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: src/sys/kern/vfs_bio.c,v 1.242.2.20 2003/05/28 18:38:10 alc Exp $ 15 * $DragonFly: src/sys/kern/vfs_bio.c,v 1.32 2004/11/09 17:36:41 dillon Exp $ 16 */ 17 18 /* 19 * this file contains a new buffer I/O scheme implementing a coherent 20 * VM object and buffer cache scheme. Pains have been taken to make 21 * sure that the performance degradation associated with schemes such 22 * as this is not realized. 23 * 24 * Author: John S. Dyson 25 * Significant help during the development and debugging phases 26 * had been provided by David Greenman, also of the FreeBSD core team. 27 * 28 * see man buf(9) for more info. 29 */ 30 31 #include <sys/param.h> 32 #include <sys/systm.h> 33 #include <sys/buf.h> 34 #include <sys/conf.h> 35 #include <sys/eventhandler.h> 36 #include <sys/lock.h> 37 #include <sys/malloc.h> 38 #include <sys/mount.h> 39 #include <sys/kernel.h> 40 #include <sys/kthread.h> 41 #include <sys/proc.h> 42 #include <sys/reboot.h> 43 #include <sys/resourcevar.h> 44 #include <sys/sysctl.h> 45 #include <sys/vmmeter.h> 46 #include <sys/vnode.h> 47 #include <sys/proc.h> 48 #include <vm/vm.h> 49 #include <vm/vm_param.h> 50 #include <vm/vm_kern.h> 51 #include <vm/vm_pageout.h> 52 #include <vm/vm_page.h> 53 #include <vm/vm_object.h> 54 #include <vm/vm_extern.h> 55 #include <vm/vm_map.h> 56 57 #include <sys/buf2.h> 58 #include <sys/thread2.h> 59 #include <vm/vm_page2.h> 60 61 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 62 63 struct bio_ops bioops; /* I/O operation notification */ 64 65 struct buf *buf; /* buffer header pool */ 66 struct swqueue bswlist; 67 68 static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 69 vm_offset_t to); 70 static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 71 vm_offset_t to); 72 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 73 int pageno, vm_page_t m); 74 static void vfs_clean_pages(struct buf * bp); 75 static void vfs_setdirty(struct buf *bp); 76 static void vfs_vmio_release(struct buf *bp); 77 static void vfs_backgroundwritedone(struct buf *bp); 78 static int flushbufqueues(void); 79 80 static int bd_request; 81 82 static void buf_daemon (void); 83 /* 84 * bogus page -- for I/O to/from partially complete buffers 85 * this is a temporary solution to the problem, but it is not 86 * really that bad. it would be better to split the buffer 87 * for input in the case of buffers partially already in memory, 88 * but the code is intricate enough already. 89 */ 90 vm_page_t bogus_page; 91 int vmiodirenable = TRUE; 92 int runningbufspace; 93 struct lwkt_token buftimetoken; /* Interlock on setting prio and timo */ 94 95 static vm_offset_t bogus_offset; 96 97 static int bufspace, maxbufspace, 98 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace; 99 static int bufreusecnt, bufdefragcnt, buffreekvacnt; 100 static int needsbuffer; 101 static int lorunningspace, hirunningspace, runningbufreq; 102 static int numdirtybuffers, lodirtybuffers, hidirtybuffers; 103 static int numfreebuffers, lofreebuffers, hifreebuffers; 104 static int getnewbufcalls; 105 static int getnewbufrestarts; 106 107 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, 108 &numdirtybuffers, 0, ""); 109 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, 110 &lodirtybuffers, 0, ""); 111 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, 112 &hidirtybuffers, 0, ""); 113 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, 114 &numfreebuffers, 0, ""); 115 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, 116 &lofreebuffers, 0, ""); 117 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, 118 &hifreebuffers, 0, ""); 119 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, 120 &runningbufspace, 0, ""); 121 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, 122 &lorunningspace, 0, ""); 123 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, 124 &hirunningspace, 0, ""); 125 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, 126 &maxbufspace, 0, ""); 127 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, 128 &hibufspace, 0, ""); 129 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, 130 &lobufspace, 0, ""); 131 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, 132 &bufspace, 0, ""); 133 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, 134 &maxbufmallocspace, 0, ""); 135 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, 136 &bufmallocspace, 0, ""); 137 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, 138 &getnewbufcalls, 0, ""); 139 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, 140 &getnewbufrestarts, 0, ""); 141 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, 142 &vmiodirenable, 0, ""); 143 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, 144 &bufdefragcnt, 0, ""); 145 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, 146 &buffreekvacnt, 0, ""); 147 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, 148 &bufreusecnt, 0, ""); 149 150 /* 151 * Disable background writes for now. There appear to be races in the 152 * flags tests and locking operations as well as races in the completion 153 * code modifying the original bp (origbp) without holding a lock, assuming 154 * splbio protection when there might not be splbio protection. 155 */ 156 static int dobkgrdwrite = 0; 157 SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 158 "Do background writes (honoring the BV_BKGRDWRITE flag)?"); 159 160 static int bufhashmask; 161 static int bufhashshift; 162 static LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash; 163 struct bqueues bufqueues[BUFFER_QUEUES] = { { 0 } }; 164 char *buf_wmesg = BUF_WMESG; 165 166 extern int vm_swap_size; 167 168 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 169 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 170 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 171 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 172 173 /* 174 * Buffer hash table code. Note that the logical block scans linearly, which 175 * gives us some L1 cache locality. 176 */ 177 178 static __inline 179 struct bufhashhdr * 180 bufhash(struct vnode *vnp, daddr_t bn) 181 { 182 u_int64_t hashkey64; 183 int hashkey; 184 185 /* 186 * A variation on the Fibonacci hash that Knuth credits to 187 * R. W. Floyd, see Knuth's _Art of Computer Programming, 188 * Volume 3 / Sorting and Searching_ 189 * 190 * We reduce the argument to 32 bits before doing the hash to 191 * avoid the need for a slow 64x64 multiply on 32 bit platforms. 192 * 193 * sizeof(struct vnode) is 168 on i386, so toss some of the lower 194 * bits of the vnode address to reduce the key range, which 195 * improves the distribution of keys across buckets. 196 * 197 * The file system cylinder group blocks are very heavily 198 * used. They are located at invervals of fbg, which is 199 * on the order of 89 to 94 * 2^10, depending on other 200 * filesystem parameters, for a 16k block size. Smaller block 201 * sizes will reduce fpg approximately proportionally. This 202 * will cause the cylinder group index to be hashed using the 203 * lower bits of the hash multiplier, which will not distribute 204 * the keys as uniformly in a classic Fibonacci hash where a 205 * relatively small number of the upper bits of the result 206 * are used. Using 2^16 as a close-enough approximation to 207 * fpg, split the hash multiplier in half, with the upper 16 208 * bits being the inverse of the golden ratio, and the lower 209 * 16 bits being a fraction between 1/3 and 3/7 (closer to 210 * 3/7 in this case), that gives good experimental results. 211 */ 212 hashkey64 = ((u_int64_t)(uintptr_t)vnp >> 3) + (u_int64_t)bn; 213 hashkey = (((u_int32_t)(hashkey64 + (hashkey64 >> 32)) * 0x9E376DB1u) >> 214 bufhashshift) & bufhashmask; 215 return(&bufhashtbl[hashkey]); 216 } 217 218 /* 219 * numdirtywakeup: 220 * 221 * If someone is blocked due to there being too many dirty buffers, 222 * and numdirtybuffers is now reasonable, wake them up. 223 */ 224 225 static __inline void 226 numdirtywakeup(int level) 227 { 228 if (numdirtybuffers <= level) { 229 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 230 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 231 wakeup(&needsbuffer); 232 } 233 } 234 } 235 236 /* 237 * bufspacewakeup: 238 * 239 * Called when buffer space is potentially available for recovery. 240 * getnewbuf() will block on this flag when it is unable to free 241 * sufficient buffer space. Buffer space becomes recoverable when 242 * bp's get placed back in the queues. 243 */ 244 245 static __inline void 246 bufspacewakeup(void) 247 { 248 /* 249 * If someone is waiting for BUF space, wake them up. Even 250 * though we haven't freed the kva space yet, the waiting 251 * process will be able to now. 252 */ 253 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 254 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 255 wakeup(&needsbuffer); 256 } 257 } 258 259 /* 260 * runningbufwakeup() - in-progress I/O accounting. 261 * 262 */ 263 static __inline void 264 runningbufwakeup(struct buf *bp) 265 { 266 if (bp->b_runningbufspace) { 267 runningbufspace -= bp->b_runningbufspace; 268 bp->b_runningbufspace = 0; 269 if (runningbufreq && runningbufspace <= lorunningspace) { 270 runningbufreq = 0; 271 wakeup(&runningbufreq); 272 } 273 } 274 } 275 276 /* 277 * bufcountwakeup: 278 * 279 * Called when a buffer has been added to one of the free queues to 280 * account for the buffer and to wakeup anyone waiting for free buffers. 281 * This typically occurs when large amounts of metadata are being handled 282 * by the buffer cache ( else buffer space runs out first, usually ). 283 */ 284 285 static __inline void 286 bufcountwakeup(void) 287 { 288 ++numfreebuffers; 289 if (needsbuffer) { 290 needsbuffer &= ~VFS_BIO_NEED_ANY; 291 if (numfreebuffers >= hifreebuffers) 292 needsbuffer &= ~VFS_BIO_NEED_FREE; 293 wakeup(&needsbuffer); 294 } 295 } 296 297 /* 298 * waitrunningbufspace() 299 * 300 * runningbufspace is a measure of the amount of I/O currently 301 * running. This routine is used in async-write situations to 302 * prevent creating huge backups of pending writes to a device. 303 * Only asynchronous writes are governed by this function. 304 * 305 * Reads will adjust runningbufspace, but will not block based on it. 306 * The read load has a side effect of reducing the allowed write load. 307 * 308 * This does NOT turn an async write into a sync write. It waits 309 * for earlier writes to complete and generally returns before the 310 * caller's write has reached the device. 311 */ 312 static __inline void 313 waitrunningbufspace(void) 314 { 315 while (runningbufspace > hirunningspace) { 316 int s; 317 318 s = splbio(); /* fix race against interrupt/biodone() */ 319 ++runningbufreq; 320 tsleep(&runningbufreq, 0, "wdrain", 0); 321 splx(s); 322 } 323 } 324 325 /* 326 * vfs_buf_test_cache: 327 * 328 * Called when a buffer is extended. This function clears the B_CACHE 329 * bit if the newly extended portion of the buffer does not contain 330 * valid data. 331 */ 332 static __inline__ 333 void 334 vfs_buf_test_cache(struct buf *bp, 335 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 336 vm_page_t m) 337 { 338 if (bp->b_flags & B_CACHE) { 339 int base = (foff + off) & PAGE_MASK; 340 if (vm_page_is_valid(m, base, size) == 0) 341 bp->b_flags &= ~B_CACHE; 342 } 343 } 344 345 static __inline__ 346 void 347 bd_wakeup(int dirtybuflevel) 348 { 349 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 350 bd_request = 1; 351 wakeup(&bd_request); 352 } 353 } 354 355 /* 356 * bd_speedup - speedup the buffer cache flushing code 357 */ 358 359 static __inline__ 360 void 361 bd_speedup(void) 362 { 363 bd_wakeup(1); 364 } 365 366 /* 367 * Initialize buffer headers and related structures. 368 */ 369 370 caddr_t 371 bufhashinit(caddr_t vaddr) 372 { 373 /* first, make a null hash table */ 374 bufhashshift = 29; 375 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 376 bufhashshift--; 377 bufhashtbl = (void *)vaddr; 378 vaddr = vaddr + sizeof(*bufhashtbl) * bufhashmask; 379 --bufhashmask; 380 return(vaddr); 381 } 382 383 void 384 bufinit(void) 385 { 386 struct buf *bp; 387 int i; 388 389 TAILQ_INIT(&bswlist); 390 LIST_INIT(&invalhash); 391 lwkt_token_init(&buftimetoken); 392 393 for (i = 0; i <= bufhashmask; i++) 394 LIST_INIT(&bufhashtbl[i]); 395 396 /* next, make a null set of free lists */ 397 for (i = 0; i < BUFFER_QUEUES; i++) 398 TAILQ_INIT(&bufqueues[i]); 399 400 /* finally, initialize each buffer header and stick on empty q */ 401 for (i = 0; i < nbuf; i++) { 402 bp = &buf[i]; 403 bzero(bp, sizeof *bp); 404 bp->b_flags = B_INVAL; /* we're just an empty header */ 405 bp->b_dev = NODEV; 406 bp->b_qindex = QUEUE_EMPTY; 407 bp->b_xflags = 0; 408 xio_init(&bp->b_xio); 409 LIST_INIT(&bp->b_dep); 410 BUF_LOCKINIT(bp); 411 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 412 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 413 } 414 415 /* 416 * maxbufspace is the absolute maximum amount of buffer space we are 417 * allowed to reserve in KVM and in real terms. The absolute maximum 418 * is nominally used by buf_daemon. hibufspace is the nominal maximum 419 * used by most other processes. The differential is required to 420 * ensure that buf_daemon is able to run when other processes might 421 * be blocked waiting for buffer space. 422 * 423 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 424 * this may result in KVM fragmentation which is not handled optimally 425 * by the system. 426 */ 427 maxbufspace = nbuf * BKVASIZE; 428 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 429 lobufspace = hibufspace - MAXBSIZE; 430 431 lorunningspace = 512 * 1024; 432 hirunningspace = 1024 * 1024; 433 434 /* 435 * Limit the amount of malloc memory since it is wired permanently into 436 * the kernel space. Even though this is accounted for in the buffer 437 * allocation, we don't want the malloced region to grow uncontrolled. 438 * The malloc scheme improves memory utilization significantly on average 439 * (small) directories. 440 */ 441 maxbufmallocspace = hibufspace / 20; 442 443 /* 444 * Reduce the chance of a deadlock occuring by limiting the number 445 * of delayed-write dirty buffers we allow to stack up. 446 */ 447 hidirtybuffers = nbuf / 4 + 20; 448 numdirtybuffers = 0; 449 /* 450 * To support extreme low-memory systems, make sure hidirtybuffers cannot 451 * eat up all available buffer space. This occurs when our minimum cannot 452 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 453 * BKVASIZE'd (8K) buffers. 454 */ 455 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 456 hidirtybuffers >>= 1; 457 } 458 lodirtybuffers = hidirtybuffers / 2; 459 460 /* 461 * Try to keep the number of free buffers in the specified range, 462 * and give special processes (e.g. like buf_daemon) access to an 463 * emergency reserve. 464 */ 465 lofreebuffers = nbuf / 18 + 5; 466 hifreebuffers = 2 * lofreebuffers; 467 numfreebuffers = nbuf; 468 469 /* 470 * Maximum number of async ops initiated per buf_daemon loop. This is 471 * somewhat of a hack at the moment, we really need to limit ourselves 472 * based on the number of bytes of I/O in-transit that were initiated 473 * from buf_daemon. 474 */ 475 476 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 477 bogus_page = vm_page_alloc(kernel_object, 478 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 479 VM_ALLOC_NORMAL); 480 vmstats.v_wire_count++; 481 482 } 483 484 /* 485 * bfreekva() - free the kva allocation for a buffer. 486 * 487 * Must be called at splbio() or higher as this is the only locking for 488 * buffer_map. 489 * 490 * Since this call frees up buffer space, we call bufspacewakeup(). 491 */ 492 static void 493 bfreekva(struct buf * bp) 494 { 495 int count; 496 497 if (bp->b_kvasize) { 498 ++buffreekvacnt; 499 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 500 vm_map_lock(buffer_map); 501 bufspace -= bp->b_kvasize; 502 vm_map_delete(buffer_map, 503 (vm_offset_t) bp->b_kvabase, 504 (vm_offset_t) bp->b_kvabase + bp->b_kvasize, 505 &count 506 ); 507 vm_map_unlock(buffer_map); 508 vm_map_entry_release(count); 509 bp->b_kvasize = 0; 510 bufspacewakeup(); 511 } 512 } 513 514 /* 515 * bremfree: 516 * 517 * Remove the buffer from the appropriate free list. 518 */ 519 void 520 bremfree(struct buf * bp) 521 { 522 int s = splbio(); 523 int old_qindex = bp->b_qindex; 524 525 if (bp->b_qindex != QUEUE_NONE) { 526 KASSERT(BUF_REFCNTNB(bp) == 1, 527 ("bremfree: bp %p not locked",bp)); 528 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 529 bp->b_qindex = QUEUE_NONE; 530 } else { 531 if (BUF_REFCNTNB(bp) <= 1) 532 panic("bremfree: removing a buffer not on a queue"); 533 } 534 535 /* 536 * Fixup numfreebuffers count. If the buffer is invalid or not 537 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 538 * the buffer was free and we must decrement numfreebuffers. 539 */ 540 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 541 switch(old_qindex) { 542 case QUEUE_DIRTY: 543 case QUEUE_CLEAN: 544 case QUEUE_EMPTY: 545 case QUEUE_EMPTYKVA: 546 --numfreebuffers; 547 break; 548 default: 549 break; 550 } 551 } 552 splx(s); 553 } 554 555 556 /* 557 * Get a buffer with the specified data. Look in the cache first. We 558 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 559 * is set, the buffer is valid and we do not have to do anything ( see 560 * getblk() ). 561 */ 562 int 563 bread(struct vnode * vp, daddr_t blkno, int size, struct buf ** bpp) 564 { 565 struct buf *bp; 566 567 bp = getblk(vp, blkno, size, 0, 0); 568 *bpp = bp; 569 570 /* if not found in cache, do some I/O */ 571 if ((bp->b_flags & B_CACHE) == 0) { 572 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp)); 573 bp->b_flags |= B_READ; 574 bp->b_flags &= ~(B_ERROR | B_INVAL); 575 vfs_busy_pages(bp, 0); 576 VOP_STRATEGY(vp, bp); 577 return (biowait(bp)); 578 } 579 return (0); 580 } 581 582 /* 583 * Operates like bread, but also starts asynchronous I/O on 584 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior 585 * to initiating I/O . If B_CACHE is set, the buffer is valid 586 * and we do not have to do anything. 587 */ 588 int 589 breadn(struct vnode * vp, daddr_t blkno, int size, daddr_t * rablkno, 590 int *rabsize, int cnt, struct buf ** bpp) 591 { 592 struct buf *bp, *rabp; 593 int i; 594 int rv = 0, readwait = 0; 595 596 *bpp = bp = getblk(vp, blkno, size, 0, 0); 597 598 /* if not found in cache, do some I/O */ 599 if ((bp->b_flags & B_CACHE) == 0) { 600 bp->b_flags |= B_READ; 601 bp->b_flags &= ~(B_ERROR | B_INVAL); 602 vfs_busy_pages(bp, 0); 603 VOP_STRATEGY(vp, bp); 604 ++readwait; 605 } 606 607 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 608 if (inmem(vp, *rablkno)) 609 continue; 610 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 611 612 if ((rabp->b_flags & B_CACHE) == 0) { 613 rabp->b_flags |= B_READ | B_ASYNC; 614 rabp->b_flags &= ~(B_ERROR | B_INVAL); 615 vfs_busy_pages(rabp, 0); 616 BUF_KERNPROC(rabp); 617 VOP_STRATEGY(vp, rabp); 618 } else { 619 brelse(rabp); 620 } 621 } 622 623 if (readwait) { 624 rv = biowait(bp); 625 } 626 return (rv); 627 } 628 629 /* 630 * Write, release buffer on completion. (Done by iodone 631 * if async). Do not bother writing anything if the buffer 632 * is invalid. 633 * 634 * Note that we set B_CACHE here, indicating that buffer is 635 * fully valid and thus cacheable. This is true even of NFS 636 * now so we set it generally. This could be set either here 637 * or in biodone() since the I/O is synchronous. We put it 638 * here. 639 */ 640 int 641 bwrite(struct buf * bp) 642 { 643 int oldflags, s; 644 struct buf *newbp; 645 646 if (bp->b_flags & B_INVAL) { 647 brelse(bp); 648 return (0); 649 } 650 651 oldflags = bp->b_flags; 652 653 if (BUF_REFCNTNB(bp) == 0) 654 panic("bwrite: buffer is not busy???"); 655 s = splbio(); 656 /* 657 * If a background write is already in progress, delay 658 * writing this block if it is asynchronous. Otherwise 659 * wait for the background write to complete. 660 */ 661 if (bp->b_xflags & BX_BKGRDINPROG) { 662 if (bp->b_flags & B_ASYNC) { 663 splx(s); 664 bdwrite(bp); 665 return (0); 666 } 667 bp->b_xflags |= BX_BKGRDWAIT; 668 tsleep(&bp->b_xflags, 0, "biord", 0); 669 if (bp->b_xflags & BX_BKGRDINPROG) 670 panic("bwrite: still writing"); 671 } 672 673 /* Mark the buffer clean */ 674 bundirty(bp); 675 676 /* 677 * If this buffer is marked for background writing and we 678 * do not have to wait for it, make a copy and write the 679 * copy so as to leave this buffer ready for further use. 680 * 681 * This optimization eats a lot of memory. If we have a page 682 * or buffer shortfull we can't do it. 683 */ 684 if (dobkgrdwrite && 685 (bp->b_xflags & BX_BKGRDWRITE) && 686 (bp->b_flags & B_ASYNC) && 687 !vm_page_count_severe() && 688 !buf_dirty_count_severe()) { 689 if (bp->b_flags & B_CALL) 690 panic("bwrite: need chained iodone"); 691 692 /* get a new block */ 693 newbp = geteblk(bp->b_bufsize); 694 695 /* set it to be identical to the old block */ 696 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 697 bgetvp(bp->b_vp, newbp); 698 newbp->b_lblkno = bp->b_lblkno; 699 newbp->b_blkno = bp->b_blkno; 700 newbp->b_offset = bp->b_offset; 701 newbp->b_iodone = vfs_backgroundwritedone; 702 newbp->b_flags |= B_ASYNC | B_CALL; 703 newbp->b_flags &= ~B_INVAL; 704 705 /* move over the dependencies */ 706 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps) 707 (*bioops.io_movedeps)(bp, newbp); 708 709 /* 710 * Initiate write on the copy, release the original to 711 * the B_LOCKED queue so that it cannot go away until 712 * the background write completes. If not locked it could go 713 * away and then be reconstituted while it was being written. 714 * If the reconstituted buffer were written, we could end up 715 * with two background copies being written at the same time. 716 */ 717 bp->b_xflags |= BX_BKGRDINPROG; 718 bp->b_flags |= B_LOCKED; 719 bqrelse(bp); 720 bp = newbp; 721 } 722 723 bp->b_flags &= ~(B_READ | B_DONE | B_ERROR); 724 bp->b_flags |= B_WRITEINPROG | B_CACHE; 725 726 bp->b_vp->v_numoutput++; 727 vfs_busy_pages(bp, 1); 728 729 /* 730 * Normal bwrites pipeline writes 731 */ 732 bp->b_runningbufspace = bp->b_bufsize; 733 runningbufspace += bp->b_runningbufspace; 734 735 splx(s); 736 if (oldflags & B_ASYNC) 737 BUF_KERNPROC(bp); 738 VOP_STRATEGY(bp->b_vp, bp); 739 740 if ((oldflags & B_ASYNC) == 0) { 741 int rtval = biowait(bp); 742 brelse(bp); 743 return (rtval); 744 } else if ((oldflags & B_NOWDRAIN) == 0) { 745 /* 746 * don't allow the async write to saturate the I/O 747 * system. Deadlocks can occur only if a device strategy 748 * routine (like in VN) turns around and issues another 749 * high-level write, in which case B_NOWDRAIN is expected 750 * to be set. Otherwise we will not deadlock here because 751 * we are blocking waiting for I/O that is already in-progress 752 * to complete. 753 */ 754 waitrunningbufspace(); 755 } 756 757 return (0); 758 } 759 760 /* 761 * Complete a background write started from bwrite. 762 */ 763 static void 764 vfs_backgroundwritedone(struct buf *bp) 765 { 766 struct buf *origbp; 767 768 /* 769 * Find the original buffer that we are writing. 770 */ 771 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 772 panic("backgroundwritedone: lost buffer"); 773 /* 774 * Process dependencies then return any unfinished ones. 775 */ 776 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 777 (*bioops.io_complete)(bp); 778 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps) 779 (*bioops.io_movedeps)(bp, origbp); 780 /* 781 * Clear the BX_BKGRDINPROG flag in the original buffer 782 * and awaken it if it is waiting for the write to complete. 783 * If BX_BKGRDINPROG is not set in the original buffer it must 784 * have been released and re-instantiated - which is not legal. 785 */ 786 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), ("backgroundwritedone: lost buffer2")); 787 origbp->b_xflags &= ~BX_BKGRDINPROG; 788 if (origbp->b_xflags & BX_BKGRDWAIT) { 789 origbp->b_xflags &= ~BX_BKGRDWAIT; 790 wakeup(&origbp->b_xflags); 791 } 792 /* 793 * Clear the B_LOCKED flag and remove it from the locked 794 * queue if it currently resides there. 795 */ 796 origbp->b_flags &= ~B_LOCKED; 797 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 798 bremfree(origbp); 799 bqrelse(origbp); 800 } 801 /* 802 * This buffer is marked B_NOCACHE, so when it is released 803 * by biodone, it will be tossed. We mark it with B_READ 804 * to avoid biodone doing a second vwakeup. 805 */ 806 bp->b_flags |= B_NOCACHE | B_READ; 807 bp->b_flags &= ~(B_CACHE | B_CALL | B_DONE); 808 bp->b_iodone = 0; 809 biodone(bp); 810 } 811 812 /* 813 * Delayed write. (Buffer is marked dirty). Do not bother writing 814 * anything if the buffer is marked invalid. 815 * 816 * Note that since the buffer must be completely valid, we can safely 817 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 818 * biodone() in order to prevent getblk from writing the buffer 819 * out synchronously. 820 */ 821 void 822 bdwrite(struct buf *bp) 823 { 824 if (BUF_REFCNTNB(bp) == 0) 825 panic("bdwrite: buffer is not busy"); 826 827 if (bp->b_flags & B_INVAL) { 828 brelse(bp); 829 return; 830 } 831 bdirty(bp); 832 833 /* 834 * Set B_CACHE, indicating that the buffer is fully valid. This is 835 * true even of NFS now. 836 */ 837 bp->b_flags |= B_CACHE; 838 839 /* 840 * This bmap keeps the system from needing to do the bmap later, 841 * perhaps when the system is attempting to do a sync. Since it 842 * is likely that the indirect block -- or whatever other datastructure 843 * that the filesystem needs is still in memory now, it is a good 844 * thing to do this. Note also, that if the pageout daemon is 845 * requesting a sync -- there might not be enough memory to do 846 * the bmap then... So, this is important to do. 847 */ 848 if (bp->b_lblkno == bp->b_blkno) { 849 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 850 } 851 852 /* 853 * Set the *dirty* buffer range based upon the VM system dirty pages. 854 */ 855 vfs_setdirty(bp); 856 857 /* 858 * We need to do this here to satisfy the vnode_pager and the 859 * pageout daemon, so that it thinks that the pages have been 860 * "cleaned". Note that since the pages are in a delayed write 861 * buffer -- the VFS layer "will" see that the pages get written 862 * out on the next sync, or perhaps the cluster will be completed. 863 */ 864 vfs_clean_pages(bp); 865 bqrelse(bp); 866 867 /* 868 * Wakeup the buffer flushing daemon if we have a lot of dirty 869 * buffers (midpoint between our recovery point and our stall 870 * point). 871 */ 872 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 873 874 /* 875 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 876 * due to the softdep code. 877 */ 878 } 879 880 /* 881 * bdirty: 882 * 883 * Turn buffer into delayed write request. We must clear B_READ and 884 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 885 * itself to properly update it in the dirty/clean lists. We mark it 886 * B_DONE to ensure that any asynchronization of the buffer properly 887 * clears B_DONE ( else a panic will occur later ). 888 * 889 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 890 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 891 * should only be called if the buffer is known-good. 892 * 893 * Since the buffer is not on a queue, we do not update the numfreebuffers 894 * count. 895 * 896 * Must be called at splbio(). 897 * The buffer must be on QUEUE_NONE. 898 */ 899 void 900 bdirty(struct buf *bp) 901 { 902 KASSERT(bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 903 bp->b_flags &= ~(B_READ|B_RELBUF); 904 905 if ((bp->b_flags & B_DELWRI) == 0) { 906 bp->b_flags |= B_DONE | B_DELWRI; 907 reassignbuf(bp, bp->b_vp); 908 ++numdirtybuffers; 909 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 910 } 911 } 912 913 /* 914 * bundirty: 915 * 916 * Clear B_DELWRI for buffer. 917 * 918 * Since the buffer is not on a queue, we do not update the numfreebuffers 919 * count. 920 * 921 * Must be called at splbio(). 922 * The buffer must be on QUEUE_NONE. 923 */ 924 925 void 926 bundirty(struct buf *bp) 927 { 928 KASSERT(bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 929 930 if (bp->b_flags & B_DELWRI) { 931 bp->b_flags &= ~B_DELWRI; 932 reassignbuf(bp, bp->b_vp); 933 --numdirtybuffers; 934 numdirtywakeup(lodirtybuffers); 935 } 936 /* 937 * Since it is now being written, we can clear its deferred write flag. 938 */ 939 bp->b_flags &= ~B_DEFERRED; 940 } 941 942 /* 943 * bawrite: 944 * 945 * Asynchronous write. Start output on a buffer, but do not wait for 946 * it to complete. The buffer is released when the output completes. 947 * 948 * bwrite() ( or the VOP routine anyway ) is responsible for handling 949 * B_INVAL buffers. Not us. 950 */ 951 void 952 bawrite(struct buf * bp) 953 { 954 bp->b_flags |= B_ASYNC; 955 (void) VOP_BWRITE(bp->b_vp, bp); 956 } 957 958 /* 959 * bowrite: 960 * 961 * Ordered write. Start output on a buffer, and flag it so that the 962 * device will write it in the order it was queued. The buffer is 963 * released when the output completes. bwrite() ( or the VOP routine 964 * anyway ) is responsible for handling B_INVAL buffers. 965 */ 966 int 967 bowrite(struct buf * bp) 968 { 969 bp->b_flags |= B_ORDERED | B_ASYNC; 970 return (VOP_BWRITE(bp->b_vp, bp)); 971 } 972 973 /* 974 * bwillwrite: 975 * 976 * Called prior to the locking of any vnodes when we are expecting to 977 * write. We do not want to starve the buffer cache with too many 978 * dirty buffers so we block here. By blocking prior to the locking 979 * of any vnodes we attempt to avoid the situation where a locked vnode 980 * prevents the various system daemons from flushing related buffers. 981 */ 982 983 void 984 bwillwrite(void) 985 { 986 if (numdirtybuffers >= hidirtybuffers) { 987 int s; 988 989 s = splbio(); 990 while (numdirtybuffers >= hidirtybuffers) { 991 bd_wakeup(1); 992 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 993 tsleep(&needsbuffer, 0, "flswai", 0); 994 } 995 splx(s); 996 } 997 } 998 999 /* 1000 * Return true if we have too many dirty buffers. 1001 */ 1002 int 1003 buf_dirty_count_severe(void) 1004 { 1005 return(numdirtybuffers >= hidirtybuffers); 1006 } 1007 1008 /* 1009 * brelse: 1010 * 1011 * Release a busy buffer and, if requested, free its resources. The 1012 * buffer will be stashed in the appropriate bufqueue[] allowing it 1013 * to be accessed later as a cache entity or reused for other purposes. 1014 */ 1015 void 1016 brelse(struct buf * bp) 1017 { 1018 int s; 1019 1020 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1021 1022 s = splbio(); 1023 1024 if (bp->b_flags & B_LOCKED) 1025 bp->b_flags &= ~B_ERROR; 1026 1027 if ((bp->b_flags & (B_READ | B_ERROR | B_INVAL)) == B_ERROR) { 1028 /* 1029 * Failed write, redirty. Must clear B_ERROR to prevent 1030 * pages from being scrapped. If B_INVAL is set then 1031 * this case is not run and the next case is run to 1032 * destroy the buffer. B_INVAL can occur if the buffer 1033 * is outside the range supported by the underlying device. 1034 */ 1035 bp->b_flags &= ~B_ERROR; 1036 bdirty(bp); 1037 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_FREEBUF)) || 1038 (bp->b_bufsize <= 0)) { 1039 /* 1040 * Either a failed I/O or we were asked to free or not 1041 * cache the buffer. 1042 */ 1043 bp->b_flags |= B_INVAL; 1044 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 1045 (*bioops.io_deallocate)(bp); 1046 if (bp->b_flags & B_DELWRI) { 1047 --numdirtybuffers; 1048 numdirtywakeup(lodirtybuffers); 1049 } 1050 bp->b_flags &= ~(B_DELWRI | B_CACHE | B_FREEBUF); 1051 if ((bp->b_flags & B_VMIO) == 0) { 1052 if (bp->b_bufsize) 1053 allocbuf(bp, 0); 1054 if (bp->b_vp) 1055 brelvp(bp); 1056 } 1057 } 1058 1059 /* 1060 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1061 * is called with B_DELWRI set, the underlying pages may wind up 1062 * getting freed causing a previous write (bdwrite()) to get 'lost' 1063 * because pages associated with a B_DELWRI bp are marked clean. 1064 * 1065 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1066 * if B_DELWRI is set. 1067 * 1068 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1069 * on pages to return pages to the VM page queues. 1070 */ 1071 if (bp->b_flags & B_DELWRI) 1072 bp->b_flags &= ~B_RELBUF; 1073 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1074 bp->b_flags |= B_RELBUF; 1075 1076 /* 1077 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1078 * constituted, not even NFS buffers now. Two flags effect this. If 1079 * B_INVAL, the struct buf is invalidated but the VM object is kept 1080 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1081 * 1082 * If B_ERROR or B_NOCACHE is set, pages in the VM object will be 1083 * invalidated. B_ERROR cannot be set for a failed write unless the 1084 * buffer is also B_INVAL because it hits the re-dirtying code above. 1085 * 1086 * Normally we can do this whether a buffer is B_DELWRI or not. If 1087 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1088 * the commit state and we cannot afford to lose the buffer. If the 1089 * buffer has a background write in progress, we need to keep it 1090 * around to prevent it from being reconstituted and starting a second 1091 * background write. 1092 */ 1093 if ((bp->b_flags & B_VMIO) 1094 && !(bp->b_vp->v_tag == VT_NFS && 1095 !vn_isdisk(bp->b_vp, NULL) && 1096 (bp->b_flags & B_DELWRI)) 1097 ) { 1098 1099 int i, j, resid; 1100 vm_page_t m; 1101 off_t foff; 1102 vm_pindex_t poff; 1103 vm_object_t obj; 1104 struct vnode *vp; 1105 1106 vp = bp->b_vp; 1107 1108 /* 1109 * Get the base offset and length of the buffer. Note that 1110 * in the VMIO case if the buffer block size is not 1111 * page-aligned then b_data pointer may not be page-aligned. 1112 * But our b_xio.xio_pages array *IS* page aligned. 1113 * 1114 * block sizes less then DEV_BSIZE (usually 512) are not 1115 * supported due to the page granularity bits (m->valid, 1116 * m->dirty, etc...). 1117 * 1118 * See man buf(9) for more information 1119 */ 1120 1121 resid = bp->b_bufsize; 1122 foff = bp->b_offset; 1123 1124 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1125 m = bp->b_xio.xio_pages[i]; 1126 vm_page_flag_clear(m, PG_ZERO); 1127 /* 1128 * If we hit a bogus page, fixup *all* of them 1129 * now. Note that we left these pages wired 1130 * when we removed them so they had better exist, 1131 * and they cannot be ripped out from under us so 1132 * no splvm() protection is necessary. 1133 */ 1134 if (m == bogus_page) { 1135 VOP_GETVOBJECT(vp, &obj); 1136 poff = OFF_TO_IDX(bp->b_offset); 1137 1138 for (j = i; j < bp->b_xio.xio_npages; j++) { 1139 vm_page_t mtmp; 1140 1141 mtmp = bp->b_xio.xio_pages[j]; 1142 if (mtmp == bogus_page) { 1143 mtmp = vm_page_lookup(obj, poff + j); 1144 if (!mtmp) { 1145 panic("brelse: page missing"); 1146 } 1147 bp->b_xio.xio_pages[j] = mtmp; 1148 } 1149 } 1150 1151 if ((bp->b_flags & B_INVAL) == 0) { 1152 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 1153 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 1154 } 1155 m = bp->b_xio.xio_pages[i]; 1156 } 1157 1158 /* 1159 * Invalidate the backing store if B_NOCACHE is set 1160 * (e.g. used with vinvalbuf()). If this is NFS 1161 * we impose a requirement that the block size be 1162 * a multiple of PAGE_SIZE and create a temporary 1163 * hack to basically invalidate the whole page. The 1164 * problem is that NFS uses really odd buffer sizes 1165 * especially when tracking piecemeal writes and 1166 * it also vinvalbuf()'s a lot, which would result 1167 * in only partial page validation and invalidation 1168 * here. If the file page is mmap()'d, however, 1169 * all the valid bits get set so after we invalidate 1170 * here we would end up with weird m->valid values 1171 * like 0xfc. nfs_getpages() can't handle this so 1172 * we clear all the valid bits for the NFS case 1173 * instead of just some of them. 1174 * 1175 * The real bug is the VM system having to set m->valid 1176 * to VM_PAGE_BITS_ALL for faulted-in pages, which 1177 * itself is an artifact of the whole 512-byte 1178 * granular mess that exists to support odd block 1179 * sizes and UFS meta-data block sizes (e.g. 6144). 1180 * A complete rewrite is required. 1181 */ 1182 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 1183 int poffset = foff & PAGE_MASK; 1184 int presid; 1185 1186 presid = PAGE_SIZE - poffset; 1187 if (bp->b_vp->v_tag == VT_NFS && 1188 bp->b_vp->v_type == VREG) { 1189 ; /* entire page */ 1190 } else if (presid > resid) { 1191 presid = resid; 1192 } 1193 KASSERT(presid >= 0, ("brelse: extra page")); 1194 vm_page_set_invalid(m, poffset, presid); 1195 } 1196 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1197 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1198 } 1199 1200 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1201 vfs_vmio_release(bp); 1202 1203 } else if (bp->b_flags & B_VMIO) { 1204 1205 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1206 vfs_vmio_release(bp); 1207 1208 } 1209 1210 if (bp->b_qindex != QUEUE_NONE) 1211 panic("brelse: free buffer onto another queue???"); 1212 if (BUF_REFCNTNB(bp) > 1) { 1213 /* Temporary panic to verify exclusive locking */ 1214 /* This panic goes away when we allow shared refs */ 1215 panic("brelse: multiple refs"); 1216 /* do not release to free list */ 1217 BUF_UNLOCK(bp); 1218 splx(s); 1219 return; 1220 } 1221 1222 /* enqueue */ 1223 1224 /* buffers with no memory */ 1225 if (bp->b_bufsize == 0) { 1226 bp->b_flags |= B_INVAL; 1227 bp->b_xflags &= ~BX_BKGRDWRITE; 1228 if (bp->b_xflags & BX_BKGRDINPROG) 1229 panic("losing buffer 1"); 1230 if (bp->b_kvasize) { 1231 bp->b_qindex = QUEUE_EMPTYKVA; 1232 } else { 1233 bp->b_qindex = QUEUE_EMPTY; 1234 } 1235 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1236 LIST_REMOVE(bp, b_hash); 1237 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1238 bp->b_dev = NODEV; 1239 /* buffers with junk contents */ 1240 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) { 1241 bp->b_flags |= B_INVAL; 1242 bp->b_xflags &= ~BX_BKGRDWRITE; 1243 if (bp->b_xflags & BX_BKGRDINPROG) 1244 panic("losing buffer 2"); 1245 bp->b_qindex = QUEUE_CLEAN; 1246 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1247 LIST_REMOVE(bp, b_hash); 1248 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1249 bp->b_dev = NODEV; 1250 1251 /* buffers that are locked */ 1252 } else if (bp->b_flags & B_LOCKED) { 1253 bp->b_qindex = QUEUE_LOCKED; 1254 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1255 1256 /* remaining buffers */ 1257 } else { 1258 switch(bp->b_flags & (B_DELWRI|B_AGE)) { 1259 case B_DELWRI | B_AGE: 1260 bp->b_qindex = QUEUE_DIRTY; 1261 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1262 break; 1263 case B_DELWRI: 1264 bp->b_qindex = QUEUE_DIRTY; 1265 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1266 break; 1267 case B_AGE: 1268 bp->b_qindex = QUEUE_CLEAN; 1269 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1270 break; 1271 default: 1272 bp->b_qindex = QUEUE_CLEAN; 1273 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1274 break; 1275 } 1276 } 1277 1278 /* 1279 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1280 * on the correct queue. 1281 */ 1282 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) 1283 bundirty(bp); 1284 1285 /* 1286 * Fixup numfreebuffers count. The bp is on an appropriate queue 1287 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1288 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1289 * if B_INVAL is set ). 1290 */ 1291 1292 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1293 bufcountwakeup(); 1294 1295 /* 1296 * Something we can maybe free or reuse 1297 */ 1298 if (bp->b_bufsize || bp->b_kvasize) 1299 bufspacewakeup(); 1300 1301 /* unlock */ 1302 BUF_UNLOCK(bp); 1303 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | 1304 B_DIRECT | B_NOWDRAIN); 1305 splx(s); 1306 } 1307 1308 /* 1309 * Release a buffer back to the appropriate queue but do not try to free 1310 * it. The buffer is expected to be used again soon. 1311 * 1312 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1313 * biodone() to requeue an async I/O on completion. It is also used when 1314 * known good buffers need to be requeued but we think we may need the data 1315 * again soon. 1316 * 1317 * XXX we should be able to leave the B_RELBUF hint set on completion. 1318 */ 1319 void 1320 bqrelse(struct buf * bp) 1321 { 1322 int s; 1323 1324 s = splbio(); 1325 1326 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1327 1328 if (bp->b_qindex != QUEUE_NONE) 1329 panic("bqrelse: free buffer onto another queue???"); 1330 if (BUF_REFCNTNB(bp) > 1) { 1331 /* do not release to free list */ 1332 panic("bqrelse: multiple refs"); 1333 BUF_UNLOCK(bp); 1334 splx(s); 1335 return; 1336 } 1337 if (bp->b_flags & B_LOCKED) { 1338 bp->b_flags &= ~B_ERROR; 1339 bp->b_qindex = QUEUE_LOCKED; 1340 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1341 /* buffers with stale but valid contents */ 1342 } else if (bp->b_flags & B_DELWRI) { 1343 bp->b_qindex = QUEUE_DIRTY; 1344 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1345 } else if (vm_page_count_severe()) { 1346 /* 1347 * We are too low on memory, we have to try to free the 1348 * buffer (most importantly: the wired pages making up its 1349 * backing store) *now*. 1350 */ 1351 splx(s); 1352 brelse(bp); 1353 return; 1354 } else { 1355 bp->b_qindex = QUEUE_CLEAN; 1356 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1357 } 1358 1359 if ((bp->b_flags & B_LOCKED) == 0 && 1360 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1361 bufcountwakeup(); 1362 } 1363 1364 /* 1365 * Something we can maybe free or reuse. 1366 */ 1367 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1368 bufspacewakeup(); 1369 1370 /* unlock */ 1371 BUF_UNLOCK(bp); 1372 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1373 splx(s); 1374 } 1375 1376 static void 1377 vfs_vmio_release(struct buf *bp) 1378 { 1379 int i, s; 1380 vm_page_t m; 1381 1382 s = splvm(); 1383 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1384 m = bp->b_xio.xio_pages[i]; 1385 bp->b_xio.xio_pages[i] = NULL; 1386 /* 1387 * In order to keep page LRU ordering consistent, put 1388 * everything on the inactive queue. 1389 */ 1390 vm_page_unwire(m, 0); 1391 /* 1392 * We don't mess with busy pages, it is 1393 * the responsibility of the process that 1394 * busied the pages to deal with them. 1395 */ 1396 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1397 continue; 1398 1399 if (m->wire_count == 0) { 1400 vm_page_flag_clear(m, PG_ZERO); 1401 /* 1402 * Might as well free the page if we can and it has 1403 * no valid data. We also free the page if the 1404 * buffer was used for direct I/O. 1405 */ 1406 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { 1407 vm_page_busy(m); 1408 vm_page_protect(m, VM_PROT_NONE); 1409 vm_page_free(m); 1410 } else if (bp->b_flags & B_DIRECT) { 1411 vm_page_try_to_free(m); 1412 } else if (vm_page_count_severe()) { 1413 vm_page_try_to_cache(m); 1414 } 1415 } 1416 } 1417 splx(s); 1418 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages); 1419 if (bp->b_bufsize) { 1420 bufspacewakeup(); 1421 bp->b_bufsize = 0; 1422 } 1423 bp->b_xio.xio_npages = 0; 1424 bp->b_flags &= ~B_VMIO; 1425 if (bp->b_vp) 1426 brelvp(bp); 1427 } 1428 1429 /* 1430 * Check to see if a block is currently memory resident. 1431 */ 1432 struct buf * 1433 gbincore(struct vnode * vp, daddr_t blkno) 1434 { 1435 struct buf *bp; 1436 struct bufhashhdr *bh; 1437 1438 bh = bufhash(vp, blkno); 1439 1440 /* Search hash chain */ 1441 LIST_FOREACH(bp, bh, b_hash) { 1442 /* hit */ 1443 if (bp->b_vp == vp && bp->b_lblkno == blkno && 1444 (bp->b_flags & B_INVAL) == 0) { 1445 break; 1446 } 1447 } 1448 return (bp); 1449 } 1450 1451 /* 1452 * vfs_bio_awrite: 1453 * 1454 * Implement clustered async writes for clearing out B_DELWRI buffers. 1455 * This is much better then the old way of writing only one buffer at 1456 * a time. Note that we may not be presented with the buffers in the 1457 * correct order, so we search for the cluster in both directions. 1458 */ 1459 int 1460 vfs_bio_awrite(struct buf * bp) 1461 { 1462 int i; 1463 int j; 1464 daddr_t lblkno = bp->b_lblkno; 1465 struct vnode *vp = bp->b_vp; 1466 int s; 1467 int ncl; 1468 struct buf *bpa; 1469 int nwritten; 1470 int size; 1471 int maxcl; 1472 1473 s = splbio(); 1474 /* 1475 * right now we support clustered writing only to regular files. If 1476 * we find a clusterable block we could be in the middle of a cluster 1477 * rather then at the beginning. 1478 */ 1479 if ((vp->v_type == VREG) && 1480 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1481 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1482 1483 size = vp->v_mount->mnt_stat.f_iosize; 1484 maxcl = MAXPHYS / size; 1485 1486 for (i = 1; i < maxcl; i++) { 1487 if ((bpa = gbincore(vp, lblkno + i)) && 1488 BUF_REFCNT(bpa) == 0 && 1489 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1490 (B_DELWRI | B_CLUSTEROK)) && 1491 (bpa->b_bufsize == size)) { 1492 if ((bpa->b_blkno == bpa->b_lblkno) || 1493 (bpa->b_blkno != 1494 bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 1495 break; 1496 } else { 1497 break; 1498 } 1499 } 1500 for (j = 1; i + j <= maxcl && j <= lblkno; j++) { 1501 if ((bpa = gbincore(vp, lblkno - j)) && 1502 BUF_REFCNT(bpa) == 0 && 1503 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1504 (B_DELWRI | B_CLUSTEROK)) && 1505 (bpa->b_bufsize == size)) { 1506 if ((bpa->b_blkno == bpa->b_lblkno) || 1507 (bpa->b_blkno != 1508 bp->b_blkno - ((j * size) >> DEV_BSHIFT))) 1509 break; 1510 } else { 1511 break; 1512 } 1513 } 1514 --j; 1515 ncl = i + j; 1516 /* 1517 * this is a possible cluster write 1518 */ 1519 if (ncl != 1) { 1520 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1521 splx(s); 1522 return nwritten; 1523 } 1524 } 1525 1526 BUF_LOCK(bp, LK_EXCLUSIVE); 1527 bremfree(bp); 1528 bp->b_flags |= B_ASYNC; 1529 1530 splx(s); 1531 /* 1532 * default (old) behavior, writing out only one block 1533 * 1534 * XXX returns b_bufsize instead of b_bcount for nwritten? 1535 */ 1536 nwritten = bp->b_bufsize; 1537 (void) VOP_BWRITE(bp->b_vp, bp); 1538 1539 return nwritten; 1540 } 1541 1542 /* 1543 * getnewbuf: 1544 * 1545 * Find and initialize a new buffer header, freeing up existing buffers 1546 * in the bufqueues as necessary. The new buffer is returned locked. 1547 * 1548 * Important: B_INVAL is not set. If the caller wishes to throw the 1549 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1550 * 1551 * We block if: 1552 * We have insufficient buffer headers 1553 * We have insufficient buffer space 1554 * buffer_map is too fragmented ( space reservation fails ) 1555 * If we have to flush dirty buffers ( but we try to avoid this ) 1556 * 1557 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1558 * Instead we ask the buf daemon to do it for us. We attempt to 1559 * avoid piecemeal wakeups of the pageout daemon. 1560 */ 1561 1562 static struct buf * 1563 getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1564 { 1565 struct buf *bp; 1566 struct buf *nbp; 1567 int defrag = 0; 1568 int nqindex; 1569 static int flushingbufs; 1570 1571 /* 1572 * We can't afford to block since we might be holding a vnode lock, 1573 * which may prevent system daemons from running. We deal with 1574 * low-memory situations by proactively returning memory and running 1575 * async I/O rather then sync I/O. 1576 */ 1577 1578 ++getnewbufcalls; 1579 --getnewbufrestarts; 1580 restart: 1581 ++getnewbufrestarts; 1582 1583 /* 1584 * Setup for scan. If we do not have enough free buffers, 1585 * we setup a degenerate case that immediately fails. Note 1586 * that if we are specially marked process, we are allowed to 1587 * dip into our reserves. 1588 * 1589 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1590 * 1591 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1592 * However, there are a number of cases (defragging, reusing, ...) 1593 * where we cannot backup. 1594 */ 1595 nqindex = QUEUE_EMPTYKVA; 1596 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1597 1598 if (nbp == NULL) { 1599 /* 1600 * If no EMPTYKVA buffers and we are either 1601 * defragging or reusing, locate a CLEAN buffer 1602 * to free or reuse. If bufspace useage is low 1603 * skip this step so we can allocate a new buffer. 1604 */ 1605 if (defrag || bufspace >= lobufspace) { 1606 nqindex = QUEUE_CLEAN; 1607 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1608 } 1609 1610 /* 1611 * If we could not find or were not allowed to reuse a 1612 * CLEAN buffer, check to see if it is ok to use an EMPTY 1613 * buffer. We can only use an EMPTY buffer if allocating 1614 * its KVA would not otherwise run us out of buffer space. 1615 */ 1616 if (nbp == NULL && defrag == 0 && 1617 bufspace + maxsize < hibufspace) { 1618 nqindex = QUEUE_EMPTY; 1619 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1620 } 1621 } 1622 1623 /* 1624 * Run scan, possibly freeing data and/or kva mappings on the fly 1625 * depending. 1626 */ 1627 1628 while ((bp = nbp) != NULL) { 1629 int qindex = nqindex; 1630 1631 /* 1632 * Calculate next bp ( we can only use it if we do not block 1633 * or do other fancy things ). 1634 */ 1635 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1636 switch(qindex) { 1637 case QUEUE_EMPTY: 1638 nqindex = QUEUE_EMPTYKVA; 1639 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1640 break; 1641 /* fall through */ 1642 case QUEUE_EMPTYKVA: 1643 nqindex = QUEUE_CLEAN; 1644 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1645 break; 1646 /* fall through */ 1647 case QUEUE_CLEAN: 1648 /* 1649 * nbp is NULL. 1650 */ 1651 break; 1652 } 1653 } 1654 1655 /* 1656 * Sanity Checks 1657 */ 1658 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1659 1660 /* 1661 * Note: we no longer distinguish between VMIO and non-VMIO 1662 * buffers. 1663 */ 1664 1665 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1666 1667 /* 1668 * If we are defragging then we need a buffer with 1669 * b_kvasize != 0. XXX this situation should no longer 1670 * occur, if defrag is non-zero the buffer's b_kvasize 1671 * should also be non-zero at this point. XXX 1672 */ 1673 if (defrag && bp->b_kvasize == 0) { 1674 printf("Warning: defrag empty buffer %p\n", bp); 1675 continue; 1676 } 1677 1678 /* 1679 * Start freeing the bp. This is somewhat involved. nbp 1680 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1681 */ 1682 1683 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1684 panic("getnewbuf: locked buf"); 1685 bremfree(bp); 1686 1687 if (qindex == QUEUE_CLEAN) { 1688 if (bp->b_flags & B_VMIO) { 1689 bp->b_flags &= ~B_ASYNC; 1690 vfs_vmio_release(bp); 1691 } 1692 if (bp->b_vp) 1693 brelvp(bp); 1694 } 1695 1696 /* 1697 * NOTE: nbp is now entirely invalid. We can only restart 1698 * the scan from this point on. 1699 * 1700 * Get the rest of the buffer freed up. b_kva* is still 1701 * valid after this operation. 1702 */ 1703 1704 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 1705 (*bioops.io_deallocate)(bp); 1706 if (bp->b_xflags & BX_BKGRDINPROG) 1707 panic("losing buffer 3"); 1708 LIST_REMOVE(bp, b_hash); 1709 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1710 1711 /* 1712 * spl protection not required when scrapping a buffer's 1713 * contents because it is already wired. 1714 */ 1715 if (bp->b_bufsize) 1716 allocbuf(bp, 0); 1717 1718 bp->b_flags = 0; 1719 bp->b_xflags = 0; 1720 bp->b_dev = NODEV; 1721 bp->b_vp = NULL; 1722 bp->b_blkno = bp->b_lblkno = 0; 1723 bp->b_offset = NOOFFSET; 1724 bp->b_iodone = 0; 1725 bp->b_error = 0; 1726 bp->b_resid = 0; 1727 bp->b_bcount = 0; 1728 bp->b_xio.xio_npages = 0; 1729 bp->b_dirtyoff = bp->b_dirtyend = 0; 1730 1731 LIST_INIT(&bp->b_dep); 1732 1733 /* 1734 * If we are defragging then free the buffer. 1735 */ 1736 if (defrag) { 1737 bp->b_flags |= B_INVAL; 1738 bfreekva(bp); 1739 brelse(bp); 1740 defrag = 0; 1741 goto restart; 1742 } 1743 1744 /* 1745 * If we are overcomitted then recover the buffer and its 1746 * KVM space. This occurs in rare situations when multiple 1747 * processes are blocked in getnewbuf() or allocbuf(). 1748 */ 1749 if (bufspace >= hibufspace) 1750 flushingbufs = 1; 1751 if (flushingbufs && bp->b_kvasize != 0) { 1752 bp->b_flags |= B_INVAL; 1753 bfreekva(bp); 1754 brelse(bp); 1755 goto restart; 1756 } 1757 if (bufspace < lobufspace) 1758 flushingbufs = 0; 1759 break; 1760 } 1761 1762 /* 1763 * If we exhausted our list, sleep as appropriate. We may have to 1764 * wakeup various daemons and write out some dirty buffers. 1765 * 1766 * Generally we are sleeping due to insufficient buffer space. 1767 */ 1768 1769 if (bp == NULL) { 1770 int flags; 1771 char *waitmsg; 1772 1773 if (defrag) { 1774 flags = VFS_BIO_NEED_BUFSPACE; 1775 waitmsg = "nbufkv"; 1776 } else if (bufspace >= hibufspace) { 1777 waitmsg = "nbufbs"; 1778 flags = VFS_BIO_NEED_BUFSPACE; 1779 } else { 1780 waitmsg = "newbuf"; 1781 flags = VFS_BIO_NEED_ANY; 1782 } 1783 1784 bd_speedup(); /* heeeelp */ 1785 1786 needsbuffer |= flags; 1787 while (needsbuffer & flags) { 1788 if (tsleep(&needsbuffer, slpflag, waitmsg, slptimeo)) 1789 return (NULL); 1790 } 1791 } else { 1792 /* 1793 * We finally have a valid bp. We aren't quite out of the 1794 * woods, we still have to reserve kva space. In order 1795 * to keep fragmentation sane we only allocate kva in 1796 * BKVASIZE chunks. 1797 */ 1798 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1799 1800 if (maxsize != bp->b_kvasize) { 1801 vm_offset_t addr = 0; 1802 int count; 1803 1804 bfreekva(bp); 1805 1806 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1807 vm_map_lock(buffer_map); 1808 1809 if (vm_map_findspace(buffer_map, 1810 vm_map_min(buffer_map), maxsize, 1811 maxsize, &addr)) { 1812 /* 1813 * Uh oh. Buffer map is to fragmented. We 1814 * must defragment the map. 1815 */ 1816 vm_map_unlock(buffer_map); 1817 vm_map_entry_release(count); 1818 ++bufdefragcnt; 1819 defrag = 1; 1820 bp->b_flags |= B_INVAL; 1821 brelse(bp); 1822 goto restart; 1823 } 1824 if (addr) { 1825 vm_map_insert(buffer_map, &count, 1826 NULL, 0, 1827 addr, addr + maxsize, 1828 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1829 1830 bp->b_kvabase = (caddr_t) addr; 1831 bp->b_kvasize = maxsize; 1832 bufspace += bp->b_kvasize; 1833 ++bufreusecnt; 1834 } 1835 vm_map_unlock(buffer_map); 1836 vm_map_entry_release(count); 1837 } 1838 bp->b_data = bp->b_kvabase; 1839 } 1840 return(bp); 1841 } 1842 1843 /* 1844 * buf_daemon: 1845 * 1846 * buffer flushing daemon. Buffers are normally flushed by the 1847 * update daemon but if it cannot keep up this process starts to 1848 * take the load in an attempt to prevent getnewbuf() from blocking. 1849 */ 1850 1851 static struct thread *bufdaemonthread; 1852 1853 static struct kproc_desc buf_kp = { 1854 "bufdaemon", 1855 buf_daemon, 1856 &bufdaemonthread 1857 }; 1858 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1859 1860 static void 1861 buf_daemon() 1862 { 1863 int s; 1864 1865 /* 1866 * This process needs to be suspended prior to shutdown sync. 1867 */ 1868 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 1869 bufdaemonthread, SHUTDOWN_PRI_LAST); 1870 1871 /* 1872 * This process is allowed to take the buffer cache to the limit 1873 */ 1874 s = splbio(); 1875 1876 for (;;) { 1877 kproc_suspend_loop(); 1878 1879 /* 1880 * Do the flush. Limit the amount of in-transit I/O we 1881 * allow to build up, otherwise we would completely saturate 1882 * the I/O system. Wakeup any waiting processes before we 1883 * normally would so they can run in parallel with our drain. 1884 */ 1885 while (numdirtybuffers > lodirtybuffers) { 1886 if (flushbufqueues() == 0) 1887 break; 1888 waitrunningbufspace(); 1889 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 1890 } 1891 1892 /* 1893 * Only clear bd_request if we have reached our low water 1894 * mark. The buf_daemon normally waits 5 seconds and 1895 * then incrementally flushes any dirty buffers that have 1896 * built up, within reason. 1897 * 1898 * If we were unable to hit our low water mark and couldn't 1899 * find any flushable buffers, we sleep half a second. 1900 * Otherwise we loop immediately. 1901 */ 1902 if (numdirtybuffers <= lodirtybuffers) { 1903 /* 1904 * We reached our low water mark, reset the 1905 * request and sleep until we are needed again. 1906 * The sleep is just so the suspend code works. 1907 */ 1908 bd_request = 0; 1909 tsleep(&bd_request, 0, "psleep", hz); 1910 } else { 1911 /* 1912 * We couldn't find any flushable dirty buffers but 1913 * still have too many dirty buffers, we 1914 * have to sleep and try again. (rare) 1915 */ 1916 tsleep(&bd_request, 0, "qsleep", hz / 2); 1917 } 1918 } 1919 } 1920 1921 /* 1922 * flushbufqueues: 1923 * 1924 * Try to flush a buffer in the dirty queue. We must be careful to 1925 * free up B_INVAL buffers instead of write them, which NFS is 1926 * particularly sensitive to. 1927 */ 1928 1929 static int 1930 flushbufqueues(void) 1931 { 1932 struct buf *bp; 1933 int r = 0; 1934 1935 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1936 1937 while (bp) { 1938 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); 1939 if ((bp->b_flags & B_DELWRI) != 0 && 1940 (bp->b_xflags & BX_BKGRDINPROG) == 0) { 1941 if (bp->b_flags & B_INVAL) { 1942 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1943 panic("flushbufqueues: locked buf"); 1944 bremfree(bp); 1945 brelse(bp); 1946 ++r; 1947 break; 1948 } 1949 if (LIST_FIRST(&bp->b_dep) != NULL && 1950 bioops.io_countdeps && 1951 (bp->b_flags & B_DEFERRED) == 0 && 1952 (*bioops.io_countdeps)(bp, 0)) { 1953 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], 1954 bp, b_freelist); 1955 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], 1956 bp, b_freelist); 1957 bp->b_flags |= B_DEFERRED; 1958 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1959 continue; 1960 } 1961 vfs_bio_awrite(bp); 1962 ++r; 1963 break; 1964 } 1965 bp = TAILQ_NEXT(bp, b_freelist); 1966 } 1967 return (r); 1968 } 1969 1970 /* 1971 * Check to see if a block is currently memory resident. 1972 */ 1973 struct buf * 1974 incore(struct vnode * vp, daddr_t blkno) 1975 { 1976 struct buf *bp; 1977 1978 crit_enter(); 1979 bp = gbincore(vp, blkno); 1980 crit_exit(); 1981 return (bp); 1982 } 1983 1984 /* 1985 * Returns true if no I/O is needed to access the associated VM object. 1986 * This is like incore except it also hunts around in the VM system for 1987 * the data. 1988 * 1989 * Note that we ignore vm_page_free() races from interrupts against our 1990 * lookup, since if the caller is not protected our return value will not 1991 * be any more valid then otherwise once we splx(). 1992 */ 1993 int 1994 inmem(struct vnode * vp, daddr_t blkno) 1995 { 1996 vm_object_t obj; 1997 vm_offset_t toff, tinc, size; 1998 vm_page_t m; 1999 vm_ooffset_t off; 2000 2001 if (incore(vp, blkno)) 2002 return 1; 2003 if (vp->v_mount == NULL) 2004 return 0; 2005 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_flag & VOBJBUF) == 0) 2006 return 0; 2007 2008 size = PAGE_SIZE; 2009 if (size > vp->v_mount->mnt_stat.f_iosize) 2010 size = vp->v_mount->mnt_stat.f_iosize; 2011 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2012 2013 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2014 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2015 if (!m) 2016 return 0; 2017 tinc = size; 2018 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2019 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2020 if (vm_page_is_valid(m, 2021 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2022 return 0; 2023 } 2024 return 1; 2025 } 2026 2027 /* 2028 * vfs_setdirty: 2029 * 2030 * Sets the dirty range for a buffer based on the status of the dirty 2031 * bits in the pages comprising the buffer. 2032 * 2033 * The range is limited to the size of the buffer. 2034 * 2035 * This routine is primarily used by NFS, but is generalized for the 2036 * B_VMIO case. 2037 */ 2038 static void 2039 vfs_setdirty(struct buf *bp) 2040 { 2041 int i; 2042 vm_object_t object; 2043 2044 /* 2045 * Degenerate case - empty buffer 2046 */ 2047 2048 if (bp->b_bufsize == 0) 2049 return; 2050 2051 /* 2052 * We qualify the scan for modified pages on whether the 2053 * object has been flushed yet. The OBJ_WRITEABLE flag 2054 * is not cleared simply by protecting pages off. 2055 */ 2056 2057 if ((bp->b_flags & B_VMIO) == 0) 2058 return; 2059 2060 object = bp->b_xio.xio_pages[0]->object; 2061 2062 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2063 printf("Warning: object %p writeable but not mightbedirty\n", object); 2064 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2065 printf("Warning: object %p mightbedirty but not writeable\n", object); 2066 2067 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2068 vm_offset_t boffset; 2069 vm_offset_t eoffset; 2070 2071 /* 2072 * test the pages to see if they have been modified directly 2073 * by users through the VM system. 2074 */ 2075 for (i = 0; i < bp->b_xio.xio_npages; i++) { 2076 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO); 2077 vm_page_test_dirty(bp->b_xio.xio_pages[i]); 2078 } 2079 2080 /* 2081 * Calculate the encompassing dirty range, boffset and eoffset, 2082 * (eoffset - boffset) bytes. 2083 */ 2084 2085 for (i = 0; i < bp->b_xio.xio_npages; i++) { 2086 if (bp->b_xio.xio_pages[i]->dirty) 2087 break; 2088 } 2089 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2090 2091 for (i = bp->b_xio.xio_npages - 1; i >= 0; --i) { 2092 if (bp->b_xio.xio_pages[i]->dirty) { 2093 break; 2094 } 2095 } 2096 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2097 2098 /* 2099 * Fit it to the buffer. 2100 */ 2101 2102 if (eoffset > bp->b_bcount) 2103 eoffset = bp->b_bcount; 2104 2105 /* 2106 * If we have a good dirty range, merge with the existing 2107 * dirty range. 2108 */ 2109 2110 if (boffset < eoffset) { 2111 if (bp->b_dirtyoff > boffset) 2112 bp->b_dirtyoff = boffset; 2113 if (bp->b_dirtyend < eoffset) 2114 bp->b_dirtyend = eoffset; 2115 } 2116 } 2117 } 2118 2119 /* 2120 * getblk: 2121 * 2122 * Get a block given a specified block and offset into a file/device. 2123 * The buffers B_DONE bit will be cleared on return, making it almost 2124 * ready for an I/O initiation. B_INVAL may or may not be set on 2125 * return. The caller should clear B_INVAL prior to initiating a 2126 * READ. 2127 * 2128 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE 2129 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ, 2130 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer 2131 * without doing any of those things the system will likely believe 2132 * the buffer to be valid (especially if it is not B_VMIO), and the 2133 * next getblk() will return the buffer with B_CACHE set. 2134 * 2135 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2136 * an existing buffer. 2137 * 2138 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2139 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2140 * and then cleared based on the backing VM. If the previous buffer is 2141 * non-0-sized but invalid, B_CACHE will be cleared. 2142 * 2143 * If getblk() must create a new buffer, the new buffer is returned with 2144 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2145 * case it is returned with B_INVAL clear and B_CACHE set based on the 2146 * backing VM. 2147 * 2148 * getblk() also forces a VOP_BWRITE() for any B_DELWRI buffer whos 2149 * B_CACHE bit is clear. 2150 * 2151 * What this means, basically, is that the caller should use B_CACHE to 2152 * determine whether the buffer is fully valid or not and should clear 2153 * B_INVAL prior to issuing a read. If the caller intends to validate 2154 * the buffer by loading its data area with something, the caller needs 2155 * to clear B_INVAL. If the caller does this without issuing an I/O, 2156 * the caller should set B_CACHE ( as an optimization ), else the caller 2157 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2158 * a write attempt or if it was a successfull read. If the caller 2159 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR 2160 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2161 */ 2162 struct buf * 2163 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2164 { 2165 struct buf *bp; 2166 int s; 2167 struct bufhashhdr *bh; 2168 2169 if (size > MAXBSIZE) 2170 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE); 2171 2172 s = splbio(); 2173 loop: 2174 /* 2175 * Block if we are low on buffers. Certain processes are allowed 2176 * to completely exhaust the buffer cache. 2177 * 2178 * If this check ever becomes a bottleneck it may be better to 2179 * move it into the else, when gbincore() fails. At the moment 2180 * it isn't a problem. 2181 * 2182 * XXX remove, we cannot afford to block anywhere if holding a vnode 2183 * lock in low-memory situation, so take it to the max. 2184 */ 2185 if (numfreebuffers == 0) { 2186 if (!curproc) 2187 return NULL; 2188 needsbuffer |= VFS_BIO_NEED_ANY; 2189 tsleep(&needsbuffer, slpflag, "newbuf", slptimeo); 2190 } 2191 2192 if ((bp = gbincore(vp, blkno))) { 2193 /* 2194 * Buffer is in-core. If the buffer is not busy, it must 2195 * be on a queue. 2196 */ 2197 2198 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2199 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, 2200 "getblk", slpflag, slptimeo) == ENOLCK) 2201 goto loop; 2202 splx(s); 2203 return (struct buf *) NULL; 2204 } 2205 2206 /* 2207 * The buffer is locked. B_CACHE is cleared if the buffer is 2208 * invalid. Ohterwise, for a non-VMIO buffer, B_CACHE is set 2209 * and for a VMIO buffer B_CACHE is adjusted according to the 2210 * backing VM cache. 2211 */ 2212 if (bp->b_flags & B_INVAL) 2213 bp->b_flags &= ~B_CACHE; 2214 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2215 bp->b_flags |= B_CACHE; 2216 bremfree(bp); 2217 2218 /* 2219 * check for size inconsistancies for non-VMIO case. 2220 */ 2221 2222 if (bp->b_bcount != size) { 2223 if ((bp->b_flags & B_VMIO) == 0 || 2224 (size > bp->b_kvasize)) { 2225 if (bp->b_flags & B_DELWRI) { 2226 bp->b_flags |= B_NOCACHE; 2227 VOP_BWRITE(bp->b_vp, bp); 2228 } else { 2229 if ((bp->b_flags & B_VMIO) && 2230 (LIST_FIRST(&bp->b_dep) == NULL)) { 2231 bp->b_flags |= B_RELBUF; 2232 brelse(bp); 2233 } else { 2234 bp->b_flags |= B_NOCACHE; 2235 VOP_BWRITE(bp->b_vp, bp); 2236 } 2237 } 2238 goto loop; 2239 } 2240 } 2241 2242 /* 2243 * If the size is inconsistant in the VMIO case, we can resize 2244 * the buffer. This might lead to B_CACHE getting set or 2245 * cleared. If the size has not changed, B_CACHE remains 2246 * unchanged from its previous state. 2247 */ 2248 2249 if (bp->b_bcount != size) 2250 allocbuf(bp, size); 2251 2252 KASSERT(bp->b_offset != NOOFFSET, 2253 ("getblk: no buffer offset")); 2254 2255 /* 2256 * A buffer with B_DELWRI set and B_CACHE clear must 2257 * be committed before we can return the buffer in 2258 * order to prevent the caller from issuing a read 2259 * ( due to B_CACHE not being set ) and overwriting 2260 * it. 2261 * 2262 * Most callers, including NFS and FFS, need this to 2263 * operate properly either because they assume they 2264 * can issue a read if B_CACHE is not set, or because 2265 * ( for example ) an uncached B_DELWRI might loop due 2266 * to softupdates re-dirtying the buffer. In the latter 2267 * case, B_CACHE is set after the first write completes, 2268 * preventing further loops. 2269 * 2270 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2271 * above while extending the buffer, we cannot allow the 2272 * buffer to remain with B_CACHE set after the write 2273 * completes or it will represent a corrupt state. To 2274 * deal with this we set B_NOCACHE to scrap the buffer 2275 * after the write. 2276 * 2277 * We might be able to do something fancy, like setting 2278 * B_CACHE in bwrite() except if B_DELWRI is already set, 2279 * so the below call doesn't set B_CACHE, but that gets real 2280 * confusing. This is much easier. 2281 */ 2282 2283 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2284 bp->b_flags |= B_NOCACHE; 2285 VOP_BWRITE(bp->b_vp, bp); 2286 goto loop; 2287 } 2288 2289 splx(s); 2290 bp->b_flags &= ~B_DONE; 2291 } else { 2292 /* 2293 * Buffer is not in-core, create new buffer. The buffer 2294 * returned by getnewbuf() is locked. Note that the returned 2295 * buffer is also considered valid (not marked B_INVAL). 2296 */ 2297 int bsize, maxsize, vmio; 2298 off_t offset; 2299 2300 if (vn_isdisk(vp, NULL)) 2301 bsize = DEV_BSIZE; 2302 else if (vp->v_mountedhere) 2303 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2304 else if (vp->v_mount) 2305 bsize = vp->v_mount->mnt_stat.f_iosize; 2306 else 2307 bsize = size; 2308 2309 offset = (off_t)blkno * bsize; 2310 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && (vp->v_flag & VOBJBUF); 2311 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2312 maxsize = imax(maxsize, bsize); 2313 2314 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2315 if (slpflag || slptimeo) { 2316 splx(s); 2317 return NULL; 2318 } 2319 goto loop; 2320 } 2321 2322 /* 2323 * This code is used to make sure that a buffer is not 2324 * created while the getnewbuf routine is blocked. 2325 * This can be a problem whether the vnode is locked or not. 2326 * If the buffer is created out from under us, we have to 2327 * throw away the one we just created. There is now window 2328 * race because we are safely running at splbio() from the 2329 * point of the duplicate buffer creation through to here, 2330 * and we've locked the buffer. 2331 */ 2332 if (gbincore(vp, blkno)) { 2333 bp->b_flags |= B_INVAL; 2334 brelse(bp); 2335 goto loop; 2336 } 2337 2338 /* 2339 * Insert the buffer into the hash, so that it can 2340 * be found by incore. 2341 */ 2342 bp->b_blkno = bp->b_lblkno = blkno; 2343 bp->b_offset = offset; 2344 2345 bgetvp(vp, bp); 2346 LIST_REMOVE(bp, b_hash); 2347 bh = bufhash(vp, blkno); 2348 LIST_INSERT_HEAD(bh, bp, b_hash); 2349 2350 /* 2351 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2352 * buffer size starts out as 0, B_CACHE will be set by 2353 * allocbuf() for the VMIO case prior to it testing the 2354 * backing store for validity. 2355 */ 2356 2357 if (vmio) { 2358 bp->b_flags |= B_VMIO; 2359 #if defined(VFS_BIO_DEBUG) 2360 if (vn_canvmio(vp) != TRUE) 2361 printf("getblk: vmioing file type %d???\n", vp->v_type); 2362 #endif 2363 } else { 2364 bp->b_flags &= ~B_VMIO; 2365 } 2366 2367 allocbuf(bp, size); 2368 2369 splx(s); 2370 bp->b_flags &= ~B_DONE; 2371 } 2372 return (bp); 2373 } 2374 2375 /* 2376 * Get an empty, disassociated buffer of given size. The buffer is initially 2377 * set to B_INVAL. 2378 * 2379 * spl protection is not required for the allocbuf() call because races are 2380 * impossible here. 2381 */ 2382 struct buf * 2383 geteblk(int size) 2384 { 2385 struct buf *bp; 2386 int s; 2387 int maxsize; 2388 2389 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2390 2391 s = splbio(); 2392 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0); 2393 splx(s); 2394 allocbuf(bp, size); 2395 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2396 return (bp); 2397 } 2398 2399 2400 /* 2401 * This code constitutes the buffer memory from either anonymous system 2402 * memory (in the case of non-VMIO operations) or from an associated 2403 * VM object (in the case of VMIO operations). This code is able to 2404 * resize a buffer up or down. 2405 * 2406 * Note that this code is tricky, and has many complications to resolve 2407 * deadlock or inconsistant data situations. Tread lightly!!! 2408 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2409 * the caller. Calling this code willy nilly can result in the loss of data. 2410 * 2411 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2412 * B_CACHE for the non-VMIO case. 2413 * 2414 * This routine does not need to be called at splbio() but you must own the 2415 * buffer. 2416 */ 2417 int 2418 allocbuf(struct buf *bp, int size) 2419 { 2420 int newbsize, mbsize; 2421 int i; 2422 2423 if (BUF_REFCNT(bp) == 0) 2424 panic("allocbuf: buffer not busy"); 2425 2426 if (bp->b_kvasize < size) 2427 panic("allocbuf: buffer too small"); 2428 2429 if ((bp->b_flags & B_VMIO) == 0) { 2430 caddr_t origbuf; 2431 int origbufsize; 2432 /* 2433 * Just get anonymous memory from the kernel. Don't 2434 * mess with B_CACHE. 2435 */ 2436 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2437 #if !defined(NO_B_MALLOC) 2438 if (bp->b_flags & B_MALLOC) 2439 newbsize = mbsize; 2440 else 2441 #endif 2442 newbsize = round_page(size); 2443 2444 if (newbsize < bp->b_bufsize) { 2445 #if !defined(NO_B_MALLOC) 2446 /* 2447 * malloced buffers are not shrunk 2448 */ 2449 if (bp->b_flags & B_MALLOC) { 2450 if (newbsize) { 2451 bp->b_bcount = size; 2452 } else { 2453 free(bp->b_data, M_BIOBUF); 2454 if (bp->b_bufsize) { 2455 bufmallocspace -= bp->b_bufsize; 2456 bufspacewakeup(); 2457 bp->b_bufsize = 0; 2458 } 2459 bp->b_data = bp->b_kvabase; 2460 bp->b_bcount = 0; 2461 bp->b_flags &= ~B_MALLOC; 2462 } 2463 return 1; 2464 } 2465 #endif 2466 vm_hold_free_pages( 2467 bp, 2468 (vm_offset_t) bp->b_data + newbsize, 2469 (vm_offset_t) bp->b_data + bp->b_bufsize); 2470 } else if (newbsize > bp->b_bufsize) { 2471 #if !defined(NO_B_MALLOC) 2472 /* 2473 * We only use malloced memory on the first allocation. 2474 * and revert to page-allocated memory when the buffer 2475 * grows. 2476 */ 2477 if ( (bufmallocspace < maxbufmallocspace) && 2478 (bp->b_bufsize == 0) && 2479 (mbsize <= PAGE_SIZE/2)) { 2480 2481 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2482 bp->b_bufsize = mbsize; 2483 bp->b_bcount = size; 2484 bp->b_flags |= B_MALLOC; 2485 bufmallocspace += mbsize; 2486 return 1; 2487 } 2488 #endif 2489 origbuf = NULL; 2490 origbufsize = 0; 2491 #if !defined(NO_B_MALLOC) 2492 /* 2493 * If the buffer is growing on its other-than-first allocation, 2494 * then we revert to the page-allocation scheme. 2495 */ 2496 if (bp->b_flags & B_MALLOC) { 2497 origbuf = bp->b_data; 2498 origbufsize = bp->b_bufsize; 2499 bp->b_data = bp->b_kvabase; 2500 if (bp->b_bufsize) { 2501 bufmallocspace -= bp->b_bufsize; 2502 bufspacewakeup(); 2503 bp->b_bufsize = 0; 2504 } 2505 bp->b_flags &= ~B_MALLOC; 2506 newbsize = round_page(newbsize); 2507 } 2508 #endif 2509 vm_hold_load_pages( 2510 bp, 2511 (vm_offset_t) bp->b_data + bp->b_bufsize, 2512 (vm_offset_t) bp->b_data + newbsize); 2513 #if !defined(NO_B_MALLOC) 2514 if (origbuf) { 2515 bcopy(origbuf, bp->b_data, origbufsize); 2516 free(origbuf, M_BIOBUF); 2517 } 2518 #endif 2519 } 2520 } else { 2521 vm_page_t m; 2522 int desiredpages; 2523 2524 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2525 desiredpages = (size == 0) ? 0 : 2526 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2527 2528 #if !defined(NO_B_MALLOC) 2529 if (bp->b_flags & B_MALLOC) 2530 panic("allocbuf: VMIO buffer can't be malloced"); 2531 #endif 2532 /* 2533 * Set B_CACHE initially if buffer is 0 length or will become 2534 * 0-length. 2535 */ 2536 if (size == 0 || bp->b_bufsize == 0) 2537 bp->b_flags |= B_CACHE; 2538 2539 if (newbsize < bp->b_bufsize) { 2540 /* 2541 * DEV_BSIZE aligned new buffer size is less then the 2542 * DEV_BSIZE aligned existing buffer size. Figure out 2543 * if we have to remove any pages. 2544 */ 2545 if (desiredpages < bp->b_xio.xio_npages) { 2546 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) { 2547 /* 2548 * the page is not freed here -- it 2549 * is the responsibility of 2550 * vnode_pager_setsize 2551 */ 2552 m = bp->b_xio.xio_pages[i]; 2553 KASSERT(m != bogus_page, 2554 ("allocbuf: bogus page found")); 2555 while (vm_page_sleep_busy(m, TRUE, "biodep")) 2556 ; 2557 2558 bp->b_xio.xio_pages[i] = NULL; 2559 vm_page_unwire(m, 0); 2560 } 2561 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2562 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages)); 2563 bp->b_xio.xio_npages = desiredpages; 2564 } 2565 } else if (size > bp->b_bcount) { 2566 /* 2567 * We are growing the buffer, possibly in a 2568 * byte-granular fashion. 2569 */ 2570 struct vnode *vp; 2571 vm_object_t obj; 2572 vm_offset_t toff; 2573 vm_offset_t tinc; 2574 2575 /* 2576 * Step 1, bring in the VM pages from the object, 2577 * allocating them if necessary. We must clear 2578 * B_CACHE if these pages are not valid for the 2579 * range covered by the buffer. 2580 * 2581 * spl protection is required to protect against 2582 * interrupts unbusying and freeing pages between 2583 * our vm_page_lookup() and our busycheck/wiring 2584 * call. 2585 */ 2586 vp = bp->b_vp; 2587 VOP_GETVOBJECT(vp, &obj); 2588 2589 crit_enter(); 2590 while (bp->b_xio.xio_npages < desiredpages) { 2591 vm_page_t m; 2592 vm_pindex_t pi; 2593 2594 pi = OFF_TO_IDX(bp->b_offset) + bp->b_xio.xio_npages; 2595 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2596 /* 2597 * note: must allocate system pages 2598 * since blocking here could intefere 2599 * with paging I/O, no matter which 2600 * process we are. 2601 */ 2602 m = vm_page_alloc(obj, pi, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM); 2603 if (m == NULL) { 2604 vm_wait(); 2605 vm_pageout_deficit += desiredpages - 2606 bp->b_xio.xio_npages; 2607 } else { 2608 vm_page_wire(m); 2609 vm_page_wakeup(m); 2610 bp->b_flags &= ~B_CACHE; 2611 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 2612 ++bp->b_xio.xio_npages; 2613 } 2614 continue; 2615 } 2616 2617 /* 2618 * We found a page. If we have to sleep on it, 2619 * retry because it might have gotten freed out 2620 * from under us. 2621 * 2622 * We can only test PG_BUSY here. Blocking on 2623 * m->busy might lead to a deadlock: 2624 * 2625 * vm_fault->getpages->cluster_read->allocbuf 2626 * 2627 */ 2628 2629 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 2630 continue; 2631 2632 /* 2633 * We have a good page. Should we wakeup the 2634 * page daemon? 2635 */ 2636 if ((curthread != pagethread) && 2637 ((m->queue - m->pc) == PQ_CACHE) && 2638 ((vmstats.v_free_count + vmstats.v_cache_count) < 2639 (vmstats.v_free_min + vmstats.v_cache_min))) { 2640 pagedaemon_wakeup(); 2641 } 2642 vm_page_flag_clear(m, PG_ZERO); 2643 vm_page_wire(m); 2644 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 2645 ++bp->b_xio.xio_npages; 2646 } 2647 crit_exit(); 2648 2649 /* 2650 * Step 2. We've loaded the pages into the buffer, 2651 * we have to figure out if we can still have B_CACHE 2652 * set. Note that B_CACHE is set according to the 2653 * byte-granular range ( bcount and size ), new the 2654 * aligned range ( newbsize ). 2655 * 2656 * The VM test is against m->valid, which is DEV_BSIZE 2657 * aligned. Needless to say, the validity of the data 2658 * needs to also be DEV_BSIZE aligned. Note that this 2659 * fails with NFS if the server or some other client 2660 * extends the file's EOF. If our buffer is resized, 2661 * B_CACHE may remain set! XXX 2662 */ 2663 2664 toff = bp->b_bcount; 2665 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2666 2667 while ((bp->b_flags & B_CACHE) && toff < size) { 2668 vm_pindex_t pi; 2669 2670 if (tinc > (size - toff)) 2671 tinc = size - toff; 2672 2673 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2674 PAGE_SHIFT; 2675 2676 vfs_buf_test_cache( 2677 bp, 2678 bp->b_offset, 2679 toff, 2680 tinc, 2681 bp->b_xio.xio_pages[pi] 2682 ); 2683 toff += tinc; 2684 tinc = PAGE_SIZE; 2685 } 2686 2687 /* 2688 * Step 3, fixup the KVM pmap. Remember that 2689 * bp->b_data is relative to bp->b_offset, but 2690 * bp->b_offset may be offset into the first page. 2691 */ 2692 2693 bp->b_data = (caddr_t) 2694 trunc_page((vm_offset_t)bp->b_data); 2695 pmap_qenter( 2696 (vm_offset_t)bp->b_data, 2697 bp->b_xio.xio_pages, 2698 bp->b_xio.xio_npages 2699 ); 2700 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2701 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2702 } 2703 } 2704 if (newbsize < bp->b_bufsize) 2705 bufspacewakeup(); 2706 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2707 bp->b_bcount = size; /* requested buffer size */ 2708 return 1; 2709 } 2710 2711 /* 2712 * biowait: 2713 * 2714 * Wait for buffer I/O completion, returning error status. The buffer 2715 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR 2716 * error and cleared. 2717 */ 2718 int 2719 biowait(struct buf * bp) 2720 { 2721 int s; 2722 2723 s = splbio(); 2724 while ((bp->b_flags & B_DONE) == 0) { 2725 #if defined(NO_SCHEDULE_MODS) 2726 tsleep(bp, 0, "biowait", 0); 2727 #else 2728 if (bp->b_flags & B_READ) 2729 tsleep(bp, 0, "biord", 0); 2730 else 2731 tsleep(bp, 0, "biowr", 0); 2732 #endif 2733 } 2734 splx(s); 2735 if (bp->b_flags & B_EINTR) { 2736 bp->b_flags &= ~B_EINTR; 2737 return (EINTR); 2738 } 2739 if (bp->b_flags & B_ERROR) { 2740 return (bp->b_error ? bp->b_error : EIO); 2741 } else { 2742 return (0); 2743 } 2744 } 2745 2746 /* 2747 * biodone: 2748 * 2749 * Finish I/O on a buffer, optionally calling a completion function. 2750 * This is usually called from an interrupt so process blocking is 2751 * not allowed. 2752 * 2753 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2754 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2755 * assuming B_INVAL is clear. 2756 * 2757 * For the VMIO case, we set B_CACHE if the op was a read and no 2758 * read error occured, or if the op was a write. B_CACHE is never 2759 * set if the buffer is invalid or otherwise uncacheable. 2760 * 2761 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2762 * initiator to leave B_INVAL set to brelse the buffer out of existance 2763 * in the biodone routine. 2764 * 2765 * b_dev is required to be reinitialized prior to the top level strategy 2766 * call in a device stack. To avoid improper reuse, biodone() sets 2767 * b_dev to NODEV. 2768 */ 2769 void 2770 biodone(struct buf *bp) 2771 { 2772 int s, error; 2773 2774 s = splbio(); 2775 2776 KASSERT(BUF_REFCNTNB(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp))); 2777 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 2778 2779 bp->b_flags |= B_DONE; 2780 bp->b_dev = NODEV; 2781 runningbufwakeup(bp); 2782 2783 if (bp->b_flags & B_FREEBUF) { 2784 brelse(bp); 2785 splx(s); 2786 return; 2787 } 2788 2789 if ((bp->b_flags & B_READ) == 0) { 2790 vwakeup(bp); 2791 } 2792 2793 /* call optional completion function if requested */ 2794 if (bp->b_flags & B_CALL) { 2795 bp->b_flags &= ~B_CALL; 2796 (*bp->b_iodone) (bp); 2797 splx(s); 2798 return; 2799 } 2800 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 2801 (*bioops.io_complete)(bp); 2802 2803 if (bp->b_flags & B_VMIO) { 2804 int i; 2805 vm_ooffset_t foff; 2806 vm_page_t m; 2807 vm_object_t obj; 2808 int iosize; 2809 struct vnode *vp = bp->b_vp; 2810 2811 error = VOP_GETVOBJECT(vp, &obj); 2812 2813 #if defined(VFS_BIO_DEBUG) 2814 if (vp->v_holdcnt == 0) { 2815 panic("biodone: zero vnode hold count"); 2816 } 2817 2818 if (error) { 2819 panic("biodone: missing VM object"); 2820 } 2821 2822 if ((vp->v_flag & VOBJBUF) == 0) { 2823 panic("biodone: vnode is not setup for merged cache"); 2824 } 2825 #endif 2826 2827 foff = bp->b_offset; 2828 KASSERT(bp->b_offset != NOOFFSET, 2829 ("biodone: no buffer offset")); 2830 2831 if (error) { 2832 panic("biodone: no object"); 2833 } 2834 #if defined(VFS_BIO_DEBUG) 2835 if (obj->paging_in_progress < bp->b_xio.xio_npages) { 2836 printf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n", 2837 obj->paging_in_progress, bp->b_xio.xio_npages); 2838 } 2839 #endif 2840 2841 /* 2842 * Set B_CACHE if the op was a normal read and no error 2843 * occured. B_CACHE is set for writes in the b*write() 2844 * routines. 2845 */ 2846 iosize = bp->b_bcount - bp->b_resid; 2847 if ((bp->b_flags & (B_READ|B_FREEBUF|B_INVAL|B_NOCACHE|B_ERROR)) == B_READ) { 2848 bp->b_flags |= B_CACHE; 2849 } 2850 2851 for (i = 0; i < bp->b_xio.xio_npages; i++) { 2852 int bogusflag = 0; 2853 int resid; 2854 2855 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2856 if (resid > iosize) 2857 resid = iosize; 2858 2859 /* 2860 * cleanup bogus pages, restoring the originals. Since 2861 * the originals should still be wired, we don't have 2862 * to worry about interrupt/freeing races destroying 2863 * the VM object association. 2864 */ 2865 m = bp->b_xio.xio_pages[i]; 2866 if (m == bogus_page) { 2867 bogusflag = 1; 2868 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2869 if (m == NULL) 2870 panic("biodone: page disappeared"); 2871 bp->b_xio.xio_pages[i] = m; 2872 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 2873 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 2874 } 2875 #if defined(VFS_BIO_DEBUG) 2876 if (OFF_TO_IDX(foff) != m->pindex) { 2877 printf( 2878 "biodone: foff(%lu)/m->pindex(%d) mismatch\n", 2879 (unsigned long)foff, m->pindex); 2880 } 2881 #endif 2882 2883 /* 2884 * In the write case, the valid and clean bits are 2885 * already changed correctly ( see bdwrite() ), so we 2886 * only need to do this here in the read case. 2887 */ 2888 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) { 2889 vfs_page_set_valid(bp, foff, i, m); 2890 } 2891 vm_page_flag_clear(m, PG_ZERO); 2892 2893 /* 2894 * when debugging new filesystems or buffer I/O methods, this 2895 * is the most common error that pops up. if you see this, you 2896 * have not set the page busy flag correctly!!! 2897 */ 2898 if (m->busy == 0) { 2899 printf("biodone: page busy < 0, " 2900 "pindex: %d, foff: 0x(%x,%x), " 2901 "resid: %d, index: %d\n", 2902 (int) m->pindex, (int)(foff >> 32), 2903 (int) foff & 0xffffffff, resid, i); 2904 if (!vn_isdisk(vp, NULL)) 2905 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 2906 bp->b_vp->v_mount->mnt_stat.f_iosize, 2907 (int) bp->b_lblkno, 2908 bp->b_flags, bp->b_xio.xio_npages); 2909 else 2910 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 2911 (int) bp->b_lblkno, 2912 bp->b_flags, bp->b_xio.xio_npages); 2913 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2914 m->valid, m->dirty, m->wire_count); 2915 panic("biodone: page busy < 0"); 2916 } 2917 vm_page_io_finish(m); 2918 vm_object_pip_subtract(obj, 1); 2919 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2920 iosize -= resid; 2921 } 2922 if (obj) 2923 vm_object_pip_wakeupn(obj, 0); 2924 } 2925 2926 /* 2927 * For asynchronous completions, release the buffer now. The brelse 2928 * will do a wakeup there if necessary - so no need to do a wakeup 2929 * here in the async case. The sync case always needs to do a wakeup. 2930 */ 2931 2932 if (bp->b_flags & B_ASYNC) { 2933 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0) 2934 brelse(bp); 2935 else 2936 bqrelse(bp); 2937 } else { 2938 wakeup(bp); 2939 } 2940 splx(s); 2941 } 2942 2943 /* 2944 * This routine is called in lieu of iodone in the case of 2945 * incomplete I/O. This keeps the busy status for pages 2946 * consistant. 2947 */ 2948 void 2949 vfs_unbusy_pages(struct buf *bp) 2950 { 2951 int i; 2952 2953 runningbufwakeup(bp); 2954 if (bp->b_flags & B_VMIO) { 2955 struct vnode *vp = bp->b_vp; 2956 vm_object_t obj; 2957 2958 VOP_GETVOBJECT(vp, &obj); 2959 2960 for (i = 0; i < bp->b_xio.xio_npages; i++) { 2961 vm_page_t m = bp->b_xio.xio_pages[i]; 2962 2963 /* 2964 * When restoring bogus changes the original pages 2965 * should still be wired, so we are in no danger of 2966 * losing the object association and do not need 2967 * spl protection particularly. 2968 */ 2969 if (m == bogus_page) { 2970 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 2971 if (!m) { 2972 panic("vfs_unbusy_pages: page missing"); 2973 } 2974 bp->b_xio.xio_pages[i] = m; 2975 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 2976 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 2977 } 2978 vm_object_pip_subtract(obj, 1); 2979 vm_page_flag_clear(m, PG_ZERO); 2980 vm_page_io_finish(m); 2981 } 2982 vm_object_pip_wakeupn(obj, 0); 2983 } 2984 } 2985 2986 /* 2987 * vfs_page_set_valid: 2988 * 2989 * Set the valid bits in a page based on the supplied offset. The 2990 * range is restricted to the buffer's size. 2991 * 2992 * This routine is typically called after a read completes. 2993 */ 2994 static void 2995 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 2996 { 2997 vm_ooffset_t soff, eoff; 2998 2999 /* 3000 * Start and end offsets in buffer. eoff - soff may not cross a 3001 * page boundry or cross the end of the buffer. The end of the 3002 * buffer, in this case, is our file EOF, not the allocation size 3003 * of the buffer. 3004 */ 3005 soff = off; 3006 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3007 if (eoff > bp->b_offset + bp->b_bcount) 3008 eoff = bp->b_offset + bp->b_bcount; 3009 3010 /* 3011 * Set valid range. This is typically the entire buffer and thus the 3012 * entire page. 3013 */ 3014 if (eoff > soff) { 3015 vm_page_set_validclean( 3016 m, 3017 (vm_offset_t) (soff & PAGE_MASK), 3018 (vm_offset_t) (eoff - soff) 3019 ); 3020 } 3021 } 3022 3023 /* 3024 * This routine is called before a device strategy routine. 3025 * It is used to tell the VM system that paging I/O is in 3026 * progress, and treat the pages associated with the buffer 3027 * almost as being PG_BUSY. Also the object paging_in_progress 3028 * flag is handled to make sure that the object doesn't become 3029 * inconsistant. 3030 * 3031 * Since I/O has not been initiated yet, certain buffer flags 3032 * such as B_ERROR or B_INVAL may be in an inconsistant state 3033 * and should be ignored. 3034 */ 3035 void 3036 vfs_busy_pages(struct buf *bp, int clear_modify) 3037 { 3038 int i, bogus; 3039 3040 if (bp->b_flags & B_VMIO) { 3041 struct vnode *vp = bp->b_vp; 3042 vm_object_t obj; 3043 vm_ooffset_t foff; 3044 3045 VOP_GETVOBJECT(vp, &obj); 3046 foff = bp->b_offset; 3047 KASSERT(bp->b_offset != NOOFFSET, 3048 ("vfs_busy_pages: no buffer offset")); 3049 vfs_setdirty(bp); 3050 3051 retry: 3052 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3053 vm_page_t m = bp->b_xio.xio_pages[i]; 3054 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 3055 goto retry; 3056 } 3057 3058 bogus = 0; 3059 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3060 vm_page_t m = bp->b_xio.xio_pages[i]; 3061 3062 vm_page_flag_clear(m, PG_ZERO); 3063 if ((bp->b_flags & B_CLUSTER) == 0) { 3064 vm_object_pip_add(obj, 1); 3065 vm_page_io_start(m); 3066 } 3067 3068 /* 3069 * When readying a buffer for a read ( i.e 3070 * clear_modify == 0 ), it is important to do 3071 * bogus_page replacement for valid pages in 3072 * partially instantiated buffers. Partially 3073 * instantiated buffers can, in turn, occur when 3074 * reconstituting a buffer from its VM backing store 3075 * base. We only have to do this if B_CACHE is 3076 * clear ( which causes the I/O to occur in the 3077 * first place ). The replacement prevents the read 3078 * I/O from overwriting potentially dirty VM-backed 3079 * pages. XXX bogus page replacement is, uh, bogus. 3080 * It may not work properly with small-block devices. 3081 * We need to find a better way. 3082 */ 3083 3084 vm_page_protect(m, VM_PROT_NONE); 3085 if (clear_modify) 3086 vfs_page_set_valid(bp, foff, i, m); 3087 else if (m->valid == VM_PAGE_BITS_ALL && 3088 (bp->b_flags & B_CACHE) == 0) { 3089 bp->b_xio.xio_pages[i] = bogus_page; 3090 bogus++; 3091 } 3092 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3093 } 3094 if (bogus) 3095 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3096 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3097 } 3098 3099 /* 3100 * This is the easiest place to put the process accounting for the I/O 3101 * for now. 3102 */ 3103 { 3104 struct proc *p; 3105 3106 if ((p = curthread->td_proc) != NULL) { 3107 if (bp->b_flags & B_READ) 3108 p->p_stats->p_ru.ru_inblock++; 3109 else 3110 p->p_stats->p_ru.ru_oublock++; 3111 } 3112 } 3113 } 3114 3115 /* 3116 * Tell the VM system that the pages associated with this buffer 3117 * are clean. This is used for delayed writes where the data is 3118 * going to go to disk eventually without additional VM intevention. 3119 * 3120 * Note that while we only really need to clean through to b_bcount, we 3121 * just go ahead and clean through to b_bufsize. 3122 */ 3123 static void 3124 vfs_clean_pages(struct buf *bp) 3125 { 3126 int i; 3127 3128 if (bp->b_flags & B_VMIO) { 3129 vm_ooffset_t foff; 3130 3131 foff = bp->b_offset; 3132 KASSERT(bp->b_offset != NOOFFSET, 3133 ("vfs_clean_pages: no buffer offset")); 3134 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3135 vm_page_t m = bp->b_xio.xio_pages[i]; 3136 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3137 vm_ooffset_t eoff = noff; 3138 3139 if (eoff > bp->b_offset + bp->b_bufsize) 3140 eoff = bp->b_offset + bp->b_bufsize; 3141 vfs_page_set_valid(bp, foff, i, m); 3142 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3143 foff = noff; 3144 } 3145 } 3146 } 3147 3148 /* 3149 * vfs_bio_set_validclean: 3150 * 3151 * Set the range within the buffer to valid and clean. The range is 3152 * relative to the beginning of the buffer, b_offset. Note that b_offset 3153 * itself may be offset from the beginning of the first page. 3154 */ 3155 3156 void 3157 vfs_bio_set_validclean(struct buf *bp, int base, int size) 3158 { 3159 if (bp->b_flags & B_VMIO) { 3160 int i; 3161 int n; 3162 3163 /* 3164 * Fixup base to be relative to beginning of first page. 3165 * Set initial n to be the maximum number of bytes in the 3166 * first page that can be validated. 3167 */ 3168 3169 base += (bp->b_offset & PAGE_MASK); 3170 n = PAGE_SIZE - (base & PAGE_MASK); 3171 3172 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_xio.xio_npages; ++i) { 3173 vm_page_t m = bp->b_xio.xio_pages[i]; 3174 3175 if (n > size) 3176 n = size; 3177 3178 vm_page_set_validclean(m, base & PAGE_MASK, n); 3179 base += n; 3180 size -= n; 3181 n = PAGE_SIZE; 3182 } 3183 } 3184 } 3185 3186 /* 3187 * vfs_bio_clrbuf: 3188 * 3189 * clear a buffer. This routine essentially fakes an I/O, so we need 3190 * to clear B_ERROR and B_INVAL. 3191 * 3192 * Note that while we only theoretically need to clear through b_bcount, 3193 * we go ahead and clear through b_bufsize. 3194 */ 3195 3196 void 3197 vfs_bio_clrbuf(struct buf *bp) 3198 { 3199 int i, mask = 0; 3200 caddr_t sa, ea; 3201 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3202 bp->b_flags &= ~(B_INVAL|B_ERROR); 3203 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3204 (bp->b_offset & PAGE_MASK) == 0) { 3205 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3206 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) { 3207 bp->b_resid = 0; 3208 return; 3209 } 3210 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) && 3211 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) { 3212 bzero(bp->b_data, bp->b_bufsize); 3213 bp->b_xio.xio_pages[0]->valid |= mask; 3214 bp->b_resid = 0; 3215 return; 3216 } 3217 } 3218 ea = sa = bp->b_data; 3219 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) { 3220 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3221 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3222 ea = (caddr_t)(vm_offset_t)ulmin( 3223 (u_long)(vm_offset_t)ea, 3224 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3225 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3226 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask) 3227 continue; 3228 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) { 3229 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) { 3230 bzero(sa, ea - sa); 3231 } 3232 } else { 3233 for (; sa < ea; sa += DEV_BSIZE, j++) { 3234 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) && 3235 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0) 3236 bzero(sa, DEV_BSIZE); 3237 } 3238 } 3239 bp->b_xio.xio_pages[i]->valid |= mask; 3240 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO); 3241 } 3242 bp->b_resid = 0; 3243 } else { 3244 clrbuf(bp); 3245 } 3246 } 3247 3248 /* 3249 * vm_hold_load_pages and vm_hold_unload pages get pages into 3250 * a buffers address space. The pages are anonymous and are 3251 * not associated with a file object. 3252 */ 3253 void 3254 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3255 { 3256 vm_offset_t pg; 3257 vm_page_t p; 3258 int index; 3259 3260 to = round_page(to); 3261 from = round_page(from); 3262 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3263 3264 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3265 3266 tryagain: 3267 3268 /* 3269 * note: must allocate system pages since blocking here 3270 * could intefere with paging I/O, no matter which 3271 * process we are. 3272 */ 3273 p = vm_page_alloc(kernel_object, 3274 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3275 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM); 3276 if (!p) { 3277 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3278 vm_wait(); 3279 goto tryagain; 3280 } 3281 vm_page_wire(p); 3282 p->valid = VM_PAGE_BITS_ALL; 3283 vm_page_flag_clear(p, PG_ZERO); 3284 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 3285 bp->b_xio.xio_pages[index] = p; 3286 vm_page_wakeup(p); 3287 } 3288 bp->b_xio.xio_npages = index; 3289 } 3290 3291 void 3292 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3293 { 3294 vm_offset_t pg; 3295 vm_page_t p; 3296 int index, newnpages; 3297 3298 from = round_page(from); 3299 to = round_page(to); 3300 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3301 3302 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3303 p = bp->b_xio.xio_pages[index]; 3304 if (p && (index < bp->b_xio.xio_npages)) { 3305 if (p->busy) { 3306 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 3307 bp->b_blkno, bp->b_lblkno); 3308 } 3309 bp->b_xio.xio_pages[index] = NULL; 3310 pmap_kremove(pg); 3311 vm_page_busy(p); 3312 vm_page_unwire(p, 0); 3313 vm_page_free(p); 3314 } 3315 } 3316 bp->b_xio.xio_npages = newnpages; 3317 } 3318 3319 /* 3320 * Map an IO request into kernel virtual address space. 3321 * 3322 * All requests are (re)mapped into kernel VA space. 3323 * Notice that we use b_bufsize for the size of the buffer 3324 * to be mapped. b_bcount might be modified by the driver. 3325 */ 3326 int 3327 vmapbuf(struct buf *bp) 3328 { 3329 caddr_t addr, v, kva; 3330 vm_paddr_t pa; 3331 int pidx; 3332 int i; 3333 struct vm_page *m; 3334 3335 if ((bp->b_flags & B_PHYS) == 0) 3336 panic("vmapbuf"); 3337 if (bp->b_bufsize < 0) 3338 return (-1); 3339 for (v = bp->b_saveaddr, 3340 addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), 3341 pidx = 0; 3342 addr < bp->b_data + bp->b_bufsize; 3343 addr += PAGE_SIZE, v += PAGE_SIZE, pidx++) { 3344 /* 3345 * Do the vm_fault if needed; do the copy-on-write thing 3346 * when reading stuff off device into memory. 3347 */ 3348 retry: 3349 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data, 3350 (bp->b_flags&B_READ)?(VM_PROT_READ|VM_PROT_WRITE):VM_PROT_READ); 3351 if (i < 0) { 3352 for (i = 0; i < pidx; ++i) { 3353 vm_page_unhold(bp->b_xio.xio_pages[i]); 3354 bp->b_xio.xio_pages[i] = NULL; 3355 } 3356 return(-1); 3357 } 3358 3359 /* 3360 * WARNING! If sparc support is MFCd in the future this will 3361 * have to be changed from pmap_kextract() to pmap_extract() 3362 * ala -current. 3363 */ 3364 #ifdef __sparc64__ 3365 #error "If MFCing sparc support use pmap_extract" 3366 #endif 3367 pa = pmap_kextract((vm_offset_t)addr); 3368 if (pa == 0) { 3369 printf("vmapbuf: warning, race against user address during I/O"); 3370 goto retry; 3371 } 3372 m = PHYS_TO_VM_PAGE(pa); 3373 vm_page_hold(m); 3374 bp->b_xio.xio_pages[pidx] = m; 3375 } 3376 if (pidx > btoc(MAXPHYS)) 3377 panic("vmapbuf: mapped more than MAXPHYS"); 3378 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_xio.xio_pages, pidx); 3379 3380 kva = bp->b_saveaddr; 3381 bp->b_xio.xio_npages = pidx; 3382 bp->b_saveaddr = bp->b_data; 3383 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3384 return(0); 3385 } 3386 3387 /* 3388 * Free the io map PTEs associated with this IO operation. 3389 * We also invalidate the TLB entries and restore the original b_addr. 3390 */ 3391 void 3392 vunmapbuf(struct buf *bp) 3393 { 3394 int pidx; 3395 int npages; 3396 vm_page_t *m; 3397 3398 if ((bp->b_flags & B_PHYS) == 0) 3399 panic("vunmapbuf"); 3400 3401 npages = bp->b_xio.xio_npages; 3402 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3403 npages); 3404 m = bp->b_xio.xio_pages; 3405 for (pidx = 0; pidx < npages; pidx++) 3406 vm_page_unhold(*m++); 3407 3408 bp->b_data = bp->b_saveaddr; 3409 } 3410 3411 #include "opt_ddb.h" 3412 #ifdef DDB 3413 #include <ddb/ddb.h> 3414 3415 DB_SHOW_COMMAND(buffer, db_show_buffer) 3416 { 3417 /* get args */ 3418 struct buf *bp = (struct buf *)addr; 3419 3420 if (!have_addr) { 3421 db_printf("usage: show buffer <addr>\n"); 3422 return; 3423 } 3424 3425 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3426 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 3427 "b_resid = %ld\nb_dev = (%d,%d), b_data = %p, " 3428 "b_blkno = %d, b_pblkno = %d\n", 3429 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3430 major(bp->b_dev), minor(bp->b_dev), 3431 bp->b_data, bp->b_blkno, bp->b_pblkno); 3432 if (bp->b_xio.xio_npages) { 3433 int i; 3434 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ", 3435 bp->b_xio.xio_npages); 3436 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3437 vm_page_t m; 3438 m = bp->b_xio.xio_pages[i]; 3439 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3440 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3441 if ((i + 1) < bp->b_xio.xio_npages) 3442 db_printf(","); 3443 } 3444 db_printf("\n"); 3445 } 3446 } 3447 #endif /* DDB */ 3448