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, 2014 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
39 /*
40 * ZFS I/O Scheduler
41 * ---------------
42 *
43 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The
44 * I/O scheduler determines when and in what order those operations are
45 * issued. The I/O scheduler divides operations into six I/O classes
46 * prioritized in the following order: sync read, sync write, async read,
47 * async write, scrub/resilver and trim. Each queue defines the minimum and
48 * maximum number of concurrent operations that may be issued to the device.
49 * In addition, the device has an aggregate maximum. Note that the sum of the
50 * per-queue minimums must not exceed the aggregate maximum, and if the
51 * aggregate maximum is equal to or greater than the sum of the per-queue
52 * maximums, the per-queue minimum has no effect.
53 *
54 * For many physical devices, throughput increases with the number of
55 * concurrent operations, but latency typically suffers. Further, physical
56 * devices typically have a limit at which more concurrent operations have no
57 * effect on throughput or can actually cause it to decrease.
58 *
59 * The scheduler selects the next operation to issue by first looking for an
60 * I/O class whose minimum has not been satisfied. Once all are satisfied and
61 * the aggregate maximum has not been hit, the scheduler looks for classes
62 * whose maximum has not been satisfied. Iteration through the I/O classes is
63 * done in the order specified above. No further operations are issued if the
64 * aggregate maximum number of concurrent operations has been hit or if there
65 * are no operations queued for an I/O class that has not hit its maximum.
66 * Every time an I/O is queued or an operation completes, the I/O scheduler
67 * looks for new operations to issue.
68 *
69 * All I/O classes have a fixed maximum number of outstanding operations
70 * except for the async write class. Asynchronous writes represent the data
71 * that is committed to stable storage during the syncing stage for
72 * transaction groups (see txg.c). Transaction groups enter the syncing state
73 * periodically so the number of queued async writes will quickly burst up and
74 * then bleed down to zero. Rather than servicing them as quickly as possible,
75 * the I/O scheduler changes the maximum number of active async write I/Os
76 * according to the amount of dirty data in the pool (see dsl_pool.c). Since
77 * both throughput and latency typically increase with the number of
78 * concurrent operations issued to physical devices, reducing the burstiness
79 * in the number of concurrent operations also stabilizes the response time of
80 * operations from other -- and in particular synchronous -- queues. In broad
81 * strokes, the I/O scheduler will issue more concurrent operations from the
82 * async write queue as there's more dirty data in the pool.
83 *
84 * Async Writes
85 *
86 * The number of concurrent operations issued for the async write I/O class
87 * follows a piece-wise linear function defined by a few adjustable points.
88 *
89 * | o---------| <-- zfs_vdev_async_write_max_active
90 * ^ | /^ |
91 * | | / | |
92 * active | / | |
93 * I/O | / | |
94 * count | / | |
95 * | / | |
96 * |------------o | | <-- zfs_vdev_async_write_min_active
97 * 0|____________^______|_________|
98 * 0% | | 100% of zfs_dirty_data_max
99 * | |
100 * | `-- zfs_vdev_async_write_active_max_dirty_percent
101 * `--------- zfs_vdev_async_write_active_min_dirty_percent
102 *
103 * Until the amount of dirty data exceeds a minimum percentage of the dirty
104 * data allowed in the pool, the I/O scheduler will limit the number of
105 * concurrent operations to the minimum. As that threshold is crossed, the
106 * number of concurrent operations issued increases linearly to the maximum at
107 * the specified maximum percentage of the dirty data allowed in the pool.
108 *
109 * Ideally, the amount of dirty data on a busy pool will stay in the sloped
110 * part of the function between zfs_vdev_async_write_active_min_dirty_percent
111 * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the
112 * maximum percentage, this indicates that the rate of incoming data is
113 * greater than the rate that the backend storage can handle. In this case, we
114 * must further throttle incoming writes (see dmu_tx_delay() for details).
115 */
116
117 /*
118 * The maximum number of I/Os active to each device. Ideally, this will be >=
119 * the sum of each queue's max_active. It must be at least the sum of each
120 * queue's min_active.
121 */
122 uint32_t zfs_vdev_max_active = 1000;
123
124 /*
125 * Per-queue limits on the number of I/Os active to each device. If the
126 * sum of the queue's max_active is < zfs_vdev_max_active, then the
127 * min_active comes into play. We will send min_active from each queue,
128 * and then select from queues in the order defined by zio_priority_t.
129 *
130 * In general, smaller max_active's will lead to lower latency of synchronous
131 * operations. Larger max_active's may lead to higher overall throughput,
132 * depending on underlying storage.
133 *
134 * The ratio of the queues' max_actives determines the balance of performance
135 * between reads, writes, and scrubs. E.g., increasing
136 * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete
137 * more quickly, but reads and writes to have higher latency and lower
138 * throughput.
139 */
140 uint32_t zfs_vdev_sync_read_min_active = 10;
141 uint32_t zfs_vdev_sync_read_max_active = 10;
142 uint32_t zfs_vdev_sync_write_min_active = 10;
143 uint32_t zfs_vdev_sync_write_max_active = 10;
144 uint32_t zfs_vdev_async_read_min_active = 1;
145 uint32_t zfs_vdev_async_read_max_active = 3;
146 uint32_t zfs_vdev_async_write_min_active = 1;
147 uint32_t zfs_vdev_async_write_max_active = 10;
148 uint32_t zfs_vdev_scrub_min_active = 1;
149 uint32_t zfs_vdev_scrub_max_active = 2;
150 uint32_t zfs_vdev_trim_min_active = 1;
151 /*
152 * TRIM max active is large in comparison to the other values due to the fact
153 * that TRIM IOs are coalesced at the device layer. This value is set such
154 * that a typical SSD can process the queued IOs in a single request.
155 */
156 uint32_t zfs_vdev_trim_max_active = 64;
157
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 */
166 int zfs_vdev_async_write_active_min_dirty_percent = 30;
167 int 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 */
175 int zfs_vdev_aggregation_limit = SPA_OLD_MAXBLOCKSIZE;
176 int zfs_vdev_read_gap_limit = 32 << 10;
177 int 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
190 int zfs_vdev_queue_depth_pct = 1000;
191 #else
192 int zfs_vdev_queue_depth_pct = 300;
193 #endif
194
195
196 #ifdef __FreeBSD__
197 #ifdef _KERNEL
198 SYSCTL_DECL(_vfs_zfs_vdev);
199
200 static int sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS);
201 SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_min_dirty_percent,
202 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
203 sysctl_zfs_async_write_active_min_dirty_percent, "I",
204 "Percentage of async write dirty data below which "
205 "async_write_min_active is used.");
206
207 static int sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS);
208 SYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_max_dirty_percent,
209 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
210 sysctl_zfs_async_write_active_max_dirty_percent, "I",
211 "Percentage of async write dirty data above which "
212 "async_write_max_active is used.");
213
214 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RWTUN,
215 &zfs_vdev_max_active, 0,
216 "The maximum number of I/Os of all types active for each device.");
217
218 #define ZFS_VDEV_QUEUE_KNOB_MIN(name) \
219 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, CTLFLAG_RWTUN,\
220 &zfs_vdev_ ## name ## _min_active, 0, \
221 "Initial number of I/O requests of type " #name \
222 " active for each device");
223
224 #define ZFS_VDEV_QUEUE_KNOB_MAX(name) \
225 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, CTLFLAG_RWTUN,\
226 &zfs_vdev_ ## name ## _max_active, 0, \
227 "Maximum number of I/O requests of type " #name \
228 " active for each device");
229
230 ZFS_VDEV_QUEUE_KNOB_MIN(sync_read);
231 ZFS_VDEV_QUEUE_KNOB_MAX(sync_read);
232 ZFS_VDEV_QUEUE_KNOB_MIN(sync_write);
233 ZFS_VDEV_QUEUE_KNOB_MAX(sync_write);
234 ZFS_VDEV_QUEUE_KNOB_MIN(async_read);
235 ZFS_VDEV_QUEUE_KNOB_MAX(async_read);
236 ZFS_VDEV_QUEUE_KNOB_MIN(async_write);
237 ZFS_VDEV_QUEUE_KNOB_MAX(async_write);
238 ZFS_VDEV_QUEUE_KNOB_MIN(scrub);
239 ZFS_VDEV_QUEUE_KNOB_MAX(scrub);
240 ZFS_VDEV_QUEUE_KNOB_MIN(trim);
241 ZFS_VDEV_QUEUE_KNOB_MAX(trim);
242
243 #undef ZFS_VDEV_QUEUE_KNOB
244
245 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RWTUN,
246 &zfs_vdev_aggregation_limit, 0,
247 "I/O requests are aggregated up to this size");
248 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RWTUN,
249 &zfs_vdev_read_gap_limit, 0,
250 "Acceptable gap between two reads being aggregated");
251 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RWTUN,
252 &zfs_vdev_write_gap_limit, 0,
253 "Acceptable gap between two writes being aggregated");
254 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, queue_depth_pct, CTLFLAG_RWTUN,
255 &zfs_vdev_queue_depth_pct, 0,
256 "Queue depth percentage for each top-level");
257
258 static int
sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS)259 sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS)
260 {
261 int val, err;
262
263 val = zfs_vdev_async_write_active_min_dirty_percent;
264 err = sysctl_handle_int(oidp, &val, 0, req);
265 if (err != 0 || req->newptr == NULL)
266 return (err);
267
268 if (val < 0 || val > 100 ||
269 val >= zfs_vdev_async_write_active_max_dirty_percent)
270 return (EINVAL);
271
272 zfs_vdev_async_write_active_min_dirty_percent = val;
273
274 return (0);
275 }
276
277 static int
sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS)278 sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS)
279 {
280 int val, err;
281
282 val = zfs_vdev_async_write_active_max_dirty_percent;
283 err = sysctl_handle_int(oidp, &val, 0, req);
284 if (err != 0 || req->newptr == NULL)
285 return (err);
286
287 if (val < 0 || val > 100 ||
288 val <= zfs_vdev_async_write_active_min_dirty_percent)
289 return (EINVAL);
290
291 zfs_vdev_async_write_active_max_dirty_percent = val;
292
293 return (0);
294 }
295 #endif
296 #endif
297
298 int
vdev_queue_offset_compare(const void * x1,const void * x2)299 vdev_queue_offset_compare(const void *x1, const void *x2)
300 {
301 const zio_t *z1 = x1;
302 const zio_t *z2 = x2;
303
304 if (z1->io_offset < z2->io_offset)
305 return (-1);
306 if (z1->io_offset > z2->io_offset)
307 return (1);
308
309 if (z1 < z2)
310 return (-1);
311 if (z1 > z2)
312 return (1);
313
314 return (0);
315 }
316
317 static inline avl_tree_t *
vdev_queue_class_tree(vdev_queue_t * vq,zio_priority_t p)318 vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p)
319 {
320 return (&vq->vq_class[p].vqc_queued_tree);
321 }
322
323 static inline avl_tree_t *
vdev_queue_type_tree(vdev_queue_t * vq,zio_type_t t)324 vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t)
325 {
326 if (t == ZIO_TYPE_READ)
327 return (&vq->vq_read_offset_tree);
328 else if (t == ZIO_TYPE_WRITE)
329 return (&vq->vq_write_offset_tree);
330 else
331 return (NULL);
332 }
333
334 int
vdev_queue_timestamp_compare(const void * x1,const void * x2)335 vdev_queue_timestamp_compare(const void *x1, const void *x2)
336 {
337 const zio_t *z1 = x1;
338 const zio_t *z2 = x2;
339
340 if (z1->io_timestamp < z2->io_timestamp)
341 return (-1);
342 if (z1->io_timestamp > z2->io_timestamp)
343 return (1);
344
345 if (z1->io_offset < z2->io_offset)
346 return (-1);
347 if (z1->io_offset > z2->io_offset)
348 return (1);
349
350 if (z1 < z2)
351 return (-1);
352 if (z1 > z2)
353 return (1);
354
355 return (0);
356 }
357
358 void
vdev_queue_init(vdev_t * vd)359 vdev_queue_init(vdev_t *vd)
360 {
361 vdev_queue_t *vq = &vd->vdev_queue;
362
363 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
364 vq->vq_vdev = vd;
365
366 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
367 sizeof (zio_t), offsetof(struct zio, io_queue_node));
368 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ),
369 vdev_queue_offset_compare, sizeof (zio_t),
370 offsetof(struct zio, io_offset_node));
371 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE),
372 vdev_queue_offset_compare, sizeof (zio_t),
373 offsetof(struct zio, io_offset_node));
374
375 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
376 int (*compfn) (const void *, const void *);
377
378 /*
379 * The synchronous i/o queues are dispatched in FIFO rather
380 * than LBA order. This provides more consistent latency for
381 * these i/os.
382 */
383 if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE)
384 compfn = vdev_queue_timestamp_compare;
385 else
386 compfn = vdev_queue_offset_compare;
387
388 avl_create(vdev_queue_class_tree(vq, p), compfn,
389 sizeof (zio_t), offsetof(struct zio, io_queue_node));
390 }
391
392 vq->vq_lastoffset = 0;
393 }
394
395 void
vdev_queue_fini(vdev_t * vd)396 vdev_queue_fini(vdev_t *vd)
397 {
398 vdev_queue_t *vq = &vd->vdev_queue;
399
400 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
401 avl_destroy(vdev_queue_class_tree(vq, p));
402 avl_destroy(&vq->vq_active_tree);
403 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
404 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE));
405
406 mutex_destroy(&vq->vq_lock);
407 }
408
409 static void
vdev_queue_io_add(vdev_queue_t * vq,zio_t * zio)410 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
411 {
412 spa_t *spa = zio->io_spa;
413 avl_tree_t *qtt;
414
415 ASSERT(MUTEX_HELD(&vq->vq_lock));
416 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
417 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
418 qtt = vdev_queue_type_tree(vq, zio->io_type);
419 if (qtt)
420 avl_add(qtt, zio);
421
422 #ifdef illumos
423 mutex_enter(&spa->spa_iokstat_lock);
424 spa->spa_queue_stats[zio->io_priority].spa_queued++;
425 if (spa->spa_iokstat != NULL)
426 kstat_waitq_enter(spa->spa_iokstat->ks_data);
427 mutex_exit(&spa->spa_iokstat_lock);
428 #endif
429 }
430
431 static void
vdev_queue_io_remove(vdev_queue_t * vq,zio_t * zio)432 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
433 {
434 spa_t *spa = zio->io_spa;
435 avl_tree_t *qtt;
436
437 ASSERT(MUTEX_HELD(&vq->vq_lock));
438 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
439 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
440 qtt = vdev_queue_type_tree(vq, zio->io_type);
441 if (qtt)
442 avl_remove(qtt, zio);
443
444 #ifdef illumos
445 mutex_enter(&spa->spa_iokstat_lock);
446 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
447 spa->spa_queue_stats[zio->io_priority].spa_queued--;
448 if (spa->spa_iokstat != NULL)
449 kstat_waitq_exit(spa->spa_iokstat->ks_data);
450 mutex_exit(&spa->spa_iokstat_lock);
451 #endif
452 }
453
454 static void
vdev_queue_pending_add(vdev_queue_t * vq,zio_t * zio)455 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
456 {
457 spa_t *spa = zio->io_spa;
458 ASSERT(MUTEX_HELD(&vq->vq_lock));
459 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
460 vq->vq_class[zio->io_priority].vqc_active++;
461 avl_add(&vq->vq_active_tree, zio);
462
463 #ifdef illumos
464 mutex_enter(&spa->spa_iokstat_lock);
465 spa->spa_queue_stats[zio->io_priority].spa_active++;
466 if (spa->spa_iokstat != NULL)
467 kstat_runq_enter(spa->spa_iokstat->ks_data);
468 mutex_exit(&spa->spa_iokstat_lock);
469 #endif
470 }
471
472 static void
vdev_queue_pending_remove(vdev_queue_t * vq,zio_t * zio)473 vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
474 {
475 spa_t *spa = zio->io_spa;
476 ASSERT(MUTEX_HELD(&vq->vq_lock));
477 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
478 vq->vq_class[zio->io_priority].vqc_active--;
479 avl_remove(&vq->vq_active_tree, zio);
480
481 #ifdef illumos
482 mutex_enter(&spa->spa_iokstat_lock);
483 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
484 spa->spa_queue_stats[zio->io_priority].spa_active--;
485 if (spa->spa_iokstat != NULL) {
486 kstat_io_t *ksio = spa->spa_iokstat->ks_data;
487
488 kstat_runq_exit(spa->spa_iokstat->ks_data);
489 if (zio->io_type == ZIO_TYPE_READ) {
490 ksio->reads++;
491 ksio->nread += zio->io_size;
492 } else if (zio->io_type == ZIO_TYPE_WRITE) {
493 ksio->writes++;
494 ksio->nwritten += zio->io_size;
495 }
496 }
497 mutex_exit(&spa->spa_iokstat_lock);
498 #endif
499 }
500
501 static void
vdev_queue_agg_io_done(zio_t * aio)502 vdev_queue_agg_io_done(zio_t *aio)
503 {
504 if (aio->io_type == ZIO_TYPE_READ) {
505 zio_t *pio;
506 zio_link_t *zl = NULL;
507 while ((pio = zio_walk_parents(aio, &zl)) != NULL) {
508 bcopy((char *)aio->io_data + (pio->io_offset -
509 aio->io_offset), pio->io_data, pio->io_size);
510 }
511 }
512
513 zio_buf_free(aio->io_data, aio->io_size);
514 }
515
516 static int
vdev_queue_class_min_active(zio_priority_t p)517 vdev_queue_class_min_active(zio_priority_t p)
518 {
519 switch (p) {
520 case ZIO_PRIORITY_SYNC_READ:
521 return (zfs_vdev_sync_read_min_active);
522 case ZIO_PRIORITY_SYNC_WRITE:
523 return (zfs_vdev_sync_write_min_active);
524 case ZIO_PRIORITY_ASYNC_READ:
525 return (zfs_vdev_async_read_min_active);
526 case ZIO_PRIORITY_ASYNC_WRITE:
527 return (zfs_vdev_async_write_min_active);
528 case ZIO_PRIORITY_SCRUB:
529 return (zfs_vdev_scrub_min_active);
530 case ZIO_PRIORITY_TRIM:
531 return (zfs_vdev_trim_min_active);
532 default:
533 panic("invalid priority %u", p);
534 return (0);
535 }
536 }
537
538 static __noinline int
vdev_queue_max_async_writes(spa_t * spa)539 vdev_queue_max_async_writes(spa_t *spa)
540 {
541 int writes;
542 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total;
543 uint64_t min_bytes = zfs_dirty_data_max *
544 zfs_vdev_async_write_active_min_dirty_percent / 100;
545 uint64_t max_bytes = zfs_dirty_data_max *
546 zfs_vdev_async_write_active_max_dirty_percent / 100;
547
548 /*
549 * Sync tasks correspond to interactive user actions. To reduce the
550 * execution time of those actions we push data out as fast as possible.
551 */
552 if (spa_has_pending_synctask(spa)) {
553 return (zfs_vdev_async_write_max_active);
554 }
555
556 if (dirty < min_bytes)
557 return (zfs_vdev_async_write_min_active);
558 if (dirty > max_bytes)
559 return (zfs_vdev_async_write_max_active);
560
561 /*
562 * linear interpolation:
563 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
564 * move right by min_bytes
565 * move up by min_writes
566 */
567 writes = (dirty - min_bytes) *
568 (zfs_vdev_async_write_max_active -
569 zfs_vdev_async_write_min_active) /
570 (max_bytes - min_bytes) +
571 zfs_vdev_async_write_min_active;
572 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
573 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
574 return (writes);
575 }
576
577 static int
vdev_queue_class_max_active(spa_t * spa,zio_priority_t p)578 vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
579 {
580 switch (p) {
581 case ZIO_PRIORITY_SYNC_READ:
582 return (zfs_vdev_sync_read_max_active);
583 case ZIO_PRIORITY_SYNC_WRITE:
584 return (zfs_vdev_sync_write_max_active);
585 case ZIO_PRIORITY_ASYNC_READ:
586 return (zfs_vdev_async_read_max_active);
587 case ZIO_PRIORITY_ASYNC_WRITE:
588 return (vdev_queue_max_async_writes(spa));
589 case ZIO_PRIORITY_SCRUB:
590 return (zfs_vdev_scrub_max_active);
591 case ZIO_PRIORITY_TRIM:
592 return (zfs_vdev_trim_max_active);
593 default:
594 panic("invalid priority %u", p);
595 return (0);
596 }
597 }
598
599 /*
600 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
601 * there is no eligible class.
602 */
603 static zio_priority_t
vdev_queue_class_to_issue(vdev_queue_t * vq)604 vdev_queue_class_to_issue(vdev_queue_t *vq)
605 {
606 spa_t *spa = vq->vq_vdev->vdev_spa;
607 zio_priority_t p;
608
609 ASSERT(MUTEX_HELD(&vq->vq_lock));
610
611 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
612 return (ZIO_PRIORITY_NUM_QUEUEABLE);
613
614 /* find a queue that has not reached its minimum # outstanding i/os */
615 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
616 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
617 vq->vq_class[p].vqc_active <
618 vdev_queue_class_min_active(p))
619 return (p);
620 }
621
622 /*
623 * If we haven't found a queue, look for one that hasn't reached its
624 * maximum # outstanding i/os.
625 */
626 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
627 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
628 vq->vq_class[p].vqc_active <
629 vdev_queue_class_max_active(spa, p))
630 return (p);
631 }
632
633 /* No eligible queued i/os */
634 return (ZIO_PRIORITY_NUM_QUEUEABLE);
635 }
636
637 /*
638 * Compute the range spanned by two i/os, which is the endpoint of the last
639 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
640 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
641 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
642 */
643 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
644 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
645
646 static zio_t *
vdev_queue_aggregate(vdev_queue_t * vq,zio_t * zio)647 vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
648 {
649 zio_t *first, *last, *aio, *dio, *mandatory, *nio;
650 void *abuf;
651 uint64_t maxgap = 0;
652 uint64_t size;
653 boolean_t stretch;
654 avl_tree_t *t;
655 enum zio_flag flags;
656
657 ASSERT(MUTEX_HELD(&vq->vq_lock));
658
659 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
660 return (NULL);
661
662 first = last = zio;
663
664 if (zio->io_type == ZIO_TYPE_READ)
665 maxgap = zfs_vdev_read_gap_limit;
666
667 /*
668 * We can aggregate I/Os that are sufficiently adjacent and of
669 * the same flavor, as expressed by the AGG_INHERIT flags.
670 * The latter requirement is necessary so that certain
671 * attributes of the I/O, such as whether it's a normal I/O
672 * or a scrub/resilver, can be preserved in the aggregate.
673 * We can include optional I/Os, but don't allow them
674 * to begin a range as they add no benefit in that situation.
675 */
676
677 /*
678 * We keep track of the last non-optional I/O.
679 */
680 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
681
682 /*
683 * Walk backwards through sufficiently contiguous I/Os
684 * recording the last non-option I/O.
685 */
686 flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
687 t = vdev_queue_type_tree(vq, zio->io_type);
688 while (t != NULL && (dio = AVL_PREV(t, first)) != NULL &&
689 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
690 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
691 IO_GAP(dio, first) <= maxgap) {
692 first = dio;
693 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
694 mandatory = first;
695 }
696
697 /*
698 * Skip any initial optional I/Os.
699 */
700 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
701 first = AVL_NEXT(t, first);
702 ASSERT(first != NULL);
703 }
704
705 /*
706 * Walk forward through sufficiently contiguous I/Os.
707 */
708 while ((dio = AVL_NEXT(t, last)) != NULL &&
709 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
710 IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit &&
711 IO_GAP(last, dio) <= maxgap) {
712 last = dio;
713 if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
714 mandatory = last;
715 }
716
717 /*
718 * Now that we've established the range of the I/O aggregation
719 * we must decide what to do with trailing optional I/Os.
720 * For reads, there's nothing to do. While we are unable to
721 * aggregate further, it's possible that a trailing optional
722 * I/O would allow the underlying device to aggregate with
723 * subsequent I/Os. We must therefore determine if the next
724 * non-optional I/O is close enough to make aggregation
725 * worthwhile.
726 */
727 stretch = B_FALSE;
728 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
729 zio_t *nio = last;
730 while ((dio = AVL_NEXT(t, nio)) != NULL &&
731 IO_GAP(nio, dio) == 0 &&
732 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
733 nio = dio;
734 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
735 stretch = B_TRUE;
736 break;
737 }
738 }
739 }
740
741 if (stretch) {
742 /* This may be a no-op. */
743 dio = AVL_NEXT(t, last);
744 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
745 } else {
746 while (last != mandatory && last != first) {
747 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
748 last = AVL_PREV(t, last);
749 ASSERT(last != NULL);
750 }
751 }
752
753 if (first == last)
754 return (NULL);
755
756 size = IO_SPAN(first, last);
757 ASSERT3U(size, <=, zfs_vdev_aggregation_limit);
758
759 abuf = zio_buf_alloc_nowait(size);
760 if (abuf == NULL)
761 return (NULL);
762
763 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
764 abuf, size, first->io_type, zio->io_priority,
765 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
766 vdev_queue_agg_io_done, NULL);
767 aio->io_timestamp = first->io_timestamp;
768
769 nio = first;
770 do {
771 dio = nio;
772 nio = AVL_NEXT(t, dio);
773 ASSERT3U(dio->io_type, ==, aio->io_type);
774
775 if (dio->io_flags & ZIO_FLAG_NODATA) {
776 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
777 bzero((char *)aio->io_data + (dio->io_offset -
778 aio->io_offset), dio->io_size);
779 } else if (dio->io_type == ZIO_TYPE_WRITE) {
780 bcopy(dio->io_data, (char *)aio->io_data +
781 (dio->io_offset - aio->io_offset),
782 dio->io_size);
783 }
784
785 zio_add_child(dio, aio);
786 vdev_queue_io_remove(vq, dio);
787 zio_vdev_io_bypass(dio);
788 zio_execute(dio);
789 } while (dio != last);
790
791 return (aio);
792 }
793
794 static zio_t *
vdev_queue_io_to_issue(vdev_queue_t * vq)795 vdev_queue_io_to_issue(vdev_queue_t *vq)
796 {
797 zio_t *zio, *aio;
798 zio_priority_t p;
799 avl_index_t idx;
800 avl_tree_t *tree;
801 zio_t *search;
802
803 again:
804 ASSERT(MUTEX_HELD(&vq->vq_lock));
805
806 p = vdev_queue_class_to_issue(vq);
807
808 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
809 /* No eligible queued i/os */
810 return (NULL);
811 }
812
813 /*
814 * For LBA-ordered queues (async / scrub), issue the i/o which follows
815 * the most recently issued i/o in LBA (offset) order.
816 *
817 * For FIFO queues (sync), issue the i/o with the lowest timestamp.
818 */
819 tree = vdev_queue_class_tree(vq, p);
820 search = kmem_zalloc(sizeof (*search), KM_NOSLEEP);
821 if (search) {
822 search->io_offset = vq->vq_last_offset + 1;
823 VERIFY3P(avl_find(tree, search, &idx), ==, NULL);
824 kmem_free(search, sizeof (*search));
825 zio = avl_nearest(tree, idx, AVL_AFTER);
826 } else {
827 /* Can't find nearest, fallback to first */
828 zio = NULL;
829 }
830 if (zio == NULL)
831 zio = avl_first(tree);
832 ASSERT3U(zio->io_priority, ==, p);
833
834 aio = vdev_queue_aggregate(vq, zio);
835 if (aio != NULL)
836 zio = aio;
837 else
838 vdev_queue_io_remove(vq, zio);
839
840 /*
841 * If the I/O is or was optional and therefore has no data, we need to
842 * simply discard it. We need to drop the vdev queue's lock to avoid a
843 * deadlock that we could encounter since this I/O will complete
844 * immediately.
845 */
846 if (zio->io_flags & ZIO_FLAG_NODATA) {
847 mutex_exit(&vq->vq_lock);
848 zio_vdev_io_bypass(zio);
849 zio_execute(zio);
850 mutex_enter(&vq->vq_lock);
851 goto again;
852 }
853
854 vdev_queue_pending_add(vq, zio);
855 vq->vq_last_offset = zio->io_offset;
856
857 return (zio);
858 }
859
860 zio_t *
vdev_queue_io(zio_t * zio)861 vdev_queue_io(zio_t *zio)
862 {
863 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
864 zio_t *nio;
865
866 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
867 return (zio);
868
869 /*
870 * Children i/os inherent their parent's priority, which might
871 * not match the child's i/o type. Fix it up here.
872 */
873 if (zio->io_type == ZIO_TYPE_READ) {
874 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
875 zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
876 zio->io_priority != ZIO_PRIORITY_SCRUB)
877 zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
878 } else if (zio->io_type == ZIO_TYPE_WRITE) {
879 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
880 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE)
881 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
882 } else {
883 ASSERT(zio->io_type == ZIO_TYPE_FREE);
884 zio->io_priority = ZIO_PRIORITY_TRIM;
885 }
886
887 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
888
889 mutex_enter(&vq->vq_lock);
890 zio->io_timestamp = gethrtime();
891 vdev_queue_io_add(vq, zio);
892 nio = vdev_queue_io_to_issue(vq);
893 mutex_exit(&vq->vq_lock);
894
895 if (nio == NULL)
896 return (NULL);
897
898 if (nio->io_done == vdev_queue_agg_io_done) {
899 zio_nowait(nio);
900 return (NULL);
901 }
902
903 return (nio);
904 }
905
906 void
vdev_queue_io_done(zio_t * zio)907 vdev_queue_io_done(zio_t *zio)
908 {
909 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
910 zio_t *nio;
911
912 mutex_enter(&vq->vq_lock);
913
914 vdev_queue_pending_remove(vq, zio);
915
916 vq->vq_io_complete_ts = gethrtime();
917
918 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
919 mutex_exit(&vq->vq_lock);
920 if (nio->io_done == vdev_queue_agg_io_done) {
921 zio_nowait(nio);
922 } else {
923 zio_vdev_io_reissue(nio);
924 zio_execute(nio);
925 }
926 mutex_enter(&vq->vq_lock);
927 }
928
929 mutex_exit(&vq->vq_lock);
930 }
931
932 /*
933 * As these three methods are only used for load calculations we're not concerned
934 * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex
935 * use here, instead we prefer to keep it lock free for performance.
936 */
937 int
vdev_queue_length(vdev_t * vd)938 vdev_queue_length(vdev_t *vd)
939 {
940 return (avl_numnodes(&vd->vdev_queue.vq_active_tree));
941 }
942
943 uint64_t
vdev_queue_lastoffset(vdev_t * vd)944 vdev_queue_lastoffset(vdev_t *vd)
945 {
946 return (vd->vdev_queue.vq_lastoffset);
947 }
948
949 void
vdev_queue_register_lastoffset(vdev_t * vd,zio_t * zio)950 vdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio)
951 {
952 vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size;
953 }
954