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.
45 int zfs_prefetch_disable = B_FALSE;
47 /* max # of streams per zfetch */
48 unsigned int zfetch_max_streams = 8;
49 /* min time before stream reclaim */
50 unsigned int zfetch_min_sec_reap = 2;
51 /* max bytes to prefetch per stream (default 8MB) */
52 unsigned int zfetch_max_distance = 8 * 1024 * 1024;
53 /* max bytes to prefetch indirects for per stream (default 64MB) */
54 unsigned int zfetch_max_idistance = 64 * 1024 * 1024;
55 /* max number of bytes in an array_read in which we allow prefetching (1MB) */
56 unsigned long zfetch_array_rd_sz = 1024 * 1024;
58 typedef struct zfetch_stats {
59 kstat_named_t zfetchstat_hits;
60 kstat_named_t zfetchstat_misses;
61 kstat_named_t zfetchstat_max_streams;
64 static zfetch_stats_t zfetch_stats = {
65 { "hits", KSTAT_DATA_UINT64 },
66 { "misses", KSTAT_DATA_UINT64 },
67 { "max_streams", KSTAT_DATA_UINT64 },
70 #define ZFETCHSTAT_BUMP(stat) \
71 atomic_inc_64(&zfetch_stats.stat.value.ui64);
78 zfetch_ksp = kstat_create("zfs", 0, "zfetchstats", "misc",
79 KSTAT_TYPE_NAMED, sizeof (zfetch_stats) / sizeof (kstat_named_t),
82 if (zfetch_ksp != NULL) {
83 zfetch_ksp->ks_data = &zfetch_stats;
84 kstat_install(zfetch_ksp);
91 if (zfetch_ksp != NULL) {
92 kstat_delete(zfetch_ksp);
98 * This takes a pointer to a zfetch structure and a dnode. It performs the
99 * necessary setup for the zfetch structure, grokking data from the
103 dmu_zfetch_init(zfetch_t *zf, dnode_t *dno)
110 list_create(&zf->zf_stream, sizeof (zstream_t),
111 offsetof(zstream_t, zs_node));
113 rw_init(&zf->zf_rwlock, NULL, RW_DEFAULT, NULL);
117 dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs)
119 ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
120 list_remove(&zf->zf_stream, zs);
121 mutex_destroy(&zs->zs_lock);
122 kmem_free(zs, sizeof (*zs));
126 * Clean-up state associated with a zfetch structure (e.g. destroy the
127 * streams). This doesn't free the zfetch_t itself, that's left to the caller.
130 dmu_zfetch_fini(zfetch_t *zf)
134 ASSERT(!RW_LOCK_HELD(&zf->zf_rwlock));
136 rw_enter(&zf->zf_rwlock, RW_WRITER);
137 while ((zs = list_head(&zf->zf_stream)) != NULL)
138 dmu_zfetch_stream_remove(zf, zs);
139 rw_exit(&zf->zf_rwlock);
140 list_destroy(&zf->zf_stream);
141 rw_destroy(&zf->zf_rwlock);
147 * If there aren't too many streams already, create a new stream.
148 * The "blkid" argument is the next block that we expect this stream to access.
149 * While we're here, clean up old streams (which haven't been
150 * accessed for at least zfetch_min_sec_reap seconds).
153 dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid)
158 uint32_t max_streams;
160 ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
163 * Clean up old streams.
165 for (zs = list_head(&zf->zf_stream);
166 zs != NULL; zs = zs_next) {
167 zs_next = list_next(&zf->zf_stream, zs);
168 if (((gethrtime() - zs->zs_atime) / NANOSEC) >
170 dmu_zfetch_stream_remove(zf, zs);
176 * The maximum number of streams is normally zfetch_max_streams,
177 * but for small files we lower it such that it's at least possible
178 * for all the streams to be non-overlapping.
180 * If we are already at the maximum number of streams for this file,
181 * even after removing old streams, then don't create this stream.
183 max_streams = MAX(1, MIN(zfetch_max_streams,
184 zf->zf_dnode->dn_maxblkid * zf->zf_dnode->dn_datablksz /
185 zfetch_max_distance));
186 if (numstreams >= max_streams) {
187 ZFETCHSTAT_BUMP(zfetchstat_max_streams);
191 zs = kmem_zalloc(sizeof (*zs), KM_SLEEP);
192 zs->zs_blkid = blkid;
193 zs->zs_pf_blkid = blkid;
194 zs->zs_ipf_blkid = blkid;
195 zs->zs_atime = gethrtime();
196 mutex_init(&zs->zs_lock, NULL, MUTEX_DEFAULT, NULL);
198 list_insert_head(&zf->zf_stream, zs);
202 * This is the predictive prefetch entry point. It associates dnode access
203 * specified with blkid and nblks arguments with prefetch stream, predicts
204 * further accesses based on that stats and initiates speculative prefetch.
205 * fetch_data argument specifies whether actual data blocks should be fetched:
206 * FALSE -- prefetch only indirect blocks for predicted data blocks;
207 * TRUE -- prefetch predicted data blocks plus following indirect blocks.
210 dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data)
213 int64_t pf_start, ipf_start, ipf_istart, ipf_iend;
214 int64_t pf_ahead_blks, max_blks, iblk;
215 int epbs, max_dist_blks, pf_nblks, ipf_nblks, i;
216 uint64_t end_of_access_blkid;
217 end_of_access_blkid = blkid + nblks;
219 if (zfs_prefetch_disable)
223 * As a fast path for small (single-block) files, ignore access
224 * to the first block.
229 rw_enter(&zf->zf_rwlock, RW_READER);
231 for (zs = list_head(&zf->zf_stream); zs != NULL;
232 zs = list_next(&zf->zf_stream, zs)) {
233 if (blkid == zs->zs_blkid) {
234 mutex_enter(&zs->zs_lock);
236 * zs_blkid could have changed before we
237 * acquired zs_lock; re-check them here.
239 if (blkid != zs->zs_blkid) {
240 mutex_exit(&zs->zs_lock);
249 * This access is not part of any existing stream. Create
250 * a new stream for it.
252 ZFETCHSTAT_BUMP(zfetchstat_misses);
253 if (rw_tryupgrade(&zf->zf_rwlock))
254 dmu_zfetch_stream_create(zf, end_of_access_blkid);
255 rw_exit(&zf->zf_rwlock);
260 * This access was to a block that we issued a prefetch for on
261 * behalf of this stream. Issue further prefetches for this stream.
263 * Normally, we start prefetching where we stopped
264 * prefetching last (zs_pf_blkid). But when we get our first
265 * hit on this stream, zs_pf_blkid == zs_blkid, we don't
266 * want to prefetch the block we just accessed. In this case,
267 * start just after the block we just accessed.
269 pf_start = MAX(zs->zs_pf_blkid, end_of_access_blkid);
272 * Double our amount of prefetched data, but don't let the
273 * prefetch get further ahead than zfetch_max_distance.
277 zfetch_max_distance >> zf->zf_dnode->dn_datablkshift;
279 * Previously, we were (zs_pf_blkid - blkid) ahead. We
280 * want to now be double that, so read that amount again,
281 * plus the amount we are catching up by (i.e. the amount
284 pf_ahead_blks = zs->zs_pf_blkid - blkid + nblks;
285 max_blks = max_dist_blks - (pf_start - end_of_access_blkid);
286 pf_nblks = MIN(pf_ahead_blks, max_blks);
291 zs->zs_pf_blkid = pf_start + pf_nblks;
294 * Do the same for indirects, starting from where we stopped last,
295 * or where we will stop reading data blocks (and the indirects
296 * that point to them).
298 ipf_start = MAX(zs->zs_ipf_blkid, zs->zs_pf_blkid);
299 max_dist_blks = zfetch_max_idistance >> zf->zf_dnode->dn_datablkshift;
301 * We want to double our distance ahead of the data prefetch
302 * (or reader, if we are not prefetching data). Previously, we
303 * were (zs_ipf_blkid - blkid) ahead. To double that, we read
304 * that amount again, plus the amount we are catching up by
305 * (i.e. the amount read now + the amount of data prefetched now).
307 pf_ahead_blks = zs->zs_ipf_blkid - blkid + nblks + pf_nblks;
308 max_blks = max_dist_blks - (ipf_start - end_of_access_blkid);
309 ipf_nblks = MIN(pf_ahead_blks, max_blks);
310 zs->zs_ipf_blkid = ipf_start + ipf_nblks;
312 epbs = zf->zf_dnode->dn_indblkshift - SPA_BLKPTRSHIFT;
313 ipf_istart = P2ROUNDUP(ipf_start, 1 << epbs) >> epbs;
314 ipf_iend = P2ROUNDUP(zs->zs_ipf_blkid, 1 << epbs) >> epbs;
316 zs->zs_atime = gethrtime();
317 zs->zs_blkid = end_of_access_blkid;
318 mutex_exit(&zs->zs_lock);
319 rw_exit(&zf->zf_rwlock);
322 * dbuf_prefetch() is asynchronous (even when it needs to read
323 * indirect blocks), but we still prefer to drop our locks before
324 * calling it to reduce the time we hold them.
327 for (i = 0; i < pf_nblks; i++) {
328 dbuf_prefetch(zf->zf_dnode, 0, pf_start + i,
329 ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
331 for (iblk = ipf_istart; iblk < ipf_iend; iblk++) {
332 dbuf_prefetch(zf->zf_dnode, 1, iblk,
333 ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
335 ZFETCHSTAT_BUMP(zfetchstat_hits);
338 #if defined(_KERNEL) && defined(HAVE_SPL)
340 module_param(zfs_prefetch_disable, int, 0644);
341 MODULE_PARM_DESC(zfs_prefetch_disable, "Disable all ZFS prefetching");
343 module_param(zfetch_max_streams, uint, 0644);
344 MODULE_PARM_DESC(zfetch_max_streams, "Max number of streams per zfetch");
346 module_param(zfetch_min_sec_reap, uint, 0644);
347 MODULE_PARM_DESC(zfetch_min_sec_reap, "Min time before stream reclaim");
349 module_param(zfetch_max_distance, uint, 0644);
350 MODULE_PARM_DESC(zfetch_max_distance,
351 "Max bytes to prefetch per stream (default 8MB)");
353 module_param(zfetch_array_rd_sz, ulong, 0644);
354 MODULE_PARM_DESC(zfetch_array_rd_sz, "Number of bytes in a array_read");