vfs_bio.c revision 146801
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 146801 2005-05-30 07:01:18Z jeff $"); 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 .bop_sync = bufsync, 83}; 84 85/* 86 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 87 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 88 */ 89struct buf *buf; /* buffer header pool */ 90 91static struct proc *bufdaemonproc; 92 93static int inmem(struct vnode *vp, daddr_t blkno); 94static void vm_hold_free_pages(struct buf *bp, vm_offset_t from, 95 vm_offset_t to); 96static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 97 vm_offset_t to); 98static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 99 int pageno, vm_page_t m); 100static void vfs_clean_pages(struct buf *bp); 101static void vfs_setdirty(struct buf *bp); 102static void vfs_vmio_release(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); 107static void 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"); 184 185/* 186 * Wakeup point for bufdaemon, as well as indicator of whether it is already 187 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 188 * is idling. 189 */ 190static int bd_request; 191 192/* 193 * This lock synchronizes access to bd_request. 194 */ 195static struct mtx bdlock; 196 197/* 198 * bogus page -- for I/O to/from partially complete buffers 199 * this is a temporary solution to the problem, but it is not 200 * really that bad. it would be better to split the buffer 201 * for input in the case of buffers partially already in memory, 202 * but the code is intricate enough already. 203 */ 204vm_page_t bogus_page; 205 206/* 207 * Synchronization (sleep/wakeup) variable for active buffer space requests. 208 * Set when wait starts, cleared prior to wakeup(). 209 * Used in runningbufwakeup() and waitrunningbufspace(). 210 */ 211static int runningbufreq; 212 213/* 214 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 215 * waitrunningbufspace(). 216 */ 217static struct mtx rbreqlock; 218 219/* 220 * Synchronization (sleep/wakeup) variable for buffer requests. 221 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 222 * by and/or. 223 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 224 * getnewbuf(), and getblk(). 225 */ 226static int needsbuffer; 227 228/* 229 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 230 */ 231static struct mtx nblock; 232 233/* 234 * Lock that protects against bwait()/bdone()/B_DONE races. 235 */ 236 237static struct mtx bdonelock; 238 239/* 240 * Definitions for the buffer free lists. 241 */ 242#define BUFFER_QUEUES 5 /* number of free buffer queues */ 243 244#define QUEUE_NONE 0 /* on no queue */ 245#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 246#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 247#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 248#define QUEUE_EMPTY 4 /* empty buffer headers */ 249 250/* Queues for free buffers with various properties */ 251static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 252 253/* Lock for the bufqueues */ 254static struct mtx bqlock; 255 256/* 257 * Single global constant for BUF_WMESG, to avoid getting multiple references. 258 * buf_wmesg is referred from macros. 259 */ 260const char *buf_wmesg = BUF_WMESG; 261 262#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 263#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 264#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 265#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 266 267#ifdef DIRECTIO 268extern void ffs_rawread_setup(void); 269#endif /* DIRECTIO */ 270/* 271 * numdirtywakeup: 272 * 273 * If someone is blocked due to there being too many dirty buffers, 274 * and numdirtybuffers is now reasonable, wake them up. 275 */ 276 277static __inline void 278numdirtywakeup(int level) 279{ 280 281 if (numdirtybuffers <= level) { 282 mtx_lock(&nblock); 283 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 284 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 285 wakeup(&needsbuffer); 286 } 287 mtx_unlock(&nblock); 288 } 289} 290 291/* 292 * bufspacewakeup: 293 * 294 * Called when buffer space is potentially available for recovery. 295 * getnewbuf() will block on this flag when it is unable to free 296 * sufficient buffer space. Buffer space becomes recoverable when 297 * bp's get placed back in the queues. 298 */ 299 300static __inline void 301bufspacewakeup(void) 302{ 303 304 /* 305 * If someone is waiting for BUF space, wake them up. Even 306 * though we haven't freed the kva space yet, the waiting 307 * process will be able to now. 308 */ 309 mtx_lock(&nblock); 310 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 311 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 312 wakeup(&needsbuffer); 313 } 314 mtx_unlock(&nblock); 315} 316 317/* 318 * runningbufwakeup() - in-progress I/O accounting. 319 * 320 */ 321static __inline void 322runningbufwakeup(struct buf *bp) 323{ 324 325 if (bp->b_runningbufspace) { 326 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 327 bp->b_runningbufspace = 0; 328 mtx_lock(&rbreqlock); 329 if (runningbufreq && runningbufspace <= lorunningspace) { 330 runningbufreq = 0; 331 wakeup(&runningbufreq); 332 } 333 mtx_unlock(&rbreqlock); 334 } 335} 336 337/* 338 * bufcountwakeup: 339 * 340 * Called when a buffer has been added to one of the free queues to 341 * account for the buffer and to wakeup anyone waiting for free buffers. 342 * This typically occurs when large amounts of metadata are being handled 343 * by the buffer cache ( else buffer space runs out first, usually ). 344 */ 345 346static __inline void 347bufcountwakeup(void) 348{ 349 350 atomic_add_int(&numfreebuffers, 1); 351 mtx_lock(&nblock); 352 if (needsbuffer) { 353 needsbuffer &= ~VFS_BIO_NEED_ANY; 354 if (numfreebuffers >= hifreebuffers) 355 needsbuffer &= ~VFS_BIO_NEED_FREE; 356 wakeup(&needsbuffer); 357 } 358 mtx_unlock(&nblock); 359} 360 361/* 362 * waitrunningbufspace() 363 * 364 * runningbufspace is a measure of the amount of I/O currently 365 * running. This routine is used in async-write situations to 366 * prevent creating huge backups of pending writes to a device. 367 * Only asynchronous writes are governed by this function. 368 * 369 * Reads will adjust runningbufspace, but will not block based on it. 370 * The read load has a side effect of reducing the allowed write load. 371 * 372 * This does NOT turn an async write into a sync write. It waits 373 * for earlier writes to complete and generally returns before the 374 * caller's write has reached the device. 375 */ 376static __inline void 377waitrunningbufspace(void) 378{ 379 380 mtx_lock(&rbreqlock); 381 while (runningbufspace > hirunningspace) { 382 ++runningbufreq; 383 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 384 } 385 mtx_unlock(&rbreqlock); 386} 387 388 389/* 390 * vfs_buf_test_cache: 391 * 392 * Called when a buffer is extended. This function clears the B_CACHE 393 * bit if the newly extended portion of the buffer does not contain 394 * valid data. 395 */ 396static __inline 397void 398vfs_buf_test_cache(struct buf *bp, 399 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 400 vm_page_t m) 401{ 402 403 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 404 if (bp->b_flags & B_CACHE) { 405 int base = (foff + off) & PAGE_MASK; 406 if (vm_page_is_valid(m, base, size) == 0) 407 bp->b_flags &= ~B_CACHE; 408 } 409} 410 411/* Wake up the buffer deamon if necessary */ 412static __inline 413void 414bd_wakeup(int dirtybuflevel) 415{ 416 417 mtx_lock(&bdlock); 418 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 419 bd_request = 1; 420 wakeup(&bd_request); 421 } 422 mtx_unlock(&bdlock); 423} 424 425/* 426 * bd_speedup - speedup the buffer cache flushing code 427 */ 428 429static __inline 430void 431bd_speedup(void) 432{ 433 434 bd_wakeup(1); 435} 436 437/* 438 * Calculating buffer cache scaling values and reserve space for buffer 439 * headers. This is called during low level kernel initialization and 440 * may be called more then once. We CANNOT write to the memory area 441 * being reserved at this time. 442 */ 443caddr_t 444kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 445{ 446 447 /* 448 * physmem_est is in pages. Convert it to kilobytes (assumes 449 * PAGE_SIZE is >= 1K) 450 */ 451 physmem_est = physmem_est * (PAGE_SIZE / 1024); 452 453 /* 454 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 455 * For the first 64MB of ram nominally allocate sufficient buffers to 456 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 457 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 458 * the buffer cache we limit the eventual kva reservation to 459 * maxbcache bytes. 460 * 461 * factor represents the 1/4 x ram conversion. 462 */ 463 if (nbuf == 0) { 464 int factor = 4 * BKVASIZE / 1024; 465 466 nbuf = 50; 467 if (physmem_est > 4096) 468 nbuf += min((physmem_est - 4096) / factor, 469 65536 / factor); 470 if (physmem_est > 65536) 471 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 472 473 if (maxbcache && nbuf > maxbcache / BKVASIZE) 474 nbuf = maxbcache / BKVASIZE; 475 } 476 477#if 0 478 /* 479 * Do not allow the buffer_map to be more then 1/2 the size of the 480 * kernel_map. 481 */ 482 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 483 (BKVASIZE * 2)) { 484 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 485 (BKVASIZE * 2); 486 printf("Warning: nbufs capped at %d\n", nbuf); 487 } 488#endif 489 490 /* 491 * swbufs are used as temporary holders for I/O, such as paging I/O. 492 * We have no less then 16 and no more then 256. 493 */ 494 nswbuf = max(min(nbuf/4, 256), 16); 495#ifdef NSWBUF_MIN 496 if (nswbuf < NSWBUF_MIN) 497 nswbuf = NSWBUF_MIN; 498#endif 499#ifdef DIRECTIO 500 ffs_rawread_setup(); 501#endif 502 503 /* 504 * Reserve space for the buffer cache buffers 505 */ 506 swbuf = (void *)v; 507 v = (caddr_t)(swbuf + nswbuf); 508 buf = (void *)v; 509 v = (caddr_t)(buf + nbuf); 510 511 return(v); 512} 513 514/* Initialize the buffer subsystem. Called before use of any buffers. */ 515void 516bufinit(void) 517{ 518 struct buf *bp; 519 int i; 520 521 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 522 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 523 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 524 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 525 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 526 527 /* next, make a null set of free lists */ 528 for (i = 0; i < BUFFER_QUEUES; i++) 529 TAILQ_INIT(&bufqueues[i]); 530 531 /* finally, initialize each buffer header and stick on empty q */ 532 for (i = 0; i < nbuf; i++) { 533 bp = &buf[i]; 534 bzero(bp, sizeof *bp); 535 bp->b_flags = B_INVAL; /* we're just an empty header */ 536 bp->b_rcred = NOCRED; 537 bp->b_wcred = NOCRED; 538 bp->b_qindex = QUEUE_EMPTY; 539 bp->b_vflags = 0; 540 bp->b_xflags = 0; 541 LIST_INIT(&bp->b_dep); 542 BUF_LOCKINIT(bp); 543 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 544 } 545 546 /* 547 * maxbufspace is the absolute maximum amount of buffer space we are 548 * allowed to reserve in KVM and in real terms. The absolute maximum 549 * is nominally used by buf_daemon. hibufspace is the nominal maximum 550 * used by most other processes. The differential is required to 551 * ensure that buf_daemon is able to run when other processes might 552 * be blocked waiting for buffer space. 553 * 554 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 555 * this may result in KVM fragmentation which is not handled optimally 556 * by the system. 557 */ 558 maxbufspace = nbuf * BKVASIZE; 559 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 560 lobufspace = hibufspace - MAXBSIZE; 561 562 lorunningspace = 512 * 1024; 563 hirunningspace = 1024 * 1024; 564 565/* 566 * Limit the amount of malloc memory since it is wired permanently into 567 * the kernel space. Even though this is accounted for in the buffer 568 * allocation, we don't want the malloced region to grow uncontrolled. 569 * The malloc scheme improves memory utilization significantly on average 570 * (small) directories. 571 */ 572 maxbufmallocspace = hibufspace / 20; 573 574/* 575 * Reduce the chance of a deadlock occuring by limiting the number 576 * of delayed-write dirty buffers we allow to stack up. 577 */ 578 hidirtybuffers = nbuf / 4 + 20; 579 dirtybufthresh = hidirtybuffers * 9 / 10; 580 numdirtybuffers = 0; 581/* 582 * To support extreme low-memory systems, make sure hidirtybuffers cannot 583 * eat up all available buffer space. This occurs when our minimum cannot 584 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 585 * BKVASIZE'd (8K) buffers. 586 */ 587 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 588 hidirtybuffers >>= 1; 589 } 590 lodirtybuffers = hidirtybuffers / 2; 591 592/* 593 * Try to keep the number of free buffers in the specified range, 594 * and give special processes (e.g. like buf_daemon) access to an 595 * emergency reserve. 596 */ 597 lofreebuffers = nbuf / 18 + 5; 598 hifreebuffers = 2 * lofreebuffers; 599 numfreebuffers = nbuf; 600 601/* 602 * Maximum number of async ops initiated per buf_daemon loop. This is 603 * somewhat of a hack at the moment, we really need to limit ourselves 604 * based on the number of bytes of I/O in-transit that were initiated 605 * from buf_daemon. 606 */ 607 608 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 609 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 610} 611 612/* 613 * bfreekva() - free the kva allocation for a buffer. 614 * 615 * Since this call frees up buffer space, we call bufspacewakeup(). 616 */ 617static void 618bfreekva(struct buf *bp) 619{ 620 621 if (bp->b_kvasize) { 622 atomic_add_int(&buffreekvacnt, 1); 623 atomic_subtract_int(&bufspace, bp->b_kvasize); 624 vm_map_lock(buffer_map); 625 vm_map_delete(buffer_map, 626 (vm_offset_t) bp->b_kvabase, 627 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 628 ); 629 vm_map_unlock(buffer_map); 630 bp->b_kvasize = 0; 631 bufspacewakeup(); 632 } 633} 634 635/* 636 * bremfree: 637 * 638 * Mark the buffer for removal from the appropriate free list in brelse. 639 * 640 */ 641void 642bremfree(struct buf *bp) 643{ 644 645 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 646 KASSERT(BUF_REFCNT(bp), ("bremfree: buf must be locked.")); 647 KASSERT((bp->b_flags & B_REMFREE) == 0 && bp->b_qindex != QUEUE_NONE, 648 ("bremfree: buffer %p not on a queue.", bp)); 649 650 bp->b_flags |= B_REMFREE; 651 /* Fixup numfreebuffers count. */ 652 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 653 atomic_subtract_int(&numfreebuffers, 1); 654} 655 656/* 657 * bremfreef: 658 * 659 * Force an immediate removal from a free list. Used only in nfs when 660 * it abuses the b_freelist pointer. 661 */ 662void 663bremfreef(struct buf *bp) 664{ 665 mtx_lock(&bqlock); 666 bremfreel(bp); 667 mtx_unlock(&bqlock); 668} 669 670/* 671 * bremfreel: 672 * 673 * Removes a buffer from the free list, must be called with the 674 * bqlock held. 675 */ 676static void 677bremfreel(struct buf *bp) 678{ 679 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 680 bp, bp->b_vp, bp->b_flags); 681 KASSERT(BUF_REFCNT(bp), ("bremfreel: buffer %p not locked.", bp)); 682 KASSERT(bp->b_qindex != QUEUE_NONE, 683 ("bremfreel: buffer %p not on a queue.", bp)); 684 mtx_assert(&bqlock, MA_OWNED); 685 686 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 687 bp->b_qindex = QUEUE_NONE; 688 /* 689 * If this was a delayed bremfree() we only need to remove the buffer 690 * from the queue and return the stats are already done. 691 */ 692 if (bp->b_flags & B_REMFREE) { 693 bp->b_flags &= ~B_REMFREE; 694 return; 695 } 696 /* 697 * Fixup numfreebuffers count. If the buffer is invalid or not 698 * delayed-write, the buffer was free and we must decrement 699 * numfreebuffers. 700 */ 701 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 702 atomic_subtract_int(&numfreebuffers, 1); 703} 704 705 706/* 707 * Get a buffer with the specified data. Look in the cache first. We 708 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 709 * is set, the buffer is valid and we do not have to do anything ( see 710 * getblk() ). This is really just a special case of breadn(). 711 */ 712int 713bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 714 struct buf **bpp) 715{ 716 717 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 718} 719 720/* 721 * Operates like bread, but also starts asynchronous I/O on 722 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 723 * to initiating I/O . If B_CACHE is set, the buffer is valid 724 * and we do not have to do anything. 725 */ 726int 727breadn(struct vnode * vp, daddr_t blkno, int size, 728 daddr_t * rablkno, int *rabsize, 729 int cnt, struct ucred * cred, struct buf **bpp) 730{ 731 struct buf *bp, *rabp; 732 int i; 733 int rv = 0, readwait = 0; 734 735 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 736 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 737 738 /* if not found in cache, do some I/O */ 739 if ((bp->b_flags & B_CACHE) == 0) { 740 if (curthread != PCPU_GET(idlethread)) 741 curthread->td_proc->p_stats->p_ru.ru_inblock++; 742 bp->b_iocmd = BIO_READ; 743 bp->b_flags &= ~B_INVAL; 744 bp->b_ioflags &= ~BIO_ERROR; 745 if (bp->b_rcred == NOCRED && cred != NOCRED) 746 bp->b_rcred = crhold(cred); 747 vfs_busy_pages(bp, 0); 748 bp->b_iooffset = dbtob(bp->b_blkno); 749 bstrategy(bp); 750 ++readwait; 751 } 752 753 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 754 if (inmem(vp, *rablkno)) 755 continue; 756 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 757 758 if ((rabp->b_flags & B_CACHE) == 0) { 759 if (curthread != PCPU_GET(idlethread)) 760 curthread->td_proc->p_stats->p_ru.ru_inblock++; 761 rabp->b_flags |= B_ASYNC; 762 rabp->b_flags &= ~B_INVAL; 763 rabp->b_ioflags &= ~BIO_ERROR; 764 rabp->b_iocmd = BIO_READ; 765 if (rabp->b_rcred == NOCRED && cred != NOCRED) 766 rabp->b_rcred = crhold(cred); 767 vfs_busy_pages(rabp, 0); 768 BUF_KERNPROC(rabp); 769 rabp->b_iooffset = dbtob(rabp->b_blkno); 770 bstrategy(rabp); 771 } else { 772 brelse(rabp); 773 } 774 } 775 776 if (readwait) { 777 rv = bufwait(bp); 778 } 779 return (rv); 780} 781 782/* 783 * Write, release buffer on completion. (Done by iodone 784 * if async). Do not bother writing anything if the buffer 785 * is invalid. 786 * 787 * Note that we set B_CACHE here, indicating that buffer is 788 * fully valid and thus cacheable. This is true even of NFS 789 * now so we set it generally. This could be set either here 790 * or in biodone() since the I/O is synchronous. We put it 791 * here. 792 */ 793int 794bufwrite(struct buf *bp) 795{ 796 int oldflags; 797 798 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 799 if (bp->b_flags & B_INVAL) { 800 brelse(bp); 801 return (0); 802 } 803 804 oldflags = bp->b_flags; 805 806 if (BUF_REFCNT(bp) == 0) 807 panic("bufwrite: buffer is not busy???"); 808 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 809 ("FFS background buffer should not get here %p", bp)); 810 811 /* Mark the buffer clean */ 812 bundirty(bp); 813 814 bp->b_flags &= ~B_DONE; 815 bp->b_ioflags &= ~BIO_ERROR; 816 bp->b_flags |= B_CACHE; 817 bp->b_iocmd = BIO_WRITE; 818 819 bufobj_wref(bp->b_bufobj); 820 vfs_busy_pages(bp, 1); 821 822 /* 823 * Normal bwrites pipeline writes 824 */ 825 bp->b_runningbufspace = bp->b_bufsize; 826 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 827 828 if (curthread != PCPU_GET(idlethread)) 829 curthread->td_proc->p_stats->p_ru.ru_oublock++; 830 if (oldflags & B_ASYNC) 831 BUF_KERNPROC(bp); 832 bp->b_iooffset = dbtob(bp->b_blkno); 833 bstrategy(bp); 834 835 if ((oldflags & B_ASYNC) == 0) { 836 int rtval = bufwait(bp); 837 brelse(bp); 838 return (rtval); 839 } else { 840 /* 841 * don't allow the async write to saturate the I/O 842 * system. We will not deadlock here because 843 * we are blocking waiting for I/O that is already in-progress 844 * to complete. We do not block here if it is the update 845 * or syncer daemon trying to clean up as that can lead 846 * to deadlock. 847 */ 848 if (curthread->td_proc != bufdaemonproc && 849 curthread->td_proc != updateproc) 850 waitrunningbufspace(); 851 } 852 853 return (0); 854} 855 856/* 857 * Delayed write. (Buffer is marked dirty). Do not bother writing 858 * anything if the buffer is marked invalid. 859 * 860 * Note that since the buffer must be completely valid, we can safely 861 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 862 * biodone() in order to prevent getblk from writing the buffer 863 * out synchronously. 864 */ 865void 866bdwrite(struct buf *bp) 867{ 868 struct thread *td = curthread; 869 struct vnode *vp; 870 struct buf *nbp; 871 struct bufobj *bo; 872 873 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 874 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 875 KASSERT(BUF_REFCNT(bp) != 0, ("bdwrite: buffer is not busy")); 876 877 if (bp->b_flags & B_INVAL) { 878 brelse(bp); 879 return; 880 } 881 882 /* 883 * If we have too many dirty buffers, don't create any more. 884 * If we are wildly over our limit, then force a complete 885 * cleanup. Otherwise, just keep the situation from getting 886 * out of control. Note that we have to avoid a recursive 887 * disaster and not try to clean up after our own cleanup! 888 */ 889 vp = bp->b_vp; 890 bo = bp->b_bufobj; 891 if ((td->td_pflags & TDP_COWINPROGRESS) == 0) { 892 BO_LOCK(bo); 893 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 894 BO_UNLOCK(bo); 895 (void) VOP_FSYNC(vp, MNT_NOWAIT, td); 896 altbufferflushes++; 897 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 898 /* 899 * Try to find a buffer to flush. 900 */ 901 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 902 if ((nbp->b_vflags & BV_BKGRDINPROG) || 903 BUF_LOCK(nbp, 904 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 905 continue; 906 if (bp == nbp) 907 panic("bdwrite: found ourselves"); 908 BO_UNLOCK(bo); 909 /* Don't countdeps with the bo lock held. */ 910 if (buf_countdeps(nbp, 0)) { 911 BO_LOCK(bo); 912 BUF_UNLOCK(nbp); 913 continue; 914 } 915 if (nbp->b_flags & B_CLUSTEROK) { 916 vfs_bio_awrite(nbp); 917 } else { 918 bremfree(nbp); 919 bawrite(nbp); 920 } 921 dirtybufferflushes++; 922 break; 923 } 924 if (nbp == NULL) 925 BO_UNLOCK(bo); 926 } else 927 BO_UNLOCK(bo); 928 } else 929 recursiveflushes++; 930 931 bdirty(bp); 932 /* 933 * Set B_CACHE, indicating that the buffer is fully valid. This is 934 * true even of NFS now. 935 */ 936 bp->b_flags |= B_CACHE; 937 938 /* 939 * This bmap keeps the system from needing to do the bmap later, 940 * perhaps when the system is attempting to do a sync. Since it 941 * is likely that the indirect block -- or whatever other datastructure 942 * that the filesystem needs is still in memory now, it is a good 943 * thing to do this. Note also, that if the pageout daemon is 944 * requesting a sync -- there might not be enough memory to do 945 * the bmap then... So, this is important to do. 946 */ 947 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 948 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 949 } 950 951 /* 952 * Set the *dirty* buffer range based upon the VM system dirty pages. 953 */ 954 vfs_setdirty(bp); 955 956 /* 957 * We need to do this here to satisfy the vnode_pager and the 958 * pageout daemon, so that it thinks that the pages have been 959 * "cleaned". Note that since the pages are in a delayed write 960 * buffer -- the VFS layer "will" see that the pages get written 961 * out on the next sync, or perhaps the cluster will be completed. 962 */ 963 vfs_clean_pages(bp); 964 bqrelse(bp); 965 966 /* 967 * Wakeup the buffer flushing daemon if we have a lot of dirty 968 * buffers (midpoint between our recovery point and our stall 969 * point). 970 */ 971 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 972 973 /* 974 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 975 * due to the softdep code. 976 */ 977} 978 979/* 980 * bdirty: 981 * 982 * Turn buffer into delayed write request. We must clear BIO_READ and 983 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 984 * itself to properly update it in the dirty/clean lists. We mark it 985 * B_DONE to ensure that any asynchronization of the buffer properly 986 * clears B_DONE ( else a panic will occur later ). 987 * 988 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 989 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 990 * should only be called if the buffer is known-good. 991 * 992 * Since the buffer is not on a queue, we do not update the numfreebuffers 993 * count. 994 * 995 * The buffer must be on QUEUE_NONE. 996 */ 997void 998bdirty(struct buf *bp) 999{ 1000 1001 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1002 bp, bp->b_vp, bp->b_flags); 1003 KASSERT(BUF_REFCNT(bp) == 1, ("bdirty: bp %p not locked",bp)); 1004 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1005 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1006 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1007 bp->b_flags &= ~(B_RELBUF); 1008 bp->b_iocmd = BIO_WRITE; 1009 1010 if ((bp->b_flags & B_DELWRI) == 0) { 1011 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1012 reassignbuf(bp); 1013 atomic_add_int(&numdirtybuffers, 1); 1014 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1015 } 1016} 1017 1018/* 1019 * bundirty: 1020 * 1021 * Clear B_DELWRI for buffer. 1022 * 1023 * Since the buffer is not on a queue, we do not update the numfreebuffers 1024 * count. 1025 * 1026 * The buffer must be on QUEUE_NONE. 1027 */ 1028 1029void 1030bundirty(struct buf *bp) 1031{ 1032 1033 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1034 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1035 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1036 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1037 KASSERT(BUF_REFCNT(bp) == 1, ("bundirty: bp %p not locked",bp)); 1038 1039 if (bp->b_flags & B_DELWRI) { 1040 bp->b_flags &= ~B_DELWRI; 1041 reassignbuf(bp); 1042 atomic_subtract_int(&numdirtybuffers, 1); 1043 numdirtywakeup(lodirtybuffers); 1044 } 1045 /* 1046 * Since it is now being written, we can clear its deferred write flag. 1047 */ 1048 bp->b_flags &= ~B_DEFERRED; 1049} 1050 1051/* 1052 * bawrite: 1053 * 1054 * Asynchronous write. Start output on a buffer, but do not wait for 1055 * it to complete. The buffer is released when the output completes. 1056 * 1057 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1058 * B_INVAL buffers. Not us. 1059 */ 1060void 1061bawrite(struct buf *bp) 1062{ 1063 1064 bp->b_flags |= B_ASYNC; 1065 (void) bwrite(bp); 1066} 1067 1068/* 1069 * bwillwrite: 1070 * 1071 * Called prior to the locking of any vnodes when we are expecting to 1072 * write. We do not want to starve the buffer cache with too many 1073 * dirty buffers so we block here. By blocking prior to the locking 1074 * of any vnodes we attempt to avoid the situation where a locked vnode 1075 * prevents the various system daemons from flushing related buffers. 1076 */ 1077 1078void 1079bwillwrite(void) 1080{ 1081 1082 if (numdirtybuffers >= hidirtybuffers) { 1083 mtx_lock(&nblock); 1084 while (numdirtybuffers >= hidirtybuffers) { 1085 bd_wakeup(1); 1086 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1087 msleep(&needsbuffer, &nblock, 1088 (PRIBIO + 4), "flswai", 0); 1089 } 1090 mtx_unlock(&nblock); 1091 } 1092} 1093 1094/* 1095 * Return true if we have too many dirty buffers. 1096 */ 1097int 1098buf_dirty_count_severe(void) 1099{ 1100 1101 return(numdirtybuffers >= hidirtybuffers); 1102} 1103 1104/* 1105 * brelse: 1106 * 1107 * Release a busy buffer and, if requested, free its resources. The 1108 * buffer will be stashed in the appropriate bufqueue[] allowing it 1109 * to be accessed later as a cache entity or reused for other purposes. 1110 */ 1111void 1112brelse(struct buf *bp) 1113{ 1114 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1115 bp, bp->b_vp, bp->b_flags); 1116 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1117 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1118 1119 if (bp->b_iocmd == BIO_WRITE && 1120 (bp->b_ioflags & BIO_ERROR) && 1121 !(bp->b_flags & B_INVAL)) { 1122 /* 1123 * Failed write, redirty. Must clear BIO_ERROR to prevent 1124 * pages from being scrapped. If B_INVAL is set then 1125 * this case is not run and the next case is run to 1126 * destroy the buffer. B_INVAL can occur if the buffer 1127 * is outside the range supported by the underlying device. 1128 */ 1129 bp->b_ioflags &= ~BIO_ERROR; 1130 bdirty(bp); 1131 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1132 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1133 /* 1134 * Either a failed I/O or we were asked to free or not 1135 * cache the buffer. 1136 */ 1137 bp->b_flags |= B_INVAL; 1138 if (LIST_FIRST(&bp->b_dep) != NULL) 1139 buf_deallocate(bp); 1140 if (bp->b_flags & B_DELWRI) { 1141 atomic_subtract_int(&numdirtybuffers, 1); 1142 numdirtywakeup(lodirtybuffers); 1143 } 1144 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1145 if ((bp->b_flags & B_VMIO) == 0) { 1146 if (bp->b_bufsize) 1147 allocbuf(bp, 0); 1148 if (bp->b_vp) 1149 brelvp(bp); 1150 } 1151 } 1152 1153 /* 1154 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1155 * is called with B_DELWRI set, the underlying pages may wind up 1156 * getting freed causing a previous write (bdwrite()) to get 'lost' 1157 * because pages associated with a B_DELWRI bp are marked clean. 1158 * 1159 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1160 * if B_DELWRI is set. 1161 * 1162 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1163 * on pages to return pages to the VM page queues. 1164 */ 1165 if (bp->b_flags & B_DELWRI) 1166 bp->b_flags &= ~B_RELBUF; 1167 else if (vm_page_count_severe()) { 1168 /* 1169 * XXX This lock may not be necessary since BKGRDINPROG 1170 * cannot be set while we hold the buf lock, it can only be 1171 * cleared if it is already pending. 1172 */ 1173 if (bp->b_vp) { 1174 BO_LOCK(bp->b_bufobj); 1175 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1176 bp->b_flags |= B_RELBUF; 1177 BO_UNLOCK(bp->b_bufobj); 1178 } else 1179 bp->b_flags |= B_RELBUF; 1180 } 1181 1182 /* 1183 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1184 * constituted, not even NFS buffers now. Two flags effect this. If 1185 * B_INVAL, the struct buf is invalidated but the VM object is kept 1186 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1187 * 1188 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1189 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1190 * buffer is also B_INVAL because it hits the re-dirtying code above. 1191 * 1192 * Normally we can do this whether a buffer is B_DELWRI or not. If 1193 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1194 * the commit state and we cannot afford to lose the buffer. If the 1195 * buffer has a background write in progress, we need to keep it 1196 * around to prevent it from being reconstituted and starting a second 1197 * background write. 1198 */ 1199 if ((bp->b_flags & B_VMIO) 1200 && !(bp->b_vp->v_mount != NULL && 1201 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1202 !vn_isdisk(bp->b_vp, NULL) && 1203 (bp->b_flags & B_DELWRI)) 1204 ) { 1205 1206 int i, j, resid; 1207 vm_page_t m; 1208 off_t foff; 1209 vm_pindex_t poff; 1210 vm_object_t obj; 1211 1212 obj = bp->b_bufobj->bo_object; 1213 1214 /* 1215 * Get the base offset and length of the buffer. Note that 1216 * in the VMIO case if the buffer block size is not 1217 * page-aligned then b_data pointer may not be page-aligned. 1218 * But our b_pages[] array *IS* page aligned. 1219 * 1220 * block sizes less then DEV_BSIZE (usually 512) are not 1221 * supported due to the page granularity bits (m->valid, 1222 * m->dirty, etc...). 1223 * 1224 * See man buf(9) for more information 1225 */ 1226 resid = bp->b_bufsize; 1227 foff = bp->b_offset; 1228 VM_OBJECT_LOCK(obj); 1229 for (i = 0; i < bp->b_npages; i++) { 1230 int had_bogus = 0; 1231 1232 m = bp->b_pages[i]; 1233 1234 /* 1235 * If we hit a bogus page, fixup *all* the bogus pages 1236 * now. 1237 */ 1238 if (m == bogus_page) { 1239 poff = OFF_TO_IDX(bp->b_offset); 1240 had_bogus = 1; 1241 1242 for (j = i; j < bp->b_npages; j++) { 1243 vm_page_t mtmp; 1244 mtmp = bp->b_pages[j]; 1245 if (mtmp == bogus_page) { 1246 mtmp = vm_page_lookup(obj, poff + j); 1247 if (!mtmp) { 1248 panic("brelse: page missing\n"); 1249 } 1250 bp->b_pages[j] = mtmp; 1251 } 1252 } 1253 1254 if ((bp->b_flags & B_INVAL) == 0) { 1255 pmap_qenter( 1256 trunc_page((vm_offset_t)bp->b_data), 1257 bp->b_pages, bp->b_npages); 1258 } 1259 m = bp->b_pages[i]; 1260 } 1261 if ((bp->b_flags & B_NOCACHE) || 1262 (bp->b_ioflags & BIO_ERROR)) { 1263 int poffset = foff & PAGE_MASK; 1264 int presid = resid > (PAGE_SIZE - poffset) ? 1265 (PAGE_SIZE - poffset) : resid; 1266 1267 KASSERT(presid >= 0, ("brelse: extra page")); 1268 vm_page_lock_queues(); 1269 vm_page_set_invalid(m, poffset, presid); 1270 vm_page_unlock_queues(); 1271 if (had_bogus) 1272 printf("avoided corruption bug in bogus_page/brelse code\n"); 1273 } 1274 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1275 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1276 } 1277 VM_OBJECT_UNLOCK(obj); 1278 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1279 vfs_vmio_release(bp); 1280 1281 } else if (bp->b_flags & B_VMIO) { 1282 1283 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1284 vfs_vmio_release(bp); 1285 } 1286 1287 } 1288 1289 if (BUF_REFCNT(bp) > 1) { 1290 /* do not release to free list */ 1291 BUF_UNLOCK(bp); 1292 return; 1293 } 1294 1295 /* enqueue */ 1296 mtx_lock(&bqlock); 1297 /* Handle delayed bremfree() processing. */ 1298 if (bp->b_flags & B_REMFREE) 1299 bremfreel(bp); 1300 if (bp->b_qindex != QUEUE_NONE) 1301 panic("brelse: free buffer onto another queue???"); 1302 1303 /* buffers with no memory */ 1304 if (bp->b_bufsize == 0) { 1305 bp->b_flags |= B_INVAL; 1306 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1307 if (bp->b_vflags & BV_BKGRDINPROG) 1308 panic("losing buffer 1"); 1309 if (bp->b_kvasize) { 1310 bp->b_qindex = QUEUE_EMPTYKVA; 1311 } else { 1312 bp->b_qindex = QUEUE_EMPTY; 1313 } 1314 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1315 /* buffers with junk contents */ 1316 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1317 (bp->b_ioflags & BIO_ERROR)) { 1318 bp->b_flags |= B_INVAL; 1319 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1320 if (bp->b_vflags & BV_BKGRDINPROG) 1321 panic("losing buffer 2"); 1322 bp->b_qindex = QUEUE_CLEAN; 1323 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1324 /* remaining buffers */ 1325 } else { 1326 if (bp->b_flags & B_DELWRI) 1327 bp->b_qindex = QUEUE_DIRTY; 1328 else 1329 bp->b_qindex = QUEUE_CLEAN; 1330 if (bp->b_flags & B_AGE) 1331 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1332 else 1333 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1334 } 1335 mtx_unlock(&bqlock); 1336 1337 /* 1338 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1339 * placed the buffer on the correct queue. We must also disassociate 1340 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1341 * find it. 1342 */ 1343 if (bp->b_flags & B_INVAL) { 1344 if (bp->b_flags & B_DELWRI) 1345 bundirty(bp); 1346 if (bp->b_vp) 1347 brelvp(bp); 1348 } 1349 1350 /* 1351 * Fixup numfreebuffers count. The bp is on an appropriate queue 1352 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1353 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1354 * if B_INVAL is set ). 1355 */ 1356 1357 if (!(bp->b_flags & B_DELWRI)) 1358 bufcountwakeup(); 1359 1360 /* 1361 * Something we can maybe free or reuse 1362 */ 1363 if (bp->b_bufsize || bp->b_kvasize) 1364 bufspacewakeup(); 1365 1366 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1367 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1368 panic("brelse: not dirty"); 1369 /* unlock */ 1370 BUF_UNLOCK(bp); 1371} 1372 1373/* 1374 * Release a buffer back to the appropriate queue but do not try to free 1375 * it. The buffer is expected to be used again soon. 1376 * 1377 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1378 * biodone() to requeue an async I/O on completion. It is also used when 1379 * known good buffers need to be requeued but we think we may need the data 1380 * again soon. 1381 * 1382 * XXX we should be able to leave the B_RELBUF hint set on completion. 1383 */ 1384void 1385bqrelse(struct buf *bp) 1386{ 1387 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1388 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1389 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1390 1391 if (BUF_REFCNT(bp) > 1) { 1392 /* do not release to free list */ 1393 BUF_UNLOCK(bp); 1394 return; 1395 } 1396 mtx_lock(&bqlock); 1397 /* Handle delayed bremfree() processing. */ 1398 if (bp->b_flags & B_REMFREE) 1399 bremfreel(bp); 1400 if (bp->b_qindex != QUEUE_NONE) 1401 panic("bqrelse: free buffer onto another queue???"); 1402 /* buffers with stale but valid contents */ 1403 if (bp->b_flags & B_DELWRI) { 1404 bp->b_qindex = QUEUE_DIRTY; 1405 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1406 } else { 1407 /* 1408 * XXX This lock may not be necessary since BKGRDINPROG 1409 * cannot be set while we hold the buf lock, it can only be 1410 * cleared if it is already pending. 1411 */ 1412 BO_LOCK(bp->b_bufobj); 1413 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1414 BO_UNLOCK(bp->b_bufobj); 1415 bp->b_qindex = QUEUE_CLEAN; 1416 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1417 b_freelist); 1418 } else { 1419 /* 1420 * We are too low on memory, we have to try to free 1421 * the buffer (most importantly: the wired pages 1422 * making up its backing store) *now*. 1423 */ 1424 BO_UNLOCK(bp->b_bufobj); 1425 mtx_unlock(&bqlock); 1426 brelse(bp); 1427 return; 1428 } 1429 } 1430 mtx_unlock(&bqlock); 1431 1432 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1433 bufcountwakeup(); 1434 1435 /* 1436 * Something we can maybe free or reuse. 1437 */ 1438 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1439 bufspacewakeup(); 1440 1441 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1442 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1443 panic("bqrelse: not dirty"); 1444 /* unlock */ 1445 BUF_UNLOCK(bp); 1446} 1447 1448/* Give pages used by the bp back to the VM system (where possible) */ 1449static void 1450vfs_vmio_release(struct buf *bp) 1451{ 1452 int i; 1453 vm_page_t m; 1454 1455 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1456 vm_page_lock_queues(); 1457 for (i = 0; i < bp->b_npages; i++) { 1458 m = bp->b_pages[i]; 1459 bp->b_pages[i] = NULL; 1460 /* 1461 * In order to keep page LRU ordering consistent, put 1462 * everything on the inactive queue. 1463 */ 1464 vm_page_unwire(m, 0); 1465 /* 1466 * We don't mess with busy pages, it is 1467 * the responsibility of the process that 1468 * busied the pages to deal with them. 1469 */ 1470 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1471 continue; 1472 1473 if (m->wire_count == 0) { 1474 /* 1475 * Might as well free the page if we can and it has 1476 * no valid data. We also free the page if the 1477 * buffer was used for direct I/O 1478 */ 1479 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1480 m->hold_count == 0) { 1481 pmap_remove_all(m); 1482 vm_page_free(m); 1483 } else if (bp->b_flags & B_DIRECT) { 1484 vm_page_try_to_free(m); 1485 } else if (vm_page_count_severe()) { 1486 vm_page_try_to_cache(m); 1487 } 1488 } 1489 } 1490 vm_page_unlock_queues(); 1491 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1492 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1493 1494 if (bp->b_bufsize) { 1495 bufspacewakeup(); 1496 bp->b_bufsize = 0; 1497 } 1498 bp->b_npages = 0; 1499 bp->b_flags &= ~B_VMIO; 1500 if (bp->b_vp) 1501 brelvp(bp); 1502} 1503 1504/* 1505 * Check to see if a block at a particular lbn is available for a clustered 1506 * write. 1507 */ 1508static int 1509vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1510{ 1511 struct buf *bpa; 1512 int match; 1513 1514 match = 0; 1515 1516 /* If the buf isn't in core skip it */ 1517 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1518 return (0); 1519 1520 /* If the buf is busy we don't want to wait for it */ 1521 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1522 return (0); 1523 1524 /* Only cluster with valid clusterable delayed write buffers */ 1525 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1526 (B_DELWRI | B_CLUSTEROK)) 1527 goto done; 1528 1529 if (bpa->b_bufsize != size) 1530 goto done; 1531 1532 /* 1533 * Check to see if it is in the expected place on disk and that the 1534 * block has been mapped. 1535 */ 1536 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1537 match = 1; 1538done: 1539 BUF_UNLOCK(bpa); 1540 return (match); 1541} 1542 1543/* 1544 * vfs_bio_awrite: 1545 * 1546 * Implement clustered async writes for clearing out B_DELWRI buffers. 1547 * This is much better then the old way of writing only one buffer at 1548 * a time. Note that we may not be presented with the buffers in the 1549 * correct order, so we search for the cluster in both directions. 1550 */ 1551int 1552vfs_bio_awrite(struct buf *bp) 1553{ 1554 int i; 1555 int j; 1556 daddr_t lblkno = bp->b_lblkno; 1557 struct vnode *vp = bp->b_vp; 1558 int ncl; 1559 int nwritten; 1560 int size; 1561 int maxcl; 1562 1563 /* 1564 * right now we support clustered writing only to regular files. If 1565 * we find a clusterable block we could be in the middle of a cluster 1566 * rather then at the beginning. 1567 */ 1568 if ((vp->v_type == VREG) && 1569 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1570 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1571 1572 size = vp->v_mount->mnt_stat.f_iosize; 1573 maxcl = MAXPHYS / size; 1574 1575 VI_LOCK(vp); 1576 for (i = 1; i < maxcl; i++) 1577 if (vfs_bio_clcheck(vp, size, lblkno + i, 1578 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1579 break; 1580 1581 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1582 if (vfs_bio_clcheck(vp, size, lblkno - j, 1583 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1584 break; 1585 1586 VI_UNLOCK(vp); 1587 --j; 1588 ncl = i + j; 1589 /* 1590 * this is a possible cluster write 1591 */ 1592 if (ncl != 1) { 1593 BUF_UNLOCK(bp); 1594 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1595 return nwritten; 1596 } 1597 } 1598 bremfree(bp); 1599 bp->b_flags |= B_ASYNC; 1600 /* 1601 * default (old) behavior, writing out only one block 1602 * 1603 * XXX returns b_bufsize instead of b_bcount for nwritten? 1604 */ 1605 nwritten = bp->b_bufsize; 1606 (void) bwrite(bp); 1607 1608 return nwritten; 1609} 1610 1611/* 1612 * getnewbuf: 1613 * 1614 * Find and initialize a new buffer header, freeing up existing buffers 1615 * in the bufqueues as necessary. The new buffer is returned locked. 1616 * 1617 * Important: B_INVAL is not set. If the caller wishes to throw the 1618 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1619 * 1620 * We block if: 1621 * We have insufficient buffer headers 1622 * We have insufficient buffer space 1623 * buffer_map is too fragmented ( space reservation fails ) 1624 * If we have to flush dirty buffers ( but we try to avoid this ) 1625 * 1626 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1627 * Instead we ask the buf daemon to do it for us. We attempt to 1628 * avoid piecemeal wakeups of the pageout daemon. 1629 */ 1630 1631static struct buf * 1632getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1633{ 1634 struct buf *bp; 1635 struct buf *nbp; 1636 int defrag = 0; 1637 int nqindex; 1638 static int flushingbufs; 1639 1640 /* 1641 * We can't afford to block since we might be holding a vnode lock, 1642 * which may prevent system daemons from running. We deal with 1643 * low-memory situations by proactively returning memory and running 1644 * async I/O rather then sync I/O. 1645 */ 1646 1647 atomic_add_int(&getnewbufcalls, 1); 1648 atomic_subtract_int(&getnewbufrestarts, 1); 1649restart: 1650 atomic_add_int(&getnewbufrestarts, 1); 1651 1652 /* 1653 * Setup for scan. If we do not have enough free buffers, 1654 * we setup a degenerate case that immediately fails. Note 1655 * that if we are specially marked process, we are allowed to 1656 * dip into our reserves. 1657 * 1658 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1659 * 1660 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1661 * However, there are a number of cases (defragging, reusing, ...) 1662 * where we cannot backup. 1663 */ 1664 mtx_lock(&bqlock); 1665 nqindex = QUEUE_EMPTYKVA; 1666 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1667 1668 if (nbp == NULL) { 1669 /* 1670 * If no EMPTYKVA buffers and we are either 1671 * defragging or reusing, locate a CLEAN buffer 1672 * to free or reuse. If bufspace useage is low 1673 * skip this step so we can allocate a new buffer. 1674 */ 1675 if (defrag || bufspace >= lobufspace) { 1676 nqindex = QUEUE_CLEAN; 1677 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1678 } 1679 1680 /* 1681 * If we could not find or were not allowed to reuse a 1682 * CLEAN buffer, check to see if it is ok to use an EMPTY 1683 * buffer. We can only use an EMPTY buffer if allocating 1684 * its KVA would not otherwise run us out of buffer space. 1685 */ 1686 if (nbp == NULL && defrag == 0 && 1687 bufspace + maxsize < hibufspace) { 1688 nqindex = QUEUE_EMPTY; 1689 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1690 } 1691 } 1692 1693 /* 1694 * Run scan, possibly freeing data and/or kva mappings on the fly 1695 * depending. 1696 */ 1697 1698 while ((bp = nbp) != NULL) { 1699 int qindex = nqindex; 1700 1701 /* 1702 * Calculate next bp ( we can only use it if we do not block 1703 * or do other fancy things ). 1704 */ 1705 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1706 switch(qindex) { 1707 case QUEUE_EMPTY: 1708 nqindex = QUEUE_EMPTYKVA; 1709 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1710 break; 1711 /* FALLTHROUGH */ 1712 case QUEUE_EMPTYKVA: 1713 nqindex = QUEUE_CLEAN; 1714 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1715 break; 1716 /* FALLTHROUGH */ 1717 case QUEUE_CLEAN: 1718 /* 1719 * nbp is NULL. 1720 */ 1721 break; 1722 } 1723 } 1724 /* 1725 * If we are defragging then we need a buffer with 1726 * b_kvasize != 0. XXX this situation should no longer 1727 * occur, if defrag is non-zero the buffer's b_kvasize 1728 * should also be non-zero at this point. XXX 1729 */ 1730 if (defrag && bp->b_kvasize == 0) { 1731 printf("Warning: defrag empty buffer %p\n", bp); 1732 continue; 1733 } 1734 1735 /* 1736 * Start freeing the bp. This is somewhat involved. nbp 1737 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1738 */ 1739 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1740 continue; 1741 if (bp->b_vp) { 1742 BO_LOCK(bp->b_bufobj); 1743 if (bp->b_vflags & BV_BKGRDINPROG) { 1744 BO_UNLOCK(bp->b_bufobj); 1745 BUF_UNLOCK(bp); 1746 continue; 1747 } 1748 BO_UNLOCK(bp->b_bufobj); 1749 } 1750 CTR6(KTR_BUF, 1751 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1752 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1753 bp->b_kvasize, bp->b_bufsize, qindex); 1754 1755 /* 1756 * Sanity Checks 1757 */ 1758 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1759 1760 /* 1761 * Note: we no longer distinguish between VMIO and non-VMIO 1762 * buffers. 1763 */ 1764 1765 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1766 1767 bremfreel(bp); 1768 mtx_unlock(&bqlock); 1769 1770 if (qindex == QUEUE_CLEAN) { 1771 if (bp->b_flags & B_VMIO) { 1772 bp->b_flags &= ~B_ASYNC; 1773 vfs_vmio_release(bp); 1774 } 1775 if (bp->b_vp) 1776 brelvp(bp); 1777 } 1778 1779 /* 1780 * NOTE: nbp is now entirely invalid. We can only restart 1781 * the scan from this point on. 1782 * 1783 * Get the rest of the buffer freed up. b_kva* is still 1784 * valid after this operation. 1785 */ 1786 1787 if (bp->b_rcred != NOCRED) { 1788 crfree(bp->b_rcred); 1789 bp->b_rcred = NOCRED; 1790 } 1791 if (bp->b_wcred != NOCRED) { 1792 crfree(bp->b_wcred); 1793 bp->b_wcred = NOCRED; 1794 } 1795 if (LIST_FIRST(&bp->b_dep) != NULL) 1796 buf_deallocate(bp); 1797 if (bp->b_vflags & BV_BKGRDINPROG) 1798 panic("losing buffer 3"); 1799 KASSERT(bp->b_vp == NULL, 1800 ("bp: %p still has vnode %p. qindex: %d", 1801 bp, bp->b_vp, qindex)); 1802 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1803 ("bp: %p still on a buffer list. xflags %X", 1804 bp, bp->b_xflags)); 1805 1806 if (bp->b_bufsize) 1807 allocbuf(bp, 0); 1808 1809 bp->b_flags = 0; 1810 bp->b_ioflags = 0; 1811 bp->b_xflags = 0; 1812 bp->b_vflags = 0; 1813 bp->b_vp = NULL; 1814 bp->b_blkno = bp->b_lblkno = 0; 1815 bp->b_offset = NOOFFSET; 1816 bp->b_iodone = 0; 1817 bp->b_error = 0; 1818 bp->b_resid = 0; 1819 bp->b_bcount = 0; 1820 bp->b_npages = 0; 1821 bp->b_dirtyoff = bp->b_dirtyend = 0; 1822 bp->b_bufobj = NULL; 1823 1824 LIST_INIT(&bp->b_dep); 1825 1826 /* 1827 * If we are defragging then free the buffer. 1828 */ 1829 if (defrag) { 1830 bp->b_flags |= B_INVAL; 1831 bfreekva(bp); 1832 brelse(bp); 1833 defrag = 0; 1834 goto restart; 1835 } 1836 1837 /* 1838 * If we are overcomitted then recover the buffer and its 1839 * KVM space. This occurs in rare situations when multiple 1840 * processes are blocked in getnewbuf() or allocbuf(). 1841 */ 1842 if (bufspace >= hibufspace) 1843 flushingbufs = 1; 1844 if (flushingbufs && bp->b_kvasize != 0) { 1845 bp->b_flags |= B_INVAL; 1846 bfreekva(bp); 1847 brelse(bp); 1848 goto restart; 1849 } 1850 if (bufspace < lobufspace) 1851 flushingbufs = 0; 1852 break; 1853 } 1854 1855 /* 1856 * If we exhausted our list, sleep as appropriate. We may have to 1857 * wakeup various daemons and write out some dirty buffers. 1858 * 1859 * Generally we are sleeping due to insufficient buffer space. 1860 */ 1861 1862 if (bp == NULL) { 1863 int flags; 1864 char *waitmsg; 1865 1866 mtx_unlock(&bqlock); 1867 if (defrag) { 1868 flags = VFS_BIO_NEED_BUFSPACE; 1869 waitmsg = "nbufkv"; 1870 } else if (bufspace >= hibufspace) { 1871 waitmsg = "nbufbs"; 1872 flags = VFS_BIO_NEED_BUFSPACE; 1873 } else { 1874 waitmsg = "newbuf"; 1875 flags = VFS_BIO_NEED_ANY; 1876 } 1877 1878 bd_speedup(); /* heeeelp */ 1879 1880 mtx_lock(&nblock); 1881 needsbuffer |= flags; 1882 while (needsbuffer & flags) { 1883 if (msleep(&needsbuffer, &nblock, 1884 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1885 mtx_unlock(&nblock); 1886 return (NULL); 1887 } 1888 } 1889 mtx_unlock(&nblock); 1890 } else { 1891 /* 1892 * We finally have a valid bp. We aren't quite out of the 1893 * woods, we still have to reserve kva space. In order 1894 * to keep fragmentation sane we only allocate kva in 1895 * BKVASIZE chunks. 1896 */ 1897 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1898 1899 if (maxsize != bp->b_kvasize) { 1900 vm_offset_t addr = 0; 1901 1902 bfreekva(bp); 1903 1904 vm_map_lock(buffer_map); 1905 if (vm_map_findspace(buffer_map, 1906 vm_map_min(buffer_map), maxsize, &addr)) { 1907 /* 1908 * Uh oh. Buffer map is to fragmented. We 1909 * must defragment the map. 1910 */ 1911 atomic_add_int(&bufdefragcnt, 1); 1912 vm_map_unlock(buffer_map); 1913 defrag = 1; 1914 bp->b_flags |= B_INVAL; 1915 brelse(bp); 1916 goto restart; 1917 } 1918 if (addr) { 1919 vm_map_insert(buffer_map, NULL, 0, 1920 addr, addr + maxsize, 1921 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1922 1923 bp->b_kvabase = (caddr_t) addr; 1924 bp->b_kvasize = maxsize; 1925 atomic_add_int(&bufspace, bp->b_kvasize); 1926 atomic_add_int(&bufreusecnt, 1); 1927 } 1928 vm_map_unlock(buffer_map); 1929 } 1930 bp->b_saveaddr = bp->b_kvabase; 1931 bp->b_data = bp->b_saveaddr; 1932 } 1933 return(bp); 1934} 1935 1936/* 1937 * buf_daemon: 1938 * 1939 * buffer flushing daemon. Buffers are normally flushed by the 1940 * update daemon but if it cannot keep up this process starts to 1941 * take the load in an attempt to prevent getnewbuf() from blocking. 1942 */ 1943 1944static struct kproc_desc buf_kp = { 1945 "bufdaemon", 1946 buf_daemon, 1947 &bufdaemonproc 1948}; 1949SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1950 1951static void 1952buf_daemon() 1953{ 1954 mtx_lock(&Giant); 1955 1956 /* 1957 * This process needs to be suspended prior to shutdown sync. 1958 */ 1959 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 1960 SHUTDOWN_PRI_LAST); 1961 1962 /* 1963 * This process is allowed to take the buffer cache to the limit 1964 */ 1965 mtx_lock(&bdlock); 1966 for (;;) { 1967 bd_request = 0; 1968 mtx_unlock(&bdlock); 1969 1970 kthread_suspend_check(bufdaemonproc); 1971 1972 /* 1973 * Do the flush. Limit the amount of in-transit I/O we 1974 * allow to build up, otherwise we would completely saturate 1975 * the I/O system. Wakeup any waiting processes before we 1976 * normally would so they can run in parallel with our drain. 1977 */ 1978 while (numdirtybuffers > lodirtybuffers) { 1979 if (flushbufqueues(0) == 0) { 1980 /* 1981 * Could not find any buffers without rollback 1982 * dependencies, so just write the first one 1983 * in the hopes of eventually making progress. 1984 */ 1985 flushbufqueues(1); 1986 break; 1987 } 1988 waitrunningbufspace(); 1989 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 1990 } 1991 1992 /* 1993 * Only clear bd_request if we have reached our low water 1994 * mark. The buf_daemon normally waits 1 second and 1995 * then incrementally flushes any dirty buffers that have 1996 * built up, within reason. 1997 * 1998 * If we were unable to hit our low water mark and couldn't 1999 * find any flushable buffers, we sleep half a second. 2000 * Otherwise we loop immediately. 2001 */ 2002 mtx_lock(&bdlock); 2003 if (numdirtybuffers <= lodirtybuffers) { 2004 /* 2005 * We reached our low water mark, reset the 2006 * request and sleep until we are needed again. 2007 * The sleep is just so the suspend code works. 2008 */ 2009 bd_request = 0; 2010 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2011 } else { 2012 /* 2013 * We couldn't find any flushable dirty buffers but 2014 * still have too many dirty buffers, we 2015 * have to sleep and try again. (rare) 2016 */ 2017 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2018 } 2019 } 2020} 2021 2022/* 2023 * flushbufqueues: 2024 * 2025 * Try to flush a buffer in the dirty queue. We must be careful to 2026 * free up B_INVAL buffers instead of write them, which NFS is 2027 * particularly sensitive to. 2028 */ 2029static int flushwithdeps = 0; 2030SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2031 0, "Number of buffers flushed with dependecies that require rollbacks"); 2032 2033static int 2034flushbufqueues(int flushdeps) 2035{ 2036 struct thread *td = curthread; 2037 struct vnode *vp; 2038 struct mount *mp; 2039 struct buf *bp; 2040 int hasdeps; 2041 2042 mtx_lock(&bqlock); 2043 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2044 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2045 continue; 2046 BO_LOCK(bp->b_bufobj); 2047 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2048 (bp->b_flags & B_DELWRI) == 0) { 2049 BO_UNLOCK(bp->b_bufobj); 2050 BUF_UNLOCK(bp); 2051 continue; 2052 } 2053 BO_UNLOCK(bp->b_bufobj); 2054 if (bp->b_flags & B_INVAL) { 2055 bremfreel(bp); 2056 mtx_unlock(&bqlock); 2057 brelse(bp); 2058 return (1); 2059 } 2060 2061 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2062 if (flushdeps == 0) { 2063 BUF_UNLOCK(bp); 2064 continue; 2065 } 2066 hasdeps = 1; 2067 } else 2068 hasdeps = 0; 2069 /* 2070 * We must hold the lock on a vnode before writing 2071 * one of its buffers. Otherwise we may confuse, or 2072 * in the case of a snapshot vnode, deadlock the 2073 * system. 2074 * 2075 * The lock order here is the reverse of the normal 2076 * of vnode followed by buf lock. This is ok because 2077 * the NOWAIT will prevent deadlock. 2078 */ 2079 vp = bp->b_vp; 2080 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2081 BUF_UNLOCK(bp); 2082 continue; 2083 } 2084 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2085 mtx_unlock(&bqlock); 2086 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2087 bp, bp->b_vp, bp->b_flags); 2088 vfs_bio_awrite(bp); 2089 vn_finished_write(mp); 2090 VOP_UNLOCK(vp, 0, td); 2091 flushwithdeps += hasdeps; 2092 return (1); 2093 } 2094 vn_finished_write(mp); 2095 BUF_UNLOCK(bp); 2096 } 2097 mtx_unlock(&bqlock); 2098 return (0); 2099} 2100 2101/* 2102 * Check to see if a block is currently memory resident. 2103 */ 2104struct buf * 2105incore(struct bufobj *bo, daddr_t blkno) 2106{ 2107 struct buf *bp; 2108 2109 BO_LOCK(bo); 2110 bp = gbincore(bo, blkno); 2111 BO_UNLOCK(bo); 2112 return (bp); 2113} 2114 2115/* 2116 * Returns true if no I/O is needed to access the 2117 * associated VM object. This is like incore except 2118 * it also hunts around in the VM system for the data. 2119 */ 2120 2121static int 2122inmem(struct vnode * vp, daddr_t blkno) 2123{ 2124 vm_object_t obj; 2125 vm_offset_t toff, tinc, size; 2126 vm_page_t m; 2127 vm_ooffset_t off; 2128 2129 ASSERT_VOP_LOCKED(vp, "inmem"); 2130 2131 if (incore(&vp->v_bufobj, blkno)) 2132 return 1; 2133 if (vp->v_mount == NULL) 2134 return 0; 2135 obj = vp->v_object; 2136 if (obj == NULL) 2137 return (0); 2138 2139 size = PAGE_SIZE; 2140 if (size > vp->v_mount->mnt_stat.f_iosize) 2141 size = vp->v_mount->mnt_stat.f_iosize; 2142 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2143 2144 VM_OBJECT_LOCK(obj); 2145 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2146 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2147 if (!m) 2148 goto notinmem; 2149 tinc = size; 2150 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2151 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2152 if (vm_page_is_valid(m, 2153 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2154 goto notinmem; 2155 } 2156 VM_OBJECT_UNLOCK(obj); 2157 return 1; 2158 2159notinmem: 2160 VM_OBJECT_UNLOCK(obj); 2161 return (0); 2162} 2163 2164/* 2165 * vfs_setdirty: 2166 * 2167 * Sets the dirty range for a buffer based on the status of the dirty 2168 * bits in the pages comprising the buffer. 2169 * 2170 * The range is limited to the size of the buffer. 2171 * 2172 * This routine is primarily used by NFS, but is generalized for the 2173 * B_VMIO case. 2174 */ 2175static void 2176vfs_setdirty(struct buf *bp) 2177{ 2178 int i; 2179 vm_object_t object; 2180 2181 /* 2182 * Degenerate case - empty buffer 2183 */ 2184 2185 if (bp->b_bufsize == 0) 2186 return; 2187 2188 /* 2189 * We qualify the scan for modified pages on whether the 2190 * object has been flushed yet. The OBJ_WRITEABLE flag 2191 * is not cleared simply by protecting pages off. 2192 */ 2193 2194 if ((bp->b_flags & B_VMIO) == 0) 2195 return; 2196 2197 object = bp->b_pages[0]->object; 2198 VM_OBJECT_LOCK(object); 2199 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2200 printf("Warning: object %p writeable but not mightbedirty\n", object); 2201 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2202 printf("Warning: object %p mightbedirty but not writeable\n", object); 2203 2204 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2205 vm_offset_t boffset; 2206 vm_offset_t eoffset; 2207 2208 vm_page_lock_queues(); 2209 /* 2210 * test the pages to see if they have been modified directly 2211 * by users through the VM system. 2212 */ 2213 for (i = 0; i < bp->b_npages; i++) 2214 vm_page_test_dirty(bp->b_pages[i]); 2215 2216 /* 2217 * Calculate the encompassing dirty range, boffset and eoffset, 2218 * (eoffset - boffset) bytes. 2219 */ 2220 2221 for (i = 0; i < bp->b_npages; i++) { 2222 if (bp->b_pages[i]->dirty) 2223 break; 2224 } 2225 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2226 2227 for (i = bp->b_npages - 1; i >= 0; --i) { 2228 if (bp->b_pages[i]->dirty) { 2229 break; 2230 } 2231 } 2232 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2233 2234 vm_page_unlock_queues(); 2235 /* 2236 * Fit it to the buffer. 2237 */ 2238 2239 if (eoffset > bp->b_bcount) 2240 eoffset = bp->b_bcount; 2241 2242 /* 2243 * If we have a good dirty range, merge with the existing 2244 * dirty range. 2245 */ 2246 2247 if (boffset < eoffset) { 2248 if (bp->b_dirtyoff > boffset) 2249 bp->b_dirtyoff = boffset; 2250 if (bp->b_dirtyend < eoffset) 2251 bp->b_dirtyend = eoffset; 2252 } 2253 } 2254 VM_OBJECT_UNLOCK(object); 2255} 2256 2257/* 2258 * getblk: 2259 * 2260 * Get a block given a specified block and offset into a file/device. 2261 * The buffers B_DONE bit will be cleared on return, making it almost 2262 * ready for an I/O initiation. B_INVAL may or may not be set on 2263 * return. The caller should clear B_INVAL prior to initiating a 2264 * READ. 2265 * 2266 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2267 * an existing buffer. 2268 * 2269 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2270 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2271 * and then cleared based on the backing VM. If the previous buffer is 2272 * non-0-sized but invalid, B_CACHE will be cleared. 2273 * 2274 * If getblk() must create a new buffer, the new buffer is returned with 2275 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2276 * case it is returned with B_INVAL clear and B_CACHE set based on the 2277 * backing VM. 2278 * 2279 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2280 * B_CACHE bit is clear. 2281 * 2282 * What this means, basically, is that the caller should use B_CACHE to 2283 * determine whether the buffer is fully valid or not and should clear 2284 * B_INVAL prior to issuing a read. If the caller intends to validate 2285 * the buffer by loading its data area with something, the caller needs 2286 * to clear B_INVAL. If the caller does this without issuing an I/O, 2287 * the caller should set B_CACHE ( as an optimization ), else the caller 2288 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2289 * a write attempt or if it was a successfull read. If the caller 2290 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2291 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2292 */ 2293struct buf * 2294getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2295 int flags) 2296{ 2297 struct buf *bp; 2298 struct bufobj *bo; 2299 int error; 2300 2301 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2302 ASSERT_VOP_LOCKED(vp, "getblk"); 2303 if (size > MAXBSIZE) 2304 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2305 2306 bo = &vp->v_bufobj; 2307loop: 2308 /* 2309 * Block if we are low on buffers. Certain processes are allowed 2310 * to completely exhaust the buffer cache. 2311 * 2312 * If this check ever becomes a bottleneck it may be better to 2313 * move it into the else, when gbincore() fails. At the moment 2314 * it isn't a problem. 2315 * 2316 * XXX remove if 0 sections (clean this up after its proven) 2317 */ 2318 if (numfreebuffers == 0) { 2319 if (curthread == PCPU_GET(idlethread)) 2320 return NULL; 2321 mtx_lock(&nblock); 2322 needsbuffer |= VFS_BIO_NEED_ANY; 2323 mtx_unlock(&nblock); 2324 } 2325 2326 VI_LOCK(vp); 2327 bp = gbincore(bo, blkno); 2328 if (bp != NULL) { 2329 int lockflags; 2330 /* 2331 * Buffer is in-core. If the buffer is not busy, it must 2332 * be on a queue. 2333 */ 2334 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2335 2336 if (flags & GB_LOCK_NOWAIT) 2337 lockflags |= LK_NOWAIT; 2338 2339 error = BUF_TIMELOCK(bp, lockflags, 2340 VI_MTX(vp), "getblk", slpflag, slptimeo); 2341 2342 /* 2343 * If we slept and got the lock we have to restart in case 2344 * the buffer changed identities. 2345 */ 2346 if (error == ENOLCK) 2347 goto loop; 2348 /* We timed out or were interrupted. */ 2349 else if (error) 2350 return (NULL); 2351 2352 /* 2353 * The buffer is locked. B_CACHE is cleared if the buffer is 2354 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2355 * and for a VMIO buffer B_CACHE is adjusted according to the 2356 * backing VM cache. 2357 */ 2358 if (bp->b_flags & B_INVAL) 2359 bp->b_flags &= ~B_CACHE; 2360 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2361 bp->b_flags |= B_CACHE; 2362 bremfree(bp); 2363 2364 /* 2365 * check for size inconsistancies for non-VMIO case. 2366 */ 2367 2368 if (bp->b_bcount != size) { 2369 if ((bp->b_flags & B_VMIO) == 0 || 2370 (size > bp->b_kvasize)) { 2371 if (bp->b_flags & B_DELWRI) { 2372 bp->b_flags |= B_NOCACHE; 2373 bwrite(bp); 2374 } else { 2375 if ((bp->b_flags & B_VMIO) && 2376 (LIST_FIRST(&bp->b_dep) == NULL)) { 2377 bp->b_flags |= B_RELBUF; 2378 brelse(bp); 2379 } else { 2380 bp->b_flags |= B_NOCACHE; 2381 bwrite(bp); 2382 } 2383 } 2384 goto loop; 2385 } 2386 } 2387 2388 /* 2389 * If the size is inconsistant in the VMIO case, we can resize 2390 * the buffer. This might lead to B_CACHE getting set or 2391 * cleared. If the size has not changed, B_CACHE remains 2392 * unchanged from its previous state. 2393 */ 2394 2395 if (bp->b_bcount != size) 2396 allocbuf(bp, size); 2397 2398 KASSERT(bp->b_offset != NOOFFSET, 2399 ("getblk: no buffer offset")); 2400 2401 /* 2402 * A buffer with B_DELWRI set and B_CACHE clear must 2403 * be committed before we can return the buffer in 2404 * order to prevent the caller from issuing a read 2405 * ( due to B_CACHE not being set ) and overwriting 2406 * it. 2407 * 2408 * Most callers, including NFS and FFS, need this to 2409 * operate properly either because they assume they 2410 * can issue a read if B_CACHE is not set, or because 2411 * ( for example ) an uncached B_DELWRI might loop due 2412 * to softupdates re-dirtying the buffer. In the latter 2413 * case, B_CACHE is set after the first write completes, 2414 * preventing further loops. 2415 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2416 * above while extending the buffer, we cannot allow the 2417 * buffer to remain with B_CACHE set after the write 2418 * completes or it will represent a corrupt state. To 2419 * deal with this we set B_NOCACHE to scrap the buffer 2420 * after the write. 2421 * 2422 * We might be able to do something fancy, like setting 2423 * B_CACHE in bwrite() except if B_DELWRI is already set, 2424 * so the below call doesn't set B_CACHE, but that gets real 2425 * confusing. This is much easier. 2426 */ 2427 2428 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2429 bp->b_flags |= B_NOCACHE; 2430 bwrite(bp); 2431 goto loop; 2432 } 2433 bp->b_flags &= ~B_DONE; 2434 } else { 2435 int bsize, maxsize, vmio; 2436 off_t offset; 2437 2438 /* 2439 * Buffer is not in-core, create new buffer. The buffer 2440 * returned by getnewbuf() is locked. Note that the returned 2441 * buffer is also considered valid (not marked B_INVAL). 2442 */ 2443 VI_UNLOCK(vp); 2444 /* 2445 * If the user does not want us to create the buffer, bail out 2446 * here. 2447 */ 2448 if (flags & GB_NOCREAT) 2449 return NULL; 2450 bsize = bo->bo_bsize; 2451 offset = blkno * bsize; 2452 vmio = vp->v_object != NULL; 2453 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2454 maxsize = imax(maxsize, bsize); 2455 2456 bp = getnewbuf(slpflag, slptimeo, size, maxsize); 2457 if (bp == NULL) { 2458 if (slpflag || slptimeo) 2459 return NULL; 2460 goto loop; 2461 } 2462 2463 /* 2464 * This code is used to make sure that a buffer is not 2465 * created while the getnewbuf routine is blocked. 2466 * This can be a problem whether the vnode is locked or not. 2467 * If the buffer is created out from under us, we have to 2468 * throw away the one we just created. 2469 * 2470 * Note: this must occur before we associate the buffer 2471 * with the vp especially considering limitations in 2472 * the splay tree implementation when dealing with duplicate 2473 * lblkno's. 2474 */ 2475 BO_LOCK(bo); 2476 if (gbincore(bo, blkno)) { 2477 BO_UNLOCK(bo); 2478 bp->b_flags |= B_INVAL; 2479 brelse(bp); 2480 goto loop; 2481 } 2482 2483 /* 2484 * Insert the buffer into the hash, so that it can 2485 * be found by incore. 2486 */ 2487 bp->b_blkno = bp->b_lblkno = blkno; 2488 bp->b_offset = offset; 2489 2490 bgetvp(vp, bp); 2491 BO_UNLOCK(bo); 2492 2493 /* 2494 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2495 * buffer size starts out as 0, B_CACHE will be set by 2496 * allocbuf() for the VMIO case prior to it testing the 2497 * backing store for validity. 2498 */ 2499 2500 if (vmio) { 2501 bp->b_flags |= B_VMIO; 2502#if defined(VFS_BIO_DEBUG) 2503 if (vn_canvmio(vp) != TRUE) 2504 printf("getblk: VMIO on vnode type %d\n", 2505 vp->v_type); 2506#endif 2507 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2508 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2509 bp, vp->v_object, bp->b_bufobj->bo_object)); 2510 } else { 2511 bp->b_flags &= ~B_VMIO; 2512 KASSERT(bp->b_bufobj->bo_object == NULL, 2513 ("ARGH! has b_bufobj->bo_object %p %p\n", 2514 bp, bp->b_bufobj->bo_object)); 2515 } 2516 2517 allocbuf(bp, size); 2518 bp->b_flags &= ~B_DONE; 2519 } 2520 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2521 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2522 KASSERT(bp->b_bufobj == bo, 2523 ("wrong b_bufobj %p should be %p", bp->b_bufobj, bo)); 2524 return (bp); 2525} 2526 2527/* 2528 * Get an empty, disassociated buffer of given size. The buffer is initially 2529 * set to B_INVAL. 2530 */ 2531struct buf * 2532geteblk(int size) 2533{ 2534 struct buf *bp; 2535 int maxsize; 2536 2537 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2538 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2539 continue; 2540 allocbuf(bp, size); 2541 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2542 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2543 return (bp); 2544} 2545 2546 2547/* 2548 * This code constitutes the buffer memory from either anonymous system 2549 * memory (in the case of non-VMIO operations) or from an associated 2550 * VM object (in the case of VMIO operations). This code is able to 2551 * resize a buffer up or down. 2552 * 2553 * Note that this code is tricky, and has many complications to resolve 2554 * deadlock or inconsistant data situations. Tread lightly!!! 2555 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2556 * the caller. Calling this code willy nilly can result in the loss of data. 2557 * 2558 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2559 * B_CACHE for the non-VMIO case. 2560 */ 2561 2562int 2563allocbuf(struct buf *bp, int size) 2564{ 2565 int newbsize, mbsize; 2566 int i; 2567 2568 if (BUF_REFCNT(bp) == 0) 2569 panic("allocbuf: buffer not busy"); 2570 2571 if (bp->b_kvasize < size) 2572 panic("allocbuf: buffer too small"); 2573 2574 if ((bp->b_flags & B_VMIO) == 0) { 2575 caddr_t origbuf; 2576 int origbufsize; 2577 /* 2578 * Just get anonymous memory from the kernel. Don't 2579 * mess with B_CACHE. 2580 */ 2581 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2582 if (bp->b_flags & B_MALLOC) 2583 newbsize = mbsize; 2584 else 2585 newbsize = round_page(size); 2586 2587 if (newbsize < bp->b_bufsize) { 2588 /* 2589 * malloced buffers are not shrunk 2590 */ 2591 if (bp->b_flags & B_MALLOC) { 2592 if (newbsize) { 2593 bp->b_bcount = size; 2594 } else { 2595 free(bp->b_data, M_BIOBUF); 2596 if (bp->b_bufsize) { 2597 atomic_subtract_int( 2598 &bufmallocspace, 2599 bp->b_bufsize); 2600 bufspacewakeup(); 2601 bp->b_bufsize = 0; 2602 } 2603 bp->b_saveaddr = bp->b_kvabase; 2604 bp->b_data = bp->b_saveaddr; 2605 bp->b_bcount = 0; 2606 bp->b_flags &= ~B_MALLOC; 2607 } 2608 return 1; 2609 } 2610 vm_hold_free_pages( 2611 bp, 2612 (vm_offset_t) bp->b_data + newbsize, 2613 (vm_offset_t) bp->b_data + bp->b_bufsize); 2614 } else if (newbsize > bp->b_bufsize) { 2615 /* 2616 * We only use malloced memory on the first allocation. 2617 * and revert to page-allocated memory when the buffer 2618 * grows. 2619 */ 2620 /* 2621 * There is a potential smp race here that could lead 2622 * to bufmallocspace slightly passing the max. It 2623 * is probably extremely rare and not worth worrying 2624 * over. 2625 */ 2626 if ( (bufmallocspace < maxbufmallocspace) && 2627 (bp->b_bufsize == 0) && 2628 (mbsize <= PAGE_SIZE/2)) { 2629 2630 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2631 bp->b_bufsize = mbsize; 2632 bp->b_bcount = size; 2633 bp->b_flags |= B_MALLOC; 2634 atomic_add_int(&bufmallocspace, mbsize); 2635 return 1; 2636 } 2637 origbuf = NULL; 2638 origbufsize = 0; 2639 /* 2640 * If the buffer is growing on its other-than-first allocation, 2641 * then we revert to the page-allocation scheme. 2642 */ 2643 if (bp->b_flags & B_MALLOC) { 2644 origbuf = bp->b_data; 2645 origbufsize = bp->b_bufsize; 2646 bp->b_data = bp->b_kvabase; 2647 if (bp->b_bufsize) { 2648 atomic_subtract_int(&bufmallocspace, 2649 bp->b_bufsize); 2650 bufspacewakeup(); 2651 bp->b_bufsize = 0; 2652 } 2653 bp->b_flags &= ~B_MALLOC; 2654 newbsize = round_page(newbsize); 2655 } 2656 vm_hold_load_pages( 2657 bp, 2658 (vm_offset_t) bp->b_data + bp->b_bufsize, 2659 (vm_offset_t) bp->b_data + newbsize); 2660 if (origbuf) { 2661 bcopy(origbuf, bp->b_data, origbufsize); 2662 free(origbuf, M_BIOBUF); 2663 } 2664 } 2665 } else { 2666 int desiredpages; 2667 2668 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2669 desiredpages = (size == 0) ? 0 : 2670 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2671 2672 if (bp->b_flags & B_MALLOC) 2673 panic("allocbuf: VMIO buffer can't be malloced"); 2674 /* 2675 * Set B_CACHE initially if buffer is 0 length or will become 2676 * 0-length. 2677 */ 2678 if (size == 0 || bp->b_bufsize == 0) 2679 bp->b_flags |= B_CACHE; 2680 2681 if (newbsize < bp->b_bufsize) { 2682 /* 2683 * DEV_BSIZE aligned new buffer size is less then the 2684 * DEV_BSIZE aligned existing buffer size. Figure out 2685 * if we have to remove any pages. 2686 */ 2687 if (desiredpages < bp->b_npages) { 2688 vm_page_t m; 2689 2690 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2691 vm_page_lock_queues(); 2692 for (i = desiredpages; i < bp->b_npages; i++) { 2693 /* 2694 * the page is not freed here -- it 2695 * is the responsibility of 2696 * vnode_pager_setsize 2697 */ 2698 m = bp->b_pages[i]; 2699 KASSERT(m != bogus_page, 2700 ("allocbuf: bogus page found")); 2701 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2702 vm_page_lock_queues(); 2703 2704 bp->b_pages[i] = NULL; 2705 vm_page_unwire(m, 0); 2706 } 2707 vm_page_unlock_queues(); 2708 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2709 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2710 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2711 bp->b_npages = desiredpages; 2712 } 2713 } else if (size > bp->b_bcount) { 2714 /* 2715 * We are growing the buffer, possibly in a 2716 * byte-granular fashion. 2717 */ 2718 struct vnode *vp; 2719 vm_object_t obj; 2720 vm_offset_t toff; 2721 vm_offset_t tinc; 2722 2723 /* 2724 * Step 1, bring in the VM pages from the object, 2725 * allocating them if necessary. We must clear 2726 * B_CACHE if these pages are not valid for the 2727 * range covered by the buffer. 2728 */ 2729 2730 vp = bp->b_vp; 2731 obj = bp->b_bufobj->bo_object; 2732 2733 VM_OBJECT_LOCK(obj); 2734 while (bp->b_npages < desiredpages) { 2735 vm_page_t m; 2736 vm_pindex_t pi; 2737 2738 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2739 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2740 /* 2741 * note: must allocate system pages 2742 * since blocking here could intefere 2743 * with paging I/O, no matter which 2744 * process we are. 2745 */ 2746 m = vm_page_alloc(obj, pi, 2747 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2748 VM_ALLOC_WIRED); 2749 if (m == NULL) { 2750 atomic_add_int(&vm_pageout_deficit, 2751 desiredpages - bp->b_npages); 2752 VM_OBJECT_UNLOCK(obj); 2753 VM_WAIT; 2754 VM_OBJECT_LOCK(obj); 2755 } else { 2756 bp->b_flags &= ~B_CACHE; 2757 bp->b_pages[bp->b_npages] = m; 2758 ++bp->b_npages; 2759 } 2760 continue; 2761 } 2762 2763 /* 2764 * We found a page. If we have to sleep on it, 2765 * retry because it might have gotten freed out 2766 * from under us. 2767 * 2768 * We can only test PG_BUSY here. Blocking on 2769 * m->busy might lead to a deadlock: 2770 * 2771 * vm_fault->getpages->cluster_read->allocbuf 2772 * 2773 */ 2774 vm_page_lock_queues(); 2775 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2776 continue; 2777 2778 /* 2779 * We have a good page. Should we wakeup the 2780 * page daemon? 2781 */ 2782 if ((curproc != pageproc) && 2783 ((m->queue - m->pc) == PQ_CACHE) && 2784 ((cnt.v_free_count + cnt.v_cache_count) < 2785 (cnt.v_free_min + cnt.v_cache_min))) { 2786 pagedaemon_wakeup(); 2787 } 2788 vm_page_wire(m); 2789 vm_page_unlock_queues(); 2790 bp->b_pages[bp->b_npages] = m; 2791 ++bp->b_npages; 2792 } 2793 2794 /* 2795 * Step 2. We've loaded the pages into the buffer, 2796 * we have to figure out if we can still have B_CACHE 2797 * set. Note that B_CACHE is set according to the 2798 * byte-granular range ( bcount and size ), new the 2799 * aligned range ( newbsize ). 2800 * 2801 * The VM test is against m->valid, which is DEV_BSIZE 2802 * aligned. Needless to say, the validity of the data 2803 * needs to also be DEV_BSIZE aligned. Note that this 2804 * fails with NFS if the server or some other client 2805 * extends the file's EOF. If our buffer is resized, 2806 * B_CACHE may remain set! XXX 2807 */ 2808 2809 toff = bp->b_bcount; 2810 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2811 2812 while ((bp->b_flags & B_CACHE) && toff < size) { 2813 vm_pindex_t pi; 2814 2815 if (tinc > (size - toff)) 2816 tinc = size - toff; 2817 2818 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2819 PAGE_SHIFT; 2820 2821 vfs_buf_test_cache( 2822 bp, 2823 bp->b_offset, 2824 toff, 2825 tinc, 2826 bp->b_pages[pi] 2827 ); 2828 toff += tinc; 2829 tinc = PAGE_SIZE; 2830 } 2831 VM_OBJECT_UNLOCK(obj); 2832 2833 /* 2834 * Step 3, fixup the KVM pmap. Remember that 2835 * bp->b_data is relative to bp->b_offset, but 2836 * bp->b_offset may be offset into the first page. 2837 */ 2838 2839 bp->b_data = (caddr_t) 2840 trunc_page((vm_offset_t)bp->b_data); 2841 pmap_qenter( 2842 (vm_offset_t)bp->b_data, 2843 bp->b_pages, 2844 bp->b_npages 2845 ); 2846 2847 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2848 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2849 } 2850 } 2851 if (newbsize < bp->b_bufsize) 2852 bufspacewakeup(); 2853 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2854 bp->b_bcount = size; /* requested buffer size */ 2855 return 1; 2856} 2857 2858void 2859biodone(struct bio *bp) 2860{ 2861 2862 mtx_lock(&bdonelock); 2863 bp->bio_flags |= BIO_DONE; 2864 if (bp->bio_done == NULL) 2865 wakeup(bp); 2866 mtx_unlock(&bdonelock); 2867 if (bp->bio_done != NULL) 2868 bp->bio_done(bp); 2869} 2870 2871/* 2872 * Wait for a BIO to finish. 2873 * 2874 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2875 * case is not yet clear. 2876 */ 2877int 2878biowait(struct bio *bp, const char *wchan) 2879{ 2880 2881 mtx_lock(&bdonelock); 2882 while ((bp->bio_flags & BIO_DONE) == 0) 2883 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 2884 mtx_unlock(&bdonelock); 2885 if (bp->bio_error != 0) 2886 return (bp->bio_error); 2887 if (!(bp->bio_flags & BIO_ERROR)) 2888 return (0); 2889 return (EIO); 2890} 2891 2892void 2893biofinish(struct bio *bp, struct devstat *stat, int error) 2894{ 2895 2896 if (error) { 2897 bp->bio_error = error; 2898 bp->bio_flags |= BIO_ERROR; 2899 } 2900 if (stat != NULL) 2901 devstat_end_transaction_bio(stat, bp); 2902 biodone(bp); 2903} 2904 2905/* 2906 * bufwait: 2907 * 2908 * Wait for buffer I/O completion, returning error status. The buffer 2909 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 2910 * error and cleared. 2911 */ 2912int 2913bufwait(struct buf *bp) 2914{ 2915 if (bp->b_iocmd == BIO_READ) 2916 bwait(bp, PRIBIO, "biord"); 2917 else 2918 bwait(bp, PRIBIO, "biowr"); 2919 if (bp->b_flags & B_EINTR) { 2920 bp->b_flags &= ~B_EINTR; 2921 return (EINTR); 2922 } 2923 if (bp->b_ioflags & BIO_ERROR) { 2924 return (bp->b_error ? bp->b_error : EIO); 2925 } else { 2926 return (0); 2927 } 2928} 2929 2930 /* 2931 * Call back function from struct bio back up to struct buf. 2932 */ 2933static void 2934bufdonebio(struct bio *bip) 2935{ 2936 struct buf *bp; 2937 2938 bp = bip->bio_caller2; 2939 bp->b_resid = bp->b_bcount - bip->bio_completed; 2940 bp->b_resid = bip->bio_resid; /* XXX: remove */ 2941 bp->b_ioflags = bip->bio_flags; 2942 bp->b_error = bip->bio_error; 2943 if (bp->b_error) 2944 bp->b_ioflags |= BIO_ERROR; 2945 bufdone(bp); 2946 g_destroy_bio(bip); 2947} 2948 2949void 2950dev_strategy(struct cdev *dev, struct buf *bp) 2951{ 2952 struct cdevsw *csw; 2953 struct bio *bip; 2954 2955 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 2956 panic("b_iocmd botch"); 2957 for (;;) { 2958 bip = g_new_bio(); 2959 if (bip != NULL) 2960 break; 2961 /* Try again later */ 2962 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 2963 } 2964 bip->bio_cmd = bp->b_iocmd; 2965 bip->bio_offset = bp->b_iooffset; 2966 bip->bio_length = bp->b_bcount; 2967 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 2968 bip->bio_data = bp->b_data; 2969 bip->bio_done = bufdonebio; 2970 bip->bio_caller2 = bp; 2971 bip->bio_dev = dev; 2972 KASSERT(dev->si_refcount > 0, 2973 ("dev_strategy on un-referenced struct cdev *(%s)", 2974 devtoname(dev))); 2975 csw = dev_refthread(dev); 2976 if (csw == NULL) { 2977 bp->b_error = ENXIO; 2978 bp->b_ioflags = BIO_ERROR; 2979 bufdone(bp); 2980 return; 2981 } 2982 (*csw->d_strategy)(bip); 2983 dev_relthread(dev); 2984} 2985 2986/* 2987 * bufdone: 2988 * 2989 * Finish I/O on a buffer, optionally calling a completion function. 2990 * This is usually called from an interrupt so process blocking is 2991 * not allowed. 2992 * 2993 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2994 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2995 * assuming B_INVAL is clear. 2996 * 2997 * For the VMIO case, we set B_CACHE if the op was a read and no 2998 * read error occured, or if the op was a write. B_CACHE is never 2999 * set if the buffer is invalid or otherwise uncacheable. 3000 * 3001 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3002 * initiator to leave B_INVAL set to brelse the buffer out of existance 3003 * in the biodone routine. 3004 */ 3005void 3006bufdone(struct buf *bp) 3007{ 3008 struct bufobj *dropobj; 3009 void (*biodone)(struct buf *); 3010 3011 3012 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3013 dropobj = NULL; 3014 3015 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3016 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3017 3018 runningbufwakeup(bp); 3019 if (bp->b_iocmd == BIO_WRITE) 3020 dropobj = bp->b_bufobj; 3021 /* call optional completion function if requested */ 3022 if (bp->b_iodone != NULL) { 3023 biodone = bp->b_iodone; 3024 bp->b_iodone = NULL; 3025 bp->b_flags |= B_DONE; 3026 (*biodone) (bp); 3027 if (dropobj) 3028 bufobj_wdrop(dropobj); 3029 return; 3030 } 3031 if (LIST_FIRST(&bp->b_dep) != NULL) 3032 buf_complete(bp); 3033 3034 if (bp->b_flags & B_VMIO) { 3035 int i; 3036 vm_ooffset_t foff; 3037 vm_page_t m; 3038 vm_object_t obj; 3039 int iosize; 3040 struct vnode *vp = bp->b_vp; 3041 3042 obj = bp->b_bufobj->bo_object; 3043 3044#if defined(VFS_BIO_DEBUG) 3045 mp_fixme("usecount and vflag accessed without locks."); 3046 if (vp->v_usecount == 0) { 3047 panic("biodone: zero vnode ref count"); 3048 } 3049 3050 KASSERT(vp->v_object != NULL, 3051 ("biodone: vnode %p has no vm_object", vp)); 3052#endif 3053 3054 foff = bp->b_offset; 3055 KASSERT(bp->b_offset != NOOFFSET, 3056 ("biodone: no buffer offset")); 3057 3058 VM_OBJECT_LOCK(obj); 3059#if defined(VFS_BIO_DEBUG) 3060 if (obj->paging_in_progress < bp->b_npages) { 3061 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3062 obj->paging_in_progress, bp->b_npages); 3063 } 3064#endif 3065 3066 /* 3067 * Set B_CACHE if the op was a normal read and no error 3068 * occured. B_CACHE is set for writes in the b*write() 3069 * routines. 3070 */ 3071 iosize = bp->b_bcount - bp->b_resid; 3072 if (bp->b_iocmd == BIO_READ && 3073 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3074 !(bp->b_ioflags & BIO_ERROR)) { 3075 bp->b_flags |= B_CACHE; 3076 } 3077 vm_page_lock_queues(); 3078 for (i = 0; i < bp->b_npages; i++) { 3079 int bogusflag = 0; 3080 int resid; 3081 3082 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3083 if (resid > iosize) 3084 resid = iosize; 3085 3086 /* 3087 * cleanup bogus pages, restoring the originals 3088 */ 3089 m = bp->b_pages[i]; 3090 if (m == bogus_page) { 3091 bogusflag = 1; 3092 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3093 if (m == NULL) 3094 panic("biodone: page disappeared!"); 3095 bp->b_pages[i] = m; 3096 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3097 } 3098#if defined(VFS_BIO_DEBUG) 3099 if (OFF_TO_IDX(foff) != m->pindex) { 3100 printf( 3101"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3102 (intmax_t)foff, (uintmax_t)m->pindex); 3103 } 3104#endif 3105 3106 /* 3107 * In the write case, the valid and clean bits are 3108 * already changed correctly ( see bdwrite() ), so we 3109 * only need to do this here in the read case. 3110 */ 3111 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3112 vfs_page_set_valid(bp, foff, i, m); 3113 } 3114 3115 /* 3116 * when debugging new filesystems or buffer I/O methods, this 3117 * is the most common error that pops up. if you see this, you 3118 * have not set the page busy flag correctly!!! 3119 */ 3120 if (m->busy == 0) { 3121 printf("biodone: page busy < 0, " 3122 "pindex: %d, foff: 0x(%x,%x), " 3123 "resid: %d, index: %d\n", 3124 (int) m->pindex, (int)(foff >> 32), 3125 (int) foff & 0xffffffff, resid, i); 3126 if (!vn_isdisk(vp, NULL)) 3127 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3128 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3129 (intmax_t) bp->b_lblkno, 3130 bp->b_flags, bp->b_npages); 3131 else 3132 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3133 (intmax_t) bp->b_lblkno, 3134 bp->b_flags, bp->b_npages); 3135 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3136 (u_long)m->valid, (u_long)m->dirty, 3137 m->wire_count); 3138 panic("biodone: page busy < 0\n"); 3139 } 3140 vm_page_io_finish(m); 3141 vm_object_pip_subtract(obj, 1); 3142 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3143 iosize -= resid; 3144 } 3145 vm_page_unlock_queues(); 3146 vm_object_pip_wakeupn(obj, 0); 3147 VM_OBJECT_UNLOCK(obj); 3148 } 3149 3150 /* 3151 * For asynchronous completions, release the buffer now. The brelse 3152 * will do a wakeup there if necessary - so no need to do a wakeup 3153 * here in the async case. The sync case always needs to do a wakeup. 3154 */ 3155 3156 if (bp->b_flags & B_ASYNC) { 3157 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3158 brelse(bp); 3159 else 3160 bqrelse(bp); 3161 } else 3162 bdone(bp); 3163 if (dropobj) 3164 bufobj_wdrop(dropobj); 3165} 3166 3167/* 3168 * This routine is called in lieu of iodone in the case of 3169 * incomplete I/O. This keeps the busy status for pages 3170 * consistant. 3171 */ 3172void 3173vfs_unbusy_pages(struct buf *bp) 3174{ 3175 int i; 3176 vm_object_t obj; 3177 vm_page_t m; 3178 3179 runningbufwakeup(bp); 3180 if (!(bp->b_flags & B_VMIO)) 3181 return; 3182 3183 obj = bp->b_bufobj->bo_object; 3184 VM_OBJECT_LOCK(obj); 3185 vm_page_lock_queues(); 3186 for (i = 0; i < bp->b_npages; i++) { 3187 m = bp->b_pages[i]; 3188 if (m == bogus_page) { 3189 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3190 if (!m) 3191 panic("vfs_unbusy_pages: page missing\n"); 3192 bp->b_pages[i] = m; 3193 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3194 bp->b_pages, bp->b_npages); 3195 } 3196 vm_object_pip_subtract(obj, 1); 3197 vm_page_io_finish(m); 3198 } 3199 vm_page_unlock_queues(); 3200 vm_object_pip_wakeupn(obj, 0); 3201 VM_OBJECT_UNLOCK(obj); 3202} 3203 3204/* 3205 * vfs_page_set_valid: 3206 * 3207 * Set the valid bits in a page based on the supplied offset. The 3208 * range is restricted to the buffer's size. 3209 * 3210 * This routine is typically called after a read completes. 3211 */ 3212static void 3213vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3214{ 3215 vm_ooffset_t soff, eoff; 3216 3217 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3218 /* 3219 * Start and end offsets in buffer. eoff - soff may not cross a 3220 * page boundry or cross the end of the buffer. The end of the 3221 * buffer, in this case, is our file EOF, not the allocation size 3222 * of the buffer. 3223 */ 3224 soff = off; 3225 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3226 if (eoff > bp->b_offset + bp->b_bcount) 3227 eoff = bp->b_offset + bp->b_bcount; 3228 3229 /* 3230 * Set valid range. This is typically the entire buffer and thus the 3231 * entire page. 3232 */ 3233 if (eoff > soff) { 3234 vm_page_set_validclean( 3235 m, 3236 (vm_offset_t) (soff & PAGE_MASK), 3237 (vm_offset_t) (eoff - soff) 3238 ); 3239 } 3240} 3241 3242/* 3243 * This routine is called before a device strategy routine. 3244 * It is used to tell the VM system that paging I/O is in 3245 * progress, and treat the pages associated with the buffer 3246 * almost as being PG_BUSY. Also the object paging_in_progress 3247 * flag is handled to make sure that the object doesn't become 3248 * inconsistant. 3249 * 3250 * Since I/O has not been initiated yet, certain buffer flags 3251 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3252 * and should be ignored. 3253 */ 3254void 3255vfs_busy_pages(struct buf *bp, int clear_modify) 3256{ 3257 int i, bogus; 3258 vm_object_t obj; 3259 vm_ooffset_t foff; 3260 vm_page_t m; 3261 3262 if (!(bp->b_flags & B_VMIO)) 3263 return; 3264 3265 obj = bp->b_bufobj->bo_object; 3266 foff = bp->b_offset; 3267 KASSERT(bp->b_offset != NOOFFSET, 3268 ("vfs_busy_pages: no buffer offset")); 3269 vfs_setdirty(bp); 3270 VM_OBJECT_LOCK(obj); 3271retry: 3272 vm_page_lock_queues(); 3273 for (i = 0; i < bp->b_npages; i++) { 3274 m = bp->b_pages[i]; 3275 3276 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3277 goto retry; 3278 } 3279 bogus = 0; 3280 for (i = 0; i < bp->b_npages; i++) { 3281 m = bp->b_pages[i]; 3282 3283 if ((bp->b_flags & B_CLUSTER) == 0) { 3284 vm_object_pip_add(obj, 1); 3285 vm_page_io_start(m); 3286 } 3287 /* 3288 * When readying a buffer for a read ( i.e 3289 * clear_modify == 0 ), it is important to do 3290 * bogus_page replacement for valid pages in 3291 * partially instantiated buffers. Partially 3292 * instantiated buffers can, in turn, occur when 3293 * reconstituting a buffer from its VM backing store 3294 * base. We only have to do this if B_CACHE is 3295 * clear ( which causes the I/O to occur in the 3296 * first place ). The replacement prevents the read 3297 * I/O from overwriting potentially dirty VM-backed 3298 * pages. XXX bogus page replacement is, uh, bogus. 3299 * It may not work properly with small-block devices. 3300 * We need to find a better way. 3301 */ 3302 pmap_remove_all(m); 3303 if (clear_modify) 3304 vfs_page_set_valid(bp, foff, i, m); 3305 else if (m->valid == VM_PAGE_BITS_ALL && 3306 (bp->b_flags & B_CACHE) == 0) { 3307 bp->b_pages[i] = bogus_page; 3308 bogus++; 3309 } 3310 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3311 } 3312 vm_page_unlock_queues(); 3313 VM_OBJECT_UNLOCK(obj); 3314 if (bogus) 3315 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3316 bp->b_pages, bp->b_npages); 3317} 3318 3319/* 3320 * Tell the VM system that the pages associated with this buffer 3321 * are clean. This is used for delayed writes where the data is 3322 * going to go to disk eventually without additional VM intevention. 3323 * 3324 * Note that while we only really need to clean through to b_bcount, we 3325 * just go ahead and clean through to b_bufsize. 3326 */ 3327static void 3328vfs_clean_pages(struct buf *bp) 3329{ 3330 int i; 3331 vm_ooffset_t foff, noff, eoff; 3332 vm_page_t m; 3333 3334 if (!(bp->b_flags & B_VMIO)) 3335 return; 3336 3337 foff = bp->b_offset; 3338 KASSERT(bp->b_offset != NOOFFSET, 3339 ("vfs_clean_pages: no buffer offset")); 3340 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3341 vm_page_lock_queues(); 3342 for (i = 0; i < bp->b_npages; i++) { 3343 m = bp->b_pages[i]; 3344 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3345 eoff = noff; 3346 3347 if (eoff > bp->b_offset + bp->b_bufsize) 3348 eoff = bp->b_offset + bp->b_bufsize; 3349 vfs_page_set_valid(bp, foff, i, m); 3350 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3351 foff = noff; 3352 } 3353 vm_page_unlock_queues(); 3354 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3355} 3356 3357/* 3358 * vfs_bio_set_validclean: 3359 * 3360 * Set the range within the buffer to valid and clean. The range is 3361 * relative to the beginning of the buffer, b_offset. Note that b_offset 3362 * itself may be offset from the beginning of the first page. 3363 * 3364 */ 3365 3366void 3367vfs_bio_set_validclean(struct buf *bp, int base, int size) 3368{ 3369 int i, n; 3370 vm_page_t m; 3371 3372 if (!(bp->b_flags & B_VMIO)) 3373 return; 3374 /* 3375 * Fixup base to be relative to beginning of first page. 3376 * Set initial n to be the maximum number of bytes in the 3377 * first page that can be validated. 3378 */ 3379 3380 base += (bp->b_offset & PAGE_MASK); 3381 n = PAGE_SIZE - (base & PAGE_MASK); 3382 3383 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3384 vm_page_lock_queues(); 3385 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3386 m = bp->b_pages[i]; 3387 if (n > size) 3388 n = size; 3389 vm_page_set_validclean(m, base & PAGE_MASK, n); 3390 base += n; 3391 size -= n; 3392 n = PAGE_SIZE; 3393 } 3394 vm_page_unlock_queues(); 3395 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3396} 3397 3398/* 3399 * vfs_bio_clrbuf: 3400 * 3401 * clear a buffer. This routine essentially fakes an I/O, so we need 3402 * to clear BIO_ERROR and B_INVAL. 3403 * 3404 * Note that while we only theoretically need to clear through b_bcount, 3405 * we go ahead and clear through b_bufsize. 3406 */ 3407 3408void 3409vfs_bio_clrbuf(struct buf *bp) 3410{ 3411 int i, j, mask = 0; 3412 caddr_t sa, ea; 3413 3414 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3415 clrbuf(bp); 3416 return; 3417 } 3418 3419 bp->b_flags &= ~B_INVAL; 3420 bp->b_ioflags &= ~BIO_ERROR; 3421 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3422 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3423 (bp->b_offset & PAGE_MASK) == 0) { 3424 if (bp->b_pages[0] == bogus_page) 3425 goto unlock; 3426 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3427 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3428 if ((bp->b_pages[0]->valid & mask) == mask) 3429 goto unlock; 3430 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3431 ((bp->b_pages[0]->valid & mask) == 0)) { 3432 bzero(bp->b_data, bp->b_bufsize); 3433 bp->b_pages[0]->valid |= mask; 3434 goto unlock; 3435 } 3436 } 3437 ea = sa = bp->b_data; 3438 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3439 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3440 ea = (caddr_t)(vm_offset_t)ulmin( 3441 (u_long)(vm_offset_t)ea, 3442 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3443 if (bp->b_pages[i] == bogus_page) 3444 continue; 3445 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3446 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3447 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3448 if ((bp->b_pages[i]->valid & mask) == mask) 3449 continue; 3450 if ((bp->b_pages[i]->valid & mask) == 0) { 3451 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3452 bzero(sa, ea - sa); 3453 } else { 3454 for (; sa < ea; sa += DEV_BSIZE, j++) { 3455 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3456 (bp->b_pages[i]->valid & (1 << j)) == 0) 3457 bzero(sa, DEV_BSIZE); 3458 } 3459 } 3460 bp->b_pages[i]->valid |= mask; 3461 } 3462unlock: 3463 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3464 bp->b_resid = 0; 3465} 3466 3467/* 3468 * vm_hold_load_pages and vm_hold_free_pages get pages into 3469 * a buffers address space. The pages are anonymous and are 3470 * not associated with a file object. 3471 */ 3472static void 3473vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3474{ 3475 vm_offset_t pg; 3476 vm_page_t p; 3477 int index; 3478 3479 to = round_page(to); 3480 from = round_page(from); 3481 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3482 3483 VM_OBJECT_LOCK(kernel_object); 3484 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3485tryagain: 3486 /* 3487 * note: must allocate system pages since blocking here 3488 * could intefere with paging I/O, no matter which 3489 * process we are. 3490 */ 3491 p = vm_page_alloc(kernel_object, 3492 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3493 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3494 if (!p) { 3495 atomic_add_int(&vm_pageout_deficit, 3496 (to - pg) >> PAGE_SHIFT); 3497 VM_OBJECT_UNLOCK(kernel_object); 3498 VM_WAIT; 3499 VM_OBJECT_LOCK(kernel_object); 3500 goto tryagain; 3501 } 3502 p->valid = VM_PAGE_BITS_ALL; 3503 pmap_qenter(pg, &p, 1); 3504 bp->b_pages[index] = p; 3505 } 3506 VM_OBJECT_UNLOCK(kernel_object); 3507 bp->b_npages = index; 3508} 3509 3510/* Return pages associated with this buf to the vm system */ 3511static void 3512vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3513{ 3514 vm_offset_t pg; 3515 vm_page_t p; 3516 int index, newnpages; 3517 3518 from = round_page(from); 3519 to = round_page(to); 3520 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3521 3522 VM_OBJECT_LOCK(kernel_object); 3523 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3524 p = bp->b_pages[index]; 3525 if (p && (index < bp->b_npages)) { 3526 if (p->busy) { 3527 printf( 3528 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3529 (intmax_t)bp->b_blkno, 3530 (intmax_t)bp->b_lblkno); 3531 } 3532 bp->b_pages[index] = NULL; 3533 pmap_qremove(pg, 1); 3534 vm_page_lock_queues(); 3535 vm_page_unwire(p, 0); 3536 vm_page_free(p); 3537 vm_page_unlock_queues(); 3538 } 3539 } 3540 VM_OBJECT_UNLOCK(kernel_object); 3541 bp->b_npages = newnpages; 3542} 3543 3544/* 3545 * Map an IO request into kernel virtual address space. 3546 * 3547 * All requests are (re)mapped into kernel VA space. 3548 * Notice that we use b_bufsize for the size of the buffer 3549 * to be mapped. b_bcount might be modified by the driver. 3550 * 3551 * Note that even if the caller determines that the address space should 3552 * be valid, a race or a smaller-file mapped into a larger space may 3553 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3554 * check the return value. 3555 */ 3556int 3557vmapbuf(struct buf *bp) 3558{ 3559 caddr_t addr, kva; 3560 vm_prot_t prot; 3561 int pidx, i; 3562 struct vm_page *m; 3563 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3564 3565 if (bp->b_bufsize < 0) 3566 return (-1); 3567 prot = VM_PROT_READ; 3568 if (bp->b_iocmd == BIO_READ) 3569 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3570 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3571 addr < bp->b_data + bp->b_bufsize; 3572 addr += PAGE_SIZE, pidx++) { 3573 /* 3574 * Do the vm_fault if needed; do the copy-on-write thing 3575 * when reading stuff off device into memory. 3576 * 3577 * NOTE! Must use pmap_extract() because addr may be in 3578 * the userland address space, and kextract is only guarenteed 3579 * to work for the kernland address space (see: sparc64 port). 3580 */ 3581retry: 3582 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3583 prot) < 0) { 3584 vm_page_lock_queues(); 3585 for (i = 0; i < pidx; ++i) { 3586 vm_page_unhold(bp->b_pages[i]); 3587 bp->b_pages[i] = NULL; 3588 } 3589 vm_page_unlock_queues(); 3590 return(-1); 3591 } 3592 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3593 if (m == NULL) 3594 goto retry; 3595 bp->b_pages[pidx] = m; 3596 } 3597 if (pidx > btoc(MAXPHYS)) 3598 panic("vmapbuf: mapped more than MAXPHYS"); 3599 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3600 3601 kva = bp->b_saveaddr; 3602 bp->b_npages = pidx; 3603 bp->b_saveaddr = bp->b_data; 3604 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3605 return(0); 3606} 3607 3608/* 3609 * Free the io map PTEs associated with this IO operation. 3610 * We also invalidate the TLB entries and restore the original b_addr. 3611 */ 3612void 3613vunmapbuf(struct buf *bp) 3614{ 3615 int pidx; 3616 int npages; 3617 3618 npages = bp->b_npages; 3619 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3620 vm_page_lock_queues(); 3621 for (pidx = 0; pidx < npages; pidx++) 3622 vm_page_unhold(bp->b_pages[pidx]); 3623 vm_page_unlock_queues(); 3624 3625 bp->b_data = bp->b_saveaddr; 3626} 3627 3628void 3629bdone(struct buf *bp) 3630{ 3631 3632 mtx_lock(&bdonelock); 3633 bp->b_flags |= B_DONE; 3634 wakeup(bp); 3635 mtx_unlock(&bdonelock); 3636} 3637 3638void 3639bwait(struct buf *bp, u_char pri, const char *wchan) 3640{ 3641 3642 mtx_lock(&bdonelock); 3643 while ((bp->b_flags & B_DONE) == 0) 3644 msleep(bp, &bdonelock, pri, wchan, 0); 3645 mtx_unlock(&bdonelock); 3646} 3647 3648int 3649bufsync(struct bufobj *bo, int waitfor, struct thread *td) 3650{ 3651 3652 return (VOP_FSYNC(bo->__bo_vnode, waitfor, td)); 3653} 3654 3655void 3656bufstrategy(struct bufobj *bo, struct buf *bp) 3657{ 3658 int i = 0; 3659 struct vnode *vp; 3660 3661 vp = bp->b_vp; 3662 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3663 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3664 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3665 i = VOP_STRATEGY(vp, bp); 3666 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3667} 3668 3669void 3670bufobj_wrefl(struct bufobj *bo) 3671{ 3672 3673 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3674 ASSERT_BO_LOCKED(bo); 3675 bo->bo_numoutput++; 3676} 3677 3678void 3679bufobj_wref(struct bufobj *bo) 3680{ 3681 3682 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3683 BO_LOCK(bo); 3684 bo->bo_numoutput++; 3685 BO_UNLOCK(bo); 3686} 3687 3688void 3689bufobj_wdrop(struct bufobj *bo) 3690{ 3691 3692 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3693 BO_LOCK(bo); 3694 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3695 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3696 bo->bo_flag &= ~BO_WWAIT; 3697 wakeup(&bo->bo_numoutput); 3698 } 3699 BO_UNLOCK(bo); 3700} 3701 3702int 3703bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3704{ 3705 int error; 3706 3707 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3708 ASSERT_BO_LOCKED(bo); 3709 error = 0; 3710 while (bo->bo_numoutput) { 3711 bo->bo_flag |= BO_WWAIT; 3712 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3713 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3714 if (error) 3715 break; 3716 } 3717 return (error); 3718} 3719 3720#include "opt_ddb.h" 3721#ifdef DDB 3722#include <ddb/ddb.h> 3723 3724/* DDB command to show buffer data */ 3725DB_SHOW_COMMAND(buffer, db_show_buffer) 3726{ 3727 /* get args */ 3728 struct buf *bp = (struct buf *)addr; 3729 3730 if (!have_addr) { 3731 db_printf("usage: show buffer <addr>\n"); 3732 return; 3733 } 3734 3735 db_printf("buf at %p\n", bp); 3736 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3737 db_printf( 3738 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3739 "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n", 3740 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3741 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno); 3742 db_printf("lockstatus = %d, excl count = %d, excl owner %p\n", 3743 lockstatus(&bp->b_lock, NULL), bp->b_lock.lk_exclusivecount, 3744 bp->b_lock.lk_lockholder); 3745 if (bp->b_npages) { 3746 int i; 3747 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3748 for (i = 0; i < bp->b_npages; i++) { 3749 vm_page_t m; 3750 m = bp->b_pages[i]; 3751 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3752 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3753 if ((i + 1) < bp->b_npages) 3754 db_printf(","); 3755 } 3756 db_printf("\n"); 3757 } 3758} 3759 3760DB_SHOW_COMMAND(lockedbufs, lockedbufs) 3761{ 3762 struct buf *bp; 3763 int i; 3764 3765 for (i = 0; i < nbuf; i++) { 3766 bp = &buf[i]; 3767 if (lockcount(&bp->b_lock)) { 3768 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 3769 db_printf("\n"); 3770 } 3771 } 3772} 3773#endif /* DDB */ 3774