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 /* 27 * Copyright (c) 2013, 2017 by Delphix. All rights reserved. 28 */ 29 30 #include <sys/zfs_context.h> 31 #include <sys/dnode.h> 32 #include <sys/dmu_objset.h> 33 #include <sys/dmu_zfetch.h> 34 #include <sys/dmu.h> 35 #include <sys/dbuf.h> 36 #include <sys/kstat.h> 37 #include <sys/wmsum.h> 38 39 /* 40 * This tunable disables predictive prefetch. Note that it leaves "prescient" 41 * prefetch (e.g. prefetch for zfs send) intact. Unlike predictive prefetch, 42 * prescient prefetch never issues i/os that end up not being needed, 43 * so it can't hurt performance. 44 */ 45 46 int zfs_prefetch_disable = B_FALSE; 47 48 /* max # of streams per zfetch */ 49 unsigned int zfetch_max_streams = 8; 50 /* min time before stream reclaim */ 51 unsigned int zfetch_min_sec_reap = 2; 52 /* max bytes to prefetch per stream (default 8MB) */ 53 unsigned int zfetch_max_distance = 8 * 1024 * 1024; 54 /* max bytes to prefetch indirects for per stream (default 64MB) */ 55 unsigned int zfetch_max_idistance = 64 * 1024 * 1024; 56 /* max number of bytes in an array_read in which we allow prefetching (1MB) */ 57 unsigned long zfetch_array_rd_sz = 1024 * 1024; 58 59 typedef struct zfetch_stats { 60 kstat_named_t zfetchstat_hits; 61 kstat_named_t zfetchstat_misses; 62 kstat_named_t zfetchstat_max_streams; 63 kstat_named_t zfetchstat_io_issued; 64 } zfetch_stats_t; 65 66 static zfetch_stats_t zfetch_stats = { 67 { "hits", KSTAT_DATA_UINT64 }, 68 { "misses", KSTAT_DATA_UINT64 }, 69 { "max_streams", KSTAT_DATA_UINT64 }, 70 { "io_issued", KSTAT_DATA_UINT64 }, 71 }; 72 73 struct { 74 wmsum_t zfetchstat_hits; 75 wmsum_t zfetchstat_misses; 76 wmsum_t zfetchstat_max_streams; 77 wmsum_t zfetchstat_io_issued; 78 } zfetch_sums; 79 80 #define ZFETCHSTAT_BUMP(stat) \ 81 wmsum_add(&zfetch_sums.stat, 1) 82 #define ZFETCHSTAT_ADD(stat, val) \ 83 wmsum_add(&zfetch_sums.stat, val) 84 85 86 kstat_t *zfetch_ksp; 87 88 static int 89 zfetch_kstats_update(kstat_t *ksp, int rw) 90 { 91 zfetch_stats_t *zs = ksp->ks_data; 92 93 if (rw == KSTAT_WRITE) 94 return (EACCES); 95 zs->zfetchstat_hits.value.ui64 = 96 wmsum_value(&zfetch_sums.zfetchstat_hits); 97 zs->zfetchstat_misses.value.ui64 = 98 wmsum_value(&zfetch_sums.zfetchstat_misses); 99 zs->zfetchstat_max_streams.value.ui64 = 100 wmsum_value(&zfetch_sums.zfetchstat_max_streams); 101 zs->zfetchstat_io_issued.value.ui64 = 102 wmsum_value(&zfetch_sums.zfetchstat_io_issued); 103 return (0); 104 } 105 106 void 107 zfetch_init(void) 108 { 109 wmsum_init(&zfetch_sums.zfetchstat_hits, 0); 110 wmsum_init(&zfetch_sums.zfetchstat_misses, 0); 111 wmsum_init(&zfetch_sums.zfetchstat_max_streams, 0); 112 wmsum_init(&zfetch_sums.zfetchstat_io_issued, 0); 113 114 zfetch_ksp = kstat_create("zfs", 0, "zfetchstats", "misc", 115 KSTAT_TYPE_NAMED, sizeof (zfetch_stats) / sizeof (kstat_named_t), 116 KSTAT_FLAG_VIRTUAL); 117 118 if (zfetch_ksp != NULL) { 119 zfetch_ksp->ks_data = &zfetch_stats; 120 zfetch_ksp->ks_update = zfetch_kstats_update; 121 kstat_install(zfetch_ksp); 122 } 123 } 124 125 void 126 zfetch_fini(void) 127 { 128 if (zfetch_ksp != NULL) { 129 kstat_delete(zfetch_ksp); 130 zfetch_ksp = NULL; 131 } 132 133 wmsum_fini(&zfetch_sums.zfetchstat_hits); 134 wmsum_fini(&zfetch_sums.zfetchstat_misses); 135 wmsum_fini(&zfetch_sums.zfetchstat_max_streams); 136 wmsum_fini(&zfetch_sums.zfetchstat_io_issued); 137 } 138 139 /* 140 * This takes a pointer to a zfetch structure and a dnode. It performs the 141 * necessary setup for the zfetch structure, grokking data from the 142 * associated dnode. 143 */ 144 void 145 dmu_zfetch_init(zfetch_t *zf, dnode_t *dno) 146 { 147 if (zf == NULL) 148 return; 149 zf->zf_dnode = dno; 150 zf->zf_numstreams = 0; 151 152 list_create(&zf->zf_stream, sizeof (zstream_t), 153 offsetof(zstream_t, zs_node)); 154 155 mutex_init(&zf->zf_lock, NULL, MUTEX_DEFAULT, NULL); 156 } 157 158 static void 159 dmu_zfetch_stream_fini(zstream_t *zs) 160 { 161 ASSERT(!list_link_active(&zs->zs_node)); 162 zfs_refcount_destroy(&zs->zs_callers); 163 zfs_refcount_destroy(&zs->zs_refs); 164 kmem_free(zs, sizeof (*zs)); 165 } 166 167 static void 168 dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs) 169 { 170 ASSERT(MUTEX_HELD(&zf->zf_lock)); 171 list_remove(&zf->zf_stream, zs); 172 zf->zf_numstreams--; 173 membar_producer(); 174 if (zfs_refcount_remove(&zs->zs_refs, NULL) == 0) 175 dmu_zfetch_stream_fini(zs); 176 } 177 178 /* 179 * Clean-up state associated with a zfetch structure (e.g. destroy the 180 * streams). This doesn't free the zfetch_t itself, that's left to the caller. 181 */ 182 void 183 dmu_zfetch_fini(zfetch_t *zf) 184 { 185 zstream_t *zs; 186 187 mutex_enter(&zf->zf_lock); 188 while ((zs = list_head(&zf->zf_stream)) != NULL) 189 dmu_zfetch_stream_remove(zf, zs); 190 mutex_exit(&zf->zf_lock); 191 list_destroy(&zf->zf_stream); 192 mutex_destroy(&zf->zf_lock); 193 194 zf->zf_dnode = NULL; 195 } 196 197 /* 198 * If there aren't too many streams already, create a new stream. 199 * The "blkid" argument is the next block that we expect this stream to access. 200 * While we're here, clean up old streams (which haven't been 201 * accessed for at least zfetch_min_sec_reap seconds). 202 */ 203 static void 204 dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid) 205 { 206 zstream_t *zs_next; 207 hrtime_t now = gethrtime(); 208 209 ASSERT(MUTEX_HELD(&zf->zf_lock)); 210 211 /* 212 * Clean up old streams. 213 */ 214 for (zstream_t *zs = list_head(&zf->zf_stream); 215 zs != NULL; zs = zs_next) { 216 zs_next = list_next(&zf->zf_stream, zs); 217 /* 218 * Skip if still active. 1 -- zf_stream reference. 219 */ 220 if (zfs_refcount_count(&zs->zs_refs) != 1) 221 continue; 222 if (((now - zs->zs_atime) / NANOSEC) > 223 zfetch_min_sec_reap) 224 dmu_zfetch_stream_remove(zf, zs); 225 } 226 227 /* 228 * The maximum number of streams is normally zfetch_max_streams, 229 * but for small files we lower it such that it's at least possible 230 * for all the streams to be non-overlapping. 231 * 232 * If we are already at the maximum number of streams for this file, 233 * even after removing old streams, then don't create this stream. 234 */ 235 uint32_t max_streams = MAX(1, MIN(zfetch_max_streams, 236 zf->zf_dnode->dn_maxblkid * zf->zf_dnode->dn_datablksz / 237 zfetch_max_distance)); 238 if (zf->zf_numstreams >= max_streams) { 239 ZFETCHSTAT_BUMP(zfetchstat_max_streams); 240 return; 241 } 242 243 zstream_t *zs = kmem_zalloc(sizeof (*zs), KM_SLEEP); 244 zs->zs_blkid = blkid; 245 zs->zs_pf_blkid1 = blkid; 246 zs->zs_pf_blkid = blkid; 247 zs->zs_ipf_blkid1 = blkid; 248 zs->zs_ipf_blkid = blkid; 249 zs->zs_atime = now; 250 zs->zs_fetch = zf; 251 zs->zs_missed = B_FALSE; 252 zfs_refcount_create(&zs->zs_callers); 253 zfs_refcount_create(&zs->zs_refs); 254 /* One reference for zf_stream. */ 255 zfs_refcount_add(&zs->zs_refs, NULL); 256 zf->zf_numstreams++; 257 list_insert_head(&zf->zf_stream, zs); 258 } 259 260 static void 261 dmu_zfetch_stream_done(void *arg, boolean_t io_issued) 262 { 263 zstream_t *zs = arg; 264 265 if (zfs_refcount_remove(&zs->zs_refs, NULL) == 0) 266 dmu_zfetch_stream_fini(zs); 267 } 268 269 /* 270 * This is the predictive prefetch entry point. dmu_zfetch_prepare() 271 * associates dnode access specified with blkid and nblks arguments with 272 * prefetch stream, predicts further accesses based on that stats and returns 273 * the stream pointer on success. That pointer must later be passed to 274 * dmu_zfetch_run() to initiate the speculative prefetch for the stream and 275 * release it. dmu_zfetch() is a wrapper for simple cases when window between 276 * prediction and prefetch initiation is not needed. 277 * fetch_data argument specifies whether actual data blocks should be fetched: 278 * FALSE -- prefetch only indirect blocks for predicted data blocks; 279 * TRUE -- prefetch predicted data blocks plus following indirect blocks. 280 */ 281 zstream_t * 282 dmu_zfetch_prepare(zfetch_t *zf, uint64_t blkid, uint64_t nblks, 283 boolean_t fetch_data, boolean_t have_lock) 284 { 285 zstream_t *zs; 286 int64_t pf_start, ipf_start; 287 int64_t pf_ahead_blks, max_blks; 288 int max_dist_blks, pf_nblks, ipf_nblks; 289 uint64_t end_of_access_blkid, maxblkid; 290 end_of_access_blkid = blkid + nblks; 291 spa_t *spa = zf->zf_dnode->dn_objset->os_spa; 292 293 if (zfs_prefetch_disable) 294 return (NULL); 295 /* 296 * If we haven't yet loaded the indirect vdevs' mappings, we 297 * can only read from blocks that we carefully ensure are on 298 * concrete vdevs (or previously-loaded indirect vdevs). So we 299 * can't allow the predictive prefetcher to attempt reads of other 300 * blocks (e.g. of the MOS's dnode object). 301 */ 302 if (!spa_indirect_vdevs_loaded(spa)) 303 return (NULL); 304 305 /* 306 * As a fast path for small (single-block) files, ignore access 307 * to the first block. 308 */ 309 if (!have_lock && blkid == 0) 310 return (NULL); 311 312 if (!have_lock) 313 rw_enter(&zf->zf_dnode->dn_struct_rwlock, RW_READER); 314 315 /* 316 * A fast path for small files for which no prefetch will 317 * happen. 318 */ 319 maxblkid = zf->zf_dnode->dn_maxblkid; 320 if (maxblkid < 2) { 321 if (!have_lock) 322 rw_exit(&zf->zf_dnode->dn_struct_rwlock); 323 return (NULL); 324 } 325 mutex_enter(&zf->zf_lock); 326 327 /* 328 * Find matching prefetch stream. Depending on whether the accesses 329 * are block-aligned, first block of the new access may either follow 330 * the last block of the previous access, or be equal to it. 331 */ 332 for (zs = list_head(&zf->zf_stream); zs != NULL; 333 zs = list_next(&zf->zf_stream, zs)) { 334 if (blkid == zs->zs_blkid) { 335 break; 336 } else if (blkid + 1 == zs->zs_blkid) { 337 blkid++; 338 nblks--; 339 break; 340 } 341 } 342 343 /* 344 * If the file is ending, remove the matching stream if found. 345 * If not found then it is too late to create a new one now. 346 */ 347 if (end_of_access_blkid >= maxblkid) { 348 if (zs != NULL) 349 dmu_zfetch_stream_remove(zf, zs); 350 mutex_exit(&zf->zf_lock); 351 if (!have_lock) 352 rw_exit(&zf->zf_dnode->dn_struct_rwlock); 353 return (NULL); 354 } 355 356 /* Exit if we already prefetched this block before. */ 357 if (nblks == 0) { 358 mutex_exit(&zf->zf_lock); 359 if (!have_lock) 360 rw_exit(&zf->zf_dnode->dn_struct_rwlock); 361 return (NULL); 362 } 363 364 if (zs == NULL) { 365 /* 366 * This access is not part of any existing stream. Create 367 * a new stream for it. 368 */ 369 dmu_zfetch_stream_create(zf, end_of_access_blkid); 370 mutex_exit(&zf->zf_lock); 371 if (!have_lock) 372 rw_exit(&zf->zf_dnode->dn_struct_rwlock); 373 ZFETCHSTAT_BUMP(zfetchstat_misses); 374 return (NULL); 375 } 376 377 /* 378 * This access was to a block that we issued a prefetch for on 379 * behalf of this stream. Issue further prefetches for this stream. 380 * 381 * Normally, we start prefetching where we stopped 382 * prefetching last (zs_pf_blkid). But when we get our first 383 * hit on this stream, zs_pf_blkid == zs_blkid, we don't 384 * want to prefetch the block we just accessed. In this case, 385 * start just after the block we just accessed. 386 */ 387 pf_start = MAX(zs->zs_pf_blkid, end_of_access_blkid); 388 if (zs->zs_pf_blkid1 < end_of_access_blkid) 389 zs->zs_pf_blkid1 = end_of_access_blkid; 390 if (zs->zs_ipf_blkid1 < end_of_access_blkid) 391 zs->zs_ipf_blkid1 = end_of_access_blkid; 392 393 /* 394 * Double our amount of prefetched data, but don't let the 395 * prefetch get further ahead than zfetch_max_distance. 396 */ 397 if (fetch_data) { 398 max_dist_blks = 399 zfetch_max_distance >> zf->zf_dnode->dn_datablkshift; 400 /* 401 * Previously, we were (zs_pf_blkid - blkid) ahead. We 402 * want to now be double that, so read that amount again, 403 * plus the amount we are catching up by (i.e. the amount 404 * read just now). 405 */ 406 pf_ahead_blks = zs->zs_pf_blkid - blkid + nblks; 407 max_blks = max_dist_blks - (pf_start - end_of_access_blkid); 408 pf_nblks = MIN(pf_ahead_blks, max_blks); 409 } else { 410 pf_nblks = 0; 411 } 412 413 zs->zs_pf_blkid = pf_start + pf_nblks; 414 415 /* 416 * Do the same for indirects, starting from where we stopped last, 417 * or where we will stop reading data blocks (and the indirects 418 * that point to them). 419 */ 420 ipf_start = MAX(zs->zs_ipf_blkid, zs->zs_pf_blkid); 421 max_dist_blks = zfetch_max_idistance >> zf->zf_dnode->dn_datablkshift; 422 /* 423 * We want to double our distance ahead of the data prefetch 424 * (or reader, if we are not prefetching data). Previously, we 425 * were (zs_ipf_blkid - blkid) ahead. To double that, we read 426 * that amount again, plus the amount we are catching up by 427 * (i.e. the amount read now + the amount of data prefetched now). 428 */ 429 pf_ahead_blks = zs->zs_ipf_blkid - blkid + nblks + pf_nblks; 430 max_blks = max_dist_blks - (ipf_start - zs->zs_pf_blkid); 431 ipf_nblks = MIN(pf_ahead_blks, max_blks); 432 zs->zs_ipf_blkid = ipf_start + ipf_nblks; 433 434 zs->zs_blkid = end_of_access_blkid; 435 /* Protect the stream from reclamation. */ 436 zs->zs_atime = gethrtime(); 437 zfs_refcount_add(&zs->zs_refs, NULL); 438 /* Count concurrent callers. */ 439 zfs_refcount_add(&zs->zs_callers, NULL); 440 mutex_exit(&zf->zf_lock); 441 442 if (!have_lock) 443 rw_exit(&zf->zf_dnode->dn_struct_rwlock); 444 445 ZFETCHSTAT_BUMP(zfetchstat_hits); 446 return (zs); 447 } 448 449 void 450 dmu_zfetch_run(zstream_t *zs, boolean_t missed, boolean_t have_lock) 451 { 452 zfetch_t *zf = zs->zs_fetch; 453 int64_t pf_start, pf_end, ipf_start, ipf_end; 454 int epbs, issued; 455 456 if (missed) 457 zs->zs_missed = missed; 458 459 /* 460 * Postpone the prefetch if there are more concurrent callers. 461 * It happens when multiple requests are waiting for the same 462 * indirect block. The last one will run the prefetch for all. 463 */ 464 if (zfs_refcount_remove(&zs->zs_callers, NULL) != 0) { 465 /* Drop reference taken in dmu_zfetch_prepare(). */ 466 if (zfs_refcount_remove(&zs->zs_refs, NULL) == 0) 467 dmu_zfetch_stream_fini(zs); 468 return; 469 } 470 471 mutex_enter(&zf->zf_lock); 472 if (zs->zs_missed) { 473 pf_start = zs->zs_pf_blkid1; 474 pf_end = zs->zs_pf_blkid1 = zs->zs_pf_blkid; 475 } else { 476 pf_start = pf_end = 0; 477 } 478 ipf_start = MAX(zs->zs_pf_blkid1, zs->zs_ipf_blkid1); 479 ipf_end = zs->zs_ipf_blkid1 = zs->zs_ipf_blkid; 480 mutex_exit(&zf->zf_lock); 481 ASSERT3S(pf_start, <=, pf_end); 482 ASSERT3S(ipf_start, <=, ipf_end); 483 484 epbs = zf->zf_dnode->dn_indblkshift - SPA_BLKPTRSHIFT; 485 ipf_start = P2ROUNDUP(ipf_start, 1 << epbs) >> epbs; 486 ipf_end = P2ROUNDUP(ipf_end, 1 << epbs) >> epbs; 487 ASSERT3S(ipf_start, <=, ipf_end); 488 issued = pf_end - pf_start + ipf_end - ipf_start; 489 if (issued > 1) { 490 /* More references on top of taken in dmu_zfetch_prepare(). */ 491 zfs_refcount_add_many(&zs->zs_refs, issued - 1, NULL); 492 } else if (issued == 0) { 493 /* Some other thread has done our work, so drop the ref. */ 494 if (zfs_refcount_remove(&zs->zs_refs, NULL) == 0) 495 dmu_zfetch_stream_fini(zs); 496 return; 497 } 498 499 if (!have_lock) 500 rw_enter(&zf->zf_dnode->dn_struct_rwlock, RW_READER); 501 502 issued = 0; 503 for (int64_t blk = pf_start; blk < pf_end; blk++) { 504 issued += dbuf_prefetch_impl(zf->zf_dnode, 0, blk, 505 ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH, 506 dmu_zfetch_stream_done, zs); 507 } 508 for (int64_t iblk = ipf_start; iblk < ipf_end; iblk++) { 509 issued += dbuf_prefetch_impl(zf->zf_dnode, 1, iblk, 510 ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH, 511 dmu_zfetch_stream_done, zs); 512 } 513 514 if (!have_lock) 515 rw_exit(&zf->zf_dnode->dn_struct_rwlock); 516 517 if (issued) 518 ZFETCHSTAT_ADD(zfetchstat_io_issued, issued); 519 } 520 521 void 522 dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data, 523 boolean_t missed, boolean_t have_lock) 524 { 525 zstream_t *zs; 526 527 zs = dmu_zfetch_prepare(zf, blkid, nblks, fetch_data, have_lock); 528 if (zs) 529 dmu_zfetch_run(zs, missed, have_lock); 530 } 531 532 /* BEGIN CSTYLED */ 533 ZFS_MODULE_PARAM(zfs_prefetch, zfs_prefetch_, disable, INT, ZMOD_RW, 534 "Disable all ZFS prefetching"); 535 536 ZFS_MODULE_PARAM(zfs_prefetch, zfetch_, max_streams, UINT, ZMOD_RW, 537 "Max number of streams per zfetch"); 538 539 ZFS_MODULE_PARAM(zfs_prefetch, zfetch_, min_sec_reap, UINT, ZMOD_RW, 540 "Min time before stream reclaim"); 541 542 ZFS_MODULE_PARAM(zfs_prefetch, zfetch_, max_distance, UINT, ZMOD_RW, 543 "Max bytes to prefetch per stream"); 544 545 ZFS_MODULE_PARAM(zfs_prefetch, zfetch_, max_idistance, UINT, ZMOD_RW, 546 "Max bytes to prefetch indirects for per stream"); 547 548 ZFS_MODULE_PARAM(zfs_prefetch, zfetch_, array_rd_sz, ULONG, ZMOD_RW, 549 "Number of bytes in a array_read"); 550 /* END CSTYLED */ 551