vfs_bio.c revision 137197
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 137197 2004-11-04 09:48:18Z phk $"); 43 44#include <sys/param.h> 45#include <sys/systm.h> 46#include <sys/bio.h> 47#include <sys/conf.h> 48#include <sys/buf.h> 49#include <sys/devicestat.h> 50#include <sys/eventhandler.h> 51#include <sys/lock.h> 52#include <sys/malloc.h> 53#include <sys/mount.h> 54#include <sys/mutex.h> 55#include <sys/kernel.h> 56#include <sys/kthread.h> 57#include <sys/proc.h> 58#include <sys/resourcevar.h> 59#include <sys/sysctl.h> 60#include <sys/vmmeter.h> 61#include <sys/vnode.h> 62#include <geom/geom.h> 63#include <vm/vm.h> 64#include <vm/vm_param.h> 65#include <vm/vm_kern.h> 66#include <vm/vm_pageout.h> 67#include <vm/vm_page.h> 68#include <vm/vm_object.h> 69#include <vm/vm_extern.h> 70#include <vm/vm_map.h> 71#include "opt_directio.h" 72#include "opt_swap.h" 73 74static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 75 76struct bio_ops bioops; /* I/O operation notification */ 77 78struct buf_ops buf_ops_bio = { 79 .bop_name = "buf_ops_bio", 80 .bop_write = bufwrite, 81 .bop_strategy = bufstrategy, 82}; 83 84/* 85 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 86 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 87 */ 88struct buf *buf; /* buffer header pool */ 89 90static struct proc *bufdaemonproc; 91 92static int inmem(struct vnode *vp, daddr_t blkno); 93static void vm_hold_free_pages(struct buf *bp, vm_offset_t from, 94 vm_offset_t to); 95static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 96 vm_offset_t to); 97static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 98 int pageno, vm_page_t m); 99static void vfs_clean_pages(struct buf *bp); 100static void vfs_setdirty(struct buf *bp); 101static void vfs_vmio_release(struct buf *bp); 102static void vfs_backgroundwritedone(struct buf *bp); 103static int vfs_bio_clcheck(struct vnode *vp, int size, 104 daddr_t lblkno, daddr_t blkno); 105static int flushbufqueues(int flushdeps); 106static void buf_daemon(void); 107void bremfreel(struct buf *bp); 108 109int vmiodirenable = TRUE; 110SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 111 "Use the VM system for directory writes"); 112int runningbufspace; 113SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 114 "Amount of presently outstanding async buffer io"); 115static int bufspace; 116SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 117 "KVA memory used for bufs"); 118static int maxbufspace; 119SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 120 "Maximum allowed value of bufspace (including buf_daemon)"); 121static int bufmallocspace; 122SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 123 "Amount of malloced memory for buffers"); 124static int maxbufmallocspace; 125SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 126 "Maximum amount of malloced memory for buffers"); 127static int lobufspace; 128SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 129 "Minimum amount of buffers we want to have"); 130static int hibufspace; 131SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 132 "Maximum allowed value of bufspace (excluding buf_daemon)"); 133static int bufreusecnt; 134SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 135 "Number of times we have reused a buffer"); 136static int buffreekvacnt; 137SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 138 "Number of times we have freed the KVA space from some buffer"); 139static int bufdefragcnt; 140SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 141 "Number of times we have had to repeat buffer allocation to defragment"); 142static int lorunningspace; 143SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 144 "Minimum preferred space used for in-progress I/O"); 145static int hirunningspace; 146SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 147 "Maximum amount of space to use for in-progress I/O"); 148static int dirtybufferflushes; 149SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 150 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 151static int altbufferflushes; 152SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 153 0, "Number of fsync flushes to limit dirty buffers"); 154static int recursiveflushes; 155SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 156 0, "Number of flushes skipped due to being recursive"); 157static int numdirtybuffers; 158SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 159 "Number of buffers that are dirty (has unwritten changes) at the moment"); 160static int lodirtybuffers; 161SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 162 "How many buffers we want to have free before bufdaemon can sleep"); 163static int hidirtybuffers; 164SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 165 "When the number of dirty buffers is considered severe"); 166static int dirtybufthresh; 167SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 168 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 169static int numfreebuffers; 170SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 171 "Number of free buffers"); 172static int lofreebuffers; 173SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 174 "XXX Unused"); 175static int hifreebuffers; 176SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 177 "XXX Complicatedly unused"); 178static int getnewbufcalls; 179SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 180 "Number of calls to getnewbuf"); 181static int getnewbufrestarts; 182SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 183 "Number of times getnewbuf has had to restart a buffer aquisition"); 184static int dobkgrdwrite = 1; 185SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 186 "Do background writes (honoring the BV_BKGRDWRITE flag)?"); 187 188/* 189 * Wakeup point for bufdaemon, as well as indicator of whether it is already 190 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 191 * is idling. 192 */ 193static int bd_request; 194 195/* 196 * This lock synchronizes access to bd_request. 197 */ 198static struct mtx bdlock; 199 200/* 201 * bogus page -- for I/O to/from partially complete buffers 202 * this is a temporary solution to the problem, but it is not 203 * really that bad. it would be better to split the buffer 204 * for input in the case of buffers partially already in memory, 205 * but the code is intricate enough already. 206 */ 207vm_page_t bogus_page; 208 209/* 210 * Synchronization (sleep/wakeup) variable for active buffer space requests. 211 * Set when wait starts, cleared prior to wakeup(). 212 * Used in runningbufwakeup() and waitrunningbufspace(). 213 */ 214static int runningbufreq; 215 216/* 217 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 218 * waitrunningbufspace(). 219 */ 220static struct mtx rbreqlock; 221 222/* 223 * Synchronization (sleep/wakeup) variable for buffer requests. 224 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 225 * by and/or. 226 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 227 * getnewbuf(), and getblk(). 228 */ 229static int needsbuffer; 230 231/* 232 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 233 */ 234static struct mtx nblock; 235 236/* 237 * Lock that protects against bwait()/bdone()/B_DONE races. 238 */ 239 240static struct mtx bdonelock; 241 242/* 243 * Definitions for the buffer free lists. 244 */ 245#define BUFFER_QUEUES 5 /* number of free buffer queues */ 246 247#define QUEUE_NONE 0 /* on no queue */ 248#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 249#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 250#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 251#define QUEUE_EMPTY 4 /* empty buffer headers */ 252 253/* Queues for free buffers with various properties */ 254static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 255 256/* Lock for the bufqueues */ 257static struct mtx bqlock; 258 259/* 260 * Single global constant for BUF_WMESG, to avoid getting multiple references. 261 * buf_wmesg is referred from macros. 262 */ 263const char *buf_wmesg = BUF_WMESG; 264 265#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 266#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 267#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 268#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 269 270#ifdef DIRECTIO 271extern void ffs_rawread_setup(void); 272#endif /* DIRECTIO */ 273/* 274 * numdirtywakeup: 275 * 276 * If someone is blocked due to there being too many dirty buffers, 277 * and numdirtybuffers is now reasonable, wake them up. 278 */ 279 280static __inline void 281numdirtywakeup(int level) 282{ 283 284 if (numdirtybuffers <= level) { 285 mtx_lock(&nblock); 286 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 287 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 288 wakeup(&needsbuffer); 289 } 290 mtx_unlock(&nblock); 291 } 292} 293 294/* 295 * bufspacewakeup: 296 * 297 * Called when buffer space is potentially available for recovery. 298 * getnewbuf() will block on this flag when it is unable to free 299 * sufficient buffer space. Buffer space becomes recoverable when 300 * bp's get placed back in the queues. 301 */ 302 303static __inline void 304bufspacewakeup(void) 305{ 306 307 /* 308 * If someone is waiting for BUF space, wake them up. Even 309 * though we haven't freed the kva space yet, the waiting 310 * process will be able to now. 311 */ 312 mtx_lock(&nblock); 313 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 314 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 315 wakeup(&needsbuffer); 316 } 317 mtx_unlock(&nblock); 318} 319 320/* 321 * runningbufwakeup() - in-progress I/O accounting. 322 * 323 */ 324static __inline void 325runningbufwakeup(struct buf *bp) 326{ 327 328 if (bp->b_runningbufspace) { 329 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 330 bp->b_runningbufspace = 0; 331 mtx_lock(&rbreqlock); 332 if (runningbufreq && runningbufspace <= lorunningspace) { 333 runningbufreq = 0; 334 wakeup(&runningbufreq); 335 } 336 mtx_unlock(&rbreqlock); 337 } 338} 339 340/* 341 * bufcountwakeup: 342 * 343 * Called when a buffer has been added to one of the free queues to 344 * account for the buffer and to wakeup anyone waiting for free buffers. 345 * This typically occurs when large amounts of metadata are being handled 346 * by the buffer cache ( else buffer space runs out first, usually ). 347 */ 348 349static __inline void 350bufcountwakeup(void) 351{ 352 353 atomic_add_int(&numfreebuffers, 1); 354 mtx_lock(&nblock); 355 if (needsbuffer) { 356 needsbuffer &= ~VFS_BIO_NEED_ANY; 357 if (numfreebuffers >= hifreebuffers) 358 needsbuffer &= ~VFS_BIO_NEED_FREE; 359 wakeup(&needsbuffer); 360 } 361 mtx_unlock(&nblock); 362} 363 364/* 365 * waitrunningbufspace() 366 * 367 * runningbufspace is a measure of the amount of I/O currently 368 * running. This routine is used in async-write situations to 369 * prevent creating huge backups of pending writes to a device. 370 * Only asynchronous writes are governed by this function. 371 * 372 * Reads will adjust runningbufspace, but will not block based on it. 373 * The read load has a side effect of reducing the allowed write load. 374 * 375 * This does NOT turn an async write into a sync write. It waits 376 * for earlier writes to complete and generally returns before the 377 * caller's write has reached the device. 378 */ 379static __inline void 380waitrunningbufspace(void) 381{ 382 383 mtx_lock(&rbreqlock); 384 while (runningbufspace > hirunningspace) { 385 ++runningbufreq; 386 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 387 } 388 mtx_unlock(&rbreqlock); 389} 390 391 392/* 393 * vfs_buf_test_cache: 394 * 395 * Called when a buffer is extended. This function clears the B_CACHE 396 * bit if the newly extended portion of the buffer does not contain 397 * valid data. 398 */ 399static __inline 400void 401vfs_buf_test_cache(struct buf *bp, 402 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 403 vm_page_t m) 404{ 405 406 GIANT_REQUIRED; 407 408 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 409 if (bp->b_flags & B_CACHE) { 410 int base = (foff + off) & PAGE_MASK; 411 if (vm_page_is_valid(m, base, size) == 0) 412 bp->b_flags &= ~B_CACHE; 413 } 414} 415 416/* Wake up the buffer deamon if necessary */ 417static __inline 418void 419bd_wakeup(int dirtybuflevel) 420{ 421 422 mtx_lock(&bdlock); 423 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 424 bd_request = 1; 425 wakeup(&bd_request); 426 } 427 mtx_unlock(&bdlock); 428} 429 430/* 431 * bd_speedup - speedup the buffer cache flushing code 432 */ 433 434static __inline 435void 436bd_speedup(void) 437{ 438 439 bd_wakeup(1); 440} 441 442/* 443 * Calculating buffer cache scaling values and reserve space for buffer 444 * headers. This is called during low level kernel initialization and 445 * may be called more then once. We CANNOT write to the memory area 446 * being reserved at this time. 447 */ 448caddr_t 449kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 450{ 451 452 /* 453 * physmem_est is in pages. Convert it to kilobytes (assumes 454 * PAGE_SIZE is >= 1K) 455 */ 456 physmem_est = physmem_est * (PAGE_SIZE / 1024); 457 458 /* 459 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 460 * For the first 64MB of ram nominally allocate sufficient buffers to 461 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 462 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 463 * the buffer cache we limit the eventual kva reservation to 464 * maxbcache bytes. 465 * 466 * factor represents the 1/4 x ram conversion. 467 */ 468 if (nbuf == 0) { 469 int factor = 4 * BKVASIZE / 1024; 470 471 nbuf = 50; 472 if (physmem_est > 4096) 473 nbuf += min((physmem_est - 4096) / factor, 474 65536 / factor); 475 if (physmem_est > 65536) 476 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 477 478 if (maxbcache && nbuf > maxbcache / BKVASIZE) 479 nbuf = maxbcache / BKVASIZE; 480 } 481 482#if 0 483 /* 484 * Do not allow the buffer_map to be more then 1/2 the size of the 485 * kernel_map. 486 */ 487 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 488 (BKVASIZE * 2)) { 489 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 490 (BKVASIZE * 2); 491 printf("Warning: nbufs capped at %d\n", nbuf); 492 } 493#endif 494 495 /* 496 * swbufs are used as temporary holders for I/O, such as paging I/O. 497 * We have no less then 16 and no more then 256. 498 */ 499 nswbuf = max(min(nbuf/4, 256), 16); 500#ifdef NSWBUF_MIN 501 if (nswbuf < NSWBUF_MIN) 502 nswbuf = NSWBUF_MIN; 503#endif 504#ifdef DIRECTIO 505 ffs_rawread_setup(); 506#endif 507 508 /* 509 * Reserve space for the buffer cache buffers 510 */ 511 swbuf = (void *)v; 512 v = (caddr_t)(swbuf + nswbuf); 513 buf = (void *)v; 514 v = (caddr_t)(buf + nbuf); 515 516 return(v); 517} 518 519/* Initialize the buffer subsystem. Called before use of any buffers. */ 520void 521bufinit(void) 522{ 523 struct buf *bp; 524 int i; 525 526 GIANT_REQUIRED; 527 528 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 529 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 530 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 531 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 532 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 533 534 /* next, make a null set of free lists */ 535 for (i = 0; i < BUFFER_QUEUES; i++) 536 TAILQ_INIT(&bufqueues[i]); 537 538 /* finally, initialize each buffer header and stick on empty q */ 539 for (i = 0; i < nbuf; i++) { 540 bp = &buf[i]; 541 bzero(bp, sizeof *bp); 542 bp->b_flags = B_INVAL; /* we're just an empty header */ 543 bp->b_rcred = NOCRED; 544 bp->b_wcred = NOCRED; 545 bp->b_qindex = QUEUE_EMPTY; 546 bp->b_vflags = 0; 547 bp->b_xflags = 0; 548 LIST_INIT(&bp->b_dep); 549 BUF_LOCKINIT(bp); 550 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 551 } 552 553 /* 554 * maxbufspace is the absolute maximum amount of buffer space we are 555 * allowed to reserve in KVM and in real terms. The absolute maximum 556 * is nominally used by buf_daemon. hibufspace is the nominal maximum 557 * used by most other processes. The differential is required to 558 * ensure that buf_daemon is able to run when other processes might 559 * be blocked waiting for buffer space. 560 * 561 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 562 * this may result in KVM fragmentation which is not handled optimally 563 * by the system. 564 */ 565 maxbufspace = nbuf * BKVASIZE; 566 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 567 lobufspace = hibufspace - MAXBSIZE; 568 569 lorunningspace = 512 * 1024; 570 hirunningspace = 1024 * 1024; 571 572/* 573 * Limit the amount of malloc memory since it is wired permanently into 574 * the kernel space. Even though this is accounted for in the buffer 575 * allocation, we don't want the malloced region to grow uncontrolled. 576 * The malloc scheme improves memory utilization significantly on average 577 * (small) directories. 578 */ 579 maxbufmallocspace = hibufspace / 20; 580 581/* 582 * Reduce the chance of a deadlock occuring by limiting the number 583 * of delayed-write dirty buffers we allow to stack up. 584 */ 585 hidirtybuffers = nbuf / 4 + 20; 586 dirtybufthresh = hidirtybuffers * 9 / 10; 587 numdirtybuffers = 0; 588/* 589 * To support extreme low-memory systems, make sure hidirtybuffers cannot 590 * eat up all available buffer space. This occurs when our minimum cannot 591 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 592 * BKVASIZE'd (8K) buffers. 593 */ 594 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 595 hidirtybuffers >>= 1; 596 } 597 lodirtybuffers = hidirtybuffers / 2; 598 599/* 600 * Try to keep the number of free buffers in the specified range, 601 * and give special processes (e.g. like buf_daemon) access to an 602 * emergency reserve. 603 */ 604 lofreebuffers = nbuf / 18 + 5; 605 hifreebuffers = 2 * lofreebuffers; 606 numfreebuffers = nbuf; 607 608/* 609 * Maximum number of async ops initiated per buf_daemon loop. This is 610 * somewhat of a hack at the moment, we really need to limit ourselves 611 * based on the number of bytes of I/O in-transit that were initiated 612 * from buf_daemon. 613 */ 614 615 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 616 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 617} 618 619/* 620 * bfreekva() - free the kva allocation for a buffer. 621 * 622 * Must be called at splbio() or higher as this is the only locking for 623 * buffer_map. 624 * 625 * Since this call frees up buffer space, we call bufspacewakeup(). 626 */ 627static void 628bfreekva(struct buf *bp) 629{ 630 631 GIANT_REQUIRED; 632 633 if (bp->b_kvasize) { 634 atomic_add_int(&buffreekvacnt, 1); 635 atomic_subtract_int(&bufspace, bp->b_kvasize); 636 vm_map_delete(buffer_map, 637 (vm_offset_t) bp->b_kvabase, 638 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 639 ); 640 bp->b_kvasize = 0; 641 bufspacewakeup(); 642 } 643} 644 645/* 646 * bremfree: 647 * 648 * Remove the buffer from the appropriate free list. 649 */ 650void 651bremfree(struct buf *bp) 652{ 653 654 mtx_lock(&bqlock); 655 bremfreel(bp); 656 mtx_unlock(&bqlock); 657} 658 659void 660bremfreel(struct buf *bp) 661{ 662 int s = splbio(); 663 int old_qindex = bp->b_qindex; 664 665 GIANT_REQUIRED; 666 667 if (bp->b_qindex != QUEUE_NONE) { 668 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 669 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 670 bp->b_qindex = QUEUE_NONE; 671 } else { 672 if (BUF_REFCNT(bp) <= 1) 673 panic("bremfree: removing a buffer not on a queue"); 674 } 675 676 /* 677 * Fixup numfreebuffers count. If the buffer is invalid or not 678 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 679 * the buffer was free and we must decrement numfreebuffers. 680 */ 681 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 682 switch(old_qindex) { 683 case QUEUE_DIRTY: 684 case QUEUE_CLEAN: 685 case QUEUE_EMPTY: 686 case QUEUE_EMPTYKVA: 687 atomic_subtract_int(&numfreebuffers, 1); 688 break; 689 default: 690 break; 691 } 692 } 693 splx(s); 694} 695 696 697/* 698 * Get a buffer with the specified data. Look in the cache first. We 699 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 700 * is set, the buffer is valid and we do not have to do anything ( see 701 * getblk() ). This is really just a special case of breadn(). 702 */ 703int 704bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 705 struct buf **bpp) 706{ 707 708 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 709} 710 711/* 712 * Operates like bread, but also starts asynchronous I/O on 713 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 714 * to initiating I/O . If B_CACHE is set, the buffer is valid 715 * and we do not have to do anything. 716 */ 717int 718breadn(struct vnode * vp, daddr_t blkno, int size, 719 daddr_t * rablkno, int *rabsize, 720 int cnt, struct ucred * cred, struct buf **bpp) 721{ 722 struct buf *bp, *rabp; 723 int i; 724 int rv = 0, readwait = 0; 725 726 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 727 728 /* if not found in cache, do some I/O */ 729 if ((bp->b_flags & B_CACHE) == 0) { 730 if (curthread != PCPU_GET(idlethread)) 731 curthread->td_proc->p_stats->p_ru.ru_inblock++; 732 bp->b_iocmd = BIO_READ; 733 bp->b_flags &= ~B_INVAL; 734 bp->b_ioflags &= ~BIO_ERROR; 735 if (bp->b_rcred == NOCRED && cred != NOCRED) 736 bp->b_rcred = crhold(cred); 737 vfs_busy_pages(bp, 0); 738 bp->b_iooffset = dbtob(bp->b_blkno); 739 bstrategy(bp); 740 ++readwait; 741 } 742 743 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 744 if (inmem(vp, *rablkno)) 745 continue; 746 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 747 748 if ((rabp->b_flags & B_CACHE) == 0) { 749 if (curthread != PCPU_GET(idlethread)) 750 curthread->td_proc->p_stats->p_ru.ru_inblock++; 751 rabp->b_flags |= B_ASYNC; 752 rabp->b_flags &= ~B_INVAL; 753 rabp->b_ioflags &= ~BIO_ERROR; 754 rabp->b_iocmd = BIO_READ; 755 if (rabp->b_rcred == NOCRED && cred != NOCRED) 756 rabp->b_rcred = crhold(cred); 757 vfs_busy_pages(rabp, 0); 758 BUF_KERNPROC(rabp); 759 rabp->b_iooffset = dbtob(rabp->b_blkno); 760 bstrategy(rabp); 761 } else { 762 brelse(rabp); 763 } 764 } 765 766 if (readwait) { 767 rv = bufwait(bp); 768 } 769 return (rv); 770} 771 772/* 773 * Write, release buffer on completion. (Done by iodone 774 * if async). Do not bother writing anything if the buffer 775 * is invalid. 776 * 777 * Note that we set B_CACHE here, indicating that buffer is 778 * fully valid and thus cacheable. This is true even of NFS 779 * now so we set it generally. This could be set either here 780 * or in biodone() since the I/O is synchronous. We put it 781 * here. 782 */ 783int 784bufwrite(struct buf *bp) 785{ 786 int oldflags, s; 787 struct buf *newbp; 788 789 if (bp->b_flags & B_INVAL) { 790 brelse(bp); 791 return (0); 792 } 793 794 oldflags = bp->b_flags; 795 796 if (BUF_REFCNT(bp) == 0) 797 panic("bufwrite: buffer is not busy???"); 798 s = splbio(); 799 /* 800 * If a background write is already in progress, delay 801 * writing this block if it is asynchronous. Otherwise 802 * wait for the background write to complete. 803 */ 804 BO_LOCK(bp->b_bufobj); 805 if (bp->b_vflags & BV_BKGRDINPROG) { 806 if (bp->b_flags & B_ASYNC) { 807 BO_UNLOCK(bp->b_bufobj); 808 splx(s); 809 bdwrite(bp); 810 return (0); 811 } 812 bp->b_vflags |= BV_BKGRDWAIT; 813 msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj), PRIBIO, "bwrbg", 0); 814 if (bp->b_vflags & BV_BKGRDINPROG) 815 panic("bufwrite: still writing"); 816 } 817 BO_UNLOCK(bp->b_bufobj); 818 819 /* Mark the buffer clean */ 820 bundirty(bp); 821 822 /* 823 * If this buffer is marked for background writing and we 824 * do not have to wait for it, make a copy and write the 825 * copy so as to leave this buffer ready for further use. 826 * 827 * This optimization eats a lot of memory. If we have a page 828 * or buffer shortfall we can't do it. 829 */ 830 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 831 (bp->b_flags & B_ASYNC) && 832 !vm_page_count_severe() && 833 !buf_dirty_count_severe()) { 834 KASSERT(bp->b_iodone == NULL, 835 ("bufwrite: needs chained iodone (%p)", bp->b_iodone)); 836 837 /* get a new block */ 838 newbp = geteblk(bp->b_bufsize); 839 840 /* 841 * set it to be identical to the old block. We have to 842 * set b_lblkno and BKGRDMARKER before calling bgetvp() 843 * to avoid confusing the splay tree and gbincore(). 844 */ 845 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 846 newbp->b_lblkno = bp->b_lblkno; 847 newbp->b_xflags |= BX_BKGRDMARKER; 848 BO_LOCK(bp->b_bufobj); 849 bp->b_vflags |= BV_BKGRDINPROG; 850 bgetvp(bp->b_vp, newbp); 851 BO_UNLOCK(bp->b_bufobj); 852 newbp->b_bufobj = &bp->b_vp->v_bufobj; 853 newbp->b_blkno = bp->b_blkno; 854 newbp->b_offset = bp->b_offset; 855 newbp->b_iodone = vfs_backgroundwritedone; 856 newbp->b_flags |= B_ASYNC; 857 newbp->b_flags &= ~B_INVAL; 858 859 /* move over the dependencies */ 860 if (LIST_FIRST(&bp->b_dep) != NULL) 861 buf_movedeps(bp, newbp); 862 863 /* 864 * Initiate write on the copy, release the original to 865 * the B_LOCKED queue so that it cannot go away until 866 * the background write completes. If not locked it could go 867 * away and then be reconstituted while it was being written. 868 * If the reconstituted buffer were written, we could end up 869 * with two background copies being written at the same time. 870 */ 871 bqrelse(bp); 872 bp = newbp; 873 } 874 875 bp->b_flags &= ~B_DONE; 876 bp->b_ioflags &= ~BIO_ERROR; 877 bp->b_flags |= B_CACHE; 878 bp->b_iocmd = BIO_WRITE; 879 880 bufobj_wref(bp->b_bufobj); 881 vfs_busy_pages(bp, 1); 882 883 /* 884 * Normal bwrites pipeline writes 885 */ 886 bp->b_runningbufspace = bp->b_bufsize; 887 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 888 889 if (curthread != PCPU_GET(idlethread)) 890 curthread->td_proc->p_stats->p_ru.ru_oublock++; 891 splx(s); 892 if (oldflags & B_ASYNC) 893 BUF_KERNPROC(bp); 894 bp->b_iooffset = dbtob(bp->b_blkno); 895 bstrategy(bp); 896 897 if ((oldflags & B_ASYNC) == 0) { 898 int rtval = bufwait(bp); 899 brelse(bp); 900 return (rtval); 901 } else { 902 /* 903 * don't allow the async write to saturate the I/O 904 * system. We will not deadlock here because 905 * we are blocking waiting for I/O that is already in-progress 906 * to complete. We do not block here if it is the update 907 * or syncer daemon trying to clean up as that can lead 908 * to deadlock. 909 */ 910 if (curthread->td_proc != bufdaemonproc && 911 curthread->td_proc != updateproc) 912 waitrunningbufspace(); 913 } 914 915 return (0); 916} 917 918/* 919 * Complete a background write started from bwrite. 920 */ 921static void 922vfs_backgroundwritedone(struct buf *bp) 923{ 924 struct buf *origbp; 925 926 /* 927 * Find the original buffer that we are writing. 928 */ 929 BO_LOCK(bp->b_bufobj); 930 if ((origbp = gbincore(bp->b_bufobj, bp->b_lblkno)) == NULL) 931 panic("backgroundwritedone: lost buffer"); 932 933 /* 934 * Clear the BV_BKGRDINPROG flag in the original buffer 935 * and awaken it if it is waiting for the write to complete. 936 * If BV_BKGRDINPROG is not set in the original buffer it must 937 * have been released and re-instantiated - which is not legal. 938 */ 939 KASSERT((origbp->b_vflags & BV_BKGRDINPROG), 940 ("backgroundwritedone: lost buffer2")); 941 origbp->b_vflags &= ~BV_BKGRDINPROG; 942 if (origbp->b_vflags & BV_BKGRDWAIT) { 943 origbp->b_vflags &= ~BV_BKGRDWAIT; 944 wakeup(&origbp->b_xflags); 945 } 946 BO_UNLOCK(bp->b_bufobj); 947 /* 948 * Process dependencies then return any unfinished ones. 949 */ 950 if (LIST_FIRST(&bp->b_dep) != NULL) 951 buf_complete(bp); 952 if (LIST_FIRST(&bp->b_dep) != NULL) 953 buf_movedeps(bp, origbp); 954 955 /* 956 * This buffer is marked B_NOCACHE, so when it is released 957 * by biodone, it will be tossed. We mark it with BIO_READ 958 * to avoid biodone doing a second bufobj_wdrop. 959 */ 960 bp->b_flags |= B_NOCACHE; 961 bp->b_iocmd = BIO_READ; 962 bp->b_flags &= ~(B_CACHE | B_DONE); 963 bp->b_iodone = 0; 964 bufdone(bp); 965} 966 967/* 968 * Delayed write. (Buffer is marked dirty). Do not bother writing 969 * anything if the buffer is marked invalid. 970 * 971 * Note that since the buffer must be completely valid, we can safely 972 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 973 * biodone() in order to prevent getblk from writing the buffer 974 * out synchronously. 975 */ 976void 977bdwrite(struct buf *bp) 978{ 979 struct thread *td = curthread; 980 struct vnode *vp; 981 struct buf *nbp; 982 struct bufobj *bo; 983 984 GIANT_REQUIRED; 985 986 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 987 KASSERT(BUF_REFCNT(bp) != 0, ("bdwrite: buffer is not busy")); 988 989 if (bp->b_flags & B_INVAL) { 990 brelse(bp); 991 return; 992 } 993 994 /* 995 * If we have too many dirty buffers, don't create any more. 996 * If we are wildly over our limit, then force a complete 997 * cleanup. Otherwise, just keep the situation from getting 998 * out of control. Note that we have to avoid a recursive 999 * disaster and not try to clean up after our own cleanup! 1000 */ 1001 vp = bp->b_vp; 1002 bo = bp->b_bufobj; 1003 BO_LOCK(bo); 1004 if (td->td_pflags & TDP_COWINPROGRESS) { 1005 recursiveflushes++; 1006 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 1007 BO_UNLOCK(bo); 1008 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 1009 BO_LOCK(bo); 1010 altbufferflushes++; 1011 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 1012 /* 1013 * Try to find a buffer to flush. 1014 */ 1015 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 1016 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1017 buf_countdeps(nbp, 0) || 1018 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1019 continue; 1020 if (bp == nbp) 1021 panic("bdwrite: found ourselves"); 1022 BO_UNLOCK(bo); 1023 if (nbp->b_flags & B_CLUSTEROK) { 1024 vfs_bio_awrite(nbp); 1025 } else { 1026 bremfree(nbp); 1027 bawrite(nbp); 1028 } 1029 BO_LOCK(bo); 1030 dirtybufferflushes++; 1031 break; 1032 } 1033 } 1034 BO_UNLOCK(bo); 1035 1036 bdirty(bp); 1037 /* 1038 * Set B_CACHE, indicating that the buffer is fully valid. This is 1039 * true even of NFS now. 1040 */ 1041 bp->b_flags |= B_CACHE; 1042 1043 /* 1044 * This bmap keeps the system from needing to do the bmap later, 1045 * perhaps when the system is attempting to do a sync. Since it 1046 * is likely that the indirect block -- or whatever other datastructure 1047 * that the filesystem needs is still in memory now, it is a good 1048 * thing to do this. Note also, that if the pageout daemon is 1049 * requesting a sync -- there might not be enough memory to do 1050 * the bmap then... So, this is important to do. 1051 */ 1052 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1053 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1054 } 1055 1056 /* 1057 * Set the *dirty* buffer range based upon the VM system dirty pages. 1058 */ 1059 vfs_setdirty(bp); 1060 1061 /* 1062 * We need to do this here to satisfy the vnode_pager and the 1063 * pageout daemon, so that it thinks that the pages have been 1064 * "cleaned". Note that since the pages are in a delayed write 1065 * buffer -- the VFS layer "will" see that the pages get written 1066 * out on the next sync, or perhaps the cluster will be completed. 1067 */ 1068 vfs_clean_pages(bp); 1069 bqrelse(bp); 1070 1071 /* 1072 * Wakeup the buffer flushing daemon if we have a lot of dirty 1073 * buffers (midpoint between our recovery point and our stall 1074 * point). 1075 */ 1076 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1077 1078 /* 1079 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1080 * due to the softdep code. 1081 */ 1082} 1083 1084/* 1085 * bdirty: 1086 * 1087 * Turn buffer into delayed write request. We must clear BIO_READ and 1088 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1089 * itself to properly update it in the dirty/clean lists. We mark it 1090 * B_DONE to ensure that any asynchronization of the buffer properly 1091 * clears B_DONE ( else a panic will occur later ). 1092 * 1093 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1094 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1095 * should only be called if the buffer is known-good. 1096 * 1097 * Since the buffer is not on a queue, we do not update the numfreebuffers 1098 * count. 1099 * 1100 * Must be called at splbio(). 1101 * The buffer must be on QUEUE_NONE. 1102 */ 1103void 1104bdirty(struct buf *bp) 1105{ 1106 1107 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1108 KASSERT(bp->b_qindex == QUEUE_NONE, 1109 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1110 bp->b_flags &= ~(B_RELBUF); 1111 bp->b_iocmd = BIO_WRITE; 1112 1113 if ((bp->b_flags & B_DELWRI) == 0) { 1114 bp->b_flags |= B_DONE | B_DELWRI; 1115 reassignbuf(bp); 1116 atomic_add_int(&numdirtybuffers, 1); 1117 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1118 } 1119} 1120 1121/* 1122 * bundirty: 1123 * 1124 * Clear B_DELWRI for buffer. 1125 * 1126 * Since the buffer is not on a queue, we do not update the numfreebuffers 1127 * count. 1128 * 1129 * Must be called at splbio(). 1130 * The buffer must be on QUEUE_NONE. 1131 */ 1132 1133void 1134bundirty(struct buf *bp) 1135{ 1136 1137 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1138 KASSERT(bp->b_qindex == QUEUE_NONE, 1139 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1140 1141 if (bp->b_flags & B_DELWRI) { 1142 bp->b_flags &= ~B_DELWRI; 1143 reassignbuf(bp); 1144 atomic_subtract_int(&numdirtybuffers, 1); 1145 numdirtywakeup(lodirtybuffers); 1146 } 1147 /* 1148 * Since it is now being written, we can clear its deferred write flag. 1149 */ 1150 bp->b_flags &= ~B_DEFERRED; 1151} 1152 1153/* 1154 * bawrite: 1155 * 1156 * Asynchronous write. Start output on a buffer, but do not wait for 1157 * it to complete. The buffer is released when the output completes. 1158 * 1159 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1160 * B_INVAL buffers. Not us. 1161 */ 1162void 1163bawrite(struct buf *bp) 1164{ 1165 1166 bp->b_flags |= B_ASYNC; 1167 (void) bwrite(bp); 1168} 1169 1170/* 1171 * bwillwrite: 1172 * 1173 * Called prior to the locking of any vnodes when we are expecting to 1174 * write. We do not want to starve the buffer cache with too many 1175 * dirty buffers so we block here. By blocking prior to the locking 1176 * of any vnodes we attempt to avoid the situation where a locked vnode 1177 * prevents the various system daemons from flushing related buffers. 1178 */ 1179 1180void 1181bwillwrite(void) 1182{ 1183 1184 if (numdirtybuffers >= hidirtybuffers) { 1185 int s; 1186 1187 mtx_lock(&Giant); 1188 s = splbio(); 1189 mtx_lock(&nblock); 1190 while (numdirtybuffers >= hidirtybuffers) { 1191 bd_wakeup(1); 1192 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1193 msleep(&needsbuffer, &nblock, 1194 (PRIBIO + 4), "flswai", 0); 1195 } 1196 splx(s); 1197 mtx_unlock(&nblock); 1198 mtx_unlock(&Giant); 1199 } 1200} 1201 1202/* 1203 * Return true if we have too many dirty buffers. 1204 */ 1205int 1206buf_dirty_count_severe(void) 1207{ 1208 1209 return(numdirtybuffers >= hidirtybuffers); 1210} 1211 1212/* 1213 * brelse: 1214 * 1215 * Release a busy buffer and, if requested, free its resources. The 1216 * buffer will be stashed in the appropriate bufqueue[] allowing it 1217 * to be accessed later as a cache entity or reused for other purposes. 1218 */ 1219void 1220brelse(struct buf *bp) 1221{ 1222 int s; 1223 1224 GIANT_REQUIRED; 1225 1226 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1227 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1228 1229 s = splbio(); 1230 1231 if (bp->b_iocmd == BIO_WRITE && 1232 (bp->b_ioflags & BIO_ERROR) && 1233 !(bp->b_flags & B_INVAL)) { 1234 /* 1235 * Failed write, redirty. Must clear BIO_ERROR to prevent 1236 * pages from being scrapped. If B_INVAL is set then 1237 * this case is not run and the next case is run to 1238 * destroy the buffer. B_INVAL can occur if the buffer 1239 * is outside the range supported by the underlying device. 1240 */ 1241 bp->b_ioflags &= ~BIO_ERROR; 1242 bdirty(bp); 1243 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1244 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1245 /* 1246 * Either a failed I/O or we were asked to free or not 1247 * cache the buffer. 1248 */ 1249 bp->b_flags |= B_INVAL; 1250 if (LIST_FIRST(&bp->b_dep) != NULL) 1251 buf_deallocate(bp); 1252 if (bp->b_flags & B_DELWRI) { 1253 atomic_subtract_int(&numdirtybuffers, 1); 1254 numdirtywakeup(lodirtybuffers); 1255 } 1256 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1257 if ((bp->b_flags & B_VMIO) == 0) { 1258 if (bp->b_bufsize) 1259 allocbuf(bp, 0); 1260 if (bp->b_vp) 1261 brelvp(bp); 1262 } 1263 } 1264 1265 /* 1266 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1267 * is called with B_DELWRI set, the underlying pages may wind up 1268 * getting freed causing a previous write (bdwrite()) to get 'lost' 1269 * because pages associated with a B_DELWRI bp are marked clean. 1270 * 1271 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1272 * if B_DELWRI is set. 1273 * 1274 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1275 * on pages to return pages to the VM page queues. 1276 */ 1277 if (bp->b_flags & B_DELWRI) 1278 bp->b_flags &= ~B_RELBUF; 1279 else if (vm_page_count_severe()) { 1280 /* 1281 * XXX This lock may not be necessary since BKGRDINPROG 1282 * cannot be set while we hold the buf lock, it can only be 1283 * cleared if it is already pending. 1284 */ 1285 if (bp->b_vp) { 1286 BO_LOCK(bp->b_bufobj); 1287 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1288 bp->b_flags |= B_RELBUF; 1289 BO_UNLOCK(bp->b_bufobj); 1290 } else 1291 bp->b_flags |= B_RELBUF; 1292 } 1293 1294 /* 1295 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1296 * constituted, not even NFS buffers now. Two flags effect this. If 1297 * B_INVAL, the struct buf is invalidated but the VM object is kept 1298 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1299 * 1300 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1301 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1302 * buffer is also B_INVAL because it hits the re-dirtying code above. 1303 * 1304 * Normally we can do this whether a buffer is B_DELWRI or not. If 1305 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1306 * the commit state and we cannot afford to lose the buffer. If the 1307 * buffer has a background write in progress, we need to keep it 1308 * around to prevent it from being reconstituted and starting a second 1309 * background write. 1310 */ 1311 if ((bp->b_flags & B_VMIO) 1312 && !(bp->b_vp->v_mount != NULL && 1313 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1314 !vn_isdisk(bp->b_vp, NULL) && 1315 (bp->b_flags & B_DELWRI)) 1316 ) { 1317 1318 int i, j, resid; 1319 vm_page_t m; 1320 off_t foff; 1321 vm_pindex_t poff; 1322 vm_object_t obj; 1323 1324 obj = bp->b_bufobj->bo_object; 1325 1326 /* 1327 * Get the base offset and length of the buffer. Note that 1328 * in the VMIO case if the buffer block size is not 1329 * page-aligned then b_data pointer may not be page-aligned. 1330 * But our b_pages[] array *IS* page aligned. 1331 * 1332 * block sizes less then DEV_BSIZE (usually 512) are not 1333 * supported due to the page granularity bits (m->valid, 1334 * m->dirty, etc...). 1335 * 1336 * See man buf(9) for more information 1337 */ 1338 resid = bp->b_bufsize; 1339 foff = bp->b_offset; 1340 VM_OBJECT_LOCK(obj); 1341 for (i = 0; i < bp->b_npages; i++) { 1342 int had_bogus = 0; 1343 1344 m = bp->b_pages[i]; 1345 1346 /* 1347 * If we hit a bogus page, fixup *all* the bogus pages 1348 * now. 1349 */ 1350 if (m == bogus_page) { 1351 poff = OFF_TO_IDX(bp->b_offset); 1352 had_bogus = 1; 1353 1354 for (j = i; j < bp->b_npages; j++) { 1355 vm_page_t mtmp; 1356 mtmp = bp->b_pages[j]; 1357 if (mtmp == bogus_page) { 1358 mtmp = vm_page_lookup(obj, poff + j); 1359 if (!mtmp) { 1360 panic("brelse: page missing\n"); 1361 } 1362 bp->b_pages[j] = mtmp; 1363 } 1364 } 1365 1366 if ((bp->b_flags & B_INVAL) == 0) { 1367 pmap_qenter( 1368 trunc_page((vm_offset_t)bp->b_data), 1369 bp->b_pages, bp->b_npages); 1370 } 1371 m = bp->b_pages[i]; 1372 } 1373 if ((bp->b_flags & B_NOCACHE) || 1374 (bp->b_ioflags & BIO_ERROR)) { 1375 int poffset = foff & PAGE_MASK; 1376 int presid = resid > (PAGE_SIZE - poffset) ? 1377 (PAGE_SIZE - poffset) : resid; 1378 1379 KASSERT(presid >= 0, ("brelse: extra page")); 1380 vm_page_lock_queues(); 1381 vm_page_set_invalid(m, poffset, presid); 1382 vm_page_unlock_queues(); 1383 if (had_bogus) 1384 printf("avoided corruption bug in bogus_page/brelse code\n"); 1385 } 1386 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1387 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1388 } 1389 VM_OBJECT_UNLOCK(obj); 1390 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1391 vfs_vmio_release(bp); 1392 1393 } else if (bp->b_flags & B_VMIO) { 1394 1395 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1396 vfs_vmio_release(bp); 1397 } 1398 1399 } 1400 1401 if (bp->b_qindex != QUEUE_NONE) 1402 panic("brelse: free buffer onto another queue???"); 1403 if (BUF_REFCNT(bp) > 1) { 1404 /* do not release to free list */ 1405 BUF_UNLOCK(bp); 1406 splx(s); 1407 return; 1408 } 1409 1410 /* enqueue */ 1411 mtx_lock(&bqlock); 1412 1413 /* buffers with no memory */ 1414 if (bp->b_bufsize == 0) { 1415 bp->b_flags |= B_INVAL; 1416 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1417 if (bp->b_vflags & BV_BKGRDINPROG) 1418 panic("losing buffer 1"); 1419 if (bp->b_kvasize) { 1420 bp->b_qindex = QUEUE_EMPTYKVA; 1421 } else { 1422 bp->b_qindex = QUEUE_EMPTY; 1423 } 1424 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1425 /* buffers with junk contents */ 1426 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1427 (bp->b_ioflags & BIO_ERROR)) { 1428 bp->b_flags |= B_INVAL; 1429 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1430 if (bp->b_vflags & BV_BKGRDINPROG) 1431 panic("losing buffer 2"); 1432 bp->b_qindex = QUEUE_CLEAN; 1433 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1434 /* remaining buffers */ 1435 } else { 1436 if (bp->b_flags & B_DELWRI) 1437 bp->b_qindex = QUEUE_DIRTY; 1438 else 1439 bp->b_qindex = QUEUE_CLEAN; 1440 if (bp->b_flags & B_AGE) 1441 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1442 else 1443 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1444 } 1445 mtx_unlock(&bqlock); 1446 1447 /* 1448 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1449 * placed the buffer on the correct queue. We must also disassociate 1450 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1451 * find it. 1452 */ 1453 if (bp->b_flags & B_INVAL) { 1454 if (bp->b_flags & B_DELWRI) 1455 bundirty(bp); 1456 if (bp->b_vp) 1457 brelvp(bp); 1458 } 1459 1460 /* 1461 * Fixup numfreebuffers count. The bp is on an appropriate queue 1462 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1463 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1464 * if B_INVAL is set ). 1465 */ 1466 1467 if (!(bp->b_flags & B_DELWRI)) 1468 bufcountwakeup(); 1469 1470 /* 1471 * Something we can maybe free or reuse 1472 */ 1473 if (bp->b_bufsize || bp->b_kvasize) 1474 bufspacewakeup(); 1475 1476 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1477 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1478 panic("brelse: not dirty"); 1479 /* unlock */ 1480 BUF_UNLOCK(bp); 1481 splx(s); 1482} 1483 1484/* 1485 * Release a buffer back to the appropriate queue but do not try to free 1486 * it. The buffer is expected to be used again soon. 1487 * 1488 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1489 * biodone() to requeue an async I/O on completion. It is also used when 1490 * known good buffers need to be requeued but we think we may need the data 1491 * again soon. 1492 * 1493 * XXX we should be able to leave the B_RELBUF hint set on completion. 1494 */ 1495void 1496bqrelse(struct buf *bp) 1497{ 1498 int s; 1499 1500 s = splbio(); 1501 1502 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1503 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1504 1505 if (bp->b_qindex != QUEUE_NONE) 1506 panic("bqrelse: free buffer onto another queue???"); 1507 if (BUF_REFCNT(bp) > 1) { 1508 /* do not release to free list */ 1509 BUF_UNLOCK(bp); 1510 splx(s); 1511 return; 1512 } 1513 mtx_lock(&bqlock); 1514 /* buffers with stale but valid contents */ 1515 if (bp->b_flags & B_DELWRI) { 1516 bp->b_qindex = QUEUE_DIRTY; 1517 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1518 } else { 1519 /* 1520 * XXX This lock may not be necessary since BKGRDINPROG 1521 * cannot be set while we hold the buf lock, it can only be 1522 * cleared if it is already pending. 1523 */ 1524 BO_LOCK(bp->b_bufobj); 1525 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1526 BO_UNLOCK(bp->b_bufobj); 1527 bp->b_qindex = QUEUE_CLEAN; 1528 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1529 b_freelist); 1530 } else { 1531 /* 1532 * We are too low on memory, we have to try to free 1533 * the buffer (most importantly: the wired pages 1534 * making up its backing store) *now*. 1535 */ 1536 BO_UNLOCK(bp->b_bufobj); 1537 mtx_unlock(&bqlock); 1538 splx(s); 1539 brelse(bp); 1540 return; 1541 } 1542 } 1543 mtx_unlock(&bqlock); 1544 1545 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1546 bufcountwakeup(); 1547 1548 /* 1549 * Something we can maybe free or reuse. 1550 */ 1551 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1552 bufspacewakeup(); 1553 1554 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1555 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1556 panic("bqrelse: not dirty"); 1557 /* unlock */ 1558 BUF_UNLOCK(bp); 1559 splx(s); 1560} 1561 1562/* Give pages used by the bp back to the VM system (where possible) */ 1563static void 1564vfs_vmio_release(struct buf *bp) 1565{ 1566 int i; 1567 vm_page_t m; 1568 1569 GIANT_REQUIRED; 1570 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1571 vm_page_lock_queues(); 1572 for (i = 0; i < bp->b_npages; i++) { 1573 m = bp->b_pages[i]; 1574 bp->b_pages[i] = NULL; 1575 /* 1576 * In order to keep page LRU ordering consistent, put 1577 * everything on the inactive queue. 1578 */ 1579 vm_page_unwire(m, 0); 1580 /* 1581 * We don't mess with busy pages, it is 1582 * the responsibility of the process that 1583 * busied the pages to deal with them. 1584 */ 1585 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1586 continue; 1587 1588 if (m->wire_count == 0) { 1589 /* 1590 * Might as well free the page if we can and it has 1591 * no valid data. We also free the page if the 1592 * buffer was used for direct I/O 1593 */ 1594 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1595 m->hold_count == 0) { 1596 pmap_remove_all(m); 1597 vm_page_free(m); 1598 } else if (bp->b_flags & B_DIRECT) { 1599 vm_page_try_to_free(m); 1600 } else if (vm_page_count_severe()) { 1601 vm_page_try_to_cache(m); 1602 } 1603 } 1604 } 1605 vm_page_unlock_queues(); 1606 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1607 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1608 1609 if (bp->b_bufsize) { 1610 bufspacewakeup(); 1611 bp->b_bufsize = 0; 1612 } 1613 bp->b_npages = 0; 1614 bp->b_flags &= ~B_VMIO; 1615 if (bp->b_vp) 1616 brelvp(bp); 1617} 1618 1619/* 1620 * Check to see if a block at a particular lbn is available for a clustered 1621 * write. 1622 */ 1623static int 1624vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1625{ 1626 struct buf *bpa; 1627 int match; 1628 1629 match = 0; 1630 1631 /* If the buf isn't in core skip it */ 1632 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1633 return (0); 1634 1635 /* If the buf is busy we don't want to wait for it */ 1636 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1637 return (0); 1638 1639 /* Only cluster with valid clusterable delayed write buffers */ 1640 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1641 (B_DELWRI | B_CLUSTEROK)) 1642 goto done; 1643 1644 if (bpa->b_bufsize != size) 1645 goto done; 1646 1647 /* 1648 * Check to see if it is in the expected place on disk and that the 1649 * block has been mapped. 1650 */ 1651 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1652 match = 1; 1653done: 1654 BUF_UNLOCK(bpa); 1655 return (match); 1656} 1657 1658/* 1659 * vfs_bio_awrite: 1660 * 1661 * Implement clustered async writes for clearing out B_DELWRI buffers. 1662 * This is much better then the old way of writing only one buffer at 1663 * a time. Note that we may not be presented with the buffers in the 1664 * correct order, so we search for the cluster in both directions. 1665 */ 1666int 1667vfs_bio_awrite(struct buf *bp) 1668{ 1669 int i; 1670 int j; 1671 daddr_t lblkno = bp->b_lblkno; 1672 struct vnode *vp = bp->b_vp; 1673 int s; 1674 int ncl; 1675 int nwritten; 1676 int size; 1677 int maxcl; 1678 1679 s = splbio(); 1680 /* 1681 * right now we support clustered writing only to regular files. If 1682 * we find a clusterable block we could be in the middle of a cluster 1683 * rather then at the beginning. 1684 */ 1685 if ((vp->v_type == VREG) && 1686 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1687 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1688 1689 size = vp->v_mount->mnt_stat.f_iosize; 1690 maxcl = MAXPHYS / size; 1691 1692 VI_LOCK(vp); 1693 for (i = 1; i < maxcl; i++) 1694 if (vfs_bio_clcheck(vp, size, lblkno + i, 1695 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1696 break; 1697 1698 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1699 if (vfs_bio_clcheck(vp, size, lblkno - j, 1700 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1701 break; 1702 1703 VI_UNLOCK(vp); 1704 --j; 1705 ncl = i + j; 1706 /* 1707 * this is a possible cluster write 1708 */ 1709 if (ncl != 1) { 1710 BUF_UNLOCK(bp); 1711 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1712 splx(s); 1713 return nwritten; 1714 } 1715 } 1716 1717 bremfree(bp); 1718 bp->b_flags |= B_ASYNC; 1719 1720 splx(s); 1721 /* 1722 * default (old) behavior, writing out only one block 1723 * 1724 * XXX returns b_bufsize instead of b_bcount for nwritten? 1725 */ 1726 nwritten = bp->b_bufsize; 1727 (void) bwrite(bp); 1728 1729 return nwritten; 1730} 1731 1732/* 1733 * getnewbuf: 1734 * 1735 * Find and initialize a new buffer header, freeing up existing buffers 1736 * in the bufqueues as necessary. The new buffer is returned locked. 1737 * 1738 * Important: B_INVAL is not set. If the caller wishes to throw the 1739 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1740 * 1741 * We block if: 1742 * We have insufficient buffer headers 1743 * We have insufficient buffer space 1744 * buffer_map is too fragmented ( space reservation fails ) 1745 * If we have to flush dirty buffers ( but we try to avoid this ) 1746 * 1747 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1748 * Instead we ask the buf daemon to do it for us. We attempt to 1749 * avoid piecemeal wakeups of the pageout daemon. 1750 */ 1751 1752static struct buf * 1753getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1754{ 1755 struct buf *bp; 1756 struct buf *nbp; 1757 int defrag = 0; 1758 int nqindex; 1759 static int flushingbufs; 1760 1761 GIANT_REQUIRED; 1762 1763 /* 1764 * We can't afford to block since we might be holding a vnode lock, 1765 * which may prevent system daemons from running. We deal with 1766 * low-memory situations by proactively returning memory and running 1767 * async I/O rather then sync I/O. 1768 */ 1769 1770 atomic_add_int(&getnewbufcalls, 1); 1771 atomic_subtract_int(&getnewbufrestarts, 1); 1772restart: 1773 atomic_add_int(&getnewbufrestarts, 1); 1774 1775 /* 1776 * Setup for scan. If we do not have enough free buffers, 1777 * we setup a degenerate case that immediately fails. Note 1778 * that if we are specially marked process, we are allowed to 1779 * dip into our reserves. 1780 * 1781 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1782 * 1783 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1784 * However, there are a number of cases (defragging, reusing, ...) 1785 * where we cannot backup. 1786 */ 1787 mtx_lock(&bqlock); 1788 nqindex = QUEUE_EMPTYKVA; 1789 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1790 1791 if (nbp == NULL) { 1792 /* 1793 * If no EMPTYKVA buffers and we are either 1794 * defragging or reusing, locate a CLEAN buffer 1795 * to free or reuse. If bufspace useage is low 1796 * skip this step so we can allocate a new buffer. 1797 */ 1798 if (defrag || bufspace >= lobufspace) { 1799 nqindex = QUEUE_CLEAN; 1800 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1801 } 1802 1803 /* 1804 * If we could not find or were not allowed to reuse a 1805 * CLEAN buffer, check to see if it is ok to use an EMPTY 1806 * buffer. We can only use an EMPTY buffer if allocating 1807 * its KVA would not otherwise run us out of buffer space. 1808 */ 1809 if (nbp == NULL && defrag == 0 && 1810 bufspace + maxsize < hibufspace) { 1811 nqindex = QUEUE_EMPTY; 1812 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1813 } 1814 } 1815 1816 /* 1817 * Run scan, possibly freeing data and/or kva mappings on the fly 1818 * depending. 1819 */ 1820 1821 while ((bp = nbp) != NULL) { 1822 int qindex = nqindex; 1823 1824 /* 1825 * Calculate next bp ( we can only use it if we do not block 1826 * or do other fancy things ). 1827 */ 1828 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1829 switch(qindex) { 1830 case QUEUE_EMPTY: 1831 nqindex = QUEUE_EMPTYKVA; 1832 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1833 break; 1834 /* FALLTHROUGH */ 1835 case QUEUE_EMPTYKVA: 1836 nqindex = QUEUE_CLEAN; 1837 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1838 break; 1839 /* FALLTHROUGH */ 1840 case QUEUE_CLEAN: 1841 /* 1842 * nbp is NULL. 1843 */ 1844 break; 1845 } 1846 } 1847 if (bp->b_vp) { 1848 BO_LOCK(bp->b_bufobj); 1849 if (bp->b_vflags & BV_BKGRDINPROG) { 1850 BO_UNLOCK(bp->b_bufobj); 1851 continue; 1852 } 1853 BO_UNLOCK(bp->b_bufobj); 1854 } 1855 1856 /* 1857 * Sanity Checks 1858 */ 1859 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1860 1861 /* 1862 * Note: we no longer distinguish between VMIO and non-VMIO 1863 * buffers. 1864 */ 1865 1866 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1867 1868 /* 1869 * If we are defragging then we need a buffer with 1870 * b_kvasize != 0. XXX this situation should no longer 1871 * occur, if defrag is non-zero the buffer's b_kvasize 1872 * should also be non-zero at this point. XXX 1873 */ 1874 if (defrag && bp->b_kvasize == 0) { 1875 printf("Warning: defrag empty buffer %p\n", bp); 1876 continue; 1877 } 1878 1879 /* 1880 * Start freeing the bp. This is somewhat involved. nbp 1881 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1882 */ 1883 1884 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1885 panic("getnewbuf: locked buf"); 1886 bremfreel(bp); 1887 mtx_unlock(&bqlock); 1888 1889 if (qindex == QUEUE_CLEAN) { 1890 if (bp->b_flags & B_VMIO) { 1891 bp->b_flags &= ~B_ASYNC; 1892 vfs_vmio_release(bp); 1893 } 1894 if (bp->b_vp) 1895 brelvp(bp); 1896 } 1897 1898 /* 1899 * NOTE: nbp is now entirely invalid. We can only restart 1900 * the scan from this point on. 1901 * 1902 * Get the rest of the buffer freed up. b_kva* is still 1903 * valid after this operation. 1904 */ 1905 1906 if (bp->b_rcred != NOCRED) { 1907 crfree(bp->b_rcred); 1908 bp->b_rcred = NOCRED; 1909 } 1910 if (bp->b_wcred != NOCRED) { 1911 crfree(bp->b_wcred); 1912 bp->b_wcred = NOCRED; 1913 } 1914 if (LIST_FIRST(&bp->b_dep) != NULL) 1915 buf_deallocate(bp); 1916 if (bp->b_vflags & BV_BKGRDINPROG) 1917 panic("losing buffer 3"); 1918 1919 if (bp->b_bufsize) 1920 allocbuf(bp, 0); 1921 1922 bp->b_flags = 0; 1923 bp->b_ioflags = 0; 1924 bp->b_xflags = 0; 1925 bp->b_vflags = 0; 1926 bp->b_vp = NULL; 1927 bp->b_blkno = bp->b_lblkno = 0; 1928 bp->b_offset = NOOFFSET; 1929 bp->b_iodone = 0; 1930 bp->b_error = 0; 1931 bp->b_resid = 0; 1932 bp->b_bcount = 0; 1933 bp->b_npages = 0; 1934 bp->b_dirtyoff = bp->b_dirtyend = 0; 1935 bp->b_bufobj = NULL; 1936 1937 LIST_INIT(&bp->b_dep); 1938 1939 /* 1940 * If we are defragging then free the buffer. 1941 */ 1942 if (defrag) { 1943 bp->b_flags |= B_INVAL; 1944 bfreekva(bp); 1945 brelse(bp); 1946 defrag = 0; 1947 goto restart; 1948 } 1949 1950 /* 1951 * If we are overcomitted then recover the buffer and its 1952 * KVM space. This occurs in rare situations when multiple 1953 * processes are blocked in getnewbuf() or allocbuf(). 1954 */ 1955 if (bufspace >= hibufspace) 1956 flushingbufs = 1; 1957 if (flushingbufs && bp->b_kvasize != 0) { 1958 bp->b_flags |= B_INVAL; 1959 bfreekva(bp); 1960 brelse(bp); 1961 goto restart; 1962 } 1963 if (bufspace < lobufspace) 1964 flushingbufs = 0; 1965 break; 1966 } 1967 1968 /* 1969 * If we exhausted our list, sleep as appropriate. We may have to 1970 * wakeup various daemons and write out some dirty buffers. 1971 * 1972 * Generally we are sleeping due to insufficient buffer space. 1973 */ 1974 1975 if (bp == NULL) { 1976 int flags; 1977 char *waitmsg; 1978 1979 mtx_unlock(&bqlock); 1980 if (defrag) { 1981 flags = VFS_BIO_NEED_BUFSPACE; 1982 waitmsg = "nbufkv"; 1983 } else if (bufspace >= hibufspace) { 1984 waitmsg = "nbufbs"; 1985 flags = VFS_BIO_NEED_BUFSPACE; 1986 } else { 1987 waitmsg = "newbuf"; 1988 flags = VFS_BIO_NEED_ANY; 1989 } 1990 1991 bd_speedup(); /* heeeelp */ 1992 1993 mtx_lock(&nblock); 1994 needsbuffer |= flags; 1995 while (needsbuffer & flags) { 1996 if (msleep(&needsbuffer, &nblock, 1997 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1998 mtx_unlock(&nblock); 1999 return (NULL); 2000 } 2001 } 2002 mtx_unlock(&nblock); 2003 } else { 2004 /* 2005 * We finally have a valid bp. We aren't quite out of the 2006 * woods, we still have to reserve kva space. In order 2007 * to keep fragmentation sane we only allocate kva in 2008 * BKVASIZE chunks. 2009 */ 2010 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2011 2012 if (maxsize != bp->b_kvasize) { 2013 vm_offset_t addr = 0; 2014 2015 bfreekva(bp); 2016 2017 if (vm_map_findspace(buffer_map, 2018 vm_map_min(buffer_map), maxsize, &addr)) { 2019 /* 2020 * Uh oh. Buffer map is to fragmented. We 2021 * must defragment the map. 2022 */ 2023 atomic_add_int(&bufdefragcnt, 1); 2024 defrag = 1; 2025 bp->b_flags |= B_INVAL; 2026 brelse(bp); 2027 goto restart; 2028 } 2029 if (addr) { 2030 vm_map_insert(buffer_map, NULL, 0, 2031 addr, addr + maxsize, 2032 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2033 2034 bp->b_kvabase = (caddr_t) addr; 2035 bp->b_kvasize = maxsize; 2036 atomic_add_int(&bufspace, bp->b_kvasize); 2037 atomic_add_int(&bufreusecnt, 1); 2038 } 2039 } 2040 bp->b_saveaddr = bp->b_kvabase; 2041 bp->b_data = bp->b_saveaddr; 2042 } 2043 return(bp); 2044} 2045 2046/* 2047 * buf_daemon: 2048 * 2049 * buffer flushing daemon. Buffers are normally flushed by the 2050 * update daemon but if it cannot keep up this process starts to 2051 * take the load in an attempt to prevent getnewbuf() from blocking. 2052 */ 2053 2054static struct kproc_desc buf_kp = { 2055 "bufdaemon", 2056 buf_daemon, 2057 &bufdaemonproc 2058}; 2059SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2060 2061static void 2062buf_daemon() 2063{ 2064 int s; 2065 2066 mtx_lock(&Giant); 2067 2068 /* 2069 * This process needs to be suspended prior to shutdown sync. 2070 */ 2071 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2072 SHUTDOWN_PRI_LAST); 2073 2074 /* 2075 * This process is allowed to take the buffer cache to the limit 2076 */ 2077 s = splbio(); 2078 mtx_lock(&bdlock); 2079 2080 for (;;) { 2081 bd_request = 0; 2082 mtx_unlock(&bdlock); 2083 2084 kthread_suspend_check(bufdaemonproc); 2085 2086 /* 2087 * Do the flush. Limit the amount of in-transit I/O we 2088 * allow to build up, otherwise we would completely saturate 2089 * the I/O system. Wakeup any waiting processes before we 2090 * normally would so they can run in parallel with our drain. 2091 */ 2092 while (numdirtybuffers > lodirtybuffers) { 2093 if (flushbufqueues(0) == 0) { 2094 /* 2095 * Could not find any buffers without rollback 2096 * dependencies, so just write the first one 2097 * in the hopes of eventually making progress. 2098 */ 2099 flushbufqueues(1); 2100 break; 2101 } 2102 waitrunningbufspace(); 2103 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2104 } 2105 2106 /* 2107 * Only clear bd_request if we have reached our low water 2108 * mark. The buf_daemon normally waits 1 second and 2109 * then incrementally flushes any dirty buffers that have 2110 * built up, within reason. 2111 * 2112 * If we were unable to hit our low water mark and couldn't 2113 * find any flushable buffers, we sleep half a second. 2114 * Otherwise we loop immediately. 2115 */ 2116 mtx_lock(&bdlock); 2117 if (numdirtybuffers <= lodirtybuffers) { 2118 /* 2119 * We reached our low water mark, reset the 2120 * request and sleep until we are needed again. 2121 * The sleep is just so the suspend code works. 2122 */ 2123 bd_request = 0; 2124 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2125 } else { 2126 /* 2127 * We couldn't find any flushable dirty buffers but 2128 * still have too many dirty buffers, we 2129 * have to sleep and try again. (rare) 2130 */ 2131 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2132 } 2133 } 2134} 2135 2136/* 2137 * flushbufqueues: 2138 * 2139 * Try to flush a buffer in the dirty queue. We must be careful to 2140 * free up B_INVAL buffers instead of write them, which NFS is 2141 * particularly sensitive to. 2142 */ 2143int flushwithdeps = 0; 2144SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2145 0, "Number of buffers flushed with dependecies that require rollbacks"); 2146 2147static int 2148flushbufqueues(int flushdeps) 2149{ 2150 struct thread *td = curthread; 2151 struct vnode *vp; 2152 struct mount *mp; 2153 struct buf *bp; 2154 int hasdeps; 2155 2156 mtx_lock(&bqlock); 2157 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2158 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2159 continue; 2160 KASSERT((bp->b_flags & B_DELWRI), 2161 ("unexpected clean buffer %p", bp)); 2162 BO_LOCK(bp->b_bufobj); 2163 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 2164 BO_UNLOCK(bp->b_bufobj); 2165 BUF_UNLOCK(bp); 2166 continue; 2167 } 2168 BO_UNLOCK(bp->b_bufobj); 2169 if (bp->b_flags & B_INVAL) { 2170 bremfreel(bp); 2171 mtx_unlock(&bqlock); 2172 brelse(bp); 2173 return (1); 2174 } 2175 2176 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2177 if (flushdeps == 0) { 2178 BUF_UNLOCK(bp); 2179 continue; 2180 } 2181 hasdeps = 1; 2182 } else 2183 hasdeps = 0; 2184 /* 2185 * We must hold the lock on a vnode before writing 2186 * one of its buffers. Otherwise we may confuse, or 2187 * in the case of a snapshot vnode, deadlock the 2188 * system. 2189 * 2190 * The lock order here is the reverse of the normal 2191 * of vnode followed by buf lock. This is ok because 2192 * the NOWAIT will prevent deadlock. 2193 */ 2194 vp = bp->b_vp; 2195 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2196 BUF_UNLOCK(bp); 2197 continue; 2198 } 2199 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2200 mtx_unlock(&bqlock); 2201 vfs_bio_awrite(bp); 2202 vn_finished_write(mp); 2203 VOP_UNLOCK(vp, 0, td); 2204 flushwithdeps += hasdeps; 2205 return (1); 2206 } 2207 vn_finished_write(mp); 2208 BUF_UNLOCK(bp); 2209 } 2210 mtx_unlock(&bqlock); 2211 return (0); 2212} 2213 2214/* 2215 * Check to see if a block is currently memory resident. 2216 */ 2217struct buf * 2218incore(struct bufobj *bo, daddr_t blkno) 2219{ 2220 struct buf *bp; 2221 2222 int s = splbio(); 2223 BO_LOCK(bo); 2224 bp = gbincore(bo, blkno); 2225 BO_UNLOCK(bo); 2226 splx(s); 2227 return (bp); 2228} 2229 2230/* 2231 * Returns true if no I/O is needed to access the 2232 * associated VM object. This is like incore except 2233 * it also hunts around in the VM system for the data. 2234 */ 2235 2236static int 2237inmem(struct vnode * vp, daddr_t blkno) 2238{ 2239 vm_object_t obj; 2240 vm_offset_t toff, tinc, size; 2241 vm_page_t m; 2242 vm_ooffset_t off; 2243 2244 GIANT_REQUIRED; 2245 ASSERT_VOP_LOCKED(vp, "inmem"); 2246 2247 if (incore(&vp->v_bufobj, blkno)) 2248 return 1; 2249 if (vp->v_mount == NULL) 2250 return 0; 2251 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2252 return 0; 2253 2254 size = PAGE_SIZE; 2255 if (size > vp->v_mount->mnt_stat.f_iosize) 2256 size = vp->v_mount->mnt_stat.f_iosize; 2257 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2258 2259 VM_OBJECT_LOCK(obj); 2260 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2261 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2262 if (!m) 2263 goto notinmem; 2264 tinc = size; 2265 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2266 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2267 if (vm_page_is_valid(m, 2268 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2269 goto notinmem; 2270 } 2271 VM_OBJECT_UNLOCK(obj); 2272 return 1; 2273 2274notinmem: 2275 VM_OBJECT_UNLOCK(obj); 2276 return (0); 2277} 2278 2279/* 2280 * vfs_setdirty: 2281 * 2282 * Sets the dirty range for a buffer based on the status of the dirty 2283 * bits in the pages comprising the buffer. 2284 * 2285 * The range is limited to the size of the buffer. 2286 * 2287 * This routine is primarily used by NFS, but is generalized for the 2288 * B_VMIO case. 2289 */ 2290static void 2291vfs_setdirty(struct buf *bp) 2292{ 2293 int i; 2294 vm_object_t object; 2295 2296 GIANT_REQUIRED; 2297 /* 2298 * Degenerate case - empty buffer 2299 */ 2300 2301 if (bp->b_bufsize == 0) 2302 return; 2303 2304 /* 2305 * We qualify the scan for modified pages on whether the 2306 * object has been flushed yet. The OBJ_WRITEABLE flag 2307 * is not cleared simply by protecting pages off. 2308 */ 2309 2310 if ((bp->b_flags & B_VMIO) == 0) 2311 return; 2312 2313 object = bp->b_pages[0]->object; 2314 VM_OBJECT_LOCK(object); 2315 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2316 printf("Warning: object %p writeable but not mightbedirty\n", object); 2317 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2318 printf("Warning: object %p mightbedirty but not writeable\n", object); 2319 2320 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2321 vm_offset_t boffset; 2322 vm_offset_t eoffset; 2323 2324 vm_page_lock_queues(); 2325 /* 2326 * test the pages to see if they have been modified directly 2327 * by users through the VM system. 2328 */ 2329 for (i = 0; i < bp->b_npages; i++) 2330 vm_page_test_dirty(bp->b_pages[i]); 2331 2332 /* 2333 * Calculate the encompassing dirty range, boffset and eoffset, 2334 * (eoffset - boffset) bytes. 2335 */ 2336 2337 for (i = 0; i < bp->b_npages; i++) { 2338 if (bp->b_pages[i]->dirty) 2339 break; 2340 } 2341 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2342 2343 for (i = bp->b_npages - 1; i >= 0; --i) { 2344 if (bp->b_pages[i]->dirty) { 2345 break; 2346 } 2347 } 2348 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2349 2350 vm_page_unlock_queues(); 2351 /* 2352 * Fit it to the buffer. 2353 */ 2354 2355 if (eoffset > bp->b_bcount) 2356 eoffset = bp->b_bcount; 2357 2358 /* 2359 * If we have a good dirty range, merge with the existing 2360 * dirty range. 2361 */ 2362 2363 if (boffset < eoffset) { 2364 if (bp->b_dirtyoff > boffset) 2365 bp->b_dirtyoff = boffset; 2366 if (bp->b_dirtyend < eoffset) 2367 bp->b_dirtyend = eoffset; 2368 } 2369 } 2370 VM_OBJECT_UNLOCK(object); 2371} 2372 2373/* 2374 * getblk: 2375 * 2376 * Get a block given a specified block and offset into a file/device. 2377 * The buffers B_DONE bit will be cleared on return, making it almost 2378 * ready for an I/O initiation. B_INVAL may or may not be set on 2379 * return. The caller should clear B_INVAL prior to initiating a 2380 * READ. 2381 * 2382 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2383 * an existing buffer. 2384 * 2385 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2386 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2387 * and then cleared based on the backing VM. If the previous buffer is 2388 * non-0-sized but invalid, B_CACHE will be cleared. 2389 * 2390 * If getblk() must create a new buffer, the new buffer is returned with 2391 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2392 * case it is returned with B_INVAL clear and B_CACHE set based on the 2393 * backing VM. 2394 * 2395 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2396 * B_CACHE bit is clear. 2397 * 2398 * What this means, basically, is that the caller should use B_CACHE to 2399 * determine whether the buffer is fully valid or not and should clear 2400 * B_INVAL prior to issuing a read. If the caller intends to validate 2401 * the buffer by loading its data area with something, the caller needs 2402 * to clear B_INVAL. If the caller does this without issuing an I/O, 2403 * the caller should set B_CACHE ( as an optimization ), else the caller 2404 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2405 * a write attempt or if it was a successfull read. If the caller 2406 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2407 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2408 */ 2409struct buf * 2410getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2411 int flags) 2412{ 2413 struct buf *bp; 2414 struct bufobj *bo; 2415 int s; 2416 int error; 2417 ASSERT_VOP_LOCKED(vp, "getblk"); 2418 struct vm_object *vmo; 2419 2420 if (size > MAXBSIZE) 2421 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2422 2423 bo = &vp->v_bufobj; 2424 s = splbio(); 2425loop: 2426 /* 2427 * Block if we are low on buffers. Certain processes are allowed 2428 * to completely exhaust the buffer cache. 2429 * 2430 * If this check ever becomes a bottleneck it may be better to 2431 * move it into the else, when gbincore() fails. At the moment 2432 * it isn't a problem. 2433 * 2434 * XXX remove if 0 sections (clean this up after its proven) 2435 */ 2436 if (numfreebuffers == 0) { 2437 if (curthread == PCPU_GET(idlethread)) 2438 return NULL; 2439 mtx_lock(&nblock); 2440 needsbuffer |= VFS_BIO_NEED_ANY; 2441 mtx_unlock(&nblock); 2442 } 2443 2444 VI_LOCK(vp); 2445 bp = gbincore(bo, blkno); 2446 if (bp != NULL) { 2447 int lockflags; 2448 /* 2449 * Buffer is in-core. If the buffer is not busy, it must 2450 * be on a queue. 2451 */ 2452 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2453 2454 if (flags & GB_LOCK_NOWAIT) 2455 lockflags |= LK_NOWAIT; 2456 2457 error = BUF_TIMELOCK(bp, lockflags, 2458 VI_MTX(vp), "getblk", slpflag, slptimeo); 2459 2460 /* 2461 * If we slept and got the lock we have to restart in case 2462 * the buffer changed identities. 2463 */ 2464 if (error == ENOLCK) 2465 goto loop; 2466 /* We timed out or were interrupted. */ 2467 else if (error) 2468 return (NULL); 2469 2470 /* 2471 * The buffer is locked. B_CACHE is cleared if the buffer is 2472 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2473 * and for a VMIO buffer B_CACHE is adjusted according to the 2474 * backing VM cache. 2475 */ 2476 if (bp->b_flags & B_INVAL) 2477 bp->b_flags &= ~B_CACHE; 2478 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2479 bp->b_flags |= B_CACHE; 2480 bremfree(bp); 2481 2482 /* 2483 * check for size inconsistancies for non-VMIO case. 2484 */ 2485 2486 if (bp->b_bcount != size) { 2487 if ((bp->b_flags & B_VMIO) == 0 || 2488 (size > bp->b_kvasize)) { 2489 if (bp->b_flags & B_DELWRI) { 2490 bp->b_flags |= B_NOCACHE; 2491 bwrite(bp); 2492 } else { 2493 if ((bp->b_flags & B_VMIO) && 2494 (LIST_FIRST(&bp->b_dep) == NULL)) { 2495 bp->b_flags |= B_RELBUF; 2496 brelse(bp); 2497 } else { 2498 bp->b_flags |= B_NOCACHE; 2499 bwrite(bp); 2500 } 2501 } 2502 goto loop; 2503 } 2504 } 2505 2506 /* 2507 * If the size is inconsistant in the VMIO case, we can resize 2508 * the buffer. This might lead to B_CACHE getting set or 2509 * cleared. If the size has not changed, B_CACHE remains 2510 * unchanged from its previous state. 2511 */ 2512 2513 if (bp->b_bcount != size) 2514 allocbuf(bp, size); 2515 2516 KASSERT(bp->b_offset != NOOFFSET, 2517 ("getblk: no buffer offset")); 2518 2519 /* 2520 * A buffer with B_DELWRI set and B_CACHE clear must 2521 * be committed before we can return the buffer in 2522 * order to prevent the caller from issuing a read 2523 * ( due to B_CACHE not being set ) and overwriting 2524 * it. 2525 * 2526 * Most callers, including NFS and FFS, need this to 2527 * operate properly either because they assume they 2528 * can issue a read if B_CACHE is not set, or because 2529 * ( for example ) an uncached B_DELWRI might loop due 2530 * to softupdates re-dirtying the buffer. In the latter 2531 * case, B_CACHE is set after the first write completes, 2532 * preventing further loops. 2533 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2534 * above while extending the buffer, we cannot allow the 2535 * buffer to remain with B_CACHE set after the write 2536 * completes or it will represent a corrupt state. To 2537 * deal with this we set B_NOCACHE to scrap the buffer 2538 * after the write. 2539 * 2540 * We might be able to do something fancy, like setting 2541 * B_CACHE in bwrite() except if B_DELWRI is already set, 2542 * so the below call doesn't set B_CACHE, but that gets real 2543 * confusing. This is much easier. 2544 */ 2545 2546 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2547 bp->b_flags |= B_NOCACHE; 2548 bwrite(bp); 2549 goto loop; 2550 } 2551 2552 splx(s); 2553 bp->b_flags &= ~B_DONE; 2554 } else { 2555 int bsize, maxsize, vmio; 2556 off_t offset; 2557 2558 /* 2559 * Buffer is not in-core, create new buffer. The buffer 2560 * returned by getnewbuf() is locked. Note that the returned 2561 * buffer is also considered valid (not marked B_INVAL). 2562 */ 2563 VI_UNLOCK(vp); 2564 /* 2565 * If the user does not want us to create the buffer, bail out 2566 * here. 2567 */ 2568 if (flags & GB_NOCREAT) { 2569 splx(s); 2570 return NULL; 2571 } 2572 2573 bsize = bo->bo_bsize; 2574 offset = blkno * bsize; 2575 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2576 (vp->v_vflag & VV_OBJBUF); 2577 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2578 maxsize = imax(maxsize, bsize); 2579 2580 bp = getnewbuf(slpflag, slptimeo, size, maxsize); 2581 if (bp == NULL) { 2582 if (slpflag || slptimeo) { 2583 splx(s); 2584 return NULL; 2585 } 2586 goto loop; 2587 } 2588 2589 /* 2590 * This code is used to make sure that a buffer is not 2591 * created while the getnewbuf routine is blocked. 2592 * This can be a problem whether the vnode is locked or not. 2593 * If the buffer is created out from under us, we have to 2594 * throw away the one we just created. There is now window 2595 * race because we are safely running at splbio() from the 2596 * point of the duplicate buffer creation through to here, 2597 * and we've locked the buffer. 2598 * 2599 * Note: this must occur before we associate the buffer 2600 * with the vp especially considering limitations in 2601 * the splay tree implementation when dealing with duplicate 2602 * lblkno's. 2603 */ 2604 BO_LOCK(bo); 2605 if (gbincore(bo, blkno)) { 2606 BO_UNLOCK(bo); 2607 bp->b_flags |= B_INVAL; 2608 brelse(bp); 2609 goto loop; 2610 } 2611 2612 /* 2613 * Insert the buffer into the hash, so that it can 2614 * be found by incore. 2615 */ 2616 bp->b_blkno = bp->b_lblkno = blkno; 2617 bp->b_offset = offset; 2618 2619 bgetvp(vp, bp); 2620 BO_UNLOCK(bo); 2621 2622 /* 2623 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2624 * buffer size starts out as 0, B_CACHE will be set by 2625 * allocbuf() for the VMIO case prior to it testing the 2626 * backing store for validity. 2627 */ 2628 2629 if (vmio) { 2630 bp->b_flags |= B_VMIO; 2631#if defined(VFS_BIO_DEBUG) 2632 if (vn_canvmio(vp) != TRUE) 2633 printf("getblk: VMIO on vnode type %d\n", 2634 vp->v_type); 2635#endif 2636 VOP_GETVOBJECT(vp, &vmo); 2637 KASSERT(vmo == bp->b_bufobj->bo_object, 2638 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2639 bp, vmo, bp->b_bufobj->bo_object)); 2640 } else { 2641 bp->b_flags &= ~B_VMIO; 2642 KASSERT(bp->b_bufobj->bo_object == NULL, 2643 ("ARGH! has b_bufobj->bo_object %p %p\n", 2644 bp, bp->b_bufobj->bo_object)); 2645 } 2646 2647 allocbuf(bp, size); 2648 2649 splx(s); 2650 bp->b_flags &= ~B_DONE; 2651 } 2652 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2653 KASSERT(bp->b_bufobj == bo, 2654 ("wrong b_bufobj %p should be %p", bp->b_bufobj, bo)); 2655 return (bp); 2656} 2657 2658/* 2659 * Get an empty, disassociated buffer of given size. The buffer is initially 2660 * set to B_INVAL. 2661 */ 2662struct buf * 2663geteblk(int size) 2664{ 2665 struct buf *bp; 2666 int s; 2667 int maxsize; 2668 2669 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2670 2671 s = splbio(); 2672 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2673 continue; 2674 splx(s); 2675 allocbuf(bp, size); 2676 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2677 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2678 return (bp); 2679} 2680 2681 2682/* 2683 * This code constitutes the buffer memory from either anonymous system 2684 * memory (in the case of non-VMIO operations) or from an associated 2685 * VM object (in the case of VMIO operations). This code is able to 2686 * resize a buffer up or down. 2687 * 2688 * Note that this code is tricky, and has many complications to resolve 2689 * deadlock or inconsistant data situations. Tread lightly!!! 2690 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2691 * the caller. Calling this code willy nilly can result in the loss of data. 2692 * 2693 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2694 * B_CACHE for the non-VMIO case. 2695 */ 2696 2697int 2698allocbuf(struct buf *bp, int size) 2699{ 2700 int newbsize, mbsize; 2701 int i; 2702 2703 GIANT_REQUIRED; 2704 2705 if (BUF_REFCNT(bp) == 0) 2706 panic("allocbuf: buffer not busy"); 2707 2708 if (bp->b_kvasize < size) 2709 panic("allocbuf: buffer too small"); 2710 2711 if ((bp->b_flags & B_VMIO) == 0) { 2712 caddr_t origbuf; 2713 int origbufsize; 2714 /* 2715 * Just get anonymous memory from the kernel. Don't 2716 * mess with B_CACHE. 2717 */ 2718 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2719 if (bp->b_flags & B_MALLOC) 2720 newbsize = mbsize; 2721 else 2722 newbsize = round_page(size); 2723 2724 if (newbsize < bp->b_bufsize) { 2725 /* 2726 * malloced buffers are not shrunk 2727 */ 2728 if (bp->b_flags & B_MALLOC) { 2729 if (newbsize) { 2730 bp->b_bcount = size; 2731 } else { 2732 free(bp->b_data, M_BIOBUF); 2733 if (bp->b_bufsize) { 2734 atomic_subtract_int( 2735 &bufmallocspace, 2736 bp->b_bufsize); 2737 bufspacewakeup(); 2738 bp->b_bufsize = 0; 2739 } 2740 bp->b_saveaddr = bp->b_kvabase; 2741 bp->b_data = bp->b_saveaddr; 2742 bp->b_bcount = 0; 2743 bp->b_flags &= ~B_MALLOC; 2744 } 2745 return 1; 2746 } 2747 vm_hold_free_pages( 2748 bp, 2749 (vm_offset_t) bp->b_data + newbsize, 2750 (vm_offset_t) bp->b_data + bp->b_bufsize); 2751 } else if (newbsize > bp->b_bufsize) { 2752 /* 2753 * We only use malloced memory on the first allocation. 2754 * and revert to page-allocated memory when the buffer 2755 * grows. 2756 */ 2757 /* 2758 * There is a potential smp race here that could lead 2759 * to bufmallocspace slightly passing the max. It 2760 * is probably extremely rare and not worth worrying 2761 * over. 2762 */ 2763 if ( (bufmallocspace < maxbufmallocspace) && 2764 (bp->b_bufsize == 0) && 2765 (mbsize <= PAGE_SIZE/2)) { 2766 2767 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2768 bp->b_bufsize = mbsize; 2769 bp->b_bcount = size; 2770 bp->b_flags |= B_MALLOC; 2771 atomic_add_int(&bufmallocspace, mbsize); 2772 return 1; 2773 } 2774 origbuf = NULL; 2775 origbufsize = 0; 2776 /* 2777 * If the buffer is growing on its other-than-first allocation, 2778 * then we revert to the page-allocation scheme. 2779 */ 2780 if (bp->b_flags & B_MALLOC) { 2781 origbuf = bp->b_data; 2782 origbufsize = bp->b_bufsize; 2783 bp->b_data = bp->b_kvabase; 2784 if (bp->b_bufsize) { 2785 atomic_subtract_int(&bufmallocspace, 2786 bp->b_bufsize); 2787 bufspacewakeup(); 2788 bp->b_bufsize = 0; 2789 } 2790 bp->b_flags &= ~B_MALLOC; 2791 newbsize = round_page(newbsize); 2792 } 2793 vm_hold_load_pages( 2794 bp, 2795 (vm_offset_t) bp->b_data + bp->b_bufsize, 2796 (vm_offset_t) bp->b_data + newbsize); 2797 if (origbuf) { 2798 bcopy(origbuf, bp->b_data, origbufsize); 2799 free(origbuf, M_BIOBUF); 2800 } 2801 } 2802 } else { 2803 int desiredpages; 2804 2805 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2806 desiredpages = (size == 0) ? 0 : 2807 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2808 2809 if (bp->b_flags & B_MALLOC) 2810 panic("allocbuf: VMIO buffer can't be malloced"); 2811 /* 2812 * Set B_CACHE initially if buffer is 0 length or will become 2813 * 0-length. 2814 */ 2815 if (size == 0 || bp->b_bufsize == 0) 2816 bp->b_flags |= B_CACHE; 2817 2818 if (newbsize < bp->b_bufsize) { 2819 /* 2820 * DEV_BSIZE aligned new buffer size is less then the 2821 * DEV_BSIZE aligned existing buffer size. Figure out 2822 * if we have to remove any pages. 2823 */ 2824 if (desiredpages < bp->b_npages) { 2825 vm_page_t m; 2826 2827 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2828 vm_page_lock_queues(); 2829 for (i = desiredpages; i < bp->b_npages; i++) { 2830 /* 2831 * the page is not freed here -- it 2832 * is the responsibility of 2833 * vnode_pager_setsize 2834 */ 2835 m = bp->b_pages[i]; 2836 KASSERT(m != bogus_page, 2837 ("allocbuf: bogus page found")); 2838 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2839 vm_page_lock_queues(); 2840 2841 bp->b_pages[i] = NULL; 2842 vm_page_unwire(m, 0); 2843 } 2844 vm_page_unlock_queues(); 2845 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2846 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2847 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2848 bp->b_npages = desiredpages; 2849 } 2850 } else if (size > bp->b_bcount) { 2851 /* 2852 * We are growing the buffer, possibly in a 2853 * byte-granular fashion. 2854 */ 2855 struct vnode *vp; 2856 vm_object_t obj; 2857 vm_offset_t toff; 2858 vm_offset_t tinc; 2859 2860 /* 2861 * Step 1, bring in the VM pages from the object, 2862 * allocating them if necessary. We must clear 2863 * B_CACHE if these pages are not valid for the 2864 * range covered by the buffer. 2865 */ 2866 2867 vp = bp->b_vp; 2868 obj = bp->b_bufobj->bo_object; 2869 2870 VM_OBJECT_LOCK(obj); 2871 while (bp->b_npages < desiredpages) { 2872 vm_page_t m; 2873 vm_pindex_t pi; 2874 2875 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2876 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2877 /* 2878 * note: must allocate system pages 2879 * since blocking here could intefere 2880 * with paging I/O, no matter which 2881 * process we are. 2882 */ 2883 m = vm_page_alloc(obj, pi, 2884 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2885 VM_ALLOC_WIRED); 2886 if (m == NULL) { 2887 atomic_add_int(&vm_pageout_deficit, 2888 desiredpages - bp->b_npages); 2889 VM_OBJECT_UNLOCK(obj); 2890 VM_WAIT; 2891 VM_OBJECT_LOCK(obj); 2892 } else { 2893 bp->b_flags &= ~B_CACHE; 2894 bp->b_pages[bp->b_npages] = m; 2895 ++bp->b_npages; 2896 } 2897 continue; 2898 } 2899 2900 /* 2901 * We found a page. If we have to sleep on it, 2902 * retry because it might have gotten freed out 2903 * from under us. 2904 * 2905 * We can only test PG_BUSY here. Blocking on 2906 * m->busy might lead to a deadlock: 2907 * 2908 * vm_fault->getpages->cluster_read->allocbuf 2909 * 2910 */ 2911 vm_page_lock_queues(); 2912 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2913 continue; 2914 2915 /* 2916 * We have a good page. Should we wakeup the 2917 * page daemon? 2918 */ 2919 if ((curproc != pageproc) && 2920 ((m->queue - m->pc) == PQ_CACHE) && 2921 ((cnt.v_free_count + cnt.v_cache_count) < 2922 (cnt.v_free_min + cnt.v_cache_min))) { 2923 pagedaemon_wakeup(); 2924 } 2925 vm_page_wire(m); 2926 vm_page_unlock_queues(); 2927 bp->b_pages[bp->b_npages] = m; 2928 ++bp->b_npages; 2929 } 2930 2931 /* 2932 * Step 2. We've loaded the pages into the buffer, 2933 * we have to figure out if we can still have B_CACHE 2934 * set. Note that B_CACHE is set according to the 2935 * byte-granular range ( bcount and size ), new the 2936 * aligned range ( newbsize ). 2937 * 2938 * The VM test is against m->valid, which is DEV_BSIZE 2939 * aligned. Needless to say, the validity of the data 2940 * needs to also be DEV_BSIZE aligned. Note that this 2941 * fails with NFS if the server or some other client 2942 * extends the file's EOF. If our buffer is resized, 2943 * B_CACHE may remain set! XXX 2944 */ 2945 2946 toff = bp->b_bcount; 2947 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2948 2949 while ((bp->b_flags & B_CACHE) && toff < size) { 2950 vm_pindex_t pi; 2951 2952 if (tinc > (size - toff)) 2953 tinc = size - toff; 2954 2955 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2956 PAGE_SHIFT; 2957 2958 vfs_buf_test_cache( 2959 bp, 2960 bp->b_offset, 2961 toff, 2962 tinc, 2963 bp->b_pages[pi] 2964 ); 2965 toff += tinc; 2966 tinc = PAGE_SIZE; 2967 } 2968 VM_OBJECT_UNLOCK(obj); 2969 2970 /* 2971 * Step 3, fixup the KVM pmap. Remember that 2972 * bp->b_data is relative to bp->b_offset, but 2973 * bp->b_offset may be offset into the first page. 2974 */ 2975 2976 bp->b_data = (caddr_t) 2977 trunc_page((vm_offset_t)bp->b_data); 2978 pmap_qenter( 2979 (vm_offset_t)bp->b_data, 2980 bp->b_pages, 2981 bp->b_npages 2982 ); 2983 2984 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2985 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2986 } 2987 } 2988 if (newbsize < bp->b_bufsize) 2989 bufspacewakeup(); 2990 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2991 bp->b_bcount = size; /* requested buffer size */ 2992 return 1; 2993} 2994 2995void 2996biodone(struct bio *bp) 2997{ 2998 2999 mtx_lock(&bdonelock); 3000 bp->bio_flags |= BIO_DONE; 3001 if (bp->bio_done == NULL) 3002 wakeup(bp); 3003 mtx_unlock(&bdonelock); 3004 if (bp->bio_done != NULL) 3005 bp->bio_done(bp); 3006} 3007 3008/* 3009 * Wait for a BIO to finish. 3010 * 3011 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3012 * case is not yet clear. 3013 */ 3014int 3015biowait(struct bio *bp, const char *wchan) 3016{ 3017 3018 mtx_lock(&bdonelock); 3019 while ((bp->bio_flags & BIO_DONE) == 0) 3020 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3021 mtx_unlock(&bdonelock); 3022 if (bp->bio_error != 0) 3023 return (bp->bio_error); 3024 if (!(bp->bio_flags & BIO_ERROR)) 3025 return (0); 3026 return (EIO); 3027} 3028 3029void 3030biofinish(struct bio *bp, struct devstat *stat, int error) 3031{ 3032 3033 if (error) { 3034 bp->bio_error = error; 3035 bp->bio_flags |= BIO_ERROR; 3036 } 3037 if (stat != NULL) 3038 devstat_end_transaction_bio(stat, bp); 3039 biodone(bp); 3040} 3041 3042/* 3043 * bufwait: 3044 * 3045 * Wait for buffer I/O completion, returning error status. The buffer 3046 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3047 * error and cleared. 3048 */ 3049int 3050bufwait(struct buf *bp) 3051{ 3052 int s; 3053 3054 s = splbio(); 3055 if (bp->b_iocmd == BIO_READ) 3056 bwait(bp, PRIBIO, "biord"); 3057 else 3058 bwait(bp, PRIBIO, "biowr"); 3059 splx(s); 3060 if (bp->b_flags & B_EINTR) { 3061 bp->b_flags &= ~B_EINTR; 3062 return (EINTR); 3063 } 3064 if (bp->b_ioflags & BIO_ERROR) { 3065 return (bp->b_error ? bp->b_error : EIO); 3066 } else { 3067 return (0); 3068 } 3069} 3070 3071 /* 3072 * Call back function from struct bio back up to struct buf. 3073 */ 3074static void 3075bufdonebio(struct bio *bip) 3076{ 3077 struct buf *bp; 3078 3079 /* Device drivers may or may not hold giant, hold it here. */ 3080 mtx_lock(&Giant); 3081 bp = bip->bio_caller2; 3082 bp->b_resid = bp->b_bcount - bip->bio_completed; 3083 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3084 bp->b_ioflags = bip->bio_flags; 3085 bp->b_error = bip->bio_error; 3086 if (bp->b_error) 3087 bp->b_ioflags |= BIO_ERROR; 3088 bufdone(bp); 3089 mtx_unlock(&Giant); 3090 g_destroy_bio(bip); 3091} 3092 3093void 3094dev_strategy(struct cdev *dev, struct buf *bp) 3095{ 3096 struct cdevsw *csw; 3097 struct bio *bip; 3098 3099 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3100 panic("b_iocmd botch"); 3101 for (;;) { 3102 bip = g_new_bio(); 3103 if (bip != NULL) 3104 break; 3105 /* Try again later */ 3106 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3107 } 3108 bip->bio_cmd = bp->b_iocmd; 3109 bip->bio_offset = bp->b_iooffset; 3110 bip->bio_length = bp->b_bcount; 3111 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3112 bip->bio_data = bp->b_data; 3113 bip->bio_done = bufdonebio; 3114 bip->bio_caller2 = bp; 3115 bip->bio_dev = dev; 3116 KASSERT(dev->si_refcount > 0, 3117 ("dev_strategy on un-referenced struct cdev *(%s)", 3118 devtoname(dev))); 3119 csw = dev_refthread(dev); 3120 if (csw == NULL) { 3121 bp->b_error = ENXIO; 3122 bp->b_ioflags = BIO_ERROR; 3123 mtx_lock(&Giant); /* XXX: too defensive ? */ 3124 bufdone(bp); 3125 mtx_unlock(&Giant); /* XXX: too defensive ? */ 3126 return; 3127 } 3128 (*csw->d_strategy)(bip); 3129 dev_relthread(dev); 3130} 3131 3132/* 3133 * bufdone: 3134 * 3135 * Finish I/O on a buffer, optionally calling a completion function. 3136 * This is usually called from an interrupt so process blocking is 3137 * not allowed. 3138 * 3139 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3140 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3141 * assuming B_INVAL is clear. 3142 * 3143 * For the VMIO case, we set B_CACHE if the op was a read and no 3144 * read error occured, or if the op was a write. B_CACHE is never 3145 * set if the buffer is invalid or otherwise uncacheable. 3146 * 3147 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3148 * initiator to leave B_INVAL set to brelse the buffer out of existance 3149 * in the biodone routine. 3150 */ 3151void 3152bufdone(struct buf *bp) 3153{ 3154 int s; 3155 void (*biodone)(struct buf *); 3156 3157 3158 s = splbio(); 3159 3160 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3161 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3162 3163 bp->b_flags |= B_DONE; 3164 runningbufwakeup(bp); 3165 3166 if (bp->b_iocmd == BIO_WRITE && bp->b_bufobj != NULL) 3167 bufobj_wdrop(bp->b_bufobj); 3168 3169 /* call optional completion function if requested */ 3170 if (bp->b_iodone != NULL) { 3171 biodone = bp->b_iodone; 3172 bp->b_iodone = NULL; 3173 (*biodone) (bp); 3174 splx(s); 3175 return; 3176 } 3177 if (LIST_FIRST(&bp->b_dep) != NULL) 3178 buf_complete(bp); 3179 3180 if (bp->b_flags & B_VMIO) { 3181 int i; 3182 vm_ooffset_t foff; 3183 vm_page_t m; 3184 vm_object_t obj; 3185 int iosize; 3186 struct vnode *vp = bp->b_vp; 3187 3188 obj = bp->b_bufobj->bo_object; 3189 3190#if defined(VFS_BIO_DEBUG) 3191 mp_fixme("usecount and vflag accessed without locks."); 3192 if (vp->v_usecount == 0) { 3193 panic("biodone: zero vnode ref count"); 3194 } 3195 3196 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3197 panic("biodone: vnode is not setup for merged cache"); 3198 } 3199#endif 3200 3201 foff = bp->b_offset; 3202 KASSERT(bp->b_offset != NOOFFSET, 3203 ("biodone: no buffer offset")); 3204 3205 VM_OBJECT_LOCK(obj); 3206#if defined(VFS_BIO_DEBUG) 3207 if (obj->paging_in_progress < bp->b_npages) { 3208 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3209 obj->paging_in_progress, bp->b_npages); 3210 } 3211#endif 3212 3213 /* 3214 * Set B_CACHE if the op was a normal read and no error 3215 * occured. B_CACHE is set for writes in the b*write() 3216 * routines. 3217 */ 3218 iosize = bp->b_bcount - bp->b_resid; 3219 if (bp->b_iocmd == BIO_READ && 3220 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3221 !(bp->b_ioflags & BIO_ERROR)) { 3222 bp->b_flags |= B_CACHE; 3223 } 3224 vm_page_lock_queues(); 3225 for (i = 0; i < bp->b_npages; i++) { 3226 int bogusflag = 0; 3227 int resid; 3228 3229 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3230 if (resid > iosize) 3231 resid = iosize; 3232 3233 /* 3234 * cleanup bogus pages, restoring the originals 3235 */ 3236 m = bp->b_pages[i]; 3237 if (m == bogus_page) { 3238 bogusflag = 1; 3239 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3240 if (m == NULL) 3241 panic("biodone: page disappeared!"); 3242 bp->b_pages[i] = m; 3243 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3244 } 3245#if defined(VFS_BIO_DEBUG) 3246 if (OFF_TO_IDX(foff) != m->pindex) { 3247 printf( 3248"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3249 (intmax_t)foff, (uintmax_t)m->pindex); 3250 } 3251#endif 3252 3253 /* 3254 * In the write case, the valid and clean bits are 3255 * already changed correctly ( see bdwrite() ), so we 3256 * only need to do this here in the read case. 3257 */ 3258 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3259 vfs_page_set_valid(bp, foff, i, m); 3260 } 3261 3262 /* 3263 * when debugging new filesystems or buffer I/O methods, this 3264 * is the most common error that pops up. if you see this, you 3265 * have not set the page busy flag correctly!!! 3266 */ 3267 if (m->busy == 0) { 3268 printf("biodone: page busy < 0, " 3269 "pindex: %d, foff: 0x(%x,%x), " 3270 "resid: %d, index: %d\n", 3271 (int) m->pindex, (int)(foff >> 32), 3272 (int) foff & 0xffffffff, resid, i); 3273 if (!vn_isdisk(vp, NULL)) 3274 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3275 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3276 (intmax_t) bp->b_lblkno, 3277 bp->b_flags, bp->b_npages); 3278 else 3279 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3280 (intmax_t) bp->b_lblkno, 3281 bp->b_flags, bp->b_npages); 3282 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3283 (u_long)m->valid, (u_long)m->dirty, 3284 m->wire_count); 3285 panic("biodone: page busy < 0\n"); 3286 } 3287 vm_page_io_finish(m); 3288 vm_object_pip_subtract(obj, 1); 3289 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3290 iosize -= resid; 3291 } 3292 vm_page_unlock_queues(); 3293 vm_object_pip_wakeupn(obj, 0); 3294 VM_OBJECT_UNLOCK(obj); 3295 } 3296 3297 /* 3298 * For asynchronous completions, release the buffer now. The brelse 3299 * will do a wakeup there if necessary - so no need to do a wakeup 3300 * here in the async case. The sync case always needs to do a wakeup. 3301 */ 3302 3303 if (bp->b_flags & B_ASYNC) { 3304 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3305 brelse(bp); 3306 else 3307 bqrelse(bp); 3308 } else { 3309 bdone(bp); 3310 } 3311 splx(s); 3312} 3313 3314/* 3315 * This routine is called in lieu of iodone in the case of 3316 * incomplete I/O. This keeps the busy status for pages 3317 * consistant. 3318 */ 3319void 3320vfs_unbusy_pages(struct buf *bp) 3321{ 3322 int i; 3323 vm_object_t obj; 3324 vm_page_t m; 3325 3326 runningbufwakeup(bp); 3327 if (!(bp->b_flags & B_VMIO)) 3328 return; 3329 3330 obj = bp->b_bufobj->bo_object; 3331 VM_OBJECT_LOCK(obj); 3332 vm_page_lock_queues(); 3333 for (i = 0; i < bp->b_npages; i++) { 3334 m = bp->b_pages[i]; 3335 if (m == bogus_page) { 3336 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3337 if (!m) 3338 panic("vfs_unbusy_pages: page missing\n"); 3339 bp->b_pages[i] = m; 3340 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3341 bp->b_pages, bp->b_npages); 3342 } 3343 vm_object_pip_subtract(obj, 1); 3344 vm_page_io_finish(m); 3345 } 3346 vm_page_unlock_queues(); 3347 vm_object_pip_wakeupn(obj, 0); 3348 VM_OBJECT_UNLOCK(obj); 3349} 3350 3351/* 3352 * vfs_page_set_valid: 3353 * 3354 * Set the valid bits in a page based on the supplied offset. The 3355 * range is restricted to the buffer's size. 3356 * 3357 * This routine is typically called after a read completes. 3358 */ 3359static void 3360vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3361{ 3362 vm_ooffset_t soff, eoff; 3363 3364 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3365 /* 3366 * Start and end offsets in buffer. eoff - soff may not cross a 3367 * page boundry or cross the end of the buffer. The end of the 3368 * buffer, in this case, is our file EOF, not the allocation size 3369 * of the buffer. 3370 */ 3371 soff = off; 3372 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3373 if (eoff > bp->b_offset + bp->b_bcount) 3374 eoff = bp->b_offset + bp->b_bcount; 3375 3376 /* 3377 * Set valid range. This is typically the entire buffer and thus the 3378 * entire page. 3379 */ 3380 if (eoff > soff) { 3381 vm_page_set_validclean( 3382 m, 3383 (vm_offset_t) (soff & PAGE_MASK), 3384 (vm_offset_t) (eoff - soff) 3385 ); 3386 } 3387} 3388 3389/* 3390 * This routine is called before a device strategy routine. 3391 * It is used to tell the VM system that paging I/O is in 3392 * progress, and treat the pages associated with the buffer 3393 * almost as being PG_BUSY. Also the object paging_in_progress 3394 * flag is handled to make sure that the object doesn't become 3395 * inconsistant. 3396 * 3397 * Since I/O has not been initiated yet, certain buffer flags 3398 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3399 * and should be ignored. 3400 */ 3401void 3402vfs_busy_pages(struct buf *bp, int clear_modify) 3403{ 3404 int i, bogus; 3405 vm_object_t obj; 3406 vm_ooffset_t foff; 3407 vm_page_t m; 3408 3409 if (!(bp->b_flags & B_VMIO)) 3410 return; 3411 3412 obj = bp->b_bufobj->bo_object; 3413 foff = bp->b_offset; 3414 KASSERT(bp->b_offset != NOOFFSET, 3415 ("vfs_busy_pages: no buffer offset")); 3416 vfs_setdirty(bp); 3417 VM_OBJECT_LOCK(obj); 3418retry: 3419 vm_page_lock_queues(); 3420 for (i = 0; i < bp->b_npages; i++) { 3421 m = bp->b_pages[i]; 3422 3423 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3424 goto retry; 3425 } 3426 bogus = 0; 3427 for (i = 0; i < bp->b_npages; i++) { 3428 m = bp->b_pages[i]; 3429 3430 if ((bp->b_flags & B_CLUSTER) == 0) { 3431 vm_object_pip_add(obj, 1); 3432 vm_page_io_start(m); 3433 } 3434 /* 3435 * When readying a buffer for a read ( i.e 3436 * clear_modify == 0 ), it is important to do 3437 * bogus_page replacement for valid pages in 3438 * partially instantiated buffers. Partially 3439 * instantiated buffers can, in turn, occur when 3440 * reconstituting a buffer from its VM backing store 3441 * base. We only have to do this if B_CACHE is 3442 * clear ( which causes the I/O to occur in the 3443 * first place ). The replacement prevents the read 3444 * I/O from overwriting potentially dirty VM-backed 3445 * pages. XXX bogus page replacement is, uh, bogus. 3446 * It may not work properly with small-block devices. 3447 * We need to find a better way. 3448 */ 3449 pmap_remove_all(m); 3450 if (clear_modify) 3451 vfs_page_set_valid(bp, foff, i, m); 3452 else if (m->valid == VM_PAGE_BITS_ALL && 3453 (bp->b_flags & B_CACHE) == 0) { 3454 bp->b_pages[i] = bogus_page; 3455 bogus++; 3456 } 3457 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3458 } 3459 vm_page_unlock_queues(); 3460 VM_OBJECT_UNLOCK(obj); 3461 if (bogus) 3462 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3463 bp->b_pages, bp->b_npages); 3464} 3465 3466/* 3467 * Tell the VM system that the pages associated with this buffer 3468 * are clean. This is used for delayed writes where the data is 3469 * going to go to disk eventually without additional VM intevention. 3470 * 3471 * Note that while we only really need to clean through to b_bcount, we 3472 * just go ahead and clean through to b_bufsize. 3473 */ 3474static void 3475vfs_clean_pages(struct buf *bp) 3476{ 3477 int i; 3478 vm_ooffset_t foff, noff, eoff; 3479 vm_page_t m; 3480 3481 if (!(bp->b_flags & B_VMIO)) 3482 return; 3483 3484 foff = bp->b_offset; 3485 KASSERT(bp->b_offset != NOOFFSET, 3486 ("vfs_clean_pages: no buffer offset")); 3487 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3488 vm_page_lock_queues(); 3489 for (i = 0; i < bp->b_npages; i++) { 3490 m = bp->b_pages[i]; 3491 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3492 eoff = noff; 3493 3494 if (eoff > bp->b_offset + bp->b_bufsize) 3495 eoff = bp->b_offset + bp->b_bufsize; 3496 vfs_page_set_valid(bp, foff, i, m); 3497 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3498 foff = noff; 3499 } 3500 vm_page_unlock_queues(); 3501 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3502} 3503 3504/* 3505 * vfs_bio_set_validclean: 3506 * 3507 * Set the range within the buffer to valid and clean. The range is 3508 * relative to the beginning of the buffer, b_offset. Note that b_offset 3509 * itself may be offset from the beginning of the first page. 3510 * 3511 */ 3512 3513void 3514vfs_bio_set_validclean(struct buf *bp, int base, int size) 3515{ 3516 int i, n; 3517 vm_page_t m; 3518 3519 if (!(bp->b_flags & B_VMIO)) 3520 return; 3521 /* 3522 * Fixup base to be relative to beginning of first page. 3523 * Set initial n to be the maximum number of bytes in the 3524 * first page that can be validated. 3525 */ 3526 3527 base += (bp->b_offset & PAGE_MASK); 3528 n = PAGE_SIZE - (base & PAGE_MASK); 3529 3530 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3531 vm_page_lock_queues(); 3532 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3533 m = bp->b_pages[i]; 3534 if (n > size) 3535 n = size; 3536 vm_page_set_validclean(m, base & PAGE_MASK, n); 3537 base += n; 3538 size -= n; 3539 n = PAGE_SIZE; 3540 } 3541 vm_page_unlock_queues(); 3542 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3543} 3544 3545/* 3546 * vfs_bio_clrbuf: 3547 * 3548 * clear a buffer. This routine essentially fakes an I/O, so we need 3549 * to clear BIO_ERROR and B_INVAL. 3550 * 3551 * Note that while we only theoretically need to clear through b_bcount, 3552 * we go ahead and clear through b_bufsize. 3553 */ 3554 3555void 3556vfs_bio_clrbuf(struct buf *bp) 3557{ 3558 int i, j, mask = 0; 3559 caddr_t sa, ea; 3560 3561 GIANT_REQUIRED; 3562 3563 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3564 clrbuf(bp); 3565 return; 3566 } 3567 3568 bp->b_flags &= ~B_INVAL; 3569 bp->b_ioflags &= ~BIO_ERROR; 3570 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3571 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3572 (bp->b_offset & PAGE_MASK) == 0) { 3573 if (bp->b_pages[0] == bogus_page) 3574 goto unlock; 3575 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3576 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3577 if ((bp->b_pages[0]->valid & mask) == mask) 3578 goto unlock; 3579 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3580 ((bp->b_pages[0]->valid & mask) == 0)) { 3581 bzero(bp->b_data, bp->b_bufsize); 3582 bp->b_pages[0]->valid |= mask; 3583 goto unlock; 3584 } 3585 } 3586 ea = sa = bp->b_data; 3587 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3588 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3589 ea = (caddr_t)(vm_offset_t)ulmin( 3590 (u_long)(vm_offset_t)ea, 3591 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3592 if (bp->b_pages[i] == bogus_page) 3593 continue; 3594 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3595 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3596 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3597 if ((bp->b_pages[i]->valid & mask) == mask) 3598 continue; 3599 if ((bp->b_pages[i]->valid & mask) == 0) { 3600 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3601 bzero(sa, ea - sa); 3602 } else { 3603 for (; sa < ea; sa += DEV_BSIZE, j++) { 3604 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3605 (bp->b_pages[i]->valid & (1 << j)) == 0) 3606 bzero(sa, DEV_BSIZE); 3607 } 3608 } 3609 bp->b_pages[i]->valid |= mask; 3610 } 3611unlock: 3612 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3613 bp->b_resid = 0; 3614} 3615 3616/* 3617 * vm_hold_load_pages and vm_hold_free_pages get pages into 3618 * a buffers address space. The pages are anonymous and are 3619 * not associated with a file object. 3620 */ 3621static void 3622vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3623{ 3624 vm_offset_t pg; 3625 vm_page_t p; 3626 int index; 3627 3628 to = round_page(to); 3629 from = round_page(from); 3630 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3631 3632 VM_OBJECT_LOCK(kernel_object); 3633 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3634tryagain: 3635 /* 3636 * note: must allocate system pages since blocking here 3637 * could intefere with paging I/O, no matter which 3638 * process we are. 3639 */ 3640 p = vm_page_alloc(kernel_object, 3641 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3642 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3643 if (!p) { 3644 atomic_add_int(&vm_pageout_deficit, 3645 (to - pg) >> PAGE_SHIFT); 3646 VM_OBJECT_UNLOCK(kernel_object); 3647 VM_WAIT; 3648 VM_OBJECT_LOCK(kernel_object); 3649 goto tryagain; 3650 } 3651 p->valid = VM_PAGE_BITS_ALL; 3652 pmap_qenter(pg, &p, 1); 3653 bp->b_pages[index] = p; 3654 } 3655 VM_OBJECT_UNLOCK(kernel_object); 3656 bp->b_npages = index; 3657} 3658 3659/* Return pages associated with this buf to the vm system */ 3660static void 3661vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3662{ 3663 vm_offset_t pg; 3664 vm_page_t p; 3665 int index, newnpages; 3666 3667 from = round_page(from); 3668 to = round_page(to); 3669 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3670 3671 VM_OBJECT_LOCK(kernel_object); 3672 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3673 p = bp->b_pages[index]; 3674 if (p && (index < bp->b_npages)) { 3675 if (p->busy) { 3676 printf( 3677 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3678 (intmax_t)bp->b_blkno, 3679 (intmax_t)bp->b_lblkno); 3680 } 3681 bp->b_pages[index] = NULL; 3682 pmap_qremove(pg, 1); 3683 vm_page_lock_queues(); 3684 vm_page_unwire(p, 0); 3685 vm_page_free(p); 3686 vm_page_unlock_queues(); 3687 } 3688 } 3689 VM_OBJECT_UNLOCK(kernel_object); 3690 bp->b_npages = newnpages; 3691} 3692 3693/* 3694 * Map an IO request into kernel virtual address space. 3695 * 3696 * All requests are (re)mapped into kernel VA space. 3697 * Notice that we use b_bufsize for the size of the buffer 3698 * to be mapped. b_bcount might be modified by the driver. 3699 * 3700 * Note that even if the caller determines that the address space should 3701 * be valid, a race or a smaller-file mapped into a larger space may 3702 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3703 * check the return value. 3704 */ 3705int 3706vmapbuf(struct buf *bp) 3707{ 3708 caddr_t addr, kva; 3709 vm_prot_t prot; 3710 int pidx, i; 3711 struct vm_page *m; 3712 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3713 3714 if (bp->b_bufsize < 0) 3715 return (-1); 3716 prot = VM_PROT_READ; 3717 if (bp->b_iocmd == BIO_READ) 3718 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3719 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3720 addr < bp->b_data + bp->b_bufsize; 3721 addr += PAGE_SIZE, pidx++) { 3722 /* 3723 * Do the vm_fault if needed; do the copy-on-write thing 3724 * when reading stuff off device into memory. 3725 * 3726 * NOTE! Must use pmap_extract() because addr may be in 3727 * the userland address space, and kextract is only guarenteed 3728 * to work for the kernland address space (see: sparc64 port). 3729 */ 3730retry: 3731 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3732 prot) < 0) { 3733 vm_page_lock_queues(); 3734 for (i = 0; i < pidx; ++i) { 3735 vm_page_unhold(bp->b_pages[i]); 3736 bp->b_pages[i] = NULL; 3737 } 3738 vm_page_unlock_queues(); 3739 return(-1); 3740 } 3741 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3742 if (m == NULL) 3743 goto retry; 3744 bp->b_pages[pidx] = m; 3745 } 3746 if (pidx > btoc(MAXPHYS)) 3747 panic("vmapbuf: mapped more than MAXPHYS"); 3748 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3749 3750 kva = bp->b_saveaddr; 3751 bp->b_npages = pidx; 3752 bp->b_saveaddr = bp->b_data; 3753 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3754 return(0); 3755} 3756 3757/* 3758 * Free the io map PTEs associated with this IO operation. 3759 * We also invalidate the TLB entries and restore the original b_addr. 3760 */ 3761void 3762vunmapbuf(struct buf *bp) 3763{ 3764 int pidx; 3765 int npages; 3766 3767 npages = bp->b_npages; 3768 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3769 vm_page_lock_queues(); 3770 for (pidx = 0; pidx < npages; pidx++) 3771 vm_page_unhold(bp->b_pages[pidx]); 3772 vm_page_unlock_queues(); 3773 3774 bp->b_data = bp->b_saveaddr; 3775} 3776 3777void 3778bdone(struct buf *bp) 3779{ 3780 3781 mtx_lock(&bdonelock); 3782 bp->b_flags |= B_DONE; 3783 wakeup(bp); 3784 mtx_unlock(&bdonelock); 3785} 3786 3787void 3788bwait(struct buf *bp, u_char pri, const char *wchan) 3789{ 3790 3791 mtx_lock(&bdonelock); 3792 while ((bp->b_flags & B_DONE) == 0) 3793 msleep(bp, &bdonelock, pri, wchan, 0); 3794 mtx_unlock(&bdonelock); 3795} 3796 3797void 3798bufstrategy(struct bufobj *bo, struct buf *bp) 3799{ 3800 int i = 0; 3801 struct vnode *vp; 3802 3803 vp = bp->b_vp; 3804 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3805 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3806 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3807 i = VOP_STRATEGY(vp, bp); 3808 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3809} 3810 3811void 3812bufobj_wref(struct bufobj *bo) 3813{ 3814 3815 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3816 BO_LOCK(bo); 3817 bo->bo_numoutput++; 3818 BO_UNLOCK(bo); 3819} 3820 3821void 3822bufobj_wdrop(struct bufobj *bo) 3823{ 3824 3825 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3826 BO_LOCK(bo); 3827 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3828 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3829 bo->bo_flag &= ~BO_WWAIT; 3830 wakeup(&bo->bo_numoutput); 3831 } 3832 BO_UNLOCK(bo); 3833} 3834 3835int 3836bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3837{ 3838 int error; 3839 3840 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3841 ASSERT_BO_LOCKED(bo); 3842 error = 0; 3843 while (bo->bo_numoutput) { 3844 bo->bo_flag |= BO_WWAIT; 3845 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3846 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3847 if (error) 3848 break; 3849 } 3850 return (error); 3851} 3852 3853#include "opt_ddb.h" 3854#ifdef DDB 3855#include <ddb/ddb.h> 3856 3857/* DDB command to show buffer data */ 3858DB_SHOW_COMMAND(buffer, db_show_buffer) 3859{ 3860 /* get args */ 3861 struct buf *bp = (struct buf *)addr; 3862 3863 if (!have_addr) { 3864 db_printf("usage: show buffer <addr>\n"); 3865 return; 3866 } 3867 3868 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3869 db_printf( 3870 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3871 "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n", 3872 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3873 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno); 3874 if (bp->b_npages) { 3875 int i; 3876 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3877 for (i = 0; i < bp->b_npages; i++) { 3878 vm_page_t m; 3879 m = bp->b_pages[i]; 3880 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3881 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3882 if ((i + 1) < bp->b_npages) 3883 db_printf(","); 3884 } 3885 db_printf("\n"); 3886 } 3887} 3888#endif /* DDB */ 3889