vfs_bio.c revision 118337
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 118337 2003-08-02 05:05:34Z alc $"); 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 int i; 497 498 GIANT_REQUIRED; 499 500 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 501 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 502 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 503 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 504 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 505 506 /* next, make a null set of free lists */ 507 for (i = 0; i < BUFFER_QUEUES; i++) 508 TAILQ_INIT(&bufqueues[i]); 509 510 /* finally, initialize each buffer header and stick on empty q */ 511 for (i = 0; i < nbuf; i++) { 512 bp = &buf[i]; 513 bzero(bp, sizeof *bp); 514 bp->b_flags = B_INVAL; /* we're just an empty header */ 515 bp->b_dev = NODEV; 516 bp->b_rcred = NOCRED; 517 bp->b_wcred = NOCRED; 518 bp->b_qindex = QUEUE_EMPTY; 519 bp->b_vflags = 0; 520 bp->b_xflags = 0; 521 LIST_INIT(&bp->b_dep); 522 BUF_LOCKINIT(bp); 523 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 524 } 525 526 /* 527 * maxbufspace is the absolute maximum amount of buffer space we are 528 * allowed to reserve in KVM and in real terms. The absolute maximum 529 * is nominally used by buf_daemon. hibufspace is the nominal maximum 530 * used by most other processes. The differential is required to 531 * ensure that buf_daemon is able to run when other processes might 532 * be blocked waiting for buffer space. 533 * 534 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 535 * this may result in KVM fragmentation which is not handled optimally 536 * by the system. 537 */ 538 maxbufspace = nbuf * BKVASIZE; 539 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 540 lobufspace = hibufspace - MAXBSIZE; 541 542 lorunningspace = 512 * 1024; 543 hirunningspace = 1024 * 1024; 544 545/* 546 * Limit the amount of malloc memory since it is wired permanently into 547 * the kernel space. Even though this is accounted for in the buffer 548 * allocation, we don't want the malloced region to grow uncontrolled. 549 * The malloc scheme improves memory utilization significantly on average 550 * (small) directories. 551 */ 552 maxbufmallocspace = hibufspace / 20; 553 554/* 555 * Reduce the chance of a deadlock occuring by limiting the number 556 * of delayed-write dirty buffers we allow to stack up. 557 */ 558 hidirtybuffers = nbuf / 4 + 20; 559 dirtybufthresh = hidirtybuffers * 9 / 10; 560 numdirtybuffers = 0; 561/* 562 * To support extreme low-memory systems, make sure hidirtybuffers cannot 563 * eat up all available buffer space. This occurs when our minimum cannot 564 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 565 * BKVASIZE'd (8K) buffers. 566 */ 567 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 568 hidirtybuffers >>= 1; 569 } 570 lodirtybuffers = hidirtybuffers / 2; 571 572/* 573 * Try to keep the number of free buffers in the specified range, 574 * and give special processes (e.g. like buf_daemon) access to an 575 * emergency reserve. 576 */ 577 lofreebuffers = nbuf / 18 + 5; 578 hifreebuffers = 2 * lofreebuffers; 579 numfreebuffers = nbuf; 580 581/* 582 * Maximum number of async ops initiated per buf_daemon loop. This is 583 * somewhat of a hack at the moment, we really need to limit ourselves 584 * based on the number of bytes of I/O in-transit that were initiated 585 * from buf_daemon. 586 */ 587 588 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 589 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 590} 591 592/* 593 * bfreekva() - free the kva allocation for a buffer. 594 * 595 * Must be called at splbio() or higher as this is the only locking for 596 * buffer_map. 597 * 598 * Since this call frees up buffer space, we call bufspacewakeup(). 599 */ 600static void 601bfreekva(struct buf * bp) 602{ 603 GIANT_REQUIRED; 604 605 if (bp->b_kvasize) { 606 atomic_add_int(&buffreekvacnt, 1); 607 atomic_subtract_int(&bufspace, bp->b_kvasize); 608 vm_map_delete(buffer_map, 609 (vm_offset_t) bp->b_kvabase, 610 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 611 ); 612 bp->b_kvasize = 0; 613 bufspacewakeup(); 614 } 615} 616 617/* 618 * bremfree: 619 * 620 * Remove the buffer from the appropriate free list. 621 */ 622void 623bremfree(struct buf * bp) 624{ 625 mtx_lock(&bqlock); 626 bremfreel(bp); 627 mtx_unlock(&bqlock); 628} 629 630void 631bremfreel(struct buf * bp) 632{ 633 int s = splbio(); 634 int old_qindex = bp->b_qindex; 635 636 GIANT_REQUIRED; 637 638 if (bp->b_qindex != QUEUE_NONE) { 639 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 640 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 641 bp->b_qindex = QUEUE_NONE; 642 } else { 643 if (BUF_REFCNT(bp) <= 1) 644 panic("bremfree: removing a buffer not on a queue"); 645 } 646 647 /* 648 * Fixup numfreebuffers count. If the buffer is invalid or not 649 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 650 * the buffer was free and we must decrement numfreebuffers. 651 */ 652 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 653 switch(old_qindex) { 654 case QUEUE_DIRTY: 655 case QUEUE_CLEAN: 656 case QUEUE_EMPTY: 657 case QUEUE_EMPTYKVA: 658 atomic_subtract_int(&numfreebuffers, 1); 659 break; 660 default: 661 break; 662 } 663 } 664 splx(s); 665} 666 667 668/* 669 * Get a buffer with the specified data. Look in the cache first. We 670 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 671 * is set, the buffer is valid and we do not have to do anything ( see 672 * getblk() ). This is really just a special case of breadn(). 673 */ 674int 675bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 676 struct buf ** bpp) 677{ 678 679 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 680} 681 682/* 683 * Operates like bread, but also starts asynchronous I/O on 684 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 685 * to initiating I/O . If B_CACHE is set, the buffer is valid 686 * and we do not have to do anything. 687 */ 688int 689breadn(struct vnode * vp, daddr_t blkno, int size, 690 daddr_t * rablkno, int *rabsize, 691 int cnt, struct ucred * cred, struct buf ** bpp) 692{ 693 struct buf *bp, *rabp; 694 int i; 695 int rv = 0, readwait = 0; 696 697 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 698 699 /* if not found in cache, do some I/O */ 700 if ((bp->b_flags & B_CACHE) == 0) { 701 if (curthread != PCPU_GET(idlethread)) 702 curthread->td_proc->p_stats->p_ru.ru_inblock++; 703 bp->b_iocmd = BIO_READ; 704 bp->b_flags &= ~B_INVAL; 705 bp->b_ioflags &= ~BIO_ERROR; 706 if (bp->b_rcred == NOCRED && cred != NOCRED) 707 bp->b_rcred = crhold(cred); 708 vfs_busy_pages(bp, 0); 709 if (vp->v_type == VCHR) 710 VOP_SPECSTRATEGY(vp, bp); 711 else 712 VOP_STRATEGY(vp, bp); 713 ++readwait; 714 } 715 716 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 717 if (inmem(vp, *rablkno)) 718 continue; 719 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 720 721 if ((rabp->b_flags & B_CACHE) == 0) { 722 if (curthread != PCPU_GET(idlethread)) 723 curthread->td_proc->p_stats->p_ru.ru_inblock++; 724 rabp->b_flags |= B_ASYNC; 725 rabp->b_flags &= ~B_INVAL; 726 rabp->b_ioflags &= ~BIO_ERROR; 727 rabp->b_iocmd = BIO_READ; 728 if (rabp->b_rcred == NOCRED && cred != NOCRED) 729 rabp->b_rcred = crhold(cred); 730 vfs_busy_pages(rabp, 0); 731 BUF_KERNPROC(rabp); 732 if (vp->v_type == VCHR) 733 VOP_SPECSTRATEGY(vp, rabp); 734 else 735 VOP_STRATEGY(vp, rabp); 736 } else { 737 brelse(rabp); 738 } 739 } 740 741 if (readwait) { 742 rv = bufwait(bp); 743 } 744 return (rv); 745} 746 747/* 748 * Write, release buffer on completion. (Done by iodone 749 * if async). Do not bother writing anything if the buffer 750 * is invalid. 751 * 752 * Note that we set B_CACHE here, indicating that buffer is 753 * fully valid and thus cacheable. This is true even of NFS 754 * now so we set it generally. This could be set either here 755 * or in biodone() since the I/O is synchronous. We put it 756 * here. 757 */ 758 759int 760bwrite(struct buf * bp) 761{ 762 int oldflags, s; 763 struct buf *newbp; 764 765 if (bp->b_flags & B_INVAL) { 766 brelse(bp); 767 return (0); 768 } 769 770 oldflags = bp->b_flags; 771 772 if (BUF_REFCNT(bp) == 0) 773 panic("bwrite: buffer is not busy???"); 774 s = splbio(); 775 /* 776 * If a background write is already in progress, delay 777 * writing this block if it is asynchronous. Otherwise 778 * wait for the background write to complete. 779 */ 780 if (bp->b_xflags & BX_BKGRDINPROG) { 781 if (bp->b_flags & B_ASYNC) { 782 splx(s); 783 bdwrite(bp); 784 return (0); 785 } 786 bp->b_xflags |= BX_BKGRDWAIT; 787 tsleep(&bp->b_xflags, PRIBIO, "bwrbg", 0); 788 if (bp->b_xflags & BX_BKGRDINPROG) 789 panic("bwrite: still writing"); 790 } 791 792 /* Mark the buffer clean */ 793 bundirty(bp); 794 795 /* 796 * If this buffer is marked for background writing and we 797 * do not have to wait for it, make a copy and write the 798 * copy so as to leave this buffer ready for further use. 799 * 800 * This optimization eats a lot of memory. If we have a page 801 * or buffer shortfall we can't do it. 802 */ 803 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 804 (bp->b_flags & B_ASYNC) && 805 !vm_page_count_severe() && 806 !buf_dirty_count_severe()) { 807 if (bp->b_iodone != NULL) { 808 printf("bp->b_iodone = %p\n", bp->b_iodone); 809 panic("bwrite: need chained iodone"); 810 } 811 812 /* get a new block */ 813 newbp = geteblk(bp->b_bufsize); 814 815 /* 816 * set it to be identical to the old block. We have to 817 * set b_lblkno and BKGRDMARKER before calling bgetvp() 818 * to avoid confusing the splay tree and gbincore(). 819 */ 820 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 821 newbp->b_lblkno = bp->b_lblkno; 822 newbp->b_xflags |= BX_BKGRDMARKER; 823 /* XXX The BX_ flags need to be protected as well */ 824 VI_LOCK(bp->b_vp); 825 bgetvp(bp->b_vp, newbp); 826 VI_UNLOCK(bp->b_vp); 827 newbp->b_blkno = bp->b_blkno; 828 newbp->b_offset = bp->b_offset; 829 newbp->b_iodone = vfs_backgroundwritedone; 830 newbp->b_flags |= B_ASYNC; 831 newbp->b_flags &= ~B_INVAL; 832 833 /* move over the dependencies */ 834 if (LIST_FIRST(&bp->b_dep) != NULL) 835 buf_movedeps(bp, newbp); 836 837 /* 838 * Initiate write on the copy, release the original to 839 * the B_LOCKED queue so that it cannot go away until 840 * the background write completes. If not locked it could go 841 * away and then be reconstituted while it was being written. 842 * If the reconstituted buffer were written, we could end up 843 * with two background copies being written at the same time. 844 */ 845 bp->b_xflags |= BX_BKGRDINPROG; 846 bp->b_flags |= B_LOCKED; 847 bqrelse(bp); 848 bp = newbp; 849 } 850 851 bp->b_flags &= ~B_DONE; 852 bp->b_ioflags &= ~BIO_ERROR; 853 bp->b_flags |= B_WRITEINPROG | B_CACHE; 854 bp->b_iocmd = BIO_WRITE; 855 856 VI_LOCK(bp->b_vp); 857 bp->b_vp->v_numoutput++; 858 VI_UNLOCK(bp->b_vp); 859 vfs_busy_pages(bp, 1); 860 861 /* 862 * Normal bwrites pipeline writes 863 */ 864 bp->b_runningbufspace = bp->b_bufsize; 865 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 866 867 if (curthread != PCPU_GET(idlethread)) 868 curthread->td_proc->p_stats->p_ru.ru_oublock++; 869 splx(s); 870 if (oldflags & B_ASYNC) 871 BUF_KERNPROC(bp); 872 if (bp->b_vp->v_type == VCHR) 873 VOP_SPECSTRATEGY(bp->b_vp, bp); 874 else 875 VOP_STRATEGY(bp->b_vp, bp); 876 877 if ((oldflags & B_ASYNC) == 0) { 878 int rtval = bufwait(bp); 879 brelse(bp); 880 return (rtval); 881 } else { 882 /* 883 * don't allow the async write to saturate the I/O 884 * system. We will not deadlock here because 885 * we are blocking waiting for I/O that is already in-progress 886 * to complete. 887 */ 888 waitrunningbufspace(); 889 } 890 891 return (0); 892} 893 894/* 895 * Complete a background write started from bwrite. 896 */ 897static void 898vfs_backgroundwritedone(bp) 899 struct buf *bp; 900{ 901 struct buf *origbp; 902 903 /* 904 * Find the original buffer that we are writing. 905 */ 906 VI_LOCK(bp->b_vp); 907 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 908 panic("backgroundwritedone: lost buffer"); 909 VI_UNLOCK(bp->b_vp); 910 /* 911 * Process dependencies then return any unfinished ones. 912 */ 913 if (LIST_FIRST(&bp->b_dep) != NULL) 914 buf_complete(bp); 915 if (LIST_FIRST(&bp->b_dep) != NULL) 916 buf_movedeps(bp, origbp); 917 918 /* XXX Find out if origbp can disappear or get inconsistent */ 919 /* 920 * Clear the BX_BKGRDINPROG flag in the original buffer 921 * and awaken it if it is waiting for the write to complete. 922 * If BX_BKGRDINPROG is not set in the original buffer it must 923 * have been released and re-instantiated - which is not legal. 924 */ 925 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), 926 ("backgroundwritedone: lost buffer2")); 927 origbp->b_xflags &= ~BX_BKGRDINPROG; 928 if (origbp->b_xflags & BX_BKGRDWAIT) { 929 origbp->b_xflags &= ~BX_BKGRDWAIT; 930 wakeup(&origbp->b_xflags); 931 } 932 /* 933 * Clear the B_LOCKED flag and remove it from the locked 934 * queue if it currently resides there. 935 */ 936 origbp->b_flags &= ~B_LOCKED; 937 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) { 938 bremfree(origbp); 939 bqrelse(origbp); 940 } 941 /* 942 * This buffer is marked B_NOCACHE, so when it is released 943 * by biodone, it will be tossed. We mark it with BIO_READ 944 * to avoid biodone doing a second vwakeup. 945 */ 946 bp->b_flags |= B_NOCACHE; 947 bp->b_iocmd = BIO_READ; 948 bp->b_flags &= ~(B_CACHE | B_DONE); 949 bp->b_iodone = 0; 950 bufdone(bp); 951} 952 953/* 954 * Delayed write. (Buffer is marked dirty). Do not bother writing 955 * anything if the buffer is marked invalid. 956 * 957 * Note that since the buffer must be completely valid, we can safely 958 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 959 * biodone() in order to prevent getblk from writing the buffer 960 * out synchronously. 961 */ 962void 963bdwrite(struct buf * bp) 964{ 965 struct thread *td = curthread; 966 struct vnode *vp; 967 struct buf *nbp; 968 969 GIANT_REQUIRED; 970 971 if (BUF_REFCNT(bp) == 0) 972 panic("bdwrite: buffer is not busy"); 973 974 if (bp->b_flags & B_INVAL) { 975 brelse(bp); 976 return; 977 } 978 979 /* 980 * If we have too many dirty buffers, don't create any more. 981 * If we are wildly over our limit, then force a complete 982 * cleanup. Otherwise, just keep the situation from getting 983 * out of control. Note that we have to avoid a recursive 984 * disaster and not try to clean up after our own cleanup! 985 */ 986 vp = bp->b_vp; 987 VI_LOCK(vp); 988 if (td->td_proc->p_flag & P_COWINPROGRESS) { 989 recursiveflushes++; 990 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) { 991 VI_UNLOCK(vp); 992 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 993 VI_LOCK(vp); 994 altbufferflushes++; 995 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) { 996 /* 997 * Try to find a buffer to flush. 998 */ 999 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) { 1000 if ((nbp->b_xflags & BX_BKGRDINPROG) || 1001 buf_countdeps(nbp, 0) || 1002 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1003 continue; 1004 if (bp == nbp) 1005 panic("bdwrite: found ourselves"); 1006 VI_UNLOCK(vp); 1007 if (nbp->b_flags & B_CLUSTEROK) { 1008 vfs_bio_awrite(nbp); 1009 } else { 1010 bremfree(nbp); 1011 bawrite(nbp); 1012 } 1013 VI_LOCK(vp); 1014 dirtybufferflushes++; 1015 break; 1016 } 1017 } 1018 VI_UNLOCK(vp); 1019 1020 bdirty(bp); 1021 /* 1022 * Set B_CACHE, indicating that the buffer is fully valid. This is 1023 * true even of NFS now. 1024 */ 1025 bp->b_flags |= B_CACHE; 1026 1027 /* 1028 * This bmap keeps the system from needing to do the bmap later, 1029 * perhaps when the system is attempting to do a sync. Since it 1030 * is likely that the indirect block -- or whatever other datastructure 1031 * that the filesystem needs is still in memory now, it is a good 1032 * thing to do this. Note also, that if the pageout daemon is 1033 * requesting a sync -- there might not be enough memory to do 1034 * the bmap then... So, this is important to do. 1035 */ 1036 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1037 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1038 } 1039 1040 /* 1041 * Set the *dirty* buffer range based upon the VM system dirty pages. 1042 */ 1043 vfs_setdirty(bp); 1044 1045 /* 1046 * We need to do this here to satisfy the vnode_pager and the 1047 * pageout daemon, so that it thinks that the pages have been 1048 * "cleaned". Note that since the pages are in a delayed write 1049 * buffer -- the VFS layer "will" see that the pages get written 1050 * out on the next sync, or perhaps the cluster will be completed. 1051 */ 1052 vfs_clean_pages(bp); 1053 bqrelse(bp); 1054 1055 /* 1056 * Wakeup the buffer flushing daemon if we have a lot of dirty 1057 * buffers (midpoint between our recovery point and our stall 1058 * point). 1059 */ 1060 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1061 1062 /* 1063 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1064 * due to the softdep code. 1065 */ 1066} 1067 1068/* 1069 * bdirty: 1070 * 1071 * Turn buffer into delayed write request. We must clear BIO_READ and 1072 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1073 * itself to properly update it in the dirty/clean lists. We mark it 1074 * B_DONE to ensure that any asynchronization of the buffer properly 1075 * clears B_DONE ( else a panic will occur later ). 1076 * 1077 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1078 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1079 * should only be called if the buffer is known-good. 1080 * 1081 * Since the buffer is not on a queue, we do not update the numfreebuffers 1082 * count. 1083 * 1084 * Must be called at splbio(). 1085 * The buffer must be on QUEUE_NONE. 1086 */ 1087void 1088bdirty(bp) 1089 struct buf *bp; 1090{ 1091 KASSERT(bp->b_qindex == QUEUE_NONE, 1092 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1093 bp->b_flags &= ~(B_RELBUF); 1094 bp->b_iocmd = BIO_WRITE; 1095 1096 if ((bp->b_flags & B_DELWRI) == 0) { 1097 bp->b_flags |= B_DONE | B_DELWRI; 1098 reassignbuf(bp, bp->b_vp); 1099 atomic_add_int(&numdirtybuffers, 1); 1100 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1101 } 1102} 1103 1104/* 1105 * bundirty: 1106 * 1107 * Clear B_DELWRI for buffer. 1108 * 1109 * Since the buffer is not on a queue, we do not update the numfreebuffers 1110 * count. 1111 * 1112 * Must be called at splbio(). 1113 * The buffer must be on QUEUE_NONE. 1114 */ 1115 1116void 1117bundirty(bp) 1118 struct buf *bp; 1119{ 1120 KASSERT(bp->b_qindex == QUEUE_NONE, 1121 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1122 1123 if (bp->b_flags & B_DELWRI) { 1124 bp->b_flags &= ~B_DELWRI; 1125 reassignbuf(bp, bp->b_vp); 1126 atomic_subtract_int(&numdirtybuffers, 1); 1127 numdirtywakeup(lodirtybuffers); 1128 } 1129 /* 1130 * Since it is now being written, we can clear its deferred write flag. 1131 */ 1132 bp->b_flags &= ~B_DEFERRED; 1133} 1134 1135/* 1136 * bawrite: 1137 * 1138 * Asynchronous write. Start output on a buffer, but do not wait for 1139 * it to complete. The buffer is released when the output completes. 1140 * 1141 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1142 * B_INVAL buffers. Not us. 1143 */ 1144void 1145bawrite(struct buf * bp) 1146{ 1147 bp->b_flags |= B_ASYNC; 1148 (void) BUF_WRITE(bp); 1149} 1150 1151/* 1152 * bwillwrite: 1153 * 1154 * Called prior to the locking of any vnodes when we are expecting to 1155 * write. We do not want to starve the buffer cache with too many 1156 * dirty buffers so we block here. By blocking prior to the locking 1157 * of any vnodes we attempt to avoid the situation where a locked vnode 1158 * prevents the various system daemons from flushing related buffers. 1159 */ 1160 1161void 1162bwillwrite(void) 1163{ 1164 if (numdirtybuffers >= hidirtybuffers) { 1165 int s; 1166 1167 mtx_lock(&Giant); 1168 s = splbio(); 1169 mtx_lock(&nblock); 1170 while (numdirtybuffers >= hidirtybuffers) { 1171 bd_wakeup(1); 1172 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1173 msleep(&needsbuffer, &nblock, 1174 (PRIBIO + 4), "flswai", 0); 1175 } 1176 splx(s); 1177 mtx_unlock(&nblock); 1178 mtx_unlock(&Giant); 1179 } 1180} 1181 1182/* 1183 * Return true if we have too many dirty buffers. 1184 */ 1185int 1186buf_dirty_count_severe(void) 1187{ 1188 return(numdirtybuffers >= hidirtybuffers); 1189} 1190 1191/* 1192 * brelse: 1193 * 1194 * Release a busy buffer and, if requested, free its resources. The 1195 * buffer will be stashed in the appropriate bufqueue[] allowing it 1196 * to be accessed later as a cache entity or reused for other purposes. 1197 */ 1198void 1199brelse(struct buf * bp) 1200{ 1201 int s; 1202 1203 GIANT_REQUIRED; 1204 1205 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1206 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1207 1208 s = splbio(); 1209 1210 if (bp->b_flags & B_LOCKED) 1211 bp->b_ioflags &= ~BIO_ERROR; 1212 1213 if (bp->b_iocmd == BIO_WRITE && 1214 (bp->b_ioflags & BIO_ERROR) && 1215 !(bp->b_flags & B_INVAL)) { 1216 /* 1217 * Failed write, redirty. Must clear BIO_ERROR to prevent 1218 * pages from being scrapped. If B_INVAL is set then 1219 * this case is not run and the next case is run to 1220 * destroy the buffer. B_INVAL can occur if the buffer 1221 * is outside the range supported by the underlying device. 1222 */ 1223 bp->b_ioflags &= ~BIO_ERROR; 1224 bdirty(bp); 1225 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1226 (bp->b_ioflags & BIO_ERROR) || 1227 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1228 /* 1229 * Either a failed I/O or we were asked to free or not 1230 * cache the buffer. 1231 */ 1232 bp->b_flags |= B_INVAL; 1233 if (LIST_FIRST(&bp->b_dep) != NULL) 1234 buf_deallocate(bp); 1235 if (bp->b_flags & B_DELWRI) { 1236 atomic_subtract_int(&numdirtybuffers, 1); 1237 numdirtywakeup(lodirtybuffers); 1238 } 1239 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1240 if ((bp->b_flags & B_VMIO) == 0) { 1241 if (bp->b_bufsize) 1242 allocbuf(bp, 0); 1243 if (bp->b_vp) 1244 brelvp(bp); 1245 } 1246 } 1247 1248 /* 1249 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1250 * is called with B_DELWRI set, the underlying pages may wind up 1251 * getting freed causing a previous write (bdwrite()) to get 'lost' 1252 * because pages associated with a B_DELWRI bp are marked clean. 1253 * 1254 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1255 * if B_DELWRI is set. 1256 * 1257 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1258 * on pages to return pages to the VM page queues. 1259 */ 1260 if (bp->b_flags & B_DELWRI) 1261 bp->b_flags &= ~B_RELBUF; 1262 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1263 bp->b_flags |= B_RELBUF; 1264 1265 /* 1266 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1267 * constituted, not even NFS buffers now. Two flags effect this. If 1268 * B_INVAL, the struct buf is invalidated but the VM object is kept 1269 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1270 * 1271 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1272 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1273 * buffer is also B_INVAL because it hits the re-dirtying code above. 1274 * 1275 * Normally we can do this whether a buffer is B_DELWRI or not. If 1276 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1277 * the commit state and we cannot afford to lose the buffer. If the 1278 * buffer has a background write in progress, we need to keep it 1279 * around to prevent it from being reconstituted and starting a second 1280 * background write. 1281 */ 1282 if ((bp->b_flags & B_VMIO) 1283 && !(bp->b_vp->v_mount != NULL && 1284 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1285 !vn_isdisk(bp->b_vp, NULL) && 1286 (bp->b_flags & B_DELWRI)) 1287 ) { 1288 1289 int i, j, resid; 1290 vm_page_t m; 1291 off_t foff; 1292 vm_pindex_t poff; 1293 vm_object_t obj; 1294 struct vnode *vp; 1295 1296 vp = bp->b_vp; 1297 obj = bp->b_object; 1298 1299 /* 1300 * Get the base offset and length of the buffer. Note that 1301 * in the VMIO case if the buffer block size is not 1302 * page-aligned then b_data pointer may not be page-aligned. 1303 * But our b_pages[] array *IS* page aligned. 1304 * 1305 * block sizes less then DEV_BSIZE (usually 512) are not 1306 * supported due to the page granularity bits (m->valid, 1307 * m->dirty, etc...). 1308 * 1309 * See man buf(9) for more information 1310 */ 1311 resid = bp->b_bufsize; 1312 foff = bp->b_offset; 1313 if (obj != NULL) 1314 VM_OBJECT_LOCK(obj); 1315 for (i = 0; i < bp->b_npages; i++) { 1316 int had_bogus = 0; 1317 1318 m = bp->b_pages[i]; 1319 vm_page_lock_queues(); 1320 vm_page_flag_clear(m, PG_ZERO); 1321 vm_page_unlock_queues(); 1322 1323 /* 1324 * If we hit a bogus page, fixup *all* the bogus pages 1325 * now. 1326 */ 1327 if (m == bogus_page) { 1328 poff = OFF_TO_IDX(bp->b_offset); 1329 had_bogus = 1; 1330 1331 for (j = i; j < bp->b_npages; j++) { 1332 vm_page_t mtmp; 1333 mtmp = bp->b_pages[j]; 1334 if (mtmp == bogus_page) { 1335 mtmp = vm_page_lookup(obj, poff + j); 1336 if (!mtmp) { 1337 panic("brelse: page missing\n"); 1338 } 1339 bp->b_pages[j] = mtmp; 1340 } 1341 } 1342 1343 if ((bp->b_flags & B_INVAL) == 0) { 1344 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1345 } 1346 m = bp->b_pages[i]; 1347 } 1348 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1349 int poffset = foff & PAGE_MASK; 1350 int presid = resid > (PAGE_SIZE - poffset) ? 1351 (PAGE_SIZE - poffset) : resid; 1352 1353 KASSERT(presid >= 0, ("brelse: extra page")); 1354 vm_page_set_invalid(m, poffset, presid); 1355 if (had_bogus) 1356 printf("avoided corruption bug in bogus_page/brelse code\n"); 1357 } 1358 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1359 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1360 } 1361 if (obj != NULL) 1362 VM_OBJECT_UNLOCK(obj); 1363 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1364 vfs_vmio_release(bp); 1365 1366 } else if (bp->b_flags & B_VMIO) { 1367 1368 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1369 vfs_vmio_release(bp); 1370 } 1371 1372 } 1373 1374 if (bp->b_qindex != QUEUE_NONE) 1375 panic("brelse: free buffer onto another queue???"); 1376 if (BUF_REFCNT(bp) > 1) { 1377 /* do not release to free list */ 1378 BUF_UNLOCK(bp); 1379 splx(s); 1380 return; 1381 } 1382 1383 /* enqueue */ 1384 mtx_lock(&bqlock); 1385 1386 /* buffers with no memory */ 1387 if (bp->b_bufsize == 0) { 1388 bp->b_flags |= B_INVAL; 1389 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1390 if (bp->b_xflags & BX_BKGRDINPROG) 1391 panic("losing buffer 1"); 1392 if (bp->b_kvasize) { 1393 bp->b_qindex = QUEUE_EMPTYKVA; 1394 } else { 1395 bp->b_qindex = QUEUE_EMPTY; 1396 } 1397 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1398 bp->b_dev = NODEV; 1399 /* buffers with junk contents */ 1400 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1401 (bp->b_ioflags & BIO_ERROR)) { 1402 bp->b_flags |= B_INVAL; 1403 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1404 if (bp->b_xflags & BX_BKGRDINPROG) 1405 panic("losing buffer 2"); 1406 bp->b_qindex = QUEUE_CLEAN; 1407 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1408 bp->b_dev = NODEV; 1409 1410 /* buffers that are locked */ 1411 } else if (bp->b_flags & B_LOCKED) { 1412 bp->b_qindex = QUEUE_LOCKED; 1413 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1414 1415 /* remaining buffers */ 1416 } else { 1417 if (bp->b_flags & B_DELWRI) 1418 bp->b_qindex = QUEUE_DIRTY; 1419 else 1420 bp->b_qindex = QUEUE_CLEAN; 1421 if (bp->b_flags & B_AGE) 1422 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1423 else 1424 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1425 } 1426 mtx_unlock(&bqlock); 1427 1428 /* 1429 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1430 * placed the buffer on the correct queue. We must also disassociate 1431 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1432 * find it. 1433 */ 1434 if (bp->b_flags & B_INVAL) { 1435 if (bp->b_flags & B_DELWRI) 1436 bundirty(bp); 1437 if (bp->b_vp) 1438 brelvp(bp); 1439 } 1440 1441 /* 1442 * Fixup numfreebuffers count. The bp is on an appropriate queue 1443 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1444 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1445 * if B_INVAL is set ). 1446 */ 1447 1448 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1449 bufcountwakeup(); 1450 1451 /* 1452 * Something we can maybe free or reuse 1453 */ 1454 if (bp->b_bufsize || bp->b_kvasize) 1455 bufspacewakeup(); 1456 1457 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1458 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1459 panic("brelse: not dirty"); 1460 /* unlock */ 1461 BUF_UNLOCK(bp); 1462 splx(s); 1463} 1464 1465/* 1466 * Release a buffer back to the appropriate queue but do not try to free 1467 * it. The buffer is expected to be used again soon. 1468 * 1469 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1470 * biodone() to requeue an async I/O on completion. It is also used when 1471 * known good buffers need to be requeued but we think we may need the data 1472 * again soon. 1473 * 1474 * XXX we should be able to leave the B_RELBUF hint set on completion. 1475 */ 1476void 1477bqrelse(struct buf * bp) 1478{ 1479 int s; 1480 1481 s = splbio(); 1482 1483 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1484 1485 if (bp->b_qindex != QUEUE_NONE) 1486 panic("bqrelse: free buffer onto another queue???"); 1487 if (BUF_REFCNT(bp) > 1) { 1488 /* do not release to free list */ 1489 BUF_UNLOCK(bp); 1490 splx(s); 1491 return; 1492 } 1493 mtx_lock(&bqlock); 1494 if (bp->b_flags & B_LOCKED) { 1495 bp->b_ioflags &= ~BIO_ERROR; 1496 bp->b_qindex = QUEUE_LOCKED; 1497 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1498 /* buffers with stale but valid contents */ 1499 } else if (bp->b_flags & B_DELWRI) { 1500 bp->b_qindex = QUEUE_DIRTY; 1501 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1502 } else if (vm_page_count_severe()) { 1503 /* 1504 * We are too low on memory, we have to try to free the 1505 * buffer (most importantly: the wired pages making up its 1506 * backing store) *now*. 1507 */ 1508 mtx_unlock(&bqlock); 1509 splx(s); 1510 brelse(bp); 1511 return; 1512 } else { 1513 bp->b_qindex = QUEUE_CLEAN; 1514 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1515 } 1516 mtx_unlock(&bqlock); 1517 1518 if ((bp->b_flags & B_LOCKED) == 0 && 1519 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1520 bufcountwakeup(); 1521 } 1522 1523 /* 1524 * Something we can maybe free or reuse. 1525 */ 1526 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1527 bufspacewakeup(); 1528 1529 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1530 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1531 panic("bqrelse: not dirty"); 1532 /* unlock */ 1533 BUF_UNLOCK(bp); 1534 splx(s); 1535} 1536 1537/* Give pages used by the bp back to the VM system (where possible) */ 1538static void 1539vfs_vmio_release(bp) 1540 struct buf *bp; 1541{ 1542 int i; 1543 vm_page_t m; 1544 1545 GIANT_REQUIRED; 1546 if (bp->b_object != NULL) 1547 VM_OBJECT_LOCK(bp->b_object); 1548 vm_page_lock_queues(); 1549 for (i = 0; i < bp->b_npages; i++) { 1550 m = bp->b_pages[i]; 1551 bp->b_pages[i] = NULL; 1552 /* 1553 * In order to keep page LRU ordering consistent, put 1554 * everything on the inactive queue. 1555 */ 1556 vm_page_unwire(m, 0); 1557 /* 1558 * We don't mess with busy pages, it is 1559 * the responsibility of the process that 1560 * busied the pages to deal with them. 1561 */ 1562 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1563 continue; 1564 1565 if (m->wire_count == 0) { 1566 vm_page_flag_clear(m, PG_ZERO); 1567 /* 1568 * Might as well free the page if we can and it has 1569 * no valid data. We also free the page if the 1570 * buffer was used for direct I/O 1571 */ 1572 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1573 m->hold_count == 0) { 1574 vm_page_busy(m); 1575 pmap_remove_all(m); 1576 vm_page_free(m); 1577 } else if (bp->b_flags & B_DIRECT) { 1578 vm_page_try_to_free(m); 1579 } else if (vm_page_count_severe()) { 1580 vm_page_try_to_cache(m); 1581 } 1582 } 1583 } 1584 vm_page_unlock_queues(); 1585 if (bp->b_object != NULL) 1586 VM_OBJECT_UNLOCK(bp->b_object); 1587 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1588 1589 if (bp->b_bufsize) { 1590 bufspacewakeup(); 1591 bp->b_bufsize = 0; 1592 } 1593 bp->b_npages = 0; 1594 bp->b_flags &= ~B_VMIO; 1595 if (bp->b_vp) 1596 brelvp(bp); 1597} 1598 1599/* 1600 * Check to see if a block at a particular lbn is available for a clustered 1601 * write. 1602 */ 1603static int 1604vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1605{ 1606 struct buf *bpa; 1607 int match; 1608 1609 match = 0; 1610 1611 /* If the buf isn't in core skip it */ 1612 if ((bpa = gbincore(vp, lblkno)) == NULL) 1613 return (0); 1614 1615 /* If the buf is busy we don't want to wait for it */ 1616 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1617 return (0); 1618 1619 /* Only cluster with valid clusterable delayed write buffers */ 1620 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1621 (B_DELWRI | B_CLUSTEROK)) 1622 goto done; 1623 1624 if (bpa->b_bufsize != size) 1625 goto done; 1626 1627 /* 1628 * Check to see if it is in the expected place on disk and that the 1629 * block has been mapped. 1630 */ 1631 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1632 match = 1; 1633done: 1634 BUF_UNLOCK(bpa); 1635 return (match); 1636} 1637 1638/* 1639 * vfs_bio_awrite: 1640 * 1641 * Implement clustered async writes for clearing out B_DELWRI buffers. 1642 * This is much better then the old way of writing only one buffer at 1643 * a time. Note that we may not be presented with the buffers in the 1644 * correct order, so we search for the cluster in both directions. 1645 */ 1646int 1647vfs_bio_awrite(struct buf * bp) 1648{ 1649 int i; 1650 int j; 1651 daddr_t lblkno = bp->b_lblkno; 1652 struct vnode *vp = bp->b_vp; 1653 int s; 1654 int ncl; 1655 int nwritten; 1656 int size; 1657 int maxcl; 1658 1659 s = splbio(); 1660 /* 1661 * right now we support clustered writing only to regular files. If 1662 * we find a clusterable block we could be in the middle of a cluster 1663 * rather then at the beginning. 1664 */ 1665 if ((vp->v_type == VREG) && 1666 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1667 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1668 1669 size = vp->v_mount->mnt_stat.f_iosize; 1670 maxcl = MAXPHYS / size; 1671 1672 VI_LOCK(vp); 1673 for (i = 1; i < maxcl; i++) 1674 if (vfs_bio_clcheck(vp, size, lblkno + i, 1675 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1676 break; 1677 1678 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1679 if (vfs_bio_clcheck(vp, size, lblkno - j, 1680 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1681 break; 1682 1683 VI_UNLOCK(vp); 1684 --j; 1685 ncl = i + j; 1686 /* 1687 * this is a possible cluster write 1688 */ 1689 if (ncl != 1) { 1690 BUF_UNLOCK(bp); 1691 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1692 splx(s); 1693 return nwritten; 1694 } 1695 } 1696 1697 bremfree(bp); 1698 bp->b_flags |= B_ASYNC; 1699 1700 splx(s); 1701 /* 1702 * default (old) behavior, writing out only one block 1703 * 1704 * XXX returns b_bufsize instead of b_bcount for nwritten? 1705 */ 1706 nwritten = bp->b_bufsize; 1707 (void) BUF_WRITE(bp); 1708 1709 return nwritten; 1710} 1711 1712/* 1713 * getnewbuf: 1714 * 1715 * Find and initialize a new buffer header, freeing up existing buffers 1716 * in the bufqueues as necessary. The new buffer is returned locked. 1717 * 1718 * Important: B_INVAL is not set. If the caller wishes to throw the 1719 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1720 * 1721 * We block if: 1722 * We have insufficient buffer headers 1723 * We have insufficient buffer space 1724 * buffer_map is too fragmented ( space reservation fails ) 1725 * If we have to flush dirty buffers ( but we try to avoid this ) 1726 * 1727 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1728 * Instead we ask the buf daemon to do it for us. We attempt to 1729 * avoid piecemeal wakeups of the pageout daemon. 1730 */ 1731 1732static struct buf * 1733getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1734{ 1735 struct buf *bp; 1736 struct buf *nbp; 1737 int defrag = 0; 1738 int nqindex; 1739 static int flushingbufs; 1740 1741 GIANT_REQUIRED; 1742 1743 /* 1744 * We can't afford to block since we might be holding a vnode lock, 1745 * which may prevent system daemons from running. We deal with 1746 * low-memory situations by proactively returning memory and running 1747 * async I/O rather then sync I/O. 1748 */ 1749 1750 atomic_add_int(&getnewbufcalls, 1); 1751 atomic_subtract_int(&getnewbufrestarts, 1); 1752restart: 1753 atomic_add_int(&getnewbufrestarts, 1); 1754 1755 /* 1756 * Setup for scan. If we do not have enough free buffers, 1757 * we setup a degenerate case that immediately fails. Note 1758 * that if we are specially marked process, we are allowed to 1759 * dip into our reserves. 1760 * 1761 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1762 * 1763 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1764 * However, there are a number of cases (defragging, reusing, ...) 1765 * where we cannot backup. 1766 */ 1767 mtx_lock(&bqlock); 1768 nqindex = QUEUE_EMPTYKVA; 1769 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1770 1771 if (nbp == NULL) { 1772 /* 1773 * If no EMPTYKVA buffers and we are either 1774 * defragging or reusing, locate a CLEAN buffer 1775 * to free or reuse. If bufspace useage is low 1776 * skip this step so we can allocate a new buffer. 1777 */ 1778 if (defrag || bufspace >= lobufspace) { 1779 nqindex = QUEUE_CLEAN; 1780 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1781 } 1782 1783 /* 1784 * If we could not find or were not allowed to reuse a 1785 * CLEAN buffer, check to see if it is ok to use an EMPTY 1786 * buffer. We can only use an EMPTY buffer if allocating 1787 * its KVA would not otherwise run us out of buffer space. 1788 */ 1789 if (nbp == NULL && defrag == 0 && 1790 bufspace + maxsize < hibufspace) { 1791 nqindex = QUEUE_EMPTY; 1792 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1793 } 1794 } 1795 1796 /* 1797 * Run scan, possibly freeing data and/or kva mappings on the fly 1798 * depending. 1799 */ 1800 1801 while ((bp = nbp) != NULL) { 1802 int qindex = nqindex; 1803 1804 /* 1805 * Calculate next bp ( we can only use it if we do not block 1806 * or do other fancy things ). 1807 */ 1808 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1809 switch(qindex) { 1810 case QUEUE_EMPTY: 1811 nqindex = QUEUE_EMPTYKVA; 1812 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1813 break; 1814 /* FALLTHROUGH */ 1815 case QUEUE_EMPTYKVA: 1816 nqindex = QUEUE_CLEAN; 1817 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1818 break; 1819 /* FALLTHROUGH */ 1820 case QUEUE_CLEAN: 1821 /* 1822 * nbp is NULL. 1823 */ 1824 break; 1825 } 1826 } 1827 1828 /* 1829 * Sanity Checks 1830 */ 1831 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1832 1833 /* 1834 * Note: we no longer distinguish between VMIO and non-VMIO 1835 * buffers. 1836 */ 1837 1838 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1839 1840 /* 1841 * If we are defragging then we need a buffer with 1842 * b_kvasize != 0. XXX this situation should no longer 1843 * occur, if defrag is non-zero the buffer's b_kvasize 1844 * should also be non-zero at this point. XXX 1845 */ 1846 if (defrag && bp->b_kvasize == 0) { 1847 printf("Warning: defrag empty buffer %p\n", bp); 1848 continue; 1849 } 1850 1851 /* 1852 * Start freeing the bp. This is somewhat involved. nbp 1853 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1854 */ 1855 1856 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1857 panic("getnewbuf: locked buf"); 1858 bremfreel(bp); 1859 mtx_unlock(&bqlock); 1860 1861 if (qindex == QUEUE_CLEAN) { 1862 if (bp->b_flags & B_VMIO) { 1863 bp->b_flags &= ~B_ASYNC; 1864 vfs_vmio_release(bp); 1865 } 1866 if (bp->b_vp) 1867 brelvp(bp); 1868 } 1869 1870 /* 1871 * NOTE: nbp is now entirely invalid. We can only restart 1872 * the scan from this point on. 1873 * 1874 * Get the rest of the buffer freed up. b_kva* is still 1875 * valid after this operation. 1876 */ 1877 1878 if (bp->b_rcred != NOCRED) { 1879 crfree(bp->b_rcred); 1880 bp->b_rcred = NOCRED; 1881 } 1882 if (bp->b_wcred != NOCRED) { 1883 crfree(bp->b_wcred); 1884 bp->b_wcred = NOCRED; 1885 } 1886 if (LIST_FIRST(&bp->b_dep) != NULL) 1887 buf_deallocate(bp); 1888 if (bp->b_xflags & BX_BKGRDINPROG) 1889 panic("losing buffer 3"); 1890 1891 if (bp->b_bufsize) 1892 allocbuf(bp, 0); 1893 1894 bp->b_flags = 0; 1895 bp->b_ioflags = 0; 1896 bp->b_xflags = 0; 1897 bp->b_vflags = 0; 1898 bp->b_dev = NODEV; 1899 bp->b_vp = NULL; 1900 bp->b_blkno = bp->b_lblkno = 0; 1901 bp->b_offset = NOOFFSET; 1902 bp->b_iodone = 0; 1903 bp->b_error = 0; 1904 bp->b_resid = 0; 1905 bp->b_bcount = 0; 1906 bp->b_npages = 0; 1907 bp->b_dirtyoff = bp->b_dirtyend = 0; 1908 bp->b_magic = B_MAGIC_BIO; 1909 bp->b_op = &buf_ops_bio; 1910 bp->b_object = NULL; 1911 1912 LIST_INIT(&bp->b_dep); 1913 1914 /* 1915 * If we are defragging then free the buffer. 1916 */ 1917 if (defrag) { 1918 bp->b_flags |= B_INVAL; 1919 bfreekva(bp); 1920 brelse(bp); 1921 defrag = 0; 1922 goto restart; 1923 } 1924 1925 /* 1926 * If we are overcomitted then recover the buffer and its 1927 * KVM space. This occurs in rare situations when multiple 1928 * processes are blocked in getnewbuf() or allocbuf(). 1929 */ 1930 if (bufspace >= hibufspace) 1931 flushingbufs = 1; 1932 if (flushingbufs && bp->b_kvasize != 0) { 1933 bp->b_flags |= B_INVAL; 1934 bfreekva(bp); 1935 brelse(bp); 1936 goto restart; 1937 } 1938 if (bufspace < lobufspace) 1939 flushingbufs = 0; 1940 break; 1941 } 1942 1943 /* 1944 * If we exhausted our list, sleep as appropriate. We may have to 1945 * wakeup various daemons and write out some dirty buffers. 1946 * 1947 * Generally we are sleeping due to insufficient buffer space. 1948 */ 1949 1950 if (bp == NULL) { 1951 int flags; 1952 char *waitmsg; 1953 1954 mtx_unlock(&bqlock); 1955 if (defrag) { 1956 flags = VFS_BIO_NEED_BUFSPACE; 1957 waitmsg = "nbufkv"; 1958 } else if (bufspace >= hibufspace) { 1959 waitmsg = "nbufbs"; 1960 flags = VFS_BIO_NEED_BUFSPACE; 1961 } else { 1962 waitmsg = "newbuf"; 1963 flags = VFS_BIO_NEED_ANY; 1964 } 1965 1966 bd_speedup(); /* heeeelp */ 1967 1968 mtx_lock(&nblock); 1969 needsbuffer |= flags; 1970 while (needsbuffer & flags) { 1971 if (msleep(&needsbuffer, &nblock, 1972 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1973 mtx_unlock(&nblock); 1974 return (NULL); 1975 } 1976 } 1977 mtx_unlock(&nblock); 1978 } else { 1979 /* 1980 * We finally have a valid bp. We aren't quite out of the 1981 * woods, we still have to reserve kva space. In order 1982 * to keep fragmentation sane we only allocate kva in 1983 * BKVASIZE chunks. 1984 */ 1985 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1986 1987 if (maxsize != bp->b_kvasize) { 1988 vm_offset_t addr = 0; 1989 1990 bfreekva(bp); 1991 1992 if (vm_map_findspace(buffer_map, 1993 vm_map_min(buffer_map), maxsize, &addr)) { 1994 /* 1995 * Uh oh. Buffer map is to fragmented. We 1996 * must defragment the map. 1997 */ 1998 atomic_add_int(&bufdefragcnt, 1); 1999 defrag = 1; 2000 bp->b_flags |= B_INVAL; 2001 brelse(bp); 2002 goto restart; 2003 } 2004 if (addr) { 2005 vm_map_insert(buffer_map, NULL, 0, 2006 addr, addr + maxsize, 2007 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2008 2009 bp->b_kvabase = (caddr_t) addr; 2010 bp->b_kvasize = maxsize; 2011 atomic_add_int(&bufspace, bp->b_kvasize); 2012 atomic_add_int(&bufreusecnt, 1); 2013 } 2014 } 2015 bp->b_saveaddr = bp->b_kvabase; 2016 bp->b_data = bp->b_saveaddr; 2017 } 2018 return(bp); 2019} 2020 2021/* 2022 * buf_daemon: 2023 * 2024 * buffer flushing daemon. Buffers are normally flushed by the 2025 * update daemon but if it cannot keep up this process starts to 2026 * take the load in an attempt to prevent getnewbuf() from blocking. 2027 */ 2028 2029static struct proc *bufdaemonproc; 2030 2031static struct kproc_desc buf_kp = { 2032 "bufdaemon", 2033 buf_daemon, 2034 &bufdaemonproc 2035}; 2036SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2037 2038static void 2039buf_daemon() 2040{ 2041 int s; 2042 2043 mtx_lock(&Giant); 2044 2045 /* 2046 * This process needs to be suspended prior to shutdown sync. 2047 */ 2048 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2049 SHUTDOWN_PRI_LAST); 2050 2051 /* 2052 * This process is allowed to take the buffer cache to the limit 2053 */ 2054 s = splbio(); 2055 mtx_lock(&bdlock); 2056 2057 for (;;) { 2058 bd_request = 0; 2059 mtx_unlock(&bdlock); 2060 2061 kthread_suspend_check(bufdaemonproc); 2062 2063 /* 2064 * Do the flush. Limit the amount of in-transit I/O we 2065 * allow to build up, otherwise we would completely saturate 2066 * the I/O system. Wakeup any waiting processes before we 2067 * normally would so they can run in parallel with our drain. 2068 */ 2069 while (numdirtybuffers > lodirtybuffers) { 2070 if (flushbufqueues(0) == 0) { 2071 /* 2072 * Could not find any buffers without rollback 2073 * dependencies, so just write the first one 2074 * in the hopes of eventually making progress. 2075 */ 2076 flushbufqueues(1); 2077 break; 2078 } 2079 waitrunningbufspace(); 2080 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2081 } 2082 2083 /* 2084 * Only clear bd_request if we have reached our low water 2085 * mark. The buf_daemon normally waits 1 second and 2086 * then incrementally flushes any dirty buffers that have 2087 * built up, within reason. 2088 * 2089 * If we were unable to hit our low water mark and couldn't 2090 * find any flushable buffers, we sleep half a second. 2091 * Otherwise we loop immediately. 2092 */ 2093 mtx_lock(&bdlock); 2094 if (numdirtybuffers <= lodirtybuffers) { 2095 /* 2096 * We reached our low water mark, reset the 2097 * request and sleep until we are needed again. 2098 * The sleep is just so the suspend code works. 2099 */ 2100 bd_request = 0; 2101 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2102 } else { 2103 /* 2104 * We couldn't find any flushable dirty buffers but 2105 * still have too many dirty buffers, we 2106 * have to sleep and try again. (rare) 2107 */ 2108 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2109 } 2110 } 2111} 2112 2113/* 2114 * flushbufqueues: 2115 * 2116 * Try to flush a buffer in the dirty queue. We must be careful to 2117 * free up B_INVAL buffers instead of write them, which NFS is 2118 * particularly sensitive to. 2119 */ 2120int flushwithdeps = 0; 2121SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2122 0, "Number of buffers flushed with dependecies that require rollbacks"); 2123static int 2124flushbufqueues(int flushdeps) 2125{ 2126 struct thread *td = curthread; 2127 struct vnode *vp; 2128 struct buf *bp; 2129 int hasdeps; 2130 2131 mtx_lock(&bqlock); 2132 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2133 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2134 continue; 2135 KASSERT((bp->b_flags & B_DELWRI), 2136 ("unexpected clean buffer %p", bp)); 2137 if ((bp->b_xflags & BX_BKGRDINPROG) != 0) { 2138 BUF_UNLOCK(bp); 2139 continue; 2140 } 2141 if (bp->b_flags & B_INVAL) { 2142 bremfreel(bp); 2143 mtx_unlock(&bqlock); 2144 brelse(bp); 2145 return (1); 2146 } 2147 2148 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2149 if (flushdeps == 0) { 2150 BUF_UNLOCK(bp); 2151 continue; 2152 } 2153 hasdeps = 1; 2154 } else 2155 hasdeps = 0; 2156 /* 2157 * We must hold the lock on a vnode before writing 2158 * one of its buffers. Otherwise we may confuse, or 2159 * in the case of a snapshot vnode, deadlock the 2160 * system. 2161 * 2162 * The lock order here is the reverse of the normal 2163 * of vnode followed by buf lock. This is ok because 2164 * the NOWAIT will prevent deadlock. 2165 */ 2166 if ((vp = bp->b_vp) == NULL || 2167 vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2168 mtx_unlock(&bqlock); 2169 vfs_bio_awrite(bp); 2170 if (vp != NULL) 2171 VOP_UNLOCK(vp, 0, td); 2172 flushwithdeps += hasdeps; 2173 return (1); 2174 } 2175 BUF_UNLOCK(bp); 2176 } 2177 mtx_unlock(&bqlock); 2178 return (0); 2179} 2180 2181/* 2182 * Check to see if a block is currently memory resident. 2183 */ 2184struct buf * 2185incore(struct vnode * vp, daddr_t blkno) 2186{ 2187 struct buf *bp; 2188 2189 int s = splbio(); 2190 VI_LOCK(vp); 2191 bp = gbincore(vp, blkno); 2192 VI_UNLOCK(vp); 2193 splx(s); 2194 return (bp); 2195} 2196 2197/* 2198 * Returns true if no I/O is needed to access the 2199 * associated VM object. This is like incore except 2200 * it also hunts around in the VM system for the data. 2201 */ 2202 2203int 2204inmem(struct vnode * vp, daddr_t blkno) 2205{ 2206 vm_object_t obj; 2207 vm_offset_t toff, tinc, size; 2208 vm_page_t m; 2209 vm_ooffset_t off; 2210 2211 GIANT_REQUIRED; 2212 ASSERT_VOP_LOCKED(vp, "inmem"); 2213 2214 if (incore(vp, blkno)) 2215 return 1; 2216 if (vp->v_mount == NULL) 2217 return 0; 2218 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2219 return 0; 2220 2221 size = PAGE_SIZE; 2222 if (size > vp->v_mount->mnt_stat.f_iosize) 2223 size = vp->v_mount->mnt_stat.f_iosize; 2224 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2225 2226 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2227 VM_OBJECT_LOCK(obj); 2228 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2229 VM_OBJECT_UNLOCK(obj); 2230 if (!m) 2231 goto notinmem; 2232 tinc = size; 2233 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2234 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2235 if (vm_page_is_valid(m, 2236 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2237 goto notinmem; 2238 } 2239 return 1; 2240 2241notinmem: 2242 return (0); 2243} 2244 2245/* 2246 * vfs_setdirty: 2247 * 2248 * Sets the dirty range for a buffer based on the status of the dirty 2249 * bits in the pages comprising the buffer. 2250 * 2251 * The range is limited to the size of the buffer. 2252 * 2253 * This routine is primarily used by NFS, but is generalized for the 2254 * B_VMIO case. 2255 */ 2256static void 2257vfs_setdirty(struct buf *bp) 2258{ 2259 int i; 2260 vm_object_t object; 2261 2262 GIANT_REQUIRED; 2263 /* 2264 * Degenerate case - empty buffer 2265 */ 2266 2267 if (bp->b_bufsize == 0) 2268 return; 2269 2270 /* 2271 * We qualify the scan for modified pages on whether the 2272 * object has been flushed yet. The OBJ_WRITEABLE flag 2273 * is not cleared simply by protecting pages off. 2274 */ 2275 2276 if ((bp->b_flags & B_VMIO) == 0) 2277 return; 2278 2279 object = bp->b_pages[0]->object; 2280 VM_OBJECT_LOCK(object); 2281 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2282 printf("Warning: object %p writeable but not mightbedirty\n", object); 2283 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2284 printf("Warning: object %p mightbedirty but not writeable\n", object); 2285 2286 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2287 vm_offset_t boffset; 2288 vm_offset_t eoffset; 2289 2290 vm_page_lock_queues(); 2291 /* 2292 * test the pages to see if they have been modified directly 2293 * by users through the VM system. 2294 */ 2295 for (i = 0; i < bp->b_npages; i++) { 2296 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2297 vm_page_test_dirty(bp->b_pages[i]); 2298 } 2299 2300 /* 2301 * Calculate the encompassing dirty range, boffset and eoffset, 2302 * (eoffset - boffset) bytes. 2303 */ 2304 2305 for (i = 0; i < bp->b_npages; i++) { 2306 if (bp->b_pages[i]->dirty) 2307 break; 2308 } 2309 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2310 2311 for (i = bp->b_npages - 1; i >= 0; --i) { 2312 if (bp->b_pages[i]->dirty) { 2313 break; 2314 } 2315 } 2316 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2317 2318 vm_page_unlock_queues(); 2319 /* 2320 * Fit it to the buffer. 2321 */ 2322 2323 if (eoffset > bp->b_bcount) 2324 eoffset = bp->b_bcount; 2325 2326 /* 2327 * If we have a good dirty range, merge with the existing 2328 * dirty range. 2329 */ 2330 2331 if (boffset < eoffset) { 2332 if (bp->b_dirtyoff > boffset) 2333 bp->b_dirtyoff = boffset; 2334 if (bp->b_dirtyend < eoffset) 2335 bp->b_dirtyend = eoffset; 2336 } 2337 } 2338 VM_OBJECT_UNLOCK(object); 2339} 2340 2341/* 2342 * getblk: 2343 * 2344 * Get a block given a specified block and offset into a file/device. 2345 * The buffers B_DONE bit will be cleared on return, making it almost 2346 * ready for an I/O initiation. B_INVAL may or may not be set on 2347 * return. The caller should clear B_INVAL prior to initiating a 2348 * READ. 2349 * 2350 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2351 * an existing buffer. 2352 * 2353 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2354 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2355 * and then cleared based on the backing VM. If the previous buffer is 2356 * non-0-sized but invalid, B_CACHE will be cleared. 2357 * 2358 * If getblk() must create a new buffer, the new buffer is returned with 2359 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2360 * case it is returned with B_INVAL clear and B_CACHE set based on the 2361 * backing VM. 2362 * 2363 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2364 * B_CACHE bit is clear. 2365 * 2366 * What this means, basically, is that the caller should use B_CACHE to 2367 * determine whether the buffer is fully valid or not and should clear 2368 * B_INVAL prior to issuing a read. If the caller intends to validate 2369 * the buffer by loading its data area with something, the caller needs 2370 * to clear B_INVAL. If the caller does this without issuing an I/O, 2371 * the caller should set B_CACHE ( as an optimization ), else the caller 2372 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2373 * a write attempt or if it was a successfull read. If the caller 2374 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2375 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2376 */ 2377struct buf * 2378getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2379 int flags) 2380{ 2381 struct buf *bp; 2382 int s; 2383 int error; 2384 ASSERT_VOP_LOCKED(vp, "getblk"); 2385 2386 if (size > MAXBSIZE) 2387 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2388 2389 s = splbio(); 2390loop: 2391 /* 2392 * Block if we are low on buffers. Certain processes are allowed 2393 * to completely exhaust the buffer cache. 2394 * 2395 * If this check ever becomes a bottleneck it may be better to 2396 * move it into the else, when gbincore() fails. At the moment 2397 * it isn't a problem. 2398 * 2399 * XXX remove if 0 sections (clean this up after its proven) 2400 */ 2401 if (numfreebuffers == 0) { 2402 if (curthread == PCPU_GET(idlethread)) 2403 return NULL; 2404 mtx_lock(&nblock); 2405 needsbuffer |= VFS_BIO_NEED_ANY; 2406 mtx_unlock(&nblock); 2407 } 2408 2409 VI_LOCK(vp); 2410 if ((bp = gbincore(vp, blkno))) { 2411 int lockflags; 2412 /* 2413 * Buffer is in-core. If the buffer is not busy, it must 2414 * be on a queue. 2415 */ 2416 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2417 2418 if (flags & GB_LOCK_NOWAIT) 2419 lockflags |= LK_NOWAIT; 2420 2421 error = BUF_TIMELOCK(bp, lockflags, 2422 VI_MTX(vp), "getblk", slpflag, slptimeo); 2423 2424 /* 2425 * If we slept and got the lock we have to restart in case 2426 * the buffer changed identities. 2427 */ 2428 if (error == ENOLCK) 2429 goto loop; 2430 /* We timed out or were interrupted. */ 2431 else if (error) 2432 return (NULL); 2433 2434 /* 2435 * The buffer is locked. B_CACHE is cleared if the buffer is 2436 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2437 * and for a VMIO buffer B_CACHE is adjusted according to the 2438 * backing VM cache. 2439 */ 2440 if (bp->b_flags & B_INVAL) 2441 bp->b_flags &= ~B_CACHE; 2442 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2443 bp->b_flags |= B_CACHE; 2444 bremfree(bp); 2445 2446 /* 2447 * check for size inconsistancies for non-VMIO case. 2448 */ 2449 2450 if (bp->b_bcount != size) { 2451 if ((bp->b_flags & B_VMIO) == 0 || 2452 (size > bp->b_kvasize)) { 2453 if (bp->b_flags & B_DELWRI) { 2454 bp->b_flags |= B_NOCACHE; 2455 BUF_WRITE(bp); 2456 } else { 2457 if ((bp->b_flags & B_VMIO) && 2458 (LIST_FIRST(&bp->b_dep) == NULL)) { 2459 bp->b_flags |= B_RELBUF; 2460 brelse(bp); 2461 } else { 2462 bp->b_flags |= B_NOCACHE; 2463 BUF_WRITE(bp); 2464 } 2465 } 2466 goto loop; 2467 } 2468 } 2469 2470 /* 2471 * If the size is inconsistant in the VMIO case, we can resize 2472 * the buffer. This might lead to B_CACHE getting set or 2473 * cleared. If the size has not changed, B_CACHE remains 2474 * unchanged from its previous state. 2475 */ 2476 2477 if (bp->b_bcount != size) 2478 allocbuf(bp, size); 2479 2480 KASSERT(bp->b_offset != NOOFFSET, 2481 ("getblk: no buffer offset")); 2482 2483 /* 2484 * A buffer with B_DELWRI set and B_CACHE clear must 2485 * be committed before we can return the buffer in 2486 * order to prevent the caller from issuing a read 2487 * ( due to B_CACHE not being set ) and overwriting 2488 * it. 2489 * 2490 * Most callers, including NFS and FFS, need this to 2491 * operate properly either because they assume they 2492 * can issue a read if B_CACHE is not set, or because 2493 * ( for example ) an uncached B_DELWRI might loop due 2494 * to softupdates re-dirtying the buffer. In the latter 2495 * case, B_CACHE is set after the first write completes, 2496 * preventing further loops. 2497 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2498 * above while extending the buffer, we cannot allow the 2499 * buffer to remain with B_CACHE set after the write 2500 * completes or it will represent a corrupt state. To 2501 * deal with this we set B_NOCACHE to scrap the buffer 2502 * after the write. 2503 * 2504 * We might be able to do something fancy, like setting 2505 * B_CACHE in bwrite() except if B_DELWRI is already set, 2506 * so the below call doesn't set B_CACHE, but that gets real 2507 * confusing. This is much easier. 2508 */ 2509 2510 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2511 bp->b_flags |= B_NOCACHE; 2512 BUF_WRITE(bp); 2513 goto loop; 2514 } 2515 2516 splx(s); 2517 bp->b_flags &= ~B_DONE; 2518 } else { 2519 int bsize, maxsize, vmio; 2520 off_t offset; 2521 2522 /* 2523 * Buffer is not in-core, create new buffer. The buffer 2524 * returned by getnewbuf() is locked. Note that the returned 2525 * buffer is also considered valid (not marked B_INVAL). 2526 */ 2527 VI_UNLOCK(vp); 2528 if (vn_isdisk(vp, NULL)) 2529 bsize = DEV_BSIZE; 2530 else if (vp->v_mountedhere) 2531 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2532 else if (vp->v_mount) 2533 bsize = vp->v_mount->mnt_stat.f_iosize; 2534 else 2535 bsize = size; 2536 2537 offset = blkno * bsize; 2538 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2539 (vp->v_vflag & VV_OBJBUF); 2540 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2541 maxsize = imax(maxsize, bsize); 2542 2543 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2544 if (slpflag || slptimeo) { 2545 splx(s); 2546 return NULL; 2547 } 2548 goto loop; 2549 } 2550 2551 /* 2552 * This code is used to make sure that a buffer is not 2553 * created while the getnewbuf routine is blocked. 2554 * This can be a problem whether the vnode is locked or not. 2555 * If the buffer is created out from under us, we have to 2556 * throw away the one we just created. There is now window 2557 * race because we are safely running at splbio() from the 2558 * point of the duplicate buffer creation through to here, 2559 * and we've locked the buffer. 2560 * 2561 * Note: this must occur before we associate the buffer 2562 * with the vp especially considering limitations in 2563 * the splay tree implementation when dealing with duplicate 2564 * lblkno's. 2565 */ 2566 VI_LOCK(vp); 2567 if (gbincore(vp, blkno)) { 2568 VI_UNLOCK(vp); 2569 bp->b_flags |= B_INVAL; 2570 brelse(bp); 2571 goto loop; 2572 } 2573 2574 /* 2575 * Insert the buffer into the hash, so that it can 2576 * be found by incore. 2577 */ 2578 bp->b_blkno = bp->b_lblkno = blkno; 2579 bp->b_offset = offset; 2580 2581 bgetvp(vp, bp); 2582 VI_UNLOCK(vp); 2583 2584 /* 2585 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2586 * buffer size starts out as 0, B_CACHE will be set by 2587 * allocbuf() for the VMIO case prior to it testing the 2588 * backing store for validity. 2589 */ 2590 2591 if (vmio) { 2592 bp->b_flags |= B_VMIO; 2593#if defined(VFS_BIO_DEBUG) 2594 if (vp->v_type != VREG) 2595 printf("getblk: vmioing file type %d???\n", vp->v_type); 2596#endif 2597 VOP_GETVOBJECT(vp, &bp->b_object); 2598 } else { 2599 bp->b_flags &= ~B_VMIO; 2600 bp->b_object = NULL; 2601 } 2602 2603 allocbuf(bp, size); 2604 2605 splx(s); 2606 bp->b_flags &= ~B_DONE; 2607 } 2608 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2609 return (bp); 2610} 2611 2612/* 2613 * Get an empty, disassociated buffer of given size. The buffer is initially 2614 * set to B_INVAL. 2615 */ 2616struct buf * 2617geteblk(int size) 2618{ 2619 struct buf *bp; 2620 int s; 2621 int maxsize; 2622 2623 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2624 2625 s = splbio(); 2626 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2627 continue; 2628 splx(s); 2629 allocbuf(bp, size); 2630 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2631 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2632 return (bp); 2633} 2634 2635 2636/* 2637 * This code constitutes the buffer memory from either anonymous system 2638 * memory (in the case of non-VMIO operations) or from an associated 2639 * VM object (in the case of VMIO operations). This code is able to 2640 * resize a buffer up or down. 2641 * 2642 * Note that this code is tricky, and has many complications to resolve 2643 * deadlock or inconsistant data situations. Tread lightly!!! 2644 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2645 * the caller. Calling this code willy nilly can result in the loss of data. 2646 * 2647 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2648 * B_CACHE for the non-VMIO case. 2649 */ 2650 2651int 2652allocbuf(struct buf *bp, int size) 2653{ 2654 int newbsize, mbsize; 2655 int i; 2656 2657 GIANT_REQUIRED; 2658 2659 if (BUF_REFCNT(bp) == 0) 2660 panic("allocbuf: buffer not busy"); 2661 2662 if (bp->b_kvasize < size) 2663 panic("allocbuf: buffer too small"); 2664 2665 if ((bp->b_flags & B_VMIO) == 0) { 2666 caddr_t origbuf; 2667 int origbufsize; 2668 /* 2669 * Just get anonymous memory from the kernel. Don't 2670 * mess with B_CACHE. 2671 */ 2672 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2673 if (bp->b_flags & B_MALLOC) 2674 newbsize = mbsize; 2675 else 2676 newbsize = round_page(size); 2677 2678 if (newbsize < bp->b_bufsize) { 2679 /* 2680 * malloced buffers are not shrunk 2681 */ 2682 if (bp->b_flags & B_MALLOC) { 2683 if (newbsize) { 2684 bp->b_bcount = size; 2685 } else { 2686 free(bp->b_data, M_BIOBUF); 2687 if (bp->b_bufsize) { 2688 atomic_subtract_int( 2689 &bufmallocspace, 2690 bp->b_bufsize); 2691 bufspacewakeup(); 2692 bp->b_bufsize = 0; 2693 } 2694 bp->b_saveaddr = bp->b_kvabase; 2695 bp->b_data = bp->b_saveaddr; 2696 bp->b_bcount = 0; 2697 bp->b_flags &= ~B_MALLOC; 2698 } 2699 return 1; 2700 } 2701 vm_hold_free_pages( 2702 bp, 2703 (vm_offset_t) bp->b_data + newbsize, 2704 (vm_offset_t) bp->b_data + bp->b_bufsize); 2705 } else if (newbsize > bp->b_bufsize) { 2706 /* 2707 * We only use malloced memory on the first allocation. 2708 * and revert to page-allocated memory when the buffer 2709 * grows. 2710 */ 2711 /* 2712 * There is a potential smp race here that could lead 2713 * to bufmallocspace slightly passing the max. It 2714 * is probably extremely rare and not worth worrying 2715 * over. 2716 */ 2717 if ( (bufmallocspace < maxbufmallocspace) && 2718 (bp->b_bufsize == 0) && 2719 (mbsize <= PAGE_SIZE/2)) { 2720 2721 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2722 bp->b_bufsize = mbsize; 2723 bp->b_bcount = size; 2724 bp->b_flags |= B_MALLOC; 2725 atomic_add_int(&bufmallocspace, mbsize); 2726 return 1; 2727 } 2728 origbuf = NULL; 2729 origbufsize = 0; 2730 /* 2731 * If the buffer is growing on its other-than-first allocation, 2732 * then we revert to the page-allocation scheme. 2733 */ 2734 if (bp->b_flags & B_MALLOC) { 2735 origbuf = bp->b_data; 2736 origbufsize = bp->b_bufsize; 2737 bp->b_data = bp->b_kvabase; 2738 if (bp->b_bufsize) { 2739 atomic_subtract_int(&bufmallocspace, 2740 bp->b_bufsize); 2741 bufspacewakeup(); 2742 bp->b_bufsize = 0; 2743 } 2744 bp->b_flags &= ~B_MALLOC; 2745 newbsize = round_page(newbsize); 2746 } 2747 vm_hold_load_pages( 2748 bp, 2749 (vm_offset_t) bp->b_data + bp->b_bufsize, 2750 (vm_offset_t) bp->b_data + newbsize); 2751 if (origbuf) { 2752 bcopy(origbuf, bp->b_data, origbufsize); 2753 free(origbuf, M_BIOBUF); 2754 } 2755 } 2756 } else { 2757 int desiredpages; 2758 2759 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2760 desiredpages = (size == 0) ? 0 : 2761 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2762 2763 if (bp->b_flags & B_MALLOC) 2764 panic("allocbuf: VMIO buffer can't be malloced"); 2765 /* 2766 * Set B_CACHE initially if buffer is 0 length or will become 2767 * 0-length. 2768 */ 2769 if (size == 0 || bp->b_bufsize == 0) 2770 bp->b_flags |= B_CACHE; 2771 2772 if (newbsize < bp->b_bufsize) { 2773 /* 2774 * DEV_BSIZE aligned new buffer size is less then the 2775 * DEV_BSIZE aligned existing buffer size. Figure out 2776 * if we have to remove any pages. 2777 */ 2778 if (desiredpages < bp->b_npages) { 2779 vm_page_t m; 2780 2781 vm_page_lock_queues(); 2782 for (i = desiredpages; i < bp->b_npages; i++) { 2783 /* 2784 * the page is not freed here -- it 2785 * is the responsibility of 2786 * vnode_pager_setsize 2787 */ 2788 m = bp->b_pages[i]; 2789 KASSERT(m != bogus_page, 2790 ("allocbuf: bogus page found")); 2791 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2792 vm_page_lock_queues(); 2793 2794 bp->b_pages[i] = NULL; 2795 vm_page_unwire(m, 0); 2796 } 2797 vm_page_unlock_queues(); 2798 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2799 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2800 bp->b_npages = desiredpages; 2801 } 2802 } else if (size > bp->b_bcount) { 2803 /* 2804 * We are growing the buffer, possibly in a 2805 * byte-granular fashion. 2806 */ 2807 struct vnode *vp; 2808 vm_object_t obj; 2809 vm_offset_t toff; 2810 vm_offset_t tinc; 2811 2812 /* 2813 * Step 1, bring in the VM pages from the object, 2814 * allocating them if necessary. We must clear 2815 * B_CACHE if these pages are not valid for the 2816 * range covered by the buffer. 2817 */ 2818 2819 vp = bp->b_vp; 2820 obj = bp->b_object; 2821 2822 VM_OBJECT_LOCK(obj); 2823 while (bp->b_npages < desiredpages) { 2824 vm_page_t m; 2825 vm_pindex_t pi; 2826 2827 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2828 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2829 /* 2830 * note: must allocate system pages 2831 * since blocking here could intefere 2832 * with paging I/O, no matter which 2833 * process we are. 2834 */ 2835 m = vm_page_alloc(obj, pi, 2836 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2837 if (m == NULL) { 2838 atomic_add_int(&vm_pageout_deficit, 2839 desiredpages - bp->b_npages); 2840 VM_OBJECT_UNLOCK(obj); 2841 VM_WAIT; 2842 VM_OBJECT_LOCK(obj); 2843 } else { 2844 vm_page_lock_queues(); 2845 vm_page_wakeup(m); 2846 vm_page_unlock_queues(); 2847 bp->b_flags &= ~B_CACHE; 2848 bp->b_pages[bp->b_npages] = m; 2849 ++bp->b_npages; 2850 } 2851 continue; 2852 } 2853 2854 /* 2855 * We found a page. If we have to sleep on it, 2856 * retry because it might have gotten freed out 2857 * from under us. 2858 * 2859 * We can only test PG_BUSY here. Blocking on 2860 * m->busy might lead to a deadlock: 2861 * 2862 * vm_fault->getpages->cluster_read->allocbuf 2863 * 2864 */ 2865 vm_page_lock_queues(); 2866 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2867 continue; 2868 2869 /* 2870 * We have a good page. Should we wakeup the 2871 * page daemon? 2872 */ 2873 if ((curproc != pageproc) && 2874 ((m->queue - m->pc) == PQ_CACHE) && 2875 ((cnt.v_free_count + cnt.v_cache_count) < 2876 (cnt.v_free_min + cnt.v_cache_min))) { 2877 pagedaemon_wakeup(); 2878 } 2879 vm_page_flag_clear(m, PG_ZERO); 2880 vm_page_wire(m); 2881 vm_page_unlock_queues(); 2882 bp->b_pages[bp->b_npages] = m; 2883 ++bp->b_npages; 2884 } 2885 VM_OBJECT_UNLOCK(obj); 2886 2887 /* 2888 * Step 2. We've loaded the pages into the buffer, 2889 * we have to figure out if we can still have B_CACHE 2890 * set. Note that B_CACHE is set according to the 2891 * byte-granular range ( bcount and size ), new the 2892 * aligned range ( newbsize ). 2893 * 2894 * The VM test is against m->valid, which is DEV_BSIZE 2895 * aligned. Needless to say, the validity of the data 2896 * needs to also be DEV_BSIZE aligned. Note that this 2897 * fails with NFS if the server or some other client 2898 * extends the file's EOF. If our buffer is resized, 2899 * B_CACHE may remain set! XXX 2900 */ 2901 2902 toff = bp->b_bcount; 2903 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2904 2905 while ((bp->b_flags & B_CACHE) && toff < size) { 2906 vm_pindex_t pi; 2907 2908 if (tinc > (size - toff)) 2909 tinc = size - toff; 2910 2911 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2912 PAGE_SHIFT; 2913 2914 vfs_buf_test_cache( 2915 bp, 2916 bp->b_offset, 2917 toff, 2918 tinc, 2919 bp->b_pages[pi] 2920 ); 2921 toff += tinc; 2922 tinc = PAGE_SIZE; 2923 } 2924 2925 /* 2926 * Step 3, fixup the KVM pmap. Remember that 2927 * bp->b_data is relative to bp->b_offset, but 2928 * bp->b_offset may be offset into the first page. 2929 */ 2930 2931 bp->b_data = (caddr_t) 2932 trunc_page((vm_offset_t)bp->b_data); 2933 pmap_qenter( 2934 (vm_offset_t)bp->b_data, 2935 bp->b_pages, 2936 bp->b_npages 2937 ); 2938 2939 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2940 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2941 } 2942 } 2943 if (newbsize < bp->b_bufsize) 2944 bufspacewakeup(); 2945 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2946 bp->b_bcount = size; /* requested buffer size */ 2947 return 1; 2948} 2949 2950void 2951biodone(struct bio *bp) 2952{ 2953 mtx_lock(&bdonelock); 2954 bp->bio_flags |= BIO_DONE; 2955 if (bp->bio_done == NULL) 2956 wakeup(bp); 2957 mtx_unlock(&bdonelock); 2958 if (bp->bio_done != NULL) 2959 bp->bio_done(bp); 2960} 2961 2962/* 2963 * Wait for a BIO to finish. 2964 * 2965 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2966 * case is not yet clear. 2967 */ 2968int 2969biowait(struct bio *bp, const char *wchan) 2970{ 2971 2972 mtx_lock(&bdonelock); 2973 while ((bp->bio_flags & BIO_DONE) == 0) 2974 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 2975 mtx_unlock(&bdonelock); 2976 if (bp->bio_error != 0) 2977 return (bp->bio_error); 2978 if (!(bp->bio_flags & BIO_ERROR)) 2979 return (0); 2980 return (EIO); 2981} 2982 2983void 2984biofinish(struct bio *bp, struct devstat *stat, int error) 2985{ 2986 2987 if (error) { 2988 bp->bio_error = error; 2989 bp->bio_flags |= BIO_ERROR; 2990 } 2991 if (stat != NULL) 2992 devstat_end_transaction_bio(stat, bp); 2993 biodone(bp); 2994} 2995 2996/* 2997 * bufwait: 2998 * 2999 * Wait for buffer I/O completion, returning error status. The buffer 3000 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3001 * error and cleared. 3002 */ 3003int 3004bufwait(register struct buf * bp) 3005{ 3006 int s; 3007 3008 s = splbio(); 3009 if (bp->b_iocmd == BIO_READ) 3010 bwait(bp, PRIBIO, "biord"); 3011 else 3012 bwait(bp, PRIBIO, "biowr"); 3013 splx(s); 3014 if (bp->b_flags & B_EINTR) { 3015 bp->b_flags &= ~B_EINTR; 3016 return (EINTR); 3017 } 3018 if (bp->b_ioflags & BIO_ERROR) { 3019 return (bp->b_error ? bp->b_error : EIO); 3020 } else { 3021 return (0); 3022 } 3023} 3024 3025 /* 3026 * Call back function from struct bio back up to struct buf. 3027 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 3028 */ 3029void 3030bufdonebio(struct bio *bp) 3031{ 3032 bufdone(bp->bio_caller2); 3033} 3034 3035/* 3036 * bufdone: 3037 * 3038 * Finish I/O on a buffer, optionally calling a completion function. 3039 * This is usually called from an interrupt so process blocking is 3040 * not allowed. 3041 * 3042 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3043 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3044 * assuming B_INVAL is clear. 3045 * 3046 * For the VMIO case, we set B_CACHE if the op was a read and no 3047 * read error occured, or if the op was a write. B_CACHE is never 3048 * set if the buffer is invalid or otherwise uncacheable. 3049 * 3050 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3051 * initiator to leave B_INVAL set to brelse the buffer out of existance 3052 * in the biodone routine. 3053 */ 3054void 3055bufdone(struct buf *bp) 3056{ 3057 int s; 3058 void (*biodone)(struct buf *); 3059 3060 GIANT_REQUIRED; 3061 3062 s = splbio(); 3063 3064 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3065 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3066 3067 bp->b_flags |= B_DONE; 3068 runningbufwakeup(bp); 3069 3070 if (bp->b_iocmd == BIO_DELETE) { 3071 brelse(bp); 3072 splx(s); 3073 return; 3074 } 3075 3076 if (bp->b_iocmd == BIO_WRITE) { 3077 vwakeup(bp); 3078 } 3079 3080 /* call optional completion function if requested */ 3081 if (bp->b_iodone != NULL) { 3082 biodone = bp->b_iodone; 3083 bp->b_iodone = NULL; 3084 (*biodone) (bp); 3085 splx(s); 3086 return; 3087 } 3088 if (LIST_FIRST(&bp->b_dep) != NULL) 3089 buf_complete(bp); 3090 3091 if (bp->b_flags & B_VMIO) { 3092 int i; 3093 vm_ooffset_t foff; 3094 vm_page_t m; 3095 vm_object_t obj; 3096 int iosize; 3097 struct vnode *vp = bp->b_vp; 3098 3099 obj = bp->b_object; 3100 3101#if defined(VFS_BIO_DEBUG) 3102 mp_fixme("usecount and vflag accessed without locks."); 3103 if (vp->v_usecount == 0) { 3104 panic("biodone: zero vnode ref count"); 3105 } 3106 3107 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3108 panic("biodone: vnode is not setup for merged cache"); 3109 } 3110#endif 3111 3112 foff = bp->b_offset; 3113 KASSERT(bp->b_offset != NOOFFSET, 3114 ("biodone: no buffer offset")); 3115 3116 if (obj != NULL) 3117 VM_OBJECT_LOCK(obj); 3118#if defined(VFS_BIO_DEBUG) 3119 if (obj->paging_in_progress < bp->b_npages) { 3120 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3121 obj->paging_in_progress, bp->b_npages); 3122 } 3123#endif 3124 3125 /* 3126 * Set B_CACHE if the op was a normal read and no error 3127 * occured. B_CACHE is set for writes in the b*write() 3128 * routines. 3129 */ 3130 iosize = bp->b_bcount - bp->b_resid; 3131 if (bp->b_iocmd == BIO_READ && 3132 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3133 !(bp->b_ioflags & BIO_ERROR)) { 3134 bp->b_flags |= B_CACHE; 3135 } 3136 vm_page_lock_queues(); 3137 for (i = 0; i < bp->b_npages; i++) { 3138 int bogusflag = 0; 3139 int resid; 3140 3141 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3142 if (resid > iosize) 3143 resid = iosize; 3144 3145 /* 3146 * cleanup bogus pages, restoring the originals 3147 */ 3148 m = bp->b_pages[i]; 3149 if (m == bogus_page) { 3150 bogusflag = 1; 3151 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3152 if (m == NULL) 3153 panic("biodone: page disappeared!"); 3154 bp->b_pages[i] = m; 3155 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3156 } 3157#if defined(VFS_BIO_DEBUG) 3158 if (OFF_TO_IDX(foff) != m->pindex) { 3159 printf( 3160"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3161 (intmax_t)foff, (uintmax_t)m->pindex); 3162 } 3163#endif 3164 3165 /* 3166 * In the write case, the valid and clean bits are 3167 * already changed correctly ( see bdwrite() ), so we 3168 * only need to do this here in the read case. 3169 */ 3170 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3171 vfs_page_set_valid(bp, foff, i, m); 3172 } 3173 vm_page_flag_clear(m, PG_ZERO); 3174 3175 /* 3176 * when debugging new filesystems or buffer I/O methods, this 3177 * is the most common error that pops up. if you see this, you 3178 * have not set the page busy flag correctly!!! 3179 */ 3180 if (m->busy == 0) { 3181 printf("biodone: page busy < 0, " 3182 "pindex: %d, foff: 0x(%x,%x), " 3183 "resid: %d, index: %d\n", 3184 (int) m->pindex, (int)(foff >> 32), 3185 (int) foff & 0xffffffff, resid, i); 3186 if (!vn_isdisk(vp, NULL)) 3187 printf(" iosize: %ld, lblkno: %jd, flags: 0x%x, npages: %d\n", 3188 bp->b_vp->v_mount->mnt_stat.f_iosize, 3189 (intmax_t) bp->b_lblkno, 3190 bp->b_flags, bp->b_npages); 3191 else 3192 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3193 (intmax_t) bp->b_lblkno, 3194 bp->b_flags, bp->b_npages); 3195 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 3196 m->valid, m->dirty, m->wire_count); 3197 panic("biodone: page busy < 0\n"); 3198 } 3199 vm_page_io_finish(m); 3200 vm_object_pip_subtract(obj, 1); 3201 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3202 iosize -= resid; 3203 } 3204 vm_page_unlock_queues(); 3205 if (obj != NULL) { 3206 vm_object_pip_wakeupn(obj, 0); 3207 VM_OBJECT_UNLOCK(obj); 3208 } 3209 } 3210 3211 /* 3212 * For asynchronous completions, release the buffer now. The brelse 3213 * will do a wakeup there if necessary - so no need to do a wakeup 3214 * here in the async case. The sync case always needs to do a wakeup. 3215 */ 3216 3217 if (bp->b_flags & B_ASYNC) { 3218 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3219 brelse(bp); 3220 else 3221 bqrelse(bp); 3222 } else { 3223 bdone(bp); 3224 } 3225 splx(s); 3226} 3227 3228/* 3229 * This routine is called in lieu of iodone in the case of 3230 * incomplete I/O. This keeps the busy status for pages 3231 * consistant. 3232 */ 3233void 3234vfs_unbusy_pages(struct buf * bp) 3235{ 3236 int i; 3237 3238 GIANT_REQUIRED; 3239 3240 runningbufwakeup(bp); 3241 if (bp->b_flags & B_VMIO) { 3242 vm_object_t obj; 3243 3244 obj = bp->b_object; 3245 VM_OBJECT_LOCK(obj); 3246 vm_page_lock_queues(); 3247 for (i = 0; i < bp->b_npages; i++) { 3248 vm_page_t m = bp->b_pages[i]; 3249 3250 if (m == bogus_page) { 3251 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3252 if (!m) { 3253 panic("vfs_unbusy_pages: page missing\n"); 3254 } 3255 bp->b_pages[i] = m; 3256 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3257 } 3258 vm_object_pip_subtract(obj, 1); 3259 vm_page_flag_clear(m, PG_ZERO); 3260 vm_page_io_finish(m); 3261 } 3262 vm_page_unlock_queues(); 3263 vm_object_pip_wakeupn(obj, 0); 3264 VM_OBJECT_UNLOCK(obj); 3265 } 3266} 3267 3268/* 3269 * vfs_page_set_valid: 3270 * 3271 * Set the valid bits in a page based on the supplied offset. The 3272 * range is restricted to the buffer's size. 3273 * 3274 * This routine is typically called after a read completes. 3275 */ 3276static void 3277vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3278{ 3279 vm_ooffset_t soff, eoff; 3280 3281 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3282 /* 3283 * Start and end offsets in buffer. eoff - soff may not cross a 3284 * page boundry or cross the end of the buffer. The end of the 3285 * buffer, in this case, is our file EOF, not the allocation size 3286 * of the buffer. 3287 */ 3288 soff = off; 3289 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3290 if (eoff > bp->b_offset + bp->b_bcount) 3291 eoff = bp->b_offset + bp->b_bcount; 3292 3293 /* 3294 * Set valid range. This is typically the entire buffer and thus the 3295 * entire page. 3296 */ 3297 if (eoff > soff) { 3298 vm_page_set_validclean( 3299 m, 3300 (vm_offset_t) (soff & PAGE_MASK), 3301 (vm_offset_t) (eoff - soff) 3302 ); 3303 } 3304} 3305 3306/* 3307 * This routine is called before a device strategy routine. 3308 * It is used to tell the VM system that paging I/O is in 3309 * progress, and treat the pages associated with the buffer 3310 * almost as being PG_BUSY. Also the object paging_in_progress 3311 * flag is handled to make sure that the object doesn't become 3312 * inconsistant. 3313 * 3314 * Since I/O has not been initiated yet, certain buffer flags 3315 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3316 * and should be ignored. 3317 */ 3318void 3319vfs_busy_pages(struct buf * bp, int clear_modify) 3320{ 3321 int i, bogus; 3322 3323 if (bp->b_flags & B_VMIO) { 3324 vm_object_t obj; 3325 vm_ooffset_t foff; 3326 3327 obj = bp->b_object; 3328 foff = bp->b_offset; 3329 KASSERT(bp->b_offset != NOOFFSET, 3330 ("vfs_busy_pages: no buffer offset")); 3331 vfs_setdirty(bp); 3332 if (obj != NULL) 3333 VM_OBJECT_LOCK(obj); 3334retry: 3335 vm_page_lock_queues(); 3336 for (i = 0; i < bp->b_npages; i++) { 3337 vm_page_t m = bp->b_pages[i]; 3338 3339 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3340 goto retry; 3341 } 3342 bogus = 0; 3343 for (i = 0; i < bp->b_npages; i++) { 3344 vm_page_t m = bp->b_pages[i]; 3345 3346 vm_page_flag_clear(m, PG_ZERO); 3347 if ((bp->b_flags & B_CLUSTER) == 0) { 3348 vm_object_pip_add(obj, 1); 3349 vm_page_io_start(m); 3350 } 3351 /* 3352 * When readying a buffer for a read ( i.e 3353 * clear_modify == 0 ), it is important to do 3354 * bogus_page replacement for valid pages in 3355 * partially instantiated buffers. Partially 3356 * instantiated buffers can, in turn, occur when 3357 * reconstituting a buffer from its VM backing store 3358 * base. We only have to do this if B_CACHE is 3359 * clear ( which causes the I/O to occur in the 3360 * first place ). The replacement prevents the read 3361 * I/O from overwriting potentially dirty VM-backed 3362 * pages. XXX bogus page replacement is, uh, bogus. 3363 * It may not work properly with small-block devices. 3364 * We need to find a better way. 3365 */ 3366 pmap_remove_all(m); 3367 if (clear_modify) 3368 vfs_page_set_valid(bp, foff, i, m); 3369 else if (m->valid == VM_PAGE_BITS_ALL && 3370 (bp->b_flags & B_CACHE) == 0) { 3371 bp->b_pages[i] = bogus_page; 3372 bogus++; 3373 } 3374 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3375 } 3376 vm_page_unlock_queues(); 3377 if (obj != NULL) 3378 VM_OBJECT_UNLOCK(obj); 3379 if (bogus) 3380 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3381 } 3382} 3383 3384/* 3385 * Tell the VM system that the pages associated with this buffer 3386 * are clean. This is used for delayed writes where the data is 3387 * going to go to disk eventually without additional VM intevention. 3388 * 3389 * Note that while we only really need to clean through to b_bcount, we 3390 * just go ahead and clean through to b_bufsize. 3391 */ 3392static void 3393vfs_clean_pages(struct buf * bp) 3394{ 3395 int i; 3396 3397 GIANT_REQUIRED; 3398 3399 if (bp->b_flags & B_VMIO) { 3400 vm_ooffset_t foff; 3401 3402 foff = bp->b_offset; 3403 KASSERT(bp->b_offset != NOOFFSET, 3404 ("vfs_clean_pages: no buffer offset")); 3405 vm_page_lock_queues(); 3406 for (i = 0; i < bp->b_npages; i++) { 3407 vm_page_t m = bp->b_pages[i]; 3408 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3409 vm_ooffset_t eoff = noff; 3410 3411 if (eoff > bp->b_offset + bp->b_bufsize) 3412 eoff = bp->b_offset + bp->b_bufsize; 3413 vfs_page_set_valid(bp, foff, i, m); 3414 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3415 foff = noff; 3416 } 3417 vm_page_unlock_queues(); 3418 } 3419} 3420 3421/* 3422 * vfs_bio_set_validclean: 3423 * 3424 * Set the range within the buffer to valid and clean. The range is 3425 * relative to the beginning of the buffer, b_offset. Note that b_offset 3426 * itself may be offset from the beginning of the first page. 3427 * 3428 */ 3429 3430void 3431vfs_bio_set_validclean(struct buf *bp, int base, int size) 3432{ 3433 if (bp->b_flags & B_VMIO) { 3434 int i; 3435 int n; 3436 3437 /* 3438 * Fixup base to be relative to beginning of first page. 3439 * Set initial n to be the maximum number of bytes in the 3440 * first page that can be validated. 3441 */ 3442 3443 base += (bp->b_offset & PAGE_MASK); 3444 n = PAGE_SIZE - (base & PAGE_MASK); 3445 3446 vm_page_lock_queues(); 3447 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3448 vm_page_t m = bp->b_pages[i]; 3449 3450 if (n > size) 3451 n = size; 3452 3453 vm_page_set_validclean(m, base & PAGE_MASK, n); 3454 base += n; 3455 size -= n; 3456 n = PAGE_SIZE; 3457 } 3458 vm_page_unlock_queues(); 3459 } 3460} 3461 3462/* 3463 * vfs_bio_clrbuf: 3464 * 3465 * clear a buffer. This routine essentially fakes an I/O, so we need 3466 * to clear BIO_ERROR and B_INVAL. 3467 * 3468 * Note that while we only theoretically need to clear through b_bcount, 3469 * we go ahead and clear through b_bufsize. 3470 */ 3471 3472void 3473vfs_bio_clrbuf(struct buf *bp) 3474{ 3475 int i, mask = 0; 3476 caddr_t sa, ea; 3477 3478 GIANT_REQUIRED; 3479 3480 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3481 bp->b_flags &= ~B_INVAL; 3482 bp->b_ioflags &= ~BIO_ERROR; 3483 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3484 (bp->b_offset & PAGE_MASK) == 0) { 3485 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3486 if ((bp->b_pages[0]->valid & mask) == mask) { 3487 bp->b_resid = 0; 3488 return; 3489 } 3490 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3491 ((bp->b_pages[0]->valid & mask) == 0)) { 3492 bzero(bp->b_data, bp->b_bufsize); 3493 bp->b_pages[0]->valid |= mask; 3494 bp->b_resid = 0; 3495 return; 3496 } 3497 } 3498 ea = sa = bp->b_data; 3499 for(i=0;i<bp->b_npages;i++,sa=ea) { 3500 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3501 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3502 ea = (caddr_t)(vm_offset_t)ulmin( 3503 (u_long)(vm_offset_t)ea, 3504 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3505 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3506 if ((bp->b_pages[i]->valid & mask) == mask) 3507 continue; 3508 if ((bp->b_pages[i]->valid & mask) == 0) { 3509 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3510 bzero(sa, ea - sa); 3511 } 3512 } else { 3513 for (; sa < ea; sa += DEV_BSIZE, j++) { 3514 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3515 (bp->b_pages[i]->valid & (1<<j)) == 0) 3516 bzero(sa, DEV_BSIZE); 3517 } 3518 } 3519 bp->b_pages[i]->valid |= mask; 3520 vm_page_lock_queues(); 3521 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3522 vm_page_unlock_queues(); 3523 } 3524 bp->b_resid = 0; 3525 } else { 3526 clrbuf(bp); 3527 } 3528} 3529 3530/* 3531 * vm_hold_load_pages and vm_hold_free_pages get pages into 3532 * a buffers address space. The pages are anonymous and are 3533 * not associated with a file object. 3534 */ 3535static void 3536vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3537{ 3538 vm_offset_t pg; 3539 vm_page_t p; 3540 int index; 3541 3542 GIANT_REQUIRED; 3543 3544 to = round_page(to); 3545 from = round_page(from); 3546 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3547 3548 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3549tryagain: 3550 /* 3551 * note: must allocate system pages since blocking here 3552 * could intefere with paging I/O, no matter which 3553 * process we are. 3554 */ 3555 VM_OBJECT_LOCK(kernel_object); 3556 p = vm_page_alloc(kernel_object, 3557 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3558 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3559 VM_OBJECT_UNLOCK(kernel_object); 3560 if (!p) { 3561 atomic_add_int(&vm_pageout_deficit, 3562 (to - pg) >> PAGE_SHIFT); 3563 VM_WAIT; 3564 goto tryagain; 3565 } 3566 vm_page_lock_queues(); 3567 p->valid = VM_PAGE_BITS_ALL; 3568 vm_page_unlock_queues(); 3569 pmap_qenter(pg, &p, 1); 3570 bp->b_pages[index] = p; 3571 vm_page_lock_queues(); 3572 vm_page_wakeup(p); 3573 vm_page_unlock_queues(); 3574 } 3575 bp->b_npages = index; 3576} 3577 3578/* Return pages associated with this buf to the vm system */ 3579static void 3580vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3581{ 3582 vm_offset_t pg; 3583 vm_page_t p; 3584 int index, newnpages; 3585 3586 GIANT_REQUIRED; 3587 3588 from = round_page(from); 3589 to = round_page(to); 3590 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3591 3592 if (bp->b_object != NULL) 3593 VM_OBJECT_LOCK(bp->b_object); 3594 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3595 p = bp->b_pages[index]; 3596 if (p && (index < bp->b_npages)) { 3597 if (p->busy) { 3598 printf( 3599 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3600 (intmax_t)bp->b_blkno, 3601 (intmax_t)bp->b_lblkno); 3602 } 3603 bp->b_pages[index] = NULL; 3604 pmap_qremove(pg, 1); 3605 vm_page_lock_queues(); 3606 vm_page_busy(p); 3607 vm_page_unwire(p, 0); 3608 vm_page_free(p); 3609 vm_page_unlock_queues(); 3610 } 3611 } 3612 if (bp->b_object != NULL) 3613 VM_OBJECT_UNLOCK(bp->b_object); 3614 bp->b_npages = newnpages; 3615} 3616 3617/* 3618 * Map an IO request into kernel virtual address space. 3619 * 3620 * All requests are (re)mapped into kernel VA space. 3621 * Notice that we use b_bufsize for the size of the buffer 3622 * to be mapped. b_bcount might be modified by the driver. 3623 * 3624 * Note that even if the caller determines that the address space should 3625 * be valid, a race or a smaller-file mapped into a larger space may 3626 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3627 * check the return value. 3628 */ 3629int 3630vmapbuf(struct buf *bp) 3631{ 3632 caddr_t addr, kva; 3633 vm_paddr_t pa; 3634 int pidx, i; 3635 struct vm_page *m; 3636 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3637 3638 GIANT_REQUIRED; 3639 3640 if ((bp->b_flags & B_PHYS) == 0) 3641 panic("vmapbuf"); 3642 if (bp->b_bufsize < 0) 3643 return (-1); 3644 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3645 addr < bp->b_data + bp->b_bufsize; 3646 addr += PAGE_SIZE, pidx++) { 3647 /* 3648 * Do the vm_fault if needed; do the copy-on-write thing 3649 * when reading stuff off device into memory. 3650 * 3651 * NOTE! Must use pmap_extract() because addr may be in 3652 * the userland address space, and kextract is only guarenteed 3653 * to work for the kernland address space (see: sparc64 port). 3654 */ 3655retry: 3656 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data, 3657 (bp->b_iocmd == BIO_READ) ? 3658 (VM_PROT_READ|VM_PROT_WRITE) : VM_PROT_READ); 3659 if (i < 0) { 3660 vm_page_lock_queues(); 3661 for (i = 0; i < pidx; ++i) { 3662 vm_page_unhold(bp->b_pages[i]); 3663 bp->b_pages[i] = NULL; 3664 } 3665 vm_page_unlock_queues(); 3666 return(-1); 3667 } 3668 pa = pmap_extract(pmap, (vm_offset_t)addr); 3669 if (pa == 0) { 3670 printf("vmapbuf: warning, race against user address during I/O"); 3671 goto retry; 3672 } 3673 m = PHYS_TO_VM_PAGE(pa); 3674 vm_page_lock_queues(); 3675 vm_page_hold(m); 3676 vm_page_unlock_queues(); 3677 bp->b_pages[pidx] = m; 3678 } 3679 if (pidx > btoc(MAXPHYS)) 3680 panic("vmapbuf: mapped more than MAXPHYS"); 3681 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3682 3683 kva = bp->b_saveaddr; 3684 bp->b_npages = pidx; 3685 bp->b_saveaddr = bp->b_data; 3686 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3687 return(0); 3688} 3689 3690/* 3691 * Free the io map PTEs associated with this IO operation. 3692 * We also invalidate the TLB entries and restore the original b_addr. 3693 */ 3694void 3695vunmapbuf(struct buf *bp) 3696{ 3697 int pidx; 3698 int npages; 3699 3700 GIANT_REQUIRED; 3701 3702 if ((bp->b_flags & B_PHYS) == 0) 3703 panic("vunmapbuf"); 3704 3705 npages = bp->b_npages; 3706 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3707 npages); 3708 vm_page_lock_queues(); 3709 for (pidx = 0; pidx < npages; pidx++) 3710 vm_page_unhold(bp->b_pages[pidx]); 3711 vm_page_unlock_queues(); 3712 3713 bp->b_data = bp->b_saveaddr; 3714} 3715 3716void 3717bdone(struct buf *bp) 3718{ 3719 mtx_lock(&bdonelock); 3720 bp->b_flags |= B_DONE; 3721 wakeup(bp); 3722 mtx_unlock(&bdonelock); 3723} 3724 3725void 3726bwait(struct buf *bp, u_char pri, const char *wchan) 3727{ 3728 mtx_lock(&bdonelock); 3729 while ((bp->b_flags & B_DONE) == 0) 3730 msleep(bp, &bdonelock, pri, wchan, 0); 3731 mtx_unlock(&bdonelock); 3732} 3733 3734#include "opt_ddb.h" 3735#ifdef DDB 3736#include <ddb/ddb.h> 3737 3738/* DDB command to show buffer data */ 3739DB_SHOW_COMMAND(buffer, db_show_buffer) 3740{ 3741 /* get args */ 3742 struct buf *bp = (struct buf *)addr; 3743 3744 if (!have_addr) { 3745 db_printf("usage: show buffer <addr>\n"); 3746 return; 3747 } 3748 3749 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3750 db_printf( 3751 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3752 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd, b_pblkno = %jd\n", 3753 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3754 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3755 (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno); 3756 if (bp->b_npages) { 3757 int i; 3758 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3759 for (i = 0; i < bp->b_npages; i++) { 3760 vm_page_t m; 3761 m = bp->b_pages[i]; 3762 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3763 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3764 if ((i + 1) < bp->b_npages) 3765 db_printf(","); 3766 } 3767 db_printf("\n"); 3768 } 3769} 3770#endif /* DDB */ 3771