xref: /titanic_41/usr/src/uts/common/fs/zfs/vdev_queue.c (revision 9a0e82381c494d655823a54fc713a8c2a571131c)
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  */
41 int zfs_vdev_max_pending = 35;
42 int zfs_vdev_min_pending = 4;
43 
44 /* deadline = pri + (lbolt >> time_shift) */
45 int zfs_vdev_time_shift = 6;
46 
47 /* exponential I/O issue ramp-up rate */
48 int 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  */
54 int zfs_vdev_aggregation_limit = SPA_MAXBLOCKSIZE;
55 int zfs_vdev_read_gap_limit = 32 << 10;
56 
57 /*
58  * Virtual device vector for disk I/O scheduling.
59  */
60 int
61 vdev_queue_deadline_compare(const void *x1, const void *x2)
62 {
63 	const zio_t *z1 = x1;
64 	const zio_t *z2 = x2;
65 
66 	if (z1->io_deadline < z2->io_deadline)
67 		return (-1);
68 	if (z1->io_deadline > z2->io_deadline)
69 		return (1);
70 
71 	if (z1->io_offset < z2->io_offset)
72 		return (-1);
73 	if (z1->io_offset > z2->io_offset)
74 		return (1);
75 
76 	if (z1 < z2)
77 		return (-1);
78 	if (z1 > z2)
79 		return (1);
80 
81 	return (0);
82 }
83 
84 int
85 vdev_queue_offset_compare(const void *x1, const void *x2)
86 {
87 	const zio_t *z1 = x1;
88 	const zio_t *z2 = x2;
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 
103 void
104 vdev_queue_init(vdev_t *vd)
105 {
106 	vdev_queue_t *vq = &vd->vdev_queue;
107 
108 	mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
109 
110 	avl_create(&vq->vq_deadline_tree, vdev_queue_deadline_compare,
111 	    sizeof (zio_t), offsetof(struct zio, io_deadline_node));
112 
113 	avl_create(&vq->vq_read_tree, vdev_queue_offset_compare,
114 	    sizeof (zio_t), offsetof(struct zio, io_offset_node));
115 
116 	avl_create(&vq->vq_write_tree, vdev_queue_offset_compare,
117 	    sizeof (zio_t), offsetof(struct zio, io_offset_node));
118 
119 	avl_create(&vq->vq_pending_tree, vdev_queue_offset_compare,
120 	    sizeof (zio_t), offsetof(struct zio, io_offset_node));
121 }
122 
123 void
124 vdev_queue_fini(vdev_t *vd)
125 {
126 	vdev_queue_t *vq = &vd->vdev_queue;
127 
128 	avl_destroy(&vq->vq_deadline_tree);
129 	avl_destroy(&vq->vq_read_tree);
130 	avl_destroy(&vq->vq_write_tree);
131 	avl_destroy(&vq->vq_pending_tree);
132 
133 	mutex_destroy(&vq->vq_lock);
134 }
135 
136 static void
137 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
138 {
139 	avl_add(&vq->vq_deadline_tree, zio);
140 	avl_add(zio->io_vdev_tree, zio);
141 }
142 
143 static void
144 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
145 {
146 	avl_remove(&vq->vq_deadline_tree, zio);
147 	avl_remove(zio->io_vdev_tree, zio);
148 }
149 
150 static void
151 vdev_queue_agg_io_done(zio_t *aio)
152 {
153 	zio_t *pio;
154 
155 	while ((pio = zio_walk_parents(aio)) != NULL)
156 		if (aio->io_type == ZIO_TYPE_READ)
157 			bcopy((char *)aio->io_data + (pio->io_offset -
158 			    aio->io_offset), pio->io_data, pio->io_size);
159 
160 	zio_buf_free(aio->io_data, aio->io_size);
161 }
162 
163 /*
164  * Compute the range spanned by two i/os, which is the endpoint of the last
165  * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
166  * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
167  * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
168  */
169 #define	IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
170 #define	IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
171 
172 static zio_t *
173 vdev_queue_io_to_issue(vdev_queue_t *vq, uint64_t pending_limit)
174 {
175 	zio_t *fio, *lio, *aio, *dio, *nio;
176 	avl_tree_t *t;
177 	int flags;
178 	uint64_t maxspan = zfs_vdev_aggregation_limit;
179 	uint64_t maxgap;
180 
181 	ASSERT(MUTEX_HELD(&vq->vq_lock));
182 
183 	if (avl_numnodes(&vq->vq_pending_tree) >= pending_limit ||
184 	    avl_numnodes(&vq->vq_deadline_tree) == 0)
185 		return (NULL);
186 
187 	fio = lio = avl_first(&vq->vq_deadline_tree);
188 
189 	t = fio->io_vdev_tree;
190 	flags = fio->io_flags & ZIO_FLAG_AGG_INHERIT;
191 	maxgap = (t == &vq->vq_read_tree) ? zfs_vdev_read_gap_limit : 0;
192 
193 	if (!(flags & ZIO_FLAG_DONT_AGGREGATE)) {
194 		/*
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.
200 		 */
201 		while ((dio = AVL_PREV(t, fio)) != NULL &&
202 		    (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
203 		    IO_SPAN(dio, lio) <= maxspan && IO_GAP(dio, fio) <= maxgap)
204 			fio = dio;
205 
206 		while ((dio = AVL_NEXT(t, lio)) != NULL &&
207 		    (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
208 		    IO_SPAN(fio, dio) <= maxspan && IO_GAP(lio, dio) <= maxgap)
209 			lio = dio;
210 	}
211 
212 	if (fio != lio) {
213 		uint64_t size = IO_SPAN(fio, lio);
214 		ASSERT(size <= zfs_vdev_aggregation_limit);
215 
216 		aio = zio_vdev_delegated_io(fio->io_vd, fio->io_offset,
217 		    zio_buf_alloc(size), size, fio->io_type, ZIO_PRIORITY_NOW,
218 		    flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
219 		    vdev_queue_agg_io_done, NULL);
220 
221 		nio = fio;
222 		do {
223 			dio = nio;
224 			nio = AVL_NEXT(t, dio);
225 			ASSERT(dio->io_type == aio->io_type);
226 			ASSERT(dio->io_vdev_tree == t);
227 
228 			if (dio->io_type == ZIO_TYPE_WRITE)
229 				bcopy(dio->io_data, (char *)aio->io_data +
230 				    (dio->io_offset - aio->io_offset),
231 				    dio->io_size);
232 
233 			zio_add_child(dio, aio);
234 			vdev_queue_io_remove(vq, dio);
235 			zio_vdev_io_bypass(dio);
236 			zio_execute(dio);
237 		} while (dio != lio);
238 
239 		avl_add(&vq->vq_pending_tree, aio);
240 
241 		return (aio);
242 	}
243 
244 	ASSERT(fio->io_vdev_tree == t);
245 	vdev_queue_io_remove(vq, fio);
246 
247 	avl_add(&vq->vq_pending_tree, fio);
248 
249 	return (fio);
250 }
251 
252 zio_t *
253 vdev_queue_io(zio_t *zio)
254 {
255 	vdev_queue_t *vq = &zio->io_vd->vdev_queue;
256 	zio_t *nio;
257 
258 	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
259 
260 	if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
261 		return (zio);
262 
263 	zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
264 
265 	if (zio->io_type == ZIO_TYPE_READ)
266 		zio->io_vdev_tree = &vq->vq_read_tree;
267 	else
268 		zio->io_vdev_tree = &vq->vq_write_tree;
269 
270 	mutex_enter(&vq->vq_lock);
271 
272 	zio->io_deadline = (lbolt64 >> zfs_vdev_time_shift) + zio->io_priority;
273 
274 	vdev_queue_io_add(vq, zio);
275 
276 	nio = vdev_queue_io_to_issue(vq, zfs_vdev_min_pending);
277 
278 	mutex_exit(&vq->vq_lock);
279 
280 	if (nio == NULL)
281 		return (NULL);
282 
283 	if (nio->io_done == vdev_queue_agg_io_done) {
284 		zio_nowait(nio);
285 		return (NULL);
286 	}
287 
288 	return (nio);
289 }
290 
291 void
292 vdev_queue_io_done(zio_t *zio)
293 {
294 	vdev_queue_t *vq = &zio->io_vd->vdev_queue;
295 
296 	mutex_enter(&vq->vq_lock);
297 
298 	avl_remove(&vq->vq_pending_tree, zio);
299 
300 	for (int i = 0; i < zfs_vdev_ramp_rate; i++) {
301 		zio_t *nio = vdev_queue_io_to_issue(vq, zfs_vdev_max_pending);
302 		if (nio == NULL)
303 			break;
304 		mutex_exit(&vq->vq_lock);
305 		if (nio->io_done == vdev_queue_agg_io_done) {
306 			zio_nowait(nio);
307 		} else {
308 			zio_vdev_io_reissue(nio);
309 			zio_execute(nio);
310 		}
311 		mutex_enter(&vq->vq_lock);
312 	}
313 
314 	mutex_exit(&vq->vq_lock);
315 }
316