vfs_bio.c revision 176708
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 176708 2008-03-01 19:47:50Z 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 bp->b_waiters = 0; 561 LIST_INIT(&bp->b_dep); 562 BUF_LOCKINIT(bp); 563 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 564 } 565 566 /* 567 * maxbufspace is the absolute maximum amount of buffer space we are 568 * allowed to reserve in KVM and in real terms. The absolute maximum 569 * is nominally used by buf_daemon. hibufspace is the nominal maximum 570 * used by most other processes. The differential is required to 571 * ensure that buf_daemon is able to run when other processes might 572 * be blocked waiting for buffer space. 573 * 574 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 575 * this may result in KVM fragmentation which is not handled optimally 576 * by the system. 577 */ 578 maxbufspace = nbuf * BKVASIZE; 579 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 580 lobufspace = hibufspace - MAXBSIZE; 581 582 lorunningspace = 512 * 1024; 583 hirunningspace = 1024 * 1024; 584 585/* 586 * Limit the amount of malloc memory since it is wired permanently into 587 * the kernel space. Even though this is accounted for in the buffer 588 * allocation, we don't want the malloced region to grow uncontrolled. 589 * The malloc scheme improves memory utilization significantly on average 590 * (small) directories. 591 */ 592 maxbufmallocspace = hibufspace / 20; 593 594/* 595 * Reduce the chance of a deadlock occuring by limiting the number 596 * of delayed-write dirty buffers we allow to stack up. 597 */ 598 hidirtybuffers = nbuf / 4 + 20; 599 dirtybufthresh = hidirtybuffers * 9 / 10; 600 numdirtybuffers = 0; 601/* 602 * To support extreme low-memory systems, make sure hidirtybuffers cannot 603 * eat up all available buffer space. This occurs when our minimum cannot 604 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 605 * BKVASIZE'd (8K) buffers. 606 */ 607 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 608 hidirtybuffers >>= 1; 609 } 610 lodirtybuffers = hidirtybuffers / 2; 611 612/* 613 * Try to keep the number of free buffers in the specified range, 614 * and give special processes (e.g. like buf_daemon) access to an 615 * emergency reserve. 616 */ 617 lofreebuffers = nbuf / 18 + 5; 618 hifreebuffers = 2 * lofreebuffers; 619 numfreebuffers = nbuf; 620 621/* 622 * Maximum number of async ops initiated per buf_daemon loop. This is 623 * somewhat of a hack at the moment, we really need to limit ourselves 624 * based on the number of bytes of I/O in-transit that were initiated 625 * from buf_daemon. 626 */ 627 628 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 629 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 630} 631 632/* 633 * bfreekva() - free the kva allocation for a buffer. 634 * 635 * Since this call frees up buffer space, we call bufspacewakeup(). 636 */ 637static void 638bfreekva(struct buf *bp) 639{ 640 641 if (bp->b_kvasize) { 642 atomic_add_int(&buffreekvacnt, 1); 643 atomic_subtract_int(&bufspace, bp->b_kvasize); 644 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase, 645 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 646 bp->b_kvasize = 0; 647 bufspacewakeup(); 648 } 649} 650 651/* 652 * bremfree: 653 * 654 * Mark the buffer for removal from the appropriate free list in brelse. 655 * 656 */ 657void 658bremfree(struct buf *bp) 659{ 660 661 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 662 KASSERT((bp->b_flags & B_REMFREE) == 0, 663 ("bremfree: buffer %p already marked for delayed removal.", bp)); 664 KASSERT(bp->b_qindex != QUEUE_NONE, 665 ("bremfree: buffer %p not on a queue.", bp)); 666 BUF_ASSERT_HELD(bp); 667 668 bp->b_flags |= B_REMFREE; 669 /* Fixup numfreebuffers count. */ 670 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 671 atomic_subtract_int(&numfreebuffers, 1); 672} 673 674/* 675 * bremfreef: 676 * 677 * Force an immediate removal from a free list. Used only in nfs when 678 * it abuses the b_freelist pointer. 679 */ 680void 681bremfreef(struct buf *bp) 682{ 683 mtx_lock(&bqlock); 684 bremfreel(bp); 685 mtx_unlock(&bqlock); 686} 687 688/* 689 * bremfreel: 690 * 691 * Removes a buffer from the free list, must be called with the 692 * bqlock held. 693 */ 694static void 695bremfreel(struct buf *bp) 696{ 697 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 698 bp, bp->b_vp, bp->b_flags); 699 KASSERT(bp->b_qindex != QUEUE_NONE, 700 ("bremfreel: buffer %p not on a queue.", bp)); 701 BUF_ASSERT_HELD(bp); 702 mtx_assert(&bqlock, MA_OWNED); 703 704 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 705 bp->b_qindex = QUEUE_NONE; 706 /* 707 * If this was a delayed bremfree() we only need to remove the buffer 708 * from the queue and return the stats are already done. 709 */ 710 if (bp->b_flags & B_REMFREE) { 711 bp->b_flags &= ~B_REMFREE; 712 return; 713 } 714 /* 715 * Fixup numfreebuffers count. If the buffer is invalid or not 716 * delayed-write, the buffer was free and we must decrement 717 * numfreebuffers. 718 */ 719 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 720 atomic_subtract_int(&numfreebuffers, 1); 721} 722 723 724/* 725 * Get a buffer with the specified data. Look in the cache first. We 726 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 727 * is set, the buffer is valid and we do not have to do anything ( see 728 * getblk() ). This is really just a special case of breadn(). 729 */ 730int 731bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 732 struct buf **bpp) 733{ 734 735 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 736} 737 738/* 739 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 740 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 741 * the buffer is valid and we do not have to do anything. 742 */ 743void 744breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 745 int cnt, struct ucred * cred) 746{ 747 struct buf *rabp; 748 int i; 749 750 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 751 if (inmem(vp, *rablkno)) 752 continue; 753 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 754 755 if ((rabp->b_flags & B_CACHE) == 0) { 756 if (!TD_IS_IDLETHREAD(curthread)) 757 curthread->td_ru.ru_inblock++; 758 rabp->b_flags |= B_ASYNC; 759 rabp->b_flags &= ~B_INVAL; 760 rabp->b_ioflags &= ~BIO_ERROR; 761 rabp->b_iocmd = BIO_READ; 762 if (rabp->b_rcred == NOCRED && cred != NOCRED) 763 rabp->b_rcred = crhold(cred); 764 vfs_busy_pages(rabp, 0); 765 BUF_KERNPROC(rabp); 766 rabp->b_iooffset = dbtob(rabp->b_blkno); 767 bstrategy(rabp); 768 } else { 769 brelse(rabp); 770 } 771 } 772} 773 774/* 775 * Operates like bread, but also starts asynchronous I/O on 776 * read-ahead blocks. 777 */ 778int 779breadn(struct vnode * vp, daddr_t blkno, int size, 780 daddr_t * rablkno, int *rabsize, 781 int cnt, struct ucred * cred, struct buf **bpp) 782{ 783 struct buf *bp; 784 int rv = 0, readwait = 0; 785 786 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 787 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 788 789 /* if not found in cache, do some I/O */ 790 if ((bp->b_flags & B_CACHE) == 0) { 791 if (!TD_IS_IDLETHREAD(curthread)) 792 curthread->td_ru.ru_inblock++; 793 bp->b_iocmd = BIO_READ; 794 bp->b_flags &= ~B_INVAL; 795 bp->b_ioflags &= ~BIO_ERROR; 796 if (bp->b_rcred == NOCRED && cred != NOCRED) 797 bp->b_rcred = crhold(cred); 798 vfs_busy_pages(bp, 0); 799 bp->b_iooffset = dbtob(bp->b_blkno); 800 bstrategy(bp); 801 ++readwait; 802 } 803 804 breada(vp, rablkno, rabsize, cnt, cred); 805 806 if (readwait) { 807 rv = bufwait(bp); 808 } 809 return (rv); 810} 811 812/* 813 * Write, release buffer on completion. (Done by iodone 814 * if async). Do not bother writing anything if the buffer 815 * is invalid. 816 * 817 * Note that we set B_CACHE here, indicating that buffer is 818 * fully valid and thus cacheable. This is true even of NFS 819 * now so we set it generally. This could be set either here 820 * or in biodone() since the I/O is synchronous. We put it 821 * here. 822 */ 823int 824bufwrite(struct buf *bp) 825{ 826 int oldflags; 827 struct vnode *vp; 828 int vp_md; 829 830 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 831 if (bp->b_flags & B_INVAL) { 832 brelse(bp); 833 return (0); 834 } 835 836 oldflags = bp->b_flags; 837 838 BUF_ASSERT_HELD(bp); 839 840 if (bp->b_pin_count > 0) 841 bunpin_wait(bp); 842 843 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 844 ("FFS background buffer should not get here %p", bp)); 845 846 vp = bp->b_vp; 847 if (vp) 848 vp_md = vp->v_vflag & VV_MD; 849 else 850 vp_md = 0; 851 852 /* Mark the buffer clean */ 853 bundirty(bp); 854 855 bp->b_flags &= ~B_DONE; 856 bp->b_ioflags &= ~BIO_ERROR; 857 bp->b_flags |= B_CACHE; 858 bp->b_iocmd = BIO_WRITE; 859 860 bufobj_wref(bp->b_bufobj); 861 vfs_busy_pages(bp, 1); 862 863 /* 864 * Normal bwrites pipeline writes 865 */ 866 bp->b_runningbufspace = bp->b_bufsize; 867 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 868 869 if (!TD_IS_IDLETHREAD(curthread)) 870 curthread->td_ru.ru_oublock++; 871 if (oldflags & B_ASYNC) 872 BUF_KERNPROC(bp); 873 bp->b_iooffset = dbtob(bp->b_blkno); 874 bstrategy(bp); 875 876 if ((oldflags & B_ASYNC) == 0) { 877 int rtval = bufwait(bp); 878 brelse(bp); 879 return (rtval); 880 } else { 881 /* 882 * don't allow the async write to saturate the I/O 883 * system. We will not deadlock here because 884 * we are blocking waiting for I/O that is already in-progress 885 * to complete. We do not block here if it is the update 886 * or syncer daemon trying to clean up as that can lead 887 * to deadlock. 888 */ 889 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 890 waitrunningbufspace(); 891 } 892 893 return (0); 894} 895 896void 897bufbdflush(struct bufobj *bo, struct buf *bp) 898{ 899 struct buf *nbp; 900 901 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 902 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 903 altbufferflushes++; 904 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 905 BO_LOCK(bo); 906 /* 907 * Try to find a buffer to flush. 908 */ 909 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 910 if ((nbp->b_vflags & BV_BKGRDINPROG) || 911 BUF_LOCK(nbp, 912 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 913 continue; 914 if (bp == nbp) 915 panic("bdwrite: found ourselves"); 916 BO_UNLOCK(bo); 917 /* Don't countdeps with the bo lock held. */ 918 if (buf_countdeps(nbp, 0)) { 919 BO_LOCK(bo); 920 BUF_UNLOCK(nbp); 921 continue; 922 } 923 if (nbp->b_flags & B_CLUSTEROK) { 924 vfs_bio_awrite(nbp); 925 } else { 926 bremfree(nbp); 927 bawrite(nbp); 928 } 929 dirtybufferflushes++; 930 break; 931 } 932 if (nbp == NULL) 933 BO_UNLOCK(bo); 934 } 935} 936 937/* 938 * Delayed write. (Buffer is marked dirty). Do not bother writing 939 * anything if the buffer is marked invalid. 940 * 941 * Note that since the buffer must be completely valid, we can safely 942 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 943 * biodone() in order to prevent getblk from writing the buffer 944 * out synchronously. 945 */ 946void 947bdwrite(struct buf *bp) 948{ 949 struct thread *td = curthread; 950 struct vnode *vp; 951 struct bufobj *bo; 952 953 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 954 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 955 BUF_ASSERT_HELD(bp); 956 957 if (bp->b_flags & B_INVAL) { 958 brelse(bp); 959 return; 960 } 961 962 /* 963 * If we have too many dirty buffers, don't create any more. 964 * If we are wildly over our limit, then force a complete 965 * cleanup. Otherwise, just keep the situation from getting 966 * out of control. Note that we have to avoid a recursive 967 * disaster and not try to clean up after our own cleanup! 968 */ 969 vp = bp->b_vp; 970 bo = bp->b_bufobj; 971 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 972 td->td_pflags |= TDP_INBDFLUSH; 973 BO_BDFLUSH(bo, bp); 974 td->td_pflags &= ~TDP_INBDFLUSH; 975 } else 976 recursiveflushes++; 977 978 bdirty(bp); 979 /* 980 * Set B_CACHE, indicating that the buffer is fully valid. This is 981 * true even of NFS now. 982 */ 983 bp->b_flags |= B_CACHE; 984 985 /* 986 * This bmap keeps the system from needing to do the bmap later, 987 * perhaps when the system is attempting to do a sync. Since it 988 * is likely that the indirect block -- or whatever other datastructure 989 * that the filesystem needs is still in memory now, it is a good 990 * thing to do this. Note also, that if the pageout daemon is 991 * requesting a sync -- there might not be enough memory to do 992 * the bmap then... So, this is important to do. 993 */ 994 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 995 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 996 } 997 998 /* 999 * Set the *dirty* buffer range based upon the VM system dirty pages. 1000 */ 1001 vfs_setdirty(bp); 1002 1003 /* 1004 * We need to do this here to satisfy the vnode_pager and the 1005 * pageout daemon, so that it thinks that the pages have been 1006 * "cleaned". Note that since the pages are in a delayed write 1007 * buffer -- the VFS layer "will" see that the pages get written 1008 * out on the next sync, or perhaps the cluster will be completed. 1009 */ 1010 vfs_clean_pages(bp); 1011 bqrelse(bp); 1012 1013 /* 1014 * Wakeup the buffer flushing daemon if we have a lot of dirty 1015 * buffers (midpoint between our recovery point and our stall 1016 * point). 1017 */ 1018 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1019 1020 /* 1021 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1022 * due to the softdep code. 1023 */ 1024} 1025 1026/* 1027 * bdirty: 1028 * 1029 * Turn buffer into delayed write request. We must clear BIO_READ and 1030 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1031 * itself to properly update it in the dirty/clean lists. We mark it 1032 * B_DONE to ensure that any asynchronization of the buffer properly 1033 * clears B_DONE ( else a panic will occur later ). 1034 * 1035 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1036 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1037 * should only be called if the buffer is known-good. 1038 * 1039 * Since the buffer is not on a queue, we do not update the numfreebuffers 1040 * count. 1041 * 1042 * The buffer must be on QUEUE_NONE. 1043 */ 1044void 1045bdirty(struct buf *bp) 1046{ 1047 1048 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1049 bp, bp->b_vp, bp->b_flags); 1050 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1051 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1052 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1053 BUF_ASSERT_HELD(bp); 1054 bp->b_flags &= ~(B_RELBUF); 1055 bp->b_iocmd = BIO_WRITE; 1056 1057 if ((bp->b_flags & B_DELWRI) == 0) { 1058 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1059 reassignbuf(bp); 1060 atomic_add_int(&numdirtybuffers, 1); 1061 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1062 } 1063} 1064 1065/* 1066 * bundirty: 1067 * 1068 * Clear B_DELWRI for buffer. 1069 * 1070 * Since the buffer is not on a queue, we do not update the numfreebuffers 1071 * count. 1072 * 1073 * The buffer must be on QUEUE_NONE. 1074 */ 1075 1076void 1077bundirty(struct buf *bp) 1078{ 1079 1080 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1081 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1082 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1083 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1084 BUF_ASSERT_HELD(bp); 1085 1086 if (bp->b_flags & B_DELWRI) { 1087 bp->b_flags &= ~B_DELWRI; 1088 reassignbuf(bp); 1089 atomic_subtract_int(&numdirtybuffers, 1); 1090 numdirtywakeup(lodirtybuffers); 1091 } 1092 /* 1093 * Since it is now being written, we can clear its deferred write flag. 1094 */ 1095 bp->b_flags &= ~B_DEFERRED; 1096} 1097 1098/* 1099 * bawrite: 1100 * 1101 * Asynchronous write. Start output on a buffer, but do not wait for 1102 * it to complete. The buffer is released when the output completes. 1103 * 1104 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1105 * B_INVAL buffers. Not us. 1106 */ 1107void 1108bawrite(struct buf *bp) 1109{ 1110 1111 bp->b_flags |= B_ASYNC; 1112 (void) bwrite(bp); 1113} 1114 1115/* 1116 * bwillwrite: 1117 * 1118 * Called prior to the locking of any vnodes when we are expecting to 1119 * write. We do not want to starve the buffer cache with too many 1120 * dirty buffers so we block here. By blocking prior to the locking 1121 * of any vnodes we attempt to avoid the situation where a locked vnode 1122 * prevents the various system daemons from flushing related buffers. 1123 */ 1124 1125void 1126bwillwrite(void) 1127{ 1128 1129 if (numdirtybuffers >= hidirtybuffers) { 1130 mtx_lock(&nblock); 1131 while (numdirtybuffers >= hidirtybuffers) { 1132 bd_wakeup(1); 1133 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1134 msleep(&needsbuffer, &nblock, 1135 (PRIBIO + 4), "flswai", 0); 1136 } 1137 mtx_unlock(&nblock); 1138 } 1139} 1140 1141/* 1142 * Return true if we have too many dirty buffers. 1143 */ 1144int 1145buf_dirty_count_severe(void) 1146{ 1147 1148 return(numdirtybuffers >= hidirtybuffers); 1149} 1150 1151/* 1152 * brelse: 1153 * 1154 * Release a busy buffer and, if requested, free its resources. The 1155 * buffer will be stashed in the appropriate bufqueue[] allowing it 1156 * to be accessed later as a cache entity or reused for other purposes. 1157 */ 1158void 1159brelse(struct buf *bp) 1160{ 1161 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1162 bp, bp->b_vp, bp->b_flags); 1163 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1164 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1165 1166 if (bp->b_flags & B_MANAGED) { 1167 bqrelse(bp); 1168 return; 1169 } 1170 1171 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1172 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1173 /* 1174 * Failed write, redirty. Must clear BIO_ERROR to prevent 1175 * pages from being scrapped. If the error is anything 1176 * other than an I/O error (EIO), assume that retryingi 1177 * is futile. 1178 */ 1179 bp->b_ioflags &= ~BIO_ERROR; 1180 bdirty(bp); 1181 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1182 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1183 /* 1184 * Either a failed I/O or we were asked to free or not 1185 * cache the buffer. 1186 */ 1187 bp->b_flags |= B_INVAL; 1188 if (!LIST_EMPTY(&bp->b_dep)) 1189 buf_deallocate(bp); 1190 if (bp->b_flags & B_DELWRI) { 1191 atomic_subtract_int(&numdirtybuffers, 1); 1192 numdirtywakeup(lodirtybuffers); 1193 } 1194 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1195 if ((bp->b_flags & B_VMIO) == 0) { 1196 if (bp->b_bufsize) 1197 allocbuf(bp, 0); 1198 if (bp->b_vp) 1199 (void) brelvp(bp); 1200 } 1201 } 1202 1203 /* 1204 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1205 * is called with B_DELWRI set, the underlying pages may wind up 1206 * getting freed causing a previous write (bdwrite()) to get 'lost' 1207 * because pages associated with a B_DELWRI bp are marked clean. 1208 * 1209 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1210 * if B_DELWRI is set. 1211 * 1212 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1213 * on pages to return pages to the VM page queues. 1214 */ 1215 if (bp->b_flags & B_DELWRI) 1216 bp->b_flags &= ~B_RELBUF; 1217 else if (vm_page_count_severe()) { 1218 /* 1219 * XXX This lock may not be necessary since BKGRDINPROG 1220 * cannot be set while we hold the buf lock, it can only be 1221 * cleared if it is already pending. 1222 */ 1223 if (bp->b_vp) { 1224 BO_LOCK(bp->b_bufobj); 1225 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1226 bp->b_flags |= B_RELBUF; 1227 BO_UNLOCK(bp->b_bufobj); 1228 } else 1229 bp->b_flags |= B_RELBUF; 1230 } 1231 1232 /* 1233 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1234 * constituted, not even NFS buffers now. Two flags effect this. If 1235 * B_INVAL, the struct buf is invalidated but the VM object is kept 1236 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1237 * 1238 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1239 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1240 * buffer is also B_INVAL because it hits the re-dirtying code above. 1241 * 1242 * Normally we can do this whether a buffer is B_DELWRI or not. If 1243 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1244 * the commit state and we cannot afford to lose the buffer. If the 1245 * buffer has a background write in progress, we need to keep it 1246 * around to prevent it from being reconstituted and starting a second 1247 * background write. 1248 */ 1249 if ((bp->b_flags & B_VMIO) 1250 && !(bp->b_vp->v_mount != NULL && 1251 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1252 !vn_isdisk(bp->b_vp, NULL) && 1253 (bp->b_flags & B_DELWRI)) 1254 ) { 1255 1256 int i, j, resid; 1257 vm_page_t m; 1258 off_t foff; 1259 vm_pindex_t poff; 1260 vm_object_t obj; 1261 1262 obj = bp->b_bufobj->bo_object; 1263 1264 /* 1265 * Get the base offset and length of the buffer. Note that 1266 * in the VMIO case if the buffer block size is not 1267 * page-aligned then b_data pointer may not be page-aligned. 1268 * But our b_pages[] array *IS* page aligned. 1269 * 1270 * block sizes less then DEV_BSIZE (usually 512) are not 1271 * supported due to the page granularity bits (m->valid, 1272 * m->dirty, etc...). 1273 * 1274 * See man buf(9) for more information 1275 */ 1276 resid = bp->b_bufsize; 1277 foff = bp->b_offset; 1278 VM_OBJECT_LOCK(obj); 1279 for (i = 0; i < bp->b_npages; i++) { 1280 int had_bogus = 0; 1281 1282 m = bp->b_pages[i]; 1283 1284 /* 1285 * If we hit a bogus page, fixup *all* the bogus pages 1286 * now. 1287 */ 1288 if (m == bogus_page) { 1289 poff = OFF_TO_IDX(bp->b_offset); 1290 had_bogus = 1; 1291 1292 for (j = i; j < bp->b_npages; j++) { 1293 vm_page_t mtmp; 1294 mtmp = bp->b_pages[j]; 1295 if (mtmp == bogus_page) { 1296 mtmp = vm_page_lookup(obj, poff + j); 1297 if (!mtmp) { 1298 panic("brelse: page missing\n"); 1299 } 1300 bp->b_pages[j] = mtmp; 1301 } 1302 } 1303 1304 if ((bp->b_flags & B_INVAL) == 0) { 1305 pmap_qenter( 1306 trunc_page((vm_offset_t)bp->b_data), 1307 bp->b_pages, bp->b_npages); 1308 } 1309 m = bp->b_pages[i]; 1310 } 1311 if ((bp->b_flags & B_NOCACHE) || 1312 (bp->b_ioflags & BIO_ERROR)) { 1313 int poffset = foff & PAGE_MASK; 1314 int presid = resid > (PAGE_SIZE - poffset) ? 1315 (PAGE_SIZE - poffset) : resid; 1316 1317 KASSERT(presid >= 0, ("brelse: extra page")); 1318 vm_page_lock_queues(); 1319 vm_page_set_invalid(m, poffset, presid); 1320 vm_page_unlock_queues(); 1321 if (had_bogus) 1322 printf("avoided corruption bug in bogus_page/brelse code\n"); 1323 } 1324 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1325 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1326 } 1327 VM_OBJECT_UNLOCK(obj); 1328 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1329 vfs_vmio_release(bp); 1330 1331 } else if (bp->b_flags & B_VMIO) { 1332 1333 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1334 vfs_vmio_release(bp); 1335 } 1336 1337 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1338 if (bp->b_bufsize != 0) 1339 allocbuf(bp, 0); 1340 if (bp->b_vp != NULL) 1341 (void) brelvp(bp); 1342 } 1343 1344 if (BUF_LOCKRECURSED(bp)) { 1345 /* do not release to free list */ 1346 BUF_UNLOCK(bp); 1347 return; 1348 } 1349 1350 /* enqueue */ 1351 mtx_lock(&bqlock); 1352 /* Handle delayed bremfree() processing. */ 1353 if (bp->b_flags & B_REMFREE) 1354 bremfreel(bp); 1355 if (bp->b_qindex != QUEUE_NONE) 1356 panic("brelse: free buffer onto another queue???"); 1357 1358 /* buffers with no memory */ 1359 if (bp->b_bufsize == 0) { 1360 bp->b_flags |= B_INVAL; 1361 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1362 if (bp->b_vflags & BV_BKGRDINPROG) 1363 panic("losing buffer 1"); 1364 if (bp->b_kvasize) { 1365 bp->b_qindex = QUEUE_EMPTYKVA; 1366 } else { 1367 bp->b_qindex = QUEUE_EMPTY; 1368 } 1369 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1370 /* buffers with junk contents */ 1371 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1372 (bp->b_ioflags & BIO_ERROR)) { 1373 bp->b_flags |= B_INVAL; 1374 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1375 if (bp->b_vflags & BV_BKGRDINPROG) 1376 panic("losing buffer 2"); 1377 bp->b_qindex = QUEUE_CLEAN; 1378 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1379 /* remaining buffers */ 1380 } else { 1381 if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) == 1382 (B_DELWRI|B_NEEDSGIANT)) 1383 bp->b_qindex = QUEUE_DIRTY_GIANT; 1384 if (bp->b_flags & B_DELWRI) 1385 bp->b_qindex = QUEUE_DIRTY; 1386 else 1387 bp->b_qindex = QUEUE_CLEAN; 1388 if (bp->b_flags & B_AGE) 1389 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1390 else 1391 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1392 } 1393 mtx_unlock(&bqlock); 1394 1395 /* 1396 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1397 * placed the buffer on the correct queue. We must also disassociate 1398 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1399 * find it. 1400 */ 1401 if (bp->b_flags & B_INVAL) { 1402 if (bp->b_flags & B_DELWRI) 1403 bundirty(bp); 1404 if (bp->b_vp) 1405 (void) brelvp(bp); 1406 } 1407 1408 /* 1409 * Fixup numfreebuffers count. The bp is on an appropriate queue 1410 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1411 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1412 * if B_INVAL is set ). 1413 */ 1414 1415 if (!(bp->b_flags & B_DELWRI)) 1416 bufcountwakeup(); 1417 1418 /* 1419 * Something we can maybe free or reuse 1420 */ 1421 if (bp->b_bufsize || bp->b_kvasize) 1422 bufspacewakeup(); 1423 1424 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1425 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1426 panic("brelse: not dirty"); 1427 /* unlock */ 1428 BUF_UNLOCK(bp); 1429} 1430 1431/* 1432 * Release a buffer back to the appropriate queue but do not try to free 1433 * it. The buffer is expected to be used again soon. 1434 * 1435 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1436 * biodone() to requeue an async I/O on completion. It is also used when 1437 * known good buffers need to be requeued but we think we may need the data 1438 * again soon. 1439 * 1440 * XXX we should be able to leave the B_RELBUF hint set on completion. 1441 */ 1442void 1443bqrelse(struct buf *bp) 1444{ 1445 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1446 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1447 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1448 1449 if (BUF_LOCKRECURSED(bp)) { 1450 /* do not release to free list */ 1451 BUF_UNLOCK(bp); 1452 return; 1453 } 1454 1455 if (bp->b_flags & B_MANAGED) { 1456 if (bp->b_flags & B_REMFREE) { 1457 mtx_lock(&bqlock); 1458 bremfreel(bp); 1459 mtx_unlock(&bqlock); 1460 } 1461 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1462 BUF_UNLOCK(bp); 1463 return; 1464 } 1465 1466 mtx_lock(&bqlock); 1467 /* Handle delayed bremfree() processing. */ 1468 if (bp->b_flags & B_REMFREE) 1469 bremfreel(bp); 1470 if (bp->b_qindex != QUEUE_NONE) 1471 panic("bqrelse: free buffer onto another queue???"); 1472 /* buffers with stale but valid contents */ 1473 if (bp->b_flags & B_DELWRI) { 1474 if (bp->b_flags & B_NEEDSGIANT) 1475 bp->b_qindex = QUEUE_DIRTY_GIANT; 1476 else 1477 bp->b_qindex = QUEUE_DIRTY; 1478 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1479 } else { 1480 /* 1481 * XXX This lock may not be necessary since BKGRDINPROG 1482 * cannot be set while we hold the buf lock, it can only be 1483 * cleared if it is already pending. 1484 */ 1485 BO_LOCK(bp->b_bufobj); 1486 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1487 BO_UNLOCK(bp->b_bufobj); 1488 bp->b_qindex = QUEUE_CLEAN; 1489 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1490 b_freelist); 1491 } else { 1492 /* 1493 * We are too low on memory, we have to try to free 1494 * the buffer (most importantly: the wired pages 1495 * making up its backing store) *now*. 1496 */ 1497 BO_UNLOCK(bp->b_bufobj); 1498 mtx_unlock(&bqlock); 1499 brelse(bp); 1500 return; 1501 } 1502 } 1503 mtx_unlock(&bqlock); 1504 1505 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1506 bufcountwakeup(); 1507 1508 /* 1509 * Something we can maybe free or reuse. 1510 */ 1511 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1512 bufspacewakeup(); 1513 1514 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1515 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1516 panic("bqrelse: not dirty"); 1517 /* unlock */ 1518 BUF_UNLOCK(bp); 1519} 1520 1521/* Give pages used by the bp back to the VM system (where possible) */ 1522static void 1523vfs_vmio_release(struct buf *bp) 1524{ 1525 int i; 1526 vm_page_t m; 1527 1528 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1529 vm_page_lock_queues(); 1530 for (i = 0; i < bp->b_npages; i++) { 1531 m = bp->b_pages[i]; 1532 bp->b_pages[i] = NULL; 1533 /* 1534 * In order to keep page LRU ordering consistent, put 1535 * everything on the inactive queue. 1536 */ 1537 vm_page_unwire(m, 0); 1538 /* 1539 * We don't mess with busy pages, it is 1540 * the responsibility of the process that 1541 * busied the pages to deal with them. 1542 */ 1543 if ((m->oflags & VPO_BUSY) || (m->busy != 0)) 1544 continue; 1545 1546 if (m->wire_count == 0) { 1547 /* 1548 * Might as well free the page if we can and it has 1549 * no valid data. We also free the page if the 1550 * buffer was used for direct I/O 1551 */ 1552 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1553 m->hold_count == 0) { 1554 vm_page_free(m); 1555 } else if (bp->b_flags & B_DIRECT) { 1556 vm_page_try_to_free(m); 1557 } else if (vm_page_count_severe()) { 1558 vm_page_try_to_cache(m); 1559 } 1560 } 1561 } 1562 vm_page_unlock_queues(); 1563 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1564 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1565 1566 if (bp->b_bufsize) { 1567 bufspacewakeup(); 1568 bp->b_bufsize = 0; 1569 } 1570 bp->b_npages = 0; 1571 bp->b_flags &= ~B_VMIO; 1572 if (bp->b_vp) 1573 (void) brelvp(bp); 1574} 1575 1576/* 1577 * Check to see if a block at a particular lbn is available for a clustered 1578 * write. 1579 */ 1580static int 1581vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1582{ 1583 struct buf *bpa; 1584 int match; 1585 1586 match = 0; 1587 1588 /* If the buf isn't in core skip it */ 1589 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1590 return (0); 1591 1592 /* If the buf is busy we don't want to wait for it */ 1593 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1594 return (0); 1595 1596 /* Only cluster with valid clusterable delayed write buffers */ 1597 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1598 (B_DELWRI | B_CLUSTEROK)) 1599 goto done; 1600 1601 if (bpa->b_bufsize != size) 1602 goto done; 1603 1604 /* 1605 * Check to see if it is in the expected place on disk and that the 1606 * block has been mapped. 1607 */ 1608 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1609 match = 1; 1610done: 1611 BUF_UNLOCK(bpa); 1612 return (match); 1613} 1614 1615/* 1616 * vfs_bio_awrite: 1617 * 1618 * Implement clustered async writes for clearing out B_DELWRI buffers. 1619 * This is much better then the old way of writing only one buffer at 1620 * a time. Note that we may not be presented with the buffers in the 1621 * correct order, so we search for the cluster in both directions. 1622 */ 1623int 1624vfs_bio_awrite(struct buf *bp) 1625{ 1626 int i; 1627 int j; 1628 daddr_t lblkno = bp->b_lblkno; 1629 struct vnode *vp = bp->b_vp; 1630 int ncl; 1631 int nwritten; 1632 int size; 1633 int maxcl; 1634 1635 /* 1636 * right now we support clustered writing only to regular files. If 1637 * we find a clusterable block we could be in the middle of a cluster 1638 * rather then at the beginning. 1639 */ 1640 if ((vp->v_type == VREG) && 1641 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1642 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1643 1644 size = vp->v_mount->mnt_stat.f_iosize; 1645 maxcl = MAXPHYS / size; 1646 1647 VI_LOCK(vp); 1648 for (i = 1; i < maxcl; i++) 1649 if (vfs_bio_clcheck(vp, size, lblkno + i, 1650 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1651 break; 1652 1653 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1654 if (vfs_bio_clcheck(vp, size, lblkno - j, 1655 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1656 break; 1657 1658 VI_UNLOCK(vp); 1659 --j; 1660 ncl = i + j; 1661 /* 1662 * this is a possible cluster write 1663 */ 1664 if (ncl != 1) { 1665 BUF_UNLOCK(bp); 1666 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1667 return nwritten; 1668 } 1669 } 1670 bremfree(bp); 1671 bp->b_flags |= B_ASYNC; 1672 /* 1673 * default (old) behavior, writing out only one block 1674 * 1675 * XXX returns b_bufsize instead of b_bcount for nwritten? 1676 */ 1677 nwritten = bp->b_bufsize; 1678 (void) bwrite(bp); 1679 1680 return nwritten; 1681} 1682 1683/* 1684 * getnewbuf: 1685 * 1686 * Find and initialize a new buffer header, freeing up existing buffers 1687 * in the bufqueues as necessary. The new buffer is returned locked. 1688 * 1689 * Important: B_INVAL is not set. If the caller wishes to throw the 1690 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1691 * 1692 * We block if: 1693 * We have insufficient buffer headers 1694 * We have insufficient buffer space 1695 * buffer_map is too fragmented ( space reservation fails ) 1696 * If we have to flush dirty buffers ( but we try to avoid this ) 1697 * 1698 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1699 * Instead we ask the buf daemon to do it for us. We attempt to 1700 * avoid piecemeal wakeups of the pageout daemon. 1701 */ 1702 1703static struct buf * 1704getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1705{ 1706 struct buf *bp; 1707 struct buf *nbp; 1708 int defrag = 0; 1709 int nqindex; 1710 int waiters = 0; 1711 static int flushingbufs; 1712 1713 /* 1714 * We can't afford to block since we might be holding a vnode lock, 1715 * which may prevent system daemons from running. We deal with 1716 * low-memory situations by proactively returning memory and running 1717 * async I/O rather then sync I/O. 1718 */ 1719 1720 atomic_add_int(&getnewbufcalls, 1); 1721 atomic_subtract_int(&getnewbufrestarts, 1); 1722restart: 1723 atomic_add_int(&getnewbufrestarts, 1); 1724 1725 /* 1726 * Setup for scan. If we do not have enough free buffers, 1727 * we setup a degenerate case that immediately fails. Note 1728 * that if we are specially marked process, we are allowed to 1729 * dip into our reserves. 1730 * 1731 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1732 * 1733 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1734 * However, there are a number of cases (defragging, reusing, ...) 1735 * where we cannot backup. 1736 */ 1737 mtx_lock(&bqlock); 1738 nqindex = QUEUE_EMPTYKVA; 1739 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1740 1741 if (nbp == NULL) { 1742 /* 1743 * If no EMPTYKVA buffers and we are either 1744 * defragging or reusing, locate a CLEAN buffer 1745 * to free or reuse. If bufspace useage is low 1746 * skip this step so we can allocate a new buffer. 1747 */ 1748 if (defrag || bufspace >= lobufspace) { 1749 nqindex = QUEUE_CLEAN; 1750 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1751 } 1752 1753 /* 1754 * If we could not find or were not allowed to reuse a 1755 * CLEAN buffer, check to see if it is ok to use an EMPTY 1756 * buffer. We can only use an EMPTY buffer if allocating 1757 * its KVA would not otherwise run us out of buffer space. 1758 */ 1759 if (nbp == NULL && defrag == 0 && 1760 bufspace + maxsize < hibufspace) { 1761 nqindex = QUEUE_EMPTY; 1762 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1763 } 1764 } 1765 1766 /* 1767 * Run scan, possibly freeing data and/or kva mappings on the fly 1768 * depending. 1769 */ 1770 1771 while ((bp = nbp) != NULL) { 1772 int qindex = nqindex; 1773 1774 /* 1775 * Calculate next bp ( we can only use it if we do not block 1776 * or do other fancy things ). 1777 */ 1778 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1779 switch(qindex) { 1780 case QUEUE_EMPTY: 1781 nqindex = QUEUE_EMPTYKVA; 1782 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1783 break; 1784 /* FALLTHROUGH */ 1785 case QUEUE_EMPTYKVA: 1786 nqindex = QUEUE_CLEAN; 1787 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1788 break; 1789 /* FALLTHROUGH */ 1790 case QUEUE_CLEAN: 1791 /* 1792 * nbp is NULL. 1793 */ 1794 break; 1795 } 1796 } 1797 /* 1798 * If we are defragging then we need a buffer with 1799 * b_kvasize != 0. XXX this situation should no longer 1800 * occur, if defrag is non-zero the buffer's b_kvasize 1801 * should also be non-zero at this point. XXX 1802 */ 1803 if (defrag && bp->b_kvasize == 0) { 1804 printf("Warning: defrag empty buffer %p\n", bp); 1805 continue; 1806 } 1807 1808 /* 1809 * Start freeing the bp. This is somewhat involved. nbp 1810 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1811 */ 1812 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1813 continue; 1814 if (bp->b_vp) { 1815 BO_LOCK(bp->b_bufobj); 1816 if (bp->b_vflags & BV_BKGRDINPROG) { 1817 BO_UNLOCK(bp->b_bufobj); 1818 BUF_UNLOCK(bp); 1819 continue; 1820 } 1821 BO_UNLOCK(bp->b_bufobj); 1822 } 1823 CTR6(KTR_BUF, 1824 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1825 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1826 bp->b_kvasize, bp->b_bufsize, qindex); 1827 1828 /* 1829 * Sanity Checks 1830 */ 1831 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1832 1833 /* 1834 * Note: we no longer distinguish between VMIO and non-VMIO 1835 * buffers. 1836 */ 1837 1838 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1839 1840 bremfreel(bp); 1841 mtx_unlock(&bqlock); 1842 1843 if (qindex == QUEUE_CLEAN) { 1844 if (bp->b_flags & B_VMIO) { 1845 bp->b_flags &= ~B_ASYNC; 1846 vfs_vmio_release(bp); 1847 } 1848 if (bp->b_vp) 1849 waiters = brelvp(bp); 1850 } 1851 1852 /* 1853 * NOTE: nbp is now entirely invalid. We can only restart 1854 * the scan from this point on. 1855 * 1856 * Get the rest of the buffer freed up. b_kva* is still 1857 * valid after this operation. 1858 */ 1859 1860 if (bp->b_rcred != NOCRED) { 1861 crfree(bp->b_rcred); 1862 bp->b_rcred = NOCRED; 1863 } 1864 if (bp->b_wcred != NOCRED) { 1865 crfree(bp->b_wcred); 1866 bp->b_wcred = NOCRED; 1867 } 1868 if (!LIST_EMPTY(&bp->b_dep)) 1869 buf_deallocate(bp); 1870 if (bp->b_vflags & BV_BKGRDINPROG) 1871 panic("losing buffer 3"); 1872 KASSERT(bp->b_vp == NULL, 1873 ("bp: %p still has vnode %p. qindex: %d", 1874 bp, bp->b_vp, qindex)); 1875 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1876 ("bp: %p still on a buffer list. xflags %X", 1877 bp, bp->b_xflags)); 1878 1879 if (bp->b_bufsize) 1880 allocbuf(bp, 0); 1881 1882 bp->b_flags = 0; 1883 bp->b_ioflags = 0; 1884 bp->b_xflags = 0; 1885 bp->b_vflags = 0; 1886 bp->b_vp = NULL; 1887 bp->b_blkno = bp->b_lblkno = 0; 1888 bp->b_offset = NOOFFSET; 1889 bp->b_iodone = 0; 1890 bp->b_error = 0; 1891 bp->b_resid = 0; 1892 bp->b_bcount = 0; 1893 bp->b_npages = 0; 1894 bp->b_dirtyoff = bp->b_dirtyend = 0; 1895 bp->b_bufobj = NULL; 1896 bp->b_pin_count = 0; 1897 bp->b_fsprivate1 = NULL; 1898 bp->b_fsprivate2 = NULL; 1899 bp->b_fsprivate3 = NULL; 1900 1901 LIST_INIT(&bp->b_dep); 1902 1903 /* 1904 * If we are defragging then free the buffer. 1905 */ 1906 if (defrag) { 1907 bp->b_flags |= B_INVAL; 1908 bfreekva(bp); 1909 brelse(bp); 1910 defrag = 0; 1911 goto restart; 1912 } 1913 1914 /* 1915 * Notify any waiters for the buffer lock about 1916 * identity change by freeing the buffer. 1917 */ 1918 if (qindex == QUEUE_CLEAN && waiters > 0) { 1919 bp->b_flags |= B_INVAL; 1920 bfreekva(bp); 1921 brelse(bp); 1922 goto restart; 1923 } 1924 1925 /* 1926 * If we are overcomitted then recover the buffer and its 1927 * KVM space. This occurs in rare situations when multiple 1928 * processes are blocked in getnewbuf() or allocbuf(). 1929 */ 1930 if (bufspace >= hibufspace) 1931 flushingbufs = 1; 1932 if (flushingbufs && bp->b_kvasize != 0) { 1933 bp->b_flags |= B_INVAL; 1934 bfreekva(bp); 1935 brelse(bp); 1936 goto restart; 1937 } 1938 if (bufspace < lobufspace) 1939 flushingbufs = 0; 1940 break; 1941 } 1942 1943 /* 1944 * If we exhausted our list, sleep as appropriate. We may have to 1945 * wakeup various daemons and write out some dirty buffers. 1946 * 1947 * Generally we are sleeping due to insufficient buffer space. 1948 */ 1949 1950 if (bp == NULL) { 1951 int flags; 1952 char *waitmsg; 1953 1954 if (defrag) { 1955 flags = VFS_BIO_NEED_BUFSPACE; 1956 waitmsg = "nbufkv"; 1957 } else if (bufspace >= hibufspace) { 1958 waitmsg = "nbufbs"; 1959 flags = VFS_BIO_NEED_BUFSPACE; 1960 } else { 1961 waitmsg = "newbuf"; 1962 flags = VFS_BIO_NEED_ANY; 1963 } 1964 mtx_lock(&nblock); 1965 needsbuffer |= flags; 1966 mtx_unlock(&nblock); 1967 mtx_unlock(&bqlock); 1968 1969 bd_speedup(); /* heeeelp */ 1970 1971 mtx_lock(&nblock); 1972 while (needsbuffer & flags) { 1973 if (msleep(&needsbuffer, &nblock, 1974 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1975 mtx_unlock(&nblock); 1976 return (NULL); 1977 } 1978 } 1979 mtx_unlock(&nblock); 1980 } else { 1981 /* 1982 * We finally have a valid bp. We aren't quite out of the 1983 * woods, we still have to reserve kva space. In order 1984 * to keep fragmentation sane we only allocate kva in 1985 * BKVASIZE chunks. 1986 */ 1987 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1988 1989 if (maxsize != bp->b_kvasize) { 1990 vm_offset_t addr = 0; 1991 1992 bfreekva(bp); 1993 1994 vm_map_lock(buffer_map); 1995 if (vm_map_findspace(buffer_map, 1996 vm_map_min(buffer_map), maxsize, &addr)) { 1997 /* 1998 * Uh oh. Buffer map is to fragmented. We 1999 * must defragment the map. 2000 */ 2001 atomic_add_int(&bufdefragcnt, 1); 2002 vm_map_unlock(buffer_map); 2003 defrag = 1; 2004 bp->b_flags |= B_INVAL; 2005 brelse(bp); 2006 goto restart; 2007 } 2008 if (addr) { 2009 vm_map_insert(buffer_map, NULL, 0, 2010 addr, addr + maxsize, 2011 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2012 2013 bp->b_kvabase = (caddr_t) addr; 2014 bp->b_kvasize = maxsize; 2015 atomic_add_int(&bufspace, bp->b_kvasize); 2016 atomic_add_int(&bufreusecnt, 1); 2017 } 2018 vm_map_unlock(buffer_map); 2019 } 2020 bp->b_saveaddr = bp->b_kvabase; 2021 bp->b_data = bp->b_saveaddr; 2022 } 2023 return(bp); 2024} 2025 2026/* 2027 * buf_daemon: 2028 * 2029 * buffer flushing daemon. Buffers are normally flushed by the 2030 * update daemon but if it cannot keep up this process starts to 2031 * take the load in an attempt to prevent getnewbuf() from blocking. 2032 */ 2033 2034static struct kproc_desc buf_kp = { 2035 "bufdaemon", 2036 buf_daemon, 2037 &bufdaemonproc 2038}; 2039SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2040 2041static void 2042buf_daemon() 2043{ 2044 2045 /* 2046 * This process needs to be suspended prior to shutdown sync. 2047 */ 2048 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2049 SHUTDOWN_PRI_LAST); 2050 2051 /* 2052 * This process is allowed to take the buffer cache to the limit 2053 */ 2054 curthread->td_pflags |= TDP_NORUNNINGBUF; 2055 mtx_lock(&bdlock); 2056 for (;;) { 2057 bd_request = 0; 2058 mtx_unlock(&bdlock); 2059 2060 kproc_suspend_check(bufdaemonproc); 2061 2062 /* 2063 * Do the flush. Limit the amount of in-transit I/O we 2064 * allow to build up, otherwise we would completely saturate 2065 * the I/O system. Wakeup any waiting processes before we 2066 * normally would so they can run in parallel with our drain. 2067 */ 2068 while (numdirtybuffers > lodirtybuffers) { 2069 int flushed; 2070 2071 flushed = flushbufqueues(QUEUE_DIRTY, 0); 2072 /* The list empty check here is slightly racy */ 2073 if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) { 2074 mtx_lock(&Giant); 2075 flushed += flushbufqueues(QUEUE_DIRTY_GIANT, 0); 2076 mtx_unlock(&Giant); 2077 } 2078 if (flushed == 0) { 2079 /* 2080 * Could not find any buffers without rollback 2081 * dependencies, so just write the first one 2082 * in the hopes of eventually making progress. 2083 */ 2084 flushbufqueues(QUEUE_DIRTY, 1); 2085 if (!TAILQ_EMPTY( 2086 &bufqueues[QUEUE_DIRTY_GIANT])) { 2087 mtx_lock(&Giant); 2088 flushbufqueues(QUEUE_DIRTY_GIANT, 1); 2089 mtx_unlock(&Giant); 2090 } 2091 break; 2092 } 2093 uio_yield(); 2094 } 2095 2096 /* 2097 * Only clear bd_request if we have reached our low water 2098 * mark. The buf_daemon normally waits 1 second and 2099 * then incrementally flushes any dirty buffers that have 2100 * built up, within reason. 2101 * 2102 * If we were unable to hit our low water mark and couldn't 2103 * find any flushable buffers, we sleep half a second. 2104 * Otherwise we loop immediately. 2105 */ 2106 mtx_lock(&bdlock); 2107 if (numdirtybuffers <= lodirtybuffers) { 2108 /* 2109 * We reached our low water mark, reset the 2110 * request and sleep until we are needed again. 2111 * The sleep is just so the suspend code works. 2112 */ 2113 bd_request = 0; 2114 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2115 } else { 2116 /* 2117 * We couldn't find any flushable dirty buffers but 2118 * still have too many dirty buffers, we 2119 * have to sleep and try again. (rare) 2120 */ 2121 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2122 } 2123 } 2124} 2125 2126/* 2127 * flushbufqueues: 2128 * 2129 * Try to flush a buffer in the dirty queue. We must be careful to 2130 * free up B_INVAL buffers instead of write them, which NFS is 2131 * particularly sensitive to. 2132 */ 2133static int flushwithdeps = 0; 2134SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2135 0, "Number of buffers flushed with dependecies that require rollbacks"); 2136 2137static int 2138flushbufqueues(int queue, int flushdeps) 2139{ 2140 struct buf sentinel; 2141 struct vnode *vp; 2142 struct mount *mp; 2143 struct buf *bp; 2144 int hasdeps; 2145 int flushed; 2146 int target; 2147 2148 target = numdirtybuffers - lodirtybuffers; 2149 if (flushdeps && target > 2) 2150 target /= 2; 2151 flushed = 0; 2152 bp = NULL; 2153 mtx_lock(&bqlock); 2154 TAILQ_INSERT_TAIL(&bufqueues[queue], &sentinel, b_freelist); 2155 while (flushed != target) { 2156 bp = TAILQ_FIRST(&bufqueues[queue]); 2157 if (bp == &sentinel) 2158 break; 2159 TAILQ_REMOVE(&bufqueues[queue], bp, b_freelist); 2160 TAILQ_INSERT_TAIL(&bufqueues[queue], bp, b_freelist); 2161 2162 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2163 continue; 2164 if (bp->b_pin_count > 0) { 2165 BUF_UNLOCK(bp); 2166 continue; 2167 } 2168 BO_LOCK(bp->b_bufobj); 2169 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2170 (bp->b_flags & B_DELWRI) == 0) { 2171 BO_UNLOCK(bp->b_bufobj); 2172 BUF_UNLOCK(bp); 2173 continue; 2174 } 2175 BO_UNLOCK(bp->b_bufobj); 2176 if (bp->b_flags & B_INVAL) { 2177 bremfreel(bp); 2178 mtx_unlock(&bqlock); 2179 brelse(bp); 2180 flushed++; 2181 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2182 mtx_lock(&bqlock); 2183 continue; 2184 } 2185 2186 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2187 if (flushdeps == 0) { 2188 BUF_UNLOCK(bp); 2189 continue; 2190 } 2191 hasdeps = 1; 2192 } else 2193 hasdeps = 0; 2194 /* 2195 * We must hold the lock on a vnode before writing 2196 * one of its buffers. Otherwise we may confuse, or 2197 * in the case of a snapshot vnode, deadlock the 2198 * system. 2199 * 2200 * The lock order here is the reverse of the normal 2201 * of vnode followed by buf lock. This is ok because 2202 * the NOWAIT will prevent deadlock. 2203 */ 2204 vp = bp->b_vp; 2205 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2206 BUF_UNLOCK(bp); 2207 continue; 2208 } 2209 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 2210 mtx_unlock(&bqlock); 2211 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2212 bp, bp->b_vp, bp->b_flags); 2213 vfs_bio_awrite(bp); 2214 vn_finished_write(mp); 2215 VOP_UNLOCK(vp, 0); 2216 flushwithdeps += hasdeps; 2217 flushed++; 2218 waitrunningbufspace(); 2219 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2220 mtx_lock(&bqlock); 2221 continue; 2222 } 2223 vn_finished_write(mp); 2224 BUF_UNLOCK(bp); 2225 } 2226 TAILQ_REMOVE(&bufqueues[queue], &sentinel, b_freelist); 2227 mtx_unlock(&bqlock); 2228 return (flushed); 2229} 2230 2231/* 2232 * Check to see if a block is currently memory resident. 2233 */ 2234struct buf * 2235incore(struct bufobj *bo, daddr_t blkno) 2236{ 2237 struct buf *bp; 2238 2239 BO_LOCK(bo); 2240 bp = gbincore(bo, blkno); 2241 BO_UNLOCK(bo); 2242 return (bp); 2243} 2244 2245/* 2246 * Returns true if no I/O is needed to access the 2247 * associated VM object. This is like incore except 2248 * it also hunts around in the VM system for the data. 2249 */ 2250 2251static int 2252inmem(struct vnode * vp, daddr_t blkno) 2253{ 2254 vm_object_t obj; 2255 vm_offset_t toff, tinc, size; 2256 vm_page_t m; 2257 vm_ooffset_t off; 2258 2259 ASSERT_VOP_LOCKED(vp, "inmem"); 2260 2261 if (incore(&vp->v_bufobj, blkno)) 2262 return 1; 2263 if (vp->v_mount == NULL) 2264 return 0; 2265 obj = vp->v_object; 2266 if (obj == NULL) 2267 return (0); 2268 2269 size = PAGE_SIZE; 2270 if (size > vp->v_mount->mnt_stat.f_iosize) 2271 size = vp->v_mount->mnt_stat.f_iosize; 2272 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2273 2274 VM_OBJECT_LOCK(obj); 2275 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2276 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2277 if (!m) 2278 goto notinmem; 2279 tinc = size; 2280 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2281 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2282 if (vm_page_is_valid(m, 2283 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2284 goto notinmem; 2285 } 2286 VM_OBJECT_UNLOCK(obj); 2287 return 1; 2288 2289notinmem: 2290 VM_OBJECT_UNLOCK(obj); 2291 return (0); 2292} 2293 2294/* 2295 * vfs_setdirty: 2296 * 2297 * Sets the dirty range for a buffer based on the status of the dirty 2298 * bits in the pages comprising the buffer. 2299 * 2300 * The range is limited to the size of the buffer. 2301 * 2302 * This routine is primarily used by NFS, but is generalized for the 2303 * B_VMIO case. 2304 */ 2305static void 2306vfs_setdirty(struct buf *bp) 2307{ 2308 2309 /* 2310 * Degenerate case - empty buffer 2311 */ 2312 2313 if (bp->b_bufsize == 0) 2314 return; 2315 2316 /* 2317 * We qualify the scan for modified pages on whether the 2318 * object has been flushed yet. 2319 */ 2320 2321 if ((bp->b_flags & B_VMIO) == 0) 2322 return; 2323 2324 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2325 vfs_setdirty_locked_object(bp); 2326 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2327} 2328 2329static void 2330vfs_setdirty_locked_object(struct buf *bp) 2331{ 2332 vm_object_t object; 2333 int i; 2334 2335 object = bp->b_bufobj->bo_object; 2336 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2337 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2338 vm_offset_t boffset; 2339 vm_offset_t eoffset; 2340 2341 vm_page_lock_queues(); 2342 /* 2343 * test the pages to see if they have been modified directly 2344 * by users through the VM system. 2345 */ 2346 for (i = 0; i < bp->b_npages; i++) 2347 vm_page_test_dirty(bp->b_pages[i]); 2348 2349 /* 2350 * Calculate the encompassing dirty range, boffset and eoffset, 2351 * (eoffset - boffset) bytes. 2352 */ 2353 2354 for (i = 0; i < bp->b_npages; i++) { 2355 if (bp->b_pages[i]->dirty) 2356 break; 2357 } 2358 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2359 2360 for (i = bp->b_npages - 1; i >= 0; --i) { 2361 if (bp->b_pages[i]->dirty) { 2362 break; 2363 } 2364 } 2365 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2366 2367 vm_page_unlock_queues(); 2368 /* 2369 * Fit it to the buffer. 2370 */ 2371 2372 if (eoffset > bp->b_bcount) 2373 eoffset = bp->b_bcount; 2374 2375 /* 2376 * If we have a good dirty range, merge with the existing 2377 * dirty range. 2378 */ 2379 2380 if (boffset < eoffset) { 2381 if (bp->b_dirtyoff > boffset) 2382 bp->b_dirtyoff = boffset; 2383 if (bp->b_dirtyend < eoffset) 2384 bp->b_dirtyend = eoffset; 2385 } 2386 } 2387} 2388 2389/* 2390 * getblk: 2391 * 2392 * Get a block given a specified block and offset into a file/device. 2393 * The buffers B_DONE bit will be cleared on return, making it almost 2394 * ready for an I/O initiation. B_INVAL may or may not be set on 2395 * return. The caller should clear B_INVAL prior to initiating a 2396 * READ. 2397 * 2398 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2399 * an existing buffer. 2400 * 2401 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2402 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2403 * and then cleared based on the backing VM. If the previous buffer is 2404 * non-0-sized but invalid, B_CACHE will be cleared. 2405 * 2406 * If getblk() must create a new buffer, the new buffer is returned with 2407 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2408 * case it is returned with B_INVAL clear and B_CACHE set based on the 2409 * backing VM. 2410 * 2411 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2412 * B_CACHE bit is clear. 2413 * 2414 * What this means, basically, is that the caller should use B_CACHE to 2415 * determine whether the buffer is fully valid or not and should clear 2416 * B_INVAL prior to issuing a read. If the caller intends to validate 2417 * the buffer by loading its data area with something, the caller needs 2418 * to clear B_INVAL. If the caller does this without issuing an I/O, 2419 * the caller should set B_CACHE ( as an optimization ), else the caller 2420 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2421 * a write attempt or if it was a successfull read. If the caller 2422 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2423 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2424 */ 2425struct buf * 2426getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2427 int flags) 2428{ 2429 struct buf *bp; 2430 struct bufobj *bo; 2431 int error; 2432 2433 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2434 ASSERT_VOP_LOCKED(vp, "getblk"); 2435 if (size > MAXBSIZE) 2436 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2437 2438 bo = &vp->v_bufobj; 2439loop: 2440 /* 2441 * Block if we are low on buffers. Certain processes are allowed 2442 * to completely exhaust the buffer cache. 2443 * 2444 * If this check ever becomes a bottleneck it may be better to 2445 * move it into the else, when gbincore() fails. At the moment 2446 * it isn't a problem. 2447 * 2448 * XXX remove if 0 sections (clean this up after its proven) 2449 */ 2450 if (numfreebuffers == 0) { 2451 if (TD_IS_IDLETHREAD(curthread)) 2452 return NULL; 2453 mtx_lock(&nblock); 2454 needsbuffer |= VFS_BIO_NEED_ANY; 2455 mtx_unlock(&nblock); 2456 } 2457 2458 BO_LOCK(bo); 2459 bp = gbincore(bo, blkno); 2460 if (bp != NULL) { 2461 int lockflags; 2462 /* 2463 * Buffer is in-core. If the buffer is not busy, it must 2464 * be on a queue. 2465 */ 2466 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2467 2468 if (flags & GB_LOCK_NOWAIT) 2469 lockflags |= LK_NOWAIT; 2470 2471 error = BUF_TIMELOCK(bp, lockflags, 2472 VI_MTX(vp), "getblk", slpflag, slptimeo); 2473 2474 /* 2475 * If we slept and got the lock we have to restart in case 2476 * the buffer changed identities. 2477 */ 2478 if (error == ENOLCK) 2479 goto loop; 2480 /* We timed out or were interrupted. */ 2481 else if (error) 2482 return (NULL); 2483 2484 /* 2485 * The buffer is locked. B_CACHE is cleared if the buffer is 2486 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2487 * and for a VMIO buffer B_CACHE is adjusted according to the 2488 * backing VM cache. 2489 */ 2490 if (bp->b_flags & B_INVAL) 2491 bp->b_flags &= ~B_CACHE; 2492 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2493 bp->b_flags |= B_CACHE; 2494 bremfree(bp); 2495 2496 /* 2497 * check for size inconsistancies for non-VMIO case. 2498 */ 2499 2500 if (bp->b_bcount != size) { 2501 if ((bp->b_flags & B_VMIO) == 0 || 2502 (size > bp->b_kvasize)) { 2503 if (bp->b_flags & B_DELWRI) { 2504 /* 2505 * If buffer is pinned and caller does 2506 * not want sleep waiting for it to be 2507 * unpinned, bail out 2508 * */ 2509 if (bp->b_pin_count > 0) { 2510 if (flags & GB_LOCK_NOWAIT) { 2511 bqrelse(bp); 2512 return (NULL); 2513 } else { 2514 bunpin_wait(bp); 2515 } 2516 } 2517 bp->b_flags |= B_NOCACHE; 2518 bwrite(bp); 2519 } else { 2520 if (LIST_EMPTY(&bp->b_dep)) { 2521 bp->b_flags |= B_RELBUF; 2522 brelse(bp); 2523 } else { 2524 bp->b_flags |= B_NOCACHE; 2525 bwrite(bp); 2526 } 2527 } 2528 goto loop; 2529 } 2530 } 2531 2532 /* 2533 * If the size is inconsistant in the VMIO case, we can resize 2534 * the buffer. This might lead to B_CACHE getting set or 2535 * cleared. If the size has not changed, B_CACHE remains 2536 * unchanged from its previous state. 2537 */ 2538 2539 if (bp->b_bcount != size) 2540 allocbuf(bp, size); 2541 2542 KASSERT(bp->b_offset != NOOFFSET, 2543 ("getblk: no buffer offset")); 2544 2545 /* 2546 * A buffer with B_DELWRI set and B_CACHE clear must 2547 * be committed before we can return the buffer in 2548 * order to prevent the caller from issuing a read 2549 * ( due to B_CACHE not being set ) and overwriting 2550 * it. 2551 * 2552 * Most callers, including NFS and FFS, need this to 2553 * operate properly either because they assume they 2554 * can issue a read if B_CACHE is not set, or because 2555 * ( for example ) an uncached B_DELWRI might loop due 2556 * to softupdates re-dirtying the buffer. In the latter 2557 * case, B_CACHE is set after the first write completes, 2558 * preventing further loops. 2559 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2560 * above while extending the buffer, we cannot allow the 2561 * buffer to remain with B_CACHE set after the write 2562 * completes or it will represent a corrupt state. To 2563 * deal with this we set B_NOCACHE to scrap the buffer 2564 * after the write. 2565 * 2566 * We might be able to do something fancy, like setting 2567 * B_CACHE in bwrite() except if B_DELWRI is already set, 2568 * so the below call doesn't set B_CACHE, but that gets real 2569 * confusing. This is much easier. 2570 */ 2571 2572 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2573 bp->b_flags |= B_NOCACHE; 2574 bwrite(bp); 2575 goto loop; 2576 } 2577 bp->b_flags &= ~B_DONE; 2578 } else { 2579 int bsize, maxsize, vmio; 2580 off_t offset; 2581 2582 /* 2583 * Buffer is not in-core, create new buffer. The buffer 2584 * returned by getnewbuf() is locked. Note that the returned 2585 * buffer is also considered valid (not marked B_INVAL). 2586 */ 2587 BO_UNLOCK(bo); 2588 /* 2589 * If the user does not want us to create the buffer, bail out 2590 * here. 2591 */ 2592 if (flags & GB_NOCREAT) 2593 return NULL; 2594 bsize = bo->bo_bsize; 2595 offset = blkno * bsize; 2596 vmio = vp->v_object != NULL; 2597 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2598 maxsize = imax(maxsize, bsize); 2599 2600 bp = getnewbuf(slpflag, slptimeo, size, maxsize); 2601 if (bp == NULL) { 2602 if (slpflag || slptimeo) 2603 return NULL; 2604 goto loop; 2605 } 2606 2607 /* 2608 * This code is used to make sure that a buffer is not 2609 * created while the getnewbuf routine is blocked. 2610 * This can be a problem whether the vnode is locked or not. 2611 * If the buffer is created out from under us, we have to 2612 * throw away the one we just created. 2613 * 2614 * Note: this must occur before we associate the buffer 2615 * with the vp especially considering limitations in 2616 * the splay tree implementation when dealing with duplicate 2617 * lblkno's. 2618 */ 2619 BO_LOCK(bo); 2620 if (gbincore(bo, blkno)) { 2621 BO_UNLOCK(bo); 2622 bp->b_flags |= B_INVAL; 2623 brelse(bp); 2624 goto loop; 2625 } 2626 2627 /* 2628 * Insert the buffer into the hash, so that it can 2629 * be found by incore. 2630 */ 2631 bp->b_blkno = bp->b_lblkno = blkno; 2632 bp->b_offset = offset; 2633 bgetvp(vp, bp); 2634 BO_UNLOCK(bo); 2635 2636 /* 2637 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2638 * buffer size starts out as 0, B_CACHE will be set by 2639 * allocbuf() for the VMIO case prior to it testing the 2640 * backing store for validity. 2641 */ 2642 2643 if (vmio) { 2644 bp->b_flags |= B_VMIO; 2645#if defined(VFS_BIO_DEBUG) 2646 if (vn_canvmio(vp) != TRUE) 2647 printf("getblk: VMIO on vnode type %d\n", 2648 vp->v_type); 2649#endif 2650 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2651 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2652 bp, vp->v_object, bp->b_bufobj->bo_object)); 2653 } else { 2654 bp->b_flags &= ~B_VMIO; 2655 KASSERT(bp->b_bufobj->bo_object == NULL, 2656 ("ARGH! has b_bufobj->bo_object %p %p\n", 2657 bp, bp->b_bufobj->bo_object)); 2658 } 2659 2660 allocbuf(bp, size); 2661 bp->b_flags &= ~B_DONE; 2662 } 2663 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2664 BUF_ASSERT_HELD(bp); 2665 KASSERT(bp->b_bufobj == bo, 2666 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2667 return (bp); 2668} 2669 2670/* 2671 * Get an empty, disassociated buffer of given size. The buffer is initially 2672 * set to B_INVAL. 2673 */ 2674struct buf * 2675geteblk(int size) 2676{ 2677 struct buf *bp; 2678 int maxsize; 2679 2680 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2681 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2682 continue; 2683 allocbuf(bp, size); 2684 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2685 BUF_ASSERT_HELD(bp); 2686 return (bp); 2687} 2688 2689 2690/* 2691 * This code constitutes the buffer memory from either anonymous system 2692 * memory (in the case of non-VMIO operations) or from an associated 2693 * VM object (in the case of VMIO operations). This code is able to 2694 * resize a buffer up or down. 2695 * 2696 * Note that this code is tricky, and has many complications to resolve 2697 * deadlock or inconsistant data situations. Tread lightly!!! 2698 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2699 * the caller. Calling this code willy nilly can result in the loss of data. 2700 * 2701 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2702 * B_CACHE for the non-VMIO case. 2703 */ 2704 2705int 2706allocbuf(struct buf *bp, int size) 2707{ 2708 int newbsize, mbsize; 2709 int i; 2710 2711 BUF_ASSERT_HELD(bp); 2712 2713 if (bp->b_kvasize < size) 2714 panic("allocbuf: buffer too small"); 2715 2716 if ((bp->b_flags & B_VMIO) == 0) { 2717 caddr_t origbuf; 2718 int origbufsize; 2719 /* 2720 * Just get anonymous memory from the kernel. Don't 2721 * mess with B_CACHE. 2722 */ 2723 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2724 if (bp->b_flags & B_MALLOC) 2725 newbsize = mbsize; 2726 else 2727 newbsize = round_page(size); 2728 2729 if (newbsize < bp->b_bufsize) { 2730 /* 2731 * malloced buffers are not shrunk 2732 */ 2733 if (bp->b_flags & B_MALLOC) { 2734 if (newbsize) { 2735 bp->b_bcount = size; 2736 } else { 2737 free(bp->b_data, M_BIOBUF); 2738 if (bp->b_bufsize) { 2739 atomic_subtract_int( 2740 &bufmallocspace, 2741 bp->b_bufsize); 2742 bufspacewakeup(); 2743 bp->b_bufsize = 0; 2744 } 2745 bp->b_saveaddr = bp->b_kvabase; 2746 bp->b_data = bp->b_saveaddr; 2747 bp->b_bcount = 0; 2748 bp->b_flags &= ~B_MALLOC; 2749 } 2750 return 1; 2751 } 2752 vm_hold_free_pages( 2753 bp, 2754 (vm_offset_t) bp->b_data + newbsize, 2755 (vm_offset_t) bp->b_data + bp->b_bufsize); 2756 } else if (newbsize > bp->b_bufsize) { 2757 /* 2758 * We only use malloced memory on the first allocation. 2759 * and revert to page-allocated memory when the buffer 2760 * grows. 2761 */ 2762 /* 2763 * There is a potential smp race here that could lead 2764 * to bufmallocspace slightly passing the max. It 2765 * is probably extremely rare and not worth worrying 2766 * over. 2767 */ 2768 if ( (bufmallocspace < maxbufmallocspace) && 2769 (bp->b_bufsize == 0) && 2770 (mbsize <= PAGE_SIZE/2)) { 2771 2772 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2773 bp->b_bufsize = mbsize; 2774 bp->b_bcount = size; 2775 bp->b_flags |= B_MALLOC; 2776 atomic_add_int(&bufmallocspace, mbsize); 2777 return 1; 2778 } 2779 origbuf = NULL; 2780 origbufsize = 0; 2781 /* 2782 * If the buffer is growing on its other-than-first allocation, 2783 * then we revert to the page-allocation scheme. 2784 */ 2785 if (bp->b_flags & B_MALLOC) { 2786 origbuf = bp->b_data; 2787 origbufsize = bp->b_bufsize; 2788 bp->b_data = bp->b_kvabase; 2789 if (bp->b_bufsize) { 2790 atomic_subtract_int(&bufmallocspace, 2791 bp->b_bufsize); 2792 bufspacewakeup(); 2793 bp->b_bufsize = 0; 2794 } 2795 bp->b_flags &= ~B_MALLOC; 2796 newbsize = round_page(newbsize); 2797 } 2798 vm_hold_load_pages( 2799 bp, 2800 (vm_offset_t) bp->b_data + bp->b_bufsize, 2801 (vm_offset_t) bp->b_data + newbsize); 2802 if (origbuf) { 2803 bcopy(origbuf, bp->b_data, origbufsize); 2804 free(origbuf, M_BIOBUF); 2805 } 2806 } 2807 } else { 2808 int desiredpages; 2809 2810 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2811 desiredpages = (size == 0) ? 0 : 2812 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2813 2814 if (bp->b_flags & B_MALLOC) 2815 panic("allocbuf: VMIO buffer can't be malloced"); 2816 /* 2817 * Set B_CACHE initially if buffer is 0 length or will become 2818 * 0-length. 2819 */ 2820 if (size == 0 || bp->b_bufsize == 0) 2821 bp->b_flags |= B_CACHE; 2822 2823 if (newbsize < bp->b_bufsize) { 2824 /* 2825 * DEV_BSIZE aligned new buffer size is less then the 2826 * DEV_BSIZE aligned existing buffer size. Figure out 2827 * if we have to remove any pages. 2828 */ 2829 if (desiredpages < bp->b_npages) { 2830 vm_page_t m; 2831 2832 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2833 vm_page_lock_queues(); 2834 for (i = desiredpages; i < bp->b_npages; i++) { 2835 /* 2836 * the page is not freed here -- it 2837 * is the responsibility of 2838 * vnode_pager_setsize 2839 */ 2840 m = bp->b_pages[i]; 2841 KASSERT(m != bogus_page, 2842 ("allocbuf: bogus page found")); 2843 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2844 vm_page_lock_queues(); 2845 2846 bp->b_pages[i] = NULL; 2847 vm_page_unwire(m, 0); 2848 } 2849 vm_page_unlock_queues(); 2850 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2851 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2852 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2853 bp->b_npages = desiredpages; 2854 } 2855 } else if (size > bp->b_bcount) { 2856 /* 2857 * We are growing the buffer, possibly in a 2858 * byte-granular fashion. 2859 */ 2860 struct vnode *vp; 2861 vm_object_t obj; 2862 vm_offset_t toff; 2863 vm_offset_t tinc; 2864 2865 /* 2866 * Step 1, bring in the VM pages from the object, 2867 * allocating them if necessary. We must clear 2868 * B_CACHE if these pages are not valid for the 2869 * range covered by the buffer. 2870 */ 2871 2872 vp = bp->b_vp; 2873 obj = bp->b_bufobj->bo_object; 2874 2875 VM_OBJECT_LOCK(obj); 2876 while (bp->b_npages < desiredpages) { 2877 vm_page_t m; 2878 vm_pindex_t pi; 2879 2880 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2881 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2882 /* 2883 * note: must allocate system pages 2884 * since blocking here could intefere 2885 * with paging I/O, no matter which 2886 * process we are. 2887 */ 2888 m = vm_page_alloc(obj, pi, 2889 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2890 VM_ALLOC_WIRED); 2891 if (m == NULL) { 2892 atomic_add_int(&vm_pageout_deficit, 2893 desiredpages - bp->b_npages); 2894 VM_OBJECT_UNLOCK(obj); 2895 VM_WAIT; 2896 VM_OBJECT_LOCK(obj); 2897 } else { 2898 if (m->valid == 0) 2899 bp->b_flags &= ~B_CACHE; 2900 bp->b_pages[bp->b_npages] = m; 2901 ++bp->b_npages; 2902 } 2903 continue; 2904 } 2905 2906 /* 2907 * We found a page. If we have to sleep on it, 2908 * retry because it might have gotten freed out 2909 * from under us. 2910 * 2911 * We can only test VPO_BUSY here. Blocking on 2912 * m->busy might lead to a deadlock: 2913 * 2914 * vm_fault->getpages->cluster_read->allocbuf 2915 * 2916 */ 2917 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2918 continue; 2919 2920 /* 2921 * We have a good page. 2922 */ 2923 vm_page_lock_queues(); 2924 vm_page_wire(m); 2925 vm_page_unlock_queues(); 2926 bp->b_pages[bp->b_npages] = m; 2927 ++bp->b_npages; 2928 } 2929 2930 /* 2931 * Step 2. We've loaded the pages into the buffer, 2932 * we have to figure out if we can still have B_CACHE 2933 * set. Note that B_CACHE is set according to the 2934 * byte-granular range ( bcount and size ), new the 2935 * aligned range ( newbsize ). 2936 * 2937 * The VM test is against m->valid, which is DEV_BSIZE 2938 * aligned. Needless to say, the validity of the data 2939 * needs to also be DEV_BSIZE aligned. Note that this 2940 * fails with NFS if the server or some other client 2941 * extends the file's EOF. If our buffer is resized, 2942 * B_CACHE may remain set! XXX 2943 */ 2944 2945 toff = bp->b_bcount; 2946 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2947 2948 while ((bp->b_flags & B_CACHE) && toff < size) { 2949 vm_pindex_t pi; 2950 2951 if (tinc > (size - toff)) 2952 tinc = size - toff; 2953 2954 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2955 PAGE_SHIFT; 2956 2957 vfs_buf_test_cache( 2958 bp, 2959 bp->b_offset, 2960 toff, 2961 tinc, 2962 bp->b_pages[pi] 2963 ); 2964 toff += tinc; 2965 tinc = PAGE_SIZE; 2966 } 2967 VM_OBJECT_UNLOCK(obj); 2968 2969 /* 2970 * Step 3, fixup the KVM pmap. Remember that 2971 * bp->b_data is relative to bp->b_offset, but 2972 * bp->b_offset may be offset into the first page. 2973 */ 2974 2975 bp->b_data = (caddr_t) 2976 trunc_page((vm_offset_t)bp->b_data); 2977 pmap_qenter( 2978 (vm_offset_t)bp->b_data, 2979 bp->b_pages, 2980 bp->b_npages 2981 ); 2982 2983 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2984 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2985 } 2986 } 2987 if (newbsize < bp->b_bufsize) 2988 bufspacewakeup(); 2989 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2990 bp->b_bcount = size; /* requested buffer size */ 2991 return 1; 2992} 2993 2994void 2995biodone(struct bio *bp) 2996{ 2997 void (*done)(struct bio *); 2998 2999 mtx_lock(&bdonelock); 3000 bp->bio_flags |= BIO_DONE; 3001 done = bp->bio_done; 3002 if (done == NULL) 3003 wakeup(bp); 3004 mtx_unlock(&bdonelock); 3005 if (done != NULL) 3006 done(bp); 3007} 3008 3009/* 3010 * Wait for a BIO to finish. 3011 * 3012 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3013 * case is not yet clear. 3014 */ 3015int 3016biowait(struct bio *bp, const char *wchan) 3017{ 3018 3019 mtx_lock(&bdonelock); 3020 while ((bp->bio_flags & BIO_DONE) == 0) 3021 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3022 mtx_unlock(&bdonelock); 3023 if (bp->bio_error != 0) 3024 return (bp->bio_error); 3025 if (!(bp->bio_flags & BIO_ERROR)) 3026 return (0); 3027 return (EIO); 3028} 3029 3030void 3031biofinish(struct bio *bp, struct devstat *stat, int error) 3032{ 3033 3034 if (error) { 3035 bp->bio_error = error; 3036 bp->bio_flags |= BIO_ERROR; 3037 } 3038 if (stat != NULL) 3039 devstat_end_transaction_bio(stat, bp); 3040 biodone(bp); 3041} 3042 3043/* 3044 * bufwait: 3045 * 3046 * Wait for buffer I/O completion, returning error status. The buffer 3047 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3048 * error and cleared. 3049 */ 3050int 3051bufwait(struct buf *bp) 3052{ 3053 if (bp->b_iocmd == BIO_READ) 3054 bwait(bp, PRIBIO, "biord"); 3055 else 3056 bwait(bp, PRIBIO, "biowr"); 3057 if (bp->b_flags & B_EINTR) { 3058 bp->b_flags &= ~B_EINTR; 3059 return (EINTR); 3060 } 3061 if (bp->b_ioflags & BIO_ERROR) { 3062 return (bp->b_error ? bp->b_error : EIO); 3063 } else { 3064 return (0); 3065 } 3066} 3067 3068 /* 3069 * Call back function from struct bio back up to struct buf. 3070 */ 3071static void 3072bufdonebio(struct bio *bip) 3073{ 3074 struct buf *bp; 3075 3076 bp = bip->bio_caller2; 3077 bp->b_resid = bp->b_bcount - bip->bio_completed; 3078 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3079 bp->b_ioflags = bip->bio_flags; 3080 bp->b_error = bip->bio_error; 3081 if (bp->b_error) 3082 bp->b_ioflags |= BIO_ERROR; 3083 bufdone(bp); 3084 g_destroy_bio(bip); 3085} 3086 3087void 3088dev_strategy(struct cdev *dev, struct buf *bp) 3089{ 3090 struct cdevsw *csw; 3091 struct bio *bip; 3092 3093 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3094 panic("b_iocmd botch"); 3095 for (;;) { 3096 bip = g_new_bio(); 3097 if (bip != NULL) 3098 break; 3099 /* Try again later */ 3100 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3101 } 3102 bip->bio_cmd = bp->b_iocmd; 3103 bip->bio_offset = bp->b_iooffset; 3104 bip->bio_length = bp->b_bcount; 3105 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3106 bip->bio_data = bp->b_data; 3107 bip->bio_done = bufdonebio; 3108 bip->bio_caller2 = bp; 3109 bip->bio_dev = dev; 3110 KASSERT(dev->si_refcount > 0, 3111 ("dev_strategy on un-referenced struct cdev *(%s)", 3112 devtoname(dev))); 3113 csw = dev_refthread(dev); 3114 if (csw == NULL) { 3115 g_destroy_bio(bip); 3116 bp->b_error = ENXIO; 3117 bp->b_ioflags = BIO_ERROR; 3118 bufdone(bp); 3119 return; 3120 } 3121 (*csw->d_strategy)(bip); 3122 dev_relthread(dev); 3123} 3124 3125/* 3126 * bufdone: 3127 * 3128 * Finish I/O on a buffer, optionally calling a completion function. 3129 * This is usually called from an interrupt so process blocking is 3130 * not allowed. 3131 * 3132 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3133 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3134 * assuming B_INVAL is clear. 3135 * 3136 * For the VMIO case, we set B_CACHE if the op was a read and no 3137 * read error occured, or if the op was a write. B_CACHE is never 3138 * set if the buffer is invalid or otherwise uncacheable. 3139 * 3140 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3141 * initiator to leave B_INVAL set to brelse the buffer out of existance 3142 * in the biodone routine. 3143 */ 3144void 3145bufdone(struct buf *bp) 3146{ 3147 struct bufobj *dropobj; 3148 void (*biodone)(struct buf *); 3149 3150 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3151 dropobj = NULL; 3152 3153 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3154 BUF_ASSERT_HELD(bp); 3155 3156 runningbufwakeup(bp); 3157 if (bp->b_iocmd == BIO_WRITE) 3158 dropobj = bp->b_bufobj; 3159 /* call optional completion function if requested */ 3160 if (bp->b_iodone != NULL) { 3161 biodone = bp->b_iodone; 3162 bp->b_iodone = NULL; 3163 (*biodone) (bp); 3164 if (dropobj) 3165 bufobj_wdrop(dropobj); 3166 return; 3167 } 3168 3169 bufdone_finish(bp); 3170 3171 if (dropobj) 3172 bufobj_wdrop(dropobj); 3173} 3174 3175void 3176bufdone_finish(struct buf *bp) 3177{ 3178 BUF_ASSERT_HELD(bp); 3179 3180 if (!LIST_EMPTY(&bp->b_dep)) 3181 buf_complete(bp); 3182 3183 if (bp->b_flags & B_VMIO) { 3184 int i; 3185 vm_ooffset_t foff; 3186 vm_page_t m; 3187 vm_object_t obj; 3188 int iosize; 3189 struct vnode *vp = bp->b_vp; 3190 boolean_t are_queues_locked; 3191 3192 obj = bp->b_bufobj->bo_object; 3193 3194#if defined(VFS_BIO_DEBUG) 3195 mp_fixme("usecount and vflag accessed without locks."); 3196 if (vp->v_usecount == 0) { 3197 panic("biodone: zero vnode ref count"); 3198 } 3199 3200 KASSERT(vp->v_object != NULL, 3201 ("biodone: vnode %p has no vm_object", vp)); 3202#endif 3203 3204 foff = bp->b_offset; 3205 KASSERT(bp->b_offset != NOOFFSET, 3206 ("biodone: no buffer offset")); 3207 3208 VM_OBJECT_LOCK(obj); 3209#if defined(VFS_BIO_DEBUG) 3210 if (obj->paging_in_progress < bp->b_npages) { 3211 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3212 obj->paging_in_progress, bp->b_npages); 3213 } 3214#endif 3215 3216 /* 3217 * Set B_CACHE if the op was a normal read and no error 3218 * occured. B_CACHE is set for writes in the b*write() 3219 * routines. 3220 */ 3221 iosize = bp->b_bcount - bp->b_resid; 3222 if (bp->b_iocmd == BIO_READ && 3223 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3224 !(bp->b_ioflags & BIO_ERROR)) { 3225 bp->b_flags |= B_CACHE; 3226 } 3227 if (bp->b_iocmd == BIO_READ) { 3228 vm_page_lock_queues(); 3229 are_queues_locked = TRUE; 3230 } else 3231 are_queues_locked = FALSE; 3232 for (i = 0; i < bp->b_npages; i++) { 3233 int bogusflag = 0; 3234 int resid; 3235 3236 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3237 if (resid > iosize) 3238 resid = iosize; 3239 3240 /* 3241 * cleanup bogus pages, restoring the originals 3242 */ 3243 m = bp->b_pages[i]; 3244 if (m == bogus_page) { 3245 bogusflag = 1; 3246 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3247 if (m == NULL) 3248 panic("biodone: page disappeared!"); 3249 bp->b_pages[i] = m; 3250 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3251 bp->b_pages, bp->b_npages); 3252 } 3253#if defined(VFS_BIO_DEBUG) 3254 if (OFF_TO_IDX(foff) != m->pindex) { 3255 printf( 3256"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3257 (intmax_t)foff, (uintmax_t)m->pindex); 3258 } 3259#endif 3260 3261 /* 3262 * In the write case, the valid and clean bits are 3263 * already changed correctly ( see bdwrite() ), so we 3264 * only need to do this here in the read case. 3265 */ 3266 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3267 vfs_page_set_valid(bp, foff, m); 3268 } 3269 3270 /* 3271 * when debugging new filesystems or buffer I/O methods, this 3272 * is the most common error that pops up. if you see this, you 3273 * have not set the page busy flag correctly!!! 3274 */ 3275 if (m->busy == 0) { 3276 printf("biodone: page busy < 0, " 3277 "pindex: %d, foff: 0x(%x,%x), " 3278 "resid: %d, index: %d\n", 3279 (int) m->pindex, (int)(foff >> 32), 3280 (int) foff & 0xffffffff, resid, i); 3281 if (!vn_isdisk(vp, NULL)) 3282 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3283 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3284 (intmax_t) bp->b_lblkno, 3285 bp->b_flags, bp->b_npages); 3286 else 3287 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3288 (intmax_t) bp->b_lblkno, 3289 bp->b_flags, bp->b_npages); 3290 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3291 (u_long)m->valid, (u_long)m->dirty, 3292 m->wire_count); 3293 panic("biodone: page busy < 0\n"); 3294 } 3295 vm_page_io_finish(m); 3296 vm_object_pip_subtract(obj, 1); 3297 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3298 iosize -= resid; 3299 } 3300 if (are_queues_locked) 3301 vm_page_unlock_queues(); 3302 vm_object_pip_wakeupn(obj, 0); 3303 VM_OBJECT_UNLOCK(obj); 3304 } 3305 3306 /* 3307 * For asynchronous completions, release the buffer now. The brelse 3308 * will do a wakeup there if necessary - so no need to do a wakeup 3309 * here in the async case. The sync case always needs to do a wakeup. 3310 */ 3311 3312 if (bp->b_flags & B_ASYNC) { 3313 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3314 brelse(bp); 3315 else 3316 bqrelse(bp); 3317 } else 3318 bdone(bp); 3319} 3320 3321/* 3322 * This routine is called in lieu of iodone in the case of 3323 * incomplete I/O. This keeps the busy status for pages 3324 * consistant. 3325 */ 3326void 3327vfs_unbusy_pages(struct buf *bp) 3328{ 3329 int i; 3330 vm_object_t obj; 3331 vm_page_t m; 3332 3333 runningbufwakeup(bp); 3334 if (!(bp->b_flags & B_VMIO)) 3335 return; 3336 3337 obj = bp->b_bufobj->bo_object; 3338 VM_OBJECT_LOCK(obj); 3339 for (i = 0; i < bp->b_npages; i++) { 3340 m = bp->b_pages[i]; 3341 if (m == bogus_page) { 3342 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3343 if (!m) 3344 panic("vfs_unbusy_pages: page missing\n"); 3345 bp->b_pages[i] = m; 3346 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3347 bp->b_pages, bp->b_npages); 3348 } 3349 vm_object_pip_subtract(obj, 1); 3350 vm_page_io_finish(m); 3351 } 3352 vm_object_pip_wakeupn(obj, 0); 3353 VM_OBJECT_UNLOCK(obj); 3354} 3355 3356/* 3357 * vfs_page_set_valid: 3358 * 3359 * Set the valid bits in a page based on the supplied offset. The 3360 * range is restricted to the buffer's size. 3361 * 3362 * This routine is typically called after a read completes. 3363 */ 3364static void 3365vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3366{ 3367 vm_ooffset_t soff, eoff; 3368 3369 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3370 /* 3371 * Start and end offsets in buffer. eoff - soff may not cross a 3372 * page boundry or cross the end of the buffer. The end of the 3373 * buffer, in this case, is our file EOF, not the allocation size 3374 * of the buffer. 3375 */ 3376 soff = off; 3377 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3378 if (eoff > bp->b_offset + bp->b_bcount) 3379 eoff = bp->b_offset + bp->b_bcount; 3380 3381 /* 3382 * Set valid range. This is typically the entire buffer and thus the 3383 * entire page. 3384 */ 3385 if (eoff > soff) { 3386 vm_page_set_validclean( 3387 m, 3388 (vm_offset_t) (soff & PAGE_MASK), 3389 (vm_offset_t) (eoff - soff) 3390 ); 3391 } 3392} 3393 3394/* 3395 * This routine is called before a device strategy routine. 3396 * It is used to tell the VM system that paging I/O is in 3397 * progress, and treat the pages associated with the buffer 3398 * almost as being VPO_BUSY. Also the object paging_in_progress 3399 * flag is handled to make sure that the object doesn't become 3400 * inconsistant. 3401 * 3402 * Since I/O has not been initiated yet, certain buffer flags 3403 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3404 * and should be ignored. 3405 */ 3406void 3407vfs_busy_pages(struct buf *bp, int clear_modify) 3408{ 3409 int i, bogus; 3410 vm_object_t obj; 3411 vm_ooffset_t foff; 3412 vm_page_t m; 3413 3414 if (!(bp->b_flags & B_VMIO)) 3415 return; 3416 3417 obj = bp->b_bufobj->bo_object; 3418 foff = bp->b_offset; 3419 KASSERT(bp->b_offset != NOOFFSET, 3420 ("vfs_busy_pages: no buffer offset")); 3421 VM_OBJECT_LOCK(obj); 3422 if (bp->b_bufsize != 0) 3423 vfs_setdirty_locked_object(bp); 3424retry: 3425 for (i = 0; i < bp->b_npages; i++) { 3426 m = bp->b_pages[i]; 3427 3428 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3429 goto retry; 3430 } 3431 bogus = 0; 3432 vm_page_lock_queues(); 3433 for (i = 0; i < bp->b_npages; i++) { 3434 m = bp->b_pages[i]; 3435 3436 if ((bp->b_flags & B_CLUSTER) == 0) { 3437 vm_object_pip_add(obj, 1); 3438 vm_page_io_start(m); 3439 } 3440 /* 3441 * When readying a buffer for a read ( i.e 3442 * clear_modify == 0 ), it is important to do 3443 * bogus_page replacement for valid pages in 3444 * partially instantiated buffers. Partially 3445 * instantiated buffers can, in turn, occur when 3446 * reconstituting a buffer from its VM backing store 3447 * base. We only have to do this if B_CACHE is 3448 * clear ( which causes the I/O to occur in the 3449 * first place ). The replacement prevents the read 3450 * I/O from overwriting potentially dirty VM-backed 3451 * pages. XXX bogus page replacement is, uh, bogus. 3452 * It may not work properly with small-block devices. 3453 * We need to find a better way. 3454 */ 3455 pmap_remove_all(m); 3456 if (clear_modify) 3457 vfs_page_set_valid(bp, foff, m); 3458 else if (m->valid == VM_PAGE_BITS_ALL && 3459 (bp->b_flags & B_CACHE) == 0) { 3460 bp->b_pages[i] = bogus_page; 3461 bogus++; 3462 } 3463 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3464 } 3465 vm_page_unlock_queues(); 3466 VM_OBJECT_UNLOCK(obj); 3467 if (bogus) 3468 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3469 bp->b_pages, bp->b_npages); 3470} 3471 3472/* 3473 * Tell the VM system that the pages associated with this buffer 3474 * are clean. This is used for delayed writes where the data is 3475 * going to go to disk eventually without additional VM intevention. 3476 * 3477 * Note that while we only really need to clean through to b_bcount, we 3478 * just go ahead and clean through to b_bufsize. 3479 */ 3480static void 3481vfs_clean_pages(struct buf *bp) 3482{ 3483 int i; 3484 vm_ooffset_t foff, noff, eoff; 3485 vm_page_t m; 3486 3487 if (!(bp->b_flags & B_VMIO)) 3488 return; 3489 3490 foff = bp->b_offset; 3491 KASSERT(bp->b_offset != NOOFFSET, 3492 ("vfs_clean_pages: no buffer offset")); 3493 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3494 vm_page_lock_queues(); 3495 for (i = 0; i < bp->b_npages; i++) { 3496 m = bp->b_pages[i]; 3497 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3498 eoff = noff; 3499 3500 if (eoff > bp->b_offset + bp->b_bufsize) 3501 eoff = bp->b_offset + bp->b_bufsize; 3502 vfs_page_set_valid(bp, foff, m); 3503 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3504 foff = noff; 3505 } 3506 vm_page_unlock_queues(); 3507 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3508} 3509 3510/* 3511 * vfs_bio_set_validclean: 3512 * 3513 * Set the range within the buffer to valid and clean. The range is 3514 * relative to the beginning of the buffer, b_offset. Note that b_offset 3515 * itself may be offset from the beginning of the first page. 3516 * 3517 */ 3518 3519void 3520vfs_bio_set_validclean(struct buf *bp, int base, int size) 3521{ 3522 int i, n; 3523 vm_page_t m; 3524 3525 if (!(bp->b_flags & B_VMIO)) 3526 return; 3527 /* 3528 * Fixup base to be relative to beginning of first page. 3529 * Set initial n to be the maximum number of bytes in the 3530 * first page that can be validated. 3531 */ 3532 3533 base += (bp->b_offset & PAGE_MASK); 3534 n = PAGE_SIZE - (base & PAGE_MASK); 3535 3536 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3537 vm_page_lock_queues(); 3538 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3539 m = bp->b_pages[i]; 3540 if (n > size) 3541 n = size; 3542 vm_page_set_validclean(m, base & PAGE_MASK, n); 3543 base += n; 3544 size -= n; 3545 n = PAGE_SIZE; 3546 } 3547 vm_page_unlock_queues(); 3548 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3549} 3550 3551/* 3552 * vfs_bio_clrbuf: 3553 * 3554 * clear a buffer. This routine essentially fakes an I/O, so we need 3555 * to clear BIO_ERROR and B_INVAL. 3556 * 3557 * Note that while we only theoretically need to clear through b_bcount, 3558 * we go ahead and clear through b_bufsize. 3559 */ 3560 3561void 3562vfs_bio_clrbuf(struct buf *bp) 3563{ 3564 int i, j, mask = 0; 3565 caddr_t sa, ea; 3566 3567 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3568 clrbuf(bp); 3569 return; 3570 } 3571 3572 bp->b_flags &= ~B_INVAL; 3573 bp->b_ioflags &= ~BIO_ERROR; 3574 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3575 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3576 (bp->b_offset & PAGE_MASK) == 0) { 3577 if (bp->b_pages[0] == bogus_page) 3578 goto unlock; 3579 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3580 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3581 if ((bp->b_pages[0]->valid & mask) == mask) 3582 goto unlock; 3583 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3584 ((bp->b_pages[0]->valid & mask) == 0)) { 3585 bzero(bp->b_data, bp->b_bufsize); 3586 bp->b_pages[0]->valid |= mask; 3587 goto unlock; 3588 } 3589 } 3590 ea = sa = bp->b_data; 3591 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3592 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3593 ea = (caddr_t)(vm_offset_t)ulmin( 3594 (u_long)(vm_offset_t)ea, 3595 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3596 if (bp->b_pages[i] == bogus_page) 3597 continue; 3598 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3599 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3600 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3601 if ((bp->b_pages[i]->valid & mask) == mask) 3602 continue; 3603 if ((bp->b_pages[i]->valid & mask) == 0) { 3604 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3605 bzero(sa, ea - sa); 3606 } else { 3607 for (; sa < ea; sa += DEV_BSIZE, j++) { 3608 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3609 (bp->b_pages[i]->valid & (1 << j)) == 0) 3610 bzero(sa, DEV_BSIZE); 3611 } 3612 } 3613 bp->b_pages[i]->valid |= mask; 3614 } 3615unlock: 3616 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3617 bp->b_resid = 0; 3618} 3619 3620/* 3621 * vm_hold_load_pages and vm_hold_free_pages get pages into 3622 * a buffers address space. The pages are anonymous and are 3623 * not associated with a file object. 3624 */ 3625static void 3626vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3627{ 3628 vm_offset_t pg; 3629 vm_page_t p; 3630 int index; 3631 3632 to = round_page(to); 3633 from = round_page(from); 3634 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3635 3636 VM_OBJECT_LOCK(kernel_object); 3637 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3638tryagain: 3639 /* 3640 * note: must allocate system pages since blocking here 3641 * could intefere with paging I/O, no matter which 3642 * process we are. 3643 */ 3644 p = vm_page_alloc(kernel_object, 3645 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3646 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3647 if (!p) { 3648 atomic_add_int(&vm_pageout_deficit, 3649 (to - pg) >> PAGE_SHIFT); 3650 VM_OBJECT_UNLOCK(kernel_object); 3651 VM_WAIT; 3652 VM_OBJECT_LOCK(kernel_object); 3653 goto tryagain; 3654 } 3655 p->valid = VM_PAGE_BITS_ALL; 3656 pmap_qenter(pg, &p, 1); 3657 bp->b_pages[index] = p; 3658 } 3659 VM_OBJECT_UNLOCK(kernel_object); 3660 bp->b_npages = index; 3661} 3662 3663/* Return pages associated with this buf to the vm system */ 3664static void 3665vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3666{ 3667 vm_offset_t pg; 3668 vm_page_t p; 3669 int index, newnpages; 3670 3671 from = round_page(from); 3672 to = round_page(to); 3673 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3674 3675 VM_OBJECT_LOCK(kernel_object); 3676 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3677 p = bp->b_pages[index]; 3678 if (p && (index < bp->b_npages)) { 3679 if (p->busy) { 3680 printf( 3681 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3682 (intmax_t)bp->b_blkno, 3683 (intmax_t)bp->b_lblkno); 3684 } 3685 bp->b_pages[index] = NULL; 3686 pmap_qremove(pg, 1); 3687 vm_page_lock_queues(); 3688 vm_page_unwire(p, 0); 3689 vm_page_free(p); 3690 vm_page_unlock_queues(); 3691 } 3692 } 3693 VM_OBJECT_UNLOCK(kernel_object); 3694 bp->b_npages = newnpages; 3695} 3696 3697/* 3698 * Map an IO request into kernel virtual address space. 3699 * 3700 * All requests are (re)mapped into kernel VA space. 3701 * Notice that we use b_bufsize for the size of the buffer 3702 * to be mapped. b_bcount might be modified by the driver. 3703 * 3704 * Note that even if the caller determines that the address space should 3705 * be valid, a race or a smaller-file mapped into a larger space may 3706 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3707 * check the return value. 3708 */ 3709int 3710vmapbuf(struct buf *bp) 3711{ 3712 caddr_t addr, kva; 3713 vm_prot_t prot; 3714 int pidx, i; 3715 struct vm_page *m; 3716 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3717 3718 if (bp->b_bufsize < 0) 3719 return (-1); 3720 prot = VM_PROT_READ; 3721 if (bp->b_iocmd == BIO_READ) 3722 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3723 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3724 addr < bp->b_data + bp->b_bufsize; 3725 addr += PAGE_SIZE, pidx++) { 3726 /* 3727 * Do the vm_fault if needed; do the copy-on-write thing 3728 * when reading stuff off device into memory. 3729 * 3730 * NOTE! Must use pmap_extract() because addr may be in 3731 * the userland address space, and kextract is only guarenteed 3732 * to work for the kernland address space (see: sparc64 port). 3733 */ 3734retry: 3735 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3736 prot) < 0) { 3737 vm_page_lock_queues(); 3738 for (i = 0; i < pidx; ++i) { 3739 vm_page_unhold(bp->b_pages[i]); 3740 bp->b_pages[i] = NULL; 3741 } 3742 vm_page_unlock_queues(); 3743 return(-1); 3744 } 3745 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3746 if (m == NULL) 3747 goto retry; 3748 bp->b_pages[pidx] = m; 3749 } 3750 if (pidx > btoc(MAXPHYS)) 3751 panic("vmapbuf: mapped more than MAXPHYS"); 3752 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3753 3754 kva = bp->b_saveaddr; 3755 bp->b_npages = pidx; 3756 bp->b_saveaddr = bp->b_data; 3757 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3758 return(0); 3759} 3760 3761/* 3762 * Free the io map PTEs associated with this IO operation. 3763 * We also invalidate the TLB entries and restore the original b_addr. 3764 */ 3765void 3766vunmapbuf(struct buf *bp) 3767{ 3768 int pidx; 3769 int npages; 3770 3771 npages = bp->b_npages; 3772 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3773 vm_page_lock_queues(); 3774 for (pidx = 0; pidx < npages; pidx++) 3775 vm_page_unhold(bp->b_pages[pidx]); 3776 vm_page_unlock_queues(); 3777 3778 bp->b_data = bp->b_saveaddr; 3779} 3780 3781void 3782bdone(struct buf *bp) 3783{ 3784 3785 mtx_lock(&bdonelock); 3786 bp->b_flags |= B_DONE; 3787 wakeup(bp); 3788 mtx_unlock(&bdonelock); 3789} 3790 3791void 3792bwait(struct buf *bp, u_char pri, const char *wchan) 3793{ 3794 3795 mtx_lock(&bdonelock); 3796 while ((bp->b_flags & B_DONE) == 0) 3797 msleep(bp, &bdonelock, pri, wchan, 0); 3798 mtx_unlock(&bdonelock); 3799} 3800 3801int 3802bufsync(struct bufobj *bo, int waitfor, struct thread *td) 3803{ 3804 3805 return (VOP_FSYNC(bo->__bo_vnode, waitfor, td)); 3806} 3807 3808void 3809bufstrategy(struct bufobj *bo, struct buf *bp) 3810{ 3811 int i = 0; 3812 struct vnode *vp; 3813 3814 vp = bp->b_vp; 3815 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3816 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3817 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3818 i = VOP_STRATEGY(vp, bp); 3819 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3820} 3821 3822void 3823bufobj_wrefl(struct bufobj *bo) 3824{ 3825 3826 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3827 ASSERT_BO_LOCKED(bo); 3828 bo->bo_numoutput++; 3829} 3830 3831void 3832bufobj_wref(struct bufobj *bo) 3833{ 3834 3835 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3836 BO_LOCK(bo); 3837 bo->bo_numoutput++; 3838 BO_UNLOCK(bo); 3839} 3840 3841void 3842bufobj_wdrop(struct bufobj *bo) 3843{ 3844 3845 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3846 BO_LOCK(bo); 3847 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3848 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3849 bo->bo_flag &= ~BO_WWAIT; 3850 wakeup(&bo->bo_numoutput); 3851 } 3852 BO_UNLOCK(bo); 3853} 3854 3855int 3856bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3857{ 3858 int error; 3859 3860 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3861 ASSERT_BO_LOCKED(bo); 3862 error = 0; 3863 while (bo->bo_numoutput) { 3864 bo->bo_flag |= BO_WWAIT; 3865 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3866 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3867 if (error) 3868 break; 3869 } 3870 return (error); 3871} 3872 3873void 3874bpin(struct buf *bp) 3875{ 3876 mtx_lock(&bpinlock); 3877 bp->b_pin_count++; 3878 mtx_unlock(&bpinlock); 3879} 3880 3881void 3882bunpin(struct buf *bp) 3883{ 3884 mtx_lock(&bpinlock); 3885 if (--bp->b_pin_count == 0) 3886 wakeup(bp); 3887 mtx_unlock(&bpinlock); 3888} 3889 3890void 3891bunpin_wait(struct buf *bp) 3892{ 3893 mtx_lock(&bpinlock); 3894 while (bp->b_pin_count > 0) 3895 msleep(bp, &bpinlock, PRIBIO, "bwunpin", 0); 3896 mtx_unlock(&bpinlock); 3897} 3898 3899#include "opt_ddb.h" 3900#ifdef DDB 3901#include <ddb/ddb.h> 3902 3903/* DDB command to show buffer data */ 3904DB_SHOW_COMMAND(buffer, db_show_buffer) 3905{ 3906 /* get args */ 3907 struct buf *bp = (struct buf *)addr; 3908 3909 if (!have_addr) { 3910 db_printf("usage: show buffer <addr>\n"); 3911 return; 3912 } 3913 3914 db_printf("buf at %p\n", bp); 3915 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3916 db_printf( 3917 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3918 "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n", 3919 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3920 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno); 3921 if (bp->b_npages) { 3922 int i; 3923 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3924 for (i = 0; i < bp->b_npages; i++) { 3925 vm_page_t m; 3926 m = bp->b_pages[i]; 3927 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3928 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3929 if ((i + 1) < bp->b_npages) 3930 db_printf(","); 3931 } 3932 db_printf("\n"); 3933 } 3934 lockmgr_printinfo(&bp->b_lock); 3935} 3936 3937DB_SHOW_COMMAND(lockedbufs, lockedbufs) 3938{ 3939 struct buf *bp; 3940 int i; 3941 3942 for (i = 0; i < nbuf; i++) { 3943 bp = &buf[i]; 3944 if (BUF_ISLOCKED(bp)) { 3945 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 3946 db_printf("\n"); 3947 } 3948 } 3949} 3950#endif /* DDB */ 3951