]>
Commit | Line | Data |
---|---|---|
87d546d8 TG |
1 | /* |
2 | * CDDL HEADER START | |
3 | * | |
4 | * The contents of this file are subject to the terms of the | |
5 | * Common Development and Distribution License (the "License"). | |
6 | * You may not use this file except in compliance with the License. | |
7 | * | |
8 | * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE | |
9 | * or http://www.opensolaris.org/os/licensing. | |
10 | * See the License for the specific language governing permissions | |
11 | * and limitations under the License. | |
12 | * | |
13 | * When distributing Covered Code, include this CDDL HEADER in each | |
14 | * file and include the License file at usr/src/OPENSOLARIS.LICENSE. | |
15 | * If applicable, add the following below this CDDL HEADER, with the | |
16 | * fields enclosed by brackets "[]" replaced with your own identifying | |
17 | * information: Portions Copyright [yyyy] [name of copyright owner] | |
18 | * | |
19 | * CDDL HEADER END | |
20 | */ | |
21 | /* | |
22 | * Copyright 2009 Sun Microsystems, Inc. All rights reserved. | |
23 | * Use is subject to license terms. | |
24 | */ | |
25 | ||
26 | /* | |
5b5f25ca | 27 | * Copyright (c) 2012, 2017 by Delphix. All rights reserved. |
87d546d8 TG |
28 | */ |
29 | ||
30 | #include <sys/zfs_context.h> | |
31 | #include <sys/vdev_impl.h> | |
32 | #include <sys/spa_impl.h> | |
33 | #include <sys/zio.h> | |
34 | #include <sys/avl.h> | |
35 | #include <sys/dsl_pool.h> | |
36 | #include <sys/spa.h> | |
37 | #include <sys/spa_impl.h> | |
38 | #include <sys/kstat.h> | |
39 | ||
40 | /* | |
41 | * ZFS I/O Scheduler | |
42 | * --------------- | |
43 | * | |
44 | * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The | |
45 | * I/O scheduler determines when and in what order those operations are | |
46 | * issued. The I/O scheduler divides operations into 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). | |
118 | */ | |
119 | ||
120 | /* | |
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. | |
124 | */ | |
125 | uint32_t zfs_vdev_max_active = 1000; | |
126 | ||
127 | /* | |
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. | |
142 | */ | |
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; | |
1ec59ff7 | 149 | uint32_t zfs_vdev_async_write_min_active = 2; |
87d546d8 TG |
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; | |
153 | ||
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; | |
163 | ||
164 | /* | |
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. | |
169 | */ | |
170 | int zfs_vdev_aggregation_limit = SPA_OLD_MAXBLOCKSIZE; | |
171 | int zfs_vdev_read_gap_limit = 32 << 10; | |
172 | int zfs_vdev_write_gap_limit = 4 << 10; | |
173 | ||
174 | int | |
175 | vdev_queue_offset_compare(const void *x1, const void *x2) | |
176 | { | |
177 | const zio_t *z1 = x1; | |
178 | const zio_t *z2 = x2; | |
179 | ||
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 | static inline avl_tree_t * | |
194 | vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p) | |
195 | { | |
196 | return (&vq->vq_class[p].vqc_queued_tree); | |
197 | } | |
198 | ||
199 | static inline avl_tree_t * | |
200 | vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t) | |
201 | { | |
202 | ASSERT(t == ZIO_TYPE_READ || t == ZIO_TYPE_WRITE); | |
203 | if (t == ZIO_TYPE_READ) | |
204 | return (&vq->vq_read_offset_tree); | |
205 | else | |
206 | return (&vq->vq_write_offset_tree); | |
207 | } | |
208 | ||
209 | int | |
210 | vdev_queue_timestamp_compare(const void *x1, const void *x2) | |
211 | { | |
212 | const zio_t *z1 = x1; | |
213 | const zio_t *z2 = x2; | |
214 | ||
215 | if (z1->io_timestamp < z2->io_timestamp) | |
216 | return (-1); | |
217 | if (z1->io_timestamp > z2->io_timestamp) | |
218 | return (1); | |
219 | ||
220 | if (z1 < z2) | |
221 | return (-1); | |
222 | if (z1 > z2) | |
223 | return (1); | |
224 | ||
225 | return (0); | |
226 | } | |
227 | ||
228 | static int | |
229 | vdev_queue_class_min_active(zio_priority_t p) | |
230 | { | |
231 | switch (p) { | |
232 | case ZIO_PRIORITY_SYNC_READ: | |
233 | return (zfs_vdev_sync_read_min_active); | |
234 | case ZIO_PRIORITY_SYNC_WRITE: | |
235 | return (zfs_vdev_sync_write_min_active); | |
236 | case ZIO_PRIORITY_ASYNC_READ: | |
237 | return (zfs_vdev_async_read_min_active); | |
238 | case ZIO_PRIORITY_ASYNC_WRITE: | |
239 | return (zfs_vdev_async_write_min_active); | |
240 | case ZIO_PRIORITY_SCRUB: | |
241 | return (zfs_vdev_scrub_min_active); | |
242 | default: | |
243 | panic("invalid priority %u", p); | |
244 | return (0); | |
245 | } | |
246 | } | |
247 | ||
248 | static int | |
249 | vdev_queue_max_async_writes(spa_t *spa) | |
250 | { | |
251 | int writes; | |
252 | uint64_t dirty = 0; | |
253 | dsl_pool_t *dp = spa_get_dsl(spa); | |
254 | uint64_t min_bytes = zfs_dirty_data_max * | |
255 | zfs_vdev_async_write_active_min_dirty_percent / 100; | |
256 | uint64_t max_bytes = zfs_dirty_data_max * | |
257 | zfs_vdev_async_write_active_max_dirty_percent / 100; | |
258 | ||
259 | /* | |
260 | * Async writes may occur before the assignment of the spa's | |
261 | * dsl_pool_t if a self-healing zio is issued prior to the | |
262 | * completion of dmu_objset_open_impl(). | |
263 | */ | |
264 | if (dp == NULL) | |
265 | return (zfs_vdev_async_write_max_active); | |
266 | ||
267 | /* | |
268 | * Sync tasks correspond to interactive user actions. To reduce the | |
269 | * execution time of those actions we push data out as fast as possible. | |
270 | */ | |
271 | if (spa_has_pending_synctask(spa)) | |
272 | return (zfs_vdev_async_write_max_active); | |
273 | ||
274 | dirty = dp->dp_dirty_total; | |
275 | if (dirty < min_bytes) | |
276 | return (zfs_vdev_async_write_min_active); | |
277 | if (dirty > max_bytes) | |
278 | return (zfs_vdev_async_write_max_active); | |
279 | ||
280 | /* | |
281 | * linear interpolation: | |
282 | * slope = (max_writes - min_writes) / (max_bytes - min_bytes) | |
283 | * move right by min_bytes | |
284 | * move up by min_writes | |
285 | */ | |
286 | writes = (dirty - min_bytes) * | |
287 | (zfs_vdev_async_write_max_active - | |
288 | zfs_vdev_async_write_min_active) / | |
289 | (max_bytes - min_bytes) + | |
290 | zfs_vdev_async_write_min_active; | |
291 | ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); | |
292 | ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); | |
293 | return (writes); | |
294 | } | |
295 | ||
296 | static int | |
297 | vdev_queue_class_max_active(spa_t *spa, zio_priority_t p) | |
298 | { | |
299 | switch (p) { | |
300 | case ZIO_PRIORITY_SYNC_READ: | |
301 | return (zfs_vdev_sync_read_max_active); | |
302 | case ZIO_PRIORITY_SYNC_WRITE: | |
303 | return (zfs_vdev_sync_write_max_active); | |
304 | case ZIO_PRIORITY_ASYNC_READ: | |
305 | return (zfs_vdev_async_read_max_active); | |
306 | case ZIO_PRIORITY_ASYNC_WRITE: | |
307 | return (vdev_queue_max_async_writes(spa)); | |
308 | case ZIO_PRIORITY_SCRUB: | |
309 | return (zfs_vdev_scrub_max_active); | |
310 | default: | |
311 | panic("invalid priority %u", p); | |
312 | return (0); | |
313 | } | |
314 | } | |
315 | ||
316 | /* | |
317 | * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if | |
318 | * there is no eligible class. | |
319 | */ | |
320 | static zio_priority_t | |
321 | vdev_queue_class_to_issue(vdev_queue_t *vq) | |
322 | { | |
323 | spa_t *spa = vq->vq_vdev->vdev_spa; | |
324 | zio_priority_t p; | |
325 | ||
326 | if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active) | |
327 | return (ZIO_PRIORITY_NUM_QUEUEABLE); | |
328 | ||
329 | /* find a queue that has not reached its minimum # outstanding i/os */ | |
330 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { | |
331 | if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && | |
332 | vq->vq_class[p].vqc_active < | |
333 | vdev_queue_class_min_active(p)) | |
334 | return (p); | |
335 | } | |
336 | ||
337 | /* | |
338 | * If we haven't found a queue, look for one that hasn't reached its | |
339 | * maximum # outstanding i/os. | |
340 | */ | |
341 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { | |
342 | if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && | |
343 | vq->vq_class[p].vqc_active < | |
344 | vdev_queue_class_max_active(spa, p)) | |
345 | return (p); | |
346 | } | |
347 | ||
348 | /* No eligible queued i/os */ | |
349 | return (ZIO_PRIORITY_NUM_QUEUEABLE); | |
350 | } | |
351 | ||
352 | void | |
353 | vdev_queue_init(vdev_t *vd) | |
354 | { | |
355 | vdev_queue_t *vq = &vd->vdev_queue; | |
356 | zio_priority_t p; | |
357 | ||
358 | mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); | |
359 | vq->vq_vdev = vd; | |
360 | taskq_init_ent(&vd->vdev_queue.vq_io_search.io_tqent); | |
361 | ||
362 | avl_create(&vq->vq_active_tree, vdev_queue_offset_compare, | |
363 | sizeof (zio_t), offsetof(struct zio, io_queue_node)); | |
364 | avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ), | |
365 | vdev_queue_offset_compare, sizeof (zio_t), | |
366 | offsetof(struct zio, io_offset_node)); | |
367 | avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE), | |
368 | vdev_queue_offset_compare, sizeof (zio_t), | |
369 | offsetof(struct zio, io_offset_node)); | |
370 | ||
371 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { | |
372 | int (*compfn) (const void *, const void *); | |
373 | ||
374 | /* | |
375 | * The synchronous i/o queues are dispatched in FIFO rather | |
376 | * than LBA order. This provides more consistent latency for | |
377 | * these i/os. | |
378 | */ | |
379 | if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE) | |
380 | compfn = vdev_queue_timestamp_compare; | |
381 | else | |
382 | compfn = vdev_queue_offset_compare; | |
383 | avl_create(vdev_queue_class_tree(vq, p), compfn, | |
384 | sizeof (zio_t), offsetof(struct zio, io_queue_node)); | |
385 | } | |
386 | } | |
387 | ||
388 | void | |
389 | vdev_queue_fini(vdev_t *vd) | |
390 | { | |
391 | vdev_queue_t *vq = &vd->vdev_queue; | |
392 | zio_priority_t p; | |
393 | ||
394 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) | |
395 | avl_destroy(vdev_queue_class_tree(vq, p)); | |
396 | avl_destroy(&vq->vq_active_tree); | |
397 | avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ)); | |
398 | avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE)); | |
399 | ||
400 | mutex_destroy(&vq->vq_lock); | |
401 | } | |
402 | ||
403 | static void | |
404 | vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) | |
405 | { | |
406 | spa_t *spa = zio->io_spa; | |
407 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
408 | ||
409 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
410 | avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio); | |
411 | avl_add(vdev_queue_type_tree(vq, zio->io_type), zio); | |
412 | ||
413 | if (ssh->kstat != NULL) { | |
414 | mutex_enter(&ssh->lock); | |
415 | kstat_waitq_enter(ssh->kstat->ks_data); | |
416 | mutex_exit(&ssh->lock); | |
417 | } | |
418 | } | |
419 | ||
420 | static void | |
421 | vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) | |
422 | { | |
423 | spa_t *spa = zio->io_spa; | |
424 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
425 | ||
426 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
427 | avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio); | |
428 | avl_remove(vdev_queue_type_tree(vq, zio->io_type), zio); | |
429 | ||
430 | if (ssh->kstat != NULL) { | |
431 | mutex_enter(&ssh->lock); | |
432 | kstat_waitq_exit(ssh->kstat->ks_data); | |
433 | mutex_exit(&ssh->lock); | |
434 | } | |
435 | } | |
436 | ||
437 | static void | |
438 | vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) | |
439 | { | |
440 | spa_t *spa = zio->io_spa; | |
441 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
442 | ||
443 | ASSERT(MUTEX_HELD(&vq->vq_lock)); | |
444 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
445 | vq->vq_class[zio->io_priority].vqc_active++; | |
446 | avl_add(&vq->vq_active_tree, zio); | |
447 | ||
448 | if (ssh->kstat != NULL) { | |
449 | mutex_enter(&ssh->lock); | |
450 | kstat_runq_enter(ssh->kstat->ks_data); | |
451 | mutex_exit(&ssh->lock); | |
452 | } | |
453 | } | |
454 | ||
455 | static void | |
456 | vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) | |
457 | { | |
458 | spa_t *spa = zio->io_spa; | |
459 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
460 | ||
461 | ASSERT(MUTEX_HELD(&vq->vq_lock)); | |
462 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
463 | vq->vq_class[zio->io_priority].vqc_active--; | |
464 | avl_remove(&vq->vq_active_tree, zio); | |
465 | ||
466 | if (ssh->kstat != NULL) { | |
467 | kstat_io_t *ksio = ssh->kstat->ks_data; | |
468 | ||
469 | mutex_enter(&ssh->lock); | |
470 | kstat_runq_exit(ksio); | |
471 | if (zio->io_type == ZIO_TYPE_READ) { | |
472 | ksio->reads++; | |
473 | ksio->nread += zio->io_size; | |
474 | } else if (zio->io_type == ZIO_TYPE_WRITE) { | |
475 | ksio->writes++; | |
476 | ksio->nwritten += zio->io_size; | |
477 | } | |
478 | mutex_exit(&ssh->lock); | |
479 | } | |
480 | } | |
481 | ||
482 | static void | |
483 | vdev_queue_agg_io_done(zio_t *aio) | |
484 | { | |
485 | if (aio->io_type == ZIO_TYPE_READ) { | |
486 | zio_t *pio; | |
487 | while ((pio = zio_walk_parents(aio)) != NULL) { | |
488 | bcopy((char *)aio->io_data + (pio->io_offset - | |
489 | aio->io_offset), pio->io_data, pio->io_size); | |
490 | } | |
491 | } | |
492 | ||
493 | zio_buf_free(aio->io_data, aio->io_size); | |
494 | } | |
495 | ||
496 | /* | |
497 | * Compute the range spanned by two i/os, which is the endpoint of the last | |
498 | * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). | |
499 | * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); | |
500 | * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. | |
501 | */ | |
502 | #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) | |
503 | #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) | |
504 | ||
505 | static zio_t * | |
506 | vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) | |
507 | { | |
508 | zio_t *first, *last, *aio, *dio, *mandatory, *nio; | |
509 | uint64_t maxgap = 0; | |
510 | uint64_t size; | |
511 | boolean_t stretch = B_FALSE; | |
512 | avl_tree_t *t = vdev_queue_type_tree(vq, zio->io_type); | |
513 | enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; | |
514 | void *buf; | |
515 | ||
516 | if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE) | |
517 | return (NULL); | |
518 | ||
519 | /* | |
520 | * Prevent users from setting the zfs_vdev_aggregation_limit | |
521 | * tuning larger than SPA_MAXBLOCKSIZE. | |
522 | */ | |
523 | zfs_vdev_aggregation_limit = | |
524 | MIN(zfs_vdev_aggregation_limit, SPA_MAXBLOCKSIZE); | |
525 | ||
526 | first = last = zio; | |
527 | ||
528 | if (zio->io_type == ZIO_TYPE_READ) | |
529 | maxgap = zfs_vdev_read_gap_limit; | |
530 | ||
531 | /* | |
532 | * We can aggregate I/Os that are sufficiently adjacent and of | |
533 | * the same flavor, as expressed by the AGG_INHERIT flags. | |
534 | * The latter requirement is necessary so that certain | |
535 | * attributes of the I/O, such as whether it's a normal I/O | |
536 | * or a scrub/resilver, can be preserved in the aggregate. | |
537 | * We can include optional I/Os, but don't allow them | |
538 | * to begin a range as they add no benefit in that situation. | |
539 | */ | |
540 | ||
541 | /* | |
542 | * We keep track of the last non-optional I/O. | |
543 | */ | |
544 | mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; | |
545 | ||
546 | /* | |
547 | * Walk backwards through sufficiently contiguous I/Os | |
5b5f25ca | 548 | * recording the last non-optional I/O. |
87d546d8 TG |
549 | */ |
550 | while ((dio = AVL_PREV(t, first)) != NULL && | |
551 | (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && | |
552 | IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit && | |
553 | IO_GAP(dio, first) <= maxgap) { | |
554 | first = dio; | |
555 | if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) | |
556 | mandatory = first; | |
557 | } | |
558 | ||
559 | /* | |
560 | * Skip any initial optional I/Os. | |
561 | */ | |
562 | while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { | |
563 | first = AVL_NEXT(t, first); | |
564 | ASSERT(first != NULL); | |
565 | } | |
566 | ||
567 | ||
568 | /* | |
569 | * Walk forward through sufficiently contiguous I/Os. | |
5b5f25ca CIK |
570 | * The aggregation limit does not apply to optional i/os, so that |
571 | * we can issue contiguous writes even if they are larger than the | |
572 | * aggregation limit. | |
87d546d8 TG |
573 | */ |
574 | while ((dio = AVL_NEXT(t, last)) != NULL && | |
575 | (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && | |
5b5f25ca CIK |
576 | (IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit || |
577 | (dio->io_flags & ZIO_FLAG_OPTIONAL)) && | |
87d546d8 TG |
578 | IO_GAP(last, dio) <= maxgap) { |
579 | last = dio; | |
580 | if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) | |
581 | mandatory = last; | |
582 | } | |
583 | ||
584 | /* | |
585 | * Now that we've established the range of the I/O aggregation | |
586 | * we must decide what to do with trailing optional I/Os. | |
587 | * For reads, there's nothing to do. While we are unable to | |
588 | * aggregate further, it's possible that a trailing optional | |
589 | * I/O would allow the underlying device to aggregate with | |
590 | * subsequent I/Os. We must therefore determine if the next | |
591 | * non-optional I/O is close enough to make aggregation | |
592 | * worthwhile. | |
593 | */ | |
594 | if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { | |
595 | zio_t *nio = last; | |
596 | while ((dio = AVL_NEXT(t, nio)) != NULL && | |
597 | IO_GAP(nio, dio) == 0 && | |
598 | IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { | |
599 | nio = dio; | |
600 | if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { | |
601 | stretch = B_TRUE; | |
602 | break; | |
603 | } | |
604 | } | |
605 | } | |
606 | ||
607 | if (stretch) { | |
608 | /* This may be a no-op. */ | |
609 | dio = AVL_NEXT(t, last); | |
610 | dio->io_flags &= ~ZIO_FLAG_OPTIONAL; | |
611 | } else { | |
5b5f25ca | 612 | /* do not include the optional i/o */ |
87d546d8 TG |
613 | while (last != mandatory && last != first) { |
614 | ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); | |
615 | last = AVL_PREV(t, last); | |
616 | ASSERT(last != NULL); | |
617 | } | |
618 | } | |
619 | ||
620 | if (first == last) | |
621 | return (NULL); | |
622 | ||
623 | size = IO_SPAN(first, last); | |
87d546d8 TG |
624 | |
625 | buf = zio_buf_alloc_flags(size, KM_NOSLEEP); | |
626 | if (buf == NULL) | |
627 | return (NULL); | |
628 | ||
629 | aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, | |
630 | buf, size, first->io_type, zio->io_priority, | |
631 | flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE, | |
632 | vdev_queue_agg_io_done, NULL); | |
633 | aio->io_timestamp = first->io_timestamp; | |
634 | ||
635 | nio = first; | |
636 | do { | |
637 | dio = nio; | |
638 | nio = AVL_NEXT(t, dio); | |
639 | ASSERT3U(dio->io_type, ==, aio->io_type); | |
640 | ||
641 | if (dio->io_flags & ZIO_FLAG_NODATA) { | |
642 | ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); | |
643 | bzero((char *)aio->io_data + (dio->io_offset - | |
644 | aio->io_offset), dio->io_size); | |
645 | } else if (dio->io_type == ZIO_TYPE_WRITE) { | |
646 | bcopy(dio->io_data, (char *)aio->io_data + | |
647 | (dio->io_offset - aio->io_offset), | |
648 | dio->io_size); | |
649 | } | |
650 | ||
651 | zio_add_child(dio, aio); | |
652 | vdev_queue_io_remove(vq, dio); | |
653 | zio_vdev_io_bypass(dio); | |
654 | zio_execute(dio); | |
655 | } while (dio != last); | |
656 | ||
657 | return (aio); | |
658 | } | |
659 | ||
660 | static zio_t * | |
661 | vdev_queue_io_to_issue(vdev_queue_t *vq) | |
662 | { | |
663 | zio_t *zio, *aio; | |
664 | zio_priority_t p; | |
665 | avl_index_t idx; | |
666 | avl_tree_t *tree; | |
667 | ||
668 | again: | |
669 | ASSERT(MUTEX_HELD(&vq->vq_lock)); | |
670 | ||
671 | p = vdev_queue_class_to_issue(vq); | |
672 | ||
673 | if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { | |
674 | /* No eligible queued i/os */ | |
675 | return (NULL); | |
676 | } | |
677 | ||
678 | /* | |
679 | * For LBA-ordered queues (async / scrub), issue the i/o which follows | |
680 | * the most recently issued i/o in LBA (offset) order. | |
681 | * | |
682 | * For FIFO queues (sync), issue the i/o with the lowest timestamp. | |
683 | */ | |
684 | tree = vdev_queue_class_tree(vq, p); | |
685 | vq->vq_io_search.io_timestamp = 0; | |
686 | vq->vq_io_search.io_offset = vq->vq_last_offset + 1; | |
687 | VERIFY3P(avl_find(tree, &vq->vq_io_search, | |
688 | &idx), ==, NULL); | |
689 | zio = avl_nearest(tree, idx, AVL_AFTER); | |
690 | if (zio == NULL) | |
691 | zio = avl_first(tree); | |
692 | ASSERT3U(zio->io_priority, ==, p); | |
693 | ||
694 | aio = vdev_queue_aggregate(vq, zio); | |
695 | if (aio != NULL) | |
696 | zio = aio; | |
697 | else | |
698 | vdev_queue_io_remove(vq, zio); | |
699 | ||
700 | /* | |
701 | * If the I/O is or was optional and therefore has no data, we need to | |
702 | * simply discard it. We need to drop the vdev queue's lock to avoid a | |
703 | * deadlock that we could encounter since this I/O will complete | |
704 | * immediately. | |
705 | */ | |
706 | if (zio->io_flags & ZIO_FLAG_NODATA) { | |
707 | mutex_exit(&vq->vq_lock); | |
708 | zio_vdev_io_bypass(zio); | |
709 | zio_execute(zio); | |
710 | mutex_enter(&vq->vq_lock); | |
711 | goto again; | |
712 | } | |
713 | ||
714 | vdev_queue_pending_add(vq, zio); | |
715 | vq->vq_last_offset = zio->io_offset; | |
716 | ||
717 | return (zio); | |
718 | } | |
719 | ||
720 | zio_t * | |
721 | vdev_queue_io(zio_t *zio) | |
722 | { | |
723 | vdev_queue_t *vq = &zio->io_vd->vdev_queue; | |
724 | zio_t *nio; | |
725 | ||
726 | if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) | |
727 | return (zio); | |
728 | ||
729 | /* | |
730 | * Children i/os inherent their parent's priority, which might | |
731 | * not match the child's i/o type. Fix it up here. | |
732 | */ | |
733 | if (zio->io_type == ZIO_TYPE_READ) { | |
734 | if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && | |
735 | zio->io_priority != ZIO_PRIORITY_ASYNC_READ && | |
736 | zio->io_priority != ZIO_PRIORITY_SCRUB) | |
737 | zio->io_priority = ZIO_PRIORITY_ASYNC_READ; | |
738 | } else { | |
739 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
740 | if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && | |
741 | zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE) | |
742 | zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; | |
743 | } | |
744 | ||
745 | zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; | |
746 | ||
747 | mutex_enter(&vq->vq_lock); | |
748 | zio->io_timestamp = gethrtime(); | |
749 | vdev_queue_io_add(vq, zio); | |
750 | nio = vdev_queue_io_to_issue(vq); | |
751 | mutex_exit(&vq->vq_lock); | |
752 | ||
753 | if (nio == NULL) | |
754 | return (NULL); | |
755 | ||
756 | if (nio->io_done == vdev_queue_agg_io_done) { | |
757 | zio_nowait(nio); | |
758 | return (NULL); | |
759 | } | |
760 | ||
761 | return (nio); | |
762 | } | |
763 | ||
764 | void | |
765 | vdev_queue_io_done(zio_t *zio) | |
766 | { | |
767 | vdev_queue_t *vq = &zio->io_vd->vdev_queue; | |
768 | zio_t *nio; | |
769 | ||
770 | if (zio_injection_enabled) | |
771 | delay(SEC_TO_TICK(zio_handle_io_delay(zio))); | |
772 | ||
773 | mutex_enter(&vq->vq_lock); | |
774 | ||
775 | vdev_queue_pending_remove(vq, zio); | |
776 | ||
777 | zio->io_delta = gethrtime() - zio->io_timestamp; | |
778 | vq->vq_io_complete_ts = gethrtime(); | |
779 | vq->vq_io_delta_ts = vq->vq_io_complete_ts - zio->io_timestamp; | |
780 | ||
781 | while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { | |
782 | mutex_exit(&vq->vq_lock); | |
783 | if (nio->io_done == vdev_queue_agg_io_done) { | |
784 | zio_nowait(nio); | |
785 | } else { | |
786 | zio_vdev_io_reissue(nio); | |
787 | zio_execute(nio); | |
788 | } | |
789 | mutex_enter(&vq->vq_lock); | |
790 | } | |
791 | ||
792 | mutex_exit(&vq->vq_lock); | |
793 | } | |
794 | ||
795 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
796 | module_param(zfs_vdev_aggregation_limit, int, 0644); | |
797 | MODULE_PARM_DESC(zfs_vdev_aggregation_limit, "Max vdev I/O aggregation size"); | |
798 | ||
799 | module_param(zfs_vdev_read_gap_limit, int, 0644); | |
800 | MODULE_PARM_DESC(zfs_vdev_read_gap_limit, "Aggregate read I/O over gap"); | |
801 | ||
802 | module_param(zfs_vdev_write_gap_limit, int, 0644); | |
803 | MODULE_PARM_DESC(zfs_vdev_write_gap_limit, "Aggregate write I/O over gap"); | |
804 | ||
805 | module_param(zfs_vdev_max_active, int, 0644); | |
806 | MODULE_PARM_DESC(zfs_vdev_max_active, "Maximum number of active I/Os per vdev"); | |
807 | ||
808 | module_param(zfs_vdev_async_write_active_max_dirty_percent, int, 0644); | |
809 | MODULE_PARM_DESC(zfs_vdev_async_write_active_max_dirty_percent, | |
810 | "Async write concurrency max threshold"); | |
811 | ||
812 | module_param(zfs_vdev_async_write_active_min_dirty_percent, int, 0644); | |
813 | MODULE_PARM_DESC(zfs_vdev_async_write_active_min_dirty_percent, | |
814 | "Async write concurrency min threshold"); | |
815 | ||
816 | module_param(zfs_vdev_async_read_max_active, int, 0644); | |
817 | MODULE_PARM_DESC(zfs_vdev_async_read_max_active, | |
818 | "Max active async read I/Os per vdev"); | |
819 | ||
820 | module_param(zfs_vdev_async_read_min_active, int, 0644); | |
821 | MODULE_PARM_DESC(zfs_vdev_async_read_min_active, | |
822 | "Min active async read I/Os per vdev"); | |
823 | ||
824 | module_param(zfs_vdev_async_write_max_active, int, 0644); | |
825 | MODULE_PARM_DESC(zfs_vdev_async_write_max_active, | |
826 | "Max active async write I/Os per vdev"); | |
827 | ||
828 | module_param(zfs_vdev_async_write_min_active, int, 0644); | |
829 | MODULE_PARM_DESC(zfs_vdev_async_write_min_active, | |
830 | "Min active async write I/Os per vdev"); | |
831 | ||
832 | module_param(zfs_vdev_scrub_max_active, int, 0644); | |
833 | MODULE_PARM_DESC(zfs_vdev_scrub_max_active, "Max active scrub I/Os per vdev"); | |
834 | ||
835 | module_param(zfs_vdev_scrub_min_active, int, 0644); | |
836 | MODULE_PARM_DESC(zfs_vdev_scrub_min_active, "Min active scrub I/Os per vdev"); | |
837 | ||
838 | module_param(zfs_vdev_sync_read_max_active, int, 0644); | |
839 | MODULE_PARM_DESC(zfs_vdev_sync_read_max_active, | |
840 | "Max active sync read I/Os per vdev"); | |
841 | ||
842 | module_param(zfs_vdev_sync_read_min_active, int, 0644); | |
843 | MODULE_PARM_DESC(zfs_vdev_sync_read_min_active, | |
844 | "Min active sync read I/Os per vdev"); | |
845 | ||
846 | module_param(zfs_vdev_sync_write_max_active, int, 0644); | |
847 | MODULE_PARM_DESC(zfs_vdev_sync_write_max_active, | |
848 | "Max active sync write I/Os per vdev"); | |
849 | ||
850 | module_param(zfs_vdev_sync_write_min_active, int, 0644); | |
851 | MODULE_PARM_DESC(zfs_vdev_sync_write_min_active, | |
852 | "Min active sync write I/Os per vdev"); | |
853 | #endif |