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 * Copyright (c) 2012-2013 Matthew Dillon. All rights reserved. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 */ 32 33 #include "opt_debug_cluster.h" 34 35 #include <sys/param.h> 36 #include <sys/systm.h> 37 #include <sys/kernel.h> 38 #include <sys/proc.h> 39 #include <sys/buf.h> 40 #include <sys/vnode.h> 41 #include <sys/malloc.h> 42 #include <sys/mount.h> 43 #include <sys/resourcevar.h> 44 #include <sys/vmmeter.h> 45 #include <vm/vm.h> 46 #include <vm/vm_object.h> 47 #include <vm/vm_page.h> 48 #include <sys/sysctl.h> 49 50 #include <sys/buf2.h> 51 #include <vm/vm_page2.h> 52 53 #include <machine/limits.h> 54 55 /* 56 * Cluster tracking cache - replaces the original vnode v_* fields which had 57 * limited utility and were not MP safe. 58 * 59 * The cluster tracking cache is a simple 4-way set-associative non-chained 60 * cache. It is capable of tracking up to four zones separated by 1MB or 61 * more per vnode. 62 * 63 * NOTE: We want this structure to be cache-line friendly so the iterator 64 * is embedded rather than in a separate array. 65 * 66 * NOTE: A cluster cache entry can become stale when a vnode is recycled. 67 * For now we treat the values as heuristical but also self-consistent. 68 * i.e. the values cannot be completely random and cannot be SMP unsafe 69 * or the cluster code might end-up clustering non-contiguous buffers 70 * at the wrong offsets. 71 */ 72 struct cluster_cache { 73 struct vnode *vp; 74 u_int locked; 75 off_t v_lastw; /* last write (write cluster) */ 76 off_t v_cstart; /* start block of cluster */ 77 off_t v_lasta; /* last allocation */ 78 u_int v_clen; /* length of current cluster */ 79 u_int iterator; 80 } __cachealign; 81 82 typedef struct cluster_cache cluster_cache_t; 83 84 #define CLUSTER_CACHE_SIZE 512 85 #define CLUSTER_CACHE_MASK (CLUSTER_CACHE_SIZE - 1) 86 87 #define CLUSTER_ZONE ((off_t)(1024 * 1024)) 88 89 cluster_cache_t cluster_array[CLUSTER_CACHE_SIZE]; 90 91 #if defined(CLUSTERDEBUG) 92 #include <sys/sysctl.h> 93 static int rcluster= 0; 94 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, ""); 95 #endif 96 97 static MALLOC_DEFINE(M_SEGMENT, "cluster_save", "cluster_save buffer"); 98 99 static struct cluster_save * 100 cluster_collectbufs (cluster_cache_t *cc, struct vnode *vp, 101 struct buf *last_bp, int blksize); 102 static struct buf * 103 cluster_rbuild (struct vnode *vp, off_t filesize, off_t loffset, 104 off_t doffset, int blksize, int run, 105 struct buf *fbp, int *srp); 106 static void cluster_callback (struct bio *); 107 static void cluster_setram (struct buf *); 108 static void cluster_clrram (struct buf *); 109 static int cluster_wbuild(struct vnode *vp, struct buf **bpp, int blksize, 110 off_t start_loffset, int bytes); 111 112 static int write_behind = 1; 113 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, 114 "Cluster write-behind setting"); 115 static quad_t write_behind_minfilesize = 10 * 1024 * 1024; 116 SYSCTL_QUAD(_vfs, OID_AUTO, write_behind_minfilesize, CTLFLAG_RW, 117 &write_behind_minfilesize, 0, "Cluster write-behind setting"); 118 static int max_readahead = 2 * 1024 * 1024; 119 SYSCTL_INT(_vfs, OID_AUTO, max_readahead, CTLFLAG_RW, &max_readahead, 0, 120 "Limit in bytes for desired cluster read-ahead"); 121 122 extern vm_page_t bogus_page; 123 124 /* 125 * nblks is our cluster_rbuild request size. The approximate number of 126 * physical read-ahead requests is maxra / nblks. The physical request 127 * size is limited by the device (maxrbuild). We also do not want to make 128 * the request size too big or it will mess up the B_RAM streaming. 129 */ 130 static __inline 131 int 132 calc_rbuild_reqsize(int maxra, int maxrbuild) 133 { 134 int nblks; 135 136 if ((nblks = maxra / 4) > maxrbuild) 137 nblks = maxrbuild; 138 if (nblks < 1) 139 nblks = maxra; 140 return nblks; 141 } 142 143 /* 144 * Acquire/release cluster cache (can return dummy entry) 145 */ 146 static 147 cluster_cache_t * 148 cluster_getcache(cluster_cache_t *dummy, struct vnode *vp, off_t loffset) 149 { 150 cluster_cache_t *cc; 151 size_t hv; 152 int i; 153 int xact; 154 155 hv = (size_t)(intptr_t)vp ^ (size_t)(intptr_t)vp / sizeof(*vp); 156 hv &= CLUSTER_CACHE_MASK & ~3; 157 cc = &cluster_array[hv]; 158 159 xact = -1; 160 for (i = 0; i < 4; ++i) { 161 if (cc[i].vp != vp) 162 continue; 163 if (((cc[i].v_cstart ^ loffset) & ~(CLUSTER_ZONE - 1)) == 0) { 164 xact = i; 165 break; 166 } 167 } 168 if (xact >= 0 && atomic_swap_int(&cc[xact].locked, 1) == 0) { 169 if (cc[xact].vp == vp && 170 ((cc[i].v_cstart ^ loffset) & ~(CLUSTER_ZONE - 1)) == 0) { 171 return(&cc[xact]); 172 } 173 atomic_swap_int(&cc[xact].locked, 0); 174 } 175 176 /* 177 * New entry. If we can't acquire the cache line then use the 178 * passed-in dummy element and reset all fields. 179 * 180 * When we are able to acquire the cache line we only clear the 181 * fields if the vp does not match. This allows us to multi-zone 182 * a vp and for excessive zones / partial clusters to be retired. 183 */ 184 i = cc->iterator++ & 3; 185 cc += i; 186 if (atomic_swap_int(&cc->locked, 1) != 0) { 187 cc = dummy; 188 cc->locked = 1; 189 cc->vp = NULL; 190 } 191 if (cc->vp != vp) { 192 cc->vp = vp; 193 cc->v_lasta = 0; 194 cc->v_clen = 0; 195 cc->v_cstart = 0; 196 cc->v_lastw = 0; 197 } 198 return(cc); 199 } 200 201 static 202 void 203 cluster_putcache(cluster_cache_t *cc) 204 { 205 atomic_swap_int(&cc->locked, 0); 206 } 207 208 /* 209 * This replaces bread(), providing a synchronous read of the requested 210 * buffer plus asynchronous read-ahead within the specified bounds. 211 * 212 * The caller may pre-populate *bpp if it already has the requested buffer 213 * in-hand, else must set *bpp to NULL. Note that the cluster_read() inline 214 * sets *bpp to NULL and then calls cluster_readx() for compatibility. 215 * 216 * filesize - read-ahead @ blksize will not cross this boundary 217 * loffset - loffset for returned *bpp 218 * blksize - blocksize for returned *bpp and read-ahead bps 219 * minreq - minimum (not a hard minimum) in bytes, typically reflects 220 * a higher level uio resid. 221 * maxreq - maximum (sequential heuristic) in bytes (highet typ ~2MB) 222 * bpp - return buffer (*bpp) for (loffset,blksize) 223 */ 224 int 225 cluster_readx(struct vnode *vp, off_t filesize, off_t loffset, 226 int blksize, size_t minreq, size_t maxreq, struct buf **bpp) 227 { 228 struct buf *bp, *rbp, *reqbp; 229 off_t origoffset; 230 off_t doffset; 231 int error; 232 int i; 233 int maxra; 234 int maxrbuild; 235 int sr; 236 237 sr = 0; 238 239 /* 240 * Calculate the desired read-ahead in blksize'd blocks (maxra). 241 * To do this we calculate maxreq. 242 * 243 * maxreq typically starts out as a sequential heuristic. If the 244 * high level uio/resid is bigger (minreq), we pop maxreq up to 245 * minreq. This represents the case where random I/O is being 246 * performed by the userland is issuing big read()'s. 247 * 248 * Then we limit maxreq to max_readahead to ensure it is a reasonable 249 * value. 250 * 251 * Finally we must ensure that (loffset + maxreq) does not cross the 252 * boundary (filesize) for the current blocksize. If we allowed it 253 * to cross we could end up with buffers past the boundary with the 254 * wrong block size (HAMMER large-data areas use mixed block sizes). 255 * minreq is also absolutely limited to filesize. 256 */ 257 if (maxreq < minreq) 258 maxreq = minreq; 259 /* minreq not used beyond this point */ 260 261 if (maxreq > max_readahead) { 262 maxreq = max_readahead; 263 if (maxreq > 16 * 1024 * 1024) 264 maxreq = 16 * 1024 * 1024; 265 } 266 if (maxreq < blksize) 267 maxreq = blksize; 268 if (loffset + maxreq > filesize) { 269 if (loffset > filesize) 270 maxreq = 0; 271 else 272 maxreq = filesize - loffset; 273 } 274 275 maxra = (int)(maxreq / blksize); 276 277 /* 278 * Get the requested block. 279 */ 280 if (*bpp) 281 reqbp = bp = *bpp; 282 else 283 *bpp = reqbp = bp = getblk(vp, loffset, blksize, 0, 0); 284 origoffset = loffset; 285 286 /* 287 * Calculate the maximum cluster size for a single I/O, used 288 * by cluster_rbuild(). 289 */ 290 maxrbuild = vmaxiosize(vp) / blksize; 291 292 /* 293 * If it is in the cache, then check to see if the reads have been 294 * sequential. If they have, then try some read-ahead, otherwise 295 * back-off on prospective read-aheads. 296 */ 297 if (bp->b_flags & B_CACHE) { 298 /* 299 * Not sequential, do not do any read-ahead 300 */ 301 if (maxra <= 1) 302 return 0; 303 304 /* 305 * No read-ahead mark, do not do any read-ahead 306 * yet. 307 */ 308 if ((bp->b_flags & B_RAM) == 0) 309 return 0; 310 311 /* 312 * We hit a read-ahead-mark, figure out how much read-ahead 313 * to do (maxra) and where to start (loffset). 314 * 315 * Typically the way this works is that B_RAM is set in the 316 * middle of the cluster and triggers an overlapping 317 * read-ahead of 1/2 a cluster more blocks. This ensures 318 * that the cluster read-ahead scales with the read-ahead 319 * count and is thus better-able to absorb the caller's 320 * latency. 321 * 322 * Estimate where the next unread block will be by assuming 323 * that the B_RAM's are placed at the half-way point. 324 */ 325 bp->b_flags &= ~B_RAM; 326 327 i = maxra / 2; 328 rbp = findblk(vp, loffset + i * blksize, FINDBLK_TEST); 329 if (rbp == NULL || (rbp->b_flags & B_CACHE) == 0) { 330 while (i) { 331 --i; 332 rbp = findblk(vp, loffset + i * blksize, 333 FINDBLK_TEST); 334 if (rbp) { 335 ++i; 336 break; 337 } 338 } 339 } else { 340 while (i < maxra) { 341 rbp = findblk(vp, loffset + i * blksize, 342 FINDBLK_TEST); 343 if (rbp == NULL) 344 break; 345 ++i; 346 } 347 } 348 349 /* 350 * We got everything or everything is in the cache, no 351 * point continuing. 352 */ 353 if (i >= maxra) 354 return 0; 355 356 /* 357 * Calculate where to start the read-ahead and how much 358 * to do. Generally speaking we want to read-ahead by 359 * (maxra) when we've found a read-ahead mark. We do 360 * not want to reduce maxra here as it will cause 361 * successive read-ahead I/O's to be smaller and smaller. 362 * 363 * However, we have to make sure we don't break the 364 * filesize limitation for the clustered operation. 365 */ 366 loffset += i * blksize; 367 reqbp = bp = NULL; 368 369 if (loffset >= filesize) 370 return 0; 371 if (loffset + maxra * blksize > filesize) { 372 maxreq = filesize - loffset; 373 maxra = (int)(maxreq / blksize); 374 } 375 376 /* 377 * Set RAM on first read-ahead block since we still have 378 * approximate maxra/2 blocks ahead of us that are already 379 * cached or in-progress. 380 */ 381 sr = 1; 382 } else { 383 /* 384 * Start block is not valid, we will want to do a 385 * full read-ahead. 386 */ 387 __debugvar off_t firstread = bp->b_loffset; 388 int nblks; 389 390 /* 391 * Set-up synchronous read for bp. 392 */ 393 bp->b_cmd = BUF_CMD_READ; 394 bp->b_bio1.bio_done = biodone_sync; 395 bp->b_bio1.bio_flags |= BIO_SYNC; 396 397 KASSERT(firstread != NOOFFSET, 398 ("cluster_read: no buffer offset")); 399 400 nblks = calc_rbuild_reqsize(maxra, maxrbuild); 401 402 /* 403 * Set RAM half-way through the full-cluster. 404 */ 405 sr = (maxra + 1) / 2; 406 407 if (nblks > 1) { 408 int burstbytes; 409 410 error = VOP_BMAP(vp, loffset, &doffset, 411 &burstbytes, NULL, BUF_CMD_READ); 412 if (error) 413 goto single_block_read; 414 if (nblks > burstbytes / blksize) 415 nblks = burstbytes / blksize; 416 if (doffset == NOOFFSET) 417 goto single_block_read; 418 if (nblks <= 1) 419 goto single_block_read; 420 421 bp = cluster_rbuild(vp, filesize, loffset, 422 doffset, blksize, nblks, bp, &sr); 423 loffset += bp->b_bufsize; 424 maxra -= bp->b_bufsize / blksize; 425 } else { 426 single_block_read: 427 /* 428 * If it isn't in the cache, then get a chunk from 429 * disk if sequential, otherwise just get the block. 430 */ 431 loffset += blksize; 432 --maxra; 433 } 434 } 435 436 /* 437 * If B_CACHE was not set issue bp. bp will either be an 438 * asynchronous cluster buf or a synchronous single-buf. 439 * If it is a single buf it will be the same as reqbp. 440 * 441 * NOTE: Once an async cluster buf is issued bp becomes invalid. 442 */ 443 if (bp) { 444 #if defined(CLUSTERDEBUG) 445 if (rcluster) 446 kprintf("S(%012jx,%d,%d)\n", 447 (intmax_t)bp->b_loffset, bp->b_bcount, maxra); 448 #endif 449 if ((bp->b_flags & B_CLUSTER) == 0) 450 vfs_busy_pages(vp, bp); 451 bp->b_flags &= ~(B_ERROR|B_INVAL); 452 vn_strategy(vp, &bp->b_bio1); 453 /* bp invalid now */ 454 bp = NULL; 455 } 456 457 #if defined(CLUSTERDEBUG) 458 if (rcluster) 459 kprintf("cluster_rd %016jx/%d maxra=%d sr=%d\n", 460 loffset, blksize, maxra, sr); 461 #endif 462 463 /* 464 * If we have been doing sequential I/O, then do some read-ahead. 465 * The code above us should have positioned us at the next likely 466 * offset. 467 * 468 * Only mess with buffers which we can immediately lock. HAMMER 469 * will do device-readahead irrespective of what the blocks 470 * represent. 471 * 472 * Set B_RAM on the first buffer (the next likely offset needing 473 * read-ahead), under the assumption that there are still 474 * approximately maxra/2 blocks good ahead of us. 475 */ 476 while (maxra > 0) { 477 int burstbytes; 478 int nblks; 479 480 rbp = getblk(vp, loffset, blksize, 481 GETBLK_SZMATCH|GETBLK_NOWAIT, 0); 482 #if defined(CLUSTERDEBUG) 483 if (rcluster) { 484 kprintf("read-ahead %016jx rbp=%p ", 485 loffset, rbp); 486 } 487 #endif 488 if (rbp == NULL) 489 goto no_read_ahead; 490 if ((rbp->b_flags & B_CACHE)) { 491 bqrelse(rbp); 492 goto no_read_ahead; 493 } 494 495 /* 496 * If BMAP is not supported or has an issue, we still do 497 * (maxra) read-ahead, but we do not try to use rbuild. 498 */ 499 error = VOP_BMAP(vp, loffset, &doffset, 500 &burstbytes, NULL, BUF_CMD_READ); 501 if (error || doffset == NOOFFSET) { 502 nblks = 1; 503 doffset = NOOFFSET; 504 } else { 505 nblks = calc_rbuild_reqsize(maxra, maxrbuild); 506 if (nblks > burstbytes / blksize) 507 nblks = burstbytes / blksize; 508 } 509 rbp->b_cmd = BUF_CMD_READ; 510 511 if (nblks > 1) { 512 rbp = cluster_rbuild(vp, filesize, loffset, 513 doffset, blksize, 514 nblks, rbp, &sr); 515 } else { 516 rbp->b_bio2.bio_offset = doffset; 517 if (--sr == 0) 518 cluster_setram(rbp); 519 } 520 521 rbp->b_flags &= ~(B_ERROR|B_INVAL); 522 523 if ((rbp->b_flags & B_CLUSTER) == 0) 524 vfs_busy_pages(vp, rbp); 525 BUF_KERNPROC(rbp); 526 loffset += rbp->b_bufsize; 527 maxra -= rbp->b_bufsize / blksize; 528 vn_strategy(vp, &rbp->b_bio1); 529 /* rbp invalid now */ 530 } 531 532 /* 533 * Wait for our original buffer to complete its I/O. reqbp will 534 * be NULL if the original buffer was B_CACHE. We are returning 535 * (*bpp) which is the same as reqbp when reqbp != NULL. 536 */ 537 no_read_ahead: 538 if (reqbp) { 539 KKASSERT(reqbp->b_bio1.bio_flags & BIO_SYNC); 540 error = biowait(&reqbp->b_bio1, "clurd"); 541 } else { 542 error = 0; 543 } 544 return (error); 545 } 546 547 /* 548 * This replaces breadcb(), providing an asynchronous read of the requested 549 * buffer with a callback, plus an asynchronous read-ahead within the 550 * specified bounds. 551 * 552 * The callback must check whether BIO_DONE is set in the bio and issue 553 * the bpdone(bp, 0) if it isn't. The callback is responsible for clearing 554 * BIO_DONE and disposing of the I/O (bqrelse()ing it). 555 * 556 * filesize - read-ahead @ blksize will not cross this boundary 557 * loffset - loffset for returned *bpp 558 * blksize - blocksize for returned *bpp and read-ahead bps 559 * minreq - minimum (not a hard minimum) in bytes, typically reflects 560 * a higher level uio resid. 561 * maxreq - maximum (sequential heuristic) in bytes (highet typ ~2MB) 562 * bpp - return buffer (*bpp) for (loffset,blksize) 563 */ 564 void 565 cluster_readcb(struct vnode *vp, off_t filesize, off_t loffset, 566 int blksize, size_t minreq, size_t maxreq, 567 void (*func)(struct bio *), void *arg) 568 { 569 struct buf *bp, *rbp, *reqbp; 570 off_t origoffset; 571 off_t doffset; 572 int i; 573 int maxra; 574 int maxrbuild; 575 int sr; 576 577 sr = 0; 578 579 /* 580 * Calculate the desired read-ahead in blksize'd blocks (maxra). 581 * To do this we calculate maxreq. 582 * 583 * maxreq typically starts out as a sequential heuristic. If the 584 * high level uio/resid is bigger (minreq), we pop maxreq up to 585 * minreq. This represents the case where random I/O is being 586 * performed by the userland is issuing big read()'s. 587 * 588 * Then we limit maxreq to max_readahead to ensure it is a reasonable 589 * value. 590 * 591 * Finally we must ensure that (loffset + maxreq) does not cross the 592 * boundary (filesize) for the current blocksize. If we allowed it 593 * to cross we could end up with buffers past the boundary with the 594 * wrong block size (HAMMER large-data areas use mixed block sizes). 595 * minreq is also absolutely limited to filesize. 596 */ 597 if (maxreq < minreq) 598 maxreq = minreq; 599 /* minreq not used beyond this point */ 600 601 if (maxreq > max_readahead) { 602 maxreq = max_readahead; 603 if (maxreq > 16 * 1024 * 1024) 604 maxreq = 16 * 1024 * 1024; 605 } 606 if (maxreq < blksize) 607 maxreq = blksize; 608 if (loffset + maxreq > filesize) { 609 if (loffset > filesize) 610 maxreq = 0; 611 else 612 maxreq = filesize - loffset; 613 } 614 615 maxra = (int)(maxreq / blksize); 616 617 /* 618 * Get the requested block. 619 */ 620 reqbp = bp = getblk(vp, loffset, blksize, 0, 0); 621 origoffset = loffset; 622 623 /* 624 * Calculate the maximum cluster size for a single I/O, used 625 * by cluster_rbuild(). 626 */ 627 maxrbuild = vmaxiosize(vp) / blksize; 628 629 /* 630 * if it is in the cache, then check to see if the reads have been 631 * sequential. If they have, then try some read-ahead, otherwise 632 * back-off on prospective read-aheads. 633 */ 634 if (bp->b_flags & B_CACHE) { 635 /* 636 * Setup for func() call whether we do read-ahead or not. 637 */ 638 bp->b_bio1.bio_caller_info1.ptr = arg; 639 bp->b_bio1.bio_flags |= BIO_DONE; 640 641 /* 642 * Not sequential, do not do any read-ahead 643 */ 644 if (maxra <= 1) 645 goto no_read_ahead; 646 647 /* 648 * No read-ahead mark, do not do any read-ahead 649 * yet. 650 */ 651 if ((bp->b_flags & B_RAM) == 0) 652 goto no_read_ahead; 653 bp->b_flags &= ~B_RAM; 654 655 /* 656 * We hit a read-ahead-mark, figure out how much read-ahead 657 * to do (maxra) and where to start (loffset). 658 * 659 * Shortcut the scan. Typically the way this works is that 660 * we've built up all the blocks inbetween except for the 661 * last in previous iterations, so if the second-to-last 662 * block is present we just skip ahead to it. 663 * 664 * This algorithm has O(1) cpu in the steady state no 665 * matter how large maxra is. 666 */ 667 if (findblk(vp, loffset + (maxra - 2) * blksize, FINDBLK_TEST)) 668 i = maxra - 1; 669 else 670 i = 1; 671 while (i < maxra) { 672 if (findblk(vp, loffset + i * blksize, 673 FINDBLK_TEST) == NULL) { 674 break; 675 } 676 ++i; 677 } 678 679 /* 680 * We got everything or everything is in the cache, no 681 * point continuing. 682 */ 683 if (i >= maxra) 684 goto no_read_ahead; 685 686 /* 687 * Calculate where to start the read-ahead and how much 688 * to do. Generally speaking we want to read-ahead by 689 * (maxra) when we've found a read-ahead mark. We do 690 * not want to reduce maxra here as it will cause 691 * successive read-ahead I/O's to be smaller and smaller. 692 * 693 * However, we have to make sure we don't break the 694 * filesize limitation for the clustered operation. 695 */ 696 loffset += i * blksize; 697 bp = NULL; 698 /* leave reqbp intact to force function callback */ 699 700 if (loffset >= filesize) 701 goto no_read_ahead; 702 if (loffset + maxra * blksize > filesize) { 703 maxreq = filesize - loffset; 704 maxra = (int)(maxreq / blksize); 705 } 706 sr = 1; 707 } else { 708 /* 709 * bp is not valid, no prior cluster in progress so get a 710 * full cluster read-ahead going. 711 */ 712 __debugvar off_t firstread = bp->b_loffset; 713 int nblks; 714 int error; 715 716 /* 717 * Set-up synchronous read for bp. 718 */ 719 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL); 720 bp->b_cmd = BUF_CMD_READ; 721 bp->b_bio1.bio_done = func; 722 bp->b_bio1.bio_caller_info1.ptr = arg; 723 BUF_KERNPROC(bp); 724 reqbp = NULL; /* don't func() reqbp, it's running async */ 725 726 KASSERT(firstread != NOOFFSET, 727 ("cluster_read: no buffer offset")); 728 729 /* 730 * nblks is our cluster_rbuild request size, limited 731 * primarily by the device. 732 */ 733 nblks = calc_rbuild_reqsize(maxra, maxrbuild); 734 735 /* 736 * Set RAM half-way through the full-cluster. 737 */ 738 sr = (maxra + 1) / 2; 739 740 if (nblks > 1) { 741 int burstbytes; 742 743 error = VOP_BMAP(vp, loffset, &doffset, 744 &burstbytes, NULL, BUF_CMD_READ); 745 if (error) 746 goto single_block_read; 747 if (nblks > burstbytes / blksize) 748 nblks = burstbytes / blksize; 749 if (doffset == NOOFFSET) 750 goto single_block_read; 751 if (nblks <= 1) 752 goto single_block_read; 753 754 bp = cluster_rbuild(vp, filesize, loffset, 755 doffset, blksize, nblks, bp, &sr); 756 loffset += bp->b_bufsize; 757 maxra -= bp->b_bufsize / blksize; 758 } else { 759 single_block_read: 760 /* 761 * If it isn't in the cache, then get a chunk from 762 * disk if sequential, otherwise just get the block. 763 */ 764 loffset += blksize; 765 --maxra; 766 } 767 } 768 769 /* 770 * If bp != NULL then B_CACHE was *NOT* set and bp must be issued. 771 * bp will either be an asynchronous cluster buf or an asynchronous 772 * single-buf. 773 * 774 * NOTE: Once an async cluster buf is issued bp becomes invalid. 775 */ 776 if (bp) { 777 #if defined(CLUSTERDEBUG) 778 if (rcluster) 779 kprintf("S(%012jx,%d,%d)\n", 780 (intmax_t)bp->b_loffset, bp->b_bcount, maxra); 781 #endif 782 if ((bp->b_flags & B_CLUSTER) == 0) 783 vfs_busy_pages(vp, bp); 784 bp->b_flags &= ~(B_ERROR|B_INVAL); 785 vn_strategy(vp, &bp->b_bio1); 786 /* bp invalid now */ 787 bp = NULL; 788 } 789 790 #if defined(CLUSTERDEBUG) 791 if (rcluster) 792 kprintf("cluster_rd %016jx/%d maxra=%d sr=%d\n", 793 loffset, blksize, maxra, sr); 794 #endif 795 796 /* 797 * If we have been doing sequential I/O, then do some read-ahead. 798 * The code above us should have positioned us at the next likely 799 * offset. 800 * 801 * Only mess with buffers which we can immediately lock. HAMMER 802 * will do device-readahead irrespective of what the blocks 803 * represent. 804 */ 805 while (maxra > 0) { 806 int burstbytes; 807 int error; 808 int nblks; 809 810 rbp = getblk(vp, loffset, blksize, 811 GETBLK_SZMATCH|GETBLK_NOWAIT, 0); 812 if (rbp == NULL) 813 goto no_read_ahead; 814 if ((rbp->b_flags & B_CACHE)) { 815 bqrelse(rbp); 816 goto no_read_ahead; 817 } 818 819 /* 820 * If BMAP is not supported or has an issue, we still do 821 * (maxra) read-ahead, but we do not try to use rbuild. 822 */ 823 error = VOP_BMAP(vp, loffset, &doffset, 824 &burstbytes, NULL, BUF_CMD_READ); 825 if (error || doffset == NOOFFSET) { 826 nblks = 1; 827 doffset = NOOFFSET; 828 } else { 829 nblks = calc_rbuild_reqsize(maxra, maxrbuild); 830 if (nblks > burstbytes / blksize) 831 nblks = burstbytes / blksize; 832 } 833 rbp->b_cmd = BUF_CMD_READ; 834 835 if (nblks > 1) { 836 rbp = cluster_rbuild(vp, filesize, loffset, 837 doffset, blksize, 838 nblks, rbp, &sr); 839 } else { 840 rbp->b_bio2.bio_offset = doffset; 841 if (--sr == 0) 842 cluster_setram(rbp); 843 } 844 845 rbp->b_flags &= ~(B_ERROR|B_INVAL); 846 847 if ((rbp->b_flags & B_CLUSTER) == 0) 848 vfs_busy_pages(vp, rbp); 849 BUF_KERNPROC(rbp); 850 loffset += rbp->b_bufsize; 851 maxra -= rbp->b_bufsize / blksize; 852 vn_strategy(vp, &rbp->b_bio1); 853 /* rbp invalid now */ 854 } 855 856 /* 857 * If reqbp is non-NULL it had B_CACHE set and we issue the 858 * function callback synchronously. 859 * 860 * Note that we may start additional asynchronous I/O before doing 861 * the func() callback for the B_CACHE case 862 */ 863 no_read_ahead: 864 if (reqbp) 865 func(&reqbp->b_bio1); 866 } 867 868 /* 869 * If blocks are contiguous on disk, use this to provide clustered 870 * read ahead. We will read as many blocks as possible sequentially 871 * and then parcel them up into logical blocks in the buffer hash table. 872 * 873 * This function either returns a cluster buf or it returns fbp. fbp is 874 * already expected to be set up as a synchronous or asynchronous request. 875 * 876 * If a cluster buf is returned it will always be async. 877 * 878 * (*srp) counts down original blocks to determine where B_RAM should be set. 879 * Set B_RAM when *srp drops to 0. If (*srp) starts at 0, B_RAM will not be 880 * set on any buffer. Make sure B_RAM is cleared on any other buffers to 881 * prevent degenerate read-aheads from being generated. 882 */ 883 static struct buf * 884 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset, off_t doffset, 885 int blksize, int run, struct buf *fbp, int *srp) 886 { 887 struct buf *bp, *tbp; 888 off_t boffset; 889 int i, j; 890 int maxiosize = vmaxiosize(vp); 891 892 /* 893 * avoid a division 894 */ 895 while (loffset + run * blksize > filesize) { 896 --run; 897 } 898 899 tbp = fbp; 900 tbp->b_bio2.bio_offset = doffset; 901 if((tbp->b_flags & B_MALLOC) || 902 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) { 903 if (--*srp == 0) 904 cluster_setram(tbp); 905 else 906 cluster_clrram(tbp); 907 return tbp; 908 } 909 910 /* 911 * Get a pbuf, limit cluster I/O on a per-device basis. If 912 * doing cluster I/O for a file, limit cluster I/O on a 913 * per-mount basis. 914 */ 915 if (vp->v_type == VCHR || vp->v_type == VBLK) 916 bp = trypbuf_kva(&vp->v_pbuf_count); 917 else 918 bp = trypbuf_kva(&vp->v_mount->mnt_pbuf_count); 919 920 if (bp == NULL) 921 return tbp; 922 923 /* 924 * We are synthesizing a buffer out of vm_page_t's, but 925 * if the block size is not page aligned then the starting 926 * address may not be either. Inherit the b_data offset 927 * from the original buffer. 928 */ 929 bp->b_vp = vp; 930 bp->b_data = (char *)((vm_offset_t)bp->b_data | 931 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 932 bp->b_flags |= B_CLUSTER | B_VMIO; 933 bp->b_cmd = BUF_CMD_READ; 934 bp->b_bio1.bio_done = cluster_callback; /* default to async */ 935 bp->b_bio1.bio_caller_info1.cluster_head = NULL; 936 bp->b_bio1.bio_caller_info2.cluster_tail = NULL; 937 bp->b_loffset = loffset; 938 bp->b_bio2.bio_offset = doffset; 939 KASSERT(bp->b_loffset != NOOFFSET, 940 ("cluster_rbuild: no buffer offset")); 941 942 bp->b_bcount = 0; 943 bp->b_bufsize = 0; 944 bp->b_xio.xio_npages = 0; 945 946 for (boffset = doffset, i = 0; i < run; ++i, boffset += blksize) { 947 if (i) { 948 if ((bp->b_xio.xio_npages * PAGE_SIZE) + 949 round_page(blksize) > maxiosize) { 950 break; 951 } 952 953 /* 954 * Shortcut some checks and try to avoid buffers that 955 * would block in the lock. The same checks have to 956 * be made again after we officially get the buffer. 957 */ 958 tbp = getblk(vp, loffset + i * blksize, blksize, 959 GETBLK_SZMATCH|GETBLK_NOWAIT, 0); 960 if (tbp == NULL) 961 break; 962 for (j = 0; j < tbp->b_xio.xio_npages; j++) { 963 if (tbp->b_xio.xio_pages[j]->valid) 964 break; 965 } 966 if (j != tbp->b_xio.xio_npages) { 967 bqrelse(tbp); 968 break; 969 } 970 971 /* 972 * Stop scanning if the buffer is fuly valid 973 * (marked B_CACHE), or locked (may be doing a 974 * background write), or if the buffer is not 975 * VMIO backed. The clustering code can only deal 976 * with VMIO-backed buffers. 977 */ 978 if ((tbp->b_flags & (B_CACHE|B_LOCKED)) || 979 (tbp->b_flags & B_VMIO) == 0 || 980 (LIST_FIRST(&tbp->b_dep) != NULL && 981 buf_checkread(tbp)) 982 ) { 983 bqrelse(tbp); 984 break; 985 } 986 987 /* 988 * The buffer must be completely invalid in order to 989 * take part in the cluster. If it is partially valid 990 * then we stop. 991 */ 992 for (j = 0;j < tbp->b_xio.xio_npages; j++) { 993 if (tbp->b_xio.xio_pages[j]->valid) 994 break; 995 } 996 if (j != tbp->b_xio.xio_npages) { 997 bqrelse(tbp); 998 break; 999 } 1000 1001 /* 1002 * Depress the priority of buffers not explicitly 1003 * requested. 1004 */ 1005 /* tbp->b_flags |= B_AGE; */ 1006 1007 /* 1008 * Set the block number if it isn't set, otherwise 1009 * if it is make sure it matches the block number we 1010 * expect. 1011 */ 1012 if (tbp->b_bio2.bio_offset == NOOFFSET) { 1013 tbp->b_bio2.bio_offset = boffset; 1014 } else if (tbp->b_bio2.bio_offset != boffset) { 1015 brelse(tbp); 1016 break; 1017 } 1018 } 1019 1020 /* 1021 * Set B_RAM if (*srp) is 1. B_RAM is only set on one buffer 1022 * in the cluster, including potentially the first buffer 1023 * once we start streaming the read-aheads. 1024 */ 1025 if (--*srp == 0) 1026 cluster_setram(tbp); 1027 else 1028 cluster_clrram(tbp); 1029 1030 /* 1031 * The passed-in tbp (i == 0) will already be set up for 1032 * async or sync operation. All other tbp's acquire in 1033 * our loop are set up for async operation. 1034 */ 1035 tbp->b_cmd = BUF_CMD_READ; 1036 BUF_KERNPROC(tbp); 1037 cluster_append(&bp->b_bio1, tbp); 1038 for (j = 0; j < tbp->b_xio.xio_npages; ++j) { 1039 vm_page_t m; 1040 1041 m = tbp->b_xio.xio_pages[j]; 1042 vm_page_busy_wait(m, FALSE, "clurpg"); 1043 vm_page_io_start(m); 1044 vm_page_wakeup(m); 1045 vm_object_pip_add(m->object, 1); 1046 if ((bp->b_xio.xio_npages == 0) || 1047 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) { 1048 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 1049 bp->b_xio.xio_npages++; 1050 } 1051 if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) { 1052 tbp->b_xio.xio_pages[j] = bogus_page; 1053 tbp->b_flags |= B_HASBOGUS; 1054 } 1055 } 1056 /* 1057 * XXX shouldn't this be += size for both, like in 1058 * cluster_wbuild()? 1059 * 1060 * Don't inherit tbp->b_bufsize as it may be larger due to 1061 * a non-page-aligned size. Instead just aggregate using 1062 * 'size'. 1063 */ 1064 if (tbp->b_bcount != blksize) 1065 kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, blksize); 1066 if (tbp->b_bufsize != blksize) 1067 kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, blksize); 1068 bp->b_bcount += blksize; 1069 bp->b_bufsize += blksize; 1070 } 1071 1072 /* 1073 * Fully valid pages in the cluster are already good and do not need 1074 * to be re-read from disk. Replace the page with bogus_page 1075 */ 1076 for (j = 0; j < bp->b_xio.xio_npages; j++) { 1077 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) == 1078 VM_PAGE_BITS_ALL) { 1079 bp->b_xio.xio_pages[j] = bogus_page; 1080 bp->b_flags |= B_HASBOGUS; 1081 } 1082 } 1083 if (bp->b_bufsize > bp->b_kvasize) { 1084 panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)", 1085 bp->b_bufsize, bp->b_kvasize); 1086 } 1087 pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 1088 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages); 1089 BUF_KERNPROC(bp); 1090 return (bp); 1091 } 1092 1093 /* 1094 * Cleanup after a clustered read or write. 1095 * This is complicated by the fact that any of the buffers might have 1096 * extra memory (if there were no empty buffer headers at allocbuf time) 1097 * that we will need to shift around. 1098 * 1099 * The returned bio is &bp->b_bio1 1100 */ 1101 static void 1102 cluster_callback(struct bio *bio) 1103 { 1104 struct buf *bp = bio->bio_buf; 1105 struct buf *tbp; 1106 struct vnode *vp; 1107 int error = 0; 1108 1109 /* 1110 * Must propogate errors to all the components. A short read (EOF) 1111 * is a critical error. 1112 */ 1113 if (bp->b_flags & B_ERROR) { 1114 error = bp->b_error; 1115 } else if (bp->b_bcount != bp->b_bufsize) { 1116 panic("cluster_callback: unexpected EOF on cluster %p!", bio); 1117 } 1118 1119 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), 1120 bp->b_xio.xio_npages); 1121 /* 1122 * Move memory from the large cluster buffer into the component 1123 * buffers and mark IO as done on these. Since the memory map 1124 * is the same, no actual copying is required. 1125 */ 1126 while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) { 1127 bio->bio_caller_info1.cluster_head = tbp->b_cluster_next; 1128 if (error) { 1129 tbp->b_flags |= B_ERROR | B_IOISSUED; 1130 tbp->b_error = error; 1131 } else { 1132 tbp->b_dirtyoff = tbp->b_dirtyend = 0; 1133 tbp->b_flags &= ~(B_ERROR|B_INVAL); 1134 tbp->b_flags |= B_IOISSUED; 1135 /* 1136 * XXX the bdwrite()/bqrelse() issued during 1137 * cluster building clears B_RELBUF (see bqrelse() 1138 * comment). If direct I/O was specified, we have 1139 * to restore it here to allow the buffer and VM 1140 * to be freed. 1141 */ 1142 if (tbp->b_flags & B_DIRECT) 1143 tbp->b_flags |= B_RELBUF; 1144 1145 /* 1146 * XXX I think biodone() below will do this, but do 1147 * it here anyway for consistency. 1148 */ 1149 if (tbp->b_cmd == BUF_CMD_WRITE) 1150 bundirty(tbp); 1151 } 1152 biodone(&tbp->b_bio1); 1153 } 1154 vp = bp->b_vp; 1155 bp->b_vp = NULL; 1156 if (vp->v_type == VCHR || vp->v_type == VBLK) 1157 relpbuf(bp, &vp->v_pbuf_count); 1158 else 1159 relpbuf(bp, &vp->v_mount->mnt_pbuf_count); 1160 } 1161 1162 /* 1163 * Implement modified write build for cluster. 1164 * 1165 * write_behind = 0 write behind disabled 1166 * write_behind = 1 write behind normal (default) 1167 * write_behind = 2 write behind backed-off 1168 * 1169 * In addition, write_behind is only activated for files that have 1170 * grown past a certain size (default 10MB). Otherwise temporary files 1171 * wind up generating a lot of unnecessary disk I/O. 1172 */ 1173 static __inline int 1174 cluster_wbuild_wb(struct vnode *vp, int blksize, off_t start_loffset, int len) 1175 { 1176 int r = 0; 1177 1178 switch(write_behind) { 1179 case 2: 1180 if (start_loffset < len) 1181 break; 1182 start_loffset -= len; 1183 /* fall through */ 1184 case 1: 1185 if (vp->v_filesize >= write_behind_minfilesize) { 1186 r = cluster_wbuild(vp, NULL, blksize, 1187 start_loffset, len); 1188 } 1189 /* fall through */ 1190 default: 1191 /* fall through */ 1192 break; 1193 } 1194 return(r); 1195 } 1196 1197 /* 1198 * Do clustered write for FFS. 1199 * 1200 * Three cases: 1201 * 1. Write is not sequential (write asynchronously) 1202 * Write is sequential: 1203 * 2. beginning of cluster - begin cluster 1204 * 3. middle of a cluster - add to cluster 1205 * 4. end of a cluster - asynchronously write cluster 1206 * 1207 * WARNING! vnode fields are not locked and must ONLY be used heuristically. 1208 */ 1209 void 1210 cluster_write(struct buf *bp, off_t filesize, int blksize, int seqcount) 1211 { 1212 struct vnode *vp; 1213 off_t loffset; 1214 int maxclen, cursize; 1215 int async; 1216 cluster_cache_t dummy; 1217 cluster_cache_t *cc; 1218 1219 vp = bp->b_vp; 1220 if (vp->v_type == VREG) 1221 async = vp->v_mount->mnt_flag & MNT_ASYNC; 1222 else 1223 async = 0; 1224 loffset = bp->b_loffset; 1225 KASSERT(bp->b_loffset != NOOFFSET, 1226 ("cluster_write: no buffer offset")); 1227 1228 cc = cluster_getcache(&dummy, vp, loffset); 1229 1230 /* 1231 * Initialize vnode to beginning of file. 1232 */ 1233 if (loffset == 0) 1234 cc->v_lasta = cc->v_clen = cc->v_cstart = cc->v_lastw = 0; 1235 1236 if (cc->v_clen == 0 || loffset != cc->v_lastw + blksize || 1237 (bp->b_bio2.bio_offset != NOOFFSET && 1238 (bp->b_bio2.bio_offset != cc->v_lasta + blksize))) { 1239 /* 1240 * Next block is not logically sequential, or, if physical 1241 * block offsets are available, not physically sequential. 1242 * 1243 * If physical block offsets are not available we only 1244 * get here if we weren't logically sequential. 1245 */ 1246 maxclen = vmaxiosize(vp); 1247 if (cc->v_clen != 0) { 1248 /* 1249 * Next block is not sequential. 1250 * 1251 * If we are not writing at end of file, the process 1252 * seeked to another point in the file since its last 1253 * write, or we have reached our maximum cluster size, 1254 * then push the previous cluster. Otherwise try 1255 * reallocating to make it sequential. 1256 * 1257 * Change to algorithm: only push previous cluster if 1258 * it was sequential from the point of view of the 1259 * seqcount heuristic, otherwise leave the buffer 1260 * intact so we can potentially optimize the I/O 1261 * later on in the buf_daemon or update daemon 1262 * flush. 1263 */ 1264 cursize = cc->v_lastw - cc->v_cstart + blksize; 1265 if (bp->b_loffset + blksize < filesize || 1266 loffset != cc->v_lastw + blksize || 1267 cc->v_clen <= cursize) { 1268 if (!async && seqcount > 0) { 1269 cluster_wbuild_wb(vp, blksize, 1270 cc->v_cstart, cursize); 1271 } 1272 } else { 1273 struct buf **bpp, **endbp; 1274 struct cluster_save *buflist; 1275 1276 buflist = cluster_collectbufs(cc, vp, 1277 bp, blksize); 1278 endbp = &buflist->bs_children 1279 [buflist->bs_nchildren - 1]; 1280 if (VOP_REALLOCBLKS(vp, buflist)) { 1281 /* 1282 * Failed, push the previous cluster 1283 * if *really* writing sequentially 1284 * in the logical file (seqcount > 1), 1285 * otherwise delay it in the hopes that 1286 * the low level disk driver can 1287 * optimize the write ordering. 1288 * 1289 * NOTE: We do not brelse the last 1290 * element which is bp, and we 1291 * do not return here. 1292 */ 1293 for (bpp = buflist->bs_children; 1294 bpp < endbp; bpp++) 1295 brelse(*bpp); 1296 kfree(buflist, M_SEGMENT); 1297 if (seqcount > 1) { 1298 cluster_wbuild_wb(vp, 1299 blksize, cc->v_cstart, 1300 cursize); 1301 } 1302 } else { 1303 /* 1304 * Succeeded, keep building cluster. 1305 */ 1306 for (bpp = buflist->bs_children; 1307 bpp <= endbp; bpp++) 1308 bdwrite(*bpp); 1309 kfree(buflist, M_SEGMENT); 1310 cc->v_lastw = loffset; 1311 cc->v_lasta = bp->b_bio2.bio_offset; 1312 cluster_putcache(cc); 1313 return; 1314 } 1315 } 1316 } 1317 1318 /* 1319 * Consider beginning a cluster. If at end of file, make 1320 * cluster as large as possible, otherwise find size of 1321 * existing cluster. 1322 */ 1323 if ((vp->v_type == VREG) && 1324 bp->b_loffset + blksize < filesize && 1325 (bp->b_bio2.bio_offset == NOOFFSET) && 1326 (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL, BUF_CMD_WRITE) || 1327 bp->b_bio2.bio_offset == NOOFFSET)) { 1328 bdwrite(bp); 1329 cc->v_clen = 0; 1330 cc->v_lasta = bp->b_bio2.bio_offset; 1331 cc->v_cstart = loffset + blksize; 1332 cc->v_lastw = loffset; 1333 cluster_putcache(cc); 1334 return; 1335 } 1336 if (maxclen > blksize) 1337 cc->v_clen = maxclen - blksize; 1338 else 1339 cc->v_clen = 0; 1340 if (!async && cc->v_clen == 0) { /* I/O not contiguous */ 1341 cc->v_cstart = loffset + blksize; 1342 bdwrite(bp); 1343 } else { /* Wait for rest of cluster */ 1344 cc->v_cstart = loffset; 1345 bdwrite(bp); 1346 } 1347 } else if (loffset == cc->v_cstart + cc->v_clen) { 1348 /* 1349 * At end of cluster, write it out if seqcount tells us we 1350 * are operating sequentially, otherwise let the buf or 1351 * update daemon handle it. 1352 */ 1353 bdwrite(bp); 1354 if (seqcount > 1) 1355 cluster_wbuild_wb(vp, blksize, cc->v_cstart, 1356 cc->v_clen + blksize); 1357 cc->v_clen = 0; 1358 cc->v_cstart = loffset + blksize; 1359 } else if (vm_page_count_severe() && 1360 bp->b_loffset + blksize < filesize) { 1361 /* 1362 * We are low on memory, get it going NOW. However, do not 1363 * try to push out a partial block at the end of the file 1364 * as this could lead to extremely non-optimal write activity. 1365 */ 1366 bawrite(bp); 1367 } else { 1368 /* 1369 * In the middle of a cluster, so just delay the I/O for now. 1370 */ 1371 bdwrite(bp); 1372 } 1373 cc->v_lastw = loffset; 1374 cc->v_lasta = bp->b_bio2.bio_offset; 1375 cluster_putcache(cc); 1376 } 1377 1378 /* 1379 * This is the clustered version of bawrite(). It works similarly to 1380 * cluster_write() except I/O on the buffer is guaranteed to occur. 1381 */ 1382 int 1383 cluster_awrite(struct buf *bp) 1384 { 1385 int total; 1386 1387 /* 1388 * Don't bother if it isn't clusterable. 1389 */ 1390 if ((bp->b_flags & B_CLUSTEROK) == 0 || 1391 bp->b_vp == NULL || 1392 (bp->b_vp->v_flag & VOBJBUF) == 0) { 1393 total = bp->b_bufsize; 1394 bawrite(bp); 1395 return (total); 1396 } 1397 1398 total = cluster_wbuild(bp->b_vp, &bp, bp->b_bufsize, 1399 bp->b_loffset, vmaxiosize(bp->b_vp)); 1400 1401 /* 1402 * If bp is still non-NULL then cluster_wbuild() did not initiate 1403 * I/O on it and we must do so here to provide the API guarantee. 1404 */ 1405 if (bp) 1406 bawrite(bp); 1407 1408 return total; 1409 } 1410 1411 /* 1412 * This is an awful lot like cluster_rbuild...wish they could be combined. 1413 * The last lbn argument is the current block on which I/O is being 1414 * performed. Check to see that it doesn't fall in the middle of 1415 * the current block (if last_bp == NULL). 1416 * 1417 * cluster_wbuild() normally does not guarantee anything. If bpp is 1418 * non-NULL and cluster_wbuild() is able to incorporate it into the 1419 * I/O it will set *bpp to NULL, otherwise it will leave it alone and 1420 * the caller must dispose of *bpp. 1421 */ 1422 static int 1423 cluster_wbuild(struct vnode *vp, struct buf **bpp, 1424 int blksize, off_t start_loffset, int bytes) 1425 { 1426 struct buf *bp, *tbp; 1427 int i, j; 1428 int totalwritten = 0; 1429 int must_initiate; 1430 int maxiosize = vmaxiosize(vp); 1431 1432 while (bytes > 0) { 1433 /* 1434 * If the buffer matches the passed locked & removed buffer 1435 * we used the passed buffer (which might not be B_DELWRI). 1436 * 1437 * Otherwise locate the buffer and determine if it is 1438 * compatible. 1439 */ 1440 if (bpp && (*bpp)->b_loffset == start_loffset) { 1441 tbp = *bpp; 1442 *bpp = NULL; 1443 bpp = NULL; 1444 } else { 1445 tbp = findblk(vp, start_loffset, FINDBLK_NBLOCK); 1446 if (tbp == NULL || 1447 (tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != 1448 B_DELWRI || 1449 (LIST_FIRST(&tbp->b_dep) && buf_checkwrite(tbp))) { 1450 if (tbp) 1451 BUF_UNLOCK(tbp); 1452 start_loffset += blksize; 1453 bytes -= blksize; 1454 continue; 1455 } 1456 bremfree(tbp); 1457 } 1458 KKASSERT(tbp->b_cmd == BUF_CMD_DONE); 1459 1460 /* 1461 * Extra memory in the buffer, punt on this buffer. 1462 * XXX we could handle this in most cases, but we would 1463 * have to push the extra memory down to after our max 1464 * possible cluster size and then potentially pull it back 1465 * up if the cluster was terminated prematurely--too much 1466 * hassle. 1467 */ 1468 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) || 1469 (tbp->b_bcount != tbp->b_bufsize) || 1470 (tbp->b_bcount != blksize) || 1471 (bytes == blksize)) { 1472 totalwritten += tbp->b_bufsize; 1473 bawrite(tbp); 1474 start_loffset += blksize; 1475 bytes -= blksize; 1476 continue; 1477 } 1478 1479 /* 1480 * Get a pbuf, limit cluster I/O on a per-device basis. If 1481 * doing cluster I/O for a file, limit cluster I/O on a 1482 * per-mount basis. 1483 * 1484 * HAMMER and other filesystems may attempt to queue a massive 1485 * amount of write I/O, using trypbuf() here easily results in 1486 * situation where the I/O stream becomes non-clustered. 1487 */ 1488 if (vp->v_type == VCHR || vp->v_type == VBLK) 1489 bp = getpbuf_kva(&vp->v_pbuf_count); 1490 else 1491 bp = getpbuf_kva(&vp->v_mount->mnt_pbuf_count); 1492 1493 /* 1494 * Set up the pbuf. Track our append point with b_bcount 1495 * and b_bufsize. b_bufsize is not used by the device but 1496 * our caller uses it to loop clusters and we use it to 1497 * detect a premature EOF on the block device. 1498 */ 1499 bp->b_bcount = 0; 1500 bp->b_bufsize = 0; 1501 bp->b_xio.xio_npages = 0; 1502 bp->b_loffset = tbp->b_loffset; 1503 bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset; 1504 bp->b_vp = vp; 1505 1506 /* 1507 * We are synthesizing a buffer out of vm_page_t's, but 1508 * if the block size is not page aligned then the starting 1509 * address may not be either. Inherit the b_data offset 1510 * from the original buffer. 1511 */ 1512 bp->b_data = (char *)((vm_offset_t)bp->b_data | 1513 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 1514 bp->b_flags &= ~B_ERROR; 1515 bp->b_flags |= B_CLUSTER | B_BNOCLIP | 1516 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT)); 1517 bp->b_bio1.bio_caller_info1.cluster_head = NULL; 1518 bp->b_bio1.bio_caller_info2.cluster_tail = NULL; 1519 1520 /* 1521 * From this location in the file, scan forward to see 1522 * if there are buffers with adjacent data that need to 1523 * be written as well. 1524 * 1525 * IO *must* be initiated on index 0 at this point 1526 * (particularly when called from cluster_awrite()). 1527 */ 1528 for (i = 0; i < bytes; (i += blksize), (start_loffset += blksize)) { 1529 if (i == 0) { 1530 must_initiate = 1; 1531 } else { 1532 /* 1533 * Not first buffer. 1534 */ 1535 must_initiate = 0; 1536 tbp = findblk(vp, start_loffset, 1537 FINDBLK_NBLOCK); 1538 /* 1539 * Buffer not found or could not be locked 1540 * non-blocking. 1541 */ 1542 if (tbp == NULL) 1543 break; 1544 1545 /* 1546 * If it IS in core, but has different 1547 * characteristics, then don't cluster 1548 * with it. 1549 */ 1550 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK | 1551 B_INVAL | B_DELWRI | B_NEEDCOMMIT)) 1552 != (B_DELWRI | B_CLUSTEROK | 1553 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) || 1554 (tbp->b_flags & B_LOCKED) 1555 ) { 1556 BUF_UNLOCK(tbp); 1557 break; 1558 } 1559 1560 /* 1561 * Check that the combined cluster 1562 * would make sense with regard to pages 1563 * and would not be too large 1564 * 1565 * WARNING! buf_checkwrite() must be the last 1566 * check made. If it returns 0 then 1567 * we must initiate the I/O. 1568 */ 1569 if ((tbp->b_bcount != blksize) || 1570 ((bp->b_bio2.bio_offset + i) != 1571 tbp->b_bio2.bio_offset) || 1572 ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) > 1573 (maxiosize / PAGE_SIZE)) || 1574 (LIST_FIRST(&tbp->b_dep) && 1575 buf_checkwrite(tbp)) 1576 ) { 1577 BUF_UNLOCK(tbp); 1578 break; 1579 } 1580 if (LIST_FIRST(&tbp->b_dep)) 1581 must_initiate = 1; 1582 /* 1583 * Ok, it's passed all the tests, 1584 * so remove it from the free list 1585 * and mark it busy. We will use it. 1586 */ 1587 bremfree(tbp); 1588 KKASSERT(tbp->b_cmd == BUF_CMD_DONE); 1589 } 1590 1591 /* 1592 * If the IO is via the VM then we do some 1593 * special VM hackery (yuck). Since the buffer's 1594 * block size may not be page-aligned it is possible 1595 * for a page to be shared between two buffers. We 1596 * have to get rid of the duplication when building 1597 * the cluster. 1598 */ 1599 if (tbp->b_flags & B_VMIO) { 1600 vm_page_t m; 1601 1602 /* 1603 * Try to avoid deadlocks with the VM system. 1604 * However, we cannot abort the I/O if 1605 * must_initiate is non-zero. 1606 */ 1607 if (must_initiate == 0) { 1608 for (j = 0; 1609 j < tbp->b_xio.xio_npages; 1610 ++j) { 1611 m = tbp->b_xio.xio_pages[j]; 1612 if (m->flags & PG_BUSY) { 1613 bqrelse(tbp); 1614 goto finishcluster; 1615 } 1616 } 1617 } 1618 1619 for (j = 0; j < tbp->b_xio.xio_npages; ++j) { 1620 m = tbp->b_xio.xio_pages[j]; 1621 vm_page_busy_wait(m, FALSE, "clurpg"); 1622 vm_page_io_start(m); 1623 vm_page_wakeup(m); 1624 vm_object_pip_add(m->object, 1); 1625 if ((bp->b_xio.xio_npages == 0) || 1626 (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) { 1627 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 1628 bp->b_xio.xio_npages++; 1629 } 1630 } 1631 } 1632 bp->b_bcount += blksize; 1633 bp->b_bufsize += blksize; 1634 1635 /* 1636 * NOTE: see bwrite/bawrite code for why we no longer 1637 * undirty tbp here. 1638 * 1639 * bundirty(tbp); REMOVED 1640 */ 1641 tbp->b_flags &= ~B_ERROR; 1642 tbp->b_cmd = BUF_CMD_WRITE; 1643 BUF_KERNPROC(tbp); 1644 cluster_append(&bp->b_bio1, tbp); 1645 1646 /* 1647 * check for latent dependencies to be handled 1648 */ 1649 if (LIST_FIRST(&tbp->b_dep) != NULL) 1650 buf_start(tbp); 1651 } 1652 finishcluster: 1653 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 1654 (vm_page_t *)bp->b_xio.xio_pages, 1655 bp->b_xio.xio_npages); 1656 if (bp->b_bufsize > bp->b_kvasize) { 1657 panic("cluster_wbuild: b_bufsize(%d) " 1658 "> b_kvasize(%d)\n", 1659 bp->b_bufsize, bp->b_kvasize); 1660 } 1661 totalwritten += bp->b_bufsize; 1662 bp->b_dirtyoff = 0; 1663 bp->b_dirtyend = bp->b_bufsize; 1664 bp->b_bio1.bio_done = cluster_callback; 1665 bp->b_cmd = BUF_CMD_WRITE; 1666 1667 vfs_busy_pages(vp, bp); 1668 bsetrunningbufspace(bp, bp->b_bufsize); 1669 BUF_KERNPROC(bp); 1670 vn_strategy(vp, &bp->b_bio1); 1671 1672 bytes -= i; 1673 } 1674 return totalwritten; 1675 } 1676 1677 /* 1678 * Collect together all the buffers in a cluster, plus add one 1679 * additional buffer passed-in. 1680 * 1681 * Only pre-existing buffers whos block size matches blksize are collected. 1682 * (this is primarily because HAMMER1 uses varying block sizes and we don't 1683 * want to override its choices). 1684 * 1685 * This code will not try to collect buffers that it cannot lock, otherwise 1686 * it might deadlock against SMP-friendly filesystems. 1687 */ 1688 static struct cluster_save * 1689 cluster_collectbufs(cluster_cache_t *cc, struct vnode *vp, 1690 struct buf *last_bp, int blksize) 1691 { 1692 struct cluster_save *buflist; 1693 struct buf *bp; 1694 off_t loffset; 1695 int i, len; 1696 int j; 1697 int k; 1698 1699 len = (int)(cc->v_lastw - cc->v_cstart + blksize) / blksize; 1700 KKASSERT(len > 0); 1701 buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist), 1702 M_SEGMENT, M_WAITOK); 1703 buflist->bs_nchildren = 0; 1704 buflist->bs_children = (struct buf **) (buflist + 1); 1705 for (loffset = cc->v_cstart, i = 0, j = 0; 1706 i < len; 1707 (loffset += blksize), i++) { 1708 bp = getcacheblk(vp, loffset, 1709 last_bp->b_bcount, GETBLK_SZMATCH | 1710 GETBLK_NOWAIT); 1711 buflist->bs_children[i] = bp; 1712 if (bp == NULL) { 1713 j = i + 1; 1714 } else if (bp->b_bio2.bio_offset == NOOFFSET) { 1715 VOP_BMAP(bp->b_vp, bp->b_loffset, 1716 &bp->b_bio2.bio_offset, 1717 NULL, NULL, BUF_CMD_WRITE); 1718 } 1719 } 1720 1721 /* 1722 * Get rid of gaps 1723 */ 1724 for (k = 0; k < j; ++k) { 1725 if (buflist->bs_children[k]) { 1726 bqrelse(buflist->bs_children[k]); 1727 buflist->bs_children[k] = NULL; 1728 } 1729 } 1730 if (j != 0) { 1731 if (j != i) { 1732 bcopy(buflist->bs_children + j, 1733 buflist->bs_children + 0, 1734 sizeof(buflist->bs_children[0]) * (i - j)); 1735 } 1736 i -= j; 1737 } 1738 buflist->bs_children[i] = bp = last_bp; 1739 if (bp->b_bio2.bio_offset == NOOFFSET) { 1740 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset, 1741 NULL, NULL, BUF_CMD_WRITE); 1742 } 1743 buflist->bs_nchildren = i + 1; 1744 return (buflist); 1745 } 1746 1747 void 1748 cluster_append(struct bio *bio, struct buf *tbp) 1749 { 1750 tbp->b_cluster_next = NULL; 1751 if (bio->bio_caller_info1.cluster_head == NULL) { 1752 bio->bio_caller_info1.cluster_head = tbp; 1753 bio->bio_caller_info2.cluster_tail = tbp; 1754 } else { 1755 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp; 1756 bio->bio_caller_info2.cluster_tail = tbp; 1757 } 1758 } 1759 1760 static 1761 void 1762 cluster_setram(struct buf *bp) 1763 { 1764 bp->b_flags |= B_RAM; 1765 if (bp->b_xio.xio_npages) 1766 vm_page_flag_set(bp->b_xio.xio_pages[0], PG_RAM); 1767 } 1768 1769 static 1770 void 1771 cluster_clrram(struct buf *bp) 1772 { 1773 bp->b_flags &= ~B_RAM; 1774 if (bp->b_xio.xio_npages) 1775 vm_page_flag_clear(bp->b_xio.xio_pages[0], PG_RAM); 1776 } 1777