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 | CTLFLAG_MPSAFE, 0,
62 SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, max_streams, CTLFLAG_RWTUN,
63 &zfetch_max_streams, 0, "Max # of streams per zfetch");
64 SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, min_sec_reap, CTLFLAG_RWTUN,
65 &zfetch_min_sec_reap, 0, "Min time before stream reclaim");
66 SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, max_distance, CTLFLAG_RWTUN,
67 &zfetch_max_distance, 0, "Max bytes to prefetch per stream");
68 SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, max_idistance, CTLFLAG_RWTUN,
69 &zfetch_max_idistance, 0, "Max bytes to prefetch indirects for per stream");
70 SYSCTL_UQUAD(_vfs_zfs_zfetch, OID_AUTO, array_rd_sz, CTLFLAG_RWTUN,
71 &zfetch_array_rd_sz, 0,
72 "Number of bytes in a array_read at which we stop prefetching");
74 typedef struct zfetch_stats {
75 kstat_named_t zfetchstat_hits;
76 kstat_named_t zfetchstat_misses;
77 kstat_named_t zfetchstat_max_streams;
80 static zfetch_stats_t zfetch_stats = {
81 { "hits", KSTAT_DATA_UINT64 },
82 { "misses", KSTAT_DATA_UINT64 },
83 { "max_streams", KSTAT_DATA_UINT64 },
86 #define ZFETCHSTAT_BUMP(stat) \
87 atomic_inc_64(&zfetch_stats.stat.value.ui64);
94 zfetch_ksp = kstat_create("zfs", 0, "zfetchstats", "misc",
95 KSTAT_TYPE_NAMED, sizeof (zfetch_stats) / sizeof (kstat_named_t),
98 if (zfetch_ksp != NULL) {
99 zfetch_ksp->ks_data = &zfetch_stats;
100 kstat_install(zfetch_ksp);
107 if (zfetch_ksp != NULL) {
108 kstat_delete(zfetch_ksp);
114 * This takes a pointer to a zfetch structure and a dnode. It performs the
115 * necessary setup for the zfetch structure, grokking data from the
119 dmu_zfetch_init(zfetch_t *zf, dnode_t *dno)
126 list_create(&zf->zf_stream, sizeof (zstream_t),
127 offsetof(zstream_t, zs_node));
129 rw_init(&zf->zf_rwlock, NULL, RW_DEFAULT, NULL);
133 dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs)
135 ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
136 list_remove(&zf->zf_stream, zs);
137 mutex_destroy(&zs->zs_lock);
138 kmem_free(zs, sizeof (*zs));
142 * Clean-up state associated with a zfetch structure (e.g. destroy the
143 * streams). This doesn't free the zfetch_t itself, that's left to the caller.
146 dmu_zfetch_fini(zfetch_t *zf)
150 ASSERT(!RW_LOCK_HELD(&zf->zf_rwlock));
152 rw_enter(&zf->zf_rwlock, RW_WRITER);
153 while ((zs = list_head(&zf->zf_stream)) != NULL)
154 dmu_zfetch_stream_remove(zf, zs);
155 rw_exit(&zf->zf_rwlock);
156 list_destroy(&zf->zf_stream);
157 rw_destroy(&zf->zf_rwlock);
163 * If there aren't too many streams already, create a new stream.
164 * The "blkid" argument is the next block that we expect this stream to access.
165 * While we're here, clean up old streams (which haven't been
166 * accessed for at least zfetch_min_sec_reap seconds).
169 dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid)
174 ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
177 * Clean up old streams.
179 for (zstream_t *zs = list_head(&zf->zf_stream);
180 zs != NULL; zs = zs_next) {
181 zs_next = list_next(&zf->zf_stream, zs);
182 if (((gethrtime() - zs->zs_atime) / NANOSEC) >
184 dmu_zfetch_stream_remove(zf, zs);
190 * The maximum number of streams is normally zfetch_max_streams,
191 * but for small files we lower it such that it's at least possible
192 * for all the streams to be non-overlapping.
194 * If we are already at the maximum number of streams for this file,
195 * even after removing old streams, then don't create this stream.
197 uint32_t max_streams = MAX(1, MIN(zfetch_max_streams,
198 zf->zf_dnode->dn_maxblkid * zf->zf_dnode->dn_datablksz /
199 zfetch_max_distance));
200 if (numstreams >= max_streams) {
201 ZFETCHSTAT_BUMP(zfetchstat_max_streams);
205 zstream_t *zs = kmem_zalloc(sizeof (*zs), KM_SLEEP);
206 zs->zs_blkid = blkid;
207 zs->zs_pf_blkid = blkid;
208 zs->zs_ipf_blkid = blkid;
209 zs->zs_atime = gethrtime();
210 mutex_init(&zs->zs_lock, NULL, MUTEX_DEFAULT, NULL);
212 list_insert_head(&zf->zf_stream, zs);
216 * This is the predictive prefetch entry point. It associates dnode access
217 * specified with blkid and nblks arguments with prefetch stream, predicts
218 * further accesses based on that stats and initiates speculative prefetch.
219 * fetch_data argument specifies whether actual data blocks should be fetched:
220 * FALSE -- prefetch only indirect blocks for predicted data blocks;
221 * TRUE -- prefetch predicted data blocks plus following indirect blocks.
224 dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data)
227 int64_t pf_start, ipf_start, ipf_istart, ipf_iend;
228 int64_t pf_ahead_blks, max_blks;
229 int epbs, max_dist_blks, pf_nblks, ipf_nblks;
230 uint64_t end_of_access_blkid = blkid + nblks;
231 spa_t *spa = zf->zf_dnode->dn_objset->os_spa;
233 if (zfs_prefetch_disable)
237 * If we haven't yet loaded the indirect vdevs' mappings, we
238 * can only read from blocks that we carefully ensure are on
239 * concrete vdevs (or previously-loaded indirect vdevs). So we
240 * can't allow the predictive prefetcher to attempt reads of other
241 * blocks (e.g. of the MOS's dnode obejct).
243 if (!spa_indirect_vdevs_loaded(spa))
247 * As a fast path for small (single-block) files, ignore access
248 * to the first block.
253 rw_enter(&zf->zf_rwlock, RW_READER);
256 * Find matching prefetch stream. Depending on whether the accesses
257 * are block-aligned, first block of the new access may either follow
258 * the last block of the previous access, or be equal to it.
260 for (zs = list_head(&zf->zf_stream); zs != NULL;
261 zs = list_next(&zf->zf_stream, zs)) {
262 if (blkid == zs->zs_blkid || blkid + 1 == zs->zs_blkid) {
263 mutex_enter(&zs->zs_lock);
265 * zs_blkid could have changed before we
266 * acquired zs_lock; re-check them here.
268 if (blkid == zs->zs_blkid) {
270 } else if (blkid + 1 == zs->zs_blkid) {
274 /* Already prefetched this before. */
275 mutex_exit(&zs->zs_lock);
276 rw_exit(&zf->zf_rwlock);
281 mutex_exit(&zs->zs_lock);
287 * This access is not part of any existing stream. Create
288 * a new stream for it.
290 ZFETCHSTAT_BUMP(zfetchstat_misses);
291 if (rw_tryupgrade(&zf->zf_rwlock))
292 dmu_zfetch_stream_create(zf, end_of_access_blkid);
293 rw_exit(&zf->zf_rwlock);
298 * This access was to a block that we issued a prefetch for on
299 * behalf of this stream. Issue further prefetches for this stream.
301 * Normally, we start prefetching where we stopped
302 * prefetching last (zs_pf_blkid). But when we get our first
303 * hit on this stream, zs_pf_blkid == zs_blkid, we don't
304 * want to prefetch the block we just accessed. In this case,
305 * start just after the block we just accessed.
307 pf_start = MAX(zs->zs_pf_blkid, end_of_access_blkid);
310 * Double our amount of prefetched data, but don't let the
311 * prefetch get further ahead than zfetch_max_distance.
315 zfetch_max_distance >> zf->zf_dnode->dn_datablkshift;
317 * Previously, we were (zs_pf_blkid - blkid) ahead. We
318 * want to now be double that, so read that amount again,
319 * plus the amount we are catching up by (i.e. the amount
322 pf_ahead_blks = zs->zs_pf_blkid - blkid + nblks;
323 max_blks = max_dist_blks - (pf_start - end_of_access_blkid);
324 pf_nblks = MIN(pf_ahead_blks, max_blks);
329 zs->zs_pf_blkid = pf_start + pf_nblks;
332 * Do the same for indirects, starting from where we stopped last,
333 * or where we will stop reading data blocks (and the indirects
334 * that point to them).
336 ipf_start = MAX(zs->zs_ipf_blkid, zs->zs_pf_blkid);
337 max_dist_blks = zfetch_max_idistance >> zf->zf_dnode->dn_datablkshift;
339 * We want to double our distance ahead of the data prefetch
340 * (or reader, if we are not prefetching data). Previously, we
341 * were (zs_ipf_blkid - blkid) ahead. To double that, we read
342 * that amount again, plus the amount we are catching up by
343 * (i.e. the amount read now + the amount of data prefetched now).
345 pf_ahead_blks = zs->zs_ipf_blkid - blkid + nblks + pf_nblks;
346 max_blks = max_dist_blks - (ipf_start - end_of_access_blkid);
347 ipf_nblks = MIN(pf_ahead_blks, max_blks);
348 zs->zs_ipf_blkid = ipf_start + ipf_nblks;
350 epbs = zf->zf_dnode->dn_indblkshift - SPA_BLKPTRSHIFT;
351 ipf_istart = P2ROUNDUP(ipf_start, 1 << epbs) >> epbs;
352 ipf_iend = P2ROUNDUP(zs->zs_ipf_blkid, 1 << epbs) >> epbs;
354 zs->zs_atime = gethrtime();
355 zs->zs_blkid = end_of_access_blkid;
356 mutex_exit(&zs->zs_lock);
357 rw_exit(&zf->zf_rwlock);
360 * dbuf_prefetch() is asynchronous (even when it needs to read
361 * indirect blocks), but we still prefer to drop our locks before
362 * calling it to reduce the time we hold them.
365 for (int i = 0; i < pf_nblks; i++) {
366 dbuf_prefetch(zf->zf_dnode, 0, pf_start + i,
367 ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
369 for (int64_t iblk = ipf_istart; iblk < ipf_iend; iblk++) {
370 dbuf_prefetch(zf->zf_dnode, 1, iblk,
371 ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
373 ZFETCHSTAT_BUMP(zfetchstat_hits);