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 38 /* 39 * This tunable disables predictive prefetch. Note that it leaves "prescient" 40 * prefetch (e.g. prefetch for zfs send) intact. Unlike predictive prefetch, 41 * prescient prefetch never issues i/os that end up not being needed, 42 * so it can't hurt performance. 43 */ 44 45 int zfs_prefetch_disable = B_FALSE; 46 47 /* max # of streams per zfetch */ 48 unsigned int zfetch_max_streams = 8; 49 /* min time before stream reclaim */ 50 unsigned int zfetch_min_sec_reap = 2; 51 /* max bytes to prefetch per stream (default 8MB) */ 52 unsigned int zfetch_max_distance = 8 * 1024 * 1024; 53 /* max bytes to prefetch indirects for per stream (default 64MB) */ 54 unsigned int zfetch_max_idistance = 64 * 1024 * 1024; 55 /* max number of bytes in an array_read in which we allow prefetching (1MB) */ 56 unsigned long zfetch_array_rd_sz = 1024 * 1024; 57 58 typedef struct zfetch_stats { 59 kstat_named_t zfetchstat_hits; 60 kstat_named_t zfetchstat_misses; 61 kstat_named_t zfetchstat_max_streams; 62 } zfetch_stats_t; 63 64 static zfetch_stats_t zfetch_stats = { 65 { "hits", KSTAT_DATA_UINT64 }, 66 { "misses", KSTAT_DATA_UINT64 }, 67 { "max_streams", KSTAT_DATA_UINT64 }, 68 }; 69 70 #define ZFETCHSTAT_BUMP(stat) \ 71 atomic_inc_64(&zfetch_stats.stat.value.ui64); 72 73 kstat_t *zfetch_ksp; 74 75 void 76 zfetch_init(void) 77 { 78 zfetch_ksp = kstat_create("zfs", 0, "zfetchstats", "misc", 79 KSTAT_TYPE_NAMED, sizeof (zfetch_stats) / sizeof (kstat_named_t), 80 KSTAT_FLAG_VIRTUAL); 81 82 if (zfetch_ksp != NULL) { 83 zfetch_ksp->ks_data = &zfetch_stats; 84 kstat_install(zfetch_ksp); 85 } 86 } 87 88 void 89 zfetch_fini(void) 90 { 91 if (zfetch_ksp != NULL) { 92 kstat_delete(zfetch_ksp); 93 zfetch_ksp = NULL; 94 } 95 } 96 97 /* 98 * This takes a pointer to a zfetch structure and a dnode. It performs the 99 * necessary setup for the zfetch structure, grokking data from the 100 * associated dnode. 101 */ 102 void 103 dmu_zfetch_init(zfetch_t *zf, dnode_t *dno) 104 { 105 if (zf == NULL) 106 return; 107 108 zf->zf_dnode = dno; 109 110 list_create(&zf->zf_stream, sizeof (zstream_t), 111 offsetof(zstream_t, zs_node)); 112 113 mutex_init(&zf->zf_lock, NULL, MUTEX_DEFAULT, NULL); 114 } 115 116 static void 117 dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs) 118 { 119 ASSERT(MUTEX_HELD(&zf->zf_lock)); 120 list_remove(&zf->zf_stream, zs); 121 mutex_destroy(&zs->zs_lock); 122 kmem_free(zs, sizeof (*zs)); 123 } 124 125 /* 126 * Clean-up state associated with a zfetch structure (e.g. destroy the 127 * streams). This doesn't free the zfetch_t itself, that's left to the caller. 128 */ 129 void 130 dmu_zfetch_fini(zfetch_t *zf) 131 { 132 zstream_t *zs; 133 134 mutex_enter(&zf->zf_lock); 135 while ((zs = list_head(&zf->zf_stream)) != NULL) 136 dmu_zfetch_stream_remove(zf, zs); 137 mutex_exit(&zf->zf_lock); 138 list_destroy(&zf->zf_stream); 139 mutex_destroy(&zf->zf_lock); 140 141 zf->zf_dnode = NULL; 142 } 143 144 /* 145 * If there aren't too many streams already, create a new stream. 146 * The "blkid" argument is the next block that we expect this stream to access. 147 * While we're here, clean up old streams (which haven't been 148 * accessed for at least zfetch_min_sec_reap seconds). 149 */ 150 static void 151 dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid) 152 { 153 zstream_t *zs_next; 154 int numstreams = 0; 155 156 ASSERT(MUTEX_HELD(&zf->zf_lock)); 157 158 /* 159 * Clean up old streams. 160 */ 161 for (zstream_t *zs = list_head(&zf->zf_stream); 162 zs != NULL; zs = zs_next) { 163 zs_next = list_next(&zf->zf_stream, zs); 164 if (((gethrtime() - zs->zs_atime) / NANOSEC) > 165 zfetch_min_sec_reap) 166 dmu_zfetch_stream_remove(zf, zs); 167 else 168 numstreams++; 169 } 170 171 /* 172 * The maximum number of streams is normally zfetch_max_streams, 173 * but for small files we lower it such that it's at least possible 174 * for all the streams to be non-overlapping. 175 * 176 * If we are already at the maximum number of streams for this file, 177 * even after removing old streams, then don't create this stream. 178 */ 179 uint32_t max_streams = MAX(1, MIN(zfetch_max_streams, 180 zf->zf_dnode->dn_maxblkid * zf->zf_dnode->dn_datablksz / 181 zfetch_max_distance)); 182 if (numstreams >= max_streams) { 183 ZFETCHSTAT_BUMP(zfetchstat_max_streams); 184 return; 185 } 186 187 zstream_t *zs = kmem_zalloc(sizeof (*zs), KM_SLEEP); 188 zs->zs_blkid = blkid; 189 zs->zs_pf_blkid = blkid; 190 zs->zs_ipf_blkid = blkid; 191 zs->zs_atime = gethrtime(); 192 mutex_init(&zs->zs_lock, NULL, MUTEX_DEFAULT, NULL); 193 194 list_insert_head(&zf->zf_stream, zs); 195 } 196 197 /* 198 * This is the predictive prefetch entry point. It associates dnode access 199 * specified with blkid and nblks arguments with prefetch stream, predicts 200 * further accesses based on that stats and initiates speculative prefetch. 201 * fetch_data argument specifies whether actual data blocks should be fetched: 202 * FALSE -- prefetch only indirect blocks for predicted data blocks; 203 * TRUE -- prefetch predicted data blocks plus following indirect blocks. 204 */ 205 void 206 dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data, 207 boolean_t have_lock) 208 { 209 zstream_t *zs; 210 int64_t pf_start, ipf_start, ipf_istart, ipf_iend; 211 int64_t pf_ahead_blks, max_blks; 212 int epbs, max_dist_blks, pf_nblks, ipf_nblks; 213 uint64_t end_of_access_blkid; 214 end_of_access_blkid = blkid + nblks; 215 spa_t *spa = zf->zf_dnode->dn_objset->os_spa; 216 217 if (zfs_prefetch_disable) 218 return; 219 /* 220 * If we haven't yet loaded the indirect vdevs' mappings, we 221 * can only read from blocks that we carefully ensure are on 222 * concrete vdevs (or previously-loaded indirect vdevs). So we 223 * can't allow the predictive prefetcher to attempt reads of other 224 * blocks (e.g. of the MOS's dnode object). 225 */ 226 if (!spa_indirect_vdevs_loaded(spa)) 227 return; 228 229 /* 230 * As a fast path for small (single-block) files, ignore access 231 * to the first block. 232 */ 233 if (blkid == 0) 234 return; 235 236 if (!have_lock) 237 rw_enter(&zf->zf_dnode->dn_struct_rwlock, RW_READER); 238 mutex_enter(&zf->zf_lock); 239 240 /* 241 * Find matching prefetch stream. Depending on whether the accesses 242 * are block-aligned, first block of the new access may either follow 243 * the last block of the previous access, or be equal to it. 244 */ 245 for (zs = list_head(&zf->zf_stream); zs != NULL; 246 zs = list_next(&zf->zf_stream, zs)) { 247 if (blkid == zs->zs_blkid || blkid + 1 == zs->zs_blkid) { 248 mutex_enter(&zs->zs_lock); 249 /* 250 * zs_blkid could have changed before we 251 * acquired zs_lock; re-check them here. 252 */ 253 if (blkid == zs->zs_blkid) { 254 break; 255 } else if (blkid + 1 == zs->zs_blkid) { 256 blkid++; 257 nblks--; 258 if (nblks == 0) { 259 /* Already prefetched this before. */ 260 mutex_exit(&zs->zs_lock); 261 mutex_exit(&zf->zf_lock); 262 if (!have_lock) { 263 rw_exit(&zf->zf_dnode-> 264 dn_struct_rwlock); 265 } 266 return; 267 } 268 break; 269 } 270 mutex_exit(&zs->zs_lock); 271 } 272 } 273 274 if (zs == NULL) { 275 /* 276 * This access is not part of any existing stream. Create 277 * a new stream for it. 278 */ 279 ZFETCHSTAT_BUMP(zfetchstat_misses); 280 281 dmu_zfetch_stream_create(zf, end_of_access_blkid); 282 mutex_exit(&zf->zf_lock); 283 if (!have_lock) 284 rw_exit(&zf->zf_dnode->dn_struct_rwlock); 285 return; 286 } 287 288 /* 289 * This access was to a block that we issued a prefetch for on 290 * behalf of this stream. Issue further prefetches for this stream. 291 * 292 * Normally, we start prefetching where we stopped 293 * prefetching last (zs_pf_blkid). But when we get our first 294 * hit on this stream, zs_pf_blkid == zs_blkid, we don't 295 * want to prefetch the block we just accessed. In this case, 296 * start just after the block we just accessed. 297 */ 298 pf_start = MAX(zs->zs_pf_blkid, end_of_access_blkid); 299 300 /* 301 * Double our amount of prefetched data, but don't let the 302 * prefetch get further ahead than zfetch_max_distance. 303 */ 304 if (fetch_data) { 305 max_dist_blks = 306 zfetch_max_distance >> zf->zf_dnode->dn_datablkshift; 307 /* 308 * Previously, we were (zs_pf_blkid - blkid) ahead. We 309 * want to now be double that, so read that amount again, 310 * plus the amount we are catching up by (i.e. the amount 311 * read just now). 312 */ 313 pf_ahead_blks = zs->zs_pf_blkid - blkid + nblks; 314 max_blks = max_dist_blks - (pf_start - end_of_access_blkid); 315 pf_nblks = MIN(pf_ahead_blks, max_blks); 316 } else { 317 pf_nblks = 0; 318 } 319 320 zs->zs_pf_blkid = pf_start + pf_nblks; 321 322 /* 323 * Do the same for indirects, starting from where we stopped last, 324 * or where we will stop reading data blocks (and the indirects 325 * that point to them). 326 */ 327 ipf_start = MAX(zs->zs_ipf_blkid, zs->zs_pf_blkid); 328 max_dist_blks = zfetch_max_idistance >> zf->zf_dnode->dn_datablkshift; 329 /* 330 * We want to double our distance ahead of the data prefetch 331 * (or reader, if we are not prefetching data). Previously, we 332 * were (zs_ipf_blkid - blkid) ahead. To double that, we read 333 * that amount again, plus the amount we are catching up by 334 * (i.e. the amount read now + the amount of data prefetched now). 335 */ 336 pf_ahead_blks = zs->zs_ipf_blkid - blkid + nblks + pf_nblks; 337 max_blks = max_dist_blks - (ipf_start - end_of_access_blkid); 338 ipf_nblks = MIN(pf_ahead_blks, max_blks); 339 zs->zs_ipf_blkid = ipf_start + ipf_nblks; 340 341 epbs = zf->zf_dnode->dn_indblkshift - SPA_BLKPTRSHIFT; 342 ipf_istart = P2ROUNDUP(ipf_start, 1 << epbs) >> epbs; 343 ipf_iend = P2ROUNDUP(zs->zs_ipf_blkid, 1 << epbs) >> epbs; 344 345 zs->zs_atime = gethrtime(); 346 zs->zs_blkid = end_of_access_blkid; 347 mutex_exit(&zs->zs_lock); 348 mutex_exit(&zf->zf_lock); 349 350 /* 351 * dbuf_prefetch() is asynchronous (even when it needs to read 352 * indirect blocks), but we still prefer to drop our locks before 353 * calling it to reduce the time we hold them. 354 */ 355 356 for (int i = 0; i < pf_nblks; i++) { 357 dbuf_prefetch(zf->zf_dnode, 0, pf_start + i, 358 ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH); 359 } 360 for (int64_t iblk = ipf_istart; iblk < ipf_iend; iblk++) { 361 dbuf_prefetch(zf->zf_dnode, 1, iblk, 362 ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH); 363 } 364 if (!have_lock) 365 rw_exit(&zf->zf_dnode->dn_struct_rwlock); 366 ZFETCHSTAT_BUMP(zfetchstat_hits); 367 } 368 369 /* BEGIN CSTYLED */ 370 ZFS_MODULE_PARAM(zfs_prefetch, zfs_prefetch_, disable, INT, ZMOD_RW, 371 "Disable all ZFS prefetching"); 372 373 ZFS_MODULE_PARAM(zfs_prefetch, zfetch_, max_streams, UINT, ZMOD_RW, 374 "Max number of streams per zfetch"); 375 376 ZFS_MODULE_PARAM(zfs_prefetch, zfetch_, min_sec_reap, UINT, ZMOD_RW, 377 "Min time before stream reclaim"); 378 379 ZFS_MODULE_PARAM(zfs_prefetch, zfetch_, max_distance, UINT, ZMOD_RW, 380 "Max bytes to prefetch per stream (default 8MB)"); 381 382 ZFS_MODULE_PARAM(zfs_prefetch, zfetch_, array_rd_sz, ULONG, ZMOD_RW, 383 "Number of bytes in a array_read"); 384 /* END CSTYLED */ 385