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 2008 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26 #include <sys/zfs_context.h> 27 #include <sys/spa.h> 28 #include <sys/vdev_impl.h> 29 #include <sys/zio.h> 30 #include <sys/kstat.h> 31 32 /* 33 * Virtual device read-ahead caching. 34 * 35 * This file implements a simple LRU read-ahead cache. When the DMU reads 36 * a given block, it will often want other, nearby blocks soon thereafter. 37 * We take advantage of this by reading a larger disk region and caching 38 * the result. In the best case, this can turn 128 back-to-back 512-byte 39 * reads into a single 64k read followed by 127 cache hits; this reduces 40 * latency dramatically. In the worst case, it can turn an isolated 512-byte 41 * read into a 64k read, which doesn't affect latency all that much but is 42 * terribly wasteful of bandwidth. A more intelligent version of the cache 43 * could keep track of access patterns and not do read-ahead unless it sees 44 * at least two temporally close I/Os to the same region. Currently, only 45 * metadata I/O is inflated. A futher enhancement could take advantage of 46 * more semantic information about the I/O. And it could use something 47 * faster than an AVL tree; that was chosen solely for convenience. 48 * 49 * There are five cache operations: allocate, fill, read, write, evict. 50 * 51 * (1) Allocate. This reserves a cache entry for the specified region. 52 * We separate the allocate and fill operations so that multiple threads 53 * don't generate I/O for the same cache miss. 54 * 55 * (2) Fill. When the I/O for a cache miss completes, the fill routine 56 * places the data in the previously allocated cache entry. 57 * 58 * (3) Read. Read data from the cache. 59 * 60 * (4) Write. Update cache contents after write completion. 61 * 62 * (5) Evict. When allocating a new entry, we evict the oldest (LRU) entry 63 * if the total cache size exceeds zfs_vdev_cache_size. 64 */ 65 66 /* 67 * These tunables are for performance analysis. 68 */ 69 /* 70 * All i/os smaller than zfs_vdev_cache_max will be turned into 71 * 1<<zfs_vdev_cache_bshift byte reads by the vdev_cache (aka software 72 * track buffer). At most zfs_vdev_cache_size bytes will be kept in each 73 * vdev's vdev_cache. 74 */ 75 int zfs_vdev_cache_max = 1<<14; /* 16KB */ 76 int zfs_vdev_cache_size = 10ULL << 20; /* 10MB */ 77 int zfs_vdev_cache_bshift = 16; 78 79 #define VCBS (1 << zfs_vdev_cache_bshift) /* 64KB */ 80 81 kstat_t *vdc_ksp = NULL; 82 83 typedef struct vdc_stats { 84 kstat_named_t vdc_stat_delegations; 85 kstat_named_t vdc_stat_hits; 86 kstat_named_t vdc_stat_misses; 87 } vdc_stats_t; 88 89 static vdc_stats_t vdc_stats = { 90 { "delegations", KSTAT_DATA_UINT64 }, 91 { "hits", KSTAT_DATA_UINT64 }, 92 { "misses", KSTAT_DATA_UINT64 } 93 }; 94 95 #define VDCSTAT_BUMP(stat) atomic_add_64(&vdc_stats.stat.value.ui64, 1); 96 97 static int 98 vdev_cache_offset_compare(const void *a1, const void *a2) 99 { 100 const vdev_cache_entry_t *ve1 = a1; 101 const vdev_cache_entry_t *ve2 = a2; 102 103 if (ve1->ve_offset < ve2->ve_offset) 104 return (-1); 105 if (ve1->ve_offset > ve2->ve_offset) 106 return (1); 107 return (0); 108 } 109 110 static int 111 vdev_cache_lastused_compare(const void *a1, const void *a2) 112 { 113 const vdev_cache_entry_t *ve1 = a1; 114 const vdev_cache_entry_t *ve2 = a2; 115 116 if (ve1->ve_lastused < ve2->ve_lastused) 117 return (-1); 118 if (ve1->ve_lastused > ve2->ve_lastused) 119 return (1); 120 121 /* 122 * Among equally old entries, sort by offset to ensure uniqueness. 123 */ 124 return (vdev_cache_offset_compare(a1, a2)); 125 } 126 127 /* 128 * Evict the specified entry from the cache. 129 */ 130 static void 131 vdev_cache_evict(vdev_cache_t *vc, vdev_cache_entry_t *ve) 132 { 133 ASSERT(MUTEX_HELD(&vc->vc_lock)); 134 ASSERT(ve->ve_fill_io == NULL); 135 ASSERT(ve->ve_data != NULL); 136 137 avl_remove(&vc->vc_lastused_tree, ve); 138 avl_remove(&vc->vc_offset_tree, ve); 139 zio_buf_free(ve->ve_data, VCBS); 140 kmem_free(ve, sizeof (vdev_cache_entry_t)); 141 } 142 143 /* 144 * Allocate an entry in the cache. At the point we don't have the data, 145 * we're just creating a placeholder so that multiple threads don't all 146 * go off and read the same blocks. 147 */ 148 static vdev_cache_entry_t * 149 vdev_cache_allocate(zio_t *zio) 150 { 151 vdev_cache_t *vc = &zio->io_vd->vdev_cache; 152 uint64_t offset = P2ALIGN(zio->io_offset, VCBS); 153 vdev_cache_entry_t *ve; 154 155 ASSERT(MUTEX_HELD(&vc->vc_lock)); 156 157 if (zfs_vdev_cache_size == 0) 158 return (NULL); 159 160 /* 161 * If adding a new entry would exceed the cache size, 162 * evict the oldest entry (LRU). 163 */ 164 if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) > 165 zfs_vdev_cache_size) { 166 ve = avl_first(&vc->vc_lastused_tree); 167 if (ve->ve_fill_io != NULL) 168 return (NULL); 169 ASSERT(ve->ve_hits != 0); 170 vdev_cache_evict(vc, ve); 171 } 172 173 ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_SLEEP); 174 ve->ve_offset = offset; 175 ve->ve_lastused = lbolt; 176 ve->ve_data = zio_buf_alloc(VCBS); 177 178 avl_add(&vc->vc_offset_tree, ve); 179 avl_add(&vc->vc_lastused_tree, ve); 180 181 return (ve); 182 } 183 184 static void 185 vdev_cache_hit(vdev_cache_t *vc, vdev_cache_entry_t *ve, zio_t *zio) 186 { 187 uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS); 188 189 ASSERT(MUTEX_HELD(&vc->vc_lock)); 190 ASSERT(ve->ve_fill_io == NULL); 191 192 if (ve->ve_lastused != lbolt) { 193 avl_remove(&vc->vc_lastused_tree, ve); 194 ve->ve_lastused = lbolt; 195 avl_add(&vc->vc_lastused_tree, ve); 196 } 197 198 ve->ve_hits++; 199 bcopy(ve->ve_data + cache_phase, zio->io_data, zio->io_size); 200 } 201 202 /* 203 * Fill a previously allocated cache entry with data. 204 */ 205 static void 206 vdev_cache_fill(zio_t *zio) 207 { 208 vdev_t *vd = zio->io_vd; 209 vdev_cache_t *vc = &vd->vdev_cache; 210 vdev_cache_entry_t *ve = zio->io_private; 211 zio_t *dio; 212 213 ASSERT(zio->io_size == VCBS); 214 215 /* 216 * Add data to the cache. 217 */ 218 mutex_enter(&vc->vc_lock); 219 220 ASSERT(ve->ve_fill_io == zio); 221 ASSERT(ve->ve_offset == zio->io_offset); 222 ASSERT(ve->ve_data == zio->io_data); 223 224 ve->ve_fill_io = NULL; 225 226 /* 227 * Even if this cache line was invalidated by a missed write update, 228 * any reads that were queued up before the missed update are still 229 * valid, so we can satisfy them from this line before we evict it. 230 */ 231 for (dio = zio->io_delegate_list; dio; dio = dio->io_delegate_next) 232 vdev_cache_hit(vc, ve, dio); 233 234 if (zio->io_error || ve->ve_missed_update) 235 vdev_cache_evict(vc, ve); 236 237 mutex_exit(&vc->vc_lock); 238 239 while ((dio = zio->io_delegate_list) != NULL) { 240 zio->io_delegate_list = dio->io_delegate_next; 241 dio->io_delegate_next = NULL; 242 dio->io_error = zio->io_error; 243 zio_execute(dio); 244 } 245 } 246 247 /* 248 * Read data from the cache. Returns 0 on cache hit, errno on a miss. 249 */ 250 int 251 vdev_cache_read(zio_t *zio) 252 { 253 vdev_cache_t *vc = &zio->io_vd->vdev_cache; 254 vdev_cache_entry_t *ve, ve_search; 255 uint64_t cache_offset = P2ALIGN(zio->io_offset, VCBS); 256 uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS); 257 zio_t *fio; 258 259 ASSERT(zio->io_type == ZIO_TYPE_READ); 260 261 if (zio->io_flags & ZIO_FLAG_DONT_CACHE) 262 return (EINVAL); 263 264 if (zio->io_size > zfs_vdev_cache_max) 265 return (EOVERFLOW); 266 267 /* 268 * If the I/O straddles two or more cache blocks, don't cache it. 269 */ 270 if (P2CROSS(zio->io_offset, zio->io_offset + zio->io_size - 1, VCBS)) 271 return (EXDEV); 272 273 ASSERT(cache_phase + zio->io_size <= VCBS); 274 275 mutex_enter(&vc->vc_lock); 276 277 ve_search.ve_offset = cache_offset; 278 ve = avl_find(&vc->vc_offset_tree, &ve_search, NULL); 279 280 if (ve != NULL) { 281 if (ve->ve_missed_update) { 282 mutex_exit(&vc->vc_lock); 283 return (ESTALE); 284 } 285 286 if ((fio = ve->ve_fill_io) != NULL) { 287 zio->io_delegate_next = fio->io_delegate_list; 288 fio->io_delegate_list = zio; 289 zio_vdev_io_bypass(zio); 290 mutex_exit(&vc->vc_lock); 291 VDCSTAT_BUMP(vdc_stat_delegations); 292 return (0); 293 } 294 295 vdev_cache_hit(vc, ve, zio); 296 zio_vdev_io_bypass(zio); 297 298 mutex_exit(&vc->vc_lock); 299 zio_execute(zio); 300 VDCSTAT_BUMP(vdc_stat_hits); 301 return (0); 302 } 303 304 ve = vdev_cache_allocate(zio); 305 306 if (ve == NULL) { 307 mutex_exit(&vc->vc_lock); 308 return (ENOMEM); 309 } 310 311 fio = zio_vdev_delegated_io(zio->io_vd, cache_offset, 312 ve->ve_data, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_CACHE_FILL, 313 ZIO_FLAG_DONT_CACHE, vdev_cache_fill, ve); 314 315 ve->ve_fill_io = fio; 316 fio->io_delegate_list = zio; 317 zio_vdev_io_bypass(zio); 318 319 mutex_exit(&vc->vc_lock); 320 zio_nowait(fio); 321 VDCSTAT_BUMP(vdc_stat_misses); 322 323 return (0); 324 } 325 326 /* 327 * Update cache contents upon write completion. 328 */ 329 void 330 vdev_cache_write(zio_t *zio) 331 { 332 vdev_cache_t *vc = &zio->io_vd->vdev_cache; 333 vdev_cache_entry_t *ve, ve_search; 334 uint64_t io_start = zio->io_offset; 335 uint64_t io_end = io_start + zio->io_size; 336 uint64_t min_offset = P2ALIGN(io_start, VCBS); 337 uint64_t max_offset = P2ROUNDUP(io_end, VCBS); 338 avl_index_t where; 339 340 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 341 342 mutex_enter(&vc->vc_lock); 343 344 ve_search.ve_offset = min_offset; 345 ve = avl_find(&vc->vc_offset_tree, &ve_search, &where); 346 347 if (ve == NULL) 348 ve = avl_nearest(&vc->vc_offset_tree, where, AVL_AFTER); 349 350 while (ve != NULL && ve->ve_offset < max_offset) { 351 uint64_t start = MAX(ve->ve_offset, io_start); 352 uint64_t end = MIN(ve->ve_offset + VCBS, io_end); 353 354 if (ve->ve_fill_io != NULL) { 355 ve->ve_missed_update = 1; 356 } else { 357 bcopy((char *)zio->io_data + start - io_start, 358 ve->ve_data + start - ve->ve_offset, end - start); 359 } 360 ve = AVL_NEXT(&vc->vc_offset_tree, ve); 361 } 362 mutex_exit(&vc->vc_lock); 363 } 364 365 void 366 vdev_cache_purge(vdev_t *vd) 367 { 368 vdev_cache_t *vc = &vd->vdev_cache; 369 vdev_cache_entry_t *ve; 370 371 mutex_enter(&vc->vc_lock); 372 while ((ve = avl_first(&vc->vc_offset_tree)) != NULL) 373 vdev_cache_evict(vc, ve); 374 mutex_exit(&vc->vc_lock); 375 } 376 377 void 378 vdev_cache_init(vdev_t *vd) 379 { 380 vdev_cache_t *vc = &vd->vdev_cache; 381 382 mutex_init(&vc->vc_lock, NULL, MUTEX_DEFAULT, NULL); 383 384 avl_create(&vc->vc_offset_tree, vdev_cache_offset_compare, 385 sizeof (vdev_cache_entry_t), 386 offsetof(struct vdev_cache_entry, ve_offset_node)); 387 388 avl_create(&vc->vc_lastused_tree, vdev_cache_lastused_compare, 389 sizeof (vdev_cache_entry_t), 390 offsetof(struct vdev_cache_entry, ve_lastused_node)); 391 } 392 393 void 394 vdev_cache_fini(vdev_t *vd) 395 { 396 vdev_cache_t *vc = &vd->vdev_cache; 397 398 vdev_cache_purge(vd); 399 400 avl_destroy(&vc->vc_offset_tree); 401 avl_destroy(&vc->vc_lastused_tree); 402 403 mutex_destroy(&vc->vc_lock); 404 } 405 406 void 407 vdev_cache_stat_init(void) 408 { 409 vdc_ksp = kstat_create("zfs", 0, "vdev_cache_stats", "misc", 410 KSTAT_TYPE_NAMED, sizeof (vdc_stats) / sizeof (kstat_named_t), 411 KSTAT_FLAG_VIRTUAL); 412 if (vdc_ksp != NULL) { 413 vdc_ksp->ks_data = &vdc_stats; 414 kstat_install(vdc_ksp); 415 } 416 } 417 418 void 419 vdev_cache_stat_fini(void) 420 { 421 if (vdc_ksp != NULL) { 422 kstat_delete(vdc_ksp); 423 vdc_ksp = NULL; 424 } 425 } 426