vfs_bio.c revision 91700
1/* 2 * Copyright (c) 1994,1997 John S. Dyson 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice immediately at the beginning of the file, without modification, 10 * this list of conditions, and the following disclaimer. 11 * 2. Absolutely no warranty of function or purpose is made by the author 12 * John S. Dyson. 13 * 14 * $FreeBSD: head/sys/kern/vfs_bio.c 91700 2002-03-05 18:20:58Z eivind $ 15 */ 16 17/* 18 * this file contains a new buffer I/O scheme implementing a coherent 19 * VM object and buffer cache scheme. Pains have been taken to make 20 * sure that the performance degradation associated with schemes such 21 * as this is not realized. 22 * 23 * Author: John S. Dyson 24 * Significant help during the development and debugging phases 25 * had been provided by David Greenman, also of the FreeBSD core team. 26 * 27 * see man buf(9) for more info. 28 */ 29 30#include <sys/param.h> 31#include <sys/systm.h> 32#include <sys/bio.h> 33#include <sys/buf.h> 34#include <sys/eventhandler.h> 35#include <sys/lock.h> 36#include <sys/malloc.h> 37#include <sys/mount.h> 38#include <sys/mutex.h> 39#include <sys/kernel.h> 40#include <sys/kthread.h> 41#include <sys/ktr.h> 42#include <sys/proc.h> 43#include <sys/reboot.h> 44#include <sys/resourcevar.h> 45#include <sys/sysctl.h> 46#include <sys/vmmeter.h> 47#include <sys/vnode.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 57static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 58 59struct bio_ops bioops; /* I/O operation notification */ 60 61struct buf_ops buf_ops_bio = { 62 "buf_ops_bio", 63 bwrite 64}; 65 66/* 67 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 68 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 69 */ 70struct buf *buf; /* buffer header pool */ 71struct mtx buftimelock; /* Interlock on setting prio and timo */ 72 73static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 74 vm_offset_t to); 75static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 76 vm_offset_t to); 77static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 78 int pageno, vm_page_t m); 79static void vfs_clean_pages(struct buf * bp); 80static void vfs_setdirty(struct buf *bp); 81static void vfs_vmio_release(struct buf *bp); 82static void vfs_backgroundwritedone(struct buf *bp); 83static int flushbufqueues(void); 84static void buf_daemon __P((void)); 85 86int vmiodirenable = TRUE; 87SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 88 "Use the VM system for directory writes"); 89int runningbufspace; 90SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 91 "Amount of presently outstanding async buffer io"); 92static int bufspace; 93SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 94 "KVA memory used for bufs"); 95static int maxbufspace; 96SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 97 "Maximum allowed value of bufspace (including buf_daemon)"); 98static int bufmallocspace; 99SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 100 "Amount of malloced memory for buffers"); 101static int maxbufmallocspace; 102SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 103 "Maximum amount of malloced memory for buffers"); 104static int lobufspace; 105SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 106 "Minimum amount of buffers we want to have"); 107static int hibufspace; 108SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 109 "Maximum allowed value of bufspace (excluding buf_daemon)"); 110static int bufreusecnt; 111SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 112 "Number of times we have reused a buffer"); 113static int buffreekvacnt; 114SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 115 "Number of times we have freed the KVA space from some buffer"); 116static int bufdefragcnt; 117SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 118 "Number of times we have had to repeat buffer allocation to defragment"); 119static int lorunningspace; 120SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 121 "Minimum preferred space used for in-progress I/O"); 122static int hirunningspace; 123SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 124 "Maximum amount of space to use for in-progress I/O"); 125static int numdirtybuffers; 126SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 127 "Number of buffers that are dirty (has unwritten changes) at the moment"); 128static int lodirtybuffers; 129SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 130 "How many buffers we want to have free before bufdaemon can sleep"); 131static int hidirtybuffers; 132SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 133 "When the number of dirty buffers is considered severe"); 134static int numfreebuffers; 135SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 136 "Number of free buffers"); 137static int lofreebuffers; 138SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 139 "XXX Unused"); 140static int hifreebuffers; 141SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 142 "XXX Complicatedly unused"); 143static int getnewbufcalls; 144SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 145 "Number of calls to getnewbuf"); 146static int getnewbufrestarts; 147SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 148 "Number of times getnewbuf has had to restart a buffer aquisition"); 149static int dobkgrdwrite = 1; 150SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 151 "Do background writes (honoring the BX_BKGRDWRITE flag)?"); 152 153/* 154 * Wakeup point for bufdaemon, as well as indicator of whether it is already 155 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 156 * is idling. 157 */ 158static int bd_request; 159 160/* 161 * bogus page -- for I/O to/from partially complete buffers 162 * this is a temporary solution to the problem, but it is not 163 * really that bad. it would be better to split the buffer 164 * for input in the case of buffers partially already in memory, 165 * but the code is intricate enough already. 166 */ 167vm_page_t bogus_page; 168 169/* 170 * Offset for bogus_page. 171 * XXX bogus_offset should be local to bufinit 172 */ 173static vm_offset_t bogus_offset; 174 175/* 176 * Synchronization (sleep/wakeup) variable for active buffer space requests. 177 * Set when wait starts, cleared prior to wakeup(). 178 * Used in runningbufwakeup() and waitrunningbufspace(). 179 */ 180static int runningbufreq; 181 182/* 183 * Synchronization (sleep/wakeup) variable for buffer requests. 184 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 185 * by and/or. 186 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 187 * getnewbuf(), and getblk(). 188 */ 189static int needsbuffer; 190 191/* 192 * Mask for index into the buffer hash table, which needs to be power of 2 in 193 * size. Set in kern_vfs_bio_buffer_alloc. 194 */ 195static int bufhashmask; 196 197/* 198 * Hash table for all buffers, with a linked list hanging from each table 199 * entry. Set in kern_vfs_bio_buffer_alloc, initialized in buf_init. 200 */ 201static LIST_HEAD(bufhashhdr, buf) *bufhashtbl; 202 203/* 204 * Somewhere to store buffers when they are not in another list, to always 205 * have them in a list (and thus being able to use the same set of operations 206 * on them.) 207 */ 208static struct bufhashhdr invalhash; 209 210/* 211 * Definitions for the buffer free lists. 212 */ 213#define BUFFER_QUEUES 6 /* number of free buffer queues */ 214 215#define QUEUE_NONE 0 /* on no queue */ 216#define QUEUE_LOCKED 1 /* locked buffers */ 217#define QUEUE_CLEAN 2 /* non-B_DELWRI buffers */ 218#define QUEUE_DIRTY 3 /* B_DELWRI buffers */ 219#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 220#define QUEUE_EMPTY 5 /* empty buffer headers */ 221 222/* Queues for free buffers with various properties */ 223static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 224/* 225 * Single global constant for BUF_WMESG, to avoid getting multiple references. 226 * buf_wmesg is referred from macros. 227 */ 228const char *buf_wmesg = BUF_WMESG; 229 230#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 231#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 232#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 233#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 234 235/* 236 * Buffer hash table code. Note that the logical block scans linearly, which 237 * gives us some L1 cache locality. 238 */ 239 240static __inline 241struct bufhashhdr * 242bufhash(struct vnode *vnp, daddr_t bn) 243{ 244 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]); 245} 246 247/* 248 * numdirtywakeup: 249 * 250 * If someone is blocked due to there being too many dirty buffers, 251 * and numdirtybuffers is now reasonable, wake them up. 252 */ 253 254static __inline void 255numdirtywakeup(int level) 256{ 257 if (numdirtybuffers <= level) { 258 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 259 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 260 wakeup(&needsbuffer); 261 } 262 } 263} 264 265/* 266 * bufspacewakeup: 267 * 268 * Called when buffer space is potentially available for recovery. 269 * getnewbuf() will block on this flag when it is unable to free 270 * sufficient buffer space. Buffer space becomes recoverable when 271 * bp's get placed back in the queues. 272 */ 273 274static __inline void 275bufspacewakeup(void) 276{ 277 /* 278 * If someone is waiting for BUF space, wake them up. Even 279 * though we haven't freed the kva space yet, the waiting 280 * process will be able to now. 281 */ 282 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 283 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 284 wakeup(&needsbuffer); 285 } 286} 287 288/* 289 * runningbufwakeup() - in-progress I/O accounting. 290 * 291 */ 292static __inline void 293runningbufwakeup(struct buf *bp) 294{ 295 if (bp->b_runningbufspace) { 296 runningbufspace -= bp->b_runningbufspace; 297 bp->b_runningbufspace = 0; 298 if (runningbufreq && runningbufspace <= lorunningspace) { 299 runningbufreq = 0; 300 wakeup(&runningbufreq); 301 } 302 } 303} 304 305/* 306 * bufcountwakeup: 307 * 308 * Called when a buffer has been added to one of the free queues to 309 * account for the buffer and to wakeup anyone waiting for free buffers. 310 * This typically occurs when large amounts of metadata are being handled 311 * by the buffer cache ( else buffer space runs out first, usually ). 312 */ 313 314static __inline void 315bufcountwakeup(void) 316{ 317 ++numfreebuffers; 318 if (needsbuffer) { 319 needsbuffer &= ~VFS_BIO_NEED_ANY; 320 if (numfreebuffers >= hifreebuffers) 321 needsbuffer &= ~VFS_BIO_NEED_FREE; 322 wakeup(&needsbuffer); 323 } 324} 325 326/* 327 * waitrunningbufspace() 328 * 329 * runningbufspace is a measure of the amount of I/O currently 330 * running. This routine is used in async-write situations to 331 * prevent creating huge backups of pending writes to a device. 332 * Only asynchronous writes are governed by this function. 333 * 334 * Reads will adjust runningbufspace, but will not block based on it. 335 * The read load has a side effect of reducing the allowed write load. 336 * 337 * This does NOT turn an async write into a sync write. It waits 338 * for earlier writes to complete and generally returns before the 339 * caller's write has reached the device. 340 */ 341static __inline void 342waitrunningbufspace(void) 343{ 344 /* 345 * XXX race against wakeup interrupt, currently 346 * protected by Giant. FIXME! 347 */ 348 while (runningbufspace > hirunningspace) { 349 ++runningbufreq; 350 tsleep(&runningbufreq, PVM, "wdrain", 0); 351 } 352} 353 354 355/* 356 * vfs_buf_test_cache: 357 * 358 * Called when a buffer is extended. This function clears the B_CACHE 359 * bit if the newly extended portion of the buffer does not contain 360 * valid data. 361 */ 362static __inline__ 363void 364vfs_buf_test_cache(struct buf *bp, 365 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 366 vm_page_t m) 367{ 368 GIANT_REQUIRED; 369 370 if (bp->b_flags & B_CACHE) { 371 int base = (foff + off) & PAGE_MASK; 372 if (vm_page_is_valid(m, base, size) == 0) 373 bp->b_flags &= ~B_CACHE; 374 } 375} 376 377/* Wake up the buffer deamon if necessary */ 378static __inline__ 379void 380bd_wakeup(int dirtybuflevel) 381{ 382 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 383 bd_request = 1; 384 wakeup(&bd_request); 385 } 386} 387 388/* 389 * bd_speedup - speedup the buffer cache flushing code 390 */ 391 392static __inline__ 393void 394bd_speedup(void) 395{ 396 bd_wakeup(1); 397} 398 399/* 400 * Calculating buffer cache scaling values and reserve space for buffer 401 * headers. This is called during low level kernel initialization and 402 * may be called more then once. We CANNOT write to the memory area 403 * being reserved at this time. 404 */ 405caddr_t 406kern_vfs_bio_buffer_alloc(caddr_t v, int physmem_est) 407{ 408 /* 409 * physmem_est is in pages. Convert it to kilobytes (assumes 410 * PAGE_SIZE is >= 1K) 411 */ 412 physmem_est = physmem_est * (PAGE_SIZE / 1024); 413 414 /* 415 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 416 * For the first 64MB of ram nominally allocate sufficient buffers to 417 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 418 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 419 * the buffer cache we limit the eventual kva reservation to 420 * maxbcache bytes. 421 * 422 * factor represents the 1/4 x ram conversion. 423 */ 424 if (nbuf == 0) { 425 int factor = 4 * BKVASIZE / 1024; 426 427 nbuf = 50; 428 if (physmem_est > 4096) 429 nbuf += min((physmem_est - 4096) / factor, 430 65536 / factor); 431 if (physmem_est > 65536) 432 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 433 434 if (maxbcache && nbuf > maxbcache / BKVASIZE) 435 nbuf = maxbcache / BKVASIZE; 436 } 437 438#if 0 439 /* 440 * Do not allow the buffer_map to be more then 1/2 the size of the 441 * kernel_map. 442 */ 443 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 444 (BKVASIZE * 2)) { 445 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 446 (BKVASIZE * 2); 447 printf("Warning: nbufs capped at %d\n", nbuf); 448 } 449#endif 450 451 /* 452 * swbufs are used as temporary holders for I/O, such as paging I/O. 453 * We have no less then 16 and no more then 256. 454 */ 455 nswbuf = max(min(nbuf/4, 256), 16); 456 457 /* 458 * Reserve space for the buffer cache buffers 459 */ 460 swbuf = (void *)v; 461 v = (caddr_t)(swbuf + nswbuf); 462 buf = (void *)v; 463 v = (caddr_t)(buf + nbuf); 464 465 /* 466 * Calculate the hash table size and reserve space 467 */ 468 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 469 ; 470 bufhashtbl = (void *)v; 471 v = (caddr_t)(bufhashtbl + bufhashmask); 472 --bufhashmask; 473 474 return(v); 475} 476 477/* Initialize the buffer subsystem. Called before use of any buffers. */ 478void 479bufinit(void) 480{ 481 struct buf *bp; 482 int i; 483 484 GIANT_REQUIRED; 485 486 LIST_INIT(&invalhash); 487 mtx_init(&buftimelock, "buftime lock", MTX_DEF); 488 489 for (i = 0; i <= bufhashmask; i++) 490 LIST_INIT(&bufhashtbl[i]); 491 492 /* next, make a null set of free lists */ 493 for (i = 0; i < BUFFER_QUEUES; i++) 494 TAILQ_INIT(&bufqueues[i]); 495 496 /* finally, initialize each buffer header and stick on empty q */ 497 for (i = 0; i < nbuf; i++) { 498 bp = &buf[i]; 499 bzero(bp, sizeof *bp); 500 bp->b_flags = B_INVAL; /* we're just an empty header */ 501 bp->b_dev = NODEV; 502 bp->b_rcred = NOCRED; 503 bp->b_wcred = NOCRED; 504 bp->b_qindex = QUEUE_EMPTY; 505 bp->b_xflags = 0; 506 LIST_INIT(&bp->b_dep); 507 BUF_LOCKINIT(bp); 508 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 509 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 510 } 511 512 /* 513 * maxbufspace is the absolute maximum amount of buffer space we are 514 * allowed to reserve in KVM and in real terms. The absolute maximum 515 * is nominally used by buf_daemon. hibufspace is the nominal maximum 516 * used by most other processes. The differential is required to 517 * ensure that buf_daemon is able to run when other processes might 518 * be blocked waiting for buffer space. 519 * 520 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 521 * this may result in KVM fragmentation which is not handled optimally 522 * by the system. 523 */ 524 maxbufspace = nbuf * BKVASIZE; 525 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 526 lobufspace = hibufspace - MAXBSIZE; 527 528 lorunningspace = 512 * 1024; 529 hirunningspace = 1024 * 1024; 530 531/* 532 * Limit the amount of malloc memory since it is wired permanently into 533 * the kernel space. Even though this is accounted for in the buffer 534 * allocation, we don't want the malloced region to grow uncontrolled. 535 * The malloc scheme improves memory utilization significantly on average 536 * (small) directories. 537 */ 538 maxbufmallocspace = hibufspace / 20; 539 540/* 541 * Reduce the chance of a deadlock occuring by limiting the number 542 * of delayed-write dirty buffers we allow to stack up. 543 */ 544 hidirtybuffers = nbuf / 4 + 20; 545 numdirtybuffers = 0; 546/* 547 * To support extreme low-memory systems, make sure hidirtybuffers cannot 548 * eat up all available buffer space. This occurs when our minimum cannot 549 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 550 * BKVASIZE'd (8K) buffers. 551 */ 552 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 553 hidirtybuffers >>= 1; 554 } 555 lodirtybuffers = hidirtybuffers / 2; 556 557/* 558 * Try to keep the number of free buffers in the specified range, 559 * and give special processes (e.g. like buf_daemon) access to an 560 * emergency reserve. 561 */ 562 lofreebuffers = nbuf / 18 + 5; 563 hifreebuffers = 2 * lofreebuffers; 564 numfreebuffers = nbuf; 565 566/* 567 * Maximum number of async ops initiated per buf_daemon loop. This is 568 * somewhat of a hack at the moment, we really need to limit ourselves 569 * based on the number of bytes of I/O in-transit that were initiated 570 * from buf_daemon. 571 */ 572 573 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 574 bogus_page = vm_page_alloc(kernel_object, 575 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 576 VM_ALLOC_NORMAL); 577 cnt.v_wire_count++; 578} 579 580/* 581 * bfreekva() - free the kva allocation for a buffer. 582 * 583 * Must be called at splbio() or higher as this is the only locking for 584 * buffer_map. 585 * 586 * Since this call frees up buffer space, we call bufspacewakeup(). 587 */ 588static void 589bfreekva(struct buf * bp) 590{ 591 GIANT_REQUIRED; 592 593 if (bp->b_kvasize) { 594 ++buffreekvacnt; 595 bufspace -= bp->b_kvasize; 596 vm_map_delete(buffer_map, 597 (vm_offset_t) bp->b_kvabase, 598 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 599 ); 600 bp->b_kvasize = 0; 601 bufspacewakeup(); 602 } 603} 604 605/* 606 * bremfree: 607 * 608 * Remove the buffer from the appropriate free list. 609 */ 610void 611bremfree(struct buf * bp) 612{ 613 int s = splbio(); 614 int old_qindex = bp->b_qindex; 615 616 GIANT_REQUIRED; 617 618 if (bp->b_qindex != QUEUE_NONE) { 619 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 620 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 621 bp->b_qindex = QUEUE_NONE; 622 } else { 623 if (BUF_REFCNT(bp) <= 1) 624 panic("bremfree: removing a buffer not on a queue"); 625 } 626 627 /* 628 * Fixup numfreebuffers count. If the buffer is invalid or not 629 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 630 * the buffer was free and we must decrement numfreebuffers. 631 */ 632 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 633 switch(old_qindex) { 634 case QUEUE_DIRTY: 635 case QUEUE_CLEAN: 636 case QUEUE_EMPTY: 637 case QUEUE_EMPTYKVA: 638 --numfreebuffers; 639 break; 640 default: 641 break; 642 } 643 } 644 splx(s); 645} 646 647 648/* 649 * Get a buffer with the specified data. Look in the cache first. We 650 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 651 * is set, the buffer is valid and we do not have to do anything ( see 652 * getblk() ). This is really just a special case of breadn(). 653 */ 654int 655bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 656 struct buf ** bpp) 657{ 658 659 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 660} 661 662/* 663 * Operates like bread, but also starts asynchronous I/O on 664 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 665 * to initiating I/O . If B_CACHE is set, the buffer is valid 666 * and we do not have to do anything. 667 */ 668int 669breadn(struct vnode * vp, daddr_t blkno, int size, 670 daddr_t * rablkno, int *rabsize, 671 int cnt, struct ucred * cred, struct buf ** bpp) 672{ 673 struct buf *bp, *rabp; 674 int i; 675 int rv = 0, readwait = 0; 676 677 *bpp = bp = getblk(vp, blkno, size, 0, 0); 678 679 /* if not found in cache, do some I/O */ 680 if ((bp->b_flags & B_CACHE) == 0) { 681 if (curthread != PCPU_GET(idlethread)) 682 curthread->td_proc->p_stats->p_ru.ru_inblock++; 683 bp->b_iocmd = BIO_READ; 684 bp->b_flags &= ~B_INVAL; 685 bp->b_ioflags &= ~BIO_ERROR; 686 if (bp->b_rcred == NOCRED && cred != NOCRED) 687 bp->b_rcred = crhold(cred); 688 vfs_busy_pages(bp, 0); 689 VOP_STRATEGY(vp, bp); 690 ++readwait; 691 } 692 693 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 694 if (inmem(vp, *rablkno)) 695 continue; 696 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 697 698 if ((rabp->b_flags & B_CACHE) == 0) { 699 if (curthread != PCPU_GET(idlethread)) 700 curthread->td_proc->p_stats->p_ru.ru_inblock++; 701 rabp->b_flags |= B_ASYNC; 702 rabp->b_flags &= ~B_INVAL; 703 rabp->b_ioflags &= ~BIO_ERROR; 704 rabp->b_iocmd = BIO_READ; 705 if (rabp->b_rcred == NOCRED && cred != NOCRED) 706 rabp->b_rcred = crhold(cred); 707 vfs_busy_pages(rabp, 0); 708 BUF_KERNPROC(rabp); 709 VOP_STRATEGY(vp, rabp); 710 } else { 711 brelse(rabp); 712 } 713 } 714 715 if (readwait) { 716 rv = bufwait(bp); 717 } 718 return (rv); 719} 720 721/* 722 * Write, release buffer on completion. (Done by iodone 723 * if async). Do not bother writing anything if the buffer 724 * is invalid. 725 * 726 * Note that we set B_CACHE here, indicating that buffer is 727 * fully valid and thus cacheable. This is true even of NFS 728 * now so we set it generally. This could be set either here 729 * or in biodone() since the I/O is synchronous. We put it 730 * here. 731 */ 732 733int 734bwrite(struct buf * bp) 735{ 736 int oldflags, s; 737 struct buf *newbp; 738 739 if (bp->b_flags & B_INVAL) { 740 brelse(bp); 741 return (0); 742 } 743 744 oldflags = bp->b_flags; 745 746 if (BUF_REFCNT(bp) == 0) 747 panic("bwrite: buffer is not busy???"); 748 s = splbio(); 749 /* 750 * If a background write is already in progress, delay 751 * writing this block if it is asynchronous. Otherwise 752 * wait for the background write to complete. 753 */ 754 if (bp->b_xflags & BX_BKGRDINPROG) { 755 if (bp->b_flags & B_ASYNC) { 756 splx(s); 757 bdwrite(bp); 758 return (0); 759 } 760 bp->b_xflags |= BX_BKGRDWAIT; 761 tsleep(&bp->b_xflags, PRIBIO, "biord", 0); 762 if (bp->b_xflags & BX_BKGRDINPROG) 763 panic("bwrite: still writing"); 764 } 765 766 /* Mark the buffer clean */ 767 bundirty(bp); 768 769 /* 770 * If this buffer is marked for background writing and we 771 * do not have to wait for it, make a copy and write the 772 * copy so as to leave this buffer ready for further use. 773 * 774 * This optimization eats a lot of memory. If we have a page 775 * or buffer shortfall we can't do it. 776 */ 777 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 778 (bp->b_flags & B_ASYNC) && 779 !vm_page_count_severe() && 780 !buf_dirty_count_severe()) { 781 if (bp->b_iodone != NULL) { 782 printf("bp->b_iodone = %p\n", bp->b_iodone); 783 panic("bwrite: need chained iodone"); 784 } 785 786 /* get a new block */ 787 newbp = geteblk(bp->b_bufsize); 788 789 /* set it to be identical to the old block */ 790 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 791 bgetvp(bp->b_vp, newbp); 792 newbp->b_lblkno = bp->b_lblkno; 793 newbp->b_blkno = bp->b_blkno; 794 newbp->b_offset = bp->b_offset; 795 newbp->b_iodone = vfs_backgroundwritedone; 796 newbp->b_flags |= B_ASYNC; 797 newbp->b_flags &= ~B_INVAL; 798 799 /* move over the dependencies */ 800 if (LIST_FIRST(&bp->b_dep) != NULL) 801 buf_movedeps(bp, newbp); 802 803 /* 804 * Initiate write on the copy, release the original to 805 * the B_LOCKED queue so that it cannot go away until 806 * the background write completes. If not locked it could go 807 * away and then be reconstituted while it was being written. 808 * If the reconstituted buffer were written, we could end up 809 * with two background copies being written at the same time. 810 */ 811 bp->b_xflags |= BX_BKGRDINPROG; 812 bp->b_flags |= B_LOCKED; 813 bqrelse(bp); 814 bp = newbp; 815 } 816 817 bp->b_flags &= ~B_DONE; 818 bp->b_ioflags &= ~BIO_ERROR; 819 bp->b_flags |= B_WRITEINPROG | B_CACHE; 820 bp->b_iocmd = BIO_WRITE; 821 822 bp->b_vp->v_numoutput++; 823 vfs_busy_pages(bp, 1); 824 825 /* 826 * Normal bwrites pipeline writes 827 */ 828 bp->b_runningbufspace = bp->b_bufsize; 829 runningbufspace += bp->b_runningbufspace; 830 831 if (curthread != PCPU_GET(idlethread)) 832 curthread->td_proc->p_stats->p_ru.ru_oublock++; 833 splx(s); 834 if (oldflags & B_ASYNC) 835 BUF_KERNPROC(bp); 836 BUF_STRATEGY(bp); 837 838 if ((oldflags & B_ASYNC) == 0) { 839 int rtval = bufwait(bp); 840 brelse(bp); 841 return (rtval); 842 } else if ((oldflags & B_NOWDRAIN) == 0) { 843 /* 844 * don't allow the async write to saturate the I/O 845 * system. Deadlocks can occur only if a device strategy 846 * routine (like in MD) turns around and issues another 847 * high-level write, in which case B_NOWDRAIN is expected 848 * to be set. Otherwise we will not deadlock here because 849 * we are blocking waiting for I/O that is already in-progress 850 * to complete. 851 */ 852 waitrunningbufspace(); 853 } 854 855 return (0); 856} 857 858/* 859 * Complete a background write started from bwrite. 860 */ 861static void 862vfs_backgroundwritedone(bp) 863 struct buf *bp; 864{ 865 struct buf *origbp; 866 867 /* 868 * Find the original buffer that we are writing. 869 */ 870 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 871 panic("backgroundwritedone: lost buffer"); 872 /* 873 * Process dependencies then return any unfinished ones. 874 */ 875 if (LIST_FIRST(&bp->b_dep) != NULL) 876 buf_complete(bp); 877 if (LIST_FIRST(&bp->b_dep) != NULL) 878 buf_movedeps(bp, origbp); 879 /* 880 * Clear the BX_BKGRDINPROG flag in the original buffer 881 * and awaken it if it is waiting for the write to complete. 882 * If BX_BKGRDINPROG is not set in the original buffer it must 883 * have been released and re-instantiated - which is not legal. 884 */ 885 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), 886 ("backgroundwritedone: lost buffer2")); 887 origbp->b_xflags &= ~BX_BKGRDINPROG; 888 if (origbp->b_xflags & BX_BKGRDWAIT) { 889 origbp->b_xflags &= ~BX_BKGRDWAIT; 890 wakeup(&origbp->b_xflags); 891 } 892 /* 893 * Clear the B_LOCKED flag and remove it from the locked 894 * queue if it currently resides there. 895 */ 896 origbp->b_flags &= ~B_LOCKED; 897 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 898 bremfree(origbp); 899 bqrelse(origbp); 900 } 901 /* 902 * This buffer is marked B_NOCACHE, so when it is released 903 * by biodone, it will be tossed. We mark it with BIO_READ 904 * to avoid biodone doing a second vwakeup. 905 */ 906 bp->b_flags |= B_NOCACHE; 907 bp->b_iocmd = BIO_READ; 908 bp->b_flags &= ~(B_CACHE | B_DONE); 909 bp->b_iodone = 0; 910 bufdone(bp); 911} 912 913/* 914 * Delayed write. (Buffer is marked dirty). Do not bother writing 915 * anything if the buffer is marked invalid. 916 * 917 * Note that since the buffer must be completely valid, we can safely 918 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 919 * biodone() in order to prevent getblk from writing the buffer 920 * out synchronously. 921 */ 922void 923bdwrite(struct buf * bp) 924{ 925 GIANT_REQUIRED; 926 927 if (BUF_REFCNT(bp) == 0) 928 panic("bdwrite: buffer is not busy"); 929 930 if (bp->b_flags & B_INVAL) { 931 brelse(bp); 932 return; 933 } 934 bdirty(bp); 935 936 /* 937 * Set B_CACHE, indicating that the buffer is fully valid. This is 938 * true even of NFS now. 939 */ 940 bp->b_flags |= B_CACHE; 941 942 /* 943 * This bmap keeps the system from needing to do the bmap later, 944 * perhaps when the system is attempting to do a sync. Since it 945 * is likely that the indirect block -- or whatever other datastructure 946 * that the filesystem needs is still in memory now, it is a good 947 * thing to do this. Note also, that if the pageout daemon is 948 * requesting a sync -- there might not be enough memory to do 949 * the bmap then... So, this is important to do. 950 */ 951 if (bp->b_lblkno == bp->b_blkno) { 952 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 953 } 954 955 /* 956 * Set the *dirty* buffer range based upon the VM system dirty pages. 957 */ 958 vfs_setdirty(bp); 959 960 /* 961 * We need to do this here to satisfy the vnode_pager and the 962 * pageout daemon, so that it thinks that the pages have been 963 * "cleaned". Note that since the pages are in a delayed write 964 * buffer -- the VFS layer "will" see that the pages get written 965 * out on the next sync, or perhaps the cluster will be completed. 966 */ 967 vfs_clean_pages(bp); 968 bqrelse(bp); 969 970 /* 971 * Wakeup the buffer flushing daemon if we have a lot of dirty 972 * buffers (midpoint between our recovery point and our stall 973 * point). 974 */ 975 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 976 977 /* 978 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 979 * due to the softdep code. 980 */ 981} 982 983/* 984 * bdirty: 985 * 986 * Turn buffer into delayed write request. We must clear BIO_READ and 987 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 988 * itself to properly update it in the dirty/clean lists. We mark it 989 * B_DONE to ensure that any asynchronization of the buffer properly 990 * clears B_DONE ( else a panic will occur later ). 991 * 992 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 993 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 994 * should only be called if the buffer is known-good. 995 * 996 * Since the buffer is not on a queue, we do not update the numfreebuffers 997 * count. 998 * 999 * Must be called at splbio(). 1000 * The buffer must be on QUEUE_NONE. 1001 */ 1002void 1003bdirty(bp) 1004 struct buf *bp; 1005{ 1006 KASSERT(bp->b_qindex == QUEUE_NONE, 1007 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1008 bp->b_flags &= ~(B_RELBUF); 1009 bp->b_iocmd = BIO_WRITE; 1010 1011 if ((bp->b_flags & B_DELWRI) == 0) { 1012 bp->b_flags |= B_DONE | B_DELWRI; 1013 reassignbuf(bp, bp->b_vp); 1014 ++numdirtybuffers; 1015 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1016 } 1017} 1018 1019/* 1020 * bundirty: 1021 * 1022 * Clear B_DELWRI for buffer. 1023 * 1024 * Since the buffer is not on a queue, we do not update the numfreebuffers 1025 * count. 1026 * 1027 * Must be called at splbio(). 1028 * The buffer must be on QUEUE_NONE. 1029 */ 1030 1031void 1032bundirty(bp) 1033 struct buf *bp; 1034{ 1035 KASSERT(bp->b_qindex == QUEUE_NONE, 1036 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1037 1038 if (bp->b_flags & B_DELWRI) { 1039 bp->b_flags &= ~B_DELWRI; 1040 reassignbuf(bp, bp->b_vp); 1041 --numdirtybuffers; 1042 numdirtywakeup(lodirtybuffers); 1043 } 1044 /* 1045 * Since it is now being written, we can clear its deferred write flag. 1046 */ 1047 bp->b_flags &= ~B_DEFERRED; 1048} 1049 1050/* 1051 * bawrite: 1052 * 1053 * Asynchronous write. Start output on a buffer, but do not wait for 1054 * it to complete. The buffer is released when the output completes. 1055 * 1056 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1057 * B_INVAL buffers. Not us. 1058 */ 1059void 1060bawrite(struct buf * bp) 1061{ 1062 bp->b_flags |= B_ASYNC; 1063 (void) BUF_WRITE(bp); 1064} 1065 1066/* 1067 * bwillwrite: 1068 * 1069 * Called prior to the locking of any vnodes when we are expecting to 1070 * write. We do not want to starve the buffer cache with too many 1071 * dirty buffers so we block here. By blocking prior to the locking 1072 * of any vnodes we attempt to avoid the situation where a locked vnode 1073 * prevents the various system daemons from flushing related buffers. 1074 */ 1075 1076void 1077bwillwrite(void) 1078{ 1079 if (numdirtybuffers >= hidirtybuffers) { 1080 int s; 1081 1082 s = splbio(); 1083 while (numdirtybuffers >= hidirtybuffers) { 1084 bd_wakeup(1); 1085 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1086 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0); 1087 } 1088 splx(s); 1089 } 1090} 1091 1092/* 1093 * Return true if we have too many dirty buffers. 1094 */ 1095int 1096buf_dirty_count_severe(void) 1097{ 1098 return(numdirtybuffers >= hidirtybuffers); 1099} 1100 1101/* 1102 * brelse: 1103 * 1104 * Release a busy buffer and, if requested, free its resources. The 1105 * buffer will be stashed in the appropriate bufqueue[] allowing it 1106 * to be accessed later as a cache entity or reused for other purposes. 1107 */ 1108void 1109brelse(struct buf * bp) 1110{ 1111 int s; 1112 1113 GIANT_REQUIRED; 1114 1115 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1116 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1117 1118 s = splbio(); 1119 1120 if (bp->b_flags & B_LOCKED) 1121 bp->b_ioflags &= ~BIO_ERROR; 1122 1123 if (bp->b_iocmd == BIO_WRITE && 1124 (bp->b_ioflags & BIO_ERROR) && 1125 !(bp->b_flags & B_INVAL)) { 1126 /* 1127 * Failed write, redirty. Must clear BIO_ERROR to prevent 1128 * pages from being scrapped. If B_INVAL is set then 1129 * this case is not run and the next case is run to 1130 * destroy the buffer. B_INVAL can occur if the buffer 1131 * is outside the range supported by the underlying device. 1132 */ 1133 bp->b_ioflags &= ~BIO_ERROR; 1134 bdirty(bp); 1135 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1136 (bp->b_ioflags & BIO_ERROR) || 1137 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1138 /* 1139 * Either a failed I/O or we were asked to free or not 1140 * cache the buffer. 1141 */ 1142 bp->b_flags |= B_INVAL; 1143 if (LIST_FIRST(&bp->b_dep) != NULL) 1144 buf_deallocate(bp); 1145 if (bp->b_flags & B_DELWRI) { 1146 --numdirtybuffers; 1147 numdirtywakeup(lodirtybuffers); 1148 } 1149 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1150 if ((bp->b_flags & B_VMIO) == 0) { 1151 if (bp->b_bufsize) 1152 allocbuf(bp, 0); 1153 if (bp->b_vp) 1154 brelvp(bp); 1155 } 1156 } 1157 1158 /* 1159 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1160 * is called with B_DELWRI set, the underlying pages may wind up 1161 * getting freed causing a previous write (bdwrite()) to get 'lost' 1162 * because pages associated with a B_DELWRI bp are marked clean. 1163 * 1164 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1165 * if B_DELWRI is set. 1166 * 1167 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1168 * on pages to return pages to the VM page queues. 1169 */ 1170 if (bp->b_flags & B_DELWRI) 1171 bp->b_flags &= ~B_RELBUF; 1172 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1173 bp->b_flags |= B_RELBUF; 1174 1175 /* 1176 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1177 * constituted, not even NFS buffers now. Two flags effect this. If 1178 * B_INVAL, the struct buf is invalidated but the VM object is kept 1179 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1180 * 1181 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1182 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1183 * buffer is also B_INVAL because it hits the re-dirtying code above. 1184 * 1185 * Normally we can do this whether a buffer is B_DELWRI or not. If 1186 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1187 * the commit state and we cannot afford to lose the buffer. If the 1188 * buffer has a background write in progress, we need to keep it 1189 * around to prevent it from being reconstituted and starting a second 1190 * background write. 1191 */ 1192 if ((bp->b_flags & B_VMIO) 1193 && !(bp->b_vp->v_tag == VT_NFS && 1194 !vn_isdisk(bp->b_vp, NULL) && 1195 (bp->b_flags & B_DELWRI)) 1196 ) { 1197 1198 int i, j, resid; 1199 vm_page_t m; 1200 off_t foff; 1201 vm_pindex_t poff; 1202 vm_object_t obj; 1203 struct vnode *vp; 1204 1205 vp = bp->b_vp; 1206 1207 /* 1208 * Get the base offset and length of the buffer. Note that 1209 * in the VMIO case if the buffer block size is not 1210 * page-aligned then b_data pointer may not be page-aligned. 1211 * But our b_pages[] array *IS* page aligned. 1212 * 1213 * block sizes less then DEV_BSIZE (usually 512) are not 1214 * supported due to the page granularity bits (m->valid, 1215 * m->dirty, etc...). 1216 * 1217 * See man buf(9) for more information 1218 */ 1219 resid = bp->b_bufsize; 1220 foff = bp->b_offset; 1221 1222 for (i = 0; i < bp->b_npages; i++) { 1223 int had_bogus = 0; 1224 1225 m = bp->b_pages[i]; 1226 vm_page_flag_clear(m, PG_ZERO); 1227 1228 /* 1229 * If we hit a bogus page, fixup *all* the bogus pages 1230 * now. 1231 */ 1232 if (m == bogus_page) { 1233 VOP_GETVOBJECT(vp, &obj); 1234 poff = OFF_TO_IDX(bp->b_offset); 1235 had_bogus = 1; 1236 1237 for (j = i; j < bp->b_npages; j++) { 1238 vm_page_t mtmp; 1239 mtmp = bp->b_pages[j]; 1240 if (mtmp == bogus_page) { 1241 mtmp = vm_page_lookup(obj, poff + j); 1242 if (!mtmp) { 1243 panic("brelse: page missing\n"); 1244 } 1245 bp->b_pages[j] = mtmp; 1246 } 1247 } 1248 1249 if ((bp->b_flags & B_INVAL) == 0) { 1250 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1251 } 1252 m = bp->b_pages[i]; 1253 } 1254 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1255 int poffset = foff & PAGE_MASK; 1256 int presid = resid > (PAGE_SIZE - poffset) ? 1257 (PAGE_SIZE - poffset) : resid; 1258 1259 KASSERT(presid >= 0, ("brelse: extra page")); 1260 vm_page_set_invalid(m, poffset, presid); 1261 if (had_bogus) 1262 printf("avoided corruption bug in bogus_page/brelse code\n"); 1263 } 1264 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1265 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1266 } 1267 1268 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1269 vfs_vmio_release(bp); 1270 1271 } else if (bp->b_flags & B_VMIO) { 1272 1273 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1274 vfs_vmio_release(bp); 1275 } 1276 1277 } 1278 1279 if (bp->b_qindex != QUEUE_NONE) 1280 panic("brelse: free buffer onto another queue???"); 1281 if (BUF_REFCNT(bp) > 1) { 1282 /* do not release to free list */ 1283 BUF_UNLOCK(bp); 1284 splx(s); 1285 return; 1286 } 1287 1288 /* enqueue */ 1289 1290 /* buffers with no memory */ 1291 if (bp->b_bufsize == 0) { 1292 bp->b_flags |= B_INVAL; 1293 bp->b_xflags &= ~BX_BKGRDWRITE; 1294 if (bp->b_xflags & BX_BKGRDINPROG) 1295 panic("losing buffer 1"); 1296 if (bp->b_kvasize) { 1297 bp->b_qindex = QUEUE_EMPTYKVA; 1298 } else { 1299 bp->b_qindex = QUEUE_EMPTY; 1300 } 1301 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1302 LIST_REMOVE(bp, b_hash); 1303 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1304 bp->b_dev = NODEV; 1305 /* buffers with junk contents */ 1306 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1307 (bp->b_ioflags & BIO_ERROR)) { 1308 bp->b_flags |= B_INVAL; 1309 bp->b_xflags &= ~BX_BKGRDWRITE; 1310 if (bp->b_xflags & BX_BKGRDINPROG) 1311 panic("losing buffer 2"); 1312 bp->b_qindex = QUEUE_CLEAN; 1313 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1314 LIST_REMOVE(bp, b_hash); 1315 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1316 bp->b_dev = NODEV; 1317 1318 /* buffers that are locked */ 1319 } else if (bp->b_flags & B_LOCKED) { 1320 bp->b_qindex = QUEUE_LOCKED; 1321 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1322 1323 /* remaining buffers */ 1324 } else { 1325 if (bp->b_flags & B_DELWRI) 1326 bp->b_qindex = QUEUE_DIRTY; 1327 else 1328 bp->b_qindex = QUEUE_CLEAN; 1329 if (bp->b_flags & B_AGE) 1330 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1331 else 1332 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1333 } 1334 1335 /* 1336 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1337 * on the correct queue. 1338 */ 1339 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) { 1340 bp->b_flags &= ~B_DELWRI; 1341 --numdirtybuffers; 1342 numdirtywakeup(lodirtybuffers); 1343 } 1344 1345 /* 1346 * Fixup numfreebuffers count. The bp is on an appropriate queue 1347 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1348 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1349 * if B_INVAL is set ). 1350 */ 1351 1352 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1353 bufcountwakeup(); 1354 1355 /* 1356 * Something we can maybe free or reuse 1357 */ 1358 if (bp->b_bufsize || bp->b_kvasize) 1359 bufspacewakeup(); 1360 1361 /* unlock */ 1362 BUF_UNLOCK(bp); 1363 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | 1364 B_DIRECT | B_NOWDRAIN); 1365 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1366 panic("brelse: not dirty"); 1367 splx(s); 1368} 1369 1370/* 1371 * Release a buffer back to the appropriate queue but do not try to free 1372 * it. The buffer is expected to be used again soon. 1373 * 1374 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1375 * biodone() to requeue an async I/O on completion. It is also used when 1376 * known good buffers need to be requeued but we think we may need the data 1377 * again soon. 1378 * 1379 * XXX we should be able to leave the B_RELBUF hint set on completion. 1380 */ 1381void 1382bqrelse(struct buf * bp) 1383{ 1384 int s; 1385 1386 s = splbio(); 1387 1388 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1389 1390 if (bp->b_qindex != QUEUE_NONE) 1391 panic("bqrelse: free buffer onto another queue???"); 1392 if (BUF_REFCNT(bp) > 1) { 1393 /* do not release to free list */ 1394 BUF_UNLOCK(bp); 1395 splx(s); 1396 return; 1397 } 1398 if (bp->b_flags & B_LOCKED) { 1399 bp->b_ioflags &= ~BIO_ERROR; 1400 bp->b_qindex = QUEUE_LOCKED; 1401 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1402 /* buffers with stale but valid contents */ 1403 } else if (bp->b_flags & B_DELWRI) { 1404 bp->b_qindex = QUEUE_DIRTY; 1405 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1406 } else if (vm_page_count_severe()) { 1407 /* 1408 * We are too low on memory, we have to try to free the 1409 * buffer (most importantly: the wired pages making up its 1410 * backing store) *now*. 1411 */ 1412 splx(s); 1413 brelse(bp); 1414 return; 1415 } else { 1416 bp->b_qindex = QUEUE_CLEAN; 1417 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1418 } 1419 1420 if ((bp->b_flags & B_LOCKED) == 0 && 1421 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1422 bufcountwakeup(); 1423 } 1424 1425 /* 1426 * Something we can maybe free or reuse. 1427 */ 1428 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1429 bufspacewakeup(); 1430 1431 /* unlock */ 1432 BUF_UNLOCK(bp); 1433 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1434 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1435 panic("bqrelse: not dirty"); 1436 splx(s); 1437} 1438 1439/* Give pages used by the bp back to the VM system (where possible) */ 1440static void 1441vfs_vmio_release(bp) 1442 struct buf *bp; 1443{ 1444 int i; 1445 vm_page_t m; 1446 1447 GIANT_REQUIRED; 1448 1449 for (i = 0; i < bp->b_npages; i++) { 1450 m = bp->b_pages[i]; 1451 bp->b_pages[i] = NULL; 1452 /* 1453 * In order to keep page LRU ordering consistent, put 1454 * everything on the inactive queue. 1455 */ 1456 vm_page_unwire(m, 0); 1457 /* 1458 * We don't mess with busy pages, it is 1459 * the responsibility of the process that 1460 * busied the pages to deal with them. 1461 */ 1462 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1463 continue; 1464 1465 if (m->wire_count == 0) { 1466 vm_page_flag_clear(m, PG_ZERO); 1467 /* 1468 * Might as well free the page if we can and it has 1469 * no valid data. We also free the page if the 1470 * buffer was used for direct I/O 1471 */ 1472 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1473 m->hold_count == 0) { 1474 vm_page_busy(m); 1475 vm_page_protect(m, VM_PROT_NONE); 1476 vm_page_free(m); 1477 } else if (bp->b_flags & B_DIRECT) { 1478 vm_page_try_to_free(m); 1479 } else if (vm_page_count_severe()) { 1480 vm_page_try_to_cache(m); 1481 } 1482 } 1483 } 1484 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1485 1486 if (bp->b_bufsize) { 1487 bufspacewakeup(); 1488 bp->b_bufsize = 0; 1489 } 1490 bp->b_npages = 0; 1491 bp->b_flags &= ~B_VMIO; 1492 if (bp->b_vp) 1493 brelvp(bp); 1494} 1495 1496/* 1497 * Check to see if a block is currently memory resident. 1498 */ 1499struct buf * 1500gbincore(struct vnode * vp, daddr_t blkno) 1501{ 1502 struct buf *bp; 1503 struct bufhashhdr *bh; 1504 1505 bh = bufhash(vp, blkno); 1506 1507 /* Search hash chain */ 1508 LIST_FOREACH(bp, bh, b_hash) { 1509 /* hit */ 1510 if (bp->b_vp == vp && bp->b_lblkno == blkno && 1511 (bp->b_flags & B_INVAL) == 0) { 1512 break; 1513 } 1514 } 1515 return (bp); 1516} 1517 1518/* 1519 * vfs_bio_awrite: 1520 * 1521 * Implement clustered async writes for clearing out B_DELWRI buffers. 1522 * This is much better then the old way of writing only one buffer at 1523 * a time. Note that we may not be presented with the buffers in the 1524 * correct order, so we search for the cluster in both directions. 1525 */ 1526int 1527vfs_bio_awrite(struct buf * bp) 1528{ 1529 int i; 1530 int j; 1531 daddr_t lblkno = bp->b_lblkno; 1532 struct vnode *vp = bp->b_vp; 1533 int s; 1534 int ncl; 1535 struct buf *bpa; 1536 int nwritten; 1537 int size; 1538 int maxcl; 1539 1540 s = splbio(); 1541 /* 1542 * right now we support clustered writing only to regular files. If 1543 * we find a clusterable block we could be in the middle of a cluster 1544 * rather then at the beginning. 1545 */ 1546 if ((vp->v_type == VREG) && 1547 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1548 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1549 1550 size = vp->v_mount->mnt_stat.f_iosize; 1551 maxcl = MAXPHYS / size; 1552 1553 for (i = 1; i < maxcl; i++) { 1554 if ((bpa = gbincore(vp, lblkno + i)) && 1555 BUF_REFCNT(bpa) == 0 && 1556 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1557 (B_DELWRI | B_CLUSTEROK)) && 1558 (bpa->b_bufsize == size)) { 1559 if ((bpa->b_blkno == bpa->b_lblkno) || 1560 (bpa->b_blkno != 1561 bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 1562 break; 1563 } else { 1564 break; 1565 } 1566 } 1567 for (j = 1; i + j <= maxcl && j <= lblkno; j++) { 1568 if ((bpa = gbincore(vp, lblkno - j)) && 1569 BUF_REFCNT(bpa) == 0 && 1570 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1571 (B_DELWRI | B_CLUSTEROK)) && 1572 (bpa->b_bufsize == size)) { 1573 if ((bpa->b_blkno == bpa->b_lblkno) || 1574 (bpa->b_blkno != 1575 bp->b_blkno - ((j * size) >> DEV_BSHIFT))) 1576 break; 1577 } else { 1578 break; 1579 } 1580 } 1581 --j; 1582 ncl = i + j; 1583 /* 1584 * this is a possible cluster write 1585 */ 1586 if (ncl != 1) { 1587 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1588 splx(s); 1589 return nwritten; 1590 } 1591 } 1592 1593 BUF_LOCK(bp, LK_EXCLUSIVE); 1594 bremfree(bp); 1595 bp->b_flags |= B_ASYNC; 1596 1597 splx(s); 1598 /* 1599 * default (old) behavior, writing out only one block 1600 * 1601 * XXX returns b_bufsize instead of b_bcount for nwritten? 1602 */ 1603 nwritten = bp->b_bufsize; 1604 (void) BUF_WRITE(bp); 1605 1606 return nwritten; 1607} 1608 1609/* 1610 * getnewbuf: 1611 * 1612 * Find and initialize a new buffer header, freeing up existing buffers 1613 * in the bufqueues as necessary. The new buffer is returned locked. 1614 * 1615 * Important: B_INVAL is not set. If the caller wishes to throw the 1616 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1617 * 1618 * We block if: 1619 * We have insufficient buffer headers 1620 * We have insufficient buffer space 1621 * buffer_map is too fragmented ( space reservation fails ) 1622 * If we have to flush dirty buffers ( but we try to avoid this ) 1623 * 1624 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1625 * Instead we ask the buf daemon to do it for us. We attempt to 1626 * avoid piecemeal wakeups of the pageout daemon. 1627 */ 1628 1629static struct buf * 1630getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1631{ 1632 struct buf *bp; 1633 struct buf *nbp; 1634 int defrag = 0; 1635 int nqindex; 1636 static int flushingbufs; 1637 1638 GIANT_REQUIRED; 1639 1640 /* 1641 * We can't afford to block since we might be holding a vnode lock, 1642 * which may prevent system daemons from running. We deal with 1643 * low-memory situations by proactively returning memory and running 1644 * async I/O rather then sync I/O. 1645 */ 1646 1647 ++getnewbufcalls; 1648 --getnewbufrestarts; 1649restart: 1650 ++getnewbufrestarts; 1651 1652 /* 1653 * Setup for scan. If we do not have enough free buffers, 1654 * we setup a degenerate case that immediately fails. Note 1655 * that if we are specially marked process, we are allowed to 1656 * dip into our reserves. 1657 * 1658 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1659 * 1660 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1661 * However, there are a number of cases (defragging, reusing, ...) 1662 * where we cannot backup. 1663 */ 1664 nqindex = QUEUE_EMPTYKVA; 1665 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1666 1667 if (nbp == NULL) { 1668 /* 1669 * If no EMPTYKVA buffers and we are either 1670 * defragging or reusing, locate a CLEAN buffer 1671 * to free or reuse. If bufspace useage is low 1672 * skip this step so we can allocate a new buffer. 1673 */ 1674 if (defrag || bufspace >= lobufspace) { 1675 nqindex = QUEUE_CLEAN; 1676 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1677 } 1678 1679 /* 1680 * If we could not find or were not allowed to reuse a 1681 * CLEAN buffer, check to see if it is ok to use an EMPTY 1682 * buffer. We can only use an EMPTY buffer if allocating 1683 * its KVA would not otherwise run us out of buffer space. 1684 */ 1685 if (nbp == NULL && defrag == 0 && 1686 bufspace + maxsize < hibufspace) { 1687 nqindex = QUEUE_EMPTY; 1688 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1689 } 1690 } 1691 1692 /* 1693 * Run scan, possibly freeing data and/or kva mappings on the fly 1694 * depending. 1695 */ 1696 1697 while ((bp = nbp) != NULL) { 1698 int qindex = nqindex; 1699 1700 /* 1701 * Calculate next bp ( we can only use it if we do not block 1702 * or do other fancy things ). 1703 */ 1704 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1705 switch(qindex) { 1706 case QUEUE_EMPTY: 1707 nqindex = QUEUE_EMPTYKVA; 1708 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1709 break; 1710 /* fall through */ 1711 case QUEUE_EMPTYKVA: 1712 nqindex = QUEUE_CLEAN; 1713 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1714 break; 1715 /* fall through */ 1716 case QUEUE_CLEAN: 1717 /* 1718 * nbp is NULL. 1719 */ 1720 break; 1721 } 1722 } 1723 1724 /* 1725 * Sanity Checks 1726 */ 1727 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1728 1729 /* 1730 * Note: we no longer distinguish between VMIO and non-VMIO 1731 * buffers. 1732 */ 1733 1734 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1735 1736 /* 1737 * If we are defragging then we need a buffer with 1738 * b_kvasize != 0. XXX this situation should no longer 1739 * occur, if defrag is non-zero the buffer's b_kvasize 1740 * should also be non-zero at this point. XXX 1741 */ 1742 if (defrag && bp->b_kvasize == 0) { 1743 printf("Warning: defrag empty buffer %p\n", bp); 1744 continue; 1745 } 1746 1747 /* 1748 * Start freeing the bp. This is somewhat involved. nbp 1749 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1750 */ 1751 1752 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1753 panic("getnewbuf: locked buf"); 1754 bremfree(bp); 1755 1756 if (qindex == QUEUE_CLEAN) { 1757 if (bp->b_flags & B_VMIO) { 1758 bp->b_flags &= ~B_ASYNC; 1759 vfs_vmio_release(bp); 1760 } 1761 if (bp->b_vp) 1762 brelvp(bp); 1763 } 1764 1765 /* 1766 * NOTE: nbp is now entirely invalid. We can only restart 1767 * the scan from this point on. 1768 * 1769 * Get the rest of the buffer freed up. b_kva* is still 1770 * valid after this operation. 1771 */ 1772 1773 if (bp->b_rcred != NOCRED) { 1774 crfree(bp->b_rcred); 1775 bp->b_rcred = NOCRED; 1776 } 1777 if (bp->b_wcred != NOCRED) { 1778 crfree(bp->b_wcred); 1779 bp->b_wcred = NOCRED; 1780 } 1781 if (LIST_FIRST(&bp->b_dep) != NULL) 1782 buf_deallocate(bp); 1783 if (bp->b_xflags & BX_BKGRDINPROG) 1784 panic("losing buffer 3"); 1785 LIST_REMOVE(bp, b_hash); 1786 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1787 1788 if (bp->b_bufsize) 1789 allocbuf(bp, 0); 1790 1791 bp->b_flags = 0; 1792 bp->b_ioflags = 0; 1793 bp->b_xflags = 0; 1794 bp->b_dev = NODEV; 1795 bp->b_vp = NULL; 1796 bp->b_blkno = bp->b_lblkno = 0; 1797 bp->b_offset = NOOFFSET; 1798 bp->b_iodone = 0; 1799 bp->b_error = 0; 1800 bp->b_resid = 0; 1801 bp->b_bcount = 0; 1802 bp->b_npages = 0; 1803 bp->b_dirtyoff = bp->b_dirtyend = 0; 1804 bp->b_magic = B_MAGIC_BIO; 1805 bp->b_op = &buf_ops_bio; 1806 1807 LIST_INIT(&bp->b_dep); 1808 1809 /* 1810 * If we are defragging then free the buffer. 1811 */ 1812 if (defrag) { 1813 bp->b_flags |= B_INVAL; 1814 bfreekva(bp); 1815 brelse(bp); 1816 defrag = 0; 1817 goto restart; 1818 } 1819 1820 /* 1821 * If we are overcomitted then recover the buffer and its 1822 * KVM space. This occurs in rare situations when multiple 1823 * processes are blocked in getnewbuf() or allocbuf(). 1824 */ 1825 if (bufspace >= hibufspace) 1826 flushingbufs = 1; 1827 if (flushingbufs && bp->b_kvasize != 0) { 1828 bp->b_flags |= B_INVAL; 1829 bfreekva(bp); 1830 brelse(bp); 1831 goto restart; 1832 } 1833 if (bufspace < lobufspace) 1834 flushingbufs = 0; 1835 break; 1836 } 1837 1838 /* 1839 * If we exhausted our list, sleep as appropriate. We may have to 1840 * wakeup various daemons and write out some dirty buffers. 1841 * 1842 * Generally we are sleeping due to insufficient buffer space. 1843 */ 1844 1845 if (bp == NULL) { 1846 int flags; 1847 char *waitmsg; 1848 1849 if (defrag) { 1850 flags = VFS_BIO_NEED_BUFSPACE; 1851 waitmsg = "nbufkv"; 1852 } else if (bufspace >= hibufspace) { 1853 waitmsg = "nbufbs"; 1854 flags = VFS_BIO_NEED_BUFSPACE; 1855 } else { 1856 waitmsg = "newbuf"; 1857 flags = VFS_BIO_NEED_ANY; 1858 } 1859 1860 bd_speedup(); /* heeeelp */ 1861 1862 needsbuffer |= flags; 1863 while (needsbuffer & flags) { 1864 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, 1865 waitmsg, slptimeo)) 1866 return (NULL); 1867 } 1868 } else { 1869 /* 1870 * We finally have a valid bp. We aren't quite out of the 1871 * woods, we still have to reserve kva space. In order 1872 * to keep fragmentation sane we only allocate kva in 1873 * BKVASIZE chunks. 1874 */ 1875 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1876 1877 if (maxsize != bp->b_kvasize) { 1878 vm_offset_t addr = 0; 1879 1880 bfreekva(bp); 1881 1882 if (vm_map_findspace(buffer_map, 1883 vm_map_min(buffer_map), maxsize, &addr)) { 1884 /* 1885 * Uh oh. Buffer map is to fragmented. We 1886 * must defragment the map. 1887 */ 1888 ++bufdefragcnt; 1889 defrag = 1; 1890 bp->b_flags |= B_INVAL; 1891 brelse(bp); 1892 goto restart; 1893 } 1894 if (addr) { 1895 vm_map_insert(buffer_map, NULL, 0, 1896 addr, addr + maxsize, 1897 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1898 1899 bp->b_kvabase = (caddr_t) addr; 1900 bp->b_kvasize = maxsize; 1901 bufspace += bp->b_kvasize; 1902 ++bufreusecnt; 1903 } 1904 } 1905 bp->b_data = bp->b_kvabase; 1906 } 1907 return(bp); 1908} 1909 1910/* 1911 * buf_daemon: 1912 * 1913 * buffer flushing daemon. Buffers are normally flushed by the 1914 * update daemon but if it cannot keep up this process starts to 1915 * take the load in an attempt to prevent getnewbuf() from blocking. 1916 */ 1917 1918static struct proc *bufdaemonproc; 1919 1920static struct kproc_desc buf_kp = { 1921 "bufdaemon", 1922 buf_daemon, 1923 &bufdaemonproc 1924}; 1925SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1926 1927static void 1928buf_daemon() 1929{ 1930 int s; 1931 1932 mtx_lock(&Giant); 1933 1934 /* 1935 * This process needs to be suspended prior to shutdown sync. 1936 */ 1937 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 1938 SHUTDOWN_PRI_LAST); 1939 1940 /* 1941 * This process is allowed to take the buffer cache to the limit 1942 */ 1943 s = splbio(); 1944 1945 for (;;) { 1946 kthread_suspend_check(bufdaemonproc); 1947 1948 bd_request = 0; 1949 1950 /* 1951 * Do the flush. Limit the amount of in-transit I/O we 1952 * allow to build up, otherwise we would completely saturate 1953 * the I/O system. Wakeup any waiting processes before we 1954 * normally would so they can run in parallel with our drain. 1955 */ 1956 while (numdirtybuffers > lodirtybuffers) { 1957 if (flushbufqueues() == 0) 1958 break; 1959 waitrunningbufspace(); 1960 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 1961 } 1962 1963 /* 1964 * Only clear bd_request if we have reached our low water 1965 * mark. The buf_daemon normally waits 1 second and 1966 * then incrementally flushes any dirty buffers that have 1967 * built up, within reason. 1968 * 1969 * If we were unable to hit our low water mark and couldn't 1970 * find any flushable buffers, we sleep half a second. 1971 * Otherwise we loop immediately. 1972 */ 1973 if (numdirtybuffers <= lodirtybuffers) { 1974 /* 1975 * We reached our low water mark, reset the 1976 * request and sleep until we are needed again. 1977 * The sleep is just so the suspend code works. 1978 */ 1979 bd_request = 0; 1980 tsleep(&bd_request, PVM, "psleep", hz); 1981 } else { 1982 /* 1983 * We couldn't find any flushable dirty buffers but 1984 * still have too many dirty buffers, we 1985 * have to sleep and try again. (rare) 1986 */ 1987 tsleep(&bd_request, PVM, "qsleep", hz / 2); 1988 } 1989 } 1990} 1991 1992/* 1993 * flushbufqueues: 1994 * 1995 * Try to flush a buffer in the dirty queue. We must be careful to 1996 * free up B_INVAL buffers instead of write them, which NFS is 1997 * particularly sensitive to. 1998 */ 1999 2000static int 2001flushbufqueues(void) 2002{ 2003 struct buf *bp; 2004 int r = 0; 2005 2006 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 2007 2008 while (bp) { 2009 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); 2010 if ((bp->b_flags & B_DELWRI) != 0 && 2011 (bp->b_xflags & BX_BKGRDINPROG) == 0) { 2012 if (bp->b_flags & B_INVAL) { 2013 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 2014 panic("flushbufqueues: locked buf"); 2015 bremfree(bp); 2016 brelse(bp); 2017 ++r; 2018 break; 2019 } 2020 if (LIST_FIRST(&bp->b_dep) != NULL && 2021 (bp->b_flags & B_DEFERRED) == 0 && 2022 buf_countdeps(bp, 0)) { 2023 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], 2024 bp, b_freelist); 2025 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], 2026 bp, b_freelist); 2027 bp->b_flags |= B_DEFERRED; 2028 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 2029 continue; 2030 } 2031 vfs_bio_awrite(bp); 2032 ++r; 2033 break; 2034 } 2035 bp = TAILQ_NEXT(bp, b_freelist); 2036 } 2037 return (r); 2038} 2039 2040/* 2041 * Check to see if a block is currently memory resident. 2042 */ 2043struct buf * 2044incore(struct vnode * vp, daddr_t blkno) 2045{ 2046 struct buf *bp; 2047 2048 int s = splbio(); 2049 bp = gbincore(vp, blkno); 2050 splx(s); 2051 return (bp); 2052} 2053 2054/* 2055 * Returns true if no I/O is needed to access the 2056 * associated VM object. This is like incore except 2057 * it also hunts around in the VM system for the data. 2058 */ 2059 2060int 2061inmem(struct vnode * vp, daddr_t blkno) 2062{ 2063 vm_object_t obj; 2064 vm_offset_t toff, tinc, size; 2065 vm_page_t m; 2066 vm_ooffset_t off; 2067 2068 GIANT_REQUIRED; 2069 2070 if (incore(vp, blkno)) 2071 return 1; 2072 if (vp->v_mount == NULL) 2073 return 0; 2074 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_flag & VOBJBUF) == 0) 2075 return 0; 2076 2077 size = PAGE_SIZE; 2078 if (size > vp->v_mount->mnt_stat.f_iosize) 2079 size = vp->v_mount->mnt_stat.f_iosize; 2080 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2081 2082 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2083 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2084 if (!m) 2085 goto notinmem; 2086 tinc = size; 2087 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2088 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2089 if (vm_page_is_valid(m, 2090 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2091 goto notinmem; 2092 } 2093 return 1; 2094 2095notinmem: 2096 return (0); 2097} 2098 2099/* 2100 * vfs_setdirty: 2101 * 2102 * Sets the dirty range for a buffer based on the status of the dirty 2103 * bits in the pages comprising the buffer. 2104 * 2105 * The range is limited to the size of the buffer. 2106 * 2107 * This routine is primarily used by NFS, but is generalized for the 2108 * B_VMIO case. 2109 */ 2110static void 2111vfs_setdirty(struct buf *bp) 2112{ 2113 int i; 2114 vm_object_t object; 2115 2116 GIANT_REQUIRED; 2117 /* 2118 * Degenerate case - empty buffer 2119 */ 2120 2121 if (bp->b_bufsize == 0) 2122 return; 2123 2124 /* 2125 * We qualify the scan for modified pages on whether the 2126 * object has been flushed yet. The OBJ_WRITEABLE flag 2127 * is not cleared simply by protecting pages off. 2128 */ 2129 2130 if ((bp->b_flags & B_VMIO) == 0) 2131 return; 2132 2133 object = bp->b_pages[0]->object; 2134 2135 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2136 printf("Warning: object %p writeable but not mightbedirty\n", object); 2137 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2138 printf("Warning: object %p mightbedirty but not writeable\n", object); 2139 2140 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2141 vm_offset_t boffset; 2142 vm_offset_t eoffset; 2143 2144 /* 2145 * test the pages to see if they have been modified directly 2146 * by users through the VM system. 2147 */ 2148 for (i = 0; i < bp->b_npages; i++) { 2149 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2150 vm_page_test_dirty(bp->b_pages[i]); 2151 } 2152 2153 /* 2154 * Calculate the encompassing dirty range, boffset and eoffset, 2155 * (eoffset - boffset) bytes. 2156 */ 2157 2158 for (i = 0; i < bp->b_npages; i++) { 2159 if (bp->b_pages[i]->dirty) 2160 break; 2161 } 2162 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2163 2164 for (i = bp->b_npages - 1; i >= 0; --i) { 2165 if (bp->b_pages[i]->dirty) { 2166 break; 2167 } 2168 } 2169 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2170 2171 /* 2172 * Fit it to the buffer. 2173 */ 2174 2175 if (eoffset > bp->b_bcount) 2176 eoffset = bp->b_bcount; 2177 2178 /* 2179 * If we have a good dirty range, merge with the existing 2180 * dirty range. 2181 */ 2182 2183 if (boffset < eoffset) { 2184 if (bp->b_dirtyoff > boffset) 2185 bp->b_dirtyoff = boffset; 2186 if (bp->b_dirtyend < eoffset) 2187 bp->b_dirtyend = eoffset; 2188 } 2189 } 2190} 2191 2192/* 2193 * getblk: 2194 * 2195 * Get a block given a specified block and offset into a file/device. 2196 * The buffers B_DONE bit will be cleared on return, making it almost 2197 * ready for an I/O initiation. B_INVAL may or may not be set on 2198 * return. The caller should clear B_INVAL prior to initiating a 2199 * READ. 2200 * 2201 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2202 * an existing buffer. 2203 * 2204 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2205 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2206 * and then cleared based on the backing VM. If the previous buffer is 2207 * non-0-sized but invalid, B_CACHE will be cleared. 2208 * 2209 * If getblk() must create a new buffer, the new buffer is returned with 2210 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2211 * case it is returned with B_INVAL clear and B_CACHE set based on the 2212 * backing VM. 2213 * 2214 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2215 * B_CACHE bit is clear. 2216 * 2217 * What this means, basically, is that the caller should use B_CACHE to 2218 * determine whether the buffer is fully valid or not and should clear 2219 * B_INVAL prior to issuing a read. If the caller intends to validate 2220 * the buffer by loading its data area with something, the caller needs 2221 * to clear B_INVAL. If the caller does this without issuing an I/O, 2222 * the caller should set B_CACHE ( as an optimization ), else the caller 2223 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2224 * a write attempt or if it was a successfull read. If the caller 2225 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2226 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2227 */ 2228struct buf * 2229getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2230{ 2231 struct buf *bp; 2232 int s; 2233 struct bufhashhdr *bh; 2234 2235 if (size > MAXBSIZE) 2236 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2237 2238 s = splbio(); 2239loop: 2240 /* 2241 * Block if we are low on buffers. Certain processes are allowed 2242 * to completely exhaust the buffer cache. 2243 * 2244 * If this check ever becomes a bottleneck it may be better to 2245 * move it into the else, when gbincore() fails. At the moment 2246 * it isn't a problem. 2247 * 2248 * XXX remove if 0 sections (clean this up after its proven) 2249 */ 2250 if (numfreebuffers == 0) { 2251 if (curthread == PCPU_GET(idlethread)) 2252 return NULL; 2253 needsbuffer |= VFS_BIO_NEED_ANY; 2254 } 2255 2256 if ((bp = gbincore(vp, blkno))) { 2257 /* 2258 * Buffer is in-core. If the buffer is not busy, it must 2259 * be on a queue. 2260 */ 2261 2262 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2263 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, 2264 "getblk", slpflag, slptimeo) == ENOLCK) 2265 goto loop; 2266 splx(s); 2267 return (struct buf *) NULL; 2268 } 2269 2270 /* 2271 * The buffer is locked. B_CACHE is cleared if the buffer is 2272 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2273 * and for a VMIO buffer B_CACHE is adjusted according to the 2274 * backing VM cache. 2275 */ 2276 if (bp->b_flags & B_INVAL) 2277 bp->b_flags &= ~B_CACHE; 2278 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2279 bp->b_flags |= B_CACHE; 2280 bremfree(bp); 2281 2282 /* 2283 * check for size inconsistancies for non-VMIO case. 2284 */ 2285 2286 if (bp->b_bcount != size) { 2287 if ((bp->b_flags & B_VMIO) == 0 || 2288 (size > bp->b_kvasize)) { 2289 if (bp->b_flags & B_DELWRI) { 2290 bp->b_flags |= B_NOCACHE; 2291 BUF_WRITE(bp); 2292 } else { 2293 if ((bp->b_flags & B_VMIO) && 2294 (LIST_FIRST(&bp->b_dep) == NULL)) { 2295 bp->b_flags |= B_RELBUF; 2296 brelse(bp); 2297 } else { 2298 bp->b_flags |= B_NOCACHE; 2299 BUF_WRITE(bp); 2300 } 2301 } 2302 goto loop; 2303 } 2304 } 2305 2306 /* 2307 * If the size is inconsistant in the VMIO case, we can resize 2308 * the buffer. This might lead to B_CACHE getting set or 2309 * cleared. If the size has not changed, B_CACHE remains 2310 * unchanged from its previous state. 2311 */ 2312 2313 if (bp->b_bcount != size) 2314 allocbuf(bp, size); 2315 2316 KASSERT(bp->b_offset != NOOFFSET, 2317 ("getblk: no buffer offset")); 2318 2319 /* 2320 * A buffer with B_DELWRI set and B_CACHE clear must 2321 * be committed before we can return the buffer in 2322 * order to prevent the caller from issuing a read 2323 * ( due to B_CACHE not being set ) and overwriting 2324 * it. 2325 * 2326 * Most callers, including NFS and FFS, need this to 2327 * operate properly either because they assume they 2328 * can issue a read if B_CACHE is not set, or because 2329 * ( for example ) an uncached B_DELWRI might loop due 2330 * to softupdates re-dirtying the buffer. In the latter 2331 * case, B_CACHE is set after the first write completes, 2332 * preventing further loops. 2333 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2334 * above while extending the buffer, we cannot allow the 2335 * buffer to remain with B_CACHE set after the write 2336 * completes or it will represent a corrupt state. To 2337 * deal with this we set B_NOCACHE to scrap the buffer 2338 * after the write. 2339 * 2340 * We might be able to do something fancy, like setting 2341 * B_CACHE in bwrite() except if B_DELWRI is already set, 2342 * so the below call doesn't set B_CACHE, but that gets real 2343 * confusing. This is much easier. 2344 */ 2345 2346 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2347 bp->b_flags |= B_NOCACHE; 2348 BUF_WRITE(bp); 2349 goto loop; 2350 } 2351 2352 splx(s); 2353 bp->b_flags &= ~B_DONE; 2354 } else { 2355 /* 2356 * Buffer is not in-core, create new buffer. The buffer 2357 * returned by getnewbuf() is locked. Note that the returned 2358 * buffer is also considered valid (not marked B_INVAL). 2359 */ 2360 int bsize, maxsize, vmio; 2361 off_t offset; 2362 2363 if (vn_isdisk(vp, NULL)) 2364 bsize = DEV_BSIZE; 2365 else if (vp->v_mountedhere) 2366 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2367 else if (vp->v_mount) 2368 bsize = vp->v_mount->mnt_stat.f_iosize; 2369 else 2370 bsize = size; 2371 2372 offset = (off_t)blkno * bsize; 2373 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && (vp->v_flag & VOBJBUF); 2374 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2375 maxsize = imax(maxsize, bsize); 2376 2377 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2378 if (slpflag || slptimeo) { 2379 splx(s); 2380 return NULL; 2381 } 2382 goto loop; 2383 } 2384 2385 /* 2386 * This code is used to make sure that a buffer is not 2387 * created while the getnewbuf routine is blocked. 2388 * This can be a problem whether the vnode is locked or not. 2389 * If the buffer is created out from under us, we have to 2390 * throw away the one we just created. There is now window 2391 * race because we are safely running at splbio() from the 2392 * point of the duplicate buffer creation through to here, 2393 * and we've locked the buffer. 2394 */ 2395 if (gbincore(vp, blkno)) { 2396 bp->b_flags |= B_INVAL; 2397 brelse(bp); 2398 goto loop; 2399 } 2400 2401 /* 2402 * Insert the buffer into the hash, so that it can 2403 * be found by incore. 2404 */ 2405 bp->b_blkno = bp->b_lblkno = blkno; 2406 bp->b_offset = offset; 2407 2408 bgetvp(vp, bp); 2409 LIST_REMOVE(bp, b_hash); 2410 bh = bufhash(vp, blkno); 2411 LIST_INSERT_HEAD(bh, bp, b_hash); 2412 2413 /* 2414 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2415 * buffer size starts out as 0, B_CACHE will be set by 2416 * allocbuf() for the VMIO case prior to it testing the 2417 * backing store for validity. 2418 */ 2419 2420 if (vmio) { 2421 bp->b_flags |= B_VMIO; 2422#if defined(VFS_BIO_DEBUG) 2423 if (vp->v_type != VREG) 2424 printf("getblk: vmioing file type %d???\n", vp->v_type); 2425#endif 2426 } else { 2427 bp->b_flags &= ~B_VMIO; 2428 } 2429 2430 allocbuf(bp, size); 2431 2432 splx(s); 2433 bp->b_flags &= ~B_DONE; 2434 } 2435 return (bp); 2436} 2437 2438/* 2439 * Get an empty, disassociated buffer of given size. The buffer is initially 2440 * set to B_INVAL. 2441 */ 2442struct buf * 2443geteblk(int size) 2444{ 2445 struct buf *bp; 2446 int s; 2447 int maxsize; 2448 2449 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2450 2451 s = splbio(); 2452 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0); 2453 splx(s); 2454 allocbuf(bp, size); 2455 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2456 return (bp); 2457} 2458 2459 2460/* 2461 * This code constitutes the buffer memory from either anonymous system 2462 * memory (in the case of non-VMIO operations) or from an associated 2463 * VM object (in the case of VMIO operations). This code is able to 2464 * resize a buffer up or down. 2465 * 2466 * Note that this code is tricky, and has many complications to resolve 2467 * deadlock or inconsistant data situations. Tread lightly!!! 2468 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2469 * the caller. Calling this code willy nilly can result in the loss of data. 2470 * 2471 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2472 * B_CACHE for the non-VMIO case. 2473 */ 2474 2475int 2476allocbuf(struct buf *bp, int size) 2477{ 2478 int newbsize, mbsize; 2479 int i; 2480 2481 GIANT_REQUIRED; 2482 2483 if (BUF_REFCNT(bp) == 0) 2484 panic("allocbuf: buffer not busy"); 2485 2486 if (bp->b_kvasize < size) 2487 panic("allocbuf: buffer too small"); 2488 2489 if ((bp->b_flags & B_VMIO) == 0) { 2490 caddr_t origbuf; 2491 int origbufsize; 2492 /* 2493 * Just get anonymous memory from the kernel. Don't 2494 * mess with B_CACHE. 2495 */ 2496 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2497#if !defined(NO_B_MALLOC) 2498 if (bp->b_flags & B_MALLOC) 2499 newbsize = mbsize; 2500 else 2501#endif 2502 newbsize = round_page(size); 2503 2504 if (newbsize < bp->b_bufsize) { 2505#if !defined(NO_B_MALLOC) 2506 /* 2507 * malloced buffers are not shrunk 2508 */ 2509 if (bp->b_flags & B_MALLOC) { 2510 if (newbsize) { 2511 bp->b_bcount = size; 2512 } else { 2513 free(bp->b_data, M_BIOBUF); 2514 if (bp->b_bufsize) { 2515 bufmallocspace -= bp->b_bufsize; 2516 bufspacewakeup(); 2517 bp->b_bufsize = 0; 2518 } 2519 bp->b_data = bp->b_kvabase; 2520 bp->b_bcount = 0; 2521 bp->b_flags &= ~B_MALLOC; 2522 } 2523 return 1; 2524 } 2525#endif 2526 vm_hold_free_pages( 2527 bp, 2528 (vm_offset_t) bp->b_data + newbsize, 2529 (vm_offset_t) bp->b_data + bp->b_bufsize); 2530 } else if (newbsize > bp->b_bufsize) { 2531#if !defined(NO_B_MALLOC) 2532 /* 2533 * We only use malloced memory on the first allocation. 2534 * and revert to page-allocated memory when the buffer 2535 * grows. 2536 */ 2537 if ( (bufmallocspace < maxbufmallocspace) && 2538 (bp->b_bufsize == 0) && 2539 (mbsize <= PAGE_SIZE/2)) { 2540 2541 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2542 bp->b_bufsize = mbsize; 2543 bp->b_bcount = size; 2544 bp->b_flags |= B_MALLOC; 2545 bufmallocspace += mbsize; 2546 return 1; 2547 } 2548#endif 2549 origbuf = NULL; 2550 origbufsize = 0; 2551#if !defined(NO_B_MALLOC) 2552 /* 2553 * If the buffer is growing on its other-than-first allocation, 2554 * then we revert to the page-allocation scheme. 2555 */ 2556 if (bp->b_flags & B_MALLOC) { 2557 origbuf = bp->b_data; 2558 origbufsize = bp->b_bufsize; 2559 bp->b_data = bp->b_kvabase; 2560 if (bp->b_bufsize) { 2561 bufmallocspace -= bp->b_bufsize; 2562 bufspacewakeup(); 2563 bp->b_bufsize = 0; 2564 } 2565 bp->b_flags &= ~B_MALLOC; 2566 newbsize = round_page(newbsize); 2567 } 2568#endif 2569 vm_hold_load_pages( 2570 bp, 2571 (vm_offset_t) bp->b_data + bp->b_bufsize, 2572 (vm_offset_t) bp->b_data + newbsize); 2573#if !defined(NO_B_MALLOC) 2574 if (origbuf) { 2575 bcopy(origbuf, bp->b_data, origbufsize); 2576 free(origbuf, M_BIOBUF); 2577 } 2578#endif 2579 } 2580 } else { 2581 vm_page_t m; 2582 int desiredpages; 2583 2584 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2585 desiredpages = (size == 0) ? 0 : 2586 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2587 2588#if !defined(NO_B_MALLOC) 2589 if (bp->b_flags & B_MALLOC) 2590 panic("allocbuf: VMIO buffer can't be malloced"); 2591#endif 2592 /* 2593 * Set B_CACHE initially if buffer is 0 length or will become 2594 * 0-length. 2595 */ 2596 if (size == 0 || bp->b_bufsize == 0) 2597 bp->b_flags |= B_CACHE; 2598 2599 if (newbsize < bp->b_bufsize) { 2600 /* 2601 * DEV_BSIZE aligned new buffer size is less then the 2602 * DEV_BSIZE aligned existing buffer size. Figure out 2603 * if we have to remove any pages. 2604 */ 2605 if (desiredpages < bp->b_npages) { 2606 for (i = desiredpages; i < bp->b_npages; i++) { 2607 /* 2608 * the page is not freed here -- it 2609 * is the responsibility of 2610 * vnode_pager_setsize 2611 */ 2612 m = bp->b_pages[i]; 2613 KASSERT(m != bogus_page, 2614 ("allocbuf: bogus page found")); 2615 while (vm_page_sleep_busy(m, TRUE, "biodep")) 2616 ; 2617 2618 bp->b_pages[i] = NULL; 2619 vm_page_unwire(m, 0); 2620 } 2621 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2622 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2623 bp->b_npages = desiredpages; 2624 } 2625 } else if (size > bp->b_bcount) { 2626 /* 2627 * We are growing the buffer, possibly in a 2628 * byte-granular fashion. 2629 */ 2630 struct vnode *vp; 2631 vm_object_t obj; 2632 vm_offset_t toff; 2633 vm_offset_t tinc; 2634 2635 /* 2636 * Step 1, bring in the VM pages from the object, 2637 * allocating them if necessary. We must clear 2638 * B_CACHE if these pages are not valid for the 2639 * range covered by the buffer. 2640 */ 2641 2642 vp = bp->b_vp; 2643 VOP_GETVOBJECT(vp, &obj); 2644 2645 while (bp->b_npages < desiredpages) { 2646 vm_page_t m; 2647 vm_pindex_t pi; 2648 2649 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2650 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2651 /* 2652 * note: must allocate system pages 2653 * since blocking here could intefere 2654 * with paging I/O, no matter which 2655 * process we are. 2656 */ 2657 m = vm_page_alloc(obj, pi, VM_ALLOC_SYSTEM); 2658 if (m == NULL) { 2659 VM_WAIT; 2660 vm_pageout_deficit += desiredpages - bp->b_npages; 2661 } else { 2662 vm_page_wire(m); 2663 vm_page_wakeup(m); 2664 bp->b_flags &= ~B_CACHE; 2665 bp->b_pages[bp->b_npages] = m; 2666 ++bp->b_npages; 2667 } 2668 continue; 2669 } 2670 2671 /* 2672 * We found a page. If we have to sleep on it, 2673 * retry because it might have gotten freed out 2674 * from under us. 2675 * 2676 * We can only test PG_BUSY here. Blocking on 2677 * m->busy might lead to a deadlock: 2678 * 2679 * vm_fault->getpages->cluster_read->allocbuf 2680 * 2681 */ 2682 2683 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 2684 continue; 2685 2686 /* 2687 * We have a good page. Should we wakeup the 2688 * page daemon? 2689 */ 2690 if ((curproc != pageproc) && 2691 ((m->queue - m->pc) == PQ_CACHE) && 2692 ((cnt.v_free_count + cnt.v_cache_count) < 2693 (cnt.v_free_min + cnt.v_cache_min))) { 2694 pagedaemon_wakeup(); 2695 } 2696 vm_page_flag_clear(m, PG_ZERO); 2697 vm_page_wire(m); 2698 bp->b_pages[bp->b_npages] = m; 2699 ++bp->b_npages; 2700 } 2701 2702 /* 2703 * Step 2. We've loaded the pages into the buffer, 2704 * we have to figure out if we can still have B_CACHE 2705 * set. Note that B_CACHE is set according to the 2706 * byte-granular range ( bcount and size ), new the 2707 * aligned range ( newbsize ). 2708 * 2709 * The VM test is against m->valid, which is DEV_BSIZE 2710 * aligned. Needless to say, the validity of the data 2711 * needs to also be DEV_BSIZE aligned. Note that this 2712 * fails with NFS if the server or some other client 2713 * extends the file's EOF. If our buffer is resized, 2714 * B_CACHE may remain set! XXX 2715 */ 2716 2717 toff = bp->b_bcount; 2718 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2719 2720 while ((bp->b_flags & B_CACHE) && toff < size) { 2721 vm_pindex_t pi; 2722 2723 if (tinc > (size - toff)) 2724 tinc = size - toff; 2725 2726 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2727 PAGE_SHIFT; 2728 2729 vfs_buf_test_cache( 2730 bp, 2731 bp->b_offset, 2732 toff, 2733 tinc, 2734 bp->b_pages[pi] 2735 ); 2736 toff += tinc; 2737 tinc = PAGE_SIZE; 2738 } 2739 2740 /* 2741 * Step 3, fixup the KVM pmap. Remember that 2742 * bp->b_data is relative to bp->b_offset, but 2743 * bp->b_offset may be offset into the first page. 2744 */ 2745 2746 bp->b_data = (caddr_t) 2747 trunc_page((vm_offset_t)bp->b_data); 2748 pmap_qenter( 2749 (vm_offset_t)bp->b_data, 2750 bp->b_pages, 2751 bp->b_npages 2752 ); 2753 2754 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2755 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2756 } 2757 } 2758 if (newbsize < bp->b_bufsize) 2759 bufspacewakeup(); 2760 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2761 bp->b_bcount = size; /* requested buffer size */ 2762 return 1; 2763} 2764 2765/* 2766 * bufwait: 2767 * 2768 * Wait for buffer I/O completion, returning error status. The buffer 2769 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR 2770 * error and cleared. 2771 */ 2772int 2773bufwait(register struct buf * bp) 2774{ 2775 int s; 2776 2777 s = splbio(); 2778 while ((bp->b_flags & B_DONE) == 0) { 2779 if (bp->b_iocmd == BIO_READ) 2780 tsleep(bp, PRIBIO, "biord", 0); 2781 else 2782 tsleep(bp, PRIBIO, "biowr", 0); 2783 } 2784 splx(s); 2785 if (bp->b_flags & B_EINTR) { 2786 bp->b_flags &= ~B_EINTR; 2787 return (EINTR); 2788 } 2789 if (bp->b_ioflags & BIO_ERROR) { 2790 return (bp->b_error ? bp->b_error : EIO); 2791 } else { 2792 return (0); 2793 } 2794} 2795 2796 /* 2797 * Call back function from struct bio back up to struct buf. 2798 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 2799 */ 2800void 2801bufdonebio(struct bio *bp) 2802{ 2803 bufdone(bp->bio_caller2); 2804} 2805 2806/* 2807 * bufdone: 2808 * 2809 * Finish I/O on a buffer, optionally calling a completion function. 2810 * This is usually called from an interrupt so process blocking is 2811 * not allowed. 2812 * 2813 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2814 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2815 * assuming B_INVAL is clear. 2816 * 2817 * For the VMIO case, we set B_CACHE if the op was a read and no 2818 * read error occured, or if the op was a write. B_CACHE is never 2819 * set if the buffer is invalid or otherwise uncacheable. 2820 * 2821 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2822 * initiator to leave B_INVAL set to brelse the buffer out of existance 2823 * in the biodone routine. 2824 */ 2825void 2826bufdone(struct buf *bp) 2827{ 2828 int s, error; 2829 void (*biodone) __P((struct buf *)); 2830 2831 GIANT_REQUIRED; 2832 2833 s = splbio(); 2834 2835 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 2836 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 2837 2838 bp->b_flags |= B_DONE; 2839 runningbufwakeup(bp); 2840 2841 if (bp->b_iocmd == BIO_DELETE) { 2842 brelse(bp); 2843 splx(s); 2844 return; 2845 } 2846 2847 if (bp->b_iocmd == BIO_WRITE) { 2848 vwakeup(bp); 2849 } 2850 2851 /* call optional completion function if requested */ 2852 if (bp->b_iodone != NULL) { 2853 biodone = bp->b_iodone; 2854 bp->b_iodone = NULL; 2855 (*biodone) (bp); 2856 splx(s); 2857 return; 2858 } 2859 if (LIST_FIRST(&bp->b_dep) != NULL) 2860 buf_complete(bp); 2861 2862 if (bp->b_flags & B_VMIO) { 2863 int i; 2864 vm_ooffset_t foff; 2865 vm_page_t m; 2866 vm_object_t obj; 2867 int iosize; 2868 struct vnode *vp = bp->b_vp; 2869 2870 error = VOP_GETVOBJECT(vp, &obj); 2871 2872#if defined(VFS_BIO_DEBUG) 2873 if (vp->v_usecount == 0) { 2874 panic("biodone: zero vnode ref count"); 2875 } 2876 2877 if (error) { 2878 panic("biodone: missing VM object"); 2879 } 2880 2881 if ((vp->v_flag & VOBJBUF) == 0) { 2882 panic("biodone: vnode is not setup for merged cache"); 2883 } 2884#endif 2885 2886 foff = bp->b_offset; 2887 KASSERT(bp->b_offset != NOOFFSET, 2888 ("biodone: no buffer offset")); 2889 2890 if (error) { 2891 panic("biodone: no object"); 2892 } 2893#if defined(VFS_BIO_DEBUG) 2894 if (obj->paging_in_progress < bp->b_npages) { 2895 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 2896 obj->paging_in_progress, bp->b_npages); 2897 } 2898#endif 2899 2900 /* 2901 * Set B_CACHE if the op was a normal read and no error 2902 * occured. B_CACHE is set for writes in the b*write() 2903 * routines. 2904 */ 2905 iosize = bp->b_bcount - bp->b_resid; 2906 if (bp->b_iocmd == BIO_READ && 2907 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 2908 !(bp->b_ioflags & BIO_ERROR)) { 2909 bp->b_flags |= B_CACHE; 2910 } 2911 2912 for (i = 0; i < bp->b_npages; i++) { 2913 int bogusflag = 0; 2914 int resid; 2915 2916 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2917 if (resid > iosize) 2918 resid = iosize; 2919 2920 /* 2921 * cleanup bogus pages, restoring the originals 2922 */ 2923 m = bp->b_pages[i]; 2924 if (m == bogus_page) { 2925 bogusflag = 1; 2926 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2927 if (m == NULL) 2928 panic("biodone: page disappeared!"); 2929 bp->b_pages[i] = m; 2930 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2931 } 2932#if defined(VFS_BIO_DEBUG) 2933 if (OFF_TO_IDX(foff) != m->pindex) { 2934 printf( 2935"biodone: foff(%lu)/m->pindex(%d) mismatch\n", 2936 (unsigned long)foff, m->pindex); 2937 } 2938#endif 2939 2940 /* 2941 * In the write case, the valid and clean bits are 2942 * already changed correctly ( see bdwrite() ), so we 2943 * only need to do this here in the read case. 2944 */ 2945 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 2946 vfs_page_set_valid(bp, foff, i, m); 2947 } 2948 vm_page_flag_clear(m, PG_ZERO); 2949 2950 /* 2951 * when debugging new filesystems or buffer I/O methods, this 2952 * is the most common error that pops up. if you see this, you 2953 * have not set the page busy flag correctly!!! 2954 */ 2955 if (m->busy == 0) { 2956 printf("biodone: page busy < 0, " 2957 "pindex: %d, foff: 0x(%x,%x), " 2958 "resid: %d, index: %d\n", 2959 (int) m->pindex, (int)(foff >> 32), 2960 (int) foff & 0xffffffff, resid, i); 2961 if (!vn_isdisk(vp, NULL)) 2962 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 2963 bp->b_vp->v_mount->mnt_stat.f_iosize, 2964 (int) bp->b_lblkno, 2965 bp->b_flags, bp->b_npages); 2966 else 2967 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 2968 (int) bp->b_lblkno, 2969 bp->b_flags, bp->b_npages); 2970 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2971 m->valid, m->dirty, m->wire_count); 2972 panic("biodone: page busy < 0\n"); 2973 } 2974 vm_page_io_finish(m); 2975 vm_object_pip_subtract(obj, 1); 2976 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2977 iosize -= resid; 2978 } 2979 if (obj) 2980 vm_object_pip_wakeupn(obj, 0); 2981 } 2982 2983 /* 2984 * For asynchronous completions, release the buffer now. The brelse 2985 * will do a wakeup there if necessary - so no need to do a wakeup 2986 * here in the async case. The sync case always needs to do a wakeup. 2987 */ 2988 2989 if (bp->b_flags & B_ASYNC) { 2990 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 2991 brelse(bp); 2992 else 2993 bqrelse(bp); 2994 } else { 2995 wakeup(bp); 2996 } 2997 splx(s); 2998} 2999 3000/* 3001 * This routine is called in lieu of iodone in the case of 3002 * incomplete I/O. This keeps the busy status for pages 3003 * consistant. 3004 */ 3005void 3006vfs_unbusy_pages(struct buf * bp) 3007{ 3008 int i; 3009 3010 GIANT_REQUIRED; 3011 3012 runningbufwakeup(bp); 3013 if (bp->b_flags & B_VMIO) { 3014 struct vnode *vp = bp->b_vp; 3015 vm_object_t obj; 3016 3017 VOP_GETVOBJECT(vp, &obj); 3018 3019 for (i = 0; i < bp->b_npages; i++) { 3020 vm_page_t m = bp->b_pages[i]; 3021 3022 if (m == bogus_page) { 3023 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3024 if (!m) { 3025 panic("vfs_unbusy_pages: page missing\n"); 3026 } 3027 bp->b_pages[i] = m; 3028 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3029 } 3030 vm_object_pip_subtract(obj, 1); 3031 vm_page_flag_clear(m, PG_ZERO); 3032 vm_page_io_finish(m); 3033 } 3034 vm_object_pip_wakeupn(obj, 0); 3035 } 3036} 3037 3038/* 3039 * vfs_page_set_valid: 3040 * 3041 * Set the valid bits in a page based on the supplied offset. The 3042 * range is restricted to the buffer's size. 3043 * 3044 * This routine is typically called after a read completes. 3045 */ 3046static void 3047vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3048{ 3049 vm_ooffset_t soff, eoff; 3050 3051 GIANT_REQUIRED; 3052 /* 3053 * Start and end offsets in buffer. eoff - soff may not cross a 3054 * page boundry or cross the end of the buffer. The end of the 3055 * buffer, in this case, is our file EOF, not the allocation size 3056 * of the buffer. 3057 */ 3058 soff = off; 3059 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3060 if (eoff > bp->b_offset + bp->b_bcount) 3061 eoff = bp->b_offset + bp->b_bcount; 3062 3063 /* 3064 * Set valid range. This is typically the entire buffer and thus the 3065 * entire page. 3066 */ 3067 if (eoff > soff) { 3068 vm_page_set_validclean( 3069 m, 3070 (vm_offset_t) (soff & PAGE_MASK), 3071 (vm_offset_t) (eoff - soff) 3072 ); 3073 } 3074} 3075 3076/* 3077 * This routine is called before a device strategy routine. 3078 * It is used to tell the VM system that paging I/O is in 3079 * progress, and treat the pages associated with the buffer 3080 * almost as being PG_BUSY. Also the object paging_in_progress 3081 * flag is handled to make sure that the object doesn't become 3082 * inconsistant. 3083 * 3084 * Since I/O has not been initiated yet, certain buffer flags 3085 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3086 * and should be ignored. 3087 */ 3088void 3089vfs_busy_pages(struct buf * bp, int clear_modify) 3090{ 3091 int i, bogus; 3092 3093 GIANT_REQUIRED; 3094 3095 if (bp->b_flags & B_VMIO) { 3096 struct vnode *vp = bp->b_vp; 3097 vm_object_t obj; 3098 vm_ooffset_t foff; 3099 3100 VOP_GETVOBJECT(vp, &obj); 3101 foff = bp->b_offset; 3102 KASSERT(bp->b_offset != NOOFFSET, 3103 ("vfs_busy_pages: no buffer offset")); 3104 vfs_setdirty(bp); 3105 3106retry: 3107 for (i = 0; i < bp->b_npages; i++) { 3108 vm_page_t m = bp->b_pages[i]; 3109 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 3110 goto retry; 3111 } 3112 3113 bogus = 0; 3114 for (i = 0; i < bp->b_npages; i++) { 3115 vm_page_t m = bp->b_pages[i]; 3116 3117 vm_page_flag_clear(m, PG_ZERO); 3118 if ((bp->b_flags & B_CLUSTER) == 0) { 3119 vm_object_pip_add(obj, 1); 3120 vm_page_io_start(m); 3121 } 3122 3123 /* 3124 * When readying a buffer for a read ( i.e 3125 * clear_modify == 0 ), it is important to do 3126 * bogus_page replacement for valid pages in 3127 * partially instantiated buffers. Partially 3128 * instantiated buffers can, in turn, occur when 3129 * reconstituting a buffer from its VM backing store 3130 * base. We only have to do this if B_CACHE is 3131 * clear ( which causes the I/O to occur in the 3132 * first place ). The replacement prevents the read 3133 * I/O from overwriting potentially dirty VM-backed 3134 * pages. XXX bogus page replacement is, uh, bogus. 3135 * It may not work properly with small-block devices. 3136 * We need to find a better way. 3137 */ 3138 3139 vm_page_protect(m, VM_PROT_NONE); 3140 if (clear_modify) 3141 vfs_page_set_valid(bp, foff, i, m); 3142 else if (m->valid == VM_PAGE_BITS_ALL && 3143 (bp->b_flags & B_CACHE) == 0) { 3144 bp->b_pages[i] = bogus_page; 3145 bogus++; 3146 } 3147 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3148 } 3149 if (bogus) 3150 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3151 } 3152} 3153 3154/* 3155 * Tell the VM system that the pages associated with this buffer 3156 * are clean. This is used for delayed writes where the data is 3157 * going to go to disk eventually without additional VM intevention. 3158 * 3159 * Note that while we only really need to clean through to b_bcount, we 3160 * just go ahead and clean through to b_bufsize. 3161 */ 3162static void 3163vfs_clean_pages(struct buf * bp) 3164{ 3165 int i; 3166 3167 GIANT_REQUIRED; 3168 3169 if (bp->b_flags & B_VMIO) { 3170 vm_ooffset_t foff; 3171 3172 foff = bp->b_offset; 3173 KASSERT(bp->b_offset != NOOFFSET, 3174 ("vfs_clean_pages: no buffer offset")); 3175 for (i = 0; i < bp->b_npages; i++) { 3176 vm_page_t m = bp->b_pages[i]; 3177 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3178 vm_ooffset_t eoff = noff; 3179 3180 if (eoff > bp->b_offset + bp->b_bufsize) 3181 eoff = bp->b_offset + bp->b_bufsize; 3182 vfs_page_set_valid(bp, foff, i, m); 3183 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3184 foff = noff; 3185 } 3186 } 3187} 3188 3189/* 3190 * vfs_bio_set_validclean: 3191 * 3192 * Set the range within the buffer to valid and clean. The range is 3193 * relative to the beginning of the buffer, b_offset. Note that b_offset 3194 * itself may be offset from the beginning of the first page. 3195 * 3196 */ 3197 3198void 3199vfs_bio_set_validclean(struct buf *bp, int base, int size) 3200{ 3201 if (bp->b_flags & B_VMIO) { 3202 int i; 3203 int n; 3204 3205 /* 3206 * Fixup base to be relative to beginning of first page. 3207 * Set initial n to be the maximum number of bytes in the 3208 * first page that can be validated. 3209 */ 3210 3211 base += (bp->b_offset & PAGE_MASK); 3212 n = PAGE_SIZE - (base & PAGE_MASK); 3213 3214 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3215 vm_page_t m = bp->b_pages[i]; 3216 3217 if (n > size) 3218 n = size; 3219 3220 vm_page_set_validclean(m, base & PAGE_MASK, n); 3221 base += n; 3222 size -= n; 3223 n = PAGE_SIZE; 3224 } 3225 } 3226} 3227 3228/* 3229 * vfs_bio_clrbuf: 3230 * 3231 * clear a buffer. This routine essentially fakes an I/O, so we need 3232 * to clear BIO_ERROR and B_INVAL. 3233 * 3234 * Note that while we only theoretically need to clear through b_bcount, 3235 * we go ahead and clear through b_bufsize. 3236 */ 3237 3238void 3239vfs_bio_clrbuf(struct buf *bp) { 3240 int i, mask = 0; 3241 caddr_t sa, ea; 3242 3243 GIANT_REQUIRED; 3244 3245 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3246 bp->b_flags &= ~B_INVAL; 3247 bp->b_ioflags &= ~BIO_ERROR; 3248 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3249 (bp->b_offset & PAGE_MASK) == 0) { 3250 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3251 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3252 ((bp->b_pages[0]->valid & mask) != mask)) { 3253 bzero(bp->b_data, bp->b_bufsize); 3254 } 3255 bp->b_pages[0]->valid |= mask; 3256 bp->b_resid = 0; 3257 return; 3258 } 3259 ea = sa = bp->b_data; 3260 for(i=0;i<bp->b_npages;i++,sa=ea) { 3261 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3262 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3263 ea = (caddr_t)(vm_offset_t)ulmin( 3264 (u_long)(vm_offset_t)ea, 3265 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3266 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3267 if ((bp->b_pages[i]->valid & mask) == mask) 3268 continue; 3269 if ((bp->b_pages[i]->valid & mask) == 0) { 3270 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3271 bzero(sa, ea - sa); 3272 } 3273 } else { 3274 for (; sa < ea; sa += DEV_BSIZE, j++) { 3275 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3276 (bp->b_pages[i]->valid & (1<<j)) == 0) 3277 bzero(sa, DEV_BSIZE); 3278 } 3279 } 3280 bp->b_pages[i]->valid |= mask; 3281 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3282 } 3283 bp->b_resid = 0; 3284 } else { 3285 clrbuf(bp); 3286 } 3287} 3288 3289/* 3290 * vm_hold_load_pages and vm_hold_free_pages get pages into 3291 * a buffers address space. The pages are anonymous and are 3292 * not associated with a file object. 3293 */ 3294static void 3295vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3296{ 3297 vm_offset_t pg; 3298 vm_page_t p; 3299 int index; 3300 3301 GIANT_REQUIRED; 3302 3303 to = round_page(to); 3304 from = round_page(from); 3305 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3306 3307 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3308tryagain: 3309 /* 3310 * note: must allocate system pages since blocking here 3311 * could intefere with paging I/O, no matter which 3312 * process we are. 3313 */ 3314 p = vm_page_alloc(kernel_object, 3315 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3316 VM_ALLOC_SYSTEM); 3317 if (!p) { 3318 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3319 VM_WAIT; 3320 goto tryagain; 3321 } 3322 vm_page_wire(p); 3323 p->valid = VM_PAGE_BITS_ALL; 3324 vm_page_flag_clear(p, PG_ZERO); 3325 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 3326 bp->b_pages[index] = p; 3327 vm_page_wakeup(p); 3328 } 3329 bp->b_npages = index; 3330} 3331 3332/* Return pages associated with this buf to the vm system */ 3333void 3334vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3335{ 3336 vm_offset_t pg; 3337 vm_page_t p; 3338 int index, newnpages; 3339 3340 GIANT_REQUIRED; 3341 3342 from = round_page(from); 3343 to = round_page(to); 3344 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3345 3346 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3347 p = bp->b_pages[index]; 3348 if (p && (index < bp->b_npages)) { 3349 if (p->busy) { 3350 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 3351 bp->b_blkno, bp->b_lblkno); 3352 } 3353 bp->b_pages[index] = NULL; 3354 pmap_kremove(pg); 3355 vm_page_busy(p); 3356 vm_page_unwire(p, 0); 3357 vm_page_free(p); 3358 } 3359 } 3360 bp->b_npages = newnpages; 3361} 3362 3363 3364#include "opt_ddb.h" 3365#ifdef DDB 3366#include <ddb/ddb.h> 3367 3368/* DDB command to show buffer data */ 3369DB_SHOW_COMMAND(buffer, db_show_buffer) 3370{ 3371 /* get args */ 3372 struct buf *bp = (struct buf *)addr; 3373 3374 if (!have_addr) { 3375 db_printf("usage: show buffer <addr>\n"); 3376 return; 3377 } 3378 3379 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3380 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 3381 "b_resid = %ld\nb_dev = (%d,%d), b_data = %p, " 3382 "b_blkno = %d, b_pblkno = %d\n", 3383 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3384 major(bp->b_dev), minor(bp->b_dev), 3385 bp->b_data, bp->b_blkno, bp->b_pblkno); 3386 if (bp->b_npages) { 3387 int i; 3388 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3389 for (i = 0; i < bp->b_npages; i++) { 3390 vm_page_t m; 3391 m = bp->b_pages[i]; 3392 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3393 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3394 if ((i + 1) < bp->b_npages) 3395 db_printf(","); 3396 } 3397 db_printf("\n"); 3398 } 3399} 3400#endif /* DDB */ 3401