xref: /titanic_41/usr/src/uts/common/fs/zfs/vdev_cache.c (revision bf56214c0556fa6864189c826d39dbe156bb22a0)
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.  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 zfs_vdev_cache_size.
64  */
65 
66 /*
67  * These tunables are for performance analysis.
68  */
69 /*
70  * All i/os smaller than zfs_vdev_cache_max will be turned into
71  * 1<<zfs_vdev_cache_bshift byte reads by the vdev_cache (aka software
72  * track buffer.  At most zfs_vdev_cache_size bytes will be kept in each
73  * vdev's vdev_cache.
74  */
75 int zfs_vdev_cache_max = 1<<14;
76 int zfs_vdev_cache_size = 10ULL << 20;
77 int zfs_vdev_cache_bshift = 16;
78 
79 #define	VCBS (1 << zfs_vdev_cache_bshift)
80 
81 static int
82 vdev_cache_offset_compare(const void *a1, const void *a2)
83 {
84 	const vdev_cache_entry_t *ve1 = a1;
85 	const vdev_cache_entry_t *ve2 = a2;
86 
87 	if (ve1->ve_offset < ve2->ve_offset)
88 		return (-1);
89 	if (ve1->ve_offset > ve2->ve_offset)
90 		return (1);
91 	return (0);
92 }
93 
94 static int
95 vdev_cache_lastused_compare(const void *a1, const void *a2)
96 {
97 	const vdev_cache_entry_t *ve1 = a1;
98 	const vdev_cache_entry_t *ve2 = a2;
99 
100 	if (ve1->ve_lastused < ve2->ve_lastused)
101 		return (-1);
102 	if (ve1->ve_lastused > ve2->ve_lastused)
103 		return (1);
104 
105 	/*
106 	 * Among equally old entries, sort by offset to ensure uniqueness.
107 	 */
108 	return (vdev_cache_offset_compare(a1, a2));
109 }
110 
111 /*
112  * Evict the specified entry from the cache.
113  */
114 static void
115 vdev_cache_evict(vdev_cache_t *vc, vdev_cache_entry_t *ve)
116 {
117 	ASSERT(MUTEX_HELD(&vc->vc_lock));
118 	ASSERT(ve->ve_fill_io == NULL);
119 	ASSERT(ve->ve_data != NULL);
120 
121 	dprintf("evicting %p, off %llx, LRU %llu, age %lu, hits %u, stale %u\n",
122 	    vc, ve->ve_offset, ve->ve_lastused, lbolt - ve->ve_lastused,
123 	    ve->ve_hits, ve->ve_missed_update);
124 
125 	avl_remove(&vc->vc_lastused_tree, ve);
126 	avl_remove(&vc->vc_offset_tree, ve);
127 	zio_buf_free(ve->ve_data, VCBS);
128 	kmem_free(ve, sizeof (vdev_cache_entry_t));
129 }
130 
131 /*
132  * Allocate an entry in the cache.  At the point we don't have the data,
133  * we're just creating a placeholder so that multiple threads don't all
134  * go off and read the same blocks.
135  */
136 static vdev_cache_entry_t *
137 vdev_cache_allocate(zio_t *zio)
138 {
139 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
140 	uint64_t offset = P2ALIGN(zio->io_offset, VCBS);
141 	vdev_cache_entry_t *ve;
142 
143 	ASSERT(MUTEX_HELD(&vc->vc_lock));
144 
145 	if (zfs_vdev_cache_size == 0)
146 		return (NULL);
147 
148 	/*
149 	 * If adding a new entry would exceed the cache size,
150 	 * evict the oldest entry (LRU).
151 	 */
152 	if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) >
153 	    zfs_vdev_cache_size) {
154 		ve = avl_first(&vc->vc_lastused_tree);
155 		if (ve->ve_fill_io != NULL) {
156 			dprintf("can't evict in %p, still filling\n", vc);
157 			return (NULL);
158 		}
159 		ASSERT(ve->ve_hits != 0);
160 		vdev_cache_evict(vc, ve);
161 	}
162 
163 	ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_SLEEP);
164 	ve->ve_offset = offset;
165 	ve->ve_lastused = lbolt;
166 	ve->ve_data = zio_buf_alloc(VCBS);
167 
168 	avl_add(&vc->vc_offset_tree, ve);
169 	avl_add(&vc->vc_lastused_tree, ve);
170 
171 	return (ve);
172 }
173 
174 static void
175 vdev_cache_hit(vdev_cache_t *vc, vdev_cache_entry_t *ve, zio_t *zio)
176 {
177 	uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
178 
179 	ASSERT(MUTEX_HELD(&vc->vc_lock));
180 	ASSERT(ve->ve_fill_io == NULL);
181 
182 	if (ve->ve_lastused != lbolt) {
183 		avl_remove(&vc->vc_lastused_tree, ve);
184 		ve->ve_lastused = lbolt;
185 		avl_add(&vc->vc_lastused_tree, ve);
186 	}
187 
188 	ve->ve_hits++;
189 	bcopy(ve->ve_data + cache_phase, zio->io_data, zio->io_size);
190 }
191 
192 /*
193  * Fill a previously allocated cache entry with data.
194  */
195 static void
196 vdev_cache_fill(zio_t *zio)
197 {
198 	vdev_t *vd = zio->io_vd;
199 	vdev_cache_t *vc = &vd->vdev_cache;
200 	vdev_cache_entry_t *ve = zio->io_private;
201 	zio_t *dio;
202 
203 	ASSERT(zio->io_size == VCBS);
204 
205 	/*
206 	 * Add data to the cache.
207 	 */
208 	mutex_enter(&vc->vc_lock);
209 
210 	ASSERT(ve->ve_fill_io == zio);
211 	ASSERT(ve->ve_offset == zio->io_offset);
212 	ASSERT(ve->ve_data == zio->io_data);
213 
214 	ve->ve_fill_io = NULL;
215 
216 	/*
217 	 * Even if this cache line was invalidated by a missed write update,
218 	 * any reads that were queued up before the missed update are still
219 	 * valid, so we can satisfy them from this line before we evict it.
220 	 */
221 	for (dio = zio->io_delegate_list; dio; dio = dio->io_delegate_next)
222 		vdev_cache_hit(vc, ve, dio);
223 
224 	if (zio->io_error || ve->ve_missed_update)
225 		vdev_cache_evict(vc, ve);
226 
227 	mutex_exit(&vc->vc_lock);
228 
229 	while ((dio = zio->io_delegate_list) != NULL) {
230 		zio->io_delegate_list = dio->io_delegate_next;
231 		dio->io_delegate_next = NULL;
232 		dio->io_error = zio->io_error;
233 		zio_next_stage(dio);
234 	}
235 }
236 
237 /*
238  * Read data from the cache.  Returns 0 on cache hit, errno on a miss.
239  */
240 int
241 vdev_cache_read(zio_t *zio)
242 {
243 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
244 	vdev_cache_entry_t *ve, ve_search;
245 	uint64_t cache_offset = P2ALIGN(zio->io_offset, VCBS);
246 	uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
247 	zio_t *fio;
248 
249 	ASSERT(zio->io_type == ZIO_TYPE_READ);
250 
251 	if (zio->io_flags & ZIO_FLAG_DONT_CACHE)
252 		return (EINVAL);
253 
254 	if (zio->io_size > zfs_vdev_cache_max)
255 		return (EOVERFLOW);
256 
257 	/*
258 	 * If the I/O straddles two or more cache blocks, don't cache it.
259 	 */
260 	if (P2CROSS(zio->io_offset, zio->io_offset + zio->io_size - 1, VCBS))
261 		return (EXDEV);
262 
263 	ASSERT(cache_phase + zio->io_size <= VCBS);
264 
265 	mutex_enter(&vc->vc_lock);
266 
267 	ve_search.ve_offset = cache_offset;
268 	ve = avl_find(&vc->vc_offset_tree, &ve_search, NULL);
269 
270 	if (ve != NULL) {
271 		if (ve->ve_missed_update) {
272 			mutex_exit(&vc->vc_lock);
273 			return (ESTALE);
274 		}
275 
276 		if ((fio = ve->ve_fill_io) != NULL) {
277 			zio->io_delegate_next = fio->io_delegate_list;
278 			fio->io_delegate_list = zio;
279 			zio_vdev_io_bypass(zio);
280 			mutex_exit(&vc->vc_lock);
281 			return (0);
282 		}
283 
284 		vdev_cache_hit(vc, ve, zio);
285 		zio_vdev_io_bypass(zio);
286 
287 		mutex_exit(&vc->vc_lock);
288 		zio_next_stage(zio);
289 		return (0);
290 	}
291 
292 	ve = vdev_cache_allocate(zio);
293 
294 	if (ve == NULL) {
295 		mutex_exit(&vc->vc_lock);
296 		return (ENOMEM);
297 	}
298 
299 	fio = zio_vdev_child_io(zio, NULL, zio->io_vd, cache_offset,
300 	    ve->ve_data, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_CACHE_FILL,
301 	    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_PROPAGATE |
302 	    ZIO_FLAG_DONT_RETRY | ZIO_FLAG_NOBOOKMARK,
303 	    vdev_cache_fill, ve);
304 
305 	ve->ve_fill_io = fio;
306 	fio->io_delegate_list = zio;
307 	zio_vdev_io_bypass(zio);
308 
309 	mutex_exit(&vc->vc_lock);
310 	zio_nowait(fio);
311 
312 	return (0);
313 }
314 
315 /*
316  * Update cache contents upon write completion.
317  */
318 void
319 vdev_cache_write(zio_t *zio)
320 {
321 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
322 	vdev_cache_entry_t *ve, ve_search;
323 	uint64_t io_start = zio->io_offset;
324 	uint64_t io_end = io_start + zio->io_size;
325 	uint64_t min_offset = P2ALIGN(io_start, VCBS);
326 	uint64_t max_offset = P2ROUNDUP(io_end, VCBS);
327 	avl_index_t where;
328 
329 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
330 
331 	mutex_enter(&vc->vc_lock);
332 
333 	ve_search.ve_offset = min_offset;
334 	ve = avl_find(&vc->vc_offset_tree, &ve_search, &where);
335 
336 	if (ve == NULL)
337 		ve = avl_nearest(&vc->vc_offset_tree, where, AVL_AFTER);
338 
339 	while (ve != NULL && ve->ve_offset < max_offset) {
340 		uint64_t start = MAX(ve->ve_offset, io_start);
341 		uint64_t end = MIN(ve->ve_offset + VCBS, io_end);
342 
343 		if (ve->ve_fill_io != NULL) {
344 			ve->ve_missed_update = 1;
345 		} else {
346 			bcopy((char *)zio->io_data + start - io_start,
347 			    ve->ve_data + start - ve->ve_offset, end - start);
348 		}
349 		ve = AVL_NEXT(&vc->vc_offset_tree, ve);
350 	}
351 	mutex_exit(&vc->vc_lock);
352 }
353 
354 void
355 vdev_cache_purge(vdev_t *vd)
356 {
357 	vdev_cache_t *vc = &vd->vdev_cache;
358 	vdev_cache_entry_t *ve;
359 
360 	mutex_enter(&vc->vc_lock);
361 	while ((ve = avl_first(&vc->vc_offset_tree)) != NULL)
362 		vdev_cache_evict(vc, ve);
363 	mutex_exit(&vc->vc_lock);
364 }
365 
366 void
367 vdev_cache_init(vdev_t *vd)
368 {
369 	vdev_cache_t *vc = &vd->vdev_cache;
370 
371 	mutex_init(&vc->vc_lock, NULL, MUTEX_DEFAULT, NULL);
372 
373 	avl_create(&vc->vc_offset_tree, vdev_cache_offset_compare,
374 	    sizeof (vdev_cache_entry_t),
375 	    offsetof(struct vdev_cache_entry, ve_offset_node));
376 
377 	avl_create(&vc->vc_lastused_tree, vdev_cache_lastused_compare,
378 	    sizeof (vdev_cache_entry_t),
379 	    offsetof(struct vdev_cache_entry, ve_lastused_node));
380 }
381 
382 void
383 vdev_cache_fini(vdev_t *vd)
384 {
385 	vdev_cache_t *vc = &vd->vdev_cache;
386 
387 	vdev_cache_purge(vd);
388 
389 	avl_destroy(&vc->vc_offset_tree);
390 	avl_destroy(&vc->vc_lastused_tree);
391 
392 	mutex_destroy(&vc->vc_lock);
393 }
394