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 by Delphix. All rights reserved.
28 */
29
30 #include <sys/zfs_context.h>
31 #include <sys/vdev_impl.h>
32 #include <sys/spa_impl.h>
33 #include <sys/zio.h>
34 #include <sys/avl.h>
35
36 /*
37 * These tunables are for performance analysis.
38 */
39
40 /* The maximum number of I/Os concurrently pending to each device. */
41 int zfs_vdev_max_pending = 10;
42
43 /*
44 * The initial number of I/Os pending to each device, before it starts ramping
45 * up to zfs_vdev_max_pending.
46 */
47 int zfs_vdev_min_pending = 4;
48
49 /*
50 * The deadlines are grouped into buckets based on zfs_vdev_time_shift:
51 * deadline = pri + gethrtime() >> time_shift)
52 */
53 int zfs_vdev_time_shift = 29; /* each bucket is 0.537 seconds */
54
55 /* exponential I/O issue ramp-up rate */
56 int zfs_vdev_ramp_rate = 2;
57
58 /*
59 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
60 * For read I/Os, we also aggregate across small adjacency gaps; for writes
61 * we include spans of optional I/Os to aid aggregation at the disk even when
62 * they aren't able to help us aggregate at this level.
63 */
64 int zfs_vdev_aggregation_limit = SPA_MAXBLOCKSIZE;
65 int zfs_vdev_read_gap_limit = 32 << 10;
66 int zfs_vdev_write_gap_limit = 4 << 10;
67
68 /*
69 * Virtual device vector for disk I/O scheduling.
70 */
71 int
72 vdev_queue_deadline_compare(const void *x1, const void *x2)
73 {
74 const zio_t *z1 = x1;
75 const zio_t *z2 = x2;
76
77 if (z1->io_deadline < z2->io_deadline)
78 return (-1);
79 if (z1->io_deadline > z2->io_deadline)
80 return (1);
81
82 if (z1->io_offset < z2->io_offset)
83 return (-1);
84 if (z1->io_offset > z2->io_offset)
85 return (1);
86
87 if (z1 < z2)
88 return (-1);
89 if (z1 > z2)
90 return (1);
91
92 return (0);
93 }
94
95 int
96 vdev_queue_offset_compare(const void *x1, const void *x2)
97 {
98 const zio_t *z1 = x1;
99 const zio_t *z2 = x2;
100
101 if (z1->io_offset < z2->io_offset)
102 return (-1);
103 if (z1->io_offset > z2->io_offset)
104 return (1);
105
106 if (z1 < z2)
107 return (-1);
108 if (z1 > z2)
109 return (1);
110
111 return (0);
112 }
113
114 void
115 vdev_queue_init(vdev_t *vd)
116 {
117 vdev_queue_t *vq = &vd->vdev_queue;
118
119 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
120
121 avl_create(&vq->vq_deadline_tree, vdev_queue_deadline_compare,
122 sizeof (zio_t), offsetof(struct zio, io_deadline_node));
123
124 avl_create(&vq->vq_read_tree, vdev_queue_offset_compare,
125 sizeof (zio_t), offsetof(struct zio, io_offset_node));
126
127 avl_create(&vq->vq_write_tree, vdev_queue_offset_compare,
128 sizeof (zio_t), offsetof(struct zio, io_offset_node));
129
130 avl_create(&vq->vq_pending_tree, vdev_queue_offset_compare,
131 sizeof (zio_t), offsetof(struct zio, io_offset_node));
132 }
133
134 void
135 vdev_queue_fini(vdev_t *vd)
136 {
137 vdev_queue_t *vq = &vd->vdev_queue;
138
139 avl_destroy(&vq->vq_deadline_tree);
140 avl_destroy(&vq->vq_read_tree);
141 avl_destroy(&vq->vq_write_tree);
142 avl_destroy(&vq->vq_pending_tree);
143
144 mutex_destroy(&vq->vq_lock);
145 }
146
147 static void
148 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
149 {
150 spa_t *spa = zio->io_spa;
151 avl_add(&vq->vq_deadline_tree, zio);
152 avl_add(zio->io_vdev_tree, zio);
153
154 if (spa->spa_iokstat != NULL) {
155 mutex_enter(&spa->spa_iokstat_lock);
156 kstat_waitq_enter(spa->spa_iokstat->ks_data);
157 mutex_exit(&spa->spa_iokstat_lock);
158 }
159 }
160
161 static void
162 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
163 {
164 spa_t *spa = zio->io_spa;
165 avl_remove(&vq->vq_deadline_tree, zio);
166 avl_remove(zio->io_vdev_tree, zio);
167
168 if (spa->spa_iokstat != NULL) {
169 mutex_enter(&spa->spa_iokstat_lock);
170 kstat_waitq_exit(spa->spa_iokstat->ks_data);
171 mutex_exit(&spa->spa_iokstat_lock);
172 }
173 }
174
175 static void
176 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
177 {
178 spa_t *spa = zio->io_spa;
179 avl_add(&vq->vq_pending_tree, zio);
180 if (spa->spa_iokstat != NULL) {
181 mutex_enter(&spa->spa_iokstat_lock);
182 kstat_runq_enter(spa->spa_iokstat->ks_data);
183 mutex_exit(&spa->spa_iokstat_lock);
184 }
185 }
186
187 static void
188 vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
189 {
190 spa_t *spa = zio->io_spa;
191 avl_remove(&vq->vq_pending_tree, zio);
192 if (spa->spa_iokstat != NULL) {
193 kstat_io_t *ksio = spa->spa_iokstat->ks_data;
194
195 mutex_enter(&spa->spa_iokstat_lock);
196 kstat_runq_exit(spa->spa_iokstat->ks_data);
197 if (zio->io_type == ZIO_TYPE_READ) {
198 ksio->reads++;
199 ksio->nread += zio->io_size;
200 } else if (zio->io_type == ZIO_TYPE_WRITE) {
201 ksio->writes++;
202 ksio->nwritten += zio->io_size;
203 }
204 mutex_exit(&spa->spa_iokstat_lock);
205 }
206 }
207
208 static void
209 vdev_queue_agg_io_done(zio_t *aio)
210 {
211 zio_t *pio;
212
213 while ((pio = zio_walk_parents(aio)) != NULL)
214 if (aio->io_type == ZIO_TYPE_READ)
215 bcopy((char *)aio->io_data + (pio->io_offset -
216 aio->io_offset), pio->io_data, pio->io_size);
217
218 zio_buf_free(aio->io_data, aio->io_size);
219 }
220
221 /*
222 * Compute the range spanned by two i/os, which is the endpoint of the last
223 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
224 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
225 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
226 */
227 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
228 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
229
230 static zio_t *
231 vdev_queue_io_to_issue(vdev_queue_t *vq, uint64_t pending_limit)
232 {
233 zio_t *fio, *lio, *aio, *dio, *nio, *mio;
234 avl_tree_t *t;
235 int flags;
236 uint64_t maxspan = zfs_vdev_aggregation_limit;
237 uint64_t maxgap;
238 int stretch;
239
240 again:
241 ASSERT(MUTEX_HELD(&vq->vq_lock));
242
243 if (avl_numnodes(&vq->vq_pending_tree) >= pending_limit ||
244 avl_numnodes(&vq->vq_deadline_tree) == 0)
245 return (NULL);
246
247 fio = lio = avl_first(&vq->vq_deadline_tree);
248
249 t = fio->io_vdev_tree;
250 flags = fio->io_flags & ZIO_FLAG_AGG_INHERIT;
251 maxgap = (t == &vq->vq_read_tree) ? zfs_vdev_read_gap_limit : 0;
252
253 if (!(flags & ZIO_FLAG_DONT_AGGREGATE)) {
254 /*
255 * We can aggregate I/Os that are sufficiently adjacent and of
256 * the same flavor, as expressed by the AGG_INHERIT flags.
257 * The latter requirement is necessary so that certain
258 * attributes of the I/O, such as whether it's a normal I/O
259 * or a scrub/resilver, can be preserved in the aggregate.
260 * We can include optional I/Os, but don't allow them
261 * to begin a range as they add no benefit in that situation.
262 */
263
264 /*
265 * We keep track of the last non-optional I/O.
266 */
267 mio = (fio->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : fio;
268
269 /*
270 * Walk backwards through sufficiently contiguous I/Os
271 * recording the last non-option I/O.
272 */
273 while ((dio = AVL_PREV(t, fio)) != NULL &&
274 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
275 IO_SPAN(dio, lio) <= maxspan &&
276 IO_GAP(dio, fio) <= maxgap) {
277 fio = dio;
278 if (mio == NULL && !(fio->io_flags & ZIO_FLAG_OPTIONAL))
279 mio = fio;
280 }
281
282 /*
283 * Skip any initial optional I/Os.
284 */
285 while ((fio->io_flags & ZIO_FLAG_OPTIONAL) && fio != lio) {
286 fio = AVL_NEXT(t, fio);
287 ASSERT(fio != NULL);
288 }
289
290 /*
291 * Walk forward through sufficiently contiguous I/Os.
292 */
293 while ((dio = AVL_NEXT(t, lio)) != NULL &&
294 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
295 IO_SPAN(fio, dio) <= maxspan &&
296 IO_GAP(lio, dio) <= maxgap) {
297 lio = dio;
298 if (!(lio->io_flags & ZIO_FLAG_OPTIONAL))
299 mio = lio;
300 }
301
302 /*
303 * Now that we've established the range of the I/O aggregation
304 * we must decide what to do with trailing optional I/Os.
305 * For reads, there's nothing to do. While we are unable to
306 * aggregate further, it's possible that a trailing optional
307 * I/O would allow the underlying device to aggregate with
308 * subsequent I/Os. We must therefore determine if the next
309 * non-optional I/O is close enough to make aggregation
310 * worthwhile.
311 */
312 stretch = B_FALSE;
313 if (t != &vq->vq_read_tree && mio != NULL) {
314 nio = lio;
315 while ((dio = AVL_NEXT(t, nio)) != NULL &&
316 IO_GAP(nio, dio) == 0 &&
317 IO_GAP(mio, dio) <= zfs_vdev_write_gap_limit) {
318 nio = dio;
319 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
320 stretch = B_TRUE;
321 break;
322 }
323 }
324 }
325
326 if (stretch) {
327 /* This may be a no-op. */
328 VERIFY((dio = AVL_NEXT(t, lio)) != NULL);
329 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
330 } else {
331 while (lio != mio && lio != fio) {
332 ASSERT(lio->io_flags & ZIO_FLAG_OPTIONAL);
333 lio = AVL_PREV(t, lio);
334 ASSERT(lio != NULL);
335 }
336 }
337 }
338
339 if (fio != lio) {
340 uint64_t size = IO_SPAN(fio, lio);
341 ASSERT(size <= zfs_vdev_aggregation_limit);
342
343 aio = zio_vdev_delegated_io(fio->io_vd, fio->io_offset,
344 zio_buf_alloc(size), size, fio->io_type, ZIO_PRIORITY_AGG,
345 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
346 vdev_queue_agg_io_done, NULL);
347 aio->io_timestamp = fio->io_timestamp;
348
349 nio = fio;
350 do {
351 dio = nio;
352 nio = AVL_NEXT(t, dio);
353 ASSERT(dio->io_type == aio->io_type);
354 ASSERT(dio->io_vdev_tree == t);
355
356 if (dio->io_flags & ZIO_FLAG_NODATA) {
357 ASSERT(dio->io_type == ZIO_TYPE_WRITE);
358 bzero((char *)aio->io_data + (dio->io_offset -
359 aio->io_offset), dio->io_size);
360 } else if (dio->io_type == ZIO_TYPE_WRITE) {
361 bcopy(dio->io_data, (char *)aio->io_data +
362 (dio->io_offset - aio->io_offset),
363 dio->io_size);
364 }
365
366 zio_add_child(dio, aio);
367 vdev_queue_io_remove(vq, dio);
368 zio_vdev_io_bypass(dio);
369 zio_execute(dio);
370 } while (dio != lio);
371
372 vdev_queue_pending_add(vq, aio);
373
374 return (aio);
375 }
376
377 ASSERT(fio->io_vdev_tree == t);
378 vdev_queue_io_remove(vq, fio);
379
380 /*
381 * If the I/O is or was optional and therefore has no data, we need to
382 * simply discard it. We need to drop the vdev queue's lock to avoid a
383 * deadlock that we could encounter since this I/O will complete
384 * immediately.
385 */
386 if (fio->io_flags & ZIO_FLAG_NODATA) {
387 mutex_exit(&vq->vq_lock);
388 zio_vdev_io_bypass(fio);
389 zio_execute(fio);
390 mutex_enter(&vq->vq_lock);
391 goto again;
392 }
393
394 vdev_queue_pending_add(vq, fio);
395
396 return (fio);
397 }
398
399 zio_t *
400 vdev_queue_io(zio_t *zio)
401 {
402 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
403 zio_t *nio;
404
405 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
406
407 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
408 return (zio);
409
410 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
411
412 if (zio->io_type == ZIO_TYPE_READ)
413 zio->io_vdev_tree = &vq->vq_read_tree;
414 else
415 zio->io_vdev_tree = &vq->vq_write_tree;
416
417 mutex_enter(&vq->vq_lock);
418
419 zio->io_timestamp = gethrtime();
420 zio->io_deadline = (zio->io_timestamp >> zfs_vdev_time_shift) +
421 zio->io_priority;
422
423 vdev_queue_io_add(vq, zio);
424
425 nio = vdev_queue_io_to_issue(vq, zfs_vdev_min_pending);
426
427 mutex_exit(&vq->vq_lock);
428
429 if (nio == NULL)
430 return (NULL);
431
432 if (nio->io_done == vdev_queue_agg_io_done) {
433 zio_nowait(nio);
434 return (NULL);
435 }
436
437 return (nio);
438 }
439
440 void
441 vdev_queue_io_done(zio_t *zio)
442 {
443 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
444
445 if (zio_injection_enabled)
446 delay(SEC_TO_TICK(zio_handle_io_delay(zio)));
447
448 mutex_enter(&vq->vq_lock);
449
450 vdev_queue_pending_remove(vq, zio);
451
452 vq->vq_io_complete_ts = gethrtime();
453
454 for (int i = 0; i < zfs_vdev_ramp_rate; i++) {
455 zio_t *nio = vdev_queue_io_to_issue(vq, zfs_vdev_max_pending);
456 if (nio == NULL)
457 break;
458 mutex_exit(&vq->vq_lock);
459 if (nio->io_done == vdev_queue_agg_io_done) {
460 zio_nowait(nio);
461 } else {
462 zio_vdev_io_reissue(nio);
463 zio_execute(nio);
464 }
465 mutex_enter(&vq->vq_lock);
466 }
467
468 mutex_exit(&vq->vq_lock);
469 }