vdev_queue.c revision 209962 
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#include <sys/zfs_context.h>
27#include <sys/spa.h>
28#include <sys/vdev_impl.h>
29#include <sys/zio.h>
30#include <sys/avl.h>
31
32/*
33 * These tunables are for performance analysis.
34 */
35/*
36 * zfs_vdev_max_pending is the maximum number of i/os concurrently
37 * pending to each device.  zfs_vdev_min_pending is the initial number
38 * of i/os pending to each device (before it starts ramping up to
39 * max_pending).
40 */
41int zfs_vdev_max_pending = 35;
42int zfs_vdev_min_pending = 4;
43
44/* deadline = pri + (LBOLT >> time_shift) */
45int zfs_vdev_time_shift = 6;
46
47/* exponential I/O issue ramp-up rate */
48int zfs_vdev_ramp_rate = 2;
49
50/*
51 * To reduce IOPs, we aggregate small adjacent i/os into one large i/o.
52 * For read i/os, we also aggregate across small adjacency gaps.
53 */
54int zfs_vdev_aggregation_limit = SPA_MAXBLOCKSIZE;
55int zfs_vdev_read_gap_limit = 32 << 10;
56
57SYSCTL_DECL(_vfs_zfs_vdev);
58TUNABLE_INT("vfs.zfs.vdev.max_pending", &zfs_vdev_max_pending);
59SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, max_pending, CTLFLAG_RDTUN,
60    &zfs_vdev_max_pending, 0, "Maximum I/O requests pending on each device");
61TUNABLE_INT("vfs.zfs.vdev.min_pending", &zfs_vdev_min_pending);
62SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, min_pending, CTLFLAG_RDTUN,
63    &zfs_vdev_min_pending, 0,
64    "Initial number of I/O requests pending to each device");
65TUNABLE_INT("vfs.zfs.vdev.time_shift", &zfs_vdev_time_shift);
66SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, time_shift, CTLFLAG_RDTUN,
67    &zfs_vdev_time_shift, 0, "Used for calculating I/O request deadline");
68TUNABLE_INT("vfs.zfs.vdev.ramp_rate", &zfs_vdev_ramp_rate);
69SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, ramp_rate, CTLFLAG_RDTUN,
70    &zfs_vdev_ramp_rate, 0, "Exponential I/O issue ramp-up rate");
71TUNABLE_INT("vfs.zfs.vdev.aggregation_limit", &zfs_vdev_aggregation_limit);
72SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RDTUN,
73    &zfs_vdev_aggregation_limit, 0,
74    "I/O requests are aggregated up to this size");
75
76/*
77 * Virtual device vector for disk I/O scheduling.
78 */
79int
80vdev_queue_deadline_compare(const void *x1, const void *x2)
81{
82	const zio_t *z1 = x1;
83	const zio_t *z2 = x2;
84
85	if (z1->io_deadline < z2->io_deadline)
86		return (-1);
87	if (z1->io_deadline > z2->io_deadline)
88		return (1);
89
90	if (z1->io_offset < z2->io_offset)
91		return (-1);
92	if (z1->io_offset > z2->io_offset)
93		return (1);
94
95	if (z1 < z2)
96		return (-1);
97	if (z1 > z2)
98		return (1);
99
100	return (0);
101}
102
103int
104vdev_queue_offset_compare(const void *x1, const void *x2)
105{
106	const zio_t *z1 = x1;
107	const zio_t *z2 = x2;
108
109	if (z1->io_offset < z2->io_offset)
110		return (-1);
111	if (z1->io_offset > z2->io_offset)
112		return (1);
113
114	if (z1 < z2)
115		return (-1);
116	if (z1 > z2)
117		return (1);
118
119	return (0);
120}
121
122void
123vdev_queue_init(vdev_t *vd)
124{
125	vdev_queue_t *vq = &vd->vdev_queue;
126
127	mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
128
129	avl_create(&vq->vq_deadline_tree, vdev_queue_deadline_compare,
130	    sizeof (zio_t), offsetof(struct zio, io_deadline_node));
131
132	avl_create(&vq->vq_read_tree, vdev_queue_offset_compare,
133	    sizeof (zio_t), offsetof(struct zio, io_offset_node));
134
135	avl_create(&vq->vq_write_tree, vdev_queue_offset_compare,
136	    sizeof (zio_t), offsetof(struct zio, io_offset_node));
137
138	avl_create(&vq->vq_pending_tree, vdev_queue_offset_compare,
139	    sizeof (zio_t), offsetof(struct zio, io_offset_node));
140}
141
142void
143vdev_queue_fini(vdev_t *vd)
144{
145	vdev_queue_t *vq = &vd->vdev_queue;
146
147	avl_destroy(&vq->vq_deadline_tree);
148	avl_destroy(&vq->vq_read_tree);
149	avl_destroy(&vq->vq_write_tree);
150	avl_destroy(&vq->vq_pending_tree);
151
152	mutex_destroy(&vq->vq_lock);
153}
154
155static void
156vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
157{
158	avl_add(&vq->vq_deadline_tree, zio);
159	avl_add(zio->io_vdev_tree, zio);
160}
161
162static void
163vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
164{
165	avl_remove(&vq->vq_deadline_tree, zio);
166	avl_remove(zio->io_vdev_tree, zio);
167}
168
169static void
170vdev_queue_agg_io_done(zio_t *aio)
171{
172	zio_t *pio;
173
174	while ((pio = zio_walk_parents(aio)) != NULL)
175		if (aio->io_type == ZIO_TYPE_READ)
176			bcopy((char *)aio->io_data + (pio->io_offset -
177			    aio->io_offset), pio->io_data, pio->io_size);
178
179	zio_buf_free(aio->io_data, aio->io_size);
180}
181
182/*
183 * Compute the range spanned by two i/os, which is the endpoint of the last
184 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
185 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
186 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
187 */
188#define	IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
189#define	IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
190
191static zio_t *
192vdev_queue_io_to_issue(vdev_queue_t *vq, uint64_t pending_limit)
193{
194	zio_t *fio, *lio, *aio, *dio, *nio;
195	avl_tree_t *t;
196	int flags;
197	uint64_t maxspan = zfs_vdev_aggregation_limit;
198	uint64_t maxgap;
199
200	ASSERT(MUTEX_HELD(&vq->vq_lock));
201
202	if (avl_numnodes(&vq->vq_pending_tree) >= pending_limit ||
203	    avl_numnodes(&vq->vq_deadline_tree) == 0)
204		return (NULL);
205
206	fio = lio = avl_first(&vq->vq_deadline_tree);
207
208	t = fio->io_vdev_tree;
209	flags = fio->io_flags & ZIO_FLAG_AGG_INHERIT;
210	maxgap = (t == &vq->vq_read_tree) ? zfs_vdev_read_gap_limit : 0;
211
212	if (!(flags & ZIO_FLAG_DONT_AGGREGATE)) {
213		/*
214		 * We can aggregate I/Os that are adjacent and of the
215		 * same flavor, as expressed by the AGG_INHERIT flags.
216		 * The latter is necessary so that certain attributes
217		 * of the I/O, such as whether it's a normal I/O or a
218		 * scrub/resilver, can be preserved in the aggregate.
219		 */
220		while ((dio = AVL_PREV(t, fio)) != NULL &&
221		    (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
222		    IO_SPAN(dio, lio) <= maxspan && IO_GAP(dio, fio) <= maxgap)
223			fio = dio;
224
225		while ((dio = AVL_NEXT(t, lio)) != NULL &&
226		    (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
227		    IO_SPAN(fio, dio) <= maxspan && IO_GAP(lio, dio) <= maxgap)
228			lio = dio;
229	}
230
231	if (fio != lio) {
232		uint64_t size = IO_SPAN(fio, lio);
233		ASSERT(size <= zfs_vdev_aggregation_limit);
234
235		aio = zio_vdev_delegated_io(fio->io_vd, fio->io_offset,
236		    zio_buf_alloc(size), size, fio->io_type, ZIO_PRIORITY_NOW,
237		    flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
238		    vdev_queue_agg_io_done, NULL);
239
240		nio = fio;
241		do {
242			dio = nio;
243			nio = AVL_NEXT(t, dio);
244			ASSERT(dio->io_type == aio->io_type);
245			ASSERT(dio->io_vdev_tree == t);
246
247			if (dio->io_type == ZIO_TYPE_WRITE)
248				bcopy(dio->io_data, (char *)aio->io_data +
249				    (dio->io_offset - aio->io_offset),
250				    dio->io_size);
251
252			zio_add_child(dio, aio);
253			vdev_queue_io_remove(vq, dio);
254			zio_vdev_io_bypass(dio);
255			zio_execute(dio);
256		} while (dio != lio);
257
258		avl_add(&vq->vq_pending_tree, aio);
259
260		return (aio);
261	}
262
263	ASSERT(fio->io_vdev_tree == t);
264	vdev_queue_io_remove(vq, fio);
265
266	avl_add(&vq->vq_pending_tree, fio);
267
268	return (fio);
269}
270
271zio_t *
272vdev_queue_io(zio_t *zio)
273{
274	vdev_queue_t *vq = &zio->io_vd->vdev_queue;
275	zio_t *nio;
276
277	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
278
279	if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
280		return (zio);
281
282	zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
283
284	if (zio->io_type == ZIO_TYPE_READ)
285		zio->io_vdev_tree = &vq->vq_read_tree;
286	else
287		zio->io_vdev_tree = &vq->vq_write_tree;
288
289	mutex_enter(&vq->vq_lock);
290
291	zio->io_deadline = (lbolt64 >> zfs_vdev_time_shift) + zio->io_priority;
292
293	vdev_queue_io_add(vq, zio);
294
295	nio = vdev_queue_io_to_issue(vq, zfs_vdev_min_pending);
296
297	mutex_exit(&vq->vq_lock);
298
299	if (nio == NULL)
300		return (NULL);
301
302	if (nio->io_done == vdev_queue_agg_io_done) {
303		zio_nowait(nio);
304		return (NULL);
305	}
306
307	return (nio);
308}
309
310void
311vdev_queue_io_done(zio_t *zio)
312{
313	vdev_queue_t *vq = &zio->io_vd->vdev_queue;
314
315	mutex_enter(&vq->vq_lock);
316
317	avl_remove(&vq->vq_pending_tree, zio);
318
319	for (int i = 0; i < zfs_vdev_ramp_rate; i++) {
320		zio_t *nio = vdev_queue_io_to_issue(vq, zfs_vdev_max_pending);
321		if (nio == NULL)
322			break;
323		mutex_exit(&vq->vq_lock);
324		if (nio->io_done == vdev_queue_agg_io_done) {
325			zio_nowait(nio);
326		} else {
327			zio_vdev_io_reissue(nio);
328			zio_execute(nio);
329		}
330		mutex_enter(&vq->vq_lock);
331	}
332
333	mutex_exit(&vq->vq_lock);
334}
335