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 */ 37 38 #include <sys/cdefs.h> 39 __FBSDID("$FreeBSD$"); 40 41 #include "opt_debug_cluster.h" 42 43 #include <sys/param.h> 44 #include <sys/systm.h> 45 #include <sys/kernel.h> 46 #include <sys/proc.h> 47 #include <sys/bio.h> 48 #include <sys/buf.h> 49 #include <sys/vnode.h> 50 #include <sys/malloc.h> 51 #include <sys/mount.h> 52 #include <sys/resourcevar.h> 53 #include <sys/vmmeter.h> 54 #include <vm/vm.h> 55 #include <vm/vm_object.h> 56 #include <vm/vm_page.h> 57 #include <sys/sysctl.h> 58 59 #if defined(CLUSTERDEBUG) 60 static int rcluster= 0; 61 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, 62 "Debug VFS clustering code"); 63 #endif 64 65 static MALLOC_DEFINE(M_SEGMENT, "cluster_save buffer", "cluster_save buffer"); 66 67 static struct cluster_save * 68 cluster_collectbufs(struct vnode *vp, struct buf *last_bp); 69 static struct buf * 70 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn, 71 daddr_t blkno, long size, int run, struct buf *fbp); 72 73 static int write_behind = 1; 74 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, 75 "Cluster write-behind; 0: disable, 1: enable, 2: backed off"); 76 77 static int read_max = 8; 78 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0, 79 "Cluster read-ahead max block count"); 80 81 /* Page expended to mark partially backed buffers */ 82 extern vm_page_t bogus_page; 83 84 /* 85 * Number of physical bufs (pbufs) this subsystem is allowed. 86 * Manipulated by vm_pager.c 87 */ 88 extern int cluster_pbuf_freecnt; 89 90 /* 91 * Read data to a buf, including read-ahead if we find this to be beneficial. 92 * cluster_read replaces bread. 93 */ 94 int 95 cluster_read(vp, filesize, lblkno, size, cred, totread, seqcount, bpp) 96 struct vnode *vp; 97 u_quad_t filesize; 98 daddr_t lblkno; 99 long size; 100 struct ucred *cred; 101 long totread; 102 int seqcount; 103 struct buf **bpp; 104 { 105 struct buf *bp, *rbp, *reqbp; 106 daddr_t blkno, origblkno; 107 int maxra, racluster; 108 int error, ncontig; 109 int i; 110 111 error = 0; 112 113 /* 114 * Try to limit the amount of read-ahead by a few 115 * ad-hoc parameters. This needs work!!! 116 */ 117 racluster = vp->v_mount->mnt_iosize_max / size; 118 maxra = seqcount; 119 maxra = min(read_max, maxra); 120 maxra = min(nbuf/8, maxra); 121 if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize) 122 maxra = (filesize / size) - lblkno; 123 124 /* 125 * get the requested block 126 */ 127 *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, 0); 128 origblkno = lblkno; 129 130 /* 131 * if it is in the cache, then check to see if the reads have been 132 * sequential. If they have, then try some read-ahead, otherwise 133 * back-off on prospective read-aheads. 134 */ 135 if (bp->b_flags & B_CACHE) { 136 if (!seqcount) { 137 return 0; 138 } else if ((bp->b_flags & B_RAM) == 0) { 139 return 0; 140 } else { 141 int s; 142 bp->b_flags &= ~B_RAM; 143 /* 144 * We do the spl here so that there is no window 145 * between the incore and the b_usecount increment 146 * below. We opt to keep the spl out of the loop 147 * for efficiency. 148 */ 149 s = splbio(); 150 VI_LOCK(vp); 151 for (i = 1; i < maxra; i++) { 152 /* 153 * Stop if the buffer does not exist or it 154 * is invalid (about to go away?) 155 */ 156 rbp = gbincore(vp, lblkno+i); 157 if (rbp == NULL || (rbp->b_flags & B_INVAL)) 158 break; 159 160 /* 161 * Set another read-ahead mark so we know 162 * to check again. 163 */ 164 if (((i % racluster) == (racluster - 1)) || 165 (i == (maxra - 1))) 166 rbp->b_flags |= B_RAM; 167 } 168 VI_UNLOCK(vp); 169 splx(s); 170 if (i >= maxra) { 171 return 0; 172 } 173 lblkno += i; 174 } 175 reqbp = bp = NULL; 176 /* 177 * If it isn't in the cache, then get a chunk from 178 * disk if sequential, otherwise just get the block. 179 */ 180 } else { 181 off_t firstread = bp->b_offset; 182 int nblks; 183 184 KASSERT(bp->b_offset != NOOFFSET, 185 ("cluster_read: no buffer offset")); 186 187 ncontig = 0; 188 189 /* 190 * Compute the total number of blocks that we should read 191 * synchronously. 192 */ 193 if (firstread + totread > filesize) 194 totread = filesize - firstread; 195 nblks = howmany(totread, size); 196 if (nblks > racluster) 197 nblks = racluster; 198 199 /* 200 * Now compute the number of contiguous blocks. 201 */ 202 if (nblks > 1) { 203 error = VOP_BMAP(vp, lblkno, NULL, 204 &blkno, &ncontig, NULL); 205 /* 206 * If this failed to map just do the original block. 207 */ 208 if (error || blkno == -1) 209 ncontig = 0; 210 } 211 212 /* 213 * If we have contiguous data available do a cluster 214 * otherwise just read the requested block. 215 */ 216 if (ncontig) { 217 /* Account for our first block. */ 218 ncontig = min(ncontig + 1, nblks); 219 if (ncontig < nblks) 220 nblks = ncontig; 221 bp = cluster_rbuild(vp, filesize, lblkno, 222 blkno, size, nblks, bp); 223 lblkno += (bp->b_bufsize / size); 224 } else { 225 bp->b_flags |= B_RAM; 226 bp->b_iocmd = BIO_READ; 227 lblkno += 1; 228 } 229 } 230 231 /* 232 * handle the synchronous read so that it is available ASAP. 233 */ 234 if (bp) { 235 if ((bp->b_flags & B_CLUSTER) == 0) { 236 vfs_busy_pages(bp, 0); 237 } 238 bp->b_flags &= ~B_INVAL; 239 bp->b_ioflags &= ~BIO_ERROR; 240 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL) 241 BUF_KERNPROC(bp); 242 bp->b_iooffset = dbtob(bp->b_blkno); 243 error = VOP_STRATEGY(vp, bp); 244 curproc->p_stats->p_ru.ru_inblock++; 245 if (error) 246 return (error); 247 } 248 249 /* 250 * If we have been doing sequential I/O, then do some read-ahead. 251 */ 252 while (lblkno < (origblkno + maxra)) { 253 error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL); 254 if (error) 255 break; 256 257 if (blkno == -1) 258 break; 259 260 /* 261 * We could throttle ncontig here by maxra but we might as 262 * well read the data if it is contiguous. We're throttled 263 * by racluster anyway. 264 */ 265 if (ncontig) { 266 ncontig = min(ncontig + 1, racluster); 267 rbp = cluster_rbuild(vp, filesize, lblkno, blkno, 268 size, ncontig, NULL); 269 lblkno += (rbp->b_bufsize / size); 270 if (rbp->b_flags & B_DELWRI) { 271 bqrelse(rbp); 272 continue; 273 } 274 } else { 275 rbp = getblk(vp, lblkno, size, 0, 0, 0); 276 lblkno += 1; 277 if (rbp->b_flags & B_DELWRI) { 278 bqrelse(rbp); 279 continue; 280 } 281 rbp->b_flags |= B_ASYNC | B_RAM; 282 rbp->b_iocmd = BIO_READ; 283 rbp->b_blkno = blkno; 284 } 285 if (rbp->b_flags & B_CACHE) { 286 rbp->b_flags &= ~B_ASYNC; 287 bqrelse(rbp); 288 continue; 289 } 290 if ((rbp->b_flags & B_CLUSTER) == 0) { 291 vfs_busy_pages(rbp, 0); 292 } 293 rbp->b_flags &= ~B_INVAL; 294 rbp->b_ioflags &= ~BIO_ERROR; 295 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL) 296 BUF_KERNPROC(rbp); 297 rbp->b_iooffset = dbtob(rbp->b_blkno); 298 (void) VOP_STRATEGY(vp, rbp); 299 curproc->p_stats->p_ru.ru_inblock++; 300 } 301 302 if (reqbp) 303 return (bufwait(reqbp)); 304 else 305 return (error); 306 } 307 308 /* 309 * If blocks are contiguous on disk, use this to provide clustered 310 * read ahead. We will read as many blocks as possible sequentially 311 * and then parcel them up into logical blocks in the buffer hash table. 312 */ 313 static struct buf * 314 cluster_rbuild(vp, filesize, lbn, blkno, size, run, fbp) 315 struct vnode *vp; 316 u_quad_t filesize; 317 daddr_t lbn; 318 daddr_t blkno; 319 long size; 320 int run; 321 struct buf *fbp; 322 { 323 struct buf *bp, *tbp; 324 daddr_t bn; 325 int i, inc, j; 326 327 GIANT_REQUIRED; 328 329 KASSERT(size == vp->v_mount->mnt_stat.f_iosize, 330 ("cluster_rbuild: size %ld != filesize %jd\n", 331 size, (intmax_t)vp->v_mount->mnt_stat.f_iosize)); 332 333 /* 334 * avoid a division 335 */ 336 while ((u_quad_t) size * (lbn + run) > filesize) { 337 --run; 338 } 339 340 if (fbp) { 341 tbp = fbp; 342 tbp->b_iocmd = BIO_READ; 343 } else { 344 tbp = getblk(vp, lbn, size, 0, 0, 0); 345 if (tbp->b_flags & B_CACHE) 346 return tbp; 347 tbp->b_flags |= B_ASYNC | B_RAM; 348 tbp->b_iocmd = BIO_READ; 349 } 350 351 tbp->b_blkno = blkno; 352 if( (tbp->b_flags & B_MALLOC) || 353 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) ) 354 return tbp; 355 356 bp = trypbuf(&cluster_pbuf_freecnt); 357 if (bp == 0) 358 return tbp; 359 360 /* 361 * We are synthesizing a buffer out of vm_page_t's, but 362 * if the block size is not page aligned then the starting 363 * address may not be either. Inherit the b_data offset 364 * from the original buffer. 365 */ 366 bp->b_data = (char *)((vm_offset_t)bp->b_data | 367 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 368 bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO; 369 bp->b_iocmd = BIO_READ; 370 bp->b_iodone = cluster_callback; 371 bp->b_blkno = blkno; 372 bp->b_lblkno = lbn; 373 bp->b_offset = tbp->b_offset; 374 KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset")); 375 pbgetvp(vp, bp); 376 377 TAILQ_INIT(&bp->b_cluster.cluster_head); 378 379 bp->b_bcount = 0; 380 bp->b_bufsize = 0; 381 bp->b_npages = 0; 382 383 inc = btodb(size); 384 for (bn = blkno, i = 0; i < run; ++i, bn += inc) { 385 if (i != 0) { 386 if ((bp->b_npages * PAGE_SIZE) + 387 round_page(size) > vp->v_mount->mnt_iosize_max) { 388 break; 389 } 390 391 tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT); 392 393 /* Don't wait around for locked bufs. */ 394 if (tbp == NULL) 395 break; 396 397 /* 398 * Stop scanning if the buffer is fully valid 399 * (marked B_CACHE), or locked (may be doing a 400 * background write), or if the buffer is not 401 * VMIO backed. The clustering code can only deal 402 * with VMIO-backed buffers. 403 */ 404 VI_LOCK(bp->b_vp); 405 if ((tbp->b_vflags & BV_BKGRDINPROG) || 406 (tbp->b_flags & B_CACHE) || 407 (tbp->b_flags & B_VMIO) == 0) { 408 VI_UNLOCK(bp->b_vp); 409 bqrelse(tbp); 410 break; 411 } 412 VI_UNLOCK(bp->b_vp); 413 414 /* 415 * The buffer must be completely invalid in order to 416 * take part in the cluster. If it is partially valid 417 * then we stop. 418 */ 419 VM_OBJECT_LOCK(tbp->b_object); 420 for (j = 0;j < tbp->b_npages; j++) { 421 VM_OBJECT_LOCK_ASSERT(tbp->b_pages[j]->object, 422 MA_OWNED); 423 if (tbp->b_pages[j]->valid) 424 break; 425 } 426 VM_OBJECT_UNLOCK(tbp->b_object); 427 if (j != tbp->b_npages) { 428 bqrelse(tbp); 429 break; 430 } 431 432 /* 433 * Set a read-ahead mark as appropriate 434 */ 435 if ((fbp && (i == 1)) || (i == (run - 1))) 436 tbp->b_flags |= B_RAM; 437 438 /* 439 * Set the buffer up for an async read (XXX should 440 * we do this only if we do not wind up brelse()ing?). 441 * Set the block number if it isn't set, otherwise 442 * if it is make sure it matches the block number we 443 * expect. 444 */ 445 tbp->b_flags |= B_ASYNC; 446 tbp->b_iocmd = BIO_READ; 447 if (tbp->b_blkno == tbp->b_lblkno) { 448 tbp->b_blkno = bn; 449 } else if (tbp->b_blkno != bn) { 450 brelse(tbp); 451 break; 452 } 453 } 454 /* 455 * XXX fbp from caller may not be B_ASYNC, but we are going 456 * to biodone() it in cluster_callback() anyway 457 */ 458 BUF_KERNPROC(tbp); 459 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, 460 tbp, b_cluster.cluster_entry); 461 VM_OBJECT_LOCK(tbp->b_object); 462 vm_page_lock_queues(); 463 for (j = 0; j < tbp->b_npages; j += 1) { 464 vm_page_t m; 465 m = tbp->b_pages[j]; 466 vm_page_io_start(m); 467 vm_object_pip_add(m->object, 1); 468 if ((bp->b_npages == 0) || 469 (bp->b_pages[bp->b_npages-1] != m)) { 470 bp->b_pages[bp->b_npages] = m; 471 bp->b_npages++; 472 } 473 if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) 474 tbp->b_pages[j] = bogus_page; 475 } 476 vm_page_unlock_queues(); 477 VM_OBJECT_UNLOCK(tbp->b_object); 478 /* 479 * XXX shouldn't this be += size for both, like in 480 * cluster_wbuild()? 481 * 482 * Don't inherit tbp->b_bufsize as it may be larger due to 483 * a non-page-aligned size. Instead just aggregate using 484 * 'size'. 485 */ 486 if (tbp->b_bcount != size) 487 printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size); 488 if (tbp->b_bufsize != size) 489 printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size); 490 bp->b_bcount += size; 491 bp->b_bufsize += size; 492 } 493 494 /* 495 * Fully valid pages in the cluster are already good and do not need 496 * to be re-read from disk. Replace the page with bogus_page 497 */ 498 VM_OBJECT_LOCK(bp->b_object); 499 for (j = 0; j < bp->b_npages; j++) { 500 VM_OBJECT_LOCK_ASSERT(bp->b_pages[j]->object, MA_OWNED); 501 if ((bp->b_pages[j]->valid & VM_PAGE_BITS_ALL) == 502 VM_PAGE_BITS_ALL) { 503 bp->b_pages[j] = bogus_page; 504 } 505 } 506 VM_OBJECT_UNLOCK(bp->b_object); 507 if (bp->b_bufsize > bp->b_kvasize) 508 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n", 509 bp->b_bufsize, bp->b_kvasize); 510 bp->b_kvasize = bp->b_bufsize; 511 512 pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 513 (vm_page_t *)bp->b_pages, bp->b_npages); 514 return (bp); 515 } 516 517 /* 518 * Cleanup after a clustered read or write. 519 * This is complicated by the fact that any of the buffers might have 520 * extra memory (if there were no empty buffer headers at allocbuf time) 521 * that we will need to shift around. 522 */ 523 void 524 cluster_callback(bp) 525 struct buf *bp; 526 { 527 struct buf *nbp, *tbp; 528 int error = 0; 529 530 GIANT_REQUIRED; 531 532 /* 533 * Must propogate errors to all the components. 534 */ 535 if (bp->b_ioflags & BIO_ERROR) 536 error = bp->b_error; 537 538 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 539 /* 540 * Move memory from the large cluster buffer into the component 541 * buffers and mark IO as done on these. 542 */ 543 for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head); 544 tbp; tbp = nbp) { 545 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry); 546 if (error) { 547 tbp->b_ioflags |= BIO_ERROR; 548 tbp->b_error = error; 549 } else { 550 tbp->b_dirtyoff = tbp->b_dirtyend = 0; 551 tbp->b_flags &= ~B_INVAL; 552 tbp->b_ioflags &= ~BIO_ERROR; 553 /* 554 * XXX the bdwrite()/bqrelse() issued during 555 * cluster building clears B_RELBUF (see bqrelse() 556 * comment). If direct I/O was specified, we have 557 * to restore it here to allow the buffer and VM 558 * to be freed. 559 */ 560 if (tbp->b_flags & B_DIRECT) 561 tbp->b_flags |= B_RELBUF; 562 } 563 bufdone(tbp); 564 } 565 relpbuf(bp, &cluster_pbuf_freecnt); 566 } 567 568 /* 569 * cluster_wbuild_wb: 570 * 571 * Implement modified write build for cluster. 572 * 573 * write_behind = 0 write behind disabled 574 * write_behind = 1 write behind normal (default) 575 * write_behind = 2 write behind backed-off 576 */ 577 578 static __inline int 579 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len) 580 { 581 int r = 0; 582 583 switch(write_behind) { 584 case 2: 585 if (start_lbn < len) 586 break; 587 start_lbn -= len; 588 /* FALLTHROUGH */ 589 case 1: 590 r = cluster_wbuild(vp, size, start_lbn, len); 591 /* FALLTHROUGH */ 592 default: 593 /* FALLTHROUGH */ 594 break; 595 } 596 return(r); 597 } 598 599 /* 600 * Do clustered write for FFS. 601 * 602 * Three cases: 603 * 1. Write is not sequential (write asynchronously) 604 * Write is sequential: 605 * 2. beginning of cluster - begin cluster 606 * 3. middle of a cluster - add to cluster 607 * 4. end of a cluster - asynchronously write cluster 608 */ 609 void 610 cluster_write(bp, filesize, seqcount) 611 struct buf *bp; 612 u_quad_t filesize; 613 int seqcount; 614 { 615 struct vnode *vp; 616 daddr_t lbn; 617 int maxclen, cursize; 618 int lblocksize; 619 int async; 620 621 vp = bp->b_vp; 622 if (vp->v_type == VREG) { 623 async = vp->v_mount->mnt_flag & MNT_ASYNC; 624 lblocksize = vp->v_mount->mnt_stat.f_iosize; 625 } else { 626 async = 0; 627 lblocksize = bp->b_bufsize; 628 } 629 lbn = bp->b_lblkno; 630 KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset")); 631 632 /* Initialize vnode to beginning of file. */ 633 if (lbn == 0) 634 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; 635 636 if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 || 637 (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) { 638 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1; 639 if (vp->v_clen != 0) { 640 /* 641 * Next block is not sequential. 642 * 643 * If we are not writing at end of file, the process 644 * seeked to another point in the file since its last 645 * write, or we have reached our maximum cluster size, 646 * then push the previous cluster. Otherwise try 647 * reallocating to make it sequential. 648 * 649 * Change to algorithm: only push previous cluster if 650 * it was sequential from the point of view of the 651 * seqcount heuristic, otherwise leave the buffer 652 * intact so we can potentially optimize the I/O 653 * later on in the buf_daemon or update daemon 654 * flush. 655 */ 656 cursize = vp->v_lastw - vp->v_cstart + 1; 657 if (((u_quad_t) bp->b_offset + lblocksize) != filesize || 658 lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) { 659 if (!async && seqcount > 0) { 660 cluster_wbuild_wb(vp, lblocksize, 661 vp->v_cstart, cursize); 662 } 663 } else { 664 struct buf **bpp, **endbp; 665 struct cluster_save *buflist; 666 667 buflist = cluster_collectbufs(vp, bp); 668 endbp = &buflist->bs_children 669 [buflist->bs_nchildren - 1]; 670 if (VOP_REALLOCBLKS(vp, buflist)) { 671 /* 672 * Failed, push the previous cluster 673 * if *really* writing sequentially 674 * in the logical file (seqcount > 1), 675 * otherwise delay it in the hopes that 676 * the low level disk driver can 677 * optimize the write ordering. 678 */ 679 for (bpp = buflist->bs_children; 680 bpp < endbp; bpp++) 681 brelse(*bpp); 682 free(buflist, M_SEGMENT); 683 if (seqcount > 1) { 684 cluster_wbuild_wb(vp, 685 lblocksize, vp->v_cstart, 686 cursize); 687 } 688 } else { 689 /* 690 * Succeeded, keep building cluster. 691 */ 692 for (bpp = buflist->bs_children; 693 bpp <= endbp; bpp++) 694 bdwrite(*bpp); 695 free(buflist, M_SEGMENT); 696 vp->v_lastw = lbn; 697 vp->v_lasta = bp->b_blkno; 698 return; 699 } 700 } 701 } 702 /* 703 * Consider beginning a cluster. If at end of file, make 704 * cluster as large as possible, otherwise find size of 705 * existing cluster. 706 */ 707 if ((vp->v_type == VREG) && 708 ((u_quad_t) bp->b_offset + lblocksize) != filesize && 709 (bp->b_blkno == bp->b_lblkno) && 710 (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) || 711 bp->b_blkno == -1)) { 712 bawrite(bp); 713 vp->v_clen = 0; 714 vp->v_lasta = bp->b_blkno; 715 vp->v_cstart = lbn + 1; 716 vp->v_lastw = lbn; 717 return; 718 } 719 vp->v_clen = maxclen; 720 if (!async && maxclen == 0) { /* I/O not contiguous */ 721 vp->v_cstart = lbn + 1; 722 bawrite(bp); 723 } else { /* Wait for rest of cluster */ 724 vp->v_cstart = lbn; 725 bdwrite(bp); 726 } 727 } else if (lbn == vp->v_cstart + vp->v_clen) { 728 /* 729 * At end of cluster, write it out if seqcount tells us we 730 * are operating sequentially, otherwise let the buf or 731 * update daemon handle it. 732 */ 733 bdwrite(bp); 734 if (seqcount > 1) 735 cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1); 736 vp->v_clen = 0; 737 vp->v_cstart = lbn + 1; 738 } else if (vm_page_count_severe()) { 739 /* 740 * We are low on memory, get it going NOW 741 */ 742 bawrite(bp); 743 } else { 744 /* 745 * In the middle of a cluster, so just delay the I/O for now. 746 */ 747 bdwrite(bp); 748 } 749 vp->v_lastw = lbn; 750 vp->v_lasta = bp->b_blkno; 751 } 752 753 754 /* 755 * This is an awful lot like cluster_rbuild...wish they could be combined. 756 * The last lbn argument is the current block on which I/O is being 757 * performed. Check to see that it doesn't fall in the middle of 758 * the current block (if last_bp == NULL). 759 */ 760 int 761 cluster_wbuild(vp, size, start_lbn, len) 762 struct vnode *vp; 763 long size; 764 daddr_t start_lbn; 765 int len; 766 { 767 struct buf *bp, *tbp; 768 int i, j, s; 769 int totalwritten = 0; 770 int dbsize = btodb(size); 771 772 GIANT_REQUIRED; 773 774 while (len > 0) { 775 s = splbio(); 776 /* 777 * If the buffer is not delayed-write (i.e. dirty), or it 778 * is delayed-write but either locked or inval, it cannot 779 * partake in the clustered write. 780 */ 781 VI_LOCK(vp); 782 if ((tbp = gbincore(vp, start_lbn)) == NULL || 783 (tbp->b_vflags & BV_BKGRDINPROG)) { 784 VI_UNLOCK(vp); 785 ++start_lbn; 786 --len; 787 splx(s); 788 continue; 789 } 790 if (BUF_LOCK(tbp, 791 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, VI_MTX(vp))) { 792 ++start_lbn; 793 --len; 794 splx(s); 795 continue; 796 } 797 if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) { 798 BUF_UNLOCK(tbp); 799 ++start_lbn; 800 --len; 801 splx(s); 802 continue; 803 } 804 bremfree(tbp); 805 tbp->b_flags &= ~B_DONE; 806 splx(s); 807 808 /* 809 * Extra memory in the buffer, punt on this buffer. 810 * XXX we could handle this in most cases, but we would 811 * have to push the extra memory down to after our max 812 * possible cluster size and then potentially pull it back 813 * up if the cluster was terminated prematurely--too much 814 * hassle. 815 */ 816 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) != 817 (B_CLUSTEROK | B_VMIO)) || 818 (tbp->b_bcount != tbp->b_bufsize) || 819 (tbp->b_bcount != size) || 820 (len == 1) || 821 ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) { 822 totalwritten += tbp->b_bufsize; 823 bawrite(tbp); 824 ++start_lbn; 825 --len; 826 continue; 827 } 828 829 /* 830 * We got a pbuf to make the cluster in. 831 * so initialise it. 832 */ 833 TAILQ_INIT(&bp->b_cluster.cluster_head); 834 bp->b_bcount = 0; 835 bp->b_magic = tbp->b_magic; 836 bp->b_op = tbp->b_op; 837 bp->b_bufsize = 0; 838 bp->b_npages = 0; 839 if (tbp->b_wcred != NOCRED) 840 bp->b_wcred = crhold(tbp->b_wcred); 841 842 bp->b_blkno = tbp->b_blkno; 843 bp->b_lblkno = tbp->b_lblkno; 844 bp->b_offset = tbp->b_offset; 845 846 /* 847 * We are synthesizing a buffer out of vm_page_t's, but 848 * if the block size is not page aligned then the starting 849 * address may not be either. Inherit the b_data offset 850 * from the original buffer. 851 */ 852 bp->b_data = (char *)((vm_offset_t)bp->b_data | 853 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 854 bp->b_flags |= B_CLUSTER | 855 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT)); 856 bp->b_iodone = cluster_callback; 857 pbgetvp(vp, bp); 858 /* 859 * From this location in the file, scan forward to see 860 * if there are buffers with adjacent data that need to 861 * be written as well. 862 */ 863 for (i = 0; i < len; ++i, ++start_lbn) { 864 if (i != 0) { /* If not the first buffer */ 865 s = splbio(); 866 /* 867 * If the adjacent data is not even in core it 868 * can't need to be written. 869 */ 870 VI_LOCK(vp); 871 if ((tbp = gbincore(vp, start_lbn)) == NULL || 872 (tbp->b_vflags & BV_BKGRDINPROG)) { 873 VI_UNLOCK(vp); 874 splx(s); 875 break; 876 } 877 878 /* 879 * If it IS in core, but has different 880 * characteristics, or is locked (which 881 * means it could be undergoing a background 882 * I/O or be in a weird state), then don't 883 * cluster with it. 884 */ 885 if (BUF_LOCK(tbp, 886 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, 887 VI_MTX(vp))) { 888 splx(s); 889 break; 890 } 891 892 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK | 893 B_INVAL | B_DELWRI | B_NEEDCOMMIT)) 894 != (B_DELWRI | B_CLUSTEROK | 895 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) || 896 tbp->b_wcred != bp->b_wcred) { 897 BUF_UNLOCK(tbp); 898 splx(s); 899 break; 900 } 901 902 /* 903 * Check that the combined cluster 904 * would make sense with regard to pages 905 * and would not be too large 906 */ 907 if ((tbp->b_bcount != size) || 908 ((bp->b_blkno + (dbsize * i)) != 909 tbp->b_blkno) || 910 ((tbp->b_npages + bp->b_npages) > 911 (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) { 912 BUF_UNLOCK(tbp); 913 splx(s); 914 break; 915 } 916 /* 917 * Ok, it's passed all the tests, 918 * so remove it from the free list 919 * and mark it busy. We will use it. 920 */ 921 bremfree(tbp); 922 tbp->b_flags &= ~B_DONE; 923 splx(s); 924 } /* end of code for non-first buffers only */ 925 /* check for latent dependencies to be handled */ 926 if ((LIST_FIRST(&tbp->b_dep)) != NULL) { 927 tbp->b_iocmd = BIO_WRITE; 928 buf_start(tbp); 929 } 930 /* 931 * If the IO is via the VM then we do some 932 * special VM hackery (yuck). Since the buffer's 933 * block size may not be page-aligned it is possible 934 * for a page to be shared between two buffers. We 935 * have to get rid of the duplication when building 936 * the cluster. 937 */ 938 if (tbp->b_flags & B_VMIO) { 939 vm_page_t m; 940 941 if (i != 0) { /* if not first buffer */ 942 for (j = 0; j < tbp->b_npages; j += 1) { 943 m = tbp->b_pages[j]; 944 if (m->flags & PG_BUSY) { 945 bqrelse(tbp); 946 goto finishcluster; 947 } 948 } 949 } 950 if (tbp->b_object != NULL) 951 VM_OBJECT_LOCK(tbp->b_object); 952 vm_page_lock_queues(); 953 for (j = 0; j < tbp->b_npages; j += 1) { 954 m = tbp->b_pages[j]; 955 vm_page_io_start(m); 956 vm_object_pip_add(m->object, 1); 957 if ((bp->b_npages == 0) || 958 (bp->b_pages[bp->b_npages - 1] != m)) { 959 bp->b_pages[bp->b_npages] = m; 960 bp->b_npages++; 961 } 962 } 963 vm_page_unlock_queues(); 964 if (tbp->b_object != NULL) 965 VM_OBJECT_UNLOCK(tbp->b_object); 966 } 967 bp->b_bcount += size; 968 bp->b_bufsize += size; 969 970 s = splbio(); 971 bundirty(tbp); 972 tbp->b_flags &= ~B_DONE; 973 tbp->b_ioflags &= ~BIO_ERROR; 974 tbp->b_flags |= B_ASYNC; 975 tbp->b_iocmd = BIO_WRITE; 976 reassignbuf(tbp, tbp->b_vp); /* put on clean list */ 977 VI_LOCK(tbp->b_vp); 978 ++tbp->b_vp->v_numoutput; 979 VI_UNLOCK(tbp->b_vp); 980 splx(s); 981 BUF_KERNPROC(tbp); 982 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, 983 tbp, b_cluster.cluster_entry); 984 } 985 finishcluster: 986 pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 987 (vm_page_t *) bp->b_pages, bp->b_npages); 988 if (bp->b_bufsize > bp->b_kvasize) 989 panic( 990 "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n", 991 bp->b_bufsize, bp->b_kvasize); 992 bp->b_kvasize = bp->b_bufsize; 993 totalwritten += bp->b_bufsize; 994 bp->b_dirtyoff = 0; 995 bp->b_dirtyend = bp->b_bufsize; 996 bawrite(bp); 997 998 len -= i; 999 } 1000 return totalwritten; 1001 } 1002 1003 /* 1004 * Collect together all the buffers in a cluster. 1005 * Plus add one additional buffer. 1006 */ 1007 static struct cluster_save * 1008 cluster_collectbufs(vp, last_bp) 1009 struct vnode *vp; 1010 struct buf *last_bp; 1011 { 1012 struct cluster_save *buflist; 1013 struct buf *bp; 1014 daddr_t lbn; 1015 int i, len; 1016 1017 len = vp->v_lastw - vp->v_cstart + 1; 1018 buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist), 1019 M_SEGMENT, M_WAITOK); 1020 buflist->bs_nchildren = 0; 1021 buflist->bs_children = (struct buf **) (buflist + 1); 1022 for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) { 1023 (void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp); 1024 buflist->bs_children[i] = bp; 1025 if (bp->b_blkno == bp->b_lblkno) 1026 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, 1027 NULL, NULL); 1028 } 1029 buflist->bs_children[i] = bp = last_bp; 1030 if (bp->b_blkno == bp->b_lblkno) 1031 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, 1032 NULL, NULL); 1033 buflist->bs_nchildren = i + 1; 1034 return (buflist); 1035 } 1036