vfs_bio.c revision 176249
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 176249 2008-02-13 20:44:19Z attilio $"); 43 44#include <sys/param.h> 45#include <sys/systm.h> 46#include <sys/bio.h> 47#include <sys/conf.h> 48#include <sys/buf.h> 49#include <sys/devicestat.h> 50#include <sys/eventhandler.h> 51#include <sys/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 * Lock that protects against bwait()/bdone()/B_DONE races. 241 */ 242 243static struct mtx bdonelock; 244 245/* 246 * Lock that protects against bwait()/bdone()/B_DONE races. 247 */ 248static struct mtx bpinlock; 249 250/* 251 * Definitions for the buffer free lists. 252 */ 253#define BUFFER_QUEUES 6 /* number of free buffer queues */ 254 255#define QUEUE_NONE 0 /* on no queue */ 256#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 257#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 258#define QUEUE_DIRTY_GIANT 3 /* B_DELWRI buffers that need giant */ 259#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 260#define QUEUE_EMPTY 5 /* empty buffer headers */ 261 262/* Queues for free buffers with various properties */ 263static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 264 265/* Lock for the bufqueues */ 266static struct mtx bqlock; 267 268/* 269 * Single global constant for BUF_WMESG, to avoid getting multiple references. 270 * buf_wmesg is referred from macros. 271 */ 272const char *buf_wmesg = BUF_WMESG; 273 274#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 275#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 276#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 277#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 278 279#ifdef DIRECTIO 280extern void ffs_rawread_setup(void); 281#endif /* DIRECTIO */ 282/* 283 * numdirtywakeup: 284 * 285 * If someone is blocked due to there being too many dirty buffers, 286 * and numdirtybuffers is now reasonable, wake them up. 287 */ 288 289static __inline void 290numdirtywakeup(int level) 291{ 292 293 if (numdirtybuffers <= level) { 294 mtx_lock(&nblock); 295 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 296 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 297 wakeup(&needsbuffer); 298 } 299 mtx_unlock(&nblock); 300 } 301} 302 303/* 304 * bufspacewakeup: 305 * 306 * Called when buffer space is potentially available for recovery. 307 * getnewbuf() will block on this flag when it is unable to free 308 * sufficient buffer space. Buffer space becomes recoverable when 309 * bp's get placed back in the queues. 310 */ 311 312static __inline void 313bufspacewakeup(void) 314{ 315 316 /* 317 * If someone is waiting for BUF space, wake them up. Even 318 * though we haven't freed the kva space yet, the waiting 319 * process will be able to now. 320 */ 321 mtx_lock(&nblock); 322 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 323 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 324 wakeup(&needsbuffer); 325 } 326 mtx_unlock(&nblock); 327} 328 329/* 330 * runningbufwakeup() - in-progress I/O accounting. 331 * 332 */ 333void 334runningbufwakeup(struct buf *bp) 335{ 336 337 if (bp->b_runningbufspace) { 338 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 339 bp->b_runningbufspace = 0; 340 mtx_lock(&rbreqlock); 341 if (runningbufreq && runningbufspace <= lorunningspace) { 342 runningbufreq = 0; 343 wakeup(&runningbufreq); 344 } 345 mtx_unlock(&rbreqlock); 346 } 347} 348 349/* 350 * bufcountwakeup: 351 * 352 * Called when a buffer has been added to one of the free queues to 353 * account for the buffer and to wakeup anyone waiting for free buffers. 354 * This typically occurs when large amounts of metadata are being handled 355 * by the buffer cache ( else buffer space runs out first, usually ). 356 */ 357 358static __inline void 359bufcountwakeup(void) 360{ 361 362 atomic_add_int(&numfreebuffers, 1); 363 mtx_lock(&nblock); 364 if (needsbuffer) { 365 needsbuffer &= ~VFS_BIO_NEED_ANY; 366 if (numfreebuffers >= hifreebuffers) 367 needsbuffer &= ~VFS_BIO_NEED_FREE; 368 wakeup(&needsbuffer); 369 } 370 mtx_unlock(&nblock); 371} 372 373/* 374 * waitrunningbufspace() 375 * 376 * runningbufspace is a measure of the amount of I/O currently 377 * running. This routine is used in async-write situations to 378 * prevent creating huge backups of pending writes to a device. 379 * Only asynchronous writes are governed by this function. 380 * 381 * Reads will adjust runningbufspace, but will not block based on it. 382 * The read load has a side effect of reducing the allowed write load. 383 * 384 * This does NOT turn an async write into a sync write. It waits 385 * for earlier writes to complete and generally returns before the 386 * caller's write has reached the device. 387 */ 388void 389waitrunningbufspace(void) 390{ 391 392 mtx_lock(&rbreqlock); 393 while (runningbufspace > hirunningspace) { 394 ++runningbufreq; 395 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 396 } 397 mtx_unlock(&rbreqlock); 398} 399 400 401/* 402 * vfs_buf_test_cache: 403 * 404 * Called when a buffer is extended. This function clears the B_CACHE 405 * bit if the newly extended portion of the buffer does not contain 406 * valid data. 407 */ 408static __inline 409void 410vfs_buf_test_cache(struct buf *bp, 411 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 412 vm_page_t m) 413{ 414 415 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 416 if (bp->b_flags & B_CACHE) { 417 int base = (foff + off) & PAGE_MASK; 418 if (vm_page_is_valid(m, base, size) == 0) 419 bp->b_flags &= ~B_CACHE; 420 } 421} 422 423/* Wake up the buffer daemon if necessary */ 424static __inline 425void 426bd_wakeup(int dirtybuflevel) 427{ 428 429 mtx_lock(&bdlock); 430 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 431 bd_request = 1; 432 wakeup(&bd_request); 433 } 434 mtx_unlock(&bdlock); 435} 436 437/* 438 * bd_speedup - speedup the buffer cache flushing code 439 */ 440 441static __inline 442void 443bd_speedup(void) 444{ 445 446 bd_wakeup(1); 447} 448 449/* 450 * Calculating buffer cache scaling values and reserve space for buffer 451 * headers. This is called during low level kernel initialization and 452 * may be called more then once. We CANNOT write to the memory area 453 * being reserved at this time. 454 */ 455caddr_t 456kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 457{ 458 int maxbuf; 459 460 /* 461 * physmem_est is in pages. Convert it to kilobytes (assumes 462 * PAGE_SIZE is >= 1K) 463 */ 464 physmem_est = physmem_est * (PAGE_SIZE / 1024); 465 466 /* 467 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 468 * For the first 64MB of ram nominally allocate sufficient buffers to 469 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 470 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 471 * the buffer cache we limit the eventual kva reservation to 472 * maxbcache bytes. 473 * 474 * factor represents the 1/4 x ram conversion. 475 */ 476 if (nbuf == 0) { 477 int factor = 4 * BKVASIZE / 1024; 478 479 nbuf = 50; 480 if (physmem_est > 4096) 481 nbuf += min((physmem_est - 4096) / factor, 482 65536 / factor); 483 if (physmem_est > 65536) 484 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 485 486 if (maxbcache && nbuf > maxbcache / BKVASIZE) 487 nbuf = maxbcache / BKVASIZE; 488 489 /* XXX Avoid integer overflows later on with maxbufspace. */ 490 maxbuf = (INT_MAX / 3) / BKVASIZE; 491 if (nbuf > maxbuf) 492 nbuf = maxbuf; 493 } 494 495#if 0 496 /* 497 * Do not allow the buffer_map to be more then 1/2 the size of the 498 * kernel_map. 499 */ 500 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 501 (BKVASIZE * 2)) { 502 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 503 (BKVASIZE * 2); 504 printf("Warning: nbufs capped at %d\n", nbuf); 505 } 506#endif 507 508 /* 509 * swbufs are used as temporary holders for I/O, such as paging I/O. 510 * We have no less then 16 and no more then 256. 511 */ 512 nswbuf = max(min(nbuf/4, 256), 16); 513#ifdef NSWBUF_MIN 514 if (nswbuf < NSWBUF_MIN) 515 nswbuf = NSWBUF_MIN; 516#endif 517#ifdef DIRECTIO 518 ffs_rawread_setup(); 519#endif 520 521 /* 522 * Reserve space for the buffer cache buffers 523 */ 524 swbuf = (void *)v; 525 v = (caddr_t)(swbuf + nswbuf); 526 buf = (void *)v; 527 v = (caddr_t)(buf + nbuf); 528 529 return(v); 530} 531 532/* Initialize the buffer subsystem. Called before use of any buffers. */ 533void 534bufinit(void) 535{ 536 struct buf *bp; 537 int i; 538 539 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 540 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 541 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 542 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 543 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 544 mtx_init(&bpinlock, "bpin lock", NULL, MTX_DEF); 545 546 /* next, make a null set of free lists */ 547 for (i = 0; i < BUFFER_QUEUES; i++) 548 TAILQ_INIT(&bufqueues[i]); 549 550 /* finally, initialize each buffer header and stick on empty q */ 551 for (i = 0; i < nbuf; i++) { 552 bp = &buf[i]; 553 bzero(bp, sizeof *bp); 554 bp->b_flags = B_INVAL; /* we're just an empty header */ 555 bp->b_rcred = NOCRED; 556 bp->b_wcred = NOCRED; 557 bp->b_qindex = QUEUE_EMPTY; 558 bp->b_vflags = 0; 559 bp->b_xflags = 0; 560 LIST_INIT(&bp->b_dep); 561 BUF_LOCKINIT(bp); 562 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 563 } 564 565 /* 566 * maxbufspace is the absolute maximum amount of buffer space we are 567 * allowed to reserve in KVM and in real terms. The absolute maximum 568 * is nominally used by buf_daemon. hibufspace is the nominal maximum 569 * used by most other processes. The differential is required to 570 * ensure that buf_daemon is able to run when other processes might 571 * be blocked waiting for buffer space. 572 * 573 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 574 * this may result in KVM fragmentation which is not handled optimally 575 * by the system. 576 */ 577 maxbufspace = nbuf * BKVASIZE; 578 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 579 lobufspace = hibufspace - MAXBSIZE; 580 581 lorunningspace = 512 * 1024; 582 hirunningspace = 1024 * 1024; 583 584/* 585 * Limit the amount of malloc memory since it is wired permanently into 586 * the kernel space. Even though this is accounted for in the buffer 587 * allocation, we don't want the malloced region to grow uncontrolled. 588 * The malloc scheme improves memory utilization significantly on average 589 * (small) directories. 590 */ 591 maxbufmallocspace = hibufspace / 20; 592 593/* 594 * Reduce the chance of a deadlock occuring by limiting the number 595 * of delayed-write dirty buffers we allow to stack up. 596 */ 597 hidirtybuffers = nbuf / 4 + 20; 598 dirtybufthresh = hidirtybuffers * 9 / 10; 599 numdirtybuffers = 0; 600/* 601 * To support extreme low-memory systems, make sure hidirtybuffers cannot 602 * eat up all available buffer space. This occurs when our minimum cannot 603 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 604 * BKVASIZE'd (8K) buffers. 605 */ 606 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 607 hidirtybuffers >>= 1; 608 } 609 lodirtybuffers = hidirtybuffers / 2; 610 611/* 612 * Try to keep the number of free buffers in the specified range, 613 * and give special processes (e.g. like buf_daemon) access to an 614 * emergency reserve. 615 */ 616 lofreebuffers = nbuf / 18 + 5; 617 hifreebuffers = 2 * lofreebuffers; 618 numfreebuffers = nbuf; 619 620/* 621 * Maximum number of async ops initiated per buf_daemon loop. This is 622 * somewhat of a hack at the moment, we really need to limit ourselves 623 * based on the number of bytes of I/O in-transit that were initiated 624 * from buf_daemon. 625 */ 626 627 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 628 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 629} 630 631/* 632 * bfreekva() - free the kva allocation for a buffer. 633 * 634 * Since this call frees up buffer space, we call bufspacewakeup(). 635 */ 636static void 637bfreekva(struct buf *bp) 638{ 639 640 if (bp->b_kvasize) { 641 atomic_add_int(&buffreekvacnt, 1); 642 atomic_subtract_int(&bufspace, bp->b_kvasize); 643 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase, 644 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 645 bp->b_kvasize = 0; 646 bufspacewakeup(); 647 } 648} 649 650/* 651 * bremfree: 652 * 653 * Mark the buffer for removal from the appropriate free list in brelse. 654 * 655 */ 656void 657bremfree(struct buf *bp) 658{ 659 660 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 661 KASSERT((bp->b_flags & B_REMFREE) == 0, 662 ("bremfree: buffer %p already marked for delayed removal.", bp)); 663 KASSERT(bp->b_qindex != QUEUE_NONE, 664 ("bremfree: buffer %p not on a queue.", bp)); 665 BUF_ASSERT_HELD(bp); 666 667 bp->b_flags |= B_REMFREE; 668 /* Fixup numfreebuffers count. */ 669 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 670 atomic_subtract_int(&numfreebuffers, 1); 671} 672 673/* 674 * bremfreef: 675 * 676 * Force an immediate removal from a free list. Used only in nfs when 677 * it abuses the b_freelist pointer. 678 */ 679void 680bremfreef(struct buf *bp) 681{ 682 mtx_lock(&bqlock); 683 bremfreel(bp); 684 mtx_unlock(&bqlock); 685} 686 687/* 688 * bremfreel: 689 * 690 * Removes a buffer from the free list, must be called with the 691 * bqlock held. 692 */ 693static void 694bremfreel(struct buf *bp) 695{ 696 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 697 bp, bp->b_vp, bp->b_flags); 698 KASSERT(bp->b_qindex != QUEUE_NONE, 699 ("bremfreel: buffer %p not on a queue.", bp)); 700 BUF_ASSERT_HELD(bp); 701 mtx_assert(&bqlock, MA_OWNED); 702 703 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 704 bp->b_qindex = QUEUE_NONE; 705 /* 706 * If this was a delayed bremfree() we only need to remove the buffer 707 * from the queue and return the stats are already done. 708 */ 709 if (bp->b_flags & B_REMFREE) { 710 bp->b_flags &= ~B_REMFREE; 711 return; 712 } 713 /* 714 * Fixup numfreebuffers count. If the buffer is invalid or not 715 * delayed-write, the buffer was free and we must decrement 716 * numfreebuffers. 717 */ 718 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 719 atomic_subtract_int(&numfreebuffers, 1); 720} 721 722 723/* 724 * Get a buffer with the specified data. Look in the cache first. We 725 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 726 * is set, the buffer is valid and we do not have to do anything ( see 727 * getblk() ). This is really just a special case of breadn(). 728 */ 729int 730bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 731 struct buf **bpp) 732{ 733 734 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 735} 736 737/* 738 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 739 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 740 * the buffer is valid and we do not have to do anything. 741 */ 742void 743breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 744 int cnt, struct ucred * cred) 745{ 746 struct buf *rabp; 747 int i; 748 749 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 750 if (inmem(vp, *rablkno)) 751 continue; 752 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 753 754 if ((rabp->b_flags & B_CACHE) == 0) { 755 if (!TD_IS_IDLETHREAD(curthread)) 756 curthread->td_ru.ru_inblock++; 757 rabp->b_flags |= B_ASYNC; 758 rabp->b_flags &= ~B_INVAL; 759 rabp->b_ioflags &= ~BIO_ERROR; 760 rabp->b_iocmd = BIO_READ; 761 if (rabp->b_rcred == NOCRED && cred != NOCRED) 762 rabp->b_rcred = crhold(cred); 763 vfs_busy_pages(rabp, 0); 764 BUF_KERNPROC(rabp); 765 rabp->b_iooffset = dbtob(rabp->b_blkno); 766 bstrategy(rabp); 767 } else { 768 brelse(rabp); 769 } 770 } 771} 772 773/* 774 * Operates like bread, but also starts asynchronous I/O on 775 * read-ahead blocks. 776 */ 777int 778breadn(struct vnode * vp, daddr_t blkno, int size, 779 daddr_t * rablkno, int *rabsize, 780 int cnt, struct ucred * cred, struct buf **bpp) 781{ 782 struct buf *bp; 783 int rv = 0, readwait = 0; 784 785 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 786 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 787 788 /* if not found in cache, do some I/O */ 789 if ((bp->b_flags & B_CACHE) == 0) { 790 if (!TD_IS_IDLETHREAD(curthread)) 791 curthread->td_ru.ru_inblock++; 792 bp->b_iocmd = BIO_READ; 793 bp->b_flags &= ~B_INVAL; 794 bp->b_ioflags &= ~BIO_ERROR; 795 if (bp->b_rcred == NOCRED && cred != NOCRED) 796 bp->b_rcred = crhold(cred); 797 vfs_busy_pages(bp, 0); 798 bp->b_iooffset = dbtob(bp->b_blkno); 799 bstrategy(bp); 800 ++readwait; 801 } 802 803 breada(vp, rablkno, rabsize, cnt, cred); 804 805 if (readwait) { 806 rv = bufwait(bp); 807 } 808 return (rv); 809} 810 811/* 812 * Write, release buffer on completion. (Done by iodone 813 * if async). Do not bother writing anything if the buffer 814 * is invalid. 815 * 816 * Note that we set B_CACHE here, indicating that buffer is 817 * fully valid and thus cacheable. This is true even of NFS 818 * now so we set it generally. This could be set either here 819 * or in biodone() since the I/O is synchronous. We put it 820 * here. 821 */ 822int 823bufwrite(struct buf *bp) 824{ 825 int oldflags; 826 struct vnode *vp; 827 int vp_md; 828 829 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 830 if (bp->b_flags & B_INVAL) { 831 brelse(bp); 832 return (0); 833 } 834 835 oldflags = bp->b_flags; 836 837 BUF_ASSERT_HELD(bp); 838 839 if (bp->b_pin_count > 0) 840 bunpin_wait(bp); 841 842 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 843 ("FFS background buffer should not get here %p", bp)); 844 845 vp = bp->b_vp; 846 if (vp) 847 vp_md = vp->v_vflag & VV_MD; 848 else 849 vp_md = 0; 850 851 /* Mark the buffer clean */ 852 bundirty(bp); 853 854 bp->b_flags &= ~B_DONE; 855 bp->b_ioflags &= ~BIO_ERROR; 856 bp->b_flags |= B_CACHE; 857 bp->b_iocmd = BIO_WRITE; 858 859 bufobj_wref(bp->b_bufobj); 860 vfs_busy_pages(bp, 1); 861 862 /* 863 * Normal bwrites pipeline writes 864 */ 865 bp->b_runningbufspace = bp->b_bufsize; 866 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 867 868 if (!TD_IS_IDLETHREAD(curthread)) 869 curthread->td_ru.ru_oublock++; 870 if (oldflags & B_ASYNC) 871 BUF_KERNPROC(bp); 872 bp->b_iooffset = dbtob(bp->b_blkno); 873 bstrategy(bp); 874 875 if ((oldflags & B_ASYNC) == 0) { 876 int rtval = bufwait(bp); 877 brelse(bp); 878 return (rtval); 879 } else { 880 /* 881 * don't allow the async write to saturate the I/O 882 * system. We will not deadlock here because 883 * we are blocking waiting for I/O that is already in-progress 884 * to complete. We do not block here if it is the update 885 * or syncer daemon trying to clean up as that can lead 886 * to deadlock. 887 */ 888 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 889 waitrunningbufspace(); 890 } 891 892 return (0); 893} 894 895void 896bufbdflush(struct bufobj *bo, struct buf *bp) 897{ 898 struct buf *nbp; 899 900 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 901 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 902 altbufferflushes++; 903 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 904 BO_LOCK(bo); 905 /* 906 * Try to find a buffer to flush. 907 */ 908 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 909 if ((nbp->b_vflags & BV_BKGRDINPROG) || 910 BUF_LOCK(nbp, 911 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 912 continue; 913 if (bp == nbp) 914 panic("bdwrite: found ourselves"); 915 BO_UNLOCK(bo); 916 /* Don't countdeps with the bo lock held. */ 917 if (buf_countdeps(nbp, 0)) { 918 BO_LOCK(bo); 919 BUF_UNLOCK(nbp); 920 continue; 921 } 922 if (nbp->b_flags & B_CLUSTEROK) { 923 vfs_bio_awrite(nbp); 924 } else { 925 bremfree(nbp); 926 bawrite(nbp); 927 } 928 dirtybufferflushes++; 929 break; 930 } 931 if (nbp == NULL) 932 BO_UNLOCK(bo); 933 } 934} 935 936/* 937 * Delayed write. (Buffer is marked dirty). Do not bother writing 938 * anything if the buffer is marked invalid. 939 * 940 * Note that since the buffer must be completely valid, we can safely 941 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 942 * biodone() in order to prevent getblk from writing the buffer 943 * out synchronously. 944 */ 945void 946bdwrite(struct buf *bp) 947{ 948 struct thread *td = curthread; 949 struct vnode *vp; 950 struct bufobj *bo; 951 952 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 953 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 954 BUF_ASSERT_HELD(bp); 955 956 if (bp->b_flags & B_INVAL) { 957 brelse(bp); 958 return; 959 } 960 961 /* 962 * If we have too many dirty buffers, don't create any more. 963 * If we are wildly over our limit, then force a complete 964 * cleanup. Otherwise, just keep the situation from getting 965 * out of control. Note that we have to avoid a recursive 966 * disaster and not try to clean up after our own cleanup! 967 */ 968 vp = bp->b_vp; 969 bo = bp->b_bufobj; 970 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 971 td->td_pflags |= TDP_INBDFLUSH; 972 BO_BDFLUSH(bo, bp); 973 td->td_pflags &= ~TDP_INBDFLUSH; 974 } else 975 recursiveflushes++; 976 977 bdirty(bp); 978 /* 979 * Set B_CACHE, indicating that the buffer is fully valid. This is 980 * true even of NFS now. 981 */ 982 bp->b_flags |= B_CACHE; 983 984 /* 985 * This bmap keeps the system from needing to do the bmap later, 986 * perhaps when the system is attempting to do a sync. Since it 987 * is likely that the indirect block -- or whatever other datastructure 988 * that the filesystem needs is still in memory now, it is a good 989 * thing to do this. Note also, that if the pageout daemon is 990 * requesting a sync -- there might not be enough memory to do 991 * the bmap then... So, this is important to do. 992 */ 993 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 994 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 995 } 996 997 /* 998 * Set the *dirty* buffer range based upon the VM system dirty pages. 999 */ 1000 vfs_setdirty(bp); 1001 1002 /* 1003 * We need to do this here to satisfy the vnode_pager and the 1004 * pageout daemon, so that it thinks that the pages have been 1005 * "cleaned". Note that since the pages are in a delayed write 1006 * buffer -- the VFS layer "will" see that the pages get written 1007 * out on the next sync, or perhaps the cluster will be completed. 1008 */ 1009 vfs_clean_pages(bp); 1010 bqrelse(bp); 1011 1012 /* 1013 * Wakeup the buffer flushing daemon if we have a lot of dirty 1014 * buffers (midpoint between our recovery point and our stall 1015 * point). 1016 */ 1017 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1018 1019 /* 1020 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1021 * due to the softdep code. 1022 */ 1023} 1024 1025/* 1026 * bdirty: 1027 * 1028 * Turn buffer into delayed write request. We must clear BIO_READ and 1029 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1030 * itself to properly update it in the dirty/clean lists. We mark it 1031 * B_DONE to ensure that any asynchronization of the buffer properly 1032 * clears B_DONE ( else a panic will occur later ). 1033 * 1034 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1035 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1036 * should only be called if the buffer is known-good. 1037 * 1038 * Since the buffer is not on a queue, we do not update the numfreebuffers 1039 * count. 1040 * 1041 * The buffer must be on QUEUE_NONE. 1042 */ 1043void 1044bdirty(struct buf *bp) 1045{ 1046 1047 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1048 bp, bp->b_vp, bp->b_flags); 1049 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1050 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1051 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1052 BUF_ASSERT_HELD(bp); 1053 bp->b_flags &= ~(B_RELBUF); 1054 bp->b_iocmd = BIO_WRITE; 1055 1056 if ((bp->b_flags & B_DELWRI) == 0) { 1057 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1058 reassignbuf(bp); 1059 atomic_add_int(&numdirtybuffers, 1); 1060 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1061 } 1062} 1063 1064/* 1065 * bundirty: 1066 * 1067 * Clear B_DELWRI for buffer. 1068 * 1069 * Since the buffer is not on a queue, we do not update the numfreebuffers 1070 * count. 1071 * 1072 * The buffer must be on QUEUE_NONE. 1073 */ 1074 1075void 1076bundirty(struct buf *bp) 1077{ 1078 1079 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1080 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1081 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1082 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1083 BUF_ASSERT_HELD(bp); 1084 1085 if (bp->b_flags & B_DELWRI) { 1086 bp->b_flags &= ~B_DELWRI; 1087 reassignbuf(bp); 1088 atomic_subtract_int(&numdirtybuffers, 1); 1089 numdirtywakeup(lodirtybuffers); 1090 } 1091 /* 1092 * Since it is now being written, we can clear its deferred write flag. 1093 */ 1094 bp->b_flags &= ~B_DEFERRED; 1095} 1096 1097/* 1098 * bawrite: 1099 * 1100 * Asynchronous write. Start output on a buffer, but do not wait for 1101 * it to complete. The buffer is released when the output completes. 1102 * 1103 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1104 * B_INVAL buffers. Not us. 1105 */ 1106void 1107bawrite(struct buf *bp) 1108{ 1109 1110 bp->b_flags |= B_ASYNC; 1111 (void) bwrite(bp); 1112} 1113 1114/* 1115 * bwillwrite: 1116 * 1117 * Called prior to the locking of any vnodes when we are expecting to 1118 * write. We do not want to starve the buffer cache with too many 1119 * dirty buffers so we block here. By blocking prior to the locking 1120 * of any vnodes we attempt to avoid the situation where a locked vnode 1121 * prevents the various system daemons from flushing related buffers. 1122 */ 1123 1124void 1125bwillwrite(void) 1126{ 1127 1128 if (numdirtybuffers >= hidirtybuffers) { 1129 mtx_lock(&nblock); 1130 while (numdirtybuffers >= hidirtybuffers) { 1131 bd_wakeup(1); 1132 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1133 msleep(&needsbuffer, &nblock, 1134 (PRIBIO + 4), "flswai", 0); 1135 } 1136 mtx_unlock(&nblock); 1137 } 1138} 1139 1140/* 1141 * Return true if we have too many dirty buffers. 1142 */ 1143int 1144buf_dirty_count_severe(void) 1145{ 1146 1147 return(numdirtybuffers >= hidirtybuffers); 1148} 1149 1150/* 1151 * brelse: 1152 * 1153 * Release a busy buffer and, if requested, free its resources. The 1154 * buffer will be stashed in the appropriate bufqueue[] allowing it 1155 * to be accessed later as a cache entity or reused for other purposes. 1156 */ 1157void 1158brelse(struct buf *bp) 1159{ 1160 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1161 bp, bp->b_vp, bp->b_flags); 1162 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1163 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1164 1165 if (bp->b_flags & B_MANAGED) { 1166 bqrelse(bp); 1167 return; 1168 } 1169 1170 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1171 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1172 /* 1173 * Failed write, redirty. Must clear BIO_ERROR to prevent 1174 * pages from being scrapped. If the error is anything 1175 * other than an I/O error (EIO), assume that retryingi 1176 * is futile. 1177 */ 1178 bp->b_ioflags &= ~BIO_ERROR; 1179 bdirty(bp); 1180 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1181 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1182 /* 1183 * Either a failed I/O or we were asked to free or not 1184 * cache the buffer. 1185 */ 1186 bp->b_flags |= B_INVAL; 1187 if (!LIST_EMPTY(&bp->b_dep)) 1188 buf_deallocate(bp); 1189 if (bp->b_flags & B_DELWRI) { 1190 atomic_subtract_int(&numdirtybuffers, 1); 1191 numdirtywakeup(lodirtybuffers); 1192 } 1193 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1194 if ((bp->b_flags & B_VMIO) == 0) { 1195 if (bp->b_bufsize) 1196 allocbuf(bp, 0); 1197 if (bp->b_vp) 1198 brelvp(bp); 1199 } 1200 } 1201 1202 /* 1203 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1204 * is called with B_DELWRI set, the underlying pages may wind up 1205 * getting freed causing a previous write (bdwrite()) to get 'lost' 1206 * because pages associated with a B_DELWRI bp are marked clean. 1207 * 1208 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1209 * if B_DELWRI is set. 1210 * 1211 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1212 * on pages to return pages to the VM page queues. 1213 */ 1214 if (bp->b_flags & B_DELWRI) 1215 bp->b_flags &= ~B_RELBUF; 1216 else if (vm_page_count_severe()) { 1217 /* 1218 * XXX This lock may not be necessary since BKGRDINPROG 1219 * cannot be set while we hold the buf lock, it can only be 1220 * cleared if it is already pending. 1221 */ 1222 if (bp->b_vp) { 1223 BO_LOCK(bp->b_bufobj); 1224 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1225 bp->b_flags |= B_RELBUF; 1226 BO_UNLOCK(bp->b_bufobj); 1227 } else 1228 bp->b_flags |= B_RELBUF; 1229 } 1230 1231 /* 1232 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1233 * constituted, not even NFS buffers now. Two flags effect this. If 1234 * B_INVAL, the struct buf is invalidated but the VM object is kept 1235 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1236 * 1237 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1238 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1239 * buffer is also B_INVAL because it hits the re-dirtying code above. 1240 * 1241 * Normally we can do this whether a buffer is B_DELWRI or not. If 1242 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1243 * the commit state and we cannot afford to lose the buffer. If the 1244 * buffer has a background write in progress, we need to keep it 1245 * around to prevent it from being reconstituted and starting a second 1246 * background write. 1247 */ 1248 if ((bp->b_flags & B_VMIO) 1249 && !(bp->b_vp->v_mount != NULL && 1250 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1251 !vn_isdisk(bp->b_vp, NULL) && 1252 (bp->b_flags & B_DELWRI)) 1253 ) { 1254 1255 int i, j, resid; 1256 vm_page_t m; 1257 off_t foff; 1258 vm_pindex_t poff; 1259 vm_object_t obj; 1260 1261 obj = bp->b_bufobj->bo_object; 1262 1263 /* 1264 * Get the base offset and length of the buffer. Note that 1265 * in the VMIO case if the buffer block size is not 1266 * page-aligned then b_data pointer may not be page-aligned. 1267 * But our b_pages[] array *IS* page aligned. 1268 * 1269 * block sizes less then DEV_BSIZE (usually 512) are not 1270 * supported due to the page granularity bits (m->valid, 1271 * m->dirty, etc...). 1272 * 1273 * See man buf(9) for more information 1274 */ 1275 resid = bp->b_bufsize; 1276 foff = bp->b_offset; 1277 VM_OBJECT_LOCK(obj); 1278 for (i = 0; i < bp->b_npages; i++) { 1279 int had_bogus = 0; 1280 1281 m = bp->b_pages[i]; 1282 1283 /* 1284 * If we hit a bogus page, fixup *all* the bogus pages 1285 * now. 1286 */ 1287 if (m == bogus_page) { 1288 poff = OFF_TO_IDX(bp->b_offset); 1289 had_bogus = 1; 1290 1291 for (j = i; j < bp->b_npages; j++) { 1292 vm_page_t mtmp; 1293 mtmp = bp->b_pages[j]; 1294 if (mtmp == bogus_page) { 1295 mtmp = vm_page_lookup(obj, poff + j); 1296 if (!mtmp) { 1297 panic("brelse: page missing\n"); 1298 } 1299 bp->b_pages[j] = mtmp; 1300 } 1301 } 1302 1303 if ((bp->b_flags & B_INVAL) == 0) { 1304 pmap_qenter( 1305 trunc_page((vm_offset_t)bp->b_data), 1306 bp->b_pages, bp->b_npages); 1307 } 1308 m = bp->b_pages[i]; 1309 } 1310 if ((bp->b_flags & B_NOCACHE) || 1311 (bp->b_ioflags & BIO_ERROR)) { 1312 int poffset = foff & PAGE_MASK; 1313 int presid = resid > (PAGE_SIZE - poffset) ? 1314 (PAGE_SIZE - poffset) : resid; 1315 1316 KASSERT(presid >= 0, ("brelse: extra page")); 1317 vm_page_lock_queues(); 1318 vm_page_set_invalid(m, poffset, presid); 1319 vm_page_unlock_queues(); 1320 if (had_bogus) 1321 printf("avoided corruption bug in bogus_page/brelse code\n"); 1322 } 1323 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1324 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1325 } 1326 VM_OBJECT_UNLOCK(obj); 1327 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1328 vfs_vmio_release(bp); 1329 1330 } else if (bp->b_flags & B_VMIO) { 1331 1332 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1333 vfs_vmio_release(bp); 1334 } 1335 1336 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1337 if (bp->b_bufsize != 0) 1338 allocbuf(bp, 0); 1339 if (bp->b_vp != NULL) 1340 brelvp(bp); 1341 } 1342 1343 if (BUF_LOCKRECURSED(bp)) { 1344 /* do not release to free list */ 1345 BUF_UNLOCK(bp); 1346 return; 1347 } 1348 1349 /* enqueue */ 1350 mtx_lock(&bqlock); 1351 /* Handle delayed bremfree() processing. */ 1352 if (bp->b_flags & B_REMFREE) 1353 bremfreel(bp); 1354 if (bp->b_qindex != QUEUE_NONE) 1355 panic("brelse: free buffer onto another queue???"); 1356 1357 /* buffers with no memory */ 1358 if (bp->b_bufsize == 0) { 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 1"); 1363 if (bp->b_kvasize) { 1364 bp->b_qindex = QUEUE_EMPTYKVA; 1365 } else { 1366 bp->b_qindex = QUEUE_EMPTY; 1367 } 1368 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1369 /* buffers with junk contents */ 1370 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1371 (bp->b_ioflags & BIO_ERROR)) { 1372 bp->b_flags |= B_INVAL; 1373 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1374 if (bp->b_vflags & BV_BKGRDINPROG) 1375 panic("losing buffer 2"); 1376 bp->b_qindex = QUEUE_CLEAN; 1377 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1378 /* remaining buffers */ 1379 } else { 1380 if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) == 1381 (B_DELWRI|B_NEEDSGIANT)) 1382 bp->b_qindex = QUEUE_DIRTY_GIANT; 1383 if (bp->b_flags & B_DELWRI) 1384 bp->b_qindex = QUEUE_DIRTY; 1385 else 1386 bp->b_qindex = QUEUE_CLEAN; 1387 if (bp->b_flags & B_AGE) 1388 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1389 else 1390 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1391 } 1392 mtx_unlock(&bqlock); 1393 1394 /* 1395 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1396 * placed the buffer on the correct queue. We must also disassociate 1397 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1398 * find it. 1399 */ 1400 if (bp->b_flags & B_INVAL) { 1401 if (bp->b_flags & B_DELWRI) 1402 bundirty(bp); 1403 if (bp->b_vp) 1404 brelvp(bp); 1405 } 1406 1407 /* 1408 * Fixup numfreebuffers count. The bp is on an appropriate queue 1409 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1410 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1411 * if B_INVAL is set ). 1412 */ 1413 1414 if (!(bp->b_flags & B_DELWRI)) 1415 bufcountwakeup(); 1416 1417 /* 1418 * Something we can maybe free or reuse 1419 */ 1420 if (bp->b_bufsize || bp->b_kvasize) 1421 bufspacewakeup(); 1422 1423 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1424 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1425 panic("brelse: not dirty"); 1426 /* unlock */ 1427 BUF_UNLOCK(bp); 1428} 1429 1430/* 1431 * Release a buffer back to the appropriate queue but do not try to free 1432 * it. The buffer is expected to be used again soon. 1433 * 1434 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1435 * biodone() to requeue an async I/O on completion. It is also used when 1436 * known good buffers need to be requeued but we think we may need the data 1437 * again soon. 1438 * 1439 * XXX we should be able to leave the B_RELBUF hint set on completion. 1440 */ 1441void 1442bqrelse(struct buf *bp) 1443{ 1444 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1445 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1446 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1447 1448 if (BUF_LOCKRECURSED(bp)) { 1449 /* do not release to free list */ 1450 BUF_UNLOCK(bp); 1451 return; 1452 } 1453 1454 if (bp->b_flags & B_MANAGED) { 1455 if (bp->b_flags & B_REMFREE) { 1456 mtx_lock(&bqlock); 1457 bremfreel(bp); 1458 mtx_unlock(&bqlock); 1459 } 1460 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1461 BUF_UNLOCK(bp); 1462 return; 1463 } 1464 1465 mtx_lock(&bqlock); 1466 /* Handle delayed bremfree() processing. */ 1467 if (bp->b_flags & B_REMFREE) 1468 bremfreel(bp); 1469 if (bp->b_qindex != QUEUE_NONE) 1470 panic("bqrelse: free buffer onto another queue???"); 1471 /* buffers with stale but valid contents */ 1472 if (bp->b_flags & B_DELWRI) { 1473 if (bp->b_flags & B_NEEDSGIANT) 1474 bp->b_qindex = QUEUE_DIRTY_GIANT; 1475 else 1476 bp->b_qindex = QUEUE_DIRTY; 1477 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1478 } else { 1479 /* 1480 * XXX This lock may not be necessary since BKGRDINPROG 1481 * cannot be set while we hold the buf lock, it can only be 1482 * cleared if it is already pending. 1483 */ 1484 BO_LOCK(bp->b_bufobj); 1485 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1486 BO_UNLOCK(bp->b_bufobj); 1487 bp->b_qindex = QUEUE_CLEAN; 1488 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1489 b_freelist); 1490 } else { 1491 /* 1492 * We are too low on memory, we have to try to free 1493 * the buffer (most importantly: the wired pages 1494 * making up its backing store) *now*. 1495 */ 1496 BO_UNLOCK(bp->b_bufobj); 1497 mtx_unlock(&bqlock); 1498 brelse(bp); 1499 return; 1500 } 1501 } 1502 mtx_unlock(&bqlock); 1503 1504 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1505 bufcountwakeup(); 1506 1507 /* 1508 * Something we can maybe free or reuse. 1509 */ 1510 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1511 bufspacewakeup(); 1512 1513 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1514 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1515 panic("bqrelse: not dirty"); 1516 /* unlock */ 1517 BUF_UNLOCK(bp); 1518} 1519 1520/* Give pages used by the bp back to the VM system (where possible) */ 1521static void 1522vfs_vmio_release(struct buf *bp) 1523{ 1524 int i; 1525 vm_page_t m; 1526 1527 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1528 vm_page_lock_queues(); 1529 for (i = 0; i < bp->b_npages; i++) { 1530 m = bp->b_pages[i]; 1531 bp->b_pages[i] = NULL; 1532 /* 1533 * In order to keep page LRU ordering consistent, put 1534 * everything on the inactive queue. 1535 */ 1536 vm_page_unwire(m, 0); 1537 /* 1538 * We don't mess with busy pages, it is 1539 * the responsibility of the process that 1540 * busied the pages to deal with them. 1541 */ 1542 if ((m->oflags & VPO_BUSY) || (m->busy != 0)) 1543 continue; 1544 1545 if (m->wire_count == 0) { 1546 /* 1547 * Might as well free the page if we can and it has 1548 * no valid data. We also free the page if the 1549 * buffer was used for direct I/O 1550 */ 1551 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1552 m->hold_count == 0) { 1553 vm_page_free(m); 1554 } else if (bp->b_flags & B_DIRECT) { 1555 vm_page_try_to_free(m); 1556 } else if (vm_page_count_severe()) { 1557 vm_page_try_to_cache(m); 1558 } 1559 } 1560 } 1561 vm_page_unlock_queues(); 1562 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1563 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1564 1565 if (bp->b_bufsize) { 1566 bufspacewakeup(); 1567 bp->b_bufsize = 0; 1568 } 1569 bp->b_npages = 0; 1570 bp->b_flags &= ~B_VMIO; 1571 if (bp->b_vp) 1572 brelvp(bp); 1573} 1574 1575/* 1576 * Check to see if a block at a particular lbn is available for a clustered 1577 * write. 1578 */ 1579static int 1580vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1581{ 1582 struct buf *bpa; 1583 int match; 1584 1585 match = 0; 1586 1587 /* If the buf isn't in core skip it */ 1588 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1589 return (0); 1590 1591 /* If the buf is busy we don't want to wait for it */ 1592 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1593 return (0); 1594 1595 /* Only cluster with valid clusterable delayed write buffers */ 1596 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1597 (B_DELWRI | B_CLUSTEROK)) 1598 goto done; 1599 1600 if (bpa->b_bufsize != size) 1601 goto done; 1602 1603 /* 1604 * Check to see if it is in the expected place on disk and that the 1605 * block has been mapped. 1606 */ 1607 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1608 match = 1; 1609done: 1610 BUF_UNLOCK(bpa); 1611 return (match); 1612} 1613 1614/* 1615 * vfs_bio_awrite: 1616 * 1617 * Implement clustered async writes for clearing out B_DELWRI buffers. 1618 * This is much better then the old way of writing only one buffer at 1619 * a time. Note that we may not be presented with the buffers in the 1620 * correct order, so we search for the cluster in both directions. 1621 */ 1622int 1623vfs_bio_awrite(struct buf *bp) 1624{ 1625 int i; 1626 int j; 1627 daddr_t lblkno = bp->b_lblkno; 1628 struct vnode *vp = bp->b_vp; 1629 int ncl; 1630 int nwritten; 1631 int size; 1632 int maxcl; 1633 1634 /* 1635 * right now we support clustered writing only to regular files. If 1636 * we find a clusterable block we could be in the middle of a cluster 1637 * rather then at the beginning. 1638 */ 1639 if ((vp->v_type == VREG) && 1640 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1641 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1642 1643 size = vp->v_mount->mnt_stat.f_iosize; 1644 maxcl = MAXPHYS / size; 1645 1646 VI_LOCK(vp); 1647 for (i = 1; i < maxcl; i++) 1648 if (vfs_bio_clcheck(vp, size, lblkno + i, 1649 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1650 break; 1651 1652 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1653 if (vfs_bio_clcheck(vp, size, lblkno - j, 1654 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1655 break; 1656 1657 VI_UNLOCK(vp); 1658 --j; 1659 ncl = i + j; 1660 /* 1661 * this is a possible cluster write 1662 */ 1663 if (ncl != 1) { 1664 BUF_UNLOCK(bp); 1665 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1666 return nwritten; 1667 } 1668 } 1669 bremfree(bp); 1670 bp->b_flags |= B_ASYNC; 1671 /* 1672 * default (old) behavior, writing out only one block 1673 * 1674 * XXX returns b_bufsize instead of b_bcount for nwritten? 1675 */ 1676 nwritten = bp->b_bufsize; 1677 (void) bwrite(bp); 1678 1679 return nwritten; 1680} 1681 1682/* 1683 * getnewbuf: 1684 * 1685 * Find and initialize a new buffer header, freeing up existing buffers 1686 * in the bufqueues as necessary. The new buffer is returned locked. 1687 * 1688 * Important: B_INVAL is not set. If the caller wishes to throw the 1689 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1690 * 1691 * We block if: 1692 * We have insufficient buffer headers 1693 * We have insufficient buffer space 1694 * buffer_map is too fragmented ( space reservation fails ) 1695 * If we have to flush dirty buffers ( but we try to avoid this ) 1696 * 1697 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1698 * Instead we ask the buf daemon to do it for us. We attempt to 1699 * avoid piecemeal wakeups of the pageout daemon. 1700 */ 1701 1702static struct buf * 1703getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1704{ 1705 struct buf *bp; 1706 struct buf *nbp; 1707 int defrag = 0; 1708 int nqindex; 1709 static int flushingbufs; 1710 1711 /* 1712 * We can't afford to block since we might be holding a vnode lock, 1713 * which may prevent system daemons from running. We deal with 1714 * low-memory situations by proactively returning memory and running 1715 * async I/O rather then sync I/O. 1716 */ 1717 1718 atomic_add_int(&getnewbufcalls, 1); 1719 atomic_subtract_int(&getnewbufrestarts, 1); 1720restart: 1721 atomic_add_int(&getnewbufrestarts, 1); 1722 1723 /* 1724 * Setup for scan. If we do not have enough free buffers, 1725 * we setup a degenerate case that immediately fails. Note 1726 * that if we are specially marked process, we are allowed to 1727 * dip into our reserves. 1728 * 1729 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1730 * 1731 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1732 * However, there are a number of cases (defragging, reusing, ...) 1733 * where we cannot backup. 1734 */ 1735 mtx_lock(&bqlock); 1736 nqindex = QUEUE_EMPTYKVA; 1737 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1738 1739 if (nbp == NULL) { 1740 /* 1741 * If no EMPTYKVA buffers and we are either 1742 * defragging or reusing, locate a CLEAN buffer 1743 * to free or reuse. If bufspace useage is low 1744 * skip this step so we can allocate a new buffer. 1745 */ 1746 if (defrag || bufspace >= lobufspace) { 1747 nqindex = QUEUE_CLEAN; 1748 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1749 } 1750 1751 /* 1752 * If we could not find or were not allowed to reuse a 1753 * CLEAN buffer, check to see if it is ok to use an EMPTY 1754 * buffer. We can only use an EMPTY buffer if allocating 1755 * its KVA would not otherwise run us out of buffer space. 1756 */ 1757 if (nbp == NULL && defrag == 0 && 1758 bufspace + maxsize < hibufspace) { 1759 nqindex = QUEUE_EMPTY; 1760 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1761 } 1762 } 1763 1764 /* 1765 * Run scan, possibly freeing data and/or kva mappings on the fly 1766 * depending. 1767 */ 1768 1769 while ((bp = nbp) != NULL) { 1770 int qindex = nqindex; 1771 1772 /* 1773 * Calculate next bp ( we can only use it if we do not block 1774 * or do other fancy things ). 1775 */ 1776 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1777 switch(qindex) { 1778 case QUEUE_EMPTY: 1779 nqindex = QUEUE_EMPTYKVA; 1780 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1781 break; 1782 /* FALLTHROUGH */ 1783 case QUEUE_EMPTYKVA: 1784 nqindex = QUEUE_CLEAN; 1785 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1786 break; 1787 /* FALLTHROUGH */ 1788 case QUEUE_CLEAN: 1789 /* 1790 * nbp is NULL. 1791 */ 1792 break; 1793 } 1794 } 1795 /* 1796 * If we are defragging then we need a buffer with 1797 * b_kvasize != 0. XXX this situation should no longer 1798 * occur, if defrag is non-zero the buffer's b_kvasize 1799 * should also be non-zero at this point. XXX 1800 */ 1801 if (defrag && bp->b_kvasize == 0) { 1802 printf("Warning: defrag empty buffer %p\n", bp); 1803 continue; 1804 } 1805 1806 /* 1807 * Start freeing the bp. This is somewhat involved. nbp 1808 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1809 */ 1810 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1811 continue; 1812 if (bp->b_vp) { 1813 BO_LOCK(bp->b_bufobj); 1814 if (bp->b_vflags & BV_BKGRDINPROG) { 1815 BO_UNLOCK(bp->b_bufobj); 1816 BUF_UNLOCK(bp); 1817 continue; 1818 } 1819 BO_UNLOCK(bp->b_bufobj); 1820 } 1821 CTR6(KTR_BUF, 1822 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1823 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1824 bp->b_kvasize, bp->b_bufsize, qindex); 1825 1826 /* 1827 * Sanity Checks 1828 */ 1829 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1830 1831 /* 1832 * Note: we no longer distinguish between VMIO and non-VMIO 1833 * buffers. 1834 */ 1835 1836 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1837 1838 bremfreel(bp); 1839 mtx_unlock(&bqlock); 1840 1841 if (qindex == QUEUE_CLEAN) { 1842 if (bp->b_flags & B_VMIO) { 1843 bp->b_flags &= ~B_ASYNC; 1844 vfs_vmio_release(bp); 1845 } 1846 if (bp->b_vp) 1847 brelvp(bp); 1848 } 1849 1850 /* 1851 * NOTE: nbp is now entirely invalid. We can only restart 1852 * the scan from this point on. 1853 * 1854 * Get the rest of the buffer freed up. b_kva* is still 1855 * valid after this operation. 1856 */ 1857 1858 if (bp->b_rcred != NOCRED) { 1859 crfree(bp->b_rcred); 1860 bp->b_rcred = NOCRED; 1861 } 1862 if (bp->b_wcred != NOCRED) { 1863 crfree(bp->b_wcred); 1864 bp->b_wcred = NOCRED; 1865 } 1866 if (!LIST_EMPTY(&bp->b_dep)) 1867 buf_deallocate(bp); 1868 if (bp->b_vflags & BV_BKGRDINPROG) 1869 panic("losing buffer 3"); 1870 KASSERT(bp->b_vp == NULL, 1871 ("bp: %p still has vnode %p. qindex: %d", 1872 bp, bp->b_vp, qindex)); 1873 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1874 ("bp: %p still on a buffer list. xflags %X", 1875 bp, bp->b_xflags)); 1876 1877 if (bp->b_bufsize) 1878 allocbuf(bp, 0); 1879 1880 bp->b_flags = 0; 1881 bp->b_ioflags = 0; 1882 bp->b_xflags = 0; 1883 bp->b_vflags = 0; 1884 bp->b_vp = NULL; 1885 bp->b_blkno = bp->b_lblkno = 0; 1886 bp->b_offset = NOOFFSET; 1887 bp->b_iodone = 0; 1888 bp->b_error = 0; 1889 bp->b_resid = 0; 1890 bp->b_bcount = 0; 1891 bp->b_npages = 0; 1892 bp->b_dirtyoff = bp->b_dirtyend = 0; 1893 bp->b_bufobj = NULL; 1894 bp->b_pin_count = 0; 1895 bp->b_fsprivate1 = NULL; 1896 bp->b_fsprivate2 = NULL; 1897 bp->b_fsprivate3 = NULL; 1898 1899 LIST_INIT(&bp->b_dep); 1900 1901 /* 1902 * If we are defragging then free the buffer. 1903 */ 1904 if (defrag) { 1905 bp->b_flags |= B_INVAL; 1906 bfreekva(bp); 1907 brelse(bp); 1908 defrag = 0; 1909 goto restart; 1910 } 1911 1912 /* 1913 * Notify any waiters for the buffer lock about 1914 * identity change by freeing the buffer. 1915 */ 1916 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp) > 0) { 1917 bp->b_flags |= B_INVAL; 1918 bfreekva(bp); 1919 brelse(bp); 1920 goto restart; 1921 } 1922 1923 /* 1924 * If we are overcomitted then recover the buffer and its 1925 * KVM space. This occurs in rare situations when multiple 1926 * processes are blocked in getnewbuf() or allocbuf(). 1927 */ 1928 if (bufspace >= hibufspace) 1929 flushingbufs = 1; 1930 if (flushingbufs && bp->b_kvasize != 0) { 1931 bp->b_flags |= B_INVAL; 1932 bfreekva(bp); 1933 brelse(bp); 1934 goto restart; 1935 } 1936 if (bufspace < lobufspace) 1937 flushingbufs = 0; 1938 break; 1939 } 1940 1941 /* 1942 * If we exhausted our list, sleep as appropriate. We may have to 1943 * wakeup various daemons and write out some dirty buffers. 1944 * 1945 * Generally we are sleeping due to insufficient buffer space. 1946 */ 1947 1948 if (bp == NULL) { 1949 int flags; 1950 char *waitmsg; 1951 1952 if (defrag) { 1953 flags = VFS_BIO_NEED_BUFSPACE; 1954 waitmsg = "nbufkv"; 1955 } else if (bufspace >= hibufspace) { 1956 waitmsg = "nbufbs"; 1957 flags = VFS_BIO_NEED_BUFSPACE; 1958 } else { 1959 waitmsg = "newbuf"; 1960 flags = VFS_BIO_NEED_ANY; 1961 } 1962 mtx_lock(&nblock); 1963 needsbuffer |= flags; 1964 mtx_unlock(&nblock); 1965 mtx_unlock(&bqlock); 1966 1967 bd_speedup(); /* heeeelp */ 1968 1969 mtx_lock(&nblock); 1970 while (needsbuffer & flags) { 1971 if (msleep(&needsbuffer, &nblock, 1972 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1973 mtx_unlock(&nblock); 1974 return (NULL); 1975 } 1976 } 1977 mtx_unlock(&nblock); 1978 } else { 1979 /* 1980 * We finally have a valid bp. We aren't quite out of the 1981 * woods, we still have to reserve kva space. In order 1982 * to keep fragmentation sane we only allocate kva in 1983 * BKVASIZE chunks. 1984 */ 1985 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1986 1987 if (maxsize != bp->b_kvasize) { 1988 vm_offset_t addr = 0; 1989 1990 bfreekva(bp); 1991 1992 vm_map_lock(buffer_map); 1993 if (vm_map_findspace(buffer_map, 1994 vm_map_min(buffer_map), maxsize, &addr)) { 1995 /* 1996 * Uh oh. Buffer map is to fragmented. We 1997 * must defragment the map. 1998 */ 1999 atomic_add_int(&bufdefragcnt, 1); 2000 vm_map_unlock(buffer_map); 2001 defrag = 1; 2002 bp->b_flags |= B_INVAL; 2003 brelse(bp); 2004 goto restart; 2005 } 2006 if (addr) { 2007 vm_map_insert(buffer_map, NULL, 0, 2008 addr, addr + maxsize, 2009 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2010 2011 bp->b_kvabase = (caddr_t) addr; 2012 bp->b_kvasize = maxsize; 2013 atomic_add_int(&bufspace, bp->b_kvasize); 2014 atomic_add_int(&bufreusecnt, 1); 2015 } 2016 vm_map_unlock(buffer_map); 2017 } 2018 bp->b_saveaddr = bp->b_kvabase; 2019 bp->b_data = bp->b_saveaddr; 2020 } 2021 return(bp); 2022} 2023 2024/* 2025 * buf_daemon: 2026 * 2027 * buffer flushing daemon. Buffers are normally flushed by the 2028 * update daemon but if it cannot keep up this process starts to 2029 * take the load in an attempt to prevent getnewbuf() from blocking. 2030 */ 2031 2032static struct kproc_desc buf_kp = { 2033 "bufdaemon", 2034 buf_daemon, 2035 &bufdaemonproc 2036}; 2037SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2038 2039static void 2040buf_daemon() 2041{ 2042 2043 /* 2044 * This process needs to be suspended prior to shutdown sync. 2045 */ 2046 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2047 SHUTDOWN_PRI_LAST); 2048 2049 /* 2050 * This process is allowed to take the buffer cache to the limit 2051 */ 2052 curthread->td_pflags |= TDP_NORUNNINGBUF; 2053 mtx_lock(&bdlock); 2054 for (;;) { 2055 bd_request = 0; 2056 mtx_unlock(&bdlock); 2057 2058 kproc_suspend_check(bufdaemonproc); 2059 2060 /* 2061 * Do the flush. Limit the amount of in-transit I/O we 2062 * allow to build up, otherwise we would completely saturate 2063 * the I/O system. Wakeup any waiting processes before we 2064 * normally would so they can run in parallel with our drain. 2065 */ 2066 while (numdirtybuffers > lodirtybuffers) { 2067 int flushed; 2068 2069 flushed = flushbufqueues(QUEUE_DIRTY, 0); 2070 /* The list empty check here is slightly racy */ 2071 if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) { 2072 mtx_lock(&Giant); 2073 flushed += flushbufqueues(QUEUE_DIRTY_GIANT, 0); 2074 mtx_unlock(&Giant); 2075 } 2076 if (flushed == 0) { 2077 /* 2078 * Could not find any buffers without rollback 2079 * dependencies, so just write the first one 2080 * in the hopes of eventually making progress. 2081 */ 2082 flushbufqueues(QUEUE_DIRTY, 1); 2083 if (!TAILQ_EMPTY( 2084 &bufqueues[QUEUE_DIRTY_GIANT])) { 2085 mtx_lock(&Giant); 2086 flushbufqueues(QUEUE_DIRTY_GIANT, 1); 2087 mtx_unlock(&Giant); 2088 } 2089 break; 2090 } 2091 uio_yield(); 2092 } 2093 2094 /* 2095 * Only clear bd_request if we have reached our low water 2096 * mark. The buf_daemon normally waits 1 second and 2097 * then incrementally flushes any dirty buffers that have 2098 * built up, within reason. 2099 * 2100 * If we were unable to hit our low water mark and couldn't 2101 * find any flushable buffers, we sleep half a second. 2102 * Otherwise we loop immediately. 2103 */ 2104 mtx_lock(&bdlock); 2105 if (numdirtybuffers <= lodirtybuffers) { 2106 /* 2107 * We reached our low water mark, reset the 2108 * request and sleep until we are needed again. 2109 * The sleep is just so the suspend code works. 2110 */ 2111 bd_request = 0; 2112 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2113 } else { 2114 /* 2115 * We couldn't find any flushable dirty buffers but 2116 * still have too many dirty buffers, we 2117 * have to sleep and try again. (rare) 2118 */ 2119 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2120 } 2121 } 2122} 2123 2124/* 2125 * flushbufqueues: 2126 * 2127 * Try to flush a buffer in the dirty queue. We must be careful to 2128 * free up B_INVAL buffers instead of write them, which NFS is 2129 * particularly sensitive to. 2130 */ 2131static int flushwithdeps = 0; 2132SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2133 0, "Number of buffers flushed with dependecies that require rollbacks"); 2134 2135static int 2136flushbufqueues(int queue, int flushdeps) 2137{ 2138 struct buf sentinel; 2139 struct vnode *vp; 2140 struct mount *mp; 2141 struct buf *bp; 2142 int hasdeps; 2143 int flushed; 2144 int target; 2145 2146 target = numdirtybuffers - lodirtybuffers; 2147 if (flushdeps && target > 2) 2148 target /= 2; 2149 flushed = 0; 2150 bp = NULL; 2151 mtx_lock(&bqlock); 2152 TAILQ_INSERT_TAIL(&bufqueues[queue], &sentinel, b_freelist); 2153 while (flushed != target) { 2154 bp = TAILQ_FIRST(&bufqueues[queue]); 2155 if (bp == &sentinel) 2156 break; 2157 TAILQ_REMOVE(&bufqueues[queue], bp, b_freelist); 2158 TAILQ_INSERT_TAIL(&bufqueues[queue], bp, b_freelist); 2159 2160 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2161 continue; 2162 if (bp->b_pin_count > 0) { 2163 BUF_UNLOCK(bp); 2164 continue; 2165 } 2166 BO_LOCK(bp->b_bufobj); 2167 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2168 (bp->b_flags & B_DELWRI) == 0) { 2169 BO_UNLOCK(bp->b_bufobj); 2170 BUF_UNLOCK(bp); 2171 continue; 2172 } 2173 BO_UNLOCK(bp->b_bufobj); 2174 if (bp->b_flags & B_INVAL) { 2175 bremfreel(bp); 2176 mtx_unlock(&bqlock); 2177 brelse(bp); 2178 flushed++; 2179 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2180 mtx_lock(&bqlock); 2181 continue; 2182 } 2183 2184 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2185 if (flushdeps == 0) { 2186 BUF_UNLOCK(bp); 2187 continue; 2188 } 2189 hasdeps = 1; 2190 } else 2191 hasdeps = 0; 2192 /* 2193 * We must hold the lock on a vnode before writing 2194 * one of its buffers. Otherwise we may confuse, or 2195 * in the case of a snapshot vnode, deadlock the 2196 * system. 2197 * 2198 * The lock order here is the reverse of the normal 2199 * of vnode followed by buf lock. This is ok because 2200 * the NOWAIT will prevent deadlock. 2201 */ 2202 vp = bp->b_vp; 2203 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2204 BUF_UNLOCK(bp); 2205 continue; 2206 } 2207 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 2208 mtx_unlock(&bqlock); 2209 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2210 bp, bp->b_vp, bp->b_flags); 2211 vfs_bio_awrite(bp); 2212 vn_finished_write(mp); 2213 VOP_UNLOCK(vp, 0); 2214 flushwithdeps += hasdeps; 2215 flushed++; 2216 waitrunningbufspace(); 2217 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2218 mtx_lock(&bqlock); 2219 continue; 2220 } 2221 vn_finished_write(mp); 2222 BUF_UNLOCK(bp); 2223 } 2224 TAILQ_REMOVE(&bufqueues[queue], &sentinel, b_freelist); 2225 mtx_unlock(&bqlock); 2226 return (flushed); 2227} 2228 2229/* 2230 * Check to see if a block is currently memory resident. 2231 */ 2232struct buf * 2233incore(struct bufobj *bo, daddr_t blkno) 2234{ 2235 struct buf *bp; 2236 2237 BO_LOCK(bo); 2238 bp = gbincore(bo, blkno); 2239 BO_UNLOCK(bo); 2240 return (bp); 2241} 2242 2243/* 2244 * Returns true if no I/O is needed to access the 2245 * associated VM object. This is like incore except 2246 * it also hunts around in the VM system for the data. 2247 */ 2248 2249static int 2250inmem(struct vnode * vp, daddr_t blkno) 2251{ 2252 vm_object_t obj; 2253 vm_offset_t toff, tinc, size; 2254 vm_page_t m; 2255 vm_ooffset_t off; 2256 2257 ASSERT_VOP_LOCKED(vp, "inmem"); 2258 2259 if (incore(&vp->v_bufobj, blkno)) 2260 return 1; 2261 if (vp->v_mount == NULL) 2262 return 0; 2263 obj = vp->v_object; 2264 if (obj == NULL) 2265 return (0); 2266 2267 size = PAGE_SIZE; 2268 if (size > vp->v_mount->mnt_stat.f_iosize) 2269 size = vp->v_mount->mnt_stat.f_iosize; 2270 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2271 2272 VM_OBJECT_LOCK(obj); 2273 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2274 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2275 if (!m) 2276 goto notinmem; 2277 tinc = size; 2278 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2279 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2280 if (vm_page_is_valid(m, 2281 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2282 goto notinmem; 2283 } 2284 VM_OBJECT_UNLOCK(obj); 2285 return 1; 2286 2287notinmem: 2288 VM_OBJECT_UNLOCK(obj); 2289 return (0); 2290} 2291 2292/* 2293 * vfs_setdirty: 2294 * 2295 * Sets the dirty range for a buffer based on the status of the dirty 2296 * bits in the pages comprising the buffer. 2297 * 2298 * The range is limited to the size of the buffer. 2299 * 2300 * This routine is primarily used by NFS, but is generalized for the 2301 * B_VMIO case. 2302 */ 2303static void 2304vfs_setdirty(struct buf *bp) 2305{ 2306 2307 /* 2308 * Degenerate case - empty buffer 2309 */ 2310 2311 if (bp->b_bufsize == 0) 2312 return; 2313 2314 /* 2315 * We qualify the scan for modified pages on whether the 2316 * object has been flushed yet. 2317 */ 2318 2319 if ((bp->b_flags & B_VMIO) == 0) 2320 return; 2321 2322 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2323 vfs_setdirty_locked_object(bp); 2324 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2325} 2326 2327static void 2328vfs_setdirty_locked_object(struct buf *bp) 2329{ 2330 vm_object_t object; 2331 int i; 2332 2333 object = bp->b_bufobj->bo_object; 2334 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2335 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2336 vm_offset_t boffset; 2337 vm_offset_t eoffset; 2338 2339 vm_page_lock_queues(); 2340 /* 2341 * test the pages to see if they have been modified directly 2342 * by users through the VM system. 2343 */ 2344 for (i = 0; i < bp->b_npages; i++) 2345 vm_page_test_dirty(bp->b_pages[i]); 2346 2347 /* 2348 * Calculate the encompassing dirty range, boffset and eoffset, 2349 * (eoffset - boffset) bytes. 2350 */ 2351 2352 for (i = 0; i < bp->b_npages; i++) { 2353 if (bp->b_pages[i]->dirty) 2354 break; 2355 } 2356 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2357 2358 for (i = bp->b_npages - 1; i >= 0; --i) { 2359 if (bp->b_pages[i]->dirty) { 2360 break; 2361 } 2362 } 2363 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2364 2365 vm_page_unlock_queues(); 2366 /* 2367 * Fit it to the buffer. 2368 */ 2369 2370 if (eoffset > bp->b_bcount) 2371 eoffset = bp->b_bcount; 2372 2373 /* 2374 * If we have a good dirty range, merge with the existing 2375 * dirty range. 2376 */ 2377 2378 if (boffset < eoffset) { 2379 if (bp->b_dirtyoff > boffset) 2380 bp->b_dirtyoff = boffset; 2381 if (bp->b_dirtyend < eoffset) 2382 bp->b_dirtyend = eoffset; 2383 } 2384 } 2385} 2386 2387/* 2388 * getblk: 2389 * 2390 * Get a block given a specified block and offset into a file/device. 2391 * The buffers B_DONE bit will be cleared on return, making it almost 2392 * ready for an I/O initiation. B_INVAL may or may not be set on 2393 * return. The caller should clear B_INVAL prior to initiating a 2394 * READ. 2395 * 2396 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2397 * an existing buffer. 2398 * 2399 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2400 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2401 * and then cleared based on the backing VM. If the previous buffer is 2402 * non-0-sized but invalid, B_CACHE will be cleared. 2403 * 2404 * If getblk() must create a new buffer, the new buffer is returned with 2405 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2406 * case it is returned with B_INVAL clear and B_CACHE set based on the 2407 * backing VM. 2408 * 2409 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2410 * B_CACHE bit is clear. 2411 * 2412 * What this means, basically, is that the caller should use B_CACHE to 2413 * determine whether the buffer is fully valid or not and should clear 2414 * B_INVAL prior to issuing a read. If the caller intends to validate 2415 * the buffer by loading its data area with something, the caller needs 2416 * to clear B_INVAL. If the caller does this without issuing an I/O, 2417 * the caller should set B_CACHE ( as an optimization ), else the caller 2418 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2419 * a write attempt or if it was a successfull read. If the caller 2420 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2421 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2422 */ 2423struct buf * 2424getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2425 int flags) 2426{ 2427 struct buf *bp; 2428 struct bufobj *bo; 2429 int error; 2430 2431 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2432 ASSERT_VOP_LOCKED(vp, "getblk"); 2433 if (size > MAXBSIZE) 2434 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2435 2436 bo = &vp->v_bufobj; 2437loop: 2438 /* 2439 * Block if we are low on buffers. Certain processes are allowed 2440 * to completely exhaust the buffer cache. 2441 * 2442 * If this check ever becomes a bottleneck it may be better to 2443 * move it into the else, when gbincore() fails. At the moment 2444 * it isn't a problem. 2445 * 2446 * XXX remove if 0 sections (clean this up after its proven) 2447 */ 2448 if (numfreebuffers == 0) { 2449 if (TD_IS_IDLETHREAD(curthread)) 2450 return NULL; 2451 mtx_lock(&nblock); 2452 needsbuffer |= VFS_BIO_NEED_ANY; 2453 mtx_unlock(&nblock); 2454 } 2455 2456 BO_LOCK(bo); 2457 bp = gbincore(bo, blkno); 2458 if (bp != NULL) { 2459 int lockflags; 2460 /* 2461 * Buffer is in-core. If the buffer is not busy, it must 2462 * be on a queue. 2463 */ 2464 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2465 2466 if (flags & GB_LOCK_NOWAIT) 2467 lockflags |= LK_NOWAIT; 2468 2469 error = BUF_TIMELOCK(bp, lockflags, 2470 VI_MTX(vp), "getblk", slpflag, slptimeo); 2471 2472 /* 2473 * If we slept and got the lock we have to restart in case 2474 * the buffer changed identities. 2475 */ 2476 if (error == ENOLCK) 2477 goto loop; 2478 /* We timed out or were interrupted. */ 2479 else if (error) 2480 return (NULL); 2481 2482 /* 2483 * The buffer is locked. B_CACHE is cleared if the buffer is 2484 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2485 * and for a VMIO buffer B_CACHE is adjusted according to the 2486 * backing VM cache. 2487 */ 2488 if (bp->b_flags & B_INVAL) 2489 bp->b_flags &= ~B_CACHE; 2490 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2491 bp->b_flags |= B_CACHE; 2492 bremfree(bp); 2493 2494 /* 2495 * check for size inconsistancies for non-VMIO case. 2496 */ 2497 2498 if (bp->b_bcount != size) { 2499 if ((bp->b_flags & B_VMIO) == 0 || 2500 (size > bp->b_kvasize)) { 2501 if (bp->b_flags & B_DELWRI) { 2502 /* 2503 * If buffer is pinned and caller does 2504 * not want sleep waiting for it to be 2505 * unpinned, bail out 2506 * */ 2507 if (bp->b_pin_count > 0) { 2508 if (flags & GB_LOCK_NOWAIT) { 2509 bqrelse(bp); 2510 return (NULL); 2511 } else { 2512 bunpin_wait(bp); 2513 } 2514 } 2515 bp->b_flags |= B_NOCACHE; 2516 bwrite(bp); 2517 } else { 2518 if (LIST_EMPTY(&bp->b_dep)) { 2519 bp->b_flags |= B_RELBUF; 2520 brelse(bp); 2521 } else { 2522 bp->b_flags |= B_NOCACHE; 2523 bwrite(bp); 2524 } 2525 } 2526 goto loop; 2527 } 2528 } 2529 2530 /* 2531 * If the size is inconsistant in the VMIO case, we can resize 2532 * the buffer. This might lead to B_CACHE getting set or 2533 * cleared. If the size has not changed, B_CACHE remains 2534 * unchanged from its previous state. 2535 */ 2536 2537 if (bp->b_bcount != size) 2538 allocbuf(bp, size); 2539 2540 KASSERT(bp->b_offset != NOOFFSET, 2541 ("getblk: no buffer offset")); 2542 2543 /* 2544 * A buffer with B_DELWRI set and B_CACHE clear must 2545 * be committed before we can return the buffer in 2546 * order to prevent the caller from issuing a read 2547 * ( due to B_CACHE not being set ) and overwriting 2548 * it. 2549 * 2550 * Most callers, including NFS and FFS, need this to 2551 * operate properly either because they assume they 2552 * can issue a read if B_CACHE is not set, or because 2553 * ( for example ) an uncached B_DELWRI might loop due 2554 * to softupdates re-dirtying the buffer. In the latter 2555 * case, B_CACHE is set after the first write completes, 2556 * preventing further loops. 2557 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2558 * above while extending the buffer, we cannot allow the 2559 * buffer to remain with B_CACHE set after the write 2560 * completes or it will represent a corrupt state. To 2561 * deal with this we set B_NOCACHE to scrap the buffer 2562 * after the write. 2563 * 2564 * We might be able to do something fancy, like setting 2565 * B_CACHE in bwrite() except if B_DELWRI is already set, 2566 * so the below call doesn't set B_CACHE, but that gets real 2567 * confusing. This is much easier. 2568 */ 2569 2570 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2571 bp->b_flags |= B_NOCACHE; 2572 bwrite(bp); 2573 goto loop; 2574 } 2575 bp->b_flags &= ~B_DONE; 2576 } else { 2577 int bsize, maxsize, vmio; 2578 off_t offset; 2579 2580 /* 2581 * Buffer is not in-core, create new buffer. The buffer 2582 * returned by getnewbuf() is locked. Note that the returned 2583 * buffer is also considered valid (not marked B_INVAL). 2584 */ 2585 BO_UNLOCK(bo); 2586 /* 2587 * If the user does not want us to create the buffer, bail out 2588 * here. 2589 */ 2590 if (flags & GB_NOCREAT) 2591 return NULL; 2592 bsize = bo->bo_bsize; 2593 offset = blkno * bsize; 2594 vmio = vp->v_object != NULL; 2595 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2596 maxsize = imax(maxsize, bsize); 2597 2598 bp = getnewbuf(slpflag, slptimeo, size, maxsize); 2599 if (bp == NULL) { 2600 if (slpflag || slptimeo) 2601 return NULL; 2602 goto loop; 2603 } 2604 2605 /* 2606 * This code is used to make sure that a buffer is not 2607 * created while the getnewbuf routine is blocked. 2608 * This can be a problem whether the vnode is locked or not. 2609 * If the buffer is created out from under us, we have to 2610 * throw away the one we just created. 2611 * 2612 * Note: this must occur before we associate the buffer 2613 * with the vp especially considering limitations in 2614 * the splay tree implementation when dealing with duplicate 2615 * lblkno's. 2616 */ 2617 BO_LOCK(bo); 2618 if (gbincore(bo, blkno)) { 2619 BO_UNLOCK(bo); 2620 bp->b_flags |= B_INVAL; 2621 brelse(bp); 2622 goto loop; 2623 } 2624 2625 /* 2626 * Insert the buffer into the hash, so that it can 2627 * be found by incore. 2628 */ 2629 bp->b_blkno = bp->b_lblkno = blkno; 2630 bp->b_offset = offset; 2631 bgetvp(vp, bp); 2632 BO_UNLOCK(bo); 2633 2634 /* 2635 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2636 * buffer size starts out as 0, B_CACHE will be set by 2637 * allocbuf() for the VMIO case prior to it testing the 2638 * backing store for validity. 2639 */ 2640 2641 if (vmio) { 2642 bp->b_flags |= B_VMIO; 2643#if defined(VFS_BIO_DEBUG) 2644 if (vn_canvmio(vp) != TRUE) 2645 printf("getblk: VMIO on vnode type %d\n", 2646 vp->v_type); 2647#endif 2648 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2649 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2650 bp, vp->v_object, bp->b_bufobj->bo_object)); 2651 } else { 2652 bp->b_flags &= ~B_VMIO; 2653 KASSERT(bp->b_bufobj->bo_object == NULL, 2654 ("ARGH! has b_bufobj->bo_object %p %p\n", 2655 bp, bp->b_bufobj->bo_object)); 2656 } 2657 2658 allocbuf(bp, size); 2659 bp->b_flags &= ~B_DONE; 2660 } 2661 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2662 BUF_ASSERT_HELD(bp); 2663 KASSERT(bp->b_bufobj == bo, 2664 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2665 return (bp); 2666} 2667 2668/* 2669 * Get an empty, disassociated buffer of given size. The buffer is initially 2670 * set to B_INVAL. 2671 */ 2672struct buf * 2673geteblk(int size) 2674{ 2675 struct buf *bp; 2676 int maxsize; 2677 2678 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2679 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2680 continue; 2681 allocbuf(bp, size); 2682 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2683 BUF_ASSERT_HELD(bp); 2684 return (bp); 2685} 2686 2687 2688/* 2689 * This code constitutes the buffer memory from either anonymous system 2690 * memory (in the case of non-VMIO operations) or from an associated 2691 * VM object (in the case of VMIO operations). This code is able to 2692 * resize a buffer up or down. 2693 * 2694 * Note that this code is tricky, and has many complications to resolve 2695 * deadlock or inconsistant data situations. Tread lightly!!! 2696 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2697 * the caller. Calling this code willy nilly can result in the loss of data. 2698 * 2699 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2700 * B_CACHE for the non-VMIO case. 2701 */ 2702 2703int 2704allocbuf(struct buf *bp, int size) 2705{ 2706 int newbsize, mbsize; 2707 int i; 2708 2709 BUF_ASSERT_HELD(bp); 2710 2711 if (bp->b_kvasize < size) 2712 panic("allocbuf: buffer too small"); 2713 2714 if ((bp->b_flags & B_VMIO) == 0) { 2715 caddr_t origbuf; 2716 int origbufsize; 2717 /* 2718 * Just get anonymous memory from the kernel. Don't 2719 * mess with B_CACHE. 2720 */ 2721 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2722 if (bp->b_flags & B_MALLOC) 2723 newbsize = mbsize; 2724 else 2725 newbsize = round_page(size); 2726 2727 if (newbsize < bp->b_bufsize) { 2728 /* 2729 * malloced buffers are not shrunk 2730 */ 2731 if (bp->b_flags & B_MALLOC) { 2732 if (newbsize) { 2733 bp->b_bcount = size; 2734 } else { 2735 free(bp->b_data, M_BIOBUF); 2736 if (bp->b_bufsize) { 2737 atomic_subtract_int( 2738 &bufmallocspace, 2739 bp->b_bufsize); 2740 bufspacewakeup(); 2741 bp->b_bufsize = 0; 2742 } 2743 bp->b_saveaddr = bp->b_kvabase; 2744 bp->b_data = bp->b_saveaddr; 2745 bp->b_bcount = 0; 2746 bp->b_flags &= ~B_MALLOC; 2747 } 2748 return 1; 2749 } 2750 vm_hold_free_pages( 2751 bp, 2752 (vm_offset_t) bp->b_data + newbsize, 2753 (vm_offset_t) bp->b_data + bp->b_bufsize); 2754 } else if (newbsize > bp->b_bufsize) { 2755 /* 2756 * We only use malloced memory on the first allocation. 2757 * and revert to page-allocated memory when the buffer 2758 * grows. 2759 */ 2760 /* 2761 * There is a potential smp race here that could lead 2762 * to bufmallocspace slightly passing the max. It 2763 * is probably extremely rare and not worth worrying 2764 * over. 2765 */ 2766 if ( (bufmallocspace < maxbufmallocspace) && 2767 (bp->b_bufsize == 0) && 2768 (mbsize <= PAGE_SIZE/2)) { 2769 2770 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2771 bp->b_bufsize = mbsize; 2772 bp->b_bcount = size; 2773 bp->b_flags |= B_MALLOC; 2774 atomic_add_int(&bufmallocspace, mbsize); 2775 return 1; 2776 } 2777 origbuf = NULL; 2778 origbufsize = 0; 2779 /* 2780 * If the buffer is growing on its other-than-first allocation, 2781 * then we revert to the page-allocation scheme. 2782 */ 2783 if (bp->b_flags & B_MALLOC) { 2784 origbuf = bp->b_data; 2785 origbufsize = bp->b_bufsize; 2786 bp->b_data = bp->b_kvabase; 2787 if (bp->b_bufsize) { 2788 atomic_subtract_int(&bufmallocspace, 2789 bp->b_bufsize); 2790 bufspacewakeup(); 2791 bp->b_bufsize = 0; 2792 } 2793 bp->b_flags &= ~B_MALLOC; 2794 newbsize = round_page(newbsize); 2795 } 2796 vm_hold_load_pages( 2797 bp, 2798 (vm_offset_t) bp->b_data + bp->b_bufsize, 2799 (vm_offset_t) bp->b_data + newbsize); 2800 if (origbuf) { 2801 bcopy(origbuf, bp->b_data, origbufsize); 2802 free(origbuf, M_BIOBUF); 2803 } 2804 } 2805 } else { 2806 int desiredpages; 2807 2808 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2809 desiredpages = (size == 0) ? 0 : 2810 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2811 2812 if (bp->b_flags & B_MALLOC) 2813 panic("allocbuf: VMIO buffer can't be malloced"); 2814 /* 2815 * Set B_CACHE initially if buffer is 0 length or will become 2816 * 0-length. 2817 */ 2818 if (size == 0 || bp->b_bufsize == 0) 2819 bp->b_flags |= B_CACHE; 2820 2821 if (newbsize < bp->b_bufsize) { 2822 /* 2823 * DEV_BSIZE aligned new buffer size is less then the 2824 * DEV_BSIZE aligned existing buffer size. Figure out 2825 * if we have to remove any pages. 2826 */ 2827 if (desiredpages < bp->b_npages) { 2828 vm_page_t m; 2829 2830 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2831 vm_page_lock_queues(); 2832 for (i = desiredpages; i < bp->b_npages; i++) { 2833 /* 2834 * the page is not freed here -- it 2835 * is the responsibility of 2836 * vnode_pager_setsize 2837 */ 2838 m = bp->b_pages[i]; 2839 KASSERT(m != bogus_page, 2840 ("allocbuf: bogus page found")); 2841 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2842 vm_page_lock_queues(); 2843 2844 bp->b_pages[i] = NULL; 2845 vm_page_unwire(m, 0); 2846 } 2847 vm_page_unlock_queues(); 2848 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2849 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2850 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2851 bp->b_npages = desiredpages; 2852 } 2853 } else if (size > bp->b_bcount) { 2854 /* 2855 * We are growing the buffer, possibly in a 2856 * byte-granular fashion. 2857 */ 2858 struct vnode *vp; 2859 vm_object_t obj; 2860 vm_offset_t toff; 2861 vm_offset_t tinc; 2862 2863 /* 2864 * Step 1, bring in the VM pages from the object, 2865 * allocating them if necessary. We must clear 2866 * B_CACHE if these pages are not valid for the 2867 * range covered by the buffer. 2868 */ 2869 2870 vp = bp->b_vp; 2871 obj = bp->b_bufobj->bo_object; 2872 2873 VM_OBJECT_LOCK(obj); 2874 while (bp->b_npages < desiredpages) { 2875 vm_page_t m; 2876 vm_pindex_t pi; 2877 2878 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2879 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2880 /* 2881 * note: must allocate system pages 2882 * since blocking here could intefere 2883 * with paging I/O, no matter which 2884 * process we are. 2885 */ 2886 m = vm_page_alloc(obj, pi, 2887 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2888 VM_ALLOC_WIRED); 2889 if (m == NULL) { 2890 atomic_add_int(&vm_pageout_deficit, 2891 desiredpages - bp->b_npages); 2892 VM_OBJECT_UNLOCK(obj); 2893 VM_WAIT; 2894 VM_OBJECT_LOCK(obj); 2895 } else { 2896 if (m->valid == 0) 2897 bp->b_flags &= ~B_CACHE; 2898 bp->b_pages[bp->b_npages] = m; 2899 ++bp->b_npages; 2900 } 2901 continue; 2902 } 2903 2904 /* 2905 * We found a page. If we have to sleep on it, 2906 * retry because it might have gotten freed out 2907 * from under us. 2908 * 2909 * We can only test VPO_BUSY here. Blocking on 2910 * m->busy might lead to a deadlock: 2911 * 2912 * vm_fault->getpages->cluster_read->allocbuf 2913 * 2914 */ 2915 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2916 continue; 2917 2918 /* 2919 * We have a good page. 2920 */ 2921 vm_page_lock_queues(); 2922 vm_page_wire(m); 2923 vm_page_unlock_queues(); 2924 bp->b_pages[bp->b_npages] = m; 2925 ++bp->b_npages; 2926 } 2927 2928 /* 2929 * Step 2. We've loaded the pages into the buffer, 2930 * we have to figure out if we can still have B_CACHE 2931 * set. Note that B_CACHE is set according to the 2932 * byte-granular range ( bcount and size ), new the 2933 * aligned range ( newbsize ). 2934 * 2935 * The VM test is against m->valid, which is DEV_BSIZE 2936 * aligned. Needless to say, the validity of the data 2937 * needs to also be DEV_BSIZE aligned. Note that this 2938 * fails with NFS if the server or some other client 2939 * extends the file's EOF. If our buffer is resized, 2940 * B_CACHE may remain set! XXX 2941 */ 2942 2943 toff = bp->b_bcount; 2944 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2945 2946 while ((bp->b_flags & B_CACHE) && toff < size) { 2947 vm_pindex_t pi; 2948 2949 if (tinc > (size - toff)) 2950 tinc = size - toff; 2951 2952 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2953 PAGE_SHIFT; 2954 2955 vfs_buf_test_cache( 2956 bp, 2957 bp->b_offset, 2958 toff, 2959 tinc, 2960 bp->b_pages[pi] 2961 ); 2962 toff += tinc; 2963 tinc = PAGE_SIZE; 2964 } 2965 VM_OBJECT_UNLOCK(obj); 2966 2967 /* 2968 * Step 3, fixup the KVM pmap. Remember that 2969 * bp->b_data is relative to bp->b_offset, but 2970 * bp->b_offset may be offset into the first page. 2971 */ 2972 2973 bp->b_data = (caddr_t) 2974 trunc_page((vm_offset_t)bp->b_data); 2975 pmap_qenter( 2976 (vm_offset_t)bp->b_data, 2977 bp->b_pages, 2978 bp->b_npages 2979 ); 2980 2981 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2982 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2983 } 2984 } 2985 if (newbsize < bp->b_bufsize) 2986 bufspacewakeup(); 2987 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2988 bp->b_bcount = size; /* requested buffer size */ 2989 return 1; 2990} 2991 2992void 2993biodone(struct bio *bp) 2994{ 2995 void (*done)(struct bio *); 2996 2997 mtx_lock(&bdonelock); 2998 bp->bio_flags |= BIO_DONE; 2999 done = bp->bio_done; 3000 if (done == NULL) 3001 wakeup(bp); 3002 mtx_unlock(&bdonelock); 3003 if (done != NULL) 3004 done(bp); 3005} 3006 3007/* 3008 * Wait for a BIO to finish. 3009 * 3010 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3011 * case is not yet clear. 3012 */ 3013int 3014biowait(struct bio *bp, const char *wchan) 3015{ 3016 3017 mtx_lock(&bdonelock); 3018 while ((bp->bio_flags & BIO_DONE) == 0) 3019 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3020 mtx_unlock(&bdonelock); 3021 if (bp->bio_error != 0) 3022 return (bp->bio_error); 3023 if (!(bp->bio_flags & BIO_ERROR)) 3024 return (0); 3025 return (EIO); 3026} 3027 3028void 3029biofinish(struct bio *bp, struct devstat *stat, int error) 3030{ 3031 3032 if (error) { 3033 bp->bio_error = error; 3034 bp->bio_flags |= BIO_ERROR; 3035 } 3036 if (stat != NULL) 3037 devstat_end_transaction_bio(stat, bp); 3038 biodone(bp); 3039} 3040 3041/* 3042 * bufwait: 3043 * 3044 * Wait for buffer I/O completion, returning error status. The buffer 3045 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3046 * error and cleared. 3047 */ 3048int 3049bufwait(struct buf *bp) 3050{ 3051 if (bp->b_iocmd == BIO_READ) 3052 bwait(bp, PRIBIO, "biord"); 3053 else 3054 bwait(bp, PRIBIO, "biowr"); 3055 if (bp->b_flags & B_EINTR) { 3056 bp->b_flags &= ~B_EINTR; 3057 return (EINTR); 3058 } 3059 if (bp->b_ioflags & BIO_ERROR) { 3060 return (bp->b_error ? bp->b_error : EIO); 3061 } else { 3062 return (0); 3063 } 3064} 3065 3066 /* 3067 * Call back function from struct bio back up to struct buf. 3068 */ 3069static void 3070bufdonebio(struct bio *bip) 3071{ 3072 struct buf *bp; 3073 3074 bp = bip->bio_caller2; 3075 bp->b_resid = bp->b_bcount - bip->bio_completed; 3076 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3077 bp->b_ioflags = bip->bio_flags; 3078 bp->b_error = bip->bio_error; 3079 if (bp->b_error) 3080 bp->b_ioflags |= BIO_ERROR; 3081 bufdone(bp); 3082 g_destroy_bio(bip); 3083} 3084 3085void 3086dev_strategy(struct cdev *dev, struct buf *bp) 3087{ 3088 struct cdevsw *csw; 3089 struct bio *bip; 3090 3091 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3092 panic("b_iocmd botch"); 3093 for (;;) { 3094 bip = g_new_bio(); 3095 if (bip != NULL) 3096 break; 3097 /* Try again later */ 3098 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3099 } 3100 bip->bio_cmd = bp->b_iocmd; 3101 bip->bio_offset = bp->b_iooffset; 3102 bip->bio_length = bp->b_bcount; 3103 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3104 bip->bio_data = bp->b_data; 3105 bip->bio_done = bufdonebio; 3106 bip->bio_caller2 = bp; 3107 bip->bio_dev = dev; 3108 KASSERT(dev->si_refcount > 0, 3109 ("dev_strategy on un-referenced struct cdev *(%s)", 3110 devtoname(dev))); 3111 csw = dev_refthread(dev); 3112 if (csw == NULL) { 3113 g_destroy_bio(bip); 3114 bp->b_error = ENXIO; 3115 bp->b_ioflags = BIO_ERROR; 3116 bufdone(bp); 3117 return; 3118 } 3119 (*csw->d_strategy)(bip); 3120 dev_relthread(dev); 3121} 3122 3123/* 3124 * bufdone: 3125 * 3126 * Finish I/O on a buffer, optionally calling a completion function. 3127 * This is usually called from an interrupt so process blocking is 3128 * not allowed. 3129 * 3130 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3131 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3132 * assuming B_INVAL is clear. 3133 * 3134 * For the VMIO case, we set B_CACHE if the op was a read and no 3135 * read error occured, or if the op was a write. B_CACHE is never 3136 * set if the buffer is invalid or otherwise uncacheable. 3137 * 3138 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3139 * initiator to leave B_INVAL set to brelse the buffer out of existance 3140 * in the biodone routine. 3141 */ 3142void 3143bufdone(struct buf *bp) 3144{ 3145 struct bufobj *dropobj; 3146 void (*biodone)(struct buf *); 3147 3148 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3149 dropobj = NULL; 3150 3151 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3152 BUF_ASSERT_HELD(bp); 3153 3154 runningbufwakeup(bp); 3155 if (bp->b_iocmd == BIO_WRITE) 3156 dropobj = bp->b_bufobj; 3157 /* call optional completion function if requested */ 3158 if (bp->b_iodone != NULL) { 3159 biodone = bp->b_iodone; 3160 bp->b_iodone = NULL; 3161 (*biodone) (bp); 3162 if (dropobj) 3163 bufobj_wdrop(dropobj); 3164 return; 3165 } 3166 3167 bufdone_finish(bp); 3168 3169 if (dropobj) 3170 bufobj_wdrop(dropobj); 3171} 3172 3173void 3174bufdone_finish(struct buf *bp) 3175{ 3176 BUF_ASSERT_HELD(bp); 3177 3178 if (!LIST_EMPTY(&bp->b_dep)) 3179 buf_complete(bp); 3180 3181 if (bp->b_flags & B_VMIO) { 3182 int i; 3183 vm_ooffset_t foff; 3184 vm_page_t m; 3185 vm_object_t obj; 3186 int iosize; 3187 struct vnode *vp = bp->b_vp; 3188 boolean_t are_queues_locked; 3189 3190 obj = bp->b_bufobj->bo_object; 3191 3192#if defined(VFS_BIO_DEBUG) 3193 mp_fixme("usecount and vflag accessed without locks."); 3194 if (vp->v_usecount == 0) { 3195 panic("biodone: zero vnode ref count"); 3196 } 3197 3198 KASSERT(vp->v_object != NULL, 3199 ("biodone: vnode %p has no vm_object", vp)); 3200#endif 3201 3202 foff = bp->b_offset; 3203 KASSERT(bp->b_offset != NOOFFSET, 3204 ("biodone: no buffer offset")); 3205 3206 VM_OBJECT_LOCK(obj); 3207#if defined(VFS_BIO_DEBUG) 3208 if (obj->paging_in_progress < bp->b_npages) { 3209 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3210 obj->paging_in_progress, bp->b_npages); 3211 } 3212#endif 3213 3214 /* 3215 * Set B_CACHE if the op was a normal read and no error 3216 * occured. B_CACHE is set for writes in the b*write() 3217 * routines. 3218 */ 3219 iosize = bp->b_bcount - bp->b_resid; 3220 if (bp->b_iocmd == BIO_READ && 3221 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3222 !(bp->b_ioflags & BIO_ERROR)) { 3223 bp->b_flags |= B_CACHE; 3224 } 3225 if (bp->b_iocmd == BIO_READ) { 3226 vm_page_lock_queues(); 3227 are_queues_locked = TRUE; 3228 } else 3229 are_queues_locked = FALSE; 3230 for (i = 0; i < bp->b_npages; i++) { 3231 int bogusflag = 0; 3232 int resid; 3233 3234 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3235 if (resid > iosize) 3236 resid = iosize; 3237 3238 /* 3239 * cleanup bogus pages, restoring the originals 3240 */ 3241 m = bp->b_pages[i]; 3242 if (m == bogus_page) { 3243 bogusflag = 1; 3244 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3245 if (m == NULL) 3246 panic("biodone: page disappeared!"); 3247 bp->b_pages[i] = m; 3248 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3249 bp->b_pages, bp->b_npages); 3250 } 3251#if defined(VFS_BIO_DEBUG) 3252 if (OFF_TO_IDX(foff) != m->pindex) { 3253 printf( 3254"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3255 (intmax_t)foff, (uintmax_t)m->pindex); 3256 } 3257#endif 3258 3259 /* 3260 * In the write case, the valid and clean bits are 3261 * already changed correctly ( see bdwrite() ), so we 3262 * only need to do this here in the read case. 3263 */ 3264 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3265 vfs_page_set_valid(bp, foff, m); 3266 } 3267 3268 /* 3269 * when debugging new filesystems or buffer I/O methods, this 3270 * is the most common error that pops up. if you see this, you 3271 * have not set the page busy flag correctly!!! 3272 */ 3273 if (m->busy == 0) { 3274 printf("biodone: page busy < 0, " 3275 "pindex: %d, foff: 0x(%x,%x), " 3276 "resid: %d, index: %d\n", 3277 (int) m->pindex, (int)(foff >> 32), 3278 (int) foff & 0xffffffff, resid, i); 3279 if (!vn_isdisk(vp, NULL)) 3280 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3281 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3282 (intmax_t) bp->b_lblkno, 3283 bp->b_flags, bp->b_npages); 3284 else 3285 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3286 (intmax_t) bp->b_lblkno, 3287 bp->b_flags, bp->b_npages); 3288 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3289 (u_long)m->valid, (u_long)m->dirty, 3290 m->wire_count); 3291 panic("biodone: page busy < 0\n"); 3292 } 3293 vm_page_io_finish(m); 3294 vm_object_pip_subtract(obj, 1); 3295 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3296 iosize -= resid; 3297 } 3298 if (are_queues_locked) 3299 vm_page_unlock_queues(); 3300 vm_object_pip_wakeupn(obj, 0); 3301 VM_OBJECT_UNLOCK(obj); 3302 } 3303 3304 /* 3305 * For asynchronous completions, release the buffer now. The brelse 3306 * will do a wakeup there if necessary - so no need to do a wakeup 3307 * here in the async case. The sync case always needs to do a wakeup. 3308 */ 3309 3310 if (bp->b_flags & B_ASYNC) { 3311 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3312 brelse(bp); 3313 else 3314 bqrelse(bp); 3315 } else 3316 bdone(bp); 3317} 3318 3319/* 3320 * This routine is called in lieu of iodone in the case of 3321 * incomplete I/O. This keeps the busy status for pages 3322 * consistant. 3323 */ 3324void 3325vfs_unbusy_pages(struct buf *bp) 3326{ 3327 int i; 3328 vm_object_t obj; 3329 vm_page_t m; 3330 3331 runningbufwakeup(bp); 3332 if (!(bp->b_flags & B_VMIO)) 3333 return; 3334 3335 obj = bp->b_bufobj->bo_object; 3336 VM_OBJECT_LOCK(obj); 3337 for (i = 0; i < bp->b_npages; i++) { 3338 m = bp->b_pages[i]; 3339 if (m == bogus_page) { 3340 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3341 if (!m) 3342 panic("vfs_unbusy_pages: page missing\n"); 3343 bp->b_pages[i] = m; 3344 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3345 bp->b_pages, bp->b_npages); 3346 } 3347 vm_object_pip_subtract(obj, 1); 3348 vm_page_io_finish(m); 3349 } 3350 vm_object_pip_wakeupn(obj, 0); 3351 VM_OBJECT_UNLOCK(obj); 3352} 3353 3354/* 3355 * vfs_page_set_valid: 3356 * 3357 * Set the valid bits in a page based on the supplied offset. The 3358 * range is restricted to the buffer's size. 3359 * 3360 * This routine is typically called after a read completes. 3361 */ 3362static void 3363vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3364{ 3365 vm_ooffset_t soff, eoff; 3366 3367 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3368 /* 3369 * Start and end offsets in buffer. eoff - soff may not cross a 3370 * page boundry or cross the end of the buffer. The end of the 3371 * buffer, in this case, is our file EOF, not the allocation size 3372 * of the buffer. 3373 */ 3374 soff = off; 3375 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3376 if (eoff > bp->b_offset + bp->b_bcount) 3377 eoff = bp->b_offset + bp->b_bcount; 3378 3379 /* 3380 * Set valid range. This is typically the entire buffer and thus the 3381 * entire page. 3382 */ 3383 if (eoff > soff) { 3384 vm_page_set_validclean( 3385 m, 3386 (vm_offset_t) (soff & PAGE_MASK), 3387 (vm_offset_t) (eoff - soff) 3388 ); 3389 } 3390} 3391 3392/* 3393 * This routine is called before a device strategy routine. 3394 * It is used to tell the VM system that paging I/O is in 3395 * progress, and treat the pages associated with the buffer 3396 * almost as being VPO_BUSY. Also the object paging_in_progress 3397 * flag is handled to make sure that the object doesn't become 3398 * inconsistant. 3399 * 3400 * Since I/O has not been initiated yet, certain buffer flags 3401 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3402 * and should be ignored. 3403 */ 3404void 3405vfs_busy_pages(struct buf *bp, int clear_modify) 3406{ 3407 int i, bogus; 3408 vm_object_t obj; 3409 vm_ooffset_t foff; 3410 vm_page_t m; 3411 3412 if (!(bp->b_flags & B_VMIO)) 3413 return; 3414 3415 obj = bp->b_bufobj->bo_object; 3416 foff = bp->b_offset; 3417 KASSERT(bp->b_offset != NOOFFSET, 3418 ("vfs_busy_pages: no buffer offset")); 3419 VM_OBJECT_LOCK(obj); 3420 if (bp->b_bufsize != 0) 3421 vfs_setdirty_locked_object(bp); 3422retry: 3423 for (i = 0; i < bp->b_npages; i++) { 3424 m = bp->b_pages[i]; 3425 3426 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3427 goto retry; 3428 } 3429 bogus = 0; 3430 vm_page_lock_queues(); 3431 for (i = 0; i < bp->b_npages; i++) { 3432 m = bp->b_pages[i]; 3433 3434 if ((bp->b_flags & B_CLUSTER) == 0) { 3435 vm_object_pip_add(obj, 1); 3436 vm_page_io_start(m); 3437 } 3438 /* 3439 * When readying a buffer for a read ( i.e 3440 * clear_modify == 0 ), it is important to do 3441 * bogus_page replacement for valid pages in 3442 * partially instantiated buffers. Partially 3443 * instantiated buffers can, in turn, occur when 3444 * reconstituting a buffer from its VM backing store 3445 * base. We only have to do this if B_CACHE is 3446 * clear ( which causes the I/O to occur in the 3447 * first place ). The replacement prevents the read 3448 * I/O from overwriting potentially dirty VM-backed 3449 * pages. XXX bogus page replacement is, uh, bogus. 3450 * It may not work properly with small-block devices. 3451 * We need to find a better way. 3452 */ 3453 pmap_remove_all(m); 3454 if (clear_modify) 3455 vfs_page_set_valid(bp, foff, m); 3456 else if (m->valid == VM_PAGE_BITS_ALL && 3457 (bp->b_flags & B_CACHE) == 0) { 3458 bp->b_pages[i] = bogus_page; 3459 bogus++; 3460 } 3461 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3462 } 3463 vm_page_unlock_queues(); 3464 VM_OBJECT_UNLOCK(obj); 3465 if (bogus) 3466 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3467 bp->b_pages, bp->b_npages); 3468} 3469 3470/* 3471 * Tell the VM system that the pages associated with this buffer 3472 * are clean. This is used for delayed writes where the data is 3473 * going to go to disk eventually without additional VM intevention. 3474 * 3475 * Note that while we only really need to clean through to b_bcount, we 3476 * just go ahead and clean through to b_bufsize. 3477 */ 3478static void 3479vfs_clean_pages(struct buf *bp) 3480{ 3481 int i; 3482 vm_ooffset_t foff, noff, eoff; 3483 vm_page_t m; 3484 3485 if (!(bp->b_flags & B_VMIO)) 3486 return; 3487 3488 foff = bp->b_offset; 3489 KASSERT(bp->b_offset != NOOFFSET, 3490 ("vfs_clean_pages: no buffer offset")); 3491 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3492 vm_page_lock_queues(); 3493 for (i = 0; i < bp->b_npages; i++) { 3494 m = bp->b_pages[i]; 3495 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3496 eoff = noff; 3497 3498 if (eoff > bp->b_offset + bp->b_bufsize) 3499 eoff = bp->b_offset + bp->b_bufsize; 3500 vfs_page_set_valid(bp, foff, m); 3501 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3502 foff = noff; 3503 } 3504 vm_page_unlock_queues(); 3505 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3506} 3507 3508/* 3509 * vfs_bio_set_validclean: 3510 * 3511 * Set the range within the buffer to valid and clean. The range is 3512 * relative to the beginning of the buffer, b_offset. Note that b_offset 3513 * itself may be offset from the beginning of the first page. 3514 * 3515 */ 3516 3517void 3518vfs_bio_set_validclean(struct buf *bp, int base, int size) 3519{ 3520 int i, n; 3521 vm_page_t m; 3522 3523 if (!(bp->b_flags & B_VMIO)) 3524 return; 3525 /* 3526 * Fixup base to be relative to beginning of first page. 3527 * Set initial n to be the maximum number of bytes in the 3528 * first page that can be validated. 3529 */ 3530 3531 base += (bp->b_offset & PAGE_MASK); 3532 n = PAGE_SIZE - (base & PAGE_MASK); 3533 3534 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3535 vm_page_lock_queues(); 3536 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3537 m = bp->b_pages[i]; 3538 if (n > size) 3539 n = size; 3540 vm_page_set_validclean(m, base & PAGE_MASK, n); 3541 base += n; 3542 size -= n; 3543 n = PAGE_SIZE; 3544 } 3545 vm_page_unlock_queues(); 3546 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3547} 3548 3549/* 3550 * vfs_bio_clrbuf: 3551 * 3552 * clear a buffer. This routine essentially fakes an I/O, so we need 3553 * to clear BIO_ERROR and B_INVAL. 3554 * 3555 * Note that while we only theoretically need to clear through b_bcount, 3556 * we go ahead and clear through b_bufsize. 3557 */ 3558 3559void 3560vfs_bio_clrbuf(struct buf *bp) 3561{ 3562 int i, j, mask = 0; 3563 caddr_t sa, ea; 3564 3565 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3566 clrbuf(bp); 3567 return; 3568 } 3569 3570 bp->b_flags &= ~B_INVAL; 3571 bp->b_ioflags &= ~BIO_ERROR; 3572 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3573 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3574 (bp->b_offset & PAGE_MASK) == 0) { 3575 if (bp->b_pages[0] == bogus_page) 3576 goto unlock; 3577 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3578 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3579 if ((bp->b_pages[0]->valid & mask) == mask) 3580 goto unlock; 3581 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3582 ((bp->b_pages[0]->valid & mask) == 0)) { 3583 bzero(bp->b_data, bp->b_bufsize); 3584 bp->b_pages[0]->valid |= mask; 3585 goto unlock; 3586 } 3587 } 3588 ea = sa = bp->b_data; 3589 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3590 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3591 ea = (caddr_t)(vm_offset_t)ulmin( 3592 (u_long)(vm_offset_t)ea, 3593 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3594 if (bp->b_pages[i] == bogus_page) 3595 continue; 3596 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3597 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3598 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3599 if ((bp->b_pages[i]->valid & mask) == mask) 3600 continue; 3601 if ((bp->b_pages[i]->valid & mask) == 0) { 3602 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3603 bzero(sa, ea - sa); 3604 } else { 3605 for (; sa < ea; sa += DEV_BSIZE, j++) { 3606 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3607 (bp->b_pages[i]->valid & (1 << j)) == 0) 3608 bzero(sa, DEV_BSIZE); 3609 } 3610 } 3611 bp->b_pages[i]->valid |= mask; 3612 } 3613unlock: 3614 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3615 bp->b_resid = 0; 3616} 3617 3618/* 3619 * vm_hold_load_pages and vm_hold_free_pages get pages into 3620 * a buffers address space. The pages are anonymous and are 3621 * not associated with a file object. 3622 */ 3623static void 3624vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3625{ 3626 vm_offset_t pg; 3627 vm_page_t p; 3628 int index; 3629 3630 to = round_page(to); 3631 from = round_page(from); 3632 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3633 3634 VM_OBJECT_LOCK(kernel_object); 3635 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3636tryagain: 3637 /* 3638 * note: must allocate system pages since blocking here 3639 * could intefere with paging I/O, no matter which 3640 * process we are. 3641 */ 3642 p = vm_page_alloc(kernel_object, 3643 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3644 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3645 if (!p) { 3646 atomic_add_int(&vm_pageout_deficit, 3647 (to - pg) >> PAGE_SHIFT); 3648 VM_OBJECT_UNLOCK(kernel_object); 3649 VM_WAIT; 3650 VM_OBJECT_LOCK(kernel_object); 3651 goto tryagain; 3652 } 3653 p->valid = VM_PAGE_BITS_ALL; 3654 pmap_qenter(pg, &p, 1); 3655 bp->b_pages[index] = p; 3656 } 3657 VM_OBJECT_UNLOCK(kernel_object); 3658 bp->b_npages = index; 3659} 3660 3661/* Return pages associated with this buf to the vm system */ 3662static void 3663vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3664{ 3665 vm_offset_t pg; 3666 vm_page_t p; 3667 int index, newnpages; 3668 3669 from = round_page(from); 3670 to = round_page(to); 3671 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3672 3673 VM_OBJECT_LOCK(kernel_object); 3674 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3675 p = bp->b_pages[index]; 3676 if (p && (index < bp->b_npages)) { 3677 if (p->busy) { 3678 printf( 3679 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3680 (intmax_t)bp->b_blkno, 3681 (intmax_t)bp->b_lblkno); 3682 } 3683 bp->b_pages[index] = NULL; 3684 pmap_qremove(pg, 1); 3685 vm_page_lock_queues(); 3686 vm_page_unwire(p, 0); 3687 vm_page_free(p); 3688 vm_page_unlock_queues(); 3689 } 3690 } 3691 VM_OBJECT_UNLOCK(kernel_object); 3692 bp->b_npages = newnpages; 3693} 3694 3695/* 3696 * Map an IO request into kernel virtual address space. 3697 * 3698 * All requests are (re)mapped into kernel VA space. 3699 * Notice that we use b_bufsize for the size of the buffer 3700 * to be mapped. b_bcount might be modified by the driver. 3701 * 3702 * Note that even if the caller determines that the address space should 3703 * be valid, a race or a smaller-file mapped into a larger space may 3704 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3705 * check the return value. 3706 */ 3707int 3708vmapbuf(struct buf *bp) 3709{ 3710 caddr_t addr, kva; 3711 vm_prot_t prot; 3712 int pidx, i; 3713 struct vm_page *m; 3714 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3715 3716 if (bp->b_bufsize < 0) 3717 return (-1); 3718 prot = VM_PROT_READ; 3719 if (bp->b_iocmd == BIO_READ) 3720 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3721 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3722 addr < bp->b_data + bp->b_bufsize; 3723 addr += PAGE_SIZE, pidx++) { 3724 /* 3725 * Do the vm_fault if needed; do the copy-on-write thing 3726 * when reading stuff off device into memory. 3727 * 3728 * NOTE! Must use pmap_extract() because addr may be in 3729 * the userland address space, and kextract is only guarenteed 3730 * to work for the kernland address space (see: sparc64 port). 3731 */ 3732retry: 3733 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3734 prot) < 0) { 3735 vm_page_lock_queues(); 3736 for (i = 0; i < pidx; ++i) { 3737 vm_page_unhold(bp->b_pages[i]); 3738 bp->b_pages[i] = NULL; 3739 } 3740 vm_page_unlock_queues(); 3741 return(-1); 3742 } 3743 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3744 if (m == NULL) 3745 goto retry; 3746 bp->b_pages[pidx] = m; 3747 } 3748 if (pidx > btoc(MAXPHYS)) 3749 panic("vmapbuf: mapped more than MAXPHYS"); 3750 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3751 3752 kva = bp->b_saveaddr; 3753 bp->b_npages = pidx; 3754 bp->b_saveaddr = bp->b_data; 3755 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3756 return(0); 3757} 3758 3759/* 3760 * Free the io map PTEs associated with this IO operation. 3761 * We also invalidate the TLB entries and restore the original b_addr. 3762 */ 3763void 3764vunmapbuf(struct buf *bp) 3765{ 3766 int pidx; 3767 int npages; 3768 3769 npages = bp->b_npages; 3770 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3771 vm_page_lock_queues(); 3772 for (pidx = 0; pidx < npages; pidx++) 3773 vm_page_unhold(bp->b_pages[pidx]); 3774 vm_page_unlock_queues(); 3775 3776 bp->b_data = bp->b_saveaddr; 3777} 3778 3779void 3780bdone(struct buf *bp) 3781{ 3782 3783 mtx_lock(&bdonelock); 3784 bp->b_flags |= B_DONE; 3785 wakeup(bp); 3786 mtx_unlock(&bdonelock); 3787} 3788 3789void 3790bwait(struct buf *bp, u_char pri, const char *wchan) 3791{ 3792 3793 mtx_lock(&bdonelock); 3794 while ((bp->b_flags & B_DONE) == 0) 3795 msleep(bp, &bdonelock, pri, wchan, 0); 3796 mtx_unlock(&bdonelock); 3797} 3798 3799int 3800bufsync(struct bufobj *bo, int waitfor, struct thread *td) 3801{ 3802 3803 return (VOP_FSYNC(bo->__bo_vnode, waitfor, td)); 3804} 3805 3806void 3807bufstrategy(struct bufobj *bo, struct buf *bp) 3808{ 3809 int i = 0; 3810 struct vnode *vp; 3811 3812 vp = bp->b_vp; 3813 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3814 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3815 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3816 i = VOP_STRATEGY(vp, bp); 3817 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3818} 3819 3820void 3821bufobj_wrefl(struct bufobj *bo) 3822{ 3823 3824 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3825 ASSERT_BO_LOCKED(bo); 3826 bo->bo_numoutput++; 3827} 3828 3829void 3830bufobj_wref(struct bufobj *bo) 3831{ 3832 3833 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3834 BO_LOCK(bo); 3835 bo->bo_numoutput++; 3836 BO_UNLOCK(bo); 3837} 3838 3839void 3840bufobj_wdrop(struct bufobj *bo) 3841{ 3842 3843 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3844 BO_LOCK(bo); 3845 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3846 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3847 bo->bo_flag &= ~BO_WWAIT; 3848 wakeup(&bo->bo_numoutput); 3849 } 3850 BO_UNLOCK(bo); 3851} 3852 3853int 3854bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3855{ 3856 int error; 3857 3858 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3859 ASSERT_BO_LOCKED(bo); 3860 error = 0; 3861 while (bo->bo_numoutput) { 3862 bo->bo_flag |= BO_WWAIT; 3863 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3864 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3865 if (error) 3866 break; 3867 } 3868 return (error); 3869} 3870 3871void 3872bpin(struct buf *bp) 3873{ 3874 mtx_lock(&bpinlock); 3875 bp->b_pin_count++; 3876 mtx_unlock(&bpinlock); 3877} 3878 3879void 3880bunpin(struct buf *bp) 3881{ 3882 mtx_lock(&bpinlock); 3883 if (--bp->b_pin_count == 0) 3884 wakeup(bp); 3885 mtx_unlock(&bpinlock); 3886} 3887 3888void 3889bunpin_wait(struct buf *bp) 3890{ 3891 mtx_lock(&bpinlock); 3892 while (bp->b_pin_count > 0) 3893 msleep(bp, &bpinlock, PRIBIO, "bwunpin", 0); 3894 mtx_unlock(&bpinlock); 3895} 3896 3897#include "opt_ddb.h" 3898#ifdef DDB 3899#include <ddb/ddb.h> 3900 3901/* DDB command to show buffer data */ 3902DB_SHOW_COMMAND(buffer, db_show_buffer) 3903{ 3904 /* get args */ 3905 struct buf *bp = (struct buf *)addr; 3906 3907 if (!have_addr) { 3908 db_printf("usage: show buffer <addr>\n"); 3909 return; 3910 } 3911 3912 db_printf("buf at %p\n", bp); 3913 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3914 db_printf( 3915 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3916 "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n", 3917 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3918 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno); 3919 if (bp->b_npages) { 3920 int i; 3921 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3922 for (i = 0; i < bp->b_npages; i++) { 3923 vm_page_t m; 3924 m = bp->b_pages[i]; 3925 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3926 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3927 if ((i + 1) < bp->b_npages) 3928 db_printf(","); 3929 } 3930 db_printf("\n"); 3931 } 3932 lockmgr_printinfo(&bp->b_lock); 3933} 3934 3935DB_SHOW_COMMAND(lockedbufs, lockedbufs) 3936{ 3937 struct buf *bp; 3938 int i; 3939 3940 for (i = 0; i < nbuf; i++) { 3941 bp = &buf[i]; 3942 if (BUF_ISLOCKED(bp)) { 3943 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 3944 db_printf("\n"); 3945 } 3946 } 3947} 3948#endif /* DDB */ 3949