vfs_bio.c revision 75573
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 75573 2001-04-17 05:37:51Z mckusick $ 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 *buf; /* buffer header pool */ 62struct swqueue bswlist; 63struct mtx buftimelock; /* Interlock on setting prio and timo */ 64 65static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 66 vm_offset_t to); 67static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 68 vm_offset_t to); 69static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 70 int pageno, vm_page_t m); 71static void vfs_clean_pages(struct buf * bp); 72static void vfs_setdirty(struct buf *bp); 73static void vfs_vmio_release(struct buf *bp); 74static void vfs_backgroundwritedone(struct buf *bp); 75static int flushbufqueues(void); 76 77static int bd_request; 78 79static void buf_daemon __P((void)); 80/* 81 * bogus page -- for I/O to/from partially complete buffers 82 * this is a temporary solution to the problem, but it is not 83 * really that bad. it would be better to split the buffer 84 * for input in the case of buffers partially already in memory, 85 * but the code is intricate enough already. 86 */ 87vm_page_t bogus_page; 88int vmiodirenable = FALSE; 89int runningbufspace; 90static vm_offset_t bogus_offset; 91 92static int bufspace, maxbufspace, 93 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace; 94static int bufreusecnt, bufdefragcnt, buffreekvacnt; 95static int needsbuffer; 96static int lorunningspace, hirunningspace, runningbufreq; 97static int numdirtybuffers, lodirtybuffers, hidirtybuffers; 98static int numfreebuffers, lofreebuffers, hifreebuffers; 99static int getnewbufcalls; 100static int getnewbufrestarts; 101 102SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, 103 &numdirtybuffers, 0, ""); 104SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, 105 &lodirtybuffers, 0, ""); 106SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, 107 &hidirtybuffers, 0, ""); 108SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, 109 &numfreebuffers, 0, ""); 110SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, 111 &lofreebuffers, 0, ""); 112SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, 113 &hifreebuffers, 0, ""); 114SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, 115 &runningbufspace, 0, ""); 116SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, 117 &lorunningspace, 0, ""); 118SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, 119 &hirunningspace, 0, ""); 120SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, 121 &maxbufspace, 0, ""); 122SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, 123 &hibufspace, 0, ""); 124SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, 125 &lobufspace, 0, ""); 126SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, 127 &bufspace, 0, ""); 128SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, 129 &maxbufmallocspace, 0, ""); 130SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, 131 &bufmallocspace, 0, ""); 132SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, 133 &getnewbufcalls, 0, ""); 134SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, 135 &getnewbufrestarts, 0, ""); 136SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, 137 &vmiodirenable, 0, ""); 138SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, 139 &bufdefragcnt, 0, ""); 140SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, 141 &buffreekvacnt, 0, ""); 142SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, 143 &bufreusecnt, 0, ""); 144 145static int bufhashmask; 146static LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash; 147struct bqueues bufqueues[BUFFER_QUEUES] = { { 0 } }; 148char *buf_wmesg = BUF_WMESG; 149 150extern int vm_swap_size; 151 152#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 153#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 154#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 155#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 156 157/* 158 * Buffer hash table code. Note that the logical block scans linearly, which 159 * gives us some L1 cache locality. 160 */ 161 162static __inline 163struct bufhashhdr * 164bufhash(struct vnode *vnp, daddr_t bn) 165{ 166 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]); 167} 168 169/* 170 * numdirtywakeup: 171 * 172 * If someone is blocked due to there being too many dirty buffers, 173 * and numdirtybuffers is now reasonable, wake them up. 174 */ 175 176static __inline void 177numdirtywakeup(int level) 178{ 179 if (numdirtybuffers <= level) { 180 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 181 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 182 wakeup(&needsbuffer); 183 } 184 } 185} 186 187/* 188 * bufspacewakeup: 189 * 190 * Called when buffer space is potentially available for recovery. 191 * getnewbuf() will block on this flag when it is unable to free 192 * sufficient buffer space. Buffer space becomes recoverable when 193 * bp's get placed back in the queues. 194 */ 195 196static __inline void 197bufspacewakeup(void) 198{ 199 /* 200 * If someone is waiting for BUF space, wake them up. Even 201 * though we haven't freed the kva space yet, the waiting 202 * process will be able to now. 203 */ 204 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 205 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 206 wakeup(&needsbuffer); 207 } 208} 209 210/* 211 * runningbufwakeup() - in-progress I/O accounting. 212 * 213 */ 214static __inline void 215runningbufwakeup(struct buf *bp) 216{ 217 if (bp->b_runningbufspace) { 218 runningbufspace -= bp->b_runningbufspace; 219 bp->b_runningbufspace = 0; 220 if (runningbufreq && runningbufspace <= lorunningspace) { 221 runningbufreq = 0; 222 wakeup(&runningbufreq); 223 } 224 } 225} 226 227/* 228 * bufcountwakeup: 229 * 230 * Called when a buffer has been added to one of the free queues to 231 * account for the buffer and to wakeup anyone waiting for free buffers. 232 * This typically occurs when large amounts of metadata are being handled 233 * by the buffer cache ( else buffer space runs out first, usually ). 234 */ 235 236static __inline void 237bufcountwakeup(void) 238{ 239 ++numfreebuffers; 240 if (needsbuffer) { 241 needsbuffer &= ~VFS_BIO_NEED_ANY; 242 if (numfreebuffers >= hifreebuffers) 243 needsbuffer &= ~VFS_BIO_NEED_FREE; 244 wakeup(&needsbuffer); 245 } 246} 247 248/* 249 * waitrunningbufspace() 250 * 251 * runningbufspace is a measure of the amount of I/O currently 252 * running. This routine is used in async-write situations to 253 * prevent creating huge backups of pending writes to a device. 254 * Only asynchronous writes are governed by this function. 255 * 256 * Reads will adjust runningbufspace, but will not block based on it. 257 * The read load has a side effect of reducing the allowed write load. 258 * 259 * This does NOT turn an async write into a sync write. It waits 260 * for earlier writes to complete and generally returns before the 261 * caller's write has reached the device. 262 */ 263static __inline void 264waitrunningbufspace(void) 265{ 266 while (runningbufspace > hirunningspace) { 267 ++runningbufreq; 268 tsleep(&runningbufreq, PVM, "wdrain", 0); 269 } 270} 271 272 273/* 274 * vfs_buf_test_cache: 275 * 276 * Called when a buffer is extended. This function clears the B_CACHE 277 * bit if the newly extended portion of the buffer does not contain 278 * valid data. 279 */ 280static __inline__ 281void 282vfs_buf_test_cache(struct buf *bp, 283 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 284 vm_page_t m) 285{ 286 if (bp->b_flags & B_CACHE) { 287 int base = (foff + off) & PAGE_MASK; 288 if (vm_page_is_valid(m, base, size) == 0) 289 bp->b_flags &= ~B_CACHE; 290 } 291} 292 293static __inline__ 294void 295bd_wakeup(int dirtybuflevel) 296{ 297 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 298 bd_request = 1; 299 wakeup(&bd_request); 300 } 301} 302 303/* 304 * bd_speedup - speedup the buffer cache flushing code 305 */ 306 307static __inline__ 308void 309bd_speedup(void) 310{ 311 bd_wakeup(1); 312} 313 314/* 315 * Initialize buffer headers and related structures. 316 */ 317 318caddr_t 319bufhashinit(caddr_t vaddr) 320{ 321 /* first, make a null hash table */ 322 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 323 ; 324 bufhashtbl = (void *)vaddr; 325 vaddr = vaddr + sizeof(*bufhashtbl) * bufhashmask; 326 --bufhashmask; 327 return(vaddr); 328} 329 330void 331bufinit(void) 332{ 333 struct buf *bp; 334 int i; 335 336 TAILQ_INIT(&bswlist); 337 LIST_INIT(&invalhash); 338 mtx_init(&buftimelock, "buftime lock", MTX_DEF); 339 340 for (i = 0; i <= bufhashmask; i++) 341 LIST_INIT(&bufhashtbl[i]); 342 343 /* next, make a null set of free lists */ 344 for (i = 0; i < BUFFER_QUEUES; i++) 345 TAILQ_INIT(&bufqueues[i]); 346 347 /* finally, initialize each buffer header and stick on empty q */ 348 for (i = 0; i < nbuf; i++) { 349 bp = &buf[i]; 350 bzero(bp, sizeof *bp); 351 bp->b_flags = B_INVAL; /* we're just an empty header */ 352 bp->b_dev = NODEV; 353 bp->b_rcred = NOCRED; 354 bp->b_wcred = NOCRED; 355 bp->b_qindex = QUEUE_EMPTY; 356 bp->b_xflags = 0; 357 LIST_INIT(&bp->b_dep); 358 BUF_LOCKINIT(bp); 359 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 360 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 361 } 362 363 /* 364 * maxbufspace is the absolute maximum amount of buffer space we are 365 * allowed to reserve in KVM and in real terms. The absolute maximum 366 * is nominally used by buf_daemon. hibufspace is the nominal maximum 367 * used by most other processes. The differential is required to 368 * ensure that buf_daemon is able to run when other processes might 369 * be blocked waiting for buffer space. 370 * 371 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 372 * this may result in KVM fragmentation which is not handled optimally 373 * by the system. 374 */ 375 maxbufspace = nbuf * BKVASIZE; 376 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 377 lobufspace = hibufspace - MAXBSIZE; 378 379 lorunningspace = 512 * 1024; 380 hirunningspace = 1024 * 1024; 381 382/* 383 * Limit the amount of malloc memory since it is wired permanently into 384 * the kernel space. Even though this is accounted for in the buffer 385 * allocation, we don't want the malloced region to grow uncontrolled. 386 * The malloc scheme improves memory utilization significantly on average 387 * (small) directories. 388 */ 389 maxbufmallocspace = hibufspace / 20; 390 391/* 392 * Reduce the chance of a deadlock occuring by limiting the number 393 * of delayed-write dirty buffers we allow to stack up. 394 */ 395 hidirtybuffers = nbuf / 4 + 20; 396 numdirtybuffers = 0; 397/* 398 * To support extreme low-memory systems, make sure hidirtybuffers cannot 399 * eat up all available buffer space. This occurs when our minimum cannot 400 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 401 * BKVASIZE'd (8K) buffers. 402 */ 403 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 404 hidirtybuffers >>= 1; 405 } 406 lodirtybuffers = hidirtybuffers / 2; 407 408/* 409 * Try to keep the number of free buffers in the specified range, 410 * and give special processes (e.g. like buf_daemon) access to an 411 * emergency reserve. 412 */ 413 lofreebuffers = nbuf / 18 + 5; 414 hifreebuffers = 2 * lofreebuffers; 415 numfreebuffers = nbuf; 416 417/* 418 * Maximum number of async ops initiated per buf_daemon loop. This is 419 * somewhat of a hack at the moment, we really need to limit ourselves 420 * based on the number of bytes of I/O in-transit that were initiated 421 * from buf_daemon. 422 */ 423 424 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 425 bogus_page = vm_page_alloc(kernel_object, 426 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 427 VM_ALLOC_NORMAL); 428 cnt.v_wire_count++; 429 430} 431 432/* 433 * bfreekva() - free the kva allocation for a buffer. 434 * 435 * Must be called at splbio() or higher as this is the only locking for 436 * buffer_map. 437 * 438 * Since this call frees up buffer space, we call bufspacewakeup(). 439 */ 440static void 441bfreekva(struct buf * bp) 442{ 443 if (bp->b_kvasize) { 444 ++buffreekvacnt; 445 bufspace -= bp->b_kvasize; 446 vm_map_delete(buffer_map, 447 (vm_offset_t) bp->b_kvabase, 448 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 449 ); 450 bp->b_kvasize = 0; 451 bufspacewakeup(); 452 } 453} 454 455/* 456 * bremfree: 457 * 458 * Remove the buffer from the appropriate free list. 459 */ 460void 461bremfree(struct buf * bp) 462{ 463 int s = splbio(); 464 int old_qindex = bp->b_qindex; 465 466 if (bp->b_qindex != QUEUE_NONE) { 467 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 468 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 469 bp->b_qindex = QUEUE_NONE; 470 } else { 471 if (BUF_REFCNT(bp) <= 1) 472 panic("bremfree: removing a buffer not on a queue"); 473 } 474 475 /* 476 * Fixup numfreebuffers count. If the buffer is invalid or not 477 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 478 * the buffer was free and we must decrement numfreebuffers. 479 */ 480 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 481 switch(old_qindex) { 482 case QUEUE_DIRTY: 483 case QUEUE_CLEAN: 484 case QUEUE_EMPTY: 485 case QUEUE_EMPTYKVA: 486 --numfreebuffers; 487 break; 488 default: 489 break; 490 } 491 } 492 splx(s); 493} 494 495 496/* 497 * Get a buffer with the specified data. Look in the cache first. We 498 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 499 * is set, the buffer is valid and we do not have to do anything ( see 500 * getblk() ). 501 */ 502int 503bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 504 struct buf ** bpp) 505{ 506 struct buf *bp; 507 508 bp = getblk(vp, blkno, size, 0, 0); 509 *bpp = bp; 510 511 /* if not found in cache, do some I/O */ 512 if ((bp->b_flags & B_CACHE) == 0) { 513 if (curproc != PCPU_GET(idleproc)) 514 curproc->p_stats->p_ru.ru_inblock++; 515 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp)); 516 bp->b_iocmd = BIO_READ; 517 bp->b_flags &= ~B_INVAL; 518 bp->b_ioflags &= ~BIO_ERROR; 519 if (bp->b_rcred == NOCRED) { 520 if (cred != NOCRED) 521 crhold(cred); 522 bp->b_rcred = cred; 523 } 524 vfs_busy_pages(bp, 0); 525 VOP_STRATEGY(vp, bp); 526 return (bufwait(bp)); 527 } 528 return (0); 529} 530 531/* 532 * Operates like bread, but also starts asynchronous I/O on 533 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 534 * to initiating I/O . If B_CACHE is set, the buffer is valid 535 * and we do not have to do anything. 536 */ 537int 538breadn(struct vnode * vp, daddr_t blkno, int size, 539 daddr_t * rablkno, int *rabsize, 540 int cnt, struct ucred * cred, struct buf ** bpp) 541{ 542 struct buf *bp, *rabp; 543 int i; 544 int rv = 0, readwait = 0; 545 546 *bpp = bp = getblk(vp, blkno, size, 0, 0); 547 548 /* if not found in cache, do some I/O */ 549 if ((bp->b_flags & B_CACHE) == 0) { 550 if (curproc != PCPU_GET(idleproc)) 551 curproc->p_stats->p_ru.ru_inblock++; 552 bp->b_iocmd = BIO_READ; 553 bp->b_flags &= ~B_INVAL; 554 bp->b_ioflags &= ~BIO_ERROR; 555 if (bp->b_rcred == NOCRED) { 556 if (cred != NOCRED) 557 crhold(cred); 558 bp->b_rcred = cred; 559 } 560 vfs_busy_pages(bp, 0); 561 VOP_STRATEGY(vp, bp); 562 ++readwait; 563 } 564 565 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 566 if (inmem(vp, *rablkno)) 567 continue; 568 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 569 570 if ((rabp->b_flags & B_CACHE) == 0) { 571 if (curproc != PCPU_GET(idleproc)) 572 curproc->p_stats->p_ru.ru_inblock++; 573 rabp->b_flags |= B_ASYNC; 574 rabp->b_flags &= ~B_INVAL; 575 rabp->b_ioflags &= ~BIO_ERROR; 576 rabp->b_iocmd = BIO_READ; 577 if (rabp->b_rcred == NOCRED) { 578 if (cred != NOCRED) 579 crhold(cred); 580 rabp->b_rcred = cred; 581 } 582 vfs_busy_pages(rabp, 0); 583 BUF_KERNPROC(rabp); 584 VOP_STRATEGY(vp, rabp); 585 } else { 586 brelse(rabp); 587 } 588 } 589 590 if (readwait) { 591 rv = bufwait(bp); 592 } 593 return (rv); 594} 595 596/* 597 * Write, release buffer on completion. (Done by iodone 598 * if async). Do not bother writing anything if the buffer 599 * is invalid. 600 * 601 * Note that we set B_CACHE here, indicating that buffer is 602 * fully valid and thus cacheable. This is true even of NFS 603 * now so we set it generally. This could be set either here 604 * or in biodone() since the I/O is synchronous. We put it 605 * here. 606 */ 607 608int dobkgrdwrite = 1; 609SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, ""); 610 611int 612bwrite(struct buf * bp) 613{ 614 int oldflags, s; 615 struct buf *newbp; 616 617 if (bp->b_flags & B_INVAL) { 618 brelse(bp); 619 return (0); 620 } 621 622 oldflags = bp->b_flags; 623 624 if (BUF_REFCNT(bp) == 0) 625 panic("bwrite: buffer is not busy???"); 626 s = splbio(); 627 /* 628 * If a background write is already in progress, delay 629 * writing this block if it is asynchronous. Otherwise 630 * wait for the background write to complete. 631 */ 632 if (bp->b_xflags & BX_BKGRDINPROG) { 633 if (bp->b_flags & B_ASYNC) { 634 splx(s); 635 bdwrite(bp); 636 return (0); 637 } 638 bp->b_xflags |= BX_BKGRDWAIT; 639 tsleep(&bp->b_xflags, PRIBIO, "biord", 0); 640 if (bp->b_xflags & BX_BKGRDINPROG) 641 panic("bwrite: still writing"); 642 } 643 644 /* Mark the buffer clean */ 645 bundirty(bp); 646 647 /* 648 * If this buffer is marked for background writing and we 649 * do not have to wait for it, make a copy and write the 650 * copy so as to leave this buffer ready for further use. 651 * 652 * This optimization eats a lot of memory. If we have a page 653 * or buffer shortfall we can't do it. 654 */ 655 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 656 (bp->b_flags & B_ASYNC) && 657 !vm_page_count_severe() && 658 !buf_dirty_count_severe()) { 659 if (bp->b_iodone != NULL) { 660 printf("bp->b_iodone = %p\n", bp->b_iodone); 661 panic("bwrite: need chained iodone"); 662 } 663 664 /* get a new block */ 665 newbp = geteblk(bp->b_bufsize); 666 667 /* set it to be identical to the old block */ 668 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 669 bgetvp(bp->b_vp, newbp); 670 newbp->b_lblkno = bp->b_lblkno; 671 newbp->b_blkno = bp->b_blkno; 672 newbp->b_offset = bp->b_offset; 673 newbp->b_iodone = vfs_backgroundwritedone; 674 newbp->b_flags |= B_ASYNC; 675 newbp->b_flags &= ~B_INVAL; 676 677 /* move over the dependencies */ 678 if (LIST_FIRST(&bp->b_dep) != NULL) 679 buf_movedeps(bp, newbp); 680 681 /* 682 * Initiate write on the copy, release the original to 683 * the B_LOCKED queue so that it cannot go away until 684 * the background write completes. If not locked it could go 685 * away and then be reconstituted while it was being written. 686 * If the reconstituted buffer were written, we could end up 687 * with two background copies being written at the same time. 688 */ 689 bp->b_xflags |= BX_BKGRDINPROG; 690 bp->b_flags |= B_LOCKED; 691 bqrelse(bp); 692 bp = newbp; 693 } 694 695 bp->b_flags &= ~B_DONE; 696 bp->b_ioflags &= ~BIO_ERROR; 697 bp->b_flags |= B_WRITEINPROG | B_CACHE; 698 bp->b_iocmd = BIO_WRITE; 699 700 bp->b_vp->v_numoutput++; 701 vfs_busy_pages(bp, 1); 702 703 /* 704 * Normal bwrites pipeline writes 705 */ 706 bp->b_runningbufspace = bp->b_bufsize; 707 runningbufspace += bp->b_runningbufspace; 708 709 if (curproc != PCPU_GET(idleproc)) 710 curproc->p_stats->p_ru.ru_oublock++; 711 splx(s); 712 if (oldflags & B_ASYNC) 713 BUF_KERNPROC(bp); 714 BUF_STRATEGY(bp); 715 716 if ((oldflags & B_ASYNC) == 0) { 717 int rtval = bufwait(bp); 718 brelse(bp); 719 return (rtval); 720 } else { 721 /* 722 * don't allow the async write to saturate the I/O 723 * system. There is no chance of deadlock here because 724 * we are blocking on I/O that is already in-progress. 725 */ 726 waitrunningbufspace(); 727 } 728 729 return (0); 730} 731 732/* 733 * Complete a background write started from bwrite. 734 */ 735static void 736vfs_backgroundwritedone(bp) 737 struct buf *bp; 738{ 739 struct buf *origbp; 740 741 /* 742 * Find the original buffer that we are writing. 743 */ 744 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 745 panic("backgroundwritedone: lost buffer"); 746 /* 747 * Process dependencies then return any unfinished ones. 748 */ 749 if (LIST_FIRST(&bp->b_dep) != NULL) 750 buf_complete(bp); 751 if (LIST_FIRST(&bp->b_dep) != NULL) 752 buf_movedeps(bp, origbp); 753 /* 754 * Clear the BX_BKGRDINPROG flag in the original buffer 755 * and awaken it if it is waiting for the write to complete. 756 * If BX_BKGRDINPROG is not set in the original buffer it must 757 * have been released and re-instantiated - which is not legal. 758 */ 759 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), ("backgroundwritedone: lost buffer2")); 760 origbp->b_xflags &= ~BX_BKGRDINPROG; 761 if (origbp->b_xflags & BX_BKGRDWAIT) { 762 origbp->b_xflags &= ~BX_BKGRDWAIT; 763 wakeup(&origbp->b_xflags); 764 } 765 /* 766 * Clear the B_LOCKED flag and remove it from the locked 767 * queue if it currently resides there. 768 */ 769 origbp->b_flags &= ~B_LOCKED; 770 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 771 bremfree(origbp); 772 bqrelse(origbp); 773 } 774 /* 775 * This buffer is marked B_NOCACHE, so when it is released 776 * by biodone, it will be tossed. We mark it with BIO_READ 777 * to avoid biodone doing a second vwakeup. 778 */ 779 bp->b_flags |= B_NOCACHE; 780 bp->b_iocmd = BIO_READ; 781 bp->b_flags &= ~(B_CACHE | B_DONE); 782 bp->b_iodone = 0; 783 bufdone(bp); 784} 785 786/* 787 * Delayed write. (Buffer is marked dirty). Do not bother writing 788 * anything if the buffer is marked invalid. 789 * 790 * Note that since the buffer must be completely valid, we can safely 791 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 792 * biodone() in order to prevent getblk from writing the buffer 793 * out synchronously. 794 */ 795void 796bdwrite(struct buf * bp) 797{ 798 if (BUF_REFCNT(bp) == 0) 799 panic("bdwrite: buffer is not busy"); 800 801 if (bp->b_flags & B_INVAL) { 802 brelse(bp); 803 return; 804 } 805 bdirty(bp); 806 807 /* 808 * Set B_CACHE, indicating that the buffer is fully valid. This is 809 * true even of NFS now. 810 */ 811 bp->b_flags |= B_CACHE; 812 813 /* 814 * This bmap keeps the system from needing to do the bmap later, 815 * perhaps when the system is attempting to do a sync. Since it 816 * is likely that the indirect block -- or whatever other datastructure 817 * that the filesystem needs is still in memory now, it is a good 818 * thing to do this. Note also, that if the pageout daemon is 819 * requesting a sync -- there might not be enough memory to do 820 * the bmap then... So, this is important to do. 821 */ 822 if (bp->b_lblkno == bp->b_blkno) { 823 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 824 } 825 826 /* 827 * Set the *dirty* buffer range based upon the VM system dirty pages. 828 */ 829 vfs_setdirty(bp); 830 831 /* 832 * We need to do this here to satisfy the vnode_pager and the 833 * pageout daemon, so that it thinks that the pages have been 834 * "cleaned". Note that since the pages are in a delayed write 835 * buffer -- the VFS layer "will" see that the pages get written 836 * out on the next sync, or perhaps the cluster will be completed. 837 */ 838 vfs_clean_pages(bp); 839 bqrelse(bp); 840 841 /* 842 * Wakeup the buffer flushing daemon if we have a lot of dirty 843 * buffers (midpoint between our recovery point and our stall 844 * point). 845 */ 846 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 847 848 /* 849 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 850 * due to the softdep code. 851 */ 852} 853 854/* 855 * bdirty: 856 * 857 * Turn buffer into delayed write request. We must clear BIO_READ and 858 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 859 * itself to properly update it in the dirty/clean lists. We mark it 860 * B_DONE to ensure that any asynchronization of the buffer properly 861 * clears B_DONE ( else a panic will occur later ). 862 * 863 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 864 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 865 * should only be called if the buffer is known-good. 866 * 867 * Since the buffer is not on a queue, we do not update the numfreebuffers 868 * count. 869 * 870 * Must be called at splbio(). 871 * The buffer must be on QUEUE_NONE. 872 */ 873void 874bdirty(bp) 875 struct buf *bp; 876{ 877 KASSERT(bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 878 bp->b_flags &= ~(B_RELBUF); 879 bp->b_iocmd = BIO_WRITE; 880 881 if ((bp->b_flags & B_DELWRI) == 0) { 882 bp->b_flags |= B_DONE | B_DELWRI; 883 reassignbuf(bp, bp->b_vp); 884 ++numdirtybuffers; 885 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 886 } 887} 888 889/* 890 * bundirty: 891 * 892 * Clear B_DELWRI for buffer. 893 * 894 * Since the buffer is not on a queue, we do not update the numfreebuffers 895 * count. 896 * 897 * Must be called at splbio(). 898 * The buffer must be on QUEUE_NONE. 899 */ 900 901void 902bundirty(bp) 903 struct buf *bp; 904{ 905 KASSERT(bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 906 907 if (bp->b_flags & B_DELWRI) { 908 bp->b_flags &= ~B_DELWRI; 909 reassignbuf(bp, bp->b_vp); 910 --numdirtybuffers; 911 numdirtywakeup(lodirtybuffers); 912 } 913 /* 914 * Since it is now being written, we can clear its deferred write flag. 915 */ 916 bp->b_flags &= ~B_DEFERRED; 917} 918 919/* 920 * bawrite: 921 * 922 * Asynchronous write. Start output on a buffer, but do not wait for 923 * it to complete. The buffer is released when the output completes. 924 * 925 * bwrite() ( or the VOP routine anyway ) is responsible for handling 926 * B_INVAL buffers. Not us. 927 */ 928void 929bawrite(struct buf * bp) 930{ 931 bp->b_flags |= B_ASYNC; 932 (void) BUF_WRITE(bp); 933} 934 935/* 936 * bowrite: 937 * 938 * Ordered write. Start output on a buffer, and flag it so that the 939 * device will write it in the order it was queued. The buffer is 940 * released when the output completes. bwrite() ( or the VOP routine 941 * anyway ) is responsible for handling B_INVAL buffers. 942 */ 943int 944bowrite(struct buf * bp) 945{ 946 bp->b_ioflags |= BIO_ORDERED; 947 bp->b_flags |= B_ASYNC; 948 return (BUF_WRITE(bp)); 949} 950 951/* 952 * bwillwrite: 953 * 954 * Called prior to the locking of any vnodes when we are expecting to 955 * write. We do not want to starve the buffer cache with too many 956 * dirty buffers so we block here. By blocking prior to the locking 957 * of any vnodes we attempt to avoid the situation where a locked vnode 958 * prevents the various system daemons from flushing related buffers. 959 */ 960 961void 962bwillwrite(void) 963{ 964 if (numdirtybuffers >= hidirtybuffers) { 965 int s; 966 967 s = splbio(); 968 while (numdirtybuffers >= hidirtybuffers) { 969 bd_wakeup(1); 970 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 971 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0); 972 } 973 splx(s); 974 } 975} 976 977/* 978 * Return true if we have too many dirty buffers. 979 */ 980int 981buf_dirty_count_severe(void) 982{ 983 return(numdirtybuffers >= hidirtybuffers); 984} 985 986/* 987 * brelse: 988 * 989 * Release a busy buffer and, if requested, free its resources. The 990 * buffer will be stashed in the appropriate bufqueue[] allowing it 991 * to be accessed later as a cache entity or reused for other purposes. 992 */ 993void 994brelse(struct buf * bp) 995{ 996 int s; 997 998 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 999 1000 s = splbio(); 1001 1002 if (bp->b_flags & B_LOCKED) 1003 bp->b_ioflags &= ~BIO_ERROR; 1004 1005 if (bp->b_iocmd == BIO_WRITE && 1006 (bp->b_ioflags & BIO_ERROR) && 1007 !(bp->b_flags & B_INVAL)) { 1008 /* 1009 * Failed write, redirty. Must clear BIO_ERROR to prevent 1010 * pages from being scrapped. If B_INVAL is set then 1011 * this case is not run and the next case is run to 1012 * destroy the buffer. B_INVAL can occur if the buffer 1013 * is outside the range supported by the underlying device. 1014 */ 1015 bp->b_ioflags &= ~BIO_ERROR; 1016 bdirty(bp); 1017 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1018 (bp->b_ioflags & BIO_ERROR) || 1019 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1020 /* 1021 * Either a failed I/O or we were asked to free or not 1022 * cache the buffer. 1023 */ 1024 bp->b_flags |= B_INVAL; 1025 if (LIST_FIRST(&bp->b_dep) != NULL) 1026 buf_deallocate(bp); 1027 if (bp->b_flags & B_DELWRI) { 1028 --numdirtybuffers; 1029 numdirtywakeup(lodirtybuffers); 1030 } 1031 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1032 if ((bp->b_flags & B_VMIO) == 0) { 1033 if (bp->b_bufsize) 1034 allocbuf(bp, 0); 1035 if (bp->b_vp) 1036 brelvp(bp); 1037 } 1038 } 1039 1040 /* 1041 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1042 * is called with B_DELWRI set, the underlying pages may wind up 1043 * getting freed causing a previous write (bdwrite()) to get 'lost' 1044 * because pages associated with a B_DELWRI bp are marked clean. 1045 * 1046 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1047 * if B_DELWRI is set. 1048 * 1049 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1050 * on pages to return pages to the VM page queues. 1051 */ 1052 if (bp->b_flags & B_DELWRI) 1053 bp->b_flags &= ~B_RELBUF; 1054 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1055 bp->b_flags |= B_RELBUF; 1056 1057 /* 1058 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1059 * constituted, not even NFS buffers now. Two flags effect this. If 1060 * B_INVAL, the struct buf is invalidated but the VM object is kept 1061 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1062 * 1063 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1064 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1065 * buffer is also B_INVAL because it hits the re-dirtying code above. 1066 * 1067 * Normally we can do this whether a buffer is B_DELWRI or not. If 1068 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1069 * the commit state and we cannot afford to lose the buffer. If the 1070 * buffer has a background write in progress, we need to keep it 1071 * around to prevent it from being reconstituted and starting a second 1072 * background write. 1073 */ 1074 if ((bp->b_flags & B_VMIO) 1075 && !(bp->b_vp->v_tag == VT_NFS && 1076 !vn_isdisk(bp->b_vp, NULL) && 1077 (bp->b_flags & B_DELWRI)) 1078 ) { 1079 1080 int i, j, resid; 1081 vm_page_t m; 1082 off_t foff; 1083 vm_pindex_t poff; 1084 vm_object_t obj; 1085 struct vnode *vp; 1086 1087 vp = bp->b_vp; 1088 1089 /* 1090 * Get the base offset and length of the buffer. Note that 1091 * for block sizes that are less then PAGE_SIZE, the b_data 1092 * base of the buffer does not represent exactly b_offset and 1093 * neither b_offset nor b_size are necessarily page aligned. 1094 * Instead, the starting position of b_offset is: 1095 * 1096 * b_data + (b_offset & PAGE_MASK) 1097 * 1098 * block sizes less then DEV_BSIZE (usually 512) are not 1099 * supported due to the page granularity bits (m->valid, 1100 * m->dirty, etc...). 1101 * 1102 * See man buf(9) for more information 1103 */ 1104 resid = bp->b_bufsize; 1105 foff = bp->b_offset; 1106 1107 for (i = 0; i < bp->b_npages; i++) { 1108 int had_bogus = 0; 1109 1110 m = bp->b_pages[i]; 1111 vm_page_flag_clear(m, PG_ZERO); 1112 1113 /* 1114 * If we hit a bogus page, fixup *all* the bogus pages 1115 * now. 1116 */ 1117 if (m == bogus_page) { 1118 VOP_GETVOBJECT(vp, &obj); 1119 poff = OFF_TO_IDX(bp->b_offset); 1120 had_bogus = 1; 1121 1122 for (j = i; j < bp->b_npages; j++) { 1123 vm_page_t mtmp; 1124 mtmp = bp->b_pages[j]; 1125 if (mtmp == bogus_page) { 1126 mtmp = vm_page_lookup(obj, poff + j); 1127 if (!mtmp) { 1128 panic("brelse: page missing\n"); 1129 } 1130 bp->b_pages[j] = mtmp; 1131 } 1132 } 1133 1134 if ((bp->b_flags & B_INVAL) == 0) { 1135 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1136 } 1137 m = bp->b_pages[i]; 1138 } 1139 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1140 int poffset = foff & PAGE_MASK; 1141 int presid = resid > (PAGE_SIZE - poffset) ? 1142 (PAGE_SIZE - poffset) : resid; 1143 1144 KASSERT(presid >= 0, ("brelse: extra page")); 1145 vm_page_set_invalid(m, poffset, presid); 1146 if (had_bogus) 1147 printf("avoided corruption bug in bogus_page/brelse code\n"); 1148 } 1149 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1150 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1151 } 1152 1153 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1154 vfs_vmio_release(bp); 1155 1156 } else if (bp->b_flags & B_VMIO) { 1157 1158 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1159 vfs_vmio_release(bp); 1160 1161 } 1162 1163 if (bp->b_qindex != QUEUE_NONE) 1164 panic("brelse: free buffer onto another queue???"); 1165 if (BUF_REFCNT(bp) > 1) { 1166 /* do not release to free list */ 1167 BUF_UNLOCK(bp); 1168 splx(s); 1169 return; 1170 } 1171 1172 /* enqueue */ 1173 1174 /* buffers with no memory */ 1175 if (bp->b_bufsize == 0) { 1176 bp->b_flags |= B_INVAL; 1177 bp->b_xflags &= ~BX_BKGRDWRITE; 1178 if (bp->b_xflags & BX_BKGRDINPROG) 1179 panic("losing buffer 1"); 1180 if (bp->b_kvasize) { 1181 bp->b_qindex = QUEUE_EMPTYKVA; 1182 } else { 1183 bp->b_qindex = QUEUE_EMPTY; 1184 } 1185 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1186 LIST_REMOVE(bp, b_hash); 1187 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1188 bp->b_dev = NODEV; 1189 /* buffers with junk contents */ 1190 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || (bp->b_ioflags & BIO_ERROR)) { 1191 bp->b_flags |= B_INVAL; 1192 bp->b_xflags &= ~BX_BKGRDWRITE; 1193 if (bp->b_xflags & BX_BKGRDINPROG) 1194 panic("losing buffer 2"); 1195 bp->b_qindex = QUEUE_CLEAN; 1196 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1197 LIST_REMOVE(bp, b_hash); 1198 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1199 bp->b_dev = NODEV; 1200 1201 /* buffers that are locked */ 1202 } else if (bp->b_flags & B_LOCKED) { 1203 bp->b_qindex = QUEUE_LOCKED; 1204 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1205 1206 /* remaining buffers */ 1207 } else { 1208 switch(bp->b_flags & (B_DELWRI|B_AGE)) { 1209 case B_DELWRI | B_AGE: 1210 bp->b_qindex = QUEUE_DIRTY; 1211 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1212 break; 1213 case B_DELWRI: 1214 bp->b_qindex = QUEUE_DIRTY; 1215 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1216 break; 1217 case B_AGE: 1218 bp->b_qindex = QUEUE_CLEAN; 1219 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1220 break; 1221 default: 1222 bp->b_qindex = QUEUE_CLEAN; 1223 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1224 break; 1225 } 1226 } 1227 1228 /* 1229 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1230 * on the correct queue. 1231 */ 1232 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) { 1233 bp->b_flags &= ~B_DELWRI; 1234 --numdirtybuffers; 1235 numdirtywakeup(lodirtybuffers); 1236 } 1237 1238 /* 1239 * Fixup numfreebuffers count. The bp is on an appropriate queue 1240 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1241 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1242 * if B_INVAL is set ). 1243 */ 1244 1245 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1246 bufcountwakeup(); 1247 1248 /* 1249 * Something we can maybe free or reuse 1250 */ 1251 if (bp->b_bufsize || bp->b_kvasize) 1252 bufspacewakeup(); 1253 1254 /* unlock */ 1255 BUF_UNLOCK(bp); 1256 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1257 bp->b_ioflags &= ~BIO_ORDERED; 1258 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1259 panic("brelse: not dirty"); 1260 splx(s); 1261} 1262 1263/* 1264 * Release a buffer back to the appropriate queue but do not try to free 1265 * it. The buffer is expected to be used again soon. 1266 * 1267 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1268 * biodone() to requeue an async I/O on completion. It is also used when 1269 * known good buffers need to be requeued but we think we may need the data 1270 * again soon. 1271 */ 1272void 1273bqrelse(struct buf * bp) 1274{ 1275 int s; 1276 1277 s = splbio(); 1278 1279 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1280 1281 if (bp->b_qindex != QUEUE_NONE) 1282 panic("bqrelse: free buffer onto another queue???"); 1283 if (BUF_REFCNT(bp) > 1) { 1284 /* do not release to free list */ 1285 BUF_UNLOCK(bp); 1286 splx(s); 1287 return; 1288 } 1289 if (bp->b_flags & B_LOCKED) { 1290 bp->b_ioflags &= ~BIO_ERROR; 1291 bp->b_qindex = QUEUE_LOCKED; 1292 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1293 /* buffers with stale but valid contents */ 1294 } else if (bp->b_flags & B_DELWRI) { 1295 bp->b_qindex = QUEUE_DIRTY; 1296 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1297 } else if (vm_page_count_severe()) { 1298 /* 1299 * We are too low on memory, we have to try to free the 1300 * buffer (most importantly: the wired pages making up its 1301 * backing store) *now*. 1302 */ 1303 splx(s); 1304 brelse(bp); 1305 return; 1306 } else { 1307 bp->b_qindex = QUEUE_CLEAN; 1308 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1309 } 1310 1311 if ((bp->b_flags & B_LOCKED) == 0 && 1312 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1313 bufcountwakeup(); 1314 } 1315 1316 /* 1317 * Something we can maybe free or reuse. 1318 */ 1319 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1320 bufspacewakeup(); 1321 1322 /* unlock */ 1323 BUF_UNLOCK(bp); 1324 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1325 bp->b_ioflags &= ~BIO_ORDERED; 1326 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1327 panic("bqrelse: not dirty"); 1328 splx(s); 1329} 1330 1331static void 1332vfs_vmio_release(bp) 1333 struct buf *bp; 1334{ 1335 int i, s; 1336 vm_page_t m; 1337 1338 s = splvm(); 1339 for (i = 0; i < bp->b_npages; i++) { 1340 m = bp->b_pages[i]; 1341 bp->b_pages[i] = NULL; 1342 /* 1343 * In order to keep page LRU ordering consistent, put 1344 * everything on the inactive queue. 1345 */ 1346 vm_page_unwire(m, 0); 1347 /* 1348 * We don't mess with busy pages, it is 1349 * the responsibility of the process that 1350 * busied the pages to deal with them. 1351 */ 1352 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1353 continue; 1354 1355 if (m->wire_count == 0) { 1356 vm_page_flag_clear(m, PG_ZERO); 1357 /* 1358 * Might as well free the page if we can and it has 1359 * no valid data. 1360 */ 1361 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { 1362 vm_page_busy(m); 1363 vm_page_protect(m, VM_PROT_NONE); 1364 vm_page_free(m); 1365 } else if (vm_page_count_severe()) { 1366 vm_page_try_to_cache(m); 1367 } 1368 } 1369 } 1370 splx(s); 1371 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1372 if (bp->b_bufsize) { 1373 bufspacewakeup(); 1374 bp->b_bufsize = 0; 1375 } 1376 bp->b_npages = 0; 1377 bp->b_flags &= ~B_VMIO; 1378 if (bp->b_vp) 1379 brelvp(bp); 1380} 1381 1382/* 1383 * Check to see if a block is currently memory resident. 1384 */ 1385struct buf * 1386gbincore(struct vnode * vp, daddr_t blkno) 1387{ 1388 struct buf *bp; 1389 struct bufhashhdr *bh; 1390 1391 bh = bufhash(vp, blkno); 1392 1393 /* Search hash chain */ 1394 LIST_FOREACH(bp, bh, b_hash) { 1395 /* hit */ 1396 if (bp->b_vp == vp && bp->b_lblkno == blkno && 1397 (bp->b_flags & B_INVAL) == 0) { 1398 break; 1399 } 1400 } 1401 return (bp); 1402} 1403 1404/* 1405 * vfs_bio_awrite: 1406 * 1407 * Implement clustered async writes for clearing out B_DELWRI buffers. 1408 * This is much better then the old way of writing only one buffer at 1409 * a time. Note that we may not be presented with the buffers in the 1410 * correct order, so we search for the cluster in both directions. 1411 */ 1412int 1413vfs_bio_awrite(struct buf * bp) 1414{ 1415 int i; 1416 int j; 1417 daddr_t lblkno = bp->b_lblkno; 1418 struct vnode *vp = bp->b_vp; 1419 int s; 1420 int ncl; 1421 struct buf *bpa; 1422 int nwritten; 1423 int size; 1424 int maxcl; 1425 1426 s = splbio(); 1427 /* 1428 * right now we support clustered writing only to regular files. If 1429 * we find a clusterable block we could be in the middle of a cluster 1430 * rather then at the beginning. 1431 */ 1432 if ((vp->v_type == VREG) && 1433 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1434 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1435 1436 size = vp->v_mount->mnt_stat.f_iosize; 1437 maxcl = MAXPHYS / size; 1438 1439 for (i = 1; i < maxcl; i++) { 1440 if ((bpa = gbincore(vp, lblkno + i)) && 1441 BUF_REFCNT(bpa) == 0 && 1442 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1443 (B_DELWRI | B_CLUSTEROK)) && 1444 (bpa->b_bufsize == size)) { 1445 if ((bpa->b_blkno == bpa->b_lblkno) || 1446 (bpa->b_blkno != 1447 bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 1448 break; 1449 } else { 1450 break; 1451 } 1452 } 1453 for (j = 1; i + j <= maxcl && j <= lblkno; j++) { 1454 if ((bpa = gbincore(vp, lblkno - j)) && 1455 BUF_REFCNT(bpa) == 0 && 1456 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1457 (B_DELWRI | B_CLUSTEROK)) && 1458 (bpa->b_bufsize == size)) { 1459 if ((bpa->b_blkno == bpa->b_lblkno) || 1460 (bpa->b_blkno != 1461 bp->b_blkno - ((j * size) >> DEV_BSHIFT))) 1462 break; 1463 } else { 1464 break; 1465 } 1466 } 1467 --j; 1468 ncl = i + j; 1469 /* 1470 * this is a possible cluster write 1471 */ 1472 if (ncl != 1) { 1473 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1474 splx(s); 1475 return nwritten; 1476 } 1477 } 1478 1479 BUF_LOCK(bp, LK_EXCLUSIVE); 1480 bremfree(bp); 1481 bp->b_flags |= B_ASYNC; 1482 1483 splx(s); 1484 /* 1485 * default (old) behavior, writing out only one block 1486 * 1487 * XXX returns b_bufsize instead of b_bcount for nwritten? 1488 */ 1489 nwritten = bp->b_bufsize; 1490 (void) BUF_WRITE(bp); 1491 1492 return nwritten; 1493} 1494 1495/* 1496 * getnewbuf: 1497 * 1498 * Find and initialize a new buffer header, freeing up existing buffers 1499 * in the bufqueues as necessary. The new buffer is returned locked. 1500 * 1501 * Important: B_INVAL is not set. If the caller wishes to throw the 1502 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1503 * 1504 * We block if: 1505 * We have insufficient buffer headers 1506 * We have insufficient buffer space 1507 * buffer_map is too fragmented ( space reservation fails ) 1508 * If we have to flush dirty buffers ( but we try to avoid this ) 1509 * 1510 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1511 * Instead we ask the buf daemon to do it for us. We attempt to 1512 * avoid piecemeal wakeups of the pageout daemon. 1513 */ 1514 1515static struct buf * 1516getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1517{ 1518 struct buf *bp; 1519 struct buf *nbp; 1520 int defrag = 0; 1521 int nqindex; 1522 static int flushingbufs; 1523 1524 /* 1525 * We can't afford to block since we might be holding a vnode lock, 1526 * which may prevent system daemons from running. We deal with 1527 * low-memory situations by proactively returning memory and running 1528 * async I/O rather then sync I/O. 1529 */ 1530 1531 ++getnewbufcalls; 1532 --getnewbufrestarts; 1533restart: 1534 ++getnewbufrestarts; 1535 1536 /* 1537 * Setup for scan. If we do not have enough free buffers, 1538 * we setup a degenerate case that immediately fails. Note 1539 * that if we are specially marked process, we are allowed to 1540 * dip into our reserves. 1541 * 1542 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1543 * 1544 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1545 * However, there are a number of cases (defragging, reusing, ...) 1546 * where we cannot backup. 1547 */ 1548 nqindex = QUEUE_EMPTYKVA; 1549 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1550 1551 if (nbp == NULL) { 1552 /* 1553 * If no EMPTYKVA buffers and we are either 1554 * defragging or reusing, locate a CLEAN buffer 1555 * to free or reuse. If bufspace useage is low 1556 * skip this step so we can allocate a new buffer. 1557 */ 1558 if (defrag || bufspace >= lobufspace) { 1559 nqindex = QUEUE_CLEAN; 1560 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1561 } 1562 1563 /* 1564 * If we could not find or were not allowed to reuse a 1565 * CLEAN buffer, check to see if it is ok to use an EMPTY 1566 * buffer. We can only use an EMPTY buffer if allocating 1567 * its KVA would not otherwise run us out of buffer space. 1568 */ 1569 if (nbp == NULL && defrag == 0 && 1570 bufspace + maxsize < hibufspace) { 1571 nqindex = QUEUE_EMPTY; 1572 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1573 } 1574 } 1575 1576 /* 1577 * Run scan, possibly freeing data and/or kva mappings on the fly 1578 * depending. 1579 */ 1580 1581 while ((bp = nbp) != NULL) { 1582 int qindex = nqindex; 1583 1584 /* 1585 * Calculate next bp ( we can only use it if we do not block 1586 * or do other fancy things ). 1587 */ 1588 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1589 switch(qindex) { 1590 case QUEUE_EMPTY: 1591 nqindex = QUEUE_EMPTYKVA; 1592 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1593 break; 1594 /* fall through */ 1595 case QUEUE_EMPTYKVA: 1596 nqindex = QUEUE_CLEAN; 1597 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1598 break; 1599 /* fall through */ 1600 case QUEUE_CLEAN: 1601 /* 1602 * nbp is NULL. 1603 */ 1604 break; 1605 } 1606 } 1607 1608 /* 1609 * Sanity Checks 1610 */ 1611 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1612 1613 /* 1614 * Note: we no longer distinguish between VMIO and non-VMIO 1615 * buffers. 1616 */ 1617 1618 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1619 1620 /* 1621 * If we are defragging then we need a buffer with 1622 * b_kvasize != 0. XXX this situation should no longer 1623 * occur, if defrag is non-zero the buffer's b_kvasize 1624 * should also be non-zero at this point. XXX 1625 */ 1626 if (defrag && bp->b_kvasize == 0) { 1627 printf("Warning: defrag empty buffer %p\n", bp); 1628 continue; 1629 } 1630 1631 /* 1632 * Start freeing the bp. This is somewhat involved. nbp 1633 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1634 */ 1635 1636 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1637 panic("getnewbuf: locked buf"); 1638 bremfree(bp); 1639 1640 if (qindex == QUEUE_CLEAN) { 1641 if (bp->b_flags & B_VMIO) { 1642 bp->b_flags &= ~B_ASYNC; 1643 vfs_vmio_release(bp); 1644 } 1645 if (bp->b_vp) 1646 brelvp(bp); 1647 } 1648 1649 /* 1650 * NOTE: nbp is now entirely invalid. We can only restart 1651 * the scan from this point on. 1652 * 1653 * Get the rest of the buffer freed up. b_kva* is still 1654 * valid after this operation. 1655 */ 1656 1657 if (bp->b_rcred != NOCRED) { 1658 crfree(bp->b_rcred); 1659 bp->b_rcred = NOCRED; 1660 } 1661 if (bp->b_wcred != NOCRED) { 1662 crfree(bp->b_wcred); 1663 bp->b_wcred = NOCRED; 1664 } 1665 if (LIST_FIRST(&bp->b_dep) != NULL) 1666 buf_deallocate(bp); 1667 if (bp->b_xflags & BX_BKGRDINPROG) 1668 panic("losing buffer 3"); 1669 LIST_REMOVE(bp, b_hash); 1670 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1671 1672 if (bp->b_bufsize) 1673 allocbuf(bp, 0); 1674 1675 bp->b_flags = 0; 1676 bp->b_ioflags = 0; 1677 bp->b_xflags = 0; 1678 bp->b_dev = NODEV; 1679 bp->b_vp = NULL; 1680 bp->b_blkno = bp->b_lblkno = 0; 1681 bp->b_offset = NOOFFSET; 1682 bp->b_iodone = 0; 1683 bp->b_error = 0; 1684 bp->b_resid = 0; 1685 bp->b_bcount = 0; 1686 bp->b_npages = 0; 1687 bp->b_dirtyoff = bp->b_dirtyend = 0; 1688 1689 LIST_INIT(&bp->b_dep); 1690 1691 /* 1692 * If we are defragging then free the buffer. 1693 */ 1694 if (defrag) { 1695 bp->b_flags |= B_INVAL; 1696 bfreekva(bp); 1697 brelse(bp); 1698 defrag = 0; 1699 goto restart; 1700 } 1701 1702 /* 1703 * If we are overcomitted then recover the buffer and its 1704 * KVM space. This occurs in rare situations when multiple 1705 * processes are blocked in getnewbuf() or allocbuf(). 1706 */ 1707 if (bufspace >= hibufspace) 1708 flushingbufs = 1; 1709 if (flushingbufs && bp->b_kvasize != 0) { 1710 bp->b_flags |= B_INVAL; 1711 bfreekva(bp); 1712 brelse(bp); 1713 goto restart; 1714 } 1715 if (bufspace < lobufspace) 1716 flushingbufs = 0; 1717 break; 1718 } 1719 1720 /* 1721 * If we exhausted our list, sleep as appropriate. We may have to 1722 * wakeup various daemons and write out some dirty buffers. 1723 * 1724 * Generally we are sleeping due to insufficient buffer space. 1725 */ 1726 1727 if (bp == NULL) { 1728 int flags; 1729 char *waitmsg; 1730 1731 if (defrag) { 1732 flags = VFS_BIO_NEED_BUFSPACE; 1733 waitmsg = "nbufkv"; 1734 } else if (bufspace >= hibufspace) { 1735 waitmsg = "nbufbs"; 1736 flags = VFS_BIO_NEED_BUFSPACE; 1737 } else { 1738 waitmsg = "newbuf"; 1739 flags = VFS_BIO_NEED_ANY; 1740 } 1741 1742 bd_speedup(); /* heeeelp */ 1743 1744 needsbuffer |= flags; 1745 while (needsbuffer & flags) { 1746 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, 1747 waitmsg, slptimeo)) 1748 return (NULL); 1749 } 1750 } else { 1751 /* 1752 * We finally have a valid bp. We aren't quite out of the 1753 * woods, we still have to reserve kva space. In order 1754 * to keep fragmentation sane we only allocate kva in 1755 * BKVASIZE chunks. 1756 */ 1757 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1758 1759 if (maxsize != bp->b_kvasize) { 1760 vm_offset_t addr = 0; 1761 1762 bfreekva(bp); 1763 1764 if (vm_map_findspace(buffer_map, 1765 vm_map_min(buffer_map), maxsize, &addr)) { 1766 /* 1767 * Uh oh. Buffer map is to fragmented. We 1768 * must defragment the map. 1769 */ 1770 ++bufdefragcnt; 1771 defrag = 1; 1772 bp->b_flags |= B_INVAL; 1773 brelse(bp); 1774 goto restart; 1775 } 1776 if (addr) { 1777 vm_map_insert(buffer_map, NULL, 0, 1778 addr, addr + maxsize, 1779 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1780 1781 bp->b_kvabase = (caddr_t) addr; 1782 bp->b_kvasize = maxsize; 1783 bufspace += bp->b_kvasize; 1784 ++bufreusecnt; 1785 } 1786 } 1787 bp->b_data = bp->b_kvabase; 1788 } 1789 return(bp); 1790} 1791 1792/* 1793 * buf_daemon: 1794 * 1795 * buffer flushing daemon. Buffers are normally flushed by the 1796 * update daemon but if it cannot keep up this process starts to 1797 * take the load in an attempt to prevent getnewbuf() from blocking. 1798 */ 1799 1800static struct proc *bufdaemonproc; 1801 1802static struct kproc_desc buf_kp = { 1803 "bufdaemon", 1804 buf_daemon, 1805 &bufdaemonproc 1806}; 1807SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1808 1809static void 1810buf_daemon() 1811{ 1812 int s; 1813 1814 mtx_lock(&Giant); 1815 1816 /* 1817 * This process needs to be suspended prior to shutdown sync. 1818 */ 1819 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 1820 SHUTDOWN_PRI_LAST); 1821 1822 /* 1823 * This process is allowed to take the buffer cache to the limit 1824 */ 1825 curproc->p_flag |= P_BUFEXHAUST; 1826 s = splbio(); 1827 1828 for (;;) { 1829 kthread_suspend_check(bufdaemonproc); 1830 1831 bd_request = 0; 1832 1833 /* 1834 * Do the flush. Limit the amount of in-transit I/O we 1835 * allow to build up, otherwise we would completely saturate 1836 * the I/O system. Wakeup any waiting processes before we 1837 * normally would so they can run in parallel with our drain. 1838 */ 1839 while (numdirtybuffers > lodirtybuffers) { 1840 if (flushbufqueues() == 0) 1841 break; 1842 waitrunningbufspace(); 1843 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 1844 } 1845 1846 /* 1847 * Only clear bd_request if we have reached our low water 1848 * mark. The buf_daemon normally waits 5 seconds and 1849 * then incrementally flushes any dirty buffers that have 1850 * built up, within reason. 1851 * 1852 * If we were unable to hit our low water mark and couldn't 1853 * find any flushable buffers, we sleep half a second. 1854 * Otherwise we loop immediately. 1855 */ 1856 if (numdirtybuffers <= lodirtybuffers) { 1857 /* 1858 * We reached our low water mark, reset the 1859 * request and sleep until we are needed again. 1860 * The sleep is just so the suspend code works. 1861 */ 1862 bd_request = 0; 1863 tsleep(&bd_request, PVM, "psleep", hz); 1864 } else { 1865 /* 1866 * We couldn't find any flushable dirty buffers but 1867 * still have too many dirty buffers, we 1868 * have to sleep and try again. (rare) 1869 */ 1870 tsleep(&bd_request, PVM, "qsleep", hz / 2); 1871 } 1872 } 1873} 1874 1875/* 1876 * flushbufqueues: 1877 * 1878 * Try to flush a buffer in the dirty queue. We must be careful to 1879 * free up B_INVAL buffers instead of write them, which NFS is 1880 * particularly sensitive to. 1881 */ 1882 1883static int 1884flushbufqueues(void) 1885{ 1886 struct buf *bp; 1887 int r = 0; 1888 1889 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1890 1891 while (bp) { 1892 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); 1893 if ((bp->b_flags & B_DELWRI) != 0 && 1894 (bp->b_xflags & BX_BKGRDINPROG) == 0) { 1895 if (bp->b_flags & B_INVAL) { 1896 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1897 panic("flushbufqueues: locked buf"); 1898 bremfree(bp); 1899 brelse(bp); 1900 ++r; 1901 break; 1902 } 1903 if (LIST_FIRST(&bp->b_dep) != NULL && 1904 (bp->b_flags & B_DEFERRED) == 0 && 1905 buf_countdeps(bp, 0)) { 1906 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], 1907 bp, b_freelist); 1908 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], 1909 bp, b_freelist); 1910 bp->b_flags |= B_DEFERRED; 1911 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1912 continue; 1913 } 1914 vfs_bio_awrite(bp); 1915 ++r; 1916 break; 1917 } 1918 bp = TAILQ_NEXT(bp, b_freelist); 1919 } 1920 return (r); 1921} 1922 1923/* 1924 * Check to see if a block is currently memory resident. 1925 */ 1926struct buf * 1927incore(struct vnode * vp, daddr_t blkno) 1928{ 1929 struct buf *bp; 1930 1931 int s = splbio(); 1932 bp = gbincore(vp, blkno); 1933 splx(s); 1934 return (bp); 1935} 1936 1937/* 1938 * Returns true if no I/O is needed to access the 1939 * associated VM object. This is like incore except 1940 * it also hunts around in the VM system for the data. 1941 */ 1942 1943int 1944inmem(struct vnode * vp, daddr_t blkno) 1945{ 1946 vm_object_t obj; 1947 vm_offset_t toff, tinc, size; 1948 vm_page_t m; 1949 vm_ooffset_t off; 1950 1951 if (incore(vp, blkno)) 1952 return 1; 1953 if (vp->v_mount == NULL) 1954 return 0; 1955 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_flag & VOBJBUF) == 0) 1956 return 0; 1957 1958 size = PAGE_SIZE; 1959 if (size > vp->v_mount->mnt_stat.f_iosize) 1960 size = vp->v_mount->mnt_stat.f_iosize; 1961 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 1962 1963 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1964 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 1965 if (!m) 1966 return 0; 1967 tinc = size; 1968 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 1969 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 1970 if (vm_page_is_valid(m, 1971 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 1972 return 0; 1973 } 1974 return 1; 1975} 1976 1977/* 1978 * vfs_setdirty: 1979 * 1980 * Sets the dirty range for a buffer based on the status of the dirty 1981 * bits in the pages comprising the buffer. 1982 * 1983 * The range is limited to the size of the buffer. 1984 * 1985 * This routine is primarily used by NFS, but is generalized for the 1986 * B_VMIO case. 1987 */ 1988static void 1989vfs_setdirty(struct buf *bp) 1990{ 1991 int i; 1992 vm_object_t object; 1993 1994 /* 1995 * Degenerate case - empty buffer 1996 */ 1997 1998 if (bp->b_bufsize == 0) 1999 return; 2000 2001 /* 2002 * We qualify the scan for modified pages on whether the 2003 * object has been flushed yet. The OBJ_WRITEABLE flag 2004 * is not cleared simply by protecting pages off. 2005 */ 2006 2007 if ((bp->b_flags & B_VMIO) == 0) 2008 return; 2009 2010 object = bp->b_pages[0]->object; 2011 2012 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2013 printf("Warning: object %p writeable but not mightbedirty\n", object); 2014 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2015 printf("Warning: object %p mightbedirty but not writeable\n", object); 2016 2017 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2018 vm_offset_t boffset; 2019 vm_offset_t eoffset; 2020 2021 /* 2022 * test the pages to see if they have been modified directly 2023 * by users through the VM system. 2024 */ 2025 for (i = 0; i < bp->b_npages; i++) { 2026 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2027 vm_page_test_dirty(bp->b_pages[i]); 2028 } 2029 2030 /* 2031 * Calculate the encompassing dirty range, boffset and eoffset, 2032 * (eoffset - boffset) bytes. 2033 */ 2034 2035 for (i = 0; i < bp->b_npages; i++) { 2036 if (bp->b_pages[i]->dirty) 2037 break; 2038 } 2039 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2040 2041 for (i = bp->b_npages - 1; i >= 0; --i) { 2042 if (bp->b_pages[i]->dirty) { 2043 break; 2044 } 2045 } 2046 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2047 2048 /* 2049 * Fit it to the buffer. 2050 */ 2051 2052 if (eoffset > bp->b_bcount) 2053 eoffset = bp->b_bcount; 2054 2055 /* 2056 * If we have a good dirty range, merge with the existing 2057 * dirty range. 2058 */ 2059 2060 if (boffset < eoffset) { 2061 if (bp->b_dirtyoff > boffset) 2062 bp->b_dirtyoff = boffset; 2063 if (bp->b_dirtyend < eoffset) 2064 bp->b_dirtyend = eoffset; 2065 } 2066 } 2067} 2068 2069/* 2070 * getblk: 2071 * 2072 * Get a block given a specified block and offset into a file/device. 2073 * The buffers B_DONE bit will be cleared on return, making it almost 2074 * ready for an I/O initiation. B_INVAL may or may not be set on 2075 * return. The caller should clear B_INVAL prior to initiating a 2076 * READ. 2077 * 2078 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2079 * an existing buffer. 2080 * 2081 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2082 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2083 * and then cleared based on the backing VM. If the previous buffer is 2084 * non-0-sized but invalid, B_CACHE will be cleared. 2085 * 2086 * If getblk() must create a new buffer, the new buffer is returned with 2087 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2088 * case it is returned with B_INVAL clear and B_CACHE set based on the 2089 * backing VM. 2090 * 2091 * getblk() also forces a VOP_BWRITE() for any B_DELWRI buffer whos 2092 * B_CACHE bit is clear. 2093 * 2094 * What this means, basically, is that the caller should use B_CACHE to 2095 * determine whether the buffer is fully valid or not and should clear 2096 * B_INVAL prior to issuing a read. If the caller intends to validate 2097 * the buffer by loading its data area with something, the caller needs 2098 * to clear B_INVAL. If the caller does this without issuing an I/O, 2099 * the caller should set B_CACHE ( as an optimization ), else the caller 2100 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2101 * a write attempt or if it was a successfull read. If the caller 2102 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2103 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2104 */ 2105struct buf * 2106getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2107{ 2108 struct buf *bp; 2109 int s; 2110 struct bufhashhdr *bh; 2111 2112 if (size > MAXBSIZE) 2113 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2114 2115 s = splbio(); 2116loop: 2117 /* 2118 * Block if we are low on buffers. Certain processes are allowed 2119 * to completely exhaust the buffer cache. 2120 * 2121 * If this check ever becomes a bottleneck it may be better to 2122 * move it into the else, when gbincore() fails. At the moment 2123 * it isn't a problem. 2124 * 2125 * XXX remove if 0 sections (clean this up after its proven) 2126 */ 2127 if (numfreebuffers == 0) { 2128 if (curproc == PCPU_GET(idleproc)) 2129 return NULL; 2130 needsbuffer |= VFS_BIO_NEED_ANY; 2131 } 2132 2133 if ((bp = gbincore(vp, blkno))) { 2134 /* 2135 * Buffer is in-core. If the buffer is not busy, it must 2136 * be on a queue. 2137 */ 2138 2139 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2140 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, 2141 "getblk", slpflag, slptimeo) == ENOLCK) 2142 goto loop; 2143 splx(s); 2144 return (struct buf *) NULL; 2145 } 2146 2147 /* 2148 * The buffer is locked. B_CACHE is cleared if the buffer is 2149 * invalid. Ohterwise, for a non-VMIO buffer, B_CACHE is set 2150 * and for a VMIO buffer B_CACHE is adjusted according to the 2151 * backing VM cache. 2152 */ 2153 if (bp->b_flags & B_INVAL) 2154 bp->b_flags &= ~B_CACHE; 2155 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2156 bp->b_flags |= B_CACHE; 2157 bremfree(bp); 2158 2159 /* 2160 * check for size inconsistancies for non-VMIO case. 2161 */ 2162 2163 if (bp->b_bcount != size) { 2164 if ((bp->b_flags & B_VMIO) == 0 || 2165 (size > bp->b_kvasize)) { 2166 if (bp->b_flags & B_DELWRI) { 2167 bp->b_flags |= B_NOCACHE; 2168 BUF_WRITE(bp); 2169 } else { 2170 if ((bp->b_flags & B_VMIO) && 2171 (LIST_FIRST(&bp->b_dep) == NULL)) { 2172 bp->b_flags |= B_RELBUF; 2173 brelse(bp); 2174 } else { 2175 bp->b_flags |= B_NOCACHE; 2176 BUF_WRITE(bp); 2177 } 2178 } 2179 goto loop; 2180 } 2181 } 2182 2183 /* 2184 * If the size is inconsistant in the VMIO case, we can resize 2185 * the buffer. This might lead to B_CACHE getting set or 2186 * cleared. If the size has not changed, B_CACHE remains 2187 * unchanged from its previous state. 2188 */ 2189 2190 if (bp->b_bcount != size) 2191 allocbuf(bp, size); 2192 2193 KASSERT(bp->b_offset != NOOFFSET, 2194 ("getblk: no buffer offset")); 2195 2196 /* 2197 * A buffer with B_DELWRI set and B_CACHE clear must 2198 * be committed before we can return the buffer in 2199 * order to prevent the caller from issuing a read 2200 * ( due to B_CACHE not being set ) and overwriting 2201 * it. 2202 * 2203 * Most callers, including NFS and FFS, need this to 2204 * operate properly either because they assume they 2205 * can issue a read if B_CACHE is not set, or because 2206 * ( for example ) an uncached B_DELWRI might loop due 2207 * to softupdates re-dirtying the buffer. In the latter 2208 * case, B_CACHE is set after the first write completes, 2209 * preventing further loops. 2210 */ 2211 2212 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2213 BUF_WRITE(bp); 2214 goto loop; 2215 } 2216 2217 splx(s); 2218 bp->b_flags &= ~B_DONE; 2219 } else { 2220 /* 2221 * Buffer is not in-core, create new buffer. The buffer 2222 * returned by getnewbuf() is locked. Note that the returned 2223 * buffer is also considered valid (not marked B_INVAL). 2224 */ 2225 int bsize, maxsize, vmio; 2226 off_t offset; 2227 2228 if (vn_isdisk(vp, NULL)) 2229 bsize = DEV_BSIZE; 2230 else if (vp->v_mountedhere) 2231 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2232 else if (vp->v_mount) 2233 bsize = vp->v_mount->mnt_stat.f_iosize; 2234 else 2235 bsize = size; 2236 2237 offset = (off_t)blkno * bsize; 2238 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && (vp->v_flag & VOBJBUF); 2239 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2240 maxsize = imax(maxsize, bsize); 2241 2242 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2243 if (slpflag || slptimeo) { 2244 splx(s); 2245 return NULL; 2246 } 2247 goto loop; 2248 } 2249 2250 /* 2251 * This code is used to make sure that a buffer is not 2252 * created while the getnewbuf routine is blocked. 2253 * This can be a problem whether the vnode is locked or not. 2254 * If the buffer is created out from under us, we have to 2255 * throw away the one we just created. There is now window 2256 * race because we are safely running at splbio() from the 2257 * point of the duplicate buffer creation through to here, 2258 * and we've locked the buffer. 2259 */ 2260 if (gbincore(vp, blkno)) { 2261 bp->b_flags |= B_INVAL; 2262 brelse(bp); 2263 goto loop; 2264 } 2265 2266 /* 2267 * Insert the buffer into the hash, so that it can 2268 * be found by incore. 2269 */ 2270 bp->b_blkno = bp->b_lblkno = blkno; 2271 bp->b_offset = offset; 2272 2273 bgetvp(vp, bp); 2274 LIST_REMOVE(bp, b_hash); 2275 bh = bufhash(vp, blkno); 2276 LIST_INSERT_HEAD(bh, bp, b_hash); 2277 2278 /* 2279 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2280 * buffer size starts out as 0, B_CACHE will be set by 2281 * allocbuf() for the VMIO case prior to it testing the 2282 * backing store for validity. 2283 */ 2284 2285 if (vmio) { 2286 bp->b_flags |= B_VMIO; 2287#if defined(VFS_BIO_DEBUG) 2288 if (vp->v_type != VREG) 2289 printf("getblk: vmioing file type %d???\n", vp->v_type); 2290#endif 2291 } else { 2292 bp->b_flags &= ~B_VMIO; 2293 } 2294 2295 allocbuf(bp, size); 2296 2297 splx(s); 2298 bp->b_flags &= ~B_DONE; 2299 } 2300 return (bp); 2301} 2302 2303/* 2304 * Get an empty, disassociated buffer of given size. The buffer is initially 2305 * set to B_INVAL. 2306 */ 2307struct buf * 2308geteblk(int size) 2309{ 2310 struct buf *bp; 2311 int s; 2312 int maxsize; 2313 2314 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2315 2316 s = splbio(); 2317 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0); 2318 splx(s); 2319 allocbuf(bp, size); 2320 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2321 return (bp); 2322} 2323 2324 2325/* 2326 * This code constitutes the buffer memory from either anonymous system 2327 * memory (in the case of non-VMIO operations) or from an associated 2328 * VM object (in the case of VMIO operations). This code is able to 2329 * resize a buffer up or down. 2330 * 2331 * Note that this code is tricky, and has many complications to resolve 2332 * deadlock or inconsistant data situations. Tread lightly!!! 2333 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2334 * the caller. Calling this code willy nilly can result in the loss of data. 2335 * 2336 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2337 * B_CACHE for the non-VMIO case. 2338 */ 2339 2340int 2341allocbuf(struct buf *bp, int size) 2342{ 2343 int newbsize, mbsize; 2344 int i; 2345 2346 if (BUF_REFCNT(bp) == 0) 2347 panic("allocbuf: buffer not busy"); 2348 2349 if (bp->b_kvasize < size) 2350 panic("allocbuf: buffer too small"); 2351 2352 if ((bp->b_flags & B_VMIO) == 0) { 2353 caddr_t origbuf; 2354 int origbufsize; 2355 /* 2356 * Just get anonymous memory from the kernel. Don't 2357 * mess with B_CACHE. 2358 */ 2359 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2360#if !defined(NO_B_MALLOC) 2361 if (bp->b_flags & B_MALLOC) 2362 newbsize = mbsize; 2363 else 2364#endif 2365 newbsize = round_page(size); 2366 2367 if (newbsize < bp->b_bufsize) { 2368#if !defined(NO_B_MALLOC) 2369 /* 2370 * malloced buffers are not shrunk 2371 */ 2372 if (bp->b_flags & B_MALLOC) { 2373 if (newbsize) { 2374 bp->b_bcount = size; 2375 } else { 2376 free(bp->b_data, M_BIOBUF); 2377 if (bp->b_bufsize) { 2378 bufmallocspace -= bp->b_bufsize; 2379 bufspacewakeup(); 2380 bp->b_bufsize = 0; 2381 } 2382 bp->b_data = bp->b_kvabase; 2383 bp->b_bcount = 0; 2384 bp->b_flags &= ~B_MALLOC; 2385 } 2386 return 1; 2387 } 2388#endif 2389 vm_hold_free_pages( 2390 bp, 2391 (vm_offset_t) bp->b_data + newbsize, 2392 (vm_offset_t) bp->b_data + bp->b_bufsize); 2393 } else if (newbsize > bp->b_bufsize) { 2394#if !defined(NO_B_MALLOC) 2395 /* 2396 * We only use malloced memory on the first allocation. 2397 * and revert to page-allocated memory when the buffer 2398 * grows. 2399 */ 2400 if ( (bufmallocspace < maxbufmallocspace) && 2401 (bp->b_bufsize == 0) && 2402 (mbsize <= PAGE_SIZE/2)) { 2403 2404 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2405 bp->b_bufsize = mbsize; 2406 bp->b_bcount = size; 2407 bp->b_flags |= B_MALLOC; 2408 bufmallocspace += mbsize; 2409 return 1; 2410 } 2411#endif 2412 origbuf = NULL; 2413 origbufsize = 0; 2414#if !defined(NO_B_MALLOC) 2415 /* 2416 * If the buffer is growing on its other-than-first allocation, 2417 * then we revert to the page-allocation scheme. 2418 */ 2419 if (bp->b_flags & B_MALLOC) { 2420 origbuf = bp->b_data; 2421 origbufsize = bp->b_bufsize; 2422 bp->b_data = bp->b_kvabase; 2423 if (bp->b_bufsize) { 2424 bufmallocspace -= bp->b_bufsize; 2425 bufspacewakeup(); 2426 bp->b_bufsize = 0; 2427 } 2428 bp->b_flags &= ~B_MALLOC; 2429 newbsize = round_page(newbsize); 2430 } 2431#endif 2432 vm_hold_load_pages( 2433 bp, 2434 (vm_offset_t) bp->b_data + bp->b_bufsize, 2435 (vm_offset_t) bp->b_data + newbsize); 2436#if !defined(NO_B_MALLOC) 2437 if (origbuf) { 2438 bcopy(origbuf, bp->b_data, origbufsize); 2439 free(origbuf, M_BIOBUF); 2440 } 2441#endif 2442 } 2443 } else { 2444 vm_page_t m; 2445 int desiredpages; 2446 2447 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2448 desiredpages = (size == 0) ? 0 : 2449 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2450 2451#if !defined(NO_B_MALLOC) 2452 if (bp->b_flags & B_MALLOC) 2453 panic("allocbuf: VMIO buffer can't be malloced"); 2454#endif 2455 /* 2456 * Set B_CACHE initially if buffer is 0 length or will become 2457 * 0-length. 2458 */ 2459 if (size == 0 || bp->b_bufsize == 0) 2460 bp->b_flags |= B_CACHE; 2461 2462 if (newbsize < bp->b_bufsize) { 2463 /* 2464 * DEV_BSIZE aligned new buffer size is less then the 2465 * DEV_BSIZE aligned existing buffer size. Figure out 2466 * if we have to remove any pages. 2467 */ 2468 if (desiredpages < bp->b_npages) { 2469 for (i = desiredpages; i < bp->b_npages; i++) { 2470 /* 2471 * the page is not freed here -- it 2472 * is the responsibility of 2473 * vnode_pager_setsize 2474 */ 2475 m = bp->b_pages[i]; 2476 KASSERT(m != bogus_page, 2477 ("allocbuf: bogus page found")); 2478 while (vm_page_sleep_busy(m, TRUE, "biodep")) 2479 ; 2480 2481 bp->b_pages[i] = NULL; 2482 vm_page_unwire(m, 0); 2483 } 2484 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2485 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2486 bp->b_npages = desiredpages; 2487 } 2488 } else if (size > bp->b_bcount) { 2489 /* 2490 * We are growing the buffer, possibly in a 2491 * byte-granular fashion. 2492 */ 2493 struct vnode *vp; 2494 vm_object_t obj; 2495 vm_offset_t toff; 2496 vm_offset_t tinc; 2497 2498 /* 2499 * Step 1, bring in the VM pages from the object, 2500 * allocating them if necessary. We must clear 2501 * B_CACHE if these pages are not valid for the 2502 * range covered by the buffer. 2503 */ 2504 2505 vp = bp->b_vp; 2506 VOP_GETVOBJECT(vp, &obj); 2507 2508 while (bp->b_npages < desiredpages) { 2509 vm_page_t m; 2510 vm_pindex_t pi; 2511 2512 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2513 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2514 /* 2515 * note: must allocate system pages 2516 * since blocking here could intefere 2517 * with paging I/O, no matter which 2518 * process we are. 2519 */ 2520 m = vm_page_alloc(obj, pi, VM_ALLOC_SYSTEM); 2521 if (m == NULL) { 2522 VM_WAIT; 2523 vm_pageout_deficit += desiredpages - bp->b_npages; 2524 } else { 2525 vm_page_wire(m); 2526 vm_page_wakeup(m); 2527 bp->b_flags &= ~B_CACHE; 2528 bp->b_pages[bp->b_npages] = m; 2529 ++bp->b_npages; 2530 } 2531 continue; 2532 } 2533 2534 /* 2535 * We found a page. If we have to sleep on it, 2536 * retry because it might have gotten freed out 2537 * from under us. 2538 * 2539 * We can only test PG_BUSY here. Blocking on 2540 * m->busy might lead to a deadlock: 2541 * 2542 * vm_fault->getpages->cluster_read->allocbuf 2543 * 2544 */ 2545 2546 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 2547 continue; 2548 2549 /* 2550 * We have a good page. Should we wakeup the 2551 * page daemon? 2552 */ 2553 if ((curproc != pageproc) && 2554 ((m->queue - m->pc) == PQ_CACHE) && 2555 ((cnt.v_free_count + cnt.v_cache_count) < 2556 (cnt.v_free_min + cnt.v_cache_min))) { 2557 pagedaemon_wakeup(); 2558 } 2559 vm_page_flag_clear(m, PG_ZERO); 2560 vm_page_wire(m); 2561 bp->b_pages[bp->b_npages] = m; 2562 ++bp->b_npages; 2563 } 2564 2565 /* 2566 * Step 2. We've loaded the pages into the buffer, 2567 * we have to figure out if we can still have B_CACHE 2568 * set. Note that B_CACHE is set according to the 2569 * byte-granular range ( bcount and size ), new the 2570 * aligned range ( newbsize ). 2571 * 2572 * The VM test is against m->valid, which is DEV_BSIZE 2573 * aligned. Needless to say, the validity of the data 2574 * needs to also be DEV_BSIZE aligned. Note that this 2575 * fails with NFS if the server or some other client 2576 * extends the file's EOF. If our buffer is resized, 2577 * B_CACHE may remain set! XXX 2578 */ 2579 2580 toff = bp->b_bcount; 2581 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2582 2583 while ((bp->b_flags & B_CACHE) && toff < size) { 2584 vm_pindex_t pi; 2585 2586 if (tinc > (size - toff)) 2587 tinc = size - toff; 2588 2589 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2590 PAGE_SHIFT; 2591 2592 vfs_buf_test_cache( 2593 bp, 2594 bp->b_offset, 2595 toff, 2596 tinc, 2597 bp->b_pages[pi] 2598 ); 2599 toff += tinc; 2600 tinc = PAGE_SIZE; 2601 } 2602 2603 /* 2604 * Step 3, fixup the KVM pmap. Remember that 2605 * bp->b_data is relative to bp->b_offset, but 2606 * bp->b_offset may be offset into the first page. 2607 */ 2608 2609 bp->b_data = (caddr_t) 2610 trunc_page((vm_offset_t)bp->b_data); 2611 pmap_qenter( 2612 (vm_offset_t)bp->b_data, 2613 bp->b_pages, 2614 bp->b_npages 2615 ); 2616 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2617 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2618 } 2619 } 2620 if (newbsize < bp->b_bufsize) 2621 bufspacewakeup(); 2622 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2623 bp->b_bcount = size; /* requested buffer size */ 2624 return 1; 2625} 2626 2627/* 2628 * bufwait: 2629 * 2630 * Wait for buffer I/O completion, returning error status. The buffer 2631 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR 2632 * error and cleared. 2633 */ 2634int 2635bufwait(register struct buf * bp) 2636{ 2637 int s; 2638 2639 s = splbio(); 2640 while ((bp->b_flags & B_DONE) == 0) { 2641 if (bp->b_iocmd == BIO_READ) 2642 tsleep(bp, PRIBIO, "biord", 0); 2643 else 2644 tsleep(bp, PRIBIO, "biowr", 0); 2645 } 2646 splx(s); 2647 if (bp->b_flags & B_EINTR) { 2648 bp->b_flags &= ~B_EINTR; 2649 return (EINTR); 2650 } 2651 if (bp->b_ioflags & BIO_ERROR) { 2652 return (bp->b_error ? bp->b_error : EIO); 2653 } else { 2654 return (0); 2655 } 2656} 2657 2658 /* 2659 * Call back function from struct bio back up to struct buf. 2660 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 2661 */ 2662void 2663bufdonebio(struct bio *bp) 2664{ 2665 bufdone(bp->bio_caller2); 2666} 2667 2668/* 2669 * bufdone: 2670 * 2671 * Finish I/O on a buffer, optionally calling a completion function. 2672 * This is usually called from an interrupt so process blocking is 2673 * not allowed. 2674 * 2675 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2676 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2677 * assuming B_INVAL is clear. 2678 * 2679 * For the VMIO case, we set B_CACHE if the op was a read and no 2680 * read error occured, or if the op was a write. B_CACHE is never 2681 * set if the buffer is invalid or otherwise uncacheable. 2682 * 2683 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2684 * initiator to leave B_INVAL set to brelse the buffer out of existance 2685 * in the biodone routine. 2686 */ 2687void 2688bufdone(struct buf *bp) 2689{ 2690 int s, error; 2691 void (*biodone) __P((struct buf *)); 2692 2693 s = splbio(); 2694 2695 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 2696 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 2697 2698 bp->b_flags |= B_DONE; 2699 runningbufwakeup(bp); 2700 2701 if (bp->b_iocmd == BIO_DELETE) { 2702 brelse(bp); 2703 splx(s); 2704 return; 2705 } 2706 2707 if (bp->b_iocmd == BIO_WRITE) { 2708 vwakeup(bp); 2709 } 2710 2711 /* call optional completion function if requested */ 2712 if (bp->b_iodone != NULL) { 2713 biodone = bp->b_iodone; 2714 bp->b_iodone = NULL; 2715 (*biodone) (bp); 2716 splx(s); 2717 return; 2718 } 2719 if (LIST_FIRST(&bp->b_dep) != NULL) 2720 buf_complete(bp); 2721 2722 if (bp->b_flags & B_VMIO) { 2723 int i; 2724 vm_ooffset_t foff; 2725 vm_page_t m; 2726 vm_object_t obj; 2727 int iosize; 2728 struct vnode *vp = bp->b_vp; 2729 2730 error = VOP_GETVOBJECT(vp, &obj); 2731 2732#if defined(VFS_BIO_DEBUG) 2733 if (vp->v_usecount == 0) { 2734 panic("biodone: zero vnode ref count"); 2735 } 2736 2737 if (error) { 2738 panic("biodone: missing VM object"); 2739 } 2740 2741 if ((vp->v_flag & VOBJBUF) == 0) { 2742 panic("biodone: vnode is not setup for merged cache"); 2743 } 2744#endif 2745 2746 foff = bp->b_offset; 2747 KASSERT(bp->b_offset != NOOFFSET, 2748 ("biodone: no buffer offset")); 2749 2750 if (error) { 2751 panic("biodone: no object"); 2752 } 2753#if defined(VFS_BIO_DEBUG) 2754 if (obj->paging_in_progress < bp->b_npages) { 2755 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 2756 obj->paging_in_progress, bp->b_npages); 2757 } 2758#endif 2759 2760 /* 2761 * Set B_CACHE if the op was a normal read and no error 2762 * occured. B_CACHE is set for writes in the b*write() 2763 * routines. 2764 */ 2765 iosize = bp->b_bcount - bp->b_resid; 2766 if (bp->b_iocmd == BIO_READ && 2767 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 2768 !(bp->b_ioflags & BIO_ERROR)) { 2769 bp->b_flags |= B_CACHE; 2770 } 2771 2772 for (i = 0; i < bp->b_npages; i++) { 2773 int bogusflag = 0; 2774 int resid; 2775 2776 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2777 if (resid > iosize) 2778 resid = iosize; 2779 2780 /* 2781 * cleanup bogus pages, restoring the originals 2782 */ 2783 m = bp->b_pages[i]; 2784 if (m == bogus_page) { 2785 bogusflag = 1; 2786 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2787 if (m == NULL) 2788 panic("biodone: page disappeared!"); 2789 bp->b_pages[i] = m; 2790 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2791 } 2792#if defined(VFS_BIO_DEBUG) 2793 if (OFF_TO_IDX(foff) != m->pindex) { 2794 printf( 2795"biodone: foff(%lu)/m->pindex(%d) mismatch\n", 2796 (unsigned long)foff, m->pindex); 2797 } 2798#endif 2799 2800 /* 2801 * In the write case, the valid and clean bits are 2802 * already changed correctly ( see bdwrite() ), so we 2803 * only need to do this here in the read case. 2804 */ 2805 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 2806 vfs_page_set_valid(bp, foff, i, m); 2807 } 2808 vm_page_flag_clear(m, PG_ZERO); 2809 2810 /* 2811 * when debugging new filesystems or buffer I/O methods, this 2812 * is the most common error that pops up. if you see this, you 2813 * have not set the page busy flag correctly!!! 2814 */ 2815 if (m->busy == 0) { 2816 printf("biodone: page busy < 0, " 2817 "pindex: %d, foff: 0x(%x,%x), " 2818 "resid: %d, index: %d\n", 2819 (int) m->pindex, (int)(foff >> 32), 2820 (int) foff & 0xffffffff, resid, i); 2821 if (!vn_isdisk(vp, NULL)) 2822 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 2823 bp->b_vp->v_mount->mnt_stat.f_iosize, 2824 (int) bp->b_lblkno, 2825 bp->b_flags, bp->b_npages); 2826 else 2827 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 2828 (int) bp->b_lblkno, 2829 bp->b_flags, bp->b_npages); 2830 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2831 m->valid, m->dirty, m->wire_count); 2832 panic("biodone: page busy < 0\n"); 2833 } 2834 vm_page_io_finish(m); 2835 vm_object_pip_subtract(obj, 1); 2836 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2837 iosize -= resid; 2838 } 2839 if (obj) 2840 vm_object_pip_wakeupn(obj, 0); 2841 } 2842 2843 /* 2844 * For asynchronous completions, release the buffer now. The brelse 2845 * will do a wakeup there if necessary - so no need to do a wakeup 2846 * here in the async case. The sync case always needs to do a wakeup. 2847 */ 2848 2849 if (bp->b_flags & B_ASYNC) { 2850 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 2851 brelse(bp); 2852 else 2853 bqrelse(bp); 2854 } else { 2855 wakeup(bp); 2856 } 2857 splx(s); 2858} 2859 2860/* 2861 * This routine is called in lieu of iodone in the case of 2862 * incomplete I/O. This keeps the busy status for pages 2863 * consistant. 2864 */ 2865void 2866vfs_unbusy_pages(struct buf * bp) 2867{ 2868 int i; 2869 2870 runningbufwakeup(bp); 2871 if (bp->b_flags & B_VMIO) { 2872 struct vnode *vp = bp->b_vp; 2873 vm_object_t obj; 2874 2875 VOP_GETVOBJECT(vp, &obj); 2876 2877 for (i = 0; i < bp->b_npages; i++) { 2878 vm_page_t m = bp->b_pages[i]; 2879 2880 if (m == bogus_page) { 2881 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 2882 if (!m) { 2883 panic("vfs_unbusy_pages: page missing\n"); 2884 } 2885 bp->b_pages[i] = m; 2886 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2887 } 2888 vm_object_pip_subtract(obj, 1); 2889 vm_page_flag_clear(m, PG_ZERO); 2890 vm_page_io_finish(m); 2891 } 2892 vm_object_pip_wakeupn(obj, 0); 2893 } 2894} 2895 2896/* 2897 * vfs_page_set_valid: 2898 * 2899 * Set the valid bits in a page based on the supplied offset. The 2900 * range is restricted to the buffer's size. 2901 * 2902 * This routine is typically called after a read completes. 2903 */ 2904static void 2905vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 2906{ 2907 vm_ooffset_t soff, eoff; 2908 2909 /* 2910 * Start and end offsets in buffer. eoff - soff may not cross a 2911 * page boundry or cross the end of the buffer. The end of the 2912 * buffer, in this case, is our file EOF, not the allocation size 2913 * of the buffer. 2914 */ 2915 soff = off; 2916 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2917 if (eoff > bp->b_offset + bp->b_bcount) 2918 eoff = bp->b_offset + bp->b_bcount; 2919 2920 /* 2921 * Set valid range. This is typically the entire buffer and thus the 2922 * entire page. 2923 */ 2924 if (eoff > soff) { 2925 vm_page_set_validclean( 2926 m, 2927 (vm_offset_t) (soff & PAGE_MASK), 2928 (vm_offset_t) (eoff - soff) 2929 ); 2930 } 2931} 2932 2933/* 2934 * This routine is called before a device strategy routine. 2935 * It is used to tell the VM system that paging I/O is in 2936 * progress, and treat the pages associated with the buffer 2937 * almost as being PG_BUSY. Also the object paging_in_progress 2938 * flag is handled to make sure that the object doesn't become 2939 * inconsistant. 2940 * 2941 * Since I/O has not been initiated yet, certain buffer flags 2942 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 2943 * and should be ignored. 2944 */ 2945void 2946vfs_busy_pages(struct buf * bp, int clear_modify) 2947{ 2948 int i, bogus; 2949 2950 if (bp->b_flags & B_VMIO) { 2951 struct vnode *vp = bp->b_vp; 2952 vm_object_t obj; 2953 vm_ooffset_t foff; 2954 2955 VOP_GETVOBJECT(vp, &obj); 2956 foff = bp->b_offset; 2957 KASSERT(bp->b_offset != NOOFFSET, 2958 ("vfs_busy_pages: no buffer offset")); 2959 vfs_setdirty(bp); 2960 2961retry: 2962 for (i = 0; i < bp->b_npages; i++) { 2963 vm_page_t m = bp->b_pages[i]; 2964 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 2965 goto retry; 2966 } 2967 2968 bogus = 0; 2969 for (i = 0; i < bp->b_npages; i++) { 2970 vm_page_t m = bp->b_pages[i]; 2971 2972 vm_page_flag_clear(m, PG_ZERO); 2973 if ((bp->b_flags & B_CLUSTER) == 0) { 2974 vm_object_pip_add(obj, 1); 2975 vm_page_io_start(m); 2976 } 2977 2978 /* 2979 * When readying a buffer for a read ( i.e 2980 * clear_modify == 0 ), it is important to do 2981 * bogus_page replacement for valid pages in 2982 * partially instantiated buffers. Partially 2983 * instantiated buffers can, in turn, occur when 2984 * reconstituting a buffer from its VM backing store 2985 * base. We only have to do this if B_CACHE is 2986 * clear ( which causes the I/O to occur in the 2987 * first place ). The replacement prevents the read 2988 * I/O from overwriting potentially dirty VM-backed 2989 * pages. XXX bogus page replacement is, uh, bogus. 2990 * It may not work properly with small-block devices. 2991 * We need to find a better way. 2992 */ 2993 2994 vm_page_protect(m, VM_PROT_NONE); 2995 if (clear_modify) 2996 vfs_page_set_valid(bp, foff, i, m); 2997 else if (m->valid == VM_PAGE_BITS_ALL && 2998 (bp->b_flags & B_CACHE) == 0) { 2999 bp->b_pages[i] = bogus_page; 3000 bogus++; 3001 } 3002 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3003 } 3004 if (bogus) 3005 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3006 } 3007} 3008 3009/* 3010 * Tell the VM system that the pages associated with this buffer 3011 * are clean. This is used for delayed writes where the data is 3012 * going to go to disk eventually without additional VM intevention. 3013 * 3014 * Note that while we only really need to clean through to b_bcount, we 3015 * just go ahead and clean through to b_bufsize. 3016 */ 3017static void 3018vfs_clean_pages(struct buf * bp) 3019{ 3020 int i; 3021 3022 if (bp->b_flags & B_VMIO) { 3023 vm_ooffset_t foff; 3024 3025 foff = bp->b_offset; 3026 KASSERT(bp->b_offset != NOOFFSET, 3027 ("vfs_clean_pages: no buffer offset")); 3028 for (i = 0; i < bp->b_npages; i++) { 3029 vm_page_t m = bp->b_pages[i]; 3030 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3031 vm_ooffset_t eoff = noff; 3032 3033 if (eoff > bp->b_offset + bp->b_bufsize) 3034 eoff = bp->b_offset + bp->b_bufsize; 3035 vfs_page_set_valid(bp, foff, i, m); 3036 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3037 foff = noff; 3038 } 3039 } 3040} 3041 3042/* 3043 * vfs_bio_set_validclean: 3044 * 3045 * Set the range within the buffer to valid and clean. The range is 3046 * relative to the beginning of the buffer, b_offset. Note that b_offset 3047 * itself may be offset from the beginning of the first page. 3048 */ 3049 3050void 3051vfs_bio_set_validclean(struct buf *bp, int base, int size) 3052{ 3053 if (bp->b_flags & B_VMIO) { 3054 int i; 3055 int n; 3056 3057 /* 3058 * Fixup base to be relative to beginning of first page. 3059 * Set initial n to be the maximum number of bytes in the 3060 * first page that can be validated. 3061 */ 3062 3063 base += (bp->b_offset & PAGE_MASK); 3064 n = PAGE_SIZE - (base & PAGE_MASK); 3065 3066 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3067 vm_page_t m = bp->b_pages[i]; 3068 3069 if (n > size) 3070 n = size; 3071 3072 vm_page_set_validclean(m, base & PAGE_MASK, n); 3073 base += n; 3074 size -= n; 3075 n = PAGE_SIZE; 3076 } 3077 } 3078} 3079 3080/* 3081 * vfs_bio_clrbuf: 3082 * 3083 * clear a buffer. This routine essentially fakes an I/O, so we need 3084 * to clear BIO_ERROR and B_INVAL. 3085 * 3086 * Note that while we only theoretically need to clear through b_bcount, 3087 * we go ahead and clear through b_bufsize. 3088 */ 3089 3090void 3091vfs_bio_clrbuf(struct buf *bp) { 3092 int i, mask = 0; 3093 caddr_t sa, ea; 3094 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3095 bp->b_flags &= ~B_INVAL; 3096 bp->b_ioflags &= ~BIO_ERROR; 3097 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3098 (bp->b_offset & PAGE_MASK) == 0) { 3099 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3100 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3101 ((bp->b_pages[0]->valid & mask) != mask)) { 3102 bzero(bp->b_data, bp->b_bufsize); 3103 } 3104 bp->b_pages[0]->valid |= mask; 3105 bp->b_resid = 0; 3106 return; 3107 } 3108 ea = sa = bp->b_data; 3109 for(i=0;i<bp->b_npages;i++,sa=ea) { 3110 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3111 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3112 ea = (caddr_t)(vm_offset_t)ulmin( 3113 (u_long)(vm_offset_t)ea, 3114 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3115 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3116 if ((bp->b_pages[i]->valid & mask) == mask) 3117 continue; 3118 if ((bp->b_pages[i]->valid & mask) == 0) { 3119 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3120 bzero(sa, ea - sa); 3121 } 3122 } else { 3123 for (; sa < ea; sa += DEV_BSIZE, j++) { 3124 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3125 (bp->b_pages[i]->valid & (1<<j)) == 0) 3126 bzero(sa, DEV_BSIZE); 3127 } 3128 } 3129 bp->b_pages[i]->valid |= mask; 3130 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3131 } 3132 bp->b_resid = 0; 3133 } else { 3134 clrbuf(bp); 3135 } 3136} 3137 3138/* 3139 * vm_hold_load_pages and vm_hold_unload pages get pages into 3140 * a buffers address space. The pages are anonymous and are 3141 * not associated with a file object. 3142 */ 3143void 3144vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3145{ 3146 vm_offset_t pg; 3147 vm_page_t p; 3148 int index; 3149 3150 to = round_page(to); 3151 from = round_page(from); 3152 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3153 3154 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3155 3156tryagain: 3157 3158 /* 3159 * note: must allocate system pages since blocking here 3160 * could intefere with paging I/O, no matter which 3161 * process we are. 3162 */ 3163 p = vm_page_alloc(kernel_object, 3164 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3165 VM_ALLOC_SYSTEM); 3166 if (!p) { 3167 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3168 VM_WAIT; 3169 goto tryagain; 3170 } 3171 vm_page_wire(p); 3172 p->valid = VM_PAGE_BITS_ALL; 3173 vm_page_flag_clear(p, PG_ZERO); 3174 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 3175 bp->b_pages[index] = p; 3176 vm_page_wakeup(p); 3177 } 3178 bp->b_npages = index; 3179} 3180 3181void 3182vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3183{ 3184 vm_offset_t pg; 3185 vm_page_t p; 3186 int index, newnpages; 3187 3188 from = round_page(from); 3189 to = round_page(to); 3190 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3191 3192 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3193 p = bp->b_pages[index]; 3194 if (p && (index < bp->b_npages)) { 3195 if (p->busy) { 3196 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 3197 bp->b_blkno, bp->b_lblkno); 3198 } 3199 bp->b_pages[index] = NULL; 3200 pmap_kremove(pg); 3201 vm_page_busy(p); 3202 vm_page_unwire(p, 0); 3203 vm_page_free(p); 3204 } 3205 } 3206 bp->b_npages = newnpages; 3207} 3208 3209 3210#include "opt_ddb.h" 3211#ifdef DDB 3212#include <ddb/ddb.h> 3213 3214DB_SHOW_COMMAND(buffer, db_show_buffer) 3215{ 3216 /* get args */ 3217 struct buf *bp = (struct buf *)addr; 3218 3219 if (!have_addr) { 3220 db_printf("usage: show buffer <addr>\n"); 3221 return; 3222 } 3223 3224 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3225 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 3226 "b_resid = %ld\nb_dev = (%d,%d), b_data = %p, " 3227 "b_blkno = %d, b_pblkno = %d\n", 3228 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3229 major(bp->b_dev), minor(bp->b_dev), 3230 bp->b_data, bp->b_blkno, bp->b_pblkno); 3231 if (bp->b_npages) { 3232 int i; 3233 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3234 for (i = 0; i < bp->b_npages; i++) { 3235 vm_page_t m; 3236 m = bp->b_pages[i]; 3237 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3238 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3239 if ((i + 1) < bp->b_npages) 3240 db_printf(","); 3241 } 3242 db_printf("\n"); 3243 } 3244} 3245#endif /* DDB */ 3246