vfs_bio.c revision 122746
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 122746 2003-11-15 08:45:36Z alc $"); 30 31#include <sys/param.h> 32#include <sys/systm.h> 33#include <sys/bio.h> 34#include <sys/conf.h> 35#include <sys/buf.h> 36#include <sys/devicestat.h> 37#include <sys/eventhandler.h> 38#include <sys/lock.h> 39#include <sys/malloc.h> 40#include <sys/mount.h> 41#include <sys/mutex.h> 42#include <sys/kernel.h> 43#include <sys/kthread.h> 44#include <sys/proc.h> 45#include <sys/resourcevar.h> 46#include <sys/sysctl.h> 47#include <sys/vmmeter.h> 48#include <sys/vnode.h> 49#include <vm/vm.h> 50#include <vm/vm_param.h> 51#include <vm/vm_kern.h> 52#include <vm/vm_pageout.h> 53#include <vm/vm_page.h> 54#include <vm/vm_object.h> 55#include <vm/vm_extern.h> 56#include <vm/vm_map.h> 57#include "opt_directio.h" 58#include "opt_swap.h" 59 60static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 61 62struct bio_ops bioops; /* I/O operation notification */ 63 64struct buf_ops buf_ops_bio = { 65 "buf_ops_bio", 66 bwrite 67}; 68 69/* 70 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 71 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 72 */ 73struct buf *buf; /* buffer header pool */ 74 75static struct proc *bufdaemonproc; 76 77static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 78 vm_offset_t to); 79static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 80 vm_offset_t to); 81static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 82 int pageno, vm_page_t m); 83static void vfs_clean_pages(struct buf * bp); 84static void vfs_setdirty(struct buf *bp); 85static void vfs_vmio_release(struct buf *bp); 86static void vfs_backgroundwritedone(struct buf *bp); 87static int vfs_bio_clcheck(struct vnode *vp, int size, 88 daddr_t lblkno, daddr_t blkno); 89static int flushbufqueues(int flushdeps); 90static void buf_daemon(void); 91void bremfreel(struct buf * bp); 92 93int vmiodirenable = TRUE; 94SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 95 "Use the VM system for directory writes"); 96int runningbufspace; 97SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 98 "Amount of presently outstanding async buffer io"); 99static int bufspace; 100SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 101 "KVA memory used for bufs"); 102static int maxbufspace; 103SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 104 "Maximum allowed value of bufspace (including buf_daemon)"); 105static int bufmallocspace; 106SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 107 "Amount of malloced memory for buffers"); 108static int maxbufmallocspace; 109SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 110 "Maximum amount of malloced memory for buffers"); 111static int lobufspace; 112SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 113 "Minimum amount of buffers we want to have"); 114static int hibufspace; 115SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 116 "Maximum allowed value of bufspace (excluding buf_daemon)"); 117static int bufreusecnt; 118SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 119 "Number of times we have reused a buffer"); 120static int buffreekvacnt; 121SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 122 "Number of times we have freed the KVA space from some buffer"); 123static int bufdefragcnt; 124SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 125 "Number of times we have had to repeat buffer allocation to defragment"); 126static int lorunningspace; 127SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 128 "Minimum preferred space used for in-progress I/O"); 129static int hirunningspace; 130SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 131 "Maximum amount of space to use for in-progress I/O"); 132static int dirtybufferflushes; 133SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 134 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 135static int altbufferflushes; 136SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 137 0, "Number of fsync flushes to limit dirty buffers"); 138static int recursiveflushes; 139SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 140 0, "Number of flushes skipped due to being recursive"); 141static int numdirtybuffers; 142SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 143 "Number of buffers that are dirty (has unwritten changes) at the moment"); 144static int lodirtybuffers; 145SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 146 "How many buffers we want to have free before bufdaemon can sleep"); 147static int hidirtybuffers; 148SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 149 "When the number of dirty buffers is considered severe"); 150static int dirtybufthresh; 151SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 152 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 153static int numfreebuffers; 154SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 155 "Number of free buffers"); 156static int lofreebuffers; 157SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 158 "XXX Unused"); 159static int hifreebuffers; 160SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 161 "XXX Complicatedly unused"); 162static int getnewbufcalls; 163SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 164 "Number of calls to getnewbuf"); 165static int getnewbufrestarts; 166SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 167 "Number of times getnewbuf has had to restart a buffer aquisition"); 168static int dobkgrdwrite = 1; 169SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 170 "Do background writes (honoring the BV_BKGRDWRITE flag)?"); 171 172/* 173 * Wakeup point for bufdaemon, as well as indicator of whether it is already 174 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 175 * is idling. 176 */ 177static int bd_request; 178 179/* 180 * This lock synchronizes access to bd_request. 181 */ 182static struct mtx bdlock; 183 184/* 185 * bogus page -- for I/O to/from partially complete buffers 186 * this is a temporary solution to the problem, but it is not 187 * really that bad. it would be better to split the buffer 188 * for input in the case of buffers partially already in memory, 189 * but the code is intricate enough already. 190 */ 191vm_page_t bogus_page; 192 193/* 194 * Synchronization (sleep/wakeup) variable for active buffer space requests. 195 * Set when wait starts, cleared prior to wakeup(). 196 * Used in runningbufwakeup() and waitrunningbufspace(). 197 */ 198static int runningbufreq; 199 200/* 201 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 202 * waitrunningbufspace(). 203 */ 204static struct mtx rbreqlock; 205 206/* 207 * Synchronization (sleep/wakeup) variable for buffer requests. 208 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 209 * by and/or. 210 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 211 * getnewbuf(), and getblk(). 212 */ 213static int needsbuffer; 214 215/* 216 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 217 */ 218static struct mtx nblock; 219 220/* 221 * Lock that protects against bwait()/bdone()/B_DONE races. 222 */ 223 224static struct mtx bdonelock; 225 226/* 227 * Definitions for the buffer free lists. 228 */ 229#define BUFFER_QUEUES 5 /* number of free buffer queues */ 230 231#define QUEUE_NONE 0 /* on no queue */ 232#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 233#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 234#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 235#define QUEUE_EMPTY 4 /* empty buffer headers */ 236 237/* Queues for free buffers with various properties */ 238static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 239 240/* Lock for the bufqueues */ 241static struct mtx bqlock; 242 243/* 244 * Single global constant for BUF_WMESG, to avoid getting multiple references. 245 * buf_wmesg is referred from macros. 246 */ 247const char *buf_wmesg = BUF_WMESG; 248 249#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 250#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 251#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 252#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 253 254#ifdef DIRECTIO 255extern void ffs_rawread_setup(void); 256#endif /* DIRECTIO */ 257/* 258 * numdirtywakeup: 259 * 260 * If someone is blocked due to there being too many dirty buffers, 261 * and numdirtybuffers is now reasonable, wake them up. 262 */ 263 264static __inline void 265numdirtywakeup(int level) 266{ 267 if (numdirtybuffers <= level) { 268 mtx_lock(&nblock); 269 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 270 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 271 wakeup(&needsbuffer); 272 } 273 mtx_unlock(&nblock); 274 } 275} 276 277/* 278 * bufspacewakeup: 279 * 280 * Called when buffer space is potentially available for recovery. 281 * getnewbuf() will block on this flag when it is unable to free 282 * sufficient buffer space. Buffer space becomes recoverable when 283 * bp's get placed back in the queues. 284 */ 285 286static __inline void 287bufspacewakeup(void) 288{ 289 /* 290 * If someone is waiting for BUF space, wake them up. Even 291 * though we haven't freed the kva space yet, the waiting 292 * process will be able to now. 293 */ 294 mtx_lock(&nblock); 295 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 296 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 297 wakeup(&needsbuffer); 298 } 299 mtx_unlock(&nblock); 300} 301 302/* 303 * runningbufwakeup() - in-progress I/O accounting. 304 * 305 */ 306static __inline void 307runningbufwakeup(struct buf *bp) 308{ 309 if (bp->b_runningbufspace) { 310 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 311 bp->b_runningbufspace = 0; 312 mtx_lock(&rbreqlock); 313 if (runningbufreq && runningbufspace <= lorunningspace) { 314 runningbufreq = 0; 315 wakeup(&runningbufreq); 316 } 317 mtx_unlock(&rbreqlock); 318 } 319} 320 321/* 322 * bufcountwakeup: 323 * 324 * Called when a buffer has been added to one of the free queues to 325 * account for the buffer and to wakeup anyone waiting for free buffers. 326 * This typically occurs when large amounts of metadata are being handled 327 * by the buffer cache ( else buffer space runs out first, usually ). 328 */ 329 330static __inline void 331bufcountwakeup(void) 332{ 333 atomic_add_int(&numfreebuffers, 1); 334 mtx_lock(&nblock); 335 if (needsbuffer) { 336 needsbuffer &= ~VFS_BIO_NEED_ANY; 337 if (numfreebuffers >= hifreebuffers) 338 needsbuffer &= ~VFS_BIO_NEED_FREE; 339 wakeup(&needsbuffer); 340 } 341 mtx_unlock(&nblock); 342} 343 344/* 345 * waitrunningbufspace() 346 * 347 * runningbufspace is a measure of the amount of I/O currently 348 * running. This routine is used in async-write situations to 349 * prevent creating huge backups of pending writes to a device. 350 * Only asynchronous writes are governed by this function. 351 * 352 * Reads will adjust runningbufspace, but will not block based on it. 353 * The read load has a side effect of reducing the allowed write load. 354 * 355 * This does NOT turn an async write into a sync write. It waits 356 * for earlier writes to complete and generally returns before the 357 * caller's write has reached the device. 358 */ 359static __inline void 360waitrunningbufspace(void) 361{ 362 mtx_lock(&rbreqlock); 363 while (runningbufspace > hirunningspace) { 364 ++runningbufreq; 365 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 366 } 367 mtx_unlock(&rbreqlock); 368} 369 370 371/* 372 * vfs_buf_test_cache: 373 * 374 * Called when a buffer is extended. This function clears the B_CACHE 375 * bit if the newly extended portion of the buffer does not contain 376 * valid data. 377 */ 378static __inline__ 379void 380vfs_buf_test_cache(struct buf *bp, 381 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 382 vm_page_t m) 383{ 384 GIANT_REQUIRED; 385 386 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 387 if (bp->b_flags & B_CACHE) { 388 int base = (foff + off) & PAGE_MASK; 389 if (vm_page_is_valid(m, base, size) == 0) 390 bp->b_flags &= ~B_CACHE; 391 } 392} 393 394/* Wake up the buffer deamon if necessary */ 395static __inline__ 396void 397bd_wakeup(int dirtybuflevel) 398{ 399 mtx_lock(&bdlock); 400 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 401 bd_request = 1; 402 wakeup(&bd_request); 403 } 404 mtx_unlock(&bdlock); 405} 406 407/* 408 * bd_speedup - speedup the buffer cache flushing code 409 */ 410 411static __inline__ 412void 413bd_speedup(void) 414{ 415 bd_wakeup(1); 416} 417 418/* 419 * Calculating buffer cache scaling values and reserve space for buffer 420 * headers. This is called during low level kernel initialization and 421 * may be called more then once. We CANNOT write to the memory area 422 * being reserved at this time. 423 */ 424caddr_t 425kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 426{ 427 /* 428 * physmem_est is in pages. Convert it to kilobytes (assumes 429 * PAGE_SIZE is >= 1K) 430 */ 431 physmem_est = physmem_est * (PAGE_SIZE / 1024); 432 433 /* 434 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 435 * For the first 64MB of ram nominally allocate sufficient buffers to 436 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 437 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 438 * the buffer cache we limit the eventual kva reservation to 439 * maxbcache bytes. 440 * 441 * factor represents the 1/4 x ram conversion. 442 */ 443 if (nbuf == 0) { 444 int factor = 4 * BKVASIZE / 1024; 445 446 nbuf = 50; 447 if (physmem_est > 4096) 448 nbuf += min((physmem_est - 4096) / factor, 449 65536 / factor); 450 if (physmem_est > 65536) 451 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 452 453 if (maxbcache && nbuf > maxbcache / BKVASIZE) 454 nbuf = maxbcache / BKVASIZE; 455 } 456 457#if 0 458 /* 459 * Do not allow the buffer_map to be more then 1/2 the size of the 460 * kernel_map. 461 */ 462 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 463 (BKVASIZE * 2)) { 464 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 465 (BKVASIZE * 2); 466 printf("Warning: nbufs capped at %d\n", nbuf); 467 } 468#endif 469 470 /* 471 * swbufs are used as temporary holders for I/O, such as paging I/O. 472 * We have no less then 16 and no more then 256. 473 */ 474 nswbuf = max(min(nbuf/4, 256), 16); 475#ifdef NSWBUF_MIN 476 if (nswbuf < NSWBUF_MIN) 477 nswbuf = NSWBUF_MIN; 478#endif 479#ifdef DIRECTIO 480 ffs_rawread_setup(); 481#endif 482 483 /* 484 * Reserve space for the buffer cache buffers 485 */ 486 swbuf = (void *)v; 487 v = (caddr_t)(swbuf + nswbuf); 488 buf = (void *)v; 489 v = (caddr_t)(buf + nbuf); 490 491 return(v); 492} 493 494/* Initialize the buffer subsystem. Called before use of any buffers. */ 495void 496bufinit(void) 497{ 498 struct buf *bp; 499 int i; 500 501 GIANT_REQUIRED; 502 503 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 504 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 505 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 506 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 507 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 508 509 /* next, make a null set of free lists */ 510 for (i = 0; i < BUFFER_QUEUES; i++) 511 TAILQ_INIT(&bufqueues[i]); 512 513 /* finally, initialize each buffer header and stick on empty q */ 514 for (i = 0; i < nbuf; i++) { 515 bp = &buf[i]; 516 bzero(bp, sizeof *bp); 517 bp->b_flags = B_INVAL; /* we're just an empty header */ 518 bp->b_dev = NODEV; 519 bp->b_rcred = NOCRED; 520 bp->b_wcred = NOCRED; 521 bp->b_qindex = QUEUE_EMPTY; 522 bp->b_vflags = 0; 523 bp->b_xflags = 0; 524 LIST_INIT(&bp->b_dep); 525 BUF_LOCKINIT(bp); 526 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 527 } 528 529 /* 530 * maxbufspace is the absolute maximum amount of buffer space we are 531 * allowed to reserve in KVM and in real terms. The absolute maximum 532 * is nominally used by buf_daemon. hibufspace is the nominal maximum 533 * used by most other processes. The differential is required to 534 * ensure that buf_daemon is able to run when other processes might 535 * be blocked waiting for buffer space. 536 * 537 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 538 * this may result in KVM fragmentation which is not handled optimally 539 * by the system. 540 */ 541 maxbufspace = nbuf * BKVASIZE; 542 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 543 lobufspace = hibufspace - MAXBSIZE; 544 545 lorunningspace = 512 * 1024; 546 hirunningspace = 1024 * 1024; 547 548/* 549 * Limit the amount of malloc memory since it is wired permanently into 550 * the kernel space. Even though this is accounted for in the buffer 551 * allocation, we don't want the malloced region to grow uncontrolled. 552 * The malloc scheme improves memory utilization significantly on average 553 * (small) directories. 554 */ 555 maxbufmallocspace = hibufspace / 20; 556 557/* 558 * Reduce the chance of a deadlock occuring by limiting the number 559 * of delayed-write dirty buffers we allow to stack up. 560 */ 561 hidirtybuffers = nbuf / 4 + 20; 562 dirtybufthresh = hidirtybuffers * 9 / 10; 563 numdirtybuffers = 0; 564/* 565 * To support extreme low-memory systems, make sure hidirtybuffers cannot 566 * eat up all available buffer space. This occurs when our minimum cannot 567 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 568 * BKVASIZE'd (8K) buffers. 569 */ 570 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 571 hidirtybuffers >>= 1; 572 } 573 lodirtybuffers = hidirtybuffers / 2; 574 575/* 576 * Try to keep the number of free buffers in the specified range, 577 * and give special processes (e.g. like buf_daemon) access to an 578 * emergency reserve. 579 */ 580 lofreebuffers = nbuf / 18 + 5; 581 hifreebuffers = 2 * lofreebuffers; 582 numfreebuffers = nbuf; 583 584/* 585 * Maximum number of async ops initiated per buf_daemon loop. This is 586 * somewhat of a hack at the moment, we really need to limit ourselves 587 * based on the number of bytes of I/O in-transit that were initiated 588 * from buf_daemon. 589 */ 590 591 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 592 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 593} 594 595/* 596 * bfreekva() - free the kva allocation for a buffer. 597 * 598 * Must be called at splbio() or higher as this is the only locking for 599 * buffer_map. 600 * 601 * Since this call frees up buffer space, we call bufspacewakeup(). 602 */ 603static void 604bfreekva(struct buf * bp) 605{ 606 GIANT_REQUIRED; 607 608 if (bp->b_kvasize) { 609 atomic_add_int(&buffreekvacnt, 1); 610 atomic_subtract_int(&bufspace, bp->b_kvasize); 611 vm_map_delete(buffer_map, 612 (vm_offset_t) bp->b_kvabase, 613 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 614 ); 615 bp->b_kvasize = 0; 616 bufspacewakeup(); 617 } 618} 619 620/* 621 * bremfree: 622 * 623 * Remove the buffer from the appropriate free list. 624 */ 625void 626bremfree(struct buf * bp) 627{ 628 mtx_lock(&bqlock); 629 bremfreel(bp); 630 mtx_unlock(&bqlock); 631} 632 633void 634bremfreel(struct buf * bp) 635{ 636 int s = splbio(); 637 int old_qindex = bp->b_qindex; 638 639 GIANT_REQUIRED; 640 641 if (bp->b_qindex != QUEUE_NONE) { 642 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 643 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 644 bp->b_qindex = QUEUE_NONE; 645 } else { 646 if (BUF_REFCNT(bp) <= 1) 647 panic("bremfree: removing a buffer not on a queue"); 648 } 649 650 /* 651 * Fixup numfreebuffers count. If the buffer is invalid or not 652 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 653 * the buffer was free and we must decrement numfreebuffers. 654 */ 655 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 656 switch(old_qindex) { 657 case QUEUE_DIRTY: 658 case QUEUE_CLEAN: 659 case QUEUE_EMPTY: 660 case QUEUE_EMPTYKVA: 661 atomic_subtract_int(&numfreebuffers, 1); 662 break; 663 default: 664 break; 665 } 666 } 667 splx(s); 668} 669 670 671/* 672 * Get a buffer with the specified data. Look in the cache first. We 673 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 674 * is set, the buffer is valid and we do not have to do anything ( see 675 * getblk() ). This is really just a special case of breadn(). 676 */ 677int 678bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 679 struct buf ** bpp) 680{ 681 682 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 683} 684 685/* 686 * Operates like bread, but also starts asynchronous I/O on 687 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 688 * to initiating I/O . If B_CACHE is set, the buffer is valid 689 * and we do not have to do anything. 690 */ 691int 692breadn(struct vnode * vp, daddr_t blkno, int size, 693 daddr_t * rablkno, int *rabsize, 694 int cnt, struct ucred * cred, struct buf ** bpp) 695{ 696 struct buf *bp, *rabp; 697 int i; 698 int rv = 0, readwait = 0; 699 700 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 701 702 /* if not found in cache, do some I/O */ 703 if ((bp->b_flags & B_CACHE) == 0) { 704 if (curthread != PCPU_GET(idlethread)) 705 curthread->td_proc->p_stats->p_ru.ru_inblock++; 706 bp->b_iocmd = BIO_READ; 707 bp->b_flags &= ~B_INVAL; 708 bp->b_ioflags &= ~BIO_ERROR; 709 if (bp->b_rcred == NOCRED && cred != NOCRED) 710 bp->b_rcred = crhold(cred); 711 vfs_busy_pages(bp, 0); 712 bp->b_iooffset = dbtob(bp->b_blkno); 713 if (vp->v_type == VCHR) 714 VOP_SPECSTRATEGY(vp, bp); 715 else 716 VOP_STRATEGY(vp, bp); 717 ++readwait; 718 } 719 720 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 721 if (inmem(vp, *rablkno)) 722 continue; 723 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 724 725 if ((rabp->b_flags & B_CACHE) == 0) { 726 if (curthread != PCPU_GET(idlethread)) 727 curthread->td_proc->p_stats->p_ru.ru_inblock++; 728 rabp->b_flags |= B_ASYNC; 729 rabp->b_flags &= ~B_INVAL; 730 rabp->b_ioflags &= ~BIO_ERROR; 731 rabp->b_iocmd = BIO_READ; 732 if (rabp->b_rcred == NOCRED && cred != NOCRED) 733 rabp->b_rcred = crhold(cred); 734 vfs_busy_pages(rabp, 0); 735 BUF_KERNPROC(rabp); 736 rabp->b_iooffset = dbtob(rabp->b_blkno); 737 if (vp->v_type == VCHR) 738 VOP_SPECSTRATEGY(vp, rabp); 739 else 740 VOP_STRATEGY(vp, rabp); 741 } else { 742 brelse(rabp); 743 } 744 } 745 746 if (readwait) { 747 rv = bufwait(bp); 748 } 749 return (rv); 750} 751 752/* 753 * Write, release buffer on completion. (Done by iodone 754 * if async). Do not bother writing anything if the buffer 755 * is invalid. 756 * 757 * Note that we set B_CACHE here, indicating that buffer is 758 * fully valid and thus cacheable. This is true even of NFS 759 * now so we set it generally. This could be set either here 760 * or in biodone() since the I/O is synchronous. We put it 761 * here. 762 */ 763 764int 765bwrite(struct buf * bp) 766{ 767 int oldflags, s; 768 struct buf *newbp; 769 770 if (bp->b_flags & B_INVAL) { 771 brelse(bp); 772 return (0); 773 } 774 775 oldflags = bp->b_flags; 776 777 if (BUF_REFCNT(bp) == 0) 778 panic("bwrite: buffer is not busy???"); 779 s = splbio(); 780 /* 781 * If a background write is already in progress, delay 782 * writing this block if it is asynchronous. Otherwise 783 * wait for the background write to complete. 784 */ 785 VI_LOCK(bp->b_vp); 786 if (bp->b_vflags & BV_BKGRDINPROG) { 787 if (bp->b_flags & B_ASYNC) { 788 VI_UNLOCK(bp->b_vp); 789 splx(s); 790 bdwrite(bp); 791 return (0); 792 } 793 bp->b_vflags |= BV_BKGRDWAIT; 794 msleep(&bp->b_xflags, VI_MTX(bp->b_vp), PRIBIO, "bwrbg", 0); 795 if (bp->b_vflags & BV_BKGRDINPROG) 796 panic("bwrite: still writing"); 797 } 798 VI_UNLOCK(bp->b_vp); 799 800 /* Mark the buffer clean */ 801 bundirty(bp); 802 803 /* 804 * If this buffer is marked for background writing and we 805 * do not have to wait for it, make a copy and write the 806 * copy so as to leave this buffer ready for further use. 807 * 808 * This optimization eats a lot of memory. If we have a page 809 * or buffer shortfall we can't do it. 810 */ 811 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 812 (bp->b_flags & B_ASYNC) && 813 !vm_page_count_severe() && 814 !buf_dirty_count_severe()) { 815 if (bp->b_iodone != NULL) { 816 printf("bp->b_iodone = %p\n", bp->b_iodone); 817 panic("bwrite: need chained iodone"); 818 } 819 820 /* get a new block */ 821 newbp = geteblk(bp->b_bufsize); 822 823 /* 824 * set it to be identical to the old block. We have to 825 * set b_lblkno and BKGRDMARKER before calling bgetvp() 826 * to avoid confusing the splay tree and gbincore(). 827 */ 828 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 829 newbp->b_lblkno = bp->b_lblkno; 830 newbp->b_xflags |= BX_BKGRDMARKER; 831 VI_LOCK(bp->b_vp); 832 bp->b_vflags |= BV_BKGRDINPROG; 833 bgetvp(bp->b_vp, newbp); 834 VI_UNLOCK(bp->b_vp); 835 newbp->b_blkno = bp->b_blkno; 836 newbp->b_offset = bp->b_offset; 837 newbp->b_iodone = vfs_backgroundwritedone; 838 newbp->b_flags |= B_ASYNC; 839 newbp->b_flags &= ~B_INVAL; 840 841 /* move over the dependencies */ 842 if (LIST_FIRST(&bp->b_dep) != NULL) 843 buf_movedeps(bp, newbp); 844 845 /* 846 * Initiate write on the copy, release the original to 847 * the B_LOCKED queue so that it cannot go away until 848 * the background write completes. If not locked it could go 849 * away and then be reconstituted while it was being written. 850 * If the reconstituted buffer were written, we could end up 851 * with two background copies being written at the same time. 852 */ 853 bqrelse(bp); 854 bp = newbp; 855 } 856 857 bp->b_flags &= ~B_DONE; 858 bp->b_ioflags &= ~BIO_ERROR; 859 bp->b_flags |= B_WRITEINPROG | B_CACHE; 860 bp->b_iocmd = BIO_WRITE; 861 862 VI_LOCK(bp->b_vp); 863 bp->b_vp->v_numoutput++; 864 VI_UNLOCK(bp->b_vp); 865 vfs_busy_pages(bp, 1); 866 867 /* 868 * Normal bwrites pipeline writes 869 */ 870 bp->b_runningbufspace = bp->b_bufsize; 871 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 872 873 if (curthread != PCPU_GET(idlethread)) 874 curthread->td_proc->p_stats->p_ru.ru_oublock++; 875 splx(s); 876 if (oldflags & B_ASYNC) 877 BUF_KERNPROC(bp); 878 bp->b_iooffset = dbtob(bp->b_blkno); 879 if (bp->b_vp->v_type == VCHR) 880 VOP_SPECSTRATEGY(bp->b_vp, bp); 881 else 882 VOP_STRATEGY(bp->b_vp, bp); 883 884 if ((oldflags & B_ASYNC) == 0) { 885 int rtval = bufwait(bp); 886 brelse(bp); 887 return (rtval); 888 } else { 889 /* 890 * don't allow the async write to saturate the I/O 891 * system. We will not deadlock here because 892 * we are blocking waiting for I/O that is already in-progress 893 * to complete. We do not block here if it is the update 894 * or syncer daemon trying to clean up as that can lead 895 * to deadlock. 896 */ 897 if (curthread->td_proc != bufdaemonproc && 898 curthread->td_proc != updateproc) 899 waitrunningbufspace(); 900 } 901 902 return (0); 903} 904 905/* 906 * Complete a background write started from bwrite. 907 */ 908static void 909vfs_backgroundwritedone(bp) 910 struct buf *bp; 911{ 912 struct buf *origbp; 913 914 /* 915 * Find the original buffer that we are writing. 916 */ 917 VI_LOCK(bp->b_vp); 918 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 919 panic("backgroundwritedone: lost buffer"); 920 921 /* 922 * Clear the BV_BKGRDINPROG flag in the original buffer 923 * and awaken it if it is waiting for the write to complete. 924 * If BV_BKGRDINPROG is not set in the original buffer it must 925 * have been released and re-instantiated - which is not legal. 926 */ 927 KASSERT((origbp->b_vflags & BV_BKGRDINPROG), 928 ("backgroundwritedone: lost buffer2")); 929 origbp->b_vflags &= ~BV_BKGRDINPROG; 930 if (origbp->b_vflags & BV_BKGRDWAIT) { 931 origbp->b_vflags &= ~BV_BKGRDWAIT; 932 wakeup(&origbp->b_xflags); 933 } 934 VI_UNLOCK(bp->b_vp); 935 /* 936 * Process dependencies then return any unfinished ones. 937 */ 938 if (LIST_FIRST(&bp->b_dep) != NULL) 939 buf_complete(bp); 940 if (LIST_FIRST(&bp->b_dep) != NULL) 941 buf_movedeps(bp, origbp); 942 943 /* 944 * This buffer is marked B_NOCACHE, so when it is released 945 * by biodone, it will be tossed. We mark it with BIO_READ 946 * to avoid biodone doing a second vwakeup. 947 */ 948 bp->b_flags |= B_NOCACHE; 949 bp->b_iocmd = BIO_READ; 950 bp->b_flags &= ~(B_CACHE | B_DONE); 951 bp->b_iodone = 0; 952 bufdone(bp); 953} 954 955/* 956 * Delayed write. (Buffer is marked dirty). Do not bother writing 957 * anything if the buffer is marked invalid. 958 * 959 * Note that since the buffer must be completely valid, we can safely 960 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 961 * biodone() in order to prevent getblk from writing the buffer 962 * out synchronously. 963 */ 964void 965bdwrite(struct buf * bp) 966{ 967 struct thread *td = curthread; 968 struct vnode *vp; 969 struct buf *nbp; 970 971 GIANT_REQUIRED; 972 973 if (BUF_REFCNT(bp) == 0) 974 panic("bdwrite: buffer is not busy"); 975 976 if (bp->b_flags & B_INVAL) { 977 brelse(bp); 978 return; 979 } 980 981 /* 982 * If we have too many dirty buffers, don't create any more. 983 * If we are wildly over our limit, then force a complete 984 * cleanup. Otherwise, just keep the situation from getting 985 * out of control. Note that we have to avoid a recursive 986 * disaster and not try to clean up after our own cleanup! 987 */ 988 vp = bp->b_vp; 989 VI_LOCK(vp); 990 if (td->td_pflags & TDP_COWINPROGRESS) { 991 recursiveflushes++; 992 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) { 993 VI_UNLOCK(vp); 994 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 995 VI_LOCK(vp); 996 altbufferflushes++; 997 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) { 998 /* 999 * Try to find a buffer to flush. 1000 */ 1001 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) { 1002 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1003 buf_countdeps(nbp, 0) || 1004 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1005 continue; 1006 if (bp == nbp) 1007 panic("bdwrite: found ourselves"); 1008 VI_UNLOCK(vp); 1009 if (nbp->b_flags & B_CLUSTEROK) { 1010 vfs_bio_awrite(nbp); 1011 } else { 1012 bremfree(nbp); 1013 bawrite(nbp); 1014 } 1015 VI_LOCK(vp); 1016 dirtybufferflushes++; 1017 break; 1018 } 1019 } 1020 VI_UNLOCK(vp); 1021 1022 bdirty(bp); 1023 /* 1024 * Set B_CACHE, indicating that the buffer is fully valid. This is 1025 * true even of NFS now. 1026 */ 1027 bp->b_flags |= B_CACHE; 1028 1029 /* 1030 * This bmap keeps the system from needing to do the bmap later, 1031 * perhaps when the system is attempting to do a sync. Since it 1032 * is likely that the indirect block -- or whatever other datastructure 1033 * that the filesystem needs is still in memory now, it is a good 1034 * thing to do this. Note also, that if the pageout daemon is 1035 * requesting a sync -- there might not be enough memory to do 1036 * the bmap then... So, this is important to do. 1037 */ 1038 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1039 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1040 } 1041 1042 /* 1043 * Set the *dirty* buffer range based upon the VM system dirty pages. 1044 */ 1045 vfs_setdirty(bp); 1046 1047 /* 1048 * We need to do this here to satisfy the vnode_pager and the 1049 * pageout daemon, so that it thinks that the pages have been 1050 * "cleaned". Note that since the pages are in a delayed write 1051 * buffer -- the VFS layer "will" see that the pages get written 1052 * out on the next sync, or perhaps the cluster will be completed. 1053 */ 1054 vfs_clean_pages(bp); 1055 bqrelse(bp); 1056 1057 /* 1058 * Wakeup the buffer flushing daemon if we have a lot of dirty 1059 * buffers (midpoint between our recovery point and our stall 1060 * point). 1061 */ 1062 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1063 1064 /* 1065 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1066 * due to the softdep code. 1067 */ 1068} 1069 1070/* 1071 * bdirty: 1072 * 1073 * Turn buffer into delayed write request. We must clear BIO_READ and 1074 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1075 * itself to properly update it in the dirty/clean lists. We mark it 1076 * B_DONE to ensure that any asynchronization of the buffer properly 1077 * clears B_DONE ( else a panic will occur later ). 1078 * 1079 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1080 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1081 * should only be called if the buffer is known-good. 1082 * 1083 * Since the buffer is not on a queue, we do not update the numfreebuffers 1084 * count. 1085 * 1086 * Must be called at splbio(). 1087 * The buffer must be on QUEUE_NONE. 1088 */ 1089void 1090bdirty(bp) 1091 struct buf *bp; 1092{ 1093 KASSERT(bp->b_qindex == QUEUE_NONE, 1094 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1095 bp->b_flags &= ~(B_RELBUF); 1096 bp->b_iocmd = BIO_WRITE; 1097 1098 if ((bp->b_flags & B_DELWRI) == 0) { 1099 bp->b_flags |= B_DONE | B_DELWRI; 1100 reassignbuf(bp, bp->b_vp); 1101 atomic_add_int(&numdirtybuffers, 1); 1102 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1103 } 1104} 1105 1106/* 1107 * bundirty: 1108 * 1109 * Clear B_DELWRI for buffer. 1110 * 1111 * Since the buffer is not on a queue, we do not update the numfreebuffers 1112 * count. 1113 * 1114 * Must be called at splbio(). 1115 * The buffer must be on QUEUE_NONE. 1116 */ 1117 1118void 1119bundirty(bp) 1120 struct buf *bp; 1121{ 1122 KASSERT(bp->b_qindex == QUEUE_NONE, 1123 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1124 1125 if (bp->b_flags & B_DELWRI) { 1126 bp->b_flags &= ~B_DELWRI; 1127 reassignbuf(bp, bp->b_vp); 1128 atomic_subtract_int(&numdirtybuffers, 1); 1129 numdirtywakeup(lodirtybuffers); 1130 } 1131 /* 1132 * Since it is now being written, we can clear its deferred write flag. 1133 */ 1134 bp->b_flags &= ~B_DEFERRED; 1135} 1136 1137/* 1138 * bawrite: 1139 * 1140 * Asynchronous write. Start output on a buffer, but do not wait for 1141 * it to complete. The buffer is released when the output completes. 1142 * 1143 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1144 * B_INVAL buffers. Not us. 1145 */ 1146void 1147bawrite(struct buf * bp) 1148{ 1149 bp->b_flags |= B_ASYNC; 1150 (void) BUF_WRITE(bp); 1151} 1152 1153/* 1154 * bwillwrite: 1155 * 1156 * Called prior to the locking of any vnodes when we are expecting to 1157 * write. We do not want to starve the buffer cache with too many 1158 * dirty buffers so we block here. By blocking prior to the locking 1159 * of any vnodes we attempt to avoid the situation where a locked vnode 1160 * prevents the various system daemons from flushing related buffers. 1161 */ 1162 1163void 1164bwillwrite(void) 1165{ 1166 if (numdirtybuffers >= hidirtybuffers) { 1167 int s; 1168 1169 mtx_lock(&Giant); 1170 s = splbio(); 1171 mtx_lock(&nblock); 1172 while (numdirtybuffers >= hidirtybuffers) { 1173 bd_wakeup(1); 1174 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1175 msleep(&needsbuffer, &nblock, 1176 (PRIBIO + 4), "flswai", 0); 1177 } 1178 splx(s); 1179 mtx_unlock(&nblock); 1180 mtx_unlock(&Giant); 1181 } 1182} 1183 1184/* 1185 * Return true if we have too many dirty buffers. 1186 */ 1187int 1188buf_dirty_count_severe(void) 1189{ 1190 return(numdirtybuffers >= hidirtybuffers); 1191} 1192 1193/* 1194 * brelse: 1195 * 1196 * Release a busy buffer and, if requested, free its resources. The 1197 * buffer will be stashed in the appropriate bufqueue[] allowing it 1198 * to be accessed later as a cache entity or reused for other purposes. 1199 */ 1200void 1201brelse(struct buf * bp) 1202{ 1203 int s; 1204 1205 GIANT_REQUIRED; 1206 1207 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1208 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1209 1210 s = splbio(); 1211 1212 if (bp->b_iocmd == BIO_WRITE && 1213 (bp->b_ioflags & BIO_ERROR) && 1214 !(bp->b_flags & B_INVAL)) { 1215 /* 1216 * Failed write, redirty. Must clear BIO_ERROR to prevent 1217 * pages from being scrapped. If B_INVAL is set then 1218 * this case is not run and the next case is run to 1219 * destroy the buffer. B_INVAL can occur if the buffer 1220 * is outside the range supported by the underlying device. 1221 */ 1222 bp->b_ioflags &= ~BIO_ERROR; 1223 bdirty(bp); 1224 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1225 (bp->b_ioflags & BIO_ERROR) || 1226 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1227 /* 1228 * Either a failed I/O or we were asked to free or not 1229 * cache the buffer. 1230 */ 1231 bp->b_flags |= B_INVAL; 1232 if (LIST_FIRST(&bp->b_dep) != NULL) 1233 buf_deallocate(bp); 1234 if (bp->b_flags & B_DELWRI) { 1235 atomic_subtract_int(&numdirtybuffers, 1); 1236 numdirtywakeup(lodirtybuffers); 1237 } 1238 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1239 if ((bp->b_flags & B_VMIO) == 0) { 1240 if (bp->b_bufsize) 1241 allocbuf(bp, 0); 1242 if (bp->b_vp) 1243 brelvp(bp); 1244 } 1245 } 1246 1247 /* 1248 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1249 * is called with B_DELWRI set, the underlying pages may wind up 1250 * getting freed causing a previous write (bdwrite()) to get 'lost' 1251 * because pages associated with a B_DELWRI bp are marked clean. 1252 * 1253 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1254 * if B_DELWRI is set. 1255 * 1256 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1257 * on pages to return pages to the VM page queues. 1258 */ 1259 if (bp->b_flags & B_DELWRI) 1260 bp->b_flags &= ~B_RELBUF; 1261 else if (vm_page_count_severe()) { 1262 /* 1263 * XXX This lock may not be necessary since BKGRDINPROG 1264 * cannot be set while we hold the buf lock, it can only be 1265 * cleared if it is already pending. 1266 */ 1267 if (bp->b_vp) { 1268 VI_LOCK(bp->b_vp); 1269 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1270 bp->b_flags |= B_RELBUF; 1271 VI_UNLOCK(bp->b_vp); 1272 } else 1273 bp->b_flags |= B_RELBUF; 1274 } 1275 1276 /* 1277 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1278 * constituted, not even NFS buffers now. Two flags effect this. If 1279 * B_INVAL, the struct buf is invalidated but the VM object is kept 1280 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1281 * 1282 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1283 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1284 * buffer is also B_INVAL because it hits the re-dirtying code above. 1285 * 1286 * Normally we can do this whether a buffer is B_DELWRI or not. If 1287 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1288 * the commit state and we cannot afford to lose the buffer. If the 1289 * buffer has a background write in progress, we need to keep it 1290 * around to prevent it from being reconstituted and starting a second 1291 * background write. 1292 */ 1293 if ((bp->b_flags & B_VMIO) 1294 && !(bp->b_vp->v_mount != NULL && 1295 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1296 !vn_isdisk(bp->b_vp, NULL) && 1297 (bp->b_flags & B_DELWRI)) 1298 ) { 1299 1300 int i, j, resid; 1301 vm_page_t m; 1302 off_t foff; 1303 vm_pindex_t poff; 1304 vm_object_t obj; 1305 struct vnode *vp; 1306 1307 vp = bp->b_vp; 1308 obj = bp->b_object; 1309 1310 /* 1311 * Get the base offset and length of the buffer. Note that 1312 * in the VMIO case if the buffer block size is not 1313 * page-aligned then b_data pointer may not be page-aligned. 1314 * But our b_pages[] array *IS* page aligned. 1315 * 1316 * block sizes less then DEV_BSIZE (usually 512) are not 1317 * supported due to the page granularity bits (m->valid, 1318 * m->dirty, etc...). 1319 * 1320 * See man buf(9) for more information 1321 */ 1322 resid = bp->b_bufsize; 1323 foff = bp->b_offset; 1324 VM_OBJECT_LOCK(obj); 1325 for (i = 0; i < bp->b_npages; i++) { 1326 int had_bogus = 0; 1327 1328 m = bp->b_pages[i]; 1329 vm_page_lock_queues(); 1330 vm_page_flag_clear(m, PG_ZERO); 1331 vm_page_unlock_queues(); 1332 1333 /* 1334 * If we hit a bogus page, fixup *all* the bogus pages 1335 * now. 1336 */ 1337 if (m == bogus_page) { 1338 poff = OFF_TO_IDX(bp->b_offset); 1339 had_bogus = 1; 1340 1341 for (j = i; j < bp->b_npages; j++) { 1342 vm_page_t mtmp; 1343 mtmp = bp->b_pages[j]; 1344 if (mtmp == bogus_page) { 1345 mtmp = vm_page_lookup(obj, poff + j); 1346 if (!mtmp) { 1347 panic("brelse: page missing\n"); 1348 } 1349 bp->b_pages[j] = mtmp; 1350 } 1351 } 1352 1353 if ((bp->b_flags & B_INVAL) == 0) { 1354 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1355 } 1356 m = bp->b_pages[i]; 1357 } 1358 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1359 int poffset = foff & PAGE_MASK; 1360 int presid = resid > (PAGE_SIZE - poffset) ? 1361 (PAGE_SIZE - poffset) : resid; 1362 1363 KASSERT(presid >= 0, ("brelse: extra page")); 1364 vm_page_lock_queues(); 1365 vm_page_set_invalid(m, poffset, presid); 1366 vm_page_unlock_queues(); 1367 if (had_bogus) 1368 printf("avoided corruption bug in bogus_page/brelse code\n"); 1369 } 1370 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1371 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1372 } 1373 VM_OBJECT_UNLOCK(obj); 1374 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1375 vfs_vmio_release(bp); 1376 1377 } else if (bp->b_flags & B_VMIO) { 1378 1379 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1380 vfs_vmio_release(bp); 1381 } 1382 1383 } 1384 1385 if (bp->b_qindex != QUEUE_NONE) 1386 panic("brelse: free buffer onto another queue???"); 1387 if (BUF_REFCNT(bp) > 1) { 1388 /* do not release to free list */ 1389 BUF_UNLOCK(bp); 1390 splx(s); 1391 return; 1392 } 1393 1394 /* enqueue */ 1395 mtx_lock(&bqlock); 1396 1397 /* buffers with no memory */ 1398 if (bp->b_bufsize == 0) { 1399 bp->b_flags |= B_INVAL; 1400 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1401 if (bp->b_vflags & BV_BKGRDINPROG) 1402 panic("losing buffer 1"); 1403 if (bp->b_kvasize) { 1404 bp->b_qindex = QUEUE_EMPTYKVA; 1405 } else { 1406 bp->b_qindex = QUEUE_EMPTY; 1407 } 1408 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1409 bp->b_dev = NODEV; 1410 /* buffers with junk contents */ 1411 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1412 (bp->b_ioflags & BIO_ERROR)) { 1413 bp->b_flags |= B_INVAL; 1414 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1415 if (bp->b_vflags & BV_BKGRDINPROG) 1416 panic("losing buffer 2"); 1417 bp->b_qindex = QUEUE_CLEAN; 1418 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1419 bp->b_dev = NODEV; 1420 /* remaining buffers */ 1421 } else { 1422 if (bp->b_flags & B_DELWRI) 1423 bp->b_qindex = QUEUE_DIRTY; 1424 else 1425 bp->b_qindex = QUEUE_CLEAN; 1426 if (bp->b_flags & B_AGE) 1427 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1428 else 1429 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1430 } 1431 mtx_unlock(&bqlock); 1432 1433 /* 1434 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1435 * placed the buffer on the correct queue. We must also disassociate 1436 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1437 * find it. 1438 */ 1439 if (bp->b_flags & B_INVAL) { 1440 if (bp->b_flags & B_DELWRI) 1441 bundirty(bp); 1442 if (bp->b_vp) 1443 brelvp(bp); 1444 } 1445 1446 /* 1447 * Fixup numfreebuffers count. The bp is on an appropriate queue 1448 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1449 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1450 * if B_INVAL is set ). 1451 */ 1452 1453 if (!(bp->b_flags & B_DELWRI)) 1454 bufcountwakeup(); 1455 1456 /* 1457 * Something we can maybe free or reuse 1458 */ 1459 if (bp->b_bufsize || bp->b_kvasize) 1460 bufspacewakeup(); 1461 1462 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1463 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1464 panic("brelse: not dirty"); 1465 /* unlock */ 1466 BUF_UNLOCK(bp); 1467 splx(s); 1468} 1469 1470/* 1471 * Release a buffer back to the appropriate queue but do not try to free 1472 * it. The buffer is expected to be used again soon. 1473 * 1474 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1475 * biodone() to requeue an async I/O on completion. It is also used when 1476 * known good buffers need to be requeued but we think we may need the data 1477 * again soon. 1478 * 1479 * XXX we should be able to leave the B_RELBUF hint set on completion. 1480 */ 1481void 1482bqrelse(struct buf * bp) 1483{ 1484 int s; 1485 1486 s = splbio(); 1487 1488 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1489 1490 if (bp->b_qindex != QUEUE_NONE) 1491 panic("bqrelse: free buffer onto another queue???"); 1492 if (BUF_REFCNT(bp) > 1) { 1493 /* do not release to free list */ 1494 BUF_UNLOCK(bp); 1495 splx(s); 1496 return; 1497 } 1498 mtx_lock(&bqlock); 1499 /* buffers with stale but valid contents */ 1500 if (bp->b_flags & B_DELWRI) { 1501 bp->b_qindex = QUEUE_DIRTY; 1502 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1503 } else { 1504 /* 1505 * XXX This lock may not be necessary since BKGRDINPROG 1506 * cannot be set while we hold the buf lock, it can only be 1507 * cleared if it is already pending. 1508 */ 1509 VI_LOCK(bp->b_vp); 1510 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1511 VI_UNLOCK(bp->b_vp); 1512 bp->b_qindex = QUEUE_CLEAN; 1513 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1514 b_freelist); 1515 } else { 1516 /* 1517 * We are too low on memory, we have to try to free 1518 * the buffer (most importantly: the wired pages 1519 * making up its backing store) *now*. 1520 */ 1521 VI_UNLOCK(bp->b_vp); 1522 mtx_unlock(&bqlock); 1523 splx(s); 1524 brelse(bp); 1525 return; 1526 } 1527 } 1528 mtx_unlock(&bqlock); 1529 1530 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1531 bufcountwakeup(); 1532 1533 /* 1534 * Something we can maybe free or reuse. 1535 */ 1536 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1537 bufspacewakeup(); 1538 1539 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1540 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1541 panic("bqrelse: not dirty"); 1542 /* unlock */ 1543 BUF_UNLOCK(bp); 1544 splx(s); 1545} 1546 1547/* Give pages used by the bp back to the VM system (where possible) */ 1548static void 1549vfs_vmio_release(bp) 1550 struct buf *bp; 1551{ 1552 int i; 1553 vm_page_t m; 1554 1555 GIANT_REQUIRED; 1556 VM_OBJECT_LOCK(bp->b_object); 1557 vm_page_lock_queues(); 1558 for (i = 0; i < bp->b_npages; i++) { 1559 m = bp->b_pages[i]; 1560 bp->b_pages[i] = NULL; 1561 /* 1562 * In order to keep page LRU ordering consistent, put 1563 * everything on the inactive queue. 1564 */ 1565 vm_page_unwire(m, 0); 1566 /* 1567 * We don't mess with busy pages, it is 1568 * the responsibility of the process that 1569 * busied the pages to deal with them. 1570 */ 1571 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1572 continue; 1573 1574 if (m->wire_count == 0) { 1575 vm_page_flag_clear(m, PG_ZERO); 1576 /* 1577 * Might as well free the page if we can and it has 1578 * no valid data. We also free the page if the 1579 * buffer was used for direct I/O 1580 */ 1581 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1582 m->hold_count == 0) { 1583 vm_page_busy(m); 1584 pmap_remove_all(m); 1585 vm_page_free(m); 1586 } else if (bp->b_flags & B_DIRECT) { 1587 vm_page_try_to_free(m); 1588 } else if (vm_page_count_severe()) { 1589 vm_page_try_to_cache(m); 1590 } 1591 } 1592 } 1593 vm_page_unlock_queues(); 1594 VM_OBJECT_UNLOCK(bp->b_object); 1595 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1596 1597 if (bp->b_bufsize) { 1598 bufspacewakeup(); 1599 bp->b_bufsize = 0; 1600 } 1601 bp->b_npages = 0; 1602 bp->b_flags &= ~B_VMIO; 1603 if (bp->b_vp) 1604 brelvp(bp); 1605} 1606 1607/* 1608 * Check to see if a block at a particular lbn is available for a clustered 1609 * write. 1610 */ 1611static int 1612vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1613{ 1614 struct buf *bpa; 1615 int match; 1616 1617 match = 0; 1618 1619 /* If the buf isn't in core skip it */ 1620 if ((bpa = gbincore(vp, lblkno)) == NULL) 1621 return (0); 1622 1623 /* If the buf is busy we don't want to wait for it */ 1624 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1625 return (0); 1626 1627 /* Only cluster with valid clusterable delayed write buffers */ 1628 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1629 (B_DELWRI | B_CLUSTEROK)) 1630 goto done; 1631 1632 if (bpa->b_bufsize != size) 1633 goto done; 1634 1635 /* 1636 * Check to see if it is in the expected place on disk and that the 1637 * block has been mapped. 1638 */ 1639 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1640 match = 1; 1641done: 1642 BUF_UNLOCK(bpa); 1643 return (match); 1644} 1645 1646/* 1647 * vfs_bio_awrite: 1648 * 1649 * Implement clustered async writes for clearing out B_DELWRI buffers. 1650 * This is much better then the old way of writing only one buffer at 1651 * a time. Note that we may not be presented with the buffers in the 1652 * correct order, so we search for the cluster in both directions. 1653 */ 1654int 1655vfs_bio_awrite(struct buf * bp) 1656{ 1657 int i; 1658 int j; 1659 daddr_t lblkno = bp->b_lblkno; 1660 struct vnode *vp = bp->b_vp; 1661 int s; 1662 int ncl; 1663 int nwritten; 1664 int size; 1665 int maxcl; 1666 1667 s = splbio(); 1668 /* 1669 * right now we support clustered writing only to regular files. If 1670 * we find a clusterable block we could be in the middle of a cluster 1671 * rather then at the beginning. 1672 */ 1673 if ((vp->v_type == VREG) && 1674 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1675 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1676 1677 size = vp->v_mount->mnt_stat.f_iosize; 1678 maxcl = MAXPHYS / size; 1679 1680 VI_LOCK(vp); 1681 for (i = 1; i < maxcl; i++) 1682 if (vfs_bio_clcheck(vp, size, lblkno + i, 1683 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1684 break; 1685 1686 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1687 if (vfs_bio_clcheck(vp, size, lblkno - j, 1688 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1689 break; 1690 1691 VI_UNLOCK(vp); 1692 --j; 1693 ncl = i + j; 1694 /* 1695 * this is a possible cluster write 1696 */ 1697 if (ncl != 1) { 1698 BUF_UNLOCK(bp); 1699 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1700 splx(s); 1701 return nwritten; 1702 } 1703 } 1704 1705 bremfree(bp); 1706 bp->b_flags |= B_ASYNC; 1707 1708 splx(s); 1709 /* 1710 * default (old) behavior, writing out only one block 1711 * 1712 * XXX returns b_bufsize instead of b_bcount for nwritten? 1713 */ 1714 nwritten = bp->b_bufsize; 1715 (void) BUF_WRITE(bp); 1716 1717 return nwritten; 1718} 1719 1720/* 1721 * getnewbuf: 1722 * 1723 * Find and initialize a new buffer header, freeing up existing buffers 1724 * in the bufqueues as necessary. The new buffer is returned locked. 1725 * 1726 * Important: B_INVAL is not set. If the caller wishes to throw the 1727 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1728 * 1729 * We block if: 1730 * We have insufficient buffer headers 1731 * We have insufficient buffer space 1732 * buffer_map is too fragmented ( space reservation fails ) 1733 * If we have to flush dirty buffers ( but we try to avoid this ) 1734 * 1735 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1736 * Instead we ask the buf daemon to do it for us. We attempt to 1737 * avoid piecemeal wakeups of the pageout daemon. 1738 */ 1739 1740static struct buf * 1741getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1742{ 1743 struct buf *bp; 1744 struct buf *nbp; 1745 int defrag = 0; 1746 int nqindex; 1747 static int flushingbufs; 1748 1749 GIANT_REQUIRED; 1750 1751 /* 1752 * We can't afford to block since we might be holding a vnode lock, 1753 * which may prevent system daemons from running. We deal with 1754 * low-memory situations by proactively returning memory and running 1755 * async I/O rather then sync I/O. 1756 */ 1757 1758 atomic_add_int(&getnewbufcalls, 1); 1759 atomic_subtract_int(&getnewbufrestarts, 1); 1760restart: 1761 atomic_add_int(&getnewbufrestarts, 1); 1762 1763 /* 1764 * Setup for scan. If we do not have enough free buffers, 1765 * we setup a degenerate case that immediately fails. Note 1766 * that if we are specially marked process, we are allowed to 1767 * dip into our reserves. 1768 * 1769 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1770 * 1771 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1772 * However, there are a number of cases (defragging, reusing, ...) 1773 * where we cannot backup. 1774 */ 1775 mtx_lock(&bqlock); 1776 nqindex = QUEUE_EMPTYKVA; 1777 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1778 1779 if (nbp == NULL) { 1780 /* 1781 * If no EMPTYKVA buffers and we are either 1782 * defragging or reusing, locate a CLEAN buffer 1783 * to free or reuse. If bufspace useage is low 1784 * skip this step so we can allocate a new buffer. 1785 */ 1786 if (defrag || bufspace >= lobufspace) { 1787 nqindex = QUEUE_CLEAN; 1788 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1789 } 1790 1791 /* 1792 * If we could not find or were not allowed to reuse a 1793 * CLEAN buffer, check to see if it is ok to use an EMPTY 1794 * buffer. We can only use an EMPTY buffer if allocating 1795 * its KVA would not otherwise run us out of buffer space. 1796 */ 1797 if (nbp == NULL && defrag == 0 && 1798 bufspace + maxsize < hibufspace) { 1799 nqindex = QUEUE_EMPTY; 1800 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1801 } 1802 } 1803 1804 /* 1805 * Run scan, possibly freeing data and/or kva mappings on the fly 1806 * depending. 1807 */ 1808 1809 while ((bp = nbp) != NULL) { 1810 int qindex = nqindex; 1811 1812 /* 1813 * Calculate next bp ( we can only use it if we do not block 1814 * or do other fancy things ). 1815 */ 1816 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1817 switch(qindex) { 1818 case QUEUE_EMPTY: 1819 nqindex = QUEUE_EMPTYKVA; 1820 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1821 break; 1822 /* FALLTHROUGH */ 1823 case QUEUE_EMPTYKVA: 1824 nqindex = QUEUE_CLEAN; 1825 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1826 break; 1827 /* FALLTHROUGH */ 1828 case QUEUE_CLEAN: 1829 /* 1830 * nbp is NULL. 1831 */ 1832 break; 1833 } 1834 } 1835 if (bp->b_vp) { 1836 VI_LOCK(bp->b_vp); 1837 if (bp->b_vflags & BV_BKGRDINPROG) { 1838 VI_UNLOCK(bp->b_vp); 1839 continue; 1840 } 1841 VI_UNLOCK(bp->b_vp); 1842 } 1843 1844 /* 1845 * Sanity Checks 1846 */ 1847 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1848 1849 /* 1850 * Note: we no longer distinguish between VMIO and non-VMIO 1851 * buffers. 1852 */ 1853 1854 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1855 1856 /* 1857 * If we are defragging then we need a buffer with 1858 * b_kvasize != 0. XXX this situation should no longer 1859 * occur, if defrag is non-zero the buffer's b_kvasize 1860 * should also be non-zero at this point. XXX 1861 */ 1862 if (defrag && bp->b_kvasize == 0) { 1863 printf("Warning: defrag empty buffer %p\n", bp); 1864 continue; 1865 } 1866 1867 /* 1868 * Start freeing the bp. This is somewhat involved. nbp 1869 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1870 */ 1871 1872 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1873 panic("getnewbuf: locked buf"); 1874 bremfreel(bp); 1875 mtx_unlock(&bqlock); 1876 1877 if (qindex == QUEUE_CLEAN) { 1878 if (bp->b_flags & B_VMIO) { 1879 bp->b_flags &= ~B_ASYNC; 1880 vfs_vmio_release(bp); 1881 } 1882 if (bp->b_vp) 1883 brelvp(bp); 1884 } 1885 1886 /* 1887 * NOTE: nbp is now entirely invalid. We can only restart 1888 * the scan from this point on. 1889 * 1890 * Get the rest of the buffer freed up. b_kva* is still 1891 * valid after this operation. 1892 */ 1893 1894 if (bp->b_rcred != NOCRED) { 1895 crfree(bp->b_rcred); 1896 bp->b_rcred = NOCRED; 1897 } 1898 if (bp->b_wcred != NOCRED) { 1899 crfree(bp->b_wcred); 1900 bp->b_wcred = NOCRED; 1901 } 1902 if (LIST_FIRST(&bp->b_dep) != NULL) 1903 buf_deallocate(bp); 1904 if (bp->b_vflags & BV_BKGRDINPROG) 1905 panic("losing buffer 3"); 1906 1907 if (bp->b_bufsize) 1908 allocbuf(bp, 0); 1909 1910 bp->b_flags = 0; 1911 bp->b_ioflags = 0; 1912 bp->b_xflags = 0; 1913 bp->b_vflags = 0; 1914 bp->b_dev = NODEV; 1915 bp->b_vp = NULL; 1916 bp->b_blkno = bp->b_lblkno = 0; 1917 bp->b_offset = NOOFFSET; 1918 bp->b_iodone = 0; 1919 bp->b_error = 0; 1920 bp->b_resid = 0; 1921 bp->b_bcount = 0; 1922 bp->b_npages = 0; 1923 bp->b_dirtyoff = bp->b_dirtyend = 0; 1924 bp->b_magic = B_MAGIC_BIO; 1925 bp->b_op = &buf_ops_bio; 1926 bp->b_object = NULL; 1927 1928 LIST_INIT(&bp->b_dep); 1929 1930 /* 1931 * If we are defragging then free the buffer. 1932 */ 1933 if (defrag) { 1934 bp->b_flags |= B_INVAL; 1935 bfreekva(bp); 1936 brelse(bp); 1937 defrag = 0; 1938 goto restart; 1939 } 1940 1941 /* 1942 * If we are overcomitted then recover the buffer and its 1943 * KVM space. This occurs in rare situations when multiple 1944 * processes are blocked in getnewbuf() or allocbuf(). 1945 */ 1946 if (bufspace >= hibufspace) 1947 flushingbufs = 1; 1948 if (flushingbufs && bp->b_kvasize != 0) { 1949 bp->b_flags |= B_INVAL; 1950 bfreekva(bp); 1951 brelse(bp); 1952 goto restart; 1953 } 1954 if (bufspace < lobufspace) 1955 flushingbufs = 0; 1956 break; 1957 } 1958 1959 /* 1960 * If we exhausted our list, sleep as appropriate. We may have to 1961 * wakeup various daemons and write out some dirty buffers. 1962 * 1963 * Generally we are sleeping due to insufficient buffer space. 1964 */ 1965 1966 if (bp == NULL) { 1967 int flags; 1968 char *waitmsg; 1969 1970 mtx_unlock(&bqlock); 1971 if (defrag) { 1972 flags = VFS_BIO_NEED_BUFSPACE; 1973 waitmsg = "nbufkv"; 1974 } else if (bufspace >= hibufspace) { 1975 waitmsg = "nbufbs"; 1976 flags = VFS_BIO_NEED_BUFSPACE; 1977 } else { 1978 waitmsg = "newbuf"; 1979 flags = VFS_BIO_NEED_ANY; 1980 } 1981 1982 bd_speedup(); /* heeeelp */ 1983 1984 mtx_lock(&nblock); 1985 needsbuffer |= flags; 1986 while (needsbuffer & flags) { 1987 if (msleep(&needsbuffer, &nblock, 1988 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1989 mtx_unlock(&nblock); 1990 return (NULL); 1991 } 1992 } 1993 mtx_unlock(&nblock); 1994 } else { 1995 /* 1996 * We finally have a valid bp. We aren't quite out of the 1997 * woods, we still have to reserve kva space. In order 1998 * to keep fragmentation sane we only allocate kva in 1999 * BKVASIZE chunks. 2000 */ 2001 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2002 2003 if (maxsize != bp->b_kvasize) { 2004 vm_offset_t addr = 0; 2005 2006 bfreekva(bp); 2007 2008 if (vm_map_findspace(buffer_map, 2009 vm_map_min(buffer_map), maxsize, &addr)) { 2010 /* 2011 * Uh oh. Buffer map is to fragmented. We 2012 * must defragment the map. 2013 */ 2014 atomic_add_int(&bufdefragcnt, 1); 2015 defrag = 1; 2016 bp->b_flags |= B_INVAL; 2017 brelse(bp); 2018 goto restart; 2019 } 2020 if (addr) { 2021 vm_map_insert(buffer_map, NULL, 0, 2022 addr, addr + maxsize, 2023 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2024 2025 bp->b_kvabase = (caddr_t) addr; 2026 bp->b_kvasize = maxsize; 2027 atomic_add_int(&bufspace, bp->b_kvasize); 2028 atomic_add_int(&bufreusecnt, 1); 2029 } 2030 } 2031 bp->b_saveaddr = bp->b_kvabase; 2032 bp->b_data = bp->b_saveaddr; 2033 } 2034 return(bp); 2035} 2036 2037/* 2038 * buf_daemon: 2039 * 2040 * buffer flushing daemon. Buffers are normally flushed by the 2041 * update daemon but if it cannot keep up this process starts to 2042 * take the load in an attempt to prevent getnewbuf() from blocking. 2043 */ 2044 2045static struct kproc_desc buf_kp = { 2046 "bufdaemon", 2047 buf_daemon, 2048 &bufdaemonproc 2049}; 2050SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2051 2052static void 2053buf_daemon() 2054{ 2055 int s; 2056 2057 mtx_lock(&Giant); 2058 2059 /* 2060 * This process needs to be suspended prior to shutdown sync. 2061 */ 2062 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2063 SHUTDOWN_PRI_LAST); 2064 2065 /* 2066 * This process is allowed to take the buffer cache to the limit 2067 */ 2068 s = splbio(); 2069 mtx_lock(&bdlock); 2070 2071 for (;;) { 2072 bd_request = 0; 2073 mtx_unlock(&bdlock); 2074 2075 kthread_suspend_check(bufdaemonproc); 2076 2077 /* 2078 * Do the flush. Limit the amount of in-transit I/O we 2079 * allow to build up, otherwise we would completely saturate 2080 * the I/O system. Wakeup any waiting processes before we 2081 * normally would so they can run in parallel with our drain. 2082 */ 2083 while (numdirtybuffers > lodirtybuffers) { 2084 if (flushbufqueues(0) == 0) { 2085 /* 2086 * Could not find any buffers without rollback 2087 * dependencies, so just write the first one 2088 * in the hopes of eventually making progress. 2089 */ 2090 flushbufqueues(1); 2091 break; 2092 } 2093 waitrunningbufspace(); 2094 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2095 } 2096 2097 /* 2098 * Only clear bd_request if we have reached our low water 2099 * mark. The buf_daemon normally waits 1 second and 2100 * then incrementally flushes any dirty buffers that have 2101 * built up, within reason. 2102 * 2103 * If we were unable to hit our low water mark and couldn't 2104 * find any flushable buffers, we sleep half a second. 2105 * Otherwise we loop immediately. 2106 */ 2107 mtx_lock(&bdlock); 2108 if (numdirtybuffers <= lodirtybuffers) { 2109 /* 2110 * We reached our low water mark, reset the 2111 * request and sleep until we are needed again. 2112 * The sleep is just so the suspend code works. 2113 */ 2114 bd_request = 0; 2115 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2116 } else { 2117 /* 2118 * We couldn't find any flushable dirty buffers but 2119 * still have too many dirty buffers, we 2120 * have to sleep and try again. (rare) 2121 */ 2122 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2123 } 2124 } 2125} 2126 2127/* 2128 * flushbufqueues: 2129 * 2130 * Try to flush a buffer in the dirty queue. We must be careful to 2131 * free up B_INVAL buffers instead of write them, which NFS is 2132 * particularly sensitive to. 2133 */ 2134int flushwithdeps = 0; 2135SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2136 0, "Number of buffers flushed with dependecies that require rollbacks"); 2137static int 2138flushbufqueues(int flushdeps) 2139{ 2140 struct thread *td = curthread; 2141 struct vnode *vp; 2142 struct mount *mp; 2143 struct buf *bp; 2144 int hasdeps; 2145 2146 mtx_lock(&bqlock); 2147 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2148 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2149 continue; 2150 KASSERT((bp->b_flags & B_DELWRI), 2151 ("unexpected clean buffer %p", bp)); 2152 VI_LOCK(bp->b_vp); 2153 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 2154 VI_UNLOCK(bp->b_vp); 2155 BUF_UNLOCK(bp); 2156 continue; 2157 } 2158 VI_UNLOCK(bp->b_vp); 2159 if (bp->b_flags & B_INVAL) { 2160 bremfreel(bp); 2161 mtx_unlock(&bqlock); 2162 brelse(bp); 2163 return (1); 2164 } 2165 2166 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2167 if (flushdeps == 0) { 2168 BUF_UNLOCK(bp); 2169 continue; 2170 } 2171 hasdeps = 1; 2172 } else 2173 hasdeps = 0; 2174 /* 2175 * We must hold the lock on a vnode before writing 2176 * one of its buffers. Otherwise we may confuse, or 2177 * in the case of a snapshot vnode, deadlock the 2178 * system. 2179 * 2180 * The lock order here is the reverse of the normal 2181 * of vnode followed by buf lock. This is ok because 2182 * the NOWAIT will prevent deadlock. 2183 */ 2184 vp = bp->b_vp; 2185 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2186 BUF_UNLOCK(bp); 2187 continue; 2188 } 2189 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2190 mtx_unlock(&bqlock); 2191 vfs_bio_awrite(bp); 2192 vn_finished_write(mp); 2193 VOP_UNLOCK(vp, 0, td); 2194 flushwithdeps += hasdeps; 2195 return (1); 2196 } 2197 vn_finished_write(mp); 2198 BUF_UNLOCK(bp); 2199 } 2200 mtx_unlock(&bqlock); 2201 return (0); 2202} 2203 2204/* 2205 * Check to see if a block is currently memory resident. 2206 */ 2207struct buf * 2208incore(struct vnode * vp, daddr_t blkno) 2209{ 2210 struct buf *bp; 2211 2212 int s = splbio(); 2213 VI_LOCK(vp); 2214 bp = gbincore(vp, blkno); 2215 VI_UNLOCK(vp); 2216 splx(s); 2217 return (bp); 2218} 2219 2220/* 2221 * Returns true if no I/O is needed to access the 2222 * associated VM object. This is like incore except 2223 * it also hunts around in the VM system for the data. 2224 */ 2225 2226int 2227inmem(struct vnode * vp, daddr_t blkno) 2228{ 2229 vm_object_t obj; 2230 vm_offset_t toff, tinc, size; 2231 vm_page_t m; 2232 vm_ooffset_t off; 2233 2234 GIANT_REQUIRED; 2235 ASSERT_VOP_LOCKED(vp, "inmem"); 2236 2237 if (incore(vp, blkno)) 2238 return 1; 2239 if (vp->v_mount == NULL) 2240 return 0; 2241 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2242 return 0; 2243 2244 size = PAGE_SIZE; 2245 if (size > vp->v_mount->mnt_stat.f_iosize) 2246 size = vp->v_mount->mnt_stat.f_iosize; 2247 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2248 2249 VM_OBJECT_LOCK(obj); 2250 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2251 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2252 if (!m) 2253 goto notinmem; 2254 tinc = size; 2255 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2256 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2257 if (vm_page_is_valid(m, 2258 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2259 goto notinmem; 2260 } 2261 VM_OBJECT_UNLOCK(obj); 2262 return 1; 2263 2264notinmem: 2265 VM_OBJECT_UNLOCK(obj); 2266 return (0); 2267} 2268 2269/* 2270 * vfs_setdirty: 2271 * 2272 * Sets the dirty range for a buffer based on the status of the dirty 2273 * bits in the pages comprising the buffer. 2274 * 2275 * The range is limited to the size of the buffer. 2276 * 2277 * This routine is primarily used by NFS, but is generalized for the 2278 * B_VMIO case. 2279 */ 2280static void 2281vfs_setdirty(struct buf *bp) 2282{ 2283 int i; 2284 vm_object_t object; 2285 2286 GIANT_REQUIRED; 2287 /* 2288 * Degenerate case - empty buffer 2289 */ 2290 2291 if (bp->b_bufsize == 0) 2292 return; 2293 2294 /* 2295 * We qualify the scan for modified pages on whether the 2296 * object has been flushed yet. The OBJ_WRITEABLE flag 2297 * is not cleared simply by protecting pages off. 2298 */ 2299 2300 if ((bp->b_flags & B_VMIO) == 0) 2301 return; 2302 2303 object = bp->b_pages[0]->object; 2304 VM_OBJECT_LOCK(object); 2305 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2306 printf("Warning: object %p writeable but not mightbedirty\n", object); 2307 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2308 printf("Warning: object %p mightbedirty but not writeable\n", object); 2309 2310 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2311 vm_offset_t boffset; 2312 vm_offset_t eoffset; 2313 2314 vm_page_lock_queues(); 2315 /* 2316 * test the pages to see if they have been modified directly 2317 * by users through the VM system. 2318 */ 2319 for (i = 0; i < bp->b_npages; i++) { 2320 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2321 vm_page_test_dirty(bp->b_pages[i]); 2322 } 2323 2324 /* 2325 * Calculate the encompassing dirty range, boffset and eoffset, 2326 * (eoffset - boffset) bytes. 2327 */ 2328 2329 for (i = 0; i < bp->b_npages; i++) { 2330 if (bp->b_pages[i]->dirty) 2331 break; 2332 } 2333 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2334 2335 for (i = bp->b_npages - 1; i >= 0; --i) { 2336 if (bp->b_pages[i]->dirty) { 2337 break; 2338 } 2339 } 2340 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2341 2342 vm_page_unlock_queues(); 2343 /* 2344 * Fit it to the buffer. 2345 */ 2346 2347 if (eoffset > bp->b_bcount) 2348 eoffset = bp->b_bcount; 2349 2350 /* 2351 * If we have a good dirty range, merge with the existing 2352 * dirty range. 2353 */ 2354 2355 if (boffset < eoffset) { 2356 if (bp->b_dirtyoff > boffset) 2357 bp->b_dirtyoff = boffset; 2358 if (bp->b_dirtyend < eoffset) 2359 bp->b_dirtyend = eoffset; 2360 } 2361 } 2362 VM_OBJECT_UNLOCK(object); 2363} 2364 2365/* 2366 * getblk: 2367 * 2368 * Get a block given a specified block and offset into a file/device. 2369 * The buffers B_DONE bit will be cleared on return, making it almost 2370 * ready for an I/O initiation. B_INVAL may or may not be set on 2371 * return. The caller should clear B_INVAL prior to initiating a 2372 * READ. 2373 * 2374 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2375 * an existing buffer. 2376 * 2377 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2378 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2379 * and then cleared based on the backing VM. If the previous buffer is 2380 * non-0-sized but invalid, B_CACHE will be cleared. 2381 * 2382 * If getblk() must create a new buffer, the new buffer is returned with 2383 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2384 * case it is returned with B_INVAL clear and B_CACHE set based on the 2385 * backing VM. 2386 * 2387 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2388 * B_CACHE bit is clear. 2389 * 2390 * What this means, basically, is that the caller should use B_CACHE to 2391 * determine whether the buffer is fully valid or not and should clear 2392 * B_INVAL prior to issuing a read. If the caller intends to validate 2393 * the buffer by loading its data area with something, the caller needs 2394 * to clear B_INVAL. If the caller does this without issuing an I/O, 2395 * the caller should set B_CACHE ( as an optimization ), else the caller 2396 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2397 * a write attempt or if it was a successfull read. If the caller 2398 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2399 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2400 */ 2401struct buf * 2402getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2403 int flags) 2404{ 2405 struct buf *bp; 2406 int s; 2407 int error; 2408 ASSERT_VOP_LOCKED(vp, "getblk"); 2409 2410 if (size > MAXBSIZE) 2411 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2412 2413 s = splbio(); 2414loop: 2415 /* 2416 * Block if we are low on buffers. Certain processes are allowed 2417 * to completely exhaust the buffer cache. 2418 * 2419 * If this check ever becomes a bottleneck it may be better to 2420 * move it into the else, when gbincore() fails. At the moment 2421 * it isn't a problem. 2422 * 2423 * XXX remove if 0 sections (clean this up after its proven) 2424 */ 2425 if (numfreebuffers == 0) { 2426 if (curthread == PCPU_GET(idlethread)) 2427 return NULL; 2428 mtx_lock(&nblock); 2429 needsbuffer |= VFS_BIO_NEED_ANY; 2430 mtx_unlock(&nblock); 2431 } 2432 2433 VI_LOCK(vp); 2434 if ((bp = gbincore(vp, blkno))) { 2435 int lockflags; 2436 /* 2437 * Buffer is in-core. If the buffer is not busy, it must 2438 * be on a queue. 2439 */ 2440 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2441 2442 if (flags & GB_LOCK_NOWAIT) 2443 lockflags |= LK_NOWAIT; 2444 2445 error = BUF_TIMELOCK(bp, lockflags, 2446 VI_MTX(vp), "getblk", slpflag, slptimeo); 2447 2448 /* 2449 * If we slept and got the lock we have to restart in case 2450 * the buffer changed identities. 2451 */ 2452 if (error == ENOLCK) 2453 goto loop; 2454 /* We timed out or were interrupted. */ 2455 else if (error) 2456 return (NULL); 2457 2458 /* 2459 * The buffer is locked. B_CACHE is cleared if the buffer is 2460 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2461 * and for a VMIO buffer B_CACHE is adjusted according to the 2462 * backing VM cache. 2463 */ 2464 if (bp->b_flags & B_INVAL) 2465 bp->b_flags &= ~B_CACHE; 2466 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2467 bp->b_flags |= B_CACHE; 2468 bremfree(bp); 2469 2470 /* 2471 * check for size inconsistancies for non-VMIO case. 2472 */ 2473 2474 if (bp->b_bcount != size) { 2475 if ((bp->b_flags & B_VMIO) == 0 || 2476 (size > bp->b_kvasize)) { 2477 if (bp->b_flags & B_DELWRI) { 2478 bp->b_flags |= B_NOCACHE; 2479 BUF_WRITE(bp); 2480 } else { 2481 if ((bp->b_flags & B_VMIO) && 2482 (LIST_FIRST(&bp->b_dep) == NULL)) { 2483 bp->b_flags |= B_RELBUF; 2484 brelse(bp); 2485 } else { 2486 bp->b_flags |= B_NOCACHE; 2487 BUF_WRITE(bp); 2488 } 2489 } 2490 goto loop; 2491 } 2492 } 2493 2494 /* 2495 * If the size is inconsistant in the VMIO case, we can resize 2496 * the buffer. This might lead to B_CACHE getting set or 2497 * cleared. If the size has not changed, B_CACHE remains 2498 * unchanged from its previous state. 2499 */ 2500 2501 if (bp->b_bcount != size) 2502 allocbuf(bp, size); 2503 2504 KASSERT(bp->b_offset != NOOFFSET, 2505 ("getblk: no buffer offset")); 2506 2507 /* 2508 * A buffer with B_DELWRI set and B_CACHE clear must 2509 * be committed before we can return the buffer in 2510 * order to prevent the caller from issuing a read 2511 * ( due to B_CACHE not being set ) and overwriting 2512 * it. 2513 * 2514 * Most callers, including NFS and FFS, need this to 2515 * operate properly either because they assume they 2516 * can issue a read if B_CACHE is not set, or because 2517 * ( for example ) an uncached B_DELWRI might loop due 2518 * to softupdates re-dirtying the buffer. In the latter 2519 * case, B_CACHE is set after the first write completes, 2520 * preventing further loops. 2521 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2522 * above while extending the buffer, we cannot allow the 2523 * buffer to remain with B_CACHE set after the write 2524 * completes or it will represent a corrupt state. To 2525 * deal with this we set B_NOCACHE to scrap the buffer 2526 * after the write. 2527 * 2528 * We might be able to do something fancy, like setting 2529 * B_CACHE in bwrite() except if B_DELWRI is already set, 2530 * so the below call doesn't set B_CACHE, but that gets real 2531 * confusing. This is much easier. 2532 */ 2533 2534 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2535 bp->b_flags |= B_NOCACHE; 2536 BUF_WRITE(bp); 2537 goto loop; 2538 } 2539 2540 splx(s); 2541 bp->b_flags &= ~B_DONE; 2542 } else { 2543 int bsize, maxsize, vmio; 2544 off_t offset; 2545 2546 /* 2547 * Buffer is not in-core, create new buffer. The buffer 2548 * returned by getnewbuf() is locked. Note that the returned 2549 * buffer is also considered valid (not marked B_INVAL). 2550 */ 2551 VI_UNLOCK(vp); 2552 /* 2553 * If the user does not want us to create the buffer, bail out 2554 * here. 2555 */ 2556 if (flags & GB_NOCREAT) { 2557 splx(s); 2558 return NULL; 2559 } 2560 if (vn_isdisk(vp, NULL)) 2561 bsize = DEV_BSIZE; 2562 else if (vp->v_mountedhere) 2563 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2564 else if (vp->v_mount) 2565 bsize = vp->v_mount->mnt_stat.f_iosize; 2566 else 2567 bsize = size; 2568 2569 offset = blkno * bsize; 2570 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2571 (vp->v_vflag & VV_OBJBUF); 2572 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2573 maxsize = imax(maxsize, bsize); 2574 2575 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2576 if (slpflag || slptimeo) { 2577 splx(s); 2578 return NULL; 2579 } 2580 goto loop; 2581 } 2582 2583 /* 2584 * This code is used to make sure that a buffer is not 2585 * created while the getnewbuf routine is blocked. 2586 * This can be a problem whether the vnode is locked or not. 2587 * If the buffer is created out from under us, we have to 2588 * throw away the one we just created. There is now window 2589 * race because we are safely running at splbio() from the 2590 * point of the duplicate buffer creation through to here, 2591 * and we've locked the buffer. 2592 * 2593 * Note: this must occur before we associate the buffer 2594 * with the vp especially considering limitations in 2595 * the splay tree implementation when dealing with duplicate 2596 * lblkno's. 2597 */ 2598 VI_LOCK(vp); 2599 if (gbincore(vp, blkno)) { 2600 VI_UNLOCK(vp); 2601 bp->b_flags |= B_INVAL; 2602 brelse(bp); 2603 goto loop; 2604 } 2605 2606 /* 2607 * Insert the buffer into the hash, so that it can 2608 * be found by incore. 2609 */ 2610 bp->b_blkno = bp->b_lblkno = blkno; 2611 bp->b_offset = offset; 2612 2613 bgetvp(vp, bp); 2614 VI_UNLOCK(vp); 2615 2616 /* 2617 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2618 * buffer size starts out as 0, B_CACHE will be set by 2619 * allocbuf() for the VMIO case prior to it testing the 2620 * backing store for validity. 2621 */ 2622 2623 if (vmio) { 2624 bp->b_flags |= B_VMIO; 2625#if defined(VFS_BIO_DEBUG) 2626 if (vp->v_type != VREG) 2627 printf("getblk: vmioing file type %d???\n", vp->v_type); 2628#endif 2629 VOP_GETVOBJECT(vp, &bp->b_object); 2630 } else { 2631 bp->b_flags &= ~B_VMIO; 2632 bp->b_object = NULL; 2633 } 2634 2635 allocbuf(bp, size); 2636 2637 splx(s); 2638 bp->b_flags &= ~B_DONE; 2639 } 2640 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2641 return (bp); 2642} 2643 2644/* 2645 * Get an empty, disassociated buffer of given size. The buffer is initially 2646 * set to B_INVAL. 2647 */ 2648struct buf * 2649geteblk(int size) 2650{ 2651 struct buf *bp; 2652 int s; 2653 int maxsize; 2654 2655 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2656 2657 s = splbio(); 2658 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2659 continue; 2660 splx(s); 2661 allocbuf(bp, size); 2662 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2663 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2664 return (bp); 2665} 2666 2667 2668/* 2669 * This code constitutes the buffer memory from either anonymous system 2670 * memory (in the case of non-VMIO operations) or from an associated 2671 * VM object (in the case of VMIO operations). This code is able to 2672 * resize a buffer up or down. 2673 * 2674 * Note that this code is tricky, and has many complications to resolve 2675 * deadlock or inconsistant data situations. Tread lightly!!! 2676 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2677 * the caller. Calling this code willy nilly can result in the loss of data. 2678 * 2679 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2680 * B_CACHE for the non-VMIO case. 2681 */ 2682 2683int 2684allocbuf(struct buf *bp, int size) 2685{ 2686 int newbsize, mbsize; 2687 int i; 2688 2689 GIANT_REQUIRED; 2690 2691 if (BUF_REFCNT(bp) == 0) 2692 panic("allocbuf: buffer not busy"); 2693 2694 if (bp->b_kvasize < size) 2695 panic("allocbuf: buffer too small"); 2696 2697 if ((bp->b_flags & B_VMIO) == 0) { 2698 caddr_t origbuf; 2699 int origbufsize; 2700 /* 2701 * Just get anonymous memory from the kernel. Don't 2702 * mess with B_CACHE. 2703 */ 2704 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2705 if (bp->b_flags & B_MALLOC) 2706 newbsize = mbsize; 2707 else 2708 newbsize = round_page(size); 2709 2710 if (newbsize < bp->b_bufsize) { 2711 /* 2712 * malloced buffers are not shrunk 2713 */ 2714 if (bp->b_flags & B_MALLOC) { 2715 if (newbsize) { 2716 bp->b_bcount = size; 2717 } else { 2718 free(bp->b_data, M_BIOBUF); 2719 if (bp->b_bufsize) { 2720 atomic_subtract_int( 2721 &bufmallocspace, 2722 bp->b_bufsize); 2723 bufspacewakeup(); 2724 bp->b_bufsize = 0; 2725 } 2726 bp->b_saveaddr = bp->b_kvabase; 2727 bp->b_data = bp->b_saveaddr; 2728 bp->b_bcount = 0; 2729 bp->b_flags &= ~B_MALLOC; 2730 } 2731 return 1; 2732 } 2733 vm_hold_free_pages( 2734 bp, 2735 (vm_offset_t) bp->b_data + newbsize, 2736 (vm_offset_t) bp->b_data + bp->b_bufsize); 2737 } else if (newbsize > bp->b_bufsize) { 2738 /* 2739 * We only use malloced memory on the first allocation. 2740 * and revert to page-allocated memory when the buffer 2741 * grows. 2742 */ 2743 /* 2744 * There is a potential smp race here that could lead 2745 * to bufmallocspace slightly passing the max. It 2746 * is probably extremely rare and not worth worrying 2747 * over. 2748 */ 2749 if ( (bufmallocspace < maxbufmallocspace) && 2750 (bp->b_bufsize == 0) && 2751 (mbsize <= PAGE_SIZE/2)) { 2752 2753 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2754 bp->b_bufsize = mbsize; 2755 bp->b_bcount = size; 2756 bp->b_flags |= B_MALLOC; 2757 atomic_add_int(&bufmallocspace, mbsize); 2758 return 1; 2759 } 2760 origbuf = NULL; 2761 origbufsize = 0; 2762 /* 2763 * If the buffer is growing on its other-than-first allocation, 2764 * then we revert to the page-allocation scheme. 2765 */ 2766 if (bp->b_flags & B_MALLOC) { 2767 origbuf = bp->b_data; 2768 origbufsize = bp->b_bufsize; 2769 bp->b_data = bp->b_kvabase; 2770 if (bp->b_bufsize) { 2771 atomic_subtract_int(&bufmallocspace, 2772 bp->b_bufsize); 2773 bufspacewakeup(); 2774 bp->b_bufsize = 0; 2775 } 2776 bp->b_flags &= ~B_MALLOC; 2777 newbsize = round_page(newbsize); 2778 } 2779 vm_hold_load_pages( 2780 bp, 2781 (vm_offset_t) bp->b_data + bp->b_bufsize, 2782 (vm_offset_t) bp->b_data + newbsize); 2783 if (origbuf) { 2784 bcopy(origbuf, bp->b_data, origbufsize); 2785 free(origbuf, M_BIOBUF); 2786 } 2787 } 2788 } else { 2789 int desiredpages; 2790 2791 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2792 desiredpages = (size == 0) ? 0 : 2793 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2794 2795 if (bp->b_flags & B_MALLOC) 2796 panic("allocbuf: VMIO buffer can't be malloced"); 2797 /* 2798 * Set B_CACHE initially if buffer is 0 length or will become 2799 * 0-length. 2800 */ 2801 if (size == 0 || bp->b_bufsize == 0) 2802 bp->b_flags |= B_CACHE; 2803 2804 if (newbsize < bp->b_bufsize) { 2805 /* 2806 * DEV_BSIZE aligned new buffer size is less then the 2807 * DEV_BSIZE aligned existing buffer size. Figure out 2808 * if we have to remove any pages. 2809 */ 2810 if (desiredpages < bp->b_npages) { 2811 vm_page_t m; 2812 2813 vm_page_lock_queues(); 2814 for (i = desiredpages; i < bp->b_npages; i++) { 2815 /* 2816 * the page is not freed here -- it 2817 * is the responsibility of 2818 * vnode_pager_setsize 2819 */ 2820 m = bp->b_pages[i]; 2821 KASSERT(m != bogus_page, 2822 ("allocbuf: bogus page found")); 2823 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2824 vm_page_lock_queues(); 2825 2826 bp->b_pages[i] = NULL; 2827 vm_page_unwire(m, 0); 2828 } 2829 vm_page_unlock_queues(); 2830 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2831 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2832 bp->b_npages = desiredpages; 2833 } 2834 } else if (size > bp->b_bcount) { 2835 /* 2836 * We are growing the buffer, possibly in a 2837 * byte-granular fashion. 2838 */ 2839 struct vnode *vp; 2840 vm_object_t obj; 2841 vm_offset_t toff; 2842 vm_offset_t tinc; 2843 2844 /* 2845 * Step 1, bring in the VM pages from the object, 2846 * allocating them if necessary. We must clear 2847 * B_CACHE if these pages are not valid for the 2848 * range covered by the buffer. 2849 */ 2850 2851 vp = bp->b_vp; 2852 obj = bp->b_object; 2853 2854 VM_OBJECT_LOCK(obj); 2855 while (bp->b_npages < desiredpages) { 2856 vm_page_t m; 2857 vm_pindex_t pi; 2858 2859 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2860 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2861 /* 2862 * note: must allocate system pages 2863 * since blocking here could intefere 2864 * with paging I/O, no matter which 2865 * process we are. 2866 */ 2867 m = vm_page_alloc(obj, pi, 2868 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2869 if (m == NULL) { 2870 atomic_add_int(&vm_pageout_deficit, 2871 desiredpages - bp->b_npages); 2872 VM_OBJECT_UNLOCK(obj); 2873 VM_WAIT; 2874 VM_OBJECT_LOCK(obj); 2875 } else { 2876 vm_page_lock_queues(); 2877 vm_page_wakeup(m); 2878 vm_page_unlock_queues(); 2879 bp->b_flags &= ~B_CACHE; 2880 bp->b_pages[bp->b_npages] = m; 2881 ++bp->b_npages; 2882 } 2883 continue; 2884 } 2885 2886 /* 2887 * We found a page. If we have to sleep on it, 2888 * retry because it might have gotten freed out 2889 * from under us. 2890 * 2891 * We can only test PG_BUSY here. Blocking on 2892 * m->busy might lead to a deadlock: 2893 * 2894 * vm_fault->getpages->cluster_read->allocbuf 2895 * 2896 */ 2897 vm_page_lock_queues(); 2898 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2899 continue; 2900 2901 /* 2902 * We have a good page. Should we wakeup the 2903 * page daemon? 2904 */ 2905 if ((curproc != pageproc) && 2906 ((m->queue - m->pc) == PQ_CACHE) && 2907 ((cnt.v_free_count + cnt.v_cache_count) < 2908 (cnt.v_free_min + cnt.v_cache_min))) { 2909 pagedaemon_wakeup(); 2910 } 2911 vm_page_flag_clear(m, PG_ZERO); 2912 vm_page_wire(m); 2913 vm_page_unlock_queues(); 2914 bp->b_pages[bp->b_npages] = m; 2915 ++bp->b_npages; 2916 } 2917 2918 /* 2919 * Step 2. We've loaded the pages into the buffer, 2920 * we have to figure out if we can still have B_CACHE 2921 * set. Note that B_CACHE is set according to the 2922 * byte-granular range ( bcount and size ), new the 2923 * aligned range ( newbsize ). 2924 * 2925 * The VM test is against m->valid, which is DEV_BSIZE 2926 * aligned. Needless to say, the validity of the data 2927 * needs to also be DEV_BSIZE aligned. Note that this 2928 * fails with NFS if the server or some other client 2929 * extends the file's EOF. If our buffer is resized, 2930 * B_CACHE may remain set! XXX 2931 */ 2932 2933 toff = bp->b_bcount; 2934 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2935 2936 while ((bp->b_flags & B_CACHE) && toff < size) { 2937 vm_pindex_t pi; 2938 2939 if (tinc > (size - toff)) 2940 tinc = size - toff; 2941 2942 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2943 PAGE_SHIFT; 2944 2945 vfs_buf_test_cache( 2946 bp, 2947 bp->b_offset, 2948 toff, 2949 tinc, 2950 bp->b_pages[pi] 2951 ); 2952 toff += tinc; 2953 tinc = PAGE_SIZE; 2954 } 2955 VM_OBJECT_UNLOCK(obj); 2956 2957 /* 2958 * Step 3, fixup the KVM pmap. Remember that 2959 * bp->b_data is relative to bp->b_offset, but 2960 * bp->b_offset may be offset into the first page. 2961 */ 2962 2963 bp->b_data = (caddr_t) 2964 trunc_page((vm_offset_t)bp->b_data); 2965 pmap_qenter( 2966 (vm_offset_t)bp->b_data, 2967 bp->b_pages, 2968 bp->b_npages 2969 ); 2970 2971 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2972 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2973 } 2974 } 2975 if (newbsize < bp->b_bufsize) 2976 bufspacewakeup(); 2977 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2978 bp->b_bcount = size; /* requested buffer size */ 2979 return 1; 2980} 2981 2982void 2983biodone(struct bio *bp) 2984{ 2985 mtx_lock(&bdonelock); 2986 bp->bio_flags |= BIO_DONE; 2987 if (bp->bio_done == NULL) 2988 wakeup(bp); 2989 mtx_unlock(&bdonelock); 2990 if (bp->bio_done != NULL) 2991 bp->bio_done(bp); 2992} 2993 2994/* 2995 * Wait for a BIO to finish. 2996 * 2997 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2998 * case is not yet clear. 2999 */ 3000int 3001biowait(struct bio *bp, const char *wchan) 3002{ 3003 3004 mtx_lock(&bdonelock); 3005 while ((bp->bio_flags & BIO_DONE) == 0) 3006 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3007 mtx_unlock(&bdonelock); 3008 if (bp->bio_error != 0) 3009 return (bp->bio_error); 3010 if (!(bp->bio_flags & BIO_ERROR)) 3011 return (0); 3012 return (EIO); 3013} 3014 3015void 3016biofinish(struct bio *bp, struct devstat *stat, int error) 3017{ 3018 3019 if (error) { 3020 bp->bio_error = error; 3021 bp->bio_flags |= BIO_ERROR; 3022 } 3023 if (stat != NULL) 3024 devstat_end_transaction_bio(stat, bp); 3025 biodone(bp); 3026} 3027 3028/* 3029 * bufwait: 3030 * 3031 * Wait for buffer I/O completion, returning error status. The buffer 3032 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3033 * error and cleared. 3034 */ 3035int 3036bufwait(register struct buf * bp) 3037{ 3038 int s; 3039 3040 s = splbio(); 3041 if (bp->b_iocmd == BIO_READ) 3042 bwait(bp, PRIBIO, "biord"); 3043 else 3044 bwait(bp, PRIBIO, "biowr"); 3045 splx(s); 3046 if (bp->b_flags & B_EINTR) { 3047 bp->b_flags &= ~B_EINTR; 3048 return (EINTR); 3049 } 3050 if (bp->b_ioflags & BIO_ERROR) { 3051 return (bp->b_error ? bp->b_error : EIO); 3052 } else { 3053 return (0); 3054 } 3055} 3056 3057 /* 3058 * Call back function from struct bio back up to struct buf. 3059 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 3060 */ 3061static void 3062bufdonebio(struct bio *bp) 3063{ 3064 3065 /* Device drivers may or may not hold giant, hold it here. */ 3066 mtx_lock(&Giant); 3067 bufdone(bp->bio_caller2); 3068 mtx_unlock(&Giant); 3069} 3070 3071void 3072dev_strategy(struct buf *bp) 3073{ 3074 3075 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3076 panic("b_iocmd botch"); 3077 if (bp->b_flags & B_PHYS) 3078 KASSERT(bp->b_io.bio_offset == bp->b_offset, 3079 ("bio_offset %jd wrong, should be %jd", 3080 (intmax_t)bp->b_io.bio_offset, (intmax_t)bp->b_offset)); 3081 else 3082 KASSERT(bp->b_io.bio_offset == dbtob(bp->b_blkno), 3083 ("bio_offset %jd wrong, should be %jd", 3084 (intmax_t)bp->b_io.bio_offset, 3085 (intmax_t)dbtob(bp->b_blkno))); 3086 bp->b_io.bio_done = bufdonebio; 3087 bp->b_io.bio_caller2 = bp; 3088 (*devsw(bp->b_io.bio_dev)->d_strategy)(&bp->b_io); 3089} 3090 3091/* 3092 * bufdone: 3093 * 3094 * Finish I/O on a buffer, optionally calling a completion function. 3095 * This is usually called from an interrupt so process blocking is 3096 * not allowed. 3097 * 3098 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3099 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3100 * assuming B_INVAL is clear. 3101 * 3102 * For the VMIO case, we set B_CACHE if the op was a read and no 3103 * read error occured, or if the op was a write. B_CACHE is never 3104 * set if the buffer is invalid or otherwise uncacheable. 3105 * 3106 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3107 * initiator to leave B_INVAL set to brelse the buffer out of existance 3108 * in the biodone routine. 3109 */ 3110void 3111bufdone(struct buf *bp) 3112{ 3113 int s; 3114 void (*biodone)(struct buf *); 3115 3116 GIANT_REQUIRED; 3117 3118 s = splbio(); 3119 3120 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3121 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3122 3123 bp->b_flags |= B_DONE; 3124 runningbufwakeup(bp); 3125 3126 if (bp->b_iocmd == BIO_DELETE) { 3127 brelse(bp); 3128 splx(s); 3129 return; 3130 } 3131 3132 if (bp->b_iocmd == BIO_WRITE) { 3133 vwakeup(bp); 3134 } 3135 3136 /* call optional completion function if requested */ 3137 if (bp->b_iodone != NULL) { 3138 biodone = bp->b_iodone; 3139 bp->b_iodone = NULL; 3140 (*biodone) (bp); 3141 splx(s); 3142 return; 3143 } 3144 if (LIST_FIRST(&bp->b_dep) != NULL) 3145 buf_complete(bp); 3146 3147 if (bp->b_flags & B_VMIO) { 3148 int i; 3149 vm_ooffset_t foff; 3150 vm_page_t m; 3151 vm_object_t obj; 3152 int iosize; 3153 struct vnode *vp = bp->b_vp; 3154 3155 obj = bp->b_object; 3156 3157#if defined(VFS_BIO_DEBUG) 3158 mp_fixme("usecount and vflag accessed without locks."); 3159 if (vp->v_usecount == 0) { 3160 panic("biodone: zero vnode ref count"); 3161 } 3162 3163 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3164 panic("biodone: vnode is not setup for merged cache"); 3165 } 3166#endif 3167 3168 foff = bp->b_offset; 3169 KASSERT(bp->b_offset != NOOFFSET, 3170 ("biodone: no buffer offset")); 3171 3172 VM_OBJECT_LOCK(obj); 3173#if defined(VFS_BIO_DEBUG) 3174 if (obj->paging_in_progress < bp->b_npages) { 3175 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3176 obj->paging_in_progress, bp->b_npages); 3177 } 3178#endif 3179 3180 /* 3181 * Set B_CACHE if the op was a normal read and no error 3182 * occured. B_CACHE is set for writes in the b*write() 3183 * routines. 3184 */ 3185 iosize = bp->b_bcount - bp->b_resid; 3186 if (bp->b_iocmd == BIO_READ && 3187 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3188 !(bp->b_ioflags & BIO_ERROR)) { 3189 bp->b_flags |= B_CACHE; 3190 } 3191 vm_page_lock_queues(); 3192 for (i = 0; i < bp->b_npages; i++) { 3193 int bogusflag = 0; 3194 int resid; 3195 3196 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3197 if (resid > iosize) 3198 resid = iosize; 3199 3200 /* 3201 * cleanup bogus pages, restoring the originals 3202 */ 3203 m = bp->b_pages[i]; 3204 if (m == bogus_page) { 3205 bogusflag = 1; 3206 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3207 if (m == NULL) 3208 panic("biodone: page disappeared!"); 3209 bp->b_pages[i] = m; 3210 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3211 } 3212#if defined(VFS_BIO_DEBUG) 3213 if (OFF_TO_IDX(foff) != m->pindex) { 3214 printf( 3215"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3216 (intmax_t)foff, (uintmax_t)m->pindex); 3217 } 3218#endif 3219 3220 /* 3221 * In the write case, the valid and clean bits are 3222 * already changed correctly ( see bdwrite() ), so we 3223 * only need to do this here in the read case. 3224 */ 3225 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3226 vfs_page_set_valid(bp, foff, i, m); 3227 } 3228 vm_page_flag_clear(m, PG_ZERO); 3229 3230 /* 3231 * when debugging new filesystems or buffer I/O methods, this 3232 * is the most common error that pops up. if you see this, you 3233 * have not set the page busy flag correctly!!! 3234 */ 3235 if (m->busy == 0) { 3236 printf("biodone: page busy < 0, " 3237 "pindex: %d, foff: 0x(%x,%x), " 3238 "resid: %d, index: %d\n", 3239 (int) m->pindex, (int)(foff >> 32), 3240 (int) foff & 0xffffffff, resid, i); 3241 if (!vn_isdisk(vp, NULL)) 3242 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3243 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3244 (intmax_t) bp->b_lblkno, 3245 bp->b_flags, bp->b_npages); 3246 else 3247 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3248 (intmax_t) bp->b_lblkno, 3249 bp->b_flags, bp->b_npages); 3250 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3251 (u_long)m->valid, (u_long)m->dirty, 3252 m->wire_count); 3253 panic("biodone: page busy < 0\n"); 3254 } 3255 vm_page_io_finish(m); 3256 vm_object_pip_subtract(obj, 1); 3257 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3258 iosize -= resid; 3259 } 3260 vm_page_unlock_queues(); 3261 vm_object_pip_wakeupn(obj, 0); 3262 VM_OBJECT_UNLOCK(obj); 3263 } 3264 3265 /* 3266 * For asynchronous completions, release the buffer now. The brelse 3267 * will do a wakeup there if necessary - so no need to do a wakeup 3268 * here in the async case. The sync case always needs to do a wakeup. 3269 */ 3270 3271 if (bp->b_flags & B_ASYNC) { 3272 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3273 brelse(bp); 3274 else 3275 bqrelse(bp); 3276 } else { 3277 bdone(bp); 3278 } 3279 splx(s); 3280} 3281 3282/* 3283 * This routine is called in lieu of iodone in the case of 3284 * incomplete I/O. This keeps the busy status for pages 3285 * consistant. 3286 */ 3287void 3288vfs_unbusy_pages(struct buf * bp) 3289{ 3290 int i; 3291 3292 runningbufwakeup(bp); 3293 if (bp->b_flags & B_VMIO) { 3294 vm_object_t obj; 3295 3296 obj = bp->b_object; 3297 VM_OBJECT_LOCK(obj); 3298 vm_page_lock_queues(); 3299 for (i = 0; i < bp->b_npages; i++) { 3300 vm_page_t m = bp->b_pages[i]; 3301 3302 if (m == bogus_page) { 3303 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3304 if (!m) { 3305 panic("vfs_unbusy_pages: page missing\n"); 3306 } 3307 bp->b_pages[i] = m; 3308 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3309 } 3310 vm_object_pip_subtract(obj, 1); 3311 vm_page_flag_clear(m, PG_ZERO); 3312 vm_page_io_finish(m); 3313 } 3314 vm_page_unlock_queues(); 3315 vm_object_pip_wakeupn(obj, 0); 3316 VM_OBJECT_UNLOCK(obj); 3317 } 3318} 3319 3320/* 3321 * vfs_page_set_valid: 3322 * 3323 * Set the valid bits in a page based on the supplied offset. The 3324 * range is restricted to the buffer's size. 3325 * 3326 * This routine is typically called after a read completes. 3327 */ 3328static void 3329vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3330{ 3331 vm_ooffset_t soff, eoff; 3332 3333 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3334 /* 3335 * Start and end offsets in buffer. eoff - soff may not cross a 3336 * page boundry or cross the end of the buffer. The end of the 3337 * buffer, in this case, is our file EOF, not the allocation size 3338 * of the buffer. 3339 */ 3340 soff = off; 3341 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3342 if (eoff > bp->b_offset + bp->b_bcount) 3343 eoff = bp->b_offset + bp->b_bcount; 3344 3345 /* 3346 * Set valid range. This is typically the entire buffer and thus the 3347 * entire page. 3348 */ 3349 if (eoff > soff) { 3350 vm_page_set_validclean( 3351 m, 3352 (vm_offset_t) (soff & PAGE_MASK), 3353 (vm_offset_t) (eoff - soff) 3354 ); 3355 } 3356} 3357 3358/* 3359 * This routine is called before a device strategy routine. 3360 * It is used to tell the VM system that paging I/O is in 3361 * progress, and treat the pages associated with the buffer 3362 * almost as being PG_BUSY. Also the object paging_in_progress 3363 * flag is handled to make sure that the object doesn't become 3364 * inconsistant. 3365 * 3366 * Since I/O has not been initiated yet, certain buffer flags 3367 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3368 * and should be ignored. 3369 */ 3370void 3371vfs_busy_pages(struct buf * bp, int clear_modify) 3372{ 3373 int i, bogus; 3374 3375 if (bp->b_flags & B_VMIO) { 3376 vm_object_t obj; 3377 vm_ooffset_t foff; 3378 3379 obj = bp->b_object; 3380 foff = bp->b_offset; 3381 KASSERT(bp->b_offset != NOOFFSET, 3382 ("vfs_busy_pages: no buffer offset")); 3383 vfs_setdirty(bp); 3384 VM_OBJECT_LOCK(obj); 3385retry: 3386 vm_page_lock_queues(); 3387 for (i = 0; i < bp->b_npages; i++) { 3388 vm_page_t m = bp->b_pages[i]; 3389 3390 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3391 goto retry; 3392 } 3393 bogus = 0; 3394 for (i = 0; i < bp->b_npages; i++) { 3395 vm_page_t m = bp->b_pages[i]; 3396 3397 vm_page_flag_clear(m, PG_ZERO); 3398 if ((bp->b_flags & B_CLUSTER) == 0) { 3399 vm_object_pip_add(obj, 1); 3400 vm_page_io_start(m); 3401 } 3402 /* 3403 * When readying a buffer for a read ( i.e 3404 * clear_modify == 0 ), it is important to do 3405 * bogus_page replacement for valid pages in 3406 * partially instantiated buffers. Partially 3407 * instantiated buffers can, in turn, occur when 3408 * reconstituting a buffer from its VM backing store 3409 * base. We only have to do this if B_CACHE is 3410 * clear ( which causes the I/O to occur in the 3411 * first place ). The replacement prevents the read 3412 * I/O from overwriting potentially dirty VM-backed 3413 * pages. XXX bogus page replacement is, uh, bogus. 3414 * It may not work properly with small-block devices. 3415 * We need to find a better way. 3416 */ 3417 pmap_remove_all(m); 3418 if (clear_modify) 3419 vfs_page_set_valid(bp, foff, i, m); 3420 else if (m->valid == VM_PAGE_BITS_ALL && 3421 (bp->b_flags & B_CACHE) == 0) { 3422 bp->b_pages[i] = bogus_page; 3423 bogus++; 3424 } 3425 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3426 } 3427 vm_page_unlock_queues(); 3428 VM_OBJECT_UNLOCK(obj); 3429 if (bogus) 3430 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3431 } 3432} 3433 3434/* 3435 * Tell the VM system that the pages associated with this buffer 3436 * are clean. This is used for delayed writes where the data is 3437 * going to go to disk eventually without additional VM intevention. 3438 * 3439 * Note that while we only really need to clean through to b_bcount, we 3440 * just go ahead and clean through to b_bufsize. 3441 */ 3442static void 3443vfs_clean_pages(struct buf * bp) 3444{ 3445 int i; 3446 3447 if (bp->b_flags & B_VMIO) { 3448 vm_ooffset_t foff; 3449 3450 foff = bp->b_offset; 3451 KASSERT(bp->b_offset != NOOFFSET, 3452 ("vfs_clean_pages: no buffer offset")); 3453 VM_OBJECT_LOCK(bp->b_object); 3454 vm_page_lock_queues(); 3455 for (i = 0; i < bp->b_npages; i++) { 3456 vm_page_t m = bp->b_pages[i]; 3457 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3458 vm_ooffset_t eoff = noff; 3459 3460 if (eoff > bp->b_offset + bp->b_bufsize) 3461 eoff = bp->b_offset + bp->b_bufsize; 3462 vfs_page_set_valid(bp, foff, i, m); 3463 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3464 foff = noff; 3465 } 3466 vm_page_unlock_queues(); 3467 VM_OBJECT_UNLOCK(bp->b_object); 3468 } 3469} 3470 3471/* 3472 * vfs_bio_set_validclean: 3473 * 3474 * Set the range within the buffer to valid and clean. The range is 3475 * relative to the beginning of the buffer, b_offset. Note that b_offset 3476 * itself may be offset from the beginning of the first page. 3477 * 3478 */ 3479 3480void 3481vfs_bio_set_validclean(struct buf *bp, int base, int size) 3482{ 3483 if (bp->b_flags & B_VMIO) { 3484 int i; 3485 int n; 3486 3487 /* 3488 * Fixup base to be relative to beginning of first page. 3489 * Set initial n to be the maximum number of bytes in the 3490 * first page that can be validated. 3491 */ 3492 3493 base += (bp->b_offset & PAGE_MASK); 3494 n = PAGE_SIZE - (base & PAGE_MASK); 3495 3496 VM_OBJECT_LOCK(bp->b_object); 3497 vm_page_lock_queues(); 3498 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3499 vm_page_t m = bp->b_pages[i]; 3500 3501 if (n > size) 3502 n = size; 3503 3504 vm_page_set_validclean(m, base & PAGE_MASK, n); 3505 base += n; 3506 size -= n; 3507 n = PAGE_SIZE; 3508 } 3509 vm_page_unlock_queues(); 3510 VM_OBJECT_UNLOCK(bp->b_object); 3511 } 3512} 3513 3514/* 3515 * vfs_bio_clrbuf: 3516 * 3517 * clear a buffer. This routine essentially fakes an I/O, so we need 3518 * to clear BIO_ERROR and B_INVAL. 3519 * 3520 * Note that while we only theoretically need to clear through b_bcount, 3521 * we go ahead and clear through b_bufsize. 3522 */ 3523 3524void 3525vfs_bio_clrbuf(struct buf *bp) 3526{ 3527 int i, mask = 0; 3528 caddr_t sa, ea; 3529 3530 GIANT_REQUIRED; 3531 3532 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3533 bp->b_flags &= ~B_INVAL; 3534 bp->b_ioflags &= ~BIO_ERROR; 3535 VM_OBJECT_LOCK(bp->b_object); 3536 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3537 (bp->b_offset & PAGE_MASK) == 0) { 3538 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3539 if (bp->b_pages[0] != bogus_page) 3540 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3541 if ((bp->b_pages[0]->valid & mask) == mask) 3542 goto unlock; 3543 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3544 ((bp->b_pages[0]->valid & mask) == 0)) { 3545 bzero(bp->b_data, bp->b_bufsize); 3546 bp->b_pages[0]->valid |= mask; 3547 goto unlock; 3548 } 3549 } 3550 ea = sa = bp->b_data; 3551 for(i=0;i<bp->b_npages;i++,sa=ea) { 3552 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3553 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3554 ea = (caddr_t)(vm_offset_t)ulmin( 3555 (u_long)(vm_offset_t)ea, 3556 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3557 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3558 if (bp->b_pages[i] != bogus_page) 3559 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3560 if ((bp->b_pages[i]->valid & mask) == mask) 3561 continue; 3562 if ((bp->b_pages[i]->valid & mask) == 0) { 3563 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3564 bzero(sa, ea - sa); 3565 } 3566 } else { 3567 for (; sa < ea; sa += DEV_BSIZE, j++) { 3568 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3569 (bp->b_pages[i]->valid & (1<<j)) == 0) 3570 bzero(sa, DEV_BSIZE); 3571 } 3572 } 3573 bp->b_pages[i]->valid |= mask; 3574 vm_page_lock_queues(); 3575 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3576 vm_page_unlock_queues(); 3577 } 3578unlock: 3579 VM_OBJECT_UNLOCK(bp->b_object); 3580 bp->b_resid = 0; 3581 } else { 3582 clrbuf(bp); 3583 } 3584} 3585 3586/* 3587 * vm_hold_load_pages and vm_hold_free_pages get pages into 3588 * a buffers address space. The pages are anonymous and are 3589 * not associated with a file object. 3590 */ 3591static void 3592vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3593{ 3594 vm_offset_t pg; 3595 vm_page_t p; 3596 int index; 3597 3598 to = round_page(to); 3599 from = round_page(from); 3600 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3601 3602 VM_OBJECT_LOCK(kernel_object); 3603 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3604tryagain: 3605 /* 3606 * note: must allocate system pages since blocking here 3607 * could intefere with paging I/O, no matter which 3608 * process we are. 3609 */ 3610 p = vm_page_alloc(kernel_object, 3611 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3612 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3613 if (!p) { 3614 atomic_add_int(&vm_pageout_deficit, 3615 (to - pg) >> PAGE_SHIFT); 3616 VM_OBJECT_UNLOCK(kernel_object); 3617 VM_WAIT; 3618 VM_OBJECT_LOCK(kernel_object); 3619 goto tryagain; 3620 } 3621 p->valid = VM_PAGE_BITS_ALL; 3622 pmap_qenter(pg, &p, 1); 3623 bp->b_pages[index] = p; 3624 vm_page_lock_queues(); 3625 vm_page_wakeup(p); 3626 vm_page_unlock_queues(); 3627 } 3628 VM_OBJECT_UNLOCK(kernel_object); 3629 bp->b_npages = index; 3630} 3631 3632/* Return pages associated with this buf to the vm system */ 3633static void 3634vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3635{ 3636 vm_offset_t pg; 3637 vm_page_t p; 3638 int index, newnpages; 3639 3640 GIANT_REQUIRED; 3641 3642 from = round_page(from); 3643 to = round_page(to); 3644 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3645 3646 VM_OBJECT_LOCK(kernel_object); 3647 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3648 p = bp->b_pages[index]; 3649 if (p && (index < bp->b_npages)) { 3650 if (p->busy) { 3651 printf( 3652 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3653 (intmax_t)bp->b_blkno, 3654 (intmax_t)bp->b_lblkno); 3655 } 3656 bp->b_pages[index] = NULL; 3657 pmap_qremove(pg, 1); 3658 vm_page_lock_queues(); 3659 vm_page_busy(p); 3660 vm_page_unwire(p, 0); 3661 vm_page_free(p); 3662 vm_page_unlock_queues(); 3663 } 3664 } 3665 VM_OBJECT_UNLOCK(kernel_object); 3666 bp->b_npages = newnpages; 3667} 3668 3669/* 3670 * Map an IO request into kernel virtual address space. 3671 * 3672 * All requests are (re)mapped into kernel VA space. 3673 * Notice that we use b_bufsize for the size of the buffer 3674 * to be mapped. b_bcount might be modified by the driver. 3675 * 3676 * Note that even if the caller determines that the address space should 3677 * be valid, a race or a smaller-file mapped into a larger space may 3678 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3679 * check the return value. 3680 */ 3681int 3682vmapbuf(struct buf *bp) 3683{ 3684 caddr_t addr, kva; 3685 vm_prot_t prot; 3686 int pidx, i; 3687 struct vm_page *m; 3688 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3689 3690 GIANT_REQUIRED; 3691 3692 if (bp->b_bufsize < 0) 3693 return (-1); 3694 prot = (bp->b_iocmd == BIO_READ) ? VM_PROT_READ | VM_PROT_WRITE : 3695 VM_PROT_READ; 3696 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3697 addr < bp->b_data + bp->b_bufsize; 3698 addr += PAGE_SIZE, pidx++) { 3699 /* 3700 * Do the vm_fault if needed; do the copy-on-write thing 3701 * when reading stuff off device into memory. 3702 * 3703 * NOTE! Must use pmap_extract() because addr may be in 3704 * the userland address space, and kextract is only guarenteed 3705 * to work for the kernland address space (see: sparc64 port). 3706 */ 3707retry: 3708 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3709 prot) < 0) { 3710 vm_page_lock_queues(); 3711 for (i = 0; i < pidx; ++i) { 3712 vm_page_unhold(bp->b_pages[i]); 3713 bp->b_pages[i] = NULL; 3714 } 3715 vm_page_unlock_queues(); 3716 return(-1); 3717 } 3718 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3719 if (m == NULL) 3720 goto retry; 3721 bp->b_pages[pidx] = m; 3722 } 3723 if (pidx > btoc(MAXPHYS)) 3724 panic("vmapbuf: mapped more than MAXPHYS"); 3725 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3726 3727 kva = bp->b_saveaddr; 3728 bp->b_npages = pidx; 3729 bp->b_saveaddr = bp->b_data; 3730 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3731 return(0); 3732} 3733 3734/* 3735 * Free the io map PTEs associated with this IO operation. 3736 * We also invalidate the TLB entries and restore the original b_addr. 3737 */ 3738void 3739vunmapbuf(struct buf *bp) 3740{ 3741 int pidx; 3742 int npages; 3743 3744 GIANT_REQUIRED; 3745 3746 npages = bp->b_npages; 3747 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3748 npages); 3749 vm_page_lock_queues(); 3750 for (pidx = 0; pidx < npages; pidx++) 3751 vm_page_unhold(bp->b_pages[pidx]); 3752 vm_page_unlock_queues(); 3753 3754 bp->b_data = bp->b_saveaddr; 3755} 3756 3757void 3758bdone(struct buf *bp) 3759{ 3760 mtx_lock(&bdonelock); 3761 bp->b_flags |= B_DONE; 3762 wakeup(bp); 3763 mtx_unlock(&bdonelock); 3764} 3765 3766void 3767bwait(struct buf *bp, u_char pri, const char *wchan) 3768{ 3769 mtx_lock(&bdonelock); 3770 while ((bp->b_flags & B_DONE) == 0) 3771 msleep(bp, &bdonelock, pri, wchan, 0); 3772 mtx_unlock(&bdonelock); 3773} 3774 3775#include "opt_ddb.h" 3776#ifdef DDB 3777#include <ddb/ddb.h> 3778 3779/* DDB command to show buffer data */ 3780DB_SHOW_COMMAND(buffer, db_show_buffer) 3781{ 3782 /* get args */ 3783 struct buf *bp = (struct buf *)addr; 3784 3785 if (!have_addr) { 3786 db_printf("usage: show buffer <addr>\n"); 3787 return; 3788 } 3789 3790 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3791 db_printf( 3792 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3793 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd\n", 3794 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3795 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3796 (intmax_t)bp->b_blkno); 3797 if (bp->b_npages) { 3798 int i; 3799 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3800 for (i = 0; i < bp->b_npages; i++) { 3801 vm_page_t m; 3802 m = bp->b_pages[i]; 3803 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3804 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3805 if ((i + 1) < bp->b_npages) 3806 db_printf(","); 3807 } 3808 db_printf("\n"); 3809 } 3810} 3811#endif /* DDB */ 3812