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 */
124 uint32_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 */
142 uint32_t zfs_vdev_sync_read_min_active = 10;
143 uint32_t zfs_vdev_sync_read_max_active = 10;
144 uint32_t zfs_vdev_sync_write_min_active = 10;
145 uint32_t zfs_vdev_sync_write_max_active = 10;
146 uint32_t zfs_vdev_async_read_min_active = 1;
147 uint32_t zfs_vdev_async_read_max_active = 3;
148 uint32_t zfs_vdev_async_write_min_active = 1;
149 uint32_t zfs_vdev_async_write_max_active = 10;
150 uint32_t zfs_vdev_scrub_min_active = 1;
151 uint32_t zfs_vdev_scrub_max_active = 2;
152 uint32_t zfs_vdev_removal_min_active = 1;
153 uint32_t zfs_vdev_removal_max_active = 2;
154 uint32_t zfs_vdev_initializing_min_active = 1;
155 uint32_t zfs_vdev_initializing_max_active = 1;
156
157 /*
158 * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent
159 * dirty data, use zfs_vdev_async_write_min_active. When it has more than
160 * zfs_vdev_async_write_active_max_dirty_percent, use
161 * zfs_vdev_async_write_max_active. The value is linearly interpolated
162 * between min and max.
163 */
164 int zfs_vdev_async_write_active_min_dirty_percent = 30;
165 int zfs_vdev_async_write_active_max_dirty_percent = 60;
166
167 /*
168 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
169 * For read I/Os, we also aggregate across small adjacency gaps; for writes
170 * we include spans of optional I/Os to aid aggregation at the disk even when
171 * they aren't able to help us aggregate at this level.
172 */
173 int zfs_vdev_aggregation_limit = 1 << 20;
174 int zfs_vdev_read_gap_limit = 32 << 10;
175 int zfs_vdev_write_gap_limit = 4 << 10;
176
177 /*
178 * Define the queue depth percentage for each top-level. This percentage is
179 * used in conjunction with zfs_vdev_async_max_active to determine how many
180 * allocations a specific top-level vdev should handle. Once the queue depth
181 * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100
182 * then allocator will stop allocating blocks on that top-level device.
183 * The default kernel setting is 1000% which will yield 100 allocations per
184 * device. For userland testing, the default setting is 300% which equates
185 * to 30 allocations per device.
186 */
187 #ifdef _KERNEL
188 int zfs_vdev_queue_depth_pct = 1000;
189 #else
190 int zfs_vdev_queue_depth_pct = 300;
191 #endif
192
193 /*
194 * When performing allocations for a given metaslab, we want to make sure that
195 * there are enough IOs to aggregate together to improve throughput. We want to
196 * ensure that there are at least 128k worth of IOs that can be aggregated, and
197 * we assume that the average allocation size is 4k, so we need the queue depth
198 * to be 32 per allocator to get good aggregation of sequential writes.
199 */
200 int zfs_vdev_def_queue_depth = 32;
201
202
203 int
204 vdev_queue_offset_compare(const void *x1, const void *x2)
205 {
206 const zio_t *z1 = (const zio_t *)x1;
207 const zio_t *z2 = (const zio_t *)x2;
208
209 int cmp = AVL_CMP(z1->io_offset, z2->io_offset);
210
211 if (likely(cmp))
212 return (cmp);
213
214 return (AVL_PCMP(z1, z2));
215 }
216
217 static inline avl_tree_t *
218 vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p)
219 {
220 return (&vq->vq_class[p].vqc_queued_tree);
221 }
222
223 static inline avl_tree_t *
224 vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t)
225 {
226 ASSERT(t == ZIO_TYPE_READ || t == ZIO_TYPE_WRITE);
227 if (t == ZIO_TYPE_READ)
228 return (&vq->vq_read_offset_tree);
229 else
230 return (&vq->vq_write_offset_tree);
231 }
232
233 int
234 vdev_queue_timestamp_compare(const void *x1, const void *x2)
235 {
236 const zio_t *z1 = (const zio_t *)x1;
237 const zio_t *z2 = (const zio_t *)x2;
238
239 int cmp = AVL_CMP(z1->io_timestamp, z2->io_timestamp);
240
241 if (likely(cmp))
242 return (cmp);
243
244 return (AVL_PCMP(z1, z2));
245 }
246
247 void
248 vdev_queue_init(vdev_t *vd)
249 {
250 vdev_queue_t *vq = &vd->vdev_queue;
251
252 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
253 vq->vq_vdev = vd;
254
255 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
256 sizeof (zio_t), offsetof(struct zio, io_queue_node));
257 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ),
258 vdev_queue_offset_compare, sizeof (zio_t),
259 offsetof(struct zio, io_offset_node));
260 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE),
261 vdev_queue_offset_compare, sizeof (zio_t),
262 offsetof(struct zio, io_offset_node));
263
264 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
265 int (*compfn) (const void *, const void *);
266
267 /*
268 * The synchronous i/o queues are dispatched in FIFO rather
269 * than LBA order. This provides more consistent latency for
270 * these i/os.
271 */
272 if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE)
273 compfn = vdev_queue_timestamp_compare;
274 else
275 compfn = vdev_queue_offset_compare;
276
277 avl_create(vdev_queue_class_tree(vq, p), compfn,
278 sizeof (zio_t), offsetof(struct zio, io_queue_node));
279 }
280
281 vq->vq_last_offset = 0;
282 }
283
284 void
285 vdev_queue_fini(vdev_t *vd)
286 {
287 vdev_queue_t *vq = &vd->vdev_queue;
288
289 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
290 avl_destroy(vdev_queue_class_tree(vq, p));
291 avl_destroy(&vq->vq_active_tree);
292 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
293 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE));
294
295 mutex_destroy(&vq->vq_lock);
296 }
297
298 static void
299 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
300 {
301 spa_t *spa = zio->io_spa;
302
303 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
304 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
305 avl_add(vdev_queue_type_tree(vq, zio->io_type), zio);
306
307 mutex_enter(&spa->spa_iokstat_lock);
308 spa->spa_queue_stats[zio->io_priority].spa_queued++;
309 if (spa->spa_iokstat != NULL)
310 kstat_waitq_enter(spa->spa_iokstat->ks_data);
311 mutex_exit(&spa->spa_iokstat_lock);
312 }
313
314 static void
315 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
316 {
317 spa_t *spa = zio->io_spa;
318
319 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
320 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
321 avl_remove(vdev_queue_type_tree(vq, zio->io_type), zio);
322
323 mutex_enter(&spa->spa_iokstat_lock);
324 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
325 spa->spa_queue_stats[zio->io_priority].spa_queued--;
326 if (spa->spa_iokstat != NULL)
327 kstat_waitq_exit(spa->spa_iokstat->ks_data);
328 mutex_exit(&spa->spa_iokstat_lock);
329 }
330
331 static void
332 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
333 {
334 spa_t *spa = zio->io_spa;
335 ASSERT(MUTEX_HELD(&vq->vq_lock));
336 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
337 vq->vq_class[zio->io_priority].vqc_active++;
338 avl_add(&vq->vq_active_tree, zio);
339
340 mutex_enter(&spa->spa_iokstat_lock);
341 spa->spa_queue_stats[zio->io_priority].spa_active++;
342 if (spa->spa_iokstat != NULL)
343 kstat_runq_enter(spa->spa_iokstat->ks_data);
344 mutex_exit(&spa->spa_iokstat_lock);
345 }
346
347 static void
348 vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
349 {
350 spa_t *spa = zio->io_spa;
351 ASSERT(MUTEX_HELD(&vq->vq_lock));
352 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
353 vq->vq_class[zio->io_priority].vqc_active--;
354 avl_remove(&vq->vq_active_tree, zio);
355
356 mutex_enter(&spa->spa_iokstat_lock);
357 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
358 spa->spa_queue_stats[zio->io_priority].spa_active--;
359 if (spa->spa_iokstat != NULL) {
360 kstat_io_t *ksio = spa->spa_iokstat->ks_data;
361
362 kstat_runq_exit(spa->spa_iokstat->ks_data);
363 if (zio->io_type == ZIO_TYPE_READ) {
364 ksio->reads++;
365 ksio->nread += zio->io_size;
366 } else if (zio->io_type == ZIO_TYPE_WRITE) {
367 ksio->writes++;
368 ksio->nwritten += zio->io_size;
369 }
370 }
371 mutex_exit(&spa->spa_iokstat_lock);
372 }
373
374 static void
375 vdev_queue_agg_io_done(zio_t *aio)
376 {
377 if (aio->io_type == ZIO_TYPE_READ) {
378 zio_t *pio;
379 zio_link_t *zl = NULL;
380 while ((pio = zio_walk_parents(aio, &zl)) != NULL) {
381 abd_copy_off(pio->io_abd, aio->io_abd,
382 0, pio->io_offset - aio->io_offset, pio->io_size);
383 }
384 }
385
386 abd_free(aio->io_abd);
387 }
388
389 static int
390 vdev_queue_class_min_active(zio_priority_t p)
391 {
392 switch (p) {
393 case ZIO_PRIORITY_SYNC_READ:
394 return (zfs_vdev_sync_read_min_active);
395 case ZIO_PRIORITY_SYNC_WRITE:
396 return (zfs_vdev_sync_write_min_active);
397 case ZIO_PRIORITY_ASYNC_READ:
398 return (zfs_vdev_async_read_min_active);
399 case ZIO_PRIORITY_ASYNC_WRITE:
400 return (zfs_vdev_async_write_min_active);
401 case ZIO_PRIORITY_SCRUB:
402 return (zfs_vdev_scrub_min_active);
403 case ZIO_PRIORITY_REMOVAL:
404 return (zfs_vdev_removal_min_active);
405 case ZIO_PRIORITY_INITIALIZING:
406 return (zfs_vdev_initializing_min_active);
407 default:
408 panic("invalid priority %u", p);
409 }
410 }
411
412 static int
413 vdev_queue_max_async_writes(spa_t *spa)
414 {
415 int writes;
416 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total;
417 uint64_t min_bytes = zfs_dirty_data_max *
418 zfs_vdev_async_write_active_min_dirty_percent / 100;
419 uint64_t max_bytes = zfs_dirty_data_max *
420 zfs_vdev_async_write_active_max_dirty_percent / 100;
421
422 /*
423 * Sync tasks correspond to interactive user actions. To reduce the
424 * execution time of those actions we push data out as fast as possible.
425 */
426 if (spa_has_pending_synctask(spa)) {
427 return (zfs_vdev_async_write_max_active);
428 }
429
430 if (dirty < min_bytes)
431 return (zfs_vdev_async_write_min_active);
432 if (dirty > max_bytes)
433 return (zfs_vdev_async_write_max_active);
434
435 /*
436 * linear interpolation:
437 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
438 * move right by min_bytes
439 * move up by min_writes
440 */
441 writes = (dirty - min_bytes) *
442 (zfs_vdev_async_write_max_active -
443 zfs_vdev_async_write_min_active) /
444 (max_bytes - min_bytes) +
445 zfs_vdev_async_write_min_active;
446 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
447 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
448 return (writes);
449 }
450
451 static int
452 vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
453 {
454 switch (p) {
455 case ZIO_PRIORITY_SYNC_READ:
456 return (zfs_vdev_sync_read_max_active);
457 case ZIO_PRIORITY_SYNC_WRITE:
458 return (zfs_vdev_sync_write_max_active);
459 case ZIO_PRIORITY_ASYNC_READ:
460 return (zfs_vdev_async_read_max_active);
461 case ZIO_PRIORITY_ASYNC_WRITE:
462 return (vdev_queue_max_async_writes(spa));
463 case ZIO_PRIORITY_SCRUB:
464 return (zfs_vdev_scrub_max_active);
465 case ZIO_PRIORITY_REMOVAL:
466 return (zfs_vdev_removal_max_active);
467 case ZIO_PRIORITY_INITIALIZING:
468 return (zfs_vdev_initializing_max_active);
469 default:
470 panic("invalid priority %u", p);
471 }
472 }
473
474 /*
475 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
476 * there is no eligible class.
477 */
478 static zio_priority_t
479 vdev_queue_class_to_issue(vdev_queue_t *vq)
480 {
481 spa_t *spa = vq->vq_vdev->vdev_spa;
482 zio_priority_t p;
483
484 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
485 return (ZIO_PRIORITY_NUM_QUEUEABLE);
486
487 /* find a queue that has not reached its minimum # outstanding i/os */
488 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
489 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
490 vq->vq_class[p].vqc_active <
491 vdev_queue_class_min_active(p))
492 return (p);
493 }
494
495 /*
496 * If we haven't found a queue, look for one that hasn't reached its
497 * maximum # outstanding i/os.
498 */
499 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
500 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
501 vq->vq_class[p].vqc_active <
502 vdev_queue_class_max_active(spa, p))
503 return (p);
504 }
505
506 /* No eligible queued i/os */
507 return (ZIO_PRIORITY_NUM_QUEUEABLE);
508 }
509
510 /*
511 * Compute the range spanned by two i/os, which is the endpoint of the last
512 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
513 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
514 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
515 */
516 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
517 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
518
519 static zio_t *
520 vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
521 {
522 zio_t *first, *last, *aio, *dio, *mandatory, *nio;
523 zio_link_t *zl = NULL;
524 uint64_t maxgap = 0;
525 uint64_t size;
526 boolean_t stretch = B_FALSE;
527 avl_tree_t *t = vdev_queue_type_tree(vq, zio->io_type);
528 enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
529
530 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
531 return (NULL);
532
533 first = last = zio;
534
535 if (zio->io_type == ZIO_TYPE_READ)
536 maxgap = zfs_vdev_read_gap_limit;
537
538 /*
539 * We can aggregate I/Os that are sufficiently adjacent and of
540 * the same flavor, as expressed by the AGG_INHERIT flags.
541 * The latter requirement is necessary so that certain
542 * attributes of the I/O, such as whether it's a normal I/O
543 * or a scrub/resilver, can be preserved in the aggregate.
544 * We can include optional I/Os, but don't allow them
545 * to begin a range as they add no benefit in that situation.
546 */
547
548 /*
549 * We keep track of the last non-optional I/O.
550 */
551 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
552
553 /*
554 * Walk backwards through sufficiently contiguous I/Os
555 * recording the last non-optional I/O.
556 */
557 while ((dio = AVL_PREV(t, first)) != NULL &&
558 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
559 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
560 IO_GAP(dio, first) <= maxgap &&
561 dio->io_type == zio->io_type) {
562 first = dio;
563 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
564 mandatory = first;
565 }
566
567 /*
568 * Skip any initial optional I/Os.
569 */
570 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
571 first = AVL_NEXT(t, first);
572 ASSERT(first != NULL);
573 }
574
575 /*
576 * Walk forward through sufficiently contiguous I/Os.
577 * The aggregation limit does not apply to optional i/os, so that
578 * we can issue contiguous writes even if they are larger than the
579 * aggregation limit.
580 */
581 while ((dio = AVL_NEXT(t, last)) != NULL &&
582 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
583 (IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit ||
584 (dio->io_flags & ZIO_FLAG_OPTIONAL)) &&
585 IO_GAP(last, dio) <= maxgap &&
586 dio->io_type == zio->io_type) {
587 last = dio;
588 if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
589 mandatory = last;
590 }
591
592 /*
593 * Now that we've established the range of the I/O aggregation
594 * we must decide what to do with trailing optional I/Os.
595 * For reads, there's nothing to do. While we are unable to
596 * aggregate further, it's possible that a trailing optional
597 * I/O would allow the underlying device to aggregate with
598 * subsequent I/Os. We must therefore determine if the next
599 * non-optional I/O is close enough to make aggregation
600 * worthwhile.
601 */
602 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
603 zio_t *nio = last;
604 while ((dio = AVL_NEXT(t, nio)) != NULL &&
605 IO_GAP(nio, dio) == 0 &&
606 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
607 nio = dio;
608 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
609 stretch = B_TRUE;
610 break;
611 }
612 }
613 }
614
615 if (stretch) {
616 /*
617 * We are going to include an optional io in our aggregated
618 * span, thus closing the write gap. Only mandatory i/os can
619 * start aggregated spans, so make sure that the next i/o
620 * after our span is mandatory.
621 */
622 dio = AVL_NEXT(t, last);
623 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
624 } else {
625 /* do not include the optional i/o */
626 while (last != mandatory && last != first) {
627 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
628 last = AVL_PREV(t, last);
629 ASSERT(last != NULL);
630 }
631 }
632
633 if (first == last)
634 return (NULL);
635
636 size = IO_SPAN(first, last);
637 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
638
639 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
640 abd_alloc_for_io(size, B_TRUE), size, first->io_type,
641 zio->io_priority, flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
642 vdev_queue_agg_io_done, NULL);
643 aio->io_timestamp = first->io_timestamp;
644
645 nio = first;
646 do {
647 dio = nio;
648 nio = AVL_NEXT(t, dio);
649 ASSERT3U(dio->io_type, ==, aio->io_type);
650
651 if (dio->io_flags & ZIO_FLAG_NODATA) {
652 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
653 abd_zero_off(aio->io_abd,
654 dio->io_offset - aio->io_offset, dio->io_size);
655 } else if (dio->io_type == ZIO_TYPE_WRITE) {
656 abd_copy_off(aio->io_abd, dio->io_abd,
657 dio->io_offset - aio->io_offset, 0, dio->io_size);
658 }
659
660 zio_add_child(dio, aio);
661 vdev_queue_io_remove(vq, dio);
662 } while (dio != last);
663
664 /*
665 * We need to drop the vdev queue's lock to avoid a deadlock that we
666 * could encounter since this I/O will complete immediately.
667 */
668 mutex_exit(&vq->vq_lock);
669 while ((dio = zio_walk_parents(aio, &zl)) != NULL) {
670 zio_vdev_io_bypass(dio);
671 zio_execute(dio);
672 }
673 mutex_enter(&vq->vq_lock);
674
675 return (aio);
676 }
677
678 static zio_t *
679 vdev_queue_io_to_issue(vdev_queue_t *vq)
680 {
681 zio_t *zio, *aio;
682 zio_priority_t p;
683 avl_index_t idx;
684 avl_tree_t *tree;
685 zio_t search;
686
687 again:
688 ASSERT(MUTEX_HELD(&vq->vq_lock));
689
690 p = vdev_queue_class_to_issue(vq);
691
692 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
693 /* No eligible queued i/os */
694 return (NULL);
695 }
696
697 /*
698 * For LBA-ordered queues (async / scrub / initializing), issue the
699 * i/o which follows the most recently issued i/o in LBA (offset) order.
700 *
701 * For FIFO queues (sync), issue the i/o with the lowest timestamp.
702 */
703 tree = vdev_queue_class_tree(vq, p);
704 search.io_timestamp = 0;
705 search.io_offset = vq->vq_last_offset - 1;
706 VERIFY3P(avl_find(tree, &search, &idx), ==, NULL);
707 zio = avl_nearest(tree, idx, AVL_AFTER);
708 if (zio == NULL)
709 zio = avl_first(tree);
710 ASSERT3U(zio->io_priority, ==, p);
711
712 aio = vdev_queue_aggregate(vq, zio);
713 if (aio != NULL)
714 zio = aio;
715 else
716 vdev_queue_io_remove(vq, zio);
717
718 /*
719 * If the I/O is or was optional and therefore has no data, we need to
720 * simply discard it. We need to drop the vdev queue's lock to avoid a
721 * deadlock that we could encounter since this I/O will complete
722 * immediately.
723 */
724 if (zio->io_flags & ZIO_FLAG_NODATA) {
725 mutex_exit(&vq->vq_lock);
726 zio_vdev_io_bypass(zio);
727 zio_execute(zio);
728 mutex_enter(&vq->vq_lock);
729 goto again;
730 }
731
732 vdev_queue_pending_add(vq, zio);
733 vq->vq_last_offset = zio->io_offset + zio->io_size;
734
735 return (zio);
736 }
737
738 zio_t *
739 vdev_queue_io(zio_t *zio)
740 {
741 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
742 zio_t *nio;
743
744 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
745 return (zio);
746
747 /*
748 * Children i/os inherent their parent's priority, which might
749 * not match the child's i/o type. Fix it up here.
750 */
751 if (zio->io_type == ZIO_TYPE_READ) {
752 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
753 zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
754 zio->io_priority != ZIO_PRIORITY_SCRUB &&
755 zio->io_priority != ZIO_PRIORITY_REMOVAL &&
756 zio->io_priority != ZIO_PRIORITY_INITIALIZING)
757 zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
758 } else {
759 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
760 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
761 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE &&
762 zio->io_priority != ZIO_PRIORITY_REMOVAL &&
763 zio->io_priority != ZIO_PRIORITY_INITIALIZING)
764 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
765 }
766
767 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
768
769 mutex_enter(&vq->vq_lock);
770 zio->io_timestamp = gethrtime();
771 vdev_queue_io_add(vq, zio);
772 nio = vdev_queue_io_to_issue(vq);
773 mutex_exit(&vq->vq_lock);
774
775 if (nio == NULL)
776 return (NULL);
777
778 if (nio->io_done == vdev_queue_agg_io_done) {
779 zio_nowait(nio);
780 return (NULL);
781 }
782
783 return (nio);
784 }
785
786 void
787 vdev_queue_io_done(zio_t *zio)
788 {
789 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
790 zio_t *nio;
791
792 mutex_enter(&vq->vq_lock);
793
794 vdev_queue_pending_remove(vq, zio);
795
796 vq->vq_io_complete_ts = gethrtime();
797
798 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
799 mutex_exit(&vq->vq_lock);
800 if (nio->io_done == vdev_queue_agg_io_done) {
801 zio_nowait(nio);
802 } else {
803 zio_vdev_io_reissue(nio);
804 zio_execute(nio);
805 }
806 mutex_enter(&vq->vq_lock);
807 }
808
809 mutex_exit(&vq->vq_lock);
810 }
811
812 void
813 vdev_queue_change_io_priority(zio_t *zio, zio_priority_t priority)
814 {
815 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
816 avl_tree_t *tree;
817
818 /*
819 * ZIO_PRIORITY_NOW is used by the vdev cache code and the aggregate zio
820 * code to issue IOs without adding them to the vdev queue. In this
821 * case, the zio is already going to be issued as quickly as possible
822 * and so it doesn't need any reprioitization to help.
823 */
824 if (zio->io_priority == ZIO_PRIORITY_NOW)
825 return;
826
827 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
828 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
829
830 if (zio->io_type == ZIO_TYPE_READ) {
831 if (priority != ZIO_PRIORITY_SYNC_READ &&
832 priority != ZIO_PRIORITY_ASYNC_READ &&
833 priority != ZIO_PRIORITY_SCRUB)
834 priority = ZIO_PRIORITY_ASYNC_READ;
835 } else {
836 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
837 if (priority != ZIO_PRIORITY_SYNC_WRITE &&
838 priority != ZIO_PRIORITY_ASYNC_WRITE)
839 priority = ZIO_PRIORITY_ASYNC_WRITE;
840 }
841
842 mutex_enter(&vq->vq_lock);
843
844 /*
845 * If the zio is in none of the queues we can simply change
846 * the priority. If the zio is waiting to be submitted we must
847 * remove it from the queue and re-insert it with the new priority.
848 * Otherwise, the zio is currently active and we cannot change its
849 * priority.
850 */
851 tree = vdev_queue_class_tree(vq, zio->io_priority);
852 if (avl_find(tree, zio, NULL) == zio) {
853 spa_t *spa = zio->io_spa;
854 zio_priority_t oldpri = zio->io_priority;
855
856 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
857 zio->io_priority = priority;
858 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
859
860 mutex_enter(&spa->spa_iokstat_lock);
861 ASSERT3U(spa->spa_queue_stats[oldpri].spa_queued, >, 0);
862 spa->spa_queue_stats[oldpri].spa_queued--;
863 spa->spa_queue_stats[zio->io_priority].spa_queued++;
864 mutex_exit(&spa->spa_iokstat_lock);
865 } else if (avl_find(&vq->vq_active_tree, zio, NULL) != zio) {
866 zio->io_priority = priority;
867 }
868
869 mutex_exit(&vq->vq_lock);
870 }
871
872 /*
873 * As these two methods are only used for load calculations we're not
874 * concerned if we get an incorrect value on 32bit platforms due to lack of
875 * vq_lock mutex use here, instead we prefer to keep it lock free for
876 * performance.
877 */
878 int
879 vdev_queue_length(vdev_t *vd)
880 {
881 return (avl_numnodes(&vd->vdev_queue.vq_active_tree));
882 }
883
884 uint64_t
885 vdev_queue_last_offset(vdev_t *vd)
886 {
887 return (vd->vdev_queue.vq_last_offset);
888 }