1 /* 2 * Copyright (c) 1994,1997 John S. Dyson 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice immediately at the beginning of the file, without modification, 10 * this list of conditions, and the following disclaimer. 11 * 2. Absolutely no warranty of function or purpose is made by the author 12 * John S. Dyson. 13 * 14 * $FreeBSD$ 15 */ 16 17 /* 18 * this file contains a new buffer I/O scheme implementing a coherent 19 * VM object and buffer cache scheme. Pains have been taken to make 20 * sure that the performance degradation associated with schemes such 21 * as this is not realized. 22 * 23 * Author: John S. Dyson 24 * Significant help during the development and debugging phases 25 * had been provided by David Greenman, also of the FreeBSD core team. 26 * 27 * see man buf(9) for more info. 28 */ 29 30 #include <sys/param.h> 31 #include <sys/systm.h> 32 #include <sys/bio.h> 33 #include <sys/buf.h> 34 #include <sys/devicestat.h> 35 #include <sys/eventhandler.h> 36 #include <sys/lock.h> 37 #include <sys/malloc.h> 38 #include <sys/mount.h> 39 #include <sys/mutex.h> 40 #include <sys/kernel.h> 41 #include <sys/kthread.h> 42 #include <sys/proc.h> 43 #include <sys/resourcevar.h> 44 #include <sys/sysctl.h> 45 #include <sys/vmmeter.h> 46 #include <sys/vnode.h> 47 #include <vm/vm.h> 48 #include <vm/vm_param.h> 49 #include <vm/vm_kern.h> 50 #include <vm/vm_pageout.h> 51 #include <vm/vm_page.h> 52 #include <vm/vm_object.h> 53 #include <vm/vm_extern.h> 54 #include <vm/vm_map.h> 55 56 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 57 58 struct bio_ops bioops; /* I/O operation notification */ 59 60 struct buf_ops buf_ops_bio = { 61 "buf_ops_bio", 62 bwrite 63 }; 64 65 /* 66 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 67 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 68 */ 69 struct buf *buf; /* buffer header pool */ 70 71 static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 72 vm_offset_t to); 73 static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 74 vm_offset_t to); 75 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 76 int pageno, vm_page_t m); 77 static void vfs_clean_pages(struct buf * bp); 78 static void vfs_setdirty(struct buf *bp); 79 static void vfs_vmio_release(struct buf *bp); 80 static void vfs_backgroundwritedone(struct buf *bp); 81 static int vfs_bio_clcheck(struct vnode *vp, int size, 82 daddr_t lblkno, daddr_t blkno); 83 static int flushbufqueues(int flushdeps); 84 static void buf_daemon(void); 85 void bremfreel(struct buf * bp); 86 87 int vmiodirenable = TRUE; 88 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 89 "Use the VM system for directory writes"); 90 int runningbufspace; 91 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 92 "Amount of presently outstanding async buffer io"); 93 static int bufspace; 94 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 95 "KVA memory used for bufs"); 96 static int maxbufspace; 97 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 98 "Maximum allowed value of bufspace (including buf_daemon)"); 99 static int bufmallocspace; 100 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 101 "Amount of malloced memory for buffers"); 102 static int maxbufmallocspace; 103 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 104 "Maximum amount of malloced memory for buffers"); 105 static int lobufspace; 106 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 107 "Minimum amount of buffers we want to have"); 108 static int hibufspace; 109 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 110 "Maximum allowed value of bufspace (excluding buf_daemon)"); 111 static int bufreusecnt; 112 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 113 "Number of times we have reused a buffer"); 114 static int buffreekvacnt; 115 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 116 "Number of times we have freed the KVA space from some buffer"); 117 static int bufdefragcnt; 118 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 119 "Number of times we have had to repeat buffer allocation to defragment"); 120 static int lorunningspace; 121 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 122 "Minimum preferred space used for in-progress I/O"); 123 static int hirunningspace; 124 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 125 "Maximum amount of space to use for in-progress I/O"); 126 static int dirtybufferflushes; 127 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 128 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 129 static int altbufferflushes; 130 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 131 0, "Number of fsync flushes to limit dirty buffers"); 132 static int recursiveflushes; 133 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 134 0, "Number of flushes skipped due to being recursive"); 135 static int numdirtybuffers; 136 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 137 "Number of buffers that are dirty (has unwritten changes) at the moment"); 138 static int lodirtybuffers; 139 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 140 "How many buffers we want to have free before bufdaemon can sleep"); 141 static int hidirtybuffers; 142 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 143 "When the number of dirty buffers is considered severe"); 144 static int dirtybufthresh; 145 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 146 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 147 static int numfreebuffers; 148 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 149 "Number of free buffers"); 150 static int lofreebuffers; 151 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 152 "XXX Unused"); 153 static int hifreebuffers; 154 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 155 "XXX Complicatedly unused"); 156 static int getnewbufcalls; 157 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 158 "Number of calls to getnewbuf"); 159 static int getnewbufrestarts; 160 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 161 "Number of times getnewbuf has had to restart a buffer aquisition"); 162 static int dobkgrdwrite = 1; 163 SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 164 "Do background writes (honoring the BX_BKGRDWRITE flag)?"); 165 166 /* 167 * Wakeup point for bufdaemon, as well as indicator of whether it is already 168 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 169 * is idling. 170 */ 171 static int bd_request; 172 173 /* 174 * This lock synchronizes access to bd_request. 175 */ 176 static struct mtx bdlock; 177 178 /* 179 * bogus page -- for I/O to/from partially complete buffers 180 * this is a temporary solution to the problem, but it is not 181 * really that bad. it would be better to split the buffer 182 * for input in the case of buffers partially already in memory, 183 * but the code is intricate enough already. 184 */ 185 vm_page_t bogus_page; 186 187 /* 188 * Synchronization (sleep/wakeup) variable for active buffer space requests. 189 * Set when wait starts, cleared prior to wakeup(). 190 * Used in runningbufwakeup() and waitrunningbufspace(). 191 */ 192 static int runningbufreq; 193 194 /* 195 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 196 * waitrunningbufspace(). 197 */ 198 static struct mtx rbreqlock; 199 200 /* 201 * Synchronization (sleep/wakeup) variable for buffer requests. 202 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 203 * by and/or. 204 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 205 * getnewbuf(), and getblk(). 206 */ 207 static int needsbuffer; 208 209 /* 210 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 211 */ 212 static struct mtx nblock; 213 214 /* 215 * Lock that protects against bwait()/bdone()/B_DONE races. 216 */ 217 218 static struct mtx bdonelock; 219 220 /* 221 * Definitions for the buffer free lists. 222 */ 223 #define BUFFER_QUEUES 6 /* number of free buffer queues */ 224 225 #define QUEUE_NONE 0 /* on no queue */ 226 #define QUEUE_LOCKED 1 /* locked buffers */ 227 #define QUEUE_CLEAN 2 /* non-B_DELWRI buffers */ 228 #define QUEUE_DIRTY 3 /* B_DELWRI buffers */ 229 #define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 230 #define QUEUE_EMPTY 5 /* empty buffer headers */ 231 232 /* Queues for free buffers with various properties */ 233 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 234 235 /* Lock for the bufqueues */ 236 static struct mtx bqlock; 237 238 /* 239 * Single global constant for BUF_WMESG, to avoid getting multiple references. 240 * buf_wmesg is referred from macros. 241 */ 242 const char *buf_wmesg = BUF_WMESG; 243 244 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 245 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 246 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 247 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 248 249 /* 250 * numdirtywakeup: 251 * 252 * If someone is blocked due to there being too many dirty buffers, 253 * and numdirtybuffers is now reasonable, wake them up. 254 */ 255 256 static __inline void 257 numdirtywakeup(int level) 258 { 259 if (numdirtybuffers <= level) { 260 mtx_lock(&nblock); 261 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 262 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 263 wakeup(&needsbuffer); 264 } 265 mtx_unlock(&nblock); 266 } 267 } 268 269 /* 270 * bufspacewakeup: 271 * 272 * Called when buffer space is potentially available for recovery. 273 * getnewbuf() will block on this flag when it is unable to free 274 * sufficient buffer space. Buffer space becomes recoverable when 275 * bp's get placed back in the queues. 276 */ 277 278 static __inline void 279 bufspacewakeup(void) 280 { 281 /* 282 * If someone is waiting for BUF space, wake them up. Even 283 * though we haven't freed the kva space yet, the waiting 284 * process will be able to now. 285 */ 286 mtx_lock(&nblock); 287 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 288 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 289 wakeup(&needsbuffer); 290 } 291 mtx_unlock(&nblock); 292 } 293 294 /* 295 * runningbufwakeup() - in-progress I/O accounting. 296 * 297 */ 298 static __inline void 299 runningbufwakeup(struct buf *bp) 300 { 301 if (bp->b_runningbufspace) { 302 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 303 bp->b_runningbufspace = 0; 304 mtx_lock(&rbreqlock); 305 if (runningbufreq && runningbufspace <= lorunningspace) { 306 runningbufreq = 0; 307 wakeup(&runningbufreq); 308 } 309 mtx_unlock(&rbreqlock); 310 } 311 } 312 313 /* 314 * bufcountwakeup: 315 * 316 * Called when a buffer has been added to one of the free queues to 317 * account for the buffer and to wakeup anyone waiting for free buffers. 318 * This typically occurs when large amounts of metadata are being handled 319 * by the buffer cache ( else buffer space runs out first, usually ). 320 */ 321 322 static __inline void 323 bufcountwakeup(void) 324 { 325 atomic_add_int(&numfreebuffers, 1); 326 mtx_lock(&nblock); 327 if (needsbuffer) { 328 needsbuffer &= ~VFS_BIO_NEED_ANY; 329 if (numfreebuffers >= hifreebuffers) 330 needsbuffer &= ~VFS_BIO_NEED_FREE; 331 wakeup(&needsbuffer); 332 } 333 mtx_unlock(&nblock); 334 } 335 336 /* 337 * waitrunningbufspace() 338 * 339 * runningbufspace is a measure of the amount of I/O currently 340 * running. This routine is used in async-write situations to 341 * prevent creating huge backups of pending writes to a device. 342 * Only asynchronous writes are governed by this function. 343 * 344 * Reads will adjust runningbufspace, but will not block based on it. 345 * The read load has a side effect of reducing the allowed write load. 346 * 347 * This does NOT turn an async write into a sync write. It waits 348 * for earlier writes to complete and generally returns before the 349 * caller's write has reached the device. 350 */ 351 static __inline void 352 waitrunningbufspace(void) 353 { 354 mtx_lock(&rbreqlock); 355 while (runningbufspace > hirunningspace) { 356 ++runningbufreq; 357 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 358 } 359 mtx_unlock(&rbreqlock); 360 } 361 362 363 /* 364 * vfs_buf_test_cache: 365 * 366 * Called when a buffer is extended. This function clears the B_CACHE 367 * bit if the newly extended portion of the buffer does not contain 368 * valid data. 369 */ 370 static __inline__ 371 void 372 vfs_buf_test_cache(struct buf *bp, 373 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 374 vm_page_t m) 375 { 376 GIANT_REQUIRED; 377 378 if (bp->b_flags & B_CACHE) { 379 int base = (foff + off) & PAGE_MASK; 380 if (vm_page_is_valid(m, base, size) == 0) 381 bp->b_flags &= ~B_CACHE; 382 } 383 } 384 385 /* Wake up the buffer deamon if necessary */ 386 static __inline__ 387 void 388 bd_wakeup(int dirtybuflevel) 389 { 390 mtx_lock(&bdlock); 391 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 392 bd_request = 1; 393 wakeup(&bd_request); 394 } 395 mtx_unlock(&bdlock); 396 } 397 398 /* 399 * bd_speedup - speedup the buffer cache flushing code 400 */ 401 402 static __inline__ 403 void 404 bd_speedup(void) 405 { 406 bd_wakeup(1); 407 } 408 409 /* 410 * Calculating buffer cache scaling values and reserve space for buffer 411 * headers. This is called during low level kernel initialization and 412 * may be called more then once. We CANNOT write to the memory area 413 * being reserved at this time. 414 */ 415 caddr_t 416 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 417 { 418 /* 419 * physmem_est is in pages. Convert it to kilobytes (assumes 420 * PAGE_SIZE is >= 1K) 421 */ 422 physmem_est = physmem_est * (PAGE_SIZE / 1024); 423 424 /* 425 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 426 * For the first 64MB of ram nominally allocate sufficient buffers to 427 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 428 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 429 * the buffer cache we limit the eventual kva reservation to 430 * maxbcache bytes. 431 * 432 * factor represents the 1/4 x ram conversion. 433 */ 434 if (nbuf == 0) { 435 int factor = 4 * BKVASIZE / 1024; 436 437 nbuf = 50; 438 if (physmem_est > 4096) 439 nbuf += min((physmem_est - 4096) / factor, 440 65536 / factor); 441 if (physmem_est > 65536) 442 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 443 444 if (maxbcache && nbuf > maxbcache / BKVASIZE) 445 nbuf = maxbcache / BKVASIZE; 446 } 447 448 #if 0 449 /* 450 * Do not allow the buffer_map to be more then 1/2 the size of the 451 * kernel_map. 452 */ 453 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 454 (BKVASIZE * 2)) { 455 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 456 (BKVASIZE * 2); 457 printf("Warning: nbufs capped at %d\n", nbuf); 458 } 459 #endif 460 461 /* 462 * swbufs are used as temporary holders for I/O, such as paging I/O. 463 * We have no less then 16 and no more then 256. 464 */ 465 nswbuf = max(min(nbuf/4, 256), 16); 466 467 /* 468 * Reserve space for the buffer cache buffers 469 */ 470 swbuf = (void *)v; 471 v = (caddr_t)(swbuf + nswbuf); 472 buf = (void *)v; 473 v = (caddr_t)(buf + nbuf); 474 475 return(v); 476 } 477 478 /* Initialize the buffer subsystem. Called before use of any buffers. */ 479 void 480 bufinit(void) 481 { 482 struct buf *bp; 483 vm_offset_t bogus_offset; 484 int i; 485 486 GIANT_REQUIRED; 487 488 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 489 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 490 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 491 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 492 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 493 494 /* next, make a null set of free lists */ 495 for (i = 0; i < BUFFER_QUEUES; i++) 496 TAILQ_INIT(&bufqueues[i]); 497 498 /* finally, initialize each buffer header and stick on empty q */ 499 for (i = 0; i < nbuf; i++) { 500 bp = &buf[i]; 501 bzero(bp, sizeof *bp); 502 bp->b_flags = B_INVAL; /* we're just an empty header */ 503 bp->b_dev = NODEV; 504 bp->b_rcred = NOCRED; 505 bp->b_wcred = NOCRED; 506 bp->b_qindex = QUEUE_EMPTY; 507 bp->b_vflags = 0; 508 bp->b_xflags = 0; 509 LIST_INIT(&bp->b_dep); 510 BUF_LOCKINIT(bp); 511 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 512 } 513 514 /* 515 * maxbufspace is the absolute maximum amount of buffer space we are 516 * allowed to reserve in KVM and in real terms. The absolute maximum 517 * is nominally used by buf_daemon. hibufspace is the nominal maximum 518 * used by most other processes. The differential is required to 519 * ensure that buf_daemon is able to run when other processes might 520 * be blocked waiting for buffer space. 521 * 522 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 523 * this may result in KVM fragmentation which is not handled optimally 524 * by the system. 525 */ 526 maxbufspace = nbuf * BKVASIZE; 527 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 528 lobufspace = hibufspace - MAXBSIZE; 529 530 lorunningspace = 512 * 1024; 531 hirunningspace = 1024 * 1024; 532 533 /* 534 * Limit the amount of malloc memory since it is wired permanently into 535 * the kernel space. Even though this is accounted for in the buffer 536 * allocation, we don't want the malloced region to grow uncontrolled. 537 * The malloc scheme improves memory utilization significantly on average 538 * (small) directories. 539 */ 540 maxbufmallocspace = hibufspace / 20; 541 542 /* 543 * Reduce the chance of a deadlock occuring by limiting the number 544 * of delayed-write dirty buffers we allow to stack up. 545 */ 546 hidirtybuffers = nbuf / 4 + 20; 547 dirtybufthresh = hidirtybuffers * 9 / 10; 548 numdirtybuffers = 0; 549 /* 550 * To support extreme low-memory systems, make sure hidirtybuffers cannot 551 * eat up all available buffer space. This occurs when our minimum cannot 552 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 553 * BKVASIZE'd (8K) buffers. 554 */ 555 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 556 hidirtybuffers >>= 1; 557 } 558 lodirtybuffers = hidirtybuffers / 2; 559 560 /* 561 * Try to keep the number of free buffers in the specified range, 562 * and give special processes (e.g. like buf_daemon) access to an 563 * emergency reserve. 564 */ 565 lofreebuffers = nbuf / 18 + 5; 566 hifreebuffers = 2 * lofreebuffers; 567 numfreebuffers = nbuf; 568 569 /* 570 * Maximum number of async ops initiated per buf_daemon loop. This is 571 * somewhat of a hack at the moment, we really need to limit ourselves 572 * based on the number of bytes of I/O in-transit that were initiated 573 * from buf_daemon. 574 */ 575 576 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 577 vm_object_lock(kernel_object); 578 bogus_page = vm_page_alloc(kernel_object, 579 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 580 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 581 vm_object_unlock(kernel_object); 582 } 583 584 /* 585 * bfreekva() - free the kva allocation for a buffer. 586 * 587 * Must be called at splbio() or higher as this is the only locking for 588 * buffer_map. 589 * 590 * Since this call frees up buffer space, we call bufspacewakeup(). 591 */ 592 static void 593 bfreekva(struct buf * bp) 594 { 595 GIANT_REQUIRED; 596 597 if (bp->b_kvasize) { 598 atomic_add_int(&buffreekvacnt, 1); 599 atomic_subtract_int(&bufspace, bp->b_kvasize); 600 vm_map_delete(buffer_map, 601 (vm_offset_t) bp->b_kvabase, 602 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 603 ); 604 bp->b_kvasize = 0; 605 bufspacewakeup(); 606 } 607 } 608 609 /* 610 * bremfree: 611 * 612 * Remove the buffer from the appropriate free list. 613 */ 614 void 615 bremfree(struct buf * bp) 616 { 617 mtx_lock(&bqlock); 618 bremfreel(bp); 619 mtx_unlock(&bqlock); 620 } 621 622 void 623 bremfreel(struct buf * bp) 624 { 625 int s = splbio(); 626 int old_qindex = bp->b_qindex; 627 628 GIANT_REQUIRED; 629 630 if (bp->b_qindex != QUEUE_NONE) { 631 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 632 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 633 bp->b_qindex = QUEUE_NONE; 634 } else { 635 if (BUF_REFCNT(bp) <= 1) 636 panic("bremfree: removing a buffer not on a queue"); 637 } 638 639 /* 640 * Fixup numfreebuffers count. If the buffer is invalid or not 641 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 642 * the buffer was free and we must decrement numfreebuffers. 643 */ 644 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 645 switch(old_qindex) { 646 case QUEUE_DIRTY: 647 case QUEUE_CLEAN: 648 case QUEUE_EMPTY: 649 case QUEUE_EMPTYKVA: 650 atomic_subtract_int(&numfreebuffers, 1); 651 break; 652 default: 653 break; 654 } 655 } 656 splx(s); 657 } 658 659 660 /* 661 * Get a buffer with the specified data. Look in the cache first. We 662 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 663 * is set, the buffer is valid and we do not have to do anything ( see 664 * getblk() ). This is really just a special case of breadn(). 665 */ 666 int 667 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 668 struct buf ** bpp) 669 { 670 671 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 672 } 673 674 /* 675 * Operates like bread, but also starts asynchronous I/O on 676 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 677 * to initiating I/O . If B_CACHE is set, the buffer is valid 678 * and we do not have to do anything. 679 */ 680 int 681 breadn(struct vnode * vp, daddr_t blkno, int size, 682 daddr_t * rablkno, int *rabsize, 683 int cnt, struct ucred * cred, struct buf ** bpp) 684 { 685 struct buf *bp, *rabp; 686 int i; 687 int rv = 0, readwait = 0; 688 689 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 690 691 /* if not found in cache, do some I/O */ 692 if ((bp->b_flags & B_CACHE) == 0) { 693 if (curthread != PCPU_GET(idlethread)) 694 curthread->td_proc->p_stats->p_ru.ru_inblock++; 695 bp->b_iocmd = BIO_READ; 696 bp->b_flags &= ~B_INVAL; 697 bp->b_ioflags &= ~BIO_ERROR; 698 if (bp->b_rcred == NOCRED && cred != NOCRED) 699 bp->b_rcred = crhold(cred); 700 vfs_busy_pages(bp, 0); 701 if (vp->v_type == VCHR) 702 VOP_SPECSTRATEGY(vp, bp); 703 else 704 VOP_STRATEGY(vp, bp); 705 ++readwait; 706 } 707 708 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 709 if (inmem(vp, *rablkno)) 710 continue; 711 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 712 713 if ((rabp->b_flags & B_CACHE) == 0) { 714 if (curthread != PCPU_GET(idlethread)) 715 curthread->td_proc->p_stats->p_ru.ru_inblock++; 716 rabp->b_flags |= B_ASYNC; 717 rabp->b_flags &= ~B_INVAL; 718 rabp->b_ioflags &= ~BIO_ERROR; 719 rabp->b_iocmd = BIO_READ; 720 if (rabp->b_rcred == NOCRED && cred != NOCRED) 721 rabp->b_rcred = crhold(cred); 722 vfs_busy_pages(rabp, 0); 723 BUF_KERNPROC(rabp); 724 if (vp->v_type == VCHR) 725 VOP_SPECSTRATEGY(vp, rabp); 726 else 727 VOP_STRATEGY(vp, rabp); 728 } else { 729 brelse(rabp); 730 } 731 } 732 733 if (readwait) { 734 rv = bufwait(bp); 735 } 736 return (rv); 737 } 738 739 /* 740 * Write, release buffer on completion. (Done by iodone 741 * if async). Do not bother writing anything if the buffer 742 * is invalid. 743 * 744 * Note that we set B_CACHE here, indicating that buffer is 745 * fully valid and thus cacheable. This is true even of NFS 746 * now so we set it generally. This could be set either here 747 * or in biodone() since the I/O is synchronous. We put it 748 * here. 749 */ 750 751 int 752 bwrite(struct buf * bp) 753 { 754 int oldflags, s; 755 struct buf *newbp; 756 757 if (bp->b_flags & B_INVAL) { 758 brelse(bp); 759 return (0); 760 } 761 762 oldflags = bp->b_flags; 763 764 if (BUF_REFCNT(bp) == 0) 765 panic("bwrite: buffer is not busy???"); 766 s = splbio(); 767 /* 768 * If a background write is already in progress, delay 769 * writing this block if it is asynchronous. Otherwise 770 * wait for the background write to complete. 771 */ 772 if (bp->b_xflags & BX_BKGRDINPROG) { 773 if (bp->b_flags & B_ASYNC) { 774 splx(s); 775 bdwrite(bp); 776 return (0); 777 } 778 bp->b_xflags |= BX_BKGRDWAIT; 779 tsleep(&bp->b_xflags, PRIBIO, "bwrbg", 0); 780 if (bp->b_xflags & BX_BKGRDINPROG) 781 panic("bwrite: still writing"); 782 } 783 784 /* Mark the buffer clean */ 785 bundirty(bp); 786 787 /* 788 * If this buffer is marked for background writing and we 789 * do not have to wait for it, make a copy and write the 790 * copy so as to leave this buffer ready for further use. 791 * 792 * This optimization eats a lot of memory. If we have a page 793 * or buffer shortfall we can't do it. 794 */ 795 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 796 (bp->b_flags & B_ASYNC) && 797 !vm_page_count_severe() && 798 !buf_dirty_count_severe()) { 799 if (bp->b_iodone != NULL) { 800 printf("bp->b_iodone = %p\n", bp->b_iodone); 801 panic("bwrite: need chained iodone"); 802 } 803 804 /* get a new block */ 805 newbp = geteblk(bp->b_bufsize); 806 807 /* 808 * set it to be identical to the old block. We have to 809 * set b_lblkno and BKGRDMARKER before calling bgetvp() 810 * to avoid confusing the splay tree and gbincore(). 811 */ 812 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 813 newbp->b_lblkno = bp->b_lblkno; 814 newbp->b_xflags |= BX_BKGRDMARKER; 815 /* XXX The BX_ flags need to be protected as well */ 816 VI_LOCK(bp->b_vp); 817 bgetvp(bp->b_vp, newbp); 818 VI_UNLOCK(bp->b_vp); 819 newbp->b_blkno = bp->b_blkno; 820 newbp->b_offset = bp->b_offset; 821 newbp->b_iodone = vfs_backgroundwritedone; 822 newbp->b_flags |= B_ASYNC; 823 newbp->b_flags &= ~B_INVAL; 824 825 /* move over the dependencies */ 826 if (LIST_FIRST(&bp->b_dep) != NULL) 827 buf_movedeps(bp, newbp); 828 829 /* 830 * Initiate write on the copy, release the original to 831 * the B_LOCKED queue so that it cannot go away until 832 * the background write completes. If not locked it could go 833 * away and then be reconstituted while it was being written. 834 * If the reconstituted buffer were written, we could end up 835 * with two background copies being written at the same time. 836 */ 837 bp->b_xflags |= BX_BKGRDINPROG; 838 bp->b_flags |= B_LOCKED; 839 bqrelse(bp); 840 bp = newbp; 841 } 842 843 bp->b_flags &= ~B_DONE; 844 bp->b_ioflags &= ~BIO_ERROR; 845 bp->b_flags |= B_WRITEINPROG | B_CACHE; 846 bp->b_iocmd = BIO_WRITE; 847 848 VI_LOCK(bp->b_vp); 849 bp->b_vp->v_numoutput++; 850 VI_UNLOCK(bp->b_vp); 851 vfs_busy_pages(bp, 1); 852 853 /* 854 * Normal bwrites pipeline writes 855 */ 856 bp->b_runningbufspace = bp->b_bufsize; 857 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 858 859 if (curthread != PCPU_GET(idlethread)) 860 curthread->td_proc->p_stats->p_ru.ru_oublock++; 861 splx(s); 862 if (oldflags & B_ASYNC) 863 BUF_KERNPROC(bp); 864 if (bp->b_vp->v_type == VCHR) 865 VOP_SPECSTRATEGY(bp->b_vp, bp); 866 else 867 VOP_STRATEGY(bp->b_vp, bp); 868 869 if ((oldflags & B_ASYNC) == 0) { 870 int rtval = bufwait(bp); 871 brelse(bp); 872 return (rtval); 873 } else if ((oldflags & B_NOWDRAIN) == 0) { 874 /* 875 * don't allow the async write to saturate the I/O 876 * system. Deadlocks can occur only if a device strategy 877 * routine (like in MD) turns around and issues another 878 * high-level write, in which case B_NOWDRAIN is expected 879 * to be set. Otherwise we will not deadlock here because 880 * we are blocking waiting for I/O that is already in-progress 881 * to complete. 882 */ 883 waitrunningbufspace(); 884 } 885 886 return (0); 887 } 888 889 /* 890 * Complete a background write started from bwrite. 891 */ 892 static void 893 vfs_backgroundwritedone(bp) 894 struct buf *bp; 895 { 896 struct buf *origbp; 897 898 /* 899 * Find the original buffer that we are writing. 900 */ 901 VI_LOCK(bp->b_vp); 902 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 903 panic("backgroundwritedone: lost buffer"); 904 VI_UNLOCK(bp->b_vp); 905 /* 906 * Process dependencies then return any unfinished ones. 907 */ 908 if (LIST_FIRST(&bp->b_dep) != NULL) 909 buf_complete(bp); 910 if (LIST_FIRST(&bp->b_dep) != NULL) 911 buf_movedeps(bp, origbp); 912 913 /* XXX Find out if origbp can disappear or get inconsistent */ 914 /* 915 * Clear the BX_BKGRDINPROG flag in the original buffer 916 * and awaken it if it is waiting for the write to complete. 917 * If BX_BKGRDINPROG is not set in the original buffer it must 918 * have been released and re-instantiated - which is not legal. 919 */ 920 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), 921 ("backgroundwritedone: lost buffer2")); 922 origbp->b_xflags &= ~BX_BKGRDINPROG; 923 if (origbp->b_xflags & BX_BKGRDWAIT) { 924 origbp->b_xflags &= ~BX_BKGRDWAIT; 925 wakeup(&origbp->b_xflags); 926 } 927 /* 928 * Clear the B_LOCKED flag and remove it from the locked 929 * queue if it currently resides there. 930 */ 931 origbp->b_flags &= ~B_LOCKED; 932 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) { 933 bremfree(origbp); 934 bqrelse(origbp); 935 } 936 /* 937 * This buffer is marked B_NOCACHE, so when it is released 938 * by biodone, it will be tossed. We mark it with BIO_READ 939 * to avoid biodone doing a second vwakeup. 940 */ 941 bp->b_flags |= B_NOCACHE; 942 bp->b_iocmd = BIO_READ; 943 bp->b_flags &= ~(B_CACHE | B_DONE); 944 bp->b_iodone = 0; 945 bufdone(bp); 946 } 947 948 /* 949 * Delayed write. (Buffer is marked dirty). Do not bother writing 950 * anything if the buffer is marked invalid. 951 * 952 * Note that since the buffer must be completely valid, we can safely 953 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 954 * biodone() in order to prevent getblk from writing the buffer 955 * out synchronously. 956 */ 957 void 958 bdwrite(struct buf * bp) 959 { 960 struct thread *td = curthread; 961 struct vnode *vp; 962 struct buf *nbp; 963 964 GIANT_REQUIRED; 965 966 if (BUF_REFCNT(bp) == 0) 967 panic("bdwrite: buffer is not busy"); 968 969 if (bp->b_flags & B_INVAL) { 970 brelse(bp); 971 return; 972 } 973 974 /* 975 * If we have too many dirty buffers, don't create any more. 976 * If we are wildly over our limit, then force a complete 977 * cleanup. Otherwise, just keep the situation from getting 978 * out of control. Note that we have to avoid a recursive 979 * disaster and not try to clean up after our own cleanup! 980 */ 981 vp = bp->b_vp; 982 VI_LOCK(vp); 983 if (td->td_proc->p_flag & P_COWINPROGRESS) { 984 recursiveflushes++; 985 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) { 986 VI_UNLOCK(vp); 987 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 988 VI_LOCK(vp); 989 altbufferflushes++; 990 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) { 991 /* 992 * Try to find a buffer to flush. 993 */ 994 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) { 995 if ((nbp->b_xflags & BX_BKGRDINPROG) || 996 buf_countdeps(nbp, 0) || 997 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 998 continue; 999 if (bp == nbp) 1000 panic("bdwrite: found ourselves"); 1001 VI_UNLOCK(vp); 1002 if (nbp->b_flags & B_CLUSTEROK) { 1003 vfs_bio_awrite(nbp); 1004 } else { 1005 bremfree(nbp); 1006 bawrite(nbp); 1007 } 1008 VI_LOCK(vp); 1009 dirtybufferflushes++; 1010 break; 1011 } 1012 } 1013 VI_UNLOCK(vp); 1014 1015 bdirty(bp); 1016 /* 1017 * Set B_CACHE, indicating that the buffer is fully valid. This is 1018 * true even of NFS now. 1019 */ 1020 bp->b_flags |= B_CACHE; 1021 1022 /* 1023 * This bmap keeps the system from needing to do the bmap later, 1024 * perhaps when the system is attempting to do a sync. Since it 1025 * is likely that the indirect block -- or whatever other datastructure 1026 * that the filesystem needs is still in memory now, it is a good 1027 * thing to do this. Note also, that if the pageout daemon is 1028 * requesting a sync -- there might not be enough memory to do 1029 * the bmap then... So, this is important to do. 1030 */ 1031 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1032 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1033 } 1034 1035 /* 1036 * Set the *dirty* buffer range based upon the VM system dirty pages. 1037 */ 1038 vfs_setdirty(bp); 1039 1040 /* 1041 * We need to do this here to satisfy the vnode_pager and the 1042 * pageout daemon, so that it thinks that the pages have been 1043 * "cleaned". Note that since the pages are in a delayed write 1044 * buffer -- the VFS layer "will" see that the pages get written 1045 * out on the next sync, or perhaps the cluster will be completed. 1046 */ 1047 vfs_clean_pages(bp); 1048 bqrelse(bp); 1049 1050 /* 1051 * Wakeup the buffer flushing daemon if we have a lot of dirty 1052 * buffers (midpoint between our recovery point and our stall 1053 * point). 1054 */ 1055 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1056 1057 /* 1058 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1059 * due to the softdep code. 1060 */ 1061 } 1062 1063 /* 1064 * bdirty: 1065 * 1066 * Turn buffer into delayed write request. We must clear BIO_READ and 1067 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1068 * itself to properly update it in the dirty/clean lists. We mark it 1069 * B_DONE to ensure that any asynchronization of the buffer properly 1070 * clears B_DONE ( else a panic will occur later ). 1071 * 1072 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1073 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1074 * should only be called if the buffer is known-good. 1075 * 1076 * Since the buffer is not on a queue, we do not update the numfreebuffers 1077 * count. 1078 * 1079 * Must be called at splbio(). 1080 * The buffer must be on QUEUE_NONE. 1081 */ 1082 void 1083 bdirty(bp) 1084 struct buf *bp; 1085 { 1086 KASSERT(bp->b_qindex == QUEUE_NONE, 1087 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1088 bp->b_flags &= ~(B_RELBUF); 1089 bp->b_iocmd = BIO_WRITE; 1090 1091 if ((bp->b_flags & B_DELWRI) == 0) { 1092 bp->b_flags |= B_DONE | B_DELWRI; 1093 reassignbuf(bp, bp->b_vp); 1094 atomic_add_int(&numdirtybuffers, 1); 1095 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1096 } 1097 } 1098 1099 /* 1100 * bundirty: 1101 * 1102 * Clear B_DELWRI for buffer. 1103 * 1104 * Since the buffer is not on a queue, we do not update the numfreebuffers 1105 * count. 1106 * 1107 * Must be called at splbio(). 1108 * The buffer must be on QUEUE_NONE. 1109 */ 1110 1111 void 1112 bundirty(bp) 1113 struct buf *bp; 1114 { 1115 KASSERT(bp->b_qindex == QUEUE_NONE, 1116 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1117 1118 if (bp->b_flags & B_DELWRI) { 1119 bp->b_flags &= ~B_DELWRI; 1120 reassignbuf(bp, bp->b_vp); 1121 atomic_subtract_int(&numdirtybuffers, 1); 1122 numdirtywakeup(lodirtybuffers); 1123 } 1124 /* 1125 * Since it is now being written, we can clear its deferred write flag. 1126 */ 1127 bp->b_flags &= ~B_DEFERRED; 1128 } 1129 1130 /* 1131 * bawrite: 1132 * 1133 * Asynchronous write. Start output on a buffer, but do not wait for 1134 * it to complete. The buffer is released when the output completes. 1135 * 1136 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1137 * B_INVAL buffers. Not us. 1138 */ 1139 void 1140 bawrite(struct buf * bp) 1141 { 1142 bp->b_flags |= B_ASYNC; 1143 (void) BUF_WRITE(bp); 1144 } 1145 1146 /* 1147 * bwillwrite: 1148 * 1149 * Called prior to the locking of any vnodes when we are expecting to 1150 * write. We do not want to starve the buffer cache with too many 1151 * dirty buffers so we block here. By blocking prior to the locking 1152 * of any vnodes we attempt to avoid the situation where a locked vnode 1153 * prevents the various system daemons from flushing related buffers. 1154 */ 1155 1156 void 1157 bwillwrite(void) 1158 { 1159 if (numdirtybuffers >= hidirtybuffers) { 1160 int s; 1161 1162 mtx_lock(&Giant); 1163 s = splbio(); 1164 mtx_lock(&nblock); 1165 while (numdirtybuffers >= hidirtybuffers) { 1166 bd_wakeup(1); 1167 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1168 msleep(&needsbuffer, &nblock, 1169 (PRIBIO + 4), "flswai", 0); 1170 } 1171 splx(s); 1172 mtx_unlock(&nblock); 1173 mtx_unlock(&Giant); 1174 } 1175 } 1176 1177 /* 1178 * Return true if we have too many dirty buffers. 1179 */ 1180 int 1181 buf_dirty_count_severe(void) 1182 { 1183 return(numdirtybuffers >= hidirtybuffers); 1184 } 1185 1186 /* 1187 * brelse: 1188 * 1189 * Release a busy buffer and, if requested, free its resources. The 1190 * buffer will be stashed in the appropriate bufqueue[] allowing it 1191 * to be accessed later as a cache entity or reused for other purposes. 1192 */ 1193 void 1194 brelse(struct buf * bp) 1195 { 1196 int s; 1197 1198 GIANT_REQUIRED; 1199 1200 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1201 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1202 1203 s = splbio(); 1204 1205 if (bp->b_flags & B_LOCKED) 1206 bp->b_ioflags &= ~BIO_ERROR; 1207 1208 if (bp->b_iocmd == BIO_WRITE && 1209 (bp->b_ioflags & BIO_ERROR) && 1210 !(bp->b_flags & B_INVAL)) { 1211 /* 1212 * Failed write, redirty. Must clear BIO_ERROR to prevent 1213 * pages from being scrapped. If B_INVAL is set then 1214 * this case is not run and the next case is run to 1215 * destroy the buffer. B_INVAL can occur if the buffer 1216 * is outside the range supported by the underlying device. 1217 */ 1218 bp->b_ioflags &= ~BIO_ERROR; 1219 bdirty(bp); 1220 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1221 (bp->b_ioflags & BIO_ERROR) || 1222 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1223 /* 1224 * Either a failed I/O or we were asked to free or not 1225 * cache the buffer. 1226 */ 1227 bp->b_flags |= B_INVAL; 1228 if (LIST_FIRST(&bp->b_dep) != NULL) 1229 buf_deallocate(bp); 1230 if (bp->b_flags & B_DELWRI) { 1231 atomic_subtract_int(&numdirtybuffers, 1); 1232 numdirtywakeup(lodirtybuffers); 1233 } 1234 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1235 if ((bp->b_flags & B_VMIO) == 0) { 1236 if (bp->b_bufsize) 1237 allocbuf(bp, 0); 1238 if (bp->b_vp) 1239 brelvp(bp); 1240 } 1241 } 1242 1243 /* 1244 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1245 * is called with B_DELWRI set, the underlying pages may wind up 1246 * getting freed causing a previous write (bdwrite()) to get 'lost' 1247 * because pages associated with a B_DELWRI bp are marked clean. 1248 * 1249 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1250 * if B_DELWRI is set. 1251 * 1252 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1253 * on pages to return pages to the VM page queues. 1254 */ 1255 if (bp->b_flags & B_DELWRI) 1256 bp->b_flags &= ~B_RELBUF; 1257 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1258 bp->b_flags |= B_RELBUF; 1259 1260 /* 1261 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1262 * constituted, not even NFS buffers now. Two flags effect this. If 1263 * B_INVAL, the struct buf is invalidated but the VM object is kept 1264 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1265 * 1266 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1267 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1268 * buffer is also B_INVAL because it hits the re-dirtying code above. 1269 * 1270 * Normally we can do this whether a buffer is B_DELWRI or not. If 1271 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1272 * the commit state and we cannot afford to lose the buffer. If the 1273 * buffer has a background write in progress, we need to keep it 1274 * around to prevent it from being reconstituted and starting a second 1275 * background write. 1276 */ 1277 if ((bp->b_flags & B_VMIO) 1278 && !(bp->b_vp->v_mount != NULL && 1279 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1280 !vn_isdisk(bp->b_vp, NULL) && 1281 (bp->b_flags & B_DELWRI)) 1282 ) { 1283 1284 int i, j, resid; 1285 vm_page_t m; 1286 off_t foff; 1287 vm_pindex_t poff; 1288 vm_object_t obj; 1289 struct vnode *vp; 1290 1291 vp = bp->b_vp; 1292 obj = bp->b_object; 1293 1294 /* 1295 * Get the base offset and length of the buffer. Note that 1296 * in the VMIO case if the buffer block size is not 1297 * page-aligned then b_data pointer may not be page-aligned. 1298 * But our b_pages[] array *IS* page aligned. 1299 * 1300 * block sizes less then DEV_BSIZE (usually 512) are not 1301 * supported due to the page granularity bits (m->valid, 1302 * m->dirty, etc...). 1303 * 1304 * See man buf(9) for more information 1305 */ 1306 resid = bp->b_bufsize; 1307 foff = bp->b_offset; 1308 1309 for (i = 0; i < bp->b_npages; i++) { 1310 int had_bogus = 0; 1311 1312 m = bp->b_pages[i]; 1313 vm_page_lock_queues(); 1314 vm_page_flag_clear(m, PG_ZERO); 1315 vm_page_unlock_queues(); 1316 1317 /* 1318 * If we hit a bogus page, fixup *all* the bogus pages 1319 * now. 1320 */ 1321 if (m == bogus_page) { 1322 poff = OFF_TO_IDX(bp->b_offset); 1323 had_bogus = 1; 1324 1325 for (j = i; j < bp->b_npages; j++) { 1326 vm_page_t mtmp; 1327 mtmp = bp->b_pages[j]; 1328 if (mtmp == bogus_page) { 1329 mtmp = vm_page_lookup(obj, poff + j); 1330 if (!mtmp) { 1331 panic("brelse: page missing\n"); 1332 } 1333 bp->b_pages[j] = mtmp; 1334 } 1335 } 1336 1337 if ((bp->b_flags & B_INVAL) == 0) { 1338 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1339 } 1340 m = bp->b_pages[i]; 1341 } 1342 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1343 int poffset = foff & PAGE_MASK; 1344 int presid = resid > (PAGE_SIZE - poffset) ? 1345 (PAGE_SIZE - poffset) : resid; 1346 1347 KASSERT(presid >= 0, ("brelse: extra page")); 1348 vm_page_set_invalid(m, poffset, presid); 1349 if (had_bogus) 1350 printf("avoided corruption bug in bogus_page/brelse code\n"); 1351 } 1352 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1353 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1354 } 1355 1356 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1357 vfs_vmio_release(bp); 1358 1359 } else if (bp->b_flags & B_VMIO) { 1360 1361 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1362 vfs_vmio_release(bp); 1363 } 1364 1365 } 1366 1367 if (bp->b_qindex != QUEUE_NONE) 1368 panic("brelse: free buffer onto another queue???"); 1369 if (BUF_REFCNT(bp) > 1) { 1370 /* do not release to free list */ 1371 BUF_UNLOCK(bp); 1372 splx(s); 1373 return; 1374 } 1375 1376 /* enqueue */ 1377 mtx_lock(&bqlock); 1378 1379 /* buffers with no memory */ 1380 if (bp->b_bufsize == 0) { 1381 bp->b_flags |= B_INVAL; 1382 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1383 if (bp->b_xflags & BX_BKGRDINPROG) 1384 panic("losing buffer 1"); 1385 if (bp->b_kvasize) { 1386 bp->b_qindex = QUEUE_EMPTYKVA; 1387 } else { 1388 bp->b_qindex = QUEUE_EMPTY; 1389 } 1390 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1391 bp->b_dev = NODEV; 1392 /* buffers with junk contents */ 1393 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1394 (bp->b_ioflags & BIO_ERROR)) { 1395 bp->b_flags |= B_INVAL; 1396 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1397 if (bp->b_xflags & BX_BKGRDINPROG) 1398 panic("losing buffer 2"); 1399 bp->b_qindex = QUEUE_CLEAN; 1400 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1401 bp->b_dev = NODEV; 1402 1403 /* buffers that are locked */ 1404 } else if (bp->b_flags & B_LOCKED) { 1405 bp->b_qindex = QUEUE_LOCKED; 1406 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1407 1408 /* remaining buffers */ 1409 } else { 1410 if (bp->b_flags & B_DELWRI) 1411 bp->b_qindex = QUEUE_DIRTY; 1412 else 1413 bp->b_qindex = QUEUE_CLEAN; 1414 if (bp->b_flags & B_AGE) 1415 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1416 else 1417 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1418 } 1419 mtx_unlock(&bqlock); 1420 1421 /* 1422 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1423 * placed the buffer on the correct queue. We must also disassociate 1424 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1425 * find it. 1426 */ 1427 if (bp->b_flags & B_INVAL) { 1428 if (bp->b_flags & B_DELWRI) 1429 bundirty(bp); 1430 if (bp->b_vp) 1431 brelvp(bp); 1432 } 1433 1434 /* 1435 * Fixup numfreebuffers count. The bp is on an appropriate queue 1436 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1437 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1438 * if B_INVAL is set ). 1439 */ 1440 1441 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1442 bufcountwakeup(); 1443 1444 /* 1445 * Something we can maybe free or reuse 1446 */ 1447 if (bp->b_bufsize || bp->b_kvasize) 1448 bufspacewakeup(); 1449 1450 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | 1451 B_DIRECT | B_NOWDRAIN); 1452 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1453 panic("brelse: not dirty"); 1454 /* unlock */ 1455 BUF_UNLOCK(bp); 1456 splx(s); 1457 } 1458 1459 /* 1460 * Release a buffer back to the appropriate queue but do not try to free 1461 * it. The buffer is expected to be used again soon. 1462 * 1463 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1464 * biodone() to requeue an async I/O on completion. It is also used when 1465 * known good buffers need to be requeued but we think we may need the data 1466 * again soon. 1467 * 1468 * XXX we should be able to leave the B_RELBUF hint set on completion. 1469 */ 1470 void 1471 bqrelse(struct buf * bp) 1472 { 1473 int s; 1474 1475 s = splbio(); 1476 1477 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1478 1479 if (bp->b_qindex != QUEUE_NONE) 1480 panic("bqrelse: free buffer onto another queue???"); 1481 if (BUF_REFCNT(bp) > 1) { 1482 /* do not release to free list */ 1483 BUF_UNLOCK(bp); 1484 splx(s); 1485 return; 1486 } 1487 mtx_lock(&bqlock); 1488 if (bp->b_flags & B_LOCKED) { 1489 bp->b_ioflags &= ~BIO_ERROR; 1490 bp->b_qindex = QUEUE_LOCKED; 1491 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1492 /* buffers with stale but valid contents */ 1493 } else if (bp->b_flags & B_DELWRI) { 1494 bp->b_qindex = QUEUE_DIRTY; 1495 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1496 } else if (vm_page_count_severe()) { 1497 /* 1498 * We are too low on memory, we have to try to free the 1499 * buffer (most importantly: the wired pages making up its 1500 * backing store) *now*. 1501 */ 1502 mtx_unlock(&bqlock); 1503 splx(s); 1504 brelse(bp); 1505 return; 1506 } else { 1507 bp->b_qindex = QUEUE_CLEAN; 1508 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1509 } 1510 mtx_unlock(&bqlock); 1511 1512 if ((bp->b_flags & B_LOCKED) == 0 && 1513 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1514 bufcountwakeup(); 1515 } 1516 1517 /* 1518 * Something we can maybe free or reuse. 1519 */ 1520 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1521 bufspacewakeup(); 1522 1523 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1524 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1525 panic("bqrelse: not dirty"); 1526 /* unlock */ 1527 BUF_UNLOCK(bp); 1528 splx(s); 1529 } 1530 1531 /* Give pages used by the bp back to the VM system (where possible) */ 1532 static void 1533 vfs_vmio_release(bp) 1534 struct buf *bp; 1535 { 1536 int i; 1537 vm_page_t m; 1538 1539 GIANT_REQUIRED; 1540 vm_page_lock_queues(); 1541 for (i = 0; i < bp->b_npages; i++) { 1542 m = bp->b_pages[i]; 1543 bp->b_pages[i] = NULL; 1544 /* 1545 * In order to keep page LRU ordering consistent, put 1546 * everything on the inactive queue. 1547 */ 1548 vm_page_unwire(m, 0); 1549 /* 1550 * We don't mess with busy pages, it is 1551 * the responsibility of the process that 1552 * busied the pages to deal with them. 1553 */ 1554 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1555 continue; 1556 1557 if (m->wire_count == 0) { 1558 vm_page_flag_clear(m, PG_ZERO); 1559 /* 1560 * Might as well free the page if we can and it has 1561 * no valid data. We also free the page if the 1562 * buffer was used for direct I/O 1563 */ 1564 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1565 m->hold_count == 0) { 1566 vm_page_busy(m); 1567 pmap_remove_all(m); 1568 vm_page_free(m); 1569 } else if (bp->b_flags & B_DIRECT) { 1570 vm_page_try_to_free(m); 1571 } else if (vm_page_count_severe()) { 1572 vm_page_try_to_cache(m); 1573 } 1574 } 1575 } 1576 vm_page_unlock_queues(); 1577 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1578 1579 if (bp->b_bufsize) { 1580 bufspacewakeup(); 1581 bp->b_bufsize = 0; 1582 } 1583 bp->b_npages = 0; 1584 bp->b_flags &= ~B_VMIO; 1585 if (bp->b_vp) 1586 brelvp(bp); 1587 } 1588 1589 /* 1590 * Check to see if a block at a particular lbn is available for a clustered 1591 * write. 1592 */ 1593 static int 1594 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1595 { 1596 struct buf *bpa; 1597 int match; 1598 1599 match = 0; 1600 1601 /* If the buf isn't in core skip it */ 1602 if ((bpa = gbincore(vp, lblkno)) == NULL) 1603 return (0); 1604 1605 /* If the buf is busy we don't want to wait for it */ 1606 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1607 return (0); 1608 1609 /* Only cluster with valid clusterable delayed write buffers */ 1610 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1611 (B_DELWRI | B_CLUSTEROK)) 1612 goto done; 1613 1614 if (bpa->b_bufsize != size) 1615 goto done; 1616 1617 /* 1618 * Check to see if it is in the expected place on disk and that the 1619 * block has been mapped. 1620 */ 1621 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1622 match = 1; 1623 done: 1624 BUF_UNLOCK(bpa); 1625 return (match); 1626 } 1627 1628 /* 1629 * vfs_bio_awrite: 1630 * 1631 * Implement clustered async writes for clearing out B_DELWRI buffers. 1632 * This is much better then the old way of writing only one buffer at 1633 * a time. Note that we may not be presented with the buffers in the 1634 * correct order, so we search for the cluster in both directions. 1635 */ 1636 int 1637 vfs_bio_awrite(struct buf * bp) 1638 { 1639 int i; 1640 int j; 1641 daddr_t lblkno = bp->b_lblkno; 1642 struct vnode *vp = bp->b_vp; 1643 int s; 1644 int ncl; 1645 int nwritten; 1646 int size; 1647 int maxcl; 1648 1649 s = splbio(); 1650 /* 1651 * right now we support clustered writing only to regular files. If 1652 * we find a clusterable block we could be in the middle of a cluster 1653 * rather then at the beginning. 1654 */ 1655 if ((vp->v_type == VREG) && 1656 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1657 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1658 1659 size = vp->v_mount->mnt_stat.f_iosize; 1660 maxcl = MAXPHYS / size; 1661 1662 VI_LOCK(vp); 1663 for (i = 1; i < maxcl; i++) 1664 if (vfs_bio_clcheck(vp, size, lblkno + i, 1665 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1666 break; 1667 1668 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1669 if (vfs_bio_clcheck(vp, size, lblkno - j, 1670 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1671 break; 1672 1673 VI_UNLOCK(vp); 1674 --j; 1675 ncl = i + j; 1676 /* 1677 * this is a possible cluster write 1678 */ 1679 if (ncl != 1) { 1680 BUF_UNLOCK(bp); 1681 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1682 splx(s); 1683 return nwritten; 1684 } 1685 } 1686 1687 bremfree(bp); 1688 bp->b_flags |= B_ASYNC; 1689 1690 splx(s); 1691 /* 1692 * default (old) behavior, writing out only one block 1693 * 1694 * XXX returns b_bufsize instead of b_bcount for nwritten? 1695 */ 1696 nwritten = bp->b_bufsize; 1697 (void) BUF_WRITE(bp); 1698 1699 return nwritten; 1700 } 1701 1702 /* 1703 * getnewbuf: 1704 * 1705 * Find and initialize a new buffer header, freeing up existing buffers 1706 * in the bufqueues as necessary. The new buffer is returned locked. 1707 * 1708 * Important: B_INVAL is not set. If the caller wishes to throw the 1709 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1710 * 1711 * We block if: 1712 * We have insufficient buffer headers 1713 * We have insufficient buffer space 1714 * buffer_map is too fragmented ( space reservation fails ) 1715 * If we have to flush dirty buffers ( but we try to avoid this ) 1716 * 1717 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1718 * Instead we ask the buf daemon to do it for us. We attempt to 1719 * avoid piecemeal wakeups of the pageout daemon. 1720 */ 1721 1722 static struct buf * 1723 getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1724 { 1725 struct buf *bp; 1726 struct buf *nbp; 1727 int defrag = 0; 1728 int nqindex; 1729 static int flushingbufs; 1730 1731 GIANT_REQUIRED; 1732 1733 /* 1734 * We can't afford to block since we might be holding a vnode lock, 1735 * which may prevent system daemons from running. We deal with 1736 * low-memory situations by proactively returning memory and running 1737 * async I/O rather then sync I/O. 1738 */ 1739 1740 atomic_add_int(&getnewbufcalls, 1); 1741 atomic_subtract_int(&getnewbufrestarts, 1); 1742 restart: 1743 atomic_add_int(&getnewbufrestarts, 1); 1744 1745 /* 1746 * Setup for scan. If we do not have enough free buffers, 1747 * we setup a degenerate case that immediately fails. Note 1748 * that if we are specially marked process, we are allowed to 1749 * dip into our reserves. 1750 * 1751 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1752 * 1753 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1754 * However, there are a number of cases (defragging, reusing, ...) 1755 * where we cannot backup. 1756 */ 1757 mtx_lock(&bqlock); 1758 nqindex = QUEUE_EMPTYKVA; 1759 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1760 1761 if (nbp == NULL) { 1762 /* 1763 * If no EMPTYKVA buffers and we are either 1764 * defragging or reusing, locate a CLEAN buffer 1765 * to free or reuse. If bufspace useage is low 1766 * skip this step so we can allocate a new buffer. 1767 */ 1768 if (defrag || bufspace >= lobufspace) { 1769 nqindex = QUEUE_CLEAN; 1770 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1771 } 1772 1773 /* 1774 * If we could not find or were not allowed to reuse a 1775 * CLEAN buffer, check to see if it is ok to use an EMPTY 1776 * buffer. We can only use an EMPTY buffer if allocating 1777 * its KVA would not otherwise run us out of buffer space. 1778 */ 1779 if (nbp == NULL && defrag == 0 && 1780 bufspace + maxsize < hibufspace) { 1781 nqindex = QUEUE_EMPTY; 1782 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1783 } 1784 } 1785 1786 /* 1787 * Run scan, possibly freeing data and/or kva mappings on the fly 1788 * depending. 1789 */ 1790 1791 while ((bp = nbp) != NULL) { 1792 int qindex = nqindex; 1793 1794 /* 1795 * Calculate next bp ( we can only use it if we do not block 1796 * or do other fancy things ). 1797 */ 1798 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1799 switch(qindex) { 1800 case QUEUE_EMPTY: 1801 nqindex = QUEUE_EMPTYKVA; 1802 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1803 break; 1804 /* FALLTHROUGH */ 1805 case QUEUE_EMPTYKVA: 1806 nqindex = QUEUE_CLEAN; 1807 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1808 break; 1809 /* FALLTHROUGH */ 1810 case QUEUE_CLEAN: 1811 /* 1812 * nbp is NULL. 1813 */ 1814 break; 1815 } 1816 } 1817 1818 /* 1819 * Sanity Checks 1820 */ 1821 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1822 1823 /* 1824 * Note: we no longer distinguish between VMIO and non-VMIO 1825 * buffers. 1826 */ 1827 1828 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1829 1830 /* 1831 * If we are defragging then we need a buffer with 1832 * b_kvasize != 0. XXX this situation should no longer 1833 * occur, if defrag is non-zero the buffer's b_kvasize 1834 * should also be non-zero at this point. XXX 1835 */ 1836 if (defrag && bp->b_kvasize == 0) { 1837 printf("Warning: defrag empty buffer %p\n", bp); 1838 continue; 1839 } 1840 1841 /* 1842 * Start freeing the bp. This is somewhat involved. nbp 1843 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1844 */ 1845 1846 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1847 panic("getnewbuf: locked buf"); 1848 bremfreel(bp); 1849 mtx_unlock(&bqlock); 1850 1851 if (qindex == QUEUE_CLEAN) { 1852 if (bp->b_flags & B_VMIO) { 1853 bp->b_flags &= ~B_ASYNC; 1854 vfs_vmio_release(bp); 1855 } 1856 if (bp->b_vp) 1857 brelvp(bp); 1858 } 1859 1860 /* 1861 * NOTE: nbp is now entirely invalid. We can only restart 1862 * the scan from this point on. 1863 * 1864 * Get the rest of the buffer freed up. b_kva* is still 1865 * valid after this operation. 1866 */ 1867 1868 if (bp->b_rcred != NOCRED) { 1869 crfree(bp->b_rcred); 1870 bp->b_rcred = NOCRED; 1871 } 1872 if (bp->b_wcred != NOCRED) { 1873 crfree(bp->b_wcred); 1874 bp->b_wcred = NOCRED; 1875 } 1876 if (LIST_FIRST(&bp->b_dep) != NULL) 1877 buf_deallocate(bp); 1878 if (bp->b_xflags & BX_BKGRDINPROG) 1879 panic("losing buffer 3"); 1880 1881 if (bp->b_bufsize) 1882 allocbuf(bp, 0); 1883 1884 bp->b_flags = 0; 1885 bp->b_ioflags = 0; 1886 bp->b_xflags = 0; 1887 bp->b_vflags = 0; 1888 bp->b_dev = NODEV; 1889 bp->b_vp = NULL; 1890 bp->b_blkno = bp->b_lblkno = 0; 1891 bp->b_offset = NOOFFSET; 1892 bp->b_iodone = 0; 1893 bp->b_error = 0; 1894 bp->b_resid = 0; 1895 bp->b_bcount = 0; 1896 bp->b_npages = 0; 1897 bp->b_dirtyoff = bp->b_dirtyend = 0; 1898 bp->b_magic = B_MAGIC_BIO; 1899 bp->b_op = &buf_ops_bio; 1900 bp->b_object = NULL; 1901 1902 LIST_INIT(&bp->b_dep); 1903 1904 /* 1905 * If we are defragging then free the buffer. 1906 */ 1907 if (defrag) { 1908 bp->b_flags |= B_INVAL; 1909 bfreekva(bp); 1910 brelse(bp); 1911 defrag = 0; 1912 goto restart; 1913 } 1914 1915 /* 1916 * If we are overcomitted then recover the buffer and its 1917 * KVM space. This occurs in rare situations when multiple 1918 * processes are blocked in getnewbuf() or allocbuf(). 1919 */ 1920 if (bufspace >= hibufspace) 1921 flushingbufs = 1; 1922 if (flushingbufs && bp->b_kvasize != 0) { 1923 bp->b_flags |= B_INVAL; 1924 bfreekva(bp); 1925 brelse(bp); 1926 goto restart; 1927 } 1928 if (bufspace < lobufspace) 1929 flushingbufs = 0; 1930 break; 1931 } 1932 1933 /* 1934 * If we exhausted our list, sleep as appropriate. We may have to 1935 * wakeup various daemons and write out some dirty buffers. 1936 * 1937 * Generally we are sleeping due to insufficient buffer space. 1938 */ 1939 1940 if (bp == NULL) { 1941 int flags; 1942 char *waitmsg; 1943 1944 mtx_unlock(&bqlock); 1945 if (defrag) { 1946 flags = VFS_BIO_NEED_BUFSPACE; 1947 waitmsg = "nbufkv"; 1948 } else if (bufspace >= hibufspace) { 1949 waitmsg = "nbufbs"; 1950 flags = VFS_BIO_NEED_BUFSPACE; 1951 } else { 1952 waitmsg = "newbuf"; 1953 flags = VFS_BIO_NEED_ANY; 1954 } 1955 1956 bd_speedup(); /* heeeelp */ 1957 1958 mtx_lock(&nblock); 1959 needsbuffer |= flags; 1960 while (needsbuffer & flags) { 1961 if (msleep(&needsbuffer, &nblock, 1962 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1963 mtx_unlock(&nblock); 1964 return (NULL); 1965 } 1966 } 1967 mtx_unlock(&nblock); 1968 } else { 1969 /* 1970 * We finally have a valid bp. We aren't quite out of the 1971 * woods, we still have to reserve kva space. In order 1972 * to keep fragmentation sane we only allocate kva in 1973 * BKVASIZE chunks. 1974 */ 1975 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1976 1977 if (maxsize != bp->b_kvasize) { 1978 vm_offset_t addr = 0; 1979 1980 bfreekva(bp); 1981 1982 if (vm_map_findspace(buffer_map, 1983 vm_map_min(buffer_map), maxsize, &addr)) { 1984 /* 1985 * Uh oh. Buffer map is to fragmented. We 1986 * must defragment the map. 1987 */ 1988 atomic_add_int(&bufdefragcnt, 1); 1989 defrag = 1; 1990 bp->b_flags |= B_INVAL; 1991 brelse(bp); 1992 goto restart; 1993 } 1994 if (addr) { 1995 vm_map_insert(buffer_map, NULL, 0, 1996 addr, addr + maxsize, 1997 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1998 1999 bp->b_kvabase = (caddr_t) addr; 2000 bp->b_kvasize = maxsize; 2001 atomic_add_int(&bufspace, bp->b_kvasize); 2002 atomic_add_int(&bufreusecnt, 1); 2003 } 2004 } 2005 bp->b_data = bp->b_kvabase; 2006 } 2007 return(bp); 2008 } 2009 2010 /* 2011 * buf_daemon: 2012 * 2013 * buffer flushing daemon. Buffers are normally flushed by the 2014 * update daemon but if it cannot keep up this process starts to 2015 * take the load in an attempt to prevent getnewbuf() from blocking. 2016 */ 2017 2018 static struct proc *bufdaemonproc; 2019 2020 static struct kproc_desc buf_kp = { 2021 "bufdaemon", 2022 buf_daemon, 2023 &bufdaemonproc 2024 }; 2025 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2026 2027 static void 2028 buf_daemon() 2029 { 2030 int s; 2031 2032 mtx_lock(&Giant); 2033 2034 /* 2035 * This process needs to be suspended prior to shutdown sync. 2036 */ 2037 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2038 SHUTDOWN_PRI_LAST); 2039 2040 /* 2041 * This process is allowed to take the buffer cache to the limit 2042 */ 2043 s = splbio(); 2044 mtx_lock(&bdlock); 2045 2046 for (;;) { 2047 bd_request = 0; 2048 mtx_unlock(&bdlock); 2049 2050 kthread_suspend_check(bufdaemonproc); 2051 2052 /* 2053 * Do the flush. Limit the amount of in-transit I/O we 2054 * allow to build up, otherwise we would completely saturate 2055 * the I/O system. Wakeup any waiting processes before we 2056 * normally would so they can run in parallel with our drain. 2057 */ 2058 while (numdirtybuffers > lodirtybuffers) { 2059 if (flushbufqueues(0) == 0) { 2060 /* 2061 * Could not find any buffers without rollback 2062 * dependencies, so just write the first one 2063 * in the hopes of eventually making progress. 2064 */ 2065 flushbufqueues(1); 2066 break; 2067 } 2068 waitrunningbufspace(); 2069 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2070 } 2071 2072 /* 2073 * Only clear bd_request if we have reached our low water 2074 * mark. The buf_daemon normally waits 1 second and 2075 * then incrementally flushes any dirty buffers that have 2076 * built up, within reason. 2077 * 2078 * If we were unable to hit our low water mark and couldn't 2079 * find any flushable buffers, we sleep half a second. 2080 * Otherwise we loop immediately. 2081 */ 2082 mtx_lock(&bdlock); 2083 if (numdirtybuffers <= lodirtybuffers) { 2084 /* 2085 * We reached our low water mark, reset the 2086 * request and sleep until we are needed again. 2087 * The sleep is just so the suspend code works. 2088 */ 2089 bd_request = 0; 2090 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2091 } else { 2092 /* 2093 * We couldn't find any flushable dirty buffers but 2094 * still have too many dirty buffers, we 2095 * have to sleep and try again. (rare) 2096 */ 2097 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2098 } 2099 } 2100 } 2101 2102 /* 2103 * flushbufqueues: 2104 * 2105 * Try to flush a buffer in the dirty queue. We must be careful to 2106 * free up B_INVAL buffers instead of write them, which NFS is 2107 * particularly sensitive to. 2108 */ 2109 int flushwithdeps = 0; 2110 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2111 0, "Number of buffers flushed with dependecies that require rollbacks"); 2112 static int 2113 flushbufqueues(int flushdeps) 2114 { 2115 struct thread *td = curthread; 2116 struct vnode *vp; 2117 struct buf *bp; 2118 int hasdeps; 2119 2120 mtx_lock(&bqlock); 2121 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2122 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2123 continue; 2124 KASSERT((bp->b_flags & B_DELWRI), 2125 ("unexpected clean buffer %p", bp)); 2126 if ((bp->b_xflags & BX_BKGRDINPROG) != 0) { 2127 BUF_UNLOCK(bp); 2128 continue; 2129 } 2130 if (bp->b_flags & B_INVAL) { 2131 bremfreel(bp); 2132 mtx_unlock(&bqlock); 2133 brelse(bp); 2134 return (1); 2135 } 2136 2137 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2138 if (flushdeps == 0) { 2139 BUF_UNLOCK(bp); 2140 continue; 2141 } 2142 hasdeps = 1; 2143 } else 2144 hasdeps = 0; 2145 /* 2146 * We must hold the lock on a vnode before writing 2147 * one of its buffers. Otherwise we may confuse, or 2148 * in the case of a snapshot vnode, deadlock the 2149 * system. 2150 * 2151 * The lock order here is the reverse of the normal 2152 * of vnode followed by buf lock. This is ok because 2153 * the NOWAIT will prevent deadlock. 2154 */ 2155 if ((vp = bp->b_vp) == NULL || 2156 vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2157 mtx_unlock(&bqlock); 2158 vfs_bio_awrite(bp); 2159 if (vp != NULL) 2160 VOP_UNLOCK(vp, 0, td); 2161 flushwithdeps += hasdeps; 2162 return (1); 2163 } 2164 BUF_UNLOCK(bp); 2165 } 2166 mtx_unlock(&bqlock); 2167 return (0); 2168 } 2169 2170 /* 2171 * Check to see if a block is currently memory resident. 2172 */ 2173 struct buf * 2174 incore(struct vnode * vp, daddr_t blkno) 2175 { 2176 struct buf *bp; 2177 2178 int s = splbio(); 2179 VI_LOCK(vp); 2180 bp = gbincore(vp, blkno); 2181 VI_UNLOCK(vp); 2182 splx(s); 2183 return (bp); 2184 } 2185 2186 /* 2187 * Returns true if no I/O is needed to access the 2188 * associated VM object. This is like incore except 2189 * it also hunts around in the VM system for the data. 2190 */ 2191 2192 int 2193 inmem(struct vnode * vp, daddr_t blkno) 2194 { 2195 vm_object_t obj; 2196 vm_offset_t toff, tinc, size; 2197 vm_page_t m; 2198 vm_ooffset_t off; 2199 2200 GIANT_REQUIRED; 2201 ASSERT_VOP_LOCKED(vp, "inmem"); 2202 2203 if (incore(vp, blkno)) 2204 return 1; 2205 if (vp->v_mount == NULL) 2206 return 0; 2207 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2208 return 0; 2209 2210 size = PAGE_SIZE; 2211 if (size > vp->v_mount->mnt_stat.f_iosize) 2212 size = vp->v_mount->mnt_stat.f_iosize; 2213 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2214 2215 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2216 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2217 if (!m) 2218 goto notinmem; 2219 tinc = size; 2220 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2221 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2222 if (vm_page_is_valid(m, 2223 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2224 goto notinmem; 2225 } 2226 return 1; 2227 2228 notinmem: 2229 return (0); 2230 } 2231 2232 /* 2233 * vfs_setdirty: 2234 * 2235 * Sets the dirty range for a buffer based on the status of the dirty 2236 * bits in the pages comprising the buffer. 2237 * 2238 * The range is limited to the size of the buffer. 2239 * 2240 * This routine is primarily used by NFS, but is generalized for the 2241 * B_VMIO case. 2242 */ 2243 static void 2244 vfs_setdirty(struct buf *bp) 2245 { 2246 int i; 2247 vm_object_t object; 2248 2249 GIANT_REQUIRED; 2250 /* 2251 * Degenerate case - empty buffer 2252 */ 2253 2254 if (bp->b_bufsize == 0) 2255 return; 2256 2257 /* 2258 * We qualify the scan for modified pages on whether the 2259 * object has been flushed yet. The OBJ_WRITEABLE flag 2260 * is not cleared simply by protecting pages off. 2261 */ 2262 2263 if ((bp->b_flags & B_VMIO) == 0) 2264 return; 2265 2266 object = bp->b_pages[0]->object; 2267 2268 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2269 printf("Warning: object %p writeable but not mightbedirty\n", object); 2270 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2271 printf("Warning: object %p mightbedirty but not writeable\n", object); 2272 2273 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2274 vm_offset_t boffset; 2275 vm_offset_t eoffset; 2276 2277 vm_page_lock_queues(); 2278 /* 2279 * test the pages to see if they have been modified directly 2280 * by users through the VM system. 2281 */ 2282 for (i = 0; i < bp->b_npages; i++) { 2283 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2284 vm_page_test_dirty(bp->b_pages[i]); 2285 } 2286 2287 /* 2288 * Calculate the encompassing dirty range, boffset and eoffset, 2289 * (eoffset - boffset) bytes. 2290 */ 2291 2292 for (i = 0; i < bp->b_npages; i++) { 2293 if (bp->b_pages[i]->dirty) 2294 break; 2295 } 2296 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2297 2298 for (i = bp->b_npages - 1; i >= 0; --i) { 2299 if (bp->b_pages[i]->dirty) { 2300 break; 2301 } 2302 } 2303 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2304 2305 vm_page_unlock_queues(); 2306 /* 2307 * Fit it to the buffer. 2308 */ 2309 2310 if (eoffset > bp->b_bcount) 2311 eoffset = bp->b_bcount; 2312 2313 /* 2314 * If we have a good dirty range, merge with the existing 2315 * dirty range. 2316 */ 2317 2318 if (boffset < eoffset) { 2319 if (bp->b_dirtyoff > boffset) 2320 bp->b_dirtyoff = boffset; 2321 if (bp->b_dirtyend < eoffset) 2322 bp->b_dirtyend = eoffset; 2323 } 2324 } 2325 } 2326 2327 /* 2328 * getblk: 2329 * 2330 * Get a block given a specified block and offset into a file/device. 2331 * The buffers B_DONE bit will be cleared on return, making it almost 2332 * ready for an I/O initiation. B_INVAL may or may not be set on 2333 * return. The caller should clear B_INVAL prior to initiating a 2334 * READ. 2335 * 2336 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2337 * an existing buffer. 2338 * 2339 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2340 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2341 * and then cleared based on the backing VM. If the previous buffer is 2342 * non-0-sized but invalid, B_CACHE will be cleared. 2343 * 2344 * If getblk() must create a new buffer, the new buffer is returned with 2345 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2346 * case it is returned with B_INVAL clear and B_CACHE set based on the 2347 * backing VM. 2348 * 2349 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2350 * B_CACHE bit is clear. 2351 * 2352 * What this means, basically, is that the caller should use B_CACHE to 2353 * determine whether the buffer is fully valid or not and should clear 2354 * B_INVAL prior to issuing a read. If the caller intends to validate 2355 * the buffer by loading its data area with something, the caller needs 2356 * to clear B_INVAL. If the caller does this without issuing an I/O, 2357 * the caller should set B_CACHE ( as an optimization ), else the caller 2358 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2359 * a write attempt or if it was a successfull read. If the caller 2360 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2361 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2362 */ 2363 struct buf * 2364 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2365 int flags) 2366 { 2367 struct buf *bp; 2368 int s; 2369 int error; 2370 ASSERT_VOP_LOCKED(vp, "getblk"); 2371 2372 if (size > MAXBSIZE) 2373 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2374 2375 s = splbio(); 2376 loop: 2377 /* 2378 * Block if we are low on buffers. Certain processes are allowed 2379 * to completely exhaust the buffer cache. 2380 * 2381 * If this check ever becomes a bottleneck it may be better to 2382 * move it into the else, when gbincore() fails. At the moment 2383 * it isn't a problem. 2384 * 2385 * XXX remove if 0 sections (clean this up after its proven) 2386 */ 2387 if (numfreebuffers == 0) { 2388 if (curthread == PCPU_GET(idlethread)) 2389 return NULL; 2390 mtx_lock(&nblock); 2391 needsbuffer |= VFS_BIO_NEED_ANY; 2392 mtx_unlock(&nblock); 2393 } 2394 2395 VI_LOCK(vp); 2396 if ((bp = gbincore(vp, blkno))) { 2397 int lockflags; 2398 /* 2399 * Buffer is in-core. If the buffer is not busy, it must 2400 * be on a queue. 2401 */ 2402 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2403 2404 if (flags & GB_LOCK_NOWAIT) 2405 lockflags |= LK_NOWAIT; 2406 2407 error = BUF_TIMELOCK(bp, lockflags, 2408 VI_MTX(vp), "getblk", slpflag, slptimeo); 2409 2410 /* 2411 * If we slept and got the lock we have to restart in case 2412 * the buffer changed identities. 2413 */ 2414 if (error == ENOLCK) 2415 goto loop; 2416 /* We timed out or were interrupted. */ 2417 else if (error) 2418 return (NULL); 2419 2420 /* 2421 * The buffer is locked. B_CACHE is cleared if the buffer is 2422 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2423 * and for a VMIO buffer B_CACHE is adjusted according to the 2424 * backing VM cache. 2425 */ 2426 if (bp->b_flags & B_INVAL) 2427 bp->b_flags &= ~B_CACHE; 2428 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2429 bp->b_flags |= B_CACHE; 2430 bremfree(bp); 2431 2432 /* 2433 * check for size inconsistancies for non-VMIO case. 2434 */ 2435 2436 if (bp->b_bcount != size) { 2437 if ((bp->b_flags & B_VMIO) == 0 || 2438 (size > bp->b_kvasize)) { 2439 if (bp->b_flags & B_DELWRI) { 2440 bp->b_flags |= B_NOCACHE; 2441 BUF_WRITE(bp); 2442 } else { 2443 if ((bp->b_flags & B_VMIO) && 2444 (LIST_FIRST(&bp->b_dep) == NULL)) { 2445 bp->b_flags |= B_RELBUF; 2446 brelse(bp); 2447 } else { 2448 bp->b_flags |= B_NOCACHE; 2449 BUF_WRITE(bp); 2450 } 2451 } 2452 goto loop; 2453 } 2454 } 2455 2456 /* 2457 * If the size is inconsistant in the VMIO case, we can resize 2458 * the buffer. This might lead to B_CACHE getting set or 2459 * cleared. If the size has not changed, B_CACHE remains 2460 * unchanged from its previous state. 2461 */ 2462 2463 if (bp->b_bcount != size) 2464 allocbuf(bp, size); 2465 2466 KASSERT(bp->b_offset != NOOFFSET, 2467 ("getblk: no buffer offset")); 2468 2469 /* 2470 * A buffer with B_DELWRI set and B_CACHE clear must 2471 * be committed before we can return the buffer in 2472 * order to prevent the caller from issuing a read 2473 * ( due to B_CACHE not being set ) and overwriting 2474 * it. 2475 * 2476 * Most callers, including NFS and FFS, need this to 2477 * operate properly either because they assume they 2478 * can issue a read if B_CACHE is not set, or because 2479 * ( for example ) an uncached B_DELWRI might loop due 2480 * to softupdates re-dirtying the buffer. In the latter 2481 * case, B_CACHE is set after the first write completes, 2482 * preventing further loops. 2483 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2484 * above while extending the buffer, we cannot allow the 2485 * buffer to remain with B_CACHE set after the write 2486 * completes or it will represent a corrupt state. To 2487 * deal with this we set B_NOCACHE to scrap the buffer 2488 * after the write. 2489 * 2490 * We might be able to do something fancy, like setting 2491 * B_CACHE in bwrite() except if B_DELWRI is already set, 2492 * so the below call doesn't set B_CACHE, but that gets real 2493 * confusing. This is much easier. 2494 */ 2495 2496 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2497 bp->b_flags |= B_NOCACHE; 2498 BUF_WRITE(bp); 2499 goto loop; 2500 } 2501 2502 splx(s); 2503 bp->b_flags &= ~B_DONE; 2504 } else { 2505 int bsize, maxsize, vmio; 2506 off_t offset; 2507 2508 /* 2509 * Buffer is not in-core, create new buffer. The buffer 2510 * returned by getnewbuf() is locked. Note that the returned 2511 * buffer is also considered valid (not marked B_INVAL). 2512 */ 2513 VI_UNLOCK(vp); 2514 if (vn_isdisk(vp, NULL)) 2515 bsize = DEV_BSIZE; 2516 else if (vp->v_mountedhere) 2517 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2518 else if (vp->v_mount) 2519 bsize = vp->v_mount->mnt_stat.f_iosize; 2520 else 2521 bsize = size; 2522 2523 offset = blkno * bsize; 2524 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2525 (vp->v_vflag & VV_OBJBUF); 2526 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2527 maxsize = imax(maxsize, bsize); 2528 2529 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2530 if (slpflag || slptimeo) { 2531 splx(s); 2532 return NULL; 2533 } 2534 goto loop; 2535 } 2536 2537 /* 2538 * This code is used to make sure that a buffer is not 2539 * created while the getnewbuf routine is blocked. 2540 * This can be a problem whether the vnode is locked or not. 2541 * If the buffer is created out from under us, we have to 2542 * throw away the one we just created. There is now window 2543 * race because we are safely running at splbio() from the 2544 * point of the duplicate buffer creation through to here, 2545 * and we've locked the buffer. 2546 * 2547 * Note: this must occur before we associate the buffer 2548 * with the vp especially considering limitations in 2549 * the splay tree implementation when dealing with duplicate 2550 * lblkno's. 2551 */ 2552 VI_LOCK(vp); 2553 if (gbincore(vp, blkno)) { 2554 VI_UNLOCK(vp); 2555 bp->b_flags |= B_INVAL; 2556 brelse(bp); 2557 goto loop; 2558 } 2559 2560 /* 2561 * Insert the buffer into the hash, so that it can 2562 * be found by incore. 2563 */ 2564 bp->b_blkno = bp->b_lblkno = blkno; 2565 bp->b_offset = offset; 2566 2567 bgetvp(vp, bp); 2568 VI_UNLOCK(vp); 2569 2570 /* 2571 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2572 * buffer size starts out as 0, B_CACHE will be set by 2573 * allocbuf() for the VMIO case prior to it testing the 2574 * backing store for validity. 2575 */ 2576 2577 if (vmio) { 2578 bp->b_flags |= B_VMIO; 2579 #if defined(VFS_BIO_DEBUG) 2580 if (vp->v_type != VREG) 2581 printf("getblk: vmioing file type %d???\n", vp->v_type); 2582 #endif 2583 VOP_GETVOBJECT(vp, &bp->b_object); 2584 } else { 2585 bp->b_flags &= ~B_VMIO; 2586 bp->b_object = NULL; 2587 } 2588 2589 allocbuf(bp, size); 2590 2591 splx(s); 2592 bp->b_flags &= ~B_DONE; 2593 } 2594 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2595 return (bp); 2596 } 2597 2598 /* 2599 * Get an empty, disassociated buffer of given size. The buffer is initially 2600 * set to B_INVAL. 2601 */ 2602 struct buf * 2603 geteblk(int size) 2604 { 2605 struct buf *bp; 2606 int s; 2607 int maxsize; 2608 2609 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2610 2611 s = splbio(); 2612 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2613 continue; 2614 splx(s); 2615 allocbuf(bp, size); 2616 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2617 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2618 return (bp); 2619 } 2620 2621 2622 /* 2623 * This code constitutes the buffer memory from either anonymous system 2624 * memory (in the case of non-VMIO operations) or from an associated 2625 * VM object (in the case of VMIO operations). This code is able to 2626 * resize a buffer up or down. 2627 * 2628 * Note that this code is tricky, and has many complications to resolve 2629 * deadlock or inconsistant data situations. Tread lightly!!! 2630 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2631 * the caller. Calling this code willy nilly can result in the loss of data. 2632 * 2633 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2634 * B_CACHE for the non-VMIO case. 2635 */ 2636 2637 int 2638 allocbuf(struct buf *bp, int size) 2639 { 2640 int newbsize, mbsize; 2641 int i; 2642 2643 GIANT_REQUIRED; 2644 2645 if (BUF_REFCNT(bp) == 0) 2646 panic("allocbuf: buffer not busy"); 2647 2648 if (bp->b_kvasize < size) 2649 panic("allocbuf: buffer too small"); 2650 2651 if ((bp->b_flags & B_VMIO) == 0) { 2652 caddr_t origbuf; 2653 int origbufsize; 2654 /* 2655 * Just get anonymous memory from the kernel. Don't 2656 * mess with B_CACHE. 2657 */ 2658 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2659 if (bp->b_flags & B_MALLOC) 2660 newbsize = mbsize; 2661 else 2662 newbsize = round_page(size); 2663 2664 if (newbsize < bp->b_bufsize) { 2665 /* 2666 * malloced buffers are not shrunk 2667 */ 2668 if (bp->b_flags & B_MALLOC) { 2669 if (newbsize) { 2670 bp->b_bcount = size; 2671 } else { 2672 free(bp->b_data, M_BIOBUF); 2673 if (bp->b_bufsize) { 2674 atomic_subtract_int( 2675 &bufmallocspace, 2676 bp->b_bufsize); 2677 bufspacewakeup(); 2678 bp->b_bufsize = 0; 2679 } 2680 bp->b_data = bp->b_kvabase; 2681 bp->b_bcount = 0; 2682 bp->b_flags &= ~B_MALLOC; 2683 } 2684 return 1; 2685 } 2686 vm_hold_free_pages( 2687 bp, 2688 (vm_offset_t) bp->b_data + newbsize, 2689 (vm_offset_t) bp->b_data + bp->b_bufsize); 2690 } else if (newbsize > bp->b_bufsize) { 2691 /* 2692 * We only use malloced memory on the first allocation. 2693 * and revert to page-allocated memory when the buffer 2694 * grows. 2695 */ 2696 /* 2697 * There is a potential smp race here that could lead 2698 * to bufmallocspace slightly passing the max. It 2699 * is probably extremely rare and not worth worrying 2700 * over. 2701 */ 2702 if ( (bufmallocspace < maxbufmallocspace) && 2703 (bp->b_bufsize == 0) && 2704 (mbsize <= PAGE_SIZE/2)) { 2705 2706 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2707 bp->b_bufsize = mbsize; 2708 bp->b_bcount = size; 2709 bp->b_flags |= B_MALLOC; 2710 atomic_add_int(&bufmallocspace, mbsize); 2711 return 1; 2712 } 2713 origbuf = NULL; 2714 origbufsize = 0; 2715 /* 2716 * If the buffer is growing on its other-than-first allocation, 2717 * then we revert to the page-allocation scheme. 2718 */ 2719 if (bp->b_flags & B_MALLOC) { 2720 origbuf = bp->b_data; 2721 origbufsize = bp->b_bufsize; 2722 bp->b_data = bp->b_kvabase; 2723 if (bp->b_bufsize) { 2724 atomic_subtract_int(&bufmallocspace, 2725 bp->b_bufsize); 2726 bufspacewakeup(); 2727 bp->b_bufsize = 0; 2728 } 2729 bp->b_flags &= ~B_MALLOC; 2730 newbsize = round_page(newbsize); 2731 } 2732 vm_hold_load_pages( 2733 bp, 2734 (vm_offset_t) bp->b_data + bp->b_bufsize, 2735 (vm_offset_t) bp->b_data + newbsize); 2736 if (origbuf) { 2737 bcopy(origbuf, bp->b_data, origbufsize); 2738 free(origbuf, M_BIOBUF); 2739 } 2740 } 2741 } else { 2742 int desiredpages; 2743 2744 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2745 desiredpages = (size == 0) ? 0 : 2746 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2747 2748 if (bp->b_flags & B_MALLOC) 2749 panic("allocbuf: VMIO buffer can't be malloced"); 2750 /* 2751 * Set B_CACHE initially if buffer is 0 length or will become 2752 * 0-length. 2753 */ 2754 if (size == 0 || bp->b_bufsize == 0) 2755 bp->b_flags |= B_CACHE; 2756 2757 if (newbsize < bp->b_bufsize) { 2758 /* 2759 * DEV_BSIZE aligned new buffer size is less then the 2760 * DEV_BSIZE aligned existing buffer size. Figure out 2761 * if we have to remove any pages. 2762 */ 2763 if (desiredpages < bp->b_npages) { 2764 vm_page_t m; 2765 2766 vm_page_lock_queues(); 2767 for (i = desiredpages; i < bp->b_npages; i++) { 2768 /* 2769 * the page is not freed here -- it 2770 * is the responsibility of 2771 * vnode_pager_setsize 2772 */ 2773 m = bp->b_pages[i]; 2774 KASSERT(m != bogus_page, 2775 ("allocbuf: bogus page found")); 2776 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2777 vm_page_lock_queues(); 2778 2779 bp->b_pages[i] = NULL; 2780 vm_page_unwire(m, 0); 2781 } 2782 vm_page_unlock_queues(); 2783 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2784 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2785 bp->b_npages = desiredpages; 2786 } 2787 } else if (size > bp->b_bcount) { 2788 /* 2789 * We are growing the buffer, possibly in a 2790 * byte-granular fashion. 2791 */ 2792 struct vnode *vp; 2793 vm_object_t obj; 2794 vm_offset_t toff; 2795 vm_offset_t tinc; 2796 2797 /* 2798 * Step 1, bring in the VM pages from the object, 2799 * allocating them if necessary. We must clear 2800 * B_CACHE if these pages are not valid for the 2801 * range covered by the buffer. 2802 */ 2803 2804 vp = bp->b_vp; 2805 obj = bp->b_object; 2806 2807 while (bp->b_npages < desiredpages) { 2808 vm_page_t m; 2809 vm_pindex_t pi; 2810 2811 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2812 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2813 /* 2814 * note: must allocate system pages 2815 * since blocking here could intefere 2816 * with paging I/O, no matter which 2817 * process we are. 2818 */ 2819 m = vm_page_alloc(obj, pi, 2820 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2821 if (m == NULL) { 2822 atomic_add_int(&vm_pageout_deficit, 2823 desiredpages - bp->b_npages); 2824 VM_WAIT; 2825 } else { 2826 vm_page_lock_queues(); 2827 vm_page_wakeup(m); 2828 vm_page_unlock_queues(); 2829 bp->b_flags &= ~B_CACHE; 2830 bp->b_pages[bp->b_npages] = m; 2831 ++bp->b_npages; 2832 } 2833 continue; 2834 } 2835 2836 /* 2837 * We found a page. If we have to sleep on it, 2838 * retry because it might have gotten freed out 2839 * from under us. 2840 * 2841 * We can only test PG_BUSY here. Blocking on 2842 * m->busy might lead to a deadlock: 2843 * 2844 * vm_fault->getpages->cluster_read->allocbuf 2845 * 2846 */ 2847 vm_page_lock_queues(); 2848 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2849 continue; 2850 2851 /* 2852 * We have a good page. Should we wakeup the 2853 * page daemon? 2854 */ 2855 if ((curproc != pageproc) && 2856 ((m->queue - m->pc) == PQ_CACHE) && 2857 ((cnt.v_free_count + cnt.v_cache_count) < 2858 (cnt.v_free_min + cnt.v_cache_min))) { 2859 pagedaemon_wakeup(); 2860 } 2861 vm_page_flag_clear(m, PG_ZERO); 2862 vm_page_wire(m); 2863 vm_page_unlock_queues(); 2864 bp->b_pages[bp->b_npages] = m; 2865 ++bp->b_npages; 2866 } 2867 2868 /* 2869 * Step 2. We've loaded the pages into the buffer, 2870 * we have to figure out if we can still have B_CACHE 2871 * set. Note that B_CACHE is set according to the 2872 * byte-granular range ( bcount and size ), new the 2873 * aligned range ( newbsize ). 2874 * 2875 * The VM test is against m->valid, which is DEV_BSIZE 2876 * aligned. Needless to say, the validity of the data 2877 * needs to also be DEV_BSIZE aligned. Note that this 2878 * fails with NFS if the server or some other client 2879 * extends the file's EOF. If our buffer is resized, 2880 * B_CACHE may remain set! XXX 2881 */ 2882 2883 toff = bp->b_bcount; 2884 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2885 2886 while ((bp->b_flags & B_CACHE) && toff < size) { 2887 vm_pindex_t pi; 2888 2889 if (tinc > (size - toff)) 2890 tinc = size - toff; 2891 2892 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2893 PAGE_SHIFT; 2894 2895 vfs_buf_test_cache( 2896 bp, 2897 bp->b_offset, 2898 toff, 2899 tinc, 2900 bp->b_pages[pi] 2901 ); 2902 toff += tinc; 2903 tinc = PAGE_SIZE; 2904 } 2905 2906 /* 2907 * Step 3, fixup the KVM pmap. Remember that 2908 * bp->b_data is relative to bp->b_offset, but 2909 * bp->b_offset may be offset into the first page. 2910 */ 2911 2912 bp->b_data = (caddr_t) 2913 trunc_page((vm_offset_t)bp->b_data); 2914 pmap_qenter( 2915 (vm_offset_t)bp->b_data, 2916 bp->b_pages, 2917 bp->b_npages 2918 ); 2919 2920 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2921 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2922 } 2923 } 2924 if (newbsize < bp->b_bufsize) 2925 bufspacewakeup(); 2926 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2927 bp->b_bcount = size; /* requested buffer size */ 2928 return 1; 2929 } 2930 2931 void 2932 biodone(struct bio *bp) 2933 { 2934 mtx_lock(&bdonelock); 2935 bp->bio_flags |= BIO_DONE; 2936 if (bp->bio_done == NULL) 2937 wakeup(bp); 2938 mtx_unlock(&bdonelock); 2939 if (bp->bio_done != NULL) 2940 bp->bio_done(bp); 2941 } 2942 2943 /* 2944 * Wait for a BIO to finish. 2945 * 2946 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2947 * case is not yet clear. 2948 */ 2949 int 2950 biowait(struct bio *bp, const char *wchan) 2951 { 2952 2953 mtx_lock(&bdonelock); 2954 while ((bp->bio_flags & BIO_DONE) == 0) 2955 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 2956 mtx_unlock(&bdonelock); 2957 if (bp->bio_error != 0) 2958 return (bp->bio_error); 2959 if (!(bp->bio_flags & BIO_ERROR)) 2960 return (0); 2961 return (EIO); 2962 } 2963 2964 void 2965 biofinish(struct bio *bp, struct devstat *stat, int error) 2966 { 2967 2968 if (error) { 2969 bp->bio_error = error; 2970 bp->bio_flags |= BIO_ERROR; 2971 } 2972 if (stat != NULL) 2973 devstat_end_transaction_bio(stat, bp); 2974 biodone(bp); 2975 } 2976 2977 void 2978 bioq_init(struct bio_queue_head *head) 2979 { 2980 TAILQ_INIT(&head->queue); 2981 head->last_pblkno = 0; 2982 head->insert_point = NULL; 2983 head->switch_point = NULL; 2984 } 2985 2986 void 2987 bioq_remove(struct bio_queue_head *head, struct bio *bp) 2988 { 2989 if (bp == head->switch_point) 2990 head->switch_point = TAILQ_NEXT(bp, bio_queue); 2991 if (bp == head->insert_point) { 2992 head->insert_point = TAILQ_PREV(bp, bio_queue, bio_queue); 2993 if (head->insert_point == NULL) 2994 head->last_pblkno = 0; 2995 } else if (bp == TAILQ_FIRST(&head->queue)) 2996 head->last_pblkno = bp->bio_pblkno; 2997 TAILQ_REMOVE(&head->queue, bp, bio_queue); 2998 if (TAILQ_FIRST(&head->queue) == head->switch_point) 2999 head->switch_point = NULL; 3000 } 3001 3002 /* 3003 * bufwait: 3004 * 3005 * Wait for buffer I/O completion, returning error status. The buffer 3006 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3007 * error and cleared. 3008 */ 3009 int 3010 bufwait(register struct buf * bp) 3011 { 3012 int s; 3013 3014 s = splbio(); 3015 if (bp->b_iocmd == BIO_READ) 3016 bwait(bp, PRIBIO, "biord"); 3017 else 3018 bwait(bp, PRIBIO, "biowr"); 3019 splx(s); 3020 if (bp->b_flags & B_EINTR) { 3021 bp->b_flags &= ~B_EINTR; 3022 return (EINTR); 3023 } 3024 if (bp->b_ioflags & BIO_ERROR) { 3025 return (bp->b_error ? bp->b_error : EIO); 3026 } else { 3027 return (0); 3028 } 3029 } 3030 3031 /* 3032 * Call back function from struct bio back up to struct buf. 3033 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 3034 */ 3035 void 3036 bufdonebio(struct bio *bp) 3037 { 3038 bufdone(bp->bio_caller2); 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 */ 3060 void 3061 bufdone(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 defined(VFS_BIO_DEBUG) 3123 if (obj->paging_in_progress < bp->b_npages) { 3124 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3125 obj->paging_in_progress, bp->b_npages); 3126 } 3127 #endif 3128 3129 /* 3130 * Set B_CACHE if the op was a normal read and no error 3131 * occured. B_CACHE is set for writes in the b*write() 3132 * routines. 3133 */ 3134 iosize = bp->b_bcount - bp->b_resid; 3135 if (bp->b_iocmd == BIO_READ && 3136 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3137 !(bp->b_ioflags & BIO_ERROR)) { 3138 bp->b_flags |= B_CACHE; 3139 } 3140 vm_page_lock_queues(); 3141 for (i = 0; i < bp->b_npages; i++) { 3142 int bogusflag = 0; 3143 int resid; 3144 3145 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3146 if (resid > iosize) 3147 resid = iosize; 3148 3149 /* 3150 * cleanup bogus pages, restoring the originals 3151 */ 3152 m = bp->b_pages[i]; 3153 if (m == bogus_page) { 3154 bogusflag = 1; 3155 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3156 if (m == NULL) 3157 panic("biodone: page disappeared!"); 3158 bp->b_pages[i] = m; 3159 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3160 } 3161 #if defined(VFS_BIO_DEBUG) 3162 if (OFF_TO_IDX(foff) != m->pindex) { 3163 printf( 3164 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3165 (intmax_t)foff, (uintmax_t)m->pindex); 3166 } 3167 #endif 3168 3169 /* 3170 * In the write case, the valid and clean bits are 3171 * already changed correctly ( see bdwrite() ), so we 3172 * only need to do this here in the read case. 3173 */ 3174 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3175 vfs_page_set_valid(bp, foff, i, m); 3176 } 3177 vm_page_flag_clear(m, PG_ZERO); 3178 3179 /* 3180 * when debugging new filesystems or buffer I/O methods, this 3181 * is the most common error that pops up. if you see this, you 3182 * have not set the page busy flag correctly!!! 3183 */ 3184 if (m->busy == 0) { 3185 printf("biodone: page busy < 0, " 3186 "pindex: %d, foff: 0x(%x,%x), " 3187 "resid: %d, index: %d\n", 3188 (int) m->pindex, (int)(foff >> 32), 3189 (int) foff & 0xffffffff, resid, i); 3190 if (!vn_isdisk(vp, NULL)) 3191 printf(" iosize: %ld, lblkno: %jd, flags: 0x%x, npages: %d\n", 3192 bp->b_vp->v_mount->mnt_stat.f_iosize, 3193 (intmax_t) bp->b_lblkno, 3194 bp->b_flags, bp->b_npages); 3195 else 3196 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3197 (intmax_t) bp->b_lblkno, 3198 bp->b_flags, bp->b_npages); 3199 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 3200 m->valid, m->dirty, m->wire_count); 3201 panic("biodone: page busy < 0\n"); 3202 } 3203 vm_page_io_finish(m); 3204 vm_object_pip_subtract(obj, 1); 3205 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3206 iosize -= resid; 3207 } 3208 vm_page_unlock_queues(); 3209 if (obj) 3210 vm_object_pip_wakeupn(obj, 0); 3211 } 3212 3213 /* 3214 * For asynchronous completions, release the buffer now. The brelse 3215 * will do a wakeup there if necessary - so no need to do a wakeup 3216 * here in the async case. The sync case always needs to do a wakeup. 3217 */ 3218 3219 if (bp->b_flags & B_ASYNC) { 3220 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3221 brelse(bp); 3222 else 3223 bqrelse(bp); 3224 } else { 3225 bdone(bp); 3226 } 3227 splx(s); 3228 } 3229 3230 /* 3231 * This routine is called in lieu of iodone in the case of 3232 * incomplete I/O. This keeps the busy status for pages 3233 * consistant. 3234 */ 3235 void 3236 vfs_unbusy_pages(struct buf * bp) 3237 { 3238 int i; 3239 3240 GIANT_REQUIRED; 3241 3242 runningbufwakeup(bp); 3243 if (bp->b_flags & B_VMIO) { 3244 vm_object_t obj; 3245 3246 obj = bp->b_object; 3247 vm_page_lock_queues(); 3248 for (i = 0; i < bp->b_npages; i++) { 3249 vm_page_t m = bp->b_pages[i]; 3250 3251 if (m == bogus_page) { 3252 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3253 if (!m) { 3254 panic("vfs_unbusy_pages: page missing\n"); 3255 } 3256 bp->b_pages[i] = m; 3257 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3258 } 3259 vm_object_pip_subtract(obj, 1); 3260 vm_page_flag_clear(m, PG_ZERO); 3261 vm_page_io_finish(m); 3262 } 3263 vm_page_unlock_queues(); 3264 vm_object_pip_wakeupn(obj, 0); 3265 } 3266 } 3267 3268 /* 3269 * vfs_page_set_valid: 3270 * 3271 * Set the valid bits in a page based on the supplied offset. The 3272 * range is restricted to the buffer's size. 3273 * 3274 * This routine is typically called after a read completes. 3275 */ 3276 static void 3277 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3278 { 3279 vm_ooffset_t soff, eoff; 3280 3281 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3282 /* 3283 * Start and end offsets in buffer. eoff - soff may not cross a 3284 * page boundry or cross the end of the buffer. The end of the 3285 * buffer, in this case, is our file EOF, not the allocation size 3286 * of the buffer. 3287 */ 3288 soff = off; 3289 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3290 if (eoff > bp->b_offset + bp->b_bcount) 3291 eoff = bp->b_offset + bp->b_bcount; 3292 3293 /* 3294 * Set valid range. This is typically the entire buffer and thus the 3295 * entire page. 3296 */ 3297 if (eoff > soff) { 3298 vm_page_set_validclean( 3299 m, 3300 (vm_offset_t) (soff & PAGE_MASK), 3301 (vm_offset_t) (eoff - soff) 3302 ); 3303 } 3304 } 3305 3306 /* 3307 * This routine is called before a device strategy routine. 3308 * It is used to tell the VM system that paging I/O is in 3309 * progress, and treat the pages associated with the buffer 3310 * almost as being PG_BUSY. Also the object paging_in_progress 3311 * flag is handled to make sure that the object doesn't become 3312 * inconsistant. 3313 * 3314 * Since I/O has not been initiated yet, certain buffer flags 3315 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3316 * and should be ignored. 3317 */ 3318 void 3319 vfs_busy_pages(struct buf * bp, int clear_modify) 3320 { 3321 int i, bogus; 3322 3323 if (bp->b_flags & B_VMIO) { 3324 vm_object_t obj; 3325 vm_ooffset_t foff; 3326 3327 obj = bp->b_object; 3328 foff = bp->b_offset; 3329 KASSERT(bp->b_offset != NOOFFSET, 3330 ("vfs_busy_pages: no buffer offset")); 3331 vfs_setdirty(bp); 3332 retry: 3333 vm_page_lock_queues(); 3334 for (i = 0; i < bp->b_npages; i++) { 3335 vm_page_t m = bp->b_pages[i]; 3336 3337 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3338 goto retry; 3339 } 3340 bogus = 0; 3341 for (i = 0; i < bp->b_npages; i++) { 3342 vm_page_t m = bp->b_pages[i]; 3343 3344 vm_page_flag_clear(m, PG_ZERO); 3345 if ((bp->b_flags & B_CLUSTER) == 0) { 3346 vm_object_pip_add(obj, 1); 3347 vm_page_io_start(m); 3348 } 3349 /* 3350 * When readying a buffer for a read ( i.e 3351 * clear_modify == 0 ), it is important to do 3352 * bogus_page replacement for valid pages in 3353 * partially instantiated buffers. Partially 3354 * instantiated buffers can, in turn, occur when 3355 * reconstituting a buffer from its VM backing store 3356 * base. We only have to do this if B_CACHE is 3357 * clear ( which causes the I/O to occur in the 3358 * first place ). The replacement prevents the read 3359 * I/O from overwriting potentially dirty VM-backed 3360 * pages. XXX bogus page replacement is, uh, bogus. 3361 * It may not work properly with small-block devices. 3362 * We need to find a better way. 3363 */ 3364 pmap_remove_all(m); 3365 if (clear_modify) 3366 vfs_page_set_valid(bp, foff, i, m); 3367 else if (m->valid == VM_PAGE_BITS_ALL && 3368 (bp->b_flags & B_CACHE) == 0) { 3369 bp->b_pages[i] = bogus_page; 3370 bogus++; 3371 } 3372 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3373 } 3374 vm_page_unlock_queues(); 3375 if (bogus) 3376 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3377 } 3378 } 3379 3380 /* 3381 * Tell the VM system that the pages associated with this buffer 3382 * are clean. This is used for delayed writes where the data is 3383 * going to go to disk eventually without additional VM intevention. 3384 * 3385 * Note that while we only really need to clean through to b_bcount, we 3386 * just go ahead and clean through to b_bufsize. 3387 */ 3388 static void 3389 vfs_clean_pages(struct buf * bp) 3390 { 3391 int i; 3392 3393 GIANT_REQUIRED; 3394 3395 if (bp->b_flags & B_VMIO) { 3396 vm_ooffset_t foff; 3397 3398 foff = bp->b_offset; 3399 KASSERT(bp->b_offset != NOOFFSET, 3400 ("vfs_clean_pages: no buffer offset")); 3401 vm_page_lock_queues(); 3402 for (i = 0; i < bp->b_npages; i++) { 3403 vm_page_t m = bp->b_pages[i]; 3404 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3405 vm_ooffset_t eoff = noff; 3406 3407 if (eoff > bp->b_offset + bp->b_bufsize) 3408 eoff = bp->b_offset + bp->b_bufsize; 3409 vfs_page_set_valid(bp, foff, i, m); 3410 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3411 foff = noff; 3412 } 3413 vm_page_unlock_queues(); 3414 } 3415 } 3416 3417 /* 3418 * vfs_bio_set_validclean: 3419 * 3420 * Set the range within the buffer to valid and clean. The range is 3421 * relative to the beginning of the buffer, b_offset. Note that b_offset 3422 * itself may be offset from the beginning of the first page. 3423 * 3424 */ 3425 3426 void 3427 vfs_bio_set_validclean(struct buf *bp, int base, int size) 3428 { 3429 if (bp->b_flags & B_VMIO) { 3430 int i; 3431 int n; 3432 3433 /* 3434 * Fixup base to be relative to beginning of first page. 3435 * Set initial n to be the maximum number of bytes in the 3436 * first page that can be validated. 3437 */ 3438 3439 base += (bp->b_offset & PAGE_MASK); 3440 n = PAGE_SIZE - (base & PAGE_MASK); 3441 3442 vm_page_lock_queues(); 3443 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3444 vm_page_t m = bp->b_pages[i]; 3445 3446 if (n > size) 3447 n = size; 3448 3449 vm_page_set_validclean(m, base & PAGE_MASK, n); 3450 base += n; 3451 size -= n; 3452 n = PAGE_SIZE; 3453 } 3454 vm_page_unlock_queues(); 3455 } 3456 } 3457 3458 /* 3459 * vfs_bio_clrbuf: 3460 * 3461 * clear a buffer. This routine essentially fakes an I/O, so we need 3462 * to clear BIO_ERROR and B_INVAL. 3463 * 3464 * Note that while we only theoretically need to clear through b_bcount, 3465 * we go ahead and clear through b_bufsize. 3466 */ 3467 3468 void 3469 vfs_bio_clrbuf(struct buf *bp) 3470 { 3471 int i, mask = 0; 3472 caddr_t sa, ea; 3473 3474 GIANT_REQUIRED; 3475 3476 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3477 bp->b_flags &= ~B_INVAL; 3478 bp->b_ioflags &= ~BIO_ERROR; 3479 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3480 (bp->b_offset & PAGE_MASK) == 0) { 3481 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3482 if ((bp->b_pages[0]->valid & mask) == mask) { 3483 bp->b_resid = 0; 3484 return; 3485 } 3486 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3487 ((bp->b_pages[0]->valid & mask) == 0)) { 3488 bzero(bp->b_data, bp->b_bufsize); 3489 bp->b_pages[0]->valid |= mask; 3490 bp->b_resid = 0; 3491 return; 3492 } 3493 } 3494 ea = sa = bp->b_data; 3495 for(i=0;i<bp->b_npages;i++,sa=ea) { 3496 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3497 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3498 ea = (caddr_t)(vm_offset_t)ulmin( 3499 (u_long)(vm_offset_t)ea, 3500 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3501 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3502 if ((bp->b_pages[i]->valid & mask) == mask) 3503 continue; 3504 if ((bp->b_pages[i]->valid & mask) == 0) { 3505 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3506 bzero(sa, ea - sa); 3507 } 3508 } else { 3509 for (; sa < ea; sa += DEV_BSIZE, j++) { 3510 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3511 (bp->b_pages[i]->valid & (1<<j)) == 0) 3512 bzero(sa, DEV_BSIZE); 3513 } 3514 } 3515 bp->b_pages[i]->valid |= mask; 3516 vm_page_lock_queues(); 3517 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3518 vm_page_unlock_queues(); 3519 } 3520 bp->b_resid = 0; 3521 } else { 3522 clrbuf(bp); 3523 } 3524 } 3525 3526 /* 3527 * vm_hold_load_pages and vm_hold_free_pages get pages into 3528 * a buffers address space. The pages are anonymous and are 3529 * not associated with a file object. 3530 */ 3531 static void 3532 vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3533 { 3534 vm_offset_t pg; 3535 vm_page_t p; 3536 int index; 3537 3538 GIANT_REQUIRED; 3539 3540 to = round_page(to); 3541 from = round_page(from); 3542 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3543 3544 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3545 tryagain: 3546 /* 3547 * note: must allocate system pages since blocking here 3548 * could intefere with paging I/O, no matter which 3549 * process we are. 3550 */ 3551 vm_object_lock(kernel_object); 3552 p = vm_page_alloc(kernel_object, 3553 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3554 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3555 vm_object_unlock(kernel_object); 3556 if (!p) { 3557 atomic_add_int(&vm_pageout_deficit, 3558 (to - pg) >> PAGE_SHIFT); 3559 VM_WAIT; 3560 goto tryagain; 3561 } 3562 vm_page_lock_queues(); 3563 p->valid = VM_PAGE_BITS_ALL; 3564 vm_page_unlock_queues(); 3565 pmap_qenter(pg, &p, 1); 3566 bp->b_pages[index] = p; 3567 vm_page_lock_queues(); 3568 vm_page_wakeup(p); 3569 vm_page_unlock_queues(); 3570 } 3571 bp->b_npages = index; 3572 } 3573 3574 /* Return pages associated with this buf to the vm system */ 3575 static void 3576 vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3577 { 3578 vm_offset_t pg; 3579 vm_page_t p; 3580 int index, newnpages; 3581 3582 GIANT_REQUIRED; 3583 3584 from = round_page(from); 3585 to = round_page(to); 3586 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3587 3588 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3589 p = bp->b_pages[index]; 3590 if (p && (index < bp->b_npages)) { 3591 if (p->busy) { 3592 printf( 3593 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3594 (intmax_t)bp->b_blkno, 3595 (intmax_t)bp->b_lblkno); 3596 } 3597 bp->b_pages[index] = NULL; 3598 pmap_qremove(pg, 1); 3599 vm_page_lock_queues(); 3600 vm_page_busy(p); 3601 vm_page_unwire(p, 0); 3602 vm_page_free(p); 3603 vm_page_unlock_queues(); 3604 } 3605 } 3606 bp->b_npages = newnpages; 3607 } 3608 3609 /* 3610 * Map an IO request into kernel virtual address space. 3611 * 3612 * All requests are (re)mapped into kernel VA space. 3613 * Notice that we use b_bufsize for the size of the buffer 3614 * to be mapped. b_bcount might be modified by the driver. 3615 * 3616 * Note that even if the caller determines that the address space should 3617 * be valid, a race or a smaller-file mapped into a larger space may 3618 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3619 * check the return value. 3620 */ 3621 int 3622 vmapbuf(struct buf *bp) 3623 { 3624 caddr_t addr, kva; 3625 vm_paddr_t pa; 3626 int pidx, i; 3627 struct vm_page *m; 3628 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3629 3630 GIANT_REQUIRED; 3631 3632 if ((bp->b_flags & B_PHYS) == 0) 3633 panic("vmapbuf"); 3634 3635 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3636 addr < bp->b_data + bp->b_bufsize; 3637 addr += PAGE_SIZE, pidx++) { 3638 /* 3639 * Do the vm_fault if needed; do the copy-on-write thing 3640 * when reading stuff off device into memory. 3641 * 3642 * NOTE! Must use pmap_extract() because addr may be in 3643 * the userland address space, and kextract is only guarenteed 3644 * to work for the kernland address space (see: sparc64 port). 3645 */ 3646 retry: 3647 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data, 3648 (bp->b_iocmd == BIO_READ) ? 3649 (VM_PROT_READ|VM_PROT_WRITE) : VM_PROT_READ); 3650 if (i < 0) { 3651 printf("vmapbuf: warning, bad user address during I/O\n"); 3652 vm_page_lock_queues(); 3653 for (i = 0; i < pidx; ++i) { 3654 vm_page_unhold(bp->b_pages[i]); 3655 bp->b_pages[i] = NULL; 3656 } 3657 vm_page_unlock_queues(); 3658 return(-1); 3659 } 3660 pa = trunc_page(pmap_extract(pmap, (vm_offset_t) addr)); 3661 if (pa == 0) { 3662 printf("vmapbuf: warning, race against user address during I/O"); 3663 goto retry; 3664 } 3665 m = PHYS_TO_VM_PAGE(pa); 3666 vm_page_lock_queues(); 3667 vm_page_hold(m); 3668 vm_page_unlock_queues(); 3669 bp->b_pages[pidx] = m; 3670 } 3671 if (pidx > btoc(MAXPHYS)) 3672 panic("vmapbuf: mapped more than MAXPHYS"); 3673 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3674 3675 kva = bp->b_saveaddr; 3676 bp->b_npages = pidx; 3677 bp->b_saveaddr = bp->b_data; 3678 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3679 return(0); 3680 } 3681 3682 /* 3683 * Free the io map PTEs associated with this IO operation. 3684 * We also invalidate the TLB entries and restore the original b_addr. 3685 */ 3686 void 3687 vunmapbuf(struct buf *bp) 3688 { 3689 int pidx; 3690 int npages; 3691 3692 GIANT_REQUIRED; 3693 3694 if ((bp->b_flags & B_PHYS) == 0) 3695 panic("vunmapbuf"); 3696 3697 npages = bp->b_npages; 3698 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3699 npages); 3700 vm_page_lock_queues(); 3701 for (pidx = 0; pidx < npages; pidx++) 3702 vm_page_unhold(bp->b_pages[pidx]); 3703 vm_page_unlock_queues(); 3704 3705 bp->b_data = bp->b_saveaddr; 3706 } 3707 3708 void 3709 bdone(struct buf *bp) 3710 { 3711 mtx_lock(&bdonelock); 3712 bp->b_flags |= B_DONE; 3713 wakeup(bp); 3714 mtx_unlock(&bdonelock); 3715 } 3716 3717 void 3718 bwait(struct buf *bp, u_char pri, const char *wchan) 3719 { 3720 mtx_lock(&bdonelock); 3721 while ((bp->b_flags & B_DONE) == 0) 3722 msleep(bp, &bdonelock, pri, wchan, 0); 3723 mtx_unlock(&bdonelock); 3724 } 3725 3726 #include "opt_ddb.h" 3727 #ifdef DDB 3728 #include <ddb/ddb.h> 3729 3730 /* DDB command to show buffer data */ 3731 DB_SHOW_COMMAND(buffer, db_show_buffer) 3732 { 3733 /* get args */ 3734 struct buf *bp = (struct buf *)addr; 3735 3736 if (!have_addr) { 3737 db_printf("usage: show buffer <addr>\n"); 3738 return; 3739 } 3740 3741 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3742 db_printf( 3743 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3744 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd, b_pblkno = %jd\n", 3745 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3746 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3747 (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno); 3748 if (bp->b_npages) { 3749 int i; 3750 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3751 for (i = 0; i < bp->b_npages; i++) { 3752 vm_page_t m; 3753 m = bp->b_pages[i]; 3754 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3755 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3756 if ((i + 1) < bp->b_npages) 3757 db_printf(","); 3758 } 3759 db_printf("\n"); 3760 } 3761 } 3762 #endif /* DDB */ 3763