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 TUNABLE_INT("vfs.zfs.vdev.async_write_active_min_dirty_percent",
201 &zfs_vdev_async_write_active_min_dirty_percent);
202 static int sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS);
203 SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_min_dirty_percent,
204 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
205 sysctl_zfs_async_write_active_min_dirty_percent, "I",
206 "Percentage of async write dirty data below which "
207 "async_write_min_active is used.");
209 TUNABLE_INT("vfs.zfs.vdev.async_write_active_max_dirty_percent",
210 &zfs_vdev_async_write_active_max_dirty_percent);
211 static int sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS);
212 SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_max_dirty_percent,
213 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
214 sysctl_zfs_async_write_active_max_dirty_percent, "I",
215 "Percentage of async write dirty data above which "
216 "async_write_max_active is used.");
218 TUNABLE_INT("vfs.zfs.vdev.max_active", &zfs_vdev_max_active);
219 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RWTUN,
220 &zfs_vdev_max_active, 0,
221 "The maximum number of I/Os of all types active for each device.");
223 #define ZFS_VDEV_QUEUE_KNOB_MIN(name) \
224 TUNABLE_INT("vfs.zfs.vdev." #name "_min_active", \
225 &zfs_vdev_ ## name ## _min_active); \
226 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, \
227 CTLFLAG_RWTUN, &zfs_vdev_ ## name ## _min_active, 0, \
228 "Initial number of I/O requests of type " #name \
229 " active for each device");
231 #define ZFS_VDEV_QUEUE_KNOB_MAX(name) \
232 TUNABLE_INT("vfs.zfs.vdev." #name "_max_active", \
233 &zfs_vdev_ ## name ## _max_active); \
234 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, \
235 CTLFLAG_RWTUN, &zfs_vdev_ ## name ## _max_active, 0, \
236 "Maximum number of I/O requests of type " #name \
237 " active for each device");
239 ZFS_VDEV_QUEUE_KNOB_MIN(sync_read);
240 ZFS_VDEV_QUEUE_KNOB_MAX(sync_read);
241 ZFS_VDEV_QUEUE_KNOB_MIN(sync_write);
242 ZFS_VDEV_QUEUE_KNOB_MAX(sync_write);
243 ZFS_VDEV_QUEUE_KNOB_MIN(async_read);
244 ZFS_VDEV_QUEUE_KNOB_MAX(async_read);
245 ZFS_VDEV_QUEUE_KNOB_MIN(async_write);
246 ZFS_VDEV_QUEUE_KNOB_MAX(async_write);
247 ZFS_VDEV_QUEUE_KNOB_MIN(scrub);
248 ZFS_VDEV_QUEUE_KNOB_MAX(scrub);
249 ZFS_VDEV_QUEUE_KNOB_MIN(trim);
250 ZFS_VDEV_QUEUE_KNOB_MAX(trim);
252 #undef ZFS_VDEV_QUEUE_KNOB
254 TUNABLE_INT("vfs.zfs.vdev.aggregation_limit", &zfs_vdev_aggregation_limit);
255 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RWTUN,
256 &zfs_vdev_aggregation_limit, 0,
257 "I/O requests are aggregated up to this size");
258 TUNABLE_INT("vfs.zfs.vdev.read_gap_limit", &zfs_vdev_read_gap_limit);
259 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RWTUN,
260 &zfs_vdev_read_gap_limit, 0,
261 "Acceptable gap between two reads being aggregated");
262 TUNABLE_INT("vfs.zfs.vdev.write_gap_limit", &zfs_vdev_write_gap_limit);
263 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RWTUN,
264 &zfs_vdev_write_gap_limit, 0,
265 "Acceptable gap between two writes being aggregated");
266 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, queue_depth_pct, CTLFLAG_RWTUN,
267 &zfs_vdev_queue_depth_pct, 0,
268 "Queue depth percentage for each top-level");
271 sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS)
275 val = zfs_vdev_async_write_active_min_dirty_percent;
276 err = sysctl_handle_int(oidp, &val, 0, req);
277 if (err != 0 || req->newptr == NULL)
280 if (val < 0 || val > 100 ||
281 val >= zfs_vdev_async_write_active_max_dirty_percent)
284 zfs_vdev_async_write_active_min_dirty_percent = val;
290 sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS)
294 val = zfs_vdev_async_write_active_max_dirty_percent;
295 err = sysctl_handle_int(oidp, &val, 0, req);
296 if (err != 0 || req->newptr == NULL)
299 if (val < 0 || val > 100 ||
300 val <= zfs_vdev_async_write_active_min_dirty_percent)
303 zfs_vdev_async_write_active_max_dirty_percent = val;
311 vdev_queue_offset_compare(const void *x1, const void *x2)
313 const zio_t *z1 = x1;
314 const zio_t *z2 = x2;
316 if (z1->io_offset < z2->io_offset)
318 if (z1->io_offset > z2->io_offset)
329 static inline avl_tree_t *
330 vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p)
332 return (&vq->vq_class[p].vqc_queued_tree);
335 static inline avl_tree_t *
336 vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t)
338 if (t == ZIO_TYPE_READ)
339 return (&vq->vq_read_offset_tree);
340 else if (t == ZIO_TYPE_WRITE)
341 return (&vq->vq_write_offset_tree);
347 vdev_queue_timestamp_compare(const void *x1, const void *x2)
349 const zio_t *z1 = x1;
350 const zio_t *z2 = x2;
352 if (z1->io_timestamp < z2->io_timestamp)
354 if (z1->io_timestamp > z2->io_timestamp)
357 if (z1->io_offset < z2->io_offset)
359 if (z1->io_offset > z2->io_offset)
371 vdev_queue_init(vdev_t *vd)
373 vdev_queue_t *vq = &vd->vdev_queue;
375 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
378 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
379 sizeof (zio_t), offsetof(struct zio, io_queue_node));
380 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ),
381 vdev_queue_offset_compare, sizeof (zio_t),
382 offsetof(struct zio, io_offset_node));
383 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE),
384 vdev_queue_offset_compare, sizeof (zio_t),
385 offsetof(struct zio, io_offset_node));
387 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
388 int (*compfn) (const void *, const void *);
391 * The synchronous i/o queues are dispatched in FIFO rather
392 * than LBA order. This provides more consistent latency for
395 if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE)
396 compfn = vdev_queue_timestamp_compare;
398 compfn = vdev_queue_offset_compare;
400 avl_create(vdev_queue_class_tree(vq, p), compfn,
401 sizeof (zio_t), offsetof(struct zio, io_queue_node));
404 vq->vq_lastoffset = 0;
408 vdev_queue_fini(vdev_t *vd)
410 vdev_queue_t *vq = &vd->vdev_queue;
412 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
413 avl_destroy(vdev_queue_class_tree(vq, p));
414 avl_destroy(&vq->vq_active_tree);
415 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
416 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE));
418 mutex_destroy(&vq->vq_lock);
422 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
424 spa_t *spa = zio->io_spa;
427 ASSERT(MUTEX_HELD(&vq->vq_lock));
428 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
429 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
430 qtt = vdev_queue_type_tree(vq, zio->io_type);
435 mutex_enter(&spa->spa_iokstat_lock);
436 spa->spa_queue_stats[zio->io_priority].spa_queued++;
437 if (spa->spa_iokstat != NULL)
438 kstat_waitq_enter(spa->spa_iokstat->ks_data);
439 mutex_exit(&spa->spa_iokstat_lock);
444 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
446 spa_t *spa = zio->io_spa;
449 ASSERT(MUTEX_HELD(&vq->vq_lock));
450 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
451 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
452 qtt = vdev_queue_type_tree(vq, zio->io_type);
454 avl_remove(qtt, zio);
457 mutex_enter(&spa->spa_iokstat_lock);
458 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
459 spa->spa_queue_stats[zio->io_priority].spa_queued--;
460 if (spa->spa_iokstat != NULL)
461 kstat_waitq_exit(spa->spa_iokstat->ks_data);
462 mutex_exit(&spa->spa_iokstat_lock);
467 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
469 spa_t *spa = zio->io_spa;
470 ASSERT(MUTEX_HELD(&vq->vq_lock));
471 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
472 vq->vq_class[zio->io_priority].vqc_active++;
473 avl_add(&vq->vq_active_tree, zio);
476 mutex_enter(&spa->spa_iokstat_lock);
477 spa->spa_queue_stats[zio->io_priority].spa_active++;
478 if (spa->spa_iokstat != NULL)
479 kstat_runq_enter(spa->spa_iokstat->ks_data);
480 mutex_exit(&spa->spa_iokstat_lock);
485 vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
487 spa_t *spa = zio->io_spa;
488 ASSERT(MUTEX_HELD(&vq->vq_lock));
489 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
490 vq->vq_class[zio->io_priority].vqc_active--;
491 avl_remove(&vq->vq_active_tree, zio);
494 mutex_enter(&spa->spa_iokstat_lock);
495 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
496 spa->spa_queue_stats[zio->io_priority].spa_active--;
497 if (spa->spa_iokstat != NULL) {
498 kstat_io_t *ksio = spa->spa_iokstat->ks_data;
500 kstat_runq_exit(spa->spa_iokstat->ks_data);
501 if (zio->io_type == ZIO_TYPE_READ) {
503 ksio->nread += zio->io_size;
504 } else if (zio->io_type == ZIO_TYPE_WRITE) {
506 ksio->nwritten += zio->io_size;
509 mutex_exit(&spa->spa_iokstat_lock);
514 vdev_queue_agg_io_done(zio_t *aio)
516 if (aio->io_type == ZIO_TYPE_READ) {
518 zio_link_t *zl = NULL;
519 while ((pio = zio_walk_parents(aio, &zl)) != NULL) {
520 bcopy((char *)aio->io_data + (pio->io_offset -
521 aio->io_offset), pio->io_data, pio->io_size);
525 zio_buf_free(aio->io_data, aio->io_size);
529 vdev_queue_class_min_active(zio_priority_t p)
532 case ZIO_PRIORITY_SYNC_READ:
533 return (zfs_vdev_sync_read_min_active);
534 case ZIO_PRIORITY_SYNC_WRITE:
535 return (zfs_vdev_sync_write_min_active);
536 case ZIO_PRIORITY_ASYNC_READ:
537 return (zfs_vdev_async_read_min_active);
538 case ZIO_PRIORITY_ASYNC_WRITE:
539 return (zfs_vdev_async_write_min_active);
540 case ZIO_PRIORITY_SCRUB:
541 return (zfs_vdev_scrub_min_active);
542 case ZIO_PRIORITY_TRIM:
543 return (zfs_vdev_trim_min_active);
545 panic("invalid priority %u", p);
550 static __noinline int
551 vdev_queue_max_async_writes(spa_t *spa)
554 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total;
555 uint64_t min_bytes = zfs_dirty_data_max *
556 zfs_vdev_async_write_active_min_dirty_percent / 100;
557 uint64_t max_bytes = zfs_dirty_data_max *
558 zfs_vdev_async_write_active_max_dirty_percent / 100;
561 * Sync tasks correspond to interactive user actions. To reduce the
562 * execution time of those actions we push data out as fast as possible.
564 if (spa_has_pending_synctask(spa)) {
565 return (zfs_vdev_async_write_max_active);
568 if (dirty < min_bytes)
569 return (zfs_vdev_async_write_min_active);
570 if (dirty > max_bytes)
571 return (zfs_vdev_async_write_max_active);
574 * linear interpolation:
575 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
576 * move right by min_bytes
577 * move up by min_writes
579 writes = (dirty - min_bytes) *
580 (zfs_vdev_async_write_max_active -
581 zfs_vdev_async_write_min_active) /
582 (max_bytes - min_bytes) +
583 zfs_vdev_async_write_min_active;
584 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
585 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
590 vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
593 case ZIO_PRIORITY_SYNC_READ:
594 return (zfs_vdev_sync_read_max_active);
595 case ZIO_PRIORITY_SYNC_WRITE:
596 return (zfs_vdev_sync_write_max_active);
597 case ZIO_PRIORITY_ASYNC_READ:
598 return (zfs_vdev_async_read_max_active);
599 case ZIO_PRIORITY_ASYNC_WRITE:
600 return (vdev_queue_max_async_writes(spa));
601 case ZIO_PRIORITY_SCRUB:
602 return (zfs_vdev_scrub_max_active);
603 case ZIO_PRIORITY_TRIM:
604 return (zfs_vdev_trim_max_active);
606 panic("invalid priority %u", p);
612 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
613 * there is no eligible class.
615 static zio_priority_t
616 vdev_queue_class_to_issue(vdev_queue_t *vq)
618 spa_t *spa = vq->vq_vdev->vdev_spa;
621 ASSERT(MUTEX_HELD(&vq->vq_lock));
623 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
624 return (ZIO_PRIORITY_NUM_QUEUEABLE);
626 /* find a queue that has not reached its minimum # outstanding i/os */
627 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
628 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
629 vq->vq_class[p].vqc_active <
630 vdev_queue_class_min_active(p))
635 * If we haven't found a queue, look for one that hasn't reached its
636 * maximum # outstanding i/os.
638 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
639 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
640 vq->vq_class[p].vqc_active <
641 vdev_queue_class_max_active(spa, p))
645 /* No eligible queued i/os */
646 return (ZIO_PRIORITY_NUM_QUEUEABLE);
650 * Compute the range spanned by two i/os, which is the endpoint of the last
651 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
652 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
653 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
655 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
656 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
659 vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
661 zio_t *first, *last, *aio, *dio, *mandatory, *nio;
668 ASSERT(MUTEX_HELD(&vq->vq_lock));
670 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
675 if (zio->io_type == ZIO_TYPE_READ)
676 maxgap = zfs_vdev_read_gap_limit;
679 * We can aggregate I/Os that are sufficiently adjacent and of
680 * the same flavor, as expressed by the AGG_INHERIT flags.
681 * The latter requirement is necessary so that certain
682 * attributes of the I/O, such as whether it's a normal I/O
683 * or a scrub/resilver, can be preserved in the aggregate.
684 * We can include optional I/Os, but don't allow them
685 * to begin a range as they add no benefit in that situation.
689 * We keep track of the last non-optional I/O.
691 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
694 * Walk backwards through sufficiently contiguous I/Os
695 * recording the last non-option I/O.
697 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
698 t = vdev_queue_type_tree(vq, zio->io_type);
699 while (t != NULL && (dio = AVL_PREV(t, first)) != NULL &&
700 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
701 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
702 IO_GAP(dio, first) <= maxgap) {
704 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
709 * Skip any initial optional I/Os.
711 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
712 first = AVL_NEXT(t, first);
713 ASSERT(first != NULL);
717 * Walk forward through sufficiently contiguous I/Os.
719 while ((dio = AVL_NEXT(t, last)) != NULL &&
720 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
721 IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit &&
722 IO_GAP(last, dio) <= maxgap) {
724 if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
729 * Now that we've established the range of the I/O aggregation
730 * we must decide what to do with trailing optional I/Os.
731 * For reads, there's nothing to do. While we are unable to
732 * aggregate further, it's possible that a trailing optional
733 * I/O would allow the underlying device to aggregate with
734 * subsequent I/Os. We must therefore determine if the next
735 * non-optional I/O is close enough to make aggregation
739 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
741 while ((dio = AVL_NEXT(t, nio)) != NULL &&
742 IO_GAP(nio, dio) == 0 &&
743 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
745 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
753 /* This may be a no-op. */
754 dio = AVL_NEXT(t, last);
755 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
757 while (last != mandatory && last != first) {
758 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
759 last = AVL_PREV(t, last);
760 ASSERT(last != NULL);
767 size = IO_SPAN(first, last);
768 ASSERT3U(size, <=, zfs_vdev_aggregation_limit);
770 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
771 zio_buf_alloc(size), size, first->io_type, zio->io_priority,
772 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
773 vdev_queue_agg_io_done, NULL);
774 aio->io_timestamp = first->io_timestamp;
779 nio = AVL_NEXT(t, dio);
780 ASSERT3U(dio->io_type, ==, aio->io_type);
782 if (dio->io_flags & ZIO_FLAG_NODATA) {
783 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
784 bzero((char *)aio->io_data + (dio->io_offset -
785 aio->io_offset), dio->io_size);
786 } else if (dio->io_type == ZIO_TYPE_WRITE) {
787 bcopy(dio->io_data, (char *)aio->io_data +
788 (dio->io_offset - aio->io_offset),
792 zio_add_child(dio, aio);
793 vdev_queue_io_remove(vq, dio);
794 zio_vdev_io_bypass(dio);
796 } while (dio != last);
802 vdev_queue_io_to_issue(vdev_queue_t *vq)
811 ASSERT(MUTEX_HELD(&vq->vq_lock));
813 p = vdev_queue_class_to_issue(vq);
815 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
816 /* No eligible queued i/os */
821 * For LBA-ordered queues (async / scrub), issue the i/o which follows
822 * the most recently issued i/o in LBA (offset) order.
824 * For FIFO queues (sync), issue the i/o with the lowest timestamp.
826 tree = vdev_queue_class_tree(vq, p);
827 search.io_timestamp = 0;
828 search.io_offset = vq->vq_last_offset + 1;
829 VERIFY3P(avl_find(tree, &search, &idx), ==, NULL);
830 zio = avl_nearest(tree, idx, AVL_AFTER);
832 zio = avl_first(tree);
833 ASSERT3U(zio->io_priority, ==, p);
835 aio = vdev_queue_aggregate(vq, zio);
839 vdev_queue_io_remove(vq, zio);
842 * If the I/O is or was optional and therefore has no data, we need to
843 * simply discard it. We need to drop the vdev queue's lock to avoid a
844 * deadlock that we could encounter since this I/O will complete
847 if (zio->io_flags & ZIO_FLAG_NODATA) {
848 mutex_exit(&vq->vq_lock);
849 zio_vdev_io_bypass(zio);
851 mutex_enter(&vq->vq_lock);
855 vdev_queue_pending_add(vq, zio);
856 vq->vq_last_offset = zio->io_offset;
862 vdev_queue_io(zio_t *zio)
864 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
867 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
871 * Children i/os inherent their parent's priority, which might
872 * not match the child's i/o type. Fix it up here.
874 if (zio->io_type == ZIO_TYPE_READ) {
875 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
876 zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
877 zio->io_priority != ZIO_PRIORITY_SCRUB)
878 zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
879 } else if (zio->io_type == ZIO_TYPE_WRITE) {
880 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
881 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE)
882 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
884 ASSERT(zio->io_type == ZIO_TYPE_FREE);
885 zio->io_priority = ZIO_PRIORITY_TRIM;
888 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
890 mutex_enter(&vq->vq_lock);
891 zio->io_timestamp = gethrtime();
892 vdev_queue_io_add(vq, zio);
893 nio = vdev_queue_io_to_issue(vq);
894 mutex_exit(&vq->vq_lock);
899 if (nio->io_done == vdev_queue_agg_io_done) {
908 vdev_queue_io_done(zio_t *zio)
910 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
913 mutex_enter(&vq->vq_lock);
915 vdev_queue_pending_remove(vq, zio);
917 vq->vq_io_complete_ts = gethrtime();
919 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
920 mutex_exit(&vq->vq_lock);
921 if (nio->io_done == vdev_queue_agg_io_done) {
924 zio_vdev_io_reissue(nio);
927 mutex_enter(&vq->vq_lock);
930 mutex_exit(&vq->vq_lock);
934 * As these three methods are only used for load calculations we're not concerned
935 * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex
936 * use here, instead we prefer to keep it lock free for performance.
939 vdev_queue_length(vdev_t *vd)
941 return (avl_numnodes(&vd->vdev_queue.vq_active_tree));
945 vdev_queue_lastoffset(vdev_t *vd)
947 return (vd->vdev_queue.vq_lastoffset);
951 vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio)
953 vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size;