vfs_bio.c revision 58926
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 58926 2000-04-02 09:26:51Z phk $ 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/buf.h> 33#include <sys/conf.h> 34#include <sys/eventhandler.h> 35#include <sys/lock.h> 36#include <sys/malloc.h> 37#include <sys/mount.h> 38#include <sys/kernel.h> 39#include <sys/kthread.h> 40#include <sys/proc.h> 41#include <sys/reboot.h> 42#include <sys/resourcevar.h> 43#include <sys/sysctl.h> 44#include <sys/vmmeter.h> 45#include <sys/vnode.h> 46#include <vm/vm.h> 47#include <vm/vm_param.h> 48#include <vm/vm_kern.h> 49#include <vm/vm_pageout.h> 50#include <vm/vm_page.h> 51#include <vm/vm_object.h> 52#include <vm/vm_extern.h> 53#include <vm/vm_map.h> 54 55static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 56 57struct bio_ops bioops; /* I/O operation notification */ 58 59struct buf *buf; /* buffer header pool */ 60struct swqueue bswlist; 61 62static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 63 vm_offset_t to); 64static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 65 vm_offset_t to); 66static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 67 int pageno, vm_page_t m); 68static void vfs_clean_pages(struct buf * bp); 69static void vfs_setdirty(struct buf *bp); 70static void vfs_vmio_release(struct buf *bp); 71static void vfs_backgroundwritedone(struct buf *bp); 72static int flushbufqueues(void); 73 74static int bd_request; 75 76static void buf_daemon __P((void)); 77/* 78 * bogus page -- for I/O to/from partially complete buffers 79 * this is a temporary solution to the problem, but it is not 80 * really that bad. it would be better to split the buffer 81 * for input in the case of buffers partially already in memory, 82 * but the code is intricate enough already. 83 */ 84vm_page_t bogus_page; 85int runningbufspace; 86int vmiodirenable = FALSE; 87static vm_offset_t bogus_offset; 88 89static int bufspace, maxbufspace, 90 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace; 91static int bufreusecnt, bufdefragcnt, buffreekvacnt; 92static int maxbdrun; 93static int needsbuffer; 94static int numdirtybuffers, hidirtybuffers; 95static int numfreebuffers, lofreebuffers, hifreebuffers; 96static int getnewbufcalls; 97static int getnewbufrestarts; 98 99SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, 100 &numdirtybuffers, 0, ""); 101SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, 102 &hidirtybuffers, 0, ""); 103SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, 104 &numfreebuffers, 0, ""); 105SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, 106 &lofreebuffers, 0, ""); 107SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, 108 &hifreebuffers, 0, ""); 109SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, 110 &runningbufspace, 0, ""); 111SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, 112 &maxbufspace, 0, ""); 113SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, 114 &hibufspace, 0, ""); 115SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, 116 &lobufspace, 0, ""); 117SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, 118 &bufspace, 0, ""); 119SYSCTL_INT(_vfs, OID_AUTO, maxbdrun, CTLFLAG_RW, 120 &maxbdrun, 0, ""); 121SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, 122 &maxbufmallocspace, 0, ""); 123SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, 124 &bufmallocspace, 0, ""); 125SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, 126 &getnewbufcalls, 0, ""); 127SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, 128 &getnewbufrestarts, 0, ""); 129SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, 130 &vmiodirenable, 0, ""); 131SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, 132 &bufdefragcnt, 0, ""); 133SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, 134 &buffreekvacnt, 0, ""); 135SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, 136 &bufreusecnt, 0, ""); 137 138static int bufhashmask; 139static LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash; 140struct bqueues bufqueues[BUFFER_QUEUES] = { { 0 } }; 141char *buf_wmesg = BUF_WMESG; 142 143extern int vm_swap_size; 144 145#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 146#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 147#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 148#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 149 150/* 151 * Buffer hash table code. Note that the logical block scans linearly, which 152 * gives us some L1 cache locality. 153 */ 154 155static __inline 156struct bufhashhdr * 157bufhash(struct vnode *vnp, daddr_t bn) 158{ 159 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]); 160} 161 162/* 163 * numdirtywakeup: 164 * 165 * If someone is blocked due to there being too many dirty buffers, 166 * and numdirtybuffers is now reasonable, wake them up. 167 */ 168 169static __inline void 170numdirtywakeup(void) 171{ 172 if (numdirtybuffers < hidirtybuffers) { 173 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 174 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 175 wakeup(&needsbuffer); 176 } 177 } 178} 179 180/* 181 * bufspacewakeup: 182 * 183 * Called when buffer space is potentially available for recovery. 184 * getnewbuf() will block on this flag when it is unable to free 185 * sufficient buffer space. Buffer space becomes recoverable when 186 * bp's get placed back in the queues. 187 */ 188 189static __inline void 190bufspacewakeup(void) 191{ 192 /* 193 * If someone is waiting for BUF space, wake them up. Even 194 * though we haven't freed the kva space yet, the waiting 195 * process will be able to now. 196 */ 197 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 198 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 199 wakeup(&needsbuffer); 200 } 201} 202 203/* 204 * bufcountwakeup: 205 * 206 * Called when a buffer has been added to one of the free queues to 207 * account for the buffer and to wakeup anyone waiting for free buffers. 208 * This typically occurs when large amounts of metadata are being handled 209 * by the buffer cache ( else buffer space runs out first, usually ). 210 */ 211 212static __inline void 213bufcountwakeup(void) 214{ 215 ++numfreebuffers; 216 if (needsbuffer) { 217 needsbuffer &= ~VFS_BIO_NEED_ANY; 218 if (numfreebuffers >= hifreebuffers) 219 needsbuffer &= ~VFS_BIO_NEED_FREE; 220 wakeup(&needsbuffer); 221 } 222} 223 224/* 225 * vfs_buf_test_cache: 226 * 227 * Called when a buffer is extended. This function clears the B_CACHE 228 * bit if the newly extended portion of the buffer does not contain 229 * valid data. 230 */ 231static __inline__ 232void 233vfs_buf_test_cache(struct buf *bp, 234 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 235 vm_page_t m) 236{ 237 if (bp->b_flags & B_CACHE) { 238 int base = (foff + off) & PAGE_MASK; 239 if (vm_page_is_valid(m, base, size) == 0) 240 bp->b_flags &= ~B_CACHE; 241 } 242} 243 244static __inline__ 245void 246bd_wakeup(int dirtybuflevel) 247{ 248 if (numdirtybuffers >= dirtybuflevel && bd_request == 0) { 249 bd_request = 1; 250 wakeup(&bd_request); 251 } 252} 253 254/* 255 * bd_speedup - speedup the buffer cache flushing code 256 */ 257 258static __inline__ 259void 260bd_speedup(void) 261{ 262 bd_wakeup(1); 263} 264 265/* 266 * Initialize buffer headers and related structures. 267 */ 268 269caddr_t 270bufhashinit(caddr_t vaddr) 271{ 272 /* first, make a null hash table */ 273 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 274 ; 275 bufhashtbl = (void *)vaddr; 276 vaddr = vaddr + sizeof(*bufhashtbl) * bufhashmask; 277 --bufhashmask; 278 return(vaddr); 279} 280 281void 282bufinit(void) 283{ 284 struct buf *bp; 285 int i; 286 287 TAILQ_INIT(&bswlist); 288 LIST_INIT(&invalhash); 289 simple_lock_init(&buftimelock); 290 291 for (i = 0; i <= bufhashmask; i++) 292 LIST_INIT(&bufhashtbl[i]); 293 294 /* next, make a null set of free lists */ 295 for (i = 0; i < BUFFER_QUEUES; i++) 296 TAILQ_INIT(&bufqueues[i]); 297 298 /* finally, initialize each buffer header and stick on empty q */ 299 for (i = 0; i < nbuf; i++) { 300 bp = &buf[i]; 301 bzero(bp, sizeof *bp); 302 bp->b_flags = B_INVAL; /* we're just an empty header */ 303 bp->b_dev = NODEV; 304 bp->b_rcred = NOCRED; 305 bp->b_wcred = NOCRED; 306 bp->b_qindex = QUEUE_EMPTY; 307 bp->b_xflags = 0; 308 LIST_INIT(&bp->b_dep); 309 BUF_LOCKINIT(bp); 310 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 311 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 312 } 313 314 /* 315 * maxbufspace is the absolute maximum amount of buffer space we are 316 * allowed to reserve in KVM and in real terms. The absolute maximum 317 * is nominally used by buf_daemon. hibufspace is the nominal maximum 318 * used by most other processes. The differential is required to 319 * ensure that buf_daemon is able to run when other processes might 320 * be blocked waiting for buffer space. 321 * 322 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 323 * this may result in KVM fragmentation which is not handled optimally 324 * by the system. 325 */ 326 maxbufspace = nbuf * BKVASIZE; 327 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 328 lobufspace = hibufspace - MAXBSIZE; 329 330/* 331 * Limit the amount of malloc memory since it is wired permanently into 332 * the kernel space. Even though this is accounted for in the buffer 333 * allocation, we don't want the malloced region to grow uncontrolled. 334 * The malloc scheme improves memory utilization significantly on average 335 * (small) directories. 336 */ 337 maxbufmallocspace = hibufspace / 20; 338 339/* 340 * Reduce the chance of a deadlock occuring by limiting the number 341 * of delayed-write dirty buffers we allow to stack up. 342 */ 343 hidirtybuffers = nbuf / 4 + 20; 344 numdirtybuffers = 0; 345/* 346 * To support extreme low-memory systems, make sure hidirtybuffers cannot 347 * eat up all available buffer space. This occurs when our minimum cannot 348 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 349 * BKVASIZE'd (8K) buffers. 350 */ 351 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 352 hidirtybuffers >>= 1; 353 } 354 355/* 356 * Try to keep the number of free buffers in the specified range, 357 * and give special processes (e.g. like buf_daemon) access to an 358 * emergency reserve. 359 */ 360 lofreebuffers = nbuf / 18 + 5; 361 hifreebuffers = 2 * lofreebuffers; 362 numfreebuffers = nbuf; 363 364/* 365 * Maximum number of async ops initiated per buf_daemon loop. This is 366 * somewhat of a hack at the moment, we really need to limit ourselves 367 * based on the number of bytes of I/O in-transit that were initiated 368 * from buf_daemon. 369 */ 370 if ((maxbdrun = nswbuf / 4) < 4) 371 maxbdrun = 4; 372 373 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 374 bogus_page = vm_page_alloc(kernel_object, 375 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 376 VM_ALLOC_NORMAL); 377 cnt.v_wire_count++; 378 379} 380 381/* 382 * bfreekva() - free the kva allocation for a buffer. 383 * 384 * Must be called at splbio() or higher as this is the only locking for 385 * buffer_map. 386 * 387 * Since this call frees up buffer space, we call bufspacewakeup(). 388 */ 389static void 390bfreekva(struct buf * bp) 391{ 392 if (bp->b_kvasize) { 393 ++buffreekvacnt; 394 bufspace -= bp->b_kvasize; 395 vm_map_delete(buffer_map, 396 (vm_offset_t) bp->b_kvabase, 397 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 398 ); 399 bp->b_kvasize = 0; 400 bufspacewakeup(); 401 } 402} 403 404/* 405 * bremfree: 406 * 407 * Remove the buffer from the appropriate free list. 408 */ 409void 410bremfree(struct buf * bp) 411{ 412 int s = splbio(); 413 int old_qindex = bp->b_qindex; 414 415 if (bp->b_qindex != QUEUE_NONE) { 416 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 417 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 418 bp->b_qindex = QUEUE_NONE; 419 runningbufspace += bp->b_bufsize; 420 } else { 421 if (BUF_REFCNT(bp) <= 1) 422 panic("bremfree: removing a buffer not on a queue"); 423 } 424 425 /* 426 * Fixup numfreebuffers count. If the buffer is invalid or not 427 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 428 * the buffer was free and we must decrement numfreebuffers. 429 */ 430 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 431 switch(old_qindex) { 432 case QUEUE_DIRTY: 433 case QUEUE_CLEAN: 434 case QUEUE_EMPTY: 435 case QUEUE_EMPTYKVA: 436 --numfreebuffers; 437 break; 438 default: 439 break; 440 } 441 } 442 splx(s); 443} 444 445 446/* 447 * Get a buffer with the specified data. Look in the cache first. We 448 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 449 * is set, the buffer is valid and we do not have to do anything ( see 450 * getblk() ). 451 */ 452int 453bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 454 struct buf ** bpp) 455{ 456 struct buf *bp; 457 458 bp = getblk(vp, blkno, size, 0, 0); 459 *bpp = bp; 460 461 /* if not found in cache, do some I/O */ 462 if ((bp->b_flags & B_CACHE) == 0) { 463 if (curproc != NULL) 464 curproc->p_stats->p_ru.ru_inblock++; 465 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp)); 466 bp->b_iocmd = BIO_READ; 467 bp->b_flags &= ~(B_ERROR | B_INVAL); 468 if (bp->b_rcred == NOCRED) { 469 if (cred != NOCRED) 470 crhold(cred); 471 bp->b_rcred = cred; 472 } 473 vfs_busy_pages(bp, 0); 474 VOP_STRATEGY(vp, bp); 475 return (biowait(bp)); 476 } 477 return (0); 478} 479 480/* 481 * Operates like bread, but also starts asynchronous I/O on 482 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior 483 * to initiating I/O . If B_CACHE is set, the buffer is valid 484 * and we do not have to do anything. 485 */ 486int 487breadn(struct vnode * vp, daddr_t blkno, int size, 488 daddr_t * rablkno, int *rabsize, 489 int cnt, struct ucred * cred, struct buf ** bpp) 490{ 491 struct buf *bp, *rabp; 492 int i; 493 int rv = 0, readwait = 0; 494 495 *bpp = bp = getblk(vp, blkno, size, 0, 0); 496 497 /* if not found in cache, do some I/O */ 498 if ((bp->b_flags & B_CACHE) == 0) { 499 if (curproc != NULL) 500 curproc->p_stats->p_ru.ru_inblock++; 501 bp->b_iocmd = BIO_READ; 502 bp->b_flags &= ~(B_ERROR | B_INVAL); 503 if (bp->b_rcred == NOCRED) { 504 if (cred != NOCRED) 505 crhold(cred); 506 bp->b_rcred = cred; 507 } 508 vfs_busy_pages(bp, 0); 509 VOP_STRATEGY(vp, bp); 510 ++readwait; 511 } 512 513 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 514 if (inmem(vp, *rablkno)) 515 continue; 516 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 517 518 if ((rabp->b_flags & B_CACHE) == 0) { 519 if (curproc != NULL) 520 curproc->p_stats->p_ru.ru_inblock++; 521 rabp->b_flags |= B_ASYNC; 522 rabp->b_flags &= ~(B_ERROR | B_INVAL); 523 rabp->b_iocmd = BIO_READ; 524 if (rabp->b_rcred == NOCRED) { 525 if (cred != NOCRED) 526 crhold(cred); 527 rabp->b_rcred = cred; 528 } 529 vfs_busy_pages(rabp, 0); 530 BUF_KERNPROC(rabp); 531 VOP_STRATEGY(vp, rabp); 532 } else { 533 brelse(rabp); 534 } 535 } 536 537 if (readwait) { 538 rv = biowait(bp); 539 } 540 return (rv); 541} 542 543/* 544 * Write, release buffer on completion. (Done by iodone 545 * if async). Do not bother writing anything if the buffer 546 * is invalid. 547 * 548 * Note that we set B_CACHE here, indicating that buffer is 549 * fully valid and thus cacheable. This is true even of NFS 550 * now so we set it generally. This could be set either here 551 * or in biodone() since the I/O is synchronous. We put it 552 * here. 553 */ 554int 555bwrite(struct buf * bp) 556{ 557 int oldflags, s; 558 struct buf *newbp; 559 560 if (bp->b_flags & B_INVAL) { 561 brelse(bp); 562 return (0); 563 } 564 565 oldflags = bp->b_flags; 566 567 if (BUF_REFCNT(bp) == 0) 568 panic("bwrite: buffer is not busy???"); 569 s = splbio(); 570 /* 571 * If a background write is already in progress, delay 572 * writing this block if it is asynchronous. Otherwise 573 * wait for the background write to complete. 574 */ 575 if (bp->b_xflags & BX_BKGRDINPROG) { 576 if (bp->b_flags & B_ASYNC) { 577 splx(s); 578 bdwrite(bp); 579 return (0); 580 } 581 bp->b_xflags |= BX_BKGRDWAIT; 582 tsleep(&bp->b_xflags, PRIBIO, "biord", 0); 583 if (bp->b_xflags & BX_BKGRDINPROG) 584 panic("bwrite: still writing"); 585 } 586 587 /* Mark the buffer clean */ 588 bundirty(bp); 589 590 /* 591 * If this buffer is marked for background writing and we 592 * do not have to wait for it, make a copy and write the 593 * copy so as to leave this buffer ready for further use. 594 */ 595 if ((bp->b_xflags & BX_BKGRDWRITE) && (bp->b_flags & B_ASYNC)) { 596 if (bp->b_iodone != NULL) { 597 printf("bp->b_iodone = %p\n", bp->b_iodone); 598 panic("bwrite: need chained iodone"); 599 } 600 601 /* get a new block */ 602 newbp = geteblk(bp->b_bufsize); 603 604 /* set it to be identical to the old block */ 605 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 606 bgetvp(bp->b_vp, newbp); 607 newbp->b_lblkno = bp->b_lblkno; 608 newbp->b_blkno = bp->b_blkno; 609 newbp->b_offset = bp->b_offset; 610 newbp->b_iodone = vfs_backgroundwritedone; 611 newbp->b_flags |= B_ASYNC; 612 newbp->b_flags &= ~B_INVAL; 613 614 /* move over the dependencies */ 615 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps) 616 (*bioops.io_movedeps)(bp, newbp); 617 618 /* 619 * Initiate write on the copy, release the original to 620 * the B_LOCKED queue so that it cannot go away until 621 * the background write completes. If not locked it could go 622 * away and then be reconstituted while it was being written. 623 * If the reconstituted buffer were written, we could end up 624 * with two background copies being written at the same time. 625 */ 626 bp->b_xflags |= BX_BKGRDINPROG; 627 bp->b_flags |= B_LOCKED; 628 bqrelse(bp); 629 bp = newbp; 630 } 631 632 bp->b_flags &= ~(B_DONE | B_ERROR); 633 bp->b_flags |= B_WRITEINPROG | B_CACHE; 634 bp->b_iocmd = BIO_WRITE; 635 636 bp->b_vp->v_numoutput++; 637 vfs_busy_pages(bp, 1); 638 if (curproc != NULL) 639 curproc->p_stats->p_ru.ru_oublock++; 640 splx(s); 641 if (oldflags & B_ASYNC) 642 BUF_KERNPROC(bp); 643 BUF_STRATEGY(bp); 644 645 if ((oldflags & B_ASYNC) == 0) { 646 int rtval = biowait(bp); 647 brelse(bp); 648 return (rtval); 649 } 650 651 return (0); 652} 653 654/* 655 * Complete a background write started from bwrite. 656 */ 657static void 658vfs_backgroundwritedone(bp) 659 struct buf *bp; 660{ 661 struct buf *origbp; 662 663 /* 664 * Find the original buffer that we are writing. 665 */ 666 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 667 panic("backgroundwritedone: lost buffer"); 668 /* 669 * Process dependencies then return any unfinished ones. 670 */ 671 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 672 (*bioops.io_complete)(bp); 673 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps) 674 (*bioops.io_movedeps)(bp, origbp); 675 /* 676 * Clear the BX_BKGRDINPROG flag in the original buffer 677 * and awaken it if it is waiting for the write to complete. 678 */ 679 origbp->b_xflags &= ~BX_BKGRDINPROG; 680 if (origbp->b_xflags & BX_BKGRDWAIT) { 681 origbp->b_xflags &= ~BX_BKGRDWAIT; 682 wakeup(&origbp->b_xflags); 683 } 684 /* 685 * Clear the B_LOCKED flag and remove it from the locked 686 * queue if it currently resides there. 687 */ 688 origbp->b_flags &= ~B_LOCKED; 689 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 690 bremfree(origbp); 691 bqrelse(origbp); 692 } 693 /* 694 * This buffer is marked B_NOCACHE, so when it is released 695 * by biodone, it will be tossed. We mark it with BIO_READ 696 * to avoid biodone doing a second vwakeup. 697 */ 698 bp->b_flags |= B_NOCACHE; 699 bp->b_iocmd = BIO_READ; 700 bp->b_flags &= ~(B_CACHE | B_DONE); 701 bp->b_iodone = 0; 702 biodone(bp); 703} 704 705/* 706 * Delayed write. (Buffer is marked dirty). Do not bother writing 707 * anything if the buffer is marked invalid. 708 * 709 * Note that since the buffer must be completely valid, we can safely 710 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 711 * biodone() in order to prevent getblk from writing the buffer 712 * out synchronously. 713 */ 714void 715bdwrite(struct buf * bp) 716{ 717 if (BUF_REFCNT(bp) == 0) 718 panic("bdwrite: buffer is not busy"); 719 720 if (bp->b_flags & B_INVAL) { 721 brelse(bp); 722 return; 723 } 724 bdirty(bp); 725 726 /* 727 * Set B_CACHE, indicating that the buffer is fully valid. This is 728 * true even of NFS now. 729 */ 730 bp->b_flags |= B_CACHE; 731 732 /* 733 * This bmap keeps the system from needing to do the bmap later, 734 * perhaps when the system is attempting to do a sync. Since it 735 * is likely that the indirect block -- or whatever other datastructure 736 * that the filesystem needs is still in memory now, it is a good 737 * thing to do this. Note also, that if the pageout daemon is 738 * requesting a sync -- there might not be enough memory to do 739 * the bmap then... So, this is important to do. 740 */ 741 if (bp->b_lblkno == bp->b_blkno) { 742 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 743 } 744 745 /* 746 * Set the *dirty* buffer range based upon the VM system dirty pages. 747 */ 748 vfs_setdirty(bp); 749 750 /* 751 * We need to do this here to satisfy the vnode_pager and the 752 * pageout daemon, so that it thinks that the pages have been 753 * "cleaned". Note that since the pages are in a delayed write 754 * buffer -- the VFS layer "will" see that the pages get written 755 * out on the next sync, or perhaps the cluster will be completed. 756 */ 757 vfs_clean_pages(bp); 758 bqrelse(bp); 759 760 /* 761 * Wakeup the buffer flushing daemon if we have saturated the 762 * buffer cache. 763 */ 764 765 bd_wakeup(hidirtybuffers); 766 767 /* 768 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 769 * due to the softdep code. 770 */ 771} 772 773/* 774 * bdirty: 775 * 776 * Turn buffer into delayed write request. We must clear BIO_READ and 777 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 778 * itself to properly update it in the dirty/clean lists. We mark it 779 * B_DONE to ensure that any asynchronization of the buffer properly 780 * clears B_DONE ( else a panic will occur later ). 781 * 782 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 783 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 784 * should only be called if the buffer is known-good. 785 * 786 * Since the buffer is not on a queue, we do not update the numfreebuffers 787 * count. 788 * 789 * Must be called at splbio(). 790 * The buffer must be on QUEUE_NONE. 791 */ 792void 793bdirty(bp) 794 struct buf *bp; 795{ 796 KASSERT(bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 797 bp->b_flags &= ~(B_RELBUF); 798 bp->b_iocmd = BIO_WRITE; 799 800 if ((bp->b_flags & B_DELWRI) == 0) { 801 bp->b_flags |= B_DONE | B_DELWRI; 802 reassignbuf(bp, bp->b_vp); 803 ++numdirtybuffers; 804 bd_wakeup(hidirtybuffers); 805 } 806} 807 808/* 809 * bundirty: 810 * 811 * Clear B_DELWRI for buffer. 812 * 813 * Since the buffer is not on a queue, we do not update the numfreebuffers 814 * count. 815 * 816 * Must be called at splbio(). 817 * The buffer must be on QUEUE_NONE. 818 */ 819 820void 821bundirty(bp) 822 struct buf *bp; 823{ 824 KASSERT(bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 825 826 if (bp->b_flags & B_DELWRI) { 827 bp->b_flags &= ~B_DELWRI; 828 reassignbuf(bp, bp->b_vp); 829 --numdirtybuffers; 830 numdirtywakeup(); 831 } 832 /* 833 * Since it is now being written, we can clear its deferred write flag. 834 */ 835 bp->b_flags &= ~B_DEFERRED; 836} 837 838/* 839 * bawrite: 840 * 841 * Asynchronous write. Start output on a buffer, but do not wait for 842 * it to complete. The buffer is released when the output completes. 843 * 844 * bwrite() ( or the VOP routine anyway ) is responsible for handling 845 * B_INVAL buffers. Not us. 846 */ 847void 848bawrite(struct buf * bp) 849{ 850 bp->b_flags |= B_ASYNC; 851 (void) BUF_WRITE(bp); 852} 853 854/* 855 * bowrite: 856 * 857 * Ordered write. Start output on a buffer, and flag it so that the 858 * device will write it in the order it was queued. The buffer is 859 * released when the output completes. bwrite() ( or the VOP routine 860 * anyway ) is responsible for handling B_INVAL buffers. 861 */ 862int 863bowrite(struct buf * bp) 864{ 865 bp->b_flags |= B_ORDERED | B_ASYNC; 866 return (BUF_WRITE(bp)); 867} 868 869/* 870 * bwillwrite: 871 * 872 * Called prior to the locking of any vnodes when we are expecting to 873 * write. We do not want to starve the buffer cache with too many 874 * dirty buffers so we block here. By blocking prior to the locking 875 * of any vnodes we attempt to avoid the situation where a locked vnode 876 * prevents the various system daemons from flushing related buffers. 877 */ 878 879void 880bwillwrite(void) 881{ 882 int slop = hidirtybuffers / 10; 883 884 if (numdirtybuffers > hidirtybuffers + slop) { 885 int s; 886 887 s = splbio(); 888 while (numdirtybuffers > hidirtybuffers) { 889 bd_wakeup(hidirtybuffers); 890 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 891 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0); 892 } 893 splx(s); 894 } 895} 896 897/* 898 * brelse: 899 * 900 * Release a busy buffer and, if requested, free its resources. The 901 * buffer will be stashed in the appropriate bufqueue[] allowing it 902 * to be accessed later as a cache entity or reused for other purposes. 903 */ 904void 905brelse(struct buf * bp) 906{ 907 int s; 908 909 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 910 911 s = splbio(); 912 913 if (bp->b_flags & B_LOCKED) 914 bp->b_flags &= ~B_ERROR; 915 916 if (bp->b_iocmd == BIO_WRITE && 917 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) { 918 /* 919 * Failed write, redirty. Must clear B_ERROR to prevent 920 * pages from being scrapped. If B_INVAL is set then 921 * this case is not run and the next case is run to 922 * destroy the buffer. B_INVAL can occur if the buffer 923 * is outside the range supported by the underlying device. 924 */ 925 bp->b_flags &= ~B_ERROR; 926 bdirty(bp); 927 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) || 928 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 929 /* 930 * Either a failed I/O or we were asked to free or not 931 * cache the buffer. 932 */ 933 bp->b_flags |= B_INVAL; 934 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 935 (*bioops.io_deallocate)(bp); 936 if (bp->b_flags & B_DELWRI) { 937 --numdirtybuffers; 938 numdirtywakeup(); 939 } 940 bp->b_flags &= ~(B_DELWRI | B_CACHE); 941 if ((bp->b_flags & B_VMIO) == 0) { 942 if (bp->b_bufsize) 943 allocbuf(bp, 0); 944 if (bp->b_vp) 945 brelvp(bp); 946 } 947 } 948 949 /* 950 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 951 * is called with B_DELWRI set, the underlying pages may wind up 952 * getting freed causing a previous write (bdwrite()) to get 'lost' 953 * because pages associated with a B_DELWRI bp are marked clean. 954 * 955 * We still allow the B_INVAL case to call vfs_vmio_release(), even 956 * if B_DELWRI is set. 957 */ 958 959 if (bp->b_flags & B_DELWRI) 960 bp->b_flags &= ~B_RELBUF; 961 962 /* 963 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 964 * constituted, not even NFS buffers now. Two flags effect this. If 965 * B_INVAL, the struct buf is invalidated but the VM object is kept 966 * around ( i.e. so it is trivial to reconstitute the buffer later ). 967 * 968 * If B_ERROR or B_NOCACHE is set, pages in the VM object will be 969 * invalidated. B_ERROR cannot be set for a failed write unless the 970 * buffer is also B_INVAL because it hits the re-dirtying code above. 971 * 972 * Normally we can do this whether a buffer is B_DELWRI or not. If 973 * the buffer is an NFS buffer, it is tracking piecemeal writes or 974 * the commit state and we cannot afford to lose the buffer. If the 975 * buffer has a background write in progress, we need to keep it 976 * around to prevent it from being reconstituted and starting a second 977 * background write. 978 */ 979 if ((bp->b_flags & B_VMIO) 980 && !(bp->b_vp->v_tag == VT_NFS && 981 !vn_isdisk(bp->b_vp, NULL) && 982 (bp->b_flags & B_DELWRI) && 983 (bp->b_xflags & BX_BKGRDINPROG)) 984 ) { 985 986 int i, j, resid; 987 vm_page_t m; 988 off_t foff; 989 vm_pindex_t poff; 990 vm_object_t obj; 991 struct vnode *vp; 992 993 vp = bp->b_vp; 994 995 /* 996 * Get the base offset and length of the buffer. Note that 997 * for block sizes that are less then PAGE_SIZE, the b_data 998 * base of the buffer does not represent exactly b_offset and 999 * neither b_offset nor b_size are necessarily page aligned. 1000 * Instead, the starting position of b_offset is: 1001 * 1002 * b_data + (b_offset & PAGE_MASK) 1003 * 1004 * block sizes less then DEV_BSIZE (usually 512) are not 1005 * supported due to the page granularity bits (m->valid, 1006 * m->dirty, etc...). 1007 * 1008 * See man buf(9) for more information 1009 */ 1010 1011 resid = bp->b_bufsize; 1012 foff = bp->b_offset; 1013 1014 for (i = 0; i < bp->b_npages; i++) { 1015 m = bp->b_pages[i]; 1016 vm_page_flag_clear(m, PG_ZERO); 1017 if (m == bogus_page) { 1018 1019 obj = (vm_object_t) vp->v_object; 1020 poff = OFF_TO_IDX(bp->b_offset); 1021 1022 for (j = i; j < bp->b_npages; j++) { 1023 m = bp->b_pages[j]; 1024 if (m == bogus_page) { 1025 m = vm_page_lookup(obj, poff + j); 1026 if (!m) { 1027 panic("brelse: page missing\n"); 1028 } 1029 bp->b_pages[j] = m; 1030 } 1031 } 1032 1033 if ((bp->b_flags & B_INVAL) == 0) { 1034 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1035 } 1036 } 1037 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 1038 int poffset = foff & PAGE_MASK; 1039 int presid = resid > (PAGE_SIZE - poffset) ? 1040 (PAGE_SIZE - poffset) : resid; 1041 1042 KASSERT(presid >= 0, ("brelse: extra page")); 1043 vm_page_set_invalid(m, poffset, presid); 1044 } 1045 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1046 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 1047 } 1048 1049 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1050 vfs_vmio_release(bp); 1051 1052 } else if (bp->b_flags & B_VMIO) { 1053 1054 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1055 vfs_vmio_release(bp); 1056 1057 } 1058 1059 if (bp->b_qindex != QUEUE_NONE) 1060 panic("brelse: free buffer onto another queue???"); 1061 if (BUF_REFCNT(bp) > 1) { 1062 /* Temporary panic to verify exclusive locking */ 1063 /* This panic goes away when we allow shared refs */ 1064 panic("brelse: multiple refs"); 1065 /* do not release to free list */ 1066 BUF_UNLOCK(bp); 1067 splx(s); 1068 return; 1069 } 1070 1071 /* enqueue */ 1072 1073 /* buffers with no memory */ 1074 if (bp->b_bufsize == 0) { 1075 bp->b_flags |= B_INVAL; 1076 bp->b_xflags &= ~BX_BKGRDWRITE; 1077 if (bp->b_xflags & BX_BKGRDINPROG) 1078 panic("losing buffer 1"); 1079 if (bp->b_kvasize) { 1080 bp->b_qindex = QUEUE_EMPTYKVA; 1081 } else { 1082 bp->b_qindex = QUEUE_EMPTY; 1083 } 1084 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1085 LIST_REMOVE(bp, b_hash); 1086 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1087 bp->b_dev = NODEV; 1088 /* buffers with junk contents */ 1089 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) { 1090 bp->b_flags |= B_INVAL; 1091 bp->b_xflags &= ~BX_BKGRDWRITE; 1092 if (bp->b_xflags & BX_BKGRDINPROG) 1093 panic("losing buffer 2"); 1094 bp->b_qindex = QUEUE_CLEAN; 1095 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1096 LIST_REMOVE(bp, b_hash); 1097 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1098 bp->b_dev = NODEV; 1099 1100 /* buffers that are locked */ 1101 } else if (bp->b_flags & B_LOCKED) { 1102 bp->b_qindex = QUEUE_LOCKED; 1103 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1104 1105 /* remaining buffers */ 1106 } else { 1107 switch(bp->b_flags & (B_DELWRI|B_AGE)) { 1108 case B_DELWRI | B_AGE: 1109 bp->b_qindex = QUEUE_DIRTY; 1110 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1111 break; 1112 case B_DELWRI: 1113 bp->b_qindex = QUEUE_DIRTY; 1114 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1115 break; 1116 case B_AGE: 1117 bp->b_qindex = QUEUE_CLEAN; 1118 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1119 break; 1120 default: 1121 bp->b_qindex = QUEUE_CLEAN; 1122 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1123 break; 1124 } 1125 } 1126 1127 /* 1128 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1129 * on the correct queue. 1130 */ 1131 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) { 1132 bp->b_flags &= ~B_DELWRI; 1133 --numdirtybuffers; 1134 numdirtywakeup(); 1135 } 1136 1137 runningbufspace -= bp->b_bufsize; 1138 1139 /* 1140 * Fixup numfreebuffers count. The bp is on an appropriate queue 1141 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1142 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1143 * if B_INVAL is set ). 1144 */ 1145 1146 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1147 bufcountwakeup(); 1148 1149 /* 1150 * Something we can maybe free. 1151 */ 1152 1153 if (bp->b_bufsize || bp->b_kvasize) 1154 bufspacewakeup(); 1155 1156 /* unlock */ 1157 BUF_UNLOCK(bp); 1158 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1159 splx(s); 1160} 1161 1162/* 1163 * Release a buffer back to the appropriate queue but do not try to free 1164 * it. 1165 * 1166 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1167 * biodone() to requeue an async I/O on completion. It is also used when 1168 * known good buffers need to be requeued but we think we may need the data 1169 * again soon. 1170 */ 1171void 1172bqrelse(struct buf * bp) 1173{ 1174 int s; 1175 1176 s = splbio(); 1177 1178 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1179 1180 if (bp->b_qindex != QUEUE_NONE) 1181 panic("bqrelse: free buffer onto another queue???"); 1182 if (BUF_REFCNT(bp) > 1) { 1183 /* do not release to free list */ 1184 panic("bqrelse: multiple refs"); 1185 BUF_UNLOCK(bp); 1186 splx(s); 1187 return; 1188 } 1189 if (bp->b_flags & B_LOCKED) { 1190 bp->b_flags &= ~B_ERROR; 1191 bp->b_qindex = QUEUE_LOCKED; 1192 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1193 /* buffers with stale but valid contents */ 1194 } else if (bp->b_flags & B_DELWRI) { 1195 bp->b_qindex = QUEUE_DIRTY; 1196 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1197 } else { 1198 bp->b_qindex = QUEUE_CLEAN; 1199 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1200 } 1201 1202 runningbufspace -= bp->b_bufsize; 1203 1204 if ((bp->b_flags & B_LOCKED) == 0 && 1205 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1206 bufcountwakeup(); 1207 } 1208 1209 /* 1210 * Something we can maybe wakeup 1211 */ 1212 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1213 bufspacewakeup(); 1214 1215 /* unlock */ 1216 BUF_UNLOCK(bp); 1217 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1218 splx(s); 1219} 1220 1221static void 1222vfs_vmio_release(bp) 1223 struct buf *bp; 1224{ 1225 int i, s; 1226 vm_page_t m; 1227 1228 s = splvm(); 1229 for (i = 0; i < bp->b_npages; i++) { 1230 m = bp->b_pages[i]; 1231 bp->b_pages[i] = NULL; 1232 /* 1233 * In order to keep page LRU ordering consistent, put 1234 * everything on the inactive queue. 1235 */ 1236 vm_page_unwire(m, 0); 1237 /* 1238 * We don't mess with busy pages, it is 1239 * the responsibility of the process that 1240 * busied the pages to deal with them. 1241 */ 1242 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1243 continue; 1244 1245 if (m->wire_count == 0) { 1246 vm_page_flag_clear(m, PG_ZERO); 1247 /* 1248 * Might as well free the page if we can and it has 1249 * no valid data. 1250 */ 1251 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { 1252 vm_page_busy(m); 1253 vm_page_protect(m, VM_PROT_NONE); 1254 vm_page_free(m); 1255 } 1256 } 1257 } 1258 runningbufspace -= bp->b_bufsize; 1259 splx(s); 1260 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1261 if (bp->b_bufsize) 1262 bufspacewakeup(); 1263 bp->b_npages = 0; 1264 bp->b_bufsize = 0; 1265 bp->b_flags &= ~B_VMIO; 1266 if (bp->b_vp) 1267 brelvp(bp); 1268} 1269 1270/* 1271 * Check to see if a block is currently memory resident. 1272 */ 1273struct buf * 1274gbincore(struct vnode * vp, daddr_t blkno) 1275{ 1276 struct buf *bp; 1277 struct bufhashhdr *bh; 1278 1279 bh = bufhash(vp, blkno); 1280 1281 /* Search hash chain */ 1282 LIST_FOREACH(bp, bh, b_hash) { 1283 /* hit */ 1284 if (bp->b_vp == vp && bp->b_lblkno == blkno && 1285 (bp->b_flags & B_INVAL) == 0) { 1286 break; 1287 } 1288 } 1289 return (bp); 1290} 1291 1292/* 1293 * vfs_bio_awrite: 1294 * 1295 * Implement clustered async writes for clearing out B_DELWRI buffers. 1296 * This is much better then the old way of writing only one buffer at 1297 * a time. Note that we may not be presented with the buffers in the 1298 * correct order, so we search for the cluster in both directions. 1299 */ 1300int 1301vfs_bio_awrite(struct buf * bp) 1302{ 1303 int i; 1304 int j; 1305 daddr_t lblkno = bp->b_lblkno; 1306 struct vnode *vp = bp->b_vp; 1307 int s; 1308 int ncl; 1309 struct buf *bpa; 1310 int nwritten; 1311 int size; 1312 int maxcl; 1313 1314 s = splbio(); 1315 /* 1316 * right now we support clustered writing only to regular files. If 1317 * we find a clusterable block we could be in the middle of a cluster 1318 * rather then at the beginning. 1319 */ 1320 if ((vp->v_type == VREG) && 1321 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1322 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1323 1324 size = vp->v_mount->mnt_stat.f_iosize; 1325 maxcl = MAXPHYS / size; 1326 1327 for (i = 1; i < maxcl; i++) { 1328 if ((bpa = gbincore(vp, lblkno + i)) && 1329 BUF_REFCNT(bpa) == 0 && 1330 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1331 (B_DELWRI | B_CLUSTEROK)) && 1332 (bpa->b_bufsize == size)) { 1333 if ((bpa->b_blkno == bpa->b_lblkno) || 1334 (bpa->b_blkno != 1335 bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 1336 break; 1337 } else { 1338 break; 1339 } 1340 } 1341 for (j = 1; i + j <= maxcl && j <= lblkno; j++) { 1342 if ((bpa = gbincore(vp, lblkno - j)) && 1343 BUF_REFCNT(bpa) == 0 && 1344 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1345 (B_DELWRI | B_CLUSTEROK)) && 1346 (bpa->b_bufsize == size)) { 1347 if ((bpa->b_blkno == bpa->b_lblkno) || 1348 (bpa->b_blkno != 1349 bp->b_blkno - ((j * size) >> DEV_BSHIFT))) 1350 break; 1351 } else { 1352 break; 1353 } 1354 } 1355 --j; 1356 ncl = i + j; 1357 /* 1358 * this is a possible cluster write 1359 */ 1360 if (ncl != 1) { 1361 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1362 splx(s); 1363 return nwritten; 1364 } 1365 } 1366 1367 BUF_LOCK(bp, LK_EXCLUSIVE); 1368 bremfree(bp); 1369 bp->b_flags |= B_ASYNC; 1370 1371 splx(s); 1372 /* 1373 * default (old) behavior, writing out only one block 1374 * 1375 * XXX returns b_bufsize instead of b_bcount for nwritten? 1376 */ 1377 nwritten = bp->b_bufsize; 1378 (void) BUF_WRITE(bp); 1379 1380 return nwritten; 1381} 1382 1383/* 1384 * getnewbuf: 1385 * 1386 * Find and initialize a new buffer header, freeing up existing buffers 1387 * in the bufqueues as necessary. The new buffer is returned locked. 1388 * 1389 * Important: B_INVAL is not set. If the caller wishes to throw the 1390 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1391 * 1392 * We block if: 1393 * We have insufficient buffer headers 1394 * We have insufficient buffer space 1395 * buffer_map is too fragmented ( space reservation fails ) 1396 * If we have to flush dirty buffers ( but we try to avoid this ) 1397 * 1398 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1399 * Instead we ask the buf daemon to do it for us. We attempt to 1400 * avoid piecemeal wakeups of the pageout daemon. 1401 */ 1402 1403static struct buf * 1404getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1405{ 1406 struct buf *bp; 1407 struct buf *nbp; 1408 int defrag = 0; 1409 int nqindex; 1410 int isspecial; 1411 static int flushingbufs; 1412 1413 if (curproc && (curproc->p_flag & P_BUFEXHAUST) == 0) 1414 isspecial = 0; 1415 else 1416 isspecial = 1; 1417 1418 ++getnewbufcalls; 1419 --getnewbufrestarts; 1420restart: 1421 ++getnewbufrestarts; 1422 1423 /* 1424 * Setup for scan. If we do not have enough free buffers, 1425 * we setup a degenerate case that immediately fails. Note 1426 * that if we are specially marked process, we are allowed to 1427 * dip into our reserves. 1428 * 1429 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1430 * 1431 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1432 * However, there are a number of cases (defragging, reusing, ...) 1433 * where we cannot backup. 1434 */ 1435 1436 if (isspecial == 0 && numfreebuffers < lofreebuffers) { 1437 /* 1438 * This will cause an immediate failure 1439 */ 1440 nqindex = QUEUE_CLEAN; 1441 nbp = NULL; 1442 } else { 1443 /* 1444 * Locate a buffer which already has KVA assigned. First 1445 * try EMPTYKVA buffers. 1446 */ 1447 nqindex = QUEUE_EMPTYKVA; 1448 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1449 1450 if (nbp == NULL) { 1451 /* 1452 * If no EMPTYKVA buffers and we are either 1453 * defragging or reusing, locate a CLEAN buffer 1454 * to free or reuse. If bufspace useage is low 1455 * skip this step so we can allocate a new buffer. 1456 */ 1457 if (defrag || bufspace >= lobufspace) { 1458 nqindex = QUEUE_CLEAN; 1459 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1460 } 1461 1462 /* 1463 * Nada. If we are allowed to allocate an EMPTY 1464 * buffer, go get one. 1465 */ 1466 if (nbp == NULL && defrag == 0 && 1467 (isspecial || bufspace < hibufspace)) { 1468 nqindex = QUEUE_EMPTY; 1469 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1470 } 1471 } 1472 } 1473 1474 /* 1475 * Run scan, possibly freeing data and/or kva mappings on the fly 1476 * depending. 1477 */ 1478 1479 while ((bp = nbp) != NULL) { 1480 int qindex = nqindex; 1481 1482 /* 1483 * Calculate next bp ( we can only use it if we do not block 1484 * or do other fancy things ). 1485 */ 1486 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1487 switch(qindex) { 1488 case QUEUE_EMPTY: 1489 nqindex = QUEUE_EMPTYKVA; 1490 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1491 break; 1492 /* fall through */ 1493 case QUEUE_EMPTYKVA: 1494 nqindex = QUEUE_CLEAN; 1495 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1496 break; 1497 /* fall through */ 1498 case QUEUE_CLEAN: 1499 /* 1500 * nbp is NULL. 1501 */ 1502 break; 1503 } 1504 } 1505 1506 /* 1507 * Sanity Checks 1508 */ 1509 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1510 1511 /* 1512 * Note: we no longer distinguish between VMIO and non-VMIO 1513 * buffers. 1514 */ 1515 1516 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1517 1518 /* 1519 * If we are defragging then we need a buffer with 1520 * b_kvasize != 0. XXX this situation should no longer 1521 * occur, if defrag is non-zero the buffer's b_kvasize 1522 * should also be non-zero at this point. XXX 1523 */ 1524 if (defrag && bp->b_kvasize == 0) { 1525 printf("Warning: defrag empty buffer %p\n", bp); 1526 continue; 1527 } 1528 1529 /* 1530 * Start freeing the bp. This is somewhat involved. nbp 1531 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1532 */ 1533 1534 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1535 panic("getnewbuf: locked buf"); 1536 bremfree(bp); 1537 1538 if (qindex == QUEUE_CLEAN) { 1539 if (bp->b_flags & B_VMIO) { 1540 bp->b_flags &= ~B_ASYNC; 1541 vfs_vmio_release(bp); 1542 } 1543 if (bp->b_vp) 1544 brelvp(bp); 1545 } 1546 1547 /* 1548 * NOTE: nbp is now entirely invalid. We can only restart 1549 * the scan from this point on. 1550 * 1551 * Get the rest of the buffer freed up. b_kva* is still 1552 * valid after this operation. 1553 */ 1554 1555 if (bp->b_rcred != NOCRED) { 1556 crfree(bp->b_rcred); 1557 bp->b_rcred = NOCRED; 1558 } 1559 if (bp->b_wcred != NOCRED) { 1560 crfree(bp->b_wcred); 1561 bp->b_wcred = NOCRED; 1562 } 1563 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 1564 (*bioops.io_deallocate)(bp); 1565 if (bp->b_xflags & BX_BKGRDINPROG) 1566 panic("losing buffer 3"); 1567 LIST_REMOVE(bp, b_hash); 1568 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1569 1570 if (bp->b_bufsize) 1571 allocbuf(bp, 0); 1572 1573 bp->b_flags = 0; 1574 bp->b_xflags = 0; 1575 bp->b_dev = NODEV; 1576 bp->b_vp = NULL; 1577 bp->b_blkno = bp->b_lblkno = 0; 1578 bp->b_offset = NOOFFSET; 1579 bp->b_iodone = 0; 1580 bp->b_error = 0; 1581 bp->b_resid = 0; 1582 bp->b_bcount = 0; 1583 bp->b_npages = 0; 1584 bp->b_dirtyoff = bp->b_dirtyend = 0; 1585 1586 LIST_INIT(&bp->b_dep); 1587 1588 /* 1589 * If we are defragging then free the buffer. 1590 */ 1591 if (defrag) { 1592 bp->b_flags |= B_INVAL; 1593 bfreekva(bp); 1594 brelse(bp); 1595 defrag = 0; 1596 goto restart; 1597 } 1598 1599 /* 1600 * If we are a normal process then deal with bufspace 1601 * hysteresis. A normal process tries to keep bufspace 1602 * between lobufspace and hibufspace. Note: if we encounter 1603 * a buffer with b_kvasize == 0 then it means we started 1604 * our scan on the EMPTY list and should allocate a new 1605 * buffer. 1606 */ 1607 if (isspecial == 0) { 1608 if (bufspace > hibufspace) 1609 flushingbufs = 1; 1610 if (flushingbufs && bp->b_kvasize != 0) { 1611 bp->b_flags |= B_INVAL; 1612 bfreekva(bp); 1613 brelse(bp); 1614 goto restart; 1615 } 1616 if (bufspace < lobufspace) 1617 flushingbufs = 0; 1618 } 1619 break; 1620 } 1621 1622 /* 1623 * If we exhausted our list, sleep as appropriate. We may have to 1624 * wakeup various daemons and write out some dirty buffers. 1625 * 1626 * Generally we are sleeping due to insufficient buffer space. 1627 */ 1628 1629 if (bp == NULL) { 1630 int flags; 1631 char *waitmsg; 1632 1633 if (defrag) { 1634 flags = VFS_BIO_NEED_BUFSPACE; 1635 waitmsg = "nbufkv"; 1636 } else if (bufspace >= hibufspace) { 1637 waitmsg = "nbufbs"; 1638 flags = VFS_BIO_NEED_BUFSPACE; 1639 } else { 1640 waitmsg = "newbuf"; 1641 flags = VFS_BIO_NEED_ANY; 1642 } 1643 1644 bd_speedup(); /* heeeelp */ 1645 1646 needsbuffer |= flags; 1647 while (needsbuffer & flags) { 1648 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, 1649 waitmsg, slptimeo)) 1650 return (NULL); 1651 } 1652 } else { 1653 /* 1654 * We finally have a valid bp. We aren't quite out of the 1655 * woods, we still have to reserve kva space. In order 1656 * to keep fragmentation sane we only allocate kva in 1657 * BKVASIZE chunks. 1658 */ 1659 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1660 1661 if (maxsize != bp->b_kvasize) { 1662 vm_offset_t addr = 0; 1663 1664 bfreekva(bp); 1665 1666 if (vm_map_findspace(buffer_map, 1667 vm_map_min(buffer_map), maxsize, &addr)) { 1668 /* 1669 * Uh oh. Buffer map is to fragmented. We 1670 * must defragment the map. 1671 */ 1672 ++bufdefragcnt; 1673 defrag = 1; 1674 bp->b_flags |= B_INVAL; 1675 brelse(bp); 1676 goto restart; 1677 } 1678 if (addr) { 1679 vm_map_insert(buffer_map, NULL, 0, 1680 addr, addr + maxsize, 1681 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1682 1683 bp->b_kvabase = (caddr_t) addr; 1684 bp->b_kvasize = maxsize; 1685 bufspace += bp->b_kvasize; 1686 ++bufreusecnt; 1687 } 1688 } 1689 bp->b_data = bp->b_kvabase; 1690 } 1691 return(bp); 1692} 1693 1694/* 1695 * waitfreebuffers: 1696 * 1697 * Wait for sufficient free buffers. Only called from normal processes. 1698 */ 1699 1700static void 1701waitfreebuffers(int slpflag, int slptimeo) 1702{ 1703 while (numfreebuffers < hifreebuffers) { 1704 if (numfreebuffers >= hifreebuffers) 1705 break; 1706 needsbuffer |= VFS_BIO_NEED_FREE; 1707 if (tsleep(&needsbuffer, (PRIBIO + 4)|slpflag, "biofre", slptimeo)) 1708 break; 1709 } 1710} 1711 1712/* 1713 * buf_daemon: 1714 * 1715 * buffer flushing daemon. Buffers are normally flushed by the 1716 * update daemon but if it cannot keep up this process starts to 1717 * take the load in an attempt to prevent getnewbuf() from blocking. 1718 */ 1719 1720static struct proc *bufdaemonproc; 1721static int bd_interval; 1722static int bd_flushto; 1723static int bd_flushinc; 1724 1725static struct kproc_desc buf_kp = { 1726 "bufdaemon", 1727 buf_daemon, 1728 &bufdaemonproc 1729}; 1730SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1731 1732static void 1733buf_daemon() 1734{ 1735 int s; 1736 1737 /* 1738 * This process needs to be suspended prior to shutdown sync. 1739 */ 1740 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, bufdaemonproc, 1741 SHUTDOWN_PRI_LAST); 1742 1743 /* 1744 * This process is allowed to take the buffer cache to the limit 1745 */ 1746 curproc->p_flag |= P_BUFEXHAUST; 1747 s = splbio(); 1748 1749 bd_interval = 5 * hz; /* dynamically adjusted */ 1750 bd_flushto = hidirtybuffers; /* dynamically adjusted */ 1751 bd_flushinc = 1; 1752 1753 for (;;) { 1754 kproc_suspend_loop(bufdaemonproc); 1755 1756 bd_request = 0; 1757 1758 /* 1759 * Do the flush. Limit the number of buffers we flush in one 1760 * go. The failure condition occurs when processes are writing 1761 * buffers faster then we can dispose of them. In this case 1762 * we may be flushing so often that the previous set of flushes 1763 * have not had time to complete, causing us to run out of 1764 * physical buffers and block. 1765 */ 1766 { 1767 int runcount = maxbdrun; 1768 1769 while (numdirtybuffers > bd_flushto && runcount) { 1770 --runcount; 1771 if (flushbufqueues() == 0) 1772 break; 1773 } 1774 } 1775 1776 if (bd_request || 1777 tsleep(&bd_request, PVM, "psleep", bd_interval) == 0) { 1778 /* 1779 * Another request is pending or we were woken up 1780 * without timing out. Flush more. 1781 */ 1782 --bd_flushto; 1783 if (bd_flushto >= numdirtybuffers - 5) { 1784 bd_flushto = numdirtybuffers - 10; 1785 bd_flushinc = 1; 1786 } 1787 if (bd_flushto < 2) 1788 bd_flushto = 2; 1789 } else { 1790 /* 1791 * We slept and timed out, we can slow down. 1792 */ 1793 bd_flushto += bd_flushinc; 1794 if (bd_flushto > hidirtybuffers) 1795 bd_flushto = hidirtybuffers; 1796 ++bd_flushinc; 1797 if (bd_flushinc > hidirtybuffers / 20 + 1) 1798 bd_flushinc = hidirtybuffers / 20 + 1; 1799 } 1800 1801 /* 1802 * Set the interval on a linear scale based on hidirtybuffers 1803 * with a maximum frequency of 1/10 second. 1804 */ 1805 bd_interval = bd_flushto * 5 * hz / hidirtybuffers; 1806 if (bd_interval < hz / 10) 1807 bd_interval = hz / 10; 1808 } 1809} 1810 1811/* 1812 * flushbufqueues: 1813 * 1814 * Try to flush a buffer in the dirty queue. We must be careful to 1815 * free up B_INVAL buffers instead of write them, which NFS is 1816 * particularly sensitive to. 1817 */ 1818 1819static int 1820flushbufqueues(void) 1821{ 1822 struct buf *bp; 1823 int r = 0; 1824 1825 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1826 1827 while (bp) { 1828 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); 1829 if ((bp->b_flags & B_DELWRI) != 0 && 1830 (bp->b_xflags & BX_BKGRDINPROG) == 0) { 1831 if (bp->b_flags & B_INVAL) { 1832 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1833 panic("flushbufqueues: locked buf"); 1834 bremfree(bp); 1835 brelse(bp); 1836 ++r; 1837 break; 1838 } 1839 if (LIST_FIRST(&bp->b_dep) != NULL && 1840 bioops.io_countdeps && 1841 (bp->b_flags & B_DEFERRED) == 0 && 1842 (*bioops.io_countdeps)(bp, 0)) { 1843 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], 1844 bp, b_freelist); 1845 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], 1846 bp, b_freelist); 1847 bp->b_flags |= B_DEFERRED; 1848 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 1849 continue; 1850 } 1851 vfs_bio_awrite(bp); 1852 ++r; 1853 break; 1854 } 1855 bp = TAILQ_NEXT(bp, b_freelist); 1856 } 1857 return (r); 1858} 1859 1860/* 1861 * Check to see if a block is currently memory resident. 1862 */ 1863struct buf * 1864incore(struct vnode * vp, daddr_t blkno) 1865{ 1866 struct buf *bp; 1867 1868 int s = splbio(); 1869 bp = gbincore(vp, blkno); 1870 splx(s); 1871 return (bp); 1872} 1873 1874/* 1875 * Returns true if no I/O is needed to access the 1876 * associated VM object. This is like incore except 1877 * it also hunts around in the VM system for the data. 1878 */ 1879 1880int 1881inmem(struct vnode * vp, daddr_t blkno) 1882{ 1883 vm_object_t obj; 1884 vm_offset_t toff, tinc, size; 1885 vm_page_t m; 1886 vm_ooffset_t off; 1887 1888 if (incore(vp, blkno)) 1889 return 1; 1890 if (vp->v_mount == NULL) 1891 return 0; 1892 if ((vp->v_object == NULL) || (vp->v_flag & VOBJBUF) == 0) 1893 return 0; 1894 1895 obj = vp->v_object; 1896 size = PAGE_SIZE; 1897 if (size > vp->v_mount->mnt_stat.f_iosize) 1898 size = vp->v_mount->mnt_stat.f_iosize; 1899 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 1900 1901 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1902 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 1903 if (!m) 1904 return 0; 1905 tinc = size; 1906 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 1907 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 1908 if (vm_page_is_valid(m, 1909 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 1910 return 0; 1911 } 1912 return 1; 1913} 1914 1915/* 1916 * vfs_setdirty: 1917 * 1918 * Sets the dirty range for a buffer based on the status of the dirty 1919 * bits in the pages comprising the buffer. 1920 * 1921 * The range is limited to the size of the buffer. 1922 * 1923 * This routine is primarily used by NFS, but is generalized for the 1924 * B_VMIO case. 1925 */ 1926static void 1927vfs_setdirty(struct buf *bp) 1928{ 1929 int i; 1930 vm_object_t object; 1931 1932 /* 1933 * Degenerate case - empty buffer 1934 */ 1935 1936 if (bp->b_bufsize == 0) 1937 return; 1938 1939 /* 1940 * We qualify the scan for modified pages on whether the 1941 * object has been flushed yet. The OBJ_WRITEABLE flag 1942 * is not cleared simply by protecting pages off. 1943 */ 1944 1945 if ((bp->b_flags & B_VMIO) == 0) 1946 return; 1947 1948 object = bp->b_pages[0]->object; 1949 1950 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 1951 printf("Warning: object %p writeable but not mightbedirty\n", object); 1952 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 1953 printf("Warning: object %p mightbedirty but not writeable\n", object); 1954 1955 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 1956 vm_offset_t boffset; 1957 vm_offset_t eoffset; 1958 1959 /* 1960 * test the pages to see if they have been modified directly 1961 * by users through the VM system. 1962 */ 1963 for (i = 0; i < bp->b_npages; i++) { 1964 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 1965 vm_page_test_dirty(bp->b_pages[i]); 1966 } 1967 1968 /* 1969 * Calculate the encompassing dirty range, boffset and eoffset, 1970 * (eoffset - boffset) bytes. 1971 */ 1972 1973 for (i = 0; i < bp->b_npages; i++) { 1974 if (bp->b_pages[i]->dirty) 1975 break; 1976 } 1977 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 1978 1979 for (i = bp->b_npages - 1; i >= 0; --i) { 1980 if (bp->b_pages[i]->dirty) { 1981 break; 1982 } 1983 } 1984 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 1985 1986 /* 1987 * Fit it to the buffer. 1988 */ 1989 1990 if (eoffset > bp->b_bcount) 1991 eoffset = bp->b_bcount; 1992 1993 /* 1994 * If we have a good dirty range, merge with the existing 1995 * dirty range. 1996 */ 1997 1998 if (boffset < eoffset) { 1999 if (bp->b_dirtyoff > boffset) 2000 bp->b_dirtyoff = boffset; 2001 if (bp->b_dirtyend < eoffset) 2002 bp->b_dirtyend = eoffset; 2003 } 2004 } 2005} 2006 2007/* 2008 * getblk: 2009 * 2010 * Get a block given a specified block and offset into a file/device. 2011 * The buffers B_DONE bit will be cleared on return, making it almost 2012 * ready for an I/O initiation. B_INVAL may or may not be set on 2013 * return. The caller should clear B_INVAL prior to initiating a 2014 * READ. 2015 * 2016 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2017 * an existing buffer. 2018 * 2019 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2020 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2021 * and then cleared based on the backing VM. If the previous buffer is 2022 * non-0-sized but invalid, B_CACHE will be cleared. 2023 * 2024 * If getblk() must create a new buffer, the new buffer is returned with 2025 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2026 * case it is returned with B_INVAL clear and B_CACHE set based on the 2027 * backing VM. 2028 * 2029 * getblk() also forces a VOP_BWRITE() for any B_DELWRI buffer whos 2030 * B_CACHE bit is clear. 2031 * 2032 * What this means, basically, is that the caller should use B_CACHE to 2033 * determine whether the buffer is fully valid or not and should clear 2034 * B_INVAL prior to issuing a read. If the caller intends to validate 2035 * the buffer by loading its data area with something, the caller needs 2036 * to clear B_INVAL. If the caller does this without issuing an I/O, 2037 * the caller should set B_CACHE ( as an optimization ), else the caller 2038 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2039 * a write attempt or if it was a successfull read. If the caller 2040 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR 2041 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2042 */ 2043struct buf * 2044getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2045{ 2046 struct buf *bp; 2047 int s; 2048 struct bufhashhdr *bh; 2049 2050 if (size > MAXBSIZE) 2051 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2052 2053 s = splbio(); 2054loop: 2055 /* 2056 * Block if we are low on buffers. Certain processes are allowed 2057 * to completely exhaust the buffer cache. 2058 * 2059 * If this check ever becomes a bottleneck it may be better to 2060 * move it into the else, when gbincore() fails. At the moment 2061 * it isn't a problem. 2062 */ 2063 if (!curproc || (curproc->p_flag & P_BUFEXHAUST)) { 2064 if (numfreebuffers == 0) { 2065 if (!curproc) 2066 return NULL; 2067 needsbuffer |= VFS_BIO_NEED_ANY; 2068 tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, "newbuf", 2069 slptimeo); 2070 } 2071 } else if (numfreebuffers < lofreebuffers) { 2072 waitfreebuffers(slpflag, slptimeo); 2073 } 2074 2075 if ((bp = gbincore(vp, blkno))) { 2076 /* 2077 * Buffer is in-core. If the buffer is not busy, it must 2078 * be on a queue. 2079 */ 2080 2081 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2082 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, 2083 "getblk", slpflag, slptimeo) == ENOLCK) 2084 goto loop; 2085 splx(s); 2086 return (struct buf *) NULL; 2087 } 2088 2089 /* 2090 * The buffer is locked. B_CACHE is cleared if the buffer is 2091 * invalid. Ohterwise, for a non-VMIO buffer, B_CACHE is set 2092 * and for a VMIO buffer B_CACHE is adjusted according to the 2093 * backing VM cache. 2094 */ 2095 if (bp->b_flags & B_INVAL) 2096 bp->b_flags &= ~B_CACHE; 2097 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2098 bp->b_flags |= B_CACHE; 2099 bremfree(bp); 2100 2101 /* 2102 * check for size inconsistancies for non-VMIO case. 2103 */ 2104 2105 if (bp->b_bcount != size) { 2106 if ((bp->b_flags & B_VMIO) == 0 || 2107 (size > bp->b_kvasize)) { 2108 if (bp->b_flags & B_DELWRI) { 2109 bp->b_flags |= B_NOCACHE; 2110 BUF_WRITE(bp); 2111 } else { 2112 if ((bp->b_flags & B_VMIO) && 2113 (LIST_FIRST(&bp->b_dep) == NULL)) { 2114 bp->b_flags |= B_RELBUF; 2115 brelse(bp); 2116 } else { 2117 bp->b_flags |= B_NOCACHE; 2118 BUF_WRITE(bp); 2119 } 2120 } 2121 goto loop; 2122 } 2123 } 2124 2125 /* 2126 * If the size is inconsistant in the VMIO case, we can resize 2127 * the buffer. This might lead to B_CACHE getting set or 2128 * cleared. If the size has not changed, B_CACHE remains 2129 * unchanged from its previous state. 2130 */ 2131 2132 if (bp->b_bcount != size) 2133 allocbuf(bp, size); 2134 2135 KASSERT(bp->b_offset != NOOFFSET, 2136 ("getblk: no buffer offset")); 2137 2138 /* 2139 * A buffer with B_DELWRI set and B_CACHE clear must 2140 * be committed before we can return the buffer in 2141 * order to prevent the caller from issuing a read 2142 * ( due to B_CACHE not being set ) and overwriting 2143 * it. 2144 * 2145 * Most callers, including NFS and FFS, need this to 2146 * operate properly either because they assume they 2147 * can issue a read if B_CACHE is not set, or because 2148 * ( for example ) an uncached B_DELWRI might loop due 2149 * to softupdates re-dirtying the buffer. In the latter 2150 * case, B_CACHE is set after the first write completes, 2151 * preventing further loops. 2152 */ 2153 2154 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2155 BUF_WRITE(bp); 2156 goto loop; 2157 } 2158 2159 splx(s); 2160 bp->b_flags &= ~B_DONE; 2161 } else { 2162 /* 2163 * Buffer is not in-core, create new buffer. The buffer 2164 * returned by getnewbuf() is locked. Note that the returned 2165 * buffer is also considered valid (not marked B_INVAL). 2166 */ 2167 int bsize, maxsize, vmio; 2168 off_t offset; 2169 2170 if (vn_isdisk(vp, NULL)) 2171 bsize = DEV_BSIZE; 2172 else if (vp->v_mountedhere) 2173 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2174 else if (vp->v_mount) 2175 bsize = vp->v_mount->mnt_stat.f_iosize; 2176 else 2177 bsize = size; 2178 2179 offset = (off_t)blkno * bsize; 2180 vmio = (vp->v_object != 0) && (vp->v_flag & VOBJBUF); 2181 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2182 maxsize = imax(maxsize, bsize); 2183 2184 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2185 if (slpflag || slptimeo) { 2186 splx(s); 2187 return NULL; 2188 } 2189 goto loop; 2190 } 2191 2192 /* 2193 * This code is used to make sure that a buffer is not 2194 * created while the getnewbuf routine is blocked. 2195 * This can be a problem whether the vnode is locked or not. 2196 * If the buffer is created out from under us, we have to 2197 * throw away the one we just created. There is now window 2198 * race because we are safely running at splbio() from the 2199 * point of the duplicate buffer creation through to here, 2200 * and we've locked the buffer. 2201 */ 2202 if (gbincore(vp, blkno)) { 2203 bp->b_flags |= B_INVAL; 2204 brelse(bp); 2205 goto loop; 2206 } 2207 2208 /* 2209 * Insert the buffer into the hash, so that it can 2210 * be found by incore. 2211 */ 2212 bp->b_blkno = bp->b_lblkno = blkno; 2213 bp->b_offset = offset; 2214 2215 bgetvp(vp, bp); 2216 LIST_REMOVE(bp, b_hash); 2217 bh = bufhash(vp, blkno); 2218 LIST_INSERT_HEAD(bh, bp, b_hash); 2219 2220 /* 2221 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2222 * buffer size starts out as 0, B_CACHE will be set by 2223 * allocbuf() for the VMIO case prior to it testing the 2224 * backing store for validity. 2225 */ 2226 2227 if (vmio) { 2228 bp->b_flags |= B_VMIO; 2229#if defined(VFS_BIO_DEBUG) 2230 if (vp->v_type != VREG && vp->v_type != VBLK) 2231 printf("getblk: vmioing file type %d???\n", vp->v_type); 2232#endif 2233 } else { 2234 bp->b_flags &= ~B_VMIO; 2235 } 2236 2237 allocbuf(bp, size); 2238 2239 splx(s); 2240 bp->b_flags &= ~B_DONE; 2241 } 2242 return (bp); 2243} 2244 2245/* 2246 * Get an empty, disassociated buffer of given size. The buffer is initially 2247 * set to B_INVAL. 2248 */ 2249struct buf * 2250geteblk(int size) 2251{ 2252 struct buf *bp; 2253 int s; 2254 int maxsize; 2255 2256 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2257 2258 s = splbio(); 2259 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0); 2260 splx(s); 2261 allocbuf(bp, size); 2262 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2263 return (bp); 2264} 2265 2266 2267/* 2268 * This code constitutes the buffer memory from either anonymous system 2269 * memory (in the case of non-VMIO operations) or from an associated 2270 * VM object (in the case of VMIO operations). This code is able to 2271 * resize a buffer up or down. 2272 * 2273 * Note that this code is tricky, and has many complications to resolve 2274 * deadlock or inconsistant data situations. Tread lightly!!! 2275 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2276 * the caller. Calling this code willy nilly can result in the loss of data. 2277 * 2278 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2279 * B_CACHE for the non-VMIO case. 2280 */ 2281 2282int 2283allocbuf(struct buf *bp, int size) 2284{ 2285 int newbsize, mbsize; 2286 int i; 2287 2288 if (BUF_REFCNT(bp) == 0) 2289 panic("allocbuf: buffer not busy"); 2290 2291 if (bp->b_kvasize < size) 2292 panic("allocbuf: buffer too small"); 2293 2294 if ((bp->b_flags & B_VMIO) == 0) { 2295 caddr_t origbuf; 2296 int origbufsize; 2297 /* 2298 * Just get anonymous memory from the kernel. Don't 2299 * mess with B_CACHE. 2300 */ 2301 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2302#if !defined(NO_B_MALLOC) 2303 if (bp->b_flags & B_MALLOC) 2304 newbsize = mbsize; 2305 else 2306#endif 2307 newbsize = round_page(size); 2308 2309 if (newbsize < bp->b_bufsize) { 2310#if !defined(NO_B_MALLOC) 2311 /* 2312 * malloced buffers are not shrunk 2313 */ 2314 if (bp->b_flags & B_MALLOC) { 2315 if (newbsize) { 2316 bp->b_bcount = size; 2317 } else { 2318 free(bp->b_data, M_BIOBUF); 2319 bufmallocspace -= bp->b_bufsize; 2320 runningbufspace -= bp->b_bufsize; 2321 if (bp->b_bufsize) 2322 bufspacewakeup(); 2323 bp->b_data = bp->b_kvabase; 2324 bp->b_bufsize = 0; 2325 bp->b_bcount = 0; 2326 bp->b_flags &= ~B_MALLOC; 2327 } 2328 return 1; 2329 } 2330#endif 2331 vm_hold_free_pages( 2332 bp, 2333 (vm_offset_t) bp->b_data + newbsize, 2334 (vm_offset_t) bp->b_data + bp->b_bufsize); 2335 } else if (newbsize > bp->b_bufsize) { 2336#if !defined(NO_B_MALLOC) 2337 /* 2338 * We only use malloced memory on the first allocation. 2339 * and revert to page-allocated memory when the buffer 2340 * grows. 2341 */ 2342 if ( (bufmallocspace < maxbufmallocspace) && 2343 (bp->b_bufsize == 0) && 2344 (mbsize <= PAGE_SIZE/2)) { 2345 2346 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2347 bp->b_bufsize = mbsize; 2348 bp->b_bcount = size; 2349 bp->b_flags |= B_MALLOC; 2350 bufmallocspace += mbsize; 2351 runningbufspace += bp->b_bufsize; 2352 return 1; 2353 } 2354#endif 2355 origbuf = NULL; 2356 origbufsize = 0; 2357#if !defined(NO_B_MALLOC) 2358 /* 2359 * If the buffer is growing on its other-than-first allocation, 2360 * then we revert to the page-allocation scheme. 2361 */ 2362 if (bp->b_flags & B_MALLOC) { 2363 origbuf = bp->b_data; 2364 origbufsize = bp->b_bufsize; 2365 bp->b_data = bp->b_kvabase; 2366 bufmallocspace -= bp->b_bufsize; 2367 runningbufspace -= bp->b_bufsize; 2368 if (bp->b_bufsize) 2369 bufspacewakeup(); 2370 bp->b_bufsize = 0; 2371 bp->b_flags &= ~B_MALLOC; 2372 newbsize = round_page(newbsize); 2373 } 2374#endif 2375 vm_hold_load_pages( 2376 bp, 2377 (vm_offset_t) bp->b_data + bp->b_bufsize, 2378 (vm_offset_t) bp->b_data + newbsize); 2379#if !defined(NO_B_MALLOC) 2380 if (origbuf) { 2381 bcopy(origbuf, bp->b_data, origbufsize); 2382 free(origbuf, M_BIOBUF); 2383 } 2384#endif 2385 } 2386 } else { 2387 vm_page_t m; 2388 int desiredpages; 2389 2390 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2391 desiredpages = (size == 0) ? 0 : 2392 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2393 2394#if !defined(NO_B_MALLOC) 2395 if (bp->b_flags & B_MALLOC) 2396 panic("allocbuf: VMIO buffer can't be malloced"); 2397#endif 2398 /* 2399 * Set B_CACHE initially if buffer is 0 length or will become 2400 * 0-length. 2401 */ 2402 if (size == 0 || bp->b_bufsize == 0) 2403 bp->b_flags |= B_CACHE; 2404 2405 if (newbsize < bp->b_bufsize) { 2406 /* 2407 * DEV_BSIZE aligned new buffer size is less then the 2408 * DEV_BSIZE aligned existing buffer size. Figure out 2409 * if we have to remove any pages. 2410 */ 2411 if (desiredpages < bp->b_npages) { 2412 for (i = desiredpages; i < bp->b_npages; i++) { 2413 /* 2414 * the page is not freed here -- it 2415 * is the responsibility of 2416 * vnode_pager_setsize 2417 */ 2418 m = bp->b_pages[i]; 2419 KASSERT(m != bogus_page, 2420 ("allocbuf: bogus page found")); 2421 while (vm_page_sleep_busy(m, TRUE, "biodep")) 2422 ; 2423 2424 bp->b_pages[i] = NULL; 2425 vm_page_unwire(m, 0); 2426 } 2427 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2428 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2429 bp->b_npages = desiredpages; 2430 } 2431 } else if (size > bp->b_bcount) { 2432 /* 2433 * We are growing the buffer, possibly in a 2434 * byte-granular fashion. 2435 */ 2436 struct vnode *vp; 2437 vm_object_t obj; 2438 vm_offset_t toff; 2439 vm_offset_t tinc; 2440 2441 /* 2442 * Step 1, bring in the VM pages from the object, 2443 * allocating them if necessary. We must clear 2444 * B_CACHE if these pages are not valid for the 2445 * range covered by the buffer. 2446 */ 2447 2448 vp = bp->b_vp; 2449 obj = vp->v_object; 2450 2451 while (bp->b_npages < desiredpages) { 2452 vm_page_t m; 2453 vm_pindex_t pi; 2454 2455 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2456 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2457 m = vm_page_alloc(obj, pi, VM_ALLOC_NORMAL); 2458 if (m == NULL) { 2459 VM_WAIT; 2460 vm_pageout_deficit += desiredpages - bp->b_npages; 2461 } else { 2462 vm_page_wire(m); 2463 vm_page_wakeup(m); 2464 bp->b_flags &= ~B_CACHE; 2465 bp->b_pages[bp->b_npages] = m; 2466 ++bp->b_npages; 2467 } 2468 continue; 2469 } 2470 2471 /* 2472 * We found a page. If we have to sleep on it, 2473 * retry because it might have gotten freed out 2474 * from under us. 2475 * 2476 * We can only test PG_BUSY here. Blocking on 2477 * m->busy might lead to a deadlock: 2478 * 2479 * vm_fault->getpages->cluster_read->allocbuf 2480 * 2481 */ 2482 2483 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 2484 continue; 2485 2486 /* 2487 * We have a good page. Should we wakeup the 2488 * page daemon? 2489 */ 2490 if ((curproc != pageproc) && 2491 ((m->queue - m->pc) == PQ_CACHE) && 2492 ((cnt.v_free_count + cnt.v_cache_count) < 2493 (cnt.v_free_min + cnt.v_cache_min))) { 2494 pagedaemon_wakeup(); 2495 } 2496 vm_page_flag_clear(m, PG_ZERO); 2497 vm_page_wire(m); 2498 bp->b_pages[bp->b_npages] = m; 2499 ++bp->b_npages; 2500 } 2501 2502 /* 2503 * Step 2. We've loaded the pages into the buffer, 2504 * we have to figure out if we can still have B_CACHE 2505 * set. Note that B_CACHE is set according to the 2506 * byte-granular range ( bcount and size ), new the 2507 * aligned range ( newbsize ). 2508 * 2509 * The VM test is against m->valid, which is DEV_BSIZE 2510 * aligned. Needless to say, the validity of the data 2511 * needs to also be DEV_BSIZE aligned. Note that this 2512 * fails with NFS if the server or some other client 2513 * extends the file's EOF. If our buffer is resized, 2514 * B_CACHE may remain set! XXX 2515 */ 2516 2517 toff = bp->b_bcount; 2518 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2519 2520 while ((bp->b_flags & B_CACHE) && toff < size) { 2521 vm_pindex_t pi; 2522 2523 if (tinc > (size - toff)) 2524 tinc = size - toff; 2525 2526 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2527 PAGE_SHIFT; 2528 2529 vfs_buf_test_cache( 2530 bp, 2531 bp->b_offset, 2532 toff, 2533 tinc, 2534 bp->b_pages[pi] 2535 ); 2536 toff += tinc; 2537 tinc = PAGE_SIZE; 2538 } 2539 2540 /* 2541 * Step 3, fixup the KVM pmap. Remember that 2542 * bp->b_data is relative to bp->b_offset, but 2543 * bp->b_offset may be offset into the first page. 2544 */ 2545 2546 bp->b_data = (caddr_t) 2547 trunc_page((vm_offset_t)bp->b_data); 2548 pmap_qenter( 2549 (vm_offset_t)bp->b_data, 2550 bp->b_pages, 2551 bp->b_npages 2552 ); 2553 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2554 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2555 } 2556 } 2557 runningbufspace += (newbsize - bp->b_bufsize); 2558 if (newbsize < bp->b_bufsize) 2559 bufspacewakeup(); 2560 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2561 bp->b_bcount = size; /* requested buffer size */ 2562 return 1; 2563} 2564 2565/* 2566 * biowait: 2567 * 2568 * Wait for buffer I/O completion, returning error status. The buffer 2569 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR 2570 * error and cleared. 2571 */ 2572int 2573biowait(register struct buf * bp) 2574{ 2575 int s; 2576 2577 s = splbio(); 2578 while ((bp->b_flags & B_DONE) == 0) { 2579#if defined(NO_SCHEDULE_MODS) 2580 tsleep(bp, PRIBIO, "biowait", 0); 2581#else 2582 if (bp->b_iocmd == BIO_READ) 2583 tsleep(bp, PRIBIO, "biord", 0); 2584 else 2585 tsleep(bp, PRIBIO, "biowr", 0); 2586#endif 2587 } 2588 splx(s); 2589 if (bp->b_flags & B_EINTR) { 2590 bp->b_flags &= ~B_EINTR; 2591 return (EINTR); 2592 } 2593 if (bp->b_flags & B_ERROR) { 2594 return (bp->b_error ? bp->b_error : EIO); 2595 } else { 2596 return (0); 2597 } 2598} 2599 2600/* 2601 * biodone: 2602 * 2603 * Finish I/O on a buffer, optionally calling a completion function. 2604 * This is usually called from an interrupt so process blocking is 2605 * not allowed. 2606 * 2607 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2608 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2609 * assuming B_INVAL is clear. 2610 * 2611 * For the VMIO case, we set B_CACHE if the op was a read and no 2612 * read error occured, or if the op was a write. B_CACHE is never 2613 * set if the buffer is invalid or otherwise uncacheable. 2614 * 2615 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2616 * initiator to leave B_INVAL set to brelse the buffer out of existance 2617 * in the biodone routine. 2618 */ 2619void 2620biodone(register struct buf * bp) 2621{ 2622 int s; 2623 void (*biodone) __P((struct buf *)); 2624 2625 s = splbio(); 2626 2627 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 2628 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 2629 2630 bp->b_flags |= B_DONE; 2631 2632 if (bp->b_iocmd == BIO_DELETE) { 2633 brelse(bp); 2634 splx(s); 2635 return; 2636 } 2637 2638 if (bp->b_iocmd == BIO_WRITE) { 2639 vwakeup(bp); 2640 } 2641 2642 /* call optional completion function if requested */ 2643 if (bp->b_iodone != NULL) { 2644 biodone = bp->b_iodone; 2645 bp->b_iodone = NULL; 2646 (*biodone) (bp); 2647 splx(s); 2648 return; 2649 } 2650 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 2651 (*bioops.io_complete)(bp); 2652 2653 if (bp->b_flags & B_VMIO) { 2654 int i, resid; 2655 vm_ooffset_t foff; 2656 vm_page_t m; 2657 vm_object_t obj; 2658 int iosize; 2659 struct vnode *vp = bp->b_vp; 2660 2661 obj = vp->v_object; 2662 2663#if defined(VFS_BIO_DEBUG) 2664 if (vp->v_usecount == 0) { 2665 panic("biodone: zero vnode ref count"); 2666 } 2667 2668 if (vp->v_object == NULL) { 2669 panic("biodone: missing VM object"); 2670 } 2671 2672 if ((vp->v_flag & VOBJBUF) == 0) { 2673 panic("biodone: vnode is not setup for merged cache"); 2674 } 2675#endif 2676 2677 foff = bp->b_offset; 2678 KASSERT(bp->b_offset != NOOFFSET, 2679 ("biodone: no buffer offset")); 2680 2681 if (!obj) { 2682 panic("biodone: no object"); 2683 } 2684#if defined(VFS_BIO_DEBUG) 2685 if (obj->paging_in_progress < bp->b_npages) { 2686 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 2687 obj->paging_in_progress, bp->b_npages); 2688 } 2689#endif 2690 2691 /* 2692 * Set B_CACHE if the op was a normal read and no error 2693 * occured. B_CACHE is set for writes in the b*write() 2694 * routines. 2695 */ 2696 iosize = bp->b_bcount - bp->b_resid; 2697 if (bp->b_iocmd == BIO_READ && 2698 !(bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR))) { 2699 bp->b_flags |= B_CACHE; 2700 } 2701 2702 for (i = 0; i < bp->b_npages; i++) { 2703 int bogusflag = 0; 2704 m = bp->b_pages[i]; 2705 if (m == bogus_page) { 2706 bogusflag = 1; 2707 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2708 if (!m) { 2709#if defined(VFS_BIO_DEBUG) 2710 printf("biodone: page disappeared\n"); 2711#endif 2712 vm_object_pip_subtract(obj, 1); 2713 bp->b_flags &= ~B_CACHE; 2714 continue; 2715 } 2716 bp->b_pages[i] = m; 2717 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2718 } 2719#if defined(VFS_BIO_DEBUG) 2720 if (OFF_TO_IDX(foff) != m->pindex) { 2721 printf( 2722"biodone: foff(%lu)/m->pindex(%d) mismatch\n", 2723 (unsigned long)foff, m->pindex); 2724 } 2725#endif 2726 resid = IDX_TO_OFF(m->pindex + 1) - foff; 2727 if (resid > iosize) 2728 resid = iosize; 2729 2730 /* 2731 * In the write case, the valid and clean bits are 2732 * already changed correctly ( see bdwrite() ), so we 2733 * only need to do this here in the read case. 2734 */ 2735 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 2736 vfs_page_set_valid(bp, foff, i, m); 2737 } 2738 vm_page_flag_clear(m, PG_ZERO); 2739 2740 /* 2741 * when debugging new filesystems or buffer I/O methods, this 2742 * is the most common error that pops up. if you see this, you 2743 * have not set the page busy flag correctly!!! 2744 */ 2745 if (m->busy == 0) { 2746 printf("biodone: page busy < 0, " 2747 "pindex: %d, foff: 0x(%x,%x), " 2748 "resid: %d, index: %d\n", 2749 (int) m->pindex, (int)(foff >> 32), 2750 (int) foff & 0xffffffff, resid, i); 2751 if (!vn_isdisk(vp, NULL)) 2752 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 2753 bp->b_vp->v_mount->mnt_stat.f_iosize, 2754 (int) bp->b_lblkno, 2755 bp->b_flags, bp->b_npages); 2756 else 2757 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 2758 (int) bp->b_lblkno, 2759 bp->b_flags, bp->b_npages); 2760 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2761 m->valid, m->dirty, m->wire_count); 2762 panic("biodone: page busy < 0\n"); 2763 } 2764 vm_page_io_finish(m); 2765 vm_object_pip_subtract(obj, 1); 2766 foff += resid; 2767 iosize -= resid; 2768 } 2769 if (obj) 2770 vm_object_pip_wakeupn(obj, 0); 2771 } 2772 /* 2773 * For asynchronous completions, release the buffer now. The brelse 2774 * will do a wakeup there if necessary - so no need to do a wakeup 2775 * here in the async case. The sync case always needs to do a wakeup. 2776 */ 2777 2778 if (bp->b_flags & B_ASYNC) { 2779 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0) 2780 brelse(bp); 2781 else 2782 bqrelse(bp); 2783 } else { 2784 wakeup(bp); 2785 } 2786 splx(s); 2787} 2788 2789/* 2790 * This routine is called in lieu of iodone in the case of 2791 * incomplete I/O. This keeps the busy status for pages 2792 * consistant. 2793 */ 2794void 2795vfs_unbusy_pages(struct buf * bp) 2796{ 2797 int i; 2798 2799 if (bp->b_flags & B_VMIO) { 2800 struct vnode *vp = bp->b_vp; 2801 vm_object_t obj = vp->v_object; 2802 2803 for (i = 0; i < bp->b_npages; i++) { 2804 vm_page_t m = bp->b_pages[i]; 2805 2806 if (m == bogus_page) { 2807 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 2808 if (!m) { 2809 panic("vfs_unbusy_pages: page missing\n"); 2810 } 2811 bp->b_pages[i] = m; 2812 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2813 } 2814 vm_object_pip_subtract(obj, 1); 2815 vm_page_flag_clear(m, PG_ZERO); 2816 vm_page_io_finish(m); 2817 } 2818 vm_object_pip_wakeupn(obj, 0); 2819 } 2820} 2821 2822/* 2823 * vfs_page_set_valid: 2824 * 2825 * Set the valid bits in a page based on the supplied offset. The 2826 * range is restricted to the buffer's size. 2827 * 2828 * This routine is typically called after a read completes. 2829 */ 2830static void 2831vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 2832{ 2833 vm_ooffset_t soff, eoff; 2834 2835 /* 2836 * Start and end offsets in buffer. eoff - soff may not cross a 2837 * page boundry or cross the end of the buffer. The end of the 2838 * buffer, in this case, is our file EOF, not the allocation size 2839 * of the buffer. 2840 */ 2841 soff = off; 2842 eoff = (off + PAGE_SIZE) & ~PAGE_MASK; 2843 if (eoff > bp->b_offset + bp->b_bcount) 2844 eoff = bp->b_offset + bp->b_bcount; 2845 2846 /* 2847 * Set valid range. This is typically the entire buffer and thus the 2848 * entire page. 2849 */ 2850 if (eoff > soff) { 2851 vm_page_set_validclean( 2852 m, 2853 (vm_offset_t) (soff & PAGE_MASK), 2854 (vm_offset_t) (eoff - soff) 2855 ); 2856 } 2857} 2858 2859/* 2860 * This routine is called before a device strategy routine. 2861 * It is used to tell the VM system that paging I/O is in 2862 * progress, and treat the pages associated with the buffer 2863 * almost as being PG_BUSY. Also the object paging_in_progress 2864 * flag is handled to make sure that the object doesn't become 2865 * inconsistant. 2866 * 2867 * Since I/O has not been initiated yet, certain buffer flags 2868 * such as B_ERROR or B_INVAL may be in an inconsistant state 2869 * and should be ignored. 2870 */ 2871void 2872vfs_busy_pages(struct buf * bp, int clear_modify) 2873{ 2874 int i, bogus; 2875 2876 if (bp->b_flags & B_VMIO) { 2877 struct vnode *vp = bp->b_vp; 2878 vm_object_t obj = vp->v_object; 2879 vm_ooffset_t foff; 2880 2881 foff = bp->b_offset; 2882 KASSERT(bp->b_offset != NOOFFSET, 2883 ("vfs_busy_pages: no buffer offset")); 2884 vfs_setdirty(bp); 2885 2886retry: 2887 for (i = 0; i < bp->b_npages; i++) { 2888 vm_page_t m = bp->b_pages[i]; 2889 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 2890 goto retry; 2891 } 2892 2893 bogus = 0; 2894 for (i = 0; i < bp->b_npages; i++) { 2895 vm_page_t m = bp->b_pages[i]; 2896 2897 vm_page_flag_clear(m, PG_ZERO); 2898 if ((bp->b_flags & B_CLUSTER) == 0) { 2899 vm_object_pip_add(obj, 1); 2900 vm_page_io_start(m); 2901 } 2902 2903 /* 2904 * When readying a buffer for a read ( i.e 2905 * clear_modify == 0 ), it is important to do 2906 * bogus_page replacement for valid pages in 2907 * partially instantiated buffers. Partially 2908 * instantiated buffers can, in turn, occur when 2909 * reconstituting a buffer from its VM backing store 2910 * base. We only have to do this if B_CACHE is 2911 * clear ( which causes the I/O to occur in the 2912 * first place ). The replacement prevents the read 2913 * I/O from overwriting potentially dirty VM-backed 2914 * pages. XXX bogus page replacement is, uh, bogus. 2915 * It may not work properly with small-block devices. 2916 * We need to find a better way. 2917 */ 2918 2919 vm_page_protect(m, VM_PROT_NONE); 2920 if (clear_modify) 2921 vfs_page_set_valid(bp, foff, i, m); 2922 else if (m->valid == VM_PAGE_BITS_ALL && 2923 (bp->b_flags & B_CACHE) == 0) { 2924 bp->b_pages[i] = bogus_page; 2925 bogus++; 2926 } 2927 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 2928 } 2929 if (bogus) 2930 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2931 } 2932} 2933 2934/* 2935 * Tell the VM system that the pages associated with this buffer 2936 * are clean. This is used for delayed writes where the data is 2937 * going to go to disk eventually without additional VM intevention. 2938 * 2939 * Note that while we only really need to clean through to b_bcount, we 2940 * just go ahead and clean through to b_bufsize. 2941 */ 2942static void 2943vfs_clean_pages(struct buf * bp) 2944{ 2945 int i; 2946 2947 if (bp->b_flags & B_VMIO) { 2948 vm_ooffset_t foff; 2949 2950 foff = bp->b_offset; 2951 KASSERT(bp->b_offset != NOOFFSET, 2952 ("vfs_clean_pages: no buffer offset")); 2953 for (i = 0; i < bp->b_npages; i++) { 2954 vm_page_t m = bp->b_pages[i]; 2955 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~PAGE_MASK; 2956 vm_ooffset_t eoff = noff; 2957 2958 if (eoff > bp->b_offset + bp->b_bufsize) 2959 eoff = bp->b_offset + bp->b_bufsize; 2960 vfs_page_set_valid(bp, foff, i, m); 2961 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2962 foff = noff; 2963 } 2964 } 2965} 2966 2967/* 2968 * vfs_bio_set_validclean: 2969 * 2970 * Set the range within the buffer to valid and clean. The range is 2971 * relative to the beginning of the buffer, b_offset. Note that b_offset 2972 * itself may be offset from the beginning of the first page. 2973 */ 2974 2975void 2976vfs_bio_set_validclean(struct buf *bp, int base, int size) 2977{ 2978 if (bp->b_flags & B_VMIO) { 2979 int i; 2980 int n; 2981 2982 /* 2983 * Fixup base to be relative to beginning of first page. 2984 * Set initial n to be the maximum number of bytes in the 2985 * first page that can be validated. 2986 */ 2987 2988 base += (bp->b_offset & PAGE_MASK); 2989 n = PAGE_SIZE - (base & PAGE_MASK); 2990 2991 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 2992 vm_page_t m = bp->b_pages[i]; 2993 2994 if (n > size) 2995 n = size; 2996 2997 vm_page_set_validclean(m, base & PAGE_MASK, n); 2998 base += n; 2999 size -= n; 3000 n = PAGE_SIZE; 3001 } 3002 } 3003} 3004 3005/* 3006 * vfs_bio_clrbuf: 3007 * 3008 * clear a buffer. This routine essentially fakes an I/O, so we need 3009 * to clear B_ERROR and B_INVAL. 3010 * 3011 * Note that while we only theoretically need to clear through b_bcount, 3012 * we go ahead and clear through b_bufsize. 3013 */ 3014 3015void 3016vfs_bio_clrbuf(struct buf *bp) { 3017 int i, mask = 0; 3018 caddr_t sa, ea; 3019 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3020 bp->b_flags &= ~(B_INVAL|B_ERROR); 3021 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3022 (bp->b_offset & PAGE_MASK) == 0) { 3023 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3024 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3025 ((bp->b_pages[0]->valid & mask) != mask)) { 3026 bzero(bp->b_data, bp->b_bufsize); 3027 } 3028 bp->b_pages[0]->valid |= mask; 3029 bp->b_resid = 0; 3030 return; 3031 } 3032 ea = sa = bp->b_data; 3033 for(i=0;i<bp->b_npages;i++,sa=ea) { 3034 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3035 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3036 ea = (caddr_t)(vm_offset_t)ulmin( 3037 (u_long)(vm_offset_t)ea, 3038 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3039 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3040 if ((bp->b_pages[i]->valid & mask) == mask) 3041 continue; 3042 if ((bp->b_pages[i]->valid & mask) == 0) { 3043 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3044 bzero(sa, ea - sa); 3045 } 3046 } else { 3047 for (; sa < ea; sa += DEV_BSIZE, j++) { 3048 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3049 (bp->b_pages[i]->valid & (1<<j)) == 0) 3050 bzero(sa, DEV_BSIZE); 3051 } 3052 } 3053 bp->b_pages[i]->valid |= mask; 3054 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3055 } 3056 bp->b_resid = 0; 3057 } else { 3058 clrbuf(bp); 3059 } 3060} 3061 3062/* 3063 * vm_hold_load_pages and vm_hold_unload pages get pages into 3064 * a buffers address space. The pages are anonymous and are 3065 * not associated with a file object. 3066 */ 3067void 3068vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3069{ 3070 vm_offset_t pg; 3071 vm_page_t p; 3072 int index; 3073 3074 to = round_page(to); 3075 from = round_page(from); 3076 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3077 3078 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3079 3080tryagain: 3081 3082 p = vm_page_alloc(kernel_object, 3083 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3084 VM_ALLOC_NORMAL); 3085 if (!p) { 3086 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3087 VM_WAIT; 3088 goto tryagain; 3089 } 3090 vm_page_wire(p); 3091 p->valid = VM_PAGE_BITS_ALL; 3092 vm_page_flag_clear(p, PG_ZERO); 3093 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 3094 bp->b_pages[index] = p; 3095 vm_page_wakeup(p); 3096 } 3097 bp->b_npages = index; 3098} 3099 3100void 3101vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3102{ 3103 vm_offset_t pg; 3104 vm_page_t p; 3105 int index, newnpages; 3106 3107 from = round_page(from); 3108 to = round_page(to); 3109 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3110 3111 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3112 p = bp->b_pages[index]; 3113 if (p && (index < bp->b_npages)) { 3114 if (p->busy) { 3115 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 3116 bp->b_blkno, bp->b_lblkno); 3117 } 3118 bp->b_pages[index] = NULL; 3119 pmap_kremove(pg); 3120 vm_page_busy(p); 3121 vm_page_unwire(p, 0); 3122 vm_page_free(p); 3123 } 3124 } 3125 bp->b_npages = newnpages; 3126} 3127 3128 3129#include "opt_ddb.h" 3130#ifdef DDB 3131#include <ddb/ddb.h> 3132 3133DB_SHOW_COMMAND(buffer, db_show_buffer) 3134{ 3135 /* get args */ 3136 struct buf *bp = (struct buf *)addr; 3137 3138 if (!have_addr) { 3139 db_printf("usage: show buffer <addr>\n"); 3140 return; 3141 } 3142 3143 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3144 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 3145 "b_resid = %ld\nb_dev = (%d,%d), b_data = %p, " 3146 "b_blkno = %d, b_pblkno = %d\n", 3147 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3148 major(bp->b_dev), minor(bp->b_dev), 3149 bp->b_data, bp->b_blkno, bp->b_pblkno); 3150 if (bp->b_npages) { 3151 int i; 3152 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3153 for (i = 0; i < bp->b_npages; i++) { 3154 vm_page_t m; 3155 m = bp->b_pages[i]; 3156 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3157 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3158 if ((i + 1) < bp->b_npages) 3159 db_printf(","); 3160 } 3161 db_printf("\n"); 3162 } 3163} 3164#endif /* DDB */ 3165