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 1.54 2006/02/17 19:18:06 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 /* 62 * Buffer queues. 63 */ 64 #define BUFFER_QUEUES 6 65 enum bufq_type { 66 BQUEUE_NONE, /* not on any queue */ 67 BQUEUE_LOCKED, /* locked buffers */ 68 BQUEUE_CLEAN, /* non-B_DELWRI buffers */ 69 BQUEUE_DIRTY, /* B_DELWRI buffers */ 70 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */ 71 BQUEUE_EMPTY /* empty buffer headers */ 72 }; 73 TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES]; 74 75 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 76 77 struct bio_ops bioops; /* I/O operation notification */ 78 79 struct buf *buf; /* buffer header pool */ 80 81 static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 82 vm_offset_t to); 83 static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 84 vm_offset_t to); 85 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 86 int pageno, vm_page_t m); 87 static void vfs_clean_pages(struct buf * bp); 88 static void vfs_setdirty(struct buf *bp); 89 static void vfs_vmio_release(struct buf *bp); 90 static int flushbufqueues(void); 91 92 static int bd_request; 93 94 static void buf_daemon (void); 95 /* 96 * bogus page -- for I/O to/from partially complete buffers 97 * this is a temporary solution to the problem, but it is not 98 * really that bad. it would be better to split the buffer 99 * for input in the case of buffers partially already in memory, 100 * but the code is intricate enough already. 101 */ 102 vm_page_t bogus_page; 103 int vmiodirenable = TRUE; 104 int runningbufspace; 105 struct spinlock buftimespinlock; /* Interlock on setting prio and timo */ 106 107 static int bufspace, maxbufspace, 108 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace; 109 static int bufreusecnt, bufdefragcnt, buffreekvacnt; 110 static int needsbuffer; 111 static int lorunningspace, hirunningspace, runningbufreq; 112 static int numdirtybuffers, lodirtybuffers, hidirtybuffers; 113 static int numfreebuffers, lofreebuffers, hifreebuffers; 114 static int getnewbufcalls; 115 static int getnewbufrestarts; 116 117 /* 118 * Sysctls for operational control of the buffer cache. 119 */ 120 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 121 "Number of dirty buffers to flush before bufdaemon becomes inactive"); 122 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 123 "High watermark used to trigger explicit flushing of dirty buffers"); 124 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 125 "Low watermark for special reserve in low-memory situations"); 126 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 127 "High watermark for special reserve in low-memory situations"); 128 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 129 "Minimum amount of buffer space required for active I/O"); 130 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 131 "Maximum amount of buffer space to usable for active I/O"); 132 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 133 "Use the VM system for performing directory writes"); 134 /* 135 * Sysctls determining current state of the buffer cache. 136 */ 137 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 138 "Pending number of dirty buffers"); 139 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 140 "Number of free buffers on the buffer cache free list"); 141 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 142 "I/O bytes currently in progress due to asynchronous writes"); 143 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 144 "Hard limit on maximum amount of memory usable for buffer space"); 145 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 146 "Soft limit on maximum amount of memory usable for buffer space"); 147 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 148 "Minimum amount of memory to reserve for system buffer space"); 149 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 150 "Amount of memory available for buffers"); 151 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace, 152 0, "Maximum amount of memory reserved for buffers using malloc"); 153 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 154 "Amount of memory left for buffers using malloc-scheme"); 155 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0, 156 "New buffer header acquisition requests"); 157 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts, 158 0, "New buffer header acquisition restarts"); 159 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0, 160 "Buffer acquisition restarts due to fragmented buffer map"); 161 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0, 162 "Amount of time KVA space was deallocated in an arbitrary buffer"); 163 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0, 164 "Amount of time buffer re-use operations were successful"); 165 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf), 166 "sizeof(struct buf)"); 167 168 static int bufhashmask; 169 static int bufhashshift; 170 static LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash; 171 char *buf_wmesg = BUF_WMESG; 172 173 extern int vm_swap_size; 174 175 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 176 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 177 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 178 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 179 180 /* 181 * Buffer hash table code. Note that the logical block scans linearly, which 182 * gives us some L1 cache locality. 183 */ 184 185 static __inline 186 struct bufhashhdr * 187 bufhash(struct vnode *vnp, daddr_t bn) 188 { 189 u_int64_t hashkey64; 190 int hashkey; 191 192 /* 193 * A variation on the Fibonacci hash that Knuth credits to 194 * R. W. Floyd, see Knuth's _Art of Computer Programming, 195 * Volume 3 / Sorting and Searching_ 196 * 197 * We reduce the argument to 32 bits before doing the hash to 198 * avoid the need for a slow 64x64 multiply on 32 bit platforms. 199 * 200 * sizeof(struct vnode) is 168 on i386, so toss some of the lower 201 * bits of the vnode address to reduce the key range, which 202 * improves the distribution of keys across buckets. 203 * 204 * The file system cylinder group blocks are very heavily 205 * used. They are located at invervals of fbg, which is 206 * on the order of 89 to 94 * 2^10, depending on other 207 * filesystem parameters, for a 16k block size. Smaller block 208 * sizes will reduce fpg approximately proportionally. This 209 * will cause the cylinder group index to be hashed using the 210 * lower bits of the hash multiplier, which will not distribute 211 * the keys as uniformly in a classic Fibonacci hash where a 212 * relatively small number of the upper bits of the result 213 * are used. Using 2^16 as a close-enough approximation to 214 * fpg, split the hash multiplier in half, with the upper 16 215 * bits being the inverse of the golden ratio, and the lower 216 * 16 bits being a fraction between 1/3 and 3/7 (closer to 217 * 3/7 in this case), that gives good experimental results. 218 */ 219 hashkey64 = ((u_int64_t)(uintptr_t)vnp >> 3) + (u_int64_t)bn; 220 hashkey = (((u_int32_t)(hashkey64 + (hashkey64 >> 32)) * 0x9E376DB1u) >> 221 bufhashshift) & bufhashmask; 222 return(&bufhashtbl[hashkey]); 223 } 224 225 /* 226 * numdirtywakeup: 227 * 228 * If someone is blocked due to there being too many dirty buffers, 229 * and numdirtybuffers is now reasonable, wake them up. 230 */ 231 232 static __inline void 233 numdirtywakeup(int level) 234 { 235 if (numdirtybuffers <= level) { 236 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 237 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 238 wakeup(&needsbuffer); 239 } 240 } 241 } 242 243 /* 244 * bufspacewakeup: 245 * 246 * Called when buffer space is potentially available for recovery. 247 * getnewbuf() will block on this flag when it is unable to free 248 * sufficient buffer space. Buffer space becomes recoverable when 249 * bp's get placed back in the queues. 250 */ 251 252 static __inline void 253 bufspacewakeup(void) 254 { 255 /* 256 * If someone is waiting for BUF space, wake them up. Even 257 * though we haven't freed the kva space yet, the waiting 258 * process will be able to now. 259 */ 260 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 261 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 262 wakeup(&needsbuffer); 263 } 264 } 265 266 /* 267 * runningbufwakeup: 268 * 269 * Accounting for I/O in progress. 270 * 271 */ 272 static __inline void 273 runningbufwakeup(struct buf *bp) 274 { 275 if (bp->b_runningbufspace) { 276 runningbufspace -= bp->b_runningbufspace; 277 bp->b_runningbufspace = 0; 278 if (runningbufreq && runningbufspace <= lorunningspace) { 279 runningbufreq = 0; 280 wakeup(&runningbufreq); 281 } 282 } 283 } 284 285 /* 286 * bufcountwakeup: 287 * 288 * Called when a buffer has been added to one of the free queues to 289 * account for the buffer and to wakeup anyone waiting for free buffers. 290 * This typically occurs when large amounts of metadata are being handled 291 * by the buffer cache ( else buffer space runs out first, usually ). 292 */ 293 294 static __inline void 295 bufcountwakeup(void) 296 { 297 ++numfreebuffers; 298 if (needsbuffer) { 299 needsbuffer &= ~VFS_BIO_NEED_ANY; 300 if (numfreebuffers >= hifreebuffers) 301 needsbuffer &= ~VFS_BIO_NEED_FREE; 302 wakeup(&needsbuffer); 303 } 304 } 305 306 /* 307 * waitrunningbufspace() 308 * 309 * runningbufspace is a measure of the amount of I/O currently 310 * running. This routine is used in async-write situations to 311 * prevent creating huge backups of pending writes to a device. 312 * Only asynchronous writes are governed by this function. 313 * 314 * Reads will adjust runningbufspace, but will not block based on it. 315 * The read load has a side effect of reducing the allowed write load. 316 * 317 * This does NOT turn an async write into a sync write. It waits 318 * for earlier writes to complete and generally returns before the 319 * caller's write has reached the device. 320 */ 321 static __inline void 322 waitrunningbufspace(void) 323 { 324 if (runningbufspace > hirunningspace) { 325 crit_enter(); 326 while (runningbufspace > hirunningspace) { 327 ++runningbufreq; 328 tsleep(&runningbufreq, 0, "wdrain", 0); 329 } 330 crit_exit(); 331 } 332 } 333 334 /* 335 * vfs_buf_test_cache: 336 * 337 * Called when a buffer is extended. This function clears the B_CACHE 338 * bit if the newly extended portion of the buffer does not contain 339 * valid data. 340 */ 341 static __inline__ 342 void 343 vfs_buf_test_cache(struct buf *bp, 344 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 345 vm_page_t m) 346 { 347 if (bp->b_flags & B_CACHE) { 348 int base = (foff + off) & PAGE_MASK; 349 if (vm_page_is_valid(m, base, size) == 0) 350 bp->b_flags &= ~B_CACHE; 351 } 352 } 353 354 /* 355 * bd_wakeup: 356 * 357 * Wake up the buffer daemon if the number of outstanding dirty buffers 358 * is above specified threshold 'dirtybuflevel'. 359 * 360 * The buffer daemon is explicitly woken up when (a) the pending number 361 * of dirty buffers exceeds the recovery and stall mid-point value, 362 * (b) during bwillwrite() or (c) buf freelist was exhausted. 363 */ 364 static __inline__ 365 void 366 bd_wakeup(int dirtybuflevel) 367 { 368 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 369 bd_request = 1; 370 wakeup(&bd_request); 371 } 372 } 373 374 /* 375 * bd_speedup: 376 * 377 * Speed up the buffer cache flushing process. 378 */ 379 380 static __inline__ 381 void 382 bd_speedup(void) 383 { 384 bd_wakeup(1); 385 } 386 387 /* 388 * bufhashinit: 389 * 390 * Initialize buffer headers and related structures. 391 */ 392 393 caddr_t 394 bufhashinit(caddr_t vaddr) 395 { 396 /* first, make a null hash table */ 397 bufhashshift = 29; 398 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 399 bufhashshift--; 400 bufhashtbl = (void *)vaddr; 401 vaddr = vaddr + sizeof(*bufhashtbl) * bufhashmask; 402 --bufhashmask; 403 return(vaddr); 404 } 405 406 /* 407 * bufinit: 408 * 409 * Load time initialisation of the buffer cache, called from machine 410 * dependant initialization code. 411 */ 412 void 413 bufinit(void) 414 { 415 struct buf *bp; 416 vm_offset_t bogus_offset; 417 int i; 418 419 LIST_INIT(&invalhash); 420 spin_init(&buftimespinlock); 421 422 for (i = 0; i <= bufhashmask; i++) 423 LIST_INIT(&bufhashtbl[i]); 424 425 /* next, make a null set of free lists */ 426 for (i = 0; i < BUFFER_QUEUES; i++) 427 TAILQ_INIT(&bufqueues[i]); 428 429 /* finally, initialize each buffer header and stick on empty q */ 430 for (i = 0; i < nbuf; i++) { 431 bp = &buf[i]; 432 bzero(bp, sizeof *bp); 433 bp->b_flags = B_INVAL; /* we're just an empty header */ 434 bp->b_qindex = BQUEUE_EMPTY; 435 initbufbio(bp); 436 xio_init(&bp->b_xio); 437 LIST_INIT(&bp->b_dep); 438 BUF_LOCKINIT(bp); 439 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist); 440 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 441 } 442 443 /* 444 * maxbufspace is the absolute maximum amount of buffer space we are 445 * allowed to reserve in KVM and in real terms. The absolute maximum 446 * is nominally used by buf_daemon. hibufspace is the nominal maximum 447 * used by most other processes. The differential is required to 448 * ensure that buf_daemon is able to run when other processes might 449 * be blocked waiting for buffer space. 450 * 451 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 452 * this may result in KVM fragmentation which is not handled optimally 453 * by the system. 454 */ 455 maxbufspace = nbuf * BKVASIZE; 456 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 457 lobufspace = hibufspace - MAXBSIZE; 458 459 lorunningspace = 512 * 1024; 460 hirunningspace = 1024 * 1024; 461 462 /* 463 * Limit the amount of malloc memory since it is wired permanently into 464 * the kernel space. Even though this is accounted for in the buffer 465 * allocation, we don't want the malloced region to grow uncontrolled. 466 * The malloc scheme improves memory utilization significantly on average 467 * (small) directories. 468 */ 469 maxbufmallocspace = hibufspace / 20; 470 471 /* 472 * Reduce the chance of a deadlock occuring by limiting the number 473 * of delayed-write dirty buffers we allow to stack up. 474 */ 475 hidirtybuffers = nbuf / 4 + 20; 476 numdirtybuffers = 0; 477 /* 478 * To support extreme low-memory systems, make sure hidirtybuffers cannot 479 * eat up all available buffer space. This occurs when our minimum cannot 480 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 481 * BKVASIZE'd (8K) buffers. 482 */ 483 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 484 hidirtybuffers >>= 1; 485 } 486 lodirtybuffers = hidirtybuffers / 2; 487 488 /* 489 * Try to keep the number of free buffers in the specified range, 490 * and give special processes (e.g. like buf_daemon) access to an 491 * emergency reserve. 492 */ 493 lofreebuffers = nbuf / 18 + 5; 494 hifreebuffers = 2 * lofreebuffers; 495 numfreebuffers = nbuf; 496 497 /* 498 * Maximum number of async ops initiated per buf_daemon loop. This is 499 * somewhat of a hack at the moment, we really need to limit ourselves 500 * based on the number of bytes of I/O in-transit that were initiated 501 * from buf_daemon. 502 */ 503 504 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 505 bogus_page = vm_page_alloc(kernel_object, 506 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 507 VM_ALLOC_NORMAL); 508 vmstats.v_wire_count++; 509 510 } 511 512 /* 513 * Initialize the embedded bio structures 514 */ 515 void 516 initbufbio(struct buf *bp) 517 { 518 bp->b_bio1.bio_buf = bp; 519 bp->b_bio1.bio_prev = NULL; 520 bp->b_bio1.bio_blkno = (daddr_t)-1; 521 bp->b_bio1.bio_offset = NOOFFSET; 522 bp->b_bio1.bio_next = &bp->b_bio2; 523 bp->b_bio1.bio_done = NULL; 524 525 bp->b_bio2.bio_buf = bp; 526 bp->b_bio2.bio_prev = &bp->b_bio1; 527 bp->b_bio2.bio_blkno = (daddr_t)-1; 528 bp->b_bio2.bio_offset = NOOFFSET; 529 bp->b_bio2.bio_next = NULL; 530 bp->b_bio2.bio_done = NULL; 531 } 532 533 /* 534 * Reinitialize the embedded bio structures as well as any additional 535 * translation cache layers. 536 */ 537 void 538 reinitbufbio(struct buf *bp) 539 { 540 struct bio *bio; 541 542 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) { 543 bio->bio_done = NULL; 544 bio->bio_blkno = (daddr_t)-1; 545 bio->bio_offset = NOOFFSET; 546 } 547 } 548 549 /* 550 * Push another BIO layer onto an existing BIO and return it. The new 551 * BIO layer may already exist, holding cached translation data. 552 */ 553 struct bio * 554 push_bio(struct bio *bio) 555 { 556 struct bio *nbio; 557 558 if ((nbio = bio->bio_next) == NULL) { 559 int index = bio - &bio->bio_buf->b_bio_array[0]; 560 if (index >= NBUF_BIO) { 561 panic("push_bio: too many layers bp %p\n", 562 bio->bio_buf); 563 } 564 nbio = &bio->bio_buf->b_bio_array[index + 1]; 565 bio->bio_next = nbio; 566 nbio->bio_prev = bio; 567 nbio->bio_buf = bio->bio_buf; 568 nbio->bio_blkno = (daddr_t)-1; 569 nbio->bio_offset = NOOFFSET; 570 nbio->bio_done = NULL; 571 nbio->bio_next = NULL; 572 } 573 KKASSERT(nbio->bio_done == NULL); 574 return(nbio); 575 } 576 577 void 578 pop_bio(struct bio *bio) 579 { 580 /* NOP */ 581 } 582 583 void 584 clearbiocache(struct bio *bio) 585 { 586 while (bio) { 587 bio->bio_blkno = (daddr_t)-1; 588 bio->bio_offset = NOOFFSET; 589 bio = bio->bio_next; 590 } 591 } 592 593 /* 594 * bfreekva: 595 * 596 * Free the KVA allocation for buffer 'bp'. 597 * 598 * Must be called from a critical section as this is the only locking for 599 * buffer_map. 600 * 601 * Since this call frees up buffer space, we call bufspacewakeup(). 602 */ 603 static void 604 bfreekva(struct buf * bp) 605 { 606 int count; 607 608 if (bp->b_kvasize) { 609 ++buffreekvacnt; 610 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 611 vm_map_lock(buffer_map); 612 bufspace -= bp->b_kvasize; 613 vm_map_delete(buffer_map, 614 (vm_offset_t) bp->b_kvabase, 615 (vm_offset_t) bp->b_kvabase + bp->b_kvasize, 616 &count 617 ); 618 vm_map_unlock(buffer_map); 619 vm_map_entry_release(count); 620 bp->b_kvasize = 0; 621 bufspacewakeup(); 622 } 623 } 624 625 /* 626 * bremfree: 627 * 628 * Remove the buffer from the appropriate free list. 629 */ 630 void 631 bremfree(struct buf * bp) 632 { 633 int old_qindex; 634 635 crit_enter(); 636 old_qindex = bp->b_qindex; 637 638 if (bp->b_qindex != BQUEUE_NONE) { 639 KASSERT(BUF_REFCNTNB(bp) == 1, 640 ("bremfree: bp %p not locked",bp)); 641 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 642 bp->b_qindex = BQUEUE_NONE; 643 } else { 644 if (BUF_REFCNTNB(bp) <= 1) 645 panic("bremfree: removing a buffer not on a queue"); 646 } 647 648 /* 649 * Fixup numfreebuffers count. If the buffer is invalid or not 650 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 651 * the buffer was free and we must decrement numfreebuffers. 652 */ 653 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 654 switch(old_qindex) { 655 case BQUEUE_DIRTY: 656 case BQUEUE_CLEAN: 657 case BQUEUE_EMPTY: 658 case BQUEUE_EMPTYKVA: 659 --numfreebuffers; 660 break; 661 default: 662 break; 663 } 664 } 665 crit_exit(); 666 } 667 668 669 /* 670 * bread: 671 * 672 * Get a buffer with the specified data. Look in the cache first. We 673 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 674 * is set, the buffer is valid and we do not have to do anything ( see 675 * getblk() ). 676 */ 677 int 678 bread(struct vnode * vp, daddr_t blkno, int size, struct buf ** bpp) 679 { 680 struct buf *bp; 681 682 bp = getblk(vp, blkno, size, 0, 0); 683 *bpp = bp; 684 685 /* if not found in cache, do some I/O */ 686 if ((bp->b_flags & B_CACHE) == 0) { 687 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp)); 688 bp->b_flags |= B_READ; 689 bp->b_flags &= ~(B_ERROR | B_INVAL); 690 vfs_busy_pages(bp, 0); 691 vn_strategy(vp, &bp->b_bio1); 692 return (biowait(bp)); 693 } 694 return (0); 695 } 696 697 /* 698 * breadn: 699 * 700 * Operates like bread, but also starts asynchronous I/O on 701 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior 702 * to initiating I/O . If B_CACHE is set, the buffer is valid 703 * and we do not have to do anything. 704 */ 705 int 706 breadn(struct vnode * vp, daddr_t blkno, int size, daddr_t * rablkno, 707 int *rabsize, int cnt, struct buf ** bpp) 708 { 709 struct buf *bp, *rabp; 710 int i; 711 int rv = 0, readwait = 0; 712 713 *bpp = bp = getblk(vp, blkno, size, 0, 0); 714 715 /* if not found in cache, do some I/O */ 716 if ((bp->b_flags & B_CACHE) == 0) { 717 bp->b_flags |= B_READ; 718 bp->b_flags &= ~(B_ERROR | B_INVAL); 719 vfs_busy_pages(bp, 0); 720 vn_strategy(vp, &bp->b_bio1); 721 ++readwait; 722 } 723 724 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 725 if (inmem(vp, *rablkno)) 726 continue; 727 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 728 729 if ((rabp->b_flags & B_CACHE) == 0) { 730 rabp->b_flags |= B_READ | B_ASYNC; 731 rabp->b_flags &= ~(B_ERROR | B_INVAL); 732 vfs_busy_pages(rabp, 0); 733 BUF_KERNPROC(rabp); 734 vn_strategy(vp, &rabp->b_bio1); 735 } else { 736 brelse(rabp); 737 } 738 } 739 740 if (readwait) { 741 rv = biowait(bp); 742 } 743 return (rv); 744 } 745 746 /* 747 * bwrite: 748 * 749 * Write, release buffer on completion. (Done by iodone 750 * if async). Do not bother writing anything if the buffer 751 * is invalid. 752 * 753 * Note that we set B_CACHE here, indicating that buffer is 754 * fully valid and thus cacheable. This is true even of NFS 755 * now so we set it generally. This could be set either here 756 * or in biodone() since the I/O is synchronous. We put it 757 * here. 758 */ 759 int 760 bwrite(struct buf * bp) 761 { 762 int oldflags; 763 764 if (bp->b_flags & B_INVAL) { 765 brelse(bp); 766 return (0); 767 } 768 769 oldflags = bp->b_flags; 770 771 if (BUF_REFCNTNB(bp) == 0) 772 panic("bwrite: buffer is not busy???"); 773 crit_enter(); 774 /* 775 * If a background write is already in progress, delay 776 * writing this block if it is asynchronous. Otherwise 777 * wait for the background write to complete. 778 */ 779 if (bp->b_xflags & BX_BKGRDINPROG) { 780 if (bp->b_flags & B_ASYNC) { 781 crit_exit(); 782 bdwrite(bp); 783 return (0); 784 } 785 bp->b_xflags |= BX_BKGRDWAIT; 786 tsleep(&bp->b_xflags, 0, "biord", 0); 787 if (bp->b_xflags & BX_BKGRDINPROG) 788 panic("bwrite: still writing"); 789 } 790 791 /* Mark the buffer clean */ 792 bundirty(bp); 793 794 bp->b_flags &= ~(B_READ | B_DONE | B_ERROR); 795 bp->b_flags |= B_CACHE; 796 797 vfs_busy_pages(bp, 1); 798 799 /* 800 * Normal bwrites pipeline writes 801 */ 802 bp->b_runningbufspace = bp->b_bufsize; 803 runningbufspace += bp->b_runningbufspace; 804 805 crit_exit(); 806 if (oldflags & B_ASYNC) 807 BUF_KERNPROC(bp); 808 vn_strategy(bp->b_vp, &bp->b_bio1); 809 810 if ((oldflags & B_ASYNC) == 0) { 811 int rtval = biowait(bp); 812 brelse(bp); 813 return (rtval); 814 } else if ((oldflags & B_NOWDRAIN) == 0) { 815 /* 816 * don't allow the async write to saturate the I/O 817 * system. Deadlocks can occur only if a device strategy 818 * routine (like in VN) turns around and issues another 819 * high-level write, in which case B_NOWDRAIN is expected 820 * to be set. Otherwise we will not deadlock here because 821 * we are blocking waiting for I/O that is already in-progress 822 * to complete. 823 */ 824 waitrunningbufspace(); 825 } 826 827 return (0); 828 } 829 830 /* 831 * bdwrite: 832 * 833 * Delayed write. (Buffer is marked dirty). Do not bother writing 834 * anything if the buffer is marked invalid. 835 * 836 * Note that since the buffer must be completely valid, we can safely 837 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 838 * biodone() in order to prevent getblk from writing the buffer 839 * out synchronously. 840 */ 841 void 842 bdwrite(struct buf *bp) 843 { 844 if (BUF_REFCNTNB(bp) == 0) 845 panic("bdwrite: buffer is not busy"); 846 847 if (bp->b_flags & B_INVAL) { 848 brelse(bp); 849 return; 850 } 851 bdirty(bp); 852 853 /* 854 * Set B_CACHE, indicating that the buffer is fully valid. This is 855 * true even of NFS now. 856 */ 857 bp->b_flags |= B_CACHE; 858 859 /* 860 * This bmap keeps the system from needing to do the bmap later, 861 * perhaps when the system is attempting to do a sync. Since it 862 * is likely that the indirect block -- or whatever other datastructure 863 * that the filesystem needs is still in memory now, it is a good 864 * thing to do this. Note also, that if the pageout daemon is 865 * requesting a sync -- there might not be enough memory to do 866 * the bmap then... So, this is important to do. 867 */ 868 if (bp->b_bio2.bio_blkno == (daddr_t)-1) { 869 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_bio2.bio_blkno, 870 NULL, NULL); 871 } 872 873 /* 874 * Set the *dirty* buffer range based upon the VM system dirty pages. 875 */ 876 vfs_setdirty(bp); 877 878 /* 879 * We need to do this here to satisfy the vnode_pager and the 880 * pageout daemon, so that it thinks that the pages have been 881 * "cleaned". Note that since the pages are in a delayed write 882 * buffer -- the VFS layer "will" see that the pages get written 883 * out on the next sync, or perhaps the cluster will be completed. 884 */ 885 vfs_clean_pages(bp); 886 bqrelse(bp); 887 888 /* 889 * Wakeup the buffer flushing daemon if we have a lot of dirty 890 * buffers (midpoint between our recovery point and our stall 891 * point). 892 */ 893 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 894 895 /* 896 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 897 * due to the softdep code. 898 */ 899 } 900 901 /* 902 * bdirty: 903 * 904 * Turn buffer into delayed write request. We must clear B_READ and 905 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 906 * itself to properly update it in the dirty/clean lists. We mark it 907 * B_DONE to ensure that any asynchronization of the buffer properly 908 * clears B_DONE ( else a panic will occur later ). 909 * 910 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 911 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 912 * should only be called if the buffer is known-good. 913 * 914 * Since the buffer is not on a queue, we do not update the numfreebuffers 915 * count. 916 * 917 * Must be called from a critical section. 918 * The buffer must be on BQUEUE_NONE. 919 */ 920 void 921 bdirty(struct buf *bp) 922 { 923 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 924 bp->b_flags &= ~(B_READ|B_RELBUF); 925 926 if ((bp->b_flags & B_DELWRI) == 0) { 927 bp->b_flags |= B_DONE | B_DELWRI; 928 reassignbuf(bp, bp->b_vp); 929 ++numdirtybuffers; 930 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 931 } 932 } 933 934 /* 935 * bundirty: 936 * 937 * Clear B_DELWRI for buffer. 938 * 939 * Since the buffer is not on a queue, we do not update the numfreebuffers 940 * count. 941 * 942 * Must be called from a critical section. 943 * 944 * The buffer is typically on BQUEUE_NONE but there is one case in 945 * brelse() that calls this function after placing the buffer on 946 * a different queue. 947 */ 948 949 void 950 bundirty(struct buf *bp) 951 { 952 if (bp->b_flags & B_DELWRI) { 953 bp->b_flags &= ~B_DELWRI; 954 reassignbuf(bp, bp->b_vp); 955 --numdirtybuffers; 956 numdirtywakeup(lodirtybuffers); 957 } 958 /* 959 * Since it is now being written, we can clear its deferred write flag. 960 */ 961 bp->b_flags &= ~B_DEFERRED; 962 } 963 964 /* 965 * bawrite: 966 * 967 * Asynchronous write. Start output on a buffer, but do not wait for 968 * it to complete. The buffer is released when the output completes. 969 * 970 * bwrite() ( or the VOP routine anyway ) is responsible for handling 971 * B_INVAL buffers. Not us. 972 */ 973 void 974 bawrite(struct buf * bp) 975 { 976 bp->b_flags |= B_ASYNC; 977 (void) VOP_BWRITE(bp->b_vp, bp); 978 } 979 980 /* 981 * bowrite: 982 * 983 * Ordered write. Start output on a buffer, and flag it so that the 984 * device will write it in the order it was queued. The buffer is 985 * released when the output completes. bwrite() ( or the VOP routine 986 * anyway ) is responsible for handling B_INVAL buffers. 987 */ 988 int 989 bowrite(struct buf * bp) 990 { 991 bp->b_flags |= B_ORDERED | B_ASYNC; 992 return (VOP_BWRITE(bp->b_vp, bp)); 993 } 994 995 /* 996 * bwillwrite: 997 * 998 * Called prior to the locking of any vnodes when we are expecting to 999 * write. We do not want to starve the buffer cache with too many 1000 * dirty buffers so we block here. By blocking prior to the locking 1001 * of any vnodes we attempt to avoid the situation where a locked vnode 1002 * prevents the various system daemons from flushing related buffers. 1003 */ 1004 1005 void 1006 bwillwrite(void) 1007 { 1008 if (numdirtybuffers >= hidirtybuffers) { 1009 crit_enter(); 1010 while (numdirtybuffers >= hidirtybuffers) { 1011 bd_wakeup(1); 1012 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1013 tsleep(&needsbuffer, 0, "flswai", 0); 1014 } 1015 crit_exit(); 1016 } 1017 } 1018 1019 /* 1020 * buf_dirty_count_severe: 1021 * 1022 * Return true if we have too many dirty buffers. 1023 */ 1024 int 1025 buf_dirty_count_severe(void) 1026 { 1027 return(numdirtybuffers >= hidirtybuffers); 1028 } 1029 1030 /* 1031 * brelse: 1032 * 1033 * Release a busy buffer and, if requested, free its resources. The 1034 * buffer will be stashed in the appropriate bufqueue[] allowing it 1035 * to be accessed later as a cache entity or reused for other purposes. 1036 */ 1037 void 1038 brelse(struct buf * bp) 1039 { 1040 #ifdef INVARIANTS 1041 int saved_flags = bp->b_flags; 1042 #endif 1043 1044 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1045 1046 crit_enter(); 1047 1048 if (bp->b_flags & B_LOCKED) 1049 bp->b_flags &= ~B_ERROR; 1050 1051 if ((bp->b_flags & (B_READ | B_ERROR | B_INVAL)) == B_ERROR) { 1052 /* 1053 * Failed write, redirty. Must clear B_ERROR to prevent 1054 * pages from being scrapped. If B_INVAL is set then 1055 * this case is not run and the next case is run to 1056 * destroy the buffer. B_INVAL can occur if the buffer 1057 * is outside the range supported by the underlying device. 1058 */ 1059 bp->b_flags &= ~B_ERROR; 1060 bdirty(bp); 1061 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_FREEBUF)) || 1062 (bp->b_bufsize <= 0)) { 1063 /* 1064 * Either a failed I/O or we were asked to free or not 1065 * cache the buffer. 1066 */ 1067 bp->b_flags |= B_INVAL; 1068 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 1069 (*bioops.io_deallocate)(bp); 1070 if (bp->b_flags & B_DELWRI) { 1071 --numdirtybuffers; 1072 numdirtywakeup(lodirtybuffers); 1073 } 1074 bp->b_flags &= ~(B_DELWRI | B_CACHE | B_FREEBUF); 1075 } 1076 1077 /* 1078 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1079 * is called with B_DELWRI set, the underlying pages may wind up 1080 * getting freed causing a previous write (bdwrite()) to get 'lost' 1081 * because pages associated with a B_DELWRI bp are marked clean. 1082 * 1083 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1084 * if B_DELWRI is set. 1085 * 1086 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1087 * on pages to return pages to the VM page queues. 1088 */ 1089 if (bp->b_flags & B_DELWRI) 1090 bp->b_flags &= ~B_RELBUF; 1091 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1092 bp->b_flags |= B_RELBUF; 1093 1094 /* 1095 * At this point destroying the buffer is governed by the B_INVAL 1096 * or B_RELBUF flags. 1097 */ 1098 1099 /* 1100 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1101 * constituted, not even NFS buffers now. Two flags effect this. If 1102 * B_INVAL, the struct buf is invalidated but the VM object is kept 1103 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1104 * 1105 * If B_ERROR or B_NOCACHE is set, pages in the VM object will be 1106 * invalidated. B_ERROR cannot be set for a failed write unless the 1107 * buffer is also B_INVAL because it hits the re-dirtying code above. 1108 * 1109 * Normally we can do this whether a buffer is B_DELWRI or not. If 1110 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1111 * the commit state and we cannot afford to lose the buffer. If the 1112 * buffer has a background write in progress, we need to keep it 1113 * around to prevent it from being reconstituted and starting a second 1114 * background write. 1115 */ 1116 if ((bp->b_flags & B_VMIO) 1117 && !(bp->b_vp->v_tag == VT_NFS && 1118 !vn_isdisk(bp->b_vp, NULL) && 1119 (bp->b_flags & B_DELWRI)) 1120 ) { 1121 /* 1122 * Rundown for VMIO buffers which are not dirty NFS buffers. 1123 */ 1124 int i, j, resid; 1125 vm_page_t m; 1126 off_t foff; 1127 vm_pindex_t poff; 1128 vm_object_t obj; 1129 struct vnode *vp; 1130 1131 vp = bp->b_vp; 1132 1133 /* 1134 * Get the base offset and length of the buffer. Note that 1135 * in the VMIO case if the buffer block size is not 1136 * page-aligned then b_data pointer may not be page-aligned. 1137 * But our b_xio.xio_pages array *IS* page aligned. 1138 * 1139 * block sizes less then DEV_BSIZE (usually 512) are not 1140 * supported due to the page granularity bits (m->valid, 1141 * m->dirty, etc...). 1142 * 1143 * See man buf(9) for more information 1144 */ 1145 1146 resid = bp->b_bufsize; 1147 foff = bp->b_loffset; 1148 1149 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1150 m = bp->b_xio.xio_pages[i]; 1151 vm_page_flag_clear(m, PG_ZERO); 1152 /* 1153 * If we hit a bogus page, fixup *all* of them 1154 * now. Note that we left these pages wired 1155 * when we removed them so they had better exist, 1156 * and they cannot be ripped out from under us so 1157 * no critical section protection is necessary. 1158 */ 1159 if (m == bogus_page) { 1160 VOP_GETVOBJECT(vp, &obj); 1161 poff = OFF_TO_IDX(bp->b_loffset); 1162 1163 for (j = i; j < bp->b_xio.xio_npages; j++) { 1164 vm_page_t mtmp; 1165 1166 mtmp = bp->b_xio.xio_pages[j]; 1167 if (mtmp == bogus_page) { 1168 mtmp = vm_page_lookup(obj, poff + j); 1169 if (!mtmp) { 1170 panic("brelse: page missing"); 1171 } 1172 bp->b_xio.xio_pages[j] = mtmp; 1173 } 1174 } 1175 1176 if ((bp->b_flags & B_INVAL) == 0) { 1177 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 1178 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 1179 } 1180 m = bp->b_xio.xio_pages[i]; 1181 } 1182 1183 /* 1184 * Invalidate the backing store if B_NOCACHE is set 1185 * (e.g. used with vinvalbuf()). If this is NFS 1186 * we impose a requirement that the block size be 1187 * a multiple of PAGE_SIZE and create a temporary 1188 * hack to basically invalidate the whole page. The 1189 * problem is that NFS uses really odd buffer sizes 1190 * especially when tracking piecemeal writes and 1191 * it also vinvalbuf()'s a lot, which would result 1192 * in only partial page validation and invalidation 1193 * here. If the file page is mmap()'d, however, 1194 * all the valid bits get set so after we invalidate 1195 * here we would end up with weird m->valid values 1196 * like 0xfc. nfs_getpages() can't handle this so 1197 * we clear all the valid bits for the NFS case 1198 * instead of just some of them. 1199 * 1200 * The real bug is the VM system having to set m->valid 1201 * to VM_PAGE_BITS_ALL for faulted-in pages, which 1202 * itself is an artifact of the whole 512-byte 1203 * granular mess that exists to support odd block 1204 * sizes and UFS meta-data block sizes (e.g. 6144). 1205 * A complete rewrite is required. 1206 */ 1207 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 1208 int poffset = foff & PAGE_MASK; 1209 int presid; 1210 1211 presid = PAGE_SIZE - poffset; 1212 if (bp->b_vp->v_tag == VT_NFS && 1213 bp->b_vp->v_type == VREG) { 1214 ; /* entire page */ 1215 } else if (presid > resid) { 1216 presid = resid; 1217 } 1218 KASSERT(presid >= 0, ("brelse: extra page")); 1219 vm_page_set_invalid(m, poffset, presid); 1220 } 1221 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1222 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1223 } 1224 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1225 vfs_vmio_release(bp); 1226 } else if (bp->b_flags & B_VMIO) { 1227 /* 1228 * Rundown for VMIO buffers which are dirty NFS buffers. Such 1229 * buffers contain tracking ranges for NFS and cannot normally 1230 * be released. Due to the dirty check above this series of 1231 * conditionals, B_RELBUF probably will never be set in this 1232 * codepath. 1233 */ 1234 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1235 vfs_vmio_release(bp); 1236 } else { 1237 /* 1238 * Rundown for non-VMIO buffers. 1239 */ 1240 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1241 #if 0 1242 if (bp->b_vp) 1243 printf("brelse bp %p %08x/%08lx: Warning, caught and fixed brelvp bug\n", bp, saved_flags, bp->b_flags); 1244 #endif 1245 if (bp->b_bufsize) 1246 allocbuf(bp, 0); 1247 if (bp->b_vp) 1248 brelvp(bp); 1249 } 1250 } 1251 1252 if (bp->b_qindex != BQUEUE_NONE) 1253 panic("brelse: free buffer onto another queue???"); 1254 if (BUF_REFCNTNB(bp) > 1) { 1255 /* Temporary panic to verify exclusive locking */ 1256 /* This panic goes away when we allow shared refs */ 1257 panic("brelse: multiple refs"); 1258 /* do not release to free list */ 1259 BUF_UNLOCK(bp); 1260 crit_exit(); 1261 return; 1262 } 1263 1264 /* 1265 * Figure out the correct queue to place the cleaned up buffer on. 1266 * Buffers placed in the EMPTY or EMPTYKVA had better already be 1267 * disassociated from their vnode. 1268 */ 1269 1270 if (bp->b_bufsize == 0) { 1271 /* 1272 * Buffers with no memory. Due to conditionals near the top 1273 * of brelse() such buffers should probably already be 1274 * marked B_INVAL and disassociated from their vnode. 1275 */ 1276 bp->b_flags |= B_INVAL; 1277 bp->b_xflags &= ~BX_BKGRDWRITE; 1278 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08lx vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp)); 1279 if (bp->b_xflags & BX_BKGRDINPROG) 1280 panic("losing buffer 1"); 1281 if (bp->b_kvasize) { 1282 bp->b_qindex = BQUEUE_EMPTYKVA; 1283 } else { 1284 bp->b_qindex = BQUEUE_EMPTY; 1285 } 1286 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1287 LIST_REMOVE(bp, b_hash); 1288 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1289 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) { 1290 /* 1291 * Buffers with junk contents. Again these buffers had better 1292 * already be disassociated from their vnode. 1293 */ 1294 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08lx vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp)); 1295 bp->b_flags |= B_INVAL; 1296 bp->b_xflags &= ~BX_BKGRDWRITE; 1297 if (bp->b_xflags & BX_BKGRDINPROG) 1298 panic("losing buffer 2"); 1299 bp->b_qindex = BQUEUE_CLEAN; 1300 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1301 LIST_REMOVE(bp, b_hash); 1302 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1303 } else if (bp->b_flags & B_LOCKED) { 1304 /* 1305 * Buffers that are locked. 1306 */ 1307 bp->b_qindex = BQUEUE_LOCKED; 1308 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist); 1309 } else { 1310 /* 1311 * Remaining buffers. These buffers are still associated with 1312 * their vnode. 1313 */ 1314 switch(bp->b_flags & (B_DELWRI|B_AGE)) { 1315 case B_DELWRI | B_AGE: 1316 bp->b_qindex = BQUEUE_DIRTY; 1317 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_DIRTY], bp, b_freelist); 1318 break; 1319 case B_DELWRI: 1320 bp->b_qindex = BQUEUE_DIRTY; 1321 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist); 1322 break; 1323 case B_AGE: 1324 bp->b_qindex = BQUEUE_CLEAN; 1325 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1326 break; 1327 default: 1328 bp->b_qindex = BQUEUE_CLEAN; 1329 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1330 break; 1331 } 1332 } 1333 1334 /* 1335 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1336 * on the correct queue. 1337 */ 1338 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) 1339 bundirty(bp); 1340 1341 /* 1342 * Fixup numfreebuffers count. The bp is on an appropriate queue 1343 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1344 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1345 * if B_INVAL is set ). 1346 */ 1347 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1348 bufcountwakeup(); 1349 1350 /* 1351 * Something we can maybe free or reuse 1352 */ 1353 if (bp->b_bufsize || bp->b_kvasize) 1354 bufspacewakeup(); 1355 1356 /* unlock */ 1357 BUF_UNLOCK(bp); 1358 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | 1359 B_DIRECT | B_NOWDRAIN); 1360 crit_exit(); 1361 } 1362 1363 /* 1364 * bqrelse: 1365 * 1366 * Release a buffer back to the appropriate queue but do not try to free 1367 * it. The buffer is expected to be used again soon. 1368 * 1369 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1370 * biodone() to requeue an async I/O on completion. It is also used when 1371 * known good buffers need to be requeued but we think we may need the data 1372 * again soon. 1373 * 1374 * XXX we should be able to leave the B_RELBUF hint set on completion. 1375 */ 1376 void 1377 bqrelse(struct buf * bp) 1378 { 1379 crit_enter(); 1380 1381 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1382 1383 if (bp->b_qindex != BQUEUE_NONE) 1384 panic("bqrelse: free buffer onto another queue???"); 1385 if (BUF_REFCNTNB(bp) > 1) { 1386 /* do not release to free list */ 1387 panic("bqrelse: multiple refs"); 1388 BUF_UNLOCK(bp); 1389 crit_exit(); 1390 return; 1391 } 1392 if (bp->b_flags & B_LOCKED) { 1393 bp->b_flags &= ~B_ERROR; 1394 bp->b_qindex = BQUEUE_LOCKED; 1395 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist); 1396 /* buffers with stale but valid contents */ 1397 } else if (bp->b_flags & B_DELWRI) { 1398 bp->b_qindex = BQUEUE_DIRTY; 1399 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist); 1400 } else if (vm_page_count_severe()) { 1401 /* 1402 * We are too low on memory, we have to try to free the 1403 * buffer (most importantly: the wired pages making up its 1404 * backing store) *now*. 1405 */ 1406 crit_exit(); 1407 brelse(bp); 1408 return; 1409 } else { 1410 bp->b_qindex = BQUEUE_CLEAN; 1411 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist); 1412 } 1413 1414 if ((bp->b_flags & B_LOCKED) == 0 && 1415 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1416 bufcountwakeup(); 1417 } 1418 1419 /* 1420 * Something we can maybe free or reuse. 1421 */ 1422 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1423 bufspacewakeup(); 1424 1425 /* 1426 * Final cleanup and unlock. Clear bits that are only used while a 1427 * buffer is actively locked. 1428 */ 1429 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1430 BUF_UNLOCK(bp); 1431 crit_exit(); 1432 } 1433 1434 /* 1435 * vfs_vmio_release: 1436 * 1437 * Return backing pages held by the buffer 'bp' back to the VM system 1438 * if possible. The pages are freed if they are no longer valid or 1439 * attempt to free if it was used for direct I/O otherwise they are 1440 * sent to the page cache. 1441 * 1442 * Pages that were marked busy are left alone and skipped. 1443 * 1444 * The KVA mapping (b_data) for the underlying pages is removed by 1445 * this function. 1446 */ 1447 static void 1448 vfs_vmio_release(struct buf *bp) 1449 { 1450 int i; 1451 vm_page_t m; 1452 1453 crit_enter(); 1454 for (i = 0; i < bp->b_xio.xio_npages; i++) { 1455 m = bp->b_xio.xio_pages[i]; 1456 bp->b_xio.xio_pages[i] = NULL; 1457 /* 1458 * In order to keep page LRU ordering consistent, put 1459 * everything on the inactive queue. 1460 */ 1461 vm_page_unwire(m, 0); 1462 /* 1463 * We don't mess with busy pages, it is 1464 * the responsibility of the process that 1465 * busied the pages to deal with them. 1466 */ 1467 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1468 continue; 1469 1470 if (m->wire_count == 0) { 1471 vm_page_flag_clear(m, PG_ZERO); 1472 /* 1473 * Might as well free the page if we can and it has 1474 * no valid data. We also free the page if the 1475 * buffer was used for direct I/O. 1476 */ 1477 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1478 m->hold_count == 0) { 1479 vm_page_busy(m); 1480 vm_page_protect(m, VM_PROT_NONE); 1481 vm_page_free(m); 1482 } else if (bp->b_flags & B_DIRECT) { 1483 vm_page_try_to_free(m); 1484 } else if (vm_page_count_severe()) { 1485 vm_page_try_to_cache(m); 1486 } 1487 } 1488 } 1489 crit_exit(); 1490 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages); 1491 if (bp->b_bufsize) { 1492 bufspacewakeup(); 1493 bp->b_bufsize = 0; 1494 } 1495 bp->b_xio.xio_npages = 0; 1496 bp->b_flags &= ~B_VMIO; 1497 if (bp->b_vp) 1498 brelvp(bp); 1499 } 1500 1501 /* 1502 * gbincore: 1503 * 1504 * Check to see if a block is currently memory resident. 1505 */ 1506 struct buf * 1507 gbincore(struct vnode * vp, daddr_t blkno) 1508 { 1509 struct buf *bp; 1510 struct bufhashhdr *bh; 1511 1512 bh = bufhash(vp, blkno); 1513 LIST_FOREACH(bp, bh, b_hash) { 1514 if (bp->b_vp == vp && bp->b_lblkno == blkno) 1515 break; 1516 } 1517 return (bp); 1518 } 1519 1520 /* 1521 * vfs_bio_awrite: 1522 * 1523 * Implement clustered async writes for clearing out B_DELWRI buffers. 1524 * This is much better then the old way of writing only one buffer at 1525 * a time. Note that we may not be presented with the buffers in the 1526 * correct order, so we search for the cluster in both directions. 1527 */ 1528 int 1529 vfs_bio_awrite(struct buf * bp) 1530 { 1531 int i; 1532 int j; 1533 daddr_t lblkno = bp->b_lblkno; 1534 struct vnode *vp = bp->b_vp; 1535 int ncl; 1536 struct buf *bpa; 1537 int nwritten; 1538 int size; 1539 int maxcl; 1540 1541 crit_enter(); 1542 /* 1543 * right now we support clustered writing only to regular files. If 1544 * we find a clusterable block we could be in the middle of a cluster 1545 * rather then at the beginning. 1546 * 1547 * NOTE: b_bio1 contains the logical loffset/lblkno and is aliased 1548 * to b_lblkno and b_loffset. b_bio2 contains the translated block 1549 * number. 1550 */ 1551 if ((vp->v_type == VREG) && 1552 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1553 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1554 1555 size = vp->v_mount->mnt_stat.f_iosize; 1556 maxcl = MAXPHYS / size; 1557 1558 for (i = 1; i < maxcl; i++) { 1559 if ((bpa = gbincore(vp, lblkno + i)) && 1560 BUF_REFCNT(bpa) == 0 && 1561 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1562 (B_DELWRI | B_CLUSTEROK)) && 1563 (bpa->b_bufsize == size)) { 1564 if ((bpa->b_bio2.bio_blkno == (daddr_t)-1) || 1565 (bpa->b_bio2.bio_blkno != 1566 bp->b_bio2.bio_blkno + ((i * size) >> DEV_BSHIFT))) 1567 break; 1568 } else { 1569 break; 1570 } 1571 } 1572 for (j = 1; i + j <= maxcl && j <= lblkno; j++) { 1573 if ((bpa = gbincore(vp, lblkno - j)) && 1574 BUF_REFCNT(bpa) == 0 && 1575 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1576 (B_DELWRI | B_CLUSTEROK)) && 1577 (bpa->b_bufsize == size)) { 1578 if ((bpa->b_bio2.bio_blkno == (daddr_t)-1) || 1579 (bpa->b_bio2.bio_blkno != 1580 bp->b_bio2.bio_blkno - ((j * size) >> DEV_BSHIFT))) 1581 break; 1582 } else { 1583 break; 1584 } 1585 } 1586 --j; 1587 ncl = i + j; 1588 /* 1589 * this is a possible cluster write 1590 */ 1591 if (ncl != 1) { 1592 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1593 crit_exit(); 1594 return nwritten; 1595 } 1596 } 1597 1598 BUF_LOCK(bp, LK_EXCLUSIVE); 1599 bremfree(bp); 1600 bp->b_flags |= B_ASYNC; 1601 1602 crit_exit(); 1603 /* 1604 * default (old) behavior, writing out only one block 1605 * 1606 * XXX returns b_bufsize instead of b_bcount for nwritten? 1607 */ 1608 nwritten = bp->b_bufsize; 1609 (void) VOP_BWRITE(bp->b_vp, bp); 1610 1611 return nwritten; 1612 } 1613 1614 /* 1615 * getnewbuf: 1616 * 1617 * Find and initialize a new buffer header, freeing up existing buffers 1618 * in the bufqueues as necessary. The new buffer is returned locked. 1619 * 1620 * Important: B_INVAL is not set. If the caller wishes to throw the 1621 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1622 * 1623 * We block if: 1624 * We have insufficient buffer headers 1625 * We have insufficient buffer space 1626 * buffer_map is too fragmented ( space reservation fails ) 1627 * If we have to flush dirty buffers ( but we try to avoid this ) 1628 * 1629 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1630 * Instead we ask the buf daemon to do it for us. We attempt to 1631 * avoid piecemeal wakeups of the pageout daemon. 1632 */ 1633 1634 static struct buf * 1635 getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1636 { 1637 struct buf *bp; 1638 struct buf *nbp; 1639 int defrag = 0; 1640 int nqindex; 1641 static int flushingbufs; 1642 1643 /* 1644 * We can't afford to block since we might be holding a vnode lock, 1645 * which may prevent system daemons from running. We deal with 1646 * low-memory situations by proactively returning memory and running 1647 * async I/O rather then sync I/O. 1648 */ 1649 1650 ++getnewbufcalls; 1651 --getnewbufrestarts; 1652 restart: 1653 ++getnewbufrestarts; 1654 1655 /* 1656 * Setup for scan. If we do not have enough free buffers, 1657 * we setup a degenerate case that immediately fails. Note 1658 * that if we are specially marked process, we are allowed to 1659 * dip into our reserves. 1660 * 1661 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1662 * 1663 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1664 * However, there are a number of cases (defragging, reusing, ...) 1665 * where we cannot backup. 1666 */ 1667 nqindex = BQUEUE_EMPTYKVA; 1668 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]); 1669 1670 if (nbp == NULL) { 1671 /* 1672 * If no EMPTYKVA buffers and we are either 1673 * defragging or reusing, locate a CLEAN buffer 1674 * to free or reuse. If bufspace useage is low 1675 * skip this step so we can allocate a new buffer. 1676 */ 1677 if (defrag || bufspace >= lobufspace) { 1678 nqindex = BQUEUE_CLEAN; 1679 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]); 1680 } 1681 1682 /* 1683 * If we could not find or were not allowed to reuse a 1684 * CLEAN buffer, check to see if it is ok to use an EMPTY 1685 * buffer. We can only use an EMPTY buffer if allocating 1686 * its KVA would not otherwise run us out of buffer space. 1687 */ 1688 if (nbp == NULL && defrag == 0 && 1689 bufspace + maxsize < hibufspace) { 1690 nqindex = BQUEUE_EMPTY; 1691 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]); 1692 } 1693 } 1694 1695 /* 1696 * Run scan, possibly freeing data and/or kva mappings on the fly 1697 * depending. 1698 */ 1699 1700 while ((bp = nbp) != NULL) { 1701 int qindex = nqindex; 1702 1703 /* 1704 * Calculate next bp ( we can only use it if we do not block 1705 * or do other fancy things ). 1706 */ 1707 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1708 switch(qindex) { 1709 case BQUEUE_EMPTY: 1710 nqindex = BQUEUE_EMPTYKVA; 1711 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]))) 1712 break; 1713 /* fall through */ 1714 case BQUEUE_EMPTYKVA: 1715 nqindex = BQUEUE_CLEAN; 1716 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]))) 1717 break; 1718 /* fall through */ 1719 case BQUEUE_CLEAN: 1720 /* 1721 * nbp is NULL. 1722 */ 1723 break; 1724 } 1725 } 1726 1727 /* 1728 * Sanity Checks 1729 */ 1730 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1731 1732 /* 1733 * Note: we no longer distinguish between VMIO and non-VMIO 1734 * buffers. 1735 */ 1736 1737 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1738 1739 /* 1740 * If we are defragging then we need a buffer with 1741 * b_kvasize != 0. XXX this situation should no longer 1742 * occur, if defrag is non-zero the buffer's b_kvasize 1743 * should also be non-zero at this point. XXX 1744 */ 1745 if (defrag && bp->b_kvasize == 0) { 1746 printf("Warning: defrag empty buffer %p\n", bp); 1747 continue; 1748 } 1749 1750 /* 1751 * Start freeing the bp. This is somewhat involved. nbp 1752 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers 1753 * on the clean list must be disassociated from their 1754 * current vnode. Buffers on the empty[kva] lists have 1755 * already been disassociated. 1756 */ 1757 1758 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { 1759 printf("getnewbuf: warning, locked buf %p, race corrected\n", bp); 1760 tsleep(&bd_request, 0, "gnbxxx", hz / 100); 1761 goto restart; 1762 } 1763 if (bp->b_qindex != qindex) { 1764 printf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex); 1765 BUF_UNLOCK(bp); 1766 goto restart; 1767 } 1768 bremfree(bp); 1769 1770 if (qindex == BQUEUE_CLEAN) { 1771 if (bp->b_flags & B_VMIO) { 1772 bp->b_flags &= ~B_ASYNC; 1773 vfs_vmio_release(bp); 1774 } 1775 if (bp->b_vp) 1776 brelvp(bp); 1777 } 1778 1779 /* 1780 * NOTE: nbp is now entirely invalid. We can only restart 1781 * the scan from this point on. 1782 * 1783 * Get the rest of the buffer freed up. b_kva* is still 1784 * valid after this operation. 1785 */ 1786 1787 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08lx vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex)); 1788 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 1789 (*bioops.io_deallocate)(bp); 1790 if (bp->b_xflags & BX_BKGRDINPROG) 1791 panic("losing buffer 3"); 1792 LIST_REMOVE(bp, b_hash); 1793 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1794 1795 /* 1796 * critical section protection is not required when 1797 * scrapping a buffer's contents because it is already 1798 * wired. 1799 */ 1800 if (bp->b_bufsize) 1801 allocbuf(bp, 0); 1802 1803 bp->b_flags = 0; 1804 bp->b_xflags = 0; 1805 bp->b_vp = NULL; 1806 bp->b_error = 0; 1807 bp->b_resid = 0; 1808 bp->b_bcount = 0; 1809 bp->b_xio.xio_npages = 0; 1810 bp->b_dirtyoff = bp->b_dirtyend = 0; 1811 reinitbufbio(bp); 1812 1813 LIST_INIT(&bp->b_dep); 1814 1815 /* 1816 * If we are defragging then free the buffer. 1817 */ 1818 if (defrag) { 1819 bp->b_flags |= B_INVAL; 1820 bfreekva(bp); 1821 brelse(bp); 1822 defrag = 0; 1823 goto restart; 1824 } 1825 1826 /* 1827 * If we are overcomitted then recover the buffer and its 1828 * KVM space. This occurs in rare situations when multiple 1829 * processes are blocked in getnewbuf() or allocbuf(). 1830 */ 1831 if (bufspace >= hibufspace) 1832 flushingbufs = 1; 1833 if (flushingbufs && bp->b_kvasize != 0) { 1834 bp->b_flags |= B_INVAL; 1835 bfreekva(bp); 1836 brelse(bp); 1837 goto restart; 1838 } 1839 if (bufspace < lobufspace) 1840 flushingbufs = 0; 1841 break; 1842 } 1843 1844 /* 1845 * If we exhausted our list, sleep as appropriate. We may have to 1846 * wakeup various daemons and write out some dirty buffers. 1847 * 1848 * Generally we are sleeping due to insufficient buffer space. 1849 */ 1850 1851 if (bp == NULL) { 1852 int flags; 1853 char *waitmsg; 1854 1855 if (defrag) { 1856 flags = VFS_BIO_NEED_BUFSPACE; 1857 waitmsg = "nbufkv"; 1858 } else if (bufspace >= hibufspace) { 1859 waitmsg = "nbufbs"; 1860 flags = VFS_BIO_NEED_BUFSPACE; 1861 } else { 1862 waitmsg = "newbuf"; 1863 flags = VFS_BIO_NEED_ANY; 1864 } 1865 1866 bd_speedup(); /* heeeelp */ 1867 1868 needsbuffer |= flags; 1869 while (needsbuffer & flags) { 1870 if (tsleep(&needsbuffer, slpflag, waitmsg, slptimeo)) 1871 return (NULL); 1872 } 1873 } else { 1874 /* 1875 * We finally have a valid bp. We aren't quite out of the 1876 * woods, we still have to reserve kva space. In order 1877 * to keep fragmentation sane we only allocate kva in 1878 * BKVASIZE chunks. 1879 */ 1880 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1881 1882 if (maxsize != bp->b_kvasize) { 1883 vm_offset_t addr = 0; 1884 int count; 1885 1886 bfreekva(bp); 1887 1888 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1889 vm_map_lock(buffer_map); 1890 1891 if (vm_map_findspace(buffer_map, 1892 vm_map_min(buffer_map), maxsize, 1893 maxsize, &addr)) { 1894 /* 1895 * Uh oh. Buffer map is too fragmented. We 1896 * must defragment the map. 1897 */ 1898 vm_map_unlock(buffer_map); 1899 vm_map_entry_release(count); 1900 ++bufdefragcnt; 1901 defrag = 1; 1902 bp->b_flags |= B_INVAL; 1903 brelse(bp); 1904 goto restart; 1905 } 1906 if (addr) { 1907 vm_map_insert(buffer_map, &count, 1908 NULL, 0, 1909 addr, addr + maxsize, 1910 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1911 1912 bp->b_kvabase = (caddr_t) addr; 1913 bp->b_kvasize = maxsize; 1914 bufspace += bp->b_kvasize; 1915 ++bufreusecnt; 1916 } 1917 vm_map_unlock(buffer_map); 1918 vm_map_entry_release(count); 1919 } 1920 bp->b_data = bp->b_kvabase; 1921 } 1922 return(bp); 1923 } 1924 1925 /* 1926 * buf_daemon: 1927 * 1928 * Buffer flushing daemon. Buffers are normally flushed by the 1929 * update daemon but if it cannot keep up this process starts to 1930 * take the load in an attempt to prevent getnewbuf() from blocking. 1931 */ 1932 1933 static struct thread *bufdaemonthread; 1934 1935 static struct kproc_desc buf_kp = { 1936 "bufdaemon", 1937 buf_daemon, 1938 &bufdaemonthread 1939 }; 1940 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1941 1942 static void 1943 buf_daemon() 1944 { 1945 /* 1946 * This process needs to be suspended prior to shutdown sync. 1947 */ 1948 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 1949 bufdaemonthread, SHUTDOWN_PRI_LAST); 1950 1951 /* 1952 * This process is allowed to take the buffer cache to the limit 1953 */ 1954 crit_enter(); 1955 1956 for (;;) { 1957 kproc_suspend_loop(); 1958 1959 /* 1960 * Do the flush. Limit the amount of in-transit I/O we 1961 * allow to build up, otherwise we would completely saturate 1962 * the I/O system. Wakeup any waiting processes before we 1963 * normally would so they can run in parallel with our drain. 1964 */ 1965 while (numdirtybuffers > lodirtybuffers) { 1966 if (flushbufqueues() == 0) 1967 break; 1968 waitrunningbufspace(); 1969 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 1970 } 1971 1972 /* 1973 * Only clear bd_request if we have reached our low water 1974 * mark. The buf_daemon normally waits 5 seconds and 1975 * then incrementally flushes any dirty buffers that have 1976 * built up, within reason. 1977 * 1978 * If we were unable to hit our low water mark and couldn't 1979 * find any flushable buffers, we sleep half a second. 1980 * Otherwise we loop immediately. 1981 */ 1982 if (numdirtybuffers <= lodirtybuffers) { 1983 /* 1984 * We reached our low water mark, reset the 1985 * request and sleep until we are needed again. 1986 * The sleep is just so the suspend code works. 1987 */ 1988 bd_request = 0; 1989 tsleep(&bd_request, 0, "psleep", hz); 1990 } else { 1991 /* 1992 * We couldn't find any flushable dirty buffers but 1993 * still have too many dirty buffers, we 1994 * have to sleep and try again. (rare) 1995 */ 1996 tsleep(&bd_request, 0, "qsleep", hz / 2); 1997 } 1998 } 1999 } 2000 2001 /* 2002 * flushbufqueues: 2003 * 2004 * Try to flush a buffer in the dirty queue. We must be careful to 2005 * free up B_INVAL buffers instead of write them, which NFS is 2006 * particularly sensitive to. 2007 */ 2008 2009 static int 2010 flushbufqueues(void) 2011 { 2012 struct buf *bp; 2013 int r = 0; 2014 2015 bp = TAILQ_FIRST(&bufqueues[BQUEUE_DIRTY]); 2016 2017 while (bp) { 2018 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); 2019 if ((bp->b_flags & B_DELWRI) != 0 && 2020 (bp->b_xflags & BX_BKGRDINPROG) == 0) { 2021 if (bp->b_flags & B_INVAL) { 2022 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 2023 panic("flushbufqueues: locked buf"); 2024 bremfree(bp); 2025 brelse(bp); 2026 ++r; 2027 break; 2028 } 2029 if (LIST_FIRST(&bp->b_dep) != NULL && 2030 bioops.io_countdeps && 2031 (bp->b_flags & B_DEFERRED) == 0 && 2032 (*bioops.io_countdeps)(bp, 0)) { 2033 TAILQ_REMOVE(&bufqueues[BQUEUE_DIRTY], 2034 bp, b_freelist); 2035 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], 2036 bp, b_freelist); 2037 bp->b_flags |= B_DEFERRED; 2038 bp = TAILQ_FIRST(&bufqueues[BQUEUE_DIRTY]); 2039 continue; 2040 } 2041 vfs_bio_awrite(bp); 2042 ++r; 2043 break; 2044 } 2045 bp = TAILQ_NEXT(bp, b_freelist); 2046 } 2047 return (r); 2048 } 2049 2050 /* 2051 * incore: 2052 * 2053 * Check to see if a block is currently resident in memory. 2054 */ 2055 struct buf * 2056 incore(struct vnode * vp, daddr_t blkno) 2057 { 2058 struct buf *bp; 2059 2060 crit_enter(); 2061 bp = gbincore(vp, blkno); 2062 crit_exit(); 2063 return (bp); 2064 } 2065 2066 /* 2067 * inmem: 2068 * 2069 * Returns true if no I/O is needed to access the associated VM object. 2070 * This is like incore except it also hunts around in the VM system for 2071 * the data. 2072 * 2073 * Note that we ignore vm_page_free() races from interrupts against our 2074 * lookup, since if the caller is not protected our return value will not 2075 * be any more valid then otherwise once we exit the critical section. 2076 */ 2077 int 2078 inmem(struct vnode * vp, daddr_t blkno) 2079 { 2080 vm_object_t obj; 2081 vm_offset_t toff, tinc, size; 2082 vm_page_t m; 2083 vm_ooffset_t off; 2084 2085 if (incore(vp, blkno)) 2086 return 1; 2087 if (vp->v_mount == NULL) 2088 return 0; 2089 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_flag & VOBJBUF) == 0) 2090 return 0; 2091 2092 size = PAGE_SIZE; 2093 if (size > vp->v_mount->mnt_stat.f_iosize) 2094 size = vp->v_mount->mnt_stat.f_iosize; 2095 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2096 2097 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2098 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2099 if (!m) 2100 return 0; 2101 tinc = size; 2102 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2103 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2104 if (vm_page_is_valid(m, 2105 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2106 return 0; 2107 } 2108 return 1; 2109 } 2110 2111 /* 2112 * vfs_setdirty: 2113 * 2114 * Sets the dirty range for a buffer based on the status of the dirty 2115 * bits in the pages comprising the buffer. 2116 * 2117 * The range is limited to the size of the buffer. 2118 * 2119 * This routine is primarily used by NFS, but is generalized for the 2120 * B_VMIO case. 2121 */ 2122 static void 2123 vfs_setdirty(struct buf *bp) 2124 { 2125 int i; 2126 vm_object_t object; 2127 2128 /* 2129 * Degenerate case - empty buffer 2130 */ 2131 2132 if (bp->b_bufsize == 0) 2133 return; 2134 2135 /* 2136 * We qualify the scan for modified pages on whether the 2137 * object has been flushed yet. The OBJ_WRITEABLE flag 2138 * is not cleared simply by protecting pages off. 2139 */ 2140 2141 if ((bp->b_flags & B_VMIO) == 0) 2142 return; 2143 2144 object = bp->b_xio.xio_pages[0]->object; 2145 2146 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2147 printf("Warning: object %p writeable but not mightbedirty\n", object); 2148 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2149 printf("Warning: object %p mightbedirty but not writeable\n", object); 2150 2151 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2152 vm_offset_t boffset; 2153 vm_offset_t eoffset; 2154 2155 /* 2156 * test the pages to see if they have been modified directly 2157 * by users through the VM system. 2158 */ 2159 for (i = 0; i < bp->b_xio.xio_npages; i++) { 2160 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO); 2161 vm_page_test_dirty(bp->b_xio.xio_pages[i]); 2162 } 2163 2164 /* 2165 * Calculate the encompassing dirty range, boffset and eoffset, 2166 * (eoffset - boffset) bytes. 2167 */ 2168 2169 for (i = 0; i < bp->b_xio.xio_npages; i++) { 2170 if (bp->b_xio.xio_pages[i]->dirty) 2171 break; 2172 } 2173 boffset = (i << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK); 2174 2175 for (i = bp->b_xio.xio_npages - 1; i >= 0; --i) { 2176 if (bp->b_xio.xio_pages[i]->dirty) { 2177 break; 2178 } 2179 } 2180 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK); 2181 2182 /* 2183 * Fit it to the buffer. 2184 */ 2185 2186 if (eoffset > bp->b_bcount) 2187 eoffset = bp->b_bcount; 2188 2189 /* 2190 * If we have a good dirty range, merge with the existing 2191 * dirty range. 2192 */ 2193 2194 if (boffset < eoffset) { 2195 if (bp->b_dirtyoff > boffset) 2196 bp->b_dirtyoff = boffset; 2197 if (bp->b_dirtyend < eoffset) 2198 bp->b_dirtyend = eoffset; 2199 } 2200 } 2201 } 2202 2203 /* 2204 * getblk: 2205 * 2206 * Get a block given a specified block and offset into a file/device. 2207 * The buffers B_DONE bit will be cleared on return, making it almost 2208 * ready for an I/O initiation. B_INVAL may or may not be set on 2209 * return. The caller should clear B_INVAL prior to initiating a 2210 * READ. 2211 * 2212 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE 2213 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ, 2214 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer 2215 * without doing any of those things the system will likely believe 2216 * the buffer to be valid (especially if it is not B_VMIO), and the 2217 * next getblk() will return the buffer with B_CACHE set. 2218 * 2219 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2220 * an existing buffer. 2221 * 2222 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2223 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2224 * and then cleared based on the backing VM. If the previous buffer is 2225 * non-0-sized but invalid, B_CACHE will be cleared. 2226 * 2227 * If getblk() must create a new buffer, the new buffer is returned with 2228 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2229 * case it is returned with B_INVAL clear and B_CACHE set based on the 2230 * backing VM. 2231 * 2232 * getblk() also forces a VOP_BWRITE() for any B_DELWRI buffer whos 2233 * B_CACHE bit is clear. 2234 * 2235 * What this means, basically, is that the caller should use B_CACHE to 2236 * determine whether the buffer is fully valid or not and should clear 2237 * B_INVAL prior to issuing a read. If the caller intends to validate 2238 * the buffer by loading its data area with something, the caller needs 2239 * to clear B_INVAL. If the caller does this without issuing an I/O, 2240 * the caller should set B_CACHE ( as an optimization ), else the caller 2241 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2242 * a write attempt or if it was a successfull read. If the caller 2243 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR 2244 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2245 */ 2246 struct buf * 2247 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2248 { 2249 struct buf *bp; 2250 struct bufhashhdr *bh; 2251 2252 if (size > MAXBSIZE) 2253 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE); 2254 2255 crit_enter(); 2256 loop: 2257 /* 2258 * Block if we are low on buffers. Certain processes are allowed 2259 * to completely exhaust the buffer cache. 2260 * 2261 * If this check ever becomes a bottleneck it may be better to 2262 * move it into the else, when gbincore() fails. At the moment 2263 * it isn't a problem. 2264 * 2265 * XXX remove, we cannot afford to block anywhere if holding a vnode 2266 * lock in low-memory situation, so take it to the max. 2267 */ 2268 if (numfreebuffers == 0) { 2269 if (!curproc) 2270 return NULL; 2271 needsbuffer |= VFS_BIO_NEED_ANY; 2272 tsleep(&needsbuffer, slpflag, "newbuf", slptimeo); 2273 } 2274 2275 if ((bp = gbincore(vp, blkno))) { 2276 /* 2277 * The buffer was found in the cache, but we need to lock it. 2278 * Even with LK_NOWAIT the lockmgr may break our critical 2279 * section, so double-check the validity of the buffer 2280 * once the lock has been obtained. 2281 */ 2282 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2283 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, 2284 "getblk", slpflag, slptimeo) == ENOLCK) 2285 goto loop; 2286 crit_exit(); 2287 return (struct buf *) NULL; 2288 } 2289 2290 /* 2291 * Once the buffer has been locked, make sure we didn't race 2292 * a buffer recyclement. Buffers that are no longer hashed 2293 * will have b_vp == NULL, so this takes care of that check 2294 * as well. 2295 */ 2296 if (bp->b_vp != vp || bp->b_lblkno != blkno) { 2297 printf("Warning buffer %p (vp %p lblkno %d) was recycled\n", bp, vp, (int)blkno); 2298 BUF_UNLOCK(bp); 2299 goto loop; 2300 } 2301 2302 /* 2303 * Make sure that B_INVAL buffers do not have a cached 2304 * block number translation. 2305 */ 2306 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_blkno != (daddr_t)-1)) { 2307 printf("Warning invalid buffer %p (vp %p lblkno %d) did not have cleared bio_blkno cache\n", bp, vp, (int)blkno); 2308 clearbiocache(&bp->b_bio2); 2309 } 2310 2311 /* 2312 * The buffer is locked. B_CACHE is cleared if the buffer is 2313 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2314 * and for a VMIO buffer B_CACHE is adjusted according to the 2315 * backing VM cache. 2316 */ 2317 if (bp->b_flags & B_INVAL) 2318 bp->b_flags &= ~B_CACHE; 2319 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2320 bp->b_flags |= B_CACHE; 2321 bremfree(bp); 2322 2323 /* 2324 * check for size inconsistancies for non-VMIO case. 2325 */ 2326 2327 if (bp->b_bcount != size) { 2328 if ((bp->b_flags & B_VMIO) == 0 || 2329 (size > bp->b_kvasize)) { 2330 if (bp->b_flags & B_DELWRI) { 2331 bp->b_flags |= B_NOCACHE; 2332 VOP_BWRITE(bp->b_vp, bp); 2333 } else { 2334 if ((bp->b_flags & B_VMIO) && 2335 (LIST_FIRST(&bp->b_dep) == NULL)) { 2336 bp->b_flags |= B_RELBUF; 2337 brelse(bp); 2338 } else { 2339 bp->b_flags |= B_NOCACHE; 2340 VOP_BWRITE(bp->b_vp, bp); 2341 } 2342 } 2343 goto loop; 2344 } 2345 } 2346 2347 /* 2348 * If the size is inconsistant in the VMIO case, we can resize 2349 * the buffer. This might lead to B_CACHE getting set or 2350 * cleared. If the size has not changed, B_CACHE remains 2351 * unchanged from its previous state. 2352 */ 2353 2354 if (bp->b_bcount != size) 2355 allocbuf(bp, size); 2356 2357 KASSERT(bp->b_loffset != NOOFFSET, 2358 ("getblk: no buffer offset")); 2359 2360 /* 2361 * A buffer with B_DELWRI set and B_CACHE clear must 2362 * be committed before we can return the buffer in 2363 * order to prevent the caller from issuing a read 2364 * ( due to B_CACHE not being set ) and overwriting 2365 * it. 2366 * 2367 * Most callers, including NFS and FFS, need this to 2368 * operate properly either because they assume they 2369 * can issue a read if B_CACHE is not set, or because 2370 * ( for example ) an uncached B_DELWRI might loop due 2371 * to softupdates re-dirtying the buffer. In the latter 2372 * case, B_CACHE is set after the first write completes, 2373 * preventing further loops. 2374 * 2375 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2376 * above while extending the buffer, we cannot allow the 2377 * buffer to remain with B_CACHE set after the write 2378 * completes or it will represent a corrupt state. To 2379 * deal with this we set B_NOCACHE to scrap the buffer 2380 * after the write. 2381 * 2382 * We might be able to do something fancy, like setting 2383 * B_CACHE in bwrite() except if B_DELWRI is already set, 2384 * so the below call doesn't set B_CACHE, but that gets real 2385 * confusing. This is much easier. 2386 */ 2387 2388 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2389 bp->b_flags |= B_NOCACHE; 2390 VOP_BWRITE(bp->b_vp, bp); 2391 goto loop; 2392 } 2393 2394 crit_exit(); 2395 bp->b_flags &= ~B_DONE; 2396 } else { 2397 /* 2398 * Buffer is not in-core, create new buffer. The buffer 2399 * returned by getnewbuf() is locked. Note that the returned 2400 * buffer is also considered valid (not marked B_INVAL). 2401 * 2402 * Calculating the offset for the I/O requires figuring out 2403 * the block size. We use DEV_BSIZE for VBLK or VCHR and 2404 * the mount's f_iosize otherwise. If the vnode does not 2405 * have an associated mount we assume that the passed size is 2406 * the block size. 2407 * 2408 * Note that vn_isdisk() cannot be used here since it may 2409 * return a failure for numerous reasons. Note that the 2410 * buffer size may be larger then the block size (the caller 2411 * will use block numbers with the proper multiple). Beware 2412 * of using any v_* fields which are part of unions. In 2413 * particular, in DragonFly the mount point overloading 2414 * mechanism is such that the underlying directory (with a 2415 * non-NULL v_mountedhere) is not a special case. 2416 */ 2417 int bsize, maxsize, vmio; 2418 off_t offset; 2419 2420 if (vp->v_type == VBLK || vp->v_type == VCHR) 2421 bsize = DEV_BSIZE; 2422 else if (vp->v_mount) 2423 bsize = vp->v_mount->mnt_stat.f_iosize; 2424 else 2425 bsize = size; 2426 2427 offset = (off_t)blkno * bsize; 2428 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && (vp->v_flag & VOBJBUF); 2429 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2430 maxsize = imax(maxsize, bsize); 2431 2432 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2433 if (slpflag || slptimeo) { 2434 crit_exit(); 2435 return NULL; 2436 } 2437 goto loop; 2438 } 2439 2440 /* 2441 * This code is used to make sure that a buffer is not 2442 * created while the getnewbuf routine is blocked. 2443 * This can be a problem whether the vnode is locked or not. 2444 * If the buffer is created out from under us, we have to 2445 * throw away the one we just created. There is now window 2446 * race because we are safely running in a critical section 2447 * from the point of the duplicate buffer creation through 2448 * to here, and we've locked the buffer. 2449 */ 2450 if (gbincore(vp, blkno)) { 2451 bp->b_flags |= B_INVAL; 2452 brelse(bp); 2453 goto loop; 2454 } 2455 2456 /* 2457 * Insert the buffer into the hash, so that it can 2458 * be found by incore. bgetvp() and bufhash() 2459 * must be synchronized with each other. Make sure the 2460 * translation layer has been cleared. 2461 */ 2462 bp->b_lblkno = blkno; 2463 bp->b_loffset = offset; 2464 bp->b_bio2.bio_blkno = (daddr_t)-1; 2465 2466 bgetvp(vp, bp); 2467 LIST_REMOVE(bp, b_hash); 2468 bh = bufhash(vp, blkno); 2469 LIST_INSERT_HEAD(bh, bp, b_hash); 2470 2471 /* 2472 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2473 * buffer size starts out as 0, B_CACHE will be set by 2474 * allocbuf() for the VMIO case prior to it testing the 2475 * backing store for validity. 2476 */ 2477 2478 if (vmio) { 2479 bp->b_flags |= B_VMIO; 2480 #if defined(VFS_BIO_DEBUG) 2481 if (vn_canvmio(vp) != TRUE) 2482 printf("getblk: vmioing file type %d???\n", vp->v_type); 2483 #endif 2484 } else { 2485 bp->b_flags &= ~B_VMIO; 2486 } 2487 2488 allocbuf(bp, size); 2489 2490 crit_exit(); 2491 bp->b_flags &= ~B_DONE; 2492 } 2493 return (bp); 2494 } 2495 2496 /* 2497 * geteblk: 2498 * 2499 * Get an empty, disassociated buffer of given size. The buffer is 2500 * initially set to B_INVAL. 2501 * 2502 * critical section protection is not required for the allocbuf() 2503 * call because races are impossible here. 2504 */ 2505 struct buf * 2506 geteblk(int size) 2507 { 2508 struct buf *bp; 2509 int maxsize; 2510 2511 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2512 2513 crit_enter(); 2514 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2515 ; 2516 crit_exit(); 2517 allocbuf(bp, size); 2518 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2519 return (bp); 2520 } 2521 2522 2523 /* 2524 * allocbuf: 2525 * 2526 * This code constitutes the buffer memory from either anonymous system 2527 * memory (in the case of non-VMIO operations) or from an associated 2528 * VM object (in the case of VMIO operations). This code is able to 2529 * resize a buffer up or down. 2530 * 2531 * Note that this code is tricky, and has many complications to resolve 2532 * deadlock or inconsistant data situations. Tread lightly!!! 2533 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2534 * the caller. Calling this code willy nilly can result in the loss of data. 2535 * 2536 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2537 * B_CACHE for the non-VMIO case. 2538 * 2539 * This routine does not need to be called from a critical section but you 2540 * must own the buffer. 2541 */ 2542 int 2543 allocbuf(struct buf *bp, int size) 2544 { 2545 int newbsize, mbsize; 2546 int i; 2547 2548 if (BUF_REFCNT(bp) == 0) 2549 panic("allocbuf: buffer not busy"); 2550 2551 if (bp->b_kvasize < size) 2552 panic("allocbuf: buffer too small"); 2553 2554 if ((bp->b_flags & B_VMIO) == 0) { 2555 caddr_t origbuf; 2556 int origbufsize; 2557 /* 2558 * Just get anonymous memory from the kernel. Don't 2559 * mess with B_CACHE. 2560 */ 2561 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2562 if (bp->b_flags & B_MALLOC) 2563 newbsize = mbsize; 2564 else 2565 newbsize = round_page(size); 2566 2567 if (newbsize < bp->b_bufsize) { 2568 /* 2569 * malloced buffers are not shrunk 2570 */ 2571 if (bp->b_flags & B_MALLOC) { 2572 if (newbsize) { 2573 bp->b_bcount = size; 2574 } else { 2575 free(bp->b_data, M_BIOBUF); 2576 if (bp->b_bufsize) { 2577 bufmallocspace -= bp->b_bufsize; 2578 bufspacewakeup(); 2579 bp->b_bufsize = 0; 2580 } 2581 bp->b_data = bp->b_kvabase; 2582 bp->b_bcount = 0; 2583 bp->b_flags &= ~B_MALLOC; 2584 } 2585 return 1; 2586 } 2587 vm_hold_free_pages( 2588 bp, 2589 (vm_offset_t) bp->b_data + newbsize, 2590 (vm_offset_t) bp->b_data + bp->b_bufsize); 2591 } else if (newbsize > bp->b_bufsize) { 2592 /* 2593 * We only use malloced memory on the first allocation. 2594 * and revert to page-allocated memory when the buffer 2595 * grows. 2596 */ 2597 if ( (bufmallocspace < maxbufmallocspace) && 2598 (bp->b_bufsize == 0) && 2599 (mbsize <= PAGE_SIZE/2)) { 2600 2601 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2602 bp->b_bufsize = mbsize; 2603 bp->b_bcount = size; 2604 bp->b_flags |= B_MALLOC; 2605 bufmallocspace += mbsize; 2606 return 1; 2607 } 2608 origbuf = NULL; 2609 origbufsize = 0; 2610 /* 2611 * If the buffer is growing on its other-than-first allocation, 2612 * then we revert to the page-allocation scheme. 2613 */ 2614 if (bp->b_flags & B_MALLOC) { 2615 origbuf = bp->b_data; 2616 origbufsize = bp->b_bufsize; 2617 bp->b_data = bp->b_kvabase; 2618 if (bp->b_bufsize) { 2619 bufmallocspace -= bp->b_bufsize; 2620 bufspacewakeup(); 2621 bp->b_bufsize = 0; 2622 } 2623 bp->b_flags &= ~B_MALLOC; 2624 newbsize = round_page(newbsize); 2625 } 2626 vm_hold_load_pages( 2627 bp, 2628 (vm_offset_t) bp->b_data + bp->b_bufsize, 2629 (vm_offset_t) bp->b_data + newbsize); 2630 if (origbuf) { 2631 bcopy(origbuf, bp->b_data, origbufsize); 2632 free(origbuf, M_BIOBUF); 2633 } 2634 } 2635 } else { 2636 vm_page_t m; 2637 int desiredpages; 2638 2639 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2640 desiredpages = (size == 0) ? 0 : 2641 num_pages((bp->b_loffset & PAGE_MASK) + newbsize); 2642 2643 if (bp->b_flags & B_MALLOC) 2644 panic("allocbuf: VMIO buffer can't be malloced"); 2645 /* 2646 * Set B_CACHE initially if buffer is 0 length or will become 2647 * 0-length. 2648 */ 2649 if (size == 0 || bp->b_bufsize == 0) 2650 bp->b_flags |= B_CACHE; 2651 2652 if (newbsize < bp->b_bufsize) { 2653 /* 2654 * DEV_BSIZE aligned new buffer size is less then the 2655 * DEV_BSIZE aligned existing buffer size. Figure out 2656 * if we have to remove any pages. 2657 */ 2658 if (desiredpages < bp->b_xio.xio_npages) { 2659 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) { 2660 /* 2661 * the page is not freed here -- it 2662 * is the responsibility of 2663 * vnode_pager_setsize 2664 */ 2665 m = bp->b_xio.xio_pages[i]; 2666 KASSERT(m != bogus_page, 2667 ("allocbuf: bogus page found")); 2668 while (vm_page_sleep_busy(m, TRUE, "biodep")) 2669 ; 2670 2671 bp->b_xio.xio_pages[i] = NULL; 2672 vm_page_unwire(m, 0); 2673 } 2674 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2675 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages)); 2676 bp->b_xio.xio_npages = desiredpages; 2677 } 2678 } else if (size > bp->b_bcount) { 2679 /* 2680 * We are growing the buffer, possibly in a 2681 * byte-granular fashion. 2682 */ 2683 struct vnode *vp; 2684 vm_object_t obj; 2685 vm_offset_t toff; 2686 vm_offset_t tinc; 2687 2688 /* 2689 * Step 1, bring in the VM pages from the object, 2690 * allocating them if necessary. We must clear 2691 * B_CACHE if these pages are not valid for the 2692 * range covered by the buffer. 2693 * 2694 * critical section protection is required to protect 2695 * against interrupts unbusying and freeing pages 2696 * between our vm_page_lookup() and our 2697 * busycheck/wiring call. 2698 */ 2699 vp = bp->b_vp; 2700 VOP_GETVOBJECT(vp, &obj); 2701 2702 crit_enter(); 2703 while (bp->b_xio.xio_npages < desiredpages) { 2704 vm_page_t m; 2705 vm_pindex_t pi; 2706 2707 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages; 2708 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2709 /* 2710 * note: must allocate system pages 2711 * since blocking here could intefere 2712 * with paging I/O, no matter which 2713 * process we are. 2714 */ 2715 m = vm_page_alloc(obj, pi, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM); 2716 if (m == NULL) { 2717 vm_wait(); 2718 vm_pageout_deficit += desiredpages - 2719 bp->b_xio.xio_npages; 2720 } else { 2721 vm_page_wire(m); 2722 vm_page_wakeup(m); 2723 bp->b_flags &= ~B_CACHE; 2724 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 2725 ++bp->b_xio.xio_npages; 2726 } 2727 continue; 2728 } 2729 2730 /* 2731 * We found a page. If we have to sleep on it, 2732 * retry because it might have gotten freed out 2733 * from under us. 2734 * 2735 * We can only test PG_BUSY here. Blocking on 2736 * m->busy might lead to a deadlock: 2737 * 2738 * vm_fault->getpages->cluster_read->allocbuf 2739 * 2740 */ 2741 2742 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 2743 continue; 2744 2745 /* 2746 * We have a good page. Should we wakeup the 2747 * page daemon? 2748 */ 2749 if ((curthread != pagethread) && 2750 ((m->queue - m->pc) == PQ_CACHE) && 2751 ((vmstats.v_free_count + vmstats.v_cache_count) < 2752 (vmstats.v_free_min + vmstats.v_cache_min))) { 2753 pagedaemon_wakeup(); 2754 } 2755 vm_page_flag_clear(m, PG_ZERO); 2756 vm_page_wire(m); 2757 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 2758 ++bp->b_xio.xio_npages; 2759 } 2760 crit_exit(); 2761 2762 /* 2763 * Step 2. We've loaded the pages into the buffer, 2764 * we have to figure out if we can still have B_CACHE 2765 * set. Note that B_CACHE is set according to the 2766 * byte-granular range ( bcount and size ), not the 2767 * aligned range ( newbsize ). 2768 * 2769 * The VM test is against m->valid, which is DEV_BSIZE 2770 * aligned. Needless to say, the validity of the data 2771 * needs to also be DEV_BSIZE aligned. Note that this 2772 * fails with NFS if the server or some other client 2773 * extends the file's EOF. If our buffer is resized, 2774 * B_CACHE may remain set! XXX 2775 */ 2776 2777 toff = bp->b_bcount; 2778 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK); 2779 2780 while ((bp->b_flags & B_CACHE) && toff < size) { 2781 vm_pindex_t pi; 2782 2783 if (tinc > (size - toff)) 2784 tinc = size - toff; 2785 2786 pi = ((bp->b_loffset & PAGE_MASK) + toff) >> 2787 PAGE_SHIFT; 2788 2789 vfs_buf_test_cache( 2790 bp, 2791 bp->b_loffset, 2792 toff, 2793 tinc, 2794 bp->b_xio.xio_pages[pi] 2795 ); 2796 toff += tinc; 2797 tinc = PAGE_SIZE; 2798 } 2799 2800 /* 2801 * Step 3, fixup the KVM pmap. Remember that 2802 * bp->b_data is relative to bp->b_loffset, but 2803 * bp->b_loffset may be offset into the first page. 2804 */ 2805 2806 bp->b_data = (caddr_t) 2807 trunc_page((vm_offset_t)bp->b_data); 2808 pmap_qenter( 2809 (vm_offset_t)bp->b_data, 2810 bp->b_xio.xio_pages, 2811 bp->b_xio.xio_npages 2812 ); 2813 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2814 (vm_offset_t)(bp->b_loffset & PAGE_MASK)); 2815 } 2816 } 2817 if (newbsize < bp->b_bufsize) 2818 bufspacewakeup(); 2819 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2820 bp->b_bcount = size; /* requested buffer size */ 2821 return 1; 2822 } 2823 2824 /* 2825 * biowait: 2826 * 2827 * Wait for buffer I/O completion, returning error status. The buffer 2828 * is left locked and B_DONE on return. B_EINTR is converted into an 2829 * EINTR error and cleared. 2830 */ 2831 int 2832 biowait(struct buf * bp) 2833 { 2834 crit_enter(); 2835 while ((bp->b_flags & B_DONE) == 0) { 2836 if (bp->b_flags & B_READ) 2837 tsleep(bp, 0, "biord", 0); 2838 else 2839 tsleep(bp, 0, "biowr", 0); 2840 } 2841 crit_exit(); 2842 if (bp->b_flags & B_EINTR) { 2843 bp->b_flags &= ~B_EINTR; 2844 return (EINTR); 2845 } 2846 if (bp->b_flags & B_ERROR) { 2847 return (bp->b_error ? bp->b_error : EIO); 2848 } else { 2849 return (0); 2850 } 2851 } 2852 2853 /* 2854 * This associates a tracking count with an I/O. vn_strategy() and 2855 * dev_dstrategy() do this automatically but there are a few cases 2856 * where a vnode or device layer is bypassed when a block translation 2857 * is cached. In such cases bio_start_transaction() may be called on 2858 * the bypassed layers so the system gets an I/O in progress indication 2859 * for those higher layers. 2860 */ 2861 void 2862 bio_start_transaction(struct bio *bio, struct bio_track *track) 2863 { 2864 bio->bio_track = track; 2865 atomic_add_int(&track->bk_active, 1); 2866 } 2867 2868 /* 2869 * Initiate I/O on a vnode. 2870 */ 2871 void 2872 vn_strategy(struct vnode *vp, struct bio *bio) 2873 { 2874 struct bio_track *track; 2875 2876 if (bio->bio_buf->b_flags & B_READ) 2877 track = &vp->v_track_read; 2878 else 2879 track = &vp->v_track_write; 2880 bio->bio_track = track; 2881 atomic_add_int(&track->bk_active, 1); 2882 vop_strategy(*vp->v_ops, vp, bio); 2883 } 2884 2885 2886 /* 2887 * biodone: 2888 * 2889 * Finish I/O on a buffer, optionally calling a completion function. 2890 * This is usually called from an interrupt so process blocking is 2891 * not allowed. 2892 * 2893 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2894 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2895 * assuming B_INVAL is clear. 2896 * 2897 * For the VMIO case, we set B_CACHE if the op was a read and no 2898 * read error occured, or if the op was a write. B_CACHE is never 2899 * set if the buffer is invalid or otherwise uncacheable. 2900 * 2901 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2902 * initiator to leave B_INVAL set to brelse the buffer out of existance 2903 * in the biodone routine. 2904 */ 2905 void 2906 biodone(struct bio *bio) 2907 { 2908 struct buf *bp = bio->bio_buf; 2909 int error; 2910 2911 crit_enter(); 2912 2913 KASSERT(BUF_REFCNTNB(bp) > 0, 2914 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp))); 2915 KASSERT(!(bp->b_flags & B_DONE), 2916 ("biodone: bp %p already done", bp)); 2917 2918 bp->b_flags |= B_DONE; 2919 runningbufwakeup(bp); 2920 2921 /* 2922 * Run up the chain of BIO's. 2923 */ 2924 while (bio) { 2925 biodone_t *done_func; 2926 struct bio_track *track; 2927 2928 /* 2929 * BIO tracking. Most but not all BIOs are tracked. 2930 */ 2931 if ((track = bio->bio_track) != NULL) { 2932 atomic_subtract_int(&track->bk_active, 1); 2933 if (track->bk_active < 0) { 2934 panic("biodone: bad active count bio %p\n", 2935 bio); 2936 } 2937 if (track->bk_waitflag) { 2938 track->bk_waitflag = 0; 2939 wakeup(track); 2940 } 2941 bio->bio_track = NULL; 2942 } 2943 2944 /* 2945 * A bio_done function terminates the loop. The function 2946 * will be responsible for any further chaining and/or 2947 * buffer management. 2948 */ 2949 if ((done_func = bio->bio_done) != NULL) { 2950 bio->bio_done = NULL; 2951 done_func(bio); 2952 crit_exit(); 2953 return; 2954 } 2955 bio = bio->bio_prev; 2956 } 2957 2958 /* 2959 * Special case (XXX) - not a read or write. 2960 */ 2961 if (bp->b_flags & B_FREEBUF) { 2962 brelse(bp); 2963 crit_exit(); 2964 return; 2965 } 2966 2967 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 2968 (*bioops.io_complete)(bp); 2969 2970 if (bp->b_flags & B_VMIO) { 2971 int i; 2972 vm_ooffset_t foff; 2973 vm_page_t m; 2974 vm_object_t obj; 2975 int iosize; 2976 struct vnode *vp = bp->b_vp; 2977 2978 error = VOP_GETVOBJECT(vp, &obj); 2979 2980 #if defined(VFS_BIO_DEBUG) 2981 if (vp->v_holdcnt == 0) { 2982 panic("biodone: zero vnode hold count"); 2983 } 2984 2985 if (error) { 2986 panic("biodone: missing VM object"); 2987 } 2988 2989 if ((vp->v_flag & VOBJBUF) == 0) { 2990 panic("biodone: vnode is not setup for merged cache"); 2991 } 2992 #endif 2993 2994 foff = bp->b_loffset; 2995 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset")); 2996 2997 if (error) { 2998 panic("biodone: no object"); 2999 } 3000 #if defined(VFS_BIO_DEBUG) 3001 if (obj->paging_in_progress < bp->b_xio.xio_npages) { 3002 printf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n", 3003 obj->paging_in_progress, bp->b_xio.xio_npages); 3004 } 3005 #endif 3006 3007 /* 3008 * Set B_CACHE if the op was a normal read and no error 3009 * occured. B_CACHE is set for writes in the b*write() 3010 * routines. 3011 */ 3012 iosize = bp->b_bcount - bp->b_resid; 3013 if ((bp->b_flags & (B_READ|B_FREEBUF|B_INVAL|B_NOCACHE|B_ERROR)) == B_READ) { 3014 bp->b_flags |= B_CACHE; 3015 } 3016 3017 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3018 int bogusflag = 0; 3019 int resid; 3020 3021 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3022 if (resid > iosize) 3023 resid = iosize; 3024 3025 /* 3026 * cleanup bogus pages, restoring the originals. Since 3027 * the originals should still be wired, we don't have 3028 * to worry about interrupt/freeing races destroying 3029 * the VM object association. 3030 */ 3031 m = bp->b_xio.xio_pages[i]; 3032 if (m == bogus_page) { 3033 bogusflag = 1; 3034 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3035 if (m == NULL) 3036 panic("biodone: page disappeared"); 3037 bp->b_xio.xio_pages[i] = m; 3038 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3039 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3040 } 3041 #if defined(VFS_BIO_DEBUG) 3042 if (OFF_TO_IDX(foff) != m->pindex) { 3043 printf( 3044 "biodone: foff(%lu)/m->pindex(%d) mismatch\n", 3045 (unsigned long)foff, m->pindex); 3046 } 3047 #endif 3048 3049 /* 3050 * In the write case, the valid and clean bits are 3051 * already changed correctly ( see bdwrite() ), so we 3052 * only need to do this here in the read case. 3053 */ 3054 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) { 3055 vfs_page_set_valid(bp, foff, i, m); 3056 } 3057 vm_page_flag_clear(m, PG_ZERO); 3058 3059 /* 3060 * when debugging new filesystems or buffer I/O methods, this 3061 * is the most common error that pops up. if you see this, you 3062 * have not set the page busy flag correctly!!! 3063 */ 3064 if (m->busy == 0) { 3065 printf("biodone: page busy < 0, " 3066 "pindex: %d, foff: 0x(%x,%x), " 3067 "resid: %d, index: %d\n", 3068 (int) m->pindex, (int)(foff >> 32), 3069 (int) foff & 0xffffffff, resid, i); 3070 if (!vn_isdisk(vp, NULL)) 3071 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 3072 bp->b_vp->v_mount->mnt_stat.f_iosize, 3073 (int) bp->b_lblkno, 3074 bp->b_flags, bp->b_xio.xio_npages); 3075 else 3076 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 3077 (int) bp->b_lblkno, 3078 bp->b_flags, bp->b_xio.xio_npages); 3079 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 3080 m->valid, m->dirty, m->wire_count); 3081 panic("biodone: page busy < 0"); 3082 } 3083 vm_page_io_finish(m); 3084 vm_object_pip_subtract(obj, 1); 3085 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3086 iosize -= resid; 3087 } 3088 if (obj) 3089 vm_object_pip_wakeupn(obj, 0); 3090 } 3091 3092 /* 3093 * For asynchronous completions, release the buffer now. The brelse 3094 * will do a wakeup there if necessary - so no need to do a wakeup 3095 * here in the async case. The sync case always needs to do a wakeup. 3096 */ 3097 3098 if (bp->b_flags & B_ASYNC) { 3099 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0) 3100 brelse(bp); 3101 else 3102 bqrelse(bp); 3103 } else { 3104 wakeup(bp); 3105 } 3106 crit_exit(); 3107 } 3108 3109 /* 3110 * vfs_unbusy_pages: 3111 * 3112 * This routine is called in lieu of iodone in the case of 3113 * incomplete I/O. This keeps the busy status for pages 3114 * consistant. 3115 */ 3116 void 3117 vfs_unbusy_pages(struct buf *bp) 3118 { 3119 int i; 3120 3121 runningbufwakeup(bp); 3122 if (bp->b_flags & B_VMIO) { 3123 struct vnode *vp = bp->b_vp; 3124 vm_object_t obj; 3125 3126 VOP_GETVOBJECT(vp, &obj); 3127 3128 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3129 vm_page_t m = bp->b_xio.xio_pages[i]; 3130 3131 /* 3132 * When restoring bogus changes the original pages 3133 * should still be wired, so we are in no danger of 3134 * losing the object association and do not need 3135 * critical section protection particularly. 3136 */ 3137 if (m == bogus_page) { 3138 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i); 3139 if (!m) { 3140 panic("vfs_unbusy_pages: page missing"); 3141 } 3142 bp->b_xio.xio_pages[i] = m; 3143 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3144 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3145 } 3146 vm_object_pip_subtract(obj, 1); 3147 vm_page_flag_clear(m, PG_ZERO); 3148 vm_page_io_finish(m); 3149 } 3150 vm_object_pip_wakeupn(obj, 0); 3151 } 3152 } 3153 3154 /* 3155 * vfs_page_set_valid: 3156 * 3157 * Set the valid bits in a page based on the supplied offset. The 3158 * range is restricted to the buffer's size. 3159 * 3160 * This routine is typically called after a read completes. 3161 */ 3162 static void 3163 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3164 { 3165 vm_ooffset_t soff, eoff; 3166 3167 /* 3168 * Start and end offsets in buffer. eoff - soff may not cross a 3169 * page boundry or cross the end of the buffer. The end of the 3170 * buffer, in this case, is our file EOF, not the allocation size 3171 * of the buffer. 3172 */ 3173 soff = off; 3174 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3175 if (eoff > bp->b_loffset + bp->b_bcount) 3176 eoff = bp->b_loffset + bp->b_bcount; 3177 3178 /* 3179 * Set valid range. This is typically the entire buffer and thus the 3180 * entire page. 3181 */ 3182 if (eoff > soff) { 3183 vm_page_set_validclean( 3184 m, 3185 (vm_offset_t) (soff & PAGE_MASK), 3186 (vm_offset_t) (eoff - soff) 3187 ); 3188 } 3189 } 3190 3191 /* 3192 * vfs_busy_pages: 3193 * 3194 * This routine is called before a device strategy routine. 3195 * It is used to tell the VM system that paging I/O is in 3196 * progress, and treat the pages associated with the buffer 3197 * almost as being PG_BUSY. Also the object 'paging_in_progress' 3198 * flag is handled to make sure that the object doesn't become 3199 * inconsistant. 3200 * 3201 * Since I/O has not been initiated yet, certain buffer flags 3202 * such as B_ERROR or B_INVAL may be in an inconsistant state 3203 * and should be ignored. 3204 */ 3205 void 3206 vfs_busy_pages(struct buf *bp, int clear_modify) 3207 { 3208 int i, bogus; 3209 struct proc *p = curthread->td_proc; 3210 3211 if (bp->b_flags & B_VMIO) { 3212 struct vnode *vp = bp->b_vp; 3213 vm_object_t obj; 3214 vm_ooffset_t foff; 3215 3216 VOP_GETVOBJECT(vp, &obj); 3217 foff = bp->b_loffset; 3218 KASSERT(bp->b_loffset != NOOFFSET, 3219 ("vfs_busy_pages: no buffer offset")); 3220 vfs_setdirty(bp); 3221 3222 retry: 3223 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3224 vm_page_t m = bp->b_xio.xio_pages[i]; 3225 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 3226 goto retry; 3227 } 3228 3229 bogus = 0; 3230 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3231 vm_page_t m = bp->b_xio.xio_pages[i]; 3232 3233 vm_page_flag_clear(m, PG_ZERO); 3234 if ((bp->b_flags & B_CLUSTER) == 0) { 3235 vm_object_pip_add(obj, 1); 3236 vm_page_io_start(m); 3237 } 3238 3239 /* 3240 * When readying a buffer for a read ( i.e 3241 * clear_modify == 0 ), it is important to do 3242 * bogus_page replacement for valid pages in 3243 * partially instantiated buffers. Partially 3244 * instantiated buffers can, in turn, occur when 3245 * reconstituting a buffer from its VM backing store 3246 * base. We only have to do this if B_CACHE is 3247 * clear ( which causes the I/O to occur in the 3248 * first place ). The replacement prevents the read 3249 * I/O from overwriting potentially dirty VM-backed 3250 * pages. XXX bogus page replacement is, uh, bogus. 3251 * It may not work properly with small-block devices. 3252 * We need to find a better way. 3253 */ 3254 3255 vm_page_protect(m, VM_PROT_NONE); 3256 if (clear_modify) 3257 vfs_page_set_valid(bp, foff, i, m); 3258 else if (m->valid == VM_PAGE_BITS_ALL && 3259 (bp->b_flags & B_CACHE) == 0) { 3260 bp->b_xio.xio_pages[i] = bogus_page; 3261 bogus++; 3262 } 3263 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3264 } 3265 if (bogus) 3266 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3267 bp->b_xio.xio_pages, bp->b_xio.xio_npages); 3268 } 3269 3270 /* 3271 * This is the easiest place to put the process accounting for the I/O 3272 * for now. 3273 */ 3274 if (p != NULL) { 3275 if (bp->b_flags & B_READ) 3276 p->p_stats->p_ru.ru_inblock++; 3277 else 3278 p->p_stats->p_ru.ru_oublock++; 3279 } 3280 } 3281 3282 /* 3283 * vfs_clean_pages: 3284 * 3285 * Tell the VM system that the pages associated with this buffer 3286 * are clean. This is used for delayed writes where the data is 3287 * going to go to disk eventually without additional VM intevention. 3288 * 3289 * Note that while we only really need to clean through to b_bcount, we 3290 * just go ahead and clean through to b_bufsize. 3291 */ 3292 static void 3293 vfs_clean_pages(struct buf *bp) 3294 { 3295 int i; 3296 3297 if (bp->b_flags & B_VMIO) { 3298 vm_ooffset_t foff; 3299 3300 foff = bp->b_loffset; 3301 KASSERT(foff != NOOFFSET, ("vfs_clean_pages: no buffer offset")); 3302 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3303 vm_page_t m = bp->b_xio.xio_pages[i]; 3304 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3305 vm_ooffset_t eoff = noff; 3306 3307 if (eoff > bp->b_loffset + bp->b_bufsize) 3308 eoff = bp->b_loffset + bp->b_bufsize; 3309 vfs_page_set_valid(bp, foff, i, m); 3310 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3311 foff = noff; 3312 } 3313 } 3314 } 3315 3316 /* 3317 * vfs_bio_set_validclean: 3318 * 3319 * Set the range within the buffer to valid and clean. The range is 3320 * relative to the beginning of the buffer, b_loffset. Note that 3321 * b_loffset itself may be offset from the beginning of the first page. 3322 */ 3323 3324 void 3325 vfs_bio_set_validclean(struct buf *bp, int base, int size) 3326 { 3327 if (bp->b_flags & B_VMIO) { 3328 int i; 3329 int n; 3330 3331 /* 3332 * Fixup base to be relative to beginning of first page. 3333 * Set initial n to be the maximum number of bytes in the 3334 * first page that can be validated. 3335 */ 3336 3337 base += (bp->b_loffset & PAGE_MASK); 3338 n = PAGE_SIZE - (base & PAGE_MASK); 3339 3340 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_xio.xio_npages; ++i) { 3341 vm_page_t m = bp->b_xio.xio_pages[i]; 3342 3343 if (n > size) 3344 n = size; 3345 3346 vm_page_set_validclean(m, base & PAGE_MASK, n); 3347 base += n; 3348 size -= n; 3349 n = PAGE_SIZE; 3350 } 3351 } 3352 } 3353 3354 /* 3355 * vfs_bio_clrbuf: 3356 * 3357 * Clear a buffer. This routine essentially fakes an I/O, so we need 3358 * to clear B_ERROR and B_INVAL. 3359 * 3360 * Note that while we only theoretically need to clear through b_bcount, 3361 * we go ahead and clear through b_bufsize. 3362 */ 3363 3364 void 3365 vfs_bio_clrbuf(struct buf *bp) 3366 { 3367 int i, mask = 0; 3368 caddr_t sa, ea; 3369 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3370 bp->b_flags &= ~(B_INVAL|B_ERROR); 3371 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3372 (bp->b_loffset & PAGE_MASK) == 0) { 3373 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3374 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) { 3375 bp->b_resid = 0; 3376 return; 3377 } 3378 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) && 3379 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) { 3380 bzero(bp->b_data, bp->b_bufsize); 3381 bp->b_xio.xio_pages[0]->valid |= mask; 3382 bp->b_resid = 0; 3383 return; 3384 } 3385 } 3386 ea = sa = bp->b_data; 3387 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) { 3388 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3389 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3390 ea = (caddr_t)(vm_offset_t)ulmin( 3391 (u_long)(vm_offset_t)ea, 3392 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3393 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3394 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask) 3395 continue; 3396 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) { 3397 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) { 3398 bzero(sa, ea - sa); 3399 } 3400 } else { 3401 for (; sa < ea; sa += DEV_BSIZE, j++) { 3402 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) && 3403 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0) 3404 bzero(sa, DEV_BSIZE); 3405 } 3406 } 3407 bp->b_xio.xio_pages[i]->valid |= mask; 3408 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO); 3409 } 3410 bp->b_resid = 0; 3411 } else { 3412 clrbuf(bp); 3413 } 3414 } 3415 3416 /* 3417 * vm_hold_load_pages: 3418 * 3419 * Load pages into the buffer's address space. The pages are 3420 * allocated from the kernel object in order to reduce interference 3421 * with the any VM paging I/O activity. The range of loaded 3422 * pages will be wired. 3423 * 3424 * If a page cannot be allocated, the 'pagedaemon' is woken up to 3425 * retrieve the full range (to - from) of pages. 3426 * 3427 */ 3428 void 3429 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3430 { 3431 vm_offset_t pg; 3432 vm_page_t p; 3433 int index; 3434 3435 to = round_page(to); 3436 from = round_page(from); 3437 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3438 3439 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3440 3441 tryagain: 3442 3443 /* 3444 * Note: must allocate system pages since blocking here 3445 * could intefere with paging I/O, no matter which 3446 * process we are. 3447 */ 3448 p = vm_page_alloc(kernel_object, 3449 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3450 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM); 3451 if (!p) { 3452 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3453 vm_wait(); 3454 goto tryagain; 3455 } 3456 vm_page_wire(p); 3457 p->valid = VM_PAGE_BITS_ALL; 3458 vm_page_flag_clear(p, PG_ZERO); 3459 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 3460 bp->b_xio.xio_pages[index] = p; 3461 vm_page_wakeup(p); 3462 } 3463 bp->b_xio.xio_npages = index; 3464 } 3465 3466 /* 3467 * vm_hold_free_pages: 3468 * 3469 * Return pages associated with the buffer back to the VM system. 3470 * 3471 * The range of pages underlying the buffer's address space will 3472 * be unmapped and un-wired. 3473 */ 3474 void 3475 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3476 { 3477 vm_offset_t pg; 3478 vm_page_t p; 3479 int index, newnpages; 3480 3481 from = round_page(from); 3482 to = round_page(to); 3483 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3484 3485 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3486 p = bp->b_xio.xio_pages[index]; 3487 if (p && (index < bp->b_xio.xio_npages)) { 3488 if (p->busy) { 3489 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 3490 bp->b_bio2.bio_blkno, bp->b_lblkno); 3491 } 3492 bp->b_xio.xio_pages[index] = NULL; 3493 pmap_kremove(pg); 3494 vm_page_busy(p); 3495 vm_page_unwire(p, 0); 3496 vm_page_free(p); 3497 } 3498 } 3499 bp->b_xio.xio_npages = newnpages; 3500 } 3501 3502 /* 3503 * vmapbuf: 3504 * 3505 * Map an IO request into kernel virtual address space. 3506 * 3507 * All requests are (re)mapped into kernel VA space. 3508 * Notice that we use b_bufsize for the size of the buffer 3509 * to be mapped. b_bcount might be modified by the driver. 3510 */ 3511 int 3512 vmapbuf(struct buf *bp) 3513 { 3514 caddr_t addr, v, kva; 3515 vm_paddr_t pa; 3516 int pidx; 3517 int i; 3518 struct vm_page *m; 3519 3520 if ((bp->b_flags & B_PHYS) == 0) 3521 panic("vmapbuf"); 3522 if (bp->b_bufsize < 0) 3523 return (-1); 3524 for (v = bp->b_saveaddr, 3525 addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), 3526 pidx = 0; 3527 addr < bp->b_data + bp->b_bufsize; 3528 addr += PAGE_SIZE, v += PAGE_SIZE, pidx++) { 3529 /* 3530 * Do the vm_fault if needed; do the copy-on-write thing 3531 * when reading stuff off device into memory. 3532 */ 3533 retry: 3534 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data, 3535 (bp->b_flags&B_READ)?(VM_PROT_READ|VM_PROT_WRITE):VM_PROT_READ); 3536 if (i < 0) { 3537 for (i = 0; i < pidx; ++i) { 3538 vm_page_unhold(bp->b_xio.xio_pages[i]); 3539 bp->b_xio.xio_pages[i] = NULL; 3540 } 3541 return(-1); 3542 } 3543 3544 /* 3545 * WARNING! If sparc support is MFCd in the future this will 3546 * have to be changed from pmap_kextract() to pmap_extract() 3547 * ala -current. 3548 */ 3549 #ifdef __sparc64__ 3550 #error "If MFCing sparc support use pmap_extract" 3551 #endif 3552 pa = pmap_kextract((vm_offset_t)addr); 3553 if (pa == 0) { 3554 printf("vmapbuf: warning, race against user address during I/O"); 3555 goto retry; 3556 } 3557 m = PHYS_TO_VM_PAGE(pa); 3558 vm_page_hold(m); 3559 bp->b_xio.xio_pages[pidx] = m; 3560 } 3561 if (pidx > btoc(MAXPHYS)) 3562 panic("vmapbuf: mapped more than MAXPHYS"); 3563 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_xio.xio_pages, pidx); 3564 3565 kva = bp->b_saveaddr; 3566 bp->b_xio.xio_npages = pidx; 3567 bp->b_saveaddr = bp->b_data; 3568 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3569 return(0); 3570 } 3571 3572 /* 3573 * vunmapbuf: 3574 * 3575 * Free the io map PTEs associated with this IO operation. 3576 * We also invalidate the TLB entries and restore the original b_addr. 3577 */ 3578 void 3579 vunmapbuf(struct buf *bp) 3580 { 3581 int pidx; 3582 int npages; 3583 vm_page_t *m; 3584 3585 if ((bp->b_flags & B_PHYS) == 0) 3586 panic("vunmapbuf"); 3587 3588 npages = bp->b_xio.xio_npages; 3589 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3590 npages); 3591 m = bp->b_xio.xio_pages; 3592 for (pidx = 0; pidx < npages; pidx++) 3593 vm_page_unhold(*m++); 3594 3595 bp->b_data = bp->b_saveaddr; 3596 } 3597 3598 /* 3599 * print out statistics from the current status of the buffer pool 3600 * this can be toggeled by the system control option debug.syncprt 3601 */ 3602 #ifdef DEBUG 3603 void 3604 vfs_bufstats(void) 3605 { 3606 int i, j, count; 3607 struct buf *bp; 3608 struct bqueues *dp; 3609 int counts[(MAXBSIZE / PAGE_SIZE) + 1]; 3610 static char *bname[3] = { "LOCKED", "LRU", "AGE" }; 3611 3612 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) { 3613 count = 0; 3614 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++) 3615 counts[j] = 0; 3616 crit_enter(); 3617 TAILQ_FOREACH(bp, dp, b_freelist) { 3618 counts[bp->b_bufsize/PAGE_SIZE]++; 3619 count++; 3620 } 3621 crit_exit(); 3622 printf("%s: total-%d", bname[i], count); 3623 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++) 3624 if (counts[j] != 0) 3625 printf(", %d-%d", j * PAGE_SIZE, counts[j]); 3626 printf("\n"); 3627 } 3628 } 3629 #endif 3630 3631 #include "opt_ddb.h" 3632 #ifdef DDB 3633 #include <ddb/ddb.h> 3634 3635 DB_SHOW_COMMAND(buffer, db_show_buffer) 3636 { 3637 /* get args */ 3638 struct buf *bp = (struct buf *)addr; 3639 3640 if (!have_addr) { 3641 db_printf("usage: show buffer <addr>\n"); 3642 return; 3643 } 3644 3645 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3646 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 3647 "b_resid = %ld\n, b_data = %p, " 3648 "bio_blkno(disk) = %d, bio_blkno(phys) = %d\n", 3649 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3650 bp->b_data, bp->b_bio2.bio_blkno, (bp->b_bio2.bio_next ? bp->b_bio2.bio_next->bio_blkno : (daddr_t)-1)); 3651 if (bp->b_xio.xio_npages) { 3652 int i; 3653 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ", 3654 bp->b_xio.xio_npages); 3655 for (i = 0; i < bp->b_xio.xio_npages; i++) { 3656 vm_page_t m; 3657 m = bp->b_xio.xio_pages[i]; 3658 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3659 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3660 if ((i + 1) < bp->b_xio.xio_npages) 3661 db_printf(","); 3662 } 3663 db_printf("\n"); 3664 } 3665 } 3666 #endif /* DDB */ 3667