1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26#include <sys/zfs_context.h> 27#include <sys/dnode.h> 28#include <sys/dmu_objset.h> 29#include <sys/dmu_zfetch.h> 30#include <sys/dmu.h> 31#include <sys/dbuf.h> 32#include <sys/kstat.h> 33 34/* 35 * I'm against tune-ables, but these should probably exist as tweakable globals 36 * until we can get this working the way we want it to. 37 */ 38 39int zfs_prefetch_disable = 0; 40 41/* max # of streams per zfetch */ 42uint32_t zfetch_max_streams = 8; 43/* min time before stream reclaim */ 44uint32_t zfetch_min_sec_reap = 2; 45/* max number of blocks to fetch at a time */ 46uint32_t zfetch_block_cap = 256; 47/* number of bytes in a array_read at which we stop prefetching (1Mb) */ 48uint64_t zfetch_array_rd_sz = 1024 * 1024; 49 50SYSCTL_DECL(_vfs_zfs); 51SYSCTL_INT(_vfs_zfs, OID_AUTO, prefetch_disable, CTLFLAG_RW, 52 &zfs_prefetch_disable, 0, "Disable prefetch"); 53SYSCTL_NODE(_vfs_zfs, OID_AUTO, zfetch, CTLFLAG_RW, 0, "ZFS ZFETCH"); 54TUNABLE_INT("vfs.zfs.zfetch.max_streams", &zfetch_max_streams); 55SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, max_streams, CTLFLAG_RW, 56 &zfetch_max_streams, 0, "Max # of streams per zfetch"); 57TUNABLE_INT("vfs.zfs.zfetch.min_sec_reap", &zfetch_min_sec_reap); 58SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, min_sec_reap, CTLFLAG_RDTUN, 59 &zfetch_min_sec_reap, 0, "Min time before stream reclaim"); 60TUNABLE_INT("vfs.zfs.zfetch.block_cap", &zfetch_block_cap); 61SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, block_cap, CTLFLAG_RDTUN, 62 &zfetch_block_cap, 0, "Max number of blocks to fetch at a time"); 63TUNABLE_QUAD("vfs.zfs.zfetch.array_rd_sz", &zfetch_array_rd_sz); 64SYSCTL_UQUAD(_vfs_zfs_zfetch, OID_AUTO, array_rd_sz, CTLFLAG_RDTUN, 65 &zfetch_array_rd_sz, 0, 66 "Number of bytes in a array_read at which we stop prefetching"); 67 68/* forward decls for static routines */ 69static boolean_t dmu_zfetch_colinear(zfetch_t *, zstream_t *); 70static void dmu_zfetch_dofetch(zfetch_t *, zstream_t *); 71static uint64_t dmu_zfetch_fetch(dnode_t *, uint64_t, uint64_t); 72static uint64_t dmu_zfetch_fetchsz(dnode_t *, uint64_t, uint64_t); 73static boolean_t dmu_zfetch_find(zfetch_t *, zstream_t *, int); 74static int dmu_zfetch_stream_insert(zfetch_t *, zstream_t *); 75static zstream_t *dmu_zfetch_stream_reclaim(zfetch_t *); 76static void dmu_zfetch_stream_remove(zfetch_t *, zstream_t *); 77static int dmu_zfetch_streams_equal(zstream_t *, zstream_t *); 78 79typedef struct zfetch_stats { 80 kstat_named_t zfetchstat_hits; 81 kstat_named_t zfetchstat_misses; 82 kstat_named_t zfetchstat_colinear_hits; 83 kstat_named_t zfetchstat_colinear_misses; 84 kstat_named_t zfetchstat_stride_hits; 85 kstat_named_t zfetchstat_stride_misses; 86 kstat_named_t zfetchstat_reclaim_successes; 87 kstat_named_t zfetchstat_reclaim_failures; 88 kstat_named_t zfetchstat_stream_resets; 89 kstat_named_t zfetchstat_stream_noresets; 90 kstat_named_t zfetchstat_bogus_streams; 91} zfetch_stats_t; 92 93static zfetch_stats_t zfetch_stats = { 94 { "hits", KSTAT_DATA_UINT64 }, 95 { "misses", KSTAT_DATA_UINT64 }, 96 { "colinear_hits", KSTAT_DATA_UINT64 }, 97 { "colinear_misses", KSTAT_DATA_UINT64 }, 98 { "stride_hits", KSTAT_DATA_UINT64 }, 99 { "stride_misses", KSTAT_DATA_UINT64 }, 100 { "reclaim_successes", KSTAT_DATA_UINT64 }, 101 { "reclaim_failures", KSTAT_DATA_UINT64 }, 102 { "streams_resets", KSTAT_DATA_UINT64 }, 103 { "streams_noresets", KSTAT_DATA_UINT64 }, 104 { "bogus_streams", KSTAT_DATA_UINT64 }, 105}; 106 107#define ZFETCHSTAT_INCR(stat, val) \ 108 atomic_add_64(&zfetch_stats.stat.value.ui64, (val)); 109 110#define ZFETCHSTAT_BUMP(stat) ZFETCHSTAT_INCR(stat, 1); 111 112kstat_t *zfetch_ksp; 113 114/* 115 * Given a zfetch structure and a zstream structure, determine whether the 116 * blocks to be read are part of a co-linear pair of existing prefetch 117 * streams. If a set is found, coalesce the streams, removing one, and 118 * configure the prefetch so it looks for a strided access pattern. 119 * 120 * In other words: if we find two sequential access streams that are 121 * the same length and distance N appart, and this read is N from the 122 * last stream, then we are probably in a strided access pattern. So 123 * combine the two sequential streams into a single strided stream. 124 * 125 * Returns whether co-linear streams were found. 126 */ 127static boolean_t 128dmu_zfetch_colinear(zfetch_t *zf, zstream_t *zh) 129{ 130 zstream_t *z_walk; 131 zstream_t *z_comp; 132 133 if (! rw_tryenter(&zf->zf_rwlock, RW_WRITER)) 134 return (0); 135 136 if (zh == NULL) { 137 rw_exit(&zf->zf_rwlock); 138 return (0); 139 } 140 141 for (z_walk = list_head(&zf->zf_stream); z_walk; 142 z_walk = list_next(&zf->zf_stream, z_walk)) { 143 for (z_comp = list_next(&zf->zf_stream, z_walk); z_comp; 144 z_comp = list_next(&zf->zf_stream, z_comp)) { 145 int64_t diff; 146 147 if (z_walk->zst_len != z_walk->zst_stride || 148 z_comp->zst_len != z_comp->zst_stride) { 149 continue; 150 } 151 152 diff = z_comp->zst_offset - z_walk->zst_offset; 153 if (z_comp->zst_offset + diff == zh->zst_offset) { 154 z_walk->zst_offset = zh->zst_offset; 155 z_walk->zst_direction = diff < 0 ? -1 : 1; 156 z_walk->zst_stride = 157 diff * z_walk->zst_direction; 158 z_walk->zst_ph_offset = 159 zh->zst_offset + z_walk->zst_stride; 160 dmu_zfetch_stream_remove(zf, z_comp); 161 mutex_destroy(&z_comp->zst_lock); 162 kmem_free(z_comp, sizeof (zstream_t)); 163 164 dmu_zfetch_dofetch(zf, z_walk); 165 166 rw_exit(&zf->zf_rwlock); 167 return (1); 168 } 169 170 diff = z_walk->zst_offset - z_comp->zst_offset; 171 if (z_walk->zst_offset + diff == zh->zst_offset) { 172 z_walk->zst_offset = zh->zst_offset; 173 z_walk->zst_direction = diff < 0 ? -1 : 1; 174 z_walk->zst_stride = 175 diff * z_walk->zst_direction; 176 z_walk->zst_ph_offset = 177 zh->zst_offset + z_walk->zst_stride; 178 dmu_zfetch_stream_remove(zf, z_comp); 179 mutex_destroy(&z_comp->zst_lock); 180 kmem_free(z_comp, sizeof (zstream_t)); 181 182 dmu_zfetch_dofetch(zf, z_walk); 183 184 rw_exit(&zf->zf_rwlock); 185 return (1); 186 } 187 } 188 } 189 190 rw_exit(&zf->zf_rwlock); 191 return (0); 192} 193 194/* 195 * Given a zstream_t, determine the bounds of the prefetch. Then call the 196 * routine that actually prefetches the individual blocks. 197 */ 198static void 199dmu_zfetch_dofetch(zfetch_t *zf, zstream_t *zs) 200{ 201 uint64_t prefetch_tail; 202 uint64_t prefetch_limit; 203 uint64_t prefetch_ofst; 204 uint64_t prefetch_len; 205 uint64_t blocks_fetched; 206 207 zs->zst_stride = MAX((int64_t)zs->zst_stride, zs->zst_len); 208 zs->zst_cap = MIN(zfetch_block_cap, 2 * zs->zst_cap); 209 210 prefetch_tail = MAX((int64_t)zs->zst_ph_offset, 211 (int64_t)(zs->zst_offset + zs->zst_stride)); 212 /* 213 * XXX: use a faster division method? 214 */ 215 prefetch_limit = zs->zst_offset + zs->zst_len + 216 (zs->zst_cap * zs->zst_stride) / zs->zst_len; 217 218 while (prefetch_tail < prefetch_limit) { 219 prefetch_ofst = zs->zst_offset + zs->zst_direction * 220 (prefetch_tail - zs->zst_offset); 221 222 prefetch_len = zs->zst_len; 223 224 /* 225 * Don't prefetch beyond the end of the file, if working 226 * backwards. 227 */ 228 if ((zs->zst_direction == ZFETCH_BACKWARD) && 229 (prefetch_ofst > prefetch_tail)) { 230 prefetch_len += prefetch_ofst; 231 prefetch_ofst = 0; 232 } 233 234 /* don't prefetch more than we're supposed to */ 235 if (prefetch_len > zs->zst_len) 236 break; 237 238 blocks_fetched = dmu_zfetch_fetch(zf->zf_dnode, 239 prefetch_ofst, zs->zst_len); 240 241 prefetch_tail += zs->zst_stride; 242 /* stop if we've run out of stuff to prefetch */ 243 if (blocks_fetched < zs->zst_len) 244 break; 245 } 246 zs->zst_ph_offset = prefetch_tail; 247 zs->zst_last = ddi_get_lbolt(); 248} 249 250void 251zfetch_init(void) 252{ 253 254 zfetch_ksp = kstat_create("zfs", 0, "zfetchstats", "misc", 255 KSTAT_TYPE_NAMED, sizeof (zfetch_stats) / sizeof (kstat_named_t), 256 KSTAT_FLAG_VIRTUAL); 257 258 if (zfetch_ksp != NULL) { 259 zfetch_ksp->ks_data = &zfetch_stats; 260 kstat_install(zfetch_ksp); 261 } 262} 263 264void 265zfetch_fini(void) 266{ 267 if (zfetch_ksp != NULL) { 268 kstat_delete(zfetch_ksp); 269 zfetch_ksp = NULL; 270 } 271} 272 273/* 274 * This takes a pointer to a zfetch structure and a dnode. It performs the 275 * necessary setup for the zfetch structure, grokking data from the 276 * associated dnode. 277 */ 278void 279dmu_zfetch_init(zfetch_t *zf, dnode_t *dno) 280{ 281 if (zf == NULL) { 282 return; 283 } 284 285 zf->zf_dnode = dno; 286 zf->zf_stream_cnt = 0; 287 zf->zf_alloc_fail = 0; 288 289 list_create(&zf->zf_stream, sizeof (zstream_t), 290 offsetof(zstream_t, zst_node)); 291 292 rw_init(&zf->zf_rwlock, NULL, RW_DEFAULT, NULL); 293} 294 295/* 296 * This function computes the actual size, in blocks, that can be prefetched, 297 * and fetches it. 298 */ 299static uint64_t 300dmu_zfetch_fetch(dnode_t *dn, uint64_t blkid, uint64_t nblks) 301{ 302 uint64_t fetchsz; 303 uint64_t i; 304 305 fetchsz = dmu_zfetch_fetchsz(dn, blkid, nblks); 306 307 for (i = 0; i < fetchsz; i++) { 308 dbuf_prefetch(dn, blkid + i); 309 } 310 311 return (fetchsz); 312} 313 314/* 315 * this function returns the number of blocks that would be prefetched, based 316 * upon the supplied dnode, blockid, and nblks. This is used so that we can 317 * update streams in place, and then prefetch with their old value after the 318 * fact. This way, we can delay the prefetch, but subsequent accesses to the 319 * stream won't result in the same data being prefetched multiple times. 320 */ 321static uint64_t 322dmu_zfetch_fetchsz(dnode_t *dn, uint64_t blkid, uint64_t nblks) 323{ 324 uint64_t fetchsz; 325 326 if (blkid > dn->dn_maxblkid) { 327 return (0); 328 } 329 330 /* compute fetch size */ 331 if (blkid + nblks + 1 > dn->dn_maxblkid) { 332 fetchsz = (dn->dn_maxblkid - blkid) + 1; 333 ASSERT(blkid + fetchsz - 1 <= dn->dn_maxblkid); 334 } else { 335 fetchsz = nblks; 336 } 337 338 339 return (fetchsz); 340} 341 342/* 343 * given a zfetch and a zstream structure, see if there is an associated zstream 344 * for this block read. If so, it starts a prefetch for the stream it 345 * located and returns true, otherwise it returns false 346 */ 347static boolean_t 348dmu_zfetch_find(zfetch_t *zf, zstream_t *zh, int prefetched) 349{ 350 zstream_t *zs; 351 int64_t diff; 352 int reset = !prefetched; 353 int rc = 0; 354 355 if (zh == NULL) 356 return (0); 357 358 /* 359 * XXX: This locking strategy is a bit coarse; however, it's impact has 360 * yet to be tested. If this turns out to be an issue, it can be 361 * modified in a number of different ways. 362 */ 363 364 rw_enter(&zf->zf_rwlock, RW_READER); 365top: 366 367 for (zs = list_head(&zf->zf_stream); zs; 368 zs = list_next(&zf->zf_stream, zs)) { 369 370 /* 371 * XXX - should this be an assert? 372 */ 373 if (zs->zst_len == 0) { 374 /* bogus stream */ 375 ZFETCHSTAT_BUMP(zfetchstat_bogus_streams); 376 continue; 377 } 378 379 /* 380 * We hit this case when we are in a strided prefetch stream: 381 * we will read "len" blocks before "striding". 382 */ 383 if (zh->zst_offset >= zs->zst_offset && 384 zh->zst_offset < zs->zst_offset + zs->zst_len) { 385 if (prefetched) { 386 /* already fetched */ 387 ZFETCHSTAT_BUMP(zfetchstat_stride_hits); 388 rc = 1; 389 goto out; 390 } else { 391 ZFETCHSTAT_BUMP(zfetchstat_stride_misses); 392 } 393 } 394 395 /* 396 * This is the forward sequential read case: we increment 397 * len by one each time we hit here, so we will enter this 398 * case on every read. 399 */ 400 if (zh->zst_offset == zs->zst_offset + zs->zst_len) { 401 402 reset = !prefetched && zs->zst_len > 1; 403 404 if (mutex_tryenter(&zs->zst_lock) == 0) { 405 rc = 1; 406 goto out; 407 } 408 409 if (zh->zst_offset != zs->zst_offset + zs->zst_len) { 410 mutex_exit(&zs->zst_lock); 411 goto top; 412 } 413 zs->zst_len += zh->zst_len; 414 diff = zs->zst_len - zfetch_block_cap; 415 if (diff > 0) { 416 zs->zst_offset += diff; 417 zs->zst_len = zs->zst_len > diff ? 418 zs->zst_len - diff : 0; 419 } 420 zs->zst_direction = ZFETCH_FORWARD; 421 422 break; 423 424 /* 425 * Same as above, but reading backwards through the file. 426 */ 427 } else if (zh->zst_offset == zs->zst_offset - zh->zst_len) { 428 /* backwards sequential access */ 429 430 reset = !prefetched && zs->zst_len > 1; 431 432 if (mutex_tryenter(&zs->zst_lock) == 0) { 433 rc = 1; 434 goto out; 435 } 436 437 if (zh->zst_offset != zs->zst_offset - zh->zst_len) { 438 mutex_exit(&zs->zst_lock); 439 goto top; 440 } 441 442 zs->zst_offset = zs->zst_offset > zh->zst_len ? 443 zs->zst_offset - zh->zst_len : 0; 444 zs->zst_ph_offset = zs->zst_ph_offset > zh->zst_len ? 445 zs->zst_ph_offset - zh->zst_len : 0; 446 zs->zst_len += zh->zst_len; 447 448 diff = zs->zst_len - zfetch_block_cap; 449 if (diff > 0) { 450 zs->zst_ph_offset = zs->zst_ph_offset > diff ? 451 zs->zst_ph_offset - diff : 0; 452 zs->zst_len = zs->zst_len > diff ? 453 zs->zst_len - diff : zs->zst_len; 454 } 455 zs->zst_direction = ZFETCH_BACKWARD; 456 457 break; 458 459 } else if ((zh->zst_offset - zs->zst_offset - zs->zst_stride < 460 zs->zst_len) && (zs->zst_len != zs->zst_stride)) { 461 /* strided forward access */ 462 463 if (mutex_tryenter(&zs->zst_lock) == 0) { 464 rc = 1; 465 goto out; 466 } 467 468 if ((zh->zst_offset - zs->zst_offset - zs->zst_stride >= 469 zs->zst_len) || (zs->zst_len == zs->zst_stride)) { 470 mutex_exit(&zs->zst_lock); 471 goto top; 472 } 473 474 zs->zst_offset += zs->zst_stride; 475 zs->zst_direction = ZFETCH_FORWARD; 476 477 break; 478 479 } else if ((zh->zst_offset - zs->zst_offset + zs->zst_stride < 480 zs->zst_len) && (zs->zst_len != zs->zst_stride)) { 481 /* strided reverse access */ 482 483 if (mutex_tryenter(&zs->zst_lock) == 0) { 484 rc = 1; 485 goto out; 486 } 487 488 if ((zh->zst_offset - zs->zst_offset + zs->zst_stride >= 489 zs->zst_len) || (zs->zst_len == zs->zst_stride)) { 490 mutex_exit(&zs->zst_lock); 491 goto top; 492 } 493 494 zs->zst_offset = zs->zst_offset > zs->zst_stride ? 495 zs->zst_offset - zs->zst_stride : 0; 496 zs->zst_ph_offset = (zs->zst_ph_offset > 497 (2 * zs->zst_stride)) ? 498 (zs->zst_ph_offset - (2 * zs->zst_stride)) : 0; 499 zs->zst_direction = ZFETCH_BACKWARD; 500 501 break; 502 } 503 } 504 505 if (zs) { 506 if (reset) { 507 zstream_t *remove = zs; 508 509 ZFETCHSTAT_BUMP(zfetchstat_stream_resets); 510 rc = 0; 511 mutex_exit(&zs->zst_lock); 512 rw_exit(&zf->zf_rwlock); 513 rw_enter(&zf->zf_rwlock, RW_WRITER); 514 /* 515 * Relocate the stream, in case someone removes 516 * it while we were acquiring the WRITER lock. 517 */ 518 for (zs = list_head(&zf->zf_stream); zs; 519 zs = list_next(&zf->zf_stream, zs)) { 520 if (zs == remove) { 521 dmu_zfetch_stream_remove(zf, zs); 522 mutex_destroy(&zs->zst_lock); 523 kmem_free(zs, sizeof (zstream_t)); 524 break; 525 } 526 } 527 } else { 528 ZFETCHSTAT_BUMP(zfetchstat_stream_noresets); 529 rc = 1; 530 dmu_zfetch_dofetch(zf, zs); 531 mutex_exit(&zs->zst_lock); 532 } 533 } 534out: 535 rw_exit(&zf->zf_rwlock); 536 return (rc); 537} 538 539/* 540 * Clean-up state associated with a zfetch structure. This frees allocated 541 * structure members, empties the zf_stream tree, and generally makes things 542 * nice. This doesn't free the zfetch_t itself, that's left to the caller. 543 */ 544void 545dmu_zfetch_rele(zfetch_t *zf) 546{ 547 zstream_t *zs; 548 zstream_t *zs_next; 549 550 ASSERT(!RW_LOCK_HELD(&zf->zf_rwlock)); 551 552 for (zs = list_head(&zf->zf_stream); zs; zs = zs_next) { 553 zs_next = list_next(&zf->zf_stream, zs); 554 555 list_remove(&zf->zf_stream, zs); 556 mutex_destroy(&zs->zst_lock); 557 kmem_free(zs, sizeof (zstream_t)); 558 } 559 list_destroy(&zf->zf_stream); 560 rw_destroy(&zf->zf_rwlock); 561 562 zf->zf_dnode = NULL; 563} 564 565/* 566 * Given a zfetch and zstream structure, insert the zstream structure into the 567 * AVL tree contained within the zfetch structure. Peform the appropriate 568 * book-keeping. It is possible that another thread has inserted a stream which 569 * matches one that we are about to insert, so we must be sure to check for this 570 * case. If one is found, return failure, and let the caller cleanup the 571 * duplicates. 572 */ 573static int 574dmu_zfetch_stream_insert(zfetch_t *zf, zstream_t *zs) 575{ 576 zstream_t *zs_walk; 577 zstream_t *zs_next; 578 579 ASSERT(RW_WRITE_HELD(&zf->zf_rwlock)); 580 581 for (zs_walk = list_head(&zf->zf_stream); zs_walk; zs_walk = zs_next) { 582 zs_next = list_next(&zf->zf_stream, zs_walk); 583 584 if (dmu_zfetch_streams_equal(zs_walk, zs)) { 585 return (0); 586 } 587 } 588 589 list_insert_head(&zf->zf_stream, zs); 590 zf->zf_stream_cnt++; 591 return (1); 592} 593 594 595/* 596 * Walk the list of zstreams in the given zfetch, find an old one (by time), and 597 * reclaim it for use by the caller. 598 */ 599static zstream_t * 600dmu_zfetch_stream_reclaim(zfetch_t *zf) 601{ 602 zstream_t *zs; 603 604 if (! rw_tryenter(&zf->zf_rwlock, RW_WRITER)) 605 return (0); 606 607 for (zs = list_head(&zf->zf_stream); zs; 608 zs = list_next(&zf->zf_stream, zs)) { 609 610 if (((ddi_get_lbolt() - zs->zst_last)/hz) > zfetch_min_sec_reap) 611 break; 612 } 613 614 if (zs) { 615 dmu_zfetch_stream_remove(zf, zs); 616 mutex_destroy(&zs->zst_lock); 617 bzero(zs, sizeof (zstream_t)); 618 } else { 619 zf->zf_alloc_fail++; 620 } 621 rw_exit(&zf->zf_rwlock); 622 623 return (zs); 624} 625 626/* 627 * Given a zfetch and zstream structure, remove the zstream structure from its 628 * container in the zfetch structure. Perform the appropriate book-keeping. 629 */ 630static void 631dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs) 632{ 633 ASSERT(RW_WRITE_HELD(&zf->zf_rwlock)); 634 635 list_remove(&zf->zf_stream, zs); 636 zf->zf_stream_cnt--; 637} 638 639static int 640dmu_zfetch_streams_equal(zstream_t *zs1, zstream_t *zs2) 641{ 642 if (zs1->zst_offset != zs2->zst_offset) 643 return (0); 644 645 if (zs1->zst_len != zs2->zst_len) 646 return (0); 647 648 if (zs1->zst_stride != zs2->zst_stride) 649 return (0); 650 651 if (zs1->zst_ph_offset != zs2->zst_ph_offset) 652 return (0); 653 654 if (zs1->zst_cap != zs2->zst_cap) 655 return (0); 656 657 if (zs1->zst_direction != zs2->zst_direction) 658 return (0); 659 660 return (1); 661} 662 663/* 664 * This is the prefetch entry point. It calls all of the other dmu_zfetch 665 * routines to create, delete, find, or operate upon prefetch streams. 666 */ 667void 668dmu_zfetch(zfetch_t *zf, uint64_t offset, uint64_t size, int prefetched) 669{ 670 zstream_t zst; 671 zstream_t *newstream; 672 boolean_t fetched; 673 int inserted; 674 unsigned int blkshft; 675 uint64_t blksz; 676 677 if (zfs_prefetch_disable) 678 return; 679 680 /* files that aren't ln2 blocksz are only one block -- nothing to do */ 681 if (!zf->zf_dnode->dn_datablkshift) 682 return; 683 684 /* convert offset and size, into blockid and nblocks */ 685 blkshft = zf->zf_dnode->dn_datablkshift; 686 blksz = (1 << blkshft); 687 688 bzero(&zst, sizeof (zstream_t)); 689 zst.zst_offset = offset >> blkshft; 690 zst.zst_len = (P2ROUNDUP(offset + size, blksz) - 691 P2ALIGN(offset, blksz)) >> blkshft; 692 693 fetched = dmu_zfetch_find(zf, &zst, prefetched); 694 if (fetched) { 695 ZFETCHSTAT_BUMP(zfetchstat_hits); 696 } else { 697 ZFETCHSTAT_BUMP(zfetchstat_misses); 698 fetched = dmu_zfetch_colinear(zf, &zst); 699 if (fetched) { 700 ZFETCHSTAT_BUMP(zfetchstat_colinear_hits); 701 } else { 702 ZFETCHSTAT_BUMP(zfetchstat_colinear_misses); 703 } 704 } 705 706 if (!fetched) { 707 newstream = dmu_zfetch_stream_reclaim(zf); 708 709 /* 710 * we still couldn't find a stream, drop the lock, and allocate 711 * one if possible. Otherwise, give up and go home. 712 */ 713 if (newstream) { 714 ZFETCHSTAT_BUMP(zfetchstat_reclaim_successes); 715 } else { 716 uint64_t maxblocks; 717 uint32_t max_streams; 718 uint32_t cur_streams; 719 720 ZFETCHSTAT_BUMP(zfetchstat_reclaim_failures); 721 cur_streams = zf->zf_stream_cnt; 722 maxblocks = zf->zf_dnode->dn_maxblkid; 723 724 max_streams = MIN(zfetch_max_streams, 725 (maxblocks / zfetch_block_cap)); 726 if (max_streams == 0) { 727 max_streams++; 728 } 729 730 if (cur_streams >= max_streams) { 731 return; 732 } 733 newstream = kmem_zalloc(sizeof (zstream_t), KM_SLEEP); 734 } 735 736 newstream->zst_offset = zst.zst_offset; 737 newstream->zst_len = zst.zst_len; 738 newstream->zst_stride = zst.zst_len; 739 newstream->zst_ph_offset = zst.zst_len + zst.zst_offset; 740 newstream->zst_cap = zst.zst_len; 741 newstream->zst_direction = ZFETCH_FORWARD; 742 newstream->zst_last = ddi_get_lbolt(); 743 744 mutex_init(&newstream->zst_lock, NULL, MUTEX_DEFAULT, NULL); 745 746 rw_enter(&zf->zf_rwlock, RW_WRITER); 747 inserted = dmu_zfetch_stream_insert(zf, newstream); 748 rw_exit(&zf->zf_rwlock); 749 750 if (!inserted) { 751 mutex_destroy(&newstream->zst_lock); 752 kmem_free(newstream, sizeof (zstream_t)); 753 } 754 } 755} 756