1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3 * CDDL HEADER START
4 *
5 * The contents of this file are subject to the terms of the
6 * Common Development and Distribution License (the "License").
7 * You may not use this file except in compliance with the License.
8 *
9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10 * or https://opensource.org/licenses/CDDL-1.0.
11 * See the License for the specific language governing permissions
12 * and limitations under the License.
13 *
14 * When distributing Covered Code, include this CDDL HEADER in each
15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16 * If applicable, add the following below this CDDL HEADER, with the
17 * fields enclosed by brackets "[]" replaced with your own identifying
18 * information: Portions Copyright [yyyy] [name of copyright owner]
19 *
20 * CDDL HEADER END
21 */
22 /*
23 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
24 * Use is subject to license terms.
25 */
26
27 /*
28 * Copyright (c) 2013, 2017 by Delphix. All rights reserved.
29 */
30
31 #include <sys/zfs_context.h>
32 #include <sys/arc_impl.h>
33 #include <sys/dnode.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dmu_zfetch.h>
36 #include <sys/dmu.h>
37 #include <sys/dbuf.h>
38 #include <sys/kstat.h>
39 #include <sys/wmsum.h>
40
41 /*
42 * This tunable disables predictive prefetch. Note that it leaves "prescient"
43 * prefetch (e.g. prefetch for zfs send) intact. Unlike predictive prefetch,
44 * prescient prefetch never issues i/os that end up not being needed,
45 * so it can't hurt performance.
46 */
47
48 static int zfs_prefetch_disable = B_FALSE;
49
50 /* max # of streams per zfetch */
51 static unsigned int zfetch_max_streams = 8;
52 /* min time before stream reclaim */
53 static unsigned int zfetch_min_sec_reap = 1;
54 /* max time before stream delete */
55 static unsigned int zfetch_max_sec_reap = 2;
56 #ifdef _ILP32
57 /* min bytes to prefetch per stream (default 2MB) */
58 static unsigned int zfetch_min_distance = 2 * 1024 * 1024;
59 /* max bytes to prefetch per stream (default 8MB) */
60 unsigned int zfetch_max_distance = 8 * 1024 * 1024;
61 #else
62 /* min bytes to prefetch per stream (default 4MB) */
63 static unsigned int zfetch_min_distance = 4 * 1024 * 1024;
64 /* max bytes to prefetch per stream (default 64MB) */
65 unsigned int zfetch_max_distance = 64 * 1024 * 1024;
66 #endif
67 /* max bytes to prefetch indirects for per stream (default 128MB) */
68 unsigned int zfetch_max_idistance = 128 * 1024 * 1024;
69 /* max request reorder distance within a stream (default 16MB) */
70 unsigned int zfetch_max_reorder = 16 * 1024 * 1024;
71 /* Max log2 fraction of holes in a stream */
72 unsigned int zfetch_hole_shift = 2;
73
74 typedef struct zfetch_stats {
75 kstat_named_t zfetchstat_hits;
76 kstat_named_t zfetchstat_future;
77 kstat_named_t zfetchstat_stride;
78 kstat_named_t zfetchstat_past;
79 kstat_named_t zfetchstat_misses;
80 kstat_named_t zfetchstat_max_streams;
81 kstat_named_t zfetchstat_io_issued;
82 kstat_named_t zfetchstat_io_active;
83 } zfetch_stats_t;
84
85 static zfetch_stats_t zfetch_stats = {
86 { "hits", KSTAT_DATA_UINT64 },
87 { "future", KSTAT_DATA_UINT64 },
88 { "stride", KSTAT_DATA_UINT64 },
89 { "past", KSTAT_DATA_UINT64 },
90 { "misses", KSTAT_DATA_UINT64 },
91 { "max_streams", KSTAT_DATA_UINT64 },
92 { "io_issued", KSTAT_DATA_UINT64 },
93 { "io_active", KSTAT_DATA_UINT64 },
94 };
95
96 struct {
97 wmsum_t zfetchstat_hits;
98 wmsum_t zfetchstat_future;
99 wmsum_t zfetchstat_stride;
100 wmsum_t zfetchstat_past;
101 wmsum_t zfetchstat_misses;
102 wmsum_t zfetchstat_max_streams;
103 wmsum_t zfetchstat_io_issued;
104 aggsum_t zfetchstat_io_active;
105 } zfetch_sums;
106
107 #define ZFETCHSTAT_BUMP(stat) \
108 wmsum_add(&zfetch_sums.stat, 1)
109 #define ZFETCHSTAT_ADD(stat, val) \
110 wmsum_add(&zfetch_sums.stat, val)
111
112
113 static kstat_t *zfetch_ksp;
114
115 static int
zfetch_kstats_update(kstat_t * ksp,int rw)116 zfetch_kstats_update(kstat_t *ksp, int rw)
117 {
118 zfetch_stats_t *zs = ksp->ks_data;
119
120 if (rw == KSTAT_WRITE)
121 return (EACCES);
122 zs->zfetchstat_hits.value.ui64 =
123 wmsum_value(&zfetch_sums.zfetchstat_hits);
124 zs->zfetchstat_future.value.ui64 =
125 wmsum_value(&zfetch_sums.zfetchstat_future);
126 zs->zfetchstat_stride.value.ui64 =
127 wmsum_value(&zfetch_sums.zfetchstat_stride);
128 zs->zfetchstat_past.value.ui64 =
129 wmsum_value(&zfetch_sums.zfetchstat_past);
130 zs->zfetchstat_misses.value.ui64 =
131 wmsum_value(&zfetch_sums.zfetchstat_misses);
132 zs->zfetchstat_max_streams.value.ui64 =
133 wmsum_value(&zfetch_sums.zfetchstat_max_streams);
134 zs->zfetchstat_io_issued.value.ui64 =
135 wmsum_value(&zfetch_sums.zfetchstat_io_issued);
136 zs->zfetchstat_io_active.value.ui64 =
137 aggsum_value(&zfetch_sums.zfetchstat_io_active);
138 return (0);
139 }
140
141 void
zfetch_init(void)142 zfetch_init(void)
143 {
144 wmsum_init(&zfetch_sums.zfetchstat_hits, 0);
145 wmsum_init(&zfetch_sums.zfetchstat_future, 0);
146 wmsum_init(&zfetch_sums.zfetchstat_stride, 0);
147 wmsum_init(&zfetch_sums.zfetchstat_past, 0);
148 wmsum_init(&zfetch_sums.zfetchstat_misses, 0);
149 wmsum_init(&zfetch_sums.zfetchstat_max_streams, 0);
150 wmsum_init(&zfetch_sums.zfetchstat_io_issued, 0);
151 aggsum_init(&zfetch_sums.zfetchstat_io_active, 0);
152
153 zfetch_ksp = kstat_create("zfs", 0, "zfetchstats", "misc",
154 KSTAT_TYPE_NAMED, sizeof (zfetch_stats) / sizeof (kstat_named_t),
155 KSTAT_FLAG_VIRTUAL);
156
157 if (zfetch_ksp != NULL) {
158 zfetch_ksp->ks_data = &zfetch_stats;
159 zfetch_ksp->ks_update = zfetch_kstats_update;
160 kstat_install(zfetch_ksp);
161 }
162 }
163
164 void
zfetch_fini(void)165 zfetch_fini(void)
166 {
167 if (zfetch_ksp != NULL) {
168 kstat_delete(zfetch_ksp);
169 zfetch_ksp = NULL;
170 }
171
172 wmsum_fini(&zfetch_sums.zfetchstat_hits);
173 wmsum_fini(&zfetch_sums.zfetchstat_future);
174 wmsum_fini(&zfetch_sums.zfetchstat_stride);
175 wmsum_fini(&zfetch_sums.zfetchstat_past);
176 wmsum_fini(&zfetch_sums.zfetchstat_misses);
177 wmsum_fini(&zfetch_sums.zfetchstat_max_streams);
178 wmsum_fini(&zfetch_sums.zfetchstat_io_issued);
179 ASSERT0(aggsum_value(&zfetch_sums.zfetchstat_io_active));
180 aggsum_fini(&zfetch_sums.zfetchstat_io_active);
181 }
182
183 /*
184 * This takes a pointer to a zfetch structure and a dnode. It performs the
185 * necessary setup for the zfetch structure, grokking data from the
186 * associated dnode.
187 */
188 void
dmu_zfetch_init(zfetch_t * zf,dnode_t * dno)189 dmu_zfetch_init(zfetch_t *zf, dnode_t *dno)
190 {
191 if (zf == NULL)
192 return;
193 zf->zf_dnode = dno;
194 zf->zf_numstreams = 0;
195
196 list_create(&zf->zf_stream, sizeof (zstream_t),
197 offsetof(zstream_t, zs_node));
198
199 mutex_init(&zf->zf_lock, NULL, MUTEX_DEFAULT, NULL);
200 }
201
202 static void
dmu_zfetch_stream_fini(zstream_t * zs)203 dmu_zfetch_stream_fini(zstream_t *zs)
204 {
205 ASSERT(!list_link_active(&zs->zs_node));
206 zfs_refcount_destroy(&zs->zs_callers);
207 zfs_refcount_destroy(&zs->zs_refs);
208 kmem_free(zs, sizeof (*zs));
209 }
210
211 static void
dmu_zfetch_stream_remove(zfetch_t * zf,zstream_t * zs)212 dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs)
213 {
214 ASSERT(MUTEX_HELD(&zf->zf_lock));
215 list_remove(&zf->zf_stream, zs);
216 zf->zf_numstreams--;
217 membar_producer();
218 if (zfs_refcount_remove(&zs->zs_refs, NULL) == 0)
219 dmu_zfetch_stream_fini(zs);
220 }
221
222 /*
223 * Clean-up state associated with a zfetch structure (e.g. destroy the
224 * streams). This doesn't free the zfetch_t itself, that's left to the caller.
225 */
226 void
dmu_zfetch_fini(zfetch_t * zf)227 dmu_zfetch_fini(zfetch_t *zf)
228 {
229 zstream_t *zs;
230
231 mutex_enter(&zf->zf_lock);
232 while ((zs = list_head(&zf->zf_stream)) != NULL)
233 dmu_zfetch_stream_remove(zf, zs);
234 mutex_exit(&zf->zf_lock);
235 list_destroy(&zf->zf_stream);
236 mutex_destroy(&zf->zf_lock);
237
238 zf->zf_dnode = NULL;
239 }
240
241 /*
242 * If there aren't too many active streams already, create one more.
243 * In process delete/reuse all streams without hits for zfetch_max_sec_reap.
244 * If needed, reuse oldest stream without hits for zfetch_min_sec_reap or ever.
245 * The "blkid" argument is the next block that we expect this stream to access.
246 */
247 static void
dmu_zfetch_stream_create(zfetch_t * zf,uint64_t blkid)248 dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid)
249 {
250 zstream_t *zs, *zs_next, *zs_old = NULL;
251 uint_t now = gethrestime_sec(), t;
252
253 ASSERT(MUTEX_HELD(&zf->zf_lock));
254
255 /*
256 * Delete too old streams, reusing the first found one.
257 */
258 t = now - zfetch_max_sec_reap;
259 for (zs = list_head(&zf->zf_stream); zs != NULL; zs = zs_next) {
260 zs_next = list_next(&zf->zf_stream, zs);
261 /*
262 * Skip if still active. 1 -- zf_stream reference.
263 */
264 if ((int)(zs->zs_atime - t) >= 0)
265 continue;
266 if (zfs_refcount_count(&zs->zs_refs) != 1)
267 continue;
268 if (zs_old)
269 dmu_zfetch_stream_remove(zf, zs);
270 else
271 zs_old = zs;
272 }
273 if (zs_old) {
274 zs = zs_old;
275 list_remove(&zf->zf_stream, zs);
276 goto reuse;
277 }
278
279 /*
280 * The maximum number of streams is normally zfetch_max_streams,
281 * but for small files we lower it such that it's at least possible
282 * for all the streams to be non-overlapping.
283 */
284 uint32_t max_streams = MAX(1, MIN(zfetch_max_streams,
285 (zf->zf_dnode->dn_maxblkid << zf->zf_dnode->dn_datablkshift) /
286 zfetch_max_distance));
287 if (zf->zf_numstreams >= max_streams) {
288 t = now - zfetch_min_sec_reap;
289 for (zs = list_head(&zf->zf_stream); zs != NULL;
290 zs = list_next(&zf->zf_stream, zs)) {
291 if ((int)(zs->zs_atime - t) >= 0)
292 continue;
293 if (zfs_refcount_count(&zs->zs_refs) != 1)
294 continue;
295 if (zs_old == NULL ||
296 (int)(zs_old->zs_atime - zs->zs_atime) >= 0)
297 zs_old = zs;
298 }
299 if (zs_old) {
300 zs = zs_old;
301 list_remove(&zf->zf_stream, zs);
302 goto reuse;
303 }
304 ZFETCHSTAT_BUMP(zfetchstat_max_streams);
305 return;
306 }
307
308 zs = kmem_zalloc(sizeof (*zs), KM_SLEEP);
309 zfs_refcount_create(&zs->zs_callers);
310 zfs_refcount_create(&zs->zs_refs);
311 /* One reference for zf_stream. */
312 zfs_refcount_add(&zs->zs_refs, NULL);
313 zf->zf_numstreams++;
314
315 reuse:
316 list_insert_head(&zf->zf_stream, zs);
317 zs->zs_blkid = blkid;
318 /* Allow immediate stream reuse until first hit. */
319 zs->zs_atime = now - zfetch_min_sec_reap;
320 memset(zs->zs_ranges, 0, sizeof (zs->zs_ranges));
321 zs->zs_pf_dist = 0;
322 zs->zs_ipf_dist = 0;
323 zs->zs_pf_start = blkid;
324 zs->zs_pf_end = blkid;
325 zs->zs_ipf_start = blkid;
326 zs->zs_ipf_end = blkid;
327 zs->zs_missed = B_FALSE;
328 zs->zs_more = B_FALSE;
329 }
330
331 static void
dmu_zfetch_done(void * arg,uint64_t level,uint64_t blkid,boolean_t io_issued)332 dmu_zfetch_done(void *arg, uint64_t level, uint64_t blkid, boolean_t io_issued)
333 {
334 zstream_t *zs = arg;
335
336 if (io_issued && level == 0 && blkid < zs->zs_blkid)
337 zs->zs_more = B_TRUE;
338 if (zfs_refcount_remove(&zs->zs_refs, NULL) == 0)
339 dmu_zfetch_stream_fini(zs);
340 aggsum_add(&zfetch_sums.zfetchstat_io_active, -1);
341 }
342
343 /*
344 * Process stream hit access for nblks blocks starting at zs_blkid. Return
345 * number of blocks to proceed for after aggregation with future ranges.
346 */
347 static uint64_t
dmu_zfetch_hit(zstream_t * zs,uint64_t nblks)348 dmu_zfetch_hit(zstream_t *zs, uint64_t nblks)
349 {
350 uint_t i, j;
351
352 /* Optimize sequential accesses (no future ranges). */
353 if (zs->zs_ranges[0].start == 0)
354 goto done;
355
356 /* Look for intersections with further ranges. */
357 for (i = 0; i < ZFETCH_RANGES; i++) {
358 zsrange_t *r = &zs->zs_ranges[i];
359 if (r->start == 0 || r->start > nblks)
360 break;
361 if (r->end >= nblks) {
362 nblks = r->end;
363 i++;
364 break;
365 }
366 }
367
368 /* Delete all found intersecting ranges, updates remaining. */
369 for (j = 0; i < ZFETCH_RANGES; i++, j++) {
370 if (zs->zs_ranges[i].start == 0)
371 break;
372 ASSERT3U(zs->zs_ranges[i].start, >, nblks);
373 ASSERT3U(zs->zs_ranges[i].end, >, nblks);
374 zs->zs_ranges[j].start = zs->zs_ranges[i].start - nblks;
375 zs->zs_ranges[j].end = zs->zs_ranges[i].end - nblks;
376 }
377 if (j < ZFETCH_RANGES) {
378 zs->zs_ranges[j].start = 0;
379 zs->zs_ranges[j].end = 0;
380 }
381
382 done:
383 zs->zs_blkid += nblks;
384 return (nblks);
385 }
386
387 /*
388 * Process future stream access for nblks blocks starting at blkid. Return
389 * number of blocks to proceed for if future ranges reach fill threshold.
390 */
391 static uint64_t
dmu_zfetch_future(zstream_t * zs,uint64_t blkid,uint64_t nblks)392 dmu_zfetch_future(zstream_t *zs, uint64_t blkid, uint64_t nblks)
393 {
394 ASSERT3U(blkid, >, zs->zs_blkid);
395 blkid -= zs->zs_blkid;
396 ASSERT3U(blkid + nblks, <=, UINT16_MAX);
397
398 /* Search for first and last intersection or insert point. */
399 uint_t f = ZFETCH_RANGES, l = 0, i;
400 for (i = 0; i < ZFETCH_RANGES; i++) {
401 zsrange_t *r = &zs->zs_ranges[i];
402 if (r->start == 0 || r->start > blkid + nblks)
403 break;
404 if (r->end < blkid)
405 continue;
406 if (f > i)
407 f = i;
408 if (l < i)
409 l = i;
410 }
411 if (f <= l) {
412 /* Got some intersecting range, expand it if needed. */
413 if (zs->zs_ranges[f].start > blkid)
414 zs->zs_ranges[f].start = blkid;
415 zs->zs_ranges[f].end = MAX(zs->zs_ranges[l].end, blkid + nblks);
416 if (f < l) {
417 /* Got more than one intersection, remove others. */
418 for (f++, l++; l < ZFETCH_RANGES; f++, l++) {
419 zs->zs_ranges[f].start = zs->zs_ranges[l].start;
420 zs->zs_ranges[f].end = zs->zs_ranges[l].end;
421 }
422 zs->zs_ranges[f].start = 0;
423 zs->zs_ranges[f].end = 0;
424 }
425 } else if (i < ZFETCH_RANGES) {
426 /* Got no intersecting ranges, insert new one. */
427 for (l = ZFETCH_RANGES - 1; l > i; l--) {
428 zs->zs_ranges[l].start = zs->zs_ranges[l - 1].start;
429 zs->zs_ranges[l].end = zs->zs_ranges[l - 1].end;
430 }
431 zs->zs_ranges[i].start = blkid;
432 zs->zs_ranges[i].end = blkid + nblks;
433 } else {
434 /* No space left to insert. Drop the range. */
435 return (0);
436 }
437
438 /* Check if with the new access addition we reached fill threshold. */
439 if (zfetch_hole_shift >= 16)
440 return (0);
441 uint_t hole = 0;
442 for (i = f = l = 0; i < ZFETCH_RANGES; i++) {
443 zsrange_t *r = &zs->zs_ranges[i];
444 if (r->start == 0)
445 break;
446 hole += r->start - f;
447 f = r->end;
448 if (hole <= r->end >> zfetch_hole_shift)
449 l = r->end;
450 }
451 if (l > 0)
452 return (dmu_zfetch_hit(zs, l));
453
454 return (0);
455 }
456
457 /*
458 * This is the predictive prefetch entry point. dmu_zfetch_prepare()
459 * associates dnode access specified with blkid and nblks arguments with
460 * prefetch stream, predicts further accesses based on that stats and returns
461 * the stream pointer on success. That pointer must later be passed to
462 * dmu_zfetch_run() to initiate the speculative prefetch for the stream and
463 * release it. dmu_zfetch() is a wrapper for simple cases when window between
464 * prediction and prefetch initiation is not needed.
465 * fetch_data argument specifies whether actual data blocks should be fetched:
466 * FALSE -- prefetch only indirect blocks for predicted data blocks;
467 * TRUE -- prefetch predicted data blocks plus following indirect blocks.
468 */
469 zstream_t *
dmu_zfetch_prepare(zfetch_t * zf,uint64_t blkid,uint64_t nblks,boolean_t fetch_data,boolean_t have_lock)470 dmu_zfetch_prepare(zfetch_t *zf, uint64_t blkid, uint64_t nblks,
471 boolean_t fetch_data, boolean_t have_lock)
472 {
473 zstream_t *zs;
474 spa_t *spa = zf->zf_dnode->dn_objset->os_spa;
475 zfs_prefetch_type_t os_prefetch = zf->zf_dnode->dn_objset->os_prefetch;
476 int64_t ipf_start, ipf_end;
477
478 if (zfs_prefetch_disable || os_prefetch == ZFS_PREFETCH_NONE)
479 return (NULL);
480
481 if (os_prefetch == ZFS_PREFETCH_METADATA)
482 fetch_data = B_FALSE;
483
484 /*
485 * If we haven't yet loaded the indirect vdevs' mappings, we
486 * can only read from blocks that we carefully ensure are on
487 * concrete vdevs (or previously-loaded indirect vdevs). So we
488 * can't allow the predictive prefetcher to attempt reads of other
489 * blocks (e.g. of the MOS's dnode object).
490 */
491 if (!spa_indirect_vdevs_loaded(spa))
492 return (NULL);
493
494 /*
495 * As a fast path for small (single-block) files, ignore access
496 * to the first block.
497 */
498 if (!have_lock && blkid == 0)
499 return (NULL);
500
501 if (!have_lock)
502 rw_enter(&zf->zf_dnode->dn_struct_rwlock, RW_READER);
503
504 /*
505 * A fast path for small files for which no prefetch will
506 * happen.
507 */
508 uint64_t maxblkid = zf->zf_dnode->dn_maxblkid;
509 if (maxblkid < 2) {
510 if (!have_lock)
511 rw_exit(&zf->zf_dnode->dn_struct_rwlock);
512 return (NULL);
513 }
514 mutex_enter(&zf->zf_lock);
515
516 /*
517 * Find perfect prefetch stream. Depending on whether the accesses
518 * are block-aligned, first block of the new access may either follow
519 * the last block of the previous access, or be equal to it.
520 */
521 unsigned int dbs = zf->zf_dnode->dn_datablkshift;
522 uint64_t end_blkid = blkid + nblks;
523 for (zs = list_head(&zf->zf_stream); zs != NULL;
524 zs = list_next(&zf->zf_stream, zs)) {
525 if (blkid == zs->zs_blkid) {
526 goto hit;
527 } else if (blkid + 1 == zs->zs_blkid) {
528 blkid++;
529 nblks--;
530 goto hit;
531 }
532 }
533
534 /*
535 * Find close enough prefetch stream. Access crossing stream position
536 * is a hit in its new part. Access ahead of stream position considered
537 * a hit for metadata prefetch, since we do not care about fill percent,
538 * or stored for future otherwise. Access behind stream position is
539 * silently ignored, since we already skipped it reaching fill percent.
540 */
541 uint_t max_reorder = MIN((zfetch_max_reorder >> dbs) + 1, UINT16_MAX);
542 uint_t t = gethrestime_sec() - zfetch_max_sec_reap;
543 for (zs = list_head(&zf->zf_stream); zs != NULL;
544 zs = list_next(&zf->zf_stream, zs)) {
545 if (blkid > zs->zs_blkid) {
546 if (end_blkid <= zs->zs_blkid + max_reorder) {
547 if (!fetch_data) {
548 nblks = dmu_zfetch_hit(zs,
549 end_blkid - zs->zs_blkid);
550 ZFETCHSTAT_BUMP(zfetchstat_stride);
551 goto future;
552 }
553 nblks = dmu_zfetch_future(zs, blkid, nblks);
554 if (nblks > 0)
555 ZFETCHSTAT_BUMP(zfetchstat_stride);
556 else
557 ZFETCHSTAT_BUMP(zfetchstat_future);
558 goto future;
559 }
560 } else if (end_blkid >= zs->zs_blkid) {
561 nblks -= zs->zs_blkid - blkid;
562 blkid += zs->zs_blkid - blkid;
563 goto hit;
564 } else if (end_blkid + max_reorder > zs->zs_blkid &&
565 (int)(zs->zs_atime - t) >= 0) {
566 ZFETCHSTAT_BUMP(zfetchstat_past);
567 zs->zs_atime = gethrestime_sec();
568 goto out;
569 }
570 }
571
572 /*
573 * This access is not part of any existing stream. Create a new
574 * stream for it unless we are at the end of file.
575 */
576 ASSERT0P(zs);
577 if (end_blkid < maxblkid)
578 dmu_zfetch_stream_create(zf, end_blkid);
579 mutex_exit(&zf->zf_lock);
580 ZFETCHSTAT_BUMP(zfetchstat_misses);
581 ipf_start = 0;
582 goto prescient;
583
584 hit:
585 nblks = dmu_zfetch_hit(zs, nblks);
586 ZFETCHSTAT_BUMP(zfetchstat_hits);
587
588 future:
589 zs->zs_atime = gethrestime_sec();
590
591 /* Exit if we already prefetched for this position before. */
592 if (nblks == 0)
593 goto out;
594
595 /* If the file is ending, remove the stream. */
596 end_blkid = zs->zs_blkid;
597 if (end_blkid >= maxblkid) {
598 dmu_zfetch_stream_remove(zf, zs);
599 out:
600 mutex_exit(&zf->zf_lock);
601 if (!have_lock)
602 rw_exit(&zf->zf_dnode->dn_struct_rwlock);
603 return (NULL);
604 }
605
606 /*
607 * This access was to a block that we issued a prefetch for on
608 * behalf of this stream. Calculate further prefetch distances.
609 *
610 * Start prefetch from the demand access size (nblks). Double the
611 * distance every access up to zfetch_min_distance. After that only
612 * if needed increase the distance by 1/8 up to zfetch_max_distance.
613 *
614 * Don't double the distance beyond single block if we have more
615 * than ~6% of ARC held by active prefetches. It should help with
616 * getting out of RAM on some badly mispredicted read patterns.
617 */
618 unsigned int nbytes = nblks << dbs;
619 unsigned int pf_nblks;
620 if (fetch_data) {
621 if (unlikely(zs->zs_pf_dist < nbytes))
622 zs->zs_pf_dist = nbytes;
623 else if (zs->zs_pf_dist < zfetch_min_distance &&
624 (zs->zs_pf_dist < (1 << dbs) ||
625 aggsum_compare(&zfetch_sums.zfetchstat_io_active,
626 arc_c_max >> (4 + dbs)) < 0))
627 zs->zs_pf_dist *= 2;
628 else if (zs->zs_more)
629 zs->zs_pf_dist += zs->zs_pf_dist / 8;
630 zs->zs_more = B_FALSE;
631 if (zs->zs_pf_dist > zfetch_max_distance)
632 zs->zs_pf_dist = zfetch_max_distance;
633 pf_nblks = zs->zs_pf_dist >> dbs;
634 } else {
635 pf_nblks = 0;
636 }
637 if (zs->zs_pf_start < end_blkid)
638 zs->zs_pf_start = end_blkid;
639 if (zs->zs_pf_end < end_blkid + pf_nblks)
640 zs->zs_pf_end = end_blkid + pf_nblks;
641
642 /*
643 * Do the same for indirects, starting where we will stop reading
644 * data blocks (and the indirects that point to them).
645 */
646 if (unlikely(zs->zs_ipf_dist < nbytes))
647 zs->zs_ipf_dist = nbytes;
648 else
649 zs->zs_ipf_dist *= 2;
650 if (zs->zs_ipf_dist > zfetch_max_idistance)
651 zs->zs_ipf_dist = zfetch_max_idistance;
652 pf_nblks = zs->zs_ipf_dist >> dbs;
653 if (zs->zs_ipf_start < zs->zs_pf_end)
654 zs->zs_ipf_start = zs->zs_pf_end;
655 ipf_start = zs->zs_ipf_end;
656 if (zs->zs_ipf_end < zs->zs_pf_end + pf_nblks)
657 zs->zs_ipf_end = zs->zs_pf_end + pf_nblks;
658
659 zfs_refcount_add(&zs->zs_refs, NULL);
660 /* Count concurrent callers. */
661 zfs_refcount_add(&zs->zs_callers, NULL);
662 mutex_exit(&zf->zf_lock);
663
664 prescient:
665 /*
666 * Prefetch the following indirect blocks for this access to reduce
667 * dbuf_hold() sync read delays in dmu_buf_hold_array_by_dnode().
668 * This covers the gap during the first couple accesses when we can
669 * not predict the future yet, but know what is needed right now.
670 * This should be very rare for reads/writes to need more than one
671 * indirect, but more useful for cloning due to much bigger accesses.
672 */
673 ipf_start = MAX(ipf_start, blkid + 1);
674 int epbs = zf->zf_dnode->dn_indblkshift - SPA_BLKPTRSHIFT;
675 ipf_start = P2ROUNDUP(ipf_start, 1 << epbs) >> epbs;
676 ipf_end = P2ROUNDUP(end_blkid, 1 << epbs) >> epbs;
677
678 int issued = 0;
679 for (int64_t iblk = ipf_start; iblk < ipf_end; iblk++) {
680 issued += dbuf_prefetch(zf->zf_dnode, 1, iblk,
681 ZIO_PRIORITY_SYNC_READ, ARC_FLAG_PRESCIENT_PREFETCH);
682 }
683
684 if (!have_lock)
685 rw_exit(&zf->zf_dnode->dn_struct_rwlock);
686 if (issued)
687 ZFETCHSTAT_ADD(zfetchstat_io_issued, issued);
688 return (zs);
689 }
690
691 void
dmu_zfetch_run(zfetch_t * zf,zstream_t * zs,boolean_t missed,boolean_t have_lock)692 dmu_zfetch_run(zfetch_t *zf, zstream_t *zs, boolean_t missed,
693 boolean_t have_lock)
694 {
695 int64_t pf_start, pf_end, ipf_start, ipf_end;
696 int epbs, issued;
697
698 if (missed)
699 zs->zs_missed = missed;
700
701 /*
702 * Postpone the prefetch if there are more concurrent callers.
703 * It happens when multiple requests are waiting for the same
704 * indirect block. The last one will run the prefetch for all.
705 */
706 if (zfs_refcount_remove(&zs->zs_callers, NULL) != 0) {
707 /* Drop reference taken in dmu_zfetch_prepare(). */
708 if (zfs_refcount_remove(&zs->zs_refs, NULL) == 0)
709 dmu_zfetch_stream_fini(zs);
710 return;
711 }
712
713 mutex_enter(&zf->zf_lock);
714 if (zs->zs_missed) {
715 pf_start = zs->zs_pf_start;
716 pf_end = zs->zs_pf_start = zs->zs_pf_end;
717 } else {
718 pf_start = pf_end = 0;
719 }
720 ipf_start = zs->zs_ipf_start;
721 ipf_end = zs->zs_ipf_start = zs->zs_ipf_end;
722 mutex_exit(&zf->zf_lock);
723 ASSERT3S(pf_start, <=, pf_end);
724 ASSERT3S(ipf_start, <=, ipf_end);
725
726 epbs = zf->zf_dnode->dn_indblkshift - SPA_BLKPTRSHIFT;
727 ipf_start = P2ROUNDUP(ipf_start, 1 << epbs) >> epbs;
728 ipf_end = P2ROUNDUP(ipf_end, 1 << epbs) >> epbs;
729 ASSERT3S(ipf_start, <=, ipf_end);
730 issued = pf_end - pf_start + ipf_end - ipf_start;
731 if (issued > 1) {
732 /* More references on top of taken in dmu_zfetch_prepare(). */
733 zfs_refcount_add_few(&zs->zs_refs, issued - 1, NULL);
734 } else if (issued == 0) {
735 /* Some other thread has done our work, so drop the ref. */
736 if (zfs_refcount_remove(&zs->zs_refs, NULL) == 0)
737 dmu_zfetch_stream_fini(zs);
738 return;
739 }
740 aggsum_add(&zfetch_sums.zfetchstat_io_active, issued);
741
742 if (!have_lock)
743 rw_enter(&zf->zf_dnode->dn_struct_rwlock, RW_READER);
744
745 issued = 0;
746 for (int64_t blk = pf_start; blk < pf_end; blk++) {
747 issued += dbuf_prefetch_impl(zf->zf_dnode, 0, blk,
748 ZIO_PRIORITY_ASYNC_READ, 0, dmu_zfetch_done, zs);
749 }
750 for (int64_t iblk = ipf_start; iblk < ipf_end; iblk++) {
751 issued += dbuf_prefetch_impl(zf->zf_dnode, 1, iblk,
752 ZIO_PRIORITY_ASYNC_READ, 0, dmu_zfetch_done, zs);
753 }
754
755 if (!have_lock)
756 rw_exit(&zf->zf_dnode->dn_struct_rwlock);
757
758 if (issued)
759 ZFETCHSTAT_ADD(zfetchstat_io_issued, issued);
760 }
761
762 void
dmu_zfetch(zfetch_t * zf,uint64_t blkid,uint64_t nblks,boolean_t fetch_data,boolean_t missed,boolean_t have_lock)763 dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data,
764 boolean_t missed, boolean_t have_lock)
765 {
766 zstream_t *zs;
767
768 zs = dmu_zfetch_prepare(zf, blkid, nblks, fetch_data, have_lock);
769 if (zs)
770 dmu_zfetch_run(zf, zs, missed, have_lock);
771 }
772
773 ZFS_MODULE_PARAM(zfs_prefetch, zfs_prefetch_, disable, INT, ZMOD_RW,
774 "Disable all ZFS prefetching");
775
776 ZFS_MODULE_PARAM(zfs_prefetch, zfetch_, max_streams, UINT, ZMOD_RW,
777 "Max number of streams per zfetch");
778
779 ZFS_MODULE_PARAM(zfs_prefetch, zfetch_, min_sec_reap, UINT, ZMOD_RW,
780 "Min time before stream reclaim");
781
782 ZFS_MODULE_PARAM(zfs_prefetch, zfetch_, max_sec_reap, UINT, ZMOD_RW,
783 "Max time before stream delete");
784
785 ZFS_MODULE_PARAM(zfs_prefetch, zfetch_, min_distance, UINT, ZMOD_RW,
786 "Min bytes to prefetch per stream");
787
788 ZFS_MODULE_PARAM(zfs_prefetch, zfetch_, max_distance, UINT, ZMOD_RW,
789 "Max bytes to prefetch per stream");
790
791 ZFS_MODULE_PARAM(zfs_prefetch, zfetch_, max_idistance, UINT, ZMOD_RW,
792 "Max bytes to prefetch indirects for per stream");
793
794 ZFS_MODULE_PARAM(zfs_prefetch, zfetch_, max_reorder, UINT, ZMOD_RW,
795 "Max request reorder distance within a stream");
796
797 ZFS_MODULE_PARAM(zfs_prefetch, zfetch_, hole_shift, UINT, ZMOD_RW,
798 "Max log2 fraction of holes in a stream");
799