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_OLD_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 TUNABLE_INT("vfs.zfs.vdev.async_write_active_min_dirty_percent",
181 &zfs_vdev_async_write_active_min_dirty_percent);
182 static int sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS);
183 SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_min_dirty_percent,
184 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
185 sysctl_zfs_async_write_active_min_dirty_percent, "I",
186 "Percentage of async write dirty data below which "
187 "async_write_min_active is used.");
189 TUNABLE_INT("vfs.zfs.vdev.async_write_active_max_dirty_percent",
190 &zfs_vdev_async_write_active_max_dirty_percent);
191 static int sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS);
192 SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_max_dirty_percent,
193 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
194 sysctl_zfs_async_write_active_max_dirty_percent, "I",
195 "Percentage of async write dirty data above which "
196 "async_write_max_active is used.");
198 TUNABLE_INT("vfs.zfs.vdev.max_active", &zfs_vdev_max_active);
199 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RWTUN,
200 &zfs_vdev_max_active, 0,
201 "The maximum number of I/Os of all types active for each device.");
203 #define ZFS_VDEV_QUEUE_KNOB_MIN(name) \
204 TUNABLE_INT("vfs.zfs.vdev." #name "_min_active", \
205 &zfs_vdev_ ## name ## _min_active); \
206 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, \
207 CTLFLAG_RWTUN, &zfs_vdev_ ## name ## _min_active, 0, \
208 "Initial number of I/O requests of type " #name \
209 " active for each device");
211 #define ZFS_VDEV_QUEUE_KNOB_MAX(name) \
212 TUNABLE_INT("vfs.zfs.vdev." #name "_max_active", \
213 &zfs_vdev_ ## name ## _max_active); \
214 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, \
215 CTLFLAG_RWTUN, &zfs_vdev_ ## name ## _max_active, 0, \
216 "Maximum number of I/O requests of type " #name \
217 " active for each device");
219 ZFS_VDEV_QUEUE_KNOB_MIN(sync_read);
220 ZFS_VDEV_QUEUE_KNOB_MAX(sync_read);
221 ZFS_VDEV_QUEUE_KNOB_MIN(sync_write);
222 ZFS_VDEV_QUEUE_KNOB_MAX(sync_write);
223 ZFS_VDEV_QUEUE_KNOB_MIN(async_read);
224 ZFS_VDEV_QUEUE_KNOB_MAX(async_read);
225 ZFS_VDEV_QUEUE_KNOB_MIN(async_write);
226 ZFS_VDEV_QUEUE_KNOB_MAX(async_write);
227 ZFS_VDEV_QUEUE_KNOB_MIN(scrub);
228 ZFS_VDEV_QUEUE_KNOB_MAX(scrub);
229 ZFS_VDEV_QUEUE_KNOB_MIN(trim);
230 ZFS_VDEV_QUEUE_KNOB_MAX(trim);
232 #undef ZFS_VDEV_QUEUE_KNOB
234 TUNABLE_INT("vfs.zfs.vdev.aggregation_limit", &zfs_vdev_aggregation_limit);
235 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RWTUN,
236 &zfs_vdev_aggregation_limit, 0,
237 "I/O requests are aggregated up to this size");
238 TUNABLE_INT("vfs.zfs.vdev.read_gap_limit", &zfs_vdev_read_gap_limit);
239 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RWTUN,
240 &zfs_vdev_read_gap_limit, 0,
241 "Acceptable gap between two reads being aggregated");
242 TUNABLE_INT("vfs.zfs.vdev.write_gap_limit", &zfs_vdev_write_gap_limit);
243 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RWTUN,
244 &zfs_vdev_write_gap_limit, 0,
245 "Acceptable gap between two writes being aggregated");
248 sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS)
252 val = zfs_vdev_async_write_active_min_dirty_percent;
253 err = sysctl_handle_int(oidp, &val, 0, req);
254 if (err != 0 || req->newptr == NULL)
257 if (val < 0 || val > 100 ||
258 val >= zfs_vdev_async_write_active_max_dirty_percent)
261 zfs_vdev_async_write_active_min_dirty_percent = val;
267 sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS)
271 val = zfs_vdev_async_write_active_max_dirty_percent;
272 err = sysctl_handle_int(oidp, &val, 0, req);
273 if (err != 0 || req->newptr == NULL)
276 if (val < 0 || val > 100 ||
277 val <= zfs_vdev_async_write_active_min_dirty_percent)
280 zfs_vdev_async_write_active_max_dirty_percent = val;
287 vdev_queue_offset_compare(const void *x1, const void *x2)
289 const zio_t *z1 = x1;
290 const zio_t *z2 = x2;
292 if (z1->io_offset < z2->io_offset)
294 if (z1->io_offset > z2->io_offset)
306 vdev_queue_timestamp_compare(const void *x1, const void *x2)
308 const zio_t *z1 = x1;
309 const zio_t *z2 = x2;
311 if (z1->io_timestamp < z2->io_timestamp)
313 if (z1->io_timestamp > z2->io_timestamp)
316 if (z1->io_offset < z2->io_offset)
318 if (z1->io_offset > z2->io_offset)
330 vdev_queue_init(vdev_t *vd)
332 vdev_queue_t *vq = &vd->vdev_queue;
334 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
337 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
338 sizeof (zio_t), offsetof(struct zio, io_queue_node));
340 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
342 * The synchronous i/o queues are FIFO rather than LBA ordered.
343 * This provides more consistent latency for these i/os, and
344 * they tend to not be tightly clustered anyway so there is
345 * little to no throughput loss.
347 boolean_t fifo = (p == ZIO_PRIORITY_SYNC_READ ||
348 p == ZIO_PRIORITY_SYNC_WRITE);
349 avl_create(&vq->vq_class[p].vqc_queued_tree,
350 fifo ? vdev_queue_timestamp_compare :
351 vdev_queue_offset_compare,
352 sizeof (zio_t), offsetof(struct zio, io_queue_node));
355 vq->vq_lastoffset = 0;
359 vdev_queue_fini(vdev_t *vd)
361 vdev_queue_t *vq = &vd->vdev_queue;
363 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
364 avl_destroy(&vq->vq_class[p].vqc_queued_tree);
365 avl_destroy(&vq->vq_active_tree);
367 mutex_destroy(&vq->vq_lock);
371 vdev_queue_io_add(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_add(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio);
379 mutex_enter(&spa->spa_iokstat_lock);
380 spa->spa_queue_stats[zio->io_priority].spa_queued++;
381 if (spa->spa_iokstat != NULL)
382 kstat_waitq_enter(spa->spa_iokstat->ks_data);
383 mutex_exit(&spa->spa_iokstat_lock);
388 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
390 spa_t *spa = zio->io_spa;
391 ASSERT(MUTEX_HELD(&vq->vq_lock));
392 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
393 avl_remove(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio);
396 mutex_enter(&spa->spa_iokstat_lock);
397 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
398 spa->spa_queue_stats[zio->io_priority].spa_queued--;
399 if (spa->spa_iokstat != NULL)
400 kstat_waitq_exit(spa->spa_iokstat->ks_data);
401 mutex_exit(&spa->spa_iokstat_lock);
406 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
408 spa_t *spa = zio->io_spa;
409 ASSERT(MUTEX_HELD(&vq->vq_lock));
410 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
411 vq->vq_class[zio->io_priority].vqc_active++;
412 avl_add(&vq->vq_active_tree, zio);
415 mutex_enter(&spa->spa_iokstat_lock);
416 spa->spa_queue_stats[zio->io_priority].spa_active++;
417 if (spa->spa_iokstat != NULL)
418 kstat_runq_enter(spa->spa_iokstat->ks_data);
419 mutex_exit(&spa->spa_iokstat_lock);
424 vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
426 spa_t *spa = zio->io_spa;
427 ASSERT(MUTEX_HELD(&vq->vq_lock));
428 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
429 vq->vq_class[zio->io_priority].vqc_active--;
430 avl_remove(&vq->vq_active_tree, zio);
433 mutex_enter(&spa->spa_iokstat_lock);
434 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
435 spa->spa_queue_stats[zio->io_priority].spa_active--;
436 if (spa->spa_iokstat != NULL) {
437 kstat_io_t *ksio = spa->spa_iokstat->ks_data;
439 kstat_runq_exit(spa->spa_iokstat->ks_data);
440 if (zio->io_type == ZIO_TYPE_READ) {
442 ksio->nread += zio->io_size;
443 } else if (zio->io_type == ZIO_TYPE_WRITE) {
445 ksio->nwritten += zio->io_size;
448 mutex_exit(&spa->spa_iokstat_lock);
453 vdev_queue_agg_io_done(zio_t *aio)
455 if (aio->io_type == ZIO_TYPE_READ) {
457 while ((pio = zio_walk_parents(aio)) != NULL) {
458 bcopy((char *)aio->io_data + (pio->io_offset -
459 aio->io_offset), pio->io_data, pio->io_size);
463 zio_buf_free(aio->io_data, aio->io_size);
467 vdev_queue_class_min_active(zio_priority_t p)
470 case ZIO_PRIORITY_SYNC_READ:
471 return (zfs_vdev_sync_read_min_active);
472 case ZIO_PRIORITY_SYNC_WRITE:
473 return (zfs_vdev_sync_write_min_active);
474 case ZIO_PRIORITY_ASYNC_READ:
475 return (zfs_vdev_async_read_min_active);
476 case ZIO_PRIORITY_ASYNC_WRITE:
477 return (zfs_vdev_async_write_min_active);
478 case ZIO_PRIORITY_SCRUB:
479 return (zfs_vdev_scrub_min_active);
480 case ZIO_PRIORITY_TRIM:
481 return (zfs_vdev_trim_min_active);
483 panic("invalid priority %u", p);
489 vdev_queue_max_async_writes(spa_t *spa)
492 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total;
493 uint64_t min_bytes = zfs_dirty_data_max *
494 zfs_vdev_async_write_active_min_dirty_percent / 100;
495 uint64_t max_bytes = zfs_dirty_data_max *
496 zfs_vdev_async_write_active_max_dirty_percent / 100;
499 * Sync tasks correspond to interactive user actions. To reduce the
500 * execution time of those actions we push data out as fast as possible.
502 if (spa_has_pending_synctask(spa)) {
503 return (zfs_vdev_async_write_max_active);
506 if (dirty < min_bytes)
507 return (zfs_vdev_async_write_min_active);
508 if (dirty > max_bytes)
509 return (zfs_vdev_async_write_max_active);
512 * linear interpolation:
513 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
514 * move right by min_bytes
515 * move up by min_writes
517 writes = (dirty - min_bytes) *
518 (zfs_vdev_async_write_max_active -
519 zfs_vdev_async_write_min_active) /
520 (max_bytes - min_bytes) +
521 zfs_vdev_async_write_min_active;
522 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
523 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
528 vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
531 case ZIO_PRIORITY_SYNC_READ:
532 return (zfs_vdev_sync_read_max_active);
533 case ZIO_PRIORITY_SYNC_WRITE:
534 return (zfs_vdev_sync_write_max_active);
535 case ZIO_PRIORITY_ASYNC_READ:
536 return (zfs_vdev_async_read_max_active);
537 case ZIO_PRIORITY_ASYNC_WRITE:
538 return (vdev_queue_max_async_writes(spa));
539 case ZIO_PRIORITY_SCRUB:
540 return (zfs_vdev_scrub_max_active);
541 case ZIO_PRIORITY_TRIM:
542 return (zfs_vdev_trim_max_active);
544 panic("invalid priority %u", p);
550 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
551 * there is no eligible class.
553 static zio_priority_t
554 vdev_queue_class_to_issue(vdev_queue_t *vq)
556 spa_t *spa = vq->vq_vdev->vdev_spa;
559 ASSERT(MUTEX_HELD(&vq->vq_lock));
561 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
562 return (ZIO_PRIORITY_NUM_QUEUEABLE);
564 /* find a queue that has not reached its minimum # outstanding i/os */
565 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
566 if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 &&
567 vq->vq_class[p].vqc_active <
568 vdev_queue_class_min_active(p))
573 * If we haven't found a queue, look for one that hasn't reached its
574 * maximum # outstanding i/os.
576 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
577 if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 &&
578 vq->vq_class[p].vqc_active <
579 vdev_queue_class_max_active(spa, p))
583 /* No eligible queued i/os */
584 return (ZIO_PRIORITY_NUM_QUEUEABLE);
588 * Compute the range spanned by two i/os, which is the endpoint of the last
589 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
590 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
591 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
593 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
594 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
597 vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
599 zio_t *first, *last, *aio, *dio, *mandatory, *nio;
606 ASSERT(MUTEX_HELD(&vq->vq_lock));
608 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
612 * The synchronous i/o queues are not sorted by LBA, so we can't
613 * find adjacent i/os. These i/os tend to not be tightly clustered,
614 * or too large to aggregate, so this has little impact on performance.
616 if (zio->io_priority == ZIO_PRIORITY_SYNC_READ ||
617 zio->io_priority == ZIO_PRIORITY_SYNC_WRITE)
622 if (zio->io_type == ZIO_TYPE_READ)
623 maxgap = zfs_vdev_read_gap_limit;
626 * We can aggregate I/Os that are sufficiently adjacent and of
627 * the same flavor, as expressed by the AGG_INHERIT flags.
628 * The latter requirement is necessary so that certain
629 * attributes of the I/O, such as whether it's a normal I/O
630 * or a scrub/resilver, can be preserved in the aggregate.
631 * We can include optional I/Os, but don't allow them
632 * to begin a range as they add no benefit in that situation.
636 * We keep track of the last non-optional I/O.
638 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
641 * Walk backwards through sufficiently contiguous I/Os
642 * recording the last non-option I/O.
644 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
645 t = &vq->vq_class[zio->io_priority].vqc_queued_tree;
646 while ((dio = AVL_PREV(t, first)) != NULL &&
647 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
648 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
649 IO_GAP(dio, first) <= maxgap) {
651 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
656 * Skip any initial optional I/Os.
658 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
659 first = AVL_NEXT(t, first);
660 ASSERT(first != NULL);
664 * Walk forward through sufficiently contiguous I/Os.
666 while ((dio = AVL_NEXT(t, last)) != NULL &&
667 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
668 IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit &&
669 IO_GAP(last, dio) <= maxgap) {
671 if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
676 * Now that we've established the range of the I/O aggregation
677 * we must decide what to do with trailing optional I/Os.
678 * For reads, there's nothing to do. While we are unable to
679 * aggregate further, it's possible that a trailing optional
680 * I/O would allow the underlying device to aggregate with
681 * subsequent I/Os. We must therefore determine if the next
682 * non-optional I/O is close enough to make aggregation
686 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
688 while ((dio = AVL_NEXT(t, nio)) != NULL &&
689 IO_GAP(nio, dio) == 0 &&
690 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
692 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
700 /* This may be a no-op. */
701 dio = AVL_NEXT(t, last);
702 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
704 while (last != mandatory && last != first) {
705 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
706 last = AVL_PREV(t, last);
707 ASSERT(last != NULL);
714 size = IO_SPAN(first, last);
715 ASSERT3U(size, <=, zfs_vdev_aggregation_limit);
717 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
718 zio_buf_alloc(size), size, first->io_type, zio->io_priority,
719 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
720 vdev_queue_agg_io_done, NULL);
721 aio->io_timestamp = first->io_timestamp;
726 nio = AVL_NEXT(t, dio);
727 ASSERT3U(dio->io_type, ==, aio->io_type);
729 if (dio->io_flags & ZIO_FLAG_NODATA) {
730 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
731 bzero((char *)aio->io_data + (dio->io_offset -
732 aio->io_offset), dio->io_size);
733 } else if (dio->io_type == ZIO_TYPE_WRITE) {
734 bcopy(dio->io_data, (char *)aio->io_data +
735 (dio->io_offset - aio->io_offset),
739 zio_add_child(dio, aio);
740 vdev_queue_io_remove(vq, dio);
741 zio_vdev_io_bypass(dio);
743 } while (dio != last);
749 vdev_queue_io_to_issue(vdev_queue_t *vq)
754 vdev_queue_class_t *vqc;
758 ASSERT(MUTEX_HELD(&vq->vq_lock));
760 p = vdev_queue_class_to_issue(vq);
762 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
763 /* No eligible queued i/os */
768 * For LBA-ordered queues (async / scrub), issue the i/o which follows
769 * the most recently issued i/o in LBA (offset) order.
771 * For FIFO queues (sync), issue the i/o with the lowest timestamp.
773 vqc = &vq->vq_class[p];
774 search.io_timestamp = 0;
775 search.io_offset = vq->vq_last_offset + 1;
776 VERIFY3P(avl_find(&vqc->vqc_queued_tree, &search, &idx), ==, NULL);
777 zio = avl_nearest(&vqc->vqc_queued_tree, idx, AVL_AFTER);
779 zio = avl_first(&vqc->vqc_queued_tree);
780 ASSERT3U(zio->io_priority, ==, p);
782 aio = vdev_queue_aggregate(vq, zio);
786 vdev_queue_io_remove(vq, zio);
789 * If the I/O is or was optional and therefore has no data, we need to
790 * simply discard it. We need to drop the vdev queue's lock to avoid a
791 * deadlock that we could encounter since this I/O will complete
794 if (zio->io_flags & ZIO_FLAG_NODATA) {
795 mutex_exit(&vq->vq_lock);
796 zio_vdev_io_bypass(zio);
798 mutex_enter(&vq->vq_lock);
802 vdev_queue_pending_add(vq, zio);
803 vq->vq_last_offset = zio->io_offset;
809 vdev_queue_io(zio_t *zio)
811 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
814 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
818 * Children i/os inherent their parent's priority, which might
819 * not match the child's i/o type. Fix it up here.
821 if (zio->io_type == ZIO_TYPE_READ) {
822 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
823 zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
824 zio->io_priority != ZIO_PRIORITY_SCRUB)
825 zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
826 } else if (zio->io_type == ZIO_TYPE_WRITE) {
827 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
828 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE)
829 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
831 ASSERT(zio->io_type == ZIO_TYPE_FREE);
832 zio->io_priority = ZIO_PRIORITY_TRIM;
835 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
837 mutex_enter(&vq->vq_lock);
838 zio->io_timestamp = gethrtime();
839 vdev_queue_io_add(vq, zio);
840 nio = vdev_queue_io_to_issue(vq);
841 mutex_exit(&vq->vq_lock);
846 if (nio->io_done == vdev_queue_agg_io_done) {
855 vdev_queue_io_done(zio_t *zio)
857 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
860 if (zio_injection_enabled)
861 delay(SEC_TO_TICK(zio_handle_io_delay(zio)));
863 mutex_enter(&vq->vq_lock);
865 vdev_queue_pending_remove(vq, zio);
867 vq->vq_io_complete_ts = gethrtime();
869 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
870 mutex_exit(&vq->vq_lock);
871 if (nio->io_done == vdev_queue_agg_io_done) {
874 zio_vdev_io_reissue(nio);
877 mutex_enter(&vq->vq_lock);
880 mutex_exit(&vq->vq_lock);
884 * As these three methods are only used for load calculations we're not concerned
885 * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex
886 * use here, instead we prefer to keep it lock free for performance.
889 vdev_queue_length(vdev_t *vd)
891 return (avl_numnodes(&vd->vdev_queue.vq_active_tree));
895 vdev_queue_lastoffset(vdev_t *vd)
897 return (vd->vdev_queue.vq_lastoffset);
901 vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio)
903 vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size;