xref: /illumos-gate/usr/src/uts/common/fs/zfs/vdev_cache.c (revision 69962b5647e4a8b9b14998733b765925381b727e)
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  * Copyright (c) 2013 by Delphix. All rights reserved.
27  */
28 
29 #include <sys/zfs_context.h>
30 #include <sys/spa.h>
31 #include <sys/vdev_impl.h>
32 #include <sys/zio.h>
33 #include <sys/kstat.h>
34 
35 /*
36  * Virtual device read-ahead caching.
37  *
38  * This file implements a simple LRU read-ahead cache.  When the DMU reads
39  * a given block, it will often want other, nearby blocks soon thereafter.
40  * We take advantage of this by reading a larger disk region and caching
41  * the result.  In the best case, this can turn 128 back-to-back 512-byte
42  * reads into a single 64k read followed by 127 cache hits; this reduces
43  * latency dramatically.  In the worst case, it can turn an isolated 512-byte
44  * read into a 64k read, which doesn't affect latency all that much but is
45  * terribly wasteful of bandwidth.  A more intelligent version of the cache
46  * could keep track of access patterns and not do read-ahead unless it sees
47  * at least two temporally close I/Os to the same region.  Currently, only
48  * metadata I/O is inflated.  A futher enhancement could take advantage of
49  * more semantic information about the I/O.  And it could use something
50  * faster than an AVL tree; that was chosen solely for convenience.
51  *
52  * There are five cache operations: allocate, fill, read, write, evict.
53  *
54  * (1) Allocate.  This reserves a cache entry for the specified region.
55  *     We separate the allocate and fill operations so that multiple threads
56  *     don't generate I/O for the same cache miss.
57  *
58  * (2) Fill.  When the I/O for a cache miss completes, the fill routine
59  *     places the data in the previously allocated cache entry.
60  *
61  * (3) Read.  Read data from the cache.
62  *
63  * (4) Write.  Update cache contents after write completion.
64  *
65  * (5) Evict.  When allocating a new entry, we evict the oldest (LRU) entry
66  *     if the total cache size exceeds zfs_vdev_cache_size.
67  */
68 
69 /*
70  * These tunables are for performance analysis.
71  */
72 /*
73  * All i/os smaller than zfs_vdev_cache_max will be turned into
74  * 1<<zfs_vdev_cache_bshift byte reads by the vdev_cache (aka software
75  * track buffer).  At most zfs_vdev_cache_size bytes will be kept in each
76  * vdev's vdev_cache.
77  *
78  * TODO: Note that with the current ZFS code, it turns out that the
79  * vdev cache is not helpful, and in some cases actually harmful.  It
80  * is better if we disable this.  Once some time has passed, we should
81  * actually remove this to simplify the code.  For now we just disable
82  * it by setting the zfs_vdev_cache_size to zero.  Note that Solaris 11
83  * has made these same changes.
84  */
85 int zfs_vdev_cache_max = 1<<14;			/* 16KB */
86 int zfs_vdev_cache_size = 0;
87 int zfs_vdev_cache_bshift = 16;
88 
89 #define	VCBS (1 << zfs_vdev_cache_bshift)	/* 64KB */
90 
91 kstat_t	*vdc_ksp = NULL;
92 
93 typedef struct vdc_stats {
94 	kstat_named_t vdc_stat_delegations;
95 	kstat_named_t vdc_stat_hits;
96 	kstat_named_t vdc_stat_misses;
97 } vdc_stats_t;
98 
99 static vdc_stats_t vdc_stats = {
100 	{ "delegations",	KSTAT_DATA_UINT64 },
101 	{ "hits",		KSTAT_DATA_UINT64 },
102 	{ "misses",		KSTAT_DATA_UINT64 }
103 };
104 
105 #define	VDCSTAT_BUMP(stat)	atomic_add_64(&vdc_stats.stat.value.ui64, 1);
106 
107 static int
108 vdev_cache_offset_compare(const void *a1, const void *a2)
109 {
110 	const vdev_cache_entry_t *ve1 = a1;
111 	const vdev_cache_entry_t *ve2 = a2;
112 
113 	if (ve1->ve_offset < ve2->ve_offset)
114 		return (-1);
115 	if (ve1->ve_offset > ve2->ve_offset)
116 		return (1);
117 	return (0);
118 }
119 
120 static int
121 vdev_cache_lastused_compare(const void *a1, const void *a2)
122 {
123 	const vdev_cache_entry_t *ve1 = a1;
124 	const vdev_cache_entry_t *ve2 = a2;
125 
126 	if (ve1->ve_lastused < ve2->ve_lastused)
127 		return (-1);
128 	if (ve1->ve_lastused > ve2->ve_lastused)
129 		return (1);
130 
131 	/*
132 	 * Among equally old entries, sort by offset to ensure uniqueness.
133 	 */
134 	return (vdev_cache_offset_compare(a1, a2));
135 }
136 
137 /*
138  * Evict the specified entry from the cache.
139  */
140 static void
141 vdev_cache_evict(vdev_cache_t *vc, vdev_cache_entry_t *ve)
142 {
143 	ASSERT(MUTEX_HELD(&vc->vc_lock));
144 	ASSERT(ve->ve_fill_io == NULL);
145 	ASSERT(ve->ve_data != NULL);
146 
147 	avl_remove(&vc->vc_lastused_tree, ve);
148 	avl_remove(&vc->vc_offset_tree, ve);
149 	zio_buf_free(ve->ve_data, VCBS);
150 	kmem_free(ve, sizeof (vdev_cache_entry_t));
151 }
152 
153 /*
154  * Allocate an entry in the cache.  At the point we don't have the data,
155  * we're just creating a placeholder so that multiple threads don't all
156  * go off and read the same blocks.
157  */
158 static vdev_cache_entry_t *
159 vdev_cache_allocate(zio_t *zio)
160 {
161 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
162 	uint64_t offset = P2ALIGN(zio->io_offset, VCBS);
163 	vdev_cache_entry_t *ve;
164 
165 	ASSERT(MUTEX_HELD(&vc->vc_lock));
166 
167 	if (zfs_vdev_cache_size == 0)
168 		return (NULL);
169 
170 	/*
171 	 * If adding a new entry would exceed the cache size,
172 	 * evict the oldest entry (LRU).
173 	 */
174 	if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) >
175 	    zfs_vdev_cache_size) {
176 		ve = avl_first(&vc->vc_lastused_tree);
177 		if (ve->ve_fill_io != NULL)
178 			return (NULL);
179 		ASSERT(ve->ve_hits != 0);
180 		vdev_cache_evict(vc, ve);
181 	}
182 
183 	ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_SLEEP);
184 	ve->ve_offset = offset;
185 	ve->ve_lastused = ddi_get_lbolt();
186 	ve->ve_data = zio_buf_alloc(VCBS);
187 
188 	avl_add(&vc->vc_offset_tree, ve);
189 	avl_add(&vc->vc_lastused_tree, ve);
190 
191 	return (ve);
192 }
193 
194 static void
195 vdev_cache_hit(vdev_cache_t *vc, vdev_cache_entry_t *ve, zio_t *zio)
196 {
197 	uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
198 
199 	ASSERT(MUTEX_HELD(&vc->vc_lock));
200 	ASSERT(ve->ve_fill_io == NULL);
201 
202 	if (ve->ve_lastused != ddi_get_lbolt()) {
203 		avl_remove(&vc->vc_lastused_tree, ve);
204 		ve->ve_lastused = ddi_get_lbolt();
205 		avl_add(&vc->vc_lastused_tree, ve);
206 	}
207 
208 	ve->ve_hits++;
209 	bcopy(ve->ve_data + cache_phase, zio->io_data, zio->io_size);
210 }
211 
212 /*
213  * Fill a previously allocated cache entry with data.
214  */
215 static void
216 vdev_cache_fill(zio_t *fio)
217 {
218 	vdev_t *vd = fio->io_vd;
219 	vdev_cache_t *vc = &vd->vdev_cache;
220 	vdev_cache_entry_t *ve = fio->io_private;
221 	zio_t *pio;
222 
223 	ASSERT(fio->io_size == VCBS);
224 
225 	/*
226 	 * Add data to the cache.
227 	 */
228 	mutex_enter(&vc->vc_lock);
229 
230 	ASSERT(ve->ve_fill_io == fio);
231 	ASSERT(ve->ve_offset == fio->io_offset);
232 	ASSERT(ve->ve_data == fio->io_data);
233 
234 	ve->ve_fill_io = NULL;
235 
236 	/*
237 	 * Even if this cache line was invalidated by a missed write update,
238 	 * any reads that were queued up before the missed update are still
239 	 * valid, so we can satisfy them from this line before we evict it.
240 	 */
241 	while ((pio = zio_walk_parents(fio)) != NULL)
242 		vdev_cache_hit(vc, ve, pio);
243 
244 	if (fio->io_error || ve->ve_missed_update)
245 		vdev_cache_evict(vc, ve);
246 
247 	mutex_exit(&vc->vc_lock);
248 }
249 
250 /*
251  * Read data from the cache.  Returns 0 on cache hit, errno on a miss.
252  */
253 int
254 vdev_cache_read(zio_t *zio)
255 {
256 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
257 	vdev_cache_entry_t *ve, ve_search;
258 	uint64_t cache_offset = P2ALIGN(zio->io_offset, VCBS);
259 	uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
260 	zio_t *fio;
261 
262 	ASSERT(zio->io_type == ZIO_TYPE_READ);
263 
264 	if (zio->io_flags & ZIO_FLAG_DONT_CACHE)
265 		return (SET_ERROR(EINVAL));
266 
267 	if (zio->io_size > zfs_vdev_cache_max)
268 		return (SET_ERROR(EOVERFLOW));
269 
270 	/*
271 	 * If the I/O straddles two or more cache blocks, don't cache it.
272 	 */
273 	if (P2BOUNDARY(zio->io_offset, zio->io_size, VCBS))
274 		return (SET_ERROR(EXDEV));
275 
276 	ASSERT(cache_phase + zio->io_size <= VCBS);
277 
278 	mutex_enter(&vc->vc_lock);
279 
280 	ve_search.ve_offset = cache_offset;
281 	ve = avl_find(&vc->vc_offset_tree, &ve_search, NULL);
282 
283 	if (ve != NULL) {
284 		if (ve->ve_missed_update) {
285 			mutex_exit(&vc->vc_lock);
286 			return (SET_ERROR(ESTALE));
287 		}
288 
289 		if ((fio = ve->ve_fill_io) != NULL) {
290 			zio_vdev_io_bypass(zio);
291 			zio_add_child(zio, fio);
292 			mutex_exit(&vc->vc_lock);
293 			VDCSTAT_BUMP(vdc_stat_delegations);
294 			return (0);
295 		}
296 
297 		vdev_cache_hit(vc, ve, zio);
298 		zio_vdev_io_bypass(zio);
299 
300 		mutex_exit(&vc->vc_lock);
301 		VDCSTAT_BUMP(vdc_stat_hits);
302 		return (0);
303 	}
304 
305 	ve = vdev_cache_allocate(zio);
306 
307 	if (ve == NULL) {
308 		mutex_exit(&vc->vc_lock);
309 		return (SET_ERROR(ENOMEM));
310 	}
311 
312 	fio = zio_vdev_delegated_io(zio->io_vd, cache_offset,
313 	    ve->ve_data, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_NOW,
314 	    ZIO_FLAG_DONT_CACHE, vdev_cache_fill, ve);
315 
316 	ve->ve_fill_io = fio;
317 	zio_vdev_io_bypass(zio);
318 	zio_add_child(zio, fio);
319 
320 	mutex_exit(&vc->vc_lock);
321 	zio_nowait(fio);
322 	VDCSTAT_BUMP(vdc_stat_misses);
323 
324 	return (0);
325 }
326 
327 /*
328  * Update cache contents upon write completion.
329  */
330 void
331 vdev_cache_write(zio_t *zio)
332 {
333 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
334 	vdev_cache_entry_t *ve, ve_search;
335 	uint64_t io_start = zio->io_offset;
336 	uint64_t io_end = io_start + zio->io_size;
337 	uint64_t min_offset = P2ALIGN(io_start, VCBS);
338 	uint64_t max_offset = P2ROUNDUP(io_end, VCBS);
339 	avl_index_t where;
340 
341 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
342 
343 	mutex_enter(&vc->vc_lock);
344 
345 	ve_search.ve_offset = min_offset;
346 	ve = avl_find(&vc->vc_offset_tree, &ve_search, &where);
347 
348 	if (ve == NULL)
349 		ve = avl_nearest(&vc->vc_offset_tree, where, AVL_AFTER);
350 
351 	while (ve != NULL && ve->ve_offset < max_offset) {
352 		uint64_t start = MAX(ve->ve_offset, io_start);
353 		uint64_t end = MIN(ve->ve_offset + VCBS, io_end);
354 
355 		if (ve->ve_fill_io != NULL) {
356 			ve->ve_missed_update = 1;
357 		} else {
358 			bcopy((char *)zio->io_data + start - io_start,
359 			    ve->ve_data + start - ve->ve_offset, end - start);
360 		}
361 		ve = AVL_NEXT(&vc->vc_offset_tree, ve);
362 	}
363 	mutex_exit(&vc->vc_lock);
364 }
365 
366 void
367 vdev_cache_purge(vdev_t *vd)
368 {
369 	vdev_cache_t *vc = &vd->vdev_cache;
370 	vdev_cache_entry_t *ve;
371 
372 	mutex_enter(&vc->vc_lock);
373 	while ((ve = avl_first(&vc->vc_offset_tree)) != NULL)
374 		vdev_cache_evict(vc, ve);
375 	mutex_exit(&vc->vc_lock);
376 }
377 
378 void
379 vdev_cache_init(vdev_t *vd)
380 {
381 	vdev_cache_t *vc = &vd->vdev_cache;
382 
383 	mutex_init(&vc->vc_lock, NULL, MUTEX_DEFAULT, NULL);
384 
385 	avl_create(&vc->vc_offset_tree, vdev_cache_offset_compare,
386 	    sizeof (vdev_cache_entry_t),
387 	    offsetof(struct vdev_cache_entry, ve_offset_node));
388 
389 	avl_create(&vc->vc_lastused_tree, vdev_cache_lastused_compare,
390 	    sizeof (vdev_cache_entry_t),
391 	    offsetof(struct vdev_cache_entry, ve_lastused_node));
392 }
393 
394 void
395 vdev_cache_fini(vdev_t *vd)
396 {
397 	vdev_cache_t *vc = &vd->vdev_cache;
398 
399 	vdev_cache_purge(vd);
400 
401 	avl_destroy(&vc->vc_offset_tree);
402 	avl_destroy(&vc->vc_lastused_tree);
403 
404 	mutex_destroy(&vc->vc_lock);
405 }
406 
407 void
408 vdev_cache_stat_init(void)
409 {
410 	vdc_ksp = kstat_create("zfs", 0, "vdev_cache_stats", "misc",
411 	    KSTAT_TYPE_NAMED, sizeof (vdc_stats) / sizeof (kstat_named_t),
412 	    KSTAT_FLAG_VIRTUAL);
413 	if (vdc_ksp != NULL) {
414 		vdc_ksp->ks_data = &vdc_stats;
415 		kstat_install(vdc_ksp);
416 	}
417 }
418 
419 void
420 vdev_cache_stat_fini(void)
421 {
422 	if (vdc_ksp != NULL) {
423 		kstat_delete(vdc_ksp);
424 		vdc_ksp = NULL;
425 	}
426 }
427