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) 2012, 2018 by Delphix. All rights reserved. 28 */ 29 30 #include <sys/zfs_context.h> 31 #include <sys/vdev_impl.h> 32 #include <sys/spa_impl.h> 33 #include <sys/zio.h> 34 #include <sys/avl.h> 35 #include <sys/dsl_pool.h> 36 #include <sys/metaslab_impl.h> 37 #include <sys/spa.h> 38 #include <sys/spa_impl.h> 39 #include <sys/kstat.h> 40 #include <sys/abd.h> 41 42 /* 43 * ZFS I/O Scheduler 44 * --------------- 45 * 46 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The 47 * I/O scheduler determines when and in what order those operations are 48 * issued. The I/O scheduler divides operations into five I/O classes 49 * prioritized in the following order: sync read, sync write, async read, 50 * async write, and scrub/resilver. Each queue defines the minimum and 51 * maximum number of concurrent operations that may be issued to the device. 52 * In addition, the device has an aggregate maximum. Note that the sum of the 53 * per-queue minimums must not exceed the aggregate maximum. If the 54 * sum of the per-queue maximums exceeds the aggregate maximum, then the 55 * number of active i/os may reach zfs_vdev_max_active, in which case no 56 * further i/os will be issued regardless of whether all per-queue 57 * minimums have been met. 58 * 59 * For many physical devices, throughput increases with the number of 60 * concurrent operations, but latency typically suffers. Further, physical 61 * devices typically have a limit at which more concurrent operations have no 62 * effect on throughput or can actually cause it to decrease. 63 * 64 * The scheduler selects the next operation to issue by first looking for an 65 * I/O class whose minimum has not been satisfied. Once all are satisfied and 66 * the aggregate maximum has not been hit, the scheduler looks for classes 67 * whose maximum has not been satisfied. Iteration through the I/O classes is 68 * done in the order specified above. No further operations are issued if the 69 * aggregate maximum number of concurrent operations has been hit or if there 70 * are no operations queued for an I/O class that has not hit its maximum. 71 * Every time an i/o is queued or an operation completes, the I/O scheduler 72 * looks for new operations to issue. 73 * 74 * All I/O classes have a fixed maximum number of outstanding operations 75 * except for the async write class. Asynchronous writes represent the data 76 * that is committed to stable storage during the syncing stage for 77 * transaction groups (see txg.c). Transaction groups enter the syncing state 78 * periodically so the number of queued async writes will quickly burst up and 79 * then bleed down to zero. Rather than servicing them as quickly as possible, 80 * the I/O scheduler changes the maximum number of active async write i/os 81 * according to the amount of dirty data in the pool (see dsl_pool.c). Since 82 * both throughput and latency typically increase with the number of 83 * concurrent operations issued to physical devices, reducing the burstiness 84 * in the number of concurrent operations also stabilizes the response time of 85 * operations from other -- and in particular synchronous -- queues. In broad 86 * strokes, the I/O scheduler will issue more concurrent operations from the 87 * async write queue as there's more dirty data in the pool. 88 * 89 * Async Writes 90 * 91 * The number of concurrent operations issued for the async write I/O class 92 * follows a piece-wise linear function defined by a few adjustable points. 93 * 94 * | o---------| <-- zfs_vdev_async_write_max_active 95 * ^ | /^ | 96 * | | / | | 97 * active | / | | 98 * I/O | / | | 99 * count | / | | 100 * | / | | 101 * |------------o | | <-- zfs_vdev_async_write_min_active 102 * 0|____________^______|_________| 103 * 0% | | 100% of zfs_dirty_data_max 104 * | | 105 * | `-- zfs_vdev_async_write_active_max_dirty_percent 106 * `--------- zfs_vdev_async_write_active_min_dirty_percent 107 * 108 * Until the amount of dirty data exceeds a minimum percentage of the dirty 109 * data allowed in the pool, the I/O scheduler will limit the number of 110 * concurrent operations to the minimum. As that threshold is crossed, the 111 * number of concurrent operations issued increases linearly to the maximum at 112 * the specified maximum percentage of the dirty data allowed in the pool. 113 * 114 * Ideally, the amount of dirty data on a busy pool will stay in the sloped 115 * part of the function between zfs_vdev_async_write_active_min_dirty_percent 116 * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the 117 * maximum percentage, this indicates that the rate of incoming data is 118 * greater than the rate that the backend storage can handle. In this case, we 119 * must further throttle incoming writes (see dmu_tx_delay() for details). 120 */ 121 122 /* 123 * The maximum number of i/os active to each device. Ideally, this will be >= 124 * the sum of each queue's max_active. It must be at least the sum of each 125 * queue's min_active. 126 */ 127 uint32_t zfs_vdev_max_active = 1000; 128 129 /* 130 * Per-queue limits on the number of i/os active to each device. If the 131 * number of active i/os is < zfs_vdev_max_active, then the min_active comes 132 * into play. We will send min_active from each queue, and then select from 133 * queues in the order defined by zio_priority_t. 134 * 135 * In general, smaller max_active's will lead to lower latency of synchronous 136 * operations. Larger max_active's may lead to higher overall throughput, 137 * depending on underlying storage. 138 * 139 * The ratio of the queues' max_actives determines the balance of performance 140 * between reads, writes, and scrubs. E.g., increasing 141 * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete 142 * more quickly, but reads and writes to have higher latency and lower 143 * throughput. 144 */ 145 uint32_t zfs_vdev_sync_read_min_active = 10; 146 uint32_t zfs_vdev_sync_read_max_active = 10; 147 uint32_t zfs_vdev_sync_write_min_active = 10; 148 uint32_t zfs_vdev_sync_write_max_active = 10; 149 uint32_t zfs_vdev_async_read_min_active = 1; 150 uint32_t zfs_vdev_async_read_max_active = 3; 151 uint32_t zfs_vdev_async_write_min_active = 2; 152 uint32_t zfs_vdev_async_write_max_active = 10; 153 uint32_t zfs_vdev_scrub_min_active = 1; 154 uint32_t zfs_vdev_scrub_max_active = 2; 155 uint32_t zfs_vdev_removal_min_active = 1; 156 uint32_t zfs_vdev_removal_max_active = 2; 157 uint32_t zfs_vdev_initializing_min_active = 1; 158 uint32_t zfs_vdev_initializing_max_active = 1; 159 uint32_t zfs_vdev_trim_min_active = 1; 160 uint32_t zfs_vdev_trim_max_active = 2; 161 uint32_t zfs_vdev_rebuild_min_active = 1; 162 uint32_t zfs_vdev_rebuild_max_active = 3; 163 164 /* 165 * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent 166 * dirty data, use zfs_vdev_async_write_min_active. When it has more than 167 * zfs_vdev_async_write_active_max_dirty_percent, use 168 * zfs_vdev_async_write_max_active. The value is linearly interpolated 169 * between min and max. 170 */ 171 int zfs_vdev_async_write_active_min_dirty_percent = 30; 172 int zfs_vdev_async_write_active_max_dirty_percent = 60; 173 174 /* 175 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O. 176 * For read I/Os, we also aggregate across small adjacency gaps; for writes 177 * we include spans of optional I/Os to aid aggregation at the disk even when 178 * they aren't able to help us aggregate at this level. 179 */ 180 int zfs_vdev_aggregation_limit = 1 << 20; 181 int zfs_vdev_aggregation_limit_non_rotating = SPA_OLD_MAXBLOCKSIZE; 182 int zfs_vdev_read_gap_limit = 32 << 10; 183 int zfs_vdev_write_gap_limit = 4 << 10; 184 185 /* 186 * Define the queue depth percentage for each top-level. This percentage is 187 * used in conjunction with zfs_vdev_async_max_active to determine how many 188 * allocations a specific top-level vdev should handle. Once the queue depth 189 * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100 190 * then allocator will stop allocating blocks on that top-level device. 191 * The default kernel setting is 1000% which will yield 100 allocations per 192 * device. For userland testing, the default setting is 300% which equates 193 * to 30 allocations per device. 194 */ 195 #ifdef _KERNEL 196 int zfs_vdev_queue_depth_pct = 1000; 197 #else 198 int zfs_vdev_queue_depth_pct = 300; 199 #endif 200 201 /* 202 * When performing allocations for a given metaslab, we want to make sure that 203 * there are enough IOs to aggregate together to improve throughput. We want to 204 * ensure that there are at least 128k worth of IOs that can be aggregated, and 205 * we assume that the average allocation size is 4k, so we need the queue depth 206 * to be 32 per allocator to get good aggregation of sequential writes. 207 */ 208 int zfs_vdev_def_queue_depth = 32; 209 210 /* 211 * Allow TRIM I/Os to be aggregated. This should normally not be needed since 212 * TRIM I/O for extents up to zfs_trim_extent_bytes_max (128M) can be submitted 213 * by the TRIM code in zfs_trim.c. 214 */ 215 int zfs_vdev_aggregate_trim = 0; 216 217 static int 218 vdev_queue_offset_compare(const void *x1, const void *x2) 219 { 220 const zio_t *z1 = (const zio_t *)x1; 221 const zio_t *z2 = (const zio_t *)x2; 222 223 int cmp = TREE_CMP(z1->io_offset, z2->io_offset); 224 225 if (likely(cmp)) 226 return (cmp); 227 228 return (TREE_PCMP(z1, z2)); 229 } 230 231 static inline avl_tree_t * 232 vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p) 233 { 234 return (&vq->vq_class[p].vqc_queued_tree); 235 } 236 237 static inline avl_tree_t * 238 vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t) 239 { 240 ASSERT(t == ZIO_TYPE_READ || t == ZIO_TYPE_WRITE || t == ZIO_TYPE_TRIM); 241 if (t == ZIO_TYPE_READ) 242 return (&vq->vq_read_offset_tree); 243 else if (t == ZIO_TYPE_WRITE) 244 return (&vq->vq_write_offset_tree); 245 else 246 return (&vq->vq_trim_offset_tree); 247 } 248 249 static int 250 vdev_queue_timestamp_compare(const void *x1, const void *x2) 251 { 252 const zio_t *z1 = (const zio_t *)x1; 253 const zio_t *z2 = (const zio_t *)x2; 254 255 int cmp = TREE_CMP(z1->io_timestamp, z2->io_timestamp); 256 257 if (likely(cmp)) 258 return (cmp); 259 260 return (TREE_PCMP(z1, z2)); 261 } 262 263 static int 264 vdev_queue_class_min_active(zio_priority_t p) 265 { 266 switch (p) { 267 case ZIO_PRIORITY_SYNC_READ: 268 return (zfs_vdev_sync_read_min_active); 269 case ZIO_PRIORITY_SYNC_WRITE: 270 return (zfs_vdev_sync_write_min_active); 271 case ZIO_PRIORITY_ASYNC_READ: 272 return (zfs_vdev_async_read_min_active); 273 case ZIO_PRIORITY_ASYNC_WRITE: 274 return (zfs_vdev_async_write_min_active); 275 case ZIO_PRIORITY_SCRUB: 276 return (zfs_vdev_scrub_min_active); 277 case ZIO_PRIORITY_REMOVAL: 278 return (zfs_vdev_removal_min_active); 279 case ZIO_PRIORITY_INITIALIZING: 280 return (zfs_vdev_initializing_min_active); 281 case ZIO_PRIORITY_TRIM: 282 return (zfs_vdev_trim_min_active); 283 case ZIO_PRIORITY_REBUILD: 284 return (zfs_vdev_rebuild_min_active); 285 default: 286 panic("invalid priority %u", p); 287 return (0); 288 } 289 } 290 291 static int 292 vdev_queue_max_async_writes(spa_t *spa) 293 { 294 int writes; 295 uint64_t dirty = 0; 296 dsl_pool_t *dp = spa_get_dsl(spa); 297 uint64_t min_bytes = zfs_dirty_data_max * 298 zfs_vdev_async_write_active_min_dirty_percent / 100; 299 uint64_t max_bytes = zfs_dirty_data_max * 300 zfs_vdev_async_write_active_max_dirty_percent / 100; 301 302 /* 303 * Async writes may occur before the assignment of the spa's 304 * dsl_pool_t if a self-healing zio is issued prior to the 305 * completion of dmu_objset_open_impl(). 306 */ 307 if (dp == NULL) 308 return (zfs_vdev_async_write_max_active); 309 310 /* 311 * Sync tasks correspond to interactive user actions. To reduce the 312 * execution time of those actions we push data out as fast as possible. 313 */ 314 if (spa_has_pending_synctask(spa)) 315 return (zfs_vdev_async_write_max_active); 316 317 dirty = dp->dp_dirty_total; 318 if (dirty < min_bytes) 319 return (zfs_vdev_async_write_min_active); 320 if (dirty > max_bytes) 321 return (zfs_vdev_async_write_max_active); 322 323 /* 324 * linear interpolation: 325 * slope = (max_writes - min_writes) / (max_bytes - min_bytes) 326 * move right by min_bytes 327 * move up by min_writes 328 */ 329 writes = (dirty - min_bytes) * 330 (zfs_vdev_async_write_max_active - 331 zfs_vdev_async_write_min_active) / 332 (max_bytes - min_bytes) + 333 zfs_vdev_async_write_min_active; 334 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); 335 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); 336 return (writes); 337 } 338 339 static int 340 vdev_queue_class_max_active(spa_t *spa, zio_priority_t p) 341 { 342 switch (p) { 343 case ZIO_PRIORITY_SYNC_READ: 344 return (zfs_vdev_sync_read_max_active); 345 case ZIO_PRIORITY_SYNC_WRITE: 346 return (zfs_vdev_sync_write_max_active); 347 case ZIO_PRIORITY_ASYNC_READ: 348 return (zfs_vdev_async_read_max_active); 349 case ZIO_PRIORITY_ASYNC_WRITE: 350 return (vdev_queue_max_async_writes(spa)); 351 case ZIO_PRIORITY_SCRUB: 352 return (zfs_vdev_scrub_max_active); 353 case ZIO_PRIORITY_REMOVAL: 354 return (zfs_vdev_removal_max_active); 355 case ZIO_PRIORITY_INITIALIZING: 356 return (zfs_vdev_initializing_max_active); 357 case ZIO_PRIORITY_TRIM: 358 return (zfs_vdev_trim_max_active); 359 case ZIO_PRIORITY_REBUILD: 360 return (zfs_vdev_rebuild_max_active); 361 default: 362 panic("invalid priority %u", p); 363 return (0); 364 } 365 } 366 367 /* 368 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if 369 * there is no eligible class. 370 */ 371 static zio_priority_t 372 vdev_queue_class_to_issue(vdev_queue_t *vq) 373 { 374 spa_t *spa = vq->vq_vdev->vdev_spa; 375 zio_priority_t p; 376 377 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active) 378 return (ZIO_PRIORITY_NUM_QUEUEABLE); 379 380 /* find a queue that has not reached its minimum # outstanding i/os */ 381 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 382 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && 383 vq->vq_class[p].vqc_active < 384 vdev_queue_class_min_active(p)) 385 return (p); 386 } 387 388 /* 389 * If we haven't found a queue, look for one that hasn't reached its 390 * maximum # outstanding i/os. 391 */ 392 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 393 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && 394 vq->vq_class[p].vqc_active < 395 vdev_queue_class_max_active(spa, p)) 396 return (p); 397 } 398 399 /* No eligible queued i/os */ 400 return (ZIO_PRIORITY_NUM_QUEUEABLE); 401 } 402 403 void 404 vdev_queue_init(vdev_t *vd) 405 { 406 vdev_queue_t *vq = &vd->vdev_queue; 407 zio_priority_t p; 408 409 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); 410 vq->vq_vdev = vd; 411 taskq_init_ent(&vd->vdev_queue.vq_io_search.io_tqent); 412 413 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare, 414 sizeof (zio_t), offsetof(struct zio, io_queue_node)); 415 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ), 416 vdev_queue_offset_compare, sizeof (zio_t), 417 offsetof(struct zio, io_offset_node)); 418 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE), 419 vdev_queue_offset_compare, sizeof (zio_t), 420 offsetof(struct zio, io_offset_node)); 421 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_TRIM), 422 vdev_queue_offset_compare, sizeof (zio_t), 423 offsetof(struct zio, io_offset_node)); 424 425 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 426 int (*compfn) (const void *, const void *); 427 428 /* 429 * The synchronous/trim i/o queues are dispatched in FIFO rather 430 * than LBA order. This provides more consistent latency for 431 * these i/os. 432 */ 433 if (p == ZIO_PRIORITY_SYNC_READ || 434 p == ZIO_PRIORITY_SYNC_WRITE || 435 p == ZIO_PRIORITY_TRIM) { 436 compfn = vdev_queue_timestamp_compare; 437 } else { 438 compfn = vdev_queue_offset_compare; 439 } 440 avl_create(vdev_queue_class_tree(vq, p), compfn, 441 sizeof (zio_t), offsetof(struct zio, io_queue_node)); 442 } 443 444 vq->vq_last_offset = 0; 445 } 446 447 void 448 vdev_queue_fini(vdev_t *vd) 449 { 450 vdev_queue_t *vq = &vd->vdev_queue; 451 452 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) 453 avl_destroy(vdev_queue_class_tree(vq, p)); 454 avl_destroy(&vq->vq_active_tree); 455 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ)); 456 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE)); 457 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_TRIM)); 458 459 mutex_destroy(&vq->vq_lock); 460 } 461 462 static void 463 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) 464 { 465 spa_t *spa = zio->io_spa; 466 spa_history_kstat_t *shk = &spa->spa_stats.io_history; 467 468 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 469 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio); 470 avl_add(vdev_queue_type_tree(vq, zio->io_type), zio); 471 472 if (shk->kstat != NULL) { 473 mutex_enter(&shk->lock); 474 kstat_waitq_enter(shk->kstat->ks_data); 475 mutex_exit(&shk->lock); 476 } 477 } 478 479 static void 480 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) 481 { 482 spa_t *spa = zio->io_spa; 483 spa_history_kstat_t *shk = &spa->spa_stats.io_history; 484 485 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 486 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio); 487 avl_remove(vdev_queue_type_tree(vq, zio->io_type), zio); 488 489 if (shk->kstat != NULL) { 490 mutex_enter(&shk->lock); 491 kstat_waitq_exit(shk->kstat->ks_data); 492 mutex_exit(&shk->lock); 493 } 494 } 495 496 static void 497 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) 498 { 499 spa_t *spa = zio->io_spa; 500 spa_history_kstat_t *shk = &spa->spa_stats.io_history; 501 502 ASSERT(MUTEX_HELD(&vq->vq_lock)); 503 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 504 vq->vq_class[zio->io_priority].vqc_active++; 505 avl_add(&vq->vq_active_tree, zio); 506 507 if (shk->kstat != NULL) { 508 mutex_enter(&shk->lock); 509 kstat_runq_enter(shk->kstat->ks_data); 510 mutex_exit(&shk->lock); 511 } 512 } 513 514 static void 515 vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) 516 { 517 spa_t *spa = zio->io_spa; 518 spa_history_kstat_t *shk = &spa->spa_stats.io_history; 519 520 ASSERT(MUTEX_HELD(&vq->vq_lock)); 521 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 522 vq->vq_class[zio->io_priority].vqc_active--; 523 avl_remove(&vq->vq_active_tree, zio); 524 525 if (shk->kstat != NULL) { 526 kstat_io_t *ksio = shk->kstat->ks_data; 527 528 mutex_enter(&shk->lock); 529 kstat_runq_exit(ksio); 530 if (zio->io_type == ZIO_TYPE_READ) { 531 ksio->reads++; 532 ksio->nread += zio->io_size; 533 } else if (zio->io_type == ZIO_TYPE_WRITE) { 534 ksio->writes++; 535 ksio->nwritten += zio->io_size; 536 } 537 mutex_exit(&shk->lock); 538 } 539 } 540 541 static void 542 vdev_queue_agg_io_done(zio_t *aio) 543 { 544 abd_free(aio->io_abd); 545 } 546 547 /* 548 * Compute the range spanned by two i/os, which is the endpoint of the last 549 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). 550 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); 551 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. 552 */ 553 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) 554 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) 555 556 /* 557 * Sufficiently adjacent io_offset's in ZIOs will be aggregated. We do this 558 * by creating a gang ABD from the adjacent ZIOs io_abd's. By using 559 * a gang ABD we avoid doing memory copies to and from the parent, 560 * child ZIOs. The gang ABD also accounts for gaps between adjacent 561 * io_offsets by simply getting the zero ABD for writes or allocating 562 * a new ABD for reads and placing them in the gang ABD as well. 563 */ 564 static zio_t * 565 vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) 566 { 567 zio_t *first, *last, *aio, *dio, *mandatory, *nio; 568 zio_link_t *zl = NULL; 569 uint64_t maxgap = 0; 570 uint64_t size; 571 uint64_t limit; 572 int maxblocksize; 573 boolean_t stretch = B_FALSE; 574 avl_tree_t *t = vdev_queue_type_tree(vq, zio->io_type); 575 enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; 576 uint64_t next_offset; 577 abd_t *abd; 578 579 maxblocksize = spa_maxblocksize(vq->vq_vdev->vdev_spa); 580 if (vq->vq_vdev->vdev_nonrot) 581 limit = zfs_vdev_aggregation_limit_non_rotating; 582 else 583 limit = zfs_vdev_aggregation_limit; 584 limit = MAX(MIN(limit, maxblocksize), 0); 585 586 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE || limit == 0) 587 return (NULL); 588 589 /* 590 * While TRIM commands could be aggregated based on offset this 591 * behavior is disabled until it's determined to be beneficial. 592 */ 593 if (zio->io_type == ZIO_TYPE_TRIM && !zfs_vdev_aggregate_trim) 594 return (NULL); 595 596 first = last = zio; 597 598 if (zio->io_type == ZIO_TYPE_READ) 599 maxgap = zfs_vdev_read_gap_limit; 600 601 /* 602 * We can aggregate I/Os that are sufficiently adjacent and of 603 * the same flavor, as expressed by the AGG_INHERIT flags. 604 * The latter requirement is necessary so that certain 605 * attributes of the I/O, such as whether it's a normal I/O 606 * or a scrub/resilver, can be preserved in the aggregate. 607 * We can include optional I/Os, but don't allow them 608 * to begin a range as they add no benefit in that situation. 609 */ 610 611 /* 612 * We keep track of the last non-optional I/O. 613 */ 614 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; 615 616 /* 617 * Walk backwards through sufficiently contiguous I/Os 618 * recording the last non-optional I/O. 619 */ 620 while ((dio = AVL_PREV(t, first)) != NULL && 621 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 622 IO_SPAN(dio, last) <= limit && 623 IO_GAP(dio, first) <= maxgap && 624 dio->io_type == zio->io_type) { 625 first = dio; 626 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) 627 mandatory = first; 628 } 629 630 /* 631 * Skip any initial optional I/Os. 632 */ 633 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { 634 first = AVL_NEXT(t, first); 635 ASSERT(first != NULL); 636 } 637 638 639 /* 640 * Walk forward through sufficiently contiguous I/Os. 641 * The aggregation limit does not apply to optional i/os, so that 642 * we can issue contiguous writes even if they are larger than the 643 * aggregation limit. 644 */ 645 while ((dio = AVL_NEXT(t, last)) != NULL && 646 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 647 (IO_SPAN(first, dio) <= limit || 648 (dio->io_flags & ZIO_FLAG_OPTIONAL)) && 649 IO_SPAN(first, dio) <= maxblocksize && 650 IO_GAP(last, dio) <= maxgap && 651 dio->io_type == zio->io_type) { 652 last = dio; 653 if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) 654 mandatory = last; 655 } 656 657 /* 658 * Now that we've established the range of the I/O aggregation 659 * we must decide what to do with trailing optional I/Os. 660 * For reads, there's nothing to do. While we are unable to 661 * aggregate further, it's possible that a trailing optional 662 * I/O would allow the underlying device to aggregate with 663 * subsequent I/Os. We must therefore determine if the next 664 * non-optional I/O is close enough to make aggregation 665 * worthwhile. 666 */ 667 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { 668 zio_t *nio = last; 669 while ((dio = AVL_NEXT(t, nio)) != NULL && 670 IO_GAP(nio, dio) == 0 && 671 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { 672 nio = dio; 673 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { 674 stretch = B_TRUE; 675 break; 676 } 677 } 678 } 679 680 if (stretch) { 681 /* 682 * We are going to include an optional io in our aggregated 683 * span, thus closing the write gap. Only mandatory i/os can 684 * start aggregated spans, so make sure that the next i/o 685 * after our span is mandatory. 686 */ 687 dio = AVL_NEXT(t, last); 688 dio->io_flags &= ~ZIO_FLAG_OPTIONAL; 689 } else { 690 /* do not include the optional i/o */ 691 while (last != mandatory && last != first) { 692 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); 693 last = AVL_PREV(t, last); 694 ASSERT(last != NULL); 695 } 696 } 697 698 if (first == last) 699 return (NULL); 700 701 size = IO_SPAN(first, last); 702 ASSERT3U(size, <=, maxblocksize); 703 704 abd = abd_alloc_gang_abd(); 705 if (abd == NULL) 706 return (NULL); 707 708 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, 709 abd, size, first->io_type, zio->io_priority, 710 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE, 711 vdev_queue_agg_io_done, NULL); 712 aio->io_timestamp = first->io_timestamp; 713 714 nio = first; 715 next_offset = first->io_offset; 716 do { 717 dio = nio; 718 nio = AVL_NEXT(t, dio); 719 zio_add_child(dio, aio); 720 vdev_queue_io_remove(vq, dio); 721 722 if (dio->io_offset != next_offset) { 723 /* allocate a buffer for a read gap */ 724 ASSERT3U(dio->io_type, ==, ZIO_TYPE_READ); 725 ASSERT3U(dio->io_offset, >, next_offset); 726 abd = abd_alloc_for_io( 727 dio->io_offset - next_offset, B_TRUE); 728 abd_gang_add(aio->io_abd, abd, B_TRUE); 729 } 730 if (dio->io_abd && 731 (dio->io_size != abd_get_size(dio->io_abd))) { 732 /* abd size not the same as IO size */ 733 ASSERT3U(abd_get_size(dio->io_abd), >, dio->io_size); 734 abd = abd_get_offset_size(dio->io_abd, 0, dio->io_size); 735 abd_gang_add(aio->io_abd, abd, B_TRUE); 736 } else { 737 if (dio->io_flags & ZIO_FLAG_NODATA) { 738 /* allocate a buffer for a write gap */ 739 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); 740 ASSERT3P(dio->io_abd, ==, NULL); 741 abd_gang_add(aio->io_abd, 742 abd_get_zeros(dio->io_size), B_TRUE); 743 } else { 744 /* 745 * We pass B_FALSE to abd_gang_add() 746 * because we did not allocate a new 747 * ABD, so it is assumed the caller 748 * will free this ABD. 749 */ 750 abd_gang_add(aio->io_abd, dio->io_abd, 751 B_FALSE); 752 } 753 } 754 next_offset = dio->io_offset + dio->io_size; 755 } while (dio != last); 756 ASSERT3U(abd_get_size(aio->io_abd), ==, aio->io_size); 757 758 /* 759 * We need to drop the vdev queue's lock during zio_execute() to 760 * avoid a deadlock that we could encounter due to lock order 761 * reversal between vq_lock and io_lock in zio_change_priority(). 762 */ 763 mutex_exit(&vq->vq_lock); 764 while ((dio = zio_walk_parents(aio, &zl)) != NULL) { 765 ASSERT3U(dio->io_type, ==, aio->io_type); 766 767 zio_vdev_io_bypass(dio); 768 zio_execute(dio); 769 } 770 mutex_enter(&vq->vq_lock); 771 772 return (aio); 773 } 774 775 static zio_t * 776 vdev_queue_io_to_issue(vdev_queue_t *vq) 777 { 778 zio_t *zio, *aio; 779 zio_priority_t p; 780 avl_index_t idx; 781 avl_tree_t *tree; 782 783 again: 784 ASSERT(MUTEX_HELD(&vq->vq_lock)); 785 786 p = vdev_queue_class_to_issue(vq); 787 788 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { 789 /* No eligible queued i/os */ 790 return (NULL); 791 } 792 793 /* 794 * For LBA-ordered queues (async / scrub / initializing), issue the 795 * i/o which follows the most recently issued i/o in LBA (offset) order. 796 * 797 * For FIFO queues (sync/trim), issue the i/o with the lowest timestamp. 798 */ 799 tree = vdev_queue_class_tree(vq, p); 800 vq->vq_io_search.io_timestamp = 0; 801 vq->vq_io_search.io_offset = vq->vq_last_offset - 1; 802 VERIFY3P(avl_find(tree, &vq->vq_io_search, &idx), ==, NULL); 803 zio = avl_nearest(tree, idx, AVL_AFTER); 804 if (zio == NULL) 805 zio = avl_first(tree); 806 ASSERT3U(zio->io_priority, ==, p); 807 808 aio = vdev_queue_aggregate(vq, zio); 809 if (aio != NULL) 810 zio = aio; 811 else 812 vdev_queue_io_remove(vq, zio); 813 814 /* 815 * If the I/O is or was optional and therefore has no data, we need to 816 * simply discard it. We need to drop the vdev queue's lock to avoid a 817 * deadlock that we could encounter since this I/O will complete 818 * immediately. 819 */ 820 if (zio->io_flags & ZIO_FLAG_NODATA) { 821 mutex_exit(&vq->vq_lock); 822 zio_vdev_io_bypass(zio); 823 zio_execute(zio); 824 mutex_enter(&vq->vq_lock); 825 goto again; 826 } 827 828 vdev_queue_pending_add(vq, zio); 829 vq->vq_last_offset = zio->io_offset + zio->io_size; 830 831 return (zio); 832 } 833 834 zio_t * 835 vdev_queue_io(zio_t *zio) 836 { 837 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 838 zio_t *nio; 839 840 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) 841 return (zio); 842 843 /* 844 * Children i/os inherent their parent's priority, which might 845 * not match the child's i/o type. Fix it up here. 846 */ 847 if (zio->io_type == ZIO_TYPE_READ) { 848 ASSERT(zio->io_priority != ZIO_PRIORITY_TRIM); 849 850 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && 851 zio->io_priority != ZIO_PRIORITY_ASYNC_READ && 852 zio->io_priority != ZIO_PRIORITY_SCRUB && 853 zio->io_priority != ZIO_PRIORITY_REMOVAL && 854 zio->io_priority != ZIO_PRIORITY_INITIALIZING && 855 zio->io_priority != ZIO_PRIORITY_REBUILD) { 856 zio->io_priority = ZIO_PRIORITY_ASYNC_READ; 857 } 858 } else if (zio->io_type == ZIO_TYPE_WRITE) { 859 ASSERT(zio->io_priority != ZIO_PRIORITY_TRIM); 860 861 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && 862 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE && 863 zio->io_priority != ZIO_PRIORITY_REMOVAL && 864 zio->io_priority != ZIO_PRIORITY_INITIALIZING && 865 zio->io_priority != ZIO_PRIORITY_REBUILD) { 866 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; 867 } 868 } else { 869 ASSERT(zio->io_type == ZIO_TYPE_TRIM); 870 ASSERT(zio->io_priority == ZIO_PRIORITY_TRIM); 871 } 872 873 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; 874 875 mutex_enter(&vq->vq_lock); 876 zio->io_timestamp = gethrtime(); 877 vdev_queue_io_add(vq, zio); 878 nio = vdev_queue_io_to_issue(vq); 879 mutex_exit(&vq->vq_lock); 880 881 if (nio == NULL) 882 return (NULL); 883 884 if (nio->io_done == vdev_queue_agg_io_done) { 885 zio_nowait(nio); 886 return (NULL); 887 } 888 889 return (nio); 890 } 891 892 void 893 vdev_queue_io_done(zio_t *zio) 894 { 895 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 896 zio_t *nio; 897 898 mutex_enter(&vq->vq_lock); 899 900 vdev_queue_pending_remove(vq, zio); 901 902 zio->io_delta = gethrtime() - zio->io_timestamp; 903 vq->vq_io_complete_ts = gethrtime(); 904 vq->vq_io_delta_ts = vq->vq_io_complete_ts - zio->io_timestamp; 905 906 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { 907 mutex_exit(&vq->vq_lock); 908 if (nio->io_done == vdev_queue_agg_io_done) { 909 zio_nowait(nio); 910 } else { 911 zio_vdev_io_reissue(nio); 912 zio_execute(nio); 913 } 914 mutex_enter(&vq->vq_lock); 915 } 916 917 mutex_exit(&vq->vq_lock); 918 } 919 920 void 921 vdev_queue_change_io_priority(zio_t *zio, zio_priority_t priority) 922 { 923 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 924 avl_tree_t *tree; 925 926 /* 927 * ZIO_PRIORITY_NOW is used by the vdev cache code and the aggregate zio 928 * code to issue IOs without adding them to the vdev queue. In this 929 * case, the zio is already going to be issued as quickly as possible 930 * and so it doesn't need any reprioritization to help. 931 */ 932 if (zio->io_priority == ZIO_PRIORITY_NOW) 933 return; 934 935 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 936 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 937 938 if (zio->io_type == ZIO_TYPE_READ) { 939 if (priority != ZIO_PRIORITY_SYNC_READ && 940 priority != ZIO_PRIORITY_ASYNC_READ && 941 priority != ZIO_PRIORITY_SCRUB) 942 priority = ZIO_PRIORITY_ASYNC_READ; 943 } else { 944 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 945 if (priority != ZIO_PRIORITY_SYNC_WRITE && 946 priority != ZIO_PRIORITY_ASYNC_WRITE) 947 priority = ZIO_PRIORITY_ASYNC_WRITE; 948 } 949 950 mutex_enter(&vq->vq_lock); 951 952 /* 953 * If the zio is in none of the queues we can simply change 954 * the priority. If the zio is waiting to be submitted we must 955 * remove it from the queue and re-insert it with the new priority. 956 * Otherwise, the zio is currently active and we cannot change its 957 * priority. 958 */ 959 tree = vdev_queue_class_tree(vq, zio->io_priority); 960 if (avl_find(tree, zio, NULL) == zio) { 961 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio); 962 zio->io_priority = priority; 963 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio); 964 } else if (avl_find(&vq->vq_active_tree, zio, NULL) != zio) { 965 zio->io_priority = priority; 966 } 967 968 mutex_exit(&vq->vq_lock); 969 } 970 971 /* 972 * As these two methods are only used for load calculations we're not 973 * concerned if we get an incorrect value on 32bit platforms due to lack of 974 * vq_lock mutex use here, instead we prefer to keep it lock free for 975 * performance. 976 */ 977 int 978 vdev_queue_length(vdev_t *vd) 979 { 980 return (avl_numnodes(&vd->vdev_queue.vq_active_tree)); 981 } 982 983 uint64_t 984 vdev_queue_last_offset(vdev_t *vd) 985 { 986 return (vd->vdev_queue.vq_last_offset); 987 } 988 989 /* BEGIN CSTYLED */ 990 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, aggregation_limit, INT, ZMOD_RW, 991 "Max vdev I/O aggregation size"); 992 993 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, aggregation_limit_non_rotating, INT, ZMOD_RW, 994 "Max vdev I/O aggregation size for non-rotating media"); 995 996 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, aggregate_trim, INT, ZMOD_RW, 997 "Allow TRIM I/O to be aggregated"); 998 999 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, read_gap_limit, INT, ZMOD_RW, 1000 "Aggregate read I/O over gap"); 1001 1002 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, write_gap_limit, INT, ZMOD_RW, 1003 "Aggregate write I/O over gap"); 1004 1005 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, max_active, INT, ZMOD_RW, 1006 "Maximum number of active I/Os per vdev"); 1007 1008 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_write_active_max_dirty_percent, INT, ZMOD_RW, 1009 "Async write concurrency max threshold"); 1010 1011 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_write_active_min_dirty_percent, INT, ZMOD_RW, 1012 "Async write concurrency min threshold"); 1013 1014 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_read_max_active, INT, ZMOD_RW, 1015 "Max active async read I/Os per vdev"); 1016 1017 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_read_min_active, INT, ZMOD_RW, 1018 "Min active async read I/Os per vdev"); 1019 1020 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_write_max_active, INT, ZMOD_RW, 1021 "Max active async write I/Os per vdev"); 1022 1023 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_write_min_active, INT, ZMOD_RW, 1024 "Min active async write I/Os per vdev"); 1025 1026 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, initializing_max_active, INT, ZMOD_RW, 1027 "Max active initializing I/Os per vdev"); 1028 1029 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, initializing_min_active, INT, ZMOD_RW, 1030 "Min active initializing I/Os per vdev"); 1031 1032 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, removal_max_active, INT, ZMOD_RW, 1033 "Max active removal I/Os per vdev"); 1034 1035 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, removal_min_active, INT, ZMOD_RW, 1036 "Min active removal I/Os per vdev"); 1037 1038 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, scrub_max_active, INT, ZMOD_RW, 1039 "Max active scrub I/Os per vdev"); 1040 1041 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, scrub_min_active, INT, ZMOD_RW, 1042 "Min active scrub I/Os per vdev"); 1043 1044 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, sync_read_max_active, INT, ZMOD_RW, 1045 "Max active sync read I/Os per vdev"); 1046 1047 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, sync_read_min_active, INT, ZMOD_RW, 1048 "Min active sync read I/Os per vdev"); 1049 1050 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, sync_write_max_active, INT, ZMOD_RW, 1051 "Max active sync write I/Os per vdev"); 1052 1053 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, sync_write_min_active, INT, ZMOD_RW, 1054 "Min active sync write I/Os per vdev"); 1055 1056 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, trim_max_active, INT, ZMOD_RW, 1057 "Max active trim/discard I/Os per vdev"); 1058 1059 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, trim_min_active, INT, ZMOD_RW, 1060 "Min active trim/discard I/Os per vdev"); 1061 1062 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, rebuild_max_active, INT, ZMOD_RW, 1063 "Max active rebuild I/Os per vdev"); 1064 1065 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, rebuild_min_active, INT, ZMOD_RW, 1066 "Min active rebuild I/Os per vdev"); 1067 1068 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, queue_depth_pct, INT, ZMOD_RW, 1069 "Queue depth percentage for each top-level vdev"); 1070 /* END CSTYLED */ 1071