vfs_bio.c revision 111466
1/* 2 * Copyright (c) 1994,1997 John S. Dyson 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice immediately at the beginning of the file, without modification, 10 * this list of conditions, and the following disclaimer. 11 * 2. Absolutely no warranty of function or purpose is made by the author 12 * John S. Dyson. 13 * 14 * $FreeBSD: head/sys/kern/vfs_bio.c 111466 2003-02-25 06:44:42Z mckusick $ 15 */ 16 17/* 18 * this file contains a new buffer I/O scheme implementing a coherent 19 * VM object and buffer cache scheme. Pains have been taken to make 20 * sure that the performance degradation associated with schemes such 21 * as this is not realized. 22 * 23 * Author: John S. Dyson 24 * Significant help during the development and debugging phases 25 * had been provided by David Greenman, also of the FreeBSD core team. 26 * 27 * see man buf(9) for more info. 28 */ 29 30#include <sys/param.h> 31#include <sys/systm.h> 32#include <sys/stdint.h> 33#include <sys/bio.h> 34#include <sys/buf.h> 35#include <sys/devicestat.h> 36#include <sys/eventhandler.h> 37#include <sys/lock.h> 38#include <sys/malloc.h> 39#include <sys/mount.h> 40#include <sys/mutex.h> 41#include <sys/kernel.h> 42#include <sys/kthread.h> 43#include <sys/proc.h> 44#include <sys/resourcevar.h> 45#include <sys/sysctl.h> 46#include <sys/vmmeter.h> 47#include <sys/vnode.h> 48#include <vm/vm.h> 49#include <vm/vm_param.h> 50#include <vm/vm_kern.h> 51#include <vm/vm_pageout.h> 52#include <vm/vm_page.h> 53#include <vm/vm_object.h> 54#include <vm/vm_extern.h> 55#include <vm/vm_map.h> 56 57static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 58 59struct bio_ops bioops; /* I/O operation notification */ 60 61struct buf_ops buf_ops_bio = { 62 "buf_ops_bio", 63 bwrite 64}; 65 66/* 67 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 68 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 69 */ 70struct buf *buf; /* buffer header pool */ 71 72static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 73 vm_offset_t to); 74static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 75 vm_offset_t to); 76static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 77 int pageno, vm_page_t m); 78static void vfs_clean_pages(struct buf * bp); 79static void vfs_setdirty(struct buf *bp); 80static void vfs_vmio_release(struct buf *bp); 81static void vfs_backgroundwritedone(struct buf *bp); 82static int vfs_bio_clcheck(struct vnode *vp, int size, 83 daddr_t lblkno, daddr_t blkno); 84static int flushbufqueues(void); 85static void buf_daemon(void); 86void bremfreel(struct buf * bp); 87 88int vmiodirenable = TRUE; 89SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 90 "Use the VM system for directory writes"); 91int runningbufspace; 92SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 93 "Amount of presently outstanding async buffer io"); 94static int bufspace; 95SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 96 "KVA memory used for bufs"); 97static int maxbufspace; 98SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 99 "Maximum allowed value of bufspace (including buf_daemon)"); 100static int bufmallocspace; 101SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 102 "Amount of malloced memory for buffers"); 103static int maxbufmallocspace; 104SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 105 "Maximum amount of malloced memory for buffers"); 106static int lobufspace; 107SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 108 "Minimum amount of buffers we want to have"); 109static int hibufspace; 110SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 111 "Maximum allowed value of bufspace (excluding buf_daemon)"); 112static int bufreusecnt; 113SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 114 "Number of times we have reused a buffer"); 115static int buffreekvacnt; 116SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 117 "Number of times we have freed the KVA space from some buffer"); 118static int bufdefragcnt; 119SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 120 "Number of times we have had to repeat buffer allocation to defragment"); 121static int lorunningspace; 122SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 123 "Minimum preferred space used for in-progress I/O"); 124static int hirunningspace; 125SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 126 "Maximum amount of space to use for in-progress I/O"); 127static int dirtybufferflushes; 128SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 129 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 130static int altbufferflushes; 131SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 132 0, "Number of fsync flushes to limit dirty buffers"); 133static int numdirtybuffers; 134SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 135 "Number of buffers that are dirty (has unwritten changes) at the moment"); 136static int lodirtybuffers; 137SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 138 "How many buffers we want to have free before bufdaemon can sleep"); 139static int hidirtybuffers; 140SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 141 "When the number of dirty buffers is considered severe"); 142static int dirtybufthresh; 143SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 144 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 145static int numfreebuffers; 146SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 147 "Number of free buffers"); 148static int lofreebuffers; 149SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 150 "XXX Unused"); 151static int hifreebuffers; 152SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 153 "XXX Complicatedly unused"); 154static int getnewbufcalls; 155SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 156 "Number of calls to getnewbuf"); 157static int getnewbufrestarts; 158SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 159 "Number of times getnewbuf has had to restart a buffer aquisition"); 160static int dobkgrdwrite = 1; 161SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 162 "Do background writes (honoring the BX_BKGRDWRITE flag)?"); 163 164/* 165 * Wakeup point for bufdaemon, as well as indicator of whether it is already 166 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 167 * is idling. 168 */ 169static int bd_request; 170 171/* 172 * This lock synchronizes access to bd_request. 173 */ 174static struct mtx bdlock; 175 176/* 177 * bogus page -- for I/O to/from partially complete buffers 178 * this is a temporary solution to the problem, but it is not 179 * really that bad. it would be better to split the buffer 180 * for input in the case of buffers partially already in memory, 181 * but the code is intricate enough already. 182 */ 183vm_page_t bogus_page; 184 185/* 186 * Synchronization (sleep/wakeup) variable for active buffer space requests. 187 * Set when wait starts, cleared prior to wakeup(). 188 * Used in runningbufwakeup() and waitrunningbufspace(). 189 */ 190static int runningbufreq; 191 192/* 193 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 194 * waitrunningbufspace(). 195 */ 196static struct mtx rbreqlock; 197 198/* 199 * Synchronization (sleep/wakeup) variable for buffer requests. 200 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 201 * by and/or. 202 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 203 * getnewbuf(), and getblk(). 204 */ 205static int needsbuffer; 206 207/* 208 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 209 */ 210static struct mtx nblock; 211 212#ifdef USE_BUFHASH 213/* 214 * Mask for index into the buffer hash table, which needs to be power of 2 in 215 * size. Set in kern_vfs_bio_buffer_alloc. 216 */ 217static int bufhashmask; 218 219/* 220 * Hash table for all buffers, with a linked list hanging from each table 221 * entry. Set in kern_vfs_bio_buffer_alloc, initialized in buf_init. 222 */ 223static LIST_HEAD(bufhashhdr, buf) *bufhashtbl; 224 225/* 226 * Somewhere to store buffers when they are not in another list, to always 227 * have them in a list (and thus being able to use the same set of operations 228 * on them.) 229 */ 230static struct bufhashhdr invalhash; 231 232#endif 233 234/* 235 * Definitions for the buffer free lists. 236 */ 237#define BUFFER_QUEUES 6 /* number of free buffer queues */ 238 239#define QUEUE_NONE 0 /* on no queue */ 240#define QUEUE_LOCKED 1 /* locked buffers */ 241#define QUEUE_CLEAN 2 /* non-B_DELWRI buffers */ 242#define QUEUE_DIRTY 3 /* B_DELWRI buffers */ 243#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 244#define QUEUE_EMPTY 5 /* empty buffer headers */ 245 246/* Queues for free buffers with various properties */ 247static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 248 249/* Lock for the bufqueues */ 250static struct mtx bqlock; 251 252/* 253 * Single global constant for BUF_WMESG, to avoid getting multiple references. 254 * buf_wmesg is referred from macros. 255 */ 256const char *buf_wmesg = BUF_WMESG; 257 258#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 259#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 260#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 261#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 262 263#ifdef USE_BUFHASH 264/* 265 * Buffer hash table code. Note that the logical block scans linearly, which 266 * gives us some L1 cache locality. 267 */ 268 269static __inline 270struct bufhashhdr * 271bufhash(struct vnode *vnp, daddr_t bn) 272{ 273 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]); 274} 275 276#endif 277 278/* 279 * numdirtywakeup: 280 * 281 * If someone is blocked due to there being too many dirty buffers, 282 * and numdirtybuffers is now reasonable, wake them up. 283 */ 284 285static __inline void 286numdirtywakeup(int level) 287{ 288 if (numdirtybuffers <= level) { 289 mtx_lock(&nblock); 290 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 291 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 292 wakeup(&needsbuffer); 293 } 294 mtx_unlock(&nblock); 295 } 296} 297 298/* 299 * bufspacewakeup: 300 * 301 * Called when buffer space is potentially available for recovery. 302 * getnewbuf() will block on this flag when it is unable to free 303 * sufficient buffer space. Buffer space becomes recoverable when 304 * bp's get placed back in the queues. 305 */ 306 307static __inline void 308bufspacewakeup(void) 309{ 310 /* 311 * If someone is waiting for BUF space, wake them up. Even 312 * though we haven't freed the kva space yet, the waiting 313 * process will be able to now. 314 */ 315 mtx_lock(&nblock); 316 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 317 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 318 wakeup(&needsbuffer); 319 } 320 mtx_unlock(&nblock); 321} 322 323/* 324 * runningbufwakeup() - in-progress I/O accounting. 325 * 326 */ 327static __inline void 328runningbufwakeup(struct buf *bp) 329{ 330 if (bp->b_runningbufspace) { 331 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 332 bp->b_runningbufspace = 0; 333 mtx_lock(&rbreqlock); 334 if (runningbufreq && runningbufspace <= lorunningspace) { 335 runningbufreq = 0; 336 wakeup(&runningbufreq); 337 } 338 mtx_unlock(&rbreqlock); 339 } 340} 341 342/* 343 * bufcountwakeup: 344 * 345 * Called when a buffer has been added to one of the free queues to 346 * account for the buffer and to wakeup anyone waiting for free buffers. 347 * This typically occurs when large amounts of metadata are being handled 348 * by the buffer cache ( else buffer space runs out first, usually ). 349 */ 350 351static __inline void 352bufcountwakeup(void) 353{ 354 atomic_add_int(&numfreebuffers, 1); 355 mtx_lock(&nblock); 356 if (needsbuffer) { 357 needsbuffer &= ~VFS_BIO_NEED_ANY; 358 if (numfreebuffers >= hifreebuffers) 359 needsbuffer &= ~VFS_BIO_NEED_FREE; 360 wakeup(&needsbuffer); 361 } 362 mtx_unlock(&nblock); 363} 364 365/* 366 * waitrunningbufspace() 367 * 368 * runningbufspace is a measure of the amount of I/O currently 369 * running. This routine is used in async-write situations to 370 * prevent creating huge backups of pending writes to a device. 371 * Only asynchronous writes are governed by this function. 372 * 373 * Reads will adjust runningbufspace, but will not block based on it. 374 * The read load has a side effect of reducing the allowed write load. 375 * 376 * This does NOT turn an async write into a sync write. It waits 377 * for earlier writes to complete and generally returns before the 378 * caller's write has reached the device. 379 */ 380static __inline void 381waitrunningbufspace(void) 382{ 383 mtx_lock(&rbreqlock); 384 while (runningbufspace > hirunningspace) { 385 ++runningbufreq; 386 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 387 } 388 mtx_unlock(&rbreqlock); 389} 390 391 392/* 393 * vfs_buf_test_cache: 394 * 395 * Called when a buffer is extended. This function clears the B_CACHE 396 * bit if the newly extended portion of the buffer does not contain 397 * valid data. 398 */ 399static __inline__ 400void 401vfs_buf_test_cache(struct buf *bp, 402 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 403 vm_page_t m) 404{ 405 GIANT_REQUIRED; 406 407 if (bp->b_flags & B_CACHE) { 408 int base = (foff + off) & PAGE_MASK; 409 if (vm_page_is_valid(m, base, size) == 0) 410 bp->b_flags &= ~B_CACHE; 411 } 412} 413 414/* Wake up the buffer deamon if necessary */ 415static __inline__ 416void 417bd_wakeup(int dirtybuflevel) 418{ 419 mtx_lock(&bdlock); 420 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 421 bd_request = 1; 422 wakeup(&bd_request); 423 } 424 mtx_unlock(&bdlock); 425} 426 427/* 428 * bd_speedup - speedup the buffer cache flushing code 429 */ 430 431static __inline__ 432void 433bd_speedup(void) 434{ 435 bd_wakeup(1); 436} 437 438/* 439 * Calculating buffer cache scaling values and reserve space for buffer 440 * headers. This is called during low level kernel initialization and 441 * may be called more then once. We CANNOT write to the memory area 442 * being reserved at this time. 443 */ 444caddr_t 445kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 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); 495 496 /* 497 * Reserve space for the buffer cache buffers 498 */ 499 swbuf = (void *)v; 500 v = (caddr_t)(swbuf + nswbuf); 501 buf = (void *)v; 502 v = (caddr_t)(buf + nbuf); 503 504#ifdef USE_BUFHASH 505 /* 506 * Calculate the hash table size and reserve space 507 */ 508 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 509 ; 510 bufhashtbl = (void *)v; 511 v = (caddr_t)(bufhashtbl + bufhashmask); 512 --bufhashmask; 513#endif 514 return(v); 515} 516 517/* Initialize the buffer subsystem. Called before use of any buffers. */ 518void 519bufinit(void) 520{ 521 struct buf *bp; 522 vm_offset_t bogus_offset; 523 int i; 524 525 GIANT_REQUIRED; 526 527#ifdef USE_BUFHASH 528 LIST_INIT(&invalhash); 529#endif 530 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 531 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 532 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 533 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 534 535#ifdef USE_BUFHASH 536 for (i = 0; i <= bufhashmask; i++) 537 LIST_INIT(&bufhashtbl[i]); 538#endif 539 540 /* next, make a null set of free lists */ 541 for (i = 0; i < BUFFER_QUEUES; i++) 542 TAILQ_INIT(&bufqueues[i]); 543 544 /* finally, initialize each buffer header and stick on empty q */ 545 for (i = 0; i < nbuf; i++) { 546 bp = &buf[i]; 547 bzero(bp, sizeof *bp); 548 bp->b_flags = B_INVAL; /* we're just an empty header */ 549 bp->b_dev = NODEV; 550 bp->b_rcred = NOCRED; 551 bp->b_wcred = NOCRED; 552 bp->b_qindex = QUEUE_EMPTY; 553 bp->b_vflags = 0; 554 bp->b_xflags = 0; 555 LIST_INIT(&bp->b_dep); 556 BUF_LOCKINIT(bp); 557 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 558#ifdef USE_BUFHASH 559 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 560#endif 561 } 562 563 /* 564 * maxbufspace is the absolute maximum amount of buffer space we are 565 * allowed to reserve in KVM and in real terms. The absolute maximum 566 * is nominally used by buf_daemon. hibufspace is the nominal maximum 567 * used by most other processes. The differential is required to 568 * ensure that buf_daemon is able to run when other processes might 569 * be blocked waiting for buffer space. 570 * 571 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 572 * this may result in KVM fragmentation which is not handled optimally 573 * by the system. 574 */ 575 maxbufspace = nbuf * BKVASIZE; 576 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 577 lobufspace = hibufspace - MAXBSIZE; 578 579 lorunningspace = 512 * 1024; 580 hirunningspace = 1024 * 1024; 581 582/* 583 * Limit the amount of malloc memory since it is wired permanently into 584 * the kernel space. Even though this is accounted for in the buffer 585 * allocation, we don't want the malloced region to grow uncontrolled. 586 * The malloc scheme improves memory utilization significantly on average 587 * (small) directories. 588 */ 589 maxbufmallocspace = hibufspace / 20; 590 591/* 592 * Reduce the chance of a deadlock occuring by limiting the number 593 * of delayed-write dirty buffers we allow to stack up. 594 */ 595 hidirtybuffers = nbuf / 4 + 20; 596 dirtybufthresh = hidirtybuffers * 9 / 10; 597 numdirtybuffers = 0; 598/* 599 * To support extreme low-memory systems, make sure hidirtybuffers cannot 600 * eat up all available buffer space. This occurs when our minimum cannot 601 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 602 * BKVASIZE'd (8K) buffers. 603 */ 604 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 605 hidirtybuffers >>= 1; 606 } 607 lodirtybuffers = hidirtybuffers / 2; 608 609/* 610 * Try to keep the number of free buffers in the specified range, 611 * and give special processes (e.g. like buf_daemon) access to an 612 * emergency reserve. 613 */ 614 lofreebuffers = nbuf / 18 + 5; 615 hifreebuffers = 2 * lofreebuffers; 616 numfreebuffers = nbuf; 617 618/* 619 * Maximum number of async ops initiated per buf_daemon loop. This is 620 * somewhat of a hack at the moment, we really need to limit ourselves 621 * based on the number of bytes of I/O in-transit that were initiated 622 * from buf_daemon. 623 */ 624 625 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 626 vm_object_lock(kernel_object); 627 bogus_page = vm_page_alloc(kernel_object, 628 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 629 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 630 vm_object_unlock(kernel_object); 631} 632 633/* 634 * bfreekva() - free the kva allocation for a buffer. 635 * 636 * Must be called at splbio() or higher as this is the only locking for 637 * buffer_map. 638 * 639 * Since this call frees up buffer space, we call bufspacewakeup(). 640 */ 641static void 642bfreekva(struct buf * bp) 643{ 644 GIANT_REQUIRED; 645 646 if (bp->b_kvasize) { 647 atomic_add_int(&buffreekvacnt, 1); 648 atomic_subtract_int(&bufspace, bp->b_kvasize); 649 vm_map_delete(buffer_map, 650 (vm_offset_t) bp->b_kvabase, 651 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 652 ); 653 bp->b_kvasize = 0; 654 bufspacewakeup(); 655 } 656} 657 658/* 659 * bremfree: 660 * 661 * Remove the buffer from the appropriate free list. 662 */ 663void 664bremfree(struct buf * bp) 665{ 666 mtx_lock(&bqlock); 667 bremfreel(bp); 668 mtx_unlock(&bqlock); 669} 670 671void 672bremfreel(struct buf * bp) 673{ 674 int s = splbio(); 675 int old_qindex = bp->b_qindex; 676 677 GIANT_REQUIRED; 678 679 if (bp->b_qindex != QUEUE_NONE) { 680 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 681 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 682 bp->b_qindex = QUEUE_NONE; 683 } else { 684 if (BUF_REFCNT(bp) <= 1) 685 panic("bremfree: removing a buffer not on a queue"); 686 } 687 688 /* 689 * Fixup numfreebuffers count. If the buffer is invalid or not 690 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 691 * the buffer was free and we must decrement numfreebuffers. 692 */ 693 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 694 switch(old_qindex) { 695 case QUEUE_DIRTY: 696 case QUEUE_CLEAN: 697 case QUEUE_EMPTY: 698 case QUEUE_EMPTYKVA: 699 atomic_subtract_int(&numfreebuffers, 1); 700 break; 701 default: 702 break; 703 } 704 } 705 splx(s); 706} 707 708 709/* 710 * Get a buffer with the specified data. Look in the cache first. We 711 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 712 * is set, the buffer is valid and we do not have to do anything ( see 713 * getblk() ). This is really just a special case of breadn(). 714 */ 715int 716bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 717 struct buf ** bpp) 718{ 719 720 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 721} 722 723/* 724 * Operates like bread, but also starts asynchronous I/O on 725 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 726 * to initiating I/O . If B_CACHE is set, the buffer is valid 727 * and we do not have to do anything. 728 */ 729int 730breadn(struct vnode * vp, daddr_t blkno, int size, 731 daddr_t * rablkno, int *rabsize, 732 int cnt, struct ucred * cred, struct buf ** bpp) 733{ 734 struct buf *bp, *rabp; 735 int i; 736 int rv = 0, readwait = 0; 737 738 *bpp = bp = getblk(vp, blkno, size, 0, 0); 739 740 /* if not found in cache, do some I/O */ 741 if ((bp->b_flags & B_CACHE) == 0) { 742 if (curthread != PCPU_GET(idlethread)) 743 curthread->td_proc->p_stats->p_ru.ru_inblock++; 744 bp->b_iocmd = BIO_READ; 745 bp->b_flags &= ~B_INVAL; 746 bp->b_ioflags &= ~BIO_ERROR; 747 if (bp->b_rcred == NOCRED && cred != NOCRED) 748 bp->b_rcred = crhold(cred); 749 vfs_busy_pages(bp, 0); 750 if (vp->v_type == VCHR) 751 VOP_SPECSTRATEGY(vp, bp); 752 else 753 VOP_STRATEGY(vp, bp); 754 ++readwait; 755 } 756 757 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 758 if (inmem(vp, *rablkno)) 759 continue; 760 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 761 762 if ((rabp->b_flags & B_CACHE) == 0) { 763 if (curthread != PCPU_GET(idlethread)) 764 curthread->td_proc->p_stats->p_ru.ru_inblock++; 765 rabp->b_flags |= B_ASYNC; 766 rabp->b_flags &= ~B_INVAL; 767 rabp->b_ioflags &= ~BIO_ERROR; 768 rabp->b_iocmd = BIO_READ; 769 if (rabp->b_rcred == NOCRED && cred != NOCRED) 770 rabp->b_rcred = crhold(cred); 771 vfs_busy_pages(rabp, 0); 772 BUF_KERNPROC(rabp); 773 if (vp->v_type == VCHR) 774 VOP_SPECSTRATEGY(vp, rabp); 775 else 776 VOP_STRATEGY(vp, rabp); 777 } else { 778 brelse(rabp); 779 } 780 } 781 782 if (readwait) { 783 rv = bufwait(bp); 784 } 785 return (rv); 786} 787 788/* 789 * Write, release buffer on completion. (Done by iodone 790 * if async). Do not bother writing anything if the buffer 791 * is invalid. 792 * 793 * Note that we set B_CACHE here, indicating that buffer is 794 * fully valid and thus cacheable. This is true even of NFS 795 * now so we set it generally. This could be set either here 796 * or in biodone() since the I/O is synchronous. We put it 797 * here. 798 */ 799 800int 801bwrite(struct buf * bp) 802{ 803 int oldflags, s; 804 struct buf *newbp; 805 806 if (bp->b_flags & B_INVAL) { 807 brelse(bp); 808 return (0); 809 } 810 811 oldflags = bp->b_flags; 812 813 if (BUF_REFCNT(bp) == 0) 814 panic("bwrite: buffer is not busy???"); 815 s = splbio(); 816 /* 817 * If a background write is already in progress, delay 818 * writing this block if it is asynchronous. Otherwise 819 * wait for the background write to complete. 820 */ 821 if (bp->b_xflags & BX_BKGRDINPROG) { 822 if (bp->b_flags & B_ASYNC) { 823 splx(s); 824 bdwrite(bp); 825 return (0); 826 } 827 bp->b_xflags |= BX_BKGRDWAIT; 828 tsleep(&bp->b_xflags, PRIBIO, "bwrbg", 0); 829 if (bp->b_xflags & BX_BKGRDINPROG) 830 panic("bwrite: still writing"); 831 } 832 833 /* Mark the buffer clean */ 834 bundirty(bp); 835 836 /* 837 * If this buffer is marked for background writing and we 838 * do not have to wait for it, make a copy and write the 839 * copy so as to leave this buffer ready for further use. 840 * 841 * This optimization eats a lot of memory. If we have a page 842 * or buffer shortfall we can't do it. 843 */ 844 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 845 (bp->b_flags & B_ASYNC) && 846 !vm_page_count_severe() && 847 !buf_dirty_count_severe()) { 848 if (bp->b_iodone != NULL) { 849 printf("bp->b_iodone = %p\n", bp->b_iodone); 850 panic("bwrite: need chained iodone"); 851 } 852 853 /* get a new block */ 854 newbp = geteblk(bp->b_bufsize); 855 856 /* 857 * set it to be identical to the old block. We have to 858 * set b_lblkno and BKGRDMARKER before calling bgetvp() 859 * to avoid confusing the splay tree and gbincore(). 860 */ 861 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 862 newbp->b_lblkno = bp->b_lblkno; 863 newbp->b_xflags |= BX_BKGRDMARKER; 864 bgetvp(bp->b_vp, newbp); 865 newbp->b_blkno = bp->b_blkno; 866 newbp->b_offset = bp->b_offset; 867 newbp->b_iodone = vfs_backgroundwritedone; 868 newbp->b_flags |= B_ASYNC; 869 newbp->b_flags &= ~B_INVAL; 870 871 /* move over the dependencies */ 872 if (LIST_FIRST(&bp->b_dep) != NULL) 873 buf_movedeps(bp, newbp); 874 875 /* 876 * Initiate write on the copy, release the original to 877 * the B_LOCKED queue so that it cannot go away until 878 * the background write completes. If not locked it could go 879 * away and then be reconstituted while it was being written. 880 * If the reconstituted buffer were written, we could end up 881 * with two background copies being written at the same time. 882 */ 883 bp->b_xflags |= BX_BKGRDINPROG; 884 bp->b_flags |= B_LOCKED; 885 bqrelse(bp); 886 bp = newbp; 887 } 888 889 bp->b_flags &= ~B_DONE; 890 bp->b_ioflags &= ~BIO_ERROR; 891 bp->b_flags |= B_WRITEINPROG | B_CACHE; 892 bp->b_iocmd = BIO_WRITE; 893 894 VI_LOCK(bp->b_vp); 895 bp->b_vp->v_numoutput++; 896 VI_UNLOCK(bp->b_vp); 897 vfs_busy_pages(bp, 1); 898 899 /* 900 * Normal bwrites pipeline writes 901 */ 902 bp->b_runningbufspace = bp->b_bufsize; 903 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 904 905 if (curthread != PCPU_GET(idlethread)) 906 curthread->td_proc->p_stats->p_ru.ru_oublock++; 907 splx(s); 908 if (oldflags & B_ASYNC) 909 BUF_KERNPROC(bp); 910 if (bp->b_vp->v_type == VCHR) 911 VOP_SPECSTRATEGY(bp->b_vp, bp); 912 else 913 VOP_STRATEGY(bp->b_vp, bp); 914 915 if ((oldflags & B_ASYNC) == 0) { 916 int rtval = bufwait(bp); 917 brelse(bp); 918 return (rtval); 919 } else if ((oldflags & B_NOWDRAIN) == 0) { 920 /* 921 * don't allow the async write to saturate the I/O 922 * system. Deadlocks can occur only if a device strategy 923 * routine (like in MD) turns around and issues another 924 * high-level write, in which case B_NOWDRAIN is expected 925 * to be set. Otherwise we will not deadlock here because 926 * we are blocking waiting for I/O that is already in-progress 927 * to complete. 928 */ 929 waitrunningbufspace(); 930 } 931 932 return (0); 933} 934 935/* 936 * Complete a background write started from bwrite. 937 */ 938static void 939vfs_backgroundwritedone(bp) 940 struct buf *bp; 941{ 942 struct buf *origbp; 943 944 /* 945 * Find the original buffer that we are writing. 946 */ 947 VI_LOCK(bp->b_vp); 948 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 949 panic("backgroundwritedone: lost buffer"); 950 VI_UNLOCK(bp->b_vp); 951 /* 952 * Process dependencies then return any unfinished ones. 953 */ 954 if (LIST_FIRST(&bp->b_dep) != NULL) 955 buf_complete(bp); 956 if (LIST_FIRST(&bp->b_dep) != NULL) 957 buf_movedeps(bp, origbp); 958 959 /* XXX Find out if origbp can disappear or get inconsistent */ 960 /* 961 * Clear the BX_BKGRDINPROG flag in the original buffer 962 * and awaken it if it is waiting for the write to complete. 963 * If BX_BKGRDINPROG is not set in the original buffer it must 964 * have been released and re-instantiated - which is not legal. 965 */ 966 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), 967 ("backgroundwritedone: lost buffer2")); 968 origbp->b_xflags &= ~BX_BKGRDINPROG; 969 if (origbp->b_xflags & BX_BKGRDWAIT) { 970 origbp->b_xflags &= ~BX_BKGRDWAIT; 971 wakeup(&origbp->b_xflags); 972 } 973 /* 974 * Clear the B_LOCKED flag and remove it from the locked 975 * queue if it currently resides there. 976 */ 977 origbp->b_flags &= ~B_LOCKED; 978 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) { 979 bremfree(origbp); 980 bqrelse(origbp); 981 } 982 /* 983 * This buffer is marked B_NOCACHE, so when it is released 984 * by biodone, it will be tossed. We mark it with BIO_READ 985 * to avoid biodone doing a second vwakeup. 986 */ 987 bp->b_flags |= B_NOCACHE; 988 bp->b_iocmd = BIO_READ; 989 bp->b_flags &= ~(B_CACHE | B_DONE); 990 bp->b_iodone = 0; 991 bufdone(bp); 992} 993 994/* 995 * Delayed write. (Buffer is marked dirty). Do not bother writing 996 * anything if the buffer is marked invalid. 997 * 998 * Note that since the buffer must be completely valid, we can safely 999 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1000 * biodone() in order to prevent getblk from writing the buffer 1001 * out synchronously. 1002 */ 1003void 1004bdwrite(struct buf * bp) 1005{ 1006 struct thread *td = curthread; 1007 struct vnode *vp; 1008 struct buf *nbp; 1009 1010 GIANT_REQUIRED; 1011 1012 if (BUF_REFCNT(bp) == 0) 1013 panic("bdwrite: buffer is not busy"); 1014 1015 if (bp->b_flags & B_INVAL) { 1016 brelse(bp); 1017 return; 1018 } 1019 1020 /* 1021 * If we have too many dirty buffers, don't create any more. 1022 * If we are wildly over our limit, then force a complete 1023 * cleanup. Otherwise, just keep the situation from getting 1024 * out of control. 1025 */ 1026 vp = bp->b_vp; 1027 VI_LOCK(vp); 1028 if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) { 1029 VI_UNLOCK(vp); 1030 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 1031 VI_LOCK(vp); 1032 altbufferflushes++; 1033 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) { 1034 /* 1035 * Try to find a buffer to flush. 1036 */ 1037 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) { 1038 if ((nbp->b_xflags & BX_BKGRDINPROG) || 1039 buf_countdeps(nbp, 0) || 1040 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT)) 1041 continue; 1042 if (bp == nbp) 1043 panic("bdwrite: found ourselves"); 1044 VI_UNLOCK(vp); 1045 if (nbp->b_flags & B_CLUSTEROK) { 1046 BUF_UNLOCK(nbp); 1047 vfs_bio_awrite(nbp); 1048 } else { 1049 bremfree(nbp); 1050 bawrite(nbp); 1051 } 1052 VI_LOCK(vp); 1053 dirtybufferflushes++; 1054 break; 1055 } 1056 } 1057 VI_UNLOCK(vp); 1058 1059 bdirty(bp); 1060 /* 1061 * Set B_CACHE, indicating that the buffer is fully valid. This is 1062 * true even of NFS now. 1063 */ 1064 bp->b_flags |= B_CACHE; 1065 1066 /* 1067 * This bmap keeps the system from needing to do the bmap later, 1068 * perhaps when the system is attempting to do a sync. Since it 1069 * is likely that the indirect block -- or whatever other datastructure 1070 * that the filesystem needs is still in memory now, it is a good 1071 * thing to do this. Note also, that if the pageout daemon is 1072 * requesting a sync -- there might not be enough memory to do 1073 * the bmap then... So, this is important to do. 1074 */ 1075 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1076 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1077 } 1078 1079 /* 1080 * Set the *dirty* buffer range based upon the VM system dirty pages. 1081 */ 1082 vfs_setdirty(bp); 1083 1084 /* 1085 * We need to do this here to satisfy the vnode_pager and the 1086 * pageout daemon, so that it thinks that the pages have been 1087 * "cleaned". Note that since the pages are in a delayed write 1088 * buffer -- the VFS layer "will" see that the pages get written 1089 * out on the next sync, or perhaps the cluster will be completed. 1090 */ 1091 vfs_clean_pages(bp); 1092 bqrelse(bp); 1093 1094 /* 1095 * Wakeup the buffer flushing daemon if we have a lot of dirty 1096 * buffers (midpoint between our recovery point and our stall 1097 * point). 1098 */ 1099 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1100 1101 /* 1102 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1103 * due to the softdep code. 1104 */ 1105} 1106 1107/* 1108 * bdirty: 1109 * 1110 * Turn buffer into delayed write request. We must clear BIO_READ and 1111 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1112 * itself to properly update it in the dirty/clean lists. We mark it 1113 * B_DONE to ensure that any asynchronization of the buffer properly 1114 * clears B_DONE ( else a panic will occur later ). 1115 * 1116 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1117 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1118 * should only be called if the buffer is known-good. 1119 * 1120 * Since the buffer is not on a queue, we do not update the numfreebuffers 1121 * count. 1122 * 1123 * Must be called at splbio(). 1124 * The buffer must be on QUEUE_NONE. 1125 */ 1126void 1127bdirty(bp) 1128 struct buf *bp; 1129{ 1130 KASSERT(bp->b_qindex == QUEUE_NONE, 1131 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1132 bp->b_flags &= ~(B_RELBUF); 1133 bp->b_iocmd = BIO_WRITE; 1134 1135 if ((bp->b_flags & B_DELWRI) == 0) { 1136 bp->b_flags |= B_DONE | B_DELWRI; 1137 reassignbuf(bp, bp->b_vp); 1138 atomic_add_int(&numdirtybuffers, 1); 1139 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1140 } 1141} 1142 1143/* 1144 * bundirty: 1145 * 1146 * Clear B_DELWRI for buffer. 1147 * 1148 * Since the buffer is not on a queue, we do not update the numfreebuffers 1149 * count. 1150 * 1151 * Must be called at splbio(). 1152 * The buffer must be on QUEUE_NONE. 1153 */ 1154 1155void 1156bundirty(bp) 1157 struct buf *bp; 1158{ 1159 KASSERT(bp->b_qindex == QUEUE_NONE, 1160 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1161 1162 if (bp->b_flags & B_DELWRI) { 1163 bp->b_flags &= ~B_DELWRI; 1164 reassignbuf(bp, bp->b_vp); 1165 atomic_subtract_int(&numdirtybuffers, 1); 1166 numdirtywakeup(lodirtybuffers); 1167 } 1168 /* 1169 * Since it is now being written, we can clear its deferred write flag. 1170 */ 1171 bp->b_flags &= ~B_DEFERRED; 1172} 1173 1174/* 1175 * bawrite: 1176 * 1177 * Asynchronous write. Start output on a buffer, but do not wait for 1178 * it to complete. The buffer is released when the output completes. 1179 * 1180 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1181 * B_INVAL buffers. Not us. 1182 */ 1183void 1184bawrite(struct buf * bp) 1185{ 1186 bp->b_flags |= B_ASYNC; 1187 (void) BUF_WRITE(bp); 1188} 1189 1190/* 1191 * bwillwrite: 1192 * 1193 * Called prior to the locking of any vnodes when we are expecting to 1194 * write. We do not want to starve the buffer cache with too many 1195 * dirty buffers so we block here. By blocking prior to the locking 1196 * of any vnodes we attempt to avoid the situation where a locked vnode 1197 * prevents the various system daemons from flushing related buffers. 1198 */ 1199 1200void 1201bwillwrite(void) 1202{ 1203 if (numdirtybuffers >= hidirtybuffers) { 1204 int s; 1205 1206 mtx_lock(&Giant); 1207 s = splbio(); 1208 mtx_lock(&nblock); 1209 while (numdirtybuffers >= hidirtybuffers) { 1210 bd_wakeup(1); 1211 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1212 msleep(&needsbuffer, &nblock, 1213 (PRIBIO + 4), "flswai", 0); 1214 } 1215 splx(s); 1216 mtx_unlock(&nblock); 1217 mtx_unlock(&Giant); 1218 } 1219} 1220 1221/* 1222 * Return true if we have too many dirty buffers. 1223 */ 1224int 1225buf_dirty_count_severe(void) 1226{ 1227 return(numdirtybuffers >= hidirtybuffers); 1228} 1229 1230/* 1231 * brelse: 1232 * 1233 * Release a busy buffer and, if requested, free its resources. The 1234 * buffer will be stashed in the appropriate bufqueue[] allowing it 1235 * to be accessed later as a cache entity or reused for other purposes. 1236 */ 1237void 1238brelse(struct buf * bp) 1239{ 1240 int s; 1241 1242 GIANT_REQUIRED; 1243 1244 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1245 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1246 1247 s = splbio(); 1248 1249 if (bp->b_flags & B_LOCKED) 1250 bp->b_ioflags &= ~BIO_ERROR; 1251 1252 if (bp->b_iocmd == BIO_WRITE && 1253 (bp->b_ioflags & BIO_ERROR) && 1254 !(bp->b_flags & B_INVAL)) { 1255 /* 1256 * Failed write, redirty. Must clear BIO_ERROR to prevent 1257 * pages from being scrapped. If B_INVAL is set then 1258 * this case is not run and the next case is run to 1259 * destroy the buffer. B_INVAL can occur if the buffer 1260 * is outside the range supported by the underlying device. 1261 */ 1262 bp->b_ioflags &= ~BIO_ERROR; 1263 bdirty(bp); 1264 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1265 (bp->b_ioflags & BIO_ERROR) || 1266 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1267 /* 1268 * Either a failed I/O or we were asked to free or not 1269 * cache the buffer. 1270 */ 1271 bp->b_flags |= B_INVAL; 1272 if (LIST_FIRST(&bp->b_dep) != NULL) 1273 buf_deallocate(bp); 1274 if (bp->b_flags & B_DELWRI) { 1275 atomic_subtract_int(&numdirtybuffers, 1); 1276 numdirtywakeup(lodirtybuffers); 1277 } 1278 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1279 if ((bp->b_flags & B_VMIO) == 0) { 1280 if (bp->b_bufsize) 1281 allocbuf(bp, 0); 1282 if (bp->b_vp) 1283 brelvp(bp); 1284 } 1285 } 1286 1287 /* 1288 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1289 * is called with B_DELWRI set, the underlying pages may wind up 1290 * getting freed causing a previous write (bdwrite()) to get 'lost' 1291 * because pages associated with a B_DELWRI bp are marked clean. 1292 * 1293 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1294 * if B_DELWRI is set. 1295 * 1296 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1297 * on pages to return pages to the VM page queues. 1298 */ 1299 if (bp->b_flags & B_DELWRI) 1300 bp->b_flags &= ~B_RELBUF; 1301 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1302 bp->b_flags |= B_RELBUF; 1303 1304 /* 1305 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1306 * constituted, not even NFS buffers now. Two flags effect this. If 1307 * B_INVAL, the struct buf is invalidated but the VM object is kept 1308 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1309 * 1310 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1311 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1312 * buffer is also B_INVAL because it hits the re-dirtying code above. 1313 * 1314 * Normally we can do this whether a buffer is B_DELWRI or not. If 1315 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1316 * the commit state and we cannot afford to lose the buffer. If the 1317 * buffer has a background write in progress, we need to keep it 1318 * around to prevent it from being reconstituted and starting a second 1319 * background write. 1320 */ 1321 if ((bp->b_flags & B_VMIO) 1322 && !(bp->b_vp->v_mount != NULL && 1323 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1324 !vn_isdisk(bp->b_vp, NULL) && 1325 (bp->b_flags & B_DELWRI)) 1326 ) { 1327 1328 int i, j, resid; 1329 vm_page_t m; 1330 off_t foff; 1331 vm_pindex_t poff; 1332 vm_object_t obj; 1333 struct vnode *vp; 1334 1335 vp = bp->b_vp; 1336 obj = bp->b_object; 1337 1338 /* 1339 * Get the base offset and length of the buffer. Note that 1340 * in the VMIO case if the buffer block size is not 1341 * page-aligned then b_data pointer may not be page-aligned. 1342 * But our b_pages[] array *IS* page aligned. 1343 * 1344 * block sizes less then DEV_BSIZE (usually 512) are not 1345 * supported due to the page granularity bits (m->valid, 1346 * m->dirty, etc...). 1347 * 1348 * See man buf(9) for more information 1349 */ 1350 resid = bp->b_bufsize; 1351 foff = bp->b_offset; 1352 1353 for (i = 0; i < bp->b_npages; i++) { 1354 int had_bogus = 0; 1355 1356 m = bp->b_pages[i]; 1357 vm_page_lock_queues(); 1358 vm_page_flag_clear(m, PG_ZERO); 1359 vm_page_unlock_queues(); 1360 1361 /* 1362 * If we hit a bogus page, fixup *all* the bogus pages 1363 * now. 1364 */ 1365 if (m == bogus_page) { 1366 poff = OFF_TO_IDX(bp->b_offset); 1367 had_bogus = 1; 1368 1369 for (j = i; j < bp->b_npages; j++) { 1370 vm_page_t mtmp; 1371 mtmp = bp->b_pages[j]; 1372 if (mtmp == bogus_page) { 1373 mtmp = vm_page_lookup(obj, poff + j); 1374 if (!mtmp) { 1375 panic("brelse: page missing\n"); 1376 } 1377 bp->b_pages[j] = mtmp; 1378 } 1379 } 1380 1381 if ((bp->b_flags & B_INVAL) == 0) { 1382 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1383 } 1384 m = bp->b_pages[i]; 1385 } 1386 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1387 int poffset = foff & PAGE_MASK; 1388 int presid = resid > (PAGE_SIZE - poffset) ? 1389 (PAGE_SIZE - poffset) : resid; 1390 1391 KASSERT(presid >= 0, ("brelse: extra page")); 1392 vm_page_set_invalid(m, poffset, presid); 1393 if (had_bogus) 1394 printf("avoided corruption bug in bogus_page/brelse code\n"); 1395 } 1396 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1397 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1398 } 1399 1400 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1401 vfs_vmio_release(bp); 1402 1403 } else if (bp->b_flags & B_VMIO) { 1404 1405 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1406 vfs_vmio_release(bp); 1407 } 1408 1409 } 1410 1411 if (bp->b_qindex != QUEUE_NONE) 1412 panic("brelse: free buffer onto another queue???"); 1413 if (BUF_REFCNT(bp) > 1) { 1414 /* do not release to free list */ 1415 BUF_UNLOCK(bp); 1416 splx(s); 1417 return; 1418 } 1419 1420 /* enqueue */ 1421 mtx_lock(&bqlock); 1422 1423 /* buffers with no memory */ 1424 if (bp->b_bufsize == 0) { 1425 bp->b_flags |= B_INVAL; 1426 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1427 if (bp->b_xflags & BX_BKGRDINPROG) 1428 panic("losing buffer 1"); 1429 if (bp->b_kvasize) { 1430 bp->b_qindex = QUEUE_EMPTYKVA; 1431 } else { 1432 bp->b_qindex = QUEUE_EMPTY; 1433 } 1434 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1435#ifdef USE_BUFHASH 1436 LIST_REMOVE(bp, b_hash); 1437 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1438#endif 1439 bp->b_dev = NODEV; 1440 /* buffers with junk contents */ 1441 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1442 (bp->b_ioflags & BIO_ERROR)) { 1443 bp->b_flags |= B_INVAL; 1444 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1445 if (bp->b_xflags & BX_BKGRDINPROG) 1446 panic("losing buffer 2"); 1447 bp->b_qindex = QUEUE_CLEAN; 1448 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1449#ifdef USE_BUFHASH 1450 LIST_REMOVE(bp, b_hash); 1451 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1452#endif 1453 bp->b_dev = NODEV; 1454 1455 /* buffers that are locked */ 1456 } else if (bp->b_flags & B_LOCKED) { 1457 bp->b_qindex = QUEUE_LOCKED; 1458 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1459 1460 /* remaining buffers */ 1461 } else { 1462 if (bp->b_flags & B_DELWRI) 1463 bp->b_qindex = QUEUE_DIRTY; 1464 else 1465 bp->b_qindex = QUEUE_CLEAN; 1466 if (bp->b_flags & B_AGE) 1467 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1468 else 1469 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1470 } 1471 mtx_unlock(&bqlock); 1472 1473 /* 1474 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1475 * placed the buffer on the correct queue. We must also disassociate 1476 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1477 * find it. 1478 */ 1479 if (bp->b_flags & B_INVAL) { 1480 if (bp->b_flags & B_DELWRI) 1481 bundirty(bp); 1482 if (bp->b_vp) 1483 brelvp(bp); 1484 } 1485 1486 /* 1487 * Fixup numfreebuffers count. The bp is on an appropriate queue 1488 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1489 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1490 * if B_INVAL is set ). 1491 */ 1492 1493 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1494 bufcountwakeup(); 1495 1496 /* 1497 * Something we can maybe free or reuse 1498 */ 1499 if (bp->b_bufsize || bp->b_kvasize) 1500 bufspacewakeup(); 1501 1502 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | 1503 B_DIRECT | B_NOWDRAIN); 1504 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1505 panic("brelse: not dirty"); 1506 /* unlock */ 1507 BUF_UNLOCK(bp); 1508 splx(s); 1509} 1510 1511/* 1512 * Release a buffer back to the appropriate queue but do not try to free 1513 * it. The buffer is expected to be used again soon. 1514 * 1515 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1516 * biodone() to requeue an async I/O on completion. It is also used when 1517 * known good buffers need to be requeued but we think we may need the data 1518 * again soon. 1519 * 1520 * XXX we should be able to leave the B_RELBUF hint set on completion. 1521 */ 1522void 1523bqrelse(struct buf * bp) 1524{ 1525 int s; 1526 1527 s = splbio(); 1528 1529 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1530 1531 if (bp->b_qindex != QUEUE_NONE) 1532 panic("bqrelse: free buffer onto another queue???"); 1533 if (BUF_REFCNT(bp) > 1) { 1534 /* do not release to free list */ 1535 BUF_UNLOCK(bp); 1536 splx(s); 1537 return; 1538 } 1539 mtx_lock(&bqlock); 1540 if (bp->b_flags & B_LOCKED) { 1541 bp->b_ioflags &= ~BIO_ERROR; 1542 bp->b_qindex = QUEUE_LOCKED; 1543 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1544 /* buffers with stale but valid contents */ 1545 } else if (bp->b_flags & B_DELWRI) { 1546 bp->b_qindex = QUEUE_DIRTY; 1547 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1548 } else if (vm_page_count_severe()) { 1549 /* 1550 * We are too low on memory, we have to try to free the 1551 * buffer (most importantly: the wired pages making up its 1552 * backing store) *now*. 1553 */ 1554 mtx_unlock(&bqlock); 1555 splx(s); 1556 brelse(bp); 1557 return; 1558 } else { 1559 bp->b_qindex = QUEUE_CLEAN; 1560 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1561 } 1562 mtx_unlock(&bqlock); 1563 1564 if ((bp->b_flags & B_LOCKED) == 0 && 1565 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1566 bufcountwakeup(); 1567 } 1568 1569 /* 1570 * Something we can maybe free or reuse. 1571 */ 1572 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1573 bufspacewakeup(); 1574 1575 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1576 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1577 panic("bqrelse: not dirty"); 1578 /* unlock */ 1579 BUF_UNLOCK(bp); 1580 splx(s); 1581} 1582 1583/* Give pages used by the bp back to the VM system (where possible) */ 1584static void 1585vfs_vmio_release(bp) 1586 struct buf *bp; 1587{ 1588 int i; 1589 vm_page_t m; 1590 1591 GIANT_REQUIRED; 1592 vm_page_lock_queues(); 1593 for (i = 0; i < bp->b_npages; i++) { 1594 m = bp->b_pages[i]; 1595 bp->b_pages[i] = NULL; 1596 /* 1597 * In order to keep page LRU ordering consistent, put 1598 * everything on the inactive queue. 1599 */ 1600 vm_page_unwire(m, 0); 1601 /* 1602 * We don't mess with busy pages, it is 1603 * the responsibility of the process that 1604 * busied the pages to deal with them. 1605 */ 1606 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1607 continue; 1608 1609 if (m->wire_count == 0) { 1610 vm_page_flag_clear(m, PG_ZERO); 1611 /* 1612 * Might as well free the page if we can and it has 1613 * no valid data. We also free the page if the 1614 * buffer was used for direct I/O 1615 */ 1616 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1617 m->hold_count == 0) { 1618 vm_page_busy(m); 1619 pmap_remove_all(m); 1620 vm_page_free(m); 1621 } else if (bp->b_flags & B_DIRECT) { 1622 vm_page_try_to_free(m); 1623 } else if (vm_page_count_severe()) { 1624 vm_page_try_to_cache(m); 1625 } 1626 } 1627 } 1628 vm_page_unlock_queues(); 1629 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1630 1631 if (bp->b_bufsize) { 1632 bufspacewakeup(); 1633 bp->b_bufsize = 0; 1634 } 1635 bp->b_npages = 0; 1636 bp->b_flags &= ~B_VMIO; 1637 if (bp->b_vp) 1638 brelvp(bp); 1639} 1640 1641#ifdef USE_BUFHASH 1642/* 1643 * XXX MOVED TO VFS_SUBR.C 1644 * 1645 * Check to see if a block is currently memory resident. 1646 */ 1647struct buf * 1648gbincore(struct vnode * vp, daddr_t blkno) 1649{ 1650 struct buf *bp; 1651 struct bufhashhdr *bh; 1652 1653 bh = bufhash(vp, blkno); 1654 1655 /* Search hash chain */ 1656 LIST_FOREACH(bp, bh, b_hash) { 1657 /* hit */ 1658 if (bp->b_vp == vp && bp->b_lblkno == blkno && 1659 (bp->b_flags & B_INVAL) == 0) { 1660 break; 1661 } 1662 } 1663 return (bp); 1664} 1665#endif 1666 1667/* 1668 * Check to see if a block at a particular lbn is available for a clustered 1669 * write. 1670 */ 1671static int 1672vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1673{ 1674 struct buf *bpa; 1675 int match; 1676 1677 match = 0; 1678 1679 /* If the buf isn't in core skip it */ 1680 if ((bpa = gbincore(vp, lblkno)) == NULL) 1681 return (0); 1682 1683 /* If the buf is busy we don't want to wait for it */ 1684 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1685 return (0); 1686 1687 /* Only cluster with valid clusterable delayed write buffers */ 1688 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1689 (B_DELWRI | B_CLUSTEROK)) 1690 goto done; 1691 1692 if (bpa->b_bufsize != size) 1693 goto done; 1694 1695 /* 1696 * Check to see if it is in the expected place on disk and that the 1697 * block has been mapped. 1698 */ 1699 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1700 match = 1; 1701done: 1702 BUF_UNLOCK(bpa); 1703 return (match); 1704} 1705 1706/* 1707 * vfs_bio_awrite: 1708 * 1709 * Implement clustered async writes for clearing out B_DELWRI buffers. 1710 * This is much better then the old way of writing only one buffer at 1711 * a time. Note that we may not be presented with the buffers in the 1712 * correct order, so we search for the cluster in both directions. 1713 */ 1714int 1715vfs_bio_awrite(struct buf * bp) 1716{ 1717 int i; 1718 int j; 1719 daddr_t lblkno = bp->b_lblkno; 1720 struct vnode *vp = bp->b_vp; 1721 int s; 1722 int ncl; 1723 int nwritten; 1724 int size; 1725 int maxcl; 1726 1727 s = splbio(); 1728 /* 1729 * right now we support clustered writing only to regular files. If 1730 * we find a clusterable block we could be in the middle of a cluster 1731 * rather then at the beginning. 1732 */ 1733 if ((vp->v_type == VREG) && 1734 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1735 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1736 1737 size = vp->v_mount->mnt_stat.f_iosize; 1738 maxcl = MAXPHYS / size; 1739 1740 VI_LOCK(vp); 1741 for (i = 1; i < maxcl; i++) 1742 if (vfs_bio_clcheck(vp, size, lblkno + i, 1743 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1744 break; 1745 1746 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1747 if (vfs_bio_clcheck(vp, size, lblkno - j, 1748 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1749 break; 1750 1751 VI_UNLOCK(vp); 1752 --j; 1753 ncl = i + j; 1754 /* 1755 * this is a possible cluster write 1756 */ 1757 if (ncl != 1) { 1758 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1759 splx(s); 1760 return nwritten; 1761 } 1762 } 1763 1764 BUF_LOCK(bp, LK_EXCLUSIVE, NULL); 1765 bremfree(bp); 1766 bp->b_flags |= B_ASYNC; 1767 1768 splx(s); 1769 /* 1770 * default (old) behavior, writing out only one block 1771 * 1772 * XXX returns b_bufsize instead of b_bcount for nwritten? 1773 */ 1774 nwritten = bp->b_bufsize; 1775 (void) BUF_WRITE(bp); 1776 1777 return nwritten; 1778} 1779 1780/* 1781 * getnewbuf: 1782 * 1783 * Find and initialize a new buffer header, freeing up existing buffers 1784 * in the bufqueues as necessary. The new buffer is returned locked. 1785 * 1786 * Important: B_INVAL is not set. If the caller wishes to throw the 1787 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1788 * 1789 * We block if: 1790 * We have insufficient buffer headers 1791 * We have insufficient buffer space 1792 * buffer_map is too fragmented ( space reservation fails ) 1793 * If we have to flush dirty buffers ( but we try to avoid this ) 1794 * 1795 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1796 * Instead we ask the buf daemon to do it for us. We attempt to 1797 * avoid piecemeal wakeups of the pageout daemon. 1798 */ 1799 1800static struct buf * 1801getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1802{ 1803 struct buf *bp; 1804 struct buf *nbp; 1805 int defrag = 0; 1806 int nqindex; 1807 static int flushingbufs; 1808 1809 GIANT_REQUIRED; 1810 1811 /* 1812 * We can't afford to block since we might be holding a vnode lock, 1813 * which may prevent system daemons from running. We deal with 1814 * low-memory situations by proactively returning memory and running 1815 * async I/O rather then sync I/O. 1816 */ 1817 1818 atomic_add_int(&getnewbufcalls, 1); 1819 atomic_subtract_int(&getnewbufrestarts, 1); 1820restart: 1821 atomic_add_int(&getnewbufrestarts, 1); 1822 1823 /* 1824 * Setup for scan. If we do not have enough free buffers, 1825 * we setup a degenerate case that immediately fails. Note 1826 * that if we are specially marked process, we are allowed to 1827 * dip into our reserves. 1828 * 1829 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1830 * 1831 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1832 * However, there are a number of cases (defragging, reusing, ...) 1833 * where we cannot backup. 1834 */ 1835 mtx_lock(&bqlock); 1836 nqindex = QUEUE_EMPTYKVA; 1837 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1838 1839 if (nbp == NULL) { 1840 /* 1841 * If no EMPTYKVA buffers and we are either 1842 * defragging or reusing, locate a CLEAN buffer 1843 * to free or reuse. If bufspace useage is low 1844 * skip this step so we can allocate a new buffer. 1845 */ 1846 if (defrag || bufspace >= lobufspace) { 1847 nqindex = QUEUE_CLEAN; 1848 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1849 } 1850 1851 /* 1852 * If we could not find or were not allowed to reuse a 1853 * CLEAN buffer, check to see if it is ok to use an EMPTY 1854 * buffer. We can only use an EMPTY buffer if allocating 1855 * its KVA would not otherwise run us out of buffer space. 1856 */ 1857 if (nbp == NULL && defrag == 0 && 1858 bufspace + maxsize < hibufspace) { 1859 nqindex = QUEUE_EMPTY; 1860 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1861 } 1862 } 1863 1864 /* 1865 * Run scan, possibly freeing data and/or kva mappings on the fly 1866 * depending. 1867 */ 1868 1869 while ((bp = nbp) != NULL) { 1870 int qindex = nqindex; 1871 1872 /* 1873 * Calculate next bp ( we can only use it if we do not block 1874 * or do other fancy things ). 1875 */ 1876 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1877 switch(qindex) { 1878 case QUEUE_EMPTY: 1879 nqindex = QUEUE_EMPTYKVA; 1880 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1881 break; 1882 /* FALLTHROUGH */ 1883 case QUEUE_EMPTYKVA: 1884 nqindex = QUEUE_CLEAN; 1885 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1886 break; 1887 /* FALLTHROUGH */ 1888 case QUEUE_CLEAN: 1889 /* 1890 * nbp is NULL. 1891 */ 1892 break; 1893 } 1894 } 1895 1896 /* 1897 * Sanity Checks 1898 */ 1899 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1900 1901 /* 1902 * Note: we no longer distinguish between VMIO and non-VMIO 1903 * buffers. 1904 */ 1905 1906 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1907 1908 /* 1909 * If we are defragging then we need a buffer with 1910 * b_kvasize != 0. XXX this situation should no longer 1911 * occur, if defrag is non-zero the buffer's b_kvasize 1912 * should also be non-zero at this point. XXX 1913 */ 1914 if (defrag && bp->b_kvasize == 0) { 1915 printf("Warning: defrag empty buffer %p\n", bp); 1916 continue; 1917 } 1918 1919 /* 1920 * Start freeing the bp. This is somewhat involved. nbp 1921 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1922 */ 1923 1924 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1925 panic("getnewbuf: locked buf"); 1926 bremfreel(bp); 1927 mtx_unlock(&bqlock); 1928 1929 if (qindex == QUEUE_CLEAN) { 1930 if (bp->b_flags & B_VMIO) { 1931 bp->b_flags &= ~B_ASYNC; 1932 vfs_vmio_release(bp); 1933 } 1934 if (bp->b_vp) 1935 brelvp(bp); 1936 } 1937 1938 /* 1939 * NOTE: nbp is now entirely invalid. We can only restart 1940 * the scan from this point on. 1941 * 1942 * Get the rest of the buffer freed up. b_kva* is still 1943 * valid after this operation. 1944 */ 1945 1946 if (bp->b_rcred != NOCRED) { 1947 crfree(bp->b_rcred); 1948 bp->b_rcred = NOCRED; 1949 } 1950 if (bp->b_wcred != NOCRED) { 1951 crfree(bp->b_wcred); 1952 bp->b_wcred = NOCRED; 1953 } 1954 if (LIST_FIRST(&bp->b_dep) != NULL) 1955 buf_deallocate(bp); 1956 if (bp->b_xflags & BX_BKGRDINPROG) 1957 panic("losing buffer 3"); 1958#ifdef USE_BUFHASH 1959 LIST_REMOVE(bp, b_hash); 1960 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1961#endif 1962 1963 if (bp->b_bufsize) 1964 allocbuf(bp, 0); 1965 1966 bp->b_flags = 0; 1967 bp->b_ioflags = 0; 1968 bp->b_xflags = 0; 1969 bp->b_vflags = 0; 1970 bp->b_dev = NODEV; 1971 bp->b_vp = NULL; 1972 bp->b_blkno = bp->b_lblkno = 0; 1973 bp->b_offset = NOOFFSET; 1974 bp->b_iodone = 0; 1975 bp->b_error = 0; 1976 bp->b_resid = 0; 1977 bp->b_bcount = 0; 1978 bp->b_npages = 0; 1979 bp->b_dirtyoff = bp->b_dirtyend = 0; 1980 bp->b_magic = B_MAGIC_BIO; 1981 bp->b_op = &buf_ops_bio; 1982 bp->b_object = NULL; 1983 1984 LIST_INIT(&bp->b_dep); 1985 1986 /* 1987 * If we are defragging then free the buffer. 1988 */ 1989 if (defrag) { 1990 bp->b_flags |= B_INVAL; 1991 bfreekva(bp); 1992 brelse(bp); 1993 defrag = 0; 1994 goto restart; 1995 } 1996 1997 /* 1998 * If we are overcomitted then recover the buffer and its 1999 * KVM space. This occurs in rare situations when multiple 2000 * processes are blocked in getnewbuf() or allocbuf(). 2001 */ 2002 if (bufspace >= hibufspace) 2003 flushingbufs = 1; 2004 if (flushingbufs && bp->b_kvasize != 0) { 2005 bp->b_flags |= B_INVAL; 2006 bfreekva(bp); 2007 brelse(bp); 2008 goto restart; 2009 } 2010 if (bufspace < lobufspace) 2011 flushingbufs = 0; 2012 break; 2013 } 2014 2015 /* 2016 * If we exhausted our list, sleep as appropriate. We may have to 2017 * wakeup various daemons and write out some dirty buffers. 2018 * 2019 * Generally we are sleeping due to insufficient buffer space. 2020 */ 2021 2022 if (bp == NULL) { 2023 int flags; 2024 char *waitmsg; 2025 2026 mtx_unlock(&bqlock); 2027 if (defrag) { 2028 flags = VFS_BIO_NEED_BUFSPACE; 2029 waitmsg = "nbufkv"; 2030 } else if (bufspace >= hibufspace) { 2031 waitmsg = "nbufbs"; 2032 flags = VFS_BIO_NEED_BUFSPACE; 2033 } else { 2034 waitmsg = "newbuf"; 2035 flags = VFS_BIO_NEED_ANY; 2036 } 2037 2038 bd_speedup(); /* heeeelp */ 2039 2040 mtx_lock(&nblock); 2041 needsbuffer |= flags; 2042 while (needsbuffer & flags) { 2043 if (msleep(&needsbuffer, &nblock, 2044 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2045 mtx_unlock(&nblock); 2046 return (NULL); 2047 } 2048 } 2049 mtx_unlock(&nblock); 2050 } else { 2051 /* 2052 * We finally have a valid bp. We aren't quite out of the 2053 * woods, we still have to reserve kva space. In order 2054 * to keep fragmentation sane we only allocate kva in 2055 * BKVASIZE chunks. 2056 */ 2057 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2058 2059 if (maxsize != bp->b_kvasize) { 2060 vm_offset_t addr = 0; 2061 2062 bfreekva(bp); 2063 2064 if (vm_map_findspace(buffer_map, 2065 vm_map_min(buffer_map), maxsize, &addr)) { 2066 /* 2067 * Uh oh. Buffer map is to fragmented. We 2068 * must defragment the map. 2069 */ 2070 atomic_add_int(&bufdefragcnt, 1); 2071 defrag = 1; 2072 bp->b_flags |= B_INVAL; 2073 brelse(bp); 2074 goto restart; 2075 } 2076 if (addr) { 2077 vm_map_insert(buffer_map, NULL, 0, 2078 addr, addr + maxsize, 2079 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2080 2081 bp->b_kvabase = (caddr_t) addr; 2082 bp->b_kvasize = maxsize; 2083 atomic_add_int(&bufspace, bp->b_kvasize); 2084 atomic_add_int(&bufreusecnt, 1); 2085 } 2086 } 2087 bp->b_data = bp->b_kvabase; 2088 } 2089 return(bp); 2090} 2091 2092/* 2093 * buf_daemon: 2094 * 2095 * buffer flushing daemon. Buffers are normally flushed by the 2096 * update daemon but if it cannot keep up this process starts to 2097 * take the load in an attempt to prevent getnewbuf() from blocking. 2098 */ 2099 2100static struct proc *bufdaemonproc; 2101 2102static struct kproc_desc buf_kp = { 2103 "bufdaemon", 2104 buf_daemon, 2105 &bufdaemonproc 2106}; 2107SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2108 2109static void 2110buf_daemon() 2111{ 2112 int s; 2113 2114 mtx_lock(&Giant); 2115 2116 /* 2117 * This process needs to be suspended prior to shutdown sync. 2118 */ 2119 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2120 SHUTDOWN_PRI_LAST); 2121 2122 /* 2123 * This process is allowed to take the buffer cache to the limit 2124 */ 2125 s = splbio(); 2126 mtx_lock(&bdlock); 2127 2128 for (;;) { 2129 bd_request = 0; 2130 mtx_unlock(&bdlock); 2131 2132 kthread_suspend_check(bufdaemonproc); 2133 2134 /* 2135 * Do the flush. Limit the amount of in-transit I/O we 2136 * allow to build up, otherwise we would completely saturate 2137 * the I/O system. Wakeup any waiting processes before we 2138 * normally would so they can run in parallel with our drain. 2139 */ 2140 while (numdirtybuffers > lodirtybuffers) { 2141 if (flushbufqueues() == 0) 2142 break; 2143 waitrunningbufspace(); 2144 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2145 } 2146 2147 /* 2148 * Only clear bd_request if we have reached our low water 2149 * mark. The buf_daemon normally waits 1 second and 2150 * then incrementally flushes any dirty buffers that have 2151 * built up, within reason. 2152 * 2153 * If we were unable to hit our low water mark and couldn't 2154 * find any flushable buffers, we sleep half a second. 2155 * Otherwise we loop immediately. 2156 */ 2157 mtx_lock(&bdlock); 2158 if (numdirtybuffers <= lodirtybuffers) { 2159 /* 2160 * We reached our low water mark, reset the 2161 * request and sleep until we are needed again. 2162 * The sleep is just so the suspend code works. 2163 */ 2164 bd_request = 0; 2165 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2166 } else { 2167 /* 2168 * We couldn't find any flushable dirty buffers but 2169 * still have too many dirty buffers, we 2170 * have to sleep and try again. (rare) 2171 */ 2172 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2173 } 2174 } 2175} 2176 2177/* 2178 * flushbufqueues: 2179 * 2180 * Try to flush a buffer in the dirty queue. We must be careful to 2181 * free up B_INVAL buffers instead of write them, which NFS is 2182 * particularly sensitive to. 2183 */ 2184int flushwithdeps = 0; 2185SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2186 0, "Number of buffers flushed with dependecies that require rollbacks"); 2187static int 2188flushbufqueues(void) 2189{ 2190 struct thread *td = curthread; 2191 struct vnode *vp; 2192 struct buf *bp; 2193 2194 mtx_lock(&bqlock); 2195 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2196 KASSERT((bp->b_flags & B_DELWRI), 2197 ("unexpected clean buffer %p", bp)); 2198 if ((bp->b_xflags & BX_BKGRDINPROG) != 0) 2199 continue; 2200 if (bp->b_flags & B_INVAL) { 2201 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2202 panic("flushbufqueues: locked buf"); 2203 bremfreel(bp); 2204 mtx_unlock(&bqlock); 2205 brelse(bp); 2206 return (1); 2207 } 2208 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) 2209 continue; 2210 /* 2211 * We must hold the lock on a vnode before writing 2212 * one of its buffers. Otherwise we may confuse, or 2213 * in the case of a snapshot vnode, deadlock the 2214 * system. 2215 */ 2216 if ((vp = bp->b_vp) == NULL || 2217 vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2218 mtx_unlock(&bqlock); 2219 vfs_bio_awrite(bp); 2220 if (vp != NULL) 2221 VOP_UNLOCK(vp, 0, td); 2222 return (1); 2223 } 2224 } 2225 /* 2226 * Could not find any buffers without rollback dependencies, 2227 * so just write the first one in the hopes of eventually 2228 * making progress. 2229 */ 2230 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2231 KASSERT((bp->b_flags & B_DELWRI), 2232 ("unexpected clean buffer %p", bp)); 2233 if ((bp->b_xflags & BX_BKGRDINPROG) != 0) 2234 continue; 2235 if (bp->b_flags & B_INVAL) { 2236 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2237 panic("flushbufqueues: locked buf"); 2238 bremfreel(bp); 2239 mtx_unlock(&bqlock); 2240 brelse(bp); 2241 return (1); 2242 } 2243 /* 2244 * We must hold the lock on a vnode before writing 2245 * one of its buffers. Otherwise we may confuse, or 2246 * in the case of a snapshot vnode, deadlock the 2247 * system. 2248 */ 2249 if ((vp = bp->b_vp) == NULL || 2250 vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2251 mtx_unlock(&bqlock); 2252 vfs_bio_awrite(bp); 2253 if (vp != NULL) 2254 VOP_UNLOCK(vp, 0, td); 2255 flushwithdeps += 1; 2256 return (0); 2257 } 2258 } 2259 mtx_unlock(&bqlock); 2260 return (0); 2261} 2262 2263/* 2264 * Check to see if a block is currently memory resident. 2265 */ 2266struct buf * 2267incore(struct vnode * vp, daddr_t blkno) 2268{ 2269 struct buf *bp; 2270 2271 int s = splbio(); 2272 VI_LOCK(vp); 2273 bp = gbincore(vp, blkno); 2274 VI_UNLOCK(vp); 2275 splx(s); 2276 return (bp); 2277} 2278 2279/* 2280 * Returns true if no I/O is needed to access the 2281 * associated VM object. This is like incore except 2282 * it also hunts around in the VM system for the data. 2283 */ 2284 2285int 2286inmem(struct vnode * vp, daddr_t blkno) 2287{ 2288 vm_object_t obj; 2289 vm_offset_t toff, tinc, size; 2290 vm_page_t m; 2291 vm_ooffset_t off; 2292 2293 GIANT_REQUIRED; 2294 ASSERT_VOP_LOCKED(vp, "inmem"); 2295 2296 if (incore(vp, blkno)) 2297 return 1; 2298 if (vp->v_mount == NULL) 2299 return 0; 2300 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2301 return 0; 2302 2303 size = PAGE_SIZE; 2304 if (size > vp->v_mount->mnt_stat.f_iosize) 2305 size = vp->v_mount->mnt_stat.f_iosize; 2306 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2307 2308 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2309 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2310 if (!m) 2311 goto notinmem; 2312 tinc = size; 2313 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2314 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2315 if (vm_page_is_valid(m, 2316 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2317 goto notinmem; 2318 } 2319 return 1; 2320 2321notinmem: 2322 return (0); 2323} 2324 2325/* 2326 * vfs_setdirty: 2327 * 2328 * Sets the dirty range for a buffer based on the status of the dirty 2329 * bits in the pages comprising the buffer. 2330 * 2331 * The range is limited to the size of the buffer. 2332 * 2333 * This routine is primarily used by NFS, but is generalized for the 2334 * B_VMIO case. 2335 */ 2336static void 2337vfs_setdirty(struct buf *bp) 2338{ 2339 int i; 2340 vm_object_t object; 2341 2342 GIANT_REQUIRED; 2343 /* 2344 * Degenerate case - empty buffer 2345 */ 2346 2347 if (bp->b_bufsize == 0) 2348 return; 2349 2350 /* 2351 * We qualify the scan for modified pages on whether the 2352 * object has been flushed yet. The OBJ_WRITEABLE flag 2353 * is not cleared simply by protecting pages off. 2354 */ 2355 2356 if ((bp->b_flags & B_VMIO) == 0) 2357 return; 2358 2359 object = bp->b_pages[0]->object; 2360 2361 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2362 printf("Warning: object %p writeable but not mightbedirty\n", object); 2363 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2364 printf("Warning: object %p mightbedirty but not writeable\n", object); 2365 2366 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2367 vm_offset_t boffset; 2368 vm_offset_t eoffset; 2369 2370 vm_page_lock_queues(); 2371 /* 2372 * test the pages to see if they have been modified directly 2373 * by users through the VM system. 2374 */ 2375 for (i = 0; i < bp->b_npages; i++) { 2376 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2377 vm_page_test_dirty(bp->b_pages[i]); 2378 } 2379 2380 /* 2381 * Calculate the encompassing dirty range, boffset and eoffset, 2382 * (eoffset - boffset) bytes. 2383 */ 2384 2385 for (i = 0; i < bp->b_npages; i++) { 2386 if (bp->b_pages[i]->dirty) 2387 break; 2388 } 2389 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2390 2391 for (i = bp->b_npages - 1; i >= 0; --i) { 2392 if (bp->b_pages[i]->dirty) { 2393 break; 2394 } 2395 } 2396 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2397 2398 vm_page_unlock_queues(); 2399 /* 2400 * Fit it to the buffer. 2401 */ 2402 2403 if (eoffset > bp->b_bcount) 2404 eoffset = bp->b_bcount; 2405 2406 /* 2407 * If we have a good dirty range, merge with the existing 2408 * dirty range. 2409 */ 2410 2411 if (boffset < eoffset) { 2412 if (bp->b_dirtyoff > boffset) 2413 bp->b_dirtyoff = boffset; 2414 if (bp->b_dirtyend < eoffset) 2415 bp->b_dirtyend = eoffset; 2416 } 2417 } 2418} 2419 2420/* 2421 * getblk: 2422 * 2423 * Get a block given a specified block and offset into a file/device. 2424 * The buffers B_DONE bit will be cleared on return, making it almost 2425 * ready for an I/O initiation. B_INVAL may or may not be set on 2426 * return. The caller should clear B_INVAL prior to initiating a 2427 * READ. 2428 * 2429 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2430 * an existing buffer. 2431 * 2432 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2433 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2434 * and then cleared based on the backing VM. If the previous buffer is 2435 * non-0-sized but invalid, B_CACHE will be cleared. 2436 * 2437 * If getblk() must create a new buffer, the new buffer is returned with 2438 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2439 * case it is returned with B_INVAL clear and B_CACHE set based on the 2440 * backing VM. 2441 * 2442 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2443 * B_CACHE bit is clear. 2444 * 2445 * What this means, basically, is that the caller should use B_CACHE to 2446 * determine whether the buffer is fully valid or not and should clear 2447 * B_INVAL prior to issuing a read. If the caller intends to validate 2448 * the buffer by loading its data area with something, the caller needs 2449 * to clear B_INVAL. If the caller does this without issuing an I/O, 2450 * the caller should set B_CACHE ( as an optimization ), else the caller 2451 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2452 * a write attempt or if it was a successfull read. If the caller 2453 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2454 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2455 */ 2456struct buf * 2457getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2458{ 2459 struct buf *bp; 2460 int s; 2461 int error; 2462#ifdef USE_BUFHASH 2463 struct bufhashhdr *bh; 2464#endif 2465 ASSERT_VOP_LOCKED(vp, "getblk"); 2466 2467 if (size > MAXBSIZE) 2468 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2469 2470 s = splbio(); 2471loop: 2472 /* 2473 * Block if we are low on buffers. Certain processes are allowed 2474 * to completely exhaust the buffer cache. 2475 * 2476 * If this check ever becomes a bottleneck it may be better to 2477 * move it into the else, when gbincore() fails. At the moment 2478 * it isn't a problem. 2479 * 2480 * XXX remove if 0 sections (clean this up after its proven) 2481 */ 2482 if (numfreebuffers == 0) { 2483 if (curthread == PCPU_GET(idlethread)) 2484 return NULL; 2485 mtx_lock(&nblock); 2486 needsbuffer |= VFS_BIO_NEED_ANY; 2487 mtx_unlock(&nblock); 2488 } 2489 2490 VI_LOCK(vp); 2491 if ((bp = gbincore(vp, blkno))) { 2492 /* 2493 * Buffer is in-core. If the buffer is not busy, it must 2494 * be on a queue. 2495 */ 2496 2497 error = BUF_TIMELOCK(bp, 2498 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 2499 VI_MTX(vp), "getblk", slpflag, slptimeo); 2500 2501 /* 2502 * If we slept and got the lock we have to restart in case 2503 * the buffer changed identities. 2504 */ 2505 if (error == ENOLCK) 2506 goto loop; 2507 /* We timed out or were interrupted. */ 2508 else if (error) 2509 return (NULL); 2510 2511 /* 2512 * The buffer is locked. B_CACHE is cleared if the buffer is 2513 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2514 * and for a VMIO buffer B_CACHE is adjusted according to the 2515 * backing VM cache. 2516 */ 2517 if (bp->b_flags & B_INVAL) 2518 bp->b_flags &= ~B_CACHE; 2519 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2520 bp->b_flags |= B_CACHE; 2521 bremfree(bp); 2522 2523 /* 2524 * check for size inconsistancies for non-VMIO case. 2525 */ 2526 2527 if (bp->b_bcount != size) { 2528 if ((bp->b_flags & B_VMIO) == 0 || 2529 (size > bp->b_kvasize)) { 2530 if (bp->b_flags & B_DELWRI) { 2531 bp->b_flags |= B_NOCACHE; 2532 BUF_WRITE(bp); 2533 } else { 2534 if ((bp->b_flags & B_VMIO) && 2535 (LIST_FIRST(&bp->b_dep) == NULL)) { 2536 bp->b_flags |= B_RELBUF; 2537 brelse(bp); 2538 } else { 2539 bp->b_flags |= B_NOCACHE; 2540 BUF_WRITE(bp); 2541 } 2542 } 2543 goto loop; 2544 } 2545 } 2546 2547 /* 2548 * If the size is inconsistant in the VMIO case, we can resize 2549 * the buffer. This might lead to B_CACHE getting set or 2550 * cleared. If the size has not changed, B_CACHE remains 2551 * unchanged from its previous state. 2552 */ 2553 2554 if (bp->b_bcount != size) 2555 allocbuf(bp, size); 2556 2557 KASSERT(bp->b_offset != NOOFFSET, 2558 ("getblk: no buffer offset")); 2559 2560 /* 2561 * A buffer with B_DELWRI set and B_CACHE clear must 2562 * be committed before we can return the buffer in 2563 * order to prevent the caller from issuing a read 2564 * ( due to B_CACHE not being set ) and overwriting 2565 * it. 2566 * 2567 * Most callers, including NFS and FFS, need this to 2568 * operate properly either because they assume they 2569 * can issue a read if B_CACHE is not set, or because 2570 * ( for example ) an uncached B_DELWRI might loop due 2571 * to softupdates re-dirtying the buffer. In the latter 2572 * case, B_CACHE is set after the first write completes, 2573 * preventing further loops. 2574 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2575 * above while extending the buffer, we cannot allow the 2576 * buffer to remain with B_CACHE set after the write 2577 * completes or it will represent a corrupt state. To 2578 * deal with this we set B_NOCACHE to scrap the buffer 2579 * after the write. 2580 * 2581 * We might be able to do something fancy, like setting 2582 * B_CACHE in bwrite() except if B_DELWRI is already set, 2583 * so the below call doesn't set B_CACHE, but that gets real 2584 * confusing. This is much easier. 2585 */ 2586 2587 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2588 bp->b_flags |= B_NOCACHE; 2589 BUF_WRITE(bp); 2590 goto loop; 2591 } 2592 2593 splx(s); 2594 bp->b_flags &= ~B_DONE; 2595 } else { 2596 int bsize, maxsize, vmio; 2597 off_t offset; 2598 2599 /* 2600 * Buffer is not in-core, create new buffer. The buffer 2601 * returned by getnewbuf() is locked. Note that the returned 2602 * buffer is also considered valid (not marked B_INVAL). 2603 */ 2604 VI_UNLOCK(vp); 2605 if (vn_isdisk(vp, NULL)) 2606 bsize = DEV_BSIZE; 2607 else if (vp->v_mountedhere) 2608 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2609 else if (vp->v_mount) 2610 bsize = vp->v_mount->mnt_stat.f_iosize; 2611 else 2612 bsize = size; 2613 2614 offset = blkno * bsize; 2615 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2616 (vp->v_vflag & VV_OBJBUF); 2617 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2618 maxsize = imax(maxsize, bsize); 2619 2620 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2621 if (slpflag || slptimeo) { 2622 splx(s); 2623 return NULL; 2624 } 2625 goto loop; 2626 } 2627 2628 /* 2629 * This code is used to make sure that a buffer is not 2630 * created while the getnewbuf routine is blocked. 2631 * This can be a problem whether the vnode is locked or not. 2632 * If the buffer is created out from under us, we have to 2633 * throw away the one we just created. There is now window 2634 * race because we are safely running at splbio() from the 2635 * point of the duplicate buffer creation through to here, 2636 * and we've locked the buffer. 2637 * 2638 * Note: this must occur before we associate the buffer 2639 * with the vp especially considering limitations in 2640 * the splay tree implementation when dealing with duplicate 2641 * lblkno's. 2642 */ 2643 VI_LOCK(vp); 2644 if (gbincore(vp, blkno)) { 2645 VI_UNLOCK(vp); 2646 bp->b_flags |= B_INVAL; 2647 brelse(bp); 2648 goto loop; 2649 } 2650 VI_UNLOCK(vp); 2651 2652 /* 2653 * Insert the buffer into the hash, so that it can 2654 * be found by incore. 2655 */ 2656 bp->b_blkno = bp->b_lblkno = blkno; 2657 bp->b_offset = offset; 2658 2659 bgetvp(vp, bp); 2660#ifdef USE_BUFHASH 2661 LIST_REMOVE(bp, b_hash); 2662 bh = bufhash(vp, blkno); 2663 LIST_INSERT_HEAD(bh, bp, b_hash); 2664#endif 2665 2666 /* 2667 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2668 * buffer size starts out as 0, B_CACHE will be set by 2669 * allocbuf() for the VMIO case prior to it testing the 2670 * backing store for validity. 2671 */ 2672 2673 if (vmio) { 2674 bp->b_flags |= B_VMIO; 2675#if defined(VFS_BIO_DEBUG) 2676 if (vp->v_type != VREG) 2677 printf("getblk: vmioing file type %d???\n", vp->v_type); 2678#endif 2679 VOP_GETVOBJECT(vp, &bp->b_object); 2680 } else { 2681 bp->b_flags &= ~B_VMIO; 2682 bp->b_object = NULL; 2683 } 2684 2685 allocbuf(bp, size); 2686 2687 splx(s); 2688 bp->b_flags &= ~B_DONE; 2689 } 2690 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2691 return (bp); 2692} 2693 2694/* 2695 * Get an empty, disassociated buffer of given size. The buffer is initially 2696 * set to B_INVAL. 2697 */ 2698struct buf * 2699geteblk(int size) 2700{ 2701 struct buf *bp; 2702 int s; 2703 int maxsize; 2704 2705 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2706 2707 s = splbio(); 2708 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2709 continue; 2710 splx(s); 2711 allocbuf(bp, size); 2712 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2713 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2714 return (bp); 2715} 2716 2717 2718/* 2719 * This code constitutes the buffer memory from either anonymous system 2720 * memory (in the case of non-VMIO operations) or from an associated 2721 * VM object (in the case of VMIO operations). This code is able to 2722 * resize a buffer up or down. 2723 * 2724 * Note that this code is tricky, and has many complications to resolve 2725 * deadlock or inconsistant data situations. Tread lightly!!! 2726 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2727 * the caller. Calling this code willy nilly can result in the loss of data. 2728 * 2729 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2730 * B_CACHE for the non-VMIO case. 2731 */ 2732 2733int 2734allocbuf(struct buf *bp, int size) 2735{ 2736 int newbsize, mbsize; 2737 int i; 2738 2739 GIANT_REQUIRED; 2740 2741 if (BUF_REFCNT(bp) == 0) 2742 panic("allocbuf: buffer not busy"); 2743 2744 if (bp->b_kvasize < size) 2745 panic("allocbuf: buffer too small"); 2746 2747 if ((bp->b_flags & B_VMIO) == 0) { 2748 caddr_t origbuf; 2749 int origbufsize; 2750 /* 2751 * Just get anonymous memory from the kernel. Don't 2752 * mess with B_CACHE. 2753 */ 2754 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2755 if (bp->b_flags & B_MALLOC) 2756 newbsize = mbsize; 2757 else 2758 newbsize = round_page(size); 2759 2760 if (newbsize < bp->b_bufsize) { 2761 /* 2762 * malloced buffers are not shrunk 2763 */ 2764 if (bp->b_flags & B_MALLOC) { 2765 if (newbsize) { 2766 bp->b_bcount = size; 2767 } else { 2768 free(bp->b_data, M_BIOBUF); 2769 if (bp->b_bufsize) { 2770 atomic_subtract_int( 2771 &bufmallocspace, 2772 bp->b_bufsize); 2773 bufspacewakeup(); 2774 bp->b_bufsize = 0; 2775 } 2776 bp->b_data = bp->b_kvabase; 2777 bp->b_bcount = 0; 2778 bp->b_flags &= ~B_MALLOC; 2779 } 2780 return 1; 2781 } 2782 vm_hold_free_pages( 2783 bp, 2784 (vm_offset_t) bp->b_data + newbsize, 2785 (vm_offset_t) bp->b_data + bp->b_bufsize); 2786 } else if (newbsize > bp->b_bufsize) { 2787 /* 2788 * We only use malloced memory on the first allocation. 2789 * and revert to page-allocated memory when the buffer 2790 * grows. 2791 */ 2792 /* 2793 * There is a potential smp race here that could lead 2794 * to bufmallocspace slightly passing the max. It 2795 * is probably extremely rare and not worth worrying 2796 * over. 2797 */ 2798 if ( (bufmallocspace < maxbufmallocspace) && 2799 (bp->b_bufsize == 0) && 2800 (mbsize <= PAGE_SIZE/2)) { 2801 2802 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2803 bp->b_bufsize = mbsize; 2804 bp->b_bcount = size; 2805 bp->b_flags |= B_MALLOC; 2806 atomic_add_int(&bufmallocspace, mbsize); 2807 return 1; 2808 } 2809 origbuf = NULL; 2810 origbufsize = 0; 2811 /* 2812 * If the buffer is growing on its other-than-first allocation, 2813 * then we revert to the page-allocation scheme. 2814 */ 2815 if (bp->b_flags & B_MALLOC) { 2816 origbuf = bp->b_data; 2817 origbufsize = bp->b_bufsize; 2818 bp->b_data = bp->b_kvabase; 2819 if (bp->b_bufsize) { 2820 atomic_subtract_int(&bufmallocspace, 2821 bp->b_bufsize); 2822 bufspacewakeup(); 2823 bp->b_bufsize = 0; 2824 } 2825 bp->b_flags &= ~B_MALLOC; 2826 newbsize = round_page(newbsize); 2827 } 2828 vm_hold_load_pages( 2829 bp, 2830 (vm_offset_t) bp->b_data + bp->b_bufsize, 2831 (vm_offset_t) bp->b_data + newbsize); 2832 if (origbuf) { 2833 bcopy(origbuf, bp->b_data, origbufsize); 2834 free(origbuf, M_BIOBUF); 2835 } 2836 } 2837 } else { 2838 int desiredpages; 2839 2840 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2841 desiredpages = (size == 0) ? 0 : 2842 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2843 2844 if (bp->b_flags & B_MALLOC) 2845 panic("allocbuf: VMIO buffer can't be malloced"); 2846 /* 2847 * Set B_CACHE initially if buffer is 0 length or will become 2848 * 0-length. 2849 */ 2850 if (size == 0 || bp->b_bufsize == 0) 2851 bp->b_flags |= B_CACHE; 2852 2853 if (newbsize < bp->b_bufsize) { 2854 /* 2855 * DEV_BSIZE aligned new buffer size is less then the 2856 * DEV_BSIZE aligned existing buffer size. Figure out 2857 * if we have to remove any pages. 2858 */ 2859 if (desiredpages < bp->b_npages) { 2860 vm_page_t m; 2861 2862 vm_page_lock_queues(); 2863 for (i = desiredpages; i < bp->b_npages; i++) { 2864 /* 2865 * the page is not freed here -- it 2866 * is the responsibility of 2867 * vnode_pager_setsize 2868 */ 2869 m = bp->b_pages[i]; 2870 KASSERT(m != bogus_page, 2871 ("allocbuf: bogus page found")); 2872 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2873 vm_page_lock_queues(); 2874 2875 bp->b_pages[i] = NULL; 2876 vm_page_unwire(m, 0); 2877 } 2878 vm_page_unlock_queues(); 2879 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2880 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2881 bp->b_npages = desiredpages; 2882 } 2883 } else if (size > bp->b_bcount) { 2884 /* 2885 * We are growing the buffer, possibly in a 2886 * byte-granular fashion. 2887 */ 2888 struct vnode *vp; 2889 vm_object_t obj; 2890 vm_offset_t toff; 2891 vm_offset_t tinc; 2892 2893 /* 2894 * Step 1, bring in the VM pages from the object, 2895 * allocating them if necessary. We must clear 2896 * B_CACHE if these pages are not valid for the 2897 * range covered by the buffer. 2898 */ 2899 2900 vp = bp->b_vp; 2901 obj = bp->b_object; 2902 2903 while (bp->b_npages < desiredpages) { 2904 vm_page_t m; 2905 vm_pindex_t pi; 2906 2907 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2908 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2909 /* 2910 * note: must allocate system pages 2911 * since blocking here could intefere 2912 * with paging I/O, no matter which 2913 * process we are. 2914 */ 2915 m = vm_page_alloc(obj, pi, 2916 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2917 if (m == NULL) { 2918 atomic_add_int(&vm_pageout_deficit, 2919 desiredpages - bp->b_npages); 2920 VM_WAIT; 2921 } else { 2922 vm_page_lock_queues(); 2923 vm_page_wakeup(m); 2924 vm_page_unlock_queues(); 2925 bp->b_flags &= ~B_CACHE; 2926 bp->b_pages[bp->b_npages] = m; 2927 ++bp->b_npages; 2928 } 2929 continue; 2930 } 2931 2932 /* 2933 * We found a page. If we have to sleep on it, 2934 * retry because it might have gotten freed out 2935 * from under us. 2936 * 2937 * We can only test PG_BUSY here. Blocking on 2938 * m->busy might lead to a deadlock: 2939 * 2940 * vm_fault->getpages->cluster_read->allocbuf 2941 * 2942 */ 2943 vm_page_lock_queues(); 2944 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2945 continue; 2946 2947 /* 2948 * We have a good page. Should we wakeup the 2949 * page daemon? 2950 */ 2951 if ((curproc != pageproc) && 2952 ((m->queue - m->pc) == PQ_CACHE) && 2953 ((cnt.v_free_count + cnt.v_cache_count) < 2954 (cnt.v_free_min + cnt.v_cache_min))) { 2955 pagedaemon_wakeup(); 2956 } 2957 vm_page_flag_clear(m, PG_ZERO); 2958 vm_page_wire(m); 2959 vm_page_unlock_queues(); 2960 bp->b_pages[bp->b_npages] = m; 2961 ++bp->b_npages; 2962 } 2963 2964 /* 2965 * Step 2. We've loaded the pages into the buffer, 2966 * we have to figure out if we can still have B_CACHE 2967 * set. Note that B_CACHE is set according to the 2968 * byte-granular range ( bcount and size ), new the 2969 * aligned range ( newbsize ). 2970 * 2971 * The VM test is against m->valid, which is DEV_BSIZE 2972 * aligned. Needless to say, the validity of the data 2973 * needs to also be DEV_BSIZE aligned. Note that this 2974 * fails with NFS if the server or some other client 2975 * extends the file's EOF. If our buffer is resized, 2976 * B_CACHE may remain set! XXX 2977 */ 2978 2979 toff = bp->b_bcount; 2980 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2981 2982 while ((bp->b_flags & B_CACHE) && toff < size) { 2983 vm_pindex_t pi; 2984 2985 if (tinc > (size - toff)) 2986 tinc = size - toff; 2987 2988 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2989 PAGE_SHIFT; 2990 2991 vfs_buf_test_cache( 2992 bp, 2993 bp->b_offset, 2994 toff, 2995 tinc, 2996 bp->b_pages[pi] 2997 ); 2998 toff += tinc; 2999 tinc = PAGE_SIZE; 3000 } 3001 3002 /* 3003 * Step 3, fixup the KVM pmap. Remember that 3004 * bp->b_data is relative to bp->b_offset, but 3005 * bp->b_offset may be offset into the first page. 3006 */ 3007 3008 bp->b_data = (caddr_t) 3009 trunc_page((vm_offset_t)bp->b_data); 3010 pmap_qenter( 3011 (vm_offset_t)bp->b_data, 3012 bp->b_pages, 3013 bp->b_npages 3014 ); 3015 3016 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3017 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 3018 } 3019 } 3020 if (newbsize < bp->b_bufsize) 3021 bufspacewakeup(); 3022 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3023 bp->b_bcount = size; /* requested buffer size */ 3024 return 1; 3025} 3026 3027void 3028biodone(struct bio *bp) 3029{ 3030 bp->bio_flags |= BIO_DONE; 3031 if (bp->bio_done != NULL) 3032 bp->bio_done(bp); 3033 else 3034 wakeup(bp); 3035} 3036 3037/* 3038 * Wait for a BIO to finish. 3039 * 3040 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3041 * case is not yet clear. 3042 */ 3043int 3044biowait(struct bio *bp, const char *wchan) 3045{ 3046 3047 while ((bp->bio_flags & BIO_DONE) == 0) 3048 msleep(bp, NULL, PRIBIO, wchan, hz / 10); 3049 if (bp->bio_error != 0) 3050 return (bp->bio_error); 3051 if (!(bp->bio_flags & BIO_ERROR)) 3052 return (0); 3053 return (EIO); 3054} 3055 3056void 3057biofinish(struct bio *bp, struct devstat *stat, int error) 3058{ 3059 3060 if (error) { 3061 bp->bio_error = error; 3062 bp->bio_flags |= BIO_ERROR; 3063 } 3064 if (stat != NULL) 3065 devstat_end_transaction_bio(stat, bp); 3066 biodone(bp); 3067} 3068 3069void 3070bioq_init(struct bio_queue_head *head) 3071{ 3072 TAILQ_INIT(&head->queue); 3073 head->last_pblkno = 0; 3074 head->insert_point = NULL; 3075 head->switch_point = NULL; 3076} 3077 3078void 3079bioq_remove(struct bio_queue_head *head, struct bio *bp) 3080{ 3081 if (bp == head->switch_point) 3082 head->switch_point = TAILQ_NEXT(bp, bio_queue); 3083 if (bp == head->insert_point) { 3084 head->insert_point = TAILQ_PREV(bp, bio_queue, bio_queue); 3085 if (head->insert_point == NULL) 3086 head->last_pblkno = 0; 3087 } else if (bp == TAILQ_FIRST(&head->queue)) 3088 head->last_pblkno = bp->bio_pblkno; 3089 TAILQ_REMOVE(&head->queue, bp, bio_queue); 3090 if (TAILQ_FIRST(&head->queue) == head->switch_point) 3091 head->switch_point = NULL; 3092} 3093 3094/* 3095 * bufwait: 3096 * 3097 * Wait for buffer I/O completion, returning error status. The buffer 3098 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3099 * error and cleared. 3100 */ 3101int 3102bufwait(register struct buf * bp) 3103{ 3104 int s; 3105 3106 s = splbio(); 3107 while ((bp->b_flags & B_DONE) == 0) { 3108 if (bp->b_iocmd == BIO_READ) 3109 tsleep(bp, PRIBIO, "biord", 0); 3110 else 3111 tsleep(bp, PRIBIO, "biowr", 0); 3112 } 3113 splx(s); 3114 if (bp->b_flags & B_EINTR) { 3115 bp->b_flags &= ~B_EINTR; 3116 return (EINTR); 3117 } 3118 if (bp->b_ioflags & BIO_ERROR) { 3119 return (bp->b_error ? bp->b_error : EIO); 3120 } else { 3121 return (0); 3122 } 3123} 3124 3125 /* 3126 * Call back function from struct bio back up to struct buf. 3127 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 3128 */ 3129void 3130bufdonebio(struct bio *bp) 3131{ 3132 bufdone(bp->bio_caller2); 3133} 3134 3135/* 3136 * bufdone: 3137 * 3138 * Finish I/O on a buffer, optionally calling a completion function. 3139 * This is usually called from an interrupt so process blocking is 3140 * not allowed. 3141 * 3142 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3143 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3144 * assuming B_INVAL is clear. 3145 * 3146 * For the VMIO case, we set B_CACHE if the op was a read and no 3147 * read error occured, or if the op was a write. B_CACHE is never 3148 * set if the buffer is invalid or otherwise uncacheable. 3149 * 3150 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3151 * initiator to leave B_INVAL set to brelse the buffer out of existance 3152 * in the biodone routine. 3153 */ 3154void 3155bufdone(struct buf *bp) 3156{ 3157 int s; 3158 void (*biodone)(struct buf *); 3159 3160 GIANT_REQUIRED; 3161 3162 s = splbio(); 3163 3164 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3165 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3166 3167 bp->b_flags |= B_DONE; 3168 runningbufwakeup(bp); 3169 3170 if (bp->b_iocmd == BIO_DELETE) { 3171 brelse(bp); 3172 splx(s); 3173 return; 3174 } 3175 3176 if (bp->b_iocmd == BIO_WRITE) { 3177 vwakeup(bp); 3178 } 3179 3180 /* call optional completion function if requested */ 3181 if (bp->b_iodone != NULL) { 3182 biodone = bp->b_iodone; 3183 bp->b_iodone = NULL; 3184 (*biodone) (bp); 3185 splx(s); 3186 return; 3187 } 3188 if (LIST_FIRST(&bp->b_dep) != NULL) 3189 buf_complete(bp); 3190 3191 if (bp->b_flags & B_VMIO) { 3192 int i; 3193 vm_ooffset_t foff; 3194 vm_page_t m; 3195 vm_object_t obj; 3196 int iosize; 3197 struct vnode *vp = bp->b_vp; 3198 3199 obj = bp->b_object; 3200 3201#if defined(VFS_BIO_DEBUG) 3202 mp_fixme("usecount and vflag accessed without locks."); 3203 if (vp->v_usecount == 0) { 3204 panic("biodone: zero vnode ref count"); 3205 } 3206 3207 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3208 panic("biodone: vnode is not setup for merged cache"); 3209 } 3210#endif 3211 3212 foff = bp->b_offset; 3213 KASSERT(bp->b_offset != NOOFFSET, 3214 ("biodone: no buffer offset")); 3215 3216#if defined(VFS_BIO_DEBUG) 3217 if (obj->paging_in_progress < bp->b_npages) { 3218 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3219 obj->paging_in_progress, bp->b_npages); 3220 } 3221#endif 3222 3223 /* 3224 * Set B_CACHE if the op was a normal read and no error 3225 * occured. B_CACHE is set for writes in the b*write() 3226 * routines. 3227 */ 3228 iosize = bp->b_bcount - bp->b_resid; 3229 if (bp->b_iocmd == BIO_READ && 3230 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3231 !(bp->b_ioflags & BIO_ERROR)) { 3232 bp->b_flags |= B_CACHE; 3233 } 3234 vm_page_lock_queues(); 3235 for (i = 0; i < bp->b_npages; i++) { 3236 int bogusflag = 0; 3237 int resid; 3238 3239 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3240 if (resid > iosize) 3241 resid = iosize; 3242 3243 /* 3244 * cleanup bogus pages, restoring the originals 3245 */ 3246 m = bp->b_pages[i]; 3247 if (m == bogus_page) { 3248 bogusflag = 1; 3249 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3250 if (m == NULL) 3251 panic("biodone: page disappeared!"); 3252 bp->b_pages[i] = m; 3253 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3254 } 3255#if defined(VFS_BIO_DEBUG) 3256 if (OFF_TO_IDX(foff) != m->pindex) { 3257 printf( 3258"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3259 (intmax_t)foff, (uintmax_t)m->pindex); 3260 } 3261#endif 3262 3263 /* 3264 * In the write case, the valid and clean bits are 3265 * already changed correctly ( see bdwrite() ), so we 3266 * only need to do this here in the read case. 3267 */ 3268 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3269 vfs_page_set_valid(bp, foff, i, m); 3270 } 3271 vm_page_flag_clear(m, PG_ZERO); 3272 3273 /* 3274 * when debugging new filesystems or buffer I/O methods, this 3275 * is the most common error that pops up. if you see this, you 3276 * have not set the page busy flag correctly!!! 3277 */ 3278 if (m->busy == 0) { 3279 printf("biodone: page busy < 0, " 3280 "pindex: %d, foff: 0x(%x,%x), " 3281 "resid: %d, index: %d\n", 3282 (int) m->pindex, (int)(foff >> 32), 3283 (int) foff & 0xffffffff, resid, i); 3284 if (!vn_isdisk(vp, NULL)) 3285 printf(" iosize: %ld, lblkno: %jd, flags: 0x%x, npages: %d\n", 3286 bp->b_vp->v_mount->mnt_stat.f_iosize, 3287 (intmax_t) bp->b_lblkno, 3288 bp->b_flags, bp->b_npages); 3289 else 3290 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3291 (intmax_t) bp->b_lblkno, 3292 bp->b_flags, bp->b_npages); 3293 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 3294 m->valid, m->dirty, m->wire_count); 3295 panic("biodone: page busy < 0\n"); 3296 } 3297 vm_page_io_finish(m); 3298 vm_object_pip_subtract(obj, 1); 3299 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3300 iosize -= resid; 3301 } 3302 vm_page_unlock_queues(); 3303 if (obj) 3304 vm_object_pip_wakeupn(obj, 0); 3305 } 3306 3307 /* 3308 * For asynchronous completions, release the buffer now. The brelse 3309 * will do a wakeup there if necessary - so no need to do a wakeup 3310 * here in the async case. The sync case always needs to do a wakeup. 3311 */ 3312 3313 if (bp->b_flags & B_ASYNC) { 3314 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3315 brelse(bp); 3316 else 3317 bqrelse(bp); 3318 } else { 3319 wakeup(bp); 3320 } 3321 splx(s); 3322} 3323 3324/* 3325 * This routine is called in lieu of iodone in the case of 3326 * incomplete I/O. This keeps the busy status for pages 3327 * consistant. 3328 */ 3329void 3330vfs_unbusy_pages(struct buf * bp) 3331{ 3332 int i; 3333 3334 GIANT_REQUIRED; 3335 3336 runningbufwakeup(bp); 3337 if (bp->b_flags & B_VMIO) { 3338 vm_object_t obj; 3339 3340 obj = bp->b_object; 3341 vm_page_lock_queues(); 3342 for (i = 0; i < bp->b_npages; i++) { 3343 vm_page_t m = bp->b_pages[i]; 3344 3345 if (m == bogus_page) { 3346 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3347 if (!m) { 3348 panic("vfs_unbusy_pages: page missing\n"); 3349 } 3350 bp->b_pages[i] = m; 3351 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3352 } 3353 vm_object_pip_subtract(obj, 1); 3354 vm_page_flag_clear(m, PG_ZERO); 3355 vm_page_io_finish(m); 3356 } 3357 vm_page_unlock_queues(); 3358 vm_object_pip_wakeupn(obj, 0); 3359 } 3360} 3361 3362/* 3363 * vfs_page_set_valid: 3364 * 3365 * Set the valid bits in a page based on the supplied offset. The 3366 * range is restricted to the buffer's size. 3367 * 3368 * This routine is typically called after a read completes. 3369 */ 3370static void 3371vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3372{ 3373 vm_ooffset_t soff, eoff; 3374 3375 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3376 /* 3377 * Start and end offsets in buffer. eoff - soff may not cross a 3378 * page boundry or cross the end of the buffer. The end of the 3379 * buffer, in this case, is our file EOF, not the allocation size 3380 * of the buffer. 3381 */ 3382 soff = off; 3383 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3384 if (eoff > bp->b_offset + bp->b_bcount) 3385 eoff = bp->b_offset + bp->b_bcount; 3386 3387 /* 3388 * Set valid range. This is typically the entire buffer and thus the 3389 * entire page. 3390 */ 3391 if (eoff > soff) { 3392 vm_page_set_validclean( 3393 m, 3394 (vm_offset_t) (soff & PAGE_MASK), 3395 (vm_offset_t) (eoff - soff) 3396 ); 3397 } 3398} 3399 3400/* 3401 * This routine is called before a device strategy routine. 3402 * It is used to tell the VM system that paging I/O is in 3403 * progress, and treat the pages associated with the buffer 3404 * almost as being PG_BUSY. Also the object paging_in_progress 3405 * flag is handled to make sure that the object doesn't become 3406 * inconsistant. 3407 * 3408 * Since I/O has not been initiated yet, certain buffer flags 3409 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3410 * and should be ignored. 3411 */ 3412void 3413vfs_busy_pages(struct buf * bp, int clear_modify) 3414{ 3415 int i, bogus; 3416 3417 if (bp->b_flags & B_VMIO) { 3418 vm_object_t obj; 3419 vm_ooffset_t foff; 3420 3421 obj = bp->b_object; 3422 foff = bp->b_offset; 3423 KASSERT(bp->b_offset != NOOFFSET, 3424 ("vfs_busy_pages: no buffer offset")); 3425 vfs_setdirty(bp); 3426retry: 3427 vm_page_lock_queues(); 3428 for (i = 0; i < bp->b_npages; i++) { 3429 vm_page_t m = bp->b_pages[i]; 3430 3431 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3432 goto retry; 3433 } 3434 bogus = 0; 3435 for (i = 0; i < bp->b_npages; i++) { 3436 vm_page_t m = bp->b_pages[i]; 3437 3438 vm_page_flag_clear(m, PG_ZERO); 3439 if ((bp->b_flags & B_CLUSTER) == 0) { 3440 vm_object_pip_add(obj, 1); 3441 vm_page_io_start(m); 3442 } 3443 /* 3444 * When readying a buffer for a read ( i.e 3445 * clear_modify == 0 ), it is important to do 3446 * bogus_page replacement for valid pages in 3447 * partially instantiated buffers. Partially 3448 * instantiated buffers can, in turn, occur when 3449 * reconstituting a buffer from its VM backing store 3450 * base. We only have to do this if B_CACHE is 3451 * clear ( which causes the I/O to occur in the 3452 * first place ). The replacement prevents the read 3453 * I/O from overwriting potentially dirty VM-backed 3454 * pages. XXX bogus page replacement is, uh, bogus. 3455 * It may not work properly with small-block devices. 3456 * We need to find a better way. 3457 */ 3458 pmap_remove_all(m); 3459 if (clear_modify) 3460 vfs_page_set_valid(bp, foff, i, m); 3461 else if (m->valid == VM_PAGE_BITS_ALL && 3462 (bp->b_flags & B_CACHE) == 0) { 3463 bp->b_pages[i] = bogus_page; 3464 bogus++; 3465 } 3466 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3467 } 3468 vm_page_unlock_queues(); 3469 if (bogus) 3470 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3471 } 3472} 3473 3474/* 3475 * Tell the VM system that the pages associated with this buffer 3476 * are clean. This is used for delayed writes where the data is 3477 * going to go to disk eventually without additional VM intevention. 3478 * 3479 * Note that while we only really need to clean through to b_bcount, we 3480 * just go ahead and clean through to b_bufsize. 3481 */ 3482static void 3483vfs_clean_pages(struct buf * bp) 3484{ 3485 int i; 3486 3487 GIANT_REQUIRED; 3488 3489 if (bp->b_flags & B_VMIO) { 3490 vm_ooffset_t foff; 3491 3492 foff = bp->b_offset; 3493 KASSERT(bp->b_offset != NOOFFSET, 3494 ("vfs_clean_pages: no buffer offset")); 3495 vm_page_lock_queues(); 3496 for (i = 0; i < bp->b_npages; i++) { 3497 vm_page_t m = bp->b_pages[i]; 3498 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3499 vm_ooffset_t eoff = noff; 3500 3501 if (eoff > bp->b_offset + bp->b_bufsize) 3502 eoff = bp->b_offset + bp->b_bufsize; 3503 vfs_page_set_valid(bp, foff, i, m); 3504 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3505 foff = noff; 3506 } 3507 vm_page_unlock_queues(); 3508 } 3509} 3510 3511/* 3512 * vfs_bio_set_validclean: 3513 * 3514 * Set the range within the buffer to valid and clean. The range is 3515 * relative to the beginning of the buffer, b_offset. Note that b_offset 3516 * itself may be offset from the beginning of the first page. 3517 * 3518 */ 3519 3520void 3521vfs_bio_set_validclean(struct buf *bp, int base, int size) 3522{ 3523 if (bp->b_flags & B_VMIO) { 3524 int i; 3525 int n; 3526 3527 /* 3528 * Fixup base to be relative to beginning of first page. 3529 * Set initial n to be the maximum number of bytes in the 3530 * first page that can be validated. 3531 */ 3532 3533 base += (bp->b_offset & PAGE_MASK); 3534 n = PAGE_SIZE - (base & PAGE_MASK); 3535 3536 vm_page_lock_queues(); 3537 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3538 vm_page_t m = bp->b_pages[i]; 3539 3540 if (n > size) 3541 n = size; 3542 3543 vm_page_set_validclean(m, base & PAGE_MASK, n); 3544 base += n; 3545 size -= n; 3546 n = PAGE_SIZE; 3547 } 3548 vm_page_unlock_queues(); 3549 } 3550} 3551 3552/* 3553 * vfs_bio_clrbuf: 3554 * 3555 * clear a buffer. This routine essentially fakes an I/O, so we need 3556 * to clear BIO_ERROR and B_INVAL. 3557 * 3558 * Note that while we only theoretically need to clear through b_bcount, 3559 * we go ahead and clear through b_bufsize. 3560 */ 3561 3562void 3563vfs_bio_clrbuf(struct buf *bp) 3564{ 3565 int i, mask = 0; 3566 caddr_t sa, ea; 3567 3568 GIANT_REQUIRED; 3569 3570 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3571 bp->b_flags &= ~B_INVAL; 3572 bp->b_ioflags &= ~BIO_ERROR; 3573 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3574 (bp->b_offset & PAGE_MASK) == 0) { 3575 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3576 if ((bp->b_pages[0]->valid & mask) == mask) { 3577 bp->b_resid = 0; 3578 return; 3579 } 3580 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3581 ((bp->b_pages[0]->valid & mask) == 0)) { 3582 bzero(bp->b_data, bp->b_bufsize); 3583 bp->b_pages[0]->valid |= mask; 3584 bp->b_resid = 0; 3585 return; 3586 } 3587 } 3588 ea = sa = bp->b_data; 3589 for(i=0;i<bp->b_npages;i++,sa=ea) { 3590 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3591 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3592 ea = (caddr_t)(vm_offset_t)ulmin( 3593 (u_long)(vm_offset_t)ea, 3594 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3595 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3596 if ((bp->b_pages[i]->valid & mask) == mask) 3597 continue; 3598 if ((bp->b_pages[i]->valid & mask) == 0) { 3599 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3600 bzero(sa, ea - sa); 3601 } 3602 } else { 3603 for (; sa < ea; sa += DEV_BSIZE, j++) { 3604 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3605 (bp->b_pages[i]->valid & (1<<j)) == 0) 3606 bzero(sa, DEV_BSIZE); 3607 } 3608 } 3609 bp->b_pages[i]->valid |= mask; 3610 vm_page_lock_queues(); 3611 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3612 vm_page_unlock_queues(); 3613 } 3614 bp->b_resid = 0; 3615 } else { 3616 clrbuf(bp); 3617 } 3618} 3619 3620/* 3621 * vm_hold_load_pages and vm_hold_free_pages get pages into 3622 * a buffers address space. The pages are anonymous and are 3623 * not associated with a file object. 3624 */ 3625static void 3626vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3627{ 3628 vm_offset_t pg; 3629 vm_page_t p; 3630 int index; 3631 3632 GIANT_REQUIRED; 3633 3634 to = round_page(to); 3635 from = round_page(from); 3636 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3637 3638 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3639tryagain: 3640 /* 3641 * note: must allocate system pages since blocking here 3642 * could intefere with paging I/O, no matter which 3643 * process we are. 3644 */ 3645 vm_object_lock(kernel_object); 3646 p = vm_page_alloc(kernel_object, 3647 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3648 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3649 vm_object_unlock(kernel_object); 3650 if (!p) { 3651 atomic_add_int(&vm_pageout_deficit, 3652 (to - pg) >> PAGE_SHIFT); 3653 VM_WAIT; 3654 goto tryagain; 3655 } 3656 vm_page_lock_queues(); 3657 p->valid = VM_PAGE_BITS_ALL; 3658 vm_page_unlock_queues(); 3659 pmap_qenter(pg, &p, 1); 3660 bp->b_pages[index] = p; 3661 vm_page_lock_queues(); 3662 vm_page_wakeup(p); 3663 vm_page_unlock_queues(); 3664 } 3665 bp->b_npages = index; 3666} 3667 3668/* Return pages associated with this buf to the vm system */ 3669static void 3670vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3671{ 3672 vm_offset_t pg; 3673 vm_page_t p; 3674 int index, newnpages; 3675 3676 GIANT_REQUIRED; 3677 3678 from = round_page(from); 3679 to = round_page(to); 3680 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3681 3682 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3683 p = bp->b_pages[index]; 3684 if (p && (index < bp->b_npages)) { 3685 if (p->busy) { 3686 printf( 3687 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3688 (intmax_t)bp->b_blkno, 3689 (intmax_t)bp->b_lblkno); 3690 } 3691 bp->b_pages[index] = NULL; 3692 pmap_qremove(pg, 1); 3693 vm_page_lock_queues(); 3694 vm_page_busy(p); 3695 vm_page_unwire(p, 0); 3696 vm_page_free(p); 3697 vm_page_unlock_queues(); 3698 } 3699 } 3700 bp->b_npages = newnpages; 3701} 3702 3703/* 3704 * Map an IO request into kernel virtual address space. 3705 * 3706 * All requests are (re)mapped into kernel VA space. 3707 * Notice that we use b_bufsize for the size of the buffer 3708 * to be mapped. b_bcount might be modified by the driver. 3709 * 3710 * Note that even if the caller determines that the address space should 3711 * be valid, a race or a smaller-file mapped into a larger space may 3712 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3713 * check the return value. 3714 */ 3715int 3716vmapbuf(struct buf *bp) 3717{ 3718 caddr_t addr, kva; 3719 vm_offset_t pa; 3720 int pidx, i; 3721 struct vm_page *m; 3722 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3723 3724 GIANT_REQUIRED; 3725 3726 if ((bp->b_flags & B_PHYS) == 0) 3727 panic("vmapbuf"); 3728 3729 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3730 addr < bp->b_data + bp->b_bufsize; 3731 addr += PAGE_SIZE, pidx++) { 3732 /* 3733 * Do the vm_fault if needed; do the copy-on-write thing 3734 * when reading stuff off device into memory. 3735 * 3736 * NOTE! Must use pmap_extract() because addr may be in 3737 * the userland address space, and kextract is only guarenteed 3738 * to work for the kernland address space (see: sparc64 port). 3739 */ 3740retry: 3741 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data, 3742 (bp->b_iocmd == BIO_READ) ? 3743 (VM_PROT_READ|VM_PROT_WRITE) : VM_PROT_READ); 3744 if (i < 0) { 3745 printf("vmapbuf: warning, bad user address during I/O\n"); 3746 vm_page_lock_queues(); 3747 for (i = 0; i < pidx; ++i) { 3748 vm_page_unhold(bp->b_pages[i]); 3749 bp->b_pages[i] = NULL; 3750 } 3751 vm_page_unlock_queues(); 3752 return(-1); 3753 } 3754 pa = trunc_page(pmap_extract(pmap, (vm_offset_t) addr)); 3755 if (pa == 0) { 3756 printf("vmapbuf: warning, race against user address during I/O"); 3757 goto retry; 3758 } 3759 m = PHYS_TO_VM_PAGE(pa); 3760 vm_page_lock_queues(); 3761 vm_page_hold(m); 3762 vm_page_unlock_queues(); 3763 bp->b_pages[pidx] = m; 3764 } 3765 if (pidx > btoc(MAXPHYS)) 3766 panic("vmapbuf: mapped more than MAXPHYS"); 3767 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3768 3769 kva = bp->b_saveaddr; 3770 bp->b_npages = pidx; 3771 bp->b_saveaddr = bp->b_data; 3772 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3773 return(0); 3774} 3775 3776/* 3777 * Free the io map PTEs associated with this IO operation. 3778 * We also invalidate the TLB entries and restore the original b_addr. 3779 */ 3780void 3781vunmapbuf(struct buf *bp) 3782{ 3783 int pidx; 3784 int npages; 3785 3786 GIANT_REQUIRED; 3787 3788 if ((bp->b_flags & B_PHYS) == 0) 3789 panic("vunmapbuf"); 3790 3791 npages = bp->b_npages; 3792 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3793 npages); 3794 vm_page_lock_queues(); 3795 for (pidx = 0; pidx < npages; pidx++) 3796 vm_page_unhold(bp->b_pages[pidx]); 3797 vm_page_unlock_queues(); 3798 3799 bp->b_data = bp->b_saveaddr; 3800} 3801 3802#include "opt_ddb.h" 3803#ifdef DDB 3804#include <ddb/ddb.h> 3805 3806/* DDB command to show buffer data */ 3807DB_SHOW_COMMAND(buffer, db_show_buffer) 3808{ 3809 /* get args */ 3810 struct buf *bp = (struct buf *)addr; 3811 3812 if (!have_addr) { 3813 db_printf("usage: show buffer <addr>\n"); 3814 return; 3815 } 3816 3817 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3818 db_printf( 3819 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3820 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd, b_pblkno = %jd\n", 3821 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3822 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3823 (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno); 3824 if (bp->b_npages) { 3825 int i; 3826 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3827 for (i = 0; i < bp->b_npages; i++) { 3828 vm_page_t m; 3829 m = bp->b_pages[i]; 3830 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3831 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3832 if ((i + 1) < bp->b_npages) 3833 db_printf(","); 3834 } 3835 db_printf("\n"); 3836 } 3837} 3838#endif /* DDB */ 3839