1 /*- 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 1993 5 * The Regents of the University of California. All rights reserved. 6 * Modifications/enhancements: 7 * Copyright (c) 1995 John S. Dyson. All rights reserved. 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 3. Neither the name of the University nor the names of its contributors 18 * may be used to endorse or promote products derived from this software 19 * without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 * 33 * @(#)vfs_cluster.c 8.7 (Berkeley) 2/13/94 34 */ 35 36 #include <sys/cdefs.h> 37 __FBSDID("$FreeBSD$"); 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/racct.h> 51 #include <sys/resourcevar.h> 52 #include <sys/rwlock.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, "cl_savebuf", "cluster_save buffer"); 66 static uma_zone_t cluster_pbuf_zone; 67 68 static void cluster_init(void *); 69 static struct cluster_save *cluster_collectbufs(struct vnode *vp, 70 struct buf *last_bp, int gbflags); 71 static struct buf *cluster_rbuild(struct vnode *vp, u_quad_t filesize, 72 daddr_t lbn, daddr_t blkno, long size, int run, int gbflags, 73 struct buf *fbp); 74 static void cluster_callback(struct buf *); 75 76 static int write_behind = 1; 77 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, 78 "Cluster write-behind; 0: disable, 1: enable, 2: backed off"); 79 80 static int read_max = 64; 81 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0, 82 "Cluster read-ahead max block count"); 83 84 static int read_min = 1; 85 SYSCTL_INT(_vfs, OID_AUTO, read_min, CTLFLAG_RW, &read_min, 0, 86 "Cluster read min block count"); 87 88 SYSINIT(cluster, SI_SUB_CPU, SI_ORDER_ANY, cluster_init, NULL); 89 90 static void 91 cluster_init(void *dummy) 92 { 93 94 cluster_pbuf_zone = pbuf_zsecond_create("clpbuf", nswbuf / 2); 95 } 96 97 /* 98 * Read data to a buf, including read-ahead if we find this to be beneficial. 99 * cluster_read replaces bread. 100 */ 101 int 102 cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, long size, 103 struct ucred *cred, long totread, int seqcount, int gbflags, 104 struct buf **bpp) 105 { 106 struct buf *bp, *rbp, *reqbp; 107 struct bufobj *bo; 108 struct thread *td; 109 daddr_t blkno, origblkno; 110 int maxra, racluster; 111 int error, ncontig; 112 int i; 113 114 error = 0; 115 td = curthread; 116 bo = &vp->v_bufobj; 117 if (!unmapped_buf_allowed) 118 gbflags &= ~GB_UNMAPPED; 119 120 /* 121 * Try to limit the amount of read-ahead by a few 122 * ad-hoc parameters. This needs work!!! 123 */ 124 racluster = vp->v_mount->mnt_iosize_max / size; 125 maxra = seqcount; 126 maxra = min(read_max, maxra); 127 maxra = min(nbuf/8, maxra); 128 if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize) 129 maxra = (filesize / size) - lblkno; 130 131 /* 132 * get the requested block 133 */ 134 error = getblkx(vp, lblkno, size, 0, 0, gbflags, &bp); 135 if (error != 0) { 136 *bpp = NULL; 137 return (error); 138 } 139 gbflags &= ~GB_NOSPARSE; 140 origblkno = lblkno; 141 *bpp = reqbp = bp; 142 143 /* 144 * if it is in the cache, then check to see if the reads have been 145 * sequential. If they have, then try some read-ahead, otherwise 146 * back-off on prospective read-aheads. 147 */ 148 if (bp->b_flags & B_CACHE) { 149 if (!seqcount) { 150 return 0; 151 } else if ((bp->b_flags & B_RAM) == 0) { 152 return 0; 153 } else { 154 bp->b_flags &= ~B_RAM; 155 BO_RLOCK(bo); 156 for (i = 1; i < maxra; i++) { 157 /* 158 * Stop if the buffer does not exist or it 159 * is invalid (about to go away?) 160 */ 161 rbp = gbincore(&vp->v_bufobj, lblkno+i); 162 if (rbp == NULL || (rbp->b_flags & B_INVAL)) 163 break; 164 165 /* 166 * Set another read-ahead mark so we know 167 * to check again. (If we can lock the 168 * buffer without waiting) 169 */ 170 if ((((i % racluster) == (racluster - 1)) || 171 (i == (maxra - 1))) 172 && (0 == BUF_LOCK(rbp, 173 LK_EXCLUSIVE | LK_NOWAIT, NULL))) { 174 rbp->b_flags |= B_RAM; 175 BUF_UNLOCK(rbp); 176 } 177 } 178 BO_RUNLOCK(bo); 179 if (i >= maxra) { 180 return 0; 181 } 182 lblkno += i; 183 } 184 reqbp = bp = NULL; 185 /* 186 * If it isn't in the cache, then get a chunk from 187 * disk if sequential, otherwise just get the block. 188 */ 189 } else { 190 off_t firstread = bp->b_offset; 191 int nblks; 192 long minread; 193 194 KASSERT(bp->b_offset != NOOFFSET, 195 ("cluster_read: no buffer offset")); 196 197 ncontig = 0; 198 199 /* 200 * Adjust totread if needed 201 */ 202 minread = read_min * size; 203 if (minread > totread) 204 totread = minread; 205 206 /* 207 * Compute the total number of blocks that we should read 208 * synchronously. 209 */ 210 if (firstread + totread > filesize) 211 totread = filesize - firstread; 212 nblks = howmany(totread, size); 213 if (nblks > racluster) 214 nblks = racluster; 215 216 /* 217 * Now compute the number of contiguous blocks. 218 */ 219 if (nblks > 1) { 220 error = VOP_BMAP(vp, lblkno, NULL, 221 &blkno, &ncontig, NULL); 222 /* 223 * If this failed to map just do the original block. 224 */ 225 if (error || blkno == -1) 226 ncontig = 0; 227 } 228 229 /* 230 * If we have contiguous data available do a cluster 231 * otherwise just read the requested block. 232 */ 233 if (ncontig) { 234 /* Account for our first block. */ 235 ncontig = min(ncontig + 1, nblks); 236 if (ncontig < nblks) 237 nblks = ncontig; 238 bp = cluster_rbuild(vp, filesize, lblkno, 239 blkno, size, nblks, gbflags, bp); 240 lblkno += (bp->b_bufsize / size); 241 } else { 242 bp->b_flags |= B_RAM; 243 bp->b_iocmd = BIO_READ; 244 lblkno += 1; 245 } 246 } 247 248 /* 249 * handle the synchronous read so that it is available ASAP. 250 */ 251 if (bp) { 252 if ((bp->b_flags & B_CLUSTER) == 0) { 253 vfs_busy_pages(bp, 0); 254 } 255 bp->b_flags &= ~B_INVAL; 256 bp->b_ioflags &= ~BIO_ERROR; 257 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL) 258 BUF_KERNPROC(bp); 259 bp->b_iooffset = dbtob(bp->b_blkno); 260 bstrategy(bp); 261 #ifdef RACCT 262 if (racct_enable) { 263 PROC_LOCK(td->td_proc); 264 racct_add_buf(td->td_proc, bp, 0); 265 PROC_UNLOCK(td->td_proc); 266 } 267 #endif /* RACCT */ 268 td->td_ru.ru_inblock++; 269 } 270 271 /* 272 * If we have been doing sequential I/O, then do some read-ahead. 273 */ 274 while (lblkno < (origblkno + maxra)) { 275 error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL); 276 if (error) 277 break; 278 279 if (blkno == -1) 280 break; 281 282 /* 283 * We could throttle ncontig here by maxra but we might as 284 * well read the data if it is contiguous. We're throttled 285 * by racluster anyway. 286 */ 287 if (ncontig) { 288 ncontig = min(ncontig + 1, racluster); 289 rbp = cluster_rbuild(vp, filesize, lblkno, blkno, 290 size, ncontig, gbflags, NULL); 291 lblkno += (rbp->b_bufsize / size); 292 if (rbp->b_flags & B_DELWRI) { 293 bqrelse(rbp); 294 continue; 295 } 296 } else { 297 rbp = getblk(vp, lblkno, size, 0, 0, gbflags); 298 lblkno += 1; 299 if (rbp->b_flags & B_DELWRI) { 300 bqrelse(rbp); 301 continue; 302 } 303 rbp->b_flags |= B_ASYNC | B_RAM; 304 rbp->b_iocmd = BIO_READ; 305 rbp->b_blkno = blkno; 306 } 307 if (rbp->b_flags & B_CACHE) { 308 rbp->b_flags &= ~B_ASYNC; 309 bqrelse(rbp); 310 continue; 311 } 312 if ((rbp->b_flags & B_CLUSTER) == 0) { 313 vfs_busy_pages(rbp, 0); 314 } 315 rbp->b_flags &= ~B_INVAL; 316 rbp->b_ioflags &= ~BIO_ERROR; 317 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL) 318 BUF_KERNPROC(rbp); 319 rbp->b_iooffset = dbtob(rbp->b_blkno); 320 bstrategy(rbp); 321 #ifdef RACCT 322 if (racct_enable) { 323 PROC_LOCK(td->td_proc); 324 racct_add_buf(td->td_proc, rbp, 0); 325 PROC_UNLOCK(td->td_proc); 326 } 327 #endif /* RACCT */ 328 td->td_ru.ru_inblock++; 329 } 330 331 if (reqbp) { 332 /* 333 * Like bread, always brelse() the buffer when 334 * returning an error. 335 */ 336 error = bufwait(reqbp); 337 if (error != 0) { 338 brelse(reqbp); 339 *bpp = NULL; 340 } 341 } 342 return (error); 343 } 344 345 /* 346 * If blocks are contiguous on disk, use this to provide clustered 347 * read ahead. We will read as many blocks as possible sequentially 348 * and then parcel them up into logical blocks in the buffer hash table. 349 */ 350 static struct buf * 351 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn, 352 daddr_t blkno, long size, int run, int gbflags, struct buf *fbp) 353 { 354 struct buf *bp, *tbp; 355 daddr_t bn; 356 off_t off; 357 long tinc, tsize; 358 int i, inc, j, k, toff; 359 360 KASSERT(size == vp->v_mount->mnt_stat.f_iosize, 361 ("cluster_rbuild: size %ld != f_iosize %jd\n", 362 size, (intmax_t)vp->v_mount->mnt_stat.f_iosize)); 363 364 /* 365 * avoid a division 366 */ 367 while ((u_quad_t) size * (lbn + run) > filesize) { 368 --run; 369 } 370 371 if (fbp) { 372 tbp = fbp; 373 tbp->b_iocmd = BIO_READ; 374 } else { 375 tbp = getblk(vp, lbn, size, 0, 0, gbflags); 376 if (tbp->b_flags & B_CACHE) 377 return tbp; 378 tbp->b_flags |= B_ASYNC | B_RAM; 379 tbp->b_iocmd = BIO_READ; 380 } 381 tbp->b_blkno = blkno; 382 if( (tbp->b_flags & B_MALLOC) || 383 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) ) 384 return tbp; 385 386 bp = uma_zalloc(cluster_pbuf_zone, M_NOWAIT); 387 if (bp == NULL) 388 return tbp; 389 390 /* 391 * We are synthesizing a buffer out of vm_page_t's, but 392 * if the block size is not page aligned then the starting 393 * address may not be either. Inherit the b_data offset 394 * from the original buffer. 395 */ 396 bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO; 397 if ((gbflags & GB_UNMAPPED) != 0) { 398 bp->b_data = unmapped_buf; 399 } else { 400 bp->b_data = (char *)((vm_offset_t)bp->b_data | 401 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 402 } 403 bp->b_iocmd = BIO_READ; 404 bp->b_iodone = cluster_callback; 405 bp->b_blkno = blkno; 406 bp->b_lblkno = lbn; 407 bp->b_offset = tbp->b_offset; 408 KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset")); 409 pbgetvp(vp, bp); 410 411 TAILQ_INIT(&bp->b_cluster.cluster_head); 412 413 bp->b_bcount = 0; 414 bp->b_bufsize = 0; 415 bp->b_npages = 0; 416 417 inc = btodb(size); 418 for (bn = blkno, i = 0; i < run; ++i, bn += inc) { 419 if (i == 0) { 420 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); 421 vm_object_pip_add(tbp->b_bufobj->bo_object, 422 tbp->b_npages); 423 vfs_busy_pages_acquire(tbp); 424 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); 425 } else { 426 if ((bp->b_npages * PAGE_SIZE) + 427 round_page(size) > vp->v_mount->mnt_iosize_max) { 428 break; 429 } 430 431 tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT | 432 (gbflags & GB_UNMAPPED)); 433 434 /* Don't wait around for locked bufs. */ 435 if (tbp == NULL) 436 break; 437 438 /* 439 * Stop scanning if the buffer is fully valid 440 * (marked B_CACHE), or locked (may be doing a 441 * background write), or if the buffer is not 442 * VMIO backed. The clustering code can only deal 443 * with VMIO-backed buffers. The bo lock is not 444 * required for the BKGRDINPROG check since it 445 * can not be set without the buf lock. 446 */ 447 if ((tbp->b_vflags & BV_BKGRDINPROG) || 448 (tbp->b_flags & B_CACHE) || 449 (tbp->b_flags & B_VMIO) == 0) { 450 bqrelse(tbp); 451 break; 452 } 453 454 /* 455 * The buffer must be completely invalid in order to 456 * take part in the cluster. If it is partially valid 457 * then we stop. 458 */ 459 off = tbp->b_offset; 460 tsize = size; 461 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); 462 for (j = 0; tsize > 0; j++) { 463 toff = off & PAGE_MASK; 464 tinc = tsize; 465 if (toff + tinc > PAGE_SIZE) 466 tinc = PAGE_SIZE - toff; 467 VM_OBJECT_ASSERT_WLOCKED(tbp->b_pages[j]->object); 468 if (vm_page_trysbusy(tbp->b_pages[j]) == 0) 469 break; 470 if ((tbp->b_pages[j]->valid & 471 vm_page_bits(toff, tinc)) != 0) { 472 vm_page_sunbusy(tbp->b_pages[j]); 473 break; 474 } 475 vm_object_pip_add(tbp->b_bufobj->bo_object, 1); 476 off += tinc; 477 tsize -= tinc; 478 } 479 if (tsize > 0) { 480 clean_sbusy: 481 vm_object_pip_wakeupn(tbp->b_bufobj->bo_object, 482 j); 483 for (k = 0; k < j; k++) 484 vm_page_sunbusy(tbp->b_pages[k]); 485 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); 486 bqrelse(tbp); 487 break; 488 } 489 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); 490 491 /* 492 * Set a read-ahead mark as appropriate 493 */ 494 if ((fbp && (i == 1)) || (i == (run - 1))) 495 tbp->b_flags |= B_RAM; 496 497 /* 498 * Set the buffer up for an async read (XXX should 499 * we do this only if we do not wind up brelse()ing?). 500 * Set the block number if it isn't set, otherwise 501 * if it is make sure it matches the block number we 502 * expect. 503 */ 504 tbp->b_flags |= B_ASYNC; 505 tbp->b_iocmd = BIO_READ; 506 if (tbp->b_blkno == tbp->b_lblkno) { 507 tbp->b_blkno = bn; 508 } else if (tbp->b_blkno != bn) { 509 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); 510 goto clean_sbusy; 511 } 512 } 513 /* 514 * XXX fbp from caller may not be B_ASYNC, but we are going 515 * to biodone() it in cluster_callback() anyway 516 */ 517 BUF_KERNPROC(tbp); 518 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, 519 tbp, b_cluster.cluster_entry); 520 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); 521 for (j = 0; j < tbp->b_npages; j += 1) { 522 vm_page_t m; 523 m = tbp->b_pages[j]; 524 if ((bp->b_npages == 0) || 525 (bp->b_pages[bp->b_npages-1] != m)) { 526 bp->b_pages[bp->b_npages] = m; 527 bp->b_npages++; 528 } 529 if (vm_page_all_valid(m)) 530 tbp->b_pages[j] = bogus_page; 531 } 532 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); 533 /* 534 * Don't inherit tbp->b_bufsize as it may be larger due to 535 * a non-page-aligned size. Instead just aggregate using 536 * 'size'. 537 */ 538 if (tbp->b_bcount != size) 539 printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size); 540 if (tbp->b_bufsize != size) 541 printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size); 542 bp->b_bcount += size; 543 bp->b_bufsize += size; 544 } 545 546 /* 547 * Fully valid pages in the cluster are already good and do not need 548 * to be re-read from disk. Replace the page with bogus_page 549 */ 550 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 551 for (j = 0; j < bp->b_npages; j++) { 552 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[j]->object); 553 if (vm_page_all_valid(bp->b_pages[j])) 554 bp->b_pages[j] = bogus_page; 555 } 556 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 557 if (bp->b_bufsize > bp->b_kvasize) 558 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n", 559 bp->b_bufsize, bp->b_kvasize); 560 561 if (buf_mapped(bp)) { 562 pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 563 (vm_page_t *)bp->b_pages, bp->b_npages); 564 } 565 return (bp); 566 } 567 568 /* 569 * Cleanup after a clustered read or write. 570 * This is complicated by the fact that any of the buffers might have 571 * extra memory (if there were no empty buffer headers at allocbuf time) 572 * that we will need to shift around. 573 */ 574 static void 575 cluster_callback(struct buf *bp) 576 { 577 struct buf *nbp, *tbp; 578 int error = 0; 579 580 /* 581 * Must propagate errors to all the components. 582 */ 583 if (bp->b_ioflags & BIO_ERROR) 584 error = bp->b_error; 585 586 if (buf_mapped(bp)) { 587 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), 588 bp->b_npages); 589 } 590 /* 591 * Move memory from the large cluster buffer into the component 592 * buffers and mark IO as done on these. 593 */ 594 for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head); 595 tbp; tbp = nbp) { 596 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry); 597 if (error) { 598 tbp->b_ioflags |= BIO_ERROR; 599 tbp->b_error = error; 600 } else { 601 tbp->b_dirtyoff = tbp->b_dirtyend = 0; 602 tbp->b_flags &= ~B_INVAL; 603 tbp->b_ioflags &= ~BIO_ERROR; 604 /* 605 * XXX the bdwrite()/bqrelse() issued during 606 * cluster building clears B_RELBUF (see bqrelse() 607 * comment). If direct I/O was specified, we have 608 * to restore it here to allow the buffer and VM 609 * to be freed. 610 */ 611 if (tbp->b_flags & B_DIRECT) 612 tbp->b_flags |= B_RELBUF; 613 } 614 bufdone(tbp); 615 } 616 pbrelvp(bp); 617 uma_zfree(cluster_pbuf_zone, bp); 618 } 619 620 /* 621 * cluster_wbuild_wb: 622 * 623 * Implement modified write build for cluster. 624 * 625 * write_behind = 0 write behind disabled 626 * write_behind = 1 write behind normal (default) 627 * write_behind = 2 write behind backed-off 628 */ 629 630 static __inline int 631 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len, 632 int gbflags) 633 { 634 int r = 0; 635 636 switch (write_behind) { 637 case 2: 638 if (start_lbn < len) 639 break; 640 start_lbn -= len; 641 /* FALLTHROUGH */ 642 case 1: 643 r = cluster_wbuild(vp, size, start_lbn, len, gbflags); 644 /* FALLTHROUGH */ 645 default: 646 /* FALLTHROUGH */ 647 break; 648 } 649 return(r); 650 } 651 652 /* 653 * Do clustered write for FFS. 654 * 655 * Three cases: 656 * 1. Write is not sequential (write asynchronously) 657 * Write is sequential: 658 * 2. beginning of cluster - begin cluster 659 * 3. middle of a cluster - add to cluster 660 * 4. end of a cluster - asynchronously write cluster 661 */ 662 void 663 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount, 664 int gbflags) 665 { 666 daddr_t lbn; 667 int maxclen, cursize; 668 int lblocksize; 669 int async; 670 671 if (!unmapped_buf_allowed) 672 gbflags &= ~GB_UNMAPPED; 673 674 if (vp->v_type == VREG) { 675 async = DOINGASYNC(vp); 676 lblocksize = vp->v_mount->mnt_stat.f_iosize; 677 } else { 678 async = 0; 679 lblocksize = bp->b_bufsize; 680 } 681 lbn = bp->b_lblkno; 682 KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset")); 683 684 /* Initialize vnode to beginning of file. */ 685 if (lbn == 0) 686 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; 687 688 if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 || 689 (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) { 690 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1; 691 if (vp->v_clen != 0) { 692 /* 693 * Next block is not sequential. 694 * 695 * If we are not writing at end of file, the process 696 * seeked to another point in the file since its last 697 * write, or we have reached our maximum cluster size, 698 * then push the previous cluster. Otherwise try 699 * reallocating to make it sequential. 700 * 701 * Change to algorithm: only push previous cluster if 702 * it was sequential from the point of view of the 703 * seqcount heuristic, otherwise leave the buffer 704 * intact so we can potentially optimize the I/O 705 * later on in the buf_daemon or update daemon 706 * flush. 707 */ 708 cursize = vp->v_lastw - vp->v_cstart + 1; 709 if (((u_quad_t) bp->b_offset + lblocksize) != filesize || 710 lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) { 711 if (!async && seqcount > 0) { 712 cluster_wbuild_wb(vp, lblocksize, 713 vp->v_cstart, cursize, gbflags); 714 } 715 } else { 716 struct buf **bpp, **endbp; 717 struct cluster_save *buflist; 718 719 buflist = cluster_collectbufs(vp, bp, gbflags); 720 if (buflist == NULL) { 721 /* 722 * Cluster build failed so just write 723 * it now. 724 */ 725 bawrite(bp); 726 return; 727 } 728 endbp = &buflist->bs_children 729 [buflist->bs_nchildren - 1]; 730 if (VOP_REALLOCBLKS(vp, buflist)) { 731 /* 732 * Failed, push the previous cluster 733 * if *really* writing sequentially 734 * in the logical file (seqcount > 1), 735 * otherwise delay it in the hopes that 736 * the low level disk driver can 737 * optimize the write ordering. 738 */ 739 for (bpp = buflist->bs_children; 740 bpp < endbp; bpp++) 741 brelse(*bpp); 742 free(buflist, M_SEGMENT); 743 if (seqcount > 1) { 744 cluster_wbuild_wb(vp, 745 lblocksize, vp->v_cstart, 746 cursize, gbflags); 747 } 748 } else { 749 /* 750 * Succeeded, keep building cluster. 751 */ 752 for (bpp = buflist->bs_children; 753 bpp <= endbp; bpp++) 754 bdwrite(*bpp); 755 free(buflist, M_SEGMENT); 756 vp->v_lastw = lbn; 757 vp->v_lasta = bp->b_blkno; 758 return; 759 } 760 } 761 } 762 /* 763 * Consider beginning a cluster. If at end of file, make 764 * cluster as large as possible, otherwise find size of 765 * existing cluster. 766 */ 767 if ((vp->v_type == VREG) && 768 ((u_quad_t) bp->b_offset + lblocksize) != filesize && 769 (bp->b_blkno == bp->b_lblkno) && 770 (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) || 771 bp->b_blkno == -1)) { 772 bawrite(bp); 773 vp->v_clen = 0; 774 vp->v_lasta = bp->b_blkno; 775 vp->v_cstart = lbn + 1; 776 vp->v_lastw = lbn; 777 return; 778 } 779 vp->v_clen = maxclen; 780 if (!async && maxclen == 0) { /* I/O not contiguous */ 781 vp->v_cstart = lbn + 1; 782 bawrite(bp); 783 } else { /* Wait for rest of cluster */ 784 vp->v_cstart = lbn; 785 bdwrite(bp); 786 } 787 } else if (lbn == vp->v_cstart + vp->v_clen) { 788 /* 789 * At end of cluster, write it out if seqcount tells us we 790 * are operating sequentially, otherwise let the buf or 791 * update daemon handle it. 792 */ 793 bdwrite(bp); 794 if (seqcount > 1) { 795 cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, 796 vp->v_clen + 1, gbflags); 797 } 798 vp->v_clen = 0; 799 vp->v_cstart = lbn + 1; 800 } else if (vm_page_count_severe()) { 801 /* 802 * We are low on memory, get it going NOW 803 */ 804 bawrite(bp); 805 } else { 806 /* 807 * In the middle of a cluster, so just delay the I/O for now. 808 */ 809 bdwrite(bp); 810 } 811 vp->v_lastw = lbn; 812 vp->v_lasta = bp->b_blkno; 813 } 814 815 816 /* 817 * This is an awful lot like cluster_rbuild...wish they could be combined. 818 * The last lbn argument is the current block on which I/O is being 819 * performed. Check to see that it doesn't fall in the middle of 820 * the current block (if last_bp == NULL). 821 */ 822 int 823 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len, 824 int gbflags) 825 { 826 struct buf *bp, *tbp; 827 struct bufobj *bo; 828 int i, j; 829 int totalwritten = 0; 830 int dbsize = btodb(size); 831 832 if (!unmapped_buf_allowed) 833 gbflags &= ~GB_UNMAPPED; 834 835 bo = &vp->v_bufobj; 836 while (len > 0) { 837 /* 838 * If the buffer is not delayed-write (i.e. dirty), or it 839 * is delayed-write but either locked or inval, it cannot 840 * partake in the clustered write. 841 */ 842 BO_LOCK(bo); 843 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL || 844 (tbp->b_vflags & BV_BKGRDINPROG)) { 845 BO_UNLOCK(bo); 846 ++start_lbn; 847 --len; 848 continue; 849 } 850 if (BUF_LOCK(tbp, 851 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) { 852 ++start_lbn; 853 --len; 854 continue; 855 } 856 if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) { 857 BUF_UNLOCK(tbp); 858 ++start_lbn; 859 --len; 860 continue; 861 } 862 bremfree(tbp); 863 tbp->b_flags &= ~B_DONE; 864 865 /* 866 * Extra memory in the buffer, punt on this buffer. 867 * XXX we could handle this in most cases, but we would 868 * have to push the extra memory down to after our max 869 * possible cluster size and then potentially pull it back 870 * up if the cluster was terminated prematurely--too much 871 * hassle. 872 */ 873 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) != 874 (B_CLUSTEROK | B_VMIO)) || 875 (tbp->b_bcount != tbp->b_bufsize) || 876 (tbp->b_bcount != size) || 877 (len == 1) || 878 ((bp = uma_zalloc(cluster_pbuf_zone, 879 (vp->v_vflag & VV_MD) != 0 ? M_NOWAIT : M_WAITOK)) == NULL)) { 880 totalwritten += tbp->b_bufsize; 881 bawrite(tbp); 882 ++start_lbn; 883 --len; 884 continue; 885 } 886 887 /* 888 * We got a pbuf to make the cluster in. 889 * so initialise it. 890 */ 891 TAILQ_INIT(&bp->b_cluster.cluster_head); 892 bp->b_bcount = 0; 893 bp->b_bufsize = 0; 894 bp->b_npages = 0; 895 if (tbp->b_wcred != NOCRED) 896 bp->b_wcred = crhold(tbp->b_wcred); 897 898 bp->b_blkno = tbp->b_blkno; 899 bp->b_lblkno = tbp->b_lblkno; 900 bp->b_offset = tbp->b_offset; 901 902 /* 903 * We are synthesizing a buffer out of vm_page_t's, but 904 * if the block size is not page aligned then the starting 905 * address may not be either. Inherit the b_data offset 906 * from the original buffer. 907 */ 908 if ((gbflags & GB_UNMAPPED) == 0 || 909 (tbp->b_flags & B_VMIO) == 0) { 910 bp->b_data = (char *)((vm_offset_t)bp->b_data | 911 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 912 } else { 913 bp->b_data = unmapped_buf; 914 } 915 bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO | 916 B_NEEDCOMMIT)); 917 bp->b_iodone = cluster_callback; 918 pbgetvp(vp, bp); 919 /* 920 * From this location in the file, scan forward to see 921 * if there are buffers with adjacent data that need to 922 * be written as well. 923 */ 924 for (i = 0; i < len; ++i, ++start_lbn) { 925 if (i != 0) { /* If not the first buffer */ 926 /* 927 * If the adjacent data is not even in core it 928 * can't need to be written. 929 */ 930 BO_LOCK(bo); 931 if ((tbp = gbincore(bo, start_lbn)) == NULL || 932 (tbp->b_vflags & BV_BKGRDINPROG)) { 933 BO_UNLOCK(bo); 934 break; 935 } 936 937 /* 938 * If it IS in core, but has different 939 * characteristics, or is locked (which 940 * means it could be undergoing a background 941 * I/O or be in a weird state), then don't 942 * cluster with it. 943 */ 944 if (BUF_LOCK(tbp, 945 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, 946 BO_LOCKPTR(bo))) 947 break; 948 949 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK | 950 B_INVAL | B_DELWRI | B_NEEDCOMMIT)) 951 != (B_DELWRI | B_CLUSTEROK | 952 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) || 953 tbp->b_wcred != bp->b_wcred) { 954 BUF_UNLOCK(tbp); 955 break; 956 } 957 958 /* 959 * Check that the combined cluster 960 * would make sense with regard to pages 961 * and would not be too large 962 */ 963 if ((tbp->b_bcount != size) || 964 ((bp->b_blkno + (dbsize * i)) != 965 tbp->b_blkno) || 966 ((tbp->b_npages + bp->b_npages) > 967 (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) { 968 BUF_UNLOCK(tbp); 969 break; 970 } 971 972 /* 973 * Ok, it's passed all the tests, 974 * so remove it from the free list 975 * and mark it busy. We will use it. 976 */ 977 bremfree(tbp); 978 tbp->b_flags &= ~B_DONE; 979 } /* end of code for non-first buffers only */ 980 /* 981 * If the IO is via the VM then we do some 982 * special VM hackery (yuck). Since the buffer's 983 * block size may not be page-aligned it is possible 984 * for a page to be shared between two buffers. We 985 * have to get rid of the duplication when building 986 * the cluster. 987 */ 988 if (tbp->b_flags & B_VMIO) { 989 vm_page_t m; 990 991 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object); 992 if (i == 0) { 993 vfs_busy_pages_acquire(tbp); 994 } else { /* if not first buffer */ 995 for (j = 0; j < tbp->b_npages; j += 1) { 996 m = tbp->b_pages[j]; 997 if (vm_page_trysbusy(m) == 0) { 998 for (j--; j >= 0; j--) 999 vm_page_sunbusy( 1000 tbp->b_pages[j]); 1001 VM_OBJECT_WUNLOCK( 1002 tbp->b_object); 1003 bqrelse(tbp); 1004 goto finishcluster; 1005 } 1006 } 1007 } 1008 vm_object_pip_add(tbp->b_bufobj->bo_object, 1009 tbp->b_npages); 1010 for (j = 0; j < tbp->b_npages; j += 1) { 1011 m = tbp->b_pages[j]; 1012 if ((bp->b_npages == 0) || 1013 (bp->b_pages[bp->b_npages - 1] != m)) { 1014 bp->b_pages[bp->b_npages] = m; 1015 bp->b_npages++; 1016 } 1017 } 1018 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object); 1019 } 1020 bp->b_bcount += size; 1021 bp->b_bufsize += size; 1022 /* 1023 * If any of the clustered buffers have their 1024 * B_BARRIER flag set, transfer that request to 1025 * the cluster. 1026 */ 1027 bp->b_flags |= (tbp->b_flags & B_BARRIER); 1028 tbp->b_flags &= ~(B_DONE | B_BARRIER); 1029 tbp->b_flags |= B_ASYNC; 1030 tbp->b_ioflags &= ~BIO_ERROR; 1031 tbp->b_iocmd = BIO_WRITE; 1032 bundirty(tbp); 1033 reassignbuf(tbp); /* put on clean list */ 1034 bufobj_wref(tbp->b_bufobj); 1035 BUF_KERNPROC(tbp); 1036 buf_track(tbp, __func__); 1037 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, 1038 tbp, b_cluster.cluster_entry); 1039 } 1040 finishcluster: 1041 if (buf_mapped(bp)) { 1042 pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 1043 (vm_page_t *)bp->b_pages, bp->b_npages); 1044 } 1045 if (bp->b_bufsize > bp->b_kvasize) 1046 panic( 1047 "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n", 1048 bp->b_bufsize, bp->b_kvasize); 1049 totalwritten += bp->b_bufsize; 1050 bp->b_dirtyoff = 0; 1051 bp->b_dirtyend = bp->b_bufsize; 1052 bawrite(bp); 1053 1054 len -= i; 1055 } 1056 return totalwritten; 1057 } 1058 1059 /* 1060 * Collect together all the buffers in a cluster. 1061 * Plus add one additional buffer. 1062 */ 1063 static struct cluster_save * 1064 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags) 1065 { 1066 struct cluster_save *buflist; 1067 struct buf *bp; 1068 daddr_t lbn; 1069 int i, j, len, error; 1070 1071 len = vp->v_lastw - vp->v_cstart + 1; 1072 buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist), 1073 M_SEGMENT, M_WAITOK); 1074 buflist->bs_nchildren = 0; 1075 buflist->bs_children = (struct buf **) (buflist + 1); 1076 for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) { 1077 error = bread_gb(vp, lbn, last_bp->b_bcount, NOCRED, 1078 gbflags, &bp); 1079 if (error != 0) { 1080 /* 1081 * If read fails, release collected buffers 1082 * and return failure. 1083 */ 1084 for (j = 0; j < i; j++) 1085 brelse(buflist->bs_children[j]); 1086 free(buflist, M_SEGMENT); 1087 return (NULL); 1088 } 1089 buflist->bs_children[i] = bp; 1090 if (bp->b_blkno == bp->b_lblkno) 1091 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, 1092 NULL, NULL); 1093 } 1094 buflist->bs_children[i] = bp = last_bp; 1095 if (bp->b_blkno == bp->b_lblkno) 1096 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1097 buflist->bs_nchildren = i + 1; 1098 return (buflist); 1099 } 1100