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3571{ 3572 vm_offset_t pg; 3573 vm_page_t p; 3574 int index; 3575 3576 to = round_page(to); 3577 from = round_page(from); 3578 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3579 3580 VM_OBJECT_LOCK(kernel_object); 3581 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3582tryagain: 3583 /* 3584 * note: must allocate system pages since blocking here 3585 * could intefere with paging I/O, no matter which 3586 * process we are. 3587 */ 3588 p = vm_page_alloc(kernel_object, 3589 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3590 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3591 if (!p) { 3592 atomic_add_int(&vm_pageout_deficit, 3593 (to - pg) >> PAGE_SHIFT); 3594 VM_OBJECT_UNLOCK(kernel_object); 3595 VM_WAIT; 3596 VM_OBJECT_LOCK(kernel_object); 3597 goto tryagain; 3598 } 3599 p->valid = VM_PAGE_BITS_ALL; 3600 pmap_qenter(pg, &p, 1); 3601 bp->b_pages[index] = p; 3602 } 3603 VM_OBJECT_UNLOCK(kernel_object); 3604 bp->b_npages = index; 3605} 3606 3607/* Return pages associated with this buf to the vm system */ 3608static void 3609vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3610{ 3611 vm_offset_t pg; 3612 vm_page_t p; 3613 int index, newnpages; 3614 3615 from = round_page(from); 3616 to = round_page(to); 3617 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3618 3619 VM_OBJECT_LOCK(kernel_object); 3620 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3621 p = bp->b_pages[index]; 3622 if (p && (index < bp->b_npages)) { 3623 if (p->busy) { 3624 printf( 3625 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3626 (intmax_t)bp->b_blkno, 3627 (intmax_t)bp->b_lblkno); 3628 } 3629 bp->b_pages[index] = NULL; 3630 pmap_qremove(pg, 1); 3631 vm_page_lock_queues(); 3632 vm_page_unwire(p, 0); 3633 vm_page_free(p); 3634 vm_page_unlock_queues(); 3635 } 3636 } 3637 VM_OBJECT_UNLOCK(kernel_object); 3638 bp->b_npages = newnpages; 3639} 3640 3641/* 3642 * Map an IO request into kernel virtual address space. 3643 * 3644 * All requests are (re)mapped into kernel VA space. 3645 * Notice that we use b_bufsize for the size of the buffer 3646 * to be mapped. b_bcount might be modified by the driver. 3647 * 3648 * Note that even if the caller determines that the address space should 3649 * be valid, a race or a smaller-file mapped into a larger space may 3650 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3651 * check the return value. 3652 */ 3653int 3654vmapbuf(struct buf *bp) 3655{ 3656 caddr_t addr, kva; 3657 vm_prot_t prot; 3658 int pidx, i; 3659 struct vm_page *m; 3660 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3661 3662 if (bp->b_bufsize < 0) 3663 return (-1); 3664 prot = VM_PROT_READ; 3665 if (bp->b_iocmd == BIO_READ) 3666 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3667 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3668 addr < bp->b_data + bp->b_bufsize; 3669 addr += PAGE_SIZE, pidx++) { 3670 /* 3671 * Do the vm_fault if needed; do the copy-on-write thing 3672 * when reading stuff off device into memory. 3673 * 3674 * NOTE! Must use pmap_extract() because addr may be in 3675 * the userland address space, and kextract is only guarenteed 3676 * to work for the kernland address space (see: sparc64 port). 3677 */ 3678retry: 3679 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3680 prot) < 0) { 3681 vm_page_lock_queues(); 3682 for (i = 0; i < pidx; ++i) { 3683 vm_page_unhold(bp->b_pages[i]); 3684 bp->b_pages[i] = NULL; 3685 } 3686 vm_page_unlock_queues(); 3687 return(-1); 3688 } 3689 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3690 if (m == NULL) 3691 goto retry; 3692 bp->b_pages[pidx] = m; 3693 } 3694 if (pidx > btoc(MAXPHYS)) 3695 panic("vmapbuf: mapped more than MAXPHYS"); 3696 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3697 3698 kva = bp->b_saveaddr; 3699 bp->b_npages = pidx; 3700 bp->b_saveaddr = bp->b_data; 3701 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3702 return(0); 3703} 3704 3705/* 3706 * Free the io map PTEs associated with this IO operation. 3707 * We also invalidate the TLB entries and restore the original b_addr. 3708 */ 3709void 3710vunmapbuf(struct buf *bp) 3711{ 3712 int pidx; 3713 int npages; 3714 3715 npages = bp->b_npages; 3716 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3717 vm_page_lock_queues(); 3718 for (pidx = 0; pidx < npages; pidx++) 3719 vm_page_unhold(bp->b_pages[pidx]); 3720 vm_page_unlock_queues(); 3721 3722 bp->b_data = bp->b_saveaddr; 3723} 3724 3725void 3726bdone(struct buf *bp) 3727{ 3728 3729 mtx_lock(&bdonelock); 3730 bp->b_flags |= B_DONE; 3731 wakeup(bp); 3732 mtx_unlock(&bdonelock); 3733} 3734 3735void 3736bwait(struct buf *bp, u_char pri, const char *wchan) 3737{ 3738 3739 mtx_lock(&bdonelock); 3740 while ((bp->b_flags & B_DONE) == 0) 3741 msleep(bp, &bdonelock, pri, wchan, 0); 3742 mtx_unlock(&bdonelock); 3743} 3744 3745int 3746bufsync(struct bufobj *bo, int waitfor, struct thread *td) 3747{ 3748 3749 return (VOP_FSYNC(bo->__bo_vnode, waitfor, td)); 3750} 3751 3752void 3753bufstrategy(struct bufobj *bo, struct buf *bp) 3754{ 3755 int i = 0; 3756 struct vnode *vp; 3757 3758 vp = bp->b_vp; 3759 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3760 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3761 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3762 i = VOP_STRATEGY(vp, bp); 3763 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3764} 3765 3766void 3767bufobj_wrefl(struct bufobj *bo) 3768{ 3769 3770 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3771 ASSERT_BO_LOCKED(bo); 3772 bo->bo_numoutput++; 3773} 3774 3775void 3776bufobj_wref(struct bufobj *bo) 3777{ 3778 3779 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3780 BO_LOCK(bo); 3781 bo->bo_numoutput++; 3782 BO_UNLOCK(bo); 3783} 3784 3785void 3786bufobj_wdrop(struct bufobj *bo) 3787{ 3788 3789 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3790 BO_LOCK(bo); 3791 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3792 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3793 bo->bo_flag &= ~BO_WWAIT; 3794 wakeup(&bo->bo_numoutput); 3795 } 3796 BO_UNLOCK(bo); 3797} 3798 3799int 3800bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3801{ 3802 int error; 3803 3804 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3805 ASSERT_BO_LOCKED(bo); 3806 error = 0; 3807 while (bo->bo_numoutput) { 3808 bo->bo_flag |= BO_WWAIT; 3809 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3810 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3811 if (error) 3812 break; 3813 } 3814 return (error); 3815} 3816 3817void 3818bpin(struct buf *bp) 3819{ 3820 mtx_lock(&bpinlock); 3821 bp->b_pin_count++; 3822 mtx_unlock(&bpinlock); 3823} 3824 3825void 3826bunpin(struct buf *bp) 3827{ 3828 mtx_lock(&bpinlock); 3829 if (--bp->b_pin_count == 0) 3830 wakeup(bp); 3831 mtx_unlock(&bpinlock); 3832} 3833 3834void 3835bunpin_wait(struct buf *bp) 3836{ 3837 mtx_lock(&bpinlock); 3838 while (bp->b_pin_count > 0) 3839 msleep(bp, &bpinlock, PRIBIO, "bwunpin", 0); 3840 mtx_unlock(&bpinlock); 3841} 3842 3843#include "opt_ddb.h" 3844#ifdef DDB 3845#include <ddb/ddb.h> 3846 3847/* DDB command to show buffer data */ 3848DB_SHOW_COMMAND(buffer, db_show_buffer) 3849{ 3850 /* get args */ 3851 struct buf *bp = (struct buf *)addr; 3852 3853 if (!have_addr) { 3854 db_printf("usage: show buffer <addr>\n"); 3855 return; 3856 } 3857 3858 db_printf("buf at %p\n", bp); 3859 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3860 db_printf( 3861 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3862 "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n", 3863 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3864 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno); 3865 if (bp->b_npages) { 3866 int i; 3867 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3868 for (i = 0; i < bp->b_npages; i++) { 3869 vm_page_t m; 3870 m = bp->b_pages[i]; 3871 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3872 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3873 if ((i + 1) < bp->b_npages) 3874 db_printf(","); 3875 } 3876 db_printf("\n"); 3877 } 3878 lockmgr_printinfo(&bp->b_lock); 3879} 3880 3881DB_SHOW_COMMAND(lockedbufs, lockedbufs) 3882{ 3883 struct buf *bp; 3884 int i; 3885 3886 for (i = 0; i < nbuf; i++) { 3887 bp = &buf[i]; 3888 if (lockcount(&bp->b_lock)) { 3889 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 3890 db_printf("\n"); 3891 } 3892 } 3893} 3894#endif /* DDB */ 3895