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