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 * $FreeBSD: src/sys/kern/vfs_cluster.c,v 1.92.2.9 2001/11/18 07:10:59 dillon Exp $ 33 * $DragonFly: src/sys/kern/vfs_cluster.c,v 1.40 2008/07/14 03:09:00 dillon Exp $ 34 */ 35 36 #include "opt_debug_cluster.h" 37 38 #include <sys/param.h> 39 #include <sys/systm.h> 40 #include <sys/kernel.h> 41 #include <sys/proc.h> 42 #include <sys/buf.h> 43 #include <sys/vnode.h> 44 #include <sys/malloc.h> 45 #include <sys/mount.h> 46 #include <sys/resourcevar.h> 47 #include <sys/vmmeter.h> 48 #include <vm/vm.h> 49 #include <vm/vm_object.h> 50 #include <vm/vm_page.h> 51 #include <sys/sysctl.h> 52 53 #include <sys/buf2.h> 54 #include <vm/vm_page2.h> 55 56 #include <machine/limits.h> 57 58 #if defined(CLUSTERDEBUG) 59 #include <sys/sysctl.h> 60 static int rcluster= 0; 61 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, ""); 62 #endif 63 64 static MALLOC_DEFINE(M_SEGMENT, "cluster_save", "cluster_save buffer"); 65 66 static struct cluster_save * 67 cluster_collectbufs (struct vnode *vp, struct buf *last_bp, 68 int blksize); 69 static struct buf * 70 cluster_rbuild (struct vnode *vp, off_t filesize, off_t loffset, 71 off_t doffset, int blksize, int run, 72 struct buf *fbp); 73 static void cluster_callback (struct bio *); 74 static void cluster_setram (struct buf *); 75 static int cluster_wbuild(struct vnode *vp, struct buf **bpp, int blksize, 76 off_t start_loffset, int bytes); 77 78 static int write_behind = 1; 79 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, 80 "Cluster write-behind setting"); 81 static quad_t write_behind_minfilesize = 10 * 1024 * 1024; 82 SYSCTL_QUAD(_vfs, OID_AUTO, write_behind_minfilesize, CTLFLAG_RW, 83 &write_behind_minfilesize, 0, "Cluster write-behind setting"); 84 static int max_readahead = 2 * 1024 * 1024; 85 SYSCTL_INT(_vfs, OID_AUTO, max_readahead, CTLFLAG_RW, &max_readahead, 0, 86 "Limit in bytes for desired cluster read-ahead"); 87 88 extern vm_page_t bogus_page; 89 90 extern int cluster_pbuf_freecnt; 91 92 /* 93 * This replaces bread. 94 * 95 * filesize - read-ahead @ blksize will not cross this boundary 96 * loffset - loffset for returned *bpp 97 * blksize - blocksize for returned *bpp and read-ahead bps 98 * minreq - minimum (not a hard minimum) in bytes, typically reflects 99 * a higher level uio resid. 100 * maxreq - maximum (sequential heuristic) in bytes (highet typ ~2MB) 101 * bpp - return buffer (*bpp) for (loffset,blksize) 102 */ 103 int 104 cluster_readx(struct vnode *vp, off_t filesize, off_t loffset, 105 int blksize, size_t minreq, size_t maxreq, struct buf **bpp) 106 { 107 struct buf *bp, *rbp, *reqbp; 108 off_t origoffset; 109 off_t doffset; 110 int error; 111 int i; 112 int maxra; 113 int maxrbuild; 114 115 error = 0; 116 117 /* 118 * Calculate the desired read-ahead in blksize'd blocks (maxra). 119 * To do this we calculate maxreq. 120 * 121 * maxreq typically starts out as a sequential heuristic. If the 122 * high level uio/resid is bigger (minreq), we pop maxreq up to 123 * minreq. This represents the case where random I/O is being 124 * performed by the userland is issuing big read()'s. 125 * 126 * Then we limit maxreq to max_readahead to ensure it is a reasonable 127 * value. 128 * 129 * Finally we must ensure that (loffset + maxreq) does not cross the 130 * boundary (filesize) for the current blocksize. If we allowed it 131 * to cross we could end up with buffers past the boundary with the 132 * wrong block size (HAMMER large-data areas use mixed block sizes). 133 * minreq is also absolutely limited to filesize. 134 */ 135 if (maxreq < minreq) 136 maxreq = minreq; 137 /* minreq not used beyond this point */ 138 139 if (maxreq > max_readahead) { 140 maxreq = max_readahead; 141 if (maxreq > 16 * 1024 * 1024) 142 maxreq = 16 * 1024 * 1024; 143 } 144 if (maxreq < blksize) 145 maxreq = blksize; 146 if (loffset + maxreq > filesize) { 147 if (loffset > filesize) 148 maxreq = 0; 149 else 150 maxreq = filesize - loffset; 151 } 152 153 maxra = (int)(maxreq / blksize); 154 155 /* 156 * Get the requested block. 157 */ 158 if (*bpp) 159 reqbp = bp = *bpp; 160 else 161 *bpp = reqbp = bp = getblk(vp, loffset, blksize, 0, 0); 162 origoffset = loffset; 163 164 /* 165 * Calculate the maximum cluster size for a single I/O, used 166 * by cluster_rbuild(). 167 */ 168 maxrbuild = vmaxiosize(vp) / blksize; 169 170 /* 171 * if it is in the cache, then check to see if the reads have been 172 * sequential. If they have, then try some read-ahead, otherwise 173 * back-off on prospective read-aheads. 174 */ 175 if (bp->b_flags & B_CACHE) { 176 /* 177 * Not sequential, do not do any read-ahead 178 */ 179 if (maxra <= 1) 180 return 0; 181 182 /* 183 * No read-ahead mark, do not do any read-ahead 184 * yet. 185 */ 186 if ((bp->b_flags & B_RAM) == 0) 187 return 0; 188 189 /* 190 * We hit a read-ahead-mark, figure out how much read-ahead 191 * to do (maxra) and where to start (loffset). 192 * 193 * Shortcut the scan. Typically the way this works is that 194 * we've built up all the blocks inbetween except for the 195 * last in previous iterations, so if the second-to-last 196 * block is present we just skip ahead to it. 197 * 198 * This algorithm has O(1) cpu in the steady state no 199 * matter how large maxra is. 200 */ 201 bp->b_flags &= ~B_RAM; 202 203 if (findblk(vp, loffset + (maxra - 2) * blksize, FINDBLK_TEST)) 204 i = maxra - 1; 205 else 206 i = 1; 207 while (i < maxra) { 208 if (findblk(vp, loffset + i * blksize, 209 FINDBLK_TEST) == NULL) { 210 break; 211 } 212 ++i; 213 } 214 215 /* 216 * We got everything or everything is in the cache, no 217 * point continuing. 218 */ 219 if (i >= maxra) 220 return 0; 221 222 /* 223 * Calculate where to start the read-ahead and how much 224 * to do. Generally speaking we want to read-ahead by 225 * (maxra) when we've found a read-ahead mark. We do 226 * not want to reduce maxra here as it will cause 227 * successive read-ahead I/O's to be smaller and smaller. 228 * 229 * However, we have to make sure we don't break the 230 * filesize limitation for the clustered operation. 231 */ 232 loffset += i * blksize; 233 reqbp = bp = NULL; 234 235 if (loffset >= filesize) 236 return 0; 237 if (loffset + maxra * blksize > filesize) { 238 maxreq = filesize - loffset; 239 maxra = (int)(maxreq / blksize); 240 } 241 } else { 242 __debugvar off_t firstread = bp->b_loffset; 243 int nblks; 244 245 /* 246 * Set-up synchronous read for bp. 247 */ 248 bp->b_cmd = BUF_CMD_READ; 249 bp->b_bio1.bio_done = biodone_sync; 250 bp->b_bio1.bio_flags |= BIO_SYNC; 251 252 KASSERT(firstread != NOOFFSET, 253 ("cluster_read: no buffer offset")); 254 255 /* 256 * nblks is our cluster_rbuild request size, limited 257 * primarily by the device. 258 */ 259 if ((nblks = maxra) > maxrbuild) 260 nblks = maxrbuild; 261 262 if (nblks > 1) { 263 int burstbytes; 264 265 error = VOP_BMAP(vp, loffset, &doffset, 266 &burstbytes, NULL, BUF_CMD_READ); 267 if (error) 268 goto single_block_read; 269 if (nblks > burstbytes / blksize) 270 nblks = burstbytes / blksize; 271 if (doffset == NOOFFSET) 272 goto single_block_read; 273 if (nblks <= 1) 274 goto single_block_read; 275 276 bp = cluster_rbuild(vp, filesize, loffset, 277 doffset, blksize, nblks, bp); 278 loffset += bp->b_bufsize; 279 maxra -= bp->b_bufsize / blksize; 280 } else { 281 single_block_read: 282 /* 283 * If it isn't in the cache, then get a chunk from 284 * disk if sequential, otherwise just get the block. 285 */ 286 cluster_setram(bp); 287 loffset += blksize; 288 --maxra; 289 } 290 } 291 292 /* 293 * If B_CACHE was not set issue bp. bp will either be an 294 * asynchronous cluster buf or a synchronous single-buf. 295 * If it is a single buf it will be the same as reqbp. 296 * 297 * NOTE: Once an async cluster buf is issued bp becomes invalid. 298 */ 299 if (bp) { 300 #if defined(CLUSTERDEBUG) 301 if (rcluster) 302 kprintf("S(%012jx,%d,%d)\n", 303 (intmax_t)bp->b_loffset, bp->b_bcount, maxra); 304 #endif 305 if ((bp->b_flags & B_CLUSTER) == 0) 306 vfs_busy_pages(vp, bp); 307 bp->b_flags &= ~(B_ERROR|B_INVAL); 308 vn_strategy(vp, &bp->b_bio1); 309 error = 0; 310 /* bp invalid now */ 311 } 312 313 /* 314 * If we have been doing sequential I/O, then do some read-ahead. 315 * The code above us should have positioned us at the next likely 316 * offset. 317 * 318 * Only mess with buffers which we can immediately lock. HAMMER 319 * will do device-readahead irrespective of what the blocks 320 * represent. 321 */ 322 while (error == 0 && maxra > 0) { 323 int burstbytes; 324 int tmp_error; 325 int nblks; 326 327 rbp = getblk(vp, loffset, blksize, 328 GETBLK_SZMATCH|GETBLK_NOWAIT, 0); 329 if (rbp == NULL) 330 goto no_read_ahead; 331 if ((rbp->b_flags & B_CACHE)) { 332 bqrelse(rbp); 333 goto no_read_ahead; 334 } 335 336 /* 337 * An error from the read-ahead bmap has nothing to do 338 * with the caller's original request. 339 */ 340 tmp_error = VOP_BMAP(vp, loffset, &doffset, 341 &burstbytes, NULL, BUF_CMD_READ); 342 if (tmp_error || doffset == NOOFFSET) { 343 rbp->b_flags |= B_INVAL; 344 brelse(rbp); 345 rbp = NULL; 346 goto no_read_ahead; 347 } 348 if ((nblks = maxra) > maxrbuild) 349 nblks = maxrbuild; 350 if (nblks > burstbytes / blksize) 351 nblks = burstbytes / blksize; 352 353 /* 354 * rbp: async read 355 */ 356 rbp->b_cmd = BUF_CMD_READ; 357 /*rbp->b_flags |= B_AGE*/; 358 cluster_setram(rbp); 359 360 if (nblks > 1) { 361 rbp = cluster_rbuild(vp, filesize, loffset, 362 doffset, blksize, 363 nblks, rbp); 364 } else { 365 rbp->b_bio2.bio_offset = doffset; 366 } 367 368 #if defined(CLUSTERDEBUG) 369 if (rcluster) { 370 if (bp) { 371 kprintf("A+(%012jx,%d,%jd) " 372 "doff=%012jx minr=%zd ra=%d\n", 373 (intmax_t)loffset, rbp->b_bcount, 374 (intmax_t)(loffset - origoffset), 375 (intmax_t)doffset, minreq, maxra); 376 } else { 377 kprintf("A-(%012jx,%d,%jd) " 378 "doff=%012jx minr=%zd ra=%d\n", 379 (intmax_t)rbp->b_loffset, rbp->b_bcount, 380 (intmax_t)(loffset - origoffset), 381 (intmax_t)doffset, minreq, maxra); 382 } 383 } 384 #endif 385 rbp->b_flags &= ~(B_ERROR|B_INVAL); 386 387 if ((rbp->b_flags & B_CLUSTER) == 0) 388 vfs_busy_pages(vp, rbp); 389 BUF_KERNPROC(rbp); 390 loffset += rbp->b_bufsize; 391 maxra -= rbp->b_bufsize / blksize; 392 vn_strategy(vp, &rbp->b_bio1); 393 /* rbp invalid now */ 394 } 395 396 /* 397 * Wait for our original buffer to complete its I/O. reqbp will 398 * be NULL if the original buffer was B_CACHE. We are returning 399 * (*bpp) which is the same as reqbp when reqbp != NULL. 400 */ 401 no_read_ahead: 402 if (reqbp) { 403 KKASSERT(reqbp->b_bio1.bio_flags & BIO_SYNC); 404 error = biowait(&reqbp->b_bio1, "clurd"); 405 } 406 return (error); 407 } 408 409 /* 410 * If blocks are contiguous on disk, use this to provide clustered 411 * read ahead. We will read as many blocks as possible sequentially 412 * and then parcel them up into logical blocks in the buffer hash table. 413 * 414 * This function either returns a cluster buf or it returns fbp. fbp is 415 * already expected to be set up as a synchronous or asynchronous request. 416 * 417 * If a cluster buf is returned it will always be async. 418 */ 419 static struct buf * 420 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset, off_t doffset, 421 int blksize, int run, struct buf *fbp) 422 { 423 struct buf *bp, *tbp; 424 off_t boffset; 425 int i, j; 426 int maxiosize = vmaxiosize(vp); 427 428 /* 429 * avoid a division 430 */ 431 while (loffset + run * blksize > filesize) { 432 --run; 433 } 434 435 tbp = fbp; 436 tbp->b_bio2.bio_offset = doffset; 437 if((tbp->b_flags & B_MALLOC) || 438 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) { 439 return tbp; 440 } 441 442 bp = trypbuf_kva(&cluster_pbuf_freecnt); 443 if (bp == NULL) { 444 return tbp; 445 } 446 447 /* 448 * We are synthesizing a buffer out of vm_page_t's, but 449 * if the block size is not page aligned then the starting 450 * address may not be either. Inherit the b_data offset 451 * from the original buffer. 452 */ 453 bp->b_data = (char *)((vm_offset_t)bp->b_data | 454 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 455 bp->b_flags |= B_CLUSTER | B_VMIO; 456 bp->b_cmd = BUF_CMD_READ; 457 bp->b_bio1.bio_done = cluster_callback; /* default to async */ 458 bp->b_bio1.bio_caller_info1.cluster_head = NULL; 459 bp->b_bio1.bio_caller_info2.cluster_tail = NULL; 460 bp->b_loffset = loffset; 461 bp->b_bio2.bio_offset = doffset; 462 KASSERT(bp->b_loffset != NOOFFSET, 463 ("cluster_rbuild: no buffer offset")); 464 465 bp->b_bcount = 0; 466 bp->b_bufsize = 0; 467 bp->b_xio.xio_npages = 0; 468 469 for (boffset = doffset, i = 0; i < run; ++i, boffset += blksize) { 470 if (i) { 471 if ((bp->b_xio.xio_npages * PAGE_SIZE) + 472 round_page(blksize) > maxiosize) { 473 break; 474 } 475 476 /* 477 * Shortcut some checks and try to avoid buffers that 478 * would block in the lock. The same checks have to 479 * be made again after we officially get the buffer. 480 */ 481 tbp = getblk(vp, loffset + i * blksize, blksize, 482 GETBLK_SZMATCH|GETBLK_NOWAIT, 0); 483 if (tbp == NULL) 484 break; 485 for (j = 0; j < tbp->b_xio.xio_npages; j++) { 486 if (tbp->b_xio.xio_pages[j]->valid) 487 break; 488 } 489 if (j != tbp->b_xio.xio_npages) { 490 bqrelse(tbp); 491 break; 492 } 493 494 /* 495 * Stop scanning if the buffer is fuly valid 496 * (marked B_CACHE), or locked (may be doing a 497 * background write), or if the buffer is not 498 * VMIO backed. The clustering code can only deal 499 * with VMIO-backed buffers. 500 */ 501 if ((tbp->b_flags & (B_CACHE|B_LOCKED)) || 502 (tbp->b_flags & B_VMIO) == 0 || 503 (LIST_FIRST(&tbp->b_dep) != NULL && 504 buf_checkread(tbp)) 505 ) { 506 bqrelse(tbp); 507 break; 508 } 509 510 /* 511 * The buffer must be completely invalid in order to 512 * take part in the cluster. If it is partially valid 513 * then we stop. 514 */ 515 for (j = 0;j < tbp->b_xio.xio_npages; j++) { 516 if (tbp->b_xio.xio_pages[j]->valid) 517 break; 518 } 519 if (j != tbp->b_xio.xio_npages) { 520 bqrelse(tbp); 521 break; 522 } 523 524 /* 525 * Set a read-ahead mark as appropriate. Always 526 * set the read-ahead mark at (run - 1). It is 527 * unclear why we were also setting it at i == 1. 528 */ 529 if (/*i == 1 ||*/ i == (run - 1)) 530 cluster_setram(tbp); 531 532 /* 533 * Depress the priority of buffers not explicitly 534 * requested. 535 */ 536 /* tbp->b_flags |= B_AGE; */ 537 538 /* 539 * Set the block number if it isn't set, otherwise 540 * if it is make sure it matches the block number we 541 * expect. 542 */ 543 if (tbp->b_bio2.bio_offset == NOOFFSET) { 544 tbp->b_bio2.bio_offset = boffset; 545 } else if (tbp->b_bio2.bio_offset != boffset) { 546 brelse(tbp); 547 break; 548 } 549 } 550 551 /* 552 * The passed-in tbp (i == 0) will already be set up for 553 * async or sync operation. All other tbp's acquire in 554 * our loop are set up for async operation. 555 */ 556 tbp->b_cmd = BUF_CMD_READ; 557 BUF_KERNPROC(tbp); 558 cluster_append(&bp->b_bio1, tbp); 559 for (j = 0; j < tbp->b_xio.xio_npages; ++j) { 560 vm_page_t m; 561 562 m = tbp->b_xio.xio_pages[j]; 563 vm_page_busy_wait(m, FALSE, "clurpg"); 564 vm_page_io_start(m); 565 vm_page_wakeup(m); 566 vm_object_pip_add(m->object, 1); 567 if ((bp->b_xio.xio_npages == 0) || 568 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) { 569 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 570 bp->b_xio.xio_npages++; 571 } 572 if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) 573 tbp->b_xio.xio_pages[j] = bogus_page; 574 } 575 /* 576 * XXX shouldn't this be += size for both, like in 577 * cluster_wbuild()? 578 * 579 * Don't inherit tbp->b_bufsize as it may be larger due to 580 * a non-page-aligned size. Instead just aggregate using 581 * 'size'. 582 */ 583 if (tbp->b_bcount != blksize) 584 kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, blksize); 585 if (tbp->b_bufsize != blksize) 586 kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, blksize); 587 bp->b_bcount += blksize; 588 bp->b_bufsize += blksize; 589 } 590 591 /* 592 * Fully valid pages in the cluster are already good and do not need 593 * to be re-read from disk. Replace the page with bogus_page 594 */ 595 for (j = 0; j < bp->b_xio.xio_npages; j++) { 596 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) == 597 VM_PAGE_BITS_ALL) { 598 bp->b_xio.xio_pages[j] = bogus_page; 599 } 600 } 601 if (bp->b_bufsize > bp->b_kvasize) { 602 panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)", 603 bp->b_bufsize, bp->b_kvasize); 604 } 605 pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 606 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages); 607 BUF_KERNPROC(bp); 608 return (bp); 609 } 610 611 /* 612 * Cleanup after a clustered read or write. 613 * This is complicated by the fact that any of the buffers might have 614 * extra memory (if there were no empty buffer headers at allocbuf time) 615 * that we will need to shift around. 616 * 617 * The returned bio is &bp->b_bio1 618 */ 619 void 620 cluster_callback(struct bio *bio) 621 { 622 struct buf *bp = bio->bio_buf; 623 struct buf *tbp; 624 int error = 0; 625 626 /* 627 * Must propogate errors to all the components. A short read (EOF) 628 * is a critical error. 629 */ 630 if (bp->b_flags & B_ERROR) { 631 error = bp->b_error; 632 } else if (bp->b_bcount != bp->b_bufsize) { 633 panic("cluster_callback: unexpected EOF on cluster %p!", bio); 634 } 635 636 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages); 637 /* 638 * Move memory from the large cluster buffer into the component 639 * buffers and mark IO as done on these. Since the memory map 640 * is the same, no actual copying is required. 641 */ 642 while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) { 643 bio->bio_caller_info1.cluster_head = tbp->b_cluster_next; 644 if (error) { 645 tbp->b_flags |= B_ERROR | B_IODEBUG; 646 tbp->b_error = error; 647 } else { 648 tbp->b_dirtyoff = tbp->b_dirtyend = 0; 649 tbp->b_flags &= ~(B_ERROR|B_INVAL); 650 tbp->b_flags |= B_IODEBUG; 651 /* 652 * XXX the bdwrite()/bqrelse() issued during 653 * cluster building clears B_RELBUF (see bqrelse() 654 * comment). If direct I/O was specified, we have 655 * to restore it here to allow the buffer and VM 656 * to be freed. 657 */ 658 if (tbp->b_flags & B_DIRECT) 659 tbp->b_flags |= B_RELBUF; 660 } 661 biodone(&tbp->b_bio1); 662 } 663 relpbuf(bp, &cluster_pbuf_freecnt); 664 } 665 666 /* 667 * Implement modified write build for cluster. 668 * 669 * write_behind = 0 write behind disabled 670 * write_behind = 1 write behind normal (default) 671 * write_behind = 2 write behind backed-off 672 * 673 * In addition, write_behind is only activated for files that have 674 * grown past a certain size (default 10MB). Otherwise temporary files 675 * wind up generating a lot of unnecessary disk I/O. 676 */ 677 static __inline int 678 cluster_wbuild_wb(struct vnode *vp, int blksize, off_t start_loffset, int len) 679 { 680 int r = 0; 681 682 switch(write_behind) { 683 case 2: 684 if (start_loffset < len) 685 break; 686 start_loffset -= len; 687 /* fall through */ 688 case 1: 689 if (vp->v_filesize >= write_behind_minfilesize) { 690 r = cluster_wbuild(vp, NULL, blksize, 691 start_loffset, len); 692 } 693 /* fall through */ 694 default: 695 /* fall through */ 696 break; 697 } 698 return(r); 699 } 700 701 /* 702 * Do clustered write for FFS. 703 * 704 * Three cases: 705 * 1. Write is not sequential (write asynchronously) 706 * Write is sequential: 707 * 2. beginning of cluster - begin cluster 708 * 3. middle of a cluster - add to cluster 709 * 4. end of a cluster - asynchronously write cluster 710 */ 711 void 712 cluster_write(struct buf *bp, off_t filesize, int blksize, int seqcount) 713 { 714 struct vnode *vp; 715 off_t loffset; 716 int maxclen, cursize; 717 int async; 718 719 vp = bp->b_vp; 720 if (vp->v_type == VREG) 721 async = vp->v_mount->mnt_flag & MNT_ASYNC; 722 else 723 async = 0; 724 loffset = bp->b_loffset; 725 KASSERT(bp->b_loffset != NOOFFSET, 726 ("cluster_write: no buffer offset")); 727 728 /* Initialize vnode to beginning of file. */ 729 if (loffset == 0) 730 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; 731 732 if (vp->v_clen == 0 || loffset != vp->v_lastw + blksize || 733 bp->b_bio2.bio_offset == NOOFFSET || 734 (bp->b_bio2.bio_offset != vp->v_lasta + blksize)) { 735 maxclen = vmaxiosize(vp); 736 if (vp->v_clen != 0) { 737 /* 738 * Next block is not sequential. 739 * 740 * If we are not writing at end of file, the process 741 * seeked to another point in the file since its last 742 * write, or we have reached our maximum cluster size, 743 * then push the previous cluster. Otherwise try 744 * reallocating to make it sequential. 745 * 746 * Change to algorithm: only push previous cluster if 747 * it was sequential from the point of view of the 748 * seqcount heuristic, otherwise leave the buffer 749 * intact so we can potentially optimize the I/O 750 * later on in the buf_daemon or update daemon 751 * flush. 752 */ 753 cursize = vp->v_lastw - vp->v_cstart + blksize; 754 if (bp->b_loffset + blksize < filesize || 755 loffset != vp->v_lastw + blksize || vp->v_clen <= cursize) { 756 if (!async && seqcount > 0) { 757 cluster_wbuild_wb(vp, blksize, 758 vp->v_cstart, cursize); 759 } 760 } else { 761 struct buf **bpp, **endbp; 762 struct cluster_save *buflist; 763 764 buflist = cluster_collectbufs(vp, bp, blksize); 765 endbp = &buflist->bs_children 766 [buflist->bs_nchildren - 1]; 767 if (VOP_REALLOCBLKS(vp, buflist)) { 768 /* 769 * Failed, push the previous cluster 770 * if *really* writing sequentially 771 * in the logical file (seqcount > 1), 772 * otherwise delay it in the hopes that 773 * the low level disk driver can 774 * optimize the write ordering. 775 */ 776 for (bpp = buflist->bs_children; 777 bpp < endbp; bpp++) 778 brelse(*bpp); 779 kfree(buflist, M_SEGMENT); 780 if (seqcount > 1) { 781 cluster_wbuild_wb(vp, 782 blksize, vp->v_cstart, 783 cursize); 784 } 785 } else { 786 /* 787 * Succeeded, keep building cluster. 788 */ 789 for (bpp = buflist->bs_children; 790 bpp <= endbp; bpp++) 791 bdwrite(*bpp); 792 kfree(buflist, M_SEGMENT); 793 vp->v_lastw = loffset; 794 vp->v_lasta = bp->b_bio2.bio_offset; 795 return; 796 } 797 } 798 } 799 /* 800 * Consider beginning a cluster. If at end of file, make 801 * cluster as large as possible, otherwise find size of 802 * existing cluster. 803 */ 804 if ((vp->v_type == VREG) && 805 bp->b_loffset + blksize < filesize && 806 (bp->b_bio2.bio_offset == NOOFFSET) && 807 (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL, BUF_CMD_WRITE) || 808 bp->b_bio2.bio_offset == NOOFFSET)) { 809 bdwrite(bp); 810 vp->v_clen = 0; 811 vp->v_lasta = bp->b_bio2.bio_offset; 812 vp->v_cstart = loffset + blksize; 813 vp->v_lastw = loffset; 814 return; 815 } 816 if (maxclen > blksize) 817 vp->v_clen = maxclen - blksize; 818 else 819 vp->v_clen = 0; 820 if (!async && vp->v_clen == 0) { /* I/O not contiguous */ 821 vp->v_cstart = loffset + blksize; 822 bdwrite(bp); 823 } else { /* Wait for rest of cluster */ 824 vp->v_cstart = loffset; 825 bdwrite(bp); 826 } 827 } else if (loffset == vp->v_cstart + vp->v_clen) { 828 /* 829 * At end of cluster, write it out if seqcount tells us we 830 * are operating sequentially, otherwise let the buf or 831 * update daemon handle it. 832 */ 833 bdwrite(bp); 834 if (seqcount > 1) 835 cluster_wbuild_wb(vp, blksize, vp->v_cstart, 836 vp->v_clen + blksize); 837 vp->v_clen = 0; 838 vp->v_cstart = loffset + blksize; 839 } else if (vm_page_count_severe() && 840 bp->b_loffset + blksize < filesize) { 841 /* 842 * We are low on memory, get it going NOW. However, do not 843 * try to push out a partial block at the end of the file 844 * as this could lead to extremely non-optimal write activity. 845 */ 846 bawrite(bp); 847 } else { 848 /* 849 * In the middle of a cluster, so just delay the I/O for now. 850 */ 851 bdwrite(bp); 852 } 853 vp->v_lastw = loffset; 854 vp->v_lasta = bp->b_bio2.bio_offset; 855 } 856 857 /* 858 * This is the clustered version of bawrite(). It works similarly to 859 * cluster_write() except I/O on the buffer is guaranteed to occur. 860 */ 861 int 862 cluster_awrite(struct buf *bp) 863 { 864 int total; 865 866 /* 867 * Don't bother if it isn't clusterable. 868 */ 869 if ((bp->b_flags & B_CLUSTEROK) == 0 || 870 bp->b_vp == NULL || 871 (bp->b_vp->v_flag & VOBJBUF) == 0) { 872 total = bp->b_bufsize; 873 bawrite(bp); 874 return (total); 875 } 876 877 total = cluster_wbuild(bp->b_vp, &bp, bp->b_bufsize, 878 bp->b_loffset, vmaxiosize(bp->b_vp)); 879 if (bp) 880 bawrite(bp); 881 882 return total; 883 } 884 885 /* 886 * This is an awful lot like cluster_rbuild...wish they could be combined. 887 * The last lbn argument is the current block on which I/O is being 888 * performed. Check to see that it doesn't fall in the middle of 889 * the current block (if last_bp == NULL). 890 * 891 * cluster_wbuild() normally does not guarantee anything. If bpp is 892 * non-NULL and cluster_wbuild() is able to incorporate it into the 893 * I/O it will set *bpp to NULL, otherwise it will leave it alone and 894 * the caller must dispose of *bpp. 895 */ 896 static int 897 cluster_wbuild(struct vnode *vp, struct buf **bpp, 898 int blksize, off_t start_loffset, int bytes) 899 { 900 struct buf *bp, *tbp; 901 int i, j; 902 int totalwritten = 0; 903 int must_initiate; 904 int maxiosize = vmaxiosize(vp); 905 906 while (bytes > 0) { 907 /* 908 * If the buffer matches the passed locked & removed buffer 909 * we used the passed buffer (which might not be B_DELWRI). 910 * 911 * Otherwise locate the buffer and determine if it is 912 * compatible. 913 */ 914 if (bpp && (*bpp)->b_loffset == start_loffset) { 915 tbp = *bpp; 916 *bpp = NULL; 917 bpp = NULL; 918 } else { 919 tbp = findblk(vp, start_loffset, FINDBLK_NBLOCK); 920 if (tbp == NULL || 921 (tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != 922 B_DELWRI || 923 (LIST_FIRST(&tbp->b_dep) && buf_checkwrite(tbp))) { 924 if (tbp) 925 BUF_UNLOCK(tbp); 926 start_loffset += blksize; 927 bytes -= blksize; 928 continue; 929 } 930 bremfree(tbp); 931 } 932 KKASSERT(tbp->b_cmd == BUF_CMD_DONE); 933 934 /* 935 * Extra memory in the buffer, punt on this buffer. 936 * XXX we could handle this in most cases, but we would 937 * have to push the extra memory down to after our max 938 * possible cluster size and then potentially pull it back 939 * up if the cluster was terminated prematurely--too much 940 * hassle. 941 */ 942 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) || 943 (tbp->b_bcount != tbp->b_bufsize) || 944 (tbp->b_bcount != blksize) || 945 (bytes == blksize) || 946 ((bp = getpbuf_kva(&cluster_pbuf_freecnt)) == NULL)) { 947 totalwritten += tbp->b_bufsize; 948 bawrite(tbp); 949 start_loffset += blksize; 950 bytes -= blksize; 951 continue; 952 } 953 954 /* 955 * Set up the pbuf. Track our append point with b_bcount 956 * and b_bufsize. b_bufsize is not used by the device but 957 * our caller uses it to loop clusters and we use it to 958 * detect a premature EOF on the block device. 959 */ 960 bp->b_bcount = 0; 961 bp->b_bufsize = 0; 962 bp->b_xio.xio_npages = 0; 963 bp->b_loffset = tbp->b_loffset; 964 bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset; 965 966 /* 967 * We are synthesizing a buffer out of vm_page_t's, but 968 * if the block size is not page aligned then the starting 969 * address may not be either. Inherit the b_data offset 970 * from the original buffer. 971 */ 972 bp->b_data = (char *)((vm_offset_t)bp->b_data | 973 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 974 bp->b_flags &= ~B_ERROR; 975 bp->b_flags |= B_CLUSTER | B_BNOCLIP | 976 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT)); 977 bp->b_bio1.bio_caller_info1.cluster_head = NULL; 978 bp->b_bio1.bio_caller_info2.cluster_tail = NULL; 979 980 /* 981 * From this location in the file, scan forward to see 982 * if there are buffers with adjacent data that need to 983 * be written as well. 984 * 985 * IO *must* be initiated on index 0 at this point 986 * (particularly when called from cluster_awrite()). 987 */ 988 for (i = 0; i < bytes; (i += blksize), (start_loffset += blksize)) { 989 if (i == 0) { 990 must_initiate = 1; 991 } else { 992 /* 993 * Not first buffer. 994 */ 995 must_initiate = 0; 996 tbp = findblk(vp, start_loffset, 997 FINDBLK_NBLOCK); 998 /* 999 * Buffer not found or could not be locked 1000 * non-blocking. 1001 */ 1002 if (tbp == NULL) 1003 break; 1004 1005 /* 1006 * If it IS in core, but has different 1007 * characteristics, then don't cluster 1008 * with it. 1009 */ 1010 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK | 1011 B_INVAL | B_DELWRI | B_NEEDCOMMIT)) 1012 != (B_DELWRI | B_CLUSTEROK | 1013 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) || 1014 (tbp->b_flags & B_LOCKED) 1015 ) { 1016 BUF_UNLOCK(tbp); 1017 break; 1018 } 1019 1020 /* 1021 * Check that the combined cluster 1022 * would make sense with regard to pages 1023 * and would not be too large 1024 * 1025 * WARNING! buf_checkwrite() must be the last 1026 * check made. If it returns 0 then 1027 * we must initiate the I/O. 1028 */ 1029 if ((tbp->b_bcount != blksize) || 1030 ((bp->b_bio2.bio_offset + i) != 1031 tbp->b_bio2.bio_offset) || 1032 ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) > 1033 (maxiosize / PAGE_SIZE)) || 1034 (LIST_FIRST(&tbp->b_dep) && 1035 buf_checkwrite(tbp)) 1036 ) { 1037 BUF_UNLOCK(tbp); 1038 break; 1039 } 1040 if (LIST_FIRST(&tbp->b_dep)) 1041 must_initiate = 1; 1042 /* 1043 * Ok, it's passed all the tests, 1044 * so remove it from the free list 1045 * and mark it busy. We will use it. 1046 */ 1047 bremfree(tbp); 1048 KKASSERT(tbp->b_cmd == BUF_CMD_DONE); 1049 } 1050 1051 /* 1052 * If the IO is via the VM then we do some 1053 * special VM hackery (yuck). Since the buffer's 1054 * block size may not be page-aligned it is possible 1055 * for a page to be shared between two buffers. We 1056 * have to get rid of the duplication when building 1057 * the cluster. 1058 */ 1059 if (tbp->b_flags & B_VMIO) { 1060 vm_page_t m; 1061 1062 /* 1063 * Try to avoid deadlocks with the VM system. 1064 * However, we cannot abort the I/O if 1065 * must_initiate is non-zero. 1066 */ 1067 if (must_initiate == 0) { 1068 for (j = 0; 1069 j < tbp->b_xio.xio_npages; 1070 ++j) { 1071 m = tbp->b_xio.xio_pages[j]; 1072 if (m->flags & PG_BUSY) { 1073 bqrelse(tbp); 1074 goto finishcluster; 1075 } 1076 } 1077 } 1078 1079 for (j = 0; j < tbp->b_xio.xio_npages; ++j) { 1080 m = tbp->b_xio.xio_pages[j]; 1081 vm_page_busy_wait(m, FALSE, "clurpg"); 1082 vm_page_io_start(m); 1083 vm_page_wakeup(m); 1084 vm_object_pip_add(m->object, 1); 1085 if ((bp->b_xio.xio_npages == 0) || 1086 (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) { 1087 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 1088 bp->b_xio.xio_npages++; 1089 } 1090 } 1091 } 1092 bp->b_bcount += blksize; 1093 bp->b_bufsize += blksize; 1094 1095 bundirty(tbp); 1096 tbp->b_flags &= ~B_ERROR; 1097 tbp->b_cmd = BUF_CMD_WRITE; 1098 BUF_KERNPROC(tbp); 1099 cluster_append(&bp->b_bio1, tbp); 1100 1101 /* 1102 * check for latent dependencies to be handled 1103 */ 1104 if (LIST_FIRST(&tbp->b_dep) != NULL) 1105 buf_start(tbp); 1106 } 1107 finishcluster: 1108 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 1109 (vm_page_t *)bp->b_xio.xio_pages, 1110 bp->b_xio.xio_npages); 1111 if (bp->b_bufsize > bp->b_kvasize) { 1112 panic("cluster_wbuild: b_bufsize(%d) " 1113 "> b_kvasize(%d)\n", 1114 bp->b_bufsize, bp->b_kvasize); 1115 } 1116 totalwritten += bp->b_bufsize; 1117 bp->b_dirtyoff = 0; 1118 bp->b_dirtyend = bp->b_bufsize; 1119 bp->b_bio1.bio_done = cluster_callback; 1120 bp->b_cmd = BUF_CMD_WRITE; 1121 1122 vfs_busy_pages(vp, bp); 1123 bsetrunningbufspace(bp, bp->b_bufsize); 1124 BUF_KERNPROC(bp); 1125 vn_strategy(vp, &bp->b_bio1); 1126 1127 bytes -= i; 1128 } 1129 return totalwritten; 1130 } 1131 1132 /* 1133 * Collect together all the buffers in a cluster, plus add one 1134 * additional buffer passed-in. 1135 * 1136 * Only pre-existing buffers whos block size matches blksize are collected. 1137 * (this is primarily because HAMMER1 uses varying block sizes and we don't 1138 * want to override its choices). 1139 */ 1140 static struct cluster_save * 1141 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int blksize) 1142 { 1143 struct cluster_save *buflist; 1144 struct buf *bp; 1145 off_t loffset; 1146 int i, len; 1147 int j; 1148 int k; 1149 1150 len = (int)(vp->v_lastw - vp->v_cstart + blksize) / blksize; 1151 buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist), 1152 M_SEGMENT, M_WAITOK); 1153 buflist->bs_nchildren = 0; 1154 buflist->bs_children = (struct buf **) (buflist + 1); 1155 for (loffset = vp->v_cstart, i = 0, j = 0; 1156 i < len; 1157 (loffset += blksize), i++) { 1158 bp = getcacheblk(vp, loffset, 1159 last_bp->b_bcount, GETBLK_SZMATCH); 1160 buflist->bs_children[i] = bp; 1161 if (bp == NULL) { 1162 j = i + 1; 1163 } else if (bp->b_bio2.bio_offset == NOOFFSET) { 1164 VOP_BMAP(bp->b_vp, bp->b_loffset, 1165 &bp->b_bio2.bio_offset, 1166 NULL, NULL, BUF_CMD_WRITE); 1167 } 1168 } 1169 1170 /* 1171 * Get rid of gaps 1172 */ 1173 for (k = 0; k < j; ++k) { 1174 if (buflist->bs_children[k]) { 1175 bqrelse(buflist->bs_children[k]); 1176 buflist->bs_children[k] = NULL; 1177 } 1178 } 1179 if (j != 0) { 1180 if (j != i) { 1181 bcopy(buflist->bs_children + j, 1182 buflist->bs_children + 0, 1183 sizeof(buflist->bs_children[0]) * (i - j)); 1184 } 1185 i -= j; 1186 } 1187 buflist->bs_children[i] = bp = last_bp; 1188 if (bp->b_bio2.bio_offset == NOOFFSET) { 1189 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset, 1190 NULL, NULL, BUF_CMD_WRITE); 1191 } 1192 buflist->bs_nchildren = i + 1; 1193 return (buflist); 1194 } 1195 1196 void 1197 cluster_append(struct bio *bio, struct buf *tbp) 1198 { 1199 tbp->b_cluster_next = NULL; 1200 if (bio->bio_caller_info1.cluster_head == NULL) { 1201 bio->bio_caller_info1.cluster_head = tbp; 1202 bio->bio_caller_info2.cluster_tail = tbp; 1203 } else { 1204 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp; 1205 bio->bio_caller_info2.cluster_tail = tbp; 1206 } 1207 } 1208 1209 static 1210 void 1211 cluster_setram (struct buf *bp) 1212 { 1213 bp->b_flags |= B_RAM; 1214 if (bp->b_xio.xio_npages) 1215 vm_page_flag_set(bp->b_xio.xio_pages[0], PG_RAM); 1216 } 1217