xref: /titanic_41/usr/src/uts/common/fs/zfs/vdev_queue.c (revision bd9f6899328e19cbb74e3ad02f5c32002285887e)
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/vdev_impl.h>
28 #include <sys/zio.h>
29 #include <sys/avl.h>
30 
31 /*
32  * These tunables are for performance analysis.
33  */
34 /*
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
38  * max_pending).
39  */
40 int zfs_vdev_max_pending = 10;
41 int zfs_vdev_min_pending = 4;
42 
43 /* deadline = pri + (lbolt >> time_shift) */
44 int zfs_vdev_time_shift = 6;
45 
46 /* exponential I/O issue ramp-up rate */
47 int zfs_vdev_ramp_rate = 2;
48 
49 /*
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.
54  */
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;
58 
59 /*
60  * Virtual device vector for disk I/O scheduling.
61  */
62 int
63 vdev_queue_deadline_compare(const void *x1, const void *x2)
64 {
65 	const zio_t *z1 = x1;
66 	const zio_t *z2 = x2;
67 
68 	if (z1->io_deadline < z2->io_deadline)
69 		return (-1);
70 	if (z1->io_deadline > z2->io_deadline)
71 		return (1);
72 
73 	if (z1->io_offset < z2->io_offset)
74 		return (-1);
75 	if (z1->io_offset > z2->io_offset)
76 		return (1);
77 
78 	if (z1 < z2)
79 		return (-1);
80 	if (z1 > z2)
81 		return (1);
82 
83 	return (0);
84 }
85 
86 int
87 vdev_queue_offset_compare(const void *x1, const void *x2)
88 {
89 	const zio_t *z1 = x1;
90 	const zio_t *z2 = x2;
91 
92 	if (z1->io_offset < z2->io_offset)
93 		return (-1);
94 	if (z1->io_offset > z2->io_offset)
95 		return (1);
96 
97 	if (z1 < z2)
98 		return (-1);
99 	if (z1 > z2)
100 		return (1);
101 
102 	return (0);
103 }
104 
105 void
106 vdev_queue_init(vdev_t *vd)
107 {
108 	vdev_queue_t *vq = &vd->vdev_queue;
109 
110 	mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
111 
112 	avl_create(&vq->vq_deadline_tree, vdev_queue_deadline_compare,
113 	    sizeof (zio_t), offsetof(struct zio, io_deadline_node));
114 
115 	avl_create(&vq->vq_read_tree, vdev_queue_offset_compare,
116 	    sizeof (zio_t), offsetof(struct zio, io_offset_node));
117 
118 	avl_create(&vq->vq_write_tree, vdev_queue_offset_compare,
119 	    sizeof (zio_t), offsetof(struct zio, io_offset_node));
120 
121 	avl_create(&vq->vq_pending_tree, vdev_queue_offset_compare,
122 	    sizeof (zio_t), offsetof(struct zio, io_offset_node));
123 }
124 
125 void
126 vdev_queue_fini(vdev_t *vd)
127 {
128 	vdev_queue_t *vq = &vd->vdev_queue;
129 
130 	avl_destroy(&vq->vq_deadline_tree);
131 	avl_destroy(&vq->vq_read_tree);
132 	avl_destroy(&vq->vq_write_tree);
133 	avl_destroy(&vq->vq_pending_tree);
134 
135 	mutex_destroy(&vq->vq_lock);
136 }
137 
138 static void
139 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
140 {
141 	avl_add(&vq->vq_deadline_tree, zio);
142 	avl_add(zio->io_vdev_tree, zio);
143 }
144 
145 static void
146 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
147 {
148 	avl_remove(&vq->vq_deadline_tree, zio);
149 	avl_remove(zio->io_vdev_tree, zio);
150 }
151 
152 static void
153 vdev_queue_agg_io_done(zio_t *aio)
154 {
155 	zio_t *pio;
156 
157 	while ((pio = zio_walk_parents(aio)) != NULL)
158 		if (aio->io_type == ZIO_TYPE_READ)
159 			bcopy((char *)aio->io_data + (pio->io_offset -
160 			    aio->io_offset), pio->io_data, pio->io_size);
161 
162 	zio_buf_free(aio->io_data, aio->io_size);
163 }
164 
165 /*
166  * Compute the range spanned by two i/os, which is the endpoint of the last
167  * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
168  * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
169  * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
170  */
171 #define	IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
172 #define	IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
173 
174 static zio_t *
175 vdev_queue_io_to_issue(vdev_queue_t *vq, uint64_t pending_limit)
176 {
177 	zio_t *fio, *lio, *aio, *dio, *nio, *mio;
178 	avl_tree_t *t;
179 	int flags;
180 	uint64_t maxspan = zfs_vdev_aggregation_limit;
181 	uint64_t maxgap;
182 	int stretch;
183 
184 again:
185 	ASSERT(MUTEX_HELD(&vq->vq_lock));
186 
187 	if (avl_numnodes(&vq->vq_pending_tree) >= pending_limit ||
188 	    avl_numnodes(&vq->vq_deadline_tree) == 0)
189 		return (NULL);
190 
191 	fio = lio = avl_first(&vq->vq_deadline_tree);
192 
193 	t = fio->io_vdev_tree;
194 	flags = fio->io_flags & ZIO_FLAG_AGG_INHERIT;
195 	maxgap = (t == &vq->vq_read_tree) ? zfs_vdev_read_gap_limit : 0;
196 
197 	if (!(flags & ZIO_FLAG_DONT_AGGREGATE)) {
198 		/*
199 		 * We can aggregate I/Os that are sufficiently adjacent and of
200 		 * the same flavor, as expressed by the AGG_INHERIT flags.
201 		 * The latter requirement is necessary so that certain
202 		 * attributes of the I/O, such as whether it's a normal I/O
203 		 * or a scrub/resilver, can be preserved in the aggregate.
204 		 * We can include optional I/Os, but don't allow them
205 		 * to begin a range as they add no benefit in that situation.
206 		 */
207 
208 		/*
209 		 * We keep track of the last non-optional I/O.
210 		 */
211 		mio = (fio->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : fio;
212 
213 		/*
214 		 * Walk backwards through sufficiently contiguous I/Os
215 		 * recording the last non-option I/O.
216 		 */
217 		while ((dio = AVL_PREV(t, fio)) != NULL &&
218 		    (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
219 		    IO_SPAN(dio, lio) <= maxspan &&
220 		    IO_GAP(dio, fio) <= maxgap) {
221 			fio = dio;
222 			if (mio == NULL && !(fio->io_flags & ZIO_FLAG_OPTIONAL))
223 				mio = fio;
224 		}
225 
226 		/*
227 		 * Skip any initial optional I/Os.
228 		 */
229 		while ((fio->io_flags & ZIO_FLAG_OPTIONAL) && fio != lio) {
230 			fio = AVL_NEXT(t, fio);
231 			ASSERT(fio != NULL);
232 		}
233 
234 		/*
235 		 * Walk forward through sufficiently contiguous I/Os.
236 		 */
237 		while ((dio = AVL_NEXT(t, lio)) != NULL &&
238 		    (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
239 		    IO_SPAN(fio, dio) <= maxspan &&
240 		    IO_GAP(lio, dio) <= maxgap) {
241 			lio = dio;
242 			if (!(lio->io_flags & ZIO_FLAG_OPTIONAL))
243 				mio = lio;
244 		}
245 
246 		/*
247 		 * Now that we've established the range of the I/O aggregation
248 		 * we must decide what to do with trailing optional I/Os.
249 		 * For reads, there's nothing to do. While we are unable to
250 		 * aggregate further, it's possible that a trailing optional
251 		 * I/O would allow the underlying device to aggregate with
252 		 * subsequent I/Os. We must therefore determine if the next
253 		 * non-optional I/O is close enough to make aggregation
254 		 * worthwhile.
255 		 */
256 		stretch = B_FALSE;
257 		if (t != &vq->vq_read_tree && mio != NULL) {
258 			nio = lio;
259 			while ((dio = AVL_NEXT(t, nio)) != NULL &&
260 			    IO_GAP(nio, dio) == 0 &&
261 			    IO_GAP(mio, dio) <= zfs_vdev_write_gap_limit) {
262 				nio = dio;
263 				if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
264 					stretch = B_TRUE;
265 					break;
266 				}
267 			}
268 		}
269 
270 		if (stretch) {
271 			/* This may be a no-op. */
272 			VERIFY((dio = AVL_NEXT(t, lio)) != NULL);
273 			dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
274 		} else {
275 			while (lio != mio && lio != fio) {
276 				ASSERT(lio->io_flags & ZIO_FLAG_OPTIONAL);
277 				lio = AVL_PREV(t, lio);
278 				ASSERT(lio != NULL);
279 			}
280 		}
281 	}
282 
283 	if (fio != lio) {
284 		uint64_t size = IO_SPAN(fio, lio);
285 		ASSERT(size <= zfs_vdev_aggregation_limit);
286 
287 		aio = zio_vdev_delegated_io(fio->io_vd, fio->io_offset,
288 		    zio_buf_alloc(size), size, fio->io_type, ZIO_PRIORITY_NOW,
289 		    flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
290 		    vdev_queue_agg_io_done, NULL);
291 
292 		nio = fio;
293 		do {
294 			dio = nio;
295 			nio = AVL_NEXT(t, dio);
296 			ASSERT(dio->io_type == aio->io_type);
297 			ASSERT(dio->io_vdev_tree == t);
298 
299 			if (dio->io_flags & ZIO_FLAG_NODATA) {
300 				ASSERT(dio->io_type == ZIO_TYPE_WRITE);
301 				bzero((char *)aio->io_data + (dio->io_offset -
302 				    aio->io_offset), dio->io_size);
303 			} else if (dio->io_type == ZIO_TYPE_WRITE) {
304 				bcopy(dio->io_data, (char *)aio->io_data +
305 				    (dio->io_offset - aio->io_offset),
306 				    dio->io_size);
307 			}
308 
309 			zio_add_child(dio, aio);
310 			vdev_queue_io_remove(vq, dio);
311 			zio_vdev_io_bypass(dio);
312 			zio_execute(dio);
313 		} while (dio != lio);
314 
315 		avl_add(&vq->vq_pending_tree, aio);
316 
317 		return (aio);
318 	}
319 
320 	ASSERT(fio->io_vdev_tree == t);
321 	vdev_queue_io_remove(vq, fio);
322 
323 	/*
324 	 * If the I/O is or was optional and therefore has no data, we need to
325 	 * simply discard it. We need to drop the vdev queue's lock to avoid a
326 	 * deadlock that we could encounter since this I/O will complete
327 	 * immediately.
328 	 */
329 	if (fio->io_flags & ZIO_FLAG_NODATA) {
330 		mutex_exit(&vq->vq_lock);
331 		zio_vdev_io_bypass(fio);
332 		zio_execute(fio);
333 		mutex_enter(&vq->vq_lock);
334 		goto again;
335 	}
336 
337 	avl_add(&vq->vq_pending_tree, fio);
338 
339 	return (fio);
340 }
341 
342 zio_t *
343 vdev_queue_io(zio_t *zio)
344 {
345 	vdev_queue_t *vq = &zio->io_vd->vdev_queue;
346 	zio_t *nio;
347 
348 	ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
349 
350 	if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
351 		return (zio);
352 
353 	zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
354 
355 	if (zio->io_type == ZIO_TYPE_READ)
356 		zio->io_vdev_tree = &vq->vq_read_tree;
357 	else
358 		zio->io_vdev_tree = &vq->vq_write_tree;
359 
360 	mutex_enter(&vq->vq_lock);
361 
362 	zio->io_deadline = (lbolt64 >> zfs_vdev_time_shift) + zio->io_priority;
363 
364 	vdev_queue_io_add(vq, zio);
365 
366 	nio = vdev_queue_io_to_issue(vq, zfs_vdev_min_pending);
367 
368 	mutex_exit(&vq->vq_lock);
369 
370 	if (nio == NULL)
371 		return (NULL);
372 
373 	if (nio->io_done == vdev_queue_agg_io_done) {
374 		zio_nowait(nio);
375 		return (NULL);
376 	}
377 
378 	return (nio);
379 }
380 
381 void
382 vdev_queue_io_done(zio_t *zio)
383 {
384 	vdev_queue_t *vq = &zio->io_vd->vdev_queue;
385 
386 	mutex_enter(&vq->vq_lock);
387 
388 	avl_remove(&vq->vq_pending_tree, zio);
389 
390 	for (int i = 0; i < zfs_vdev_ramp_rate; i++) {
391 		zio_t *nio = vdev_queue_io_to_issue(vq, zfs_vdev_max_pending);
392 		if (nio == NULL)
393 			break;
394 		mutex_exit(&vq->vq_lock);
395 		if (nio->io_done == vdev_queue_agg_io_done) {
396 			zio_nowait(nio);
397 		} else {
398 			zio_vdev_io_reissue(nio);
399 			zio_execute(nio);
400 		}
401 		mutex_enter(&vq->vq_lock);
402 	}
403 
404 	mutex_exit(&vq->vq_lock);
405 }
406