xref: /illumos-gate/usr/src/uts/common/fs/zfs/vdev_cache.c (revision ea8dc4b6d2251b437950c0056bc626b311c73c27)
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 2006 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/zfs_context.h>
29 #include <sys/spa.h>
30 #include <sys/vdev_impl.h>
31 #include <sys/zio.h>
32 
33 /*
34  * Virtual device read-ahead caching.
35  *
36  * This file implements a simple LRU read-ahead cache.  When the DMU reads
37  * a given block, it will often want other, nearby blocks soon thereafter.
38  * We take advantage of this by reading a larger disk region and caching
39  * the result.  In the best case, this can turn 256 back-to-back 512-byte
40  * reads into a single 128k read followed by 255 cache hits; this reduces
41  * latency dramatically.  In the worst case, it can turn an isolated 512-byte
42  * read into a 128k read, which doesn't affect latency all that much but is
43  * terribly wasteful of bandwidth.  A more intelligent version of the cache
44  * could keep track of access patterns and not do read-ahead unless it sees
45  * at least two temporally close I/Os to the same region.  It could also
46  * take advantage of semantic information about the I/O.  And it could use
47  * something faster than an AVL tree; that was chosen solely for convenience.
48  *
49  * There are five cache operations: allocate, fill, read, write, evict.
50  *
51  * (1) Allocate.  This reserves a cache entry for the specified region.
52  *     We separate the allocate and fill operations so that multiple threads
53  *     don't generate I/O for the same cache miss.
54  *
55  * (2) Fill.  When the I/O for a cache miss completes, the fill routine
56  *     places the data in the previously allocated cache entry.
57  *
58  * (3) Read.  Read data from the cache.
59  *
60  * (4) Write.  Update cache contents after write completion.
61  *
62  * (5) Evict.  When allocating a new entry, we evict the oldest (LRU) entry
63  *     if the total cache size exceeds vc_size.
64  */
65 
66 static int
67 vdev_cache_offset_compare(const void *a1, const void *a2)
68 {
69 	const vdev_cache_entry_t *ve1 = a1;
70 	const vdev_cache_entry_t *ve2 = a2;
71 
72 	if (ve1->ve_offset < ve2->ve_offset)
73 		return (-1);
74 	if (ve1->ve_offset > ve2->ve_offset)
75 		return (1);
76 	return (0);
77 }
78 
79 static int
80 vdev_cache_lastused_compare(const void *a1, const void *a2)
81 {
82 	const vdev_cache_entry_t *ve1 = a1;
83 	const vdev_cache_entry_t *ve2 = a2;
84 
85 	if (ve1->ve_lastused < ve2->ve_lastused)
86 		return (-1);
87 	if (ve1->ve_lastused > ve2->ve_lastused)
88 		return (1);
89 
90 	/*
91 	 * Among equally old entries, sort by offset to ensure uniqueness.
92 	 */
93 	return (vdev_cache_offset_compare(a1, a2));
94 }
95 
96 /*
97  * Evict the specified entry from the cache.
98  */
99 static void
100 vdev_cache_evict(vdev_cache_t *vc, vdev_cache_entry_t *ve)
101 {
102 	ASSERT(MUTEX_HELD(&vc->vc_lock));
103 	ASSERT(ve->ve_fill_io == NULL);
104 	ASSERT(ve->ve_data != NULL);
105 
106 	dprintf("evicting %p, off %llx, LRU %llu, age %lu, hits %u, stale %u\n",
107 	    vc, ve->ve_offset, ve->ve_lastused, lbolt - ve->ve_lastused,
108 	    ve->ve_hits, ve->ve_missed_update);
109 
110 	avl_remove(&vc->vc_lastused_tree, ve);
111 	avl_remove(&vc->vc_offset_tree, ve);
112 	zio_buf_free(ve->ve_data, vc->vc_blocksize);
113 	kmem_free(ve, sizeof (vdev_cache_entry_t));
114 }
115 
116 /*
117  * Allocate an entry in the cache.  At the point we don't have the data,
118  * we're just creating a placeholder so that multiple threads don't all
119  * go off and read the same blocks.
120  */
121 static vdev_cache_entry_t *
122 vdev_cache_allocate(zio_t *zio)
123 {
124 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
125 	uint64_t offset = P2ALIGN(zio->io_offset, vc->vc_blocksize);
126 	vdev_cache_entry_t *ve;
127 
128 	ASSERT(MUTEX_HELD(&vc->vc_lock));
129 
130 	if (vc->vc_size == 0)
131 		return (NULL);
132 
133 	/*
134 	 * If adding a new entry would exceed the cache size,
135 	 * evict the oldest entry (LRU).
136 	 */
137 	if ((avl_numnodes(&vc->vc_lastused_tree) << vc->vc_bshift) >
138 	    vc->vc_size) {
139 		ve = avl_first(&vc->vc_lastused_tree);
140 		if (ve->ve_fill_io != NULL) {
141 			dprintf("can't evict in %p, still filling\n", vc);
142 			return (NULL);
143 		}
144 		ASSERT(ve->ve_hits != 0);
145 		vdev_cache_evict(vc, ve);
146 	}
147 
148 	ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_SLEEP);
149 	ve->ve_offset = offset;
150 	ve->ve_lastused = lbolt;
151 	ve->ve_data = zio_buf_alloc(vc->vc_blocksize);
152 
153 	avl_add(&vc->vc_offset_tree, ve);
154 	avl_add(&vc->vc_lastused_tree, ve);
155 
156 	return (ve);
157 }
158 
159 static void
160 vdev_cache_hit(vdev_cache_t *vc, vdev_cache_entry_t *ve, zio_t *zio)
161 {
162 	uint64_t cache_phase = P2PHASE(zio->io_offset, vc->vc_blocksize);
163 
164 	ASSERT(MUTEX_HELD(&vc->vc_lock));
165 	ASSERT(ve->ve_fill_io == NULL);
166 
167 	if (ve->ve_lastused != lbolt) {
168 		avl_remove(&vc->vc_lastused_tree, ve);
169 		ve->ve_lastused = lbolt;
170 		avl_add(&vc->vc_lastused_tree, ve);
171 	}
172 
173 	ve->ve_hits++;
174 	bcopy(ve->ve_data + cache_phase, zio->io_data, zio->io_size);
175 }
176 
177 /*
178  * Fill a previously allocated cache entry with data.
179  */
180 static void
181 vdev_cache_fill(zio_t *zio)
182 {
183 	vdev_t *vd = zio->io_vd;
184 	vdev_cache_t *vc = &vd->vdev_cache;
185 	vdev_cache_entry_t *ve = zio->io_private;
186 	zio_t *dio;
187 
188 	ASSERT(zio->io_size == vc->vc_blocksize);
189 
190 	/*
191 	 * Add data to the cache.
192 	 */
193 	mutex_enter(&vc->vc_lock);
194 
195 	ASSERT(ve->ve_fill_io == zio);
196 	ASSERT(ve->ve_offset == zio->io_offset);
197 	ASSERT(ve->ve_data == zio->io_data);
198 
199 	ve->ve_fill_io = NULL;
200 
201 	/*
202 	 * Even if this cache line was invalidated by a missed write update,
203 	 * any reads that were queued up before the missed update are still
204 	 * valid, so we can satisfy them from this line before we evict it.
205 	 */
206 	for (dio = zio->io_delegate_list; dio; dio = dio->io_delegate_next)
207 		vdev_cache_hit(vc, ve, dio);
208 
209 	if (zio->io_error || ve->ve_missed_update)
210 		vdev_cache_evict(vc, ve);
211 
212 	mutex_exit(&vc->vc_lock);
213 
214 	while ((dio = zio->io_delegate_list) != NULL) {
215 		zio->io_delegate_list = dio->io_delegate_next;
216 		dio->io_delegate_next = NULL;
217 		dio->io_error = zio->io_error;
218 		zio_next_stage(dio);
219 	}
220 }
221 
222 /*
223  * Read data from the cache.  Returns 0 on cache hit, errno on a miss.
224  */
225 int
226 vdev_cache_read(zio_t *zio)
227 {
228 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
229 	vdev_cache_entry_t *ve, ve_search;
230 	uint64_t cache_offset = P2ALIGN(zio->io_offset, vc->vc_blocksize);
231 	uint64_t cache_phase = P2PHASE(zio->io_offset, vc->vc_blocksize);
232 	zio_t *fio;
233 
234 	ASSERT(zio->io_type == ZIO_TYPE_READ);
235 
236 	if (zio->io_flags & ZIO_FLAG_DONT_CACHE)
237 		return (EINVAL);
238 
239 	if (zio->io_size > vc->vc_max)
240 		return (EOVERFLOW);
241 
242 	/*
243 	 * If the I/O straddles two or more cache blocks, don't cache it.
244 	 */
245 	if (P2CROSS(zio->io_offset, zio->io_offset + zio->io_size - 1,
246 	    vc->vc_blocksize))
247 		return (EXDEV);
248 
249 	ASSERT(cache_phase + zio->io_size <= vc->vc_blocksize);
250 
251 	mutex_enter(&vc->vc_lock);
252 
253 	ve_search.ve_offset = cache_offset;
254 	ve = avl_find(&vc->vc_offset_tree, &ve_search, NULL);
255 
256 	if (ve != NULL) {
257 		if (ve->ve_missed_update) {
258 			mutex_exit(&vc->vc_lock);
259 			return (ESTALE);
260 		}
261 
262 		if ((fio = ve->ve_fill_io) != NULL) {
263 			zio->io_delegate_next = fio->io_delegate_list;
264 			fio->io_delegate_list = zio;
265 			zio_vdev_io_bypass(zio);
266 			mutex_exit(&vc->vc_lock);
267 			return (0);
268 		}
269 
270 		vdev_cache_hit(vc, ve, zio);
271 		zio_vdev_io_bypass(zio);
272 
273 		mutex_exit(&vc->vc_lock);
274 		zio_next_stage(zio);
275 		return (0);
276 	}
277 
278 	ve = vdev_cache_allocate(zio);
279 
280 	if (ve == NULL) {
281 		mutex_exit(&vc->vc_lock);
282 		return (ENOMEM);
283 	}
284 
285 	fio = zio_vdev_child_io(zio, NULL, zio->io_vd, cache_offset,
286 	    ve->ve_data, vc->vc_blocksize, ZIO_TYPE_READ,
287 	    ZIO_PRIORITY_CACHE_FILL,
288 	    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_PROPAGATE |
289 	    ZIO_FLAG_DONT_RETRY | ZIO_FLAG_NOBOOKMARK,
290 	    vdev_cache_fill, ve);
291 
292 	ve->ve_fill_io = fio;
293 	fio->io_delegate_list = zio;
294 	zio_vdev_io_bypass(zio);
295 
296 	mutex_exit(&vc->vc_lock);
297 	zio_nowait(fio);
298 
299 	return (0);
300 }
301 
302 /*
303  * Update cache contents upon write completion.
304  */
305 void
306 vdev_cache_write(zio_t *zio)
307 {
308 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
309 	vdev_cache_entry_t *ve, ve_search;
310 	uint64_t io_start = zio->io_offset;
311 	uint64_t io_end = io_start + zio->io_size;
312 	uint64_t min_offset = P2ALIGN(io_start, vc->vc_blocksize);
313 	uint64_t max_offset = P2ROUNDUP(io_end, vc->vc_blocksize);
314 	avl_index_t where;
315 
316 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
317 
318 	mutex_enter(&vc->vc_lock);
319 
320 	ve_search.ve_offset = min_offset;
321 	ve = avl_find(&vc->vc_offset_tree, &ve_search, &where);
322 
323 	if (ve == NULL)
324 		ve = avl_nearest(&vc->vc_offset_tree, where, AVL_AFTER);
325 
326 	while (ve != NULL && ve->ve_offset < max_offset) {
327 		uint64_t start = MAX(ve->ve_offset, io_start);
328 		uint64_t end = MIN(ve->ve_offset + vc->vc_blocksize, io_end);
329 
330 		if (ve->ve_fill_io != NULL) {
331 			ve->ve_missed_update = 1;
332 		} else {
333 			bcopy((char *)zio->io_data + start - io_start,
334 			    ve->ve_data + start - ve->ve_offset, end - start);
335 		}
336 		ve = AVL_NEXT(&vc->vc_offset_tree, ve);
337 	}
338 	mutex_exit(&vc->vc_lock);
339 }
340 
341 void
342 vdev_cache_init(vdev_t *vd)
343 {
344 	vdev_cache_t *vc = &vd->vdev_cache;
345 
346 	mutex_init(&vc->vc_lock, NULL, MUTEX_DEFAULT, NULL);
347 
348 	avl_create(&vc->vc_offset_tree, vdev_cache_offset_compare,
349 	    sizeof (vdev_cache_entry_t),
350 	    offsetof(struct vdev_cache_entry, ve_offset_node));
351 
352 	avl_create(&vc->vc_lastused_tree, vdev_cache_lastused_compare,
353 	    sizeof (vdev_cache_entry_t),
354 	    offsetof(struct vdev_cache_entry, ve_lastused_node));
355 
356 	vc->vc_blocksize = 1ULL << vc->vc_bshift;
357 }
358 
359 void
360 vdev_cache_fini(vdev_t *vd)
361 {
362 	vdev_cache_t *vc = &vd->vdev_cache;
363 	vdev_cache_entry_t *ve;
364 
365 	mutex_enter(&vc->vc_lock);
366 	while ((ve = avl_first(&vc->vc_offset_tree)) != NULL)
367 		vdev_cache_evict(vc, ve);
368 	mutex_exit(&vc->vc_lock);
369 
370 	avl_destroy(&vc->vc_offset_tree);
371 	avl_destroy(&vc->vc_lastused_tree);
372 
373 	mutex_destroy(&vc->vc_lock);
374 }
375