vdev_queue.c revision 305273
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26/* 27 * Copyright (c) 2012, 2014 by Delphix. All rights reserved. 28 * Copyright (c) 2014 Integros [integros.com] 29 */ 30 31#include <sys/zfs_context.h> 32#include <sys/vdev_impl.h> 33#include <sys/spa_impl.h> 34#include <sys/zio.h> 35#include <sys/avl.h> 36#include <sys/dsl_pool.h> 37 38/* 39 * ZFS I/O Scheduler 40 * --------------- 41 * 42 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The 43 * I/O scheduler determines when and in what order those operations are 44 * issued. The I/O scheduler divides operations into six I/O classes 45 * prioritized in the following order: sync read, sync write, async read, 46 * async write, scrub/resilver and trim. Each queue defines the minimum and 47 * maximum number of concurrent operations that may be issued to the device. 48 * In addition, the device has an aggregate maximum. Note that the sum of the 49 * per-queue minimums must not exceed the aggregate maximum, and if the 50 * aggregate maximum is equal to or greater than the sum of the per-queue 51 * maximums, the per-queue minimum has no effect. 52 * 53 * For many physical devices, throughput increases with the number of 54 * concurrent operations, but latency typically suffers. Further, physical 55 * devices typically have a limit at which more concurrent operations have no 56 * effect on throughput or can actually cause it to decrease. 57 * 58 * The scheduler selects the next operation to issue by first looking for an 59 * I/O class whose minimum has not been satisfied. Once all are satisfied and 60 * the aggregate maximum has not been hit, the scheduler looks for classes 61 * whose maximum has not been satisfied. Iteration through the I/O classes is 62 * done in the order specified above. No further operations are issued if the 63 * aggregate maximum number of concurrent operations has been hit or if there 64 * are no operations queued for an I/O class that has not hit its maximum. 65 * Every time an I/O is queued or an operation completes, the I/O scheduler 66 * looks for new operations to issue. 67 * 68 * All I/O classes have a fixed maximum number of outstanding operations 69 * except for the async write class. Asynchronous writes represent the data 70 * that is committed to stable storage during the syncing stage for 71 * transaction groups (see txg.c). Transaction groups enter the syncing state 72 * periodically so the number of queued async writes will quickly burst up and 73 * then bleed down to zero. Rather than servicing them as quickly as possible, 74 * the I/O scheduler changes the maximum number of active async write I/Os 75 * according to the amount of dirty data in the pool (see dsl_pool.c). Since 76 * both throughput and latency typically increase with the number of 77 * concurrent operations issued to physical devices, reducing the burstiness 78 * in the number of concurrent operations also stabilizes the response time of 79 * operations from other -- and in particular synchronous -- queues. In broad 80 * strokes, the I/O scheduler will issue more concurrent operations from the 81 * async write queue as there's more dirty data in the pool. 82 * 83 * Async Writes 84 * 85 * The number of concurrent operations issued for the async write I/O class 86 * follows a piece-wise linear function defined by a few adjustable points. 87 * 88 * | o---------| <-- zfs_vdev_async_write_max_active 89 * ^ | /^ | 90 * | | / | | 91 * active | / | | 92 * I/O | / | | 93 * count | / | | 94 * | / | | 95 * |------------o | | <-- zfs_vdev_async_write_min_active 96 * 0|____________^______|_________| 97 * 0% | | 100% of zfs_dirty_data_max 98 * | | 99 * | `-- zfs_vdev_async_write_active_max_dirty_percent 100 * `--------- zfs_vdev_async_write_active_min_dirty_percent 101 * 102 * Until the amount of dirty data exceeds a minimum percentage of the dirty 103 * data allowed in the pool, the I/O scheduler will limit the number of 104 * concurrent operations to the minimum. As that threshold is crossed, the 105 * number of concurrent operations issued increases linearly to the maximum at 106 * the specified maximum percentage of the dirty data allowed in the pool. 107 * 108 * Ideally, the amount of dirty data on a busy pool will stay in the sloped 109 * part of the function between zfs_vdev_async_write_active_min_dirty_percent 110 * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the 111 * maximum percentage, this indicates that the rate of incoming data is 112 * greater than the rate that the backend storage can handle. In this case, we 113 * must further throttle incoming writes (see dmu_tx_delay() for details). 114 */ 115 116/* 117 * The maximum number of I/Os active to each device. Ideally, this will be >= 118 * the sum of each queue's max_active. It must be at least the sum of each 119 * queue's min_active. 120 */ 121uint32_t zfs_vdev_max_active = 1000; 122 123/* 124 * Per-queue limits on the number of I/Os active to each device. If the 125 * sum of the queue's max_active is < zfs_vdev_max_active, then the 126 * min_active comes into play. We will send min_active from each queue, 127 * and then select from queues in the order defined by zio_priority_t. 128 * 129 * In general, smaller max_active's will lead to lower latency of synchronous 130 * operations. Larger max_active's may lead to higher overall throughput, 131 * depending on underlying storage. 132 * 133 * The ratio of the queues' max_actives determines the balance of performance 134 * between reads, writes, and scrubs. E.g., increasing 135 * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete 136 * more quickly, but reads and writes to have higher latency and lower 137 * throughput. 138 */ 139uint32_t zfs_vdev_sync_read_min_active = 10; 140uint32_t zfs_vdev_sync_read_max_active = 10; 141uint32_t zfs_vdev_sync_write_min_active = 10; 142uint32_t zfs_vdev_sync_write_max_active = 10; 143uint32_t zfs_vdev_async_read_min_active = 1; 144uint32_t zfs_vdev_async_read_max_active = 3; 145uint32_t zfs_vdev_async_write_min_active = 1; 146uint32_t zfs_vdev_async_write_max_active = 10; 147uint32_t zfs_vdev_scrub_min_active = 1; 148uint32_t zfs_vdev_scrub_max_active = 2; 149uint32_t zfs_vdev_trim_min_active = 1; 150/* 151 * TRIM max active is large in comparison to the other values due to the fact 152 * that TRIM IOs are coalesced at the device layer. This value is set such 153 * that a typical SSD can process the queued IOs in a single request. 154 */ 155uint32_t zfs_vdev_trim_max_active = 64; 156 157 158/* 159 * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent 160 * dirty data, use zfs_vdev_async_write_min_active. When it has more than 161 * zfs_vdev_async_write_active_max_dirty_percent, use 162 * zfs_vdev_async_write_max_active. The value is linearly interpolated 163 * between min and max. 164 */ 165int zfs_vdev_async_write_active_min_dirty_percent = 30; 166int zfs_vdev_async_write_active_max_dirty_percent = 60; 167 168/* 169 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O. 170 * For read I/Os, we also aggregate across small adjacency gaps; for writes 171 * we include spans of optional I/Os to aid aggregation at the disk even when 172 * they aren't able to help us aggregate at this level. 173 */ 174int zfs_vdev_aggregation_limit = SPA_OLD_MAXBLOCKSIZE; 175int zfs_vdev_read_gap_limit = 32 << 10; 176int zfs_vdev_write_gap_limit = 4 << 10; 177 178#ifdef __FreeBSD__ 179#ifdef _KERNEL 180SYSCTL_DECL(_vfs_zfs_vdev); 181 182TUNABLE_INT("vfs.zfs.vdev.async_write_active_min_dirty_percent", 183 &zfs_vdev_async_write_active_min_dirty_percent); 184static int sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS); 185SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_min_dirty_percent, 186 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int), 187 sysctl_zfs_async_write_active_min_dirty_percent, "I", 188 "Percentage of async write dirty data below which " 189 "async_write_min_active is used."); 190 191TUNABLE_INT("vfs.zfs.vdev.async_write_active_max_dirty_percent", 192 &zfs_vdev_async_write_active_max_dirty_percent); 193static int sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS); 194SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_max_dirty_percent, 195 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int), 196 sysctl_zfs_async_write_active_max_dirty_percent, "I", 197 "Percentage of async write dirty data above which " 198 "async_write_max_active is used."); 199 200TUNABLE_INT("vfs.zfs.vdev.max_active", &zfs_vdev_max_active); 201SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RWTUN, 202 &zfs_vdev_max_active, 0, 203 "The maximum number of I/Os of all types active for each device."); 204 205#define ZFS_VDEV_QUEUE_KNOB_MIN(name) \ 206TUNABLE_INT("vfs.zfs.vdev." #name "_min_active", \ 207 &zfs_vdev_ ## name ## _min_active); \ 208SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, \ 209 CTLFLAG_RWTUN, &zfs_vdev_ ## name ## _min_active, 0, \ 210 "Initial number of I/O requests of type " #name \ 211 " active for each device"); 212 213#define ZFS_VDEV_QUEUE_KNOB_MAX(name) \ 214TUNABLE_INT("vfs.zfs.vdev." #name "_max_active", \ 215 &zfs_vdev_ ## name ## _max_active); \ 216SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, \ 217 CTLFLAG_RWTUN, &zfs_vdev_ ## name ## _max_active, 0, \ 218 "Maximum number of I/O requests of type " #name \ 219 " active for each device"); 220 221ZFS_VDEV_QUEUE_KNOB_MIN(sync_read); 222ZFS_VDEV_QUEUE_KNOB_MAX(sync_read); 223ZFS_VDEV_QUEUE_KNOB_MIN(sync_write); 224ZFS_VDEV_QUEUE_KNOB_MAX(sync_write); 225ZFS_VDEV_QUEUE_KNOB_MIN(async_read); 226ZFS_VDEV_QUEUE_KNOB_MAX(async_read); 227ZFS_VDEV_QUEUE_KNOB_MIN(async_write); 228ZFS_VDEV_QUEUE_KNOB_MAX(async_write); 229ZFS_VDEV_QUEUE_KNOB_MIN(scrub); 230ZFS_VDEV_QUEUE_KNOB_MAX(scrub); 231ZFS_VDEV_QUEUE_KNOB_MIN(trim); 232ZFS_VDEV_QUEUE_KNOB_MAX(trim); 233 234#undef ZFS_VDEV_QUEUE_KNOB 235 236TUNABLE_INT("vfs.zfs.vdev.aggregation_limit", &zfs_vdev_aggregation_limit); 237SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RWTUN, 238 &zfs_vdev_aggregation_limit, 0, 239 "I/O requests are aggregated up to this size"); 240TUNABLE_INT("vfs.zfs.vdev.read_gap_limit", &zfs_vdev_read_gap_limit); 241SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RWTUN, 242 &zfs_vdev_read_gap_limit, 0, 243 "Acceptable gap between two reads being aggregated"); 244TUNABLE_INT("vfs.zfs.vdev.write_gap_limit", &zfs_vdev_write_gap_limit); 245SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RWTUN, 246 &zfs_vdev_write_gap_limit, 0, 247 "Acceptable gap between two writes being aggregated"); 248 249static int 250sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS) 251{ 252 int val, err; 253 254 val = zfs_vdev_async_write_active_min_dirty_percent; 255 err = sysctl_handle_int(oidp, &val, 0, req); 256 if (err != 0 || req->newptr == NULL) 257 return (err); 258 259 if (val < 0 || val > 100 || 260 val >= zfs_vdev_async_write_active_max_dirty_percent) 261 return (EINVAL); 262 263 zfs_vdev_async_write_active_min_dirty_percent = val; 264 265 return (0); 266} 267 268static int 269sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS) 270{ 271 int val, err; 272 273 val = zfs_vdev_async_write_active_max_dirty_percent; 274 err = sysctl_handle_int(oidp, &val, 0, req); 275 if (err != 0 || req->newptr == NULL) 276 return (err); 277 278 if (val < 0 || val > 100 || 279 val <= zfs_vdev_async_write_active_min_dirty_percent) 280 return (EINVAL); 281 282 zfs_vdev_async_write_active_max_dirty_percent = val; 283 284 return (0); 285} 286#endif 287#endif 288 289int 290vdev_queue_offset_compare(const void *x1, const void *x2) 291{ 292 const zio_t *z1 = x1; 293 const zio_t *z2 = x2; 294 295 if (z1->io_offset < z2->io_offset) 296 return (-1); 297 if (z1->io_offset > z2->io_offset) 298 return (1); 299 300 if (z1 < z2) 301 return (-1); 302 if (z1 > z2) 303 return (1); 304 305 return (0); 306} 307 308static inline avl_tree_t * 309vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p) 310{ 311 return (&vq->vq_class[p].vqc_queued_tree); 312} 313 314static inline avl_tree_t * 315vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t) 316{ 317 if (t == ZIO_TYPE_READ) 318 return (&vq->vq_read_offset_tree); 319 else if (t == ZIO_TYPE_WRITE) 320 return (&vq->vq_write_offset_tree); 321 else 322 return (NULL); 323} 324 325int 326vdev_queue_timestamp_compare(const void *x1, const void *x2) 327{ 328 const zio_t *z1 = x1; 329 const zio_t *z2 = x2; 330 331 if (z1->io_timestamp < z2->io_timestamp) 332 return (-1); 333 if (z1->io_timestamp > z2->io_timestamp) 334 return (1); 335 336 if (z1->io_offset < z2->io_offset) 337 return (-1); 338 if (z1->io_offset > z2->io_offset) 339 return (1); 340 341 if (z1 < z2) 342 return (-1); 343 if (z1 > z2) 344 return (1); 345 346 return (0); 347} 348 349void 350vdev_queue_init(vdev_t *vd) 351{ 352 vdev_queue_t *vq = &vd->vdev_queue; 353 354 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); 355 vq->vq_vdev = vd; 356 357 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare, 358 sizeof (zio_t), offsetof(struct zio, io_queue_node)); 359 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ), 360 vdev_queue_offset_compare, sizeof (zio_t), 361 offsetof(struct zio, io_offset_node)); 362 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE), 363 vdev_queue_offset_compare, sizeof (zio_t), 364 offsetof(struct zio, io_offset_node)); 365 366 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 367 int (*compfn) (const void *, const void *); 368 369 /* 370 * The synchronous i/o queues are dispatched in FIFO rather 371 * than LBA order. This provides more consistent latency for 372 * these i/os. 373 */ 374 if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE) 375 compfn = vdev_queue_timestamp_compare; 376 else 377 compfn = vdev_queue_offset_compare; 378 379 avl_create(vdev_queue_class_tree(vq, p), compfn, 380 sizeof (zio_t), offsetof(struct zio, io_queue_node)); 381 } 382 383 vq->vq_lastoffset = 0; 384} 385 386void 387vdev_queue_fini(vdev_t *vd) 388{ 389 vdev_queue_t *vq = &vd->vdev_queue; 390 391 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) 392 avl_destroy(vdev_queue_class_tree(vq, p)); 393 avl_destroy(&vq->vq_active_tree); 394 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ)); 395 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE)); 396 397 mutex_destroy(&vq->vq_lock); 398} 399 400static void 401vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) 402{ 403 spa_t *spa = zio->io_spa; 404 avl_tree_t *qtt; 405 ASSERT(MUTEX_HELD(&vq->vq_lock)); 406 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 407 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio); 408 qtt = vdev_queue_type_tree(vq, zio->io_type); 409 if (qtt) 410 avl_add(qtt, zio); 411 412#ifdef illumos 413 mutex_enter(&spa->spa_iokstat_lock); 414 spa->spa_queue_stats[zio->io_priority].spa_queued++; 415 if (spa->spa_iokstat != NULL) 416 kstat_waitq_enter(spa->spa_iokstat->ks_data); 417 mutex_exit(&spa->spa_iokstat_lock); 418#endif 419} 420 421static void 422vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) 423{ 424 spa_t *spa = zio->io_spa; 425 avl_tree_t *qtt; 426 ASSERT(MUTEX_HELD(&vq->vq_lock)); 427 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 428 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio); 429 qtt = vdev_queue_type_tree(vq, zio->io_type); 430 if (qtt) 431 avl_remove(qtt, zio); 432 433#ifdef illumos 434 mutex_enter(&spa->spa_iokstat_lock); 435 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0); 436 spa->spa_queue_stats[zio->io_priority].spa_queued--; 437 if (spa->spa_iokstat != NULL) 438 kstat_waitq_exit(spa->spa_iokstat->ks_data); 439 mutex_exit(&spa->spa_iokstat_lock); 440#endif 441} 442 443static void 444vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) 445{ 446 spa_t *spa = zio->io_spa; 447 ASSERT(MUTEX_HELD(&vq->vq_lock)); 448 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 449 vq->vq_class[zio->io_priority].vqc_active++; 450 avl_add(&vq->vq_active_tree, zio); 451 452#ifdef illumos 453 mutex_enter(&spa->spa_iokstat_lock); 454 spa->spa_queue_stats[zio->io_priority].spa_active++; 455 if (spa->spa_iokstat != NULL) 456 kstat_runq_enter(spa->spa_iokstat->ks_data); 457 mutex_exit(&spa->spa_iokstat_lock); 458#endif 459} 460 461static void 462vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) 463{ 464 spa_t *spa = zio->io_spa; 465 ASSERT(MUTEX_HELD(&vq->vq_lock)); 466 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 467 vq->vq_class[zio->io_priority].vqc_active--; 468 avl_remove(&vq->vq_active_tree, zio); 469 470#ifdef illumos 471 mutex_enter(&spa->spa_iokstat_lock); 472 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0); 473 spa->spa_queue_stats[zio->io_priority].spa_active--; 474 if (spa->spa_iokstat != NULL) { 475 kstat_io_t *ksio = spa->spa_iokstat->ks_data; 476 477 kstat_runq_exit(spa->spa_iokstat->ks_data); 478 if (zio->io_type == ZIO_TYPE_READ) { 479 ksio->reads++; 480 ksio->nread += zio->io_size; 481 } else if (zio->io_type == ZIO_TYPE_WRITE) { 482 ksio->writes++; 483 ksio->nwritten += zio->io_size; 484 } 485 } 486 mutex_exit(&spa->spa_iokstat_lock); 487#endif 488} 489 490static void 491vdev_queue_agg_io_done(zio_t *aio) 492{ 493 if (aio->io_type == ZIO_TYPE_READ) { 494 zio_t *pio; 495 while ((pio = zio_walk_parents(aio)) != NULL) { 496 bcopy((char *)aio->io_data + (pio->io_offset - 497 aio->io_offset), pio->io_data, pio->io_size); 498 } 499 } 500 501 zio_buf_free(aio->io_data, aio->io_size); 502} 503 504static int 505vdev_queue_class_min_active(zio_priority_t p) 506{ 507 switch (p) { 508 case ZIO_PRIORITY_SYNC_READ: 509 return (zfs_vdev_sync_read_min_active); 510 case ZIO_PRIORITY_SYNC_WRITE: 511 return (zfs_vdev_sync_write_min_active); 512 case ZIO_PRIORITY_ASYNC_READ: 513 return (zfs_vdev_async_read_min_active); 514 case ZIO_PRIORITY_ASYNC_WRITE: 515 return (zfs_vdev_async_write_min_active); 516 case ZIO_PRIORITY_SCRUB: 517 return (zfs_vdev_scrub_min_active); 518 case ZIO_PRIORITY_TRIM: 519 return (zfs_vdev_trim_min_active); 520 default: 521 panic("invalid priority %u", p); 522 return (0); 523 } 524} 525 526static __noinline int 527vdev_queue_max_async_writes(spa_t *spa) 528{ 529 int writes; 530 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total; 531 uint64_t min_bytes = zfs_dirty_data_max * 532 zfs_vdev_async_write_active_min_dirty_percent / 100; 533 uint64_t max_bytes = zfs_dirty_data_max * 534 zfs_vdev_async_write_active_max_dirty_percent / 100; 535 536 /* 537 * Sync tasks correspond to interactive user actions. To reduce the 538 * execution time of those actions we push data out as fast as possible. 539 */ 540 if (spa_has_pending_synctask(spa)) { 541 return (zfs_vdev_async_write_max_active); 542 } 543 544 if (dirty < min_bytes) 545 return (zfs_vdev_async_write_min_active); 546 if (dirty > max_bytes) 547 return (zfs_vdev_async_write_max_active); 548 549 /* 550 * linear interpolation: 551 * slope = (max_writes - min_writes) / (max_bytes - min_bytes) 552 * move right by min_bytes 553 * move up by min_writes 554 */ 555 writes = (dirty - min_bytes) * 556 (zfs_vdev_async_write_max_active - 557 zfs_vdev_async_write_min_active) / 558 (max_bytes - min_bytes) + 559 zfs_vdev_async_write_min_active; 560 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); 561 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); 562 return (writes); 563} 564 565static int 566vdev_queue_class_max_active(spa_t *spa, zio_priority_t p) 567{ 568 switch (p) { 569 case ZIO_PRIORITY_SYNC_READ: 570 return (zfs_vdev_sync_read_max_active); 571 case ZIO_PRIORITY_SYNC_WRITE: 572 return (zfs_vdev_sync_write_max_active); 573 case ZIO_PRIORITY_ASYNC_READ: 574 return (zfs_vdev_async_read_max_active); 575 case ZIO_PRIORITY_ASYNC_WRITE: 576 return (vdev_queue_max_async_writes(spa)); 577 case ZIO_PRIORITY_SCRUB: 578 return (zfs_vdev_scrub_max_active); 579 case ZIO_PRIORITY_TRIM: 580 return (zfs_vdev_trim_max_active); 581 default: 582 panic("invalid priority %u", p); 583 return (0); 584 } 585} 586 587/* 588 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if 589 * there is no eligible class. 590 */ 591static zio_priority_t 592vdev_queue_class_to_issue(vdev_queue_t *vq) 593{ 594 spa_t *spa = vq->vq_vdev->vdev_spa; 595 zio_priority_t p; 596 597 ASSERT(MUTEX_HELD(&vq->vq_lock)); 598 599 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active) 600 return (ZIO_PRIORITY_NUM_QUEUEABLE); 601 602 /* find a queue that has not reached its minimum # outstanding i/os */ 603 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 604 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && 605 vq->vq_class[p].vqc_active < 606 vdev_queue_class_min_active(p)) 607 return (p); 608 } 609 610 /* 611 * If we haven't found a queue, look for one that hasn't reached its 612 * maximum # outstanding i/os. 613 */ 614 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 615 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && 616 vq->vq_class[p].vqc_active < 617 vdev_queue_class_max_active(spa, p)) 618 return (p); 619 } 620 621 /* No eligible queued i/os */ 622 return (ZIO_PRIORITY_NUM_QUEUEABLE); 623} 624 625/* 626 * Compute the range spanned by two i/os, which is the endpoint of the last 627 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). 628 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); 629 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. 630 */ 631#define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) 632#define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) 633 634static zio_t * 635vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) 636{ 637 zio_t *first, *last, *aio, *dio, *mandatory, *nio; 638 uint64_t maxgap = 0; 639 uint64_t size; 640 boolean_t stretch; 641 avl_tree_t *t; 642 enum zio_flag flags; 643 644 ASSERT(MUTEX_HELD(&vq->vq_lock)); 645 646 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE) 647 return (NULL); 648 649 first = last = zio; 650 651 if (zio->io_type == ZIO_TYPE_READ) 652 maxgap = zfs_vdev_read_gap_limit; 653 654 /* 655 * We can aggregate I/Os that are sufficiently adjacent and of 656 * the same flavor, as expressed by the AGG_INHERIT flags. 657 * The latter requirement is necessary so that certain 658 * attributes of the I/O, such as whether it's a normal I/O 659 * or a scrub/resilver, can be preserved in the aggregate. 660 * We can include optional I/Os, but don't allow them 661 * to begin a range as they add no benefit in that situation. 662 */ 663 664 /* 665 * We keep track of the last non-optional I/O. 666 */ 667 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; 668 669 /* 670 * Walk backwards through sufficiently contiguous I/Os 671 * recording the last non-option I/O. 672 */ 673 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; 674 t = vdev_queue_type_tree(vq, zio->io_type); 675 while (t != NULL && (dio = AVL_PREV(t, first)) != NULL && 676 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 677 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit && 678 IO_GAP(dio, first) <= maxgap) { 679 first = dio; 680 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) 681 mandatory = first; 682 } 683 684 /* 685 * Skip any initial optional I/Os. 686 */ 687 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { 688 first = AVL_NEXT(t, first); 689 ASSERT(first != NULL); 690 } 691 692 /* 693 * Walk forward through sufficiently contiguous I/Os. 694 */ 695 while ((dio = AVL_NEXT(t, last)) != NULL && 696 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 697 IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit && 698 IO_GAP(last, dio) <= maxgap) { 699 last = dio; 700 if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) 701 mandatory = last; 702 } 703 704 /* 705 * Now that we've established the range of the I/O aggregation 706 * we must decide what to do with trailing optional I/Os. 707 * For reads, there's nothing to do. While we are unable to 708 * aggregate further, it's possible that a trailing optional 709 * I/O would allow the underlying device to aggregate with 710 * subsequent I/Os. We must therefore determine if the next 711 * non-optional I/O is close enough to make aggregation 712 * worthwhile. 713 */ 714 stretch = B_FALSE; 715 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { 716 zio_t *nio = last; 717 while ((dio = AVL_NEXT(t, nio)) != NULL && 718 IO_GAP(nio, dio) == 0 && 719 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { 720 nio = dio; 721 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { 722 stretch = B_TRUE; 723 break; 724 } 725 } 726 } 727 728 if (stretch) { 729 /* This may be a no-op. */ 730 dio = AVL_NEXT(t, last); 731 dio->io_flags &= ~ZIO_FLAG_OPTIONAL; 732 } else { 733 while (last != mandatory && last != first) { 734 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); 735 last = AVL_PREV(t, last); 736 ASSERT(last != NULL); 737 } 738 } 739 740 if (first == last) 741 return (NULL); 742 743 size = IO_SPAN(first, last); 744 ASSERT3U(size, <=, zfs_vdev_aggregation_limit); 745 746 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, 747 zio_buf_alloc(size), size, first->io_type, zio->io_priority, 748 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE, 749 vdev_queue_agg_io_done, NULL); 750 aio->io_timestamp = first->io_timestamp; 751 752 nio = first; 753 do { 754 dio = nio; 755 nio = AVL_NEXT(t, dio); 756 ASSERT3U(dio->io_type, ==, aio->io_type); 757 758 if (dio->io_flags & ZIO_FLAG_NODATA) { 759 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); 760 bzero((char *)aio->io_data + (dio->io_offset - 761 aio->io_offset), dio->io_size); 762 } else if (dio->io_type == ZIO_TYPE_WRITE) { 763 bcopy(dio->io_data, (char *)aio->io_data + 764 (dio->io_offset - aio->io_offset), 765 dio->io_size); 766 } 767 768 zio_add_child(dio, aio); 769 vdev_queue_io_remove(vq, dio); 770 zio_vdev_io_bypass(dio); 771 zio_execute(dio); 772 } while (dio != last); 773 774 return (aio); 775} 776 777static zio_t * 778vdev_queue_io_to_issue(vdev_queue_t *vq) 779{ 780 zio_t *zio, *aio; 781 zio_priority_t p; 782 avl_index_t idx; 783 avl_tree_t *tree; 784 zio_t search; 785 786again: 787 ASSERT(MUTEX_HELD(&vq->vq_lock)); 788 789 p = vdev_queue_class_to_issue(vq); 790 791 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { 792 /* No eligible queued i/os */ 793 return (NULL); 794 } 795 796 /* 797 * For LBA-ordered queues (async / scrub), issue the i/o which follows 798 * the most recently issued i/o in LBA (offset) order. 799 * 800 * For FIFO queues (sync), issue the i/o with the lowest timestamp. 801 */ 802 tree = vdev_queue_class_tree(vq, p); 803 search.io_timestamp = 0; 804 search.io_offset = vq->vq_last_offset + 1; 805 VERIFY3P(avl_find(tree, &search, &idx), ==, NULL); 806 zio = avl_nearest(tree, idx, AVL_AFTER); 807 if (zio == NULL) 808 zio = avl_first(tree); 809 ASSERT3U(zio->io_priority, ==, p); 810 811 aio = vdev_queue_aggregate(vq, zio); 812 if (aio != NULL) 813 zio = aio; 814 else 815 vdev_queue_io_remove(vq, zio); 816 817 /* 818 * If the I/O is or was optional and therefore has no data, we need to 819 * simply discard it. We need to drop the vdev queue's lock to avoid a 820 * deadlock that we could encounter since this I/O will complete 821 * immediately. 822 */ 823 if (zio->io_flags & ZIO_FLAG_NODATA) { 824 mutex_exit(&vq->vq_lock); 825 zio_vdev_io_bypass(zio); 826 zio_execute(zio); 827 mutex_enter(&vq->vq_lock); 828 goto again; 829 } 830 831 vdev_queue_pending_add(vq, zio); 832 vq->vq_last_offset = zio->io_offset; 833 834 return (zio); 835} 836 837zio_t * 838vdev_queue_io(zio_t *zio) 839{ 840 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 841 zio_t *nio; 842 843 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) 844 return (zio); 845 846 /* 847 * Children i/os inherent their parent's priority, which might 848 * not match the child's i/o type. Fix it up here. 849 */ 850 if (zio->io_type == ZIO_TYPE_READ) { 851 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && 852 zio->io_priority != ZIO_PRIORITY_ASYNC_READ && 853 zio->io_priority != ZIO_PRIORITY_SCRUB) 854 zio->io_priority = ZIO_PRIORITY_ASYNC_READ; 855 } else if (zio->io_type == ZIO_TYPE_WRITE) { 856 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && 857 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE) 858 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; 859 } else { 860 ASSERT(zio->io_type == ZIO_TYPE_FREE); 861 zio->io_priority = ZIO_PRIORITY_TRIM; 862 } 863 864 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; 865 866 mutex_enter(&vq->vq_lock); 867 zio->io_timestamp = gethrtime(); 868 vdev_queue_io_add(vq, zio); 869 nio = vdev_queue_io_to_issue(vq); 870 mutex_exit(&vq->vq_lock); 871 872 if (nio == NULL) 873 return (NULL); 874 875 if (nio->io_done == vdev_queue_agg_io_done) { 876 zio_nowait(nio); 877 return (NULL); 878 } 879 880 return (nio); 881} 882 883void 884vdev_queue_io_done(zio_t *zio) 885{ 886 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 887 zio_t *nio; 888 889 mutex_enter(&vq->vq_lock); 890 891 vdev_queue_pending_remove(vq, zio); 892 893 vq->vq_io_complete_ts = gethrtime(); 894 895 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { 896 mutex_exit(&vq->vq_lock); 897 if (nio->io_done == vdev_queue_agg_io_done) { 898 zio_nowait(nio); 899 } else { 900 zio_vdev_io_reissue(nio); 901 zio_execute(nio); 902 } 903 mutex_enter(&vq->vq_lock); 904 } 905 906 mutex_exit(&vq->vq_lock); 907} 908 909/* 910 * As these three methods are only used for load calculations we're not concerned 911 * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex 912 * use here, instead we prefer to keep it lock free for performance. 913 */ 914int 915vdev_queue_length(vdev_t *vd) 916{ 917 return (avl_numnodes(&vd->vdev_queue.vq_active_tree)); 918} 919 920uint64_t 921vdev_queue_lastoffset(vdev_t *vd) 922{ 923 return (vd->vdev_queue.vq_lastoffset); 924} 925 926void 927vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio) 928{ 929 vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size; 930} 931