vfs_bio.c revision 137168
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 137168 2004-11-03 20:17:31Z alc $"); 43 44#include <sys/param.h> 45#include <sys/systm.h> 46#include <sys/bio.h> 47#include <sys/conf.h> 48#include <sys/buf.h> 49#include <sys/devicestat.h> 50#include <sys/eventhandler.h> 51#include <sys/lock.h> 52#include <sys/malloc.h> 53#include <sys/mount.h> 54#include <sys/mutex.h> 55#include <sys/kernel.h> 56#include <sys/kthread.h> 57#include <sys/proc.h> 58#include <sys/resourcevar.h> 59#include <sys/sysctl.h> 60#include <sys/vmmeter.h> 61#include <sys/vnode.h> 62#include <geom/geom.h> 63#include <vm/vm.h> 64#include <vm/vm_param.h> 65#include <vm/vm_kern.h> 66#include <vm/vm_pageout.h> 67#include <vm/vm_page.h> 68#include <vm/vm_object.h> 69#include <vm/vm_extern.h> 70#include <vm/vm_map.h> 71#include "opt_directio.h" 72#include "opt_swap.h" 73 74static MALLOC_DEFINE(M_BIOBUF, "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_dev = NULL; 544 bp->b_rcred = NOCRED; 545 bp->b_wcred = NOCRED; 546 bp->b_qindex = QUEUE_EMPTY; 547 bp->b_vflags = 0; 548 bp->b_xflags = 0; 549 LIST_INIT(&bp->b_dep); 550 BUF_LOCKINIT(bp); 551 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 552 } 553 554 /* 555 * maxbufspace is the absolute maximum amount of buffer space we are 556 * allowed to reserve in KVM and in real terms. The absolute maximum 557 * is nominally used by buf_daemon. hibufspace is the nominal maximum 558 * used by most other processes. The differential is required to 559 * ensure that buf_daemon is able to run when other processes might 560 * be blocked waiting for buffer space. 561 * 562 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 563 * this may result in KVM fragmentation which is not handled optimally 564 * by the system. 565 */ 566 maxbufspace = nbuf * BKVASIZE; 567 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 568 lobufspace = hibufspace - MAXBSIZE; 569 570 lorunningspace = 512 * 1024; 571 hirunningspace = 1024 * 1024; 572 573/* 574 * Limit the amount of malloc memory since it is wired permanently into 575 * the kernel space. Even though this is accounted for in the buffer 576 * allocation, we don't want the malloced region to grow uncontrolled. 577 * The malloc scheme improves memory utilization significantly on average 578 * (small) directories. 579 */ 580 maxbufmallocspace = hibufspace / 20; 581 582/* 583 * Reduce the chance of a deadlock occuring by limiting the number 584 * of delayed-write dirty buffers we allow to stack up. 585 */ 586 hidirtybuffers = nbuf / 4 + 20; 587 dirtybufthresh = hidirtybuffers * 9 / 10; 588 numdirtybuffers = 0; 589/* 590 * To support extreme low-memory systems, make sure hidirtybuffers cannot 591 * eat up all available buffer space. This occurs when our minimum cannot 592 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 593 * BKVASIZE'd (8K) buffers. 594 */ 595 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 596 hidirtybuffers >>= 1; 597 } 598 lodirtybuffers = hidirtybuffers / 2; 599 600/* 601 * Try to keep the number of free buffers in the specified range, 602 * and give special processes (e.g. like buf_daemon) access to an 603 * emergency reserve. 604 */ 605 lofreebuffers = nbuf / 18 + 5; 606 hifreebuffers = 2 * lofreebuffers; 607 numfreebuffers = nbuf; 608 609/* 610 * Maximum number of async ops initiated per buf_daemon loop. This is 611 * somewhat of a hack at the moment, we really need to limit ourselves 612 * based on the number of bytes of I/O in-transit that were initiated 613 * from buf_daemon. 614 */ 615 616 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 617 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 618} 619 620/* 621 * bfreekva() - free the kva allocation for a buffer. 622 * 623 * Must be called at splbio() or higher as this is the only locking for 624 * buffer_map. 625 * 626 * Since this call frees up buffer space, we call bufspacewakeup(). 627 */ 628static void 629bfreekva(struct buf *bp) 630{ 631 632 GIANT_REQUIRED; 633 634 if (bp->b_kvasize) { 635 atomic_add_int(&buffreekvacnt, 1); 636 atomic_subtract_int(&bufspace, bp->b_kvasize); 637 vm_map_delete(buffer_map, 638 (vm_offset_t) bp->b_kvabase, 639 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 640 ); 641 bp->b_kvasize = 0; 642 bufspacewakeup(); 643 } 644} 645 646/* 647 * bremfree: 648 * 649 * Remove the buffer from the appropriate free list. 650 */ 651void 652bremfree(struct buf *bp) 653{ 654 655 mtx_lock(&bqlock); 656 bremfreel(bp); 657 mtx_unlock(&bqlock); 658} 659 660void 661bremfreel(struct buf *bp) 662{ 663 int s = splbio(); 664 int old_qindex = bp->b_qindex; 665 666 GIANT_REQUIRED; 667 668 if (bp->b_qindex != QUEUE_NONE) { 669 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 670 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 671 bp->b_qindex = QUEUE_NONE; 672 } else { 673 if (BUF_REFCNT(bp) <= 1) 674 panic("bremfree: removing a buffer not on a queue"); 675 } 676 677 /* 678 * Fixup numfreebuffers count. If the buffer is invalid or not 679 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 680 * the buffer was free and we must decrement numfreebuffers. 681 */ 682 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 683 switch(old_qindex) { 684 case QUEUE_DIRTY: 685 case QUEUE_CLEAN: 686 case QUEUE_EMPTY: 687 case QUEUE_EMPTYKVA: 688 atomic_subtract_int(&numfreebuffers, 1); 689 break; 690 default: 691 break; 692 } 693 } 694 splx(s); 695} 696 697 698/* 699 * Get a buffer with the specified data. Look in the cache first. We 700 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 701 * is set, the buffer is valid and we do not have to do anything ( see 702 * getblk() ). This is really just a special case of breadn(). 703 */ 704int 705bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 706 struct buf **bpp) 707{ 708 709 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 710} 711 712/* 713 * Operates like bread, but also starts asynchronous I/O on 714 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 715 * to initiating I/O . If B_CACHE is set, the buffer is valid 716 * and we do not have to do anything. 717 */ 718int 719breadn(struct vnode * vp, daddr_t blkno, int size, 720 daddr_t * rablkno, int *rabsize, 721 int cnt, struct ucred * cred, struct buf **bpp) 722{ 723 struct buf *bp, *rabp; 724 int i; 725 int rv = 0, readwait = 0; 726 727 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 728 729 /* if not found in cache, do some I/O */ 730 if ((bp->b_flags & B_CACHE) == 0) { 731 if (curthread != PCPU_GET(idlethread)) 732 curthread->td_proc->p_stats->p_ru.ru_inblock++; 733 bp->b_iocmd = BIO_READ; 734 bp->b_flags &= ~B_INVAL; 735 bp->b_ioflags &= ~BIO_ERROR; 736 if (bp->b_rcred == NOCRED && cred != NOCRED) 737 bp->b_rcred = crhold(cred); 738 vfs_busy_pages(bp, 0); 739 bp->b_iooffset = dbtob(bp->b_blkno); 740 bstrategy(bp); 741 ++readwait; 742 } 743 744 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 745 if (inmem(vp, *rablkno)) 746 continue; 747 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 748 749 if ((rabp->b_flags & B_CACHE) == 0) { 750 if (curthread != PCPU_GET(idlethread)) 751 curthread->td_proc->p_stats->p_ru.ru_inblock++; 752 rabp->b_flags |= B_ASYNC; 753 rabp->b_flags &= ~B_INVAL; 754 rabp->b_ioflags &= ~BIO_ERROR; 755 rabp->b_iocmd = BIO_READ; 756 if (rabp->b_rcred == NOCRED && cred != NOCRED) 757 rabp->b_rcred = crhold(cred); 758 vfs_busy_pages(rabp, 0); 759 BUF_KERNPROC(rabp); 760 rabp->b_iooffset = dbtob(rabp->b_blkno); 761 bstrategy(rabp); 762 } else { 763 brelse(rabp); 764 } 765 } 766 767 if (readwait) { 768 rv = bufwait(bp); 769 } 770 return (rv); 771} 772 773/* 774 * Write, release buffer on completion. (Done by iodone 775 * if async). Do not bother writing anything if the buffer 776 * is invalid. 777 * 778 * Note that we set B_CACHE here, indicating that buffer is 779 * fully valid and thus cacheable. This is true even of NFS 780 * now so we set it generally. This could be set either here 781 * or in biodone() since the I/O is synchronous. We put it 782 * here. 783 */ 784int 785bufwrite(struct buf *bp) 786{ 787 int oldflags, s; 788 struct buf *newbp; 789 790 if (bp->b_flags & B_INVAL) { 791 brelse(bp); 792 return (0); 793 } 794 795 oldflags = bp->b_flags; 796 797 if (BUF_REFCNT(bp) == 0) 798 panic("bufwrite: buffer is not busy???"); 799 s = splbio(); 800 /* 801 * If a background write is already in progress, delay 802 * writing this block if it is asynchronous. Otherwise 803 * wait for the background write to complete. 804 */ 805 BO_LOCK(bp->b_bufobj); 806 if (bp->b_vflags & BV_BKGRDINPROG) { 807 if (bp->b_flags & B_ASYNC) { 808 BO_UNLOCK(bp->b_bufobj); 809 splx(s); 810 bdwrite(bp); 811 return (0); 812 } 813 bp->b_vflags |= BV_BKGRDWAIT; 814 msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj), PRIBIO, "bwrbg", 0); 815 if (bp->b_vflags & BV_BKGRDINPROG) 816 panic("bufwrite: still writing"); 817 } 818 BO_UNLOCK(bp->b_bufobj); 819 820 /* Mark the buffer clean */ 821 bundirty(bp); 822 823 /* 824 * If this buffer is marked for background writing and we 825 * do not have to wait for it, make a copy and write the 826 * copy so as to leave this buffer ready for further use. 827 * 828 * This optimization eats a lot of memory. If we have a page 829 * or buffer shortfall we can't do it. 830 */ 831 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 832 (bp->b_flags & B_ASYNC) && 833 !vm_page_count_severe() && 834 !buf_dirty_count_severe()) { 835 KASSERT(bp->b_iodone == NULL, 836 ("bufwrite: needs chained iodone (%p)", bp->b_iodone)); 837 838 /* get a new block */ 839 newbp = geteblk(bp->b_bufsize); 840 841 /* 842 * set it to be identical to the old block. We have to 843 * set b_lblkno and BKGRDMARKER before calling bgetvp() 844 * to avoid confusing the splay tree and gbincore(). 845 */ 846 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 847 newbp->b_lblkno = bp->b_lblkno; 848 newbp->b_xflags |= BX_BKGRDMARKER; 849 BO_LOCK(bp->b_bufobj); 850 bp->b_vflags |= BV_BKGRDINPROG; 851 bgetvp(bp->b_vp, newbp); 852 BO_UNLOCK(bp->b_bufobj); 853 newbp->b_bufobj = &bp->b_vp->v_bufobj; 854 newbp->b_blkno = bp->b_blkno; 855 newbp->b_offset = bp->b_offset; 856 newbp->b_iodone = vfs_backgroundwritedone; 857 newbp->b_flags |= B_ASYNC; 858 newbp->b_flags &= ~B_INVAL; 859 860 /* move over the dependencies */ 861 if (LIST_FIRST(&bp->b_dep) != NULL) 862 buf_movedeps(bp, newbp); 863 864 /* 865 * Initiate write on the copy, release the original to 866 * the B_LOCKED queue so that it cannot go away until 867 * the background write completes. If not locked it could go 868 * away and then be reconstituted while it was being written. 869 * If the reconstituted buffer were written, we could end up 870 * with two background copies being written at the same time. 871 */ 872 bqrelse(bp); 873 bp = newbp; 874 } 875 876 bp->b_flags &= ~B_DONE; 877 bp->b_ioflags &= ~BIO_ERROR; 878 bp->b_flags |= B_CACHE; 879 bp->b_iocmd = BIO_WRITE; 880 881 bufobj_wref(bp->b_bufobj); 882 vfs_busy_pages(bp, 1); 883 884 /* 885 * Normal bwrites pipeline writes 886 */ 887 bp->b_runningbufspace = bp->b_bufsize; 888 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 889 890 if (curthread != PCPU_GET(idlethread)) 891 curthread->td_proc->p_stats->p_ru.ru_oublock++; 892 splx(s); 893 if (oldflags & B_ASYNC) 894 BUF_KERNPROC(bp); 895 bp->b_iooffset = dbtob(bp->b_blkno); 896 bstrategy(bp); 897 898 if ((oldflags & B_ASYNC) == 0) { 899 int rtval = bufwait(bp); 900 brelse(bp); 901 return (rtval); 902 } else { 903 /* 904 * don't allow the async write to saturate the I/O 905 * system. We will not deadlock here because 906 * we are blocking waiting for I/O that is already in-progress 907 * to complete. We do not block here if it is the update 908 * or syncer daemon trying to clean up as that can lead 909 * to deadlock. 910 */ 911 if (curthread->td_proc != bufdaemonproc && 912 curthread->td_proc != updateproc) 913 waitrunningbufspace(); 914 } 915 916 return (0); 917} 918 919/* 920 * Complete a background write started from bwrite. 921 */ 922static void 923vfs_backgroundwritedone(struct buf *bp) 924{ 925 struct buf *origbp; 926 927 /* 928 * Find the original buffer that we are writing. 929 */ 930 BO_LOCK(bp->b_bufobj); 931 if ((origbp = gbincore(bp->b_bufobj, bp->b_lblkno)) == NULL) 932 panic("backgroundwritedone: lost buffer"); 933 934 /* 935 * Clear the BV_BKGRDINPROG flag in the original buffer 936 * and awaken it if it is waiting for the write to complete. 937 * If BV_BKGRDINPROG is not set in the original buffer it must 938 * have been released and re-instantiated - which is not legal. 939 */ 940 KASSERT((origbp->b_vflags & BV_BKGRDINPROG), 941 ("backgroundwritedone: lost buffer2")); 942 origbp->b_vflags &= ~BV_BKGRDINPROG; 943 if (origbp->b_vflags & BV_BKGRDWAIT) { 944 origbp->b_vflags &= ~BV_BKGRDWAIT; 945 wakeup(&origbp->b_xflags); 946 } 947 BO_UNLOCK(bp->b_bufobj); 948 /* 949 * Process dependencies then return any unfinished ones. 950 */ 951 if (LIST_FIRST(&bp->b_dep) != NULL) 952 buf_complete(bp); 953 if (LIST_FIRST(&bp->b_dep) != NULL) 954 buf_movedeps(bp, origbp); 955 956 /* 957 * This buffer is marked B_NOCACHE, so when it is released 958 * by biodone, it will be tossed. We mark it with BIO_READ 959 * to avoid biodone doing a second bufobj_wdrop. 960 */ 961 bp->b_flags |= B_NOCACHE; 962 bp->b_iocmd = BIO_READ; 963 bp->b_flags &= ~(B_CACHE | B_DONE); 964 bp->b_iodone = 0; 965 bufdone(bp); 966} 967 968/* 969 * Delayed write. (Buffer is marked dirty). Do not bother writing 970 * anything if the buffer is marked invalid. 971 * 972 * Note that since the buffer must be completely valid, we can safely 973 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 974 * biodone() in order to prevent getblk from writing the buffer 975 * out synchronously. 976 */ 977void 978bdwrite(struct buf *bp) 979{ 980 struct thread *td = curthread; 981 struct vnode *vp; 982 struct buf *nbp; 983 struct bufobj *bo; 984 985 GIANT_REQUIRED; 986 987 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 988 KASSERT(BUF_REFCNT(bp) != 0, ("bdwrite: buffer is not busy")); 989 990 if (bp->b_flags & B_INVAL) { 991 brelse(bp); 992 return; 993 } 994 995 /* 996 * If we have too many dirty buffers, don't create any more. 997 * If we are wildly over our limit, then force a complete 998 * cleanup. Otherwise, just keep the situation from getting 999 * out of control. Note that we have to avoid a recursive 1000 * disaster and not try to clean up after our own cleanup! 1001 */ 1002 vp = bp->b_vp; 1003 bo = bp->b_bufobj; 1004 BO_LOCK(bo); 1005 if (td->td_pflags & TDP_COWINPROGRESS) { 1006 recursiveflushes++; 1007 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 1008 BO_UNLOCK(bo); 1009 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 1010 BO_LOCK(bo); 1011 altbufferflushes++; 1012 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 1013 /* 1014 * Try to find a buffer to flush. 1015 */ 1016 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 1017 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1018 buf_countdeps(nbp, 0) || 1019 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1020 continue; 1021 if (bp == nbp) 1022 panic("bdwrite: found ourselves"); 1023 BO_UNLOCK(bo); 1024 if (nbp->b_flags & B_CLUSTEROK) { 1025 vfs_bio_awrite(nbp); 1026 } else { 1027 bremfree(nbp); 1028 bawrite(nbp); 1029 } 1030 BO_LOCK(bo); 1031 dirtybufferflushes++; 1032 break; 1033 } 1034 } 1035 BO_UNLOCK(bo); 1036 1037 bdirty(bp); 1038 /* 1039 * Set B_CACHE, indicating that the buffer is fully valid. This is 1040 * true even of NFS now. 1041 */ 1042 bp->b_flags |= B_CACHE; 1043 1044 /* 1045 * This bmap keeps the system from needing to do the bmap later, 1046 * perhaps when the system is attempting to do a sync. Since it 1047 * is likely that the indirect block -- or whatever other datastructure 1048 * that the filesystem needs is still in memory now, it is a good 1049 * thing to do this. Note also, that if the pageout daemon is 1050 * requesting a sync -- there might not be enough memory to do 1051 * the bmap then... So, this is important to do. 1052 */ 1053 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1054 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1055 } 1056 1057 /* 1058 * Set the *dirty* buffer range based upon the VM system dirty pages. 1059 */ 1060 vfs_setdirty(bp); 1061 1062 /* 1063 * We need to do this here to satisfy the vnode_pager and the 1064 * pageout daemon, so that it thinks that the pages have been 1065 * "cleaned". Note that since the pages are in a delayed write 1066 * buffer -- the VFS layer "will" see that the pages get written 1067 * out on the next sync, or perhaps the cluster will be completed. 1068 */ 1069 vfs_clean_pages(bp); 1070 bqrelse(bp); 1071 1072 /* 1073 * Wakeup the buffer flushing daemon if we have a lot of dirty 1074 * buffers (midpoint between our recovery point and our stall 1075 * point). 1076 */ 1077 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1078 1079 /* 1080 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1081 * due to the softdep code. 1082 */ 1083} 1084 1085/* 1086 * bdirty: 1087 * 1088 * Turn buffer into delayed write request. We must clear BIO_READ and 1089 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1090 * itself to properly update it in the dirty/clean lists. We mark it 1091 * B_DONE to ensure that any asynchronization of the buffer properly 1092 * clears B_DONE ( else a panic will occur later ). 1093 * 1094 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1095 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1096 * should only be called if the buffer is known-good. 1097 * 1098 * Since the buffer is not on a queue, we do not update the numfreebuffers 1099 * count. 1100 * 1101 * Must be called at splbio(). 1102 * The buffer must be on QUEUE_NONE. 1103 */ 1104void 1105bdirty(struct buf *bp) 1106{ 1107 1108 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1109 KASSERT(bp->b_qindex == QUEUE_NONE, 1110 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1111 bp->b_flags &= ~(B_RELBUF); 1112 bp->b_iocmd = BIO_WRITE; 1113 1114 if ((bp->b_flags & B_DELWRI) == 0) { 1115 bp->b_flags |= B_DONE | B_DELWRI; 1116 reassignbuf(bp); 1117 atomic_add_int(&numdirtybuffers, 1); 1118 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1119 } 1120} 1121 1122/* 1123 * bundirty: 1124 * 1125 * Clear B_DELWRI for buffer. 1126 * 1127 * Since the buffer is not on a queue, we do not update the numfreebuffers 1128 * count. 1129 * 1130 * Must be called at splbio(). 1131 * The buffer must be on QUEUE_NONE. 1132 */ 1133 1134void 1135bundirty(struct buf *bp) 1136{ 1137 1138 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1139 KASSERT(bp->b_qindex == QUEUE_NONE, 1140 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1141 1142 if (bp->b_flags & B_DELWRI) { 1143 bp->b_flags &= ~B_DELWRI; 1144 reassignbuf(bp); 1145 atomic_subtract_int(&numdirtybuffers, 1); 1146 numdirtywakeup(lodirtybuffers); 1147 } 1148 /* 1149 * Since it is now being written, we can clear its deferred write flag. 1150 */ 1151 bp->b_flags &= ~B_DEFERRED; 1152} 1153 1154/* 1155 * bawrite: 1156 * 1157 * Asynchronous write. Start output on a buffer, but do not wait for 1158 * it to complete. The buffer is released when the output completes. 1159 * 1160 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1161 * B_INVAL buffers. Not us. 1162 */ 1163void 1164bawrite(struct buf *bp) 1165{ 1166 1167 bp->b_flags |= B_ASYNC; 1168 (void) bwrite(bp); 1169} 1170 1171/* 1172 * bwillwrite: 1173 * 1174 * Called prior to the locking of any vnodes when we are expecting to 1175 * write. We do not want to starve the buffer cache with too many 1176 * dirty buffers so we block here. By blocking prior to the locking 1177 * of any vnodes we attempt to avoid the situation where a locked vnode 1178 * prevents the various system daemons from flushing related buffers. 1179 */ 1180 1181void 1182bwillwrite(void) 1183{ 1184 1185 if (numdirtybuffers >= hidirtybuffers) { 1186 int s; 1187 1188 mtx_lock(&Giant); 1189 s = splbio(); 1190 mtx_lock(&nblock); 1191 while (numdirtybuffers >= hidirtybuffers) { 1192 bd_wakeup(1); 1193 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1194 msleep(&needsbuffer, &nblock, 1195 (PRIBIO + 4), "flswai", 0); 1196 } 1197 splx(s); 1198 mtx_unlock(&nblock); 1199 mtx_unlock(&Giant); 1200 } 1201} 1202 1203/* 1204 * Return true if we have too many dirty buffers. 1205 */ 1206int 1207buf_dirty_count_severe(void) 1208{ 1209 1210 return(numdirtybuffers >= hidirtybuffers); 1211} 1212 1213/* 1214 * brelse: 1215 * 1216 * Release a busy buffer and, if requested, free its resources. The 1217 * buffer will be stashed in the appropriate bufqueue[] allowing it 1218 * to be accessed later as a cache entity or reused for other purposes. 1219 */ 1220void 1221brelse(struct buf *bp) 1222{ 1223 int s; 1224 1225 GIANT_REQUIRED; 1226 1227 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1228 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1229 1230 s = splbio(); 1231 1232 if (bp->b_iocmd == BIO_WRITE && 1233 (bp->b_ioflags & BIO_ERROR) && 1234 !(bp->b_flags & B_INVAL)) { 1235 /* 1236 * Failed write, redirty. Must clear BIO_ERROR to prevent 1237 * pages from being scrapped. If B_INVAL is set then 1238 * this case is not run and the next case is run to 1239 * destroy the buffer. B_INVAL can occur if the buffer 1240 * is outside the range supported by the underlying device. 1241 */ 1242 bp->b_ioflags &= ~BIO_ERROR; 1243 bdirty(bp); 1244 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1245 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1246 /* 1247 * Either a failed I/O or we were asked to free or not 1248 * cache the buffer. 1249 */ 1250 bp->b_flags |= B_INVAL; 1251 if (LIST_FIRST(&bp->b_dep) != NULL) 1252 buf_deallocate(bp); 1253 if (bp->b_flags & B_DELWRI) { 1254 atomic_subtract_int(&numdirtybuffers, 1); 1255 numdirtywakeup(lodirtybuffers); 1256 } 1257 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1258 if ((bp->b_flags & B_VMIO) == 0) { 1259 if (bp->b_bufsize) 1260 allocbuf(bp, 0); 1261 if (bp->b_vp) 1262 brelvp(bp); 1263 } 1264 } 1265 1266 /* 1267 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1268 * is called with B_DELWRI set, the underlying pages may wind up 1269 * getting freed causing a previous write (bdwrite()) to get 'lost' 1270 * because pages associated with a B_DELWRI bp are marked clean. 1271 * 1272 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1273 * if B_DELWRI is set. 1274 * 1275 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1276 * on pages to return pages to the VM page queues. 1277 */ 1278 if (bp->b_flags & B_DELWRI) 1279 bp->b_flags &= ~B_RELBUF; 1280 else if (vm_page_count_severe()) { 1281 /* 1282 * XXX This lock may not be necessary since BKGRDINPROG 1283 * cannot be set while we hold the buf lock, it can only be 1284 * cleared if it is already pending. 1285 */ 1286 if (bp->b_vp) { 1287 BO_LOCK(bp->b_bufobj); 1288 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1289 bp->b_flags |= B_RELBUF; 1290 BO_UNLOCK(bp->b_bufobj); 1291 } else 1292 bp->b_flags |= B_RELBUF; 1293 } 1294 1295 /* 1296 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1297 * constituted, not even NFS buffers now. Two flags effect this. If 1298 * B_INVAL, the struct buf is invalidated but the VM object is kept 1299 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1300 * 1301 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1302 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1303 * buffer is also B_INVAL because it hits the re-dirtying code above. 1304 * 1305 * Normally we can do this whether a buffer is B_DELWRI or not. If 1306 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1307 * the commit state and we cannot afford to lose the buffer. If the 1308 * buffer has a background write in progress, we need to keep it 1309 * around to prevent it from being reconstituted and starting a second 1310 * background write. 1311 */ 1312 if ((bp->b_flags & B_VMIO) 1313 && !(bp->b_vp->v_mount != NULL && 1314 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1315 !vn_isdisk(bp->b_vp, NULL) && 1316 (bp->b_flags & B_DELWRI)) 1317 ) { 1318 1319 int i, j, resid; 1320 vm_page_t m; 1321 off_t foff; 1322 vm_pindex_t poff; 1323 vm_object_t obj; 1324 1325 obj = bp->b_object; 1326 1327 /* 1328 * Get the base offset and length of the buffer. Note that 1329 * in the VMIO case if the buffer block size is not 1330 * page-aligned then b_data pointer may not be page-aligned. 1331 * But our b_pages[] array *IS* page aligned. 1332 * 1333 * block sizes less then DEV_BSIZE (usually 512) are not 1334 * supported due to the page granularity bits (m->valid, 1335 * m->dirty, etc...). 1336 * 1337 * See man buf(9) for more information 1338 */ 1339 resid = bp->b_bufsize; 1340 foff = bp->b_offset; 1341 VM_OBJECT_LOCK(obj); 1342 for (i = 0; i < bp->b_npages; i++) { 1343 int had_bogus = 0; 1344 1345 m = bp->b_pages[i]; 1346 1347 /* 1348 * If we hit a bogus page, fixup *all* the bogus pages 1349 * now. 1350 */ 1351 if (m == bogus_page) { 1352 poff = OFF_TO_IDX(bp->b_offset); 1353 had_bogus = 1; 1354 1355 for (j = i; j < bp->b_npages; j++) { 1356 vm_page_t mtmp; 1357 mtmp = bp->b_pages[j]; 1358 if (mtmp == bogus_page) { 1359 mtmp = vm_page_lookup(obj, poff + j); 1360 if (!mtmp) { 1361 panic("brelse: page missing\n"); 1362 } 1363 bp->b_pages[j] = mtmp; 1364 } 1365 } 1366 1367 if ((bp->b_flags & B_INVAL) == 0) { 1368 pmap_qenter( 1369 trunc_page((vm_offset_t)bp->b_data), 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 bp->b_dev = NULL; 1426 /* buffers with junk contents */ 1427 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1428 (bp->b_ioflags & BIO_ERROR)) { 1429 bp->b_flags |= B_INVAL; 1430 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1431 if (bp->b_vflags & BV_BKGRDINPROG) 1432 panic("losing buffer 2"); 1433 bp->b_qindex = QUEUE_CLEAN; 1434 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1435 bp->b_dev = NULL; 1436 /* remaining buffers */ 1437 } else { 1438 if (bp->b_flags & B_DELWRI) 1439 bp->b_qindex = QUEUE_DIRTY; 1440 else 1441 bp->b_qindex = QUEUE_CLEAN; 1442 if (bp->b_flags & B_AGE) 1443 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1444 else 1445 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1446 } 1447 mtx_unlock(&bqlock); 1448 1449 /* 1450 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1451 * placed the buffer on the correct queue. We must also disassociate 1452 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1453 * find it. 1454 */ 1455 if (bp->b_flags & B_INVAL) { 1456 if (bp->b_flags & B_DELWRI) 1457 bundirty(bp); 1458 if (bp->b_vp) 1459 brelvp(bp); 1460 } 1461 1462 /* 1463 * Fixup numfreebuffers count. The bp is on an appropriate queue 1464 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1465 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1466 * if B_INVAL is set ). 1467 */ 1468 1469 if (!(bp->b_flags & B_DELWRI)) 1470 bufcountwakeup(); 1471 1472 /* 1473 * Something we can maybe free or reuse 1474 */ 1475 if (bp->b_bufsize || bp->b_kvasize) 1476 bufspacewakeup(); 1477 1478 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1479 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1480 panic("brelse: not dirty"); 1481 /* unlock */ 1482 BUF_UNLOCK(bp); 1483 splx(s); 1484} 1485 1486/* 1487 * Release a buffer back to the appropriate queue but do not try to free 1488 * it. The buffer is expected to be used again soon. 1489 * 1490 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1491 * biodone() to requeue an async I/O on completion. It is also used when 1492 * known good buffers need to be requeued but we think we may need the data 1493 * again soon. 1494 * 1495 * XXX we should be able to leave the B_RELBUF hint set on completion. 1496 */ 1497void 1498bqrelse(struct buf *bp) 1499{ 1500 int s; 1501 1502 s = splbio(); 1503 1504 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1505 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1506 1507 if (bp->b_qindex != QUEUE_NONE) 1508 panic("bqrelse: free buffer onto another queue???"); 1509 if (BUF_REFCNT(bp) > 1) { 1510 /* do not release to free list */ 1511 BUF_UNLOCK(bp); 1512 splx(s); 1513 return; 1514 } 1515 mtx_lock(&bqlock); 1516 /* buffers with stale but valid contents */ 1517 if (bp->b_flags & B_DELWRI) { 1518 bp->b_qindex = QUEUE_DIRTY; 1519 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1520 } else { 1521 /* 1522 * XXX This lock may not be necessary since BKGRDINPROG 1523 * cannot be set while we hold the buf lock, it can only be 1524 * cleared if it is already pending. 1525 */ 1526 BO_LOCK(bp->b_bufobj); 1527 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1528 BO_UNLOCK(bp->b_bufobj); 1529 bp->b_qindex = QUEUE_CLEAN; 1530 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1531 b_freelist); 1532 } else { 1533 /* 1534 * We are too low on memory, we have to try to free 1535 * the buffer (most importantly: the wired pages 1536 * making up its backing store) *now*. 1537 */ 1538 BO_UNLOCK(bp->b_bufobj); 1539 mtx_unlock(&bqlock); 1540 splx(s); 1541 brelse(bp); 1542 return; 1543 } 1544 } 1545 mtx_unlock(&bqlock); 1546 1547 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1548 bufcountwakeup(); 1549 1550 /* 1551 * Something we can maybe free or reuse. 1552 */ 1553 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1554 bufspacewakeup(); 1555 1556 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1557 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1558 panic("bqrelse: not dirty"); 1559 /* unlock */ 1560 BUF_UNLOCK(bp); 1561 splx(s); 1562} 1563 1564/* Give pages used by the bp back to the VM system (where possible) */ 1565static void 1566vfs_vmio_release(struct buf *bp) 1567{ 1568 int i; 1569 vm_page_t m; 1570 1571 GIANT_REQUIRED; 1572 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1573 vm_page_lock_queues(); 1574 for (i = 0; i < bp->b_npages; i++) { 1575 m = bp->b_pages[i]; 1576 bp->b_pages[i] = NULL; 1577 /* 1578 * In order to keep page LRU ordering consistent, put 1579 * everything on the inactive queue. 1580 */ 1581 vm_page_unwire(m, 0); 1582 /* 1583 * We don't mess with busy pages, it is 1584 * the responsibility of the process that 1585 * busied the pages to deal with them. 1586 */ 1587 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1588 continue; 1589 1590 if (m->wire_count == 0) { 1591 /* 1592 * Might as well free the page if we can and it has 1593 * no valid data. We also free the page if the 1594 * buffer was used for direct I/O 1595 */ 1596 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1597 m->hold_count == 0) { 1598 pmap_remove_all(m); 1599 vm_page_free(m); 1600 } else if (bp->b_flags & B_DIRECT) { 1601 vm_page_try_to_free(m); 1602 } else if (vm_page_count_severe()) { 1603 vm_page_try_to_cache(m); 1604 } 1605 } 1606 } 1607 vm_page_unlock_queues(); 1608 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1609 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1610 1611 if (bp->b_bufsize) { 1612 bufspacewakeup(); 1613 bp->b_bufsize = 0; 1614 } 1615 bp->b_npages = 0; 1616 bp->b_flags &= ~B_VMIO; 1617 if (bp->b_vp) 1618 brelvp(bp); 1619} 1620 1621/* 1622 * Check to see if a block at a particular lbn is available for a clustered 1623 * write. 1624 */ 1625static int 1626vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1627{ 1628 struct buf *bpa; 1629 int match; 1630 1631 match = 0; 1632 1633 /* If the buf isn't in core skip it */ 1634 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1635 return (0); 1636 1637 /* If the buf is busy we don't want to wait for it */ 1638 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1639 return (0); 1640 1641 /* Only cluster with valid clusterable delayed write buffers */ 1642 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1643 (B_DELWRI | B_CLUSTEROK)) 1644 goto done; 1645 1646 if (bpa->b_bufsize != size) 1647 goto done; 1648 1649 /* 1650 * Check to see if it is in the expected place on disk and that the 1651 * block has been mapped. 1652 */ 1653 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1654 match = 1; 1655done: 1656 BUF_UNLOCK(bpa); 1657 return (match); 1658} 1659 1660/* 1661 * vfs_bio_awrite: 1662 * 1663 * Implement clustered async writes for clearing out B_DELWRI buffers. 1664 * This is much better then the old way of writing only one buffer at 1665 * a time. Note that we may not be presented with the buffers in the 1666 * correct order, so we search for the cluster in both directions. 1667 */ 1668int 1669vfs_bio_awrite(struct buf *bp) 1670{ 1671 int i; 1672 int j; 1673 daddr_t lblkno = bp->b_lblkno; 1674 struct vnode *vp = bp->b_vp; 1675 int s; 1676 int ncl; 1677 int nwritten; 1678 int size; 1679 int maxcl; 1680 1681 s = splbio(); 1682 /* 1683 * right now we support clustered writing only to regular files. If 1684 * we find a clusterable block we could be in the middle of a cluster 1685 * rather then at the beginning. 1686 */ 1687 if ((vp->v_type == VREG) && 1688 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1689 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1690 1691 size = vp->v_mount->mnt_stat.f_iosize; 1692 maxcl = MAXPHYS / size; 1693 1694 VI_LOCK(vp); 1695 for (i = 1; i < maxcl; i++) 1696 if (vfs_bio_clcheck(vp, size, lblkno + i, 1697 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1698 break; 1699 1700 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1701 if (vfs_bio_clcheck(vp, size, lblkno - j, 1702 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1703 break; 1704 1705 VI_UNLOCK(vp); 1706 --j; 1707 ncl = i + j; 1708 /* 1709 * this is a possible cluster write 1710 */ 1711 if (ncl != 1) { 1712 BUF_UNLOCK(bp); 1713 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1714 splx(s); 1715 return nwritten; 1716 } 1717 } 1718 1719 bremfree(bp); 1720 bp->b_flags |= B_ASYNC; 1721 1722 splx(s); 1723 /* 1724 * default (old) behavior, writing out only one block 1725 * 1726 * XXX returns b_bufsize instead of b_bcount for nwritten? 1727 */ 1728 nwritten = bp->b_bufsize; 1729 (void) bwrite(bp); 1730 1731 return nwritten; 1732} 1733 1734/* 1735 * getnewbuf: 1736 * 1737 * Find and initialize a new buffer header, freeing up existing buffers 1738 * in the bufqueues as necessary. The new buffer is returned locked. 1739 * 1740 * Important: B_INVAL is not set. If the caller wishes to throw the 1741 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1742 * 1743 * We block if: 1744 * We have insufficient buffer headers 1745 * We have insufficient buffer space 1746 * buffer_map is too fragmented ( space reservation fails ) 1747 * If we have to flush dirty buffers ( but we try to avoid this ) 1748 * 1749 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1750 * Instead we ask the buf daemon to do it for us. We attempt to 1751 * avoid piecemeal wakeups of the pageout daemon. 1752 */ 1753 1754static struct buf * 1755getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1756{ 1757 struct buf *bp; 1758 struct buf *nbp; 1759 int defrag = 0; 1760 int nqindex; 1761 static int flushingbufs; 1762 1763 GIANT_REQUIRED; 1764 1765 /* 1766 * We can't afford to block since we might be holding a vnode lock, 1767 * which may prevent system daemons from running. We deal with 1768 * low-memory situations by proactively returning memory and running 1769 * async I/O rather then sync I/O. 1770 */ 1771 1772 atomic_add_int(&getnewbufcalls, 1); 1773 atomic_subtract_int(&getnewbufrestarts, 1); 1774restart: 1775 atomic_add_int(&getnewbufrestarts, 1); 1776 1777 /* 1778 * Setup for scan. If we do not have enough free buffers, 1779 * we setup a degenerate case that immediately fails. Note 1780 * that if we are specially marked process, we are allowed to 1781 * dip into our reserves. 1782 * 1783 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1784 * 1785 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1786 * However, there are a number of cases (defragging, reusing, ...) 1787 * where we cannot backup. 1788 */ 1789 mtx_lock(&bqlock); 1790 nqindex = QUEUE_EMPTYKVA; 1791 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1792 1793 if (nbp == NULL) { 1794 /* 1795 * If no EMPTYKVA buffers and we are either 1796 * defragging or reusing, locate a CLEAN buffer 1797 * to free or reuse. If bufspace useage is low 1798 * skip this step so we can allocate a new buffer. 1799 */ 1800 if (defrag || bufspace >= lobufspace) { 1801 nqindex = QUEUE_CLEAN; 1802 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1803 } 1804 1805 /* 1806 * If we could not find or were not allowed to reuse a 1807 * CLEAN buffer, check to see if it is ok to use an EMPTY 1808 * buffer. We can only use an EMPTY buffer if allocating 1809 * its KVA would not otherwise run us out of buffer space. 1810 */ 1811 if (nbp == NULL && defrag == 0 && 1812 bufspace + maxsize < hibufspace) { 1813 nqindex = QUEUE_EMPTY; 1814 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1815 } 1816 } 1817 1818 /* 1819 * Run scan, possibly freeing data and/or kva mappings on the fly 1820 * depending. 1821 */ 1822 1823 while ((bp = nbp) != NULL) { 1824 int qindex = nqindex; 1825 1826 /* 1827 * Calculate next bp ( we can only use it if we do not block 1828 * or do other fancy things ). 1829 */ 1830 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1831 switch(qindex) { 1832 case QUEUE_EMPTY: 1833 nqindex = QUEUE_EMPTYKVA; 1834 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1835 break; 1836 /* FALLTHROUGH */ 1837 case QUEUE_EMPTYKVA: 1838 nqindex = QUEUE_CLEAN; 1839 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1840 break; 1841 /* FALLTHROUGH */ 1842 case QUEUE_CLEAN: 1843 /* 1844 * nbp is NULL. 1845 */ 1846 break; 1847 } 1848 } 1849 if (bp->b_vp) { 1850 BO_LOCK(bp->b_bufobj); 1851 if (bp->b_vflags & BV_BKGRDINPROG) { 1852 BO_UNLOCK(bp->b_bufobj); 1853 continue; 1854 } 1855 BO_UNLOCK(bp->b_bufobj); 1856 } 1857 1858 /* 1859 * Sanity Checks 1860 */ 1861 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1862 1863 /* 1864 * Note: we no longer distinguish between VMIO and non-VMIO 1865 * buffers. 1866 */ 1867 1868 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1869 1870 /* 1871 * If we are defragging then we need a buffer with 1872 * b_kvasize != 0. XXX this situation should no longer 1873 * occur, if defrag is non-zero the buffer's b_kvasize 1874 * should also be non-zero at this point. XXX 1875 */ 1876 if (defrag && bp->b_kvasize == 0) { 1877 printf("Warning: defrag empty buffer %p\n", bp); 1878 continue; 1879 } 1880 1881 /* 1882 * Start freeing the bp. This is somewhat involved. nbp 1883 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1884 */ 1885 1886 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1887 panic("getnewbuf: locked buf"); 1888 bremfreel(bp); 1889 mtx_unlock(&bqlock); 1890 1891 if (qindex == QUEUE_CLEAN) { 1892 if (bp->b_flags & B_VMIO) { 1893 bp->b_flags &= ~B_ASYNC; 1894 vfs_vmio_release(bp); 1895 } 1896 if (bp->b_vp) 1897 brelvp(bp); 1898 } 1899 1900 /* 1901 * NOTE: nbp is now entirely invalid. We can only restart 1902 * the scan from this point on. 1903 * 1904 * Get the rest of the buffer freed up. b_kva* is still 1905 * valid after this operation. 1906 */ 1907 1908 if (bp->b_rcred != NOCRED) { 1909 crfree(bp->b_rcred); 1910 bp->b_rcred = NOCRED; 1911 } 1912 if (bp->b_wcred != NOCRED) { 1913 crfree(bp->b_wcred); 1914 bp->b_wcred = NOCRED; 1915 } 1916 if (LIST_FIRST(&bp->b_dep) != NULL) 1917 buf_deallocate(bp); 1918 if (bp->b_vflags & BV_BKGRDINPROG) 1919 panic("losing buffer 3"); 1920 1921 if (bp->b_bufsize) 1922 allocbuf(bp, 0); 1923 1924 bp->b_flags = 0; 1925 bp->b_ioflags = 0; 1926 bp->b_xflags = 0; 1927 bp->b_vflags = 0; 1928 bp->b_dev = NULL; 1929 bp->b_vp = NULL; 1930 bp->b_blkno = bp->b_lblkno = 0; 1931 bp->b_offset = NOOFFSET; 1932 bp->b_iodone = 0; 1933 bp->b_error = 0; 1934 bp->b_resid = 0; 1935 bp->b_bcount = 0; 1936 bp->b_npages = 0; 1937 bp->b_dirtyoff = bp->b_dirtyend = 0; 1938 bp->b_magic = B_MAGIC_BIO; 1939 bp->b_bufobj = NULL; 1940 1941 LIST_INIT(&bp->b_dep); 1942 1943 /* 1944 * If we are defragging then free the buffer. 1945 */ 1946 if (defrag) { 1947 bp->b_flags |= B_INVAL; 1948 bfreekva(bp); 1949 brelse(bp); 1950 defrag = 0; 1951 goto restart; 1952 } 1953 1954 /* 1955 * If we are overcomitted then recover the buffer and its 1956 * KVM space. This occurs in rare situations when multiple 1957 * processes are blocked in getnewbuf() or allocbuf(). 1958 */ 1959 if (bufspace >= hibufspace) 1960 flushingbufs = 1; 1961 if (flushingbufs && bp->b_kvasize != 0) { 1962 bp->b_flags |= B_INVAL; 1963 bfreekva(bp); 1964 brelse(bp); 1965 goto restart; 1966 } 1967 if (bufspace < lobufspace) 1968 flushingbufs = 0; 1969 break; 1970 } 1971 1972 /* 1973 * If we exhausted our list, sleep as appropriate. We may have to 1974 * wakeup various daemons and write out some dirty buffers. 1975 * 1976 * Generally we are sleeping due to insufficient buffer space. 1977 */ 1978 1979 if (bp == NULL) { 1980 int flags; 1981 char *waitmsg; 1982 1983 mtx_unlock(&bqlock); 1984 if (defrag) { 1985 flags = VFS_BIO_NEED_BUFSPACE; 1986 waitmsg = "nbufkv"; 1987 } else if (bufspace >= hibufspace) { 1988 waitmsg = "nbufbs"; 1989 flags = VFS_BIO_NEED_BUFSPACE; 1990 } else { 1991 waitmsg = "newbuf"; 1992 flags = VFS_BIO_NEED_ANY; 1993 } 1994 1995 bd_speedup(); /* heeeelp */ 1996 1997 mtx_lock(&nblock); 1998 needsbuffer |= flags; 1999 while (needsbuffer & flags) { 2000 if (msleep(&needsbuffer, &nblock, 2001 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2002 mtx_unlock(&nblock); 2003 return (NULL); 2004 } 2005 } 2006 mtx_unlock(&nblock); 2007 } else { 2008 /* 2009 * We finally have a valid bp. We aren't quite out of the 2010 * woods, we still have to reserve kva space. In order 2011 * to keep fragmentation sane we only allocate kva in 2012 * BKVASIZE chunks. 2013 */ 2014 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2015 2016 if (maxsize != bp->b_kvasize) { 2017 vm_offset_t addr = 0; 2018 2019 bfreekva(bp); 2020 2021 if (vm_map_findspace(buffer_map, 2022 vm_map_min(buffer_map), maxsize, &addr)) { 2023 /* 2024 * Uh oh. Buffer map is to fragmented. We 2025 * must defragment the map. 2026 */ 2027 atomic_add_int(&bufdefragcnt, 1); 2028 defrag = 1; 2029 bp->b_flags |= B_INVAL; 2030 brelse(bp); 2031 goto restart; 2032 } 2033 if (addr) { 2034 vm_map_insert(buffer_map, NULL, 0, 2035 addr, addr + maxsize, 2036 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2037 2038 bp->b_kvabase = (caddr_t) addr; 2039 bp->b_kvasize = maxsize; 2040 atomic_add_int(&bufspace, bp->b_kvasize); 2041 atomic_add_int(&bufreusecnt, 1); 2042 } 2043 } 2044 bp->b_saveaddr = bp->b_kvabase; 2045 bp->b_data = bp->b_saveaddr; 2046 } 2047 return(bp); 2048} 2049 2050/* 2051 * buf_daemon: 2052 * 2053 * buffer flushing daemon. Buffers are normally flushed by the 2054 * update daemon but if it cannot keep up this process starts to 2055 * take the load in an attempt to prevent getnewbuf() from blocking. 2056 */ 2057 2058static struct kproc_desc buf_kp = { 2059 "bufdaemon", 2060 buf_daemon, 2061 &bufdaemonproc 2062}; 2063SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2064 2065static void 2066buf_daemon() 2067{ 2068 int s; 2069 2070 mtx_lock(&Giant); 2071 2072 /* 2073 * This process needs to be suspended prior to shutdown sync. 2074 */ 2075 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2076 SHUTDOWN_PRI_LAST); 2077 2078 /* 2079 * This process is allowed to take the buffer cache to the limit 2080 */ 2081 s = splbio(); 2082 mtx_lock(&bdlock); 2083 2084 for (;;) { 2085 bd_request = 0; 2086 mtx_unlock(&bdlock); 2087 2088 kthread_suspend_check(bufdaemonproc); 2089 2090 /* 2091 * Do the flush. Limit the amount of in-transit I/O we 2092 * allow to build up, otherwise we would completely saturate 2093 * the I/O system. Wakeup any waiting processes before we 2094 * normally would so they can run in parallel with our drain. 2095 */ 2096 while (numdirtybuffers > lodirtybuffers) { 2097 if (flushbufqueues(0) == 0) { 2098 /* 2099 * Could not find any buffers without rollback 2100 * dependencies, so just write the first one 2101 * in the hopes of eventually making progress. 2102 */ 2103 flushbufqueues(1); 2104 break; 2105 } 2106 waitrunningbufspace(); 2107 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2108 } 2109 2110 /* 2111 * Only clear bd_request if we have reached our low water 2112 * mark. The buf_daemon normally waits 1 second and 2113 * then incrementally flushes any dirty buffers that have 2114 * built up, within reason. 2115 * 2116 * If we were unable to hit our low water mark and couldn't 2117 * find any flushable buffers, we sleep half a second. 2118 * Otherwise we loop immediately. 2119 */ 2120 mtx_lock(&bdlock); 2121 if (numdirtybuffers <= lodirtybuffers) { 2122 /* 2123 * We reached our low water mark, reset the 2124 * request and sleep until we are needed again. 2125 * The sleep is just so the suspend code works. 2126 */ 2127 bd_request = 0; 2128 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2129 } else { 2130 /* 2131 * We couldn't find any flushable dirty buffers but 2132 * still have too many dirty buffers, we 2133 * have to sleep and try again. (rare) 2134 */ 2135 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2136 } 2137 } 2138} 2139 2140/* 2141 * flushbufqueues: 2142 * 2143 * Try to flush a buffer in the dirty queue. We must be careful to 2144 * free up B_INVAL buffers instead of write them, which NFS is 2145 * particularly sensitive to. 2146 */ 2147int flushwithdeps = 0; 2148SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2149 0, "Number of buffers flushed with dependecies that require rollbacks"); 2150 2151static int 2152flushbufqueues(int flushdeps) 2153{ 2154 struct thread *td = curthread; 2155 struct vnode *vp; 2156 struct mount *mp; 2157 struct buf *bp; 2158 int hasdeps; 2159 2160 mtx_lock(&bqlock); 2161 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2162 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2163 continue; 2164 KASSERT((bp->b_flags & B_DELWRI), 2165 ("unexpected clean buffer %p", bp)); 2166 BO_LOCK(bp->b_bufobj); 2167 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 2168 BO_UNLOCK(bp->b_bufobj); 2169 BUF_UNLOCK(bp); 2170 continue; 2171 } 2172 BO_UNLOCK(bp->b_bufobj); 2173 if (bp->b_flags & B_INVAL) { 2174 bremfreel(bp); 2175 mtx_unlock(&bqlock); 2176 brelse(bp); 2177 return (1); 2178 } 2179 2180 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2181 if (flushdeps == 0) { 2182 BUF_UNLOCK(bp); 2183 continue; 2184 } 2185 hasdeps = 1; 2186 } else 2187 hasdeps = 0; 2188 /* 2189 * We must hold the lock on a vnode before writing 2190 * one of its buffers. Otherwise we may confuse, or 2191 * in the case of a snapshot vnode, deadlock the 2192 * system. 2193 * 2194 * The lock order here is the reverse of the normal 2195 * of vnode followed by buf lock. This is ok because 2196 * the NOWAIT will prevent deadlock. 2197 */ 2198 vp = bp->b_vp; 2199 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2200 BUF_UNLOCK(bp); 2201 continue; 2202 } 2203 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2204 mtx_unlock(&bqlock); 2205 vfs_bio_awrite(bp); 2206 vn_finished_write(mp); 2207 VOP_UNLOCK(vp, 0, td); 2208 flushwithdeps += hasdeps; 2209 return (1); 2210 } 2211 vn_finished_write(mp); 2212 BUF_UNLOCK(bp); 2213 } 2214 mtx_unlock(&bqlock); 2215 return (0); 2216} 2217 2218/* 2219 * Check to see if a block is currently memory resident. 2220 */ 2221struct buf * 2222incore(struct bufobj *bo, daddr_t blkno) 2223{ 2224 struct buf *bp; 2225 2226 int s = splbio(); 2227 BO_LOCK(bo); 2228 bp = gbincore(bo, blkno); 2229 BO_UNLOCK(bo); 2230 splx(s); 2231 return (bp); 2232} 2233 2234/* 2235 * Returns true if no I/O is needed to access the 2236 * associated VM object. This is like incore except 2237 * it also hunts around in the VM system for the data. 2238 */ 2239 2240static int 2241inmem(struct vnode * vp, daddr_t blkno) 2242{ 2243 vm_object_t obj; 2244 vm_offset_t toff, tinc, size; 2245 vm_page_t m; 2246 vm_ooffset_t off; 2247 2248 GIANT_REQUIRED; 2249 ASSERT_VOP_LOCKED(vp, "inmem"); 2250 2251 if (incore(&vp->v_bufobj, blkno)) 2252 return 1; 2253 if (vp->v_mount == NULL) 2254 return 0; 2255 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2256 return 0; 2257 2258 size = PAGE_SIZE; 2259 if (size > vp->v_mount->mnt_stat.f_iosize) 2260 size = vp->v_mount->mnt_stat.f_iosize; 2261 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2262 2263 VM_OBJECT_LOCK(obj); 2264 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2265 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2266 if (!m) 2267 goto notinmem; 2268 tinc = size; 2269 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2270 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2271 if (vm_page_is_valid(m, 2272 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2273 goto notinmem; 2274 } 2275 VM_OBJECT_UNLOCK(obj); 2276 return 1; 2277 2278notinmem: 2279 VM_OBJECT_UNLOCK(obj); 2280 return (0); 2281} 2282 2283/* 2284 * vfs_setdirty: 2285 * 2286 * Sets the dirty range for a buffer based on the status of the dirty 2287 * bits in the pages comprising the buffer. 2288 * 2289 * The range is limited to the size of the buffer. 2290 * 2291 * This routine is primarily used by NFS, but is generalized for the 2292 * B_VMIO case. 2293 */ 2294static void 2295vfs_setdirty(struct buf *bp) 2296{ 2297 int i; 2298 vm_object_t object; 2299 2300 GIANT_REQUIRED; 2301 /* 2302 * Degenerate case - empty buffer 2303 */ 2304 2305 if (bp->b_bufsize == 0) 2306 return; 2307 2308 /* 2309 * We qualify the scan for modified pages on whether the 2310 * object has been flushed yet. The OBJ_WRITEABLE flag 2311 * is not cleared simply by protecting pages off. 2312 */ 2313 2314 if ((bp->b_flags & B_VMIO) == 0) 2315 return; 2316 2317 object = bp->b_pages[0]->object; 2318 VM_OBJECT_LOCK(object); 2319 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2320 printf("Warning: object %p writeable but not mightbedirty\n", object); 2321 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2322 printf("Warning: object %p mightbedirty but not writeable\n", object); 2323 2324 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2325 vm_offset_t boffset; 2326 vm_offset_t eoffset; 2327 2328 vm_page_lock_queues(); 2329 /* 2330 * test the pages to see if they have been modified directly 2331 * by users through the VM system. 2332 */ 2333 for (i = 0; i < bp->b_npages; i++) 2334 vm_page_test_dirty(bp->b_pages[i]); 2335 2336 /* 2337 * Calculate the encompassing dirty range, boffset and eoffset, 2338 * (eoffset - boffset) bytes. 2339 */ 2340 2341 for (i = 0; i < bp->b_npages; i++) { 2342 if (bp->b_pages[i]->dirty) 2343 break; 2344 } 2345 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2346 2347 for (i = bp->b_npages - 1; i >= 0; --i) { 2348 if (bp->b_pages[i]->dirty) { 2349 break; 2350 } 2351 } 2352 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2353 2354 vm_page_unlock_queues(); 2355 /* 2356 * Fit it to the buffer. 2357 */ 2358 2359 if (eoffset > bp->b_bcount) 2360 eoffset = bp->b_bcount; 2361 2362 /* 2363 * If we have a good dirty range, merge with the existing 2364 * dirty range. 2365 */ 2366 2367 if (boffset < eoffset) { 2368 if (bp->b_dirtyoff > boffset) 2369 bp->b_dirtyoff = boffset; 2370 if (bp->b_dirtyend < eoffset) 2371 bp->b_dirtyend = eoffset; 2372 } 2373 } 2374 VM_OBJECT_UNLOCK(object); 2375} 2376 2377/* 2378 * getblk: 2379 * 2380 * Get a block given a specified block and offset into a file/device. 2381 * The buffers B_DONE bit will be cleared on return, making it almost 2382 * ready for an I/O initiation. B_INVAL may or may not be set on 2383 * return. The caller should clear B_INVAL prior to initiating a 2384 * READ. 2385 * 2386 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2387 * an existing buffer. 2388 * 2389 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2390 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2391 * and then cleared based on the backing VM. If the previous buffer is 2392 * non-0-sized but invalid, B_CACHE will be cleared. 2393 * 2394 * If getblk() must create a new buffer, the new buffer is returned with 2395 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2396 * case it is returned with B_INVAL clear and B_CACHE set based on the 2397 * backing VM. 2398 * 2399 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2400 * B_CACHE bit is clear. 2401 * 2402 * What this means, basically, is that the caller should use B_CACHE to 2403 * determine whether the buffer is fully valid or not and should clear 2404 * B_INVAL prior to issuing a read. If the caller intends to validate 2405 * the buffer by loading its data area with something, the caller needs 2406 * to clear B_INVAL. If the caller does this without issuing an I/O, 2407 * the caller should set B_CACHE ( as an optimization ), else the caller 2408 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2409 * a write attempt or if it was a successfull read. If the caller 2410 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2411 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2412 */ 2413struct buf * 2414getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2415 int flags) 2416{ 2417 struct buf *bp; 2418 struct bufobj *bo; 2419 int s; 2420 int error; 2421 ASSERT_VOP_LOCKED(vp, "getblk"); 2422 struct vm_object *vmo; 2423 2424 if (size > MAXBSIZE) 2425 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2426 2427 bo = &vp->v_bufobj; 2428 s = splbio(); 2429loop: 2430 /* 2431 * Block if we are low on buffers. Certain processes are allowed 2432 * to completely exhaust the buffer cache. 2433 * 2434 * If this check ever becomes a bottleneck it may be better to 2435 * move it into the else, when gbincore() fails. At the moment 2436 * it isn't a problem. 2437 * 2438 * XXX remove if 0 sections (clean this up after its proven) 2439 */ 2440 if (numfreebuffers == 0) { 2441 if (curthread == PCPU_GET(idlethread)) 2442 return NULL; 2443 mtx_lock(&nblock); 2444 needsbuffer |= VFS_BIO_NEED_ANY; 2445 mtx_unlock(&nblock); 2446 } 2447 2448 VI_LOCK(vp); 2449 bp = gbincore(bo, blkno); 2450 if (bp != NULL) { 2451 int lockflags; 2452 /* 2453 * Buffer is in-core. If the buffer is not busy, it must 2454 * be on a queue. 2455 */ 2456 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2457 2458 if (flags & GB_LOCK_NOWAIT) 2459 lockflags |= LK_NOWAIT; 2460 2461 error = BUF_TIMELOCK(bp, lockflags, 2462 VI_MTX(vp), "getblk", slpflag, slptimeo); 2463 2464 /* 2465 * If we slept and got the lock we have to restart in case 2466 * the buffer changed identities. 2467 */ 2468 if (error == ENOLCK) 2469 goto loop; 2470 /* We timed out or were interrupted. */ 2471 else if (error) 2472 return (NULL); 2473 2474 /* 2475 * The buffer is locked. B_CACHE is cleared if the buffer is 2476 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2477 * and for a VMIO buffer B_CACHE is adjusted according to the 2478 * backing VM cache. 2479 */ 2480 if (bp->b_flags & B_INVAL) 2481 bp->b_flags &= ~B_CACHE; 2482 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2483 bp->b_flags |= B_CACHE; 2484 bremfree(bp); 2485 2486 /* 2487 * check for size inconsistancies for non-VMIO case. 2488 */ 2489 2490 if (bp->b_bcount != size) { 2491 if ((bp->b_flags & B_VMIO) == 0 || 2492 (size > bp->b_kvasize)) { 2493 if (bp->b_flags & B_DELWRI) { 2494 bp->b_flags |= B_NOCACHE; 2495 bwrite(bp); 2496 } else { 2497 if ((bp->b_flags & B_VMIO) && 2498 (LIST_FIRST(&bp->b_dep) == NULL)) { 2499 bp->b_flags |= B_RELBUF; 2500 brelse(bp); 2501 } else { 2502 bp->b_flags |= B_NOCACHE; 2503 bwrite(bp); 2504 } 2505 } 2506 goto loop; 2507 } 2508 } 2509 2510 /* 2511 * If the size is inconsistant in the VMIO case, we can resize 2512 * the buffer. This might lead to B_CACHE getting set or 2513 * cleared. If the size has not changed, B_CACHE remains 2514 * unchanged from its previous state. 2515 */ 2516 2517 if (bp->b_bcount != size) 2518 allocbuf(bp, size); 2519 2520 KASSERT(bp->b_offset != NOOFFSET, 2521 ("getblk: no buffer offset")); 2522 2523 /* 2524 * A buffer with B_DELWRI set and B_CACHE clear must 2525 * be committed before we can return the buffer in 2526 * order to prevent the caller from issuing a read 2527 * ( due to B_CACHE not being set ) and overwriting 2528 * it. 2529 * 2530 * Most callers, including NFS and FFS, need this to 2531 * operate properly either because they assume they 2532 * can issue a read if B_CACHE is not set, or because 2533 * ( for example ) an uncached B_DELWRI might loop due 2534 * to softupdates re-dirtying the buffer. In the latter 2535 * case, B_CACHE is set after the first write completes, 2536 * preventing further loops. 2537 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2538 * above while extending the buffer, we cannot allow the 2539 * buffer to remain with B_CACHE set after the write 2540 * completes or it will represent a corrupt state. To 2541 * deal with this we set B_NOCACHE to scrap the buffer 2542 * after the write. 2543 * 2544 * We might be able to do something fancy, like setting 2545 * B_CACHE in bwrite() except if B_DELWRI is already set, 2546 * so the below call doesn't set B_CACHE, but that gets real 2547 * confusing. This is much easier. 2548 */ 2549 2550 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2551 bp->b_flags |= B_NOCACHE; 2552 bwrite(bp); 2553 goto loop; 2554 } 2555 2556 splx(s); 2557 bp->b_flags &= ~B_DONE; 2558 } else { 2559 int bsize, maxsize, vmio; 2560 off_t offset; 2561 2562 /* 2563 * Buffer is not in-core, create new buffer. The buffer 2564 * returned by getnewbuf() is locked. Note that the returned 2565 * buffer is also considered valid (not marked B_INVAL). 2566 */ 2567 VI_UNLOCK(vp); 2568 /* 2569 * If the user does not want us to create the buffer, bail out 2570 * here. 2571 */ 2572 if (flags & GB_NOCREAT) { 2573 splx(s); 2574 return NULL; 2575 } 2576 2577 bsize = bo->bo_bsize; 2578 offset = blkno * bsize; 2579 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2580 (vp->v_vflag & VV_OBJBUF); 2581 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2582 maxsize = imax(maxsize, bsize); 2583 2584 bp = getnewbuf(slpflag, slptimeo, size, maxsize); 2585 if (bp == NULL) { 2586 if (slpflag || slptimeo) { 2587 splx(s); 2588 return NULL; 2589 } 2590 goto loop; 2591 } 2592 2593 /* 2594 * This code is used to make sure that a buffer is not 2595 * created while the getnewbuf routine is blocked. 2596 * This can be a problem whether the vnode is locked or not. 2597 * If the buffer is created out from under us, we have to 2598 * throw away the one we just created. There is now window 2599 * race because we are safely running at splbio() from the 2600 * point of the duplicate buffer creation through to here, 2601 * and we've locked the buffer. 2602 * 2603 * Note: this must occur before we associate the buffer 2604 * with the vp especially considering limitations in 2605 * the splay tree implementation when dealing with duplicate 2606 * lblkno's. 2607 */ 2608 BO_LOCK(bo); 2609 if (gbincore(bo, blkno)) { 2610 BO_UNLOCK(bo); 2611 bp->b_flags |= B_INVAL; 2612 brelse(bp); 2613 goto loop; 2614 } 2615 2616 /* 2617 * Insert the buffer into the hash, so that it can 2618 * be found by incore. 2619 */ 2620 bp->b_blkno = bp->b_lblkno = blkno; 2621 bp->b_offset = offset; 2622 2623 bgetvp(vp, bp); 2624 BO_UNLOCK(bo); 2625 2626 /* 2627 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2628 * buffer size starts out as 0, B_CACHE will be set by 2629 * allocbuf() for the VMIO case prior to it testing the 2630 * backing store for validity. 2631 */ 2632 2633 if (vmio) { 2634 bp->b_flags |= B_VMIO; 2635#if defined(VFS_BIO_DEBUG) 2636 if (vn_canvmio(vp) != TRUE) 2637 printf("getblk: VMIO on vnode type %d\n", 2638 vp->v_type); 2639#endif 2640 VOP_GETVOBJECT(vp, &vmo); 2641 KASSERT(vmo == bp->b_object, 2642 ("ARGH! different b_object %p %p %p\n", bp, vmo, bp->b_object)); 2643 } else { 2644 bp->b_flags &= ~B_VMIO; 2645 KASSERT(bp->b_object == NULL, 2646 ("ARGH! has b_object %p %p\n", bp, bp->b_object)); 2647 } 2648 2649 allocbuf(bp, size); 2650 2651 splx(s); 2652 bp->b_flags &= ~B_DONE; 2653 } 2654 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2655 KASSERT(bp->b_bufobj == bo, 2656 ("wrong b_bufobj %p should be %p", bp->b_bufobj, bo)); 2657 return (bp); 2658} 2659 2660/* 2661 * Get an empty, disassociated buffer of given size. The buffer is initially 2662 * set to B_INVAL. 2663 */ 2664struct buf * 2665geteblk(int size) 2666{ 2667 struct buf *bp; 2668 int s; 2669 int maxsize; 2670 2671 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2672 2673 s = splbio(); 2674 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2675 continue; 2676 splx(s); 2677 allocbuf(bp, size); 2678 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2679 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2680 return (bp); 2681} 2682 2683 2684/* 2685 * This code constitutes the buffer memory from either anonymous system 2686 * memory (in the case of non-VMIO operations) or from an associated 2687 * VM object (in the case of VMIO operations). This code is able to 2688 * resize a buffer up or down. 2689 * 2690 * Note that this code is tricky, and has many complications to resolve 2691 * deadlock or inconsistant data situations. Tread lightly!!! 2692 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2693 * the caller. Calling this code willy nilly can result in the loss of data. 2694 * 2695 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2696 * B_CACHE for the non-VMIO case. 2697 */ 2698 2699int 2700allocbuf(struct buf *bp, int size) 2701{ 2702 int newbsize, mbsize; 2703 int i; 2704 2705 GIANT_REQUIRED; 2706 2707 if (BUF_REFCNT(bp) == 0) 2708 panic("allocbuf: buffer not busy"); 2709 2710 if (bp->b_kvasize < size) 2711 panic("allocbuf: buffer too small"); 2712 2713 if ((bp->b_flags & B_VMIO) == 0) { 2714 caddr_t origbuf; 2715 int origbufsize; 2716 /* 2717 * Just get anonymous memory from the kernel. Don't 2718 * mess with B_CACHE. 2719 */ 2720 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2721 if (bp->b_flags & B_MALLOC) 2722 newbsize = mbsize; 2723 else 2724 newbsize = round_page(size); 2725 2726 if (newbsize < bp->b_bufsize) { 2727 /* 2728 * malloced buffers are not shrunk 2729 */ 2730 if (bp->b_flags & B_MALLOC) { 2731 if (newbsize) { 2732 bp->b_bcount = size; 2733 } else { 2734 free(bp->b_data, M_BIOBUF); 2735 if (bp->b_bufsize) { 2736 atomic_subtract_int( 2737 &bufmallocspace, 2738 bp->b_bufsize); 2739 bufspacewakeup(); 2740 bp->b_bufsize = 0; 2741 } 2742 bp->b_saveaddr = bp->b_kvabase; 2743 bp->b_data = bp->b_saveaddr; 2744 bp->b_bcount = 0; 2745 bp->b_flags &= ~B_MALLOC; 2746 } 2747 return 1; 2748 } 2749 vm_hold_free_pages( 2750 bp, 2751 (vm_offset_t) bp->b_data + newbsize, 2752 (vm_offset_t) bp->b_data + bp->b_bufsize); 2753 } else if (newbsize > bp->b_bufsize) { 2754 /* 2755 * We only use malloced memory on the first allocation. 2756 * and revert to page-allocated memory when the buffer 2757 * grows. 2758 */ 2759 /* 2760 * There is a potential smp race here that could lead 2761 * to bufmallocspace slightly passing the max. It 2762 * is probably extremely rare and not worth worrying 2763 * over. 2764 */ 2765 if ( (bufmallocspace < maxbufmallocspace) && 2766 (bp->b_bufsize == 0) && 2767 (mbsize <= PAGE_SIZE/2)) { 2768 2769 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2770 bp->b_bufsize = mbsize; 2771 bp->b_bcount = size; 2772 bp->b_flags |= B_MALLOC; 2773 atomic_add_int(&bufmallocspace, mbsize); 2774 return 1; 2775 } 2776 origbuf = NULL; 2777 origbufsize = 0; 2778 /* 2779 * If the buffer is growing on its other-than-first allocation, 2780 * then we revert to the page-allocation scheme. 2781 */ 2782 if (bp->b_flags & B_MALLOC) { 2783 origbuf = bp->b_data; 2784 origbufsize = bp->b_bufsize; 2785 bp->b_data = bp->b_kvabase; 2786 if (bp->b_bufsize) { 2787 atomic_subtract_int(&bufmallocspace, 2788 bp->b_bufsize); 2789 bufspacewakeup(); 2790 bp->b_bufsize = 0; 2791 } 2792 bp->b_flags &= ~B_MALLOC; 2793 newbsize = round_page(newbsize); 2794 } 2795 vm_hold_load_pages( 2796 bp, 2797 (vm_offset_t) bp->b_data + bp->b_bufsize, 2798 (vm_offset_t) bp->b_data + newbsize); 2799 if (origbuf) { 2800 bcopy(origbuf, bp->b_data, origbufsize); 2801 free(origbuf, M_BIOBUF); 2802 } 2803 } 2804 } else { 2805 int desiredpages; 2806 2807 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2808 desiredpages = (size == 0) ? 0 : 2809 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2810 2811 if (bp->b_flags & B_MALLOC) 2812 panic("allocbuf: VMIO buffer can't be malloced"); 2813 /* 2814 * Set B_CACHE initially if buffer is 0 length or will become 2815 * 0-length. 2816 */ 2817 if (size == 0 || bp->b_bufsize == 0) 2818 bp->b_flags |= B_CACHE; 2819 2820 if (newbsize < bp->b_bufsize) { 2821 /* 2822 * DEV_BSIZE aligned new buffer size is less then the 2823 * DEV_BSIZE aligned existing buffer size. Figure out 2824 * if we have to remove any pages. 2825 */ 2826 if (desiredpages < bp->b_npages) { 2827 vm_page_t m; 2828 2829 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2830 vm_page_lock_queues(); 2831 for (i = desiredpages; i < bp->b_npages; i++) { 2832 /* 2833 * the page is not freed here -- it 2834 * is the responsibility of 2835 * vnode_pager_setsize 2836 */ 2837 m = bp->b_pages[i]; 2838 KASSERT(m != bogus_page, 2839 ("allocbuf: bogus page found")); 2840 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2841 vm_page_lock_queues(); 2842 2843 bp->b_pages[i] = NULL; 2844 vm_page_unwire(m, 0); 2845 } 2846 vm_page_unlock_queues(); 2847 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2848 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2849 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2850 bp->b_npages = desiredpages; 2851 } 2852 } else if (size > bp->b_bcount) { 2853 /* 2854 * We are growing the buffer, possibly in a 2855 * byte-granular fashion. 2856 */ 2857 struct vnode *vp; 2858 vm_object_t obj; 2859 vm_offset_t toff; 2860 vm_offset_t tinc; 2861 2862 /* 2863 * Step 1, bring in the VM pages from the object, 2864 * allocating them if necessary. We must clear 2865 * B_CACHE if these pages are not valid for the 2866 * range covered by the buffer. 2867 */ 2868 2869 vp = bp->b_vp; 2870 obj = bp->b_object; 2871 2872 VM_OBJECT_LOCK(obj); 2873 while (bp->b_npages < desiredpages) { 2874 vm_page_t m; 2875 vm_pindex_t pi; 2876 2877 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2878 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2879 /* 2880 * note: must allocate system pages 2881 * since blocking here could intefere 2882 * with paging I/O, no matter which 2883 * process we are. 2884 */ 2885 m = vm_page_alloc(obj, pi, 2886 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2887 VM_ALLOC_WIRED); 2888 if (m == NULL) { 2889 atomic_add_int(&vm_pageout_deficit, 2890 desiredpages - bp->b_npages); 2891 VM_OBJECT_UNLOCK(obj); 2892 VM_WAIT; 2893 VM_OBJECT_LOCK(obj); 2894 } else { 2895 bp->b_flags &= ~B_CACHE; 2896 bp->b_pages[bp->b_npages] = m; 2897 ++bp->b_npages; 2898 } 2899 continue; 2900 } 2901 2902 /* 2903 * We found a page. If we have to sleep on it, 2904 * retry because it might have gotten freed out 2905 * from under us. 2906 * 2907 * We can only test PG_BUSY here. Blocking on 2908 * m->busy might lead to a deadlock: 2909 * 2910 * vm_fault->getpages->cluster_read->allocbuf 2911 * 2912 */ 2913 vm_page_lock_queues(); 2914 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2915 continue; 2916 2917 /* 2918 * We have a good page. Should we wakeup the 2919 * page daemon? 2920 */ 2921 if ((curproc != pageproc) && 2922 ((m->queue - m->pc) == PQ_CACHE) && 2923 ((cnt.v_free_count + cnt.v_cache_count) < 2924 (cnt.v_free_min + cnt.v_cache_min))) { 2925 pagedaemon_wakeup(); 2926 } 2927 vm_page_wire(m); 2928 vm_page_unlock_queues(); 2929 bp->b_pages[bp->b_npages] = m; 2930 ++bp->b_npages; 2931 } 2932 2933 /* 2934 * Step 2. We've loaded the pages into the buffer, 2935 * we have to figure out if we can still have B_CACHE 2936 * set. Note that B_CACHE is set according to the 2937 * byte-granular range ( bcount and size ), new the 2938 * aligned range ( newbsize ). 2939 * 2940 * The VM test is against m->valid, which is DEV_BSIZE 2941 * aligned. Needless to say, the validity of the data 2942 * needs to also be DEV_BSIZE aligned. Note that this 2943 * fails with NFS if the server or some other client 2944 * extends the file's EOF. If our buffer is resized, 2945 * B_CACHE may remain set! XXX 2946 */ 2947 2948 toff = bp->b_bcount; 2949 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2950 2951 while ((bp->b_flags & B_CACHE) && toff < size) { 2952 vm_pindex_t pi; 2953 2954 if (tinc > (size - toff)) 2955 tinc = size - toff; 2956 2957 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2958 PAGE_SHIFT; 2959 2960 vfs_buf_test_cache( 2961 bp, 2962 bp->b_offset, 2963 toff, 2964 tinc, 2965 bp->b_pages[pi] 2966 ); 2967 toff += tinc; 2968 tinc = PAGE_SIZE; 2969 } 2970 VM_OBJECT_UNLOCK(obj); 2971 2972 /* 2973 * Step 3, fixup the KVM pmap. Remember that 2974 * bp->b_data is relative to bp->b_offset, but 2975 * bp->b_offset may be offset into the first page. 2976 */ 2977 2978 bp->b_data = (caddr_t) 2979 trunc_page((vm_offset_t)bp->b_data); 2980 pmap_qenter( 2981 (vm_offset_t)bp->b_data, 2982 bp->b_pages, 2983 bp->b_npages 2984 ); 2985 2986 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2987 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2988 } 2989 } 2990 if (newbsize < bp->b_bufsize) 2991 bufspacewakeup(); 2992 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2993 bp->b_bcount = size; /* requested buffer size */ 2994 return 1; 2995} 2996 2997void 2998biodone(struct bio *bp) 2999{ 3000 3001 mtx_lock(&bdonelock); 3002 bp->bio_flags |= BIO_DONE; 3003 if (bp->bio_done == NULL) 3004 wakeup(bp); 3005 mtx_unlock(&bdonelock); 3006 if (bp->bio_done != NULL) 3007 bp->bio_done(bp); 3008} 3009 3010/* 3011 * Wait for a BIO to finish. 3012 * 3013 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3014 * case is not yet clear. 3015 */ 3016int 3017biowait(struct bio *bp, const char *wchan) 3018{ 3019 3020 mtx_lock(&bdonelock); 3021 while ((bp->bio_flags & BIO_DONE) == 0) 3022 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3023 mtx_unlock(&bdonelock); 3024 if (bp->bio_error != 0) 3025 return (bp->bio_error); 3026 if (!(bp->bio_flags & BIO_ERROR)) 3027 return (0); 3028 return (EIO); 3029} 3030 3031void 3032biofinish(struct bio *bp, struct devstat *stat, int error) 3033{ 3034 3035 if (error) { 3036 bp->bio_error = error; 3037 bp->bio_flags |= BIO_ERROR; 3038 } 3039 if (stat != NULL) 3040 devstat_end_transaction_bio(stat, bp); 3041 biodone(bp); 3042} 3043 3044/* 3045 * bufwait: 3046 * 3047 * Wait for buffer I/O completion, returning error status. The buffer 3048 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3049 * error and cleared. 3050 */ 3051int 3052bufwait(struct buf *bp) 3053{ 3054 int s; 3055 3056 s = splbio(); 3057 if (bp->b_iocmd == BIO_READ) 3058 bwait(bp, PRIBIO, "biord"); 3059 else 3060 bwait(bp, PRIBIO, "biowr"); 3061 splx(s); 3062 if (bp->b_flags & B_EINTR) { 3063 bp->b_flags &= ~B_EINTR; 3064 return (EINTR); 3065 } 3066 if (bp->b_ioflags & BIO_ERROR) { 3067 return (bp->b_error ? bp->b_error : EIO); 3068 } else { 3069 return (0); 3070 } 3071} 3072 3073 /* 3074 * Call back function from struct bio back up to struct buf. 3075 */ 3076static void 3077bufdonebio(struct bio *bip) 3078{ 3079 struct buf *bp; 3080 3081 /* Device drivers may or may not hold giant, hold it here. */ 3082 mtx_lock(&Giant); 3083 bp = bip->bio_caller2; 3084 bp->b_resid = bp->b_bcount - bip->bio_completed; 3085 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3086 bp->b_ioflags = bip->bio_flags; 3087 bp->b_error = bip->bio_error; 3088 if (bp->b_error) 3089 bp->b_ioflags |= BIO_ERROR; 3090 bufdone(bp); 3091 mtx_unlock(&Giant); 3092 g_destroy_bio(bip); 3093} 3094 3095void 3096dev_strategy(struct cdev *dev, struct buf *bp) 3097{ 3098 struct cdevsw *csw; 3099 struct bio *bip; 3100 3101 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3102 panic("b_iocmd botch"); 3103 for (;;) { 3104 bip = g_new_bio(); 3105 if (bip != NULL) 3106 break; 3107 /* Try again later */ 3108 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3109 } 3110 bip->bio_cmd = bp->b_iocmd; 3111 bip->bio_offset = bp->b_iooffset; 3112 bip->bio_length = bp->b_bcount; 3113 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3114 bip->bio_data = bp->b_data; 3115 bip->bio_done = bufdonebio; 3116 bip->bio_caller2 = bp; 3117 bip->bio_dev = dev; 3118 3119 KASSERT(dev->si_refcount > 0, 3120 ("dev_strategy on un-referenced struct cdev *(%s)", 3121 devtoname(dev))); 3122 csw = dev_refthread(dev); 3123 if (csw == NULL) { 3124 bp->b_error = ENXIO; 3125 bp->b_ioflags = BIO_ERROR; 3126 mtx_lock(&Giant); /* XXX: too defensive ? */ 3127 bufdone(bp); 3128 mtx_unlock(&Giant); /* XXX: too defensive ? */ 3129 return; 3130 } 3131 (*csw->d_strategy)(bip); 3132 dev_relthread(dev); 3133} 3134 3135/* 3136 * bufdone: 3137 * 3138 * Finish I/O on a buffer, optionally calling a completion function. 3139 * This is usually called from an interrupt so process blocking is 3140 * not allowed. 3141 * 3142 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3143 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3144 * assuming B_INVAL is clear. 3145 * 3146 * For the VMIO case, we set B_CACHE if the op was a read and no 3147 * read error occured, or if the op was a write. B_CACHE is never 3148 * set if the buffer is invalid or otherwise uncacheable. 3149 * 3150 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3151 * initiator to leave B_INVAL set to brelse the buffer out of existance 3152 * in the biodone routine. 3153 */ 3154void 3155bufdone(struct buf *bp) 3156{ 3157 int s; 3158 void (*biodone)(struct buf *); 3159 3160 3161 s = splbio(); 3162 3163 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3164 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3165 3166 bp->b_flags |= B_DONE; 3167 runningbufwakeup(bp); 3168 3169 if (bp->b_iocmd == BIO_WRITE && bp->b_bufobj != NULL) 3170 bufobj_wdrop(bp->b_bufobj); 3171 3172 /* call optional completion function if requested */ 3173 if (bp->b_iodone != NULL) { 3174 biodone = bp->b_iodone; 3175 bp->b_iodone = NULL; 3176 (*biodone) (bp); 3177 splx(s); 3178 return; 3179 } 3180 if (LIST_FIRST(&bp->b_dep) != NULL) 3181 buf_complete(bp); 3182 3183 if (bp->b_flags & B_VMIO) { 3184 int i; 3185 vm_ooffset_t foff; 3186 vm_page_t m; 3187 vm_object_t obj; 3188 int iosize; 3189 struct vnode *vp = bp->b_vp; 3190 3191 obj = bp->b_object; 3192 3193#if defined(VFS_BIO_DEBUG) 3194 mp_fixme("usecount and vflag accessed without locks."); 3195 if (vp->v_usecount == 0) { 3196 panic("biodone: zero vnode ref count"); 3197 } 3198 3199 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3200 panic("biodone: vnode is not setup for merged cache"); 3201 } 3202#endif 3203 3204 foff = bp->b_offset; 3205 KASSERT(bp->b_offset != NOOFFSET, 3206 ("biodone: no buffer offset")); 3207 3208 VM_OBJECT_LOCK(obj); 3209#if defined(VFS_BIO_DEBUG) 3210 if (obj->paging_in_progress < bp->b_npages) { 3211 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3212 obj->paging_in_progress, bp->b_npages); 3213 } 3214#endif 3215 3216 /* 3217 * Set B_CACHE if the op was a normal read and no error 3218 * occured. B_CACHE is set for writes in the b*write() 3219 * routines. 3220 */ 3221 iosize = bp->b_bcount - bp->b_resid; 3222 if (bp->b_iocmd == BIO_READ && 3223 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3224 !(bp->b_ioflags & BIO_ERROR)) { 3225 bp->b_flags |= B_CACHE; 3226 } 3227 vm_page_lock_queues(); 3228 for (i = 0; i < bp->b_npages; i++) { 3229 int bogusflag = 0; 3230 int resid; 3231 3232 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3233 if (resid > iosize) 3234 resid = iosize; 3235 3236 /* 3237 * cleanup bogus pages, restoring the originals 3238 */ 3239 m = bp->b_pages[i]; 3240 if (m == bogus_page) { 3241 bogusflag = 1; 3242 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3243 if (m == NULL) 3244 panic("biodone: page disappeared!"); 3245 bp->b_pages[i] = m; 3246 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3247 } 3248#if defined(VFS_BIO_DEBUG) 3249 if (OFF_TO_IDX(foff) != m->pindex) { 3250 printf( 3251"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3252 (intmax_t)foff, (uintmax_t)m->pindex); 3253 } 3254#endif 3255 3256 /* 3257 * In the write case, the valid and clean bits are 3258 * already changed correctly ( see bdwrite() ), so we 3259 * only need to do this here in the read case. 3260 */ 3261 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3262 vfs_page_set_valid(bp, foff, i, m); 3263 } 3264 3265 /* 3266 * when debugging new filesystems or buffer I/O methods, this 3267 * is the most common error that pops up. if you see this, you 3268 * have not set the page busy flag correctly!!! 3269 */ 3270 if (m->busy == 0) { 3271 printf("biodone: page busy < 0, " 3272 "pindex: %d, foff: 0x(%x,%x), " 3273 "resid: %d, index: %d\n", 3274 (int) m->pindex, (int)(foff >> 32), 3275 (int) foff & 0xffffffff, resid, i); 3276 if (!vn_isdisk(vp, NULL)) 3277 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3278 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3279 (intmax_t) bp->b_lblkno, 3280 bp->b_flags, bp->b_npages); 3281 else 3282 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3283 (intmax_t) bp->b_lblkno, 3284 bp->b_flags, bp->b_npages); 3285 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3286 (u_long)m->valid, (u_long)m->dirty, 3287 m->wire_count); 3288 panic("biodone: page busy < 0\n"); 3289 } 3290 vm_page_io_finish(m); 3291 vm_object_pip_subtract(obj, 1); 3292 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3293 iosize -= resid; 3294 } 3295 vm_page_unlock_queues(); 3296 vm_object_pip_wakeupn(obj, 0); 3297 VM_OBJECT_UNLOCK(obj); 3298 } 3299 3300 /* 3301 * For asynchronous completions, release the buffer now. The brelse 3302 * will do a wakeup there if necessary - so no need to do a wakeup 3303 * here in the async case. The sync case always needs to do a wakeup. 3304 */ 3305 3306 if (bp->b_flags & B_ASYNC) { 3307 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3308 brelse(bp); 3309 else 3310 bqrelse(bp); 3311 } else { 3312 bdone(bp); 3313 } 3314 splx(s); 3315} 3316 3317/* 3318 * This routine is called in lieu of iodone in the case of 3319 * incomplete I/O. This keeps the busy status for pages 3320 * consistant. 3321 */ 3322void 3323vfs_unbusy_pages(struct buf *bp) 3324{ 3325 int i; 3326 vm_object_t obj; 3327 vm_page_t m; 3328 3329 runningbufwakeup(bp); 3330 if (!(bp->b_flags & B_VMIO)) 3331 return; 3332 3333 obj = bp->b_object; 3334 VM_OBJECT_LOCK(obj); 3335 vm_page_lock_queues(); 3336 for (i = 0; i < bp->b_npages; i++) { 3337 m = bp->b_pages[i]; 3338 if (m == bogus_page) { 3339 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3340 if (!m) 3341 panic("vfs_unbusy_pages: page missing\n"); 3342 bp->b_pages[i] = m; 3343 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3344 bp->b_pages, bp->b_npages); 3345 } 3346 vm_object_pip_subtract(obj, 1); 3347 vm_page_io_finish(m); 3348 } 3349 vm_page_unlock_queues(); 3350 vm_object_pip_wakeupn(obj, 0); 3351 VM_OBJECT_UNLOCK(obj); 3352} 3353 3354/* 3355 * vfs_page_set_valid: 3356 * 3357 * Set the valid bits in a page based on the supplied offset. The 3358 * range is restricted to the buffer's size. 3359 * 3360 * This routine is typically called after a read completes. 3361 */ 3362static void 3363vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3364{ 3365 vm_ooffset_t soff, eoff; 3366 3367 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3368 /* 3369 * Start and end offsets in buffer. eoff - soff may not cross a 3370 * page boundry or cross the end of the buffer. The end of the 3371 * buffer, in this case, is our file EOF, not the allocation size 3372 * of the buffer. 3373 */ 3374 soff = off; 3375 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3376 if (eoff > bp->b_offset + bp->b_bcount) 3377 eoff = bp->b_offset + bp->b_bcount; 3378 3379 /* 3380 * Set valid range. This is typically the entire buffer and thus the 3381 * entire page. 3382 */ 3383 if (eoff > soff) { 3384 vm_page_set_validclean( 3385 m, 3386 (vm_offset_t) (soff & PAGE_MASK), 3387 (vm_offset_t) (eoff - soff) 3388 ); 3389 } 3390} 3391 3392/* 3393 * This routine is called before a device strategy routine. 3394 * It is used to tell the VM system that paging I/O is in 3395 * progress, and treat the pages associated with the buffer 3396 * almost as being PG_BUSY. Also the object paging_in_progress 3397 * flag is handled to make sure that the object doesn't become 3398 * inconsistant. 3399 * 3400 * Since I/O has not been initiated yet, certain buffer flags 3401 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3402 * and should be ignored. 3403 */ 3404void 3405vfs_busy_pages(struct buf *bp, int clear_modify) 3406{ 3407 int i, bogus; 3408 vm_object_t obj; 3409 vm_ooffset_t foff; 3410 vm_page_t m; 3411 3412 if (!(bp->b_flags & B_VMIO)) 3413 return; 3414 3415 obj = bp->b_object; 3416 foff = bp->b_offset; 3417 KASSERT(bp->b_offset != NOOFFSET, 3418 ("vfs_busy_pages: no buffer offset")); 3419 vfs_setdirty(bp); 3420 VM_OBJECT_LOCK(obj); 3421retry: 3422 vm_page_lock_queues(); 3423 for (i = 0; i < bp->b_npages; i++) { 3424 m = bp->b_pages[i]; 3425 3426 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3427 goto retry; 3428 } 3429 bogus = 0; 3430 for (i = 0; i < bp->b_npages; i++) { 3431 m = bp->b_pages[i]; 3432 3433 if ((bp->b_flags & B_CLUSTER) == 0) { 3434 vm_object_pip_add(obj, 1); 3435 vm_page_io_start(m); 3436 } 3437 /* 3438 * When readying a buffer for a read ( i.e 3439 * clear_modify == 0 ), it is important to do 3440 * bogus_page replacement for valid pages in 3441 * partially instantiated buffers. Partially 3442 * instantiated buffers can, in turn, occur when 3443 * reconstituting a buffer from its VM backing store 3444 * base. We only have to do this if B_CACHE is 3445 * clear ( which causes the I/O to occur in the 3446 * first place ). The replacement prevents the read 3447 * I/O from overwriting potentially dirty VM-backed 3448 * pages. XXX bogus page replacement is, uh, bogus. 3449 * It may not work properly with small-block devices. 3450 * We need to find a better way. 3451 */ 3452 pmap_remove_all(m); 3453 if (clear_modify) 3454 vfs_page_set_valid(bp, foff, i, m); 3455 else if (m->valid == VM_PAGE_BITS_ALL && 3456 (bp->b_flags & B_CACHE) == 0) { 3457 bp->b_pages[i] = bogus_page; 3458 bogus++; 3459 } 3460 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3461 } 3462 vm_page_unlock_queues(); 3463 VM_OBJECT_UNLOCK(obj); 3464 if (bogus) 3465 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3466 bp->b_pages, bp->b_npages); 3467} 3468 3469/* 3470 * Tell the VM system that the pages associated with this buffer 3471 * are clean. This is used for delayed writes where the data is 3472 * going to go to disk eventually without additional VM intevention. 3473 * 3474 * Note that while we only really need to clean through to b_bcount, we 3475 * just go ahead and clean through to b_bufsize. 3476 */ 3477static void 3478vfs_clean_pages(struct buf *bp) 3479{ 3480 int i; 3481 vm_ooffset_t foff, noff, eoff; 3482 vm_page_t m; 3483 3484 if (!(bp->b_flags & B_VMIO)) 3485 return; 3486 3487 foff = bp->b_offset; 3488 KASSERT(bp->b_offset != NOOFFSET, 3489 ("vfs_clean_pages: no buffer offset")); 3490 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3491 vm_page_lock_queues(); 3492 for (i = 0; i < bp->b_npages; i++) { 3493 m = bp->b_pages[i]; 3494 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3495 eoff = noff; 3496 3497 if (eoff > bp->b_offset + bp->b_bufsize) 3498 eoff = bp->b_offset + bp->b_bufsize; 3499 vfs_page_set_valid(bp, foff, i, m); 3500 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3501 foff = noff; 3502 } 3503 vm_page_unlock_queues(); 3504 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3505} 3506 3507/* 3508 * vfs_bio_set_validclean: 3509 * 3510 * Set the range within the buffer to valid and clean. The range is 3511 * relative to the beginning of the buffer, b_offset. Note that b_offset 3512 * itself may be offset from the beginning of the first page. 3513 * 3514 */ 3515 3516void 3517vfs_bio_set_validclean(struct buf *bp, int base, int size) 3518{ 3519 int i, n; 3520 vm_page_t m; 3521 3522 if (!(bp->b_flags & B_VMIO)) 3523 return; 3524 3525 /* 3526 * Fixup base to be relative to beginning of first page. 3527 * Set initial n to be the maximum number of bytes in the 3528 * first page that can be validated. 3529 */ 3530 3531 base += (bp->b_offset & PAGE_MASK); 3532 n = PAGE_SIZE - (base & PAGE_MASK); 3533 3534 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3535 vm_page_lock_queues(); 3536 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3537 m = bp->b_pages[i]; 3538 3539 if (n > size) 3540 n = size; 3541 3542 vm_page_set_validclean(m, base & PAGE_MASK, n); 3543 base += n; 3544 size -= n; 3545 n = PAGE_SIZE; 3546 } 3547 vm_page_unlock_queues(); 3548 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3549} 3550 3551/* 3552 * vfs_bio_clrbuf: 3553 * 3554 * clear a buffer. This routine essentially fakes an I/O, so we need 3555 * to clear BIO_ERROR and B_INVAL. 3556 * 3557 * Note that while we only theoretically need to clear through b_bcount, 3558 * we go ahead and clear through b_bufsize. 3559 */ 3560 3561void 3562vfs_bio_clrbuf(struct buf *bp) 3563{ 3564 int i, j, mask = 0; 3565 caddr_t sa, ea; 3566 3567 GIANT_REQUIRED; 3568 3569 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3570 clrbuf(bp); 3571 return; 3572 } 3573 bp->b_flags &= ~B_INVAL; 3574 bp->b_ioflags &= ~BIO_ERROR; 3575 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3576 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3577 (bp->b_offset & PAGE_MASK) == 0) { 3578 if (bp->b_pages[0] == bogus_page) 3579 goto unlock; 3580 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3581 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3582 if ((bp->b_pages[0]->valid & mask) == mask) 3583 goto unlock; 3584 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3585 ((bp->b_pages[0]->valid & mask) == 0)) { 3586 bzero(bp->b_data, bp->b_bufsize); 3587 bp->b_pages[0]->valid |= mask; 3588 goto unlock; 3589 } 3590 } 3591 ea = sa = bp->b_data; 3592 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3593 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3594 ea = (caddr_t)(vm_offset_t)ulmin( 3595 (u_long)(vm_offset_t)ea, 3596 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3597 if (bp->b_pages[i] == bogus_page) 3598 continue; 3599 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3600 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3601 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3602 if ((bp->b_pages[i]->valid & mask) == mask) 3603 continue; 3604 if ((bp->b_pages[i]->valid & mask) == 0) { 3605 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3606 bzero(sa, ea - sa); 3607 } else { 3608 for (; sa < ea; sa += DEV_BSIZE, j++) { 3609 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3610 (bp->b_pages[i]->valid & (1<<j)) == 0) 3611 bzero(sa, DEV_BSIZE); 3612 } 3613 } 3614 bp->b_pages[i]->valid |= mask; 3615 } 3616unlock: 3617 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3618 bp->b_resid = 0; 3619} 3620 3621/* 3622 * vm_hold_load_pages and vm_hold_free_pages get pages into 3623 * a buffers address space. The pages are anonymous and are 3624 * not associated with a file object. 3625 */ 3626static void 3627vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3628{ 3629 vm_offset_t pg; 3630 vm_page_t p; 3631 int index; 3632 3633 to = round_page(to); 3634 from = round_page(from); 3635 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3636 3637 VM_OBJECT_LOCK(kernel_object); 3638 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3639tryagain: 3640 /* 3641 * note: must allocate system pages since blocking here 3642 * could intefere with paging I/O, no matter which 3643 * process we are. 3644 */ 3645 p = vm_page_alloc(kernel_object, 3646 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3647 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3648 if (!p) { 3649 atomic_add_int(&vm_pageout_deficit, 3650 (to - pg) >> PAGE_SHIFT); 3651 VM_OBJECT_UNLOCK(kernel_object); 3652 VM_WAIT; 3653 VM_OBJECT_LOCK(kernel_object); 3654 goto tryagain; 3655 } 3656 p->valid = VM_PAGE_BITS_ALL; 3657 pmap_qenter(pg, &p, 1); 3658 bp->b_pages[index] = p; 3659 } 3660 VM_OBJECT_UNLOCK(kernel_object); 3661 bp->b_npages = index; 3662} 3663 3664/* Return pages associated with this buf to the vm system */ 3665static void 3666vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3667{ 3668 vm_offset_t pg; 3669 vm_page_t p; 3670 int index, newnpages; 3671 3672 from = round_page(from); 3673 to = round_page(to); 3674 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3675 3676 VM_OBJECT_LOCK(kernel_object); 3677 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3678 p = bp->b_pages[index]; 3679 if (p && (index < bp->b_npages)) { 3680 if (p->busy) { 3681 printf( 3682 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3683 (intmax_t)bp->b_blkno, 3684 (intmax_t)bp->b_lblkno); 3685 } 3686 bp->b_pages[index] = NULL; 3687 pmap_qremove(pg, 1); 3688 vm_page_lock_queues(); 3689 vm_page_unwire(p, 0); 3690 vm_page_free(p); 3691 vm_page_unlock_queues(); 3692 } 3693 } 3694 VM_OBJECT_UNLOCK(kernel_object); 3695 bp->b_npages = newnpages; 3696} 3697 3698/* 3699 * Map an IO request into kernel virtual address space. 3700 * 3701 * All requests are (re)mapped into kernel VA space. 3702 * Notice that we use b_bufsize for the size of the buffer 3703 * to be mapped. b_bcount might be modified by the driver. 3704 * 3705 * Note that even if the caller determines that the address space should 3706 * be valid, a race or a smaller-file mapped into a larger space may 3707 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3708 * check the return value. 3709 */ 3710int 3711vmapbuf(struct buf *bp) 3712{ 3713 caddr_t addr, kva; 3714 vm_prot_t prot; 3715 int pidx, i; 3716 struct vm_page *m; 3717 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3718 3719 if (bp->b_bufsize < 0) 3720 return (-1); 3721 prot = VM_PROT_READ; 3722 if (bp->b_iocmd == BIO_READ) 3723 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3724 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3725 addr < bp->b_data + bp->b_bufsize; 3726 addr += PAGE_SIZE, pidx++) { 3727 /* 3728 * Do the vm_fault if needed; do the copy-on-write thing 3729 * when reading stuff off device into memory. 3730 * 3731 * NOTE! Must use pmap_extract() because addr may be in 3732 * the userland address space, and kextract is only guarenteed 3733 * to work for the kernland address space (see: sparc64 port). 3734 */ 3735retry: 3736 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3737 prot) < 0) { 3738 vm_page_lock_queues(); 3739 for (i = 0; i < pidx; ++i) { 3740 vm_page_unhold(bp->b_pages[i]); 3741 bp->b_pages[i] = NULL; 3742 } 3743 vm_page_unlock_queues(); 3744 return(-1); 3745 } 3746 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3747 if (m == NULL) 3748 goto retry; 3749 bp->b_pages[pidx] = m; 3750 } 3751 if (pidx > btoc(MAXPHYS)) 3752 panic("vmapbuf: mapped more than MAXPHYS"); 3753 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3754 3755 kva = bp->b_saveaddr; 3756 bp->b_npages = pidx; 3757 bp->b_saveaddr = bp->b_data; 3758 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3759 return(0); 3760} 3761 3762/* 3763 * Free the io map PTEs associated with this IO operation. 3764 * We also invalidate the TLB entries and restore the original b_addr. 3765 */ 3766void 3767vunmapbuf(struct buf *bp) 3768{ 3769 int pidx; 3770 int npages; 3771 3772 npages = bp->b_npages; 3773 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3774 vm_page_lock_queues(); 3775 for (pidx = 0; pidx < npages; pidx++) 3776 vm_page_unhold(bp->b_pages[pidx]); 3777 vm_page_unlock_queues(); 3778 3779 bp->b_data = bp->b_saveaddr; 3780} 3781 3782void 3783bdone(struct buf *bp) 3784{ 3785 3786 mtx_lock(&bdonelock); 3787 bp->b_flags |= B_DONE; 3788 wakeup(bp); 3789 mtx_unlock(&bdonelock); 3790} 3791 3792void 3793bwait(struct buf *bp, u_char pri, const char *wchan) 3794{ 3795 3796 mtx_lock(&bdonelock); 3797 while ((bp->b_flags & B_DONE) == 0) 3798 msleep(bp, &bdonelock, pri, wchan, 0); 3799 mtx_unlock(&bdonelock); 3800} 3801 3802void 3803bufstrategy(struct bufobj *bo, struct buf *bp) 3804{ 3805 int i = 0; 3806 struct vnode *vp; 3807 3808 vp = bp->b_vp; 3809 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3810 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3811 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3812 i = VOP_STRATEGY(vp, bp); 3813 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3814} 3815 3816void 3817bufobj_wref(struct bufobj *bo) 3818{ 3819 3820 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3821 BO_LOCK(bo); 3822 bo->bo_numoutput++; 3823 BO_UNLOCK(bo); 3824} 3825 3826void 3827bufobj_wdrop(struct bufobj *bo) 3828{ 3829 3830 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3831 BO_LOCK(bo); 3832 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3833 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3834 bo->bo_flag &= ~BO_WWAIT; 3835 wakeup(&bo->bo_numoutput); 3836 } 3837 BO_UNLOCK(bo); 3838} 3839 3840int 3841bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3842{ 3843 int error; 3844 3845 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3846 ASSERT_BO_LOCKED(bo); 3847 error = 0; 3848 while (bo->bo_numoutput) { 3849 bo->bo_flag |= BO_WWAIT; 3850 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3851 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3852 if (error) 3853 break; 3854 } 3855 return (error); 3856} 3857 3858#include "opt_ddb.h" 3859#ifdef DDB 3860#include <ddb/ddb.h> 3861 3862/* DDB command to show buffer data */ 3863DB_SHOW_COMMAND(buffer, db_show_buffer) 3864{ 3865 /* get args */ 3866 struct buf *bp = (struct buf *)addr; 3867 3868 if (!have_addr) { 3869 db_printf("usage: show buffer <addr>\n"); 3870 return; 3871 } 3872 3873 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3874 db_printf( 3875 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3876 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd\n", 3877 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3878 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3879 (intmax_t)bp->b_blkno); 3880 if (bp->b_npages) { 3881 int i; 3882 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3883 for (i = 0; i < bp->b_npages; i++) { 3884 vm_page_t m; 3885 m = bp->b_pages[i]; 3886 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3887 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3888 if ((i + 1) < bp->b_npages) 3889 db_printf(","); 3890 } 3891 db_printf("\n"); 3892 } 3893} 3894#endif /* DDB */ 3895