]>
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 | |
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 | /* | |
d164b209 | 22 | * Copyright 2009 Sun Microsystems, Inc. All rights reserved. |
34dc7c2f BB |
23 | * Use is subject to license terms. |
24 | */ | |
25 | ||
cc92e9d0 | 26 | /* |
acbad6ff | 27 | * Copyright (c) 2012, 2014 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 MA |
35 | #include <sys/dsl_pool.h> |
36 | #include <sys/spa.h> | |
37 | #include <sys/spa_impl.h> | |
330847ff | 38 | #include <sys/kstat.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 MA |
121 | * The maximum number of i/os active to each device. Ideally, this will be >= |
122 | * the sum of each queue's max_active. It must be at least the sum of each | |
123 | * queue's min_active. | |
34dc7c2f | 124 | */ |
e8b96c60 | 125 | uint32_t zfs_vdev_max_active = 1000; |
34dc7c2f | 126 | |
cb682a17 | 127 | /* |
e8b96c60 MA |
128 | * Per-queue limits on the number of i/os active to each device. If the |
129 | * number of active i/os is < zfs_vdev_max_active, then the min_active comes | |
130 | * into play. We will send min_active from each queue, and then select from | |
131 | * queues in the order defined by zio_priority_t. | |
132 | * | |
133 | * In general, smaller max_active's will lead to lower latency of synchronous | |
134 | * operations. Larger max_active's may lead to higher overall throughput, | |
135 | * depending on underlying storage. | |
136 | * | |
137 | * The ratio of the queues' max_actives determines the balance of performance | |
138 | * between reads, writes, and scrubs. E.g., increasing | |
139 | * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete | |
140 | * more quickly, but reads and writes to have higher latency and lower | |
141 | * throughput. | |
cb682a17 | 142 | */ |
e8b96c60 MA |
143 | uint32_t zfs_vdev_sync_read_min_active = 10; |
144 | uint32_t zfs_vdev_sync_read_max_active = 10; | |
145 | uint32_t zfs_vdev_sync_write_min_active = 10; | |
146 | uint32_t zfs_vdev_sync_write_max_active = 10; | |
147 | uint32_t zfs_vdev_async_read_min_active = 1; | |
148 | uint32_t zfs_vdev_async_read_max_active = 3; | |
149 | uint32_t zfs_vdev_async_write_min_active = 1; | |
150 | uint32_t zfs_vdev_async_write_max_active = 10; | |
151 | uint32_t zfs_vdev_scrub_min_active = 1; | |
152 | uint32_t zfs_vdev_scrub_max_active = 2; | |
34dc7c2f | 153 | |
e8b96c60 MA |
154 | /* |
155 | * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent | |
156 | * dirty data, use zfs_vdev_async_write_min_active. When it has more than | |
157 | * zfs_vdev_async_write_active_max_dirty_percent, use | |
158 | * zfs_vdev_async_write_max_active. The value is linearly interpolated | |
159 | * between min and max. | |
160 | */ | |
161 | int zfs_vdev_async_write_active_min_dirty_percent = 30; | |
162 | int zfs_vdev_async_write_active_max_dirty_percent = 60; | |
34dc7c2f BB |
163 | |
164 | /* | |
45d1cae3 BB |
165 | * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O. |
166 | * For read I/Os, we also aggregate across small adjacency gaps; for writes | |
167 | * we include spans of optional I/Os to aid aggregation at the disk even when | |
168 | * they aren't able to help us aggregate at this level. | |
34dc7c2f | 169 | */ |
f1512ee6 | 170 | int zfs_vdev_aggregation_limit = SPA_OLD_MAXBLOCKSIZE; |
9babb374 | 171 | int zfs_vdev_read_gap_limit = 32 << 10; |
45d1cae3 | 172 | int zfs_vdev_write_gap_limit = 4 << 10; |
34dc7c2f | 173 | |
34dc7c2f | 174 | int |
e8b96c60 | 175 | vdev_queue_offset_compare(const void *x1, const void *x2) |
34dc7c2f | 176 | { |
ee36c709 GN |
177 | const zio_t *z1 = (const zio_t *)x1; |
178 | const zio_t *z2 = (const zio_t *)x2; | |
34dc7c2f | 179 | |
ee36c709 | 180 | int cmp = AVL_CMP(z1->io_offset, z2->io_offset); |
34dc7c2f | 181 | |
ee36c709 GN |
182 | if (likely(cmp)) |
183 | return (cmp); | |
34dc7c2f | 184 | |
ee36c709 | 185 | return (AVL_PCMP(z1, z2)); |
34dc7c2f BB |
186 | } |
187 | ||
ec8501ee JG |
188 | static inline avl_tree_t * |
189 | vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p) | |
190 | { | |
191 | return (&vq->vq_class[p].vqc_queued_tree); | |
192 | } | |
193 | ||
194 | static inline avl_tree_t * | |
195 | vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t) | |
196 | { | |
197 | ASSERT(t == ZIO_TYPE_READ || t == ZIO_TYPE_WRITE); | |
198 | if (t == ZIO_TYPE_READ) | |
199 | return (&vq->vq_read_offset_tree); | |
200 | else | |
201 | return (&vq->vq_write_offset_tree); | |
202 | } | |
203 | ||
34dc7c2f | 204 | int |
e8b96c60 | 205 | vdev_queue_timestamp_compare(const void *x1, const void *x2) |
34dc7c2f | 206 | { |
ee36c709 GN |
207 | const zio_t *z1 = (const zio_t *)x1; |
208 | const zio_t *z2 = (const zio_t *)x2; | |
34dc7c2f | 209 | |
ee36c709 | 210 | int cmp = AVL_CMP(z1->io_timestamp, z2->io_timestamp); |
34dc7c2f | 211 | |
ee36c709 GN |
212 | if (likely(cmp)) |
213 | return (cmp); | |
34dc7c2f | 214 | |
ee36c709 | 215 | return (AVL_PCMP(z1, z2)); |
34dc7c2f BB |
216 | } |
217 | ||
e8b96c60 MA |
218 | static int |
219 | vdev_queue_class_min_active(zio_priority_t p) | |
220 | { | |
221 | switch (p) { | |
222 | case ZIO_PRIORITY_SYNC_READ: | |
223 | return (zfs_vdev_sync_read_min_active); | |
224 | case ZIO_PRIORITY_SYNC_WRITE: | |
225 | return (zfs_vdev_sync_write_min_active); | |
226 | case ZIO_PRIORITY_ASYNC_READ: | |
227 | return (zfs_vdev_async_read_min_active); | |
228 | case ZIO_PRIORITY_ASYNC_WRITE: | |
229 | return (zfs_vdev_async_write_min_active); | |
230 | case ZIO_PRIORITY_SCRUB: | |
231 | return (zfs_vdev_scrub_min_active); | |
232 | default: | |
233 | panic("invalid priority %u", p); | |
234 | return (0); | |
235 | } | |
236 | } | |
237 | ||
238 | static int | |
acbad6ff | 239 | vdev_queue_max_async_writes(spa_t *spa) |
e8b96c60 MA |
240 | { |
241 | int writes; | |
b7faa7aa G |
242 | uint64_t dirty = 0; |
243 | dsl_pool_t *dp = spa_get_dsl(spa); | |
e8b96c60 MA |
244 | uint64_t min_bytes = zfs_dirty_data_max * |
245 | zfs_vdev_async_write_active_min_dirty_percent / 100; | |
246 | uint64_t max_bytes = zfs_dirty_data_max * | |
247 | zfs_vdev_async_write_active_max_dirty_percent / 100; | |
248 | ||
b7faa7aa G |
249 | /* |
250 | * Async writes may occur before the assignment of the spa's | |
251 | * dsl_pool_t if a self-healing zio is issued prior to the | |
252 | * completion of dmu_objset_open_impl(). | |
253 | */ | |
254 | if (dp == NULL) | |
255 | return (zfs_vdev_async_write_max_active); | |
256 | ||
acbad6ff AR |
257 | /* |
258 | * Sync tasks correspond to interactive user actions. To reduce the | |
259 | * execution time of those actions we push data out as fast as possible. | |
260 | */ | |
b7faa7aa | 261 | if (spa_has_pending_synctask(spa)) |
acbad6ff | 262 | return (zfs_vdev_async_write_max_active); |
acbad6ff | 263 | |
b7faa7aa | 264 | dirty = dp->dp_dirty_total; |
e8b96c60 MA |
265 | if (dirty < min_bytes) |
266 | return (zfs_vdev_async_write_min_active); | |
267 | if (dirty > max_bytes) | |
268 | return (zfs_vdev_async_write_max_active); | |
269 | ||
270 | /* | |
271 | * linear interpolation: | |
272 | * slope = (max_writes - min_writes) / (max_bytes - min_bytes) | |
273 | * move right by min_bytes | |
274 | * move up by min_writes | |
275 | */ | |
276 | writes = (dirty - min_bytes) * | |
277 | (zfs_vdev_async_write_max_active - | |
278 | zfs_vdev_async_write_min_active) / | |
279 | (max_bytes - min_bytes) + | |
280 | zfs_vdev_async_write_min_active; | |
281 | ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); | |
282 | ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); | |
283 | return (writes); | |
284 | } | |
285 | ||
286 | static int | |
287 | vdev_queue_class_max_active(spa_t *spa, zio_priority_t p) | |
288 | { | |
289 | switch (p) { | |
290 | case ZIO_PRIORITY_SYNC_READ: | |
291 | return (zfs_vdev_sync_read_max_active); | |
292 | case ZIO_PRIORITY_SYNC_WRITE: | |
293 | return (zfs_vdev_sync_write_max_active); | |
294 | case ZIO_PRIORITY_ASYNC_READ: | |
295 | return (zfs_vdev_async_read_max_active); | |
296 | case ZIO_PRIORITY_ASYNC_WRITE: | |
acbad6ff | 297 | return (vdev_queue_max_async_writes(spa)); |
e8b96c60 MA |
298 | case ZIO_PRIORITY_SCRUB: |
299 | return (zfs_vdev_scrub_max_active); | |
300 | default: | |
301 | panic("invalid priority %u", p); | |
302 | return (0); | |
303 | } | |
304 | } | |
305 | ||
306 | /* | |
307 | * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if | |
308 | * there is no eligible class. | |
309 | */ | |
310 | static zio_priority_t | |
311 | vdev_queue_class_to_issue(vdev_queue_t *vq) | |
312 | { | |
313 | spa_t *spa = vq->vq_vdev->vdev_spa; | |
314 | zio_priority_t p; | |
315 | ||
316 | if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active) | |
317 | return (ZIO_PRIORITY_NUM_QUEUEABLE); | |
318 | ||
319 | /* find a queue that has not reached its minimum # outstanding i/os */ | |
320 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { | |
ec8501ee | 321 | if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && |
e8b96c60 MA |
322 | vq->vq_class[p].vqc_active < |
323 | vdev_queue_class_min_active(p)) | |
324 | return (p); | |
325 | } | |
326 | ||
327 | /* | |
328 | * If we haven't found a queue, look for one that hasn't reached its | |
329 | * maximum # outstanding i/os. | |
330 | */ | |
331 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { | |
ec8501ee | 332 | if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && |
e8b96c60 MA |
333 | vq->vq_class[p].vqc_active < |
334 | vdev_queue_class_max_active(spa, p)) | |
335 | return (p); | |
336 | } | |
337 | ||
338 | /* No eligible queued i/os */ | |
339 | return (ZIO_PRIORITY_NUM_QUEUEABLE); | |
340 | } | |
341 | ||
34dc7c2f BB |
342 | void |
343 | vdev_queue_init(vdev_t *vd) | |
344 | { | |
345 | vdev_queue_t *vq = &vd->vdev_queue; | |
e8b96c60 | 346 | zio_priority_t p; |
34dc7c2f BB |
347 | |
348 | mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); | |
e8b96c60 | 349 | vq->vq_vdev = vd; |
36b454ab | 350 | taskq_init_ent(&vd->vdev_queue.vq_io_search.io_tqent); |
34dc7c2f | 351 | |
e8b96c60 MA |
352 | avl_create(&vq->vq_active_tree, vdev_queue_offset_compare, |
353 | sizeof (zio_t), offsetof(struct zio, io_queue_node)); | |
ec8501ee JG |
354 | avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ), |
355 | vdev_queue_offset_compare, sizeof (zio_t), | |
356 | offsetof(struct zio, io_offset_node)); | |
357 | avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE), | |
358 | vdev_queue_offset_compare, sizeof (zio_t), | |
359 | offsetof(struct zio, io_offset_node)); | |
34dc7c2f | 360 | |
e8b96c60 | 361 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { |
ec8501ee JG |
362 | int (*compfn) (const void *, const void *); |
363 | ||
e8b96c60 | 364 | /* |
ec8501ee JG |
365 | * The synchronous i/o queues are dispatched in FIFO rather |
366 | * than LBA order. This provides more consistent latency for | |
367 | * these i/os. | |
e8b96c60 | 368 | */ |
ec8501ee JG |
369 | if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE) |
370 | compfn = vdev_queue_timestamp_compare; | |
371 | else | |
372 | compfn = vdev_queue_offset_compare; | |
373 | avl_create(vdev_queue_class_tree(vq, p), compfn, | |
374 | sizeof (zio_t), offsetof(struct zio, io_queue_node)); | |
e8b96c60 | 375 | } |
9f500936 | 376 | |
377 | vq->vq_lastoffset = 0; | |
34dc7c2f BB |
378 | } |
379 | ||
380 | void | |
381 | vdev_queue_fini(vdev_t *vd) | |
382 | { | |
383 | vdev_queue_t *vq = &vd->vdev_queue; | |
e8b96c60 | 384 | zio_priority_t p; |
34dc7c2f | 385 | |
e8b96c60 | 386 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) |
ec8501ee | 387 | avl_destroy(vdev_queue_class_tree(vq, p)); |
e8b96c60 | 388 | avl_destroy(&vq->vq_active_tree); |
ec8501ee JG |
389 | avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ)); |
390 | avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE)); | |
34dc7c2f BB |
391 | |
392 | mutex_destroy(&vq->vq_lock); | |
393 | } | |
394 | ||
395 | static void | |
396 | vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) | |
397 | { | |
330847ff MA |
398 | spa_t *spa = zio->io_spa; |
399 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
400 | ||
e8b96c60 | 401 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); |
ec8501ee JG |
402 | avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio); |
403 | avl_add(vdev_queue_type_tree(vq, zio->io_type), zio); | |
330847ff MA |
404 | |
405 | if (ssh->kstat != NULL) { | |
406 | mutex_enter(&ssh->lock); | |
407 | kstat_waitq_enter(ssh->kstat->ks_data); | |
408 | mutex_exit(&ssh->lock); | |
409 | } | |
34dc7c2f BB |
410 | } |
411 | ||
412 | static void | |
413 | vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) | |
414 | { | |
330847ff MA |
415 | spa_t *spa = zio->io_spa; |
416 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
417 | ||
e8b96c60 | 418 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); |
ec8501ee JG |
419 | avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio); |
420 | avl_remove(vdev_queue_type_tree(vq, zio->io_type), zio); | |
330847ff MA |
421 | |
422 | if (ssh->kstat != NULL) { | |
423 | mutex_enter(&ssh->lock); | |
424 | kstat_waitq_exit(ssh->kstat->ks_data); | |
425 | mutex_exit(&ssh->lock); | |
426 | } | |
427 | } | |
428 | ||
429 | static void | |
430 | vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) | |
431 | { | |
432 | spa_t *spa = zio->io_spa; | |
433 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
434 | ||
e8b96c60 MA |
435 | ASSERT(MUTEX_HELD(&vq->vq_lock)); |
436 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
437 | vq->vq_class[zio->io_priority].vqc_active++; | |
438 | avl_add(&vq->vq_active_tree, zio); | |
330847ff MA |
439 | |
440 | if (ssh->kstat != NULL) { | |
441 | mutex_enter(&ssh->lock); | |
442 | kstat_runq_enter(ssh->kstat->ks_data); | |
443 | mutex_exit(&ssh->lock); | |
444 | } | |
445 | } | |
446 | ||
447 | static void | |
448 | vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) | |
449 | { | |
450 | spa_t *spa = zio->io_spa; | |
451 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
452 | ||
e8b96c60 MA |
453 | ASSERT(MUTEX_HELD(&vq->vq_lock)); |
454 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
455 | vq->vq_class[zio->io_priority].vqc_active--; | |
456 | avl_remove(&vq->vq_active_tree, zio); | |
330847ff MA |
457 | |
458 | if (ssh->kstat != NULL) { | |
459 | kstat_io_t *ksio = ssh->kstat->ks_data; | |
460 | ||
461 | mutex_enter(&ssh->lock); | |
462 | kstat_runq_exit(ksio); | |
463 | if (zio->io_type == ZIO_TYPE_READ) { | |
464 | ksio->reads++; | |
465 | ksio->nread += zio->io_size; | |
466 | } else if (zio->io_type == ZIO_TYPE_WRITE) { | |
467 | ksio->writes++; | |
468 | ksio->nwritten += zio->io_size; | |
469 | } | |
470 | mutex_exit(&ssh->lock); | |
471 | } | |
34dc7c2f BB |
472 | } |
473 | ||
474 | static void | |
475 | vdev_queue_agg_io_done(zio_t *aio) | |
476 | { | |
e8b96c60 MA |
477 | if (aio->io_type == ZIO_TYPE_READ) { |
478 | zio_t *pio; | |
479 | while ((pio = zio_walk_parents(aio)) != NULL) { | |
d164b209 BB |
480 | bcopy((char *)aio->io_data + (pio->io_offset - |
481 | aio->io_offset), pio->io_data, pio->io_size); | |
e8b96c60 MA |
482 | } |
483 | } | |
34dc7c2f | 484 | |
285b29d9 | 485 | zio_buf_free(aio->io_data, aio->io_size); |
34dc7c2f BB |
486 | } |
487 | ||
9babb374 BB |
488 | /* |
489 | * Compute the range spanned by two i/os, which is the endpoint of the last | |
490 | * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). | |
491 | * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); | |
492 | * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. | |
493 | */ | |
494 | #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) | |
495 | #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) | |
34dc7c2f BB |
496 | |
497 | static zio_t * | |
e8b96c60 | 498 | vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) |
34dc7c2f | 499 | { |
e8b96c60 MA |
500 | zio_t *first, *last, *aio, *dio, *mandatory, *nio; |
501 | uint64_t maxgap = 0; | |
502 | uint64_t size; | |
a58df6f5 | 503 | uint64_t limit; |
e8b96c60 | 504 | boolean_t stretch = B_FALSE; |
ec8501ee | 505 | avl_tree_t *t = vdev_queue_type_tree(vq, zio->io_type); |
e8b96c60 | 506 | enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; |
6fe53787 | 507 | void *buf; |
e8b96c60 | 508 | |
a58df6f5 BB |
509 | limit = MAX(MIN(zfs_vdev_aggregation_limit, |
510 | spa_maxblocksize(vq->vq_vdev->vdev_spa)), 0); | |
34dc7c2f | 511 | |
a58df6f5 BB |
512 | if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE || limit == 0) |
513 | return (NULL); | |
34dc7c2f | 514 | |
e8b96c60 | 515 | first = last = zio; |
34dc7c2f | 516 | |
e8b96c60 MA |
517 | if (zio->io_type == ZIO_TYPE_READ) |
518 | maxgap = zfs_vdev_read_gap_limit; | |
fb5f0bc8 | 519 | |
e8b96c60 MA |
520 | /* |
521 | * We can aggregate I/Os that are sufficiently adjacent and of | |
522 | * the same flavor, as expressed by the AGG_INHERIT flags. | |
523 | * The latter requirement is necessary so that certain | |
524 | * attributes of the I/O, such as whether it's a normal I/O | |
525 | * or a scrub/resilver, can be preserved in the aggregate. | |
526 | * We can include optional I/Os, but don't allow them | |
527 | * to begin a range as they add no benefit in that situation. | |
528 | */ | |
45d1cae3 | 529 | |
e8b96c60 MA |
530 | /* |
531 | * We keep track of the last non-optional I/O. | |
532 | */ | |
533 | mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; | |
45d1cae3 | 534 | |
e8b96c60 MA |
535 | /* |
536 | * Walk backwards through sufficiently contiguous I/Os | |
537 | * recording the last non-option I/O. | |
538 | */ | |
539 | while ((dio = AVL_PREV(t, first)) != NULL && | |
540 | (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && | |
a58df6f5 | 541 | IO_SPAN(dio, last) <= limit && |
e8b96c60 MA |
542 | IO_GAP(dio, first) <= maxgap) { |
543 | first = dio; | |
544 | if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) | |
545 | mandatory = first; | |
546 | } | |
45d1cae3 | 547 | |
e8b96c60 MA |
548 | /* |
549 | * Skip any initial optional I/Os. | |
550 | */ | |
551 | while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { | |
552 | first = AVL_NEXT(t, first); | |
553 | ASSERT(first != NULL); | |
554 | } | |
9babb374 | 555 | |
45d1cae3 | 556 | |
e8b96c60 MA |
557 | /* |
558 | * Walk forward through sufficiently contiguous I/Os. | |
559 | */ | |
560 | while ((dio = AVL_NEXT(t, last)) != NULL && | |
561 | (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && | |
a58df6f5 | 562 | IO_SPAN(first, dio) <= limit && |
e8b96c60 MA |
563 | IO_GAP(last, dio) <= maxgap) { |
564 | last = dio; | |
565 | if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) | |
566 | mandatory = last; | |
567 | } | |
568 | ||
569 | /* | |
570 | * Now that we've established the range of the I/O aggregation | |
571 | * we must decide what to do with trailing optional I/Os. | |
572 | * For reads, there's nothing to do. While we are unable to | |
573 | * aggregate further, it's possible that a trailing optional | |
574 | * I/O would allow the underlying device to aggregate with | |
575 | * subsequent I/Os. We must therefore determine if the next | |
576 | * non-optional I/O is close enough to make aggregation | |
577 | * worthwhile. | |
578 | */ | |
579 | if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { | |
580 | zio_t *nio = last; | |
581 | while ((dio = AVL_NEXT(t, nio)) != NULL && | |
582 | IO_GAP(nio, dio) == 0 && | |
583 | IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { | |
584 | nio = dio; | |
585 | if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { | |
586 | stretch = B_TRUE; | |
587 | break; | |
45d1cae3 BB |
588 | } |
589 | } | |
e8b96c60 | 590 | } |
45d1cae3 | 591 | |
e8b96c60 MA |
592 | if (stretch) { |
593 | /* This may be a no-op. */ | |
594 | dio = AVL_NEXT(t, last); | |
595 | dio->io_flags &= ~ZIO_FLAG_OPTIONAL; | |
596 | } else { | |
597 | while (last != mandatory && last != first) { | |
598 | ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); | |
599 | last = AVL_PREV(t, last); | |
600 | ASSERT(last != NULL); | |
45d1cae3 | 601 | } |
34dc7c2f BB |
602 | } |
603 | ||
e8b96c60 MA |
604 | if (first == last) |
605 | return (NULL); | |
606 | ||
e8b96c60 | 607 | size = IO_SPAN(first, last); |
a58df6f5 | 608 | ASSERT3U(size, <=, limit); |
e8b96c60 | 609 | |
6fe53787 BB |
610 | buf = zio_buf_alloc_flags(size, KM_NOSLEEP); |
611 | if (buf == NULL) | |
612 | return (NULL); | |
613 | ||
e8b96c60 | 614 | aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, |
6fe53787 | 615 | buf, size, first->io_type, zio->io_priority, |
e8b96c60 MA |
616 | flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE, |
617 | vdev_queue_agg_io_done, NULL); | |
618 | aio->io_timestamp = first->io_timestamp; | |
619 | ||
620 | nio = first; | |
621 | do { | |
622 | dio = nio; | |
623 | nio = AVL_NEXT(t, dio); | |
624 | ASSERT3U(dio->io_type, ==, aio->io_type); | |
625 | ||
626 | if (dio->io_flags & ZIO_FLAG_NODATA) { | |
627 | ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); | |
628 | bzero((char *)aio->io_data + (dio->io_offset - | |
629 | aio->io_offset), dio->io_size); | |
630 | } else if (dio->io_type == ZIO_TYPE_WRITE) { | |
631 | bcopy(dio->io_data, (char *)aio->io_data + | |
632 | (dio->io_offset - aio->io_offset), | |
633 | dio->io_size); | |
634 | } | |
d164b209 | 635 | |
e8b96c60 MA |
636 | zio_add_child(dio, aio); |
637 | vdev_queue_io_remove(vq, dio); | |
638 | zio_vdev_io_bypass(dio); | |
639 | zio_execute(dio); | |
640 | } while (dio != last); | |
34dc7c2f | 641 | |
e8b96c60 MA |
642 | return (aio); |
643 | } | |
644 | ||
645 | static zio_t * | |
646 | vdev_queue_io_to_issue(vdev_queue_t *vq) | |
647 | { | |
648 | zio_t *zio, *aio; | |
649 | zio_priority_t p; | |
650 | avl_index_t idx; | |
ec8501ee | 651 | avl_tree_t *tree; |
e8b96c60 MA |
652 | |
653 | again: | |
654 | ASSERT(MUTEX_HELD(&vq->vq_lock)); | |
655 | ||
656 | p = vdev_queue_class_to_issue(vq); | |
34dc7c2f | 657 | |
e8b96c60 MA |
658 | if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { |
659 | /* No eligible queued i/os */ | |
660 | return (NULL); | |
34dc7c2f BB |
661 | } |
662 | ||
e8b96c60 MA |
663 | /* |
664 | * For LBA-ordered queues (async / scrub), issue the i/o which follows | |
665 | * the most recently issued i/o in LBA (offset) order. | |
666 | * | |
667 | * For FIFO queues (sync), issue the i/o with the lowest timestamp. | |
668 | */ | |
ec8501ee | 669 | tree = vdev_queue_class_tree(vq, p); |
50b25b21 BB |
670 | vq->vq_io_search.io_timestamp = 0; |
671 | vq->vq_io_search.io_offset = vq->vq_last_offset + 1; | |
ec8501ee | 672 | VERIFY3P(avl_find(tree, &vq->vq_io_search, |
50b25b21 | 673 | &idx), ==, NULL); |
ec8501ee | 674 | zio = avl_nearest(tree, idx, AVL_AFTER); |
e8b96c60 | 675 | if (zio == NULL) |
ec8501ee | 676 | zio = avl_first(tree); |
e8b96c60 MA |
677 | ASSERT3U(zio->io_priority, ==, p); |
678 | ||
679 | aio = vdev_queue_aggregate(vq, zio); | |
680 | if (aio != NULL) | |
681 | zio = aio; | |
682 | else | |
683 | vdev_queue_io_remove(vq, zio); | |
34dc7c2f | 684 | |
45d1cae3 BB |
685 | /* |
686 | * If the I/O is or was optional and therefore has no data, we need to | |
687 | * simply discard it. We need to drop the vdev queue's lock to avoid a | |
688 | * deadlock that we could encounter since this I/O will complete | |
689 | * immediately. | |
690 | */ | |
e8b96c60 | 691 | if (zio->io_flags & ZIO_FLAG_NODATA) { |
45d1cae3 | 692 | mutex_exit(&vq->vq_lock); |
e8b96c60 MA |
693 | zio_vdev_io_bypass(zio); |
694 | zio_execute(zio); | |
45d1cae3 BB |
695 | mutex_enter(&vq->vq_lock); |
696 | goto again; | |
697 | } | |
698 | ||
e8b96c60 MA |
699 | vdev_queue_pending_add(vq, zio); |
700 | vq->vq_last_offset = zio->io_offset; | |
34dc7c2f | 701 | |
e8b96c60 | 702 | return (zio); |
34dc7c2f BB |
703 | } |
704 | ||
705 | zio_t * | |
706 | vdev_queue_io(zio_t *zio) | |
707 | { | |
708 | vdev_queue_t *vq = &zio->io_vd->vdev_queue; | |
709 | zio_t *nio; | |
710 | ||
34dc7c2f BB |
711 | if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) |
712 | return (zio); | |
713 | ||
e8b96c60 MA |
714 | /* |
715 | * Children i/os inherent their parent's priority, which might | |
716 | * not match the child's i/o type. Fix it up here. | |
717 | */ | |
718 | if (zio->io_type == ZIO_TYPE_READ) { | |
719 | if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && | |
720 | zio->io_priority != ZIO_PRIORITY_ASYNC_READ && | |
721 | zio->io_priority != ZIO_PRIORITY_SCRUB) | |
722 | zio->io_priority = ZIO_PRIORITY_ASYNC_READ; | |
723 | } else { | |
724 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
725 | if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && | |
726 | zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE) | |
727 | zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; | |
728 | } | |
34dc7c2f | 729 | |
e8b96c60 | 730 | zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; |
34dc7c2f BB |
731 | |
732 | mutex_enter(&vq->vq_lock); | |
cb682a17 | 733 | zio->io_timestamp = gethrtime(); |
34dc7c2f | 734 | vdev_queue_io_add(vq, zio); |
e8b96c60 | 735 | nio = vdev_queue_io_to_issue(vq); |
34dc7c2f BB |
736 | mutex_exit(&vq->vq_lock); |
737 | ||
738 | if (nio == NULL) | |
739 | return (NULL); | |
740 | ||
741 | if (nio->io_done == vdev_queue_agg_io_done) { | |
742 | zio_nowait(nio); | |
743 | return (NULL); | |
744 | } | |
745 | ||
746 | return (nio); | |
747 | } | |
748 | ||
749 | void | |
750 | vdev_queue_io_done(zio_t *zio) | |
751 | { | |
752 | vdev_queue_t *vq = &zio->io_vd->vdev_queue; | |
e8b96c60 | 753 | zio_t *nio; |
34dc7c2f BB |
754 | |
755 | mutex_enter(&vq->vq_lock); | |
756 | ||
330847ff | 757 | vdev_queue_pending_remove(vq, zio); |
34dc7c2f | 758 | |
cb682a17 MA |
759 | zio->io_delta = gethrtime() - zio->io_timestamp; |
760 | vq->vq_io_complete_ts = gethrtime(); | |
cc92e9d0 GW |
761 | vq->vq_io_delta_ts = vq->vq_io_complete_ts - zio->io_timestamp; |
762 | ||
e8b96c60 | 763 | while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { |
34dc7c2f BB |
764 | mutex_exit(&vq->vq_lock); |
765 | if (nio->io_done == vdev_queue_agg_io_done) { | |
766 | zio_nowait(nio); | |
767 | } else { | |
768 | zio_vdev_io_reissue(nio); | |
769 | zio_execute(nio); | |
770 | } | |
771 | mutex_enter(&vq->vq_lock); | |
772 | } | |
773 | ||
774 | mutex_exit(&vq->vq_lock); | |
775 | } | |
c28b2279 | 776 | |
9f500936 | 777 | /* |
778 | * As these three methods are only used for load calculations we're not | |
779 | * concerned if we get an incorrect value on 32bit platforms due to lack of | |
780 | * vq_lock mutex use here, instead we prefer to keep it lock free for | |
781 | * performance. | |
782 | */ | |
783 | int | |
784 | vdev_queue_length(vdev_t *vd) | |
785 | { | |
786 | return (avl_numnodes(&vd->vdev_queue.vq_active_tree)); | |
787 | } | |
788 | ||
789 | uint64_t | |
790 | vdev_queue_lastoffset(vdev_t *vd) | |
791 | { | |
792 | return (vd->vdev_queue.vq_lastoffset); | |
793 | } | |
794 | ||
795 | void | |
796 | vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio) | |
797 | { | |
798 | vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size; | |
799 | } | |
800 | ||
c28b2279 | 801 | #if defined(_KERNEL) && defined(HAVE_SPL) |
c28b2279 | 802 | module_param(zfs_vdev_aggregation_limit, int, 0644); |
c409e464 BB |
803 | MODULE_PARM_DESC(zfs_vdev_aggregation_limit, "Max vdev I/O aggregation size"); |
804 | ||
c409e464 BB |
805 | module_param(zfs_vdev_read_gap_limit, int, 0644); |
806 | MODULE_PARM_DESC(zfs_vdev_read_gap_limit, "Aggregate read I/O over gap"); | |
807 | ||
808 | module_param(zfs_vdev_write_gap_limit, int, 0644); | |
809 | MODULE_PARM_DESC(zfs_vdev_write_gap_limit, "Aggregate write I/O over gap"); | |
e8b96c60 MA |
810 | |
811 | module_param(zfs_vdev_max_active, int, 0644); | |
812 | MODULE_PARM_DESC(zfs_vdev_max_active, "Maximum number of active I/Os per vdev"); | |
813 | ||
814 | module_param(zfs_vdev_async_write_active_max_dirty_percent, int, 0644); | |
815 | MODULE_PARM_DESC(zfs_vdev_async_write_active_max_dirty_percent, | |
d1d7e268 | 816 | "Async write concurrency max threshold"); |
e8b96c60 MA |
817 | |
818 | module_param(zfs_vdev_async_write_active_min_dirty_percent, int, 0644); | |
819 | MODULE_PARM_DESC(zfs_vdev_async_write_active_min_dirty_percent, | |
d1d7e268 | 820 | "Async write concurrency min threshold"); |
e8b96c60 MA |
821 | |
822 | module_param(zfs_vdev_async_read_max_active, int, 0644); | |
823 | MODULE_PARM_DESC(zfs_vdev_async_read_max_active, | |
d1d7e268 | 824 | "Max active async read I/Os per vdev"); |
e8b96c60 MA |
825 | |
826 | module_param(zfs_vdev_async_read_min_active, int, 0644); | |
827 | MODULE_PARM_DESC(zfs_vdev_async_read_min_active, | |
d1d7e268 | 828 | "Min active async read I/Os per vdev"); |
e8b96c60 MA |
829 | |
830 | module_param(zfs_vdev_async_write_max_active, int, 0644); | |
831 | MODULE_PARM_DESC(zfs_vdev_async_write_max_active, | |
d1d7e268 | 832 | "Max active async write I/Os per vdev"); |
e8b96c60 MA |
833 | |
834 | module_param(zfs_vdev_async_write_min_active, int, 0644); | |
835 | MODULE_PARM_DESC(zfs_vdev_async_write_min_active, | |
d1d7e268 | 836 | "Min active async write I/Os per vdev"); |
e8b96c60 MA |
837 | |
838 | module_param(zfs_vdev_scrub_max_active, int, 0644); | |
839 | MODULE_PARM_DESC(zfs_vdev_scrub_max_active, "Max active scrub I/Os per vdev"); | |
840 | ||
841 | module_param(zfs_vdev_scrub_min_active, int, 0644); | |
842 | MODULE_PARM_DESC(zfs_vdev_scrub_min_active, "Min active scrub I/Os per vdev"); | |
843 | ||
844 | module_param(zfs_vdev_sync_read_max_active, int, 0644); | |
845 | MODULE_PARM_DESC(zfs_vdev_sync_read_max_active, | |
d1d7e268 | 846 | "Max active sync read I/Os per vdev"); |
e8b96c60 MA |
847 | |
848 | module_param(zfs_vdev_sync_read_min_active, int, 0644); | |
849 | MODULE_PARM_DESC(zfs_vdev_sync_read_min_active, | |
d1d7e268 | 850 | "Min active sync read I/Os per vdev"); |
e8b96c60 MA |
851 | |
852 | module_param(zfs_vdev_sync_write_max_active, int, 0644); | |
853 | MODULE_PARM_DESC(zfs_vdev_sync_write_max_active, | |
d1d7e268 | 854 | "Max active sync write I/Os per vdev"); |
e8b96c60 MA |
855 | |
856 | module_param(zfs_vdev_sync_write_min_active, int, 0644); | |
857 | MODULE_PARM_DESC(zfs_vdev_sync_write_min_active, | |
3757bff3 | 858 | "Min active sync write I/Os per vdev"); |
c28b2279 | 859 | #endif |