1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3 * This file and its contents are supplied under the terms of the
4 * Common Development and Distribution License ("CDDL"), version 1.0.
5 * You may only use this file in accordance with the terms of version
6 * 1.0 of the CDDL.
7 *
8 * A full copy of the text of the CDDL should have accompanied this
9 * source. A copy of the CDDL is also available via the Internet at
10 * https://opensource.org/license/CDDL-1.0.
11 */
12 /*
13 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
14 * Use is subject to license terms.
15 */
16
17 /*
18 * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
19 */
20
21 #include <sys/zfs_context.h>
22 #include <sys/vdev_impl.h>
23 #include <sys/spa_impl.h>
24 #include <sys/zio.h>
25 #include <sys/avl.h>
26 #include <sys/dsl_pool.h>
27 #include <sys/metaslab_impl.h>
28 #include <sys/spa.h>
29 #include <sys/abd.h>
30
31 /*
32 * ZFS I/O Scheduler
33 * ---------------
34 *
35 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The
36 * I/O scheduler determines when and in what order those operations are
37 * issued. The I/O scheduler divides operations into five I/O classes
38 * prioritized in the following order: sync read, sync write, async read,
39 * async write, and scrub/resilver. Each queue defines the minimum and
40 * maximum number of concurrent operations that may be issued to the device.
41 * In addition, the device has an aggregate maximum. Note that the sum of the
42 * per-queue minimums must not exceed the aggregate maximum. If the
43 * sum of the per-queue maximums exceeds the aggregate maximum, then the
44 * number of active i/os may reach zfs_vdev_max_active, in which case no
45 * further i/os will be issued regardless of whether all per-queue
46 * minimums have been met.
47 *
48 * For many physical devices, throughput increases with the number of
49 * concurrent operations, but latency typically suffers. Further, physical
50 * devices typically have a limit at which more concurrent operations have no
51 * effect on throughput or can actually cause it to decrease.
52 *
53 * The scheduler selects the next operation to issue by first looking for an
54 * I/O class whose minimum has not been satisfied. Once all are satisfied and
55 * the aggregate maximum has not been hit, the scheduler looks for classes
56 * whose maximum has not been satisfied. Iteration through the I/O classes is
57 * done in the order specified above. No further operations are issued if the
58 * aggregate maximum number of concurrent operations has been hit or if there
59 * are no operations queued for an I/O class that has not hit its maximum.
60 * Every time an i/o is queued or an operation completes, the I/O scheduler
61 * looks for new operations to issue.
62 *
63 * All I/O classes have a fixed maximum number of outstanding operations
64 * except for the async write class. Asynchronous writes represent the data
65 * that is committed to stable storage during the syncing stage for
66 * transaction groups (see txg.c). Transaction groups enter the syncing state
67 * periodically so the number of queued async writes will quickly burst up and
68 * then bleed down to zero. Rather than servicing them as quickly as possible,
69 * the I/O scheduler changes the maximum number of active async write i/os
70 * according to the amount of dirty data in the pool (see dsl_pool.c). Since
71 * both throughput and latency typically increase with the number of
72 * concurrent operations issued to physical devices, reducing the burstiness
73 * in the number of concurrent operations also stabilizes the response time of
74 * operations from other -- and in particular synchronous -- queues. In broad
75 * strokes, the I/O scheduler will issue more concurrent operations from the
76 * async write queue as there's more dirty data in the pool.
77 *
78 * Async Writes
79 *
80 * The number of concurrent operations issued for the async write I/O class
81 * follows a piece-wise linear function defined by a few adjustable points.
82 *
83 * | o---------| <-- zfs_vdev_async_write_max_active
84 * ^ | /^ |
85 * | | / | |
86 * active | / | |
87 * I/O | / | |
88 * count | / | |
89 * | / | |
90 * |------------o | | <-- zfs_vdev_async_write_min_active
91 * 0|____________^______|_________|
92 * 0% | | 100% of zfs_dirty_data_max
93 * | |
94 * | `-- zfs_vdev_async_write_active_max_dirty_percent
95 * `--------- zfs_vdev_async_write_active_min_dirty_percent
96 *
97 * Until the amount of dirty data exceeds a minimum percentage of the dirty
98 * data allowed in the pool, the I/O scheduler will limit the number of
99 * concurrent operations to the minimum. As that threshold is crossed, the
100 * number of concurrent operations issued increases linearly to the maximum at
101 * the specified maximum percentage of the dirty data allowed in the pool.
102 *
103 * Ideally, the amount of dirty data on a busy pool will stay in the sloped
104 * part of the function between zfs_vdev_async_write_active_min_dirty_percent
105 * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the
106 * maximum percentage, this indicates that the rate of incoming data is
107 * greater than the rate that the backend storage can handle. In this case, we
108 * must further throttle incoming writes (see dmu_tx_delay() for details).
109 */
110
111 /*
112 * The maximum number of i/os active to each device. Ideally, this will be >=
113 * the sum of each queue's max_active.
114 */
115 uint_t zfs_vdev_max_active = 1000;
116
117 /*
118 * Per-queue limits on the number of i/os active to each device. If the
119 * number of active i/os is < zfs_vdev_max_active, then the min_active comes
120 * into play. We will send min_active from each queue round-robin, and then
121 * send from queues in the order defined by zio_priority_t up to max_active.
122 * Some queues have additional mechanisms to limit number of active I/Os in
123 * addition to min_active and max_active, see below.
124 *
125 * In general, smaller max_active's will lead to lower latency of synchronous
126 * operations. Larger max_active's may lead to higher overall throughput,
127 * depending on underlying storage.
128 *
129 * The ratio of the queues' max_actives determines the balance of performance
130 * between reads, writes, and scrubs. E.g., increasing
131 * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete
132 * more quickly, but reads and writes to have higher latency and lower
133 * throughput.
134 */
135 static uint_t zfs_vdev_sync_read_min_active = 10;
136 static uint_t zfs_vdev_sync_read_max_active = 10;
137 static uint_t zfs_vdev_sync_write_min_active = 10;
138 static uint_t zfs_vdev_sync_write_max_active = 10;
139 static uint_t zfs_vdev_async_read_min_active = 1;
140 /* */ uint_t zfs_vdev_async_read_max_active = 3;
141 static uint_t zfs_vdev_async_write_min_active = 2;
142 static uint_t zfs_vdev_async_write_max_active = 10;
143 static uint_t zfs_vdev_scrub_min_active = 1;
144 static uint_t zfs_vdev_scrub_max_active = 3;
145 static uint_t zfs_vdev_removal_min_active = 1;
146 static uint_t zfs_vdev_removal_max_active = 2;
147 static uint_t zfs_vdev_initializing_min_active = 1;
148 static uint_t zfs_vdev_initializing_max_active = 1;
149 static uint_t zfs_vdev_trim_min_active = 1;
150 static uint_t zfs_vdev_trim_max_active = 2;
151 static uint_t zfs_vdev_rebuild_min_active = 1;
152 static uint_t zfs_vdev_rebuild_max_active = 3;
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 uint_t zfs_vdev_async_write_active_min_dirty_percent = 30;
162 uint_t zfs_vdev_async_write_active_max_dirty_percent = 60;
163
164 /*
165 * For non-interactive I/O (scrub, resilver, removal, initialize and rebuild),
166 * the number of concurrently-active I/O's is limited to *_min_active, unless
167 * the vdev is "idle". When there are no interactive I/Os active (sync or
168 * async), and zfs_vdev_nia_delay I/Os have completed since the last
169 * interactive I/O, then the vdev is considered to be "idle", and the number
170 * of concurrently-active non-interactive I/O's is increased to *_max_active.
171 */
172 static uint_t zfs_vdev_nia_delay = 5;
173
174 /*
175 * Some HDDs tend to prioritize sequential I/O so high that concurrent
176 * random I/O latency reaches several seconds. On some HDDs it happens
177 * even if sequential I/Os are submitted one at a time, and so setting
178 * *_max_active to 1 does not help. To prevent non-interactive I/Os, like
179 * scrub, from monopolizing the device no more than zfs_vdev_nia_credit
180 * I/Os can be sent while there are outstanding incomplete interactive
181 * I/Os. This enforced wait ensures the HDD services the interactive I/O
182 * within a reasonable amount of time.
183 */
184 static uint_t zfs_vdev_nia_credit = 5;
185
186 /*
187 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
188 * For read I/Os, we also aggregate across small adjacency gaps; for writes
189 * we include spans of optional I/Os to aid aggregation at the disk even when
190 * they aren't able to help us aggregate at this level.
191 */
192 static uint_t zfs_vdev_aggregation_limit = 1 << 20;
193 static uint_t zfs_vdev_aggregation_limit_non_rotating = SPA_OLD_MAXBLOCKSIZE;
194 static uint_t zfs_vdev_read_gap_limit = 32 << 10;
195 static uint_t zfs_vdev_write_gap_limit = 4 << 10;
196
197 static int
vdev_queue_offset_compare(const void * x1,const void * x2)198 vdev_queue_offset_compare(const void *x1, const void *x2)
199 {
200 const zio_t *z1 = (const zio_t *)x1;
201 const zio_t *z2 = (const zio_t *)x2;
202
203 int cmp = TREE_CMP(z1->io_offset, z2->io_offset);
204
205 if (likely(cmp))
206 return (cmp);
207
208 return (TREE_PCMP(z1, z2));
209 }
210
211 #define VDQ_T_SHIFT 29
212
213 static int
vdev_queue_to_compare(const void * x1,const void * x2)214 vdev_queue_to_compare(const void *x1, const void *x2)
215 {
216 const zio_t *z1 = (const zio_t *)x1;
217 const zio_t *z2 = (const zio_t *)x2;
218
219 int cmp = TREE_CMP(z1->io_timestamp >> VDQ_T_SHIFT,
220 z2->io_timestamp >> VDQ_T_SHIFT);
221 if (cmp == 0)
222 cmp = TREE_CMP(z1->io_offset, z2->io_offset);
223
224 if (likely(cmp | (z1->io_queue_state == ZIO_QS_NONE)))
225 return (cmp);
226
227 return (TREE_PCMP(z1, z2));
228 }
229
230 static inline boolean_t
vdev_queue_class_fifo(zio_priority_t p)231 vdev_queue_class_fifo(zio_priority_t p)
232 {
233 return (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE ||
234 p == ZIO_PRIORITY_TRIM);
235 }
236
237 static void
vdev_queue_class_add(vdev_queue_t * vq,zio_t * zio)238 vdev_queue_class_add(vdev_queue_t *vq, zio_t *zio)
239 {
240 zio_priority_t p = zio->io_priority;
241 vq->vq_cqueued |= 1U << p;
242 if (vdev_queue_class_fifo(p)) {
243 list_insert_tail(&vq->vq_class[p].vqc_list, zio);
244 vq->vq_class[p].vqc_list_numnodes++;
245 }
246 else
247 avl_add(&vq->vq_class[p].vqc_tree, zio);
248 }
249
250 static void
vdev_queue_class_remove(vdev_queue_t * vq,zio_t * zio)251 vdev_queue_class_remove(vdev_queue_t *vq, zio_t *zio)
252 {
253 zio_priority_t p = zio->io_priority;
254 uint32_t empty;
255 if (vdev_queue_class_fifo(p)) {
256 list_t *list = &vq->vq_class[p].vqc_list;
257 list_remove(list, zio);
258 empty = list_is_empty(list);
259 vq->vq_class[p].vqc_list_numnodes--;
260 } else {
261 avl_tree_t *tree = &vq->vq_class[p].vqc_tree;
262 avl_remove(tree, zio);
263 empty = avl_is_empty(tree);
264 }
265 vq->vq_cqueued &= ~(empty << p);
266 }
267
268 static uint_t
vdev_queue_class_min_active(vdev_queue_t * vq,zio_priority_t p)269 vdev_queue_class_min_active(vdev_queue_t *vq, zio_priority_t p)
270 {
271 switch (p) {
272 case ZIO_PRIORITY_SYNC_READ:
273 return (zfs_vdev_sync_read_min_active);
274 case ZIO_PRIORITY_SYNC_WRITE:
275 return (zfs_vdev_sync_write_min_active);
276 case ZIO_PRIORITY_ASYNC_READ:
277 return (zfs_vdev_async_read_min_active);
278 case ZIO_PRIORITY_ASYNC_WRITE:
279 return (zfs_vdev_async_write_min_active);
280 case ZIO_PRIORITY_SCRUB:
281 return (vq->vq_ia_active == 0 ? zfs_vdev_scrub_min_active :
282 MIN(vq->vq_nia_credit, zfs_vdev_scrub_min_active));
283 case ZIO_PRIORITY_REMOVAL:
284 return (vq->vq_ia_active == 0 ? zfs_vdev_removal_min_active :
285 MIN(vq->vq_nia_credit, zfs_vdev_removal_min_active));
286 case ZIO_PRIORITY_INITIALIZING:
287 return (vq->vq_ia_active == 0 ?zfs_vdev_initializing_min_active:
288 MIN(vq->vq_nia_credit, zfs_vdev_initializing_min_active));
289 case ZIO_PRIORITY_TRIM:
290 return (zfs_vdev_trim_min_active);
291 case ZIO_PRIORITY_REBUILD:
292 return (vq->vq_ia_active == 0 ? zfs_vdev_rebuild_min_active :
293 MIN(vq->vq_nia_credit, zfs_vdev_rebuild_min_active));
294 default:
295 panic("invalid priority %u", p);
296 return (0);
297 }
298 }
299
300 static uint_t
vdev_queue_max_async_writes(spa_t * spa)301 vdev_queue_max_async_writes(spa_t *spa)
302 {
303 uint_t writes;
304 uint64_t dirty = 0;
305 dsl_pool_t *dp = spa_get_dsl(spa);
306 uint64_t min_bytes = zfs_dirty_data_max *
307 zfs_vdev_async_write_active_min_dirty_percent / 100;
308 uint64_t max_bytes = zfs_dirty_data_max *
309 zfs_vdev_async_write_active_max_dirty_percent / 100;
310
311 /*
312 * Async writes may occur before the assignment of the spa's
313 * dsl_pool_t if a self-healing zio is issued prior to the
314 * completion of dmu_objset_open_impl().
315 */
316 if (dp == NULL)
317 return (zfs_vdev_async_write_max_active);
318
319 /*
320 * Sync tasks correspond to interactive user actions. To reduce the
321 * execution time of those actions we push data out as fast as possible.
322 */
323 dirty = dp->dp_dirty_total;
324 if (dirty > max_bytes || spa_has_pending_synctask(spa))
325 return (zfs_vdev_async_write_max_active);
326
327 if (dirty < min_bytes)
328 return (zfs_vdev_async_write_min_active);
329
330 /*
331 * linear interpolation:
332 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
333 * move right by min_bytes
334 * move up by min_writes
335 */
336 writes = (dirty - min_bytes) *
337 (zfs_vdev_async_write_max_active -
338 zfs_vdev_async_write_min_active) /
339 (max_bytes - min_bytes) +
340 zfs_vdev_async_write_min_active;
341 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
342 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
343 return (writes);
344 }
345
346 static uint_t
vdev_queue_class_max_active(vdev_queue_t * vq,zio_priority_t p)347 vdev_queue_class_max_active(vdev_queue_t *vq, zio_priority_t p)
348 {
349 switch (p) {
350 case ZIO_PRIORITY_SYNC_READ:
351 return (zfs_vdev_sync_read_max_active);
352 case ZIO_PRIORITY_SYNC_WRITE:
353 return (zfs_vdev_sync_write_max_active);
354 case ZIO_PRIORITY_ASYNC_READ:
355 return (zfs_vdev_async_read_max_active);
356 case ZIO_PRIORITY_ASYNC_WRITE:
357 return (vdev_queue_max_async_writes(vq->vq_vdev->vdev_spa));
358 case ZIO_PRIORITY_SCRUB:
359 if (vq->vq_ia_active > 0) {
360 return (MIN(vq->vq_nia_credit,
361 zfs_vdev_scrub_min_active));
362 } else if (vq->vq_nia_credit < zfs_vdev_nia_delay)
363 return (MAX(1, zfs_vdev_scrub_min_active));
364 return (zfs_vdev_scrub_max_active);
365 case ZIO_PRIORITY_REMOVAL:
366 if (vq->vq_ia_active > 0) {
367 return (MIN(vq->vq_nia_credit,
368 zfs_vdev_removal_min_active));
369 } else if (vq->vq_nia_credit < zfs_vdev_nia_delay)
370 return (MAX(1, zfs_vdev_removal_min_active));
371 return (zfs_vdev_removal_max_active);
372 case ZIO_PRIORITY_INITIALIZING:
373 if (vq->vq_ia_active > 0) {
374 return (MIN(vq->vq_nia_credit,
375 zfs_vdev_initializing_min_active));
376 } else if (vq->vq_nia_credit < zfs_vdev_nia_delay)
377 return (MAX(1, zfs_vdev_initializing_min_active));
378 return (zfs_vdev_initializing_max_active);
379 case ZIO_PRIORITY_TRIM:
380 return (zfs_vdev_trim_max_active);
381 case ZIO_PRIORITY_REBUILD:
382 if (vq->vq_ia_active > 0) {
383 return (MIN(vq->vq_nia_credit,
384 zfs_vdev_rebuild_min_active));
385 } else if (vq->vq_nia_credit < zfs_vdev_nia_delay)
386 return (MAX(1, zfs_vdev_rebuild_min_active));
387 return (zfs_vdev_rebuild_max_active);
388 default:
389 panic("invalid priority %u", p);
390 return (0);
391 }
392 }
393
394 /*
395 * Return the i/o class to issue from, or ZIO_PRIORITY_NUM_QUEUEABLE if
396 * there is no eligible class.
397 */
398 static zio_priority_t
vdev_queue_class_to_issue(vdev_queue_t * vq)399 vdev_queue_class_to_issue(vdev_queue_t *vq)
400 {
401 uint32_t cq = vq->vq_cqueued;
402 zio_priority_t p, p1;
403
404 if (cq == 0 || vq->vq_active >= zfs_vdev_max_active)
405 return (ZIO_PRIORITY_NUM_QUEUEABLE);
406
407 /*
408 * Find a queue that has not reached its minimum # outstanding i/os.
409 * Do round-robin to reduce starvation due to zfs_vdev_max_active
410 * and vq_nia_credit limits.
411 */
412 p1 = vq->vq_last_prio + 1;
413 if (p1 >= ZIO_PRIORITY_NUM_QUEUEABLE)
414 p1 = 0;
415 for (p = p1; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
416 if ((cq & (1U << p)) != 0 && vq->vq_cactive[p] <
417 vdev_queue_class_min_active(vq, p))
418 goto found;
419 }
420 for (p = 0; p < p1; p++) {
421 if ((cq & (1U << p)) != 0 && vq->vq_cactive[p] <
422 vdev_queue_class_min_active(vq, p))
423 goto found;
424 }
425
426 /*
427 * If we haven't found a queue, look for one that hasn't reached its
428 * maximum # outstanding i/os.
429 */
430 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
431 if ((cq & (1U << p)) != 0 && vq->vq_cactive[p] <
432 vdev_queue_class_max_active(vq, p))
433 break;
434 }
435
436 found:
437 vq->vq_last_prio = p;
438 return (p);
439 }
440
441 void
vdev_queue_init(vdev_t * vd)442 vdev_queue_init(vdev_t *vd)
443 {
444 vdev_queue_t *vq = &vd->vdev_queue;
445 zio_priority_t p;
446
447 vq->vq_vdev = vd;
448
449 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
450 if (vdev_queue_class_fifo(p)) {
451 list_create(&vq->vq_class[p].vqc_list,
452 sizeof (zio_t),
453 offsetof(struct zio, io_queue_node.l));
454 } else {
455 avl_create(&vq->vq_class[p].vqc_tree,
456 vdev_queue_to_compare, sizeof (zio_t),
457 offsetof(struct zio, io_queue_node.a));
458 }
459 }
460 avl_create(&vq->vq_read_offset_tree,
461 vdev_queue_offset_compare, sizeof (zio_t),
462 offsetof(struct zio, io_offset_node));
463 avl_create(&vq->vq_write_offset_tree,
464 vdev_queue_offset_compare, sizeof (zio_t),
465 offsetof(struct zio, io_offset_node));
466
467 vq->vq_last_offset = 0;
468 list_create(&vq->vq_active_list, sizeof (struct zio),
469 offsetof(struct zio, io_queue_node.l));
470 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
471 }
472
473 void
vdev_queue_fini(vdev_t * vd)474 vdev_queue_fini(vdev_t *vd)
475 {
476 vdev_queue_t *vq = &vd->vdev_queue;
477
478 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
479 if (vdev_queue_class_fifo(p))
480 list_destroy(&vq->vq_class[p].vqc_list);
481 else
482 avl_destroy(&vq->vq_class[p].vqc_tree);
483 }
484 avl_destroy(&vq->vq_read_offset_tree);
485 avl_destroy(&vq->vq_write_offset_tree);
486
487 list_destroy(&vq->vq_active_list);
488 mutex_destroy(&vq->vq_lock);
489 }
490
491 static void
vdev_queue_io_add(vdev_queue_t * vq,zio_t * zio)492 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
493 {
494 zio->io_queue_state = ZIO_QS_QUEUED;
495 vdev_queue_class_add(vq, zio);
496 if (zio->io_type == ZIO_TYPE_READ)
497 avl_add(&vq->vq_read_offset_tree, zio);
498 else if (zio->io_type == ZIO_TYPE_WRITE)
499 avl_add(&vq->vq_write_offset_tree, zio);
500 }
501
502 static void
vdev_queue_io_remove(vdev_queue_t * vq,zio_t * zio)503 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
504 {
505 vdev_queue_class_remove(vq, zio);
506 if (zio->io_type == ZIO_TYPE_READ)
507 avl_remove(&vq->vq_read_offset_tree, zio);
508 else if (zio->io_type == ZIO_TYPE_WRITE)
509 avl_remove(&vq->vq_write_offset_tree, zio);
510 zio->io_queue_state = ZIO_QS_NONE;
511 }
512
513 static boolean_t
vdev_queue_is_interactive(zio_priority_t p)514 vdev_queue_is_interactive(zio_priority_t p)
515 {
516 switch (p) {
517 case ZIO_PRIORITY_SCRUB:
518 case ZIO_PRIORITY_REMOVAL:
519 case ZIO_PRIORITY_INITIALIZING:
520 case ZIO_PRIORITY_REBUILD:
521 return (B_FALSE);
522 default:
523 return (B_TRUE);
524 }
525 }
526
527 static void
vdev_queue_pending_add(vdev_queue_t * vq,zio_t * zio)528 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
529 {
530 ASSERT(MUTEX_HELD(&vq->vq_lock));
531 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
532 vq->vq_cactive[zio->io_priority]++;
533 vq->vq_active++;
534 if (vdev_queue_is_interactive(zio->io_priority)) {
535 if (++vq->vq_ia_active == 1)
536 vq->vq_nia_credit = 1;
537 } else if (vq->vq_ia_active > 0) {
538 vq->vq_nia_credit--;
539 }
540 zio->io_queue_state = ZIO_QS_ACTIVE;
541 list_insert_tail(&vq->vq_active_list, zio);
542 }
543
544 static void
vdev_queue_pending_remove(vdev_queue_t * vq,zio_t * zio)545 vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
546 {
547 ASSERT(MUTEX_HELD(&vq->vq_lock));
548 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
549 vq->vq_cactive[zio->io_priority]--;
550 vq->vq_active--;
551 if (vdev_queue_is_interactive(zio->io_priority)) {
552 if (--vq->vq_ia_active == 0)
553 vq->vq_nia_credit = 0;
554 else
555 vq->vq_nia_credit = zfs_vdev_nia_credit;
556 } else if (vq->vq_ia_active == 0)
557 vq->vq_nia_credit++;
558 list_remove(&vq->vq_active_list, zio);
559 zio->io_queue_state = ZIO_QS_NONE;
560 }
561
562 static void
vdev_queue_agg_io_done(zio_t * aio)563 vdev_queue_agg_io_done(zio_t *aio)
564 {
565 abd_free(aio->io_abd);
566 }
567
568 /*
569 * Compute the range spanned by two i/os, which is the endpoint of the last
570 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
571 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
572 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
573 */
574 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
575 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
576
577 /*
578 * Sufficiently adjacent io_offset's in ZIOs will be aggregated. We do this
579 * by creating a gang ABD from the adjacent ZIOs io_abd's. By using
580 * a gang ABD we avoid doing memory copies to and from the parent,
581 * child ZIOs. The gang ABD also accounts for gaps between adjacent
582 * io_offsets by simply getting the zero ABD for writes or allocating
583 * a new ABD for reads and placing them in the gang ABD as well.
584 */
585 static zio_t *
vdev_queue_aggregate(vdev_queue_t * vq,zio_t * zio)586 vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
587 {
588 zio_t *first, *last, *aio, *dio, *mandatory, *nio;
589 uint64_t maxgap = 0;
590 uint64_t size;
591 uint64_t limit;
592 boolean_t stretch = B_FALSE;
593 uint64_t next_offset;
594 abd_t *abd;
595 avl_tree_t *t;
596
597 /*
598 * TRIM aggregation should not be needed since code in zfs_trim.c can
599 * submit TRIM I/O for extents up to zfs_trim_extent_bytes_max (128M).
600 */
601 if (zio->io_type == ZIO_TYPE_TRIM)
602 return (NULL);
603
604 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
605 return (NULL);
606
607 if (vq->vq_vdev->vdev_nonrot)
608 limit = zfs_vdev_aggregation_limit_non_rotating;
609 else
610 limit = zfs_vdev_aggregation_limit;
611 if (limit == 0)
612 return (NULL);
613 limit = MIN(limit, SPA_MAXBLOCKSIZE);
614
615 /*
616 * I/Os to distributed spares are directly dispatched to the dRAID
617 * leaf vdevs for aggregation. See the comment at the end of the
618 * zio_vdev_io_start() function.
619 */
620 ASSERT(vq->vq_vdev->vdev_ops != &vdev_draid_spare_ops);
621
622 first = last = zio;
623
624 if (zio->io_type == ZIO_TYPE_READ) {
625 maxgap = zfs_vdev_read_gap_limit;
626 t = &vq->vq_read_offset_tree;
627 } else {
628 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
629 t = &vq->vq_write_offset_tree;
630 }
631
632 /*
633 * We can aggregate I/Os that are sufficiently adjacent and of
634 * the same flavor, as expressed by the AGG_INHERIT flags.
635 * The latter requirement is necessary so that certain
636 * attributes of the I/O, such as whether it's a normal I/O
637 * or a scrub/resilver, can be preserved in the aggregate.
638 * We can include optional I/Os, but don't allow them
639 * to begin a range as they add no benefit in that situation.
640 */
641
642 /*
643 * We keep track of the last non-optional I/O.
644 */
645 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
646
647 /*
648 * Walk backwards through sufficiently contiguous I/Os
649 * recording the last non-optional I/O.
650 */
651 zio_flag_t flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
652 while ((dio = AVL_PREV(t, first)) != NULL &&
653 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
654 IO_SPAN(dio, last) <= limit &&
655 IO_GAP(dio, first) <= maxgap &&
656 dio->io_type == zio->io_type) {
657 first = dio;
658 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
659 mandatory = first;
660 }
661
662 /*
663 * Skip any initial optional I/Os.
664 */
665 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
666 first = AVL_NEXT(t, first);
667 ASSERT(first != NULL);
668 }
669
670
671 /*
672 * Walk forward through sufficiently contiguous I/Os.
673 * The aggregation limit does not apply to optional i/os, so that
674 * we can issue contiguous writes even if they are larger than the
675 * aggregation limit.
676 */
677 while ((dio = AVL_NEXT(t, last)) != NULL &&
678 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
679 (IO_SPAN(first, dio) <= limit ||
680 (dio->io_flags & ZIO_FLAG_OPTIONAL)) &&
681 IO_SPAN(first, dio) <= SPA_MAXBLOCKSIZE &&
682 IO_GAP(last, dio) <= maxgap &&
683 dio->io_type == zio->io_type) {
684 last = dio;
685 if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
686 mandatory = last;
687 }
688
689 /*
690 * Now that we've established the range of the I/O aggregation
691 * we must decide what to do with trailing optional I/Os.
692 * For reads, there's nothing to do. While we are unable to
693 * aggregate further, it's possible that a trailing optional
694 * I/O would allow the underlying device to aggregate with
695 * subsequent I/Os. We must therefore determine if the next
696 * non-optional I/O is close enough to make aggregation
697 * worthwhile.
698 */
699 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
700 zio_t *nio = last;
701 while ((dio = AVL_NEXT(t, nio)) != NULL &&
702 IO_GAP(nio, dio) == 0 &&
703 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
704 nio = dio;
705 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
706 stretch = B_TRUE;
707 break;
708 }
709 }
710 }
711
712 if (stretch) {
713 /*
714 * We are going to include an optional io in our aggregated
715 * span, thus closing the write gap. Only mandatory i/os can
716 * start aggregated spans, so make sure that the next i/o
717 * after our span is mandatory.
718 */
719 dio = AVL_NEXT(t, last);
720 ASSERT3P(dio, !=, NULL);
721 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
722 } else {
723 /* do not include the optional i/o */
724 while (last != mandatory && last != first) {
725 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
726 last = AVL_PREV(t, last);
727 ASSERT(last != NULL);
728 }
729 }
730
731 if (first == last)
732 return (NULL);
733
734 size = IO_SPAN(first, last);
735 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
736
737 abd = abd_alloc_gang();
738 if (abd == NULL)
739 return (NULL);
740
741 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
742 abd, size, first->io_type, zio->io_priority,
743 flags | ZIO_FLAG_DONT_QUEUE, vdev_queue_agg_io_done, NULL);
744 aio->io_timestamp = first->io_timestamp;
745
746 nio = first;
747 next_offset = first->io_offset;
748 do {
749 dio = nio;
750 nio = AVL_NEXT(t, dio);
751 ASSERT3P(dio, !=, NULL);
752 zio_add_child(dio, aio);
753 vdev_queue_io_remove(vq, dio);
754
755 if (dio->io_offset != next_offset) {
756 /* allocate a buffer for a read gap */
757 ASSERT3U(dio->io_type, ==, ZIO_TYPE_READ);
758 ASSERT3U(dio->io_offset, >, next_offset);
759 abd = abd_alloc_for_io(
760 dio->io_offset - next_offset, B_TRUE);
761 abd_gang_add(aio->io_abd, abd, B_TRUE);
762 }
763 if (dio->io_abd &&
764 (dio->io_size != abd_get_size(dio->io_abd))) {
765 /* abd size not the same as IO size */
766 ASSERT3U(abd_get_size(dio->io_abd), >, dio->io_size);
767 abd = abd_get_offset_size(dio->io_abd, 0, dio->io_size);
768 abd_gang_add(aio->io_abd, abd, B_TRUE);
769 } else {
770 if (dio->io_flags & ZIO_FLAG_NODATA) {
771 /* allocate a buffer for a write gap */
772 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
773 ASSERT0P(dio->io_abd);
774 abd_gang_add(aio->io_abd,
775 abd_get_zeros(dio->io_size), B_TRUE);
776 } else {
777 /*
778 * We pass B_FALSE to abd_gang_add()
779 * because we did not allocate a new
780 * ABD, so it is assumed the caller
781 * will free this ABD.
782 */
783 abd_gang_add(aio->io_abd, dio->io_abd,
784 B_FALSE);
785 }
786 }
787 next_offset = dio->io_offset + dio->io_size;
788 } while (dio != last);
789 ASSERT3U(abd_get_size(aio->io_abd), ==, aio->io_size);
790
791 /*
792 * Callers must call zio_vdev_io_bypass() and zio_execute() for
793 * aggregated (parent) I/Os so that we could avoid dropping the
794 * queue's lock here to avoid a deadlock that we could encounter
795 * due to lock order reversal between vq_lock and io_lock in
796 * zio_change_priority().
797 */
798 return (aio);
799 }
800
801 static zio_t *
vdev_queue_io_to_issue(vdev_queue_t * vq)802 vdev_queue_io_to_issue(vdev_queue_t *vq)
803 {
804 zio_t *zio, *aio;
805 zio_priority_t p;
806 avl_index_t idx;
807 avl_tree_t *tree;
808
809 again:
810 ASSERT(MUTEX_HELD(&vq->vq_lock));
811
812 p = vdev_queue_class_to_issue(vq);
813
814 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
815 /* No eligible queued i/os */
816 return (NULL);
817 }
818
819 if (vdev_queue_class_fifo(p)) {
820 zio = list_head(&vq->vq_class[p].vqc_list);
821 } else {
822 /*
823 * For LBA-ordered queues (async / scrub / initializing),
824 * issue the I/O which follows the most recently issued I/O
825 * in LBA (offset) order, but to avoid starvation only within
826 * the same 0.5 second interval as the first I/O.
827 */
828 tree = &vq->vq_class[p].vqc_tree;
829 zio = aio = avl_first(tree);
830 if (zio->io_offset < vq->vq_last_offset) {
831 vq->vq_io_search.io_timestamp = zio->io_timestamp;
832 vq->vq_io_search.io_offset = vq->vq_last_offset;
833 zio = avl_find(tree, &vq->vq_io_search, &idx);
834 if (zio == NULL) {
835 zio = avl_nearest(tree, idx, AVL_AFTER);
836 if (zio == NULL ||
837 (zio->io_timestamp >> VDQ_T_SHIFT) !=
838 (aio->io_timestamp >> VDQ_T_SHIFT))
839 zio = aio;
840 }
841 }
842 }
843 ASSERT3U(zio->io_priority, ==, p);
844
845 aio = vdev_queue_aggregate(vq, zio);
846 if (aio != NULL) {
847 zio = aio;
848 } else {
849 vdev_queue_io_remove(vq, zio);
850
851 /*
852 * If the I/O is or was optional and therefore has no data, we
853 * need to simply discard it. We need to drop the vdev queue's
854 * lock to avoid a deadlock that we could encounter since this
855 * I/O will complete immediately.
856 */
857 if (zio->io_flags & ZIO_FLAG_NODATA) {
858 mutex_exit(&vq->vq_lock);
859 zio_vdev_io_bypass(zio);
860 zio_execute(zio);
861 mutex_enter(&vq->vq_lock);
862 goto again;
863 }
864 }
865
866 vdev_queue_pending_add(vq, zio);
867 vq->vq_last_offset = zio->io_offset + zio->io_size;
868
869 return (zio);
870 }
871
872 static boolean_t
vdev_should_queue_io(zio_t * zio)873 vdev_should_queue_io(zio_t *zio)
874 {
875 vdev_t *vd = zio->io_vd;
876 boolean_t should_queue = B_TRUE;
877
878 /*
879 * Add zio with ZIO_FLAG_NODATA to queue as bypass code
880 * currently does not handle certain cases (gang abd, raidz
881 * write aggregation).
882 */
883 if (zio->io_flags & ZIO_FLAG_NODATA)
884 return (B_TRUE);
885
886 switch (vd->vdev_scheduler) {
887 case VDEV_SCHEDULER_AUTO:
888 if (vd->vdev_nonrot && vd->vdev_is_blkdev)
889 should_queue = B_FALSE;
890 break;
891 case VDEV_SCHEDULER_ON:
892 should_queue = B_TRUE;
893 break;
894 case VDEV_SCHEDULER_OFF:
895 should_queue = B_FALSE;
896 break;
897 default:
898 should_queue = B_TRUE;
899 break;
900 }
901 return (should_queue);
902 }
903
904 zio_t *
vdev_queue_io(zio_t * zio)905 vdev_queue_io(zio_t *zio)
906 {
907 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
908 zio_t *dio, *nio;
909 zio_link_t *zl = NULL;
910
911 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
912 return (zio);
913
914 /*
915 * Children i/os inherent their parent's priority, which might
916 * not match the child's i/o type. Fix it up here.
917 */
918 if (zio->io_type == ZIO_TYPE_READ) {
919 ASSERT(zio->io_priority != ZIO_PRIORITY_TRIM);
920
921 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
922 zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
923 zio->io_priority != ZIO_PRIORITY_SCRUB &&
924 zio->io_priority != ZIO_PRIORITY_REMOVAL &&
925 zio->io_priority != ZIO_PRIORITY_INITIALIZING &&
926 zio->io_priority != ZIO_PRIORITY_REBUILD) {
927 zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
928 }
929 } else if (zio->io_type == ZIO_TYPE_WRITE) {
930 ASSERT(zio->io_priority != ZIO_PRIORITY_TRIM);
931
932 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
933 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE &&
934 zio->io_priority != ZIO_PRIORITY_REMOVAL &&
935 zio->io_priority != ZIO_PRIORITY_INITIALIZING &&
936 zio->io_priority != ZIO_PRIORITY_REBUILD) {
937 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
938 }
939 } else {
940 ASSERT(zio->io_type == ZIO_TYPE_TRIM);
941 ASSERT(zio->io_priority == ZIO_PRIORITY_TRIM);
942 }
943
944 zio->io_flags |= ZIO_FLAG_DONT_QUEUE;
945 zio->io_timestamp = gethrtime();
946
947 if (!vdev_should_queue_io(zio)) {
948 zio->io_queue_state = ZIO_QS_NONE;
949 zio->io_flags |= ZIO_FLAG_BYPASSED_QUEUE;
950 return (zio);
951 }
952
953 mutex_enter(&vq->vq_lock);
954 vdev_queue_io_add(vq, zio);
955 nio = vdev_queue_io_to_issue(vq);
956 mutex_exit(&vq->vq_lock);
957
958 if (nio == NULL)
959 return (NULL);
960
961 if (nio->io_done == vdev_queue_agg_io_done) {
962 while ((dio = zio_walk_parents(nio, &zl)) != NULL) {
963 ASSERT3U(dio->io_type, ==, nio->io_type);
964 zio_vdev_io_bypass(dio);
965 zio_execute(dio);
966 }
967 zio_nowait(nio);
968 return (NULL);
969 }
970
971 return (nio);
972 }
973
974 void
vdev_queue_io_done(zio_t * zio)975 vdev_queue_io_done(zio_t *zio)
976 {
977 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
978 zio_t *dio, *nio;
979 zio_link_t *zl = NULL;
980
981 hrtime_t now = gethrtime();
982 vq->vq_io_complete_ts = now;
983 vq->vq_io_delta_ts = zio->io_delta = now - zio->io_timestamp;
984
985 if (zio->io_queue_state == ZIO_QS_NONE)
986 return;
987
988 mutex_enter(&vq->vq_lock);
989 vdev_queue_pending_remove(vq, zio);
990
991 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
992 mutex_exit(&vq->vq_lock);
993 if (nio->io_done == vdev_queue_agg_io_done) {
994 while ((dio = zio_walk_parents(nio, &zl)) != NULL) {
995 ASSERT3U(dio->io_type, ==, nio->io_type);
996 zio_vdev_io_bypass(dio);
997 zio_execute(dio);
998 }
999 zio_nowait(nio);
1000 } else {
1001 zio_vdev_io_reissue(nio);
1002 zio_execute(nio);
1003 }
1004 mutex_enter(&vq->vq_lock);
1005 }
1006
1007 mutex_exit(&vq->vq_lock);
1008 }
1009
1010 void
vdev_queue_change_io_priority(zio_t * zio,zio_priority_t priority)1011 vdev_queue_change_io_priority(zio_t *zio, zio_priority_t priority)
1012 {
1013 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
1014
1015 /*
1016 * ZIO_PRIORITY_NOW is used by the vdev cache code and the aggregate zio
1017 * code to issue IOs without adding them to the vdev queue. In this
1018 * case, the zio is already going to be issued as quickly as possible
1019 * and so it doesn't need any reprioritization to help.
1020 */
1021 if (zio->io_priority == ZIO_PRIORITY_NOW)
1022 return;
1023
1024 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
1025 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
1026
1027 if (zio->io_type == ZIO_TYPE_READ) {
1028 if (priority != ZIO_PRIORITY_SYNC_READ &&
1029 priority != ZIO_PRIORITY_ASYNC_READ &&
1030 priority != ZIO_PRIORITY_SCRUB)
1031 priority = ZIO_PRIORITY_ASYNC_READ;
1032 } else {
1033 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1034 if (priority != ZIO_PRIORITY_SYNC_WRITE &&
1035 priority != ZIO_PRIORITY_ASYNC_WRITE)
1036 priority = ZIO_PRIORITY_ASYNC_WRITE;
1037 }
1038
1039 mutex_enter(&vq->vq_lock);
1040
1041 /*
1042 * If the zio is in none of the queues we can simply change
1043 * the priority. If the zio is waiting to be submitted we must
1044 * remove it from the queue and re-insert it with the new priority.
1045 * Otherwise, the zio is currently active and we cannot change its
1046 * priority.
1047 */
1048 if (zio->io_queue_state == ZIO_QS_QUEUED) {
1049 vdev_queue_class_remove(vq, zio);
1050 zio->io_priority = priority;
1051 vdev_queue_class_add(vq, zio);
1052 } else if (zio->io_queue_state == ZIO_QS_NONE) {
1053 zio->io_priority = priority;
1054 }
1055
1056 mutex_exit(&vq->vq_lock);
1057 }
1058
1059 boolean_t
vdev_queue_pool_busy(spa_t * spa)1060 vdev_queue_pool_busy(spa_t *spa)
1061 {
1062 dsl_pool_t *dp = spa_get_dsl(spa);
1063 uint64_t min_bytes = zfs_dirty_data_max *
1064 zfs_vdev_async_write_active_min_dirty_percent / 100;
1065
1066 return (dp->dp_dirty_total > min_bytes);
1067 }
1068
1069 /*
1070 * As these two methods are only used for load calculations we're not
1071 * concerned if we get an incorrect value on 32bit platforms due to lack of
1072 * vq_lock mutex use here, instead we prefer to keep it lock free for
1073 * performance.
1074 */
1075 uint32_t
vdev_queue_length(vdev_t * vd)1076 vdev_queue_length(vdev_t *vd)
1077 {
1078 return (vd->vdev_queue.vq_active);
1079 }
1080
1081 uint64_t
vdev_queue_last_offset(vdev_t * vd)1082 vdev_queue_last_offset(vdev_t *vd)
1083 {
1084 return (vd->vdev_queue.vq_last_offset);
1085 }
1086
1087 uint64_t
vdev_queue_class_length(vdev_t * vd,zio_priority_t p)1088 vdev_queue_class_length(vdev_t *vd, zio_priority_t p)
1089 {
1090 vdev_queue_t *vq = &vd->vdev_queue;
1091 if (vdev_queue_class_fifo(p))
1092 return (vq->vq_class[p].vqc_list_numnodes);
1093 else
1094 return (avl_numnodes(&vq->vq_class[p].vqc_tree));
1095 }
1096
1097 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, aggregation_limit, UINT, ZMOD_RW,
1098 "Max vdev I/O aggregation size");
1099
1100 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, aggregation_limit_non_rotating, UINT,
1101 ZMOD_RW, "Max vdev I/O aggregation size for non-rotating media");
1102
1103 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, read_gap_limit, UINT, ZMOD_RW,
1104 "Aggregate read I/O over gap");
1105
1106 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, write_gap_limit, UINT, ZMOD_RW,
1107 "Aggregate write I/O over gap");
1108
1109 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, max_active, UINT, ZMOD_RW,
1110 "Maximum number of active I/Os per vdev");
1111
1112 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_write_active_max_dirty_percent,
1113 UINT, ZMOD_RW, "Async write concurrency max threshold");
1114
1115 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_write_active_min_dirty_percent,
1116 UINT, ZMOD_RW, "Async write concurrency min threshold");
1117
1118 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_read_max_active, UINT, ZMOD_RW,
1119 "Max active async read I/Os per vdev");
1120
1121 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_read_min_active, UINT, ZMOD_RW,
1122 "Min active async read I/Os per vdev");
1123
1124 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_write_max_active, UINT, ZMOD_RW,
1125 "Max active async write I/Os per vdev");
1126
1127 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_write_min_active, UINT, ZMOD_RW,
1128 "Min active async write I/Os per vdev");
1129
1130 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, initializing_max_active, UINT, ZMOD_RW,
1131 "Max active initializing I/Os per vdev");
1132
1133 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, initializing_min_active, UINT, ZMOD_RW,
1134 "Min active initializing I/Os per vdev");
1135
1136 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, removal_max_active, UINT, ZMOD_RW,
1137 "Max active removal I/Os per vdev");
1138
1139 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, removal_min_active, UINT, ZMOD_RW,
1140 "Min active removal I/Os per vdev");
1141
1142 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, scrub_max_active, UINT, ZMOD_RW,
1143 "Max active scrub I/Os per vdev");
1144
1145 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, scrub_min_active, UINT, ZMOD_RW,
1146 "Min active scrub I/Os per vdev");
1147
1148 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, sync_read_max_active, UINT, ZMOD_RW,
1149 "Max active sync read I/Os per vdev");
1150
1151 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, sync_read_min_active, UINT, ZMOD_RW,
1152 "Min active sync read I/Os per vdev");
1153
1154 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, sync_write_max_active, UINT, ZMOD_RW,
1155 "Max active sync write I/Os per vdev");
1156
1157 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, sync_write_min_active, UINT, ZMOD_RW,
1158 "Min active sync write I/Os per vdev");
1159
1160 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, trim_max_active, UINT, ZMOD_RW,
1161 "Max active trim/discard I/Os per vdev");
1162
1163 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, trim_min_active, UINT, ZMOD_RW,
1164 "Min active trim/discard I/Os per vdev");
1165
1166 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, rebuild_max_active, UINT, ZMOD_RW,
1167 "Max active rebuild I/Os per vdev");
1168
1169 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, rebuild_min_active, UINT, ZMOD_RW,
1170 "Min active rebuild I/Os per vdev");
1171
1172 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, nia_credit, UINT, ZMOD_RW,
1173 "Number of non-interactive I/Os to allow in sequence");
1174
1175 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, nia_delay, UINT, ZMOD_RW,
1176 "Number of non-interactive I/Os before _max_active");
1177