4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
27 * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
30 #include <sys/zfs_context.h>
31 #include <sys/vdev_impl.h>
32 #include <sys/spa_impl.h>
35 #include <sys/dsl_pool.h>
36 #include <sys/metaslab_impl.h>
38 #include <sys/spa_impl.h>
39 #include <sys/kstat.h>
46 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The
47 * I/O scheduler determines when and in what order those operations are
48 * issued. The I/O scheduler divides operations into five I/O classes
49 * prioritized in the following order: sync read, sync write, async read,
50 * async write, and scrub/resilver. Each queue defines the minimum and
51 * maximum number of concurrent operations that may be issued to the device.
52 * In addition, the device has an aggregate maximum. Note that the sum of the
53 * per-queue minimums must not exceed the aggregate maximum. If the
54 * sum of the per-queue maximums exceeds the aggregate maximum, then the
55 * number of active i/os may reach zfs_vdev_max_active, in which case no
56 * further i/os will be issued regardless of whether all per-queue
57 * minimums have been met.
59 * For many physical devices, throughput increases with the number of
60 * concurrent operations, but latency typically suffers. Further, physical
61 * devices typically have a limit at which more concurrent operations have no
62 * effect on throughput or can actually cause it to decrease.
64 * The scheduler selects the next operation to issue by first looking for an
65 * I/O class whose minimum has not been satisfied. Once all are satisfied and
66 * the aggregate maximum has not been hit, the scheduler looks for classes
67 * whose maximum has not been satisfied. Iteration through the I/O classes is
68 * done in the order specified above. No further operations are issued if the
69 * aggregate maximum number of concurrent operations has been hit or if there
70 * are no operations queued for an I/O class that has not hit its maximum.
71 * Every time an i/o is queued or an operation completes, the I/O scheduler
72 * looks for new operations to issue.
74 * All I/O classes have a fixed maximum number of outstanding operations
75 * except for the async write class. Asynchronous writes represent the data
76 * that is committed to stable storage during the syncing stage for
77 * transaction groups (see txg.c). Transaction groups enter the syncing state
78 * periodically so the number of queued async writes will quickly burst up and
79 * then bleed down to zero. Rather than servicing them as quickly as possible,
80 * the I/O scheduler changes the maximum number of active async write i/os
81 * according to the amount of dirty data in the pool (see dsl_pool.c). Since
82 * both throughput and latency typically increase with the number of
83 * concurrent operations issued to physical devices, reducing the burstiness
84 * in the number of concurrent operations also stabilizes the response time of
85 * operations from other -- and in particular synchronous -- queues. In broad
86 * strokes, the I/O scheduler will issue more concurrent operations from the
87 * async write queue as there's more dirty data in the pool.
91 * The number of concurrent operations issued for the async write I/O class
92 * follows a piece-wise linear function defined by a few adjustable points.
94 * | o---------| <-- zfs_vdev_async_write_max_active
101 * |------------o | | <-- zfs_vdev_async_write_min_active
102 * 0|____________^______|_________|
103 * 0% | | 100% of zfs_dirty_data_max
105 * | `-- zfs_vdev_async_write_active_max_dirty_percent
106 * `--------- zfs_vdev_async_write_active_min_dirty_percent
108 * Until the amount of dirty data exceeds a minimum percentage of the dirty
109 * data allowed in the pool, the I/O scheduler will limit the number of
110 * concurrent operations to the minimum. As that threshold is crossed, the
111 * number of concurrent operations issued increases linearly to the maximum at
112 * the specified maximum percentage of the dirty data allowed in the pool.
114 * Ideally, the amount of dirty data on a busy pool will stay in the sloped
115 * part of the function between zfs_vdev_async_write_active_min_dirty_percent
116 * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the
117 * maximum percentage, this indicates that the rate of incoming data is
118 * greater than the rate that the backend storage can handle. In this case, we
119 * must further throttle incoming writes (see dmu_tx_delay() for details).
123 * The maximum number of i/os active to each device. Ideally, this will be >=
124 * the sum of each queue's max_active. It must be at least the sum of each
125 * queue's min_active.
127 uint32_t zfs_vdev_max_active
= 1000;
130 * Per-queue limits on the number of i/os active to each device. If the
131 * number of active i/os is < zfs_vdev_max_active, then the min_active comes
132 * into play. We will send min_active from each queue, and then select from
133 * queues in the order defined by zio_priority_t.
135 * In general, smaller max_active's will lead to lower latency of synchronous
136 * operations. Larger max_active's may lead to higher overall throughput,
137 * depending on underlying storage.
139 * The ratio of the queues' max_actives determines the balance of performance
140 * between reads, writes, and scrubs. E.g., increasing
141 * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete
142 * more quickly, but reads and writes to have higher latency and lower
145 uint32_t zfs_vdev_sync_read_min_active
= 10;
146 uint32_t zfs_vdev_sync_read_max_active
= 10;
147 uint32_t zfs_vdev_sync_write_min_active
= 10;
148 uint32_t zfs_vdev_sync_write_max_active
= 10;
149 uint32_t zfs_vdev_async_read_min_active
= 1;
150 uint32_t zfs_vdev_async_read_max_active
= 3;
151 uint32_t zfs_vdev_async_write_min_active
= 2;
152 uint32_t zfs_vdev_async_write_max_active
= 10;
153 uint32_t zfs_vdev_scrub_min_active
= 1;
154 uint32_t zfs_vdev_scrub_max_active
= 2;
155 uint32_t zfs_vdev_removal_min_active
= 1;
156 uint32_t zfs_vdev_removal_max_active
= 2;
157 uint32_t zfs_vdev_initializing_min_active
= 1;
158 uint32_t zfs_vdev_initializing_max_active
= 1;
159 uint32_t zfs_vdev_trim_min_active
= 1;
160 uint32_t zfs_vdev_trim_max_active
= 2;
161 uint32_t zfs_vdev_rebuild_min_active
= 1;
162 uint32_t zfs_vdev_rebuild_max_active
= 3;
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_aggregation_limit_non_rotating
= SPA_OLD_MAXBLOCKSIZE
;
182 int zfs_vdev_read_gap_limit
= 32 << 10;
183 int zfs_vdev_write_gap_limit
= 4 << 10;
186 * Define the queue depth percentage for each top-level. This percentage is
187 * used in conjunction with zfs_vdev_async_max_active to determine how many
188 * allocations a specific top-level vdev should handle. Once the queue depth
189 * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100
190 * then allocator will stop allocating blocks on that top-level device.
191 * The default kernel setting is 1000% which will yield 100 allocations per
192 * device. For userland testing, the default setting is 300% which equates
193 * to 30 allocations per device.
196 int zfs_vdev_queue_depth_pct
= 1000;
198 int zfs_vdev_queue_depth_pct
= 300;
202 * When performing allocations for a given metaslab, we want to make sure that
203 * there are enough IOs to aggregate together to improve throughput. We want to
204 * ensure that there are at least 128k worth of IOs that can be aggregated, and
205 * we assume that the average allocation size is 4k, so we need the queue depth
206 * to be 32 per allocator to get good aggregation of sequential writes.
208 int zfs_vdev_def_queue_depth
= 32;
211 * Allow TRIM I/Os to be aggregated. This should normally not be needed since
212 * TRIM I/O for extents up to zfs_trim_extent_bytes_max (128M) can be submitted
213 * by the TRIM code in zfs_trim.c.
215 int zfs_vdev_aggregate_trim
= 0;
218 vdev_queue_offset_compare(const void *x1
, const void *x2
)
220 const zio_t
*z1
= (const zio_t
*)x1
;
221 const zio_t
*z2
= (const zio_t
*)x2
;
223 int cmp
= TREE_CMP(z1
->io_offset
, z2
->io_offset
);
228 return (TREE_PCMP(z1
, z2
));
231 static inline avl_tree_t
*
232 vdev_queue_class_tree(vdev_queue_t
*vq
, zio_priority_t p
)
234 return (&vq
->vq_class
[p
].vqc_queued_tree
);
237 static inline avl_tree_t
*
238 vdev_queue_type_tree(vdev_queue_t
*vq
, zio_type_t t
)
240 ASSERT(t
== ZIO_TYPE_READ
|| t
== ZIO_TYPE_WRITE
|| t
== ZIO_TYPE_TRIM
);
241 if (t
== ZIO_TYPE_READ
)
242 return (&vq
->vq_read_offset_tree
);
243 else if (t
== ZIO_TYPE_WRITE
)
244 return (&vq
->vq_write_offset_tree
);
246 return (&vq
->vq_trim_offset_tree
);
250 vdev_queue_timestamp_compare(const void *x1
, const void *x2
)
252 const zio_t
*z1
= (const zio_t
*)x1
;
253 const zio_t
*z2
= (const zio_t
*)x2
;
255 int cmp
= TREE_CMP(z1
->io_timestamp
, z2
->io_timestamp
);
260 return (TREE_PCMP(z1
, z2
));
264 vdev_queue_class_min_active(zio_priority_t p
)
267 case ZIO_PRIORITY_SYNC_READ
:
268 return (zfs_vdev_sync_read_min_active
);
269 case ZIO_PRIORITY_SYNC_WRITE
:
270 return (zfs_vdev_sync_write_min_active
);
271 case ZIO_PRIORITY_ASYNC_READ
:
272 return (zfs_vdev_async_read_min_active
);
273 case ZIO_PRIORITY_ASYNC_WRITE
:
274 return (zfs_vdev_async_write_min_active
);
275 case ZIO_PRIORITY_SCRUB
:
276 return (zfs_vdev_scrub_min_active
);
277 case ZIO_PRIORITY_REMOVAL
:
278 return (zfs_vdev_removal_min_active
);
279 case ZIO_PRIORITY_INITIALIZING
:
280 return (zfs_vdev_initializing_min_active
);
281 case ZIO_PRIORITY_TRIM
:
282 return (zfs_vdev_trim_min_active
);
283 case ZIO_PRIORITY_REBUILD
:
284 return (zfs_vdev_rebuild_min_active
);
286 panic("invalid priority %u", p
);
292 vdev_queue_max_async_writes(spa_t
*spa
)
296 dsl_pool_t
*dp
= spa_get_dsl(spa
);
297 uint64_t min_bytes
= zfs_dirty_data_max
*
298 zfs_vdev_async_write_active_min_dirty_percent
/ 100;
299 uint64_t max_bytes
= zfs_dirty_data_max
*
300 zfs_vdev_async_write_active_max_dirty_percent
/ 100;
303 * Async writes may occur before the assignment of the spa's
304 * dsl_pool_t if a self-healing zio is issued prior to the
305 * completion of dmu_objset_open_impl().
308 return (zfs_vdev_async_write_max_active
);
311 * Sync tasks correspond to interactive user actions. To reduce the
312 * execution time of those actions we push data out as fast as possible.
314 if (spa_has_pending_synctask(spa
))
315 return (zfs_vdev_async_write_max_active
);
317 dirty
= dp
->dp_dirty_total
;
318 if (dirty
< min_bytes
)
319 return (zfs_vdev_async_write_min_active
);
320 if (dirty
> max_bytes
)
321 return (zfs_vdev_async_write_max_active
);
324 * linear interpolation:
325 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
326 * move right by min_bytes
327 * move up by min_writes
329 writes
= (dirty
- min_bytes
) *
330 (zfs_vdev_async_write_max_active
-
331 zfs_vdev_async_write_min_active
) /
332 (max_bytes
- min_bytes
) +
333 zfs_vdev_async_write_min_active
;
334 ASSERT3U(writes
, >=, zfs_vdev_async_write_min_active
);
335 ASSERT3U(writes
, <=, zfs_vdev_async_write_max_active
);
340 vdev_queue_class_max_active(spa_t
*spa
, zio_priority_t p
)
343 case ZIO_PRIORITY_SYNC_READ
:
344 return (zfs_vdev_sync_read_max_active
);
345 case ZIO_PRIORITY_SYNC_WRITE
:
346 return (zfs_vdev_sync_write_max_active
);
347 case ZIO_PRIORITY_ASYNC_READ
:
348 return (zfs_vdev_async_read_max_active
);
349 case ZIO_PRIORITY_ASYNC_WRITE
:
350 return (vdev_queue_max_async_writes(spa
));
351 case ZIO_PRIORITY_SCRUB
:
352 return (zfs_vdev_scrub_max_active
);
353 case ZIO_PRIORITY_REMOVAL
:
354 return (zfs_vdev_removal_max_active
);
355 case ZIO_PRIORITY_INITIALIZING
:
356 return (zfs_vdev_initializing_max_active
);
357 case ZIO_PRIORITY_TRIM
:
358 return (zfs_vdev_trim_max_active
);
359 case ZIO_PRIORITY_REBUILD
:
360 return (zfs_vdev_rebuild_max_active
);
362 panic("invalid priority %u", p
);
368 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
369 * there is no eligible class.
371 static zio_priority_t
372 vdev_queue_class_to_issue(vdev_queue_t
*vq
)
374 spa_t
*spa
= vq
->vq_vdev
->vdev_spa
;
377 if (avl_numnodes(&vq
->vq_active_tree
) >= zfs_vdev_max_active
)
378 return (ZIO_PRIORITY_NUM_QUEUEABLE
);
380 /* find a queue that has not reached its minimum # outstanding i/os */
381 for (p
= 0; p
< ZIO_PRIORITY_NUM_QUEUEABLE
; p
++) {
382 if (avl_numnodes(vdev_queue_class_tree(vq
, p
)) > 0 &&
383 vq
->vq_class
[p
].vqc_active
<
384 vdev_queue_class_min_active(p
))
389 * If we haven't found a queue, look for one that hasn't reached its
390 * maximum # outstanding i/os.
392 for (p
= 0; p
< ZIO_PRIORITY_NUM_QUEUEABLE
; p
++) {
393 if (avl_numnodes(vdev_queue_class_tree(vq
, p
)) > 0 &&
394 vq
->vq_class
[p
].vqc_active
<
395 vdev_queue_class_max_active(spa
, p
))
399 /* No eligible queued i/os */
400 return (ZIO_PRIORITY_NUM_QUEUEABLE
);
404 vdev_queue_init(vdev_t
*vd
)
406 vdev_queue_t
*vq
= &vd
->vdev_queue
;
409 mutex_init(&vq
->vq_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
411 taskq_init_ent(&vd
->vdev_queue
.vq_io_search
.io_tqent
);
413 avl_create(&vq
->vq_active_tree
, vdev_queue_offset_compare
,
414 sizeof (zio_t
), offsetof(struct zio
, io_queue_node
));
415 avl_create(vdev_queue_type_tree(vq
, ZIO_TYPE_READ
),
416 vdev_queue_offset_compare
, sizeof (zio_t
),
417 offsetof(struct zio
, io_offset_node
));
418 avl_create(vdev_queue_type_tree(vq
, ZIO_TYPE_WRITE
),
419 vdev_queue_offset_compare
, sizeof (zio_t
),
420 offsetof(struct zio
, io_offset_node
));
421 avl_create(vdev_queue_type_tree(vq
, ZIO_TYPE_TRIM
),
422 vdev_queue_offset_compare
, sizeof (zio_t
),
423 offsetof(struct zio
, io_offset_node
));
425 for (p
= 0; p
< ZIO_PRIORITY_NUM_QUEUEABLE
; p
++) {
426 int (*compfn
) (const void *, const void *);
429 * The synchronous/trim i/o queues are dispatched in FIFO rather
430 * than LBA order. This provides more consistent latency for
433 if (p
== ZIO_PRIORITY_SYNC_READ
||
434 p
== ZIO_PRIORITY_SYNC_WRITE
||
435 p
== ZIO_PRIORITY_TRIM
) {
436 compfn
= vdev_queue_timestamp_compare
;
438 compfn
= vdev_queue_offset_compare
;
440 avl_create(vdev_queue_class_tree(vq
, p
), compfn
,
441 sizeof (zio_t
), offsetof(struct zio
, io_queue_node
));
444 vq
->vq_last_offset
= 0;
448 vdev_queue_fini(vdev_t
*vd
)
450 vdev_queue_t
*vq
= &vd
->vdev_queue
;
452 for (zio_priority_t p
= 0; p
< ZIO_PRIORITY_NUM_QUEUEABLE
; p
++)
453 avl_destroy(vdev_queue_class_tree(vq
, p
));
454 avl_destroy(&vq
->vq_active_tree
);
455 avl_destroy(vdev_queue_type_tree(vq
, ZIO_TYPE_READ
));
456 avl_destroy(vdev_queue_type_tree(vq
, ZIO_TYPE_WRITE
));
457 avl_destroy(vdev_queue_type_tree(vq
, ZIO_TYPE_TRIM
));
459 mutex_destroy(&vq
->vq_lock
);
463 vdev_queue_io_add(vdev_queue_t
*vq
, zio_t
*zio
)
465 spa_t
*spa
= zio
->io_spa
;
466 spa_history_kstat_t
*shk
= &spa
->spa_stats
.io_history
;
468 ASSERT3U(zio
->io_priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
469 avl_add(vdev_queue_class_tree(vq
, zio
->io_priority
), zio
);
470 avl_add(vdev_queue_type_tree(vq
, zio
->io_type
), zio
);
472 if (shk
->kstat
!= NULL
) {
473 mutex_enter(&shk
->lock
);
474 kstat_waitq_enter(shk
->kstat
->ks_data
);
475 mutex_exit(&shk
->lock
);
480 vdev_queue_io_remove(vdev_queue_t
*vq
, zio_t
*zio
)
482 spa_t
*spa
= zio
->io_spa
;
483 spa_history_kstat_t
*shk
= &spa
->spa_stats
.io_history
;
485 ASSERT3U(zio
->io_priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
486 avl_remove(vdev_queue_class_tree(vq
, zio
->io_priority
), zio
);
487 avl_remove(vdev_queue_type_tree(vq
, zio
->io_type
), zio
);
489 if (shk
->kstat
!= NULL
) {
490 mutex_enter(&shk
->lock
);
491 kstat_waitq_exit(shk
->kstat
->ks_data
);
492 mutex_exit(&shk
->lock
);
497 vdev_queue_pending_add(vdev_queue_t
*vq
, zio_t
*zio
)
499 spa_t
*spa
= zio
->io_spa
;
500 spa_history_kstat_t
*shk
= &spa
->spa_stats
.io_history
;
502 ASSERT(MUTEX_HELD(&vq
->vq_lock
));
503 ASSERT3U(zio
->io_priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
504 vq
->vq_class
[zio
->io_priority
].vqc_active
++;
505 avl_add(&vq
->vq_active_tree
, zio
);
507 if (shk
->kstat
!= NULL
) {
508 mutex_enter(&shk
->lock
);
509 kstat_runq_enter(shk
->kstat
->ks_data
);
510 mutex_exit(&shk
->lock
);
515 vdev_queue_pending_remove(vdev_queue_t
*vq
, zio_t
*zio
)
517 spa_t
*spa
= zio
->io_spa
;
518 spa_history_kstat_t
*shk
= &spa
->spa_stats
.io_history
;
520 ASSERT(MUTEX_HELD(&vq
->vq_lock
));
521 ASSERT3U(zio
->io_priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
522 vq
->vq_class
[zio
->io_priority
].vqc_active
--;
523 avl_remove(&vq
->vq_active_tree
, zio
);
525 if (shk
->kstat
!= NULL
) {
526 kstat_io_t
*ksio
= shk
->kstat
->ks_data
;
528 mutex_enter(&shk
->lock
);
529 kstat_runq_exit(ksio
);
530 if (zio
->io_type
== ZIO_TYPE_READ
) {
532 ksio
->nread
+= zio
->io_size
;
533 } else if (zio
->io_type
== ZIO_TYPE_WRITE
) {
535 ksio
->nwritten
+= zio
->io_size
;
537 mutex_exit(&shk
->lock
);
542 vdev_queue_agg_io_done(zio_t
*aio
)
544 abd_free(aio
->io_abd
);
548 * Compute the range spanned by two i/os, which is the endpoint of the last
549 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
550 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
551 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
553 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
554 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
557 * Sufficiently adjacent io_offset's in ZIOs will be aggregated. We do this
558 * by creating a gang ABD from the adjacent ZIOs io_abd's. By using
559 * a gang ABD we avoid doing memory copies to and from the parent,
560 * child ZIOs. The gang ABD also accounts for gaps between adjacent
561 * io_offsets by simply getting the zero ABD for writes or allocating
562 * a new ABD for reads and placing them in the gang ABD as well.
565 vdev_queue_aggregate(vdev_queue_t
*vq
, zio_t
*zio
)
567 zio_t
*first
, *last
, *aio
, *dio
, *mandatory
, *nio
;
568 zio_link_t
*zl
= NULL
;
573 boolean_t stretch
= B_FALSE
;
574 avl_tree_t
*t
= vdev_queue_type_tree(vq
, zio
->io_type
);
575 enum zio_flag flags
= zio
->io_flags
& ZIO_FLAG_AGG_INHERIT
;
576 uint64_t next_offset
;
579 maxblocksize
= spa_maxblocksize(vq
->vq_vdev
->vdev_spa
);
580 if (vq
->vq_vdev
->vdev_nonrot
)
581 limit
= zfs_vdev_aggregation_limit_non_rotating
;
583 limit
= zfs_vdev_aggregation_limit
;
584 limit
= MAX(MIN(limit
, maxblocksize
), 0);
586 if (zio
->io_flags
& ZIO_FLAG_DONT_AGGREGATE
|| limit
== 0)
590 * While TRIM commands could be aggregated based on offset this
591 * behavior is disabled until it's determined to be beneficial.
593 if (zio
->io_type
== ZIO_TYPE_TRIM
&& !zfs_vdev_aggregate_trim
)
598 if (zio
->io_type
== ZIO_TYPE_READ
)
599 maxgap
= zfs_vdev_read_gap_limit
;
602 * We can aggregate I/Os that are sufficiently adjacent and of
603 * the same flavor, as expressed by the AGG_INHERIT flags.
604 * The latter requirement is necessary so that certain
605 * attributes of the I/O, such as whether it's a normal I/O
606 * or a scrub/resilver, can be preserved in the aggregate.
607 * We can include optional I/Os, but don't allow them
608 * to begin a range as they add no benefit in that situation.
612 * We keep track of the last non-optional I/O.
614 mandatory
= (first
->io_flags
& ZIO_FLAG_OPTIONAL
) ? NULL
: first
;
617 * Walk backwards through sufficiently contiguous I/Os
618 * recording the last non-optional I/O.
620 while ((dio
= AVL_PREV(t
, first
)) != NULL
&&
621 (dio
->io_flags
& ZIO_FLAG_AGG_INHERIT
) == flags
&&
622 IO_SPAN(dio
, last
) <= limit
&&
623 IO_GAP(dio
, first
) <= maxgap
&&
624 dio
->io_type
== zio
->io_type
) {
626 if (mandatory
== NULL
&& !(first
->io_flags
& ZIO_FLAG_OPTIONAL
))
631 * Skip any initial optional I/Os.
633 while ((first
->io_flags
& ZIO_FLAG_OPTIONAL
) && first
!= last
) {
634 first
= AVL_NEXT(t
, first
);
635 ASSERT(first
!= NULL
);
640 * Walk forward through sufficiently contiguous I/Os.
641 * The aggregation limit does not apply to optional i/os, so that
642 * we can issue contiguous writes even if they are larger than the
645 while ((dio
= AVL_NEXT(t
, last
)) != NULL
&&
646 (dio
->io_flags
& ZIO_FLAG_AGG_INHERIT
) == flags
&&
647 (IO_SPAN(first
, dio
) <= limit
||
648 (dio
->io_flags
& ZIO_FLAG_OPTIONAL
)) &&
649 IO_SPAN(first
, dio
) <= maxblocksize
&&
650 IO_GAP(last
, dio
) <= maxgap
&&
651 dio
->io_type
== zio
->io_type
) {
653 if (!(last
->io_flags
& ZIO_FLAG_OPTIONAL
))
658 * Now that we've established the range of the I/O aggregation
659 * we must decide what to do with trailing optional I/Os.
660 * For reads, there's nothing to do. While we are unable to
661 * aggregate further, it's possible that a trailing optional
662 * I/O would allow the underlying device to aggregate with
663 * subsequent I/Os. We must therefore determine if the next
664 * non-optional I/O is close enough to make aggregation
667 if (zio
->io_type
== ZIO_TYPE_WRITE
&& mandatory
!= NULL
) {
669 while ((dio
= AVL_NEXT(t
, nio
)) != NULL
&&
670 IO_GAP(nio
, dio
) == 0 &&
671 IO_GAP(mandatory
, dio
) <= zfs_vdev_write_gap_limit
) {
673 if (!(nio
->io_flags
& ZIO_FLAG_OPTIONAL
)) {
682 * We are going to include an optional io in our aggregated
683 * span, thus closing the write gap. Only mandatory i/os can
684 * start aggregated spans, so make sure that the next i/o
685 * after our span is mandatory.
687 dio
= AVL_NEXT(t
, last
);
688 dio
->io_flags
&= ~ZIO_FLAG_OPTIONAL
;
690 /* do not include the optional i/o */
691 while (last
!= mandatory
&& last
!= first
) {
692 ASSERT(last
->io_flags
& ZIO_FLAG_OPTIONAL
);
693 last
= AVL_PREV(t
, last
);
694 ASSERT(last
!= NULL
);
701 size
= IO_SPAN(first
, last
);
702 ASSERT3U(size
, <=, maxblocksize
);
704 abd
= abd_alloc_gang_abd();
708 aio
= zio_vdev_delegated_io(first
->io_vd
, first
->io_offset
,
709 abd
, size
, first
->io_type
, zio
->io_priority
,
710 flags
| ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_QUEUE
,
711 vdev_queue_agg_io_done
, NULL
);
712 aio
->io_timestamp
= first
->io_timestamp
;
715 next_offset
= first
->io_offset
;
718 nio
= AVL_NEXT(t
, dio
);
719 zio_add_child(dio
, aio
);
720 vdev_queue_io_remove(vq
, dio
);
722 if (dio
->io_offset
!= next_offset
) {
723 /* allocate a buffer for a read gap */
724 ASSERT3U(dio
->io_type
, ==, ZIO_TYPE_READ
);
725 ASSERT3U(dio
->io_offset
, >, next_offset
);
726 abd
= abd_alloc_for_io(
727 dio
->io_offset
- next_offset
, B_TRUE
);
728 abd_gang_add(aio
->io_abd
, abd
, B_TRUE
);
731 (dio
->io_size
!= abd_get_size(dio
->io_abd
))) {
732 /* abd size not the same as IO size */
733 ASSERT3U(abd_get_size(dio
->io_abd
), >, dio
->io_size
);
734 abd
= abd_get_offset_size(dio
->io_abd
, 0, dio
->io_size
);
735 abd_gang_add(aio
->io_abd
, abd
, B_TRUE
);
737 if (dio
->io_flags
& ZIO_FLAG_NODATA
) {
738 /* allocate a buffer for a write gap */
739 ASSERT3U(dio
->io_type
, ==, ZIO_TYPE_WRITE
);
740 ASSERT3P(dio
->io_abd
, ==, NULL
);
741 abd_gang_add(aio
->io_abd
,
742 abd_get_zeros(dio
->io_size
), B_TRUE
);
745 * We pass B_FALSE to abd_gang_add()
746 * because we did not allocate a new
747 * ABD, so it is assumed the caller
748 * will free this ABD.
750 abd_gang_add(aio
->io_abd
, dio
->io_abd
,
754 next_offset
= dio
->io_offset
+ dio
->io_size
;
755 } while (dio
!= last
);
756 ASSERT3U(abd_get_size(aio
->io_abd
), ==, aio
->io_size
);
759 * We need to drop the vdev queue's lock during zio_execute() to
760 * avoid a deadlock that we could encounter due to lock order
761 * reversal between vq_lock and io_lock in zio_change_priority().
763 mutex_exit(&vq
->vq_lock
);
764 while ((dio
= zio_walk_parents(aio
, &zl
)) != NULL
) {
765 ASSERT3U(dio
->io_type
, ==, aio
->io_type
);
767 zio_vdev_io_bypass(dio
);
770 mutex_enter(&vq
->vq_lock
);
776 vdev_queue_io_to_issue(vdev_queue_t
*vq
)
784 ASSERT(MUTEX_HELD(&vq
->vq_lock
));
786 p
= vdev_queue_class_to_issue(vq
);
788 if (p
== ZIO_PRIORITY_NUM_QUEUEABLE
) {
789 /* No eligible queued i/os */
794 * For LBA-ordered queues (async / scrub / initializing), issue the
795 * i/o which follows the most recently issued i/o in LBA (offset) order.
797 * For FIFO queues (sync/trim), issue the i/o with the lowest timestamp.
799 tree
= vdev_queue_class_tree(vq
, p
);
800 vq
->vq_io_search
.io_timestamp
= 0;
801 vq
->vq_io_search
.io_offset
= vq
->vq_last_offset
- 1;
802 VERIFY3P(avl_find(tree
, &vq
->vq_io_search
, &idx
), ==, NULL
);
803 zio
= avl_nearest(tree
, idx
, AVL_AFTER
);
805 zio
= avl_first(tree
);
806 ASSERT3U(zio
->io_priority
, ==, p
);
808 aio
= vdev_queue_aggregate(vq
, zio
);
812 vdev_queue_io_remove(vq
, zio
);
815 * If the I/O is or was optional and therefore has no data, we need to
816 * simply discard it. We need to drop the vdev queue's lock to avoid a
817 * deadlock that we could encounter since this I/O will complete
820 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
821 mutex_exit(&vq
->vq_lock
);
822 zio_vdev_io_bypass(zio
);
824 mutex_enter(&vq
->vq_lock
);
828 vdev_queue_pending_add(vq
, zio
);
829 vq
->vq_last_offset
= zio
->io_offset
+ zio
->io_size
;
835 vdev_queue_io(zio_t
*zio
)
837 vdev_queue_t
*vq
= &zio
->io_vd
->vdev_queue
;
840 if (zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)
844 * Children i/os inherent their parent's priority, which might
845 * not match the child's i/o type. Fix it up here.
847 if (zio
->io_type
== ZIO_TYPE_READ
) {
848 ASSERT(zio
->io_priority
!= ZIO_PRIORITY_TRIM
);
850 if (zio
->io_priority
!= ZIO_PRIORITY_SYNC_READ
&&
851 zio
->io_priority
!= ZIO_PRIORITY_ASYNC_READ
&&
852 zio
->io_priority
!= ZIO_PRIORITY_SCRUB
&&
853 zio
->io_priority
!= ZIO_PRIORITY_REMOVAL
&&
854 zio
->io_priority
!= ZIO_PRIORITY_INITIALIZING
&&
855 zio
->io_priority
!= ZIO_PRIORITY_REBUILD
) {
856 zio
->io_priority
= ZIO_PRIORITY_ASYNC_READ
;
858 } else if (zio
->io_type
== ZIO_TYPE_WRITE
) {
859 ASSERT(zio
->io_priority
!= ZIO_PRIORITY_TRIM
);
861 if (zio
->io_priority
!= ZIO_PRIORITY_SYNC_WRITE
&&
862 zio
->io_priority
!= ZIO_PRIORITY_ASYNC_WRITE
&&
863 zio
->io_priority
!= ZIO_PRIORITY_REMOVAL
&&
864 zio
->io_priority
!= ZIO_PRIORITY_INITIALIZING
&&
865 zio
->io_priority
!= ZIO_PRIORITY_REBUILD
) {
866 zio
->io_priority
= ZIO_PRIORITY_ASYNC_WRITE
;
869 ASSERT(zio
->io_type
== ZIO_TYPE_TRIM
);
870 ASSERT(zio
->io_priority
== ZIO_PRIORITY_TRIM
);
873 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_QUEUE
;
875 mutex_enter(&vq
->vq_lock
);
876 zio
->io_timestamp
= gethrtime();
877 vdev_queue_io_add(vq
, zio
);
878 nio
= vdev_queue_io_to_issue(vq
);
879 mutex_exit(&vq
->vq_lock
);
884 if (nio
->io_done
== vdev_queue_agg_io_done
) {
893 vdev_queue_io_done(zio_t
*zio
)
895 vdev_queue_t
*vq
= &zio
->io_vd
->vdev_queue
;
898 mutex_enter(&vq
->vq_lock
);
900 vdev_queue_pending_remove(vq
, zio
);
902 zio
->io_delta
= gethrtime() - zio
->io_timestamp
;
903 vq
->vq_io_complete_ts
= gethrtime();
904 vq
->vq_io_delta_ts
= vq
->vq_io_complete_ts
- zio
->io_timestamp
;
906 while ((nio
= vdev_queue_io_to_issue(vq
)) != NULL
) {
907 mutex_exit(&vq
->vq_lock
);
908 if (nio
->io_done
== vdev_queue_agg_io_done
) {
911 zio_vdev_io_reissue(nio
);
914 mutex_enter(&vq
->vq_lock
);
917 mutex_exit(&vq
->vq_lock
);
921 vdev_queue_change_io_priority(zio_t
*zio
, zio_priority_t priority
)
923 vdev_queue_t
*vq
= &zio
->io_vd
->vdev_queue
;
927 * ZIO_PRIORITY_NOW is used by the vdev cache code and the aggregate zio
928 * code to issue IOs without adding them to the vdev queue. In this
929 * case, the zio is already going to be issued as quickly as possible
930 * and so it doesn't need any reprioritization to help.
932 if (zio
->io_priority
== ZIO_PRIORITY_NOW
)
935 ASSERT3U(zio
->io_priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
936 ASSERT3U(priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
938 if (zio
->io_type
== ZIO_TYPE_READ
) {
939 if (priority
!= ZIO_PRIORITY_SYNC_READ
&&
940 priority
!= ZIO_PRIORITY_ASYNC_READ
&&
941 priority
!= ZIO_PRIORITY_SCRUB
)
942 priority
= ZIO_PRIORITY_ASYNC_READ
;
944 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
945 if (priority
!= ZIO_PRIORITY_SYNC_WRITE
&&
946 priority
!= ZIO_PRIORITY_ASYNC_WRITE
)
947 priority
= ZIO_PRIORITY_ASYNC_WRITE
;
950 mutex_enter(&vq
->vq_lock
);
953 * If the zio is in none of the queues we can simply change
954 * the priority. If the zio is waiting to be submitted we must
955 * remove it from the queue and re-insert it with the new priority.
956 * Otherwise, the zio is currently active and we cannot change its
959 tree
= vdev_queue_class_tree(vq
, zio
->io_priority
);
960 if (avl_find(tree
, zio
, NULL
) == zio
) {
961 avl_remove(vdev_queue_class_tree(vq
, zio
->io_priority
), zio
);
962 zio
->io_priority
= priority
;
963 avl_add(vdev_queue_class_tree(vq
, zio
->io_priority
), zio
);
964 } else if (avl_find(&vq
->vq_active_tree
, zio
, NULL
) != zio
) {
965 zio
->io_priority
= priority
;
968 mutex_exit(&vq
->vq_lock
);
972 * As these two methods are only used for load calculations we're not
973 * concerned if we get an incorrect value on 32bit platforms due to lack of
974 * vq_lock mutex use here, instead we prefer to keep it lock free for
978 vdev_queue_length(vdev_t
*vd
)
980 return (avl_numnodes(&vd
->vdev_queue
.vq_active_tree
));
984 vdev_queue_last_offset(vdev_t
*vd
)
986 return (vd
->vdev_queue
.vq_last_offset
);
990 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, aggregation_limit
, INT
, ZMOD_RW
,
991 "Max vdev I/O aggregation size");
993 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, aggregation_limit_non_rotating
, INT
, ZMOD_RW
,
994 "Max vdev I/O aggregation size for non-rotating media");
996 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, aggregate_trim
, INT
, ZMOD_RW
,
997 "Allow TRIM I/O to be aggregated");
999 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, read_gap_limit
, INT
, ZMOD_RW
,
1000 "Aggregate read I/O over gap");
1002 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, write_gap_limit
, INT
, ZMOD_RW
,
1003 "Aggregate write I/O over gap");
1005 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, max_active
, INT
, ZMOD_RW
,
1006 "Maximum number of active I/Os per vdev");
1008 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, async_write_active_max_dirty_percent
, INT
, ZMOD_RW
,
1009 "Async write concurrency max threshold");
1011 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, async_write_active_min_dirty_percent
, INT
, ZMOD_RW
,
1012 "Async write concurrency min threshold");
1014 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, async_read_max_active
, INT
, ZMOD_RW
,
1015 "Max active async read I/Os per vdev");
1017 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, async_read_min_active
, INT
, ZMOD_RW
,
1018 "Min active async read I/Os per vdev");
1020 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, async_write_max_active
, INT
, ZMOD_RW
,
1021 "Max active async write I/Os per vdev");
1023 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, async_write_min_active
, INT
, ZMOD_RW
,
1024 "Min active async write I/Os per vdev");
1026 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, initializing_max_active
, INT
, ZMOD_RW
,
1027 "Max active initializing I/Os per vdev");
1029 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, initializing_min_active
, INT
, ZMOD_RW
,
1030 "Min active initializing I/Os per vdev");
1032 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, removal_max_active
, INT
, ZMOD_RW
,
1033 "Max active removal I/Os per vdev");
1035 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, removal_min_active
, INT
, ZMOD_RW
,
1036 "Min active removal I/Os per vdev");
1038 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, scrub_max_active
, INT
, ZMOD_RW
,
1039 "Max active scrub I/Os per vdev");
1041 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, scrub_min_active
, INT
, ZMOD_RW
,
1042 "Min active scrub I/Os per vdev");
1044 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, sync_read_max_active
, INT
, ZMOD_RW
,
1045 "Max active sync read I/Os per vdev");
1047 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, sync_read_min_active
, INT
, ZMOD_RW
,
1048 "Min active sync read I/Os per vdev");
1050 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, sync_write_max_active
, INT
, ZMOD_RW
,
1051 "Max active sync write I/Os per vdev");
1053 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, sync_write_min_active
, INT
, ZMOD_RW
,
1054 "Min active sync write I/Os per vdev");
1056 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, trim_max_active
, INT
, ZMOD_RW
,
1057 "Max active trim/discard I/Os per vdev");
1059 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, trim_min_active
, INT
, ZMOD_RW
,
1060 "Min active trim/discard I/Os per vdev");
1062 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, rebuild_max_active
, INT
, ZMOD_RW
,
1063 "Max active rebuild I/Os per vdev");
1065 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, rebuild_min_active
, INT
, ZMOD_RW
,
1066 "Min active rebuild I/Os per vdev");
1068 ZFS_MODULE_PARAM(zfs_vdev
, zfs_vdev_
, queue_depth_pct
, INT
, ZMOD_RW
,
1069 "Queue depth percentage for each top-level vdev");