vdev_queue.c revision 270312
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 294void 295vdev_queue_fini(vdev_t *vd) 296{ 297 vdev_queue_t *vq = &vd->vdev_queue; 298 299 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) 300 avl_destroy(&vq->vq_class[p].vqc_queued_tree); 301 avl_destroy(&vq->vq_active_tree); 302 303 mutex_destroy(&vq->vq_lock); 304} 305 306static void 307vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) 308{ 309 spa_t *spa = zio->io_spa; 310 ASSERT(MUTEX_HELD(&vq->vq_lock)); 311 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 312 avl_add(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio); 313 314#ifdef illumos 315 mutex_enter(&spa->spa_iokstat_lock); 316 spa->spa_queue_stats[zio->io_priority].spa_queued++; 317 if (spa->spa_iokstat != NULL) 318 kstat_waitq_enter(spa->spa_iokstat->ks_data); 319 mutex_exit(&spa->spa_iokstat_lock); 320#endif 321} 322 323static void 324vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) 325{ 326 spa_t *spa = zio->io_spa; 327 ASSERT(MUTEX_HELD(&vq->vq_lock)); 328 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 329 avl_remove(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio); 330 331#ifdef illumos 332 mutex_enter(&spa->spa_iokstat_lock); 333 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0); 334 spa->spa_queue_stats[zio->io_priority].spa_queued--; 335 if (spa->spa_iokstat != NULL) 336 kstat_waitq_exit(spa->spa_iokstat->ks_data); 337 mutex_exit(&spa->spa_iokstat_lock); 338#endif 339} 340 341static void 342vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) 343{ 344 spa_t *spa = zio->io_spa; 345 ASSERT(MUTEX_HELD(&vq->vq_lock)); 346 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 347 vq->vq_class[zio->io_priority].vqc_active++; 348 avl_add(&vq->vq_active_tree, zio); 349 350#ifdef illumos 351 mutex_enter(&spa->spa_iokstat_lock); 352 spa->spa_queue_stats[zio->io_priority].spa_active++; 353 if (spa->spa_iokstat != NULL) 354 kstat_runq_enter(spa->spa_iokstat->ks_data); 355 mutex_exit(&spa->spa_iokstat_lock); 356#endif 357} 358 359static void 360vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) 361{ 362 spa_t *spa = zio->io_spa; 363 ASSERT(MUTEX_HELD(&vq->vq_lock)); 364 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 365 vq->vq_class[zio->io_priority].vqc_active--; 366 avl_remove(&vq->vq_active_tree, zio); 367 368#ifdef illumos 369 mutex_enter(&spa->spa_iokstat_lock); 370 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0); 371 spa->spa_queue_stats[zio->io_priority].spa_active--; 372 if (spa->spa_iokstat != NULL) { 373 kstat_io_t *ksio = spa->spa_iokstat->ks_data; 374 375 kstat_runq_exit(spa->spa_iokstat->ks_data); 376 if (zio->io_type == ZIO_TYPE_READ) { 377 ksio->reads++; 378 ksio->nread += zio->io_size; 379 } else if (zio->io_type == ZIO_TYPE_WRITE) { 380 ksio->writes++; 381 ksio->nwritten += zio->io_size; 382 } 383 } 384 mutex_exit(&spa->spa_iokstat_lock); 385#endif 386} 387 388static void 389vdev_queue_agg_io_done(zio_t *aio) 390{ 391 if (aio->io_type == ZIO_TYPE_READ) { 392 zio_t *pio; 393 while ((pio = zio_walk_parents(aio)) != NULL) { 394 bcopy((char *)aio->io_data + (pio->io_offset - 395 aio->io_offset), pio->io_data, pio->io_size); 396 } 397 } 398 399 zio_buf_free(aio->io_data, aio->io_size); 400} 401 402static int 403vdev_queue_class_min_active(zio_priority_t p) 404{ 405 switch (p) { 406 case ZIO_PRIORITY_SYNC_READ: 407 return (zfs_vdev_sync_read_min_active); 408 case ZIO_PRIORITY_SYNC_WRITE: 409 return (zfs_vdev_sync_write_min_active); 410 case ZIO_PRIORITY_ASYNC_READ: 411 return (zfs_vdev_async_read_min_active); 412 case ZIO_PRIORITY_ASYNC_WRITE: 413 return (zfs_vdev_async_write_min_active); 414 case ZIO_PRIORITY_SCRUB: 415 return (zfs_vdev_scrub_min_active); 416 case ZIO_PRIORITY_TRIM: 417 return (zfs_vdev_trim_min_active); 418 default: 419 panic("invalid priority %u", p); 420 return (0); 421 } 422} 423 424static int 425vdev_queue_max_async_writes(spa_t *spa) 426{ 427 int writes; 428 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total; 429 uint64_t min_bytes = zfs_dirty_data_max * 430 zfs_vdev_async_write_active_min_dirty_percent / 100; 431 uint64_t max_bytes = zfs_dirty_data_max * 432 zfs_vdev_async_write_active_max_dirty_percent / 100; 433 434 /* 435 * Sync tasks correspond to interactive user actions. To reduce the 436 * execution time of those actions we push data out as fast as possible. 437 */ 438 if (spa_has_pending_synctask(spa)) { 439 return (zfs_vdev_async_write_max_active); 440 } 441 442 if (dirty < min_bytes) 443 return (zfs_vdev_async_write_min_active); 444 if (dirty > max_bytes) 445 return (zfs_vdev_async_write_max_active); 446 447 /* 448 * linear interpolation: 449 * slope = (max_writes - min_writes) / (max_bytes - min_bytes) 450 * move right by min_bytes 451 * move up by min_writes 452 */ 453 writes = (dirty - min_bytes) * 454 (zfs_vdev_async_write_max_active - 455 zfs_vdev_async_write_min_active) / 456 (max_bytes - min_bytes) + 457 zfs_vdev_async_write_min_active; 458 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); 459 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); 460 return (writes); 461} 462 463static int 464vdev_queue_class_max_active(spa_t *spa, zio_priority_t p) 465{ 466 switch (p) { 467 case ZIO_PRIORITY_SYNC_READ: 468 return (zfs_vdev_sync_read_max_active); 469 case ZIO_PRIORITY_SYNC_WRITE: 470 return (zfs_vdev_sync_write_max_active); 471 case ZIO_PRIORITY_ASYNC_READ: 472 return (zfs_vdev_async_read_max_active); 473 case ZIO_PRIORITY_ASYNC_WRITE: 474 return (vdev_queue_max_async_writes(spa)); 475 case ZIO_PRIORITY_SCRUB: 476 return (zfs_vdev_scrub_max_active); 477 case ZIO_PRIORITY_TRIM: 478 return (zfs_vdev_trim_max_active); 479 default: 480 panic("invalid priority %u", p); 481 return (0); 482 } 483} 484 485/* 486 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if 487 * there is no eligible class. 488 */ 489static zio_priority_t 490vdev_queue_class_to_issue(vdev_queue_t *vq) 491{ 492 spa_t *spa = vq->vq_vdev->vdev_spa; 493 zio_priority_t p; 494 495 ASSERT(MUTEX_HELD(&vq->vq_lock)); 496 497 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active) 498 return (ZIO_PRIORITY_NUM_QUEUEABLE); 499 500 /* find a queue that has not reached its minimum # outstanding i/os */ 501 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 502 if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 && 503 vq->vq_class[p].vqc_active < 504 vdev_queue_class_min_active(p)) 505 return (p); 506 } 507 508 /* 509 * If we haven't found a queue, look for one that hasn't reached its 510 * maximum # outstanding i/os. 511 */ 512 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 513 if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 && 514 vq->vq_class[p].vqc_active < 515 vdev_queue_class_max_active(spa, p)) 516 return (p); 517 } 518 519 /* No eligible queued i/os */ 520 return (ZIO_PRIORITY_NUM_QUEUEABLE); 521} 522 523/* 524 * Compute the range spanned by two i/os, which is the endpoint of the last 525 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). 526 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); 527 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. 528 */ 529#define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) 530#define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) 531 532static zio_t * 533vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) 534{ 535 zio_t *first, *last, *aio, *dio, *mandatory, *nio; 536 uint64_t maxgap = 0; 537 uint64_t size; 538 boolean_t stretch; 539 avl_tree_t *t; 540 enum zio_flag flags; 541 542 ASSERT(MUTEX_HELD(&vq->vq_lock)); 543 544 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE) 545 return (NULL); 546 547 /* 548 * The synchronous i/o queues are not sorted by LBA, so we can't 549 * find adjacent i/os. These i/os tend to not be tightly clustered, 550 * or too large to aggregate, so this has little impact on performance. 551 */ 552 if (zio->io_priority == ZIO_PRIORITY_SYNC_READ || 553 zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) 554 return (NULL); 555 556 first = last = zio; 557 558 if (zio->io_type == ZIO_TYPE_READ) 559 maxgap = zfs_vdev_read_gap_limit; 560 561 /* 562 * We can aggregate I/Os that are sufficiently adjacent and of 563 * the same flavor, as expressed by the AGG_INHERIT flags. 564 * The latter requirement is necessary so that certain 565 * attributes of the I/O, such as whether it's a normal I/O 566 * or a scrub/resilver, can be preserved in the aggregate. 567 * We can include optional I/Os, but don't allow them 568 * to begin a range as they add no benefit in that situation. 569 */ 570 571 /* 572 * We keep track of the last non-optional I/O. 573 */ 574 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; 575 576 /* 577 * Walk backwards through sufficiently contiguous I/Os 578 * recording the last non-option I/O. 579 */ 580 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; 581 t = &vq->vq_class[zio->io_priority].vqc_queued_tree; 582 while ((dio = AVL_PREV(t, first)) != NULL && 583 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 584 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit && 585 IO_GAP(dio, first) <= maxgap) { 586 first = dio; 587 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) 588 mandatory = first; 589 } 590 591 /* 592 * Skip any initial optional I/Os. 593 */ 594 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { 595 first = AVL_NEXT(t, first); 596 ASSERT(first != NULL); 597 } 598 599 /* 600 * Walk forward through sufficiently contiguous I/Os. 601 */ 602 while ((dio = AVL_NEXT(t, last)) != NULL && 603 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 604 IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit && 605 IO_GAP(last, dio) <= maxgap) { 606 last = dio; 607 if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) 608 mandatory = last; 609 } 610 611 /* 612 * Now that we've established the range of the I/O aggregation 613 * we must decide what to do with trailing optional I/Os. 614 * For reads, there's nothing to do. While we are unable to 615 * aggregate further, it's possible that a trailing optional 616 * I/O would allow the underlying device to aggregate with 617 * subsequent I/Os. We must therefore determine if the next 618 * non-optional I/O is close enough to make aggregation 619 * worthwhile. 620 */ 621 stretch = B_FALSE; 622 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { 623 zio_t *nio = last; 624 while ((dio = AVL_NEXT(t, nio)) != NULL && 625 IO_GAP(nio, dio) == 0 && 626 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { 627 nio = dio; 628 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { 629 stretch = B_TRUE; 630 break; 631 } 632 } 633 } 634 635 if (stretch) { 636 /* This may be a no-op. */ 637 dio = AVL_NEXT(t, last); 638 dio->io_flags &= ~ZIO_FLAG_OPTIONAL; 639 } else { 640 while (last != mandatory && last != first) { 641 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); 642 last = AVL_PREV(t, last); 643 ASSERT(last != NULL); 644 } 645 } 646 647 if (first == last) 648 return (NULL); 649 650 size = IO_SPAN(first, last); 651 ASSERT3U(size, <=, zfs_vdev_aggregation_limit); 652 653 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, 654 zio_buf_alloc(size), size, first->io_type, zio->io_priority, 655 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE, 656 vdev_queue_agg_io_done, NULL); 657 aio->io_timestamp = first->io_timestamp; 658 659 nio = first; 660 do { 661 dio = nio; 662 nio = AVL_NEXT(t, dio); 663 ASSERT3U(dio->io_type, ==, aio->io_type); 664 665 if (dio->io_flags & ZIO_FLAG_NODATA) { 666 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); 667 bzero((char *)aio->io_data + (dio->io_offset - 668 aio->io_offset), dio->io_size); 669 } else if (dio->io_type == ZIO_TYPE_WRITE) { 670 bcopy(dio->io_data, (char *)aio->io_data + 671 (dio->io_offset - aio->io_offset), 672 dio->io_size); 673 } 674 675 zio_add_child(dio, aio); 676 vdev_queue_io_remove(vq, dio); 677 zio_vdev_io_bypass(dio); 678 zio_execute(dio); 679 } while (dio != last); 680 681 return (aio); 682} 683 684static zio_t * 685vdev_queue_io_to_issue(vdev_queue_t *vq) 686{ 687 zio_t *zio, *aio; 688 zio_priority_t p; 689 avl_index_t idx; 690 vdev_queue_class_t *vqc; 691 zio_t search; 692 693again: 694 ASSERT(MUTEX_HELD(&vq->vq_lock)); 695 696 p = vdev_queue_class_to_issue(vq); 697 698 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { 699 /* No eligible queued i/os */ 700 return (NULL); 701 } 702 703 /* 704 * For LBA-ordered queues (async / scrub), issue the i/o which follows 705 * the most recently issued i/o in LBA (offset) order. 706 * 707 * For FIFO queues (sync), issue the i/o with the lowest timestamp. 708 */ 709 vqc = &vq->vq_class[p]; 710 search.io_timestamp = 0; 711 search.io_offset = vq->vq_last_offset + 1; 712 VERIFY3P(avl_find(&vqc->vqc_queued_tree, &search, &idx), ==, NULL); 713 zio = avl_nearest(&vqc->vqc_queued_tree, idx, AVL_AFTER); 714 if (zio == NULL) 715 zio = avl_first(&vqc->vqc_queued_tree); 716 ASSERT3U(zio->io_priority, ==, p); 717 718 aio = vdev_queue_aggregate(vq, zio); 719 if (aio != NULL) 720 zio = aio; 721 else 722 vdev_queue_io_remove(vq, zio); 723 724 /* 725 * If the I/O is or was optional and therefore has no data, we need to 726 * simply discard it. We need to drop the vdev queue's lock to avoid a 727 * deadlock that we could encounter since this I/O will complete 728 * immediately. 729 */ 730 if (zio->io_flags & ZIO_FLAG_NODATA) { 731 mutex_exit(&vq->vq_lock); 732 zio_vdev_io_bypass(zio); 733 zio_execute(zio); 734 mutex_enter(&vq->vq_lock); 735 goto again; 736 } 737 738 vdev_queue_pending_add(vq, zio); 739 vq->vq_last_offset = zio->io_offset; 740 741 return (zio); 742} 743 744zio_t * 745vdev_queue_io(zio_t *zio) 746{ 747 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 748 zio_t *nio; 749 750 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) 751 return (zio); 752 753 /* 754 * Children i/os inherent their parent's priority, which might 755 * not match the child's i/o type. Fix it up here. 756 */ 757 if (zio->io_type == ZIO_TYPE_READ) { 758 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && 759 zio->io_priority != ZIO_PRIORITY_ASYNC_READ && 760 zio->io_priority != ZIO_PRIORITY_SCRUB) 761 zio->io_priority = ZIO_PRIORITY_ASYNC_READ; 762 } else if (zio->io_type == ZIO_TYPE_WRITE) { 763 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && 764 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE) 765 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; 766 } else { 767 ASSERT(zio->io_type == ZIO_TYPE_FREE); 768 zio->io_priority = ZIO_PRIORITY_TRIM; 769 } 770 771 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; 772 773 mutex_enter(&vq->vq_lock); 774 zio->io_timestamp = gethrtime(); 775 vdev_queue_io_add(vq, zio); 776 nio = vdev_queue_io_to_issue(vq); 777 mutex_exit(&vq->vq_lock); 778 779 if (nio == NULL) 780 return (NULL); 781 782 if (nio->io_done == vdev_queue_agg_io_done) { 783 zio_nowait(nio); 784 return (NULL); 785 } 786 787 return (nio); 788} 789 790void 791vdev_queue_io_done(zio_t *zio) 792{ 793 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 794 zio_t *nio; 795 796 if (zio_injection_enabled) 797 delay(SEC_TO_TICK(zio_handle_io_delay(zio))); 798 799 mutex_enter(&vq->vq_lock); 800 801 vdev_queue_pending_remove(vq, zio); 802 803 vq->vq_io_complete_ts = gethrtime(); 804 805 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { 806 mutex_exit(&vq->vq_lock); 807 if (nio->io_done == vdev_queue_agg_io_done) { 808 zio_nowait(nio); 809 } else { 810 zio_vdev_io_reissue(nio); 811 zio_execute(nio); 812 } 813 mutex_enter(&vq->vq_lock); 814 } 815 816 mutex_exit(&vq->vq_lock); 817} 818