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#include <sys/metaslab_impl.h> 38 39/* 40 * ZFS I/O Scheduler 41 * --------------- 42 * 43 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The 44 * I/O scheduler determines when and in what order those operations are 45 * issued. The I/O scheduler divides operations into six I/O classes 46 * prioritized in the following order: sync read, sync write, async read, 47 * async write, scrub/resilver and trim. Each queue defines the minimum and 48 * maximum number of concurrent operations that may be issued to the device. 49 * In addition, the device has an aggregate maximum. Note that the sum of the 50 * per-queue minimums must not exceed the aggregate maximum, and if the 51 * aggregate maximum is equal to or greater than the sum of the per-queue 52 * maximums, the per-queue minimum has no effect. 53 * 54 * For many physical devices, throughput increases with the number of 55 * concurrent operations, but latency typically suffers. Further, physical 56 * devices typically have a limit at which more concurrent operations have no 57 * effect on throughput or can actually cause it to decrease. 58 * 59 * The scheduler selects the next operation to issue by first looking for an 60 * I/O class whose minimum has not been satisfied. Once all are satisfied and 61 * the aggregate maximum has not been hit, the scheduler looks for classes 62 * whose maximum has not been satisfied. Iteration through the I/O classes is 63 * done in the order specified above. No further operations are issued if the 64 * aggregate maximum number of concurrent operations has been hit or if there 65 * are no operations queued for an I/O class that has not hit its maximum. 66 * Every time an I/O is queued or an operation completes, the I/O scheduler 67 * looks for new operations to issue. 68 * 69 * All I/O classes have a fixed maximum number of outstanding operations 70 * except for the async write class. Asynchronous writes represent the data 71 * that is committed to stable storage during the syncing stage for 72 * transaction groups (see txg.c). Transaction groups enter the syncing state 73 * periodically so the number of queued async writes will quickly burst up and 74 * then bleed down to zero. Rather than servicing them as quickly as possible, 75 * the I/O scheduler changes the maximum number of active async write I/Os 76 * according to the amount of dirty data in the pool (see dsl_pool.c). Since 77 * both throughput and latency typically increase with the number of 78 * concurrent operations issued to physical devices, reducing the burstiness 79 * in the number of concurrent operations also stabilizes the response time of 80 * operations from other -- and in particular synchronous -- queues. In broad 81 * strokes, the I/O scheduler will issue more concurrent operations from the 82 * async write queue as there's more dirty data in the pool. 83 * 84 * Async Writes 85 * 86 * The number of concurrent operations issued for the async write I/O class 87 * follows a piece-wise linear function defined by a few adjustable points. 88 * 89 * | o---------| <-- zfs_vdev_async_write_max_active 90 * ^ | /^ | 91 * | | / | | 92 * active | / | | 93 * I/O | / | | 94 * count | / | | 95 * | / | | 96 * |------------o | | <-- zfs_vdev_async_write_min_active 97 * 0|____________^______|_________| 98 * 0% | | 100% of zfs_dirty_data_max 99 * | | 100 * | `-- zfs_vdev_async_write_active_max_dirty_percent 101 * `--------- zfs_vdev_async_write_active_min_dirty_percent 102 * 103 * Until the amount of dirty data exceeds a minimum percentage of the dirty 104 * data allowed in the pool, the I/O scheduler will limit the number of 105 * concurrent operations to the minimum. As that threshold is crossed, the 106 * number of concurrent operations issued increases linearly to the maximum at 107 * the specified maximum percentage of the dirty data allowed in the pool. 108 * 109 * Ideally, the amount of dirty data on a busy pool will stay in the sloped 110 * part of the function between zfs_vdev_async_write_active_min_dirty_percent 111 * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the 112 * maximum percentage, this indicates that the rate of incoming data is 113 * greater than the rate that the backend storage can handle. In this case, we 114 * must further throttle incoming writes (see dmu_tx_delay() for details). 115 */ 116 117/* 118 * The maximum number of I/Os active to each device. Ideally, this will be >= 119 * the sum of each queue's max_active. It must be at least the sum of each 120 * queue's min_active. 121 */ 122uint32_t zfs_vdev_max_active = 1000; 123 124/* 125 * Per-queue limits on the number of I/Os active to each device. If the 126 * sum of the queue's max_active is < zfs_vdev_max_active, then the 127 * min_active comes into play. We will send min_active from each queue, 128 * and then select from queues in the order defined by zio_priority_t. 129 * 130 * In general, smaller max_active's will lead to lower latency of synchronous 131 * operations. Larger max_active's may lead to higher overall throughput, 132 * depending on underlying storage. 133 * 134 * The ratio of the queues' max_actives determines the balance of performance 135 * between reads, writes, and scrubs. E.g., increasing 136 * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete 137 * more quickly, but reads and writes to have higher latency and lower 138 * throughput. 139 */ 140uint32_t zfs_vdev_sync_read_min_active = 10; 141uint32_t zfs_vdev_sync_read_max_active = 10; 142uint32_t zfs_vdev_sync_write_min_active = 10; 143uint32_t zfs_vdev_sync_write_max_active = 10; 144uint32_t zfs_vdev_async_read_min_active = 1; 145uint32_t zfs_vdev_async_read_max_active = 3; 146uint32_t zfs_vdev_async_write_min_active = 1; 147uint32_t zfs_vdev_async_write_max_active = 10; 148uint32_t zfs_vdev_scrub_min_active = 1; 149uint32_t zfs_vdev_scrub_max_active = 2; 150uint32_t zfs_vdev_trim_min_active = 1; 151/* 152 * TRIM max active is large in comparison to the other values due to the fact 153 * that TRIM IOs are coalesced at the device layer. This value is set such 154 * that a typical SSD can process the queued IOs in a single request. 155 */ 156uint32_t zfs_vdev_trim_max_active = 64; 157 158 159/* 160 * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent 161 * dirty data, use zfs_vdev_async_write_min_active. When it has more than 162 * zfs_vdev_async_write_active_max_dirty_percent, use 163 * zfs_vdev_async_write_max_active. The value is linearly interpolated 164 * between min and max. 165 */ 166int zfs_vdev_async_write_active_min_dirty_percent = 30; 167int zfs_vdev_async_write_active_max_dirty_percent = 60; 168 169/* 170 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O. 171 * For read I/Os, we also aggregate across small adjacency gaps; for writes 172 * we include spans of optional I/Os to aid aggregation at the disk even when 173 * they aren't able to help us aggregate at this level. 174 */ 175int zfs_vdev_aggregation_limit = SPA_OLD_MAXBLOCKSIZE; 176int zfs_vdev_read_gap_limit = 32 << 10; 177int zfs_vdev_write_gap_limit = 4 << 10; 178 179/* 180 * Define the queue depth percentage for each top-level. This percentage is 181 * used in conjunction with zfs_vdev_async_max_active to determine how many 182 * allocations a specific top-level vdev should handle. Once the queue depth 183 * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100 184 * then allocator will stop allocating blocks on that top-level device. 185 * The default kernel setting is 1000% which will yield 100 allocations per 186 * device. For userland testing, the default setting is 300% which equates 187 * to 30 allocations per device. 188 */ 189#ifdef _KERNEL 190int zfs_vdev_queue_depth_pct = 1000; 191#else 192int zfs_vdev_queue_depth_pct = 300; 193#endif 194 195 196#ifdef __FreeBSD__ 197#ifdef _KERNEL 198SYSCTL_DECL(_vfs_zfs_vdev); 199 200TUNABLE_INT("vfs.zfs.vdev.async_write_active_min_dirty_percent", 201 &zfs_vdev_async_write_active_min_dirty_percent); 202static int sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS); 203SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_min_dirty_percent, 204 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int), 205 sysctl_zfs_async_write_active_min_dirty_percent, "I", 206 "Percentage of async write dirty data below which " 207 "async_write_min_active is used."); 208 209TUNABLE_INT("vfs.zfs.vdev.async_write_active_max_dirty_percent", 210 &zfs_vdev_async_write_active_max_dirty_percent); 211static int sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS); 212SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_max_dirty_percent, 213 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int), 214 sysctl_zfs_async_write_active_max_dirty_percent, "I", 215 "Percentage of async write dirty data above which " 216 "async_write_max_active is used."); 217 218TUNABLE_INT("vfs.zfs.vdev.max_active", &zfs_vdev_max_active); 219SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RWTUN, 220 &zfs_vdev_max_active, 0, 221 "The maximum number of I/Os of all types active for each device."); 222 223#define ZFS_VDEV_QUEUE_KNOB_MIN(name) \ 224TUNABLE_INT("vfs.zfs.vdev." #name "_min_active", \ 225 &zfs_vdev_ ## name ## _min_active); \ 226SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, \ 227 CTLFLAG_RWTUN, &zfs_vdev_ ## name ## _min_active, 0, \ 228 "Initial number of I/O requests of type " #name \ 229 " active for each device"); 230 231#define ZFS_VDEV_QUEUE_KNOB_MAX(name) \ 232TUNABLE_INT("vfs.zfs.vdev." #name "_max_active", \ 233 &zfs_vdev_ ## name ## _max_active); \ 234SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, \ 235 CTLFLAG_RWTUN, &zfs_vdev_ ## name ## _max_active, 0, \ 236 "Maximum number of I/O requests of type " #name \ 237 " active for each device"); 238 239ZFS_VDEV_QUEUE_KNOB_MIN(sync_read); 240ZFS_VDEV_QUEUE_KNOB_MAX(sync_read); 241ZFS_VDEV_QUEUE_KNOB_MIN(sync_write); 242ZFS_VDEV_QUEUE_KNOB_MAX(sync_write); 243ZFS_VDEV_QUEUE_KNOB_MIN(async_read); 244ZFS_VDEV_QUEUE_KNOB_MAX(async_read); 245ZFS_VDEV_QUEUE_KNOB_MIN(async_write); 246ZFS_VDEV_QUEUE_KNOB_MAX(async_write); 247ZFS_VDEV_QUEUE_KNOB_MIN(scrub); 248ZFS_VDEV_QUEUE_KNOB_MAX(scrub); 249ZFS_VDEV_QUEUE_KNOB_MIN(trim); 250ZFS_VDEV_QUEUE_KNOB_MAX(trim); 251 252#undef ZFS_VDEV_QUEUE_KNOB 253 254TUNABLE_INT("vfs.zfs.vdev.aggregation_limit", &zfs_vdev_aggregation_limit); 255SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RWTUN, 256 &zfs_vdev_aggregation_limit, 0, 257 "I/O requests are aggregated up to this size"); 258TUNABLE_INT("vfs.zfs.vdev.read_gap_limit", &zfs_vdev_read_gap_limit); 259SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RWTUN, 260 &zfs_vdev_read_gap_limit, 0, 261 "Acceptable gap between two reads being aggregated"); 262TUNABLE_INT("vfs.zfs.vdev.write_gap_limit", &zfs_vdev_write_gap_limit); 263SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RWTUN, 264 &zfs_vdev_write_gap_limit, 0, 265 "Acceptable gap between two writes being aggregated"); 266SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, queue_depth_pct, CTLFLAG_RWTUN, 267 &zfs_vdev_queue_depth_pct, 0, 268 "Queue depth percentage for each top-level"); 269 270static int 271sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS) 272{ 273 int val, err; 274 275 val = zfs_vdev_async_write_active_min_dirty_percent; 276 err = sysctl_handle_int(oidp, &val, 0, req); 277 if (err != 0 || req->newptr == NULL) 278 return (err); 279 280 if (val < 0 || val > 100 || 281 val >= zfs_vdev_async_write_active_max_dirty_percent) 282 return (EINVAL); 283 284 zfs_vdev_async_write_active_min_dirty_percent = val; 285 286 return (0); 287} 288 289static int 290sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS) 291{ 292 int val, err; 293 294 val = zfs_vdev_async_write_active_max_dirty_percent; 295 err = sysctl_handle_int(oidp, &val, 0, req); 296 if (err != 0 || req->newptr == NULL) 297 return (err); 298 299 if (val < 0 || val > 100 || 300 val <= zfs_vdev_async_write_active_min_dirty_percent) 301 return (EINVAL); 302 303 zfs_vdev_async_write_active_max_dirty_percent = val; 304 305 return (0); 306} 307#endif 308#endif 309 310int 311vdev_queue_offset_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_offset < z2->io_offset) 317 return (-1); 318 if (z1->io_offset > z2->io_offset) 319 return (1); 320 321 if (z1 < z2) 322 return (-1); 323 if (z1 > z2) 324 return (1); 325 326 return (0); 327} 328 329static inline avl_tree_t * 330vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p) 331{ 332 return (&vq->vq_class[p].vqc_queued_tree); 333} 334 335static inline avl_tree_t * 336vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t) 337{ 338 if (t == ZIO_TYPE_READ) 339 return (&vq->vq_read_offset_tree); 340 else if (t == ZIO_TYPE_WRITE) 341 return (&vq->vq_write_offset_tree); 342 else 343 return (NULL); 344} 345 346int 347vdev_queue_timestamp_compare(const void *x1, const void *x2) 348{ 349 const zio_t *z1 = x1; 350 const zio_t *z2 = x2; 351 352 if (z1->io_timestamp < z2->io_timestamp) 353 return (-1); 354 if (z1->io_timestamp > z2->io_timestamp) 355 return (1); 356 357 if (z1->io_offset < z2->io_offset) 358 return (-1); 359 if (z1->io_offset > z2->io_offset) 360 return (1); 361 362 if (z1 < z2) 363 return (-1); 364 if (z1 > z2) 365 return (1); 366 367 return (0); 368} 369 370void 371vdev_queue_init(vdev_t *vd) 372{ 373 vdev_queue_t *vq = &vd->vdev_queue; 374 375 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); 376 vq->vq_vdev = vd; 377 378 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare, 379 sizeof (zio_t), offsetof(struct zio, io_queue_node)); 380 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ), 381 vdev_queue_offset_compare, sizeof (zio_t), 382 offsetof(struct zio, io_offset_node)); 383 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE), 384 vdev_queue_offset_compare, sizeof (zio_t), 385 offsetof(struct zio, io_offset_node)); 386 387 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 388 int (*compfn) (const void *, const void *); 389 390 /* 391 * The synchronous i/o queues are dispatched in FIFO rather 392 * than LBA order. This provides more consistent latency for 393 * these i/os. 394 */ 395 if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE) 396 compfn = vdev_queue_timestamp_compare; 397 else 398 compfn = vdev_queue_offset_compare; 399 400 avl_create(vdev_queue_class_tree(vq, p), compfn, 401 sizeof (zio_t), offsetof(struct zio, io_queue_node)); 402 } 403 404 vq->vq_lastoffset = 0; 405} 406 407void 408vdev_queue_fini(vdev_t *vd) 409{ 410 vdev_queue_t *vq = &vd->vdev_queue; 411 412 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) 413 avl_destroy(vdev_queue_class_tree(vq, p)); 414 avl_destroy(&vq->vq_active_tree); 415 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ)); 416 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE)); 417 418 mutex_destroy(&vq->vq_lock); 419} 420 421static void 422vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) 423{ 424 spa_t *spa = zio->io_spa; 425 avl_tree_t *qtt; 426 427 ASSERT(MUTEX_HELD(&vq->vq_lock)); 428 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 429 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio); 430 qtt = vdev_queue_type_tree(vq, zio->io_type); 431 if (qtt) 432 avl_add(qtt, zio); 433 434#ifdef illumos 435 mutex_enter(&spa->spa_iokstat_lock); 436 spa->spa_queue_stats[zio->io_priority].spa_queued++; 437 if (spa->spa_iokstat != NULL) 438 kstat_waitq_enter(spa->spa_iokstat->ks_data); 439 mutex_exit(&spa->spa_iokstat_lock); 440#endif 441} 442 443static void 444vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) 445{ 446 spa_t *spa = zio->io_spa; 447 avl_tree_t *qtt; 448 449 ASSERT(MUTEX_HELD(&vq->vq_lock)); 450 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 451 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio); 452 qtt = vdev_queue_type_tree(vq, zio->io_type); 453 if (qtt) 454 avl_remove(qtt, zio); 455 456#ifdef illumos 457 mutex_enter(&spa->spa_iokstat_lock); 458 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0); 459 spa->spa_queue_stats[zio->io_priority].spa_queued--; 460 if (spa->spa_iokstat != NULL) 461 kstat_waitq_exit(spa->spa_iokstat->ks_data); 462 mutex_exit(&spa->spa_iokstat_lock); 463#endif 464} 465 466static void 467vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) 468{ 469 spa_t *spa = zio->io_spa; 470 ASSERT(MUTEX_HELD(&vq->vq_lock)); 471 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 472 vq->vq_class[zio->io_priority].vqc_active++; 473 avl_add(&vq->vq_active_tree, zio); 474 475#ifdef illumos 476 mutex_enter(&spa->spa_iokstat_lock); 477 spa->spa_queue_stats[zio->io_priority].spa_active++; 478 if (spa->spa_iokstat != NULL) 479 kstat_runq_enter(spa->spa_iokstat->ks_data); 480 mutex_exit(&spa->spa_iokstat_lock); 481#endif 482} 483 484static void 485vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) 486{ 487 spa_t *spa = zio->io_spa; 488 ASSERT(MUTEX_HELD(&vq->vq_lock)); 489 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 490 vq->vq_class[zio->io_priority].vqc_active--; 491 avl_remove(&vq->vq_active_tree, zio); 492 493#ifdef illumos 494 mutex_enter(&spa->spa_iokstat_lock); 495 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0); 496 spa->spa_queue_stats[zio->io_priority].spa_active--; 497 if (spa->spa_iokstat != NULL) { 498 kstat_io_t *ksio = spa->spa_iokstat->ks_data; 499 500 kstat_runq_exit(spa->spa_iokstat->ks_data); 501 if (zio->io_type == ZIO_TYPE_READ) { 502 ksio->reads++; 503 ksio->nread += zio->io_size; 504 } else if (zio->io_type == ZIO_TYPE_WRITE) { 505 ksio->writes++; 506 ksio->nwritten += zio->io_size; 507 } 508 } 509 mutex_exit(&spa->spa_iokstat_lock); 510#endif 511} 512 513static void 514vdev_queue_agg_io_done(zio_t *aio) 515{ 516 if (aio->io_type == ZIO_TYPE_READ) { 517 zio_t *pio; 518 zio_link_t *zl = NULL; 519 while ((pio = zio_walk_parents(aio, &zl)) != NULL) { 520 bcopy((char *)aio->io_data + (pio->io_offset - 521 aio->io_offset), pio->io_data, pio->io_size); 522 } 523 } 524 525 zio_buf_free(aio->io_data, aio->io_size); 526} 527 528static int 529vdev_queue_class_min_active(zio_priority_t p) 530{ 531 switch (p) { 532 case ZIO_PRIORITY_SYNC_READ: 533 return (zfs_vdev_sync_read_min_active); 534 case ZIO_PRIORITY_SYNC_WRITE: 535 return (zfs_vdev_sync_write_min_active); 536 case ZIO_PRIORITY_ASYNC_READ: 537 return (zfs_vdev_async_read_min_active); 538 case ZIO_PRIORITY_ASYNC_WRITE: 539 return (zfs_vdev_async_write_min_active); 540 case ZIO_PRIORITY_SCRUB: 541 return (zfs_vdev_scrub_min_active); 542 case ZIO_PRIORITY_TRIM: 543 return (zfs_vdev_trim_min_active); 544 default: 545 panic("invalid priority %u", p); 546 return (0); 547 } 548} 549 550static __noinline int 551vdev_queue_max_async_writes(spa_t *spa) 552{ 553 int writes; 554 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total; 555 uint64_t min_bytes = zfs_dirty_data_max * 556 zfs_vdev_async_write_active_min_dirty_percent / 100; 557 uint64_t max_bytes = zfs_dirty_data_max * 558 zfs_vdev_async_write_active_max_dirty_percent / 100; 559 560 /* 561 * Sync tasks correspond to interactive user actions. To reduce the 562 * execution time of those actions we push data out as fast as possible. 563 */ 564 if (spa_has_pending_synctask(spa)) { 565 return (zfs_vdev_async_write_max_active); 566 } 567 568 if (dirty < min_bytes) 569 return (zfs_vdev_async_write_min_active); 570 if (dirty > max_bytes) 571 return (zfs_vdev_async_write_max_active); 572 573 /* 574 * linear interpolation: 575 * slope = (max_writes - min_writes) / (max_bytes - min_bytes) 576 * move right by min_bytes 577 * move up by min_writes 578 */ 579 writes = (dirty - min_bytes) * 580 (zfs_vdev_async_write_max_active - 581 zfs_vdev_async_write_min_active) / 582 (max_bytes - min_bytes) + 583 zfs_vdev_async_write_min_active; 584 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); 585 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); 586 return (writes); 587} 588 589static int 590vdev_queue_class_max_active(spa_t *spa, zio_priority_t p) 591{ 592 switch (p) { 593 case ZIO_PRIORITY_SYNC_READ: 594 return (zfs_vdev_sync_read_max_active); 595 case ZIO_PRIORITY_SYNC_WRITE: 596 return (zfs_vdev_sync_write_max_active); 597 case ZIO_PRIORITY_ASYNC_READ: 598 return (zfs_vdev_async_read_max_active); 599 case ZIO_PRIORITY_ASYNC_WRITE: 600 return (vdev_queue_max_async_writes(spa)); 601 case ZIO_PRIORITY_SCRUB: 602 return (zfs_vdev_scrub_max_active); 603 case ZIO_PRIORITY_TRIM: 604 return (zfs_vdev_trim_max_active); 605 default: 606 panic("invalid priority %u", p); 607 return (0); 608 } 609} 610 611/* 612 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if 613 * there is no eligible class. 614 */ 615static zio_priority_t 616vdev_queue_class_to_issue(vdev_queue_t *vq) 617{ 618 spa_t *spa = vq->vq_vdev->vdev_spa; 619 zio_priority_t p; 620 621 ASSERT(MUTEX_HELD(&vq->vq_lock)); 622 623 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active) 624 return (ZIO_PRIORITY_NUM_QUEUEABLE); 625 626 /* find a queue that has not reached its minimum # outstanding i/os */ 627 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 628 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && 629 vq->vq_class[p].vqc_active < 630 vdev_queue_class_min_active(p)) 631 return (p); 632 } 633 634 /* 635 * If we haven't found a queue, look for one that hasn't reached its 636 * maximum # outstanding i/os. 637 */ 638 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 639 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && 640 vq->vq_class[p].vqc_active < 641 vdev_queue_class_max_active(spa, p)) 642 return (p); 643 } 644 645 /* No eligible queued i/os */ 646 return (ZIO_PRIORITY_NUM_QUEUEABLE); 647} 648 649/* 650 * Compute the range spanned by two i/os, which is the endpoint of the last 651 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). 652 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); 653 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. 654 */ 655#define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) 656#define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) 657 658static zio_t * 659vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) 660{ 661 zio_t *first, *last, *aio, *dio, *mandatory, *nio; 662 void *abuf; 663 uint64_t maxgap = 0; 664 uint64_t size; 665 boolean_t stretch; 666 avl_tree_t *t; 667 enum zio_flag flags; 668 669 ASSERT(MUTEX_HELD(&vq->vq_lock)); 670 671 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE) 672 return (NULL); 673 674 first = last = zio; 675 676 if (zio->io_type == ZIO_TYPE_READ) 677 maxgap = zfs_vdev_read_gap_limit; 678 679 /* 680 * We can aggregate I/Os that are sufficiently adjacent and of 681 * the same flavor, as expressed by the AGG_INHERIT flags. 682 * The latter requirement is necessary so that certain 683 * attributes of the I/O, such as whether it's a normal I/O 684 * or a scrub/resilver, can be preserved in the aggregate. 685 * We can include optional I/Os, but don't allow them 686 * to begin a range as they add no benefit in that situation. 687 */ 688 689 /* 690 * We keep track of the last non-optional I/O. 691 */ 692 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; 693 694 /* 695 * Walk backwards through sufficiently contiguous I/Os 696 * recording the last non-option I/O. 697 */ 698 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; 699 t = vdev_queue_type_tree(vq, zio->io_type); 700 while (t != NULL && (dio = AVL_PREV(t, first)) != NULL && 701 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 702 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit && 703 IO_GAP(dio, first) <= maxgap) { 704 first = dio; 705 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) 706 mandatory = first; 707 } 708 709 /* 710 * Skip any initial optional I/Os. 711 */ 712 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { 713 first = AVL_NEXT(t, first); 714 ASSERT(first != NULL); 715 } 716 717 /* 718 * Walk forward through sufficiently contiguous I/Os. 719 */ 720 while ((dio = AVL_NEXT(t, last)) != NULL && 721 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 722 IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit && 723 IO_GAP(last, dio) <= maxgap) { 724 last = dio; 725 if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) 726 mandatory = last; 727 } 728 729 /* 730 * Now that we've established the range of the I/O aggregation 731 * we must decide what to do with trailing optional I/Os. 732 * For reads, there's nothing to do. While we are unable to 733 * aggregate further, it's possible that a trailing optional 734 * I/O would allow the underlying device to aggregate with 735 * subsequent I/Os. We must therefore determine if the next 736 * non-optional I/O is close enough to make aggregation 737 * worthwhile. 738 */ 739 stretch = B_FALSE; 740 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { 741 zio_t *nio = last; 742 while ((dio = AVL_NEXT(t, nio)) != NULL && 743 IO_GAP(nio, dio) == 0 && 744 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { 745 nio = dio; 746 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { 747 stretch = B_TRUE; 748 break; 749 } 750 } 751 } 752 753 if (stretch) { 754 /* This may be a no-op. */ 755 dio = AVL_NEXT(t, last); 756 dio->io_flags &= ~ZIO_FLAG_OPTIONAL; 757 } else { 758 while (last != mandatory && last != first) { 759 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); 760 last = AVL_PREV(t, last); 761 ASSERT(last != NULL); 762 } 763 } 764 765 if (first == last) 766 return (NULL); 767 768 size = IO_SPAN(first, last); 769 ASSERT3U(size, <=, zfs_vdev_aggregation_limit); 770 771 abuf = zio_buf_alloc_nowait(size); 772 if (abuf == NULL) 773 return (NULL); 774 775 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, 776 abuf, size, first->io_type, zio->io_priority, 777 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE, 778 vdev_queue_agg_io_done, NULL); 779 aio->io_timestamp = first->io_timestamp; 780 781 nio = first; 782 do { 783 dio = nio; 784 nio = AVL_NEXT(t, dio); 785 ASSERT3U(dio->io_type, ==, aio->io_type); 786 787 if (dio->io_flags & ZIO_FLAG_NODATA) { 788 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); 789 bzero((char *)aio->io_data + (dio->io_offset - 790 aio->io_offset), dio->io_size); 791 } else if (dio->io_type == ZIO_TYPE_WRITE) { 792 bcopy(dio->io_data, (char *)aio->io_data + 793 (dio->io_offset - aio->io_offset), 794 dio->io_size); 795 } 796 797 zio_add_child(dio, aio); 798 vdev_queue_io_remove(vq, dio); 799 zio_vdev_io_bypass(dio); 800 zio_execute(dio); 801 } while (dio != last); 802 803 return (aio); 804} 805 806static zio_t * 807vdev_queue_io_to_issue(vdev_queue_t *vq) 808{ 809 zio_t *zio, *aio; 810 zio_priority_t p; 811 avl_index_t idx; 812 avl_tree_t *tree; 813 zio_t search; 814 815again: 816 ASSERT(MUTEX_HELD(&vq->vq_lock)); 817 818 p = vdev_queue_class_to_issue(vq); 819 820 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { 821 /* No eligible queued i/os */ 822 return (NULL); 823 } 824 825 /* 826 * For LBA-ordered queues (async / scrub), issue the i/o which follows 827 * the most recently issued i/o in LBA (offset) order. 828 * 829 * For FIFO queues (sync), issue the i/o with the lowest timestamp. 830 */ 831 tree = vdev_queue_class_tree(vq, p); 832 search.io_timestamp = 0; 833 search.io_offset = vq->vq_last_offset + 1; 834 VERIFY3P(avl_find(tree, &search, &idx), ==, NULL); 835 zio = avl_nearest(tree, idx, AVL_AFTER); 836 if (zio == NULL) 837 zio = avl_first(tree); 838 ASSERT3U(zio->io_priority, ==, p); 839 840 aio = vdev_queue_aggregate(vq, zio); 841 if (aio != NULL) 842 zio = aio; 843 else 844 vdev_queue_io_remove(vq, zio); 845 846 /* 847 * If the I/O is or was optional and therefore has no data, we need to 848 * simply discard it. We need to drop the vdev queue's lock to avoid a 849 * deadlock that we could encounter since this I/O will complete 850 * immediately. 851 */ 852 if (zio->io_flags & ZIO_FLAG_NODATA) { 853 mutex_exit(&vq->vq_lock); 854 zio_vdev_io_bypass(zio); 855 zio_execute(zio); 856 mutex_enter(&vq->vq_lock); 857 goto again; 858 } 859 860 vdev_queue_pending_add(vq, zio); 861 vq->vq_last_offset = zio->io_offset; 862 863 return (zio); 864} 865 866zio_t * 867vdev_queue_io(zio_t *zio) 868{ 869 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 870 zio_t *nio; 871 872 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) 873 return (zio); 874 875 /* 876 * Children i/os inherent their parent's priority, which might 877 * not match the child's i/o type. Fix it up here. 878 */ 879 if (zio->io_type == ZIO_TYPE_READ) { 880 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && 881 zio->io_priority != ZIO_PRIORITY_ASYNC_READ && 882 zio->io_priority != ZIO_PRIORITY_SCRUB) 883 zio->io_priority = ZIO_PRIORITY_ASYNC_READ; 884 } else if (zio->io_type == ZIO_TYPE_WRITE) { 885 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && 886 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE) 887 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; 888 } else { 889 ASSERT(zio->io_type == ZIO_TYPE_FREE); 890 zio->io_priority = ZIO_PRIORITY_TRIM; 891 } 892 893 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; 894 895 mutex_enter(&vq->vq_lock); 896 zio->io_timestamp = gethrtime(); 897 vdev_queue_io_add(vq, zio); 898 nio = vdev_queue_io_to_issue(vq); 899 mutex_exit(&vq->vq_lock); 900 901 if (nio == NULL) 902 return (NULL); 903 904 if (nio->io_done == vdev_queue_agg_io_done) { 905 zio_nowait(nio); 906 return (NULL); 907 } 908 909 return (nio); 910} 911 912void 913vdev_queue_io_done(zio_t *zio) 914{ 915 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 916 zio_t *nio; 917 918 mutex_enter(&vq->vq_lock); 919 920 vdev_queue_pending_remove(vq, zio); 921 922 vq->vq_io_complete_ts = gethrtime(); 923 924 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { 925 mutex_exit(&vq->vq_lock); 926 if (nio->io_done == vdev_queue_agg_io_done) { 927 zio_nowait(nio); 928 } else { 929 zio_vdev_io_reissue(nio); 930 zio_execute(nio); 931 } 932 mutex_enter(&vq->vq_lock); 933 } 934 935 mutex_exit(&vq->vq_lock); 936} 937 938/* 939 * As these three methods are only used for load calculations we're not concerned 940 * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex 941 * use here, instead we prefer to keep it lock free for performance. 942 */ 943int 944vdev_queue_length(vdev_t *vd) 945{ 946 return (avl_numnodes(&vd->vdev_queue.vq_active_tree)); 947} 948 949uint64_t 950vdev_queue_lastoffset(vdev_t *vd) 951{ 952 return (vd->vdev_queue.vq_lastoffset); 953} 954 955void 956vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio) 957{ 958 vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size; 959} 960