vfs_bio.c revision 177475
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 177475 2008-03-21 12:38:44Z kib $"); 43 44#include <sys/param.h> 45#include <sys/systm.h> 46#include <sys/bio.h> 47#include <sys/conf.h> 48#include <sys/buf.h> 49#include <sys/devicestat.h> 50#include <sys/eventhandler.h> 51#include <sys/limits.h> 52#include <sys/lock.h> 53#include <sys/malloc.h> 54#include <sys/mount.h> 55#include <sys/mutex.h> 56#include <sys/kernel.h> 57#include <sys/kthread.h> 58#include <sys/proc.h> 59#include <sys/resourcevar.h> 60#include <sys/sysctl.h> 61#include <sys/vmmeter.h> 62#include <sys/vnode.h> 63#include <geom/geom.h> 64#include <vm/vm.h> 65#include <vm/vm_param.h> 66#include <vm/vm_kern.h> 67#include <vm/vm_pageout.h> 68#include <vm/vm_page.h> 69#include <vm/vm_object.h> 70#include <vm/vm_extern.h> 71#include <vm/vm_map.h> 72#include "opt_directio.h" 73#include "opt_swap.h" 74 75static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 76 77struct bio_ops bioops; /* I/O operation notification */ 78 79struct buf_ops buf_ops_bio = { 80 .bop_name = "buf_ops_bio", 81 .bop_write = bufwrite, 82 .bop_strategy = bufstrategy, 83 .bop_sync = bufsync, 84 .bop_bdflush = bufbdflush, 85}; 86 87/* 88 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 89 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 90 */ 91struct buf *buf; /* buffer header pool */ 92 93static struct proc *bufdaemonproc; 94 95static int inmem(struct vnode *vp, daddr_t blkno); 96static void vm_hold_free_pages(struct buf *bp, vm_offset_t from, 97 vm_offset_t to); 98static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 99 vm_offset_t to); 100static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 101 vm_page_t m); 102static void vfs_clean_pages(struct buf *bp); 103static void vfs_setdirty(struct buf *bp); 104static void vfs_setdirty_locked_object(struct buf *bp); 105static void vfs_vmio_release(struct buf *bp); 106static int vfs_bio_clcheck(struct vnode *vp, int size, 107 daddr_t lblkno, daddr_t blkno); 108static int flushbufqueues(int, int); 109static void buf_daemon(void); 110static void bremfreel(struct buf *bp); 111 112int vmiodirenable = TRUE; 113SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 114 "Use the VM system for directory writes"); 115int runningbufspace; 116SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 117 "Amount of presently outstanding async buffer io"); 118static int bufspace; 119SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 120 "KVA memory used for bufs"); 121static int maxbufspace; 122SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 123 "Maximum allowed value of bufspace (including buf_daemon)"); 124static int bufmallocspace; 125SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 126 "Amount of malloced memory for buffers"); 127static int maxbufmallocspace; 128SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 129 "Maximum amount of malloced memory for buffers"); 130static int lobufspace; 131SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 132 "Minimum amount of buffers we want to have"); 133int hibufspace; 134SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 135 "Maximum allowed value of bufspace (excluding buf_daemon)"); 136static int bufreusecnt; 137SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 138 "Number of times we have reused a buffer"); 139static int buffreekvacnt; 140SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 141 "Number of times we have freed the KVA space from some buffer"); 142static int bufdefragcnt; 143SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 144 "Number of times we have had to repeat buffer allocation to defragment"); 145static int lorunningspace; 146SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 147 "Minimum preferred space used for in-progress I/O"); 148static int hirunningspace; 149SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 150 "Maximum amount of space to use for in-progress I/O"); 151int dirtybufferflushes; 152SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 153 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 154int bdwriteskip; 155SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 156 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 157int altbufferflushes; 158SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 159 0, "Number of fsync flushes to limit dirty buffers"); 160static int recursiveflushes; 161SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 162 0, "Number of flushes skipped due to being recursive"); 163static int numdirtybuffers; 164SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 165 "Number of buffers that are dirty (has unwritten changes) at the moment"); 166static int lodirtybuffers; 167SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 168 "How many buffers we want to have free before bufdaemon can sleep"); 169static int hidirtybuffers; 170SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 171 "When the number of dirty buffers is considered severe"); 172int dirtybufthresh; 173SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 174 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 175static int numfreebuffers; 176SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 177 "Number of free buffers"); 178static int lofreebuffers; 179SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 180 "XXX Unused"); 181static int hifreebuffers; 182SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 183 "XXX Complicatedly unused"); 184static int getnewbufcalls; 185SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 186 "Number of calls to getnewbuf"); 187static int getnewbufrestarts; 188SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 189 "Number of times getnewbuf has had to restart a buffer aquisition"); 190 191/* 192 * Wakeup point for bufdaemon, as well as indicator of whether it is already 193 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 194 * is idling. 195 */ 196static int bd_request; 197 198/* 199 * This lock synchronizes access to bd_request. 200 */ 201static struct mtx bdlock; 202 203/* 204 * bogus page -- for I/O to/from partially complete buffers 205 * this is a temporary solution to the problem, but it is not 206 * really that bad. it would be better to split the buffer 207 * for input in the case of buffers partially already in memory, 208 * but the code is intricate enough already. 209 */ 210vm_page_t bogus_page; 211 212/* 213 * Synchronization (sleep/wakeup) variable for active buffer space requests. 214 * Set when wait starts, cleared prior to wakeup(). 215 * Used in runningbufwakeup() and waitrunningbufspace(). 216 */ 217static int runningbufreq; 218 219/* 220 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 221 * waitrunningbufspace(). 222 */ 223static struct mtx rbreqlock; 224 225/* 226 * Synchronization (sleep/wakeup) variable for buffer requests. 227 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 228 * by and/or. 229 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 230 * getnewbuf(), and getblk(). 231 */ 232static int needsbuffer; 233 234/* 235 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 236 */ 237static struct mtx nblock; 238 239/* 240 * Definitions for the buffer free lists. 241 */ 242#define BUFFER_QUEUES 6 /* number of free buffer queues */ 243 244#define QUEUE_NONE 0 /* on no queue */ 245#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 246#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 247#define QUEUE_DIRTY_GIANT 3 /* B_DELWRI buffers that need giant */ 248#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 249#define QUEUE_EMPTY 5 /* empty buffer headers */ 250 251/* Queues for free buffers with various properties */ 252static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 253 254/* Lock for the bufqueues */ 255static struct mtx bqlock; 256 257/* 258 * Single global constant for BUF_WMESG, to avoid getting multiple references. 259 * buf_wmesg is referred from macros. 260 */ 261const char *buf_wmesg = BUF_WMESG; 262 263#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 264#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 265#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 266#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 267 268#ifdef DIRECTIO 269extern void ffs_rawread_setup(void); 270#endif /* DIRECTIO */ 271/* 272 * numdirtywakeup: 273 * 274 * If someone is blocked due to there being too many dirty buffers, 275 * and numdirtybuffers is now reasonable, wake them up. 276 */ 277 278static __inline void 279numdirtywakeup(int level) 280{ 281 282 if (numdirtybuffers <= level) { 283 mtx_lock(&nblock); 284 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 285 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 286 wakeup(&needsbuffer); 287 } 288 mtx_unlock(&nblock); 289 } 290} 291 292/* 293 * bufspacewakeup: 294 * 295 * Called when buffer space is potentially available for recovery. 296 * getnewbuf() will block on this flag when it is unable to free 297 * sufficient buffer space. Buffer space becomes recoverable when 298 * bp's get placed back in the queues. 299 */ 300 301static __inline void 302bufspacewakeup(void) 303{ 304 305 /* 306 * If someone is waiting for BUF space, wake them up. Even 307 * though we haven't freed the kva space yet, the waiting 308 * process will be able to now. 309 */ 310 mtx_lock(&nblock); 311 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 312 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 313 wakeup(&needsbuffer); 314 } 315 mtx_unlock(&nblock); 316} 317 318/* 319 * runningbufwakeup() - in-progress I/O accounting. 320 * 321 */ 322void 323runningbufwakeup(struct buf *bp) 324{ 325 326 if (bp->b_runningbufspace) { 327 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 328 bp->b_runningbufspace = 0; 329 mtx_lock(&rbreqlock); 330 if (runningbufreq && runningbufspace <= lorunningspace) { 331 runningbufreq = 0; 332 wakeup(&runningbufreq); 333 } 334 mtx_unlock(&rbreqlock); 335 } 336} 337 338/* 339 * bufcountwakeup: 340 * 341 * Called when a buffer has been added to one of the free queues to 342 * account for the buffer and to wakeup anyone waiting for free buffers. 343 * This typically occurs when large amounts of metadata are being handled 344 * by the buffer cache ( else buffer space runs out first, usually ). 345 */ 346 347static __inline void 348bufcountwakeup(void) 349{ 350 351 atomic_add_int(&numfreebuffers, 1); 352 mtx_lock(&nblock); 353 if (needsbuffer) { 354 needsbuffer &= ~VFS_BIO_NEED_ANY; 355 if (numfreebuffers >= hifreebuffers) 356 needsbuffer &= ~VFS_BIO_NEED_FREE; 357 wakeup(&needsbuffer); 358 } 359 mtx_unlock(&nblock); 360} 361 362/* 363 * waitrunningbufspace() 364 * 365 * runningbufspace is a measure of the amount of I/O currently 366 * running. This routine is used in async-write situations to 367 * prevent creating huge backups of pending writes to a device. 368 * Only asynchronous writes are governed by this function. 369 * 370 * Reads will adjust runningbufspace, but will not block based on it. 371 * The read load has a side effect of reducing the allowed write load. 372 * 373 * This does NOT turn an async write into a sync write. It waits 374 * for earlier writes to complete and generally returns before the 375 * caller's write has reached the device. 376 */ 377void 378waitrunningbufspace(void) 379{ 380 381 mtx_lock(&rbreqlock); 382 while (runningbufspace > hirunningspace) { 383 ++runningbufreq; 384 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 385 } 386 mtx_unlock(&rbreqlock); 387} 388 389 390/* 391 * vfs_buf_test_cache: 392 * 393 * Called when a buffer is extended. This function clears the B_CACHE 394 * bit if the newly extended portion of the buffer does not contain 395 * valid data. 396 */ 397static __inline 398void 399vfs_buf_test_cache(struct buf *bp, 400 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 401 vm_page_t m) 402{ 403 404 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 405 if (bp->b_flags & B_CACHE) { 406 int base = (foff + off) & PAGE_MASK; 407 if (vm_page_is_valid(m, base, size) == 0) 408 bp->b_flags &= ~B_CACHE; 409 } 410} 411 412/* Wake up the buffer daemon if necessary */ 413static __inline 414void 415bd_wakeup(int dirtybuflevel) 416{ 417 418 mtx_lock(&bdlock); 419 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 420 bd_request = 1; 421 wakeup(&bd_request); 422 } 423 mtx_unlock(&bdlock); 424} 425 426/* 427 * bd_speedup - speedup the buffer cache flushing code 428 */ 429 430static __inline 431void 432bd_speedup(void) 433{ 434 435 bd_wakeup(1); 436} 437 438/* 439 * Calculating buffer cache scaling values and reserve space for buffer 440 * headers. This is called during low level kernel initialization and 441 * may be called more then once. We CANNOT write to the memory area 442 * being reserved at this time. 443 */ 444caddr_t 445kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 446{ 447 int maxbuf; 448 449 /* 450 * physmem_est is in pages. Convert it to kilobytes (assumes 451 * PAGE_SIZE is >= 1K) 452 */ 453 physmem_est = physmem_est * (PAGE_SIZE / 1024); 454 455 /* 456 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 457 * For the first 64MB of ram nominally allocate sufficient buffers to 458 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 459 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 460 * the buffer cache we limit the eventual kva reservation to 461 * maxbcache bytes. 462 * 463 * factor represents the 1/4 x ram conversion. 464 */ 465 if (nbuf == 0) { 466 int factor = 4 * BKVASIZE / 1024; 467 468 nbuf = 50; 469 if (physmem_est > 4096) 470 nbuf += min((physmem_est - 4096) / factor, 471 65536 / factor); 472 if (physmem_est > 65536) 473 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 474 475 if (maxbcache && nbuf > maxbcache / BKVASIZE) 476 nbuf = maxbcache / BKVASIZE; 477 478 /* XXX Avoid integer overflows later on with maxbufspace. */ 479 maxbuf = (INT_MAX / 3) / BKVASIZE; 480 if (nbuf > maxbuf) 481 nbuf = maxbuf; 482 } 483 484#if 0 485 /* 486 * Do not allow the buffer_map to be more then 1/2 the size of the 487 * kernel_map. 488 */ 489 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 490 (BKVASIZE * 2)) { 491 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 492 (BKVASIZE * 2); 493 printf("Warning: nbufs capped at %d\n", nbuf); 494 } 495#endif 496 497 /* 498 * swbufs are used as temporary holders for I/O, such as paging I/O. 499 * We have no less then 16 and no more then 256. 500 */ 501 nswbuf = max(min(nbuf/4, 256), 16); 502#ifdef NSWBUF_MIN 503 if (nswbuf < NSWBUF_MIN) 504 nswbuf = NSWBUF_MIN; 505#endif 506#ifdef DIRECTIO 507 ffs_rawread_setup(); 508#endif 509 510 /* 511 * Reserve space for the buffer cache buffers 512 */ 513 swbuf = (void *)v; 514 v = (caddr_t)(swbuf + nswbuf); 515 buf = (void *)v; 516 v = (caddr_t)(buf + nbuf); 517 518 return(v); 519} 520 521/* Initialize the buffer subsystem. Called before use of any buffers. */ 522void 523bufinit(void) 524{ 525 struct buf *bp; 526 int i; 527 528 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 529 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 530 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 531 mtx_init(&bdlock, "buffer daemon 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 bp->b_waiters = 0; 548 LIST_INIT(&bp->b_dep); 549 BUF_LOCKINIT(bp); 550 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 551 } 552 553 /* 554 * maxbufspace is the absolute maximum amount of buffer space we are 555 * allowed to reserve in KVM and in real terms. The absolute maximum 556 * is nominally used by buf_daemon. hibufspace is the nominal maximum 557 * used by most other processes. The differential is required to 558 * ensure that buf_daemon is able to run when other processes might 559 * be blocked waiting for buffer space. 560 * 561 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 562 * this may result in KVM fragmentation which is not handled optimally 563 * by the system. 564 */ 565 maxbufspace = nbuf * BKVASIZE; 566 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 567 lobufspace = hibufspace - MAXBSIZE; 568 569 lorunningspace = 512 * 1024; 570 hirunningspace = 1024 * 1024; 571 572/* 573 * Limit the amount of malloc memory since it is wired permanently into 574 * the kernel space. Even though this is accounted for in the buffer 575 * allocation, we don't want the malloced region to grow uncontrolled. 576 * The malloc scheme improves memory utilization significantly on average 577 * (small) directories. 578 */ 579 maxbufmallocspace = hibufspace / 20; 580 581/* 582 * Reduce the chance of a deadlock occuring by limiting the number 583 * of delayed-write dirty buffers we allow to stack up. 584 */ 585 hidirtybuffers = nbuf / 4 + 20; 586 dirtybufthresh = hidirtybuffers * 9 / 10; 587 numdirtybuffers = 0; 588/* 589 * To support extreme low-memory systems, make sure hidirtybuffers cannot 590 * eat up all available buffer space. This occurs when our minimum cannot 591 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 592 * BKVASIZE'd (8K) buffers. 593 */ 594 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 595 hidirtybuffers >>= 1; 596 } 597 lodirtybuffers = hidirtybuffers / 2; 598 599/* 600 * Try to keep the number of free buffers in the specified range, 601 * and give special processes (e.g. like buf_daemon) access to an 602 * emergency reserve. 603 */ 604 lofreebuffers = nbuf / 18 + 5; 605 hifreebuffers = 2 * lofreebuffers; 606 numfreebuffers = nbuf; 607 608/* 609 * Maximum number of async ops initiated per buf_daemon loop. This is 610 * somewhat of a hack at the moment, we really need to limit ourselves 611 * based on the number of bytes of I/O in-transit that were initiated 612 * from buf_daemon. 613 */ 614 615 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 616 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 617} 618 619/* 620 * bfreekva() - free the kva allocation for a buffer. 621 * 622 * Since this call frees up buffer space, we call bufspacewakeup(). 623 */ 624static void 625bfreekva(struct buf *bp) 626{ 627 628 if (bp->b_kvasize) { 629 atomic_add_int(&buffreekvacnt, 1); 630 atomic_subtract_int(&bufspace, bp->b_kvasize); 631 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase, 632 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 633 bp->b_kvasize = 0; 634 bufspacewakeup(); 635 } 636} 637 638/* 639 * bremfree: 640 * 641 * Mark the buffer for removal from the appropriate free list in brelse. 642 * 643 */ 644void 645bremfree(struct buf *bp) 646{ 647 648 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 649 KASSERT((bp->b_flags & B_REMFREE) == 0, 650 ("bremfree: buffer %p already marked for delayed removal.", bp)); 651 KASSERT(bp->b_qindex != QUEUE_NONE, 652 ("bremfree: buffer %p not on a queue.", bp)); 653 BUF_ASSERT_HELD(bp); 654 655 bp->b_flags |= B_REMFREE; 656 /* Fixup numfreebuffers count. */ 657 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 658 atomic_subtract_int(&numfreebuffers, 1); 659} 660 661/* 662 * bremfreef: 663 * 664 * Force an immediate removal from a free list. Used only in nfs when 665 * it abuses the b_freelist pointer. 666 */ 667void 668bremfreef(struct buf *bp) 669{ 670 mtx_lock(&bqlock); 671 bremfreel(bp); 672 mtx_unlock(&bqlock); 673} 674 675/* 676 * bremfreel: 677 * 678 * Removes a buffer from the free list, must be called with the 679 * bqlock held. 680 */ 681static void 682bremfreel(struct buf *bp) 683{ 684 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 685 bp, bp->b_vp, bp->b_flags); 686 KASSERT(bp->b_qindex != QUEUE_NONE, 687 ("bremfreel: buffer %p not on a queue.", bp)); 688 BUF_ASSERT_HELD(bp); 689 mtx_assert(&bqlock, MA_OWNED); 690 691 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 692 bp->b_qindex = QUEUE_NONE; 693 /* 694 * If this was a delayed bremfree() we only need to remove the buffer 695 * from the queue and return the stats are already done. 696 */ 697 if (bp->b_flags & B_REMFREE) { 698 bp->b_flags &= ~B_REMFREE; 699 return; 700 } 701 /* 702 * Fixup numfreebuffers count. If the buffer is invalid or not 703 * delayed-write, the buffer was free and we must decrement 704 * numfreebuffers. 705 */ 706 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 707 atomic_subtract_int(&numfreebuffers, 1); 708} 709 710 711/* 712 * Get a buffer with the specified data. Look in the cache first. We 713 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 714 * is set, the buffer is valid and we do not have to do anything ( see 715 * getblk() ). This is really just a special case of breadn(). 716 */ 717int 718bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 719 struct buf **bpp) 720{ 721 722 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 723} 724 725/* 726 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 727 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 728 * the buffer is valid and we do not have to do anything. 729 */ 730void 731breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 732 int cnt, struct ucred * cred) 733{ 734 struct buf *rabp; 735 int i; 736 737 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 738 if (inmem(vp, *rablkno)) 739 continue; 740 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 741 742 if ((rabp->b_flags & B_CACHE) == 0) { 743 if (!TD_IS_IDLETHREAD(curthread)) 744 curthread->td_ru.ru_inblock++; 745 rabp->b_flags |= B_ASYNC; 746 rabp->b_flags &= ~B_INVAL; 747 rabp->b_ioflags &= ~BIO_ERROR; 748 rabp->b_iocmd = BIO_READ; 749 if (rabp->b_rcred == NOCRED && cred != NOCRED) 750 rabp->b_rcred = crhold(cred); 751 vfs_busy_pages(rabp, 0); 752 BUF_KERNPROC(rabp); 753 rabp->b_iooffset = dbtob(rabp->b_blkno); 754 bstrategy(rabp); 755 } else { 756 brelse(rabp); 757 } 758 } 759} 760 761/* 762 * Operates like bread, but also starts asynchronous I/O on 763 * read-ahead blocks. 764 */ 765int 766breadn(struct vnode * vp, daddr_t blkno, int size, 767 daddr_t * rablkno, int *rabsize, 768 int cnt, struct ucred * cred, struct buf **bpp) 769{ 770 struct buf *bp; 771 int rv = 0, readwait = 0; 772 773 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 774 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 775 776 /* if not found in cache, do some I/O */ 777 if ((bp->b_flags & B_CACHE) == 0) { 778 if (!TD_IS_IDLETHREAD(curthread)) 779 curthread->td_ru.ru_inblock++; 780 bp->b_iocmd = BIO_READ; 781 bp->b_flags &= ~B_INVAL; 782 bp->b_ioflags &= ~BIO_ERROR; 783 if (bp->b_rcred == NOCRED && cred != NOCRED) 784 bp->b_rcred = crhold(cred); 785 vfs_busy_pages(bp, 0); 786 bp->b_iooffset = dbtob(bp->b_blkno); 787 bstrategy(bp); 788 ++readwait; 789 } 790 791 breada(vp, rablkno, rabsize, cnt, cred); 792 793 if (readwait) { 794 rv = bufwait(bp); 795 } 796 return (rv); 797} 798 799/* 800 * Write, release buffer on completion. (Done by iodone 801 * if async). Do not bother writing anything if the buffer 802 * is invalid. 803 * 804 * Note that we set B_CACHE here, indicating that buffer is 805 * fully valid and thus cacheable. This is true even of NFS 806 * now so we set it generally. This could be set either here 807 * or in biodone() since the I/O is synchronous. We put it 808 * here. 809 */ 810int 811bufwrite(struct buf *bp) 812{ 813 int oldflags; 814 struct vnode *vp; 815 int vp_md; 816 817 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 818 if (bp->b_flags & B_INVAL) { 819 brelse(bp); 820 return (0); 821 } 822 823 oldflags = bp->b_flags; 824 825 BUF_ASSERT_HELD(bp); 826 827 if (bp->b_pin_count > 0) 828 bunpin_wait(bp); 829 830 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 831 ("FFS background buffer should not get here %p", bp)); 832 833 vp = bp->b_vp; 834 if (vp) 835 vp_md = vp->v_vflag & VV_MD; 836 else 837 vp_md = 0; 838 839 /* Mark the buffer clean */ 840 bundirty(bp); 841 842 bp->b_flags &= ~B_DONE; 843 bp->b_ioflags &= ~BIO_ERROR; 844 bp->b_flags |= B_CACHE; 845 bp->b_iocmd = BIO_WRITE; 846 847 bufobj_wref(bp->b_bufobj); 848 vfs_busy_pages(bp, 1); 849 850 /* 851 * Normal bwrites pipeline writes 852 */ 853 bp->b_runningbufspace = bp->b_bufsize; 854 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 855 856 if (!TD_IS_IDLETHREAD(curthread)) 857 curthread->td_ru.ru_oublock++; 858 if (oldflags & B_ASYNC) 859 BUF_KERNPROC(bp); 860 bp->b_iooffset = dbtob(bp->b_blkno); 861 bstrategy(bp); 862 863 if ((oldflags & B_ASYNC) == 0) { 864 int rtval = bufwait(bp); 865 brelse(bp); 866 return (rtval); 867 } else { 868 /* 869 * don't allow the async write to saturate the I/O 870 * system. We will not deadlock here because 871 * we are blocking waiting for I/O that is already in-progress 872 * to complete. We do not block here if it is the update 873 * or syncer daemon trying to clean up as that can lead 874 * to deadlock. 875 */ 876 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 877 waitrunningbufspace(); 878 } 879 880 return (0); 881} 882 883void 884bufbdflush(struct bufobj *bo, struct buf *bp) 885{ 886 struct buf *nbp; 887 888 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 889 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 890 altbufferflushes++; 891 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 892 BO_LOCK(bo); 893 /* 894 * Try to find a buffer to flush. 895 */ 896 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 897 if ((nbp->b_vflags & BV_BKGRDINPROG) || 898 BUF_LOCK(nbp, 899 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 900 continue; 901 if (bp == nbp) 902 panic("bdwrite: found ourselves"); 903 BO_UNLOCK(bo); 904 /* Don't countdeps with the bo lock held. */ 905 if (buf_countdeps(nbp, 0)) { 906 BO_LOCK(bo); 907 BUF_UNLOCK(nbp); 908 continue; 909 } 910 if (nbp->b_flags & B_CLUSTEROK) { 911 vfs_bio_awrite(nbp); 912 } else { 913 bremfree(nbp); 914 bawrite(nbp); 915 } 916 dirtybufferflushes++; 917 break; 918 } 919 if (nbp == NULL) 920 BO_UNLOCK(bo); 921 } 922} 923 924/* 925 * Delayed write. (Buffer is marked dirty). Do not bother writing 926 * anything if the buffer is marked invalid. 927 * 928 * Note that since the buffer must be completely valid, we can safely 929 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 930 * biodone() in order to prevent getblk from writing the buffer 931 * out synchronously. 932 */ 933void 934bdwrite(struct buf *bp) 935{ 936 struct thread *td = curthread; 937 struct vnode *vp; 938 struct bufobj *bo; 939 940 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 941 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 942 BUF_ASSERT_HELD(bp); 943 944 if (bp->b_flags & B_INVAL) { 945 brelse(bp); 946 return; 947 } 948 949 /* 950 * If we have too many dirty buffers, don't create any more. 951 * If we are wildly over our limit, then force a complete 952 * cleanup. Otherwise, just keep the situation from getting 953 * out of control. Note that we have to avoid a recursive 954 * disaster and not try to clean up after our own cleanup! 955 */ 956 vp = bp->b_vp; 957 bo = bp->b_bufobj; 958 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 959 td->td_pflags |= TDP_INBDFLUSH; 960 BO_BDFLUSH(bo, bp); 961 td->td_pflags &= ~TDP_INBDFLUSH; 962 } else 963 recursiveflushes++; 964 965 bdirty(bp); 966 /* 967 * Set B_CACHE, indicating that the buffer is fully valid. This is 968 * true even of NFS now. 969 */ 970 bp->b_flags |= B_CACHE; 971 972 /* 973 * This bmap keeps the system from needing to do the bmap later, 974 * perhaps when the system is attempting to do a sync. Since it 975 * is likely that the indirect block -- or whatever other datastructure 976 * that the filesystem needs is still in memory now, it is a good 977 * thing to do this. Note also, that if the pageout daemon is 978 * requesting a sync -- there might not be enough memory to do 979 * the bmap then... So, this is important to do. 980 */ 981 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 982 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 983 } 984 985 /* 986 * Set the *dirty* buffer range based upon the VM system dirty pages. 987 */ 988 vfs_setdirty(bp); 989 990 /* 991 * We need to do this here to satisfy the vnode_pager and the 992 * pageout daemon, so that it thinks that the pages have been 993 * "cleaned". Note that since the pages are in a delayed write 994 * buffer -- the VFS layer "will" see that the pages get written 995 * out on the next sync, or perhaps the cluster will be completed. 996 */ 997 vfs_clean_pages(bp); 998 bqrelse(bp); 999 1000 /* 1001 * Wakeup the buffer flushing daemon if we have a lot of dirty 1002 * buffers (midpoint between our recovery point and our stall 1003 * point). 1004 */ 1005 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1006 1007 /* 1008 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1009 * due to the softdep code. 1010 */ 1011} 1012 1013/* 1014 * bdirty: 1015 * 1016 * Turn buffer into delayed write request. We must clear BIO_READ and 1017 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1018 * itself to properly update it in the dirty/clean lists. We mark it 1019 * B_DONE to ensure that any asynchronization of the buffer properly 1020 * clears B_DONE ( else a panic will occur later ). 1021 * 1022 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1023 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1024 * should only be called if the buffer is known-good. 1025 * 1026 * Since the buffer is not on a queue, we do not update the numfreebuffers 1027 * count. 1028 * 1029 * The buffer must be on QUEUE_NONE. 1030 */ 1031void 1032bdirty(struct buf *bp) 1033{ 1034 1035 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1036 bp, bp->b_vp, bp->b_flags); 1037 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1038 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1039 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1040 BUF_ASSERT_HELD(bp); 1041 bp->b_flags &= ~(B_RELBUF); 1042 bp->b_iocmd = BIO_WRITE; 1043 1044 if ((bp->b_flags & B_DELWRI) == 0) { 1045 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1046 reassignbuf(bp); 1047 atomic_add_int(&numdirtybuffers, 1); 1048 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1049 } 1050} 1051 1052/* 1053 * bundirty: 1054 * 1055 * Clear B_DELWRI for buffer. 1056 * 1057 * Since the buffer is not on a queue, we do not update the numfreebuffers 1058 * count. 1059 * 1060 * The buffer must be on QUEUE_NONE. 1061 */ 1062 1063void 1064bundirty(struct buf *bp) 1065{ 1066 1067 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1068 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1069 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1070 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1071 BUF_ASSERT_HELD(bp); 1072 1073 if (bp->b_flags & B_DELWRI) { 1074 bp->b_flags &= ~B_DELWRI; 1075 reassignbuf(bp); 1076 atomic_subtract_int(&numdirtybuffers, 1); 1077 numdirtywakeup(lodirtybuffers); 1078 } 1079 /* 1080 * Since it is now being written, we can clear its deferred write flag. 1081 */ 1082 bp->b_flags &= ~B_DEFERRED; 1083} 1084 1085/* 1086 * bawrite: 1087 * 1088 * Asynchronous write. Start output on a buffer, but do not wait for 1089 * it to complete. The buffer is released when the output completes. 1090 * 1091 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1092 * B_INVAL buffers. Not us. 1093 */ 1094void 1095bawrite(struct buf *bp) 1096{ 1097 1098 bp->b_flags |= B_ASYNC; 1099 (void) bwrite(bp); 1100} 1101 1102/* 1103 * bwillwrite: 1104 * 1105 * Called prior to the locking of any vnodes when we are expecting to 1106 * write. We do not want to starve the buffer cache with too many 1107 * dirty buffers so we block here. By blocking prior to the locking 1108 * of any vnodes we attempt to avoid the situation where a locked vnode 1109 * prevents the various system daemons from flushing related buffers. 1110 */ 1111 1112void 1113bwillwrite(void) 1114{ 1115 1116 if (numdirtybuffers >= hidirtybuffers) { 1117 mtx_lock(&nblock); 1118 while (numdirtybuffers >= hidirtybuffers) { 1119 bd_wakeup(1); 1120 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1121 msleep(&needsbuffer, &nblock, 1122 (PRIBIO + 4), "flswai", 0); 1123 } 1124 mtx_unlock(&nblock); 1125 } 1126} 1127 1128/* 1129 * Return true if we have too many dirty buffers. 1130 */ 1131int 1132buf_dirty_count_severe(void) 1133{ 1134 1135 return(numdirtybuffers >= hidirtybuffers); 1136} 1137 1138/* 1139 * brelse: 1140 * 1141 * Release a busy buffer and, if requested, free its resources. The 1142 * buffer will be stashed in the appropriate bufqueue[] allowing it 1143 * to be accessed later as a cache entity or reused for other purposes. 1144 */ 1145void 1146brelse(struct buf *bp) 1147{ 1148 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1149 bp, bp->b_vp, bp->b_flags); 1150 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1151 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1152 1153 if (bp->b_flags & B_MANAGED) { 1154 bqrelse(bp); 1155 return; 1156 } 1157 1158 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1159 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1160 /* 1161 * Failed write, redirty. Must clear BIO_ERROR to prevent 1162 * pages from being scrapped. If the error is anything 1163 * other than an I/O error (EIO), assume that retryingi 1164 * is futile. 1165 */ 1166 bp->b_ioflags &= ~BIO_ERROR; 1167 bdirty(bp); 1168 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1169 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1170 /* 1171 * Either a failed I/O or we were asked to free or not 1172 * cache the buffer. 1173 */ 1174 bp->b_flags |= B_INVAL; 1175 if (!LIST_EMPTY(&bp->b_dep)) 1176 buf_deallocate(bp); 1177 if (bp->b_flags & B_DELWRI) { 1178 atomic_subtract_int(&numdirtybuffers, 1); 1179 numdirtywakeup(lodirtybuffers); 1180 } 1181 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1182 if ((bp->b_flags & B_VMIO) == 0) { 1183 if (bp->b_bufsize) 1184 allocbuf(bp, 0); 1185 if (bp->b_vp) 1186 (void) brelvp(bp); 1187 } 1188 } 1189 1190 /* 1191 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1192 * is called with B_DELWRI set, the underlying pages may wind up 1193 * getting freed causing a previous write (bdwrite()) to get 'lost' 1194 * because pages associated with a B_DELWRI bp are marked clean. 1195 * 1196 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1197 * if B_DELWRI is set. 1198 * 1199 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1200 * on pages to return pages to the VM page queues. 1201 */ 1202 if (bp->b_flags & B_DELWRI) 1203 bp->b_flags &= ~B_RELBUF; 1204 else if (vm_page_count_severe()) { 1205 /* 1206 * The locking of the BO_LOCK is not necessary since 1207 * BKGRDINPROG cannot be set while we hold the buf 1208 * lock, it can only be cleared if it is already 1209 * pending. 1210 */ 1211 if (bp->b_vp) { 1212 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1213 bp->b_flags |= B_RELBUF; 1214 } else 1215 bp->b_flags |= B_RELBUF; 1216 } 1217 1218 /* 1219 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1220 * constituted, not even NFS buffers now. Two flags effect this. If 1221 * B_INVAL, the struct buf is invalidated but the VM object is kept 1222 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1223 * 1224 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1225 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1226 * buffer is also B_INVAL because it hits the re-dirtying code above. 1227 * 1228 * Normally we can do this whether a buffer is B_DELWRI or not. If 1229 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1230 * the commit state and we cannot afford to lose the buffer. If the 1231 * buffer has a background write in progress, we need to keep it 1232 * around to prevent it from being reconstituted and starting a second 1233 * background write. 1234 */ 1235 if ((bp->b_flags & B_VMIO) 1236 && !(bp->b_vp->v_mount != NULL && 1237 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1238 !vn_isdisk(bp->b_vp, NULL) && 1239 (bp->b_flags & B_DELWRI)) 1240 ) { 1241 1242 int i, j, resid; 1243 vm_page_t m; 1244 off_t foff; 1245 vm_pindex_t poff; 1246 vm_object_t obj; 1247 1248 obj = bp->b_bufobj->bo_object; 1249 1250 /* 1251 * Get the base offset and length of the buffer. Note that 1252 * in the VMIO case if the buffer block size is not 1253 * page-aligned then b_data pointer may not be page-aligned. 1254 * But our b_pages[] array *IS* page aligned. 1255 * 1256 * block sizes less then DEV_BSIZE (usually 512) are not 1257 * supported due to the page granularity bits (m->valid, 1258 * m->dirty, etc...). 1259 * 1260 * See man buf(9) for more information 1261 */ 1262 resid = bp->b_bufsize; 1263 foff = bp->b_offset; 1264 VM_OBJECT_LOCK(obj); 1265 for (i = 0; i < bp->b_npages; i++) { 1266 int had_bogus = 0; 1267 1268 m = bp->b_pages[i]; 1269 1270 /* 1271 * If we hit a bogus page, fixup *all* the bogus pages 1272 * now. 1273 */ 1274 if (m == bogus_page) { 1275 poff = OFF_TO_IDX(bp->b_offset); 1276 had_bogus = 1; 1277 1278 for (j = i; j < bp->b_npages; j++) { 1279 vm_page_t mtmp; 1280 mtmp = bp->b_pages[j]; 1281 if (mtmp == bogus_page) { 1282 mtmp = vm_page_lookup(obj, poff + j); 1283 if (!mtmp) { 1284 panic("brelse: page missing\n"); 1285 } 1286 bp->b_pages[j] = mtmp; 1287 } 1288 } 1289 1290 if ((bp->b_flags & B_INVAL) == 0) { 1291 pmap_qenter( 1292 trunc_page((vm_offset_t)bp->b_data), 1293 bp->b_pages, bp->b_npages); 1294 } 1295 m = bp->b_pages[i]; 1296 } 1297 if ((bp->b_flags & B_NOCACHE) || 1298 (bp->b_ioflags & BIO_ERROR)) { 1299 int poffset = foff & PAGE_MASK; 1300 int presid = resid > (PAGE_SIZE - poffset) ? 1301 (PAGE_SIZE - poffset) : resid; 1302 1303 KASSERT(presid >= 0, ("brelse: extra page")); 1304 vm_page_lock_queues(); 1305 vm_page_set_invalid(m, poffset, presid); 1306 vm_page_unlock_queues(); 1307 if (had_bogus) 1308 printf("avoided corruption bug in bogus_page/brelse code\n"); 1309 } 1310 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1311 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1312 } 1313 VM_OBJECT_UNLOCK(obj); 1314 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1315 vfs_vmio_release(bp); 1316 1317 } else if (bp->b_flags & B_VMIO) { 1318 1319 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1320 vfs_vmio_release(bp); 1321 } 1322 1323 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1324 if (bp->b_bufsize != 0) 1325 allocbuf(bp, 0); 1326 if (bp->b_vp != NULL) 1327 (void) brelvp(bp); 1328 } 1329 1330 if (BUF_LOCKRECURSED(bp)) { 1331 /* do not release to free list */ 1332 BUF_UNLOCK(bp); 1333 return; 1334 } 1335 1336 /* enqueue */ 1337 mtx_lock(&bqlock); 1338 /* Handle delayed bremfree() processing. */ 1339 if (bp->b_flags & B_REMFREE) 1340 bremfreel(bp); 1341 if (bp->b_qindex != QUEUE_NONE) 1342 panic("brelse: free buffer onto another queue???"); 1343 1344 /* buffers with no memory */ 1345 if (bp->b_bufsize == 0) { 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 1"); 1350 if (bp->b_kvasize) { 1351 bp->b_qindex = QUEUE_EMPTYKVA; 1352 } else { 1353 bp->b_qindex = QUEUE_EMPTY; 1354 } 1355 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1356 /* buffers with junk contents */ 1357 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1358 (bp->b_ioflags & BIO_ERROR)) { 1359 bp->b_flags |= B_INVAL; 1360 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1361 if (bp->b_vflags & BV_BKGRDINPROG) 1362 panic("losing buffer 2"); 1363 bp->b_qindex = QUEUE_CLEAN; 1364 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1365 /* remaining buffers */ 1366 } else { 1367 if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) == 1368 (B_DELWRI|B_NEEDSGIANT)) 1369 bp->b_qindex = QUEUE_DIRTY_GIANT; 1370 if (bp->b_flags & B_DELWRI) 1371 bp->b_qindex = QUEUE_DIRTY; 1372 else 1373 bp->b_qindex = QUEUE_CLEAN; 1374 if (bp->b_flags & B_AGE) 1375 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1376 else 1377 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1378 } 1379 mtx_unlock(&bqlock); 1380 1381 /* 1382 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1383 * placed the buffer on the correct queue. We must also disassociate 1384 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1385 * find it. 1386 */ 1387 if (bp->b_flags & B_INVAL) { 1388 if (bp->b_flags & B_DELWRI) 1389 bundirty(bp); 1390 if (bp->b_vp) 1391 (void) brelvp(bp); 1392 } 1393 1394 /* 1395 * Fixup numfreebuffers count. The bp is on an appropriate queue 1396 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1397 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1398 * if B_INVAL is set ). 1399 */ 1400 1401 if (!(bp->b_flags & B_DELWRI)) 1402 bufcountwakeup(); 1403 1404 /* 1405 * Something we can maybe free or reuse 1406 */ 1407 if (bp->b_bufsize || bp->b_kvasize) 1408 bufspacewakeup(); 1409 1410 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1411 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1412 panic("brelse: not dirty"); 1413 /* unlock */ 1414 BUF_UNLOCK(bp); 1415} 1416 1417/* 1418 * Release a buffer back to the appropriate queue but do not try to free 1419 * it. The buffer is expected to be used again soon. 1420 * 1421 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1422 * biodone() to requeue an async I/O on completion. It is also used when 1423 * known good buffers need to be requeued but we think we may need the data 1424 * again soon. 1425 * 1426 * XXX we should be able to leave the B_RELBUF hint set on completion. 1427 */ 1428void 1429bqrelse(struct buf *bp) 1430{ 1431 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1432 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1433 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1434 1435 if (BUF_LOCKRECURSED(bp)) { 1436 /* do not release to free list */ 1437 BUF_UNLOCK(bp); 1438 return; 1439 } 1440 1441 if (bp->b_flags & B_MANAGED) { 1442 if (bp->b_flags & B_REMFREE) { 1443 mtx_lock(&bqlock); 1444 bremfreel(bp); 1445 mtx_unlock(&bqlock); 1446 } 1447 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1448 BUF_UNLOCK(bp); 1449 return; 1450 } 1451 1452 mtx_lock(&bqlock); 1453 /* Handle delayed bremfree() processing. */ 1454 if (bp->b_flags & B_REMFREE) 1455 bremfreel(bp); 1456 if (bp->b_qindex != QUEUE_NONE) 1457 panic("bqrelse: free buffer onto another queue???"); 1458 /* buffers with stale but valid contents */ 1459 if (bp->b_flags & B_DELWRI) { 1460 if (bp->b_flags & B_NEEDSGIANT) 1461 bp->b_qindex = QUEUE_DIRTY_GIANT; 1462 else 1463 bp->b_qindex = QUEUE_DIRTY; 1464 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1465 } else { 1466 /* 1467 * The locking of the BO_LOCK for checking of the 1468 * BV_BKGRDINPROG is not necessary since the 1469 * BV_BKGRDINPROG cannot be set while we hold the buf 1470 * lock, it can only be cleared if it is already 1471 * pending. 1472 */ 1473 if (!vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) { 1474 bp->b_qindex = QUEUE_CLEAN; 1475 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1476 b_freelist); 1477 } else { 1478 /* 1479 * We are too low on memory, we have to try to free 1480 * the buffer (most importantly: the wired pages 1481 * making up its backing store) *now*. 1482 */ 1483 mtx_unlock(&bqlock); 1484 brelse(bp); 1485 return; 1486 } 1487 } 1488 mtx_unlock(&bqlock); 1489 1490 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1491 bufcountwakeup(); 1492 1493 /* 1494 * Something we can maybe free or reuse. 1495 */ 1496 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1497 bufspacewakeup(); 1498 1499 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1500 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1501 panic("bqrelse: not dirty"); 1502 /* unlock */ 1503 BUF_UNLOCK(bp); 1504} 1505 1506/* Give pages used by the bp back to the VM system (where possible) */ 1507static void 1508vfs_vmio_release(struct buf *bp) 1509{ 1510 int i; 1511 vm_page_t m; 1512 1513 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1514 vm_page_lock_queues(); 1515 for (i = 0; i < bp->b_npages; i++) { 1516 m = bp->b_pages[i]; 1517 bp->b_pages[i] = NULL; 1518 /* 1519 * In order to keep page LRU ordering consistent, put 1520 * everything on the inactive queue. 1521 */ 1522 vm_page_unwire(m, 0); 1523 /* 1524 * We don't mess with busy pages, it is 1525 * the responsibility of the process that 1526 * busied the pages to deal with them. 1527 */ 1528 if ((m->oflags & VPO_BUSY) || (m->busy != 0)) 1529 continue; 1530 1531 if (m->wire_count == 0) { 1532 /* 1533 * Might as well free the page if we can and it has 1534 * no valid data. We also free the page if the 1535 * buffer was used for direct I/O 1536 */ 1537 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1538 m->hold_count == 0) { 1539 vm_page_free(m); 1540 } else if (bp->b_flags & B_DIRECT) { 1541 vm_page_try_to_free(m); 1542 } else if (vm_page_count_severe()) { 1543 vm_page_try_to_cache(m); 1544 } 1545 } 1546 } 1547 vm_page_unlock_queues(); 1548 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1549 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1550 1551 if (bp->b_bufsize) { 1552 bufspacewakeup(); 1553 bp->b_bufsize = 0; 1554 } 1555 bp->b_npages = 0; 1556 bp->b_flags &= ~B_VMIO; 1557 if (bp->b_vp) 1558 (void) brelvp(bp); 1559} 1560 1561/* 1562 * Check to see if a block at a particular lbn is available for a clustered 1563 * write. 1564 */ 1565static int 1566vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1567{ 1568 struct buf *bpa; 1569 int match; 1570 1571 match = 0; 1572 1573 /* If the buf isn't in core skip it */ 1574 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1575 return (0); 1576 1577 /* If the buf is busy we don't want to wait for it */ 1578 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1579 return (0); 1580 1581 /* Only cluster with valid clusterable delayed write buffers */ 1582 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1583 (B_DELWRI | B_CLUSTEROK)) 1584 goto done; 1585 1586 if (bpa->b_bufsize != size) 1587 goto done; 1588 1589 /* 1590 * Check to see if it is in the expected place on disk and that the 1591 * block has been mapped. 1592 */ 1593 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1594 match = 1; 1595done: 1596 BUF_UNLOCK(bpa); 1597 return (match); 1598} 1599 1600/* 1601 * vfs_bio_awrite: 1602 * 1603 * Implement clustered async writes for clearing out B_DELWRI buffers. 1604 * This is much better then the old way of writing only one buffer at 1605 * a time. Note that we may not be presented with the buffers in the 1606 * correct order, so we search for the cluster in both directions. 1607 */ 1608int 1609vfs_bio_awrite(struct buf *bp) 1610{ 1611 int i; 1612 int j; 1613 daddr_t lblkno = bp->b_lblkno; 1614 struct vnode *vp = bp->b_vp; 1615 int ncl; 1616 int nwritten; 1617 int size; 1618 int maxcl; 1619 1620 /* 1621 * right now we support clustered writing only to regular files. If 1622 * we find a clusterable block we could be in the middle of a cluster 1623 * rather then at the beginning. 1624 */ 1625 if ((vp->v_type == VREG) && 1626 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1627 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1628 1629 size = vp->v_mount->mnt_stat.f_iosize; 1630 maxcl = MAXPHYS / size; 1631 1632 VI_LOCK(vp); 1633 for (i = 1; i < maxcl; i++) 1634 if (vfs_bio_clcheck(vp, size, lblkno + i, 1635 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1636 break; 1637 1638 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1639 if (vfs_bio_clcheck(vp, size, lblkno - j, 1640 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1641 break; 1642 1643 VI_UNLOCK(vp); 1644 --j; 1645 ncl = i + j; 1646 /* 1647 * this is a possible cluster write 1648 */ 1649 if (ncl != 1) { 1650 BUF_UNLOCK(bp); 1651 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1652 return nwritten; 1653 } 1654 } 1655 bremfree(bp); 1656 bp->b_flags |= B_ASYNC; 1657 /* 1658 * default (old) behavior, writing out only one block 1659 * 1660 * XXX returns b_bufsize instead of b_bcount for nwritten? 1661 */ 1662 nwritten = bp->b_bufsize; 1663 (void) bwrite(bp); 1664 1665 return nwritten; 1666} 1667 1668/* 1669 * getnewbuf: 1670 * 1671 * Find and initialize a new buffer header, freeing up existing buffers 1672 * in the bufqueues as necessary. The new buffer is returned locked. 1673 * 1674 * Important: B_INVAL is not set. If the caller wishes to throw the 1675 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1676 * 1677 * We block if: 1678 * We have insufficient buffer headers 1679 * We have insufficient buffer space 1680 * buffer_map is too fragmented ( space reservation fails ) 1681 * If we have to flush dirty buffers ( but we try to avoid this ) 1682 * 1683 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1684 * Instead we ask the buf daemon to do it for us. We attempt to 1685 * avoid piecemeal wakeups of the pageout daemon. 1686 */ 1687 1688static struct buf * 1689getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1690{ 1691 struct buf *bp; 1692 struct buf *nbp; 1693 int defrag = 0; 1694 int nqindex; 1695 int waiters = 0; 1696 static int flushingbufs; 1697 1698 /* 1699 * We can't afford to block since we might be holding a vnode lock, 1700 * which may prevent system daemons from running. We deal with 1701 * low-memory situations by proactively returning memory and running 1702 * async I/O rather then sync I/O. 1703 */ 1704 1705 atomic_add_int(&getnewbufcalls, 1); 1706 atomic_subtract_int(&getnewbufrestarts, 1); 1707restart: 1708 atomic_add_int(&getnewbufrestarts, 1); 1709 1710 /* 1711 * Setup for scan. If we do not have enough free buffers, 1712 * we setup a degenerate case that immediately fails. Note 1713 * that if we are specially marked process, we are allowed to 1714 * dip into our reserves. 1715 * 1716 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1717 * 1718 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1719 * However, there are a number of cases (defragging, reusing, ...) 1720 * where we cannot backup. 1721 */ 1722 mtx_lock(&bqlock); 1723 nqindex = QUEUE_EMPTYKVA; 1724 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1725 1726 if (nbp == NULL) { 1727 /* 1728 * If no EMPTYKVA buffers and we are either 1729 * defragging or reusing, locate a CLEAN buffer 1730 * to free or reuse. If bufspace useage is low 1731 * skip this step so we can allocate a new buffer. 1732 */ 1733 if (defrag || bufspace >= lobufspace) { 1734 nqindex = QUEUE_CLEAN; 1735 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1736 } 1737 1738 /* 1739 * If we could not find or were not allowed to reuse a 1740 * CLEAN buffer, check to see if it is ok to use an EMPTY 1741 * buffer. We can only use an EMPTY buffer if allocating 1742 * its KVA would not otherwise run us out of buffer space. 1743 */ 1744 if (nbp == NULL && defrag == 0 && 1745 bufspace + maxsize < hibufspace) { 1746 nqindex = QUEUE_EMPTY; 1747 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1748 } 1749 } 1750 1751 /* 1752 * Run scan, possibly freeing data and/or kva mappings on the fly 1753 * depending. 1754 */ 1755 1756 while ((bp = nbp) != NULL) { 1757 int qindex = nqindex; 1758 1759 /* 1760 * Calculate next bp ( we can only use it if we do not block 1761 * or do other fancy things ). 1762 */ 1763 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1764 switch(qindex) { 1765 case QUEUE_EMPTY: 1766 nqindex = QUEUE_EMPTYKVA; 1767 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1768 break; 1769 /* FALLTHROUGH */ 1770 case QUEUE_EMPTYKVA: 1771 nqindex = QUEUE_CLEAN; 1772 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1773 break; 1774 /* FALLTHROUGH */ 1775 case QUEUE_CLEAN: 1776 /* 1777 * nbp is NULL. 1778 */ 1779 break; 1780 } 1781 } 1782 /* 1783 * If we are defragging then we need a buffer with 1784 * b_kvasize != 0. XXX this situation should no longer 1785 * occur, if defrag is non-zero the buffer's b_kvasize 1786 * should also be non-zero at this point. XXX 1787 */ 1788 if (defrag && bp->b_kvasize == 0) { 1789 printf("Warning: defrag empty buffer %p\n", bp); 1790 continue; 1791 } 1792 1793 /* 1794 * Start freeing the bp. This is somewhat involved. nbp 1795 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1796 */ 1797 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1798 continue; 1799 if (bp->b_vp) { 1800 BO_LOCK(bp->b_bufobj); 1801 if (bp->b_vflags & BV_BKGRDINPROG) { 1802 BO_UNLOCK(bp->b_bufobj); 1803 BUF_UNLOCK(bp); 1804 continue; 1805 } 1806 BO_UNLOCK(bp->b_bufobj); 1807 } 1808 CTR6(KTR_BUF, 1809 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1810 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1811 bp->b_kvasize, bp->b_bufsize, qindex); 1812 1813 /* 1814 * Sanity Checks 1815 */ 1816 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1817 1818 /* 1819 * Note: we no longer distinguish between VMIO and non-VMIO 1820 * buffers. 1821 */ 1822 1823 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1824 1825 bremfreel(bp); 1826 mtx_unlock(&bqlock); 1827 1828 if (qindex == QUEUE_CLEAN) { 1829 if (bp->b_flags & B_VMIO) { 1830 bp->b_flags &= ~B_ASYNC; 1831 vfs_vmio_release(bp); 1832 } 1833 if (bp->b_vp) 1834 waiters = brelvp(bp); 1835 } 1836 1837 /* 1838 * NOTE: nbp is now entirely invalid. We can only restart 1839 * the scan from this point on. 1840 * 1841 * Get the rest of the buffer freed up. b_kva* is still 1842 * valid after this operation. 1843 */ 1844 1845 if (bp->b_rcred != NOCRED) { 1846 crfree(bp->b_rcred); 1847 bp->b_rcred = NOCRED; 1848 } 1849 if (bp->b_wcred != NOCRED) { 1850 crfree(bp->b_wcred); 1851 bp->b_wcred = NOCRED; 1852 } 1853 if (!LIST_EMPTY(&bp->b_dep)) 1854 buf_deallocate(bp); 1855 if (bp->b_vflags & BV_BKGRDINPROG) 1856 panic("losing buffer 3"); 1857 KASSERT(bp->b_vp == NULL, 1858 ("bp: %p still has vnode %p. qindex: %d", 1859 bp, bp->b_vp, qindex)); 1860 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1861 ("bp: %p still on a buffer list. xflags %X", 1862 bp, bp->b_xflags)); 1863 1864 if (bp->b_bufsize) 1865 allocbuf(bp, 0); 1866 1867 bp->b_flags = 0; 1868 bp->b_ioflags = 0; 1869 bp->b_xflags = 0; 1870 bp->b_vflags = 0; 1871 bp->b_vp = NULL; 1872 bp->b_blkno = bp->b_lblkno = 0; 1873 bp->b_offset = NOOFFSET; 1874 bp->b_iodone = 0; 1875 bp->b_error = 0; 1876 bp->b_resid = 0; 1877 bp->b_bcount = 0; 1878 bp->b_npages = 0; 1879 bp->b_dirtyoff = bp->b_dirtyend = 0; 1880 bp->b_bufobj = NULL; 1881 bp->b_pin_count = 0; 1882 bp->b_fsprivate1 = NULL; 1883 bp->b_fsprivate2 = NULL; 1884 bp->b_fsprivate3 = NULL; 1885 1886 LIST_INIT(&bp->b_dep); 1887 1888 /* 1889 * If we are defragging then free the buffer. 1890 */ 1891 if (defrag) { 1892 bp->b_flags |= B_INVAL; 1893 bfreekva(bp); 1894 brelse(bp); 1895 defrag = 0; 1896 goto restart; 1897 } 1898 1899 /* 1900 * Notify any waiters for the buffer lock about 1901 * identity change by freeing the buffer. 1902 */ 1903 if (qindex == QUEUE_CLEAN && waiters > 0) { 1904 bp->b_flags |= B_INVAL; 1905 bfreekva(bp); 1906 brelse(bp); 1907 goto restart; 1908 } 1909 1910 /* 1911 * If we are overcomitted then recover the buffer and its 1912 * KVM space. This occurs in rare situations when multiple 1913 * processes are blocked in getnewbuf() or allocbuf(). 1914 */ 1915 if (bufspace >= hibufspace) 1916 flushingbufs = 1; 1917 if (flushingbufs && bp->b_kvasize != 0) { 1918 bp->b_flags |= B_INVAL; 1919 bfreekva(bp); 1920 brelse(bp); 1921 goto restart; 1922 } 1923 if (bufspace < lobufspace) 1924 flushingbufs = 0; 1925 break; 1926 } 1927 1928 /* 1929 * If we exhausted our list, sleep as appropriate. We may have to 1930 * wakeup various daemons and write out some dirty buffers. 1931 * 1932 * Generally we are sleeping due to insufficient buffer space. 1933 */ 1934 1935 if (bp == NULL) { 1936 int flags; 1937 char *waitmsg; 1938 1939 if (defrag) { 1940 flags = VFS_BIO_NEED_BUFSPACE; 1941 waitmsg = "nbufkv"; 1942 } else if (bufspace >= hibufspace) { 1943 waitmsg = "nbufbs"; 1944 flags = VFS_BIO_NEED_BUFSPACE; 1945 } else { 1946 waitmsg = "newbuf"; 1947 flags = VFS_BIO_NEED_ANY; 1948 } 1949 mtx_lock(&nblock); 1950 needsbuffer |= flags; 1951 mtx_unlock(&nblock); 1952 mtx_unlock(&bqlock); 1953 1954 bd_speedup(); /* heeeelp */ 1955 1956 mtx_lock(&nblock); 1957 while (needsbuffer & flags) { 1958 if (msleep(&needsbuffer, &nblock, 1959 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1960 mtx_unlock(&nblock); 1961 return (NULL); 1962 } 1963 } 1964 mtx_unlock(&nblock); 1965 } else { 1966 /* 1967 * We finally have a valid bp. We aren't quite out of the 1968 * woods, we still have to reserve kva space. In order 1969 * to keep fragmentation sane we only allocate kva in 1970 * BKVASIZE chunks. 1971 */ 1972 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1973 1974 if (maxsize != bp->b_kvasize) { 1975 vm_offset_t addr = 0; 1976 1977 bfreekva(bp); 1978 1979 vm_map_lock(buffer_map); 1980 if (vm_map_findspace(buffer_map, 1981 vm_map_min(buffer_map), maxsize, &addr)) { 1982 /* 1983 * Uh oh. Buffer map is to fragmented. We 1984 * must defragment the map. 1985 */ 1986 atomic_add_int(&bufdefragcnt, 1); 1987 vm_map_unlock(buffer_map); 1988 defrag = 1; 1989 bp->b_flags |= B_INVAL; 1990 brelse(bp); 1991 goto restart; 1992 } 1993 if (addr) { 1994 vm_map_insert(buffer_map, NULL, 0, 1995 addr, addr + maxsize, 1996 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1997 1998 bp->b_kvabase = (caddr_t) addr; 1999 bp->b_kvasize = maxsize; 2000 atomic_add_int(&bufspace, bp->b_kvasize); 2001 atomic_add_int(&bufreusecnt, 1); 2002 } 2003 vm_map_unlock(buffer_map); 2004 } 2005 bp->b_saveaddr = bp->b_kvabase; 2006 bp->b_data = bp->b_saveaddr; 2007 } 2008 return(bp); 2009} 2010 2011/* 2012 * buf_daemon: 2013 * 2014 * buffer flushing daemon. Buffers are normally flushed by the 2015 * update daemon but if it cannot keep up this process starts to 2016 * take the load in an attempt to prevent getnewbuf() from blocking. 2017 */ 2018 2019static struct kproc_desc buf_kp = { 2020 "bufdaemon", 2021 buf_daemon, 2022 &bufdaemonproc 2023}; 2024SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2025 2026static void 2027buf_daemon() 2028{ 2029 2030 /* 2031 * This process needs to be suspended prior to shutdown sync. 2032 */ 2033 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2034 SHUTDOWN_PRI_LAST); 2035 2036 /* 2037 * This process is allowed to take the buffer cache to the limit 2038 */ 2039 curthread->td_pflags |= TDP_NORUNNINGBUF; 2040 mtx_lock(&bdlock); 2041 for (;;) { 2042 bd_request = 0; 2043 mtx_unlock(&bdlock); 2044 2045 kproc_suspend_check(bufdaemonproc); 2046 2047 /* 2048 * Do the flush. Limit the amount of in-transit I/O we 2049 * allow to build up, otherwise we would completely saturate 2050 * the I/O system. Wakeup any waiting processes before we 2051 * normally would so they can run in parallel with our drain. 2052 */ 2053 while (numdirtybuffers > lodirtybuffers) { 2054 int flushed; 2055 2056 flushed = flushbufqueues(QUEUE_DIRTY, 0); 2057 /* The list empty check here is slightly racy */ 2058 if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) { 2059 mtx_lock(&Giant); 2060 flushed += flushbufqueues(QUEUE_DIRTY_GIANT, 0); 2061 mtx_unlock(&Giant); 2062 } 2063 if (flushed == 0) { 2064 /* 2065 * Could not find any buffers without rollback 2066 * dependencies, so just write the first one 2067 * in the hopes of eventually making progress. 2068 */ 2069 flushbufqueues(QUEUE_DIRTY, 1); 2070 if (!TAILQ_EMPTY( 2071 &bufqueues[QUEUE_DIRTY_GIANT])) { 2072 mtx_lock(&Giant); 2073 flushbufqueues(QUEUE_DIRTY_GIANT, 1); 2074 mtx_unlock(&Giant); 2075 } 2076 break; 2077 } 2078 uio_yield(); 2079 } 2080 2081 /* 2082 * Only clear bd_request if we have reached our low water 2083 * mark. The buf_daemon normally waits 1 second and 2084 * then incrementally flushes any dirty buffers that have 2085 * built up, within reason. 2086 * 2087 * If we were unable to hit our low water mark and couldn't 2088 * find any flushable buffers, we sleep half a second. 2089 * Otherwise we loop immediately. 2090 */ 2091 mtx_lock(&bdlock); 2092 if (numdirtybuffers <= lodirtybuffers) { 2093 /* 2094 * We reached our low water mark, reset the 2095 * request and sleep until we are needed again. 2096 * The sleep is just so the suspend code works. 2097 */ 2098 bd_request = 0; 2099 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2100 } else { 2101 /* 2102 * We couldn't find any flushable dirty buffers but 2103 * still have too many dirty buffers, we 2104 * have to sleep and try again. (rare) 2105 */ 2106 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2107 } 2108 } 2109} 2110 2111/* 2112 * flushbufqueues: 2113 * 2114 * Try to flush a buffer in the dirty queue. We must be careful to 2115 * free up B_INVAL buffers instead of write them, which NFS is 2116 * particularly sensitive to. 2117 */ 2118static int flushwithdeps = 0; 2119SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2120 0, "Number of buffers flushed with dependecies that require rollbacks"); 2121 2122static int 2123flushbufqueues(int queue, int flushdeps) 2124{ 2125 struct buf sentinel; 2126 struct vnode *vp; 2127 struct mount *mp; 2128 struct buf *bp; 2129 int hasdeps; 2130 int flushed; 2131 int target; 2132 2133 target = numdirtybuffers - lodirtybuffers; 2134 if (flushdeps && target > 2) 2135 target /= 2; 2136 flushed = 0; 2137 bp = NULL; 2138 mtx_lock(&bqlock); 2139 TAILQ_INSERT_TAIL(&bufqueues[queue], &sentinel, b_freelist); 2140 while (flushed != target) { 2141 bp = TAILQ_FIRST(&bufqueues[queue]); 2142 if (bp == &sentinel) 2143 break; 2144 TAILQ_REMOVE(&bufqueues[queue], bp, b_freelist); 2145 TAILQ_INSERT_TAIL(&bufqueues[queue], bp, b_freelist); 2146 2147 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2148 continue; 2149 if (bp->b_pin_count > 0) { 2150 BUF_UNLOCK(bp); 2151 continue; 2152 } 2153 BO_LOCK(bp->b_bufobj); 2154 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2155 (bp->b_flags & B_DELWRI) == 0) { 2156 BO_UNLOCK(bp->b_bufobj); 2157 BUF_UNLOCK(bp); 2158 continue; 2159 } 2160 BO_UNLOCK(bp->b_bufobj); 2161 if (bp->b_flags & B_INVAL) { 2162 bremfreel(bp); 2163 mtx_unlock(&bqlock); 2164 brelse(bp); 2165 flushed++; 2166 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2167 mtx_lock(&bqlock); 2168 continue; 2169 } 2170 2171 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2172 if (flushdeps == 0) { 2173 BUF_UNLOCK(bp); 2174 continue; 2175 } 2176 hasdeps = 1; 2177 } else 2178 hasdeps = 0; 2179 /* 2180 * We must hold the lock on a vnode before writing 2181 * one of its buffers. Otherwise we may confuse, or 2182 * in the case of a snapshot vnode, deadlock the 2183 * system. 2184 * 2185 * The lock order here is the reverse of the normal 2186 * of vnode followed by buf lock. This is ok because 2187 * the NOWAIT will prevent deadlock. 2188 */ 2189 vp = bp->b_vp; 2190 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2191 BUF_UNLOCK(bp); 2192 continue; 2193 } 2194 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 2195 mtx_unlock(&bqlock); 2196 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2197 bp, bp->b_vp, bp->b_flags); 2198 vfs_bio_awrite(bp); 2199 vn_finished_write(mp); 2200 VOP_UNLOCK(vp, 0); 2201 flushwithdeps += hasdeps; 2202 flushed++; 2203 waitrunningbufspace(); 2204 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2205 mtx_lock(&bqlock); 2206 continue; 2207 } 2208 vn_finished_write(mp); 2209 BUF_UNLOCK(bp); 2210 } 2211 TAILQ_REMOVE(&bufqueues[queue], &sentinel, b_freelist); 2212 mtx_unlock(&bqlock); 2213 return (flushed); 2214} 2215 2216/* 2217 * Check to see if a block is currently memory resident. 2218 */ 2219struct buf * 2220incore(struct bufobj *bo, daddr_t blkno) 2221{ 2222 struct buf *bp; 2223 2224 BO_LOCK(bo); 2225 bp = gbincore(bo, blkno); 2226 BO_UNLOCK(bo); 2227 return (bp); 2228} 2229 2230/* 2231 * Returns true if no I/O is needed to access the 2232 * associated VM object. This is like incore except 2233 * it also hunts around in the VM system for the data. 2234 */ 2235 2236static int 2237inmem(struct vnode * vp, daddr_t blkno) 2238{ 2239 vm_object_t obj; 2240 vm_offset_t toff, tinc, size; 2241 vm_page_t m; 2242 vm_ooffset_t off; 2243 2244 ASSERT_VOP_LOCKED(vp, "inmem"); 2245 2246 if (incore(&vp->v_bufobj, blkno)) 2247 return 1; 2248 if (vp->v_mount == NULL) 2249 return 0; 2250 obj = vp->v_object; 2251 if (obj == NULL) 2252 return (0); 2253 2254 size = PAGE_SIZE; 2255 if (size > vp->v_mount->mnt_stat.f_iosize) 2256 size = vp->v_mount->mnt_stat.f_iosize; 2257 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2258 2259 VM_OBJECT_LOCK(obj); 2260 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2261 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2262 if (!m) 2263 goto notinmem; 2264 tinc = size; 2265 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2266 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2267 if (vm_page_is_valid(m, 2268 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2269 goto notinmem; 2270 } 2271 VM_OBJECT_UNLOCK(obj); 2272 return 1; 2273 2274notinmem: 2275 VM_OBJECT_UNLOCK(obj); 2276 return (0); 2277} 2278 2279/* 2280 * vfs_setdirty: 2281 * 2282 * Sets the dirty range for a buffer based on the status of the dirty 2283 * bits in the pages comprising the buffer. 2284 * 2285 * The range is limited to the size of the buffer. 2286 * 2287 * This routine is primarily used by NFS, but is generalized for the 2288 * B_VMIO case. 2289 */ 2290static void 2291vfs_setdirty(struct buf *bp) 2292{ 2293 2294 /* 2295 * Degenerate case - empty buffer 2296 */ 2297 2298 if (bp->b_bufsize == 0) 2299 return; 2300 2301 /* 2302 * We qualify the scan for modified pages on whether the 2303 * object has been flushed yet. 2304 */ 2305 2306 if ((bp->b_flags & B_VMIO) == 0) 2307 return; 2308 2309 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2310 vfs_setdirty_locked_object(bp); 2311 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2312} 2313 2314static void 2315vfs_setdirty_locked_object(struct buf *bp) 2316{ 2317 vm_object_t object; 2318 int i; 2319 2320 object = bp->b_bufobj->bo_object; 2321 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2322 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2323 vm_offset_t boffset; 2324 vm_offset_t eoffset; 2325 2326 vm_page_lock_queues(); 2327 /* 2328 * test the pages to see if they have been modified directly 2329 * by users through the VM system. 2330 */ 2331 for (i = 0; i < bp->b_npages; i++) 2332 vm_page_test_dirty(bp->b_pages[i]); 2333 2334 /* 2335 * Calculate the encompassing dirty range, boffset and eoffset, 2336 * (eoffset - boffset) bytes. 2337 */ 2338 2339 for (i = 0; i < bp->b_npages; i++) { 2340 if (bp->b_pages[i]->dirty) 2341 break; 2342 } 2343 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2344 2345 for (i = bp->b_npages - 1; i >= 0; --i) { 2346 if (bp->b_pages[i]->dirty) { 2347 break; 2348 } 2349 } 2350 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2351 2352 vm_page_unlock_queues(); 2353 /* 2354 * Fit it to the buffer. 2355 */ 2356 2357 if (eoffset > bp->b_bcount) 2358 eoffset = bp->b_bcount; 2359 2360 /* 2361 * If we have a good dirty range, merge with the existing 2362 * dirty range. 2363 */ 2364 2365 if (boffset < eoffset) { 2366 if (bp->b_dirtyoff > boffset) 2367 bp->b_dirtyoff = boffset; 2368 if (bp->b_dirtyend < eoffset) 2369 bp->b_dirtyend = eoffset; 2370 } 2371 } 2372} 2373 2374/* 2375 * getblk: 2376 * 2377 * Get a block given a specified block and offset into a file/device. 2378 * The buffers B_DONE bit will be cleared on return, making it almost 2379 * ready for an I/O initiation. B_INVAL may or may not be set on 2380 * return. The caller should clear B_INVAL prior to initiating a 2381 * READ. 2382 * 2383 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2384 * an existing buffer. 2385 * 2386 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2387 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2388 * and then cleared based on the backing VM. If the previous buffer is 2389 * non-0-sized but invalid, B_CACHE will be cleared. 2390 * 2391 * If getblk() must create a new buffer, the new buffer is returned with 2392 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2393 * case it is returned with B_INVAL clear and B_CACHE set based on the 2394 * backing VM. 2395 * 2396 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2397 * B_CACHE bit is clear. 2398 * 2399 * What this means, basically, is that the caller should use B_CACHE to 2400 * determine whether the buffer is fully valid or not and should clear 2401 * B_INVAL prior to issuing a read. If the caller intends to validate 2402 * the buffer by loading its data area with something, the caller needs 2403 * to clear B_INVAL. If the caller does this without issuing an I/O, 2404 * the caller should set B_CACHE ( as an optimization ), else the caller 2405 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2406 * a write attempt or if it was a successfull read. If the caller 2407 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2408 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2409 */ 2410struct buf * 2411getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2412 int flags) 2413{ 2414 struct buf *bp; 2415 struct bufobj *bo; 2416 int error; 2417 2418 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2419 ASSERT_VOP_LOCKED(vp, "getblk"); 2420 if (size > MAXBSIZE) 2421 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2422 2423 bo = &vp->v_bufobj; 2424loop: 2425 /* 2426 * Block if we are low on buffers. Certain processes are allowed 2427 * to completely exhaust the buffer cache. 2428 * 2429 * If this check ever becomes a bottleneck it may be better to 2430 * move it into the else, when gbincore() fails. At the moment 2431 * it isn't a problem. 2432 * 2433 * XXX remove if 0 sections (clean this up after its proven) 2434 */ 2435 if (numfreebuffers == 0) { 2436 if (TD_IS_IDLETHREAD(curthread)) 2437 return NULL; 2438 mtx_lock(&nblock); 2439 needsbuffer |= VFS_BIO_NEED_ANY; 2440 mtx_unlock(&nblock); 2441 } 2442 2443 BO_LOCK(bo); 2444 bp = gbincore(bo, blkno); 2445 if (bp != NULL) { 2446 int lockflags; 2447 /* 2448 * Buffer is in-core. If the buffer is not busy, it must 2449 * be on a queue. 2450 */ 2451 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2452 2453 if (flags & GB_LOCK_NOWAIT) 2454 lockflags |= LK_NOWAIT; 2455 2456 error = BUF_TIMELOCK(bp, lockflags, 2457 VI_MTX(vp), "getblk", slpflag, slptimeo); 2458 2459 /* 2460 * If we slept and got the lock we have to restart in case 2461 * the buffer changed identities. 2462 */ 2463 if (error == ENOLCK) 2464 goto loop; 2465 /* We timed out or were interrupted. */ 2466 else if (error) 2467 return (NULL); 2468 2469 /* 2470 * The buffer is locked. B_CACHE is cleared if the buffer is 2471 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2472 * and for a VMIO buffer B_CACHE is adjusted according to the 2473 * backing VM cache. 2474 */ 2475 if (bp->b_flags & B_INVAL) 2476 bp->b_flags &= ~B_CACHE; 2477 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2478 bp->b_flags |= B_CACHE; 2479 bremfree(bp); 2480 2481 /* 2482 * check for size inconsistancies for non-VMIO case. 2483 */ 2484 2485 if (bp->b_bcount != size) { 2486 if ((bp->b_flags & B_VMIO) == 0 || 2487 (size > bp->b_kvasize)) { 2488 if (bp->b_flags & B_DELWRI) { 2489 /* 2490 * If buffer is pinned and caller does 2491 * not want sleep waiting for it to be 2492 * unpinned, bail out 2493 * */ 2494 if (bp->b_pin_count > 0) { 2495 if (flags & GB_LOCK_NOWAIT) { 2496 bqrelse(bp); 2497 return (NULL); 2498 } else { 2499 bunpin_wait(bp); 2500 } 2501 } 2502 bp->b_flags |= B_NOCACHE; 2503 bwrite(bp); 2504 } else { 2505 if (LIST_EMPTY(&bp->b_dep)) { 2506 bp->b_flags |= B_RELBUF; 2507 brelse(bp); 2508 } else { 2509 bp->b_flags |= B_NOCACHE; 2510 bwrite(bp); 2511 } 2512 } 2513 goto loop; 2514 } 2515 } 2516 2517 /* 2518 * If the size is inconsistant in the VMIO case, we can resize 2519 * the buffer. This might lead to B_CACHE getting set or 2520 * cleared. If the size has not changed, B_CACHE remains 2521 * unchanged from its previous state. 2522 */ 2523 2524 if (bp->b_bcount != size) 2525 allocbuf(bp, size); 2526 2527 KASSERT(bp->b_offset != NOOFFSET, 2528 ("getblk: no buffer offset")); 2529 2530 /* 2531 * A buffer with B_DELWRI set and B_CACHE clear must 2532 * be committed before we can return the buffer in 2533 * order to prevent the caller from issuing a read 2534 * ( due to B_CACHE not being set ) and overwriting 2535 * it. 2536 * 2537 * Most callers, including NFS and FFS, need this to 2538 * operate properly either because they assume they 2539 * can issue a read if B_CACHE is not set, or because 2540 * ( for example ) an uncached B_DELWRI might loop due 2541 * to softupdates re-dirtying the buffer. In the latter 2542 * case, B_CACHE is set after the first write completes, 2543 * preventing further loops. 2544 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2545 * above while extending the buffer, we cannot allow the 2546 * buffer to remain with B_CACHE set after the write 2547 * completes or it will represent a corrupt state. To 2548 * deal with this we set B_NOCACHE to scrap the buffer 2549 * after the write. 2550 * 2551 * We might be able to do something fancy, like setting 2552 * B_CACHE in bwrite() except if B_DELWRI is already set, 2553 * so the below call doesn't set B_CACHE, but that gets real 2554 * confusing. This is much easier. 2555 */ 2556 2557 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2558 bp->b_flags |= B_NOCACHE; 2559 bwrite(bp); 2560 goto loop; 2561 } 2562 bp->b_flags &= ~B_DONE; 2563 } else { 2564 int bsize, maxsize, vmio; 2565 off_t offset; 2566 2567 /* 2568 * Buffer is not in-core, create new buffer. The buffer 2569 * returned by getnewbuf() is locked. Note that the returned 2570 * buffer is also considered valid (not marked B_INVAL). 2571 */ 2572 BO_UNLOCK(bo); 2573 /* 2574 * If the user does not want us to create the buffer, bail out 2575 * here. 2576 */ 2577 if (flags & GB_NOCREAT) 2578 return NULL; 2579 bsize = bo->bo_bsize; 2580 offset = blkno * bsize; 2581 vmio = vp->v_object != NULL; 2582 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2583 maxsize = imax(maxsize, bsize); 2584 2585 bp = getnewbuf(slpflag, slptimeo, size, maxsize); 2586 if (bp == NULL) { 2587 if (slpflag || slptimeo) 2588 return NULL; 2589 goto loop; 2590 } 2591 2592 /* 2593 * This code is used to make sure that a buffer is not 2594 * created while the getnewbuf routine is blocked. 2595 * This can be a problem whether the vnode is locked or not. 2596 * If the buffer is created out from under us, we have to 2597 * throw away the one we just created. 2598 * 2599 * Note: this must occur before we associate the buffer 2600 * with the vp especially considering limitations in 2601 * the splay tree implementation when dealing with duplicate 2602 * lblkno's. 2603 */ 2604 BO_LOCK(bo); 2605 if (gbincore(bo, blkno)) { 2606 BO_UNLOCK(bo); 2607 bp->b_flags |= B_INVAL; 2608 brelse(bp); 2609 goto loop; 2610 } 2611 2612 /* 2613 * Insert the buffer into the hash, so that it can 2614 * be found by incore. 2615 */ 2616 bp->b_blkno = bp->b_lblkno = blkno; 2617 bp->b_offset = offset; 2618 bgetvp(vp, bp); 2619 BO_UNLOCK(bo); 2620 2621 /* 2622 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2623 * buffer size starts out as 0, B_CACHE will be set by 2624 * allocbuf() for the VMIO case prior to it testing the 2625 * backing store for validity. 2626 */ 2627 2628 if (vmio) { 2629 bp->b_flags |= B_VMIO; 2630#if defined(VFS_BIO_DEBUG) 2631 if (vn_canvmio(vp) != TRUE) 2632 printf("getblk: VMIO on vnode type %d\n", 2633 vp->v_type); 2634#endif 2635 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2636 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2637 bp, vp->v_object, bp->b_bufobj->bo_object)); 2638 } else { 2639 bp->b_flags &= ~B_VMIO; 2640 KASSERT(bp->b_bufobj->bo_object == NULL, 2641 ("ARGH! has b_bufobj->bo_object %p %p\n", 2642 bp, bp->b_bufobj->bo_object)); 2643 } 2644 2645 allocbuf(bp, size); 2646 bp->b_flags &= ~B_DONE; 2647 } 2648 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2649 BUF_ASSERT_HELD(bp); 2650 KASSERT(bp->b_bufobj == bo, 2651 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2652 return (bp); 2653} 2654 2655/* 2656 * Get an empty, disassociated buffer of given size. The buffer is initially 2657 * set to B_INVAL. 2658 */ 2659struct buf * 2660geteblk(int size) 2661{ 2662 struct buf *bp; 2663 int maxsize; 2664 2665 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2666 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2667 continue; 2668 allocbuf(bp, size); 2669 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2670 BUF_ASSERT_HELD(bp); 2671 return (bp); 2672} 2673 2674 2675/* 2676 * This code constitutes the buffer memory from either anonymous system 2677 * memory (in the case of non-VMIO operations) or from an associated 2678 * VM object (in the case of VMIO operations). This code is able to 2679 * resize a buffer up or down. 2680 * 2681 * Note that this code is tricky, and has many complications to resolve 2682 * deadlock or inconsistant data situations. Tread lightly!!! 2683 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2684 * the caller. Calling this code willy nilly can result in the loss of data. 2685 * 2686 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2687 * B_CACHE for the non-VMIO case. 2688 */ 2689 2690int 2691allocbuf(struct buf *bp, int size) 2692{ 2693 int newbsize, mbsize; 2694 int i; 2695 2696 BUF_ASSERT_HELD(bp); 2697 2698 if (bp->b_kvasize < size) 2699 panic("allocbuf: buffer too small"); 2700 2701 if ((bp->b_flags & B_VMIO) == 0) { 2702 caddr_t origbuf; 2703 int origbufsize; 2704 /* 2705 * Just get anonymous memory from the kernel. Don't 2706 * mess with B_CACHE. 2707 */ 2708 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2709 if (bp->b_flags & B_MALLOC) 2710 newbsize = mbsize; 2711 else 2712 newbsize = round_page(size); 2713 2714 if (newbsize < bp->b_bufsize) { 2715 /* 2716 * malloced buffers are not shrunk 2717 */ 2718 if (bp->b_flags & B_MALLOC) { 2719 if (newbsize) { 2720 bp->b_bcount = size; 2721 } else { 2722 free(bp->b_data, M_BIOBUF); 2723 if (bp->b_bufsize) { 2724 atomic_subtract_int( 2725 &bufmallocspace, 2726 bp->b_bufsize); 2727 bufspacewakeup(); 2728 bp->b_bufsize = 0; 2729 } 2730 bp->b_saveaddr = bp->b_kvabase; 2731 bp->b_data = bp->b_saveaddr; 2732 bp->b_bcount = 0; 2733 bp->b_flags &= ~B_MALLOC; 2734 } 2735 return 1; 2736 } 2737 vm_hold_free_pages( 2738 bp, 2739 (vm_offset_t) bp->b_data + newbsize, 2740 (vm_offset_t) bp->b_data + bp->b_bufsize); 2741 } else if (newbsize > bp->b_bufsize) { 2742 /* 2743 * We only use malloced memory on the first allocation. 2744 * and revert to page-allocated memory when the buffer 2745 * grows. 2746 */ 2747 /* 2748 * There is a potential smp race here that could lead 2749 * to bufmallocspace slightly passing the max. It 2750 * is probably extremely rare and not worth worrying 2751 * over. 2752 */ 2753 if ( (bufmallocspace < maxbufmallocspace) && 2754 (bp->b_bufsize == 0) && 2755 (mbsize <= PAGE_SIZE/2)) { 2756 2757 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2758 bp->b_bufsize = mbsize; 2759 bp->b_bcount = size; 2760 bp->b_flags |= B_MALLOC; 2761 atomic_add_int(&bufmallocspace, mbsize); 2762 return 1; 2763 } 2764 origbuf = NULL; 2765 origbufsize = 0; 2766 /* 2767 * If the buffer is growing on its other-than-first allocation, 2768 * then we revert to the page-allocation scheme. 2769 */ 2770 if (bp->b_flags & B_MALLOC) { 2771 origbuf = bp->b_data; 2772 origbufsize = bp->b_bufsize; 2773 bp->b_data = bp->b_kvabase; 2774 if (bp->b_bufsize) { 2775 atomic_subtract_int(&bufmallocspace, 2776 bp->b_bufsize); 2777 bufspacewakeup(); 2778 bp->b_bufsize = 0; 2779 } 2780 bp->b_flags &= ~B_MALLOC; 2781 newbsize = round_page(newbsize); 2782 } 2783 vm_hold_load_pages( 2784 bp, 2785 (vm_offset_t) bp->b_data + bp->b_bufsize, 2786 (vm_offset_t) bp->b_data + newbsize); 2787 if (origbuf) { 2788 bcopy(origbuf, bp->b_data, origbufsize); 2789 free(origbuf, M_BIOBUF); 2790 } 2791 } 2792 } else { 2793 int desiredpages; 2794 2795 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2796 desiredpages = (size == 0) ? 0 : 2797 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2798 2799 if (bp->b_flags & B_MALLOC) 2800 panic("allocbuf: VMIO buffer can't be malloced"); 2801 /* 2802 * Set B_CACHE initially if buffer is 0 length or will become 2803 * 0-length. 2804 */ 2805 if (size == 0 || bp->b_bufsize == 0) 2806 bp->b_flags |= B_CACHE; 2807 2808 if (newbsize < bp->b_bufsize) { 2809 /* 2810 * DEV_BSIZE aligned new buffer size is less then the 2811 * DEV_BSIZE aligned existing buffer size. Figure out 2812 * if we have to remove any pages. 2813 */ 2814 if (desiredpages < bp->b_npages) { 2815 vm_page_t m; 2816 2817 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2818 vm_page_lock_queues(); 2819 for (i = desiredpages; i < bp->b_npages; i++) { 2820 /* 2821 * the page is not freed here -- it 2822 * is the responsibility of 2823 * vnode_pager_setsize 2824 */ 2825 m = bp->b_pages[i]; 2826 KASSERT(m != bogus_page, 2827 ("allocbuf: bogus page found")); 2828 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2829 vm_page_lock_queues(); 2830 2831 bp->b_pages[i] = NULL; 2832 vm_page_unwire(m, 0); 2833 } 2834 vm_page_unlock_queues(); 2835 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2836 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2837 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2838 bp->b_npages = desiredpages; 2839 } 2840 } else if (size > bp->b_bcount) { 2841 /* 2842 * We are growing the buffer, possibly in a 2843 * byte-granular fashion. 2844 */ 2845 struct vnode *vp; 2846 vm_object_t obj; 2847 vm_offset_t toff; 2848 vm_offset_t tinc; 2849 2850 /* 2851 * Step 1, bring in the VM pages from the object, 2852 * allocating them if necessary. We must clear 2853 * B_CACHE if these pages are not valid for the 2854 * range covered by the buffer. 2855 */ 2856 2857 vp = bp->b_vp; 2858 obj = bp->b_bufobj->bo_object; 2859 2860 VM_OBJECT_LOCK(obj); 2861 while (bp->b_npages < desiredpages) { 2862 vm_page_t m; 2863 vm_pindex_t pi; 2864 2865 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2866 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2867 /* 2868 * note: must allocate system pages 2869 * since blocking here could intefere 2870 * with paging I/O, no matter which 2871 * process we are. 2872 */ 2873 m = vm_page_alloc(obj, pi, 2874 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2875 VM_ALLOC_WIRED); 2876 if (m == NULL) { 2877 atomic_add_int(&vm_pageout_deficit, 2878 desiredpages - bp->b_npages); 2879 VM_OBJECT_UNLOCK(obj); 2880 VM_WAIT; 2881 VM_OBJECT_LOCK(obj); 2882 } else { 2883 if (m->valid == 0) 2884 bp->b_flags &= ~B_CACHE; 2885 bp->b_pages[bp->b_npages] = m; 2886 ++bp->b_npages; 2887 } 2888 continue; 2889 } 2890 2891 /* 2892 * We found a page. If we have to sleep on it, 2893 * retry because it might have gotten freed out 2894 * from under us. 2895 * 2896 * We can only test VPO_BUSY here. Blocking on 2897 * m->busy might lead to a deadlock: 2898 * 2899 * vm_fault->getpages->cluster_read->allocbuf 2900 * 2901 */ 2902 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2903 continue; 2904 2905 /* 2906 * We have a good page. 2907 */ 2908 vm_page_lock_queues(); 2909 vm_page_wire(m); 2910 vm_page_unlock_queues(); 2911 bp->b_pages[bp->b_npages] = m; 2912 ++bp->b_npages; 2913 } 2914 2915 /* 2916 * Step 2. We've loaded the pages into the buffer, 2917 * we have to figure out if we can still have B_CACHE 2918 * set. Note that B_CACHE is set according to the 2919 * byte-granular range ( bcount and size ), new the 2920 * aligned range ( newbsize ). 2921 * 2922 * The VM test is against m->valid, which is DEV_BSIZE 2923 * aligned. Needless to say, the validity of the data 2924 * needs to also be DEV_BSIZE aligned. Note that this 2925 * fails with NFS if the server or some other client 2926 * extends the file's EOF. If our buffer is resized, 2927 * B_CACHE may remain set! XXX 2928 */ 2929 2930 toff = bp->b_bcount; 2931 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2932 2933 while ((bp->b_flags & B_CACHE) && toff < size) { 2934 vm_pindex_t pi; 2935 2936 if (tinc > (size - toff)) 2937 tinc = size - toff; 2938 2939 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2940 PAGE_SHIFT; 2941 2942 vfs_buf_test_cache( 2943 bp, 2944 bp->b_offset, 2945 toff, 2946 tinc, 2947 bp->b_pages[pi] 2948 ); 2949 toff += tinc; 2950 tinc = PAGE_SIZE; 2951 } 2952 VM_OBJECT_UNLOCK(obj); 2953 2954 /* 2955 * Step 3, fixup the KVM pmap. Remember that 2956 * bp->b_data is relative to bp->b_offset, but 2957 * bp->b_offset may be offset into the first page. 2958 */ 2959 2960 bp->b_data = (caddr_t) 2961 trunc_page((vm_offset_t)bp->b_data); 2962 pmap_qenter( 2963 (vm_offset_t)bp->b_data, 2964 bp->b_pages, 2965 bp->b_npages 2966 ); 2967 2968 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2969 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2970 } 2971 } 2972 if (newbsize < bp->b_bufsize) 2973 bufspacewakeup(); 2974 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2975 bp->b_bcount = size; /* requested buffer size */ 2976 return 1; 2977} 2978 2979void 2980biodone(struct bio *bp) 2981{ 2982 struct mtx *mtxp; 2983 void (*done)(struct bio *); 2984 2985 mtxp = mtx_pool_find(mtxpool_sleep, bp); 2986 mtx_lock(mtxp); 2987 bp->bio_flags |= BIO_DONE; 2988 done = bp->bio_done; 2989 if (done == NULL) 2990 wakeup(bp); 2991 mtx_unlock(mtxp); 2992 if (done != NULL) 2993 done(bp); 2994} 2995 2996/* 2997 * Wait for a BIO to finish. 2998 * 2999 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3000 * case is not yet clear. 3001 */ 3002int 3003biowait(struct bio *bp, const char *wchan) 3004{ 3005 struct mtx *mtxp; 3006 3007 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3008 mtx_lock(mtxp); 3009 while ((bp->bio_flags & BIO_DONE) == 0) 3010 msleep(bp, mtxp, PRIBIO, wchan, hz / 10); 3011 mtx_unlock(mtxp); 3012 if (bp->bio_error != 0) 3013 return (bp->bio_error); 3014 if (!(bp->bio_flags & BIO_ERROR)) 3015 return (0); 3016 return (EIO); 3017} 3018 3019void 3020biofinish(struct bio *bp, struct devstat *stat, int error) 3021{ 3022 3023 if (error) { 3024 bp->bio_error = error; 3025 bp->bio_flags |= BIO_ERROR; 3026 } 3027 if (stat != NULL) 3028 devstat_end_transaction_bio(stat, bp); 3029 biodone(bp); 3030} 3031 3032/* 3033 * bufwait: 3034 * 3035 * Wait for buffer I/O completion, returning error status. The buffer 3036 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3037 * error and cleared. 3038 */ 3039int 3040bufwait(struct buf *bp) 3041{ 3042 if (bp->b_iocmd == BIO_READ) 3043 bwait(bp, PRIBIO, "biord"); 3044 else 3045 bwait(bp, PRIBIO, "biowr"); 3046 if (bp->b_flags & B_EINTR) { 3047 bp->b_flags &= ~B_EINTR; 3048 return (EINTR); 3049 } 3050 if (bp->b_ioflags & BIO_ERROR) { 3051 return (bp->b_error ? bp->b_error : EIO); 3052 } else { 3053 return (0); 3054 } 3055} 3056 3057 /* 3058 * Call back function from struct bio back up to struct buf. 3059 */ 3060static void 3061bufdonebio(struct bio *bip) 3062{ 3063 struct buf *bp; 3064 3065 bp = bip->bio_caller2; 3066 bp->b_resid = bp->b_bcount - bip->bio_completed; 3067 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3068 bp->b_ioflags = bip->bio_flags; 3069 bp->b_error = bip->bio_error; 3070 if (bp->b_error) 3071 bp->b_ioflags |= BIO_ERROR; 3072 bufdone(bp); 3073 g_destroy_bio(bip); 3074} 3075 3076void 3077dev_strategy(struct cdev *dev, struct buf *bp) 3078{ 3079 struct cdevsw *csw; 3080 struct bio *bip; 3081 3082 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3083 panic("b_iocmd botch"); 3084 for (;;) { 3085 bip = g_new_bio(); 3086 if (bip != NULL) 3087 break; 3088 /* Try again later */ 3089 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3090 } 3091 bip->bio_cmd = bp->b_iocmd; 3092 bip->bio_offset = bp->b_iooffset; 3093 bip->bio_length = bp->b_bcount; 3094 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3095 bip->bio_data = bp->b_data; 3096 bip->bio_done = bufdonebio; 3097 bip->bio_caller2 = bp; 3098 bip->bio_dev = dev; 3099 KASSERT(dev->si_refcount > 0, 3100 ("dev_strategy on un-referenced struct cdev *(%s)", 3101 devtoname(dev))); 3102 csw = dev_refthread(dev); 3103 if (csw == NULL) { 3104 g_destroy_bio(bip); 3105 bp->b_error = ENXIO; 3106 bp->b_ioflags = BIO_ERROR; 3107 bufdone(bp); 3108 return; 3109 } 3110 (*csw->d_strategy)(bip); 3111 dev_relthread(dev); 3112} 3113 3114/* 3115 * bufdone: 3116 * 3117 * Finish I/O on a buffer, optionally calling a completion function. 3118 * This is usually called from an interrupt so process blocking is 3119 * not allowed. 3120 * 3121 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3122 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3123 * assuming B_INVAL is clear. 3124 * 3125 * For the VMIO case, we set B_CACHE if the op was a read and no 3126 * read error occured, or if the op was a write. B_CACHE is never 3127 * set if the buffer is invalid or otherwise uncacheable. 3128 * 3129 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3130 * initiator to leave B_INVAL set to brelse the buffer out of existance 3131 * in the biodone routine. 3132 */ 3133void 3134bufdone(struct buf *bp) 3135{ 3136 struct bufobj *dropobj; 3137 void (*biodone)(struct buf *); 3138 3139 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3140 dropobj = NULL; 3141 3142 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3143 BUF_ASSERT_HELD(bp); 3144 3145 runningbufwakeup(bp); 3146 if (bp->b_iocmd == BIO_WRITE) 3147 dropobj = bp->b_bufobj; 3148 /* call optional completion function if requested */ 3149 if (bp->b_iodone != NULL) { 3150 biodone = bp->b_iodone; 3151 bp->b_iodone = NULL; 3152 (*biodone) (bp); 3153 if (dropobj) 3154 bufobj_wdrop(dropobj); 3155 return; 3156 } 3157 3158 bufdone_finish(bp); 3159 3160 if (dropobj) 3161 bufobj_wdrop(dropobj); 3162} 3163 3164void 3165bufdone_finish(struct buf *bp) 3166{ 3167 BUF_ASSERT_HELD(bp); 3168 3169 if (!LIST_EMPTY(&bp->b_dep)) 3170 buf_complete(bp); 3171 3172 if (bp->b_flags & B_VMIO) { 3173 int i; 3174 vm_ooffset_t foff; 3175 vm_page_t m; 3176 vm_object_t obj; 3177 int iosize; 3178 struct vnode *vp = bp->b_vp; 3179 boolean_t are_queues_locked; 3180 3181 obj = bp->b_bufobj->bo_object; 3182 3183#if defined(VFS_BIO_DEBUG) 3184 mp_fixme("usecount and vflag accessed without locks."); 3185 if (vp->v_usecount == 0) { 3186 panic("biodone: zero vnode ref count"); 3187 } 3188 3189 KASSERT(vp->v_object != NULL, 3190 ("biodone: vnode %p has no vm_object", vp)); 3191#endif 3192 3193 foff = bp->b_offset; 3194 KASSERT(bp->b_offset != NOOFFSET, 3195 ("biodone: no buffer offset")); 3196 3197 VM_OBJECT_LOCK(obj); 3198#if defined(VFS_BIO_DEBUG) 3199 if (obj->paging_in_progress < bp->b_npages) { 3200 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3201 obj->paging_in_progress, bp->b_npages); 3202 } 3203#endif 3204 3205 /* 3206 * Set B_CACHE if the op was a normal read and no error 3207 * occured. B_CACHE is set for writes in the b*write() 3208 * routines. 3209 */ 3210 iosize = bp->b_bcount - bp->b_resid; 3211 if (bp->b_iocmd == BIO_READ && 3212 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3213 !(bp->b_ioflags & BIO_ERROR)) { 3214 bp->b_flags |= B_CACHE; 3215 } 3216 if (bp->b_iocmd == BIO_READ) { 3217 vm_page_lock_queues(); 3218 are_queues_locked = TRUE; 3219 } else 3220 are_queues_locked = FALSE; 3221 for (i = 0; i < bp->b_npages; i++) { 3222 int bogusflag = 0; 3223 int resid; 3224 3225 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3226 if (resid > iosize) 3227 resid = iosize; 3228 3229 /* 3230 * cleanup bogus pages, restoring the originals 3231 */ 3232 m = bp->b_pages[i]; 3233 if (m == bogus_page) { 3234 bogusflag = 1; 3235 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3236 if (m == NULL) 3237 panic("biodone: page disappeared!"); 3238 bp->b_pages[i] = m; 3239 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3240 bp->b_pages, bp->b_npages); 3241 } 3242#if defined(VFS_BIO_DEBUG) 3243 if (OFF_TO_IDX(foff) != m->pindex) { 3244 printf( 3245"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3246 (intmax_t)foff, (uintmax_t)m->pindex); 3247 } 3248#endif 3249 3250 /* 3251 * In the write case, the valid and clean bits are 3252 * already changed correctly ( see bdwrite() ), so we 3253 * only need to do this here in the read case. 3254 */ 3255 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3256 vfs_page_set_valid(bp, foff, m); 3257 } 3258 3259 /* 3260 * when debugging new filesystems or buffer I/O methods, this 3261 * is the most common error that pops up. if you see this, you 3262 * have not set the page busy flag correctly!!! 3263 */ 3264 if (m->busy == 0) { 3265 printf("biodone: page busy < 0, " 3266 "pindex: %d, foff: 0x(%x,%x), " 3267 "resid: %d, index: %d\n", 3268 (int) m->pindex, (int)(foff >> 32), 3269 (int) foff & 0xffffffff, resid, i); 3270 if (!vn_isdisk(vp, NULL)) 3271 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3272 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3273 (intmax_t) bp->b_lblkno, 3274 bp->b_flags, bp->b_npages); 3275 else 3276 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3277 (intmax_t) bp->b_lblkno, 3278 bp->b_flags, bp->b_npages); 3279 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3280 (u_long)m->valid, (u_long)m->dirty, 3281 m->wire_count); 3282 panic("biodone: page busy < 0\n"); 3283 } 3284 vm_page_io_finish(m); 3285 vm_object_pip_subtract(obj, 1); 3286 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3287 iosize -= resid; 3288 } 3289 if (are_queues_locked) 3290 vm_page_unlock_queues(); 3291 vm_object_pip_wakeupn(obj, 0); 3292 VM_OBJECT_UNLOCK(obj); 3293 } 3294 3295 /* 3296 * For asynchronous completions, release the buffer now. The brelse 3297 * will do a wakeup there if necessary - so no need to do a wakeup 3298 * here in the async case. The sync case always needs to do a wakeup. 3299 */ 3300 3301 if (bp->b_flags & B_ASYNC) { 3302 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3303 brelse(bp); 3304 else 3305 bqrelse(bp); 3306 } else 3307 bdone(bp); 3308} 3309 3310/* 3311 * This routine is called in lieu of iodone in the case of 3312 * incomplete I/O. This keeps the busy status for pages 3313 * consistant. 3314 */ 3315void 3316vfs_unbusy_pages(struct buf *bp) 3317{ 3318 int i; 3319 vm_object_t obj; 3320 vm_page_t m; 3321 3322 runningbufwakeup(bp); 3323 if (!(bp->b_flags & B_VMIO)) 3324 return; 3325 3326 obj = bp->b_bufobj->bo_object; 3327 VM_OBJECT_LOCK(obj); 3328 for (i = 0; i < bp->b_npages; i++) { 3329 m = bp->b_pages[i]; 3330 if (m == bogus_page) { 3331 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3332 if (!m) 3333 panic("vfs_unbusy_pages: page missing\n"); 3334 bp->b_pages[i] = m; 3335 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3336 bp->b_pages, bp->b_npages); 3337 } 3338 vm_object_pip_subtract(obj, 1); 3339 vm_page_io_finish(m); 3340 } 3341 vm_object_pip_wakeupn(obj, 0); 3342 VM_OBJECT_UNLOCK(obj); 3343} 3344 3345/* 3346 * vfs_page_set_valid: 3347 * 3348 * Set the valid bits in a page based on the supplied offset. The 3349 * range is restricted to the buffer's size. 3350 * 3351 * This routine is typically called after a read completes. 3352 */ 3353static void 3354vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3355{ 3356 vm_ooffset_t soff, eoff; 3357 3358 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3359 /* 3360 * Start and end offsets in buffer. eoff - soff may not cross a 3361 * page boundry or cross the end of the buffer. The end of the 3362 * buffer, in this case, is our file EOF, not the allocation size 3363 * of the buffer. 3364 */ 3365 soff = off; 3366 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3367 if (eoff > bp->b_offset + bp->b_bcount) 3368 eoff = bp->b_offset + bp->b_bcount; 3369 3370 /* 3371 * Set valid range. This is typically the entire buffer and thus the 3372 * entire page. 3373 */ 3374 if (eoff > soff) { 3375 vm_page_set_validclean( 3376 m, 3377 (vm_offset_t) (soff & PAGE_MASK), 3378 (vm_offset_t) (eoff - soff) 3379 ); 3380 } 3381} 3382 3383/* 3384 * This routine is called before a device strategy routine. 3385 * It is used to tell the VM system that paging I/O is in 3386 * progress, and treat the pages associated with the buffer 3387 * almost as being VPO_BUSY. Also the object paging_in_progress 3388 * flag is handled to make sure that the object doesn't become 3389 * inconsistant. 3390 * 3391 * Since I/O has not been initiated yet, certain buffer flags 3392 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3393 * and should be ignored. 3394 */ 3395void 3396vfs_busy_pages(struct buf *bp, int clear_modify) 3397{ 3398 int i, bogus; 3399 vm_object_t obj; 3400 vm_ooffset_t foff; 3401 vm_page_t m; 3402 3403 if (!(bp->b_flags & B_VMIO)) 3404 return; 3405 3406 obj = bp->b_bufobj->bo_object; 3407 foff = bp->b_offset; 3408 KASSERT(bp->b_offset != NOOFFSET, 3409 ("vfs_busy_pages: no buffer offset")); 3410 VM_OBJECT_LOCK(obj); 3411 if (bp->b_bufsize != 0) 3412 vfs_setdirty_locked_object(bp); 3413retry: 3414 for (i = 0; i < bp->b_npages; i++) { 3415 m = bp->b_pages[i]; 3416 3417 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3418 goto retry; 3419 } 3420 bogus = 0; 3421 vm_page_lock_queues(); 3422 for (i = 0; i < bp->b_npages; i++) { 3423 m = bp->b_pages[i]; 3424 3425 if ((bp->b_flags & B_CLUSTER) == 0) { 3426 vm_object_pip_add(obj, 1); 3427 vm_page_io_start(m); 3428 } 3429 /* 3430 * When readying a buffer for a read ( i.e 3431 * clear_modify == 0 ), it is important to do 3432 * bogus_page replacement for valid pages in 3433 * partially instantiated buffers. Partially 3434 * instantiated buffers can, in turn, occur when 3435 * reconstituting a buffer from its VM backing store 3436 * base. We only have to do this if B_CACHE is 3437 * clear ( which causes the I/O to occur in the 3438 * first place ). The replacement prevents the read 3439 * I/O from overwriting potentially dirty VM-backed 3440 * pages. XXX bogus page replacement is, uh, bogus. 3441 * It may not work properly with small-block devices. 3442 * We need to find a better way. 3443 */ 3444 pmap_remove_all(m); 3445 if (clear_modify) 3446 vfs_page_set_valid(bp, foff, m); 3447 else if (m->valid == VM_PAGE_BITS_ALL && 3448 (bp->b_flags & B_CACHE) == 0) { 3449 bp->b_pages[i] = bogus_page; 3450 bogus++; 3451 } 3452 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3453 } 3454 vm_page_unlock_queues(); 3455 VM_OBJECT_UNLOCK(obj); 3456 if (bogus) 3457 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3458 bp->b_pages, bp->b_npages); 3459} 3460 3461/* 3462 * Tell the VM system that the pages associated with this buffer 3463 * are clean. This is used for delayed writes where the data is 3464 * going to go to disk eventually without additional VM intevention. 3465 * 3466 * Note that while we only really need to clean through to b_bcount, we 3467 * just go ahead and clean through to b_bufsize. 3468 */ 3469static void 3470vfs_clean_pages(struct buf *bp) 3471{ 3472 int i; 3473 vm_ooffset_t foff, noff, eoff; 3474 vm_page_t m; 3475 3476 if (!(bp->b_flags & B_VMIO)) 3477 return; 3478 3479 foff = bp->b_offset; 3480 KASSERT(bp->b_offset != NOOFFSET, 3481 ("vfs_clean_pages: no buffer offset")); 3482 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3483 vm_page_lock_queues(); 3484 for (i = 0; i < bp->b_npages; i++) { 3485 m = bp->b_pages[i]; 3486 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3487 eoff = noff; 3488 3489 if (eoff > bp->b_offset + bp->b_bufsize) 3490 eoff = bp->b_offset + bp->b_bufsize; 3491 vfs_page_set_valid(bp, foff, m); 3492 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3493 foff = noff; 3494 } 3495 vm_page_unlock_queues(); 3496 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3497} 3498 3499/* 3500 * vfs_bio_set_validclean: 3501 * 3502 * Set the range within the buffer to valid and clean. The range is 3503 * relative to the beginning of the buffer, b_offset. Note that b_offset 3504 * itself may be offset from the beginning of the first page. 3505 * 3506 */ 3507 3508void 3509vfs_bio_set_validclean(struct buf *bp, int base, int size) 3510{ 3511 int i, n; 3512 vm_page_t m; 3513 3514 if (!(bp->b_flags & B_VMIO)) 3515 return; 3516 /* 3517 * Fixup base to be relative to beginning of first page. 3518 * Set initial n to be the maximum number of bytes in the 3519 * first page that can be validated. 3520 */ 3521 3522 base += (bp->b_offset & PAGE_MASK); 3523 n = PAGE_SIZE - (base & PAGE_MASK); 3524 3525 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3526 vm_page_lock_queues(); 3527 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3528 m = bp->b_pages[i]; 3529 if (n > size) 3530 n = size; 3531 vm_page_set_validclean(m, base & PAGE_MASK, n); 3532 base += n; 3533 size -= n; 3534 n = PAGE_SIZE; 3535 } 3536 vm_page_unlock_queues(); 3537 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3538} 3539 3540/* 3541 * vfs_bio_clrbuf: 3542 * 3543 * clear a buffer. This routine essentially fakes an I/O, so we need 3544 * to clear BIO_ERROR and B_INVAL. 3545 * 3546 * Note that while we only theoretically need to clear through b_bcount, 3547 * we go ahead and clear through b_bufsize. 3548 */ 3549 3550void 3551vfs_bio_clrbuf(struct buf *bp) 3552{ 3553 int i, j, mask = 0; 3554 caddr_t sa, ea; 3555 3556 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3557 clrbuf(bp); 3558 return; 3559 } 3560 3561 bp->b_flags &= ~B_INVAL; 3562 bp->b_ioflags &= ~BIO_ERROR; 3563 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3564 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3565 (bp->b_offset & PAGE_MASK) == 0) { 3566 if (bp->b_pages[0] == bogus_page) 3567 goto unlock; 3568 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3569 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3570 if ((bp->b_pages[0]->valid & mask) == mask) 3571 goto unlock; 3572 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3573 ((bp->b_pages[0]->valid & mask) == 0)) { 3574 bzero(bp->b_data, bp->b_bufsize); 3575 bp->b_pages[0]->valid |= mask; 3576 goto unlock; 3577 } 3578 } 3579 ea = sa = bp->b_data; 3580 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3581 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3582 ea = (caddr_t)(vm_offset_t)ulmin( 3583 (u_long)(vm_offset_t)ea, 3584 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3585 if (bp->b_pages[i] == bogus_page) 3586 continue; 3587 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3588 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3589 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3590 if ((bp->b_pages[i]->valid & mask) == mask) 3591 continue; 3592 if ((bp->b_pages[i]->valid & mask) == 0) { 3593 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3594 bzero(sa, ea - sa); 3595 } else { 3596 for (; sa < ea; sa += DEV_BSIZE, j++) { 3597 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3598 (bp->b_pages[i]->valid & (1 << j)) == 0) 3599 bzero(sa, DEV_BSIZE); 3600 } 3601 } 3602 bp->b_pages[i]->valid |= mask; 3603 } 3604unlock: 3605 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3606 bp->b_resid = 0; 3607} 3608 3609/* 3610 * vm_hold_load_pages and vm_hold_free_pages get pages into 3611 * a buffers address space. The pages are anonymous and are 3612 * not associated with a file object. 3613 */ 3614static void 3615vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3616{ 3617 vm_offset_t pg; 3618 vm_page_t p; 3619 int index; 3620 3621 to = round_page(to); 3622 from = round_page(from); 3623 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3624 3625 VM_OBJECT_LOCK(kernel_object); 3626 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3627tryagain: 3628 /* 3629 * note: must allocate system pages since blocking here 3630 * could intefere with paging I/O, no matter which 3631 * process we are. 3632 */ 3633 p = vm_page_alloc(kernel_object, 3634 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3635 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3636 if (!p) { 3637 atomic_add_int(&vm_pageout_deficit, 3638 (to - pg) >> PAGE_SHIFT); 3639 VM_OBJECT_UNLOCK(kernel_object); 3640 VM_WAIT; 3641 VM_OBJECT_LOCK(kernel_object); 3642 goto tryagain; 3643 } 3644 p->valid = VM_PAGE_BITS_ALL; 3645 pmap_qenter(pg, &p, 1); 3646 bp->b_pages[index] = p; 3647 } 3648 VM_OBJECT_UNLOCK(kernel_object); 3649 bp->b_npages = index; 3650} 3651 3652/* Return pages associated with this buf to the vm system */ 3653static void 3654vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3655{ 3656 vm_offset_t pg; 3657 vm_page_t p; 3658 int index, newnpages; 3659 3660 from = round_page(from); 3661 to = round_page(to); 3662 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3663 3664 VM_OBJECT_LOCK(kernel_object); 3665 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3666 p = bp->b_pages[index]; 3667 if (p && (index < bp->b_npages)) { 3668 if (p->busy) { 3669 printf( 3670 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3671 (intmax_t)bp->b_blkno, 3672 (intmax_t)bp->b_lblkno); 3673 } 3674 bp->b_pages[index] = NULL; 3675 pmap_qremove(pg, 1); 3676 vm_page_lock_queues(); 3677 vm_page_unwire(p, 0); 3678 vm_page_free(p); 3679 vm_page_unlock_queues(); 3680 } 3681 } 3682 VM_OBJECT_UNLOCK(kernel_object); 3683 bp->b_npages = newnpages; 3684} 3685 3686/* 3687 * Map an IO request into kernel virtual address space. 3688 * 3689 * All requests are (re)mapped into kernel VA space. 3690 * Notice that we use b_bufsize for the size of the buffer 3691 * to be mapped. b_bcount might be modified by the driver. 3692 * 3693 * Note that even if the caller determines that the address space should 3694 * be valid, a race or a smaller-file mapped into a larger space may 3695 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3696 * check the return value. 3697 */ 3698int 3699vmapbuf(struct buf *bp) 3700{ 3701 caddr_t addr, kva; 3702 vm_prot_t prot; 3703 int pidx, i; 3704 struct vm_page *m; 3705 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3706 3707 if (bp->b_bufsize < 0) 3708 return (-1); 3709 prot = VM_PROT_READ; 3710 if (bp->b_iocmd == BIO_READ) 3711 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3712 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3713 addr < bp->b_data + bp->b_bufsize; 3714 addr += PAGE_SIZE, pidx++) { 3715 /* 3716 * Do the vm_fault if needed; do the copy-on-write thing 3717 * when reading stuff off device into memory. 3718 * 3719 * NOTE! Must use pmap_extract() because addr may be in 3720 * the userland address space, and kextract is only guarenteed 3721 * to work for the kernland address space (see: sparc64 port). 3722 */ 3723retry: 3724 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3725 prot) < 0) { 3726 vm_page_lock_queues(); 3727 for (i = 0; i < pidx; ++i) { 3728 vm_page_unhold(bp->b_pages[i]); 3729 bp->b_pages[i] = NULL; 3730 } 3731 vm_page_unlock_queues(); 3732 return(-1); 3733 } 3734 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3735 if (m == NULL) 3736 goto retry; 3737 bp->b_pages[pidx] = m; 3738 } 3739 if (pidx > btoc(MAXPHYS)) 3740 panic("vmapbuf: mapped more than MAXPHYS"); 3741 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3742 3743 kva = bp->b_saveaddr; 3744 bp->b_npages = pidx; 3745 bp->b_saveaddr = bp->b_data; 3746 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3747 return(0); 3748} 3749 3750/* 3751 * Free the io map PTEs associated with this IO operation. 3752 * We also invalidate the TLB entries and restore the original b_addr. 3753 */ 3754void 3755vunmapbuf(struct buf *bp) 3756{ 3757 int pidx; 3758 int npages; 3759 3760 npages = bp->b_npages; 3761 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3762 vm_page_lock_queues(); 3763 for (pidx = 0; pidx < npages; pidx++) 3764 vm_page_unhold(bp->b_pages[pidx]); 3765 vm_page_unlock_queues(); 3766 3767 bp->b_data = bp->b_saveaddr; 3768} 3769 3770void 3771bdone(struct buf *bp) 3772{ 3773 struct mtx *mtxp; 3774 3775 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3776 mtx_lock(mtxp); 3777 bp->b_flags |= B_DONE; 3778 wakeup(bp); 3779 mtx_unlock(mtxp); 3780} 3781 3782void 3783bwait(struct buf *bp, u_char pri, const char *wchan) 3784{ 3785 struct mtx *mtxp; 3786 3787 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3788 mtx_lock(mtxp); 3789 while ((bp->b_flags & B_DONE) == 0) 3790 msleep(bp, mtxp, pri, wchan, 0); 3791 mtx_unlock(mtxp); 3792} 3793 3794int 3795bufsync(struct bufobj *bo, int waitfor, struct thread *td) 3796{ 3797 3798 return (VOP_FSYNC(bo->__bo_vnode, waitfor, td)); 3799} 3800 3801void 3802bufstrategy(struct bufobj *bo, struct buf *bp) 3803{ 3804 int i = 0; 3805 struct vnode *vp; 3806 3807 vp = bp->b_vp; 3808 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3809 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3810 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3811 i = VOP_STRATEGY(vp, bp); 3812 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3813} 3814 3815void 3816bufobj_wrefl(struct bufobj *bo) 3817{ 3818 3819 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3820 ASSERT_BO_LOCKED(bo); 3821 bo->bo_numoutput++; 3822} 3823 3824void 3825bufobj_wref(struct bufobj *bo) 3826{ 3827 3828 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3829 BO_LOCK(bo); 3830 bo->bo_numoutput++; 3831 BO_UNLOCK(bo); 3832} 3833 3834void 3835bufobj_wdrop(struct bufobj *bo) 3836{ 3837 3838 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3839 BO_LOCK(bo); 3840 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3841 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3842 bo->bo_flag &= ~BO_WWAIT; 3843 wakeup(&bo->bo_numoutput); 3844 } 3845 BO_UNLOCK(bo); 3846} 3847 3848int 3849bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3850{ 3851 int error; 3852 3853 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3854 ASSERT_BO_LOCKED(bo); 3855 error = 0; 3856 while (bo->bo_numoutput) { 3857 bo->bo_flag |= BO_WWAIT; 3858 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3859 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3860 if (error) 3861 break; 3862 } 3863 return (error); 3864} 3865 3866void 3867bpin(struct buf *bp) 3868{ 3869 struct mtx *mtxp; 3870 3871 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3872 mtx_lock(mtxp); 3873 bp->b_pin_count++; 3874 mtx_unlock(mtxp); 3875} 3876 3877void 3878bunpin(struct buf *bp) 3879{ 3880 struct mtx *mtxp; 3881 3882 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3883 mtx_lock(mtxp); 3884 if (--bp->b_pin_count == 0) 3885 wakeup(bp); 3886 mtx_unlock(mtxp); 3887} 3888 3889void 3890bunpin_wait(struct buf *bp) 3891{ 3892 struct mtx *mtxp; 3893 3894 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3895 mtx_lock(mtxp); 3896 while (bp->b_pin_count > 0) 3897 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 3898 mtx_unlock(mtxp); 3899} 3900 3901#include "opt_ddb.h" 3902#ifdef DDB 3903#include <ddb/ddb.h> 3904 3905/* DDB command to show buffer data */ 3906DB_SHOW_COMMAND(buffer, db_show_buffer) 3907{ 3908 /* get args */ 3909 struct buf *bp = (struct buf *)addr; 3910 3911 if (!have_addr) { 3912 db_printf("usage: show buffer <addr>\n"); 3913 return; 3914 } 3915 3916 db_printf("buf at %p\n", bp); 3917 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3918 db_printf( 3919 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3920 "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n", 3921 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3922 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno); 3923 if (bp->b_npages) { 3924 int i; 3925 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3926 for (i = 0; i < bp->b_npages; i++) { 3927 vm_page_t m; 3928 m = bp->b_pages[i]; 3929 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3930 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3931 if ((i + 1) < bp->b_npages) 3932 db_printf(","); 3933 } 3934 db_printf("\n"); 3935 } 3936 lockmgr_printinfo(&bp->b_lock); 3937} 3938 3939DB_SHOW_COMMAND(lockedbufs, lockedbufs) 3940{ 3941 struct buf *bp; 3942 int i; 3943 3944 for (i = 0; i < nbuf; i++) { 3945 bp = &buf[i]; 3946 if (BUF_ISLOCKED(bp)) { 3947 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 3948 db_printf("\n"); 3949 } 3950 } 3951} 3952#endif /* DDB */ 3953