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.
28 * Copyright (c) 2014 Integros [integros.com]
31 #include <sys/zfs_context.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/spa_impl.h>
36 #include <sys/dsl_pool.h>
37 #include <sys/metaslab_impl.h>
44 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The
45 * I/O scheduler determines when and in what order those operations are
46 * issued. The I/O scheduler divides operations into six I/O classes
47 * prioritized in the following order: sync read, sync write, async read,
48 * async write, scrub/resilver and trim. Each queue defines the minimum and
49 * maximum number of concurrent operations that may be issued to the device.
50 * In addition, the device has an aggregate maximum. Note that the sum of the
51 * per-queue minimums must not exceed the aggregate maximum, and if the
52 * aggregate maximum is equal to or greater than the sum of the per-queue
53 * maximums, the per-queue minimum has no effect.
55 * For many physical devices, throughput increases with the number of
56 * concurrent operations, but latency typically suffers. Further, physical
57 * devices typically have a limit at which more concurrent operations have no
58 * effect on throughput or can actually cause it to decrease.
60 * The scheduler selects the next operation to issue by first looking for an
61 * I/O class whose minimum has not been satisfied. Once all are satisfied and
62 * the aggregate maximum has not been hit, the scheduler looks for classes
63 * whose maximum has not been satisfied. Iteration through the I/O classes is
64 * done in the order specified above. No further operations are issued if the
65 * aggregate maximum number of concurrent operations has been hit or if there
66 * are no operations queued for an I/O class that has not hit its maximum.
67 * Every time an I/O is queued or an operation completes, the I/O scheduler
68 * looks for new operations to issue.
70 * All I/O classes have a fixed maximum number of outstanding operations
71 * except for the async write class. Asynchronous writes represent the data
72 * that is committed to stable storage during the syncing stage for
73 * transaction groups (see txg.c). Transaction groups enter the syncing state
74 * periodically so the number of queued async writes will quickly burst up and
75 * then bleed down to zero. Rather than servicing them as quickly as possible,
76 * the I/O scheduler changes the maximum number of active async write I/Os
77 * according to the amount of dirty data in the pool (see dsl_pool.c). Since
78 * both throughput and latency typically increase with the number of
79 * concurrent operations issued to physical devices, reducing the burstiness
80 * in the number of concurrent operations also stabilizes the response time of
81 * operations from other -- and in particular synchronous -- queues. In broad
82 * strokes, the I/O scheduler will issue more concurrent operations from the
83 * async write queue as there's more dirty data in the pool.
87 * The number of concurrent operations issued for the async write I/O class
88 * follows a piece-wise linear function defined by a few adjustable points.
90 * | o---------| <-- zfs_vdev_async_write_max_active
97 * |------------o | | <-- zfs_vdev_async_write_min_active
98 * 0|____________^______|_________|
99 * 0% | | 100% of zfs_dirty_data_max
101 * | `-- zfs_vdev_async_write_active_max_dirty_percent
102 * `--------- zfs_vdev_async_write_active_min_dirty_percent
104 * Until the amount of dirty data exceeds a minimum percentage of the dirty
105 * data allowed in the pool, the I/O scheduler will limit the number of
106 * concurrent operations to the minimum. As that threshold is crossed, the
107 * number of concurrent operations issued increases linearly to the maximum at
108 * the specified maximum percentage of the dirty data allowed in the pool.
110 * Ideally, the amount of dirty data on a busy pool will stay in the sloped
111 * part of the function between zfs_vdev_async_write_active_min_dirty_percent
112 * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the
113 * maximum percentage, this indicates that the rate of incoming data is
114 * greater than the rate that the backend storage can handle. In this case, we
115 * must further throttle incoming writes (see dmu_tx_delay() for details).
119 * The maximum number of I/Os active to each device. Ideally, this will be >=
120 * the sum of each queue's max_active. It must be at least the sum of each
121 * queue's min_active.
123 uint32_t zfs_vdev_max_active = 1000;
126 * Per-queue limits on the number of I/Os active to each device. If the
127 * sum of the queue's max_active is < zfs_vdev_max_active, then the
128 * min_active comes into play. We will send min_active from each queue,
129 * and then select from queues in the order defined by zio_priority_t.
131 * In general, smaller max_active's will lead to lower latency of synchronous
132 * operations. Larger max_active's may lead to higher overall throughput,
133 * depending on underlying storage.
135 * The ratio of the queues' max_actives determines the balance of performance
136 * between reads, writes, and scrubs. E.g., increasing
137 * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete
138 * more quickly, but reads and writes to have higher latency and lower
141 uint32_t zfs_vdev_sync_read_min_active = 10;
142 uint32_t zfs_vdev_sync_read_max_active = 10;
143 uint32_t zfs_vdev_sync_write_min_active = 10;
144 uint32_t zfs_vdev_sync_write_max_active = 10;
145 uint32_t zfs_vdev_async_read_min_active = 1;
146 uint32_t zfs_vdev_async_read_max_active = 3;
147 uint32_t zfs_vdev_async_write_min_active = 1;
148 uint32_t zfs_vdev_async_write_max_active = 10;
149 uint32_t zfs_vdev_scrub_min_active = 1;
150 uint32_t zfs_vdev_scrub_max_active = 2;
151 uint32_t zfs_vdev_trim_min_active = 1;
153 * TRIM max active is large in comparison to the other values due to the fact
154 * that TRIM IOs are coalesced at the device layer. This value is set such
155 * that a typical SSD can process the queued IOs in a single request.
157 uint32_t zfs_vdev_trim_max_active = 64;
158 uint32_t zfs_vdev_removal_min_active = 1;
159 uint32_t zfs_vdev_removal_max_active = 2;
160 uint32_t zfs_vdev_initializing_min_active = 1;
161 uint32_t zfs_vdev_initializing_max_active = 1;
165 * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent
166 * dirty data, use zfs_vdev_async_write_min_active. When it has more than
167 * zfs_vdev_async_write_active_max_dirty_percent, use
168 * zfs_vdev_async_write_max_active. The value is linearly interpolated
169 * between min and max.
171 int zfs_vdev_async_write_active_min_dirty_percent = 30;
172 int zfs_vdev_async_write_active_max_dirty_percent = 60;
175 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
176 * For read I/Os, we also aggregate across small adjacency gaps; for writes
177 * we include spans of optional I/Os to aid aggregation at the disk even when
178 * they aren't able to help us aggregate at this level.
180 int zfs_vdev_aggregation_limit = 1 << 20;
181 int zfs_vdev_read_gap_limit = 32 << 10;
182 int zfs_vdev_write_gap_limit = 4 << 10;
185 * Define the queue depth percentage for each top-level. This percentage is
186 * used in conjunction with zfs_vdev_async_max_active to determine how many
187 * allocations a specific top-level vdev should handle. Once the queue depth
188 * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100
189 * then allocator will stop allocating blocks on that top-level device.
190 * The default kernel setting is 1000% which will yield 100 allocations per
191 * device. For userland testing, the default setting is 300% which equates
192 * to 30 allocations per device.
195 int zfs_vdev_queue_depth_pct = 1000;
197 int zfs_vdev_queue_depth_pct = 300;
201 * When performing allocations for a given metaslab, we want to make sure that
202 * there are enough IOs to aggregate together to improve throughput. We want to
203 * ensure that there are at least 128k worth of IOs that can be aggregated, and
204 * we assume that the average allocation size is 4k, so we need the queue depth
205 * to be 32 per allocator to get good aggregation of sequential writes.
207 int zfs_vdev_def_queue_depth = 32;
211 SYSCTL_DECL(_vfs_zfs_vdev);
213 static int sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS);
214 SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_min_dirty_percent,
215 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
216 sysctl_zfs_async_write_active_min_dirty_percent, "I",
217 "Percentage of async write dirty data below which "
218 "async_write_min_active is used.");
220 static int sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS);
221 SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_max_dirty_percent,
222 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
223 sysctl_zfs_async_write_active_max_dirty_percent, "I",
224 "Percentage of async write dirty data above which "
225 "async_write_max_active is used.");
227 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RWTUN,
228 &zfs_vdev_max_active, 0,
229 "The maximum number of I/Os of all types active for each device.");
231 #define ZFS_VDEV_QUEUE_KNOB_MIN(name) \
232 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, CTLFLAG_RWTUN,\
233 &zfs_vdev_ ## name ## _min_active, 0, \
234 "Initial number of I/O requests of type " #name \
235 " active for each device");
237 #define ZFS_VDEV_QUEUE_KNOB_MAX(name) \
238 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, CTLFLAG_RWTUN,\
239 &zfs_vdev_ ## name ## _max_active, 0, \
240 "Maximum number of I/O requests of type " #name \
241 " active for each device");
243 ZFS_VDEV_QUEUE_KNOB_MIN(sync_read);
244 ZFS_VDEV_QUEUE_KNOB_MAX(sync_read);
245 ZFS_VDEV_QUEUE_KNOB_MIN(sync_write);
246 ZFS_VDEV_QUEUE_KNOB_MAX(sync_write);
247 ZFS_VDEV_QUEUE_KNOB_MIN(async_read);
248 ZFS_VDEV_QUEUE_KNOB_MAX(async_read);
249 ZFS_VDEV_QUEUE_KNOB_MIN(async_write);
250 ZFS_VDEV_QUEUE_KNOB_MAX(async_write);
251 ZFS_VDEV_QUEUE_KNOB_MIN(scrub);
252 ZFS_VDEV_QUEUE_KNOB_MAX(scrub);
253 ZFS_VDEV_QUEUE_KNOB_MIN(trim);
254 ZFS_VDEV_QUEUE_KNOB_MAX(trim);
256 #undef ZFS_VDEV_QUEUE_KNOB
258 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RWTUN,
259 &zfs_vdev_aggregation_limit, 0,
260 "I/O requests are aggregated up to this size");
261 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RWTUN,
262 &zfs_vdev_read_gap_limit, 0,
263 "Acceptable gap between two reads being aggregated");
264 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RWTUN,
265 &zfs_vdev_write_gap_limit, 0,
266 "Acceptable gap between two writes being aggregated");
267 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, queue_depth_pct, CTLFLAG_RWTUN,
268 &zfs_vdev_queue_depth_pct, 0,
269 "Queue depth percentage for each top-level");
272 sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS)
276 val = zfs_vdev_async_write_active_min_dirty_percent;
277 err = sysctl_handle_int(oidp, &val, 0, req);
278 if (err != 0 || req->newptr == NULL)
281 if (val < 0 || val > 100 ||
282 val >= zfs_vdev_async_write_active_max_dirty_percent)
285 zfs_vdev_async_write_active_min_dirty_percent = val;
291 sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS)
295 val = zfs_vdev_async_write_active_max_dirty_percent;
296 err = sysctl_handle_int(oidp, &val, 0, req);
297 if (err != 0 || req->newptr == NULL)
300 if (val < 0 || val > 100 ||
301 val <= zfs_vdev_async_write_active_min_dirty_percent)
304 zfs_vdev_async_write_active_max_dirty_percent = val;
312 vdev_queue_offset_compare(const void *x1, const void *x2)
314 const zio_t *z1 = x1;
315 const zio_t *z2 = x2;
317 if (z1->io_offset < z2->io_offset)
319 if (z1->io_offset > z2->io_offset)
330 static inline avl_tree_t *
331 vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p)
333 return (&vq->vq_class[p].vqc_queued_tree);
336 static inline avl_tree_t *
337 vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t)
339 if (t == ZIO_TYPE_READ)
340 return (&vq->vq_read_offset_tree);
341 else if (t == ZIO_TYPE_WRITE)
342 return (&vq->vq_write_offset_tree);
348 vdev_queue_timestamp_compare(const void *x1, const void *x2)
350 const zio_t *z1 = x1;
351 const zio_t *z2 = x2;
353 if (z1->io_timestamp < z2->io_timestamp)
355 if (z1->io_timestamp > z2->io_timestamp)
358 if (z1->io_offset < z2->io_offset)
360 if (z1->io_offset > z2->io_offset)
372 vdev_queue_init(vdev_t *vd)
374 vdev_queue_t *vq = &vd->vdev_queue;
376 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
379 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
380 sizeof (zio_t), offsetof(struct zio, io_queue_node));
381 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ),
382 vdev_queue_offset_compare, sizeof (zio_t),
383 offsetof(struct zio, io_offset_node));
384 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE),
385 vdev_queue_offset_compare, sizeof (zio_t),
386 offsetof(struct zio, io_offset_node));
388 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
389 int (*compfn) (const void *, const void *);
392 * The synchronous i/o queues are dispatched in FIFO rather
393 * than LBA order. This provides more consistent latency for
396 if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE)
397 compfn = vdev_queue_timestamp_compare;
399 compfn = vdev_queue_offset_compare;
401 avl_create(vdev_queue_class_tree(vq, p), compfn,
402 sizeof (zio_t), offsetof(struct zio, io_queue_node));
405 vq->vq_lastoffset = 0;
409 vdev_queue_fini(vdev_t *vd)
411 vdev_queue_t *vq = &vd->vdev_queue;
413 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
414 avl_destroy(vdev_queue_class_tree(vq, p));
415 avl_destroy(&vq->vq_active_tree);
416 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
417 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE));
419 mutex_destroy(&vq->vq_lock);
423 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
425 spa_t *spa = zio->io_spa;
428 ASSERT(MUTEX_HELD(&vq->vq_lock));
429 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
430 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
431 qtt = vdev_queue_type_tree(vq, zio->io_type);
436 mutex_enter(&spa->spa_iokstat_lock);
437 spa->spa_queue_stats[zio->io_priority].spa_queued++;
438 if (spa->spa_iokstat != NULL)
439 kstat_waitq_enter(spa->spa_iokstat->ks_data);
440 mutex_exit(&spa->spa_iokstat_lock);
445 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
447 spa_t *spa = zio->io_spa;
450 ASSERT(MUTEX_HELD(&vq->vq_lock));
451 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
452 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
453 qtt = vdev_queue_type_tree(vq, zio->io_type);
455 avl_remove(qtt, zio);
458 mutex_enter(&spa->spa_iokstat_lock);
459 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
460 spa->spa_queue_stats[zio->io_priority].spa_queued--;
461 if (spa->spa_iokstat != NULL)
462 kstat_waitq_exit(spa->spa_iokstat->ks_data);
463 mutex_exit(&spa->spa_iokstat_lock);
468 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
470 spa_t *spa = zio->io_spa;
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_add(&vq->vq_active_tree, zio);
477 mutex_enter(&spa->spa_iokstat_lock);
478 spa->spa_queue_stats[zio->io_priority].spa_active++;
479 if (spa->spa_iokstat != NULL)
480 kstat_runq_enter(spa->spa_iokstat->ks_data);
481 mutex_exit(&spa->spa_iokstat_lock);
486 vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
488 spa_t *spa = zio->io_spa;
489 ASSERT(MUTEX_HELD(&vq->vq_lock));
490 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
491 vq->vq_class[zio->io_priority].vqc_active--;
492 avl_remove(&vq->vq_active_tree, zio);
495 mutex_enter(&spa->spa_iokstat_lock);
496 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
497 spa->spa_queue_stats[zio->io_priority].spa_active--;
498 if (spa->spa_iokstat != NULL) {
499 kstat_io_t *ksio = spa->spa_iokstat->ks_data;
501 kstat_runq_exit(spa->spa_iokstat->ks_data);
502 if (zio->io_type == ZIO_TYPE_READ) {
504 ksio->nread += zio->io_size;
505 } else if (zio->io_type == ZIO_TYPE_WRITE) {
507 ksio->nwritten += zio->io_size;
510 mutex_exit(&spa->spa_iokstat_lock);
515 vdev_queue_agg_io_done(zio_t *aio)
517 if (aio->io_type == ZIO_TYPE_READ) {
519 zio_link_t *zl = NULL;
520 while ((pio = zio_walk_parents(aio, &zl)) != NULL) {
521 abd_copy_off(pio->io_abd, aio->io_abd,
522 0, pio->io_offset - aio->io_offset, pio->io_size);
526 abd_free(aio->io_abd);
530 vdev_queue_class_min_active(zio_priority_t p)
533 case ZIO_PRIORITY_SYNC_READ:
534 return (zfs_vdev_sync_read_min_active);
535 case ZIO_PRIORITY_SYNC_WRITE:
536 return (zfs_vdev_sync_write_min_active);
537 case ZIO_PRIORITY_ASYNC_READ:
538 return (zfs_vdev_async_read_min_active);
539 case ZIO_PRIORITY_ASYNC_WRITE:
540 return (zfs_vdev_async_write_min_active);
541 case ZIO_PRIORITY_SCRUB:
542 return (zfs_vdev_scrub_min_active);
543 case ZIO_PRIORITY_TRIM:
544 return (zfs_vdev_trim_min_active);
545 case ZIO_PRIORITY_REMOVAL:
546 return (zfs_vdev_removal_min_active);
547 case ZIO_PRIORITY_INITIALIZING:
548 return (zfs_vdev_initializing_min_active);
550 panic("invalid priority %u", p);
555 static __noinline int
556 vdev_queue_max_async_writes(spa_t *spa)
559 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total;
560 uint64_t min_bytes = zfs_dirty_data_max *
561 zfs_vdev_async_write_active_min_dirty_percent / 100;
562 uint64_t max_bytes = zfs_dirty_data_max *
563 zfs_vdev_async_write_active_max_dirty_percent / 100;
566 * Sync tasks correspond to interactive user actions. To reduce the
567 * execution time of those actions we push data out as fast as possible.
569 if (spa_has_pending_synctask(spa)) {
570 return (zfs_vdev_async_write_max_active);
573 if (dirty < min_bytes)
574 return (zfs_vdev_async_write_min_active);
575 if (dirty > max_bytes)
576 return (zfs_vdev_async_write_max_active);
579 * linear interpolation:
580 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
581 * move right by min_bytes
582 * move up by min_writes
584 writes = (dirty - min_bytes) *
585 (zfs_vdev_async_write_max_active -
586 zfs_vdev_async_write_min_active) /
587 (max_bytes - min_bytes) +
588 zfs_vdev_async_write_min_active;
589 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
590 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
595 vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
598 case ZIO_PRIORITY_SYNC_READ:
599 return (zfs_vdev_sync_read_max_active);
600 case ZIO_PRIORITY_SYNC_WRITE:
601 return (zfs_vdev_sync_write_max_active);
602 case ZIO_PRIORITY_ASYNC_READ:
603 return (zfs_vdev_async_read_max_active);
604 case ZIO_PRIORITY_ASYNC_WRITE:
605 return (vdev_queue_max_async_writes(spa));
606 case ZIO_PRIORITY_SCRUB:
607 return (zfs_vdev_scrub_max_active);
608 case ZIO_PRIORITY_TRIM:
609 return (zfs_vdev_trim_max_active);
610 case ZIO_PRIORITY_REMOVAL:
611 return (zfs_vdev_removal_max_active);
612 case ZIO_PRIORITY_INITIALIZING:
613 return (zfs_vdev_initializing_max_active);
615 panic("invalid priority %u", p);
621 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
622 * there is no eligible class.
624 static zio_priority_t
625 vdev_queue_class_to_issue(vdev_queue_t *vq)
627 spa_t *spa = vq->vq_vdev->vdev_spa;
630 ASSERT(MUTEX_HELD(&vq->vq_lock));
632 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
633 return (ZIO_PRIORITY_NUM_QUEUEABLE);
635 /* find a queue that has not reached its minimum # outstanding i/os */
636 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
637 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
638 vq->vq_class[p].vqc_active <
639 vdev_queue_class_min_active(p))
644 * If we haven't found a queue, look for one that hasn't reached its
645 * maximum # outstanding i/os.
647 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
648 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
649 vq->vq_class[p].vqc_active <
650 vdev_queue_class_max_active(spa, p))
654 /* No eligible queued i/os */
655 return (ZIO_PRIORITY_NUM_QUEUEABLE);
659 * Compute the range spanned by two i/os, which is the endpoint of the last
660 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
661 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
662 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
664 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
665 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
668 vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
670 zio_t *first, *last, *aio, *dio, *mandatory, *nio;
677 ASSERT(MUTEX_HELD(&vq->vq_lock));
679 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
684 if (zio->io_type == ZIO_TYPE_READ)
685 maxgap = zfs_vdev_read_gap_limit;
688 * We can aggregate I/Os that are sufficiently adjacent and of
689 * the same flavor, as expressed by the AGG_INHERIT flags.
690 * The latter requirement is necessary so that certain
691 * attributes of the I/O, such as whether it's a normal I/O
692 * or a scrub/resilver, can be preserved in the aggregate.
693 * We can include optional I/Os, but don't allow them
694 * to begin a range as they add no benefit in that situation.
698 * We keep track of the last non-optional I/O.
700 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
703 * Walk backwards through sufficiently contiguous I/Os
704 * recording the last non-optional I/O.
706 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
707 t = vdev_queue_type_tree(vq, zio->io_type);
708 while (t != NULL && (dio = AVL_PREV(t, first)) != NULL &&
709 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
710 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
711 IO_GAP(dio, first) <= maxgap &&
712 dio->io_type == zio->io_type) {
714 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
719 * Skip any initial optional I/Os.
721 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
722 first = AVL_NEXT(t, first);
723 ASSERT(first != NULL);
727 * Walk forward through sufficiently contiguous I/Os.
728 * The aggregation limit does not apply to optional i/os, so that
729 * we can issue contiguous writes even if they are larger than the
732 while ((dio = AVL_NEXT(t, last)) != NULL &&
733 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
734 (IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit ||
735 (dio->io_flags & ZIO_FLAG_OPTIONAL)) &&
736 IO_GAP(last, dio) <= maxgap &&
737 dio->io_type == zio->io_type) {
739 if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
744 * Now that we've established the range of the I/O aggregation
745 * we must decide what to do with trailing optional I/Os.
746 * For reads, there's nothing to do. While we are unable to
747 * aggregate further, it's possible that a trailing optional
748 * I/O would allow the underlying device to aggregate with
749 * subsequent I/Os. We must therefore determine if the next
750 * non-optional I/O is close enough to make aggregation
754 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
756 while ((dio = AVL_NEXT(t, nio)) != NULL &&
757 IO_GAP(nio, dio) == 0 &&
758 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
760 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
769 * We are going to include an optional io in our aggregated
770 * span, thus closing the write gap. Only mandatory i/os can
771 * start aggregated spans, so make sure that the next i/o
772 * after our span is mandatory.
774 dio = AVL_NEXT(t, last);
775 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
777 /* do not include the optional i/o */
778 while (last != mandatory && last != first) {
779 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
780 last = AVL_PREV(t, last);
781 ASSERT(last != NULL);
788 size = IO_SPAN(first, last);
789 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
791 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
792 abd_alloc_for_io(size, B_TRUE), size, first->io_type,
793 zio->io_priority, flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
794 vdev_queue_agg_io_done, NULL);
795 aio->io_timestamp = first->io_timestamp;
800 nio = AVL_NEXT(t, dio);
801 ASSERT3U(dio->io_type, ==, aio->io_type);
803 if (dio->io_flags & ZIO_FLAG_NODATA) {
804 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
805 abd_zero_off(aio->io_abd,
806 dio->io_offset - aio->io_offset, dio->io_size);
807 } else if (dio->io_type == ZIO_TYPE_WRITE) {
808 abd_copy_off(aio->io_abd, dio->io_abd,
809 dio->io_offset - aio->io_offset, 0, dio->io_size);
812 zio_add_child(dio, aio);
813 vdev_queue_io_remove(vq, dio);
814 zio_vdev_io_bypass(dio);
816 } while (dio != last);
822 vdev_queue_io_to_issue(vdev_queue_t *vq)
831 ASSERT(MUTEX_HELD(&vq->vq_lock));
833 p = vdev_queue_class_to_issue(vq);
835 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
836 /* No eligible queued i/os */
841 * For LBA-ordered queues (async / scrub / initializing), issue the
842 * i/o which follows the most recently issued i/o in LBA (offset) order.
844 * For FIFO queues (sync), issue the i/o with the lowest timestamp.
846 tree = vdev_queue_class_tree(vq, p);
847 search.io_timestamp = 0;
848 search.io_offset = vq->vq_last_offset + 1;
849 VERIFY3P(avl_find(tree, &search, &idx), ==, NULL);
850 zio = avl_nearest(tree, idx, AVL_AFTER);
852 zio = avl_first(tree);
853 ASSERT3U(zio->io_priority, ==, p);
855 aio = vdev_queue_aggregate(vq, zio);
859 vdev_queue_io_remove(vq, zio);
862 * If the I/O is or was optional and therefore has no data, we need to
863 * simply discard it. We need to drop the vdev queue's lock to avoid a
864 * deadlock that we could encounter since this I/O will complete
867 if (zio->io_flags & ZIO_FLAG_NODATA) {
868 mutex_exit(&vq->vq_lock);
869 zio_vdev_io_bypass(zio);
871 mutex_enter(&vq->vq_lock);
875 vdev_queue_pending_add(vq, zio);
876 vq->vq_last_offset = zio->io_offset;
882 vdev_queue_io(zio_t *zio)
884 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
887 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
891 * Children i/os inherent their parent's priority, which might
892 * not match the child's i/o type. Fix it up here.
894 if (zio->io_type == ZIO_TYPE_READ) {
895 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
896 zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
897 zio->io_priority != ZIO_PRIORITY_SCRUB &&
898 zio->io_priority != ZIO_PRIORITY_REMOVAL &&
899 zio->io_priority != ZIO_PRIORITY_INITIALIZING)
900 zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
901 } else if (zio->io_type == ZIO_TYPE_WRITE) {
902 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
903 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE &&
904 zio->io_priority != ZIO_PRIORITY_REMOVAL &&
905 zio->io_priority != ZIO_PRIORITY_INITIALIZING)
906 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
908 ASSERT(zio->io_type == ZIO_TYPE_FREE);
909 zio->io_priority = ZIO_PRIORITY_TRIM;
912 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
914 mutex_enter(&vq->vq_lock);
915 zio->io_timestamp = gethrtime();
916 vdev_queue_io_add(vq, zio);
917 nio = vdev_queue_io_to_issue(vq);
918 mutex_exit(&vq->vq_lock);
923 if (nio->io_done == vdev_queue_agg_io_done) {
932 vdev_queue_io_done(zio_t *zio)
934 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
937 mutex_enter(&vq->vq_lock);
939 vdev_queue_pending_remove(vq, zio);
941 vq->vq_io_complete_ts = gethrtime();
943 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
944 mutex_exit(&vq->vq_lock);
945 if (nio->io_done == vdev_queue_agg_io_done) {
948 zio_vdev_io_reissue(nio);
951 mutex_enter(&vq->vq_lock);
954 mutex_exit(&vq->vq_lock);
958 vdev_queue_change_io_priority(zio_t *zio, zio_priority_t priority)
960 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
963 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
964 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
966 if (zio->io_type == ZIO_TYPE_READ) {
967 if (priority != ZIO_PRIORITY_SYNC_READ &&
968 priority != ZIO_PRIORITY_ASYNC_READ &&
969 priority != ZIO_PRIORITY_SCRUB)
970 priority = ZIO_PRIORITY_ASYNC_READ;
972 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
973 if (priority != ZIO_PRIORITY_SYNC_WRITE &&
974 priority != ZIO_PRIORITY_ASYNC_WRITE)
975 priority = ZIO_PRIORITY_ASYNC_WRITE;
978 mutex_enter(&vq->vq_lock);
981 * If the zio is in none of the queues we can simply change
982 * the priority. If the zio is waiting to be submitted we must
983 * remove it from the queue and re-insert it with the new priority.
984 * Otherwise, the zio is currently active and we cannot change its
987 tree = vdev_queue_class_tree(vq, zio->io_priority);
988 if (avl_find(tree, zio, NULL) == zio) {
989 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
990 zio->io_priority = priority;
991 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
992 } else if (avl_find(&vq->vq_active_tree, zio, NULL) != zio) {
993 zio->io_priority = priority;
996 mutex_exit(&vq->vq_lock);
1000 * As these three methods are only used for load calculations we're not concerned
1001 * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex
1002 * use here, instead we prefer to keep it lock free for performance.
1005 vdev_queue_length(vdev_t *vd)
1007 return (avl_numnodes(&vd->vdev_queue.vq_active_tree));
1011 vdev_queue_lastoffset(vdev_t *vd)
1013 return (vd->vdev_queue.vq_lastoffset);
1017 vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio)
1019 vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size;