4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
27 * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
28 * Copyright (c) 2014 Integros [integros.com]
31 #include <sys/zfs_context.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/spa_impl.h>
36 #include <sys/dsl_pool.h>
42 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The
43 * I/O scheduler determines when and in what order those operations are
44 * issued. The I/O scheduler divides operations into six I/O classes
45 * prioritized in the following order: sync read, sync write, async read,
46 * async write, scrub/resilver and trim. Each queue defines the minimum and
47 * maximum number of concurrent operations that may be issued to the device.
48 * In addition, the device has an aggregate maximum. Note that the sum of the
49 * per-queue minimums must not exceed the aggregate maximum, and if the
50 * aggregate maximum is equal to or greater than the sum of the per-queue
51 * maximums, the per-queue minimum has no effect.
53 * For many physical devices, throughput increases with the number of
54 * concurrent operations, but latency typically suffers. Further, physical
55 * devices typically have a limit at which more concurrent operations have no
56 * effect on throughput or can actually cause it to decrease.
58 * The scheduler selects the next operation to issue by first looking for an
59 * I/O class whose minimum has not been satisfied. Once all are satisfied and
60 * the aggregate maximum has not been hit, the scheduler looks for classes
61 * whose maximum has not been satisfied. Iteration through the I/O classes is
62 * done in the order specified above. No further operations are issued if the
63 * aggregate maximum number of concurrent operations has been hit or if there
64 * are no operations queued for an I/O class that has not hit its maximum.
65 * Every time an I/O is queued or an operation completes, the I/O scheduler
66 * looks for new operations to issue.
68 * All I/O classes have a fixed maximum number of outstanding operations
69 * except for the async write class. Asynchronous writes represent the data
70 * that is committed to stable storage during the syncing stage for
71 * transaction groups (see txg.c). Transaction groups enter the syncing state
72 * periodically so the number of queued async writes will quickly burst up and
73 * then bleed down to zero. Rather than servicing them as quickly as possible,
74 * the I/O scheduler changes the maximum number of active async write I/Os
75 * according to the amount of dirty data in the pool (see dsl_pool.c). Since
76 * both throughput and latency typically increase with the number of
77 * concurrent operations issued to physical devices, reducing the burstiness
78 * in the number of concurrent operations also stabilizes the response time of
79 * operations from other -- and in particular synchronous -- queues. In broad
80 * strokes, the I/O scheduler will issue more concurrent operations from the
81 * async write queue as there's more dirty data in the pool.
85 * The number of concurrent operations issued for the async write I/O class
86 * follows a piece-wise linear function defined by a few adjustable points.
88 * | o---------| <-- zfs_vdev_async_write_max_active
95 * |------------o | | <-- zfs_vdev_async_write_min_active
96 * 0|____________^______|_________|
97 * 0% | | 100% of zfs_dirty_data_max
99 * | `-- zfs_vdev_async_write_active_max_dirty_percent
100 * `--------- zfs_vdev_async_write_active_min_dirty_percent
102 * Until the amount of dirty data exceeds a minimum percentage of the dirty
103 * data allowed in the pool, the I/O scheduler will limit the number of
104 * concurrent operations to the minimum. As that threshold is crossed, the
105 * number of concurrent operations issued increases linearly to the maximum at
106 * the specified maximum percentage of the dirty data allowed in the pool.
108 * Ideally, the amount of dirty data on a busy pool will stay in the sloped
109 * part of the function between zfs_vdev_async_write_active_min_dirty_percent
110 * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the
111 * maximum percentage, this indicates that the rate of incoming data is
112 * greater than the rate that the backend storage can handle. In this case, we
113 * must further throttle incoming writes (see dmu_tx_delay() for details).
117 * The maximum number of I/Os active to each device. Ideally, this will be >=
118 * the sum of each queue's max_active. It must be at least the sum of each
119 * queue's min_active.
121 uint32_t zfs_vdev_max_active = 1000;
124 * Per-queue limits on the number of I/Os active to each device. If the
125 * sum of the queue's max_active is < zfs_vdev_max_active, then the
126 * min_active comes into play. We will send min_active from each queue,
127 * and then select from queues in the order defined by zio_priority_t.
129 * In general, smaller max_active's will lead to lower latency of synchronous
130 * operations. Larger max_active's may lead to higher overall throughput,
131 * depending on underlying storage.
133 * The ratio of the queues' max_actives determines the balance of performance
134 * between reads, writes, and scrubs. E.g., increasing
135 * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete
136 * more quickly, but reads and writes to have higher latency and lower
139 uint32_t zfs_vdev_sync_read_min_active = 10;
140 uint32_t zfs_vdev_sync_read_max_active = 10;
141 uint32_t zfs_vdev_sync_write_min_active = 10;
142 uint32_t zfs_vdev_sync_write_max_active = 10;
143 uint32_t zfs_vdev_async_read_min_active = 1;
144 uint32_t zfs_vdev_async_read_max_active = 3;
145 uint32_t zfs_vdev_async_write_min_active = 1;
146 uint32_t zfs_vdev_async_write_max_active = 10;
147 uint32_t zfs_vdev_scrub_min_active = 1;
148 uint32_t zfs_vdev_scrub_max_active = 2;
149 uint32_t zfs_vdev_trim_min_active = 1;
151 * TRIM max active is large in comparison to the other values due to the fact
152 * that TRIM IOs are coalesced at the device layer. This value is set such
153 * that a typical SSD can process the queued IOs in a single request.
155 uint32_t zfs_vdev_trim_max_active = 64;
159 * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent
160 * dirty data, use zfs_vdev_async_write_min_active. When it has more than
161 * zfs_vdev_async_write_active_max_dirty_percent, use
162 * zfs_vdev_async_write_max_active. The value is linearly interpolated
163 * between min and max.
165 int zfs_vdev_async_write_active_min_dirty_percent = 30;
166 int zfs_vdev_async_write_active_max_dirty_percent = 60;
169 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
170 * For read I/Os, we also aggregate across small adjacency gaps; for writes
171 * we include spans of optional I/Os to aid aggregation at the disk even when
172 * they aren't able to help us aggregate at this level.
174 int zfs_vdev_aggregation_limit = SPA_OLD_MAXBLOCKSIZE;
175 int zfs_vdev_read_gap_limit = 32 << 10;
176 int zfs_vdev_write_gap_limit = 4 << 10;
179 SYSCTL_DECL(_vfs_zfs_vdev);
181 TUNABLE_INT("vfs.zfs.vdev.async_write_active_min_dirty_percent",
182 &zfs_vdev_async_write_active_min_dirty_percent);
183 static int sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS);
184 SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_min_dirty_percent,
185 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
186 sysctl_zfs_async_write_active_min_dirty_percent, "I",
187 "Percentage of async write dirty data below which "
188 "async_write_min_active is used.");
190 TUNABLE_INT("vfs.zfs.vdev.async_write_active_max_dirty_percent",
191 &zfs_vdev_async_write_active_max_dirty_percent);
192 static int sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS);
193 SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_max_dirty_percent,
194 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
195 sysctl_zfs_async_write_active_max_dirty_percent, "I",
196 "Percentage of async write dirty data above which "
197 "async_write_max_active is used.");
199 TUNABLE_INT("vfs.zfs.vdev.max_active", &zfs_vdev_max_active);
200 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RWTUN,
201 &zfs_vdev_max_active, 0,
202 "The maximum number of I/Os of all types active for each device.");
204 #define ZFS_VDEV_QUEUE_KNOB_MIN(name) \
205 TUNABLE_INT("vfs.zfs.vdev." #name "_min_active", \
206 &zfs_vdev_ ## name ## _min_active); \
207 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, \
208 CTLFLAG_RWTUN, &zfs_vdev_ ## name ## _min_active, 0, \
209 "Initial number of I/O requests of type " #name \
210 " active for each device");
212 #define ZFS_VDEV_QUEUE_KNOB_MAX(name) \
213 TUNABLE_INT("vfs.zfs.vdev." #name "_max_active", \
214 &zfs_vdev_ ## name ## _max_active); \
215 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, \
216 CTLFLAG_RWTUN, &zfs_vdev_ ## name ## _max_active, 0, \
217 "Maximum number of I/O requests of type " #name \
218 " active for each device");
220 ZFS_VDEV_QUEUE_KNOB_MIN(sync_read);
221 ZFS_VDEV_QUEUE_KNOB_MAX(sync_read);
222 ZFS_VDEV_QUEUE_KNOB_MIN(sync_write);
223 ZFS_VDEV_QUEUE_KNOB_MAX(sync_write);
224 ZFS_VDEV_QUEUE_KNOB_MIN(async_read);
225 ZFS_VDEV_QUEUE_KNOB_MAX(async_read);
226 ZFS_VDEV_QUEUE_KNOB_MIN(async_write);
227 ZFS_VDEV_QUEUE_KNOB_MAX(async_write);
228 ZFS_VDEV_QUEUE_KNOB_MIN(scrub);
229 ZFS_VDEV_QUEUE_KNOB_MAX(scrub);
230 ZFS_VDEV_QUEUE_KNOB_MIN(trim);
231 ZFS_VDEV_QUEUE_KNOB_MAX(trim);
233 #undef ZFS_VDEV_QUEUE_KNOB
235 TUNABLE_INT("vfs.zfs.vdev.aggregation_limit", &zfs_vdev_aggregation_limit);
236 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RWTUN,
237 &zfs_vdev_aggregation_limit, 0,
238 "I/O requests are aggregated up to this size");
239 TUNABLE_INT("vfs.zfs.vdev.read_gap_limit", &zfs_vdev_read_gap_limit);
240 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RWTUN,
241 &zfs_vdev_read_gap_limit, 0,
242 "Acceptable gap between two reads being aggregated");
243 TUNABLE_INT("vfs.zfs.vdev.write_gap_limit", &zfs_vdev_write_gap_limit);
244 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RWTUN,
245 &zfs_vdev_write_gap_limit, 0,
246 "Acceptable gap between two writes being aggregated");
249 sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS)
253 val = zfs_vdev_async_write_active_min_dirty_percent;
254 err = sysctl_handle_int(oidp, &val, 0, req);
255 if (err != 0 || req->newptr == NULL)
258 if (val < 0 || val > 100 ||
259 val >= zfs_vdev_async_write_active_max_dirty_percent)
262 zfs_vdev_async_write_active_min_dirty_percent = val;
268 sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS)
272 val = zfs_vdev_async_write_active_max_dirty_percent;
273 err = sysctl_handle_int(oidp, &val, 0, req);
274 if (err != 0 || req->newptr == NULL)
277 if (val < 0 || val > 100 ||
278 val <= zfs_vdev_async_write_active_min_dirty_percent)
281 zfs_vdev_async_write_active_max_dirty_percent = val;
288 vdev_queue_offset_compare(const void *x1, const void *x2)
290 const zio_t *z1 = x1;
291 const zio_t *z2 = x2;
293 if (z1->io_offset < z2->io_offset)
295 if (z1->io_offset > z2->io_offset)
306 static inline avl_tree_t *
307 vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p)
309 return (&vq->vq_class[p].vqc_queued_tree);
312 static inline avl_tree_t *
313 vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t)
315 if (t == ZIO_TYPE_READ)
316 return (&vq->vq_read_offset_tree);
317 else if (t == ZIO_TYPE_WRITE)
318 return (&vq->vq_write_offset_tree);
324 vdev_queue_timestamp_compare(const void *x1, const void *x2)
326 const zio_t *z1 = x1;
327 const zio_t *z2 = x2;
329 if (z1->io_timestamp < z2->io_timestamp)
331 if (z1->io_timestamp > z2->io_timestamp)
334 if (z1->io_offset < z2->io_offset)
336 if (z1->io_offset > z2->io_offset)
348 vdev_queue_init(vdev_t *vd)
350 vdev_queue_t *vq = &vd->vdev_queue;
352 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
355 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
356 sizeof (zio_t), offsetof(struct zio, io_queue_node));
357 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ),
358 vdev_queue_offset_compare, sizeof (zio_t),
359 offsetof(struct zio, io_offset_node));
360 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE),
361 vdev_queue_offset_compare, sizeof (zio_t),
362 offsetof(struct zio, io_offset_node));
364 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
365 int (*compfn) (const void *, const void *);
368 * The synchronous i/o queues are dispatched in FIFO rather
369 * than LBA order. This provides more consistent latency for
372 if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE)
373 compfn = vdev_queue_timestamp_compare;
375 compfn = vdev_queue_offset_compare;
377 avl_create(vdev_queue_class_tree(vq, p), compfn,
378 sizeof (zio_t), offsetof(struct zio, io_queue_node));
381 vq->vq_lastoffset = 0;
385 vdev_queue_fini(vdev_t *vd)
387 vdev_queue_t *vq = &vd->vdev_queue;
389 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
390 avl_destroy(vdev_queue_class_tree(vq, p));
391 avl_destroy(&vq->vq_active_tree);
392 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
393 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE));
395 mutex_destroy(&vq->vq_lock);
399 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
401 spa_t *spa = zio->io_spa;
403 ASSERT(MUTEX_HELD(&vq->vq_lock));
404 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
405 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
406 qtt = vdev_queue_type_tree(vq, zio->io_type);
411 mutex_enter(&spa->spa_iokstat_lock);
412 spa->spa_queue_stats[zio->io_priority].spa_queued++;
413 if (spa->spa_iokstat != NULL)
414 kstat_waitq_enter(spa->spa_iokstat->ks_data);
415 mutex_exit(&spa->spa_iokstat_lock);
420 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
422 spa_t *spa = zio->io_spa;
424 ASSERT(MUTEX_HELD(&vq->vq_lock));
425 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
426 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
427 qtt = vdev_queue_type_tree(vq, zio->io_type);
429 avl_remove(qtt, zio);
432 mutex_enter(&spa->spa_iokstat_lock);
433 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
434 spa->spa_queue_stats[zio->io_priority].spa_queued--;
435 if (spa->spa_iokstat != NULL)
436 kstat_waitq_exit(spa->spa_iokstat->ks_data);
437 mutex_exit(&spa->spa_iokstat_lock);
442 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
444 spa_t *spa = zio->io_spa;
445 ASSERT(MUTEX_HELD(&vq->vq_lock));
446 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
447 vq->vq_class[zio->io_priority].vqc_active++;
448 avl_add(&vq->vq_active_tree, zio);
451 mutex_enter(&spa->spa_iokstat_lock);
452 spa->spa_queue_stats[zio->io_priority].spa_active++;
453 if (spa->spa_iokstat != NULL)
454 kstat_runq_enter(spa->spa_iokstat->ks_data);
455 mutex_exit(&spa->spa_iokstat_lock);
460 vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
462 spa_t *spa = zio->io_spa;
463 ASSERT(MUTEX_HELD(&vq->vq_lock));
464 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
465 vq->vq_class[zio->io_priority].vqc_active--;
466 avl_remove(&vq->vq_active_tree, zio);
469 mutex_enter(&spa->spa_iokstat_lock);
470 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
471 spa->spa_queue_stats[zio->io_priority].spa_active--;
472 if (spa->spa_iokstat != NULL) {
473 kstat_io_t *ksio = spa->spa_iokstat->ks_data;
475 kstat_runq_exit(spa->spa_iokstat->ks_data);
476 if (zio->io_type == ZIO_TYPE_READ) {
478 ksio->nread += zio->io_size;
479 } else if (zio->io_type == ZIO_TYPE_WRITE) {
481 ksio->nwritten += zio->io_size;
484 mutex_exit(&spa->spa_iokstat_lock);
489 vdev_queue_agg_io_done(zio_t *aio)
491 if (aio->io_type == ZIO_TYPE_READ) {
493 while ((pio = zio_walk_parents(aio)) != NULL) {
494 bcopy((char *)aio->io_data + (pio->io_offset -
495 aio->io_offset), pio->io_data, pio->io_size);
499 zio_buf_free(aio->io_data, aio->io_size);
503 vdev_queue_class_min_active(zio_priority_t p)
506 case ZIO_PRIORITY_SYNC_READ:
507 return (zfs_vdev_sync_read_min_active);
508 case ZIO_PRIORITY_SYNC_WRITE:
509 return (zfs_vdev_sync_write_min_active);
510 case ZIO_PRIORITY_ASYNC_READ:
511 return (zfs_vdev_async_read_min_active);
512 case ZIO_PRIORITY_ASYNC_WRITE:
513 return (zfs_vdev_async_write_min_active);
514 case ZIO_PRIORITY_SCRUB:
515 return (zfs_vdev_scrub_min_active);
516 case ZIO_PRIORITY_TRIM:
517 return (zfs_vdev_trim_min_active);
519 panic("invalid priority %u", p);
524 static __noinline int
525 vdev_queue_max_async_writes(spa_t *spa)
528 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total;
529 uint64_t min_bytes = zfs_dirty_data_max *
530 zfs_vdev_async_write_active_min_dirty_percent / 100;
531 uint64_t max_bytes = zfs_dirty_data_max *
532 zfs_vdev_async_write_active_max_dirty_percent / 100;
535 * Sync tasks correspond to interactive user actions. To reduce the
536 * execution time of those actions we push data out as fast as possible.
538 if (spa_has_pending_synctask(spa)) {
539 return (zfs_vdev_async_write_max_active);
542 if (dirty < min_bytes)
543 return (zfs_vdev_async_write_min_active);
544 if (dirty > max_bytes)
545 return (zfs_vdev_async_write_max_active);
548 * linear interpolation:
549 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
550 * move right by min_bytes
551 * move up by min_writes
553 writes = (dirty - min_bytes) *
554 (zfs_vdev_async_write_max_active -
555 zfs_vdev_async_write_min_active) /
556 (max_bytes - min_bytes) +
557 zfs_vdev_async_write_min_active;
558 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
559 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
564 vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
567 case ZIO_PRIORITY_SYNC_READ:
568 return (zfs_vdev_sync_read_max_active);
569 case ZIO_PRIORITY_SYNC_WRITE:
570 return (zfs_vdev_sync_write_max_active);
571 case ZIO_PRIORITY_ASYNC_READ:
572 return (zfs_vdev_async_read_max_active);
573 case ZIO_PRIORITY_ASYNC_WRITE:
574 return (vdev_queue_max_async_writes(spa));
575 case ZIO_PRIORITY_SCRUB:
576 return (zfs_vdev_scrub_max_active);
577 case ZIO_PRIORITY_TRIM:
578 return (zfs_vdev_trim_max_active);
580 panic("invalid priority %u", p);
586 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
587 * there is no eligible class.
589 static zio_priority_t
590 vdev_queue_class_to_issue(vdev_queue_t *vq)
592 spa_t *spa = vq->vq_vdev->vdev_spa;
595 ASSERT(MUTEX_HELD(&vq->vq_lock));
597 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
598 return (ZIO_PRIORITY_NUM_QUEUEABLE);
600 /* find a queue that has not reached its minimum # outstanding i/os */
601 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
602 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
603 vq->vq_class[p].vqc_active <
604 vdev_queue_class_min_active(p))
609 * If we haven't found a queue, look for one that hasn't reached its
610 * maximum # outstanding i/os.
612 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
613 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
614 vq->vq_class[p].vqc_active <
615 vdev_queue_class_max_active(spa, p))
619 /* No eligible queued i/os */
620 return (ZIO_PRIORITY_NUM_QUEUEABLE);
624 * Compute the range spanned by two i/os, which is the endpoint of the last
625 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
626 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
627 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
629 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
630 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
633 vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
635 zio_t *first, *last, *aio, *dio, *mandatory, *nio;
642 ASSERT(MUTEX_HELD(&vq->vq_lock));
644 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
649 if (zio->io_type == ZIO_TYPE_READ)
650 maxgap = zfs_vdev_read_gap_limit;
653 * We can aggregate I/Os that are sufficiently adjacent and of
654 * the same flavor, as expressed by the AGG_INHERIT flags.
655 * The latter requirement is necessary so that certain
656 * attributes of the I/O, such as whether it's a normal I/O
657 * or a scrub/resilver, can be preserved in the aggregate.
658 * We can include optional I/Os, but don't allow them
659 * to begin a range as they add no benefit in that situation.
663 * We keep track of the last non-optional I/O.
665 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
668 * Walk backwards through sufficiently contiguous I/Os
669 * recording the last non-option I/O.
671 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
672 t = vdev_queue_type_tree(vq, zio->io_type);
673 while (t != NULL && (dio = AVL_PREV(t, first)) != NULL &&
674 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
675 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
676 IO_GAP(dio, first) <= maxgap) {
678 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
683 * Skip any initial optional I/Os.
685 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
686 first = AVL_NEXT(t, first);
687 ASSERT(first != NULL);
691 * Walk forward through sufficiently contiguous I/Os.
693 while ((dio = AVL_NEXT(t, last)) != NULL &&
694 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
695 IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit &&
696 IO_GAP(last, dio) <= maxgap) {
698 if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
703 * Now that we've established the range of the I/O aggregation
704 * we must decide what to do with trailing optional I/Os.
705 * For reads, there's nothing to do. While we are unable to
706 * aggregate further, it's possible that a trailing optional
707 * I/O would allow the underlying device to aggregate with
708 * subsequent I/Os. We must therefore determine if the next
709 * non-optional I/O is close enough to make aggregation
713 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
715 while ((dio = AVL_NEXT(t, nio)) != NULL &&
716 IO_GAP(nio, dio) == 0 &&
717 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
719 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
727 /* This may be a no-op. */
728 dio = AVL_NEXT(t, last);
729 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
731 while (last != mandatory && last != first) {
732 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
733 last = AVL_PREV(t, last);
734 ASSERT(last != NULL);
741 size = IO_SPAN(first, last);
742 ASSERT3U(size, <=, zfs_vdev_aggregation_limit);
744 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
745 zio_buf_alloc(size), size, first->io_type, zio->io_priority,
746 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
747 vdev_queue_agg_io_done, NULL);
748 aio->io_timestamp = first->io_timestamp;
753 nio = AVL_NEXT(t, dio);
754 ASSERT3U(dio->io_type, ==, aio->io_type);
756 if (dio->io_flags & ZIO_FLAG_NODATA) {
757 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
758 bzero((char *)aio->io_data + (dio->io_offset -
759 aio->io_offset), dio->io_size);
760 } else if (dio->io_type == ZIO_TYPE_WRITE) {
761 bcopy(dio->io_data, (char *)aio->io_data +
762 (dio->io_offset - aio->io_offset),
766 zio_add_child(dio, aio);
767 vdev_queue_io_remove(vq, dio);
768 zio_vdev_io_bypass(dio);
770 } while (dio != last);
776 vdev_queue_io_to_issue(vdev_queue_t *vq)
785 ASSERT(MUTEX_HELD(&vq->vq_lock));
787 p = vdev_queue_class_to_issue(vq);
789 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
790 /* No eligible queued i/os */
795 * For LBA-ordered queues (async / scrub), issue the i/o which follows
796 * the most recently issued i/o in LBA (offset) order.
798 * For FIFO queues (sync), issue the i/o with the lowest timestamp.
800 tree = vdev_queue_class_tree(vq, p);
801 search.io_timestamp = 0;
802 search.io_offset = vq->vq_last_offset + 1;
803 VERIFY3P(avl_find(tree, &search, &idx), ==, NULL);
804 zio = avl_nearest(tree, idx, AVL_AFTER);
806 zio = avl_first(tree);
807 ASSERT3U(zio->io_priority, ==, p);
809 aio = vdev_queue_aggregate(vq, zio);
813 vdev_queue_io_remove(vq, zio);
816 * If the I/O is or was optional and therefore has no data, we need to
817 * simply discard it. We need to drop the vdev queue's lock to avoid a
818 * deadlock that we could encounter since this I/O will complete
821 if (zio->io_flags & ZIO_FLAG_NODATA) {
822 mutex_exit(&vq->vq_lock);
823 zio_vdev_io_bypass(zio);
825 mutex_enter(&vq->vq_lock);
829 vdev_queue_pending_add(vq, zio);
830 vq->vq_last_offset = zio->io_offset;
836 vdev_queue_io(zio_t *zio)
838 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
841 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
845 * Children i/os inherent their parent's priority, which might
846 * not match the child's i/o type. Fix it up here.
848 if (zio->io_type == ZIO_TYPE_READ) {
849 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
850 zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
851 zio->io_priority != ZIO_PRIORITY_SCRUB)
852 zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
853 } else if (zio->io_type == ZIO_TYPE_WRITE) {
854 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
855 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE)
856 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
858 ASSERT(zio->io_type == ZIO_TYPE_FREE);
859 zio->io_priority = ZIO_PRIORITY_TRIM;
862 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
864 mutex_enter(&vq->vq_lock);
865 zio->io_timestamp = gethrtime();
866 vdev_queue_io_add(vq, zio);
867 nio = vdev_queue_io_to_issue(vq);
868 mutex_exit(&vq->vq_lock);
873 if (nio->io_done == vdev_queue_agg_io_done) {
882 vdev_queue_io_done(zio_t *zio)
884 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
887 mutex_enter(&vq->vq_lock);
889 vdev_queue_pending_remove(vq, zio);
891 vq->vq_io_complete_ts = gethrtime();
893 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
894 mutex_exit(&vq->vq_lock);
895 if (nio->io_done == vdev_queue_agg_io_done) {
898 zio_vdev_io_reissue(nio);
901 mutex_enter(&vq->vq_lock);
904 mutex_exit(&vq->vq_lock);
908 * As these three methods are only used for load calculations we're not concerned
909 * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex
910 * use here, instead we prefer to keep it lock free for performance.
913 vdev_queue_length(vdev_t *vd)
915 return (avl_numnodes(&vd->vdev_queue.vq_active_tree));
919 vdev_queue_lastoffset(vdev_t *vd)
921 return (vd->vdev_queue.vq_lastoffset);
925 vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio)
927 vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size;