2 * Copyright (c) 2010 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Ilya Dryomov <idryomov@gmail.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 #include <crypto/sha2/sha2.h>
40 #define DEDUP_BUF (64 * 1024)
42 /* Sorted list of block CRCs - light version for dedup-simulate */
43 struct sim_dedup_entry_rb_tree
;
44 RB_HEAD(sim_dedup_entry_rb_tree
, sim_dedup_entry
) sim_dedup_tree
=
45 RB_INITIALIZER(&sim_dedup_tree
);
46 RB_PROTOTYPE2(sim_dedup_entry_rb_tree
, sim_dedup_entry
, rb_entry
,
47 rb_sim_dedup_entry_compare
, hammer_crc_t
);
49 struct sim_dedup_entry
{
51 uint64_t ref_blks
; /* number of blocks referenced */
52 uint64_t ref_size
; /* size of data referenced */
53 RB_ENTRY(sim_dedup_entry
) rb_entry
;
57 struct hammer_btree_leaf_elm leaf
;
63 RB_HEAD(sha_dedup_entry_rb_tree
, sha_dedup_entry
) fict_root
;
66 RB_ENTRY(dedup_entry
) rb_entry
;
69 #define HAMMER_DEDUP_ENTRY_FICTITIOUS 0x0001
71 struct sha_dedup_entry
{
72 struct hammer_btree_leaf_elm leaf
;
75 uint8_t sha_hash
[SHA256_DIGEST_LENGTH
];
76 RB_ENTRY(sha_dedup_entry
) fict_entry
;
79 /* Sorted list of HAMMER B-Tree keys */
80 struct dedup_entry_rb_tree
;
81 struct sha_dedup_entry_rb_tree
;
83 RB_HEAD(dedup_entry_rb_tree
, dedup_entry
) dedup_tree
=
84 RB_INITIALIZER(&dedup_tree
);
85 RB_PROTOTYPE2(dedup_entry_rb_tree
, dedup_entry
, rb_entry
,
86 rb_dedup_entry_compare
, hammer_crc_t
);
88 RB_PROTOTYPE(sha_dedup_entry_rb_tree
, sha_dedup_entry
, fict_entry
,
89 rb_sha_dedup_entry_compare
);
92 * Pass2 list - contains entries that were not dedup'ed because ioctl failed
94 STAILQ_HEAD(, pass2_dedup_entry
) pass2_dedup_queue
=
95 STAILQ_HEAD_INITIALIZER(pass2_dedup_queue
);
97 struct pass2_dedup_entry
{
98 struct hammer_btree_leaf_elm leaf
;
99 STAILQ_ENTRY(pass2_dedup_entry
) sq_entry
;
102 #define DEDUP_PASS2 0x0001 /* process_btree_elm() mode */
104 static int SigInfoFlag
;
105 static int SigAlrmFlag
;
106 static int64_t DedupDataReads
;
107 static int64_t DedupCurrentRecords
;
108 static int64_t DedupTotalRecords
;
109 static uint32_t DedupCrcStart
;
110 static uint32_t DedupCrcEnd
;
111 static uint64_t MemoryUse
;
113 /* PFS global ids - we deal with just one PFS at a run */
115 static struct hammer_ioc_pseudofs_rw glob_pfs
;
118 * Global accounting variables
120 * Last three don't have to be 64-bit, just to be safe..
122 static uint64_t dedup_alloc_size
;
123 static uint64_t dedup_ref_size
;
124 static uint64_t dedup_skipped_size
;
125 static uint64_t dedup_crc_failures
;
126 static uint64_t dedup_sha_failures
;
127 static uint64_t dedup_underflows
;
128 static uint64_t dedup_successes_count
;
129 static uint64_t dedup_successes_bytes
;
131 static int rb_sim_dedup_entry_compare(struct sim_dedup_entry
*sim_de1
,
132 struct sim_dedup_entry
*sim_de2
);
133 static int rb_dedup_entry_compare(struct dedup_entry
*de1
,
134 struct dedup_entry
*de2
);
135 static int rb_sha_dedup_entry_compare(struct sha_dedup_entry
*sha_de1
,
136 struct sha_dedup_entry
*sha_de2
);
137 typedef int (*scan_pfs_cb_t
)(hammer_btree_leaf_elm_t scan_leaf
, int flags
);
138 static void scan_pfs(char *filesystem
, scan_pfs_cb_t func
, const char *id
);
139 static int collect_btree_elm(hammer_btree_leaf_elm_t scan_leaf
, int flags
);
140 static int count_btree_elm(hammer_btree_leaf_elm_t scan_leaf
, int flags
);
141 static int process_btree_elm(hammer_btree_leaf_elm_t scan_leaf
, int flags
);
142 static int upgrade_chksum(hammer_btree_leaf_elm_t leaf
, uint8_t *sha_hash
);
143 static void dump_simulated_dedup(void);
144 static void dump_real_dedup(void);
145 static void dedup_usage(int code
);
147 RB_GENERATE2(sim_dedup_entry_rb_tree
, sim_dedup_entry
, rb_entry
,
148 rb_sim_dedup_entry_compare
, hammer_crc_t
, crc
);
149 RB_GENERATE2(dedup_entry_rb_tree
, dedup_entry
, rb_entry
,
150 rb_dedup_entry_compare
, hammer_crc_t
, leaf
.data_crc
);
151 RB_GENERATE(sha_dedup_entry_rb_tree
, sha_dedup_entry
, fict_entry
,
152 rb_sha_dedup_entry_compare
);
155 rb_sim_dedup_entry_compare(struct sim_dedup_entry
*sim_de1
,
156 struct sim_dedup_entry
*sim_de2
)
158 if (sim_de1
->crc
< sim_de2
->crc
)
160 if (sim_de1
->crc
> sim_de2
->crc
)
167 rb_dedup_entry_compare(struct dedup_entry
*de1
, struct dedup_entry
*de2
)
169 if (de1
->leaf
.data_crc
< de2
->leaf
.data_crc
)
171 if (de1
->leaf
.data_crc
> de2
->leaf
.data_crc
)
178 rb_sha_dedup_entry_compare(struct sha_dedup_entry
*sha_de1
,
179 struct sha_dedup_entry
*sha_de2
)
181 unsigned long *h1
= (unsigned long *)&sha_de1
->sha_hash
;
182 unsigned long *h2
= (unsigned long *)&sha_de2
->sha_hash
;
185 for (i
= 0; i
< SHA256_DIGEST_LENGTH
/ (int)sizeof(unsigned long); ++i
) {
196 * dedup-simulate <filesystem>
199 hammer_cmd_dedup_simulate(char **av
, int ac
)
201 struct sim_dedup_entry
*sim_de
;
208 glob_fd
= getpfs(&glob_pfs
, av
[0]);
211 * Collection passes (memory limited)
213 printf("Dedup-simulate running\n");
215 DedupCrcStart
= DedupCrcEnd
;
220 printf("B-Tree pass crc-range %08x-max\n",
224 scan_pfs(av
[0], collect_btree_elm
, "simu-pass");
227 dump_simulated_dedup();
230 * Calculate simulated dedup ratio and get rid of the tree
232 while ((sim_de
= RB_ROOT(&sim_dedup_tree
)) != NULL
) {
233 assert(sim_de
->ref_blks
!= 0);
234 dedup_ref_size
+= sim_de
->ref_size
;
235 dedup_alloc_size
+= sim_de
->ref_size
/ sim_de
->ref_blks
;
237 RB_REMOVE(sim_dedup_entry_rb_tree
, &sim_dedup_tree
, sim_de
);
240 if (DedupCrcEnd
&& VerboseOpt
== 0)
242 } while (DedupCrcEnd
);
244 printf("Dedup-simulate %s succeeded\n", av
[0]);
245 relpfs(glob_fd
, &glob_pfs
);
247 printf("Simulated dedup ratio = %.2f\n",
248 (dedup_alloc_size
!= 0) ?
249 (double)dedup_ref_size
/ dedup_alloc_size
: 0);
256 hammer_cmd_dedup(char **av
, int ac
)
258 struct dedup_entry
*de
;
259 struct sha_dedup_entry
*sha_de
;
260 struct pass2_dedup_entry
*pass2_de
;
273 STAILQ_INIT(&pass2_dedup_queue
);
275 glob_fd
= getpfs(&glob_pfs
, av
[0]);
278 * A cycle file is _required_ for resuming dedup after the timeout
279 * specified with -t has expired. If no -c option, then place a
280 * .dedup.cycle file either in the PFS snapshots directory or in
281 * the default snapshots directory.
284 if (glob_pfs
.ondisk
->snapshots
[0] != '/')
285 asprintf(&tmp
, "%s/%s/.dedup.cycle",
286 SNAPSHOTS_BASE
, av
[0]);
288 asprintf(&tmp
, "%s/.dedup.cycle",
289 glob_pfs
.ondisk
->snapshots
);
295 * Pre-pass to cache the btree
297 scan_pfs(av
[0], count_btree_elm
, "pre-pass ");
298 DedupTotalRecords
= DedupCurrentRecords
;
301 * Collection passes (memory limited)
303 printf("Dedup running\n");
305 DedupCrcStart
= DedupCrcEnd
;
310 printf("B-Tree pass crc-range %08x-max\n",
314 scan_pfs(av
[0], process_btree_elm
, "main-pass");
316 while ((pass2_de
= STAILQ_FIRST(&pass2_dedup_queue
)) != NULL
) {
317 if (process_btree_elm(&pass2_de
->leaf
, DEDUP_PASS2
))
318 dedup_skipped_size
-= pass2_de
->leaf
.data_len
;
320 STAILQ_REMOVE_HEAD(&pass2_dedup_queue
, sq_entry
);
323 assert(STAILQ_EMPTY(&pass2_dedup_queue
));
329 * Calculate dedup ratio and get rid of the trees
331 while ((de
= RB_ROOT(&dedup_tree
)) != NULL
) {
332 if (de
->flags
& HAMMER_DEDUP_ENTRY_FICTITIOUS
) {
333 while ((sha_de
= RB_ROOT(&de
->u
.fict_root
)) != NULL
) {
334 assert(sha_de
->ref_blks
!= 0);
335 dedup_ref_size
+= sha_de
->ref_size
;
336 dedup_alloc_size
+= sha_de
->ref_size
/ sha_de
->ref_blks
;
338 RB_REMOVE(sha_dedup_entry_rb_tree
,
339 &de
->u
.fict_root
, sha_de
);
342 assert(RB_EMPTY(&de
->u
.fict_root
));
344 assert(de
->u
.de
.ref_blks
!= 0);
345 dedup_ref_size
+= de
->u
.de
.ref_size
;
346 dedup_alloc_size
+= de
->u
.de
.ref_size
/ de
->u
.de
.ref_blks
;
349 RB_REMOVE(dedup_entry_rb_tree
, &dedup_tree
, de
);
352 assert(RB_EMPTY(&dedup_tree
));
353 if (DedupCrcEnd
&& VerboseOpt
== 0)
355 } while (DedupCrcEnd
);
357 printf("Dedup %s succeeded\n", av
[0]);
358 relpfs(glob_fd
, &glob_pfs
);
360 humanize_unsigned(buf
, sizeof(buf
), dedup_ref_size
, "B", 1024);
361 printf("Dedup ratio = %.2f (in this run)\n"
363 ((dedup_alloc_size
!= 0) ?
364 (double)dedup_ref_size
/ dedup_alloc_size
: 0),
367 humanize_unsigned(buf
, sizeof(buf
), dedup_alloc_size
, "B", 1024);
368 printf(" %8s allocated\n", buf
);
369 humanize_unsigned(buf
, sizeof(buf
), dedup_skipped_size
, "B", 1024);
370 printf(" %8s skipped\n", buf
);
371 printf(" %8jd CRC collisions\n"
372 " %8jd SHA collisions\n"
373 " %8jd big-block underflows\n"
374 " %8jd new dedup records\n"
375 " %8jd new dedup bytes\n",
376 (intmax_t)dedup_crc_failures
,
377 (intmax_t)dedup_sha_failures
,
378 (intmax_t)dedup_underflows
,
379 (intmax_t)dedup_successes_count
,
380 (intmax_t)dedup_successes_bytes
383 /* Once completed remove cycle file */
384 hammer_reset_cycle();
386 /* We don't want to mess up with other directives */
394 count_btree_elm(hammer_btree_leaf_elm_t scan_leaf __unused
, int flags __unused
)
400 collect_btree_elm(hammer_btree_leaf_elm_t scan_leaf
, int flags __unused
)
402 struct sim_dedup_entry
*sim_de
;
405 * If we are using too much memory we have to clean some out, which
406 * will cause the run to use multiple passes. Be careful of integer
409 if (MemoryUse
> MemoryLimit
) {
410 DedupCrcEnd
= DedupCrcStart
+
411 (uint32_t)(DedupCrcEnd
- DedupCrcStart
- 1) / 2;
413 printf("memory limit crc-range %08x-%08x\n",
414 DedupCrcStart
, DedupCrcEnd
);
418 sim_de
= RB_MAX(sim_dedup_entry_rb_tree
,
420 if (sim_de
== NULL
|| sim_de
->crc
< DedupCrcEnd
)
422 RB_REMOVE(sim_dedup_entry_rb_tree
,
423 &sim_dedup_tree
, sim_de
);
424 MemoryUse
-= sizeof(*sim_de
);
430 * Collect statistics based on the CRC only, do not try to read
431 * any data blocks or run SHA hashes.
433 sim_de
= RB_LOOKUP(sim_dedup_entry_rb_tree
, &sim_dedup_tree
,
434 scan_leaf
->data_crc
);
436 if (sim_de
== NULL
) {
437 sim_de
= calloc(1, sizeof(*sim_de
));
438 sim_de
->crc
= scan_leaf
->data_crc
;
439 RB_INSERT(sim_dedup_entry_rb_tree
, &sim_dedup_tree
, sim_de
);
440 MemoryUse
+= sizeof(*sim_de
);
443 sim_de
->ref_blks
+= 1;
444 sim_de
->ref_size
+= scan_leaf
->data_len
;
449 validate_dedup_pair(hammer_btree_leaf_elm_t p
, hammer_btree_leaf_elm_t q
)
451 if (HAMMER_ZONE(p
->data_offset
) != HAMMER_ZONE(q
->data_offset
))
453 if (p
->data_len
!= q
->data_len
)
459 #define DEDUP_TECH_FAILURE 1
460 #define DEDUP_CMP_FAILURE 2
461 #define DEDUP_INVALID_ZONE 3
462 #define DEDUP_UNDERFLOW 4
463 #define DEDUP_VERS_FAILURE 5
466 deduplicate(hammer_btree_leaf_elm_t p
, hammer_btree_leaf_elm_t q
)
468 struct hammer_ioc_dedup dedup
;
470 bzero(&dedup
, sizeof(dedup
));
473 * If data_offset fields are the same there is no need to run ioctl,
474 * candidate is already dedup'ed.
476 if (p
->data_offset
== q
->data_offset
)
479 dedup
.elm1
= p
->base
;
480 dedup
.elm2
= q
->base
;
482 if (ioctl(glob_fd
, HAMMERIOC_DEDUP
, &dedup
) < 0) {
483 if (errno
== EOPNOTSUPP
)
484 return (DEDUP_VERS_FAILURE
); /* must be at least version 5 */
485 /* Technical failure - locking or w/e */
486 return (DEDUP_TECH_FAILURE
);
488 if (dedup
.head
.flags
& HAMMER_IOC_DEDUP_CMP_FAILURE
)
489 return (DEDUP_CMP_FAILURE
);
490 if (dedup
.head
.flags
& HAMMER_IOC_DEDUP_INVALID_ZONE
)
491 return (DEDUP_INVALID_ZONE
);
492 if (dedup
.head
.flags
& HAMMER_IOC_DEDUP_UNDERFLOW
)
493 return (DEDUP_UNDERFLOW
);
495 ++dedup_successes_count
;
496 dedup_successes_bytes
+= p
->data_len
;
501 process_btree_elm(hammer_btree_leaf_elm_t scan_leaf
, int flags
)
503 struct dedup_entry
*de
;
504 struct sha_dedup_entry
*sha_de
, temp
;
505 struct pass2_dedup_entry
*pass2_de
;
509 * If we are using too much memory we have to clean some out, which
510 * will cause the run to use multiple passes. Be careful of integer
513 while (MemoryUse
> MemoryLimit
) {
514 DedupCrcEnd
= DedupCrcStart
+
515 (uint32_t)(DedupCrcEnd
- DedupCrcStart
- 1) / 2;
517 printf("memory limit crc-range %08x-%08x\n",
518 DedupCrcStart
, DedupCrcEnd
);
523 de
= RB_MAX(dedup_entry_rb_tree
, &dedup_tree
);
524 if (de
== NULL
|| de
->leaf
.data_crc
< DedupCrcEnd
)
526 if (de
->flags
& HAMMER_DEDUP_ENTRY_FICTITIOUS
) {
527 while ((sha_de
= RB_ROOT(&de
->u
.fict_root
)) !=
529 RB_REMOVE(sha_dedup_entry_rb_tree
,
530 &de
->u
.fict_root
, sha_de
);
531 MemoryUse
-= sizeof(*sha_de
);
535 RB_REMOVE(dedup_entry_rb_tree
, &dedup_tree
, de
);
536 MemoryUse
-= sizeof(*de
);
542 * Collect statistics based on the CRC. Colliding CRCs usually
543 * cause a SHA sub-tree to be created under the de.
545 * Trivial case if de not found.
547 de
= RB_LOOKUP(dedup_entry_rb_tree
, &dedup_tree
, scan_leaf
->data_crc
);
549 de
= calloc(1, sizeof(*de
));
550 de
->leaf
= *scan_leaf
;
551 RB_INSERT(dedup_entry_rb_tree
, &dedup_tree
, de
);
552 MemoryUse
+= sizeof(*de
);
557 * Found entry in CRC tree
559 if (de
->flags
& HAMMER_DEDUP_ENTRY_FICTITIOUS
) {
561 * Optimize the case where a CRC failure results in multiple
562 * SHA entries. If we unconditionally issue a data-read a
563 * degenerate situation where a colliding CRC's second SHA
564 * entry contains the lion's share of the deduplication
565 * candidates will result in excessive data block reads.
567 * Deal with the degenerate case by looking for a matching
568 * data_offset/data_len in the SHA elements we already have
569 * before reading the data block and generating a new SHA.
571 RB_FOREACH(sha_de
, sha_dedup_entry_rb_tree
, &de
->u
.fict_root
) {
572 if (sha_de
->leaf
.data_offset
==
573 scan_leaf
->data_offset
&&
574 sha_de
->leaf
.data_len
== scan_leaf
->data_len
) {
575 memcpy(temp
.sha_hash
, sha_de
->sha_hash
,
576 SHA256_DIGEST_LENGTH
);
582 * Entry in CRC tree is fictitious, so we already had problems
583 * with this CRC. Upgrade (compute SHA) the candidate and
584 * dive into SHA subtree. If upgrade fails insert the candidate
585 * into Pass2 list (it will be processed later).
587 if (sha_de
== NULL
) {
588 if (upgrade_chksum(scan_leaf
, temp
.sha_hash
))
591 sha_de
= RB_FIND(sha_dedup_entry_rb_tree
,
592 &de
->u
.fict_root
, &temp
);
596 * Nothing in SHA subtree so far, so this is a new
597 * 'dataset'. Insert new entry into SHA subtree.
599 if (sha_de
== NULL
) {
600 sha_de
= calloc(1, sizeof(*sha_de
));
601 sha_de
->leaf
= *scan_leaf
;
602 memcpy(sha_de
->sha_hash
, temp
.sha_hash
,
603 SHA256_DIGEST_LENGTH
);
604 RB_INSERT(sha_dedup_entry_rb_tree
, &de
->u
.fict_root
,
606 MemoryUse
+= sizeof(*sha_de
);
607 goto upgrade_stats_sha
;
611 * Found entry in SHA subtree, it means we have a potential
612 * dedup pair. Validate it (zones have to match and data_len
613 * field have to be the same too. If validation fails, treat
614 * it as a SHA collision (jump to sha256_failure).
616 if (validate_dedup_pair(&sha_de
->leaf
, scan_leaf
))
620 * We have a valid dedup pair (SHA match, validated).
622 * In case of technical failure (dedup pair was good, but
623 * ioctl failed anyways) insert the candidate into Pass2 list
624 * (we will try to dedup it after we are done with the rest of
627 * If ioctl fails because either of blocks is in the non-dedup
628 * zone (we can dedup only in LARGE_DATA and SMALL_DATA) don't
629 * bother with the candidate and terminate early.
631 * If ioctl fails because of big-block underflow replace the
632 * leaf node that found dedup entry represents with scan_leaf.
634 error
= deduplicate(&sha_de
->leaf
, scan_leaf
);
637 goto upgrade_stats_sha
;
638 case DEDUP_TECH_FAILURE
:
640 case DEDUP_CMP_FAILURE
:
642 case DEDUP_INVALID_ZONE
:
643 goto terminate_early
;
644 case DEDUP_UNDERFLOW
:
646 sha_de
->leaf
= *scan_leaf
;
647 memcpy(sha_de
->sha_hash
, temp
.sha_hash
,
648 SHA256_DIGEST_LENGTH
);
649 goto upgrade_stats_sha
;
650 case DEDUP_VERS_FAILURE
:
651 errx(1, "HAMMER filesystem must be at least "
652 "version 5 to dedup");
654 fprintf(stderr
, "Unknown error\n");
655 goto terminate_early
;
659 * Ooh la la.. SHA-256 collision. Terminate early, there's
660 * nothing we can do here.
663 ++dedup_sha_failures
;
664 goto terminate_early
;
667 * Candidate CRC is good for now (we found an entry in CRC
668 * tree and it's not fictitious). This means we have a
669 * potential dedup pair.
671 if (validate_dedup_pair(&de
->leaf
, scan_leaf
))
675 * We have a valid dedup pair (CRC match, validated)
677 error
= deduplicate(&de
->leaf
, scan_leaf
);
681 case DEDUP_TECH_FAILURE
:
683 case DEDUP_CMP_FAILURE
:
685 case DEDUP_INVALID_ZONE
:
686 goto terminate_early
;
687 case DEDUP_UNDERFLOW
:
689 de
->leaf
= *scan_leaf
;
691 case DEDUP_VERS_FAILURE
:
692 errx(1, "HAMMER filesystem must be at least "
693 "version 5 to dedup");
695 fprintf(stderr
, "Unknown error\n");
696 goto terminate_early
;
701 * We got a CRC collision - either ioctl failed because of
702 * the comparison failure or validation of the potential
703 * dedup pair went bad. In all cases insert both blocks
704 * into SHA subtree (this requires checksum upgrade) and mark
705 * entry that corresponds to this CRC in the CRC tree
706 * fictitious, so that all futher operations with this CRC go
707 * through SHA subtree.
709 ++dedup_crc_failures
;
712 * Insert block that was represented by now fictitious dedup
713 * entry (create a new SHA entry and preserve stats of the
714 * old CRC one). If checksum upgrade fails insert the
715 * candidate into Pass2 list and return - keep both trees
718 sha_de
= calloc(1, sizeof(*sha_de
));
719 sha_de
->leaf
= de
->leaf
;
720 sha_de
->ref_blks
= de
->u
.de
.ref_blks
;
721 sha_de
->ref_size
= de
->u
.de
.ref_size
;
722 if (upgrade_chksum(&sha_de
->leaf
, sha_de
->sha_hash
)) {
726 MemoryUse
+= sizeof(*sha_de
);
728 RB_INIT(&de
->u
.fict_root
);
730 * Here we can insert without prior checking because the tree
731 * is empty at this point
733 RB_INSERT(sha_dedup_entry_rb_tree
, &de
->u
.fict_root
, sha_de
);
736 * Mark entry in CRC tree fictitious
738 de
->flags
|= HAMMER_DEDUP_ENTRY_FICTITIOUS
;
741 * Upgrade checksum of the candidate and insert it into
742 * SHA subtree. If upgrade fails insert the candidate into
745 if (upgrade_chksum(scan_leaf
, temp
.sha_hash
))
747 sha_de
= RB_FIND(sha_dedup_entry_rb_tree
, &de
->u
.fict_root
,
750 /* There is an entry with this SHA already, but the only
751 * RB-tree element at this point is that entry we just
752 * added. We know for sure these blocks are different
753 * (this is crc_failure branch) so treat it as SHA
758 sha_de
= calloc(1, sizeof(*sha_de
));
759 sha_de
->leaf
= *scan_leaf
;
760 memcpy(sha_de
->sha_hash
, temp
.sha_hash
, SHA256_DIGEST_LENGTH
);
761 RB_INSERT(sha_dedup_entry_rb_tree
, &de
->u
.fict_root
, sha_de
);
762 MemoryUse
+= sizeof(*sha_de
);
763 goto upgrade_stats_sha
;
767 de
->u
.de
.ref_blks
+= 1;
768 de
->u
.de
.ref_size
+= scan_leaf
->data_len
;
772 sha_de
->ref_blks
+= 1;
773 sha_de
->ref_size
+= scan_leaf
->data_len
;
778 * If in pass2 mode don't insert anything, fall through to
781 if ((flags
& DEDUP_PASS2
) == 0) {
782 pass2_de
= calloc(1, sizeof(*pass2_de
));
783 pass2_de
->leaf
= *scan_leaf
;
784 STAILQ_INSERT_TAIL(&pass2_dedup_queue
, pass2_de
, sq_entry
);
785 dedup_skipped_size
+= scan_leaf
->data_len
;
791 * Early termination path. Fixup stats.
793 dedup_alloc_size
+= scan_leaf
->data_len
;
794 dedup_ref_size
+= scan_leaf
->data_len
;
799 upgrade_chksum(hammer_btree_leaf_elm_t leaf
, uint8_t *sha_hash
)
801 struct hammer_ioc_data data
;
802 char *buf
= malloc(DEDUP_BUF
);
806 bzero(&data
, sizeof(data
));
807 data
.elm
= leaf
->base
;
809 data
.size
= DEDUP_BUF
;
812 if (ioctl(glob_fd
, HAMMERIOC_GET_DATA
, &data
) < 0) {
813 fprintf(stderr
, "Get-data failed: %s\n", strerror(errno
));
817 DedupDataReads
+= leaf
->data_len
;
819 if (data
.leaf
.data_len
!= leaf
->data_len
) {
824 if (data
.leaf
.base
.btype
== HAMMER_BTREE_TYPE_RECORD
&&
825 data
.leaf
.base
.rec_type
== HAMMER_RECTYPE_DATA
) {
827 SHA256_Update(&ctx
, (void *)buf
, data
.leaf
.data_len
);
828 SHA256_Final(sha_hash
, &ctx
);
837 sigAlrm(int signo __unused
)
843 sigInfo(int signo __unused
)
849 scan_pfs(char *filesystem
, scan_pfs_cb_t func
, const char *id
)
851 struct hammer_ioc_mirror_rw mirror
;
852 hammer_ioc_mrecord_any_t mrec
;
853 struct hammer_btree_leaf_elm elm
;
854 char *buf
= malloc(DEDUP_BUF
);
861 DedupCurrentRecords
= 0;
862 signal(SIGINFO
, sigInfo
);
863 signal(SIGALRM
, sigAlrm
);
866 * Deduplication happens per element so hammer(8) is in full
867 * control of the ioctl()s to actually perform it. SIGALRM
868 * needs to be handled within hammer(8) but a checkpoint
869 * is needed for resuming. Use cycle file for that.
871 * Try to obtain the previous obj_id from the cycle file and
872 * if not available just start from the beginning.
874 bzero(&mirror
, sizeof(mirror
));
875 hammer_key_beg_init(&mirror
.key_beg
);
876 hammer_get_cycle(&mirror
.key_beg
, &mirror
.tid_beg
);
878 if (mirror
.key_beg
.obj_id
!= (int64_t)HAMMER_MIN_OBJID
) {
880 fprintf(stderr
, "%s: mirror-read: Resuming at object %016jx\n",
881 id
, (uintmax_t)mirror
.key_beg
.obj_id
);
884 hammer_key_end_init(&mirror
.key_end
);
886 mirror
.tid_beg
= glob_pfs
.ondisk
->sync_beg_tid
;
887 mirror
.tid_end
= glob_pfs
.ondisk
->sync_end_tid
;
888 mirror
.head
.flags
|= HAMMER_IOC_MIRROR_NODATA
; /* we want only keys */
890 mirror
.size
= DEDUP_BUF
;
891 mirror
.pfs_id
= glob_pfs
.pfs_id
;
892 mirror
.shared_uuid
= glob_pfs
.ondisk
->shared_uuid
;
894 if (VerboseOpt
&& DedupCrcStart
== 0) {
895 printf("%s %s: objspace %016jx:%04x %016jx:%04x\n",
897 (uintmax_t)mirror
.key_beg
.obj_id
,
898 mirror
.key_beg
.localization
,
899 (uintmax_t)mirror
.key_end
.obj_id
,
900 mirror
.key_end
.localization
);
901 printf("%s %s: pfs_id %d\n",
902 id
, filesystem
, glob_pfs
.pfs_id
);
909 mirror
.pfs_id
= glob_pfs
.pfs_id
;
910 mirror
.shared_uuid
= glob_pfs
.ondisk
->shared_uuid
;
911 if (ioctl(glob_fd
, HAMMERIOC_MIRROR_READ
, &mirror
) < 0) {
912 err(1, "Mirror-read %s failed", filesystem
);
915 if (mirror
.head
.flags
& HAMMER_IOC_HEAD_ERROR
) {
916 errx(1, "Mirror-read %s fatal error %d",
917 filesystem
, mirror
.head
.error
);
922 while (offset
< mirror
.count
) {
923 mrec
= (void *)((char *)buf
+ offset
);
924 bytes
= HAMMER_HEAD_DOALIGN(mrec
->head
.rec_size
);
925 if (offset
+ bytes
> mirror
.count
)
926 errx(1, "Misaligned record");
927 assert((mrec
->head
.type
&
928 HAMMER_MRECF_TYPE_MASK
) ==
929 HAMMER_MREC_TYPE_REC
);
931 elm
= mrec
->rec
.leaf
;
932 if (elm
.base
.btype
!= HAMMER_BTREE_TYPE_RECORD
)
934 if (elm
.base
.rec_type
!= HAMMER_RECTYPE_DATA
)
936 ++DedupCurrentRecords
;
937 if (DedupCrcStart
!= DedupCrcEnd
) {
938 if (elm
.data_crc
< DedupCrcStart
)
941 elm
.data_crc
>= DedupCrcEnd
) {
948 mirror
.key_beg
= mirror
.key_cur
;
949 if (DidInterrupt
|| SigAlrmFlag
) {
951 fprintf(stderr
, "%s\n",
952 (DidInterrupt
? "Interrupted" : "Timeout"));
953 hammer_set_cycle(&mirror
.key_cur
, mirror
.tid_beg
);
955 fprintf(stderr
, "Cyclefile %s updated for "
956 "continuation\n", CyclePath
);
960 if (DedupTotalRecords
) {
961 humanize_unsigned(buf1
, sizeof(buf1
),
964 humanize_unsigned(buf2
, sizeof(buf2
),
965 dedup_successes_bytes
,
967 fprintf(stderr
, "%s count %7jd/%jd "
969 "ioread %s newddup %s\n",
971 (intmax_t)DedupCurrentRecords
,
972 (intmax_t)DedupTotalRecords
,
973 (int)(DedupCurrentRecords
* 100 /
975 (int)(DedupCurrentRecords
* 10000 /
976 DedupTotalRecords
% 100),
979 fprintf(stderr
, "%s count %-7jd\n",
981 (intmax_t)DedupCurrentRecords
);
985 } while (mirror
.count
!= 0);
987 signal(SIGINFO
, SIG_IGN
);
988 signal(SIGALRM
, SIG_IGN
);
994 dump_simulated_dedup(void)
996 struct sim_dedup_entry
*sim_de
;
998 printf("=== Dumping simulated dedup entries:\n");
999 RB_FOREACH(sim_de
, sim_dedup_entry_rb_tree
, &sim_dedup_tree
) {
1000 printf("\tcrc=%08x cnt=%ju size=%ju\n",
1002 (intmax_t)sim_de
->ref_blks
,
1003 (intmax_t)sim_de
->ref_size
);
1005 printf("end of dump ===\n");
1009 dump_real_dedup(void)
1011 struct dedup_entry
*de
;
1012 struct sha_dedup_entry
*sha_de
;
1015 printf("=== Dumping dedup entries:\n");
1016 RB_FOREACH(de
, dedup_entry_rb_tree
, &dedup_tree
) {
1017 if (de
->flags
& HAMMER_DEDUP_ENTRY_FICTITIOUS
) {
1018 printf("\tcrc=%08x fictitious\n", de
->leaf
.data_crc
);
1020 RB_FOREACH(sha_de
, sha_dedup_entry_rb_tree
, &de
->u
.fict_root
) {
1021 printf("\t\tcrc=%08x cnt=%ju size=%ju\n\t"
1023 sha_de
->leaf
.data_crc
,
1024 (intmax_t)sha_de
->ref_blks
,
1025 (intmax_t)sha_de
->ref_size
);
1026 for (i
= 0; i
< SHA256_DIGEST_LENGTH
; ++i
)
1027 printf("%02x", sha_de
->sha_hash
[i
]);
1031 printf("\tcrc=%08x cnt=%ju size=%ju\n",
1033 (intmax_t)de
->u
.de
.ref_blks
,
1034 (intmax_t)de
->u
.de
.ref_size
);
1037 printf("end of dump ===\n");
1041 dedup_usage(int code
)
1044 "hammer dedup-simulate <filesystem>\n"
1045 "hammer dedup <filesystem>\n"