vdev_queue.c revision 287337
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 180static int sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS); 181SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_min_dirty_percent, 182 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int), 183 sysctl_zfs_async_write_active_min_dirty_percent, "I", 184 "Percentage of async write dirty data below which " 185 "async_write_min_active is used."); 186 187static int sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS); 188SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_max_dirty_percent, 189 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int), 190 sysctl_zfs_async_write_active_max_dirty_percent, "I", 191 "Percentage of async write dirty data above which " 192 "async_write_max_active is used."); 193 194SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RWTUN, 195 &zfs_vdev_max_active, 0, 196 "The maximum number of I/Os of all types active for each device."); 197 198#define ZFS_VDEV_QUEUE_KNOB_MIN(name) \ 199SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, CTLFLAG_RWTUN,\ 200 &zfs_vdev_ ## name ## _min_active, 0, \ 201 "Initial number of I/O requests of type " #name \ 202 " active for each device"); 203 204#define ZFS_VDEV_QUEUE_KNOB_MAX(name) \ 205SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, CTLFLAG_RWTUN,\ 206 &zfs_vdev_ ## name ## _max_active, 0, \ 207 "Maximum number of I/O requests of type " #name \ 208 " active for each device"); 209 210ZFS_VDEV_QUEUE_KNOB_MIN(sync_read); 211ZFS_VDEV_QUEUE_KNOB_MAX(sync_read); 212ZFS_VDEV_QUEUE_KNOB_MIN(sync_write); 213ZFS_VDEV_QUEUE_KNOB_MAX(sync_write); 214ZFS_VDEV_QUEUE_KNOB_MIN(async_read); 215ZFS_VDEV_QUEUE_KNOB_MAX(async_read); 216ZFS_VDEV_QUEUE_KNOB_MIN(async_write); 217ZFS_VDEV_QUEUE_KNOB_MAX(async_write); 218ZFS_VDEV_QUEUE_KNOB_MIN(scrub); 219ZFS_VDEV_QUEUE_KNOB_MAX(scrub); 220ZFS_VDEV_QUEUE_KNOB_MIN(trim); 221ZFS_VDEV_QUEUE_KNOB_MAX(trim); 222 223#undef ZFS_VDEV_QUEUE_KNOB 224 225SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RWTUN, 226 &zfs_vdev_aggregation_limit, 0, 227 "I/O requests are aggregated up to this size"); 228SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RWTUN, 229 &zfs_vdev_read_gap_limit, 0, 230 "Acceptable gap between two reads being aggregated"); 231SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RWTUN, 232 &zfs_vdev_write_gap_limit, 0, 233 "Acceptable gap between two writes being aggregated"); 234 235static int 236sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS) 237{ 238 int val, err; 239 240 val = zfs_vdev_async_write_active_min_dirty_percent; 241 err = sysctl_handle_int(oidp, &val, 0, req); 242 if (err != 0 || req->newptr == NULL) 243 return (err); 244 245 if (val < 0 || val > 100 || 246 val >= zfs_vdev_async_write_active_max_dirty_percent) 247 return (EINVAL); 248 249 zfs_vdev_async_write_active_min_dirty_percent = val; 250 251 return (0); 252} 253 254static int 255sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS) 256{ 257 int val, err; 258 259 val = zfs_vdev_async_write_active_max_dirty_percent; 260 err = sysctl_handle_int(oidp, &val, 0, req); 261 if (err != 0 || req->newptr == NULL) 262 return (err); 263 264 if (val < 0 || val > 100 || 265 val <= zfs_vdev_async_write_active_min_dirty_percent) 266 return (EINVAL); 267 268 zfs_vdev_async_write_active_max_dirty_percent = val; 269 270 return (0); 271} 272#endif 273 274int 275vdev_queue_offset_compare(const void *x1, const void *x2) 276{ 277 const zio_t *z1 = x1; 278 const zio_t *z2 = x2; 279 280 if (z1->io_offset < z2->io_offset) 281 return (-1); 282 if (z1->io_offset > z2->io_offset) 283 return (1); 284 285 if (z1 < z2) 286 return (-1); 287 if (z1 > z2) 288 return (1); 289 290 return (0); 291} 292 293static inline avl_tree_t * 294vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p) 295{ 296 return (&vq->vq_class[p].vqc_queued_tree); 297} 298 299static inline avl_tree_t * 300vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t) 301{ 302 if (t == ZIO_TYPE_READ) 303 return (&vq->vq_read_offset_tree); 304 else if (t == ZIO_TYPE_WRITE) 305 return (&vq->vq_write_offset_tree); 306 else 307 return (NULL); 308} 309 310int 311vdev_queue_timestamp_compare(const void *x1, const void *x2) 312{ 313 const zio_t *z1 = x1; 314 const zio_t *z2 = x2; 315 316 if (z1->io_timestamp < z2->io_timestamp) 317 return (-1); 318 if (z1->io_timestamp > z2->io_timestamp) 319 return (1); 320 321 if (z1->io_offset < z2->io_offset) 322 return (-1); 323 if (z1->io_offset > z2->io_offset) 324 return (1); 325 326 if (z1 < z2) 327 return (-1); 328 if (z1 > z2) 329 return (1); 330 331 return (0); 332} 333 334void 335vdev_queue_init(vdev_t *vd) 336{ 337 vdev_queue_t *vq = &vd->vdev_queue; 338 339 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); 340 vq->vq_vdev = vd; 341 342 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare, 343 sizeof (zio_t), offsetof(struct zio, io_queue_node)); 344 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ), 345 vdev_queue_offset_compare, sizeof (zio_t), 346 offsetof(struct zio, io_offset_node)); 347 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE), 348 vdev_queue_offset_compare, sizeof (zio_t), 349 offsetof(struct zio, io_offset_node)); 350 351 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 352 int (*compfn) (const void *, const void *); 353 354 /* 355 * The synchronous i/o queues are dispatched in FIFO rather 356 * than LBA order. This provides more consistent latency for 357 * these i/os. 358 */ 359 if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE) 360 compfn = vdev_queue_timestamp_compare; 361 else 362 compfn = vdev_queue_offset_compare; 363 364 avl_create(vdev_queue_class_tree(vq, p), compfn, 365 sizeof (zio_t), offsetof(struct zio, io_queue_node)); 366 } 367 368 vq->vq_lastoffset = 0; 369} 370 371void 372vdev_queue_fini(vdev_t *vd) 373{ 374 vdev_queue_t *vq = &vd->vdev_queue; 375 376 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) 377 avl_destroy(vdev_queue_class_tree(vq, p)); 378 avl_destroy(&vq->vq_active_tree); 379 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ)); 380 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE)); 381 382 mutex_destroy(&vq->vq_lock); 383} 384 385static void 386vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) 387{ 388 spa_t *spa = zio->io_spa; 389 avl_tree_t *qtt; 390 ASSERT(MUTEX_HELD(&vq->vq_lock)); 391 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 392 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio); 393 qtt = vdev_queue_type_tree(vq, zio->io_type); 394 if (qtt) 395 avl_add(qtt, zio); 396 397#ifdef illumos 398 mutex_enter(&spa->spa_iokstat_lock); 399 spa->spa_queue_stats[zio->io_priority].spa_queued++; 400 if (spa->spa_iokstat != NULL) 401 kstat_waitq_enter(spa->spa_iokstat->ks_data); 402 mutex_exit(&spa->spa_iokstat_lock); 403#endif 404} 405 406static void 407vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) 408{ 409 spa_t *spa = zio->io_spa; 410 avl_tree_t *qtt; 411 ASSERT(MUTEX_HELD(&vq->vq_lock)); 412 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 413 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio); 414 qtt = vdev_queue_type_tree(vq, zio->io_type); 415 if (qtt) 416 avl_remove(qtt, zio); 417 418#ifdef illumos 419 mutex_enter(&spa->spa_iokstat_lock); 420 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0); 421 spa->spa_queue_stats[zio->io_priority].spa_queued--; 422 if (spa->spa_iokstat != NULL) 423 kstat_waitq_exit(spa->spa_iokstat->ks_data); 424 mutex_exit(&spa->spa_iokstat_lock); 425#endif 426} 427 428static void 429vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) 430{ 431 spa_t *spa = zio->io_spa; 432 ASSERT(MUTEX_HELD(&vq->vq_lock)); 433 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 434 vq->vq_class[zio->io_priority].vqc_active++; 435 avl_add(&vq->vq_active_tree, zio); 436 437#ifdef illumos 438 mutex_enter(&spa->spa_iokstat_lock); 439 spa->spa_queue_stats[zio->io_priority].spa_active++; 440 if (spa->spa_iokstat != NULL) 441 kstat_runq_enter(spa->spa_iokstat->ks_data); 442 mutex_exit(&spa->spa_iokstat_lock); 443#endif 444} 445 446static void 447vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) 448{ 449 spa_t *spa = zio->io_spa; 450 ASSERT(MUTEX_HELD(&vq->vq_lock)); 451 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 452 vq->vq_class[zio->io_priority].vqc_active--; 453 avl_remove(&vq->vq_active_tree, zio); 454 455#ifdef illumos 456 mutex_enter(&spa->spa_iokstat_lock); 457 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0); 458 spa->spa_queue_stats[zio->io_priority].spa_active--; 459 if (spa->spa_iokstat != NULL) { 460 kstat_io_t *ksio = spa->spa_iokstat->ks_data; 461 462 kstat_runq_exit(spa->spa_iokstat->ks_data); 463 if (zio->io_type == ZIO_TYPE_READ) { 464 ksio->reads++; 465 ksio->nread += zio->io_size; 466 } else if (zio->io_type == ZIO_TYPE_WRITE) { 467 ksio->writes++; 468 ksio->nwritten += zio->io_size; 469 } 470 } 471 mutex_exit(&spa->spa_iokstat_lock); 472#endif 473} 474 475static void 476vdev_queue_agg_io_done(zio_t *aio) 477{ 478 if (aio->io_type == ZIO_TYPE_READ) { 479 zio_t *pio; 480 while ((pio = zio_walk_parents(aio)) != NULL) { 481 bcopy((char *)aio->io_data + (pio->io_offset - 482 aio->io_offset), pio->io_data, pio->io_size); 483 } 484 } 485 486 zio_buf_free(aio->io_data, aio->io_size); 487} 488 489static int 490vdev_queue_class_min_active(zio_priority_t p) 491{ 492 switch (p) { 493 case ZIO_PRIORITY_SYNC_READ: 494 return (zfs_vdev_sync_read_min_active); 495 case ZIO_PRIORITY_SYNC_WRITE: 496 return (zfs_vdev_sync_write_min_active); 497 case ZIO_PRIORITY_ASYNC_READ: 498 return (zfs_vdev_async_read_min_active); 499 case ZIO_PRIORITY_ASYNC_WRITE: 500 return (zfs_vdev_async_write_min_active); 501 case ZIO_PRIORITY_SCRUB: 502 return (zfs_vdev_scrub_min_active); 503 case ZIO_PRIORITY_TRIM: 504 return (zfs_vdev_trim_min_active); 505 default: 506 panic("invalid priority %u", p); 507 return (0); 508 } 509} 510 511static __noinline int 512vdev_queue_max_async_writes(spa_t *spa) 513{ 514 int writes; 515 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total; 516 uint64_t min_bytes = zfs_dirty_data_max * 517 zfs_vdev_async_write_active_min_dirty_percent / 100; 518 uint64_t max_bytes = zfs_dirty_data_max * 519 zfs_vdev_async_write_active_max_dirty_percent / 100; 520 521 /* 522 * Sync tasks correspond to interactive user actions. To reduce the 523 * execution time of those actions we push data out as fast as possible. 524 */ 525 if (spa_has_pending_synctask(spa)) { 526 return (zfs_vdev_async_write_max_active); 527 } 528 529 if (dirty < min_bytes) 530 return (zfs_vdev_async_write_min_active); 531 if (dirty > max_bytes) 532 return (zfs_vdev_async_write_max_active); 533 534 /* 535 * linear interpolation: 536 * slope = (max_writes - min_writes) / (max_bytes - min_bytes) 537 * move right by min_bytes 538 * move up by min_writes 539 */ 540 writes = (dirty - min_bytes) * 541 (zfs_vdev_async_write_max_active - 542 zfs_vdev_async_write_min_active) / 543 (max_bytes - min_bytes) + 544 zfs_vdev_async_write_min_active; 545 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); 546 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); 547 return (writes); 548} 549 550static int 551vdev_queue_class_max_active(spa_t *spa, zio_priority_t p) 552{ 553 switch (p) { 554 case ZIO_PRIORITY_SYNC_READ: 555 return (zfs_vdev_sync_read_max_active); 556 case ZIO_PRIORITY_SYNC_WRITE: 557 return (zfs_vdev_sync_write_max_active); 558 case ZIO_PRIORITY_ASYNC_READ: 559 return (zfs_vdev_async_read_max_active); 560 case ZIO_PRIORITY_ASYNC_WRITE: 561 return (vdev_queue_max_async_writes(spa)); 562 case ZIO_PRIORITY_SCRUB: 563 return (zfs_vdev_scrub_max_active); 564 case ZIO_PRIORITY_TRIM: 565 return (zfs_vdev_trim_max_active); 566 default: 567 panic("invalid priority %u", p); 568 return (0); 569 } 570} 571 572/* 573 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if 574 * there is no eligible class. 575 */ 576static zio_priority_t 577vdev_queue_class_to_issue(vdev_queue_t *vq) 578{ 579 spa_t *spa = vq->vq_vdev->vdev_spa; 580 zio_priority_t p; 581 582 ASSERT(MUTEX_HELD(&vq->vq_lock)); 583 584 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active) 585 return (ZIO_PRIORITY_NUM_QUEUEABLE); 586 587 /* find a queue that has not reached its minimum # outstanding i/os */ 588 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 589 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && 590 vq->vq_class[p].vqc_active < 591 vdev_queue_class_min_active(p)) 592 return (p); 593 } 594 595 /* 596 * If we haven't found a queue, look for one that hasn't reached its 597 * maximum # outstanding i/os. 598 */ 599 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 600 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && 601 vq->vq_class[p].vqc_active < 602 vdev_queue_class_max_active(spa, p)) 603 return (p); 604 } 605 606 /* No eligible queued i/os */ 607 return (ZIO_PRIORITY_NUM_QUEUEABLE); 608} 609 610/* 611 * Compute the range spanned by two i/os, which is the endpoint of the last 612 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). 613 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); 614 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. 615 */ 616#define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) 617#define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) 618 619static zio_t * 620vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) 621{ 622 zio_t *first, *last, *aio, *dio, *mandatory, *nio; 623 uint64_t maxgap = 0; 624 uint64_t size; 625 boolean_t stretch; 626 avl_tree_t *t; 627 enum zio_flag flags; 628 629 ASSERT(MUTEX_HELD(&vq->vq_lock)); 630 631 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE) 632 return (NULL); 633 634 first = last = zio; 635 636 if (zio->io_type == ZIO_TYPE_READ) 637 maxgap = zfs_vdev_read_gap_limit; 638 639 /* 640 * We can aggregate I/Os that are sufficiently adjacent and of 641 * the same flavor, as expressed by the AGG_INHERIT flags. 642 * The latter requirement is necessary so that certain 643 * attributes of the I/O, such as whether it's a normal I/O 644 * or a scrub/resilver, can be preserved in the aggregate. 645 * We can include optional I/Os, but don't allow them 646 * to begin a range as they add no benefit in that situation. 647 */ 648 649 /* 650 * We keep track of the last non-optional I/O. 651 */ 652 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; 653 654 /* 655 * Walk backwards through sufficiently contiguous I/Os 656 * recording the last non-option I/O. 657 */ 658 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; 659 t = vdev_queue_type_tree(vq, zio->io_type); 660 while (t != NULL && (dio = AVL_PREV(t, first)) != NULL && 661 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 662 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit && 663 IO_GAP(dio, first) <= maxgap) { 664 first = dio; 665 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) 666 mandatory = first; 667 } 668 669 /* 670 * Skip any initial optional I/Os. 671 */ 672 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { 673 first = AVL_NEXT(t, first); 674 ASSERT(first != NULL); 675 } 676 677 /* 678 * Walk forward through sufficiently contiguous I/Os. 679 */ 680 while ((dio = AVL_NEXT(t, last)) != NULL && 681 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 682 IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit && 683 IO_GAP(last, dio) <= maxgap) { 684 last = dio; 685 if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) 686 mandatory = last; 687 } 688 689 /* 690 * Now that we've established the range of the I/O aggregation 691 * we must decide what to do with trailing optional I/Os. 692 * For reads, there's nothing to do. While we are unable to 693 * aggregate further, it's possible that a trailing optional 694 * I/O would allow the underlying device to aggregate with 695 * subsequent I/Os. We must therefore determine if the next 696 * non-optional I/O is close enough to make aggregation 697 * worthwhile. 698 */ 699 stretch = B_FALSE; 700 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { 701 zio_t *nio = last; 702 while ((dio = AVL_NEXT(t, nio)) != NULL && 703 IO_GAP(nio, dio) == 0 && 704 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { 705 nio = dio; 706 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { 707 stretch = B_TRUE; 708 break; 709 } 710 } 711 } 712 713 if (stretch) { 714 /* This may be a no-op. */ 715 dio = AVL_NEXT(t, last); 716 dio->io_flags &= ~ZIO_FLAG_OPTIONAL; 717 } else { 718 while (last != mandatory && last != first) { 719 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); 720 last = AVL_PREV(t, last); 721 ASSERT(last != NULL); 722 } 723 } 724 725 if (first == last) 726 return (NULL); 727 728 size = IO_SPAN(first, last); 729 ASSERT3U(size, <=, zfs_vdev_aggregation_limit); 730 731 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, 732 zio_buf_alloc(size), size, first->io_type, zio->io_priority, 733 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE, 734 vdev_queue_agg_io_done, NULL); 735 aio->io_timestamp = first->io_timestamp; 736 737 nio = first; 738 do { 739 dio = nio; 740 nio = AVL_NEXT(t, dio); 741 ASSERT3U(dio->io_type, ==, aio->io_type); 742 743 if (dio->io_flags & ZIO_FLAG_NODATA) { 744 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); 745 bzero((char *)aio->io_data + (dio->io_offset - 746 aio->io_offset), dio->io_size); 747 } else if (dio->io_type == ZIO_TYPE_WRITE) { 748 bcopy(dio->io_data, (char *)aio->io_data + 749 (dio->io_offset - aio->io_offset), 750 dio->io_size); 751 } 752 753 zio_add_child(dio, aio); 754 vdev_queue_io_remove(vq, dio); 755 zio_vdev_io_bypass(dio); 756 zio_execute(dio); 757 } while (dio != last); 758 759 return (aio); 760} 761 762static zio_t * 763vdev_queue_io_to_issue(vdev_queue_t *vq) 764{ 765 zio_t *zio, *aio; 766 zio_priority_t p; 767 avl_index_t idx; 768 avl_tree_t *tree; 769 zio_t search; 770 771again: 772 ASSERT(MUTEX_HELD(&vq->vq_lock)); 773 774 p = vdev_queue_class_to_issue(vq); 775 776 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { 777 /* No eligible queued i/os */ 778 return (NULL); 779 } 780 781 /* 782 * For LBA-ordered queues (async / scrub), issue the i/o which follows 783 * the most recently issued i/o in LBA (offset) order. 784 * 785 * For FIFO queues (sync), issue the i/o with the lowest timestamp. 786 */ 787 tree = vdev_queue_class_tree(vq, p); 788 search.io_timestamp = 0; 789 search.io_offset = vq->vq_last_offset + 1; 790 VERIFY3P(avl_find(tree, &search, &idx), ==, NULL); 791 zio = avl_nearest(tree, idx, AVL_AFTER); 792 if (zio == NULL) 793 zio = avl_first(tree); 794 ASSERT3U(zio->io_priority, ==, p); 795 796 aio = vdev_queue_aggregate(vq, zio); 797 if (aio != NULL) 798 zio = aio; 799 else 800 vdev_queue_io_remove(vq, zio); 801 802 /* 803 * If the I/O is or was optional and therefore has no data, we need to 804 * simply discard it. We need to drop the vdev queue's lock to avoid a 805 * deadlock that we could encounter since this I/O will complete 806 * immediately. 807 */ 808 if (zio->io_flags & ZIO_FLAG_NODATA) { 809 mutex_exit(&vq->vq_lock); 810 zio_vdev_io_bypass(zio); 811 zio_execute(zio); 812 mutex_enter(&vq->vq_lock); 813 goto again; 814 } 815 816 vdev_queue_pending_add(vq, zio); 817 vq->vq_last_offset = zio->io_offset; 818 819 return (zio); 820} 821 822zio_t * 823vdev_queue_io(zio_t *zio) 824{ 825 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 826 zio_t *nio; 827 828 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) 829 return (zio); 830 831 /* 832 * Children i/os inherent their parent's priority, which might 833 * not match the child's i/o type. Fix it up here. 834 */ 835 if (zio->io_type == ZIO_TYPE_READ) { 836 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && 837 zio->io_priority != ZIO_PRIORITY_ASYNC_READ && 838 zio->io_priority != ZIO_PRIORITY_SCRUB) 839 zio->io_priority = ZIO_PRIORITY_ASYNC_READ; 840 } else if (zio->io_type == ZIO_TYPE_WRITE) { 841 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && 842 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE) 843 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; 844 } else { 845 ASSERT(zio->io_type == ZIO_TYPE_FREE); 846 zio->io_priority = ZIO_PRIORITY_TRIM; 847 } 848 849 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; 850 851 mutex_enter(&vq->vq_lock); 852 zio->io_timestamp = gethrtime(); 853 vdev_queue_io_add(vq, zio); 854 nio = vdev_queue_io_to_issue(vq); 855 mutex_exit(&vq->vq_lock); 856 857 if (nio == NULL) 858 return (NULL); 859 860 if (nio->io_done == vdev_queue_agg_io_done) { 861 zio_nowait(nio); 862 return (NULL); 863 } 864 865 return (nio); 866} 867 868void 869vdev_queue_io_done(zio_t *zio) 870{ 871 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 872 zio_t *nio; 873 874 if (zio_injection_enabled) 875 delay(SEC_TO_TICK(zio_handle_io_delay(zio))); 876 877 mutex_enter(&vq->vq_lock); 878 879 vdev_queue_pending_remove(vq, zio); 880 881 vq->vq_io_complete_ts = gethrtime(); 882 883 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { 884 mutex_exit(&vq->vq_lock); 885 if (nio->io_done == vdev_queue_agg_io_done) { 886 zio_nowait(nio); 887 } else { 888 zio_vdev_io_reissue(nio); 889 zio_execute(nio); 890 } 891 mutex_enter(&vq->vq_lock); 892 } 893 894 mutex_exit(&vq->vq_lock); 895} 896 897/* 898 * As these three methods are only used for load calculations we're not concerned 899 * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex 900 * use here, instead we prefer to keep it lock free for performance. 901 */ 902int 903vdev_queue_length(vdev_t *vd) 904{ 905 return (avl_numnodes(&vd->vdev_queue.vq_active_tree)); 906} 907 908uint64_t 909vdev_queue_lastoffset(vdev_t *vd) 910{ 911 return (vd->vdev_queue.vq_lastoffset); 912} 913 914void 915vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio) 916{ 917 vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size; 918} 919