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.
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.
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]
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
26 * Copyright (c) 2013 by Delphix. All rights reserved.
29 #include <sys/zfs_context.h>
31 #include <sys/vdev_impl.h>
33 #include <sys/kstat.h>
36 * Virtual device read-ahead caching.
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.
52 * There are five cache operations: allocate, fill, read, write, evict.
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.
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.
61 * (3) Read. Read data from the cache.
63 * (4) Write. Update cache contents after write completion.
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.
70 * These tunables are for performance analysis.
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
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.
85 int zfs_vdev_cache_max = 1<<14; /* 16KB */
86 int zfs_vdev_cache_size = 0;
87 int zfs_vdev_cache_bshift = 16;
89 #define VCBS (1 << zfs_vdev_cache_bshift) /* 64KB */
91 SYSCTL_DECL(_vfs_zfs_vdev);
92 SYSCTL_NODE(_vfs_zfs_vdev, OID_AUTO, cache, CTLFLAG_RW, 0, "ZFS VDEV Cache");
93 TUNABLE_INT("vfs.zfs.vdev.cache.max", &zfs_vdev_cache_max);
94 SYSCTL_INT(_vfs_zfs_vdev_cache, OID_AUTO, max, CTLFLAG_RDTUN,
95 &zfs_vdev_cache_max, 0, "Maximum I/O request size that increase read size");
96 TUNABLE_INT("vfs.zfs.vdev.cache.size", &zfs_vdev_cache_size);
97 SYSCTL_INT(_vfs_zfs_vdev_cache, OID_AUTO, size, CTLFLAG_RDTUN,
98 &zfs_vdev_cache_size, 0, "Size of VDEV cache");
99 TUNABLE_INT("vfs.zfs.vdev.cache.bshift", &zfs_vdev_cache_bshift);
100 SYSCTL_INT(_vfs_zfs_vdev_cache, OID_AUTO, bshift, CTLFLAG_RDTUN,
101 &zfs_vdev_cache_bshift, 0, "Turn too small requests into 1 << this value");
103 kstat_t *vdc_ksp = NULL;
105 typedef struct vdc_stats {
106 kstat_named_t vdc_stat_delegations;
107 kstat_named_t vdc_stat_hits;
108 kstat_named_t vdc_stat_misses;
111 static vdc_stats_t vdc_stats = {
112 { "delegations", KSTAT_DATA_UINT64 },
113 { "hits", KSTAT_DATA_UINT64 },
114 { "misses", KSTAT_DATA_UINT64 }
117 #define VDCSTAT_BUMP(stat) atomic_add_64(&vdc_stats.stat.value.ui64, 1);
120 vdev_cache_offset_compare(const void *a1, const void *a2)
122 const vdev_cache_entry_t *ve1 = a1;
123 const vdev_cache_entry_t *ve2 = a2;
125 if (ve1->ve_offset < ve2->ve_offset)
127 if (ve1->ve_offset > ve2->ve_offset)
133 vdev_cache_lastused_compare(const void *a1, const void *a2)
135 const vdev_cache_entry_t *ve1 = a1;
136 const vdev_cache_entry_t *ve2 = a2;
138 if (ve1->ve_lastused < ve2->ve_lastused)
140 if (ve1->ve_lastused > ve2->ve_lastused)
144 * Among equally old entries, sort by offset to ensure uniqueness.
146 return (vdev_cache_offset_compare(a1, a2));
150 * Evict the specified entry from the cache.
153 vdev_cache_evict(vdev_cache_t *vc, vdev_cache_entry_t *ve)
155 ASSERT(MUTEX_HELD(&vc->vc_lock));
156 ASSERT(ve->ve_fill_io == NULL);
157 ASSERT(ve->ve_data != NULL);
159 avl_remove(&vc->vc_lastused_tree, ve);
160 avl_remove(&vc->vc_offset_tree, ve);
161 zio_buf_free(ve->ve_data, VCBS);
162 kmem_free(ve, sizeof (vdev_cache_entry_t));
166 * Allocate an entry in the cache. At the point we don't have the data,
167 * we're just creating a placeholder so that multiple threads don't all
168 * go off and read the same blocks.
170 static vdev_cache_entry_t *
171 vdev_cache_allocate(zio_t *zio)
173 vdev_cache_t *vc = &zio->io_vd->vdev_cache;
174 uint64_t offset = P2ALIGN(zio->io_offset, VCBS);
175 vdev_cache_entry_t *ve;
177 ASSERT(MUTEX_HELD(&vc->vc_lock));
179 if (zfs_vdev_cache_size == 0)
183 * If adding a new entry would exceed the cache size,
184 * evict the oldest entry (LRU).
186 if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) >
187 zfs_vdev_cache_size) {
188 ve = avl_first(&vc->vc_lastused_tree);
189 if (ve->ve_fill_io != NULL)
191 ASSERT(ve->ve_hits != 0);
192 vdev_cache_evict(vc, ve);
195 ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_SLEEP);
196 ve->ve_offset = offset;
197 ve->ve_lastused = ddi_get_lbolt();
198 ve->ve_data = zio_buf_alloc(VCBS);
200 avl_add(&vc->vc_offset_tree, ve);
201 avl_add(&vc->vc_lastused_tree, ve);
207 vdev_cache_hit(vdev_cache_t *vc, vdev_cache_entry_t *ve, zio_t *zio)
209 uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
211 ASSERT(MUTEX_HELD(&vc->vc_lock));
212 ASSERT(ve->ve_fill_io == NULL);
214 if (ve->ve_lastused != ddi_get_lbolt()) {
215 avl_remove(&vc->vc_lastused_tree, ve);
216 ve->ve_lastused = ddi_get_lbolt();
217 avl_add(&vc->vc_lastused_tree, ve);
221 bcopy(ve->ve_data + cache_phase, zio->io_data, zio->io_size);
225 * Fill a previously allocated cache entry with data.
228 vdev_cache_fill(zio_t *fio)
230 vdev_t *vd = fio->io_vd;
231 vdev_cache_t *vc = &vd->vdev_cache;
232 vdev_cache_entry_t *ve = fio->io_private;
235 ASSERT(fio->io_size == VCBS);
238 * Add data to the cache.
240 mutex_enter(&vc->vc_lock);
242 ASSERT(ve->ve_fill_io == fio);
243 ASSERT(ve->ve_offset == fio->io_offset);
244 ASSERT(ve->ve_data == fio->io_data);
246 ve->ve_fill_io = NULL;
249 * Even if this cache line was invalidated by a missed write update,
250 * any reads that were queued up before the missed update are still
251 * valid, so we can satisfy them from this line before we evict it.
253 while ((pio = zio_walk_parents(fio)) != NULL)
254 vdev_cache_hit(vc, ve, pio);
256 if (fio->io_error || ve->ve_missed_update)
257 vdev_cache_evict(vc, ve);
259 mutex_exit(&vc->vc_lock);
263 * Read data from the cache. Returns 0 on cache hit, errno on a miss.
266 vdev_cache_read(zio_t *zio)
268 vdev_cache_t *vc = &zio->io_vd->vdev_cache;
269 vdev_cache_entry_t *ve, ve_search;
270 uint64_t cache_offset = P2ALIGN(zio->io_offset, VCBS);
271 uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
274 ASSERT(zio->io_type == ZIO_TYPE_READ);
276 if (zio->io_flags & ZIO_FLAG_DONT_CACHE)
277 return (SET_ERROR(EINVAL));
279 if (zio->io_size > zfs_vdev_cache_max)
280 return (SET_ERROR(EOVERFLOW));
283 * If the I/O straddles two or more cache blocks, don't cache it.
285 if (P2BOUNDARY(zio->io_offset, zio->io_size, VCBS))
286 return (SET_ERROR(EXDEV));
288 ASSERT(cache_phase + zio->io_size <= VCBS);
290 mutex_enter(&vc->vc_lock);
292 ve_search.ve_offset = cache_offset;
293 ve = avl_find(&vc->vc_offset_tree, &ve_search, NULL);
296 if (ve->ve_missed_update) {
297 mutex_exit(&vc->vc_lock);
298 return (SET_ERROR(ESTALE));
301 if ((fio = ve->ve_fill_io) != NULL) {
302 zio_vdev_io_bypass(zio);
303 zio_add_child(zio, fio);
304 mutex_exit(&vc->vc_lock);
305 VDCSTAT_BUMP(vdc_stat_delegations);
309 vdev_cache_hit(vc, ve, zio);
310 zio_vdev_io_bypass(zio);
312 mutex_exit(&vc->vc_lock);
313 VDCSTAT_BUMP(vdc_stat_hits);
317 ve = vdev_cache_allocate(zio);
320 mutex_exit(&vc->vc_lock);
321 return (SET_ERROR(ENOMEM));
324 fio = zio_vdev_delegated_io(zio->io_vd, cache_offset,
325 ve->ve_data, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_CACHE_FILL,
326 ZIO_FLAG_DONT_CACHE, vdev_cache_fill, ve);
328 ve->ve_fill_io = fio;
329 zio_vdev_io_bypass(zio);
330 zio_add_child(zio, fio);
332 mutex_exit(&vc->vc_lock);
334 VDCSTAT_BUMP(vdc_stat_misses);
340 * Update cache contents upon write completion.
343 vdev_cache_write(zio_t *zio)
345 vdev_cache_t *vc = &zio->io_vd->vdev_cache;
346 vdev_cache_entry_t *ve, ve_search;
347 uint64_t io_start = zio->io_offset;
348 uint64_t io_end = io_start + zio->io_size;
349 uint64_t min_offset = P2ALIGN(io_start, VCBS);
350 uint64_t max_offset = P2ROUNDUP(io_end, VCBS);
353 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
355 mutex_enter(&vc->vc_lock);
357 ve_search.ve_offset = min_offset;
358 ve = avl_find(&vc->vc_offset_tree, &ve_search, &where);
361 ve = avl_nearest(&vc->vc_offset_tree, where, AVL_AFTER);
363 while (ve != NULL && ve->ve_offset < max_offset) {
364 uint64_t start = MAX(ve->ve_offset, io_start);
365 uint64_t end = MIN(ve->ve_offset + VCBS, io_end);
367 if (ve->ve_fill_io != NULL) {
368 ve->ve_missed_update = 1;
370 bcopy((char *)zio->io_data + start - io_start,
371 ve->ve_data + start - ve->ve_offset, end - start);
373 ve = AVL_NEXT(&vc->vc_offset_tree, ve);
375 mutex_exit(&vc->vc_lock);
379 vdev_cache_purge(vdev_t *vd)
381 vdev_cache_t *vc = &vd->vdev_cache;
382 vdev_cache_entry_t *ve;
384 mutex_enter(&vc->vc_lock);
385 while ((ve = avl_first(&vc->vc_offset_tree)) != NULL)
386 vdev_cache_evict(vc, ve);
387 mutex_exit(&vc->vc_lock);
391 vdev_cache_init(vdev_t *vd)
393 vdev_cache_t *vc = &vd->vdev_cache;
395 mutex_init(&vc->vc_lock, NULL, MUTEX_DEFAULT, NULL);
397 avl_create(&vc->vc_offset_tree, vdev_cache_offset_compare,
398 sizeof (vdev_cache_entry_t),
399 offsetof(struct vdev_cache_entry, ve_offset_node));
401 avl_create(&vc->vc_lastused_tree, vdev_cache_lastused_compare,
402 sizeof (vdev_cache_entry_t),
403 offsetof(struct vdev_cache_entry, ve_lastused_node));
407 vdev_cache_fini(vdev_t *vd)
409 vdev_cache_t *vc = &vd->vdev_cache;
411 vdev_cache_purge(vd);
413 avl_destroy(&vc->vc_offset_tree);
414 avl_destroy(&vc->vc_lastused_tree);
416 mutex_destroy(&vc->vc_lock);
420 vdev_cache_stat_init(void)
422 vdc_ksp = kstat_create("zfs", 0, "vdev_cache_stats", "misc",
423 KSTAT_TYPE_NAMED, sizeof (vdc_stats) / sizeof (kstat_named_t),
425 if (vdc_ksp != NULL) {
426 vdc_ksp->ks_data = &vdc_stats;
427 kstat_install(vdc_ksp);
432 vdev_cache_stat_fini(void)
434 if (vdc_ksp != NULL) {
435 kstat_delete(vdc_ksp);