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, 2017 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;
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_read_gap_limit = 32 << 10;
180 int zfs_vdev_write_gap_limit = 4 << 10;
183 * Define the queue depth percentage for each top-level. This percentage is
184 * used in conjunction with zfs_vdev_async_max_active to determine how many
185 * allocations a specific top-level vdev should handle. Once the queue depth
186 * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100
187 * then allocator will stop allocating blocks on that top-level device.
188 * The default kernel setting is 1000% which will yield 100 allocations per
189 * device. For userland testing, the default setting is 300% which equates
190 * to 30 allocations per device.
193 int zfs_vdev_queue_depth_pct = 1000;
195 int zfs_vdev_queue_depth_pct = 300;
201 SYSCTL_DECL(_vfs_zfs_vdev);
203 static int sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS);
204 SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_min_dirty_percent,
205 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
206 sysctl_zfs_async_write_active_min_dirty_percent, "I",
207 "Percentage of async write dirty data below which "
208 "async_write_min_active is used.");
210 static int sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS);
211 SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_max_dirty_percent,
212 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
213 sysctl_zfs_async_write_active_max_dirty_percent, "I",
214 "Percentage of async write dirty data above which "
215 "async_write_max_active is used.");
217 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RWTUN,
218 &zfs_vdev_max_active, 0,
219 "The maximum number of I/Os of all types active for each device.");
221 #define ZFS_VDEV_QUEUE_KNOB_MIN(name) \
222 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, CTLFLAG_RWTUN,\
223 &zfs_vdev_ ## name ## _min_active, 0, \
224 "Initial number of I/O requests of type " #name \
225 " active for each device");
227 #define ZFS_VDEV_QUEUE_KNOB_MAX(name) \
228 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, CTLFLAG_RWTUN,\
229 &zfs_vdev_ ## name ## _max_active, 0, \
230 "Maximum number of I/O requests of type " #name \
231 " active for each device");
233 ZFS_VDEV_QUEUE_KNOB_MIN(sync_read);
234 ZFS_VDEV_QUEUE_KNOB_MAX(sync_read);
235 ZFS_VDEV_QUEUE_KNOB_MIN(sync_write);
236 ZFS_VDEV_QUEUE_KNOB_MAX(sync_write);
237 ZFS_VDEV_QUEUE_KNOB_MIN(async_read);
238 ZFS_VDEV_QUEUE_KNOB_MAX(async_read);
239 ZFS_VDEV_QUEUE_KNOB_MIN(async_write);
240 ZFS_VDEV_QUEUE_KNOB_MAX(async_write);
241 ZFS_VDEV_QUEUE_KNOB_MIN(scrub);
242 ZFS_VDEV_QUEUE_KNOB_MAX(scrub);
243 ZFS_VDEV_QUEUE_KNOB_MIN(trim);
244 ZFS_VDEV_QUEUE_KNOB_MAX(trim);
246 #undef ZFS_VDEV_QUEUE_KNOB
248 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RWTUN,
249 &zfs_vdev_aggregation_limit, 0,
250 "I/O requests are aggregated up to this size");
251 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RWTUN,
252 &zfs_vdev_read_gap_limit, 0,
253 "Acceptable gap between two reads being aggregated");
254 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RWTUN,
255 &zfs_vdev_write_gap_limit, 0,
256 "Acceptable gap between two writes being aggregated");
257 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, queue_depth_pct, CTLFLAG_RWTUN,
258 &zfs_vdev_queue_depth_pct, 0,
259 "Queue depth percentage for each top-level");
262 sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS)
266 val = zfs_vdev_async_write_active_min_dirty_percent;
267 err = sysctl_handle_int(oidp, &val, 0, req);
268 if (err != 0 || req->newptr == NULL)
271 if (val < 0 || val > 100 ||
272 val >= zfs_vdev_async_write_active_max_dirty_percent)
275 zfs_vdev_async_write_active_min_dirty_percent = val;
281 sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS)
285 val = zfs_vdev_async_write_active_max_dirty_percent;
286 err = sysctl_handle_int(oidp, &val, 0, req);
287 if (err != 0 || req->newptr == NULL)
290 if (val < 0 || val > 100 ||
291 val <= zfs_vdev_async_write_active_min_dirty_percent)
294 zfs_vdev_async_write_active_max_dirty_percent = val;
302 vdev_queue_offset_compare(const void *x1, const void *x2)
304 const zio_t *z1 = x1;
305 const zio_t *z2 = x2;
307 if (z1->io_offset < z2->io_offset)
309 if (z1->io_offset > z2->io_offset)
320 static inline avl_tree_t *
321 vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p)
323 return (&vq->vq_class[p].vqc_queued_tree);
326 static inline avl_tree_t *
327 vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t)
329 if (t == ZIO_TYPE_READ)
330 return (&vq->vq_read_offset_tree);
331 else if (t == ZIO_TYPE_WRITE)
332 return (&vq->vq_write_offset_tree);
338 vdev_queue_timestamp_compare(const void *x1, const void *x2)
340 const zio_t *z1 = x1;
341 const zio_t *z2 = x2;
343 if (z1->io_timestamp < z2->io_timestamp)
345 if (z1->io_timestamp > z2->io_timestamp)
348 if (z1->io_offset < z2->io_offset)
350 if (z1->io_offset > z2->io_offset)
362 vdev_queue_init(vdev_t *vd)
364 vdev_queue_t *vq = &vd->vdev_queue;
366 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
369 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
370 sizeof (zio_t), offsetof(struct zio, io_queue_node));
371 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ),
372 vdev_queue_offset_compare, sizeof (zio_t),
373 offsetof(struct zio, io_offset_node));
374 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE),
375 vdev_queue_offset_compare, sizeof (zio_t),
376 offsetof(struct zio, io_offset_node));
378 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
379 int (*compfn) (const void *, const void *);
382 * The synchronous i/o queues are dispatched in FIFO rather
383 * than LBA order. This provides more consistent latency for
386 if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE)
387 compfn = vdev_queue_timestamp_compare;
389 compfn = vdev_queue_offset_compare;
391 avl_create(vdev_queue_class_tree(vq, p), compfn,
392 sizeof (zio_t), offsetof(struct zio, io_queue_node));
395 vq->vq_lastoffset = 0;
399 vdev_queue_fini(vdev_t *vd)
401 vdev_queue_t *vq = &vd->vdev_queue;
403 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
404 avl_destroy(vdev_queue_class_tree(vq, p));
405 avl_destroy(&vq->vq_active_tree);
406 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
407 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE));
409 mutex_destroy(&vq->vq_lock);
413 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
415 spa_t *spa = zio->io_spa;
418 ASSERT(MUTEX_HELD(&vq->vq_lock));
419 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
420 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
421 qtt = vdev_queue_type_tree(vq, zio->io_type);
426 mutex_enter(&spa->spa_iokstat_lock);
427 spa->spa_queue_stats[zio->io_priority].spa_queued++;
428 if (spa->spa_iokstat != NULL)
429 kstat_waitq_enter(spa->spa_iokstat->ks_data);
430 mutex_exit(&spa->spa_iokstat_lock);
435 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
437 spa_t *spa = zio->io_spa;
440 ASSERT(MUTEX_HELD(&vq->vq_lock));
441 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
442 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
443 qtt = vdev_queue_type_tree(vq, zio->io_type);
445 avl_remove(qtt, zio);
448 mutex_enter(&spa->spa_iokstat_lock);
449 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
450 spa->spa_queue_stats[zio->io_priority].spa_queued--;
451 if (spa->spa_iokstat != NULL)
452 kstat_waitq_exit(spa->spa_iokstat->ks_data);
453 mutex_exit(&spa->spa_iokstat_lock);
458 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
460 spa_t *spa = zio->io_spa;
461 ASSERT(MUTEX_HELD(&vq->vq_lock));
462 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
463 vq->vq_class[zio->io_priority].vqc_active++;
464 avl_add(&vq->vq_active_tree, zio);
467 mutex_enter(&spa->spa_iokstat_lock);
468 spa->spa_queue_stats[zio->io_priority].spa_active++;
469 if (spa->spa_iokstat != NULL)
470 kstat_runq_enter(spa->spa_iokstat->ks_data);
471 mutex_exit(&spa->spa_iokstat_lock);
476 vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
478 spa_t *spa = zio->io_spa;
479 ASSERT(MUTEX_HELD(&vq->vq_lock));
480 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
481 vq->vq_class[zio->io_priority].vqc_active--;
482 avl_remove(&vq->vq_active_tree, zio);
485 mutex_enter(&spa->spa_iokstat_lock);
486 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
487 spa->spa_queue_stats[zio->io_priority].spa_active--;
488 if (spa->spa_iokstat != NULL) {
489 kstat_io_t *ksio = spa->spa_iokstat->ks_data;
491 kstat_runq_exit(spa->spa_iokstat->ks_data);
492 if (zio->io_type == ZIO_TYPE_READ) {
494 ksio->nread += zio->io_size;
495 } else if (zio->io_type == ZIO_TYPE_WRITE) {
497 ksio->nwritten += zio->io_size;
500 mutex_exit(&spa->spa_iokstat_lock);
505 vdev_queue_agg_io_done(zio_t *aio)
507 if (aio->io_type == ZIO_TYPE_READ) {
509 zio_link_t *zl = NULL;
510 while ((pio = zio_walk_parents(aio, &zl)) != NULL) {
511 abd_copy_off(pio->io_abd, aio->io_abd,
512 0, pio->io_offset - aio->io_offset, pio->io_size);
516 abd_free(aio->io_abd);
520 vdev_queue_class_min_active(zio_priority_t p)
523 case ZIO_PRIORITY_SYNC_READ:
524 return (zfs_vdev_sync_read_min_active);
525 case ZIO_PRIORITY_SYNC_WRITE:
526 return (zfs_vdev_sync_write_min_active);
527 case ZIO_PRIORITY_ASYNC_READ:
528 return (zfs_vdev_async_read_min_active);
529 case ZIO_PRIORITY_ASYNC_WRITE:
530 return (zfs_vdev_async_write_min_active);
531 case ZIO_PRIORITY_SCRUB:
532 return (zfs_vdev_scrub_min_active);
533 case ZIO_PRIORITY_TRIM:
534 return (zfs_vdev_trim_min_active);
535 case ZIO_PRIORITY_REMOVAL:
536 return (zfs_vdev_removal_min_active);
538 panic("invalid priority %u", p);
543 static __noinline int
544 vdev_queue_max_async_writes(spa_t *spa)
547 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total;
548 uint64_t min_bytes = zfs_dirty_data_max *
549 zfs_vdev_async_write_active_min_dirty_percent / 100;
550 uint64_t max_bytes = zfs_dirty_data_max *
551 zfs_vdev_async_write_active_max_dirty_percent / 100;
554 * Sync tasks correspond to interactive user actions. To reduce the
555 * execution time of those actions we push data out as fast as possible.
557 if (spa_has_pending_synctask(spa)) {
558 return (zfs_vdev_async_write_max_active);
561 if (dirty < min_bytes)
562 return (zfs_vdev_async_write_min_active);
563 if (dirty > max_bytes)
564 return (zfs_vdev_async_write_max_active);
567 * linear interpolation:
568 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
569 * move right by min_bytes
570 * move up by min_writes
572 writes = (dirty - min_bytes) *
573 (zfs_vdev_async_write_max_active -
574 zfs_vdev_async_write_min_active) /
575 (max_bytes - min_bytes) +
576 zfs_vdev_async_write_min_active;
577 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
578 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
583 vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
586 case ZIO_PRIORITY_SYNC_READ:
587 return (zfs_vdev_sync_read_max_active);
588 case ZIO_PRIORITY_SYNC_WRITE:
589 return (zfs_vdev_sync_write_max_active);
590 case ZIO_PRIORITY_ASYNC_READ:
591 return (zfs_vdev_async_read_max_active);
592 case ZIO_PRIORITY_ASYNC_WRITE:
593 return (vdev_queue_max_async_writes(spa));
594 case ZIO_PRIORITY_SCRUB:
595 return (zfs_vdev_scrub_max_active);
596 case ZIO_PRIORITY_TRIM:
597 return (zfs_vdev_trim_max_active);
598 case ZIO_PRIORITY_REMOVAL:
599 return (zfs_vdev_removal_max_active);
601 panic("invalid priority %u", p);
607 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
608 * there is no eligible class.
610 static zio_priority_t
611 vdev_queue_class_to_issue(vdev_queue_t *vq)
613 spa_t *spa = vq->vq_vdev->vdev_spa;
616 ASSERT(MUTEX_HELD(&vq->vq_lock));
618 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
619 return (ZIO_PRIORITY_NUM_QUEUEABLE);
621 /* find a queue that has not reached its minimum # outstanding i/os */
622 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
623 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
624 vq->vq_class[p].vqc_active <
625 vdev_queue_class_min_active(p))
630 * If we haven't found a queue, look for one that hasn't reached its
631 * maximum # outstanding i/os.
633 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
634 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
635 vq->vq_class[p].vqc_active <
636 vdev_queue_class_max_active(spa, p))
640 /* No eligible queued i/os */
641 return (ZIO_PRIORITY_NUM_QUEUEABLE);
645 * Compute the range spanned by two i/os, which is the endpoint of the last
646 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
647 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
648 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
650 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
651 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
654 vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
656 zio_t *first, *last, *aio, *dio, *mandatory, *nio;
663 ASSERT(MUTEX_HELD(&vq->vq_lock));
665 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
670 if (zio->io_type == ZIO_TYPE_READ)
671 maxgap = zfs_vdev_read_gap_limit;
674 * We can aggregate I/Os that are sufficiently adjacent and of
675 * the same flavor, as expressed by the AGG_INHERIT flags.
676 * The latter requirement is necessary so that certain
677 * attributes of the I/O, such as whether it's a normal I/O
678 * or a scrub/resilver, can be preserved in the aggregate.
679 * We can include optional I/Os, but don't allow them
680 * to begin a range as they add no benefit in that situation.
684 * We keep track of the last non-optional I/O.
686 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
689 * Walk backwards through sufficiently contiguous I/Os
690 * recording the last non-optional I/O.
692 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
693 t = vdev_queue_type_tree(vq, zio->io_type);
694 while (t != NULL && (dio = AVL_PREV(t, first)) != NULL &&
695 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
696 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
697 IO_GAP(dio, first) <= maxgap &&
698 dio->io_type == zio->io_type) {
700 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
705 * Skip any initial optional I/Os.
707 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
708 first = AVL_NEXT(t, first);
709 ASSERT(first != NULL);
713 * Walk forward through sufficiently contiguous I/Os.
714 * The aggregation limit does not apply to optional i/os, so that
715 * we can issue contiguous writes even if they are larger than the
718 while ((dio = AVL_NEXT(t, last)) != NULL &&
719 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
720 (IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit ||
721 (dio->io_flags & ZIO_FLAG_OPTIONAL)) &&
722 IO_GAP(last, dio) <= maxgap &&
723 dio->io_type == zio->io_type) {
725 if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
730 * Now that we've established the range of the I/O aggregation
731 * we must decide what to do with trailing optional I/Os.
732 * For reads, there's nothing to do. While we are unable to
733 * aggregate further, it's possible that a trailing optional
734 * I/O would allow the underlying device to aggregate with
735 * subsequent I/Os. We must therefore determine if the next
736 * non-optional I/O is close enough to make aggregation
740 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
742 while ((dio = AVL_NEXT(t, nio)) != NULL &&
743 IO_GAP(nio, dio) == 0 &&
744 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
746 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
755 * We are going to include an optional io in our aggregated
756 * span, thus closing the write gap. Only mandatory i/os can
757 * start aggregated spans, so make sure that the next i/o
758 * after our span is mandatory.
760 dio = AVL_NEXT(t, last);
761 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
763 /* do not include the optional i/o */
764 while (last != mandatory && last != first) {
765 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
766 last = AVL_PREV(t, last);
767 ASSERT(last != NULL);
774 size = IO_SPAN(first, last);
775 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
777 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
778 abd_alloc_for_io(size, B_TRUE), size, first->io_type,
779 zio->io_priority, flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
780 vdev_queue_agg_io_done, NULL);
781 aio->io_timestamp = first->io_timestamp;
786 nio = AVL_NEXT(t, dio);
787 ASSERT3U(dio->io_type, ==, aio->io_type);
789 if (dio->io_flags & ZIO_FLAG_NODATA) {
790 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
791 abd_zero_off(aio->io_abd,
792 dio->io_offset - aio->io_offset, dio->io_size);
793 } else if (dio->io_type == ZIO_TYPE_WRITE) {
794 abd_copy_off(aio->io_abd, dio->io_abd,
795 dio->io_offset - aio->io_offset, 0, dio->io_size);
798 zio_add_child(dio, aio);
799 vdev_queue_io_remove(vq, dio);
800 zio_vdev_io_bypass(dio);
802 } while (dio != last);
808 vdev_queue_io_to_issue(vdev_queue_t *vq)
817 ASSERT(MUTEX_HELD(&vq->vq_lock));
819 p = vdev_queue_class_to_issue(vq);
821 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
822 /* No eligible queued i/os */
827 * For LBA-ordered queues (async / scrub), issue the i/o which follows
828 * the most recently issued i/o in LBA (offset) order.
830 * For FIFO queues (sync), issue the i/o with the lowest timestamp.
832 tree = vdev_queue_class_tree(vq, p);
833 search.io_timestamp = 0;
834 search.io_offset = vq->vq_last_offset + 1;
835 VERIFY3P(avl_find(tree, &search, &idx), ==, NULL);
836 zio = avl_nearest(tree, idx, AVL_AFTER);
838 zio = avl_first(tree);
839 ASSERT3U(zio->io_priority, ==, p);
841 aio = vdev_queue_aggregate(vq, zio);
845 vdev_queue_io_remove(vq, zio);
848 * If the I/O is or was optional and therefore has no data, we need to
849 * simply discard it. We need to drop the vdev queue's lock to avoid a
850 * deadlock that we could encounter since this I/O will complete
853 if (zio->io_flags & ZIO_FLAG_NODATA) {
854 mutex_exit(&vq->vq_lock);
855 zio_vdev_io_bypass(zio);
857 mutex_enter(&vq->vq_lock);
861 vdev_queue_pending_add(vq, zio);
862 vq->vq_last_offset = zio->io_offset;
868 vdev_queue_io(zio_t *zio)
870 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
873 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
877 * Children i/os inherent their parent's priority, which might
878 * not match the child's i/o type. Fix it up here.
880 if (zio->io_type == ZIO_TYPE_READ) {
881 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
882 zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
883 zio->io_priority != ZIO_PRIORITY_SCRUB &&
884 zio->io_priority != ZIO_PRIORITY_REMOVAL)
885 zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
886 } else if (zio->io_type == ZIO_TYPE_WRITE) {
887 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
888 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE &&
889 zio->io_priority != ZIO_PRIORITY_REMOVAL)
890 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
892 ASSERT(zio->io_type == ZIO_TYPE_FREE);
893 zio->io_priority = ZIO_PRIORITY_TRIM;
896 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
898 mutex_enter(&vq->vq_lock);
899 zio->io_timestamp = gethrtime();
900 vdev_queue_io_add(vq, zio);
901 nio = vdev_queue_io_to_issue(vq);
902 mutex_exit(&vq->vq_lock);
907 if (nio->io_done == vdev_queue_agg_io_done) {
916 vdev_queue_io_done(zio_t *zio)
918 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
921 mutex_enter(&vq->vq_lock);
923 vdev_queue_pending_remove(vq, zio);
925 vq->vq_io_complete_ts = gethrtime();
927 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
928 mutex_exit(&vq->vq_lock);
929 if (nio->io_done == vdev_queue_agg_io_done) {
932 zio_vdev_io_reissue(nio);
935 mutex_enter(&vq->vq_lock);
938 mutex_exit(&vq->vq_lock);
942 vdev_queue_change_io_priority(zio_t *zio, zio_priority_t priority)
944 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
947 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
948 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
950 if (zio->io_type == ZIO_TYPE_READ) {
951 if (priority != ZIO_PRIORITY_SYNC_READ &&
952 priority != ZIO_PRIORITY_ASYNC_READ &&
953 priority != ZIO_PRIORITY_SCRUB)
954 priority = ZIO_PRIORITY_ASYNC_READ;
956 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
957 if (priority != ZIO_PRIORITY_SYNC_WRITE &&
958 priority != ZIO_PRIORITY_ASYNC_WRITE)
959 priority = ZIO_PRIORITY_ASYNC_WRITE;
962 mutex_enter(&vq->vq_lock);
965 * If the zio is in none of the queues we can simply change
966 * the priority. If the zio is waiting to be submitted we must
967 * remove it from the queue and re-insert it with the new priority.
968 * Otherwise, the zio is currently active and we cannot change its
971 tree = vdev_queue_class_tree(vq, zio->io_priority);
972 if (avl_find(tree, zio, NULL) == zio) {
973 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
974 zio->io_priority = priority;
975 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
976 } else if (avl_find(&vq->vq_active_tree, zio, NULL) != zio) {
977 zio->io_priority = priority;
980 mutex_exit(&vq->vq_lock);
984 * As these three methods are only used for load calculations we're not concerned
985 * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex
986 * use here, instead we prefer to keep it lock free for performance.
989 vdev_queue_length(vdev_t *vd)
991 return (avl_numnodes(&vd->vdev_queue.vq_active_tree));
995 vdev_queue_lastoffset(vdev_t *vd)
997 return (vd->vdev_queue.vq_lastoffset);
1001 vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio)
1003 vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size;