]>
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 | ||
c06d4368 | 26 | /* |
ea04106b | 27 | * Copyright (c) 2012, 2014 by Delphix. All rights reserved. |
c06d4368 AX |
28 | */ |
29 | ||
34dc7c2f | 30 | #include <sys/zfs_context.h> |
34dc7c2f | 31 | #include <sys/vdev_impl.h> |
a08ee875 | 32 | #include <sys/spa_impl.h> |
34dc7c2f BB |
33 | #include <sys/zio.h> |
34 | #include <sys/avl.h> | |
a08ee875 LG |
35 | #include <sys/dsl_pool.h> |
36 | #include <sys/spa.h> | |
37 | #include <sys/spa_impl.h> | |
38 | #include <sys/kstat.h> | |
34dc7c2f BB |
39 | |
40 | /* | |
a08ee875 LG |
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 | */ |
a08ee875 | 119 | |
34dc7c2f | 120 | /* |
a08ee875 LG |
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 | */ |
a08ee875 | 125 | uint32_t zfs_vdev_max_active = 1000; |
34dc7c2f | 126 | |
c06d4368 | 127 | /* |
a08ee875 LG |
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. | |
c06d4368 | 142 | */ |
a08ee875 LG |
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 | |
a08ee875 LG |
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 BB |
169 | */ |
170 | int zfs_vdev_aggregation_limit = SPA_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 |
a08ee875 | 175 | vdev_queue_offset_compare(const void *x1, const void *x2) |
34dc7c2f BB |
176 | { |
177 | const zio_t *z1 = x1; | |
178 | const zio_t *z2 = x2; | |
179 | ||
34dc7c2f BB |
180 | if (z1->io_offset < z2->io_offset) |
181 | return (-1); | |
182 | if (z1->io_offset > z2->io_offset) | |
183 | return (1); | |
184 | ||
185 | if (z1 < z2) | |
186 | return (-1); | |
187 | if (z1 > z2) | |
188 | return (1); | |
189 | ||
190 | return (0); | |
191 | } | |
192 | ||
193 | int | |
a08ee875 | 194 | vdev_queue_timestamp_compare(const void *x1, const void *x2) |
34dc7c2f BB |
195 | { |
196 | const zio_t *z1 = x1; | |
197 | const zio_t *z2 = x2; | |
198 | ||
a08ee875 | 199 | if (z1->io_timestamp < z2->io_timestamp) |
34dc7c2f | 200 | return (-1); |
a08ee875 | 201 | if (z1->io_timestamp > z2->io_timestamp) |
34dc7c2f BB |
202 | return (1); |
203 | ||
204 | if (z1 < z2) | |
205 | return (-1); | |
206 | if (z1 > z2) | |
207 | return (1); | |
208 | ||
209 | return (0); | |
210 | } | |
211 | ||
a08ee875 LG |
212 | static int |
213 | vdev_queue_class_min_active(zio_priority_t p) | |
214 | { | |
215 | switch (p) { | |
216 | case ZIO_PRIORITY_SYNC_READ: | |
217 | return (zfs_vdev_sync_read_min_active); | |
218 | case ZIO_PRIORITY_SYNC_WRITE: | |
219 | return (zfs_vdev_sync_write_min_active); | |
220 | case ZIO_PRIORITY_ASYNC_READ: | |
221 | return (zfs_vdev_async_read_min_active); | |
222 | case ZIO_PRIORITY_ASYNC_WRITE: | |
223 | return (zfs_vdev_async_write_min_active); | |
224 | case ZIO_PRIORITY_SCRUB: | |
225 | return (zfs_vdev_scrub_min_active); | |
226 | default: | |
227 | panic("invalid priority %u", p); | |
228 | return (0); | |
229 | } | |
230 | } | |
231 | ||
232 | static int | |
ea04106b | 233 | vdev_queue_max_async_writes(spa_t *spa) |
a08ee875 LG |
234 | { |
235 | int writes; | |
ea04106b | 236 | uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total; |
a08ee875 LG |
237 | uint64_t min_bytes = zfs_dirty_data_max * |
238 | zfs_vdev_async_write_active_min_dirty_percent / 100; | |
239 | uint64_t max_bytes = zfs_dirty_data_max * | |
240 | zfs_vdev_async_write_active_max_dirty_percent / 100; | |
241 | ||
ea04106b AX |
242 | /* |
243 | * Sync tasks correspond to interactive user actions. To reduce the | |
244 | * execution time of those actions we push data out as fast as possible. | |
245 | */ | |
246 | if (spa_has_pending_synctask(spa)) { | |
247 | return (zfs_vdev_async_write_max_active); | |
248 | } | |
249 | ||
a08ee875 LG |
250 | if (dirty < min_bytes) |
251 | return (zfs_vdev_async_write_min_active); | |
252 | if (dirty > max_bytes) | |
253 | return (zfs_vdev_async_write_max_active); | |
254 | ||
255 | /* | |
256 | * linear interpolation: | |
257 | * slope = (max_writes - min_writes) / (max_bytes - min_bytes) | |
258 | * move right by min_bytes | |
259 | * move up by min_writes | |
260 | */ | |
261 | writes = (dirty - min_bytes) * | |
262 | (zfs_vdev_async_write_max_active - | |
263 | zfs_vdev_async_write_min_active) / | |
264 | (max_bytes - min_bytes) + | |
265 | zfs_vdev_async_write_min_active; | |
266 | ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); | |
267 | ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); | |
268 | return (writes); | |
269 | } | |
270 | ||
271 | static int | |
272 | vdev_queue_class_max_active(spa_t *spa, zio_priority_t p) | |
273 | { | |
274 | switch (p) { | |
275 | case ZIO_PRIORITY_SYNC_READ: | |
276 | return (zfs_vdev_sync_read_max_active); | |
277 | case ZIO_PRIORITY_SYNC_WRITE: | |
278 | return (zfs_vdev_sync_write_max_active); | |
279 | case ZIO_PRIORITY_ASYNC_READ: | |
280 | return (zfs_vdev_async_read_max_active); | |
281 | case ZIO_PRIORITY_ASYNC_WRITE: | |
ea04106b | 282 | return (vdev_queue_max_async_writes(spa)); |
a08ee875 LG |
283 | case ZIO_PRIORITY_SCRUB: |
284 | return (zfs_vdev_scrub_max_active); | |
285 | default: | |
286 | panic("invalid priority %u", p); | |
287 | return (0); | |
288 | } | |
289 | } | |
290 | ||
291 | /* | |
292 | * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if | |
293 | * there is no eligible class. | |
294 | */ | |
295 | static zio_priority_t | |
296 | vdev_queue_class_to_issue(vdev_queue_t *vq) | |
297 | { | |
298 | spa_t *spa = vq->vq_vdev->vdev_spa; | |
299 | zio_priority_t p; | |
300 | ||
301 | if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active) | |
302 | return (ZIO_PRIORITY_NUM_QUEUEABLE); | |
303 | ||
304 | /* find a queue that has not reached its minimum # outstanding i/os */ | |
305 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { | |
306 | if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 && | |
307 | vq->vq_class[p].vqc_active < | |
308 | vdev_queue_class_min_active(p)) | |
309 | return (p); | |
310 | } | |
311 | ||
312 | /* | |
313 | * If we haven't found a queue, look for one that hasn't reached its | |
314 | * maximum # outstanding i/os. | |
315 | */ | |
316 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { | |
317 | if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 && | |
318 | vq->vq_class[p].vqc_active < | |
319 | vdev_queue_class_max_active(spa, p)) | |
320 | return (p); | |
321 | } | |
322 | ||
323 | /* No eligible queued i/os */ | |
324 | return (ZIO_PRIORITY_NUM_QUEUEABLE); | |
325 | } | |
326 | ||
34dc7c2f BB |
327 | void |
328 | vdev_queue_init(vdev_t *vd) | |
329 | { | |
330 | vdev_queue_t *vq = &vd->vdev_queue; | |
a08ee875 | 331 | zio_priority_t p; |
34dc7c2f BB |
332 | |
333 | mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); | |
a08ee875 | 334 | vq->vq_vdev = vd; |
34dc7c2f | 335 | |
a08ee875 LG |
336 | avl_create(&vq->vq_active_tree, vdev_queue_offset_compare, |
337 | sizeof (zio_t), offsetof(struct zio, io_queue_node)); | |
34dc7c2f | 338 | |
a08ee875 LG |
339 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { |
340 | /* | |
341 | * The synchronous i/o queues are FIFO rather than LBA ordered. | |
342 | * This provides more consistent latency for these i/os, and | |
343 | * they tend to not be tightly clustered anyway so there is | |
344 | * little to no throughput loss. | |
345 | */ | |
346 | boolean_t fifo = (p == ZIO_PRIORITY_SYNC_READ || | |
347 | p == ZIO_PRIORITY_SYNC_WRITE); | |
348 | avl_create(&vq->vq_class[p].vqc_queued_tree, | |
349 | fifo ? vdev_queue_timestamp_compare : | |
350 | vdev_queue_offset_compare, | |
351 | sizeof (zio_t), offsetof(struct zio, io_queue_node)); | |
352 | } | |
34dc7c2f BB |
353 | } |
354 | ||
355 | void | |
356 | vdev_queue_fini(vdev_t *vd) | |
357 | { | |
358 | vdev_queue_t *vq = &vd->vdev_queue; | |
a08ee875 | 359 | zio_priority_t p; |
34dc7c2f | 360 | |
a08ee875 LG |
361 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) |
362 | avl_destroy(&vq->vq_class[p].vqc_queued_tree); | |
363 | avl_destroy(&vq->vq_active_tree); | |
34dc7c2f BB |
364 | |
365 | mutex_destroy(&vq->vq_lock); | |
366 | } | |
367 | ||
368 | static void | |
369 | vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) | |
370 | { | |
a08ee875 LG |
371 | spa_t *spa = zio->io_spa; |
372 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
373 | ||
374 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
375 | avl_add(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio); | |
376 | ||
377 | if (ssh->kstat != NULL) { | |
378 | mutex_enter(&ssh->lock); | |
379 | kstat_waitq_enter(ssh->kstat->ks_data); | |
380 | mutex_exit(&ssh->lock); | |
381 | } | |
34dc7c2f BB |
382 | } |
383 | ||
384 | static void | |
385 | vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) | |
386 | { | |
a08ee875 LG |
387 | spa_t *spa = zio->io_spa; |
388 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
389 | ||
390 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
391 | avl_remove(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio); | |
392 | ||
393 | if (ssh->kstat != NULL) { | |
394 | mutex_enter(&ssh->lock); | |
395 | kstat_waitq_exit(ssh->kstat->ks_data); | |
396 | mutex_exit(&ssh->lock); | |
397 | } | |
398 | } | |
399 | ||
400 | static void | |
401 | vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) | |
402 | { | |
403 | spa_t *spa = zio->io_spa; | |
404 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
405 | ||
406 | ASSERT(MUTEX_HELD(&vq->vq_lock)); | |
407 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
408 | vq->vq_class[zio->io_priority].vqc_active++; | |
409 | avl_add(&vq->vq_active_tree, zio); | |
410 | ||
411 | if (ssh->kstat != NULL) { | |
412 | mutex_enter(&ssh->lock); | |
413 | kstat_runq_enter(ssh->kstat->ks_data); | |
414 | mutex_exit(&ssh->lock); | |
415 | } | |
416 | } | |
417 | ||
418 | static void | |
419 | vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) | |
420 | { | |
421 | spa_t *spa = zio->io_spa; | |
422 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
423 | ||
424 | ASSERT(MUTEX_HELD(&vq->vq_lock)); | |
425 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
426 | vq->vq_class[zio->io_priority].vqc_active--; | |
427 | avl_remove(&vq->vq_active_tree, zio); | |
428 | ||
429 | if (ssh->kstat != NULL) { | |
430 | kstat_io_t *ksio = ssh->kstat->ks_data; | |
431 | ||
432 | mutex_enter(&ssh->lock); | |
433 | kstat_runq_exit(ksio); | |
434 | if (zio->io_type == ZIO_TYPE_READ) { | |
435 | ksio->reads++; | |
436 | ksio->nread += zio->io_size; | |
437 | } else if (zio->io_type == ZIO_TYPE_WRITE) { | |
438 | ksio->writes++; | |
439 | ksio->nwritten += zio->io_size; | |
440 | } | |
441 | mutex_exit(&ssh->lock); | |
442 | } | |
34dc7c2f BB |
443 | } |
444 | ||
445 | static void | |
446 | vdev_queue_agg_io_done(zio_t *aio) | |
447 | { | |
a08ee875 LG |
448 | if (aio->io_type == ZIO_TYPE_READ) { |
449 | zio_t *pio; | |
450 | while ((pio = zio_walk_parents(aio)) != NULL) { | |
d164b209 BB |
451 | bcopy((char *)aio->io_data + (pio->io_offset - |
452 | aio->io_offset), pio->io_data, pio->io_size); | |
a08ee875 LG |
453 | } |
454 | } | |
34dc7c2f | 455 | |
ea04106b | 456 | zio_buf_free(aio->io_data, aio->io_size); |
34dc7c2f BB |
457 | } |
458 | ||
9babb374 BB |
459 | /* |
460 | * Compute the range spanned by two i/os, which is the endpoint of the last | |
461 | * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). | |
462 | * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); | |
463 | * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. | |
464 | */ | |
465 | #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) | |
466 | #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) | |
34dc7c2f BB |
467 | |
468 | static zio_t * | |
a08ee875 | 469 | vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) |
34dc7c2f | 470 | { |
a08ee875 LG |
471 | zio_t *first, *last, *aio, *dio, *mandatory, *nio; |
472 | uint64_t maxgap = 0; | |
473 | uint64_t size; | |
474 | boolean_t stretch = B_FALSE; | |
475 | vdev_queue_class_t *vqc = &vq->vq_class[zio->io_priority]; | |
476 | avl_tree_t *t = &vqc->vqc_queued_tree; | |
477 | enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; | |
478 | ||
479 | if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE) | |
480 | return (NULL); | |
34dc7c2f | 481 | |
a08ee875 LG |
482 | /* |
483 | * Prevent users from setting the zfs_vdev_aggregation_limit | |
484 | * tuning larger than SPA_MAXBLOCKSIZE. | |
485 | */ | |
486 | zfs_vdev_aggregation_limit = | |
487 | MIN(zfs_vdev_aggregation_limit, SPA_MAXBLOCKSIZE); | |
34dc7c2f | 488 | |
a08ee875 LG |
489 | /* |
490 | * The synchronous i/o queues are not sorted by LBA, so we can't | |
491 | * find adjacent i/os. These i/os tend to not be tightly clustered, | |
492 | * or too large to aggregate, so this has little impact on performance. | |
493 | */ | |
494 | if (zio->io_priority == ZIO_PRIORITY_SYNC_READ || | |
495 | zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) | |
34dc7c2f BB |
496 | return (NULL); |
497 | ||
a08ee875 | 498 | first = last = zio; |
34dc7c2f | 499 | |
a08ee875 LG |
500 | if (zio->io_type == ZIO_TYPE_READ) |
501 | maxgap = zfs_vdev_read_gap_limit; | |
fb5f0bc8 | 502 | |
a08ee875 LG |
503 | /* |
504 | * We can aggregate I/Os that are sufficiently adjacent and of | |
505 | * the same flavor, as expressed by the AGG_INHERIT flags. | |
506 | * The latter requirement is necessary so that certain | |
507 | * attributes of the I/O, such as whether it's a normal I/O | |
508 | * or a scrub/resilver, can be preserved in the aggregate. | |
509 | * We can include optional I/Os, but don't allow them | |
510 | * to begin a range as they add no benefit in that situation. | |
511 | */ | |
45d1cae3 | 512 | |
a08ee875 LG |
513 | /* |
514 | * We keep track of the last non-optional I/O. | |
515 | */ | |
516 | mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; | |
45d1cae3 | 517 | |
a08ee875 LG |
518 | /* |
519 | * Walk backwards through sufficiently contiguous I/Os | |
520 | * recording the last non-option I/O. | |
521 | */ | |
522 | while ((dio = AVL_PREV(t, first)) != NULL && | |
523 | (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && | |
524 | IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit && | |
525 | IO_GAP(dio, first) <= maxgap) { | |
526 | first = dio; | |
527 | if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) | |
528 | mandatory = first; | |
529 | } | |
45d1cae3 | 530 | |
a08ee875 LG |
531 | /* |
532 | * Skip any initial optional I/Os. | |
533 | */ | |
534 | while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { | |
535 | first = AVL_NEXT(t, first); | |
536 | ASSERT(first != NULL); | |
537 | } | |
9babb374 | 538 | |
45d1cae3 | 539 | |
a08ee875 LG |
540 | /* |
541 | * Walk forward through sufficiently contiguous I/Os. | |
542 | */ | |
543 | while ((dio = AVL_NEXT(t, last)) != NULL && | |
544 | (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && | |
545 | IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit && | |
546 | IO_GAP(last, dio) <= maxgap) { | |
547 | last = dio; | |
548 | if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) | |
549 | mandatory = last; | |
550 | } | |
551 | ||
552 | /* | |
553 | * Now that we've established the range of the I/O aggregation | |
554 | * we must decide what to do with trailing optional I/Os. | |
555 | * For reads, there's nothing to do. While we are unable to | |
556 | * aggregate further, it's possible that a trailing optional | |
557 | * I/O would allow the underlying device to aggregate with | |
558 | * subsequent I/Os. We must therefore determine if the next | |
559 | * non-optional I/O is close enough to make aggregation | |
560 | * worthwhile. | |
561 | */ | |
562 | if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { | |
563 | zio_t *nio = last; | |
564 | while ((dio = AVL_NEXT(t, nio)) != NULL && | |
565 | IO_GAP(nio, dio) == 0 && | |
566 | IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { | |
567 | nio = dio; | |
568 | if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { | |
569 | stretch = B_TRUE; | |
570 | break; | |
45d1cae3 BB |
571 | } |
572 | } | |
a08ee875 | 573 | } |
45d1cae3 | 574 | |
a08ee875 LG |
575 | if (stretch) { |
576 | /* This may be a no-op. */ | |
577 | dio = AVL_NEXT(t, last); | |
578 | dio->io_flags &= ~ZIO_FLAG_OPTIONAL; | |
579 | } else { | |
580 | while (last != mandatory && last != first) { | |
581 | ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); | |
582 | last = AVL_PREV(t, last); | |
583 | ASSERT(last != NULL); | |
45d1cae3 | 584 | } |
34dc7c2f BB |
585 | } |
586 | ||
a08ee875 LG |
587 | if (first == last) |
588 | return (NULL); | |
d164b209 | 589 | |
a08ee875 LG |
590 | size = IO_SPAN(first, last); |
591 | ASSERT3U(size, <=, zfs_vdev_aggregation_limit); | |
592 | ||
593 | aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, | |
ea04106b | 594 | zio_buf_alloc(size), size, first->io_type, zio->io_priority, |
a08ee875 LG |
595 | flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE, |
596 | vdev_queue_agg_io_done, NULL); | |
597 | aio->io_timestamp = first->io_timestamp; | |
598 | ||
599 | nio = first; | |
600 | do { | |
601 | dio = nio; | |
602 | nio = AVL_NEXT(t, dio); | |
603 | ASSERT3U(dio->io_type, ==, aio->io_type); | |
604 | ||
605 | if (dio->io_flags & ZIO_FLAG_NODATA) { | |
606 | ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); | |
607 | bzero((char *)aio->io_data + (dio->io_offset - | |
608 | aio->io_offset), dio->io_size); | |
609 | } else if (dio->io_type == ZIO_TYPE_WRITE) { | |
610 | bcopy(dio->io_data, (char *)aio->io_data + | |
611 | (dio->io_offset - aio->io_offset), | |
612 | dio->io_size); | |
613 | } | |
34dc7c2f | 614 | |
a08ee875 LG |
615 | zio_add_child(dio, aio); |
616 | vdev_queue_io_remove(vq, dio); | |
617 | zio_vdev_io_bypass(dio); | |
618 | zio_execute(dio); | |
619 | } while (dio != last); | |
620 | ||
a08ee875 LG |
621 | return (aio); |
622 | } | |
34dc7c2f | 623 | |
a08ee875 LG |
624 | static zio_t * |
625 | vdev_queue_io_to_issue(vdev_queue_t *vq) | |
626 | { | |
627 | zio_t *zio, *aio; | |
628 | zio_priority_t p; | |
629 | avl_index_t idx; | |
630 | vdev_queue_class_t *vqc; | |
a08ee875 LG |
631 | |
632 | again: | |
633 | ASSERT(MUTEX_HELD(&vq->vq_lock)); | |
634 | ||
635 | p = vdev_queue_class_to_issue(vq); | |
636 | ||
637 | if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { | |
638 | /* No eligible queued i/os */ | |
639 | return (NULL); | |
34dc7c2f BB |
640 | } |
641 | ||
a08ee875 LG |
642 | /* |
643 | * For LBA-ordered queues (async / scrub), issue the i/o which follows | |
644 | * the most recently issued i/o in LBA (offset) order. | |
645 | * | |
646 | * For FIFO queues (sync), issue the i/o with the lowest timestamp. | |
647 | */ | |
648 | vqc = &vq->vq_class[p]; | |
ea04106b AX |
649 | vq->vq_io_search.io_timestamp = 0; |
650 | vq->vq_io_search.io_offset = vq->vq_last_offset + 1; | |
651 | VERIFY3P(avl_find(&vqc->vqc_queued_tree, &vq->vq_io_search, | |
652 | &idx), ==, NULL); | |
a08ee875 LG |
653 | zio = avl_nearest(&vqc->vqc_queued_tree, idx, AVL_AFTER); |
654 | if (zio == NULL) | |
655 | zio = avl_first(&vqc->vqc_queued_tree); | |
656 | ASSERT3U(zio->io_priority, ==, p); | |
657 | ||
658 | aio = vdev_queue_aggregate(vq, zio); | |
659 | if (aio != NULL) | |
660 | zio = aio; | |
661 | else | |
662 | vdev_queue_io_remove(vq, zio); | |
34dc7c2f | 663 | |
45d1cae3 BB |
664 | /* |
665 | * If the I/O is or was optional and therefore has no data, we need to | |
666 | * simply discard it. We need to drop the vdev queue's lock to avoid a | |
667 | * deadlock that we could encounter since this I/O will complete | |
668 | * immediately. | |
669 | */ | |
a08ee875 | 670 | if (zio->io_flags & ZIO_FLAG_NODATA) { |
45d1cae3 | 671 | mutex_exit(&vq->vq_lock); |
a08ee875 LG |
672 | zio_vdev_io_bypass(zio); |
673 | zio_execute(zio); | |
45d1cae3 BB |
674 | mutex_enter(&vq->vq_lock); |
675 | goto again; | |
676 | } | |
677 | ||
a08ee875 LG |
678 | vdev_queue_pending_add(vq, zio); |
679 | vq->vq_last_offset = zio->io_offset; | |
34dc7c2f | 680 | |
a08ee875 | 681 | return (zio); |
34dc7c2f BB |
682 | } |
683 | ||
684 | zio_t * | |
685 | vdev_queue_io(zio_t *zio) | |
686 | { | |
687 | vdev_queue_t *vq = &zio->io_vd->vdev_queue; | |
688 | zio_t *nio; | |
689 | ||
34dc7c2f BB |
690 | if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) |
691 | return (zio); | |
692 | ||
a08ee875 LG |
693 | /* |
694 | * Children i/os inherent their parent's priority, which might | |
695 | * not match the child's i/o type. Fix it up here. | |
696 | */ | |
697 | if (zio->io_type == ZIO_TYPE_READ) { | |
698 | if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && | |
699 | zio->io_priority != ZIO_PRIORITY_ASYNC_READ && | |
700 | zio->io_priority != ZIO_PRIORITY_SCRUB) | |
701 | zio->io_priority = ZIO_PRIORITY_ASYNC_READ; | |
702 | } else { | |
703 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
704 | if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && | |
705 | zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE) | |
706 | zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; | |
707 | } | |
34dc7c2f | 708 | |
a08ee875 | 709 | zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; |
34dc7c2f BB |
710 | |
711 | mutex_enter(&vq->vq_lock); | |
c06d4368 | 712 | zio->io_timestamp = gethrtime(); |
34dc7c2f | 713 | vdev_queue_io_add(vq, zio); |
a08ee875 | 714 | nio = vdev_queue_io_to_issue(vq); |
34dc7c2f BB |
715 | mutex_exit(&vq->vq_lock); |
716 | ||
717 | if (nio == NULL) | |
718 | return (NULL); | |
719 | ||
720 | if (nio->io_done == vdev_queue_agg_io_done) { | |
721 | zio_nowait(nio); | |
722 | return (NULL); | |
723 | } | |
724 | ||
725 | return (nio); | |
726 | } | |
727 | ||
728 | void | |
729 | vdev_queue_io_done(zio_t *zio) | |
730 | { | |
731 | vdev_queue_t *vq = &zio->io_vd->vdev_queue; | |
a08ee875 | 732 | zio_t *nio; |
34dc7c2f | 733 | |
c06d4368 AX |
734 | if (zio_injection_enabled) |
735 | delay(SEC_TO_TICK(zio_handle_io_delay(zio))); | |
736 | ||
34dc7c2f BB |
737 | mutex_enter(&vq->vq_lock); |
738 | ||
a08ee875 | 739 | vdev_queue_pending_remove(vq, zio); |
34dc7c2f | 740 | |
c06d4368 AX |
741 | zio->io_delta = gethrtime() - zio->io_timestamp; |
742 | vq->vq_io_complete_ts = gethrtime(); | |
743 | vq->vq_io_delta_ts = vq->vq_io_complete_ts - zio->io_timestamp; | |
744 | ||
a08ee875 | 745 | while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { |
34dc7c2f BB |
746 | mutex_exit(&vq->vq_lock); |
747 | if (nio->io_done == vdev_queue_agg_io_done) { | |
748 | zio_nowait(nio); | |
749 | } else { | |
750 | zio_vdev_io_reissue(nio); | |
751 | zio_execute(nio); | |
752 | } | |
753 | mutex_enter(&vq->vq_lock); | |
754 | } | |
755 | ||
756 | mutex_exit(&vq->vq_lock); | |
757 | } | |
c28b2279 BB |
758 | |
759 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
c28b2279 | 760 | module_param(zfs_vdev_aggregation_limit, int, 0644); |
c409e464 BB |
761 | MODULE_PARM_DESC(zfs_vdev_aggregation_limit, "Max vdev I/O aggregation size"); |
762 | ||
c409e464 BB |
763 | module_param(zfs_vdev_read_gap_limit, int, 0644); |
764 | MODULE_PARM_DESC(zfs_vdev_read_gap_limit, "Aggregate read I/O over gap"); | |
765 | ||
766 | module_param(zfs_vdev_write_gap_limit, int, 0644); | |
767 | MODULE_PARM_DESC(zfs_vdev_write_gap_limit, "Aggregate write I/O over gap"); | |
a08ee875 LG |
768 | |
769 | module_param(zfs_vdev_max_active, int, 0644); | |
770 | MODULE_PARM_DESC(zfs_vdev_max_active, "Maximum number of active I/Os per vdev"); | |
771 | ||
772 | module_param(zfs_vdev_async_write_active_max_dirty_percent, int, 0644); | |
773 | MODULE_PARM_DESC(zfs_vdev_async_write_active_max_dirty_percent, | |
774 | "Async write concurrency max threshold"); | |
775 | ||
776 | module_param(zfs_vdev_async_write_active_min_dirty_percent, int, 0644); | |
777 | MODULE_PARM_DESC(zfs_vdev_async_write_active_min_dirty_percent, | |
778 | "Async write concurrency min threshold"); | |
779 | ||
780 | module_param(zfs_vdev_async_read_max_active, int, 0644); | |
781 | MODULE_PARM_DESC(zfs_vdev_async_read_max_active, | |
782 | "Max active async read I/Os per vdev"); | |
783 | ||
784 | module_param(zfs_vdev_async_read_min_active, int, 0644); | |
785 | MODULE_PARM_DESC(zfs_vdev_async_read_min_active, | |
786 | "Min active async read I/Os per vdev"); | |
787 | ||
788 | module_param(zfs_vdev_async_write_max_active, int, 0644); | |
789 | MODULE_PARM_DESC(zfs_vdev_async_write_max_active, | |
790 | "Max active async write I/Os per vdev"); | |
791 | ||
792 | module_param(zfs_vdev_async_write_min_active, int, 0644); | |
793 | MODULE_PARM_DESC(zfs_vdev_async_write_min_active, | |
794 | "Min active async write I/Os per vdev"); | |
795 | ||
796 | module_param(zfs_vdev_scrub_max_active, int, 0644); | |
797 | MODULE_PARM_DESC(zfs_vdev_scrub_max_active, "Max active scrub I/Os per vdev"); | |
798 | ||
799 | module_param(zfs_vdev_scrub_min_active, int, 0644); | |
800 | MODULE_PARM_DESC(zfs_vdev_scrub_min_active, "Min active scrub I/Os per vdev"); | |
801 | ||
802 | module_param(zfs_vdev_sync_read_max_active, int, 0644); | |
803 | MODULE_PARM_DESC(zfs_vdev_sync_read_max_active, | |
804 | "Max active sync read I/Os per vdev"); | |
805 | ||
806 | module_param(zfs_vdev_sync_read_min_active, int, 0644); | |
807 | MODULE_PARM_DESC(zfs_vdev_sync_read_min_active, | |
808 | "Min active sync read I/Os per vdev"); | |
809 | ||
810 | module_param(zfs_vdev_sync_write_max_active, int, 0644); | |
811 | MODULE_PARM_DESC(zfs_vdev_sync_write_max_active, | |
812 | "Max active sync write I/Os per vdev"); | |
813 | ||
814 | module_param(zfs_vdev_sync_write_min_active, int, 0644); | |
815 | MODULE_PARM_DESC(zfs_vdev_sync_write_min_active, | |
816 | "Min active sync write I/Osper vdev"); | |
c28b2279 | 817 | #endif |