vfs_bio.c revision 115456
1/* 2 * Copyright (c) 1994,1997 John S. Dyson 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice immediately at the beginning of the file, without modification, 10 * this list of conditions, and the following disclaimer. 11 * 2. Absolutely no warranty of function or purpose is made by the author 12 * John S. Dyson. 13 * 14 * $FreeBSD: head/sys/kern/vfs_bio.c 115456 2003-05-31 16:42:45Z phk $ 15 */ 16 17/* 18 * this file contains a new buffer I/O scheme implementing a coherent 19 * VM object and buffer cache scheme. Pains have been taken to make 20 * sure that the performance degradation associated with schemes such 21 * as this is not realized. 22 * 23 * Author: John S. Dyson 24 * Significant help during the development and debugging phases 25 * had been provided by David Greenman, also of the FreeBSD core team. 26 * 27 * see man buf(9) for more info. 28 */ 29 30#include <sys/param.h> 31#include <sys/systm.h> 32#include <sys/bio.h> 33#include <sys/buf.h> 34#include <sys/devicestat.h> 35#include <sys/eventhandler.h> 36#include <sys/lock.h> 37#include <sys/malloc.h> 38#include <sys/mount.h> 39#include <sys/mutex.h> 40#include <sys/kernel.h> 41#include <sys/kthread.h> 42#include <sys/proc.h> 43#include <sys/resourcevar.h> 44#include <sys/sysctl.h> 45#include <sys/vmmeter.h> 46#include <sys/vnode.h> 47#include <vm/vm.h> 48#include <vm/vm_param.h> 49#include <vm/vm_kern.h> 50#include <vm/vm_pageout.h> 51#include <vm/vm_page.h> 52#include <vm/vm_object.h> 53#include <vm/vm_extern.h> 54#include <vm/vm_map.h> 55#include "opt_directio.h" 56#include "opt_swap.h" 57 58static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 59 60struct bio_ops bioops; /* I/O operation notification */ 61 62struct buf_ops buf_ops_bio = { 63 "buf_ops_bio", 64 bwrite 65}; 66 67/* 68 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 69 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 70 */ 71struct buf *buf; /* buffer header pool */ 72 73static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 74 vm_offset_t to); 75static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 76 vm_offset_t to); 77static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 78 int pageno, vm_page_t m); 79static void vfs_clean_pages(struct buf * bp); 80static void vfs_setdirty(struct buf *bp); 81static void vfs_vmio_release(struct buf *bp); 82static void vfs_backgroundwritedone(struct buf *bp); 83static int vfs_bio_clcheck(struct vnode *vp, int size, 84 daddr_t lblkno, daddr_t blkno); 85static int flushbufqueues(int flushdeps); 86static void buf_daemon(void); 87void bremfreel(struct buf * bp); 88 89int vmiodirenable = TRUE; 90SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 91 "Use the VM system for directory writes"); 92int runningbufspace; 93SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 94 "Amount of presently outstanding async buffer io"); 95static int bufspace; 96SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 97 "KVA memory used for bufs"); 98static int maxbufspace; 99SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 100 "Maximum allowed value of bufspace (including buf_daemon)"); 101static int bufmallocspace; 102SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 103 "Amount of malloced memory for buffers"); 104static int maxbufmallocspace; 105SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 106 "Maximum amount of malloced memory for buffers"); 107static int lobufspace; 108SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 109 "Minimum amount of buffers we want to have"); 110static int hibufspace; 111SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 112 "Maximum allowed value of bufspace (excluding buf_daemon)"); 113static int bufreusecnt; 114SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 115 "Number of times we have reused a buffer"); 116static int buffreekvacnt; 117SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 118 "Number of times we have freed the KVA space from some buffer"); 119static int bufdefragcnt; 120SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 121 "Number of times we have had to repeat buffer allocation to defragment"); 122static int lorunningspace; 123SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 124 "Minimum preferred space used for in-progress I/O"); 125static int hirunningspace; 126SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 127 "Maximum amount of space to use for in-progress I/O"); 128static int dirtybufferflushes; 129SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 130 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 131static int altbufferflushes; 132SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 133 0, "Number of fsync flushes to limit dirty buffers"); 134static int recursiveflushes; 135SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 136 0, "Number of flushes skipped due to being recursive"); 137static int numdirtybuffers; 138SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 139 "Number of buffers that are dirty (has unwritten changes) at the moment"); 140static int lodirtybuffers; 141SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 142 "How many buffers we want to have free before bufdaemon can sleep"); 143static int hidirtybuffers; 144SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 145 "When the number of dirty buffers is considered severe"); 146static int dirtybufthresh; 147SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 148 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 149static int numfreebuffers; 150SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 151 "Number of free buffers"); 152static int lofreebuffers; 153SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 154 "XXX Unused"); 155static int hifreebuffers; 156SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 157 "XXX Complicatedly unused"); 158static int getnewbufcalls; 159SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 160 "Number of calls to getnewbuf"); 161static int getnewbufrestarts; 162SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 163 "Number of times getnewbuf has had to restart a buffer aquisition"); 164static int dobkgrdwrite = 1; 165SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 166 "Do background writes (honoring the BX_BKGRDWRITE flag)?"); 167 168/* 169 * Wakeup point for bufdaemon, as well as indicator of whether it is already 170 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 171 * is idling. 172 */ 173static int bd_request; 174 175/* 176 * This lock synchronizes access to bd_request. 177 */ 178static struct mtx bdlock; 179 180/* 181 * bogus page -- for I/O to/from partially complete buffers 182 * this is a temporary solution to the problem, but it is not 183 * really that bad. it would be better to split the buffer 184 * for input in the case of buffers partially already in memory, 185 * but the code is intricate enough already. 186 */ 187vm_page_t bogus_page; 188 189/* 190 * Synchronization (sleep/wakeup) variable for active buffer space requests. 191 * Set when wait starts, cleared prior to wakeup(). 192 * Used in runningbufwakeup() and waitrunningbufspace(). 193 */ 194static int runningbufreq; 195 196/* 197 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 198 * waitrunningbufspace(). 199 */ 200static struct mtx rbreqlock; 201 202/* 203 * Synchronization (sleep/wakeup) variable for buffer requests. 204 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 205 * by and/or. 206 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 207 * getnewbuf(), and getblk(). 208 */ 209static int needsbuffer; 210 211/* 212 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 213 */ 214static struct mtx nblock; 215 216/* 217 * Lock that protects against bwait()/bdone()/B_DONE races. 218 */ 219 220static struct mtx bdonelock; 221 222/* 223 * Definitions for the buffer free lists. 224 */ 225#define BUFFER_QUEUES 6 /* number of free buffer queues */ 226 227#define QUEUE_NONE 0 /* on no queue */ 228#define QUEUE_LOCKED 1 /* locked buffers */ 229#define QUEUE_CLEAN 2 /* non-B_DELWRI buffers */ 230#define QUEUE_DIRTY 3 /* B_DELWRI buffers */ 231#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 232#define QUEUE_EMPTY 5 /* empty buffer headers */ 233 234/* Queues for free buffers with various properties */ 235static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 236 237/* Lock for the bufqueues */ 238static struct mtx bqlock; 239 240/* 241 * Single global constant for BUF_WMESG, to avoid getting multiple references. 242 * buf_wmesg is referred from macros. 243 */ 244const char *buf_wmesg = BUF_WMESG; 245 246#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 247#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 248#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 249#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 250 251#ifdef DIRECTIO 252extern void ffs_rawread_setup(void); 253#endif /* DIRECTIO */ 254/* 255 * numdirtywakeup: 256 * 257 * If someone is blocked due to there being too many dirty buffers, 258 * and numdirtybuffers is now reasonable, wake them up. 259 */ 260 261static __inline void 262numdirtywakeup(int level) 263{ 264 if (numdirtybuffers <= level) { 265 mtx_lock(&nblock); 266 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 267 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 268 wakeup(&needsbuffer); 269 } 270 mtx_unlock(&nblock); 271 } 272} 273 274/* 275 * bufspacewakeup: 276 * 277 * Called when buffer space is potentially available for recovery. 278 * getnewbuf() will block on this flag when it is unable to free 279 * sufficient buffer space. Buffer space becomes recoverable when 280 * bp's get placed back in the queues. 281 */ 282 283static __inline void 284bufspacewakeup(void) 285{ 286 /* 287 * If someone is waiting for BUF space, wake them up. Even 288 * though we haven't freed the kva space yet, the waiting 289 * process will be able to now. 290 */ 291 mtx_lock(&nblock); 292 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 293 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 294 wakeup(&needsbuffer); 295 } 296 mtx_unlock(&nblock); 297} 298 299/* 300 * runningbufwakeup() - in-progress I/O accounting. 301 * 302 */ 303static __inline void 304runningbufwakeup(struct buf *bp) 305{ 306 if (bp->b_runningbufspace) { 307 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 308 bp->b_runningbufspace = 0; 309 mtx_lock(&rbreqlock); 310 if (runningbufreq && runningbufspace <= lorunningspace) { 311 runningbufreq = 0; 312 wakeup(&runningbufreq); 313 } 314 mtx_unlock(&rbreqlock); 315 } 316} 317 318/* 319 * bufcountwakeup: 320 * 321 * Called when a buffer has been added to one of the free queues to 322 * account for the buffer and to wakeup anyone waiting for free buffers. 323 * This typically occurs when large amounts of metadata are being handled 324 * by the buffer cache ( else buffer space runs out first, usually ). 325 */ 326 327static __inline void 328bufcountwakeup(void) 329{ 330 atomic_add_int(&numfreebuffers, 1); 331 mtx_lock(&nblock); 332 if (needsbuffer) { 333 needsbuffer &= ~VFS_BIO_NEED_ANY; 334 if (numfreebuffers >= hifreebuffers) 335 needsbuffer &= ~VFS_BIO_NEED_FREE; 336 wakeup(&needsbuffer); 337 } 338 mtx_unlock(&nblock); 339} 340 341/* 342 * waitrunningbufspace() 343 * 344 * runningbufspace is a measure of the amount of I/O currently 345 * running. This routine is used in async-write situations to 346 * prevent creating huge backups of pending writes to a device. 347 * Only asynchronous writes are governed by this function. 348 * 349 * Reads will adjust runningbufspace, but will not block based on it. 350 * The read load has a side effect of reducing the allowed write load. 351 * 352 * This does NOT turn an async write into a sync write. It waits 353 * for earlier writes to complete and generally returns before the 354 * caller's write has reached the device. 355 */ 356static __inline void 357waitrunningbufspace(void) 358{ 359 mtx_lock(&rbreqlock); 360 while (runningbufspace > hirunningspace) { 361 ++runningbufreq; 362 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 363 } 364 mtx_unlock(&rbreqlock); 365} 366 367 368/* 369 * vfs_buf_test_cache: 370 * 371 * Called when a buffer is extended. This function clears the B_CACHE 372 * bit if the newly extended portion of the buffer does not contain 373 * valid data. 374 */ 375static __inline__ 376void 377vfs_buf_test_cache(struct buf *bp, 378 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 379 vm_page_t m) 380{ 381 GIANT_REQUIRED; 382 383 if (bp->b_flags & B_CACHE) { 384 int base = (foff + off) & PAGE_MASK; 385 if (vm_page_is_valid(m, base, size) == 0) 386 bp->b_flags &= ~B_CACHE; 387 } 388} 389 390/* Wake up the buffer deamon if necessary */ 391static __inline__ 392void 393bd_wakeup(int dirtybuflevel) 394{ 395 mtx_lock(&bdlock); 396 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 397 bd_request = 1; 398 wakeup(&bd_request); 399 } 400 mtx_unlock(&bdlock); 401} 402 403/* 404 * bd_speedup - speedup the buffer cache flushing code 405 */ 406 407static __inline__ 408void 409bd_speedup(void) 410{ 411 bd_wakeup(1); 412} 413 414/* 415 * Calculating buffer cache scaling values and reserve space for buffer 416 * headers. This is called during low level kernel initialization and 417 * may be called more then once. We CANNOT write to the memory area 418 * being reserved at this time. 419 */ 420caddr_t 421kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 422{ 423 /* 424 * physmem_est is in pages. Convert it to kilobytes (assumes 425 * PAGE_SIZE is >= 1K) 426 */ 427 physmem_est = physmem_est * (PAGE_SIZE / 1024); 428 429 /* 430 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 431 * For the first 64MB of ram nominally allocate sufficient buffers to 432 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 433 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 434 * the buffer cache we limit the eventual kva reservation to 435 * maxbcache bytes. 436 * 437 * factor represents the 1/4 x ram conversion. 438 */ 439 if (nbuf == 0) { 440 int factor = 4 * BKVASIZE / 1024; 441 442 nbuf = 50; 443 if (physmem_est > 4096) 444 nbuf += min((physmem_est - 4096) / factor, 445 65536 / factor); 446 if (physmem_est > 65536) 447 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 448 449 if (maxbcache && nbuf > maxbcache / BKVASIZE) 450 nbuf = maxbcache / BKVASIZE; 451 } 452 453#if 0 454 /* 455 * Do not allow the buffer_map to be more then 1/2 the size of the 456 * kernel_map. 457 */ 458 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 459 (BKVASIZE * 2)) { 460 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 461 (BKVASIZE * 2); 462 printf("Warning: nbufs capped at %d\n", nbuf); 463 } 464#endif 465 466 /* 467 * swbufs are used as temporary holders for I/O, such as paging I/O. 468 * We have no less then 16 and no more then 256. 469 */ 470 nswbuf = max(min(nbuf/4, 256), 16); 471#ifdef NSWBUF_MIN 472 if (nswbuf < NSWBUF_MIN) 473 nswbuf = NSWBUF_MIN; 474#endif 475#ifdef DIRECTIO 476 ffs_rawread_setup(); 477#endif 478 479 /* 480 * Reserve space for the buffer cache buffers 481 */ 482 swbuf = (void *)v; 483 v = (caddr_t)(swbuf + nswbuf); 484 buf = (void *)v; 485 v = (caddr_t)(buf + nbuf); 486 487 return(v); 488} 489 490/* Initialize the buffer subsystem. Called before use of any buffers. */ 491void 492bufinit(void) 493{ 494 struct buf *bp; 495 vm_offset_t bogus_offset; 496 int i; 497 498 GIANT_REQUIRED; 499 500 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 501 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 502 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 503 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 504 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 505 506 /* next, make a null set of free lists */ 507 for (i = 0; i < BUFFER_QUEUES; i++) 508 TAILQ_INIT(&bufqueues[i]); 509 510 /* finally, initialize each buffer header and stick on empty q */ 511 for (i = 0; i < nbuf; i++) { 512 bp = &buf[i]; 513 bzero(bp, sizeof *bp); 514 bp->b_flags = B_INVAL; /* we're just an empty header */ 515 bp->b_dev = NODEV; 516 bp->b_rcred = NOCRED; 517 bp->b_wcred = NOCRED; 518 bp->b_qindex = QUEUE_EMPTY; 519 bp->b_vflags = 0; 520 bp->b_xflags = 0; 521 LIST_INIT(&bp->b_dep); 522 BUF_LOCKINIT(bp); 523 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 524 } 525 526 /* 527 * maxbufspace is the absolute maximum amount of buffer space we are 528 * allowed to reserve in KVM and in real terms. The absolute maximum 529 * is nominally used by buf_daemon. hibufspace is the nominal maximum 530 * used by most other processes. The differential is required to 531 * ensure that buf_daemon is able to run when other processes might 532 * be blocked waiting for buffer space. 533 * 534 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 535 * this may result in KVM fragmentation which is not handled optimally 536 * by the system. 537 */ 538 maxbufspace = nbuf * BKVASIZE; 539 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 540 lobufspace = hibufspace - MAXBSIZE; 541 542 lorunningspace = 512 * 1024; 543 hirunningspace = 1024 * 1024; 544 545/* 546 * Limit the amount of malloc memory since it is wired permanently into 547 * the kernel space. Even though this is accounted for in the buffer 548 * allocation, we don't want the malloced region to grow uncontrolled. 549 * The malloc scheme improves memory utilization significantly on average 550 * (small) directories. 551 */ 552 maxbufmallocspace = hibufspace / 20; 553 554/* 555 * Reduce the chance of a deadlock occuring by limiting the number 556 * of delayed-write dirty buffers we allow to stack up. 557 */ 558 hidirtybuffers = nbuf / 4 + 20; 559 dirtybufthresh = hidirtybuffers * 9 / 10; 560 numdirtybuffers = 0; 561/* 562 * To support extreme low-memory systems, make sure hidirtybuffers cannot 563 * eat up all available buffer space. This occurs when our minimum cannot 564 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 565 * BKVASIZE'd (8K) buffers. 566 */ 567 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 568 hidirtybuffers >>= 1; 569 } 570 lodirtybuffers = hidirtybuffers / 2; 571 572/* 573 * Try to keep the number of free buffers in the specified range, 574 * and give special processes (e.g. like buf_daemon) access to an 575 * emergency reserve. 576 */ 577 lofreebuffers = nbuf / 18 + 5; 578 hifreebuffers = 2 * lofreebuffers; 579 numfreebuffers = nbuf; 580 581/* 582 * Maximum number of async ops initiated per buf_daemon loop. This is 583 * somewhat of a hack at the moment, we really need to limit ourselves 584 * based on the number of bytes of I/O in-transit that were initiated 585 * from buf_daemon. 586 */ 587 588 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 589 VM_OBJECT_LOCK(kernel_object); 590 bogus_page = vm_page_alloc(kernel_object, 591 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 592 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 593 VM_OBJECT_UNLOCK(kernel_object); 594} 595 596/* 597 * bfreekva() - free the kva allocation for a buffer. 598 * 599 * Must be called at splbio() or higher as this is the only locking for 600 * buffer_map. 601 * 602 * Since this call frees up buffer space, we call bufspacewakeup(). 603 */ 604static void 605bfreekva(struct buf * bp) 606{ 607 GIANT_REQUIRED; 608 609 if (bp->b_kvasize) { 610 atomic_add_int(&buffreekvacnt, 1); 611 atomic_subtract_int(&bufspace, bp->b_kvasize); 612 vm_map_delete(buffer_map, 613 (vm_offset_t) bp->b_kvabase, 614 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 615 ); 616 bp->b_kvasize = 0; 617 bufspacewakeup(); 618 } 619} 620 621/* 622 * bremfree: 623 * 624 * Remove the buffer from the appropriate free list. 625 */ 626void 627bremfree(struct buf * bp) 628{ 629 mtx_lock(&bqlock); 630 bremfreel(bp); 631 mtx_unlock(&bqlock); 632} 633 634void 635bremfreel(struct buf * bp) 636{ 637 int s = splbio(); 638 int old_qindex = bp->b_qindex; 639 640 GIANT_REQUIRED; 641 642 if (bp->b_qindex != QUEUE_NONE) { 643 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 644 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 645 bp->b_qindex = QUEUE_NONE; 646 } else { 647 if (BUF_REFCNT(bp) <= 1) 648 panic("bremfree: removing a buffer not on a queue"); 649 } 650 651 /* 652 * Fixup numfreebuffers count. If the buffer is invalid or not 653 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 654 * the buffer was free and we must decrement numfreebuffers. 655 */ 656 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 657 switch(old_qindex) { 658 case QUEUE_DIRTY: 659 case QUEUE_CLEAN: 660 case QUEUE_EMPTY: 661 case QUEUE_EMPTYKVA: 662 atomic_subtract_int(&numfreebuffers, 1); 663 break; 664 default: 665 break; 666 } 667 } 668 splx(s); 669} 670 671 672/* 673 * Get a buffer with the specified data. Look in the cache first. We 674 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 675 * is set, the buffer is valid and we do not have to do anything ( see 676 * getblk() ). This is really just a special case of breadn(). 677 */ 678int 679bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 680 struct buf ** bpp) 681{ 682 683 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 684} 685 686/* 687 * Operates like bread, but also starts asynchronous I/O on 688 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 689 * to initiating I/O . If B_CACHE is set, the buffer is valid 690 * and we do not have to do anything. 691 */ 692int 693breadn(struct vnode * vp, daddr_t blkno, int size, 694 daddr_t * rablkno, int *rabsize, 695 int cnt, struct ucred * cred, struct buf ** bpp) 696{ 697 struct buf *bp, *rabp; 698 int i; 699 int rv = 0, readwait = 0; 700 701 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 702 703 /* if not found in cache, do some I/O */ 704 if ((bp->b_flags & B_CACHE) == 0) { 705 if (curthread != PCPU_GET(idlethread)) 706 curthread->td_proc->p_stats->p_ru.ru_inblock++; 707 bp->b_iocmd = BIO_READ; 708 bp->b_flags &= ~B_INVAL; 709 bp->b_ioflags &= ~BIO_ERROR; 710 if (bp->b_rcred == NOCRED && cred != NOCRED) 711 bp->b_rcred = crhold(cred); 712 vfs_busy_pages(bp, 0); 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 if (vp->v_type == VCHR) 737 VOP_SPECSTRATEGY(vp, rabp); 738 else 739 VOP_STRATEGY(vp, rabp); 740 } else { 741 brelse(rabp); 742 } 743 } 744 745 if (readwait) { 746 rv = bufwait(bp); 747 } 748 return (rv); 749} 750 751/* 752 * Write, release buffer on completion. (Done by iodone 753 * if async). Do not bother writing anything if the buffer 754 * is invalid. 755 * 756 * Note that we set B_CACHE here, indicating that buffer is 757 * fully valid and thus cacheable. This is true even of NFS 758 * now so we set it generally. This could be set either here 759 * or in biodone() since the I/O is synchronous. We put it 760 * here. 761 */ 762 763int 764bwrite(struct buf * bp) 765{ 766 int oldflags, s; 767 struct buf *newbp; 768 769 if (bp->b_flags & B_INVAL) { 770 brelse(bp); 771 return (0); 772 } 773 774 oldflags = bp->b_flags; 775 776 if (BUF_REFCNT(bp) == 0) 777 panic("bwrite: buffer is not busy???"); 778 s = splbio(); 779 /* 780 * If a background write is already in progress, delay 781 * writing this block if it is asynchronous. Otherwise 782 * wait for the background write to complete. 783 */ 784 if (bp->b_xflags & BX_BKGRDINPROG) { 785 if (bp->b_flags & B_ASYNC) { 786 splx(s); 787 bdwrite(bp); 788 return (0); 789 } 790 bp->b_xflags |= BX_BKGRDWAIT; 791 tsleep(&bp->b_xflags, PRIBIO, "bwrbg", 0); 792 if (bp->b_xflags & BX_BKGRDINPROG) 793 panic("bwrite: still writing"); 794 } 795 796 /* Mark the buffer clean */ 797 bundirty(bp); 798 799 /* 800 * If this buffer is marked for background writing and we 801 * do not have to wait for it, make a copy and write the 802 * copy so as to leave this buffer ready for further use. 803 * 804 * This optimization eats a lot of memory. If we have a page 805 * or buffer shortfall we can't do it. 806 */ 807 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 808 (bp->b_flags & B_ASYNC) && 809 !vm_page_count_severe() && 810 !buf_dirty_count_severe()) { 811 if (bp->b_iodone != NULL) { 812 printf("bp->b_iodone = %p\n", bp->b_iodone); 813 panic("bwrite: need chained iodone"); 814 } 815 816 /* get a new block */ 817 newbp = geteblk(bp->b_bufsize); 818 819 /* 820 * set it to be identical to the old block. We have to 821 * set b_lblkno and BKGRDMARKER before calling bgetvp() 822 * to avoid confusing the splay tree and gbincore(). 823 */ 824 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 825 newbp->b_lblkno = bp->b_lblkno; 826 newbp->b_xflags |= BX_BKGRDMARKER; 827 /* XXX The BX_ flags need to be protected as well */ 828 VI_LOCK(bp->b_vp); 829 bgetvp(bp->b_vp, newbp); 830 VI_UNLOCK(bp->b_vp); 831 newbp->b_blkno = bp->b_blkno; 832 newbp->b_offset = bp->b_offset; 833 newbp->b_iodone = vfs_backgroundwritedone; 834 newbp->b_flags |= B_ASYNC; 835 newbp->b_flags &= ~B_INVAL; 836 837 /* move over the dependencies */ 838 if (LIST_FIRST(&bp->b_dep) != NULL) 839 buf_movedeps(bp, newbp); 840 841 /* 842 * Initiate write on the copy, release the original to 843 * the B_LOCKED queue so that it cannot go away until 844 * the background write completes. If not locked it could go 845 * away and then be reconstituted while it was being written. 846 * If the reconstituted buffer were written, we could end up 847 * with two background copies being written at the same time. 848 */ 849 bp->b_xflags |= BX_BKGRDINPROG; 850 bp->b_flags |= B_LOCKED; 851 bqrelse(bp); 852 bp = newbp; 853 } 854 855 bp->b_flags &= ~B_DONE; 856 bp->b_ioflags &= ~BIO_ERROR; 857 bp->b_flags |= B_WRITEINPROG | B_CACHE; 858 bp->b_iocmd = BIO_WRITE; 859 860 VI_LOCK(bp->b_vp); 861 bp->b_vp->v_numoutput++; 862 VI_UNLOCK(bp->b_vp); 863 vfs_busy_pages(bp, 1); 864 865 /* 866 * Normal bwrites pipeline writes 867 */ 868 bp->b_runningbufspace = bp->b_bufsize; 869 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 870 871 if (curthread != PCPU_GET(idlethread)) 872 curthread->td_proc->p_stats->p_ru.ru_oublock++; 873 splx(s); 874 if (oldflags & B_ASYNC) 875 BUF_KERNPROC(bp); 876 if (bp->b_vp->v_type == VCHR) 877 VOP_SPECSTRATEGY(bp->b_vp, bp); 878 else 879 VOP_STRATEGY(bp->b_vp, bp); 880 881 if ((oldflags & B_ASYNC) == 0) { 882 int rtval = bufwait(bp); 883 brelse(bp); 884 return (rtval); 885 } else { 886 /* 887 * don't allow the async write to saturate the I/O 888 * system. We will not deadlock here because 889 * we are blocking waiting for I/O that is already in-progress 890 * to complete. 891 */ 892 waitrunningbufspace(); 893 } 894 895 return (0); 896} 897 898/* 899 * Complete a background write started from bwrite. 900 */ 901static void 902vfs_backgroundwritedone(bp) 903 struct buf *bp; 904{ 905 struct buf *origbp; 906 907 /* 908 * Find the original buffer that we are writing. 909 */ 910 VI_LOCK(bp->b_vp); 911 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 912 panic("backgroundwritedone: lost buffer"); 913 VI_UNLOCK(bp->b_vp); 914 /* 915 * Process dependencies then return any unfinished ones. 916 */ 917 if (LIST_FIRST(&bp->b_dep) != NULL) 918 buf_complete(bp); 919 if (LIST_FIRST(&bp->b_dep) != NULL) 920 buf_movedeps(bp, origbp); 921 922 /* XXX Find out if origbp can disappear or get inconsistent */ 923 /* 924 * Clear the BX_BKGRDINPROG flag in the original buffer 925 * and awaken it if it is waiting for the write to complete. 926 * If BX_BKGRDINPROG is not set in the original buffer it must 927 * have been released and re-instantiated - which is not legal. 928 */ 929 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), 930 ("backgroundwritedone: lost buffer2")); 931 origbp->b_xflags &= ~BX_BKGRDINPROG; 932 if (origbp->b_xflags & BX_BKGRDWAIT) { 933 origbp->b_xflags &= ~BX_BKGRDWAIT; 934 wakeup(&origbp->b_xflags); 935 } 936 /* 937 * Clear the B_LOCKED flag and remove it from the locked 938 * queue if it currently resides there. 939 */ 940 origbp->b_flags &= ~B_LOCKED; 941 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) { 942 bremfree(origbp); 943 bqrelse(origbp); 944 } 945 /* 946 * This buffer is marked B_NOCACHE, so when it is released 947 * by biodone, it will be tossed. We mark it with BIO_READ 948 * to avoid biodone doing a second vwakeup. 949 */ 950 bp->b_flags |= B_NOCACHE; 951 bp->b_iocmd = BIO_READ; 952 bp->b_flags &= ~(B_CACHE | B_DONE); 953 bp->b_iodone = 0; 954 bufdone(bp); 955} 956 957/* 958 * Delayed write. (Buffer is marked dirty). Do not bother writing 959 * anything if the buffer is marked invalid. 960 * 961 * Note that since the buffer must be completely valid, we can safely 962 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 963 * biodone() in order to prevent getblk from writing the buffer 964 * out synchronously. 965 */ 966void 967bdwrite(struct buf * bp) 968{ 969 struct thread *td = curthread; 970 struct vnode *vp; 971 struct buf *nbp; 972 973 GIANT_REQUIRED; 974 975 if (BUF_REFCNT(bp) == 0) 976 panic("bdwrite: buffer is not busy"); 977 978 if (bp->b_flags & B_INVAL) { 979 brelse(bp); 980 return; 981 } 982 983 /* 984 * If we have too many dirty buffers, don't create any more. 985 * If we are wildly over our limit, then force a complete 986 * cleanup. Otherwise, just keep the situation from getting 987 * out of control. Note that we have to avoid a recursive 988 * disaster and not try to clean up after our own cleanup! 989 */ 990 vp = bp->b_vp; 991 VI_LOCK(vp); 992 if (td->td_proc->p_flag & P_COWINPROGRESS) { 993 recursiveflushes++; 994 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) { 995 VI_UNLOCK(vp); 996 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 997 VI_LOCK(vp); 998 altbufferflushes++; 999 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) { 1000 /* 1001 * Try to find a buffer to flush. 1002 */ 1003 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) { 1004 if ((nbp->b_xflags & BX_BKGRDINPROG) || 1005 buf_countdeps(nbp, 0) || 1006 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1007 continue; 1008 if (bp == nbp) 1009 panic("bdwrite: found ourselves"); 1010 VI_UNLOCK(vp); 1011 if (nbp->b_flags & B_CLUSTEROK) { 1012 vfs_bio_awrite(nbp); 1013 } else { 1014 bremfree(nbp); 1015 bawrite(nbp); 1016 } 1017 VI_LOCK(vp); 1018 dirtybufferflushes++; 1019 break; 1020 } 1021 } 1022 VI_UNLOCK(vp); 1023 1024 bdirty(bp); 1025 /* 1026 * Set B_CACHE, indicating that the buffer is fully valid. This is 1027 * true even of NFS now. 1028 */ 1029 bp->b_flags |= B_CACHE; 1030 1031 /* 1032 * This bmap keeps the system from needing to do the bmap later, 1033 * perhaps when the system is attempting to do a sync. Since it 1034 * is likely that the indirect block -- or whatever other datastructure 1035 * that the filesystem needs is still in memory now, it is a good 1036 * thing to do this. Note also, that if the pageout daemon is 1037 * requesting a sync -- there might not be enough memory to do 1038 * the bmap then... So, this is important to do. 1039 */ 1040 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1041 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1042 } 1043 1044 /* 1045 * Set the *dirty* buffer range based upon the VM system dirty pages. 1046 */ 1047 vfs_setdirty(bp); 1048 1049 /* 1050 * We need to do this here to satisfy the vnode_pager and the 1051 * pageout daemon, so that it thinks that the pages have been 1052 * "cleaned". Note that since the pages are in a delayed write 1053 * buffer -- the VFS layer "will" see that the pages get written 1054 * out on the next sync, or perhaps the cluster will be completed. 1055 */ 1056 vfs_clean_pages(bp); 1057 bqrelse(bp); 1058 1059 /* 1060 * Wakeup the buffer flushing daemon if we have a lot of dirty 1061 * buffers (midpoint between our recovery point and our stall 1062 * point). 1063 */ 1064 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1065 1066 /* 1067 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1068 * due to the softdep code. 1069 */ 1070} 1071 1072/* 1073 * bdirty: 1074 * 1075 * Turn buffer into delayed write request. We must clear BIO_READ and 1076 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1077 * itself to properly update it in the dirty/clean lists. We mark it 1078 * B_DONE to ensure that any asynchronization of the buffer properly 1079 * clears B_DONE ( else a panic will occur later ). 1080 * 1081 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1082 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1083 * should only be called if the buffer is known-good. 1084 * 1085 * Since the buffer is not on a queue, we do not update the numfreebuffers 1086 * count. 1087 * 1088 * Must be called at splbio(). 1089 * The buffer must be on QUEUE_NONE. 1090 */ 1091void 1092bdirty(bp) 1093 struct buf *bp; 1094{ 1095 KASSERT(bp->b_qindex == QUEUE_NONE, 1096 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1097 bp->b_flags &= ~(B_RELBUF); 1098 bp->b_iocmd = BIO_WRITE; 1099 1100 if ((bp->b_flags & B_DELWRI) == 0) { 1101 bp->b_flags |= B_DONE | B_DELWRI; 1102 reassignbuf(bp, bp->b_vp); 1103 atomic_add_int(&numdirtybuffers, 1); 1104 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1105 } 1106} 1107 1108/* 1109 * bundirty: 1110 * 1111 * Clear B_DELWRI for buffer. 1112 * 1113 * Since the buffer is not on a queue, we do not update the numfreebuffers 1114 * count. 1115 * 1116 * Must be called at splbio(). 1117 * The buffer must be on QUEUE_NONE. 1118 */ 1119 1120void 1121bundirty(bp) 1122 struct buf *bp; 1123{ 1124 KASSERT(bp->b_qindex == QUEUE_NONE, 1125 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1126 1127 if (bp->b_flags & B_DELWRI) { 1128 bp->b_flags &= ~B_DELWRI; 1129 reassignbuf(bp, bp->b_vp); 1130 atomic_subtract_int(&numdirtybuffers, 1); 1131 numdirtywakeup(lodirtybuffers); 1132 } 1133 /* 1134 * Since it is now being written, we can clear its deferred write flag. 1135 */ 1136 bp->b_flags &= ~B_DEFERRED; 1137} 1138 1139/* 1140 * bawrite: 1141 * 1142 * Asynchronous write. Start output on a buffer, but do not wait for 1143 * it to complete. The buffer is released when the output completes. 1144 * 1145 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1146 * B_INVAL buffers. Not us. 1147 */ 1148void 1149bawrite(struct buf * bp) 1150{ 1151 bp->b_flags |= B_ASYNC; 1152 (void) BUF_WRITE(bp); 1153} 1154 1155/* 1156 * bwillwrite: 1157 * 1158 * Called prior to the locking of any vnodes when we are expecting to 1159 * write. We do not want to starve the buffer cache with too many 1160 * dirty buffers so we block here. By blocking prior to the locking 1161 * of any vnodes we attempt to avoid the situation where a locked vnode 1162 * prevents the various system daemons from flushing related buffers. 1163 */ 1164 1165void 1166bwillwrite(void) 1167{ 1168 if (numdirtybuffers >= hidirtybuffers) { 1169 int s; 1170 1171 mtx_lock(&Giant); 1172 s = splbio(); 1173 mtx_lock(&nblock); 1174 while (numdirtybuffers >= hidirtybuffers) { 1175 bd_wakeup(1); 1176 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1177 msleep(&needsbuffer, &nblock, 1178 (PRIBIO + 4), "flswai", 0); 1179 } 1180 splx(s); 1181 mtx_unlock(&nblock); 1182 mtx_unlock(&Giant); 1183 } 1184} 1185 1186/* 1187 * Return true if we have too many dirty buffers. 1188 */ 1189int 1190buf_dirty_count_severe(void) 1191{ 1192 return(numdirtybuffers >= hidirtybuffers); 1193} 1194 1195/* 1196 * brelse: 1197 * 1198 * Release a busy buffer and, if requested, free its resources. The 1199 * buffer will be stashed in the appropriate bufqueue[] allowing it 1200 * to be accessed later as a cache entity or reused for other purposes. 1201 */ 1202void 1203brelse(struct buf * bp) 1204{ 1205 int s; 1206 1207 GIANT_REQUIRED; 1208 1209 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1210 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1211 1212 s = splbio(); 1213 1214 if (bp->b_flags & B_LOCKED) 1215 bp->b_ioflags &= ~BIO_ERROR; 1216 1217 if (bp->b_iocmd == BIO_WRITE && 1218 (bp->b_ioflags & BIO_ERROR) && 1219 !(bp->b_flags & B_INVAL)) { 1220 /* 1221 * Failed write, redirty. Must clear BIO_ERROR to prevent 1222 * pages from being scrapped. If B_INVAL is set then 1223 * this case is not run and the next case is run to 1224 * destroy the buffer. B_INVAL can occur if the buffer 1225 * is outside the range supported by the underlying device. 1226 */ 1227 bp->b_ioflags &= ~BIO_ERROR; 1228 bdirty(bp); 1229 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1230 (bp->b_ioflags & BIO_ERROR) || 1231 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1232 /* 1233 * Either a failed I/O or we were asked to free or not 1234 * cache the buffer. 1235 */ 1236 bp->b_flags |= B_INVAL; 1237 if (LIST_FIRST(&bp->b_dep) != NULL) 1238 buf_deallocate(bp); 1239 if (bp->b_flags & B_DELWRI) { 1240 atomic_subtract_int(&numdirtybuffers, 1); 1241 numdirtywakeup(lodirtybuffers); 1242 } 1243 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1244 if ((bp->b_flags & B_VMIO) == 0) { 1245 if (bp->b_bufsize) 1246 allocbuf(bp, 0); 1247 if (bp->b_vp) 1248 brelvp(bp); 1249 } 1250 } 1251 1252 /* 1253 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1254 * is called with B_DELWRI set, the underlying pages may wind up 1255 * getting freed causing a previous write (bdwrite()) to get 'lost' 1256 * because pages associated with a B_DELWRI bp are marked clean. 1257 * 1258 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1259 * if B_DELWRI is set. 1260 * 1261 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1262 * on pages to return pages to the VM page queues. 1263 */ 1264 if (bp->b_flags & B_DELWRI) 1265 bp->b_flags &= ~B_RELBUF; 1266 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1267 bp->b_flags |= B_RELBUF; 1268 1269 /* 1270 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1271 * constituted, not even NFS buffers now. Two flags effect this. If 1272 * B_INVAL, the struct buf is invalidated but the VM object is kept 1273 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1274 * 1275 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1276 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1277 * buffer is also B_INVAL because it hits the re-dirtying code above. 1278 * 1279 * Normally we can do this whether a buffer is B_DELWRI or not. If 1280 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1281 * the commit state and we cannot afford to lose the buffer. If the 1282 * buffer has a background write in progress, we need to keep it 1283 * around to prevent it from being reconstituted and starting a second 1284 * background write. 1285 */ 1286 if ((bp->b_flags & B_VMIO) 1287 && !(bp->b_vp->v_mount != NULL && 1288 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1289 !vn_isdisk(bp->b_vp, NULL) && 1290 (bp->b_flags & B_DELWRI)) 1291 ) { 1292 1293 int i, j, resid; 1294 vm_page_t m; 1295 off_t foff; 1296 vm_pindex_t poff; 1297 vm_object_t obj; 1298 struct vnode *vp; 1299 1300 vp = bp->b_vp; 1301 obj = bp->b_object; 1302 1303 /* 1304 * Get the base offset and length of the buffer. Note that 1305 * in the VMIO case if the buffer block size is not 1306 * page-aligned then b_data pointer may not be page-aligned. 1307 * But our b_pages[] array *IS* page aligned. 1308 * 1309 * block sizes less then DEV_BSIZE (usually 512) are not 1310 * supported due to the page granularity bits (m->valid, 1311 * m->dirty, etc...). 1312 * 1313 * See man buf(9) for more information 1314 */ 1315 resid = bp->b_bufsize; 1316 foff = bp->b_offset; 1317 if (obj != NULL) 1318 VM_OBJECT_LOCK(obj); 1319 for (i = 0; i < bp->b_npages; i++) { 1320 int had_bogus = 0; 1321 1322 m = bp->b_pages[i]; 1323 vm_page_lock_queues(); 1324 vm_page_flag_clear(m, PG_ZERO); 1325 vm_page_unlock_queues(); 1326 1327 /* 1328 * If we hit a bogus page, fixup *all* the bogus pages 1329 * now. 1330 */ 1331 if (m == bogus_page) { 1332 poff = OFF_TO_IDX(bp->b_offset); 1333 had_bogus = 1; 1334 1335 for (j = i; j < bp->b_npages; j++) { 1336 vm_page_t mtmp; 1337 mtmp = bp->b_pages[j]; 1338 if (mtmp == bogus_page) { 1339 mtmp = vm_page_lookup(obj, poff + j); 1340 if (!mtmp) { 1341 panic("brelse: page missing\n"); 1342 } 1343 bp->b_pages[j] = mtmp; 1344 } 1345 } 1346 1347 if ((bp->b_flags & B_INVAL) == 0) { 1348 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1349 } 1350 m = bp->b_pages[i]; 1351 } 1352 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1353 int poffset = foff & PAGE_MASK; 1354 int presid = resid > (PAGE_SIZE - poffset) ? 1355 (PAGE_SIZE - poffset) : resid; 1356 1357 KASSERT(presid >= 0, ("brelse: extra page")); 1358 vm_page_set_invalid(m, poffset, presid); 1359 if (had_bogus) 1360 printf("avoided corruption bug in bogus_page/brelse code\n"); 1361 } 1362 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1363 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1364 } 1365 if (obj != NULL) 1366 VM_OBJECT_UNLOCK(obj); 1367 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1368 vfs_vmio_release(bp); 1369 1370 } else if (bp->b_flags & B_VMIO) { 1371 1372 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1373 vfs_vmio_release(bp); 1374 } 1375 1376 } 1377 1378 if (bp->b_qindex != QUEUE_NONE) 1379 panic("brelse: free buffer onto another queue???"); 1380 if (BUF_REFCNT(bp) > 1) { 1381 /* do not release to free list */ 1382 BUF_UNLOCK(bp); 1383 splx(s); 1384 return; 1385 } 1386 1387 /* enqueue */ 1388 mtx_lock(&bqlock); 1389 1390 /* buffers with no memory */ 1391 if (bp->b_bufsize == 0) { 1392 bp->b_flags |= B_INVAL; 1393 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1394 if (bp->b_xflags & BX_BKGRDINPROG) 1395 panic("losing buffer 1"); 1396 if (bp->b_kvasize) { 1397 bp->b_qindex = QUEUE_EMPTYKVA; 1398 } else { 1399 bp->b_qindex = QUEUE_EMPTY; 1400 } 1401 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1402 bp->b_dev = NODEV; 1403 /* buffers with junk contents */ 1404 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1405 (bp->b_ioflags & BIO_ERROR)) { 1406 bp->b_flags |= B_INVAL; 1407 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1408 if (bp->b_xflags & BX_BKGRDINPROG) 1409 panic("losing buffer 2"); 1410 bp->b_qindex = QUEUE_CLEAN; 1411 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1412 bp->b_dev = NODEV; 1413 1414 /* buffers that are locked */ 1415 } else if (bp->b_flags & B_LOCKED) { 1416 bp->b_qindex = QUEUE_LOCKED; 1417 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1418 1419 /* remaining buffers */ 1420 } else { 1421 if (bp->b_flags & B_DELWRI) 1422 bp->b_qindex = QUEUE_DIRTY; 1423 else 1424 bp->b_qindex = QUEUE_CLEAN; 1425 if (bp->b_flags & B_AGE) 1426 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1427 else 1428 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1429 } 1430 mtx_unlock(&bqlock); 1431 1432 /* 1433 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1434 * placed the buffer on the correct queue. We must also disassociate 1435 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1436 * find it. 1437 */ 1438 if (bp->b_flags & B_INVAL) { 1439 if (bp->b_flags & B_DELWRI) 1440 bundirty(bp); 1441 if (bp->b_vp) 1442 brelvp(bp); 1443 } 1444 1445 /* 1446 * Fixup numfreebuffers count. The bp is on an appropriate queue 1447 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1448 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1449 * if B_INVAL is set ). 1450 */ 1451 1452 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1453 bufcountwakeup(); 1454 1455 /* 1456 * Something we can maybe free or reuse 1457 */ 1458 if (bp->b_bufsize || bp->b_kvasize) 1459 bufspacewakeup(); 1460 1461 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1462 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1463 panic("brelse: not dirty"); 1464 /* unlock */ 1465 BUF_UNLOCK(bp); 1466 splx(s); 1467} 1468 1469/* 1470 * Release a buffer back to the appropriate queue but do not try to free 1471 * it. The buffer is expected to be used again soon. 1472 * 1473 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1474 * biodone() to requeue an async I/O on completion. It is also used when 1475 * known good buffers need to be requeued but we think we may need the data 1476 * again soon. 1477 * 1478 * XXX we should be able to leave the B_RELBUF hint set on completion. 1479 */ 1480void 1481bqrelse(struct buf * bp) 1482{ 1483 int s; 1484 1485 s = splbio(); 1486 1487 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1488 1489 if (bp->b_qindex != QUEUE_NONE) 1490 panic("bqrelse: free buffer onto another queue???"); 1491 if (BUF_REFCNT(bp) > 1) { 1492 /* do not release to free list */ 1493 BUF_UNLOCK(bp); 1494 splx(s); 1495 return; 1496 } 1497 mtx_lock(&bqlock); 1498 if (bp->b_flags & B_LOCKED) { 1499 bp->b_ioflags &= ~BIO_ERROR; 1500 bp->b_qindex = QUEUE_LOCKED; 1501 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1502 /* buffers with stale but valid contents */ 1503 } else if (bp->b_flags & B_DELWRI) { 1504 bp->b_qindex = QUEUE_DIRTY; 1505 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1506 } else if (vm_page_count_severe()) { 1507 /* 1508 * We are too low on memory, we have to try to free the 1509 * buffer (most importantly: the wired pages making up its 1510 * backing store) *now*. 1511 */ 1512 mtx_unlock(&bqlock); 1513 splx(s); 1514 brelse(bp); 1515 return; 1516 } else { 1517 bp->b_qindex = QUEUE_CLEAN; 1518 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1519 } 1520 mtx_unlock(&bqlock); 1521 1522 if ((bp->b_flags & B_LOCKED) == 0 && 1523 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1524 bufcountwakeup(); 1525 } 1526 1527 /* 1528 * Something we can maybe free or reuse. 1529 */ 1530 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1531 bufspacewakeup(); 1532 1533 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1534 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1535 panic("bqrelse: not dirty"); 1536 /* unlock */ 1537 BUF_UNLOCK(bp); 1538 splx(s); 1539} 1540 1541/* Give pages used by the bp back to the VM system (where possible) */ 1542static void 1543vfs_vmio_release(bp) 1544 struct buf *bp; 1545{ 1546 int i; 1547 vm_page_t m; 1548 1549 GIANT_REQUIRED; 1550 vm_page_lock_queues(); 1551 for (i = 0; i < bp->b_npages; i++) { 1552 m = bp->b_pages[i]; 1553 bp->b_pages[i] = NULL; 1554 /* 1555 * In order to keep page LRU ordering consistent, put 1556 * everything on the inactive queue. 1557 */ 1558 vm_page_unwire(m, 0); 1559 /* 1560 * We don't mess with busy pages, it is 1561 * the responsibility of the process that 1562 * busied the pages to deal with them. 1563 */ 1564 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1565 continue; 1566 1567 if (m->wire_count == 0) { 1568 vm_page_flag_clear(m, PG_ZERO); 1569 /* 1570 * Might as well free the page if we can and it has 1571 * no valid data. We also free the page if the 1572 * buffer was used for direct I/O 1573 */ 1574 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1575 m->hold_count == 0) { 1576 vm_page_busy(m); 1577 pmap_remove_all(m); 1578 vm_page_free(m); 1579 } else if (bp->b_flags & B_DIRECT) { 1580 vm_page_try_to_free(m); 1581 } else if (vm_page_count_severe()) { 1582 vm_page_try_to_cache(m); 1583 } 1584 } 1585 } 1586 vm_page_unlock_queues(); 1587 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1588 1589 if (bp->b_bufsize) { 1590 bufspacewakeup(); 1591 bp->b_bufsize = 0; 1592 } 1593 bp->b_npages = 0; 1594 bp->b_flags &= ~B_VMIO; 1595 if (bp->b_vp) 1596 brelvp(bp); 1597} 1598 1599/* 1600 * Check to see if a block at a particular lbn is available for a clustered 1601 * write. 1602 */ 1603static int 1604vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1605{ 1606 struct buf *bpa; 1607 int match; 1608 1609 match = 0; 1610 1611 /* If the buf isn't in core skip it */ 1612 if ((bpa = gbincore(vp, lblkno)) == NULL) 1613 return (0); 1614 1615 /* If the buf is busy we don't want to wait for it */ 1616 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1617 return (0); 1618 1619 /* Only cluster with valid clusterable delayed write buffers */ 1620 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1621 (B_DELWRI | B_CLUSTEROK)) 1622 goto done; 1623 1624 if (bpa->b_bufsize != size) 1625 goto done; 1626 1627 /* 1628 * Check to see if it is in the expected place on disk and that the 1629 * block has been mapped. 1630 */ 1631 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1632 match = 1; 1633done: 1634 BUF_UNLOCK(bpa); 1635 return (match); 1636} 1637 1638/* 1639 * vfs_bio_awrite: 1640 * 1641 * Implement clustered async writes for clearing out B_DELWRI buffers. 1642 * This is much better then the old way of writing only one buffer at 1643 * a time. Note that we may not be presented with the buffers in the 1644 * correct order, so we search for the cluster in both directions. 1645 */ 1646int 1647vfs_bio_awrite(struct buf * bp) 1648{ 1649 int i; 1650 int j; 1651 daddr_t lblkno = bp->b_lblkno; 1652 struct vnode *vp = bp->b_vp; 1653 int s; 1654 int ncl; 1655 int nwritten; 1656 int size; 1657 int maxcl; 1658 1659 s = splbio(); 1660 /* 1661 * right now we support clustered writing only to regular files. If 1662 * we find a clusterable block we could be in the middle of a cluster 1663 * rather then at the beginning. 1664 */ 1665 if ((vp->v_type == VREG) && 1666 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1667 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1668 1669 size = vp->v_mount->mnt_stat.f_iosize; 1670 maxcl = MAXPHYS / size; 1671 1672 VI_LOCK(vp); 1673 for (i = 1; i < maxcl; i++) 1674 if (vfs_bio_clcheck(vp, size, lblkno + i, 1675 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1676 break; 1677 1678 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1679 if (vfs_bio_clcheck(vp, size, lblkno - j, 1680 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1681 break; 1682 1683 VI_UNLOCK(vp); 1684 --j; 1685 ncl = i + j; 1686 /* 1687 * this is a possible cluster write 1688 */ 1689 if (ncl != 1) { 1690 BUF_UNLOCK(bp); 1691 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1692 splx(s); 1693 return nwritten; 1694 } 1695 } 1696 1697 bremfree(bp); 1698 bp->b_flags |= B_ASYNC; 1699 1700 splx(s); 1701 /* 1702 * default (old) behavior, writing out only one block 1703 * 1704 * XXX returns b_bufsize instead of b_bcount for nwritten? 1705 */ 1706 nwritten = bp->b_bufsize; 1707 (void) BUF_WRITE(bp); 1708 1709 return nwritten; 1710} 1711 1712/* 1713 * getnewbuf: 1714 * 1715 * Find and initialize a new buffer header, freeing up existing buffers 1716 * in the bufqueues as necessary. The new buffer is returned locked. 1717 * 1718 * Important: B_INVAL is not set. If the caller wishes to throw the 1719 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1720 * 1721 * We block if: 1722 * We have insufficient buffer headers 1723 * We have insufficient buffer space 1724 * buffer_map is too fragmented ( space reservation fails ) 1725 * If we have to flush dirty buffers ( but we try to avoid this ) 1726 * 1727 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1728 * Instead we ask the buf daemon to do it for us. We attempt to 1729 * avoid piecemeal wakeups of the pageout daemon. 1730 */ 1731 1732static struct buf * 1733getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1734{ 1735 struct buf *bp; 1736 struct buf *nbp; 1737 int defrag = 0; 1738 int nqindex; 1739 static int flushingbufs; 1740 1741 GIANT_REQUIRED; 1742 1743 /* 1744 * We can't afford to block since we might be holding a vnode lock, 1745 * which may prevent system daemons from running. We deal with 1746 * low-memory situations by proactively returning memory and running 1747 * async I/O rather then sync I/O. 1748 */ 1749 1750 atomic_add_int(&getnewbufcalls, 1); 1751 atomic_subtract_int(&getnewbufrestarts, 1); 1752restart: 1753 atomic_add_int(&getnewbufrestarts, 1); 1754 1755 /* 1756 * Setup for scan. If we do not have enough free buffers, 1757 * we setup a degenerate case that immediately fails. Note 1758 * that if we are specially marked process, we are allowed to 1759 * dip into our reserves. 1760 * 1761 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1762 * 1763 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1764 * However, there are a number of cases (defragging, reusing, ...) 1765 * where we cannot backup. 1766 */ 1767 mtx_lock(&bqlock); 1768 nqindex = QUEUE_EMPTYKVA; 1769 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1770 1771 if (nbp == NULL) { 1772 /* 1773 * If no EMPTYKVA buffers and we are either 1774 * defragging or reusing, locate a CLEAN buffer 1775 * to free or reuse. If bufspace useage is low 1776 * skip this step so we can allocate a new buffer. 1777 */ 1778 if (defrag || bufspace >= lobufspace) { 1779 nqindex = QUEUE_CLEAN; 1780 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1781 } 1782 1783 /* 1784 * If we could not find or were not allowed to reuse a 1785 * CLEAN buffer, check to see if it is ok to use an EMPTY 1786 * buffer. We can only use an EMPTY buffer if allocating 1787 * its KVA would not otherwise run us out of buffer space. 1788 */ 1789 if (nbp == NULL && defrag == 0 && 1790 bufspace + maxsize < hibufspace) { 1791 nqindex = QUEUE_EMPTY; 1792 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1793 } 1794 } 1795 1796 /* 1797 * Run scan, possibly freeing data and/or kva mappings on the fly 1798 * depending. 1799 */ 1800 1801 while ((bp = nbp) != NULL) { 1802 int qindex = nqindex; 1803 1804 /* 1805 * Calculate next bp ( we can only use it if we do not block 1806 * or do other fancy things ). 1807 */ 1808 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1809 switch(qindex) { 1810 case QUEUE_EMPTY: 1811 nqindex = QUEUE_EMPTYKVA; 1812 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1813 break; 1814 /* FALLTHROUGH */ 1815 case QUEUE_EMPTYKVA: 1816 nqindex = QUEUE_CLEAN; 1817 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1818 break; 1819 /* FALLTHROUGH */ 1820 case QUEUE_CLEAN: 1821 /* 1822 * nbp is NULL. 1823 */ 1824 break; 1825 } 1826 } 1827 1828 /* 1829 * Sanity Checks 1830 */ 1831 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1832 1833 /* 1834 * Note: we no longer distinguish between VMIO and non-VMIO 1835 * buffers. 1836 */ 1837 1838 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1839 1840 /* 1841 * If we are defragging then we need a buffer with 1842 * b_kvasize != 0. XXX this situation should no longer 1843 * occur, if defrag is non-zero the buffer's b_kvasize 1844 * should also be non-zero at this point. XXX 1845 */ 1846 if (defrag && bp->b_kvasize == 0) { 1847 printf("Warning: defrag empty buffer %p\n", bp); 1848 continue; 1849 } 1850 1851 /* 1852 * Start freeing the bp. This is somewhat involved. nbp 1853 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1854 */ 1855 1856 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1857 panic("getnewbuf: locked buf"); 1858 bremfreel(bp); 1859 mtx_unlock(&bqlock); 1860 1861 if (qindex == QUEUE_CLEAN) { 1862 if (bp->b_flags & B_VMIO) { 1863 bp->b_flags &= ~B_ASYNC; 1864 vfs_vmio_release(bp); 1865 } 1866 if (bp->b_vp) 1867 brelvp(bp); 1868 } 1869 1870 /* 1871 * NOTE: nbp is now entirely invalid. We can only restart 1872 * the scan from this point on. 1873 * 1874 * Get the rest of the buffer freed up. b_kva* is still 1875 * valid after this operation. 1876 */ 1877 1878 if (bp->b_rcred != NOCRED) { 1879 crfree(bp->b_rcred); 1880 bp->b_rcred = NOCRED; 1881 } 1882 if (bp->b_wcred != NOCRED) { 1883 crfree(bp->b_wcred); 1884 bp->b_wcred = NOCRED; 1885 } 1886 if (LIST_FIRST(&bp->b_dep) != NULL) 1887 buf_deallocate(bp); 1888 if (bp->b_xflags & BX_BKGRDINPROG) 1889 panic("losing buffer 3"); 1890 1891 if (bp->b_bufsize) 1892 allocbuf(bp, 0); 1893 1894 bp->b_flags = 0; 1895 bp->b_ioflags = 0; 1896 bp->b_xflags = 0; 1897 bp->b_vflags = 0; 1898 bp->b_dev = NODEV; 1899 bp->b_vp = NULL; 1900 bp->b_blkno = bp->b_lblkno = 0; 1901 bp->b_offset = NOOFFSET; 1902 bp->b_iodone = 0; 1903 bp->b_error = 0; 1904 bp->b_resid = 0; 1905 bp->b_bcount = 0; 1906 bp->b_npages = 0; 1907 bp->b_dirtyoff = bp->b_dirtyend = 0; 1908 bp->b_magic = B_MAGIC_BIO; 1909 bp->b_op = &buf_ops_bio; 1910 bp->b_object = NULL; 1911 1912 LIST_INIT(&bp->b_dep); 1913 1914 /* 1915 * If we are defragging then free the buffer. 1916 */ 1917 if (defrag) { 1918 bp->b_flags |= B_INVAL; 1919 bfreekva(bp); 1920 brelse(bp); 1921 defrag = 0; 1922 goto restart; 1923 } 1924 1925 /* 1926 * If we are overcomitted then recover the buffer and its 1927 * KVM space. This occurs in rare situations when multiple 1928 * processes are blocked in getnewbuf() or allocbuf(). 1929 */ 1930 if (bufspace >= hibufspace) 1931 flushingbufs = 1; 1932 if (flushingbufs && bp->b_kvasize != 0) { 1933 bp->b_flags |= B_INVAL; 1934 bfreekva(bp); 1935 brelse(bp); 1936 goto restart; 1937 } 1938 if (bufspace < lobufspace) 1939 flushingbufs = 0; 1940 break; 1941 } 1942 1943 /* 1944 * If we exhausted our list, sleep as appropriate. We may have to 1945 * wakeup various daemons and write out some dirty buffers. 1946 * 1947 * Generally we are sleeping due to insufficient buffer space. 1948 */ 1949 1950 if (bp == NULL) { 1951 int flags; 1952 char *waitmsg; 1953 1954 mtx_unlock(&bqlock); 1955 if (defrag) { 1956 flags = VFS_BIO_NEED_BUFSPACE; 1957 waitmsg = "nbufkv"; 1958 } else if (bufspace >= hibufspace) { 1959 waitmsg = "nbufbs"; 1960 flags = VFS_BIO_NEED_BUFSPACE; 1961 } else { 1962 waitmsg = "newbuf"; 1963 flags = VFS_BIO_NEED_ANY; 1964 } 1965 1966 bd_speedup(); /* heeeelp */ 1967 1968 mtx_lock(&nblock); 1969 needsbuffer |= flags; 1970 while (needsbuffer & flags) { 1971 if (msleep(&needsbuffer, &nblock, 1972 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1973 mtx_unlock(&nblock); 1974 return (NULL); 1975 } 1976 } 1977 mtx_unlock(&nblock); 1978 } else { 1979 /* 1980 * We finally have a valid bp. We aren't quite out of the 1981 * woods, we still have to reserve kva space. In order 1982 * to keep fragmentation sane we only allocate kva in 1983 * BKVASIZE chunks. 1984 */ 1985 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1986 1987 if (maxsize != bp->b_kvasize) { 1988 vm_offset_t addr = 0; 1989 1990 bfreekva(bp); 1991 1992 if (vm_map_findspace(buffer_map, 1993 vm_map_min(buffer_map), maxsize, &addr)) { 1994 /* 1995 * Uh oh. Buffer map is to fragmented. We 1996 * must defragment the map. 1997 */ 1998 atomic_add_int(&bufdefragcnt, 1); 1999 defrag = 1; 2000 bp->b_flags |= B_INVAL; 2001 brelse(bp); 2002 goto restart; 2003 } 2004 if (addr) { 2005 vm_map_insert(buffer_map, NULL, 0, 2006 addr, addr + maxsize, 2007 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2008 2009 bp->b_kvabase = (caddr_t) addr; 2010 bp->b_kvasize = maxsize; 2011 atomic_add_int(&bufspace, bp->b_kvasize); 2012 atomic_add_int(&bufreusecnt, 1); 2013 } 2014 } 2015 bp->b_data = bp->b_kvabase; 2016 } 2017 return(bp); 2018} 2019 2020/* 2021 * buf_daemon: 2022 * 2023 * buffer flushing daemon. Buffers are normally flushed by the 2024 * update daemon but if it cannot keep up this process starts to 2025 * take the load in an attempt to prevent getnewbuf() from blocking. 2026 */ 2027 2028static struct proc *bufdaemonproc; 2029 2030static struct kproc_desc buf_kp = { 2031 "bufdaemon", 2032 buf_daemon, 2033 &bufdaemonproc 2034}; 2035SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2036 2037static void 2038buf_daemon() 2039{ 2040 int s; 2041 2042 mtx_lock(&Giant); 2043 2044 /* 2045 * This process needs to be suspended prior to shutdown sync. 2046 */ 2047 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2048 SHUTDOWN_PRI_LAST); 2049 2050 /* 2051 * This process is allowed to take the buffer cache to the limit 2052 */ 2053 s = splbio(); 2054 mtx_lock(&bdlock); 2055 2056 for (;;) { 2057 bd_request = 0; 2058 mtx_unlock(&bdlock); 2059 2060 kthread_suspend_check(bufdaemonproc); 2061 2062 /* 2063 * Do the flush. Limit the amount of in-transit I/O we 2064 * allow to build up, otherwise we would completely saturate 2065 * the I/O system. Wakeup any waiting processes before we 2066 * normally would so they can run in parallel with our drain. 2067 */ 2068 while (numdirtybuffers > lodirtybuffers) { 2069 if (flushbufqueues(0) == 0) { 2070 /* 2071 * Could not find any buffers without rollback 2072 * dependencies, so just write the first one 2073 * in the hopes of eventually making progress. 2074 */ 2075 flushbufqueues(1); 2076 break; 2077 } 2078 waitrunningbufspace(); 2079 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2080 } 2081 2082 /* 2083 * Only clear bd_request if we have reached our low water 2084 * mark. The buf_daemon normally waits 1 second and 2085 * then incrementally flushes any dirty buffers that have 2086 * built up, within reason. 2087 * 2088 * If we were unable to hit our low water mark and couldn't 2089 * find any flushable buffers, we sleep half a second. 2090 * Otherwise we loop immediately. 2091 */ 2092 mtx_lock(&bdlock); 2093 if (numdirtybuffers <= lodirtybuffers) { 2094 /* 2095 * We reached our low water mark, reset the 2096 * request and sleep until we are needed again. 2097 * The sleep is just so the suspend code works. 2098 */ 2099 bd_request = 0; 2100 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2101 } else { 2102 /* 2103 * We couldn't find any flushable dirty buffers but 2104 * still have too many dirty buffers, we 2105 * have to sleep and try again. (rare) 2106 */ 2107 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2108 } 2109 } 2110} 2111 2112/* 2113 * flushbufqueues: 2114 * 2115 * Try to flush a buffer in the dirty queue. We must be careful to 2116 * free up B_INVAL buffers instead of write them, which NFS is 2117 * particularly sensitive to. 2118 */ 2119int flushwithdeps = 0; 2120SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2121 0, "Number of buffers flushed with dependecies that require rollbacks"); 2122static int 2123flushbufqueues(int flushdeps) 2124{ 2125 struct thread *td = curthread; 2126 struct vnode *vp; 2127 struct buf *bp; 2128 int hasdeps; 2129 2130 mtx_lock(&bqlock); 2131 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2132 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2133 continue; 2134 KASSERT((bp->b_flags & B_DELWRI), 2135 ("unexpected clean buffer %p", bp)); 2136 if ((bp->b_xflags & BX_BKGRDINPROG) != 0) { 2137 BUF_UNLOCK(bp); 2138 continue; 2139 } 2140 if (bp->b_flags & B_INVAL) { 2141 bremfreel(bp); 2142 mtx_unlock(&bqlock); 2143 brelse(bp); 2144 return (1); 2145 } 2146 2147 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2148 if (flushdeps == 0) { 2149 BUF_UNLOCK(bp); 2150 continue; 2151 } 2152 hasdeps = 1; 2153 } else 2154 hasdeps = 0; 2155 /* 2156 * We must hold the lock on a vnode before writing 2157 * one of its buffers. Otherwise we may confuse, or 2158 * in the case of a snapshot vnode, deadlock the 2159 * system. 2160 * 2161 * The lock order here is the reverse of the normal 2162 * of vnode followed by buf lock. This is ok because 2163 * the NOWAIT will prevent deadlock. 2164 */ 2165 if ((vp = bp->b_vp) == NULL || 2166 vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2167 mtx_unlock(&bqlock); 2168 vfs_bio_awrite(bp); 2169 if (vp != NULL) 2170 VOP_UNLOCK(vp, 0, td); 2171 flushwithdeps += hasdeps; 2172 return (1); 2173 } 2174 BUF_UNLOCK(bp); 2175 } 2176 mtx_unlock(&bqlock); 2177 return (0); 2178} 2179 2180/* 2181 * Check to see if a block is currently memory resident. 2182 */ 2183struct buf * 2184incore(struct vnode * vp, daddr_t blkno) 2185{ 2186 struct buf *bp; 2187 2188 int s = splbio(); 2189 VI_LOCK(vp); 2190 bp = gbincore(vp, blkno); 2191 VI_UNLOCK(vp); 2192 splx(s); 2193 return (bp); 2194} 2195 2196/* 2197 * Returns true if no I/O is needed to access the 2198 * associated VM object. This is like incore except 2199 * it also hunts around in the VM system for the data. 2200 */ 2201 2202int 2203inmem(struct vnode * vp, daddr_t blkno) 2204{ 2205 vm_object_t obj; 2206 vm_offset_t toff, tinc, size; 2207 vm_page_t m; 2208 vm_ooffset_t off; 2209 2210 GIANT_REQUIRED; 2211 ASSERT_VOP_LOCKED(vp, "inmem"); 2212 2213 if (incore(vp, blkno)) 2214 return 1; 2215 if (vp->v_mount == NULL) 2216 return 0; 2217 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2218 return 0; 2219 2220 size = PAGE_SIZE; 2221 if (size > vp->v_mount->mnt_stat.f_iosize) 2222 size = vp->v_mount->mnt_stat.f_iosize; 2223 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2224 2225 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2226 VM_OBJECT_LOCK(obj); 2227 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2228 VM_OBJECT_UNLOCK(obj); 2229 if (!m) 2230 goto notinmem; 2231 tinc = size; 2232 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2233 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2234 if (vm_page_is_valid(m, 2235 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2236 goto notinmem; 2237 } 2238 return 1; 2239 2240notinmem: 2241 return (0); 2242} 2243 2244/* 2245 * vfs_setdirty: 2246 * 2247 * Sets the dirty range for a buffer based on the status of the dirty 2248 * bits in the pages comprising the buffer. 2249 * 2250 * The range is limited to the size of the buffer. 2251 * 2252 * This routine is primarily used by NFS, but is generalized for the 2253 * B_VMIO case. 2254 */ 2255static void 2256vfs_setdirty(struct buf *bp) 2257{ 2258 int i; 2259 vm_object_t object; 2260 2261 GIANT_REQUIRED; 2262 /* 2263 * Degenerate case - empty buffer 2264 */ 2265 2266 if (bp->b_bufsize == 0) 2267 return; 2268 2269 /* 2270 * We qualify the scan for modified pages on whether the 2271 * object has been flushed yet. The OBJ_WRITEABLE flag 2272 * is not cleared simply by protecting pages off. 2273 */ 2274 2275 if ((bp->b_flags & B_VMIO) == 0) 2276 return; 2277 2278 object = bp->b_pages[0]->object; 2279 VM_OBJECT_LOCK(object); 2280 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2281 printf("Warning: object %p writeable but not mightbedirty\n", object); 2282 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2283 printf("Warning: object %p mightbedirty but not writeable\n", object); 2284 2285 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2286 vm_offset_t boffset; 2287 vm_offset_t eoffset; 2288 2289 vm_page_lock_queues(); 2290 /* 2291 * test the pages to see if they have been modified directly 2292 * by users through the VM system. 2293 */ 2294 for (i = 0; i < bp->b_npages; i++) { 2295 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2296 vm_page_test_dirty(bp->b_pages[i]); 2297 } 2298 2299 /* 2300 * Calculate the encompassing dirty range, boffset and eoffset, 2301 * (eoffset - boffset) bytes. 2302 */ 2303 2304 for (i = 0; i < bp->b_npages; i++) { 2305 if (bp->b_pages[i]->dirty) 2306 break; 2307 } 2308 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2309 2310 for (i = bp->b_npages - 1; i >= 0; --i) { 2311 if (bp->b_pages[i]->dirty) { 2312 break; 2313 } 2314 } 2315 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2316 2317 vm_page_unlock_queues(); 2318 /* 2319 * Fit it to the buffer. 2320 */ 2321 2322 if (eoffset > bp->b_bcount) 2323 eoffset = bp->b_bcount; 2324 2325 /* 2326 * If we have a good dirty range, merge with the existing 2327 * dirty range. 2328 */ 2329 2330 if (boffset < eoffset) { 2331 if (bp->b_dirtyoff > boffset) 2332 bp->b_dirtyoff = boffset; 2333 if (bp->b_dirtyend < eoffset) 2334 bp->b_dirtyend = eoffset; 2335 } 2336 } 2337 VM_OBJECT_UNLOCK(object); 2338} 2339 2340/* 2341 * getblk: 2342 * 2343 * Get a block given a specified block and offset into a file/device. 2344 * The buffers B_DONE bit will be cleared on return, making it almost 2345 * ready for an I/O initiation. B_INVAL may or may not be set on 2346 * return. The caller should clear B_INVAL prior to initiating a 2347 * READ. 2348 * 2349 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2350 * an existing buffer. 2351 * 2352 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2353 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2354 * and then cleared based on the backing VM. If the previous buffer is 2355 * non-0-sized but invalid, B_CACHE will be cleared. 2356 * 2357 * If getblk() must create a new buffer, the new buffer is returned with 2358 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2359 * case it is returned with B_INVAL clear and B_CACHE set based on the 2360 * backing VM. 2361 * 2362 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2363 * B_CACHE bit is clear. 2364 * 2365 * What this means, basically, is that the caller should use B_CACHE to 2366 * determine whether the buffer is fully valid or not and should clear 2367 * B_INVAL prior to issuing a read. If the caller intends to validate 2368 * the buffer by loading its data area with something, the caller needs 2369 * to clear B_INVAL. If the caller does this without issuing an I/O, 2370 * the caller should set B_CACHE ( as an optimization ), else the caller 2371 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2372 * a write attempt or if it was a successfull read. If the caller 2373 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2374 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2375 */ 2376struct buf * 2377getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2378 int flags) 2379{ 2380 struct buf *bp; 2381 int s; 2382 int error; 2383 ASSERT_VOP_LOCKED(vp, "getblk"); 2384 2385 if (size > MAXBSIZE) 2386 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2387 2388 s = splbio(); 2389loop: 2390 /* 2391 * Block if we are low on buffers. Certain processes are allowed 2392 * to completely exhaust the buffer cache. 2393 * 2394 * If this check ever becomes a bottleneck it may be better to 2395 * move it into the else, when gbincore() fails. At the moment 2396 * it isn't a problem. 2397 * 2398 * XXX remove if 0 sections (clean this up after its proven) 2399 */ 2400 if (numfreebuffers == 0) { 2401 if (curthread == PCPU_GET(idlethread)) 2402 return NULL; 2403 mtx_lock(&nblock); 2404 needsbuffer |= VFS_BIO_NEED_ANY; 2405 mtx_unlock(&nblock); 2406 } 2407 2408 VI_LOCK(vp); 2409 if ((bp = gbincore(vp, blkno))) { 2410 int lockflags; 2411 /* 2412 * Buffer is in-core. If the buffer is not busy, it must 2413 * be on a queue. 2414 */ 2415 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2416 2417 if (flags & GB_LOCK_NOWAIT) 2418 lockflags |= LK_NOWAIT; 2419 2420 error = BUF_TIMELOCK(bp, lockflags, 2421 VI_MTX(vp), "getblk", slpflag, slptimeo); 2422 2423 /* 2424 * If we slept and got the lock we have to restart in case 2425 * the buffer changed identities. 2426 */ 2427 if (error == ENOLCK) 2428 goto loop; 2429 /* We timed out or were interrupted. */ 2430 else if (error) 2431 return (NULL); 2432 2433 /* 2434 * The buffer is locked. B_CACHE is cleared if the buffer is 2435 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2436 * and for a VMIO buffer B_CACHE is adjusted according to the 2437 * backing VM cache. 2438 */ 2439 if (bp->b_flags & B_INVAL) 2440 bp->b_flags &= ~B_CACHE; 2441 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2442 bp->b_flags |= B_CACHE; 2443 bremfree(bp); 2444 2445 /* 2446 * check for size inconsistancies for non-VMIO case. 2447 */ 2448 2449 if (bp->b_bcount != size) { 2450 if ((bp->b_flags & B_VMIO) == 0 || 2451 (size > bp->b_kvasize)) { 2452 if (bp->b_flags & B_DELWRI) { 2453 bp->b_flags |= B_NOCACHE; 2454 BUF_WRITE(bp); 2455 } else { 2456 if ((bp->b_flags & B_VMIO) && 2457 (LIST_FIRST(&bp->b_dep) == NULL)) { 2458 bp->b_flags |= B_RELBUF; 2459 brelse(bp); 2460 } else { 2461 bp->b_flags |= B_NOCACHE; 2462 BUF_WRITE(bp); 2463 } 2464 } 2465 goto loop; 2466 } 2467 } 2468 2469 /* 2470 * If the size is inconsistant in the VMIO case, we can resize 2471 * the buffer. This might lead to B_CACHE getting set or 2472 * cleared. If the size has not changed, B_CACHE remains 2473 * unchanged from its previous state. 2474 */ 2475 2476 if (bp->b_bcount != size) 2477 allocbuf(bp, size); 2478 2479 KASSERT(bp->b_offset != NOOFFSET, 2480 ("getblk: no buffer offset")); 2481 2482 /* 2483 * A buffer with B_DELWRI set and B_CACHE clear must 2484 * be committed before we can return the buffer in 2485 * order to prevent the caller from issuing a read 2486 * ( due to B_CACHE not being set ) and overwriting 2487 * it. 2488 * 2489 * Most callers, including NFS and FFS, need this to 2490 * operate properly either because they assume they 2491 * can issue a read if B_CACHE is not set, or because 2492 * ( for example ) an uncached B_DELWRI might loop due 2493 * to softupdates re-dirtying the buffer. In the latter 2494 * case, B_CACHE is set after the first write completes, 2495 * preventing further loops. 2496 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2497 * above while extending the buffer, we cannot allow the 2498 * buffer to remain with B_CACHE set after the write 2499 * completes or it will represent a corrupt state. To 2500 * deal with this we set B_NOCACHE to scrap the buffer 2501 * after the write. 2502 * 2503 * We might be able to do something fancy, like setting 2504 * B_CACHE in bwrite() except if B_DELWRI is already set, 2505 * so the below call doesn't set B_CACHE, but that gets real 2506 * confusing. This is much easier. 2507 */ 2508 2509 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2510 bp->b_flags |= B_NOCACHE; 2511 BUF_WRITE(bp); 2512 goto loop; 2513 } 2514 2515 splx(s); 2516 bp->b_flags &= ~B_DONE; 2517 } else { 2518 int bsize, maxsize, vmio; 2519 off_t offset; 2520 2521 /* 2522 * Buffer is not in-core, create new buffer. The buffer 2523 * returned by getnewbuf() is locked. Note that the returned 2524 * buffer is also considered valid (not marked B_INVAL). 2525 */ 2526 VI_UNLOCK(vp); 2527 if (vn_isdisk(vp, NULL)) 2528 bsize = DEV_BSIZE; 2529 else if (vp->v_mountedhere) 2530 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2531 else if (vp->v_mount) 2532 bsize = vp->v_mount->mnt_stat.f_iosize; 2533 else 2534 bsize = size; 2535 2536 offset = blkno * bsize; 2537 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2538 (vp->v_vflag & VV_OBJBUF); 2539 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2540 maxsize = imax(maxsize, bsize); 2541 2542 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2543 if (slpflag || slptimeo) { 2544 splx(s); 2545 return NULL; 2546 } 2547 goto loop; 2548 } 2549 2550 /* 2551 * This code is used to make sure that a buffer is not 2552 * created while the getnewbuf routine is blocked. 2553 * This can be a problem whether the vnode is locked or not. 2554 * If the buffer is created out from under us, we have to 2555 * throw away the one we just created. There is now window 2556 * race because we are safely running at splbio() from the 2557 * point of the duplicate buffer creation through to here, 2558 * and we've locked the buffer. 2559 * 2560 * Note: this must occur before we associate the buffer 2561 * with the vp especially considering limitations in 2562 * the splay tree implementation when dealing with duplicate 2563 * lblkno's. 2564 */ 2565 VI_LOCK(vp); 2566 if (gbincore(vp, blkno)) { 2567 VI_UNLOCK(vp); 2568 bp->b_flags |= B_INVAL; 2569 brelse(bp); 2570 goto loop; 2571 } 2572 2573 /* 2574 * Insert the buffer into the hash, so that it can 2575 * be found by incore. 2576 */ 2577 bp->b_blkno = bp->b_lblkno = blkno; 2578 bp->b_offset = offset; 2579 2580 bgetvp(vp, bp); 2581 VI_UNLOCK(vp); 2582 2583 /* 2584 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2585 * buffer size starts out as 0, B_CACHE will be set by 2586 * allocbuf() for the VMIO case prior to it testing the 2587 * backing store for validity. 2588 */ 2589 2590 if (vmio) { 2591 bp->b_flags |= B_VMIO; 2592#if defined(VFS_BIO_DEBUG) 2593 if (vp->v_type != VREG) 2594 printf("getblk: vmioing file type %d???\n", vp->v_type); 2595#endif 2596 VOP_GETVOBJECT(vp, &bp->b_object); 2597 } else { 2598 bp->b_flags &= ~B_VMIO; 2599 bp->b_object = NULL; 2600 } 2601 2602 allocbuf(bp, size); 2603 2604 splx(s); 2605 bp->b_flags &= ~B_DONE; 2606 } 2607 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2608 return (bp); 2609} 2610 2611/* 2612 * Get an empty, disassociated buffer of given size. The buffer is initially 2613 * set to B_INVAL. 2614 */ 2615struct buf * 2616geteblk(int size) 2617{ 2618 struct buf *bp; 2619 int s; 2620 int maxsize; 2621 2622 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2623 2624 s = splbio(); 2625 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2626 continue; 2627 splx(s); 2628 allocbuf(bp, size); 2629 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2630 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2631 return (bp); 2632} 2633 2634 2635/* 2636 * This code constitutes the buffer memory from either anonymous system 2637 * memory (in the case of non-VMIO operations) or from an associated 2638 * VM object (in the case of VMIO operations). This code is able to 2639 * resize a buffer up or down. 2640 * 2641 * Note that this code is tricky, and has many complications to resolve 2642 * deadlock or inconsistant data situations. Tread lightly!!! 2643 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2644 * the caller. Calling this code willy nilly can result in the loss of data. 2645 * 2646 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2647 * B_CACHE for the non-VMIO case. 2648 */ 2649 2650int 2651allocbuf(struct buf *bp, int size) 2652{ 2653 int newbsize, mbsize; 2654 int i; 2655 2656 GIANT_REQUIRED; 2657 2658 if (BUF_REFCNT(bp) == 0) 2659 panic("allocbuf: buffer not busy"); 2660 2661 if (bp->b_kvasize < size) 2662 panic("allocbuf: buffer too small"); 2663 2664 if ((bp->b_flags & B_VMIO) == 0) { 2665 caddr_t origbuf; 2666 int origbufsize; 2667 /* 2668 * Just get anonymous memory from the kernel. Don't 2669 * mess with B_CACHE. 2670 */ 2671 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2672 if (bp->b_flags & B_MALLOC) 2673 newbsize = mbsize; 2674 else 2675 newbsize = round_page(size); 2676 2677 if (newbsize < bp->b_bufsize) { 2678 /* 2679 * malloced buffers are not shrunk 2680 */ 2681 if (bp->b_flags & B_MALLOC) { 2682 if (newbsize) { 2683 bp->b_bcount = size; 2684 } else { 2685 free(bp->b_data, M_BIOBUF); 2686 if (bp->b_bufsize) { 2687 atomic_subtract_int( 2688 &bufmallocspace, 2689 bp->b_bufsize); 2690 bufspacewakeup(); 2691 bp->b_bufsize = 0; 2692 } 2693 bp->b_data = bp->b_kvabase; 2694 bp->b_bcount = 0; 2695 bp->b_flags &= ~B_MALLOC; 2696 } 2697 return 1; 2698 } 2699 vm_hold_free_pages( 2700 bp, 2701 (vm_offset_t) bp->b_data + newbsize, 2702 (vm_offset_t) bp->b_data + bp->b_bufsize); 2703 } else if (newbsize > bp->b_bufsize) { 2704 /* 2705 * We only use malloced memory on the first allocation. 2706 * and revert to page-allocated memory when the buffer 2707 * grows. 2708 */ 2709 /* 2710 * There is a potential smp race here that could lead 2711 * to bufmallocspace slightly passing the max. It 2712 * is probably extremely rare and not worth worrying 2713 * over. 2714 */ 2715 if ( (bufmallocspace < maxbufmallocspace) && 2716 (bp->b_bufsize == 0) && 2717 (mbsize <= PAGE_SIZE/2)) { 2718 2719 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2720 bp->b_bufsize = mbsize; 2721 bp->b_bcount = size; 2722 bp->b_flags |= B_MALLOC; 2723 atomic_add_int(&bufmallocspace, mbsize); 2724 return 1; 2725 } 2726 origbuf = NULL; 2727 origbufsize = 0; 2728 /* 2729 * If the buffer is growing on its other-than-first allocation, 2730 * then we revert to the page-allocation scheme. 2731 */ 2732 if (bp->b_flags & B_MALLOC) { 2733 origbuf = bp->b_data; 2734 origbufsize = bp->b_bufsize; 2735 bp->b_data = bp->b_kvabase; 2736 if (bp->b_bufsize) { 2737 atomic_subtract_int(&bufmallocspace, 2738 bp->b_bufsize); 2739 bufspacewakeup(); 2740 bp->b_bufsize = 0; 2741 } 2742 bp->b_flags &= ~B_MALLOC; 2743 newbsize = round_page(newbsize); 2744 } 2745 vm_hold_load_pages( 2746 bp, 2747 (vm_offset_t) bp->b_data + bp->b_bufsize, 2748 (vm_offset_t) bp->b_data + newbsize); 2749 if (origbuf) { 2750 bcopy(origbuf, bp->b_data, origbufsize); 2751 free(origbuf, M_BIOBUF); 2752 } 2753 } 2754 } else { 2755 int desiredpages; 2756 2757 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2758 desiredpages = (size == 0) ? 0 : 2759 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2760 2761 if (bp->b_flags & B_MALLOC) 2762 panic("allocbuf: VMIO buffer can't be malloced"); 2763 /* 2764 * Set B_CACHE initially if buffer is 0 length or will become 2765 * 0-length. 2766 */ 2767 if (size == 0 || bp->b_bufsize == 0) 2768 bp->b_flags |= B_CACHE; 2769 2770 if (newbsize < bp->b_bufsize) { 2771 /* 2772 * DEV_BSIZE aligned new buffer size is less then the 2773 * DEV_BSIZE aligned existing buffer size. Figure out 2774 * if we have to remove any pages. 2775 */ 2776 if (desiredpages < bp->b_npages) { 2777 vm_page_t m; 2778 2779 vm_page_lock_queues(); 2780 for (i = desiredpages; i < bp->b_npages; i++) { 2781 /* 2782 * the page is not freed here -- it 2783 * is the responsibility of 2784 * vnode_pager_setsize 2785 */ 2786 m = bp->b_pages[i]; 2787 KASSERT(m != bogus_page, 2788 ("allocbuf: bogus page found")); 2789 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2790 vm_page_lock_queues(); 2791 2792 bp->b_pages[i] = NULL; 2793 vm_page_unwire(m, 0); 2794 } 2795 vm_page_unlock_queues(); 2796 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2797 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2798 bp->b_npages = desiredpages; 2799 } 2800 } else if (size > bp->b_bcount) { 2801 /* 2802 * We are growing the buffer, possibly in a 2803 * byte-granular fashion. 2804 */ 2805 struct vnode *vp; 2806 vm_object_t obj; 2807 vm_offset_t toff; 2808 vm_offset_t tinc; 2809 2810 /* 2811 * Step 1, bring in the VM pages from the object, 2812 * allocating them if necessary. We must clear 2813 * B_CACHE if these pages are not valid for the 2814 * range covered by the buffer. 2815 */ 2816 2817 vp = bp->b_vp; 2818 obj = bp->b_object; 2819 2820 VM_OBJECT_LOCK(obj); 2821 while (bp->b_npages < desiredpages) { 2822 vm_page_t m; 2823 vm_pindex_t pi; 2824 2825 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2826 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2827 /* 2828 * note: must allocate system pages 2829 * since blocking here could intefere 2830 * with paging I/O, no matter which 2831 * process we are. 2832 */ 2833 m = vm_page_alloc(obj, pi, 2834 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2835 if (m == NULL) { 2836 atomic_add_int(&vm_pageout_deficit, 2837 desiredpages - bp->b_npages); 2838 VM_OBJECT_UNLOCK(obj); 2839 VM_WAIT; 2840 VM_OBJECT_LOCK(obj); 2841 } else { 2842 vm_page_lock_queues(); 2843 vm_page_wakeup(m); 2844 vm_page_unlock_queues(); 2845 bp->b_flags &= ~B_CACHE; 2846 bp->b_pages[bp->b_npages] = m; 2847 ++bp->b_npages; 2848 } 2849 continue; 2850 } 2851 2852 /* 2853 * We found a page. If we have to sleep on it, 2854 * retry because it might have gotten freed out 2855 * from under us. 2856 * 2857 * We can only test PG_BUSY here. Blocking on 2858 * m->busy might lead to a deadlock: 2859 * 2860 * vm_fault->getpages->cluster_read->allocbuf 2861 * 2862 */ 2863 vm_page_lock_queues(); 2864 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2865 continue; 2866 2867 /* 2868 * We have a good page. Should we wakeup the 2869 * page daemon? 2870 */ 2871 if ((curproc != pageproc) && 2872 ((m->queue - m->pc) == PQ_CACHE) && 2873 ((cnt.v_free_count + cnt.v_cache_count) < 2874 (cnt.v_free_min + cnt.v_cache_min))) { 2875 pagedaemon_wakeup(); 2876 } 2877 vm_page_flag_clear(m, PG_ZERO); 2878 vm_page_wire(m); 2879 vm_page_unlock_queues(); 2880 bp->b_pages[bp->b_npages] = m; 2881 ++bp->b_npages; 2882 } 2883 VM_OBJECT_UNLOCK(obj); 2884 2885 /* 2886 * Step 2. We've loaded the pages into the buffer, 2887 * we have to figure out if we can still have B_CACHE 2888 * set. Note that B_CACHE is set according to the 2889 * byte-granular range ( bcount and size ), new the 2890 * aligned range ( newbsize ). 2891 * 2892 * The VM test is against m->valid, which is DEV_BSIZE 2893 * aligned. Needless to say, the validity of the data 2894 * needs to also be DEV_BSIZE aligned. Note that this 2895 * fails with NFS if the server or some other client 2896 * extends the file's EOF. If our buffer is resized, 2897 * B_CACHE may remain set! XXX 2898 */ 2899 2900 toff = bp->b_bcount; 2901 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2902 2903 while ((bp->b_flags & B_CACHE) && toff < size) { 2904 vm_pindex_t pi; 2905 2906 if (tinc > (size - toff)) 2907 tinc = size - toff; 2908 2909 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2910 PAGE_SHIFT; 2911 2912 vfs_buf_test_cache( 2913 bp, 2914 bp->b_offset, 2915 toff, 2916 tinc, 2917 bp->b_pages[pi] 2918 ); 2919 toff += tinc; 2920 tinc = PAGE_SIZE; 2921 } 2922 2923 /* 2924 * Step 3, fixup the KVM pmap. Remember that 2925 * bp->b_data is relative to bp->b_offset, but 2926 * bp->b_offset may be offset into the first page. 2927 */ 2928 2929 bp->b_data = (caddr_t) 2930 trunc_page((vm_offset_t)bp->b_data); 2931 pmap_qenter( 2932 (vm_offset_t)bp->b_data, 2933 bp->b_pages, 2934 bp->b_npages 2935 ); 2936 2937 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2938 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2939 } 2940 } 2941 if (newbsize < bp->b_bufsize) 2942 bufspacewakeup(); 2943 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2944 bp->b_bcount = size; /* requested buffer size */ 2945 return 1; 2946} 2947 2948void 2949biodone(struct bio *bp) 2950{ 2951 mtx_lock(&bdonelock); 2952 bp->bio_flags |= BIO_DONE; 2953 if (bp->bio_done == NULL) 2954 wakeup(bp); 2955 mtx_unlock(&bdonelock); 2956 if (bp->bio_done != NULL) 2957 bp->bio_done(bp); 2958} 2959 2960/* 2961 * Wait for a BIO to finish. 2962 * 2963 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2964 * case is not yet clear. 2965 */ 2966int 2967biowait(struct bio *bp, const char *wchan) 2968{ 2969 2970 mtx_lock(&bdonelock); 2971 while ((bp->bio_flags & BIO_DONE) == 0) 2972 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 2973 mtx_unlock(&bdonelock); 2974 if (bp->bio_error != 0) 2975 return (bp->bio_error); 2976 if (!(bp->bio_flags & BIO_ERROR)) 2977 return (0); 2978 return (EIO); 2979} 2980 2981void 2982biofinish(struct bio *bp, struct devstat *stat, int error) 2983{ 2984 2985 if (error) { 2986 bp->bio_error = error; 2987 bp->bio_flags |= BIO_ERROR; 2988 } 2989 if (stat != NULL) 2990 devstat_end_transaction_bio(stat, bp); 2991 biodone(bp); 2992} 2993 2994/* 2995 * bufwait: 2996 * 2997 * Wait for buffer I/O completion, returning error status. The buffer 2998 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 2999 * error and cleared. 3000 */ 3001int 3002bufwait(register struct buf * bp) 3003{ 3004 int s; 3005 3006 s = splbio(); 3007 if (bp->b_iocmd == BIO_READ) 3008 bwait(bp, PRIBIO, "biord"); 3009 else 3010 bwait(bp, PRIBIO, "biowr"); 3011 splx(s); 3012 if (bp->b_flags & B_EINTR) { 3013 bp->b_flags &= ~B_EINTR; 3014 return (EINTR); 3015 } 3016 if (bp->b_ioflags & BIO_ERROR) { 3017 return (bp->b_error ? bp->b_error : EIO); 3018 } else { 3019 return (0); 3020 } 3021} 3022 3023 /* 3024 * Call back function from struct bio back up to struct buf. 3025 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 3026 */ 3027void 3028bufdonebio(struct bio *bp) 3029{ 3030 bufdone(bp->bio_caller2); 3031} 3032 3033/* 3034 * bufdone: 3035 * 3036 * Finish I/O on a buffer, optionally calling a completion function. 3037 * This is usually called from an interrupt so process blocking is 3038 * not allowed. 3039 * 3040 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3041 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3042 * assuming B_INVAL is clear. 3043 * 3044 * For the VMIO case, we set B_CACHE if the op was a read and no 3045 * read error occured, or if the op was a write. B_CACHE is never 3046 * set if the buffer is invalid or otherwise uncacheable. 3047 * 3048 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3049 * initiator to leave B_INVAL set to brelse the buffer out of existance 3050 * in the biodone routine. 3051 */ 3052void 3053bufdone(struct buf *bp) 3054{ 3055 int s; 3056 void (*biodone)(struct buf *); 3057 3058 GIANT_REQUIRED; 3059 3060 s = splbio(); 3061 3062 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3063 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3064 3065 bp->b_flags |= B_DONE; 3066 runningbufwakeup(bp); 3067 3068 if (bp->b_iocmd == BIO_DELETE) { 3069 brelse(bp); 3070 splx(s); 3071 return; 3072 } 3073 3074 if (bp->b_iocmd == BIO_WRITE) { 3075 vwakeup(bp); 3076 } 3077 3078 /* call optional completion function if requested */ 3079 if (bp->b_iodone != NULL) { 3080 biodone = bp->b_iodone; 3081 bp->b_iodone = NULL; 3082 (*biodone) (bp); 3083 splx(s); 3084 return; 3085 } 3086 if (LIST_FIRST(&bp->b_dep) != NULL) 3087 buf_complete(bp); 3088 3089 if (bp->b_flags & B_VMIO) { 3090 int i; 3091 vm_ooffset_t foff; 3092 vm_page_t m; 3093 vm_object_t obj; 3094 int iosize; 3095 struct vnode *vp = bp->b_vp; 3096 3097 obj = bp->b_object; 3098 3099#if defined(VFS_BIO_DEBUG) 3100 mp_fixme("usecount and vflag accessed without locks."); 3101 if (vp->v_usecount == 0) { 3102 panic("biodone: zero vnode ref count"); 3103 } 3104 3105 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3106 panic("biodone: vnode is not setup for merged cache"); 3107 } 3108#endif 3109 3110 foff = bp->b_offset; 3111 KASSERT(bp->b_offset != NOOFFSET, 3112 ("biodone: no buffer offset")); 3113 3114 if (obj != NULL) 3115 VM_OBJECT_LOCK(obj); 3116#if defined(VFS_BIO_DEBUG) 3117 if (obj->paging_in_progress < bp->b_npages) { 3118 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3119 obj->paging_in_progress, bp->b_npages); 3120 } 3121#endif 3122 3123 /* 3124 * Set B_CACHE if the op was a normal read and no error 3125 * occured. B_CACHE is set for writes in the b*write() 3126 * routines. 3127 */ 3128 iosize = bp->b_bcount - bp->b_resid; 3129 if (bp->b_iocmd == BIO_READ && 3130 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3131 !(bp->b_ioflags & BIO_ERROR)) { 3132 bp->b_flags |= B_CACHE; 3133 } 3134 vm_page_lock_queues(); 3135 for (i = 0; i < bp->b_npages; i++) { 3136 int bogusflag = 0; 3137 int resid; 3138 3139 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3140 if (resid > iosize) 3141 resid = iosize; 3142 3143 /* 3144 * cleanup bogus pages, restoring the originals 3145 */ 3146 m = bp->b_pages[i]; 3147 if (m == bogus_page) { 3148 bogusflag = 1; 3149 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3150 if (m == NULL) 3151 panic("biodone: page disappeared!"); 3152 bp->b_pages[i] = m; 3153 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3154 } 3155#if defined(VFS_BIO_DEBUG) 3156 if (OFF_TO_IDX(foff) != m->pindex) { 3157 printf( 3158"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3159 (intmax_t)foff, (uintmax_t)m->pindex); 3160 } 3161#endif 3162 3163 /* 3164 * In the write case, the valid and clean bits are 3165 * already changed correctly ( see bdwrite() ), so we 3166 * only need to do this here in the read case. 3167 */ 3168 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3169 vfs_page_set_valid(bp, foff, i, m); 3170 } 3171 vm_page_flag_clear(m, PG_ZERO); 3172 3173 /* 3174 * when debugging new filesystems or buffer I/O methods, this 3175 * is the most common error that pops up. if you see this, you 3176 * have not set the page busy flag correctly!!! 3177 */ 3178 if (m->busy == 0) { 3179 printf("biodone: page busy < 0, " 3180 "pindex: %d, foff: 0x(%x,%x), " 3181 "resid: %d, index: %d\n", 3182 (int) m->pindex, (int)(foff >> 32), 3183 (int) foff & 0xffffffff, resid, i); 3184 if (!vn_isdisk(vp, NULL)) 3185 printf(" iosize: %ld, lblkno: %jd, flags: 0x%x, npages: %d\n", 3186 bp->b_vp->v_mount->mnt_stat.f_iosize, 3187 (intmax_t) bp->b_lblkno, 3188 bp->b_flags, bp->b_npages); 3189 else 3190 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3191 (intmax_t) bp->b_lblkno, 3192 bp->b_flags, bp->b_npages); 3193 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 3194 m->valid, m->dirty, m->wire_count); 3195 panic("biodone: page busy < 0\n"); 3196 } 3197 vm_page_io_finish(m); 3198 vm_object_pip_subtract(obj, 1); 3199 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3200 iosize -= resid; 3201 } 3202 vm_page_unlock_queues(); 3203 if (obj != NULL) { 3204 vm_object_pip_wakeupn(obj, 0); 3205 VM_OBJECT_UNLOCK(obj); 3206 } 3207 } 3208 3209 /* 3210 * For asynchronous completions, release the buffer now. The brelse 3211 * will do a wakeup there if necessary - so no need to do a wakeup 3212 * here in the async case. The sync case always needs to do a wakeup. 3213 */ 3214 3215 if (bp->b_flags & B_ASYNC) { 3216 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3217 brelse(bp); 3218 else 3219 bqrelse(bp); 3220 } else { 3221 bdone(bp); 3222 } 3223 splx(s); 3224} 3225 3226/* 3227 * This routine is called in lieu of iodone in the case of 3228 * incomplete I/O. This keeps the busy status for pages 3229 * consistant. 3230 */ 3231void 3232vfs_unbusy_pages(struct buf * bp) 3233{ 3234 int i; 3235 3236 GIANT_REQUIRED; 3237 3238 runningbufwakeup(bp); 3239 if (bp->b_flags & B_VMIO) { 3240 vm_object_t obj; 3241 3242 obj = bp->b_object; 3243 VM_OBJECT_LOCK(obj); 3244 vm_page_lock_queues(); 3245 for (i = 0; i < bp->b_npages; i++) { 3246 vm_page_t m = bp->b_pages[i]; 3247 3248 if (m == bogus_page) { 3249 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3250 if (!m) { 3251 panic("vfs_unbusy_pages: page missing\n"); 3252 } 3253 bp->b_pages[i] = m; 3254 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3255 } 3256 vm_object_pip_subtract(obj, 1); 3257 vm_page_flag_clear(m, PG_ZERO); 3258 vm_page_io_finish(m); 3259 } 3260 vm_page_unlock_queues(); 3261 vm_object_pip_wakeupn(obj, 0); 3262 VM_OBJECT_UNLOCK(obj); 3263 } 3264} 3265 3266/* 3267 * vfs_page_set_valid: 3268 * 3269 * Set the valid bits in a page based on the supplied offset. The 3270 * range is restricted to the buffer's size. 3271 * 3272 * This routine is typically called after a read completes. 3273 */ 3274static void 3275vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3276{ 3277 vm_ooffset_t soff, eoff; 3278 3279 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3280 /* 3281 * Start and end offsets in buffer. eoff - soff may not cross a 3282 * page boundry or cross the end of the buffer. The end of the 3283 * buffer, in this case, is our file EOF, not the allocation size 3284 * of the buffer. 3285 */ 3286 soff = off; 3287 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3288 if (eoff > bp->b_offset + bp->b_bcount) 3289 eoff = bp->b_offset + bp->b_bcount; 3290 3291 /* 3292 * Set valid range. This is typically the entire buffer and thus the 3293 * entire page. 3294 */ 3295 if (eoff > soff) { 3296 vm_page_set_validclean( 3297 m, 3298 (vm_offset_t) (soff & PAGE_MASK), 3299 (vm_offset_t) (eoff - soff) 3300 ); 3301 } 3302} 3303 3304/* 3305 * This routine is called before a device strategy routine. 3306 * It is used to tell the VM system that paging I/O is in 3307 * progress, and treat the pages associated with the buffer 3308 * almost as being PG_BUSY. Also the object paging_in_progress 3309 * flag is handled to make sure that the object doesn't become 3310 * inconsistant. 3311 * 3312 * Since I/O has not been initiated yet, certain buffer flags 3313 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3314 * and should be ignored. 3315 */ 3316void 3317vfs_busy_pages(struct buf * bp, int clear_modify) 3318{ 3319 int i, bogus; 3320 3321 if (bp->b_flags & B_VMIO) { 3322 vm_object_t obj; 3323 vm_ooffset_t foff; 3324 3325 obj = bp->b_object; 3326 foff = bp->b_offset; 3327 KASSERT(bp->b_offset != NOOFFSET, 3328 ("vfs_busy_pages: no buffer offset")); 3329 vfs_setdirty(bp); 3330 if (obj != NULL) 3331 VM_OBJECT_LOCK(obj); 3332retry: 3333 vm_page_lock_queues(); 3334 for (i = 0; i < bp->b_npages; i++) { 3335 vm_page_t m = bp->b_pages[i]; 3336 3337 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3338 goto retry; 3339 } 3340 bogus = 0; 3341 for (i = 0; i < bp->b_npages; i++) { 3342 vm_page_t m = bp->b_pages[i]; 3343 3344 vm_page_flag_clear(m, PG_ZERO); 3345 if ((bp->b_flags & B_CLUSTER) == 0) { 3346 vm_object_pip_add(obj, 1); 3347 vm_page_io_start(m); 3348 } 3349 /* 3350 * When readying a buffer for a read ( i.e 3351 * clear_modify == 0 ), it is important to do 3352 * bogus_page replacement for valid pages in 3353 * partially instantiated buffers. Partially 3354 * instantiated buffers can, in turn, occur when 3355 * reconstituting a buffer from its VM backing store 3356 * base. We only have to do this if B_CACHE is 3357 * clear ( which causes the I/O to occur in the 3358 * first place ). The replacement prevents the read 3359 * I/O from overwriting potentially dirty VM-backed 3360 * pages. XXX bogus page replacement is, uh, bogus. 3361 * It may not work properly with small-block devices. 3362 * We need to find a better way. 3363 */ 3364 pmap_remove_all(m); 3365 if (clear_modify) 3366 vfs_page_set_valid(bp, foff, i, m); 3367 else if (m->valid == VM_PAGE_BITS_ALL && 3368 (bp->b_flags & B_CACHE) == 0) { 3369 bp->b_pages[i] = bogus_page; 3370 bogus++; 3371 } 3372 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3373 } 3374 vm_page_unlock_queues(); 3375 if (obj != NULL) 3376 VM_OBJECT_UNLOCK(obj); 3377 if (bogus) 3378 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3379 } 3380} 3381 3382/* 3383 * Tell the VM system that the pages associated with this buffer 3384 * are clean. This is used for delayed writes where the data is 3385 * going to go to disk eventually without additional VM intevention. 3386 * 3387 * Note that while we only really need to clean through to b_bcount, we 3388 * just go ahead and clean through to b_bufsize. 3389 */ 3390static void 3391vfs_clean_pages(struct buf * bp) 3392{ 3393 int i; 3394 3395 GIANT_REQUIRED; 3396 3397 if (bp->b_flags & B_VMIO) { 3398 vm_ooffset_t foff; 3399 3400 foff = bp->b_offset; 3401 KASSERT(bp->b_offset != NOOFFSET, 3402 ("vfs_clean_pages: no buffer offset")); 3403 vm_page_lock_queues(); 3404 for (i = 0; i < bp->b_npages; i++) { 3405 vm_page_t m = bp->b_pages[i]; 3406 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3407 vm_ooffset_t eoff = noff; 3408 3409 if (eoff > bp->b_offset + bp->b_bufsize) 3410 eoff = bp->b_offset + bp->b_bufsize; 3411 vfs_page_set_valid(bp, foff, i, m); 3412 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3413 foff = noff; 3414 } 3415 vm_page_unlock_queues(); 3416 } 3417} 3418 3419/* 3420 * vfs_bio_set_validclean: 3421 * 3422 * Set the range within the buffer to valid and clean. The range is 3423 * relative to the beginning of the buffer, b_offset. Note that b_offset 3424 * itself may be offset from the beginning of the first page. 3425 * 3426 */ 3427 3428void 3429vfs_bio_set_validclean(struct buf *bp, int base, int size) 3430{ 3431 if (bp->b_flags & B_VMIO) { 3432 int i; 3433 int n; 3434 3435 /* 3436 * Fixup base to be relative to beginning of first page. 3437 * Set initial n to be the maximum number of bytes in the 3438 * first page that can be validated. 3439 */ 3440 3441 base += (bp->b_offset & PAGE_MASK); 3442 n = PAGE_SIZE - (base & PAGE_MASK); 3443 3444 vm_page_lock_queues(); 3445 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3446 vm_page_t m = bp->b_pages[i]; 3447 3448 if (n > size) 3449 n = size; 3450 3451 vm_page_set_validclean(m, base & PAGE_MASK, n); 3452 base += n; 3453 size -= n; 3454 n = PAGE_SIZE; 3455 } 3456 vm_page_unlock_queues(); 3457 } 3458} 3459 3460/* 3461 * vfs_bio_clrbuf: 3462 * 3463 * clear a buffer. This routine essentially fakes an I/O, so we need 3464 * to clear BIO_ERROR and B_INVAL. 3465 * 3466 * Note that while we only theoretically need to clear through b_bcount, 3467 * we go ahead and clear through b_bufsize. 3468 */ 3469 3470void 3471vfs_bio_clrbuf(struct buf *bp) 3472{ 3473 int i, mask = 0; 3474 caddr_t sa, ea; 3475 3476 GIANT_REQUIRED; 3477 3478 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3479 bp->b_flags &= ~B_INVAL; 3480 bp->b_ioflags &= ~BIO_ERROR; 3481 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3482 (bp->b_offset & PAGE_MASK) == 0) { 3483 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3484 if ((bp->b_pages[0]->valid & mask) == mask) { 3485 bp->b_resid = 0; 3486 return; 3487 } 3488 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3489 ((bp->b_pages[0]->valid & mask) == 0)) { 3490 bzero(bp->b_data, bp->b_bufsize); 3491 bp->b_pages[0]->valid |= mask; 3492 bp->b_resid = 0; 3493 return; 3494 } 3495 } 3496 ea = sa = bp->b_data; 3497 for(i=0;i<bp->b_npages;i++,sa=ea) { 3498 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3499 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3500 ea = (caddr_t)(vm_offset_t)ulmin( 3501 (u_long)(vm_offset_t)ea, 3502 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3503 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3504 if ((bp->b_pages[i]->valid & mask) == mask) 3505 continue; 3506 if ((bp->b_pages[i]->valid & mask) == 0) { 3507 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3508 bzero(sa, ea - sa); 3509 } 3510 } else { 3511 for (; sa < ea; sa += DEV_BSIZE, j++) { 3512 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3513 (bp->b_pages[i]->valid & (1<<j)) == 0) 3514 bzero(sa, DEV_BSIZE); 3515 } 3516 } 3517 bp->b_pages[i]->valid |= mask; 3518 vm_page_lock_queues(); 3519 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3520 vm_page_unlock_queues(); 3521 } 3522 bp->b_resid = 0; 3523 } else { 3524 clrbuf(bp); 3525 } 3526} 3527 3528/* 3529 * vm_hold_load_pages and vm_hold_free_pages get pages into 3530 * a buffers address space. The pages are anonymous and are 3531 * not associated with a file object. 3532 */ 3533static void 3534vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3535{ 3536 vm_offset_t pg; 3537 vm_page_t p; 3538 int index; 3539 3540 GIANT_REQUIRED; 3541 3542 to = round_page(to); 3543 from = round_page(from); 3544 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3545 3546 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3547tryagain: 3548 /* 3549 * note: must allocate system pages since blocking here 3550 * could intefere with paging I/O, no matter which 3551 * process we are. 3552 */ 3553 VM_OBJECT_LOCK(kernel_object); 3554 p = vm_page_alloc(kernel_object, 3555 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3556 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3557 VM_OBJECT_UNLOCK(kernel_object); 3558 if (!p) { 3559 atomic_add_int(&vm_pageout_deficit, 3560 (to - pg) >> PAGE_SHIFT); 3561 VM_WAIT; 3562 goto tryagain; 3563 } 3564 vm_page_lock_queues(); 3565 p->valid = VM_PAGE_BITS_ALL; 3566 vm_page_unlock_queues(); 3567 pmap_qenter(pg, &p, 1); 3568 bp->b_pages[index] = p; 3569 vm_page_lock_queues(); 3570 vm_page_wakeup(p); 3571 vm_page_unlock_queues(); 3572 } 3573 bp->b_npages = index; 3574} 3575 3576/* Return pages associated with this buf to the vm system */ 3577static void 3578vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3579{ 3580 vm_offset_t pg; 3581 vm_page_t p; 3582 int index, newnpages; 3583 3584 GIANT_REQUIRED; 3585 3586 from = round_page(from); 3587 to = round_page(to); 3588 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3589 3590 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3591 p = bp->b_pages[index]; 3592 if (p && (index < bp->b_npages)) { 3593 if (p->busy) { 3594 printf( 3595 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3596 (intmax_t)bp->b_blkno, 3597 (intmax_t)bp->b_lblkno); 3598 } 3599 bp->b_pages[index] = NULL; 3600 pmap_qremove(pg, 1); 3601 vm_page_lock_queues(); 3602 vm_page_busy(p); 3603 vm_page_unwire(p, 0); 3604 vm_page_free(p); 3605 vm_page_unlock_queues(); 3606 } 3607 } 3608 bp->b_npages = newnpages; 3609} 3610 3611/* 3612 * Map an IO request into kernel virtual address space. 3613 * 3614 * All requests are (re)mapped into kernel VA space. 3615 * Notice that we use b_bufsize for the size of the buffer 3616 * to be mapped. b_bcount might be modified by the driver. 3617 * 3618 * Note that even if the caller determines that the address space should 3619 * be valid, a race or a smaller-file mapped into a larger space may 3620 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3621 * check the return value. 3622 */ 3623int 3624vmapbuf(struct buf *bp) 3625{ 3626 caddr_t addr, kva; 3627 vm_paddr_t pa; 3628 int pidx, i; 3629 struct vm_page *m; 3630 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3631 3632 GIANT_REQUIRED; 3633 3634 if ((bp->b_flags & B_PHYS) == 0) 3635 panic("vmapbuf"); 3636 if (bp->b_bufsize < 0) 3637 return (-1); 3638 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3639 addr < bp->b_data + bp->b_bufsize; 3640 addr += PAGE_SIZE, pidx++) { 3641 /* 3642 * Do the vm_fault if needed; do the copy-on-write thing 3643 * when reading stuff off device into memory. 3644 * 3645 * NOTE! Must use pmap_extract() because addr may be in 3646 * the userland address space, and kextract is only guarenteed 3647 * to work for the kernland address space (see: sparc64 port). 3648 */ 3649retry: 3650 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data, 3651 (bp->b_iocmd == BIO_READ) ? 3652 (VM_PROT_READ|VM_PROT_WRITE) : VM_PROT_READ); 3653 if (i < 0) { 3654 vm_page_lock_queues(); 3655 for (i = 0; i < pidx; ++i) { 3656 vm_page_unhold(bp->b_pages[i]); 3657 bp->b_pages[i] = NULL; 3658 } 3659 vm_page_unlock_queues(); 3660 return(-1); 3661 } 3662 pa = pmap_extract(pmap, (vm_offset_t)addr); 3663 if (pa == 0) { 3664 printf("vmapbuf: warning, race against user address during I/O"); 3665 goto retry; 3666 } 3667 m = PHYS_TO_VM_PAGE(pa); 3668 vm_page_lock_queues(); 3669 vm_page_hold(m); 3670 vm_page_unlock_queues(); 3671 bp->b_pages[pidx] = m; 3672 } 3673 if (pidx > btoc(MAXPHYS)) 3674 panic("vmapbuf: mapped more than MAXPHYS"); 3675 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3676 3677 kva = bp->b_saveaddr; 3678 bp->b_npages = pidx; 3679 bp->b_saveaddr = bp->b_data; 3680 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3681 return(0); 3682} 3683 3684/* 3685 * Free the io map PTEs associated with this IO operation. 3686 * We also invalidate the TLB entries and restore the original b_addr. 3687 */ 3688void 3689vunmapbuf(struct buf *bp) 3690{ 3691 int pidx; 3692 int npages; 3693 3694 GIANT_REQUIRED; 3695 3696 if ((bp->b_flags & B_PHYS) == 0) 3697 panic("vunmapbuf"); 3698 3699 npages = bp->b_npages; 3700 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3701 npages); 3702 vm_page_lock_queues(); 3703 for (pidx = 0; pidx < npages; pidx++) 3704 vm_page_unhold(bp->b_pages[pidx]); 3705 vm_page_unlock_queues(); 3706 3707 bp->b_data = bp->b_saveaddr; 3708} 3709 3710void 3711bdone(struct buf *bp) 3712{ 3713 mtx_lock(&bdonelock); 3714 bp->b_flags |= B_DONE; 3715 wakeup(bp); 3716 mtx_unlock(&bdonelock); 3717} 3718 3719void 3720bwait(struct buf *bp, u_char pri, const char *wchan) 3721{ 3722 mtx_lock(&bdonelock); 3723 while ((bp->b_flags & B_DONE) == 0) 3724 msleep(bp, &bdonelock, pri, wchan, 0); 3725 mtx_unlock(&bdonelock); 3726} 3727 3728#include "opt_ddb.h" 3729#ifdef DDB 3730#include <ddb/ddb.h> 3731 3732/* DDB command to show buffer data */ 3733DB_SHOW_COMMAND(buffer, db_show_buffer) 3734{ 3735 /* get args */ 3736 struct buf *bp = (struct buf *)addr; 3737 3738 if (!have_addr) { 3739 db_printf("usage: show buffer <addr>\n"); 3740 return; 3741 } 3742 3743 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3744 db_printf( 3745 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3746 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd, b_pblkno = %jd\n", 3747 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3748 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3749 (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno); 3750 if (bp->b_npages) { 3751 int i; 3752 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3753 for (i = 0; i < bp->b_npages; i++) { 3754 vm_page_t m; 3755 m = bp->b_pages[i]; 3756 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3757 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3758 if ((i + 1) < bp->b_npages) 3759 db_printf(","); 3760 } 3761 db_printf("\n"); 3762 } 3763} 3764#endif /* DDB */ 3765