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 #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 * TODO: Note that with the current ZFS code, it turns out that the
76 * vdev cache is not helpful, and in some cases actually harmful. It
77 * is better if we disable this. Once some time has passed, we should
78 * actually remove this to simplify the code. For now we just disable
79 * it by setting the zfs_vdev_cache_size to zero. Note that Solaris 11
80 * has made these same changes.
82 int zfs_vdev_cache_max = 1<<14; /* 16KB */
83 int zfs_vdev_cache_size = 0;
84 int zfs_vdev_cache_bshift = 16;
86 #define VCBS (1 << zfs_vdev_cache_bshift) /* 64KB */
88 kstat_t *vdc_ksp = NULL;
90 typedef struct vdc_stats {
91 kstat_named_t vdc_stat_delegations;
92 kstat_named_t vdc_stat_hits;
93 kstat_named_t vdc_stat_misses;
96 static vdc_stats_t vdc_stats = {
97 { "delegations", KSTAT_DATA_UINT64 },
98 { "hits", KSTAT_DATA_UINT64 },
99 { "misses", KSTAT_DATA_UINT64 }
102 #define VDCSTAT_BUMP(stat) atomic_add_64(&vdc_stats.stat.value.ui64, 1);
105 vdev_cache_offset_compare(const void *a1, const void *a2)
107 const vdev_cache_entry_t *ve1 = a1;
108 const vdev_cache_entry_t *ve2 = a2;
110 if (ve1->ve_offset < ve2->ve_offset)
112 if (ve1->ve_offset > ve2->ve_offset)
118 vdev_cache_lastused_compare(const void *a1, const void *a2)
120 const vdev_cache_entry_t *ve1 = a1;
121 const vdev_cache_entry_t *ve2 = a2;
123 if (ve1->ve_lastused < ve2->ve_lastused)
125 if (ve1->ve_lastused > ve2->ve_lastused)
129 * Among equally old entries, sort by offset to ensure uniqueness.
131 return (vdev_cache_offset_compare(a1, a2));
135 * Evict the specified entry from the cache.
138 vdev_cache_evict(vdev_cache_t *vc, vdev_cache_entry_t *ve)
140 ASSERT(MUTEX_HELD(&vc->vc_lock));
141 ASSERT(ve->ve_fill_io == NULL);
142 ASSERT(ve->ve_data != NULL);
144 avl_remove(&vc->vc_lastused_tree, ve);
145 avl_remove(&vc->vc_offset_tree, ve);
146 zio_buf_free(ve->ve_data, VCBS);
147 kmem_free(ve, sizeof (vdev_cache_entry_t));
151 * Allocate an entry in the cache. At the point we don't have the data,
152 * we're just creating a placeholder so that multiple threads don't all
153 * go off and read the same blocks.
155 static vdev_cache_entry_t *
156 vdev_cache_allocate(zio_t *zio)
158 vdev_cache_t *vc = &zio->io_vd->vdev_cache;
159 uint64_t offset = P2ALIGN(zio->io_offset, VCBS);
160 vdev_cache_entry_t *ve;
162 ASSERT(MUTEX_HELD(&vc->vc_lock));
164 if (zfs_vdev_cache_size == 0)
168 * If adding a new entry would exceed the cache size,
169 * evict the oldest entry (LRU).
171 if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) >
172 zfs_vdev_cache_size) {
173 ve = avl_first(&vc->vc_lastused_tree);
174 if (ve->ve_fill_io != NULL)
176 ASSERT(ve->ve_hits != 0);
177 vdev_cache_evict(vc, ve);
180 ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_PUSHPAGE);
181 ve->ve_offset = offset;
182 ve->ve_lastused = ddi_get_lbolt();
183 ve->ve_data = zio_buf_alloc(VCBS);
185 avl_add(&vc->vc_offset_tree, ve);
186 avl_add(&vc->vc_lastused_tree, ve);
192 vdev_cache_hit(vdev_cache_t *vc, vdev_cache_entry_t *ve, zio_t *zio)
194 uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
196 ASSERT(MUTEX_HELD(&vc->vc_lock));
197 ASSERT(ve->ve_fill_io == NULL);
199 if (ve->ve_lastused != ddi_get_lbolt()) {
200 avl_remove(&vc->vc_lastused_tree, ve);
201 ve->ve_lastused = ddi_get_lbolt();
202 avl_add(&vc->vc_lastused_tree, ve);
206 bcopy(ve->ve_data + cache_phase, zio->io_data, zio->io_size);
210 * Fill a previously allocated cache entry with data.
213 vdev_cache_fill(zio_t *fio)
215 vdev_t *vd = fio->io_vd;
216 vdev_cache_t *vc = &vd->vdev_cache;
217 vdev_cache_entry_t *ve = fio->io_private;
220 ASSERT(fio->io_size == VCBS);
223 * Add data to the cache.
225 mutex_enter(&vc->vc_lock);
227 ASSERT(ve->ve_fill_io == fio);
228 ASSERT(ve->ve_offset == fio->io_offset);
229 ASSERT(ve->ve_data == fio->io_data);
231 ve->ve_fill_io = NULL;
234 * Even if this cache line was invalidated by a missed write update,
235 * any reads that were queued up before the missed update are still
236 * valid, so we can satisfy them from this line before we evict it.
238 while ((pio = zio_walk_parents(fio)) != NULL)
239 vdev_cache_hit(vc, ve, pio);
241 if (fio->io_error || ve->ve_missed_update)
242 vdev_cache_evict(vc, ve);
244 mutex_exit(&vc->vc_lock);
248 * Read data from the cache. Returns 0 on cache hit, errno on a miss.
251 vdev_cache_read(zio_t *zio)
253 vdev_cache_t *vc = &zio->io_vd->vdev_cache;
254 vdev_cache_entry_t *ve, *ve_search;
255 uint64_t cache_offset = P2ALIGN(zio->io_offset, VCBS);
256 ASSERTV(uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);)
259 ASSERT(zio->io_type == ZIO_TYPE_READ);
261 if (zio->io_flags & ZIO_FLAG_DONT_CACHE)
264 if (zio->io_size > zfs_vdev_cache_max)
268 * If the I/O straddles two or more cache blocks, don't cache it.
270 if (P2BOUNDARY(zio->io_offset, zio->io_size, VCBS))
273 ASSERT(cache_phase + zio->io_size <= VCBS);
275 mutex_enter(&vc->vc_lock);
277 ve_search = kmem_alloc(sizeof(vdev_cache_entry_t), KM_PUSHPAGE);
278 ve_search->ve_offset = cache_offset;
279 ve = avl_find(&vc->vc_offset_tree, ve_search, NULL);
280 kmem_free(ve_search, sizeof(vdev_cache_entry_t));
283 if (ve->ve_missed_update) {
284 mutex_exit(&vc->vc_lock);
288 if ((fio = ve->ve_fill_io) != NULL) {
289 zio_vdev_io_bypass(zio);
290 zio_add_child(zio, fio);
291 mutex_exit(&vc->vc_lock);
292 VDCSTAT_BUMP(vdc_stat_delegations);
296 vdev_cache_hit(vc, ve, zio);
297 zio_vdev_io_bypass(zio);
299 mutex_exit(&vc->vc_lock);
300 VDCSTAT_BUMP(vdc_stat_hits);
304 ve = vdev_cache_allocate(zio);
307 mutex_exit(&vc->vc_lock);
311 fio = zio_vdev_delegated_io(zio->io_vd, cache_offset,
312 ve->ve_data, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_CACHE_FILL,
313 ZIO_FLAG_DONT_CACHE, vdev_cache_fill, ve);
315 ve->ve_fill_io = fio;
316 zio_vdev_io_bypass(zio);
317 zio_add_child(zio, fio);
319 mutex_exit(&vc->vc_lock);
321 VDCSTAT_BUMP(vdc_stat_misses);
327 * Update cache contents upon write completion.
330 vdev_cache_write(zio_t *zio)
332 vdev_cache_t *vc = &zio->io_vd->vdev_cache;
333 vdev_cache_entry_t *ve, ve_search;
334 uint64_t io_start = zio->io_offset;
335 uint64_t io_end = io_start + zio->io_size;
336 uint64_t min_offset = P2ALIGN(io_start, VCBS);
337 uint64_t max_offset = P2ROUNDUP(io_end, VCBS);
340 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
342 mutex_enter(&vc->vc_lock);
344 ve_search.ve_offset = min_offset;
345 ve = avl_find(&vc->vc_offset_tree, &ve_search, &where);
348 ve = avl_nearest(&vc->vc_offset_tree, where, AVL_AFTER);
350 while (ve != NULL && ve->ve_offset < max_offset) {
351 uint64_t start = MAX(ve->ve_offset, io_start);
352 uint64_t end = MIN(ve->ve_offset + VCBS, io_end);
354 if (ve->ve_fill_io != NULL) {
355 ve->ve_missed_update = 1;
357 bcopy((char *)zio->io_data + start - io_start,
358 ve->ve_data + start - ve->ve_offset, end - start);
360 ve = AVL_NEXT(&vc->vc_offset_tree, ve);
362 mutex_exit(&vc->vc_lock);
366 vdev_cache_purge(vdev_t *vd)
368 vdev_cache_t *vc = &vd->vdev_cache;
369 vdev_cache_entry_t *ve;
371 mutex_enter(&vc->vc_lock);
372 while ((ve = avl_first(&vc->vc_offset_tree)) != NULL)
373 vdev_cache_evict(vc, ve);
374 mutex_exit(&vc->vc_lock);
378 vdev_cache_init(vdev_t *vd)
380 vdev_cache_t *vc = &vd->vdev_cache;
382 mutex_init(&vc->vc_lock, NULL, MUTEX_DEFAULT, NULL);
384 avl_create(&vc->vc_offset_tree, vdev_cache_offset_compare,
385 sizeof (vdev_cache_entry_t),
386 offsetof(struct vdev_cache_entry, ve_offset_node));
388 avl_create(&vc->vc_lastused_tree, vdev_cache_lastused_compare,
389 sizeof (vdev_cache_entry_t),
390 offsetof(struct vdev_cache_entry, ve_lastused_node));
394 vdev_cache_fini(vdev_t *vd)
396 vdev_cache_t *vc = &vd->vdev_cache;
398 vdev_cache_purge(vd);
400 avl_destroy(&vc->vc_offset_tree);
401 avl_destroy(&vc->vc_lastused_tree);
403 mutex_destroy(&vc->vc_lock);
407 vdev_cache_stat_init(void)
409 vdc_ksp = kstat_create("zfs", 0, "vdev_cache_stats", "misc",
410 KSTAT_TYPE_NAMED, sizeof (vdc_stats) / sizeof (kstat_named_t),
412 if (vdc_ksp != NULL) {
413 vdc_ksp->ks_data = &vdc_stats;
414 kstat_install(vdc_ksp);
419 vdev_cache_stat_fini(void)
421 if (vdc_ksp != NULL) {
422 kstat_delete(vdc_ksp);
427 #if defined(_KERNEL) && defined(HAVE_SPL)
428 module_param(zfs_vdev_cache_max, int, 0644);
429 MODULE_PARM_DESC(zfs_vdev_cache_max, "Inflate reads small than max");
431 module_param(zfs_vdev_cache_size, int, 0444);
432 MODULE_PARM_DESC(zfs_vdev_cache_size, "Total size of the per-disk cache");
434 module_param(zfs_vdev_cache_bshift, int, 0644);
435 MODULE_PARM_DESC(zfs_vdev_cache_bshift, "Shift size to inflate reads too");