xref: /illumos-gate/usr/src/uts/common/fs/zfs/vdev_queue.c (revision f78cdc34af236a6199dd9e21376f4a46348c0d56)
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  */
30 
31 #include <sys/zfs_context.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/spa_impl.h>
34 #include <sys/zio.h>
35 #include <sys/avl.h>
36 #include <sys/dsl_pool.h>
37 #include <sys/metaslab_impl.h>
38 #include <sys/abd.h>
39 
40 /*
41  * ZFS I/O Scheduler
42  * ---------------
43  *
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.
54  *
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.
59  *
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.
69  *
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.
84  *
85  * Async Writes
86  *
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.
89  *
90  *        |                   o---------| <-- zfs_vdev_async_write_max_active
91  *   ^    |                  /^         |
92  *   |    |                 / |         |
93  * active |                /  |         |
94  *  I/O   |               /   |         |
95  * count  |              /    |         |
96  *        |             /     |         |
97  *        |------------o      |         | <-- zfs_vdev_async_write_min_active
98  *       0|____________^______|_________|
99  *        0%           |      |       100% of zfs_dirty_data_max
100  *                     |      |
101  *                     |      `-- zfs_vdev_async_write_active_max_dirty_percent
102  *                     `--------- zfs_vdev_async_write_active_min_dirty_percent
103  *
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.
109  *
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).
116  */
117 
118 /*
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.
122  */
123 uint32_t zfs_vdev_max_active = 1000;
124 
125 /*
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.
130  *
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.
134  *
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
139  * throughput.
140  */
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 
154 /*
155  * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent
156  * dirty data, use zfs_vdev_async_write_min_active.  When it has more than
157  * zfs_vdev_async_write_active_max_dirty_percent, use
158  * zfs_vdev_async_write_max_active. The value is linearly interpolated
159  * between min and max.
160  */
161 int zfs_vdev_async_write_active_min_dirty_percent = 30;
162 int zfs_vdev_async_write_active_max_dirty_percent = 60;
163 
164 /*
165  * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
166  * For read I/Os, we also aggregate across small adjacency gaps; for writes
167  * we include spans of optional I/Os to aid aggregation at the disk even when
168  * they aren't able to help us aggregate at this level.
169  */
170 int zfs_vdev_aggregation_limit = SPA_OLD_MAXBLOCKSIZE;
171 int zfs_vdev_read_gap_limit = 32 << 10;
172 int zfs_vdev_write_gap_limit = 4 << 10;
173 
174 /*
175  * Define the queue depth percentage for each top-level. This percentage is
176  * used in conjunction with zfs_vdev_async_max_active to determine how many
177  * allocations a specific top-level vdev should handle. Once the queue depth
178  * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100
179  * then allocator will stop allocating blocks on that top-level device.
180  * The default kernel setting is 1000% which will yield 100 allocations per
181  * device. For userland testing, the default setting is 300% which equates
182  * to 30 allocations per device.
183  */
184 #ifdef _KERNEL
185 int zfs_vdev_queue_depth_pct = 1000;
186 #else
187 int zfs_vdev_queue_depth_pct = 300;
188 #endif
189 
190 /*
191  * When performing allocations for a given metaslab, we want to make sure that
192  * there are enough IOs to aggregate together to improve throughput. We want to
193  * ensure that there are at least 128k worth of IOs that can be aggregated, and
194  * we assume that the average allocation size is 4k, so we need the queue depth
195  * to be 32 per allocator to get good aggregation of sequential writes.
196  */
197 int zfs_vdev_def_queue_depth = 32;
198 
199 
200 int
201 vdev_queue_offset_compare(const void *x1, const void *x2)
202 {
203 	const zio_t *z1 = x1;
204 	const zio_t *z2 = x2;
205 
206 	if (z1->io_offset < z2->io_offset)
207 		return (-1);
208 	if (z1->io_offset > z2->io_offset)
209 		return (1);
210 
211 	if (z1 < z2)
212 		return (-1);
213 	if (z1 > z2)
214 		return (1);
215 
216 	return (0);
217 }
218 
219 static inline avl_tree_t *
220 vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p)
221 {
222 	return (&vq->vq_class[p].vqc_queued_tree);
223 }
224 
225 static inline avl_tree_t *
226 vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t)
227 {
228 	ASSERT(t == ZIO_TYPE_READ || t == ZIO_TYPE_WRITE);
229 	if (t == ZIO_TYPE_READ)
230 		return (&vq->vq_read_offset_tree);
231 	else
232 		return (&vq->vq_write_offset_tree);
233 }
234 
235 int
236 vdev_queue_timestamp_compare(const void *x1, const void *x2)
237 {
238 	const zio_t *z1 = x1;
239 	const zio_t *z2 = x2;
240 
241 	if (z1->io_timestamp < z2->io_timestamp)
242 		return (-1);
243 	if (z1->io_timestamp > z2->io_timestamp)
244 		return (1);
245 
246 	if (z1 < z2)
247 		return (-1);
248 	if (z1 > z2)
249 		return (1);
250 
251 	return (0);
252 }
253 
254 void
255 vdev_queue_init(vdev_t *vd)
256 {
257 	vdev_queue_t *vq = &vd->vdev_queue;
258 
259 	mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
260 	vq->vq_vdev = vd;
261 
262 	avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
263 	    sizeof (zio_t), offsetof(struct zio, io_queue_node));
264 	avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ),
265 	    vdev_queue_offset_compare, sizeof (zio_t),
266 	    offsetof(struct zio, io_offset_node));
267 	avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE),
268 	    vdev_queue_offset_compare, sizeof (zio_t),
269 	    offsetof(struct zio, io_offset_node));
270 
271 	for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
272 		int (*compfn) (const void *, const void *);
273 
274 		/*
275 		 * The synchronous i/o queues are dispatched in FIFO rather
276 		 * than LBA order.  This provides more consistent latency for
277 		 * these i/os.
278 		 */
279 		if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE)
280 			compfn = vdev_queue_timestamp_compare;
281 		else
282 			compfn = vdev_queue_offset_compare;
283 
284 		avl_create(vdev_queue_class_tree(vq, p), compfn,
285 		    sizeof (zio_t), offsetof(struct zio, io_queue_node));
286 	}
287 }
288 
289 void
290 vdev_queue_fini(vdev_t *vd)
291 {
292 	vdev_queue_t *vq = &vd->vdev_queue;
293 
294 	for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
295 		avl_destroy(vdev_queue_class_tree(vq, p));
296 	avl_destroy(&vq->vq_active_tree);
297 	avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
298 	avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE));
299 
300 	mutex_destroy(&vq->vq_lock);
301 }
302 
303 static void
304 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
305 {
306 	spa_t *spa = zio->io_spa;
307 
308 	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
309 	avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
310 	avl_add(vdev_queue_type_tree(vq, zio->io_type), zio);
311 
312 	mutex_enter(&spa->spa_iokstat_lock);
313 	spa->spa_queue_stats[zio->io_priority].spa_queued++;
314 	if (spa->spa_iokstat != NULL)
315 		kstat_waitq_enter(spa->spa_iokstat->ks_data);
316 	mutex_exit(&spa->spa_iokstat_lock);
317 }
318 
319 static void
320 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
321 {
322 	spa_t *spa = zio->io_spa;
323 
324 	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
325 	avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
326 	avl_remove(vdev_queue_type_tree(vq, zio->io_type), zio);
327 
328 	mutex_enter(&spa->spa_iokstat_lock);
329 	ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
330 	spa->spa_queue_stats[zio->io_priority].spa_queued--;
331 	if (spa->spa_iokstat != NULL)
332 		kstat_waitq_exit(spa->spa_iokstat->ks_data);
333 	mutex_exit(&spa->spa_iokstat_lock);
334 }
335 
336 static void
337 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
338 {
339 	spa_t *spa = zio->io_spa;
340 	ASSERT(MUTEX_HELD(&vq->vq_lock));
341 	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
342 	vq->vq_class[zio->io_priority].vqc_active++;
343 	avl_add(&vq->vq_active_tree, zio);
344 
345 	mutex_enter(&spa->spa_iokstat_lock);
346 	spa->spa_queue_stats[zio->io_priority].spa_active++;
347 	if (spa->spa_iokstat != NULL)
348 		kstat_runq_enter(spa->spa_iokstat->ks_data);
349 	mutex_exit(&spa->spa_iokstat_lock);
350 }
351 
352 static void
353 vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
354 {
355 	spa_t *spa = zio->io_spa;
356 	ASSERT(MUTEX_HELD(&vq->vq_lock));
357 	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
358 	vq->vq_class[zio->io_priority].vqc_active--;
359 	avl_remove(&vq->vq_active_tree, zio);
360 
361 	mutex_enter(&spa->spa_iokstat_lock);
362 	ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
363 	spa->spa_queue_stats[zio->io_priority].spa_active--;
364 	if (spa->spa_iokstat != NULL) {
365 		kstat_io_t *ksio = spa->spa_iokstat->ks_data;
366 
367 		kstat_runq_exit(spa->spa_iokstat->ks_data);
368 		if (zio->io_type == ZIO_TYPE_READ) {
369 			ksio->reads++;
370 			ksio->nread += zio->io_size;
371 		} else if (zio->io_type == ZIO_TYPE_WRITE) {
372 			ksio->writes++;
373 			ksio->nwritten += zio->io_size;
374 		}
375 	}
376 	mutex_exit(&spa->spa_iokstat_lock);
377 }
378 
379 static void
380 vdev_queue_agg_io_done(zio_t *aio)
381 {
382 	if (aio->io_type == ZIO_TYPE_READ) {
383 		zio_t *pio;
384 		zio_link_t *zl = NULL;
385 		while ((pio = zio_walk_parents(aio, &zl)) != NULL) {
386 			abd_copy_off(pio->io_abd, aio->io_abd,
387 			    0, pio->io_offset - aio->io_offset, pio->io_size);
388 		}
389 	}
390 
391 	abd_free(aio->io_abd);
392 }
393 
394 static int
395 vdev_queue_class_min_active(zio_priority_t p)
396 {
397 	switch (p) {
398 	case ZIO_PRIORITY_SYNC_READ:
399 		return (zfs_vdev_sync_read_min_active);
400 	case ZIO_PRIORITY_SYNC_WRITE:
401 		return (zfs_vdev_sync_write_min_active);
402 	case ZIO_PRIORITY_ASYNC_READ:
403 		return (zfs_vdev_async_read_min_active);
404 	case ZIO_PRIORITY_ASYNC_WRITE:
405 		return (zfs_vdev_async_write_min_active);
406 	case ZIO_PRIORITY_SCRUB:
407 		return (zfs_vdev_scrub_min_active);
408 	case ZIO_PRIORITY_REMOVAL:
409 		return (zfs_vdev_removal_min_active);
410 	default:
411 		panic("invalid priority %u", p);
412 		return (0);
413 	}
414 }
415 
416 static int
417 vdev_queue_max_async_writes(spa_t *spa)
418 {
419 	int writes;
420 	uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total;
421 	uint64_t min_bytes = zfs_dirty_data_max *
422 	    zfs_vdev_async_write_active_min_dirty_percent / 100;
423 	uint64_t max_bytes = zfs_dirty_data_max *
424 	    zfs_vdev_async_write_active_max_dirty_percent / 100;
425 
426 	/*
427 	 * Sync tasks correspond to interactive user actions. To reduce the
428 	 * execution time of those actions we push data out as fast as possible.
429 	 */
430 	if (spa_has_pending_synctask(spa)) {
431 		return (zfs_vdev_async_write_max_active);
432 	}
433 
434 	if (dirty < min_bytes)
435 		return (zfs_vdev_async_write_min_active);
436 	if (dirty > max_bytes)
437 		return (zfs_vdev_async_write_max_active);
438 
439 	/*
440 	 * linear interpolation:
441 	 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
442 	 * move right by min_bytes
443 	 * move up by min_writes
444 	 */
445 	writes = (dirty - min_bytes) *
446 	    (zfs_vdev_async_write_max_active -
447 	    zfs_vdev_async_write_min_active) /
448 	    (max_bytes - min_bytes) +
449 	    zfs_vdev_async_write_min_active;
450 	ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
451 	ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
452 	return (writes);
453 }
454 
455 static int
456 vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
457 {
458 	switch (p) {
459 	case ZIO_PRIORITY_SYNC_READ:
460 		return (zfs_vdev_sync_read_max_active);
461 	case ZIO_PRIORITY_SYNC_WRITE:
462 		return (zfs_vdev_sync_write_max_active);
463 	case ZIO_PRIORITY_ASYNC_READ:
464 		return (zfs_vdev_async_read_max_active);
465 	case ZIO_PRIORITY_ASYNC_WRITE:
466 		return (vdev_queue_max_async_writes(spa));
467 	case ZIO_PRIORITY_SCRUB:
468 		return (zfs_vdev_scrub_max_active);
469 	case ZIO_PRIORITY_REMOVAL:
470 		return (zfs_vdev_removal_max_active);
471 	default:
472 		panic("invalid priority %u", p);
473 		return (0);
474 	}
475 }
476 
477 /*
478  * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
479  * there is no eligible class.
480  */
481 static zio_priority_t
482 vdev_queue_class_to_issue(vdev_queue_t *vq)
483 {
484 	spa_t *spa = vq->vq_vdev->vdev_spa;
485 	zio_priority_t p;
486 
487 	if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
488 		return (ZIO_PRIORITY_NUM_QUEUEABLE);
489 
490 	/* find a queue that has not reached its minimum # outstanding i/os */
491 	for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
492 		if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
493 		    vq->vq_class[p].vqc_active <
494 		    vdev_queue_class_min_active(p))
495 			return (p);
496 	}
497 
498 	/*
499 	 * If we haven't found a queue, look for one that hasn't reached its
500 	 * maximum # outstanding i/os.
501 	 */
502 	for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
503 		if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
504 		    vq->vq_class[p].vqc_active <
505 		    vdev_queue_class_max_active(spa, p))
506 			return (p);
507 	}
508 
509 	/* No eligible queued i/os */
510 	return (ZIO_PRIORITY_NUM_QUEUEABLE);
511 }
512 
513 /*
514  * Compute the range spanned by two i/os, which is the endpoint of the last
515  * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
516  * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
517  * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
518  */
519 #define	IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
520 #define	IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
521 
522 static zio_t *
523 vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
524 {
525 	zio_t *first, *last, *aio, *dio, *mandatory, *nio;
526 	uint64_t maxgap = 0;
527 	uint64_t size;
528 	boolean_t stretch = B_FALSE;
529 	avl_tree_t *t = vdev_queue_type_tree(vq, zio->io_type);
530 	enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
531 
532 	if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
533 		return (NULL);
534 
535 	first = last = zio;
536 
537 	if (zio->io_type == ZIO_TYPE_READ)
538 		maxgap = zfs_vdev_read_gap_limit;
539 
540 	/*
541 	 * We can aggregate I/Os that are sufficiently adjacent and of
542 	 * the same flavor, as expressed by the AGG_INHERIT flags.
543 	 * The latter requirement is necessary so that certain
544 	 * attributes of the I/O, such as whether it's a normal I/O
545 	 * or a scrub/resilver, can be preserved in the aggregate.
546 	 * We can include optional I/Os, but don't allow them
547 	 * to begin a range as they add no benefit in that situation.
548 	 */
549 
550 	/*
551 	 * We keep track of the last non-optional I/O.
552 	 */
553 	mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
554 
555 	/*
556 	 * Walk backwards through sufficiently contiguous I/Os
557 	 * recording the last non-optional I/O.
558 	 */
559 	while ((dio = AVL_PREV(t, first)) != NULL &&
560 	    (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
561 	    IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
562 	    IO_GAP(dio, first) <= maxgap &&
563 	    dio->io_type == zio->io_type) {
564 		first = dio;
565 		if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
566 			mandatory = first;
567 	}
568 
569 	/*
570 	 * Skip any initial optional I/Os.
571 	 */
572 	while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
573 		first = AVL_NEXT(t, first);
574 		ASSERT(first != NULL);
575 	}
576 
577 	/*
578 	 * Walk forward through sufficiently contiguous I/Os.
579 	 * The aggregation limit does not apply to optional i/os, so that
580 	 * we can issue contiguous writes even if they are larger than the
581 	 * aggregation limit.
582 	 */
583 	while ((dio = AVL_NEXT(t, last)) != NULL &&
584 	    (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
585 	    (IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit ||
586 	    (dio->io_flags & ZIO_FLAG_OPTIONAL)) &&
587 	    IO_GAP(last, dio) <= maxgap &&
588 	    dio->io_type == zio->io_type) {
589 		last = dio;
590 		if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
591 			mandatory = last;
592 	}
593 
594 	/*
595 	 * Now that we've established the range of the I/O aggregation
596 	 * we must decide what to do with trailing optional I/Os.
597 	 * For reads, there's nothing to do. While we are unable to
598 	 * aggregate further, it's possible that a trailing optional
599 	 * I/O would allow the underlying device to aggregate with
600 	 * subsequent I/Os. We must therefore determine if the next
601 	 * non-optional I/O is close enough to make aggregation
602 	 * worthwhile.
603 	 */
604 	if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
605 		zio_t *nio = last;
606 		while ((dio = AVL_NEXT(t, nio)) != NULL &&
607 		    IO_GAP(nio, dio) == 0 &&
608 		    IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
609 			nio = dio;
610 			if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
611 				stretch = B_TRUE;
612 				break;
613 			}
614 		}
615 	}
616 
617 	if (stretch) {
618 		/*
619 		 * We are going to include an optional io in our aggregated
620 		 * span, thus closing the write gap.  Only mandatory i/os can
621 		 * start aggregated spans, so make sure that the next i/o
622 		 * after our span is mandatory.
623 		 */
624 		dio = AVL_NEXT(t, last);
625 		dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
626 	} else {
627 		/* do not include the optional i/o */
628 		while (last != mandatory && last != first) {
629 			ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
630 			last = AVL_PREV(t, last);
631 			ASSERT(last != NULL);
632 		}
633 	}
634 
635 	if (first == last)
636 		return (NULL);
637 
638 	size = IO_SPAN(first, last);
639 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
640 
641 	aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
642 	    abd_alloc_for_io(size, B_TRUE), size, first->io_type,
643 	    zio->io_priority, flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
644 	    vdev_queue_agg_io_done, NULL);
645 	aio->io_timestamp = first->io_timestamp;
646 
647 	nio = first;
648 	do {
649 		dio = nio;
650 		nio = AVL_NEXT(t, dio);
651 		ASSERT3U(dio->io_type, ==, aio->io_type);
652 
653 		if (dio->io_flags & ZIO_FLAG_NODATA) {
654 			ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
655 			abd_zero_off(aio->io_abd,
656 			    dio->io_offset - aio->io_offset, dio->io_size);
657 		} else if (dio->io_type == ZIO_TYPE_WRITE) {
658 			abd_copy_off(aio->io_abd, dio->io_abd,
659 			    dio->io_offset - aio->io_offset, 0, dio->io_size);
660 		}
661 
662 		zio_add_child(dio, aio);
663 		vdev_queue_io_remove(vq, dio);
664 		zio_vdev_io_bypass(dio);
665 		zio_execute(dio);
666 	} while (dio != last);
667 
668 	return (aio);
669 }
670 
671 static zio_t *
672 vdev_queue_io_to_issue(vdev_queue_t *vq)
673 {
674 	zio_t *zio, *aio;
675 	zio_priority_t p;
676 	avl_index_t idx;
677 	avl_tree_t *tree;
678 	zio_t search;
679 
680 again:
681 	ASSERT(MUTEX_HELD(&vq->vq_lock));
682 
683 	p = vdev_queue_class_to_issue(vq);
684 
685 	if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
686 		/* No eligible queued i/os */
687 		return (NULL);
688 	}
689 
690 	/*
691 	 * For LBA-ordered queues (async / scrub), issue the i/o which follows
692 	 * the most recently issued i/o in LBA (offset) order.
693 	 *
694 	 * For FIFO queues (sync), issue the i/o with the lowest timestamp.
695 	 */
696 	tree = vdev_queue_class_tree(vq, p);
697 	search.io_timestamp = 0;
698 	search.io_offset = vq->vq_last_offset + 1;
699 	VERIFY3P(avl_find(tree, &search, &idx), ==, NULL);
700 	zio = avl_nearest(tree, idx, AVL_AFTER);
701 	if (zio == NULL)
702 		zio = avl_first(tree);
703 	ASSERT3U(zio->io_priority, ==, p);
704 
705 	aio = vdev_queue_aggregate(vq, zio);
706 	if (aio != NULL)
707 		zio = aio;
708 	else
709 		vdev_queue_io_remove(vq, zio);
710 
711 	/*
712 	 * If the I/O is or was optional and therefore has no data, we need to
713 	 * simply discard it. We need to drop the vdev queue's lock to avoid a
714 	 * deadlock that we could encounter since this I/O will complete
715 	 * immediately.
716 	 */
717 	if (zio->io_flags & ZIO_FLAG_NODATA) {
718 		mutex_exit(&vq->vq_lock);
719 		zio_vdev_io_bypass(zio);
720 		zio_execute(zio);
721 		mutex_enter(&vq->vq_lock);
722 		goto again;
723 	}
724 
725 	vdev_queue_pending_add(vq, zio);
726 	vq->vq_last_offset = zio->io_offset;
727 
728 	return (zio);
729 }
730 
731 zio_t *
732 vdev_queue_io(zio_t *zio)
733 {
734 	vdev_queue_t *vq = &zio->io_vd->vdev_queue;
735 	zio_t *nio;
736 
737 	if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
738 		return (zio);
739 
740 	/*
741 	 * Children i/os inherent their parent's priority, which might
742 	 * not match the child's i/o type.  Fix it up here.
743 	 */
744 	if (zio->io_type == ZIO_TYPE_READ) {
745 		if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
746 		    zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
747 		    zio->io_priority != ZIO_PRIORITY_SCRUB &&
748 		    zio->io_priority != ZIO_PRIORITY_REMOVAL)
749 			zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
750 	} else {
751 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
752 		if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
753 		    zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE &&
754 		    zio->io_priority != ZIO_PRIORITY_REMOVAL)
755 			zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
756 	}
757 
758 	zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
759 
760 	mutex_enter(&vq->vq_lock);
761 	zio->io_timestamp = gethrtime();
762 	vdev_queue_io_add(vq, zio);
763 	nio = vdev_queue_io_to_issue(vq);
764 	mutex_exit(&vq->vq_lock);
765 
766 	if (nio == NULL)
767 		return (NULL);
768 
769 	if (nio->io_done == vdev_queue_agg_io_done) {
770 		zio_nowait(nio);
771 		return (NULL);
772 	}
773 
774 	return (nio);
775 }
776 
777 void
778 vdev_queue_io_done(zio_t *zio)
779 {
780 	vdev_queue_t *vq = &zio->io_vd->vdev_queue;
781 	zio_t *nio;
782 
783 	mutex_enter(&vq->vq_lock);
784 
785 	vdev_queue_pending_remove(vq, zio);
786 
787 	vq->vq_io_complete_ts = gethrtime();
788 
789 	while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
790 		mutex_exit(&vq->vq_lock);
791 		if (nio->io_done == vdev_queue_agg_io_done) {
792 			zio_nowait(nio);
793 		} else {
794 			zio_vdev_io_reissue(nio);
795 			zio_execute(nio);
796 		}
797 		mutex_enter(&vq->vq_lock);
798 	}
799 
800 	mutex_exit(&vq->vq_lock);
801 }
802