xref: /titanic_41/usr/src/uts/common/fs/zfs/vdev_cache.c (revision 2b24ab6b3865caeede9eeb9db6b83e1d89dcd1ea)
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.h>
28 #include <sys/vdev_impl.h>
29 #include <sys/zio.h>
30 #include <sys/kstat.h>
31 
32 /*
33  * Virtual device read-ahead caching.
34  *
35  * This file implements a simple LRU read-ahead cache.  When the DMU reads
36  * a given block, it will often want other, nearby blocks soon thereafter.
37  * We take advantage of this by reading a larger disk region and caching
38  * the result.  In the best case, this can turn 128 back-to-back 512-byte
39  * reads into a single 64k read followed by 127 cache hits; this reduces
40  * latency dramatically.  In the worst case, it can turn an isolated 512-byte
41  * read into a 64k read, which doesn't affect latency all that much but is
42  * terribly wasteful of bandwidth.  A more intelligent version of the cache
43  * could keep track of access patterns and not do read-ahead unless it sees
44  * at least two temporally close I/Os to the same region.  Currently, only
45  * metadata I/O is inflated.  A futher enhancement could take advantage of
46  * more semantic information about the I/O.  And it could use something
47  * 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;			/* 16KB */
76 int zfs_vdev_cache_size = 10ULL << 20;		/* 10MB */
77 int zfs_vdev_cache_bshift = 16;
78 
79 #define	VCBS (1 << zfs_vdev_cache_bshift)	/* 64KB */
80 
81 kstat_t	*vdc_ksp = NULL;
82 
83 typedef struct vdc_stats {
84 	kstat_named_t vdc_stat_delegations;
85 	kstat_named_t vdc_stat_hits;
86 	kstat_named_t vdc_stat_misses;
87 } vdc_stats_t;
88 
89 static vdc_stats_t vdc_stats = {
90 	{ "delegations",	KSTAT_DATA_UINT64 },
91 	{ "hits",		KSTAT_DATA_UINT64 },
92 	{ "misses",		KSTAT_DATA_UINT64 }
93 };
94 
95 #define	VDCSTAT_BUMP(stat)	atomic_add_64(&vdc_stats.stat.value.ui64, 1);
96 
97 static int
98 vdev_cache_offset_compare(const void *a1, const void *a2)
99 {
100 	const vdev_cache_entry_t *ve1 = a1;
101 	const vdev_cache_entry_t *ve2 = a2;
102 
103 	if (ve1->ve_offset < ve2->ve_offset)
104 		return (-1);
105 	if (ve1->ve_offset > ve2->ve_offset)
106 		return (1);
107 	return (0);
108 }
109 
110 static int
111 vdev_cache_lastused_compare(const void *a1, const void *a2)
112 {
113 	const vdev_cache_entry_t *ve1 = a1;
114 	const vdev_cache_entry_t *ve2 = a2;
115 
116 	if (ve1->ve_lastused < ve2->ve_lastused)
117 		return (-1);
118 	if (ve1->ve_lastused > ve2->ve_lastused)
119 		return (1);
120 
121 	/*
122 	 * Among equally old entries, sort by offset to ensure uniqueness.
123 	 */
124 	return (vdev_cache_offset_compare(a1, a2));
125 }
126 
127 /*
128  * Evict the specified entry from the cache.
129  */
130 static void
131 vdev_cache_evict(vdev_cache_t *vc, vdev_cache_entry_t *ve)
132 {
133 	ASSERT(MUTEX_HELD(&vc->vc_lock));
134 	ASSERT(ve->ve_fill_io == NULL);
135 	ASSERT(ve->ve_data != NULL);
136 
137 	avl_remove(&vc->vc_lastused_tree, ve);
138 	avl_remove(&vc->vc_offset_tree, ve);
139 	zio_buf_free(ve->ve_data, VCBS);
140 	kmem_free(ve, sizeof (vdev_cache_entry_t));
141 }
142 
143 /*
144  * Allocate an entry in the cache.  At the point we don't have the data,
145  * we're just creating a placeholder so that multiple threads don't all
146  * go off and read the same blocks.
147  */
148 static vdev_cache_entry_t *
149 vdev_cache_allocate(zio_t *zio)
150 {
151 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
152 	uint64_t offset = P2ALIGN(zio->io_offset, VCBS);
153 	vdev_cache_entry_t *ve;
154 
155 	ASSERT(MUTEX_HELD(&vc->vc_lock));
156 
157 	if (zfs_vdev_cache_size == 0)
158 		return (NULL);
159 
160 	/*
161 	 * If adding a new entry would exceed the cache size,
162 	 * evict the oldest entry (LRU).
163 	 */
164 	if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) >
165 	    zfs_vdev_cache_size) {
166 		ve = avl_first(&vc->vc_lastused_tree);
167 		if (ve->ve_fill_io != NULL)
168 			return (NULL);
169 		ASSERT(ve->ve_hits != 0);
170 		vdev_cache_evict(vc, ve);
171 	}
172 
173 	ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_SLEEP);
174 	ve->ve_offset = offset;
175 	ve->ve_lastused = lbolt;
176 	ve->ve_data = zio_buf_alloc(VCBS);
177 
178 	avl_add(&vc->vc_offset_tree, ve);
179 	avl_add(&vc->vc_lastused_tree, ve);
180 
181 	return (ve);
182 }
183 
184 static void
185 vdev_cache_hit(vdev_cache_t *vc, vdev_cache_entry_t *ve, zio_t *zio)
186 {
187 	uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
188 
189 	ASSERT(MUTEX_HELD(&vc->vc_lock));
190 	ASSERT(ve->ve_fill_io == NULL);
191 
192 	if (ve->ve_lastused != lbolt) {
193 		avl_remove(&vc->vc_lastused_tree, ve);
194 		ve->ve_lastused = lbolt;
195 		avl_add(&vc->vc_lastused_tree, ve);
196 	}
197 
198 	ve->ve_hits++;
199 	bcopy(ve->ve_data + cache_phase, zio->io_data, zio->io_size);
200 }
201 
202 /*
203  * Fill a previously allocated cache entry with data.
204  */
205 static void
206 vdev_cache_fill(zio_t *fio)
207 {
208 	vdev_t *vd = fio->io_vd;
209 	vdev_cache_t *vc = &vd->vdev_cache;
210 	vdev_cache_entry_t *ve = fio->io_private;
211 	zio_t *pio;
212 
213 	ASSERT(fio->io_size == VCBS);
214 
215 	/*
216 	 * Add data to the cache.
217 	 */
218 	mutex_enter(&vc->vc_lock);
219 
220 	ASSERT(ve->ve_fill_io == fio);
221 	ASSERT(ve->ve_offset == fio->io_offset);
222 	ASSERT(ve->ve_data == fio->io_data);
223 
224 	ve->ve_fill_io = NULL;
225 
226 	/*
227 	 * Even if this cache line was invalidated by a missed write update,
228 	 * any reads that were queued up before the missed update are still
229 	 * valid, so we can satisfy them from this line before we evict it.
230 	 */
231 	while ((pio = zio_walk_parents(fio)) != NULL)
232 		vdev_cache_hit(vc, ve, pio);
233 
234 	if (fio->io_error || ve->ve_missed_update)
235 		vdev_cache_evict(vc, ve);
236 
237 	mutex_exit(&vc->vc_lock);
238 }
239 
240 /*
241  * Read data from the cache.  Returns 0 on cache hit, errno on a miss.
242  */
243 int
244 vdev_cache_read(zio_t *zio)
245 {
246 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
247 	vdev_cache_entry_t *ve, ve_search;
248 	uint64_t cache_offset = P2ALIGN(zio->io_offset, VCBS);
249 	uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
250 	zio_t *fio;
251 
252 	ASSERT(zio->io_type == ZIO_TYPE_READ);
253 
254 	if (zio->io_flags & ZIO_FLAG_DONT_CACHE)
255 		return (EINVAL);
256 
257 	if (zio->io_size > zfs_vdev_cache_max)
258 		return (EOVERFLOW);
259 
260 	/*
261 	 * If the I/O straddles two or more cache blocks, don't cache it.
262 	 */
263 	if (P2BOUNDARY(zio->io_offset, zio->io_size, VCBS))
264 		return (EXDEV);
265 
266 	ASSERT(cache_phase + zio->io_size <= VCBS);
267 
268 	mutex_enter(&vc->vc_lock);
269 
270 	ve_search.ve_offset = cache_offset;
271 	ve = avl_find(&vc->vc_offset_tree, &ve_search, NULL);
272 
273 	if (ve != NULL) {
274 		if (ve->ve_missed_update) {
275 			mutex_exit(&vc->vc_lock);
276 			return (ESTALE);
277 		}
278 
279 		if ((fio = ve->ve_fill_io) != NULL) {
280 			zio_vdev_io_bypass(zio);
281 			zio_add_child(zio, fio);
282 			mutex_exit(&vc->vc_lock);
283 			VDCSTAT_BUMP(vdc_stat_delegations);
284 			return (0);
285 		}
286 
287 		vdev_cache_hit(vc, ve, zio);
288 		zio_vdev_io_bypass(zio);
289 
290 		mutex_exit(&vc->vc_lock);
291 		VDCSTAT_BUMP(vdc_stat_hits);
292 		return (0);
293 	}
294 
295 	ve = vdev_cache_allocate(zio);
296 
297 	if (ve == NULL) {
298 		mutex_exit(&vc->vc_lock);
299 		return (ENOMEM);
300 	}
301 
302 	fio = zio_vdev_delegated_io(zio->io_vd, cache_offset,
303 	    ve->ve_data, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_CACHE_FILL,
304 	    ZIO_FLAG_DONT_CACHE, vdev_cache_fill, ve);
305 
306 	ve->ve_fill_io = fio;
307 	zio_vdev_io_bypass(zio);
308 	zio_add_child(zio, fio);
309 
310 	mutex_exit(&vc->vc_lock);
311 	zio_nowait(fio);
312 	VDCSTAT_BUMP(vdc_stat_misses);
313 
314 	return (0);
315 }
316 
317 /*
318  * Update cache contents upon write completion.
319  */
320 void
321 vdev_cache_write(zio_t *zio)
322 {
323 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
324 	vdev_cache_entry_t *ve, ve_search;
325 	uint64_t io_start = zio->io_offset;
326 	uint64_t io_end = io_start + zio->io_size;
327 	uint64_t min_offset = P2ALIGN(io_start, VCBS);
328 	uint64_t max_offset = P2ROUNDUP(io_end, VCBS);
329 	avl_index_t where;
330 
331 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
332 
333 	mutex_enter(&vc->vc_lock);
334 
335 	ve_search.ve_offset = min_offset;
336 	ve = avl_find(&vc->vc_offset_tree, &ve_search, &where);
337 
338 	if (ve == NULL)
339 		ve = avl_nearest(&vc->vc_offset_tree, where, AVL_AFTER);
340 
341 	while (ve != NULL && ve->ve_offset < max_offset) {
342 		uint64_t start = MAX(ve->ve_offset, io_start);
343 		uint64_t end = MIN(ve->ve_offset + VCBS, io_end);
344 
345 		if (ve->ve_fill_io != NULL) {
346 			ve->ve_missed_update = 1;
347 		} else {
348 			bcopy((char *)zio->io_data + start - io_start,
349 			    ve->ve_data + start - ve->ve_offset, end - start);
350 		}
351 		ve = AVL_NEXT(&vc->vc_offset_tree, ve);
352 	}
353 	mutex_exit(&vc->vc_lock);
354 }
355 
356 void
357 vdev_cache_purge(vdev_t *vd)
358 {
359 	vdev_cache_t *vc = &vd->vdev_cache;
360 	vdev_cache_entry_t *ve;
361 
362 	mutex_enter(&vc->vc_lock);
363 	while ((ve = avl_first(&vc->vc_offset_tree)) != NULL)
364 		vdev_cache_evict(vc, ve);
365 	mutex_exit(&vc->vc_lock);
366 }
367 
368 void
369 vdev_cache_init(vdev_t *vd)
370 {
371 	vdev_cache_t *vc = &vd->vdev_cache;
372 
373 	mutex_init(&vc->vc_lock, NULL, MUTEX_DEFAULT, NULL);
374 
375 	avl_create(&vc->vc_offset_tree, vdev_cache_offset_compare,
376 	    sizeof (vdev_cache_entry_t),
377 	    offsetof(struct vdev_cache_entry, ve_offset_node));
378 
379 	avl_create(&vc->vc_lastused_tree, vdev_cache_lastused_compare,
380 	    sizeof (vdev_cache_entry_t),
381 	    offsetof(struct vdev_cache_entry, ve_lastused_node));
382 }
383 
384 void
385 vdev_cache_fini(vdev_t *vd)
386 {
387 	vdev_cache_t *vc = &vd->vdev_cache;
388 
389 	vdev_cache_purge(vd);
390 
391 	avl_destroy(&vc->vc_offset_tree);
392 	avl_destroy(&vc->vc_lastused_tree);
393 
394 	mutex_destroy(&vc->vc_lock);
395 }
396 
397 void
398 vdev_cache_stat_init(void)
399 {
400 	vdc_ksp = kstat_create("zfs", 0, "vdev_cache_stats", "misc",
401 	    KSTAT_TYPE_NAMED, sizeof (vdc_stats) / sizeof (kstat_named_t),
402 	    KSTAT_FLAG_VIRTUAL);
403 	if (vdc_ksp != NULL) {
404 		vdc_ksp->ks_data = &vdc_stats;
405 		kstat_install(vdc_ksp);
406 	}
407 }
408 
409 void
410 vdev_cache_stat_fini(void)
411 {
412 	if (vdc_ksp != NULL) {
413 		kstat_delete(vdc_ksp);
414 		vdc_ksp = NULL;
415 	}
416 }
417