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.
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.
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]
22 * Copyright 2008 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
26 #include <sys/zfs_context.h>
28 #include <sys/vdev_impl.h>
33 * These tunables are for performance analysis.
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
41 int zfs_vdev_max_pending = 35;
42 int zfs_vdev_min_pending = 4;
44 /* deadline = pri + (lbolt >> time_shift) */
45 int zfs_vdev_time_shift = 6;
47 /* exponential I/O issue ramp-up rate */
48 int zfs_vdev_ramp_rate = 2;
51 * i/os will be aggregated into a single large i/o up to
52 * zfs_vdev_aggregation_limit bytes long.
54 int zfs_vdev_aggregation_limit = SPA_MAXBLOCKSIZE;
57 * Virtual device vector for disk I/O scheduling.
60 vdev_queue_deadline_compare(const void *x1, const void *x2)
65 if (z1->io_deadline < z2->io_deadline)
67 if (z1->io_deadline > z2->io_deadline)
70 if (z1->io_offset < z2->io_offset)
72 if (z1->io_offset > z2->io_offset)
84 vdev_queue_offset_compare(const void *x1, const void *x2)
89 if (z1->io_offset < z2->io_offset)
91 if (z1->io_offset > z2->io_offset)
103 vdev_queue_init(vdev_t *vd)
105 vdev_queue_t *vq = &vd->vdev_queue;
107 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
109 avl_create(&vq->vq_deadline_tree, vdev_queue_deadline_compare,
110 sizeof (zio_t), offsetof(struct zio, io_deadline_node));
112 avl_create(&vq->vq_read_tree, vdev_queue_offset_compare,
113 sizeof (zio_t), offsetof(struct zio, io_offset_node));
115 avl_create(&vq->vq_write_tree, vdev_queue_offset_compare,
116 sizeof (zio_t), offsetof(struct zio, io_offset_node));
118 avl_create(&vq->vq_pending_tree, vdev_queue_offset_compare,
119 sizeof (zio_t), offsetof(struct zio, io_offset_node));
123 vdev_queue_fini(vdev_t *vd)
125 vdev_queue_t *vq = &vd->vdev_queue;
127 avl_destroy(&vq->vq_deadline_tree);
128 avl_destroy(&vq->vq_read_tree);
129 avl_destroy(&vq->vq_write_tree);
130 avl_destroy(&vq->vq_pending_tree);
132 mutex_destroy(&vq->vq_lock);
136 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
138 avl_add(&vq->vq_deadline_tree, zio);
139 avl_add(zio->io_vdev_tree, zio);
143 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
145 avl_remove(&vq->vq_deadline_tree, zio);
146 avl_remove(zio->io_vdev_tree, zio);
150 vdev_queue_agg_io_done(zio_t *aio)
155 while ((dio = aio->io_delegate_list) != NULL) {
156 if (aio->io_type == ZIO_TYPE_READ)
157 bcopy((char *)aio->io_data + offset, dio->io_data,
159 offset += dio->io_size;
160 aio->io_delegate_list = dio->io_delegate_next;
161 dio->io_delegate_next = NULL;
162 dio->io_error = aio->io_error;
165 ASSERT3U(offset, ==, aio->io_size);
167 zio_buf_free(aio->io_data, aio->io_size);
170 #define IS_ADJACENT(io, nio) \
171 ((io)->io_offset + (io)->io_size == (nio)->io_offset)
174 vdev_queue_io_to_issue(vdev_queue_t *vq, uint64_t pending_limit)
176 zio_t *fio, *lio, *aio, *dio;
181 ASSERT(MUTEX_HELD(&vq->vq_lock));
183 if (avl_numnodes(&vq->vq_pending_tree) >= pending_limit ||
184 avl_numnodes(&vq->vq_deadline_tree) == 0)
187 fio = lio = avl_first(&vq->vq_deadline_tree);
189 tree = fio->io_vdev_tree;
191 flags = fio->io_flags & ZIO_FLAG_AGG_INHERIT;
193 if (!(flags & ZIO_FLAG_DONT_AGGREGATE)) {
195 * We can aggregate I/Os that are adjacent and of the
196 * same flavor, as expressed by the AGG_INHERIT flags.
197 * The latter is necessary so that certain attributes
198 * of the I/O, such as whether it's a normal I/O or a
199 * scrub/resilver, can be preserved in the aggregate.
201 while ((dio = AVL_PREV(tree, fio)) != NULL &&
202 IS_ADJACENT(dio, fio) &&
203 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
204 size + dio->io_size <= zfs_vdev_aggregation_limit) {
205 dio->io_delegate_next = fio;
207 size += dio->io_size;
209 while ((dio = AVL_NEXT(tree, lio)) != NULL &&
210 IS_ADJACENT(lio, dio) &&
211 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
212 size + dio->io_size <= zfs_vdev_aggregation_limit) {
213 lio->io_delegate_next = dio;
215 size += dio->io_size;
220 char *buf = zio_buf_alloc(size);
223 ASSERT(size <= zfs_vdev_aggregation_limit);
225 aio = zio_vdev_delegated_io(fio->io_vd, fio->io_offset,
226 buf, size, fio->io_type, ZIO_PRIORITY_NOW,
227 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
228 vdev_queue_agg_io_done, NULL);
230 aio->io_delegate_list = fio;
232 for (dio = fio; dio != NULL; dio = dio->io_delegate_next) {
233 ASSERT(dio->io_type == aio->io_type);
234 ASSERT(dio->io_vdev_tree == tree);
235 if (dio->io_type == ZIO_TYPE_WRITE)
236 bcopy(dio->io_data, buf + offset, dio->io_size);
237 offset += dio->io_size;
238 vdev_queue_io_remove(vq, dio);
239 zio_vdev_io_bypass(dio);
242 ASSERT(offset == size);
244 avl_add(&vq->vq_pending_tree, aio);
249 ASSERT(fio->io_vdev_tree == tree);
250 vdev_queue_io_remove(vq, fio);
252 avl_add(&vq->vq_pending_tree, fio);
258 vdev_queue_io(zio_t *zio)
260 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
263 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
265 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
268 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
270 if (zio->io_type == ZIO_TYPE_READ)
271 zio->io_vdev_tree = &vq->vq_read_tree;
273 zio->io_vdev_tree = &vq->vq_write_tree;
275 mutex_enter(&vq->vq_lock);
277 zio->io_deadline = (lbolt64 >> zfs_vdev_time_shift) + zio->io_priority;
279 vdev_queue_io_add(vq, zio);
281 nio = vdev_queue_io_to_issue(vq, zfs_vdev_min_pending);
283 mutex_exit(&vq->vq_lock);
288 if (nio->io_done == vdev_queue_agg_io_done) {
297 vdev_queue_io_done(zio_t *zio)
299 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
301 mutex_enter(&vq->vq_lock);
303 avl_remove(&vq->vq_pending_tree, zio);
305 for (int i = 0; i < zfs_vdev_ramp_rate; i++) {
306 zio_t *nio = vdev_queue_io_to_issue(vq, zfs_vdev_max_pending);
309 mutex_exit(&vq->vq_lock);
310 if (nio->io_done == vdev_queue_agg_io_done) {
313 zio_vdev_io_reissue(nio);
316 mutex_enter(&vq->vq_lock);
319 mutex_exit(&vq->vq_lock);