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.
26 #include <sys/zfs_context.h>
27 #include <sys/dnode.h>
28 #include <sys/dmu_objset.h>
29 #include <sys/dmu_zfetch.h>
32 #include <sys/kstat.h>
35 * I'm against tune-ables, but these should probably exist as tweakable globals
36 * until we can get this working the way we want it to.
39 int zfs_prefetch_disable = 0;
41 /* max # of streams per zfetch */
42 unsigned int zfetch_max_streams = 8;
43 /* min time before stream reclaim */
44 unsigned int zfetch_min_sec_reap = 2;
45 /* max number of blocks to fetch at a time */
46 unsigned int zfetch_block_cap = 256;
47 /* number of bytes in a array_read at which we stop prefetching (1Mb) */
48 unsigned long zfetch_array_rd_sz = 1024 * 1024;
50 /* forward decls for static routines */
51 static int dmu_zfetch_colinear(zfetch_t *, zstream_t *);
52 static void dmu_zfetch_dofetch(zfetch_t *, zstream_t *);
53 static uint64_t dmu_zfetch_fetch(dnode_t *, uint64_t, uint64_t);
54 static uint64_t dmu_zfetch_fetchsz(dnode_t *, uint64_t, uint64_t);
55 static int dmu_zfetch_find(zfetch_t *, zstream_t *, int);
56 static int dmu_zfetch_stream_insert(zfetch_t *, zstream_t *);
57 static zstream_t *dmu_zfetch_stream_reclaim(zfetch_t *);
58 static void dmu_zfetch_stream_remove(zfetch_t *, zstream_t *);
59 static int dmu_zfetch_streams_equal(zstream_t *, zstream_t *);
61 typedef struct zfetch_stats {
62 kstat_named_t zfetchstat_hits;
63 kstat_named_t zfetchstat_misses;
64 kstat_named_t zfetchstat_colinear_hits;
65 kstat_named_t zfetchstat_colinear_misses;
66 kstat_named_t zfetchstat_stride_hits;
67 kstat_named_t zfetchstat_stride_misses;
68 kstat_named_t zfetchstat_reclaim_successes;
69 kstat_named_t zfetchstat_reclaim_failures;
70 kstat_named_t zfetchstat_stream_resets;
71 kstat_named_t zfetchstat_stream_noresets;
72 kstat_named_t zfetchstat_bogus_streams;
75 static zfetch_stats_t zfetch_stats = {
76 { "hits", KSTAT_DATA_UINT64 },
77 { "misses", KSTAT_DATA_UINT64 },
78 { "colinear_hits", KSTAT_DATA_UINT64 },
79 { "colinear_misses", KSTAT_DATA_UINT64 },
80 { "stride_hits", KSTAT_DATA_UINT64 },
81 { "stride_misses", KSTAT_DATA_UINT64 },
82 { "reclaim_successes", KSTAT_DATA_UINT64 },
83 { "reclaim_failures", KSTAT_DATA_UINT64 },
84 { "streams_resets", KSTAT_DATA_UINT64 },
85 { "streams_noresets", KSTAT_DATA_UINT64 },
86 { "bogus_streams", KSTAT_DATA_UINT64 },
89 #define ZFETCHSTAT_INCR(stat, val) \
90 atomic_add_64(&zfetch_stats.stat.value.ui64, (val));
92 #define ZFETCHSTAT_BUMP(stat) ZFETCHSTAT_INCR(stat, 1);
97 * Given a zfetch structure and a zstream structure, determine whether the
98 * blocks to be read are part of a co-linear pair of existing prefetch
99 * streams. If a set is found, coalesce the streams, removing one, and
100 * configure the prefetch so it looks for a strided access pattern.
102 * In other words: if we find two sequential access streams that are
103 * the same length and distance N appart, and this read is N from the
104 * last stream, then we are probably in a strided access pattern. So
105 * combine the two sequential streams into a single strided stream.
107 * If no co-linear streams are found, return NULL.
110 dmu_zfetch_colinear(zfetch_t *zf, zstream_t *zh)
115 if (! rw_tryenter(&zf->zf_rwlock, RW_WRITER))
119 rw_exit(&zf->zf_rwlock);
123 for (z_walk = list_head(&zf->zf_stream); z_walk;
124 z_walk = list_next(&zf->zf_stream, z_walk)) {
125 for (z_comp = list_next(&zf->zf_stream, z_walk); z_comp;
126 z_comp = list_next(&zf->zf_stream, z_comp)) {
129 if (z_walk->zst_len != z_walk->zst_stride ||
130 z_comp->zst_len != z_comp->zst_stride) {
134 diff = z_comp->zst_offset - z_walk->zst_offset;
135 if (z_comp->zst_offset + diff == zh->zst_offset) {
136 z_walk->zst_offset = zh->zst_offset;
137 z_walk->zst_direction = diff < 0 ? -1 : 1;
139 diff * z_walk->zst_direction;
140 z_walk->zst_ph_offset =
141 zh->zst_offset + z_walk->zst_stride;
142 dmu_zfetch_stream_remove(zf, z_comp);
143 mutex_destroy(&z_comp->zst_lock);
144 kmem_free(z_comp, sizeof (zstream_t));
146 dmu_zfetch_dofetch(zf, z_walk);
148 rw_exit(&zf->zf_rwlock);
152 diff = z_walk->zst_offset - z_comp->zst_offset;
153 if (z_walk->zst_offset + diff == zh->zst_offset) {
154 z_walk->zst_offset = zh->zst_offset;
155 z_walk->zst_direction = diff < 0 ? -1 : 1;
157 diff * z_walk->zst_direction;
158 z_walk->zst_ph_offset =
159 zh->zst_offset + z_walk->zst_stride;
160 dmu_zfetch_stream_remove(zf, z_comp);
161 mutex_destroy(&z_comp->zst_lock);
162 kmem_free(z_comp, sizeof (zstream_t));
164 dmu_zfetch_dofetch(zf, z_walk);
166 rw_exit(&zf->zf_rwlock);
172 rw_exit(&zf->zf_rwlock);
177 * Given a zstream_t, determine the bounds of the prefetch. Then call the
178 * routine that actually prefetches the individual blocks.
181 dmu_zfetch_dofetch(zfetch_t *zf, zstream_t *zs)
183 uint64_t prefetch_tail;
184 uint64_t prefetch_limit;
185 uint64_t prefetch_ofst;
186 uint64_t prefetch_len;
187 uint64_t blocks_fetched;
189 zs->zst_stride = MAX((int64_t)zs->zst_stride, zs->zst_len);
190 zs->zst_cap = MIN(zfetch_block_cap, 2 * zs->zst_cap);
192 prefetch_tail = MAX((int64_t)zs->zst_ph_offset,
193 (int64_t)(zs->zst_offset + zs->zst_stride));
195 * XXX: use a faster division method?
197 prefetch_limit = zs->zst_offset + zs->zst_len +
198 (zs->zst_cap * zs->zst_stride) / zs->zst_len;
200 while (prefetch_tail < prefetch_limit) {
201 prefetch_ofst = zs->zst_offset + zs->zst_direction *
202 (prefetch_tail - zs->zst_offset);
204 prefetch_len = zs->zst_len;
207 * Don't prefetch beyond the end of the file, if working
210 if ((zs->zst_direction == ZFETCH_BACKWARD) &&
211 (prefetch_ofst > prefetch_tail)) {
212 prefetch_len += prefetch_ofst;
216 /* don't prefetch more than we're supposed to */
217 if (prefetch_len > zs->zst_len)
220 blocks_fetched = dmu_zfetch_fetch(zf->zf_dnode,
221 prefetch_ofst, zs->zst_len);
223 prefetch_tail += zs->zst_stride;
224 /* stop if we've run out of stuff to prefetch */
225 if (blocks_fetched < zs->zst_len)
228 zs->zst_ph_offset = prefetch_tail;
229 zs->zst_last = ddi_get_lbolt();
236 zfetch_ksp = kstat_create("zfs", 0, "zfetchstats", "misc",
237 KSTAT_TYPE_NAMED, sizeof (zfetch_stats) / sizeof (kstat_named_t),
240 if (zfetch_ksp != NULL) {
241 zfetch_ksp->ks_data = &zfetch_stats;
242 kstat_install(zfetch_ksp);
249 if (zfetch_ksp != NULL) {
250 kstat_delete(zfetch_ksp);
256 * This takes a pointer to a zfetch structure and a dnode. It performs the
257 * necessary setup for the zfetch structure, grokking data from the
261 dmu_zfetch_init(zfetch_t *zf, dnode_t *dno)
268 zf->zf_stream_cnt = 0;
269 zf->zf_alloc_fail = 0;
271 list_create(&zf->zf_stream, sizeof (zstream_t),
272 offsetof(zstream_t, zst_node));
274 rw_init(&zf->zf_rwlock, NULL, RW_DEFAULT, NULL);
278 * This function computes the actual size, in blocks, that can be prefetched,
282 dmu_zfetch_fetch(dnode_t *dn, uint64_t blkid, uint64_t nblks)
287 fetchsz = dmu_zfetch_fetchsz(dn, blkid, nblks);
289 for (i = 0; i < fetchsz; i++) {
290 dbuf_prefetch(dn, blkid + i);
297 * this function returns the number of blocks that would be prefetched, based
298 * upon the supplied dnode, blockid, and nblks. This is used so that we can
299 * update streams in place, and then prefetch with their old value after the
300 * fact. This way, we can delay the prefetch, but subsequent accesses to the
301 * stream won't result in the same data being prefetched multiple times.
304 dmu_zfetch_fetchsz(dnode_t *dn, uint64_t blkid, uint64_t nblks)
308 if (blkid > dn->dn_maxblkid) {
312 /* compute fetch size */
313 if (blkid + nblks + 1 > dn->dn_maxblkid) {
314 fetchsz = (dn->dn_maxblkid - blkid) + 1;
315 ASSERT(blkid + fetchsz - 1 <= dn->dn_maxblkid);
325 * given a zfetch and a zstream structure, see if there is an associated zstream
326 * for this block read. If so, it starts a prefetch for the stream it
327 * located and returns true, otherwise it returns false
330 dmu_zfetch_find(zfetch_t *zf, zstream_t *zh, int prefetched)
334 int reset = !prefetched;
341 * XXX: This locking strategy is a bit coarse; however, it's impact has
342 * yet to be tested. If this turns out to be an issue, it can be
343 * modified in a number of different ways.
346 rw_enter(&zf->zf_rwlock, RW_READER);
349 for (zs = list_head(&zf->zf_stream); zs;
350 zs = list_next(&zf->zf_stream, zs)) {
353 * XXX - should this be an assert?
355 if (zs->zst_len == 0) {
357 ZFETCHSTAT_BUMP(zfetchstat_bogus_streams);
362 * We hit this case when we are in a strided prefetch stream:
363 * we will read "len" blocks before "striding".
365 if (zh->zst_offset >= zs->zst_offset &&
366 zh->zst_offset < zs->zst_offset + zs->zst_len) {
368 /* already fetched */
369 ZFETCHSTAT_BUMP(zfetchstat_stride_hits);
373 ZFETCHSTAT_BUMP(zfetchstat_stride_misses);
378 * This is the forward sequential read case: we increment
379 * len by one each time we hit here, so we will enter this
380 * case on every read.
382 if (zh->zst_offset == zs->zst_offset + zs->zst_len) {
384 reset = !prefetched && zs->zst_len > 1;
386 mutex_enter(&zs->zst_lock);
388 if (zh->zst_offset != zs->zst_offset + zs->zst_len) {
389 mutex_exit(&zs->zst_lock);
392 zs->zst_len += zh->zst_len;
393 diff = zs->zst_len - zfetch_block_cap;
395 zs->zst_offset += diff;
396 zs->zst_len = zs->zst_len > diff ?
397 zs->zst_len - diff : 0;
399 zs->zst_direction = ZFETCH_FORWARD;
404 * Same as above, but reading backwards through the file.
406 } else if (zh->zst_offset == zs->zst_offset - zh->zst_len) {
407 /* backwards sequential access */
409 reset = !prefetched && zs->zst_len > 1;
411 mutex_enter(&zs->zst_lock);
413 if (zh->zst_offset != zs->zst_offset - zh->zst_len) {
414 mutex_exit(&zs->zst_lock);
418 zs->zst_offset = zs->zst_offset > zh->zst_len ?
419 zs->zst_offset - zh->zst_len : 0;
420 zs->zst_ph_offset = zs->zst_ph_offset > zh->zst_len ?
421 zs->zst_ph_offset - zh->zst_len : 0;
422 zs->zst_len += zh->zst_len;
424 diff = zs->zst_len - zfetch_block_cap;
426 zs->zst_ph_offset = zs->zst_ph_offset > diff ?
427 zs->zst_ph_offset - diff : 0;
428 zs->zst_len = zs->zst_len > diff ?
429 zs->zst_len - diff : zs->zst_len;
431 zs->zst_direction = ZFETCH_BACKWARD;
435 } else if ((zh->zst_offset - zs->zst_offset - zs->zst_stride <
436 zs->zst_len) && (zs->zst_len != zs->zst_stride)) {
437 /* strided forward access */
439 mutex_enter(&zs->zst_lock);
441 if ((zh->zst_offset - zs->zst_offset - zs->zst_stride >=
442 zs->zst_len) || (zs->zst_len == zs->zst_stride)) {
443 mutex_exit(&zs->zst_lock);
447 zs->zst_offset += zs->zst_stride;
448 zs->zst_direction = ZFETCH_FORWARD;
452 } else if ((zh->zst_offset - zs->zst_offset + zs->zst_stride <
453 zs->zst_len) && (zs->zst_len != zs->zst_stride)) {
454 /* strided reverse access */
456 mutex_enter(&zs->zst_lock);
458 if ((zh->zst_offset - zs->zst_offset + zs->zst_stride >=
459 zs->zst_len) || (zs->zst_len == zs->zst_stride)) {
460 mutex_exit(&zs->zst_lock);
464 zs->zst_offset = zs->zst_offset > zs->zst_stride ?
465 zs->zst_offset - zs->zst_stride : 0;
466 zs->zst_ph_offset = (zs->zst_ph_offset >
467 (2 * zs->zst_stride)) ?
468 (zs->zst_ph_offset - (2 * zs->zst_stride)) : 0;
469 zs->zst_direction = ZFETCH_BACKWARD;
477 zstream_t *remove = zs;
479 ZFETCHSTAT_BUMP(zfetchstat_stream_resets);
481 mutex_exit(&zs->zst_lock);
482 rw_exit(&zf->zf_rwlock);
483 rw_enter(&zf->zf_rwlock, RW_WRITER);
485 * Relocate the stream, in case someone removes
486 * it while we were acquiring the WRITER lock.
488 for (zs = list_head(&zf->zf_stream); zs;
489 zs = list_next(&zf->zf_stream, zs)) {
491 dmu_zfetch_stream_remove(zf, zs);
492 mutex_destroy(&zs->zst_lock);
493 kmem_free(zs, sizeof (zstream_t));
498 ZFETCHSTAT_BUMP(zfetchstat_stream_noresets);
500 dmu_zfetch_dofetch(zf, zs);
501 mutex_exit(&zs->zst_lock);
505 rw_exit(&zf->zf_rwlock);
510 * Clean-up state associated with a zfetch structure. This frees allocated
511 * structure members, empties the zf_stream tree, and generally makes things
512 * nice. This doesn't free the zfetch_t itself, that's left to the caller.
515 dmu_zfetch_rele(zfetch_t *zf)
520 ASSERT(!RW_LOCK_HELD(&zf->zf_rwlock));
522 for (zs = list_head(&zf->zf_stream); zs; zs = zs_next) {
523 zs_next = list_next(&zf->zf_stream, zs);
525 list_remove(&zf->zf_stream, zs);
526 mutex_destroy(&zs->zst_lock);
527 kmem_free(zs, sizeof (zstream_t));
529 list_destroy(&zf->zf_stream);
530 rw_destroy(&zf->zf_rwlock);
536 * Given a zfetch and zstream structure, insert the zstream structure into the
537 * AVL tree contained within the zfetch structure. Peform the appropriate
538 * book-keeping. It is possible that another thread has inserted a stream which
539 * matches one that we are about to insert, so we must be sure to check for this
540 * case. If one is found, return failure, and let the caller cleanup the
544 dmu_zfetch_stream_insert(zfetch_t *zf, zstream_t *zs)
549 ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
551 for (zs_walk = list_head(&zf->zf_stream); zs_walk; zs_walk = zs_next) {
552 zs_next = list_next(&zf->zf_stream, zs_walk);
554 if (dmu_zfetch_streams_equal(zs_walk, zs)) {
559 list_insert_head(&zf->zf_stream, zs);
566 * Walk the list of zstreams in the given zfetch, find an old one (by time), and
567 * reclaim it for use by the caller.
570 dmu_zfetch_stream_reclaim(zfetch_t *zf)
574 if (! rw_tryenter(&zf->zf_rwlock, RW_WRITER))
577 for (zs = list_head(&zf->zf_stream); zs;
578 zs = list_next(&zf->zf_stream, zs)) {
580 if (((ddi_get_lbolt() - zs->zst_last)/hz) > zfetch_min_sec_reap)
585 dmu_zfetch_stream_remove(zf, zs);
586 mutex_destroy(&zs->zst_lock);
587 bzero(zs, sizeof (zstream_t));
591 rw_exit(&zf->zf_rwlock);
597 * Given a zfetch and zstream structure, remove the zstream structure from its
598 * container in the zfetch structure. Perform the appropriate book-keeping.
601 dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs)
603 ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
605 list_remove(&zf->zf_stream, zs);
610 dmu_zfetch_streams_equal(zstream_t *zs1, zstream_t *zs2)
612 if (zs1->zst_offset != zs2->zst_offset)
615 if (zs1->zst_len != zs2->zst_len)
618 if (zs1->zst_stride != zs2->zst_stride)
621 if (zs1->zst_ph_offset != zs2->zst_ph_offset)
624 if (zs1->zst_cap != zs2->zst_cap)
627 if (zs1->zst_direction != zs2->zst_direction)
634 * This is the prefetch entry point. It calls all of the other dmu_zfetch
635 * routines to create, delete, find, or operate upon prefetch streams.
638 dmu_zfetch(zfetch_t *zf, uint64_t offset, uint64_t size, int prefetched)
641 zstream_t *newstream;
644 unsigned int blkshft;
647 if (zfs_prefetch_disable)
650 /* files that aren't ln2 blocksz are only one block -- nothing to do */
651 if (!zf->zf_dnode->dn_datablkshift)
654 /* convert offset and size, into blockid and nblocks */
655 blkshft = zf->zf_dnode->dn_datablkshift;
656 blksz = (1 << blkshft);
658 bzero(&zst, sizeof (zstream_t));
659 zst.zst_offset = offset >> blkshft;
660 zst.zst_len = (P2ROUNDUP(offset + size, blksz) -
661 P2ALIGN(offset, blksz)) >> blkshft;
663 fetched = dmu_zfetch_find(zf, &zst, prefetched);
665 ZFETCHSTAT_BUMP(zfetchstat_hits);
667 ZFETCHSTAT_BUMP(zfetchstat_misses);
668 if ((fetched = dmu_zfetch_colinear(zf, &zst))) {
669 ZFETCHSTAT_BUMP(zfetchstat_colinear_hits);
671 ZFETCHSTAT_BUMP(zfetchstat_colinear_misses);
676 newstream = dmu_zfetch_stream_reclaim(zf);
679 * we still couldn't find a stream, drop the lock, and allocate
680 * one if possible. Otherwise, give up and go home.
683 ZFETCHSTAT_BUMP(zfetchstat_reclaim_successes);
686 uint32_t max_streams;
687 uint32_t cur_streams;
689 ZFETCHSTAT_BUMP(zfetchstat_reclaim_failures);
690 cur_streams = zf->zf_stream_cnt;
691 maxblocks = zf->zf_dnode->dn_maxblkid;
693 max_streams = MIN(zfetch_max_streams,
694 (maxblocks / zfetch_block_cap));
695 if (max_streams == 0) {
699 if (cur_streams >= max_streams) {
702 newstream = kmem_zalloc(sizeof (zstream_t), KM_PUSHPAGE);
705 newstream->zst_offset = zst.zst_offset;
706 newstream->zst_len = zst.zst_len;
707 newstream->zst_stride = zst.zst_len;
708 newstream->zst_ph_offset = zst.zst_len + zst.zst_offset;
709 newstream->zst_cap = zst.zst_len;
710 newstream->zst_direction = ZFETCH_FORWARD;
711 newstream->zst_last = ddi_get_lbolt();
713 mutex_init(&newstream->zst_lock, NULL, MUTEX_DEFAULT, NULL);
715 rw_enter(&zf->zf_rwlock, RW_WRITER);
716 inserted = dmu_zfetch_stream_insert(zf, newstream);
717 rw_exit(&zf->zf_rwlock);
720 mutex_destroy(&newstream->zst_lock);
721 kmem_free(newstream, sizeof (zstream_t));
726 #if defined(_KERNEL) && defined(HAVE_SPL)
727 module_param(zfs_prefetch_disable, int, 0644);
728 MODULE_PARM_DESC(zfs_prefetch_disable, "Disable all ZFS prefetching");
730 module_param(zfetch_max_streams, uint, 0644);
731 MODULE_PARM_DESC(zfetch_max_streams, "Max number of streams per zfetch");
733 module_param(zfetch_min_sec_reap, uint, 0644);
734 MODULE_PARM_DESC(zfetch_min_sec_reap, "Min time before stream reclaim");
736 module_param(zfetch_block_cap, uint, 0644);
737 MODULE_PARM_DESC(zfetch_block_cap, "Max number of blocks to fetch at a time");
739 module_param(zfetch_array_rd_sz, ulong, 0644);
740 MODULE_PARM_DESC(zfetch_array_rd_sz, "Number of bytes in a array_read");