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, 2014 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>
43 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The
44 * I/O scheduler determines when and in what order those operations are
45 * issued. The I/O scheduler divides operations into six I/O classes
46 * prioritized in the following order: sync read, sync write, async read,
47 * async write, scrub/resilver and trim. Each queue defines the minimum and
48 * maximum number of concurrent operations that may be issued to the device.
49 * In addition, the device has an aggregate maximum. Note that the sum of the
50 * per-queue minimums must not exceed the aggregate maximum, and if the
51 * aggregate maximum is equal to or greater than the sum of the per-queue
52 * maximums, the per-queue minimum has no effect.
54 * For many physical devices, throughput increases with the number of
55 * concurrent operations, but latency typically suffers. Further, physical
56 * devices typically have a limit at which more concurrent operations have no
57 * effect on throughput or can actually cause it to decrease.
59 * The scheduler selects the next operation to issue by first looking for an
60 * I/O class whose minimum has not been satisfied. Once all are satisfied and
61 * the aggregate maximum has not been hit, the scheduler looks for classes
62 * whose maximum has not been satisfied. Iteration through the I/O classes is
63 * done in the order specified above. No further operations are issued if the
64 * aggregate maximum number of concurrent operations has been hit or if there
65 * are no operations queued for an I/O class that has not hit its maximum.
66 * Every time an I/O is queued or an operation completes, the I/O scheduler
67 * looks for new operations to issue.
69 * All I/O classes have a fixed maximum number of outstanding operations
70 * except for the async write class. Asynchronous writes represent the data
71 * that is committed to stable storage during the syncing stage for
72 * transaction groups (see txg.c). Transaction groups enter the syncing state
73 * periodically so the number of queued async writes will quickly burst up and
74 * then bleed down to zero. Rather than servicing them as quickly as possible,
75 * the I/O scheduler changes the maximum number of active async write I/Os
76 * according to the amount of dirty data in the pool (see dsl_pool.c). Since
77 * both throughput and latency typically increase with the number of
78 * concurrent operations issued to physical devices, reducing the burstiness
79 * in the number of concurrent operations also stabilizes the response time of
80 * operations from other -- and in particular synchronous -- queues. In broad
81 * strokes, the I/O scheduler will issue more concurrent operations from the
82 * async write queue as there's more dirty data in the pool.
86 * The number of concurrent operations issued for the async write I/O class
87 * follows a piece-wise linear function defined by a few adjustable points.
89 * | o---------| <-- zfs_vdev_async_write_max_active
96 * |------------o | | <-- zfs_vdev_async_write_min_active
97 * 0|____________^______|_________|
98 * 0% | | 100% of zfs_dirty_data_max
100 * | `-- zfs_vdev_async_write_active_max_dirty_percent
101 * `--------- zfs_vdev_async_write_active_min_dirty_percent
103 * Until the amount of dirty data exceeds a minimum percentage of the dirty
104 * data allowed in the pool, the I/O scheduler will limit the number of
105 * concurrent operations to the minimum. As that threshold is crossed, the
106 * number of concurrent operations issued increases linearly to the maximum at
107 * the specified maximum percentage of the dirty data allowed in the pool.
109 * Ideally, the amount of dirty data on a busy pool will stay in the sloped
110 * part of the function between zfs_vdev_async_write_active_min_dirty_percent
111 * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the
112 * maximum percentage, this indicates that the rate of incoming data is
113 * greater than the rate that the backend storage can handle. In this case, we
114 * must further throttle incoming writes (see dmu_tx_delay() for details).
118 * The maximum number of I/Os active to each device. Ideally, this will be >=
119 * the sum of each queue's max_active. It must be at least the sum of each
120 * queue's min_active.
122 uint32_t zfs_vdev_max_active = 1000;
125 * Per-queue limits on the number of I/Os active to each device. If the
126 * sum of the queue's max_active is < zfs_vdev_max_active, then the
127 * min_active comes into play. We will send min_active from each queue,
128 * and then select from queues in the order defined by zio_priority_t.
130 * In general, smaller max_active's will lead to lower latency of synchronous
131 * operations. Larger max_active's may lead to higher overall throughput,
132 * depending on underlying storage.
134 * The ratio of the queues' max_actives determines the balance of performance
135 * between reads, writes, and scrubs. E.g., increasing
136 * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete
137 * more quickly, but reads and writes to have higher latency and lower
140 uint32_t zfs_vdev_sync_read_min_active = 10;
141 uint32_t zfs_vdev_sync_read_max_active = 10;
142 uint32_t zfs_vdev_sync_write_min_active = 10;
143 uint32_t zfs_vdev_sync_write_max_active = 10;
144 uint32_t zfs_vdev_async_read_min_active = 1;
145 uint32_t zfs_vdev_async_read_max_active = 3;
146 uint32_t zfs_vdev_async_write_min_active = 1;
147 uint32_t zfs_vdev_async_write_max_active = 10;
148 uint32_t zfs_vdev_scrub_min_active = 1;
149 uint32_t zfs_vdev_scrub_max_active = 2;
150 uint32_t zfs_vdev_trim_min_active = 1;
152 * TRIM max active is large in comparison to the other values due to the fact
153 * that TRIM IOs are coalesced at the device layer. This value is set such
154 * that a typical SSD can process the queued IOs in a single request.
156 uint32_t zfs_vdev_trim_max_active = 64;
160 * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent
161 * dirty data, use zfs_vdev_async_write_min_active. When it has more than
162 * zfs_vdev_async_write_active_max_dirty_percent, use
163 * zfs_vdev_async_write_max_active. The value is linearly interpolated
164 * between min and max.
166 int zfs_vdev_async_write_active_min_dirty_percent = 30;
167 int zfs_vdev_async_write_active_max_dirty_percent = 60;
170 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
171 * For read I/Os, we also aggregate across small adjacency gaps; for writes
172 * we include spans of optional I/Os to aid aggregation at the disk even when
173 * they aren't able to help us aggregate at this level.
175 int zfs_vdev_aggregation_limit = SPA_OLD_MAXBLOCKSIZE;
176 int zfs_vdev_read_gap_limit = 32 << 10;
177 int zfs_vdev_write_gap_limit = 4 << 10;
180 * Define the queue depth percentage for each top-level. This percentage is
181 * used in conjunction with zfs_vdev_async_max_active to determine how many
182 * allocations a specific top-level vdev should handle. Once the queue depth
183 * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100
184 * then allocator will stop allocating blocks on that top-level device.
185 * The default kernel setting is 1000% which will yield 100 allocations per
186 * device. For userland testing, the default setting is 300% which equates
187 * to 30 allocations per device.
190 int zfs_vdev_queue_depth_pct = 1000;
192 int zfs_vdev_queue_depth_pct = 300;
198 SYSCTL_DECL(_vfs_zfs_vdev);
200 static int sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS);
201 SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_min_dirty_percent,
202 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
203 sysctl_zfs_async_write_active_min_dirty_percent, "I",
204 "Percentage of async write dirty data below which "
205 "async_write_min_active is used.");
207 static int sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS);
208 SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_max_dirty_percent,
209 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
210 sysctl_zfs_async_write_active_max_dirty_percent, "I",
211 "Percentage of async write dirty data above which "
212 "async_write_max_active is used.");
214 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RWTUN,
215 &zfs_vdev_max_active, 0,
216 "The maximum number of I/Os of all types active for each device.");
218 #define ZFS_VDEV_QUEUE_KNOB_MIN(name) \
219 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, CTLFLAG_RWTUN,\
220 &zfs_vdev_ ## name ## _min_active, 0, \
221 "Initial number of I/O requests of type " #name \
222 " active for each device");
224 #define ZFS_VDEV_QUEUE_KNOB_MAX(name) \
225 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, CTLFLAG_RWTUN,\
226 &zfs_vdev_ ## name ## _max_active, 0, \
227 "Maximum number of I/O requests of type " #name \
228 " active for each device");
230 ZFS_VDEV_QUEUE_KNOB_MIN(sync_read);
231 ZFS_VDEV_QUEUE_KNOB_MAX(sync_read);
232 ZFS_VDEV_QUEUE_KNOB_MIN(sync_write);
233 ZFS_VDEV_QUEUE_KNOB_MAX(sync_write);
234 ZFS_VDEV_QUEUE_KNOB_MIN(async_read);
235 ZFS_VDEV_QUEUE_KNOB_MAX(async_read);
236 ZFS_VDEV_QUEUE_KNOB_MIN(async_write);
237 ZFS_VDEV_QUEUE_KNOB_MAX(async_write);
238 ZFS_VDEV_QUEUE_KNOB_MIN(scrub);
239 ZFS_VDEV_QUEUE_KNOB_MAX(scrub);
240 ZFS_VDEV_QUEUE_KNOB_MIN(trim);
241 ZFS_VDEV_QUEUE_KNOB_MAX(trim);
243 #undef ZFS_VDEV_QUEUE_KNOB
245 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RWTUN,
246 &zfs_vdev_aggregation_limit, 0,
247 "I/O requests are aggregated up to this size");
248 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RWTUN,
249 &zfs_vdev_read_gap_limit, 0,
250 "Acceptable gap between two reads being aggregated");
251 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RWTUN,
252 &zfs_vdev_write_gap_limit, 0,
253 "Acceptable gap between two writes being aggregated");
254 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, queue_depth_pct, CTLFLAG_RWTUN,
255 &zfs_vdev_queue_depth_pct, 0,
256 "Queue depth percentage for each top-level");
259 sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS)
263 val = zfs_vdev_async_write_active_min_dirty_percent;
264 err = sysctl_handle_int(oidp, &val, 0, req);
265 if (err != 0 || req->newptr == NULL)
268 if (val < 0 || val > 100 ||
269 val >= zfs_vdev_async_write_active_max_dirty_percent)
272 zfs_vdev_async_write_active_min_dirty_percent = val;
278 sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS)
282 val = zfs_vdev_async_write_active_max_dirty_percent;
283 err = sysctl_handle_int(oidp, &val, 0, req);
284 if (err != 0 || req->newptr == NULL)
287 if (val < 0 || val > 100 ||
288 val <= zfs_vdev_async_write_active_min_dirty_percent)
291 zfs_vdev_async_write_active_max_dirty_percent = val;
299 vdev_queue_offset_compare(const void *x1, const void *x2)
301 const zio_t *z1 = x1;
302 const zio_t *z2 = x2;
304 if (z1->io_offset < z2->io_offset)
306 if (z1->io_offset > z2->io_offset)
317 static inline avl_tree_t *
318 vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p)
320 return (&vq->vq_class[p].vqc_queued_tree);
323 static inline avl_tree_t *
324 vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t)
326 if (t == ZIO_TYPE_READ)
327 return (&vq->vq_read_offset_tree);
328 else if (t == ZIO_TYPE_WRITE)
329 return (&vq->vq_write_offset_tree);
335 vdev_queue_timestamp_compare(const void *x1, const void *x2)
337 const zio_t *z1 = x1;
338 const zio_t *z2 = x2;
340 if (z1->io_timestamp < z2->io_timestamp)
342 if (z1->io_timestamp > z2->io_timestamp)
345 if (z1->io_offset < z2->io_offset)
347 if (z1->io_offset > z2->io_offset)
359 vdev_queue_init(vdev_t *vd)
361 vdev_queue_t *vq = &vd->vdev_queue;
363 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
366 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
367 sizeof (zio_t), offsetof(struct zio, io_queue_node));
368 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ),
369 vdev_queue_offset_compare, sizeof (zio_t),
370 offsetof(struct zio, io_offset_node));
371 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE),
372 vdev_queue_offset_compare, sizeof (zio_t),
373 offsetof(struct zio, io_offset_node));
375 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
376 int (*compfn) (const void *, const void *);
379 * The synchronous i/o queues are dispatched in FIFO rather
380 * than LBA order. This provides more consistent latency for
383 if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE)
384 compfn = vdev_queue_timestamp_compare;
386 compfn = vdev_queue_offset_compare;
388 avl_create(vdev_queue_class_tree(vq, p), compfn,
389 sizeof (zio_t), offsetof(struct zio, io_queue_node));
392 vq->vq_lastoffset = 0;
396 vdev_queue_fini(vdev_t *vd)
398 vdev_queue_t *vq = &vd->vdev_queue;
400 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
401 avl_destroy(vdev_queue_class_tree(vq, p));
402 avl_destroy(&vq->vq_active_tree);
403 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
404 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE));
406 mutex_destroy(&vq->vq_lock);
410 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
412 spa_t *spa = zio->io_spa;
415 ASSERT(MUTEX_HELD(&vq->vq_lock));
416 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
417 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
418 qtt = vdev_queue_type_tree(vq, zio->io_type);
423 mutex_enter(&spa->spa_iokstat_lock);
424 spa->spa_queue_stats[zio->io_priority].spa_queued++;
425 if (spa->spa_iokstat != NULL)
426 kstat_waitq_enter(spa->spa_iokstat->ks_data);
427 mutex_exit(&spa->spa_iokstat_lock);
432 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
434 spa_t *spa = zio->io_spa;
437 ASSERT(MUTEX_HELD(&vq->vq_lock));
438 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
439 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
440 qtt = vdev_queue_type_tree(vq, zio->io_type);
442 avl_remove(qtt, zio);
445 mutex_enter(&spa->spa_iokstat_lock);
446 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
447 spa->spa_queue_stats[zio->io_priority].spa_queued--;
448 if (spa->spa_iokstat != NULL)
449 kstat_waitq_exit(spa->spa_iokstat->ks_data);
450 mutex_exit(&spa->spa_iokstat_lock);
455 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
457 spa_t *spa = zio->io_spa;
458 ASSERT(MUTEX_HELD(&vq->vq_lock));
459 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
460 vq->vq_class[zio->io_priority].vqc_active++;
461 avl_add(&vq->vq_active_tree, zio);
464 mutex_enter(&spa->spa_iokstat_lock);
465 spa->spa_queue_stats[zio->io_priority].spa_active++;
466 if (spa->spa_iokstat != NULL)
467 kstat_runq_enter(spa->spa_iokstat->ks_data);
468 mutex_exit(&spa->spa_iokstat_lock);
473 vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
475 spa_t *spa = zio->io_spa;
476 ASSERT(MUTEX_HELD(&vq->vq_lock));
477 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
478 vq->vq_class[zio->io_priority].vqc_active--;
479 avl_remove(&vq->vq_active_tree, zio);
482 mutex_enter(&spa->spa_iokstat_lock);
483 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
484 spa->spa_queue_stats[zio->io_priority].spa_active--;
485 if (spa->spa_iokstat != NULL) {
486 kstat_io_t *ksio = spa->spa_iokstat->ks_data;
488 kstat_runq_exit(spa->spa_iokstat->ks_data);
489 if (zio->io_type == ZIO_TYPE_READ) {
491 ksio->nread += zio->io_size;
492 } else if (zio->io_type == ZIO_TYPE_WRITE) {
494 ksio->nwritten += zio->io_size;
497 mutex_exit(&spa->spa_iokstat_lock);
502 vdev_queue_agg_io_done(zio_t *aio)
504 if (aio->io_type == ZIO_TYPE_READ) {
506 zio_link_t *zl = NULL;
507 while ((pio = zio_walk_parents(aio, &zl)) != NULL) {
508 bcopy((char *)aio->io_data + (pio->io_offset -
509 aio->io_offset), pio->io_data, pio->io_size);
513 zio_buf_free(aio->io_data, aio->io_size);
517 vdev_queue_class_min_active(zio_priority_t p)
520 case ZIO_PRIORITY_SYNC_READ:
521 return (zfs_vdev_sync_read_min_active);
522 case ZIO_PRIORITY_SYNC_WRITE:
523 return (zfs_vdev_sync_write_min_active);
524 case ZIO_PRIORITY_ASYNC_READ:
525 return (zfs_vdev_async_read_min_active);
526 case ZIO_PRIORITY_ASYNC_WRITE:
527 return (zfs_vdev_async_write_min_active);
528 case ZIO_PRIORITY_SCRUB:
529 return (zfs_vdev_scrub_min_active);
530 case ZIO_PRIORITY_TRIM:
531 return (zfs_vdev_trim_min_active);
533 panic("invalid priority %u", p);
538 static __noinline int
539 vdev_queue_max_async_writes(spa_t *spa)
542 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total;
543 uint64_t min_bytes = zfs_dirty_data_max *
544 zfs_vdev_async_write_active_min_dirty_percent / 100;
545 uint64_t max_bytes = zfs_dirty_data_max *
546 zfs_vdev_async_write_active_max_dirty_percent / 100;
549 * Sync tasks correspond to interactive user actions. To reduce the
550 * execution time of those actions we push data out as fast as possible.
552 if (spa_has_pending_synctask(spa)) {
553 return (zfs_vdev_async_write_max_active);
556 if (dirty < min_bytes)
557 return (zfs_vdev_async_write_min_active);
558 if (dirty > max_bytes)
559 return (zfs_vdev_async_write_max_active);
562 * linear interpolation:
563 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
564 * move right by min_bytes
565 * move up by min_writes
567 writes = (dirty - min_bytes) *
568 (zfs_vdev_async_write_max_active -
569 zfs_vdev_async_write_min_active) /
570 (max_bytes - min_bytes) +
571 zfs_vdev_async_write_min_active;
572 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
573 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
578 vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
581 case ZIO_PRIORITY_SYNC_READ:
582 return (zfs_vdev_sync_read_max_active);
583 case ZIO_PRIORITY_SYNC_WRITE:
584 return (zfs_vdev_sync_write_max_active);
585 case ZIO_PRIORITY_ASYNC_READ:
586 return (zfs_vdev_async_read_max_active);
587 case ZIO_PRIORITY_ASYNC_WRITE:
588 return (vdev_queue_max_async_writes(spa));
589 case ZIO_PRIORITY_SCRUB:
590 return (zfs_vdev_scrub_max_active);
591 case ZIO_PRIORITY_TRIM:
592 return (zfs_vdev_trim_max_active);
594 panic("invalid priority %u", p);
600 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
601 * there is no eligible class.
603 static zio_priority_t
604 vdev_queue_class_to_issue(vdev_queue_t *vq)
606 spa_t *spa = vq->vq_vdev->vdev_spa;
609 ASSERT(MUTEX_HELD(&vq->vq_lock));
611 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
612 return (ZIO_PRIORITY_NUM_QUEUEABLE);
614 /* find a queue that has not reached its minimum # outstanding i/os */
615 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
616 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
617 vq->vq_class[p].vqc_active <
618 vdev_queue_class_min_active(p))
623 * If we haven't found a queue, look for one that hasn't reached its
624 * maximum # outstanding i/os.
626 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
627 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
628 vq->vq_class[p].vqc_active <
629 vdev_queue_class_max_active(spa, p))
633 /* No eligible queued i/os */
634 return (ZIO_PRIORITY_NUM_QUEUEABLE);
638 * Compute the range spanned by two i/os, which is the endpoint of the last
639 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
640 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
641 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
643 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
644 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
647 vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
649 zio_t *first, *last, *aio, *dio, *mandatory, *nio;
656 ASSERT(MUTEX_HELD(&vq->vq_lock));
658 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
663 if (zio->io_type == ZIO_TYPE_READ)
664 maxgap = zfs_vdev_read_gap_limit;
667 * We can aggregate I/Os that are sufficiently adjacent and of
668 * the same flavor, as expressed by the AGG_INHERIT flags.
669 * The latter requirement is necessary so that certain
670 * attributes of the I/O, such as whether it's a normal I/O
671 * or a scrub/resilver, can be preserved in the aggregate.
672 * We can include optional I/Os, but don't allow them
673 * to begin a range as they add no benefit in that situation.
677 * We keep track of the last non-optional I/O.
679 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
682 * Walk backwards through sufficiently contiguous I/Os
683 * recording the last non-option I/O.
685 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
686 t = vdev_queue_type_tree(vq, zio->io_type);
687 while (t != NULL && (dio = AVL_PREV(t, first)) != NULL &&
688 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
689 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
690 IO_GAP(dio, first) <= maxgap) {
692 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
697 * Skip any initial optional I/Os.
699 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
700 first = AVL_NEXT(t, first);
701 ASSERT(first != NULL);
705 * Walk forward through sufficiently contiguous I/Os.
707 while ((dio = AVL_NEXT(t, last)) != NULL &&
708 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
709 IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit &&
710 IO_GAP(last, dio) <= maxgap) {
712 if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
717 * Now that we've established the range of the I/O aggregation
718 * we must decide what to do with trailing optional I/Os.
719 * For reads, there's nothing to do. While we are unable to
720 * aggregate further, it's possible that a trailing optional
721 * I/O would allow the underlying device to aggregate with
722 * subsequent I/Os. We must therefore determine if the next
723 * non-optional I/O is close enough to make aggregation
727 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
729 while ((dio = AVL_NEXT(t, nio)) != NULL &&
730 IO_GAP(nio, dio) == 0 &&
731 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
733 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
741 /* This may be a no-op. */
742 dio = AVL_NEXT(t, last);
743 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
745 while (last != mandatory && last != first) {
746 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
747 last = AVL_PREV(t, last);
748 ASSERT(last != NULL);
755 size = IO_SPAN(first, last);
756 ASSERT3U(size, <=, zfs_vdev_aggregation_limit);
758 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
759 zio_buf_alloc(size), size, first->io_type, zio->io_priority,
760 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
761 vdev_queue_agg_io_done, NULL);
762 aio->io_timestamp = first->io_timestamp;
767 nio = AVL_NEXT(t, dio);
768 ASSERT3U(dio->io_type, ==, aio->io_type);
770 if (dio->io_flags & ZIO_FLAG_NODATA) {
771 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
772 bzero((char *)aio->io_data + (dio->io_offset -
773 aio->io_offset), dio->io_size);
774 } else if (dio->io_type == ZIO_TYPE_WRITE) {
775 bcopy(dio->io_data, (char *)aio->io_data +
776 (dio->io_offset - aio->io_offset),
780 zio_add_child(dio, aio);
781 vdev_queue_io_remove(vq, dio);
782 zio_vdev_io_bypass(dio);
784 } while (dio != last);
790 vdev_queue_io_to_issue(vdev_queue_t *vq)
799 ASSERT(MUTEX_HELD(&vq->vq_lock));
801 p = vdev_queue_class_to_issue(vq);
803 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
804 /* No eligible queued i/os */
809 * For LBA-ordered queues (async / scrub), issue the i/o which follows
810 * the most recently issued i/o in LBA (offset) order.
812 * For FIFO queues (sync), issue the i/o with the lowest timestamp.
814 tree = vdev_queue_class_tree(vq, p);
815 search.io_timestamp = 0;
816 search.io_offset = vq->vq_last_offset + 1;
817 VERIFY3P(avl_find(tree, &search, &idx), ==, NULL);
818 zio = avl_nearest(tree, idx, AVL_AFTER);
820 zio = avl_first(tree);
821 ASSERT3U(zio->io_priority, ==, p);
823 aio = vdev_queue_aggregate(vq, zio);
827 vdev_queue_io_remove(vq, zio);
830 * If the I/O is or was optional and therefore has no data, we need to
831 * simply discard it. We need to drop the vdev queue's lock to avoid a
832 * deadlock that we could encounter since this I/O will complete
835 if (zio->io_flags & ZIO_FLAG_NODATA) {
836 mutex_exit(&vq->vq_lock);
837 zio_vdev_io_bypass(zio);
839 mutex_enter(&vq->vq_lock);
843 vdev_queue_pending_add(vq, zio);
844 vq->vq_last_offset = zio->io_offset;
850 vdev_queue_io(zio_t *zio)
852 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
855 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
859 * Children i/os inherent their parent's priority, which might
860 * not match the child's i/o type. Fix it up here.
862 if (zio->io_type == ZIO_TYPE_READ) {
863 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
864 zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
865 zio->io_priority != ZIO_PRIORITY_SCRUB)
866 zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
867 } else if (zio->io_type == ZIO_TYPE_WRITE) {
868 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
869 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE)
870 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
872 ASSERT(zio->io_type == ZIO_TYPE_FREE);
873 zio->io_priority = ZIO_PRIORITY_TRIM;
876 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
878 mutex_enter(&vq->vq_lock);
879 zio->io_timestamp = gethrtime();
880 vdev_queue_io_add(vq, zio);
881 nio = vdev_queue_io_to_issue(vq);
882 mutex_exit(&vq->vq_lock);
887 if (nio->io_done == vdev_queue_agg_io_done) {
896 vdev_queue_io_done(zio_t *zio)
898 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
901 mutex_enter(&vq->vq_lock);
903 vdev_queue_pending_remove(vq, zio);
905 vq->vq_io_complete_ts = gethrtime();
907 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
908 mutex_exit(&vq->vq_lock);
909 if (nio->io_done == vdev_queue_agg_io_done) {
912 zio_vdev_io_reissue(nio);
915 mutex_enter(&vq->vq_lock);
918 mutex_exit(&vq->vq_lock);
922 * As these three methods are only used for load calculations we're not concerned
923 * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex
924 * use here, instead we prefer to keep it lock free for performance.
927 vdev_queue_length(vdev_t *vd)
929 return (avl_numnodes(&vd->vdev_queue.vq_active_tree));
933 vdev_queue_lastoffset(vdev_t *vd)
935 return (vd->vdev_queue.vq_lastoffset);
939 vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio)
941 vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size;