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 * Copyright (c) 2013, 2015 by Delphix. All rights reserved. 27 */ 28 29 #include <sys/zfs_context.h> 30 #include <sys/spa.h> 31 #include <sys/vdev_impl.h> 32 #include <sys/zio.h> 33 #include <sys/kstat.h> 34 35 /* 36 * Virtual device read-ahead caching. 37 * 38 * This file implements a simple LRU read-ahead cache. When the DMU reads 39 * a given block, it will often want other, nearby blocks soon thereafter. 40 * We take advantage of this by reading a larger disk region and caching 41 * the result. In the best case, this can turn 128 back-to-back 512-byte 42 * reads into a single 64k read followed by 127 cache hits; this reduces 43 * latency dramatically. In the worst case, it can turn an isolated 512-byte 44 * read into a 64k read, which doesn't affect latency all that much but is 45 * terribly wasteful of bandwidth. A more intelligent version of the cache 46 * could keep track of access patterns and not do read-ahead unless it sees 47 * at least two temporally close I/Os to the same region. Currently, only 48 * metadata I/O is inflated. A futher enhancement could take advantage of 49 * more semantic information about the I/O. And it could use something 50 * faster than an AVL tree; that was chosen solely for convenience. 51 * 52 * There are five cache operations: allocate, fill, read, write, evict. 53 * 54 * (1) Allocate. This reserves a cache entry for the specified region. 55 * We separate the allocate and fill operations so that multiple threads 56 * don't generate I/O for the same cache miss. 57 * 58 * (2) Fill. When the I/O for a cache miss completes, the fill routine 59 * places the data in the previously allocated cache entry. 60 * 61 * (3) Read. Read data from the cache. 62 * 63 * (4) Write. Update cache contents after write completion. 64 * 65 * (5) Evict. When allocating a new entry, we evict the oldest (LRU) entry 66 * if the total cache size exceeds zfs_vdev_cache_size. 67 */ 68 69 /* 70 * These tunables are for performance analysis. 71 */ 72 /* 73 * All i/os smaller than zfs_vdev_cache_max will be turned into 74 * 1<<zfs_vdev_cache_bshift byte reads by the vdev_cache (aka software 75 * track buffer). At most zfs_vdev_cache_size bytes will be kept in each 76 * vdev's vdev_cache. 77 * 78 * TODO: Note that with the current ZFS code, it turns out that the 79 * vdev cache is not helpful, and in some cases actually harmful. It 80 * is better if we disable this. Once some time has passed, we should 81 * actually remove this to simplify the code. For now we just disable 82 * it by setting the zfs_vdev_cache_size to zero. Note that Solaris 11 83 * has made these same changes. 84 */ 85 int zfs_vdev_cache_max = 1<<14; /* 16KB */ 86 int zfs_vdev_cache_size = 0; 87 int zfs_vdev_cache_bshift = 16; 88 89 #define VCBS (1 << zfs_vdev_cache_bshift) /* 64KB */ 90 91 SYSCTL_DECL(_vfs_zfs_vdev); 92 SYSCTL_NODE(_vfs_zfs_vdev, OID_AUTO, cache, CTLFLAG_RW, 0, "ZFS VDEV Cache"); 93 SYSCTL_INT(_vfs_zfs_vdev_cache, OID_AUTO, max, CTLFLAG_RDTUN, 94 &zfs_vdev_cache_max, 0, "Maximum I/O request size that increase read size"); 95 SYSCTL_INT(_vfs_zfs_vdev_cache, OID_AUTO, size, CTLFLAG_RDTUN, 96 &zfs_vdev_cache_size, 0, "Size of VDEV cache"); 97 SYSCTL_INT(_vfs_zfs_vdev_cache, OID_AUTO, bshift, CTLFLAG_RDTUN, 98 &zfs_vdev_cache_bshift, 0, "Turn too small requests into 1 << this value"); 99 100 kstat_t *vdc_ksp = NULL; 101 102 typedef struct vdc_stats { 103 kstat_named_t vdc_stat_delegations; 104 kstat_named_t vdc_stat_hits; 105 kstat_named_t vdc_stat_misses; 106 } vdc_stats_t; 107 108 static vdc_stats_t vdc_stats = { 109 { "delegations", KSTAT_DATA_UINT64 }, 110 { "hits", KSTAT_DATA_UINT64 }, 111 { "misses", KSTAT_DATA_UINT64 } 112 }; 113 114 #define VDCSTAT_BUMP(stat) atomic_inc_64(&vdc_stats.stat.value.ui64); 115 116 static int 117 vdev_cache_offset_compare(const void *a1, const void *a2) 118 { 119 const vdev_cache_entry_t *ve1 = a1; 120 const vdev_cache_entry_t *ve2 = a2; 121 122 if (ve1->ve_offset < ve2->ve_offset) 123 return (-1); 124 if (ve1->ve_offset > ve2->ve_offset) 125 return (1); 126 return (0); 127 } 128 129 static int 130 vdev_cache_lastused_compare(const void *a1, const void *a2) 131 { 132 const vdev_cache_entry_t *ve1 = a1; 133 const vdev_cache_entry_t *ve2 = a2; 134 135 if (ve1->ve_lastused < ve2->ve_lastused) 136 return (-1); 137 if (ve1->ve_lastused > ve2->ve_lastused) 138 return (1); 139 140 /* 141 * Among equally old entries, sort by offset to ensure uniqueness. 142 */ 143 return (vdev_cache_offset_compare(a1, a2)); 144 } 145 146 /* 147 * Evict the specified entry from the cache. 148 */ 149 static void 150 vdev_cache_evict(vdev_cache_t *vc, vdev_cache_entry_t *ve) 151 { 152 ASSERT(MUTEX_HELD(&vc->vc_lock)); 153 ASSERT(ve->ve_fill_io == NULL); 154 ASSERT(ve->ve_data != NULL); 155 156 avl_remove(&vc->vc_lastused_tree, ve); 157 avl_remove(&vc->vc_offset_tree, ve); 158 zio_buf_free(ve->ve_data, VCBS); 159 kmem_free(ve, sizeof (vdev_cache_entry_t)); 160 } 161 162 /* 163 * Allocate an entry in the cache. At the point we don't have the data, 164 * we're just creating a placeholder so that multiple threads don't all 165 * go off and read the same blocks. 166 */ 167 static vdev_cache_entry_t * 168 vdev_cache_allocate(zio_t *zio) 169 { 170 vdev_cache_t *vc = &zio->io_vd->vdev_cache; 171 uint64_t offset = P2ALIGN(zio->io_offset, VCBS); 172 vdev_cache_entry_t *ve; 173 174 ASSERT(MUTEX_HELD(&vc->vc_lock)); 175 176 if (zfs_vdev_cache_size == 0) 177 return (NULL); 178 179 /* 180 * If adding a new entry would exceed the cache size, 181 * evict the oldest entry (LRU). 182 */ 183 if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) > 184 zfs_vdev_cache_size) { 185 ve = avl_first(&vc->vc_lastused_tree); 186 if (ve->ve_fill_io != NULL) 187 return (NULL); 188 ASSERT(ve->ve_hits != 0); 189 vdev_cache_evict(vc, ve); 190 } 191 192 ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_SLEEP); 193 ve->ve_offset = offset; 194 ve->ve_lastused = ddi_get_lbolt(); 195 ve->ve_data = zio_buf_alloc(VCBS); 196 197 avl_add(&vc->vc_offset_tree, ve); 198 avl_add(&vc->vc_lastused_tree, ve); 199 200 return (ve); 201 } 202 203 static void 204 vdev_cache_hit(vdev_cache_t *vc, vdev_cache_entry_t *ve, zio_t *zio) 205 { 206 uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS); 207 208 ASSERT(MUTEX_HELD(&vc->vc_lock)); 209 ASSERT(ve->ve_fill_io == NULL); 210 211 if (ve->ve_lastused != ddi_get_lbolt()) { 212 avl_remove(&vc->vc_lastused_tree, ve); 213 ve->ve_lastused = ddi_get_lbolt(); 214 avl_add(&vc->vc_lastused_tree, ve); 215 } 216 217 ve->ve_hits++; 218 bcopy(ve->ve_data + cache_phase, zio->io_data, zio->io_size); 219 } 220 221 /* 222 * Fill a previously allocated cache entry with data. 223 */ 224 static void 225 vdev_cache_fill(zio_t *fio) 226 { 227 vdev_t *vd = fio->io_vd; 228 vdev_cache_t *vc = &vd->vdev_cache; 229 vdev_cache_entry_t *ve = fio->io_private; 230 zio_t *pio; 231 232 ASSERT(fio->io_size == VCBS); 233 234 /* 235 * Add data to the cache. 236 */ 237 mutex_enter(&vc->vc_lock); 238 239 ASSERT(ve->ve_fill_io == fio); 240 ASSERT(ve->ve_offset == fio->io_offset); 241 ASSERT(ve->ve_data == fio->io_data); 242 243 ve->ve_fill_io = NULL; 244 245 /* 246 * Even if this cache line was invalidated by a missed write update, 247 * any reads that were queued up before the missed update are still 248 * valid, so we can satisfy them from this line before we evict it. 249 */ 250 zio_link_t *zl = NULL; 251 while ((pio = zio_walk_parents(fio, &zl)) != NULL) 252 vdev_cache_hit(vc, ve, pio); 253 254 if (fio->io_error || ve->ve_missed_update) 255 vdev_cache_evict(vc, ve); 256 257 mutex_exit(&vc->vc_lock); 258 } 259 260 /* 261 * Read data from the cache. Returns B_TRUE cache hit, B_FALSE on miss. 262 */ 263 boolean_t 264 vdev_cache_read(zio_t *zio) 265 { 266 vdev_cache_t *vc = &zio->io_vd->vdev_cache; 267 vdev_cache_entry_t *ve, ve_search; 268 uint64_t cache_offset = P2ALIGN(zio->io_offset, VCBS); 269 uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS); 270 zio_t *fio; 271 272 ASSERT(zio->io_type == ZIO_TYPE_READ); 273 274 if (zio->io_flags & ZIO_FLAG_DONT_CACHE) 275 return (B_FALSE); 276 277 if (zio->io_size > zfs_vdev_cache_max) 278 return (B_FALSE); 279 280 /* 281 * If the I/O straddles two or more cache blocks, don't cache it. 282 */ 283 if (P2BOUNDARY(zio->io_offset, zio->io_size, VCBS)) 284 return (B_FALSE); 285 286 ASSERT(cache_phase + zio->io_size <= VCBS); 287 288 mutex_enter(&vc->vc_lock); 289 290 ve_search.ve_offset = cache_offset; 291 ve = avl_find(&vc->vc_offset_tree, &ve_search, NULL); 292 293 if (ve != NULL) { 294 if (ve->ve_missed_update) { 295 mutex_exit(&vc->vc_lock); 296 return (B_FALSE); 297 } 298 299 if ((fio = ve->ve_fill_io) != NULL) { 300 zio_vdev_io_bypass(zio); 301 zio_add_child(zio, fio); 302 mutex_exit(&vc->vc_lock); 303 VDCSTAT_BUMP(vdc_stat_delegations); 304 return (B_TRUE); 305 } 306 307 vdev_cache_hit(vc, ve, zio); 308 zio_vdev_io_bypass(zio); 309 310 mutex_exit(&vc->vc_lock); 311 VDCSTAT_BUMP(vdc_stat_hits); 312 return (B_TRUE); 313 } 314 315 ve = vdev_cache_allocate(zio); 316 317 if (ve == NULL) { 318 mutex_exit(&vc->vc_lock); 319 return (B_FALSE); 320 } 321 322 fio = zio_vdev_delegated_io(zio->io_vd, cache_offset, 323 ve->ve_data, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_NOW, 324 ZIO_FLAG_DONT_CACHE, vdev_cache_fill, ve); 325 326 ve->ve_fill_io = fio; 327 zio_vdev_io_bypass(zio); 328 zio_add_child(zio, fio); 329 330 mutex_exit(&vc->vc_lock); 331 zio_nowait(fio); 332 VDCSTAT_BUMP(vdc_stat_misses); 333 334 return (B_TRUE); 335 } 336 337 /* 338 * Update cache contents upon write completion. 339 */ 340 void 341 vdev_cache_write(zio_t *zio) 342 { 343 vdev_cache_t *vc = &zio->io_vd->vdev_cache; 344 vdev_cache_entry_t *ve, ve_search; 345 uint64_t io_start = zio->io_offset; 346 uint64_t io_end = io_start + zio->io_size; 347 uint64_t min_offset = P2ALIGN(io_start, VCBS); 348 uint64_t max_offset = P2ROUNDUP(io_end, VCBS); 349 avl_index_t where; 350 351 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 352 353 mutex_enter(&vc->vc_lock); 354 355 ve_search.ve_offset = min_offset; 356 ve = avl_find(&vc->vc_offset_tree, &ve_search, &where); 357 358 if (ve == NULL) 359 ve = avl_nearest(&vc->vc_offset_tree, where, AVL_AFTER); 360 361 while (ve != NULL && ve->ve_offset < max_offset) { 362 uint64_t start = MAX(ve->ve_offset, io_start); 363 uint64_t end = MIN(ve->ve_offset + VCBS, io_end); 364 365 if (ve->ve_fill_io != NULL) { 366 ve->ve_missed_update = 1; 367 } else { 368 bcopy((char *)zio->io_data + start - io_start, 369 ve->ve_data + start - ve->ve_offset, end - start); 370 } 371 ve = AVL_NEXT(&vc->vc_offset_tree, ve); 372 } 373 mutex_exit(&vc->vc_lock); 374 } 375 376 void 377 vdev_cache_purge(vdev_t *vd) 378 { 379 vdev_cache_t *vc = &vd->vdev_cache; 380 vdev_cache_entry_t *ve; 381 382 mutex_enter(&vc->vc_lock); 383 while ((ve = avl_first(&vc->vc_offset_tree)) != NULL) 384 vdev_cache_evict(vc, ve); 385 mutex_exit(&vc->vc_lock); 386 } 387 388 void 389 vdev_cache_init(vdev_t *vd) 390 { 391 vdev_cache_t *vc = &vd->vdev_cache; 392 393 mutex_init(&vc->vc_lock, NULL, MUTEX_DEFAULT, NULL); 394 395 avl_create(&vc->vc_offset_tree, vdev_cache_offset_compare, 396 sizeof (vdev_cache_entry_t), 397 offsetof(struct vdev_cache_entry, ve_offset_node)); 398 399 avl_create(&vc->vc_lastused_tree, vdev_cache_lastused_compare, 400 sizeof (vdev_cache_entry_t), 401 offsetof(struct vdev_cache_entry, ve_lastused_node)); 402 } 403 404 void 405 vdev_cache_fini(vdev_t *vd) 406 { 407 vdev_cache_t *vc = &vd->vdev_cache; 408 409 vdev_cache_purge(vd); 410 411 avl_destroy(&vc->vc_offset_tree); 412 avl_destroy(&vc->vc_lastused_tree); 413 414 mutex_destroy(&vc->vc_lock); 415 } 416 417 void 418 vdev_cache_stat_init(void) 419 { 420 vdc_ksp = kstat_create("zfs", 0, "vdev_cache_stats", "misc", 421 KSTAT_TYPE_NAMED, sizeof (vdc_stats) / sizeof (kstat_named_t), 422 KSTAT_FLAG_VIRTUAL); 423 if (vdc_ksp != NULL) { 424 vdc_ksp->ks_data = &vdc_stats; 425 kstat_install(vdc_ksp); 426 } 427 } 428 429 void 430 vdev_cache_stat_fini(void) 431 { 432 if (vdc_ksp != NULL) { 433 kstat_delete(vdc_ksp); 434 vdc_ksp = NULL; 435 } 436 } 437