vfs_bio.c revision 119370
155714Skris/* 255714Skris * Copyright (c) 1994,1997 John S. Dyson 355714Skris * All rights reserved. 455714Skris * 555714Skris * Redistribution and use in source and binary forms, with or without 655714Skris * modification, are permitted provided that the following conditions 755714Skris * are met: 8280297Sjkim * 1. Redistributions of source code must retain the above copyright 955714Skris * notice immediately at the beginning of the file, without modification, 1055714Skris * this list of conditions, and the following disclaimer. 1155714Skris * 2. Absolutely no warranty of function or purpose is made by the author 1255714Skris * John S. Dyson. 1355714Skris */ 1455714Skris 15280297Sjkim/* 1655714Skris * this file contains a new buffer I/O scheme implementing a coherent 1755714Skris * VM object and buffer cache scheme. Pains have been taken to make 1855714Skris * sure that the performance degradation associated with schemes such 1955714Skris * as this is not realized. 2055714Skris * 2155714Skris * Author: John S. Dyson 22280297Sjkim * Significant help during the development and debugging phases 2355714Skris * had been provided by David Greenman, also of the FreeBSD core team. 2455714Skris * 2555714Skris * see man buf(9) for more info. 2655714Skris */ 2755714Skris 2855714Skris#include <sys/cdefs.h> 2955714Skris__FBSDID("$FreeBSD: head/sys/kern/vfs_bio.c 119370 2003-08-23 18:11:53Z alc $"); 3055714Skris 3155714Skris#include <sys/param.h> 3255714Skris#include <sys/systm.h> 3355714Skris#include <sys/bio.h> 3455714Skris#include <sys/buf.h> 3555714Skris#include <sys/devicestat.h> 3655714Skris#include <sys/eventhandler.h> 37280297Sjkim#include <sys/lock.h> 3855714Skris#include <sys/malloc.h> 3955714Skris#include <sys/mount.h> 40280297Sjkim#include <sys/mutex.h> 4155714Skris#include <sys/kernel.h> 4255714Skris#include <sys/kthread.h> 4355714Skris#include <sys/proc.h> 4455714Skris#include <sys/resourcevar.h> 4555714Skris#include <sys/sysctl.h> 4655714Skris#include <sys/vmmeter.h> 4755714Skris#include <sys/vnode.h> 4855714Skris#include <vm/vm.h> 4955714Skris#include <vm/vm_param.h> 5055714Skris#include <vm/vm_kern.h> 5155714Skris#include <vm/vm_pageout.h> 52280297Sjkim#include <vm/vm_page.h> 5355714Skris#include <vm/vm_object.h> 5455714Skris#include <vm/vm_extern.h> 5555714Skris#include <vm/vm_map.h> 5655714Skris#include "opt_directio.h" 5755714Skris#include "opt_swap.h" 5855714Skris 5955714Skrisstatic MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 6055714Skris 6155714Skrisstruct bio_ops bioops; /* I/O operation notification */ 6255714Skris 63160814Ssimonstruct buf_ops buf_ops_bio = { 64280297Sjkim "buf_ops_bio", 65280297Sjkim bwrite 66280297Sjkim}; 67280297Sjkim 6855714Skris/* 69280297Sjkim * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 70280297Sjkim * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 71280297Sjkim */ 72280297Sjkimstruct buf *buf; /* buffer header pool */ 7355714Skris 74280297Sjkimstatic void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 75280297Sjkim vm_offset_t to); 7655714Skrisstatic void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 77280297Sjkim vm_offset_t to); 78280297Sjkimstatic void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 79280297Sjkim int pageno, vm_page_t m); 80280297Sjkimstatic void vfs_clean_pages(struct buf * bp); 81280297Sjkimstatic void vfs_setdirty(struct buf *bp); 82280297Sjkimstatic void vfs_vmio_release(struct buf *bp); 83280297Sjkimstatic void vfs_backgroundwritedone(struct buf *bp); 84280297Sjkimstatic int vfs_bio_clcheck(struct vnode *vp, int size, 85280297Sjkim daddr_t lblkno, daddr_t blkno); 86280297Sjkimstatic int flushbufqueues(int flushdeps); 87280297Sjkimstatic void buf_daemon(void); 88280297Sjkimvoid bremfreel(struct buf * bp); 8955714Skris 90280297Sjkimint vmiodirenable = TRUE; 91280297SjkimSYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 92280297Sjkim "Use the VM system for directory writes"); 93280297Sjkimint runningbufspace; 9455714SkrisSYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 95280297Sjkim "Amount of presently outstanding async buffer io"); 96280297Sjkimstatic int bufspace; 97280297SjkimSYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 98280297Sjkim "KVA memory used for bufs"); 99280297Sjkimstatic int maxbufspace; 100280297SjkimSYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 101280297Sjkim "Maximum allowed value of bufspace (including buf_daemon)"); 10255714Skrisstatic int bufmallocspace; 10355714SkrisSYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 104280297Sjkim "Amount of malloced memory for buffers"); 105280297Sjkimstatic int maxbufmallocspace; 106280297SjkimSYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 10755714Skris "Maximum amount of malloced memory for buffers"); 108280297Sjkimstatic int lobufspace; 109280297SjkimSYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 110280297Sjkim "Minimum amount of buffers we want to have"); 111280297Sjkimstatic int hibufspace; 112280297SjkimSYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 113280297Sjkim "Maximum allowed value of bufspace (excluding buf_daemon)"); 114280297Sjkimstatic int bufreusecnt; 115280297SjkimSYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 116280297Sjkim "Number of times we have reused a buffer"); 117280297Sjkimstatic int buffreekvacnt; 118280297SjkimSYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 119280297Sjkim "Number of times we have freed the KVA space from some buffer"); 120280297Sjkimstatic int bufdefragcnt; 121280297SjkimSYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 122280297Sjkim "Number of times we have had to repeat buffer allocation to defragment"); 123280297Sjkimstatic int lorunningspace; 124280297SjkimSYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 125280297Sjkim "Minimum preferred space used for in-progress I/O"); 126280297Sjkimstatic int hirunningspace; 127280297SjkimSYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 128280297Sjkim "Maximum amount of space to use for in-progress I/O"); 129280297Sjkimstatic int dirtybufferflushes; 130SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 131 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 132static int altbufferflushes; 133SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 134 0, "Number of fsync flushes to limit dirty buffers"); 135static int recursiveflushes; 136SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 137 0, "Number of flushes skipped due to being recursive"); 138static int numdirtybuffers; 139SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 140 "Number of buffers that are dirty (has unwritten changes) at the moment"); 141static int lodirtybuffers; 142SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 143 "How many buffers we want to have free before bufdaemon can sleep"); 144static int hidirtybuffers; 145SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 146 "When the number of dirty buffers is considered severe"); 147static int dirtybufthresh; 148SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 149 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 150static int numfreebuffers; 151SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 152 "Number of free buffers"); 153static int lofreebuffers; 154SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 155 "XXX Unused"); 156static int hifreebuffers; 157SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 158 "XXX Complicatedly unused"); 159static int getnewbufcalls; 160SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 161 "Number of calls to getnewbuf"); 162static int getnewbufrestarts; 163SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 164 "Number of times getnewbuf has had to restart a buffer aquisition"); 165static int dobkgrdwrite = 1; 166SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 167 "Do background writes (honoring the BX_BKGRDWRITE flag)?"); 168 169/* 170 * Wakeup point for bufdaemon, as well as indicator of whether it is already 171 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 172 * is idling. 173 */ 174static int bd_request; 175 176/* 177 * This lock synchronizes access to bd_request. 178 */ 179static struct mtx bdlock; 180 181/* 182 * bogus page -- for I/O to/from partially complete buffers 183 * this is a temporary solution to the problem, but it is not 184 * really that bad. it would be better to split the buffer 185 * for input in the case of buffers partially already in memory, 186 * but the code is intricate enough already. 187 */ 188vm_page_t bogus_page; 189 190/* 191 * Synchronization (sleep/wakeup) variable for active buffer space requests. 192 * Set when wait starts, cleared prior to wakeup(). 193 * Used in runningbufwakeup() and waitrunningbufspace(). 194 */ 195static int runningbufreq; 196 197/* 198 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 199 * waitrunningbufspace(). 200 */ 201static struct mtx rbreqlock; 202 203/* 204 * Synchronization (sleep/wakeup) variable for buffer requests. 205 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 206 * by and/or. 207 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 208 * getnewbuf(), and getblk(). 209 */ 210static int needsbuffer; 211 212/* 213 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 214 */ 215static struct mtx nblock; 216 217/* 218 * Lock that protects against bwait()/bdone()/B_DONE races. 219 */ 220 221static struct mtx bdonelock; 222 223/* 224 * Definitions for the buffer free lists. 225 */ 226#define BUFFER_QUEUES 6 /* number of free buffer queues */ 227 228#define QUEUE_NONE 0 /* on no queue */ 229#define QUEUE_LOCKED 1 /* locked buffers */ 230#define QUEUE_CLEAN 2 /* non-B_DELWRI buffers */ 231#define QUEUE_DIRTY 3 /* B_DELWRI buffers */ 232#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 233#define QUEUE_EMPTY 5 /* empty buffer headers */ 234 235/* Queues for free buffers with various properties */ 236static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 237 238/* Lock for the bufqueues */ 239static struct mtx bqlock; 240 241/* 242 * Single global constant for BUF_WMESG, to avoid getting multiple references. 243 * buf_wmesg is referred from macros. 244 */ 245const char *buf_wmesg = BUF_WMESG; 246 247#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 248#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 249#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 250#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 251 252#ifdef DIRECTIO 253extern void ffs_rawread_setup(void); 254#endif /* DIRECTIO */ 255/* 256 * numdirtywakeup: 257 * 258 * If someone is blocked due to there being too many dirty buffers, 259 * and numdirtybuffers is now reasonable, wake them up. 260 */ 261 262static __inline void 263numdirtywakeup(int level) 264{ 265 if (numdirtybuffers <= level) { 266 mtx_lock(&nblock); 267 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 268 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 269 wakeup(&needsbuffer); 270 } 271 mtx_unlock(&nblock); 272 } 273} 274 275/* 276 * bufspacewakeup: 277 * 278 * Called when buffer space is potentially available for recovery. 279 * getnewbuf() will block on this flag when it is unable to free 280 * sufficient buffer space. Buffer space becomes recoverable when 281 * bp's get placed back in the queues. 282 */ 283 284static __inline void 285bufspacewakeup(void) 286{ 287 /* 288 * If someone is waiting for BUF space, wake them up. Even 289 * though we haven't freed the kva space yet, the waiting 290 * process will be able to now. 291 */ 292 mtx_lock(&nblock); 293 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 294 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 295 wakeup(&needsbuffer); 296 } 297 mtx_unlock(&nblock); 298} 299 300/* 301 * runningbufwakeup() - in-progress I/O accounting. 302 * 303 */ 304static __inline void 305runningbufwakeup(struct buf *bp) 306{ 307 if (bp->b_runningbufspace) { 308 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 309 bp->b_runningbufspace = 0; 310 mtx_lock(&rbreqlock); 311 if (runningbufreq && runningbufspace <= lorunningspace) { 312 runningbufreq = 0; 313 wakeup(&runningbufreq); 314 } 315 mtx_unlock(&rbreqlock); 316 } 317} 318 319/* 320 * bufcountwakeup: 321 * 322 * Called when a buffer has been added to one of the free queues to 323 * account for the buffer and to wakeup anyone waiting for free buffers. 324 * This typically occurs when large amounts of metadata are being handled 325 * by the buffer cache ( else buffer space runs out first, usually ). 326 */ 327 328static __inline void 329bufcountwakeup(void) 330{ 331 atomic_add_int(&numfreebuffers, 1); 332 mtx_lock(&nblock); 333 if (needsbuffer) { 334 needsbuffer &= ~VFS_BIO_NEED_ANY; 335 if (numfreebuffers >= hifreebuffers) 336 needsbuffer &= ~VFS_BIO_NEED_FREE; 337 wakeup(&needsbuffer); 338 } 339 mtx_unlock(&nblock); 340} 341 342/* 343 * waitrunningbufspace() 344 * 345 * runningbufspace is a measure of the amount of I/O currently 346 * running. This routine is used in async-write situations to 347 * prevent creating huge backups of pending writes to a device. 348 * Only asynchronous writes are governed by this function. 349 * 350 * Reads will adjust runningbufspace, but will not block based on it. 351 * The read load has a side effect of reducing the allowed write load. 352 * 353 * This does NOT turn an async write into a sync write. It waits 354 * for earlier writes to complete and generally returns before the 355 * caller's write has reached the device. 356 */ 357static __inline void 358waitrunningbufspace(void) 359{ 360 mtx_lock(&rbreqlock); 361 while (runningbufspace > hirunningspace) { 362 ++runningbufreq; 363 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 364 } 365 mtx_unlock(&rbreqlock); 366} 367 368 369/* 370 * vfs_buf_test_cache: 371 * 372 * Called when a buffer is extended. This function clears the B_CACHE 373 * bit if the newly extended portion of the buffer does not contain 374 * valid data. 375 */ 376static __inline__ 377void 378vfs_buf_test_cache(struct buf *bp, 379 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 380 vm_page_t m) 381{ 382 GIANT_REQUIRED; 383 384 if (bp->b_flags & B_CACHE) { 385 int base = (foff + off) & PAGE_MASK; 386 if (vm_page_is_valid(m, base, size) == 0) 387 bp->b_flags &= ~B_CACHE; 388 } 389} 390 391/* Wake up the buffer deamon if necessary */ 392static __inline__ 393void 394bd_wakeup(int dirtybuflevel) 395{ 396 mtx_lock(&bdlock); 397 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 398 bd_request = 1; 399 wakeup(&bd_request); 400 } 401 mtx_unlock(&bdlock); 402} 403 404/* 405 * bd_speedup - speedup the buffer cache flushing code 406 */ 407 408static __inline__ 409void 410bd_speedup(void) 411{ 412 bd_wakeup(1); 413} 414 415/* 416 * Calculating buffer cache scaling values and reserve space for buffer 417 * headers. This is called during low level kernel initialization and 418 * may be called more then once. We CANNOT write to the memory area 419 * being reserved at this time. 420 */ 421caddr_t 422kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 423{ 424 /* 425 * physmem_est is in pages. Convert it to kilobytes (assumes 426 * PAGE_SIZE is >= 1K) 427 */ 428 physmem_est = physmem_est * (PAGE_SIZE / 1024); 429 430 /* 431 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 432 * For the first 64MB of ram nominally allocate sufficient buffers to 433 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 434 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 435 * the buffer cache we limit the eventual kva reservation to 436 * maxbcache bytes. 437 * 438 * factor represents the 1/4 x ram conversion. 439 */ 440 if (nbuf == 0) { 441 int factor = 4 * BKVASIZE / 1024; 442 443 nbuf = 50; 444 if (physmem_est > 4096) 445 nbuf += min((physmem_est - 4096) / factor, 446 65536 / factor); 447 if (physmem_est > 65536) 448 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 449 450 if (maxbcache && nbuf > maxbcache / BKVASIZE) 451 nbuf = maxbcache / BKVASIZE; 452 } 453 454#if 0 455 /* 456 * Do not allow the buffer_map to be more then 1/2 the size of the 457 * kernel_map. 458 */ 459 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 460 (BKVASIZE * 2)) { 461 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 462 (BKVASIZE * 2); 463 printf("Warning: nbufs capped at %d\n", nbuf); 464 } 465#endif 466 467 /* 468 * swbufs are used as temporary holders for I/O, such as paging I/O. 469 * We have no less then 16 and no more then 256. 470 */ 471 nswbuf = max(min(nbuf/4, 256), 16); 472#ifdef NSWBUF_MIN 473 if (nswbuf < NSWBUF_MIN) 474 nswbuf = NSWBUF_MIN; 475#endif 476#ifdef DIRECTIO 477 ffs_rawread_setup(); 478#endif 479 480 /* 481 * Reserve space for the buffer cache buffers 482 */ 483 swbuf = (void *)v; 484 v = (caddr_t)(swbuf + nswbuf); 485 buf = (void *)v; 486 v = (caddr_t)(buf + nbuf); 487 488 return(v); 489} 490 491/* Initialize the buffer subsystem. Called before use of any buffers. */ 492void 493bufinit(void) 494{ 495 struct buf *bp; 496 int i; 497 498 GIANT_REQUIRED; 499 500 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 501 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 502 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 503 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 504 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 505 506 /* next, make a null set of free lists */ 507 for (i = 0; i < BUFFER_QUEUES; i++) 508 TAILQ_INIT(&bufqueues[i]); 509 510 /* finally, initialize each buffer header and stick on empty q */ 511 for (i = 0; i < nbuf; i++) { 512 bp = &buf[i]; 513 bzero(bp, sizeof *bp); 514 bp->b_flags = B_INVAL; /* we're just an empty header */ 515 bp->b_dev = NODEV; 516 bp->b_rcred = NOCRED; 517 bp->b_wcred = NOCRED; 518 bp->b_qindex = QUEUE_EMPTY; 519 bp->b_vflags = 0; 520 bp->b_xflags = 0; 521 LIST_INIT(&bp->b_dep); 522 BUF_LOCKINIT(bp); 523 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 524 } 525 526 /* 527 * maxbufspace is the absolute maximum amount of buffer space we are 528 * allowed to reserve in KVM and in real terms. The absolute maximum 529 * is nominally used by buf_daemon. hibufspace is the nominal maximum 530 * used by most other processes. The differential is required to 531 * ensure that buf_daemon is able to run when other processes might 532 * be blocked waiting for buffer space. 533 * 534 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 535 * this may result in KVM fragmentation which is not handled optimally 536 * by the system. 537 */ 538 maxbufspace = nbuf * BKVASIZE; 539 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 540 lobufspace = hibufspace - MAXBSIZE; 541 542 lorunningspace = 512 * 1024; 543 hirunningspace = 1024 * 1024; 544 545/* 546 * Limit the amount of malloc memory since it is wired permanently into 547 * the kernel space. Even though this is accounted for in the buffer 548 * allocation, we don't want the malloced region to grow uncontrolled. 549 * The malloc scheme improves memory utilization significantly on average 550 * (small) directories. 551 */ 552 maxbufmallocspace = hibufspace / 20; 553 554/* 555 * Reduce the chance of a deadlock occuring by limiting the number 556 * of delayed-write dirty buffers we allow to stack up. 557 */ 558 hidirtybuffers = nbuf / 4 + 20; 559 dirtybufthresh = hidirtybuffers * 9 / 10; 560 numdirtybuffers = 0; 561/* 562 * To support extreme low-memory systems, make sure hidirtybuffers cannot 563 * eat up all available buffer space. This occurs when our minimum cannot 564 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 565 * BKVASIZE'd (8K) buffers. 566 */ 567 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 568 hidirtybuffers >>= 1; 569 } 570 lodirtybuffers = hidirtybuffers / 2; 571 572/* 573 * Try to keep the number of free buffers in the specified range, 574 * and give special processes (e.g. like buf_daemon) access to an 575 * emergency reserve. 576 */ 577 lofreebuffers = nbuf / 18 + 5; 578 hifreebuffers = 2 * lofreebuffers; 579 numfreebuffers = nbuf; 580 581/* 582 * Maximum number of async ops initiated per buf_daemon loop. This is 583 * somewhat of a hack at the moment, we really need to limit ourselves 584 * based on the number of bytes of I/O in-transit that were initiated 585 * from buf_daemon. 586 */ 587 588 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 589 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 590} 591 592/* 593 * bfreekva() - free the kva allocation for a buffer. 594 * 595 * Must be called at splbio() or higher as this is the only locking for 596 * buffer_map. 597 * 598 * Since this call frees up buffer space, we call bufspacewakeup(). 599 */ 600static void 601bfreekva(struct buf * bp) 602{ 603 GIANT_REQUIRED; 604 605 if (bp->b_kvasize) { 606 atomic_add_int(&buffreekvacnt, 1); 607 atomic_subtract_int(&bufspace, bp->b_kvasize); 608 vm_map_delete(buffer_map, 609 (vm_offset_t) bp->b_kvabase, 610 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 611 ); 612 bp->b_kvasize = 0; 613 bufspacewakeup(); 614 } 615} 616 617/* 618 * bremfree: 619 * 620 * Remove the buffer from the appropriate free list. 621 */ 622void 623bremfree(struct buf * bp) 624{ 625 mtx_lock(&bqlock); 626 bremfreel(bp); 627 mtx_unlock(&bqlock); 628} 629 630void 631bremfreel(struct buf * bp) 632{ 633 int s = splbio(); 634 int old_qindex = bp->b_qindex; 635 636 GIANT_REQUIRED; 637 638 if (bp->b_qindex != QUEUE_NONE) { 639 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 640 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 641 bp->b_qindex = QUEUE_NONE; 642 } else { 643 if (BUF_REFCNT(bp) <= 1) 644 panic("bremfree: removing a buffer not on a queue"); 645 } 646 647 /* 648 * Fixup numfreebuffers count. If the buffer is invalid or not 649 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 650 * the buffer was free and we must decrement numfreebuffers. 651 */ 652 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 653 switch(old_qindex) { 654 case QUEUE_DIRTY: 655 case QUEUE_CLEAN: 656 case QUEUE_EMPTY: 657 case QUEUE_EMPTYKVA: 658 atomic_subtract_int(&numfreebuffers, 1); 659 break; 660 default: 661 break; 662 } 663 } 664 splx(s); 665} 666 667 668/* 669 * Get a buffer with the specified data. Look in the cache first. We 670 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 671 * is set, the buffer is valid and we do not have to do anything ( see 672 * getblk() ). This is really just a special case of breadn(). 673 */ 674int 675bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 676 struct buf ** bpp) 677{ 678 679 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 680} 681 682/* 683 * Operates like bread, but also starts asynchronous I/O on 684 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 685 * to initiating I/O . If B_CACHE is set, the buffer is valid 686 * and we do not have to do anything. 687 */ 688int 689breadn(struct vnode * vp, daddr_t blkno, int size, 690 daddr_t * rablkno, int *rabsize, 691 int cnt, struct ucred * cred, struct buf ** bpp) 692{ 693 struct buf *bp, *rabp; 694 int i; 695 int rv = 0, readwait = 0; 696 697 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 698 699 /* if not found in cache, do some I/O */ 700 if ((bp->b_flags & B_CACHE) == 0) { 701 if (curthread != PCPU_GET(idlethread)) 702 curthread->td_proc->p_stats->p_ru.ru_inblock++; 703 bp->b_iocmd = BIO_READ; 704 bp->b_flags &= ~B_INVAL; 705 bp->b_ioflags &= ~BIO_ERROR; 706 if (bp->b_rcred == NOCRED && cred != NOCRED) 707 bp->b_rcred = crhold(cred); 708 vfs_busy_pages(bp, 0); 709 if (vp->v_type == VCHR) 710 VOP_SPECSTRATEGY(vp, bp); 711 else 712 VOP_STRATEGY(vp, bp); 713 ++readwait; 714 } 715 716 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 717 if (inmem(vp, *rablkno)) 718 continue; 719 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 720 721 if ((rabp->b_flags & B_CACHE) == 0) { 722 if (curthread != PCPU_GET(idlethread)) 723 curthread->td_proc->p_stats->p_ru.ru_inblock++; 724 rabp->b_flags |= B_ASYNC; 725 rabp->b_flags &= ~B_INVAL; 726 rabp->b_ioflags &= ~BIO_ERROR; 727 rabp->b_iocmd = BIO_READ; 728 if (rabp->b_rcred == NOCRED && cred != NOCRED) 729 rabp->b_rcred = crhold(cred); 730 vfs_busy_pages(rabp, 0); 731 BUF_KERNPROC(rabp); 732 if (vp->v_type == VCHR) 733 VOP_SPECSTRATEGY(vp, rabp); 734 else 735 VOP_STRATEGY(vp, rabp); 736 } else { 737 brelse(rabp); 738 } 739 } 740 741 if (readwait) { 742 rv = bufwait(bp); 743 } 744 return (rv); 745} 746 747/* 748 * Write, release buffer on completion. (Done by iodone 749 * if async). Do not bother writing anything if the buffer 750 * is invalid. 751 * 752 * Note that we set B_CACHE here, indicating that buffer is 753 * fully valid and thus cacheable. This is true even of NFS 754 * now so we set it generally. This could be set either here 755 * or in biodone() since the I/O is synchronous. We put it 756 * here. 757 */ 758 759int 760bwrite(struct buf * bp) 761{ 762 int oldflags, s; 763 struct buf *newbp; 764 765 if (bp->b_flags & B_INVAL) { 766 brelse(bp); 767 return (0); 768 } 769 770 oldflags = bp->b_flags; 771 772 if (BUF_REFCNT(bp) == 0) 773 panic("bwrite: buffer is not busy???"); 774 s = splbio(); 775 /* 776 * If a background write is already in progress, delay 777 * writing this block if it is asynchronous. Otherwise 778 * wait for the background write to complete. 779 */ 780 if (bp->b_xflags & BX_BKGRDINPROG) { 781 if (bp->b_flags & B_ASYNC) { 782 splx(s); 783 bdwrite(bp); 784 return (0); 785 } 786 bp->b_xflags |= BX_BKGRDWAIT; 787 tsleep(&bp->b_xflags, PRIBIO, "bwrbg", 0); 788 if (bp->b_xflags & BX_BKGRDINPROG) 789 panic("bwrite: still writing"); 790 } 791 792 /* Mark the buffer clean */ 793 bundirty(bp); 794 795 /* 796 * If this buffer is marked for background writing and we 797 * do not have to wait for it, make a copy and write the 798 * copy so as to leave this buffer ready for further use. 799 * 800 * This optimization eats a lot of memory. If we have a page 801 * or buffer shortfall we can't do it. 802 */ 803 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 804 (bp->b_flags & B_ASYNC) && 805 !vm_page_count_severe() && 806 !buf_dirty_count_severe()) { 807 if (bp->b_iodone != NULL) { 808 printf("bp->b_iodone = %p\n", bp->b_iodone); 809 panic("bwrite: need chained iodone"); 810 } 811 812 /* get a new block */ 813 newbp = geteblk(bp->b_bufsize); 814 815 /* 816 * set it to be identical to the old block. We have to 817 * set b_lblkno and BKGRDMARKER before calling bgetvp() 818 * to avoid confusing the splay tree and gbincore(). 819 */ 820 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 821 newbp->b_lblkno = bp->b_lblkno; 822 newbp->b_xflags |= BX_BKGRDMARKER; 823 /* XXX The BX_ flags need to be protected as well */ 824 VI_LOCK(bp->b_vp); 825 bgetvp(bp->b_vp, newbp); 826 VI_UNLOCK(bp->b_vp); 827 newbp->b_blkno = bp->b_blkno; 828 newbp->b_offset = bp->b_offset; 829 newbp->b_iodone = vfs_backgroundwritedone; 830 newbp->b_flags |= B_ASYNC; 831 newbp->b_flags &= ~B_INVAL; 832 833 /* move over the dependencies */ 834 if (LIST_FIRST(&bp->b_dep) != NULL) 835 buf_movedeps(bp, newbp); 836 837 /* 838 * Initiate write on the copy, release the original to 839 * the B_LOCKED queue so that it cannot go away until 840 * the background write completes. If not locked it could go 841 * away and then be reconstituted while it was being written. 842 * If the reconstituted buffer were written, we could end up 843 * with two background copies being written at the same time. 844 */ 845 bp->b_xflags |= BX_BKGRDINPROG; 846 bp->b_flags |= B_LOCKED; 847 bqrelse(bp); 848 bp = newbp; 849 } 850 851 bp->b_flags &= ~B_DONE; 852 bp->b_ioflags &= ~BIO_ERROR; 853 bp->b_flags |= B_WRITEINPROG | B_CACHE; 854 bp->b_iocmd = BIO_WRITE; 855 856 VI_LOCK(bp->b_vp); 857 bp->b_vp->v_numoutput++; 858 VI_UNLOCK(bp->b_vp); 859 vfs_busy_pages(bp, 1); 860 861 /* 862 * Normal bwrites pipeline writes 863 */ 864 bp->b_runningbufspace = bp->b_bufsize; 865 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 866 867 if (curthread != PCPU_GET(idlethread)) 868 curthread->td_proc->p_stats->p_ru.ru_oublock++; 869 splx(s); 870 if (oldflags & B_ASYNC) 871 BUF_KERNPROC(bp); 872 if (bp->b_vp->v_type == VCHR) 873 VOP_SPECSTRATEGY(bp->b_vp, bp); 874 else 875 VOP_STRATEGY(bp->b_vp, bp); 876 877 if ((oldflags & B_ASYNC) == 0) { 878 int rtval = bufwait(bp); 879 brelse(bp); 880 return (rtval); 881 } else { 882 /* 883 * don't allow the async write to saturate the I/O 884 * system. We will not deadlock here because 885 * we are blocking waiting for I/O that is already in-progress 886 * to complete. 887 */ 888 waitrunningbufspace(); 889 } 890 891 return (0); 892} 893 894/* 895 * Complete a background write started from bwrite. 896 */ 897static void 898vfs_backgroundwritedone(bp) 899 struct buf *bp; 900{ 901 struct buf *origbp; 902 903 /* 904 * Find the original buffer that we are writing. 905 */ 906 VI_LOCK(bp->b_vp); 907 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 908 panic("backgroundwritedone: lost buffer"); 909 VI_UNLOCK(bp->b_vp); 910 /* 911 * Process dependencies then return any unfinished ones. 912 */ 913 if (LIST_FIRST(&bp->b_dep) != NULL) 914 buf_complete(bp); 915 if (LIST_FIRST(&bp->b_dep) != NULL) 916 buf_movedeps(bp, origbp); 917 918 /* XXX Find out if origbp can disappear or get inconsistent */ 919 /* 920 * Clear the BX_BKGRDINPROG flag in the original buffer 921 * and awaken it if it is waiting for the write to complete. 922 * If BX_BKGRDINPROG is not set in the original buffer it must 923 * have been released and re-instantiated - which is not legal. 924 */ 925 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), 926 ("backgroundwritedone: lost buffer2")); 927 origbp->b_xflags &= ~BX_BKGRDINPROG; 928 if (origbp->b_xflags & BX_BKGRDWAIT) { 929 origbp->b_xflags &= ~BX_BKGRDWAIT; 930 wakeup(&origbp->b_xflags); 931 } 932 /* 933 * Clear the B_LOCKED flag and remove it from the locked 934 * queue if it currently resides there. 935 */ 936 origbp->b_flags &= ~B_LOCKED; 937 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) { 938 bremfree(origbp); 939 bqrelse(origbp); 940 } 941 /* 942 * This buffer is marked B_NOCACHE, so when it is released 943 * by biodone, it will be tossed. We mark it with BIO_READ 944 * to avoid biodone doing a second vwakeup. 945 */ 946 bp->b_flags |= B_NOCACHE; 947 bp->b_iocmd = BIO_READ; 948 bp->b_flags &= ~(B_CACHE | B_DONE); 949 bp->b_iodone = 0; 950 bufdone(bp); 951} 952 953/* 954 * Delayed write. (Buffer is marked dirty). Do not bother writing 955 * anything if the buffer is marked invalid. 956 * 957 * Note that since the buffer must be completely valid, we can safely 958 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 959 * biodone() in order to prevent getblk from writing the buffer 960 * out synchronously. 961 */ 962void 963bdwrite(struct buf * bp) 964{ 965 struct thread *td = curthread; 966 struct vnode *vp; 967 struct buf *nbp; 968 969 GIANT_REQUIRED; 970 971 if (BUF_REFCNT(bp) == 0) 972 panic("bdwrite: buffer is not busy"); 973 974 if (bp->b_flags & B_INVAL) { 975 brelse(bp); 976 return; 977 } 978 979 /* 980 * If we have too many dirty buffers, don't create any more. 981 * If we are wildly over our limit, then force a complete 982 * cleanup. Otherwise, just keep the situation from getting 983 * out of control. Note that we have to avoid a recursive 984 * disaster and not try to clean up after our own cleanup! 985 */ 986 vp = bp->b_vp; 987 VI_LOCK(vp); 988 if (td->td_proc->p_flag & P_COWINPROGRESS) { 989 recursiveflushes++; 990 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) { 991 VI_UNLOCK(vp); 992 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 993 VI_LOCK(vp); 994 altbufferflushes++; 995 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) { 996 /* 997 * Try to find a buffer to flush. 998 */ 999 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) { 1000 if ((nbp->b_xflags & BX_BKGRDINPROG) || 1001 buf_countdeps(nbp, 0) || 1002 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1003 continue; 1004 if (bp == nbp) 1005 panic("bdwrite: found ourselves"); 1006 VI_UNLOCK(vp); 1007 if (nbp->b_flags & B_CLUSTEROK) { 1008 vfs_bio_awrite(nbp); 1009 } else { 1010 bremfree(nbp); 1011 bawrite(nbp); 1012 } 1013 VI_LOCK(vp); 1014 dirtybufferflushes++; 1015 break; 1016 } 1017 } 1018 VI_UNLOCK(vp); 1019 1020 bdirty(bp); 1021 /* 1022 * Set B_CACHE, indicating that the buffer is fully valid. This is 1023 * true even of NFS now. 1024 */ 1025 bp->b_flags |= B_CACHE; 1026 1027 /* 1028 * This bmap keeps the system from needing to do the bmap later, 1029 * perhaps when the system is attempting to do a sync. Since it 1030 * is likely that the indirect block -- or whatever other datastructure 1031 * that the filesystem needs is still in memory now, it is a good 1032 * thing to do this. Note also, that if the pageout daemon is 1033 * requesting a sync -- there might not be enough memory to do 1034 * the bmap then... So, this is important to do. 1035 */ 1036 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1037 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1038 } 1039 1040 /* 1041 * Set the *dirty* buffer range based upon the VM system dirty pages. 1042 */ 1043 vfs_setdirty(bp); 1044 1045 /* 1046 * We need to do this here to satisfy the vnode_pager and the 1047 * pageout daemon, so that it thinks that the pages have been 1048 * "cleaned". Note that since the pages are in a delayed write 1049 * buffer -- the VFS layer "will" see that the pages get written 1050 * out on the next sync, or perhaps the cluster will be completed. 1051 */ 1052 vfs_clean_pages(bp); 1053 bqrelse(bp); 1054 1055 /* 1056 * Wakeup the buffer flushing daemon if we have a lot of dirty 1057 * buffers (midpoint between our recovery point and our stall 1058 * point). 1059 */ 1060 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1061 1062 /* 1063 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1064 * due to the softdep code. 1065 */ 1066} 1067 1068/* 1069 * bdirty: 1070 * 1071 * Turn buffer into delayed write request. We must clear BIO_READ and 1072 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1073 * itself to properly update it in the dirty/clean lists. We mark it 1074 * B_DONE to ensure that any asynchronization of the buffer properly 1075 * clears B_DONE ( else a panic will occur later ). 1076 * 1077 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1078 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1079 * should only be called if the buffer is known-good. 1080 * 1081 * Since the buffer is not on a queue, we do not update the numfreebuffers 1082 * count. 1083 * 1084 * Must be called at splbio(). 1085 * The buffer must be on QUEUE_NONE. 1086 */ 1087void 1088bdirty(bp) 1089 struct buf *bp; 1090{ 1091 KASSERT(bp->b_qindex == QUEUE_NONE, 1092 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1093 bp->b_flags &= ~(B_RELBUF); 1094 bp->b_iocmd = BIO_WRITE; 1095 1096 if ((bp->b_flags & B_DELWRI) == 0) { 1097 bp->b_flags |= B_DONE | B_DELWRI; 1098 reassignbuf(bp, bp->b_vp); 1099 atomic_add_int(&numdirtybuffers, 1); 1100 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1101 } 1102} 1103 1104/* 1105 * bundirty: 1106 * 1107 * Clear B_DELWRI for buffer. 1108 * 1109 * Since the buffer is not on a queue, we do not update the numfreebuffers 1110 * count. 1111 * 1112 * Must be called at splbio(). 1113 * The buffer must be on QUEUE_NONE. 1114 */ 1115 1116void 1117bundirty(bp) 1118 struct buf *bp; 1119{ 1120 KASSERT(bp->b_qindex == QUEUE_NONE, 1121 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1122 1123 if (bp->b_flags & B_DELWRI) { 1124 bp->b_flags &= ~B_DELWRI; 1125 reassignbuf(bp, bp->b_vp); 1126 atomic_subtract_int(&numdirtybuffers, 1); 1127 numdirtywakeup(lodirtybuffers); 1128 } 1129 /* 1130 * Since it is now being written, we can clear its deferred write flag. 1131 */ 1132 bp->b_flags &= ~B_DEFERRED; 1133} 1134 1135/* 1136 * bawrite: 1137 * 1138 * Asynchronous write. Start output on a buffer, but do not wait for 1139 * it to complete. The buffer is released when the output completes. 1140 * 1141 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1142 * B_INVAL buffers. Not us. 1143 */ 1144void 1145bawrite(struct buf * bp) 1146{ 1147 bp->b_flags |= B_ASYNC; 1148 (void) BUF_WRITE(bp); 1149} 1150 1151/* 1152 * bwillwrite: 1153 * 1154 * Called prior to the locking of any vnodes when we are expecting to 1155 * write. We do not want to starve the buffer cache with too many 1156 * dirty buffers so we block here. By blocking prior to the locking 1157 * of any vnodes we attempt to avoid the situation where a locked vnode 1158 * prevents the various system daemons from flushing related buffers. 1159 */ 1160 1161void 1162bwillwrite(void) 1163{ 1164 if (numdirtybuffers >= hidirtybuffers) { 1165 int s; 1166 1167 mtx_lock(&Giant); 1168 s = splbio(); 1169 mtx_lock(&nblock); 1170 while (numdirtybuffers >= hidirtybuffers) { 1171 bd_wakeup(1); 1172 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1173 msleep(&needsbuffer, &nblock, 1174 (PRIBIO + 4), "flswai", 0); 1175 } 1176 splx(s); 1177 mtx_unlock(&nblock); 1178 mtx_unlock(&Giant); 1179 } 1180} 1181 1182/* 1183 * Return true if we have too many dirty buffers. 1184 */ 1185int 1186buf_dirty_count_severe(void) 1187{ 1188 return(numdirtybuffers >= hidirtybuffers); 1189} 1190 1191/* 1192 * brelse: 1193 * 1194 * Release a busy buffer and, if requested, free its resources. The 1195 * buffer will be stashed in the appropriate bufqueue[] allowing it 1196 * to be accessed later as a cache entity or reused for other purposes. 1197 */ 1198void 1199brelse(struct buf * bp) 1200{ 1201 int s; 1202 1203 GIANT_REQUIRED; 1204 1205 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1206 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1207 1208 s = splbio(); 1209 1210 if (bp->b_flags & B_LOCKED) 1211 bp->b_ioflags &= ~BIO_ERROR; 1212 1213 if (bp->b_iocmd == BIO_WRITE && 1214 (bp->b_ioflags & BIO_ERROR) && 1215 !(bp->b_flags & B_INVAL)) { 1216 /* 1217 * Failed write, redirty. Must clear BIO_ERROR to prevent 1218 * pages from being scrapped. If B_INVAL is set then 1219 * this case is not run and the next case is run to 1220 * destroy the buffer. B_INVAL can occur if the buffer 1221 * is outside the range supported by the underlying device. 1222 */ 1223 bp->b_ioflags &= ~BIO_ERROR; 1224 bdirty(bp); 1225 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1226 (bp->b_ioflags & BIO_ERROR) || 1227 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1228 /* 1229 * Either a failed I/O or we were asked to free or not 1230 * cache the buffer. 1231 */ 1232 bp->b_flags |= B_INVAL; 1233 if (LIST_FIRST(&bp->b_dep) != NULL) 1234 buf_deallocate(bp); 1235 if (bp->b_flags & B_DELWRI) { 1236 atomic_subtract_int(&numdirtybuffers, 1); 1237 numdirtywakeup(lodirtybuffers); 1238 } 1239 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1240 if ((bp->b_flags & B_VMIO) == 0) { 1241 if (bp->b_bufsize) 1242 allocbuf(bp, 0); 1243 if (bp->b_vp) 1244 brelvp(bp); 1245 } 1246 } 1247 1248 /* 1249 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1250 * is called with B_DELWRI set, the underlying pages may wind up 1251 * getting freed causing a previous write (bdwrite()) to get 'lost' 1252 * because pages associated with a B_DELWRI bp are marked clean. 1253 * 1254 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1255 * if B_DELWRI is set. 1256 * 1257 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1258 * on pages to return pages to the VM page queues. 1259 */ 1260 if (bp->b_flags & B_DELWRI) 1261 bp->b_flags &= ~B_RELBUF; 1262 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1263 bp->b_flags |= B_RELBUF; 1264 1265 /* 1266 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1267 * constituted, not even NFS buffers now. Two flags effect this. If 1268 * B_INVAL, the struct buf is invalidated but the VM object is kept 1269 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1270 * 1271 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1272 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1273 * buffer is also B_INVAL because it hits the re-dirtying code above. 1274 * 1275 * Normally we can do this whether a buffer is B_DELWRI or not. If 1276 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1277 * the commit state and we cannot afford to lose the buffer. If the 1278 * buffer has a background write in progress, we need to keep it 1279 * around to prevent it from being reconstituted and starting a second 1280 * background write. 1281 */ 1282 if ((bp->b_flags & B_VMIO) 1283 && !(bp->b_vp->v_mount != NULL && 1284 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1285 !vn_isdisk(bp->b_vp, NULL) && 1286 (bp->b_flags & B_DELWRI)) 1287 ) { 1288 1289 int i, j, resid; 1290 vm_page_t m; 1291 off_t foff; 1292 vm_pindex_t poff; 1293 vm_object_t obj; 1294 struct vnode *vp; 1295 1296 vp = bp->b_vp; 1297 obj = bp->b_object; 1298 1299 /* 1300 * Get the base offset and length of the buffer. Note that 1301 * in the VMIO case if the buffer block size is not 1302 * page-aligned then b_data pointer may not be page-aligned. 1303 * But our b_pages[] array *IS* page aligned. 1304 * 1305 * block sizes less then DEV_BSIZE (usually 512) are not 1306 * supported due to the page granularity bits (m->valid, 1307 * m->dirty, etc...). 1308 * 1309 * See man buf(9) for more information 1310 */ 1311 resid = bp->b_bufsize; 1312 foff = bp->b_offset; 1313 if (obj != NULL) 1314 VM_OBJECT_LOCK(obj); 1315 for (i = 0; i < bp->b_npages; i++) { 1316 int had_bogus = 0; 1317 1318 m = bp->b_pages[i]; 1319 vm_page_lock_queues(); 1320 vm_page_flag_clear(m, PG_ZERO); 1321 vm_page_unlock_queues(); 1322 1323 /* 1324 * If we hit a bogus page, fixup *all* the bogus pages 1325 * now. 1326 */ 1327 if (m == bogus_page) { 1328 poff = OFF_TO_IDX(bp->b_offset); 1329 had_bogus = 1; 1330 1331 for (j = i; j < bp->b_npages; j++) { 1332 vm_page_t mtmp; 1333 mtmp = bp->b_pages[j]; 1334 if (mtmp == bogus_page) { 1335 mtmp = vm_page_lookup(obj, poff + j); 1336 if (!mtmp) { 1337 panic("brelse: page missing\n"); 1338 } 1339 bp->b_pages[j] = mtmp; 1340 } 1341 } 1342 1343 if ((bp->b_flags & B_INVAL) == 0) { 1344 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1345 } 1346 m = bp->b_pages[i]; 1347 } 1348 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1349 int poffset = foff & PAGE_MASK; 1350 int presid = resid > (PAGE_SIZE - poffset) ? 1351 (PAGE_SIZE - poffset) : resid; 1352 1353 KASSERT(presid >= 0, ("brelse: extra page")); 1354 vm_page_lock_queues(); 1355 vm_page_set_invalid(m, poffset, presid); 1356 vm_page_unlock_queues(); 1357 if (had_bogus) 1358 printf("avoided corruption bug in bogus_page/brelse code\n"); 1359 } 1360 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1361 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1362 } 1363 if (obj != NULL) 1364 VM_OBJECT_UNLOCK(obj); 1365 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1366 vfs_vmio_release(bp); 1367 1368 } else if (bp->b_flags & B_VMIO) { 1369 1370 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1371 vfs_vmio_release(bp); 1372 } 1373 1374 } 1375 1376 if (bp->b_qindex != QUEUE_NONE) 1377 panic("brelse: free buffer onto another queue???"); 1378 if (BUF_REFCNT(bp) > 1) { 1379 /* do not release to free list */ 1380 BUF_UNLOCK(bp); 1381 splx(s); 1382 return; 1383 } 1384 1385 /* enqueue */ 1386 mtx_lock(&bqlock); 1387 1388 /* buffers with no memory */ 1389 if (bp->b_bufsize == 0) { 1390 bp->b_flags |= B_INVAL; 1391 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1392 if (bp->b_xflags & BX_BKGRDINPROG) 1393 panic("losing buffer 1"); 1394 if (bp->b_kvasize) { 1395 bp->b_qindex = QUEUE_EMPTYKVA; 1396 } else { 1397 bp->b_qindex = QUEUE_EMPTY; 1398 } 1399 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1400 bp->b_dev = NODEV; 1401 /* buffers with junk contents */ 1402 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1403 (bp->b_ioflags & BIO_ERROR)) { 1404 bp->b_flags |= B_INVAL; 1405 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1406 if (bp->b_xflags & BX_BKGRDINPROG) 1407 panic("losing buffer 2"); 1408 bp->b_qindex = QUEUE_CLEAN; 1409 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1410 bp->b_dev = NODEV; 1411 1412 /* buffers that are locked */ 1413 } else if (bp->b_flags & B_LOCKED) { 1414 bp->b_qindex = QUEUE_LOCKED; 1415 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1416 1417 /* remaining buffers */ 1418 } else { 1419 if (bp->b_flags & B_DELWRI) 1420 bp->b_qindex = QUEUE_DIRTY; 1421 else 1422 bp->b_qindex = QUEUE_CLEAN; 1423 if (bp->b_flags & B_AGE) 1424 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1425 else 1426 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1427 } 1428 mtx_unlock(&bqlock); 1429 1430 /* 1431 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1432 * placed the buffer on the correct queue. We must also disassociate 1433 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1434 * find it. 1435 */ 1436 if (bp->b_flags & B_INVAL) { 1437 if (bp->b_flags & B_DELWRI) 1438 bundirty(bp); 1439 if (bp->b_vp) 1440 brelvp(bp); 1441 } 1442 1443 /* 1444 * Fixup numfreebuffers count. The bp is on an appropriate queue 1445 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1446 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1447 * if B_INVAL is set ). 1448 */ 1449 1450 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1451 bufcountwakeup(); 1452 1453 /* 1454 * Something we can maybe free or reuse 1455 */ 1456 if (bp->b_bufsize || bp->b_kvasize) 1457 bufspacewakeup(); 1458 1459 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1460 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1461 panic("brelse: not dirty"); 1462 /* unlock */ 1463 BUF_UNLOCK(bp); 1464 splx(s); 1465} 1466 1467/* 1468 * Release a buffer back to the appropriate queue but do not try to free 1469 * it. The buffer is expected to be used again soon. 1470 * 1471 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1472 * biodone() to requeue an async I/O on completion. It is also used when 1473 * known good buffers need to be requeued but we think we may need the data 1474 * again soon. 1475 * 1476 * XXX we should be able to leave the B_RELBUF hint set on completion. 1477 */ 1478void 1479bqrelse(struct buf * bp) 1480{ 1481 int s; 1482 1483 s = splbio(); 1484 1485 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1486 1487 if (bp->b_qindex != QUEUE_NONE) 1488 panic("bqrelse: free buffer onto another queue???"); 1489 if (BUF_REFCNT(bp) > 1) { 1490 /* do not release to free list */ 1491 BUF_UNLOCK(bp); 1492 splx(s); 1493 return; 1494 } 1495 mtx_lock(&bqlock); 1496 if (bp->b_flags & B_LOCKED) { 1497 bp->b_ioflags &= ~BIO_ERROR; 1498 bp->b_qindex = QUEUE_LOCKED; 1499 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1500 /* buffers with stale but valid contents */ 1501 } else if (bp->b_flags & B_DELWRI) { 1502 bp->b_qindex = QUEUE_DIRTY; 1503 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1504 } else if (vm_page_count_severe()) { 1505 /* 1506 * We are too low on memory, we have to try to free the 1507 * buffer (most importantly: the wired pages making up its 1508 * backing store) *now*. 1509 */ 1510 mtx_unlock(&bqlock); 1511 splx(s); 1512 brelse(bp); 1513 return; 1514 } else { 1515 bp->b_qindex = QUEUE_CLEAN; 1516 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1517 } 1518 mtx_unlock(&bqlock); 1519 1520 if ((bp->b_flags & B_LOCKED) == 0 && 1521 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1522 bufcountwakeup(); 1523 } 1524 1525 /* 1526 * Something we can maybe free or reuse. 1527 */ 1528 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1529 bufspacewakeup(); 1530 1531 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1532 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1533 panic("bqrelse: not dirty"); 1534 /* unlock */ 1535 BUF_UNLOCK(bp); 1536 splx(s); 1537} 1538 1539/* Give pages used by the bp back to the VM system (where possible) */ 1540static void 1541vfs_vmio_release(bp) 1542 struct buf *bp; 1543{ 1544 int i; 1545 vm_page_t m; 1546 1547 GIANT_REQUIRED; 1548 if (bp->b_object != NULL) 1549 VM_OBJECT_LOCK(bp->b_object); 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 if (bp->b_object != NULL) 1588 VM_OBJECT_UNLOCK(bp->b_object); 1589 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1590 1591 if (bp->b_bufsize) { 1592 bufspacewakeup(); 1593 bp->b_bufsize = 0; 1594 } 1595 bp->b_npages = 0; 1596 bp->b_flags &= ~B_VMIO; 1597 if (bp->b_vp) 1598 brelvp(bp); 1599} 1600 1601/* 1602 * Check to see if a block at a particular lbn is available for a clustered 1603 * write. 1604 */ 1605static int 1606vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1607{ 1608 struct buf *bpa; 1609 int match; 1610 1611 match = 0; 1612 1613 /* If the buf isn't in core skip it */ 1614 if ((bpa = gbincore(vp, lblkno)) == NULL) 1615 return (0); 1616 1617 /* If the buf is busy we don't want to wait for it */ 1618 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1619 return (0); 1620 1621 /* Only cluster with valid clusterable delayed write buffers */ 1622 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1623 (B_DELWRI | B_CLUSTEROK)) 1624 goto done; 1625 1626 if (bpa->b_bufsize != size) 1627 goto done; 1628 1629 /* 1630 * Check to see if it is in the expected place on disk and that the 1631 * block has been mapped. 1632 */ 1633 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1634 match = 1; 1635done: 1636 BUF_UNLOCK(bpa); 1637 return (match); 1638} 1639 1640/* 1641 * vfs_bio_awrite: 1642 * 1643 * Implement clustered async writes for clearing out B_DELWRI buffers. 1644 * This is much better then the old way of writing only one buffer at 1645 * a time. Note that we may not be presented with the buffers in the 1646 * correct order, so we search for the cluster in both directions. 1647 */ 1648int 1649vfs_bio_awrite(struct buf * bp) 1650{ 1651 int i; 1652 int j; 1653 daddr_t lblkno = bp->b_lblkno; 1654 struct vnode *vp = bp->b_vp; 1655 int s; 1656 int ncl; 1657 int nwritten; 1658 int size; 1659 int maxcl; 1660 1661 s = splbio(); 1662 /* 1663 * right now we support clustered writing only to regular files. If 1664 * we find a clusterable block we could be in the middle of a cluster 1665 * rather then at the beginning. 1666 */ 1667 if ((vp->v_type == VREG) && 1668 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1669 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1670 1671 size = vp->v_mount->mnt_stat.f_iosize; 1672 maxcl = MAXPHYS / size; 1673 1674 VI_LOCK(vp); 1675 for (i = 1; i < maxcl; i++) 1676 if (vfs_bio_clcheck(vp, size, lblkno + i, 1677 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1678 break; 1679 1680 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1681 if (vfs_bio_clcheck(vp, size, lblkno - j, 1682 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1683 break; 1684 1685 VI_UNLOCK(vp); 1686 --j; 1687 ncl = i + j; 1688 /* 1689 * this is a possible cluster write 1690 */ 1691 if (ncl != 1) { 1692 BUF_UNLOCK(bp); 1693 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1694 splx(s); 1695 return nwritten; 1696 } 1697 } 1698 1699 bremfree(bp); 1700 bp->b_flags |= B_ASYNC; 1701 1702 splx(s); 1703 /* 1704 * default (old) behavior, writing out only one block 1705 * 1706 * XXX returns b_bufsize instead of b_bcount for nwritten? 1707 */ 1708 nwritten = bp->b_bufsize; 1709 (void) BUF_WRITE(bp); 1710 1711 return nwritten; 1712} 1713 1714/* 1715 * getnewbuf: 1716 * 1717 * Find and initialize a new buffer header, freeing up existing buffers 1718 * in the bufqueues as necessary. The new buffer is returned locked. 1719 * 1720 * Important: B_INVAL is not set. If the caller wishes to throw the 1721 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1722 * 1723 * We block if: 1724 * We have insufficient buffer headers 1725 * We have insufficient buffer space 1726 * buffer_map is too fragmented ( space reservation fails ) 1727 * If we have to flush dirty buffers ( but we try to avoid this ) 1728 * 1729 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1730 * Instead we ask the buf daemon to do it for us. We attempt to 1731 * avoid piecemeal wakeups of the pageout daemon. 1732 */ 1733 1734static struct buf * 1735getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1736{ 1737 struct buf *bp; 1738 struct buf *nbp; 1739 int defrag = 0; 1740 int nqindex; 1741 static int flushingbufs; 1742 1743 GIANT_REQUIRED; 1744 1745 /* 1746 * We can't afford to block since we might be holding a vnode lock, 1747 * which may prevent system daemons from running. We deal with 1748 * low-memory situations by proactively returning memory and running 1749 * async I/O rather then sync I/O. 1750 */ 1751 1752 atomic_add_int(&getnewbufcalls, 1); 1753 atomic_subtract_int(&getnewbufrestarts, 1); 1754restart: 1755 atomic_add_int(&getnewbufrestarts, 1); 1756 1757 /* 1758 * Setup for scan. If we do not have enough free buffers, 1759 * we setup a degenerate case that immediately fails. Note 1760 * that if we are specially marked process, we are allowed to 1761 * dip into our reserves. 1762 * 1763 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1764 * 1765 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1766 * However, there are a number of cases (defragging, reusing, ...) 1767 * where we cannot backup. 1768 */ 1769 mtx_lock(&bqlock); 1770 nqindex = QUEUE_EMPTYKVA; 1771 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1772 1773 if (nbp == NULL) { 1774 /* 1775 * If no EMPTYKVA buffers and we are either 1776 * defragging or reusing, locate a CLEAN buffer 1777 * to free or reuse. If bufspace useage is low 1778 * skip this step so we can allocate a new buffer. 1779 */ 1780 if (defrag || bufspace >= lobufspace) { 1781 nqindex = QUEUE_CLEAN; 1782 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1783 } 1784 1785 /* 1786 * If we could not find or were not allowed to reuse a 1787 * CLEAN buffer, check to see if it is ok to use an EMPTY 1788 * buffer. We can only use an EMPTY buffer if allocating 1789 * its KVA would not otherwise run us out of buffer space. 1790 */ 1791 if (nbp == NULL && defrag == 0 && 1792 bufspace + maxsize < hibufspace) { 1793 nqindex = QUEUE_EMPTY; 1794 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1795 } 1796 } 1797 1798 /* 1799 * Run scan, possibly freeing data and/or kva mappings on the fly 1800 * depending. 1801 */ 1802 1803 while ((bp = nbp) != NULL) { 1804 int qindex = nqindex; 1805 1806 /* 1807 * Calculate next bp ( we can only use it if we do not block 1808 * or do other fancy things ). 1809 */ 1810 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1811 switch(qindex) { 1812 case QUEUE_EMPTY: 1813 nqindex = QUEUE_EMPTYKVA; 1814 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1815 break; 1816 /* FALLTHROUGH */ 1817 case QUEUE_EMPTYKVA: 1818 nqindex = QUEUE_CLEAN; 1819 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1820 break; 1821 /* FALLTHROUGH */ 1822 case QUEUE_CLEAN: 1823 /* 1824 * nbp is NULL. 1825 */ 1826 break; 1827 } 1828 } 1829 1830 /* 1831 * Sanity Checks 1832 */ 1833 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1834 1835 /* 1836 * Note: we no longer distinguish between VMIO and non-VMIO 1837 * buffers. 1838 */ 1839 1840 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1841 1842 /* 1843 * If we are defragging then we need a buffer with 1844 * b_kvasize != 0. XXX this situation should no longer 1845 * occur, if defrag is non-zero the buffer's b_kvasize 1846 * should also be non-zero at this point. XXX 1847 */ 1848 if (defrag && bp->b_kvasize == 0) { 1849 printf("Warning: defrag empty buffer %p\n", bp); 1850 continue; 1851 } 1852 1853 /* 1854 * Start freeing the bp. This is somewhat involved. nbp 1855 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1856 */ 1857 1858 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1859 panic("getnewbuf: locked buf"); 1860 bremfreel(bp); 1861 mtx_unlock(&bqlock); 1862 1863 if (qindex == QUEUE_CLEAN) { 1864 if (bp->b_flags & B_VMIO) { 1865 bp->b_flags &= ~B_ASYNC; 1866 vfs_vmio_release(bp); 1867 } 1868 if (bp->b_vp) 1869 brelvp(bp); 1870 } 1871 1872 /* 1873 * NOTE: nbp is now entirely invalid. We can only restart 1874 * the scan from this point on. 1875 * 1876 * Get the rest of the buffer freed up. b_kva* is still 1877 * valid after this operation. 1878 */ 1879 1880 if (bp->b_rcred != NOCRED) { 1881 crfree(bp->b_rcred); 1882 bp->b_rcred = NOCRED; 1883 } 1884 if (bp->b_wcred != NOCRED) { 1885 crfree(bp->b_wcred); 1886 bp->b_wcred = NOCRED; 1887 } 1888 if (LIST_FIRST(&bp->b_dep) != NULL) 1889 buf_deallocate(bp); 1890 if (bp->b_xflags & BX_BKGRDINPROG) 1891 panic("losing buffer 3"); 1892 1893 if (bp->b_bufsize) 1894 allocbuf(bp, 0); 1895 1896 bp->b_flags = 0; 1897 bp->b_ioflags = 0; 1898 bp->b_xflags = 0; 1899 bp->b_vflags = 0; 1900 bp->b_dev = NODEV; 1901 bp->b_vp = NULL; 1902 bp->b_blkno = bp->b_lblkno = 0; 1903 bp->b_offset = NOOFFSET; 1904 bp->b_iodone = 0; 1905 bp->b_error = 0; 1906 bp->b_resid = 0; 1907 bp->b_bcount = 0; 1908 bp->b_npages = 0; 1909 bp->b_dirtyoff = bp->b_dirtyend = 0; 1910 bp->b_magic = B_MAGIC_BIO; 1911 bp->b_op = &buf_ops_bio; 1912 bp->b_object = NULL; 1913 1914 LIST_INIT(&bp->b_dep); 1915 1916 /* 1917 * If we are defragging then free the buffer. 1918 */ 1919 if (defrag) { 1920 bp->b_flags |= B_INVAL; 1921 bfreekva(bp); 1922 brelse(bp); 1923 defrag = 0; 1924 goto restart; 1925 } 1926 1927 /* 1928 * If we are overcomitted then recover the buffer and its 1929 * KVM space. This occurs in rare situations when multiple 1930 * processes are blocked in getnewbuf() or allocbuf(). 1931 */ 1932 if (bufspace >= hibufspace) 1933 flushingbufs = 1; 1934 if (flushingbufs && bp->b_kvasize != 0) { 1935 bp->b_flags |= B_INVAL; 1936 bfreekva(bp); 1937 brelse(bp); 1938 goto restart; 1939 } 1940 if (bufspace < lobufspace) 1941 flushingbufs = 0; 1942 break; 1943 } 1944 1945 /* 1946 * If we exhausted our list, sleep as appropriate. We may have to 1947 * wakeup various daemons and write out some dirty buffers. 1948 * 1949 * Generally we are sleeping due to insufficient buffer space. 1950 */ 1951 1952 if (bp == NULL) { 1953 int flags; 1954 char *waitmsg; 1955 1956 mtx_unlock(&bqlock); 1957 if (defrag) { 1958 flags = VFS_BIO_NEED_BUFSPACE; 1959 waitmsg = "nbufkv"; 1960 } else if (bufspace >= hibufspace) { 1961 waitmsg = "nbufbs"; 1962 flags = VFS_BIO_NEED_BUFSPACE; 1963 } else { 1964 waitmsg = "newbuf"; 1965 flags = VFS_BIO_NEED_ANY; 1966 } 1967 1968 bd_speedup(); /* heeeelp */ 1969 1970 mtx_lock(&nblock); 1971 needsbuffer |= flags; 1972 while (needsbuffer & flags) { 1973 if (msleep(&needsbuffer, &nblock, 1974 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1975 mtx_unlock(&nblock); 1976 return (NULL); 1977 } 1978 } 1979 mtx_unlock(&nblock); 1980 } else { 1981 /* 1982 * We finally have a valid bp. We aren't quite out of the 1983 * woods, we still have to reserve kva space. In order 1984 * to keep fragmentation sane we only allocate kva in 1985 * BKVASIZE chunks. 1986 */ 1987 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1988 1989 if (maxsize != bp->b_kvasize) { 1990 vm_offset_t addr = 0; 1991 1992 bfreekva(bp); 1993 1994 if (vm_map_findspace(buffer_map, 1995 vm_map_min(buffer_map), maxsize, &addr)) { 1996 /* 1997 * Uh oh. Buffer map is to fragmented. We 1998 * must defragment the map. 1999 */ 2000 atomic_add_int(&bufdefragcnt, 1); 2001 defrag = 1; 2002 bp->b_flags |= B_INVAL; 2003 brelse(bp); 2004 goto restart; 2005 } 2006 if (addr) { 2007 vm_map_insert(buffer_map, NULL, 0, 2008 addr, addr + maxsize, 2009 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2010 2011 bp->b_kvabase = (caddr_t) addr; 2012 bp->b_kvasize = maxsize; 2013 atomic_add_int(&bufspace, bp->b_kvasize); 2014 atomic_add_int(&bufreusecnt, 1); 2015 } 2016 } 2017 bp->b_saveaddr = bp->b_kvabase; 2018 bp->b_data = bp->b_saveaddr; 2019 } 2020 return(bp); 2021} 2022 2023/* 2024 * buf_daemon: 2025 * 2026 * buffer flushing daemon. Buffers are normally flushed by the 2027 * update daemon but if it cannot keep up this process starts to 2028 * take the load in an attempt to prevent getnewbuf() from blocking. 2029 */ 2030 2031static struct proc *bufdaemonproc; 2032 2033static struct kproc_desc buf_kp = { 2034 "bufdaemon", 2035 buf_daemon, 2036 &bufdaemonproc 2037}; 2038SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2039 2040static void 2041buf_daemon() 2042{ 2043 int s; 2044 2045 mtx_lock(&Giant); 2046 2047 /* 2048 * This process needs to be suspended prior to shutdown sync. 2049 */ 2050 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2051 SHUTDOWN_PRI_LAST); 2052 2053 /* 2054 * This process is allowed to take the buffer cache to the limit 2055 */ 2056 s = splbio(); 2057 mtx_lock(&bdlock); 2058 2059 for (;;) { 2060 bd_request = 0; 2061 mtx_unlock(&bdlock); 2062 2063 kthread_suspend_check(bufdaemonproc); 2064 2065 /* 2066 * Do the flush. Limit the amount of in-transit I/O we 2067 * allow to build up, otherwise we would completely saturate 2068 * the I/O system. Wakeup any waiting processes before we 2069 * normally would so they can run in parallel with our drain. 2070 */ 2071 while (numdirtybuffers > lodirtybuffers) { 2072 if (flushbufqueues(0) == 0) { 2073 /* 2074 * Could not find any buffers without rollback 2075 * dependencies, so just write the first one 2076 * in the hopes of eventually making progress. 2077 */ 2078 flushbufqueues(1); 2079 break; 2080 } 2081 waitrunningbufspace(); 2082 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2083 } 2084 2085 /* 2086 * Only clear bd_request if we have reached our low water 2087 * mark. The buf_daemon normally waits 1 second and 2088 * then incrementally flushes any dirty buffers that have 2089 * built up, within reason. 2090 * 2091 * If we were unable to hit our low water mark and couldn't 2092 * find any flushable buffers, we sleep half a second. 2093 * Otherwise we loop immediately. 2094 */ 2095 mtx_lock(&bdlock); 2096 if (numdirtybuffers <= lodirtybuffers) { 2097 /* 2098 * We reached our low water mark, reset the 2099 * request and sleep until we are needed again. 2100 * The sleep is just so the suspend code works. 2101 */ 2102 bd_request = 0; 2103 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2104 } else { 2105 /* 2106 * We couldn't find any flushable dirty buffers but 2107 * still have too many dirty buffers, we 2108 * have to sleep and try again. (rare) 2109 */ 2110 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2111 } 2112 } 2113} 2114 2115/* 2116 * flushbufqueues: 2117 * 2118 * Try to flush a buffer in the dirty queue. We must be careful to 2119 * free up B_INVAL buffers instead of write them, which NFS is 2120 * particularly sensitive to. 2121 */ 2122int flushwithdeps = 0; 2123SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2124 0, "Number of buffers flushed with dependecies that require rollbacks"); 2125static int 2126flushbufqueues(int flushdeps) 2127{ 2128 struct thread *td = curthread; 2129 struct vnode *vp; 2130 struct buf *bp; 2131 int hasdeps; 2132 2133 mtx_lock(&bqlock); 2134 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2135 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2136 continue; 2137 KASSERT((bp->b_flags & B_DELWRI), 2138 ("unexpected clean buffer %p", bp)); 2139 if ((bp->b_xflags & BX_BKGRDINPROG) != 0) { 2140 BUF_UNLOCK(bp); 2141 continue; 2142 } 2143 if (bp->b_flags & B_INVAL) { 2144 bremfreel(bp); 2145 mtx_unlock(&bqlock); 2146 brelse(bp); 2147 return (1); 2148 } 2149 2150 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2151 if (flushdeps == 0) { 2152 BUF_UNLOCK(bp); 2153 continue; 2154 } 2155 hasdeps = 1; 2156 } else 2157 hasdeps = 0; 2158 /* 2159 * We must hold the lock on a vnode before writing 2160 * one of its buffers. Otherwise we may confuse, or 2161 * in the case of a snapshot vnode, deadlock the 2162 * system. 2163 * 2164 * The lock order here is the reverse of the normal 2165 * of vnode followed by buf lock. This is ok because 2166 * the NOWAIT will prevent deadlock. 2167 */ 2168 if ((vp = bp->b_vp) == NULL || 2169 vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2170 mtx_unlock(&bqlock); 2171 vfs_bio_awrite(bp); 2172 if (vp != NULL) 2173 VOP_UNLOCK(vp, 0, td); 2174 flushwithdeps += hasdeps; 2175 return (1); 2176 } 2177 BUF_UNLOCK(bp); 2178 } 2179 mtx_unlock(&bqlock); 2180 return (0); 2181} 2182 2183/* 2184 * Check to see if a block is currently memory resident. 2185 */ 2186struct buf * 2187incore(struct vnode * vp, daddr_t blkno) 2188{ 2189 struct buf *bp; 2190 2191 int s = splbio(); 2192 VI_LOCK(vp); 2193 bp = gbincore(vp, blkno); 2194 VI_UNLOCK(vp); 2195 splx(s); 2196 return (bp); 2197} 2198 2199/* 2200 * Returns true if no I/O is needed to access the 2201 * associated VM object. This is like incore except 2202 * it also hunts around in the VM system for the data. 2203 */ 2204 2205int 2206inmem(struct vnode * vp, daddr_t blkno) 2207{ 2208 vm_object_t obj; 2209 vm_offset_t toff, tinc, size; 2210 vm_page_t m; 2211 vm_ooffset_t off; 2212 2213 GIANT_REQUIRED; 2214 ASSERT_VOP_LOCKED(vp, "inmem"); 2215 2216 if (incore(vp, blkno)) 2217 return 1; 2218 if (vp->v_mount == NULL) 2219 return 0; 2220 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2221 return 0; 2222 2223 size = PAGE_SIZE; 2224 if (size > vp->v_mount->mnt_stat.f_iosize) 2225 size = vp->v_mount->mnt_stat.f_iosize; 2226 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2227 2228 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2229 VM_OBJECT_LOCK(obj); 2230 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2231 VM_OBJECT_UNLOCK(obj); 2232 if (!m) 2233 goto notinmem; 2234 tinc = size; 2235 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2236 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2237 if (vm_page_is_valid(m, 2238 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2239 goto notinmem; 2240 } 2241 return 1; 2242 2243notinmem: 2244 return (0); 2245} 2246 2247/* 2248 * vfs_setdirty: 2249 * 2250 * Sets the dirty range for a buffer based on the status of the dirty 2251 * bits in the pages comprising the buffer. 2252 * 2253 * The range is limited to the size of the buffer. 2254 * 2255 * This routine is primarily used by NFS, but is generalized for the 2256 * B_VMIO case. 2257 */ 2258static void 2259vfs_setdirty(struct buf *bp) 2260{ 2261 int i; 2262 vm_object_t object; 2263 2264 GIANT_REQUIRED; 2265 /* 2266 * Degenerate case - empty buffer 2267 */ 2268 2269 if (bp->b_bufsize == 0) 2270 return; 2271 2272 /* 2273 * We qualify the scan for modified pages on whether the 2274 * object has been flushed yet. The OBJ_WRITEABLE flag 2275 * is not cleared simply by protecting pages off. 2276 */ 2277 2278 if ((bp->b_flags & B_VMIO) == 0) 2279 return; 2280 2281 object = bp->b_pages[0]->object; 2282 VM_OBJECT_LOCK(object); 2283 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2284 printf("Warning: object %p writeable but not mightbedirty\n", object); 2285 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2286 printf("Warning: object %p mightbedirty but not writeable\n", object); 2287 2288 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2289 vm_offset_t boffset; 2290 vm_offset_t eoffset; 2291 2292 vm_page_lock_queues(); 2293 /* 2294 * test the pages to see if they have been modified directly 2295 * by users through the VM system. 2296 */ 2297 for (i = 0; i < bp->b_npages; i++) { 2298 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2299 vm_page_test_dirty(bp->b_pages[i]); 2300 } 2301 2302 /* 2303 * Calculate the encompassing dirty range, boffset and eoffset, 2304 * (eoffset - boffset) bytes. 2305 */ 2306 2307 for (i = 0; i < bp->b_npages; i++) { 2308 if (bp->b_pages[i]->dirty) 2309 break; 2310 } 2311 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2312 2313 for (i = bp->b_npages - 1; i >= 0; --i) { 2314 if (bp->b_pages[i]->dirty) { 2315 break; 2316 } 2317 } 2318 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2319 2320 vm_page_unlock_queues(); 2321 /* 2322 * Fit it to the buffer. 2323 */ 2324 2325 if (eoffset > bp->b_bcount) 2326 eoffset = bp->b_bcount; 2327 2328 /* 2329 * If we have a good dirty range, merge with the existing 2330 * dirty range. 2331 */ 2332 2333 if (boffset < eoffset) { 2334 if (bp->b_dirtyoff > boffset) 2335 bp->b_dirtyoff = boffset; 2336 if (bp->b_dirtyend < eoffset) 2337 bp->b_dirtyend = eoffset; 2338 } 2339 } 2340 VM_OBJECT_UNLOCK(object); 2341} 2342 2343/* 2344 * getblk: 2345 * 2346 * Get a block given a specified block and offset into a file/device. 2347 * The buffers B_DONE bit will be cleared on return, making it almost 2348 * ready for an I/O initiation. B_INVAL may or may not be set on 2349 * return. The caller should clear B_INVAL prior to initiating a 2350 * READ. 2351 * 2352 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2353 * an existing buffer. 2354 * 2355 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2356 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2357 * and then cleared based on the backing VM. If the previous buffer is 2358 * non-0-sized but invalid, B_CACHE will be cleared. 2359 * 2360 * If getblk() must create a new buffer, the new buffer is returned with 2361 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2362 * case it is returned with B_INVAL clear and B_CACHE set based on the 2363 * backing VM. 2364 * 2365 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2366 * B_CACHE bit is clear. 2367 * 2368 * What this means, basically, is that the caller should use B_CACHE to 2369 * determine whether the buffer is fully valid or not and should clear 2370 * B_INVAL prior to issuing a read. If the caller intends to validate 2371 * the buffer by loading its data area with something, the caller needs 2372 * to clear B_INVAL. If the caller does this without issuing an I/O, 2373 * the caller should set B_CACHE ( as an optimization ), else the caller 2374 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2375 * a write attempt or if it was a successfull read. If the caller 2376 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2377 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2378 */ 2379struct buf * 2380getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2381 int flags) 2382{ 2383 struct buf *bp; 2384 int s; 2385 int error; 2386 ASSERT_VOP_LOCKED(vp, "getblk"); 2387 2388 if (size > MAXBSIZE) 2389 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2390 2391 s = splbio(); 2392loop: 2393 /* 2394 * Block if we are low on buffers. Certain processes are allowed 2395 * to completely exhaust the buffer cache. 2396 * 2397 * If this check ever becomes a bottleneck it may be better to 2398 * move it into the else, when gbincore() fails. At the moment 2399 * it isn't a problem. 2400 * 2401 * XXX remove if 0 sections (clean this up after its proven) 2402 */ 2403 if (numfreebuffers == 0) { 2404 if (curthread == PCPU_GET(idlethread)) 2405 return NULL; 2406 mtx_lock(&nblock); 2407 needsbuffer |= VFS_BIO_NEED_ANY; 2408 mtx_unlock(&nblock); 2409 } 2410 2411 VI_LOCK(vp); 2412 if ((bp = gbincore(vp, blkno))) { 2413 int lockflags; 2414 /* 2415 * Buffer is in-core. If the buffer is not busy, it must 2416 * be on a queue. 2417 */ 2418 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2419 2420 if (flags & GB_LOCK_NOWAIT) 2421 lockflags |= LK_NOWAIT; 2422 2423 error = BUF_TIMELOCK(bp, lockflags, 2424 VI_MTX(vp), "getblk", slpflag, slptimeo); 2425 2426 /* 2427 * If we slept and got the lock we have to restart in case 2428 * the buffer changed identities. 2429 */ 2430 if (error == ENOLCK) 2431 goto loop; 2432 /* We timed out or were interrupted. */ 2433 else if (error) 2434 return (NULL); 2435 2436 /* 2437 * The buffer is locked. B_CACHE is cleared if the buffer is 2438 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2439 * and for a VMIO buffer B_CACHE is adjusted according to the 2440 * backing VM cache. 2441 */ 2442 if (bp->b_flags & B_INVAL) 2443 bp->b_flags &= ~B_CACHE; 2444 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2445 bp->b_flags |= B_CACHE; 2446 bremfree(bp); 2447 2448 /* 2449 * check for size inconsistancies for non-VMIO case. 2450 */ 2451 2452 if (bp->b_bcount != size) { 2453 if ((bp->b_flags & B_VMIO) == 0 || 2454 (size > bp->b_kvasize)) { 2455 if (bp->b_flags & B_DELWRI) { 2456 bp->b_flags |= B_NOCACHE; 2457 BUF_WRITE(bp); 2458 } else { 2459 if ((bp->b_flags & B_VMIO) && 2460 (LIST_FIRST(&bp->b_dep) == NULL)) { 2461 bp->b_flags |= B_RELBUF; 2462 brelse(bp); 2463 } else { 2464 bp->b_flags |= B_NOCACHE; 2465 BUF_WRITE(bp); 2466 } 2467 } 2468 goto loop; 2469 } 2470 } 2471 2472 /* 2473 * If the size is inconsistant in the VMIO case, we can resize 2474 * the buffer. This might lead to B_CACHE getting set or 2475 * cleared. If the size has not changed, B_CACHE remains 2476 * unchanged from its previous state. 2477 */ 2478 2479 if (bp->b_bcount != size) 2480 allocbuf(bp, size); 2481 2482 KASSERT(bp->b_offset != NOOFFSET, 2483 ("getblk: no buffer offset")); 2484 2485 /* 2486 * A buffer with B_DELWRI set and B_CACHE clear must 2487 * be committed before we can return the buffer in 2488 * order to prevent the caller from issuing a read 2489 * ( due to B_CACHE not being set ) and overwriting 2490 * it. 2491 * 2492 * Most callers, including NFS and FFS, need this to 2493 * operate properly either because they assume they 2494 * can issue a read if B_CACHE is not set, or because 2495 * ( for example ) an uncached B_DELWRI might loop due 2496 * to softupdates re-dirtying the buffer. In the latter 2497 * case, B_CACHE is set after the first write completes, 2498 * preventing further loops. 2499 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2500 * above while extending the buffer, we cannot allow the 2501 * buffer to remain with B_CACHE set after the write 2502 * completes or it will represent a corrupt state. To 2503 * deal with this we set B_NOCACHE to scrap the buffer 2504 * after the write. 2505 * 2506 * We might be able to do something fancy, like setting 2507 * B_CACHE in bwrite() except if B_DELWRI is already set, 2508 * so the below call doesn't set B_CACHE, but that gets real 2509 * confusing. This is much easier. 2510 */ 2511 2512 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2513 bp->b_flags |= B_NOCACHE; 2514 BUF_WRITE(bp); 2515 goto loop; 2516 } 2517 2518 splx(s); 2519 bp->b_flags &= ~B_DONE; 2520 } else { 2521 int bsize, maxsize, vmio; 2522 off_t offset; 2523 2524 /* 2525 * Buffer is not in-core, create new buffer. The buffer 2526 * returned by getnewbuf() is locked. Note that the returned 2527 * buffer is also considered valid (not marked B_INVAL). 2528 */ 2529 VI_UNLOCK(vp); 2530 if (vn_isdisk(vp, NULL)) 2531 bsize = DEV_BSIZE; 2532 else if (vp->v_mountedhere) 2533 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2534 else if (vp->v_mount) 2535 bsize = vp->v_mount->mnt_stat.f_iosize; 2536 else 2537 bsize = size; 2538 2539 offset = blkno * bsize; 2540 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2541 (vp->v_vflag & VV_OBJBUF); 2542 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2543 maxsize = imax(maxsize, bsize); 2544 2545 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2546 if (slpflag || slptimeo) { 2547 splx(s); 2548 return NULL; 2549 } 2550 goto loop; 2551 } 2552 2553 /* 2554 * This code is used to make sure that a buffer is not 2555 * created while the getnewbuf routine is blocked. 2556 * This can be a problem whether the vnode is locked or not. 2557 * If the buffer is created out from under us, we have to 2558 * throw away the one we just created. There is now window 2559 * race because we are safely running at splbio() from the 2560 * point of the duplicate buffer creation through to here, 2561 * and we've locked the buffer. 2562 * 2563 * Note: this must occur before we associate the buffer 2564 * with the vp especially considering limitations in 2565 * the splay tree implementation when dealing with duplicate 2566 * lblkno's. 2567 */ 2568 VI_LOCK(vp); 2569 if (gbincore(vp, blkno)) { 2570 VI_UNLOCK(vp); 2571 bp->b_flags |= B_INVAL; 2572 brelse(bp); 2573 goto loop; 2574 } 2575 2576 /* 2577 * Insert the buffer into the hash, so that it can 2578 * be found by incore. 2579 */ 2580 bp->b_blkno = bp->b_lblkno = blkno; 2581 bp->b_offset = offset; 2582 2583 bgetvp(vp, bp); 2584 VI_UNLOCK(vp); 2585 2586 /* 2587 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2588 * buffer size starts out as 0, B_CACHE will be set by 2589 * allocbuf() for the VMIO case prior to it testing the 2590 * backing store for validity. 2591 */ 2592 2593 if (vmio) { 2594 bp->b_flags |= B_VMIO; 2595#if defined(VFS_BIO_DEBUG) 2596 if (vp->v_type != VREG) 2597 printf("getblk: vmioing file type %d???\n", vp->v_type); 2598#endif 2599 VOP_GETVOBJECT(vp, &bp->b_object); 2600 } else { 2601 bp->b_flags &= ~B_VMIO; 2602 bp->b_object = NULL; 2603 } 2604 2605 allocbuf(bp, size); 2606 2607 splx(s); 2608 bp->b_flags &= ~B_DONE; 2609 } 2610 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2611 return (bp); 2612} 2613 2614/* 2615 * Get an empty, disassociated buffer of given size. The buffer is initially 2616 * set to B_INVAL. 2617 */ 2618struct buf * 2619geteblk(int size) 2620{ 2621 struct buf *bp; 2622 int s; 2623 int maxsize; 2624 2625 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2626 2627 s = splbio(); 2628 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2629 continue; 2630 splx(s); 2631 allocbuf(bp, size); 2632 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2633 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2634 return (bp); 2635} 2636 2637 2638/* 2639 * This code constitutes the buffer memory from either anonymous system 2640 * memory (in the case of non-VMIO operations) or from an associated 2641 * VM object (in the case of VMIO operations). This code is able to 2642 * resize a buffer up or down. 2643 * 2644 * Note that this code is tricky, and has many complications to resolve 2645 * deadlock or inconsistant data situations. Tread lightly!!! 2646 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2647 * the caller. Calling this code willy nilly can result in the loss of data. 2648 * 2649 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2650 * B_CACHE for the non-VMIO case. 2651 */ 2652 2653int 2654allocbuf(struct buf *bp, int size) 2655{ 2656 int newbsize, mbsize; 2657 int i; 2658 2659 GIANT_REQUIRED; 2660 2661 if (BUF_REFCNT(bp) == 0) 2662 panic("allocbuf: buffer not busy"); 2663 2664 if (bp->b_kvasize < size) 2665 panic("allocbuf: buffer too small"); 2666 2667 if ((bp->b_flags & B_VMIO) == 0) { 2668 caddr_t origbuf; 2669 int origbufsize; 2670 /* 2671 * Just get anonymous memory from the kernel. Don't 2672 * mess with B_CACHE. 2673 */ 2674 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2675 if (bp->b_flags & B_MALLOC) 2676 newbsize = mbsize; 2677 else 2678 newbsize = round_page(size); 2679 2680 if (newbsize < bp->b_bufsize) { 2681 /* 2682 * malloced buffers are not shrunk 2683 */ 2684 if (bp->b_flags & B_MALLOC) { 2685 if (newbsize) { 2686 bp->b_bcount = size; 2687 } else { 2688 free(bp->b_data, M_BIOBUF); 2689 if (bp->b_bufsize) { 2690 atomic_subtract_int( 2691 &bufmallocspace, 2692 bp->b_bufsize); 2693 bufspacewakeup(); 2694 bp->b_bufsize = 0; 2695 } 2696 bp->b_saveaddr = bp->b_kvabase; 2697 bp->b_data = bp->b_saveaddr; 2698 bp->b_bcount = 0; 2699 bp->b_flags &= ~B_MALLOC; 2700 } 2701 return 1; 2702 } 2703 vm_hold_free_pages( 2704 bp, 2705 (vm_offset_t) bp->b_data + newbsize, 2706 (vm_offset_t) bp->b_data + bp->b_bufsize); 2707 } else if (newbsize > bp->b_bufsize) { 2708 /* 2709 * We only use malloced memory on the first allocation. 2710 * and revert to page-allocated memory when the buffer 2711 * grows. 2712 */ 2713 /* 2714 * There is a potential smp race here that could lead 2715 * to bufmallocspace slightly passing the max. It 2716 * is probably extremely rare and not worth worrying 2717 * over. 2718 */ 2719 if ( (bufmallocspace < maxbufmallocspace) && 2720 (bp->b_bufsize == 0) && 2721 (mbsize <= PAGE_SIZE/2)) { 2722 2723 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2724 bp->b_bufsize = mbsize; 2725 bp->b_bcount = size; 2726 bp->b_flags |= B_MALLOC; 2727 atomic_add_int(&bufmallocspace, mbsize); 2728 return 1; 2729 } 2730 origbuf = NULL; 2731 origbufsize = 0; 2732 /* 2733 * If the buffer is growing on its other-than-first allocation, 2734 * then we revert to the page-allocation scheme. 2735 */ 2736 if (bp->b_flags & B_MALLOC) { 2737 origbuf = bp->b_data; 2738 origbufsize = bp->b_bufsize; 2739 bp->b_data = bp->b_kvabase; 2740 if (bp->b_bufsize) { 2741 atomic_subtract_int(&bufmallocspace, 2742 bp->b_bufsize); 2743 bufspacewakeup(); 2744 bp->b_bufsize = 0; 2745 } 2746 bp->b_flags &= ~B_MALLOC; 2747 newbsize = round_page(newbsize); 2748 } 2749 vm_hold_load_pages( 2750 bp, 2751 (vm_offset_t) bp->b_data + bp->b_bufsize, 2752 (vm_offset_t) bp->b_data + newbsize); 2753 if (origbuf) { 2754 bcopy(origbuf, bp->b_data, origbufsize); 2755 free(origbuf, M_BIOBUF); 2756 } 2757 } 2758 } else { 2759 int desiredpages; 2760 2761 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2762 desiredpages = (size == 0) ? 0 : 2763 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2764 2765 if (bp->b_flags & B_MALLOC) 2766 panic("allocbuf: VMIO buffer can't be malloced"); 2767 /* 2768 * Set B_CACHE initially if buffer is 0 length or will become 2769 * 0-length. 2770 */ 2771 if (size == 0 || bp->b_bufsize == 0) 2772 bp->b_flags |= B_CACHE; 2773 2774 if (newbsize < bp->b_bufsize) { 2775 /* 2776 * DEV_BSIZE aligned new buffer size is less then the 2777 * DEV_BSIZE aligned existing buffer size. Figure out 2778 * if we have to remove any pages. 2779 */ 2780 if (desiredpages < bp->b_npages) { 2781 vm_page_t m; 2782 2783 vm_page_lock_queues(); 2784 for (i = desiredpages; i < bp->b_npages; i++) { 2785 /* 2786 * the page is not freed here -- it 2787 * is the responsibility of 2788 * vnode_pager_setsize 2789 */ 2790 m = bp->b_pages[i]; 2791 KASSERT(m != bogus_page, 2792 ("allocbuf: bogus page found")); 2793 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2794 vm_page_lock_queues(); 2795 2796 bp->b_pages[i] = NULL; 2797 vm_page_unwire(m, 0); 2798 } 2799 vm_page_unlock_queues(); 2800 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2801 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2802 bp->b_npages = desiredpages; 2803 } 2804 } else if (size > bp->b_bcount) { 2805 /* 2806 * We are growing the buffer, possibly in a 2807 * byte-granular fashion. 2808 */ 2809 struct vnode *vp; 2810 vm_object_t obj; 2811 vm_offset_t toff; 2812 vm_offset_t tinc; 2813 2814 /* 2815 * Step 1, bring in the VM pages from the object, 2816 * allocating them if necessary. We must clear 2817 * B_CACHE if these pages are not valid for the 2818 * range covered by the buffer. 2819 */ 2820 2821 vp = bp->b_vp; 2822 obj = bp->b_object; 2823 2824 VM_OBJECT_LOCK(obj); 2825 while (bp->b_npages < desiredpages) { 2826 vm_page_t m; 2827 vm_pindex_t pi; 2828 2829 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2830 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2831 /* 2832 * note: must allocate system pages 2833 * since blocking here could intefere 2834 * with paging I/O, no matter which 2835 * process we are. 2836 */ 2837 m = vm_page_alloc(obj, pi, 2838 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2839 if (m == NULL) { 2840 atomic_add_int(&vm_pageout_deficit, 2841 desiredpages - bp->b_npages); 2842 VM_OBJECT_UNLOCK(obj); 2843 VM_WAIT; 2844 VM_OBJECT_LOCK(obj); 2845 } else { 2846 vm_page_lock_queues(); 2847 vm_page_wakeup(m); 2848 vm_page_unlock_queues(); 2849 bp->b_flags &= ~B_CACHE; 2850 bp->b_pages[bp->b_npages] = m; 2851 ++bp->b_npages; 2852 } 2853 continue; 2854 } 2855 2856 /* 2857 * We found a page. If we have to sleep on it, 2858 * retry because it might have gotten freed out 2859 * from under us. 2860 * 2861 * We can only test PG_BUSY here. Blocking on 2862 * m->busy might lead to a deadlock: 2863 * 2864 * vm_fault->getpages->cluster_read->allocbuf 2865 * 2866 */ 2867 vm_page_lock_queues(); 2868 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2869 continue; 2870 2871 /* 2872 * We have a good page. Should we wakeup the 2873 * page daemon? 2874 */ 2875 if ((curproc != pageproc) && 2876 ((m->queue - m->pc) == PQ_CACHE) && 2877 ((cnt.v_free_count + cnt.v_cache_count) < 2878 (cnt.v_free_min + cnt.v_cache_min))) { 2879 pagedaemon_wakeup(); 2880 } 2881 vm_page_flag_clear(m, PG_ZERO); 2882 vm_page_wire(m); 2883 vm_page_unlock_queues(); 2884 bp->b_pages[bp->b_npages] = m; 2885 ++bp->b_npages; 2886 } 2887 VM_OBJECT_UNLOCK(obj); 2888 2889 /* 2890 * Step 2. We've loaded the pages into the buffer, 2891 * we have to figure out if we can still have B_CACHE 2892 * set. Note that B_CACHE is set according to the 2893 * byte-granular range ( bcount and size ), new the 2894 * aligned range ( newbsize ). 2895 * 2896 * The VM test is against m->valid, which is DEV_BSIZE 2897 * aligned. Needless to say, the validity of the data 2898 * needs to also be DEV_BSIZE aligned. Note that this 2899 * fails with NFS if the server or some other client 2900 * extends the file's EOF. If our buffer is resized, 2901 * B_CACHE may remain set! XXX 2902 */ 2903 2904 toff = bp->b_bcount; 2905 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2906 2907 while ((bp->b_flags & B_CACHE) && toff < size) { 2908 vm_pindex_t pi; 2909 2910 if (tinc > (size - toff)) 2911 tinc = size - toff; 2912 2913 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2914 PAGE_SHIFT; 2915 2916 vfs_buf_test_cache( 2917 bp, 2918 bp->b_offset, 2919 toff, 2920 tinc, 2921 bp->b_pages[pi] 2922 ); 2923 toff += tinc; 2924 tinc = PAGE_SIZE; 2925 } 2926 2927 /* 2928 * Step 3, fixup the KVM pmap. Remember that 2929 * bp->b_data is relative to bp->b_offset, but 2930 * bp->b_offset may be offset into the first page. 2931 */ 2932 2933 bp->b_data = (caddr_t) 2934 trunc_page((vm_offset_t)bp->b_data); 2935 pmap_qenter( 2936 (vm_offset_t)bp->b_data, 2937 bp->b_pages, 2938 bp->b_npages 2939 ); 2940 2941 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2942 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2943 } 2944 } 2945 if (newbsize < bp->b_bufsize) 2946 bufspacewakeup(); 2947 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2948 bp->b_bcount = size; /* requested buffer size */ 2949 return 1; 2950} 2951 2952void 2953biodone(struct bio *bp) 2954{ 2955 mtx_lock(&bdonelock); 2956 bp->bio_flags |= BIO_DONE; 2957 if (bp->bio_done == NULL) 2958 wakeup(bp); 2959 mtx_unlock(&bdonelock); 2960 if (bp->bio_done != NULL) 2961 bp->bio_done(bp); 2962} 2963 2964/* 2965 * Wait for a BIO to finish. 2966 * 2967 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2968 * case is not yet clear. 2969 */ 2970int 2971biowait(struct bio *bp, const char *wchan) 2972{ 2973 2974 mtx_lock(&bdonelock); 2975 while ((bp->bio_flags & BIO_DONE) == 0) 2976 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 2977 mtx_unlock(&bdonelock); 2978 if (bp->bio_error != 0) 2979 return (bp->bio_error); 2980 if (!(bp->bio_flags & BIO_ERROR)) 2981 return (0); 2982 return (EIO); 2983} 2984 2985void 2986biofinish(struct bio *bp, struct devstat *stat, int error) 2987{ 2988 2989 if (error) { 2990 bp->bio_error = error; 2991 bp->bio_flags |= BIO_ERROR; 2992 } 2993 if (stat != NULL) 2994 devstat_end_transaction_bio(stat, bp); 2995 biodone(bp); 2996} 2997 2998/* 2999 * bufwait: 3000 * 3001 * Wait for buffer I/O completion, returning error status. The buffer 3002 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3003 * error and cleared. 3004 */ 3005int 3006bufwait(register struct buf * bp) 3007{ 3008 int s; 3009 3010 s = splbio(); 3011 if (bp->b_iocmd == BIO_READ) 3012 bwait(bp, PRIBIO, "biord"); 3013 else 3014 bwait(bp, PRIBIO, "biowr"); 3015 splx(s); 3016 if (bp->b_flags & B_EINTR) { 3017 bp->b_flags &= ~B_EINTR; 3018 return (EINTR); 3019 } 3020 if (bp->b_ioflags & BIO_ERROR) { 3021 return (bp->b_error ? bp->b_error : EIO); 3022 } else { 3023 return (0); 3024 } 3025} 3026 3027 /* 3028 * Call back function from struct bio back up to struct buf. 3029 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 3030 */ 3031void 3032bufdonebio(struct bio *bp) 3033{ 3034 3035 /* Device drivers may or may not hold giant, hold it here. */ 3036 mtx_lock(&Giant); 3037 bufdone(bp->bio_caller2); 3038 mtx_unlock(&Giant); 3039} 3040 3041/* 3042 * bufdone: 3043 * 3044 * Finish I/O on a buffer, optionally calling a completion function. 3045 * This is usually called from an interrupt so process blocking is 3046 * not allowed. 3047 * 3048 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3049 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3050 * assuming B_INVAL is clear. 3051 * 3052 * For the VMIO case, we set B_CACHE if the op was a read and no 3053 * read error occured, or if the op was a write. B_CACHE is never 3054 * set if the buffer is invalid or otherwise uncacheable. 3055 * 3056 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3057 * initiator to leave B_INVAL set to brelse the buffer out of existance 3058 * in the biodone routine. 3059 */ 3060void 3061bufdone(struct buf *bp) 3062{ 3063 int s; 3064 void (*biodone)(struct buf *); 3065 3066 GIANT_REQUIRED; 3067 3068 s = splbio(); 3069 3070 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3071 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3072 3073 bp->b_flags |= B_DONE; 3074 runningbufwakeup(bp); 3075 3076 if (bp->b_iocmd == BIO_DELETE) { 3077 brelse(bp); 3078 splx(s); 3079 return; 3080 } 3081 3082 if (bp->b_iocmd == BIO_WRITE) { 3083 vwakeup(bp); 3084 } 3085 3086 /* call optional completion function if requested */ 3087 if (bp->b_iodone != NULL) { 3088 biodone = bp->b_iodone; 3089 bp->b_iodone = NULL; 3090 (*biodone) (bp); 3091 splx(s); 3092 return; 3093 } 3094 if (LIST_FIRST(&bp->b_dep) != NULL) 3095 buf_complete(bp); 3096 3097 if (bp->b_flags & B_VMIO) { 3098 int i; 3099 vm_ooffset_t foff; 3100 vm_page_t m; 3101 vm_object_t obj; 3102 int iosize; 3103 struct vnode *vp = bp->b_vp; 3104 3105 obj = bp->b_object; 3106 3107#if defined(VFS_BIO_DEBUG) 3108 mp_fixme("usecount and vflag accessed without locks."); 3109 if (vp->v_usecount == 0) { 3110 panic("biodone: zero vnode ref count"); 3111 } 3112 3113 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3114 panic("biodone: vnode is not setup for merged cache"); 3115 } 3116#endif 3117 3118 foff = bp->b_offset; 3119 KASSERT(bp->b_offset != NOOFFSET, 3120 ("biodone: no buffer offset")); 3121 3122 if (obj != NULL) 3123 VM_OBJECT_LOCK(obj); 3124#if defined(VFS_BIO_DEBUG) 3125 if (obj->paging_in_progress < bp->b_npages) { 3126 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3127 obj->paging_in_progress, bp->b_npages); 3128 } 3129#endif 3130 3131 /* 3132 * Set B_CACHE if the op was a normal read and no error 3133 * occured. B_CACHE is set for writes in the b*write() 3134 * routines. 3135 */ 3136 iosize = bp->b_bcount - bp->b_resid; 3137 if (bp->b_iocmd == BIO_READ && 3138 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3139 !(bp->b_ioflags & BIO_ERROR)) { 3140 bp->b_flags |= B_CACHE; 3141 } 3142 vm_page_lock_queues(); 3143 for (i = 0; i < bp->b_npages; i++) { 3144 int bogusflag = 0; 3145 int resid; 3146 3147 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3148 if (resid > iosize) 3149 resid = iosize; 3150 3151 /* 3152 * cleanup bogus pages, restoring the originals 3153 */ 3154 m = bp->b_pages[i]; 3155 if (m == bogus_page) { 3156 bogusflag = 1; 3157 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3158 if (m == NULL) 3159 panic("biodone: page disappeared!"); 3160 bp->b_pages[i] = m; 3161 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3162 } 3163#if defined(VFS_BIO_DEBUG) 3164 if (OFF_TO_IDX(foff) != m->pindex) { 3165 printf( 3166"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3167 (intmax_t)foff, (uintmax_t)m->pindex); 3168 } 3169#endif 3170 3171 /* 3172 * In the write case, the valid and clean bits are 3173 * already changed correctly ( see bdwrite() ), so we 3174 * only need to do this here in the read case. 3175 */ 3176 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3177 vfs_page_set_valid(bp, foff, i, m); 3178 } 3179 vm_page_flag_clear(m, PG_ZERO); 3180 3181 /* 3182 * when debugging new filesystems or buffer I/O methods, this 3183 * is the most common error that pops up. if you see this, you 3184 * have not set the page busy flag correctly!!! 3185 */ 3186 if (m->busy == 0) { 3187 printf("biodone: page busy < 0, " 3188 "pindex: %d, foff: 0x(%x,%x), " 3189 "resid: %d, index: %d\n", 3190 (int) m->pindex, (int)(foff >> 32), 3191 (int) foff & 0xffffffff, resid, i); 3192 if (!vn_isdisk(vp, NULL)) 3193 printf(" iosize: %ld, lblkno: %jd, flags: 0x%x, npages: %d\n", 3194 bp->b_vp->v_mount->mnt_stat.f_iosize, 3195 (intmax_t) bp->b_lblkno, 3196 bp->b_flags, bp->b_npages); 3197 else 3198 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3199 (intmax_t) bp->b_lblkno, 3200 bp->b_flags, bp->b_npages); 3201 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 3202 m->valid, m->dirty, m->wire_count); 3203 panic("biodone: page busy < 0\n"); 3204 } 3205 vm_page_io_finish(m); 3206 vm_object_pip_subtract(obj, 1); 3207 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3208 iosize -= resid; 3209 } 3210 vm_page_unlock_queues(); 3211 if (obj != NULL) { 3212 vm_object_pip_wakeupn(obj, 0); 3213 VM_OBJECT_UNLOCK(obj); 3214 } 3215 } 3216 3217 /* 3218 * For asynchronous completions, release the buffer now. The brelse 3219 * will do a wakeup there if necessary - so no need to do a wakeup 3220 * here in the async case. The sync case always needs to do a wakeup. 3221 */ 3222 3223 if (bp->b_flags & B_ASYNC) { 3224 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3225 brelse(bp); 3226 else 3227 bqrelse(bp); 3228 } else { 3229 bdone(bp); 3230 } 3231 splx(s); 3232} 3233 3234/* 3235 * This routine is called in lieu of iodone in the case of 3236 * incomplete I/O. This keeps the busy status for pages 3237 * consistant. 3238 */ 3239void 3240vfs_unbusy_pages(struct buf * bp) 3241{ 3242 int i; 3243 3244 GIANT_REQUIRED; 3245 3246 runningbufwakeup(bp); 3247 if (bp->b_flags & B_VMIO) { 3248 vm_object_t obj; 3249 3250 obj = bp->b_object; 3251 VM_OBJECT_LOCK(obj); 3252 vm_page_lock_queues(); 3253 for (i = 0; i < bp->b_npages; i++) { 3254 vm_page_t m = bp->b_pages[i]; 3255 3256 if (m == bogus_page) { 3257 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3258 if (!m) { 3259 panic("vfs_unbusy_pages: page missing\n"); 3260 } 3261 bp->b_pages[i] = m; 3262 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3263 } 3264 vm_object_pip_subtract(obj, 1); 3265 vm_page_flag_clear(m, PG_ZERO); 3266 vm_page_io_finish(m); 3267 } 3268 vm_page_unlock_queues(); 3269 vm_object_pip_wakeupn(obj, 0); 3270 VM_OBJECT_UNLOCK(obj); 3271 } 3272} 3273 3274/* 3275 * vfs_page_set_valid: 3276 * 3277 * Set the valid bits in a page based on the supplied offset. The 3278 * range is restricted to the buffer's size. 3279 * 3280 * This routine is typically called after a read completes. 3281 */ 3282static void 3283vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3284{ 3285 vm_ooffset_t soff, eoff; 3286 3287 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3288 /* 3289 * Start and end offsets in buffer. eoff - soff may not cross a 3290 * page boundry or cross the end of the buffer. The end of the 3291 * buffer, in this case, is our file EOF, not the allocation size 3292 * of the buffer. 3293 */ 3294 soff = off; 3295 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3296 if (eoff > bp->b_offset + bp->b_bcount) 3297 eoff = bp->b_offset + bp->b_bcount; 3298 3299 /* 3300 * Set valid range. This is typically the entire buffer and thus the 3301 * entire page. 3302 */ 3303 if (eoff > soff) { 3304 vm_page_set_validclean( 3305 m, 3306 (vm_offset_t) (soff & PAGE_MASK), 3307 (vm_offset_t) (eoff - soff) 3308 ); 3309 } 3310} 3311 3312/* 3313 * This routine is called before a device strategy routine. 3314 * It is used to tell the VM system that paging I/O is in 3315 * progress, and treat the pages associated with the buffer 3316 * almost as being PG_BUSY. Also the object paging_in_progress 3317 * flag is handled to make sure that the object doesn't become 3318 * inconsistant. 3319 * 3320 * Since I/O has not been initiated yet, certain buffer flags 3321 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3322 * and should be ignored. 3323 */ 3324void 3325vfs_busy_pages(struct buf * bp, int clear_modify) 3326{ 3327 int i, bogus; 3328 3329 if (bp->b_flags & B_VMIO) { 3330 vm_object_t obj; 3331 vm_ooffset_t foff; 3332 3333 obj = bp->b_object; 3334 foff = bp->b_offset; 3335 KASSERT(bp->b_offset != NOOFFSET, 3336 ("vfs_busy_pages: no buffer offset")); 3337 vfs_setdirty(bp); 3338 if (obj != NULL) 3339 VM_OBJECT_LOCK(obj); 3340retry: 3341 vm_page_lock_queues(); 3342 for (i = 0; i < bp->b_npages; i++) { 3343 vm_page_t m = bp->b_pages[i]; 3344 3345 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3346 goto retry; 3347 } 3348 bogus = 0; 3349 for (i = 0; i < bp->b_npages; i++) { 3350 vm_page_t m = bp->b_pages[i]; 3351 3352 vm_page_flag_clear(m, PG_ZERO); 3353 if ((bp->b_flags & B_CLUSTER) == 0) { 3354 vm_object_pip_add(obj, 1); 3355 vm_page_io_start(m); 3356 } 3357 /* 3358 * When readying a buffer for a read ( i.e 3359 * clear_modify == 0 ), it is important to do 3360 * bogus_page replacement for valid pages in 3361 * partially instantiated buffers. Partially 3362 * instantiated buffers can, in turn, occur when 3363 * reconstituting a buffer from its VM backing store 3364 * base. We only have to do this if B_CACHE is 3365 * clear ( which causes the I/O to occur in the 3366 * first place ). The replacement prevents the read 3367 * I/O from overwriting potentially dirty VM-backed 3368 * pages. XXX bogus page replacement is, uh, bogus. 3369 * It may not work properly with small-block devices. 3370 * We need to find a better way. 3371 */ 3372 pmap_remove_all(m); 3373 if (clear_modify) 3374 vfs_page_set_valid(bp, foff, i, m); 3375 else if (m->valid == VM_PAGE_BITS_ALL && 3376 (bp->b_flags & B_CACHE) == 0) { 3377 bp->b_pages[i] = bogus_page; 3378 bogus++; 3379 } 3380 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3381 } 3382 vm_page_unlock_queues(); 3383 if (obj != NULL) 3384 VM_OBJECT_UNLOCK(obj); 3385 if (bogus) 3386 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3387 } 3388} 3389 3390/* 3391 * Tell the VM system that the pages associated with this buffer 3392 * are clean. This is used for delayed writes where the data is 3393 * going to go to disk eventually without additional VM intevention. 3394 * 3395 * Note that while we only really need to clean through to b_bcount, we 3396 * just go ahead and clean through to b_bufsize. 3397 */ 3398static void 3399vfs_clean_pages(struct buf * bp) 3400{ 3401 int i; 3402 3403 GIANT_REQUIRED; 3404 3405 if (bp->b_flags & B_VMIO) { 3406 vm_ooffset_t foff; 3407 3408 foff = bp->b_offset; 3409 KASSERT(bp->b_offset != NOOFFSET, 3410 ("vfs_clean_pages: no buffer offset")); 3411 vm_page_lock_queues(); 3412 for (i = 0; i < bp->b_npages; i++) { 3413 vm_page_t m = bp->b_pages[i]; 3414 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3415 vm_ooffset_t eoff = noff; 3416 3417 if (eoff > bp->b_offset + bp->b_bufsize) 3418 eoff = bp->b_offset + bp->b_bufsize; 3419 vfs_page_set_valid(bp, foff, i, m); 3420 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3421 foff = noff; 3422 } 3423 vm_page_unlock_queues(); 3424 } 3425} 3426 3427/* 3428 * vfs_bio_set_validclean: 3429 * 3430 * Set the range within the buffer to valid and clean. The range is 3431 * relative to the beginning of the buffer, b_offset. Note that b_offset 3432 * itself may be offset from the beginning of the first page. 3433 * 3434 */ 3435 3436void 3437vfs_bio_set_validclean(struct buf *bp, int base, int size) 3438{ 3439 if (bp->b_flags & B_VMIO) { 3440 int i; 3441 int n; 3442 3443 /* 3444 * Fixup base to be relative to beginning of first page. 3445 * Set initial n to be the maximum number of bytes in the 3446 * first page that can be validated. 3447 */ 3448 3449 base += (bp->b_offset & PAGE_MASK); 3450 n = PAGE_SIZE - (base & PAGE_MASK); 3451 3452 vm_page_lock_queues(); 3453 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3454 vm_page_t m = bp->b_pages[i]; 3455 3456 if (n > size) 3457 n = size; 3458 3459 vm_page_set_validclean(m, base & PAGE_MASK, n); 3460 base += n; 3461 size -= n; 3462 n = PAGE_SIZE; 3463 } 3464 vm_page_unlock_queues(); 3465 } 3466} 3467 3468/* 3469 * vfs_bio_clrbuf: 3470 * 3471 * clear a buffer. This routine essentially fakes an I/O, so we need 3472 * to clear BIO_ERROR and B_INVAL. 3473 * 3474 * Note that while we only theoretically need to clear through b_bcount, 3475 * we go ahead and clear through b_bufsize. 3476 */ 3477 3478void 3479vfs_bio_clrbuf(struct buf *bp) 3480{ 3481 int i, mask = 0; 3482 caddr_t sa, ea; 3483 3484 GIANT_REQUIRED; 3485 3486 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3487 bp->b_flags &= ~B_INVAL; 3488 bp->b_ioflags &= ~BIO_ERROR; 3489 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3490 (bp->b_offset & PAGE_MASK) == 0) { 3491 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3492 if ((bp->b_pages[0]->valid & mask) == mask) { 3493 bp->b_resid = 0; 3494 return; 3495 } 3496 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3497 ((bp->b_pages[0]->valid & mask) == 0)) { 3498 bzero(bp->b_data, bp->b_bufsize); 3499 bp->b_pages[0]->valid |= mask; 3500 bp->b_resid = 0; 3501 return; 3502 } 3503 } 3504 ea = sa = bp->b_data; 3505 for(i=0;i<bp->b_npages;i++,sa=ea) { 3506 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3507 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3508 ea = (caddr_t)(vm_offset_t)ulmin( 3509 (u_long)(vm_offset_t)ea, 3510 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3511 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3512 if ((bp->b_pages[i]->valid & mask) == mask) 3513 continue; 3514 if ((bp->b_pages[i]->valid & mask) == 0) { 3515 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3516 bzero(sa, ea - sa); 3517 } 3518 } else { 3519 for (; sa < ea; sa += DEV_BSIZE, j++) { 3520 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3521 (bp->b_pages[i]->valid & (1<<j)) == 0) 3522 bzero(sa, DEV_BSIZE); 3523 } 3524 } 3525 bp->b_pages[i]->valid |= mask; 3526 vm_page_lock_queues(); 3527 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3528 vm_page_unlock_queues(); 3529 } 3530 bp->b_resid = 0; 3531 } else { 3532 clrbuf(bp); 3533 } 3534} 3535 3536/* 3537 * vm_hold_load_pages and vm_hold_free_pages get pages into 3538 * a buffers address space. The pages are anonymous and are 3539 * not associated with a file object. 3540 */ 3541static void 3542vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3543{ 3544 vm_offset_t pg; 3545 vm_page_t p; 3546 int index; 3547 3548 GIANT_REQUIRED; 3549 3550 to = round_page(to); 3551 from = round_page(from); 3552 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3553 3554 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3555tryagain: 3556 /* 3557 * note: must allocate system pages since blocking here 3558 * could intefere with paging I/O, no matter which 3559 * process we are. 3560 */ 3561 VM_OBJECT_LOCK(kernel_object); 3562 p = vm_page_alloc(kernel_object, 3563 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3564 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3565 VM_OBJECT_UNLOCK(kernel_object); 3566 if (!p) { 3567 atomic_add_int(&vm_pageout_deficit, 3568 (to - pg) >> PAGE_SHIFT); 3569 VM_WAIT; 3570 goto tryagain; 3571 } 3572 vm_page_lock_queues(); 3573 p->valid = VM_PAGE_BITS_ALL; 3574 vm_page_unlock_queues(); 3575 pmap_qenter(pg, &p, 1); 3576 bp->b_pages[index] = p; 3577 vm_page_lock_queues(); 3578 vm_page_wakeup(p); 3579 vm_page_unlock_queues(); 3580 } 3581 bp->b_npages = index; 3582} 3583 3584/* Return pages associated with this buf to the vm system */ 3585static void 3586vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3587{ 3588 vm_offset_t pg; 3589 vm_page_t p; 3590 int index, newnpages; 3591 3592 GIANT_REQUIRED; 3593 3594 from = round_page(from); 3595 to = round_page(to); 3596 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3597 3598 if (bp->b_object != NULL) 3599 VM_OBJECT_LOCK(bp->b_object); 3600 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3601 p = bp->b_pages[index]; 3602 if (p && (index < bp->b_npages)) { 3603 if (p->busy) { 3604 printf( 3605 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3606 (intmax_t)bp->b_blkno, 3607 (intmax_t)bp->b_lblkno); 3608 } 3609 bp->b_pages[index] = NULL; 3610 pmap_qremove(pg, 1); 3611 vm_page_lock_queues(); 3612 vm_page_busy(p); 3613 vm_page_unwire(p, 0); 3614 vm_page_free(p); 3615 vm_page_unlock_queues(); 3616 } 3617 } 3618 if (bp->b_object != NULL) 3619 VM_OBJECT_UNLOCK(bp->b_object); 3620 bp->b_npages = newnpages; 3621} 3622 3623/* 3624 * Map an IO request into kernel virtual address space. 3625 * 3626 * All requests are (re)mapped into kernel VA space. 3627 * Notice that we use b_bufsize for the size of the buffer 3628 * to be mapped. b_bcount might be modified by the driver. 3629 * 3630 * Note that even if the caller determines that the address space should 3631 * be valid, a race or a smaller-file mapped into a larger space may 3632 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3633 * check the return value. 3634 */ 3635int 3636vmapbuf(struct buf *bp) 3637{ 3638 caddr_t addr, kva; 3639 vm_paddr_t pa; 3640 int pidx, i; 3641 struct vm_page *m; 3642 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3643 3644 GIANT_REQUIRED; 3645 3646 if ((bp->b_flags & B_PHYS) == 0) 3647 panic("vmapbuf"); 3648 if (bp->b_bufsize < 0) 3649 return (-1); 3650 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3651 addr < bp->b_data + bp->b_bufsize; 3652 addr += PAGE_SIZE, pidx++) { 3653 /* 3654 * Do the vm_fault if needed; do the copy-on-write thing 3655 * when reading stuff off device into memory. 3656 * 3657 * NOTE! Must use pmap_extract() because addr may be in 3658 * the userland address space, and kextract is only guarenteed 3659 * to work for the kernland address space (see: sparc64 port). 3660 */ 3661retry: 3662 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data, 3663 (bp->b_iocmd == BIO_READ) ? 3664 (VM_PROT_READ|VM_PROT_WRITE) : VM_PROT_READ); 3665 if (i < 0) { 3666 vm_page_lock_queues(); 3667 for (i = 0; i < pidx; ++i) { 3668 vm_page_unhold(bp->b_pages[i]); 3669 bp->b_pages[i] = NULL; 3670 } 3671 vm_page_unlock_queues(); 3672 return(-1); 3673 } 3674 pa = pmap_extract(pmap, (vm_offset_t)addr); 3675 if (pa == 0) { 3676 printf("vmapbuf: warning, race against user address during I/O"); 3677 goto retry; 3678 } 3679 m = PHYS_TO_VM_PAGE(pa); 3680 vm_page_lock_queues(); 3681 vm_page_hold(m); 3682 vm_page_unlock_queues(); 3683 bp->b_pages[pidx] = m; 3684 } 3685 if (pidx > btoc(MAXPHYS)) 3686 panic("vmapbuf: mapped more than MAXPHYS"); 3687 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3688 3689 kva = bp->b_saveaddr; 3690 bp->b_npages = pidx; 3691 bp->b_saveaddr = bp->b_data; 3692 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3693 return(0); 3694} 3695 3696/* 3697 * Free the io map PTEs associated with this IO operation. 3698 * We also invalidate the TLB entries and restore the original b_addr. 3699 */ 3700void 3701vunmapbuf(struct buf *bp) 3702{ 3703 int pidx; 3704 int npages; 3705 3706 GIANT_REQUIRED; 3707 3708 if ((bp->b_flags & B_PHYS) == 0) 3709 panic("vunmapbuf"); 3710 3711 npages = bp->b_npages; 3712 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3713 npages); 3714 vm_page_lock_queues(); 3715 for (pidx = 0; pidx < npages; pidx++) 3716 vm_page_unhold(bp->b_pages[pidx]); 3717 vm_page_unlock_queues(); 3718 3719 bp->b_data = bp->b_saveaddr; 3720} 3721 3722void 3723bdone(struct buf *bp) 3724{ 3725 mtx_lock(&bdonelock); 3726 bp->b_flags |= B_DONE; 3727 wakeup(bp); 3728 mtx_unlock(&bdonelock); 3729} 3730 3731void 3732bwait(struct buf *bp, u_char pri, const char *wchan) 3733{ 3734 mtx_lock(&bdonelock); 3735 while ((bp->b_flags & B_DONE) == 0) 3736 msleep(bp, &bdonelock, pri, wchan, 0); 3737 mtx_unlock(&bdonelock); 3738} 3739 3740#include "opt_ddb.h" 3741#ifdef DDB 3742#include <ddb/ddb.h> 3743 3744/* DDB command to show buffer data */ 3745DB_SHOW_COMMAND(buffer, db_show_buffer) 3746{ 3747 /* get args */ 3748 struct buf *bp = (struct buf *)addr; 3749 3750 if (!have_addr) { 3751 db_printf("usage: show buffer <addr>\n"); 3752 return; 3753 } 3754 3755 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3756 db_printf( 3757 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3758 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd, b_pblkno = %jd\n", 3759 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3760 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3761 (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno); 3762 if (bp->b_npages) { 3763 int i; 3764 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3765 for (i = 0; i < bp->b_npages; i++) { 3766 vm_page_t m; 3767 m = bp->b_pages[i]; 3768 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3769 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3770 if ((i + 1) < bp->b_npages) 3771 db_printf(","); 3772 } 3773 db_printf("\n"); 3774 } 3775} 3776#endif /* DDB */ 3777