1/*
2 * CDDL HEADER START
3 *
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.
7 *
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.
12 *
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]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23 * Use is subject to license terms.
24 */
25
26/*
27 * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
28 * Copyright (c) 2014 Integros [integros.com]
29 * Copyright 2019 Joyent, Inc.
30 */
31
32#include <sys/zfs_context.h>
33#include <sys/vdev_impl.h>
34#include <sys/spa_impl.h>
35#include <sys/zio.h>
36#include <sys/avl.h>
37#include <sys/dsl_pool.h>
38#include <sys/metaslab_impl.h>
39#include <sys/abd.h>
40
41/*
42 * ZFS I/O Scheduler
43 * ---------------
44 *
45 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios.  The
46 * I/O scheduler determines when and in what order those operations are
47 * issued.  The I/O scheduler divides operations into five I/O classes
48 * prioritized in the following order: sync read, sync write, async read,
49 * async write, and scrub/resilver.  Each queue defines the minimum and
50 * maximum number of concurrent operations that may be issued to the device.
51 * In addition, the device has an aggregate maximum. Note that the sum of the
52 * per-queue minimums must not exceed the aggregate maximum, and if the
53 * aggregate maximum is equal to or greater than the sum of the per-queue
54 * maximums, the per-queue minimum has no effect.
55 *
56 * For many physical devices, throughput increases with the number of
57 * concurrent operations, but latency typically suffers. Further, physical
58 * devices typically have a limit at which more concurrent operations have no
59 * effect on throughput or can actually cause it to decrease.
60 *
61 * The scheduler selects the next operation to issue by first looking for an
62 * I/O class whose minimum has not been satisfied. Once all are satisfied and
63 * the aggregate maximum has not been hit, the scheduler looks for classes
64 * whose maximum has not been satisfied. Iteration through the I/O classes is
65 * done in the order specified above. No further operations are issued if the
66 * aggregate maximum number of concurrent operations has been hit or if there
67 * are no operations queued for an I/O class that has not hit its maximum.
68 * Every time an i/o is queued or an operation completes, the I/O scheduler
69 * looks for new operations to issue.
70 *
71 * All I/O classes have a fixed maximum number of outstanding operations
72 * except for the async write class. Asynchronous writes represent the data
73 * that is committed to stable storage during the syncing stage for
74 * transaction groups (see txg.c). Transaction groups enter the syncing state
75 * periodically so the number of queued async writes will quickly burst up and
76 * then bleed down to zero. Rather than servicing them as quickly as possible,
77 * the I/O scheduler changes the maximum number of active async write i/os
78 * according to the amount of dirty data in the pool (see dsl_pool.c). Since
79 * both throughput and latency typically increase with the number of
80 * concurrent operations issued to physical devices, reducing the burstiness
81 * in the number of concurrent operations also stabilizes the response time of
82 * operations from other -- and in particular synchronous -- queues. In broad
83 * strokes, the I/O scheduler will issue more concurrent operations from the
84 * async write queue as there's more dirty data in the pool.
85 *
86 * Async Writes
87 *
88 * The number of concurrent operations issued for the async write I/O class
89 * follows a piece-wise linear function defined by a few adjustable points.
90 *
91 *        |                   o---------| <-- zfs_vdev_async_write_max_active
92 *   ^    |                  /^         |
93 *   |    |                 / |         |
94 * active |                /  |         |
95 *  I/O   |               /   |         |
96 * count  |              /    |         |
97 *        |             /     |         |
98 *        |------------o      |         | <-- zfs_vdev_async_write_min_active
99 *       0|____________^______|_________|
100 *        0%           |      |       100% of zfs_dirty_data_max
101 *                     |      |
102 *                     |      `-- zfs_vdev_async_write_active_max_dirty_percent
103 *                     `--------- zfs_vdev_async_write_active_min_dirty_percent
104 *
105 * Until the amount of dirty data exceeds a minimum percentage of the dirty
106 * data allowed in the pool, the I/O scheduler will limit the number of
107 * concurrent operations to the minimum. As that threshold is crossed, the
108 * number of concurrent operations issued increases linearly to the maximum at
109 * the specified maximum percentage of the dirty data allowed in the pool.
110 *
111 * Ideally, the amount of dirty data on a busy pool will stay in the sloped
112 * part of the function between zfs_vdev_async_write_active_min_dirty_percent
113 * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the
114 * maximum percentage, this indicates that the rate of incoming data is
115 * greater than the rate that the backend storage can handle. In this case, we
116 * must further throttle incoming writes (see dmu_tx_delay() for details).
117 */
118
119/*
120 * The maximum number of i/os active to each device.  Ideally, this will be >=
121 * the sum of each queue's max_active.  It must be at least the sum of each
122 * queue's min_active.
123 */
124uint32_t zfs_vdev_max_active = 1000;
125
126/*
127 * Per-queue limits on the number of i/os active to each device.  If the
128 * sum of the queue's max_active is < zfs_vdev_max_active, then the
129 * min_active comes into play.  We will send min_active from each queue,
130 * and then select from queues in the order defined by zio_priority_t.
131 *
132 * In general, smaller max_active's will lead to lower latency of synchronous
133 * operations.  Larger max_active's may lead to higher overall throughput,
134 * depending on underlying storage.
135 *
136 * The ratio of the queues' max_actives determines the balance of performance
137 * between reads, writes, and scrubs.  E.g., increasing
138 * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete
139 * more quickly, but reads and writes to have higher latency and lower
140 * throughput.
141 */
142uint32_t zfs_vdev_sync_read_min_active = 10;
143uint32_t zfs_vdev_sync_read_max_active = 10;
144uint32_t zfs_vdev_sync_write_min_active = 10;
145uint32_t zfs_vdev_sync_write_max_active = 10;
146uint32_t zfs_vdev_async_read_min_active = 1;
147uint32_t zfs_vdev_async_read_max_active = 3;
148uint32_t zfs_vdev_async_write_min_active = 1;
149uint32_t zfs_vdev_async_write_max_active = 10;
150uint32_t zfs_vdev_scrub_min_active = 1;
151uint32_t zfs_vdev_scrub_max_active = 2;
152uint32_t zfs_vdev_removal_min_active = 1;
153uint32_t zfs_vdev_removal_max_active = 2;
154uint32_t zfs_vdev_initializing_min_active = 1;
155uint32_t zfs_vdev_initializing_max_active = 1;
156uint32_t zfs_vdev_trim_min_active = 1;
157uint32_t zfs_vdev_trim_max_active = 2;
158
159/*
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.
165 */
166int zfs_vdev_async_write_active_min_dirty_percent = 30;
167int zfs_vdev_async_write_active_max_dirty_percent = 60;
168
169/*
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.
174 */
175int zfs_vdev_aggregation_limit = 1 << 20;
176int zfs_vdev_read_gap_limit = 32 << 10;
177int zfs_vdev_write_gap_limit = 4 << 10;
178
179/*
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.
188 */
189#ifdef _KERNEL
190int zfs_vdev_queue_depth_pct = 1000;
191#else
192int zfs_vdev_queue_depth_pct = 300;
193#endif
194
195/*
196 * When performing allocations for a given metaslab, we want to make sure that
197 * there are enough IOs to aggregate together to improve throughput. We want to
198 * ensure that there are at least 128k worth of IOs that can be aggregated, and
199 * we assume that the average allocation size is 4k, so we need the queue depth
200 * to be 32 per allocator to get good aggregation of sequential writes.
201 */
202int zfs_vdev_def_queue_depth = 32;
203
204/*
205 * Allow TRIM I/Os to be aggregated.  This should normally not be needed since
206 * TRIM I/O for extents up to zfs_trim_extent_bytes_max (128M) can be submitted
207 * by the TRIM code in zfs_trim.c.
208 */
209int zfs_vdev_aggregate_trim = 0;
210
211int
212vdev_queue_offset_compare(const void *x1, const void *x2)
213{
214	const zio_t *z1 = (const zio_t *)x1;
215	const zio_t *z2 = (const zio_t *)x2;
216
217	int cmp = AVL_CMP(z1->io_offset, z2->io_offset);
218
219	if (likely(cmp))
220		return (cmp);
221
222	return (AVL_PCMP(z1, z2));
223}
224
225static inline avl_tree_t *
226vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p)
227{
228	return (&vq->vq_class[p].vqc_queued_tree);
229}
230
231static inline avl_tree_t *
232vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t)
233{
234	ASSERT(t == ZIO_TYPE_READ || t == ZIO_TYPE_WRITE || t == ZIO_TYPE_TRIM);
235	if (t == ZIO_TYPE_READ)
236		return (&vq->vq_read_offset_tree);
237	else if (t == ZIO_TYPE_WRITE)
238		return (&vq->vq_write_offset_tree);
239	else
240		return (&vq->vq_trim_offset_tree);
241}
242
243int
244vdev_queue_timestamp_compare(const void *x1, const void *x2)
245{
246	const zio_t *z1 = (const zio_t *)x1;
247	const zio_t *z2 = (const zio_t *)x2;
248
249	int cmp = AVL_CMP(z1->io_timestamp, z2->io_timestamp);
250
251	if (likely(cmp))
252		return (cmp);
253
254	return (AVL_PCMP(z1, z2));
255}
256
257void
258vdev_queue_init(vdev_t *vd)
259{
260	vdev_queue_t *vq = &vd->vdev_queue;
261
262	mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
263	vq->vq_vdev = vd;
264
265	avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
266	    sizeof (zio_t), offsetof(struct zio, io_queue_node));
267	avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ),
268	    vdev_queue_offset_compare, sizeof (zio_t),
269	    offsetof(struct zio, io_offset_node));
270	avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE),
271	    vdev_queue_offset_compare, sizeof (zio_t),
272	    offsetof(struct zio, io_offset_node));
273	avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_TRIM),
274	    vdev_queue_offset_compare, sizeof (zio_t),
275	    offsetof(struct zio, io_offset_node));
276
277	for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
278		int (*compfn) (const void *, const void *);
279
280		/*
281		 * The synchronous/trim i/o queues are dispatched in FIFO rather
282		 * than LBA order. This provides more consistent latency for
283		 * these i/os.
284		 */
285		if (p == ZIO_PRIORITY_SYNC_READ ||
286		    p == ZIO_PRIORITY_SYNC_WRITE ||
287		    p == ZIO_PRIORITY_TRIM) {
288			compfn = vdev_queue_timestamp_compare;
289		} else {
290			compfn = vdev_queue_offset_compare;
291		}
292
293		avl_create(vdev_queue_class_tree(vq, p), compfn,
294		    sizeof (zio_t), offsetof(struct zio, io_queue_node));
295	}
296
297	vq->vq_last_offset = 0;
298}
299
300void
301vdev_queue_fini(vdev_t *vd)
302{
303	vdev_queue_t *vq = &vd->vdev_queue;
304
305	for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
306		avl_destroy(vdev_queue_class_tree(vq, p));
307	avl_destroy(&vq->vq_active_tree);
308	avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
309	avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE));
310	avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_TRIM));
311
312	mutex_destroy(&vq->vq_lock);
313}
314
315static void
316vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
317{
318	spa_t *spa = zio->io_spa;
319
320	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
321	avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
322	avl_add(vdev_queue_type_tree(vq, zio->io_type), zio);
323
324	mutex_enter(&spa->spa_iokstat_lock);
325	spa->spa_queue_stats[zio->io_priority].spa_queued++;
326	if (spa->spa_iokstat != NULL)
327		kstat_waitq_enter(spa->spa_iokstat->ks_data);
328	mutex_exit(&spa->spa_iokstat_lock);
329}
330
331static void
332vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
333{
334	spa_t *spa = zio->io_spa;
335
336	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
337	avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
338	avl_remove(vdev_queue_type_tree(vq, zio->io_type), zio);
339
340	mutex_enter(&spa->spa_iokstat_lock);
341	ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
342	spa->spa_queue_stats[zio->io_priority].spa_queued--;
343	if (spa->spa_iokstat != NULL)
344		kstat_waitq_exit(spa->spa_iokstat->ks_data);
345	mutex_exit(&spa->spa_iokstat_lock);
346}
347
348static void
349vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
350{
351	spa_t *spa = zio->io_spa;
352	ASSERT(MUTEX_HELD(&vq->vq_lock));
353	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
354	vq->vq_class[zio->io_priority].vqc_active++;
355	avl_add(&vq->vq_active_tree, zio);
356
357	mutex_enter(&spa->spa_iokstat_lock);
358	spa->spa_queue_stats[zio->io_priority].spa_active++;
359	if (spa->spa_iokstat != NULL)
360		kstat_runq_enter(spa->spa_iokstat->ks_data);
361	mutex_exit(&spa->spa_iokstat_lock);
362}
363
364static void
365vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
366{
367	spa_t *spa = zio->io_spa;
368	ASSERT(MUTEX_HELD(&vq->vq_lock));
369	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
370	vq->vq_class[zio->io_priority].vqc_active--;
371	avl_remove(&vq->vq_active_tree, zio);
372
373	mutex_enter(&spa->spa_iokstat_lock);
374	ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
375	spa->spa_queue_stats[zio->io_priority].spa_active--;
376	if (spa->spa_iokstat != NULL) {
377		kstat_io_t *ksio = spa->spa_iokstat->ks_data;
378
379		kstat_runq_exit(spa->spa_iokstat->ks_data);
380		if (zio->io_type == ZIO_TYPE_READ) {
381			ksio->reads++;
382			ksio->nread += zio->io_size;
383		} else if (zio->io_type == ZIO_TYPE_WRITE) {
384			ksio->writes++;
385			ksio->nwritten += zio->io_size;
386		}
387	}
388	mutex_exit(&spa->spa_iokstat_lock);
389}
390
391static void
392vdev_queue_agg_io_done(zio_t *aio)
393{
394	if (aio->io_type == ZIO_TYPE_READ) {
395		zio_t *pio;
396		zio_link_t *zl = NULL;
397		while ((pio = zio_walk_parents(aio, &zl)) != NULL) {
398			abd_copy_off(pio->io_abd, aio->io_abd,
399			    0, pio->io_offset - aio->io_offset, pio->io_size);
400		}
401	}
402
403	abd_free(aio->io_abd);
404}
405
406static int
407vdev_queue_class_min_active(zio_priority_t p)
408{
409	switch (p) {
410	case ZIO_PRIORITY_SYNC_READ:
411		return (zfs_vdev_sync_read_min_active);
412	case ZIO_PRIORITY_SYNC_WRITE:
413		return (zfs_vdev_sync_write_min_active);
414	case ZIO_PRIORITY_ASYNC_READ:
415		return (zfs_vdev_async_read_min_active);
416	case ZIO_PRIORITY_ASYNC_WRITE:
417		return (zfs_vdev_async_write_min_active);
418	case ZIO_PRIORITY_SCRUB:
419		return (zfs_vdev_scrub_min_active);
420	case ZIO_PRIORITY_REMOVAL:
421		return (zfs_vdev_removal_min_active);
422	case ZIO_PRIORITY_INITIALIZING:
423		return (zfs_vdev_initializing_min_active);
424	case ZIO_PRIORITY_TRIM:
425		return (zfs_vdev_trim_min_active);
426	default:
427		panic("invalid priority %u", p);
428	}
429}
430
431static int
432vdev_queue_max_async_writes(spa_t *spa)
433{
434	int writes;
435	uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total;
436	uint64_t min_bytes = zfs_dirty_data_max *
437	    zfs_vdev_async_write_active_min_dirty_percent / 100;
438	uint64_t max_bytes = zfs_dirty_data_max *
439	    zfs_vdev_async_write_active_max_dirty_percent / 100;
440
441	/*
442	 * Sync tasks correspond to interactive user actions. To reduce the
443	 * execution time of those actions we push data out as fast as possible.
444	 */
445	if (spa_has_pending_synctask(spa)) {
446		return (zfs_vdev_async_write_max_active);
447	}
448
449	if (dirty < min_bytes)
450		return (zfs_vdev_async_write_min_active);
451	if (dirty > max_bytes)
452		return (zfs_vdev_async_write_max_active);
453
454	/*
455	 * linear interpolation:
456	 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
457	 * move right by min_bytes
458	 * move up by min_writes
459	 */
460	writes = (dirty - min_bytes) *
461	    (zfs_vdev_async_write_max_active -
462	    zfs_vdev_async_write_min_active) /
463	    (max_bytes - min_bytes) +
464	    zfs_vdev_async_write_min_active;
465	ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
466	ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
467	return (writes);
468}
469
470static int
471vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
472{
473	switch (p) {
474	case ZIO_PRIORITY_SYNC_READ:
475		return (zfs_vdev_sync_read_max_active);
476	case ZIO_PRIORITY_SYNC_WRITE:
477		return (zfs_vdev_sync_write_max_active);
478	case ZIO_PRIORITY_ASYNC_READ:
479		return (zfs_vdev_async_read_max_active);
480	case ZIO_PRIORITY_ASYNC_WRITE:
481		return (vdev_queue_max_async_writes(spa));
482	case ZIO_PRIORITY_SCRUB:
483		return (zfs_vdev_scrub_max_active);
484	case ZIO_PRIORITY_REMOVAL:
485		return (zfs_vdev_removal_max_active);
486	case ZIO_PRIORITY_INITIALIZING:
487		return (zfs_vdev_initializing_max_active);
488	case ZIO_PRIORITY_TRIM:
489		return (zfs_vdev_trim_max_active);
490	default:
491		panic("invalid priority %u", p);
492	}
493}
494
495/*
496 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
497 * there is no eligible class.
498 */
499static zio_priority_t
500vdev_queue_class_to_issue(vdev_queue_t *vq)
501{
502	spa_t *spa = vq->vq_vdev->vdev_spa;
503	zio_priority_t p;
504
505	if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
506		return (ZIO_PRIORITY_NUM_QUEUEABLE);
507
508	/* find a queue that has not reached its minimum # outstanding i/os */
509	for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
510		if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
511		    vq->vq_class[p].vqc_active <
512		    vdev_queue_class_min_active(p))
513			return (p);
514	}
515
516	/*
517	 * If we haven't found a queue, look for one that hasn't reached its
518	 * maximum # outstanding i/os.
519	 */
520	for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
521		if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
522		    vq->vq_class[p].vqc_active <
523		    vdev_queue_class_max_active(spa, p))
524			return (p);
525	}
526
527	/* No eligible queued i/os */
528	return (ZIO_PRIORITY_NUM_QUEUEABLE);
529}
530
531/*
532 * Compute the range spanned by two i/os, which is the endpoint of the last
533 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
534 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
535 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
536 */
537#define	IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
538#define	IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
539
540static zio_t *
541vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
542{
543	zio_t *first, *last, *aio, *dio, *mandatory, *nio;
544	zio_link_t *zl = NULL;
545	uint64_t maxgap = 0;
546	uint64_t size;
547	boolean_t stretch = B_FALSE;
548	avl_tree_t *t = vdev_queue_type_tree(vq, zio->io_type);
549	enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
550
551	if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
552		return (NULL);
553
554	/*
555	 * While TRIM commands could be aggregated based on offset this
556	 * behavior is disabled until it's determined to be beneficial.
557	 */
558	if (zio->io_type == ZIO_TYPE_TRIM && !zfs_vdev_aggregate_trim)
559		return (NULL);
560
561	first = last = zio;
562
563	if (zio->io_type == ZIO_TYPE_READ)
564		maxgap = zfs_vdev_read_gap_limit;
565
566	/*
567	 * We can aggregate I/Os that are sufficiently adjacent and of
568	 * the same flavor, as expressed by the AGG_INHERIT flags.
569	 * The latter requirement is necessary so that certain
570	 * attributes of the I/O, such as whether it's a normal I/O
571	 * or a scrub/resilver, can be preserved in the aggregate.
572	 * We can include optional I/Os, but don't allow them
573	 * to begin a range as they add no benefit in that situation.
574	 */
575
576	/*
577	 * We keep track of the last non-optional I/O.
578	 */
579	mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
580
581	/*
582	 * Walk backwards through sufficiently contiguous I/Os
583	 * recording the last non-optional I/O.
584	 */
585	while ((dio = AVL_PREV(t, first)) != NULL &&
586	    (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
587	    IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
588	    IO_GAP(dio, first) <= maxgap &&
589	    dio->io_type == zio->io_type) {
590		first = dio;
591		if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
592			mandatory = first;
593	}
594
595	/*
596	 * Skip any initial optional I/Os.
597	 */
598	while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
599		first = AVL_NEXT(t, first);
600		ASSERT(first != NULL);
601	}
602
603	/*
604	 * Walk forward through sufficiently contiguous I/Os.
605	 * The aggregation limit does not apply to optional i/os, so that
606	 * we can issue contiguous writes even if they are larger than the
607	 * aggregation limit.
608	 */
609	while ((dio = AVL_NEXT(t, last)) != NULL &&
610	    (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
611	    (IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit ||
612	    (dio->io_flags & ZIO_FLAG_OPTIONAL)) &&
613	    IO_GAP(last, dio) <= maxgap &&
614	    dio->io_type == zio->io_type) {
615		last = dio;
616		if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
617			mandatory = last;
618	}
619
620	/*
621	 * Now that we've established the range of the I/O aggregation
622	 * we must decide what to do with trailing optional I/Os.
623	 * For reads, there's nothing to do. While we are unable to
624	 * aggregate further, it's possible that a trailing optional
625	 * I/O would allow the underlying device to aggregate with
626	 * subsequent I/Os. We must therefore determine if the next
627	 * non-optional I/O is close enough to make aggregation
628	 * worthwhile.
629	 */
630	if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
631		zio_t *nio = last;
632		while ((dio = AVL_NEXT(t, nio)) != NULL &&
633		    IO_GAP(nio, dio) == 0 &&
634		    IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
635			nio = dio;
636			if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
637				stretch = B_TRUE;
638				break;
639			}
640		}
641	}
642
643	if (stretch) {
644		/*
645		 * We are going to include an optional io in our aggregated
646		 * span, thus closing the write gap.  Only mandatory i/os can
647		 * start aggregated spans, so make sure that the next i/o
648		 * after our span is mandatory.
649		 */
650		dio = AVL_NEXT(t, last);
651		dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
652	} else {
653		/* do not include the optional i/o */
654		while (last != mandatory && last != first) {
655			ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
656			last = AVL_PREV(t, last);
657			ASSERT(last != NULL);
658		}
659	}
660
661	if (first == last)
662		return (NULL);
663
664	size = IO_SPAN(first, last);
665	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
666
667	aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
668	    abd_alloc_for_io(size, B_TRUE), size, first->io_type,
669	    zio->io_priority, flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
670	    vdev_queue_agg_io_done, NULL);
671	aio->io_timestamp = first->io_timestamp;
672
673	nio = first;
674	do {
675		dio = nio;
676		nio = AVL_NEXT(t, dio);
677		ASSERT3U(dio->io_type, ==, aio->io_type);
678
679		if (dio->io_flags & ZIO_FLAG_NODATA) {
680			ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
681			abd_zero_off(aio->io_abd,
682			    dio->io_offset - aio->io_offset, dio->io_size);
683		} else if (dio->io_type == ZIO_TYPE_WRITE) {
684			abd_copy_off(aio->io_abd, dio->io_abd,
685			    dio->io_offset - aio->io_offset, 0, dio->io_size);
686		}
687
688		zio_add_child(dio, aio);
689		vdev_queue_io_remove(vq, dio);
690	} while (dio != last);
691
692	/*
693	 * We need to drop the vdev queue's lock to avoid a deadlock that we
694	 * could encounter since this I/O will complete immediately.
695	 */
696	mutex_exit(&vq->vq_lock);
697	while ((dio = zio_walk_parents(aio, &zl)) != NULL) {
698		zio_vdev_io_bypass(dio);
699		zio_execute(dio);
700	}
701	mutex_enter(&vq->vq_lock);
702
703	return (aio);
704}
705
706static zio_t *
707vdev_queue_io_to_issue(vdev_queue_t *vq)
708{
709	zio_t *zio, *aio;
710	zio_priority_t p;
711	avl_index_t idx;
712	avl_tree_t *tree;
713	zio_t search;
714
715again:
716	ASSERT(MUTEX_HELD(&vq->vq_lock));
717
718	p = vdev_queue_class_to_issue(vq);
719
720	if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
721		/* No eligible queued i/os */
722		return (NULL);
723	}
724
725	/*
726	 * For LBA-ordered queues (async / scrub / initializing), issue the
727	 * i/o which follows the most recently issued i/o in LBA (offset) order.
728	 *
729	 * For FIFO queues (sync/trim), issue the i/o with the lowest timestamp.
730	 */
731	tree = vdev_queue_class_tree(vq, p);
732	search.io_timestamp = 0;
733	search.io_offset = vq->vq_last_offset - 1;
734	VERIFY3P(avl_find(tree, &search, &idx), ==, NULL);
735	zio = avl_nearest(tree, idx, AVL_AFTER);
736	if (zio == NULL)
737		zio = avl_first(tree);
738	ASSERT3U(zio->io_priority, ==, p);
739
740	aio = vdev_queue_aggregate(vq, zio);
741	if (aio != NULL)
742		zio = aio;
743	else
744		vdev_queue_io_remove(vq, zio);
745
746	/*
747	 * If the I/O is or was optional and therefore has no data, we need to
748	 * simply discard it. We need to drop the vdev queue's lock to avoid a
749	 * deadlock that we could encounter since this I/O will complete
750	 * immediately.
751	 */
752	if (zio->io_flags & ZIO_FLAG_NODATA) {
753		mutex_exit(&vq->vq_lock);
754		zio_vdev_io_bypass(zio);
755		zio_execute(zio);
756		mutex_enter(&vq->vq_lock);
757		goto again;
758	}
759
760	vdev_queue_pending_add(vq, zio);
761	vq->vq_last_offset = zio->io_offset + zio->io_size;
762
763	return (zio);
764}
765
766zio_t *
767vdev_queue_io(zio_t *zio)
768{
769	vdev_queue_t *vq = &zio->io_vd->vdev_queue;
770	zio_t *nio;
771
772	if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
773		return (zio);
774
775	/*
776	 * Children i/os inherent their parent's priority, which might
777	 * not match the child's i/o type.  Fix it up here.
778	 */
779	if (zio->io_type == ZIO_TYPE_READ) {
780		ASSERT(zio->io_priority != ZIO_PRIORITY_TRIM);
781
782		if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
783		    zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
784		    zio->io_priority != ZIO_PRIORITY_SCRUB &&
785		    zio->io_priority != ZIO_PRIORITY_REMOVAL &&
786		    zio->io_priority != ZIO_PRIORITY_INITIALIZING) {
787			zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
788		}
789	} else if (zio->io_type == ZIO_TYPE_WRITE) {
790		ASSERT(zio->io_priority != ZIO_PRIORITY_TRIM);
791
792		if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
793		    zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE &&
794		    zio->io_priority != ZIO_PRIORITY_REMOVAL &&
795		    zio->io_priority != ZIO_PRIORITY_INITIALIZING) {
796			zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
797		}
798	} else {
799		ASSERT(zio->io_type == ZIO_TYPE_TRIM);
800		ASSERT(zio->io_priority == ZIO_PRIORITY_TRIM);
801	}
802
803	zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
804
805	mutex_enter(&vq->vq_lock);
806	zio->io_timestamp = gethrtime();
807	vdev_queue_io_add(vq, zio);
808	nio = vdev_queue_io_to_issue(vq);
809	mutex_exit(&vq->vq_lock);
810
811	if (nio == NULL)
812		return (NULL);
813
814	if (nio->io_done == vdev_queue_agg_io_done) {
815		zio_nowait(nio);
816		return (NULL);
817	}
818
819	return (nio);
820}
821
822void
823vdev_queue_io_done(zio_t *zio)
824{
825	vdev_queue_t *vq = &zio->io_vd->vdev_queue;
826	zio_t *nio;
827
828	mutex_enter(&vq->vq_lock);
829
830	vdev_queue_pending_remove(vq, zio);
831
832	vq->vq_io_complete_ts = gethrtime();
833
834	while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
835		mutex_exit(&vq->vq_lock);
836		if (nio->io_done == vdev_queue_agg_io_done) {
837			zio_nowait(nio);
838		} else {
839			zio_vdev_io_reissue(nio);
840			zio_execute(nio);
841		}
842		mutex_enter(&vq->vq_lock);
843	}
844
845	mutex_exit(&vq->vq_lock);
846}
847
848void
849vdev_queue_change_io_priority(zio_t *zio, zio_priority_t priority)
850{
851	vdev_queue_t *vq = &zio->io_vd->vdev_queue;
852	avl_tree_t *tree;
853
854	/*
855	 * ZIO_PRIORITY_NOW is used by the vdev cache code and the aggregate zio
856	 * code to issue IOs without adding them to the vdev queue. In this
857	 * case, the zio is already going to be issued as quickly as possible
858	 * and so it doesn't need any reprioitization to help.
859	 */
860	if (zio->io_priority == ZIO_PRIORITY_NOW)
861		return;
862
863	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
864	ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
865
866	if (zio->io_type == ZIO_TYPE_READ) {
867		if (priority != ZIO_PRIORITY_SYNC_READ &&
868		    priority != ZIO_PRIORITY_ASYNC_READ &&
869		    priority != ZIO_PRIORITY_SCRUB)
870			priority = ZIO_PRIORITY_ASYNC_READ;
871	} else {
872		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
873		if (priority != ZIO_PRIORITY_SYNC_WRITE &&
874		    priority != ZIO_PRIORITY_ASYNC_WRITE)
875			priority = ZIO_PRIORITY_ASYNC_WRITE;
876	}
877
878	mutex_enter(&vq->vq_lock);
879
880	/*
881	 * If the zio is in none of the queues we can simply change
882	 * the priority. If the zio is waiting to be submitted we must
883	 * remove it from the queue and re-insert it with the new priority.
884	 * Otherwise, the zio is currently active and we cannot change its
885	 * priority.
886	 */
887	tree = vdev_queue_class_tree(vq, zio->io_priority);
888	if (avl_find(tree, zio, NULL) == zio) {
889		spa_t *spa = zio->io_spa;
890		zio_priority_t oldpri = zio->io_priority;
891
892		avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
893		zio->io_priority = priority;
894		avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
895
896		mutex_enter(&spa->spa_iokstat_lock);
897		ASSERT3U(spa->spa_queue_stats[oldpri].spa_queued, >, 0);
898		spa->spa_queue_stats[oldpri].spa_queued--;
899		spa->spa_queue_stats[zio->io_priority].spa_queued++;
900		mutex_exit(&spa->spa_iokstat_lock);
901	} else if (avl_find(&vq->vq_active_tree, zio, NULL) != zio) {
902		zio->io_priority = priority;
903	}
904
905	mutex_exit(&vq->vq_lock);
906}
907
908/*
909 * As these two methods are only used for load calculations we're not
910 * concerned if we get an incorrect value on 32bit platforms due to lack of
911 * vq_lock mutex use here, instead we prefer to keep it lock free for
912 * performance.
913 */
914int
915vdev_queue_length(vdev_t *vd)
916{
917	return (avl_numnodes(&vd->vdev_queue.vq_active_tree));
918}
919
920uint64_t
921vdev_queue_last_offset(vdev_t *vd)
922{
923	return (vd->vdev_queue.vq_last_offset);
924}
925