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.
28 * Copyright (c) 2014 Integros [integros.com]
31 #include <sys/zfs_context.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/spa_impl.h>
36 #include <sys/dsl_pool.h>
42 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The
43 * I/O scheduler determines when and in what order those operations are
44 * issued. The I/O scheduler divides operations into six I/O classes
45 * prioritized in the following order: sync read, sync write, async read,
46 * async write, scrub/resilver and trim. Each queue defines the minimum and
47 * maximum number of concurrent operations that may be issued to the device.
48 * In addition, the device has an aggregate maximum. Note that the sum of the
49 * per-queue minimums must not exceed the aggregate maximum, and if the
50 * aggregate maximum is equal to or greater than the sum of the per-queue
51 * maximums, the per-queue minimum has no effect.
53 * For many physical devices, throughput increases with the number of
54 * concurrent operations, but latency typically suffers. Further, physical
55 * devices typically have a limit at which more concurrent operations have no
56 * effect on throughput or can actually cause it to decrease.
58 * The scheduler selects the next operation to issue by first looking for an
59 * I/O class whose minimum has not been satisfied. Once all are satisfied and
60 * the aggregate maximum has not been hit, the scheduler looks for classes
61 * whose maximum has not been satisfied. Iteration through the I/O classes is
62 * done in the order specified above. No further operations are issued if the
63 * aggregate maximum number of concurrent operations has been hit or if there
64 * are no operations queued for an I/O class that has not hit its maximum.
65 * Every time an I/O is queued or an operation completes, the I/O scheduler
66 * looks for new operations to issue.
68 * All I/O classes have a fixed maximum number of outstanding operations
69 * except for the async write class. Asynchronous writes represent the data
70 * that is committed to stable storage during the syncing stage for
71 * transaction groups (see txg.c). Transaction groups enter the syncing state
72 * periodically so the number of queued async writes will quickly burst up and
73 * then bleed down to zero. Rather than servicing them as quickly as possible,
74 * the I/O scheduler changes the maximum number of active async write I/Os
75 * according to the amount of dirty data in the pool (see dsl_pool.c). Since
76 * both throughput and latency typically increase with the number of
77 * concurrent operations issued to physical devices, reducing the burstiness
78 * in the number of concurrent operations also stabilizes the response time of
79 * operations from other -- and in particular synchronous -- queues. In broad
80 * strokes, the I/O scheduler will issue more concurrent operations from the
81 * async write queue as there's more dirty data in the pool.
85 * The number of concurrent operations issued for the async write I/O class
86 * follows a piece-wise linear function defined by a few adjustable points.
88 * | o---------| <-- zfs_vdev_async_write_max_active
95 * |------------o | | <-- zfs_vdev_async_write_min_active
96 * 0|____________^______|_________|
97 * 0% | | 100% of zfs_dirty_data_max
99 * | `-- zfs_vdev_async_write_active_max_dirty_percent
100 * `--------- zfs_vdev_async_write_active_min_dirty_percent
102 * Until the amount of dirty data exceeds a minimum percentage of the dirty
103 * data allowed in the pool, the I/O scheduler will limit the number of
104 * concurrent operations to the minimum. As that threshold is crossed, the
105 * number of concurrent operations issued increases linearly to the maximum at
106 * the specified maximum percentage of the dirty data allowed in the pool.
108 * Ideally, the amount of dirty data on a busy pool will stay in the sloped
109 * part of the function between zfs_vdev_async_write_active_min_dirty_percent
110 * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the
111 * maximum percentage, this indicates that the rate of incoming data is
112 * greater than the rate that the backend storage can handle. In this case, we
113 * must further throttle incoming writes (see dmu_tx_delay() for details).
117 * The maximum number of I/Os active to each device. Ideally, this will be >=
118 * the sum of each queue's max_active. It must be at least the sum of each
119 * queue's min_active.
121 uint32_t zfs_vdev_max_active = 1000;
124 * Per-queue limits on the number of I/Os active to each device. If the
125 * sum of the queue's max_active is < zfs_vdev_max_active, then the
126 * min_active comes into play. We will send min_active from each queue,
127 * and then select from queues in the order defined by zio_priority_t.
129 * In general, smaller max_active's will lead to lower latency of synchronous
130 * operations. Larger max_active's may lead to higher overall throughput,
131 * depending on underlying storage.
133 * The ratio of the queues' max_actives determines the balance of performance
134 * between reads, writes, and scrubs. E.g., increasing
135 * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete
136 * more quickly, but reads and writes to have higher latency and lower
139 uint32_t zfs_vdev_sync_read_min_active = 10;
140 uint32_t zfs_vdev_sync_read_max_active = 10;
141 uint32_t zfs_vdev_sync_write_min_active = 10;
142 uint32_t zfs_vdev_sync_write_max_active = 10;
143 uint32_t zfs_vdev_async_read_min_active = 1;
144 uint32_t zfs_vdev_async_read_max_active = 3;
145 uint32_t zfs_vdev_async_write_min_active = 1;
146 uint32_t zfs_vdev_async_write_max_active = 10;
147 uint32_t zfs_vdev_scrub_min_active = 1;
148 uint32_t zfs_vdev_scrub_max_active = 2;
149 uint32_t zfs_vdev_trim_min_active = 1;
151 * TRIM max active is large in comparison to the other values due to the fact
152 * that TRIM IOs are coalesced at the device layer. This value is set such
153 * that a typical SSD can process the queued IOs in a single request.
155 uint32_t zfs_vdev_trim_max_active = 64;
159 * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent
160 * dirty data, use zfs_vdev_async_write_min_active. When it has more than
161 * zfs_vdev_async_write_active_max_dirty_percent, use
162 * zfs_vdev_async_write_max_active. The value is linearly interpolated
163 * between min and max.
165 int zfs_vdev_async_write_active_min_dirty_percent = 30;
166 int zfs_vdev_async_write_active_max_dirty_percent = 60;
169 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
170 * For read I/Os, we also aggregate across small adjacency gaps; for writes
171 * we include spans of optional I/Os to aid aggregation at the disk even when
172 * they aren't able to help us aggregate at this level.
174 int zfs_vdev_aggregation_limit = SPA_OLD_MAXBLOCKSIZE;
175 int zfs_vdev_read_gap_limit = 32 << 10;
176 int zfs_vdev_write_gap_limit = 4 << 10;
180 SYSCTL_DECL(_vfs_zfs_vdev);
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 static int sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS);
190 SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_max_dirty_percent,
191 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
192 sysctl_zfs_async_write_active_max_dirty_percent, "I",
193 "Percentage of async write dirty data above which "
194 "async_write_max_active is used.");
196 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RWTUN,
197 &zfs_vdev_max_active, 0,
198 "The maximum number of I/Os of all types active for each device.");
200 #define ZFS_VDEV_QUEUE_KNOB_MIN(name) \
201 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, CTLFLAG_RWTUN,\
202 &zfs_vdev_ ## name ## _min_active, 0, \
203 "Initial number of I/O requests of type " #name \
204 " active for each device");
206 #define ZFS_VDEV_QUEUE_KNOB_MAX(name) \
207 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, CTLFLAG_RWTUN,\
208 &zfs_vdev_ ## name ## _max_active, 0, \
209 "Maximum number of I/O requests of type " #name \
210 " active for each device");
212 ZFS_VDEV_QUEUE_KNOB_MIN(sync_read);
213 ZFS_VDEV_QUEUE_KNOB_MAX(sync_read);
214 ZFS_VDEV_QUEUE_KNOB_MIN(sync_write);
215 ZFS_VDEV_QUEUE_KNOB_MAX(sync_write);
216 ZFS_VDEV_QUEUE_KNOB_MIN(async_read);
217 ZFS_VDEV_QUEUE_KNOB_MAX(async_read);
218 ZFS_VDEV_QUEUE_KNOB_MIN(async_write);
219 ZFS_VDEV_QUEUE_KNOB_MAX(async_write);
220 ZFS_VDEV_QUEUE_KNOB_MIN(scrub);
221 ZFS_VDEV_QUEUE_KNOB_MAX(scrub);
222 ZFS_VDEV_QUEUE_KNOB_MIN(trim);
223 ZFS_VDEV_QUEUE_KNOB_MAX(trim);
225 #undef ZFS_VDEV_QUEUE_KNOB
227 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RWTUN,
228 &zfs_vdev_aggregation_limit, 0,
229 "I/O requests are aggregated up to this size");
230 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RWTUN,
231 &zfs_vdev_read_gap_limit, 0,
232 "Acceptable gap between two reads being aggregated");
233 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RWTUN,
234 &zfs_vdev_write_gap_limit, 0,
235 "Acceptable gap between two writes being aggregated");
238 sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS)
242 val = zfs_vdev_async_write_active_min_dirty_percent;
243 err = sysctl_handle_int(oidp, &val, 0, req);
244 if (err != 0 || req->newptr == NULL)
247 if (val < 0 || val > 100 ||
248 val >= zfs_vdev_async_write_active_max_dirty_percent)
251 zfs_vdev_async_write_active_min_dirty_percent = val;
257 sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS)
261 val = zfs_vdev_async_write_active_max_dirty_percent;
262 err = sysctl_handle_int(oidp, &val, 0, req);
263 if (err != 0 || req->newptr == NULL)
266 if (val < 0 || val > 100 ||
267 val <= zfs_vdev_async_write_active_min_dirty_percent)
270 zfs_vdev_async_write_active_max_dirty_percent = val;
278 vdev_queue_offset_compare(const void *x1, const void *x2)
280 const zio_t *z1 = x1;
281 const zio_t *z2 = x2;
283 if (z1->io_offset < z2->io_offset)
285 if (z1->io_offset > z2->io_offset)
296 static inline avl_tree_t *
297 vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p)
299 return (&vq->vq_class[p].vqc_queued_tree);
302 static inline avl_tree_t *
303 vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t)
305 if (t == ZIO_TYPE_READ)
306 return (&vq->vq_read_offset_tree);
307 else if (t == ZIO_TYPE_WRITE)
308 return (&vq->vq_write_offset_tree);
314 vdev_queue_timestamp_compare(const void *x1, const void *x2)
316 const zio_t *z1 = x1;
317 const zio_t *z2 = x2;
319 if (z1->io_timestamp < z2->io_timestamp)
321 if (z1->io_timestamp > z2->io_timestamp)
324 if (z1->io_offset < z2->io_offset)
326 if (z1->io_offset > z2->io_offset)
338 vdev_queue_init(vdev_t *vd)
340 vdev_queue_t *vq = &vd->vdev_queue;
342 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
345 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
346 sizeof (zio_t), offsetof(struct zio, io_queue_node));
347 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ),
348 vdev_queue_offset_compare, sizeof (zio_t),
349 offsetof(struct zio, io_offset_node));
350 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE),
351 vdev_queue_offset_compare, sizeof (zio_t),
352 offsetof(struct zio, io_offset_node));
354 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
355 int (*compfn) (const void *, const void *);
358 * The synchronous i/o queues are dispatched in FIFO rather
359 * than LBA order. This provides more consistent latency for
362 if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE)
363 compfn = vdev_queue_timestamp_compare;
365 compfn = vdev_queue_offset_compare;
367 avl_create(vdev_queue_class_tree(vq, p), compfn,
368 sizeof (zio_t), offsetof(struct zio, io_queue_node));
371 vq->vq_lastoffset = 0;
375 vdev_queue_fini(vdev_t *vd)
377 vdev_queue_t *vq = &vd->vdev_queue;
379 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
380 avl_destroy(vdev_queue_class_tree(vq, p));
381 avl_destroy(&vq->vq_active_tree);
382 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
383 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE));
385 mutex_destroy(&vq->vq_lock);
389 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
391 spa_t *spa = zio->io_spa;
393 ASSERT(MUTEX_HELD(&vq->vq_lock));
394 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
395 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
396 qtt = vdev_queue_type_tree(vq, zio->io_type);
401 mutex_enter(&spa->spa_iokstat_lock);
402 spa->spa_queue_stats[zio->io_priority].spa_queued++;
403 if (spa->spa_iokstat != NULL)
404 kstat_waitq_enter(spa->spa_iokstat->ks_data);
405 mutex_exit(&spa->spa_iokstat_lock);
410 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
412 spa_t *spa = zio->io_spa;
414 ASSERT(MUTEX_HELD(&vq->vq_lock));
415 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
416 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
417 qtt = vdev_queue_type_tree(vq, zio->io_type);
419 avl_remove(qtt, zio);
422 mutex_enter(&spa->spa_iokstat_lock);
423 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
424 spa->spa_queue_stats[zio->io_priority].spa_queued--;
425 if (spa->spa_iokstat != NULL)
426 kstat_waitq_exit(spa->spa_iokstat->ks_data);
427 mutex_exit(&spa->spa_iokstat_lock);
432 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
434 spa_t *spa = zio->io_spa;
435 ASSERT(MUTEX_HELD(&vq->vq_lock));
436 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
437 vq->vq_class[zio->io_priority].vqc_active++;
438 avl_add(&vq->vq_active_tree, zio);
441 mutex_enter(&spa->spa_iokstat_lock);
442 spa->spa_queue_stats[zio->io_priority].spa_active++;
443 if (spa->spa_iokstat != NULL)
444 kstat_runq_enter(spa->spa_iokstat->ks_data);
445 mutex_exit(&spa->spa_iokstat_lock);
450 vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
452 spa_t *spa = zio->io_spa;
453 ASSERT(MUTEX_HELD(&vq->vq_lock));
454 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
455 vq->vq_class[zio->io_priority].vqc_active--;
456 avl_remove(&vq->vq_active_tree, zio);
459 mutex_enter(&spa->spa_iokstat_lock);
460 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
461 spa->spa_queue_stats[zio->io_priority].spa_active--;
462 if (spa->spa_iokstat != NULL) {
463 kstat_io_t *ksio = spa->spa_iokstat->ks_data;
465 kstat_runq_exit(spa->spa_iokstat->ks_data);
466 if (zio->io_type == ZIO_TYPE_READ) {
468 ksio->nread += zio->io_size;
469 } else if (zio->io_type == ZIO_TYPE_WRITE) {
471 ksio->nwritten += zio->io_size;
474 mutex_exit(&spa->spa_iokstat_lock);
479 vdev_queue_agg_io_done(zio_t *aio)
481 if (aio->io_type == ZIO_TYPE_READ) {
483 while ((pio = zio_walk_parents(aio)) != NULL) {
484 bcopy((char *)aio->io_data + (pio->io_offset -
485 aio->io_offset), pio->io_data, pio->io_size);
489 zio_buf_free(aio->io_data, aio->io_size);
493 vdev_queue_class_min_active(zio_priority_t p)
496 case ZIO_PRIORITY_SYNC_READ:
497 return (zfs_vdev_sync_read_min_active);
498 case ZIO_PRIORITY_SYNC_WRITE:
499 return (zfs_vdev_sync_write_min_active);
500 case ZIO_PRIORITY_ASYNC_READ:
501 return (zfs_vdev_async_read_min_active);
502 case ZIO_PRIORITY_ASYNC_WRITE:
503 return (zfs_vdev_async_write_min_active);
504 case ZIO_PRIORITY_SCRUB:
505 return (zfs_vdev_scrub_min_active);
506 case ZIO_PRIORITY_TRIM:
507 return (zfs_vdev_trim_min_active);
509 panic("invalid priority %u", p);
514 static __noinline int
515 vdev_queue_max_async_writes(spa_t *spa)
518 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total;
519 uint64_t min_bytes = zfs_dirty_data_max *
520 zfs_vdev_async_write_active_min_dirty_percent / 100;
521 uint64_t max_bytes = zfs_dirty_data_max *
522 zfs_vdev_async_write_active_max_dirty_percent / 100;
525 * Sync tasks correspond to interactive user actions. To reduce the
526 * execution time of those actions we push data out as fast as possible.
528 if (spa_has_pending_synctask(spa)) {
529 return (zfs_vdev_async_write_max_active);
532 if (dirty < min_bytes)
533 return (zfs_vdev_async_write_min_active);
534 if (dirty > max_bytes)
535 return (zfs_vdev_async_write_max_active);
538 * linear interpolation:
539 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
540 * move right by min_bytes
541 * move up by min_writes
543 writes = (dirty - min_bytes) *
544 (zfs_vdev_async_write_max_active -
545 zfs_vdev_async_write_min_active) /
546 (max_bytes - min_bytes) +
547 zfs_vdev_async_write_min_active;
548 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
549 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
554 vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
557 case ZIO_PRIORITY_SYNC_READ:
558 return (zfs_vdev_sync_read_max_active);
559 case ZIO_PRIORITY_SYNC_WRITE:
560 return (zfs_vdev_sync_write_max_active);
561 case ZIO_PRIORITY_ASYNC_READ:
562 return (zfs_vdev_async_read_max_active);
563 case ZIO_PRIORITY_ASYNC_WRITE:
564 return (vdev_queue_max_async_writes(spa));
565 case ZIO_PRIORITY_SCRUB:
566 return (zfs_vdev_scrub_max_active);
567 case ZIO_PRIORITY_TRIM:
568 return (zfs_vdev_trim_max_active);
570 panic("invalid priority %u", p);
576 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
577 * there is no eligible class.
579 static zio_priority_t
580 vdev_queue_class_to_issue(vdev_queue_t *vq)
582 spa_t *spa = vq->vq_vdev->vdev_spa;
585 ASSERT(MUTEX_HELD(&vq->vq_lock));
587 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
588 return (ZIO_PRIORITY_NUM_QUEUEABLE);
590 /* find a queue that has not reached its minimum # outstanding i/os */
591 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
592 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
593 vq->vq_class[p].vqc_active <
594 vdev_queue_class_min_active(p))
599 * If we haven't found a queue, look for one that hasn't reached its
600 * maximum # outstanding i/os.
602 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
603 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
604 vq->vq_class[p].vqc_active <
605 vdev_queue_class_max_active(spa, p))
609 /* No eligible queued i/os */
610 return (ZIO_PRIORITY_NUM_QUEUEABLE);
614 * Compute the range spanned by two i/os, which is the endpoint of the last
615 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
616 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
617 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
619 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
620 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
623 vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
625 zio_t *first, *last, *aio, *dio, *mandatory, *nio;
632 ASSERT(MUTEX_HELD(&vq->vq_lock));
634 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
639 if (zio->io_type == ZIO_TYPE_READ)
640 maxgap = zfs_vdev_read_gap_limit;
643 * We can aggregate I/Os that are sufficiently adjacent and of
644 * the same flavor, as expressed by the AGG_INHERIT flags.
645 * The latter requirement is necessary so that certain
646 * attributes of the I/O, such as whether it's a normal I/O
647 * or a scrub/resilver, can be preserved in the aggregate.
648 * We can include optional I/Os, but don't allow them
649 * to begin a range as they add no benefit in that situation.
653 * We keep track of the last non-optional I/O.
655 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
658 * Walk backwards through sufficiently contiguous I/Os
659 * recording the last non-option I/O.
661 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
662 t = vdev_queue_type_tree(vq, zio->io_type);
663 while (t != NULL && (dio = AVL_PREV(t, first)) != NULL &&
664 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
665 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
666 IO_GAP(dio, first) <= maxgap) {
668 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
673 * Skip any initial optional I/Os.
675 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
676 first = AVL_NEXT(t, first);
677 ASSERT(first != NULL);
681 * Walk forward through sufficiently contiguous I/Os.
683 while ((dio = AVL_NEXT(t, last)) != NULL &&
684 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
685 IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit &&
686 IO_GAP(last, dio) <= maxgap) {
688 if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
693 * Now that we've established the range of the I/O aggregation
694 * we must decide what to do with trailing optional I/Os.
695 * For reads, there's nothing to do. While we are unable to
696 * aggregate further, it's possible that a trailing optional
697 * I/O would allow the underlying device to aggregate with
698 * subsequent I/Os. We must therefore determine if the next
699 * non-optional I/O is close enough to make aggregation
703 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
705 while ((dio = AVL_NEXT(t, nio)) != NULL &&
706 IO_GAP(nio, dio) == 0 &&
707 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
709 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
717 /* This may be a no-op. */
718 dio = AVL_NEXT(t, last);
719 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
721 while (last != mandatory && last != first) {
722 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
723 last = AVL_PREV(t, last);
724 ASSERT(last != NULL);
731 size = IO_SPAN(first, last);
732 ASSERT3U(size, <=, zfs_vdev_aggregation_limit);
734 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
735 zio_buf_alloc(size), size, first->io_type, zio->io_priority,
736 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
737 vdev_queue_agg_io_done, NULL);
738 aio->io_timestamp = first->io_timestamp;
743 nio = AVL_NEXT(t, dio);
744 ASSERT3U(dio->io_type, ==, aio->io_type);
746 if (dio->io_flags & ZIO_FLAG_NODATA) {
747 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
748 bzero((char *)aio->io_data + (dio->io_offset -
749 aio->io_offset), dio->io_size);
750 } else if (dio->io_type == ZIO_TYPE_WRITE) {
751 bcopy(dio->io_data, (char *)aio->io_data +
752 (dio->io_offset - aio->io_offset),
756 zio_add_child(dio, aio);
757 vdev_queue_io_remove(vq, dio);
758 zio_vdev_io_bypass(dio);
760 } while (dio != last);
766 vdev_queue_io_to_issue(vdev_queue_t *vq)
775 ASSERT(MUTEX_HELD(&vq->vq_lock));
777 p = vdev_queue_class_to_issue(vq);
779 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
780 /* No eligible queued i/os */
785 * For LBA-ordered queues (async / scrub), issue the i/o which follows
786 * the most recently issued i/o in LBA (offset) order.
788 * For FIFO queues (sync), issue the i/o with the lowest timestamp.
790 tree = vdev_queue_class_tree(vq, p);
791 search.io_timestamp = 0;
792 search.io_offset = vq->vq_last_offset + 1;
793 VERIFY3P(avl_find(tree, &search, &idx), ==, NULL);
794 zio = avl_nearest(tree, idx, AVL_AFTER);
796 zio = avl_first(tree);
797 ASSERT3U(zio->io_priority, ==, p);
799 aio = vdev_queue_aggregate(vq, zio);
803 vdev_queue_io_remove(vq, zio);
806 * If the I/O is or was optional and therefore has no data, we need to
807 * simply discard it. We need to drop the vdev queue's lock to avoid a
808 * deadlock that we could encounter since this I/O will complete
811 if (zio->io_flags & ZIO_FLAG_NODATA) {
812 mutex_exit(&vq->vq_lock);
813 zio_vdev_io_bypass(zio);
815 mutex_enter(&vq->vq_lock);
819 vdev_queue_pending_add(vq, zio);
820 vq->vq_last_offset = zio->io_offset;
826 vdev_queue_io(zio_t *zio)
828 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
831 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
835 * Children i/os inherent their parent's priority, which might
836 * not match the child's i/o type. Fix it up here.
838 if (zio->io_type == ZIO_TYPE_READ) {
839 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
840 zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
841 zio->io_priority != ZIO_PRIORITY_SCRUB)
842 zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
843 } else if (zio->io_type == ZIO_TYPE_WRITE) {
844 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
845 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE)
846 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
848 ASSERT(zio->io_type == ZIO_TYPE_FREE);
849 zio->io_priority = ZIO_PRIORITY_TRIM;
852 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
854 mutex_enter(&vq->vq_lock);
855 zio->io_timestamp = gethrtime();
856 vdev_queue_io_add(vq, zio);
857 nio = vdev_queue_io_to_issue(vq);
858 mutex_exit(&vq->vq_lock);
863 if (nio->io_done == vdev_queue_agg_io_done) {
872 vdev_queue_io_done(zio_t *zio)
874 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
877 mutex_enter(&vq->vq_lock);
879 vdev_queue_pending_remove(vq, zio);
881 vq->vq_io_complete_ts = gethrtime();
883 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
884 mutex_exit(&vq->vq_lock);
885 if (nio->io_done == vdev_queue_agg_io_done) {
888 zio_vdev_io_reissue(nio);
891 mutex_enter(&vq->vq_lock);
894 mutex_exit(&vq->vq_lock);
898 * As these three methods are only used for load calculations we're not concerned
899 * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex
900 * use here, instead we prefer to keep it lock free for performance.
903 vdev_queue_length(vdev_t *vd)
905 return (avl_numnodes(&vd->vdev_queue.vq_active_tree));
909 vdev_queue_lastoffset(vdev_t *vd)
911 return (vd->vdev_queue.vq_lastoffset);
915 vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio)
917 vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size;