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.
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.
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]
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
27 * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
30 #include <sys/zfs_context.h>
31 #include <sys/vdev_impl.h>
32 #include <sys/spa_impl.h>
35 #include <sys/dsl_pool.h>
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.
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.
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.
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.
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.
87 * | o---------| <-- zfs_vdev_async_write_max_active
94 * |------------o | | <-- zfs_vdev_async_write_min_active
95 * 0|____________^______|_________|
96 * 0% | | 100% of zfs_dirty_data_max
98 * | `-- zfs_vdev_async_write_active_max_dirty_percent
99 * `--------- zfs_vdev_async_write_active_min_dirty_percent
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.
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).
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.
120 uint32_t zfs_vdev_max_active = 1000;
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.
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.
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
138 uint32_t zfs_vdev_sync_read_min_active = 10;
139 uint32_t zfs_vdev_sync_read_max_active = 10;
140 uint32_t zfs_vdev_sync_write_min_active = 10;
141 uint32_t zfs_vdev_sync_write_max_active = 10;
142 uint32_t zfs_vdev_async_read_min_active = 1;
143 uint32_t zfs_vdev_async_read_max_active = 3;
144 uint32_t zfs_vdev_async_write_min_active = 1;
145 uint32_t zfs_vdev_async_write_max_active = 10;
146 uint32_t zfs_vdev_scrub_min_active = 1;
147 uint32_t zfs_vdev_scrub_max_active = 2;
148 uint32_t zfs_vdev_trim_min_active = 1;
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.
154 uint32_t zfs_vdev_trim_max_active = 64;
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.
164 int zfs_vdev_async_write_active_min_dirty_percent = 30;
165 int zfs_vdev_async_write_active_max_dirty_percent = 60;
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.
173 int zfs_vdev_aggregation_limit = SPA_MAXBLOCKSIZE;
174 int zfs_vdev_read_gap_limit = 32 << 10;
175 int zfs_vdev_write_gap_limit = 4 << 10;
178 SYSCTL_DECL(_vfs_zfs_vdev);
180 static int sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS);
181 SYSCTL_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.");
187 static int sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS);
188 SYSCTL_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.");
194 SYSCTL_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.");
198 #define ZFS_VDEV_QUEUE_KNOB_MIN(name) \
199 SYSCTL_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");
204 #define ZFS_VDEV_QUEUE_KNOB_MAX(name) \
205 SYSCTL_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");
210 ZFS_VDEV_QUEUE_KNOB_MIN(sync_read);
211 ZFS_VDEV_QUEUE_KNOB_MAX(sync_read);
212 ZFS_VDEV_QUEUE_KNOB_MIN(sync_write);
213 ZFS_VDEV_QUEUE_KNOB_MAX(sync_write);
214 ZFS_VDEV_QUEUE_KNOB_MIN(async_read);
215 ZFS_VDEV_QUEUE_KNOB_MAX(async_read);
216 ZFS_VDEV_QUEUE_KNOB_MIN(async_write);
217 ZFS_VDEV_QUEUE_KNOB_MAX(async_write);
218 ZFS_VDEV_QUEUE_KNOB_MIN(scrub);
219 ZFS_VDEV_QUEUE_KNOB_MAX(scrub);
220 ZFS_VDEV_QUEUE_KNOB_MIN(trim);
221 ZFS_VDEV_QUEUE_KNOB_MAX(trim);
223 #undef ZFS_VDEV_QUEUE_KNOB
225 SYSCTL_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");
228 SYSCTL_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");
231 SYSCTL_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");
236 sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS)
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)
245 if (val < 0 || val > 100 ||
246 val >= zfs_vdev_async_write_active_max_dirty_percent)
249 zfs_vdev_async_write_active_min_dirty_percent = val;
255 sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS)
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)
264 if (val < 0 || val > 100 ||
265 val <= zfs_vdev_async_write_active_min_dirty_percent)
268 zfs_vdev_async_write_active_max_dirty_percent = val;
275 vdev_queue_offset_compare(const void *x1, const void *x2)
277 const zio_t *z1 = x1;
278 const zio_t *z2 = x2;
280 if (z1->io_offset < z2->io_offset)
282 if (z1->io_offset > z2->io_offset)
294 vdev_queue_timestamp_compare(const void *x1, const void *x2)
296 const zio_t *z1 = x1;
297 const zio_t *z2 = x2;
299 if (z1->io_timestamp < z2->io_timestamp)
301 if (z1->io_timestamp > z2->io_timestamp)
313 vdev_queue_init(vdev_t *vd)
315 vdev_queue_t *vq = &vd->vdev_queue;
317 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
320 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
321 sizeof (zio_t), offsetof(struct zio, io_queue_node));
323 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
325 * The synchronous i/o queues are FIFO rather than LBA ordered.
326 * This provides more consistent latency for these i/os, and
327 * they tend to not be tightly clustered anyway so there is
328 * little to no throughput loss.
330 boolean_t fifo = (p == ZIO_PRIORITY_SYNC_READ ||
331 p == ZIO_PRIORITY_SYNC_WRITE);
332 avl_create(&vq->vq_class[p].vqc_queued_tree,
333 fifo ? vdev_queue_timestamp_compare :
334 vdev_queue_offset_compare,
335 sizeof (zio_t), offsetof(struct zio, io_queue_node));
338 vq->vq_lastoffset = 0;
342 vdev_queue_fini(vdev_t *vd)
344 vdev_queue_t *vq = &vd->vdev_queue;
346 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
347 avl_destroy(&vq->vq_class[p].vqc_queued_tree);
348 avl_destroy(&vq->vq_active_tree);
350 mutex_destroy(&vq->vq_lock);
354 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
356 spa_t *spa = zio->io_spa;
357 ASSERT(MUTEX_HELD(&vq->vq_lock));
358 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
359 avl_add(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio);
362 mutex_enter(&spa->spa_iokstat_lock);
363 spa->spa_queue_stats[zio->io_priority].spa_queued++;
364 if (spa->spa_iokstat != NULL)
365 kstat_waitq_enter(spa->spa_iokstat->ks_data);
366 mutex_exit(&spa->spa_iokstat_lock);
371 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
373 spa_t *spa = zio->io_spa;
374 ASSERT(MUTEX_HELD(&vq->vq_lock));
375 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
376 avl_remove(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio);
379 mutex_enter(&spa->spa_iokstat_lock);
380 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
381 spa->spa_queue_stats[zio->io_priority].spa_queued--;
382 if (spa->spa_iokstat != NULL)
383 kstat_waitq_exit(spa->spa_iokstat->ks_data);
384 mutex_exit(&spa->spa_iokstat_lock);
389 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
391 spa_t *spa = zio->io_spa;
392 ASSERT(MUTEX_HELD(&vq->vq_lock));
393 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
394 vq->vq_class[zio->io_priority].vqc_active++;
395 avl_add(&vq->vq_active_tree, zio);
398 mutex_enter(&spa->spa_iokstat_lock);
399 spa->spa_queue_stats[zio->io_priority].spa_active++;
400 if (spa->spa_iokstat != NULL)
401 kstat_runq_enter(spa->spa_iokstat->ks_data);
402 mutex_exit(&spa->spa_iokstat_lock);
407 vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
409 spa_t *spa = zio->io_spa;
410 ASSERT(MUTEX_HELD(&vq->vq_lock));
411 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
412 vq->vq_class[zio->io_priority].vqc_active--;
413 avl_remove(&vq->vq_active_tree, zio);
416 mutex_enter(&spa->spa_iokstat_lock);
417 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
418 spa->spa_queue_stats[zio->io_priority].spa_active--;
419 if (spa->spa_iokstat != NULL) {
420 kstat_io_t *ksio = spa->spa_iokstat->ks_data;
422 kstat_runq_exit(spa->spa_iokstat->ks_data);
423 if (zio->io_type == ZIO_TYPE_READ) {
425 ksio->nread += zio->io_size;
426 } else if (zio->io_type == ZIO_TYPE_WRITE) {
428 ksio->nwritten += zio->io_size;
431 mutex_exit(&spa->spa_iokstat_lock);
436 vdev_queue_agg_io_done(zio_t *aio)
438 if (aio->io_type == ZIO_TYPE_READ) {
440 while ((pio = zio_walk_parents(aio)) != NULL) {
441 bcopy((char *)aio->io_data + (pio->io_offset -
442 aio->io_offset), pio->io_data, pio->io_size);
446 zio_buf_free(aio->io_data, aio->io_size);
450 vdev_queue_class_min_active(zio_priority_t p)
453 case ZIO_PRIORITY_SYNC_READ:
454 return (zfs_vdev_sync_read_min_active);
455 case ZIO_PRIORITY_SYNC_WRITE:
456 return (zfs_vdev_sync_write_min_active);
457 case ZIO_PRIORITY_ASYNC_READ:
458 return (zfs_vdev_async_read_min_active);
459 case ZIO_PRIORITY_ASYNC_WRITE:
460 return (zfs_vdev_async_write_min_active);
461 case ZIO_PRIORITY_SCRUB:
462 return (zfs_vdev_scrub_min_active);
463 case ZIO_PRIORITY_TRIM:
464 return (zfs_vdev_trim_min_active);
466 panic("invalid priority %u", p);
472 vdev_queue_max_async_writes(spa_t *spa)
475 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total;
476 uint64_t min_bytes = zfs_dirty_data_max *
477 zfs_vdev_async_write_active_min_dirty_percent / 100;
478 uint64_t max_bytes = zfs_dirty_data_max *
479 zfs_vdev_async_write_active_max_dirty_percent / 100;
482 * Sync tasks correspond to interactive user actions. To reduce the
483 * execution time of those actions we push data out as fast as possible.
485 if (spa_has_pending_synctask(spa)) {
486 return (zfs_vdev_async_write_max_active);
489 if (dirty < min_bytes)
490 return (zfs_vdev_async_write_min_active);
491 if (dirty > max_bytes)
492 return (zfs_vdev_async_write_max_active);
495 * linear interpolation:
496 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
497 * move right by min_bytes
498 * move up by min_writes
500 writes = (dirty - min_bytes) *
501 (zfs_vdev_async_write_max_active -
502 zfs_vdev_async_write_min_active) /
503 (max_bytes - min_bytes) +
504 zfs_vdev_async_write_min_active;
505 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
506 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
511 vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
514 case ZIO_PRIORITY_SYNC_READ:
515 return (zfs_vdev_sync_read_max_active);
516 case ZIO_PRIORITY_SYNC_WRITE:
517 return (zfs_vdev_sync_write_max_active);
518 case ZIO_PRIORITY_ASYNC_READ:
519 return (zfs_vdev_async_read_max_active);
520 case ZIO_PRIORITY_ASYNC_WRITE:
521 return (vdev_queue_max_async_writes(spa));
522 case ZIO_PRIORITY_SCRUB:
523 return (zfs_vdev_scrub_max_active);
524 case ZIO_PRIORITY_TRIM:
525 return (zfs_vdev_trim_max_active);
527 panic("invalid priority %u", p);
533 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
534 * there is no eligible class.
536 static zio_priority_t
537 vdev_queue_class_to_issue(vdev_queue_t *vq)
539 spa_t *spa = vq->vq_vdev->vdev_spa;
542 ASSERT(MUTEX_HELD(&vq->vq_lock));
544 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
545 return (ZIO_PRIORITY_NUM_QUEUEABLE);
547 /* find a queue that has not reached its minimum # outstanding i/os */
548 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
549 if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 &&
550 vq->vq_class[p].vqc_active <
551 vdev_queue_class_min_active(p))
556 * If we haven't found a queue, look for one that hasn't reached its
557 * maximum # outstanding i/os.
559 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
560 if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 &&
561 vq->vq_class[p].vqc_active <
562 vdev_queue_class_max_active(spa, p))
566 /* No eligible queued i/os */
567 return (ZIO_PRIORITY_NUM_QUEUEABLE);
571 * Compute the range spanned by two i/os, which is the endpoint of the last
572 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
573 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
574 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
576 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
577 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
580 vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
582 zio_t *first, *last, *aio, *dio, *mandatory, *nio;
589 ASSERT(MUTEX_HELD(&vq->vq_lock));
591 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
595 * The synchronous i/o queues are not sorted by LBA, so we can't
596 * find adjacent i/os. These i/os tend to not be tightly clustered,
597 * or too large to aggregate, so this has little impact on performance.
599 if (zio->io_priority == ZIO_PRIORITY_SYNC_READ ||
600 zio->io_priority == ZIO_PRIORITY_SYNC_WRITE)
605 if (zio->io_type == ZIO_TYPE_READ)
606 maxgap = zfs_vdev_read_gap_limit;
609 * We can aggregate I/Os that are sufficiently adjacent and of
610 * the same flavor, as expressed by the AGG_INHERIT flags.
611 * The latter requirement is necessary so that certain
612 * attributes of the I/O, such as whether it's a normal I/O
613 * or a scrub/resilver, can be preserved in the aggregate.
614 * We can include optional I/Os, but don't allow them
615 * to begin a range as they add no benefit in that situation.
619 * We keep track of the last non-optional I/O.
621 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
624 * Walk backwards through sufficiently contiguous I/Os
625 * recording the last non-option I/O.
627 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
628 t = &vq->vq_class[zio->io_priority].vqc_queued_tree;
629 while ((dio = AVL_PREV(t, first)) != NULL &&
630 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
631 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
632 IO_GAP(dio, first) <= maxgap) {
634 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
639 * Skip any initial optional I/Os.
641 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
642 first = AVL_NEXT(t, first);
643 ASSERT(first != NULL);
647 * Walk forward through sufficiently contiguous I/Os.
649 while ((dio = AVL_NEXT(t, last)) != NULL &&
650 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
651 IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit &&
652 IO_GAP(last, dio) <= maxgap) {
654 if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
659 * Now that we've established the range of the I/O aggregation
660 * we must decide what to do with trailing optional I/Os.
661 * For reads, there's nothing to do. While we are unable to
662 * aggregate further, it's possible that a trailing optional
663 * I/O would allow the underlying device to aggregate with
664 * subsequent I/Os. We must therefore determine if the next
665 * non-optional I/O is close enough to make aggregation
669 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
671 while ((dio = AVL_NEXT(t, nio)) != NULL &&
672 IO_GAP(nio, dio) == 0 &&
673 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
675 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
683 /* This may be a no-op. */
684 dio = AVL_NEXT(t, last);
685 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
687 while (last != mandatory && last != first) {
688 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
689 last = AVL_PREV(t, last);
690 ASSERT(last != NULL);
697 size = IO_SPAN(first, last);
698 ASSERT3U(size, <=, zfs_vdev_aggregation_limit);
700 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
701 zio_buf_alloc(size), size, first->io_type, zio->io_priority,
702 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
703 vdev_queue_agg_io_done, NULL);
704 aio->io_timestamp = first->io_timestamp;
709 nio = AVL_NEXT(t, dio);
710 ASSERT3U(dio->io_type, ==, aio->io_type);
712 if (dio->io_flags & ZIO_FLAG_NODATA) {
713 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
714 bzero((char *)aio->io_data + (dio->io_offset -
715 aio->io_offset), dio->io_size);
716 } else if (dio->io_type == ZIO_TYPE_WRITE) {
717 bcopy(dio->io_data, (char *)aio->io_data +
718 (dio->io_offset - aio->io_offset),
722 zio_add_child(dio, aio);
723 vdev_queue_io_remove(vq, dio);
724 zio_vdev_io_bypass(dio);
726 } while (dio != last);
732 vdev_queue_io_to_issue(vdev_queue_t *vq)
737 vdev_queue_class_t *vqc;
741 ASSERT(MUTEX_HELD(&vq->vq_lock));
743 p = vdev_queue_class_to_issue(vq);
745 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
746 /* No eligible queued i/os */
751 * For LBA-ordered queues (async / scrub), issue the i/o which follows
752 * the most recently issued i/o in LBA (offset) order.
754 * For FIFO queues (sync), issue the i/o with the lowest timestamp.
756 vqc = &vq->vq_class[p];
757 search.io_timestamp = 0;
758 search.io_offset = vq->vq_last_offset + 1;
759 VERIFY3P(avl_find(&vqc->vqc_queued_tree, &search, &idx), ==, NULL);
760 zio = avl_nearest(&vqc->vqc_queued_tree, idx, AVL_AFTER);
762 zio = avl_first(&vqc->vqc_queued_tree);
763 ASSERT3U(zio->io_priority, ==, p);
765 aio = vdev_queue_aggregate(vq, zio);
769 vdev_queue_io_remove(vq, zio);
772 * If the I/O is or was optional and therefore has no data, we need to
773 * simply discard it. We need to drop the vdev queue's lock to avoid a
774 * deadlock that we could encounter since this I/O will complete
777 if (zio->io_flags & ZIO_FLAG_NODATA) {
778 mutex_exit(&vq->vq_lock);
779 zio_vdev_io_bypass(zio);
781 mutex_enter(&vq->vq_lock);
785 vdev_queue_pending_add(vq, zio);
786 vq->vq_last_offset = zio->io_offset;
792 vdev_queue_io(zio_t *zio)
794 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
797 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
801 * Children i/os inherent their parent's priority, which might
802 * not match the child's i/o type. Fix it up here.
804 if (zio->io_type == ZIO_TYPE_READ) {
805 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
806 zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
807 zio->io_priority != ZIO_PRIORITY_SCRUB)
808 zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
809 } else if (zio->io_type == ZIO_TYPE_WRITE) {
810 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
811 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE)
812 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
814 ASSERT(zio->io_type == ZIO_TYPE_FREE);
815 zio->io_priority = ZIO_PRIORITY_TRIM;
818 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
820 mutex_enter(&vq->vq_lock);
821 zio->io_timestamp = gethrtime();
822 vdev_queue_io_add(vq, zio);
823 nio = vdev_queue_io_to_issue(vq);
824 mutex_exit(&vq->vq_lock);
829 if (nio->io_done == vdev_queue_agg_io_done) {
838 vdev_queue_io_done(zio_t *zio)
840 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
843 if (zio_injection_enabled)
844 delay(SEC_TO_TICK(zio_handle_io_delay(zio)));
846 mutex_enter(&vq->vq_lock);
848 vdev_queue_pending_remove(vq, zio);
850 vq->vq_io_complete_ts = gethrtime();
852 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
853 mutex_exit(&vq->vq_lock);
854 if (nio->io_done == vdev_queue_agg_io_done) {
857 zio_vdev_io_reissue(nio);
860 mutex_enter(&vq->vq_lock);
863 mutex_exit(&vq->vq_lock);
867 * As these three methods are only used for load calculations we're not concerned
868 * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex
869 * use here, instead we prefer to keep it lock free for performance.
872 vdev_queue_length(vdev_t *vd)
874 return (avl_numnodes(&vd->vdev_queue.vq_active_tree));
878 vdev_queue_lastoffset(vdev_t *vd)
880 return (vd->vdev_queue.vq_lastoffset);
884 vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio)
886 vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size;