vfs_cluster.c revision 86089
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: head/sys/kern/vfs_cluster.c 86089 2001-11-05 18:48:54Z dillon $ 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#include <sys/sysctl.h> 59static int rcluster= 0; 60SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, ""); 61#endif 62 63static MALLOC_DEFINE(M_SEGMENT, "cluster_save buffer", "cluster_save buffer"); 64 65static struct cluster_save * 66 cluster_collectbufs __P((struct vnode *vp, struct buf *last_bp)); 67static struct buf * 68 cluster_rbuild __P((struct vnode *vp, u_quad_t filesize, daddr_t lbn, 69 daddr_t blkno, long size, int run, struct buf *fbp)); 70 71static int write_behind = 1; 72SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, ""); 73 74extern vm_page_t bogus_page; 75 76extern int cluster_pbuf_freecnt; 77 78/* 79 * Maximum number of blocks for read-ahead. 80 */ 81#define MAXRA 32 82 83/* 84 * This replaces bread. 85 */ 86int 87cluster_read(vp, filesize, lblkno, size, cred, totread, seqcount, bpp) 88 struct vnode *vp; 89 u_quad_t filesize; 90 daddr_t lblkno; 91 long size; 92 struct ucred *cred; 93 long totread; 94 int seqcount; 95 struct buf **bpp; 96{ 97 struct buf *bp, *rbp, *reqbp; 98 daddr_t blkno, origblkno; 99 int error, num_ra; 100 int i; 101 int maxra, racluster; 102 long origtotread; 103 104 error = 0; 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 = 2 * racluster + (totread / size); 112 if (maxra > MAXRA) 113 maxra = MAXRA; 114 if (maxra > nbuf/8) 115 maxra = nbuf/8; 116 117 /* 118 * get the requested block 119 */ 120 *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0); 121 origblkno = lblkno; 122 origtotread = totread; 123 124 /* 125 * if it is in the cache, then check to see if the reads have been 126 * sequential. If they have, then try some read-ahead, otherwise 127 * back-off on prospective read-aheads. 128 */ 129 if (bp->b_flags & B_CACHE) { 130 if (!seqcount) { 131 return 0; 132 } else if ((bp->b_flags & B_RAM) == 0) { 133 return 0; 134 } else { 135 int s; 136 struct buf *tbp; 137 bp->b_flags &= ~B_RAM; 138 /* 139 * We do the spl here so that there is no window 140 * between the incore and the b_usecount increment 141 * below. We opt to keep the spl out of the loop 142 * for efficiency. 143 */ 144 s = splbio(); 145 for (i = 1; i < maxra; i++) { 146 147 if (!(tbp = incore(vp, lblkno+i))) { 148 break; 149 } 150 151 /* 152 * Set another read-ahead mark so we know 153 * to check again. 154 */ 155 if (((i % racluster) == (racluster - 1)) || 156 (i == (maxra - 1))) 157 tbp->b_flags |= B_RAM; 158 } 159 splx(s); 160 if (i >= maxra) { 161 return 0; 162 } 163 lblkno += i; 164 } 165 reqbp = bp = NULL; 166 } else { 167 off_t firstread = bp->b_offset; 168 169 KASSERT(bp->b_offset != NOOFFSET, 170 ("cluster_read: no buffer offset")); 171 if (firstread + totread > filesize) 172 totread = filesize - firstread; 173 if (totread > size) { 174 int nblks = 0; 175 int ncontigafter; 176 while (totread > 0) { 177 nblks++; 178 totread -= size; 179 } 180 if (nblks == 1) 181 goto single_block_read; 182 if (nblks > racluster) 183 nblks = racluster; 184 185 error = VOP_BMAP(vp, lblkno, NULL, 186 &blkno, &ncontigafter, NULL); 187 if (error) 188 goto single_block_read; 189 if (blkno == -1) 190 goto single_block_read; 191 if (ncontigafter == 0) 192 goto single_block_read; 193 if (ncontigafter + 1 < nblks) 194 nblks = ncontigafter + 1; 195 196 bp = cluster_rbuild(vp, filesize, lblkno, 197 blkno, size, nblks, bp); 198 lblkno += (bp->b_bufsize / size); 199 } else { 200single_block_read: 201 /* 202 * if it isn't in the cache, then get a chunk from 203 * disk if sequential, otherwise just get the block. 204 */ 205 bp->b_flags |= B_RAM; 206 bp->b_iocmd = BIO_READ; 207 lblkno += 1; 208 } 209 } 210 211 /* 212 * if we have been doing sequential I/O, then do some read-ahead 213 */ 214 rbp = NULL; 215 if (seqcount && (lblkno < (origblkno + seqcount))) { 216 /* 217 * we now build the read-ahead buffer if it is desirable. 218 */ 219 if (((u_quad_t)(lblkno + 1) * size) <= filesize && 220 !(error = VOP_BMAP(vp, lblkno, NULL, &blkno, &num_ra, NULL)) && 221 blkno != -1) { 222 int nblksread; 223 int ntoread = num_ra + 1; 224 nblksread = (origtotread + size - 1) / size; 225 if (seqcount < nblksread) 226 seqcount = nblksread; 227 if (seqcount < ntoread) 228 ntoread = seqcount; 229 if (num_ra) { 230 rbp = cluster_rbuild(vp, filesize, lblkno, 231 blkno, size, ntoread, NULL); 232 } else { 233 rbp = getblk(vp, lblkno, size, 0, 0); 234 rbp->b_flags |= B_ASYNC | B_RAM; 235 rbp->b_iocmd = BIO_READ; 236 rbp->b_blkno = blkno; 237 } 238 } 239 } 240 241 /* 242 * handle the synchronous read 243 */ 244 if (bp) { 245#if defined(CLUSTERDEBUG) 246 if (rcluster) 247 printf("S(%ld,%ld,%d) ", 248 (long)bp->b_lblkno, bp->b_bcount, seqcount); 249#endif 250 if ((bp->b_flags & B_CLUSTER) == 0) { 251 vfs_busy_pages(bp, 0); 252 } 253 bp->b_flags &= ~B_INVAL; 254 bp->b_ioflags &= ~BIO_ERROR; 255 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL) 256 BUF_KERNPROC(bp); 257 error = VOP_STRATEGY(vp, bp); 258 curproc->p_stats->p_ru.ru_inblock++; 259 } 260 261 /* 262 * and if we have read-aheads, do them too 263 */ 264 if (rbp) { 265 if (error) { 266 rbp->b_flags &= ~B_ASYNC; 267 brelse(rbp); 268 } else if (rbp->b_flags & B_CACHE) { 269 rbp->b_flags &= ~B_ASYNC; 270 bqrelse(rbp); 271 } else { 272#if defined(CLUSTERDEBUG) 273 if (rcluster) { 274 if (bp) 275 printf("A+(%ld,%ld,%ld,%d) ", 276 (long)rbp->b_lblkno, rbp->b_bcount, 277 (long)(rbp->b_lblkno - origblkno), 278 seqcount); 279 else 280 printf("A(%ld,%ld,%ld,%d) ", 281 (long)rbp->b_lblkno, rbp->b_bcount, 282 (long)(rbp->b_lblkno - origblkno), 283 seqcount); 284 } 285#endif 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 */ 309static struct buf * 310cluster_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); 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 /* 388 * Shortcut some checks and try to avoid buffers that 389 * would block in the lock. The same checks have to 390 * be made again after we officially get the buffer. 391 */ 392 if ((tbp = incore(vp, lbn + i)) != NULL) { 393 if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) 394 break; 395 BUF_UNLOCK(tbp); 396 397 for (j = 0; j < tbp->b_npages; j++) { 398 if (tbp->b_pages[j]->valid) 399 break; 400 } 401 402 if (j != tbp->b_npages) 403 break; 404 405 if (tbp->b_bcount != size) 406 break; 407 } 408 409 tbp = getblk(vp, lbn + i, size, 0, 0); 410 411 /* 412 * Stop scanning if the buffer is fully valid 413 * (marked B_CACHE), or locked (may be doing a 414 * background write), or if the buffer is not 415 * VMIO backed. The clustering code can only deal 416 * with VMIO-backed buffers. 417 */ 418 if ((tbp->b_flags & (B_CACHE|B_LOCKED)) || 419 (tbp->b_flags & B_VMIO) == 0) { 420 bqrelse(tbp); 421 break; 422 } 423 424 /* 425 * The buffer must be completely invalid in order to 426 * take part in the cluster. If it is partially valid 427 * then we stop. 428 */ 429 for (j = 0;j < tbp->b_npages; j++) { 430 if (tbp->b_pages[j]->valid) 431 break; 432 } 433 if (j != tbp->b_npages) { 434 bqrelse(tbp); 435 break; 436 } 437 438 /* 439 * Set a read-ahead mark as appropriate 440 */ 441 if ((fbp && (i == 1)) || (i == (run - 1))) 442 tbp->b_flags |= B_RAM; 443 444 /* 445 * Set the buffer up for an async read (XXX should 446 * we do this only if we do not wind up brelse()ing?). 447 * Set the block number if it isn't set, otherwise 448 * if it is make sure it matches the block number we 449 * expect. 450 */ 451 tbp->b_flags |= B_ASYNC; 452 tbp->b_iocmd = BIO_READ; 453 if (tbp->b_blkno == tbp->b_lblkno) { 454 tbp->b_blkno = bn; 455 } else if (tbp->b_blkno != bn) { 456 brelse(tbp); 457 break; 458 } 459 } 460 /* 461 * XXX fbp from caller may not be B_ASYNC, but we are going 462 * to biodone() it in cluster_callback() anyway 463 */ 464 BUF_KERNPROC(tbp); 465 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, 466 tbp, b_cluster.cluster_entry); 467 for (j = 0; j < tbp->b_npages; j += 1) { 468 vm_page_t m; 469 m = tbp->b_pages[j]; 470 vm_page_io_start(m); 471 vm_object_pip_add(m->object, 1); 472 if ((bp->b_npages == 0) || 473 (bp->b_pages[bp->b_npages-1] != m)) { 474 bp->b_pages[bp->b_npages] = m; 475 bp->b_npages++; 476 } 477 if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) 478 tbp->b_pages[j] = bogus_page; 479 } 480 /* 481 * XXX shouldn't this be += size for both, like in 482 * cluster_wbuild()? 483 * 484 * Don't inherit tbp->b_bufsize as it may be larger due to 485 * a non-page-aligned size. Instead just aggregate using 486 * 'size'. 487 */ 488 if (tbp->b_bcount != size) 489 printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size); 490 if (tbp->b_bufsize != size) 491 printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size); 492 bp->b_bcount += size; 493 bp->b_bufsize += size; 494 } 495 496 /* 497 * Fully valid pages in the cluster are already good and do not need 498 * to be re-read from disk. Replace the page with bogus_page 499 */ 500 for (j = 0; j < bp->b_npages; j++) { 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 if (bp->b_bufsize > bp->b_kvasize) 507 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n", 508 bp->b_bufsize, bp->b_kvasize); 509 bp->b_kvasize = bp->b_bufsize; 510 511 pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 512 (vm_page_t *)bp->b_pages, bp->b_npages); 513 return (bp); 514} 515 516/* 517 * Cleanup after a clustered read or write. 518 * This is complicated by the fact that any of the buffers might have 519 * extra memory (if there were no empty buffer headers at allocbuf time) 520 * that we will need to shift around. 521 */ 522void 523cluster_callback(bp) 524 struct buf *bp; 525{ 526 struct buf *nbp, *tbp; 527 int error = 0; 528 529 GIANT_REQUIRED; 530 531 /* 532 * Must propogate errors to all the components. 533 */ 534 if (bp->b_ioflags & BIO_ERROR) 535 error = bp->b_error; 536 537 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 538 /* 539 * Move memory from the large cluster buffer into the component 540 * buffers and mark IO as done on these. 541 */ 542 for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head); 543 tbp; tbp = nbp) { 544 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry); 545 if (error) { 546 tbp->b_ioflags |= BIO_ERROR; 547 tbp->b_error = error; 548 } else { 549 tbp->b_dirtyoff = tbp->b_dirtyend = 0; 550 tbp->b_flags &= ~B_INVAL; 551 tbp->b_ioflags &= ~BIO_ERROR; 552 /* 553 * XXX the bdwrite()/bqrelse() issued during 554 * cluster building clears B_RELBUF (see bqrelse() 555 * comment). If direct I/O was specified, we have 556 * to restore it here to allow the buffer and VM 557 * to be freed. 558 */ 559 if (tbp->b_flags & B_DIRECT) 560 tbp->b_flags |= B_RELBUF; 561 } 562 bufdone(tbp); 563 } 564 relpbuf(bp, &cluster_pbuf_freecnt); 565} 566 567/* 568 * cluster_wbuild_wb: 569 * 570 * Implement modified write build for cluster. 571 * 572 * write_behind = 0 write behind disabled 573 * write_behind = 1 write behind normal (default) 574 * write_behind = 2 write behind backed-off 575 */ 576 577static __inline int 578cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len) 579{ 580 int r = 0; 581 582 switch(write_behind) { 583 case 2: 584 if (start_lbn < len) 585 break; 586 start_lbn -= len; 587 /* fall through */ 588 case 1: 589 r = cluster_wbuild(vp, size, start_lbn, len); 590 /* fall through */ 591 default: 592 /* fall through */ 593 break; 594 } 595 return(r); 596} 597 598/* 599 * Do clustered write for FFS. 600 * 601 * Three cases: 602 * 1. Write is not sequential (write asynchronously) 603 * Write is sequential: 604 * 2. beginning of cluster - begin cluster 605 * 3. middle of a cluster - add to cluster 606 * 4. end of a cluster - asynchronously write cluster 607 */ 608void 609cluster_write(bp, filesize, seqcount) 610 struct buf *bp; 611 u_quad_t filesize; 612 int seqcount; 613{ 614 struct vnode *vp; 615 daddr_t lbn; 616 int maxclen, cursize; 617 int lblocksize; 618 int async; 619 620 vp = bp->b_vp; 621 if (vp->v_type == VREG) { 622 async = vp->v_mount->mnt_flag & MNT_ASYNC; 623 lblocksize = vp->v_mount->mnt_stat.f_iosize; 624 } else { 625 async = 0; 626 lblocksize = bp->b_bufsize; 627 } 628 lbn = bp->b_lblkno; 629 KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset")); 630 631 /* Initialize vnode to beginning of file. */ 632 if (lbn == 0) 633 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; 634 635 if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 || 636 (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) { 637 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1; 638 if (vp->v_clen != 0) { 639 /* 640 * Next block is not sequential. 641 * 642 * If we are not writing at end of file, the process 643 * seeked to another point in the file since its last 644 * write, or we have reached our maximum cluster size, 645 * then push the previous cluster. Otherwise try 646 * reallocating to make it sequential. 647 * 648 * Change to algorithm: only push previous cluster if 649 * it was sequential from the point of view of the 650 * seqcount heuristic, otherwise leave the buffer 651 * intact so we can potentially optimize the I/O 652 * later on in the buf_daemon or update daemon 653 * flush. 654 */ 655 cursize = vp->v_lastw - vp->v_cstart + 1; 656 if (((u_quad_t) bp->b_offset + lblocksize) != filesize || 657 lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) { 658 if (!async && seqcount > 0) { 659 cluster_wbuild_wb(vp, lblocksize, 660 vp->v_cstart, cursize); 661 } 662 } else { 663 struct buf **bpp, **endbp; 664 struct cluster_save *buflist; 665 666 buflist = cluster_collectbufs(vp, bp); 667 endbp = &buflist->bs_children 668 [buflist->bs_nchildren - 1]; 669 if (VOP_REALLOCBLKS(vp, buflist)) { 670 /* 671 * Failed, push the previous cluster 672 * if *really* writing sequentially 673 * in the logical file (seqcount > 1), 674 * otherwise delay it in the hopes that 675 * the low level disk driver can 676 * optimize the write ordering. 677 */ 678 for (bpp = buflist->bs_children; 679 bpp < endbp; bpp++) 680 brelse(*bpp); 681 free(buflist, M_SEGMENT); 682 if (seqcount > 1) { 683 cluster_wbuild_wb(vp, 684 lblocksize, vp->v_cstart, 685 cursize); 686 } 687 } else { 688 /* 689 * Succeeded, keep building cluster. 690 */ 691 for (bpp = buflist->bs_children; 692 bpp <= endbp; bpp++) 693 bdwrite(*bpp); 694 free(buflist, M_SEGMENT); 695 vp->v_lastw = lbn; 696 vp->v_lasta = bp->b_blkno; 697 return; 698 } 699 } 700 } 701 /* 702 * Consider beginning a cluster. If at end of file, make 703 * cluster as large as possible, otherwise find size of 704 * existing cluster. 705 */ 706 if ((vp->v_type == VREG) && 707 ((u_quad_t) bp->b_offset + lblocksize) != filesize && 708 (bp->b_blkno == bp->b_lblkno) && 709 (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) || 710 bp->b_blkno == -1)) { 711 bawrite(bp); 712 vp->v_clen = 0; 713 vp->v_lasta = bp->b_blkno; 714 vp->v_cstart = lbn + 1; 715 vp->v_lastw = lbn; 716 return; 717 } 718 vp->v_clen = maxclen; 719 if (!async && maxclen == 0) { /* I/O not contiguous */ 720 vp->v_cstart = lbn + 1; 721 bawrite(bp); 722 } else { /* Wait for rest of cluster */ 723 vp->v_cstart = lbn; 724 bdwrite(bp); 725 } 726 } else if (lbn == vp->v_cstart + vp->v_clen) { 727 /* 728 * At end of cluster, write it out if seqcount tells us we 729 * are operating sequentially, otherwise let the buf or 730 * update daemon handle it. 731 */ 732 bdwrite(bp); 733 if (seqcount > 1) 734 cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1); 735 vp->v_clen = 0; 736 vp->v_cstart = lbn + 1; 737 } else if (vm_page_count_severe()) { 738 /* 739 * We are low on memory, get it going NOW 740 */ 741 bawrite(bp); 742 } else { 743 /* 744 * In the middle of a cluster, so just delay the I/O for now. 745 */ 746 bdwrite(bp); 747 } 748 vp->v_lastw = lbn; 749 vp->v_lasta = bp->b_blkno; 750} 751 752 753/* 754 * This is an awful lot like cluster_rbuild...wish they could be combined. 755 * The last lbn argument is the current block on which I/O is being 756 * performed. Check to see that it doesn't fall in the middle of 757 * the current block (if last_bp == NULL). 758 */ 759int 760cluster_wbuild(vp, size, start_lbn, len) 761 struct vnode *vp; 762 long size; 763 daddr_t start_lbn; 764 int len; 765{ 766 struct buf *bp, *tbp; 767 int i, j, s; 768 int totalwritten = 0; 769 int dbsize = btodb(size); 770 771 GIANT_REQUIRED; 772 773 while (len > 0) { 774 s = splbio(); 775 /* 776 * If the buffer is not delayed-write (i.e. dirty), or it 777 * is delayed-write but either locked or inval, it cannot 778 * partake in the clustered write. 779 */ 780 if (((tbp = gbincore(vp, start_lbn)) == NULL) || 781 ((tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI) || 782 BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) { 783 ++start_lbn; 784 --len; 785 splx(s); 786 continue; 787 } 788 bremfree(tbp); 789 tbp->b_flags &= ~B_DONE; 790 splx(s); 791 792 /* 793 * Extra memory in the buffer, punt on this buffer. 794 * XXX we could handle this in most cases, but we would 795 * have to push the extra memory down to after our max 796 * possible cluster size and then potentially pull it back 797 * up if the cluster was terminated prematurely--too much 798 * hassle. 799 */ 800 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) != 801 (B_CLUSTEROK | B_VMIO)) || 802 (tbp->b_bcount != tbp->b_bufsize) || 803 (tbp->b_bcount != size) || 804 (len == 1) || 805 ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) { 806 totalwritten += tbp->b_bufsize; 807 bawrite(tbp); 808 ++start_lbn; 809 --len; 810 continue; 811 } 812 813 /* 814 * We got a pbuf to make the cluster in. 815 * so initialise it. 816 */ 817 TAILQ_INIT(&bp->b_cluster.cluster_head); 818 bp->b_bcount = 0; 819 bp->b_magic = tbp->b_magic; 820 bp->b_op = tbp->b_op; 821 bp->b_bufsize = 0; 822 bp->b_npages = 0; 823 if (tbp->b_wcred != NOCRED) 824 bp->b_wcred = crhold(tbp->b_wcred); 825 826 bp->b_blkno = tbp->b_blkno; 827 bp->b_lblkno = tbp->b_lblkno; 828 bp->b_offset = tbp->b_offset; 829 830 /* 831 * We are synthesizing a buffer out of vm_page_t's, but 832 * if the block size is not page aligned then the starting 833 * address may not be either. Inherit the b_data offset 834 * from the original buffer. 835 */ 836 bp->b_data = (char *)((vm_offset_t)bp->b_data | 837 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 838 bp->b_flags |= B_CLUSTER | 839 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT | B_NOWDRAIN)); 840 bp->b_iodone = cluster_callback; 841 pbgetvp(vp, bp); 842 /* 843 * From this location in the file, scan forward to see 844 * if there are buffers with adjacent data that need to 845 * be written as well. 846 */ 847 for (i = 0; i < len; ++i, ++start_lbn) { 848 if (i != 0) { /* If not the first buffer */ 849 s = splbio(); 850 /* 851 * If the adjacent data is not even in core it 852 * can't need to be written. 853 */ 854 if ((tbp = gbincore(vp, start_lbn)) == NULL) { 855 splx(s); 856 break; 857 } 858 859 /* 860 * If it IS in core, but has different 861 * characteristics, or is locked (which 862 * means it could be undergoing a background 863 * I/O or be in a weird state), then don't 864 * cluster with it. 865 */ 866 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK | 867 B_INVAL | B_DELWRI | B_NEEDCOMMIT)) 868 != (B_DELWRI | B_CLUSTEROK | 869 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) || 870 (tbp->b_flags & B_LOCKED) || 871 tbp->b_wcred != bp->b_wcred || 872 BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) { 873 splx(s); 874 break; 875 } 876 877 /* 878 * Check that the combined cluster 879 * would make sense with regard to pages 880 * and would not be too large 881 */ 882 if ((tbp->b_bcount != size) || 883 ((bp->b_blkno + (dbsize * i)) != 884 tbp->b_blkno) || 885 ((tbp->b_npages + bp->b_npages) > 886 (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) { 887 BUF_UNLOCK(tbp); 888 splx(s); 889 break; 890 } 891 /* 892 * Ok, it's passed all the tests, 893 * so remove it from the free list 894 * and mark it busy. We will use it. 895 */ 896 bremfree(tbp); 897 tbp->b_flags &= ~B_DONE; 898 splx(s); 899 } /* end of code for non-first buffers only */ 900 /* check for latent dependencies to be handled */ 901 if ((LIST_FIRST(&tbp->b_dep)) != NULL) 902 buf_start(tbp); 903 /* 904 * If the IO is via the VM then we do some 905 * special VM hackery (yuck). Since the buffer's 906 * block size may not be page-aligned it is possible 907 * for a page to be shared between two buffers. We 908 * have to get rid of the duplication when building 909 * the cluster. 910 */ 911 if (tbp->b_flags & B_VMIO) { 912 vm_page_t m; 913 914 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->flags & PG_BUSY) { 918 bqrelse(tbp); 919 goto finishcluster; 920 } 921 } 922 } 923 924 for (j = 0; j < tbp->b_npages; j += 1) { 925 m = tbp->b_pages[j]; 926 vm_page_io_start(m); 927 vm_object_pip_add(m->object, 1); 928 if ((bp->b_npages == 0) || 929 (bp->b_pages[bp->b_npages - 1] != m)) { 930 bp->b_pages[bp->b_npages] = m; 931 bp->b_npages++; 932 } 933 } 934 } 935 bp->b_bcount += size; 936 bp->b_bufsize += size; 937 938 s = splbio(); 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, tbp->b_vp); /* put on clean list */ 945 ++tbp->b_vp->v_numoutput; 946 splx(s); 947 BUF_KERNPROC(tbp); 948 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, 949 tbp, b_cluster.cluster_entry); 950 } 951 finishcluster: 952 pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 953 (vm_page_t *) bp->b_pages, bp->b_npages); 954 if (bp->b_bufsize > bp->b_kvasize) 955 panic( 956 "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n", 957 bp->b_bufsize, bp->b_kvasize); 958 bp->b_kvasize = bp->b_bufsize; 959 totalwritten += bp->b_bufsize; 960 bp->b_dirtyoff = 0; 961 bp->b_dirtyend = bp->b_bufsize; 962 bawrite(bp); 963 964 len -= i; 965 } 966 return totalwritten; 967} 968 969/* 970 * Collect together all the buffers in a cluster. 971 * Plus add one additional buffer. 972 */ 973static struct cluster_save * 974cluster_collectbufs(vp, last_bp) 975 struct vnode *vp; 976 struct buf *last_bp; 977{ 978 struct cluster_save *buflist; 979 struct buf *bp; 980 daddr_t lbn; 981 int i, len; 982 983 len = vp->v_lastw - vp->v_cstart + 1; 984 buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist), 985 M_SEGMENT, M_WAITOK); 986 buflist->bs_nchildren = 0; 987 buflist->bs_children = (struct buf **) (buflist + 1); 988 for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) { 989 (void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp); 990 buflist->bs_children[i] = bp; 991 if (bp->b_blkno == bp->b_lblkno) 992 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, 993 NULL, NULL); 994 } 995 buflist->bs_children[i] = bp = last_bp; 996 if (bp->b_blkno == bp->b_lblkno) 997 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, 998 NULL, NULL); 999 buflist->bs_nchildren = i + 1; 1000 return (buflist); 1001} 1002