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
57typedef struct zfetch_stats {
58	kstat_named_t zfetchstat_hits;
59	kstat_named_t zfetchstat_misses;
60	kstat_named_t zfetchstat_max_streams;
61} zfetch_stats_t;
62
63static zfetch_stats_t zfetch_stats = {
64	{ "hits",			KSTAT_DATA_UINT64 },
65	{ "misses",			KSTAT_DATA_UINT64 },
66	{ "max_streams",		KSTAT_DATA_UINT64 },
67};
68
69#define	ZFETCHSTAT_BUMP(stat) \
70	atomic_inc_64(&zfetch_stats.stat.value.ui64);
71
72kstat_t		*zfetch_ksp;
73
74void
75zfetch_init(void)
76{
77	zfetch_ksp = kstat_create("zfs", 0, "zfetchstats", "misc",
78	    KSTAT_TYPE_NAMED, sizeof (zfetch_stats) / sizeof (kstat_named_t),
79	    KSTAT_FLAG_VIRTUAL);
80
81	if (zfetch_ksp != NULL) {
82		zfetch_ksp->ks_data = &zfetch_stats;
83		kstat_install(zfetch_ksp);
84	}
85}
86
87void
88zfetch_fini(void)
89{
90	if (zfetch_ksp != NULL) {
91		kstat_delete(zfetch_ksp);
92		zfetch_ksp = NULL;
93	}
94}
95
96/*
97 * This takes a pointer to a zfetch structure and a dnode.  It performs the
98 * necessary setup for the zfetch structure, grokking data from the
99 * associated dnode.
100 */
101void
102dmu_zfetch_init(zfetch_t *zf, dnode_t *dno)
103{
104	if (zf == NULL)
105		return;
106
107	zf->zf_dnode = dno;
108
109	list_create(&zf->zf_stream, sizeof (zstream_t),
110	    offsetof(zstream_t, zs_node));
111
112	rw_init(&zf->zf_rwlock, NULL, RW_DEFAULT, NULL);
113}
114
115static void
116dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs)
117{
118	ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
119	list_remove(&zf->zf_stream, zs);
120	mutex_destroy(&zs->zs_lock);
121	kmem_free(zs, sizeof (*zs));
122}
123
124/*
125 * Clean-up state associated with a zfetch structure (e.g. destroy the
126 * streams).  This doesn't free the zfetch_t itself, that's left to the caller.
127 */
128void
129dmu_zfetch_fini(zfetch_t *zf)
130{
131	zstream_t *zs;
132
133	ASSERT(!RW_LOCK_HELD(&zf->zf_rwlock));
134
135	rw_enter(&zf->zf_rwlock, RW_WRITER);
136	while ((zs = list_head(&zf->zf_stream)) != NULL)
137		dmu_zfetch_stream_remove(zf, zs);
138	rw_exit(&zf->zf_rwlock);
139	list_destroy(&zf->zf_stream);
140	rw_destroy(&zf->zf_rwlock);
141
142	zf->zf_dnode = NULL;
143}
144
145/*
146 * If there aren't too many streams already, create a new stream.
147 * The "blkid" argument is the next block that we expect this stream to access.
148 * While we're here, clean up old streams (which haven't been
149 * accessed for at least zfetch_min_sec_reap seconds).
150 */
151static void
152dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid)
153{
154	zstream_t *zs_next;
155	int numstreams = 0;
156
157	ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
158
159	/*
160	 * Clean up old streams.
161	 */
162	for (zstream_t *zs = list_head(&zf->zf_stream);
163	    zs != NULL; zs = zs_next) {
164		zs_next = list_next(&zf->zf_stream, zs);
165		if (((gethrtime() - zs->zs_atime) / NANOSEC) >
166		    zfetch_min_sec_reap)
167			dmu_zfetch_stream_remove(zf, zs);
168		else
169			numstreams++;
170	}
171
172	/*
173	 * The maximum number of streams is normally zfetch_max_streams,
174	 * but for small files we lower it such that it's at least possible
175	 * for all the streams to be non-overlapping.
176	 *
177	 * If we are already at the maximum number of streams for this file,
178	 * even after removing old streams, then don't create this stream.
179	 */
180	uint32_t max_streams = MAX(1, MIN(zfetch_max_streams,
181	    zf->zf_dnode->dn_maxblkid * zf->zf_dnode->dn_datablksz /
182	    zfetch_max_distance));
183	if (numstreams >= max_streams) {
184		ZFETCHSTAT_BUMP(zfetchstat_max_streams);
185		return;
186	}
187
188	zstream_t *zs = kmem_zalloc(sizeof (*zs), KM_SLEEP);
189	zs->zs_blkid = blkid;
190	zs->zs_pf_blkid = blkid;
191	zs->zs_ipf_blkid = blkid;
192	zs->zs_atime = gethrtime();
193	mutex_init(&zs->zs_lock, NULL, MUTEX_DEFAULT, NULL);
194
195	list_insert_head(&zf->zf_stream, zs);
196}
197
198/*
199 * This is the predictive prefetch entry point.  It associates dnode access
200 * specified with blkid and nblks arguments with prefetch stream, predicts
201 * further accesses based on that stats and initiates speculative prefetch.
202 * fetch_data argument specifies whether actual data blocks should be fetched:
203 *   FALSE -- prefetch only indirect blocks for predicted data blocks;
204 *   TRUE -- prefetch predicted data blocks plus following indirect blocks.
205 */
206void
207dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data)
208{
209	zstream_t *zs;
210	int64_t pf_start, ipf_start, ipf_istart, ipf_iend;
211	int64_t pf_ahead_blks, max_blks;
212	int epbs, max_dist_blks, pf_nblks, ipf_nblks;
213	uint64_t end_of_access_blkid = blkid + nblks;
214	spa_t *spa = zf->zf_dnode->dn_objset->os_spa;
215
216	if (zfs_prefetch_disable)
217		return;
218
219	/*
220	 * If we haven't yet loaded the indirect vdevs' mappings, we
221	 * can only read from blocks that we carefully ensure are on
222	 * concrete vdevs (or previously-loaded indirect vdevs).  So we
223	 * can't allow the predictive prefetcher to attempt reads of other
224	 * blocks (e.g. of the MOS's dnode obejct).
225	 */
226	if (!spa_indirect_vdevs_loaded(spa))
227		return;
228
229	/*
230	 * As a fast path for small (single-block) files, ignore access
231	 * to the first block.
232	 */
233	if (blkid == 0)
234		return;
235
236	rw_enter(&zf->zf_rwlock, RW_READER);
237
238	/*
239	 * Find matching prefetch stream.  Depending on whether the accesses
240	 * are block-aligned, first block of the new access may either follow
241	 * the last block of the previous access, or be equal to it.
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 || blkid + 1 == 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				break;
253			} else if (blkid + 1 == zs->zs_blkid) {
254				blkid++;
255				nblks--;
256				if (nblks == 0) {
257					/* Already prefetched this before. */
258					mutex_exit(&zs->zs_lock);
259					rw_exit(&zf->zf_rwlock);
260					return;
261				}
262				break;
263			}
264			mutex_exit(&zs->zs_lock);
265		}
266	}
267
268	if (zs == NULL) {
269		/*
270		 * This access is not part of any existing stream.  Create
271		 * a new stream for it.
272		 */
273		ZFETCHSTAT_BUMP(zfetchstat_misses);
274		if (rw_tryupgrade(&zf->zf_rwlock))
275			dmu_zfetch_stream_create(zf, end_of_access_blkid);
276		rw_exit(&zf->zf_rwlock);
277		return;
278	}
279
280	/*
281	 * This access was to a block that we issued a prefetch for on
282	 * behalf of this stream. Issue further prefetches for this stream.
283	 *
284	 * Normally, we start prefetching where we stopped
285	 * prefetching last (zs_pf_blkid).  But when we get our first
286	 * hit on this stream, zs_pf_blkid == zs_blkid, we don't
287	 * want to prefetch the block we just accessed.  In this case,
288	 * start just after the block we just accessed.
289	 */
290	pf_start = MAX(zs->zs_pf_blkid, end_of_access_blkid);
291
292	/*
293	 * Double our amount of prefetched data, but don't let the
294	 * prefetch get further ahead than zfetch_max_distance.
295	 */
296	if (fetch_data) {
297		max_dist_blks =
298		    zfetch_max_distance >> zf->zf_dnode->dn_datablkshift;
299		/*
300		 * Previously, we were (zs_pf_blkid - blkid) ahead.  We
301		 * want to now be double that, so read that amount again,
302		 * plus the amount we are catching up by (i.e. the amount
303		 * read just now).
304		 */
305		pf_ahead_blks = zs->zs_pf_blkid - blkid + nblks;
306		max_blks = max_dist_blks - (pf_start - end_of_access_blkid);
307		pf_nblks = MIN(pf_ahead_blks, max_blks);
308	} else {
309		pf_nblks = 0;
310	}
311
312	zs->zs_pf_blkid = pf_start + pf_nblks;
313
314	/*
315	 * Do the same for indirects, starting from where we stopped last,
316	 * or where we will stop reading data blocks (and the indirects
317	 * that point to them).
318	 */
319	ipf_start = MAX(zs->zs_ipf_blkid, zs->zs_pf_blkid);
320	max_dist_blks = zfetch_max_idistance >> zf->zf_dnode->dn_datablkshift;
321	/*
322	 * We want to double our distance ahead of the data prefetch
323	 * (or reader, if we are not prefetching data).  Previously, we
324	 * were (zs_ipf_blkid - blkid) ahead.  To double that, we read
325	 * that amount again, plus the amount we are catching up by
326	 * (i.e. the amount read now + the amount of data prefetched now).
327	 */
328	pf_ahead_blks = zs->zs_ipf_blkid - blkid + nblks + pf_nblks;
329	max_blks = max_dist_blks - (ipf_start - end_of_access_blkid);
330	ipf_nblks = MIN(pf_ahead_blks, max_blks);
331	zs->zs_ipf_blkid = ipf_start + ipf_nblks;
332
333	epbs = zf->zf_dnode->dn_indblkshift - SPA_BLKPTRSHIFT;
334	ipf_istart = P2ROUNDUP(ipf_start, 1 << epbs) >> epbs;
335	ipf_iend = P2ROUNDUP(zs->zs_ipf_blkid, 1 << epbs) >> epbs;
336
337	zs->zs_atime = gethrtime();
338	zs->zs_blkid = end_of_access_blkid;
339	mutex_exit(&zs->zs_lock);
340	rw_exit(&zf->zf_rwlock);
341
342	/*
343	 * dbuf_prefetch() is asynchronous (even when it needs to read
344	 * indirect blocks), but we still prefer to drop our locks before
345	 * calling it to reduce the time we hold them.
346	 */
347
348	for (int i = 0; i < pf_nblks; i++) {
349		dbuf_prefetch(zf->zf_dnode, 0, pf_start + i,
350		    ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
351	}
352	for (int64_t iblk = ipf_istart; iblk < ipf_iend; iblk++) {
353		dbuf_prefetch(zf->zf_dnode, 1, iblk,
354		    ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
355	}
356	ZFETCHSTAT_BUMP(zfetchstat_hits);
357}
358