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.
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.
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]
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
27 * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
30 #include <sys/zfs_context.h>
31 #include <sys/vdev_impl.h>
32 #include <sys/spa_impl.h>
35 #include <sys/dsl_pool.h>
36 #include <sys/metaslab_impl.h>
38 #include <sys/spa_impl.h>
39 #include <sys/kstat.h>
46 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The
47 * I/O scheduler determines when and in what order those operations are
48 * issued. The I/O scheduler divides operations into five I/O classes
49 * prioritized in the following order: sync read, sync write, async read,
50 * async write, and scrub/resilver. Each queue defines the minimum and
51 * maximum number of concurrent operations that may be issued to the device.
52 * In addition, the device has an aggregate maximum. Note that the sum of the
53 * per-queue minimums must not exceed the aggregate maximum. If the
54 * sum of the per-queue maximums exceeds the aggregate maximum, then the
55 * number of active i/os may reach zfs_vdev_max_active, in which case no
56 * further i/os will be issued regardless of whether all per-queue
57 * minimums have been met.
59 * For many physical devices, throughput increases with the number of
60 * concurrent operations, but latency typically suffers. Further, physical
61 * devices typically have a limit at which more concurrent operations have no
62 * effect on throughput or can actually cause it to decrease.
64 * The scheduler selects the next operation to issue by first looking for an
65 * I/O class whose minimum has not been satisfied. Once all are satisfied and
66 * the aggregate maximum has not been hit, the scheduler looks for classes
67 * whose maximum has not been satisfied. Iteration through the I/O classes is
68 * done in the order specified above. No further operations are issued if the
69 * aggregate maximum number of concurrent operations has been hit or if there
70 * are no operations queued for an I/O class that has not hit its maximum.
71 * Every time an i/o is queued or an operation completes, the I/O scheduler
72 * looks for new operations to issue.
74 * All I/O classes have a fixed maximum number of outstanding operations
75 * except for the async write class. Asynchronous writes represent the data
76 * that is committed to stable storage during the syncing stage for
77 * transaction groups (see txg.c). Transaction groups enter the syncing state
78 * periodically so the number of queued async writes will quickly burst up and
79 * then bleed down to zero. Rather than servicing them as quickly as possible,
80 * the I/O scheduler changes the maximum number of active async write i/os
81 * according to the amount of dirty data in the pool (see dsl_pool.c). Since
82 * both throughput and latency typically increase with the number of
83 * concurrent operations issued to physical devices, reducing the burstiness
84 * in the number of concurrent operations also stabilizes the response time of
85 * operations from other -- and in particular synchronous -- queues. In broad
86 * strokes, the I/O scheduler will issue more concurrent operations from the
87 * async write queue as there's more dirty data in the pool.
91 * The number of concurrent operations issued for the async write I/O class
92 * follows a piece-wise linear function defined by a few adjustable points.
94 * | o---------| <-- zfs_vdev_async_write_max_active
101 * |------------o | | <-- zfs_vdev_async_write_min_active
102 * 0|____________^______|_________|
103 * 0% | | 100% of zfs_dirty_data_max
105 * | `-- zfs_vdev_async_write_active_max_dirty_percent
106 * `--------- zfs_vdev_async_write_active_min_dirty_percent
108 * Until the amount of dirty data exceeds a minimum percentage of the dirty
109 * data allowed in the pool, the I/O scheduler will limit the number of
110 * concurrent operations to the minimum. As that threshold is crossed, the
111 * number of concurrent operations issued increases linearly to the maximum at
112 * the specified maximum percentage of the dirty data allowed in the pool.
114 * Ideally, the amount of dirty data on a busy pool will stay in the sloped
115 * part of the function between zfs_vdev_async_write_active_min_dirty_percent
116 * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the
117 * maximum percentage, this indicates that the rate of incoming data is
118 * greater than the rate that the backend storage can handle. In this case, we
119 * must further throttle incoming writes (see dmu_tx_delay() for details).
123 * The maximum number of i/os active to each device. Ideally, this will be >=
124 * the sum of each queue's max_active. It must be at least the sum of each
125 * queue's min_active.
127 uint32_t zfs_vdev_max_active = 1000;
130 * Per-queue limits on the number of i/os active to each device. If the
131 * number of active i/os is < zfs_vdev_max_active, then the min_active comes
132 * into play. We will send min_active from each queue, and then select from
133 * queues in the order defined by zio_priority_t.
135 * In general, smaller max_active's will lead to lower latency of synchronous
136 * operations. Larger max_active's may lead to higher overall throughput,
137 * depending on underlying storage.
139 * The ratio of the queues' max_actives determines the balance of performance
140 * between reads, writes, and scrubs. E.g., increasing
141 * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete
142 * more quickly, but reads and writes to have higher latency and lower
145 uint32_t zfs_vdev_sync_read_min_active = 10;
146 uint32_t zfs_vdev_sync_read_max_active = 10;
147 uint32_t zfs_vdev_sync_write_min_active = 10;
148 uint32_t zfs_vdev_sync_write_max_active = 10;
149 uint32_t zfs_vdev_async_read_min_active = 1;
150 uint32_t zfs_vdev_async_read_max_active = 3;
151 uint32_t zfs_vdev_async_write_min_active = 2;
152 uint32_t zfs_vdev_async_write_max_active = 10;
153 uint32_t zfs_vdev_scrub_min_active = 1;
154 uint32_t zfs_vdev_scrub_max_active = 2;
155 uint32_t zfs_vdev_removal_min_active = 1;
156 uint32_t zfs_vdev_removal_max_active = 2;
157 uint32_t zfs_vdev_initializing_min_active = 1;
158 uint32_t zfs_vdev_initializing_max_active = 1;
159 uint32_t zfs_vdev_trim_min_active = 1;
160 uint32_t zfs_vdev_trim_max_active = 2;
163 * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent
164 * dirty data, use zfs_vdev_async_write_min_active. When it has more than
165 * zfs_vdev_async_write_active_max_dirty_percent, use
166 * zfs_vdev_async_write_max_active. The value is linearly interpolated
167 * between min and max.
169 int zfs_vdev_async_write_active_min_dirty_percent = 30;
170 int zfs_vdev_async_write_active_max_dirty_percent = 60;
173 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
174 * For read I/Os, we also aggregate across small adjacency gaps; for writes
175 * we include spans of optional I/Os to aid aggregation at the disk even when
176 * they aren't able to help us aggregate at this level.
178 int zfs_vdev_aggregation_limit = 1 << 20;
179 int zfs_vdev_aggregation_limit_non_rotating = SPA_OLD_MAXBLOCKSIZE;
180 int zfs_vdev_read_gap_limit = 32 << 10;
181 int zfs_vdev_write_gap_limit = 4 << 10;
184 * Define the queue depth percentage for each top-level. This percentage is
185 * used in conjunction with zfs_vdev_async_max_active to determine how many
186 * allocations a specific top-level vdev should handle. Once the queue depth
187 * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100
188 * then allocator will stop allocating blocks on that top-level device.
189 * The default kernel setting is 1000% which will yield 100 allocations per
190 * device. For userland testing, the default setting is 300% which equates
191 * to 30 allocations per device.
194 int zfs_vdev_queue_depth_pct = 1000;
196 int zfs_vdev_queue_depth_pct = 300;
200 * When performing allocations for a given metaslab, we want to make sure that
201 * there are enough IOs to aggregate together to improve throughput. We want to
202 * ensure that there are at least 128k worth of IOs that can be aggregated, and
203 * we assume that the average allocation size is 4k, so we need the queue depth
204 * to be 32 per allocator to get good aggregation of sequential writes.
206 int zfs_vdev_def_queue_depth = 32;
209 * Allow TRIM I/Os to be aggregated. This should normally not be needed since
210 * TRIM I/O for extents up to zfs_trim_extent_bytes_max (128M) can be submitted
211 * by the TRIM code in zfs_trim.c.
213 int zfs_vdev_aggregate_trim = 0;
216 vdev_queue_offset_compare(const void *x1, const void *x2)
218 const zio_t *z1 = (const zio_t *)x1;
219 const zio_t *z2 = (const zio_t *)x2;
221 int cmp = TREE_CMP(z1->io_offset, z2->io_offset);
226 return (TREE_PCMP(z1, z2));
229 static inline avl_tree_t *
230 vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p)
232 return (&vq->vq_class[p].vqc_queued_tree);
235 static inline avl_tree_t *
236 vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t)
238 ASSERT(t == ZIO_TYPE_READ || t == ZIO_TYPE_WRITE || t == ZIO_TYPE_TRIM);
239 if (t == ZIO_TYPE_READ)
240 return (&vq->vq_read_offset_tree);
241 else if (t == ZIO_TYPE_WRITE)
242 return (&vq->vq_write_offset_tree);
244 return (&vq->vq_trim_offset_tree);
248 vdev_queue_timestamp_compare(const void *x1, const void *x2)
250 const zio_t *z1 = (const zio_t *)x1;
251 const zio_t *z2 = (const zio_t *)x2;
253 int cmp = TREE_CMP(z1->io_timestamp, z2->io_timestamp);
258 return (TREE_PCMP(z1, z2));
262 vdev_queue_class_min_active(zio_priority_t p)
265 case ZIO_PRIORITY_SYNC_READ:
266 return (zfs_vdev_sync_read_min_active);
267 case ZIO_PRIORITY_SYNC_WRITE:
268 return (zfs_vdev_sync_write_min_active);
269 case ZIO_PRIORITY_ASYNC_READ:
270 return (zfs_vdev_async_read_min_active);
271 case ZIO_PRIORITY_ASYNC_WRITE:
272 return (zfs_vdev_async_write_min_active);
273 case ZIO_PRIORITY_SCRUB:
274 return (zfs_vdev_scrub_min_active);
275 case ZIO_PRIORITY_REMOVAL:
276 return (zfs_vdev_removal_min_active);
277 case ZIO_PRIORITY_INITIALIZING:
278 return (zfs_vdev_initializing_min_active);
279 case ZIO_PRIORITY_TRIM:
280 return (zfs_vdev_trim_min_active);
282 panic("invalid priority %u", p);
288 vdev_queue_max_async_writes(spa_t *spa)
292 dsl_pool_t *dp = spa_get_dsl(spa);
293 uint64_t min_bytes = zfs_dirty_data_max *
294 zfs_vdev_async_write_active_min_dirty_percent / 100;
295 uint64_t max_bytes = zfs_dirty_data_max *
296 zfs_vdev_async_write_active_max_dirty_percent / 100;
299 * Async writes may occur before the assignment of the spa's
300 * dsl_pool_t if a self-healing zio is issued prior to the
301 * completion of dmu_objset_open_impl().
304 return (zfs_vdev_async_write_max_active);
307 * Sync tasks correspond to interactive user actions. To reduce the
308 * execution time of those actions we push data out as fast as possible.
310 if (spa_has_pending_synctask(spa))
311 return (zfs_vdev_async_write_max_active);
313 dirty = dp->dp_dirty_total;
314 if (dirty < min_bytes)
315 return (zfs_vdev_async_write_min_active);
316 if (dirty > max_bytes)
317 return (zfs_vdev_async_write_max_active);
320 * linear interpolation:
321 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
322 * move right by min_bytes
323 * move up by min_writes
325 writes = (dirty - min_bytes) *
326 (zfs_vdev_async_write_max_active -
327 zfs_vdev_async_write_min_active) /
328 (max_bytes - min_bytes) +
329 zfs_vdev_async_write_min_active;
330 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
331 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
336 vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
339 case ZIO_PRIORITY_SYNC_READ:
340 return (zfs_vdev_sync_read_max_active);
341 case ZIO_PRIORITY_SYNC_WRITE:
342 return (zfs_vdev_sync_write_max_active);
343 case ZIO_PRIORITY_ASYNC_READ:
344 return (zfs_vdev_async_read_max_active);
345 case ZIO_PRIORITY_ASYNC_WRITE:
346 return (vdev_queue_max_async_writes(spa));
347 case ZIO_PRIORITY_SCRUB:
348 return (zfs_vdev_scrub_max_active);
349 case ZIO_PRIORITY_REMOVAL:
350 return (zfs_vdev_removal_max_active);
351 case ZIO_PRIORITY_INITIALIZING:
352 return (zfs_vdev_initializing_max_active);
353 case ZIO_PRIORITY_TRIM:
354 return (zfs_vdev_trim_max_active);
356 panic("invalid priority %u", p);
362 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
363 * there is no eligible class.
365 static zio_priority_t
366 vdev_queue_class_to_issue(vdev_queue_t *vq)
368 spa_t *spa = vq->vq_vdev->vdev_spa;
371 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
372 return (ZIO_PRIORITY_NUM_QUEUEABLE);
374 /* find a queue that has not reached its minimum # outstanding i/os */
375 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
376 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
377 vq->vq_class[p].vqc_active <
378 vdev_queue_class_min_active(p))
383 * If we haven't found a queue, look for one that hasn't reached its
384 * maximum # outstanding i/os.
386 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
387 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
388 vq->vq_class[p].vqc_active <
389 vdev_queue_class_max_active(spa, p))
393 /* No eligible queued i/os */
394 return (ZIO_PRIORITY_NUM_QUEUEABLE);
398 vdev_queue_init(vdev_t *vd)
400 vdev_queue_t *vq = &vd->vdev_queue;
403 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
405 taskq_init_ent(&vd->vdev_queue.vq_io_search.io_tqent);
407 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
408 sizeof (zio_t), offsetof(struct zio, io_queue_node));
409 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ),
410 vdev_queue_offset_compare, sizeof (zio_t),
411 offsetof(struct zio, io_offset_node));
412 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE),
413 vdev_queue_offset_compare, sizeof (zio_t),
414 offsetof(struct zio, io_offset_node));
415 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_TRIM),
416 vdev_queue_offset_compare, sizeof (zio_t),
417 offsetof(struct zio, io_offset_node));
419 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
420 int (*compfn) (const void *, const void *);
423 * The synchronous/trim i/o queues are dispatched in FIFO rather
424 * than LBA order. This provides more consistent latency for
427 if (p == ZIO_PRIORITY_SYNC_READ ||
428 p == ZIO_PRIORITY_SYNC_WRITE ||
429 p == ZIO_PRIORITY_TRIM) {
430 compfn = vdev_queue_timestamp_compare;
432 compfn = vdev_queue_offset_compare;
434 avl_create(vdev_queue_class_tree(vq, p), compfn,
435 sizeof (zio_t), offsetof(struct zio, io_queue_node));
438 vq->vq_last_offset = 0;
442 vdev_queue_fini(vdev_t *vd)
444 vdev_queue_t *vq = &vd->vdev_queue;
446 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
447 avl_destroy(vdev_queue_class_tree(vq, p));
448 avl_destroy(&vq->vq_active_tree);
449 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
450 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE));
451 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_TRIM));
453 mutex_destroy(&vq->vq_lock);
457 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
459 spa_t *spa = zio->io_spa;
460 spa_history_kstat_t *shk = &spa->spa_stats.io_history;
462 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
463 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
464 avl_add(vdev_queue_type_tree(vq, zio->io_type), zio);
466 if (shk->kstat != NULL) {
467 mutex_enter(&shk->lock);
468 kstat_waitq_enter(shk->kstat->ks_data);
469 mutex_exit(&shk->lock);
474 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
476 spa_t *spa = zio->io_spa;
477 spa_history_kstat_t *shk = &spa->spa_stats.io_history;
479 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
480 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
481 avl_remove(vdev_queue_type_tree(vq, zio->io_type), zio);
483 if (shk->kstat != NULL) {
484 mutex_enter(&shk->lock);
485 kstat_waitq_exit(shk->kstat->ks_data);
486 mutex_exit(&shk->lock);
491 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
493 spa_t *spa = zio->io_spa;
494 spa_history_kstat_t *shk = &spa->spa_stats.io_history;
496 ASSERT(MUTEX_HELD(&vq->vq_lock));
497 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
498 vq->vq_class[zio->io_priority].vqc_active++;
499 avl_add(&vq->vq_active_tree, zio);
501 if (shk->kstat != NULL) {
502 mutex_enter(&shk->lock);
503 kstat_runq_enter(shk->kstat->ks_data);
504 mutex_exit(&shk->lock);
509 vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
511 spa_t *spa = zio->io_spa;
512 spa_history_kstat_t *shk = &spa->spa_stats.io_history;
514 ASSERT(MUTEX_HELD(&vq->vq_lock));
515 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
516 vq->vq_class[zio->io_priority].vqc_active--;
517 avl_remove(&vq->vq_active_tree, zio);
519 if (shk->kstat != NULL) {
520 kstat_io_t *ksio = shk->kstat->ks_data;
522 mutex_enter(&shk->lock);
523 kstat_runq_exit(ksio);
524 if (zio->io_type == ZIO_TYPE_READ) {
526 ksio->nread += zio->io_size;
527 } else if (zio->io_type == ZIO_TYPE_WRITE) {
529 ksio->nwritten += zio->io_size;
531 mutex_exit(&shk->lock);
536 vdev_queue_agg_io_done(zio_t *aio)
538 if (aio->io_type == ZIO_TYPE_READ) {
540 zio_link_t *zl = NULL;
541 while ((pio = zio_walk_parents(aio, &zl)) != NULL) {
542 abd_copy_off(pio->io_abd, aio->io_abd,
543 0, pio->io_offset - aio->io_offset, pio->io_size);
547 abd_free(aio->io_abd);
551 * Compute the range spanned by two i/os, which is the endpoint of the last
552 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
553 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
554 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
556 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
557 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
560 vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
562 zio_t *first, *last, *aio, *dio, *mandatory, *nio;
563 zio_link_t *zl = NULL;
568 boolean_t stretch = B_FALSE;
569 avl_tree_t *t = vdev_queue_type_tree(vq, zio->io_type);
570 enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
573 maxblocksize = spa_maxblocksize(vq->vq_vdev->vdev_spa);
574 if (vq->vq_vdev->vdev_nonrot)
575 limit = zfs_vdev_aggregation_limit_non_rotating;
577 limit = zfs_vdev_aggregation_limit;
578 limit = MAX(MIN(limit, maxblocksize), 0);
580 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE || limit == 0)
584 * While TRIM commands could be aggregated based on offset this
585 * behavior is disabled until it's determined to be beneficial.
587 if (zio->io_type == ZIO_TYPE_TRIM && !zfs_vdev_aggregate_trim)
592 if (zio->io_type == ZIO_TYPE_READ)
593 maxgap = zfs_vdev_read_gap_limit;
596 * We can aggregate I/Os that are sufficiently adjacent and of
597 * the same flavor, as expressed by the AGG_INHERIT flags.
598 * The latter requirement is necessary so that certain
599 * attributes of the I/O, such as whether it's a normal I/O
600 * or a scrub/resilver, can be preserved in the aggregate.
601 * We can include optional I/Os, but don't allow them
602 * to begin a range as they add no benefit in that situation.
606 * We keep track of the last non-optional I/O.
608 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
611 * Walk backwards through sufficiently contiguous I/Os
612 * recording the last non-optional I/O.
614 while ((dio = AVL_PREV(t, first)) != NULL &&
615 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
616 IO_SPAN(dio, last) <= limit &&
617 IO_GAP(dio, first) <= maxgap &&
618 dio->io_type == zio->io_type) {
620 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
625 * Skip any initial optional I/Os.
627 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
628 first = AVL_NEXT(t, first);
629 ASSERT(first != NULL);
634 * Walk forward through sufficiently contiguous I/Os.
635 * The aggregation limit does not apply to optional i/os, so that
636 * we can issue contiguous writes even if they are larger than the
639 while ((dio = AVL_NEXT(t, last)) != NULL &&
640 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
641 (IO_SPAN(first, dio) <= limit ||
642 (dio->io_flags & ZIO_FLAG_OPTIONAL)) &&
643 IO_SPAN(first, dio) <= maxblocksize &&
644 IO_GAP(last, dio) <= maxgap &&
645 dio->io_type == zio->io_type) {
647 if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
652 * Now that we've established the range of the I/O aggregation
653 * we must decide what to do with trailing optional I/Os.
654 * For reads, there's nothing to do. While we are unable to
655 * aggregate further, it's possible that a trailing optional
656 * I/O would allow the underlying device to aggregate with
657 * subsequent I/Os. We must therefore determine if the next
658 * non-optional I/O is close enough to make aggregation
661 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
663 while ((dio = AVL_NEXT(t, nio)) != NULL &&
664 IO_GAP(nio, dio) == 0 &&
665 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
667 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
676 * We are going to include an optional io in our aggregated
677 * span, thus closing the write gap. Only mandatory i/os can
678 * start aggregated spans, so make sure that the next i/o
679 * after our span is mandatory.
681 dio = AVL_NEXT(t, last);
682 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
684 /* do not include the optional i/o */
685 while (last != mandatory && last != first) {
686 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
687 last = AVL_PREV(t, last);
688 ASSERT(last != NULL);
695 size = IO_SPAN(first, last);
696 ASSERT3U(size, <=, maxblocksize);
698 abd = abd_alloc_for_io(size, B_TRUE);
702 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
703 abd, size, first->io_type, zio->io_priority,
704 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
705 vdev_queue_agg_io_done, NULL);
706 aio->io_timestamp = first->io_timestamp;
711 nio = AVL_NEXT(t, dio);
712 zio_add_child(dio, aio);
713 vdev_queue_io_remove(vq, dio);
714 } while (dio != last);
717 * We need to drop the vdev queue's lock during zio_execute() to
718 * avoid a deadlock that we could encounter due to lock order
719 * reversal between vq_lock and io_lock in zio_change_priority().
720 * Use the dropped lock to do memory copy without congestion.
722 mutex_exit(&vq->vq_lock);
723 while ((dio = zio_walk_parents(aio, &zl)) != NULL) {
724 ASSERT3U(dio->io_type, ==, aio->io_type);
726 if (dio->io_flags & ZIO_FLAG_NODATA) {
727 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
728 abd_zero_off(aio->io_abd,
729 dio->io_offset - aio->io_offset, dio->io_size);
730 } else if (dio->io_type == ZIO_TYPE_WRITE) {
731 abd_copy_off(aio->io_abd, dio->io_abd,
732 dio->io_offset - aio->io_offset, 0, dio->io_size);
735 zio_vdev_io_bypass(dio);
738 mutex_enter(&vq->vq_lock);
744 vdev_queue_io_to_issue(vdev_queue_t *vq)
752 ASSERT(MUTEX_HELD(&vq->vq_lock));
754 p = vdev_queue_class_to_issue(vq);
756 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
757 /* No eligible queued i/os */
762 * For LBA-ordered queues (async / scrub / initializing), issue the
763 * i/o which follows the most recently issued i/o in LBA (offset) order.
765 * For FIFO queues (sync/trim), issue the i/o with the lowest timestamp.
767 tree = vdev_queue_class_tree(vq, p);
768 vq->vq_io_search.io_timestamp = 0;
769 vq->vq_io_search.io_offset = vq->vq_last_offset - 1;
770 VERIFY3P(avl_find(tree, &vq->vq_io_search, &idx), ==, NULL);
771 zio = avl_nearest(tree, idx, AVL_AFTER);
773 zio = avl_first(tree);
774 ASSERT3U(zio->io_priority, ==, p);
776 aio = vdev_queue_aggregate(vq, zio);
780 vdev_queue_io_remove(vq, zio);
783 * If the I/O is or was optional and therefore has no data, we need to
784 * simply discard it. We need to drop the vdev queue's lock to avoid a
785 * deadlock that we could encounter since this I/O will complete
788 if (zio->io_flags & ZIO_FLAG_NODATA) {
789 mutex_exit(&vq->vq_lock);
790 zio_vdev_io_bypass(zio);
792 mutex_enter(&vq->vq_lock);
796 vdev_queue_pending_add(vq, zio);
797 vq->vq_last_offset = zio->io_offset + zio->io_size;
803 vdev_queue_io(zio_t *zio)
805 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
808 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
812 * Children i/os inherent their parent's priority, which might
813 * not match the child's i/o type. Fix it up here.
815 if (zio->io_type == ZIO_TYPE_READ) {
816 ASSERT(zio->io_priority != ZIO_PRIORITY_TRIM);
818 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
819 zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
820 zio->io_priority != ZIO_PRIORITY_SCRUB &&
821 zio->io_priority != ZIO_PRIORITY_REMOVAL &&
822 zio->io_priority != ZIO_PRIORITY_INITIALIZING) {
823 zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
825 } else if (zio->io_type == ZIO_TYPE_WRITE) {
826 ASSERT(zio->io_priority != ZIO_PRIORITY_TRIM);
828 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
829 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE &&
830 zio->io_priority != ZIO_PRIORITY_REMOVAL &&
831 zio->io_priority != ZIO_PRIORITY_INITIALIZING) {
832 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
835 ASSERT(zio->io_type == ZIO_TYPE_TRIM);
836 ASSERT(zio->io_priority == ZIO_PRIORITY_TRIM);
839 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
841 mutex_enter(&vq->vq_lock);
842 zio->io_timestamp = gethrtime();
843 vdev_queue_io_add(vq, zio);
844 nio = vdev_queue_io_to_issue(vq);
845 mutex_exit(&vq->vq_lock);
850 if (nio->io_done == vdev_queue_agg_io_done) {
859 vdev_queue_io_done(zio_t *zio)
861 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
864 mutex_enter(&vq->vq_lock);
866 vdev_queue_pending_remove(vq, zio);
868 zio->io_delta = gethrtime() - zio->io_timestamp;
869 vq->vq_io_complete_ts = gethrtime();
870 vq->vq_io_delta_ts = vq->vq_io_complete_ts - zio->io_timestamp;
872 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
873 mutex_exit(&vq->vq_lock);
874 if (nio->io_done == vdev_queue_agg_io_done) {
877 zio_vdev_io_reissue(nio);
880 mutex_enter(&vq->vq_lock);
883 mutex_exit(&vq->vq_lock);
887 vdev_queue_change_io_priority(zio_t *zio, zio_priority_t priority)
889 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
893 * ZIO_PRIORITY_NOW is used by the vdev cache code and the aggregate zio
894 * code to issue IOs without adding them to the vdev queue. In this
895 * case, the zio is already going to be issued as quickly as possible
896 * and so it doesn't need any reprioritization to help.
898 if (zio->io_priority == ZIO_PRIORITY_NOW)
901 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
902 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
904 if (zio->io_type == ZIO_TYPE_READ) {
905 if (priority != ZIO_PRIORITY_SYNC_READ &&
906 priority != ZIO_PRIORITY_ASYNC_READ &&
907 priority != ZIO_PRIORITY_SCRUB)
908 priority = ZIO_PRIORITY_ASYNC_READ;
910 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
911 if (priority != ZIO_PRIORITY_SYNC_WRITE &&
912 priority != ZIO_PRIORITY_ASYNC_WRITE)
913 priority = ZIO_PRIORITY_ASYNC_WRITE;
916 mutex_enter(&vq->vq_lock);
919 * If the zio is in none of the queues we can simply change
920 * the priority. If the zio is waiting to be submitted we must
921 * remove it from the queue and re-insert it with the new priority.
922 * Otherwise, the zio is currently active and we cannot change its
925 tree = vdev_queue_class_tree(vq, zio->io_priority);
926 if (avl_find(tree, zio, NULL) == zio) {
927 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
928 zio->io_priority = priority;
929 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
930 } else if (avl_find(&vq->vq_active_tree, zio, NULL) != zio) {
931 zio->io_priority = priority;
934 mutex_exit(&vq->vq_lock);
938 * As these two methods are only used for load calculations we're not
939 * concerned if we get an incorrect value on 32bit platforms due to lack of
940 * vq_lock mutex use here, instead we prefer to keep it lock free for
944 vdev_queue_length(vdev_t *vd)
946 return (avl_numnodes(&vd->vdev_queue.vq_active_tree));
950 vdev_queue_last_offset(vdev_t *vd)
952 return (vd->vdev_queue.vq_last_offset);
956 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, aggregation_limit, INT, ZMOD_RW,
957 "Max vdev I/O aggregation size");
959 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, aggregation_limit_non_rotating, INT, ZMOD_RW,
960 "Max vdev I/O aggregation size for non-rotating media");
962 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, aggregate_trim, INT, ZMOD_RW,
963 "Allow TRIM I/O to be aggregated");
965 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, read_gap_limit, INT, ZMOD_RW,
966 "Aggregate read I/O over gap");
968 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, write_gap_limit, INT, ZMOD_RW,
969 "Aggregate write I/O over gap");
971 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, max_active, INT, ZMOD_RW,
972 "Maximum number of active I/Os per vdev");
974 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_write_active_max_dirty_percent, INT, ZMOD_RW,
975 "Async write concurrency max threshold");
977 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_write_active_min_dirty_percent, INT, ZMOD_RW,
978 "Async write concurrency min threshold");
980 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_read_max_active, INT, ZMOD_RW,
981 "Max active async read I/Os per vdev");
983 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_read_min_active, INT, ZMOD_RW,
984 "Min active async read I/Os per vdev");
986 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_write_max_active, INT, ZMOD_RW,
987 "Max active async write I/Os per vdev");
989 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, async_write_min_active, INT, ZMOD_RW,
990 "Min active async write I/Os per vdev");
992 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, initializing_max_active, INT, ZMOD_RW,
993 "Max active initializing I/Os per vdev");
995 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, initializing_min_active, INT, ZMOD_RW,
996 "Min active initializing I/Os per vdev");
998 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, removal_max_active, INT, ZMOD_RW,
999 "Max active removal I/Os per vdev");
1001 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, removal_min_active, INT, ZMOD_RW,
1002 "Min active removal I/Os per vdev");
1004 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, scrub_max_active, INT, ZMOD_RW,
1005 "Max active scrub I/Os per vdev");
1007 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, scrub_min_active, INT, ZMOD_RW,
1008 "Min active scrub I/Os per vdev");
1010 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, sync_read_max_active, INT, ZMOD_RW,
1011 "Max active sync read I/Os per vdev");
1013 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, sync_read_min_active, INT, ZMOD_RW,
1014 "Min active sync read I/Os per vdev");
1016 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, sync_write_max_active, INT, ZMOD_RW,
1017 "Max active sync write I/Os per vdev");
1019 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, sync_write_min_active, INT, ZMOD_RW,
1020 "Min active sync write I/Os per vdev");
1022 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, trim_max_active, INT, ZMOD_RW,
1023 "Max active trim/discard I/Os per vdev");
1025 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, trim_min_active, INT, ZMOD_RW,
1026 "Min active trim/discard I/Os per vdev");
1028 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, queue_depth_pct, INT, ZMOD_RW,
1029 "Queue depth percentage for each top-level vdev");