1fa9e406ahrens/*
2fa9e406ahrens * CDDL HEADER START
3fa9e406ahrens *
4fa9e406ahrens * The contents of this file are subject to the terms of the
5ea8dc4beschrock * Common Development and Distribution License (the "License").
6ea8dc4beschrock * You may not use this file except in compliance with the License.
7fa9e406ahrens *
8fa9e406ahrens * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9fa9e406ahrens * or http://www.opensolaris.org/os/licensing.
10fa9e406ahrens * See the License for the specific language governing permissions
11fa9e406ahrens * and limitations under the License.
12fa9e406ahrens *
13fa9e406ahrens * When distributing Covered Code, include this CDDL HEADER in each
14fa9e406ahrens * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15fa9e406ahrens * If applicable, add the following below this CDDL HEADER, with the
16fa9e406ahrens * fields enclosed by brackets "[]" replaced with your own identifying
17fa9e406ahrens * information: Portions Copyright [yyyy] [name of copyright owner]
18fa9e406ahrens *
19fa9e406ahrens * CDDL HEADER END
20fa9e406ahrens */
21fa9e406ahrens/*
22a3f829aBill Moore * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23fa9e406ahrens * Use is subject to license terms.
24fa9e406ahrens */
25fa9e406ahrens
26283b846George.Wilson/*
27f78cdc3Paul Dagnelie * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
28c3d26abMatthew Ahrens * Copyright (c) 2014 Integros [integros.com]
2944bf619John Levon * Copyright 2019 Joyent, Inc.
30283b846George.Wilson */
31283b846George.Wilson
32fa9e406ahrens#include <sys/zfs_context.h>
33fa9e406ahrens#include <sys/vdev_impl.h>
34c3a6601Matthew Ahrens#include <sys/spa_impl.h>
35fa9e406ahrens#include <sys/zio.h>
36fa9e406ahrens#include <sys/avl.h>
3769962b5Matthew Ahrens#include <sys/dsl_pool.h>
380f7643cGeorge Wilson#include <sys/metaslab_impl.h>
39770499eDan Kimmel#include <sys/abd.h>
40fa9e406ahrens
41fa9e406ahrens/*
4269962b5Matthew Ahrens * ZFS I/O Scheduler
4369962b5Matthew Ahrens * ---------------
4469962b5Matthew Ahrens *
4569962b5Matthew Ahrens * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios.  The
4669962b5Matthew Ahrens * I/O scheduler determines when and in what order those operations are
4769962b5Matthew Ahrens * issued.  The I/O scheduler divides operations into five I/O classes
4869962b5Matthew Ahrens * prioritized in the following order: sync read, sync write, async read,
4969962b5Matthew Ahrens * async write, and scrub/resilver.  Each queue defines the minimum and
5069962b5Matthew Ahrens * maximum number of concurrent operations that may be issued to the device.
5169962b5Matthew Ahrens * In addition, the device has an aggregate maximum. Note that the sum of the
5269962b5Matthew Ahrens * per-queue minimums must not exceed the aggregate maximum, and if the
5369962b5Matthew Ahrens * aggregate maximum is equal to or greater than the sum of the per-queue
5469962b5Matthew Ahrens * maximums, the per-queue minimum has no effect.
5569962b5Matthew Ahrens *
5669962b5Matthew Ahrens * For many physical devices, throughput increases with the number of
5769962b5Matthew Ahrens * concurrent operations, but latency typically suffers. Further, physical
5869962b5Matthew Ahrens * devices typically have a limit at which more concurrent operations have no
5969962b5Matthew Ahrens * effect on throughput or can actually cause it to decrease.
6069962b5Matthew Ahrens *
6169962b5Matthew Ahrens * The scheduler selects the next operation to issue by first looking for an
6269962b5Matthew Ahrens * I/O class whose minimum has not been satisfied. Once all are satisfied and
6369962b5Matthew Ahrens * the aggregate maximum has not been hit, the scheduler looks for classes
6469962b5Matthew Ahrens * whose maximum has not been satisfied. Iteration through the I/O classes is
6569962b5Matthew Ahrens * done in the order specified above. No further operations are issued if the
6669962b5Matthew Ahrens * aggregate maximum number of concurrent operations has been hit or if there
6769962b5Matthew Ahrens * are no operations queued for an I/O class that has not hit its maximum.
6869962b5Matthew Ahrens * Every time an i/o is queued or an operation completes, the I/O scheduler
6969962b5Matthew Ahrens * looks for new operations to issue.
7069962b5Matthew Ahrens *
7169962b5Matthew Ahrens * All I/O classes have a fixed maximum number of outstanding operations
7269962b5Matthew Ahrens * except for the async write class. Asynchronous writes represent the data
7369962b5Matthew Ahrens * that is committed to stable storage during the syncing stage for
7469962b5Matthew Ahrens * transaction groups (see txg.c). Transaction groups enter the syncing state
7569962b5Matthew Ahrens * periodically so the number of queued async writes will quickly burst up and
7669962b5Matthew Ahrens * then bleed down to zero. Rather than servicing them as quickly as possible,
7769962b5Matthew Ahrens * the I/O scheduler changes the maximum number of active async write i/os
7869962b5Matthew Ahrens * according to the amount of dirty data in the pool (see dsl_pool.c). Since
7969962b5Matthew Ahrens * both throughput and latency typically increase with the number of
8069962b5Matthew Ahrens * concurrent operations issued to physical devices, reducing the burstiness
8169962b5Matthew Ahrens * in the number of concurrent operations also stabilizes the response time of
8269962b5Matthew Ahrens * operations from other -- and in particular synchronous -- queues. In broad
8369962b5Matthew Ahrens * strokes, the I/O scheduler will issue more concurrent operations from the
8469962b5Matthew Ahrens * async write queue as there's more dirty data in the pool.
8569962b5Matthew Ahrens *
8669962b5Matthew Ahrens * Async Writes
8769962b5Matthew Ahrens *
8869962b5Matthew Ahrens * The number of concurrent operations issued for the async write I/O class
8969962b5Matthew Ahrens * follows a piece-wise linear function defined by a few adjustable points.
9069962b5Matthew Ahrens *
9169962b5Matthew Ahrens *        |                   o---------| <-- zfs_vdev_async_write_max_active
9269962b5Matthew Ahrens *   ^    |                  /^         |
9369962b5Matthew Ahrens *   |    |                 / |         |
9469962b5Matthew Ahrens * active |                /  |         |
9569962b5Matthew Ahrens *  I/O   |               /   |         |
9669962b5Matthew Ahrens * count  |              /    |         |
9769962b5Matthew Ahrens *        |             /     |         |
9869962b5Matthew Ahrens *        |------------o      |         | <-- zfs_vdev_async_write_min_active
9969962b5Matthew Ahrens *       0|____________^______|_________|
10069962b5Matthew Ahrens *        0%           |      |       100% of zfs_dirty_data_max
10169962b5Matthew Ahrens *                     |      |
10269962b5Matthew Ahrens *                     |      `-- zfs_vdev_async_write_active_max_dirty_percent
10369962b5Matthew Ahrens *                     `--------- zfs_vdev_async_write_active_min_dirty_percent
10469962b5Matthew Ahrens *
10569962b5Matthew Ahrens * Until the amount of dirty data exceeds a minimum percentage of the dirty
10669962b5Matthew Ahrens * data allowed in the pool, the I/O scheduler will limit the number of
10769962b5Matthew Ahrens * concurrent operations to the minimum. As that threshold is crossed, the
10869962b5Matthew Ahrens * number of concurrent operations issued increases linearly to the maximum at
10969962b5Matthew Ahrens * the specified maximum percentage of the dirty data allowed in the pool.
11069962b5Matthew Ahrens *
11169962b5Matthew Ahrens * Ideally, the amount of dirty data on a busy pool will stay in the sloped
11269962b5Matthew Ahrens * part of the function between zfs_vdev_async_write_active_min_dirty_percent
11369962b5Matthew Ahrens * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the
11469962b5Matthew Ahrens * maximum percentage, this indicates that the rate of incoming data is
11569962b5Matthew Ahrens * greater than the rate that the backend storage can handle. In this case, we
11669962b5Matthew Ahrens * must further throttle incoming writes (see dmu_tx_delay() for details).
117614409bahrens */
118f717074Will Andrews
119614409bahrens/*
12069962b5Matthew Ahrens * The maximum number of i/os active to each device.  Ideally, this will be >=
12169962b5Matthew Ahrens * the sum of each queue's max_active.  It must be at least the sum of each
12269962b5Matthew Ahrens * queue's min_active.
123614409bahrens */
12469962b5Matthew Ahrensuint32_t zfs_vdev_max_active = 1000;
125614409bahrens
126c55e05cMatthew Ahrens/*
12769962b5Matthew Ahrens * Per-queue limits on the number of i/os active to each device.  If the
12869962b5Matthew Ahrens * sum of the queue's max_active is < zfs_vdev_max_active, then the
12969962b5Matthew Ahrens * min_active comes into play.  We will send min_active from each queue,
13069962b5Matthew Ahrens * and then select from queues in the order defined by zio_priority_t.
13169962b5Matthew Ahrens *
13269962b5Matthew Ahrens * In general, smaller max_active's will lead to lower latency of synchronous
13369962b5Matthew Ahrens * operations.  Larger max_active's may lead to higher overall throughput,
13469962b5Matthew Ahrens * depending on underlying storage.
13569962b5Matthew Ahrens *
13669962b5Matthew Ahrens * The ratio of the queues' max_actives determines the balance of performance
13769962b5Matthew Ahrens * between reads, writes, and scrubs.  E.g., increasing
13869962b5Matthew Ahrens * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete
13969962b5Matthew Ahrens * more quickly, but reads and writes to have higher latency and lower
14069962b5Matthew Ahrens * throughput.
141c55e05cMatthew Ahrens */
14269962b5Matthew Ahrensuint32_t zfs_vdev_sync_read_min_active = 10;
14369962b5Matthew Ahrensuint32_t zfs_vdev_sync_read_max_active = 10;
14469962b5Matthew Ahrensuint32_t zfs_vdev_sync_write_min_active = 10;
14569962b5Matthew Ahrensuint32_t zfs_vdev_sync_write_max_active = 10;
14669962b5Matthew Ahrensuint32_t zfs_vdev_async_read_min_active = 1;
14769962b5Matthew Ahrensuint32_t zfs_vdev_async_read_max_active = 3;
14869962b5Matthew Ahrensuint32_t zfs_vdev_async_write_min_active = 1;
14969962b5Matthew Ahrensuint32_t zfs_vdev_async_write_max_active = 10;
15069962b5Matthew Ahrensuint32_t zfs_vdev_scrub_min_active = 1;
15169962b5Matthew Ahrensuint32_t zfs_vdev_scrub_max_active = 2;
1525cabbc6Prashanth Sreenivasauint32_t zfs_vdev_removal_min_active = 1;
1535cabbc6Prashanth Sreenivasauint32_t zfs_vdev_removal_max_active = 2;
154094e47eGeorge Wilsonuint32_t zfs_vdev_initializing_min_active = 1;
155094e47eGeorge Wilsonuint32_t zfs_vdev_initializing_max_active = 1;
156084fd14Brian Behlendorfuint32_t zfs_vdev_trim_min_active = 1;
157084fd14Brian Behlendorfuint32_t zfs_vdev_trim_max_active = 2;
158614409bahrens
15969962b5Matthew Ahrens/*
16069962b5Matthew Ahrens * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent
16169962b5Matthew Ahrens * dirty data, use zfs_vdev_async_write_min_active.  When it has more than
16269962b5Matthew Ahrens * zfs_vdev_async_write_active_max_dirty_percent, use
16369962b5Matthew Ahrens * zfs_vdev_async_write_max_active. The value is linearly interpolated
16469962b5Matthew Ahrens * between min and max.
16569962b5Matthew Ahrens */
16669962b5Matthew Ahrensint zfs_vdev_async_write_active_min_dirty_percent = 30;
16769962b5Matthew Ahrensint zfs_vdev_async_write_active_max_dirty_percent = 60;
168614409bahrens
169614409bahrens/*
170f94275cAdam Leventhal * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
171f94275cAdam Leventhal * For read I/Os, we also aggregate across small adjacency gaps; for writes
172f94275cAdam Leventhal * we include spans of optional I/Os to aid aggregation at the disk even when
173f94275cAdam Leventhal * they aren't able to help us aggregate at this level.
174614409bahrens */
175a3874b8Toomas Soomeint zfs_vdev_aggregation_limit = 1 << 20;
1766f708f7Jeff Bonwickint zfs_vdev_read_gap_limit = 32 << 10;
177f94275cAdam Leventhalint zfs_vdev_write_gap_limit = 4 << 10;
178614409bahrens
1790f7643cGeorge Wilson/*
1800f7643cGeorge Wilson * Define the queue depth percentage for each top-level. This percentage is
1810f7643cGeorge Wilson * used in conjunction with zfs_vdev_async_max_active to determine how many
1820f7643cGeorge Wilson * allocations a specific top-level vdev should handle. Once the queue depth
1830f7643cGeorge Wilson * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100
1840f7643cGeorge Wilson * then allocator will stop allocating blocks on that top-level device.
1850f7643cGeorge Wilson * The default kernel setting is 1000% which will yield 100 allocations per
1860f7643cGeorge Wilson * device. For userland testing, the default setting is 300% which equates
1870f7643cGeorge Wilson * to 30 allocations per device.
1880f7643cGeorge Wilson */
1890f7643cGeorge Wilson#ifdef _KERNEL
1900f7643cGeorge Wilsonint zfs_vdev_queue_depth_pct = 1000;
1910f7643cGeorge Wilson#else
1920f7643cGeorge Wilsonint zfs_vdev_queue_depth_pct = 300;
1930f7643cGeorge Wilson#endif
1940f7643cGeorge Wilson
195f78cdc3Paul Dagnelie/*
196f78cdc3Paul Dagnelie * When performing allocations for a given metaslab, we want to make sure that
197f78cdc3Paul Dagnelie * there are enough IOs to aggregate together to improve throughput. We want to
198f78cdc3Paul Dagnelie * ensure that there are at least 128k worth of IOs that can be aggregated, and
199f78cdc3Paul Dagnelie * we assume that the average allocation size is 4k, so we need the queue depth
200f78cdc3Paul Dagnelie * to be 32 per allocator to get good aggregation of sequential writes.
201f78cdc3Paul Dagnelie */
202f78cdc3Paul Dagnelieint zfs_vdev_def_queue_depth = 32;
203f78cdc3Paul Dagnelie
204084fd14Brian Behlendorf/*
205084fd14Brian Behlendorf * Allow TRIM I/Os to be aggregated.  This should normally not be needed since
206084fd14Brian Behlendorf * TRIM I/O for extents up to zfs_trim_extent_bytes_max (128M) can be submitted
207084fd14Brian Behlendorf * by the TRIM code in zfs_trim.c.
208084fd14Brian Behlendorf */
209084fd14Brian Behlendorfint zfs_vdev_aggregate_trim = 0;
2100f7643cGeorge Wilson
211fa9e406ahrensint
21269962b5Matthew Ahrensvdev_queue_offset_compare(const void *x1, const void *x2)
213fa9e406ahrens{
214c4ab0d3Gvozden Neskovic	const zio_t *z1 = (const zio_t *)x1;
215c4ab0d3Gvozden Neskovic	const zio_t *z2 = (const zio_t *)x2;
216fa9e406ahrens
2174d7988dPaul Dagnelie	int cmp = TREE_CMP(z1->io_offset, z2->io_offset);
218fa9e406ahrens
219c4ab0d3Gvozden Neskovic	if (likely(cmp))
220c4ab0d3Gvozden Neskovic		return (cmp);
221fa9e406ahrens
2224d7988dPaul Dagnelie	return (TREE_PCMP(z1, z2));
223fa9e406ahrens}
224fa9e406ahrens
225fe31923Justin T. Gibbsstatic inline avl_tree_t *
226fe31923Justin T. Gibbsvdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p)
227fe31923Justin T. Gibbs{
228fe31923Justin T. Gibbs	return (&vq->vq_class[p].vqc_queued_tree);
229fe31923Justin T. Gibbs}
230fe31923Justin T. Gibbs
231fe31923Justin T. Gibbsstatic inline avl_tree_t *
232fe31923Justin T. Gibbsvdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t)
233fe31923Justin T. Gibbs{
234084fd14Brian Behlendorf	ASSERT(t == ZIO_TYPE_READ || t == ZIO_TYPE_WRITE || t == ZIO_TYPE_TRIM);
235fe31923Justin T. Gibbs	if (t == ZIO_TYPE_READ)
236fe31923Justin T. Gibbs		return (&vq->vq_read_offset_tree);
237084fd14Brian Behlendorf	else if (t == ZIO_TYPE_WRITE)
238fe31923Justin T. Gibbs		return (&vq->vq_write_offset_tree);
239084fd14Brian Behlendorf	else
240084fd14Brian Behlendorf		return (&vq->vq_trim_offset_tree);
241fe31923Justin T. Gibbs}
242fe31923Justin T. Gibbs
243fa9e406ahrensint
24469962b5Matthew Ahrensvdev_queue_timestamp_compare(const void *x1, const void *x2)
245fa9e406ahrens{
246c4ab0d3Gvozden Neskovic	const zio_t *z1 = (const zio_t *)x1;
247c4ab0d3Gvozden Neskovic	const zio_t *z2 = (const zio_t *)x2;
248fa9e406ahrens
2494d7988dPaul Dagnelie	int cmp = TREE_CMP(z1->io_timestamp, z2->io_timestamp);
250fa9e406ahrens
251c4ab0d3Gvozden Neskovic	if (likely(cmp))
252c4ab0d3Gvozden Neskovic		return (cmp);
253fa9e406ahrens
2544d7988dPaul Dagnelie	return (TREE_PCMP(z1, z2));
255fa9e406ahrens}
256fa9e406ahrens
257fa9e406ahrensvoid
258fa9e406ahrensvdev_queue_init(vdev_t *vd)
259fa9e406ahrens{
260fa9e406ahrens	vdev_queue_t *vq = &vd->vdev_queue;
261fa9e406ahrens
262fa9e406ahrens	mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
26369962b5Matthew Ahrens	vq->vq_vdev = vd;
264fa9e406ahrens
26569962b5Matthew Ahrens	avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
26669962b5Matthew Ahrens	    sizeof (zio_t), offsetof(struct zio, io_queue_node));
267fe31923Justin T. Gibbs	avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ),
268fe31923Justin T. Gibbs	    vdev_queue_offset_compare, sizeof (zio_t),
269fe31923Justin T. Gibbs	    offsetof(struct zio, io_offset_node));
270fe31923Justin T. Gibbs	avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE),
271fe31923Justin T. Gibbs	    vdev_queue_offset_compare, sizeof (zio_t),
272fe31923Justin T. Gibbs	    offsetof(struct zio, io_offset_node));
273084fd14Brian Behlendorf	avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_TRIM),
274084fd14Brian Behlendorf	    vdev_queue_offset_compare, sizeof (zio_t),
275084fd14Brian Behlendorf	    offsetof(struct zio, io_offset_node));
276fa9e406ahrens
27769962b5Matthew Ahrens	for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
278fe31923Justin T. Gibbs		int (*compfn) (const void *, const void *);
279fe31923Justin T. Gibbs
28069962b5Matthew Ahrens		/*
281084fd14Brian Behlendorf		 * The synchronous/trim i/o queues are dispatched in FIFO rather
282084fd14Brian Behlendorf		 * than LBA order. This provides more consistent latency for
283fe31923Justin T. Gibbs		 * these i/os.
28469962b5Matthew Ahrens		 */
285084fd14Brian Behlendorf		if (p == ZIO_PRIORITY_SYNC_READ ||
286084fd14Brian Behlendorf		    p == ZIO_PRIORITY_SYNC_WRITE ||
287084fd14Brian Behlendorf		    p == ZIO_PRIORITY_TRIM) {
288fe31923Justin T. Gibbs			compfn = vdev_queue_timestamp_compare;
289084fd14Brian Behlendorf		} else {
290fe31923Justin T. Gibbs			compfn = vdev_queue_offset_compare;
291084fd14Brian Behlendorf		}
292fe31923Justin T. Gibbs
293fe31923Justin T. Gibbs		avl_create(vdev_queue_class_tree(vq, p), compfn,
29469962b5Matthew Ahrens		    sizeof (zio_t), offsetof(struct zio, io_queue_node));
29569962b5Matthew Ahrens	}
29612a8814Tom Caputi
29712a8814Tom Caputi	vq->vq_last_offset = 0;
298fa9e406ahrens}
299fa9e406ahrens
300fa9e406ahrensvoid
301fa9e406ahrensvdev_queue_fini(vdev_t *vd)
302fa9e406ahrens{
303fa9e406ahrens	vdev_queue_t *vq = &vd->vdev_queue;
304fa9e406ahrens
30569962b5Matthew Ahrens	for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
306fe31923Justin T. Gibbs		avl_destroy(vdev_queue_class_tree(vq, p));
30769962b5Matthew Ahrens	avl_destroy(&vq->vq_active_tree);
308fe31923Justin T. Gibbs	avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
309fe31923Justin T. Gibbs	avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE));
310084fd14Brian Behlendorf	avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_TRIM));
311fa9e406ahrens
312fa9e406ahrens	mutex_destroy(&vq->vq_lock);
313fa9e406ahrens}
314fa9e406ahrens
315fa9e406ahrensstatic void
316ea8dc4beschrockvdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
317ea8dc4beschrock{
318c3a6601Matthew Ahrens	spa_t *spa = zio->io_spa;
3190f7643cGeorge Wilson
32069962b5Matthew Ahrens	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
321fe31923Justin T. Gibbs	avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
322fe31923Justin T. Gibbs	avl_add(vdev_queue_type_tree(vq, zio->io_type), zio);
323c3a6601Matthew Ahrens
32469962b5Matthew Ahrens	mutex_enter(&spa->spa_iokstat_lock);
32569962b5Matthew Ahrens	spa->spa_queue_stats[zio->io_priority].spa_queued++;
32669962b5Matthew Ahrens	if (spa->spa_iokstat != NULL)
327c3a6601Matthew Ahrens		kstat_waitq_enter(spa->spa_iokstat->ks_data);
32869962b5Matthew Ahrens	mutex_exit(&spa->spa_iokstat_lock);
329ea8dc4beschrock}
330ea8dc4beschrock
331ea8dc4beschrockstatic void
332ea8dc4beschrockvdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
333ea8dc4beschrock{
334c3a6601Matthew Ahrens	spa_t *spa = zio->io_spa;
3350f7643cGeorge Wilson
33669962b5Matthew Ahrens	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
337fe31923Justin T. Gibbs	avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
338fe31923Justin T. Gibbs	avl_remove(vdev_queue_type_tree(vq, zio->io_type), zio);
339c3a6601Matthew Ahrens
34069962b5Matthew Ahrens	mutex_enter(&spa->spa_iokstat_lock);
34169962b5Matthew Ahrens	ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
34269962b5Matthew Ahrens	spa->spa_queue_stats[zio->io_priority].spa_queued--;
34369962b5Matthew Ahrens	if (spa->spa_iokstat != NULL)
344c3a6601Matthew Ahrens		kstat_waitq_exit(spa->spa_iokstat->ks_data);
34569962b5Matthew Ahrens	mutex_exit(&spa->spa_iokstat_lock);
346c3a6601Matthew Ahrens}
347c3a6601Matthew Ahrens
348c3a6601Matthew Ahrensstatic void
349c3a6601Matthew Ahrensvdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
350c3a6601Matthew Ahrens{
351c3a6601Matthew Ahrens	spa_t *spa = zio->io_spa;
35269962b5Matthew Ahrens	ASSERT(MUTEX_HELD(&vq->vq_lock));
35369962b5Matthew Ahrens	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
35469962b5Matthew Ahrens	vq->vq_class[zio->io_priority].vqc_active++;
35569962b5Matthew Ahrens	avl_add(&vq->vq_active_tree, zio);
35669962b5Matthew Ahrens
35769962b5Matthew Ahrens	mutex_enter(&spa->spa_iokstat_lock);
35869962b5Matthew Ahrens	spa->spa_queue_stats[zio->io_priority].spa_active++;
35969962b5Matthew Ahrens	if (spa->spa_iokstat != NULL)
360c3a6601Matthew Ahrens		kstat_runq_enter(spa->spa_iokstat->ks_data);
36169962b5Matthew Ahrens	mutex_exit(&spa->spa_iokstat_lock);
362c3a6601Matthew Ahrens}
363c3a6601Matthew Ahrens
364c3a6601Matthew Ahrensstatic void
365c3a6601Matthew Ahrensvdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
366c3a6601Matthew Ahrens{
367c3a6601Matthew Ahrens	spa_t *spa = zio->io_spa;
36869962b5Matthew Ahrens	ASSERT(MUTEX_HELD(&vq->vq_lock));
36969962b5Matthew Ahrens	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
37069962b5Matthew Ahrens	vq->vq_class[zio->io_priority].vqc_active--;
37169962b5Matthew Ahrens	avl_remove(&vq->vq_active_tree, zio);
37269962b5Matthew Ahrens
37369962b5Matthew Ahrens	mutex_enter(&spa->spa_iokstat_lock);
37469962b5Matthew Ahrens	ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
37569962b5Matthew Ahrens	spa->spa_queue_stats[zio->io_priority].spa_active--;
376c3a6601Matthew Ahrens	if (spa->spa_iokstat != NULL) {
377c3a6601Matthew Ahrens		kstat_io_t *ksio = spa->spa_iokstat->ks_data;
378c3a6601Matthew Ahrens
379c3a6601Matthew Ahrens		kstat_runq_exit(spa->spa_iokstat->ks_data);
380c3a6601Matthew Ahrens		if (zio->io_type == ZIO_TYPE_READ) {
381c3a6601Matthew Ahrens			ksio->reads++;
382c3a6601Matthew Ahrens			ksio->nread += zio->io_size;
383c3a6601Matthew Ahrens		} else if (zio->io_type == ZIO_TYPE_WRITE) {
384c3a6601Matthew Ahrens			ksio->writes++;
385c3a6601Matthew Ahrens			ksio->nwritten += zio->io_size;
386c3a6601Matthew Ahrens		}
387c3a6601Matthew Ahrens	}
38869962b5Matthew Ahrens	mutex_exit(&spa->spa_iokstat_lock);
389ea8dc4beschrock}
390ea8dc4beschrock
391ea8dc4beschrockstatic void
392fa9e406ahrensvdev_queue_agg_io_done(zio_t *aio)
393fa9e406ahrens{
39469962b5Matthew Ahrens	if (aio->io_type == ZIO_TYPE_READ) {
39569962b5Matthew Ahrens		zio_t *pio;
3960f7643cGeorge Wilson		zio_link_t *zl = NULL;
3970f7643cGeorge Wilson		while ((pio = zio_walk_parents(aio, &zl)) != NULL) {
398770499eDan Kimmel			abd_copy_off(pio->io_abd, aio->io_abd,
399770499eDan Kimmel			    0, pio->io_offset - aio->io_offset, pio->io_size);
40069962b5Matthew Ahrens		}
40169962b5Matthew Ahrens	}
402fa9e406ahrens
403770499eDan Kimmel	abd_free(aio->io_abd);
404fa9e406ahrens}
405fa9e406ahrens
40669962b5Matthew Ahrensstatic int
40769962b5Matthew Ahrensvdev_queue_class_min_active(zio_priority_t p)
40869962b5Matthew Ahrens{
40969962b5Matthew Ahrens	switch (p) {
41069962b5Matthew Ahrens	case ZIO_PRIORITY_SYNC_READ:
41169962b5Matthew Ahrens		return (zfs_vdev_sync_read_min_active);
41269962b5Matthew Ahrens	case ZIO_PRIORITY_SYNC_WRITE:
41369962b5Matthew Ahrens		return (zfs_vdev_sync_write_min_active);
41469962b5Matthew Ahrens	case ZIO_PRIORITY_ASYNC_READ:
41569962b5Matthew Ahrens		return (zfs_vdev_async_read_min_active);
41669962b5Matthew Ahrens	case ZIO_PRIORITY_ASYNC_WRITE:
41769962b5Matthew Ahrens		return (zfs_vdev_async_write_min_active);
41869962b5Matthew Ahrens	case ZIO_PRIORITY_SCRUB:
41969962b5Matthew Ahrens		return (zfs_vdev_scrub_min_active);
4205cabbc6Prashanth Sreenivasa	case ZIO_PRIORITY_REMOVAL:
4215cabbc6Prashanth Sreenivasa		return (zfs_vdev_removal_min_active);
422094e47eGeorge Wilson	case ZIO_PRIORITY_INITIALIZING:
423094e47eGeorge Wilson		return (zfs_vdev_initializing_min_active);
424084fd14Brian Behlendorf	case ZIO_PRIORITY_TRIM:
425084fd14Brian Behlendorf		return (zfs_vdev_trim_min_active);
42669962b5Matthew Ahrens	default:
42769962b5Matthew Ahrens		panic("invalid priority %u", p);
42869962b5Matthew Ahrens	}
42969962b5Matthew Ahrens}
43069962b5Matthew Ahrens
43169962b5Matthew Ahrensstatic int
43273527f4Alex Reecevdev_queue_max_async_writes(spa_t *spa)
43369962b5Matthew Ahrens{
43469962b5Matthew Ahrens	int writes;
43573527f4Alex Reece	uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total;
43669962b5Matthew Ahrens	uint64_t min_bytes = zfs_dirty_data_max *
43769962b5Matthew Ahrens	    zfs_vdev_async_write_active_min_dirty_percent / 100;
43869962b5Matthew Ahrens	uint64_t max_bytes = zfs_dirty_data_max *
43969962b5Matthew Ahrens	    zfs_vdev_async_write_active_max_dirty_percent / 100;
44069962b5Matthew Ahrens
44173527f4Alex Reece	/*
44273527f4Alex Reece	 * Sync tasks correspond to interactive user actions. To reduce the
44373527f4Alex Reece	 * execution time of those actions we push data out as fast as possible.
44473527f4Alex Reece	 */
44573527f4Alex Reece	if (spa_has_pending_synctask(spa)) {
44673527f4Alex Reece		return (zfs_vdev_async_write_max_active);
44773527f4Alex Reece	}
44873527f4Alex Reece
44969962b5Matthew Ahrens	if (dirty < min_bytes)
45069962b5Matthew Ahrens		return (zfs_vdev_async_write_min_active);
45169962b5Matthew Ahrens	if (dirty > max_bytes)
45269962b5Matthew Ahrens		return (zfs_vdev_async_write_max_active);
45369962b5Matthew Ahrens
45469962b5Matthew Ahrens	/*
45569962b5Matthew Ahrens	 * linear interpolation:
45669962b5Matthew Ahrens	 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
45769962b5Matthew Ahrens	 * move right by min_bytes
45869962b5Matthew Ahrens	 * move up by min_writes
45969962b5Matthew Ahrens	 */
46069962b5Matthew Ahrens	writes = (dirty - min_bytes) *
46169962b5Matthew Ahrens	    (zfs_vdev_async_write_max_active -
46269962b5Matthew Ahrens	    zfs_vdev_async_write_min_active) /
46369962b5Matthew Ahrens	    (max_bytes - min_bytes) +
46469962b5Matthew Ahrens	    zfs_vdev_async_write_min_active;
46569962b5Matthew Ahrens	ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
46669962b5Matthew Ahrens	ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
46769962b5Matthew Ahrens	return (writes);
46869962b5Matthew Ahrens}
46969962b5Matthew Ahrens
47069962b5Matthew Ahrensstatic int
47169962b5Matthew Ahrensvdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
47269962b5Matthew Ahrens{
47369962b5Matthew Ahrens	switch (p) {
47469962b5Matthew Ahrens	case ZIO_PRIORITY_SYNC_READ:
47569962b5Matthew Ahrens		return (zfs_vdev_sync_read_max_active);
47669962b5Matthew Ahrens	case ZIO_PRIORITY_SYNC_WRITE:
47769962b5Matthew Ahrens		return (zfs_vdev_sync_write_max_active);
47869962b5Matthew Ahrens	case ZIO_PRIORITY_ASYNC_READ:
47969962b5Matthew Ahrens		return (zfs_vdev_async_read_max_active);
48069962b5Matthew Ahrens	case ZIO_PRIORITY_ASYNC_WRITE:
48173527f4Alex Reece		return (vdev_queue_max_async_writes(spa));
48269962b5Matthew Ahrens	case ZIO_PRIORITY_SCRUB:
48369962b5Matthew Ahrens		return (zfs_vdev_scrub_max_active);
4845cabbc6Prashanth Sreenivasa	case ZIO_PRIORITY_REMOVAL:
4855cabbc6Prashanth Sreenivasa		return (zfs_vdev_removal_max_active);
486094e47eGeorge Wilson	case ZIO_PRIORITY_INITIALIZING:
487094e47eGeorge Wilson		return (zfs_vdev_initializing_max_active);
488084fd14Brian Behlendorf	case ZIO_PRIORITY_TRIM:
489084fd14Brian Behlendorf		return (zfs_vdev_trim_max_active);
49069962b5Matthew Ahrens	default:
49169962b5Matthew Ahrens		panic("invalid priority %u", p);
49269962b5Matthew Ahrens	}
49369962b5Matthew Ahrens}
49469962b5Matthew Ahrens
49569962b5Matthew Ahrens/*
49669962b5Matthew Ahrens * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
49769962b5Matthew Ahrens * there is no eligible class.
49869962b5Matthew Ahrens */
49969962b5Matthew Ahrensstatic zio_priority_t
50069962b5Matthew Ahrensvdev_queue_class_to_issue(vdev_queue_t *vq)
50169962b5Matthew Ahrens{
50269962b5Matthew Ahrens	spa_t *spa = vq->vq_vdev->vdev_spa;
50369962b5Matthew Ahrens	zio_priority_t p;
50469962b5Matthew Ahrens
50569962b5Matthew Ahrens	if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
50669962b5Matthew Ahrens		return (ZIO_PRIORITY_NUM_QUEUEABLE);
50769962b5Matthew Ahrens
50869962b5Matthew Ahrens	/* find a queue that has not reached its minimum # outstanding i/os */
50969962b5Matthew Ahrens	for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
510fe31923Justin T. Gibbs		if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
51169962b5Matthew Ahrens		    vq->vq_class[p].vqc_active <
51269962b5Matthew Ahrens		    vdev_queue_class_min_active(p))
51369962b5Matthew Ahrens			return (p);
51469962b5Matthew Ahrens	}
51569962b5Matthew Ahrens
51669962b5Matthew Ahrens	/*
51769962b5Matthew Ahrens	 * If we haven't found a queue, look for one that hasn't reached its
51869962b5Matthew Ahrens	 * maximum # outstanding i/os.
51969962b5Matthew Ahrens	 */
52069962b5Matthew Ahrens	for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
521fe31923Justin T. Gibbs		if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
52269962b5Matthew Ahrens		    vq->vq_class[p].vqc_active <
52369962b5Matthew Ahrens		    vdev_queue_class_max_active(spa, p))
52469962b5Matthew Ahrens			return (p);
52569962b5Matthew Ahrens	}
52669962b5Matthew Ahrens
52769962b5Matthew Ahrens	/* No eligible queued i/os */
52869962b5Matthew Ahrens	return (ZIO_PRIORITY_NUM_QUEUEABLE);
52969962b5Matthew Ahrens}
53069962b5Matthew Ahrens
5316f708f7Jeff Bonwick/*
5326f708f7Jeff Bonwick * Compute the range spanned by two i/os, which is the endpoint of the last
5336f708f7Jeff Bonwick * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
5346f708f7Jeff Bonwick * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
5356f708f7Jeff Bonwick * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
5366f708f7Jeff Bonwick */
5376f708f7Jeff Bonwick#define	IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
5386f708f7Jeff Bonwick#define	IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
539fa9e406ahrens
540fa9e406ahrensstatic zio_t *
54169962b5Matthew Ahrensvdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
542fa9e406ahrens{
54369962b5Matthew Ahrens	zio_t *first, *last, *aio, *dio, *mandatory, *nio;
544a3874b8Toomas Soome	zio_link_t *zl = NULL;
54569962b5Matthew Ahrens	uint64_t maxgap = 0;
54669962b5Matthew Ahrens	uint64_t size;
54769962b5Matthew Ahrens	boolean_t stretch = B_FALSE;
548fe31923Justin T. Gibbs	avl_tree_t *t = vdev_queue_type_tree(vq, zio->io_type);
54969962b5Matthew Ahrens	enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
55069962b5Matthew Ahrens
55169962b5Matthew Ahrens	if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
55269962b5Matthew Ahrens		return (NULL);
553fa9e406ahrens
554084fd14Brian Behlendorf	/*
555084fd14Brian Behlendorf	 * While TRIM commands could be aggregated based on offset this
556084fd14Brian Behlendorf	 * behavior is disabled until it's determined to be beneficial.
557084fd14Brian Behlendorf	 */
558084fd14Brian Behlendorf	if (zio->io_type == ZIO_TYPE_TRIM && !zfs_vdev_aggregate_trim)
559084fd14Brian Behlendorf		return (NULL);
560084fd14Brian Behlendorf
56169962b5Matthew Ahrens	first = last = zio;
562fa9e406ahrens
56369962b5Matthew Ahrens	if (zio->io_type == ZIO_TYPE_READ)
56469962b5Matthew Ahrens		maxgap = zfs_vdev_read_gap_limit;
5658ad4d6dJeff Bonwick
56669962b5Matthew Ahrens	/*
56769962b5Matthew Ahrens	 * We can aggregate I/Os that are sufficiently adjacent and of
56869962b5Matthew Ahrens	 * the same flavor, as expressed by the AGG_INHERIT flags.
56969962b5Matthew Ahrens	 * The latter requirement is necessary so that certain
57069962b5Matthew Ahrens	 * attributes of the I/O, such as whether it's a normal I/O
57169962b5Matthew Ahrens	 * or a scrub/resilver, can be preserved in the aggregate.
57269962b5Matthew Ahrens	 * We can include optional I/Os, but don't allow them
57369962b5Matthew Ahrens	 * to begin a range as they add no benefit in that situation.
57469962b5Matthew Ahrens	 */
575f94275cAdam Leventhal
57669962b5Matthew Ahrens	/*
57769962b5Matthew Ahrens	 * We keep track of the last non-optional I/O.
57869962b5Matthew Ahrens	 */
57969962b5Matthew Ahrens	mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
580f94275cAdam Leventhal
58169962b5Matthew Ahrens	/*
58269962b5Matthew Ahrens	 * Walk backwards through sufficiently contiguous I/Os
5835b06278Matthew Ahrens	 * recording the last non-optional I/O.
58469962b5Matthew Ahrens	 */
58569962b5Matthew Ahrens	while ((dio = AVL_PREV(t, first)) != NULL &&
58669962b5Matthew Ahrens	    (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
58769962b5Matthew Ahrens	    IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
5885cabbc6Prashanth Sreenivasa	    IO_GAP(dio, first) <= maxgap &&
5895cabbc6Prashanth Sreenivasa	    dio->io_type == zio->io_type) {
59069962b5Matthew Ahrens		first = dio;
59169962b5Matthew Ahrens		if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
59269962b5Matthew Ahrens			mandatory = first;
59369962b5Matthew Ahrens	}
594f94275cAdam Leventhal
59569962b5Matthew Ahrens	/*
59669962b5Matthew Ahrens	 * Skip any initial optional I/Os.
59769962b5Matthew Ahrens	 */
59869962b5Matthew Ahrens	while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
59969962b5Matthew Ahrens		first = AVL_NEXT(t, first);
60069962b5Matthew Ahrens		ASSERT(first != NULL);
60169962b5Matthew Ahrens	}
6026f708f7Jeff Bonwick
60369962b5Matthew Ahrens	/*
60469962b5Matthew Ahrens	 * Walk forward through sufficiently contiguous I/Os.
6055b06278Matthew Ahrens	 * The aggregation limit does not apply to optional i/os, so that
6065b06278Matthew Ahrens	 * we can issue contiguous writes even if they are larger than the
6075b06278Matthew Ahrens	 * aggregation limit.
60869962b5Matthew Ahrens	 */
60969962b5Matthew Ahrens	while ((dio = AVL_NEXT(t, last)) != NULL &&
61069962b5Matthew Ahrens	    (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
6115b06278Matthew Ahrens	    (IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit ||
6125b06278Matthew Ahrens	    (dio->io_flags & ZIO_FLAG_OPTIONAL)) &&
6135cabbc6Prashanth Sreenivasa	    IO_GAP(last, dio) <= maxgap &&
6145cabbc6Prashanth Sreenivasa	    dio->io_type == zio->io_type) {
61569962b5Matthew Ahrens		last = dio;
61669962b5Matthew Ahrens		if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
61769962b5Matthew Ahrens			mandatory = last;
61869962b5Matthew Ahrens	}
619f94275cAdam Leventhal
62069962b5Matthew Ahrens	/*
62169962b5Matthew Ahrens	 * Now that we've established the range of the I/O aggregation
62269962b5Matthew Ahrens	 * we must decide what to do with trailing optional I/Os.
62369962b5Matthew Ahrens	 * For reads, there's nothing to do. While we are unable to
62469962b5Matthew Ahrens	 * aggregate further, it's possible that a trailing optional
62569962b5Matthew Ahrens	 * I/O would allow the underlying device to aggregate with
62669962b5Matthew Ahrens	 * subsequent I/Os. We must therefore determine if the next
62769962b5Matthew Ahrens	 * non-optional I/O is close enough to make aggregation
62869962b5Matthew Ahrens	 * worthwhile.
62969962b5Matthew Ahrens	 */
63069962b5Matthew Ahrens	if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
63169962b5Matthew Ahrens		zio_t *nio = last;
63269962b5Matthew Ahrens		while ((dio = AVL_NEXT(t, nio)) != NULL &&
63369962b5Matthew Ahrens		    IO_GAP(nio, dio) == 0 &&
63469962b5Matthew Ahrens		    IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
63569962b5Matthew Ahrens			nio = dio;
63669962b5Matthew Ahrens			if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
63769962b5Matthew Ahrens				stretch = B_TRUE;
63869962b5Matthew Ahrens				break;
639f94275cAdam Leventhal			}
640f94275cAdam Leventhal		}
64169962b5Matthew Ahrens	}
642f94275cAdam Leventhal
64369962b5Matthew Ahrens	if (stretch) {
6445b06278Matthew Ahrens		/*
6455b06278Matthew Ahrens		 * We are going to include an optional io in our aggregated
6465b06278Matthew Ahrens		 * span, thus closing the write gap.  Only mandatory i/os can
6475b06278Matthew Ahrens		 * start aggregated spans, so make sure that the next i/o
6485b06278Matthew Ahrens		 * after our span is mandatory.
6495b06278Matthew Ahrens		 */
65069962b5Matthew Ahrens		dio = AVL_NEXT(t, last);
65169962b5Matthew Ahrens		dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
65269962b5Matthew Ahrens	} else {
6535b06278Matthew Ahrens		/* do not include the optional i/o */
65469962b5Matthew Ahrens		while (last != mandatory && last != first) {
65569962b5Matthew Ahrens			ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
65669962b5Matthew Ahrens			last = AVL_PREV(t, last);
65769962b5Matthew Ahrens			ASSERT(last != NULL);
658f94275cAdam Leventhal		}
659fa9e406ahrens	}
660fa9e406ahrens
66169962b5Matthew Ahrens	if (first == last)
66269962b5Matthew Ahrens		return (NULL);
66369962b5Matthew Ahrens
66469962b5Matthew Ahrens	size = IO_SPAN(first, last);
6655b06278Matthew Ahrens	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
66669962b5Matthew Ahrens
66769962b5Matthew Ahrens	aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
668770499eDan Kimmel	    abd_alloc_for_io(size, B_TRUE), size, first->io_type,
669770499eDan Kimmel	    zio->io_priority, flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
67069962b5Matthew Ahrens	    vdev_queue_agg_io_done, NULL);
67169962b5Matthew Ahrens	aio->io_timestamp = first->io_timestamp;
67269962b5Matthew Ahrens
67369962b5Matthew Ahrens	nio = first;
67469962b5Matthew Ahrens	do {
67569962b5Matthew Ahrens		dio = nio;
67669962b5Matthew Ahrens		nio = AVL_NEXT(t, dio);
67769962b5Matthew Ahrens		ASSERT3U(dio->io_type, ==, aio->io_type);
67869962b5Matthew Ahrens
67969962b5Matthew Ahrens		if (dio->io_flags & ZIO_FLAG_NODATA) {
68069962b5Matthew Ahrens			ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
681770499eDan Kimmel			abd_zero_off(aio->io_abd,
682770499eDan Kimmel			    dio->io_offset - aio->io_offset, dio->io_size);
68369962b5Matthew Ahrens		} else if (dio->io_type == ZIO_TYPE_WRITE) {
684770499eDan Kimmel			abd_copy_off(aio->io_abd, dio->io_abd,
685770499eDan Kimmel			    dio->io_offset - aio->io_offset, 0, dio->io_size);
68669962b5Matthew Ahrens		}
687a3f829aBill Moore
68869962b5Matthew Ahrens		zio_add_child(dio, aio);
68969962b5Matthew Ahrens		vdev_queue_io_remove(vq, dio);
690a3874b8Toomas Soome	} while (dio != last);
691a3874b8Toomas Soome
692a3874b8Toomas Soome	/*
693a3874b8Toomas Soome	 * We need to drop the vdev queue's lock to avoid a deadlock that we
694a3874b8Toomas Soome	 * could encounter since this I/O will complete immediately.
695a3874b8Toomas Soome	 */
696a3874b8Toomas Soome	mutex_exit(&vq->vq_lock);
697a3874b8Toomas Soome	while ((dio = zio_walk_parents(aio, &zl)) != NULL) {
69869962b5Matthew Ahrens		zio_vdev_io_bypass(dio);
69969962b5Matthew Ahrens		zio_execute(dio);
700a3874b8Toomas Soome	}
701a3874b8Toomas Soome	mutex_enter(&vq->vq_lock);
70269962b5Matthew Ahrens
70369962b5Matthew Ahrens	return (aio);
70469962b5Matthew Ahrens}
70569962b5Matthew Ahrens
70669962b5Matthew Ahrensstatic zio_t *
70769962b5Matthew Ahrensvdev_queue_io_to_issue(vdev_queue_t *vq)
70869962b5Matthew Ahrens{
70969962b5Matthew Ahrens	zio_t *zio, *aio;
71069962b5Matthew Ahrens	zio_priority_t p;
71169962b5Matthew Ahrens	avl_index_t idx;
712fe31923Justin T. Gibbs	avl_tree_t *tree;
71369962b5Matthew Ahrens	zio_t search;
71469962b5Matthew Ahrens
71569962b5Matthew Ahrensagain:
71669962b5Matthew Ahrens	ASSERT(MUTEX_HELD(&vq->vq_lock));
717fa9e406ahrens
71869962b5Matthew Ahrens	p = vdev_queue_class_to_issue(vq);
719fa9e406ahrens
72069962b5Matthew Ahrens	if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
72169962b5Matthew Ahrens		/* No eligible queued i/os */
722