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, 2014 by Delphix. All rights reserved. 28 * Copyright (c) 2014 Integros [integros.com] 29 */ 30 31#include <sys/zfs_context.h> 32#include <sys/vdev_impl.h> 33#include <sys/spa_impl.h> 34#include <sys/zio.h> 35#include <sys/avl.h> 36#include <sys/dsl_pool.h> 37 38/* 39 * ZFS I/O Scheduler 40 * --------------- 41 * 42 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The 43 * I/O scheduler determines when and in what order those operations are 44 * issued. The I/O scheduler divides operations into six I/O classes 45 * prioritized in the following order: sync read, sync write, async read, 46 * async write, scrub/resilver and trim. Each queue defines the minimum and 47 * maximum number of concurrent operations that may be issued to the device. 48 * In addition, the device has an aggregate maximum. Note that the sum of the 49 * per-queue minimums must not exceed the aggregate maximum, and if the 50 * aggregate maximum is equal to or greater than the sum of the per-queue 51 * maximums, the per-queue minimum has no effect. 52 * 53 * For many physical devices, throughput increases with the number of 54 * concurrent operations, but latency typically suffers. Further, physical 55 * devices typically have a limit at which more concurrent operations have no 56 * effect on throughput or can actually cause it to decrease. 57 * 58 * The scheduler selects the next operation to issue by first looking for an 59 * I/O class whose minimum has not been satisfied. Once all are satisfied and 60 * the aggregate maximum has not been hit, the scheduler looks for classes 61 * whose maximum has not been satisfied. Iteration through the I/O classes is 62 * done in the order specified above. No further operations are issued if the 63 * aggregate maximum number of concurrent operations has been hit or if there 64 * are no operations queued for an I/O class that has not hit its maximum. 65 * Every time an I/O is queued or an operation completes, the I/O scheduler 66 * looks for new operations to issue. 67 * 68 * All I/O classes have a fixed maximum number of outstanding operations 69 * except for the async write class. Asynchronous writes represent the data 70 * that is committed to stable storage during the syncing stage for 71 * transaction groups (see txg.c). Transaction groups enter the syncing state 72 * periodically so the number of queued async writes will quickly burst up and 73 * then bleed down to zero. Rather than servicing them as quickly as possible, 74 * the I/O scheduler changes the maximum number of active async write I/Os 75 * according to the amount of dirty data in the pool (see dsl_pool.c). Since 76 * both throughput and latency typically increase with the number of 77 * concurrent operations issued to physical devices, reducing the burstiness 78 * in the number of concurrent operations also stabilizes the response time of 79 * operations from other -- and in particular synchronous -- queues. In broad 80 * strokes, the I/O scheduler will issue more concurrent operations from the 81 * async write queue as there's more dirty data in the pool. 82 * 83 * Async Writes 84 * 85 * The number of concurrent operations issued for the async write I/O class 86 * follows a piece-wise linear function defined by a few adjustable points. 87 * 88 * | o---------| <-- zfs_vdev_async_write_max_active 89 * ^ | /^ | 90 * | | / | | 91 * active | / | | 92 * I/O | / | | 93 * count | / | | 94 * | / | | 95 * |------------o | | <-- zfs_vdev_async_write_min_active 96 * 0|____________^______|_________| 97 * 0% | | 100% of zfs_dirty_data_max 98 * | | 99 * | `-- zfs_vdev_async_write_active_max_dirty_percent 100 * `--------- zfs_vdev_async_write_active_min_dirty_percent 101 * 102 * Until the amount of dirty data exceeds a minimum percentage of the dirty 103 * data allowed in the pool, the I/O scheduler will limit the number of 104 * concurrent operations to the minimum. As that threshold is crossed, the 105 * number of concurrent operations issued increases linearly to the maximum at 106 * the specified maximum percentage of the dirty data allowed in the pool. 107 * 108 * Ideally, the amount of dirty data on a busy pool will stay in the sloped 109 * part of the function between zfs_vdev_async_write_active_min_dirty_percent 110 * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the 111 * maximum percentage, this indicates that the rate of incoming data is 112 * greater than the rate that the backend storage can handle. In this case, we 113 * must further throttle incoming writes (see dmu_tx_delay() for details). 114 */ 115 116/* 117 * The maximum number of I/Os active to each device. Ideally, this will be >= 118 * the sum of each queue's max_active. It must be at least the sum of each 119 * queue's min_active. 120 */ 121uint32_t zfs_vdev_max_active = 1000; 122 123/* 124 * Per-queue limits on the number of I/Os active to each device. If the 125 * sum of the queue's max_active is < zfs_vdev_max_active, then the 126 * min_active comes into play. We will send min_active from each queue, 127 * and then select from queues in the order defined by zio_priority_t. 128 * 129 * In general, smaller max_active's will lead to lower latency of synchronous 130 * operations. Larger max_active's may lead to higher overall throughput, 131 * depending on underlying storage. 132 * 133 * The ratio of the queues' max_actives determines the balance of performance 134 * between reads, writes, and scrubs. E.g., increasing 135 * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete 136 * more quickly, but reads and writes to have higher latency and lower 137 * throughput. 138 */ 139uint32_t zfs_vdev_sync_read_min_active = 10; 140uint32_t zfs_vdev_sync_read_max_active = 10; 141uint32_t zfs_vdev_sync_write_min_active = 10; 142uint32_t zfs_vdev_sync_write_max_active = 10; 143uint32_t zfs_vdev_async_read_min_active = 1; 144uint32_t zfs_vdev_async_read_max_active = 3; 145uint32_t zfs_vdev_async_write_min_active = 1; 146uint32_t zfs_vdev_async_write_max_active = 10; 147uint32_t zfs_vdev_scrub_min_active = 1; 148uint32_t zfs_vdev_scrub_max_active = 2; 149uint32_t zfs_vdev_trim_min_active = 1; 150/* 151 * TRIM max active is large in comparison to the other values due to the fact 152 * that TRIM IOs are coalesced at the device layer. This value is set such 153 * that a typical SSD can process the queued IOs in a single request. 154 */ 155uint32_t zfs_vdev_trim_max_active = 64; 156 157 158/* 159 * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent 160 * dirty data, use zfs_vdev_async_write_min_active. When it has more than 161 * zfs_vdev_async_write_active_max_dirty_percent, use 162 * zfs_vdev_async_write_max_active. The value is linearly interpolated 163 * between min and max. 164 */ 165int zfs_vdev_async_write_active_min_dirty_percent = 30; 166int zfs_vdev_async_write_active_max_dirty_percent = 60; 167 168/* 169 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O. 170 * For read I/Os, we also aggregate across small adjacency gaps; for writes 171 * we include spans of optional I/Os to aid aggregation at the disk even when 172 * they aren't able to help us aggregate at this level. 173 */ 174int zfs_vdev_aggregation_limit = SPA_OLD_MAXBLOCKSIZE; 175int zfs_vdev_read_gap_limit = 32 << 10; 176int zfs_vdev_write_gap_limit = 4 << 10; 177 178#ifdef __FreeBSD__ 179SYSCTL_DECL(_vfs_zfs_vdev); 180 181static int sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS); 182SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_min_dirty_percent, 183 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int), 184 sysctl_zfs_async_write_active_min_dirty_percent, "I", 185 "Percentage of async write dirty data below which " 186 "async_write_min_active is used."); 187 188static int sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS); 189SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_max_dirty_percent, 190 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int), 191 sysctl_zfs_async_write_active_max_dirty_percent, "I", 192 "Percentage of async write dirty data above which " 193 "async_write_max_active is used."); 194 195SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RWTUN, 196 &zfs_vdev_max_active, 0, 197 "The maximum number of I/Os of all types active for each device."); 198 199#define ZFS_VDEV_QUEUE_KNOB_MIN(name) \ 200SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, CTLFLAG_RWTUN,\ 201 &zfs_vdev_ ## name ## _min_active, 0, \ 202 "Initial number of I/O requests of type " #name \ 203 " active for each device"); 204 205#define ZFS_VDEV_QUEUE_KNOB_MAX(name) \ 206SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, CTLFLAG_RWTUN,\ 207 &zfs_vdev_ ## name ## _max_active, 0, \ 208 "Maximum number of I/O requests of type " #name \ 209 " active for each device"); 210 211ZFS_VDEV_QUEUE_KNOB_MIN(sync_read); 212ZFS_VDEV_QUEUE_KNOB_MAX(sync_read); 213ZFS_VDEV_QUEUE_KNOB_MIN(sync_write); 214ZFS_VDEV_QUEUE_KNOB_MAX(sync_write); 215ZFS_VDEV_QUEUE_KNOB_MIN(async_read); 216ZFS_VDEV_QUEUE_KNOB_MAX(async_read); 217ZFS_VDEV_QUEUE_KNOB_MIN(async_write); 218ZFS_VDEV_QUEUE_KNOB_MAX(async_write); 219ZFS_VDEV_QUEUE_KNOB_MIN(scrub); 220ZFS_VDEV_QUEUE_KNOB_MAX(scrub); 221ZFS_VDEV_QUEUE_KNOB_MIN(trim); 222ZFS_VDEV_QUEUE_KNOB_MAX(trim); 223 224#undef ZFS_VDEV_QUEUE_KNOB 225 226SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RWTUN, 227 &zfs_vdev_aggregation_limit, 0, 228 "I/O requests are aggregated up to this size"); 229SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RWTUN, 230 &zfs_vdev_read_gap_limit, 0, 231 "Acceptable gap between two reads being aggregated"); 232SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RWTUN, 233 &zfs_vdev_write_gap_limit, 0, 234 "Acceptable gap between two writes being aggregated"); 235 236static int 237sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS) 238{ 239 int val, err; 240 241 val = zfs_vdev_async_write_active_min_dirty_percent; 242 err = sysctl_handle_int(oidp, &val, 0, req); 243 if (err != 0 || req->newptr == NULL) 244 return (err); 245 246 if (val < 0 || val > 100 || 247 val >= zfs_vdev_async_write_active_max_dirty_percent) 248 return (EINVAL); 249 250 zfs_vdev_async_write_active_min_dirty_percent = val; 251 252 return (0); 253} 254 255static int 256sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS) 257{ 258 int val, err; 259 260 val = zfs_vdev_async_write_active_max_dirty_percent; 261 err = sysctl_handle_int(oidp, &val, 0, req); 262 if (err != 0 || req->newptr == NULL) 263 return (err); 264 265 if (val < 0 || val > 100 || 266 val <= zfs_vdev_async_write_active_min_dirty_percent) 267 return (EINVAL); 268 269 zfs_vdev_async_write_active_max_dirty_percent = val; 270 271 return (0); 272} 273#endif 274 275int 276vdev_queue_offset_compare(const void *x1, const void *x2) 277{ 278 const zio_t *z1 = x1; 279 const zio_t *z2 = x2; 280 281 if (z1->io_offset < z2->io_offset) 282 return (-1); 283 if (z1->io_offset > z2->io_offset) 284 return (1); 285 286 if (z1 < z2) 287 return (-1); 288 if (z1 > z2) 289 return (1); 290 291 return (0); 292} 293 294static inline avl_tree_t * 295vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p) 296{ 297 return (&vq->vq_class[p].vqc_queued_tree); 298} 299 300static inline avl_tree_t * 301vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t) 302{ 303 if (t == ZIO_TYPE_READ) 304 return (&vq->vq_read_offset_tree); 305 else if (t == ZIO_TYPE_WRITE) 306 return (&vq->vq_write_offset_tree); 307 else 308 return (NULL); 309} 310 311int 312vdev_queue_timestamp_compare(const void *x1, const void *x2) 313{ 314 const zio_t *z1 = x1; 315 const zio_t *z2 = x2; 316 317 if (z1->io_timestamp < z2->io_timestamp) 318 return (-1); 319 if (z1->io_timestamp > z2->io_timestamp) 320 return (1); 321 322 if (z1->io_offset < z2->io_offset) 323 return (-1); 324 if (z1->io_offset > z2->io_offset) 325 return (1); 326 327 if (z1 < z2) 328 return (-1); 329 if (z1 > z2) 330 return (1); 331 332 return (0); 333} 334 335void 336vdev_queue_init(vdev_t *vd) 337{ 338 vdev_queue_t *vq = &vd->vdev_queue; 339 340 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); 341 vq->vq_vdev = vd; 342 343 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare, 344 sizeof (zio_t), offsetof(struct zio, io_queue_node)); 345 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ), 346 vdev_queue_offset_compare, sizeof (zio_t), 347 offsetof(struct zio, io_offset_node)); 348 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE), 349 vdev_queue_offset_compare, sizeof (zio_t), 350 offsetof(struct zio, io_offset_node)); 351 352 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 353 int (*compfn) (const void *, const void *); 354 355 /* 356 * The synchronous i/o queues are dispatched in FIFO rather 357 * than LBA order. This provides more consistent latency for 358 * these i/os. 359 */ 360 if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE) 361 compfn = vdev_queue_timestamp_compare; 362 else 363 compfn = vdev_queue_offset_compare; 364 365 avl_create(vdev_queue_class_tree(vq, p), compfn, 366 sizeof (zio_t), offsetof(struct zio, io_queue_node)); 367 } 368 369 vq->vq_lastoffset = 0; 370} 371 372void 373vdev_queue_fini(vdev_t *vd) 374{ 375 vdev_queue_t *vq = &vd->vdev_queue; 376 377 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) 378 avl_destroy(vdev_queue_class_tree(vq, p)); 379 avl_destroy(&vq->vq_active_tree); 380 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ)); 381 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE)); 382 383 mutex_destroy(&vq->vq_lock); 384} 385 386static void 387vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) 388{ 389 spa_t *spa = zio->io_spa; 390 avl_tree_t *qtt; 391 ASSERT(MUTEX_HELD(&vq->vq_lock)); 392 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 393 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio); 394 qtt = vdev_queue_type_tree(vq, zio->io_type); 395 if (qtt) 396 avl_add(qtt, zio); 397 398#ifdef illumos 399 mutex_enter(&spa->spa_iokstat_lock); 400 spa->spa_queue_stats[zio->io_priority].spa_queued++; 401 if (spa->spa_iokstat != NULL) 402 kstat_waitq_enter(spa->spa_iokstat->ks_data); 403 mutex_exit(&spa->spa_iokstat_lock); 404#endif 405} 406 407static void 408vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) 409{ 410 spa_t *spa = zio->io_spa; 411 avl_tree_t *qtt; 412 ASSERT(MUTEX_HELD(&vq->vq_lock)); 413 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 414 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio); 415 qtt = vdev_queue_type_tree(vq, zio->io_type); 416 if (qtt) 417 avl_remove(qtt, zio); 418 419#ifdef illumos 420 mutex_enter(&spa->spa_iokstat_lock); 421 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0); 422 spa->spa_queue_stats[zio->io_priority].spa_queued--; 423 if (spa->spa_iokstat != NULL) 424 kstat_waitq_exit(spa->spa_iokstat->ks_data); 425 mutex_exit(&spa->spa_iokstat_lock); 426#endif 427} 428 429static void 430vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) 431{ 432 spa_t *spa = zio->io_spa; 433 ASSERT(MUTEX_HELD(&vq->vq_lock)); 434 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 435 vq->vq_class[zio->io_priority].vqc_active++; 436 avl_add(&vq->vq_active_tree, zio); 437 438#ifdef illumos 439 mutex_enter(&spa->spa_iokstat_lock); 440 spa->spa_queue_stats[zio->io_priority].spa_active++; 441 if (spa->spa_iokstat != NULL) 442 kstat_runq_enter(spa->spa_iokstat->ks_data); 443 mutex_exit(&spa->spa_iokstat_lock); 444#endif 445} 446 447static void 448vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) 449{ 450 spa_t *spa = zio->io_spa; 451 ASSERT(MUTEX_HELD(&vq->vq_lock)); 452 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 453 vq->vq_class[zio->io_priority].vqc_active--; 454 avl_remove(&vq->vq_active_tree, zio); 455 456#ifdef illumos 457 mutex_enter(&spa->spa_iokstat_lock); 458 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0); 459 spa->spa_queue_stats[zio->io_priority].spa_active--; 460 if (spa->spa_iokstat != NULL) { 461 kstat_io_t *ksio = spa->spa_iokstat->ks_data; 462 463 kstat_runq_exit(spa->spa_iokstat->ks_data); 464 if (zio->io_type == ZIO_TYPE_READ) { 465 ksio->reads++; 466 ksio->nread += zio->io_size; 467 } else if (zio->io_type == ZIO_TYPE_WRITE) { 468 ksio->writes++; 469 ksio->nwritten += zio->io_size; 470 } 471 } 472 mutex_exit(&spa->spa_iokstat_lock); 473#endif 474} 475 476static void 477vdev_queue_agg_io_done(zio_t *aio) 478{ 479 if (aio->io_type == ZIO_TYPE_READ) { 480 zio_t *pio; 481 while ((pio = zio_walk_parents(aio)) != NULL) { 482 bcopy((char *)aio->io_data + (pio->io_offset - 483 aio->io_offset), pio->io_data, pio->io_size); 484 } 485 } 486 487 zio_buf_free(aio->io_data, aio->io_size); 488} 489 490static int 491vdev_queue_class_min_active(zio_priority_t p) 492{ 493 switch (p) { 494 case ZIO_PRIORITY_SYNC_READ: 495 return (zfs_vdev_sync_read_min_active); 496 case ZIO_PRIORITY_SYNC_WRITE: 497 return (zfs_vdev_sync_write_min_active); 498 case ZIO_PRIORITY_ASYNC_READ: 499 return (zfs_vdev_async_read_min_active); 500 case ZIO_PRIORITY_ASYNC_WRITE: 501 return (zfs_vdev_async_write_min_active); 502 case ZIO_PRIORITY_SCRUB: 503 return (zfs_vdev_scrub_min_active); 504 case ZIO_PRIORITY_TRIM: 505 return (zfs_vdev_trim_min_active); 506 default: 507 panic("invalid priority %u", p); 508 return (0); 509 } 510} 511 512static __noinline int 513vdev_queue_max_async_writes(spa_t *spa) 514{ 515 int writes; 516 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total; 517 uint64_t min_bytes = zfs_dirty_data_max * 518 zfs_vdev_async_write_active_min_dirty_percent / 100; 519 uint64_t max_bytes = zfs_dirty_data_max * 520 zfs_vdev_async_write_active_max_dirty_percent / 100; 521 522 /* 523 * Sync tasks correspond to interactive user actions. To reduce the 524 * execution time of those actions we push data out as fast as possible. 525 */ 526 if (spa_has_pending_synctask(spa)) { 527 return (zfs_vdev_async_write_max_active); 528 } 529 530 if (dirty < min_bytes) 531 return (zfs_vdev_async_write_min_active); 532 if (dirty > max_bytes) 533 return (zfs_vdev_async_write_max_active); 534 535 /* 536 * linear interpolation: 537 * slope = (max_writes - min_writes) / (max_bytes - min_bytes) 538 * move right by min_bytes 539 * move up by min_writes 540 */ 541 writes = (dirty - min_bytes) * 542 (zfs_vdev_async_write_max_active - 543 zfs_vdev_async_write_min_active) / 544 (max_bytes - min_bytes) + 545 zfs_vdev_async_write_min_active; 546 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); 547 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); 548 return (writes); 549} 550 551static int 552vdev_queue_class_max_active(spa_t *spa, zio_priority_t p) 553{ 554 switch (p) { 555 case ZIO_PRIORITY_SYNC_READ: 556 return (zfs_vdev_sync_read_max_active); 557 case ZIO_PRIORITY_SYNC_WRITE: 558 return (zfs_vdev_sync_write_max_active); 559 case ZIO_PRIORITY_ASYNC_READ: 560 return (zfs_vdev_async_read_max_active); 561 case ZIO_PRIORITY_ASYNC_WRITE: 562 return (vdev_queue_max_async_writes(spa)); 563 case ZIO_PRIORITY_SCRUB: 564 return (zfs_vdev_scrub_max_active); 565 case ZIO_PRIORITY_TRIM: 566 return (zfs_vdev_trim_max_active); 567 default: 568 panic("invalid priority %u", p); 569 return (0); 570 } 571} 572 573/* 574 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if 575 * there is no eligible class. 576 */ 577static zio_priority_t 578vdev_queue_class_to_issue(vdev_queue_t *vq) 579{ 580 spa_t *spa = vq->vq_vdev->vdev_spa; 581 zio_priority_t p; 582 583 ASSERT(MUTEX_HELD(&vq->vq_lock)); 584 585 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active) 586 return (ZIO_PRIORITY_NUM_QUEUEABLE); 587 588 /* find a queue that has not reached its minimum # outstanding i/os */ 589 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 590 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && 591 vq->vq_class[p].vqc_active < 592 vdev_queue_class_min_active(p)) 593 return (p); 594 } 595 596 /* 597 * If we haven't found a queue, look for one that hasn't reached its 598 * maximum # outstanding i/os. 599 */ 600 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 601 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && 602 vq->vq_class[p].vqc_active < 603 vdev_queue_class_max_active(spa, p)) 604 return (p); 605 } 606 607 /* No eligible queued i/os */ 608 return (ZIO_PRIORITY_NUM_QUEUEABLE); 609} 610 611/* 612 * Compute the range spanned by two i/os, which is the endpoint of the last 613 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). 614 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); 615 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. 616 */ 617#define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) 618#define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) 619 620static zio_t * 621vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) 622{ 623 zio_t *first, *last, *aio, *dio, *mandatory, *nio; 624 uint64_t maxgap = 0; 625 uint64_t size; 626 boolean_t stretch; 627 avl_tree_t *t; 628 enum zio_flag flags; 629 630 ASSERT(MUTEX_HELD(&vq->vq_lock)); 631 632 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE) 633 return (NULL); 634 635 first = last = zio; 636 637 if (zio->io_type == ZIO_TYPE_READ) 638 maxgap = zfs_vdev_read_gap_limit; 639 640 /* 641 * We can aggregate I/Os that are sufficiently adjacent and of 642 * the same flavor, as expressed by the AGG_INHERIT flags. 643 * The latter requirement is necessary so that certain 644 * attributes of the I/O, such as whether it's a normal I/O 645 * or a scrub/resilver, can be preserved in the aggregate. 646 * We can include optional I/Os, but don't allow them 647 * to begin a range as they add no benefit in that situation. 648 */ 649 650 /* 651 * We keep track of the last non-optional I/O. 652 */ 653 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; 654 655 /* 656 * Walk backwards through sufficiently contiguous I/Os 657 * recording the last non-option I/O. 658 */ 659 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; 660 t = vdev_queue_type_tree(vq, zio->io_type); 661 while (t != NULL && (dio = AVL_PREV(t, first)) != NULL && 662 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 663 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit && 664 IO_GAP(dio, first) <= maxgap) { 665 first = dio; 666 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) 667 mandatory = first; 668 } 669 670 /* 671 * Skip any initial optional I/Os. 672 */ 673 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { 674 first = AVL_NEXT(t, first); 675 ASSERT(first != NULL); 676 } 677 678 /* 679 * Walk forward through sufficiently contiguous I/Os. 680 */ 681 while ((dio = AVL_NEXT(t, last)) != NULL && 682 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 683 IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit && 684 IO_GAP(last, dio) <= maxgap) { 685 last = dio; 686 if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) 687 mandatory = last; 688 } 689 690 /* 691 * Now that we've established the range of the I/O aggregation 692 * we must decide what to do with trailing optional I/Os. 693 * For reads, there's nothing to do. While we are unable to 694 * aggregate further, it's possible that a trailing optional 695 * I/O would allow the underlying device to aggregate with 696 * subsequent I/Os. We must therefore determine if the next 697 * non-optional I/O is close enough to make aggregation 698 * worthwhile. 699 */ 700 stretch = B_FALSE; 701 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { 702 zio_t *nio = last; 703 while ((dio = AVL_NEXT(t, nio)) != NULL && 704 IO_GAP(nio, dio) == 0 && 705 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { 706 nio = dio; 707 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { 708 stretch = B_TRUE; 709 break; 710 } 711 } 712 } 713 714 if (stretch) { 715 /* This may be a no-op. */ 716 dio = AVL_NEXT(t, last); 717 dio->io_flags &= ~ZIO_FLAG_OPTIONAL; 718 } else { 719 while (last != mandatory && last != first) { 720 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); 721 last = AVL_PREV(t, last); 722 ASSERT(last != NULL); 723 } 724 } 725 726 if (first == last) 727 return (NULL); 728 729 size = IO_SPAN(first, last); 730 ASSERT3U(size, <=, zfs_vdev_aggregation_limit); 731 732 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, 733 zio_buf_alloc(size), size, first->io_type, zio->io_priority, 734 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE, 735 vdev_queue_agg_io_done, NULL); 736 aio->io_timestamp = first->io_timestamp; 737 738 nio = first; 739 do { 740 dio = nio; 741 nio = AVL_NEXT(t, dio); 742 ASSERT3U(dio->io_type, ==, aio->io_type); 743 744 if (dio->io_flags & ZIO_FLAG_NODATA) { 745 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); 746 bzero((char *)aio->io_data + (dio->io_offset - 747 aio->io_offset), dio->io_size); 748 } else if (dio->io_type == ZIO_TYPE_WRITE) { 749 bcopy(dio->io_data, (char *)aio->io_data + 750 (dio->io_offset - aio->io_offset), 751 dio->io_size); 752 } 753 754 zio_add_child(dio, aio); 755 vdev_queue_io_remove(vq, dio); 756 zio_vdev_io_bypass(dio); 757 zio_execute(dio); 758 } while (dio != last); 759 760 return (aio); 761} 762 763static zio_t * 764vdev_queue_io_to_issue(vdev_queue_t *vq) 765{ 766 zio_t *zio, *aio; 767 zio_priority_t p; 768 avl_index_t idx; 769 avl_tree_t *tree; 770 zio_t search; 771 772again: 773 ASSERT(MUTEX_HELD(&vq->vq_lock)); 774 775 p = vdev_queue_class_to_issue(vq); 776 777 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { 778 /* No eligible queued i/os */ 779 return (NULL); 780 } 781 782 /* 783 * For LBA-ordered queues (async / scrub), issue the i/o which follows 784 * the most recently issued i/o in LBA (offset) order. 785 * 786 * For FIFO queues (sync), issue the i/o with the lowest timestamp. 787 */ 788 tree = vdev_queue_class_tree(vq, p); 789 search.io_timestamp = 0; 790 search.io_offset = vq->vq_last_offset + 1; 791 VERIFY3P(avl_find(tree, &search, &idx), ==, NULL); 792 zio = avl_nearest(tree, idx, AVL_AFTER); 793 if (zio == NULL) 794 zio = avl_first(tree); 795 ASSERT3U(zio->io_priority, ==, p); 796 797 aio = vdev_queue_aggregate(vq, zio); 798 if (aio != NULL) 799 zio = aio; 800 else 801 vdev_queue_io_remove(vq, zio); 802 803 /* 804 * If the I/O is or was optional and therefore has no data, we need to 805 * simply discard it. We need to drop the vdev queue's lock to avoid a 806 * deadlock that we could encounter since this I/O will complete 807 * immediately. 808 */ 809 if (zio->io_flags & ZIO_FLAG_NODATA) { 810 mutex_exit(&vq->vq_lock); 811 zio_vdev_io_bypass(zio); 812 zio_execute(zio); 813 mutex_enter(&vq->vq_lock); 814 goto again; 815 } 816 817 vdev_queue_pending_add(vq, zio); 818 vq->vq_last_offset = zio->io_offset; 819 820 return (zio); 821} 822 823zio_t * 824vdev_queue_io(zio_t *zio) 825{ 826 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 827 zio_t *nio; 828 829 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) 830 return (zio); 831 832 /* 833 * Children i/os inherent their parent's priority, which might 834 * not match the child's i/o type. Fix it up here. 835 */ 836 if (zio->io_type == ZIO_TYPE_READ) { 837 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && 838 zio->io_priority != ZIO_PRIORITY_ASYNC_READ && 839 zio->io_priority != ZIO_PRIORITY_SCRUB) 840 zio->io_priority = ZIO_PRIORITY_ASYNC_READ; 841 } else if (zio->io_type == ZIO_TYPE_WRITE) { 842 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && 843 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE) 844 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; 845 } else { 846 ASSERT(zio->io_type == ZIO_TYPE_FREE); 847 zio->io_priority = ZIO_PRIORITY_TRIM; 848 } 849 850 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; 851 852 mutex_enter(&vq->vq_lock); 853 zio->io_timestamp = gethrtime(); 854 vdev_queue_io_add(vq, zio); 855 nio = vdev_queue_io_to_issue(vq); 856 mutex_exit(&vq->vq_lock); 857 858 if (nio == NULL) 859 return (NULL); 860 861 if (nio->io_done == vdev_queue_agg_io_done) { 862 zio_nowait(nio); 863 return (NULL); 864 } 865 866 return (nio); 867} 868 869void 870vdev_queue_io_done(zio_t *zio) 871{ 872 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 873 zio_t *nio; 874 875 mutex_enter(&vq->vq_lock); 876 877 vdev_queue_pending_remove(vq, zio); 878 879 vq->vq_io_complete_ts = gethrtime(); 880 881 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { 882 mutex_exit(&vq->vq_lock); 883 if (nio->io_done == vdev_queue_agg_io_done) { 884 zio_nowait(nio); 885 } else { 886 zio_vdev_io_reissue(nio); 887 zio_execute(nio); 888 } 889 mutex_enter(&vq->vq_lock); 890 } 891 892 mutex_exit(&vq->vq_lock); 893} 894 895/* 896 * As these three methods are only used for load calculations we're not concerned 897 * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex 898 * use here, instead we prefer to keep it lock free for performance. 899 */ 900int 901vdev_queue_length(vdev_t *vd) 902{ 903 return (avl_numnodes(&vd->vdev_queue.vq_active_tree)); 904} 905 906uint64_t 907vdev_queue_lastoffset(vdev_t *vd) 908{ 909 return (vd->vdev_queue.vq_lastoffset); 910} 911 912void 913vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio) 914{ 915 vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size; 916} 917