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