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 */
44boolean_t zfs_prefetch_disable = B_FALSE;
45
46/* max # of streams per zfetch */
47uint32_t	zfetch_max_streams = 8;
48/* min time before stream reclaim */
49uint32_t	zfetch_min_sec_reap = 2;
50/* max bytes to prefetch per stream (default 8MB) */
51uint32_t	zfetch_max_distance = 8 * 1024 * 1024;
52/* max bytes to prefetch indirects for per stream (default 64MB) */
53uint32_t	zfetch_max_idistance = 64 * 1024 * 1024;
54/* max number of bytes in an array_read in which we allow prefetching (1MB) */
55uint64_t	zfetch_array_rd_sz = 1024 * 1024;
56
57SYSCTL_DECL(_vfs_zfs);
58SYSCTL_INT(_vfs_zfs, OID_AUTO, prefetch_disable, CTLFLAG_RW,
59    &zfs_prefetch_disable, 0, "Disable prefetch");
60SYSCTL_NODE(_vfs_zfs, OID_AUTO, zfetch, CTLFLAG_RW, 0, "ZFS ZFETCH");
61TUNABLE_INT("vfs.zfs.zfetch.max_streams", &zfetch_max_streams);
62SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, max_streams, CTLFLAG_RW,
63    &zfetch_max_streams, 0, "Max # of streams per zfetch");
64TUNABLE_INT("vfs.zfs.zfetch.min_sec_reap", &zfetch_min_sec_reap);
65SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, min_sec_reap, CTLFLAG_RWTUN,
66    &zfetch_min_sec_reap, 0, "Min time before stream reclaim");
67TUNABLE_INT("vfs.zfs.zfetch.max_distance", &zfetch_max_distance);
68SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, max_distance, CTLFLAG_RWTUN,
69    &zfetch_max_distance, 0, "Max bytes to prefetch per stream");
70TUNABLE_QUAD("vfs.zfs.zfetch.array_rd_sz", &zfetch_array_rd_sz);
71SYSCTL_UQUAD(_vfs_zfs_zfetch, OID_AUTO, array_rd_sz, CTLFLAG_RWTUN,
72    &zfetch_array_rd_sz, 0,
73    "Number of bytes in a array_read at which we stop prefetching");
74
75typedef struct zfetch_stats {
76	kstat_named_t zfetchstat_hits;
77	kstat_named_t zfetchstat_misses;
78	kstat_named_t zfetchstat_max_streams;
79} zfetch_stats_t;
80
81static zfetch_stats_t zfetch_stats = {
82	{ "hits",			KSTAT_DATA_UINT64 },
83	{ "misses",			KSTAT_DATA_UINT64 },
84	{ "max_streams",		KSTAT_DATA_UINT64 },
85};
86
87#define	ZFETCHSTAT_BUMP(stat) \
88	atomic_inc_64(&zfetch_stats.stat.value.ui64);
89
90kstat_t		*zfetch_ksp;
91
92void
93zfetch_init(void)
94{
95	zfetch_ksp = kstat_create("zfs", 0, "zfetchstats", "misc",
96	    KSTAT_TYPE_NAMED, sizeof (zfetch_stats) / sizeof (kstat_named_t),
97	    KSTAT_FLAG_VIRTUAL);
98
99	if (zfetch_ksp != NULL) {
100		zfetch_ksp->ks_data = &zfetch_stats;
101		kstat_install(zfetch_ksp);
102	}
103}
104
105void
106zfetch_fini(void)
107{
108	if (zfetch_ksp != NULL) {
109		kstat_delete(zfetch_ksp);
110		zfetch_ksp = NULL;
111	}
112}
113
114/*
115 * This takes a pointer to a zfetch structure and a dnode.  It performs the
116 * necessary setup for the zfetch structure, grokking data from the
117 * associated dnode.
118 */
119void
120dmu_zfetch_init(zfetch_t *zf, dnode_t *dno)
121{
122	if (zf == NULL)
123		return;
124
125	zf->zf_dnode = dno;
126
127	list_create(&zf->zf_stream, sizeof (zstream_t),
128	    offsetof(zstream_t, zs_node));
129
130	rw_init(&zf->zf_rwlock, NULL, RW_DEFAULT, NULL);
131}
132
133static void
134dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs)
135{
136	ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
137	list_remove(&zf->zf_stream, zs);
138	mutex_destroy(&zs->zs_lock);
139	kmem_free(zs, sizeof (*zs));
140}
141
142/*
143 * Clean-up state associated with a zfetch structure (e.g. destroy the
144 * streams).  This doesn't free the zfetch_t itself, that's left to the caller.
145 */
146void
147dmu_zfetch_fini(zfetch_t *zf)
148{
149	zstream_t *zs;
150
151	ASSERT(!RW_LOCK_HELD(&zf->zf_rwlock));
152
153	rw_enter(&zf->zf_rwlock, RW_WRITER);
154	while ((zs = list_head(&zf->zf_stream)) != NULL)
155		dmu_zfetch_stream_remove(zf, zs);
156	rw_exit(&zf->zf_rwlock);
157	list_destroy(&zf->zf_stream);
158	rw_destroy(&zf->zf_rwlock);
159
160	zf->zf_dnode = NULL;
161}
162
163/*
164 * If there aren't too many streams already, create a new stream.
165 * The "blkid" argument is the next block that we expect this stream to access.
166 * While we're here, clean up old streams (which haven't been
167 * accessed for at least zfetch_min_sec_reap seconds).
168 */
169static void
170dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid)
171{
172	zstream_t *zs_next;
173	int numstreams = 0;
174
175	ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
176
177	/*
178	 * Clean up old streams.
179	 */
180	for (zstream_t *zs = list_head(&zf->zf_stream);
181	    zs != NULL; zs = zs_next) {
182		zs_next = list_next(&zf->zf_stream, zs);
183		if (((gethrtime() - zs->zs_atime) / NANOSEC) >
184		    zfetch_min_sec_reap)
185			dmu_zfetch_stream_remove(zf, zs);
186		else
187			numstreams++;
188	}
189
190	/*
191	 * The maximum number of streams is normally zfetch_max_streams,
192	 * but for small files we lower it such that it's at least possible
193	 * for all the streams to be non-overlapping.
194	 *
195	 * If we are already at the maximum number of streams for this file,
196	 * even after removing old streams, then don't create this stream.
197	 */
198	uint32_t max_streams = MAX(1, MIN(zfetch_max_streams,
199	    zf->zf_dnode->dn_maxblkid * zf->zf_dnode->dn_datablksz /
200	    zfetch_max_distance));
201	if (numstreams >= max_streams) {
202		ZFETCHSTAT_BUMP(zfetchstat_max_streams);
203		return;
204	}
205
206	zstream_t *zs = kmem_zalloc(sizeof (*zs), KM_SLEEP);
207	zs->zs_blkid = blkid;
208	zs->zs_pf_blkid = blkid;
209	zs->zs_ipf_blkid = blkid;
210	zs->zs_atime = gethrtime();
211	mutex_init(&zs->zs_lock, NULL, MUTEX_DEFAULT, NULL);
212
213	list_insert_head(&zf->zf_stream, zs);
214}
215
216/*
217 * This is the predictive prefetch entry point.  It associates dnode access
218 * specified with blkid and nblks arguments with prefetch stream, predicts
219 * further accesses based on that stats and initiates speculative prefetch.
220 * fetch_data argument specifies whether actual data blocks should be fetched:
221 *   FALSE -- prefetch only indirect blocks for predicted data blocks;
222 *   TRUE -- prefetch predicted data blocks plus following indirect blocks.
223 */
224void
225dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data)
226{
227	zstream_t *zs;
228	int64_t pf_start, ipf_start, ipf_istart, ipf_iend;
229	int64_t pf_ahead_blks, max_blks;
230	int epbs, max_dist_blks, pf_nblks, ipf_nblks;
231	uint64_t end_of_access_blkid = blkid + nblks;
232
233	if (zfs_prefetch_disable)
234		return;
235
236	/*
237	 * As a fast path for small (single-block) files, ignore access
238	 * to the first block.
239	 */
240	if (blkid == 0)
241		return;
242
243	rw_enter(&zf->zf_rwlock, RW_READER);
244
245	for (zs = list_head(&zf->zf_stream); zs != NULL;
246	    zs = list_next(&zf->zf_stream, zs)) {
247		if (blkid == zs->zs_blkid) {
248			mutex_enter(&zs->zs_lock);
249			/*
250			 * zs_blkid could have changed before we
251			 * acquired zs_lock; re-check them here.
252			 */
253			if (blkid != zs->zs_blkid) {
254				mutex_exit(&zs->zs_lock);
255				continue;
256			}
257			break;
258		}
259	}
260
261	if (zs == NULL) {
262		/*
263		 * This access is not part of any existing stream.  Create
264		 * a new stream for it.
265		 */
266		ZFETCHSTAT_BUMP(zfetchstat_misses);
267		if (rw_tryupgrade(&zf->zf_rwlock))
268			dmu_zfetch_stream_create(zf, end_of_access_blkid);
269		rw_exit(&zf->zf_rwlock);
270		return;
271	}
272
273	/*
274	 * This access was to a block that we issued a prefetch for on
275	 * behalf of this stream. Issue further prefetches for this stream.
276	 *
277	 * Normally, we start prefetching where we stopped
278	 * prefetching last (zs_pf_blkid).  But when we get our first
279	 * hit on this stream, zs_pf_blkid == zs_blkid, we don't
280	 * want to prefetch the block we just accessed.  In this case,
281	 * start just after the block we just accessed.
282	 */
283	pf_start = MAX(zs->zs_pf_blkid, end_of_access_blkid);
284
285	/*
286	 * Double our amount of prefetched data, but don't let the
287	 * prefetch get further ahead than zfetch_max_distance.
288	 */
289	if (fetch_data) {
290		max_dist_blks =
291		    zfetch_max_distance >> zf->zf_dnode->dn_datablkshift;
292		/*
293		 * Previously, we were (zs_pf_blkid - blkid) ahead.  We
294		 * want to now be double that, so read that amount again,
295		 * plus the amount we are catching up by (i.e. the amount
296		 * read just now).
297		 */
298		pf_ahead_blks = zs->zs_pf_blkid - blkid + nblks;
299		max_blks = max_dist_blks - (pf_start - end_of_access_blkid);
300		pf_nblks = MIN(pf_ahead_blks, max_blks);
301	} else {
302		pf_nblks = 0;
303	}
304
305	zs->zs_pf_blkid = pf_start + pf_nblks;
306
307	/*
308	 * Do the same for indirects, starting from where we stopped last,
309	 * or where we will stop reading data blocks (and the indirects
310	 * that point to them).
311	 */
312	ipf_start = MAX(zs->zs_ipf_blkid, zs->zs_pf_blkid);
313	max_dist_blks = zfetch_max_idistance >> zf->zf_dnode->dn_datablkshift;
314	/*
315	 * We want to double our distance ahead of the data prefetch
316	 * (or reader, if we are not prefetching data).  Previously, we
317	 * were (zs_ipf_blkid - blkid) ahead.  To double that, we read
318	 * that amount again, plus the amount we are catching up by
319	 * (i.e. the amount read now + the amount of data prefetched now).
320	 */
321	pf_ahead_blks = zs->zs_ipf_blkid - blkid + nblks + pf_nblks;
322	max_blks = max_dist_blks - (ipf_start - end_of_access_blkid);
323	ipf_nblks = MIN(pf_ahead_blks, max_blks);
324	zs->zs_ipf_blkid = ipf_start + ipf_nblks;
325
326	epbs = zf->zf_dnode->dn_indblkshift - SPA_BLKPTRSHIFT;
327	ipf_istart = P2ROUNDUP(ipf_start, 1 << epbs) >> epbs;
328	ipf_iend = P2ROUNDUP(zs->zs_ipf_blkid, 1 << epbs) >> epbs;
329
330	zs->zs_atime = gethrtime();
331	zs->zs_blkid = end_of_access_blkid;
332	mutex_exit(&zs->zs_lock);
333	rw_exit(&zf->zf_rwlock);
334
335	/*
336	 * dbuf_prefetch() is asynchronous (even when it needs to read
337	 * indirect blocks), but we still prefer to drop our locks before
338	 * calling it to reduce the time we hold them.
339	 */
340
341	for (int i = 0; i < pf_nblks; i++) {
342		dbuf_prefetch(zf->zf_dnode, 0, pf_start + i,
343		    ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
344	}
345	for (int64_t iblk = ipf_istart; iblk < ipf_iend; iblk++) {
346		dbuf_prefetch(zf->zf_dnode, 1, iblk,
347		    ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
348	}
349	ZFETCHSTAT_BUMP(zfetchstat_hits);
350}
351