xref: /illumos-gate/usr/src/uts/common/fs/zfs/vdev_cache.c (revision fdb2e90695550660e500ed27553dfc73b1f3f902)
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 2007 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.  Currently, only
46  * metadata I/O is inflated.  A futher enhancement could take advantage of
47  * more semantic information about the I/O.  And it could use something
48  * faster than an AVL tree; that was chosen solely for convenience.
49  *
50  * There are five cache operations: allocate, fill, read, write, evict.
51  *
52  * (1) Allocate.  This reserves a cache entry for the specified region.
53  *     We separate the allocate and fill operations so that multiple threads
54  *     don't generate I/O for the same cache miss.
55  *
56  * (2) Fill.  When the I/O for a cache miss completes, the fill routine
57  *     places the data in the previously allocated cache entry.
58  *
59  * (3) Read.  Read data from the cache.
60  *
61  * (4) Write.  Update cache contents after write completion.
62  *
63  * (5) Evict.  When allocating a new entry, we evict the oldest (LRU) entry
64  *     if the total cache size exceeds zfs_vdev_cache_size.
65  */
66 
67 /*
68  * These tunables are for performance analysis.
69  */
70 /*
71  * All i/os smaller than zfs_vdev_cache_max will be turned into
72  * 1<<zfs_vdev_cache_bshift byte reads by the vdev_cache (aka software
73  * track buffer.  At most zfs_vdev_cache_size bytes will be kept in each
74  * vdev's vdev_cache.
75  */
76 int zfs_vdev_cache_max = 1<<14;
77 int zfs_vdev_cache_size = 10ULL << 20;
78 int zfs_vdev_cache_bshift = 16;
79 
80 #define	VCBS (1 << zfs_vdev_cache_bshift)
81 
82 static int
83 vdev_cache_offset_compare(const void *a1, const void *a2)
84 {
85 	const vdev_cache_entry_t *ve1 = a1;
86 	const vdev_cache_entry_t *ve2 = a2;
87 
88 	if (ve1->ve_offset < ve2->ve_offset)
89 		return (-1);
90 	if (ve1->ve_offset > ve2->ve_offset)
91 		return (1);
92 	return (0);
93 }
94 
95 static int
96 vdev_cache_lastused_compare(const void *a1, const void *a2)
97 {
98 	const vdev_cache_entry_t *ve1 = a1;
99 	const vdev_cache_entry_t *ve2 = a2;
100 
101 	if (ve1->ve_lastused < ve2->ve_lastused)
102 		return (-1);
103 	if (ve1->ve_lastused > ve2->ve_lastused)
104 		return (1);
105 
106 	/*
107 	 * Among equally old entries, sort by offset to ensure uniqueness.
108 	 */
109 	return (vdev_cache_offset_compare(a1, a2));
110 }
111 
112 /*
113  * Evict the specified entry from the cache.
114  */
115 static void
116 vdev_cache_evict(vdev_cache_t *vc, vdev_cache_entry_t *ve)
117 {
118 	ASSERT(MUTEX_HELD(&vc->vc_lock));
119 	ASSERT(ve->ve_fill_io == NULL);
120 	ASSERT(ve->ve_data != NULL);
121 
122 	dprintf("evicting %p, off %llx, LRU %llu, age %lu, hits %u, stale %u\n",
123 	    vc, ve->ve_offset, ve->ve_lastused, lbolt - ve->ve_lastused,
124 	    ve->ve_hits, ve->ve_missed_update);
125 
126 	avl_remove(&vc->vc_lastused_tree, ve);
127 	avl_remove(&vc->vc_offset_tree, ve);
128 	zio_buf_free(ve->ve_data, VCBS);
129 	kmem_free(ve, sizeof (vdev_cache_entry_t));
130 }
131 
132 /*
133  * Allocate an entry in the cache.  At the point we don't have the data,
134  * we're just creating a placeholder so that multiple threads don't all
135  * go off and read the same blocks.
136  */
137 static vdev_cache_entry_t *
138 vdev_cache_allocate(zio_t *zio)
139 {
140 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
141 	uint64_t offset = P2ALIGN(zio->io_offset, VCBS);
142 	vdev_cache_entry_t *ve;
143 
144 	ASSERT(MUTEX_HELD(&vc->vc_lock));
145 
146 	if (zfs_vdev_cache_size == 0)
147 		return (NULL);
148 
149 	/*
150 	 * If adding a new entry would exceed the cache size,
151 	 * evict the oldest entry (LRU).
152 	 */
153 	if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) >
154 	    zfs_vdev_cache_size) {
155 		ve = avl_first(&vc->vc_lastused_tree);
156 		if (ve->ve_fill_io != NULL) {
157 			dprintf("can't evict in %p, still filling\n", vc);
158 			return (NULL);
159 		}
160 		ASSERT(ve->ve_hits != 0);
161 		vdev_cache_evict(vc, ve);
162 	}
163 
164 	ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_SLEEP);
165 	ve->ve_offset = offset;
166 	ve->ve_lastused = lbolt;
167 	ve->ve_data = zio_buf_alloc(VCBS);
168 
169 	avl_add(&vc->vc_offset_tree, ve);
170 	avl_add(&vc->vc_lastused_tree, ve);
171 
172 	return (ve);
173 }
174 
175 static void
176 vdev_cache_hit(vdev_cache_t *vc, vdev_cache_entry_t *ve, zio_t *zio)
177 {
178 	uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
179 
180 	ASSERT(MUTEX_HELD(&vc->vc_lock));
181 	ASSERT(ve->ve_fill_io == NULL);
182 
183 	if (ve->ve_lastused != lbolt) {
184 		avl_remove(&vc->vc_lastused_tree, ve);
185 		ve->ve_lastused = lbolt;
186 		avl_add(&vc->vc_lastused_tree, ve);
187 	}
188 
189 	ve->ve_hits++;
190 	bcopy(ve->ve_data + cache_phase, zio->io_data, zio->io_size);
191 }
192 
193 /*
194  * Fill a previously allocated cache entry with data.
195  */
196 static void
197 vdev_cache_fill(zio_t *zio)
198 {
199 	vdev_t *vd = zio->io_vd;
200 	vdev_cache_t *vc = &vd->vdev_cache;
201 	vdev_cache_entry_t *ve = zio->io_private;
202 	zio_t *dio;
203 
204 	ASSERT(zio->io_size == VCBS);
205 
206 	/*
207 	 * Add data to the cache.
208 	 */
209 	mutex_enter(&vc->vc_lock);
210 
211 	ASSERT(ve->ve_fill_io == zio);
212 	ASSERT(ve->ve_offset == zio->io_offset);
213 	ASSERT(ve->ve_data == zio->io_data);
214 
215 	ve->ve_fill_io = NULL;
216 
217 	/*
218 	 * Even if this cache line was invalidated by a missed write update,
219 	 * any reads that were queued up before the missed update are still
220 	 * valid, so we can satisfy them from this line before we evict it.
221 	 */
222 	for (dio = zio->io_delegate_list; dio; dio = dio->io_delegate_next)
223 		vdev_cache_hit(vc, ve, dio);
224 
225 	if (zio->io_error || ve->ve_missed_update)
226 		vdev_cache_evict(vc, ve);
227 
228 	mutex_exit(&vc->vc_lock);
229 
230 	while ((dio = zio->io_delegate_list) != NULL) {
231 		zio->io_delegate_list = dio->io_delegate_next;
232 		dio->io_delegate_next = NULL;
233 		dio->io_error = zio->io_error;
234 		zio_next_stage(dio);
235 	}
236 }
237 
238 /*
239  * Read data from the cache.  Returns 0 on cache hit, errno on a miss.
240  */
241 int
242 vdev_cache_read(zio_t *zio)
243 {
244 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
245 	vdev_cache_entry_t *ve, ve_search;
246 	uint64_t cache_offset = P2ALIGN(zio->io_offset, VCBS);
247 	uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
248 	zio_t *fio;
249 
250 	ASSERT(zio->io_type == ZIO_TYPE_READ);
251 
252 	if (zio->io_flags & ZIO_FLAG_DONT_CACHE)
253 		return (EINVAL);
254 
255 	if (zio->io_size > zfs_vdev_cache_max)
256 		return (EOVERFLOW);
257 
258 	/*
259 	 * If the I/O straddles two or more cache blocks, don't cache it.
260 	 */
261 	if (P2CROSS(zio->io_offset, zio->io_offset + zio->io_size - 1, VCBS))
262 		return (EXDEV);
263 
264 	ASSERT(cache_phase + zio->io_size <= VCBS);
265 
266 	mutex_enter(&vc->vc_lock);
267 
268 	ve_search.ve_offset = cache_offset;
269 	ve = avl_find(&vc->vc_offset_tree, &ve_search, NULL);
270 
271 	if (ve != NULL) {
272 		if (ve->ve_missed_update) {
273 			mutex_exit(&vc->vc_lock);
274 			return (ESTALE);
275 		}
276 
277 		if ((fio = ve->ve_fill_io) != NULL) {
278 			zio->io_delegate_next = fio->io_delegate_list;
279 			fio->io_delegate_list = zio;
280 			zio_vdev_io_bypass(zio);
281 			mutex_exit(&vc->vc_lock);
282 			return (0);
283 		}
284 
285 		vdev_cache_hit(vc, ve, zio);
286 		zio_vdev_io_bypass(zio);
287 
288 		mutex_exit(&vc->vc_lock);
289 		zio_next_stage(zio);
290 		return (0);
291 	}
292 
293 	if (!(zio->io_flags & ZIO_FLAG_METADATA)) {
294 		mutex_exit(&vc->vc_lock);
295 		return (EINVAL);
296 	}
297 
298 	ve = vdev_cache_allocate(zio);
299 
300 	if (ve == NULL) {
301 		mutex_exit(&vc->vc_lock);
302 		return (ENOMEM);
303 	}
304 
305 	fio = zio_vdev_child_io(zio, NULL, zio->io_vd, cache_offset,
306 	    ve->ve_data, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_CACHE_FILL,
307 	    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_PROPAGATE |
308 	    ZIO_FLAG_DONT_RETRY | ZIO_FLAG_NOBOOKMARK,
309 	    vdev_cache_fill, ve);
310 
311 	ve->ve_fill_io = fio;
312 	fio->io_delegate_list = zio;
313 	zio_vdev_io_bypass(zio);
314 
315 	mutex_exit(&vc->vc_lock);
316 	zio_nowait(fio);
317 
318 	return (0);
319 }
320 
321 /*
322  * Update cache contents upon write completion.
323  */
324 void
325 vdev_cache_write(zio_t *zio)
326 {
327 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
328 	vdev_cache_entry_t *ve, ve_search;
329 	uint64_t io_start = zio->io_offset;
330 	uint64_t io_end = io_start + zio->io_size;
331 	uint64_t min_offset = P2ALIGN(io_start, VCBS);
332 	uint64_t max_offset = P2ROUNDUP(io_end, VCBS);
333 	avl_index_t where;
334 
335 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
336 
337 	mutex_enter(&vc->vc_lock);
338 
339 	ve_search.ve_offset = min_offset;
340 	ve = avl_find(&vc->vc_offset_tree, &ve_search, &where);
341 
342 	if (ve == NULL)
343 		ve = avl_nearest(&vc->vc_offset_tree, where, AVL_AFTER);
344 
345 	while (ve != NULL && ve->ve_offset < max_offset) {
346 		uint64_t start = MAX(ve->ve_offset, io_start);
347 		uint64_t end = MIN(ve->ve_offset + VCBS, io_end);
348 
349 		if (ve->ve_fill_io != NULL) {
350 			ve->ve_missed_update = 1;
351 		} else {
352 			bcopy((char *)zio->io_data + start - io_start,
353 			    ve->ve_data + start - ve->ve_offset, end - start);
354 		}
355 		ve = AVL_NEXT(&vc->vc_offset_tree, ve);
356 	}
357 	mutex_exit(&vc->vc_lock);
358 }
359 
360 void
361 vdev_cache_purge(vdev_t *vd)
362 {
363 	vdev_cache_t *vc = &vd->vdev_cache;
364 	vdev_cache_entry_t *ve;
365 
366 	mutex_enter(&vc->vc_lock);
367 	while ((ve = avl_first(&vc->vc_offset_tree)) != NULL)
368 		vdev_cache_evict(vc, ve);
369 	mutex_exit(&vc->vc_lock);
370 }
371 
372 void
373 vdev_cache_init(vdev_t *vd)
374 {
375 	vdev_cache_t *vc = &vd->vdev_cache;
376 
377 	mutex_init(&vc->vc_lock, NULL, MUTEX_DEFAULT, NULL);
378 
379 	avl_create(&vc->vc_offset_tree, vdev_cache_offset_compare,
380 	    sizeof (vdev_cache_entry_t),
381 	    offsetof(struct vdev_cache_entry, ve_offset_node));
382 
383 	avl_create(&vc->vc_lastused_tree, vdev_cache_lastused_compare,
384 	    sizeof (vdev_cache_entry_t),
385 	    offsetof(struct vdev_cache_entry, ve_lastused_node));
386 }
387 
388 void
389 vdev_cache_fini(vdev_t *vd)
390 {
391 	vdev_cache_t *vc = &vd->vdev_cache;
392 
393 	vdev_cache_purge(vd);
394 
395 	avl_destroy(&vc->vc_offset_tree);
396 	avl_destroy(&vc->vc_lastused_tree);
397 
398 	mutex_destroy(&vc->vc_lock);
399 }
400