1 /*
2 * CDDL HEADER START
3 *
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.
7 *
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.
12 *
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]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
24 */
25
26 /*
27 * Copyright (c) 2013, 2015 by Delphix. All rights reserved.
28 */
29
30 #include <sys/zfs_context.h>
31 #include <sys/dnode.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dmu_zfetch.h>
34 #include <sys/dmu.h>
35 #include <sys/dbuf.h>
36 #include <sys/kstat.h>
37
38 /*
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.
43 */
44 boolean_t zfs_prefetch_disable = B_FALSE;
45
46 /* max # of streams per zfetch */
47 uint32_t zfetch_max_streams = 8;
48 /* min time before stream reclaim */
49 uint32_t zfetch_min_sec_reap = 2;
50 /* max bytes to prefetch per stream (default 8MB) */
51 uint32_t zfetch_max_distance = 8 * 1024 * 1024;
52 /* max number of bytes in an array_read in which we allow prefetching (1MB) */
53 uint64_t zfetch_array_rd_sz = 1024 * 1024;
54
55 SYSCTL_DECL(_vfs_zfs);
56 SYSCTL_INT(_vfs_zfs, OID_AUTO, prefetch_disable, CTLFLAG_RW,
57 &zfs_prefetch_disable, 0, "Disable prefetch");
58 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zfetch, CTLFLAG_RW, 0, "ZFS ZFETCH");
59 SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, max_streams, CTLFLAG_RWTUN,
60 &zfetch_max_streams, 0, "Max # of streams per zfetch");
61 SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, min_sec_reap, CTLFLAG_RWTUN,
62 &zfetch_min_sec_reap, 0, "Min time before stream reclaim");
63 SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, max_distance, CTLFLAG_RWTUN,
64 &zfetch_max_distance, 0, "Max bytes to prefetch per stream");
65 SYSCTL_UQUAD(_vfs_zfs_zfetch, OID_AUTO, array_rd_sz, CTLFLAG_RWTUN,
66 &zfetch_array_rd_sz, 0,
67 "Number of bytes in a array_read at which we stop prefetching");
68
69 typedef struct zfetch_stats {
70 kstat_named_t zfetchstat_hits;
71 kstat_named_t zfetchstat_misses;
72 kstat_named_t zfetchstat_max_streams;
73 } zfetch_stats_t;
74
75 static zfetch_stats_t zfetch_stats = {
76 { "hits", KSTAT_DATA_UINT64 },
77 { "misses", KSTAT_DATA_UINT64 },
78 { "max_streams", KSTAT_DATA_UINT64 },
79 };
80
81 #define ZFETCHSTAT_BUMP(stat) \
82 atomic_inc_64(&zfetch_stats.stat.value.ui64);
83
84 kstat_t *zfetch_ksp;
85
86 void
zfetch_init(void)87 zfetch_init(void)
88 {
89 zfetch_ksp = kstat_create("zfs", 0, "zfetchstats", "misc",
90 KSTAT_TYPE_NAMED, sizeof (zfetch_stats) / sizeof (kstat_named_t),
91 KSTAT_FLAG_VIRTUAL);
92
93 if (zfetch_ksp != NULL) {
94 zfetch_ksp->ks_data = &zfetch_stats;
95 kstat_install(zfetch_ksp);
96 }
97 }
98
99 void
zfetch_fini(void)100 zfetch_fini(void)
101 {
102 if (zfetch_ksp != NULL) {
103 kstat_delete(zfetch_ksp);
104 zfetch_ksp = NULL;
105 }
106 }
107
108 /*
109 * This takes a pointer to a zfetch structure and a dnode. It performs the
110 * necessary setup for the zfetch structure, grokking data from the
111 * associated dnode.
112 */
113 void
dmu_zfetch_init(zfetch_t * zf,dnode_t * dno)114 dmu_zfetch_init(zfetch_t *zf, dnode_t *dno)
115 {
116 if (zf == NULL)
117 return;
118
119 zf->zf_dnode = dno;
120
121 list_create(&zf->zf_stream, sizeof (zstream_t),
122 offsetof(zstream_t, zs_node));
123
124 rw_init(&zf->zf_rwlock, NULL, RW_DEFAULT, NULL);
125 }
126
127 static void
dmu_zfetch_stream_remove(zfetch_t * zf,zstream_t * zs)128 dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs)
129 {
130 ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
131 list_remove(&zf->zf_stream, zs);
132 mutex_destroy(&zs->zs_lock);
133 kmem_free(zs, sizeof (*zs));
134 }
135
136 /*
137 * Clean-up state associated with a zfetch structure (e.g. destroy the
138 * streams). This doesn't free the zfetch_t itself, that's left to the caller.
139 */
140 void
dmu_zfetch_fini(zfetch_t * zf)141 dmu_zfetch_fini(zfetch_t *zf)
142 {
143 zstream_t *zs;
144
145 ASSERT(!RW_LOCK_HELD(&zf->zf_rwlock));
146
147 rw_enter(&zf->zf_rwlock, RW_WRITER);
148 while ((zs = list_head(&zf->zf_stream)) != NULL)
149 dmu_zfetch_stream_remove(zf, zs);
150 rw_exit(&zf->zf_rwlock);
151 list_destroy(&zf->zf_stream);
152 rw_destroy(&zf->zf_rwlock);
153
154 zf->zf_dnode = NULL;
155 }
156
157 /*
158 * If there aren't too many streams already, create a new stream.
159 * The "blkid" argument is the next block that we expect this stream to access.
160 * While we're here, clean up old streams (which haven't been
161 * accessed for at least zfetch_min_sec_reap seconds).
162 */
163 static void
dmu_zfetch_stream_create(zfetch_t * zf,uint64_t blkid)164 dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid)
165 {
166 zstream_t *zs_next;
167 int numstreams = 0;
168
169 ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
170
171 /*
172 * Clean up old streams.
173 */
174 for (zstream_t *zs = list_head(&zf->zf_stream);
175 zs != NULL; zs = zs_next) {
176 zs_next = list_next(&zf->zf_stream, zs);
177 if (((gethrtime() - zs->zs_atime) / NANOSEC) >
178 zfetch_min_sec_reap)
179 dmu_zfetch_stream_remove(zf, zs);
180 else
181 numstreams++;
182 }
183
184 /*
185 * The maximum number of streams is normally zfetch_max_streams,
186 * but for small files we lower it such that it's at least possible
187 * for all the streams to be non-overlapping.
188 *
189 * If we are already at the maximum number of streams for this file,
190 * even after removing old streams, then don't create this stream.
191 */
192 uint32_t max_streams = MAX(1, MIN(zfetch_max_streams,
193 zf->zf_dnode->dn_maxblkid * zf->zf_dnode->dn_datablksz /
194 zfetch_max_distance));
195 if (numstreams >= max_streams) {
196 ZFETCHSTAT_BUMP(zfetchstat_max_streams);
197 return;
198 }
199
200 zstream_t *zs = kmem_zalloc(sizeof (*zs), KM_SLEEP);
201 zs->zs_blkid = blkid;
202 zs->zs_pf_blkid = blkid;
203 zs->zs_atime = gethrtime();
204 mutex_init(&zs->zs_lock, NULL, MUTEX_DEFAULT, NULL);
205
206 list_insert_head(&zf->zf_stream, zs);
207 }
208
209 /*
210 * This is the prefetch entry point. It calls all of the other dmu_zfetch
211 * routines to create, delete, find, or operate upon prefetch streams.
212 */
213 void
dmu_zfetch(zfetch_t * zf,uint64_t blkid,uint64_t nblks)214 dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks)
215 {
216 zstream_t *zs;
217
218 if (zfs_prefetch_disable)
219 return;
220
221 /*
222 * As a fast path for small (single-block) files, ignore access
223 * to the first block.
224 */
225 if (blkid == 0)
226 return;
227
228 rw_enter(&zf->zf_rwlock, RW_READER);
229
230 for (zs = list_head(&zf->zf_stream); zs != NULL;
231 zs = list_next(&zf->zf_stream, zs)) {
232 if (blkid == zs->zs_blkid) {
233 mutex_enter(&zs->zs_lock);
234 /*
235 * zs_blkid could have changed before we
236 * acquired zs_lock; re-check them here.
237 */
238 if (blkid != zs->zs_blkid) {
239 mutex_exit(&zs->zs_lock);
240 continue;
241 }
242 break;
243 }
244 }
245
246 if (zs == NULL) {
247 /*
248 * This access is not part of any existing stream. Create
249 * a new stream for it.
250 */
251 ZFETCHSTAT_BUMP(zfetchstat_misses);
252 if (rw_tryupgrade(&zf->zf_rwlock))
253 dmu_zfetch_stream_create(zf, blkid + nblks);
254 rw_exit(&zf->zf_rwlock);
255 return;
256 }
257
258 /*
259 * This access was to a block that we issued a prefetch for on
260 * behalf of this stream. Issue further prefetches for this stream.
261 *
262 * Normally, we start prefetching where we stopped
263 * prefetching last (zs_pf_blkid). But when we get our first
264 * hit on this stream, zs_pf_blkid == zs_blkid, we don't
265 * want to prefetch to block we just accessed. In this case,
266 * start just after the block we just accessed.
267 */
268 int64_t pf_start = MAX(zs->zs_pf_blkid, blkid + nblks);
269
270 /*
271 * Double our amount of prefetched data, but don't let the
272 * prefetch get further ahead than zfetch_max_distance.
273 */
274 int pf_nblks =
275 MIN((int64_t)zs->zs_pf_blkid - zs->zs_blkid + nblks,
276 zs->zs_blkid + nblks +
277 (zfetch_max_distance >> zf->zf_dnode->dn_datablkshift) - pf_start);
278
279 zs->zs_pf_blkid = pf_start + pf_nblks;
280 zs->zs_atime = gethrtime();
281 zs->zs_blkid = blkid + nblks;
282
283 /*
284 * dbuf_prefetch() issues the prefetch i/o
285 * asynchronously, but it may need to wait for an
286 * indirect block to be read from disk. Therefore
287 * we do not want to hold any locks while we call it.
288 */
289 mutex_exit(&zs->zs_lock);
290 rw_exit(&zf->zf_rwlock);
291 for (int i = 0; i < pf_nblks; i++) {
292 dbuf_prefetch(zf->zf_dnode, 0, pf_start + i,
293 ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
294 }
295 ZFETCHSTAT_BUMP(zfetchstat_hits);
296 }
297