Deleted Added
sdiff udiff text old ( 339034 ) new ( 339105 )
full compact
vdev_queue.c (339034) vdev_queue.c (339105)
1/*
2 * CDDL HEADER START
3 *
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.
7 *
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.
12 *
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]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
24 */
25
26/*
1/*
2 * CDDL HEADER START
3 *
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.
7 *
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.
12 *
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]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
24 */
25
26/*
27 * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
27 * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
28 * Copyright (c) 2014 Integros [integros.com]
29 */
30
31#include <sys/zfs_context.h>
32#include <sys/vdev_impl.h>
33#include <sys/spa_impl.h>
34#include <sys/zio.h>
35#include <sys/avl.h>
36#include <sys/dsl_pool.h>
37#include <sys/metaslab_impl.h>
38#include <sys/abd.h>
39
40/*
41 * ZFS I/O Scheduler
42 * ---------------
43 *
44 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The
45 * I/O scheduler determines when and in what order those operations are
46 * issued. The I/O scheduler divides operations into six I/O classes
47 * prioritized in the following order: sync read, sync write, async read,
48 * async write, scrub/resilver and trim. Each queue defines the minimum and
49 * maximum number of concurrent operations that may be issued to the device.
50 * In addition, the device has an aggregate maximum. Note that the sum of the
51 * per-queue minimums must not exceed the aggregate maximum, and if the
52 * aggregate maximum is equal to or greater than the sum of the per-queue
53 * maximums, the per-queue minimum has no effect.
54 *
55 * For many physical devices, throughput increases with the number of
56 * concurrent operations, but latency typically suffers. Further, physical
57 * devices typically have a limit at which more concurrent operations have no
58 * effect on throughput or can actually cause it to decrease.
59 *
60 * The scheduler selects the next operation to issue by first looking for an
61 * I/O class whose minimum has not been satisfied. Once all are satisfied and
62 * the aggregate maximum has not been hit, the scheduler looks for classes
63 * whose maximum has not been satisfied. Iteration through the I/O classes is
64 * done in the order specified above. No further operations are issued if the
65 * aggregate maximum number of concurrent operations has been hit or if there
66 * are no operations queued for an I/O class that has not hit its maximum.
67 * Every time an I/O is queued or an operation completes, the I/O scheduler
68 * looks for new operations to issue.
69 *
70 * All I/O classes have a fixed maximum number of outstanding operations
71 * except for the async write class. Asynchronous writes represent the data
72 * that is committed to stable storage during the syncing stage for
73 * transaction groups (see txg.c). Transaction groups enter the syncing state
74 * periodically so the number of queued async writes will quickly burst up and
75 * then bleed down to zero. Rather than servicing them as quickly as possible,
76 * the I/O scheduler changes the maximum number of active async write I/Os
77 * according to the amount of dirty data in the pool (see dsl_pool.c). Since
78 * both throughput and latency typically increase with the number of
79 * concurrent operations issued to physical devices, reducing the burstiness
80 * in the number of concurrent operations also stabilizes the response time of
81 * operations from other -- and in particular synchronous -- queues. In broad
82 * strokes, the I/O scheduler will issue more concurrent operations from the
83 * async write queue as there's more dirty data in the pool.
84 *
85 * Async Writes
86 *
87 * The number of concurrent operations issued for the async write I/O class
88 * follows a piece-wise linear function defined by a few adjustable points.
89 *
90 * | o---------| <-- zfs_vdev_async_write_max_active
91 * ^ | /^ |
92 * | | / | |
93 * active | / | |
94 * I/O | / | |
95 * count | / | |
96 * | / | |
97 * |------------o | | <-- zfs_vdev_async_write_min_active
98 * 0|____________^______|_________|
99 * 0% | | 100% of zfs_dirty_data_max
100 * | |
101 * | `-- zfs_vdev_async_write_active_max_dirty_percent
102 * `--------- zfs_vdev_async_write_active_min_dirty_percent
103 *
104 * Until the amount of dirty data exceeds a minimum percentage of the dirty
105 * data allowed in the pool, the I/O scheduler will limit the number of
106 * concurrent operations to the minimum. As that threshold is crossed, the
107 * number of concurrent operations issued increases linearly to the maximum at
108 * the specified maximum percentage of the dirty data allowed in the pool.
109 *
110 * Ideally, the amount of dirty data on a busy pool will stay in the sloped
111 * part of the function between zfs_vdev_async_write_active_min_dirty_percent
112 * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the
113 * maximum percentage, this indicates that the rate of incoming data is
114 * greater than the rate that the backend storage can handle. In this case, we
115 * must further throttle incoming writes (see dmu_tx_delay() for details).
116 */
117
118/*
119 * The maximum number of I/Os active to each device. Ideally, this will be >=
120 * the sum of each queue's max_active. It must be at least the sum of each
121 * queue's min_active.
122 */
123uint32_t zfs_vdev_max_active = 1000;
124
125/*
126 * Per-queue limits on the number of I/Os active to each device. If the
127 * sum of the queue's max_active is < zfs_vdev_max_active, then the
128 * min_active comes into play. We will send min_active from each queue,
129 * and then select from queues in the order defined by zio_priority_t.
130 *
131 * In general, smaller max_active's will lead to lower latency of synchronous
132 * operations. Larger max_active's may lead to higher overall throughput,
133 * depending on underlying storage.
134 *
135 * The ratio of the queues' max_actives determines the balance of performance
136 * between reads, writes, and scrubs. E.g., increasing
137 * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete
138 * more quickly, but reads and writes to have higher latency and lower
139 * throughput.
140 */
141uint32_t zfs_vdev_sync_read_min_active = 10;
142uint32_t zfs_vdev_sync_read_max_active = 10;
143uint32_t zfs_vdev_sync_write_min_active = 10;
144uint32_t zfs_vdev_sync_write_max_active = 10;
145uint32_t zfs_vdev_async_read_min_active = 1;
146uint32_t zfs_vdev_async_read_max_active = 3;
147uint32_t zfs_vdev_async_write_min_active = 1;
148uint32_t zfs_vdev_async_write_max_active = 10;
149uint32_t zfs_vdev_scrub_min_active = 1;
150uint32_t zfs_vdev_scrub_max_active = 2;
151uint32_t zfs_vdev_trim_min_active = 1;
152/*
153 * TRIM max active is large in comparison to the other values due to the fact
154 * that TRIM IOs are coalesced at the device layer. This value is set such
155 * that a typical SSD can process the queued IOs in a single request.
156 */
157uint32_t zfs_vdev_trim_max_active = 64;
158uint32_t zfs_vdev_removal_min_active = 1;
159uint32_t zfs_vdev_removal_max_active = 2;
160
161
162/*
163 * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent
164 * dirty data, use zfs_vdev_async_write_min_active. When it has more than
165 * zfs_vdev_async_write_active_max_dirty_percent, use
166 * zfs_vdev_async_write_max_active. The value is linearly interpolated
167 * between min and max.
168 */
169int zfs_vdev_async_write_active_min_dirty_percent = 30;
170int zfs_vdev_async_write_active_max_dirty_percent = 60;
171
172/*
173 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
174 * For read I/Os, we also aggregate across small adjacency gaps; for writes
175 * we include spans of optional I/Os to aid aggregation at the disk even when
176 * they aren't able to help us aggregate at this level.
177 */
178int zfs_vdev_aggregation_limit = 1 << 20;
179int zfs_vdev_read_gap_limit = 32 << 10;
180int zfs_vdev_write_gap_limit = 4 << 10;
181
182/*
183 * Define the queue depth percentage for each top-level. This percentage is
184 * used in conjunction with zfs_vdev_async_max_active to determine how many
185 * allocations a specific top-level vdev should handle. Once the queue depth
186 * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100
187 * then allocator will stop allocating blocks on that top-level device.
188 * The default kernel setting is 1000% which will yield 100 allocations per
189 * device. For userland testing, the default setting is 300% which equates
190 * to 30 allocations per device.
191 */
192#ifdef _KERNEL
193int zfs_vdev_queue_depth_pct = 1000;
194#else
195int zfs_vdev_queue_depth_pct = 300;
196#endif
197
28 * Copyright (c) 2014 Integros [integros.com]
29 */
30
31#include <sys/zfs_context.h>
32#include <sys/vdev_impl.h>
33#include <sys/spa_impl.h>
34#include <sys/zio.h>
35#include <sys/avl.h>
36#include <sys/dsl_pool.h>
37#include <sys/metaslab_impl.h>
38#include <sys/abd.h>
39
40/*
41 * ZFS I/O Scheduler
42 * ---------------
43 *
44 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The
45 * I/O scheduler determines when and in what order those operations are
46 * issued. The I/O scheduler divides operations into six I/O classes
47 * prioritized in the following order: sync read, sync write, async read,
48 * async write, scrub/resilver and trim. Each queue defines the minimum and
49 * maximum number of concurrent operations that may be issued to the device.
50 * In addition, the device has an aggregate maximum. Note that the sum of the
51 * per-queue minimums must not exceed the aggregate maximum, and if the
52 * aggregate maximum is equal to or greater than the sum of the per-queue
53 * maximums, the per-queue minimum has no effect.
54 *
55 * For many physical devices, throughput increases with the number of
56 * concurrent operations, but latency typically suffers. Further, physical
57 * devices typically have a limit at which more concurrent operations have no
58 * effect on throughput or can actually cause it to decrease.
59 *
60 * The scheduler selects the next operation to issue by first looking for an
61 * I/O class whose minimum has not been satisfied. Once all are satisfied and
62 * the aggregate maximum has not been hit, the scheduler looks for classes
63 * whose maximum has not been satisfied. Iteration through the I/O classes is
64 * done in the order specified above. No further operations are issued if the
65 * aggregate maximum number of concurrent operations has been hit or if there
66 * are no operations queued for an I/O class that has not hit its maximum.
67 * Every time an I/O is queued or an operation completes, the I/O scheduler
68 * looks for new operations to issue.
69 *
70 * All I/O classes have a fixed maximum number of outstanding operations
71 * except for the async write class. Asynchronous writes represent the data
72 * that is committed to stable storage during the syncing stage for
73 * transaction groups (see txg.c). Transaction groups enter the syncing state
74 * periodically so the number of queued async writes will quickly burst up and
75 * then bleed down to zero. Rather than servicing them as quickly as possible,
76 * the I/O scheduler changes the maximum number of active async write I/Os
77 * according to the amount of dirty data in the pool (see dsl_pool.c). Since
78 * both throughput and latency typically increase with the number of
79 * concurrent operations issued to physical devices, reducing the burstiness
80 * in the number of concurrent operations also stabilizes the response time of
81 * operations from other -- and in particular synchronous -- queues. In broad
82 * strokes, the I/O scheduler will issue more concurrent operations from the
83 * async write queue as there's more dirty data in the pool.
84 *
85 * Async Writes
86 *
87 * The number of concurrent operations issued for the async write I/O class
88 * follows a piece-wise linear function defined by a few adjustable points.
89 *
90 * | o---------| <-- zfs_vdev_async_write_max_active
91 * ^ | /^ |
92 * | | / | |
93 * active | / | |
94 * I/O | / | |
95 * count | / | |
96 * | / | |
97 * |------------o | | <-- zfs_vdev_async_write_min_active
98 * 0|____________^______|_________|
99 * 0% | | 100% of zfs_dirty_data_max
100 * | |
101 * | `-- zfs_vdev_async_write_active_max_dirty_percent
102 * `--------- zfs_vdev_async_write_active_min_dirty_percent
103 *
104 * Until the amount of dirty data exceeds a minimum percentage of the dirty
105 * data allowed in the pool, the I/O scheduler will limit the number of
106 * concurrent operations to the minimum. As that threshold is crossed, the
107 * number of concurrent operations issued increases linearly to the maximum at
108 * the specified maximum percentage of the dirty data allowed in the pool.
109 *
110 * Ideally, the amount of dirty data on a busy pool will stay in the sloped
111 * part of the function between zfs_vdev_async_write_active_min_dirty_percent
112 * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the
113 * maximum percentage, this indicates that the rate of incoming data is
114 * greater than the rate that the backend storage can handle. In this case, we
115 * must further throttle incoming writes (see dmu_tx_delay() for details).
116 */
117
118/*
119 * The maximum number of I/Os active to each device. Ideally, this will be >=
120 * the sum of each queue's max_active. It must be at least the sum of each
121 * queue's min_active.
122 */
123uint32_t zfs_vdev_max_active = 1000;
124
125/*
126 * Per-queue limits on the number of I/Os active to each device. If the
127 * sum of the queue's max_active is < zfs_vdev_max_active, then the
128 * min_active comes into play. We will send min_active from each queue,
129 * and then select from queues in the order defined by zio_priority_t.
130 *
131 * In general, smaller max_active's will lead to lower latency of synchronous
132 * operations. Larger max_active's may lead to higher overall throughput,
133 * depending on underlying storage.
134 *
135 * The ratio of the queues' max_actives determines the balance of performance
136 * between reads, writes, and scrubs. E.g., increasing
137 * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete
138 * more quickly, but reads and writes to have higher latency and lower
139 * throughput.
140 */
141uint32_t zfs_vdev_sync_read_min_active = 10;
142uint32_t zfs_vdev_sync_read_max_active = 10;
143uint32_t zfs_vdev_sync_write_min_active = 10;
144uint32_t zfs_vdev_sync_write_max_active = 10;
145uint32_t zfs_vdev_async_read_min_active = 1;
146uint32_t zfs_vdev_async_read_max_active = 3;
147uint32_t zfs_vdev_async_write_min_active = 1;
148uint32_t zfs_vdev_async_write_max_active = 10;
149uint32_t zfs_vdev_scrub_min_active = 1;
150uint32_t zfs_vdev_scrub_max_active = 2;
151uint32_t zfs_vdev_trim_min_active = 1;
152/*
153 * TRIM max active is large in comparison to the other values due to the fact
154 * that TRIM IOs are coalesced at the device layer. This value is set such
155 * that a typical SSD can process the queued IOs in a single request.
156 */
157uint32_t zfs_vdev_trim_max_active = 64;
158uint32_t zfs_vdev_removal_min_active = 1;
159uint32_t zfs_vdev_removal_max_active = 2;
160
161
162/*
163 * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent
164 * dirty data, use zfs_vdev_async_write_min_active. When it has more than
165 * zfs_vdev_async_write_active_max_dirty_percent, use
166 * zfs_vdev_async_write_max_active. The value is linearly interpolated
167 * between min and max.
168 */
169int zfs_vdev_async_write_active_min_dirty_percent = 30;
170int zfs_vdev_async_write_active_max_dirty_percent = 60;
171
172/*
173 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
174 * For read I/Os, we also aggregate across small adjacency gaps; for writes
175 * we include spans of optional I/Os to aid aggregation at the disk even when
176 * they aren't able to help us aggregate at this level.
177 */
178int zfs_vdev_aggregation_limit = 1 << 20;
179int zfs_vdev_read_gap_limit = 32 << 10;
180int zfs_vdev_write_gap_limit = 4 << 10;
181
182/*
183 * Define the queue depth percentage for each top-level. This percentage is
184 * used in conjunction with zfs_vdev_async_max_active to determine how many
185 * allocations a specific top-level vdev should handle. Once the queue depth
186 * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100
187 * then allocator will stop allocating blocks on that top-level device.
188 * The default kernel setting is 1000% which will yield 100 allocations per
189 * device. For userland testing, the default setting is 300% which equates
190 * to 30 allocations per device.
191 */
192#ifdef _KERNEL
193int zfs_vdev_queue_depth_pct = 1000;
194#else
195int zfs_vdev_queue_depth_pct = 300;
196#endif
197
198/*
199 * When performing allocations for a given metaslab, we want to make sure that
200 * there are enough IOs to aggregate together to improve throughput. We want to
201 * ensure that there are at least 128k worth of IOs that can be aggregated, and
202 * we assume that the average allocation size is 4k, so we need the queue depth
203 * to be 32 per allocator to get good aggregation of sequential writes.
204 */
205int zfs_vdev_def_queue_depth = 32;
198
199#ifdef __FreeBSD__
200#ifdef _KERNEL
201SYSCTL_DECL(_vfs_zfs_vdev);
202
203static int sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS);
204SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_min_dirty_percent,
205 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
206 sysctl_zfs_async_write_active_min_dirty_percent, "I",
207 "Percentage of async write dirty data below which "
208 "async_write_min_active is used.");
209
210static int sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS);
211SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_max_dirty_percent,
212 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
213 sysctl_zfs_async_write_active_max_dirty_percent, "I",
214 "Percentage of async write dirty data above which "
215 "async_write_max_active is used.");
216
217SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RWTUN,
218 &zfs_vdev_max_active, 0,
219 "The maximum number of I/Os of all types active for each device.");
220
221#define ZFS_VDEV_QUEUE_KNOB_MIN(name) \
222SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, CTLFLAG_RWTUN,\
223 &zfs_vdev_ ## name ## _min_active, 0, \
224 "Initial number of I/O requests of type " #name \
225 " active for each device");
226
227#define ZFS_VDEV_QUEUE_KNOB_MAX(name) \
228SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, CTLFLAG_RWTUN,\
229 &zfs_vdev_ ## name ## _max_active, 0, \
230 "Maximum number of I/O requests of type " #name \
231 " active for each device");
232
233ZFS_VDEV_QUEUE_KNOB_MIN(sync_read);
234ZFS_VDEV_QUEUE_KNOB_MAX(sync_read);
235ZFS_VDEV_QUEUE_KNOB_MIN(sync_write);
236ZFS_VDEV_QUEUE_KNOB_MAX(sync_write);
237ZFS_VDEV_QUEUE_KNOB_MIN(async_read);
238ZFS_VDEV_QUEUE_KNOB_MAX(async_read);
239ZFS_VDEV_QUEUE_KNOB_MIN(async_write);
240ZFS_VDEV_QUEUE_KNOB_MAX(async_write);
241ZFS_VDEV_QUEUE_KNOB_MIN(scrub);
242ZFS_VDEV_QUEUE_KNOB_MAX(scrub);
243ZFS_VDEV_QUEUE_KNOB_MIN(trim);
244ZFS_VDEV_QUEUE_KNOB_MAX(trim);
245
246#undef ZFS_VDEV_QUEUE_KNOB
247
248SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RWTUN,
249 &zfs_vdev_aggregation_limit, 0,
250 "I/O requests are aggregated up to this size");
251SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RWTUN,
252 &zfs_vdev_read_gap_limit, 0,
253 "Acceptable gap between two reads being aggregated");
254SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RWTUN,
255 &zfs_vdev_write_gap_limit, 0,
256 "Acceptable gap between two writes being aggregated");
257SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, queue_depth_pct, CTLFLAG_RWTUN,
258 &zfs_vdev_queue_depth_pct, 0,
259 "Queue depth percentage for each top-level");
260
261static int
262sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS)
263{
264 int val, err;
265
266 val = zfs_vdev_async_write_active_min_dirty_percent;
267 err = sysctl_handle_int(oidp, &val, 0, req);
268 if (err != 0 || req->newptr == NULL)
269 return (err);
270
271 if (val < 0 || val > 100 ||
272 val >= zfs_vdev_async_write_active_max_dirty_percent)
273 return (EINVAL);
274
275 zfs_vdev_async_write_active_min_dirty_percent = val;
276
277 return (0);
278}
279
280static int
281sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS)
282{
283 int val, err;
284
285 val = zfs_vdev_async_write_active_max_dirty_percent;
286 err = sysctl_handle_int(oidp, &val, 0, req);
287 if (err != 0 || req->newptr == NULL)
288 return (err);
289
290 if (val < 0 || val > 100 ||
291 val <= zfs_vdev_async_write_active_min_dirty_percent)
292 return (EINVAL);
293
294 zfs_vdev_async_write_active_max_dirty_percent = val;
295
296 return (0);
297}
298#endif
299#endif
300
301int
302vdev_queue_offset_compare(const void *x1, const void *x2)
303{
304 const zio_t *z1 = x1;
305 const zio_t *z2 = x2;
306
307 if (z1->io_offset < z2->io_offset)
308 return (-1);
309 if (z1->io_offset > z2->io_offset)
310 return (1);
311
312 if (z1 < z2)
313 return (-1);
314 if (z1 > z2)
315 return (1);
316
317 return (0);
318}
319
320static inline avl_tree_t *
321vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p)
322{
323 return (&vq->vq_class[p].vqc_queued_tree);
324}
325
326static inline avl_tree_t *
327vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t)
328{
329 if (t == ZIO_TYPE_READ)
330 return (&vq->vq_read_offset_tree);
331 else if (t == ZIO_TYPE_WRITE)
332 return (&vq->vq_write_offset_tree);
333 else
334 return (NULL);
335}
336
337int
338vdev_queue_timestamp_compare(const void *x1, const void *x2)
339{
340 const zio_t *z1 = x1;
341 const zio_t *z2 = x2;
342
343 if (z1->io_timestamp < z2->io_timestamp)
344 return (-1);
345 if (z1->io_timestamp > z2->io_timestamp)
346 return (1);
347
348 if (z1->io_offset < z2->io_offset)
349 return (-1);
350 if (z1->io_offset > z2->io_offset)
351 return (1);
352
353 if (z1 < z2)
354 return (-1);
355 if (z1 > z2)
356 return (1);
357
358 return (0);
359}
360
361void
362vdev_queue_init(vdev_t *vd)
363{
364 vdev_queue_t *vq = &vd->vdev_queue;
365
366 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
367 vq->vq_vdev = vd;
368
369 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
370 sizeof (zio_t), offsetof(struct zio, io_queue_node));
371 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ),
372 vdev_queue_offset_compare, sizeof (zio_t),
373 offsetof(struct zio, io_offset_node));
374 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE),
375 vdev_queue_offset_compare, sizeof (zio_t),
376 offsetof(struct zio, io_offset_node));
377
378 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
379 int (*compfn) (const void *, const void *);
380
381 /*
382 * The synchronous i/o queues are dispatched in FIFO rather
383 * than LBA order. This provides more consistent latency for
384 * these i/os.
385 */
386 if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE)
387 compfn = vdev_queue_timestamp_compare;
388 else
389 compfn = vdev_queue_offset_compare;
390
391 avl_create(vdev_queue_class_tree(vq, p), compfn,
392 sizeof (zio_t), offsetof(struct zio, io_queue_node));
393 }
394
395 vq->vq_lastoffset = 0;
396}
397
398void
399vdev_queue_fini(vdev_t *vd)
400{
401 vdev_queue_t *vq = &vd->vdev_queue;
402
403 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
404 avl_destroy(vdev_queue_class_tree(vq, p));
405 avl_destroy(&vq->vq_active_tree);
406 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
407 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE));
408
409 mutex_destroy(&vq->vq_lock);
410}
411
412static void
413vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
414{
415 spa_t *spa = zio->io_spa;
416 avl_tree_t *qtt;
417
418 ASSERT(MUTEX_HELD(&vq->vq_lock));
419 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
420 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
421 qtt = vdev_queue_type_tree(vq, zio->io_type);
422 if (qtt)
423 avl_add(qtt, zio);
424
425#ifdef illumos
426 mutex_enter(&spa->spa_iokstat_lock);
427 spa->spa_queue_stats[zio->io_priority].spa_queued++;
428 if (spa->spa_iokstat != NULL)
429 kstat_waitq_enter(spa->spa_iokstat->ks_data);
430 mutex_exit(&spa->spa_iokstat_lock);
431#endif
432}
433
434static void
435vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
436{
437 spa_t *spa = zio->io_spa;
438 avl_tree_t *qtt;
439
440 ASSERT(MUTEX_HELD(&vq->vq_lock));
441 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
442 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
443 qtt = vdev_queue_type_tree(vq, zio->io_type);
444 if (qtt)
445 avl_remove(qtt, zio);
446
447#ifdef illumos
448 mutex_enter(&spa->spa_iokstat_lock);
449 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
450 spa->spa_queue_stats[zio->io_priority].spa_queued--;
451 if (spa->spa_iokstat != NULL)
452 kstat_waitq_exit(spa->spa_iokstat->ks_data);
453 mutex_exit(&spa->spa_iokstat_lock);
454#endif
455}
456
457static void
458vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
459{
460 spa_t *spa = zio->io_spa;
461 ASSERT(MUTEX_HELD(&vq->vq_lock));
462 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
463 vq->vq_class[zio->io_priority].vqc_active++;
464 avl_add(&vq->vq_active_tree, zio);
465
466#ifdef illumos
467 mutex_enter(&spa->spa_iokstat_lock);
468 spa->spa_queue_stats[zio->io_priority].spa_active++;
469 if (spa->spa_iokstat != NULL)
470 kstat_runq_enter(spa->spa_iokstat->ks_data);
471 mutex_exit(&spa->spa_iokstat_lock);
472#endif
473}
474
475static void
476vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
477{
478 spa_t *spa = zio->io_spa;
479 ASSERT(MUTEX_HELD(&vq->vq_lock));
480 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
481 vq->vq_class[zio->io_priority].vqc_active--;
482 avl_remove(&vq->vq_active_tree, zio);
483
484#ifdef illumos
485 mutex_enter(&spa->spa_iokstat_lock);
486 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
487 spa->spa_queue_stats[zio->io_priority].spa_active--;
488 if (spa->spa_iokstat != NULL) {
489 kstat_io_t *ksio = spa->spa_iokstat->ks_data;
490
491 kstat_runq_exit(spa->spa_iokstat->ks_data);
492 if (zio->io_type == ZIO_TYPE_READ) {
493 ksio->reads++;
494 ksio->nread += zio->io_size;
495 } else if (zio->io_type == ZIO_TYPE_WRITE) {
496 ksio->writes++;
497 ksio->nwritten += zio->io_size;
498 }
499 }
500 mutex_exit(&spa->spa_iokstat_lock);
501#endif
502}
503
504static void
505vdev_queue_agg_io_done(zio_t *aio)
506{
507 if (aio->io_type == ZIO_TYPE_READ) {
508 zio_t *pio;
509 zio_link_t *zl = NULL;
510 while ((pio = zio_walk_parents(aio, &zl)) != NULL) {
511 abd_copy_off(pio->io_abd, aio->io_abd,
512 0, pio->io_offset - aio->io_offset, pio->io_size);
513 }
514 }
515
516 abd_free(aio->io_abd);
517}
518
519static int
520vdev_queue_class_min_active(zio_priority_t p)
521{
522 switch (p) {
523 case ZIO_PRIORITY_SYNC_READ:
524 return (zfs_vdev_sync_read_min_active);
525 case ZIO_PRIORITY_SYNC_WRITE:
526 return (zfs_vdev_sync_write_min_active);
527 case ZIO_PRIORITY_ASYNC_READ:
528 return (zfs_vdev_async_read_min_active);
529 case ZIO_PRIORITY_ASYNC_WRITE:
530 return (zfs_vdev_async_write_min_active);
531 case ZIO_PRIORITY_SCRUB:
532 return (zfs_vdev_scrub_min_active);
533 case ZIO_PRIORITY_TRIM:
534 return (zfs_vdev_trim_min_active);
535 case ZIO_PRIORITY_REMOVAL:
536 return (zfs_vdev_removal_min_active);
537 default:
538 panic("invalid priority %u", p);
539 return (0);
540 }
541}
542
543static __noinline int
544vdev_queue_max_async_writes(spa_t *spa)
545{
546 int writes;
547 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total;
548 uint64_t min_bytes = zfs_dirty_data_max *
549 zfs_vdev_async_write_active_min_dirty_percent / 100;
550 uint64_t max_bytes = zfs_dirty_data_max *
551 zfs_vdev_async_write_active_max_dirty_percent / 100;
552
553 /*
554 * Sync tasks correspond to interactive user actions. To reduce the
555 * execution time of those actions we push data out as fast as possible.
556 */
557 if (spa_has_pending_synctask(spa)) {
558 return (zfs_vdev_async_write_max_active);
559 }
560
561 if (dirty < min_bytes)
562 return (zfs_vdev_async_write_min_active);
563 if (dirty > max_bytes)
564 return (zfs_vdev_async_write_max_active);
565
566 /*
567 * linear interpolation:
568 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
569 * move right by min_bytes
570 * move up by min_writes
571 */
572 writes = (dirty - min_bytes) *
573 (zfs_vdev_async_write_max_active -
574 zfs_vdev_async_write_min_active) /
575 (max_bytes - min_bytes) +
576 zfs_vdev_async_write_min_active;
577 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
578 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
579 return (writes);
580}
581
582static int
583vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
584{
585 switch (p) {
586 case ZIO_PRIORITY_SYNC_READ:
587 return (zfs_vdev_sync_read_max_active);
588 case ZIO_PRIORITY_SYNC_WRITE:
589 return (zfs_vdev_sync_write_max_active);
590 case ZIO_PRIORITY_ASYNC_READ:
591 return (zfs_vdev_async_read_max_active);
592 case ZIO_PRIORITY_ASYNC_WRITE:
593 return (vdev_queue_max_async_writes(spa));
594 case ZIO_PRIORITY_SCRUB:
595 return (zfs_vdev_scrub_max_active);
596 case ZIO_PRIORITY_TRIM:
597 return (zfs_vdev_trim_max_active);
598 case ZIO_PRIORITY_REMOVAL:
599 return (zfs_vdev_removal_max_active);
600 default:
601 panic("invalid priority %u", p);
602 return (0);
603 }
604}
605
606/*
607 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
608 * there is no eligible class.
609 */
610static zio_priority_t
611vdev_queue_class_to_issue(vdev_queue_t *vq)
612{
613 spa_t *spa = vq->vq_vdev->vdev_spa;
614 zio_priority_t p;
615
616 ASSERT(MUTEX_HELD(&vq->vq_lock));
617
618 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
619 return (ZIO_PRIORITY_NUM_QUEUEABLE);
620
621 /* find a queue that has not reached its minimum # outstanding i/os */
622 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
623 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
624 vq->vq_class[p].vqc_active <
625 vdev_queue_class_min_active(p))
626 return (p);
627 }
628
629 /*
630 * If we haven't found a queue, look for one that hasn't reached its
631 * maximum # outstanding i/os.
632 */
633 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
634 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
635 vq->vq_class[p].vqc_active <
636 vdev_queue_class_max_active(spa, p))
637 return (p);
638 }
639
640 /* No eligible queued i/os */
641 return (ZIO_PRIORITY_NUM_QUEUEABLE);
642}
643
644/*
645 * Compute the range spanned by two i/os, which is the endpoint of the last
646 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
647 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
648 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
649 */
650#define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
651#define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
652
653static zio_t *
654vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
655{
656 zio_t *first, *last, *aio, *dio, *mandatory, *nio;
657 uint64_t maxgap = 0;
658 uint64_t size;
659 boolean_t stretch;
660 avl_tree_t *t;
661 enum zio_flag flags;
662
663 ASSERT(MUTEX_HELD(&vq->vq_lock));
664
665 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
666 return (NULL);
667
668 first = last = zio;
669
670 if (zio->io_type == ZIO_TYPE_READ)
671 maxgap = zfs_vdev_read_gap_limit;
672
673 /*
674 * We can aggregate I/Os that are sufficiently adjacent and of
675 * the same flavor, as expressed by the AGG_INHERIT flags.
676 * The latter requirement is necessary so that certain
677 * attributes of the I/O, such as whether it's a normal I/O
678 * or a scrub/resilver, can be preserved in the aggregate.
679 * We can include optional I/Os, but don't allow them
680 * to begin a range as they add no benefit in that situation.
681 */
682
683 /*
684 * We keep track of the last non-optional I/O.
685 */
686 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
687
688 /*
689 * Walk backwards through sufficiently contiguous I/Os
690 * recording the last non-optional I/O.
691 */
692 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
693 t = vdev_queue_type_tree(vq, zio->io_type);
694 while (t != NULL && (dio = AVL_PREV(t, first)) != NULL &&
695 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
696 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
697 IO_GAP(dio, first) <= maxgap &&
698 dio->io_type == zio->io_type) {
699 first = dio;
700 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
701 mandatory = first;
702 }
703
704 /*
705 * Skip any initial optional I/Os.
706 */
707 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
708 first = AVL_NEXT(t, first);
709 ASSERT(first != NULL);
710 }
711
712 /*
713 * Walk forward through sufficiently contiguous I/Os.
714 * The aggregation limit does not apply to optional i/os, so that
715 * we can issue contiguous writes even if they are larger than the
716 * aggregation limit.
717 */
718 while ((dio = AVL_NEXT(t, last)) != NULL &&
719 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
720 (IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit ||
721 (dio->io_flags & ZIO_FLAG_OPTIONAL)) &&
722 IO_GAP(last, dio) <= maxgap &&
723 dio->io_type == zio->io_type) {
724 last = dio;
725 if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
726 mandatory = last;
727 }
728
729 /*
730 * Now that we've established the range of the I/O aggregation
731 * we must decide what to do with trailing optional I/Os.
732 * For reads, there's nothing to do. While we are unable to
733 * aggregate further, it's possible that a trailing optional
734 * I/O would allow the underlying device to aggregate with
735 * subsequent I/Os. We must therefore determine if the next
736 * non-optional I/O is close enough to make aggregation
737 * worthwhile.
738 */
739 stretch = B_FALSE;
740 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
741 zio_t *nio = last;
742 while ((dio = AVL_NEXT(t, nio)) != NULL &&
743 IO_GAP(nio, dio) == 0 &&
744 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
745 nio = dio;
746 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
747 stretch = B_TRUE;
748 break;
749 }
750 }
751 }
752
753 if (stretch) {
754 /*
755 * We are going to include an optional io in our aggregated
756 * span, thus closing the write gap. Only mandatory i/os can
757 * start aggregated spans, so make sure that the next i/o
758 * after our span is mandatory.
759 */
760 dio = AVL_NEXT(t, last);
761 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
762 } else {
763 /* do not include the optional i/o */
764 while (last != mandatory && last != first) {
765 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
766 last = AVL_PREV(t, last);
767 ASSERT(last != NULL);
768 }
769 }
770
771 if (first == last)
772 return (NULL);
773
774 size = IO_SPAN(first, last);
775 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
776
777 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
778 abd_alloc_for_io(size, B_TRUE), size, first->io_type,
779 zio->io_priority, flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
780 vdev_queue_agg_io_done, NULL);
781 aio->io_timestamp = first->io_timestamp;
782
783 nio = first;
784 do {
785 dio = nio;
786 nio = AVL_NEXT(t, dio);
787 ASSERT3U(dio->io_type, ==, aio->io_type);
788
789 if (dio->io_flags & ZIO_FLAG_NODATA) {
790 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
791 abd_zero_off(aio->io_abd,
792 dio->io_offset - aio->io_offset, dio->io_size);
793 } else if (dio->io_type == ZIO_TYPE_WRITE) {
794 abd_copy_off(aio->io_abd, dio->io_abd,
795 dio->io_offset - aio->io_offset, 0, dio->io_size);
796 }
797
798 zio_add_child(dio, aio);
799 vdev_queue_io_remove(vq, dio);
800 zio_vdev_io_bypass(dio);
801 zio_execute(dio);
802 } while (dio != last);
803
804 return (aio);
805}
806
807static zio_t *
808vdev_queue_io_to_issue(vdev_queue_t *vq)
809{
810 zio_t *zio, *aio;
811 zio_priority_t p;
812 avl_index_t idx;
813 avl_tree_t *tree;
814 zio_t search;
815
816again:
817 ASSERT(MUTEX_HELD(&vq->vq_lock));
818
819 p = vdev_queue_class_to_issue(vq);
820
821 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
822 /* No eligible queued i/os */
823 return (NULL);
824 }
825
826 /*
827 * For LBA-ordered queues (async / scrub), issue the i/o which follows
828 * the most recently issued i/o in LBA (offset) order.
829 *
830 * For FIFO queues (sync), issue the i/o with the lowest timestamp.
831 */
832 tree = vdev_queue_class_tree(vq, p);
833 search.io_timestamp = 0;
834 search.io_offset = vq->vq_last_offset + 1;
835 VERIFY3P(avl_find(tree, &search, &idx), ==, NULL);
836 zio = avl_nearest(tree, idx, AVL_AFTER);
837 if (zio == NULL)
838 zio = avl_first(tree);
839 ASSERT3U(zio->io_priority, ==, p);
840
841 aio = vdev_queue_aggregate(vq, zio);
842 if (aio != NULL)
843 zio = aio;
844 else
845 vdev_queue_io_remove(vq, zio);
846
847 /*
848 * If the I/O is or was optional and therefore has no data, we need to
849 * simply discard it. We need to drop the vdev queue's lock to avoid a
850 * deadlock that we could encounter since this I/O will complete
851 * immediately.
852 */
853 if (zio->io_flags & ZIO_FLAG_NODATA) {
854 mutex_exit(&vq->vq_lock);
855 zio_vdev_io_bypass(zio);
856 zio_execute(zio);
857 mutex_enter(&vq->vq_lock);
858 goto again;
859 }
860
861 vdev_queue_pending_add(vq, zio);
862 vq->vq_last_offset = zio->io_offset;
863
864 return (zio);
865}
866
867zio_t *
868vdev_queue_io(zio_t *zio)
869{
870 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
871 zio_t *nio;
872
873 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
874 return (zio);
875
876 /*
877 * Children i/os inherent their parent's priority, which might
878 * not match the child's i/o type. Fix it up here.
879 */
880 if (zio->io_type == ZIO_TYPE_READ) {
881 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
882 zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
883 zio->io_priority != ZIO_PRIORITY_SCRUB &&
884 zio->io_priority != ZIO_PRIORITY_REMOVAL)
885 zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
886 } else if (zio->io_type == ZIO_TYPE_WRITE) {
887 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
888 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE &&
889 zio->io_priority != ZIO_PRIORITY_REMOVAL)
890 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
891 } else {
892 ASSERT(zio->io_type == ZIO_TYPE_FREE);
893 zio->io_priority = ZIO_PRIORITY_TRIM;
894 }
895
896 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
897
898 mutex_enter(&vq->vq_lock);
899 zio->io_timestamp = gethrtime();
900 vdev_queue_io_add(vq, zio);
901 nio = vdev_queue_io_to_issue(vq);
902 mutex_exit(&vq->vq_lock);
903
904 if (nio == NULL)
905 return (NULL);
906
907 if (nio->io_done == vdev_queue_agg_io_done) {
908 zio_nowait(nio);
909 return (NULL);
910 }
911
912 return (nio);
913}
914
915void
916vdev_queue_io_done(zio_t *zio)
917{
918 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
919 zio_t *nio;
920
921 mutex_enter(&vq->vq_lock);
922
923 vdev_queue_pending_remove(vq, zio);
924
925 vq->vq_io_complete_ts = gethrtime();
926
927 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
928 mutex_exit(&vq->vq_lock);
929 if (nio->io_done == vdev_queue_agg_io_done) {
930 zio_nowait(nio);
931 } else {
932 zio_vdev_io_reissue(nio);
933 zio_execute(nio);
934 }
935 mutex_enter(&vq->vq_lock);
936 }
937
938 mutex_exit(&vq->vq_lock);
939}
940
941void
942vdev_queue_change_io_priority(zio_t *zio, zio_priority_t priority)
943{
944 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
945 avl_tree_t *tree;
946
947 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
948 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
949
950 if (zio->io_type == ZIO_TYPE_READ) {
951 if (priority != ZIO_PRIORITY_SYNC_READ &&
952 priority != ZIO_PRIORITY_ASYNC_READ &&
953 priority != ZIO_PRIORITY_SCRUB)
954 priority = ZIO_PRIORITY_ASYNC_READ;
955 } else {
956 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
957 if (priority != ZIO_PRIORITY_SYNC_WRITE &&
958 priority != ZIO_PRIORITY_ASYNC_WRITE)
959 priority = ZIO_PRIORITY_ASYNC_WRITE;
960 }
961
962 mutex_enter(&vq->vq_lock);
963
964 /*
965 * If the zio is in none of the queues we can simply change
966 * the priority. If the zio is waiting to be submitted we must
967 * remove it from the queue and re-insert it with the new priority.
968 * Otherwise, the zio is currently active and we cannot change its
969 * priority.
970 */
971 tree = vdev_queue_class_tree(vq, zio->io_priority);
972 if (avl_find(tree, zio, NULL) == zio) {
973 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
974 zio->io_priority = priority;
975 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
976 } else if (avl_find(&vq->vq_active_tree, zio, NULL) != zio) {
977 zio->io_priority = priority;
978 }
979
980 mutex_exit(&vq->vq_lock);
981}
982
983/*
984 * As these three methods are only used for load calculations we're not concerned
985 * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex
986 * use here, instead we prefer to keep it lock free for performance.
987 */
988int
989vdev_queue_length(vdev_t *vd)
990{
991 return (avl_numnodes(&vd->vdev_queue.vq_active_tree));
992}
993
994uint64_t
995vdev_queue_lastoffset(vdev_t *vd)
996{
997 return (vd->vdev_queue.vq_lastoffset);
998}
999
1000void
1001vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio)
1002{
1003 vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size;
1004}
206
207#ifdef __FreeBSD__
208#ifdef _KERNEL
209SYSCTL_DECL(_vfs_zfs_vdev);
210
211static int sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS);
212SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_min_dirty_percent,
213 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
214 sysctl_zfs_async_write_active_min_dirty_percent, "I",
215 "Percentage of async write dirty data below which "
216 "async_write_min_active is used.");
217
218static int sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS);
219SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_max_dirty_percent,
220 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
221 sysctl_zfs_async_write_active_max_dirty_percent, "I",
222 "Percentage of async write dirty data above which "
223 "async_write_max_active is used.");
224
225SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RWTUN,
226 &zfs_vdev_max_active, 0,
227 "The maximum number of I/Os of all types active for each device.");
228
229#define ZFS_VDEV_QUEUE_KNOB_MIN(name) \
230SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, CTLFLAG_RWTUN,\
231 &zfs_vdev_ ## name ## _min_active, 0, \
232 "Initial number of I/O requests of type " #name \
233 " active for each device");
234
235#define ZFS_VDEV_QUEUE_KNOB_MAX(name) \
236SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, CTLFLAG_RWTUN,\
237 &zfs_vdev_ ## name ## _max_active, 0, \
238 "Maximum number of I/O requests of type " #name \
239 " active for each device");
240
241ZFS_VDEV_QUEUE_KNOB_MIN(sync_read);
242ZFS_VDEV_QUEUE_KNOB_MAX(sync_read);
243ZFS_VDEV_QUEUE_KNOB_MIN(sync_write);
244ZFS_VDEV_QUEUE_KNOB_MAX(sync_write);
245ZFS_VDEV_QUEUE_KNOB_MIN(async_read);
246ZFS_VDEV_QUEUE_KNOB_MAX(async_read);
247ZFS_VDEV_QUEUE_KNOB_MIN(async_write);
248ZFS_VDEV_QUEUE_KNOB_MAX(async_write);
249ZFS_VDEV_QUEUE_KNOB_MIN(scrub);
250ZFS_VDEV_QUEUE_KNOB_MAX(scrub);
251ZFS_VDEV_QUEUE_KNOB_MIN(trim);
252ZFS_VDEV_QUEUE_KNOB_MAX(trim);
253
254#undef ZFS_VDEV_QUEUE_KNOB
255
256SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RWTUN,
257 &zfs_vdev_aggregation_limit, 0,
258 "I/O requests are aggregated up to this size");
259SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RWTUN,
260 &zfs_vdev_read_gap_limit, 0,
261 "Acceptable gap between two reads being aggregated");
262SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RWTUN,
263 &zfs_vdev_write_gap_limit, 0,
264 "Acceptable gap between two writes being aggregated");
265SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, queue_depth_pct, CTLFLAG_RWTUN,
266 &zfs_vdev_queue_depth_pct, 0,
267 "Queue depth percentage for each top-level");
268
269static int
270sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS)
271{
272 int val, err;
273
274 val = zfs_vdev_async_write_active_min_dirty_percent;
275 err = sysctl_handle_int(oidp, &val, 0, req);
276 if (err != 0 || req->newptr == NULL)
277 return (err);
278
279 if (val < 0 || val > 100 ||
280 val >= zfs_vdev_async_write_active_max_dirty_percent)
281 return (EINVAL);
282
283 zfs_vdev_async_write_active_min_dirty_percent = val;
284
285 return (0);
286}
287
288static int
289sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS)
290{
291 int val, err;
292
293 val = zfs_vdev_async_write_active_max_dirty_percent;
294 err = sysctl_handle_int(oidp, &val, 0, req);
295 if (err != 0 || req->newptr == NULL)
296 return (err);
297
298 if (val < 0 || val > 100 ||
299 val <= zfs_vdev_async_write_active_min_dirty_percent)
300 return (EINVAL);
301
302 zfs_vdev_async_write_active_max_dirty_percent = val;
303
304 return (0);
305}
306#endif
307#endif
308
309int
310vdev_queue_offset_compare(const void *x1, const void *x2)
311{
312 const zio_t *z1 = x1;
313 const zio_t *z2 = x2;
314
315 if (z1->io_offset < z2->io_offset)
316 return (-1);
317 if (z1->io_offset > z2->io_offset)
318 return (1);
319
320 if (z1 < z2)
321 return (-1);
322 if (z1 > z2)
323 return (1);
324
325 return (0);
326}
327
328static inline avl_tree_t *
329vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p)
330{
331 return (&vq->vq_class[p].vqc_queued_tree);
332}
333
334static inline avl_tree_t *
335vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t)
336{
337 if (t == ZIO_TYPE_READ)
338 return (&vq->vq_read_offset_tree);
339 else if (t == ZIO_TYPE_WRITE)
340 return (&vq->vq_write_offset_tree);
341 else
342 return (NULL);
343}
344
345int
346vdev_queue_timestamp_compare(const void *x1, const void *x2)
347{
348 const zio_t *z1 = x1;
349 const zio_t *z2 = x2;
350
351 if (z1->io_timestamp < z2->io_timestamp)
352 return (-1);
353 if (z1->io_timestamp > z2->io_timestamp)
354 return (1);
355
356 if (z1->io_offset < z2->io_offset)
357 return (-1);
358 if (z1->io_offset > z2->io_offset)
359 return (1);
360
361 if (z1 < z2)
362 return (-1);
363 if (z1 > z2)
364 return (1);
365
366 return (0);
367}
368
369void
370vdev_queue_init(vdev_t *vd)
371{
372 vdev_queue_t *vq = &vd->vdev_queue;
373
374 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
375 vq->vq_vdev = vd;
376
377 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
378 sizeof (zio_t), offsetof(struct zio, io_queue_node));
379 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ),
380 vdev_queue_offset_compare, sizeof (zio_t),
381 offsetof(struct zio, io_offset_node));
382 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE),
383 vdev_queue_offset_compare, sizeof (zio_t),
384 offsetof(struct zio, io_offset_node));
385
386 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
387 int (*compfn) (const void *, const void *);
388
389 /*
390 * The synchronous i/o queues are dispatched in FIFO rather
391 * than LBA order. This provides more consistent latency for
392 * these i/os.
393 */
394 if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE)
395 compfn = vdev_queue_timestamp_compare;
396 else
397 compfn = vdev_queue_offset_compare;
398
399 avl_create(vdev_queue_class_tree(vq, p), compfn,
400 sizeof (zio_t), offsetof(struct zio, io_queue_node));
401 }
402
403 vq->vq_lastoffset = 0;
404}
405
406void
407vdev_queue_fini(vdev_t *vd)
408{
409 vdev_queue_t *vq = &vd->vdev_queue;
410
411 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
412 avl_destroy(vdev_queue_class_tree(vq, p));
413 avl_destroy(&vq->vq_active_tree);
414 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
415 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE));
416
417 mutex_destroy(&vq->vq_lock);
418}
419
420static void
421vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
422{
423 spa_t *spa = zio->io_spa;
424 avl_tree_t *qtt;
425
426 ASSERT(MUTEX_HELD(&vq->vq_lock));
427 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
428 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
429 qtt = vdev_queue_type_tree(vq, zio->io_type);
430 if (qtt)
431 avl_add(qtt, zio);
432
433#ifdef illumos
434 mutex_enter(&spa->spa_iokstat_lock);
435 spa->spa_queue_stats[zio->io_priority].spa_queued++;
436 if (spa->spa_iokstat != NULL)
437 kstat_waitq_enter(spa->spa_iokstat->ks_data);
438 mutex_exit(&spa->spa_iokstat_lock);
439#endif
440}
441
442static void
443vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
444{
445 spa_t *spa = zio->io_spa;
446 avl_tree_t *qtt;
447
448 ASSERT(MUTEX_HELD(&vq->vq_lock));
449 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
450 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
451 qtt = vdev_queue_type_tree(vq, zio->io_type);
452 if (qtt)
453 avl_remove(qtt, zio);
454
455#ifdef illumos
456 mutex_enter(&spa->spa_iokstat_lock);
457 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
458 spa->spa_queue_stats[zio->io_priority].spa_queued--;
459 if (spa->spa_iokstat != NULL)
460 kstat_waitq_exit(spa->spa_iokstat->ks_data);
461 mutex_exit(&spa->spa_iokstat_lock);
462#endif
463}
464
465static void
466vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
467{
468 spa_t *spa = zio->io_spa;
469 ASSERT(MUTEX_HELD(&vq->vq_lock));
470 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
471 vq->vq_class[zio->io_priority].vqc_active++;
472 avl_add(&vq->vq_active_tree, zio);
473
474#ifdef illumos
475 mutex_enter(&spa->spa_iokstat_lock);
476 spa->spa_queue_stats[zio->io_priority].spa_active++;
477 if (spa->spa_iokstat != NULL)
478 kstat_runq_enter(spa->spa_iokstat->ks_data);
479 mutex_exit(&spa->spa_iokstat_lock);
480#endif
481}
482
483static void
484vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
485{
486 spa_t *spa = zio->io_spa;
487 ASSERT(MUTEX_HELD(&vq->vq_lock));
488 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
489 vq->vq_class[zio->io_priority].vqc_active--;
490 avl_remove(&vq->vq_active_tree, zio);
491
492#ifdef illumos
493 mutex_enter(&spa->spa_iokstat_lock);
494 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
495 spa->spa_queue_stats[zio->io_priority].spa_active--;
496 if (spa->spa_iokstat != NULL) {
497 kstat_io_t *ksio = spa->spa_iokstat->ks_data;
498
499 kstat_runq_exit(spa->spa_iokstat->ks_data);
500 if (zio->io_type == ZIO_TYPE_READ) {
501 ksio->reads++;
502 ksio->nread += zio->io_size;
503 } else if (zio->io_type == ZIO_TYPE_WRITE) {
504 ksio->writes++;
505 ksio->nwritten += zio->io_size;
506 }
507 }
508 mutex_exit(&spa->spa_iokstat_lock);
509#endif
510}
511
512static void
513vdev_queue_agg_io_done(zio_t *aio)
514{
515 if (aio->io_type == ZIO_TYPE_READ) {
516 zio_t *pio;
517 zio_link_t *zl = NULL;
518 while ((pio = zio_walk_parents(aio, &zl)) != NULL) {
519 abd_copy_off(pio->io_abd, aio->io_abd,
520 0, pio->io_offset - aio->io_offset, pio->io_size);
521 }
522 }
523
524 abd_free(aio->io_abd);
525}
526
527static int
528vdev_queue_class_min_active(zio_priority_t p)
529{
530 switch (p) {
531 case ZIO_PRIORITY_SYNC_READ:
532 return (zfs_vdev_sync_read_min_active);
533 case ZIO_PRIORITY_SYNC_WRITE:
534 return (zfs_vdev_sync_write_min_active);
535 case ZIO_PRIORITY_ASYNC_READ:
536 return (zfs_vdev_async_read_min_active);
537 case ZIO_PRIORITY_ASYNC_WRITE:
538 return (zfs_vdev_async_write_min_active);
539 case ZIO_PRIORITY_SCRUB:
540 return (zfs_vdev_scrub_min_active);
541 case ZIO_PRIORITY_TRIM:
542 return (zfs_vdev_trim_min_active);
543 case ZIO_PRIORITY_REMOVAL:
544 return (zfs_vdev_removal_min_active);
545 default:
546 panic("invalid priority %u", p);
547 return (0);
548 }
549}
550
551static __noinline int
552vdev_queue_max_async_writes(spa_t *spa)
553{
554 int writes;
555 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total;
556 uint64_t min_bytes = zfs_dirty_data_max *
557 zfs_vdev_async_write_active_min_dirty_percent / 100;
558 uint64_t max_bytes = zfs_dirty_data_max *
559 zfs_vdev_async_write_active_max_dirty_percent / 100;
560
561 /*
562 * Sync tasks correspond to interactive user actions. To reduce the
563 * execution time of those actions we push data out as fast as possible.
564 */
565 if (spa_has_pending_synctask(spa)) {
566 return (zfs_vdev_async_write_max_active);
567 }
568
569 if (dirty < min_bytes)
570 return (zfs_vdev_async_write_min_active);
571 if (dirty > max_bytes)
572 return (zfs_vdev_async_write_max_active);
573
574 /*
575 * linear interpolation:
576 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
577 * move right by min_bytes
578 * move up by min_writes
579 */
580 writes = (dirty - min_bytes) *
581 (zfs_vdev_async_write_max_active -
582 zfs_vdev_async_write_min_active) /
583 (max_bytes - min_bytes) +
584 zfs_vdev_async_write_min_active;
585 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
586 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
587 return (writes);
588}
589
590static int
591vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
592{
593 switch (p) {
594 case ZIO_PRIORITY_SYNC_READ:
595 return (zfs_vdev_sync_read_max_active);
596 case ZIO_PRIORITY_SYNC_WRITE:
597 return (zfs_vdev_sync_write_max_active);
598 case ZIO_PRIORITY_ASYNC_READ:
599 return (zfs_vdev_async_read_max_active);
600 case ZIO_PRIORITY_ASYNC_WRITE:
601 return (vdev_queue_max_async_writes(spa));
602 case ZIO_PRIORITY_SCRUB:
603 return (zfs_vdev_scrub_max_active);
604 case ZIO_PRIORITY_TRIM:
605 return (zfs_vdev_trim_max_active);
606 case ZIO_PRIORITY_REMOVAL:
607 return (zfs_vdev_removal_max_active);
608 default:
609 panic("invalid priority %u", p);
610 return (0);
611 }
612}
613
614/*
615 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
616 * there is no eligible class.
617 */
618static zio_priority_t
619vdev_queue_class_to_issue(vdev_queue_t *vq)
620{
621 spa_t *spa = vq->vq_vdev->vdev_spa;
622 zio_priority_t p;
623
624 ASSERT(MUTEX_HELD(&vq->vq_lock));
625
626 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
627 return (ZIO_PRIORITY_NUM_QUEUEABLE);
628
629 /* find a queue that has not reached its minimum # outstanding i/os */
630 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
631 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
632 vq->vq_class[p].vqc_active <
633 vdev_queue_class_min_active(p))
634 return (p);
635 }
636
637 /*
638 * If we haven't found a queue, look for one that hasn't reached its
639 * maximum # outstanding i/os.
640 */
641 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
642 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
643 vq->vq_class[p].vqc_active <
644 vdev_queue_class_max_active(spa, p))
645 return (p);
646 }
647
648 /* No eligible queued i/os */
649 return (ZIO_PRIORITY_NUM_QUEUEABLE);
650}
651
652/*
653 * Compute the range spanned by two i/os, which is the endpoint of the last
654 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
655 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
656 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
657 */
658#define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
659#define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
660
661static zio_t *
662vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
663{
664 zio_t *first, *last, *aio, *dio, *mandatory, *nio;
665 uint64_t maxgap = 0;
666 uint64_t size;
667 boolean_t stretch;
668 avl_tree_t *t;
669 enum zio_flag flags;
670
671 ASSERT(MUTEX_HELD(&vq->vq_lock));
672
673 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
674 return (NULL);
675
676 first = last = zio;
677
678 if (zio->io_type == ZIO_TYPE_READ)
679 maxgap = zfs_vdev_read_gap_limit;
680
681 /*
682 * We can aggregate I/Os that are sufficiently adjacent and of
683 * the same flavor, as expressed by the AGG_INHERIT flags.
684 * The latter requirement is necessary so that certain
685 * attributes of the I/O, such as whether it's a normal I/O
686 * or a scrub/resilver, can be preserved in the aggregate.
687 * We can include optional I/Os, but don't allow them
688 * to begin a range as they add no benefit in that situation.
689 */
690
691 /*
692 * We keep track of the last non-optional I/O.
693 */
694 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
695
696 /*
697 * Walk backwards through sufficiently contiguous I/Os
698 * recording the last non-optional I/O.
699 */
700 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
701 t = vdev_queue_type_tree(vq, zio->io_type);
702 while (t != NULL && (dio = AVL_PREV(t, first)) != NULL &&
703 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
704 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
705 IO_GAP(dio, first) <= maxgap &&
706 dio->io_type == zio->io_type) {
707 first = dio;
708 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
709 mandatory = first;
710 }
711
712 /*
713 * Skip any initial optional I/Os.
714 */
715 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
716 first = AVL_NEXT(t, first);
717 ASSERT(first != NULL);
718 }
719
720 /*
721 * Walk forward through sufficiently contiguous I/Os.
722 * The aggregation limit does not apply to optional i/os, so that
723 * we can issue contiguous writes even if they are larger than the
724 * aggregation limit.
725 */
726 while ((dio = AVL_NEXT(t, last)) != NULL &&
727 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
728 (IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit ||
729 (dio->io_flags & ZIO_FLAG_OPTIONAL)) &&
730 IO_GAP(last, dio) <= maxgap &&
731 dio->io_type == zio->io_type) {
732 last = dio;
733 if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
734 mandatory = last;
735 }
736
737 /*
738 * Now that we've established the range of the I/O aggregation
739 * we must decide what to do with trailing optional I/Os.
740 * For reads, there's nothing to do. While we are unable to
741 * aggregate further, it's possible that a trailing optional
742 * I/O would allow the underlying device to aggregate with
743 * subsequent I/Os. We must therefore determine if the next
744 * non-optional I/O is close enough to make aggregation
745 * worthwhile.
746 */
747 stretch = B_FALSE;
748 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
749 zio_t *nio = last;
750 while ((dio = AVL_NEXT(t, nio)) != NULL &&
751 IO_GAP(nio, dio) == 0 &&
752 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
753 nio = dio;
754 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
755 stretch = B_TRUE;
756 break;
757 }
758 }
759 }
760
761 if (stretch) {
762 /*
763 * We are going to include an optional io in our aggregated
764 * span, thus closing the write gap. Only mandatory i/os can
765 * start aggregated spans, so make sure that the next i/o
766 * after our span is mandatory.
767 */
768 dio = AVL_NEXT(t, last);
769 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
770 } else {
771 /* do not include the optional i/o */
772 while (last != mandatory && last != first) {
773 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
774 last = AVL_PREV(t, last);
775 ASSERT(last != NULL);
776 }
777 }
778
779 if (first == last)
780 return (NULL);
781
782 size = IO_SPAN(first, last);
783 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
784
785 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
786 abd_alloc_for_io(size, B_TRUE), size, first->io_type,
787 zio->io_priority, flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
788 vdev_queue_agg_io_done, NULL);
789 aio->io_timestamp = first->io_timestamp;
790
791 nio = first;
792 do {
793 dio = nio;
794 nio = AVL_NEXT(t, dio);
795 ASSERT3U(dio->io_type, ==, aio->io_type);
796
797 if (dio->io_flags & ZIO_FLAG_NODATA) {
798 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
799 abd_zero_off(aio->io_abd,
800 dio->io_offset - aio->io_offset, dio->io_size);
801 } else if (dio->io_type == ZIO_TYPE_WRITE) {
802 abd_copy_off(aio->io_abd, dio->io_abd,
803 dio->io_offset - aio->io_offset, 0, dio->io_size);
804 }
805
806 zio_add_child(dio, aio);
807 vdev_queue_io_remove(vq, dio);
808 zio_vdev_io_bypass(dio);
809 zio_execute(dio);
810 } while (dio != last);
811
812 return (aio);
813}
814
815static zio_t *
816vdev_queue_io_to_issue(vdev_queue_t *vq)
817{
818 zio_t *zio, *aio;
819 zio_priority_t p;
820 avl_index_t idx;
821 avl_tree_t *tree;
822 zio_t search;
823
824again:
825 ASSERT(MUTEX_HELD(&vq->vq_lock));
826
827 p = vdev_queue_class_to_issue(vq);
828
829 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
830 /* No eligible queued i/os */
831 return (NULL);
832 }
833
834 /*
835 * For LBA-ordered queues (async / scrub), issue the i/o which follows
836 * the most recently issued i/o in LBA (offset) order.
837 *
838 * For FIFO queues (sync), issue the i/o with the lowest timestamp.
839 */
840 tree = vdev_queue_class_tree(vq, p);
841 search.io_timestamp = 0;
842 search.io_offset = vq->vq_last_offset + 1;
843 VERIFY3P(avl_find(tree, &search, &idx), ==, NULL);
844 zio = avl_nearest(tree, idx, AVL_AFTER);
845 if (zio == NULL)
846 zio = avl_first(tree);
847 ASSERT3U(zio->io_priority, ==, p);
848
849 aio = vdev_queue_aggregate(vq, zio);
850 if (aio != NULL)
851 zio = aio;
852 else
853 vdev_queue_io_remove(vq, zio);
854
855 /*
856 * If the I/O is or was optional and therefore has no data, we need to
857 * simply discard it. We need to drop the vdev queue's lock to avoid a
858 * deadlock that we could encounter since this I/O will complete
859 * immediately.
860 */
861 if (zio->io_flags & ZIO_FLAG_NODATA) {
862 mutex_exit(&vq->vq_lock);
863 zio_vdev_io_bypass(zio);
864 zio_execute(zio);
865 mutex_enter(&vq->vq_lock);
866 goto again;
867 }
868
869 vdev_queue_pending_add(vq, zio);
870 vq->vq_last_offset = zio->io_offset;
871
872 return (zio);
873}
874
875zio_t *
876vdev_queue_io(zio_t *zio)
877{
878 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
879 zio_t *nio;
880
881 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
882 return (zio);
883
884 /*
885 * Children i/os inherent their parent's priority, which might
886 * not match the child's i/o type. Fix it up here.
887 */
888 if (zio->io_type == ZIO_TYPE_READ) {
889 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
890 zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
891 zio->io_priority != ZIO_PRIORITY_SCRUB &&
892 zio->io_priority != ZIO_PRIORITY_REMOVAL)
893 zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
894 } else if (zio->io_type == ZIO_TYPE_WRITE) {
895 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
896 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE &&
897 zio->io_priority != ZIO_PRIORITY_REMOVAL)
898 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
899 } else {
900 ASSERT(zio->io_type == ZIO_TYPE_FREE);
901 zio->io_priority = ZIO_PRIORITY_TRIM;
902 }
903
904 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
905
906 mutex_enter(&vq->vq_lock);
907 zio->io_timestamp = gethrtime();
908 vdev_queue_io_add(vq, zio);
909 nio = vdev_queue_io_to_issue(vq);
910 mutex_exit(&vq->vq_lock);
911
912 if (nio == NULL)
913 return (NULL);
914
915 if (nio->io_done == vdev_queue_agg_io_done) {
916 zio_nowait(nio);
917 return (NULL);
918 }
919
920 return (nio);
921}
922
923void
924vdev_queue_io_done(zio_t *zio)
925{
926 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
927 zio_t *nio;
928
929 mutex_enter(&vq->vq_lock);
930
931 vdev_queue_pending_remove(vq, zio);
932
933 vq->vq_io_complete_ts = gethrtime();
934
935 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
936 mutex_exit(&vq->vq_lock);
937 if (nio->io_done == vdev_queue_agg_io_done) {
938 zio_nowait(nio);
939 } else {
940 zio_vdev_io_reissue(nio);
941 zio_execute(nio);
942 }
943 mutex_enter(&vq->vq_lock);
944 }
945
946 mutex_exit(&vq->vq_lock);
947}
948
949void
950vdev_queue_change_io_priority(zio_t *zio, zio_priority_t priority)
951{
952 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
953 avl_tree_t *tree;
954
955 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
956 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
957
958 if (zio->io_type == ZIO_TYPE_READ) {
959 if (priority != ZIO_PRIORITY_SYNC_READ &&
960 priority != ZIO_PRIORITY_ASYNC_READ &&
961 priority != ZIO_PRIORITY_SCRUB)
962 priority = ZIO_PRIORITY_ASYNC_READ;
963 } else {
964 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
965 if (priority != ZIO_PRIORITY_SYNC_WRITE &&
966 priority != ZIO_PRIORITY_ASYNC_WRITE)
967 priority = ZIO_PRIORITY_ASYNC_WRITE;
968 }
969
970 mutex_enter(&vq->vq_lock);
971
972 /*
973 * If the zio is in none of the queues we can simply change
974 * the priority. If the zio is waiting to be submitted we must
975 * remove it from the queue and re-insert it with the new priority.
976 * Otherwise, the zio is currently active and we cannot change its
977 * priority.
978 */
979 tree = vdev_queue_class_tree(vq, zio->io_priority);
980 if (avl_find(tree, zio, NULL) == zio) {
981 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
982 zio->io_priority = priority;
983 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
984 } else if (avl_find(&vq->vq_active_tree, zio, NULL) != zio) {
985 zio->io_priority = priority;
986 }
987
988 mutex_exit(&vq->vq_lock);
989}
990
991/*
992 * As these three methods are only used for load calculations we're not concerned
993 * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex
994 * use here, instead we prefer to keep it lock free for performance.
995 */
996int
997vdev_queue_length(vdev_t *vd)
998{
999 return (avl_numnodes(&vd->vdev_queue.vq_active_tree));
1000}
1001
1002uint64_t
1003vdev_queue_lastoffset(vdev_t *vd)
1004{
1005 return (vd->vdev_queue.vq_lastoffset);
1006}
1007
1008void
1009vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio)
1010{
1011 vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size;
1012}