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 2006 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26 #pragma ident "%Z%%M% %I% %E% SMI" 27 28 #include <sys/zfs_context.h> 29 #include <sys/dnode.h> 30 #include <sys/dmu_objset.h> 31 #include <sys/dmu_zfetch.h> 32 #include <sys/dmu.h> 33 #include <sys/dbuf.h> 34 35 /* 36 * I'm against tune-ables, but these should probably exist as tweakable globals 37 * until we can get this working the way we want it to. 38 */ 39 40 int zfs_prefetch_disable = 0; 41 42 /* max # of streams per zfetch */ 43 uint32_t zfetch_max_streams = 8; 44 /* min time before stream reclaim */ 45 uint32_t zfetch_min_sec_reap = 2; 46 /* max number of blocks to fetch at a time */ 47 uint32_t zfetch_block_cap = 256; 48 /* number of bytes in a array_read at which we stop prefetching (1Mb) */ 49 uint64_t zfetch_array_rd_sz = 1024 * 1024; 50 51 /* forward decls for static routines */ 52 static int dmu_zfetch_colinear(zfetch_t *, zstream_t *); 53 static void dmu_zfetch_dofetch(zfetch_t *, zstream_t *); 54 static uint64_t dmu_zfetch_fetch(dnode_t *, uint64_t, uint64_t); 55 static uint64_t dmu_zfetch_fetchsz(dnode_t *, uint64_t, uint64_t); 56 static int dmu_zfetch_find(zfetch_t *, zstream_t *, int); 57 static int dmu_zfetch_stream_insert(zfetch_t *, zstream_t *); 58 static zstream_t *dmu_zfetch_stream_reclaim(zfetch_t *); 59 static void dmu_zfetch_stream_remove(zfetch_t *, zstream_t *); 60 static int dmu_zfetch_streams_equal(zstream_t *, zstream_t *); 61 62 /* 63 * Given a zfetch structure and a zstream structure, determine whether the 64 * blocks to be read are part of a co-linear pair of existing prefetch 65 * streams. If a set is found, coalesce the streams, removing one, and 66 * configure the prefetch so it looks for a strided access pattern. 67 * 68 * In other words: if we find two sequential access streams that are 69 * the same length and distance N appart, and this read is N from the 70 * last stream, then we are probably in a strided access pattern. So 71 * combine the two sequential streams into a single strided stream. 72 * 73 * If no co-linear streams are found, return NULL. 74 */ 75 static int 76 dmu_zfetch_colinear(zfetch_t *zf, zstream_t *zh) 77 { 78 zstream_t *z_walk; 79 zstream_t *z_comp; 80 81 if (! rw_tryenter(&zf->zf_rwlock, RW_WRITER)) 82 return (0); 83 84 if (zh == NULL) { 85 rw_exit(&zf->zf_rwlock); 86 return (0); 87 } 88 89 for (z_walk = list_head(&zf->zf_stream); z_walk; 90 z_walk = list_next(&zf->zf_stream, z_walk)) { 91 for (z_comp = list_next(&zf->zf_stream, z_walk); z_comp; 92 z_comp = list_next(&zf->zf_stream, z_comp)) { 93 int64_t diff; 94 95 if (z_walk->zst_len != z_walk->zst_stride || 96 z_comp->zst_len != z_comp->zst_stride) { 97 continue; 98 } 99 100 diff = z_comp->zst_offset - z_walk->zst_offset; 101 if (z_comp->zst_offset + diff == zh->zst_offset) { 102 z_walk->zst_offset = zh->zst_offset; 103 z_walk->zst_direction = diff < 0 ? -1 : 1; 104 z_walk->zst_stride = 105 diff * z_walk->zst_direction; 106 z_walk->zst_ph_offset = 107 zh->zst_offset + z_walk->zst_stride; 108 dmu_zfetch_stream_remove(zf, z_comp); 109 mutex_destroy(&z_comp->zst_lock); 110 kmem_free(z_comp, sizeof (zstream_t)); 111 112 dmu_zfetch_dofetch(zf, z_walk); 113 114 rw_exit(&zf->zf_rwlock); 115 return (1); 116 } 117 118 diff = z_walk->zst_offset - z_comp->zst_offset; 119 if (z_walk->zst_offset + diff == zh->zst_offset) { 120 z_walk->zst_offset = zh->zst_offset; 121 z_walk->zst_direction = diff < 0 ? -1 : 1; 122 z_walk->zst_stride = 123 diff * z_walk->zst_direction; 124 z_walk->zst_ph_offset = 125 zh->zst_offset + z_walk->zst_stride; 126 dmu_zfetch_stream_remove(zf, z_comp); 127 mutex_destroy(&z_comp->zst_lock); 128 kmem_free(z_comp, sizeof (zstream_t)); 129 130 dmu_zfetch_dofetch(zf, z_walk); 131 132 rw_exit(&zf->zf_rwlock); 133 return (1); 134 } 135 } 136 } 137 138 rw_exit(&zf->zf_rwlock); 139 return (0); 140 } 141 142 /* 143 * Given a zstream_t, determine the bounds of the prefetch. Then call the 144 * routine that actually prefetches the individual blocks. 145 */ 146 static void 147 dmu_zfetch_dofetch(zfetch_t *zf, zstream_t *zs) 148 { 149 uint64_t prefetch_tail; 150 uint64_t prefetch_limit; 151 uint64_t prefetch_ofst; 152 uint64_t prefetch_len; 153 uint64_t blocks_fetched; 154 155 zs->zst_stride = MAX((int64_t)zs->zst_stride, zs->zst_len); 156 zs->zst_cap = MIN(zfetch_block_cap, 2 * zs->zst_cap); 157 158 prefetch_tail = MAX((int64_t)zs->zst_ph_offset, 159 (int64_t)(zs->zst_offset + zs->zst_stride)); 160 /* 161 * XXX: use a faster division method? 162 */ 163 prefetch_limit = zs->zst_offset + zs->zst_len + 164 (zs->zst_cap * zs->zst_stride) / zs->zst_len; 165 166 while (prefetch_tail < prefetch_limit) { 167 prefetch_ofst = zs->zst_offset + zs->zst_direction * 168 (prefetch_tail - zs->zst_offset); 169 170 prefetch_len = zs->zst_len; 171 172 /* 173 * Don't prefetch beyond the end of the file, if working 174 * backwards. 175 */ 176 if ((zs->zst_direction == ZFETCH_BACKWARD) && 177 (prefetch_ofst > prefetch_tail)) { 178 prefetch_len += prefetch_ofst; 179 prefetch_ofst = 0; 180 } 181 182 /* don't prefetch more than we're supposed to */ 183 if (prefetch_len > zs->zst_len) 184 break; 185 186 blocks_fetched = dmu_zfetch_fetch(zf->zf_dnode, 187 prefetch_ofst, zs->zst_len); 188 189 prefetch_tail += zs->zst_stride; 190 /* stop if we've run out of stuff to prefetch */ 191 if (blocks_fetched < zs->zst_len) 192 break; 193 } 194 zs->zst_ph_offset = prefetch_tail; 195 zs->zst_last = lbolt; 196 } 197 198 /* 199 * This takes a pointer to a zfetch structure and a dnode. It performs the 200 * necessary setup for the zfetch structure, grokking data from the 201 * associated dnode. 202 */ 203 void 204 dmu_zfetch_init(zfetch_t *zf, dnode_t *dno) 205 { 206 if (zf == NULL) { 207 return; 208 } 209 210 zf->zf_dnode = dno; 211 zf->zf_stream_cnt = 0; 212 zf->zf_alloc_fail = 0; 213 214 list_create(&zf->zf_stream, sizeof (zstream_t), 215 offsetof(zstream_t, zst_node)); 216 217 rw_init(&zf->zf_rwlock, NULL, RW_DEFAULT, NULL); 218 } 219 220 /* 221 * This function computes the actual size, in blocks, that can be prefetched, 222 * and fetches it. 223 */ 224 static uint64_t 225 dmu_zfetch_fetch(dnode_t *dn, uint64_t blkid, uint64_t nblks) 226 { 227 uint64_t fetchsz; 228 uint64_t i; 229 230 fetchsz = dmu_zfetch_fetchsz(dn, blkid, nblks); 231 232 for (i = 0; i < fetchsz; i++) { 233 dbuf_prefetch(dn, blkid + i); 234 } 235 236 return (fetchsz); 237 } 238 239 /* 240 * this function returns the number of blocks that would be prefetched, based 241 * upon the supplied dnode, blockid, and nblks. This is used so that we can 242 * update streams in place, and then prefetch with their old value after the 243 * fact. This way, we can delay the prefetch, but subsequent accesses to the 244 * stream won't result in the same data being prefetched multiple times. 245 */ 246 static uint64_t 247 dmu_zfetch_fetchsz(dnode_t *dn, uint64_t blkid, uint64_t nblks) 248 { 249 uint64_t fetchsz; 250 251 if (blkid > dn->dn_maxblkid) { 252 return (0); 253 } 254 255 /* compute fetch size */ 256 if (blkid + nblks + 1 > dn->dn_maxblkid) { 257 fetchsz = (dn->dn_maxblkid - blkid) + 1; 258 ASSERT(blkid + fetchsz - 1 <= dn->dn_maxblkid); 259 } else { 260 fetchsz = nblks; 261 } 262 263 264 return (fetchsz); 265 } 266 267 /* 268 * given a zfetch and a zsearch structure, see if there is an associated zstream 269 * for this block read. If so, it starts a prefetch for the stream it 270 * located and returns true, otherwise it returns false 271 */ 272 static int 273 dmu_zfetch_find(zfetch_t *zf, zstream_t *zh, int prefetched) 274 { 275 zstream_t *zs; 276 int64_t diff; 277 int reset = !prefetched; 278 int rc = 0; 279 280 if (zh == NULL) 281 return (0); 282 283 /* 284 * XXX: This locking strategy is a bit coarse; however, it's impact has 285 * yet to be tested. If this turns out to be an issue, it can be 286 * modified in a number of different ways. 287 */ 288 289 rw_enter(&zf->zf_rwlock, RW_READER); 290 top: 291 292 for (zs = list_head(&zf->zf_stream); zs; 293 zs = list_next(&zf->zf_stream, zs)) { 294 295 /* 296 * XXX - should this be an assert? 297 */ 298 if (zs->zst_len == 0) { 299 /* bogus stream */ 300 continue; 301 } 302 303 /* 304 * We hit this case when we are in a strided prefetch stream: 305 * we will read "len" blocks before "striding". 306 */ 307 if (zh->zst_offset >= zs->zst_offset && 308 zh->zst_offset < zs->zst_offset + zs->zst_len) { 309 /* already fetched */ 310 rc = 1; 311 goto out; 312 } 313 314 /* 315 * This is the forward sequential read case: we increment 316 * len by one each time we hit here, so we will enter this 317 * case on every read. 318 */ 319 if (zh->zst_offset == zs->zst_offset + zs->zst_len) { 320 321 reset = !prefetched && zs->zst_len > 1; 322 323 mutex_enter(&zs->zst_lock); 324 325 if (zh->zst_offset != zs->zst_offset + zs->zst_len) { 326 mutex_exit(&zs->zst_lock); 327 goto top; 328 } 329 zs->zst_len += zh->zst_len; 330 diff = zs->zst_len - zfetch_block_cap; 331 if (diff > 0) { 332 zs->zst_offset += diff; 333 zs->zst_len = zs->zst_len > diff ? 334 zs->zst_len - diff : 0; 335 } 336 zs->zst_direction = ZFETCH_FORWARD; 337 338 break; 339 340 /* 341 * Same as above, but reading backwards through the file. 342 */ 343 } else if (zh->zst_offset == zs->zst_offset - zh->zst_len) { 344 /* backwards sequential access */ 345 346 reset = !prefetched && zs->zst_len > 1; 347 348 mutex_enter(&zs->zst_lock); 349 350 if (zh->zst_offset != zs->zst_offset - zh->zst_len) { 351 mutex_exit(&zs->zst_lock); 352 goto top; 353 } 354 355 zs->zst_offset = zs->zst_offset > zh->zst_len ? 356 zs->zst_offset - zh->zst_len : 0; 357 zs->zst_ph_offset = zs->zst_ph_offset > zh->zst_len ? 358 zs->zst_ph_offset - zh->zst_len : 0; 359 zs->zst_len += zh->zst_len; 360 361 diff = zs->zst_len - zfetch_block_cap; 362 if (diff > 0) { 363 zs->zst_ph_offset = zs->zst_ph_offset > diff ? 364 zs->zst_ph_offset - diff : 0; 365 zs->zst_len = zs->zst_len > diff ? 366 zs->zst_len - diff : zs->zst_len; 367 } 368 zs->zst_direction = ZFETCH_BACKWARD; 369 370 break; 371 372 } else if ((zh->zst_offset - zs->zst_offset - zs->zst_stride < 373 zs->zst_len) && (zs->zst_len != zs->zst_stride)) { 374 /* strided forward access */ 375 376 mutex_enter(&zs->zst_lock); 377 378 if ((zh->zst_offset - zs->zst_offset - zs->zst_stride >= 379 zs->zst_len) || (zs->zst_len == zs->zst_stride)) { 380 mutex_exit(&zs->zst_lock); 381 goto top; 382 } 383 384 zs->zst_offset += zs->zst_stride; 385 zs->zst_direction = ZFETCH_FORWARD; 386 387 break; 388 389 } else if ((zh->zst_offset - zs->zst_offset + zs->zst_stride < 390 zs->zst_len) && (zs->zst_len != zs->zst_stride)) { 391 /* strided reverse access */ 392 393 mutex_enter(&zs->zst_lock); 394 395 if ((zh->zst_offset - zs->zst_offset + zs->zst_stride >= 396 zs->zst_len) || (zs->zst_len == zs->zst_stride)) { 397 mutex_exit(&zs->zst_lock); 398 goto top; 399 } 400 401 zs->zst_offset = zs->zst_offset > zs->zst_stride ? 402 zs->zst_offset - zs->zst_stride : 0; 403 zs->zst_ph_offset = (zs->zst_ph_offset > 404 (2 * zs->zst_stride)) ? 405 (zs->zst_ph_offset - (2 * zs->zst_stride)) : 0; 406 zs->zst_direction = ZFETCH_BACKWARD; 407 408 break; 409 } 410 } 411 412 if (zs) { 413 if (reset) { 414 zstream_t *remove = zs; 415 416 rc = 0; 417 mutex_exit(&zs->zst_lock); 418 rw_exit(&zf->zf_rwlock); 419 rw_enter(&zf->zf_rwlock, RW_WRITER); 420 /* 421 * Relocate the stream, in case someone removes 422 * it while we were acquiring the WRITER lock. 423 */ 424 for (zs = list_head(&zf->zf_stream); zs; 425 zs = list_next(&zf->zf_stream, zs)) { 426 if (zs == remove) { 427 dmu_zfetch_stream_remove(zf, zs); 428 mutex_destroy(&zs->zst_lock); 429 kmem_free(zs, sizeof (zstream_t)); 430 break; 431 } 432 } 433 } else { 434 rc = 1; 435 dmu_zfetch_dofetch(zf, zs); 436 mutex_exit(&zs->zst_lock); 437 } 438 } 439 out: 440 rw_exit(&zf->zf_rwlock); 441 return (rc); 442 } 443 444 /* 445 * Clean-up state associated with a zfetch structure. This frees allocated 446 * structure members, empties the zf_stream tree, and generally makes things 447 * nice. This doesn't free the zfetch_t itself, that's left to the caller. 448 */ 449 void 450 dmu_zfetch_rele(zfetch_t *zf) 451 { 452 zstream_t *zs; 453 zstream_t *zs_next; 454 455 ASSERT(!RW_LOCK_HELD(&zf->zf_rwlock)); 456 457 for (zs = list_head(&zf->zf_stream); zs; zs = zs_next) { 458 zs_next = list_next(&zf->zf_stream, zs); 459 460 list_remove(&zf->zf_stream, zs); 461 mutex_destroy(&zs->zst_lock); 462 kmem_free(zs, sizeof (zstream_t)); 463 } 464 list_destroy(&zf->zf_stream); 465 rw_destroy(&zf->zf_rwlock); 466 467 zf->zf_dnode = NULL; 468 } 469 470 /* 471 * Given a zfetch and zstream structure, insert the zstream structure into the 472 * AVL tree contained within the zfetch structure. Peform the appropriate 473 * book-keeping. It is possible that another thread has inserted a stream which 474 * matches one that we are about to insert, so we must be sure to check for this 475 * case. If one is found, return failure, and let the caller cleanup the 476 * duplicates. 477 */ 478 static int 479 dmu_zfetch_stream_insert(zfetch_t *zf, zstream_t *zs) 480 { 481 zstream_t *zs_walk; 482 zstream_t *zs_next; 483 484 ASSERT(RW_WRITE_HELD(&zf->zf_rwlock)); 485 486 for (zs_walk = list_head(&zf->zf_stream); zs_walk; zs_walk = zs_next) { 487 zs_next = list_next(&zf->zf_stream, zs_walk); 488 489 if (dmu_zfetch_streams_equal(zs_walk, zs)) { 490 return (0); 491 } 492 } 493 494 list_insert_head(&zf->zf_stream, zs); 495 zf->zf_stream_cnt++; 496 497 return (1); 498 } 499 500 501 /* 502 * Walk the list of zstreams in the given zfetch, find an old one (by time), and 503 * reclaim it for use by the caller. 504 */ 505 static zstream_t * 506 dmu_zfetch_stream_reclaim(zfetch_t *zf) 507 { 508 zstream_t *zs; 509 510 if (! rw_tryenter(&zf->zf_rwlock, RW_WRITER)) 511 return (0); 512 513 for (zs = list_head(&zf->zf_stream); zs; 514 zs = list_next(&zf->zf_stream, zs)) { 515 516 if (((lbolt - zs->zst_last) / hz) > zfetch_min_sec_reap) 517 break; 518 } 519 520 if (zs) { 521 dmu_zfetch_stream_remove(zf, zs); 522 mutex_destroy(&zs->zst_lock); 523 bzero(zs, sizeof (zstream_t)); 524 } else { 525 zf->zf_alloc_fail++; 526 } 527 rw_exit(&zf->zf_rwlock); 528 529 return (zs); 530 } 531 532 /* 533 * Given a zfetch and zstream structure, remove the zstream structure from its 534 * container in the zfetch structure. Perform the appropriate book-keeping. 535 */ 536 static void 537 dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs) 538 { 539 ASSERT(RW_WRITE_HELD(&zf->zf_rwlock)); 540 541 list_remove(&zf->zf_stream, zs); 542 zf->zf_stream_cnt--; 543 } 544 545 static int 546 dmu_zfetch_streams_equal(zstream_t *zs1, zstream_t *zs2) 547 { 548 if (zs1->zst_offset != zs2->zst_offset) 549 return (0); 550 551 if (zs1->zst_len != zs2->zst_len) 552 return (0); 553 554 if (zs1->zst_stride != zs2->zst_stride) 555 return (0); 556 557 if (zs1->zst_ph_offset != zs2->zst_ph_offset) 558 return (0); 559 560 if (zs1->zst_cap != zs2->zst_cap) 561 return (0); 562 563 if (zs1->zst_direction != zs2->zst_direction) 564 return (0); 565 566 return (1); 567 } 568 569 /* 570 * This is the prefetch entry point. It calls all of the other dmu_zfetch 571 * routines to create, delete, find, or operate upon prefetch streams. 572 */ 573 void 574 dmu_zfetch(zfetch_t *zf, uint64_t offset, uint64_t size, int prefetched) 575 { 576 zstream_t zst; 577 zstream_t *newstream; 578 int fetched; 579 int inserted; 580 unsigned int blkshft; 581 uint64_t blksz; 582 583 if (zfs_prefetch_disable) 584 return; 585 586 /* files that aren't ln2 blocksz are only one block -- nothing to do */ 587 if (!zf->zf_dnode->dn_datablkshift) 588 return; 589 590 /* convert offset and size, into blockid and nblocks */ 591 blkshft = zf->zf_dnode->dn_datablkshift; 592 blksz = (1 << blkshft); 593 594 bzero(&zst, sizeof (zstream_t)); 595 zst.zst_offset = offset >> blkshft; 596 zst.zst_len = (P2ROUNDUP(offset + size, blksz) - 597 P2ALIGN(offset, blksz)) >> blkshft; 598 599 fetched = dmu_zfetch_find(zf, &zst, prefetched); 600 if (!fetched) { 601 fetched = dmu_zfetch_colinear(zf, &zst); 602 } 603 604 if (!fetched) { 605 newstream = dmu_zfetch_stream_reclaim(zf); 606 607 /* 608 * we still couldn't find a stream, drop the lock, and allocate 609 * one if possible. Otherwise, give up and go home. 610 */ 611 if (newstream == NULL) { 612 uint64_t maxblocks; 613 uint32_t max_streams; 614 uint32_t cur_streams; 615 616 cur_streams = zf->zf_stream_cnt; 617 maxblocks = zf->zf_dnode->dn_maxblkid; 618 619 max_streams = MIN(zfetch_max_streams, 620 (maxblocks / zfetch_block_cap)); 621 if (max_streams == 0) { 622 max_streams++; 623 } 624 625 if (cur_streams >= max_streams) { 626 return; 627 } 628 629 newstream = kmem_zalloc(sizeof (zstream_t), KM_SLEEP); 630 } 631 632 newstream->zst_offset = zst.zst_offset; 633 newstream->zst_len = zst.zst_len; 634 newstream->zst_stride = zst.zst_len; 635 newstream->zst_ph_offset = zst.zst_len + zst.zst_offset; 636 newstream->zst_cap = zst.zst_len; 637 newstream->zst_direction = ZFETCH_FORWARD; 638 newstream->zst_last = lbolt; 639 640 mutex_init(&newstream->zst_lock, NULL, MUTEX_DEFAULT, NULL); 641 642 rw_enter(&zf->zf_rwlock, RW_WRITER); 643 inserted = dmu_zfetch_stream_insert(zf, newstream); 644 rw_exit(&zf->zf_rwlock); 645 646 if (!inserted) { 647 mutex_destroy(&newstream->zst_lock); 648 kmem_free(newstream, sizeof (zstream_t)); 649 } 650 } 651 } 652