vdev_queue.c revision 260763
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) 2013 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 five I/O classes 44 * prioritized in the following order: sync read, sync write, async read, 45 * async write, and scrub/resilver. 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; 148 149/* 150 * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent 151 * dirty data, use zfs_vdev_async_write_min_active. When it has more than 152 * zfs_vdev_async_write_active_max_dirty_percent, use 153 * zfs_vdev_async_write_max_active. The value is linearly interpolated 154 * between min and max. 155 */ 156int zfs_vdev_async_write_active_min_dirty_percent = 30; 157int zfs_vdev_async_write_active_max_dirty_percent = 60; 158 159/* 160 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O. 161 * For read I/Os, we also aggregate across small adjacency gaps; for writes 162 * we include spans of optional I/Os to aid aggregation at the disk even when 163 * they aren't able to help us aggregate at this level. 164 */ 165int zfs_vdev_aggregation_limit = SPA_MAXBLOCKSIZE; 166int zfs_vdev_read_gap_limit = 32 << 10; 167int zfs_vdev_write_gap_limit = 4 << 10; 168 169#ifdef __FreeBSD__ 170SYSCTL_DECL(_vfs_zfs_vdev); 171TUNABLE_INT("vfs.zfs.vdev.max_active", &zfs_vdev_max_active); 172SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RW, 173 &zfs_vdev_max_active, 0, 174 "The maximum number of i/os of all types active for each device."); 175 176#define ZFS_VDEV_QUEUE_KNOB_MIN(name) \ 177TUNABLE_INT("vfs.zfs.vdev." #name "_min_active", \ 178 &zfs_vdev_ ## name ## _min_active); \ 179SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, CTLFLAG_RW, \ 180 &zfs_vdev_ ## name ## _min_active, 0, \ 181 "Initial number of I/O requests of type " #name \ 182 " active for each device"); 183 184#define ZFS_VDEV_QUEUE_KNOB_MAX(name) \ 185TUNABLE_INT("vfs.zfs.vdev." #name "_max_active", \ 186 &zfs_vdev_ ## name ## _max_active); \ 187SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, CTLFLAG_RW, \ 188 &zfs_vdev_ ## name ## _max_active, 0, \ 189 "Maximum number of I/O requests of type " #name \ 190 " active for each device"); 191 192ZFS_VDEV_QUEUE_KNOB_MIN(sync_read); 193ZFS_VDEV_QUEUE_KNOB_MAX(sync_read); 194ZFS_VDEV_QUEUE_KNOB_MIN(sync_write); 195ZFS_VDEV_QUEUE_KNOB_MAX(sync_write); 196ZFS_VDEV_QUEUE_KNOB_MIN(async_read); 197ZFS_VDEV_QUEUE_KNOB_MAX(async_read); 198ZFS_VDEV_QUEUE_KNOB_MIN(async_write); 199ZFS_VDEV_QUEUE_KNOB_MAX(async_write); 200ZFS_VDEV_QUEUE_KNOB_MIN(scrub); 201ZFS_VDEV_QUEUE_KNOB_MAX(scrub); 202 203#undef ZFS_VDEV_QUEUE_KNOB 204 205TUNABLE_INT("vfs.zfs.vdev.aggregation_limit", &zfs_vdev_aggregation_limit); 206SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RW, 207 &zfs_vdev_aggregation_limit, 0, 208 "I/O requests are aggregated up to this size"); 209TUNABLE_INT("vfs.zfs.vdev.read_gap_limit", &zfs_vdev_read_gap_limit); 210SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RW, 211 &zfs_vdev_read_gap_limit, 0, 212 "Acceptable gap between two reads being aggregated"); 213TUNABLE_INT("vfs.zfs.vdev.write_gap_limit", &zfs_vdev_write_gap_limit); 214SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RW, 215 &zfs_vdev_write_gap_limit, 0, 216 "Acceptable gap between two writes being aggregated"); 217#endif 218 219int 220vdev_queue_offset_compare(const void *x1, const void *x2) 221{ 222 const zio_t *z1 = x1; 223 const zio_t *z2 = x2; 224 225 if (z1->io_offset < z2->io_offset) 226 return (-1); 227 if (z1->io_offset > z2->io_offset) 228 return (1); 229 230 if (z1 < z2) 231 return (-1); 232 if (z1 > z2) 233 return (1); 234 235 return (0); 236} 237 238int 239vdev_queue_timestamp_compare(const void *x1, const void *x2) 240{ 241 const zio_t *z1 = x1; 242 const zio_t *z2 = x2; 243 244 if (z1->io_timestamp < z2->io_timestamp) 245 return (-1); 246 if (z1->io_timestamp > z2->io_timestamp) 247 return (1); 248 249 if (z1 < z2) 250 return (-1); 251 if (z1 > z2) 252 return (1); 253 254 return (0); 255} 256 257void 258vdev_queue_init(vdev_t *vd) 259{ 260 vdev_queue_t *vq = &vd->vdev_queue; 261 262 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); 263 vq->vq_vdev = vd; 264 265 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare, 266 sizeof (zio_t), offsetof(struct zio, io_queue_node)); 267 268 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 269 /* 270 * The synchronous i/o queues are FIFO rather than LBA ordered. 271 * This provides more consistent latency for these i/os, and 272 * they tend to not be tightly clustered anyway so there is 273 * little to no throughput loss. 274 */ 275 boolean_t fifo = (p == ZIO_PRIORITY_SYNC_READ || 276 p == ZIO_PRIORITY_SYNC_WRITE); 277 avl_create(&vq->vq_class[p].vqc_queued_tree, 278 fifo ? vdev_queue_timestamp_compare : 279 vdev_queue_offset_compare, 280 sizeof (zio_t), offsetof(struct zio, io_queue_node)); 281 } 282} 283 284void 285vdev_queue_fini(vdev_t *vd) 286{ 287 vdev_queue_t *vq = &vd->vdev_queue; 288 289 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) 290 avl_destroy(&vq->vq_class[p].vqc_queued_tree); 291 avl_destroy(&vq->vq_active_tree); 292 293 mutex_destroy(&vq->vq_lock); 294} 295 296static void 297vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) 298{ 299 spa_t *spa = zio->io_spa; 300 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 301 avl_add(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio); 302 303#ifdef illumos 304 mutex_enter(&spa->spa_iokstat_lock); 305 spa->spa_queue_stats[zio->io_priority].spa_queued++; 306 if (spa->spa_iokstat != NULL) 307 kstat_waitq_enter(spa->spa_iokstat->ks_data); 308 mutex_exit(&spa->spa_iokstat_lock); 309#endif 310} 311 312static void 313vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) 314{ 315 spa_t *spa = zio->io_spa; 316 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 317 avl_remove(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio); 318 319#ifdef illumos 320 mutex_enter(&spa->spa_iokstat_lock); 321 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0); 322 spa->spa_queue_stats[zio->io_priority].spa_queued--; 323 if (spa->spa_iokstat != NULL) 324 kstat_waitq_exit(spa->spa_iokstat->ks_data); 325 mutex_exit(&spa->spa_iokstat_lock); 326#endif 327} 328 329static void 330vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) 331{ 332 spa_t *spa = zio->io_spa; 333 ASSERT(MUTEX_HELD(&vq->vq_lock)); 334 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 335 vq->vq_class[zio->io_priority].vqc_active++; 336 avl_add(&vq->vq_active_tree, zio); 337 338#ifdef illumos 339 mutex_enter(&spa->spa_iokstat_lock); 340 spa->spa_queue_stats[zio->io_priority].spa_active++; 341 if (spa->spa_iokstat != NULL) 342 kstat_runq_enter(spa->spa_iokstat->ks_data); 343 mutex_exit(&spa->spa_iokstat_lock); 344#endif 345} 346 347static void 348vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) 349{ 350 spa_t *spa = zio->io_spa; 351 ASSERT(MUTEX_HELD(&vq->vq_lock)); 352 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 353 vq->vq_class[zio->io_priority].vqc_active--; 354 avl_remove(&vq->vq_active_tree, zio); 355 356#ifdef illumos 357 mutex_enter(&spa->spa_iokstat_lock); 358 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0); 359 spa->spa_queue_stats[zio->io_priority].spa_active--; 360 if (spa->spa_iokstat != NULL) { 361 kstat_io_t *ksio = spa->spa_iokstat->ks_data; 362 363 kstat_runq_exit(spa->spa_iokstat->ks_data); 364 if (zio->io_type == ZIO_TYPE_READ) { 365 ksio->reads++; 366 ksio->nread += zio->io_size; 367 } else if (zio->io_type == ZIO_TYPE_WRITE) { 368 ksio->writes++; 369 ksio->nwritten += zio->io_size; 370 } 371 } 372 mutex_exit(&spa->spa_iokstat_lock); 373#endif 374} 375 376static void 377vdev_queue_agg_io_done(zio_t *aio) 378{ 379 if (aio->io_type == ZIO_TYPE_READ) { 380 zio_t *pio; 381 while ((pio = zio_walk_parents(aio)) != NULL) { 382 bcopy((char *)aio->io_data + (pio->io_offset - 383 aio->io_offset), pio->io_data, pio->io_size); 384 } 385 } 386 387 zio_buf_free(aio->io_data, aio->io_size); 388} 389 390static int 391vdev_queue_class_min_active(zio_priority_t p) 392{ 393 switch (p) { 394 case ZIO_PRIORITY_SYNC_READ: 395 return (zfs_vdev_sync_read_min_active); 396 case ZIO_PRIORITY_SYNC_WRITE: 397 return (zfs_vdev_sync_write_min_active); 398 case ZIO_PRIORITY_ASYNC_READ: 399 return (zfs_vdev_async_read_min_active); 400 case ZIO_PRIORITY_ASYNC_WRITE: 401 return (zfs_vdev_async_write_min_active); 402 case ZIO_PRIORITY_SCRUB: 403 return (zfs_vdev_scrub_min_active); 404 default: 405 panic("invalid priority %u", p); 406 return (0); 407 } 408} 409 410static int 411vdev_queue_max_async_writes(uint64_t dirty) 412{ 413 int writes; 414 uint64_t min_bytes = zfs_dirty_data_max * 415 zfs_vdev_async_write_active_min_dirty_percent / 100; 416 uint64_t max_bytes = zfs_dirty_data_max * 417 zfs_vdev_async_write_active_max_dirty_percent / 100; 418 419 if (dirty < min_bytes) 420 return (zfs_vdev_async_write_min_active); 421 if (dirty > max_bytes) 422 return (zfs_vdev_async_write_max_active); 423 424 /* 425 * linear interpolation: 426 * slope = (max_writes - min_writes) / (max_bytes - min_bytes) 427 * move right by min_bytes 428 * move up by min_writes 429 */ 430 writes = (dirty - min_bytes) * 431 (zfs_vdev_async_write_max_active - 432 zfs_vdev_async_write_min_active) / 433 (max_bytes - min_bytes) + 434 zfs_vdev_async_write_min_active; 435 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); 436 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); 437 return (writes); 438} 439 440static int 441vdev_queue_class_max_active(spa_t *spa, zio_priority_t p) 442{ 443 switch (p) { 444 case ZIO_PRIORITY_SYNC_READ: 445 return (zfs_vdev_sync_read_max_active); 446 case ZIO_PRIORITY_SYNC_WRITE: 447 return (zfs_vdev_sync_write_max_active); 448 case ZIO_PRIORITY_ASYNC_READ: 449 return (zfs_vdev_async_read_max_active); 450 case ZIO_PRIORITY_ASYNC_WRITE: 451 return (vdev_queue_max_async_writes( 452 spa->spa_dsl_pool->dp_dirty_total)); 453 case ZIO_PRIORITY_SCRUB: 454 return (zfs_vdev_scrub_max_active); 455 default: 456 panic("invalid priority %u", p); 457 return (0); 458 } 459} 460 461/* 462 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if 463 * there is no eligible class. 464 */ 465static zio_priority_t 466vdev_queue_class_to_issue(vdev_queue_t *vq) 467{ 468 spa_t *spa = vq->vq_vdev->vdev_spa; 469 zio_priority_t p; 470 471 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active) 472 return (ZIO_PRIORITY_NUM_QUEUEABLE); 473 474 /* find a queue that has not reached its minimum # outstanding i/os */ 475 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 476 if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 && 477 vq->vq_class[p].vqc_active < 478 vdev_queue_class_min_active(p)) 479 return (p); 480 } 481 482 /* 483 * If we haven't found a queue, look for one that hasn't reached its 484 * maximum # outstanding i/os. 485 */ 486 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 487 if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 && 488 vq->vq_class[p].vqc_active < 489 vdev_queue_class_max_active(spa, p)) 490 return (p); 491 } 492 493 /* No eligible queued i/os */ 494 return (ZIO_PRIORITY_NUM_QUEUEABLE); 495} 496 497/* 498 * Compute the range spanned by two i/os, which is the endpoint of the last 499 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). 500 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); 501 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. 502 */ 503#define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) 504#define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) 505 506static zio_t * 507vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) 508{ 509 zio_t *first, *last, *aio, *dio, *mandatory, *nio; 510 uint64_t maxgap = 0; 511 uint64_t size; 512 boolean_t stretch = B_FALSE; 513 vdev_queue_class_t *vqc = &vq->vq_class[zio->io_priority]; 514 avl_tree_t *t = &vqc->vqc_queued_tree; 515 enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; 516 517 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE) 518 return (NULL); 519 520 /* 521 * The synchronous i/o queues are not sorted by LBA, so we can't 522 * find adjacent i/os. These i/os tend to not be tightly clustered, 523 * or too large to aggregate, so this has little impact on performance. 524 */ 525 if (zio->io_priority == ZIO_PRIORITY_SYNC_READ || 526 zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) 527 return (NULL); 528 529 first = last = zio; 530 531 if (zio->io_type == ZIO_TYPE_READ) 532 maxgap = zfs_vdev_read_gap_limit; 533 534 /* 535 * We can aggregate I/Os that are sufficiently adjacent and of 536 * the same flavor, as expressed by the AGG_INHERIT flags. 537 * The latter requirement is necessary so that certain 538 * attributes of the I/O, such as whether it's a normal I/O 539 * or a scrub/resilver, can be preserved in the aggregate. 540 * We can include optional I/Os, but don't allow them 541 * to begin a range as they add no benefit in that situation. 542 */ 543 544 /* 545 * We keep track of the last non-optional I/O. 546 */ 547 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; 548 549 /* 550 * Walk backwards through sufficiently contiguous I/Os 551 * recording the last non-option I/O. 552 */ 553 while ((dio = AVL_PREV(t, first)) != NULL && 554 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 555 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit && 556 IO_GAP(dio, first) <= maxgap) { 557 first = dio; 558 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) 559 mandatory = first; 560 } 561 562 /* 563 * Skip any initial optional I/Os. 564 */ 565 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { 566 first = AVL_NEXT(t, first); 567 ASSERT(first != NULL); 568 } 569 570 /* 571 * Walk forward through sufficiently contiguous I/Os. 572 */ 573 while ((dio = AVL_NEXT(t, last)) != NULL && 574 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 575 IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit && 576 IO_GAP(last, dio) <= maxgap) { 577 last = dio; 578 if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) 579 mandatory = last; 580 } 581 582 /* 583 * Now that we've established the range of the I/O aggregation 584 * we must decide what to do with trailing optional I/Os. 585 * For reads, there's nothing to do. While we are unable to 586 * aggregate further, it's possible that a trailing optional 587 * I/O would allow the underlying device to aggregate with 588 * subsequent I/Os. We must therefore determine if the next 589 * non-optional I/O is close enough to make aggregation 590 * worthwhile. 591 */ 592 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { 593 zio_t *nio = last; 594 while ((dio = AVL_NEXT(t, nio)) != NULL && 595 IO_GAP(nio, dio) == 0 && 596 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { 597 nio = dio; 598 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { 599 stretch = B_TRUE; 600 break; 601 } 602 } 603 } 604 605 if (stretch) { 606 /* This may be a no-op. */ 607 dio = AVL_NEXT(t, last); 608 dio->io_flags &= ~ZIO_FLAG_OPTIONAL; 609 } else { 610 while (last != mandatory && last != first) { 611 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); 612 last = AVL_PREV(t, last); 613 ASSERT(last != NULL); 614 } 615 } 616 617 if (first == last) 618 return (NULL); 619 620 size = IO_SPAN(first, last); 621 ASSERT3U(size, <=, zfs_vdev_aggregation_limit); 622 623 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, 624 zio_buf_alloc(size), size, first->io_type, zio->io_priority, 625 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE, 626 vdev_queue_agg_io_done, NULL); 627 aio->io_timestamp = first->io_timestamp; 628 629 nio = first; 630 do { 631 dio = nio; 632 nio = AVL_NEXT(t, dio); 633 ASSERT3U(dio->io_type, ==, aio->io_type); 634 635 if (dio->io_flags & ZIO_FLAG_NODATA) { 636 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); 637 bzero((char *)aio->io_data + (dio->io_offset - 638 aio->io_offset), dio->io_size); 639 } else if (dio->io_type == ZIO_TYPE_WRITE) { 640 bcopy(dio->io_data, (char *)aio->io_data + 641 (dio->io_offset - aio->io_offset), 642 dio->io_size); 643 } 644 645 zio_add_child(dio, aio); 646 vdev_queue_io_remove(vq, dio); 647 zio_vdev_io_bypass(dio); 648 zio_execute(dio); 649 } while (dio != last); 650 651 return (aio); 652} 653 654static zio_t * 655vdev_queue_io_to_issue(vdev_queue_t *vq) 656{ 657 zio_t *zio, *aio; 658 zio_priority_t p; 659 avl_index_t idx; 660 vdev_queue_class_t *vqc; 661 zio_t search; 662 663again: 664 ASSERT(MUTEX_HELD(&vq->vq_lock)); 665 666 p = vdev_queue_class_to_issue(vq); 667 668 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { 669 /* No eligible queued i/os */ 670 return (NULL); 671 } 672 673 /* 674 * For LBA-ordered queues (async / scrub), issue the i/o which follows 675 * the most recently issued i/o in LBA (offset) order. 676 * 677 * For FIFO queues (sync), issue the i/o with the lowest timestamp. 678 */ 679 vqc = &vq->vq_class[p]; 680 search.io_timestamp = 0; 681 search.io_offset = vq->vq_last_offset + 1; 682 VERIFY3P(avl_find(&vqc->vqc_queued_tree, &search, &idx), ==, NULL); 683 zio = avl_nearest(&vqc->vqc_queued_tree, idx, AVL_AFTER); 684 if (zio == NULL) 685 zio = avl_first(&vqc->vqc_queued_tree); 686 ASSERT3U(zio->io_priority, ==, p); 687 688 aio = vdev_queue_aggregate(vq, zio); 689 if (aio != NULL) 690 zio = aio; 691 else 692 vdev_queue_io_remove(vq, zio); 693 694 /* 695 * If the I/O is or was optional and therefore has no data, we need to 696 * simply discard it. We need to drop the vdev queue's lock to avoid a 697 * deadlock that we could encounter since this I/O will complete 698 * immediately. 699 */ 700 if (zio->io_flags & ZIO_FLAG_NODATA) { 701 mutex_exit(&vq->vq_lock); 702 zio_vdev_io_bypass(zio); 703 zio_execute(zio); 704 mutex_enter(&vq->vq_lock); 705 goto again; 706 } 707 708 vdev_queue_pending_add(vq, zio); 709 vq->vq_last_offset = zio->io_offset; 710 711 return (zio); 712} 713 714zio_t * 715vdev_queue_io(zio_t *zio) 716{ 717 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 718 zio_t *nio; 719 720 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) 721 return (zio); 722 723 /* 724 * Children i/os inherent their parent's priority, which might 725 * not match the child's i/o type. Fix it up here. 726 */ 727 if (zio->io_type == ZIO_TYPE_READ) { 728 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && 729 zio->io_priority != ZIO_PRIORITY_ASYNC_READ && 730 zio->io_priority != ZIO_PRIORITY_SCRUB) 731 zio->io_priority = ZIO_PRIORITY_ASYNC_READ; 732 } else { 733 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 734 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && 735 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE) 736 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; 737 } 738 739 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; 740 741 mutex_enter(&vq->vq_lock); 742 zio->io_timestamp = gethrtime(); 743 vdev_queue_io_add(vq, zio); 744 nio = vdev_queue_io_to_issue(vq); 745 mutex_exit(&vq->vq_lock); 746 747 if (nio == NULL) 748 return (NULL); 749 750 if (nio->io_done == vdev_queue_agg_io_done) { 751 zio_nowait(nio); 752 return (NULL); 753 } 754 755 return (nio); 756} 757 758void 759vdev_queue_io_done(zio_t *zio) 760{ 761 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 762 zio_t *nio; 763 764 if (zio_injection_enabled) 765 delay(SEC_TO_TICK(zio_handle_io_delay(zio))); 766 767 mutex_enter(&vq->vq_lock); 768 769 vdev_queue_pending_remove(vq, zio); 770 771 vq->vq_io_complete_ts = gethrtime(); 772 773 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { 774 mutex_exit(&vq->vq_lock); 775 if (nio->io_done == vdev_queue_agg_io_done) { 776 zio_nowait(nio); 777 } else { 778 zio_vdev_io_reissue(nio); 779 zio_execute(nio); 780 } 781 mutex_enter(&vq->vq_lock); 782 } 783 784 mutex_exit(&vq->vq_lock); 785} 786