xref: /illumos-gate/usr/src/uts/common/fs/zfs/vdev_queue.c (revision dd50e0cc)
1fa9e4066Sahrens /*
2fa9e4066Sahrens  * CDDL HEADER START
3fa9e4066Sahrens  *
4fa9e4066Sahrens  * The contents of this file are subject to the terms of the
5ea8dc4b6Seschrock  * Common Development and Distribution License (the "License").
6ea8dc4b6Seschrock  * You may not use this file except in compliance with the License.
7fa9e4066Sahrens  *
8fa9e4066Sahrens  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9fa9e4066Sahrens  * or http://www.opensolaris.org/os/licensing.
10fa9e4066Sahrens  * See the License for the specific language governing permissions
11fa9e4066Sahrens  * and limitations under the License.
12fa9e4066Sahrens  *
13fa9e4066Sahrens  * When distributing Covered Code, include this CDDL HEADER in each
14fa9e4066Sahrens  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15fa9e4066Sahrens  * If applicable, add the following below this CDDL HEADER, with the
16fa9e4066Sahrens  * fields enclosed by brackets "[]" replaced with your own identifying
17fa9e4066Sahrens  * information: Portions Copyright [yyyy] [name of copyright owner]
18fa9e4066Sahrens  *
19fa9e4066Sahrens  * CDDL HEADER END
20fa9e4066Sahrens  */
21fa9e4066Sahrens /*
22a3f829aeSBill Moore  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23fa9e4066Sahrens  * Use is subject to license terms.
24fa9e4066Sahrens  */
26283b8460SGeorge.Wilson /*
27f78cdc34SPaul Dagnelie  * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
28c3d26abcSMatthew Ahrens  * Copyright (c) 2014 Integros [integros.com]
2944bf619dSJohn Levon  * Copyright 2019 Joyent, Inc.
30283b8460SGeorge.Wilson  */
32fa9e4066Sahrens #include <sys/zfs_context.h>
33fa9e4066Sahrens #include <sys/vdev_impl.h>
34c3a66015SMatthew Ahrens #include <sys/spa_impl.h>
35fa9e4066Sahrens #include <sys/zio.h>
36fa9e4066Sahrens #include <sys/avl.h>
3769962b56SMatthew Ahrens #include <sys/dsl_pool.h>
380f7643c7SGeorge Wilson #include <sys/metaslab_impl.h>
39770499e1SDan Kimmel #include <sys/abd.h>
41614409b5Sahrens /*
4269962b56SMatthew Ahrens  * ZFS I/O Scheduler
4369962b56SMatthew Ahrens  * ---------------
4469962b56SMatthew Ahrens  *
4569962b56SMatthew Ahrens  * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios.  The
4669962b56SMatthew Ahrens  * I/O scheduler determines when and in what order those operations are
4769962b56SMatthew Ahrens  * issued.  The I/O scheduler divides operations into five I/O classes
4869962b56SMatthew Ahrens  * prioritized in the following order: sync read, sync write, async read,
4969962b56SMatthew Ahrens  * async write, and scrub/resilver.  Each queue defines the minimum and
5069962b56SMatthew Ahrens  * maximum number of concurrent operations that may be issued to the device.
5169962b56SMatthew Ahrens  * In addition, the device has an aggregate maximum. Note that the sum of the
5269962b56SMatthew Ahrens  * per-queue minimums must not exceed the aggregate maximum, and if the
5369962b56SMatthew Ahrens  * aggregate maximum is equal to or greater than the sum of the per-queue
5469962b56SMatthew Ahrens  * maximums, the per-queue minimum has no effect.
5569962b56SMatthew Ahrens  *
5669962b56SMatthew Ahrens  * For many physical devices, throughput increases with the number of
5769962b56SMatthew Ahrens  * concurrent operations, but latency typically suffers. Further, physical
5869962b56SMatthew Ahrens  * devices typically have a limit at which more concurrent operations have no
5969962b56SMatthew Ahrens  * effect on throughput or can actually cause it to decrease.
6069962b56SMatthew Ahrens  *
6169962b56SMatthew Ahrens  * The scheduler selects the next operation to issue by first looking for an
6269962b56SMatthew Ahrens  * I/O class whose minimum has not been satisfied. Once all are satisfied and
6369962b56SMatthew Ahrens  * the aggregate maximum has not been hit, the scheduler looks for classes
6469962b56SMatthew Ahrens  * whose maximum has not been satisfied. Iteration through the I/O classes is
6569962b56SMatthew Ahrens  * done in the order specified above. No further operations are issued if the
6669962b56SMatthew Ahrens  * aggregate maximum number of concurrent operations has been hit or if there
6769962b56SMatthew Ahrens  * are no operations queued for an I/O class that has not hit its maximum.
6869962b56SMatthew Ahrens  * Every time an i/o is queued or an operation completes, the I/O scheduler
6969962b56SMatthew Ahrens  * looks for new operations to issue.
7069962b56SMatthew Ahrens  *
7169962b56SMatthew Ahrens  * All I/O classes have a fixed maximum number of outstanding operations
7269962b56SMatthew Ahrens  * except for the async write class. Asynchronous writes represent the data
7369962b56SMatthew Ahrens  * that is committed to stable storage during the syncing stage for
7469962b56SMatthew Ahrens  * transaction groups (see txg.c). Transaction groups enter the syncing state
7569962b56SMatthew Ahrens  * periodically so the number of queued async writes will quickly burst up and
7669962b56SMatthew Ahrens  * then bleed down to zero. Rather than servicing them as quickly as possible,
7769962b56SMatthew Ahrens  * the I/O scheduler changes the maximum number of active async write i/os
7869962b56SMatthew Ahrens  * according to the amount of dirty data in the pool (see dsl_pool.c). Since
7969962b56SMatthew Ahrens  * both throughput and latency typically increase with the number of
8069962b56SMatthew Ahrens  * concurrent operations issued to physical devices, reducing the burstiness
8169962b56SMatthew Ahrens  * in the number of concurrent operations also stabilizes the response time of
8269962b56SMatthew Ahrens  * operations from other -- and in particular synchronous -- queues. In broad
8369962b56SMatthew Ahrens  * strokes, the I/O scheduler will issue more concurrent operations from the
8469962b56SMatthew Ahrens  * async write queue as there's more dirty data in the pool.
8569962b56SMatthew Ahrens  *
8669962b56SMatthew Ahrens  * Async Writes
8769962b56SMatthew Ahrens  *
8869962b56SMatthew Ahrens  * The number of concurrent operations issued for the async write I/O class
8969962b56SMatthew Ahrens  * follows a piece-wise linear function defined by a few adjustable points.
9069962b56SMatthew Ahrens  *
9169962b56SMatthew Ahrens  *        |                   o---------| <-- zfs_vdev_async_write_max_active
9269962b56SMatthew Ahrens  *   ^    |                  /^         |
9369962b56SMatthew Ahrens  *   |    |                 / |         |
9469962b56SMatthew Ahrens  * active |                /  |         |
9569962b56SMatthew Ahrens  *  I/O   |               /   |         |
9669962b56SMatthew Ahrens  * count  |              /    |         |
9769962b56SMatthew Ahrens  *        |             /     |         |
9869962b56SMatthew Ahrens  *        |------------o      |         | <-- zfs_vdev_async_write_min_active
9969962b56SMatthew Ahrens  *       0|____________^______|_________|
10069962b56SMatthew Ahrens  *        0%           |      |       100% of zfs_dirty_data_max
10169962b56SMatthew Ahrens  *                     |      |
10269962b56SMatthew Ahrens  *                     |      `-- zfs_vdev_async_write_active_max_dirty_percent
10369962b56SMatthew Ahrens  *                     `--------- zfs_vdev_async_write_active_min_dirty_percent
10469962b56SMatthew Ahrens  *
10569962b56SMatthew Ahrens  * Until the amount of dirty data exceeds a minimum percentage of the dirty
10669962b56SMatthew Ahrens  * data allowed in the pool, the I/O scheduler will limit the number of
10769962b56SMatthew Ahrens  * concurrent operations to the minimum. As that threshold is crossed, the
10869962b56SMatthew Ahrens  * number of concurrent operations issued increases linearly to the maximum at
10969962b56SMatthew Ahrens  * the specified maximum percentage of the dirty data allowed in the pool.
11069962b56SMatthew Ahrens  *
11169962b56SMatthew Ahrens  * Ideally, the amount of dirty data on a busy pool will stay in the sloped
11269962b56SMatthew Ahrens  * part of the function between zfs_vdev_async_write_active_min_dirty_percent
11369962b56SMatthew Ahrens  * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the
11469962b56SMatthew Ahrens  * maximum percentage, this indicates that the rate of incoming data is
11569962b56SMatthew Ahrens  * greater than the rate that the backend storage can handle. In this case, we
11669962b56SMatthew Ahrens  * must further throttle incoming writes (see dmu_tx_delay() for details).
117614409b5Sahrens  */
118f7170741SWill Andrews 
119614409b5Sahrens /*
12069962b56SMatthew Ahrens  * The maximum number of i/os active to each device.  Ideally, this will be >=
12169962b56SMatthew Ahrens  * the sum of each queue's max_active.  It must be at least the sum of each
12269962b56SMatthew Ahrens  * queue's min_active.
123614409b5Sahrens  */
12469962b56SMatthew Ahrens uint32_t zfs_vdev_max_active = 1000;
126c55e05cbSMatthew Ahrens /*
12769962b56SMatthew Ahrens  * Per-queue limits on the number of i/os active to each device.  If the
12869962b56SMatthew Ahrens  * sum of the queue's max_active is < zfs_vdev_max_active, then the
12969962b56SMatthew Ahrens  * min_active comes into play.  We will send min_active from each queue,
13069962b56SMatthew Ahrens  * and then select from queues in the order defined by zio_priority_t.
13169962b56SMatthew Ahrens  *
13269962b56SMatthew Ahrens  * In general, smaller max_active's will lead to lower latency of synchronous
13369962b56SMatthew Ahrens  * operations.  Larger max_active's may lead to higher overall throughput,
13469962b56SMatthew Ahrens  * depending on underlying storage.
13569962b56SMatthew Ahrens  *
13669962b56SMatthew Ahrens  * The ratio of the queues' max_actives determines the balance of performance
13769962b56SMatthew Ahrens  * between reads, writes, and scrubs.  E.g., increasing
13869962b56SMatthew Ahrens  * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete
13969962b56SMatthew Ahrens  * more quickly, but reads and writes to have higher latency and lower
14069962b56SMatthew Ahrens  * throughput.
141c55e05cbSMatthew Ahrens  */
14269962b56SMatthew Ahrens uint32_t zfs_vdev_sync_read_min_active = 10;
14369962b56SMatthew Ahrens uint32_t zfs_vdev_sync_read_max_active = 10;
14469962b56SMatthew Ahrens uint32_t zfs_vdev_sync_write_min_active = 10;
14569962b56SMatthew Ahrens uint32_t zfs_vdev_sync_write_max_active = 10;
14669962b56SMatthew Ahrens uint32_t zfs_vdev_async_read_min_active = 1;
14769962b56SMatthew Ahrens uint32_t zfs_vdev_async_read_max_active = 3;
14869962b56SMatthew Ahrens uint32_t zfs_vdev_async_write_min_active = 1;
14969962b56SMatthew Ahrens uint32_t zfs_vdev_async_write_max_active = 10;
15069962b56SMatthew Ahrens uint32_t zfs_vdev_scrub_min_active = 1;
15169962b56SMatthew Ahrens uint32_t zfs_vdev_scrub_max_active = 2;
1525cabbc6bSPrashanth Sreenivasa uint32_t zfs_vdev_removal_min_active = 1;
1535cabbc6bSPrashanth Sreenivasa uint32_t zfs_vdev_removal_max_active = 2;
154094e47e9SGeorge Wilson uint32_t zfs_vdev_initializing_min_active = 1;
155094e47e9SGeorge Wilson uint32_t zfs_vdev_initializing_max_active = 1;
156084fd14fSBrian Behlendorf uint32_t zfs_vdev_trim_min_active = 1;
157084fd14fSBrian Behlendorf uint32_t zfs_vdev_trim_max_active = 2;
15969962b56SMatthew Ahrens /*
16069962b56SMatthew Ahrens  * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent
16169962b56SMatthew Ahrens  * dirty data, use zfs_vdev_async_write_min_active.  When it has more than
16269962b56SMatthew Ahrens  * zfs_vdev_async_write_active_max_dirty_percent, use
16369962b56SMatthew Ahrens  * zfs_vdev_async_write_max_active. The value is linearly interpolated
16469962b56SMatthew Ahrens  * between min and max.
16569962b56SMatthew Ahrens  */
16669962b56SMatthew Ahrens int zfs_vdev_async_write_active_min_dirty_percent = 30;
16769962b56SMatthew Ahrens int zfs_vdev_async_write_active_max_dirty_percent = 60;
169614409b5Sahrens /*
170f94275ceSAdam Leventhal  * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
171f94275ceSAdam Leventhal  * For read I/Os, we also aggregate across small adjacency gaps; for writes
172f94275ceSAdam Leventhal  * we include spans of optional I/Os to aid aggregation at the disk even when
173f94275ceSAdam Leventhal  * they aren't able to help us aggregate at this level.
174614409b5Sahrens  */
175a3874b8bSToomas Soome int zfs_vdev_aggregation_limit = 1 << 20;
1766f708f7cSJeff Bonwick int zfs_vdev_read_gap_limit = 32 << 10;
177f94275ceSAdam Leventhal int zfs_vdev_write_gap_limit = 4 << 10;
1790f7643c7SGeorge Wilson /*
1800f7643c7SGeorge Wilson  * Define the queue depth percentage for each top-level. This percentage is
1810f7643c7SGeorge Wilson  * used in conjunction with zfs_vdev_async_max_active to determine how many
1820f7643c7SGeorge Wilson  * allocations a specific top-level vdev should handle. Once the queue depth
1830f7643c7SGeorge Wilson  * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100
1840f7643c7SGeorge Wilson  * then allocator will stop allocating blocks on that top-level device.
1850f7643c7SGeorge Wilson  * The default kernel setting is 1000% which will yield 100 allocations per
1860f7643c7SGeorge Wilson  * device. For userland testing, the default setting is 300% which equates
1870f7643c7SGeorge Wilson  * to 30 allocations per device.
1880f7643c7SGeorge Wilson  */
1890f7643c7SGeorge Wilson #ifdef _KERNEL
1900f7643c7SGeorge Wilson int zfs_vdev_queue_depth_pct = 1000;
1910f7643c7SGeorge Wilson #else
1920f7643c7SGeorge Wilson int zfs_vdev_queue_depth_pct = 300;
1930f7643c7SGeorge Wilson #endif
1940f7643c7SGeorge Wilson 
195f78cdc34SPaul Dagnelie /*
196f78cdc34SPaul Dagnelie  * When performing allocations for a given metaslab, we want to make sure that
197f78cdc34SPaul Dagnelie  * there are enough IOs to aggregate together to improve throughput. We want to
198f78cdc34SPaul Dagnelie  * ensure that there are at least 128k worth of IOs that can be aggregated, and
199f78cdc34SPaul Dagnelie  * we assume that the average allocation size is 4k, so we need the queue depth
200f78cdc34SPaul Dagnelie  * to be 32 per allocator to get good aggregation of sequential writes.
201f78cdc34SPaul Dagnelie  */
202f78cdc34SPaul Dagnelie int zfs_vdev_def_queue_depth = 32;
203f78cdc34SPaul Dagnelie 
204084fd14fSBrian Behlendorf /*
205084fd14fSBrian Behlendorf  * Allow TRIM I/Os to be aggregated.  This should normally not be needed since
206084fd14fSBrian Behlendorf  * TRIM I/O for extents up to zfs_trim_extent_bytes_max (128M) can be submitted
207084fd14fSBrian Behlendorf  * by the TRIM code in zfs_trim.c.
208084fd14fSBrian Behlendorf  */
209084fd14fSBrian Behlendorf int zfs_vdev_aggregate_trim = 0;
2100f7643c7SGeorge Wilson 
211fa9e4066Sahrens int
vdev_queue_offset_compare(const void * x1,const void * x2)21269962b56SMatthew Ahrens vdev_queue_offset_compare(const void *x1, const void *x2)
213fa9e4066Sahrens {
214c4ab0d3fSGvozden Neskovic 	const zio_t *z1 = (const zio_t *)x1;
215c4ab0d3fSGvozden Neskovic 	const zio_t *z2 = (const zio_t *)x2;
2174d7988d6SPaul Dagnelie 	int cmp = TREE_CMP(z1->io_offset, z2->io_offset);
219c4ab0d3fSGvozden Neskovic 	if (likely(cmp))
220c4ab0d3fSGvozden Neskovic 		return (cmp);
2224d7988d6SPaul Dagnelie 	return (TREE_PCMP(z1, z2));
223fa9e4066Sahrens }
225fe319232SJustin T. Gibbs static inline avl_tree_t *
vdev_queue_class_tree(vdev_queue_t * vq,zio_priority_t p)226fe319232SJustin T. Gibbs vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p)
227fe319232SJustin T. Gibbs {
228fe319232SJustin T. Gibbs 	return (&vq->vq_class[p].vqc_queued_tree);
229fe319232SJustin T. Gibbs }
230fe319232SJustin T. Gibbs 
231fe319232SJustin T. Gibbs static inline avl_tree_t *
vdev_queue_type_tree(vdev_queue_t * vq,zio_type_t t)232fe319232SJustin T. Gibbs vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t)
233fe319232SJustin T. Gibbs {
234084fd14fSBrian Behlendorf 	ASSERT(t == ZIO_TYPE_READ || t == ZIO_TYPE_WRITE || t == ZIO_TYPE_TRIM);
235fe319232SJustin T. Gibbs 	if (t == ZIO_TYPE_READ)
236fe319232SJustin T. Gibbs 		return (&vq->vq_read_offset_tree);
237084fd14fSBrian Behlendorf 	else if (t == ZIO_TYPE_WRITE)
238fe319232SJustin T. Gibbs 		return (&vq->vq_write_offset_tree);
239084fd14fSBrian Behlendorf 	else
240084fd14fSBrian Behlendorf 		return (&vq->vq_trim_offset_tree);
241fe319232SJustin T. Gibbs }
242fe319232SJustin T. Gibbs 
243fa9e4066Sahrens int
vdev_queue_timestamp_compare(const void * x1,const void * x2)24469962b56SMatthew Ahrens vdev_queue_timestamp_compare(const void *x1, const void *x2)
245fa9e4066Sahrens {
246c4ab0d3fSGvozden Neskovic 	const zio_t *z1 = (const zio_t *)x1;
247c4ab0d3fSGvozden Neskovic 	const zio_t *z2 = (const zio_t *)x2;
2494d7988d6SPaul Dagnelie 	int cmp = TREE_CMP(z1->io_timestamp, z2->io_timestamp);
251c4ab0d3fSGvozden Neskovic 	if (likely(cmp))
252c4ab0d3fSGvozden Neskovic 		return (cmp);
2544d7988d6SPaul Dagnelie 	return (TREE_PCMP(z1, z2));
255fa9e4066Sahrens }
257fa9e4066Sahrens void
vdev_queue_init(vdev_t * vd)258fa9e4066Sahrens vdev_queue_init(vdev_t *vd)
259fa9e4066Sahrens {
260fa9e4066Sahrens 	vdev_queue_t *vq = &vd->vdev_queue;
262fa9e4066Sahrens 	mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
26369962b56SMatthew Ahrens 	vq->vq_vdev = vd;
26569962b56SMatthew Ahrens 	avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
26669962b56SMatthew Ahrens 	    sizeof (zio_t), offsetof(struct zio, io_queue_node));
267fe319232SJustin T. Gibbs 	avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ),
268fe319232SJustin T. Gibbs 	    vdev_queue_offset_compare, sizeof (zio_t),
269fe319232SJustin T. Gibbs 	    offsetof(struct zio, io_offset_node));
270fe319232SJustin T. Gibbs 	avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE),
271fe319232SJustin T. Gibbs 	    vdev_queue_offset_compare, sizeof (zio_t),
272fe319232SJustin T. Gibbs 	    offsetof(struct zio, io_offset_node));
273084fd14fSBrian Behlendorf 	avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_TRIM),
274084fd14fSBrian Behlendorf 	    vdev_queue_offset_compare, sizeof (zio_t),
275084fd14fSBrian Behlendorf 	    offsetof(struct zio, io_offset_node));
27769962b56SMatthew Ahrens 	for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
278fe319232SJustin T. Gibbs 		int (*compfn) (const void *, const void *);
279fe319232SJustin T. Gibbs 
28069962b56SMatthew Ahrens 		/*
281084fd14fSBrian Behlendorf 		 * The synchronous/trim i/o queues are dispatched in FIFO rather
282084fd14fSBrian Behlendorf 		 * than LBA order. This provides more consistent latency for
283fe319232SJustin T. Gibbs 		 * these i/os.
28469962b56SMatthew Ahrens 		 */
285084fd14fSBrian Behlendorf 		if (p == ZIO_PRIORITY_SYNC_READ ||
286084fd14fSBrian Behlendorf 		    p == ZIO_PRIORITY_SYNC_WRITE ||
287084fd14fSBrian Behlendorf 		    p == ZIO_PRIORITY_TRIM) {
288fe319232SJustin T. Gibbs 			compfn = vdev_queue_timestamp_compare;
289084fd14fSBrian Behlendorf 		} else {
290fe319232SJustin T. Gibbs 			compfn = vdev_queue_offset_compare;
291084fd14fSBrian Behlendorf 		}
292fe319232SJustin T. Gibbs 
293fe319232SJustin T. Gibbs 		avl_create(vdev_queue_class_tree(vq, p), compfn,
29469962b56SMatthew Ahrens 		    sizeof (zio_t), offsetof(struct zio, io_queue_node));
29569962b56SMatthew Ahrens 	}
29612a8814cSTom Caputi 
29712a8814cSTom Caputi 	vq->vq_last_offset = 0;
298fa9e4066Sahrens }
300fa9e4066Sahrens void
vdev_queue_fini(vdev_t * vd)301fa9e4066Sahrens vdev_queue_fini(vdev_t *vd)
302fa9e4066Sahrens {
303fa9e4066Sahrens 	vdev_queue_t *vq = &vd->vdev_queue;
30569962b56SMatthew Ahrens 	for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
306fe319232SJustin T. Gibbs 		avl_destroy(vdev_queue_class_tree(vq, p));
30769962b56SMatthew Ahrens 	avl_destroy(&vq->vq_active_tree);
308fe319232SJustin T. Gibbs 	avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
309fe319232SJustin T. Gibbs 	avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE));
310084fd14fSBrian Behlendorf 	avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_TRIM));
312fa9e4066Sahrens 	mutex_destroy(&vq->vq_lock);
313fa9e4066Sahrens }
315ea8dc4b6Seschrock static void
vdev_queue_io_add(vdev_queue_t * vq,zio_t * zio)316ea8dc4b6Seschrock vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
317ea8dc4b6Seschrock {
318c3a66015SMatthew Ahrens 	spa_t *spa = zio->io_spa;
3190f7643c7SGeorge Wilson 
32069962b56SMatthew Ahrens 	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
321fe319232SJustin T. Gibbs 	avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
322fe319232SJustin T. Gibbs 	avl_add(vdev_queue_type_tree(vq, zio->io_type), zio);
323c3a66015SMatthew Ahrens 
32469962b56SMatthew Ahrens 	mutex_enter(&spa->spa_iokstat_lock);
32569962b56SMatthew Ahrens 	spa->spa_queue_stats[zio->io_priority].spa_queued++;
32669962b56SMatthew Ahrens 	if (spa->spa_iokstat != NULL)
327c3a66015SMatthew Ahrens 		kstat_waitq_enter(spa->spa_iokstat->ks_data);
32869962b56SMatthew Ahrens 	mutex_exit(&spa->spa_iokstat_lock);
329ea8dc4b6Seschrock }
331ea8dc4b6Seschrock static void
vdev_queue_io_remove(vdev_queue_t * vq,zio_t * zio)332ea8dc4b6Seschrock vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
333ea8dc4b6Seschrock {
334c3a66015SMatthew Ahrens 	spa_t *spa = zio->io_spa;
3350f7643c7SGeorge Wilson 
33669962b56SMatthew Ahrens 	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
337fe319232SJustin T. Gibbs 	avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
338fe319232SJustin T. Gibbs 	avl_remove(vdev_queue_type_tree(vq, zio->io_type), zio);
339c3a66015SMatthew Ahrens 
34069962b56SMatthew Ahrens 	mutex_enter(&spa->spa_iokstat_lock);
34169962b56SMatthew Ahrens 	ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
34269962b56SMatthew Ahrens 	spa->spa_queue_stats[zio->io_priority].spa_queued--;
34369962b56SMatthew Ahrens 	if (spa->spa_iokstat != NULL)
344c3a66015SMatthew Ahrens 		kstat_waitq_exit(spa->spa_iokstat->ks_data);
34569962b56SMatthew Ahrens 	mutex_exit(&spa->spa_iokstat_lock);
346c3a66015SMatthew Ahrens }
347c3a66015SMatthew Ahrens 
348c3a66015SMatthew Ahrens static void
vdev_queue_pending_add(vdev_queue_t * vq,zio_t * zio)349c3a66015SMatthew Ahrens vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
350c3a66015SMatthew Ahrens {
351c3a66015SMatthew Ahrens 	spa_t *spa = zio->io_spa;
35269962b56SMatthew Ahrens 	ASSERT(MUTEX_HELD(&vq->vq_lock));
35369962b56SMatthew Ahrens 	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
35469962b56SMatthew Ahrens 	vq->vq_class[zio->io_priority].vqc_active++;
35569962b56SMatthew Ahrens 	avl_add(&vq->vq_active_tree, zio);
35669962b56SMatthew Ahrens 
35769962b56SMatthew Ahrens 	mutex_enter(&spa->spa_iokstat_lock);
35869962b56SMatthew Ahrens 	spa->spa_queue_stats[zio->io_priority].spa_active++;
35969962b56SMatthew Ahrens 	if (spa->spa_iokstat != NULL)
360c3a66015SMatthew Ahrens 		kstat_runq_enter(spa->spa_iokstat->ks_data);
36169962b56SMatthew Ahrens 	mutex_exit(&spa->spa_iokstat_lock);
362c3a66015SMatthew Ahrens }
363c3a66015SMatthew Ahrens 
364c3a66015SMatthew Ahrens static void
vdev_queue_pending_remove(vdev_queue_t * vq,zio_t * zio)365c3a66015SMatthew Ahrens vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
366c3a66015SMatthew Ahrens {
367c3a66015SMatthew Ahrens 	spa_t *spa = zio->io_spa;
36869962b56SMatthew Ahrens 	ASSERT(MUTEX_HELD(&vq->vq_lock));
36969962b56SMatthew Ahrens 	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
37069962b56SMatthew Ahrens 	vq->vq_class[zio->io_priority].vqc_active--;
37169962b56SMatthew Ahrens 	avl_remove(&vq->vq_active_tree, zio);
37269962b56SMatthew Ahrens 
37369962b56SMatthew Ahrens 	mutex_enter(&spa->spa_iokstat_lock);
37469962b56SMatthew Ahrens 	ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
37569962b56SMatthew Ahrens 	spa->spa_queue_stats[zio->io_priority].spa_active--;
376c3a66015SMatthew Ahrens 	if (spa->spa_iokstat != NULL) {
377c3a66015SMatthew Ahrens 		kstat_io_t *ksio = spa->spa_iokstat->ks_data;
378c3a66015SMatthew Ahrens 
379c3a66015SMatthew Ahrens 		kstat_runq_exit(spa->spa_iokstat->ks_data);
380c3a66015SMatthew Ahrens 		if (zio->io_type == ZIO_TYPE_READ) {
381c3a66015SMatthew Ahrens 			ksio->reads++;
382c3a66015SMatthew Ahrens 			ksio->nread += zio->io_size;
383c3a66015SMatthew Ahrens 		} else if (zio->io_type == ZIO_TYPE_WRITE) {
384c3a66015SMatthew Ahrens 			ksio->writes++;
385c3a66015SMatthew Ahrens 			ksio->nwritten += zio->io_size;
386c3a66015SMatthew Ahrens 		}
387c3a66015SMatthew Ahrens 	}
38869962b56SMatthew Ahrens 	mutex_exit(&spa->spa_iokstat_lock);
389ea8dc4b6Seschrock }
391fa9e4066Sahrens static void
vdev_queue_agg_io_done(zio_t * aio)392fa9e4066Sahrens vdev_queue_agg_io_done(zio_t *aio)
393fa9e4066Sahrens {
39469962b56SMatthew Ahrens 	if (aio->io_type == ZIO_TYPE_READ) {
39569962b56SMatthew Ahrens 		zio_t *pio;
3960f7643c7SGeorge Wilson 		zio_link_t *zl = NULL;
3970f7643c7SGeorge Wilson 		while ((pio = zio_walk_parents(aio, &zl)) != NULL) {
398770499e1SDan Kimmel 			abd_copy_off(pio->io_abd, aio->io_abd,
399770499e1SDan Kimmel 			    0, pio->io_offset - aio->io_offset, pio->io_size);
40069962b56SMatthew Ahrens 		}
40169962b56SMatthew Ahrens 	}
403770499e1SDan Kimmel 	abd_free(aio->io_abd);
404fa9e4066Sahrens }
40669962b56SMatthew Ahrens static int
vdev_queue_class_min_active(zio_priority_t p)40769962b56SMatthew Ahrens vdev_queue_class_min_active(zio_priority_t p)
40869962b56SMatthew Ahrens {
40969962b56SMatthew Ahrens 	switch (p) {
41069962b56SMatthew Ahrens 	case ZIO_PRIORITY_SYNC_READ:
41169962b56SMatthew Ahrens 		return (zfs_vdev_sync_read_min_active);
41269962b56SMatthew Ahrens 	case ZIO_PRIORITY_SYNC_WRITE:
41369962b56SMatthew Ahrens 		return (zfs_vdev_sync_write_min_active);
41469962b56SMatthew Ahrens 	case ZIO_PRIORITY_ASYNC_READ:
41569962b56SMatthew Ahrens 		return (zfs_vdev_async_read_min_active);
41669962b56SMatthew Ahrens 	case ZIO_PRIORITY_ASYNC_WRITE:
41769962b56SMatthew Ahrens 		return (zfs_vdev_async_write_min_active);
41869962b56SMatthew Ahrens 	case ZIO_PRIORITY_SCRUB:
41969962b56SMatthew Ahrens 		return (zfs_vdev_scrub_min_active);
4205cabbc6bSPrashanth Sreenivasa 	case ZIO_PRIORITY_REMOVAL:
4215cabbc6bSPrashanth Sreenivasa 		return (zfs_vdev_removal_min_active);
422094e47e9SGeorge Wilson 	case ZIO_PRIORITY_INITIALIZING:
423094e47e9SGeorge Wilson 		return (zfs_vdev_initializing_min_active);
424084fd14fSBrian Behlendorf 	case ZIO_PRIORITY_TRIM:
425084fd14fSBrian Behlendorf 		return (zfs_vdev_trim_min_active);
42669962b56SMatthew Ahrens 	default:
42769962b56SMatthew Ahrens 		panic("invalid priority %u", p);
42869962b56SMatthew Ahrens 	}
42969962b56SMatthew Ahrens }
43069962b56SMatthew Ahrens 
43169962b56SMatthew Ahrens static int
vdev_queue_max_async_writes(spa_t * spa)43273527f44SAlex Reece vdev_queue_max_async_writes(spa_t *spa)
43369962b56SMatthew Ahrens {
43469962b56SMatthew Ahrens 	int writes;
43573527f44SAlex Reece 	uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total;
43669962b56SMatthew Ahrens 	uint64_t min_bytes = zfs_dirty_data_max *
43769962b56SMatthew Ahrens 	    zfs_vdev_async_write_active_min_dirty_percent / 100;
43869962b56SMatthew Ahrens 	uint64_t max_bytes = zfs_dirty_data_max *
43969962b56SMatthew Ahrens 	    zfs_vdev_async_write_active_max_dirty_percent / 100;
44069962b56SMatthew Ahrens 
44173527f44SAlex Reece 	/*
44273527f44SAlex Reece 	 * Sync tasks correspond to interactive user actions. To reduce the
44373527f44SAlex Reece 	 * execution time of those actions we push data out as fast as possible.
44473527f44SAlex Reece 	 */
44573527f44SAlex Reece 	if (spa_has_pending_synctask(spa)) {
44673527f44SAlex Reece 		return (zfs_vdev_async_write_max_active);
44773527f44SAlex Reece 	}
44873527f44SAlex Reece 
44969962b56SMatthew Ahrens 	if (dirty < min_bytes)
45069962b56SMatthew Ahrens 		return (zfs_vdev_async_write_min_active);
45169962b56SMatthew Ahrens 	if (dirty > max_bytes)
45269962b56SMatthew Ahrens 		return (zfs_vdev_async_write_max_active);
45369962b56SMatthew Ahrens 
45469962b56SMatthew Ahrens 	/*
45569962b56SMatthew Ahrens 	 * linear interpolation:
45669962b56SMatthew Ahrens 	 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
45769962b56SMatthew Ahrens 	 * move right by min_bytes
45869962b56SMatthew Ahrens 	 * move up by min_writes
45969962b56SMatthew Ahrens 	 */
46069962b56SMatthew Ahrens 	writes = (dirty - min_bytes) *
46169962b56SMatthew Ahrens 	    (zfs_vdev_async_write_max_active -
46269962b56SMatthew Ahrens 	    zfs_vdev_async_write_min_active) /
46369962b56SMatthew Ahrens 	    (max_bytes - min_bytes) +
46469962b56SMatthew Ahrens 	    zfs_vdev_async_write_min_active;
46569962b56SMatthew Ahrens 	ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
46669962b56SMatthew Ahrens 	ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
46769962b56SMatthew Ahrens 	return (writes);
46869962b56SMatthew Ahrens }
46969962b56SMatthew Ahrens 
47069962b56SMatthew Ahrens static int
vdev_queue_class_max_active(spa_t * spa,zio_priority_t p)47169962b56SMatthew Ahrens vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
47269962b56SMatthew Ahrens {
47369962b56SMatthew Ahrens 	switch (p) {
47469962b56SMatthew Ahrens 	case ZIO_PRIORITY_SYNC_READ:
47569962b56SMatthew Ahrens 		return (zfs_vdev_sync_read_max_active);
47669962b56SMatthew Ahrens 	case ZIO_PRIORITY_SYNC_WRITE:
47769962b56SMatthew Ahrens 		return (zfs_vdev_sync_write_max_active);
47869962b56SMatthew Ahrens 	case ZIO_PRIORITY_ASYNC_READ:
47969962b56SMatthew Ahrens 		return (zfs_vdev_async_read_max_active);
48069962b56SMatthew Ahrens 	case ZIO_PRIORITY_ASYNC_WRITE:
48173527f44SAlex Reece 		return (vdev_queue_max_async_writes(spa));
48269962b56SMatthew Ahrens 	case ZIO_PRIORITY_SCRUB:
48369962b56SMatthew Ahrens 		return (zfs_vdev_scrub_max_active);
4845cabbc6bSPrashanth Sreenivasa 	case ZIO_PRIORITY_REMOVAL:
4855cabbc6bSPrashanth Sreenivasa 		return (zfs_vdev_removal_max_active);
486094e47e9SGeorge Wilson 	case ZIO_PRIORITY_INITIALIZING:
487094e47e9SGeorge Wilson 		return (zfs_vdev_initializing_max_active);
488084fd14fSBrian Behlendorf 	case ZIO_PRIORITY_TRIM:
489084fd14fSBrian Behlendorf 		return (zfs_vdev_trim_max_active);
49069962b56SMatthew Ahrens 	default:
49169962b56SMatthew Ahrens 		panic("invalid priority %u", p);
49269962b56SMatthew Ahrens 	}
49369962b56SMatthew Ahrens }
49469962b56SMatthew Ahrens 
49569962b56SMatthew Ahrens /*
49669962b56SMatthew Ahrens  * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
49769962b56SMatthew Ahrens  * there is no eligible class.
49869962b56SMatthew Ahrens  */
49969962b56SMatthew Ahrens static zio_priority_t
vdev_queue_class_to_issue(vdev_queue_t * vq)50069962b56SMatthew Ahrens vdev_queue_class_to_issue(vdev_queue_t *vq)
50169962b56SMatthew Ahrens {
50269962b56SMatthew Ahrens 	spa_t *spa = vq->vq_vdev->vdev_spa;
50369962b56SMatthew Ahrens 	zio_priority_t p;
50469962b56SMatthew Ahrens 
50569962b56SMatthew Ahrens 	if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
50669962b56SMatthew Ahrens 		return (ZIO_PRIORITY_NUM_QUEUEABLE);
50769962b56SMatthew Ahrens 
50869962b56SMatthew Ahrens 	/* find a queue that has not reached its minimum # outstanding i/os */
50969962b56SMatthew Ahrens 	for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
510fe319232SJustin T. Gibbs 		if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
51169962b56SMatthew Ahrens 		    vq->vq_class[p].vqc_active <
51269962b56SMatthew Ahrens 		    vdev_queue_class_min_active(p))
51369962b56SMatthew Ahrens 			return (p);
51469962b56SMatthew Ahrens 	}
51569962b56SMatthew Ahrens 
51669962b56SMatthew Ahrens 	/*
51769962b56SMatthew Ahrens 	 * If we haven't found a queue, look for one that hasn't reached its
51869962b56SMatthew Ahrens 	 * maximum # outstanding i/os.
51969962b56SMatthew Ahrens 	 */
52069962b56SMatthew Ahrens 	for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
521fe319232SJustin T. Gibbs 		if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
52269962b56SMatthew Ahrens 		    vq->vq_class[p].vqc_active <
52369962b56SMatthew Ahrens 		    vdev_queue_class_max_active(spa, p))
52469962b56SMatthew Ahrens 			return (p);
52569962b56SMatthew Ahrens 	}
52669962b56SMatthew Ahrens 
52769962b56SMatthew Ahrens 	/* No eligible queued i/os */
52869962b56SMatthew Ahrens 	return (ZIO_PRIORITY_NUM_QUEUEABLE);
52969962b56SMatthew Ahrens }
53069962b56SMatthew Ahrens 
5316f708f7cSJeff Bonwick /*
5326f708f7cSJeff Bonwick  * Compute the range spanned by two i/os, which is the endpoint of the last
5336f708f7cSJeff Bonwick  * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
5346f708f7cSJeff Bonwick  * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
5356f708f7cSJeff Bonwick  * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
5366f708f7cSJeff Bonwick  */
5376f708f7cSJeff Bonwick #define	IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
5386f708f7cSJeff Bonwick #define	IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
540fa9e4066Sahrens static zio_t *
vdev_queue_aggregate(vdev_queue_t * vq,zio_t * zio)54169962b56SMatthew Ahrens vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
542fa9e4066Sahrens {
54369962b56SMatthew Ahrens 	zio_t *first, *last, *aio, *dio, *mandatory, *nio;
544a3874b8bSToomas Soome 	zio_link_t *zl = NULL;
54569962b56SMatthew Ahrens 	uint64_t maxgap = 0;
54669962b56SMatthew Ahrens 	uint64_t size;
54769962b56SMatthew Ahrens 	boolean_t stretch = B_FALSE;
548fe319232SJustin T. Gibbs 	avl_tree_t *t = vdev_queue_type_tree(vq, zio->io_type);
54969962b56SMatthew Ahrens 	enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
55069962b56SMatthew Ahrens 
55169962b56SMatthew Ahrens 	if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
55269962b56SMatthew Ahrens 		return (NULL);
554084fd14fSBrian Behlendorf 	/*
555084fd14fSBrian Behlendorf 	 * While TRIM commands could be aggregated based on offset this
556084fd14fSBrian Behlendorf 	 * behavior is disabled until it's determined to be beneficial.
557084fd14fSBrian Behlendorf 	 */
558084fd14fSBrian Behlendorf 	if (zio->io_type == ZIO_TYPE_TRIM && !zfs_vdev_aggregate_trim)
559084fd14fSBrian Behlendorf 		return (NULL);
560084fd14fSBrian Behlendorf 
56169962b56SMatthew Ahrens 	first = last = zio;
56369962b56SMatthew Ahrens 	if (zio->io_type == ZIO_TYPE_READ)
56469962b56SMatthew Ahrens 		maxgap = zfs_vdev_read_gap_limit;
5658ad4d6ddSJeff Bonwick 
56669962b56SMatthew Ahrens 	/*
56769962b56SMatthew Ahrens 	 * We can aggregate I/Os that are sufficiently adjacent and of
56869962b56SMatthew Ahrens 	 * the same flavor, as expressed by the AGG_INHERIT flags.
56969962b56SMatthew Ahrens 	 * The latter requirement is necessary so that certain
57069962b56SMatthew Ahrens 	 * attributes of the I/O, such as whether it's a normal I/O
57169962b56SMatthew Ahrens 	 * or a scrub/resilver, can be preserved in the aggregate.
57269962b56SMatthew Ahrens 	 * We can include optional I/Os, but don't allow them
57369962b56SMatthew Ahrens 	 * to begin a range as they add no benefit in that situation.
57469962b56SMatthew Ahrens 	 */
575f94275ceSAdam Leventhal 
57669962b56SMatthew Ahrens 	/*
57769962b56SMatthew Ahrens 	 * We keep track of the last non-optional I/O.
57869962b56SMatthew Ahrens 	 */
57969962b56SMatthew Ahrens 	mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
580f94275ceSAdam Leventhal 
58169962b56SMatthew Ahrens 	/*
58269962b56SMatthew Ahrens 	 * Walk backwards through sufficiently contiguous I/Os
5835b062782SMatthew Ahrens 	 * recording the last non-optional I/O.
58469962b56SMatthew Ahrens 	 */
58569962b56SMatthew Ahrens 	while ((dio = AVL_PREV(t, first)) != NULL &&
58669962b56SMatthew Ahrens 	    (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
58769962b56SMatthew Ahrens 	    IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
5885cabbc6bSPrashanth Sreenivasa 	    IO_GAP(dio, first) <= maxgap &&
5895cabbc6bSPrashanth Sreenivasa 	    dio->io_type == zio->io_type) {
59069962b56SMatthew Ahrens 		first = dio;
59169962b56SMatthew Ahrens 		if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
59269962b56SMatthew Ahrens 			mandatory = first;
59369962b56SMatthew Ahrens 	}
594f94275ceSAdam Leventhal 
59569962b56SMatthew Ahrens 	/*
59669962b56SMatthew Ahrens 	 * Skip any initial optional I/Os.
59769962b56SMatthew Ahrens 	 */
59869962b56SMatthew Ahrens 	while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
59969962b56SMatthew Ahrens 		first = AVL_NEXT(t, first);
60069962b56SMatthew Ahrens 		ASSERT(first != NULL);
60169962b56SMatthew Ahrens 	}
6026f708f7cSJeff Bonwick