]>
Commit | Line | Data |
---|---|---|
34dc7c2f BB |
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 | |
1d3ba0bf | 9 | * or https://opensource.org/licenses/CDDL-1.0. |
34dc7c2f BB |
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 | /* | |
d164b209 | 22 | * Copyright 2009 Sun Microsystems, Inc. All rights reserved. |
34dc7c2f BB |
23 | * Use is subject to license terms. |
24 | */ | |
25 | ||
cc92e9d0 | 26 | /* |
492f64e9 | 27 | * Copyright (c) 2012, 2018 by Delphix. All rights reserved. |
cc92e9d0 GW |
28 | */ |
29 | ||
34dc7c2f | 30 | #include <sys/zfs_context.h> |
34dc7c2f | 31 | #include <sys/vdev_impl.h> |
330847ff | 32 | #include <sys/spa_impl.h> |
34dc7c2f BB |
33 | #include <sys/zio.h> |
34 | #include <sys/avl.h> | |
e8b96c60 | 35 | #include <sys/dsl_pool.h> |
3dfb57a3 | 36 | #include <sys/metaslab_impl.h> |
e8b96c60 | 37 | #include <sys/spa.h> |
a6255b7f | 38 | #include <sys/abd.h> |
34dc7c2f BB |
39 | |
40 | /* | |
e8b96c60 MA |
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 five I/O classes | |
47 | * prioritized in the following order: sync read, sync write, async read, | |
48 | * async write, and scrub/resilver. 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. If the | |
52 | * sum of the per-queue maximums exceeds the aggregate maximum, then the | |
53 | * number of active i/os may reach zfs_vdev_max_active, in which case no | |
54 | * further i/os will be issued regardless of whether all per-queue | |
55 | * minimums have been met. | |
56 | * | |
57 | * For many physical devices, throughput increases with the number of | |
58 | * concurrent operations, but latency typically suffers. Further, physical | |
59 | * devices typically have a limit at which more concurrent operations have no | |
60 | * effect on throughput or can actually cause it to decrease. | |
61 | * | |
62 | * The scheduler selects the next operation to issue by first looking for an | |
63 | * I/O class whose minimum has not been satisfied. Once all are satisfied and | |
64 | * the aggregate maximum has not been hit, the scheduler looks for classes | |
65 | * whose maximum has not been satisfied. Iteration through the I/O classes is | |
66 | * done in the order specified above. No further operations are issued if the | |
67 | * aggregate maximum number of concurrent operations has been hit or if there | |
68 | * are no operations queued for an I/O class that has not hit its maximum. | |
69 | * Every time an i/o is queued or an operation completes, the I/O scheduler | |
70 | * looks for new operations to issue. | |
71 | * | |
72 | * All I/O classes have a fixed maximum number of outstanding operations | |
73 | * except for the async write class. Asynchronous writes represent the data | |
74 | * that is committed to stable storage during the syncing stage for | |
75 | * transaction groups (see txg.c). Transaction groups enter the syncing state | |
76 | * periodically so the number of queued async writes will quickly burst up and | |
77 | * then bleed down to zero. Rather than servicing them as quickly as possible, | |
78 | * the I/O scheduler changes the maximum number of active async write i/os | |
79 | * according to the amount of dirty data in the pool (see dsl_pool.c). Since | |
80 | * both throughput and latency typically increase with the number of | |
81 | * concurrent operations issued to physical devices, reducing the burstiness | |
82 | * in the number of concurrent operations also stabilizes the response time of | |
83 | * operations from other -- and in particular synchronous -- queues. In broad | |
84 | * strokes, the I/O scheduler will issue more concurrent operations from the | |
85 | * async write queue as there's more dirty data in the pool. | |
86 | * | |
87 | * Async Writes | |
88 | * | |
89 | * The number of concurrent operations issued for the async write I/O class | |
90 | * follows a piece-wise linear function defined by a few adjustable points. | |
91 | * | |
92 | * | o---------| <-- zfs_vdev_async_write_max_active | |
93 | * ^ | /^ | | |
94 | * | | / | | | |
95 | * active | / | | | |
96 | * I/O | / | | | |
97 | * count | / | | | |
98 | * | / | | | |
99 | * |------------o | | <-- zfs_vdev_async_write_min_active | |
100 | * 0|____________^______|_________| | |
101 | * 0% | | 100% of zfs_dirty_data_max | |
102 | * | | | |
103 | * | `-- zfs_vdev_async_write_active_max_dirty_percent | |
104 | * `--------- zfs_vdev_async_write_active_min_dirty_percent | |
105 | * | |
106 | * Until the amount of dirty data exceeds a minimum percentage of the dirty | |
107 | * data allowed in the pool, the I/O scheduler will limit the number of | |
108 | * concurrent operations to the minimum. As that threshold is crossed, the | |
109 | * number of concurrent operations issued increases linearly to the maximum at | |
110 | * the specified maximum percentage of the dirty data allowed in the pool. | |
111 | * | |
112 | * Ideally, the amount of dirty data on a busy pool will stay in the sloped | |
113 | * part of the function between zfs_vdev_async_write_active_min_dirty_percent | |
114 | * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the | |
115 | * maximum percentage, this indicates that the rate of incoming data is | |
116 | * greater than the rate that the backend storage can handle. In this case, we | |
117 | * must further throttle incoming writes (see dmu_tx_delay() for details). | |
34dc7c2f | 118 | */ |
d3cc8b15 | 119 | |
34dc7c2f | 120 | /* |
e8b96c60 | 121 | * The maximum number of i/os active to each device. Ideally, this will be >= |
6f5aac3c | 122 | * the sum of each queue's max_active. |
34dc7c2f | 123 | */ |
fdc2d303 | 124 | uint_t zfs_vdev_max_active = 1000; |
34dc7c2f | 125 | |
cb682a17 | 126 | /* |
e8b96c60 MA |
127 | * Per-queue limits on the number of i/os active to each device. If the |
128 | * number of active i/os is < zfs_vdev_max_active, then the min_active comes | |
6f5aac3c AM |
129 | * into play. We will send min_active from each queue round-robin, and then |
130 | * send from queues in the order defined by zio_priority_t up to max_active. | |
131 | * Some queues have additional mechanisms to limit number of active I/Os in | |
132 | * addition to min_active and max_active, see below. | |
e8b96c60 MA |
133 | * |
134 | * In general, smaller max_active's will lead to lower latency of synchronous | |
135 | * operations. Larger max_active's may lead to higher overall throughput, | |
136 | * depending on underlying storage. | |
137 | * | |
138 | * The ratio of the queues' max_actives determines the balance of performance | |
139 | * between reads, writes, and scrubs. E.g., increasing | |
140 | * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete | |
141 | * more quickly, but reads and writes to have higher latency and lower | |
142 | * throughput. | |
cb682a17 | 143 | */ |
fdc2d303 RY |
144 | static uint_t zfs_vdev_sync_read_min_active = 10; |
145 | static uint_t zfs_vdev_sync_read_max_active = 10; | |
146 | static uint_t zfs_vdev_sync_write_min_active = 10; | |
147 | static uint_t zfs_vdev_sync_write_max_active = 10; | |
148 | static uint_t zfs_vdev_async_read_min_active = 1; | |
149 | /* */ uint_t zfs_vdev_async_read_max_active = 3; | |
150 | static uint_t zfs_vdev_async_write_min_active = 2; | |
151 | /* */ uint_t zfs_vdev_async_write_max_active = 10; | |
152 | static uint_t zfs_vdev_scrub_min_active = 1; | |
153 | static uint_t zfs_vdev_scrub_max_active = 3; | |
154 | static uint_t zfs_vdev_removal_min_active = 1; | |
155 | static uint_t zfs_vdev_removal_max_active = 2; | |
156 | static uint_t zfs_vdev_initializing_min_active = 1; | |
157 | static uint_t zfs_vdev_initializing_max_active = 1; | |
158 | static uint_t zfs_vdev_trim_min_active = 1; | |
159 | static uint_t zfs_vdev_trim_max_active = 2; | |
160 | static uint_t zfs_vdev_rebuild_min_active = 1; | |
161 | static uint_t zfs_vdev_rebuild_max_active = 3; | |
34dc7c2f | 162 | |
e8b96c60 MA |
163 | /* |
164 | * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent | |
165 | * dirty data, use zfs_vdev_async_write_min_active. When it has more than | |
166 | * zfs_vdev_async_write_active_max_dirty_percent, use | |
167 | * zfs_vdev_async_write_max_active. The value is linearly interpolated | |
168 | * between min and max. | |
169 | */ | |
fdc2d303 RY |
170 | uint_t zfs_vdev_async_write_active_min_dirty_percent = 30; |
171 | uint_t zfs_vdev_async_write_active_max_dirty_percent = 60; | |
34dc7c2f | 172 | |
6f5aac3c AM |
173 | /* |
174 | * For non-interactive I/O (scrub, resilver, removal, initialize and rebuild), | |
175 | * the number of concurrently-active I/O's is limited to *_min_active, unless | |
176 | * the vdev is "idle". When there are no interactive I/Os active (sync or | |
177 | * async), and zfs_vdev_nia_delay I/Os have completed since the last | |
178 | * interactive I/O, then the vdev is considered to be "idle", and the number | |
179 | * of concurrently-active non-interactive I/O's is increased to *_max_active. | |
180 | */ | |
18168da7 | 181 | static uint_t zfs_vdev_nia_delay = 5; |
6f5aac3c AM |
182 | |
183 | /* | |
184 | * Some HDDs tend to prioritize sequential I/O so high that concurrent | |
185 | * random I/O latency reaches several seconds. On some HDDs it happens | |
186 | * even if sequential I/Os are submitted one at a time, and so setting | |
187 | * *_max_active to 1 does not help. To prevent non-interactive I/Os, like | |
188 | * scrub, from monopolizing the device no more than zfs_vdev_nia_credit | |
189 | * I/Os can be sent while there are outstanding incomplete interactive | |
190 | * I/Os. This enforced wait ensures the HDD services the interactive I/O | |
191 | * within a reasonable amount of time. | |
192 | */ | |
18168da7 | 193 | static uint_t zfs_vdev_nia_credit = 5; |
6f5aac3c | 194 | |
34dc7c2f | 195 | /* |
45d1cae3 BB |
196 | * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O. |
197 | * For read I/Os, we also aggregate across small adjacency gaps; for writes | |
198 | * we include spans of optional I/Os to aid aggregation at the disk even when | |
199 | * they aren't able to help us aggregate at this level. | |
34dc7c2f | 200 | */ |
fdc2d303 RY |
201 | static uint_t zfs_vdev_aggregation_limit = 1 << 20; |
202 | static uint_t zfs_vdev_aggregation_limit_non_rotating = SPA_OLD_MAXBLOCKSIZE; | |
203 | static uint_t zfs_vdev_read_gap_limit = 32 << 10; | |
204 | static uint_t zfs_vdev_write_gap_limit = 4 << 10; | |
34dc7c2f | 205 | |
3dfb57a3 DB |
206 | /* |
207 | * Define the queue depth percentage for each top-level. This percentage is | |
208 | * used in conjunction with zfs_vdev_async_max_active to determine how many | |
209 | * allocations a specific top-level vdev should handle. Once the queue depth | |
210 | * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100 | |
211 | * then allocator will stop allocating blocks on that top-level device. | |
212 | * The default kernel setting is 1000% which will yield 100 allocations per | |
213 | * device. For userland testing, the default setting is 300% which equates | |
214 | * to 30 allocations per device. | |
215 | */ | |
216 | #ifdef _KERNEL | |
fdc2d303 | 217 | uint_t zfs_vdev_queue_depth_pct = 1000; |
3dfb57a3 | 218 | #else |
fdc2d303 | 219 | uint_t zfs_vdev_queue_depth_pct = 300; |
3dfb57a3 DB |
220 | #endif |
221 | ||
492f64e9 PD |
222 | /* |
223 | * When performing allocations for a given metaslab, we want to make sure that | |
224 | * there are enough IOs to aggregate together to improve throughput. We want to | |
225 | * ensure that there are at least 128k worth of IOs that can be aggregated, and | |
226 | * we assume that the average allocation size is 4k, so we need the queue depth | |
227 | * to be 32 per allocator to get good aggregation of sequential writes. | |
228 | */ | |
fdc2d303 | 229 | uint_t zfs_vdev_def_queue_depth = 32; |
492f64e9 | 230 | |
65c7cc49 | 231 | static int |
e8b96c60 | 232 | vdev_queue_offset_compare(const void *x1, const void *x2) |
34dc7c2f | 233 | { |
ee36c709 GN |
234 | const zio_t *z1 = (const zio_t *)x1; |
235 | const zio_t *z2 = (const zio_t *)x2; | |
34dc7c2f | 236 | |
ca577779 | 237 | int cmp = TREE_CMP(z1->io_offset, z2->io_offset); |
34dc7c2f | 238 | |
ee36c709 GN |
239 | if (likely(cmp)) |
240 | return (cmp); | |
34dc7c2f | 241 | |
ca577779 | 242 | return (TREE_PCMP(z1, z2)); |
34dc7c2f BB |
243 | } |
244 | ||
8469b5aa | 245 | #define VDQ_T_SHIFT 29 |
ec8501ee | 246 | |
65c7cc49 | 247 | static int |
8469b5aa | 248 | vdev_queue_to_compare(const void *x1, const void *x2) |
34dc7c2f | 249 | { |
ee36c709 GN |
250 | const zio_t *z1 = (const zio_t *)x1; |
251 | const zio_t *z2 = (const zio_t *)x2; | |
34dc7c2f | 252 | |
8469b5aa AM |
253 | int tcmp = TREE_CMP(z1->io_timestamp >> VDQ_T_SHIFT, |
254 | z2->io_timestamp >> VDQ_T_SHIFT); | |
255 | int ocmp = TREE_CMP(z1->io_offset, z2->io_offset); | |
256 | int cmp = tcmp ? tcmp : ocmp; | |
34dc7c2f | 257 | |
8469b5aa | 258 | if (likely(cmp | (z1->io_queue_state == ZIO_QS_NONE))) |
ee36c709 | 259 | return (cmp); |
34dc7c2f | 260 | |
ca577779 | 261 | return (TREE_PCMP(z1, z2)); |
34dc7c2f BB |
262 | } |
263 | ||
8469b5aa AM |
264 | static inline boolean_t |
265 | vdev_queue_class_fifo(zio_priority_t p) | |
266 | { | |
267 | return (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE || | |
268 | p == ZIO_PRIORITY_TRIM); | |
269 | } | |
270 | ||
271 | static void | |
272 | vdev_queue_class_add(vdev_queue_t *vq, zio_t *zio) | |
273 | { | |
274 | zio_priority_t p = zio->io_priority; | |
275 | vq->vq_cqueued |= 1U << p; | |
2a154b84 | 276 | if (vdev_queue_class_fifo(p)) { |
8469b5aa | 277 | list_insert_tail(&vq->vq_class[p].vqc_list, zio); |
2a154b84 M |
278 | vq->vq_class[p].vqc_list_numnodes++; |
279 | } | |
8469b5aa AM |
280 | else |
281 | avl_add(&vq->vq_class[p].vqc_tree, zio); | |
282 | } | |
283 | ||
284 | static void | |
285 | vdev_queue_class_remove(vdev_queue_t *vq, zio_t *zio) | |
286 | { | |
287 | zio_priority_t p = zio->io_priority; | |
288 | uint32_t empty; | |
289 | if (vdev_queue_class_fifo(p)) { | |
290 | list_t *list = &vq->vq_class[p].vqc_list; | |
291 | list_remove(list, zio); | |
292 | empty = list_is_empty(list); | |
2a154b84 | 293 | vq->vq_class[p].vqc_list_numnodes--; |
8469b5aa AM |
294 | } else { |
295 | avl_tree_t *tree = &vq->vq_class[p].vqc_tree; | |
296 | avl_remove(tree, zio); | |
297 | empty = avl_is_empty(tree); | |
298 | } | |
299 | vq->vq_cqueued &= ~(empty << p); | |
300 | } | |
301 | ||
fdc2d303 | 302 | static uint_t |
6f5aac3c | 303 | vdev_queue_class_min_active(vdev_queue_t *vq, zio_priority_t p) |
e8b96c60 MA |
304 | { |
305 | switch (p) { | |
306 | case ZIO_PRIORITY_SYNC_READ: | |
307 | return (zfs_vdev_sync_read_min_active); | |
308 | case ZIO_PRIORITY_SYNC_WRITE: | |
309 | return (zfs_vdev_sync_write_min_active); | |
310 | case ZIO_PRIORITY_ASYNC_READ: | |
311 | return (zfs_vdev_async_read_min_active); | |
312 | case ZIO_PRIORITY_ASYNC_WRITE: | |
313 | return (zfs_vdev_async_write_min_active); | |
314 | case ZIO_PRIORITY_SCRUB: | |
6f5aac3c AM |
315 | return (vq->vq_ia_active == 0 ? zfs_vdev_scrub_min_active : |
316 | MIN(vq->vq_nia_credit, zfs_vdev_scrub_min_active)); | |
a1d477c2 | 317 | case ZIO_PRIORITY_REMOVAL: |
6f5aac3c AM |
318 | return (vq->vq_ia_active == 0 ? zfs_vdev_removal_min_active : |
319 | MIN(vq->vq_nia_credit, zfs_vdev_removal_min_active)); | |
619f0976 | 320 | case ZIO_PRIORITY_INITIALIZING: |
6f5aac3c AM |
321 | return (vq->vq_ia_active == 0 ?zfs_vdev_initializing_min_active: |
322 | MIN(vq->vq_nia_credit, zfs_vdev_initializing_min_active)); | |
1b939560 BB |
323 | case ZIO_PRIORITY_TRIM: |
324 | return (zfs_vdev_trim_min_active); | |
9a49d3f3 | 325 | case ZIO_PRIORITY_REBUILD: |
6f5aac3c AM |
326 | return (vq->vq_ia_active == 0 ? zfs_vdev_rebuild_min_active : |
327 | MIN(vq->vq_nia_credit, zfs_vdev_rebuild_min_active)); | |
e8b96c60 MA |
328 | default: |
329 | panic("invalid priority %u", p); | |
330 | return (0); | |
331 | } | |
332 | } | |
333 | ||
fdc2d303 | 334 | static uint_t |
acbad6ff | 335 | vdev_queue_max_async_writes(spa_t *spa) |
e8b96c60 | 336 | { |
fdc2d303 | 337 | uint_t writes; |
b7faa7aa G |
338 | uint64_t dirty = 0; |
339 | dsl_pool_t *dp = spa_get_dsl(spa); | |
e8b96c60 MA |
340 | uint64_t min_bytes = zfs_dirty_data_max * |
341 | zfs_vdev_async_write_active_min_dirty_percent / 100; | |
342 | uint64_t max_bytes = zfs_dirty_data_max * | |
343 | zfs_vdev_async_write_active_max_dirty_percent / 100; | |
344 | ||
b7faa7aa G |
345 | /* |
346 | * Async writes may occur before the assignment of the spa's | |
347 | * dsl_pool_t if a self-healing zio is issued prior to the | |
348 | * completion of dmu_objset_open_impl(). | |
349 | */ | |
350 | if (dp == NULL) | |
351 | return (zfs_vdev_async_write_max_active); | |
352 | ||
acbad6ff AR |
353 | /* |
354 | * Sync tasks correspond to interactive user actions. To reduce the | |
355 | * execution time of those actions we push data out as fast as possible. | |
356 | */ | |
8136b9d7 AM |
357 | dirty = dp->dp_dirty_total; |
358 | if (dirty > max_bytes || spa_has_pending_synctask(spa)) | |
acbad6ff | 359 | return (zfs_vdev_async_write_max_active); |
acbad6ff | 360 | |
e8b96c60 MA |
361 | if (dirty < min_bytes) |
362 | return (zfs_vdev_async_write_min_active); | |
e8b96c60 MA |
363 | |
364 | /* | |
365 | * linear interpolation: | |
366 | * slope = (max_writes - min_writes) / (max_bytes - min_bytes) | |
367 | * move right by min_bytes | |
368 | * move up by min_writes | |
369 | */ | |
370 | writes = (dirty - min_bytes) * | |
371 | (zfs_vdev_async_write_max_active - | |
372 | zfs_vdev_async_write_min_active) / | |
373 | (max_bytes - min_bytes) + | |
374 | zfs_vdev_async_write_min_active; | |
375 | ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); | |
376 | ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); | |
377 | return (writes); | |
378 | } | |
379 | ||
fdc2d303 | 380 | static uint_t |
8469b5aa | 381 | vdev_queue_class_max_active(vdev_queue_t *vq, zio_priority_t p) |
e8b96c60 MA |
382 | { |
383 | switch (p) { | |
384 | case ZIO_PRIORITY_SYNC_READ: | |
385 | return (zfs_vdev_sync_read_max_active); | |
386 | case ZIO_PRIORITY_SYNC_WRITE: | |
387 | return (zfs_vdev_sync_write_max_active); | |
388 | case ZIO_PRIORITY_ASYNC_READ: | |
389 | return (zfs_vdev_async_read_max_active); | |
390 | case ZIO_PRIORITY_ASYNC_WRITE: | |
8469b5aa | 391 | return (vdev_queue_max_async_writes(vq->vq_vdev->vdev_spa)); |
e8b96c60 | 392 | case ZIO_PRIORITY_SCRUB: |
6f5aac3c AM |
393 | if (vq->vq_ia_active > 0) { |
394 | return (MIN(vq->vq_nia_credit, | |
395 | zfs_vdev_scrub_min_active)); | |
396 | } else if (vq->vq_nia_credit < zfs_vdev_nia_delay) | |
dcf70445 | 397 | return (MAX(1, zfs_vdev_scrub_min_active)); |
e8b96c60 | 398 | return (zfs_vdev_scrub_max_active); |
a1d477c2 | 399 | case ZIO_PRIORITY_REMOVAL: |
6f5aac3c AM |
400 | if (vq->vq_ia_active > 0) { |
401 | return (MIN(vq->vq_nia_credit, | |
402 | zfs_vdev_removal_min_active)); | |
403 | } else if (vq->vq_nia_credit < zfs_vdev_nia_delay) | |
dcf70445 | 404 | return (MAX(1, zfs_vdev_removal_min_active)); |
a1d477c2 | 405 | return (zfs_vdev_removal_max_active); |
619f0976 | 406 | case ZIO_PRIORITY_INITIALIZING: |
6f5aac3c AM |
407 | if (vq->vq_ia_active > 0) { |
408 | return (MIN(vq->vq_nia_credit, | |
409 | zfs_vdev_initializing_min_active)); | |
410 | } else if (vq->vq_nia_credit < zfs_vdev_nia_delay) | |
dcf70445 | 411 | return (MAX(1, zfs_vdev_initializing_min_active)); |
619f0976 | 412 | return (zfs_vdev_initializing_max_active); |
1b939560 BB |
413 | case ZIO_PRIORITY_TRIM: |
414 | return (zfs_vdev_trim_max_active); | |
9a49d3f3 | 415 | case ZIO_PRIORITY_REBUILD: |
6f5aac3c AM |
416 | if (vq->vq_ia_active > 0) { |
417 | return (MIN(vq->vq_nia_credit, | |
418 | zfs_vdev_rebuild_min_active)); | |
419 | } else if (vq->vq_nia_credit < zfs_vdev_nia_delay) | |
dcf70445 | 420 | return (MAX(1, zfs_vdev_rebuild_min_active)); |
9a49d3f3 | 421 | return (zfs_vdev_rebuild_max_active); |
e8b96c60 MA |
422 | default: |
423 | panic("invalid priority %u", p); | |
424 | return (0); | |
425 | } | |
426 | } | |
427 | ||
428 | /* | |
547df816 | 429 | * Return the i/o class to issue from, or ZIO_PRIORITY_NUM_QUEUEABLE if |
e8b96c60 MA |
430 | * there is no eligible class. |
431 | */ | |
432 | static zio_priority_t | |
433 | vdev_queue_class_to_issue(vdev_queue_t *vq) | |
434 | { | |
8469b5aa AM |
435 | uint32_t cq = vq->vq_cqueued; |
436 | zio_priority_t p, p1; | |
e8b96c60 | 437 | |
8469b5aa | 438 | if (cq == 0 || vq->vq_active >= zfs_vdev_max_active) |
e8b96c60 MA |
439 | return (ZIO_PRIORITY_NUM_QUEUEABLE); |
440 | ||
6f5aac3c AM |
441 | /* |
442 | * Find a queue that has not reached its minimum # outstanding i/os. | |
443 | * Do round-robin to reduce starvation due to zfs_vdev_max_active | |
444 | * and vq_nia_credit limits. | |
445 | */ | |
8469b5aa AM |
446 | p1 = vq->vq_last_prio + 1; |
447 | if (p1 >= ZIO_PRIORITY_NUM_QUEUEABLE) | |
448 | p1 = 0; | |
449 | for (p = p1; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { | |
450 | if ((cq & (1U << p)) != 0 && vq->vq_cactive[p] < | |
451 | vdev_queue_class_min_active(vq, p)) | |
452 | goto found; | |
453 | } | |
454 | for (p = 0; p < p1; p++) { | |
455 | if ((cq & (1U << p)) != 0 && vq->vq_cactive[p] < | |
456 | vdev_queue_class_min_active(vq, p)) | |
457 | goto found; | |
e8b96c60 MA |
458 | } |
459 | ||
460 | /* | |
461 | * If we haven't found a queue, look for one that hasn't reached its | |
462 | * maximum # outstanding i/os. | |
463 | */ | |
464 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { | |
8469b5aa AM |
465 | if ((cq & (1U << p)) != 0 && vq->vq_cactive[p] < |
466 | vdev_queue_class_max_active(vq, p)) | |
467 | break; | |
e8b96c60 MA |
468 | } |
469 | ||
8469b5aa AM |
470 | found: |
471 | vq->vq_last_prio = p; | |
472 | return (p); | |
e8b96c60 MA |
473 | } |
474 | ||
34dc7c2f BB |
475 | void |
476 | vdev_queue_init(vdev_t *vd) | |
477 | { | |
478 | vdev_queue_t *vq = &vd->vdev_queue; | |
e8b96c60 | 479 | zio_priority_t p; |
34dc7c2f | 480 | |
e8b96c60 | 481 | vq->vq_vdev = vd; |
34dc7c2f | 482 | |
e8b96c60 | 483 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { |
8469b5aa AM |
484 | if (vdev_queue_class_fifo(p)) { |
485 | list_create(&vq->vq_class[p].vqc_list, | |
486 | sizeof (zio_t), | |
487 | offsetof(struct zio, io_queue_node.l)); | |
1b939560 | 488 | } else { |
8469b5aa AM |
489 | avl_create(&vq->vq_class[p].vqc_tree, |
490 | vdev_queue_to_compare, sizeof (zio_t), | |
491 | offsetof(struct zio, io_queue_node.a)); | |
1b939560 | 492 | } |
e8b96c60 | 493 | } |
8469b5aa AM |
494 | avl_create(&vq->vq_read_offset_tree, |
495 | vdev_queue_offset_compare, sizeof (zio_t), | |
496 | offsetof(struct zio, io_offset_node)); | |
497 | avl_create(&vq->vq_write_offset_tree, | |
498 | vdev_queue_offset_compare, sizeof (zio_t), | |
499 | offsetof(struct zio, io_offset_node)); | |
9f500936 | 500 | |
d6c6590c | 501 | vq->vq_last_offset = 0; |
8469b5aa AM |
502 | list_create(&vq->vq_active_list, sizeof (struct zio), |
503 | offsetof(struct zio, io_queue_node.l)); | |
504 | mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); | |
34dc7c2f BB |
505 | } |
506 | ||
507 | void | |
508 | vdev_queue_fini(vdev_t *vd) | |
509 | { | |
510 | vdev_queue_t *vq = &vd->vdev_queue; | |
511 | ||
8469b5aa AM |
512 | for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { |
513 | if (vdev_queue_class_fifo(p)) | |
514 | list_destroy(&vq->vq_class[p].vqc_list); | |
515 | else | |
516 | avl_destroy(&vq->vq_class[p].vqc_tree); | |
517 | } | |
518 | avl_destroy(&vq->vq_read_offset_tree); | |
519 | avl_destroy(&vq->vq_write_offset_tree); | |
34dc7c2f | 520 | |
8469b5aa | 521 | list_destroy(&vq->vq_active_list); |
34dc7c2f BB |
522 | mutex_destroy(&vq->vq_lock); |
523 | } | |
524 | ||
525 | static void | |
526 | vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) | |
527 | { | |
8469b5aa AM |
528 | zio->io_queue_state = ZIO_QS_QUEUED; |
529 | vdev_queue_class_add(vq, zio); | |
530 | if (zio->io_type == ZIO_TYPE_READ) | |
531 | avl_add(&vq->vq_read_offset_tree, zio); | |
532 | else if (zio->io_type == ZIO_TYPE_WRITE) | |
533 | avl_add(&vq->vq_write_offset_tree, zio); | |
34dc7c2f BB |
534 | } |
535 | ||
536 | static void | |
537 | vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) | |
538 | { | |
8469b5aa AM |
539 | vdev_queue_class_remove(vq, zio); |
540 | if (zio->io_type == ZIO_TYPE_READ) | |
541 | avl_remove(&vq->vq_read_offset_tree, zio); | |
542 | else if (zio->io_type == ZIO_TYPE_WRITE) | |
543 | avl_remove(&vq->vq_write_offset_tree, zio); | |
544 | zio->io_queue_state = ZIO_QS_NONE; | |
330847ff MA |
545 | } |
546 | ||
6f5aac3c AM |
547 | static boolean_t |
548 | vdev_queue_is_interactive(zio_priority_t p) | |
549 | { | |
550 | switch (p) { | |
551 | case ZIO_PRIORITY_SCRUB: | |
552 | case ZIO_PRIORITY_REMOVAL: | |
553 | case ZIO_PRIORITY_INITIALIZING: | |
554 | case ZIO_PRIORITY_REBUILD: | |
555 | return (B_FALSE); | |
556 | default: | |
557 | return (B_TRUE); | |
558 | } | |
559 | } | |
560 | ||
330847ff MA |
561 | static void |
562 | vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) | |
563 | { | |
e8b96c60 MA |
564 | ASSERT(MUTEX_HELD(&vq->vq_lock)); |
565 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
8469b5aa AM |
566 | vq->vq_cactive[zio->io_priority]++; |
567 | vq->vq_active++; | |
6f5aac3c AM |
568 | if (vdev_queue_is_interactive(zio->io_priority)) { |
569 | if (++vq->vq_ia_active == 1) | |
570 | vq->vq_nia_credit = 1; | |
571 | } else if (vq->vq_ia_active > 0) { | |
572 | vq->vq_nia_credit--; | |
573 | } | |
8469b5aa AM |
574 | zio->io_queue_state = ZIO_QS_ACTIVE; |
575 | list_insert_tail(&vq->vq_active_list, zio); | |
330847ff MA |
576 | } |
577 | ||
578 | static void | |
579 | vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) | |
580 | { | |
e8b96c60 MA |
581 | ASSERT(MUTEX_HELD(&vq->vq_lock)); |
582 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
8469b5aa AM |
583 | vq->vq_cactive[zio->io_priority]--; |
584 | vq->vq_active--; | |
6f5aac3c AM |
585 | if (vdev_queue_is_interactive(zio->io_priority)) { |
586 | if (--vq->vq_ia_active == 0) | |
587 | vq->vq_nia_credit = 0; | |
588 | else | |
589 | vq->vq_nia_credit = zfs_vdev_nia_credit; | |
590 | } else if (vq->vq_ia_active == 0) | |
591 | vq->vq_nia_credit++; | |
8469b5aa AM |
592 | list_remove(&vq->vq_active_list, zio); |
593 | zio->io_queue_state = ZIO_QS_NONE; | |
34dc7c2f BB |
594 | } |
595 | ||
596 | static void | |
597 | vdev_queue_agg_io_done(zio_t *aio) | |
598 | { | |
a6255b7f | 599 | abd_free(aio->io_abd); |
34dc7c2f BB |
600 | } |
601 | ||
9babb374 BB |
602 | /* |
603 | * Compute the range spanned by two i/os, which is the endpoint of the last | |
604 | * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). | |
605 | * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); | |
606 | * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. | |
607 | */ | |
608 | #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) | |
609 | #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) | |
34dc7c2f | 610 | |
fb822260 BA |
611 | /* |
612 | * Sufficiently adjacent io_offset's in ZIOs will be aggregated. We do this | |
613 | * by creating a gang ABD from the adjacent ZIOs io_abd's. By using | |
614 | * a gang ABD we avoid doing memory copies to and from the parent, | |
615 | * child ZIOs. The gang ABD also accounts for gaps between adjacent | |
616 | * io_offsets by simply getting the zero ABD for writes or allocating | |
617 | * a new ABD for reads and placing them in the gang ABD as well. | |
618 | */ | |
34dc7c2f | 619 | static zio_t * |
e8b96c60 | 620 | vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) |
34dc7c2f | 621 | { |
e8b96c60 MA |
622 | zio_t *first, *last, *aio, *dio, *mandatory, *nio; |
623 | uint64_t maxgap = 0; | |
624 | uint64_t size; | |
a58df6f5 | 625 | uint64_t limit; |
e8b96c60 | 626 | boolean_t stretch = B_FALSE; |
fb822260 | 627 | uint64_t next_offset; |
a6255b7f | 628 | abd_t *abd; |
8469b5aa AM |
629 | avl_tree_t *t; |
630 | ||
631 | /* | |
632 | * TRIM aggregation should not be needed since code in zfs_trim.c can | |
633 | * submit TRIM I/O for extents up to zfs_trim_extent_bytes_max (128M). | |
634 | */ | |
635 | if (zio->io_type == ZIO_TYPE_TRIM) | |
636 | return (NULL); | |
637 | ||
638 | if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE) | |
639 | return (NULL); | |
e8b96c60 | 640 | |
1af240f3 AM |
641 | if (vq->vq_vdev->vdev_nonrot) |
642 | limit = zfs_vdev_aggregation_limit_non_rotating; | |
643 | else | |
644 | limit = zfs_vdev_aggregation_limit; | |
8469b5aa | 645 | if (limit == 0) |
1b939560 | 646 | return (NULL); |
8469b5aa | 647 | limit = MIN(limit, SPA_MAXBLOCKSIZE); |
1b939560 | 648 | |
b2255edc BB |
649 | /* |
650 | * I/Os to distributed spares are directly dispatched to the dRAID | |
651 | * leaf vdevs for aggregation. See the comment at the end of the | |
652 | * zio_vdev_io_start() function. | |
653 | */ | |
654 | ASSERT(vq->vq_vdev->vdev_ops != &vdev_draid_spare_ops); | |
655 | ||
e8b96c60 | 656 | first = last = zio; |
34dc7c2f | 657 | |
8469b5aa | 658 | if (zio->io_type == ZIO_TYPE_READ) { |
e8b96c60 | 659 | maxgap = zfs_vdev_read_gap_limit; |
8469b5aa AM |
660 | t = &vq->vq_read_offset_tree; |
661 | } else { | |
662 | ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); | |
663 | t = &vq->vq_write_offset_tree; | |
664 | } | |
fb5f0bc8 | 665 | |
e8b96c60 MA |
666 | /* |
667 | * We can aggregate I/Os that are sufficiently adjacent and of | |
668 | * the same flavor, as expressed by the AGG_INHERIT flags. | |
669 | * The latter requirement is necessary so that certain | |
670 | * attributes of the I/O, such as whether it's a normal I/O | |
671 | * or a scrub/resilver, can be preserved in the aggregate. | |
672 | * We can include optional I/Os, but don't allow them | |
673 | * to begin a range as they add no benefit in that situation. | |
674 | */ | |
45d1cae3 | 675 | |
e8b96c60 MA |
676 | /* |
677 | * We keep track of the last non-optional I/O. | |
678 | */ | |
679 | mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; | |
45d1cae3 | 680 | |
e8b96c60 MA |
681 | /* |
682 | * Walk backwards through sufficiently contiguous I/Os | |
8542ef85 | 683 | * recording the last non-optional I/O. |
e8b96c60 | 684 | */ |
8469b5aa | 685 | zio_flag_t flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; |
e8b96c60 MA |
686 | while ((dio = AVL_PREV(t, first)) != NULL && |
687 | (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && | |
a58df6f5 | 688 | IO_SPAN(dio, last) <= limit && |
a1d477c2 MA |
689 | IO_GAP(dio, first) <= maxgap && |
690 | dio->io_type == zio->io_type) { | |
e8b96c60 MA |
691 | first = dio; |
692 | if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) | |
693 | mandatory = first; | |
694 | } | |
45d1cae3 | 695 | |
e8b96c60 MA |
696 | /* |
697 | * Skip any initial optional I/Os. | |
698 | */ | |
699 | while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { | |
700 | first = AVL_NEXT(t, first); | |
701 | ASSERT(first != NULL); | |
702 | } | |
9babb374 | 703 | |
45d1cae3 | 704 | |
e8b96c60 MA |
705 | /* |
706 | * Walk forward through sufficiently contiguous I/Os. | |
8542ef85 MA |
707 | * The aggregation limit does not apply to optional i/os, so that |
708 | * we can issue contiguous writes even if they are larger than the | |
709 | * aggregation limit. | |
e8b96c60 MA |
710 | */ |
711 | while ((dio = AVL_NEXT(t, last)) != NULL && | |
712 | (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && | |
8542ef85 MA |
713 | (IO_SPAN(first, dio) <= limit || |
714 | (dio->io_flags & ZIO_FLAG_OPTIONAL)) && | |
8469b5aa | 715 | IO_SPAN(first, dio) <= SPA_MAXBLOCKSIZE && |
a1d477c2 MA |
716 | IO_GAP(last, dio) <= maxgap && |
717 | dio->io_type == zio->io_type) { | |
e8b96c60 MA |
718 | last = dio; |
719 | if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) | |
720 | mandatory = last; | |
721 | } | |
722 | ||
723 | /* | |
724 | * Now that we've established the range of the I/O aggregation | |
725 | * we must decide what to do with trailing optional I/Os. | |
726 | * For reads, there's nothing to do. While we are unable to | |
727 | * aggregate further, it's possible that a trailing optional | |
728 | * I/O would allow the underlying device to aggregate with | |
729 | * subsequent I/Os. We must therefore determine if the next | |
730 | * non-optional I/O is close enough to make aggregation | |
731 | * worthwhile. | |
732 | */ | |
733 | if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { | |
734 | zio_t *nio = last; | |
735 | while ((dio = AVL_NEXT(t, nio)) != NULL && | |
736 | IO_GAP(nio, dio) == 0 && | |
737 | IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { | |
738 | nio = dio; | |
739 | if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { | |
740 | stretch = B_TRUE; | |
741 | break; | |
45d1cae3 BB |
742 | } |
743 | } | |
e8b96c60 | 744 | } |
45d1cae3 | 745 | |
e8b96c60 | 746 | if (stretch) { |
8542ef85 MA |
747 | /* |
748 | * We are going to include an optional io in our aggregated | |
749 | * span, thus closing the write gap. Only mandatory i/os can | |
750 | * start aggregated spans, so make sure that the next i/o | |
751 | * after our span is mandatory. | |
752 | */ | |
e8b96c60 | 753 | dio = AVL_NEXT(t, last); |
411d327c | 754 | ASSERT3P(dio, !=, NULL); |
e8b96c60 MA |
755 | dio->io_flags &= ~ZIO_FLAG_OPTIONAL; |
756 | } else { | |
8542ef85 | 757 | /* do not include the optional i/o */ |
e8b96c60 MA |
758 | while (last != mandatory && last != first) { |
759 | ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); | |
760 | last = AVL_PREV(t, last); | |
761 | ASSERT(last != NULL); | |
45d1cae3 | 762 | } |
34dc7c2f BB |
763 | } |
764 | ||
e8b96c60 MA |
765 | if (first == last) |
766 | return (NULL); | |
767 | ||
e8b96c60 | 768 | size = IO_SPAN(first, last); |
8469b5aa | 769 | ASSERT3U(size, <=, SPA_MAXBLOCKSIZE); |
e8b96c60 | 770 | |
e2af2acc | 771 | abd = abd_alloc_gang(); |
a6255b7f | 772 | if (abd == NULL) |
6fe53787 BB |
773 | return (NULL); |
774 | ||
e8b96c60 | 775 | aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, |
a6255b7f | 776 | abd, size, first->io_type, zio->io_priority, |
70ea484e | 777 | flags | ZIO_FLAG_DONT_QUEUE, vdev_queue_agg_io_done, NULL); |
e8b96c60 MA |
778 | aio->io_timestamp = first->io_timestamp; |
779 | ||
780 | nio = first; | |
fb822260 | 781 | next_offset = first->io_offset; |
e8b96c60 MA |
782 | do { |
783 | dio = nio; | |
784 | nio = AVL_NEXT(t, dio); | |
a6ccb36b | 785 | ASSERT3P(dio, !=, NULL); |
ae5c78e0 AM |
786 | zio_add_child(dio, aio); |
787 | vdev_queue_io_remove(vq, dio); | |
fb822260 BA |
788 | |
789 | if (dio->io_offset != next_offset) { | |
790 | /* allocate a buffer for a read gap */ | |
791 | ASSERT3U(dio->io_type, ==, ZIO_TYPE_READ); | |
792 | ASSERT3U(dio->io_offset, >, next_offset); | |
793 | abd = abd_alloc_for_io( | |
794 | dio->io_offset - next_offset, B_TRUE); | |
795 | abd_gang_add(aio->io_abd, abd, B_TRUE); | |
796 | } | |
797 | if (dio->io_abd && | |
798 | (dio->io_size != abd_get_size(dio->io_abd))) { | |
799 | /* abd size not the same as IO size */ | |
800 | ASSERT3U(abd_get_size(dio->io_abd), >, dio->io_size); | |
801 | abd = abd_get_offset_size(dio->io_abd, 0, dio->io_size); | |
802 | abd_gang_add(aio->io_abd, abd, B_TRUE); | |
803 | } else { | |
804 | if (dio->io_flags & ZIO_FLAG_NODATA) { | |
805 | /* allocate a buffer for a write gap */ | |
806 | ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); | |
807 | ASSERT3P(dio->io_abd, ==, NULL); | |
808 | abd_gang_add(aio->io_abd, | |
809 | abd_get_zeros(dio->io_size), B_TRUE); | |
810 | } else { | |
811 | /* | |
812 | * We pass B_FALSE to abd_gang_add() | |
813 | * because we did not allocate a new | |
814 | * ABD, so it is assumed the caller | |
815 | * will free this ABD. | |
816 | */ | |
817 | abd_gang_add(aio->io_abd, dio->io_abd, | |
818 | B_FALSE); | |
819 | } | |
820 | } | |
821 | next_offset = dio->io_offset + dio->io_size; | |
ae5c78e0 | 822 | } while (dio != last); |
fb822260 | 823 | ASSERT3U(abd_get_size(aio->io_abd), ==, aio->io_size); |
ae5c78e0 AM |
824 | |
825 | /* | |
7f9d9e6f AM |
826 | * Callers must call zio_vdev_io_bypass() and zio_execute() for |
827 | * aggregated (parent) I/Os so that we could avoid dropping the | |
828 | * queue's lock here to avoid a deadlock that we could encounter | |
829 | * due to lock order reversal between vq_lock and io_lock in | |
830 | * zio_change_priority(). | |
ae5c78e0 | 831 | */ |
e8b96c60 MA |
832 | return (aio); |
833 | } | |
834 | ||
835 | static zio_t * | |
836 | vdev_queue_io_to_issue(vdev_queue_t *vq) | |
837 | { | |
838 | zio_t *zio, *aio; | |
839 | zio_priority_t p; | |
840 | avl_index_t idx; | |
ec8501ee | 841 | avl_tree_t *tree; |
e8b96c60 MA |
842 | |
843 | again: | |
844 | ASSERT(MUTEX_HELD(&vq->vq_lock)); | |
845 | ||
846 | p = vdev_queue_class_to_issue(vq); | |
34dc7c2f | 847 | |
e8b96c60 MA |
848 | if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { |
849 | /* No eligible queued i/os */ | |
850 | return (NULL); | |
34dc7c2f BB |
851 | } |
852 | ||
8469b5aa AM |
853 | if (vdev_queue_class_fifo(p)) { |
854 | zio = list_head(&vq->vq_class[p].vqc_list); | |
855 | } else { | |
856 | /* | |
857 | * For LBA-ordered queues (async / scrub / initializing), | |
858 | * issue the I/O which follows the most recently issued I/O | |
859 | * in LBA (offset) order, but to avoid starvation only within | |
860 | * the same 0.5 second interval as the first I/O. | |
861 | */ | |
862 | tree = &vq->vq_class[p].vqc_tree; | |
863 | zio = aio = avl_first(tree); | |
864 | if (zio->io_offset < vq->vq_last_offset) { | |
865 | vq->vq_io_search.io_timestamp = zio->io_timestamp; | |
866 | vq->vq_io_search.io_offset = vq->vq_last_offset; | |
867 | zio = avl_find(tree, &vq->vq_io_search, &idx); | |
868 | if (zio == NULL) { | |
869 | zio = avl_nearest(tree, idx, AVL_AFTER); | |
870 | if (zio == NULL || | |
871 | (zio->io_timestamp >> VDQ_T_SHIFT) != | |
872 | (aio->io_timestamp >> VDQ_T_SHIFT)) | |
873 | zio = aio; | |
874 | } | |
875 | } | |
876 | } | |
e8b96c60 MA |
877 | ASSERT3U(zio->io_priority, ==, p); |
878 | ||
879 | aio = vdev_queue_aggregate(vq, zio); | |
7f9d9e6f | 880 | if (aio != NULL) { |
e8b96c60 | 881 | zio = aio; |
7f9d9e6f | 882 | } else { |
e8b96c60 | 883 | vdev_queue_io_remove(vq, zio); |
34dc7c2f | 884 | |
7f9d9e6f AM |
885 | /* |
886 | * If the I/O is or was optional and therefore has no data, we | |
887 | * need to simply discard it. We need to drop the vdev queue's | |
888 | * lock to avoid a deadlock that we could encounter since this | |
889 | * I/O will complete immediately. | |
890 | */ | |
891 | if (zio->io_flags & ZIO_FLAG_NODATA) { | |
892 | mutex_exit(&vq->vq_lock); | |
893 | zio_vdev_io_bypass(zio); | |
894 | zio_execute(zio); | |
895 | mutex_enter(&vq->vq_lock); | |
896 | goto again; | |
897 | } | |
45d1cae3 BB |
898 | } |
899 | ||
e8b96c60 | 900 | vdev_queue_pending_add(vq, zio); |
d6c6590c | 901 | vq->vq_last_offset = zio->io_offset + zio->io_size; |
34dc7c2f | 902 | |
e8b96c60 | 903 | return (zio); |
34dc7c2f BB |
904 | } |
905 | ||
906 | zio_t * | |
907 | vdev_queue_io(zio_t *zio) | |
908 | { | |
909 | vdev_queue_t *vq = &zio->io_vd->vdev_queue; | |
7f9d9e6f AM |
910 | zio_t *dio, *nio; |
911 | zio_link_t *zl = NULL; | |
34dc7c2f | 912 | |
34dc7c2f BB |
913 | if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) |
914 | return (zio); | |
915 | ||
e8b96c60 MA |
916 | /* |
917 | * Children i/os inherent their parent's priority, which might | |
918 | * not match the child's i/o type. Fix it up here. | |
919 | */ | |
920 | if (zio->io_type == ZIO_TYPE_READ) { | |
1b939560 BB |
921 | ASSERT(zio->io_priority != ZIO_PRIORITY_TRIM); |
922 | ||
e8b96c60 MA |
923 | if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && |
924 | zio->io_priority != ZIO_PRIORITY_ASYNC_READ && | |
a1d477c2 | 925 | zio->io_priority != ZIO_PRIORITY_SCRUB && |
619f0976 | 926 | zio->io_priority != ZIO_PRIORITY_REMOVAL && |
9a49d3f3 BB |
927 | zio->io_priority != ZIO_PRIORITY_INITIALIZING && |
928 | zio->io_priority != ZIO_PRIORITY_REBUILD) { | |
e8b96c60 | 929 | zio->io_priority = ZIO_PRIORITY_ASYNC_READ; |
1b939560 BB |
930 | } |
931 | } else if (zio->io_type == ZIO_TYPE_WRITE) { | |
932 | ASSERT(zio->io_priority != ZIO_PRIORITY_TRIM); | |
933 | ||
e8b96c60 | 934 | if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && |
a1d477c2 | 935 | zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE && |
619f0976 | 936 | zio->io_priority != ZIO_PRIORITY_REMOVAL && |
9a49d3f3 BB |
937 | zio->io_priority != ZIO_PRIORITY_INITIALIZING && |
938 | zio->io_priority != ZIO_PRIORITY_REBUILD) { | |
e8b96c60 | 939 | zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; |
1b939560 BB |
940 | } |
941 | } else { | |
942 | ASSERT(zio->io_type == ZIO_TYPE_TRIM); | |
943 | ASSERT(zio->io_priority == ZIO_PRIORITY_TRIM); | |
e8b96c60 | 944 | } |
34dc7c2f | 945 | |
70ea484e | 946 | zio->io_flags |= ZIO_FLAG_DONT_QUEUE; |
97752ba2 | 947 | zio->io_timestamp = gethrtime(); |
34dc7c2f BB |
948 | |
949 | mutex_enter(&vq->vq_lock); | |
34dc7c2f | 950 | vdev_queue_io_add(vq, zio); |
e8b96c60 | 951 | nio = vdev_queue_io_to_issue(vq); |
34dc7c2f BB |
952 | mutex_exit(&vq->vq_lock); |
953 | ||
954 | if (nio == NULL) | |
955 | return (NULL); | |
956 | ||
957 | if (nio->io_done == vdev_queue_agg_io_done) { | |
7f9d9e6f AM |
958 | while ((dio = zio_walk_parents(nio, &zl)) != NULL) { |
959 | ASSERT3U(dio->io_type, ==, nio->io_type); | |
960 | zio_vdev_io_bypass(dio); | |
961 | zio_execute(dio); | |
962 | } | |
34dc7c2f BB |
963 | zio_nowait(nio); |
964 | return (NULL); | |
965 | } | |
966 | ||
967 | return (nio); | |
968 | } | |
969 | ||
970 | void | |
971 | vdev_queue_io_done(zio_t *zio) | |
972 | { | |
973 | vdev_queue_t *vq = &zio->io_vd->vdev_queue; | |
7f9d9e6f AM |
974 | zio_t *dio, *nio; |
975 | zio_link_t *zl = NULL; | |
34dc7c2f | 976 | |
97752ba2 AM |
977 | hrtime_t now = gethrtime(); |
978 | vq->vq_io_complete_ts = now; | |
979 | vq->vq_io_delta_ts = zio->io_delta = now - zio->io_timestamp; | |
34dc7c2f | 980 | |
97752ba2 | 981 | mutex_enter(&vq->vq_lock); |
330847ff | 982 | vdev_queue_pending_remove(vq, zio); |
34dc7c2f | 983 | |
e8b96c60 | 984 | while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { |
34dc7c2f BB |
985 | mutex_exit(&vq->vq_lock); |
986 | if (nio->io_done == vdev_queue_agg_io_done) { | |
7f9d9e6f AM |
987 | while ((dio = zio_walk_parents(nio, &zl)) != NULL) { |
988 | ASSERT3U(dio->io_type, ==, nio->io_type); | |
989 | zio_vdev_io_bypass(dio); | |
990 | zio_execute(dio); | |
991 | } | |
34dc7c2f BB |
992 | zio_nowait(nio); |
993 | } else { | |
994 | zio_vdev_io_reissue(nio); | |
995 | zio_execute(nio); | |
996 | } | |
997 | mutex_enter(&vq->vq_lock); | |
998 | } | |
999 | ||
1000 | mutex_exit(&vq->vq_lock); | |
1001 | } | |
c28b2279 | 1002 | |
a8b2e306 TC |
1003 | void |
1004 | vdev_queue_change_io_priority(zio_t *zio, zio_priority_t priority) | |
1005 | { | |
1006 | vdev_queue_t *vq = &zio->io_vd->vdev_queue; | |
a8b2e306 | 1007 | |
c26cf096 TH |
1008 | /* |
1009 | * ZIO_PRIORITY_NOW is used by the vdev cache code and the aggregate zio | |
1010 | * code to issue IOs without adding them to the vdev queue. In this | |
1011 | * case, the zio is already going to be issued as quickly as possible | |
e1cfd73f | 1012 | * and so it doesn't need any reprioritization to help. |
c26cf096 TH |
1013 | */ |
1014 | if (zio->io_priority == ZIO_PRIORITY_NOW) | |
1015 | return; | |
1016 | ||
a8b2e306 TC |
1017 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); |
1018 | ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
1019 | ||
1020 | if (zio->io_type == ZIO_TYPE_READ) { | |
1021 | if (priority != ZIO_PRIORITY_SYNC_READ && | |
1022 | priority != ZIO_PRIORITY_ASYNC_READ && | |
1023 | priority != ZIO_PRIORITY_SCRUB) | |
1024 | priority = ZIO_PRIORITY_ASYNC_READ; | |
1025 | } else { | |
1026 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
1027 | if (priority != ZIO_PRIORITY_SYNC_WRITE && | |
1028 | priority != ZIO_PRIORITY_ASYNC_WRITE) | |
1029 | priority = ZIO_PRIORITY_ASYNC_WRITE; | |
1030 | } | |
1031 | ||
1032 | mutex_enter(&vq->vq_lock); | |
1033 | ||
1034 | /* | |
1035 | * If the zio is in none of the queues we can simply change | |
1036 | * the priority. If the zio is waiting to be submitted we must | |
1037 | * remove it from the queue and re-insert it with the new priority. | |
1038 | * Otherwise, the zio is currently active and we cannot change its | |
1039 | * priority. | |
1040 | */ | |
8469b5aa AM |
1041 | if (zio->io_queue_state == ZIO_QS_QUEUED) { |
1042 | vdev_queue_class_remove(vq, zio); | |
a8b2e306 | 1043 | zio->io_priority = priority; |
8469b5aa AM |
1044 | vdev_queue_class_add(vq, zio); |
1045 | } else if (zio->io_queue_state == ZIO_QS_NONE) { | |
a8b2e306 TC |
1046 | zio->io_priority = priority; |
1047 | } | |
1048 | ||
1049 | mutex_exit(&vq->vq_lock); | |
1050 | } | |
1051 | ||
9f500936 | 1052 | /* |
d6c6590c | 1053 | * As these two methods are only used for load calculations we're not |
9f500936 | 1054 | * concerned if we get an incorrect value on 32bit platforms due to lack of |
1055 | * vq_lock mutex use here, instead we prefer to keep it lock free for | |
1056 | * performance. | |
1057 | */ | |
8469b5aa | 1058 | uint32_t |
9f500936 | 1059 | vdev_queue_length(vdev_t *vd) |
1060 | { | |
8469b5aa | 1061 | return (vd->vdev_queue.vq_active); |
9f500936 | 1062 | } |
1063 | ||
1064 | uint64_t | |
d6c6590c | 1065 | vdev_queue_last_offset(vdev_t *vd) |
9f500936 | 1066 | { |
d6c6590c | 1067 | return (vd->vdev_queue.vq_last_offset); |
9f500936 | 1068 | } |
1069 | ||
8469b5aa AM |
1070 | uint64_t |
1071 | vdev_queue_class_length(vdev_t *vd, zio_priority_t p) | |
1072 | { | |
1073 | vdev_queue_t *vq = &vd->vdev_queue; | |
1074 | if (vdev_queue_class_fifo(p)) | |
2a154b84 | 1075 | return (vq->vq_class[p].vqc_list_numnodes); |
8469b5aa AM |
1076 | else |
1077 | return (avl_numnodes(&vq->vq_class[p].vqc_tree)); | |
1078 | } | |
1079 | ||
fdc2d303 | 1080 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, aggregation_limit, UINT, ZMOD_RW, |
03fdcb9a | 1081 | "Max vdev I/O aggregation size"); |
c409e464 | 1082 | |
fdc2d303 | 1083 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, aggregation_limit_non_rotating, UINT, |
7ada752a | 1084 | ZMOD_RW, "Max vdev I/O aggregation size for non-rotating media"); |
1af240f3 | 1085 | |
fdc2d303 | 1086 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, read_gap_limit, UINT, ZMOD_RW, |
03fdcb9a | 1087 | "Aggregate read I/O over gap"); |
c409e464 | 1088 | |
fdc2d303 | 1089 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, write_gap_limit, UINT, ZMOD_RW, |
03fdcb9a | 1090 | "Aggregate write I/O over gap"); |
e8b96c60 | 1091 | |
fdc2d303 | 1092 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, max_active, UINT, ZMOD_RW, |
03fdcb9a | 1093 | "Maximum number of active I/Os per vdev"); |
e8b96c60 | 1094 | |
fdc2d303 RY |
1095 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_write_active_max_dirty_percent, |
1096 | UINT, ZMOD_RW, "Async write concurrency max threshold"); | |
e8b96c60 | 1097 | |
fdc2d303 RY |
1098 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_write_active_min_dirty_percent, |
1099 | UINT, ZMOD_RW, "Async write concurrency min threshold"); | |
e8b96c60 | 1100 | |
fdc2d303 | 1101 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_read_max_active, UINT, ZMOD_RW, |
d1d7e268 | 1102 | "Max active async read I/Os per vdev"); |
e8b96c60 | 1103 | |
fdc2d303 | 1104 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_read_min_active, UINT, ZMOD_RW, |
d1d7e268 | 1105 | "Min active async read I/Os per vdev"); |
e8b96c60 | 1106 | |
fdc2d303 | 1107 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_write_max_active, UINT, ZMOD_RW, |
d1d7e268 | 1108 | "Max active async write I/Os per vdev"); |
e8b96c60 | 1109 | |
fdc2d303 | 1110 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_write_min_active, UINT, ZMOD_RW, |
d1d7e268 | 1111 | "Min active async write I/Os per vdev"); |
e8b96c60 | 1112 | |
fdc2d303 | 1113 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, initializing_max_active, UINT, ZMOD_RW, |
619f0976 GW |
1114 | "Max active initializing I/Os per vdev"); |
1115 | ||
fdc2d303 | 1116 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, initializing_min_active, UINT, ZMOD_RW, |
619f0976 GW |
1117 | "Min active initializing I/Os per vdev"); |
1118 | ||
fdc2d303 | 1119 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, removal_max_active, UINT, ZMOD_RW, |
619f0976 GW |
1120 | "Max active removal I/Os per vdev"); |
1121 | ||
fdc2d303 | 1122 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, removal_min_active, UINT, ZMOD_RW, |
619f0976 GW |
1123 | "Min active removal I/Os per vdev"); |
1124 | ||
fdc2d303 | 1125 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, scrub_max_active, UINT, ZMOD_RW, |
619f0976 | 1126 | "Max active scrub I/Os per vdev"); |
e8b96c60 | 1127 | |
fdc2d303 | 1128 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, scrub_min_active, UINT, ZMOD_RW, |
619f0976 | 1129 | "Min active scrub I/Os per vdev"); |
e8b96c60 | 1130 | |
fdc2d303 | 1131 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, sync_read_max_active, UINT, ZMOD_RW, |
d1d7e268 | 1132 | "Max active sync read I/Os per vdev"); |
e8b96c60 | 1133 | |
fdc2d303 | 1134 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, sync_read_min_active, UINT, ZMOD_RW, |
d1d7e268 | 1135 | "Min active sync read I/Os per vdev"); |
e8b96c60 | 1136 | |
fdc2d303 | 1137 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, sync_write_max_active, UINT, ZMOD_RW, |
d1d7e268 | 1138 | "Max active sync write I/Os per vdev"); |
e8b96c60 | 1139 | |
fdc2d303 | 1140 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, sync_write_min_active, UINT, ZMOD_RW, |
3757bff3 | 1141 | "Min active sync write I/Os per vdev"); |
3dfb57a3 | 1142 | |
fdc2d303 | 1143 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, trim_max_active, UINT, ZMOD_RW, |
1b939560 BB |
1144 | "Max active trim/discard I/Os per vdev"); |
1145 | ||
fdc2d303 | 1146 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, trim_min_active, UINT, ZMOD_RW, |
1b939560 BB |
1147 | "Min active trim/discard I/Os per vdev"); |
1148 | ||
fdc2d303 | 1149 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, rebuild_max_active, UINT, ZMOD_RW, |
9a49d3f3 BB |
1150 | "Max active rebuild I/Os per vdev"); |
1151 | ||
fdc2d303 | 1152 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, rebuild_min_active, UINT, ZMOD_RW, |
9a49d3f3 BB |
1153 | "Min active rebuild I/Os per vdev"); |
1154 | ||
fdc2d303 | 1155 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, nia_credit, UINT, ZMOD_RW, |
6f5aac3c AM |
1156 | "Number of non-interactive I/Os to allow in sequence"); |
1157 | ||
fdc2d303 | 1158 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, nia_delay, UINT, ZMOD_RW, |
6f5aac3c AM |
1159 | "Number of non-interactive I/Os before _max_active"); |
1160 | ||
fdc2d303 | 1161 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, queue_depth_pct, UINT, ZMOD_RW, |
3dfb57a3 | 1162 | "Queue depth percentage for each top-level vdev"); |
ece7ab7e RN |
1163 | |
1164 | ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, def_queue_depth, UINT, ZMOD_RW, | |
1165 | "Default queue depth for each allocator"); |