Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]
FreeBSD/Linux Kernel Cross Reference
sys/kern/vfs_cluster.c
Version:
- FREEBSD - FREEBSD-13-STABLE - FREEBSD-13-0 - FREEBSD-12-STABLE - FREEBSD-12-0 - FREEBSD-11-STABLE - FREEBSD-11-0 - FREEBSD-10-STABLE - FREEBSD-10-0 - FREEBSD-9-STABLE - FREEBSD-9-0 - FREEBSD-8-STABLE - FREEBSD-8-0 - FREEBSD-7-STABLE - FREEBSD-7-0 - FREEBSD-6-STABLE - FREEBSD-6-0 - FREEBSD-5-STABLE - FREEBSD-5-0 - FREEBSD-4-STABLE - FREEBSD-3-STABLE - FREEBSD22 - l41 - OPENBSD - linux-2.6 - MK84 - PLAN9 - xnu-8792
SearchContext: - none - 3 - 10
SearchContext: - none - 3 - 10
1 /*- 2 * Copyright (c) 1993 3 * The Regents of the University of California. All rights reserved. 4 * Modifications/enhancements: 5 * Copyright (c) 1995 John S. Dyson. All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 4. Neither the name of the University nor the names of its contributors 16 * may be used to endorse or promote products derived from this software 17 * without specific prior written permission. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 29 * SUCH DAMAGE. 30 * 31 * @(#)vfs_cluster.c 8.7 (Berkeley) 2/13/94 32 */ 33 34 #include <sys/cdefs.h> 35 __FBSDID("$FreeBSD: releng/5.4/sys/kern/vfs_cluster.c 145335 2005-04-20 19:11:07Z cvs2svn $"); 36 37 #include "opt_debug_cluster.h" 38 39 #include <sys/param.h> 40 #include <sys/systm.h> 41 #include <sys/kernel.h> 42 #include <sys/proc.h> 43 #include <sys/bio.h> 44 #include <sys/buf.h> 45 #include <sys/vnode.h> 46 #include <sys/malloc.h> 47 #include <sys/mount.h> 48 #include <sys/resourcevar.h> 49 #include <sys/vmmeter.h> 50 #include <vm/vm.h> 51 #include <vm/vm_object.h> 52 #include <vm/vm_page.h> 53 #include <sys/sysctl.h> 54 55 #if defined(CLUSTERDEBUG) 56 static int rcluster= 0; 57 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, 58 "Debug VFS clustering code"); 59 #endif 60 61 static MALLOC_DEFINE(M_SEGMENT, "cluster_save buffer", "cluster_save buffer"); 62 63 static struct cluster_save * 64 cluster_collectbufs(struct vnode *vp, struct buf *last_bp); 65 static struct buf * 66 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn, 67 daddr_t blkno, long size, int run, struct buf *fbp); 68 69 static int write_behind = 1; 70 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, 71 "Cluster write-behind; 0: disable, 1: enable, 2: backed off"); 72 73 static int read_max = 8; 74 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0, 75 "Cluster read-ahead max block count"); 76 77 /* Page expended to mark partially backed buffers */ 78 extern vm_page_t bogus_page; 79 80 /* 81 * Number of physical bufs (pbufs) this subsystem is allowed. 82 * Manipulated by vm_pager.c 83 */ 84 extern int cluster_pbuf_freecnt; 85 86 /* 87 * Read data to a buf, including read-ahead if we find this to be beneficial. 88 * cluster_read replaces bread. 89 */ 90 int 91 cluster_read(vp, filesize, lblkno, size, cred, totread, seqcount, bpp) 92 struct vnode *vp; 93 u_quad_t filesize; 94 daddr_t lblkno; 95 long size; 96 struct ucred *cred; 97 long totread; 98 int seqcount; 99 struct buf **bpp; 100 { 101 struct buf *bp, *rbp, *reqbp; 102 daddr_t blkno, origblkno; 103 int maxra, racluster; 104 int error, ncontig; 105 int i; 106 107 error = 0; 108 109 /* 110 * Try to limit the amount of read-ahead by a few 111 * ad-hoc parameters. This needs work!!! 112 */ 113 racluster = vp->v_mount->mnt_iosize_max / size; 114 maxra = seqcount; 115 maxra = min(read_max, maxra); 116 maxra = min(nbuf/8, maxra); 117 if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize) 118 maxra = (filesize / size) - lblkno; 119 120 /* 121 * get the requested block 122 */ 123 *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, 0); 124 origblkno = lblkno; 125 126 /* 127 * if it is in the cache, then check to see if the reads have been 128 * sequential. If they have, then try some read-ahead, otherwise 129 * back-off on prospective read-aheads. 130 */ 131 if (bp->b_flags & B_CACHE) { 132 if (!seqcount) { 133 return 0; 134 } else if ((bp->b_flags & B_RAM) == 0) { 135 return 0; 136 } else { 137 int s; 138 bp->b_flags &= ~B_RAM; 139 /* 140 * We do the spl here so that there is no window 141 * between the incore and the b_usecount increment 142 * below. We opt to keep the spl out of the loop 143 * for efficiency. 144 */ 145 s = splbio(); 146 VI_LOCK(vp); 147 for (i = 1; i < maxra; i++) { 148 /* 149 * Stop if the buffer does not exist or it 150 * is invalid (about to go away?) 151 */ 152 rbp = gbincore(vp, lblkno+i); 153 if (rbp == NULL || (rbp->b_flags & B_INVAL)) 154 break; 155 156 /* 157 * Set another read-ahead mark so we know 158 * to check again. 159 */ 160 if (((i % racluster) == (racluster - 1)) || 161 (i == (maxra - 1))) 162 rbp->b_flags |= B_RAM; 163 } 164 VI_UNLOCK(vp); 165 splx(s); 166 if (i >= maxra) { 167 return 0; 168 } 169 lblkno += i; 170 } 171 reqbp = bp = NULL; 172 /* 173 * If it isn't in the cache, then get a chunk from 174 * disk if sequential, otherwise just get the block. 175 */ 176 } else { 177 off_t firstread = bp->b_offset; 178 int nblks; 179 180 KASSERT(bp->b_offset != NOOFFSET, 181 ("cluster_read: no buffer offset")); 182 183 ncontig = 0; 184 185 /* 186 * Compute the total number of blocks that we should read 187 * synchronously. 188 */ 189 if (firstread + totread > filesize) 190 totread = filesize - firstread; 191 nblks = howmany(totread, size); 192 if (nblks > racluster) 193 nblks = racluster; 194 195 /* 196 * Now compute the number of contiguous blocks. 197 */ 198 if (nblks > 1) { 199 error = VOP_BMAP(vp, lblkno, NULL, 200 &blkno, &ncontig, NULL); 201 /* 202 * If this failed to map just do the original block. 203 */ 204 if (error || blkno == -1) 205 ncontig = 0; 206 } 207 208 /* 209 * If we have contiguous data available do a cluster 210 * otherwise just read the requested block. 211 */ 212 if (ncontig) { 213 /* Account for our first block. */ 214 ncontig = min(ncontig + 1, nblks); 215 if (ncontig < nblks) 216 nblks = ncontig; 217 bp = cluster_rbuild(vp, filesize, lblkno, 218 blkno, size, nblks, bp); 219 lblkno += (bp->b_bufsize / size); 220 } else { 221 bp->b_flags |= B_RAM; 222 bp->b_iocmd = BIO_READ; 223 lblkno += 1; 224 } 225 } 226 227 /* 228 * handle the synchronous read so that it is available ASAP. 229 */ 230 if (bp) { 231 if ((bp->b_flags & B_CLUSTER) == 0) { 232 vfs_busy_pages(bp, 0); 233 } 234 bp->b_flags &= ~B_INVAL; 235 bp->b_ioflags &= ~BIO_ERROR; 236 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL) 237 BUF_KERNPROC(bp); 238 bp->b_iooffset = dbtob(bp->b_blkno); 239 error = VOP_STRATEGY(vp, bp); 240 curproc->p_stats->p_ru.ru_inblock++; 241 if (error) 242 return (error); 243 } 244 245 /* 246 * If we have been doing sequential I/O, then do some read-ahead. 247 */ 248 while (lblkno < (origblkno + maxra)) { 249 error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL); 250 if (error) 251 break; 252 253 if (blkno == -1) 254 break; 255 256 /* 257 * We could throttle ncontig here by maxra but we might as 258 * well read the data if it is contiguous. We're throttled 259 * by racluster anyway. 260 */ 261 if (ncontig) { 262 ncontig = min(ncontig + 1, racluster); 263 rbp = cluster_rbuild(vp, filesize, lblkno, blkno, 264 size, ncontig, NULL); 265 lblkno += (rbp->b_bufsize / size); 266 if (rbp->b_flags & B_DELWRI) { 267 bqrelse(rbp); 268 continue; 269 } 270 } else { 271 rbp = getblk(vp, lblkno, size, 0, 0, 0); 272 lblkno += 1; 273 if (rbp->b_flags & B_DELWRI) { 274 bqrelse(rbp); 275 continue; 276 } 277 rbp->b_flags |= B_ASYNC | B_RAM; 278 rbp->b_iocmd = BIO_READ; 279 rbp->b_blkno = blkno; 280 } 281 if (rbp->b_flags & B_CACHE) { 282 rbp->b_flags &= ~B_ASYNC; 283 bqrelse(rbp); 284 continue; 285 } 286 if ((rbp->b_flags & B_CLUSTER) == 0) { 287 vfs_busy_pages(rbp, 0); 288 } 289 rbp->b_flags &= ~B_INVAL; 290 rbp->b_ioflags &= ~BIO_ERROR; 291 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL) 292 BUF_KERNPROC(rbp); 293 rbp->b_iooffset = dbtob(rbp->b_blkno); 294 (void) VOP_STRATEGY(vp, rbp); 295 curproc->p_stats->p_ru.ru_inblock++; 296 } 297 298 if (reqbp) 299 return (bufwait(reqbp)); 300 else 301 return (error); 302 } 303 304 /* 305 * If blocks are contiguous on disk, use this to provide clustered 306 * read ahead. We will read as many blocks as possible sequentially 307 * and then parcel them up into logical blocks in the buffer hash table. 308 */ 309 static struct buf * 310 cluster_rbuild(vp, filesize, lbn, blkno, size, run, fbp) 311 struct vnode *vp; 312 u_quad_t filesize; 313 daddr_t lbn; 314 daddr_t blkno; 315 long size; 316 int run; 317 struct buf *fbp; 318 { 319 struct buf *bp, *tbp; 320 daddr_t bn; 321 int i, inc, j; 322 323 GIANT_REQUIRED; 324 325 KASSERT(size == vp->v_mount->mnt_stat.f_iosize, 326 ("cluster_rbuild: size %ld != filesize %jd\n", 327 size, (intmax_t)vp->v_mount->mnt_stat.f_iosize)); 328 329 /* 330 * avoid a division 331 */ 332 while ((u_quad_t) size * (lbn + run) > filesize) { 333 --run; 334 } 335 336 if (fbp) { 337 tbp = fbp; 338 tbp->b_iocmd = BIO_READ; 339 } else { 340 tbp = getblk(vp, lbn, size, 0, 0, 0); 341 if (tbp->b_flags & B_CACHE) 342 return tbp; 343 tbp->b_flags |= B_ASYNC | B_RAM; 344 tbp->b_iocmd = BIO_READ; 345 } 346 347 tbp->b_blkno = blkno; 348 if( (tbp->b_flags & B_MALLOC) || 349 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) ) 350 return tbp; 351 352 bp = trypbuf(&cluster_pbuf_freecnt); 353 if (bp == 0) 354 return tbp; 355 356 /* 357 * We are synthesizing a buffer out of vm_page_t's, but 358 * if the block size is not page aligned then the starting 359 * address may not be either. Inherit the b_data offset 360 * from the original buffer. 361 */ 362 bp->b_data = (char *)((vm_offset_t)bp->b_data | 363 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 364 bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO; 365 bp->b_iocmd = BIO_READ; 366 bp->b_iodone = cluster_callback; 367 bp->b_blkno = blkno; 368 bp->b_lblkno = lbn; 369 bp->b_offset = tbp->b_offset; 370 KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset")); 371 pbgetvp(vp, bp); 372 373 TAILQ_INIT(&bp->b_cluster.cluster_head); 374 375 bp->b_bcount = 0; 376 bp->b_bufsize = 0; 377 bp->b_npages = 0; 378 379 inc = btodb(size); 380 for (bn = blkno, i = 0; i < run; ++i, bn += inc) { 381 if (i != 0) { 382 if ((bp->b_npages * PAGE_SIZE) + 383 round_page(size) > vp->v_mount->mnt_iosize_max) { 384 break; 385 } 386 387 tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT); 388 389 /* Don't wait around for locked bufs. */ 390 if (tbp == NULL) 391 break; 392 393 /* 394 * Stop scanning if the buffer is fully valid 395 * (marked B_CACHE), or locked (may be doing a 396 * background write), or if the buffer is not 397 * VMIO backed. The clustering code can only deal 398 * with VMIO-backed buffers. 399 */ 400 VI_LOCK(bp->b_vp); 401 if ((tbp->b_vflags & BV_BKGRDINPROG) || 402 (tbp->b_flags & B_CACHE) || 403 (tbp->b_flags & B_VMIO) == 0) { 404 VI_UNLOCK(bp->b_vp); 405 bqrelse(tbp); 406 break; 407 } 408 VI_UNLOCK(bp->b_vp); 409 410 /* 411 * The buffer must be completely invalid in order to 412 * take part in the cluster. If it is partially valid 413 * then we stop. 414 */ 415 VM_OBJECT_LOCK(tbp->b_object); 416 for (j = 0;j < tbp->b_npages; j++) { 417 VM_OBJECT_LOCK_ASSERT(tbp->b_pages[j]->object, 418 MA_OWNED); 419 if (tbp->b_pages[j]->valid) 420 break; 421 } 422 VM_OBJECT_UNLOCK(tbp->b_object); 423 if (j != tbp->b_npages) { 424 bqrelse(tbp); 425 break; 426 } 427 428 /* 429 * Set a read-ahead mark as appropriate 430 */ 431 if ((fbp && (i == 1)) || (i == (run - 1))) 432 tbp->b_flags |= B_RAM; 433 434 /* 435 * Set the buffer up for an async read (XXX should 436 * we do this only if we do not wind up brelse()ing?). 437 * Set the block number if it isn't set, otherwise 438 * if it is make sure it matches the block number we 439 * expect. 440 */ 441 tbp->b_flags |= B_ASYNC; 442 tbp->b_iocmd = BIO_READ; 443 if (tbp->b_blkno == tbp->b_lblkno) { 444 tbp->b_blkno = bn; 445 } else if (tbp->b_blkno != bn) { 446 brelse(tbp); 447 break; 448 } 449 } 450 /* 451 * XXX fbp from caller may not be B_ASYNC, but we are going 452 * to biodone() it in cluster_callback() anyway 453 */ 454 BUF_KERNPROC(tbp); 455 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, 456 tbp, b_cluster.cluster_entry); 457 VM_OBJECT_LOCK(tbp->b_object); 458 for (j = 0; j < tbp->b_npages; j += 1) { 459 vm_page_t m; 460 m = tbp->b_pages[j]; 461 vm_page_io_start(m); 462 vm_object_pip_add(m->object, 1); 463 if ((bp->b_npages == 0) || 464 (bp->b_pages[bp->b_npages-1] != m)) { 465 bp->b_pages[bp->b_npages] = m; 466 bp->b_npages++; 467 } 468 if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) 469 tbp->b_pages[j] = bogus_page; 470 } 471 VM_OBJECT_UNLOCK(tbp->b_object); 472 /* 473 * XXX shouldn't this be += size for both, like in 474 * cluster_wbuild()? 475 * 476 * Don't inherit tbp->b_bufsize as it may be larger due to 477 * a non-page-aligned size. Instead just aggregate using 478 * 'size'. 479 */ 480 if (tbp->b_bcount != size) 481 printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size); 482 if (tbp->b_bufsize != size) 483 printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size); 484 bp->b_bcount += size; 485 bp->b_bufsize += size; 486 } 487 488 /* 489 * Fully valid pages in the cluster are already good and do not need 490 * to be re-read from disk. Replace the page with bogus_page 491 */ 492 VM_OBJECT_LOCK(bp->b_object); 493 for (j = 0; j < bp->b_npages; j++) { 494 VM_OBJECT_LOCK_ASSERT(bp->b_pages[j]->object, MA_OWNED); 495 if ((bp->b_pages[j]->valid & VM_PAGE_BITS_ALL) == 496 VM_PAGE_BITS_ALL) { 497 bp->b_pages[j] = bogus_page; 498 } 499 } 500 VM_OBJECT_UNLOCK(bp->b_object); 501 if (bp->b_bufsize > bp->b_kvasize) 502 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n", 503 bp->b_bufsize, bp->b_kvasize); 504 bp->b_kvasize = bp->b_bufsize; 505 506 pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 507 (vm_page_t *)bp->b_pages, bp->b_npages); 508 return (bp); 509 } 510 511 /* 512 * Cleanup after a clustered read or write. 513 * This is complicated by the fact that any of the buffers might have 514 * extra memory (if there were no empty buffer headers at allocbuf time) 515 * that we will need to shift around. 516 */ 517 void 518 cluster_callback(bp) 519 struct buf *bp; 520 { 521 struct buf *nbp, *tbp; 522 int error = 0; 523 524 GIANT_REQUIRED; 525 526 /* 527 * Must propogate errors to all the components. 528 */ 529 if (bp->b_ioflags & BIO_ERROR) 530 error = bp->b_error; 531 532 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 533 /* 534 * Move memory from the large cluster buffer into the component 535 * buffers and mark IO as done on these. 536 */ 537 for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head); 538 tbp; tbp = nbp) { 539 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry); 540 if (error) { 541 tbp->b_ioflags |= BIO_ERROR; 542 tbp->b_error = error; 543 } else { 544 tbp->b_dirtyoff = tbp->b_dirtyend = 0; 545 tbp->b_flags &= ~B_INVAL; 546 tbp->b_ioflags &= ~BIO_ERROR; 547 /* 548 * XXX the bdwrite()/bqrelse() issued during 549 * cluster building clears B_RELBUF (see bqrelse() 550 * comment). If direct I/O was specified, we have 551 * to restore it here to allow the buffer and VM 552 * to be freed. 553 */ 554 if (tbp->b_flags & B_DIRECT) 555 tbp->b_flags |= B_RELBUF; 556 } 557 bufdone(tbp); 558 } 559 relpbuf(bp, &cluster_pbuf_freecnt); 560 } 561 562 /* 563 * cluster_wbuild_wb: 564 * 565 * Implement modified write build for cluster. 566 * 567 * write_behind = 0 write behind disabled 568 * write_behind = 1 write behind normal (default) 569 * write_behind = 2 write behind backed-off 570 */ 571 572 static __inline int 573 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len) 574 { 575 int r = 0; 576 577 switch(write_behind) { 578 case 2: 579 if (start_lbn < len) 580 break; 581 start_lbn -= len; 582 /* FALLTHROUGH */ 583 case 1: 584 r = cluster_wbuild(vp, size, start_lbn, len); 585 /* FALLTHROUGH */ 586 default: 587 /* FALLTHROUGH */ 588 break; 589 } 590 return(r); 591 } 592 593 /* 594 * Do clustered write for FFS. 595 * 596 * Three cases: 597 * 1. Write is not sequential (write asynchronously) 598 * Write is sequential: 599 * 2. beginning of cluster - begin cluster 600 * 3. middle of a cluster - add to cluster 601 * 4. end of a cluster - asynchronously write cluster 602 */ 603 void 604 cluster_write(bp, filesize, seqcount) 605 struct buf *bp; 606 u_quad_t filesize; 607 int seqcount; 608 { 609 struct vnode *vp; 610 daddr_t lbn; 611 int maxclen, cursize; 612 int lblocksize; 613 int async; 614 615 vp = bp->b_vp; 616 if (vp->v_type == VREG) { 617 async = vp->v_mount->mnt_flag & MNT_ASYNC; 618 lblocksize = vp->v_mount->mnt_stat.f_iosize; 619 } else { 620 async = 0; 621 lblocksize = bp->b_bufsize; 622 } 623 lbn = bp->b_lblkno; 624 KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset")); 625 626 /* Initialize vnode to beginning of file. */ 627 if (lbn == 0) 628 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; 629 630 if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 || 631 (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) { 632 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1; 633 if (vp->v_clen != 0) { 634 /* 635 * Next block is not sequential. 636 * 637 * If we are not writing at end of file, the process 638 * seeked to another point in the file since its last 639 * write, or we have reached our maximum cluster size, 640 * then push the previous cluster. Otherwise try 641 * reallocating to make it sequential. 642 * 643 * Change to algorithm: only push previous cluster if 644 * it was sequential from the point of view of the 645 * seqcount heuristic, otherwise leave the buffer 646 * intact so we can potentially optimize the I/O 647 * later on in the buf_daemon or update daemon 648 * flush. 649 */ 650 cursize = vp->v_lastw - vp->v_cstart + 1; 651 if (((u_quad_t) bp->b_offset + lblocksize) != filesize || 652 lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) { 653 if (!async && seqcount > 0) { 654 cluster_wbuild_wb(vp, lblocksize, 655 vp->v_cstart, cursize); 656 } 657 } else { 658 struct buf **bpp, **endbp; 659 struct cluster_save *buflist; 660 661 buflist = cluster_collectbufs(vp, bp); 662 endbp = &buflist->bs_children 663 [buflist->bs_nchildren - 1]; 664 if (VOP_REALLOCBLKS(vp, buflist)) { 665 /* 666 * Failed, push the previous cluster 667 * if *really* writing sequentially 668 * in the logical file (seqcount > 1), 669 * otherwise delay it in the hopes that 670 * the low level disk driver can 671 * optimize the write ordering. 672 */ 673 for (bpp = buflist->bs_children; 674 bpp < endbp; bpp++) 675 brelse(*bpp); 676 free(buflist, M_SEGMENT); 677 if (seqcount > 1) { 678 cluster_wbuild_wb(vp, 679 lblocksize, vp->v_cstart, 680 cursize); 681 } 682 } else { 683 /* 684 * Succeeded, keep building cluster. 685 */ 686 for (bpp = buflist->bs_children; 687 bpp <= endbp; bpp++) 688 bdwrite(*bpp); 689 free(buflist, M_SEGMENT); 690 vp->v_lastw = lbn; 691 vp->v_lasta = bp->b_blkno; 692 return; 693 } 694 } 695 } 696 /* 697 * Consider beginning a cluster. If at end of file, make 698 * cluster as large as possible, otherwise find size of 699 * existing cluster. 700 */ 701 if ((vp->v_type == VREG) && 702 ((u_quad_t) bp->b_offset + lblocksize) != filesize && 703 (bp->b_blkno == bp->b_lblkno) && 704 (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) || 705 bp->b_blkno == -1)) { 706 bawrite(bp); 707 vp->v_clen = 0; 708 vp->v_lasta = bp->b_blkno; 709 vp->v_cstart = lbn + 1; 710 vp->v_lastw = lbn; 711 return; 712 } 713 vp->v_clen = maxclen; 714 if (!async && maxclen == 0) { /* I/O not contiguous */ 715 vp->v_cstart = lbn + 1; 716 bawrite(bp); 717 } else { /* Wait for rest of cluster */ 718 vp->v_cstart = lbn; 719 bdwrite(bp); 720 } 721 } else if (lbn == vp->v_cstart + vp->v_clen) { 722 /* 723 * At end of cluster, write it out if seqcount tells us we 724 * are operating sequentially, otherwise let the buf or 725 * update daemon handle it. 726 */ 727 bdwrite(bp); 728 if (seqcount > 1) 729 cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1); 730 vp->v_clen = 0; 731 vp->v_cstart = lbn + 1; 732 } else if (vm_page_count_severe()) { 733 /* 734 * We are low on memory, get it going NOW 735 */ 736 bawrite(bp); 737 } else { 738 /* 739 * In the middle of a cluster, so just delay the I/O for now. 740 */ 741 bdwrite(bp); 742 } 743 vp->v_lastw = lbn; 744 vp->v_lasta = bp->b_blkno; 745 } 746 747 748 /* 749 * This is an awful lot like cluster_rbuild...wish they could be combined. 750 * The last lbn argument is the current block on which I/O is being 751 * performed. Check to see that it doesn't fall in the middle of 752 * the current block (if last_bp == NULL). 753 */ 754 int 755 cluster_wbuild(vp, size, start_lbn, len) 756 struct vnode *vp; 757 long size; 758 daddr_t start_lbn; 759 int len; 760 { 761 struct buf *bp, *tbp; 762 int i, j, s; 763 int totalwritten = 0; 764 int dbsize = btodb(size); 765 766 GIANT_REQUIRED; 767 768 while (len > 0) { 769 s = splbio(); 770 /* 771 * If the buffer is not delayed-write (i.e. dirty), or it 772 * is delayed-write but either locked or inval, it cannot 773 * partake in the clustered write. 774 */ 775 VI_LOCK(vp); 776 if ((tbp = gbincore(vp, start_lbn)) == NULL || 777 (tbp->b_vflags & BV_BKGRDINPROG)) { 778 VI_UNLOCK(vp); 779 ++start_lbn; 780 --len; 781 splx(s); 782 continue; 783 } 784 if (BUF_LOCK(tbp, 785 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, VI_MTX(vp))) { 786 ++start_lbn; 787 --len; 788 splx(s); 789 continue; 790 } 791 if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) { 792 BUF_UNLOCK(tbp); 793 ++start_lbn; 794 --len; 795 splx(s); 796 continue; 797 } 798 bremfree(tbp); 799 tbp->b_flags &= ~B_DONE; 800 splx(s); 801 802 /* 803 * Extra memory in the buffer, punt on this buffer. 804 * XXX we could handle this in most cases, but we would 805 * have to push the extra memory down to after our max 806 * possible cluster size and then potentially pull it back 807 * up if the cluster was terminated prematurely--too much 808 * hassle. 809 */ 810 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) != 811 (B_CLUSTEROK | B_VMIO)) || 812 (tbp->b_bcount != tbp->b_bufsize) || 813 (tbp->b_bcount != size) || 814 (len == 1) || 815 ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) { 816 totalwritten += tbp->b_bufsize; 817 bawrite(tbp); 818 ++start_lbn; 819 --len; 820 continue; 821 } 822 823 /* 824 * We got a pbuf to make the cluster in. 825 * so initialise it. 826 */ 827 TAILQ_INIT(&bp->b_cluster.cluster_head); 828 bp->b_bcount = 0; 829 bp->b_magic = tbp->b_magic; 830 bp->b_op = tbp->b_op; 831 bp->b_bufsize = 0; 832 bp->b_npages = 0; 833 if (tbp->b_wcred != NOCRED) 834 bp->b_wcred = crhold(tbp->b_wcred); 835 836 bp->b_blkno = tbp->b_blkno; 837 bp->b_lblkno = tbp->b_lblkno; 838 bp->b_offset = tbp->b_offset; 839 840 /* 841 * We are synthesizing a buffer out of vm_page_t's, but 842 * if the block size is not page aligned then the starting 843 * address may not be either. Inherit the b_data offset 844 * from the original buffer. 845 */ 846 bp->b_data = (char *)((vm_offset_t)bp->b_data | 847 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 848 bp->b_flags |= B_CLUSTER | 849 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT)); 850 bp->b_iodone = cluster_callback; 851 pbgetvp(vp, bp); 852 /* 853 * From this location in the file, scan forward to see 854 * if there are buffers with adjacent data that need to 855 * be written as well. 856 */ 857 for (i = 0; i < len; ++i, ++start_lbn) { 858 if (i != 0) { /* If not the first buffer */ 859 s = splbio(); 860 /* 861 * If the adjacent data is not even in core it 862 * can't need to be written. 863 */ 864 VI_LOCK(vp); 865 if ((tbp = gbincore(vp, start_lbn)) == NULL || 866 (tbp->b_vflags & BV_BKGRDINPROG)) { 867 VI_UNLOCK(vp); 868 splx(s); 869 break; 870 } 871 872 /* 873 * If it IS in core, but has different 874 * characteristics, or is locked (which 875 * means it could be undergoing a background 876 * I/O or be in a weird state), then don't 877 * cluster with it. 878 */ 879 if (BUF_LOCK(tbp, 880 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, 881 VI_MTX(vp))) { 882 splx(s); 883 break; 884 } 885 886 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK | 887 B_INVAL | B_DELWRI | B_NEEDCOMMIT)) 888 != (B_DELWRI | B_CLUSTEROK | 889 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) || 890 tbp->b_wcred != bp->b_wcred) { 891 BUF_UNLOCK(tbp); 892 splx(s); 893 break; 894 } 895 896 /* 897 * Check that the combined cluster 898 * would make sense with regard to pages 899 * and would not be too large 900 */ 901 if ((tbp->b_bcount != size) || 902 ((bp->b_blkno + (dbsize * i)) != 903 tbp->b_blkno) || 904 ((tbp->b_npages + bp->b_npages) > 905 (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) { 906 BUF_UNLOCK(tbp); 907 splx(s); 908 break; 909 } 910 /* 911 * Ok, it's passed all the tests, 912 * so remove it from the free list 913 * and mark it busy. We will use it. 914 */ 915 bremfree(tbp); 916 tbp->b_flags &= ~B_DONE; 917 splx(s); 918 } /* end of code for non-first buffers only */ 919 /* check for latent dependencies to be handled */ 920 if ((LIST_FIRST(&tbp->b_dep)) != NULL) { 921 tbp->b_iocmd = BIO_WRITE; 922 buf_start(tbp); 923 } 924 /* 925 * If the IO is via the VM then we do some 926 * special VM hackery (yuck). Since the buffer's 927 * block size may not be page-aligned it is possible 928 * for a page to be shared between two buffers. We 929 * have to get rid of the duplication when building 930 * the cluster. 931 */ 932 if (tbp->b_flags & B_VMIO) { 933 vm_page_t m; 934 935 VM_OBJECT_LOCK(tbp->b_object); 936 if (i != 0) { /* if not first buffer */ 937 for (j = 0; j < tbp->b_npages; j += 1) { 938 m = tbp->b_pages[j]; 939 if (m->flags & PG_BUSY) { 940 VM_OBJECT_UNLOCK( 941 tbp->b_object); 942 bqrelse(tbp); 943 goto finishcluster; 944 } 945 } 946 } 947 for (j = 0; j < tbp->b_npages; j += 1) { 948 m = tbp->b_pages[j]; 949 vm_page_io_start(m); 950 vm_object_pip_add(m->object, 1); 951 if ((bp->b_npages == 0) || 952 (bp->b_pages[bp->b_npages - 1] != m)) { 953 bp->b_pages[bp->b_npages] = m; 954 bp->b_npages++; 955 } 956 } 957 VM_OBJECT_UNLOCK(tbp->b_object); 958 } 959 bp->b_bcount += size; 960 bp->b_bufsize += size; 961 962 s = splbio(); 963 bundirty(tbp); 964 tbp->b_flags &= ~B_DONE; 965 tbp->b_ioflags &= ~BIO_ERROR; 966 tbp->b_flags |= B_ASYNC; 967 tbp->b_iocmd = BIO_WRITE; 968 reassignbuf(tbp); /* put on clean list */ 969 VI_LOCK(tbp->b_vp); 970 ++tbp->b_vp->v_numoutput; 971 VI_UNLOCK(tbp->b_vp); 972 splx(s); 973 BUF_KERNPROC(tbp); 974 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, 975 tbp, b_cluster.cluster_entry); 976 } 977 finishcluster: 978 pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 979 (vm_page_t *) bp->b_pages, bp->b_npages); 980 if (bp->b_bufsize > bp->b_kvasize) 981 panic( 982 "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n", 983 bp->b_bufsize, bp->b_kvasize); 984 bp->b_kvasize = bp->b_bufsize; 985 totalwritten += bp->b_bufsize; 986 bp->b_dirtyoff = 0; 987 bp->b_dirtyend = bp->b_bufsize; 988 bawrite(bp); 989 990 len -= i; 991 } 992 return totalwritten; 993 } 994 995 /* 996 * Collect together all the buffers in a cluster. 997 * Plus add one additional buffer. 998 */ 999 static struct cluster_save * 1000 cluster_collectbufs(vp, last_bp) 1001 struct vnode *vp; 1002 struct buf *last_bp; 1003 { 1004 struct cluster_save *buflist; 1005 struct buf *bp; 1006 daddr_t lbn; 1007 int i, len; 1008 1009 len = vp->v_lastw - vp->v_cstart + 1; 1010 buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist), 1011 M_SEGMENT, M_WAITOK); 1012 buflist->bs_nchildren = 0; 1013 buflist->bs_children = (struct buf **) (buflist + 1); 1014 for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) { 1015 (void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp); 1016 buflist->bs_children[i] = bp; 1017 if (bp->b_blkno == bp->b_lblkno) 1018 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, 1019 NULL, NULL); 1020 } 1021 buflist->bs_children[i] = bp = last_bp; 1022 if (bp->b_blkno == bp->b_lblkno) 1023 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, 1024 NULL, NULL); 1025 buflist->bs_nchildren = i + 1; 1026 return (buflist); 1027 }
Cache object: 630ff29172e1219d704d67365ceaf6ea
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]