vdev_queue.c revision 276081
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 */ 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 37/* 38 * ZFS I/O Scheduler 39 * --------------- 40 * 41 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The 42 * I/O scheduler determines when and in what order those operations are 43 * issued. The I/O scheduler divides operations into six I/O classes 44 * prioritized in the following order: sync read, sync write, async read, 45 * async write, scrub/resilver and trim. Each queue defines the minimum and 46 * maximum number of concurrent operations that may be issued to the device. 47 * In addition, the device has an aggregate maximum. Note that the sum of the 48 * per-queue minimums must not exceed the aggregate maximum, and if the 49 * aggregate maximum is equal to or greater than the sum of the per-queue 50 * maximums, the per-queue minimum has no effect. 51 * 52 * For many physical devices, throughput increases with the number of 53 * concurrent operations, but latency typically suffers. Further, physical 54 * devices typically have a limit at which more concurrent operations have no 55 * effect on throughput or can actually cause it to decrease. 56 * 57 * The scheduler selects the next operation to issue by first looking for an 58 * I/O class whose minimum has not been satisfied. Once all are satisfied and 59 * the aggregate maximum has not been hit, the scheduler looks for classes 60 * whose maximum has not been satisfied. Iteration through the I/O classes is 61 * done in the order specified above. No further operations are issued if the 62 * aggregate maximum number of concurrent operations has been hit or if there 63 * are no operations queued for an I/O class that has not hit its maximum. 64 * Every time an I/O is queued or an operation completes, the I/O scheduler 65 * looks for new operations to issue. 66 * 67 * All I/O classes have a fixed maximum number of outstanding operations 68 * except for the async write class. Asynchronous writes represent the data 69 * that is committed to stable storage during the syncing stage for 70 * transaction groups (see txg.c). Transaction groups enter the syncing state 71 * periodically so the number of queued async writes will quickly burst up and 72 * then bleed down to zero. Rather than servicing them as quickly as possible, 73 * the I/O scheduler changes the maximum number of active async write I/Os 74 * according to the amount of dirty data in the pool (see dsl_pool.c). Since 75 * both throughput and latency typically increase with the number of 76 * concurrent operations issued to physical devices, reducing the burstiness 77 * in the number of concurrent operations also stabilizes the response time of 78 * operations from other -- and in particular synchronous -- queues. In broad 79 * strokes, the I/O scheduler will issue more concurrent operations from the 80 * async write queue as there's more dirty data in the pool. 81 * 82 * Async Writes 83 * 84 * The number of concurrent operations issued for the async write I/O class 85 * follows a piece-wise linear function defined by a few adjustable points. 86 * 87 * | o---------| <-- zfs_vdev_async_write_max_active 88 * ^ | /^ | 89 * | | / | | 90 * active | / | | 91 * I/O | / | | 92 * count | / | | 93 * | / | | 94 * |------------o | | <-- zfs_vdev_async_write_min_active 95 * 0|____________^______|_________| 96 * 0% | | 100% of zfs_dirty_data_max 97 * | | 98 * | `-- zfs_vdev_async_write_active_max_dirty_percent 99 * `--------- zfs_vdev_async_write_active_min_dirty_percent 100 * 101 * Until the amount of dirty data exceeds a minimum percentage of the dirty 102 * data allowed in the pool, the I/O scheduler will limit the number of 103 * concurrent operations to the minimum. As that threshold is crossed, the 104 * number of concurrent operations issued increases linearly to the maximum at 105 * the specified maximum percentage of the dirty data allowed in the pool. 106 * 107 * Ideally, the amount of dirty data on a busy pool will stay in the sloped 108 * part of the function between zfs_vdev_async_write_active_min_dirty_percent 109 * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the 110 * maximum percentage, this indicates that the rate of incoming data is 111 * greater than the rate that the backend storage can handle. In this case, we 112 * must further throttle incoming writes (see dmu_tx_delay() for details). 113 */ 114 115/* 116 * The maximum number of I/Os active to each device. Ideally, this will be >= 117 * the sum of each queue's max_active. It must be at least the sum of each 118 * queue's min_active. 119 */ 120uint32_t zfs_vdev_max_active = 1000; 121 122/* 123 * Per-queue limits on the number of I/Os active to each device. If the 124 * sum of the queue's max_active is < zfs_vdev_max_active, then the 125 * min_active comes into play. We will send min_active from each queue, 126 * and then select from queues in the order defined by zio_priority_t. 127 * 128 * In general, smaller max_active's will lead to lower latency of synchronous 129 * operations. Larger max_active's may lead to higher overall throughput, 130 * depending on underlying storage. 131 * 132 * The ratio of the queues' max_actives determines the balance of performance 133 * between reads, writes, and scrubs. E.g., increasing 134 * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete 135 * more quickly, but reads and writes to have higher latency and lower 136 * throughput. 137 */ 138uint32_t zfs_vdev_sync_read_min_active = 10; 139uint32_t zfs_vdev_sync_read_max_active = 10; 140uint32_t zfs_vdev_sync_write_min_active = 10; 141uint32_t zfs_vdev_sync_write_max_active = 10; 142uint32_t zfs_vdev_async_read_min_active = 1; 143uint32_t zfs_vdev_async_read_max_active = 3; 144uint32_t zfs_vdev_async_write_min_active = 1; 145uint32_t zfs_vdev_async_write_max_active = 10; 146uint32_t zfs_vdev_scrub_min_active = 1; 147uint32_t zfs_vdev_scrub_max_active = 2; 148uint32_t zfs_vdev_trim_min_active = 1; 149/* 150 * TRIM max active is large in comparison to the other values due to the fact 151 * that TRIM IOs are coalesced at the device layer. This value is set such 152 * that a typical SSD can process the queued IOs in a single request. 153 */ 154uint32_t zfs_vdev_trim_max_active = 64; 155 156 157/* 158 * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent 159 * dirty data, use zfs_vdev_async_write_min_active. When it has more than 160 * zfs_vdev_async_write_active_max_dirty_percent, use 161 * zfs_vdev_async_write_max_active. The value is linearly interpolated 162 * between min and max. 163 */ 164int zfs_vdev_async_write_active_min_dirty_percent = 30; 165int zfs_vdev_async_write_active_max_dirty_percent = 60; 166 167/* 168 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O. 169 * For read I/Os, we also aggregate across small adjacency gaps; for writes 170 * we include spans of optional I/Os to aid aggregation at the disk even when 171 * they aren't able to help us aggregate at this level. 172 */ 173int zfs_vdev_aggregation_limit = SPA_OLD_MAXBLOCKSIZE; 174int zfs_vdev_read_gap_limit = 32 << 10; 175int zfs_vdev_write_gap_limit = 4 << 10; 176 177#ifdef __FreeBSD__ 178SYSCTL_DECL(_vfs_zfs_vdev); 179 180TUNABLE_INT("vfs.zfs.vdev.async_write_active_min_dirty_percent", 181 &zfs_vdev_async_write_active_min_dirty_percent); 182static int sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS); 183SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_min_dirty_percent, 184 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int), 185 sysctl_zfs_async_write_active_min_dirty_percent, "I", 186 "Percentage of async write dirty data below which " 187 "async_write_min_active is used."); 188 189TUNABLE_INT("vfs.zfs.vdev.async_write_active_max_dirty_percent", 190 &zfs_vdev_async_write_active_max_dirty_percent); 191static int sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS); 192SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_max_dirty_percent, 193 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int), 194 sysctl_zfs_async_write_active_max_dirty_percent, "I", 195 "Percentage of async write dirty data above which " 196 "async_write_max_active is used."); 197 198TUNABLE_INT("vfs.zfs.vdev.max_active", &zfs_vdev_max_active); 199SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RWTUN, 200 &zfs_vdev_max_active, 0, 201 "The maximum number of I/Os of all types active for each device."); 202 203#define ZFS_VDEV_QUEUE_KNOB_MIN(name) \ 204TUNABLE_INT("vfs.zfs.vdev." #name "_min_active", \ 205 &zfs_vdev_ ## name ## _min_active); \ 206SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, \ 207 CTLFLAG_RWTUN, &zfs_vdev_ ## name ## _min_active, 0, \ 208 "Initial number of I/O requests of type " #name \ 209 " active for each device"); 210 211#define ZFS_VDEV_QUEUE_KNOB_MAX(name) \ 212TUNABLE_INT("vfs.zfs.vdev." #name "_max_active", \ 213 &zfs_vdev_ ## name ## _max_active); \ 214SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, \ 215 CTLFLAG_RWTUN, &zfs_vdev_ ## name ## _max_active, 0, \ 216 "Maximum number of I/O requests of type " #name \ 217 " active for each device"); 218 219ZFS_VDEV_QUEUE_KNOB_MIN(sync_read); 220ZFS_VDEV_QUEUE_KNOB_MAX(sync_read); 221ZFS_VDEV_QUEUE_KNOB_MIN(sync_write); 222ZFS_VDEV_QUEUE_KNOB_MAX(sync_write); 223ZFS_VDEV_QUEUE_KNOB_MIN(async_read); 224ZFS_VDEV_QUEUE_KNOB_MAX(async_read); 225ZFS_VDEV_QUEUE_KNOB_MIN(async_write); 226ZFS_VDEV_QUEUE_KNOB_MAX(async_write); 227ZFS_VDEV_QUEUE_KNOB_MIN(scrub); 228ZFS_VDEV_QUEUE_KNOB_MAX(scrub); 229ZFS_VDEV_QUEUE_KNOB_MIN(trim); 230ZFS_VDEV_QUEUE_KNOB_MAX(trim); 231 232#undef ZFS_VDEV_QUEUE_KNOB 233 234TUNABLE_INT("vfs.zfs.vdev.aggregation_limit", &zfs_vdev_aggregation_limit); 235SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RWTUN, 236 &zfs_vdev_aggregation_limit, 0, 237 "I/O requests are aggregated up to this size"); 238TUNABLE_INT("vfs.zfs.vdev.read_gap_limit", &zfs_vdev_read_gap_limit); 239SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RWTUN, 240 &zfs_vdev_read_gap_limit, 0, 241 "Acceptable gap between two reads being aggregated"); 242TUNABLE_INT("vfs.zfs.vdev.write_gap_limit", &zfs_vdev_write_gap_limit); 243SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RWTUN, 244 &zfs_vdev_write_gap_limit, 0, 245 "Acceptable gap between two writes being aggregated"); 246 247static int 248sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS) 249{ 250 int val, err; 251 252 val = zfs_vdev_async_write_active_min_dirty_percent; 253 err = sysctl_handle_int(oidp, &val, 0, req); 254 if (err != 0 || req->newptr == NULL) 255 return (err); 256 257 if (val < 0 || val > 100 || 258 val >= zfs_vdev_async_write_active_max_dirty_percent) 259 return (EINVAL); 260 261 zfs_vdev_async_write_active_min_dirty_percent = val; 262 263 return (0); 264} 265 266static int 267sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS) 268{ 269 int val, err; 270 271 val = zfs_vdev_async_write_active_max_dirty_percent; 272 err = sysctl_handle_int(oidp, &val, 0, req); 273 if (err != 0 || req->newptr == NULL) 274 return (err); 275 276 if (val < 0 || val > 100 || 277 val <= zfs_vdev_async_write_active_min_dirty_percent) 278 return (EINVAL); 279 280 zfs_vdev_async_write_active_max_dirty_percent = val; 281 282 return (0); 283} 284#endif 285 286int 287vdev_queue_offset_compare(const void *x1, const void *x2) 288{ 289 const zio_t *z1 = x1; 290 const zio_t *z2 = x2; 291 292 if (z1->io_offset < z2->io_offset) 293 return (-1); 294 if (z1->io_offset > z2->io_offset) 295 return (1); 296 297 if (z1 < z2) 298 return (-1); 299 if (z1 > z2) 300 return (1); 301 302 return (0); 303} 304 305int 306vdev_queue_timestamp_compare(const void *x1, const void *x2) 307{ 308 const zio_t *z1 = x1; 309 const zio_t *z2 = x2; 310 311 if (z1->io_timestamp < z2->io_timestamp) 312 return (-1); 313 if (z1->io_timestamp > z2->io_timestamp) 314 return (1); 315 316 if (z1 < z2) 317 return (-1); 318 if (z1 > z2) 319 return (1); 320 321 return (0); 322} 323 324void 325vdev_queue_init(vdev_t *vd) 326{ 327 vdev_queue_t *vq = &vd->vdev_queue; 328 329 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); 330 vq->vq_vdev = vd; 331 332 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare, 333 sizeof (zio_t), offsetof(struct zio, io_queue_node)); 334 335 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 336 /* 337 * The synchronous i/o queues are FIFO rather than LBA ordered. 338 * This provides more consistent latency for these i/os, and 339 * they tend to not be tightly clustered anyway so there is 340 * little to no throughput loss. 341 */ 342 boolean_t fifo = (p == ZIO_PRIORITY_SYNC_READ || 343 p == ZIO_PRIORITY_SYNC_WRITE); 344 avl_create(&vq->vq_class[p].vqc_queued_tree, 345 fifo ? vdev_queue_timestamp_compare : 346 vdev_queue_offset_compare, 347 sizeof (zio_t), offsetof(struct zio, io_queue_node)); 348 } 349 350 vq->vq_lastoffset = 0; 351} 352 353void 354vdev_queue_fini(vdev_t *vd) 355{ 356 vdev_queue_t *vq = &vd->vdev_queue; 357 358 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) 359 avl_destroy(&vq->vq_class[p].vqc_queued_tree); 360 avl_destroy(&vq->vq_active_tree); 361 362 mutex_destroy(&vq->vq_lock); 363} 364 365static void 366vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) 367{ 368 spa_t *spa = zio->io_spa; 369 ASSERT(MUTEX_HELD(&vq->vq_lock)); 370 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 371 avl_add(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio); 372 373#ifdef illumos 374 mutex_enter(&spa->spa_iokstat_lock); 375 spa->spa_queue_stats[zio->io_priority].spa_queued++; 376 if (spa->spa_iokstat != NULL) 377 kstat_waitq_enter(spa->spa_iokstat->ks_data); 378 mutex_exit(&spa->spa_iokstat_lock); 379#endif 380} 381 382static void 383vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) 384{ 385 spa_t *spa = zio->io_spa; 386 ASSERT(MUTEX_HELD(&vq->vq_lock)); 387 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 388 avl_remove(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio); 389 390#ifdef illumos 391 mutex_enter(&spa->spa_iokstat_lock); 392 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0); 393 spa->spa_queue_stats[zio->io_priority].spa_queued--; 394 if (spa->spa_iokstat != NULL) 395 kstat_waitq_exit(spa->spa_iokstat->ks_data); 396 mutex_exit(&spa->spa_iokstat_lock); 397#endif 398} 399 400static void 401vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) 402{ 403 spa_t *spa = zio->io_spa; 404 ASSERT(MUTEX_HELD(&vq->vq_lock)); 405 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 406 vq->vq_class[zio->io_priority].vqc_active++; 407 avl_add(&vq->vq_active_tree, zio); 408 409#ifdef illumos 410 mutex_enter(&spa->spa_iokstat_lock); 411 spa->spa_queue_stats[zio->io_priority].spa_active++; 412 if (spa->spa_iokstat != NULL) 413 kstat_runq_enter(spa->spa_iokstat->ks_data); 414 mutex_exit(&spa->spa_iokstat_lock); 415#endif 416} 417 418static void 419vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) 420{ 421 spa_t *spa = zio->io_spa; 422 ASSERT(MUTEX_HELD(&vq->vq_lock)); 423 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 424 vq->vq_class[zio->io_priority].vqc_active--; 425 avl_remove(&vq->vq_active_tree, zio); 426 427#ifdef illumos 428 mutex_enter(&spa->spa_iokstat_lock); 429 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0); 430 spa->spa_queue_stats[zio->io_priority].spa_active--; 431 if (spa->spa_iokstat != NULL) { 432 kstat_io_t *ksio = spa->spa_iokstat->ks_data; 433 434 kstat_runq_exit(spa->spa_iokstat->ks_data); 435 if (zio->io_type == ZIO_TYPE_READ) { 436 ksio->reads++; 437 ksio->nread += zio->io_size; 438 } else if (zio->io_type == ZIO_TYPE_WRITE) { 439 ksio->writes++; 440 ksio->nwritten += zio->io_size; 441 } 442 } 443 mutex_exit(&spa->spa_iokstat_lock); 444#endif 445} 446 447static void 448vdev_queue_agg_io_done(zio_t *aio) 449{ 450 if (aio->io_type == ZIO_TYPE_READ) { 451 zio_t *pio; 452 while ((pio = zio_walk_parents(aio)) != NULL) { 453 bcopy((char *)aio->io_data + (pio->io_offset - 454 aio->io_offset), pio->io_data, pio->io_size); 455 } 456 } 457 458 zio_buf_free(aio->io_data, aio->io_size); 459} 460 461static int 462vdev_queue_class_min_active(zio_priority_t p) 463{ 464 switch (p) { 465 case ZIO_PRIORITY_SYNC_READ: 466 return (zfs_vdev_sync_read_min_active); 467 case ZIO_PRIORITY_SYNC_WRITE: 468 return (zfs_vdev_sync_write_min_active); 469 case ZIO_PRIORITY_ASYNC_READ: 470 return (zfs_vdev_async_read_min_active); 471 case ZIO_PRIORITY_ASYNC_WRITE: 472 return (zfs_vdev_async_write_min_active); 473 case ZIO_PRIORITY_SCRUB: 474 return (zfs_vdev_scrub_min_active); 475 case ZIO_PRIORITY_TRIM: 476 return (zfs_vdev_trim_min_active); 477 default: 478 panic("invalid priority %u", p); 479 return (0); 480 } 481} 482 483static int 484vdev_queue_max_async_writes(spa_t *spa) 485{ 486 int writes; 487 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total; 488 uint64_t min_bytes = zfs_dirty_data_max * 489 zfs_vdev_async_write_active_min_dirty_percent / 100; 490 uint64_t max_bytes = zfs_dirty_data_max * 491 zfs_vdev_async_write_active_max_dirty_percent / 100; 492 493 /* 494 * Sync tasks correspond to interactive user actions. To reduce the 495 * execution time of those actions we push data out as fast as possible. 496 */ 497 if (spa_has_pending_synctask(spa)) { 498 return (zfs_vdev_async_write_max_active); 499 } 500 501 if (dirty < min_bytes) 502 return (zfs_vdev_async_write_min_active); 503 if (dirty > max_bytes) 504 return (zfs_vdev_async_write_max_active); 505 506 /* 507 * linear interpolation: 508 * slope = (max_writes - min_writes) / (max_bytes - min_bytes) 509 * move right by min_bytes 510 * move up by min_writes 511 */ 512 writes = (dirty - min_bytes) * 513 (zfs_vdev_async_write_max_active - 514 zfs_vdev_async_write_min_active) / 515 (max_bytes - min_bytes) + 516 zfs_vdev_async_write_min_active; 517 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); 518 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); 519 return (writes); 520} 521 522static int 523vdev_queue_class_max_active(spa_t *spa, zio_priority_t p) 524{ 525 switch (p) { 526 case ZIO_PRIORITY_SYNC_READ: 527 return (zfs_vdev_sync_read_max_active); 528 case ZIO_PRIORITY_SYNC_WRITE: 529 return (zfs_vdev_sync_write_max_active); 530 case ZIO_PRIORITY_ASYNC_READ: 531 return (zfs_vdev_async_read_max_active); 532 case ZIO_PRIORITY_ASYNC_WRITE: 533 return (vdev_queue_max_async_writes(spa)); 534 case ZIO_PRIORITY_SCRUB: 535 return (zfs_vdev_scrub_max_active); 536 case ZIO_PRIORITY_TRIM: 537 return (zfs_vdev_trim_max_active); 538 default: 539 panic("invalid priority %u", p); 540 return (0); 541 } 542} 543 544/* 545 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if 546 * there is no eligible class. 547 */ 548static zio_priority_t 549vdev_queue_class_to_issue(vdev_queue_t *vq) 550{ 551 spa_t *spa = vq->vq_vdev->vdev_spa; 552 zio_priority_t p; 553 554 ASSERT(MUTEX_HELD(&vq->vq_lock)); 555 556 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active) 557 return (ZIO_PRIORITY_NUM_QUEUEABLE); 558 559 /* find a queue that has not reached its minimum # outstanding i/os */ 560 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 561 if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 && 562 vq->vq_class[p].vqc_active < 563 vdev_queue_class_min_active(p)) 564 return (p); 565 } 566 567 /* 568 * If we haven't found a queue, look for one that hasn't reached its 569 * maximum # outstanding i/os. 570 */ 571 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 572 if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 && 573 vq->vq_class[p].vqc_active < 574 vdev_queue_class_max_active(spa, p)) 575 return (p); 576 } 577 578 /* No eligible queued i/os */ 579 return (ZIO_PRIORITY_NUM_QUEUEABLE); 580} 581 582/* 583 * Compute the range spanned by two i/os, which is the endpoint of the last 584 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). 585 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); 586 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. 587 */ 588#define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) 589#define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) 590 591static zio_t * 592vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) 593{ 594 zio_t *first, *last, *aio, *dio, *mandatory, *nio; 595 uint64_t maxgap = 0; 596 uint64_t size; 597 boolean_t stretch; 598 avl_tree_t *t; 599 enum zio_flag flags; 600 601 ASSERT(MUTEX_HELD(&vq->vq_lock)); 602 603 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE) 604 return (NULL); 605 606 /* 607 * The synchronous i/o queues are not sorted by LBA, so we can't 608 * find adjacent i/os. These i/os tend to not be tightly clustered, 609 * or too large to aggregate, so this has little impact on performance. 610 */ 611 if (zio->io_priority == ZIO_PRIORITY_SYNC_READ || 612 zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) 613 return (NULL); 614 615 first = last = zio; 616 617 if (zio->io_type == ZIO_TYPE_READ) 618 maxgap = zfs_vdev_read_gap_limit; 619 620 /* 621 * We can aggregate I/Os that are sufficiently adjacent and of 622 * the same flavor, as expressed by the AGG_INHERIT flags. 623 * The latter requirement is necessary so that certain 624 * attributes of the I/O, such as whether it's a normal I/O 625 * or a scrub/resilver, can be preserved in the aggregate. 626 * We can include optional I/Os, but don't allow them 627 * to begin a range as they add no benefit in that situation. 628 */ 629 630 /* 631 * We keep track of the last non-optional I/O. 632 */ 633 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; 634 635 /* 636 * Walk backwards through sufficiently contiguous I/Os 637 * recording the last non-option I/O. 638 */ 639 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; 640 t = &vq->vq_class[zio->io_priority].vqc_queued_tree; 641 while ((dio = AVL_PREV(t, first)) != NULL && 642 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 643 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit && 644 IO_GAP(dio, first) <= maxgap) { 645 first = dio; 646 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) 647 mandatory = first; 648 } 649 650 /* 651 * Skip any initial optional I/Os. 652 */ 653 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { 654 first = AVL_NEXT(t, first); 655 ASSERT(first != NULL); 656 } 657 658 /* 659 * Walk forward through sufficiently contiguous I/Os. 660 */ 661 while ((dio = AVL_NEXT(t, last)) != NULL && 662 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 663 IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit && 664 IO_GAP(last, dio) <= maxgap) { 665 last = dio; 666 if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) 667 mandatory = last; 668 } 669 670 /* 671 * Now that we've established the range of the I/O aggregation 672 * we must decide what to do with trailing optional I/Os. 673 * For reads, there's nothing to do. While we are unable to 674 * aggregate further, it's possible that a trailing optional 675 * I/O would allow the underlying device to aggregate with 676 * subsequent I/Os. We must therefore determine if the next 677 * non-optional I/O is close enough to make aggregation 678 * worthwhile. 679 */ 680 stretch = B_FALSE; 681 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { 682 zio_t *nio = last; 683 while ((dio = AVL_NEXT(t, nio)) != NULL && 684 IO_GAP(nio, dio) == 0 && 685 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { 686 nio = dio; 687 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { 688 stretch = B_TRUE; 689 break; 690 } 691 } 692 } 693 694 if (stretch) { 695 /* This may be a no-op. */ 696 dio = AVL_NEXT(t, last); 697 dio->io_flags &= ~ZIO_FLAG_OPTIONAL; 698 } else { 699 while (last != mandatory && last != first) { 700 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); 701 last = AVL_PREV(t, last); 702 ASSERT(last != NULL); 703 } 704 } 705 706 if (first == last) 707 return (NULL); 708 709 size = IO_SPAN(first, last); 710 ASSERT3U(size, <=, zfs_vdev_aggregation_limit); 711 712 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, 713 zio_buf_alloc(size), size, first->io_type, zio->io_priority, 714 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE, 715 vdev_queue_agg_io_done, NULL); 716 aio->io_timestamp = first->io_timestamp; 717 718 nio = first; 719 do { 720 dio = nio; 721 nio = AVL_NEXT(t, dio); 722 ASSERT3U(dio->io_type, ==, aio->io_type); 723 724 if (dio->io_flags & ZIO_FLAG_NODATA) { 725 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); 726 bzero((char *)aio->io_data + (dio->io_offset - 727 aio->io_offset), dio->io_size); 728 } else if (dio->io_type == ZIO_TYPE_WRITE) { 729 bcopy(dio->io_data, (char *)aio->io_data + 730 (dio->io_offset - aio->io_offset), 731 dio->io_size); 732 } 733 734 zio_add_child(dio, aio); 735 vdev_queue_io_remove(vq, dio); 736 zio_vdev_io_bypass(dio); 737 zio_execute(dio); 738 } while (dio != last); 739 740 return (aio); 741} 742 743static zio_t * 744vdev_queue_io_to_issue(vdev_queue_t *vq) 745{ 746 zio_t *zio, *aio; 747 zio_priority_t p; 748 avl_index_t idx; 749 vdev_queue_class_t *vqc; 750 zio_t search; 751 752again: 753 ASSERT(MUTEX_HELD(&vq->vq_lock)); 754 755 p = vdev_queue_class_to_issue(vq); 756 757 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { 758 /* No eligible queued i/os */ 759 return (NULL); 760 } 761 762 /* 763 * For LBA-ordered queues (async / scrub), issue the i/o which follows 764 * the most recently issued i/o in LBA (offset) order. 765 * 766 * For FIFO queues (sync), issue the i/o with the lowest timestamp. 767 */ 768 vqc = &vq->vq_class[p]; 769 search.io_timestamp = 0; 770 search.io_offset = vq->vq_last_offset + 1; 771 VERIFY3P(avl_find(&vqc->vqc_queued_tree, &search, &idx), ==, NULL); 772 zio = avl_nearest(&vqc->vqc_queued_tree, idx, AVL_AFTER); 773 if (zio == NULL) 774 zio = avl_first(&vqc->vqc_queued_tree); 775 ASSERT3U(zio->io_priority, ==, p); 776 777 aio = vdev_queue_aggregate(vq, zio); 778 if (aio != NULL) 779 zio = aio; 780 else 781 vdev_queue_io_remove(vq, zio); 782 783 /* 784 * If the I/O is or was optional and therefore has no data, we need to 785 * simply discard it. We need to drop the vdev queue's lock to avoid a 786 * deadlock that we could encounter since this I/O will complete 787 * immediately. 788 */ 789 if (zio->io_flags & ZIO_FLAG_NODATA) { 790 mutex_exit(&vq->vq_lock); 791 zio_vdev_io_bypass(zio); 792 zio_execute(zio); 793 mutex_enter(&vq->vq_lock); 794 goto again; 795 } 796 797 vdev_queue_pending_add(vq, zio); 798 vq->vq_last_offset = zio->io_offset; 799 800 return (zio); 801} 802 803zio_t * 804vdev_queue_io(zio_t *zio) 805{ 806 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 807 zio_t *nio; 808 809 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) 810 return (zio); 811 812 /* 813 * Children i/os inherent their parent's priority, which might 814 * not match the child's i/o type. Fix it up here. 815 */ 816 if (zio->io_type == ZIO_TYPE_READ) { 817 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && 818 zio->io_priority != ZIO_PRIORITY_ASYNC_READ && 819 zio->io_priority != ZIO_PRIORITY_SCRUB) 820 zio->io_priority = ZIO_PRIORITY_ASYNC_READ; 821 } else if (zio->io_type == ZIO_TYPE_WRITE) { 822 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && 823 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE) 824 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; 825 } else { 826 ASSERT(zio->io_type == ZIO_TYPE_FREE); 827 zio->io_priority = ZIO_PRIORITY_TRIM; 828 } 829 830 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; 831 832 mutex_enter(&vq->vq_lock); 833 zio->io_timestamp = gethrtime(); 834 vdev_queue_io_add(vq, zio); 835 nio = vdev_queue_io_to_issue(vq); 836 mutex_exit(&vq->vq_lock); 837 838 if (nio == NULL) 839 return (NULL); 840 841 if (nio->io_done == vdev_queue_agg_io_done) { 842 zio_nowait(nio); 843 return (NULL); 844 } 845 846 return (nio); 847} 848 849void 850vdev_queue_io_done(zio_t *zio) 851{ 852 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 853 zio_t *nio; 854 855 if (zio_injection_enabled) 856 delay(SEC_TO_TICK(zio_handle_io_delay(zio))); 857 858 mutex_enter(&vq->vq_lock); 859 860 vdev_queue_pending_remove(vq, zio); 861 862 vq->vq_io_complete_ts = gethrtime(); 863 864 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { 865 mutex_exit(&vq->vq_lock); 866 if (nio->io_done == vdev_queue_agg_io_done) { 867 zio_nowait(nio); 868 } else { 869 zio_vdev_io_reissue(nio); 870 zio_execute(nio); 871 } 872 mutex_enter(&vq->vq_lock); 873 } 874 875 mutex_exit(&vq->vq_lock); 876} 877 878/* 879 * As these three methods are only used for load calculations we're not concerned 880 * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex 881 * use here, instead we prefer to keep it lock free for performance. 882 */ 883int 884vdev_queue_length(vdev_t *vd) 885{ 886 return (avl_numnodes(&vd->vdev_queue.vq_active_tree)); 887} 888 889uint64_t 890vdev_queue_lastoffset(vdev_t *vd) 891{ 892 return (vd->vdev_queue.vq_lastoffset); 893} 894 895void 896vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio) 897{ 898 vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size; 899} 900