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