vdev_queue.c revision 271238
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_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); 179TUNABLE_INT("vfs.zfs.vdev.max_active", &zfs_vdev_max_active); 180SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RW, 181 &zfs_vdev_max_active, 0, 182 "The maximum number of I/Os of all types active for each device."); 183 184#define ZFS_VDEV_QUEUE_KNOB_MIN(name) \ 185TUNABLE_INT("vfs.zfs.vdev." #name "_min_active", \ 186 &zfs_vdev_ ## name ## _min_active); \ 187SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, CTLFLAG_RW, \ 188 &zfs_vdev_ ## name ## _min_active, 0, \ 189 "Initial number of I/O requests of type " #name \ 190 " active for each device"); 191 192#define ZFS_VDEV_QUEUE_KNOB_MAX(name) \ 193TUNABLE_INT("vfs.zfs.vdev." #name "_max_active", \ 194 &zfs_vdev_ ## name ## _max_active); \ 195SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, CTLFLAG_RW, \ 196 &zfs_vdev_ ## name ## _max_active, 0, \ 197 "Maximum number of I/O requests of type " #name \ 198 " active for each device"); 199 200ZFS_VDEV_QUEUE_KNOB_MIN(sync_read); 201ZFS_VDEV_QUEUE_KNOB_MAX(sync_read); 202ZFS_VDEV_QUEUE_KNOB_MIN(sync_write); 203ZFS_VDEV_QUEUE_KNOB_MAX(sync_write); 204ZFS_VDEV_QUEUE_KNOB_MIN(async_read); 205ZFS_VDEV_QUEUE_KNOB_MAX(async_read); 206ZFS_VDEV_QUEUE_KNOB_MIN(async_write); 207ZFS_VDEV_QUEUE_KNOB_MAX(async_write); 208ZFS_VDEV_QUEUE_KNOB_MIN(scrub); 209ZFS_VDEV_QUEUE_KNOB_MAX(scrub); 210ZFS_VDEV_QUEUE_KNOB_MIN(trim); 211ZFS_VDEV_QUEUE_KNOB_MAX(trim); 212 213#undef ZFS_VDEV_QUEUE_KNOB 214 215TUNABLE_INT("vfs.zfs.vdev.aggregation_limit", &zfs_vdev_aggregation_limit); 216SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RW, 217 &zfs_vdev_aggregation_limit, 0, 218 "I/O requests are aggregated up to this size"); 219TUNABLE_INT("vfs.zfs.vdev.read_gap_limit", &zfs_vdev_read_gap_limit); 220SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RW, 221 &zfs_vdev_read_gap_limit, 0, 222 "Acceptable gap between two reads being aggregated"); 223TUNABLE_INT("vfs.zfs.vdev.write_gap_limit", &zfs_vdev_write_gap_limit); 224SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RW, 225 &zfs_vdev_write_gap_limit, 0, 226 "Acceptable gap between two writes being aggregated"); 227#endif 228 229int 230vdev_queue_offset_compare(const void *x1, const void *x2) 231{ 232 const zio_t *z1 = x1; 233 const zio_t *z2 = x2; 234 235 if (z1->io_offset < z2->io_offset) 236 return (-1); 237 if (z1->io_offset > z2->io_offset) 238 return (1); 239 240 if (z1 < z2) 241 return (-1); 242 if (z1 > z2) 243 return (1); 244 245 return (0); 246} 247 248int 249vdev_queue_timestamp_compare(const void *x1, const void *x2) 250{ 251 const zio_t *z1 = x1; 252 const zio_t *z2 = x2; 253 254 if (z1->io_timestamp < z2->io_timestamp) 255 return (-1); 256 if (z1->io_timestamp > z2->io_timestamp) 257 return (1); 258 259 if (z1 < z2) 260 return (-1); 261 if (z1 > z2) 262 return (1); 263 264 return (0); 265} 266 267void 268vdev_queue_init(vdev_t *vd) 269{ 270 vdev_queue_t *vq = &vd->vdev_queue; 271 272 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); 273 vq->vq_vdev = vd; 274 275 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare, 276 sizeof (zio_t), offsetof(struct zio, io_queue_node)); 277 278 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 279 /* 280 * The synchronous i/o queues are FIFO rather than LBA ordered. 281 * This provides more consistent latency for these i/os, and 282 * they tend to not be tightly clustered anyway so there is 283 * little to no throughput loss. 284 */ 285 boolean_t fifo = (p == ZIO_PRIORITY_SYNC_READ || 286 p == ZIO_PRIORITY_SYNC_WRITE); 287 avl_create(&vq->vq_class[p].vqc_queued_tree, 288 fifo ? vdev_queue_timestamp_compare : 289 vdev_queue_offset_compare, 290 sizeof (zio_t), offsetof(struct zio, io_queue_node)); 291 } 292 293 vq->vq_lastoffset = 0; 294} 295 296void 297vdev_queue_fini(vdev_t *vd) 298{ 299 vdev_queue_t *vq = &vd->vdev_queue; 300 301 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) 302 avl_destroy(&vq->vq_class[p].vqc_queued_tree); 303 avl_destroy(&vq->vq_active_tree); 304 305 mutex_destroy(&vq->vq_lock); 306} 307 308static void 309vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) 310{ 311 spa_t *spa = zio->io_spa; 312 ASSERT(MUTEX_HELD(&vq->vq_lock)); 313 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 314 avl_add(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio); 315 316#ifdef illumos 317 mutex_enter(&spa->spa_iokstat_lock); 318 spa->spa_queue_stats[zio->io_priority].spa_queued++; 319 if (spa->spa_iokstat != NULL) 320 kstat_waitq_enter(spa->spa_iokstat->ks_data); 321 mutex_exit(&spa->spa_iokstat_lock); 322#endif 323} 324 325static void 326vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) 327{ 328 spa_t *spa = zio->io_spa; 329 ASSERT(MUTEX_HELD(&vq->vq_lock)); 330 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 331 avl_remove(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio); 332 333#ifdef illumos 334 mutex_enter(&spa->spa_iokstat_lock); 335 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0); 336 spa->spa_queue_stats[zio->io_priority].spa_queued--; 337 if (spa->spa_iokstat != NULL) 338 kstat_waitq_exit(spa->spa_iokstat->ks_data); 339 mutex_exit(&spa->spa_iokstat_lock); 340#endif 341} 342 343static void 344vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) 345{ 346 spa_t *spa = zio->io_spa; 347 ASSERT(MUTEX_HELD(&vq->vq_lock)); 348 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 349 vq->vq_class[zio->io_priority].vqc_active++; 350 avl_add(&vq->vq_active_tree, zio); 351 352#ifdef illumos 353 mutex_enter(&spa->spa_iokstat_lock); 354 spa->spa_queue_stats[zio->io_priority].spa_active++; 355 if (spa->spa_iokstat != NULL) 356 kstat_runq_enter(spa->spa_iokstat->ks_data); 357 mutex_exit(&spa->spa_iokstat_lock); 358#endif 359} 360 361static void 362vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) 363{ 364 spa_t *spa = zio->io_spa; 365 ASSERT(MUTEX_HELD(&vq->vq_lock)); 366 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 367 vq->vq_class[zio->io_priority].vqc_active--; 368 avl_remove(&vq->vq_active_tree, zio); 369 370#ifdef illumos 371 mutex_enter(&spa->spa_iokstat_lock); 372 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0); 373 spa->spa_queue_stats[zio->io_priority].spa_active--; 374 if (spa->spa_iokstat != NULL) { 375 kstat_io_t *ksio = spa->spa_iokstat->ks_data; 376 377 kstat_runq_exit(spa->spa_iokstat->ks_data); 378 if (zio->io_type == ZIO_TYPE_READ) { 379 ksio->reads++; 380 ksio->nread += zio->io_size; 381 } else if (zio->io_type == ZIO_TYPE_WRITE) { 382 ksio->writes++; 383 ksio->nwritten += zio->io_size; 384 } 385 } 386 mutex_exit(&spa->spa_iokstat_lock); 387#endif 388} 389 390static void 391vdev_queue_agg_io_done(zio_t *aio) 392{ 393 if (aio->io_type == ZIO_TYPE_READ) { 394 zio_t *pio; 395 while ((pio = zio_walk_parents(aio)) != NULL) { 396 bcopy((char *)aio->io_data + (pio->io_offset - 397 aio->io_offset), pio->io_data, pio->io_size); 398 } 399 } 400 401 zio_buf_free(aio->io_data, aio->io_size); 402} 403 404static int 405vdev_queue_class_min_active(zio_priority_t p) 406{ 407 switch (p) { 408 case ZIO_PRIORITY_SYNC_READ: 409 return (zfs_vdev_sync_read_min_active); 410 case ZIO_PRIORITY_SYNC_WRITE: 411 return (zfs_vdev_sync_write_min_active); 412 case ZIO_PRIORITY_ASYNC_READ: 413 return (zfs_vdev_async_read_min_active); 414 case ZIO_PRIORITY_ASYNC_WRITE: 415 return (zfs_vdev_async_write_min_active); 416 case ZIO_PRIORITY_SCRUB: 417 return (zfs_vdev_scrub_min_active); 418 case ZIO_PRIORITY_TRIM: 419 return (zfs_vdev_trim_min_active); 420 default: 421 panic("invalid priority %u", p); 422 return (0); 423 } 424} 425 426static int 427vdev_queue_max_async_writes(spa_t *spa) 428{ 429 int writes; 430 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total; 431 uint64_t min_bytes = zfs_dirty_data_max * 432 zfs_vdev_async_write_active_min_dirty_percent / 100; 433 uint64_t max_bytes = zfs_dirty_data_max * 434 zfs_vdev_async_write_active_max_dirty_percent / 100; 435 436 /* 437 * Sync tasks correspond to interactive user actions. To reduce the 438 * execution time of those actions we push data out as fast as possible. 439 */ 440 if (spa_has_pending_synctask(spa)) { 441 return (zfs_vdev_async_write_max_active); 442 } 443 444 if (dirty < min_bytes) 445 return (zfs_vdev_async_write_min_active); 446 if (dirty > max_bytes) 447 return (zfs_vdev_async_write_max_active); 448 449 /* 450 * linear interpolation: 451 * slope = (max_writes - min_writes) / (max_bytes - min_bytes) 452 * move right by min_bytes 453 * move up by min_writes 454 */ 455 writes = (dirty - min_bytes) * 456 (zfs_vdev_async_write_max_active - 457 zfs_vdev_async_write_min_active) / 458 (max_bytes - min_bytes) + 459 zfs_vdev_async_write_min_active; 460 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); 461 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); 462 return (writes); 463} 464 465static int 466vdev_queue_class_max_active(spa_t *spa, zio_priority_t p) 467{ 468 switch (p) { 469 case ZIO_PRIORITY_SYNC_READ: 470 return (zfs_vdev_sync_read_max_active); 471 case ZIO_PRIORITY_SYNC_WRITE: 472 return (zfs_vdev_sync_write_max_active); 473 case ZIO_PRIORITY_ASYNC_READ: 474 return (zfs_vdev_async_read_max_active); 475 case ZIO_PRIORITY_ASYNC_WRITE: 476 return (vdev_queue_max_async_writes(spa)); 477 case ZIO_PRIORITY_SCRUB: 478 return (zfs_vdev_scrub_max_active); 479 case ZIO_PRIORITY_TRIM: 480 return (zfs_vdev_trim_max_active); 481 default: 482 panic("invalid priority %u", p); 483 return (0); 484 } 485} 486 487/* 488 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if 489 * there is no eligible class. 490 */ 491static zio_priority_t 492vdev_queue_class_to_issue(vdev_queue_t *vq) 493{ 494 spa_t *spa = vq->vq_vdev->vdev_spa; 495 zio_priority_t p; 496 497 ASSERT(MUTEX_HELD(&vq->vq_lock)); 498 499 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active) 500 return (ZIO_PRIORITY_NUM_QUEUEABLE); 501 502 /* find a queue that has not reached its minimum # outstanding i/os */ 503 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 504 if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 && 505 vq->vq_class[p].vqc_active < 506 vdev_queue_class_min_active(p)) 507 return (p); 508 } 509 510 /* 511 * If we haven't found a queue, look for one that hasn't reached its 512 * maximum # outstanding i/os. 513 */ 514 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 515 if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 && 516 vq->vq_class[p].vqc_active < 517 vdev_queue_class_max_active(spa, p)) 518 return (p); 519 } 520 521 /* No eligible queued i/os */ 522 return (ZIO_PRIORITY_NUM_QUEUEABLE); 523} 524 525/* 526 * Compute the range spanned by two i/os, which is the endpoint of the last 527 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). 528 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); 529 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. 530 */ 531#define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) 532#define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) 533 534static zio_t * 535vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) 536{ 537 zio_t *first, *last, *aio, *dio, *mandatory, *nio; 538 uint64_t maxgap = 0; 539 uint64_t size; 540 boolean_t stretch; 541 avl_tree_t *t; 542 enum zio_flag flags; 543 544 ASSERT(MUTEX_HELD(&vq->vq_lock)); 545 546 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE) 547 return (NULL); 548 549 /* 550 * The synchronous i/o queues are not sorted by LBA, so we can't 551 * find adjacent i/os. These i/os tend to not be tightly clustered, 552 * or too large to aggregate, so this has little impact on performance. 553 */ 554 if (zio->io_priority == ZIO_PRIORITY_SYNC_READ || 555 zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) 556 return (NULL); 557 558 first = last = zio; 559 560 if (zio->io_type == ZIO_TYPE_READ) 561 maxgap = zfs_vdev_read_gap_limit; 562 563 /* 564 * We can aggregate I/Os that are sufficiently adjacent and of 565 * the same flavor, as expressed by the AGG_INHERIT flags. 566 * The latter requirement is necessary so that certain 567 * attributes of the I/O, such as whether it's a normal I/O 568 * or a scrub/resilver, can be preserved in the aggregate. 569 * We can include optional I/Os, but don't allow them 570 * to begin a range as they add no benefit in that situation. 571 */ 572 573 /* 574 * We keep track of the last non-optional I/O. 575 */ 576 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; 577 578 /* 579 * Walk backwards through sufficiently contiguous I/Os 580 * recording the last non-option I/O. 581 */ 582 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; 583 t = &vq->vq_class[zio->io_priority].vqc_queued_tree; 584 while ((dio = AVL_PREV(t, first)) != NULL && 585 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 586 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit && 587 IO_GAP(dio, first) <= maxgap) { 588 first = dio; 589 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) 590 mandatory = first; 591 } 592 593 /* 594 * Skip any initial optional I/Os. 595 */ 596 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { 597 first = AVL_NEXT(t, first); 598 ASSERT(first != NULL); 599 } 600 601 /* 602 * Walk forward through sufficiently contiguous I/Os. 603 */ 604 while ((dio = AVL_NEXT(t, last)) != NULL && 605 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 606 IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit && 607 IO_GAP(last, dio) <= maxgap) { 608 last = dio; 609 if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) 610 mandatory = last; 611 } 612 613 /* 614 * Now that we've established the range of the I/O aggregation 615 * we must decide what to do with trailing optional I/Os. 616 * For reads, there's nothing to do. While we are unable to 617 * aggregate further, it's possible that a trailing optional 618 * I/O would allow the underlying device to aggregate with 619 * subsequent I/Os. We must therefore determine if the next 620 * non-optional I/O is close enough to make aggregation 621 * worthwhile. 622 */ 623 stretch = B_FALSE; 624 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { 625 zio_t *nio = last; 626 while ((dio = AVL_NEXT(t, nio)) != NULL && 627 IO_GAP(nio, dio) == 0 && 628 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { 629 nio = dio; 630 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { 631 stretch = B_TRUE; 632 break; 633 } 634 } 635 } 636 637 if (stretch) { 638 /* This may be a no-op. */ 639 dio = AVL_NEXT(t, last); 640 dio->io_flags &= ~ZIO_FLAG_OPTIONAL; 641 } else { 642 while (last != mandatory && last != first) { 643 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); 644 last = AVL_PREV(t, last); 645 ASSERT(last != NULL); 646 } 647 } 648 649 if (first == last) 650 return (NULL); 651 652 size = IO_SPAN(first, last); 653 ASSERT3U(size, <=, zfs_vdev_aggregation_limit); 654 655 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, 656 zio_buf_alloc(size), size, first->io_type, zio->io_priority, 657 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE, 658 vdev_queue_agg_io_done, NULL); 659 aio->io_timestamp = first->io_timestamp; 660 661 nio = first; 662 do { 663 dio = nio; 664 nio = AVL_NEXT(t, dio); 665 ASSERT3U(dio->io_type, ==, aio->io_type); 666 667 if (dio->io_flags & ZIO_FLAG_NODATA) { 668 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); 669 bzero((char *)aio->io_data + (dio->io_offset - 670 aio->io_offset), dio->io_size); 671 } else if (dio->io_type == ZIO_TYPE_WRITE) { 672 bcopy(dio->io_data, (char *)aio->io_data + 673 (dio->io_offset - aio->io_offset), 674 dio->io_size); 675 } 676 677 zio_add_child(dio, aio); 678 vdev_queue_io_remove(vq, dio); 679 zio_vdev_io_bypass(dio); 680 zio_execute(dio); 681 } while (dio != last); 682 683 return (aio); 684} 685 686static zio_t * 687vdev_queue_io_to_issue(vdev_queue_t *vq) 688{ 689 zio_t *zio, *aio; 690 zio_priority_t p; 691 avl_index_t idx; 692 vdev_queue_class_t *vqc; 693 zio_t search; 694 695again: 696 ASSERT(MUTEX_HELD(&vq->vq_lock)); 697 698 p = vdev_queue_class_to_issue(vq); 699 700 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { 701 /* No eligible queued i/os */ 702 return (NULL); 703 } 704 705 /* 706 * For LBA-ordered queues (async / scrub), issue the i/o which follows 707 * the most recently issued i/o in LBA (offset) order. 708 * 709 * For FIFO queues (sync), issue the i/o with the lowest timestamp. 710 */ 711 vqc = &vq->vq_class[p]; 712 search.io_timestamp = 0; 713 search.io_offset = vq->vq_last_offset + 1; 714 VERIFY3P(avl_find(&vqc->vqc_queued_tree, &search, &idx), ==, NULL); 715 zio = avl_nearest(&vqc->vqc_queued_tree, idx, AVL_AFTER); 716 if (zio == NULL) 717 zio = avl_first(&vqc->vqc_queued_tree); 718 ASSERT3U(zio->io_priority, ==, p); 719 720 aio = vdev_queue_aggregate(vq, zio); 721 if (aio != NULL) 722 zio = aio; 723 else 724 vdev_queue_io_remove(vq, zio); 725 726 /* 727 * If the I/O is or was optional and therefore has no data, we need to 728 * simply discard it. We need to drop the vdev queue's lock to avoid a 729 * deadlock that we could encounter since this I/O will complete 730 * immediately. 731 */ 732 if (zio->io_flags & ZIO_FLAG_NODATA) { 733 mutex_exit(&vq->vq_lock); 734 zio_vdev_io_bypass(zio); 735 zio_execute(zio); 736 mutex_enter(&vq->vq_lock); 737 goto again; 738 } 739 740 vdev_queue_pending_add(vq, zio); 741 vq->vq_last_offset = zio->io_offset; 742 743 return (zio); 744} 745 746zio_t * 747vdev_queue_io(zio_t *zio) 748{ 749 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 750 zio_t *nio; 751 752 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) 753 return (zio); 754 755 /* 756 * Children i/os inherent their parent's priority, which might 757 * not match the child's i/o type. Fix it up here. 758 */ 759 if (zio->io_type == ZIO_TYPE_READ) { 760 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && 761 zio->io_priority != ZIO_PRIORITY_ASYNC_READ && 762 zio->io_priority != ZIO_PRIORITY_SCRUB) 763 zio->io_priority = ZIO_PRIORITY_ASYNC_READ; 764 } else if (zio->io_type == ZIO_TYPE_WRITE) { 765 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && 766 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE) 767 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; 768 } else { 769 ASSERT(zio->io_type == ZIO_TYPE_FREE); 770 zio->io_priority = ZIO_PRIORITY_TRIM; 771 } 772 773 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; 774 775 mutex_enter(&vq->vq_lock); 776 zio->io_timestamp = gethrtime(); 777 vdev_queue_io_add(vq, zio); 778 nio = vdev_queue_io_to_issue(vq); 779 mutex_exit(&vq->vq_lock); 780 781 if (nio == NULL) 782 return (NULL); 783 784 if (nio->io_done == vdev_queue_agg_io_done) { 785 zio_nowait(nio); 786 return (NULL); 787 } 788 789 return (nio); 790} 791 792void 793vdev_queue_io_done(zio_t *zio) 794{ 795 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 796 zio_t *nio; 797 798 if (zio_injection_enabled) 799 delay(SEC_TO_TICK(zio_handle_io_delay(zio))); 800 801 mutex_enter(&vq->vq_lock); 802 803 vdev_queue_pending_remove(vq, zio); 804 805 vq->vq_io_complete_ts = gethrtime(); 806 807 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { 808 mutex_exit(&vq->vq_lock); 809 if (nio->io_done == vdev_queue_agg_io_done) { 810 zio_nowait(nio); 811 } else { 812 zio_vdev_io_reissue(nio); 813 zio_execute(nio); 814 } 815 mutex_enter(&vq->vq_lock); 816 } 817 818 mutex_exit(&vq->vq_lock); 819} 820 821/* 822 * As these three methods are only used for load calculations we're not concerned 823 * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex 824 * use here, instead we prefer to keep it lock free for performance. 825 */ 826int 827vdev_queue_length(vdev_t *vd) 828{ 829 return (avl_numnodes(&vd->vdev_queue.vq_active_tree)); 830} 831 832uint64_t 833vdev_queue_lastoffset(vdev_t *vd) 834{ 835 return (vd->vdev_queue.vq_lastoffset); 836} 837 838void 839vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio) 840{ 841 vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size; 842} 843