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 (end) (write cluster) */ 76 off_t v_cstart; /* start block (beg) of cluster */ 77 off_t v_lasta; /* last allocation (end) */ 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, int blksize, 226 int bflags, size_t minreq, size_t maxreq, 227 struct buf **bpp) 228 { 229 struct buf *bp, *rbp, *reqbp; 230 off_t origoffset; 231 off_t doffset; 232 int error; 233 int i; 234 int maxra; 235 int maxrbuild; 236 int sr; 237 238 sr = 0; 239 240 /* 241 * Calculate the desired read-ahead in blksize'd blocks (maxra). 242 * To do this we calculate maxreq. 243 * 244 * maxreq typically starts out as a sequential heuristic. If the 245 * high level uio/resid is bigger (minreq), we pop maxreq up to 246 * minreq. This represents the case where random I/O is being 247 * performed by the userland is issuing big read()'s. 248 * 249 * Then we limit maxreq to max_readahead to ensure it is a reasonable 250 * value. 251 * 252 * Finally we must ensure that (loffset + maxreq) does not cross the 253 * boundary (filesize) for the current blocksize. If we allowed it 254 * to cross we could end up with buffers past the boundary with the 255 * wrong block size (HAMMER large-data areas use mixed block sizes). 256 * minreq is also absolutely limited to filesize. 257 */ 258 if (maxreq < minreq) 259 maxreq = minreq; 260 /* minreq not used beyond this point */ 261 262 if (maxreq > max_readahead) { 263 maxreq = max_readahead; 264 if (maxreq > 16 * 1024 * 1024) 265 maxreq = 16 * 1024 * 1024; 266 } 267 if (maxreq < blksize) 268 maxreq = blksize; 269 if (loffset + maxreq > filesize) { 270 if (loffset > filesize) 271 maxreq = 0; 272 else 273 maxreq = filesize - loffset; 274 } 275 276 maxra = (int)(maxreq / blksize); 277 278 /* 279 * Get the requested block. 280 */ 281 if (*bpp) 282 reqbp = bp = *bpp; 283 else 284 *bpp = reqbp = bp = getblk(vp, loffset, blksize, 0, 0); 285 origoffset = loffset; 286 287 /* 288 * Calculate the maximum cluster size for a single I/O, used 289 * by cluster_rbuild(). 290 */ 291 maxrbuild = vmaxiosize(vp) / blksize; 292 293 /* 294 * If it is in the cache, then check to see if the reads have been 295 * sequential. If they have, then try some read-ahead, otherwise 296 * back-off on prospective read-aheads. 297 */ 298 if (bp->b_flags & B_CACHE) { 299 /* 300 * Not sequential, do not do any read-ahead 301 */ 302 if (maxra <= 1) 303 return 0; 304 305 /* 306 * No read-ahead mark, do not do any read-ahead 307 * yet. 308 */ 309 if ((bp->b_flags & B_RAM) == 0) 310 return 0; 311 312 /* 313 * We hit a read-ahead-mark, figure out how much read-ahead 314 * to do (maxra) and where to start (loffset). 315 * 316 * Typically the way this works is that B_RAM is set in the 317 * middle of the cluster and triggers an overlapping 318 * read-ahead of 1/2 a cluster more blocks. This ensures 319 * that the cluster read-ahead scales with the read-ahead 320 * count and is thus better-able to absorb the caller's 321 * latency. 322 * 323 * Estimate where the next unread block will be by assuming 324 * that the B_RAM's are placed at the half-way point. 325 */ 326 bp->b_flags &= ~B_RAM; 327 328 i = maxra / 2; 329 rbp = findblk(vp, loffset + i * blksize, FINDBLK_TEST); 330 if (rbp == NULL || (rbp->b_flags & B_CACHE) == 0) { 331 while (i) { 332 --i; 333 rbp = findblk(vp, loffset + i * blksize, 334 FINDBLK_TEST); 335 if (rbp) { 336 ++i; 337 break; 338 } 339 } 340 } else { 341 while (i < maxra) { 342 rbp = findblk(vp, loffset + i * blksize, 343 FINDBLK_TEST); 344 if (rbp == NULL) 345 break; 346 ++i; 347 } 348 } 349 350 /* 351 * We got everything or everything is in the cache, no 352 * point continuing. 353 */ 354 if (i >= maxra) 355 return 0; 356 357 /* 358 * Calculate where to start the read-ahead and how much 359 * to do. Generally speaking we want to read-ahead by 360 * (maxra) when we've found a read-ahead mark. We do 361 * not want to reduce maxra here as it will cause 362 * successive read-ahead I/O's to be smaller and smaller. 363 * 364 * However, we have to make sure we don't break the 365 * filesize limitation for the clustered operation. 366 */ 367 loffset += i * blksize; 368 reqbp = bp = NULL; 369 370 if (loffset >= filesize) 371 return 0; 372 if (loffset + maxra * blksize > filesize) { 373 maxreq = filesize - loffset; 374 maxra = (int)(maxreq / blksize); 375 } 376 377 /* 378 * Set RAM on first read-ahead block since we still have 379 * approximate maxra/2 blocks ahead of us that are already 380 * cached or in-progress. 381 */ 382 sr = 1; 383 } else { 384 /* 385 * Start block is not valid, we will want to do a 386 * full read-ahead. 387 */ 388 __debugvar off_t firstread = bp->b_loffset; 389 int nblks; 390 391 /* 392 * Set-up synchronous read for bp. 393 */ 394 bp->b_cmd = BUF_CMD_READ; 395 bp->b_bio1.bio_done = biodone_sync; 396 bp->b_bio1.bio_flags |= BIO_SYNC; 397 398 KASSERT(firstread != NOOFFSET, 399 ("cluster_read: no buffer offset")); 400 401 nblks = calc_rbuild_reqsize(maxra, maxrbuild); 402 403 /* 404 * Set RAM half-way through the full-cluster. 405 */ 406 sr = (maxra + 1) / 2; 407 408 if (nblks > 1) { 409 int burstbytes; 410 411 error = VOP_BMAP(vp, loffset, &doffset, 412 &burstbytes, NULL, BUF_CMD_READ); 413 if (error) 414 goto single_block_read; 415 if (nblks > burstbytes / blksize) 416 nblks = burstbytes / blksize; 417 if (doffset == NOOFFSET) 418 goto single_block_read; 419 if (nblks <= 1) 420 goto single_block_read; 421 422 bp = cluster_rbuild(vp, filesize, loffset, 423 doffset, blksize, nblks, bp, &sr); 424 loffset += bp->b_bufsize; 425 maxra -= bp->b_bufsize / blksize; 426 } else { 427 single_block_read: 428 /* 429 * If it isn't in the cache, then get a chunk from 430 * disk if sequential, otherwise just get the block. 431 */ 432 loffset += blksize; 433 --maxra; 434 } 435 } 436 437 /* 438 * If B_CACHE was not set issue bp. bp will either be an 439 * asynchronous cluster buf or a synchronous single-buf. 440 * If it is a single buf it will be the same as reqbp. 441 * 442 * NOTE: Once an async cluster buf is issued bp becomes invalid. 443 */ 444 if (bp) { 445 #if defined(CLUSTERDEBUG) 446 if (rcluster) 447 kprintf("S(%012jx,%d,%d)\n", 448 (intmax_t)bp->b_loffset, bp->b_bcount, maxra); 449 #endif 450 if ((bp->b_flags & B_CLUSTER) == 0) 451 vfs_busy_pages(vp, bp); 452 bp->b_flags &= ~(B_ERROR | B_INVAL | B_NOTMETA); 453 bp->b_flags |= bflags; 454 vn_strategy(vp, &bp->b_bio1); 455 /* bp invalid now */ 456 bp = NULL; 457 } 458 459 #if defined(CLUSTERDEBUG) 460 if (rcluster) 461 kprintf("cluster_rd %016jx/%d maxra=%d sr=%d\n", 462 loffset, blksize, maxra, sr); 463 #endif 464 465 /* 466 * If we have been doing sequential I/O, then do some read-ahead. 467 * The code above us should have positioned us at the next likely 468 * offset. 469 * 470 * Only mess with buffers which we can immediately lock. HAMMER 471 * will do device-readahead irrespective of what the blocks 472 * represent. 473 * 474 * Set B_RAM on the first buffer (the next likely offset needing 475 * read-ahead), under the assumption that there are still 476 * approximately maxra/2 blocks good ahead of us. 477 */ 478 while (maxra > 0) { 479 int burstbytes; 480 int nblks; 481 482 rbp = getblk(vp, loffset, blksize, 483 GETBLK_SZMATCH|GETBLK_NOWAIT, 0); 484 #if defined(CLUSTERDEBUG) 485 if (rcluster) { 486 kprintf("read-ahead %016jx rbp=%p ", 487 loffset, rbp); 488 } 489 #endif 490 if (rbp == NULL) 491 goto no_read_ahead; 492 if ((rbp->b_flags & B_CACHE)) { 493 bqrelse(rbp); 494 goto no_read_ahead; 495 } 496 497 /* 498 * If BMAP is not supported or has an issue, we still do 499 * (maxra) read-ahead, but we do not try to use rbuild. 500 */ 501 error = VOP_BMAP(vp, loffset, &doffset, 502 &burstbytes, NULL, BUF_CMD_READ); 503 if (error || doffset == NOOFFSET) { 504 nblks = 1; 505 doffset = NOOFFSET; 506 } else { 507 nblks = calc_rbuild_reqsize(maxra, maxrbuild); 508 if (nblks > burstbytes / blksize) 509 nblks = burstbytes / blksize; 510 } 511 rbp->b_cmd = BUF_CMD_READ; 512 513 if (nblks > 1) { 514 rbp = cluster_rbuild(vp, filesize, loffset, 515 doffset, blksize, 516 nblks, rbp, &sr); 517 } else { 518 rbp->b_bio2.bio_offset = doffset; 519 if (--sr == 0) 520 cluster_setram(rbp); 521 } 522 523 rbp->b_flags &= ~(B_ERROR | B_INVAL | B_NOTMETA); 524 rbp->b_flags |= bflags; 525 526 if ((rbp->b_flags & B_CLUSTER) == 0) 527 vfs_busy_pages(vp, rbp); 528 BUF_KERNPROC(rbp); 529 loffset += rbp->b_bufsize; 530 maxra -= rbp->b_bufsize / blksize; 531 vn_strategy(vp, &rbp->b_bio1); 532 /* rbp invalid now */ 533 } 534 535 /* 536 * Wait for our original buffer to complete its I/O. reqbp will 537 * be NULL if the original buffer was B_CACHE. We are returning 538 * (*bpp) which is the same as reqbp when reqbp != NULL. 539 */ 540 no_read_ahead: 541 if (reqbp) { 542 KKASSERT(reqbp->b_bio1.bio_flags & BIO_SYNC); 543 error = biowait(&reqbp->b_bio1, "clurd"); 544 } else { 545 error = 0; 546 } 547 return (error); 548 } 549 550 /* 551 * This replaces breadcb(), providing an asynchronous read of the requested 552 * buffer with a callback, plus an asynchronous read-ahead within the 553 * specified bounds. 554 * 555 * The callback must check whether BIO_DONE is set in the bio and issue 556 * the bpdone(bp, 0) if it isn't. The callback is responsible for clearing 557 * BIO_DONE and disposing of the I/O (bqrelse()ing it). 558 * 559 * filesize - read-ahead @ blksize will not cross this boundary 560 * loffset - loffset for returned *bpp 561 * blksize - blocksize for returned *bpp and read-ahead bps 562 * minreq - minimum (not a hard minimum) in bytes, typically reflects 563 * a higher level uio resid. 564 * maxreq - maximum (sequential heuristic) in bytes (highet typ ~2MB) 565 * bpp - return buffer (*bpp) for (loffset,blksize) 566 */ 567 void 568 cluster_readcb(struct vnode *vp, off_t filesize, off_t loffset, int blksize, 569 int bflags, size_t minreq, size_t maxreq, 570 void (*func)(struct bio *), void *arg) 571 { 572 struct buf *bp, *rbp, *reqbp; 573 off_t origoffset; 574 off_t doffset; 575 int i; 576 int maxra; 577 int maxrbuild; 578 int sr; 579 580 sr = 0; 581 582 /* 583 * Calculate the desired read-ahead in blksize'd blocks (maxra). 584 * To do this we calculate maxreq. 585 * 586 * maxreq typically starts out as a sequential heuristic. If the 587 * high level uio/resid is bigger (minreq), we pop maxreq up to 588 * minreq. This represents the case where random I/O is being 589 * performed by the userland is issuing big read()'s. 590 * 591 * Then we limit maxreq to max_readahead to ensure it is a reasonable 592 * value. 593 * 594 * Finally we must ensure that (loffset + maxreq) does not cross the 595 * boundary (filesize) for the current blocksize. If we allowed it 596 * to cross we could end up with buffers past the boundary with the 597 * wrong block size (HAMMER large-data areas use mixed block sizes). 598 * minreq is also absolutely limited to filesize. 599 */ 600 if (maxreq < minreq) 601 maxreq = minreq; 602 /* minreq not used beyond this point */ 603 604 if (maxreq > max_readahead) { 605 maxreq = max_readahead; 606 if (maxreq > 16 * 1024 * 1024) 607 maxreq = 16 * 1024 * 1024; 608 } 609 if (maxreq < blksize) 610 maxreq = blksize; 611 if (loffset + maxreq > filesize) { 612 if (loffset > filesize) 613 maxreq = 0; 614 else 615 maxreq = filesize - loffset; 616 } 617 618 maxra = (int)(maxreq / blksize); 619 620 /* 621 * Get the requested block. 622 */ 623 reqbp = bp = getblk(vp, loffset, blksize, 0, 0); 624 origoffset = loffset; 625 626 /* 627 * Calculate the maximum cluster size for a single I/O, used 628 * by cluster_rbuild(). 629 */ 630 maxrbuild = vmaxiosize(vp) / blksize; 631 632 /* 633 * if it is in the cache, then check to see if the reads have been 634 * sequential. If they have, then try some read-ahead, otherwise 635 * back-off on prospective read-aheads. 636 */ 637 if (bp->b_flags & B_CACHE) { 638 /* 639 * Setup for func() call whether we do read-ahead or not. 640 */ 641 bp->b_bio1.bio_caller_info1.ptr = arg; 642 bp->b_bio1.bio_flags |= BIO_DONE; 643 644 /* 645 * Not sequential, do not do any read-ahead 646 */ 647 if (maxra <= 1) 648 goto no_read_ahead; 649 650 /* 651 * No read-ahead mark, do not do any read-ahead 652 * yet. 653 */ 654 if ((bp->b_flags & B_RAM) == 0) 655 goto no_read_ahead; 656 bp->b_flags &= ~B_RAM; 657 658 /* 659 * We hit a read-ahead-mark, figure out how much read-ahead 660 * to do (maxra) and where to start (loffset). 661 * 662 * Shortcut the scan. Typically the way this works is that 663 * we've built up all the blocks inbetween except for the 664 * last in previous iterations, so if the second-to-last 665 * block is present we just skip ahead to it. 666 * 667 * This algorithm has O(1) cpu in the steady state no 668 * matter how large maxra is. 669 */ 670 if (findblk(vp, loffset + (maxra - 2) * blksize, FINDBLK_TEST)) 671 i = maxra - 1; 672 else 673 i = 1; 674 while (i < maxra) { 675 if (findblk(vp, loffset + i * blksize, 676 FINDBLK_TEST) == NULL) { 677 break; 678 } 679 ++i; 680 } 681 682 /* 683 * We got everything or everything is in the cache, no 684 * point continuing. 685 */ 686 if (i >= maxra) 687 goto no_read_ahead; 688 689 /* 690 * Calculate where to start the read-ahead and how much 691 * to do. Generally speaking we want to read-ahead by 692 * (maxra) when we've found a read-ahead mark. We do 693 * not want to reduce maxra here as it will cause 694 * successive read-ahead I/O's to be smaller and smaller. 695 * 696 * However, we have to make sure we don't break the 697 * filesize limitation for the clustered operation. 698 */ 699 loffset += i * blksize; 700 bp = NULL; 701 /* leave reqbp intact to force function callback */ 702 703 if (loffset >= filesize) 704 goto no_read_ahead; 705 if (loffset + maxra * blksize > filesize) { 706 maxreq = filesize - loffset; 707 maxra = (int)(maxreq / blksize); 708 } 709 sr = 1; 710 } else { 711 /* 712 * bp is not valid, no prior cluster in progress so get a 713 * full cluster read-ahead going. 714 */ 715 __debugvar off_t firstread = bp->b_loffset; 716 int nblks; 717 int error; 718 719 /* 720 * Set-up synchronous read for bp. 721 */ 722 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL | B_NOTMETA); 723 bp->b_flags |= bflags; 724 bp->b_cmd = BUF_CMD_READ; 725 bp->b_bio1.bio_done = func; 726 bp->b_bio1.bio_caller_info1.ptr = arg; 727 BUF_KERNPROC(bp); 728 reqbp = NULL; /* don't func() reqbp, it's running async */ 729 730 KASSERT(firstread != NOOFFSET, 731 ("cluster_read: no buffer offset")); 732 733 /* 734 * nblks is our cluster_rbuild request size, limited 735 * primarily by the device. 736 */ 737 nblks = calc_rbuild_reqsize(maxra, maxrbuild); 738 739 /* 740 * Set RAM half-way through the full-cluster. 741 */ 742 sr = (maxra + 1) / 2; 743 744 if (nblks > 1) { 745 int burstbytes; 746 747 error = VOP_BMAP(vp, loffset, &doffset, 748 &burstbytes, NULL, BUF_CMD_READ); 749 if (error) 750 goto single_block_read; 751 if (nblks > burstbytes / blksize) 752 nblks = burstbytes / blksize; 753 if (doffset == NOOFFSET) 754 goto single_block_read; 755 if (nblks <= 1) 756 goto single_block_read; 757 758 bp = cluster_rbuild(vp, filesize, loffset, 759 doffset, blksize, nblks, bp, &sr); 760 loffset += bp->b_bufsize; 761 maxra -= bp->b_bufsize / blksize; 762 } else { 763 single_block_read: 764 /* 765 * If it isn't in the cache, then get a chunk from 766 * disk if sequential, otherwise just get the block. 767 */ 768 loffset += blksize; 769 --maxra; 770 } 771 } 772 773 /* 774 * If bp != NULL then B_CACHE was *NOT* set and bp must be issued. 775 * bp will either be an asynchronous cluster buf or an asynchronous 776 * single-buf. 777 * 778 * NOTE: Once an async cluster buf is issued bp becomes invalid. 779 */ 780 if (bp) { 781 #if defined(CLUSTERDEBUG) 782 if (rcluster) 783 kprintf("S(%012jx,%d,%d)\n", 784 (intmax_t)bp->b_loffset, bp->b_bcount, maxra); 785 #endif 786 if ((bp->b_flags & B_CLUSTER) == 0) 787 vfs_busy_pages(vp, bp); 788 bp->b_flags &= ~(B_ERROR | B_INVAL | B_NOTMETA); 789 bp->b_flags |= bflags; 790 vn_strategy(vp, &bp->b_bio1); 791 /* bp invalid now */ 792 bp = NULL; 793 } 794 795 #if defined(CLUSTERDEBUG) 796 if (rcluster) 797 kprintf("cluster_rd %016jx/%d maxra=%d sr=%d\n", 798 loffset, blksize, maxra, sr); 799 #endif 800 801 /* 802 * If we have been doing sequential I/O, then do some read-ahead. 803 * The code above us should have positioned us at the next likely 804 * offset. 805 * 806 * Only mess with buffers which we can immediately lock. HAMMER 807 * will do device-readahead irrespective of what the blocks 808 * represent. 809 */ 810 while (maxra > 0) { 811 int burstbytes; 812 int error; 813 int nblks; 814 815 rbp = getblk(vp, loffset, blksize, 816 GETBLK_SZMATCH|GETBLK_NOWAIT, 0); 817 if (rbp == NULL) 818 goto no_read_ahead; 819 if ((rbp->b_flags & B_CACHE)) { 820 bqrelse(rbp); 821 goto no_read_ahead; 822 } 823 824 /* 825 * If BMAP is not supported or has an issue, we still do 826 * (maxra) read-ahead, but we do not try to use rbuild. 827 */ 828 error = VOP_BMAP(vp, loffset, &doffset, 829 &burstbytes, NULL, BUF_CMD_READ); 830 if (error || doffset == NOOFFSET) { 831 nblks = 1; 832 doffset = NOOFFSET; 833 } else { 834 nblks = calc_rbuild_reqsize(maxra, maxrbuild); 835 if (nblks > burstbytes / blksize) 836 nblks = burstbytes / blksize; 837 } 838 rbp->b_cmd = BUF_CMD_READ; 839 840 if (nblks > 1) { 841 rbp = cluster_rbuild(vp, filesize, loffset, 842 doffset, blksize, 843 nblks, rbp, &sr); 844 } else { 845 rbp->b_bio2.bio_offset = doffset; 846 if (--sr == 0) 847 cluster_setram(rbp); 848 } 849 850 rbp->b_flags &= ~(B_ERROR | B_INVAL | B_NOTMETA); 851 rbp->b_flags |= bflags; 852 853 if ((rbp->b_flags & B_CLUSTER) == 0) 854 vfs_busy_pages(vp, rbp); 855 BUF_KERNPROC(rbp); 856 loffset += rbp->b_bufsize; 857 maxra -= rbp->b_bufsize / blksize; 858 vn_strategy(vp, &rbp->b_bio1); 859 /* rbp invalid now */ 860 } 861 862 /* 863 * If reqbp is non-NULL it had B_CACHE set and we issue the 864 * function callback synchronously. 865 * 866 * Note that we may start additional asynchronous I/O before doing 867 * the func() callback for the B_CACHE case 868 */ 869 no_read_ahead: 870 if (reqbp) 871 func(&reqbp->b_bio1); 872 } 873 874 /* 875 * If blocks are contiguous on disk, use this to provide clustered 876 * read ahead. We will read as many blocks as possible sequentially 877 * and then parcel them up into logical blocks in the buffer hash table. 878 * 879 * This function either returns a cluster buf or it returns fbp. fbp is 880 * already expected to be set up as a synchronous or asynchronous request. 881 * 882 * If a cluster buf is returned it will always be async. 883 * 884 * (*srp) counts down original blocks to determine where B_RAM should be set. 885 * Set B_RAM when *srp drops to 0. If (*srp) starts at 0, B_RAM will not be 886 * set on any buffer. Make sure B_RAM is cleared on any other buffers to 887 * prevent degenerate read-aheads from being generated. 888 */ 889 static struct buf * 890 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset, off_t doffset, 891 int blksize, int run, struct buf *fbp, int *srp) 892 { 893 struct buf *bp, *tbp; 894 off_t boffset; 895 int i, j; 896 int maxiosize = vmaxiosize(vp); 897 898 /* 899 * avoid a division 900 */ 901 while (loffset + run * blksize > filesize) { 902 --run; 903 } 904 905 tbp = fbp; 906 tbp->b_bio2.bio_offset = doffset; 907 if((tbp->b_flags & B_MALLOC) || 908 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) { 909 if (--*srp == 0) 910 cluster_setram(tbp); 911 else 912 cluster_clrram(tbp); 913 return tbp; 914 } 915 916 /* 917 * Get a pbuf, limit cluster I/O on a per-device basis. If 918 * doing cluster I/O for a file, limit cluster I/O on a 919 * per-mount basis. 920 */ 921 if (vp->v_type == VCHR || vp->v_type == VBLK) 922 bp = trypbuf_kva(&vp->v_pbuf_count); 923 else 924 bp = trypbuf_kva(&vp->v_mount->mnt_pbuf_count); 925 926 if (bp == NULL) 927 return tbp; 928 929 /* 930 * We are synthesizing a buffer out of vm_page_t's, but 931 * if the block size is not page aligned then the starting 932 * address may not be either. Inherit the b_data offset 933 * from the original buffer. 934 */ 935 bp->b_vp = vp; 936 bp->b_data = (char *)((vm_offset_t)bp->b_data | 937 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 938 bp->b_flags |= B_CLUSTER | B_VMIO; 939 bp->b_cmd = BUF_CMD_READ; 940 bp->b_bio1.bio_done = cluster_callback; /* default to async */ 941 bp->b_bio1.bio_caller_info1.cluster_head = NULL; 942 bp->b_bio1.bio_caller_info2.cluster_tail = NULL; 943 bp->b_loffset = loffset; 944 bp->b_bio2.bio_offset = doffset; 945 KASSERT(bp->b_loffset != NOOFFSET, 946 ("cluster_rbuild: no buffer offset")); 947 948 bp->b_bcount = 0; 949 bp->b_bufsize = 0; 950 bp->b_xio.xio_npages = 0; 951 952 for (boffset = doffset, i = 0; i < run; ++i, boffset += blksize) { 953 if (i) { 954 if ((bp->b_xio.xio_npages * PAGE_SIZE) + 955 round_page(blksize) > maxiosize) { 956 break; 957 } 958 959 /* 960 * Shortcut some checks and try to avoid buffers that 961 * would block in the lock. The same checks have to 962 * be made again after we officially get the buffer. 963 */ 964 tbp = getblk(vp, loffset + i * blksize, blksize, 965 GETBLK_SZMATCH|GETBLK_NOWAIT, 0); 966 if (tbp == NULL) 967 break; 968 for (j = 0; j < tbp->b_xio.xio_npages; j++) { 969 if (tbp->b_xio.xio_pages[j]->valid) 970 break; 971 } 972 if (j != tbp->b_xio.xio_npages) { 973 bqrelse(tbp); 974 break; 975 } 976 977 /* 978 * Stop scanning if the buffer is fuly valid 979 * (marked B_CACHE), or locked (may be doing a 980 * background write), or if the buffer is not 981 * VMIO backed. The clustering code can only deal 982 * with VMIO-backed buffers. 983 */ 984 if ((tbp->b_flags & (B_CACHE|B_LOCKED)) || 985 (tbp->b_flags & B_VMIO) == 0 || 986 (LIST_FIRST(&tbp->b_dep) != NULL && 987 buf_checkread(tbp)) 988 ) { 989 bqrelse(tbp); 990 break; 991 } 992 993 /* 994 * The buffer must be completely invalid in order to 995 * take part in the cluster. If it is partially valid 996 * then we stop. 997 */ 998 for (j = 0;j < tbp->b_xio.xio_npages; j++) { 999 if (tbp->b_xio.xio_pages[j]->valid) 1000 break; 1001 } 1002 if (j != tbp->b_xio.xio_npages) { 1003 bqrelse(tbp); 1004 break; 1005 } 1006 1007 /* 1008 * Depress the priority of buffers not explicitly 1009 * requested. 1010 */ 1011 /* tbp->b_flags |= B_AGE; */ 1012 1013 /* 1014 * Set the block number if it isn't set, otherwise 1015 * if it is make sure it matches the block number we 1016 * expect. 1017 */ 1018 if (tbp->b_bio2.bio_offset == NOOFFSET) { 1019 tbp->b_bio2.bio_offset = boffset; 1020 } else if (tbp->b_bio2.bio_offset != boffset) { 1021 brelse(tbp); 1022 break; 1023 } 1024 } 1025 1026 /* 1027 * Set B_RAM if (*srp) is 1. B_RAM is only set on one buffer 1028 * in the cluster, including potentially the first buffer 1029 * once we start streaming the read-aheads. 1030 */ 1031 if (--*srp == 0) 1032 cluster_setram(tbp); 1033 else 1034 cluster_clrram(tbp); 1035 1036 /* 1037 * The passed-in tbp (i == 0) will already be set up for 1038 * async or sync operation. All other tbp's acquire in 1039 * our loop are set up for async operation. 1040 */ 1041 tbp->b_cmd = BUF_CMD_READ; 1042 BUF_KERNPROC(tbp); 1043 cluster_append(&bp->b_bio1, tbp); 1044 for (j = 0; j < tbp->b_xio.xio_npages; ++j) { 1045 vm_page_t m; 1046 1047 m = tbp->b_xio.xio_pages[j]; 1048 vm_page_busy_wait(m, FALSE, "clurpg"); 1049 vm_page_io_start(m); 1050 vm_page_wakeup(m); 1051 vm_object_pip_add(m->object, 1); 1052 if ((bp->b_xio.xio_npages == 0) || 1053 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) { 1054 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 1055 bp->b_xio.xio_npages++; 1056 } 1057 if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) { 1058 tbp->b_xio.xio_pages[j] = bogus_page; 1059 tbp->b_flags |= B_HASBOGUS; 1060 } 1061 } 1062 /* 1063 * XXX shouldn't this be += size for both, like in 1064 * cluster_wbuild()? 1065 * 1066 * Don't inherit tbp->b_bufsize as it may be larger due to 1067 * a non-page-aligned size. Instead just aggregate using 1068 * 'size'. 1069 */ 1070 if (tbp->b_bcount != blksize) 1071 kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, blksize); 1072 if (tbp->b_bufsize != blksize) 1073 kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, blksize); 1074 bp->b_bcount += blksize; 1075 bp->b_bufsize += blksize; 1076 } 1077 1078 /* 1079 * Fully valid pages in the cluster are already good and do not need 1080 * to be re-read from disk. Replace the page with bogus_page 1081 */ 1082 for (j = 0; j < bp->b_xio.xio_npages; j++) { 1083 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) == 1084 VM_PAGE_BITS_ALL) { 1085 bp->b_xio.xio_pages[j] = bogus_page; 1086 bp->b_flags |= B_HASBOGUS; 1087 } 1088 } 1089 if (bp->b_bufsize > bp->b_kvasize) { 1090 panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)", 1091 bp->b_bufsize, bp->b_kvasize); 1092 } 1093 pmap_qenter(trunc_page((vm_offset_t) bp->b_data), 1094 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages); 1095 BUF_KERNPROC(bp); 1096 return (bp); 1097 } 1098 1099 /* 1100 * Cleanup after a clustered read or write. 1101 * This is complicated by the fact that any of the buffers might have 1102 * extra memory (if there were no empty buffer headers at allocbuf time) 1103 * that we will need to shift around. 1104 * 1105 * The returned bio is &bp->b_bio1 1106 */ 1107 static void 1108 cluster_callback(struct bio *bio) 1109 { 1110 struct buf *bp = bio->bio_buf; 1111 struct buf *tbp; 1112 struct vnode *vp; 1113 int error = 0; 1114 1115 /* 1116 * Must propogate errors to all the components. A short read (EOF) 1117 * is a critical error. 1118 */ 1119 if (bp->b_flags & B_ERROR) { 1120 error = bp->b_error; 1121 } else if (bp->b_bcount != bp->b_bufsize) { 1122 panic("cluster_callback: unexpected EOF on cluster %p!", bio); 1123 } 1124 1125 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), 1126 bp->b_xio.xio_npages); 1127 /* 1128 * Move memory from the large cluster buffer into the component 1129 * buffers and mark IO as done on these. Since the memory map 1130 * is the same, no actual copying is required. 1131 */ 1132 while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) { 1133 bio->bio_caller_info1.cluster_head = tbp->b_cluster_next; 1134 if (error) { 1135 tbp->b_flags |= B_ERROR | B_IOISSUED; 1136 tbp->b_error = error; 1137 } else { 1138 tbp->b_dirtyoff = tbp->b_dirtyend = 0; 1139 tbp->b_flags &= ~(B_ERROR | B_INVAL); 1140 if (tbp->b_cmd == BUF_CMD_READ) { 1141 tbp->b_flags = (tbp->b_flags & ~B_NOTMETA) | 1142 (bp->b_flags & B_NOTMETA); 1143 } 1144 tbp->b_flags |= B_IOISSUED; 1145 /* 1146 * XXX the bdwrite()/bqrelse() issued during 1147 * cluster building clears B_RELBUF (see bqrelse() 1148 * comment). If direct I/O was specified, we have 1149 * to restore it here to allow the buffer and VM 1150 * to be freed. 1151 */ 1152 if (tbp->b_flags & B_DIRECT) 1153 tbp->b_flags |= B_RELBUF; 1154 1155 /* 1156 * XXX I think biodone() below will do this, but do 1157 * it here anyway for consistency. 1158 */ 1159 if (tbp->b_cmd == BUF_CMD_WRITE) 1160 bundirty(tbp); 1161 } 1162 biodone(&tbp->b_bio1); 1163 } 1164 vp = bp->b_vp; 1165 bp->b_vp = NULL; 1166 if (vp->v_type == VCHR || vp->v_type == VBLK) 1167 relpbuf(bp, &vp->v_pbuf_count); 1168 else 1169 relpbuf(bp, &vp->v_mount->mnt_pbuf_count); 1170 } 1171 1172 /* 1173 * Implement modified write build for cluster. 1174 * 1175 * write_behind = 0 write behind disabled 1176 * write_behind = 1 write behind normal (default) 1177 * write_behind = 2 write behind backed-off 1178 * 1179 * In addition, write_behind is only activated for files that have 1180 * grown past a certain size (default 10MB). Otherwise temporary files 1181 * wind up generating a lot of unnecessary disk I/O. 1182 */ 1183 static __inline int 1184 cluster_wbuild_wb(struct vnode *vp, int blksize, off_t start_loffset, int len) 1185 { 1186 int r = 0; 1187 1188 switch(write_behind) { 1189 case 2: 1190 if (start_loffset < len) 1191 break; 1192 start_loffset -= len; 1193 /* fall through */ 1194 case 1: 1195 if (vp->v_filesize >= write_behind_minfilesize) { 1196 r = cluster_wbuild(vp, NULL, blksize, 1197 start_loffset, len); 1198 } 1199 /* fall through */ 1200 default: 1201 /* fall through */ 1202 break; 1203 } 1204 return(r); 1205 } 1206 1207 /* 1208 * Do clustered write for FFS. 1209 * 1210 * Three cases: 1211 * 1. Write is not sequential (write asynchronously) 1212 * Write is sequential: 1213 * 2. beginning of cluster - begin cluster 1214 * 3. middle of a cluster - add to cluster 1215 * 4. end of a cluster - asynchronously write cluster 1216 * 1217 * WARNING! vnode fields are not locked and must ONLY be used heuristically. 1218 */ 1219 void 1220 cluster_write(struct buf *bp, off_t filesize, int blksize, int seqcount) 1221 { 1222 struct vnode *vp; 1223 off_t loffset; 1224 int maxclen, cursize; 1225 int async; 1226 cluster_cache_t dummy; 1227 cluster_cache_t *cc; 1228 1229 vp = bp->b_vp; 1230 if (vp->v_type == VREG) 1231 async = vp->v_mount->mnt_flag & MNT_ASYNC; 1232 else 1233 async = 0; 1234 loffset = bp->b_loffset; 1235 KASSERT(bp->b_loffset != NOOFFSET, 1236 ("cluster_write: no buffer offset")); 1237 1238 cc = cluster_getcache(&dummy, vp, loffset); 1239 1240 /* 1241 * Initialize vnode to beginning of file. 1242 */ 1243 if (loffset == 0) 1244 cc->v_lasta = cc->v_clen = cc->v_cstart = cc->v_lastw = 0; 1245 1246 if (cc->v_clen == 0 || loffset != cc->v_lastw || 1247 (bp->b_bio2.bio_offset != NOOFFSET && 1248 (bp->b_bio2.bio_offset != cc->v_lasta))) { 1249 /* 1250 * Next block is not logically sequential, or, if physical 1251 * block offsets are available, not physically sequential. 1252 * 1253 * If physical block offsets are not available we only 1254 * get here if we weren't logically sequential. 1255 */ 1256 maxclen = vmaxiosize(vp); 1257 if (cc->v_clen != 0) { 1258 /* 1259 * Next block is not sequential. 1260 * 1261 * If we are not writing at end of file, the process 1262 * seeked to another point in the file since its last 1263 * write, or we have reached our maximum cluster size, 1264 * then push the previous cluster. Otherwise try 1265 * reallocating to make it sequential. 1266 * 1267 * Change to algorithm: only push previous cluster if 1268 * it was sequential from the point of view of the 1269 * seqcount heuristic, otherwise leave the buffer 1270 * intact so we can potentially optimize the I/O 1271 * later on in the buf_daemon or update daemon 1272 * flush. 1273 */ 1274 cursize = cc->v_lastw - cc->v_cstart; 1275 if (bp->b_loffset + blksize < filesize || 1276 loffset != cc->v_lastw || 1277 cc->v_clen <= cursize) { 1278 if (!async && seqcount > 0) { 1279 cluster_wbuild_wb(vp, blksize, 1280 cc->v_cstart, cursize); 1281 } 1282 } else { 1283 struct buf **bpp, **endbp; 1284 struct cluster_save *buflist; 1285 1286 buflist = cluster_collectbufs(cc, vp, 1287 bp, blksize); 1288 endbp = &buflist->bs_children 1289 [buflist->bs_nchildren - 1]; 1290 if (VOP_REALLOCBLKS(vp, buflist)) { 1291 /* 1292 * Failed, push the previous cluster 1293 * if *really* writing sequentially 1294 * in the logical file (seqcount > 1), 1295 * otherwise delay it in the hopes that 1296 * the low level disk driver can 1297 * optimize the write ordering. 1298 * 1299 * NOTE: We do not brelse the last 1300 * element which is bp, and we 1301 * do not return here. 1302 */ 1303 for (bpp = buflist->bs_children; 1304 bpp < endbp; bpp++) 1305 brelse(*bpp); 1306 kfree(buflist, M_SEGMENT); 1307 if (seqcount > 1) { 1308 cluster_wbuild_wb(vp, 1309 blksize, cc->v_cstart, 1310 cursize); 1311 } 1312 } else { 1313 /* 1314 * Succeeded, keep building cluster. 1315 */ 1316 for (bpp = buflist->bs_children; 1317 bpp <= endbp; bpp++) 1318 bdwrite(*bpp); 1319 kfree(buflist, M_SEGMENT); 1320 cc->v_lastw = loffset + blksize; 1321 cc->v_lasta = bp->b_bio2.bio_offset + 1322 blksize; 1323 cluster_putcache(cc); 1324 return; 1325 } 1326 } 1327 } 1328 1329 /* 1330 * Consider beginning a cluster. If at end of file, make 1331 * cluster as large as possible, otherwise find size of 1332 * existing cluster. 1333 */ 1334 if ((vp->v_type == VREG) && 1335 bp->b_loffset + blksize < filesize && 1336 (bp->b_bio2.bio_offset == NOOFFSET) && 1337 (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL, BUF_CMD_WRITE) || 1338 bp->b_bio2.bio_offset == NOOFFSET)) { 1339 bdwrite(bp); 1340 cc->v_clen = 0; 1341 cc->v_lasta = bp->b_bio2.bio_offset + blksize; 1342 cc->v_cstart = loffset; 1343 cc->v_lastw = loffset + blksize; 1344 cluster_putcache(cc); 1345 return; 1346 } 1347 if (maxclen > blksize) 1348 cc->v_clen = maxclen; 1349 else 1350 cc->v_clen = blksize; 1351 if (!async && cc->v_clen == 0) { /* I/O not contiguous */ 1352 cc->v_cstart = loffset; 1353 bdwrite(bp); 1354 } else { /* Wait for rest of cluster */ 1355 cc->v_cstart = loffset; 1356 bdwrite(bp); 1357 } 1358 } else if (loffset == cc->v_cstart + cc->v_clen) { 1359 /* 1360 * At end of cluster, write it out if seqcount tells us we 1361 * are operating sequentially, otherwise let the buf or 1362 * update daemon handle it. 1363 */ 1364 bdwrite(bp); 1365 if (seqcount > 1) 1366 cluster_wbuild_wb(vp, blksize, cc->v_cstart, 1367 cc->v_clen + blksize); 1368 cc->v_clen = 0; 1369 cc->v_cstart = loffset; 1370 } else if (vm_page_count_severe() && 1371 bp->b_loffset + blksize < filesize) { 1372 /* 1373 * We are low on memory, get it going NOW. However, do not 1374 * try to push out a partial block at the end of the file 1375 * as this could lead to extremely non-optimal write activity. 1376 */ 1377 bawrite(bp); 1378 } else { 1379 /* 1380 * In the middle of a cluster, so just delay the I/O for now. 1381 */ 1382 bdwrite(bp); 1383 } 1384 cc->v_lastw = loffset + blksize; 1385 cc->v_lasta = bp->b_bio2.bio_offset + blksize; 1386 cluster_putcache(cc); 1387 } 1388 1389 /* 1390 * This is the clustered version of bawrite(). It works similarly to 1391 * cluster_write() except I/O on the buffer is guaranteed to occur. 1392 */ 1393 int 1394 cluster_awrite(struct buf *bp) 1395 { 1396 int total; 1397 1398 /* 1399 * Don't bother if it isn't clusterable. 1400 */ 1401 if ((bp->b_flags & B_CLUSTEROK) == 0 || 1402 bp->b_vp == NULL || 1403 (bp->b_vp->v_flag & VOBJBUF) == 0) { 1404 total = bp->b_bufsize; 1405 bawrite(bp); 1406 return (total); 1407 } 1408 1409 total = cluster_wbuild(bp->b_vp, &bp, bp->b_bufsize, 1410 bp->b_loffset, vmaxiosize(bp->b_vp)); 1411 1412 /* 1413 * If bp is still non-NULL then cluster_wbuild() did not initiate 1414 * I/O on it and we must do so here to provide the API guarantee. 1415 */ 1416 if (bp) 1417 bawrite(bp); 1418 1419 return total; 1420 } 1421 1422 /* 1423 * This is an awful lot like cluster_rbuild...wish they could be combined. 1424 * The last lbn argument is the current block on which I/O is being 1425 * performed. Check to see that it doesn't fall in the middle of 1426 * the current block (if last_bp == NULL). 1427 * 1428 * cluster_wbuild() normally does not guarantee anything. If bpp is 1429 * non-NULL and cluster_wbuild() is able to incorporate it into the 1430 * I/O it will set *bpp to NULL, otherwise it will leave it alone and 1431 * the caller must dispose of *bpp. 1432 */ 1433 static int 1434 cluster_wbuild(struct vnode *vp, struct buf **bpp, 1435 int blksize, off_t start_loffset, int bytes) 1436 { 1437 struct buf *bp, *tbp; 1438 int i, j; 1439 int totalwritten = 0; 1440 int must_initiate; 1441 int maxiosize = vmaxiosize(vp); 1442 1443 while (bytes > 0) { 1444 /* 1445 * If the buffer matches the passed locked & removed buffer 1446 * we used the passed buffer (which might not be B_DELWRI). 1447 * 1448 * Otherwise locate the buffer and determine if it is 1449 * compatible. 1450 */ 1451 if (bpp && (*bpp)->b_loffset == start_loffset) { 1452 tbp = *bpp; 1453 *bpp = NULL; 1454 bpp = NULL; 1455 } else { 1456 tbp = findblk(vp, start_loffset, FINDBLK_NBLOCK); 1457 if (tbp == NULL || 1458 (tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != 1459 B_DELWRI || 1460 (LIST_FIRST(&tbp->b_dep) && buf_checkwrite(tbp))) { 1461 if (tbp) 1462 BUF_UNLOCK(tbp); 1463 start_loffset += blksize; 1464 bytes -= blksize; 1465 continue; 1466 } 1467 bremfree(tbp); 1468 } 1469 KKASSERT(tbp->b_cmd == BUF_CMD_DONE); 1470 1471 /* 1472 * Extra memory in the buffer, punt on this buffer. 1473 * XXX we could handle this in most cases, but we would 1474 * have to push the extra memory down to after our max 1475 * possible cluster size and then potentially pull it back 1476 * up if the cluster was terminated prematurely--too much 1477 * hassle. 1478 */ 1479 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) || 1480 (tbp->b_bcount != tbp->b_bufsize) || 1481 (tbp->b_bcount != blksize) || 1482 (bytes == blksize)) { 1483 totalwritten += tbp->b_bufsize; 1484 bawrite(tbp); 1485 start_loffset += blksize; 1486 bytes -= blksize; 1487 continue; 1488 } 1489 1490 /* 1491 * Get a pbuf, limit cluster I/O on a per-device basis. If 1492 * doing cluster I/O for a file, limit cluster I/O on a 1493 * per-mount basis. 1494 * 1495 * HAMMER and other filesystems may attempt to queue a massive 1496 * amount of write I/O, using trypbuf() here easily results in 1497 * situation where the I/O stream becomes non-clustered. 1498 */ 1499 if (vp->v_type == VCHR || vp->v_type == VBLK) 1500 bp = getpbuf_kva(&vp->v_pbuf_count); 1501 else 1502 bp = getpbuf_kva(&vp->v_mount->mnt_pbuf_count); 1503 1504 /* 1505 * Set up the pbuf. Track our append point with b_bcount 1506 * and b_bufsize. b_bufsize is not used by the device but 1507 * our caller uses it to loop clusters and we use it to 1508 * detect a premature EOF on the block device. 1509 */ 1510 bp->b_bcount = 0; 1511 bp->b_bufsize = 0; 1512 bp->b_xio.xio_npages = 0; 1513 bp->b_loffset = tbp->b_loffset; 1514 bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset; 1515 bp->b_vp = vp; 1516 1517 /* 1518 * We are synthesizing a buffer out of vm_page_t's, but 1519 * if the block size is not page aligned then the starting 1520 * address may not be either. Inherit the b_data offset 1521 * from the original buffer. 1522 */ 1523 bp->b_data = (char *)((vm_offset_t)bp->b_data | 1524 ((vm_offset_t)tbp->b_data & PAGE_MASK)); 1525 bp->b_flags &= ~(B_ERROR | B_NOTMETA); 1526 bp->b_flags |= B_CLUSTER | B_BNOCLIP | 1527 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT | 1528 B_NOTMETA)); 1529 bp->b_bio1.bio_caller_info1.cluster_head = NULL; 1530 bp->b_bio1.bio_caller_info2.cluster_tail = NULL; 1531 1532 /* 1533 * From this location in the file, scan forward to see 1534 * if there are buffers with adjacent data that need to 1535 * be written as well. 1536 * 1537 * IO *must* be initiated on index 0 at this point 1538 * (particularly when called from cluster_awrite()). 1539 */ 1540 for (i = 0; i < bytes; (i += blksize), (start_loffset += blksize)) { 1541 if (i == 0) { 1542 must_initiate = 1; 1543 } else { 1544 /* 1545 * Not first buffer. 1546 */ 1547 must_initiate = 0; 1548 tbp = findblk(vp, start_loffset, 1549 FINDBLK_NBLOCK); 1550 /* 1551 * Buffer not found or could not be locked 1552 * non-blocking. 1553 */ 1554 if (tbp == NULL) 1555 break; 1556 1557 /* 1558 * If it IS in core, but has different 1559 * characteristics, then don't cluster 1560 * with it. 1561 */ 1562 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK | 1563 B_INVAL | B_DELWRI | B_NEEDCOMMIT)) 1564 != (B_DELWRI | B_CLUSTEROK | 1565 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) || 1566 (tbp->b_flags & B_LOCKED) 1567 ) { 1568 BUF_UNLOCK(tbp); 1569 break; 1570 } 1571 1572 /* 1573 * Check that the combined cluster 1574 * would make sense with regard to pages 1575 * and would not be too large 1576 * 1577 * WARNING! buf_checkwrite() must be the last 1578 * check made. If it returns 0 then 1579 * we must initiate the I/O. 1580 */ 1581 if ((tbp->b_bcount != blksize) || 1582 ((bp->b_bio2.bio_offset + i) != 1583 tbp->b_bio2.bio_offset) || 1584 ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) > 1585 (maxiosize / PAGE_SIZE)) || 1586 (LIST_FIRST(&tbp->b_dep) && 1587 buf_checkwrite(tbp)) 1588 ) { 1589 BUF_UNLOCK(tbp); 1590 break; 1591 } 1592 if (LIST_FIRST(&tbp->b_dep)) 1593 must_initiate = 1; 1594 /* 1595 * Ok, it's passed all the tests, 1596 * so remove it from the free list 1597 * and mark it busy. We will use it. 1598 */ 1599 bremfree(tbp); 1600 KKASSERT(tbp->b_cmd == BUF_CMD_DONE); 1601 } 1602 1603 /* 1604 * If the IO is via the VM then we do some 1605 * special VM hackery (yuck). Since the buffer's 1606 * block size may not be page-aligned it is possible 1607 * for a page to be shared between two buffers. We 1608 * have to get rid of the duplication when building 1609 * the cluster. 1610 */ 1611 if (tbp->b_flags & B_VMIO) { 1612 vm_page_t m; 1613 1614 /* 1615 * Try to avoid deadlocks with the VM system. 1616 * However, we cannot abort the I/O if 1617 * must_initiate is non-zero. 1618 */ 1619 if (must_initiate == 0) { 1620 for (j = 0; 1621 j < tbp->b_xio.xio_npages; 1622 ++j) { 1623 m = tbp->b_xio.xio_pages[j]; 1624 if (m->flags & PG_BUSY) { 1625 bqrelse(tbp); 1626 goto finishcluster; 1627 } 1628 } 1629 } 1630 1631 for (j = 0; j < tbp->b_xio.xio_npages; ++j) { 1632 m = tbp->b_xio.xio_pages[j]; 1633 vm_page_busy_wait(m, FALSE, "clurpg"); 1634 vm_page_io_start(m); 1635 vm_page_wakeup(m); 1636 vm_object_pip_add(m->object, 1); 1637 if ((bp->b_xio.xio_npages == 0) || 1638 (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) { 1639 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m; 1640 bp->b_xio.xio_npages++; 1641 } 1642 } 1643 } 1644 bp->b_bcount += blksize; 1645 bp->b_bufsize += blksize; 1646 1647 /* 1648 * NOTE: see bwrite/bawrite code for why we no longer 1649 * undirty tbp here. 1650 * 1651 * bundirty(tbp); REMOVED 1652 */ 1653 tbp->b_flags &= ~B_ERROR; 1654 tbp->b_cmd = BUF_CMD_WRITE; 1655 BUF_KERNPROC(tbp); 1656 cluster_append(&bp->b_bio1, tbp); 1657 1658 /* 1659 * check for latent dependencies to be handled 1660 */ 1661 if (LIST_FIRST(&tbp->b_dep) != NULL) 1662 buf_start(tbp); 1663 } 1664 finishcluster: 1665 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 1666 (vm_page_t *)bp->b_xio.xio_pages, 1667 bp->b_xio.xio_npages); 1668 if (bp->b_bufsize > bp->b_kvasize) { 1669 panic("cluster_wbuild: b_bufsize(%d) " 1670 "> b_kvasize(%d)\n", 1671 bp->b_bufsize, bp->b_kvasize); 1672 } 1673 totalwritten += bp->b_bufsize; 1674 bp->b_dirtyoff = 0; 1675 bp->b_dirtyend = bp->b_bufsize; 1676 bp->b_bio1.bio_done = cluster_callback; 1677 bp->b_cmd = BUF_CMD_WRITE; 1678 1679 vfs_busy_pages(vp, bp); 1680 bsetrunningbufspace(bp, bp->b_bufsize); 1681 BUF_KERNPROC(bp); 1682 vn_strategy(vp, &bp->b_bio1); 1683 1684 bytes -= i; 1685 } 1686 return totalwritten; 1687 } 1688 1689 /* 1690 * Collect together all the buffers in a cluster, plus add one 1691 * additional buffer passed-in. 1692 * 1693 * Only pre-existing buffers whos block size matches blksize are collected. 1694 * (this is primarily because HAMMER1 uses varying block sizes and we don't 1695 * want to override its choices). 1696 * 1697 * This code will not try to collect buffers that it cannot lock, otherwise 1698 * it might deadlock against SMP-friendly filesystems. 1699 */ 1700 static struct cluster_save * 1701 cluster_collectbufs(cluster_cache_t *cc, struct vnode *vp, 1702 struct buf *last_bp, int blksize) 1703 { 1704 struct cluster_save *buflist; 1705 struct buf *bp; 1706 off_t loffset; 1707 int i, len; 1708 int j; 1709 int k; 1710 1711 len = (int)(cc->v_lastw - cc->v_cstart) / blksize; 1712 KKASSERT(len > 0); 1713 buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist), 1714 M_SEGMENT, M_WAITOK); 1715 buflist->bs_nchildren = 0; 1716 buflist->bs_children = (struct buf **) (buflist + 1); 1717 for (loffset = cc->v_cstart, i = 0, j = 0; 1718 i < len; 1719 (loffset += blksize), i++) { 1720 bp = getcacheblk(vp, loffset, 1721 last_bp->b_bcount, GETBLK_SZMATCH | 1722 GETBLK_NOWAIT); 1723 buflist->bs_children[i] = bp; 1724 if (bp == NULL) { 1725 j = i + 1; 1726 } else if (bp->b_bio2.bio_offset == NOOFFSET) { 1727 VOP_BMAP(bp->b_vp, bp->b_loffset, 1728 &bp->b_bio2.bio_offset, 1729 NULL, NULL, BUF_CMD_WRITE); 1730 } 1731 } 1732 1733 /* 1734 * Get rid of gaps 1735 */ 1736 for (k = 0; k < j; ++k) { 1737 if (buflist->bs_children[k]) { 1738 bqrelse(buflist->bs_children[k]); 1739 buflist->bs_children[k] = NULL; 1740 } 1741 } 1742 if (j != 0) { 1743 if (j != i) { 1744 bcopy(buflist->bs_children + j, 1745 buflist->bs_children + 0, 1746 sizeof(buflist->bs_children[0]) * (i - j)); 1747 } 1748 i -= j; 1749 } 1750 buflist->bs_children[i] = bp = last_bp; 1751 if (bp->b_bio2.bio_offset == NOOFFSET) { 1752 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset, 1753 NULL, NULL, BUF_CMD_WRITE); 1754 } 1755 buflist->bs_nchildren = i + 1; 1756 return (buflist); 1757 } 1758 1759 void 1760 cluster_append(struct bio *bio, struct buf *tbp) 1761 { 1762 tbp->b_cluster_next = NULL; 1763 if (bio->bio_caller_info1.cluster_head == NULL) { 1764 bio->bio_caller_info1.cluster_head = tbp; 1765 bio->bio_caller_info2.cluster_tail = tbp; 1766 } else { 1767 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp; 1768 bio->bio_caller_info2.cluster_tail = tbp; 1769 } 1770 } 1771 1772 static 1773 void 1774 cluster_setram(struct buf *bp) 1775 { 1776 bp->b_flags |= B_RAM; 1777 if (bp->b_xio.xio_npages) 1778 vm_page_flag_set(bp->b_xio.xio_pages[0], PG_RAM); 1779 } 1780 1781 static 1782 void 1783 cluster_clrram(struct buf *bp) 1784 { 1785 bp->b_flags &= ~B_RAM; 1786 if (bp->b_xio.xio_npages) 1787 vm_page_flag_clear(bp->b_xio.xio_pages[0], PG_RAM); 1788 } 1789