xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu_zfetch.c (revision 6446bd46ed1b4e9f69da153665f82181ccaedad5)
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, 2017 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  */
44 boolean_t zfs_prefetch_disable = B_FALSE;
45 
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;
56 
57 typedef struct zfetch_stats {
58 	kstat_named_t zfetchstat_hits;
59 	kstat_named_t zfetchstat_misses;
60 	kstat_named_t zfetchstat_max_streams;
61 	kstat_named_t zfetchstat_max_completion_us;
62 	kstat_named_t zfetchstat_last_completion_us;
63 	kstat_named_t zfetchstat_io_issued;
64 } zfetch_stats_t;
65 
66 static zfetch_stats_t zfetch_stats = {
67 	{ "hits",			KSTAT_DATA_UINT64 },
68 	{ "misses",			KSTAT_DATA_UINT64 },
69 	{ "max_streams",		KSTAT_DATA_UINT64 },
70 	{ "max_completion_us",		KSTAT_DATA_UINT64 },
71 	{ "last_completion_us",		KSTAT_DATA_UINT64 },
72 	{ "io_issued",		KSTAT_DATA_UINT64 },
73 };
74 
75 #define	ZFETCHSTAT_BUMP(stat) \
76 	atomic_inc_64(&zfetch_stats.stat.value.ui64)
77 #define	ZFETCHSTAT_ADD(stat, val)				\
78 	atomic_add_64(&zfetch_stats.stat.value.ui64, val)
79 #define	ZFETCHSTAT_SET(stat, val)				\
80 	zfetch_stats.stat.value.ui64 = val
81 #define	ZFETCHSTAT_GET(stat)					\
82 	zfetch_stats.stat.value.ui64
83 
84 
85 kstat_t		*zfetch_ksp;
86 
87 void
88 zfetch_init(void)
89 {
90 	zfetch_ksp = kstat_create("zfs", 0, "zfetchstats", "misc",
91 	    KSTAT_TYPE_NAMED, sizeof (zfetch_stats) / sizeof (kstat_named_t),
92 	    KSTAT_FLAG_VIRTUAL);
93 
94 	if (zfetch_ksp != NULL) {
95 		zfetch_ksp->ks_data = &zfetch_stats;
96 		kstat_install(zfetch_ksp);
97 	}
98 }
99 
100 void
101 zfetch_fini(void)
102 {
103 	if (zfetch_ksp != NULL) {
104 		kstat_delete(zfetch_ksp);
105 		zfetch_ksp = NULL;
106 	}
107 }
108 
109 /*
110  * This takes a pointer to a zfetch structure and a dnode.  It performs the
111  * necessary setup for the zfetch structure, grokking data from the
112  * associated dnode.
113  */
114 void
115 dmu_zfetch_init(zfetch_t *zf, dnode_t *dno)
116 {
117 	if (zf == NULL)
118 		return;
119 	zf->zf_dnode = dno;
120 	zf->zf_numstreams = 0;
121 
122 	list_create(&zf->zf_stream, sizeof (zstream_t),
123 	    offsetof(zstream_t, zs_node));
124 
125 	rw_init(&zf->zf_rwlock, NULL, RW_DEFAULT, NULL);
126 }
127 
128 static void
129 dmu_zfetch_stream_fini(zstream_t *zs)
130 {
131 	mutex_destroy(&zs->zs_lock);
132 	kmem_free(zs, sizeof (*zs));
133 }
134 
135 static void
136 dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs)
137 {
138 	ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
139 	list_remove(&zf->zf_stream, zs);
140 	dmu_zfetch_stream_fini(zs);
141 	zf->zf_numstreams--;
142 }
143 
144 static void
145 dmu_zfetch_stream_orphan(zfetch_t *zf, zstream_t *zs)
146 {
147 	ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
148 	list_remove(&zf->zf_stream, zs);
149 	zs->zs_fetch = NULL;
150 	zf->zf_numstreams--;
151 }
152 
153 /*
154  * Clean-up state associated with a zfetch structure (e.g. destroy the
155  * streams).  This doesn't free the zfetch_t itself, that's left to the caller.
156  */
157 void
158 dmu_zfetch_fini(zfetch_t *zf)
159 {
160 	zstream_t *zs;
161 
162 	ASSERT(!RW_LOCK_HELD(&zf->zf_rwlock));
163 
164 	rw_enter(&zf->zf_rwlock, RW_WRITER);
165 	while ((zs = list_head(&zf->zf_stream)) != NULL) {
166 		if (zfs_refcount_count(&zs->zs_blocks) != 0)
167 			dmu_zfetch_stream_orphan(zf, zs);
168 		else
169 			dmu_zfetch_stream_remove(zf, zs);
170 	}
171 	rw_exit(&zf->zf_rwlock);
172 	list_destroy(&zf->zf_stream);
173 	rw_destroy(&zf->zf_rwlock);
174 
175 	zf->zf_dnode = NULL;
176 }
177 
178 /*
179  * If there aren't too many streams already, create a new stream.
180  * The "blkid" argument is the next block that we expect this stream to access.
181  * While we're here, clean up old streams (which haven't been
182  * accessed for at least zfetch_min_sec_reap seconds).
183  */
184 static void
185 dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid)
186 {
187 	zstream_t *zs_next;
188 	hrtime_t now = gethrtime();
189 
190 	ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
191 
192 	/*
193 	 * Clean up old streams.
194 	 */
195 	for (zstream_t *zs = list_head(&zf->zf_stream);
196 	    zs != NULL; zs = zs_next) {
197 		zs_next = list_next(&zf->zf_stream, zs);
198 		/*
199 		 * Skip gethrtime() call if there are still references
200 		 */
201 		if (zfs_refcount_count(&zs->zs_blocks) != 0)
202 			continue;
203 		if (((now - zs->zs_atime) / NANOSEC) >
204 		    zfetch_min_sec_reap)
205 			dmu_zfetch_stream_remove(zf, zs);
206 	}
207 
208 	/*
209 	 * The maximum number of streams is normally zfetch_max_streams,
210 	 * but for small files we lower it such that it's at least possible
211 	 * for all the streams to be non-overlapping.
212 	 *
213 	 * If we are already at the maximum number of streams for this file,
214 	 * even after removing old streams, then don't create this stream.
215 	 */
216 	uint32_t max_streams = MAX(1, MIN(zfetch_max_streams,
217 	    zf->zf_dnode->dn_maxblkid * zf->zf_dnode->dn_datablksz /
218 	    zfetch_max_distance));
219 	if (zf->zf_numstreams >= max_streams) {
220 		ZFETCHSTAT_BUMP(zfetchstat_max_streams);
221 		return;
222 	}
223 
224 	zstream_t *zs = kmem_zalloc(sizeof (*zs), KM_SLEEP);
225 	zs->zs_blkid = blkid;
226 	zs->zs_pf_blkid = blkid;
227 	zs->zs_ipf_blkid = blkid;
228 	zs->zs_atime = now;
229 	zs->zs_fetch = zf;
230 	zfs_refcount_create(&zs->zs_blocks);
231 	mutex_init(&zs->zs_lock, NULL, MUTEX_DEFAULT, NULL);
232 	zf->zf_numstreams++;
233 	list_insert_head(&zf->zf_stream, zs);
234 }
235 
236 static void
237 dmu_zfetch_stream_done(void *arg, boolean_t io_issued)
238 {
239 	zstream_t *zs = arg;
240 
241 	if (zs->zs_start_time && io_issued) {
242 		hrtime_t now = gethrtime();
243 		hrtime_t delta = NSEC2USEC(now - zs->zs_start_time);
244 
245 		zs->zs_start_time = 0;
246 		ZFETCHSTAT_SET(zfetchstat_last_completion_us, delta);
247 		if (delta > ZFETCHSTAT_GET(zfetchstat_max_completion_us))
248 			ZFETCHSTAT_SET(zfetchstat_max_completion_us, delta);
249 	}
250 
251 	if (zfs_refcount_remove(&zs->zs_blocks, NULL) != 0)
252 		return;
253 
254 	/*
255 	 * The parent fetch structure has gone away
256 	 */
257 	if (zs->zs_fetch == NULL)
258 		dmu_zfetch_stream_fini(zs);
259 }
260 
261 /*
262  * This is the predictive prefetch entry point.  It associates dnode access
263  * specified with blkid and nblks arguments with prefetch stream, predicts
264  * further accesses based on that stats and initiates speculative prefetch.
265  * fetch_data argument specifies whether actual data blocks should be fetched:
266  *   FALSE -- prefetch only indirect blocks for predicted data blocks;
267  *   TRUE -- prefetch predicted data blocks plus following indirect blocks.
268  */
269 void
270 dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data,
271     boolean_t have_lock)
272 {
273 	zstream_t *zs;
274 	int64_t pf_start, ipf_start, ipf_istart, ipf_iend;
275 	int64_t pf_ahead_blks, max_blks;
276 	int epbs, max_dist_blks, pf_nblks, ipf_nblks, issued;
277 	uint64_t end_of_access_blkid = blkid + nblks;
278 	spa_t *spa = zf->zf_dnode->dn_objset->os_spa;
279 
280 	if (zfs_prefetch_disable)
281 		return;
282 
283 	/*
284 	 * If we haven't yet loaded the indirect vdevs' mappings, we
285 	 * can only read from blocks that we carefully ensure are on
286 	 * concrete vdevs (or previously-loaded indirect vdevs).  So we
287 	 * can't allow the predictive prefetcher to attempt reads of other
288 	 * blocks (e.g. of the MOS's dnode obejct).
289 	 */
290 	if (!spa_indirect_vdevs_loaded(spa))
291 		return;
292 
293 	/*
294 	 * As a fast path for small (single-block) files, ignore access
295 	 * to the first block.
296 	 */
297 	if (!have_lock && blkid == 0)
298 		return;
299 
300 	if (!have_lock)
301 		rw_enter(&zf->zf_dnode->dn_struct_rwlock, RW_READER);
302 
303 
304 	/*
305 	 * A fast path for small files for which no prefetch will
306 	 * happen.
307 	 */
308 	if (zf->zf_dnode->dn_maxblkid < 2) {
309 		if (!have_lock)
310 			rw_exit(&zf->zf_dnode->dn_struct_rwlock);
311 		return;
312 	}
313 	rw_enter(&zf->zf_rwlock, RW_READER);
314 
315 	/*
316 	 * Find matching prefetch stream.  Depending on whether the accesses
317 	 * are block-aligned, first block of the new access may either follow
318 	 * the last block of the previous access, or be equal to it.
319 	 */
320 	for (zs = list_head(&zf->zf_stream); zs != NULL;
321 	    zs = list_next(&zf->zf_stream, zs)) {
322 		if (blkid == zs->zs_blkid || blkid + 1 == zs->zs_blkid) {
323 			mutex_enter(&zs->zs_lock);
324 			/*
325 			 * zs_blkid could have changed before we
326 			 * acquired zs_lock; re-check them here.
327 			 */
328 			if (blkid == zs->zs_blkid) {
329 				break;
330 			} else if (blkid + 1 == zs->zs_blkid) {
331 				blkid++;
332 				nblks--;
333 				if (nblks == 0) {
334 					/* Already prefetched this before. */
335 					mutex_exit(&zs->zs_lock);
336 					rw_exit(&zf->zf_rwlock);
337 					if (!have_lock) {
338 						rw_exit(&zf->zf_dnode->
339 						    dn_struct_rwlock);
340 					}
341 					return;
342 				}
343 				break;
344 			}
345 			mutex_exit(&zs->zs_lock);
346 		}
347 	}
348 
349 	if (zs == NULL) {
350 		/*
351 		 * This access is not part of any existing stream.  Create
352 		 * a new stream for it.
353 		 */
354 		ZFETCHSTAT_BUMP(zfetchstat_misses);
355 		if (rw_tryupgrade(&zf->zf_rwlock))
356 			dmu_zfetch_stream_create(zf, end_of_access_blkid);
357 		rw_exit(&zf->zf_rwlock);
358 		if (!have_lock)
359 			rw_exit(&zf->zf_dnode->dn_struct_rwlock);
360 		return;
361 	}
362 
363 	/*
364 	 * This access was to a block that we issued a prefetch for on
365 	 * behalf of this stream. Issue further prefetches for this stream.
366 	 *
367 	 * Normally, we start prefetching where we stopped
368 	 * prefetching last (zs_pf_blkid).  But when we get our first
369 	 * hit on this stream, zs_pf_blkid == zs_blkid, we don't
370 	 * want to prefetch the block we just accessed.  In this case,
371 	 * start just after the block we just accessed.
372 	 */
373 	pf_start = MAX(zs->zs_pf_blkid, end_of_access_blkid);
374 
375 	/*
376 	 * Double our amount of prefetched data, but don't let the
377 	 * prefetch get further ahead than zfetch_max_distance.
378 	 */
379 	if (fetch_data) {
380 		max_dist_blks =
381 		    zfetch_max_distance >> zf->zf_dnode->dn_datablkshift;
382 		/*
383 		 * Previously, we were (zs_pf_blkid - blkid) ahead.  We
384 		 * want to now be double that, so read that amount again,
385 		 * plus the amount we are catching up by (i.e. the amount
386 		 * read just now).
387 		 */
388 		pf_ahead_blks = zs->zs_pf_blkid - blkid + nblks;
389 		max_blks = max_dist_blks - (pf_start - end_of_access_blkid);
390 		pf_nblks = MIN(pf_ahead_blks, max_blks);
391 	} else {
392 		pf_nblks = 0;
393 	}
394 
395 	zs->zs_pf_blkid = pf_start + pf_nblks;
396 
397 	/*
398 	 * Do the same for indirects, starting from where we stopped last,
399 	 * or where we will stop reading data blocks (and the indirects
400 	 * that point to them).
401 	 */
402 	ipf_start = MAX(zs->zs_ipf_blkid, zs->zs_pf_blkid);
403 	max_dist_blks = zfetch_max_idistance >> zf->zf_dnode->dn_datablkshift;
404 	/*
405 	 * We want to double our distance ahead of the data prefetch
406 	 * (or reader, if we are not prefetching data).  Previously, we
407 	 * were (zs_ipf_blkid - blkid) ahead.  To double that, we read
408 	 * that amount again, plus the amount we are catching up by
409 	 * (i.e. the amount read now + the amount of data prefetched now).
410 	 */
411 	pf_ahead_blks = zs->zs_ipf_blkid - blkid + nblks + pf_nblks;
412 	max_blks = max_dist_blks - (ipf_start - end_of_access_blkid);
413 	ipf_nblks = MIN(pf_ahead_blks, max_blks);
414 	zs->zs_ipf_blkid = ipf_start + ipf_nblks;
415 
416 	epbs = zf->zf_dnode->dn_indblkshift - SPA_BLKPTRSHIFT;
417 	ipf_istart = P2ROUNDUP(ipf_start, 1 << epbs) >> epbs;
418 	ipf_iend = P2ROUNDUP(zs->zs_ipf_blkid, 1 << epbs) >> epbs;
419 
420 	zs->zs_atime = gethrtime();
421 	/* no prior reads in progress */
422 	if (zfs_refcount_count(&zs->zs_blocks) == 0)
423 		zs->zs_start_time = zs->zs_atime;
424 	zs->zs_blkid = end_of_access_blkid;
425 	zfs_refcount_add_many(&zs->zs_blocks, pf_nblks + ipf_iend - ipf_istart,
426 	    NULL);
427 	mutex_exit(&zs->zs_lock);
428 	rw_exit(&zf->zf_rwlock);
429 	issued = 0;
430 
431 	/*
432 	 * dbuf_prefetch() is asynchronous (even when it needs to read
433 	 * indirect blocks), but we still prefer to drop our locks before
434 	 * calling it to reduce the time we hold them.
435 	 */
436 
437 	for (int i = 0; i < pf_nblks; i++) {
438 		issued += dbuf_prefetch_impl(zf->zf_dnode, 0, pf_start + i,
439 		    ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH,
440 		    dmu_zfetch_stream_done, zs);
441 	}
442 	for (int64_t iblk = ipf_istart; iblk < ipf_iend; iblk++) {
443 		issued += dbuf_prefetch_impl(zf->zf_dnode, 1, iblk,
444 		    ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH,
445 		    dmu_zfetch_stream_done, zs);
446 	}
447 	if (!have_lock)
448 		rw_exit(&zf->zf_dnode->dn_struct_rwlock);
449 	ZFETCHSTAT_BUMP(zfetchstat_hits);
450 
451 	if (issued)
452 		ZFETCHSTAT_ADD(zfetchstat_io_issued, issued);
453 }
454