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: src/sys/kern/vfs_cluster.c,v 1.92.2.9 2001/11/18 07:10:59 dillon Exp $ 37 * $DragonFly: src/sys/kern/vfs_cluster.c,v 1.39 2008/06/20 05:40:04 dillon Exp $ 38 */ 39 40 #include "opt_debug_cluster.h" 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/kernel.h> 45 #include <sys/proc.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 #include <sys/buf2.h> 57 #include <vm/vm_page2.h> 58 59 #if defined(CLUSTERDEBUG) 60 #include <sys/sysctl.h> 61 static int rcluster= 0; 62 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, ""); 63 #endif 64 65 static MALLOC_DEFINE(M_SEGMENT, "cluster_save", "cluster_save buffer"); 66 67 static struct cluster_save * 68 cluster_collectbufs (struct vnode *vp, struct buf *last_bp, 69 int blksize); 70 static struct buf * 71 cluster_rbuild (struct vnode *vp, off_t filesize, off_t loffset, 72 off_t doffset, int blksize, int run, 73 struct buf *fbp, int doasync); 74 static void cluster_callback (struct bio *); 75 76 77 static int write_behind = 1; 78 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, ""); 79 80 extern vm_page_t bogus_page; 81 82 extern int cluster_pbuf_freecnt; 83 84 /* 85 * Maximum number of blocks for read-ahead. 86 */ 87 #define MAXRA 32 88 89 /* 90 * This replaces bread. 91 */ 92 int 93 cluster_read(struct vnode *vp, off_t filesize, off_t loffset, 94 int blksize, int totread, int seqcount, struct buf **bpp) 95 { 96 struct buf *bp, *rbp, *reqbp; 97 off_t origoffset; 98 off_t doffset; 99 int error; 100 int i; 101 int maxra, racluster; 102 103 error = 0; 104 105 /* 106 * Try to limit the amount of read-ahead by a few 107 * ad-hoc parameters. This needs work!!! 108 */ 109 racluster = vmaxiosize(vp) / blksize; 110 maxra = 2 * racluster + (totread / blksize); 111 if (maxra > MAXRA) 112 maxra = MAXRA; 113 if (maxra > nbuf/8) 114 maxra = nbuf/8; 115 116 /* 117 * get the requested block 118 */ 119 *bpp = reqbp = bp = getblk(vp, loffset, blksize, 0, 0); 120 origoffset = loffset; 121 122 /* 123 * if it is in the cache, then check to see if the reads have been 124 * sequential. If they have, then try some read-ahead, otherwise 125 * back-off on prospective read-aheads. 126 */ 127 if (bp->b_flags & B_CACHE) { 128 if (!seqcount) { 129 return 0; 130 } else if ((bp->b_flags & B_RAM) == 0) { 131 return 0; 132 } else { 133 struct buf *tbp; 134 bp->b_flags &= ~B_RAM; 135 /* 136 * We do the crit here so that there is no window 137 * between the findblk and the b_usecount increment 138 * below. We opt to keep the crit out of the loop 139 * for efficiency. 140 */ 141 crit_enter(); 142 for (i = 1; i < maxra; i++) { 143 if (!(tbp = findblk(vp, loffset + i * blksize))) { 144 break; 145 } 146 147 /* 148 * Set another read-ahead mark so we know 149 * to check again. 150 */ 151 if (((i % racluster) == (racluster - 1)) || 152 (i == (maxra - 1))) 153 tbp->b_flags |= B_RAM; 154 } 155 crit_exit(); 156 if (i >= maxra) { 157 return 0; 158 } 159 loffset += i * blksize; 160 } 161 reqbp = bp = NULL; 162 } else { 163 off_t firstread = bp->b_loffset; 164 int nblks; 165 166 KASSERT(firstread != NOOFFSET, 167 ("cluster_read: no buffer offset")); 168 if (firstread + totread > filesize) 169 totread = (int)(filesize - firstread); 170 nblks = totread / blksize; 171 if (nblks) { 172 int burstbytes; 173 174 if (nblks > racluster) 175 nblks = racluster; 176 177 error = VOP_BMAP(vp, loffset, &doffset, 178 &burstbytes, NULL, BUF_CMD_READ); 179 if (error) 180 goto single_block_read; 181 if (doffset == NOOFFSET) 182 goto single_block_read; 183 if (burstbytes < blksize * 2) 184 goto single_block_read; 185 if (nblks > burstbytes / blksize) 186 nblks = burstbytes / blksize; 187 188 bp = cluster_rbuild(vp, filesize, loffset, 189 doffset, blksize, nblks, bp, 0); 190 loffset += bp->b_bufsize; 191 } else { 192 single_block_read: 193 /* 194 * if it isn't in the cache, then get a chunk from 195 * disk if sequential, otherwise just get the block. 196 */ 197 bp->b_flags |= B_RAM; 198 loffset += blksize; 199 } 200 } 201 202 /* 203 * Handle the synchronous read. This only occurs if B_CACHE was 204 * not set. bp (and rbp) could be either a cluster bp or a normal 205 * bp depending on the what cluster_rbuild() decided to do. If 206 * it is a cluster bp, vfs_busy_pages() has already been called. 207 */ 208 if (bp) { 209 #if defined(CLUSTERDEBUG) 210 if (rcluster) 211 kprintf("S(%lld,%d,%d) ", 212 bp->b_loffset, bp->b_bcount, seqcount); 213 #endif 214 bp->b_cmd = BUF_CMD_READ; 215 if ((bp->b_flags & B_CLUSTER) == 0) 216 vfs_busy_pages(vp, bp); 217 bp->b_flags &= ~(B_ERROR|B_INVAL); 218 if ((bp->b_flags & B_ASYNC) || bp->b_bio1.bio_done != NULL) 219 BUF_KERNPROC(bp); 220 vn_strategy(vp, &bp->b_bio1); 221 if (bp->b_flags & B_ERROR) { 222 if ((error = bp->b_error) == 0) 223 error = EIO; 224 } else { 225 error = 0; 226 } 227 } 228 229 /* 230 * If we have been doing sequential I/O, then do some read-ahead. 231 * 232 * Only mess with buffers which we can immediately lock. HAMMER 233 * will do device-readahead irrespective of what the blocks 234 * represent. 235 */ 236 rbp = NULL; 237 if (!error && 238 seqcount && 239 loffset < origoffset + seqcount * blksize && 240 loffset + blksize <= filesize 241 ) { 242 int nblksread; 243 int ntoread; 244 int burstbytes; 245 int tmp_error; 246 247 if ((rbp = findblk(vp, loffset)) != NULL) { 248 if (BUF_LOCK(rbp, LK_EXCLUSIVE | LK_NOWAIT)) { 249 goto no_read_ahead; 250 } 251 BUF_UNLOCK(rbp); 252 } 253 rbp = getblk(vp, loffset, blksize, 0, 0); 254 if ((rbp->b_flags & B_CACHE)) { 255 bqrelse(rbp); 256 goto no_read_ahead; 257 } 258 259 /* 260 * An error from the read-ahead bmap has nothing to do 261 * with the caller's original request. 262 */ 263 tmp_error = VOP_BMAP(vp, loffset, &doffset, 264 &burstbytes, NULL, BUF_CMD_READ); 265 if (tmp_error || doffset == NOOFFSET) { 266 rbp->b_flags |= B_INVAL; 267 brelse(rbp); 268 rbp = NULL; 269 goto no_read_ahead; 270 } 271 ntoread = burstbytes / blksize; 272 nblksread = (totread + blksize - 1) / blksize; 273 if (seqcount < nblksread) 274 seqcount = nblksread; 275 if (ntoread > seqcount) 276 ntoread = seqcount; 277 278 rbp->b_flags |= B_RAM/* | B_AGE*/; 279 if (burstbytes) { 280 rbp = cluster_rbuild(vp, filesize, loffset, 281 doffset, blksize, 282 ntoread, rbp, 1); 283 } else { 284 rbp->b_bio2.bio_offset = doffset; 285 } 286 #if defined(CLUSTERDEBUG) 287 if (rcluster) { 288 if (bp) 289 kprintf("A+(%lld,%d,%lld,%d) ", 290 rbp->b_loffset, rbp->b_bcount, 291 rbp->b_loffset - origoffset, 292 seqcount); 293 else 294 kprintf("A(%lld,%d,%lld,%d) ", 295 rbp->b_loffset, rbp->b_bcount, 296 rbp->b_loffset - origoffset, 297 seqcount); 298 } 299 #endif 300 rbp->b_flags &= ~(B_ERROR|B_INVAL); 301 rbp->b_flags |= B_ASYNC; 302 rbp->b_cmd = BUF_CMD_READ; 303 304 if ((rbp->b_flags & B_CLUSTER) == 0) 305 vfs_busy_pages(vp, rbp); 306 BUF_KERNPROC(rbp); /* B_ASYNC */ 307 vn_strategy(vp, &rbp->b_bio1); 308 } 309 no_read_ahead: 310 311 if (reqbp) 312 return (biowait(reqbp)); 313 else 314 return (error); 315 } 316 317 /* 318 * If blocks are contiguous on disk, use this to provide clustered 319 * read ahead. We will read as many blocks as possible sequentially 320 * and then parcel them up into logical blocks in the buffer hash table. 321 */ 322 static struct buf * 323 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset, 324 off_t doffset, int blksize, int run, struct buf *fbp, int doasync) 325 { 326 struct buf *bp, *tbp; 327 off_t boffset; 328 int i, j; 329 int maxiosize = vmaxiosize(vp); 330 331 /* 332 * avoid a division 333 */ 334 while (loffset + run * blksize > filesize) { 335 --run; 336 } 337 338 tbp = fbp; 339 tbp->b_bio2.bio_offset = doffset; 340 if((tbp->b_flags & B_MALLOC) || 341 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) { 342 return tbp; 343 } 344 345 bp = trypbuf(&cluster_pbuf_freecnt); 346 if (bp == NULL) 347 return tbp; 348 349 /* 350 * We are synthesizing a buffer out of vm_page_t's, but 351 * if the block size is not page aligned then the starting 352 * address may not be either. Inherit the b_data offset 353 * from the original buffer. 354 */ 355 bp->b_data = (char *)((vm_offset_t)bp->b_data | 356 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 357 bp->b_flags |= B_ASYNC | B_CLUSTER | B_VMIO; 358 bp->b_cmd = BUF_CMD_READ; 359 bp->b_bio1.bio_done = cluster_callback; 360 bp->b_bio1.bio_caller_info1.cluster_head = NULL; 361 bp->b_bio1.bio_caller_info2.cluster_tail = NULL; 362 bp->b_loffset = loffset; 363 bp->b_bio2.bio_offset = doffset; 364 KASSERT(bp->b_loffset != NOOFFSET, 365 ("cluster_rbuild: no buffer offset")); 366 367 bp->b_bcount = 0; 368 bp->b_bufsize = 0; 369 bp->b_xio.xio_npages = 0; 370 371 for (boffset = doffset, i = 0; i < run; ++i, boffset += blksize) { 372 if (i) { 373 if ((bp->b_xio.xio_npages * PAGE_SIZE) + 374 round_page(blksize) > maxiosize) { 375 break; 376 } 377 378 /* 379 * Shortcut some checks and try to avoid buffers that 380 * would block in the lock. The same checks have to 381 * be made again after we officially get the buffer. 382 */ 383 if ((tbp = findblk(vp, loffset + i * blksize)) != NULL) { 384 if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) 385 break; 386 BUF_UNLOCK(tbp); 387 388 for (j = 0; j < tbp->b_xio.xio_npages; j++) { 389 if (tbp->b_xio.xio_pages[j]->valid) 390 break; 391 } 392 393 if (j != tbp->b_xio.xio_npages) 394 break; 395 396 if (tbp->b_bcount != blksize) 397 break; 398 } 399 400 tbp = getblk(vp, loffset + i * blksize, blksize, 0, 0); 401 402 /* 403 * Stop scanning if the buffer is fuly valid 404 * (marked B_CACHE), or locked (may be doing a 405 * background write), or if the buffer is not 406 * VMIO backed. The clustering code can only deal 407 * with VMIO-backed buffers. 408 */ 409 if ((tbp->b_flags & (B_CACHE|B_LOCKED)) || 410 (tbp->b_flags & B_VMIO) == 0 || 411 (LIST_FIRST(&tbp->b_dep) != NULL && 412 buf_checkread(tbp)) 413 ) { 414 bqrelse(tbp); 415 break; 416 } 417 418 /* 419 * The buffer must be completely invalid in order to 420 * take part in the cluster. If it is partially valid 421 * then we stop. 422 */ 423 for (j = 0;j < tbp->b_xio.xio_npages; j++) { 424 if (tbp->b_xio.xio_pages[j]->valid) 425 break; 426 } 427 if (j != tbp->b_xio.xio_npages) { 428 bqrelse(tbp); 429 break; 430 } 431 432 /* 433 * Set a read-ahead mark as appropriate 434 */ 435 if (i == 1 || i == (run - 1)) 436 tbp->b_flags |= B_RAM; 437 438 /* 439 * Depress the priority of buffers not explicitly 440 * requested. 441 */ 442 /* tbp->b_flags |= B_AGE; */ 443 444 /* 445 * Set the block number if it isn't set, otherwise 446 * if it is make sure it matches the block number we 447 * expect. 448 */ 449 if (tbp->b_bio2.bio_offset == NOOFFSET) { 450 tbp->b_bio2.bio_offset = boffset; 451 } else if (tbp->b_bio2.bio_offset != boffset) { 452 brelse(tbp); 453 break; 454 } 455 } 456 /* 457 * The first buffer is setup async if doasync is specified. 458 * All other buffers in the cluster are setup async. This 459 * way the caller can decide how to deal with the requested 460 * buffer. 461 */ 462 if (i || doasync) 463 tbp->b_flags |= B_ASYNC; 464 tbp->b_cmd = BUF_CMD_READ; 465 BUF_KERNPROC(tbp); 466 cluster_append(&bp->b_bio1, tbp); 467 for (j = 0; j < tbp->b_xio.xio_npages; ++j) { 468 vm_page_t m; 469 m = tbp->b_xio.xio_pages[j]; 470 vm_page_io_start(m); 471 vm_object_pip_add(m->object, 1); 472 if ((bp->b_xio.xio_npages == 0) || 473 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) { 474 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 475 bp->b_xio.xio_npages++; 476 } 477 if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) 478 tbp->b_xio.xio_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 != blksize) 489 kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, blksize); 490 if (tbp->b_bufsize != blksize) 491 kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, blksize); 492 bp->b_bcount += blksize; 493 bp->b_bufsize += blksize; 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_xio.xio_npages; j++) { 501 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) == 502 VM_PAGE_BITS_ALL) { 503 bp->b_xio.xio_pages[j] = bogus_page; 504 } 505 } 506 if (bp->b_bufsize > bp->b_kvasize) { 507 panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)", 508 bp->b_bufsize, bp->b_kvasize); 509 } 510 511 pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 512 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_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 * 522 * The returned bio is &bp->b_bio1 523 */ 524 void 525 cluster_callback(struct bio *bio) 526 { 527 struct buf *bp = bio->bio_buf; 528 struct buf *tbp; 529 int error = 0; 530 531 /* 532 * Must propogate errors to all the components. A short read (EOF) 533 * is a critical error. 534 */ 535 if (bp->b_flags & B_ERROR) { 536 error = bp->b_error; 537 } else if (bp->b_bcount != bp->b_bufsize) { 538 panic("cluster_callback: unexpected EOF on cluster %p!", bio); 539 } 540 541 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages); 542 /* 543 * Move memory from the large cluster buffer into the component 544 * buffers and mark IO as done on these. Since the memory map 545 * is the same, no actual copying is required. 546 */ 547 while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) { 548 bio->bio_caller_info1.cluster_head = tbp->b_cluster_next; 549 if (error) { 550 tbp->b_flags |= B_ERROR; 551 tbp->b_error = error; 552 } else { 553 tbp->b_dirtyoff = tbp->b_dirtyend = 0; 554 tbp->b_flags &= ~(B_ERROR|B_INVAL); 555 /* 556 * XXX the bdwrite()/bqrelse() issued during 557 * cluster building clears B_RELBUF (see bqrelse() 558 * comment). If direct I/O was specified, we have 559 * to restore it here to allow the buffer and VM 560 * to be freed. 561 */ 562 if (tbp->b_flags & B_DIRECT) 563 tbp->b_flags |= B_RELBUF; 564 } 565 biodone(&tbp->b_bio1); 566 } 567 relpbuf(bp, &cluster_pbuf_freecnt); 568 } 569 570 /* 571 * cluster_wbuild_wb: 572 * 573 * Implement modified write build for cluster. 574 * 575 * write_behind = 0 write behind disabled 576 * write_behind = 1 write behind normal (default) 577 * write_behind = 2 write behind backed-off 578 */ 579 580 static __inline int 581 cluster_wbuild_wb(struct vnode *vp, int blksize, off_t start_loffset, int len) 582 { 583 int r = 0; 584 585 switch(write_behind) { 586 case 2: 587 if (start_loffset < len) 588 break; 589 start_loffset -= len; 590 /* fall through */ 591 case 1: 592 r = cluster_wbuild(vp, blksize, start_loffset, len); 593 /* fall through */ 594 default: 595 /* fall through */ 596 break; 597 } 598 return(r); 599 } 600 601 /* 602 * Do clustered write for FFS. 603 * 604 * Three cases: 605 * 1. Write is not sequential (write asynchronously) 606 * Write is sequential: 607 * 2. beginning of cluster - begin cluster 608 * 3. middle of a cluster - add to cluster 609 * 4. end of a cluster - asynchronously write cluster 610 */ 611 void 612 cluster_write(struct buf *bp, off_t filesize, int blksize, int seqcount) 613 { 614 struct vnode *vp; 615 off_t loffset; 616 int maxclen, cursize; 617 int async; 618 619 vp = bp->b_vp; 620 if (vp->v_type == VREG) 621 async = vp->v_mount->mnt_flag & MNT_ASYNC; 622 else 623 async = 0; 624 loffset = bp->b_loffset; 625 KASSERT(bp->b_loffset != NOOFFSET, 626 ("cluster_write: no buffer offset")); 627 628 /* Initialize vnode to beginning of file. */ 629 if (loffset == 0) 630 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; 631 632 if (vp->v_clen == 0 || loffset != vp->v_lastw + blksize || 633 bp->b_bio2.bio_offset == NOOFFSET || 634 (bp->b_bio2.bio_offset != vp->v_lasta + blksize)) { 635 maxclen = vmaxiosize(vp); 636 if (vp->v_clen != 0) { 637 /* 638 * Next block is not sequential. 639 * 640 * If we are not writing at end of file, the process 641 * seeked to another point in the file since its last 642 * write, or we have reached our maximum cluster size, 643 * then push the previous cluster. Otherwise try 644 * reallocating to make it sequential. 645 * 646 * Change to algorithm: only push previous cluster if 647 * it was sequential from the point of view of the 648 * seqcount heuristic, otherwise leave the buffer 649 * intact so we can potentially optimize the I/O 650 * later on in the buf_daemon or update daemon 651 * flush. 652 */ 653 cursize = vp->v_lastw - vp->v_cstart + blksize; 654 if (bp->b_loffset + blksize != filesize || 655 loffset != vp->v_lastw + blksize || vp->v_clen <= cursize) { 656 if (!async && seqcount > 0) { 657 cluster_wbuild_wb(vp, blksize, 658 vp->v_cstart, cursize); 659 } 660 } else { 661 struct buf **bpp, **endbp; 662 struct cluster_save *buflist; 663 664 buflist = cluster_collectbufs(vp, bp, blksize); 665 endbp = &buflist->bs_children 666 [buflist->bs_nchildren - 1]; 667 if (VOP_REALLOCBLKS(vp, buflist)) { 668 /* 669 * Failed, push the previous cluster 670 * if *really* writing sequentially 671 * in the logical file (seqcount > 1), 672 * otherwise delay it in the hopes that 673 * the low level disk driver can 674 * optimize the write ordering. 675 */ 676 for (bpp = buflist->bs_children; 677 bpp < endbp; bpp++) 678 brelse(*bpp); 679 kfree(buflist, M_SEGMENT); 680 if (seqcount > 1) { 681 cluster_wbuild_wb(vp, 682 blksize, vp->v_cstart, 683 cursize); 684 } 685 } else { 686 /* 687 * Succeeded, keep building cluster. 688 */ 689 for (bpp = buflist->bs_children; 690 bpp <= endbp; bpp++) 691 bdwrite(*bpp); 692 kfree(buflist, M_SEGMENT); 693 vp->v_lastw = loffset; 694 vp->v_lasta = bp->b_bio2.bio_offset; 695 return; 696 } 697 } 698 } 699 /* 700 * Consider beginning a cluster. If at end of file, make 701 * cluster as large as possible, otherwise find size of 702 * existing cluster. 703 */ 704 if ((vp->v_type == VREG) && 705 bp->b_loffset + blksize != filesize && 706 (bp->b_bio2.bio_offset == NOOFFSET) && 707 (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL, BUF_CMD_WRITE) || 708 bp->b_bio2.bio_offset == NOOFFSET)) { 709 bawrite(bp); 710 vp->v_clen = 0; 711 vp->v_lasta = bp->b_bio2.bio_offset; 712 vp->v_cstart = loffset + blksize; 713 vp->v_lastw = loffset; 714 return; 715 } 716 if (maxclen > blksize) 717 vp->v_clen = maxclen - blksize; 718 else 719 vp->v_clen = 0; 720 if (!async && vp->v_clen == 0) { /* I/O not contiguous */ 721 vp->v_cstart = loffset + blksize; 722 bawrite(bp); 723 } else { /* Wait for rest of cluster */ 724 vp->v_cstart = loffset; 725 bdwrite(bp); 726 } 727 } else if (loffset == 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, blksize, vp->v_cstart, 736 vp->v_clen + blksize); 737 vp->v_clen = 0; 738 vp->v_cstart = loffset + blksize; 739 } else if (vm_page_count_severe()) { 740 /* 741 * We are low on memory, get it going NOW 742 */ 743 bawrite(bp); 744 } else { 745 /* 746 * In the middle of a cluster, so just delay the I/O for now. 747 */ 748 bdwrite(bp); 749 } 750 vp->v_lastw = loffset; 751 vp->v_lasta = bp->b_bio2.bio_offset; 752 } 753 754 755 /* 756 * This is an awful lot like cluster_rbuild...wish they could be combined. 757 * The last lbn argument is the current block on which I/O is being 758 * performed. Check to see that it doesn't fall in the middle of 759 * the current block (if last_bp == NULL). 760 */ 761 int 762 cluster_wbuild(struct vnode *vp, int blksize, off_t start_loffset, int bytes) 763 { 764 struct buf *bp, *tbp; 765 int i, j; 766 int totalwritten = 0; 767 int maxiosize = vmaxiosize(vp); 768 769 while (bytes > 0) { 770 crit_enter(); 771 /* 772 * If the buffer is not delayed-write (i.e. dirty), or it 773 * is delayed-write but either locked or inval, it cannot 774 * partake in the clustered write. 775 */ 776 if (((tbp = findblk(vp, start_loffset)) == NULL) || 777 ((tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI) || 778 (LIST_FIRST(&tbp->b_dep) != NULL && buf_checkwrite(tbp)) || 779 BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) { 780 start_loffset += blksize; 781 bytes -= blksize; 782 crit_exit(); 783 continue; 784 } 785 bremfree(tbp); 786 KKASSERT(tbp->b_cmd == BUF_CMD_DONE); 787 crit_exit(); 788 789 /* 790 * Extra memory in the buffer, punt on this buffer. 791 * XXX we could handle this in most cases, but we would 792 * have to push the extra memory down to after our max 793 * possible cluster size and then potentially pull it back 794 * up if the cluster was terminated prematurely--too much 795 * hassle. 796 */ 797 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) || 798 (tbp->b_bcount != tbp->b_bufsize) || 799 (tbp->b_bcount != blksize) || 800 (bytes == blksize) || 801 ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) { 802 totalwritten += tbp->b_bufsize; 803 bawrite(tbp); 804 start_loffset += blksize; 805 bytes -= blksize; 806 continue; 807 } 808 809 /* 810 * Set up the pbuf. Track our append point with b_bcount 811 * and b_bufsize. b_bufsize is not used by the device but 812 * our caller uses it to loop clusters and we use it to 813 * detect a premature EOF on the block device. 814 */ 815 bp->b_bcount = 0; 816 bp->b_bufsize = 0; 817 bp->b_xio.xio_npages = 0; 818 bp->b_loffset = tbp->b_loffset; 819 bp->b_bio2.bio_offset = tbp->b_bio2.bio_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_ERROR; 830 bp->b_flags |= B_CLUSTER | B_BNOCLIP | 831 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT)); 832 bp->b_bio1.bio_done = cluster_callback; 833 bp->b_bio1.bio_caller_info1.cluster_head = NULL; 834 bp->b_bio1.bio_caller_info2.cluster_tail = NULL; 835 /* 836 * From this location in the file, scan forward to see 837 * if there are buffers with adjacent data that need to 838 * be written as well. 839 */ 840 for (i = 0; i < bytes; (i += blksize), (start_loffset += blksize)) { 841 if (i != 0) { /* If not the first buffer */ 842 crit_enter(); 843 /* 844 * If the adjacent data is not even in core it 845 * can't need to be written. 846 */ 847 if ((tbp = findblk(vp, start_loffset)) == NULL) { 848 crit_exit(); 849 break; 850 } 851 852 /* 853 * If it IS in core, but has different 854 * characteristics, or is locked (which 855 * means it could be undergoing a background 856 * I/O or be in a weird state), then don't 857 * cluster with it. 858 */ 859 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK | 860 B_INVAL | B_DELWRI | B_NEEDCOMMIT)) 861 != (B_DELWRI | B_CLUSTEROK | 862 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) || 863 (tbp->b_flags & B_LOCKED) || 864 (LIST_FIRST(&tbp->b_dep) != NULL && buf_checkwrite(tbp)) || 865 BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) { 866 crit_exit(); 867 break; 868 } 869 870 /* 871 * Check that the combined cluster 872 * would make sense with regard to pages 873 * and would not be too large 874 */ 875 if ((tbp->b_bcount != blksize) || 876 ((bp->b_bio2.bio_offset + i) != 877 tbp->b_bio2.bio_offset) || 878 ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) > 879 (maxiosize / PAGE_SIZE))) { 880 BUF_UNLOCK(tbp); 881 crit_exit(); 882 break; 883 } 884 /* 885 * Ok, it's passed all the tests, 886 * so remove it from the free list 887 * and mark it busy. We will use it. 888 */ 889 bremfree(tbp); 890 KKASSERT(tbp->b_cmd == BUF_CMD_DONE); 891 crit_exit(); 892 } /* end of code for non-first buffers only */ 893 894 /* 895 * If the IO is via the VM then we do some 896 * special VM hackery (yuck). Since the buffer's 897 * block size may not be page-aligned it is possible 898 * for a page to be shared between two buffers. We 899 * have to get rid of the duplication when building 900 * the cluster. 901 */ 902 if (tbp->b_flags & B_VMIO) { 903 vm_page_t m; 904 905 if (i != 0) { /* if not first buffer */ 906 for (j = 0; j < tbp->b_xio.xio_npages; ++j) { 907 m = tbp->b_xio.xio_pages[j]; 908 if (m->flags & PG_BUSY) { 909 bqrelse(tbp); 910 goto finishcluster; 911 } 912 } 913 } 914 915 for (j = 0; j < tbp->b_xio.xio_npages; ++j) { 916 m = tbp->b_xio.xio_pages[j]; 917 vm_page_io_start(m); 918 vm_object_pip_add(m->object, 1); 919 if ((bp->b_xio.xio_npages == 0) || 920 (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) { 921 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 922 bp->b_xio.xio_npages++; 923 } 924 } 925 } 926 bp->b_bcount += blksize; 927 bp->b_bufsize += blksize; 928 929 crit_enter(); 930 bundirty(tbp); 931 tbp->b_flags &= ~B_ERROR; 932 tbp->b_flags |= B_ASYNC; 933 tbp->b_cmd = BUF_CMD_WRITE; 934 crit_exit(); 935 BUF_KERNPROC(tbp); 936 cluster_append(&bp->b_bio1, tbp); 937 938 /* 939 * check for latent dependencies to be handled 940 */ 941 if (LIST_FIRST(&tbp->b_dep) != NULL) 942 buf_start(tbp); 943 } 944 finishcluster: 945 pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 946 (vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages); 947 if (bp->b_bufsize > bp->b_kvasize) { 948 panic( 949 "cluster_wbuild: b_bufsize(%d) > b_kvasize(%d)\n", 950 bp->b_bufsize, bp->b_kvasize); 951 } 952 totalwritten += bp->b_bufsize; 953 bp->b_dirtyoff = 0; 954 bp->b_dirtyend = bp->b_bufsize; 955 bp->b_flags |= B_ASYNC; 956 bp->b_cmd = BUF_CMD_WRITE; 957 vfs_busy_pages(vp, bp); 958 bp->b_runningbufspace = bp->b_bufsize; 959 if (bp->b_runningbufspace) { 960 runningbufspace += bp->b_runningbufspace; 961 ++runningbufcount; 962 } 963 BUF_KERNPROC(bp); /* B_ASYNC */ 964 vn_strategy(vp, &bp->b_bio1); 965 966 bytes -= i; 967 } 968 return totalwritten; 969 } 970 971 /* 972 * Collect together all the buffers in a cluster. 973 * Plus add one additional buffer. 974 */ 975 static struct cluster_save * 976 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int blksize) 977 { 978 struct cluster_save *buflist; 979 struct buf *bp; 980 off_t loffset; 981 int i, len; 982 983 len = (int)(vp->v_lastw - vp->v_cstart + blksize) / blksize; 984 buflist = kmalloc(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 (loffset = vp->v_cstart, i = 0; i < len; (loffset += blksize), i++) { 989 (void) bread(vp, loffset, last_bp->b_bcount, &bp); 990 buflist->bs_children[i] = bp; 991 if (bp->b_bio2.bio_offset == NOOFFSET) { 992 VOP_BMAP(bp->b_vp, bp->b_loffset, 993 &bp->b_bio2.bio_offset, 994 NULL, NULL, BUF_CMD_WRITE); 995 } 996 } 997 buflist->bs_children[i] = bp = last_bp; 998 if (bp->b_bio2.bio_offset == NOOFFSET) { 999 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset, 1000 NULL, NULL, BUF_CMD_WRITE); 1001 } 1002 buflist->bs_nchildren = i + 1; 1003 return (buflist); 1004 } 1005 1006 void 1007 cluster_append(struct bio *bio, struct buf *tbp) 1008 { 1009 tbp->b_cluster_next = NULL; 1010 if (bio->bio_caller_info1.cluster_head == NULL) { 1011 bio->bio_caller_info1.cluster_head = tbp; 1012 bio->bio_caller_info2.cluster_tail = tbp; 1013 } else { 1014 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp; 1015 bio->bio_caller_info2.cluster_tail = tbp; 1016 } 1017 } 1018 1019