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.
27 * Copyright (c) 2013, 2014 by Delphix. All rights reserved.
30 #include <sys/zfs_context.h>
31 #include <sys/dnode.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dmu_zfetch.h>
36 #include <sys/kstat.h>
39 * This tunable disables predictive prefetch. Note that it leaves "prescient"
40 * prefetch (e.g. prefetch for zfs send) intact. Unlike predictive prefetch,
41 * prescient prefetch never issues i/os that end up not being needed,
42 * so it can't hurt performance.
45 int zfs_prefetch_disable
= B_FALSE
;
47 /* max # of streams per zfetch */
48 unsigned int zfetch_max_streams
= 8;
49 /* min time before stream reclaim */
50 unsigned int zfetch_min_sec_reap
= 2;
51 /* max bytes to prefetch per stream (default 8MB) */
52 unsigned int zfetch_max_distance
= 8 * 1024 * 1024;
53 /* number of bytes in a array_read at which we stop prefetching (1MB) */
54 unsigned long zfetch_array_rd_sz
= 1024 * 1024;
56 typedef struct zfetch_stats
{
57 kstat_named_t zfetchstat_hits
;
58 kstat_named_t zfetchstat_misses
;
59 kstat_named_t zfetchstat_max_streams
;
62 static zfetch_stats_t zfetch_stats
= {
63 { "hits", KSTAT_DATA_UINT64
},
64 { "misses", KSTAT_DATA_UINT64
},
65 { "max_streams", KSTAT_DATA_UINT64
},
68 #define ZFETCHSTAT_BUMP(stat) \
69 atomic_inc_64(&zfetch_stats.stat.value.ui64);
76 zfetch_ksp
= kstat_create("zfs", 0, "zfetchstats", "misc",
77 KSTAT_TYPE_NAMED
, sizeof (zfetch_stats
) / sizeof (kstat_named_t
),
80 if (zfetch_ksp
!= NULL
) {
81 zfetch_ksp
->ks_data
= &zfetch_stats
;
82 kstat_install(zfetch_ksp
);
89 if (zfetch_ksp
!= NULL
) {
90 kstat_delete(zfetch_ksp
);
96 * This takes a pointer to a zfetch structure and a dnode. It performs the
97 * necessary setup for the zfetch structure, grokking data from the
101 dmu_zfetch_init(zfetch_t
*zf
, dnode_t
*dno
)
108 list_create(&zf
->zf_stream
, sizeof (zstream_t
),
109 offsetof(zstream_t
, zs_node
));
111 rw_init(&zf
->zf_rwlock
, NULL
, RW_DEFAULT
, NULL
);
115 dmu_zfetch_stream_remove(zfetch_t
*zf
, zstream_t
*zs
)
117 ASSERT(RW_WRITE_HELD(&zf
->zf_rwlock
));
118 list_remove(&zf
->zf_stream
, zs
);
119 mutex_destroy(&zs
->zs_lock
);
120 kmem_free(zs
, sizeof (*zs
));
124 * Clean-up state associated with a zfetch structure (e.g. destroy the
125 * streams). This doesn't free the zfetch_t itself, that's left to the caller.
128 dmu_zfetch_fini(zfetch_t
*zf
)
132 ASSERT(!RW_LOCK_HELD(&zf
->zf_rwlock
));
134 rw_enter(&zf
->zf_rwlock
, RW_WRITER
);
135 while ((zs
= list_head(&zf
->zf_stream
)) != NULL
)
136 dmu_zfetch_stream_remove(zf
, zs
);
137 rw_exit(&zf
->zf_rwlock
);
138 list_destroy(&zf
->zf_stream
);
139 rw_destroy(&zf
->zf_rwlock
);
145 * If there aren't too many streams already, create a new stream.
146 * The "blkid" argument is the next block that we expect this stream to access.
147 * While we're here, clean up old streams (which haven't been
148 * accessed for at least zfetch_min_sec_reap seconds).
151 dmu_zfetch_stream_create(zfetch_t
*zf
, uint64_t blkid
)
156 uint32_t max_streams
;
158 ASSERT(RW_WRITE_HELD(&zf
->zf_rwlock
));
161 * Clean up old streams.
163 for (zs
= list_head(&zf
->zf_stream
);
164 zs
!= NULL
; zs
= zs_next
) {
165 zs_next
= list_next(&zf
->zf_stream
, zs
);
166 if (((gethrtime() - zs
->zs_atime
) / NANOSEC
) >
168 dmu_zfetch_stream_remove(zf
, zs
);
174 * The maximum number of streams is normally zfetch_max_streams,
175 * but for small files we lower it such that it's at least possible
176 * for all the streams to be non-overlapping.
178 * If we are already at the maximum number of streams for this file,
179 * even after removing old streams, then don't create this stream.
181 max_streams
= MAX(1, MIN(zfetch_max_streams
,
182 zf
->zf_dnode
->dn_maxblkid
* zf
->zf_dnode
->dn_datablksz
/
183 zfetch_max_distance
));
184 if (numstreams
>= max_streams
) {
185 ZFETCHSTAT_BUMP(zfetchstat_max_streams
);
189 zs
= kmem_zalloc(sizeof (*zs
), KM_SLEEP
);
190 zs
->zs_blkid
= blkid
;
191 zs
->zs_pf_blkid
= blkid
;
192 zs
->zs_atime
= gethrtime();
193 mutex_init(&zs
->zs_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
195 list_insert_head(&zf
->zf_stream
, zs
);
199 * This is the prefetch entry point. It calls all of the other dmu_zfetch
200 * routines to create, delete, find, or operate upon prefetch streams.
203 dmu_zfetch(zfetch_t
*zf
, uint64_t blkid
, uint64_t nblks
)
210 if (zfs_prefetch_disable
)
214 * As a fast path for small (single-block) files, ignore access
215 * to the first block.
220 rw_enter(&zf
->zf_rwlock
, RW_READER
);
222 for (zs
= list_head(&zf
->zf_stream
); zs
!= NULL
;
223 zs
= list_next(&zf
->zf_stream
, zs
)) {
224 if (blkid
== zs
->zs_blkid
) {
225 mutex_enter(&zs
->zs_lock
);
227 * zs_blkid could have changed before we
228 * acquired zs_lock; re-check them here.
230 if (blkid
!= zs
->zs_blkid
) {
231 mutex_exit(&zs
->zs_lock
);
240 * This access is not part of any existing stream. Create
241 * a new stream for it.
243 ZFETCHSTAT_BUMP(zfetchstat_misses
);
244 if (rw_tryupgrade(&zf
->zf_rwlock
))
245 dmu_zfetch_stream_create(zf
, blkid
+ nblks
);
246 rw_exit(&zf
->zf_rwlock
);
251 * This access was to a block that we issued a prefetch for on
252 * behalf of this stream. Issue further prefetches for this stream.
254 * Normally, we start prefetching where we stopped
255 * prefetching last (zs_pf_blkid). But when we get our first
256 * hit on this stream, zs_pf_blkid == zs_blkid, we don't
257 * want to prefetch to block we just accessed. In this case,
258 * start just after the block we just accessed.
260 pf_start
= MAX(zs
->zs_pf_blkid
, blkid
+ nblks
);
263 * Double our amount of prefetched data, but don't let the
264 * prefetch get further ahead than zfetch_max_distance.
267 MIN((int64_t)zs
->zs_pf_blkid
- zs
->zs_blkid
+ nblks
,
268 zs
->zs_blkid
+ nblks
+
269 (zfetch_max_distance
>> zf
->zf_dnode
->dn_datablkshift
) - pf_start
);
271 zs
->zs_pf_blkid
= pf_start
+ pf_nblks
;
272 zs
->zs_atime
= gethrtime();
273 zs
->zs_blkid
= blkid
+ nblks
;
276 * dbuf_prefetch() issues the prefetch i/o
277 * asynchronously, but it may need to wait for an
278 * indirect block to be read from disk. Therefore
279 * we do not want to hold any locks while we call it.
281 mutex_exit(&zs
->zs_lock
);
282 rw_exit(&zf
->zf_rwlock
);
283 for (i
= 0; i
< pf_nblks
; i
++) {
284 dbuf_prefetch(zf
->zf_dnode
, 0, pf_start
+ i
,
285 ZIO_PRIORITY_ASYNC_READ
, ARC_FLAG_PREDICTIVE_PREFETCH
);
287 ZFETCHSTAT_BUMP(zfetchstat_hits
);
290 #if defined(_KERNEL) && defined(HAVE_SPL)
291 module_param(zfs_prefetch_disable
, int, 0644);
292 MODULE_PARM_DESC(zfs_prefetch_disable
, "Disable all ZFS prefetching");
294 module_param(zfetch_max_streams
, uint
, 0644);
295 MODULE_PARM_DESC(zfetch_max_streams
, "Max number of streams per zfetch");
297 module_param(zfetch_min_sec_reap
, uint
, 0644);
298 MODULE_PARM_DESC(zfetch_min_sec_reap
, "Min time before stream reclaim");
300 module_param(zfetch_max_distance
, uint
, 0644);
301 MODULE_PARM_DESC(zfetch_max_distance
,
302 "Max bytes to prefetch per stream (default 8MB)");
304 module_param(zfetch_array_rd_sz
, ulong
, 0644);
305 MODULE_PARM_DESC(zfetch_array_rd_sz
, "Number of bytes in a array_read");