4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
27 * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
28 * Copyright (c) 2014 Integros [integros.com]
31 #include <sys/zfs_context.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/spa_impl.h>
36 #include <sys/dsl_pool.h>
37 #include <sys/metaslab_impl.h>
44 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The
45 * I/O scheduler determines when and in what order those operations are
46 * issued. The I/O scheduler divides operations into five I/O classes
47 * prioritized in the following order: sync read, sync write, async read,
48 * async write, and scrub/resilver. 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_removal_min_active
= 1;
152 uint32_t zfs_vdev_removal_max_active
= 2;
153 uint32_t zfs_vdev_initializing_min_active
= 1;
154 uint32_t zfs_vdev_initializing_max_active
= 1;
157 * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent
158 * dirty data, use zfs_vdev_async_write_min_active. When it has more than
159 * zfs_vdev_async_write_active_max_dirty_percent, use
160 * zfs_vdev_async_write_max_active. The value is linearly interpolated
161 * between min and max.
163 int zfs_vdev_async_write_active_min_dirty_percent
= 30;
164 int zfs_vdev_async_write_active_max_dirty_percent
= 60;
167 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
168 * For read I/Os, we also aggregate across small adjacency gaps; for writes
169 * we include spans of optional I/Os to aid aggregation at the disk even when
170 * they aren't able to help us aggregate at this level.
172 int zfs_vdev_aggregation_limit
= SPA_OLD_MAXBLOCKSIZE
;
173 int zfs_vdev_read_gap_limit
= 32 << 10;
174 int zfs_vdev_write_gap_limit
= 4 << 10;
177 * Define the queue depth percentage for each top-level. This percentage is
178 * used in conjunction with zfs_vdev_async_max_active to determine how many
179 * allocations a specific top-level vdev should handle. Once the queue depth
180 * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100
181 * then allocator will stop allocating blocks on that top-level device.
182 * The default kernel setting is 1000% which will yield 100 allocations per
183 * device. For userland testing, the default setting is 300% which equates
184 * to 30 allocations per device.
187 int zfs_vdev_queue_depth_pct
= 1000;
189 int zfs_vdev_queue_depth_pct
= 300;
193 * When performing allocations for a given metaslab, we want to make sure that
194 * there are enough IOs to aggregate together to improve throughput. We want to
195 * ensure that there are at least 128k worth of IOs that can be aggregated, and
196 * we assume that the average allocation size is 4k, so we need the queue depth
197 * to be 32 per allocator to get good aggregation of sequential writes.
199 int zfs_vdev_def_queue_depth
= 32;
203 vdev_queue_offset_compare(const void *x1
, const void *x2
)
205 const zio_t
*z1
= x1
;
206 const zio_t
*z2
= x2
;
208 if (z1
->io_offset
< z2
->io_offset
)
210 if (z1
->io_offset
> z2
->io_offset
)
221 static inline avl_tree_t
*
222 vdev_queue_class_tree(vdev_queue_t
*vq
, zio_priority_t p
)
224 return (&vq
->vq_class
[p
].vqc_queued_tree
);
227 static inline avl_tree_t
*
228 vdev_queue_type_tree(vdev_queue_t
*vq
, zio_type_t t
)
230 ASSERT(t
== ZIO_TYPE_READ
|| t
== ZIO_TYPE_WRITE
);
231 if (t
== ZIO_TYPE_READ
)
232 return (&vq
->vq_read_offset_tree
);
234 return (&vq
->vq_write_offset_tree
);
238 vdev_queue_timestamp_compare(const void *x1
, const void *x2
)
240 const zio_t
*z1
= x1
;
241 const zio_t
*z2
= x2
;
243 if (z1
->io_timestamp
< z2
->io_timestamp
)
245 if (z1
->io_timestamp
> z2
->io_timestamp
)
257 vdev_queue_init(vdev_t
*vd
)
259 vdev_queue_t
*vq
= &vd
->vdev_queue
;
261 mutex_init(&vq
->vq_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
264 avl_create(&vq
->vq_active_tree
, vdev_queue_offset_compare
,
265 sizeof (zio_t
), offsetof(struct zio
, io_queue_node
));
266 avl_create(vdev_queue_type_tree(vq
, ZIO_TYPE_READ
),
267 vdev_queue_offset_compare
, sizeof (zio_t
),
268 offsetof(struct zio
, io_offset_node
));
269 avl_create(vdev_queue_type_tree(vq
, ZIO_TYPE_WRITE
),
270 vdev_queue_offset_compare
, sizeof (zio_t
),
271 offsetof(struct zio
, io_offset_node
));
273 for (zio_priority_t p
= 0; p
< ZIO_PRIORITY_NUM_QUEUEABLE
; p
++) {
274 int (*compfn
) (const void *, const void *);
277 * The synchronous i/o queues are dispatched in FIFO rather
278 * than LBA order. This provides more consistent latency for
281 if (p
== ZIO_PRIORITY_SYNC_READ
|| p
== ZIO_PRIORITY_SYNC_WRITE
)
282 compfn
= vdev_queue_timestamp_compare
;
284 compfn
= vdev_queue_offset_compare
;
286 avl_create(vdev_queue_class_tree(vq
, p
), compfn
,
287 sizeof (zio_t
), offsetof(struct zio
, io_queue_node
));
292 vdev_queue_fini(vdev_t
*vd
)
294 vdev_queue_t
*vq
= &vd
->vdev_queue
;
296 for (zio_priority_t p
= 0; p
< ZIO_PRIORITY_NUM_QUEUEABLE
; p
++)
297 avl_destroy(vdev_queue_class_tree(vq
, p
));
298 avl_destroy(&vq
->vq_active_tree
);
299 avl_destroy(vdev_queue_type_tree(vq
, ZIO_TYPE_READ
));
300 avl_destroy(vdev_queue_type_tree(vq
, ZIO_TYPE_WRITE
));
302 mutex_destroy(&vq
->vq_lock
);
306 vdev_queue_io_add(vdev_queue_t
*vq
, zio_t
*zio
)
308 spa_t
*spa
= zio
->io_spa
;
310 ASSERT3U(zio
->io_priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
311 avl_add(vdev_queue_class_tree(vq
, zio
->io_priority
), zio
);
312 avl_add(vdev_queue_type_tree(vq
, zio
->io_type
), zio
);
314 mutex_enter(&spa
->spa_iokstat_lock
);
315 spa
->spa_queue_stats
[zio
->io_priority
].spa_queued
++;
316 if (spa
->spa_iokstat
!= NULL
)
317 kstat_waitq_enter(spa
->spa_iokstat
->ks_data
);
318 mutex_exit(&spa
->spa_iokstat_lock
);
322 vdev_queue_io_remove(vdev_queue_t
*vq
, zio_t
*zio
)
324 spa_t
*spa
= zio
->io_spa
;
326 ASSERT3U(zio
->io_priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
327 avl_remove(vdev_queue_class_tree(vq
, zio
->io_priority
), zio
);
328 avl_remove(vdev_queue_type_tree(vq
, zio
->io_type
), zio
);
330 mutex_enter(&spa
->spa_iokstat_lock
);
331 ASSERT3U(spa
->spa_queue_stats
[zio
->io_priority
].spa_queued
, >, 0);
332 spa
->spa_queue_stats
[zio
->io_priority
].spa_queued
--;
333 if (spa
->spa_iokstat
!= NULL
)
334 kstat_waitq_exit(spa
->spa_iokstat
->ks_data
);
335 mutex_exit(&spa
->spa_iokstat_lock
);
339 vdev_queue_pending_add(vdev_queue_t
*vq
, zio_t
*zio
)
341 spa_t
*spa
= zio
->io_spa
;
342 ASSERT(MUTEX_HELD(&vq
->vq_lock
));
343 ASSERT3U(zio
->io_priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
344 vq
->vq_class
[zio
->io_priority
].vqc_active
++;
345 avl_add(&vq
->vq_active_tree
, zio
);
347 mutex_enter(&spa
->spa_iokstat_lock
);
348 spa
->spa_queue_stats
[zio
->io_priority
].spa_active
++;
349 if (spa
->spa_iokstat
!= NULL
)
350 kstat_runq_enter(spa
->spa_iokstat
->ks_data
);
351 mutex_exit(&spa
->spa_iokstat_lock
);
355 vdev_queue_pending_remove(vdev_queue_t
*vq
, zio_t
*zio
)
357 spa_t
*spa
= zio
->io_spa
;
358 ASSERT(MUTEX_HELD(&vq
->vq_lock
));
359 ASSERT3U(zio
->io_priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
360 vq
->vq_class
[zio
->io_priority
].vqc_active
--;
361 avl_remove(&vq
->vq_active_tree
, zio
);
363 mutex_enter(&spa
->spa_iokstat_lock
);
364 ASSERT3U(spa
->spa_queue_stats
[zio
->io_priority
].spa_active
, >, 0);
365 spa
->spa_queue_stats
[zio
->io_priority
].spa_active
--;
366 if (spa
->spa_iokstat
!= NULL
) {
367 kstat_io_t
*ksio
= spa
->spa_iokstat
->ks_data
;
369 kstat_runq_exit(spa
->spa_iokstat
->ks_data
);
370 if (zio
->io_type
== ZIO_TYPE_READ
) {
372 ksio
->nread
+= zio
->io_size
;
373 } else if (zio
->io_type
== ZIO_TYPE_WRITE
) {
375 ksio
->nwritten
+= zio
->io_size
;
378 mutex_exit(&spa
->spa_iokstat_lock
);
382 vdev_queue_agg_io_done(zio_t
*aio
)
384 if (aio
->io_type
== ZIO_TYPE_READ
) {
386 zio_link_t
*zl
= NULL
;
387 while ((pio
= zio_walk_parents(aio
, &zl
)) != NULL
) {
388 abd_copy_off(pio
->io_abd
, aio
->io_abd
,
389 0, pio
->io_offset
- aio
->io_offset
, pio
->io_size
);
393 abd_free(aio
->io_abd
);
397 vdev_queue_class_min_active(zio_priority_t p
)
400 case ZIO_PRIORITY_SYNC_READ
:
401 return (zfs_vdev_sync_read_min_active
);
402 case ZIO_PRIORITY_SYNC_WRITE
:
403 return (zfs_vdev_sync_write_min_active
);
404 case ZIO_PRIORITY_ASYNC_READ
:
405 return (zfs_vdev_async_read_min_active
);
406 case ZIO_PRIORITY_ASYNC_WRITE
:
407 return (zfs_vdev_async_write_min_active
);
408 case ZIO_PRIORITY_SCRUB
:
409 return (zfs_vdev_scrub_min_active
);
410 case ZIO_PRIORITY_REMOVAL
:
411 return (zfs_vdev_removal_min_active
);
412 case ZIO_PRIORITY_INITIALIZING
:
413 return (zfs_vdev_initializing_min_active
);
415 panic("invalid priority %u", p
);
421 vdev_queue_max_async_writes(spa_t
*spa
)
424 uint64_t dirty
= spa
->spa_dsl_pool
->dp_dirty_total
;
425 uint64_t min_bytes
= zfs_dirty_data_max
*
426 zfs_vdev_async_write_active_min_dirty_percent
/ 100;
427 uint64_t max_bytes
= zfs_dirty_data_max
*
428 zfs_vdev_async_write_active_max_dirty_percent
/ 100;
431 * Sync tasks correspond to interactive user actions. To reduce the
432 * execution time of those actions we push data out as fast as possible.
434 if (spa_has_pending_synctask(spa
)) {
435 return (zfs_vdev_async_write_max_active
);
438 if (dirty
< min_bytes
)
439 return (zfs_vdev_async_write_min_active
);
440 if (dirty
> max_bytes
)
441 return (zfs_vdev_async_write_max_active
);
444 * linear interpolation:
445 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
446 * move right by min_bytes
447 * move up by min_writes
449 writes
= (dirty
- min_bytes
) *
450 (zfs_vdev_async_write_max_active
-
451 zfs_vdev_async_write_min_active
) /
452 (max_bytes
- min_bytes
) +
453 zfs_vdev_async_write_min_active
;
454 ASSERT3U(writes
, >=, zfs_vdev_async_write_min_active
);
455 ASSERT3U(writes
, <=, zfs_vdev_async_write_max_active
);
460 vdev_queue_class_max_active(spa_t
*spa
, zio_priority_t p
)
463 case ZIO_PRIORITY_SYNC_READ
:
464 return (zfs_vdev_sync_read_max_active
);
465 case ZIO_PRIORITY_SYNC_WRITE
:
466 return (zfs_vdev_sync_write_max_active
);
467 case ZIO_PRIORITY_ASYNC_READ
:
468 return (zfs_vdev_async_read_max_active
);
469 case ZIO_PRIORITY_ASYNC_WRITE
:
470 return (vdev_queue_max_async_writes(spa
));
471 case ZIO_PRIORITY_SCRUB
:
472 return (zfs_vdev_scrub_max_active
);
473 case ZIO_PRIORITY_REMOVAL
:
474 return (zfs_vdev_removal_max_active
);
475 case ZIO_PRIORITY_INITIALIZING
:
476 return (zfs_vdev_initializing_max_active
);
478 panic("invalid priority %u", p
);
484 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
485 * there is no eligible class.
487 static zio_priority_t
488 vdev_queue_class_to_issue(vdev_queue_t
*vq
)
490 spa_t
*spa
= vq
->vq_vdev
->vdev_spa
;
493 if (avl_numnodes(&vq
->vq_active_tree
) >= zfs_vdev_max_active
)
494 return (ZIO_PRIORITY_NUM_QUEUEABLE
);
496 /* find a queue that has not reached its minimum # outstanding i/os */
497 for (p
= 0; p
< ZIO_PRIORITY_NUM_QUEUEABLE
; p
++) {
498 if (avl_numnodes(vdev_queue_class_tree(vq
, p
)) > 0 &&
499 vq
->vq_class
[p
].vqc_active
<
500 vdev_queue_class_min_active(p
))
505 * If we haven't found a queue, look for one that hasn't reached its
506 * maximum # outstanding i/os.
508 for (p
= 0; p
< ZIO_PRIORITY_NUM_QUEUEABLE
; p
++) {
509 if (avl_numnodes(vdev_queue_class_tree(vq
, p
)) > 0 &&
510 vq
->vq_class
[p
].vqc_active
<
511 vdev_queue_class_max_active(spa
, p
))
515 /* No eligible queued i/os */
516 return (ZIO_PRIORITY_NUM_QUEUEABLE
);
520 * Compute the range spanned by two i/os, which is the endpoint of the last
521 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
522 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
523 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
525 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
526 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
529 vdev_queue_aggregate(vdev_queue_t
*vq
, zio_t
*zio
)
531 zio_t
*first
, *last
, *aio
, *dio
, *mandatory
, *nio
;
534 boolean_t stretch
= B_FALSE
;
535 avl_tree_t
*t
= vdev_queue_type_tree(vq
, zio
->io_type
);
536 enum zio_flag flags
= zio
->io_flags
& ZIO_FLAG_AGG_INHERIT
;
538 if (zio
->io_flags
& ZIO_FLAG_DONT_AGGREGATE
)
543 if (zio
->io_type
== ZIO_TYPE_READ
)
544 maxgap
= zfs_vdev_read_gap_limit
;
547 * We can aggregate I/Os that are sufficiently adjacent and of
548 * the same flavor, as expressed by the AGG_INHERIT flags.
549 * The latter requirement is necessary so that certain
550 * attributes of the I/O, such as whether it's a normal I/O
551 * or a scrub/resilver, can be preserved in the aggregate.
552 * We can include optional I/Os, but don't allow them
553 * to begin a range as they add no benefit in that situation.
557 * We keep track of the last non-optional I/O.
559 mandatory
= (first
->io_flags
& ZIO_FLAG_OPTIONAL
) ? NULL
: first
;
562 * Walk backwards through sufficiently contiguous I/Os
563 * recording the last non-optional I/O.
565 while ((dio
= AVL_PREV(t
, first
)) != NULL
&&
566 (dio
->io_flags
& ZIO_FLAG_AGG_INHERIT
) == flags
&&
567 IO_SPAN(dio
, last
) <= zfs_vdev_aggregation_limit
&&
568 IO_GAP(dio
, first
) <= maxgap
&&
569 dio
->io_type
== zio
->io_type
) {
571 if (mandatory
== NULL
&& !(first
->io_flags
& ZIO_FLAG_OPTIONAL
))
576 * Skip any initial optional I/Os.
578 while ((first
->io_flags
& ZIO_FLAG_OPTIONAL
) && first
!= last
) {
579 first
= AVL_NEXT(t
, first
);
580 ASSERT(first
!= NULL
);
584 * Walk forward through sufficiently contiguous I/Os.
585 * The aggregation limit does not apply to optional i/os, so that
586 * we can issue contiguous writes even if they are larger than the
589 while ((dio
= AVL_NEXT(t
, last
)) != NULL
&&
590 (dio
->io_flags
& ZIO_FLAG_AGG_INHERIT
) == flags
&&
591 (IO_SPAN(first
, dio
) <= zfs_vdev_aggregation_limit
||
592 (dio
->io_flags
& ZIO_FLAG_OPTIONAL
)) &&
593 IO_GAP(last
, dio
) <= maxgap
&&
594 dio
->io_type
== zio
->io_type
) {
596 if (!(last
->io_flags
& ZIO_FLAG_OPTIONAL
))
601 * Now that we've established the range of the I/O aggregation
602 * we must decide what to do with trailing optional I/Os.
603 * For reads, there's nothing to do. While we are unable to
604 * aggregate further, it's possible that a trailing optional
605 * I/O would allow the underlying device to aggregate with
606 * subsequent I/Os. We must therefore determine if the next
607 * non-optional I/O is close enough to make aggregation
610 if (zio
->io_type
== ZIO_TYPE_WRITE
&& mandatory
!= NULL
) {
612 while ((dio
= AVL_NEXT(t
, nio
)) != NULL
&&
613 IO_GAP(nio
, dio
) == 0 &&
614 IO_GAP(mandatory
, dio
) <= zfs_vdev_write_gap_limit
) {
616 if (!(nio
->io_flags
& ZIO_FLAG_OPTIONAL
)) {
625 * We are going to include an optional io in our aggregated
626 * span, thus closing the write gap. Only mandatory i/os can
627 * start aggregated spans, so make sure that the next i/o
628 * after our span is mandatory.
630 dio
= AVL_NEXT(t
, last
);
631 dio
->io_flags
&= ~ZIO_FLAG_OPTIONAL
;
633 /* do not include the optional i/o */
634 while (last
!= mandatory
&& last
!= first
) {
635 ASSERT(last
->io_flags
& ZIO_FLAG_OPTIONAL
);
636 last
= AVL_PREV(t
, last
);
637 ASSERT(last
!= NULL
);
644 size
= IO_SPAN(first
, last
);
645 ASSERT3U(size
, <=, SPA_MAXBLOCKSIZE
);
647 aio
= zio_vdev_delegated_io(first
->io_vd
, first
->io_offset
,
648 abd_alloc_for_io(size
, B_TRUE
), size
, first
->io_type
,
649 zio
->io_priority
, flags
| ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_QUEUE
,
650 vdev_queue_agg_io_done
, NULL
);
651 aio
->io_timestamp
= first
->io_timestamp
;
656 nio
= AVL_NEXT(t
, dio
);
657 ASSERT3U(dio
->io_type
, ==, aio
->io_type
);
659 if (dio
->io_flags
& ZIO_FLAG_NODATA
) {
660 ASSERT3U(dio
->io_type
, ==, ZIO_TYPE_WRITE
);
661 abd_zero_off(aio
->io_abd
,
662 dio
->io_offset
- aio
->io_offset
, dio
->io_size
);
663 } else if (dio
->io_type
== ZIO_TYPE_WRITE
) {
664 abd_copy_off(aio
->io_abd
, dio
->io_abd
,
665 dio
->io_offset
- aio
->io_offset
, 0, dio
->io_size
);
668 zio_add_child(dio
, aio
);
669 vdev_queue_io_remove(vq
, dio
);
670 zio_vdev_io_bypass(dio
);
672 } while (dio
!= last
);
678 vdev_queue_io_to_issue(vdev_queue_t
*vq
)
687 ASSERT(MUTEX_HELD(&vq
->vq_lock
));
689 p
= vdev_queue_class_to_issue(vq
);
691 if (p
== ZIO_PRIORITY_NUM_QUEUEABLE
) {
692 /* No eligible queued i/os */
697 * For LBA-ordered queues (async / scrub / initializing), issue the
698 * i/o which follows the most recently issued i/o in LBA (offset) order.
700 * For FIFO queues (sync), issue the i/o with the lowest timestamp.
702 tree
= vdev_queue_class_tree(vq
, p
);
703 search
.io_timestamp
= 0;
704 search
.io_offset
= vq
->vq_last_offset
+ 1;
705 VERIFY3P(avl_find(tree
, &search
, &idx
), ==, NULL
);
706 zio
= avl_nearest(tree
, idx
, AVL_AFTER
);
708 zio
= avl_first(tree
);
709 ASSERT3U(zio
->io_priority
, ==, p
);
711 aio
= vdev_queue_aggregate(vq
, zio
);
715 vdev_queue_io_remove(vq
, zio
);
718 * If the I/O is or was optional and therefore has no data, we need to
719 * simply discard it. We need to drop the vdev queue's lock to avoid a
720 * deadlock that we could encounter since this I/O will complete
723 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
724 mutex_exit(&vq
->vq_lock
);
725 zio_vdev_io_bypass(zio
);
727 mutex_enter(&vq
->vq_lock
);
731 vdev_queue_pending_add(vq
, zio
);
732 vq
->vq_last_offset
= zio
->io_offset
;
738 vdev_queue_io(zio_t
*zio
)
740 vdev_queue_t
*vq
= &zio
->io_vd
->vdev_queue
;
743 if (zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)
747 * Children i/os inherent their parent's priority, which might
748 * not match the child's i/o type. Fix it up here.
750 if (zio
->io_type
== ZIO_TYPE_READ
) {
751 if (zio
->io_priority
!= ZIO_PRIORITY_SYNC_READ
&&
752 zio
->io_priority
!= ZIO_PRIORITY_ASYNC_READ
&&
753 zio
->io_priority
!= ZIO_PRIORITY_SCRUB
&&
754 zio
->io_priority
!= ZIO_PRIORITY_REMOVAL
&&
755 zio
->io_priority
!= ZIO_PRIORITY_INITIALIZING
)
756 zio
->io_priority
= ZIO_PRIORITY_ASYNC_READ
;
758 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
759 if (zio
->io_priority
!= ZIO_PRIORITY_SYNC_WRITE
&&
760 zio
->io_priority
!= ZIO_PRIORITY_ASYNC_WRITE
&&
761 zio
->io_priority
!= ZIO_PRIORITY_REMOVAL
&&
762 zio
->io_priority
!= ZIO_PRIORITY_INITIALIZING
)
763 zio
->io_priority
= ZIO_PRIORITY_ASYNC_WRITE
;
766 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_QUEUE
;
768 mutex_enter(&vq
->vq_lock
);
769 zio
->io_timestamp
= gethrtime();
770 vdev_queue_io_add(vq
, zio
);
771 nio
= vdev_queue_io_to_issue(vq
);
772 mutex_exit(&vq
->vq_lock
);
777 if (nio
->io_done
== vdev_queue_agg_io_done
) {
786 vdev_queue_io_done(zio_t
*zio
)
788 vdev_queue_t
*vq
= &zio
->io_vd
->vdev_queue
;
791 mutex_enter(&vq
->vq_lock
);
793 vdev_queue_pending_remove(vq
, zio
);
795 vq
->vq_io_complete_ts
= gethrtime();
797 while ((nio
= vdev_queue_io_to_issue(vq
)) != NULL
) {
798 mutex_exit(&vq
->vq_lock
);
799 if (nio
->io_done
== vdev_queue_agg_io_done
) {
802 zio_vdev_io_reissue(nio
);
805 mutex_enter(&vq
->vq_lock
);
808 mutex_exit(&vq
->vq_lock
);