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, 2015 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.
44 boolean_t zfs_prefetch_disable = B_FALSE;
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;
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_UQUAD(_vfs_zfs_zfetch, OID_AUTO, array_rd_sz, CTLFLAG_RWTUN,
68 &zfetch_array_rd_sz, 0,
69 "Number of bytes in a array_read at which we stop prefetching");
71 typedef struct zfetch_stats {
72 kstat_named_t zfetchstat_hits;
73 kstat_named_t zfetchstat_misses;
74 kstat_named_t zfetchstat_max_streams;
77 static zfetch_stats_t zfetch_stats = {
78 { "hits", KSTAT_DATA_UINT64 },
79 { "misses", KSTAT_DATA_UINT64 },
80 { "max_streams", KSTAT_DATA_UINT64 },
83 #define ZFETCHSTAT_BUMP(stat) \
84 atomic_inc_64(&zfetch_stats.stat.value.ui64);
91 zfetch_ksp = kstat_create("zfs", 0, "zfetchstats", "misc",
92 KSTAT_TYPE_NAMED, sizeof (zfetch_stats) / sizeof (kstat_named_t),
95 if (zfetch_ksp != NULL) {
96 zfetch_ksp->ks_data = &zfetch_stats;
97 kstat_install(zfetch_ksp);
104 if (zfetch_ksp != NULL) {
105 kstat_delete(zfetch_ksp);
111 * This takes a pointer to a zfetch structure and a dnode. It performs the
112 * necessary setup for the zfetch structure, grokking data from the
116 dmu_zfetch_init(zfetch_t *zf, dnode_t *dno)
123 list_create(&zf->zf_stream, sizeof (zstream_t),
124 offsetof(zstream_t, zs_node));
126 rw_init(&zf->zf_rwlock, NULL, RW_DEFAULT, NULL);
130 dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs)
132 ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
133 list_remove(&zf->zf_stream, zs);
134 mutex_destroy(&zs->zs_lock);
135 kmem_free(zs, sizeof (*zs));
139 * Clean-up state associated with a zfetch structure (e.g. destroy the
140 * streams). This doesn't free the zfetch_t itself, that's left to the caller.
143 dmu_zfetch_fini(zfetch_t *zf)
147 ASSERT(!RW_LOCK_HELD(&zf->zf_rwlock));
149 rw_enter(&zf->zf_rwlock, RW_WRITER);
150 while ((zs = list_head(&zf->zf_stream)) != NULL)
151 dmu_zfetch_stream_remove(zf, zs);
152 rw_exit(&zf->zf_rwlock);
153 list_destroy(&zf->zf_stream);
154 rw_destroy(&zf->zf_rwlock);
160 * If there aren't too many streams already, create a new stream.
161 * The "blkid" argument is the next block that we expect this stream to access.
162 * While we're here, clean up old streams (which haven't been
163 * accessed for at least zfetch_min_sec_reap seconds).
166 dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid)
171 ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
174 * Clean up old streams.
176 for (zstream_t *zs = list_head(&zf->zf_stream);
177 zs != NULL; zs = zs_next) {
178 zs_next = list_next(&zf->zf_stream, zs);
179 if (((gethrtime() - zs->zs_atime) / NANOSEC) >
181 dmu_zfetch_stream_remove(zf, zs);
187 * The maximum number of streams is normally zfetch_max_streams,
188 * but for small files we lower it such that it's at least possible
189 * for all the streams to be non-overlapping.
191 * If we are already at the maximum number of streams for this file,
192 * even after removing old streams, then don't create this stream.
194 uint32_t max_streams = MAX(1, MIN(zfetch_max_streams,
195 zf->zf_dnode->dn_maxblkid * zf->zf_dnode->dn_datablksz /
196 zfetch_max_distance));
197 if (numstreams >= max_streams) {
198 ZFETCHSTAT_BUMP(zfetchstat_max_streams);
202 zstream_t *zs = kmem_zalloc(sizeof (*zs), KM_SLEEP);
203 zs->zs_blkid = blkid;
204 zs->zs_pf_blkid = blkid;
205 zs->zs_ipf_blkid = blkid;
206 zs->zs_atime = gethrtime();
207 mutex_init(&zs->zs_lock, NULL, MUTEX_DEFAULT, NULL);
209 list_insert_head(&zf->zf_stream, zs);
213 * This is the predictive prefetch entry point. It associates dnode access
214 * specified with blkid and nblks arguments with prefetch stream, predicts
215 * further accesses based on that stats and initiates speculative prefetch.
216 * fetch_data argument specifies whether actual data blocks should be fetched:
217 * FALSE -- prefetch only indirect blocks for predicted data blocks;
218 * TRUE -- prefetch predicted data blocks plus following indirect blocks.
221 dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data)
224 int64_t pf_start, ipf_start, ipf_istart, ipf_iend;
225 int64_t pf_ahead_blks, max_blks;
226 int epbs, max_dist_blks, pf_nblks, ipf_nblks;
227 uint64_t end_of_access_blkid = blkid + nblks;
229 if (zfs_prefetch_disable)
233 * As a fast path for small (single-block) files, ignore access
234 * to the first block.
239 rw_enter(&zf->zf_rwlock, RW_READER);
241 for (zs = list_head(&zf->zf_stream); zs != NULL;
242 zs = list_next(&zf->zf_stream, zs)) {
243 if (blkid == zs->zs_blkid) {
244 mutex_enter(&zs->zs_lock);
246 * zs_blkid could have changed before we
247 * acquired zs_lock; re-check them here.
249 if (blkid != zs->zs_blkid) {
250 mutex_exit(&zs->zs_lock);
259 * This access is not part of any existing stream. Create
260 * a new stream for it.
262 ZFETCHSTAT_BUMP(zfetchstat_misses);
263 if (rw_tryupgrade(&zf->zf_rwlock))
264 dmu_zfetch_stream_create(zf, end_of_access_blkid);
265 rw_exit(&zf->zf_rwlock);
270 * This access was to a block that we issued a prefetch for on
271 * behalf of this stream. Issue further prefetches for this stream.
273 * Normally, we start prefetching where we stopped
274 * prefetching last (zs_pf_blkid). But when we get our first
275 * hit on this stream, zs_pf_blkid == zs_blkid, we don't
276 * want to prefetch the block we just accessed. In this case,
277 * start just after the block we just accessed.
279 pf_start = MAX(zs->zs_pf_blkid, end_of_access_blkid);
282 * Double our amount of prefetched data, but don't let the
283 * prefetch get further ahead than zfetch_max_distance.
287 zfetch_max_distance >> zf->zf_dnode->dn_datablkshift;
289 * Previously, we were (zs_pf_blkid - blkid) ahead. We
290 * want to now be double that, so read that amount again,
291 * plus the amount we are catching up by (i.e. the amount
294 pf_ahead_blks = zs->zs_pf_blkid - blkid + nblks;
295 max_blks = max_dist_blks - (pf_start - end_of_access_blkid);
296 pf_nblks = MIN(pf_ahead_blks, max_blks);
301 zs->zs_pf_blkid = pf_start + pf_nblks;
304 * Do the same for indirects, starting from where we stopped last,
305 * or where we will stop reading data blocks (and the indirects
306 * that point to them).
308 ipf_start = MAX(zs->zs_ipf_blkid, zs->zs_pf_blkid);
309 max_dist_blks = zfetch_max_idistance >> zf->zf_dnode->dn_datablkshift;
311 * We want to double our distance ahead of the data prefetch
312 * (or reader, if we are not prefetching data). Previously, we
313 * were (zs_ipf_blkid - blkid) ahead. To double that, we read
314 * that amount again, plus the amount we are catching up by
315 * (i.e. the amount read now + the amount of data prefetched now).
317 pf_ahead_blks = zs->zs_ipf_blkid - blkid + nblks + pf_nblks;
318 max_blks = max_dist_blks - (ipf_start - end_of_access_blkid);
319 ipf_nblks = MIN(pf_ahead_blks, max_blks);
320 zs->zs_ipf_blkid = ipf_start + ipf_nblks;
322 epbs = zf->zf_dnode->dn_indblkshift - SPA_BLKPTRSHIFT;
323 ipf_istart = P2ROUNDUP(ipf_start, 1 << epbs) >> epbs;
324 ipf_iend = P2ROUNDUP(zs->zs_ipf_blkid, 1 << epbs) >> epbs;
326 zs->zs_atime = gethrtime();
327 zs->zs_blkid = end_of_access_blkid;
328 mutex_exit(&zs->zs_lock);
329 rw_exit(&zf->zf_rwlock);
332 * dbuf_prefetch() is asynchronous (even when it needs to read
333 * indirect blocks), but we still prefer to drop our locks before
334 * calling it to reduce the time we hold them.
337 for (int i = 0; i < pf_nblks; i++) {
338 dbuf_prefetch(zf->zf_dnode, 0, pf_start + i,
339 ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
341 for (int64_t iblk = ipf_istart; iblk < ipf_iend; iblk++) {
342 dbuf_prefetch(zf->zf_dnode, 1, iblk,
343 ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
345 ZFETCHSTAT_BUMP(zfetchstat_hits);