xref: /titanic_41/usr/src/uts/common/fs/zfs/vdev_queue.c (revision aff4bce51ecc47df7e5a6351b7cee6bc20408c63)
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_impl.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; for writes
53  * we include spans of optional I/Os to aid aggregation at the disk even when
54  * they aren't able to help us aggregate at this level.
55  */
56 int zfs_vdev_aggregation_limit = SPA_MAXBLOCKSIZE;
57 int zfs_vdev_read_gap_limit = 32 << 10;
58 int zfs_vdev_write_gap_limit = 4 << 10;
59 
60 /*
61  * Virtual device vector for disk I/O scheduling.
62  */
63 int
64 vdev_queue_deadline_compare(const void *x1, const void *x2)
65 {
66 	const zio_t *z1 = x1;
67 	const zio_t *z2 = x2;
68 
69 	if (z1->io_deadline < z2->io_deadline)
70 		return (-1);
71 	if (z1->io_deadline > z2->io_deadline)
72 		return (1);
73 
74 	if (z1->io_offset < z2->io_offset)
75 		return (-1);
76 	if (z1->io_offset > z2->io_offset)
77 		return (1);
78 
79 	if (z1 < z2)
80 		return (-1);
81 	if (z1 > z2)
82 		return (1);
83 
84 	return (0);
85 }
86 
87 int
88 vdev_queue_offset_compare(const void *x1, const void *x2)
89 {
90 	const zio_t *z1 = x1;
91 	const zio_t *z2 = x2;
92 
93 	if (z1->io_offset < z2->io_offset)
94 		return (-1);
95 	if (z1->io_offset > z2->io_offset)
96 		return (1);
97 
98 	if (z1 < z2)
99 		return (-1);
100 	if (z1 > z2)
101 		return (1);
102 
103 	return (0);
104 }
105 
106 void
107 vdev_queue_init(vdev_t *vd)
108 {
109 	vdev_queue_t *vq = &vd->vdev_queue;
110 
111 	mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
112 
113 	avl_create(&vq->vq_deadline_tree, vdev_queue_deadline_compare,
114 	    sizeof (zio_t), offsetof(struct zio, io_deadline_node));
115 
116 	avl_create(&vq->vq_read_tree, vdev_queue_offset_compare,
117 	    sizeof (zio_t), offsetof(struct zio, io_offset_node));
118 
119 	avl_create(&vq->vq_write_tree, vdev_queue_offset_compare,
120 	    sizeof (zio_t), offsetof(struct zio, io_offset_node));
121 
122 	avl_create(&vq->vq_pending_tree, vdev_queue_offset_compare,
123 	    sizeof (zio_t), offsetof(struct zio, io_offset_node));
124 }
125 
126 void
127 vdev_queue_fini(vdev_t *vd)
128 {
129 	vdev_queue_t *vq = &vd->vdev_queue;
130 
131 	avl_destroy(&vq->vq_deadline_tree);
132 	avl_destroy(&vq->vq_read_tree);
133 	avl_destroy(&vq->vq_write_tree);
134 	avl_destroy(&vq->vq_pending_tree);
135 
136 	mutex_destroy(&vq->vq_lock);
137 }
138 
139 static void
140 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
141 {
142 	avl_add(&vq->vq_deadline_tree, zio);
143 	avl_add(zio->io_vdev_tree, zio);
144 }
145 
146 static void
147 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
148 {
149 	avl_remove(&vq->vq_deadline_tree, zio);
150 	avl_remove(zio->io_vdev_tree, zio);
151 }
152 
153 static void
154 vdev_queue_agg_io_done(zio_t *aio)
155 {
156 	zio_t *pio;
157 
158 	while ((pio = zio_walk_parents(aio)) != NULL)
159 		if (aio->io_type == ZIO_TYPE_READ)
160 			bcopy((char *)aio->io_data + (pio->io_offset -
161 			    aio->io_offset), pio->io_data, pio->io_size);
162 
163 	zio_buf_free(aio->io_data, aio->io_size);
164 }
165 
166 /*
167  * Compute the range spanned by two i/os, which is the endpoint of the last
168  * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
169  * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
170  * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
171  */
172 #define	IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
173 #define	IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
174 
175 static zio_t *
176 vdev_queue_io_to_issue(vdev_queue_t *vq, uint64_t pending_limit)
177 {
178 	zio_t *fio, *lio, *aio, *dio, *nio, *mio;
179 	avl_tree_t *t;
180 	int flags;
181 	uint64_t maxspan = zfs_vdev_aggregation_limit;
182 	uint64_t maxgap;
183 	int stretch;
184 
185 again:
186 	ASSERT(MUTEX_HELD(&vq->vq_lock));
187 
188 	if (avl_numnodes(&vq->vq_pending_tree) >= pending_limit ||
189 	    avl_numnodes(&vq->vq_deadline_tree) == 0)
190 		return (NULL);
191 
192 	fio = lio = avl_first(&vq->vq_deadline_tree);
193 
194 	t = fio->io_vdev_tree;
195 	flags = fio->io_flags & ZIO_FLAG_AGG_INHERIT;
196 	maxgap = (t == &vq->vq_read_tree) ? zfs_vdev_read_gap_limit : 0;
197 
198 	if (!(flags & ZIO_FLAG_DONT_AGGREGATE)) {
199 		/*
200 		 * We can aggregate I/Os that are sufficiently adjacent and of
201 		 * the same flavor, as expressed by the AGG_INHERIT flags.
202 		 * The latter requirement is necessary so that certain
203 		 * attributes of the I/O, such as whether it's a normal I/O
204 		 * or a scrub/resilver, can be preserved in the aggregate.
205 		 * We can include optional I/Os, but don't allow them
206 		 * to begin a range as they add no benefit in that situation.
207 		 */
208 
209 		/*
210 		 * We keep track of the last non-optional I/O.
211 		 */
212 		mio = (fio->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : fio;
213 
214 		/*
215 		 * Walk backwards through sufficiently contiguous I/Os
216 		 * recording the last non-option I/O.
217 		 */
218 		while ((dio = AVL_PREV(t, fio)) != NULL &&
219 		    (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
220 		    IO_SPAN(dio, lio) <= maxspan &&
221 		    IO_GAP(dio, fio) <= maxgap) {
222 			fio = dio;
223 			if (mio == NULL && !(fio->io_flags & ZIO_FLAG_OPTIONAL))
224 				mio = fio;
225 		}
226 
227 		/*
228 		 * Skip any initial optional I/Os.
229 		 */
230 		while ((fio->io_flags & ZIO_FLAG_OPTIONAL) && fio != lio) {
231 			fio = AVL_NEXT(t, fio);
232 			ASSERT(fio != NULL);
233 		}
234 
235 		/*
236 		 * Walk forward through sufficiently contiguous I/Os.
237 		 */
238 		while ((dio = AVL_NEXT(t, lio)) != NULL &&
239 		    (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
240 		    IO_SPAN(fio, dio) <= maxspan &&
241 		    IO_GAP(lio, dio) <= maxgap) {
242 			lio = dio;
243 			if (!(lio->io_flags & ZIO_FLAG_OPTIONAL))
244 				mio = lio;
245 		}
246 
247 		/*
248 		 * Now that we've established the range of the I/O aggregation
249 		 * we must decide what to do with trailing optional I/Os.
250 		 * For reads, there's nothing to do. While we are unable to
251 		 * aggregate further, it's possible that a trailing optional
252 		 * I/O would allow the underlying device to aggregate with
253 		 * subsequent I/Os. We must therefore determine if the next
254 		 * non-optional I/O is close enough to make aggregation
255 		 * worthwhile.
256 		 */
257 		stretch = B_FALSE;
258 		if (t != &vq->vq_read_tree && mio != NULL) {
259 			nio = lio;
260 			while ((dio = AVL_NEXT(t, nio)) != NULL &&
261 			    IO_GAP(nio, dio) == 0 &&
262 			    IO_GAP(mio, dio) <= zfs_vdev_write_gap_limit) {
263 				nio = dio;
264 				if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
265 					stretch = B_TRUE;
266 					break;
267 				}
268 			}
269 		}
270 
271 		if (stretch) {
272 			/* This may be a no-op. */
273 			VERIFY((dio = AVL_NEXT(t, lio)) != NULL);
274 			dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
275 		} else {
276 			while (lio != mio && lio != fio) {
277 				ASSERT(lio->io_flags & ZIO_FLAG_OPTIONAL);
278 				lio = AVL_PREV(t, lio);
279 				ASSERT(lio != NULL);
280 			}
281 		}
282 	}
283 
284 	if (fio != lio) {
285 		uint64_t size = IO_SPAN(fio, lio);
286 		ASSERT(size <= zfs_vdev_aggregation_limit);
287 
288 		aio = zio_vdev_delegated_io(fio->io_vd, fio->io_offset,
289 		    zio_buf_alloc(size), size, fio->io_type, ZIO_PRIORITY_NOW,
290 		    flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
291 		    vdev_queue_agg_io_done, NULL);
292 
293 		nio = fio;
294 		do {
295 			dio = nio;
296 			nio = AVL_NEXT(t, dio);
297 			ASSERT(dio->io_type == aio->io_type);
298 			ASSERT(dio->io_vdev_tree == t);
299 
300 			if (dio->io_flags & ZIO_FLAG_NODATA) {
301 				ASSERT(dio->io_type == ZIO_TYPE_WRITE);
302 				bzero((char *)aio->io_data + (dio->io_offset -
303 				    aio->io_offset), dio->io_size);
304 			} else if (dio->io_type == ZIO_TYPE_WRITE) {
305 				bcopy(dio->io_data, (char *)aio->io_data +
306 				    (dio->io_offset - aio->io_offset),
307 				    dio->io_size);
308 			}
309 
310 			zio_add_child(dio, aio);
311 			vdev_queue_io_remove(vq, dio);
312 			zio_vdev_io_bypass(dio);
313 			zio_execute(dio);
314 		} while (dio != lio);
315 
316 		avl_add(&vq->vq_pending_tree, aio);
317 
318 		return (aio);
319 	}
320 
321 	ASSERT(fio->io_vdev_tree == t);
322 	vdev_queue_io_remove(vq, fio);
323 
324 	/*
325 	 * If the I/O is or was optional and therefore has no data, we need to
326 	 * simply discard it. We need to drop the vdev queue's lock to avoid a
327 	 * deadlock that we could encounter since this I/O will complete
328 	 * immediately.
329 	 */
330 	if (fio->io_flags & ZIO_FLAG_NODATA) {
331 		mutex_exit(&vq->vq_lock);
332 		zio_vdev_io_bypass(fio);
333 		zio_execute(fio);
334 		mutex_enter(&vq->vq_lock);
335 		goto again;
336 	}
337 
338 	avl_add(&vq->vq_pending_tree, fio);
339 
340 	return (fio);
341 }
342 
343 zio_t *
344 vdev_queue_io(zio_t *zio)
345 {
346 	vdev_queue_t *vq = &zio->io_vd->vdev_queue;
347 	zio_t *nio;
348 
349 	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
350 
351 	if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
352 		return (zio);
353 
354 	zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
355 
356 	if (zio->io_type == ZIO_TYPE_READ)
357 		zio->io_vdev_tree = &vq->vq_read_tree;
358 	else
359 		zio->io_vdev_tree = &vq->vq_write_tree;
360 
361 	mutex_enter(&vq->vq_lock);
362 
363 	zio->io_deadline = (lbolt64 >> zfs_vdev_time_shift) + zio->io_priority;
364 
365 	vdev_queue_io_add(vq, zio);
366 
367 	nio = vdev_queue_io_to_issue(vq, zfs_vdev_min_pending);
368 
369 	mutex_exit(&vq->vq_lock);
370 
371 	if (nio == NULL)
372 		return (NULL);
373 
374 	if (nio->io_done == vdev_queue_agg_io_done) {
375 		zio_nowait(nio);
376 		return (NULL);
377 	}
378 
379 	return (nio);
380 }
381 
382 void
383 vdev_queue_io_done(zio_t *zio)
384 {
385 	vdev_queue_t *vq = &zio->io_vd->vdev_queue;
386 
387 	mutex_enter(&vq->vq_lock);
388 
389 	avl_remove(&vq->vq_pending_tree, zio);
390 
391 	for (int i = 0; i < zfs_vdev_ramp_rate; i++) {
392 		zio_t *nio = vdev_queue_io_to_issue(vq, zfs_vdev_max_pending);
393 		if (nio == NULL)
394 			break;
395 		mutex_exit(&vq->vq_lock);
396 		if (nio->io_done == vdev_queue_agg_io_done) {
397 			zio_nowait(nio);
398 		} else {
399 			zio_vdev_io_reissue(nio);
400 			zio_execute(nio);
401 		}
402 		mutex_enter(&vq->vq_lock);
403 	}
404 
405 	mutex_exit(&vq->vq_lock);
406 }
407