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 2008 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
26 #include <sys/zfs_context.h>
28 #include <sys/vdev_impl.h>
30 #include <sys/kstat.h>
33 * Virtual device read-ahead caching.
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.
49 * There are five cache operations: allocate, fill, read, write, evict.
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.
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.
58 * (3) Read. Read data from the cache.
60 * (4) Write. Update cache contents after write completion.
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.
67 * These tunables are for performance analysis.
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
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;
79 #define VCBS (1 << zfs_vdev_cache_bshift) /* 64KB */
81 SYSCTL_DECL(_vfs_zfs_vdev);
82 SYSCTL_NODE(_vfs_zfs_vdev, OID_AUTO, cache, CTLFLAG_RW, 0, "ZFS VDEV Cache");
83 TUNABLE_INT("vfs.zfs.vdev.cache.max", &zfs_vdev_cache_max);
84 SYSCTL_INT(_vfs_zfs_vdev_cache, OID_AUTO, max, CTLFLAG_RDTUN,
85 &zfs_vdev_cache_max, 0, "Maximum I/O request size that increase read size");
86 TUNABLE_INT("vfs.zfs.vdev.cache.size", &zfs_vdev_cache_size);
87 SYSCTL_INT(_vfs_zfs_vdev_cache, OID_AUTO, size, CTLFLAG_RDTUN,
88 &zfs_vdev_cache_size, 0, "Size of VDEV cache");
89 TUNABLE_INT("vfs.zfs.vdev.cache.bshift", &zfs_vdev_cache_bshift);
90 SYSCTL_INT(_vfs_zfs_vdev_cache, OID_AUTO, bshift, CTLFLAG_RDTUN,
91 &zfs_vdev_cache_bshift, 0, "Turn too small requests into 1 << this value");
93 kstat_t *vdc_ksp = NULL;
95 typedef struct vdc_stats {
96 kstat_named_t vdc_stat_delegations;
97 kstat_named_t vdc_stat_hits;
98 kstat_named_t vdc_stat_misses;
101 static vdc_stats_t vdc_stats = {
102 { "delegations", KSTAT_DATA_UINT64 },
103 { "hits", KSTAT_DATA_UINT64 },
104 { "misses", KSTAT_DATA_UINT64 }
107 #define VDCSTAT_BUMP(stat) atomic_add_64(&vdc_stats.stat.value.ui64, 1);
110 vdev_cache_offset_compare(const void *a1, const void *a2)
112 const vdev_cache_entry_t *ve1 = a1;
113 const vdev_cache_entry_t *ve2 = a2;
115 if (ve1->ve_offset < ve2->ve_offset)
117 if (ve1->ve_offset > ve2->ve_offset)
123 vdev_cache_lastused_compare(const void *a1, const void *a2)
125 const vdev_cache_entry_t *ve1 = a1;
126 const vdev_cache_entry_t *ve2 = a2;
128 if (ve1->ve_lastused < ve2->ve_lastused)
130 if (ve1->ve_lastused > ve2->ve_lastused)
134 * Among equally old entries, sort by offset to ensure uniqueness.
136 return (vdev_cache_offset_compare(a1, a2));
140 * Evict the specified entry from the cache.
143 vdev_cache_evict(vdev_cache_t *vc, vdev_cache_entry_t *ve)
145 ASSERT(MUTEX_HELD(&vc->vc_lock));
146 ASSERT(ve->ve_fill_io == NULL);
147 ASSERT(ve->ve_data != NULL);
149 avl_remove(&vc->vc_lastused_tree, ve);
150 avl_remove(&vc->vc_offset_tree, ve);
151 zio_buf_free(ve->ve_data, VCBS);
152 kmem_free(ve, sizeof (vdev_cache_entry_t));
156 * Allocate an entry in the cache. At the point we don't have the data,
157 * we're just creating a placeholder so that multiple threads don't all
158 * go off and read the same blocks.
160 static vdev_cache_entry_t *
161 vdev_cache_allocate(zio_t *zio)
163 vdev_cache_t *vc = &zio->io_vd->vdev_cache;
164 uint64_t offset = P2ALIGN(zio->io_offset, VCBS);
165 vdev_cache_entry_t *ve;
167 ASSERT(MUTEX_HELD(&vc->vc_lock));
169 if (zfs_vdev_cache_size == 0)
173 * If adding a new entry would exceed the cache size,
174 * evict the oldest entry (LRU).
176 if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) >
177 zfs_vdev_cache_size) {
178 ve = avl_first(&vc->vc_lastused_tree);
179 if (ve->ve_fill_io != NULL)
181 ASSERT(ve->ve_hits != 0);
182 vdev_cache_evict(vc, ve);
185 ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_SLEEP);
186 ve->ve_offset = offset;
187 ve->ve_lastused = LBOLT;
188 ve->ve_data = zio_buf_alloc(VCBS);
190 avl_add(&vc->vc_offset_tree, ve);
191 avl_add(&vc->vc_lastused_tree, ve);
197 vdev_cache_hit(vdev_cache_t *vc, vdev_cache_entry_t *ve, zio_t *zio)
199 uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
201 ASSERT(MUTEX_HELD(&vc->vc_lock));
202 ASSERT(ve->ve_fill_io == NULL);
204 if (ve->ve_lastused != LBOLT) {
205 avl_remove(&vc->vc_lastused_tree, ve);
206 ve->ve_lastused = LBOLT;
207 avl_add(&vc->vc_lastused_tree, ve);
211 bcopy(ve->ve_data + cache_phase, zio->io_data, zio->io_size);
215 * Fill a previously allocated cache entry with data.
218 vdev_cache_fill(zio_t *zio)
220 vdev_t *vd = zio->io_vd;
221 vdev_cache_t *vc = &vd->vdev_cache;
222 vdev_cache_entry_t *ve = zio->io_private;
225 ASSERT(zio->io_size == VCBS);
228 * Add data to the cache.
230 mutex_enter(&vc->vc_lock);
232 ASSERT(ve->ve_fill_io == zio);
233 ASSERT(ve->ve_offset == zio->io_offset);
234 ASSERT(ve->ve_data == zio->io_data);
236 ve->ve_fill_io = NULL;
239 * Even if this cache line was invalidated by a missed write update,
240 * any reads that were queued up before the missed update are still
241 * valid, so we can satisfy them from this line before we evict it.
243 for (dio = zio->io_delegate_list; dio; dio = dio->io_delegate_next)
244 vdev_cache_hit(vc, ve, dio);
246 if (zio->io_error || ve->ve_missed_update)
247 vdev_cache_evict(vc, ve);
249 mutex_exit(&vc->vc_lock);
251 while ((dio = zio->io_delegate_list) != NULL) {
252 zio->io_delegate_list = dio->io_delegate_next;
253 dio->io_delegate_next = NULL;
254 dio->io_error = zio->io_error;
260 * Read data from the cache. Returns 0 on cache hit, errno on a miss.
263 vdev_cache_read(zio_t *zio)
265 vdev_cache_t *vc = &zio->io_vd->vdev_cache;
266 vdev_cache_entry_t *ve, ve_search;
267 uint64_t cache_offset = P2ALIGN(zio->io_offset, VCBS);
268 uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
271 ASSERT(zio->io_type == ZIO_TYPE_READ);
273 if (zio->io_flags & ZIO_FLAG_DONT_CACHE)
276 if (zio->io_size > zfs_vdev_cache_max)
280 * If the I/O straddles two or more cache blocks, don't cache it.
282 if (P2CROSS(zio->io_offset, zio->io_offset + zio->io_size - 1, VCBS))
285 ASSERT(cache_phase + zio->io_size <= VCBS);
287 mutex_enter(&vc->vc_lock);
289 ve_search.ve_offset = cache_offset;
290 ve = avl_find(&vc->vc_offset_tree, &ve_search, NULL);
293 if (ve->ve_missed_update) {
294 mutex_exit(&vc->vc_lock);
298 if ((fio = ve->ve_fill_io) != NULL) {
299 zio->io_delegate_next = fio->io_delegate_list;
300 fio->io_delegate_list = zio;
301 zio_vdev_io_bypass(zio);
302 mutex_exit(&vc->vc_lock);
303 VDCSTAT_BUMP(vdc_stat_delegations);
307 vdev_cache_hit(vc, ve, zio);
308 zio_vdev_io_bypass(zio);
310 mutex_exit(&vc->vc_lock);
312 VDCSTAT_BUMP(vdc_stat_hits);
316 ve = vdev_cache_allocate(zio);
319 mutex_exit(&vc->vc_lock);
323 fio = zio_vdev_delegated_io(zio->io_vd, cache_offset,
324 ve->ve_data, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_CACHE_FILL,
325 ZIO_FLAG_DONT_CACHE, vdev_cache_fill, ve);
327 ve->ve_fill_io = fio;
328 fio->io_delegate_list = zio;
329 zio_vdev_io_bypass(zio);
331 mutex_exit(&vc->vc_lock);
333 VDCSTAT_BUMP(vdc_stat_misses);
339 * Update cache contents upon write completion.
342 vdev_cache_write(zio_t *zio)
344 vdev_cache_t *vc = &zio->io_vd->vdev_cache;
345 vdev_cache_entry_t *ve, ve_search;
346 uint64_t io_start = zio->io_offset;
347 uint64_t io_end = io_start + zio->io_size;
348 uint64_t min_offset = P2ALIGN(io_start, VCBS);
349 uint64_t max_offset = P2ROUNDUP(io_end, VCBS);
352 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
354 mutex_enter(&vc->vc_lock);
356 ve_search.ve_offset = min_offset;
357 ve = avl_find(&vc->vc_offset_tree, &ve_search, &where);
360 ve = avl_nearest(&vc->vc_offset_tree, where, AVL_AFTER);
362 while (ve != NULL && ve->ve_offset < max_offset) {
363 uint64_t start = MAX(ve->ve_offset, io_start);
364 uint64_t end = MIN(ve->ve_offset + VCBS, io_end);
366 if (ve->ve_fill_io != NULL) {
367 ve->ve_missed_update = 1;
369 bcopy((char *)zio->io_data + start - io_start,
370 ve->ve_data + start - ve->ve_offset, end - start);
372 ve = AVL_NEXT(&vc->vc_offset_tree, ve);
374 mutex_exit(&vc->vc_lock);
378 vdev_cache_purge(vdev_t *vd)
380 vdev_cache_t *vc = &vd->vdev_cache;
381 vdev_cache_entry_t *ve;
383 mutex_enter(&vc->vc_lock);
384 while ((ve = avl_first(&vc->vc_offset_tree)) != NULL)
385 vdev_cache_evict(vc, ve);
386 mutex_exit(&vc->vc_lock);
390 vdev_cache_init(vdev_t *vd)
392 vdev_cache_t *vc = &vd->vdev_cache;
394 mutex_init(&vc->vc_lock, NULL, MUTEX_DEFAULT, NULL);
396 avl_create(&vc->vc_offset_tree, vdev_cache_offset_compare,
397 sizeof (vdev_cache_entry_t),
398 offsetof(struct vdev_cache_entry, ve_offset_node));
400 avl_create(&vc->vc_lastused_tree, vdev_cache_lastused_compare,
401 sizeof (vdev_cache_entry_t),
402 offsetof(struct vdev_cache_entry, ve_lastused_node));
406 vdev_cache_fini(vdev_t *vd)
408 vdev_cache_t *vc = &vd->vdev_cache;
410 vdev_cache_purge(vd);
412 avl_destroy(&vc->vc_offset_tree);
413 avl_destroy(&vc->vc_lastused_tree);
415 mutex_destroy(&vc->vc_lock);
419 vdev_cache_stat_init(void)
421 vdc_ksp = kstat_create("zfs", 0, "vdev_cache_stats", "misc",
422 KSTAT_TYPE_NAMED, sizeof (vdc_stats) / sizeof (kstat_named_t),
424 if (vdc_ksp != NULL) {
425 vdc_ksp->ks_data = &vdc_stats;
426 kstat_install(vdc_ksp);
431 vdev_cache_stat_fini(void)
433 if (vdc_ksp != NULL) {
434 kstat_delete(vdc_ksp);