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, 2017 by Delphix. All rights reserved.
29 #include <sys/zfs_context.h>
31 #include <sys/vdev_impl.h>
33 #include <sys/kstat.h>
37 * Virtual device read-ahead caching.
39 * This file implements a simple LRU read-ahead cache. When the DMU reads
40 * a given block, it will often want other, nearby blocks soon thereafter.
41 * We take advantage of this by reading a larger disk region and caching
42 * the result. In the best case, this can turn 128 back-to-back 512-byte
43 * reads into a single 64k read followed by 127 cache hits; this reduces
44 * latency dramatically. In the worst case, it can turn an isolated 512-byte
45 * read into a 64k read, which doesn't affect latency all that much but is
46 * terribly wasteful of bandwidth. A more intelligent version of the cache
47 * could keep track of access patterns and not do read-ahead unless it sees
48 * at least two temporally close I/Os to the same region. Currently, only
49 * metadata I/O is inflated. A futher enhancement could take advantage of
50 * more semantic information about the I/O. And it could use something
51 * faster than an AVL tree; that was chosen solely for convenience.
53 * There are five cache operations: allocate, fill, read, write, evict.
55 * (1) Allocate. This reserves a cache entry for the specified region.
56 * We separate the allocate and fill operations so that multiple threads
57 * don't generate I/O for the same cache miss.
59 * (2) Fill. When the I/O for a cache miss completes, the fill routine
60 * places the data in the previously allocated cache entry.
62 * (3) Read. Read data from the cache.
64 * (4) Write. Update cache contents after write completion.
66 * (5) Evict. When allocating a new entry, we evict the oldest (LRU) entry
67 * if the total cache size exceeds zfs_vdev_cache_size.
71 * These tunables are for performance analysis.
74 * All i/os smaller than zfs_vdev_cache_max will be turned into
75 * 1<<zfs_vdev_cache_bshift byte reads by the vdev_cache (aka software
76 * track buffer). At most zfs_vdev_cache_size bytes will be kept in each
79 * TODO: Note that with the current ZFS code, it turns out that the
80 * vdev cache is not helpful, and in some cases actually harmful. It
81 * is better if we disable this. Once some time has passed, we should
82 * actually remove this to simplify the code. For now we just disable
83 * it by setting the zfs_vdev_cache_size to zero. Note that Solaris 11
84 * has made these same changes.
86 int zfs_vdev_cache_max = 1<<14; /* 16KB */
87 int zfs_vdev_cache_size = 0;
88 int zfs_vdev_cache_bshift = 16;
90 #define VCBS (1 << zfs_vdev_cache_bshift) /* 64KB */
92 SYSCTL_DECL(_vfs_zfs_vdev);
93 SYSCTL_NODE(_vfs_zfs_vdev, OID_AUTO, cache, CTLFLAG_RW, 0, "ZFS VDEV Cache");
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 SYSCTL_INT(_vfs_zfs_vdev_cache, OID_AUTO, size, CTLFLAG_RDTUN,
97 &zfs_vdev_cache_size, 0, "Size of VDEV cache");
98 SYSCTL_INT(_vfs_zfs_vdev_cache, OID_AUTO, bshift, CTLFLAG_RDTUN,
99 &zfs_vdev_cache_bshift, 0, "Turn too small requests into 1 << this value");
101 kstat_t *vdc_ksp = NULL;
103 typedef struct vdc_stats {
104 kstat_named_t vdc_stat_delegations;
105 kstat_named_t vdc_stat_hits;
106 kstat_named_t vdc_stat_misses;
109 static vdc_stats_t vdc_stats = {
110 { "delegations", KSTAT_DATA_UINT64 },
111 { "hits", KSTAT_DATA_UINT64 },
112 { "misses", KSTAT_DATA_UINT64 }
115 #define VDCSTAT_BUMP(stat) atomic_inc_64(&vdc_stats.stat.value.ui64);
118 vdev_cache_offset_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_offset < ve2->ve_offset)
125 if (ve1->ve_offset > ve2->ve_offset)
131 vdev_cache_lastused_compare(const void *a1, const void *a2)
133 const vdev_cache_entry_t *ve1 = a1;
134 const vdev_cache_entry_t *ve2 = a2;
136 if (ve1->ve_lastused < ve2->ve_lastused)
138 if (ve1->ve_lastused > ve2->ve_lastused)
142 * Among equally old entries, sort by offset to ensure uniqueness.
144 return (vdev_cache_offset_compare(a1, a2));
148 * Evict the specified entry from the cache.
151 vdev_cache_evict(vdev_cache_t *vc, vdev_cache_entry_t *ve)
153 ASSERT(MUTEX_HELD(&vc->vc_lock));
154 ASSERT3P(ve->ve_fill_io, ==, NULL);
155 ASSERT3P(ve->ve_abd, !=, NULL);
157 avl_remove(&vc->vc_lastused_tree, ve);
158 avl_remove(&vc->vc_offset_tree, ve);
159 abd_free(ve->ve_abd);
160 kmem_free(ve, sizeof (vdev_cache_entry_t));
164 * Allocate an entry in the cache. At the point we don't have the data,
165 * we're just creating a placeholder so that multiple threads don't all
166 * go off and read the same blocks.
168 static vdev_cache_entry_t *
169 vdev_cache_allocate(zio_t *zio)
171 vdev_cache_t *vc = &zio->io_vd->vdev_cache;
172 uint64_t offset = P2ALIGN(zio->io_offset, VCBS);
173 vdev_cache_entry_t *ve;
175 ASSERT(MUTEX_HELD(&vc->vc_lock));
177 if (zfs_vdev_cache_size == 0)
181 * If adding a new entry would exceed the cache size,
182 * evict the oldest entry (LRU).
184 if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) >
185 zfs_vdev_cache_size) {
186 ve = avl_first(&vc->vc_lastused_tree);
187 if (ve->ve_fill_io != NULL)
189 ASSERT3U(ve->ve_hits, !=, 0);
190 vdev_cache_evict(vc, ve);
193 ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_SLEEP);
194 ve->ve_offset = offset;
195 ve->ve_lastused = ddi_get_lbolt();
196 ve->ve_abd = abd_alloc_for_io(VCBS, B_TRUE);
198 avl_add(&vc->vc_offset_tree, ve);
199 avl_add(&vc->vc_lastused_tree, ve);
205 vdev_cache_hit(vdev_cache_t *vc, vdev_cache_entry_t *ve, zio_t *zio)
207 uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
209 ASSERT(MUTEX_HELD(&vc->vc_lock));
210 ASSERT3P(ve->ve_fill_io, ==, NULL);
212 if (ve->ve_lastused != ddi_get_lbolt()) {
213 avl_remove(&vc->vc_lastused_tree, ve);
214 ve->ve_lastused = ddi_get_lbolt();
215 avl_add(&vc->vc_lastused_tree, ve);
219 abd_copy_off(zio->io_abd, ve->ve_abd, 0, cache_phase, zio->io_size);
223 * Fill a previously allocated cache entry with data.
226 vdev_cache_fill(zio_t *fio)
228 vdev_t *vd = fio->io_vd;
229 vdev_cache_t *vc = &vd->vdev_cache;
230 vdev_cache_entry_t *ve = fio->io_private;
233 ASSERT3U(fio->io_size, ==, VCBS);
236 * Add data to the cache.
238 mutex_enter(&vc->vc_lock);
240 ASSERT3P(ve->ve_fill_io, ==, fio);
241 ASSERT3U(ve->ve_offset, ==, fio->io_offset);
242 ASSERT3P(ve->ve_abd, ==, fio->io_abd);
244 ve->ve_fill_io = NULL;
247 * Even if this cache line was invalidated by a missed write update,
248 * any reads that were queued up before the missed update are still
249 * valid, so we can satisfy them from this line before we evict it.
251 zio_link_t *zl = NULL;
252 while ((pio = zio_walk_parents(fio, &zl)) != NULL)
253 vdev_cache_hit(vc, ve, pio);
255 if (fio->io_error || ve->ve_missed_update)
256 vdev_cache_evict(vc, ve);
258 mutex_exit(&vc->vc_lock);
262 * Read data from the cache. Returns B_TRUE cache hit, B_FALSE on miss.
265 vdev_cache_read(zio_t *zio)
267 vdev_cache_t *vc = &zio->io_vd->vdev_cache;
268 vdev_cache_entry_t *ve, ve_search;
269 uint64_t cache_offset = P2ALIGN(zio->io_offset, VCBS);
270 uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
273 ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
275 if (zio->io_flags & ZIO_FLAG_DONT_CACHE)
278 if (zio->io_size > zfs_vdev_cache_max)
282 * If the I/O straddles two or more cache blocks, don't cache it.
284 if (P2BOUNDARY(zio->io_offset, zio->io_size, VCBS))
287 ASSERT3U(cache_phase + zio->io_size, <=, VCBS);
289 mutex_enter(&vc->vc_lock);
291 ve_search.ve_offset = cache_offset;
292 ve = avl_find(&vc->vc_offset_tree, &ve_search, NULL);
295 if (ve->ve_missed_update) {
296 mutex_exit(&vc->vc_lock);
300 if ((fio = ve->ve_fill_io) != NULL) {
301 zio_vdev_io_bypass(zio);
302 zio_add_child(zio, fio);
303 mutex_exit(&vc->vc_lock);
304 VDCSTAT_BUMP(vdc_stat_delegations);
308 vdev_cache_hit(vc, ve, zio);
309 zio_vdev_io_bypass(zio);
311 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_abd, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_NOW,
325 ZIO_FLAG_DONT_CACHE, vdev_cache_fill, ve);
327 ve->ve_fill_io = fio;
328 zio_vdev_io_bypass(zio);
329 zio_add_child(zio, fio);
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 ASSERT3U(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 abd_copy_off(ve->ve_abd, zio->io_abd,
370 start - ve->ve_offset, start - io_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);