vfs_bio.c revision 137042
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 137042 2004-10-29 10:52:31Z phk $"); 43 44#include <sys/param.h> 45#include <sys/systm.h> 46#include <sys/bio.h> 47#include <sys/conf.h> 48#include <sys/buf.h> 49#include <sys/devicestat.h> 50#include <sys/eventhandler.h> 51#include <sys/lock.h> 52#include <sys/malloc.h> 53#include <sys/mount.h> 54#include <sys/mutex.h> 55#include <sys/kernel.h> 56#include <sys/kthread.h> 57#include <sys/proc.h> 58#include <sys/resourcevar.h> 59#include <sys/sysctl.h> 60#include <sys/vmmeter.h> 61#include <sys/vnode.h> 62#include <geom/geom.h> 63#include <vm/vm.h> 64#include <vm/vm_param.h> 65#include <vm/vm_kern.h> 66#include <vm/vm_pageout.h> 67#include <vm/vm_page.h> 68#include <vm/vm_object.h> 69#include <vm/vm_extern.h> 70#include <vm/vm_map.h> 71#include "opt_directio.h" 72#include "opt_swap.h" 73 74static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 75 76struct bio_ops bioops; /* I/O operation notification */ 77 78struct buf_ops buf_ops_bio = { 79 .bop_name = "buf_ops_bio", 80 .bop_write = bufwrite, 81 .bop_strategy = bufstrategy, 82}; 83 84/* 85 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 86 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 87 */ 88struct buf *buf; /* buffer header pool */ 89 90static struct proc *bufdaemonproc; 91 92static int inmem(struct vnode * vp, daddr_t blkno); 93static void vm_hold_free_pages(struct buf *bp, vm_offset_t from, 94 vm_offset_t to); 95static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 96 vm_offset_t to); 97static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 98 int pageno, vm_page_t m); 99static void vfs_clean_pages(struct buf *bp); 100static void vfs_setdirty(struct buf *bp); 101static void vfs_vmio_release(struct buf *bp); 102static void vfs_backgroundwritedone(struct buf *bp); 103static int vfs_bio_clcheck(struct vnode *vp, int size, 104 daddr_t lblkno, daddr_t blkno); 105static int flushbufqueues(int flushdeps); 106static void buf_daemon(void); 107void bremfreel(struct buf *bp); 108 109int vmiodirenable = TRUE; 110SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 111 "Use the VM system for directory writes"); 112int runningbufspace; 113SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 114 "Amount of presently outstanding async buffer io"); 115static int bufspace; 116SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 117 "KVA memory used for bufs"); 118static int maxbufspace; 119SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 120 "Maximum allowed value of bufspace (including buf_daemon)"); 121static int bufmallocspace; 122SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 123 "Amount of malloced memory for buffers"); 124static int maxbufmallocspace; 125SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 126 "Maximum amount of malloced memory for buffers"); 127static int lobufspace; 128SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 129 "Minimum amount of buffers we want to have"); 130static int hibufspace; 131SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 132 "Maximum allowed value of bufspace (excluding buf_daemon)"); 133static int bufreusecnt; 134SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 135 "Number of times we have reused a buffer"); 136static int buffreekvacnt; 137SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 138 "Number of times we have freed the KVA space from some buffer"); 139static int bufdefragcnt; 140SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 141 "Number of times we have had to repeat buffer allocation to defragment"); 142static int lorunningspace; 143SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 144 "Minimum preferred space used for in-progress I/O"); 145static int hirunningspace; 146SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 147 "Maximum amount of space to use for in-progress I/O"); 148static int dirtybufferflushes; 149SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 150 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 151static int altbufferflushes; 152SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 153 0, "Number of fsync flushes to limit dirty buffers"); 154static int recursiveflushes; 155SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 156 0, "Number of flushes skipped due to being recursive"); 157static int numdirtybuffers; 158SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 159 "Number of buffers that are dirty (has unwritten changes) at the moment"); 160static int lodirtybuffers; 161SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 162 "How many buffers we want to have free before bufdaemon can sleep"); 163static int hidirtybuffers; 164SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 165 "When the number of dirty buffers is considered severe"); 166static int dirtybufthresh; 167SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 168 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 169static int numfreebuffers; 170SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 171 "Number of free buffers"); 172static int lofreebuffers; 173SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 174 "XXX Unused"); 175static int hifreebuffers; 176SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 177 "XXX Complicatedly unused"); 178static int getnewbufcalls; 179SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 180 "Number of calls to getnewbuf"); 181static int getnewbufrestarts; 182SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 183 "Number of times getnewbuf has had to restart a buffer aquisition"); 184static int dobkgrdwrite = 1; 185SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 186 "Do background writes (honoring the BV_BKGRDWRITE flag)?"); 187 188/* 189 * Wakeup point for bufdaemon, as well as indicator of whether it is already 190 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 191 * is idling. 192 */ 193static int bd_request; 194 195/* 196 * This lock synchronizes access to bd_request. 197 */ 198static struct mtx bdlock; 199 200/* 201 * bogus page -- for I/O to/from partially complete buffers 202 * this is a temporary solution to the problem, but it is not 203 * really that bad. it would be better to split the buffer 204 * for input in the case of buffers partially already in memory, 205 * but the code is intricate enough already. 206 */ 207vm_page_t bogus_page; 208 209/* 210 * Synchronization (sleep/wakeup) variable for active buffer space requests. 211 * Set when wait starts, cleared prior to wakeup(). 212 * Used in runningbufwakeup() and waitrunningbufspace(). 213 */ 214static int runningbufreq; 215 216/* 217 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 218 * waitrunningbufspace(). 219 */ 220static struct mtx rbreqlock; 221 222/* 223 * Synchronization (sleep/wakeup) variable for buffer requests. 224 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 225 * by and/or. 226 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 227 * getnewbuf(), and getblk(). 228 */ 229static int needsbuffer; 230 231/* 232 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 233 */ 234static struct mtx nblock; 235 236/* 237 * Lock that protects against bwait()/bdone()/B_DONE races. 238 */ 239 240static struct mtx bdonelock; 241 242/* 243 * Definitions for the buffer free lists. 244 */ 245#define BUFFER_QUEUES 5 /* number of free buffer queues */ 246 247#define QUEUE_NONE 0 /* on no queue */ 248#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 249#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 250#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 251#define QUEUE_EMPTY 4 /* empty buffer headers */ 252 253/* Queues for free buffers with various properties */ 254static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 255 256/* Lock for the bufqueues */ 257static struct mtx bqlock; 258 259/* 260 * Single global constant for BUF_WMESG, to avoid getting multiple references. 261 * buf_wmesg is referred from macros. 262 */ 263const char *buf_wmesg = BUF_WMESG; 264 265#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 266#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 267#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 268#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 269 270#ifdef DIRECTIO 271extern void ffs_rawread_setup(void); 272#endif /* DIRECTIO */ 273/* 274 * numdirtywakeup: 275 * 276 * If someone is blocked due to there being too many dirty buffers, 277 * and numdirtybuffers is now reasonable, wake them up. 278 */ 279 280static __inline void 281numdirtywakeup(int level) 282{ 283 284 if (numdirtybuffers <= level) { 285 mtx_lock(&nblock); 286 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 287 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 288 wakeup(&needsbuffer); 289 } 290 mtx_unlock(&nblock); 291 } 292} 293 294/* 295 * bufspacewakeup: 296 * 297 * Called when buffer space is potentially available for recovery. 298 * getnewbuf() will block on this flag when it is unable to free 299 * sufficient buffer space. Buffer space becomes recoverable when 300 * bp's get placed back in the queues. 301 */ 302 303static __inline void 304bufspacewakeup(void) 305{ 306 307 /* 308 * If someone is waiting for BUF space, wake them up. Even 309 * though we haven't freed the kva space yet, the waiting 310 * process will be able to now. 311 */ 312 mtx_lock(&nblock); 313 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 314 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 315 wakeup(&needsbuffer); 316 } 317 mtx_unlock(&nblock); 318} 319 320/* 321 * runningbufwakeup() - in-progress I/O accounting. 322 * 323 */ 324static __inline void 325runningbufwakeup(struct buf *bp) 326{ 327 328 if (bp->b_runningbufspace) { 329 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 330 bp->b_runningbufspace = 0; 331 mtx_lock(&rbreqlock); 332 if (runningbufreq && runningbufspace <= lorunningspace) { 333 runningbufreq = 0; 334 wakeup(&runningbufreq); 335 } 336 mtx_unlock(&rbreqlock); 337 } 338} 339 340/* 341 * bufcountwakeup: 342 * 343 * Called when a buffer has been added to one of the free queues to 344 * account for the buffer and to wakeup anyone waiting for free buffers. 345 * This typically occurs when large amounts of metadata are being handled 346 * by the buffer cache ( else buffer space runs out first, usually ). 347 */ 348 349static __inline void 350bufcountwakeup(void) 351{ 352 353 atomic_add_int(&numfreebuffers, 1); 354 mtx_lock(&nblock); 355 if (needsbuffer) { 356 needsbuffer &= ~VFS_BIO_NEED_ANY; 357 if (numfreebuffers >= hifreebuffers) 358 needsbuffer &= ~VFS_BIO_NEED_FREE; 359 wakeup(&needsbuffer); 360 } 361 mtx_unlock(&nblock); 362} 363 364/* 365 * waitrunningbufspace() 366 * 367 * runningbufspace is a measure of the amount of I/O currently 368 * running. This routine is used in async-write situations to 369 * prevent creating huge backups of pending writes to a device. 370 * Only asynchronous writes are governed by this function. 371 * 372 * Reads will adjust runningbufspace, but will not block based on it. 373 * The read load has a side effect of reducing the allowed write load. 374 * 375 * This does NOT turn an async write into a sync write. It waits 376 * for earlier writes to complete and generally returns before the 377 * caller's write has reached the device. 378 */ 379static __inline void 380waitrunningbufspace(void) 381{ 382 383 mtx_lock(&rbreqlock); 384 while (runningbufspace > hirunningspace) { 385 ++runningbufreq; 386 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 387 } 388 mtx_unlock(&rbreqlock); 389} 390 391 392/* 393 * vfs_buf_test_cache: 394 * 395 * Called when a buffer is extended. This function clears the B_CACHE 396 * bit if the newly extended portion of the buffer does not contain 397 * valid data. 398 */ 399static __inline 400void 401vfs_buf_test_cache(struct buf *bp, 402 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 403 vm_page_t m) 404{ 405 406 GIANT_REQUIRED; 407 408 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 409 if (bp->b_flags & B_CACHE) { 410 int base = (foff + off) & PAGE_MASK; 411 if (vm_page_is_valid(m, base, size) == 0) 412 bp->b_flags &= ~B_CACHE; 413 } 414} 415 416/* Wake up the buffer deamon if necessary */ 417static __inline 418void 419bd_wakeup(int dirtybuflevel) 420{ 421 422 mtx_lock(&bdlock); 423 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 424 bd_request = 1; 425 wakeup(&bd_request); 426 } 427 mtx_unlock(&bdlock); 428} 429 430/* 431 * bd_speedup - speedup the buffer cache flushing code 432 */ 433 434static __inline 435void 436bd_speedup(void) 437{ 438 439 bd_wakeup(1); 440} 441 442/* 443 * Calculating buffer cache scaling values and reserve space for buffer 444 * headers. This is called during low level kernel initialization and 445 * may be called more then once. We CANNOT write to the memory area 446 * being reserved at this time. 447 */ 448caddr_t 449kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 450{ 451 452 /* 453 * physmem_est is in pages. Convert it to kilobytes (assumes 454 * PAGE_SIZE is >= 1K) 455 */ 456 physmem_est = physmem_est * (PAGE_SIZE / 1024); 457 458 /* 459 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 460 * For the first 64MB of ram nominally allocate sufficient buffers to 461 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 462 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 463 * the buffer cache we limit the eventual kva reservation to 464 * maxbcache bytes. 465 * 466 * factor represents the 1/4 x ram conversion. 467 */ 468 if (nbuf == 0) { 469 int factor = 4 * BKVASIZE / 1024; 470 471 nbuf = 50; 472 if (physmem_est > 4096) 473 nbuf += min((physmem_est - 4096) / factor, 474 65536 / factor); 475 if (physmem_est > 65536) 476 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 477 478 if (maxbcache && nbuf > maxbcache / BKVASIZE) 479 nbuf = maxbcache / BKVASIZE; 480 } 481 482#if 0 483 /* 484 * Do not allow the buffer_map to be more then 1/2 the size of the 485 * kernel_map. 486 */ 487 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 488 (BKVASIZE * 2)) { 489 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 490 (BKVASIZE * 2); 491 printf("Warning: nbufs capped at %d\n", nbuf); 492 } 493#endif 494 495 /* 496 * swbufs are used as temporary holders for I/O, such as paging I/O. 497 * We have no less then 16 and no more then 256. 498 */ 499 nswbuf = max(min(nbuf/4, 256), 16); 500#ifdef NSWBUF_MIN 501 if (nswbuf < NSWBUF_MIN) 502 nswbuf = NSWBUF_MIN; 503#endif 504#ifdef DIRECTIO 505 ffs_rawread_setup(); 506#endif 507 508 /* 509 * Reserve space for the buffer cache buffers 510 */ 511 swbuf = (void *)v; 512 v = (caddr_t)(swbuf + nswbuf); 513 buf = (void *)v; 514 v = (caddr_t)(buf + nbuf); 515 516 return(v); 517} 518 519/* Initialize the buffer subsystem. Called before use of any buffers. */ 520void 521bufinit(void) 522{ 523 struct buf *bp; 524 int i; 525 526 GIANT_REQUIRED; 527 528 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 529 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 530 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 531 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 532 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 533 534 /* next, make a null set of free lists */ 535 for (i = 0; i < BUFFER_QUEUES; i++) 536 TAILQ_INIT(&bufqueues[i]); 537 538 /* finally, initialize each buffer header and stick on empty q */ 539 for (i = 0; i < nbuf; i++) { 540 bp = &buf[i]; 541 bzero(bp, sizeof *bp); 542 bp->b_flags = B_INVAL; /* we're just an empty header */ 543 bp->b_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 vm_page_busy(m); 1599 pmap_remove_all(m); 1600 vm_page_free(m); 1601 } else if (bp->b_flags & B_DIRECT) { 1602 vm_page_try_to_free(m); 1603 } else if (vm_page_count_severe()) { 1604 vm_page_try_to_cache(m); 1605 } 1606 } 1607 } 1608 vm_page_unlock_queues(); 1609 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1610 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1611 1612 if (bp->b_bufsize) { 1613 bufspacewakeup(); 1614 bp->b_bufsize = 0; 1615 } 1616 bp->b_npages = 0; 1617 bp->b_flags &= ~B_VMIO; 1618 if (bp->b_vp) 1619 brelvp(bp); 1620} 1621 1622/* 1623 * Check to see if a block at a particular lbn is available for a clustered 1624 * write. 1625 */ 1626static int 1627vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1628{ 1629 struct buf *bpa; 1630 int match; 1631 1632 match = 0; 1633 1634 /* If the buf isn't in core skip it */ 1635 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1636 return (0); 1637 1638 /* If the buf is busy we don't want to wait for it */ 1639 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1640 return (0); 1641 1642 /* Only cluster with valid clusterable delayed write buffers */ 1643 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1644 (B_DELWRI | B_CLUSTEROK)) 1645 goto done; 1646 1647 if (bpa->b_bufsize != size) 1648 goto done; 1649 1650 /* 1651 * Check to see if it is in the expected place on disk and that the 1652 * block has been mapped. 1653 */ 1654 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1655 match = 1; 1656done: 1657 BUF_UNLOCK(bpa); 1658 return (match); 1659} 1660 1661/* 1662 * vfs_bio_awrite: 1663 * 1664 * Implement clustered async writes for clearing out B_DELWRI buffers. 1665 * This is much better then the old way of writing only one buffer at 1666 * a time. Note that we may not be presented with the buffers in the 1667 * correct order, so we search for the cluster in both directions. 1668 */ 1669int 1670vfs_bio_awrite(struct buf *bp) 1671{ 1672 int i; 1673 int j; 1674 daddr_t lblkno = bp->b_lblkno; 1675 struct vnode *vp = bp->b_vp; 1676 int s; 1677 int ncl; 1678 int nwritten; 1679 int size; 1680 int maxcl; 1681 1682 s = splbio(); 1683 /* 1684 * right now we support clustered writing only to regular files. If 1685 * we find a clusterable block we could be in the middle of a cluster 1686 * rather then at the beginning. 1687 */ 1688 if ((vp->v_type == VREG) && 1689 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1690 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1691 1692 size = vp->v_mount->mnt_stat.f_iosize; 1693 maxcl = MAXPHYS / size; 1694 1695 VI_LOCK(vp); 1696 for (i = 1; i < maxcl; i++) 1697 if (vfs_bio_clcheck(vp, size, lblkno + i, 1698 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1699 break; 1700 1701 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1702 if (vfs_bio_clcheck(vp, size, lblkno - j, 1703 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1704 break; 1705 1706 VI_UNLOCK(vp); 1707 --j; 1708 ncl = i + j; 1709 /* 1710 * this is a possible cluster write 1711 */ 1712 if (ncl != 1) { 1713 BUF_UNLOCK(bp); 1714 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1715 splx(s); 1716 return nwritten; 1717 } 1718 } 1719 1720 bremfree(bp); 1721 bp->b_flags |= B_ASYNC; 1722 1723 splx(s); 1724 /* 1725 * default (old) behavior, writing out only one block 1726 * 1727 * XXX returns b_bufsize instead of b_bcount for nwritten? 1728 */ 1729 nwritten = bp->b_bufsize; 1730 (void) bwrite(bp); 1731 1732 return nwritten; 1733} 1734 1735/* 1736 * getnewbuf: 1737 * 1738 * Find and initialize a new buffer header, freeing up existing buffers 1739 * in the bufqueues as necessary. The new buffer is returned locked. 1740 * 1741 * Important: B_INVAL is not set. If the caller wishes to throw the 1742 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1743 * 1744 * We block if: 1745 * We have insufficient buffer headers 1746 * We have insufficient buffer space 1747 * buffer_map is too fragmented ( space reservation fails ) 1748 * If we have to flush dirty buffers ( but we try to avoid this ) 1749 * 1750 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1751 * Instead we ask the buf daemon to do it for us. We attempt to 1752 * avoid piecemeal wakeups of the pageout daemon. 1753 */ 1754 1755static struct buf * 1756getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1757{ 1758 struct buf *bp; 1759 struct buf *nbp; 1760 int defrag = 0; 1761 int nqindex; 1762 static int flushingbufs; 1763 1764 GIANT_REQUIRED; 1765 1766 /* 1767 * We can't afford to block since we might be holding a vnode lock, 1768 * which may prevent system daemons from running. We deal with 1769 * low-memory situations by proactively returning memory and running 1770 * async I/O rather then sync I/O. 1771 */ 1772 1773 atomic_add_int(&getnewbufcalls, 1); 1774 atomic_subtract_int(&getnewbufrestarts, 1); 1775restart: 1776 atomic_add_int(&getnewbufrestarts, 1); 1777 1778 /* 1779 * Setup for scan. If we do not have enough free buffers, 1780 * we setup a degenerate case that immediately fails. Note 1781 * that if we are specially marked process, we are allowed to 1782 * dip into our reserves. 1783 * 1784 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1785 * 1786 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1787 * However, there are a number of cases (defragging, reusing, ...) 1788 * where we cannot backup. 1789 */ 1790 mtx_lock(&bqlock); 1791 nqindex = QUEUE_EMPTYKVA; 1792 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1793 1794 if (nbp == NULL) { 1795 /* 1796 * If no EMPTYKVA buffers and we are either 1797 * defragging or reusing, locate a CLEAN buffer 1798 * to free or reuse. If bufspace useage is low 1799 * skip this step so we can allocate a new buffer. 1800 */ 1801 if (defrag || bufspace >= lobufspace) { 1802 nqindex = QUEUE_CLEAN; 1803 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1804 } 1805 1806 /* 1807 * If we could not find or were not allowed to reuse a 1808 * CLEAN buffer, check to see if it is ok to use an EMPTY 1809 * buffer. We can only use an EMPTY buffer if allocating 1810 * its KVA would not otherwise run us out of buffer space. 1811 */ 1812 if (nbp == NULL && defrag == 0 && 1813 bufspace + maxsize < hibufspace) { 1814 nqindex = QUEUE_EMPTY; 1815 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1816 } 1817 } 1818 1819 /* 1820 * Run scan, possibly freeing data and/or kva mappings on the fly 1821 * depending. 1822 */ 1823 1824 while ((bp = nbp) != NULL) { 1825 int qindex = nqindex; 1826 1827 /* 1828 * Calculate next bp ( we can only use it if we do not block 1829 * or do other fancy things ). 1830 */ 1831 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1832 switch(qindex) { 1833 case QUEUE_EMPTY: 1834 nqindex = QUEUE_EMPTYKVA; 1835 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1836 break; 1837 /* FALLTHROUGH */ 1838 case QUEUE_EMPTYKVA: 1839 nqindex = QUEUE_CLEAN; 1840 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1841 break; 1842 /* FALLTHROUGH */ 1843 case QUEUE_CLEAN: 1844 /* 1845 * nbp is NULL. 1846 */ 1847 break; 1848 } 1849 } 1850 if (bp->b_vp) { 1851 BO_LOCK(bp->b_bufobj); 1852 if (bp->b_vflags & BV_BKGRDINPROG) { 1853 BO_UNLOCK(bp->b_bufobj); 1854 continue; 1855 } 1856 BO_UNLOCK(bp->b_bufobj); 1857 } 1858 1859 /* 1860 * Sanity Checks 1861 */ 1862 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1863 1864 /* 1865 * Note: we no longer distinguish between VMIO and non-VMIO 1866 * buffers. 1867 */ 1868 1869 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1870 1871 /* 1872 * If we are defragging then we need a buffer with 1873 * b_kvasize != 0. XXX this situation should no longer 1874 * occur, if defrag is non-zero the buffer's b_kvasize 1875 * should also be non-zero at this point. XXX 1876 */ 1877 if (defrag && bp->b_kvasize == 0) { 1878 printf("Warning: defrag empty buffer %p\n", bp); 1879 continue; 1880 } 1881 1882 /* 1883 * Start freeing the bp. This is somewhat involved. nbp 1884 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1885 */ 1886 1887 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1888 panic("getnewbuf: locked buf"); 1889 bremfreel(bp); 1890 mtx_unlock(&bqlock); 1891 1892 if (qindex == QUEUE_CLEAN) { 1893 if (bp->b_flags & B_VMIO) { 1894 bp->b_flags &= ~B_ASYNC; 1895 vfs_vmio_release(bp); 1896 } 1897 if (bp->b_vp) 1898 brelvp(bp); 1899 } 1900 1901 /* 1902 * NOTE: nbp is now entirely invalid. We can only restart 1903 * the scan from this point on. 1904 * 1905 * Get the rest of the buffer freed up. b_kva* is still 1906 * valid after this operation. 1907 */ 1908 1909 if (bp->b_rcred != NOCRED) { 1910 crfree(bp->b_rcred); 1911 bp->b_rcred = NOCRED; 1912 } 1913 if (bp->b_wcred != NOCRED) { 1914 crfree(bp->b_wcred); 1915 bp->b_wcred = NOCRED; 1916 } 1917 if (LIST_FIRST(&bp->b_dep) != NULL) 1918 buf_deallocate(bp); 1919 if (bp->b_vflags & BV_BKGRDINPROG) 1920 panic("losing buffer 3"); 1921 1922 if (bp->b_bufsize) 1923 allocbuf(bp, 0); 1924 1925 bp->b_flags = 0; 1926 bp->b_ioflags = 0; 1927 bp->b_xflags = 0; 1928 bp->b_vflags = 0; 1929 bp->b_dev = NULL; 1930 bp->b_vp = NULL; 1931 bp->b_blkno = bp->b_lblkno = 0; 1932 bp->b_offset = NOOFFSET; 1933 bp->b_iodone = 0; 1934 bp->b_error = 0; 1935 bp->b_resid = 0; 1936 bp->b_bcount = 0; 1937 bp->b_npages = 0; 1938 bp->b_dirtyoff = bp->b_dirtyend = 0; 1939 bp->b_magic = B_MAGIC_BIO; 1940 bp->b_bufobj = NULL; 1941 1942 LIST_INIT(&bp->b_dep); 1943 1944 /* 1945 * If we are defragging then free the buffer. 1946 */ 1947 if (defrag) { 1948 bp->b_flags |= B_INVAL; 1949 bfreekva(bp); 1950 brelse(bp); 1951 defrag = 0; 1952 goto restart; 1953 } 1954 1955 /* 1956 * If we are overcomitted then recover the buffer and its 1957 * KVM space. This occurs in rare situations when multiple 1958 * processes are blocked in getnewbuf() or allocbuf(). 1959 */ 1960 if (bufspace >= hibufspace) 1961 flushingbufs = 1; 1962 if (flushingbufs && bp->b_kvasize != 0) { 1963 bp->b_flags |= B_INVAL; 1964 bfreekva(bp); 1965 brelse(bp); 1966 goto restart; 1967 } 1968 if (bufspace < lobufspace) 1969 flushingbufs = 0; 1970 break; 1971 } 1972 1973 /* 1974 * If we exhausted our list, sleep as appropriate. We may have to 1975 * wakeup various daemons and write out some dirty buffers. 1976 * 1977 * Generally we are sleeping due to insufficient buffer space. 1978 */ 1979 1980 if (bp == NULL) { 1981 int flags; 1982 char *waitmsg; 1983 1984 mtx_unlock(&bqlock); 1985 if (defrag) { 1986 flags = VFS_BIO_NEED_BUFSPACE; 1987 waitmsg = "nbufkv"; 1988 } else if (bufspace >= hibufspace) { 1989 waitmsg = "nbufbs"; 1990 flags = VFS_BIO_NEED_BUFSPACE; 1991 } else { 1992 waitmsg = "newbuf"; 1993 flags = VFS_BIO_NEED_ANY; 1994 } 1995 1996 bd_speedup(); /* heeeelp */ 1997 1998 mtx_lock(&nblock); 1999 needsbuffer |= flags; 2000 while (needsbuffer & flags) { 2001 if (msleep(&needsbuffer, &nblock, 2002 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2003 mtx_unlock(&nblock); 2004 return (NULL); 2005 } 2006 } 2007 mtx_unlock(&nblock); 2008 } else { 2009 /* 2010 * We finally have a valid bp. We aren't quite out of the 2011 * woods, we still have to reserve kva space. In order 2012 * to keep fragmentation sane we only allocate kva in 2013 * BKVASIZE chunks. 2014 */ 2015 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2016 2017 if (maxsize != bp->b_kvasize) { 2018 vm_offset_t addr = 0; 2019 2020 bfreekva(bp); 2021 2022 if (vm_map_findspace(buffer_map, 2023 vm_map_min(buffer_map), maxsize, &addr)) { 2024 /* 2025 * Uh oh. Buffer map is to fragmented. We 2026 * must defragment the map. 2027 */ 2028 atomic_add_int(&bufdefragcnt, 1); 2029 defrag = 1; 2030 bp->b_flags |= B_INVAL; 2031 brelse(bp); 2032 goto restart; 2033 } 2034 if (addr) { 2035 vm_map_insert(buffer_map, NULL, 0, 2036 addr, addr + maxsize, 2037 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2038 2039 bp->b_kvabase = (caddr_t) addr; 2040 bp->b_kvasize = maxsize; 2041 atomic_add_int(&bufspace, bp->b_kvasize); 2042 atomic_add_int(&bufreusecnt, 1); 2043 } 2044 } 2045 bp->b_saveaddr = bp->b_kvabase; 2046 bp->b_data = bp->b_saveaddr; 2047 } 2048 return(bp); 2049} 2050 2051/* 2052 * buf_daemon: 2053 * 2054 * buffer flushing daemon. Buffers are normally flushed by the 2055 * update daemon but if it cannot keep up this process starts to 2056 * take the load in an attempt to prevent getnewbuf() from blocking. 2057 */ 2058 2059static struct kproc_desc buf_kp = { 2060 "bufdaemon", 2061 buf_daemon, 2062 &bufdaemonproc 2063}; 2064SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2065 2066static void 2067buf_daemon() 2068{ 2069 int s; 2070 2071 mtx_lock(&Giant); 2072 2073 /* 2074 * This process needs to be suspended prior to shutdown sync. 2075 */ 2076 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2077 SHUTDOWN_PRI_LAST); 2078 2079 /* 2080 * This process is allowed to take the buffer cache to the limit 2081 */ 2082 s = splbio(); 2083 mtx_lock(&bdlock); 2084 2085 for (;;) { 2086 bd_request = 0; 2087 mtx_unlock(&bdlock); 2088 2089 kthread_suspend_check(bufdaemonproc); 2090 2091 /* 2092 * Do the flush. Limit the amount of in-transit I/O we 2093 * allow to build up, otherwise we would completely saturate 2094 * the I/O system. Wakeup any waiting processes before we 2095 * normally would so they can run in parallel with our drain. 2096 */ 2097 while (numdirtybuffers > lodirtybuffers) { 2098 if (flushbufqueues(0) == 0) { 2099 /* 2100 * Could not find any buffers without rollback 2101 * dependencies, so just write the first one 2102 * in the hopes of eventually making progress. 2103 */ 2104 flushbufqueues(1); 2105 break; 2106 } 2107 waitrunningbufspace(); 2108 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2109 } 2110 2111 /* 2112 * Only clear bd_request if we have reached our low water 2113 * mark. The buf_daemon normally waits 1 second and 2114 * then incrementally flushes any dirty buffers that have 2115 * built up, within reason. 2116 * 2117 * If we were unable to hit our low water mark and couldn't 2118 * find any flushable buffers, we sleep half a second. 2119 * Otherwise we loop immediately. 2120 */ 2121 mtx_lock(&bdlock); 2122 if (numdirtybuffers <= lodirtybuffers) { 2123 /* 2124 * We reached our low water mark, reset the 2125 * request and sleep until we are needed again. 2126 * The sleep is just so the suspend code works. 2127 */ 2128 bd_request = 0; 2129 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2130 } else { 2131 /* 2132 * We couldn't find any flushable dirty buffers but 2133 * still have too many dirty buffers, we 2134 * have to sleep and try again. (rare) 2135 */ 2136 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2137 } 2138 } 2139} 2140 2141/* 2142 * flushbufqueues: 2143 * 2144 * Try to flush a buffer in the dirty queue. We must be careful to 2145 * free up B_INVAL buffers instead of write them, which NFS is 2146 * particularly sensitive to. 2147 */ 2148int flushwithdeps = 0; 2149SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2150 0, "Number of buffers flushed with dependecies that require rollbacks"); 2151 2152static int 2153flushbufqueues(int flushdeps) 2154{ 2155 struct thread *td = curthread; 2156 struct vnode *vp; 2157 struct mount *mp; 2158 struct buf *bp; 2159 int hasdeps; 2160 2161 mtx_lock(&bqlock); 2162 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2163 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2164 continue; 2165 KASSERT((bp->b_flags & B_DELWRI), 2166 ("unexpected clean buffer %p", bp)); 2167 BO_LOCK(bp->b_bufobj); 2168 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 2169 BO_UNLOCK(bp->b_bufobj); 2170 BUF_UNLOCK(bp); 2171 continue; 2172 } 2173 BO_UNLOCK(bp->b_bufobj); 2174 if (bp->b_flags & B_INVAL) { 2175 bremfreel(bp); 2176 mtx_unlock(&bqlock); 2177 brelse(bp); 2178 return (1); 2179 } 2180 2181 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2182 if (flushdeps == 0) { 2183 BUF_UNLOCK(bp); 2184 continue; 2185 } 2186 hasdeps = 1; 2187 } else 2188 hasdeps = 0; 2189 /* 2190 * We must hold the lock on a vnode before writing 2191 * one of its buffers. Otherwise we may confuse, or 2192 * in the case of a snapshot vnode, deadlock the 2193 * system. 2194 * 2195 * The lock order here is the reverse of the normal 2196 * of vnode followed by buf lock. This is ok because 2197 * the NOWAIT will prevent deadlock. 2198 */ 2199 vp = bp->b_vp; 2200 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2201 BUF_UNLOCK(bp); 2202 continue; 2203 } 2204 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2205 mtx_unlock(&bqlock); 2206 vfs_bio_awrite(bp); 2207 vn_finished_write(mp); 2208 VOP_UNLOCK(vp, 0, td); 2209 flushwithdeps += hasdeps; 2210 return (1); 2211 } 2212 vn_finished_write(mp); 2213 BUF_UNLOCK(bp); 2214 } 2215 mtx_unlock(&bqlock); 2216 return (0); 2217} 2218 2219/* 2220 * Check to see if a block is currently memory resident. 2221 */ 2222struct buf * 2223incore(struct bufobj *bo, daddr_t blkno) 2224{ 2225 struct buf *bp; 2226 2227 int s = splbio(); 2228 BO_LOCK(bo); 2229 bp = gbincore(bo, blkno); 2230 BO_UNLOCK(bo); 2231 splx(s); 2232 return (bp); 2233} 2234 2235/* 2236 * Returns true if no I/O is needed to access the 2237 * associated VM object. This is like incore except 2238 * it also hunts around in the VM system for the data. 2239 */ 2240 2241static int 2242inmem(struct vnode * vp, daddr_t blkno) 2243{ 2244 vm_object_t obj; 2245 vm_offset_t toff, tinc, size; 2246 vm_page_t m; 2247 vm_ooffset_t off; 2248 2249 GIANT_REQUIRED; 2250 ASSERT_VOP_LOCKED(vp, "inmem"); 2251 2252 if (incore(&vp->v_bufobj, blkno)) 2253 return 1; 2254 if (vp->v_mount == NULL) 2255 return 0; 2256 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2257 return 0; 2258 2259 size = PAGE_SIZE; 2260 if (size > vp->v_mount->mnt_stat.f_iosize) 2261 size = vp->v_mount->mnt_stat.f_iosize; 2262 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2263 2264 VM_OBJECT_LOCK(obj); 2265 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2266 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2267 if (!m) 2268 goto notinmem; 2269 tinc = size; 2270 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2271 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2272 if (vm_page_is_valid(m, 2273 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2274 goto notinmem; 2275 } 2276 VM_OBJECT_UNLOCK(obj); 2277 return 1; 2278 2279notinmem: 2280 VM_OBJECT_UNLOCK(obj); 2281 return (0); 2282} 2283 2284/* 2285 * vfs_setdirty: 2286 * 2287 * Sets the dirty range for a buffer based on the status of the dirty 2288 * bits in the pages comprising the buffer. 2289 * 2290 * The range is limited to the size of the buffer. 2291 * 2292 * This routine is primarily used by NFS, but is generalized for the 2293 * B_VMIO case. 2294 */ 2295static void 2296vfs_setdirty(struct buf *bp) 2297{ 2298 int i; 2299 vm_object_t object; 2300 2301 GIANT_REQUIRED; 2302 /* 2303 * Degenerate case - empty buffer 2304 */ 2305 2306 if (bp->b_bufsize == 0) 2307 return; 2308 2309 /* 2310 * We qualify the scan for modified pages on whether the 2311 * object has been flushed yet. The OBJ_WRITEABLE flag 2312 * is not cleared simply by protecting pages off. 2313 */ 2314 2315 if ((bp->b_flags & B_VMIO) == 0) 2316 return; 2317 2318 object = bp->b_pages[0]->object; 2319 VM_OBJECT_LOCK(object); 2320 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2321 printf("Warning: object %p writeable but not mightbedirty\n", object); 2322 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2323 printf("Warning: object %p mightbedirty but not writeable\n", object); 2324 2325 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2326 vm_offset_t boffset; 2327 vm_offset_t eoffset; 2328 2329 vm_page_lock_queues(); 2330 /* 2331 * test the pages to see if they have been modified directly 2332 * by users through the VM system. 2333 */ 2334 for (i = 0; i < bp->b_npages; i++) 2335 vm_page_test_dirty(bp->b_pages[i]); 2336 2337 /* 2338 * Calculate the encompassing dirty range, boffset and eoffset, 2339 * (eoffset - boffset) bytes. 2340 */ 2341 2342 for (i = 0; i < bp->b_npages; i++) { 2343 if (bp->b_pages[i]->dirty) 2344 break; 2345 } 2346 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2347 2348 for (i = bp->b_npages - 1; i >= 0; --i) { 2349 if (bp->b_pages[i]->dirty) { 2350 break; 2351 } 2352 } 2353 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2354 2355 vm_page_unlock_queues(); 2356 /* 2357 * Fit it to the buffer. 2358 */ 2359 2360 if (eoffset > bp->b_bcount) 2361 eoffset = bp->b_bcount; 2362 2363 /* 2364 * If we have a good dirty range, merge with the existing 2365 * dirty range. 2366 */ 2367 2368 if (boffset < eoffset) { 2369 if (bp->b_dirtyoff > boffset) 2370 bp->b_dirtyoff = boffset; 2371 if (bp->b_dirtyend < eoffset) 2372 bp->b_dirtyend = eoffset; 2373 } 2374 } 2375 VM_OBJECT_UNLOCK(object); 2376} 2377 2378/* 2379 * getblk: 2380 * 2381 * Get a block given a specified block and offset into a file/device. 2382 * The buffers B_DONE bit will be cleared on return, making it almost 2383 * ready for an I/O initiation. B_INVAL may or may not be set on 2384 * return. The caller should clear B_INVAL prior to initiating a 2385 * READ. 2386 * 2387 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2388 * an existing buffer. 2389 * 2390 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2391 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2392 * and then cleared based on the backing VM. If the previous buffer is 2393 * non-0-sized but invalid, B_CACHE will be cleared. 2394 * 2395 * If getblk() must create a new buffer, the new buffer is returned with 2396 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2397 * case it is returned with B_INVAL clear and B_CACHE set based on the 2398 * backing VM. 2399 * 2400 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2401 * B_CACHE bit is clear. 2402 * 2403 * What this means, basically, is that the caller should use B_CACHE to 2404 * determine whether the buffer is fully valid or not and should clear 2405 * B_INVAL prior to issuing a read. If the caller intends to validate 2406 * the buffer by loading its data area with something, the caller needs 2407 * to clear B_INVAL. If the caller does this without issuing an I/O, 2408 * the caller should set B_CACHE ( as an optimization ), else the caller 2409 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2410 * a write attempt or if it was a successfull read. If the caller 2411 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2412 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2413 */ 2414struct buf * 2415getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2416 int flags) 2417{ 2418 struct buf *bp; 2419 struct bufobj *bo; 2420 int s; 2421 int error; 2422 ASSERT_VOP_LOCKED(vp, "getblk"); 2423 struct vm_object *vmo; 2424 2425 if (size > MAXBSIZE) 2426 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2427 2428 bo = &vp->v_bufobj; 2429 s = splbio(); 2430loop: 2431 /* 2432 * Block if we are low on buffers. Certain processes are allowed 2433 * to completely exhaust the buffer cache. 2434 * 2435 * If this check ever becomes a bottleneck it may be better to 2436 * move it into the else, when gbincore() fails. At the moment 2437 * it isn't a problem. 2438 * 2439 * XXX remove if 0 sections (clean this up after its proven) 2440 */ 2441 if (numfreebuffers == 0) { 2442 if (curthread == PCPU_GET(idlethread)) 2443 return NULL; 2444 mtx_lock(&nblock); 2445 needsbuffer |= VFS_BIO_NEED_ANY; 2446 mtx_unlock(&nblock); 2447 } 2448 2449 VI_LOCK(vp); 2450 bp = gbincore(bo, blkno); 2451 if (bp != NULL) { 2452 int lockflags; 2453 /* 2454 * Buffer is in-core. If the buffer is not busy, it must 2455 * be on a queue. 2456 */ 2457 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2458 2459 if (flags & GB_LOCK_NOWAIT) 2460 lockflags |= LK_NOWAIT; 2461 2462 error = BUF_TIMELOCK(bp, lockflags, 2463 VI_MTX(vp), "getblk", slpflag, slptimeo); 2464 2465 /* 2466 * If we slept and got the lock we have to restart in case 2467 * the buffer changed identities. 2468 */ 2469 if (error == ENOLCK) 2470 goto loop; 2471 /* We timed out or were interrupted. */ 2472 else if (error) 2473 return (NULL); 2474 2475 /* 2476 * The buffer is locked. B_CACHE is cleared if the buffer is 2477 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2478 * and for a VMIO buffer B_CACHE is adjusted according to the 2479 * backing VM cache. 2480 */ 2481 if (bp->b_flags & B_INVAL) 2482 bp->b_flags &= ~B_CACHE; 2483 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2484 bp->b_flags |= B_CACHE; 2485 bremfree(bp); 2486 2487 /* 2488 * check for size inconsistancies for non-VMIO case. 2489 */ 2490 2491 if (bp->b_bcount != size) { 2492 if ((bp->b_flags & B_VMIO) == 0 || 2493 (size > bp->b_kvasize)) { 2494 if (bp->b_flags & B_DELWRI) { 2495 bp->b_flags |= B_NOCACHE; 2496 bwrite(bp); 2497 } else { 2498 if ((bp->b_flags & B_VMIO) && 2499 (LIST_FIRST(&bp->b_dep) == NULL)) { 2500 bp->b_flags |= B_RELBUF; 2501 brelse(bp); 2502 } else { 2503 bp->b_flags |= B_NOCACHE; 2504 bwrite(bp); 2505 } 2506 } 2507 goto loop; 2508 } 2509 } 2510 2511 /* 2512 * If the size is inconsistant in the VMIO case, we can resize 2513 * the buffer. This might lead to B_CACHE getting set or 2514 * cleared. If the size has not changed, B_CACHE remains 2515 * unchanged from its previous state. 2516 */ 2517 2518 if (bp->b_bcount != size) 2519 allocbuf(bp, size); 2520 2521 KASSERT(bp->b_offset != NOOFFSET, 2522 ("getblk: no buffer offset")); 2523 2524 /* 2525 * A buffer with B_DELWRI set and B_CACHE clear must 2526 * be committed before we can return the buffer in 2527 * order to prevent the caller from issuing a read 2528 * ( due to B_CACHE not being set ) and overwriting 2529 * it. 2530 * 2531 * Most callers, including NFS and FFS, need this to 2532 * operate properly either because they assume they 2533 * can issue a read if B_CACHE is not set, or because 2534 * ( for example ) an uncached B_DELWRI might loop due 2535 * to softupdates re-dirtying the buffer. In the latter 2536 * case, B_CACHE is set after the first write completes, 2537 * preventing further loops. 2538 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2539 * above while extending the buffer, we cannot allow the 2540 * buffer to remain with B_CACHE set after the write 2541 * completes or it will represent a corrupt state. To 2542 * deal with this we set B_NOCACHE to scrap the buffer 2543 * after the write. 2544 * 2545 * We might be able to do something fancy, like setting 2546 * B_CACHE in bwrite() except if B_DELWRI is already set, 2547 * so the below call doesn't set B_CACHE, but that gets real 2548 * confusing. This is much easier. 2549 */ 2550 2551 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2552 bp->b_flags |= B_NOCACHE; 2553 bwrite(bp); 2554 goto loop; 2555 } 2556 2557 splx(s); 2558 bp->b_flags &= ~B_DONE; 2559 } else { 2560 int bsize, maxsize, vmio; 2561 off_t offset; 2562 2563 /* 2564 * Buffer is not in-core, create new buffer. The buffer 2565 * returned by getnewbuf() is locked. Note that the returned 2566 * buffer is also considered valid (not marked B_INVAL). 2567 */ 2568 VI_UNLOCK(vp); 2569 /* 2570 * If the user does not want us to create the buffer, bail out 2571 * here. 2572 */ 2573 if (flags & GB_NOCREAT) { 2574 splx(s); 2575 return NULL; 2576 } 2577 2578 bsize = bo->bo_bsize; 2579 offset = blkno * bsize; 2580 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2581 (vp->v_vflag & VV_OBJBUF); 2582 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2583 maxsize = imax(maxsize, bsize); 2584 2585 bp = getnewbuf(slpflag, slptimeo, size, maxsize); 2586 if (bp == NULL) { 2587 if (slpflag || slptimeo) { 2588 splx(s); 2589 return NULL; 2590 } 2591 goto loop; 2592 } 2593 2594 /* 2595 * This code is used to make sure that a buffer is not 2596 * created while the getnewbuf routine is blocked. 2597 * This can be a problem whether the vnode is locked or not. 2598 * If the buffer is created out from under us, we have to 2599 * throw away the one we just created. There is now window 2600 * race because we are safely running at splbio() from the 2601 * point of the duplicate buffer creation through to here, 2602 * and we've locked the buffer. 2603 * 2604 * Note: this must occur before we associate the buffer 2605 * with the vp especially considering limitations in 2606 * the splay tree implementation when dealing with duplicate 2607 * lblkno's. 2608 */ 2609 BO_LOCK(bo); 2610 if (gbincore(bo, blkno)) { 2611 BO_UNLOCK(bo); 2612 bp->b_flags |= B_INVAL; 2613 brelse(bp); 2614 goto loop; 2615 } 2616 2617 /* 2618 * Insert the buffer into the hash, so that it can 2619 * be found by incore. 2620 */ 2621 bp->b_blkno = bp->b_lblkno = blkno; 2622 bp->b_offset = offset; 2623 2624 bgetvp(vp, bp); 2625 BO_UNLOCK(bo); 2626 2627 /* 2628 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2629 * buffer size starts out as 0, B_CACHE will be set by 2630 * allocbuf() for the VMIO case prior to it testing the 2631 * backing store for validity. 2632 */ 2633 2634 if (vmio) { 2635 bp->b_flags |= B_VMIO; 2636#if defined(VFS_BIO_DEBUG) 2637 if (vn_canvmio(vp) != TRUE) 2638 printf("getblk: VMIO on vnode type %d\n", 2639 vp->v_type); 2640#endif 2641 VOP_GETVOBJECT(vp, &vmo); 2642 KASSERT(vmo == bp->b_object, 2643 ("ARGH! different b_object %p %p %p\n", bp, vmo, bp->b_object)); 2644 } else { 2645 bp->b_flags &= ~B_VMIO; 2646 KASSERT(bp->b_object == NULL, 2647 ("ARGH! has b_object %p %p\n", bp, bp->b_object)); 2648 } 2649 2650 allocbuf(bp, size); 2651 2652 splx(s); 2653 bp->b_flags &= ~B_DONE; 2654 } 2655 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2656 KASSERT(bp->b_bufobj == bo, 2657 ("wrong b_bufobj %p should be %p", bp->b_bufobj, bo)); 2658 return (bp); 2659} 2660 2661/* 2662 * Get an empty, disassociated buffer of given size. The buffer is initially 2663 * set to B_INVAL. 2664 */ 2665struct buf * 2666geteblk(int size) 2667{ 2668 struct buf *bp; 2669 int s; 2670 int maxsize; 2671 2672 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2673 2674 s = splbio(); 2675 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2676 continue; 2677 splx(s); 2678 allocbuf(bp, size); 2679 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2680 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2681 return (bp); 2682} 2683 2684 2685/* 2686 * This code constitutes the buffer memory from either anonymous system 2687 * memory (in the case of non-VMIO operations) or from an associated 2688 * VM object (in the case of VMIO operations). This code is able to 2689 * resize a buffer up or down. 2690 * 2691 * Note that this code is tricky, and has many complications to resolve 2692 * deadlock or inconsistant data situations. Tread lightly!!! 2693 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2694 * the caller. Calling this code willy nilly can result in the loss of data. 2695 * 2696 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2697 * B_CACHE for the non-VMIO case. 2698 */ 2699 2700int 2701allocbuf(struct buf *bp, int size) 2702{ 2703 int newbsize, mbsize; 2704 int i; 2705 2706 GIANT_REQUIRED; 2707 2708 if (BUF_REFCNT(bp) == 0) 2709 panic("allocbuf: buffer not busy"); 2710 2711 if (bp->b_kvasize < size) 2712 panic("allocbuf: buffer too small"); 2713 2714 if ((bp->b_flags & B_VMIO) == 0) { 2715 caddr_t origbuf; 2716 int origbufsize; 2717 /* 2718 * Just get anonymous memory from the kernel. Don't 2719 * mess with B_CACHE. 2720 */ 2721 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2722 if (bp->b_flags & B_MALLOC) 2723 newbsize = mbsize; 2724 else 2725 newbsize = round_page(size); 2726 2727 if (newbsize < bp->b_bufsize) { 2728 /* 2729 * malloced buffers are not shrunk 2730 */ 2731 if (bp->b_flags & B_MALLOC) { 2732 if (newbsize) { 2733 bp->b_bcount = size; 2734 } else { 2735 free(bp->b_data, M_BIOBUF); 2736 if (bp->b_bufsize) { 2737 atomic_subtract_int( 2738 &bufmallocspace, 2739 bp->b_bufsize); 2740 bufspacewakeup(); 2741 bp->b_bufsize = 0; 2742 } 2743 bp->b_saveaddr = bp->b_kvabase; 2744 bp->b_data = bp->b_saveaddr; 2745 bp->b_bcount = 0; 2746 bp->b_flags &= ~B_MALLOC; 2747 } 2748 return 1; 2749 } 2750 vm_hold_free_pages( 2751 bp, 2752 (vm_offset_t) bp->b_data + newbsize, 2753 (vm_offset_t) bp->b_data + bp->b_bufsize); 2754 } else if (newbsize > bp->b_bufsize) { 2755 /* 2756 * We only use malloced memory on the first allocation. 2757 * and revert to page-allocated memory when the buffer 2758 * grows. 2759 */ 2760 /* 2761 * There is a potential smp race here that could lead 2762 * to bufmallocspace slightly passing the max. It 2763 * is probably extremely rare and not worth worrying 2764 * over. 2765 */ 2766 if ( (bufmallocspace < maxbufmallocspace) && 2767 (bp->b_bufsize == 0) && 2768 (mbsize <= PAGE_SIZE/2)) { 2769 2770 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2771 bp->b_bufsize = mbsize; 2772 bp->b_bcount = size; 2773 bp->b_flags |= B_MALLOC; 2774 atomic_add_int(&bufmallocspace, mbsize); 2775 return 1; 2776 } 2777 origbuf = NULL; 2778 origbufsize = 0; 2779 /* 2780 * If the buffer is growing on its other-than-first allocation, 2781 * then we revert to the page-allocation scheme. 2782 */ 2783 if (bp->b_flags & B_MALLOC) { 2784 origbuf = bp->b_data; 2785 origbufsize = bp->b_bufsize; 2786 bp->b_data = bp->b_kvabase; 2787 if (bp->b_bufsize) { 2788 atomic_subtract_int(&bufmallocspace, 2789 bp->b_bufsize); 2790 bufspacewakeup(); 2791 bp->b_bufsize = 0; 2792 } 2793 bp->b_flags &= ~B_MALLOC; 2794 newbsize = round_page(newbsize); 2795 } 2796 vm_hold_load_pages( 2797 bp, 2798 (vm_offset_t) bp->b_data + bp->b_bufsize, 2799 (vm_offset_t) bp->b_data + newbsize); 2800 if (origbuf) { 2801 bcopy(origbuf, bp->b_data, origbufsize); 2802 free(origbuf, M_BIOBUF); 2803 } 2804 } 2805 } else { 2806 int desiredpages; 2807 2808 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2809 desiredpages = (size == 0) ? 0 : 2810 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2811 2812 if (bp->b_flags & B_MALLOC) 2813 panic("allocbuf: VMIO buffer can't be malloced"); 2814 /* 2815 * Set B_CACHE initially if buffer is 0 length or will become 2816 * 0-length. 2817 */ 2818 if (size == 0 || bp->b_bufsize == 0) 2819 bp->b_flags |= B_CACHE; 2820 2821 if (newbsize < bp->b_bufsize) { 2822 /* 2823 * DEV_BSIZE aligned new buffer size is less then the 2824 * DEV_BSIZE aligned existing buffer size. Figure out 2825 * if we have to remove any pages. 2826 */ 2827 if (desiredpages < bp->b_npages) { 2828 vm_page_t m; 2829 2830 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2831 vm_page_lock_queues(); 2832 for (i = desiredpages; i < bp->b_npages; i++) { 2833 /* 2834 * the page is not freed here -- it 2835 * is the responsibility of 2836 * vnode_pager_setsize 2837 */ 2838 m = bp->b_pages[i]; 2839 KASSERT(m != bogus_page, 2840 ("allocbuf: bogus page found")); 2841 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2842 vm_page_lock_queues(); 2843 2844 bp->b_pages[i] = NULL; 2845 vm_page_unwire(m, 0); 2846 } 2847 vm_page_unlock_queues(); 2848 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2849 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2850 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2851 bp->b_npages = desiredpages; 2852 } 2853 } else if (size > bp->b_bcount) { 2854 /* 2855 * We are growing the buffer, possibly in a 2856 * byte-granular fashion. 2857 */ 2858 struct vnode *vp; 2859 vm_object_t obj; 2860 vm_offset_t toff; 2861 vm_offset_t tinc; 2862 2863 /* 2864 * Step 1, bring in the VM pages from the object, 2865 * allocating them if necessary. We must clear 2866 * B_CACHE if these pages are not valid for the 2867 * range covered by the buffer. 2868 */ 2869 2870 vp = bp->b_vp; 2871 obj = bp->b_object; 2872 2873 VM_OBJECT_LOCK(obj); 2874 while (bp->b_npages < desiredpages) { 2875 vm_page_t m; 2876 vm_pindex_t pi; 2877 2878 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2879 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2880 /* 2881 * note: must allocate system pages 2882 * since blocking here could intefere 2883 * with paging I/O, no matter which 2884 * process we are. 2885 */ 2886 m = vm_page_alloc(obj, pi, 2887 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2888 VM_ALLOC_WIRED); 2889 if (m == NULL) { 2890 atomic_add_int(&vm_pageout_deficit, 2891 desiredpages - bp->b_npages); 2892 VM_OBJECT_UNLOCK(obj); 2893 VM_WAIT; 2894 VM_OBJECT_LOCK(obj); 2895 } else { 2896 bp->b_flags &= ~B_CACHE; 2897 bp->b_pages[bp->b_npages] = m; 2898 ++bp->b_npages; 2899 } 2900 continue; 2901 } 2902 2903 /* 2904 * We found a page. If we have to sleep on it, 2905 * retry because it might have gotten freed out 2906 * from under us. 2907 * 2908 * We can only test PG_BUSY here. Blocking on 2909 * m->busy might lead to a deadlock: 2910 * 2911 * vm_fault->getpages->cluster_read->allocbuf 2912 * 2913 */ 2914 vm_page_lock_queues(); 2915 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2916 continue; 2917 2918 /* 2919 * We have a good page. Should we wakeup the 2920 * page daemon? 2921 */ 2922 if ((curproc != pageproc) && 2923 ((m->queue - m->pc) == PQ_CACHE) && 2924 ((cnt.v_free_count + cnt.v_cache_count) < 2925 (cnt.v_free_min + cnt.v_cache_min))) { 2926 pagedaemon_wakeup(); 2927 } 2928 vm_page_wire(m); 2929 vm_page_unlock_queues(); 2930 bp->b_pages[bp->b_npages] = m; 2931 ++bp->b_npages; 2932 } 2933 2934 /* 2935 * Step 2. We've loaded the pages into the buffer, 2936 * we have to figure out if we can still have B_CACHE 2937 * set. Note that B_CACHE is set according to the 2938 * byte-granular range ( bcount and size ), new the 2939 * aligned range ( newbsize ). 2940 * 2941 * The VM test is against m->valid, which is DEV_BSIZE 2942 * aligned. Needless to say, the validity of the data 2943 * needs to also be DEV_BSIZE aligned. Note that this 2944 * fails with NFS if the server or some other client 2945 * extends the file's EOF. If our buffer is resized, 2946 * B_CACHE may remain set! XXX 2947 */ 2948 2949 toff = bp->b_bcount; 2950 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2951 2952 while ((bp->b_flags & B_CACHE) && toff < size) { 2953 vm_pindex_t pi; 2954 2955 if (tinc > (size - toff)) 2956 tinc = size - toff; 2957 2958 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2959 PAGE_SHIFT; 2960 2961 vfs_buf_test_cache( 2962 bp, 2963 bp->b_offset, 2964 toff, 2965 tinc, 2966 bp->b_pages[pi] 2967 ); 2968 toff += tinc; 2969 tinc = PAGE_SIZE; 2970 } 2971 VM_OBJECT_UNLOCK(obj); 2972 2973 /* 2974 * Step 3, fixup the KVM pmap. Remember that 2975 * bp->b_data is relative to bp->b_offset, but 2976 * bp->b_offset may be offset into the first page. 2977 */ 2978 2979 bp->b_data = (caddr_t) 2980 trunc_page((vm_offset_t)bp->b_data); 2981 pmap_qenter( 2982 (vm_offset_t)bp->b_data, 2983 bp->b_pages, 2984 bp->b_npages 2985 ); 2986 2987 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2988 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2989 } 2990 } 2991 if (newbsize < bp->b_bufsize) 2992 bufspacewakeup(); 2993 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2994 bp->b_bcount = size; /* requested buffer size */ 2995 return 1; 2996} 2997 2998void 2999biodone(struct bio *bp) 3000{ 3001 3002 mtx_lock(&bdonelock); 3003 bp->bio_flags |= BIO_DONE; 3004 if (bp->bio_done == NULL) 3005 wakeup(bp); 3006 mtx_unlock(&bdonelock); 3007 if (bp->bio_done != NULL) 3008 bp->bio_done(bp); 3009} 3010 3011/* 3012 * Wait for a BIO to finish. 3013 * 3014 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3015 * case is not yet clear. 3016 */ 3017int 3018biowait(struct bio *bp, const char *wchan) 3019{ 3020 3021 mtx_lock(&bdonelock); 3022 while ((bp->bio_flags & BIO_DONE) == 0) 3023 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3024 mtx_unlock(&bdonelock); 3025 if (bp->bio_error != 0) 3026 return (bp->bio_error); 3027 if (!(bp->bio_flags & BIO_ERROR)) 3028 return (0); 3029 return (EIO); 3030} 3031 3032void 3033biofinish(struct bio *bp, struct devstat *stat, int error) 3034{ 3035 3036 if (error) { 3037 bp->bio_error = error; 3038 bp->bio_flags |= BIO_ERROR; 3039 } 3040 if (stat != NULL) 3041 devstat_end_transaction_bio(stat, bp); 3042 biodone(bp); 3043} 3044 3045/* 3046 * bufwait: 3047 * 3048 * Wait for buffer I/O completion, returning error status. The buffer 3049 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3050 * error and cleared. 3051 */ 3052int 3053bufwait(struct buf *bp) 3054{ 3055 int s; 3056 3057 s = splbio(); 3058 if (bp->b_iocmd == BIO_READ) 3059 bwait(bp, PRIBIO, "biord"); 3060 else 3061 bwait(bp, PRIBIO, "biowr"); 3062 splx(s); 3063 if (bp->b_flags & B_EINTR) { 3064 bp->b_flags &= ~B_EINTR; 3065 return (EINTR); 3066 } 3067 if (bp->b_ioflags & BIO_ERROR) { 3068 return (bp->b_error ? bp->b_error : EIO); 3069 } else { 3070 return (0); 3071 } 3072} 3073 3074 /* 3075 * Call back function from struct bio back up to struct buf. 3076 */ 3077static void 3078bufdonebio(struct bio *bip) 3079{ 3080 struct buf *bp; 3081 3082 /* Device drivers may or may not hold giant, hold it here. */ 3083 mtx_lock(&Giant); 3084 bp = bip->bio_caller2; 3085 bp->b_resid = bp->b_bcount - bip->bio_completed; 3086 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3087 bp->b_ioflags = bip->bio_flags; 3088 bp->b_error = bip->bio_error; 3089 if (bp->b_error) 3090 bp->b_ioflags |= BIO_ERROR; 3091 bufdone(bp); 3092 mtx_unlock(&Giant); 3093 g_destroy_bio(bip); 3094} 3095 3096void 3097dev_strategy(struct cdev *dev, struct buf *bp) 3098{ 3099 struct cdevsw *csw; 3100 struct bio *bip; 3101 3102 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3103 panic("b_iocmd botch"); 3104 for (;;) { 3105 bip = g_new_bio(); 3106 if (bip != NULL) 3107 break; 3108 /* Try again later */ 3109 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3110 } 3111 bip->bio_cmd = bp->b_iocmd; 3112 bip->bio_offset = bp->b_iooffset; 3113 bip->bio_length = bp->b_bcount; 3114 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3115 bip->bio_data = bp->b_data; 3116 bip->bio_done = bufdonebio; 3117 bip->bio_caller2 = bp; 3118 bip->bio_dev = dev; 3119 3120 KASSERT(dev->si_refcount > 0, 3121 ("dev_strategy on un-referenced struct cdev *(%s)", 3122 devtoname(dev))); 3123 csw = dev_refthread(dev); 3124 if (csw == NULL) { 3125 bp->b_error = ENXIO; 3126 bp->b_ioflags = BIO_ERROR; 3127 mtx_lock(&Giant); /* XXX: too defensive ? */ 3128 bufdone(bp); 3129 mtx_unlock(&Giant); /* XXX: too defensive ? */ 3130 return; 3131 } 3132 (*csw->d_strategy)(bip); 3133 dev_relthread(dev); 3134} 3135 3136/* 3137 * bufdone: 3138 * 3139 * Finish I/O on a buffer, optionally calling a completion function. 3140 * This is usually called from an interrupt so process blocking is 3141 * not allowed. 3142 * 3143 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3144 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3145 * assuming B_INVAL is clear. 3146 * 3147 * For the VMIO case, we set B_CACHE if the op was a read and no 3148 * read error occured, or if the op was a write. B_CACHE is never 3149 * set if the buffer is invalid or otherwise uncacheable. 3150 * 3151 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3152 * initiator to leave B_INVAL set to brelse the buffer out of existance 3153 * in the biodone routine. 3154 */ 3155void 3156bufdone(struct buf *bp) 3157{ 3158 int s; 3159 void (*biodone)(struct buf *); 3160 3161 3162 s = splbio(); 3163 3164 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3165 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3166 3167 bp->b_flags |= B_DONE; 3168 runningbufwakeup(bp); 3169 3170 if (bp->b_iocmd == BIO_WRITE && bp->b_bufobj != NULL) 3171 bufobj_wdrop(bp->b_bufobj); 3172 3173 /* call optional completion function if requested */ 3174 if (bp->b_iodone != NULL) { 3175 biodone = bp->b_iodone; 3176 bp->b_iodone = NULL; 3177 (*biodone) (bp); 3178 splx(s); 3179 return; 3180 } 3181 if (LIST_FIRST(&bp->b_dep) != NULL) 3182 buf_complete(bp); 3183 3184 if (bp->b_flags & B_VMIO) { 3185 int i; 3186 vm_ooffset_t foff; 3187 vm_page_t m; 3188 vm_object_t obj; 3189 int iosize; 3190 struct vnode *vp = bp->b_vp; 3191 3192 obj = bp->b_object; 3193 3194#if defined(VFS_BIO_DEBUG) 3195 mp_fixme("usecount and vflag accessed without locks."); 3196 if (vp->v_usecount == 0) { 3197 panic("biodone: zero vnode ref count"); 3198 } 3199 3200 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3201 panic("biodone: vnode is not setup for merged cache"); 3202 } 3203#endif 3204 3205 foff = bp->b_offset; 3206 KASSERT(bp->b_offset != NOOFFSET, 3207 ("biodone: no buffer offset")); 3208 3209 VM_OBJECT_LOCK(obj); 3210#if defined(VFS_BIO_DEBUG) 3211 if (obj->paging_in_progress < bp->b_npages) { 3212 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3213 obj->paging_in_progress, bp->b_npages); 3214 } 3215#endif 3216 3217 /* 3218 * Set B_CACHE if the op was a normal read and no error 3219 * occured. B_CACHE is set for writes in the b*write() 3220 * routines. 3221 */ 3222 iosize = bp->b_bcount - bp->b_resid; 3223 if (bp->b_iocmd == BIO_READ && 3224 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3225 !(bp->b_ioflags & BIO_ERROR)) { 3226 bp->b_flags |= B_CACHE; 3227 } 3228 vm_page_lock_queues(); 3229 for (i = 0; i < bp->b_npages; i++) { 3230 int bogusflag = 0; 3231 int resid; 3232 3233 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3234 if (resid > iosize) 3235 resid = iosize; 3236 3237 /* 3238 * cleanup bogus pages, restoring the originals 3239 */ 3240 m = bp->b_pages[i]; 3241 if (m == bogus_page) { 3242 bogusflag = 1; 3243 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3244 if (m == NULL) 3245 panic("biodone: page disappeared!"); 3246 bp->b_pages[i] = m; 3247 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3248 } 3249#if defined(VFS_BIO_DEBUG) 3250 if (OFF_TO_IDX(foff) != m->pindex) { 3251 printf( 3252"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3253 (intmax_t)foff, (uintmax_t)m->pindex); 3254 } 3255#endif 3256 3257 /* 3258 * In the write case, the valid and clean bits are 3259 * already changed correctly ( see bdwrite() ), so we 3260 * only need to do this here in the read case. 3261 */ 3262 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3263 vfs_page_set_valid(bp, foff, i, m); 3264 } 3265 3266 /* 3267 * when debugging new filesystems or buffer I/O methods, this 3268 * is the most common error that pops up. if you see this, you 3269 * have not set the page busy flag correctly!!! 3270 */ 3271 if (m->busy == 0) { 3272 printf("biodone: page busy < 0, " 3273 "pindex: %d, foff: 0x(%x,%x), " 3274 "resid: %d, index: %d\n", 3275 (int) m->pindex, (int)(foff >> 32), 3276 (int) foff & 0xffffffff, resid, i); 3277 if (!vn_isdisk(vp, NULL)) 3278 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3279 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3280 (intmax_t) bp->b_lblkno, 3281 bp->b_flags, bp->b_npages); 3282 else 3283 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3284 (intmax_t) bp->b_lblkno, 3285 bp->b_flags, bp->b_npages); 3286 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3287 (u_long)m->valid, (u_long)m->dirty, 3288 m->wire_count); 3289 panic("biodone: page busy < 0\n"); 3290 } 3291 vm_page_io_finish(m); 3292 vm_object_pip_subtract(obj, 1); 3293 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3294 iosize -= resid; 3295 } 3296 vm_page_unlock_queues(); 3297 vm_object_pip_wakeupn(obj, 0); 3298 VM_OBJECT_UNLOCK(obj); 3299 } 3300 3301 /* 3302 * For asynchronous completions, release the buffer now. The brelse 3303 * will do a wakeup there if necessary - so no need to do a wakeup 3304 * here in the async case. The sync case always needs to do a wakeup. 3305 */ 3306 3307 if (bp->b_flags & B_ASYNC) { 3308 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3309 brelse(bp); 3310 else 3311 bqrelse(bp); 3312 } else { 3313 bdone(bp); 3314 } 3315 splx(s); 3316} 3317 3318/* 3319 * This routine is called in lieu of iodone in the case of 3320 * incomplete I/O. This keeps the busy status for pages 3321 * consistant. 3322 */ 3323void 3324vfs_unbusy_pages(struct buf *bp) 3325{ 3326 int i; 3327 vm_object_t obj; 3328 vm_page_t m; 3329 3330 runningbufwakeup(bp); 3331 if (!(bp->b_flags & B_VMIO)) 3332 return; 3333 3334 obj = bp->b_object; 3335 VM_OBJECT_LOCK(obj); 3336 vm_page_lock_queues(); 3337 for (i = 0; i < bp->b_npages; i++) { 3338 m = bp->b_pages[i]; 3339 if (m == bogus_page) { 3340 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3341 if (!m) 3342 panic("vfs_unbusy_pages: page missing\n"); 3343 bp->b_pages[i] = m; 3344 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3345 bp->b_pages, bp->b_npages); 3346 } 3347 vm_object_pip_subtract(obj, 1); 3348 vm_page_io_finish(m); 3349 } 3350 vm_page_unlock_queues(); 3351 vm_object_pip_wakeupn(obj, 0); 3352 VM_OBJECT_UNLOCK(obj); 3353} 3354 3355/* 3356 * vfs_page_set_valid: 3357 * 3358 * Set the valid bits in a page based on the supplied offset. The 3359 * range is restricted to the buffer's size. 3360 * 3361 * This routine is typically called after a read completes. 3362 */ 3363static void 3364vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3365{ 3366 vm_ooffset_t soff, eoff; 3367 3368 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3369 /* 3370 * Start and end offsets in buffer. eoff - soff may not cross a 3371 * page boundry or cross the end of the buffer. The end of the 3372 * buffer, in this case, is our file EOF, not the allocation size 3373 * of the buffer. 3374 */ 3375 soff = off; 3376 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3377 if (eoff > bp->b_offset + bp->b_bcount) 3378 eoff = bp->b_offset + bp->b_bcount; 3379 3380 /* 3381 * Set valid range. This is typically the entire buffer and thus the 3382 * entire page. 3383 */ 3384 if (eoff > soff) { 3385 vm_page_set_validclean( 3386 m, 3387 (vm_offset_t) (soff & PAGE_MASK), 3388 (vm_offset_t) (eoff - soff) 3389 ); 3390 } 3391} 3392 3393/* 3394 * This routine is called before a device strategy routine. 3395 * It is used to tell the VM system that paging I/O is in 3396 * progress, and treat the pages associated with the buffer 3397 * almost as being PG_BUSY. Also the object paging_in_progress 3398 * flag is handled to make sure that the object doesn't become 3399 * inconsistant. 3400 * 3401 * Since I/O has not been initiated yet, certain buffer flags 3402 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3403 * and should be ignored. 3404 */ 3405void 3406vfs_busy_pages(struct buf *bp, int clear_modify) 3407{ 3408 int i, bogus; 3409 vm_object_t obj; 3410 vm_ooffset_t foff; 3411 vm_page_t m; 3412 3413 if (!(bp->b_flags & B_VMIO)) 3414 return; 3415 3416 obj = bp->b_object; 3417 foff = bp->b_offset; 3418 KASSERT(bp->b_offset != NOOFFSET, 3419 ("vfs_busy_pages: no buffer offset")); 3420 vfs_setdirty(bp); 3421 VM_OBJECT_LOCK(obj); 3422retry: 3423 vm_page_lock_queues(); 3424 for (i = 0; i < bp->b_npages; i++) { 3425 m = bp->b_pages[i]; 3426 3427 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3428 goto retry; 3429 } 3430 bogus = 0; 3431 for (i = 0; i < bp->b_npages; i++) { 3432 m = bp->b_pages[i]; 3433 3434 if ((bp->b_flags & B_CLUSTER) == 0) { 3435 vm_object_pip_add(obj, 1); 3436 vm_page_io_start(m); 3437 } 3438 /* 3439 * When readying a buffer for a read ( i.e 3440 * clear_modify == 0 ), it is important to do 3441 * bogus_page replacement for valid pages in 3442 * partially instantiated buffers. Partially 3443 * instantiated buffers can, in turn, occur when 3444 * reconstituting a buffer from its VM backing store 3445 * base. We only have to do this if B_CACHE is 3446 * clear ( which causes the I/O to occur in the 3447 * first place ). The replacement prevents the read 3448 * I/O from overwriting potentially dirty VM-backed 3449 * pages. XXX bogus page replacement is, uh, bogus. 3450 * It may not work properly with small-block devices. 3451 * We need to find a better way. 3452 */ 3453 pmap_remove_all(m); 3454 if (clear_modify) 3455 vfs_page_set_valid(bp, foff, i, m); 3456 else if (m->valid == VM_PAGE_BITS_ALL && 3457 (bp->b_flags & B_CACHE) == 0) { 3458 bp->b_pages[i] = bogus_page; 3459 bogus++; 3460 } 3461 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3462 } 3463 vm_page_unlock_queues(); 3464 VM_OBJECT_UNLOCK(obj); 3465 if (bogus) 3466 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3467 bp->b_pages, bp->b_npages); 3468} 3469 3470/* 3471 * Tell the VM system that the pages associated with this buffer 3472 * are clean. This is used for delayed writes where the data is 3473 * going to go to disk eventually without additional VM intevention. 3474 * 3475 * Note that while we only really need to clean through to b_bcount, we 3476 * just go ahead and clean through to b_bufsize. 3477 */ 3478static void 3479vfs_clean_pages(struct buf *bp) 3480{ 3481 int i; 3482 vm_ooffset_t foff, noff, eoff; 3483 vm_page_t m; 3484 3485 if (!(bp->b_flags & B_VMIO)) 3486 return; 3487 3488 foff = bp->b_offset; 3489 KASSERT(bp->b_offset != NOOFFSET, 3490 ("vfs_clean_pages: no buffer offset")); 3491 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3492 vm_page_lock_queues(); 3493 for (i = 0; i < bp->b_npages; i++) { 3494 m = bp->b_pages[i]; 3495 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3496 eoff = noff; 3497 3498 if (eoff > bp->b_offset + bp->b_bufsize) 3499 eoff = bp->b_offset + bp->b_bufsize; 3500 vfs_page_set_valid(bp, foff, i, m); 3501 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3502 foff = noff; 3503 } 3504 vm_page_unlock_queues(); 3505 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3506} 3507 3508/* 3509 * vfs_bio_set_validclean: 3510 * 3511 * Set the range within the buffer to valid and clean. The range is 3512 * relative to the beginning of the buffer, b_offset. Note that b_offset 3513 * itself may be offset from the beginning of the first page. 3514 * 3515 */ 3516 3517void 3518vfs_bio_set_validclean(struct buf *bp, int base, int size) 3519{ 3520 int i, n; 3521 vm_page_t m; 3522 3523 if (!(bp->b_flags & B_VMIO)) 3524 return; 3525 3526 /* 3527 * Fixup base to be relative to beginning of first page. 3528 * Set initial n to be the maximum number of bytes in the 3529 * first page that can be validated. 3530 */ 3531 3532 base += (bp->b_offset & PAGE_MASK); 3533 n = PAGE_SIZE - (base & PAGE_MASK); 3534 3535 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3536 vm_page_lock_queues(); 3537 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3538 m = bp->b_pages[i]; 3539 3540 if (n > size) 3541 n = size; 3542 3543 vm_page_set_validclean(m, base & PAGE_MASK, n); 3544 base += n; 3545 size -= n; 3546 n = PAGE_SIZE; 3547 } 3548 vm_page_unlock_queues(); 3549 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3550} 3551 3552/* 3553 * vfs_bio_clrbuf: 3554 * 3555 * clear a buffer. This routine essentially fakes an I/O, so we need 3556 * to clear BIO_ERROR and B_INVAL. 3557 * 3558 * Note that while we only theoretically need to clear through b_bcount, 3559 * we go ahead and clear through b_bufsize. 3560 */ 3561 3562void 3563vfs_bio_clrbuf(struct buf *bp) 3564{ 3565 int i, j, mask = 0; 3566 caddr_t sa, ea; 3567 3568 GIANT_REQUIRED; 3569 3570 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3571 clrbuf(bp); 3572 return; 3573 } 3574 bp->b_flags &= ~B_INVAL; 3575 bp->b_ioflags &= ~BIO_ERROR; 3576 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3577 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3578 (bp->b_offset & PAGE_MASK) == 0) { 3579 if (bp->b_pages[0] == bogus_page) 3580 goto unlock; 3581 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3582 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3583 if ((bp->b_pages[0]->valid & mask) == mask) 3584 goto unlock; 3585 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3586 ((bp->b_pages[0]->valid & mask) == 0)) { 3587 bzero(bp->b_data, bp->b_bufsize); 3588 bp->b_pages[0]->valid |= mask; 3589 goto unlock; 3590 } 3591 } 3592 ea = sa = bp->b_data; 3593 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3594 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3595 ea = (caddr_t)(vm_offset_t)ulmin( 3596 (u_long)(vm_offset_t)ea, 3597 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3598 if (bp->b_pages[i] == bogus_page) 3599 continue; 3600 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3601 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3602 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3603 if ((bp->b_pages[i]->valid & mask) == mask) 3604 continue; 3605 if ((bp->b_pages[i]->valid & mask) == 0) { 3606 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3607 bzero(sa, ea - sa); 3608 } else { 3609 for (; sa < ea; sa += DEV_BSIZE, j++) { 3610 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3611 (bp->b_pages[i]->valid & (1<<j)) == 0) 3612 bzero(sa, DEV_BSIZE); 3613 } 3614 } 3615 bp->b_pages[i]->valid |= mask; 3616 } 3617unlock: 3618 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3619 bp->b_resid = 0; 3620} 3621 3622/* 3623 * vm_hold_load_pages and vm_hold_free_pages get pages into 3624 * a buffers address space. The pages are anonymous and are 3625 * not associated with a file object. 3626 */ 3627static void 3628vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3629{ 3630 vm_offset_t pg; 3631 vm_page_t p; 3632 int index; 3633 3634 to = round_page(to); 3635 from = round_page(from); 3636 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3637 3638 VM_OBJECT_LOCK(kernel_object); 3639 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3640tryagain: 3641 /* 3642 * note: must allocate system pages since blocking here 3643 * could intefere with paging I/O, no matter which 3644 * process we are. 3645 */ 3646 p = vm_page_alloc(kernel_object, 3647 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3648 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3649 if (!p) { 3650 atomic_add_int(&vm_pageout_deficit, 3651 (to - pg) >> PAGE_SHIFT); 3652 VM_OBJECT_UNLOCK(kernel_object); 3653 VM_WAIT; 3654 VM_OBJECT_LOCK(kernel_object); 3655 goto tryagain; 3656 } 3657 p->valid = VM_PAGE_BITS_ALL; 3658 pmap_qenter(pg, &p, 1); 3659 bp->b_pages[index] = p; 3660 } 3661 VM_OBJECT_UNLOCK(kernel_object); 3662 bp->b_npages = index; 3663} 3664 3665/* Return pages associated with this buf to the vm system */ 3666static void 3667vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3668{ 3669 vm_offset_t pg; 3670 vm_page_t p; 3671 int index, newnpages; 3672 3673 from = round_page(from); 3674 to = round_page(to); 3675 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3676 3677 VM_OBJECT_LOCK(kernel_object); 3678 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3679 p = bp->b_pages[index]; 3680 if (p && (index < bp->b_npages)) { 3681 if (p->busy) { 3682 printf( 3683 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3684 (intmax_t)bp->b_blkno, 3685 (intmax_t)bp->b_lblkno); 3686 } 3687 bp->b_pages[index] = NULL; 3688 pmap_qremove(pg, 1); 3689 vm_page_lock_queues(); 3690 vm_page_busy(p); 3691 vm_page_unwire(p, 0); 3692 vm_page_free(p); 3693 vm_page_unlock_queues(); 3694 } 3695 } 3696 VM_OBJECT_UNLOCK(kernel_object); 3697 bp->b_npages = newnpages; 3698} 3699 3700/* 3701 * Map an IO request into kernel virtual address space. 3702 * 3703 * All requests are (re)mapped into kernel VA space. 3704 * Notice that we use b_bufsize for the size of the buffer 3705 * to be mapped. b_bcount might be modified by the driver. 3706 * 3707 * Note that even if the caller determines that the address space should 3708 * be valid, a race or a smaller-file mapped into a larger space may 3709 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3710 * check the return value. 3711 */ 3712int 3713vmapbuf(struct buf *bp) 3714{ 3715 caddr_t addr, kva; 3716 vm_prot_t prot; 3717 int pidx, i; 3718 struct vm_page *m; 3719 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3720 3721 if (bp->b_bufsize < 0) 3722 return (-1); 3723 prot = VM_PROT_READ; 3724 if (bp->b_iocmd == BIO_READ) 3725 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3726 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3727 addr < bp->b_data + bp->b_bufsize; 3728 addr += PAGE_SIZE, pidx++) { 3729 /* 3730 * Do the vm_fault if needed; do the copy-on-write thing 3731 * when reading stuff off device into memory. 3732 * 3733 * NOTE! Must use pmap_extract() because addr may be in 3734 * the userland address space, and kextract is only guarenteed 3735 * to work for the kernland address space (see: sparc64 port). 3736 */ 3737retry: 3738 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3739 prot) < 0) { 3740 vm_page_lock_queues(); 3741 for (i = 0; i < pidx; ++i) { 3742 vm_page_unhold(bp->b_pages[i]); 3743 bp->b_pages[i] = NULL; 3744 } 3745 vm_page_unlock_queues(); 3746 return(-1); 3747 } 3748 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3749 if (m == NULL) 3750 goto retry; 3751 bp->b_pages[pidx] = m; 3752 } 3753 if (pidx > btoc(MAXPHYS)) 3754 panic("vmapbuf: mapped more than MAXPHYS"); 3755 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3756 3757 kva = bp->b_saveaddr; 3758 bp->b_npages = pidx; 3759 bp->b_saveaddr = bp->b_data; 3760 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3761 return(0); 3762} 3763 3764/* 3765 * Free the io map PTEs associated with this IO operation. 3766 * We also invalidate the TLB entries and restore the original b_addr. 3767 */ 3768void 3769vunmapbuf(struct buf *bp) 3770{ 3771 int pidx; 3772 int npages; 3773 3774 npages = bp->b_npages; 3775 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3776 vm_page_lock_queues(); 3777 for (pidx = 0; pidx < npages; pidx++) 3778 vm_page_unhold(bp->b_pages[pidx]); 3779 vm_page_unlock_queues(); 3780 3781 bp->b_data = bp->b_saveaddr; 3782} 3783 3784void 3785bdone(struct buf *bp) 3786{ 3787 3788 mtx_lock(&bdonelock); 3789 bp->b_flags |= B_DONE; 3790 wakeup(bp); 3791 mtx_unlock(&bdonelock); 3792} 3793 3794void 3795bwait(struct buf *bp, u_char pri, const char *wchan) 3796{ 3797 3798 mtx_lock(&bdonelock); 3799 while ((bp->b_flags & B_DONE) == 0) 3800 msleep(bp, &bdonelock, pri, wchan, 0); 3801 mtx_unlock(&bdonelock); 3802} 3803 3804void 3805bufstrategy(struct bufobj *bo, struct buf *bp) 3806{ 3807 int i = 0; 3808 struct vnode *vp; 3809 3810 vp = bp->b_vp; 3811 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3812 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3813 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3814 i = VOP_STRATEGY(vp, bp); 3815 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3816} 3817 3818void 3819bufobj_wref(struct bufobj *bo) 3820{ 3821 3822 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3823 BO_LOCK(bo); 3824 bo->bo_numoutput++; 3825 BO_UNLOCK(bo); 3826} 3827 3828void 3829bufobj_wdrop(struct bufobj *bo) 3830{ 3831 3832 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3833 BO_LOCK(bo); 3834 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3835 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3836 bo->bo_flag &= ~BO_WWAIT; 3837 wakeup(&bo->bo_numoutput); 3838 } 3839 BO_UNLOCK(bo); 3840} 3841 3842int 3843bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3844{ 3845 int error; 3846 3847 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3848 ASSERT_BO_LOCKED(bo); 3849 error = 0; 3850 while (bo->bo_numoutput) { 3851 bo->bo_flag |= BO_WWAIT; 3852 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3853 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3854 if (error) 3855 break; 3856 } 3857 return (error); 3858} 3859 3860#include "opt_ddb.h" 3861#ifdef DDB 3862#include <ddb/ddb.h> 3863 3864/* DDB command to show buffer data */ 3865DB_SHOW_COMMAND(buffer, db_show_buffer) 3866{ 3867 /* get args */ 3868 struct buf *bp = (struct buf *)addr; 3869 3870 if (!have_addr) { 3871 db_printf("usage: show buffer <addr>\n"); 3872 return; 3873 } 3874 3875 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3876 db_printf( 3877 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3878 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd\n", 3879 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3880 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3881 (intmax_t)bp->b_blkno); 3882 if (bp->b_npages) { 3883 int i; 3884 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3885 for (i = 0; i < bp->b_npages; i++) { 3886 vm_page_t m; 3887 m = bp->b_pages[i]; 3888 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3889 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3890 if ((i + 1) < bp->b_npages) 3891 db_printf(","); 3892 } 3893 db_printf("\n"); 3894 } 3895} 3896#endif /* DDB */ 3897