1 /* $NetBSD: vfs_bio.c,v 1.81 2002/05/12 23:06:27 matt Exp $ */ 2 3 /*- 4 * Copyright (c) 1994 Christopher G. Demetriou 5 * Copyright (c) 1982, 1986, 1989, 1993 6 * The Regents of the University of California. All rights reserved. 7 * (c) UNIX System Laboratories, Inc. 8 * All or some portions of this file are derived from material licensed 9 * to the University of California by American Telephone and Telegraph 10 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 11 * the permission of UNIX System Laboratories, Inc. 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 3. All advertising materials mentioning features or use of this software 22 * must display the following acknowledgement: 23 * This product includes software developed by the University of 24 * California, Berkeley and its contributors. 25 * 4. Neither the name of the University nor the names of its contributors 26 * may be used to endorse or promote products derived from this software 27 * without specific prior written permission. 28 * 29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 39 * SUCH DAMAGE. 40 * 41 * @(#)vfs_bio.c 8.6 (Berkeley) 1/11/94 42 */ 43 44 /* 45 * Some references: 46 * Bach: The Design of the UNIX Operating System (Prentice Hall, 1986) 47 * Leffler, et al.: The Design and Implementation of the 4.3BSD 48 * UNIX Operating System (Addison Welley, 1989) 49 */ 50 51 #include "opt_softdep.h" 52 53 #include <sys/cdefs.h> 54 __KERNEL_RCSID(0, "$NetBSD: vfs_bio.c,v 1.81 2002/05/12 23:06:27 matt Exp $"); 55 56 #include <sys/param.h> 57 #include <sys/systm.h> 58 #include <sys/proc.h> 59 #include <sys/buf.h> 60 #include <sys/vnode.h> 61 #include <sys/mount.h> 62 #include <sys/malloc.h> 63 #include <sys/resourcevar.h> 64 #include <sys/conf.h> 65 66 #include <uvm/uvm.h> 67 68 #include <miscfs/specfs/specdev.h> 69 70 /* Macros to clear/set/test flags. */ 71 #define SET(t, f) (t) |= (f) 72 #define CLR(t, f) (t) &= ~(f) 73 #define ISSET(t, f) ((t) & (f)) 74 75 /* 76 * Definitions for the buffer hash lists. 77 */ 78 #define BUFHASH(dvp, lbn) \ 79 (&bufhashtbl[(((long)(dvp) >> 8) + (int)(lbn)) & bufhash]) 80 LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash; 81 u_long bufhash; 82 #ifndef SOFTDEP 83 struct bio_ops bioops; /* I/O operation notification */ 84 #endif 85 86 /* 87 * Insq/Remq for the buffer hash lists. 88 */ 89 #define binshash(bp, dp) LIST_INSERT_HEAD(dp, bp, b_hash) 90 #define bremhash(bp) LIST_REMOVE(bp, b_hash) 91 92 /* 93 * Definitions for the buffer free lists. 94 */ 95 #define BQUEUES 4 /* number of free buffer queues */ 96 97 #define BQ_LOCKED 0 /* super-blocks &c */ 98 #define BQ_LRU 1 /* lru, useful buffers */ 99 #define BQ_AGE 2 /* rubbish */ 100 #define BQ_EMPTY 3 /* buffer headers with no memory */ 101 102 TAILQ_HEAD(bqueues, buf) bufqueues[BQUEUES]; 103 int needbuffer; 104 105 /* 106 * Buffer pool for I/O buffers. 107 */ 108 struct pool bufpool; 109 110 /* 111 * Insq/Remq for the buffer free lists. 112 */ 113 #define binsheadfree(bp, dp) TAILQ_INSERT_HEAD(dp, bp, b_freelist) 114 #define binstailfree(bp, dp) TAILQ_INSERT_TAIL(dp, bp, b_freelist) 115 116 static __inline struct buf *bio_doread __P((struct vnode *, daddr_t, int, 117 struct ucred *, int)); 118 int count_lock_queue __P((void)); 119 120 void 121 bremfree(bp) 122 struct buf *bp; 123 { 124 int s = splbio(); 125 126 struct bqueues *dp = NULL; 127 128 /* 129 * We only calculate the head of the freelist when removing 130 * the last element of the list as that is the only time that 131 * it is needed (e.g. to reset the tail pointer). 132 * 133 * NB: This makes an assumption about how tailq's are implemented. 134 */ 135 if (bp->b_freelist.tqe_next == NULL) { 136 for (dp = bufqueues; dp < &bufqueues[BQUEUES]; dp++) 137 if (dp->tqh_last == &bp->b_freelist.tqe_next) 138 break; 139 if (dp == &bufqueues[BQUEUES]) 140 panic("bremfree: lost tail"); 141 } 142 TAILQ_REMOVE(dp, bp, b_freelist); 143 splx(s); 144 } 145 146 /* 147 * Initialize buffers and hash links for buffers. 148 */ 149 void 150 bufinit() 151 { 152 struct buf *bp; 153 struct bqueues *dp; 154 int i; 155 int base, residual; 156 157 /* 158 * Initialize the buffer pool. This pool is used for buffers 159 * which are strictly I/O control blocks, not buffer cache 160 * buffers. 161 */ 162 pool_init(&bufpool, sizeof(struct buf), 0, 0, 0, "bufpl", NULL); 163 164 for (dp = bufqueues; dp < &bufqueues[BQUEUES]; dp++) 165 TAILQ_INIT(dp); 166 bufhashtbl = hashinit(nbuf, HASH_LIST, M_CACHE, M_WAITOK, &bufhash); 167 base = bufpages / nbuf; 168 residual = bufpages % nbuf; 169 for (i = 0; i < nbuf; i++) { 170 bp = &buf[i]; 171 memset((char *)bp, 0, sizeof(*bp)); 172 bp->b_dev = NODEV; 173 bp->b_vnbufs.le_next = NOLIST; 174 LIST_INIT(&bp->b_dep); 175 bp->b_data = buffers + i * MAXBSIZE; 176 if (i < residual) 177 bp->b_bufsize = (base + 1) * PAGE_SIZE; 178 else 179 bp->b_bufsize = base * PAGE_SIZE; 180 bp->b_flags = B_INVAL; 181 dp = bp->b_bufsize ? &bufqueues[BQ_AGE] : &bufqueues[BQ_EMPTY]; 182 binsheadfree(bp, dp); 183 binshash(bp, &invalhash); 184 } 185 } 186 187 static __inline struct buf * 188 bio_doread(vp, blkno, size, cred, async) 189 struct vnode *vp; 190 daddr_t blkno; 191 int size; 192 struct ucred *cred; 193 int async; 194 { 195 struct buf *bp; 196 struct proc *p = (curproc != NULL ? curproc : &proc0); /* XXX */ 197 198 bp = getblk(vp, blkno, size, 0, 0); 199 200 /* 201 * If buffer does not have data valid, start a read. 202 * Note that if buffer is B_INVAL, getblk() won't return it. 203 * Therefore, it's valid if it's I/O has completed or been delayed. 204 */ 205 if (!ISSET(bp->b_flags, (B_DONE | B_DELWRI))) { 206 /* Start I/O for the buffer. */ 207 SET(bp->b_flags, B_READ | async); 208 VOP_STRATEGY(bp); 209 210 /* Pay for the read. */ 211 p->p_stats->p_ru.ru_inblock++; 212 } else if (async) { 213 brelse(bp); 214 } 215 216 return (bp); 217 } 218 219 /* 220 * Read a disk block. 221 * This algorithm described in Bach (p.54). 222 */ 223 int 224 bread(vp, blkno, size, cred, bpp) 225 struct vnode *vp; 226 daddr_t blkno; 227 int size; 228 struct ucred *cred; 229 struct buf **bpp; 230 { 231 struct buf *bp; 232 233 /* Get buffer for block. */ 234 bp = *bpp = bio_doread(vp, blkno, size, cred, 0); 235 236 /* Wait for the read to complete, and return result. */ 237 return (biowait(bp)); 238 } 239 240 /* 241 * Read-ahead multiple disk blocks. The first is sync, the rest async. 242 * Trivial modification to the breada algorithm presented in Bach (p.55). 243 */ 244 int 245 breadn(vp, blkno, size, rablks, rasizes, nrablks, cred, bpp) 246 struct vnode *vp; 247 daddr_t blkno; int size; 248 daddr_t rablks[]; int rasizes[]; 249 int nrablks; 250 struct ucred *cred; 251 struct buf **bpp; 252 { 253 struct buf *bp; 254 int i; 255 256 bp = *bpp = bio_doread(vp, blkno, size, cred, 0); 257 258 /* 259 * For each of the read-ahead blocks, start a read, if necessary. 260 */ 261 for (i = 0; i < nrablks; i++) { 262 /* If it's in the cache, just go on to next one. */ 263 if (incore(vp, rablks[i])) 264 continue; 265 266 /* Get a buffer for the read-ahead block */ 267 (void) bio_doread(vp, rablks[i], rasizes[i], cred, B_ASYNC); 268 } 269 270 /* Otherwise, we had to start a read for it; wait until it's valid. */ 271 return (biowait(bp)); 272 } 273 274 /* 275 * Read with single-block read-ahead. Defined in Bach (p.55), but 276 * implemented as a call to breadn(). 277 * XXX for compatibility with old file systems. 278 */ 279 int 280 breada(vp, blkno, size, rablkno, rabsize, cred, bpp) 281 struct vnode *vp; 282 daddr_t blkno; int size; 283 daddr_t rablkno; int rabsize; 284 struct ucred *cred; 285 struct buf **bpp; 286 { 287 288 return (breadn(vp, blkno, size, &rablkno, &rabsize, 1, cred, bpp)); 289 } 290 291 /* 292 * Block write. Described in Bach (p.56) 293 */ 294 int 295 bwrite(bp) 296 struct buf *bp; 297 { 298 int rv, sync, wasdelayed, s; 299 struct proc *p = (curproc != NULL ? curproc : &proc0); /* XXX */ 300 struct vnode *vp; 301 struct mount *mp; 302 303 vp = bp->b_vp; 304 if (vp != NULL) { 305 if (vp->v_type == VBLK) 306 mp = vp->v_specmountpoint; 307 else 308 mp = vp->v_mount; 309 } else { 310 mp = NULL; 311 } 312 313 /* 314 * Remember buffer type, to switch on it later. If the write was 315 * synchronous, but the file system was mounted with MNT_ASYNC, 316 * convert it to a delayed write. 317 * XXX note that this relies on delayed tape writes being converted 318 * to async, not sync writes (which is safe, but ugly). 319 */ 320 sync = !ISSET(bp->b_flags, B_ASYNC); 321 if (sync && mp != NULL && ISSET(mp->mnt_flag, MNT_ASYNC)) { 322 bdwrite(bp); 323 return (0); 324 } 325 326 /* 327 * Collect statistics on synchronous and asynchronous writes. 328 * Writes to block devices are charged to their associated 329 * filesystem (if any). 330 */ 331 if (mp != NULL) { 332 if (sync) 333 mp->mnt_stat.f_syncwrites++; 334 else 335 mp->mnt_stat.f_asyncwrites++; 336 } 337 338 wasdelayed = ISSET(bp->b_flags, B_DELWRI); 339 340 s = splbio(); 341 342 CLR(bp->b_flags, (B_READ | B_DONE | B_ERROR | B_DELWRI)); 343 344 /* 345 * Pay for the I/O operation and make sure the buf is on the correct 346 * vnode queue. 347 */ 348 if (wasdelayed) 349 reassignbuf(bp, bp->b_vp); 350 else 351 p->p_stats->p_ru.ru_oublock++; 352 353 /* Initiate disk write. Make sure the appropriate party is charged. */ 354 bp->b_vp->v_numoutput++; 355 splx(s); 356 357 VOP_STRATEGY(bp); 358 359 if (sync) { 360 /* If I/O was synchronous, wait for it to complete. */ 361 rv = biowait(bp); 362 363 /* Release the buffer. */ 364 brelse(bp); 365 366 return (rv); 367 } else { 368 return (0); 369 } 370 } 371 372 int 373 vn_bwrite(v) 374 void *v; 375 { 376 struct vop_bwrite_args *ap = v; 377 378 return (bwrite(ap->a_bp)); 379 } 380 381 /* 382 * Delayed write. 383 * 384 * The buffer is marked dirty, but is not queued for I/O. 385 * This routine should be used when the buffer is expected 386 * to be modified again soon, typically a small write that 387 * partially fills a buffer. 388 * 389 * NB: magnetic tapes cannot be delayed; they must be 390 * written in the order that the writes are requested. 391 * 392 * Described in Leffler, et al. (pp. 208-213). 393 */ 394 void 395 bdwrite(bp) 396 struct buf *bp; 397 { 398 struct proc *p = (curproc != NULL ? curproc : &proc0); /* XXX */ 399 int s; 400 401 /* If this is a tape block, write the block now. */ 402 /* XXX NOTE: the memory filesystem usurpes major device */ 403 /* XXX number 255, which is a bad idea. */ 404 if (bp->b_dev != NODEV && 405 major(bp->b_dev) != 255 && /* XXX - MFS buffers! */ 406 bdevsw[major(bp->b_dev)].d_type == D_TAPE) { 407 bawrite(bp); 408 return; 409 } 410 411 /* 412 * If the block hasn't been seen before: 413 * (1) Mark it as having been seen, 414 * (2) Charge for the write, 415 * (3) Make sure it's on its vnode's correct block list. 416 */ 417 s = splbio(); 418 419 if (!ISSET(bp->b_flags, B_DELWRI)) { 420 SET(bp->b_flags, B_DELWRI); 421 p->p_stats->p_ru.ru_oublock++; 422 reassignbuf(bp, bp->b_vp); 423 } 424 425 /* Otherwise, the "write" is done, so mark and release the buffer. */ 426 CLR(bp->b_flags, B_NEEDCOMMIT|B_DONE); 427 splx(s); 428 429 brelse(bp); 430 } 431 432 /* 433 * Asynchronous block write; just an asynchronous bwrite(). 434 */ 435 void 436 bawrite(bp) 437 struct buf *bp; 438 { 439 440 SET(bp->b_flags, B_ASYNC); 441 VOP_BWRITE(bp); 442 } 443 444 /* 445 * Ordered block write; asynchronous, but I/O will occur in order queued. 446 */ 447 void 448 bowrite(bp) 449 struct buf *bp; 450 { 451 452 SET(bp->b_flags, B_ASYNC | B_ORDERED); 453 VOP_BWRITE(bp); 454 } 455 456 /* 457 * Same as first half of bdwrite, mark buffer dirty, but do not release it. 458 */ 459 void 460 bdirty(bp) 461 struct buf *bp; 462 { 463 struct proc *p = (curproc != NULL ? curproc : &proc0); /* XXX */ 464 int s; 465 466 s = splbio(); 467 468 CLR(bp->b_flags, B_AGE); 469 470 if (!ISSET(bp->b_flags, B_DELWRI)) { 471 SET(bp->b_flags, B_DELWRI); 472 p->p_stats->p_ru.ru_oublock++; 473 reassignbuf(bp, bp->b_vp); 474 } 475 476 splx(s); 477 } 478 479 /* 480 * Release a buffer on to the free lists. 481 * Described in Bach (p. 46). 482 */ 483 void 484 brelse(bp) 485 struct buf *bp; 486 { 487 struct bqueues *bufq; 488 int s; 489 490 KASSERT(ISSET(bp->b_flags, B_BUSY)); 491 492 /* Wake up any processes waiting for any buffer to become free. */ 493 if (needbuffer) { 494 needbuffer = 0; 495 wakeup(&needbuffer); 496 } 497 498 /* Block disk interrupts. */ 499 s = splbio(); 500 501 /* Wake up any proceeses waiting for _this_ buffer to become free. */ 502 if (ISSET(bp->b_flags, B_WANTED)) { 503 CLR(bp->b_flags, B_WANTED|B_AGE); 504 wakeup(bp); 505 } 506 507 /* 508 * Determine which queue the buffer should be on, then put it there. 509 */ 510 511 /* If it's locked, don't report an error; try again later. */ 512 if (ISSET(bp->b_flags, (B_LOCKED|B_ERROR)) == (B_LOCKED|B_ERROR)) 513 CLR(bp->b_flags, B_ERROR); 514 515 /* If it's not cacheable, or an error, mark it invalid. */ 516 if (ISSET(bp->b_flags, (B_NOCACHE|B_ERROR))) 517 SET(bp->b_flags, B_INVAL); 518 519 if (ISSET(bp->b_flags, B_VFLUSH)) { 520 /* 521 * This is a delayed write buffer that was just flushed to 522 * disk. It is still on the LRU queue. If it's become 523 * invalid, then we need to move it to a different queue; 524 * otherwise leave it in its current position. 525 */ 526 CLR(bp->b_flags, B_VFLUSH); 527 if (!ISSET(bp->b_flags, B_ERROR|B_INVAL|B_LOCKED|B_AGE)) 528 goto already_queued; 529 else 530 bremfree(bp); 531 } 532 533 if ((bp->b_bufsize <= 0) || ISSET(bp->b_flags, B_INVAL)) { 534 /* 535 * If it's invalid or empty, dissociate it from its vnode 536 * and put on the head of the appropriate queue. 537 */ 538 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 539 (*bioops.io_deallocate)(bp); 540 CLR(bp->b_flags, B_DONE|B_DELWRI); 541 if (bp->b_vp) { 542 reassignbuf(bp, bp->b_vp); 543 brelvp(bp); 544 } 545 if (bp->b_bufsize <= 0) 546 /* no data */ 547 bufq = &bufqueues[BQ_EMPTY]; 548 else 549 /* invalid data */ 550 bufq = &bufqueues[BQ_AGE]; 551 binsheadfree(bp, bufq); 552 } else { 553 /* 554 * It has valid data. Put it on the end of the appropriate 555 * queue, so that it'll stick around for as long as possible. 556 * If buf is AGE, but has dependencies, must put it on last 557 * bufqueue to be scanned, ie LRU. This protects against the 558 * livelock where BQ_AGE only has buffers with dependencies, 559 * and we thus never get to the dependent buffers in BQ_LRU. 560 */ 561 if (ISSET(bp->b_flags, B_LOCKED)) 562 /* locked in core */ 563 bufq = &bufqueues[BQ_LOCKED]; 564 else if (!ISSET(bp->b_flags, B_AGE)) 565 /* valid data */ 566 bufq = &bufqueues[BQ_LRU]; 567 else { 568 /* stale but valid data */ 569 int has_deps; 570 571 if (LIST_FIRST(&bp->b_dep) != NULL && 572 bioops.io_countdeps) 573 has_deps = (*bioops.io_countdeps)(bp, 0); 574 else 575 has_deps = 0; 576 bufq = has_deps ? &bufqueues[BQ_LRU] : 577 &bufqueues[BQ_AGE]; 578 } 579 binstailfree(bp, bufq); 580 } 581 582 already_queued: 583 /* Unlock the buffer. */ 584 CLR(bp->b_flags, B_AGE|B_ASYNC|B_BUSY|B_NOCACHE|B_ORDERED); 585 SET(bp->b_flags, B_CACHE); 586 587 /* Allow disk interrupts. */ 588 splx(s); 589 } 590 591 /* 592 * Determine if a block is in the cache. 593 * Just look on what would be its hash chain. If it's there, return 594 * a pointer to it, unless it's marked invalid. If it's marked invalid, 595 * we normally don't return the buffer, unless the caller explicitly 596 * wants us to. 597 */ 598 struct buf * 599 incore(vp, blkno) 600 struct vnode *vp; 601 daddr_t blkno; 602 { 603 struct buf *bp; 604 605 bp = BUFHASH(vp, blkno)->lh_first; 606 607 /* Search hash chain */ 608 for (; bp != NULL; bp = bp->b_hash.le_next) { 609 if (bp->b_lblkno == blkno && bp->b_vp == vp && 610 !ISSET(bp->b_flags, B_INVAL)) 611 return (bp); 612 } 613 614 return (NULL); 615 } 616 617 /* 618 * Get a block of requested size that is associated with 619 * a given vnode and block offset. If it is found in the 620 * block cache, mark it as having been found, make it busy 621 * and return it. Otherwise, return an empty block of the 622 * correct size. It is up to the caller to insure that the 623 * cached blocks be of the correct size. 624 */ 625 struct buf * 626 getblk(vp, blkno, size, slpflag, slptimeo) 627 struct vnode *vp; 628 daddr_t blkno; 629 int size, slpflag, slptimeo; 630 { 631 struct buf *bp; 632 int s, err; 633 634 start: 635 bp = incore(vp, blkno); 636 if (bp != NULL) { 637 s = splbio(); 638 if (ISSET(bp->b_flags, B_BUSY)) { 639 if (curproc == uvm.pagedaemon_proc) { 640 splx(s); 641 return NULL; 642 } 643 SET(bp->b_flags, B_WANTED); 644 err = tsleep(bp, slpflag | (PRIBIO + 1), "getblk", 645 slptimeo); 646 splx(s); 647 if (err) 648 return (NULL); 649 goto start; 650 } 651 #ifdef DIAGNOSTIC 652 if (ISSET(bp->b_flags, B_DONE|B_DELWRI) && 653 bp->b_bcount < size && vp->v_type != VBLK) 654 panic("getblk: block size invariant failed"); 655 #endif 656 SET(bp->b_flags, B_BUSY); 657 bremfree(bp); 658 splx(s); 659 } else { 660 if ((bp = getnewbuf(slpflag, slptimeo)) == NULL) 661 goto start; 662 663 binshash(bp, BUFHASH(vp, blkno)); 664 bp->b_blkno = bp->b_lblkno = bp->b_rawblkno = blkno; 665 s = splbio(); 666 bgetvp(vp, bp); 667 splx(s); 668 } 669 allocbuf(bp, size); 670 return (bp); 671 } 672 673 /* 674 * Get an empty, disassociated buffer of given size. 675 */ 676 struct buf * 677 geteblk(size) 678 int size; 679 { 680 struct buf *bp; 681 682 while ((bp = getnewbuf(0, 0)) == 0) 683 ; 684 SET(bp->b_flags, B_INVAL); 685 binshash(bp, &invalhash); 686 allocbuf(bp, size); 687 return (bp); 688 } 689 690 /* 691 * Expand or contract the actual memory allocated to a buffer. 692 * 693 * If the buffer shrinks, data is lost, so it's up to the 694 * caller to have written it out *first*; this routine will not 695 * start a write. If the buffer grows, it's the callers 696 * responsibility to fill out the buffer's additional contents. 697 */ 698 void 699 allocbuf(bp, size) 700 struct buf *bp; 701 int size; 702 { 703 struct buf *nbp; 704 vsize_t desired_size; 705 int s; 706 707 desired_size = round_page((vsize_t)size); 708 if (desired_size > MAXBSIZE) 709 panic("allocbuf: buffer larger than MAXBSIZE requested"); 710 711 if (bp->b_bufsize == desired_size) 712 goto out; 713 714 /* 715 * If the buffer is smaller than the desired size, we need to snarf 716 * it from other buffers. Get buffers (via getnewbuf()), and 717 * steal their pages. 718 */ 719 while (bp->b_bufsize < desired_size) { 720 int amt; 721 722 /* find a buffer */ 723 while ((nbp = getnewbuf(0, 0)) == NULL) 724 ; 725 726 SET(nbp->b_flags, B_INVAL); 727 binshash(nbp, &invalhash); 728 729 /* and steal its pages, up to the amount we need */ 730 amt = min(nbp->b_bufsize, (desired_size - bp->b_bufsize)); 731 pagemove((nbp->b_data + nbp->b_bufsize - amt), 732 bp->b_data + bp->b_bufsize, amt); 733 bp->b_bufsize += amt; 734 nbp->b_bufsize -= amt; 735 736 /* reduce transfer count if we stole some data */ 737 if (nbp->b_bcount > nbp->b_bufsize) 738 nbp->b_bcount = nbp->b_bufsize; 739 740 #ifdef DIAGNOSTIC 741 if (nbp->b_bufsize < 0) 742 panic("allocbuf: negative bufsize"); 743 #endif 744 745 brelse(nbp); 746 } 747 748 /* 749 * If we want a buffer smaller than the current size, 750 * shrink this buffer. Grab a buf head from the EMPTY queue, 751 * move a page onto it, and put it on front of the AGE queue. 752 * If there are no free buffer headers, leave the buffer alone. 753 */ 754 if (bp->b_bufsize > desired_size) { 755 s = splbio(); 756 if ((nbp = bufqueues[BQ_EMPTY].tqh_first) == NULL) { 757 /* No free buffer head */ 758 splx(s); 759 goto out; 760 } 761 bremfree(nbp); 762 SET(nbp->b_flags, B_BUSY); 763 splx(s); 764 765 /* move the page to it and note this change */ 766 pagemove(bp->b_data + desired_size, 767 nbp->b_data, bp->b_bufsize - desired_size); 768 nbp->b_bufsize = bp->b_bufsize - desired_size; 769 bp->b_bufsize = desired_size; 770 nbp->b_bcount = 0; 771 SET(nbp->b_flags, B_INVAL); 772 773 /* release the newly-filled buffer and leave */ 774 brelse(nbp); 775 } 776 777 out: 778 bp->b_bcount = size; 779 } 780 781 /* 782 * Find a buffer which is available for use. 783 * Select something from a free list. 784 * Preference is to AGE list, then LRU list. 785 */ 786 struct buf * 787 getnewbuf(slpflag, slptimeo) 788 int slpflag, slptimeo; 789 { 790 struct buf *bp; 791 int s; 792 793 start: 794 s = splbio(); 795 if ((bp = bufqueues[BQ_AGE].tqh_first) != NULL || 796 (bp = bufqueues[BQ_LRU].tqh_first) != NULL) { 797 bremfree(bp); 798 } else { 799 /* wait for a free buffer of any kind */ 800 needbuffer = 1; 801 tsleep(&needbuffer, slpflag|(PRIBIO+1), "getnewbuf", slptimeo); 802 splx(s); 803 return (NULL); 804 } 805 806 if (ISSET(bp->b_flags, B_VFLUSH)) { 807 /* 808 * This is a delayed write buffer being flushed to disk. Make 809 * sure it gets aged out of the queue when it's finished, and 810 * leave it off the LRU queue. 811 */ 812 CLR(bp->b_flags, B_VFLUSH); 813 SET(bp->b_flags, B_AGE); 814 splx(s); 815 goto start; 816 } 817 818 /* Buffer is no longer on free lists. */ 819 SET(bp->b_flags, B_BUSY); 820 821 /* 822 * If buffer was a delayed write, start it and return NULL 823 * (since we might sleep while starting the write). 824 */ 825 if (ISSET(bp->b_flags, B_DELWRI)) { 826 splx(s); 827 /* 828 * This buffer has gone through the LRU, so make sure it gets 829 * reused ASAP. 830 */ 831 SET(bp->b_flags, B_AGE); 832 bawrite(bp); 833 return (NULL); 834 } 835 836 /* disassociate us from our vnode, if we had one... */ 837 if (bp->b_vp) 838 brelvp(bp); 839 splx(s); 840 841 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 842 (*bioops.io_deallocate)(bp); 843 844 /* clear out various other fields */ 845 bp->b_flags = B_BUSY; 846 bp->b_dev = NODEV; 847 bp->b_blkno = bp->b_lblkno = bp->b_rawblkno = 0; 848 bp->b_iodone = 0; 849 bp->b_error = 0; 850 bp->b_resid = 0; 851 bp->b_bcount = 0; 852 853 bremhash(bp); 854 return (bp); 855 } 856 857 /* 858 * Wait for operations on the buffer to complete. 859 * When they do, extract and return the I/O's error value. 860 */ 861 int 862 biowait(bp) 863 struct buf *bp; 864 { 865 int s; 866 867 s = splbio(); 868 while (!ISSET(bp->b_flags, B_DONE | B_DELWRI)) 869 tsleep(bp, PRIBIO + 1, "biowait", 0); 870 splx(s); 871 872 /* check for interruption of I/O (e.g. via NFS), then errors. */ 873 if (ISSET(bp->b_flags, B_EINTR)) { 874 CLR(bp->b_flags, B_EINTR); 875 return (EINTR); 876 } else if (ISSET(bp->b_flags, B_ERROR)) 877 return (bp->b_error ? bp->b_error : EIO); 878 else 879 return (0); 880 } 881 882 /* 883 * Mark I/O complete on a buffer. 884 * 885 * If a callback has been requested, e.g. the pageout 886 * daemon, do so. Otherwise, awaken waiting processes. 887 * 888 * [ Leffler, et al., says on p.247: 889 * "This routine wakes up the blocked process, frees the buffer 890 * for an asynchronous write, or, for a request by the pagedaemon 891 * process, invokes a procedure specified in the buffer structure" ] 892 * 893 * In real life, the pagedaemon (or other system processes) wants 894 * to do async stuff to, and doesn't want the buffer brelse()'d. 895 * (for swap pager, that puts swap buffers on the free lists (!!!), 896 * for the vn device, that puts malloc'd buffers on the free lists!) 897 */ 898 void 899 biodone(bp) 900 struct buf *bp; 901 { 902 int s = splbio(); 903 904 if (ISSET(bp->b_flags, B_DONE)) 905 panic("biodone already"); 906 SET(bp->b_flags, B_DONE); /* note that it's done */ 907 908 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 909 (*bioops.io_complete)(bp); 910 911 if (!ISSET(bp->b_flags, B_READ)) /* wake up reader */ 912 vwakeup(bp); 913 914 if (ISSET(bp->b_flags, B_CALL)) { /* if necessary, call out */ 915 CLR(bp->b_flags, B_CALL); /* but note callout done */ 916 (*bp->b_iodone)(bp); 917 } else { 918 if (ISSET(bp->b_flags, B_ASYNC)) /* if async, release */ 919 brelse(bp); 920 else { /* or just wakeup the buffer */ 921 CLR(bp->b_flags, B_WANTED); 922 wakeup(bp); 923 } 924 } 925 926 splx(s); 927 } 928 929 /* 930 * Return a count of buffers on the "locked" queue. 931 */ 932 int 933 count_lock_queue() 934 { 935 struct buf *bp; 936 int n = 0; 937 938 for (bp = bufqueues[BQ_LOCKED].tqh_first; bp; 939 bp = bp->b_freelist.tqe_next) 940 n++; 941 return (n); 942 } 943 944 #ifdef DEBUG 945 /* 946 * Print out statistics on the current allocation of the buffer pool. 947 * Can be enabled to print out on every ``sync'' by setting "syncprt" 948 * in vfs_syscalls.c using sysctl. 949 */ 950 void 951 vfs_bufstats() 952 { 953 int s, i, j, count; 954 struct buf *bp; 955 struct bqueues *dp; 956 int counts[(MAXBSIZE / PAGE_SIZE) + 1]; 957 static char *bname[BQUEUES] = { "LOCKED", "LRU", "AGE", "EMPTY" }; 958 959 for (dp = bufqueues, i = 0; dp < &bufqueues[BQUEUES]; dp++, i++) { 960 count = 0; 961 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++) 962 counts[j] = 0; 963 s = splbio(); 964 for (bp = dp->tqh_first; bp; bp = bp->b_freelist.tqe_next) { 965 counts[bp->b_bufsize/PAGE_SIZE]++; 966 count++; 967 } 968 splx(s); 969 printf("%s: total-%d", bname[i], count); 970 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++) 971 if (counts[j] != 0) 972 printf(", %d-%d", j * PAGE_SIZE, counts[j]); 973 printf("\n"); 974 } 975 } 976 #endif /* DEBUG */ 977