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