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 bytes to prefetch indirects for per stream (default 64MB) */
53 uint32_t zfetch_max_idistance = 64 * 1024 * 1024;
54 /* max number of bytes in an array_read in which we allow prefetching (1MB) */
55 uint64_t zfetch_array_rd_sz = 1024 * 1024;
56
57 SYSCTL_DECL(_vfs_zfs);
58 SYSCTL_INT(_vfs_zfs, OID_AUTO, prefetch_disable, CTLFLAG_RW,
59 &zfs_prefetch_disable, 0, "Disable prefetch");
60 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zfetch, CTLFLAG_RW, 0, "ZFS ZFETCH");
61 SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, max_streams, CTLFLAG_RWTUN,
62 &zfetch_max_streams, 0, "Max # of streams per zfetch");
63 SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, min_sec_reap, CTLFLAG_RWTUN,
64 &zfetch_min_sec_reap, 0, "Min time before stream reclaim");
65 SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, max_distance, CTLFLAG_RWTUN,
66 &zfetch_max_distance, 0, "Max bytes to prefetch per stream");
67 SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, max_idistance, CTLFLAG_RWTUN,
68 &zfetch_max_idistance, 0, "Max bytes to prefetch indirects for per stream");
69 SYSCTL_UQUAD(_vfs_zfs_zfetch, OID_AUTO, array_rd_sz, CTLFLAG_RWTUN,
70 &zfetch_array_rd_sz, 0,
71 "Number of bytes in a array_read at which we stop prefetching");
72
73 typedef struct zfetch_stats {
74 kstat_named_t zfetchstat_hits;
75 kstat_named_t zfetchstat_misses;
76 kstat_named_t zfetchstat_max_streams;
77 } zfetch_stats_t;
78
79 static zfetch_stats_t zfetch_stats = {
80 { "hits", KSTAT_DATA_UINT64 },
81 { "misses", KSTAT_DATA_UINT64 },
82 { "max_streams", KSTAT_DATA_UINT64 },
83 };
84
85 #define ZFETCHSTAT_BUMP(stat) \
86 atomic_inc_64(&zfetch_stats.stat.value.ui64);
87
88 kstat_t *zfetch_ksp;
89
90 void
zfetch_init(void)91 zfetch_init(void)
92 {
93 zfetch_ksp = kstat_create("zfs", 0, "zfetchstats", "misc",
94 KSTAT_TYPE_NAMED, sizeof (zfetch_stats) / sizeof (kstat_named_t),
95 KSTAT_FLAG_VIRTUAL);
96
97 if (zfetch_ksp != NULL) {
98 zfetch_ksp->ks_data = &zfetch_stats;
99 kstat_install(zfetch_ksp);
100 }
101 }
102
103 void
zfetch_fini(void)104 zfetch_fini(void)
105 {
106 if (zfetch_ksp != NULL) {
107 kstat_delete(zfetch_ksp);
108 zfetch_ksp = NULL;
109 }
110 }
111
112 /*
113 * This takes a pointer to a zfetch structure and a dnode. It performs the
114 * necessary setup for the zfetch structure, grokking data from the
115 * associated dnode.
116 */
117 void
dmu_zfetch_init(zfetch_t * zf,dnode_t * dno)118 dmu_zfetch_init(zfetch_t *zf, dnode_t *dno)
119 {
120 if (zf == NULL)
121 return;
122
123 zf->zf_dnode = dno;
124
125 list_create(&zf->zf_stream, sizeof (zstream_t),
126 offsetof(zstream_t, zs_node));
127
128 rw_init(&zf->zf_rwlock, NULL, RW_DEFAULT, NULL);
129 }
130
131 static void
dmu_zfetch_stream_remove(zfetch_t * zf,zstream_t * zs)132 dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs)
133 {
134 ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
135 list_remove(&zf->zf_stream, zs);
136 mutex_destroy(&zs->zs_lock);
137 kmem_free(zs, sizeof (*zs));
138 }
139
140 /*
141 * Clean-up state associated with a zfetch structure (e.g. destroy the
142 * streams). This doesn't free the zfetch_t itself, that's left to the caller.
143 */
144 void
dmu_zfetch_fini(zfetch_t * zf)145 dmu_zfetch_fini(zfetch_t *zf)
146 {
147 zstream_t *zs;
148
149 ASSERT(!RW_LOCK_HELD(&zf->zf_rwlock));
150
151 rw_enter(&zf->zf_rwlock, RW_WRITER);
152 while ((zs = list_head(&zf->zf_stream)) != NULL)
153 dmu_zfetch_stream_remove(zf, zs);
154 rw_exit(&zf->zf_rwlock);
155 list_destroy(&zf->zf_stream);
156 rw_destroy(&zf->zf_rwlock);
157
158 zf->zf_dnode = NULL;
159 }
160
161 /*
162 * If there aren't too many streams already, create a new stream.
163 * The "blkid" argument is the next block that we expect this stream to access.
164 * While we're here, clean up old streams (which haven't been
165 * accessed for at least zfetch_min_sec_reap seconds).
166 */
167 static void
dmu_zfetch_stream_create(zfetch_t * zf,uint64_t blkid)168 dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid)
169 {
170 zstream_t *zs_next;
171 int numstreams = 0;
172
173 ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
174
175 /*
176 * Clean up old streams.
177 */
178 for (zstream_t *zs = list_head(&zf->zf_stream);
179 zs != NULL; zs = zs_next) {
180 zs_next = list_next(&zf->zf_stream, zs);
181 if (((gethrtime() - zs->zs_atime) / NANOSEC) >
182 zfetch_min_sec_reap)
183 dmu_zfetch_stream_remove(zf, zs);
184 else
185 numstreams++;
186 }
187
188 /*
189 * The maximum number of streams is normally zfetch_max_streams,
190 * but for small files we lower it such that it's at least possible
191 * for all the streams to be non-overlapping.
192 *
193 * If we are already at the maximum number of streams for this file,
194 * even after removing old streams, then don't create this stream.
195 */
196 uint32_t max_streams = MAX(1, MIN(zfetch_max_streams,
197 zf->zf_dnode->dn_maxblkid * zf->zf_dnode->dn_datablksz /
198 zfetch_max_distance));
199 if (numstreams >= max_streams) {
200 ZFETCHSTAT_BUMP(zfetchstat_max_streams);
201 return;
202 }
203
204 zstream_t *zs = kmem_zalloc(sizeof (*zs), KM_SLEEP);
205 zs->zs_blkid = blkid;
206 zs->zs_pf_blkid = blkid;
207 zs->zs_ipf_blkid = blkid;
208 zs->zs_atime = gethrtime();
209 mutex_init(&zs->zs_lock, NULL, MUTEX_DEFAULT, NULL);
210
211 list_insert_head(&zf->zf_stream, zs);
212 }
213
214 /*
215 * This is the predictive prefetch entry point. It associates dnode access
216 * specified with blkid and nblks arguments with prefetch stream, predicts
217 * further accesses based on that stats and initiates speculative prefetch.
218 * fetch_data argument specifies whether actual data blocks should be fetched:
219 * FALSE -- prefetch only indirect blocks for predicted data blocks;
220 * TRUE -- prefetch predicted data blocks plus following indirect blocks.
221 */
222 void
dmu_zfetch(zfetch_t * zf,uint64_t blkid,uint64_t nblks,boolean_t fetch_data)223 dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data)
224 {
225 zstream_t *zs;
226 int64_t pf_start, ipf_start, ipf_istart, ipf_iend;
227 int64_t pf_ahead_blks, max_blks;
228 int epbs, max_dist_blks, pf_nblks, ipf_nblks;
229 uint64_t end_of_access_blkid = blkid + nblks;
230
231 if (zfs_prefetch_disable)
232 return;
233
234 /*
235 * As a fast path for small (single-block) files, ignore access
236 * to the first block.
237 */
238 if (blkid == 0)
239 return;
240
241 rw_enter(&zf->zf_rwlock, RW_READER);
242
243 for (zs = list_head(&zf->zf_stream); zs != NULL;
244 zs = list_next(&zf->zf_stream, zs)) {
245 if (blkid == zs->zs_blkid) {
246 mutex_enter(&zs->zs_lock);
247 /*
248 * zs_blkid could have changed before we
249 * acquired zs_lock; re-check them here.
250 */
251 if (blkid != zs->zs_blkid) {
252 mutex_exit(&zs->zs_lock);
253 continue;
254 }
255 break;
256 }
257 }
258
259 if (zs == NULL) {
260 /*
261 * This access is not part of any existing stream. Create
262 * a new stream for it.
263 */
264 ZFETCHSTAT_BUMP(zfetchstat_misses);
265 if (rw_tryupgrade(&zf->zf_rwlock))
266 dmu_zfetch_stream_create(zf, end_of_access_blkid);
267 rw_exit(&zf->zf_rwlock);
268 return;
269 }
270
271 /*
272 * This access was to a block that we issued a prefetch for on
273 * behalf of this stream. Issue further prefetches for this stream.
274 *
275 * Normally, we start prefetching where we stopped
276 * prefetching last (zs_pf_blkid). But when we get our first
277 * hit on this stream, zs_pf_blkid == zs_blkid, we don't
278 * want to prefetch the block we just accessed. In this case,
279 * start just after the block we just accessed.
280 */
281 pf_start = MAX(zs->zs_pf_blkid, end_of_access_blkid);
282
283 /*
284 * Double our amount of prefetched data, but don't let the
285 * prefetch get further ahead than zfetch_max_distance.
286 */
287 if (fetch_data) {
288 max_dist_blks =
289 zfetch_max_distance >> zf->zf_dnode->dn_datablkshift;
290 /*
291 * Previously, we were (zs_pf_blkid - blkid) ahead. We
292 * want to now be double that, so read that amount again,
293 * plus the amount we are catching up by (i.e. the amount
294 * read just now).
295 */
296 pf_ahead_blks = zs->zs_pf_blkid - blkid + nblks;
297 max_blks = max_dist_blks - (pf_start - end_of_access_blkid);
298 pf_nblks = MIN(pf_ahead_blks, max_blks);
299 } else {
300 pf_nblks = 0;
301 }
302
303 zs->zs_pf_blkid = pf_start + pf_nblks;
304
305 /*
306 * Do the same for indirects, starting from where we stopped last,
307 * or where we will stop reading data blocks (and the indirects
308 * that point to them).
309 */
310 ipf_start = MAX(zs->zs_ipf_blkid, zs->zs_pf_blkid);
311 max_dist_blks = zfetch_max_idistance >> zf->zf_dnode->dn_datablkshift;
312 /*
313 * We want to double our distance ahead of the data prefetch
314 * (or reader, if we are not prefetching data). Previously, we
315 * were (zs_ipf_blkid - blkid) ahead. To double that, we read
316 * that amount again, plus the amount we are catching up by
317 * (i.e. the amount read now + the amount of data prefetched now).
318 */
319 pf_ahead_blks = zs->zs_ipf_blkid - blkid + nblks + pf_nblks;
320 max_blks = max_dist_blks - (ipf_start - end_of_access_blkid);
321 ipf_nblks = MIN(pf_ahead_blks, max_blks);
322 zs->zs_ipf_blkid = ipf_start + ipf_nblks;
323
324 epbs = zf->zf_dnode->dn_indblkshift - SPA_BLKPTRSHIFT;
325 ipf_istart = P2ROUNDUP(ipf_start, 1 << epbs) >> epbs;
326 ipf_iend = P2ROUNDUP(zs->zs_ipf_blkid, 1 << epbs) >> epbs;
327
328 zs->zs_atime = gethrtime();
329 zs->zs_blkid = end_of_access_blkid;
330 mutex_exit(&zs->zs_lock);
331 rw_exit(&zf->zf_rwlock);
332
333 /*
334 * dbuf_prefetch() is asynchronous (even when it needs to read
335 * indirect blocks), but we still prefer to drop our locks before
336 * calling it to reduce the time we hold them.
337 */
338
339 for (int i = 0; i < pf_nblks; i++) {
340 dbuf_prefetch(zf->zf_dnode, 0, pf_start + i,
341 ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
342 }
343 for (int64_t iblk = ipf_istart; iblk < ipf_iend; iblk++) {
344 dbuf_prefetch(zf->zf_dnode, 1, iblk,
345 ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
346 }
347 ZFETCHSTAT_BUMP(zfetchstat_hits);
348 }
349