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>
27 #include <sys/vdev_impl.h>
32 * These tunables are for performance analysis.
35 * zfs_vdev_max_pending is the maximum number of i/os concurrently
36 * pending to each device. zfs_vdev_min_pending is the initial number
37 * of i/os pending to each device (before it starts ramping up to
40 int zfs_vdev_max_pending = 10;
41 int zfs_vdev_min_pending = 4;
43 /* deadline = pri + ddi_get_lbolt64() >> time_shift) */
44 int zfs_vdev_time_shift = 6;
46 /* exponential I/O issue ramp-up rate */
47 int zfs_vdev_ramp_rate = 2;
50 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
51 * For read I/Os, we also aggregate across small adjacency gaps; for writes
52 * we include spans of optional I/Os to aid aggregation at the disk even when
53 * they aren't able to help us aggregate at this level.
55 int zfs_vdev_aggregation_limit = SPA_MAXBLOCKSIZE;
56 int zfs_vdev_read_gap_limit = 32 << 10;
57 int zfs_vdev_write_gap_limit = 4 << 10;
59 SYSCTL_DECL(_vfs_zfs_vdev);
60 TUNABLE_INT("vfs.zfs.vdev.max_pending", &zfs_vdev_max_pending);
61 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, max_pending, CTLFLAG_RW,
62 &zfs_vdev_max_pending, 0, "Maximum I/O requests pending on each device");
63 TUNABLE_INT("vfs.zfs.vdev.min_pending", &zfs_vdev_min_pending);
64 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, min_pending, CTLFLAG_RW,
65 &zfs_vdev_min_pending, 0,
66 "Initial number of I/O requests pending to each device");
67 TUNABLE_INT("vfs.zfs.vdev.time_shift", &zfs_vdev_time_shift);
68 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, time_shift, CTLFLAG_RW,
69 &zfs_vdev_time_shift, 0, "Used for calculating I/O request deadline");
70 TUNABLE_INT("vfs.zfs.vdev.ramp_rate", &zfs_vdev_ramp_rate);
71 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, ramp_rate, CTLFLAG_RW,
72 &zfs_vdev_ramp_rate, 0, "Exponential I/O issue ramp-up rate");
73 TUNABLE_INT("vfs.zfs.vdev.aggregation_limit", &zfs_vdev_aggregation_limit);
74 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RW,
75 &zfs_vdev_aggregation_limit, 0,
76 "I/O requests are aggregated up to this size");
77 TUNABLE_INT("vfs.zfs.vdev.read_gap_limit", &zfs_vdev_read_gap_limit);
78 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RW,
79 &zfs_vdev_read_gap_limit, 0,
80 "Acceptable gap between two reads being aggregated");
81 TUNABLE_INT("vfs.zfs.vdev.write_gap_limit", &zfs_vdev_write_gap_limit);
82 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RW,
83 &zfs_vdev_write_gap_limit, 0,
84 "Acceptable gap between two writes being aggregated");
87 * Virtual device vector for disk I/O scheduling.
90 vdev_queue_deadline_compare(const void *x1, const void *x2)
95 if (z1->io_deadline < z2->io_deadline)
97 if (z1->io_deadline > z2->io_deadline)
100 if (z1->io_offset < z2->io_offset)
102 if (z1->io_offset > z2->io_offset)
114 vdev_queue_offset_compare(const void *x1, const void *x2)
116 const zio_t *z1 = x1;
117 const zio_t *z2 = x2;
119 if (z1->io_offset < z2->io_offset)
121 if (z1->io_offset > z2->io_offset)
133 vdev_queue_init(vdev_t *vd)
135 vdev_queue_t *vq = &vd->vdev_queue;
137 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
139 avl_create(&vq->vq_deadline_tree, vdev_queue_deadline_compare,
140 sizeof (zio_t), offsetof(struct zio, io_deadline_node));
142 avl_create(&vq->vq_read_tree, vdev_queue_offset_compare,
143 sizeof (zio_t), offsetof(struct zio, io_offset_node));
145 avl_create(&vq->vq_write_tree, vdev_queue_offset_compare,
146 sizeof (zio_t), offsetof(struct zio, io_offset_node));
148 avl_create(&vq->vq_pending_tree, vdev_queue_offset_compare,
149 sizeof (zio_t), offsetof(struct zio, io_offset_node));
153 vdev_queue_fini(vdev_t *vd)
155 vdev_queue_t *vq = &vd->vdev_queue;
157 avl_destroy(&vq->vq_deadline_tree);
158 avl_destroy(&vq->vq_read_tree);
159 avl_destroy(&vq->vq_write_tree);
160 avl_destroy(&vq->vq_pending_tree);
162 mutex_destroy(&vq->vq_lock);
166 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
168 avl_add(&vq->vq_deadline_tree, zio);
169 avl_add(zio->io_vdev_tree, zio);
173 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
175 avl_remove(&vq->vq_deadline_tree, zio);
176 avl_remove(zio->io_vdev_tree, zio);
180 vdev_queue_agg_io_done(zio_t *aio)
184 while ((pio = zio_walk_parents(aio)) != NULL)
185 if (aio->io_type == ZIO_TYPE_READ)
186 bcopy((char *)aio->io_data + (pio->io_offset -
187 aio->io_offset), pio->io_data, pio->io_size);
189 zio_buf_free(aio->io_data, aio->io_size);
193 * Compute the range spanned by two i/os, which is the endpoint of the last
194 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
195 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
196 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
198 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
199 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
202 vdev_queue_io_to_issue(vdev_queue_t *vq, uint64_t pending_limit)
204 zio_t *fio, *lio, *aio, *dio, *nio, *mio;
207 uint64_t maxspan = zfs_vdev_aggregation_limit;
212 ASSERT(MUTEX_HELD(&vq->vq_lock));
214 if (avl_numnodes(&vq->vq_pending_tree) >= pending_limit ||
215 avl_numnodes(&vq->vq_deadline_tree) == 0)
218 fio = lio = avl_first(&vq->vq_deadline_tree);
220 t = fio->io_vdev_tree;
221 flags = fio->io_flags & ZIO_FLAG_AGG_INHERIT;
222 maxgap = (t == &vq->vq_read_tree) ? zfs_vdev_read_gap_limit : 0;
224 if (!(flags & ZIO_FLAG_DONT_AGGREGATE)) {
226 * We can aggregate I/Os that are sufficiently adjacent and of
227 * the same flavor, as expressed by the AGG_INHERIT flags.
228 * The latter requirement is necessary so that certain
229 * attributes of the I/O, such as whether it's a normal I/O
230 * or a scrub/resilver, can be preserved in the aggregate.
231 * We can include optional I/Os, but don't allow them
232 * to begin a range as they add no benefit in that situation.
236 * We keep track of the last non-optional I/O.
238 mio = (fio->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : fio;
241 * Walk backwards through sufficiently contiguous I/Os
242 * recording the last non-option I/O.
244 while ((dio = AVL_PREV(t, fio)) != NULL &&
245 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
246 IO_SPAN(dio, lio) <= maxspan &&
247 IO_GAP(dio, fio) <= maxgap) {
249 if (mio == NULL && !(fio->io_flags & ZIO_FLAG_OPTIONAL))
254 * Skip any initial optional I/Os.
256 while ((fio->io_flags & ZIO_FLAG_OPTIONAL) && fio != lio) {
257 fio = AVL_NEXT(t, fio);
262 * Walk forward through sufficiently contiguous I/Os.
264 while ((dio = AVL_NEXT(t, lio)) != NULL &&
265 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
266 IO_SPAN(fio, dio) <= maxspan &&
267 IO_GAP(lio, dio) <= maxgap) {
269 if (!(lio->io_flags & ZIO_FLAG_OPTIONAL))
274 * Now that we've established the range of the I/O aggregation
275 * we must decide what to do with trailing optional I/Os.
276 * For reads, there's nothing to do. While we are unable to
277 * aggregate further, it's possible that a trailing optional
278 * I/O would allow the underlying device to aggregate with
279 * subsequent I/Os. We must therefore determine if the next
280 * non-optional I/O is close enough to make aggregation
284 if (t != &vq->vq_read_tree && mio != NULL) {
286 while ((dio = AVL_NEXT(t, nio)) != NULL &&
287 IO_GAP(nio, dio) == 0 &&
288 IO_GAP(mio, dio) <= zfs_vdev_write_gap_limit) {
290 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
298 /* This may be a no-op. */
299 VERIFY((dio = AVL_NEXT(t, lio)) != NULL);
300 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
302 while (lio != mio && lio != fio) {
303 ASSERT(lio->io_flags & ZIO_FLAG_OPTIONAL);
304 lio = AVL_PREV(t, lio);
311 uint64_t size = IO_SPAN(fio, lio);
312 ASSERT(size <= zfs_vdev_aggregation_limit);
314 aio = zio_vdev_delegated_io(fio->io_vd, fio->io_offset,
315 zio_buf_alloc(size), size, fio->io_type, ZIO_PRIORITY_AGG,
316 flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
317 vdev_queue_agg_io_done, NULL);
322 nio = AVL_NEXT(t, dio);
323 ASSERT(dio->io_type == aio->io_type);
324 ASSERT(dio->io_vdev_tree == t);
326 if (dio->io_flags & ZIO_FLAG_NODATA) {
327 ASSERT(dio->io_type == ZIO_TYPE_WRITE);
328 bzero((char *)aio->io_data + (dio->io_offset -
329 aio->io_offset), dio->io_size);
330 } else if (dio->io_type == ZIO_TYPE_WRITE) {
331 bcopy(dio->io_data, (char *)aio->io_data +
332 (dio->io_offset - aio->io_offset),
336 zio_add_child(dio, aio);
337 vdev_queue_io_remove(vq, dio);
338 zio_vdev_io_bypass(dio);
340 } while (dio != lio);
342 avl_add(&vq->vq_pending_tree, aio);
347 ASSERT(fio->io_vdev_tree == t);
348 vdev_queue_io_remove(vq, fio);
351 * If the I/O is or was optional and therefore has no data, we need to
352 * simply discard it. We need to drop the vdev queue's lock to avoid a
353 * deadlock that we could encounter since this I/O will complete
356 if (fio->io_flags & ZIO_FLAG_NODATA) {
357 mutex_exit(&vq->vq_lock);
358 zio_vdev_io_bypass(fio);
360 mutex_enter(&vq->vq_lock);
364 avl_add(&vq->vq_pending_tree, fio);
370 vdev_queue_io(zio_t *zio)
372 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
375 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
377 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
380 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
382 if (zio->io_type == ZIO_TYPE_READ)
383 zio->io_vdev_tree = &vq->vq_read_tree;
385 zio->io_vdev_tree = &vq->vq_write_tree;
387 mutex_enter(&vq->vq_lock);
389 zio->io_deadline = (ddi_get_lbolt64() >> zfs_vdev_time_shift) +
392 vdev_queue_io_add(vq, zio);
394 nio = vdev_queue_io_to_issue(vq, zfs_vdev_min_pending);
396 mutex_exit(&vq->vq_lock);
401 if (nio->io_done == vdev_queue_agg_io_done) {
410 vdev_queue_io_done(zio_t *zio)
412 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
414 mutex_enter(&vq->vq_lock);
416 avl_remove(&vq->vq_pending_tree, zio);
418 for (int i = 0; i < zfs_vdev_ramp_rate; i++) {
419 zio_t *nio = vdev_queue_io_to_issue(vq, zfs_vdev_max_pending);
422 mutex_exit(&vq->vq_lock);
423 if (nio->io_done == vdev_queue_agg_io_done) {
426 zio_vdev_io_reissue(nio);
429 mutex_enter(&vq->vq_lock);
432 mutex_exit(&vq->vq_lock);