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 kstat_t
*vdc_ksp
= NULL
;
94 typedef struct vdc_stats
{
95 kstat_named_t vdc_stat_delegations
;
96 kstat_named_t vdc_stat_hits
;
97 kstat_named_t vdc_stat_misses
;
100 static vdc_stats_t vdc_stats
= {
101 { "delegations", KSTAT_DATA_UINT64
},
102 { "hits", KSTAT_DATA_UINT64
},
103 { "misses", KSTAT_DATA_UINT64
}
106 #define VDCSTAT_BUMP(stat) atomic_inc_64(&vdc_stats.stat.value.ui64);
109 vdev_cache_offset_compare(const void *a1
, const void *a2
)
111 const vdev_cache_entry_t
*ve1
= a1
;
112 const vdev_cache_entry_t
*ve2
= a2
;
114 if (ve1
->ve_offset
< ve2
->ve_offset
)
116 if (ve1
->ve_offset
> ve2
->ve_offset
)
122 vdev_cache_lastused_compare(const void *a1
, const void *a2
)
124 const vdev_cache_entry_t
*ve1
= a1
;
125 const vdev_cache_entry_t
*ve2
= a2
;
127 if (ve1
->ve_lastused
< ve2
->ve_lastused
)
129 if (ve1
->ve_lastused
> ve2
->ve_lastused
)
133 * Among equally old entries, sort by offset to ensure uniqueness.
135 return (vdev_cache_offset_compare(a1
, a2
));
139 * Evict the specified entry from the cache.
142 vdev_cache_evict(vdev_cache_t
*vc
, vdev_cache_entry_t
*ve
)
144 ASSERT(MUTEX_HELD(&vc
->vc_lock
));
145 ASSERT3P(ve
->ve_fill_io
, ==, NULL
);
146 ASSERT3P(ve
->ve_abd
, !=, NULL
);
148 avl_remove(&vc
->vc_lastused_tree
, ve
);
149 avl_remove(&vc
->vc_offset_tree
, ve
);
150 abd_free(ve
->ve_abd
);
151 kmem_free(ve
, sizeof (vdev_cache_entry_t
));
155 * Allocate an entry in the cache. At the point we don't have the data,
156 * we're just creating a placeholder so that multiple threads don't all
157 * go off and read the same blocks.
159 static vdev_cache_entry_t
*
160 vdev_cache_allocate(zio_t
*zio
)
162 vdev_cache_t
*vc
= &zio
->io_vd
->vdev_cache
;
163 uint64_t offset
= P2ALIGN(zio
->io_offset
, VCBS
);
164 vdev_cache_entry_t
*ve
;
166 ASSERT(MUTEX_HELD(&vc
->vc_lock
));
168 if (zfs_vdev_cache_size
== 0)
172 * If adding a new entry would exceed the cache size,
173 * evict the oldest entry (LRU).
175 if ((avl_numnodes(&vc
->vc_lastused_tree
) << zfs_vdev_cache_bshift
) >
176 zfs_vdev_cache_size
) {
177 ve
= avl_first(&vc
->vc_lastused_tree
);
178 if (ve
->ve_fill_io
!= NULL
)
180 ASSERT3U(ve
->ve_hits
, !=, 0);
181 vdev_cache_evict(vc
, ve
);
184 ve
= kmem_zalloc(sizeof (vdev_cache_entry_t
), KM_SLEEP
);
185 ve
->ve_offset
= offset
;
186 ve
->ve_lastused
= ddi_get_lbolt();
187 ve
->ve_abd
= abd_alloc_for_io(VCBS
, B_TRUE
);
189 avl_add(&vc
->vc_offset_tree
, ve
);
190 avl_add(&vc
->vc_lastused_tree
, ve
);
196 vdev_cache_hit(vdev_cache_t
*vc
, vdev_cache_entry_t
*ve
, zio_t
*zio
)
198 uint64_t cache_phase
= P2PHASE(zio
->io_offset
, VCBS
);
200 ASSERT(MUTEX_HELD(&vc
->vc_lock
));
201 ASSERT3P(ve
->ve_fill_io
, ==, NULL
);
203 if (ve
->ve_lastused
!= ddi_get_lbolt()) {
204 avl_remove(&vc
->vc_lastused_tree
, ve
);
205 ve
->ve_lastused
= ddi_get_lbolt();
206 avl_add(&vc
->vc_lastused_tree
, ve
);
210 abd_copy_off(zio
->io_abd
, ve
->ve_abd
, 0, cache_phase
, zio
->io_size
);
214 * Fill a previously allocated cache entry with data.
217 vdev_cache_fill(zio_t
*fio
)
219 vdev_t
*vd
= fio
->io_vd
;
220 vdev_cache_t
*vc
= &vd
->vdev_cache
;
221 vdev_cache_entry_t
*ve
= fio
->io_private
;
224 ASSERT3U(fio
->io_size
, ==, VCBS
);
227 * Add data to the cache.
229 mutex_enter(&vc
->vc_lock
);
231 ASSERT3P(ve
->ve_fill_io
, ==, fio
);
232 ASSERT3U(ve
->ve_offset
, ==, fio
->io_offset
);
233 ASSERT3P(ve
->ve_abd
, ==, fio
->io_abd
);
235 ve
->ve_fill_io
= NULL
;
238 * Even if this cache line was invalidated by a missed write update,
239 * any reads that were queued up before the missed update are still
240 * valid, so we can satisfy them from this line before we evict it.
242 zio_link_t
*zl
= NULL
;
243 while ((pio
= zio_walk_parents(fio
, &zl
)) != NULL
)
244 vdev_cache_hit(vc
, ve
, pio
);
246 if (fio
->io_error
|| ve
->ve_missed_update
)
247 vdev_cache_evict(vc
, ve
);
249 mutex_exit(&vc
->vc_lock
);
253 * Read data from the cache. Returns B_TRUE cache hit, B_FALSE on miss.
256 vdev_cache_read(zio_t
*zio
)
258 vdev_cache_t
*vc
= &zio
->io_vd
->vdev_cache
;
259 vdev_cache_entry_t
*ve
, ve_search
;
260 uint64_t cache_offset
= P2ALIGN(zio
->io_offset
, VCBS
);
261 uint64_t cache_phase
= P2PHASE(zio
->io_offset
, VCBS
);
264 ASSERT3U(zio
->io_type
, ==, ZIO_TYPE_READ
);
266 if (zio
->io_flags
& ZIO_FLAG_DONT_CACHE
)
269 if (zio
->io_size
> zfs_vdev_cache_max
)
273 * If the I/O straddles two or more cache blocks, don't cache it.
275 if (P2BOUNDARY(zio
->io_offset
, zio
->io_size
, VCBS
))
278 ASSERT3U(cache_phase
+ zio
->io_size
, <=, VCBS
);
280 mutex_enter(&vc
->vc_lock
);
282 ve_search
.ve_offset
= cache_offset
;
283 ve
= avl_find(&vc
->vc_offset_tree
, &ve_search
, NULL
);
286 if (ve
->ve_missed_update
) {
287 mutex_exit(&vc
->vc_lock
);
291 if ((fio
= ve
->ve_fill_io
) != NULL
) {
292 zio_vdev_io_bypass(zio
);
293 zio_add_child(zio
, fio
);
294 mutex_exit(&vc
->vc_lock
);
295 VDCSTAT_BUMP(vdc_stat_delegations
);
299 vdev_cache_hit(vc
, ve
, zio
);
300 zio_vdev_io_bypass(zio
);
302 mutex_exit(&vc
->vc_lock
);
303 VDCSTAT_BUMP(vdc_stat_hits
);
307 ve
= vdev_cache_allocate(zio
);
310 mutex_exit(&vc
->vc_lock
);
314 fio
= zio_vdev_delegated_io(zio
->io_vd
, cache_offset
,
315 ve
->ve_abd
, VCBS
, ZIO_TYPE_READ
, ZIO_PRIORITY_NOW
,
316 ZIO_FLAG_DONT_CACHE
, vdev_cache_fill
, ve
);
318 ve
->ve_fill_io
= fio
;
319 zio_vdev_io_bypass(zio
);
320 zio_add_child(zio
, fio
);
322 mutex_exit(&vc
->vc_lock
);
324 VDCSTAT_BUMP(vdc_stat_misses
);
330 * Update cache contents upon write completion.
333 vdev_cache_write(zio_t
*zio
)
335 vdev_cache_t
*vc
= &zio
->io_vd
->vdev_cache
;
336 vdev_cache_entry_t
*ve
, ve_search
;
337 uint64_t io_start
= zio
->io_offset
;
338 uint64_t io_end
= io_start
+ zio
->io_size
;
339 uint64_t min_offset
= P2ALIGN(io_start
, VCBS
);
340 uint64_t max_offset
= P2ROUNDUP(io_end
, VCBS
);
343 ASSERT3U(zio
->io_type
, ==, ZIO_TYPE_WRITE
);
345 mutex_enter(&vc
->vc_lock
);
347 ve_search
.ve_offset
= min_offset
;
348 ve
= avl_find(&vc
->vc_offset_tree
, &ve_search
, &where
);
351 ve
= avl_nearest(&vc
->vc_offset_tree
, where
, AVL_AFTER
);
353 while (ve
!= NULL
&& ve
->ve_offset
< max_offset
) {
354 uint64_t start
= MAX(ve
->ve_offset
, io_start
);
355 uint64_t end
= MIN(ve
->ve_offset
+ VCBS
, io_end
);
357 if (ve
->ve_fill_io
!= NULL
) {
358 ve
->ve_missed_update
= 1;
360 abd_copy_off(ve
->ve_abd
, zio
->io_abd
,
361 start
- ve
->ve_offset
, start
- io_start
,
364 ve
= AVL_NEXT(&vc
->vc_offset_tree
, ve
);
366 mutex_exit(&vc
->vc_lock
);
370 vdev_cache_purge(vdev_t
*vd
)
372 vdev_cache_t
*vc
= &vd
->vdev_cache
;
373 vdev_cache_entry_t
*ve
;
375 mutex_enter(&vc
->vc_lock
);
376 while ((ve
= avl_first(&vc
->vc_offset_tree
)) != NULL
)
377 vdev_cache_evict(vc
, ve
);
378 mutex_exit(&vc
->vc_lock
);
382 vdev_cache_init(vdev_t
*vd
)
384 vdev_cache_t
*vc
= &vd
->vdev_cache
;
386 mutex_init(&vc
->vc_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
388 avl_create(&vc
->vc_offset_tree
, vdev_cache_offset_compare
,
389 sizeof (vdev_cache_entry_t
),
390 offsetof(struct vdev_cache_entry
, ve_offset_node
));
392 avl_create(&vc
->vc_lastused_tree
, vdev_cache_lastused_compare
,
393 sizeof (vdev_cache_entry_t
),
394 offsetof(struct vdev_cache_entry
, ve_lastused_node
));
398 vdev_cache_fini(vdev_t
*vd
)
400 vdev_cache_t
*vc
= &vd
->vdev_cache
;
402 vdev_cache_purge(vd
);
404 avl_destroy(&vc
->vc_offset_tree
);
405 avl_destroy(&vc
->vc_lastused_tree
);
407 mutex_destroy(&vc
->vc_lock
);
411 vdev_cache_stat_init(void)
413 vdc_ksp
= kstat_create("zfs", 0, "vdev_cache_stats", "misc",
414 KSTAT_TYPE_NAMED
, sizeof (vdc_stats
) / sizeof (kstat_named_t
),
416 if (vdc_ksp
!= NULL
) {
417 vdc_ksp
->ks_data
= &vdc_stats
;
418 kstat_install(vdc_ksp
);
423 vdev_cache_stat_fini(void)
425 if (vdc_ksp
!= NULL
) {
426 kstat_delete(vdc_ksp
);