vfs_bio.c revision 116604
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 15/* 16 * this file contains a new buffer I/O scheme implementing a coherent 17 * VM object and buffer cache scheme. Pains have been taken to make 18 * sure that the performance degradation associated with schemes such 19 * as this is not realized. 20 * 21 * Author: John S. Dyson 22 * Significant help during the development and debugging phases 23 * had been provided by David Greenman, also of the FreeBSD core team. 24 * 25 * see man buf(9) for more info. 26 */ 27 28#include <sys/cdefs.h> 29__FBSDID("$FreeBSD: head/sys/kern/vfs_bio.c 116604 2003-06-20 08:26:38Z phk $"); 30 31#include <sys/param.h> 32#include <sys/systm.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#include "opt_directio.h" 57#include "opt_swap.h" 58 59static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 60 61struct bio_ops bioops; /* I/O operation notification */ 62 63struct buf_ops buf_ops_bio = { 64 "buf_ops_bio", 65 bwrite 66}; 67 68/* 69 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 70 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 71 */ 72struct buf *buf; /* buffer header pool */ 73 74static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 75 vm_offset_t to); 76static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 77 vm_offset_t to); 78static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 79 int pageno, vm_page_t m); 80static void vfs_clean_pages(struct buf * bp); 81static void vfs_setdirty(struct buf *bp); 82static void vfs_vmio_release(struct buf *bp); 83static void vfs_backgroundwritedone(struct buf *bp); 84static int vfs_bio_clcheck(struct vnode *vp, int size, 85 daddr_t lblkno, daddr_t blkno); 86static int flushbufqueues(int flushdeps); 87static void buf_daemon(void); 88void bremfreel(struct buf * bp); 89 90int vmiodirenable = TRUE; 91SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 92 "Use the VM system for directory writes"); 93int runningbufspace; 94SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 95 "Amount of presently outstanding async buffer io"); 96static int bufspace; 97SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 98 "KVA memory used for bufs"); 99static int maxbufspace; 100SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 101 "Maximum allowed value of bufspace (including buf_daemon)"); 102static int bufmallocspace; 103SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 104 "Amount of malloced memory for buffers"); 105static int maxbufmallocspace; 106SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 107 "Maximum amount of malloced memory for buffers"); 108static int lobufspace; 109SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 110 "Minimum amount of buffers we want to have"); 111static int hibufspace; 112SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 113 "Maximum allowed value of bufspace (excluding buf_daemon)"); 114static int bufreusecnt; 115SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 116 "Number of times we have reused a buffer"); 117static int buffreekvacnt; 118SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 119 "Number of times we have freed the KVA space from some buffer"); 120static int bufdefragcnt; 121SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 122 "Number of times we have had to repeat buffer allocation to defragment"); 123static int lorunningspace; 124SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 125 "Minimum preferred space used for in-progress I/O"); 126static int hirunningspace; 127SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 128 "Maximum amount of space to use for in-progress I/O"); 129static int dirtybufferflushes; 130SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 131 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 132static int altbufferflushes; 133SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 134 0, "Number of fsync flushes to limit dirty buffers"); 135static int recursiveflushes; 136SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 137 0, "Number of flushes skipped due to being recursive"); 138static int numdirtybuffers; 139SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 140 "Number of buffers that are dirty (has unwritten changes) at the moment"); 141static int lodirtybuffers; 142SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 143 "How many buffers we want to have free before bufdaemon can sleep"); 144static int hidirtybuffers; 145SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 146 "When the number of dirty buffers is considered severe"); 147static int dirtybufthresh; 148SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 149 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 150static int numfreebuffers; 151SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 152 "Number of free buffers"); 153static int lofreebuffers; 154SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 155 "XXX Unused"); 156static int hifreebuffers; 157SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 158 "XXX Complicatedly unused"); 159static int getnewbufcalls; 160SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 161 "Number of calls to getnewbuf"); 162static int getnewbufrestarts; 163SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 164 "Number of times getnewbuf has had to restart a buffer aquisition"); 165static int dobkgrdwrite = 1; 166SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 167 "Do background writes (honoring the BX_BKGRDWRITE flag)?"); 168 169/* 170 * Wakeup point for bufdaemon, as well as indicator of whether it is already 171 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 172 * is idling. 173 */ 174static int bd_request; 175 176/* 177 * This lock synchronizes access to bd_request. 178 */ 179static struct mtx bdlock; 180 181/* 182 * bogus page -- for I/O to/from partially complete buffers 183 * this is a temporary solution to the problem, but it is not 184 * really that bad. it would be better to split the buffer 185 * for input in the case of buffers partially already in memory, 186 * but the code is intricate enough already. 187 */ 188vm_page_t bogus_page; 189 190/* 191 * Synchronization (sleep/wakeup) variable for active buffer space requests. 192 * Set when wait starts, cleared prior to wakeup(). 193 * Used in runningbufwakeup() and waitrunningbufspace(). 194 */ 195static int runningbufreq; 196 197/* 198 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 199 * waitrunningbufspace(). 200 */ 201static struct mtx rbreqlock; 202 203/* 204 * Synchronization (sleep/wakeup) variable for buffer requests. 205 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 206 * by and/or. 207 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 208 * getnewbuf(), and getblk(). 209 */ 210static int needsbuffer; 211 212/* 213 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 214 */ 215static struct mtx nblock; 216 217/* 218 * Lock that protects against bwait()/bdone()/B_DONE races. 219 */ 220 221static struct mtx bdonelock; 222 223/* 224 * Definitions for the buffer free lists. 225 */ 226#define BUFFER_QUEUES 6 /* number of free buffer queues */ 227 228#define QUEUE_NONE 0 /* on no queue */ 229#define QUEUE_LOCKED 1 /* locked buffers */ 230#define QUEUE_CLEAN 2 /* non-B_DELWRI buffers */ 231#define QUEUE_DIRTY 3 /* B_DELWRI buffers */ 232#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 233#define QUEUE_EMPTY 5 /* empty buffer headers */ 234 235/* Queues for free buffers with various properties */ 236static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 237 238/* Lock for the bufqueues */ 239static struct mtx bqlock; 240 241/* 242 * Single global constant for BUF_WMESG, to avoid getting multiple references. 243 * buf_wmesg is referred from macros. 244 */ 245const char *buf_wmesg = BUF_WMESG; 246 247#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 248#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 249#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 250#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 251 252#ifdef DIRECTIO 253extern void ffs_rawread_setup(void); 254#endif /* DIRECTIO */ 255/* 256 * numdirtywakeup: 257 * 258 * If someone is blocked due to there being too many dirty buffers, 259 * and numdirtybuffers is now reasonable, wake them up. 260 */ 261 262static __inline void 263numdirtywakeup(int level) 264{ 265 if (numdirtybuffers <= level) { 266 mtx_lock(&nblock); 267 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 268 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 269 wakeup(&needsbuffer); 270 } 271 mtx_unlock(&nblock); 272 } 273} 274 275/* 276 * bufspacewakeup: 277 * 278 * Called when buffer space is potentially available for recovery. 279 * getnewbuf() will block on this flag when it is unable to free 280 * sufficient buffer space. Buffer space becomes recoverable when 281 * bp's get placed back in the queues. 282 */ 283 284static __inline void 285bufspacewakeup(void) 286{ 287 /* 288 * If someone is waiting for BUF space, wake them up. Even 289 * though we haven't freed the kva space yet, the waiting 290 * process will be able to now. 291 */ 292 mtx_lock(&nblock); 293 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 294 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 295 wakeup(&needsbuffer); 296 } 297 mtx_unlock(&nblock); 298} 299 300/* 301 * runningbufwakeup() - in-progress I/O accounting. 302 * 303 */ 304static __inline void 305runningbufwakeup(struct buf *bp) 306{ 307 if (bp->b_runningbufspace) { 308 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 309 bp->b_runningbufspace = 0; 310 mtx_lock(&rbreqlock); 311 if (runningbufreq && runningbufspace <= lorunningspace) { 312 runningbufreq = 0; 313 wakeup(&runningbufreq); 314 } 315 mtx_unlock(&rbreqlock); 316 } 317} 318 319/* 320 * bufcountwakeup: 321 * 322 * Called when a buffer has been added to one of the free queues to 323 * account for the buffer and to wakeup anyone waiting for free buffers. 324 * This typically occurs when large amounts of metadata are being handled 325 * by the buffer cache ( else buffer space runs out first, usually ). 326 */ 327 328static __inline void 329bufcountwakeup(void) 330{ 331 atomic_add_int(&numfreebuffers, 1); 332 mtx_lock(&nblock); 333 if (needsbuffer) { 334 needsbuffer &= ~VFS_BIO_NEED_ANY; 335 if (numfreebuffers >= hifreebuffers) 336 needsbuffer &= ~VFS_BIO_NEED_FREE; 337 wakeup(&needsbuffer); 338 } 339 mtx_unlock(&nblock); 340} 341 342/* 343 * waitrunningbufspace() 344 * 345 * runningbufspace is a measure of the amount of I/O currently 346 * running. This routine is used in async-write situations to 347 * prevent creating huge backups of pending writes to a device. 348 * Only asynchronous writes are governed by this function. 349 * 350 * Reads will adjust runningbufspace, but will not block based on it. 351 * The read load has a side effect of reducing the allowed write load. 352 * 353 * This does NOT turn an async write into a sync write. It waits 354 * for earlier writes to complete and generally returns before the 355 * caller's write has reached the device. 356 */ 357static __inline void 358waitrunningbufspace(void) 359{ 360 mtx_lock(&rbreqlock); 361 while (runningbufspace > hirunningspace) { 362 ++runningbufreq; 363 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 364 } 365 mtx_unlock(&rbreqlock); 366} 367 368 369/* 370 * vfs_buf_test_cache: 371 * 372 * Called when a buffer is extended. This function clears the B_CACHE 373 * bit if the newly extended portion of the buffer does not contain 374 * valid data. 375 */ 376static __inline__ 377void 378vfs_buf_test_cache(struct buf *bp, 379 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 380 vm_page_t m) 381{ 382 GIANT_REQUIRED; 383 384 if (bp->b_flags & B_CACHE) { 385 int base = (foff + off) & PAGE_MASK; 386 if (vm_page_is_valid(m, base, size) == 0) 387 bp->b_flags &= ~B_CACHE; 388 } 389} 390 391/* Wake up the buffer deamon if necessary */ 392static __inline__ 393void 394bd_wakeup(int dirtybuflevel) 395{ 396 mtx_lock(&bdlock); 397 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 398 bd_request = 1; 399 wakeup(&bd_request); 400 } 401 mtx_unlock(&bdlock); 402} 403 404/* 405 * bd_speedup - speedup the buffer cache flushing code 406 */ 407 408static __inline__ 409void 410bd_speedup(void) 411{ 412 bd_wakeup(1); 413} 414 415/* 416 * Calculating buffer cache scaling values and reserve space for buffer 417 * headers. This is called during low level kernel initialization and 418 * may be called more then once. We CANNOT write to the memory area 419 * being reserved at this time. 420 */ 421caddr_t 422kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 423{ 424 /* 425 * physmem_est is in pages. Convert it to kilobytes (assumes 426 * PAGE_SIZE is >= 1K) 427 */ 428 physmem_est = physmem_est * (PAGE_SIZE / 1024); 429 430 /* 431 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 432 * For the first 64MB of ram nominally allocate sufficient buffers to 433 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 434 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 435 * the buffer cache we limit the eventual kva reservation to 436 * maxbcache bytes. 437 * 438 * factor represents the 1/4 x ram conversion. 439 */ 440 if (nbuf == 0) { 441 int factor = 4 * BKVASIZE / 1024; 442 443 nbuf = 50; 444 if (physmem_est > 4096) 445 nbuf += min((physmem_est - 4096) / factor, 446 65536 / factor); 447 if (physmem_est > 65536) 448 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 449 450 if (maxbcache && nbuf > maxbcache / BKVASIZE) 451 nbuf = maxbcache / BKVASIZE; 452 } 453 454#if 0 455 /* 456 * Do not allow the buffer_map to be more then 1/2 the size of the 457 * kernel_map. 458 */ 459 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 460 (BKVASIZE * 2)) { 461 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 462 (BKVASIZE * 2); 463 printf("Warning: nbufs capped at %d\n", nbuf); 464 } 465#endif 466 467 /* 468 * swbufs are used as temporary holders for I/O, such as paging I/O. 469 * We have no less then 16 and no more then 256. 470 */ 471 nswbuf = max(min(nbuf/4, 256), 16); 472#ifdef NSWBUF_MIN 473 if (nswbuf < NSWBUF_MIN) 474 nswbuf = NSWBUF_MIN; 475#endif 476#ifdef DIRECTIO 477 ffs_rawread_setup(); 478#endif 479 480 /* 481 * Reserve space for the buffer cache buffers 482 */ 483 swbuf = (void *)v; 484 v = (caddr_t)(swbuf + nswbuf); 485 buf = (void *)v; 486 v = (caddr_t)(buf + nbuf); 487 488 return(v); 489} 490 491/* Initialize the buffer subsystem. Called before use of any buffers. */ 492void 493bufinit(void) 494{ 495 struct buf *bp; 496 vm_offset_t bogus_offset; 497 int i; 498 499 GIANT_REQUIRED; 500 501 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 502 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 503 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 504 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 505 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 506 507 /* next, make a null set of free lists */ 508 for (i = 0; i < BUFFER_QUEUES; i++) 509 TAILQ_INIT(&bufqueues[i]); 510 511 /* finally, initialize each buffer header and stick on empty q */ 512 for (i = 0; i < nbuf; i++) { 513 bp = &buf[i]; 514 bzero(bp, sizeof *bp); 515 bp->b_flags = B_INVAL; /* we're just an empty header */ 516 bp->b_dev = NODEV; 517 bp->b_rcred = NOCRED; 518 bp->b_wcred = NOCRED; 519 bp->b_qindex = QUEUE_EMPTY; 520 bp->b_vflags = 0; 521 bp->b_xflags = 0; 522 LIST_INIT(&bp->b_dep); 523 BUF_LOCKINIT(bp); 524 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 525 } 526 527 /* 528 * maxbufspace is the absolute maximum amount of buffer space we are 529 * allowed to reserve in KVM and in real terms. The absolute maximum 530 * is nominally used by buf_daemon. hibufspace is the nominal maximum 531 * used by most other processes. The differential is required to 532 * ensure that buf_daemon is able to run when other processes might 533 * be blocked waiting for buffer space. 534 * 535 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 536 * this may result in KVM fragmentation which is not handled optimally 537 * by the system. 538 */ 539 maxbufspace = nbuf * BKVASIZE; 540 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 541 lobufspace = hibufspace - MAXBSIZE; 542 543 lorunningspace = 512 * 1024; 544 hirunningspace = 1024 * 1024; 545 546/* 547 * Limit the amount of malloc memory since it is wired permanently into 548 * the kernel space. Even though this is accounted for in the buffer 549 * allocation, we don't want the malloced region to grow uncontrolled. 550 * The malloc scheme improves memory utilization significantly on average 551 * (small) directories. 552 */ 553 maxbufmallocspace = hibufspace / 20; 554 555/* 556 * Reduce the chance of a deadlock occuring by limiting the number 557 * of delayed-write dirty buffers we allow to stack up. 558 */ 559 hidirtybuffers = nbuf / 4 + 20; 560 dirtybufthresh = hidirtybuffers * 9 / 10; 561 numdirtybuffers = 0; 562/* 563 * To support extreme low-memory systems, make sure hidirtybuffers cannot 564 * eat up all available buffer space. This occurs when our minimum cannot 565 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 566 * BKVASIZE'd (8K) buffers. 567 */ 568 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 569 hidirtybuffers >>= 1; 570 } 571 lodirtybuffers = hidirtybuffers / 2; 572 573/* 574 * Try to keep the number of free buffers in the specified range, 575 * and give special processes (e.g. like buf_daemon) access to an 576 * emergency reserve. 577 */ 578 lofreebuffers = nbuf / 18 + 5; 579 hifreebuffers = 2 * lofreebuffers; 580 numfreebuffers = nbuf; 581 582/* 583 * Maximum number of async ops initiated per buf_daemon loop. This is 584 * somewhat of a hack at the moment, we really need to limit ourselves 585 * based on the number of bytes of I/O in-transit that were initiated 586 * from buf_daemon. 587 */ 588 589 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 590 VM_OBJECT_LOCK(kernel_object); 591 bogus_page = vm_page_alloc(kernel_object, 592 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 593 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 594 VM_OBJECT_UNLOCK(kernel_object); 595} 596 597/* 598 * bfreekva() - free the kva allocation for a buffer. 599 * 600 * Must be called at splbio() or higher as this is the only locking for 601 * buffer_map. 602 * 603 * Since this call frees up buffer space, we call bufspacewakeup(). 604 */ 605static void 606bfreekva(struct buf * bp) 607{ 608 GIANT_REQUIRED; 609 610 if (bp->b_kvasize) { 611 atomic_add_int(&buffreekvacnt, 1); 612 atomic_subtract_int(&bufspace, bp->b_kvasize); 613 vm_map_delete(buffer_map, 614 (vm_offset_t) bp->b_kvabase, 615 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 616 ); 617 bp->b_kvasize = 0; 618 bufspacewakeup(); 619 } 620} 621 622/* 623 * bremfree: 624 * 625 * Remove the buffer from the appropriate free list. 626 */ 627void 628bremfree(struct buf * bp) 629{ 630 mtx_lock(&bqlock); 631 bremfreel(bp); 632 mtx_unlock(&bqlock); 633} 634 635void 636bremfreel(struct buf * bp) 637{ 638 int s = splbio(); 639 int old_qindex = bp->b_qindex; 640 641 GIANT_REQUIRED; 642 643 if (bp->b_qindex != QUEUE_NONE) { 644 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 645 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 646 bp->b_qindex = QUEUE_NONE; 647 } else { 648 if (BUF_REFCNT(bp) <= 1) 649 panic("bremfree: removing a buffer not on a queue"); 650 } 651 652 /* 653 * Fixup numfreebuffers count. If the buffer is invalid or not 654 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 655 * the buffer was free and we must decrement numfreebuffers. 656 */ 657 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 658 switch(old_qindex) { 659 case QUEUE_DIRTY: 660 case QUEUE_CLEAN: 661 case QUEUE_EMPTY: 662 case QUEUE_EMPTYKVA: 663 atomic_subtract_int(&numfreebuffers, 1); 664 break; 665 default: 666 break; 667 } 668 } 669 splx(s); 670} 671 672 673/* 674 * Get a buffer with the specified data. Look in the cache first. We 675 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 676 * is set, the buffer is valid and we do not have to do anything ( see 677 * getblk() ). This is really just a special case of breadn(). 678 */ 679int 680bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 681 struct buf ** bpp) 682{ 683 684 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 685} 686 687/* 688 * Operates like bread, but also starts asynchronous I/O on 689 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 690 * to initiating I/O . If B_CACHE is set, the buffer is valid 691 * and we do not have to do anything. 692 */ 693int 694breadn(struct vnode * vp, daddr_t blkno, int size, 695 daddr_t * rablkno, int *rabsize, 696 int cnt, struct ucred * cred, struct buf ** bpp) 697{ 698 struct buf *bp, *rabp; 699 int i; 700 int rv = 0, readwait = 0; 701 702 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 703 704 /* if not found in cache, do some I/O */ 705 if ((bp->b_flags & B_CACHE) == 0) { 706 if (curthread != PCPU_GET(idlethread)) 707 curthread->td_proc->p_stats->p_ru.ru_inblock++; 708 bp->b_iocmd = BIO_READ; 709 bp->b_flags &= ~B_INVAL; 710 bp->b_ioflags &= ~BIO_ERROR; 711 if (bp->b_rcred == NOCRED && cred != NOCRED) 712 bp->b_rcred = crhold(cred); 713 vfs_busy_pages(bp, 0); 714 if (vp->v_type == VCHR) 715 VOP_SPECSTRATEGY(vp, bp); 716 else 717 VOP_STRATEGY(vp, bp); 718 ++readwait; 719 } 720 721 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 722 if (inmem(vp, *rablkno)) 723 continue; 724 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 725 726 if ((rabp->b_flags & B_CACHE) == 0) { 727 if (curthread != PCPU_GET(idlethread)) 728 curthread->td_proc->p_stats->p_ru.ru_inblock++; 729 rabp->b_flags |= B_ASYNC; 730 rabp->b_flags &= ~B_INVAL; 731 rabp->b_ioflags &= ~BIO_ERROR; 732 rabp->b_iocmd = BIO_READ; 733 if (rabp->b_rcred == NOCRED && cred != NOCRED) 734 rabp->b_rcred = crhold(cred); 735 vfs_busy_pages(rabp, 0); 736 BUF_KERNPROC(rabp); 737 if (vp->v_type == VCHR) 738 VOP_SPECSTRATEGY(vp, rabp); 739 else 740 VOP_STRATEGY(vp, rabp); 741 } else { 742 brelse(rabp); 743 } 744 } 745 746 if (readwait) { 747 rv = bufwait(bp); 748 } 749 return (rv); 750} 751 752/* 753 * Write, release buffer on completion. (Done by iodone 754 * if async). Do not bother writing anything if the buffer 755 * is invalid. 756 * 757 * Note that we set B_CACHE here, indicating that buffer is 758 * fully valid and thus cacheable. This is true even of NFS 759 * now so we set it generally. This could be set either here 760 * or in biodone() since the I/O is synchronous. We put it 761 * here. 762 */ 763 764int 765bwrite(struct buf * bp) 766{ 767 int oldflags, s; 768 struct buf *newbp; 769 770 if (bp->b_flags & B_INVAL) { 771 brelse(bp); 772 return (0); 773 } 774 775 oldflags = bp->b_flags; 776 777 if (BUF_REFCNT(bp) == 0) 778 panic("bwrite: buffer is not busy???"); 779 s = splbio(); 780 /* 781 * If a background write is already in progress, delay 782 * writing this block if it is asynchronous. Otherwise 783 * wait for the background write to complete. 784 */ 785 if (bp->b_xflags & BX_BKGRDINPROG) { 786 if (bp->b_flags & B_ASYNC) { 787 splx(s); 788 bdwrite(bp); 789 return (0); 790 } 791 bp->b_xflags |= BX_BKGRDWAIT; 792 tsleep(&bp->b_xflags, PRIBIO, "bwrbg", 0); 793 if (bp->b_xflags & BX_BKGRDINPROG) 794 panic("bwrite: still writing"); 795 } 796 797 /* Mark the buffer clean */ 798 bundirty(bp); 799 800 /* 801 * If this buffer is marked for background writing and we 802 * do not have to wait for it, make a copy and write the 803 * copy so as to leave this buffer ready for further use. 804 * 805 * This optimization eats a lot of memory. If we have a page 806 * or buffer shortfall we can't do it. 807 */ 808 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 809 (bp->b_flags & B_ASYNC) && 810 !vm_page_count_severe() && 811 !buf_dirty_count_severe()) { 812 if (bp->b_iodone != NULL) { 813 printf("bp->b_iodone = %p\n", bp->b_iodone); 814 panic("bwrite: need chained iodone"); 815 } 816 817 /* get a new block */ 818 newbp = geteblk(bp->b_bufsize); 819 820 /* 821 * set it to be identical to the old block. We have to 822 * set b_lblkno and BKGRDMARKER before calling bgetvp() 823 * to avoid confusing the splay tree and gbincore(). 824 */ 825 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 826 newbp->b_lblkno = bp->b_lblkno; 827 newbp->b_xflags |= BX_BKGRDMARKER; 828 /* XXX The BX_ flags need to be protected as well */ 829 VI_LOCK(bp->b_vp); 830 bgetvp(bp->b_vp, newbp); 831 VI_UNLOCK(bp->b_vp); 832 newbp->b_blkno = bp->b_blkno; 833 newbp->b_offset = bp->b_offset; 834 newbp->b_iodone = vfs_backgroundwritedone; 835 newbp->b_flags |= B_ASYNC; 836 newbp->b_flags &= ~B_INVAL; 837 838 /* move over the dependencies */ 839 if (LIST_FIRST(&bp->b_dep) != NULL) 840 buf_movedeps(bp, newbp); 841 842 /* 843 * Initiate write on the copy, release the original to 844 * the B_LOCKED queue so that it cannot go away until 845 * the background write completes. If not locked it could go 846 * away and then be reconstituted while it was being written. 847 * If the reconstituted buffer were written, we could end up 848 * with two background copies being written at the same time. 849 */ 850 bp->b_xflags |= BX_BKGRDINPROG; 851 bp->b_flags |= B_LOCKED; 852 bqrelse(bp); 853 bp = newbp; 854 } 855 856 bp->b_flags &= ~B_DONE; 857 bp->b_ioflags &= ~BIO_ERROR; 858 bp->b_flags |= B_WRITEINPROG | B_CACHE; 859 bp->b_iocmd = BIO_WRITE; 860 861 VI_LOCK(bp->b_vp); 862 bp->b_vp->v_numoutput++; 863 VI_UNLOCK(bp->b_vp); 864 vfs_busy_pages(bp, 1); 865 866 /* 867 * Normal bwrites pipeline writes 868 */ 869 bp->b_runningbufspace = bp->b_bufsize; 870 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 871 872 if (curthread != PCPU_GET(idlethread)) 873 curthread->td_proc->p_stats->p_ru.ru_oublock++; 874 splx(s); 875 if (oldflags & B_ASYNC) 876 BUF_KERNPROC(bp); 877 if (bp->b_vp->v_type == VCHR) 878 VOP_SPECSTRATEGY(bp->b_vp, bp); 879 else 880 VOP_STRATEGY(bp->b_vp, bp); 881 882 if ((oldflags & B_ASYNC) == 0) { 883 int rtval = bufwait(bp); 884 brelse(bp); 885 return (rtval); 886 } else { 887 /* 888 * don't allow the async write to saturate the I/O 889 * system. We will not deadlock here because 890 * we are blocking waiting for I/O that is already in-progress 891 * to complete. 892 */ 893 waitrunningbufspace(); 894 } 895 896 return (0); 897} 898 899/* 900 * Complete a background write started from bwrite. 901 */ 902static void 903vfs_backgroundwritedone(bp) 904 struct buf *bp; 905{ 906 struct buf *origbp; 907 908 /* 909 * Find the original buffer that we are writing. 910 */ 911 VI_LOCK(bp->b_vp); 912 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 913 panic("backgroundwritedone: lost buffer"); 914 VI_UNLOCK(bp->b_vp); 915 /* 916 * Process dependencies then return any unfinished ones. 917 */ 918 if (LIST_FIRST(&bp->b_dep) != NULL) 919 buf_complete(bp); 920 if (LIST_FIRST(&bp->b_dep) != NULL) 921 buf_movedeps(bp, origbp); 922 923 /* XXX Find out if origbp can disappear or get inconsistent */ 924 /* 925 * Clear the BX_BKGRDINPROG flag in the original buffer 926 * and awaken it if it is waiting for the write to complete. 927 * If BX_BKGRDINPROG is not set in the original buffer it must 928 * have been released and re-instantiated - which is not legal. 929 */ 930 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), 931 ("backgroundwritedone: lost buffer2")); 932 origbp->b_xflags &= ~BX_BKGRDINPROG; 933 if (origbp->b_xflags & BX_BKGRDWAIT) { 934 origbp->b_xflags &= ~BX_BKGRDWAIT; 935 wakeup(&origbp->b_xflags); 936 } 937 /* 938 * Clear the B_LOCKED flag and remove it from the locked 939 * queue if it currently resides there. 940 */ 941 origbp->b_flags &= ~B_LOCKED; 942 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) { 943 bremfree(origbp); 944 bqrelse(origbp); 945 } 946 /* 947 * This buffer is marked B_NOCACHE, so when it is released 948 * by biodone, it will be tossed. We mark it with BIO_READ 949 * to avoid biodone doing a second vwakeup. 950 */ 951 bp->b_flags |= B_NOCACHE; 952 bp->b_iocmd = BIO_READ; 953 bp->b_flags &= ~(B_CACHE | B_DONE); 954 bp->b_iodone = 0; 955 bufdone(bp); 956} 957 958/* 959 * Delayed write. (Buffer is marked dirty). Do not bother writing 960 * anything if the buffer is marked invalid. 961 * 962 * Note that since the buffer must be completely valid, we can safely 963 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 964 * biodone() in order to prevent getblk from writing the buffer 965 * out synchronously. 966 */ 967void 968bdwrite(struct buf * bp) 969{ 970 struct thread *td = curthread; 971 struct vnode *vp; 972 struct buf *nbp; 973 974 GIANT_REQUIRED; 975 976 if (BUF_REFCNT(bp) == 0) 977 panic("bdwrite: buffer is not busy"); 978 979 if (bp->b_flags & B_INVAL) { 980 brelse(bp); 981 return; 982 } 983 984 /* 985 * If we have too many dirty buffers, don't create any more. 986 * If we are wildly over our limit, then force a complete 987 * cleanup. Otherwise, just keep the situation from getting 988 * out of control. Note that we have to avoid a recursive 989 * disaster and not try to clean up after our own cleanup! 990 */ 991 vp = bp->b_vp; 992 VI_LOCK(vp); 993 if (td->td_proc->p_flag & P_COWINPROGRESS) { 994 recursiveflushes++; 995 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) { 996 VI_UNLOCK(vp); 997 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 998 VI_LOCK(vp); 999 altbufferflushes++; 1000 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) { 1001 /* 1002 * Try to find a buffer to flush. 1003 */ 1004 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) { 1005 if ((nbp->b_xflags & BX_BKGRDINPROG) || 1006 buf_countdeps(nbp, 0) || 1007 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1008 continue; 1009 if (bp == nbp) 1010 panic("bdwrite: found ourselves"); 1011 VI_UNLOCK(vp); 1012 if (nbp->b_flags & B_CLUSTEROK) { 1013 vfs_bio_awrite(nbp); 1014 } else { 1015 bremfree(nbp); 1016 bawrite(nbp); 1017 } 1018 VI_LOCK(vp); 1019 dirtybufferflushes++; 1020 break; 1021 } 1022 } 1023 VI_UNLOCK(vp); 1024 1025 bdirty(bp); 1026 /* 1027 * Set B_CACHE, indicating that the buffer is fully valid. This is 1028 * true even of NFS now. 1029 */ 1030 bp->b_flags |= B_CACHE; 1031 1032 /* 1033 * This bmap keeps the system from needing to do the bmap later, 1034 * perhaps when the system is attempting to do a sync. Since it 1035 * is likely that the indirect block -- or whatever other datastructure 1036 * that the filesystem needs is still in memory now, it is a good 1037 * thing to do this. Note also, that if the pageout daemon is 1038 * requesting a sync -- there might not be enough memory to do 1039 * the bmap then... So, this is important to do. 1040 */ 1041 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1042 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1043 } 1044 1045 /* 1046 * Set the *dirty* buffer range based upon the VM system dirty pages. 1047 */ 1048 vfs_setdirty(bp); 1049 1050 /* 1051 * We need to do this here to satisfy the vnode_pager and the 1052 * pageout daemon, so that it thinks that the pages have been 1053 * "cleaned". Note that since the pages are in a delayed write 1054 * buffer -- the VFS layer "will" see that the pages get written 1055 * out on the next sync, or perhaps the cluster will be completed. 1056 */ 1057 vfs_clean_pages(bp); 1058 bqrelse(bp); 1059 1060 /* 1061 * Wakeup the buffer flushing daemon if we have a lot of dirty 1062 * buffers (midpoint between our recovery point and our stall 1063 * point). 1064 */ 1065 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1066 1067 /* 1068 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1069 * due to the softdep code. 1070 */ 1071} 1072 1073/* 1074 * bdirty: 1075 * 1076 * Turn buffer into delayed write request. We must clear BIO_READ and 1077 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1078 * itself to properly update it in the dirty/clean lists. We mark it 1079 * B_DONE to ensure that any asynchronization of the buffer properly 1080 * clears B_DONE ( else a panic will occur later ). 1081 * 1082 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1083 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1084 * should only be called if the buffer is known-good. 1085 * 1086 * Since the buffer is not on a queue, we do not update the numfreebuffers 1087 * count. 1088 * 1089 * Must be called at splbio(). 1090 * The buffer must be on QUEUE_NONE. 1091 */ 1092void 1093bdirty(bp) 1094 struct buf *bp; 1095{ 1096 KASSERT(bp->b_qindex == QUEUE_NONE, 1097 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1098 bp->b_flags &= ~(B_RELBUF); 1099 bp->b_iocmd = BIO_WRITE; 1100 1101 if ((bp->b_flags & B_DELWRI) == 0) { 1102 bp->b_flags |= B_DONE | B_DELWRI; 1103 reassignbuf(bp, bp->b_vp); 1104 atomic_add_int(&numdirtybuffers, 1); 1105 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1106 } 1107} 1108 1109/* 1110 * bundirty: 1111 * 1112 * Clear B_DELWRI for buffer. 1113 * 1114 * Since the buffer is not on a queue, we do not update the numfreebuffers 1115 * count. 1116 * 1117 * Must be called at splbio(). 1118 * The buffer must be on QUEUE_NONE. 1119 */ 1120 1121void 1122bundirty(bp) 1123 struct buf *bp; 1124{ 1125 KASSERT(bp->b_qindex == QUEUE_NONE, 1126 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1127 1128 if (bp->b_flags & B_DELWRI) { 1129 bp->b_flags &= ~B_DELWRI; 1130 reassignbuf(bp, bp->b_vp); 1131 atomic_subtract_int(&numdirtybuffers, 1); 1132 numdirtywakeup(lodirtybuffers); 1133 } 1134 /* 1135 * Since it is now being written, we can clear its deferred write flag. 1136 */ 1137 bp->b_flags &= ~B_DEFERRED; 1138} 1139 1140/* 1141 * bawrite: 1142 * 1143 * Asynchronous write. Start output on a buffer, but do not wait for 1144 * it to complete. The buffer is released when the output completes. 1145 * 1146 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1147 * B_INVAL buffers. Not us. 1148 */ 1149void 1150bawrite(struct buf * bp) 1151{ 1152 bp->b_flags |= B_ASYNC; 1153 (void) BUF_WRITE(bp); 1154} 1155 1156/* 1157 * bwillwrite: 1158 * 1159 * Called prior to the locking of any vnodes when we are expecting to 1160 * write. We do not want to starve the buffer cache with too many 1161 * dirty buffers so we block here. By blocking prior to the locking 1162 * of any vnodes we attempt to avoid the situation where a locked vnode 1163 * prevents the various system daemons from flushing related buffers. 1164 */ 1165 1166void 1167bwillwrite(void) 1168{ 1169 if (numdirtybuffers >= hidirtybuffers) { 1170 int s; 1171 1172 mtx_lock(&Giant); 1173 s = splbio(); 1174 mtx_lock(&nblock); 1175 while (numdirtybuffers >= hidirtybuffers) { 1176 bd_wakeup(1); 1177 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1178 msleep(&needsbuffer, &nblock, 1179 (PRIBIO + 4), "flswai", 0); 1180 } 1181 splx(s); 1182 mtx_unlock(&nblock); 1183 mtx_unlock(&Giant); 1184 } 1185} 1186 1187/* 1188 * Return true if we have too many dirty buffers. 1189 */ 1190int 1191buf_dirty_count_severe(void) 1192{ 1193 return(numdirtybuffers >= hidirtybuffers); 1194} 1195 1196/* 1197 * brelse: 1198 * 1199 * Release a busy buffer and, if requested, free its resources. The 1200 * buffer will be stashed in the appropriate bufqueue[] allowing it 1201 * to be accessed later as a cache entity or reused for other purposes. 1202 */ 1203void 1204brelse(struct buf * bp) 1205{ 1206 int s; 1207 1208 GIANT_REQUIRED; 1209 1210 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1211 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1212 1213 s = splbio(); 1214 1215 if (bp->b_flags & B_LOCKED) 1216 bp->b_ioflags &= ~BIO_ERROR; 1217 1218 if (bp->b_iocmd == BIO_WRITE && 1219 (bp->b_ioflags & BIO_ERROR) && 1220 !(bp->b_flags & B_INVAL)) { 1221 /* 1222 * Failed write, redirty. Must clear BIO_ERROR to prevent 1223 * pages from being scrapped. If B_INVAL is set then 1224 * this case is not run and the next case is run to 1225 * destroy the buffer. B_INVAL can occur if the buffer 1226 * is outside the range supported by the underlying device. 1227 */ 1228 bp->b_ioflags &= ~BIO_ERROR; 1229 bdirty(bp); 1230 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1231 (bp->b_ioflags & BIO_ERROR) || 1232 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1233 /* 1234 * Either a failed I/O or we were asked to free or not 1235 * cache the buffer. 1236 */ 1237 bp->b_flags |= B_INVAL; 1238 if (LIST_FIRST(&bp->b_dep) != NULL) 1239 buf_deallocate(bp); 1240 if (bp->b_flags & B_DELWRI) { 1241 atomic_subtract_int(&numdirtybuffers, 1); 1242 numdirtywakeup(lodirtybuffers); 1243 } 1244 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1245 if ((bp->b_flags & B_VMIO) == 0) { 1246 if (bp->b_bufsize) 1247 allocbuf(bp, 0); 1248 if (bp->b_vp) 1249 brelvp(bp); 1250 } 1251 } 1252 1253 /* 1254 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1255 * is called with B_DELWRI set, the underlying pages may wind up 1256 * getting freed causing a previous write (bdwrite()) to get 'lost' 1257 * because pages associated with a B_DELWRI bp are marked clean. 1258 * 1259 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1260 * if B_DELWRI is set. 1261 * 1262 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1263 * on pages to return pages to the VM page queues. 1264 */ 1265 if (bp->b_flags & B_DELWRI) 1266 bp->b_flags &= ~B_RELBUF; 1267 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1268 bp->b_flags |= B_RELBUF; 1269 1270 /* 1271 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1272 * constituted, not even NFS buffers now. Two flags effect this. If 1273 * B_INVAL, the struct buf is invalidated but the VM object is kept 1274 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1275 * 1276 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1277 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1278 * buffer is also B_INVAL because it hits the re-dirtying code above. 1279 * 1280 * Normally we can do this whether a buffer is B_DELWRI or not. If 1281 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1282 * the commit state and we cannot afford to lose the buffer. If the 1283 * buffer has a background write in progress, we need to keep it 1284 * around to prevent it from being reconstituted and starting a second 1285 * background write. 1286 */ 1287 if ((bp->b_flags & B_VMIO) 1288 && !(bp->b_vp->v_mount != NULL && 1289 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1290 !vn_isdisk(bp->b_vp, NULL) && 1291 (bp->b_flags & B_DELWRI)) 1292 ) { 1293 1294 int i, j, resid; 1295 vm_page_t m; 1296 off_t foff; 1297 vm_pindex_t poff; 1298 vm_object_t obj; 1299 struct vnode *vp; 1300 1301 vp = bp->b_vp; 1302 obj = bp->b_object; 1303 1304 /* 1305 * Get the base offset and length of the buffer. Note that 1306 * in the VMIO case if the buffer block size is not 1307 * page-aligned then b_data pointer may not be page-aligned. 1308 * But our b_pages[] array *IS* page aligned. 1309 * 1310 * block sizes less then DEV_BSIZE (usually 512) are not 1311 * supported due to the page granularity bits (m->valid, 1312 * m->dirty, etc...). 1313 * 1314 * See man buf(9) for more information 1315 */ 1316 resid = bp->b_bufsize; 1317 foff = bp->b_offset; 1318 if (obj != NULL) 1319 VM_OBJECT_LOCK(obj); 1320 for (i = 0; i < bp->b_npages; i++) { 1321 int had_bogus = 0; 1322 1323 m = bp->b_pages[i]; 1324 vm_page_lock_queues(); 1325 vm_page_flag_clear(m, PG_ZERO); 1326 vm_page_unlock_queues(); 1327 1328 /* 1329 * If we hit a bogus page, fixup *all* the bogus pages 1330 * now. 1331 */ 1332 if (m == bogus_page) { 1333 poff = OFF_TO_IDX(bp->b_offset); 1334 had_bogus = 1; 1335 1336 for (j = i; j < bp->b_npages; j++) { 1337 vm_page_t mtmp; 1338 mtmp = bp->b_pages[j]; 1339 if (mtmp == bogus_page) { 1340 mtmp = vm_page_lookup(obj, poff + j); 1341 if (!mtmp) { 1342 panic("brelse: page missing\n"); 1343 } 1344 bp->b_pages[j] = mtmp; 1345 } 1346 } 1347 1348 if ((bp->b_flags & B_INVAL) == 0) { 1349 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1350 } 1351 m = bp->b_pages[i]; 1352 } 1353 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1354 int poffset = foff & PAGE_MASK; 1355 int presid = resid > (PAGE_SIZE - poffset) ? 1356 (PAGE_SIZE - poffset) : resid; 1357 1358 KASSERT(presid >= 0, ("brelse: extra page")); 1359 vm_page_set_invalid(m, poffset, presid); 1360 if (had_bogus) 1361 printf("avoided corruption bug in bogus_page/brelse code\n"); 1362 } 1363 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1364 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1365 } 1366 if (obj != NULL) 1367 VM_OBJECT_UNLOCK(obj); 1368 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1369 vfs_vmio_release(bp); 1370 1371 } else if (bp->b_flags & B_VMIO) { 1372 1373 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1374 vfs_vmio_release(bp); 1375 } 1376 1377 } 1378 1379 if (bp->b_qindex != QUEUE_NONE) 1380 panic("brelse: free buffer onto another queue???"); 1381 if (BUF_REFCNT(bp) > 1) { 1382 /* do not release to free list */ 1383 BUF_UNLOCK(bp); 1384 splx(s); 1385 return; 1386 } 1387 1388 /* enqueue */ 1389 mtx_lock(&bqlock); 1390 1391 /* buffers with no memory */ 1392 if (bp->b_bufsize == 0) { 1393 bp->b_flags |= B_INVAL; 1394 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1395 if (bp->b_xflags & BX_BKGRDINPROG) 1396 panic("losing buffer 1"); 1397 if (bp->b_kvasize) { 1398 bp->b_qindex = QUEUE_EMPTYKVA; 1399 } else { 1400 bp->b_qindex = QUEUE_EMPTY; 1401 } 1402 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1403 bp->b_dev = NODEV; 1404 /* buffers with junk contents */ 1405 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1406 (bp->b_ioflags & BIO_ERROR)) { 1407 bp->b_flags |= B_INVAL; 1408 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1409 if (bp->b_xflags & BX_BKGRDINPROG) 1410 panic("losing buffer 2"); 1411 bp->b_qindex = QUEUE_CLEAN; 1412 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1413 bp->b_dev = NODEV; 1414 1415 /* buffers that are locked */ 1416 } else if (bp->b_flags & B_LOCKED) { 1417 bp->b_qindex = QUEUE_LOCKED; 1418 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1419 1420 /* remaining buffers */ 1421 } else { 1422 if (bp->b_flags & B_DELWRI) 1423 bp->b_qindex = QUEUE_DIRTY; 1424 else 1425 bp->b_qindex = QUEUE_CLEAN; 1426 if (bp->b_flags & B_AGE) 1427 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1428 else 1429 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1430 } 1431 mtx_unlock(&bqlock); 1432 1433 /* 1434 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1435 * placed the buffer on the correct queue. We must also disassociate 1436 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1437 * find it. 1438 */ 1439 if (bp->b_flags & B_INVAL) { 1440 if (bp->b_flags & B_DELWRI) 1441 bundirty(bp); 1442 if (bp->b_vp) 1443 brelvp(bp); 1444 } 1445 1446 /* 1447 * Fixup numfreebuffers count. The bp is on an appropriate queue 1448 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1449 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1450 * if B_INVAL is set ). 1451 */ 1452 1453 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1454 bufcountwakeup(); 1455 1456 /* 1457 * Something we can maybe free or reuse 1458 */ 1459 if (bp->b_bufsize || bp->b_kvasize) 1460 bufspacewakeup(); 1461 1462 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1463 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1464 panic("brelse: not dirty"); 1465 /* unlock */ 1466 BUF_UNLOCK(bp); 1467 splx(s); 1468} 1469 1470/* 1471 * Release a buffer back to the appropriate queue but do not try to free 1472 * it. The buffer is expected to be used again soon. 1473 * 1474 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1475 * biodone() to requeue an async I/O on completion. It is also used when 1476 * known good buffers need to be requeued but we think we may need the data 1477 * again soon. 1478 * 1479 * XXX we should be able to leave the B_RELBUF hint set on completion. 1480 */ 1481void 1482bqrelse(struct buf * bp) 1483{ 1484 int s; 1485 1486 s = splbio(); 1487 1488 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1489 1490 if (bp->b_qindex != QUEUE_NONE) 1491 panic("bqrelse: free buffer onto another queue???"); 1492 if (BUF_REFCNT(bp) > 1) { 1493 /* do not release to free list */ 1494 BUF_UNLOCK(bp); 1495 splx(s); 1496 return; 1497 } 1498 mtx_lock(&bqlock); 1499 if (bp->b_flags & B_LOCKED) { 1500 bp->b_ioflags &= ~BIO_ERROR; 1501 bp->b_qindex = QUEUE_LOCKED; 1502 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1503 /* buffers with stale but valid contents */ 1504 } else if (bp->b_flags & B_DELWRI) { 1505 bp->b_qindex = QUEUE_DIRTY; 1506 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1507 } else if (vm_page_count_severe()) { 1508 /* 1509 * We are too low on memory, we have to try to free the 1510 * buffer (most importantly: the wired pages making up its 1511 * backing store) *now*. 1512 */ 1513 mtx_unlock(&bqlock); 1514 splx(s); 1515 brelse(bp); 1516 return; 1517 } else { 1518 bp->b_qindex = QUEUE_CLEAN; 1519 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1520 } 1521 mtx_unlock(&bqlock); 1522 1523 if ((bp->b_flags & B_LOCKED) == 0 && 1524 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1525 bufcountwakeup(); 1526 } 1527 1528 /* 1529 * Something we can maybe free or reuse. 1530 */ 1531 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1532 bufspacewakeup(); 1533 1534 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1535 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1536 panic("bqrelse: not dirty"); 1537 /* unlock */ 1538 BUF_UNLOCK(bp); 1539 splx(s); 1540} 1541 1542/* Give pages used by the bp back to the VM system (where possible) */ 1543static void 1544vfs_vmio_release(bp) 1545 struct buf *bp; 1546{ 1547 int i; 1548 vm_page_t m; 1549 1550 GIANT_REQUIRED; 1551 if (bp->b_object != NULL) 1552 VM_OBJECT_LOCK(bp->b_object); 1553 vm_page_lock_queues(); 1554 for (i = 0; i < bp->b_npages; i++) { 1555 m = bp->b_pages[i]; 1556 bp->b_pages[i] = NULL; 1557 /* 1558 * In order to keep page LRU ordering consistent, put 1559 * everything on the inactive queue. 1560 */ 1561 vm_page_unwire(m, 0); 1562 /* 1563 * We don't mess with busy pages, it is 1564 * the responsibility of the process that 1565 * busied the pages to deal with them. 1566 */ 1567 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1568 continue; 1569 1570 if (m->wire_count == 0) { 1571 vm_page_flag_clear(m, PG_ZERO); 1572 /* 1573 * Might as well free the page if we can and it has 1574 * no valid data. We also free the page if the 1575 * buffer was used for direct I/O 1576 */ 1577 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1578 m->hold_count == 0) { 1579 vm_page_busy(m); 1580 pmap_remove_all(m); 1581 vm_page_free(m); 1582 } else if (bp->b_flags & B_DIRECT) { 1583 vm_page_try_to_free(m); 1584 } else if (vm_page_count_severe()) { 1585 vm_page_try_to_cache(m); 1586 } 1587 } 1588 } 1589 vm_page_unlock_queues(); 1590 if (bp->b_object != NULL) 1591 VM_OBJECT_UNLOCK(bp->b_object); 1592 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1593 1594 if (bp->b_bufsize) { 1595 bufspacewakeup(); 1596 bp->b_bufsize = 0; 1597 } 1598 bp->b_npages = 0; 1599 bp->b_flags &= ~B_VMIO; 1600 if (bp->b_vp) 1601 brelvp(bp); 1602} 1603 1604/* 1605 * Check to see if a block at a particular lbn is available for a clustered 1606 * write. 1607 */ 1608static int 1609vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1610{ 1611 struct buf *bpa; 1612 int match; 1613 1614 match = 0; 1615 1616 /* If the buf isn't in core skip it */ 1617 if ((bpa = gbincore(vp, lblkno)) == NULL) 1618 return (0); 1619 1620 /* If the buf is busy we don't want to wait for it */ 1621 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1622 return (0); 1623 1624 /* Only cluster with valid clusterable delayed write buffers */ 1625 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1626 (B_DELWRI | B_CLUSTEROK)) 1627 goto done; 1628 1629 if (bpa->b_bufsize != size) 1630 goto done; 1631 1632 /* 1633 * Check to see if it is in the expected place on disk and that the 1634 * block has been mapped. 1635 */ 1636 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1637 match = 1; 1638done: 1639 BUF_UNLOCK(bpa); 1640 return (match); 1641} 1642 1643/* 1644 * vfs_bio_awrite: 1645 * 1646 * Implement clustered async writes for clearing out B_DELWRI buffers. 1647 * This is much better then the old way of writing only one buffer at 1648 * a time. Note that we may not be presented with the buffers in the 1649 * correct order, so we search for the cluster in both directions. 1650 */ 1651int 1652vfs_bio_awrite(struct buf * bp) 1653{ 1654 int i; 1655 int j; 1656 daddr_t lblkno = bp->b_lblkno; 1657 struct vnode *vp = bp->b_vp; 1658 int s; 1659 int ncl; 1660 int nwritten; 1661 int size; 1662 int maxcl; 1663 1664 s = splbio(); 1665 /* 1666 * right now we support clustered writing only to regular files. If 1667 * we find a clusterable block we could be in the middle of a cluster 1668 * rather then at the beginning. 1669 */ 1670 if ((vp->v_type == VREG) && 1671 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1672 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1673 1674 size = vp->v_mount->mnt_stat.f_iosize; 1675 maxcl = MAXPHYS / size; 1676 1677 VI_LOCK(vp); 1678 for (i = 1; i < maxcl; i++) 1679 if (vfs_bio_clcheck(vp, size, lblkno + i, 1680 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1681 break; 1682 1683 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1684 if (vfs_bio_clcheck(vp, size, lblkno - j, 1685 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1686 break; 1687 1688 VI_UNLOCK(vp); 1689 --j; 1690 ncl = i + j; 1691 /* 1692 * this is a possible cluster write 1693 */ 1694 if (ncl != 1) { 1695 BUF_UNLOCK(bp); 1696 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1697 splx(s); 1698 return nwritten; 1699 } 1700 } 1701 1702 bremfree(bp); 1703 bp->b_flags |= B_ASYNC; 1704 1705 splx(s); 1706 /* 1707 * default (old) behavior, writing out only one block 1708 * 1709 * XXX returns b_bufsize instead of b_bcount for nwritten? 1710 */ 1711 nwritten = bp->b_bufsize; 1712 (void) BUF_WRITE(bp); 1713 1714 return nwritten; 1715} 1716 1717/* 1718 * getnewbuf: 1719 * 1720 * Find and initialize a new buffer header, freeing up existing buffers 1721 * in the bufqueues as necessary. The new buffer is returned locked. 1722 * 1723 * Important: B_INVAL is not set. If the caller wishes to throw the 1724 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1725 * 1726 * We block if: 1727 * We have insufficient buffer headers 1728 * We have insufficient buffer space 1729 * buffer_map is too fragmented ( space reservation fails ) 1730 * If we have to flush dirty buffers ( but we try to avoid this ) 1731 * 1732 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1733 * Instead we ask the buf daemon to do it for us. We attempt to 1734 * avoid piecemeal wakeups of the pageout daemon. 1735 */ 1736 1737static struct buf * 1738getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1739{ 1740 struct buf *bp; 1741 struct buf *nbp; 1742 int defrag = 0; 1743 int nqindex; 1744 static int flushingbufs; 1745 1746 GIANT_REQUIRED; 1747 1748 /* 1749 * We can't afford to block since we might be holding a vnode lock, 1750 * which may prevent system daemons from running. We deal with 1751 * low-memory situations by proactively returning memory and running 1752 * async I/O rather then sync I/O. 1753 */ 1754 1755 atomic_add_int(&getnewbufcalls, 1); 1756 atomic_subtract_int(&getnewbufrestarts, 1); 1757restart: 1758 atomic_add_int(&getnewbufrestarts, 1); 1759 1760 /* 1761 * Setup for scan. If we do not have enough free buffers, 1762 * we setup a degenerate case that immediately fails. Note 1763 * that if we are specially marked process, we are allowed to 1764 * dip into our reserves. 1765 * 1766 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1767 * 1768 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1769 * However, there are a number of cases (defragging, reusing, ...) 1770 * where we cannot backup. 1771 */ 1772 mtx_lock(&bqlock); 1773 nqindex = QUEUE_EMPTYKVA; 1774 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1775 1776 if (nbp == NULL) { 1777 /* 1778 * If no EMPTYKVA buffers and we are either 1779 * defragging or reusing, locate a CLEAN buffer 1780 * to free or reuse. If bufspace useage is low 1781 * skip this step so we can allocate a new buffer. 1782 */ 1783 if (defrag || bufspace >= lobufspace) { 1784 nqindex = QUEUE_CLEAN; 1785 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1786 } 1787 1788 /* 1789 * If we could not find or were not allowed to reuse a 1790 * CLEAN buffer, check to see if it is ok to use an EMPTY 1791 * buffer. We can only use an EMPTY buffer if allocating 1792 * its KVA would not otherwise run us out of buffer space. 1793 */ 1794 if (nbp == NULL && defrag == 0 && 1795 bufspace + maxsize < hibufspace) { 1796 nqindex = QUEUE_EMPTY; 1797 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1798 } 1799 } 1800 1801 /* 1802 * Run scan, possibly freeing data and/or kva mappings on the fly 1803 * depending. 1804 */ 1805 1806 while ((bp = nbp) != NULL) { 1807 int qindex = nqindex; 1808 1809 /* 1810 * Calculate next bp ( we can only use it if we do not block 1811 * or do other fancy things ). 1812 */ 1813 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1814 switch(qindex) { 1815 case QUEUE_EMPTY: 1816 nqindex = QUEUE_EMPTYKVA; 1817 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1818 break; 1819 /* FALLTHROUGH */ 1820 case QUEUE_EMPTYKVA: 1821 nqindex = QUEUE_CLEAN; 1822 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1823 break; 1824 /* FALLTHROUGH */ 1825 case QUEUE_CLEAN: 1826 /* 1827 * nbp is NULL. 1828 */ 1829 break; 1830 } 1831 } 1832 1833 /* 1834 * Sanity Checks 1835 */ 1836 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1837 1838 /* 1839 * Note: we no longer distinguish between VMIO and non-VMIO 1840 * buffers. 1841 */ 1842 1843 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1844 1845 /* 1846 * If we are defragging then we need a buffer with 1847 * b_kvasize != 0. XXX this situation should no longer 1848 * occur, if defrag is non-zero the buffer's b_kvasize 1849 * should also be non-zero at this point. XXX 1850 */ 1851 if (defrag && bp->b_kvasize == 0) { 1852 printf("Warning: defrag empty buffer %p\n", bp); 1853 continue; 1854 } 1855 1856 /* 1857 * Start freeing the bp. This is somewhat involved. nbp 1858 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1859 */ 1860 1861 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1862 panic("getnewbuf: locked buf"); 1863 bremfreel(bp); 1864 mtx_unlock(&bqlock); 1865 1866 if (qindex == QUEUE_CLEAN) { 1867 if (bp->b_flags & B_VMIO) { 1868 bp->b_flags &= ~B_ASYNC; 1869 vfs_vmio_release(bp); 1870 } 1871 if (bp->b_vp) 1872 brelvp(bp); 1873 } 1874 1875 /* 1876 * NOTE: nbp is now entirely invalid. We can only restart 1877 * the scan from this point on. 1878 * 1879 * Get the rest of the buffer freed up. b_kva* is still 1880 * valid after this operation. 1881 */ 1882 1883 if (bp->b_rcred != NOCRED) { 1884 crfree(bp->b_rcred); 1885 bp->b_rcred = NOCRED; 1886 } 1887 if (bp->b_wcred != NOCRED) { 1888 crfree(bp->b_wcred); 1889 bp->b_wcred = NOCRED; 1890 } 1891 if (LIST_FIRST(&bp->b_dep) != NULL) 1892 buf_deallocate(bp); 1893 if (bp->b_xflags & BX_BKGRDINPROG) 1894 panic("losing buffer 3"); 1895 1896 if (bp->b_bufsize) 1897 allocbuf(bp, 0); 1898 1899 bp->b_flags = 0; 1900 bp->b_ioflags = 0; 1901 bp->b_xflags = 0; 1902 bp->b_vflags = 0; 1903 bp->b_dev = NODEV; 1904 bp->b_vp = NULL; 1905 bp->b_blkno = bp->b_lblkno = 0; 1906 bp->b_offset = NOOFFSET; 1907 bp->b_iodone = 0; 1908 bp->b_error = 0; 1909 bp->b_resid = 0; 1910 bp->b_bcount = 0; 1911 bp->b_npages = 0; 1912 bp->b_dirtyoff = bp->b_dirtyend = 0; 1913 bp->b_magic = B_MAGIC_BIO; 1914 bp->b_op = &buf_ops_bio; 1915 bp->b_object = NULL; 1916 1917 LIST_INIT(&bp->b_dep); 1918 1919 /* 1920 * If we are defragging then free the buffer. 1921 */ 1922 if (defrag) { 1923 bp->b_flags |= B_INVAL; 1924 bfreekva(bp); 1925 brelse(bp); 1926 defrag = 0; 1927 goto restart; 1928 } 1929 1930 /* 1931 * If we are overcomitted then recover the buffer and its 1932 * KVM space. This occurs in rare situations when multiple 1933 * processes are blocked in getnewbuf() or allocbuf(). 1934 */ 1935 if (bufspace >= hibufspace) 1936 flushingbufs = 1; 1937 if (flushingbufs && bp->b_kvasize != 0) { 1938 bp->b_flags |= B_INVAL; 1939 bfreekva(bp); 1940 brelse(bp); 1941 goto restart; 1942 } 1943 if (bufspace < lobufspace) 1944 flushingbufs = 0; 1945 break; 1946 } 1947 1948 /* 1949 * If we exhausted our list, sleep as appropriate. We may have to 1950 * wakeup various daemons and write out some dirty buffers. 1951 * 1952 * Generally we are sleeping due to insufficient buffer space. 1953 */ 1954 1955 if (bp == NULL) { 1956 int flags; 1957 char *waitmsg; 1958 1959 mtx_unlock(&bqlock); 1960 if (defrag) { 1961 flags = VFS_BIO_NEED_BUFSPACE; 1962 waitmsg = "nbufkv"; 1963 } else if (bufspace >= hibufspace) { 1964 waitmsg = "nbufbs"; 1965 flags = VFS_BIO_NEED_BUFSPACE; 1966 } else { 1967 waitmsg = "newbuf"; 1968 flags = VFS_BIO_NEED_ANY; 1969 } 1970 1971 bd_speedup(); /* heeeelp */ 1972 1973 mtx_lock(&nblock); 1974 needsbuffer |= flags; 1975 while (needsbuffer & flags) { 1976 if (msleep(&needsbuffer, &nblock, 1977 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1978 mtx_unlock(&nblock); 1979 return (NULL); 1980 } 1981 } 1982 mtx_unlock(&nblock); 1983 } else { 1984 /* 1985 * We finally have a valid bp. We aren't quite out of the 1986 * woods, we still have to reserve kva space. In order 1987 * to keep fragmentation sane we only allocate kva in 1988 * BKVASIZE chunks. 1989 */ 1990 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1991 1992 if (maxsize != bp->b_kvasize) { 1993 vm_offset_t addr = 0; 1994 1995 bfreekva(bp); 1996 1997 if (vm_map_findspace(buffer_map, 1998 vm_map_min(buffer_map), maxsize, &addr)) { 1999 /* 2000 * Uh oh. Buffer map is to fragmented. We 2001 * must defragment the map. 2002 */ 2003 atomic_add_int(&bufdefragcnt, 1); 2004 defrag = 1; 2005 bp->b_flags |= B_INVAL; 2006 brelse(bp); 2007 goto restart; 2008 } 2009 if (addr) { 2010 vm_map_insert(buffer_map, NULL, 0, 2011 addr, addr + maxsize, 2012 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2013 2014 bp->b_kvabase = (caddr_t) addr; 2015 bp->b_kvasize = maxsize; 2016 atomic_add_int(&bufspace, bp->b_kvasize); 2017 atomic_add_int(&bufreusecnt, 1); 2018 } 2019 } 2020 bp->b_saveaddr = bp->b_kvabase; 2021 bp->b_data = bp->b_saveaddr; 2022 } 2023 return(bp); 2024} 2025 2026/* 2027 * buf_daemon: 2028 * 2029 * buffer flushing daemon. Buffers are normally flushed by the 2030 * update daemon but if it cannot keep up this process starts to 2031 * take the load in an attempt to prevent getnewbuf() from blocking. 2032 */ 2033 2034static struct proc *bufdaemonproc; 2035 2036static struct kproc_desc buf_kp = { 2037 "bufdaemon", 2038 buf_daemon, 2039 &bufdaemonproc 2040}; 2041SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2042 2043static void 2044buf_daemon() 2045{ 2046 int s; 2047 2048 mtx_lock(&Giant); 2049 2050 /* 2051 * This process needs to be suspended prior to shutdown sync. 2052 */ 2053 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2054 SHUTDOWN_PRI_LAST); 2055 2056 /* 2057 * This process is allowed to take the buffer cache to the limit 2058 */ 2059 s = splbio(); 2060 mtx_lock(&bdlock); 2061 2062 for (;;) { 2063 bd_request = 0; 2064 mtx_unlock(&bdlock); 2065 2066 kthread_suspend_check(bufdaemonproc); 2067 2068 /* 2069 * Do the flush. Limit the amount of in-transit I/O we 2070 * allow to build up, otherwise we would completely saturate 2071 * the I/O system. Wakeup any waiting processes before we 2072 * normally would so they can run in parallel with our drain. 2073 */ 2074 while (numdirtybuffers > lodirtybuffers) { 2075 if (flushbufqueues(0) == 0) { 2076 /* 2077 * Could not find any buffers without rollback 2078 * dependencies, so just write the first one 2079 * in the hopes of eventually making progress. 2080 */ 2081 flushbufqueues(1); 2082 break; 2083 } 2084 waitrunningbufspace(); 2085 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2086 } 2087 2088 /* 2089 * Only clear bd_request if we have reached our low water 2090 * mark. The buf_daemon normally waits 1 second and 2091 * then incrementally flushes any dirty buffers that have 2092 * built up, within reason. 2093 * 2094 * If we were unable to hit our low water mark and couldn't 2095 * find any flushable buffers, we sleep half a second. 2096 * Otherwise we loop immediately. 2097 */ 2098 mtx_lock(&bdlock); 2099 if (numdirtybuffers <= lodirtybuffers) { 2100 /* 2101 * We reached our low water mark, reset the 2102 * request and sleep until we are needed again. 2103 * The sleep is just so the suspend code works. 2104 */ 2105 bd_request = 0; 2106 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2107 } else { 2108 /* 2109 * We couldn't find any flushable dirty buffers but 2110 * still have too many dirty buffers, we 2111 * have to sleep and try again. (rare) 2112 */ 2113 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2114 } 2115 } 2116} 2117 2118/* 2119 * flushbufqueues: 2120 * 2121 * Try to flush a buffer in the dirty queue. We must be careful to 2122 * free up B_INVAL buffers instead of write them, which NFS is 2123 * particularly sensitive to. 2124 */ 2125int flushwithdeps = 0; 2126SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2127 0, "Number of buffers flushed with dependecies that require rollbacks"); 2128static int 2129flushbufqueues(int flushdeps) 2130{ 2131 struct thread *td = curthread; 2132 struct vnode *vp; 2133 struct buf *bp; 2134 int hasdeps; 2135 2136 mtx_lock(&bqlock); 2137 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2138 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2139 continue; 2140 KASSERT((bp->b_flags & B_DELWRI), 2141 ("unexpected clean buffer %p", bp)); 2142 if ((bp->b_xflags & BX_BKGRDINPROG) != 0) { 2143 BUF_UNLOCK(bp); 2144 continue; 2145 } 2146 if (bp->b_flags & B_INVAL) { 2147 bremfreel(bp); 2148 mtx_unlock(&bqlock); 2149 brelse(bp); 2150 return (1); 2151 } 2152 2153 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2154 if (flushdeps == 0) { 2155 BUF_UNLOCK(bp); 2156 continue; 2157 } 2158 hasdeps = 1; 2159 } else 2160 hasdeps = 0; 2161 /* 2162 * We must hold the lock on a vnode before writing 2163 * one of its buffers. Otherwise we may confuse, or 2164 * in the case of a snapshot vnode, deadlock the 2165 * system. 2166 * 2167 * The lock order here is the reverse of the normal 2168 * of vnode followed by buf lock. This is ok because 2169 * the NOWAIT will prevent deadlock. 2170 */ 2171 if ((vp = bp->b_vp) == NULL || 2172 vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2173 mtx_unlock(&bqlock); 2174 vfs_bio_awrite(bp); 2175 if (vp != NULL) 2176 VOP_UNLOCK(vp, 0, td); 2177 flushwithdeps += hasdeps; 2178 return (1); 2179 } 2180 BUF_UNLOCK(bp); 2181 } 2182 mtx_unlock(&bqlock); 2183 return (0); 2184} 2185 2186/* 2187 * Check to see if a block is currently memory resident. 2188 */ 2189struct buf * 2190incore(struct vnode * vp, daddr_t blkno) 2191{ 2192 struct buf *bp; 2193 2194 int s = splbio(); 2195 VI_LOCK(vp); 2196 bp = gbincore(vp, blkno); 2197 VI_UNLOCK(vp); 2198 splx(s); 2199 return (bp); 2200} 2201 2202/* 2203 * Returns true if no I/O is needed to access the 2204 * associated VM object. This is like incore except 2205 * it also hunts around in the VM system for the data. 2206 */ 2207 2208int 2209inmem(struct vnode * vp, daddr_t blkno) 2210{ 2211 vm_object_t obj; 2212 vm_offset_t toff, tinc, size; 2213 vm_page_t m; 2214 vm_ooffset_t off; 2215 2216 GIANT_REQUIRED; 2217 ASSERT_VOP_LOCKED(vp, "inmem"); 2218 2219 if (incore(vp, blkno)) 2220 return 1; 2221 if (vp->v_mount == NULL) 2222 return 0; 2223 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2224 return 0; 2225 2226 size = PAGE_SIZE; 2227 if (size > vp->v_mount->mnt_stat.f_iosize) 2228 size = vp->v_mount->mnt_stat.f_iosize; 2229 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2230 2231 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2232 VM_OBJECT_LOCK(obj); 2233 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2234 VM_OBJECT_UNLOCK(obj); 2235 if (!m) 2236 goto notinmem; 2237 tinc = size; 2238 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2239 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2240 if (vm_page_is_valid(m, 2241 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2242 goto notinmem; 2243 } 2244 return 1; 2245 2246notinmem: 2247 return (0); 2248} 2249 2250/* 2251 * vfs_setdirty: 2252 * 2253 * Sets the dirty range for a buffer based on the status of the dirty 2254 * bits in the pages comprising the buffer. 2255 * 2256 * The range is limited to the size of the buffer. 2257 * 2258 * This routine is primarily used by NFS, but is generalized for the 2259 * B_VMIO case. 2260 */ 2261static void 2262vfs_setdirty(struct buf *bp) 2263{ 2264 int i; 2265 vm_object_t object; 2266 2267 GIANT_REQUIRED; 2268 /* 2269 * Degenerate case - empty buffer 2270 */ 2271 2272 if (bp->b_bufsize == 0) 2273 return; 2274 2275 /* 2276 * We qualify the scan for modified pages on whether the 2277 * object has been flushed yet. The OBJ_WRITEABLE flag 2278 * is not cleared simply by protecting pages off. 2279 */ 2280 2281 if ((bp->b_flags & B_VMIO) == 0) 2282 return; 2283 2284 object = bp->b_pages[0]->object; 2285 VM_OBJECT_LOCK(object); 2286 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2287 printf("Warning: object %p writeable but not mightbedirty\n", object); 2288 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2289 printf("Warning: object %p mightbedirty but not writeable\n", object); 2290 2291 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2292 vm_offset_t boffset; 2293 vm_offset_t eoffset; 2294 2295 vm_page_lock_queues(); 2296 /* 2297 * test the pages to see if they have been modified directly 2298 * by users through the VM system. 2299 */ 2300 for (i = 0; i < bp->b_npages; i++) { 2301 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2302 vm_page_test_dirty(bp->b_pages[i]); 2303 } 2304 2305 /* 2306 * Calculate the encompassing dirty range, boffset and eoffset, 2307 * (eoffset - boffset) bytes. 2308 */ 2309 2310 for (i = 0; i < bp->b_npages; i++) { 2311 if (bp->b_pages[i]->dirty) 2312 break; 2313 } 2314 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2315 2316 for (i = bp->b_npages - 1; i >= 0; --i) { 2317 if (bp->b_pages[i]->dirty) { 2318 break; 2319 } 2320 } 2321 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2322 2323 vm_page_unlock_queues(); 2324 /* 2325 * Fit it to the buffer. 2326 */ 2327 2328 if (eoffset > bp->b_bcount) 2329 eoffset = bp->b_bcount; 2330 2331 /* 2332 * If we have a good dirty range, merge with the existing 2333 * dirty range. 2334 */ 2335 2336 if (boffset < eoffset) { 2337 if (bp->b_dirtyoff > boffset) 2338 bp->b_dirtyoff = boffset; 2339 if (bp->b_dirtyend < eoffset) 2340 bp->b_dirtyend = eoffset; 2341 } 2342 } 2343 VM_OBJECT_UNLOCK(object); 2344} 2345 2346/* 2347 * getblk: 2348 * 2349 * Get a block given a specified block and offset into a file/device. 2350 * The buffers B_DONE bit will be cleared on return, making it almost 2351 * ready for an I/O initiation. B_INVAL may or may not be set on 2352 * return. The caller should clear B_INVAL prior to initiating a 2353 * READ. 2354 * 2355 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2356 * an existing buffer. 2357 * 2358 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2359 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2360 * and then cleared based on the backing VM. If the previous buffer is 2361 * non-0-sized but invalid, B_CACHE will be cleared. 2362 * 2363 * If getblk() must create a new buffer, the new buffer is returned with 2364 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2365 * case it is returned with B_INVAL clear and B_CACHE set based on the 2366 * backing VM. 2367 * 2368 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2369 * B_CACHE bit is clear. 2370 * 2371 * What this means, basically, is that the caller should use B_CACHE to 2372 * determine whether the buffer is fully valid or not and should clear 2373 * B_INVAL prior to issuing a read. If the caller intends to validate 2374 * the buffer by loading its data area with something, the caller needs 2375 * to clear B_INVAL. If the caller does this without issuing an I/O, 2376 * the caller should set B_CACHE ( as an optimization ), else the caller 2377 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2378 * a write attempt or if it was a successfull read. If the caller 2379 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2380 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2381 */ 2382struct buf * 2383getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2384 int flags) 2385{ 2386 struct buf *bp; 2387 int s; 2388 int error; 2389 ASSERT_VOP_LOCKED(vp, "getblk"); 2390 2391 if (size > MAXBSIZE) 2392 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2393 2394 s = splbio(); 2395loop: 2396 /* 2397 * Block if we are low on buffers. Certain processes are allowed 2398 * to completely exhaust the buffer cache. 2399 * 2400 * If this check ever becomes a bottleneck it may be better to 2401 * move it into the else, when gbincore() fails. At the moment 2402 * it isn't a problem. 2403 * 2404 * XXX remove if 0 sections (clean this up after its proven) 2405 */ 2406 if (numfreebuffers == 0) { 2407 if (curthread == PCPU_GET(idlethread)) 2408 return NULL; 2409 mtx_lock(&nblock); 2410 needsbuffer |= VFS_BIO_NEED_ANY; 2411 mtx_unlock(&nblock); 2412 } 2413 2414 VI_LOCK(vp); 2415 if ((bp = gbincore(vp, blkno))) { 2416 int lockflags; 2417 /* 2418 * Buffer is in-core. If the buffer is not busy, it must 2419 * be on a queue. 2420 */ 2421 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2422 2423 if (flags & GB_LOCK_NOWAIT) 2424 lockflags |= LK_NOWAIT; 2425 2426 error = BUF_TIMELOCK(bp, lockflags, 2427 VI_MTX(vp), "getblk", slpflag, slptimeo); 2428 2429 /* 2430 * If we slept and got the lock we have to restart in case 2431 * the buffer changed identities. 2432 */ 2433 if (error == ENOLCK) 2434 goto loop; 2435 /* We timed out or were interrupted. */ 2436 else if (error) 2437 return (NULL); 2438 2439 /* 2440 * The buffer is locked. B_CACHE is cleared if the buffer is 2441 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2442 * and for a VMIO buffer B_CACHE is adjusted according to the 2443 * backing VM cache. 2444 */ 2445 if (bp->b_flags & B_INVAL) 2446 bp->b_flags &= ~B_CACHE; 2447 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2448 bp->b_flags |= B_CACHE; 2449 bremfree(bp); 2450 2451 /* 2452 * check for size inconsistancies for non-VMIO case. 2453 */ 2454 2455 if (bp->b_bcount != size) { 2456 if ((bp->b_flags & B_VMIO) == 0 || 2457 (size > bp->b_kvasize)) { 2458 if (bp->b_flags & B_DELWRI) { 2459 bp->b_flags |= B_NOCACHE; 2460 BUF_WRITE(bp); 2461 } else { 2462 if ((bp->b_flags & B_VMIO) && 2463 (LIST_FIRST(&bp->b_dep) == NULL)) { 2464 bp->b_flags |= B_RELBUF; 2465 brelse(bp); 2466 } else { 2467 bp->b_flags |= B_NOCACHE; 2468 BUF_WRITE(bp); 2469 } 2470 } 2471 goto loop; 2472 } 2473 } 2474 2475 /* 2476 * If the size is inconsistant in the VMIO case, we can resize 2477 * the buffer. This might lead to B_CACHE getting set or 2478 * cleared. If the size has not changed, B_CACHE remains 2479 * unchanged from its previous state. 2480 */ 2481 2482 if (bp->b_bcount != size) 2483 allocbuf(bp, size); 2484 2485 KASSERT(bp->b_offset != NOOFFSET, 2486 ("getblk: no buffer offset")); 2487 2488 /* 2489 * A buffer with B_DELWRI set and B_CACHE clear must 2490 * be committed before we can return the buffer in 2491 * order to prevent the caller from issuing a read 2492 * ( due to B_CACHE not being set ) and overwriting 2493 * it. 2494 * 2495 * Most callers, including NFS and FFS, need this to 2496 * operate properly either because they assume they 2497 * can issue a read if B_CACHE is not set, or because 2498 * ( for example ) an uncached B_DELWRI might loop due 2499 * to softupdates re-dirtying the buffer. In the latter 2500 * case, B_CACHE is set after the first write completes, 2501 * preventing further loops. 2502 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2503 * above while extending the buffer, we cannot allow the 2504 * buffer to remain with B_CACHE set after the write 2505 * completes or it will represent a corrupt state. To 2506 * deal with this we set B_NOCACHE to scrap the buffer 2507 * after the write. 2508 * 2509 * We might be able to do something fancy, like setting 2510 * B_CACHE in bwrite() except if B_DELWRI is already set, 2511 * so the below call doesn't set B_CACHE, but that gets real 2512 * confusing. This is much easier. 2513 */ 2514 2515 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2516 bp->b_flags |= B_NOCACHE; 2517 BUF_WRITE(bp); 2518 goto loop; 2519 } 2520 2521 splx(s); 2522 bp->b_flags &= ~B_DONE; 2523 } else { 2524 int bsize, maxsize, vmio; 2525 off_t offset; 2526 2527 /* 2528 * Buffer is not in-core, create new buffer. The buffer 2529 * returned by getnewbuf() is locked. Note that the returned 2530 * buffer is also considered valid (not marked B_INVAL). 2531 */ 2532 VI_UNLOCK(vp); 2533 if (vn_isdisk(vp, NULL)) 2534 bsize = DEV_BSIZE; 2535 else if (vp->v_mountedhere) 2536 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2537 else if (vp->v_mount) 2538 bsize = vp->v_mount->mnt_stat.f_iosize; 2539 else 2540 bsize = size; 2541 2542 offset = blkno * bsize; 2543 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2544 (vp->v_vflag & VV_OBJBUF); 2545 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2546 maxsize = imax(maxsize, bsize); 2547 2548 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2549 if (slpflag || slptimeo) { 2550 splx(s); 2551 return NULL; 2552 } 2553 goto loop; 2554 } 2555 2556 /* 2557 * This code is used to make sure that a buffer is not 2558 * created while the getnewbuf routine is blocked. 2559 * This can be a problem whether the vnode is locked or not. 2560 * If the buffer is created out from under us, we have to 2561 * throw away the one we just created. There is now window 2562 * race because we are safely running at splbio() from the 2563 * point of the duplicate buffer creation through to here, 2564 * and we've locked the buffer. 2565 * 2566 * Note: this must occur before we associate the buffer 2567 * with the vp especially considering limitations in 2568 * the splay tree implementation when dealing with duplicate 2569 * lblkno's. 2570 */ 2571 VI_LOCK(vp); 2572 if (gbincore(vp, blkno)) { 2573 VI_UNLOCK(vp); 2574 bp->b_flags |= B_INVAL; 2575 brelse(bp); 2576 goto loop; 2577 } 2578 2579 /* 2580 * Insert the buffer into the hash, so that it can 2581 * be found by incore. 2582 */ 2583 bp->b_blkno = bp->b_lblkno = blkno; 2584 bp->b_offset = offset; 2585 2586 bgetvp(vp, bp); 2587 VI_UNLOCK(vp); 2588 2589 /* 2590 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2591 * buffer size starts out as 0, B_CACHE will be set by 2592 * allocbuf() for the VMIO case prior to it testing the 2593 * backing store for validity. 2594 */ 2595 2596 if (vmio) { 2597 bp->b_flags |= B_VMIO; 2598#if defined(VFS_BIO_DEBUG) 2599 if (vp->v_type != VREG) 2600 printf("getblk: vmioing file type %d???\n", vp->v_type); 2601#endif 2602 VOP_GETVOBJECT(vp, &bp->b_object); 2603 } else { 2604 bp->b_flags &= ~B_VMIO; 2605 bp->b_object = NULL; 2606 } 2607 2608 allocbuf(bp, size); 2609 2610 splx(s); 2611 bp->b_flags &= ~B_DONE; 2612 } 2613 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2614 return (bp); 2615} 2616 2617/* 2618 * Get an empty, disassociated buffer of given size. The buffer is initially 2619 * set to B_INVAL. 2620 */ 2621struct buf * 2622geteblk(int size) 2623{ 2624 struct buf *bp; 2625 int s; 2626 int maxsize; 2627 2628 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2629 2630 s = splbio(); 2631 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2632 continue; 2633 splx(s); 2634 allocbuf(bp, size); 2635 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2636 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2637 return (bp); 2638} 2639 2640 2641/* 2642 * This code constitutes the buffer memory from either anonymous system 2643 * memory (in the case of non-VMIO operations) or from an associated 2644 * VM object (in the case of VMIO operations). This code is able to 2645 * resize a buffer up or down. 2646 * 2647 * Note that this code is tricky, and has many complications to resolve 2648 * deadlock or inconsistant data situations. Tread lightly!!! 2649 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2650 * the caller. Calling this code willy nilly can result in the loss of data. 2651 * 2652 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2653 * B_CACHE for the non-VMIO case. 2654 */ 2655 2656int 2657allocbuf(struct buf *bp, int size) 2658{ 2659 int newbsize, mbsize; 2660 int i; 2661 2662 GIANT_REQUIRED; 2663 2664 if (BUF_REFCNT(bp) == 0) 2665 panic("allocbuf: buffer not busy"); 2666 2667 if (bp->b_kvasize < size) 2668 panic("allocbuf: buffer too small"); 2669 2670 if ((bp->b_flags & B_VMIO) == 0) { 2671 caddr_t origbuf; 2672 int origbufsize; 2673 /* 2674 * Just get anonymous memory from the kernel. Don't 2675 * mess with B_CACHE. 2676 */ 2677 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2678 if (bp->b_flags & B_MALLOC) 2679 newbsize = mbsize; 2680 else 2681 newbsize = round_page(size); 2682 2683 if (newbsize < bp->b_bufsize) { 2684 /* 2685 * malloced buffers are not shrunk 2686 */ 2687 if (bp->b_flags & B_MALLOC) { 2688 if (newbsize) { 2689 bp->b_bcount = size; 2690 } else { 2691 free(bp->b_data, M_BIOBUF); 2692 if (bp->b_bufsize) { 2693 atomic_subtract_int( 2694 &bufmallocspace, 2695 bp->b_bufsize); 2696 bufspacewakeup(); 2697 bp->b_bufsize = 0; 2698 } 2699 bp->b_saveaddr = bp->b_kvabase; 2700 bp->b_data = bp->b_saveaddr; 2701 bp->b_bcount = 0; 2702 bp->b_flags &= ~B_MALLOC; 2703 } 2704 return 1; 2705 } 2706 vm_hold_free_pages( 2707 bp, 2708 (vm_offset_t) bp->b_data + newbsize, 2709 (vm_offset_t) bp->b_data + bp->b_bufsize); 2710 } else if (newbsize > bp->b_bufsize) { 2711 /* 2712 * We only use malloced memory on the first allocation. 2713 * and revert to page-allocated memory when the buffer 2714 * grows. 2715 */ 2716 /* 2717 * There is a potential smp race here that could lead 2718 * to bufmallocspace slightly passing the max. It 2719 * is probably extremely rare and not worth worrying 2720 * over. 2721 */ 2722 if ( (bufmallocspace < maxbufmallocspace) && 2723 (bp->b_bufsize == 0) && 2724 (mbsize <= PAGE_SIZE/2)) { 2725 2726 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2727 bp->b_bufsize = mbsize; 2728 bp->b_bcount = size; 2729 bp->b_flags |= B_MALLOC; 2730 atomic_add_int(&bufmallocspace, mbsize); 2731 return 1; 2732 } 2733 origbuf = NULL; 2734 origbufsize = 0; 2735 /* 2736 * If the buffer is growing on its other-than-first allocation, 2737 * then we revert to the page-allocation scheme. 2738 */ 2739 if (bp->b_flags & B_MALLOC) { 2740 origbuf = bp->b_data; 2741 origbufsize = bp->b_bufsize; 2742 bp->b_data = bp->b_kvabase; 2743 if (bp->b_bufsize) { 2744 atomic_subtract_int(&bufmallocspace, 2745 bp->b_bufsize); 2746 bufspacewakeup(); 2747 bp->b_bufsize = 0; 2748 } 2749 bp->b_flags &= ~B_MALLOC; 2750 newbsize = round_page(newbsize); 2751 } 2752 vm_hold_load_pages( 2753 bp, 2754 (vm_offset_t) bp->b_data + bp->b_bufsize, 2755 (vm_offset_t) bp->b_data + newbsize); 2756 if (origbuf) { 2757 bcopy(origbuf, bp->b_data, origbufsize); 2758 free(origbuf, M_BIOBUF); 2759 } 2760 } 2761 } else { 2762 int desiredpages; 2763 2764 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2765 desiredpages = (size == 0) ? 0 : 2766 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2767 2768 if (bp->b_flags & B_MALLOC) 2769 panic("allocbuf: VMIO buffer can't be malloced"); 2770 /* 2771 * Set B_CACHE initially if buffer is 0 length or will become 2772 * 0-length. 2773 */ 2774 if (size == 0 || bp->b_bufsize == 0) 2775 bp->b_flags |= B_CACHE; 2776 2777 if (newbsize < bp->b_bufsize) { 2778 /* 2779 * DEV_BSIZE aligned new buffer size is less then the 2780 * DEV_BSIZE aligned existing buffer size. Figure out 2781 * if we have to remove any pages. 2782 */ 2783 if (desiredpages < bp->b_npages) { 2784 vm_page_t m; 2785 2786 vm_page_lock_queues(); 2787 for (i = desiredpages; i < bp->b_npages; i++) { 2788 /* 2789 * the page is not freed here -- it 2790 * is the responsibility of 2791 * vnode_pager_setsize 2792 */ 2793 m = bp->b_pages[i]; 2794 KASSERT(m != bogus_page, 2795 ("allocbuf: bogus page found")); 2796 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2797 vm_page_lock_queues(); 2798 2799 bp->b_pages[i] = NULL; 2800 vm_page_unwire(m, 0); 2801 } 2802 vm_page_unlock_queues(); 2803 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2804 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2805 bp->b_npages = desiredpages; 2806 } 2807 } else if (size > bp->b_bcount) { 2808 /* 2809 * We are growing the buffer, possibly in a 2810 * byte-granular fashion. 2811 */ 2812 struct vnode *vp; 2813 vm_object_t obj; 2814 vm_offset_t toff; 2815 vm_offset_t tinc; 2816 2817 /* 2818 * Step 1, bring in the VM pages from the object, 2819 * allocating them if necessary. We must clear 2820 * B_CACHE if these pages are not valid for the 2821 * range covered by the buffer. 2822 */ 2823 2824 vp = bp->b_vp; 2825 obj = bp->b_object; 2826 2827 VM_OBJECT_LOCK(obj); 2828 while (bp->b_npages < desiredpages) { 2829 vm_page_t m; 2830 vm_pindex_t pi; 2831 2832 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2833 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2834 /* 2835 * note: must allocate system pages 2836 * since blocking here could intefere 2837 * with paging I/O, no matter which 2838 * process we are. 2839 */ 2840 m = vm_page_alloc(obj, pi, 2841 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2842 if (m == NULL) { 2843 atomic_add_int(&vm_pageout_deficit, 2844 desiredpages - bp->b_npages); 2845 VM_OBJECT_UNLOCK(obj); 2846 VM_WAIT; 2847 VM_OBJECT_LOCK(obj); 2848 } else { 2849 vm_page_lock_queues(); 2850 vm_page_wakeup(m); 2851 vm_page_unlock_queues(); 2852 bp->b_flags &= ~B_CACHE; 2853 bp->b_pages[bp->b_npages] = m; 2854 ++bp->b_npages; 2855 } 2856 continue; 2857 } 2858 2859 /* 2860 * We found a page. If we have to sleep on it, 2861 * retry because it might have gotten freed out 2862 * from under us. 2863 * 2864 * We can only test PG_BUSY here. Blocking on 2865 * m->busy might lead to a deadlock: 2866 * 2867 * vm_fault->getpages->cluster_read->allocbuf 2868 * 2869 */ 2870 vm_page_lock_queues(); 2871 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2872 continue; 2873 2874 /* 2875 * We have a good page. Should we wakeup the 2876 * page daemon? 2877 */ 2878 if ((curproc != pageproc) && 2879 ((m->queue - m->pc) == PQ_CACHE) && 2880 ((cnt.v_free_count + cnt.v_cache_count) < 2881 (cnt.v_free_min + cnt.v_cache_min))) { 2882 pagedaemon_wakeup(); 2883 } 2884 vm_page_flag_clear(m, PG_ZERO); 2885 vm_page_wire(m); 2886 vm_page_unlock_queues(); 2887 bp->b_pages[bp->b_npages] = m; 2888 ++bp->b_npages; 2889 } 2890 VM_OBJECT_UNLOCK(obj); 2891 2892 /* 2893 * Step 2. We've loaded the pages into the buffer, 2894 * we have to figure out if we can still have B_CACHE 2895 * set. Note that B_CACHE is set according to the 2896 * byte-granular range ( bcount and size ), new the 2897 * aligned range ( newbsize ). 2898 * 2899 * The VM test is against m->valid, which is DEV_BSIZE 2900 * aligned. Needless to say, the validity of the data 2901 * needs to also be DEV_BSIZE aligned. Note that this 2902 * fails with NFS if the server or some other client 2903 * extends the file's EOF. If our buffer is resized, 2904 * B_CACHE may remain set! XXX 2905 */ 2906 2907 toff = bp->b_bcount; 2908 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2909 2910 while ((bp->b_flags & B_CACHE) && toff < size) { 2911 vm_pindex_t pi; 2912 2913 if (tinc > (size - toff)) 2914 tinc = size - toff; 2915 2916 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2917 PAGE_SHIFT; 2918 2919 vfs_buf_test_cache( 2920 bp, 2921 bp->b_offset, 2922 toff, 2923 tinc, 2924 bp->b_pages[pi] 2925 ); 2926 toff += tinc; 2927 tinc = PAGE_SIZE; 2928 } 2929 2930 /* 2931 * Step 3, fixup the KVM pmap. Remember that 2932 * bp->b_data is relative to bp->b_offset, but 2933 * bp->b_offset may be offset into the first page. 2934 */ 2935 2936 bp->b_data = (caddr_t) 2937 trunc_page((vm_offset_t)bp->b_data); 2938 pmap_qenter( 2939 (vm_offset_t)bp->b_data, 2940 bp->b_pages, 2941 bp->b_npages 2942 ); 2943 2944 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2945 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2946 } 2947 } 2948 if (newbsize < bp->b_bufsize) 2949 bufspacewakeup(); 2950 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2951 bp->b_bcount = size; /* requested buffer size */ 2952 return 1; 2953} 2954 2955void 2956biodone(struct bio *bp) 2957{ 2958 mtx_lock(&bdonelock); 2959 bp->bio_flags |= BIO_DONE; 2960 if (bp->bio_done == NULL) 2961 wakeup(bp); 2962 mtx_unlock(&bdonelock); 2963 if (bp->bio_done != NULL) 2964 bp->bio_done(bp); 2965} 2966 2967/* 2968 * Wait for a BIO to finish. 2969 * 2970 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2971 * case is not yet clear. 2972 */ 2973int 2974biowait(struct bio *bp, const char *wchan) 2975{ 2976 2977 mtx_lock(&bdonelock); 2978 while ((bp->bio_flags & BIO_DONE) == 0) 2979 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 2980 mtx_unlock(&bdonelock); 2981 if (bp->bio_error != 0) 2982 return (bp->bio_error); 2983 if (!(bp->bio_flags & BIO_ERROR)) 2984 return (0); 2985 return (EIO); 2986} 2987 2988void 2989biofinish(struct bio *bp, struct devstat *stat, int error) 2990{ 2991 2992 if (error) { 2993 bp->bio_error = error; 2994 bp->bio_flags |= BIO_ERROR; 2995 } 2996 if (stat != NULL) 2997 devstat_end_transaction_bio(stat, bp); 2998 biodone(bp); 2999} 3000 3001/* 3002 * bufwait: 3003 * 3004 * Wait for buffer I/O completion, returning error status. The buffer 3005 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3006 * error and cleared. 3007 */ 3008int 3009bufwait(register struct buf * bp) 3010{ 3011 int s; 3012 3013 s = splbio(); 3014 if (bp->b_iocmd == BIO_READ) 3015 bwait(bp, PRIBIO, "biord"); 3016 else 3017 bwait(bp, PRIBIO, "biowr"); 3018 splx(s); 3019 if (bp->b_flags & B_EINTR) { 3020 bp->b_flags &= ~B_EINTR; 3021 return (EINTR); 3022 } 3023 if (bp->b_ioflags & BIO_ERROR) { 3024 return (bp->b_error ? bp->b_error : EIO); 3025 } else { 3026 return (0); 3027 } 3028} 3029 3030 /* 3031 * Call back function from struct bio back up to struct buf. 3032 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 3033 */ 3034void 3035bufdonebio(struct bio *bp) 3036{ 3037 bufdone(bp->bio_caller2); 3038} 3039 3040/* 3041 * bufdone: 3042 * 3043 * Finish I/O on a buffer, optionally calling a completion function. 3044 * This is usually called from an interrupt so process blocking is 3045 * not allowed. 3046 * 3047 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3048 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3049 * assuming B_INVAL is clear. 3050 * 3051 * For the VMIO case, we set B_CACHE if the op was a read and no 3052 * read error occured, or if the op was a write. B_CACHE is never 3053 * set if the buffer is invalid or otherwise uncacheable. 3054 * 3055 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3056 * initiator to leave B_INVAL set to brelse the buffer out of existance 3057 * in the biodone routine. 3058 */ 3059void 3060bufdone(struct buf *bp) 3061{ 3062 int s; 3063 void (*biodone)(struct buf *); 3064 3065 GIANT_REQUIRED; 3066 3067 s = splbio(); 3068 3069 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3070 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3071 3072 bp->b_flags |= B_DONE; 3073 runningbufwakeup(bp); 3074 3075 if (bp->b_iocmd == BIO_DELETE) { 3076 brelse(bp); 3077 splx(s); 3078 return; 3079 } 3080 3081 if (bp->b_iocmd == BIO_WRITE) { 3082 vwakeup(bp); 3083 } 3084 3085 /* call optional completion function if requested */ 3086 if (bp->b_iodone != NULL) { 3087 biodone = bp->b_iodone; 3088 bp->b_iodone = NULL; 3089 (*biodone) (bp); 3090 splx(s); 3091 return; 3092 } 3093 if (LIST_FIRST(&bp->b_dep) != NULL) 3094 buf_complete(bp); 3095 3096 if (bp->b_flags & B_VMIO) { 3097 int i; 3098 vm_ooffset_t foff; 3099 vm_page_t m; 3100 vm_object_t obj; 3101 int iosize; 3102 struct vnode *vp = bp->b_vp; 3103 3104 obj = bp->b_object; 3105 3106#if defined(VFS_BIO_DEBUG) 3107 mp_fixme("usecount and vflag accessed without locks."); 3108 if (vp->v_usecount == 0) { 3109 panic("biodone: zero vnode ref count"); 3110 } 3111 3112 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3113 panic("biodone: vnode is not setup for merged cache"); 3114 } 3115#endif 3116 3117 foff = bp->b_offset; 3118 KASSERT(bp->b_offset != NOOFFSET, 3119 ("biodone: no buffer offset")); 3120 3121 if (obj != NULL) 3122 VM_OBJECT_LOCK(obj); 3123#if defined(VFS_BIO_DEBUG) 3124 if (obj->paging_in_progress < bp->b_npages) { 3125 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3126 obj->paging_in_progress, bp->b_npages); 3127 } 3128#endif 3129 3130 /* 3131 * Set B_CACHE if the op was a normal read and no error 3132 * occured. B_CACHE is set for writes in the b*write() 3133 * routines. 3134 */ 3135 iosize = bp->b_bcount - bp->b_resid; 3136 if (bp->b_iocmd == BIO_READ && 3137 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3138 !(bp->b_ioflags & BIO_ERROR)) { 3139 bp->b_flags |= B_CACHE; 3140 } 3141 vm_page_lock_queues(); 3142 for (i = 0; i < bp->b_npages; i++) { 3143 int bogusflag = 0; 3144 int resid; 3145 3146 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3147 if (resid > iosize) 3148 resid = iosize; 3149 3150 /* 3151 * cleanup bogus pages, restoring the originals 3152 */ 3153 m = bp->b_pages[i]; 3154 if (m == bogus_page) { 3155 bogusflag = 1; 3156 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3157 if (m == NULL) 3158 panic("biodone: page disappeared!"); 3159 bp->b_pages[i] = m; 3160 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3161 } 3162#if defined(VFS_BIO_DEBUG) 3163 if (OFF_TO_IDX(foff) != m->pindex) { 3164 printf( 3165"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3166 (intmax_t)foff, (uintmax_t)m->pindex); 3167 } 3168#endif 3169 3170 /* 3171 * In the write case, the valid and clean bits are 3172 * already changed correctly ( see bdwrite() ), so we 3173 * only need to do this here in the read case. 3174 */ 3175 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3176 vfs_page_set_valid(bp, foff, i, m); 3177 } 3178 vm_page_flag_clear(m, PG_ZERO); 3179 3180 /* 3181 * when debugging new filesystems or buffer I/O methods, this 3182 * is the most common error that pops up. if you see this, you 3183 * have not set the page busy flag correctly!!! 3184 */ 3185 if (m->busy == 0) { 3186 printf("biodone: page busy < 0, " 3187 "pindex: %d, foff: 0x(%x,%x), " 3188 "resid: %d, index: %d\n", 3189 (int) m->pindex, (int)(foff >> 32), 3190 (int) foff & 0xffffffff, resid, i); 3191 if (!vn_isdisk(vp, NULL)) 3192 printf(" iosize: %ld, lblkno: %jd, flags: 0x%x, npages: %d\n", 3193 bp->b_vp->v_mount->mnt_stat.f_iosize, 3194 (intmax_t) bp->b_lblkno, 3195 bp->b_flags, bp->b_npages); 3196 else 3197 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3198 (intmax_t) bp->b_lblkno, 3199 bp->b_flags, bp->b_npages); 3200 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 3201 m->valid, m->dirty, m->wire_count); 3202 panic("biodone: page busy < 0\n"); 3203 } 3204 vm_page_io_finish(m); 3205 vm_object_pip_subtract(obj, 1); 3206 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3207 iosize -= resid; 3208 } 3209 vm_page_unlock_queues(); 3210 if (obj != NULL) { 3211 vm_object_pip_wakeupn(obj, 0); 3212 VM_OBJECT_UNLOCK(obj); 3213 } 3214 } 3215 3216 /* 3217 * For asynchronous completions, release the buffer now. The brelse 3218 * will do a wakeup there if necessary - so no need to do a wakeup 3219 * here in the async case. The sync case always needs to do a wakeup. 3220 */ 3221 3222 if (bp->b_flags & B_ASYNC) { 3223 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3224 brelse(bp); 3225 else 3226 bqrelse(bp); 3227 } else { 3228 bdone(bp); 3229 } 3230 splx(s); 3231} 3232 3233/* 3234 * This routine is called in lieu of iodone in the case of 3235 * incomplete I/O. This keeps the busy status for pages 3236 * consistant. 3237 */ 3238void 3239vfs_unbusy_pages(struct buf * bp) 3240{ 3241 int i; 3242 3243 GIANT_REQUIRED; 3244 3245 runningbufwakeup(bp); 3246 if (bp->b_flags & B_VMIO) { 3247 vm_object_t obj; 3248 3249 obj = bp->b_object; 3250 VM_OBJECT_LOCK(obj); 3251 vm_page_lock_queues(); 3252 for (i = 0; i < bp->b_npages; i++) { 3253 vm_page_t m = bp->b_pages[i]; 3254 3255 if (m == bogus_page) { 3256 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3257 if (!m) { 3258 panic("vfs_unbusy_pages: page missing\n"); 3259 } 3260 bp->b_pages[i] = m; 3261 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3262 } 3263 vm_object_pip_subtract(obj, 1); 3264 vm_page_flag_clear(m, PG_ZERO); 3265 vm_page_io_finish(m); 3266 } 3267 vm_page_unlock_queues(); 3268 vm_object_pip_wakeupn(obj, 0); 3269 VM_OBJECT_UNLOCK(obj); 3270 } 3271} 3272 3273/* 3274 * vfs_page_set_valid: 3275 * 3276 * Set the valid bits in a page based on the supplied offset. The 3277 * range is restricted to the buffer's size. 3278 * 3279 * This routine is typically called after a read completes. 3280 */ 3281static void 3282vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3283{ 3284 vm_ooffset_t soff, eoff; 3285 3286 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3287 /* 3288 * Start and end offsets in buffer. eoff - soff may not cross a 3289 * page boundry or cross the end of the buffer. The end of the 3290 * buffer, in this case, is our file EOF, not the allocation size 3291 * of the buffer. 3292 */ 3293 soff = off; 3294 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3295 if (eoff > bp->b_offset + bp->b_bcount) 3296 eoff = bp->b_offset + bp->b_bcount; 3297 3298 /* 3299 * Set valid range. This is typically the entire buffer and thus the 3300 * entire page. 3301 */ 3302 if (eoff > soff) { 3303 vm_page_set_validclean( 3304 m, 3305 (vm_offset_t) (soff & PAGE_MASK), 3306 (vm_offset_t) (eoff - soff) 3307 ); 3308 } 3309} 3310 3311/* 3312 * This routine is called before a device strategy routine. 3313 * It is used to tell the VM system that paging I/O is in 3314 * progress, and treat the pages associated with the buffer 3315 * almost as being PG_BUSY. Also the object paging_in_progress 3316 * flag is handled to make sure that the object doesn't become 3317 * inconsistant. 3318 * 3319 * Since I/O has not been initiated yet, certain buffer flags 3320 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3321 * and should be ignored. 3322 */ 3323void 3324vfs_busy_pages(struct buf * bp, int clear_modify) 3325{ 3326 int i, bogus; 3327 3328 if (bp->b_flags & B_VMIO) { 3329 vm_object_t obj; 3330 vm_ooffset_t foff; 3331 3332 obj = bp->b_object; 3333 foff = bp->b_offset; 3334 KASSERT(bp->b_offset != NOOFFSET, 3335 ("vfs_busy_pages: no buffer offset")); 3336 vfs_setdirty(bp); 3337 if (obj != NULL) 3338 VM_OBJECT_LOCK(obj); 3339retry: 3340 vm_page_lock_queues(); 3341 for (i = 0; i < bp->b_npages; i++) { 3342 vm_page_t m = bp->b_pages[i]; 3343 3344 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3345 goto retry; 3346 } 3347 bogus = 0; 3348 for (i = 0; i < bp->b_npages; i++) { 3349 vm_page_t m = bp->b_pages[i]; 3350 3351 vm_page_flag_clear(m, PG_ZERO); 3352 if ((bp->b_flags & B_CLUSTER) == 0) { 3353 vm_object_pip_add(obj, 1); 3354 vm_page_io_start(m); 3355 } 3356 /* 3357 * When readying a buffer for a read ( i.e 3358 * clear_modify == 0 ), it is important to do 3359 * bogus_page replacement for valid pages in 3360 * partially instantiated buffers. Partially 3361 * instantiated buffers can, in turn, occur when 3362 * reconstituting a buffer from its VM backing store 3363 * base. We only have to do this if B_CACHE is 3364 * clear ( which causes the I/O to occur in the 3365 * first place ). The replacement prevents the read 3366 * I/O from overwriting potentially dirty VM-backed 3367 * pages. XXX bogus page replacement is, uh, bogus. 3368 * It may not work properly with small-block devices. 3369 * We need to find a better way. 3370 */ 3371 pmap_remove_all(m); 3372 if (clear_modify) 3373 vfs_page_set_valid(bp, foff, i, m); 3374 else if (m->valid == VM_PAGE_BITS_ALL && 3375 (bp->b_flags & B_CACHE) == 0) { 3376 bp->b_pages[i] = bogus_page; 3377 bogus++; 3378 } 3379 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3380 } 3381 vm_page_unlock_queues(); 3382 if (obj != NULL) 3383 VM_OBJECT_UNLOCK(obj); 3384 if (bogus) 3385 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3386 } 3387} 3388 3389/* 3390 * Tell the VM system that the pages associated with this buffer 3391 * are clean. This is used for delayed writes where the data is 3392 * going to go to disk eventually without additional VM intevention. 3393 * 3394 * Note that while we only really need to clean through to b_bcount, we 3395 * just go ahead and clean through to b_bufsize. 3396 */ 3397static void 3398vfs_clean_pages(struct buf * bp) 3399{ 3400 int i; 3401 3402 GIANT_REQUIRED; 3403 3404 if (bp->b_flags & B_VMIO) { 3405 vm_ooffset_t foff; 3406 3407 foff = bp->b_offset; 3408 KASSERT(bp->b_offset != NOOFFSET, 3409 ("vfs_clean_pages: no buffer offset")); 3410 vm_page_lock_queues(); 3411 for (i = 0; i < bp->b_npages; i++) { 3412 vm_page_t m = bp->b_pages[i]; 3413 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3414 vm_ooffset_t eoff = noff; 3415 3416 if (eoff > bp->b_offset + bp->b_bufsize) 3417 eoff = bp->b_offset + bp->b_bufsize; 3418 vfs_page_set_valid(bp, foff, i, m); 3419 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3420 foff = noff; 3421 } 3422 vm_page_unlock_queues(); 3423 } 3424} 3425 3426/* 3427 * vfs_bio_set_validclean: 3428 * 3429 * Set the range within the buffer to valid and clean. The range is 3430 * relative to the beginning of the buffer, b_offset. Note that b_offset 3431 * itself may be offset from the beginning of the first page. 3432 * 3433 */ 3434 3435void 3436vfs_bio_set_validclean(struct buf *bp, int base, int size) 3437{ 3438 if (bp->b_flags & B_VMIO) { 3439 int i; 3440 int n; 3441 3442 /* 3443 * Fixup base to be relative to beginning of first page. 3444 * Set initial n to be the maximum number of bytes in the 3445 * first page that can be validated. 3446 */ 3447 3448 base += (bp->b_offset & PAGE_MASK); 3449 n = PAGE_SIZE - (base & PAGE_MASK); 3450 3451 vm_page_lock_queues(); 3452 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3453 vm_page_t m = bp->b_pages[i]; 3454 3455 if (n > size) 3456 n = size; 3457 3458 vm_page_set_validclean(m, base & PAGE_MASK, n); 3459 base += n; 3460 size -= n; 3461 n = PAGE_SIZE; 3462 } 3463 vm_page_unlock_queues(); 3464 } 3465} 3466 3467/* 3468 * vfs_bio_clrbuf: 3469 * 3470 * clear a buffer. This routine essentially fakes an I/O, so we need 3471 * to clear BIO_ERROR and B_INVAL. 3472 * 3473 * Note that while we only theoretically need to clear through b_bcount, 3474 * we go ahead and clear through b_bufsize. 3475 */ 3476 3477void 3478vfs_bio_clrbuf(struct buf *bp) 3479{ 3480 int i, mask = 0; 3481 caddr_t sa, ea; 3482 3483 GIANT_REQUIRED; 3484 3485 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3486 bp->b_flags &= ~B_INVAL; 3487 bp->b_ioflags &= ~BIO_ERROR; 3488 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3489 (bp->b_offset & PAGE_MASK) == 0) { 3490 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3491 if ((bp->b_pages[0]->valid & mask) == mask) { 3492 bp->b_resid = 0; 3493 return; 3494 } 3495 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3496 ((bp->b_pages[0]->valid & mask) == 0)) { 3497 bzero(bp->b_data, bp->b_bufsize); 3498 bp->b_pages[0]->valid |= mask; 3499 bp->b_resid = 0; 3500 return; 3501 } 3502 } 3503 ea = sa = bp->b_data; 3504 for(i=0;i<bp->b_npages;i++,sa=ea) { 3505 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3506 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3507 ea = (caddr_t)(vm_offset_t)ulmin( 3508 (u_long)(vm_offset_t)ea, 3509 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3510 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3511 if ((bp->b_pages[i]->valid & mask) == mask) 3512 continue; 3513 if ((bp->b_pages[i]->valid & mask) == 0) { 3514 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3515 bzero(sa, ea - sa); 3516 } 3517 } else { 3518 for (; sa < ea; sa += DEV_BSIZE, j++) { 3519 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3520 (bp->b_pages[i]->valid & (1<<j)) == 0) 3521 bzero(sa, DEV_BSIZE); 3522 } 3523 } 3524 bp->b_pages[i]->valid |= mask; 3525 vm_page_lock_queues(); 3526 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3527 vm_page_unlock_queues(); 3528 } 3529 bp->b_resid = 0; 3530 } else { 3531 clrbuf(bp); 3532 } 3533} 3534 3535/* 3536 * vm_hold_load_pages and vm_hold_free_pages get pages into 3537 * a buffers address space. The pages are anonymous and are 3538 * not associated with a file object. 3539 */ 3540static void 3541vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3542{ 3543 vm_offset_t pg; 3544 vm_page_t p; 3545 int index; 3546 3547 GIANT_REQUIRED; 3548 3549 to = round_page(to); 3550 from = round_page(from); 3551 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3552 3553 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3554tryagain: 3555 /* 3556 * note: must allocate system pages since blocking here 3557 * could intefere with paging I/O, no matter which 3558 * process we are. 3559 */ 3560 VM_OBJECT_LOCK(kernel_object); 3561 p = vm_page_alloc(kernel_object, 3562 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3563 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3564 VM_OBJECT_UNLOCK(kernel_object); 3565 if (!p) { 3566 atomic_add_int(&vm_pageout_deficit, 3567 (to - pg) >> PAGE_SHIFT); 3568 VM_WAIT; 3569 goto tryagain; 3570 } 3571 vm_page_lock_queues(); 3572 p->valid = VM_PAGE_BITS_ALL; 3573 vm_page_unlock_queues(); 3574 pmap_qenter(pg, &p, 1); 3575 bp->b_pages[index] = p; 3576 vm_page_lock_queues(); 3577 vm_page_wakeup(p); 3578 vm_page_unlock_queues(); 3579 } 3580 bp->b_npages = index; 3581} 3582 3583/* Return pages associated with this buf to the vm system */ 3584static void 3585vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3586{ 3587 vm_offset_t pg; 3588 vm_page_t p; 3589 int index, newnpages; 3590 3591 GIANT_REQUIRED; 3592 3593 from = round_page(from); 3594 to = round_page(to); 3595 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3596 3597 if (bp->b_object != NULL) 3598 VM_OBJECT_LOCK(bp->b_object); 3599 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3600 p = bp->b_pages[index]; 3601 if (p && (index < bp->b_npages)) { 3602 if (p->busy) { 3603 printf( 3604 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3605 (intmax_t)bp->b_blkno, 3606 (intmax_t)bp->b_lblkno); 3607 } 3608 bp->b_pages[index] = NULL; 3609 pmap_qremove(pg, 1); 3610 vm_page_lock_queues(); 3611 vm_page_busy(p); 3612 vm_page_unwire(p, 0); 3613 vm_page_free(p); 3614 vm_page_unlock_queues(); 3615 } 3616 } 3617 if (bp->b_object != NULL) 3618 VM_OBJECT_UNLOCK(bp->b_object); 3619 bp->b_npages = newnpages; 3620} 3621 3622/* 3623 * Map an IO request into kernel virtual address space. 3624 * 3625 * All requests are (re)mapped into kernel VA space. 3626 * Notice that we use b_bufsize for the size of the buffer 3627 * to be mapped. b_bcount might be modified by the driver. 3628 * 3629 * Note that even if the caller determines that the address space should 3630 * be valid, a race or a smaller-file mapped into a larger space may 3631 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3632 * check the return value. 3633 */ 3634int 3635vmapbuf(struct buf *bp) 3636{ 3637 caddr_t addr, kva; 3638 vm_paddr_t pa; 3639 int pidx, i; 3640 struct vm_page *m; 3641 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3642 3643 GIANT_REQUIRED; 3644 3645 if ((bp->b_flags & B_PHYS) == 0) 3646 panic("vmapbuf"); 3647 if (bp->b_bufsize < 0) 3648 return (-1); 3649 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3650 addr < bp->b_data + bp->b_bufsize; 3651 addr += PAGE_SIZE, pidx++) { 3652 /* 3653 * Do the vm_fault if needed; do the copy-on-write thing 3654 * when reading stuff off device into memory. 3655 * 3656 * NOTE! Must use pmap_extract() because addr may be in 3657 * the userland address space, and kextract is only guarenteed 3658 * to work for the kernland address space (see: sparc64 port). 3659 */ 3660retry: 3661 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data, 3662 (bp->b_iocmd == BIO_READ) ? 3663 (VM_PROT_READ|VM_PROT_WRITE) : VM_PROT_READ); 3664 if (i < 0) { 3665 vm_page_lock_queues(); 3666 for (i = 0; i < pidx; ++i) { 3667 vm_page_unhold(bp->b_pages[i]); 3668 bp->b_pages[i] = NULL; 3669 } 3670 vm_page_unlock_queues(); 3671 return(-1); 3672 } 3673 pa = pmap_extract(pmap, (vm_offset_t)addr); 3674 if (pa == 0) { 3675 printf("vmapbuf: warning, race against user address during I/O"); 3676 goto retry; 3677 } 3678 m = PHYS_TO_VM_PAGE(pa); 3679 vm_page_lock_queues(); 3680 vm_page_hold(m); 3681 vm_page_unlock_queues(); 3682 bp->b_pages[pidx] = m; 3683 } 3684 if (pidx > btoc(MAXPHYS)) 3685 panic("vmapbuf: mapped more than MAXPHYS"); 3686 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3687 3688 kva = bp->b_saveaddr; 3689 bp->b_npages = pidx; 3690 bp->b_saveaddr = bp->b_data; 3691 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3692 return(0); 3693} 3694 3695/* 3696 * Free the io map PTEs associated with this IO operation. 3697 * We also invalidate the TLB entries and restore the original b_addr. 3698 */ 3699void 3700vunmapbuf(struct buf *bp) 3701{ 3702 int pidx; 3703 int npages; 3704 3705 GIANT_REQUIRED; 3706 3707 if ((bp->b_flags & B_PHYS) == 0) 3708 panic("vunmapbuf"); 3709 3710 npages = bp->b_npages; 3711 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3712 npages); 3713 vm_page_lock_queues(); 3714 for (pidx = 0; pidx < npages; pidx++) 3715 vm_page_unhold(bp->b_pages[pidx]); 3716 vm_page_unlock_queues(); 3717 3718 bp->b_data = bp->b_saveaddr; 3719} 3720 3721void 3722bdone(struct buf *bp) 3723{ 3724 mtx_lock(&bdonelock); 3725 bp->b_flags |= B_DONE; 3726 wakeup(bp); 3727 mtx_unlock(&bdonelock); 3728} 3729 3730void 3731bwait(struct buf *bp, u_char pri, const char *wchan) 3732{ 3733 mtx_lock(&bdonelock); 3734 while ((bp->b_flags & B_DONE) == 0) 3735 msleep(bp, &bdonelock, pri, wchan, 0); 3736 mtx_unlock(&bdonelock); 3737} 3738 3739#include "opt_ddb.h" 3740#ifdef DDB 3741#include <ddb/ddb.h> 3742 3743/* DDB command to show buffer data */ 3744DB_SHOW_COMMAND(buffer, db_show_buffer) 3745{ 3746 /* get args */ 3747 struct buf *bp = (struct buf *)addr; 3748 3749 if (!have_addr) { 3750 db_printf("usage: show buffer <addr>\n"); 3751 return; 3752 } 3753 3754 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3755 db_printf( 3756 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3757 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd, b_pblkno = %jd\n", 3758 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3759 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3760 (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno); 3761 if (bp->b_npages) { 3762 int i; 3763 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3764 for (i = 0; i < bp->b_npages; i++) { 3765 vm_page_t m; 3766 m = bp->b_pages[i]; 3767 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3768 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3769 if ((i + 1) < bp->b_npages) 3770 db_printf(","); 3771 } 3772 db_printf("\n"); 3773 } 3774} 3775#endif /* DDB */ 3776