]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * CDDL HEADER START | |
3 | * | |
4 | * The contents of this file are subject to the terms of the | |
5 | * Common Development and Distribution License (the "License"). | |
6 | * You may not use this file except in compliance with the License. | |
7 | * | |
8 | * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE | |
9 | * or http://www.opensolaris.org/os/licensing. | |
10 | * See the License for the specific language governing permissions | |
11 | * and limitations under the License. | |
12 | * | |
13 | * When distributing Covered Code, include this CDDL HEADER in each | |
14 | * file and include the License file at usr/src/OPENSOLARIS.LICENSE. | |
15 | * If applicable, add the following below this CDDL HEADER, with the | |
16 | * fields enclosed by brackets "[]" replaced with your own identifying | |
17 | * information: Portions Copyright [yyyy] [name of copyright owner] | |
18 | * | |
19 | * CDDL HEADER END | |
20 | */ | |
21 | /* | |
22 | * Copyright 2009 Sun Microsystems, Inc. All rights reserved. | |
23 | * Use is subject to license terms. | |
24 | */ | |
25 | ||
26 | /* | |
27 | * Copyright (c) 2013 by Delphix. All rights reserved. | |
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; | |
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; | |
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_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 | int | |
194 | vdev_queue_timestamp_compare(const void *x1, const void *x2) | |
195 | { | |
196 | const zio_t *z1 = x1; | |
197 | const zio_t *z2 = x2; | |
198 | ||
199 | if (z1->io_timestamp < z2->io_timestamp) | |
200 | return (-1); | |
201 | if (z1->io_timestamp > z2->io_timestamp) | |
202 | return (1); | |
203 | ||
204 | if (z1 < z2) | |
205 | return (-1); | |
206 | if (z1 > z2) | |
207 | return (1); | |
208 | ||
209 | return (0); | |
210 | } | |
211 | ||
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 | |
233 | vdev_queue_max_async_writes(uint64_t dirty) | |
234 | { | |
235 | int writes; | |
236 | uint64_t min_bytes = zfs_dirty_data_max * | |
237 | zfs_vdev_async_write_active_min_dirty_percent / 100; | |
238 | uint64_t max_bytes = zfs_dirty_data_max * | |
239 | zfs_vdev_async_write_active_max_dirty_percent / 100; | |
240 | ||
241 | if (dirty < min_bytes) | |
242 | return (zfs_vdev_async_write_min_active); | |
243 | if (dirty > max_bytes) | |
244 | return (zfs_vdev_async_write_max_active); | |
245 | ||
246 | /* | |
247 | * linear interpolation: | |
248 | * slope = (max_writes - min_writes) / (max_bytes - min_bytes) | |
249 | * move right by min_bytes | |
250 | * move up by min_writes | |
251 | */ | |
252 | writes = (dirty - min_bytes) * | |
253 | (zfs_vdev_async_write_max_active - | |
254 | zfs_vdev_async_write_min_active) / | |
255 | (max_bytes - min_bytes) + | |
256 | zfs_vdev_async_write_min_active; | |
257 | ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); | |
258 | ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); | |
259 | return (writes); | |
260 | } | |
261 | ||
262 | static int | |
263 | vdev_queue_class_max_active(spa_t *spa, zio_priority_t p) | |
264 | { | |
265 | switch (p) { | |
266 | case ZIO_PRIORITY_SYNC_READ: | |
267 | return (zfs_vdev_sync_read_max_active); | |
268 | case ZIO_PRIORITY_SYNC_WRITE: | |
269 | return (zfs_vdev_sync_write_max_active); | |
270 | case ZIO_PRIORITY_ASYNC_READ: | |
271 | return (zfs_vdev_async_read_max_active); | |
272 | case ZIO_PRIORITY_ASYNC_WRITE: | |
273 | return (vdev_queue_max_async_writes( | |
274 | spa->spa_dsl_pool->dp_dirty_total)); | |
275 | case ZIO_PRIORITY_SCRUB: | |
276 | return (zfs_vdev_scrub_max_active); | |
277 | default: | |
278 | panic("invalid priority %u", p); | |
279 | return (0); | |
280 | } | |
281 | } | |
282 | ||
283 | /* | |
284 | * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if | |
285 | * there is no eligible class. | |
286 | */ | |
287 | static zio_priority_t | |
288 | vdev_queue_class_to_issue(vdev_queue_t *vq) | |
289 | { | |
290 | spa_t *spa = vq->vq_vdev->vdev_spa; | |
291 | zio_priority_t p; | |
292 | ||
293 | if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active) | |
294 | return (ZIO_PRIORITY_NUM_QUEUEABLE); | |
295 | ||
296 | /* find a queue that has not reached its minimum # outstanding i/os */ | |
297 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { | |
298 | if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 && | |
299 | vq->vq_class[p].vqc_active < | |
300 | vdev_queue_class_min_active(p)) | |
301 | return (p); | |
302 | } | |
303 | ||
304 | /* | |
305 | * If we haven't found a queue, look for one that hasn't reached its | |
306 | * maximum # outstanding i/os. | |
307 | */ | |
308 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { | |
309 | if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 && | |
310 | vq->vq_class[p].vqc_active < | |
311 | vdev_queue_class_max_active(spa, p)) | |
312 | return (p); | |
313 | } | |
314 | ||
315 | /* No eligible queued i/os */ | |
316 | return (ZIO_PRIORITY_NUM_QUEUEABLE); | |
317 | } | |
318 | ||
319 | void | |
320 | vdev_queue_init(vdev_t *vd) | |
321 | { | |
322 | vdev_queue_t *vq = &vd->vdev_queue; | |
323 | int max_active_sum; | |
324 | zio_priority_t p; | |
325 | int i; | |
326 | ||
327 | mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); | |
328 | vq->vq_vdev = vd; | |
329 | ||
330 | avl_create(&vq->vq_active_tree, vdev_queue_offset_compare, | |
331 | sizeof (zio_t), offsetof(struct zio, io_queue_node)); | |
332 | ||
333 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { | |
334 | /* | |
335 | * The synchronous i/o queues are FIFO rather than LBA ordered. | |
336 | * This provides more consistent latency for these i/os, and | |
337 | * they tend to not be tightly clustered anyway so there is | |
338 | * little to no throughput loss. | |
339 | */ | |
340 | boolean_t fifo = (p == ZIO_PRIORITY_SYNC_READ || | |
341 | p == ZIO_PRIORITY_SYNC_WRITE); | |
342 | avl_create(&vq->vq_class[p].vqc_queued_tree, | |
343 | fifo ? vdev_queue_timestamp_compare : | |
344 | vdev_queue_offset_compare, | |
345 | sizeof (zio_t), offsetof(struct zio, io_queue_node)); | |
346 | } | |
347 | ||
348 | /* | |
349 | * A list of buffers which can be used for aggregate I/O, this | |
350 | * avoids the need to allocate them on demand when memory is low. | |
351 | */ | |
352 | list_create(&vq->vq_io_list, sizeof (vdev_io_t), | |
353 | offsetof(vdev_io_t, vi_node)); | |
354 | ||
355 | max_active_sum = zfs_vdev_sync_read_max_active + | |
356 | zfs_vdev_sync_write_max_active + zfs_vdev_async_read_max_active + | |
357 | zfs_vdev_async_write_max_active + zfs_vdev_scrub_max_active; | |
358 | for (i = 0; i < max_active_sum; i++) | |
359 | list_insert_tail(&vq->vq_io_list, zio_vdev_alloc()); | |
360 | } | |
361 | ||
362 | void | |
363 | vdev_queue_fini(vdev_t *vd) | |
364 | { | |
365 | vdev_queue_t *vq = &vd->vdev_queue; | |
366 | vdev_io_t *vi; | |
367 | zio_priority_t p; | |
368 | ||
369 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) | |
370 | avl_destroy(&vq->vq_class[p].vqc_queued_tree); | |
371 | avl_destroy(&vq->vq_active_tree); | |
372 | ||
373 | while ((vi = list_head(&vq->vq_io_list)) != NULL) { | |
374 | list_remove(&vq->vq_io_list, vi); | |
375 | zio_vdev_free(vi); | |
376 | } | |
377 | ||
378 | list_destroy(&vq->vq_io_list); | |
379 | ||
380 | mutex_destroy(&vq->vq_lock); | |
381 | } | |
382 | ||
383 | static void | |
384 | vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) | |
385 | { | |
386 | spa_t *spa = zio->io_spa; | |
387 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
388 | ||
389 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
390 | avl_add(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio); | |
391 | ||
392 | if (ssh->kstat != NULL) { | |
393 | mutex_enter(&ssh->lock); | |
394 | kstat_waitq_enter(ssh->kstat->ks_data); | |
395 | mutex_exit(&ssh->lock); | |
396 | } | |
397 | } | |
398 | ||
399 | static void | |
400 | vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) | |
401 | { | |
402 | spa_t *spa = zio->io_spa; | |
403 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
404 | ||
405 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
406 | avl_remove(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio); | |
407 | ||
408 | if (ssh->kstat != NULL) { | |
409 | mutex_enter(&ssh->lock); | |
410 | kstat_waitq_exit(ssh->kstat->ks_data); | |
411 | mutex_exit(&ssh->lock); | |
412 | } | |
413 | } | |
414 | ||
415 | static void | |
416 | vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) | |
417 | { | |
418 | spa_t *spa = zio->io_spa; | |
419 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
420 | ||
421 | ASSERT(MUTEX_HELD(&vq->vq_lock)); | |
422 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
423 | vq->vq_class[zio->io_priority].vqc_active++; | |
424 | avl_add(&vq->vq_active_tree, zio); | |
425 | ||
426 | if (ssh->kstat != NULL) { | |
427 | mutex_enter(&ssh->lock); | |
428 | kstat_runq_enter(ssh->kstat->ks_data); | |
429 | mutex_exit(&ssh->lock); | |
430 | } | |
431 | } | |
432 | ||
433 | static void | |
434 | vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) | |
435 | { | |
436 | spa_t *spa = zio->io_spa; | |
437 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
438 | ||
439 | ASSERT(MUTEX_HELD(&vq->vq_lock)); | |
440 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
441 | vq->vq_class[zio->io_priority].vqc_active--; | |
442 | avl_remove(&vq->vq_active_tree, zio); | |
443 | ||
444 | if (ssh->kstat != NULL) { | |
445 | kstat_io_t *ksio = ssh->kstat->ks_data; | |
446 | ||
447 | mutex_enter(&ssh->lock); | |
448 | kstat_runq_exit(ksio); | |
449 | if (zio->io_type == ZIO_TYPE_READ) { | |
450 | ksio->reads++; | |
451 | ksio->nread += zio->io_size; | |
452 | } else if (zio->io_type == ZIO_TYPE_WRITE) { | |
453 | ksio->writes++; | |
454 | ksio->nwritten += zio->io_size; | |
455 | } | |
456 | mutex_exit(&ssh->lock); | |
457 | } | |
458 | } | |
459 | ||
460 | static void | |
461 | vdev_queue_agg_io_done(zio_t *aio) | |
462 | { | |
463 | vdev_queue_t *vq = &aio->io_vd->vdev_queue; | |
464 | vdev_io_t *vi = aio->io_data; | |
465 | ||
466 | if (aio->io_type == ZIO_TYPE_READ) { | |
467 | zio_t *pio; | |
468 | while ((pio = zio_walk_parents(aio)) != NULL) { | |
469 | bcopy((char *)aio->io_data + (pio->io_offset - | |
470 | aio->io_offset), pio->io_data, pio->io_size); | |
471 | } | |
472 | } | |
473 | ||
474 | mutex_enter(&vq->vq_lock); | |
475 | list_insert_tail(&vq->vq_io_list, vi); | |
476 | mutex_exit(&vq->vq_lock); | |
477 | } | |
478 | ||
479 | /* | |
480 | * Compute the range spanned by two i/os, which is the endpoint of the last | |
481 | * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). | |
482 | * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); | |
483 | * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. | |
484 | */ | |
485 | #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) | |
486 | #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) | |
487 | ||
488 | static zio_t * | |
489 | vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) | |
490 | { | |
491 | vdev_io_t *vi; | |
492 | zio_t *first, *last, *aio, *dio, *mandatory, *nio; | |
493 | uint64_t maxgap = 0; | |
494 | uint64_t size; | |
495 | boolean_t stretch = B_FALSE; | |
496 | vdev_queue_class_t *vqc = &vq->vq_class[zio->io_priority]; | |
497 | avl_tree_t *t = &vqc->vqc_queued_tree; | |
498 | enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; | |
499 | ||
500 | if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE) | |
501 | return (NULL); | |
502 | ||
503 | /* | |
504 | * Prevent users from setting the zfs_vdev_aggregation_limit | |
505 | * tuning larger than SPA_MAXBLOCKSIZE. | |
506 | */ | |
507 | zfs_vdev_aggregation_limit = | |
508 | MIN(zfs_vdev_aggregation_limit, SPA_MAXBLOCKSIZE); | |
509 | ||
510 | /* | |
511 | * The synchronous i/o queues are not sorted by LBA, so we can't | |
512 | * find adjacent i/os. These i/os tend to not be tightly clustered, | |
513 | * or too large to aggregate, so this has little impact on performance. | |
514 | */ | |
515 | if (zio->io_priority == ZIO_PRIORITY_SYNC_READ || | |
516 | zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) | |
517 | return (NULL); | |
518 | ||
519 | first = last = zio; | |
520 | ||
521 | if (zio->io_type == ZIO_TYPE_READ) | |
522 | maxgap = zfs_vdev_read_gap_limit; | |
523 | ||
524 | vi = list_head(&vq->vq_io_list); | |
525 | if (vi == NULL) { | |
526 | vi = zio_vdev_alloc(); | |
527 | list_insert_head(&vq->vq_io_list, vi); | |
528 | } | |
529 | ||
530 | /* | |
531 | * We can aggregate I/Os that are sufficiently adjacent and of | |
532 | * the same flavor, as expressed by the AGG_INHERIT flags. | |
533 | * The latter requirement is necessary so that certain | |
534 | * attributes of the I/O, such as whether it's a normal I/O | |
535 | * or a scrub/resilver, can be preserved in the aggregate. | |
536 | * We can include optional I/Os, but don't allow them | |
537 | * to begin a range as they add no benefit in that situation. | |
538 | */ | |
539 | ||
540 | /* | |
541 | * We keep track of the last non-optional I/O. | |
542 | */ | |
543 | mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; | |
544 | ||
545 | /* | |
546 | * Walk backwards through sufficiently contiguous I/Os | |
547 | * recording the last non-option I/O. | |
548 | */ | |
549 | while ((dio = AVL_PREV(t, first)) != NULL && | |
550 | (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && | |
551 | IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit && | |
552 | IO_GAP(dio, first) <= maxgap) { | |
553 | first = dio; | |
554 | if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) | |
555 | mandatory = first; | |
556 | } | |
557 | ||
558 | /* | |
559 | * Skip any initial optional I/Os. | |
560 | */ | |
561 | while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { | |
562 | first = AVL_NEXT(t, first); | |
563 | ASSERT(first != NULL); | |
564 | } | |
565 | ||
566 | ||
567 | /* | |
568 | * Walk forward through sufficiently contiguous I/Os. | |
569 | */ | |
570 | while ((dio = AVL_NEXT(t, last)) != NULL && | |
571 | (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && | |
572 | IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit && | |
573 | IO_GAP(last, dio) <= maxgap) { | |
574 | last = dio; | |
575 | if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) | |
576 | mandatory = last; | |
577 | } | |
578 | ||
579 | /* | |
580 | * Now that we've established the range of the I/O aggregation | |
581 | * we must decide what to do with trailing optional I/Os. | |
582 | * For reads, there's nothing to do. While we are unable to | |
583 | * aggregate further, it's possible that a trailing optional | |
584 | * I/O would allow the underlying device to aggregate with | |
585 | * subsequent I/Os. We must therefore determine if the next | |
586 | * non-optional I/O is close enough to make aggregation | |
587 | * worthwhile. | |
588 | */ | |
589 | if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { | |
590 | zio_t *nio = last; | |
591 | while ((dio = AVL_NEXT(t, nio)) != NULL && | |
592 | IO_GAP(nio, dio) == 0 && | |
593 | IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { | |
594 | nio = dio; | |
595 | if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { | |
596 | stretch = B_TRUE; | |
597 | break; | |
598 | } | |
599 | } | |
600 | } | |
601 | ||
602 | if (stretch) { | |
603 | /* This may be a no-op. */ | |
604 | dio = AVL_NEXT(t, last); | |
605 | dio->io_flags &= ~ZIO_FLAG_OPTIONAL; | |
606 | } else { | |
607 | while (last != mandatory && last != first) { | |
608 | ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); | |
609 | last = AVL_PREV(t, last); | |
610 | ASSERT(last != NULL); | |
611 | } | |
612 | } | |
613 | ||
614 | if (first == last) | |
615 | return (NULL); | |
616 | ||
617 | ASSERT(vi != NULL); | |
618 | ||
619 | size = IO_SPAN(first, last); | |
620 | ASSERT3U(size, <=, zfs_vdev_aggregation_limit); | |
621 | ||
622 | aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, | |
623 | vi, size, first->io_type, zio->io_priority, | |
624 | flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE, | |
625 | vdev_queue_agg_io_done, NULL); | |
626 | aio->io_timestamp = first->io_timestamp; | |
627 | ||
628 | nio = first; | |
629 | do { | |
630 | dio = nio; | |
631 | nio = AVL_NEXT(t, dio); | |
632 | ASSERT3U(dio->io_type, ==, aio->io_type); | |
633 | ||
634 | if (dio->io_flags & ZIO_FLAG_NODATA) { | |
635 | ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); | |
636 | bzero((char *)aio->io_data + (dio->io_offset - | |
637 | aio->io_offset), dio->io_size); | |
638 | } else if (dio->io_type == ZIO_TYPE_WRITE) { | |
639 | bcopy(dio->io_data, (char *)aio->io_data + | |
640 | (dio->io_offset - aio->io_offset), | |
641 | dio->io_size); | |
642 | } | |
643 | ||
644 | zio_add_child(dio, aio); | |
645 | vdev_queue_io_remove(vq, dio); | |
646 | zio_vdev_io_bypass(dio); | |
647 | zio_execute(dio); | |
648 | } while (dio != last); | |
649 | ||
650 | list_remove(&vq->vq_io_list, vi); | |
651 | ||
652 | return (aio); | |
653 | } | |
654 | ||
655 | static zio_t * | |
656 | vdev_queue_io_to_issue(vdev_queue_t *vq) | |
657 | { | |
658 | zio_t *zio, *aio; | |
659 | zio_priority_t p; | |
660 | avl_index_t idx; | |
661 | vdev_queue_class_t *vqc; | |
662 | zio_t *search; | |
663 | ||
664 | again: | |
665 | ASSERT(MUTEX_HELD(&vq->vq_lock)); | |
666 | ||
667 | p = vdev_queue_class_to_issue(vq); | |
668 | ||
669 | if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { | |
670 | /* No eligible queued i/os */ | |
671 | return (NULL); | |
672 | } | |
673 | ||
674 | /* | |
675 | * For LBA-ordered queues (async / scrub), issue the i/o which follows | |
676 | * the most recently issued i/o in LBA (offset) order. | |
677 | * | |
678 | * For FIFO queues (sync), issue the i/o with the lowest timestamp. | |
679 | */ | |
680 | vqc = &vq->vq_class[p]; | |
681 | search = zio_buf_alloc(sizeof (*search)); | |
682 | search->io_timestamp = 0; | |
683 | search->io_offset = vq->vq_last_offset + 1; | |
684 | VERIFY3P(avl_find(&vqc->vqc_queued_tree, search, &idx), ==, NULL); | |
685 | zio_buf_free(search, sizeof (*search)); | |
686 | zio = avl_nearest(&vqc->vqc_queued_tree, idx, AVL_AFTER); | |
687 | if (zio == NULL) | |
688 | zio = avl_first(&vqc->vqc_queued_tree); | |
689 | ASSERT3U(zio->io_priority, ==, p); | |
690 | ||
691 | aio = vdev_queue_aggregate(vq, zio); | |
692 | if (aio != NULL) | |
693 | zio = aio; | |
694 | else | |
695 | vdev_queue_io_remove(vq, zio); | |
696 | ||
697 | /* | |
698 | * If the I/O is or was optional and therefore has no data, we need to | |
699 | * simply discard it. We need to drop the vdev queue's lock to avoid a | |
700 | * deadlock that we could encounter since this I/O will complete | |
701 | * immediately. | |
702 | */ | |
703 | if (zio->io_flags & ZIO_FLAG_NODATA) { | |
704 | mutex_exit(&vq->vq_lock); | |
705 | zio_vdev_io_bypass(zio); | |
706 | zio_execute(zio); | |
707 | mutex_enter(&vq->vq_lock); | |
708 | goto again; | |
709 | } | |
710 | ||
711 | vdev_queue_pending_add(vq, zio); | |
712 | vq->vq_last_offset = zio->io_offset; | |
713 | ||
714 | return (zio); | |
715 | } | |
716 | ||
717 | zio_t * | |
718 | vdev_queue_io(zio_t *zio) | |
719 | { | |
720 | vdev_queue_t *vq = &zio->io_vd->vdev_queue; | |
721 | zio_t *nio; | |
722 | ||
723 | if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) | |
724 | return (zio); | |
725 | ||
726 | /* | |
727 | * Children i/os inherent their parent's priority, which might | |
728 | * not match the child's i/o type. Fix it up here. | |
729 | */ | |
730 | if (zio->io_type == ZIO_TYPE_READ) { | |
731 | if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && | |
732 | zio->io_priority != ZIO_PRIORITY_ASYNC_READ && | |
733 | zio->io_priority != ZIO_PRIORITY_SCRUB) | |
734 | zio->io_priority = ZIO_PRIORITY_ASYNC_READ; | |
735 | } else { | |
736 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
737 | if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && | |
738 | zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE) | |
739 | zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; | |
740 | } | |
741 | ||
742 | zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; | |
743 | ||
744 | mutex_enter(&vq->vq_lock); | |
745 | zio->io_timestamp = gethrtime(); | |
746 | vdev_queue_io_add(vq, zio); | |
747 | nio = vdev_queue_io_to_issue(vq); | |
748 | mutex_exit(&vq->vq_lock); | |
749 | ||
750 | if (nio == NULL) | |
751 | return (NULL); | |
752 | ||
753 | if (nio->io_done == vdev_queue_agg_io_done) { | |
754 | zio_nowait(nio); | |
755 | return (NULL); | |
756 | } | |
757 | ||
758 | return (nio); | |
759 | } | |
760 | ||
761 | void | |
762 | vdev_queue_io_done(zio_t *zio) | |
763 | { | |
764 | vdev_queue_t *vq = &zio->io_vd->vdev_queue; | |
765 | zio_t *nio; | |
766 | ||
767 | if (zio_injection_enabled) | |
768 | delay(SEC_TO_TICK(zio_handle_io_delay(zio))); | |
769 | ||
770 | mutex_enter(&vq->vq_lock); | |
771 | ||
772 | vdev_queue_pending_remove(vq, zio); | |
773 | ||
774 | zio->io_delta = gethrtime() - zio->io_timestamp; | |
775 | vq->vq_io_complete_ts = gethrtime(); | |
776 | vq->vq_io_delta_ts = vq->vq_io_complete_ts - zio->io_timestamp; | |
777 | ||
778 | while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { | |
779 | mutex_exit(&vq->vq_lock); | |
780 | if (nio->io_done == vdev_queue_agg_io_done) { | |
781 | zio_nowait(nio); | |
782 | } else { | |
783 | zio_vdev_io_reissue(nio); | |
784 | zio_execute(nio); | |
785 | } | |
786 | mutex_enter(&vq->vq_lock); | |
787 | } | |
788 | ||
789 | mutex_exit(&vq->vq_lock); | |
790 | } | |
791 | ||
792 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
793 | module_param(zfs_vdev_aggregation_limit, int, 0644); | |
794 | MODULE_PARM_DESC(zfs_vdev_aggregation_limit, "Max vdev I/O aggregation size"); | |
795 | ||
796 | module_param(zfs_vdev_read_gap_limit, int, 0644); | |
797 | MODULE_PARM_DESC(zfs_vdev_read_gap_limit, "Aggregate read I/O over gap"); | |
798 | ||
799 | module_param(zfs_vdev_write_gap_limit, int, 0644); | |
800 | MODULE_PARM_DESC(zfs_vdev_write_gap_limit, "Aggregate write I/O over gap"); | |
801 | ||
802 | module_param(zfs_vdev_max_active, int, 0644); | |
803 | MODULE_PARM_DESC(zfs_vdev_max_active, "Maximum number of active I/Os per vdev"); | |
804 | ||
805 | module_param(zfs_vdev_async_write_active_max_dirty_percent, int, 0644); | |
806 | MODULE_PARM_DESC(zfs_vdev_async_write_active_max_dirty_percent, | |
807 | "Async write concurrency max threshold"); | |
808 | ||
809 | module_param(zfs_vdev_async_write_active_min_dirty_percent, int, 0644); | |
810 | MODULE_PARM_DESC(zfs_vdev_async_write_active_min_dirty_percent, | |
811 | "Async write concurrency min threshold"); | |
812 | ||
813 | module_param(zfs_vdev_async_read_max_active, int, 0644); | |
814 | MODULE_PARM_DESC(zfs_vdev_async_read_max_active, | |
815 | "Max active async read I/Os per vdev"); | |
816 | ||
817 | module_param(zfs_vdev_async_read_min_active, int, 0644); | |
818 | MODULE_PARM_DESC(zfs_vdev_async_read_min_active, | |
819 | "Min active async read I/Os per vdev"); | |
820 | ||
821 | module_param(zfs_vdev_async_write_max_active, int, 0644); | |
822 | MODULE_PARM_DESC(zfs_vdev_async_write_max_active, | |
823 | "Max active async write I/Os per vdev"); | |
824 | ||
825 | module_param(zfs_vdev_async_write_min_active, int, 0644); | |
826 | MODULE_PARM_DESC(zfs_vdev_async_write_min_active, | |
827 | "Min active async write I/Os per vdev"); | |
828 | ||
829 | module_param(zfs_vdev_scrub_max_active, int, 0644); | |
830 | MODULE_PARM_DESC(zfs_vdev_scrub_max_active, "Max active scrub I/Os per vdev"); | |
831 | ||
832 | module_param(zfs_vdev_scrub_min_active, int, 0644); | |
833 | MODULE_PARM_DESC(zfs_vdev_scrub_min_active, "Min active scrub I/Os per vdev"); | |
834 | ||
835 | module_param(zfs_vdev_sync_read_max_active, int, 0644); | |
836 | MODULE_PARM_DESC(zfs_vdev_sync_read_max_active, | |
837 | "Max active sync read I/Os per vdev"); | |
838 | ||
839 | module_param(zfs_vdev_sync_read_min_active, int, 0644); | |
840 | MODULE_PARM_DESC(zfs_vdev_sync_read_min_active, | |
841 | "Min active sync read I/Os per vdev"); | |
842 | ||
843 | module_param(zfs_vdev_sync_write_max_active, int, 0644); | |
844 | MODULE_PARM_DESC(zfs_vdev_sync_write_max_active, | |
845 | "Max active sync write I/Os per vdev"); | |
846 | ||
847 | module_param(zfs_vdev_sync_write_min_active, int, 0644); | |
848 | MODULE_PARM_DESC(zfs_vdev_sync_write_min_active, | |
849 | "Min active sync write I/Osper vdev"); | |
850 | #endif |