]>
Commit | Line | Data |
---|---|---|
34dc7c2f BB |
1 | /* |
2 | * CDDL HEADER START | |
3 | * | |
4 | * The contents of this file are subject to the terms of the | |
5 | * Common Development and Distribution License (the "License"). | |
6 | * You may not use this file except in compliance with the License. | |
7 | * | |
8 | * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE | |
9 | * or http://www.opensolaris.org/os/licensing. | |
10 | * See the License for the specific language governing permissions | |
11 | * and limitations under the License. | |
12 | * | |
13 | * When distributing Covered Code, include this CDDL HEADER in each | |
14 | * file and include the License file at usr/src/OPENSOLARIS.LICENSE. | |
15 | * If applicable, add the following below this CDDL HEADER, with the | |
16 | * fields enclosed by brackets "[]" replaced with your own identifying | |
17 | * information: Portions Copyright [yyyy] [name of copyright owner] | |
18 | * | |
19 | * CDDL HEADER END | |
20 | */ | |
21 | /* | |
d164b209 | 22 | * Copyright 2009 Sun Microsystems, Inc. All rights reserved. |
34dc7c2f BB |
23 | * Use is subject to license terms. |
24 | */ | |
25 | ||
cc92e9d0 | 26 | /* |
acbad6ff | 27 | * Copyright (c) 2012, 2014 by Delphix. All rights reserved. |
cc92e9d0 GW |
28 | */ |
29 | ||
34dc7c2f | 30 | #include <sys/zfs_context.h> |
34dc7c2f | 31 | #include <sys/vdev_impl.h> |
330847ff | 32 | #include <sys/spa_impl.h> |
34dc7c2f BB |
33 | #include <sys/zio.h> |
34 | #include <sys/avl.h> | |
e8b96c60 | 35 | #include <sys/dsl_pool.h> |
3dfb57a3 | 36 | #include <sys/metaslab_impl.h> |
e8b96c60 MA |
37 | #include <sys/spa.h> |
38 | #include <sys/spa_impl.h> | |
330847ff | 39 | #include <sys/kstat.h> |
34dc7c2f BB |
40 | |
41 | /* | |
e8b96c60 MA |
42 | * ZFS I/O Scheduler |
43 | * --------------- | |
44 | * | |
45 | * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The | |
46 | * I/O scheduler determines when and in what order those operations are | |
47 | * issued. The I/O scheduler divides operations into five I/O classes | |
48 | * prioritized in the following order: sync read, sync write, async read, | |
49 | * async write, and scrub/resilver. Each queue defines the minimum and | |
50 | * maximum number of concurrent operations that may be issued to the device. | |
51 | * In addition, the device has an aggregate maximum. Note that the sum of the | |
52 | * per-queue minimums must not exceed the aggregate maximum. If the | |
53 | * sum of the per-queue maximums exceeds the aggregate maximum, then the | |
54 | * number of active i/os may reach zfs_vdev_max_active, in which case no | |
55 | * further i/os will be issued regardless of whether all per-queue | |
56 | * minimums have been met. | |
57 | * | |
58 | * For many physical devices, throughput increases with the number of | |
59 | * concurrent operations, but latency typically suffers. Further, physical | |
60 | * devices typically have a limit at which more concurrent operations have no | |
61 | * effect on throughput or can actually cause it to decrease. | |
62 | * | |
63 | * The scheduler selects the next operation to issue by first looking for an | |
64 | * I/O class whose minimum has not been satisfied. Once all are satisfied and | |
65 | * the aggregate maximum has not been hit, the scheduler looks for classes | |
66 | * whose maximum has not been satisfied. Iteration through the I/O classes is | |
67 | * done in the order specified above. No further operations are issued if the | |
68 | * aggregate maximum number of concurrent operations has been hit or if there | |
69 | * are no operations queued for an I/O class that has not hit its maximum. | |
70 | * Every time an i/o is queued or an operation completes, the I/O scheduler | |
71 | * looks for new operations to issue. | |
72 | * | |
73 | * All I/O classes have a fixed maximum number of outstanding operations | |
74 | * except for the async write class. Asynchronous writes represent the data | |
75 | * that is committed to stable storage during the syncing stage for | |
76 | * transaction groups (see txg.c). Transaction groups enter the syncing state | |
77 | * periodically so the number of queued async writes will quickly burst up and | |
78 | * then bleed down to zero. Rather than servicing them as quickly as possible, | |
79 | * the I/O scheduler changes the maximum number of active async write i/os | |
80 | * according to the amount of dirty data in the pool (see dsl_pool.c). Since | |
81 | * both throughput and latency typically increase with the number of | |
82 | * concurrent operations issued to physical devices, reducing the burstiness | |
83 | * in the number of concurrent operations also stabilizes the response time of | |
84 | * operations from other -- and in particular synchronous -- queues. In broad | |
85 | * strokes, the I/O scheduler will issue more concurrent operations from the | |
86 | * async write queue as there's more dirty data in the pool. | |
87 | * | |
88 | * Async Writes | |
89 | * | |
90 | * The number of concurrent operations issued for the async write I/O class | |
91 | * follows a piece-wise linear function defined by a few adjustable points. | |
92 | * | |
93 | * | o---------| <-- zfs_vdev_async_write_max_active | |
94 | * ^ | /^ | | |
95 | * | | / | | | |
96 | * active | / | | | |
97 | * I/O | / | | | |
98 | * count | / | | | |
99 | * | / | | | |
100 | * |------------o | | <-- zfs_vdev_async_write_min_active | |
101 | * 0|____________^______|_________| | |
102 | * 0% | | 100% of zfs_dirty_data_max | |
103 | * | | | |
104 | * | `-- zfs_vdev_async_write_active_max_dirty_percent | |
105 | * `--------- zfs_vdev_async_write_active_min_dirty_percent | |
106 | * | |
107 | * Until the amount of dirty data exceeds a minimum percentage of the dirty | |
108 | * data allowed in the pool, the I/O scheduler will limit the number of | |
109 | * concurrent operations to the minimum. As that threshold is crossed, the | |
110 | * number of concurrent operations issued increases linearly to the maximum at | |
111 | * the specified maximum percentage of the dirty data allowed in the pool. | |
112 | * | |
113 | * Ideally, the amount of dirty data on a busy pool will stay in the sloped | |
114 | * part of the function between zfs_vdev_async_write_active_min_dirty_percent | |
115 | * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the | |
116 | * maximum percentage, this indicates that the rate of incoming data is | |
117 | * greater than the rate that the backend storage can handle. In this case, we | |
118 | * must further throttle incoming writes (see dmu_tx_delay() for details). | |
34dc7c2f | 119 | */ |
d3cc8b15 | 120 | |
34dc7c2f | 121 | /* |
e8b96c60 MA |
122 | * The maximum number of i/os active to each device. Ideally, this will be >= |
123 | * the sum of each queue's max_active. It must be at least the sum of each | |
124 | * queue's min_active. | |
34dc7c2f | 125 | */ |
e8b96c60 | 126 | uint32_t zfs_vdev_max_active = 1000; |
34dc7c2f | 127 | |
cb682a17 | 128 | /* |
e8b96c60 MA |
129 | * Per-queue limits on the number of i/os active to each device. If the |
130 | * number of active i/os is < zfs_vdev_max_active, then the min_active comes | |
131 | * into play. We will send min_active from each queue, and then select from | |
132 | * queues in the order defined by zio_priority_t. | |
133 | * | |
134 | * In general, smaller max_active's will lead to lower latency of synchronous | |
135 | * operations. Larger max_active's may lead to higher overall throughput, | |
136 | * depending on underlying storage. | |
137 | * | |
138 | * The ratio of the queues' max_actives determines the balance of performance | |
139 | * between reads, writes, and scrubs. E.g., increasing | |
140 | * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete | |
141 | * more quickly, but reads and writes to have higher latency and lower | |
142 | * throughput. | |
cb682a17 | 143 | */ |
e8b96c60 MA |
144 | uint32_t zfs_vdev_sync_read_min_active = 10; |
145 | uint32_t zfs_vdev_sync_read_max_active = 10; | |
146 | uint32_t zfs_vdev_sync_write_min_active = 10; | |
147 | uint32_t zfs_vdev_sync_write_max_active = 10; | |
148 | uint32_t zfs_vdev_async_read_min_active = 1; | |
149 | uint32_t zfs_vdev_async_read_max_active = 3; | |
150 | uint32_t zfs_vdev_async_write_min_active = 1; | |
151 | uint32_t zfs_vdev_async_write_max_active = 10; | |
152 | uint32_t zfs_vdev_scrub_min_active = 1; | |
153 | uint32_t zfs_vdev_scrub_max_active = 2; | |
34dc7c2f | 154 | |
e8b96c60 MA |
155 | /* |
156 | * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent | |
157 | * dirty data, use zfs_vdev_async_write_min_active. When it has more than | |
158 | * zfs_vdev_async_write_active_max_dirty_percent, use | |
159 | * zfs_vdev_async_write_max_active. The value is linearly interpolated | |
160 | * between min and max. | |
161 | */ | |
162 | int zfs_vdev_async_write_active_min_dirty_percent = 30; | |
163 | int zfs_vdev_async_write_active_max_dirty_percent = 60; | |
34dc7c2f BB |
164 | |
165 | /* | |
45d1cae3 BB |
166 | * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O. |
167 | * For read I/Os, we also aggregate across small adjacency gaps; for writes | |
168 | * we include spans of optional I/Os to aid aggregation at the disk even when | |
169 | * they aren't able to help us aggregate at this level. | |
34dc7c2f | 170 | */ |
f1512ee6 | 171 | int zfs_vdev_aggregation_limit = SPA_OLD_MAXBLOCKSIZE; |
9babb374 | 172 | int zfs_vdev_read_gap_limit = 32 << 10; |
45d1cae3 | 173 | int zfs_vdev_write_gap_limit = 4 << 10; |
34dc7c2f | 174 | |
3dfb57a3 DB |
175 | /* |
176 | * Define the queue depth percentage for each top-level. This percentage is | |
177 | * used in conjunction with zfs_vdev_async_max_active to determine how many | |
178 | * allocations a specific top-level vdev should handle. Once the queue depth | |
179 | * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100 | |
180 | * then allocator will stop allocating blocks on that top-level device. | |
181 | * The default kernel setting is 1000% which will yield 100 allocations per | |
182 | * device. For userland testing, the default setting is 300% which equates | |
183 | * to 30 allocations per device. | |
184 | */ | |
185 | #ifdef _KERNEL | |
186 | int zfs_vdev_queue_depth_pct = 1000; | |
187 | #else | |
188 | int zfs_vdev_queue_depth_pct = 300; | |
189 | #endif | |
190 | ||
191 | ||
34dc7c2f | 192 | int |
e8b96c60 | 193 | vdev_queue_offset_compare(const void *x1, const void *x2) |
34dc7c2f | 194 | { |
ee36c709 GN |
195 | const zio_t *z1 = (const zio_t *)x1; |
196 | const zio_t *z2 = (const zio_t *)x2; | |
34dc7c2f | 197 | |
ee36c709 | 198 | int cmp = AVL_CMP(z1->io_offset, z2->io_offset); |
34dc7c2f | 199 | |
ee36c709 GN |
200 | if (likely(cmp)) |
201 | return (cmp); | |
34dc7c2f | 202 | |
ee36c709 | 203 | return (AVL_PCMP(z1, z2)); |
34dc7c2f BB |
204 | } |
205 | ||
ec8501ee JG |
206 | static inline avl_tree_t * |
207 | vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p) | |
208 | { | |
209 | return (&vq->vq_class[p].vqc_queued_tree); | |
210 | } | |
211 | ||
212 | static inline avl_tree_t * | |
213 | vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t) | |
214 | { | |
215 | ASSERT(t == ZIO_TYPE_READ || t == ZIO_TYPE_WRITE); | |
216 | if (t == ZIO_TYPE_READ) | |
217 | return (&vq->vq_read_offset_tree); | |
218 | else | |
219 | return (&vq->vq_write_offset_tree); | |
220 | } | |
221 | ||
34dc7c2f | 222 | int |
e8b96c60 | 223 | vdev_queue_timestamp_compare(const void *x1, const void *x2) |
34dc7c2f | 224 | { |
ee36c709 GN |
225 | const zio_t *z1 = (const zio_t *)x1; |
226 | const zio_t *z2 = (const zio_t *)x2; | |
34dc7c2f | 227 | |
ee36c709 | 228 | int cmp = AVL_CMP(z1->io_timestamp, z2->io_timestamp); |
34dc7c2f | 229 | |
ee36c709 GN |
230 | if (likely(cmp)) |
231 | return (cmp); | |
34dc7c2f | 232 | |
ee36c709 | 233 | return (AVL_PCMP(z1, z2)); |
34dc7c2f BB |
234 | } |
235 | ||
e8b96c60 MA |
236 | static int |
237 | vdev_queue_class_min_active(zio_priority_t p) | |
238 | { | |
239 | switch (p) { | |
240 | case ZIO_PRIORITY_SYNC_READ: | |
241 | return (zfs_vdev_sync_read_min_active); | |
242 | case ZIO_PRIORITY_SYNC_WRITE: | |
243 | return (zfs_vdev_sync_write_min_active); | |
244 | case ZIO_PRIORITY_ASYNC_READ: | |
245 | return (zfs_vdev_async_read_min_active); | |
246 | case ZIO_PRIORITY_ASYNC_WRITE: | |
247 | return (zfs_vdev_async_write_min_active); | |
248 | case ZIO_PRIORITY_SCRUB: | |
249 | return (zfs_vdev_scrub_min_active); | |
250 | default: | |
251 | panic("invalid priority %u", p); | |
252 | return (0); | |
253 | } | |
254 | } | |
255 | ||
256 | static int | |
acbad6ff | 257 | vdev_queue_max_async_writes(spa_t *spa) |
e8b96c60 MA |
258 | { |
259 | int writes; | |
b7faa7aa G |
260 | uint64_t dirty = 0; |
261 | dsl_pool_t *dp = spa_get_dsl(spa); | |
e8b96c60 MA |
262 | uint64_t min_bytes = zfs_dirty_data_max * |
263 | zfs_vdev_async_write_active_min_dirty_percent / 100; | |
264 | uint64_t max_bytes = zfs_dirty_data_max * | |
265 | zfs_vdev_async_write_active_max_dirty_percent / 100; | |
266 | ||
b7faa7aa G |
267 | /* |
268 | * Async writes may occur before the assignment of the spa's | |
269 | * dsl_pool_t if a self-healing zio is issued prior to the | |
270 | * completion of dmu_objset_open_impl(). | |
271 | */ | |
272 | if (dp == NULL) | |
273 | return (zfs_vdev_async_write_max_active); | |
274 | ||
acbad6ff AR |
275 | /* |
276 | * Sync tasks correspond to interactive user actions. To reduce the | |
277 | * execution time of those actions we push data out as fast as possible. | |
278 | */ | |
b7faa7aa | 279 | if (spa_has_pending_synctask(spa)) |
acbad6ff | 280 | return (zfs_vdev_async_write_max_active); |
acbad6ff | 281 | |
b7faa7aa | 282 | dirty = dp->dp_dirty_total; |
e8b96c60 MA |
283 | if (dirty < min_bytes) |
284 | return (zfs_vdev_async_write_min_active); | |
285 | if (dirty > max_bytes) | |
286 | return (zfs_vdev_async_write_max_active); | |
287 | ||
288 | /* | |
289 | * linear interpolation: | |
290 | * slope = (max_writes - min_writes) / (max_bytes - min_bytes) | |
291 | * move right by min_bytes | |
292 | * move up by min_writes | |
293 | */ | |
294 | writes = (dirty - min_bytes) * | |
295 | (zfs_vdev_async_write_max_active - | |
296 | zfs_vdev_async_write_min_active) / | |
297 | (max_bytes - min_bytes) + | |
298 | zfs_vdev_async_write_min_active; | |
299 | ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); | |
300 | ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); | |
301 | return (writes); | |
302 | } | |
303 | ||
304 | static int | |
305 | vdev_queue_class_max_active(spa_t *spa, zio_priority_t p) | |
306 | { | |
307 | switch (p) { | |
308 | case ZIO_PRIORITY_SYNC_READ: | |
309 | return (zfs_vdev_sync_read_max_active); | |
310 | case ZIO_PRIORITY_SYNC_WRITE: | |
311 | return (zfs_vdev_sync_write_max_active); | |
312 | case ZIO_PRIORITY_ASYNC_READ: | |
313 | return (zfs_vdev_async_read_max_active); | |
314 | case ZIO_PRIORITY_ASYNC_WRITE: | |
acbad6ff | 315 | return (vdev_queue_max_async_writes(spa)); |
e8b96c60 MA |
316 | case ZIO_PRIORITY_SCRUB: |
317 | return (zfs_vdev_scrub_max_active); | |
318 | default: | |
319 | panic("invalid priority %u", p); | |
320 | return (0); | |
321 | } | |
322 | } | |
323 | ||
324 | /* | |
325 | * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if | |
326 | * there is no eligible class. | |
327 | */ | |
328 | static zio_priority_t | |
329 | vdev_queue_class_to_issue(vdev_queue_t *vq) | |
330 | { | |
331 | spa_t *spa = vq->vq_vdev->vdev_spa; | |
332 | zio_priority_t p; | |
333 | ||
334 | if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active) | |
335 | return (ZIO_PRIORITY_NUM_QUEUEABLE); | |
336 | ||
337 | /* find a queue that has not reached its minimum # outstanding i/os */ | |
338 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { | |
ec8501ee | 339 | if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && |
e8b96c60 MA |
340 | vq->vq_class[p].vqc_active < |
341 | vdev_queue_class_min_active(p)) | |
342 | return (p); | |
343 | } | |
344 | ||
345 | /* | |
346 | * If we haven't found a queue, look for one that hasn't reached its | |
347 | * maximum # outstanding i/os. | |
348 | */ | |
349 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { | |
ec8501ee | 350 | if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && |
e8b96c60 MA |
351 | vq->vq_class[p].vqc_active < |
352 | vdev_queue_class_max_active(spa, p)) | |
353 | return (p); | |
354 | } | |
355 | ||
356 | /* No eligible queued i/os */ | |
357 | return (ZIO_PRIORITY_NUM_QUEUEABLE); | |
358 | } | |
359 | ||
34dc7c2f BB |
360 | void |
361 | vdev_queue_init(vdev_t *vd) | |
362 | { | |
363 | vdev_queue_t *vq = &vd->vdev_queue; | |
e8b96c60 | 364 | zio_priority_t p; |
34dc7c2f BB |
365 | |
366 | mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); | |
e8b96c60 | 367 | vq->vq_vdev = vd; |
36b454ab | 368 | taskq_init_ent(&vd->vdev_queue.vq_io_search.io_tqent); |
34dc7c2f | 369 | |
e8b96c60 MA |
370 | avl_create(&vq->vq_active_tree, vdev_queue_offset_compare, |
371 | sizeof (zio_t), offsetof(struct zio, io_queue_node)); | |
ec8501ee JG |
372 | avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ), |
373 | vdev_queue_offset_compare, sizeof (zio_t), | |
374 | offsetof(struct zio, io_offset_node)); | |
375 | avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE), | |
376 | vdev_queue_offset_compare, sizeof (zio_t), | |
377 | offsetof(struct zio, io_offset_node)); | |
34dc7c2f | 378 | |
e8b96c60 | 379 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { |
ec8501ee JG |
380 | int (*compfn) (const void *, const void *); |
381 | ||
e8b96c60 | 382 | /* |
ec8501ee JG |
383 | * The synchronous i/o queues are dispatched in FIFO rather |
384 | * than LBA order. This provides more consistent latency for | |
385 | * these i/os. | |
e8b96c60 | 386 | */ |
ec8501ee JG |
387 | if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE) |
388 | compfn = vdev_queue_timestamp_compare; | |
389 | else | |
390 | compfn = vdev_queue_offset_compare; | |
391 | avl_create(vdev_queue_class_tree(vq, p), compfn, | |
392 | sizeof (zio_t), offsetof(struct zio, io_queue_node)); | |
e8b96c60 | 393 | } |
9f500936 | 394 | |
395 | vq->vq_lastoffset = 0; | |
34dc7c2f BB |
396 | } |
397 | ||
398 | void | |
399 | vdev_queue_fini(vdev_t *vd) | |
400 | { | |
401 | vdev_queue_t *vq = &vd->vdev_queue; | |
e8b96c60 | 402 | zio_priority_t p; |
34dc7c2f | 403 | |
e8b96c60 | 404 | for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) |
ec8501ee | 405 | avl_destroy(vdev_queue_class_tree(vq, p)); |
e8b96c60 | 406 | avl_destroy(&vq->vq_active_tree); |
ec8501ee JG |
407 | avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ)); |
408 | avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE)); | |
34dc7c2f BB |
409 | |
410 | mutex_destroy(&vq->vq_lock); | |
411 | } | |
412 | ||
413 | static void | |
414 | vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) | |
415 | { | |
330847ff MA |
416 | spa_t *spa = zio->io_spa; |
417 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
418 | ||
e8b96c60 | 419 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); |
ec8501ee JG |
420 | avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio); |
421 | avl_add(vdev_queue_type_tree(vq, zio->io_type), zio); | |
330847ff MA |
422 | |
423 | if (ssh->kstat != NULL) { | |
424 | mutex_enter(&ssh->lock); | |
425 | kstat_waitq_enter(ssh->kstat->ks_data); | |
426 | mutex_exit(&ssh->lock); | |
427 | } | |
34dc7c2f BB |
428 | } |
429 | ||
430 | static void | |
431 | vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) | |
432 | { | |
330847ff MA |
433 | spa_t *spa = zio->io_spa; |
434 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
435 | ||
e8b96c60 | 436 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); |
ec8501ee JG |
437 | avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio); |
438 | avl_remove(vdev_queue_type_tree(vq, zio->io_type), zio); | |
330847ff MA |
439 | |
440 | if (ssh->kstat != NULL) { | |
441 | mutex_enter(&ssh->lock); | |
442 | kstat_waitq_exit(ssh->kstat->ks_data); | |
443 | mutex_exit(&ssh->lock); | |
444 | } | |
445 | } | |
446 | ||
447 | static void | |
448 | vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) | |
449 | { | |
450 | spa_t *spa = zio->io_spa; | |
451 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
452 | ||
e8b96c60 MA |
453 | ASSERT(MUTEX_HELD(&vq->vq_lock)); |
454 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
455 | vq->vq_class[zio->io_priority].vqc_active++; | |
456 | avl_add(&vq->vq_active_tree, zio); | |
330847ff MA |
457 | |
458 | if (ssh->kstat != NULL) { | |
459 | mutex_enter(&ssh->lock); | |
460 | kstat_runq_enter(ssh->kstat->ks_data); | |
461 | mutex_exit(&ssh->lock); | |
462 | } | |
463 | } | |
464 | ||
465 | static void | |
466 | vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) | |
467 | { | |
468 | spa_t *spa = zio->io_spa; | |
469 | spa_stats_history_t *ssh = &spa->spa_stats.io_history; | |
470 | ||
e8b96c60 MA |
471 | ASSERT(MUTEX_HELD(&vq->vq_lock)); |
472 | ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
473 | vq->vq_class[zio->io_priority].vqc_active--; | |
474 | avl_remove(&vq->vq_active_tree, zio); | |
330847ff MA |
475 | |
476 | if (ssh->kstat != NULL) { | |
477 | kstat_io_t *ksio = ssh->kstat->ks_data; | |
478 | ||
479 | mutex_enter(&ssh->lock); | |
480 | kstat_runq_exit(ksio); | |
481 | if (zio->io_type == ZIO_TYPE_READ) { | |
482 | ksio->reads++; | |
483 | ksio->nread += zio->io_size; | |
484 | } else if (zio->io_type == ZIO_TYPE_WRITE) { | |
485 | ksio->writes++; | |
486 | ksio->nwritten += zio->io_size; | |
487 | } | |
488 | mutex_exit(&ssh->lock); | |
489 | } | |
34dc7c2f BB |
490 | } |
491 | ||
492 | static void | |
493 | vdev_queue_agg_io_done(zio_t *aio) | |
494 | { | |
e8b96c60 MA |
495 | if (aio->io_type == ZIO_TYPE_READ) { |
496 | zio_t *pio; | |
3dfb57a3 DB |
497 | zio_link_t *zl = NULL; |
498 | while ((pio = zio_walk_parents(aio, &zl)) != NULL) { | |
d164b209 BB |
499 | bcopy((char *)aio->io_data + (pio->io_offset - |
500 | aio->io_offset), pio->io_data, pio->io_size); | |
e8b96c60 MA |
501 | } |
502 | } | |
34dc7c2f | 503 | |
285b29d9 | 504 | zio_buf_free(aio->io_data, aio->io_size); |
34dc7c2f BB |
505 | } |
506 | ||
9babb374 BB |
507 | /* |
508 | * Compute the range spanned by two i/os, which is the endpoint of the last | |
509 | * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). | |
510 | * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); | |
511 | * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. | |
512 | */ | |
513 | #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) | |
514 | #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) | |
34dc7c2f BB |
515 | |
516 | static zio_t * | |
e8b96c60 | 517 | vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) |
34dc7c2f | 518 | { |
e8b96c60 MA |
519 | zio_t *first, *last, *aio, *dio, *mandatory, *nio; |
520 | uint64_t maxgap = 0; | |
521 | uint64_t size; | |
a58df6f5 | 522 | uint64_t limit; |
e8b96c60 | 523 | boolean_t stretch = B_FALSE; |
ec8501ee | 524 | avl_tree_t *t = vdev_queue_type_tree(vq, zio->io_type); |
e8b96c60 | 525 | enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; |
6fe53787 | 526 | void *buf; |
e8b96c60 | 527 | |
a58df6f5 BB |
528 | limit = MAX(MIN(zfs_vdev_aggregation_limit, |
529 | spa_maxblocksize(vq->vq_vdev->vdev_spa)), 0); | |
34dc7c2f | 530 | |
a58df6f5 BB |
531 | if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE || limit == 0) |
532 | return (NULL); | |
34dc7c2f | 533 | |
e8b96c60 | 534 | first = last = zio; |
34dc7c2f | 535 | |
e8b96c60 MA |
536 | if (zio->io_type == ZIO_TYPE_READ) |
537 | maxgap = zfs_vdev_read_gap_limit; | |
fb5f0bc8 | 538 | |
e8b96c60 MA |
539 | /* |
540 | * We can aggregate I/Os that are sufficiently adjacent and of | |
541 | * the same flavor, as expressed by the AGG_INHERIT flags. | |
542 | * The latter requirement is necessary so that certain | |
543 | * attributes of the I/O, such as whether it's a normal I/O | |
544 | * or a scrub/resilver, can be preserved in the aggregate. | |
545 | * We can include optional I/Os, but don't allow them | |
546 | * to begin a range as they add no benefit in that situation. | |
547 | */ | |
45d1cae3 | 548 | |
e8b96c60 MA |
549 | /* |
550 | * We keep track of the last non-optional I/O. | |
551 | */ | |
552 | mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; | |
45d1cae3 | 553 | |
e8b96c60 MA |
554 | /* |
555 | * Walk backwards through sufficiently contiguous I/Os | |
556 | * recording the last non-option I/O. | |
557 | */ | |
558 | while ((dio = AVL_PREV(t, first)) != NULL && | |
559 | (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && | |
a58df6f5 | 560 | IO_SPAN(dio, last) <= limit && |
e8b96c60 MA |
561 | IO_GAP(dio, first) <= maxgap) { |
562 | first = dio; | |
563 | if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) | |
564 | mandatory = first; | |
565 | } | |
45d1cae3 | 566 | |
e8b96c60 MA |
567 | /* |
568 | * Skip any initial optional I/Os. | |
569 | */ | |
570 | while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { | |
571 | first = AVL_NEXT(t, first); | |
572 | ASSERT(first != NULL); | |
573 | } | |
9babb374 | 574 | |
45d1cae3 | 575 | |
e8b96c60 MA |
576 | /* |
577 | * Walk forward through sufficiently contiguous I/Os. | |
578 | */ | |
579 | while ((dio = AVL_NEXT(t, last)) != NULL && | |
580 | (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && | |
a58df6f5 | 581 | IO_SPAN(first, dio) <= limit && |
e8b96c60 MA |
582 | IO_GAP(last, dio) <= maxgap) { |
583 | last = dio; | |
584 | if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) | |
585 | mandatory = last; | |
586 | } | |
587 | ||
588 | /* | |
589 | * Now that we've established the range of the I/O aggregation | |
590 | * we must decide what to do with trailing optional I/Os. | |
591 | * For reads, there's nothing to do. While we are unable to | |
592 | * aggregate further, it's possible that a trailing optional | |
593 | * I/O would allow the underlying device to aggregate with | |
594 | * subsequent I/Os. We must therefore determine if the next | |
595 | * non-optional I/O is close enough to make aggregation | |
596 | * worthwhile. | |
597 | */ | |
598 | if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { | |
599 | zio_t *nio = last; | |
600 | while ((dio = AVL_NEXT(t, nio)) != NULL && | |
601 | IO_GAP(nio, dio) == 0 && | |
602 | IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { | |
603 | nio = dio; | |
604 | if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { | |
605 | stretch = B_TRUE; | |
606 | break; | |
45d1cae3 BB |
607 | } |
608 | } | |
e8b96c60 | 609 | } |
45d1cae3 | 610 | |
e8b96c60 MA |
611 | if (stretch) { |
612 | /* This may be a no-op. */ | |
613 | dio = AVL_NEXT(t, last); | |
614 | dio->io_flags &= ~ZIO_FLAG_OPTIONAL; | |
615 | } else { | |
616 | while (last != mandatory && last != first) { | |
617 | ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); | |
618 | last = AVL_PREV(t, last); | |
619 | ASSERT(last != NULL); | |
45d1cae3 | 620 | } |
34dc7c2f BB |
621 | } |
622 | ||
e8b96c60 MA |
623 | if (first == last) |
624 | return (NULL); | |
625 | ||
e8b96c60 | 626 | size = IO_SPAN(first, last); |
a58df6f5 | 627 | ASSERT3U(size, <=, limit); |
e8b96c60 | 628 | |
6fe53787 BB |
629 | buf = zio_buf_alloc_flags(size, KM_NOSLEEP); |
630 | if (buf == NULL) | |
631 | return (NULL); | |
632 | ||
e8b96c60 | 633 | aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, |
6fe53787 | 634 | buf, size, first->io_type, zio->io_priority, |
e8b96c60 MA |
635 | flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE, |
636 | vdev_queue_agg_io_done, NULL); | |
637 | aio->io_timestamp = first->io_timestamp; | |
638 | ||
639 | nio = first; | |
640 | do { | |
641 | dio = nio; | |
642 | nio = AVL_NEXT(t, dio); | |
643 | ASSERT3U(dio->io_type, ==, aio->io_type); | |
644 | ||
645 | if (dio->io_flags & ZIO_FLAG_NODATA) { | |
646 | ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); | |
647 | bzero((char *)aio->io_data + (dio->io_offset - | |
648 | aio->io_offset), dio->io_size); | |
649 | } else if (dio->io_type == ZIO_TYPE_WRITE) { | |
650 | bcopy(dio->io_data, (char *)aio->io_data + | |
651 | (dio->io_offset - aio->io_offset), | |
652 | dio->io_size); | |
653 | } | |
d164b209 | 654 | |
e8b96c60 MA |
655 | zio_add_child(dio, aio); |
656 | vdev_queue_io_remove(vq, dio); | |
657 | zio_vdev_io_bypass(dio); | |
658 | zio_execute(dio); | |
659 | } while (dio != last); | |
34dc7c2f | 660 | |
e8b96c60 MA |
661 | return (aio); |
662 | } | |
663 | ||
664 | static zio_t * | |
665 | vdev_queue_io_to_issue(vdev_queue_t *vq) | |
666 | { | |
667 | zio_t *zio, *aio; | |
668 | zio_priority_t p; | |
669 | avl_index_t idx; | |
ec8501ee | 670 | avl_tree_t *tree; |
e8b96c60 MA |
671 | |
672 | again: | |
673 | ASSERT(MUTEX_HELD(&vq->vq_lock)); | |
674 | ||
675 | p = vdev_queue_class_to_issue(vq); | |
34dc7c2f | 676 | |
e8b96c60 MA |
677 | if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { |
678 | /* No eligible queued i/os */ | |
679 | return (NULL); | |
34dc7c2f BB |
680 | } |
681 | ||
e8b96c60 MA |
682 | /* |
683 | * For LBA-ordered queues (async / scrub), issue the i/o which follows | |
684 | * the most recently issued i/o in LBA (offset) order. | |
685 | * | |
686 | * For FIFO queues (sync), issue the i/o with the lowest timestamp. | |
687 | */ | |
ec8501ee | 688 | tree = vdev_queue_class_tree(vq, p); |
50b25b21 BB |
689 | vq->vq_io_search.io_timestamp = 0; |
690 | vq->vq_io_search.io_offset = vq->vq_last_offset + 1; | |
ec8501ee | 691 | VERIFY3P(avl_find(tree, &vq->vq_io_search, |
50b25b21 | 692 | &idx), ==, NULL); |
ec8501ee | 693 | zio = avl_nearest(tree, idx, AVL_AFTER); |
e8b96c60 | 694 | if (zio == NULL) |
ec8501ee | 695 | zio = avl_first(tree); |
e8b96c60 MA |
696 | ASSERT3U(zio->io_priority, ==, p); |
697 | ||
698 | aio = vdev_queue_aggregate(vq, zio); | |
699 | if (aio != NULL) | |
700 | zio = aio; | |
701 | else | |
702 | vdev_queue_io_remove(vq, zio); | |
34dc7c2f | 703 | |
45d1cae3 BB |
704 | /* |
705 | * If the I/O is or was optional and therefore has no data, we need to | |
706 | * simply discard it. We need to drop the vdev queue's lock to avoid a | |
707 | * deadlock that we could encounter since this I/O will complete | |
708 | * immediately. | |
709 | */ | |
e8b96c60 | 710 | if (zio->io_flags & ZIO_FLAG_NODATA) { |
45d1cae3 | 711 | mutex_exit(&vq->vq_lock); |
e8b96c60 MA |
712 | zio_vdev_io_bypass(zio); |
713 | zio_execute(zio); | |
45d1cae3 BB |
714 | mutex_enter(&vq->vq_lock); |
715 | goto again; | |
716 | } | |
717 | ||
e8b96c60 MA |
718 | vdev_queue_pending_add(vq, zio); |
719 | vq->vq_last_offset = zio->io_offset; | |
34dc7c2f | 720 | |
e8b96c60 | 721 | return (zio); |
34dc7c2f BB |
722 | } |
723 | ||
724 | zio_t * | |
725 | vdev_queue_io(zio_t *zio) | |
726 | { | |
727 | vdev_queue_t *vq = &zio->io_vd->vdev_queue; | |
728 | zio_t *nio; | |
729 | ||
34dc7c2f BB |
730 | if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) |
731 | return (zio); | |
732 | ||
e8b96c60 MA |
733 | /* |
734 | * Children i/os inherent their parent's priority, which might | |
735 | * not match the child's i/o type. Fix it up here. | |
736 | */ | |
737 | if (zio->io_type == ZIO_TYPE_READ) { | |
738 | if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && | |
739 | zio->io_priority != ZIO_PRIORITY_ASYNC_READ && | |
740 | zio->io_priority != ZIO_PRIORITY_SCRUB) | |
741 | zio->io_priority = ZIO_PRIORITY_ASYNC_READ; | |
742 | } else { | |
743 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
744 | if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && | |
745 | zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE) | |
746 | zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; | |
747 | } | |
34dc7c2f | 748 | |
e8b96c60 | 749 | zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; |
34dc7c2f BB |
750 | |
751 | mutex_enter(&vq->vq_lock); | |
cb682a17 | 752 | zio->io_timestamp = gethrtime(); |
34dc7c2f | 753 | vdev_queue_io_add(vq, zio); |
e8b96c60 | 754 | nio = vdev_queue_io_to_issue(vq); |
34dc7c2f BB |
755 | mutex_exit(&vq->vq_lock); |
756 | ||
757 | if (nio == NULL) | |
758 | return (NULL); | |
759 | ||
760 | if (nio->io_done == vdev_queue_agg_io_done) { | |
761 | zio_nowait(nio); | |
762 | return (NULL); | |
763 | } | |
764 | ||
765 | return (nio); | |
766 | } | |
767 | ||
768 | void | |
769 | vdev_queue_io_done(zio_t *zio) | |
770 | { | |
771 | vdev_queue_t *vq = &zio->io_vd->vdev_queue; | |
e8b96c60 | 772 | zio_t *nio; |
34dc7c2f BB |
773 | |
774 | mutex_enter(&vq->vq_lock); | |
775 | ||
330847ff | 776 | vdev_queue_pending_remove(vq, zio); |
34dc7c2f | 777 | |
cb682a17 MA |
778 | zio->io_delta = gethrtime() - zio->io_timestamp; |
779 | vq->vq_io_complete_ts = gethrtime(); | |
cc92e9d0 GW |
780 | vq->vq_io_delta_ts = vq->vq_io_complete_ts - zio->io_timestamp; |
781 | ||
e8b96c60 | 782 | while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { |
34dc7c2f BB |
783 | mutex_exit(&vq->vq_lock); |
784 | if (nio->io_done == vdev_queue_agg_io_done) { | |
785 | zio_nowait(nio); | |
786 | } else { | |
787 | zio_vdev_io_reissue(nio); | |
788 | zio_execute(nio); | |
789 | } | |
790 | mutex_enter(&vq->vq_lock); | |
791 | } | |
792 | ||
793 | mutex_exit(&vq->vq_lock); | |
794 | } | |
c28b2279 | 795 | |
9f500936 | 796 | /* |
797 | * As these three methods are only used for load calculations we're not | |
798 | * concerned if we get an incorrect value on 32bit platforms due to lack of | |
799 | * vq_lock mutex use here, instead we prefer to keep it lock free for | |
800 | * performance. | |
801 | */ | |
802 | int | |
803 | vdev_queue_length(vdev_t *vd) | |
804 | { | |
805 | return (avl_numnodes(&vd->vdev_queue.vq_active_tree)); | |
806 | } | |
807 | ||
808 | uint64_t | |
809 | vdev_queue_lastoffset(vdev_t *vd) | |
810 | { | |
811 | return (vd->vdev_queue.vq_lastoffset); | |
812 | } | |
813 | ||
814 | void | |
815 | vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio) | |
816 | { | |
817 | vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size; | |
818 | } | |
819 | ||
c28b2279 | 820 | #if defined(_KERNEL) && defined(HAVE_SPL) |
c28b2279 | 821 | module_param(zfs_vdev_aggregation_limit, int, 0644); |
c409e464 BB |
822 | MODULE_PARM_DESC(zfs_vdev_aggregation_limit, "Max vdev I/O aggregation size"); |
823 | ||
c409e464 BB |
824 | module_param(zfs_vdev_read_gap_limit, int, 0644); |
825 | MODULE_PARM_DESC(zfs_vdev_read_gap_limit, "Aggregate read I/O over gap"); | |
826 | ||
827 | module_param(zfs_vdev_write_gap_limit, int, 0644); | |
828 | MODULE_PARM_DESC(zfs_vdev_write_gap_limit, "Aggregate write I/O over gap"); | |
e8b96c60 MA |
829 | |
830 | module_param(zfs_vdev_max_active, int, 0644); | |
831 | MODULE_PARM_DESC(zfs_vdev_max_active, "Maximum number of active I/Os per vdev"); | |
832 | ||
833 | module_param(zfs_vdev_async_write_active_max_dirty_percent, int, 0644); | |
834 | MODULE_PARM_DESC(zfs_vdev_async_write_active_max_dirty_percent, | |
d1d7e268 | 835 | "Async write concurrency max threshold"); |
e8b96c60 MA |
836 | |
837 | module_param(zfs_vdev_async_write_active_min_dirty_percent, int, 0644); | |
838 | MODULE_PARM_DESC(zfs_vdev_async_write_active_min_dirty_percent, | |
d1d7e268 | 839 | "Async write concurrency min threshold"); |
e8b96c60 MA |
840 | |
841 | module_param(zfs_vdev_async_read_max_active, int, 0644); | |
842 | MODULE_PARM_DESC(zfs_vdev_async_read_max_active, | |
d1d7e268 | 843 | "Max active async read I/Os per vdev"); |
e8b96c60 MA |
844 | |
845 | module_param(zfs_vdev_async_read_min_active, int, 0644); | |
846 | MODULE_PARM_DESC(zfs_vdev_async_read_min_active, | |
d1d7e268 | 847 | "Min active async read I/Os per vdev"); |
e8b96c60 MA |
848 | |
849 | module_param(zfs_vdev_async_write_max_active, int, 0644); | |
850 | MODULE_PARM_DESC(zfs_vdev_async_write_max_active, | |
d1d7e268 | 851 | "Max active async write I/Os per vdev"); |
e8b96c60 MA |
852 | |
853 | module_param(zfs_vdev_async_write_min_active, int, 0644); | |
854 | MODULE_PARM_DESC(zfs_vdev_async_write_min_active, | |
d1d7e268 | 855 | "Min active async write I/Os per vdev"); |
e8b96c60 MA |
856 | |
857 | module_param(zfs_vdev_scrub_max_active, int, 0644); | |
858 | MODULE_PARM_DESC(zfs_vdev_scrub_max_active, "Max active scrub I/Os per vdev"); | |
859 | ||
860 | module_param(zfs_vdev_scrub_min_active, int, 0644); | |
861 | MODULE_PARM_DESC(zfs_vdev_scrub_min_active, "Min active scrub I/Os per vdev"); | |
862 | ||
863 | module_param(zfs_vdev_sync_read_max_active, int, 0644); | |
864 | MODULE_PARM_DESC(zfs_vdev_sync_read_max_active, | |
d1d7e268 | 865 | "Max active sync read I/Os per vdev"); |
e8b96c60 MA |
866 | |
867 | module_param(zfs_vdev_sync_read_min_active, int, 0644); | |
868 | MODULE_PARM_DESC(zfs_vdev_sync_read_min_active, | |
d1d7e268 | 869 | "Min active sync read I/Os per vdev"); |
e8b96c60 MA |
870 | |
871 | module_param(zfs_vdev_sync_write_max_active, int, 0644); | |
872 | MODULE_PARM_DESC(zfs_vdev_sync_write_max_active, | |
d1d7e268 | 873 | "Max active sync write I/Os per vdev"); |
e8b96c60 MA |
874 | |
875 | module_param(zfs_vdev_sync_write_min_active, int, 0644); | |
876 | MODULE_PARM_DESC(zfs_vdev_sync_write_min_active, | |
3757bff3 | 877 | "Min active sync write I/Os per vdev"); |
3dfb57a3 DB |
878 | |
879 | module_param(zfs_vdev_queue_depth_pct, int, 0644); | |
880 | MODULE_PARM_DESC(zfs_vdev_queue_depth_pct, | |
881 | "Queue depth percentage for each top-level vdev"); | |
c28b2279 | 882 | #endif |