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 2009 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 * 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.
54 int zfs_vdev_aggregation_limit = SPA_MAXBLOCKSIZE;
55 int zfs_vdev_read_gap_limit = 32 << 10;
57 SYSCTL_DECL(_vfs_zfs_vdev);
58 TUNABLE_INT("vfs.zfs.vdev.max_pending", &zfs_vdev_max_pending);
59 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, max_pending, CTLFLAG_RDTUN,
60 &zfs_vdev_max_pending, 0, "Maximum I/O requests pending on each device");
61 TUNABLE_INT("vfs.zfs.vdev.min_pending", &zfs_vdev_min_pending);
62 SYSCTL_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");
65 TUNABLE_INT("vfs.zfs.vdev.time_shift", &zfs_vdev_time_shift);
66 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, time_shift, CTLFLAG_RDTUN,
67 &zfs_vdev_time_shift, 0, "Used for calculating I/O request deadline");
68 TUNABLE_INT("vfs.zfs.vdev.ramp_rate", &zfs_vdev_ramp_rate);
69 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, ramp_rate, CTLFLAG_RDTUN,
70 &zfs_vdev_ramp_rate, 0, "Exponential I/O issue ramp-up rate");
71 TUNABLE_INT("vfs.zfs.vdev.aggregation_limit", &zfs_vdev_aggregation_limit);
72 SYSCTL_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");
77 * Virtual device vector for disk I/O scheduling.
80 vdev_queue_deadline_compare(const void *x1, const void *x2)
85 if (z1->io_deadline < z2->io_deadline)
87 if (z1->io_deadline > z2->io_deadline)
90 if (z1->io_offset < z2->io_offset)
92 if (z1->io_offset > z2->io_offset)
104 vdev_queue_offset_compare(const void *x1, const void *x2)
106 const zio_t *z1 = x1;
107 const zio_t *z2 = x2;
109 if (z1->io_offset < z2->io_offset)
111 if (z1->io_offset > z2->io_offset)
123 vdev_queue_init(vdev_t *vd)
125 vdev_queue_t *vq = &vd->vdev_queue;
127 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
129 avl_create(&vq->vq_deadline_tree, vdev_queue_deadline_compare,
130 sizeof (zio_t), offsetof(struct zio, io_deadline_node));
132 avl_create(&vq->vq_read_tree, vdev_queue_offset_compare,
133 sizeof (zio_t), offsetof(struct zio, io_offset_node));
135 avl_create(&vq->vq_write_tree, vdev_queue_offset_compare,
136 sizeof (zio_t), offsetof(struct zio, io_offset_node));
138 avl_create(&vq->vq_pending_tree, vdev_queue_offset_compare,
139 sizeof (zio_t), offsetof(struct zio, io_offset_node));
143 vdev_queue_fini(vdev_t *vd)
145 vdev_queue_t *vq = &vd->vdev_queue;
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);
152 mutex_destroy(&vq->vq_lock);
156 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
158 avl_add(&vq->vq_deadline_tree, zio);
159 avl_add(zio->io_vdev_tree, zio);
163 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
165 avl_remove(&vq->vq_deadline_tree, zio);
166 avl_remove(zio->io_vdev_tree, zio);
170 vdev_queue_agg_io_done(zio_t *aio)
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);
179 zio_buf_free(aio->io_data, aio->io_size);
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.
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))
192 vdev_queue_io_to_issue(vdev_queue_t *vq, uint64_t pending_limit)
194 zio_t *fio, *lio, *aio, *dio, *nio;
197 uint64_t maxspan = zfs_vdev_aggregation_limit;
200 ASSERT(MUTEX_HELD(&vq->vq_lock));
202 if (avl_numnodes(&vq->vq_pending_tree) >= pending_limit ||
203 avl_numnodes(&vq->vq_deadline_tree) == 0)
206 fio = lio = avl_first(&vq->vq_deadline_tree);
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;
212 if (!(flags & ZIO_FLAG_DONT_AGGREGATE)) {
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.
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)
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)
232 uint64_t size = IO_SPAN(fio, lio);
233 ASSERT(size <= zfs_vdev_aggregation_limit);
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);
243 nio = AVL_NEXT(t, dio);
244 ASSERT(dio->io_type == aio->io_type);
245 ASSERT(dio->io_vdev_tree == t);
247 if (dio->io_type == ZIO_TYPE_WRITE)
248 bcopy(dio->io_data, (char *)aio->io_data +
249 (dio->io_offset - aio->io_offset),
252 zio_add_child(dio, aio);
253 vdev_queue_io_remove(vq, dio);
254 zio_vdev_io_bypass(dio);
256 } while (dio != lio);
258 avl_add(&vq->vq_pending_tree, aio);
263 ASSERT(fio->io_vdev_tree == t);
264 vdev_queue_io_remove(vq, fio);
266 avl_add(&vq->vq_pending_tree, fio);
272 vdev_queue_io(zio_t *zio)
274 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
277 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
279 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
282 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
284 if (zio->io_type == ZIO_TYPE_READ)
285 zio->io_vdev_tree = &vq->vq_read_tree;
287 zio->io_vdev_tree = &vq->vq_write_tree;
289 mutex_enter(&vq->vq_lock);
291 zio->io_deadline = (lbolt64 >> zfs_vdev_time_shift) + zio->io_priority;
293 vdev_queue_io_add(vq, zio);
295 nio = vdev_queue_io_to_issue(vq, zfs_vdev_min_pending);
297 mutex_exit(&vq->vq_lock);
302 if (nio->io_done == vdev_queue_agg_io_done) {
311 vdev_queue_io_done(zio_t *zio)
313 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
315 mutex_enter(&vq->vq_lock);
317 avl_remove(&vq->vq_pending_tree, zio);
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);
323 mutex_exit(&vq->vq_lock);
324 if (nio->io_done == vdev_queue_agg_io_done) {
327 zio_vdev_io_reissue(nio);
330 mutex_enter(&vq->vq_lock);
333 mutex_exit(&vq->vq_lock);