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);
156 rb_sim_dedup_entry_compare(struct sim_dedup_entry *sim_de1,
157 struct sim_dedup_entry *sim_de2)
159 if (sim_de1->crc < sim_de2->crc)
161 if (sim_de1->crc > sim_de2->crc)
169 rb_dedup_entry_compare(struct dedup_entry *de1, struct dedup_entry *de2)
171 if (de1->leaf.data_crc < de2->leaf.data_crc)
173 if (de1->leaf.data_crc > de2->leaf.data_crc)
181 rb_sha_dedup_entry_compare(struct sha_dedup_entry *sha_de1,
182 struct sha_dedup_entry *sha_de2)
184 unsigned long *h1 = (unsigned long *)&sha_de1->sha_hash;
185 unsigned long *h2 = (unsigned long *)&sha_de2->sha_hash;
188 for (i = 0; i < SHA256_DIGEST_LENGTH / (int)sizeof(unsigned long); ++i) {
199 * dedup-simulate <filesystem>
202 hammer_cmd_dedup_simulate(char **av, int ac)
204 struct sim_dedup_entry *sim_de;
211 glob_fd = getpfs(&glob_pfs, av[0]);
214 * Collection passes (memory limited)
216 printf("Dedup-simulate running\n");
218 DedupCrcStart = DedupCrcEnd;
223 printf("B-Tree pass crc-range %08x-max\n",
227 scan_pfs(av[0], collect_btree_elm, "simu-pass");
230 dump_simulated_dedup();
233 * Calculate simulated dedup ratio and get rid of the tree
235 while ((sim_de = RB_ROOT(&sim_dedup_tree)) != NULL) {
236 assert(sim_de->ref_blks != 0);
237 dedup_ref_size += sim_de->ref_size;
238 dedup_alloc_size += sim_de->ref_size / sim_de->ref_blks;
240 RB_REMOVE(sim_dedup_entry_rb_tree, &sim_dedup_tree, sim_de);
243 if (DedupCrcEnd && VerboseOpt == 0)
245 } while (DedupCrcEnd);
247 printf("Dedup-simulate %s succeeded\n", av[0]);
248 relpfs(glob_fd, &glob_pfs);
250 printf("Simulated dedup ratio = %.2f\n",
251 (dedup_alloc_size != 0) ?
252 (double)dedup_ref_size / dedup_alloc_size : 0);
259 hammer_cmd_dedup(char **av, int ac)
261 struct dedup_entry *de;
262 struct sha_dedup_entry *sha_de;
263 struct pass2_dedup_entry *pass2_de;
276 STAILQ_INIT(&pass2_dedup_queue);
278 glob_fd = getpfs(&glob_pfs, av[0]);
281 * A cycle file is _required_ for resuming dedup after the timeout
282 * specified with -t has expired. If no -c option, then place a
283 * .dedup.cycle file either in the PFS snapshots directory or in
284 * the default snapshots directory.
287 if (glob_pfs.ondisk->snapshots[0] != '/')
288 asprintf(&tmp, "%s/%s/.dedup.cycle",
289 SNAPSHOTS_BASE, av[0]);
291 asprintf(&tmp, "%s/.dedup.cycle",
292 glob_pfs.ondisk->snapshots);
298 * Pre-pass to cache the btree
300 scan_pfs(av[0], count_btree_elm, "pre-pass ");
301 DedupTotalRecords = DedupCurrentRecords;
304 * Collection passes (memory limited)
306 printf("Dedup running\n");
308 DedupCrcStart = DedupCrcEnd;
313 printf("B-Tree pass crc-range %08x-max\n",
317 scan_pfs(av[0], process_btree_elm, "main-pass");
319 while ((pass2_de = STAILQ_FIRST(&pass2_dedup_queue)) != NULL) {
320 if (process_btree_elm(&pass2_de->leaf, DEDUP_PASS2))
321 dedup_skipped_size -= pass2_de->leaf.data_len;
323 STAILQ_REMOVE_HEAD(&pass2_dedup_queue, sq_entry);
326 assert(STAILQ_EMPTY(&pass2_dedup_queue));
332 * Calculate dedup ratio and get rid of the trees
334 while ((de = RB_ROOT(&dedup_tree)) != NULL) {
335 if (de->flags & HAMMER_DEDUP_ENTRY_FICTITIOUS) {
336 while ((sha_de = RB_ROOT(&de->u.fict_root)) != NULL) {
337 assert(sha_de->ref_blks != 0);
338 dedup_ref_size += sha_de->ref_size;
339 dedup_alloc_size += sha_de->ref_size / sha_de->ref_blks;
341 RB_REMOVE(sha_dedup_entry_rb_tree,
342 &de->u.fict_root, sha_de);
345 assert(RB_EMPTY(&de->u.fict_root));
347 assert(de->u.de.ref_blks != 0);
348 dedup_ref_size += de->u.de.ref_size;
349 dedup_alloc_size += de->u.de.ref_size / de->u.de.ref_blks;
352 RB_REMOVE(dedup_entry_rb_tree, &dedup_tree, de);
355 assert(RB_EMPTY(&dedup_tree));
356 if (DedupCrcEnd && VerboseOpt == 0)
358 } while (DedupCrcEnd);
360 printf("Dedup %s succeeded\n", av[0]);
361 relpfs(glob_fd, &glob_pfs);
363 humanize_unsigned(buf, sizeof(buf), dedup_ref_size, "B", 1024);
364 printf("Dedup ratio = %.2f (in this run)\n"
366 ((dedup_alloc_size != 0) ?
367 (double)dedup_ref_size / dedup_alloc_size : 0),
370 humanize_unsigned(buf, sizeof(buf), dedup_alloc_size, "B", 1024);
371 printf(" %8s allocated\n", buf);
372 humanize_unsigned(buf, sizeof(buf), dedup_skipped_size, "B", 1024);
373 printf(" %8s skipped\n", buf);
374 printf(" %8jd CRC collisions\n"
375 " %8jd SHA collisions\n"
376 " %8jd big-block underflows\n"
377 " %8jd new dedup records\n"
378 " %8jd new dedup bytes\n",
379 (intmax_t)dedup_crc_failures,
380 (intmax_t)dedup_sha_failures,
381 (intmax_t)dedup_underflows,
382 (intmax_t)dedup_successes_count,
383 (intmax_t)dedup_successes_bytes
386 /* Once completed remove cycle file */
387 hammer_reset_cycle();
389 /* We don't want to mess up with other directives */
398 count_btree_elm(hammer_btree_leaf_elm_t scan_leaf __unused, int flags __unused)
405 collect_btree_elm(hammer_btree_leaf_elm_t scan_leaf, int flags __unused)
407 struct sim_dedup_entry *sim_de;
410 * If we are using too much memory we have to clean some out, which
411 * will cause the run to use multiple passes. Be careful of integer
414 if (MemoryUse > MemoryLimit) {
415 DedupCrcEnd = DedupCrcStart +
416 (uint32_t)(DedupCrcEnd - DedupCrcStart - 1) / 2;
418 printf("memory limit crc-range %08x-%08x\n",
419 DedupCrcStart, DedupCrcEnd);
423 sim_de = RB_MAX(sim_dedup_entry_rb_tree,
425 if (sim_de == NULL || sim_de->crc < DedupCrcEnd)
427 RB_REMOVE(sim_dedup_entry_rb_tree,
428 &sim_dedup_tree, sim_de);
429 MemoryUse -= sizeof(*sim_de);
435 * Collect statistics based on the CRC only, do not try to read
436 * any data blocks or run SHA hashes.
438 sim_de = RB_LOOKUP(sim_dedup_entry_rb_tree, &sim_dedup_tree,
439 scan_leaf->data_crc);
441 if (sim_de == NULL) {
442 sim_de = calloc(1, sizeof(*sim_de));
443 sim_de->crc = scan_leaf->data_crc;
444 RB_INSERT(sim_dedup_entry_rb_tree, &sim_dedup_tree, sim_de);
445 MemoryUse += sizeof(*sim_de);
448 sim_de->ref_blks += 1;
449 sim_de->ref_size += scan_leaf->data_len;
455 validate_dedup_pair(hammer_btree_leaf_elm_t p, hammer_btree_leaf_elm_t q)
457 if (HAMMER_ZONE(p->data_offset) != HAMMER_ZONE(q->data_offset))
459 if (p->data_len != q->data_len)
465 #define DEDUP_TECH_FAILURE 1
466 #define DEDUP_CMP_FAILURE 2
467 #define DEDUP_INVALID_ZONE 3
468 #define DEDUP_UNDERFLOW 4
469 #define DEDUP_VERS_FAILURE 5
473 deduplicate(hammer_btree_leaf_elm_t p, hammer_btree_leaf_elm_t q)
475 struct hammer_ioc_dedup dedup;
477 bzero(&dedup, sizeof(dedup));
480 * If data_offset fields are the same there is no need to run ioctl,
481 * candidate is already dedup'ed.
483 if (p->data_offset == q->data_offset)
486 dedup.elm1 = p->base;
487 dedup.elm2 = q->base;
489 if (ioctl(glob_fd, HAMMERIOC_DEDUP, &dedup) < 0) {
490 if (errno == EOPNOTSUPP)
491 return (DEDUP_VERS_FAILURE); /* must be at least version 5 */
492 /* Technical failure - locking or w/e */
493 return (DEDUP_TECH_FAILURE);
495 if (dedup.head.flags & HAMMER_IOC_DEDUP_CMP_FAILURE)
496 return (DEDUP_CMP_FAILURE);
497 if (dedup.head.flags & HAMMER_IOC_DEDUP_INVALID_ZONE)
498 return (DEDUP_INVALID_ZONE);
499 if (dedup.head.flags & HAMMER_IOC_DEDUP_UNDERFLOW)
500 return (DEDUP_UNDERFLOW);
502 ++dedup_successes_count;
503 dedup_successes_bytes += p->data_len;
509 process_btree_elm(hammer_btree_leaf_elm_t scan_leaf, int flags)
511 struct dedup_entry *de;
512 struct sha_dedup_entry *sha_de, temp;
513 struct pass2_dedup_entry *pass2_de;
517 * If we are using too much memory we have to clean some out, which
518 * will cause the run to use multiple passes. Be careful of integer
521 while (MemoryUse > MemoryLimit) {
522 DedupCrcEnd = DedupCrcStart +
523 (uint32_t)(DedupCrcEnd - DedupCrcStart - 1) / 2;
525 printf("memory limit crc-range %08x-%08x\n",
526 DedupCrcStart, DedupCrcEnd);
531 de = RB_MAX(dedup_entry_rb_tree, &dedup_tree);
532 if (de == NULL || de->leaf.data_crc < DedupCrcEnd)
534 if (de->flags & HAMMER_DEDUP_ENTRY_FICTITIOUS) {
535 while ((sha_de = RB_ROOT(&de->u.fict_root)) !=
537 RB_REMOVE(sha_dedup_entry_rb_tree,
538 &de->u.fict_root, sha_de);
539 MemoryUse -= sizeof(*sha_de);
543 RB_REMOVE(dedup_entry_rb_tree, &dedup_tree, de);
544 MemoryUse -= sizeof(*de);
550 * Collect statistics based on the CRC. Colliding CRCs usually
551 * cause a SHA sub-tree to be created under the de.
553 * Trivial case if de not found.
555 de = RB_LOOKUP(dedup_entry_rb_tree, &dedup_tree, scan_leaf->data_crc);
557 de = calloc(1, sizeof(*de));
558 de->leaf = *scan_leaf;
559 RB_INSERT(dedup_entry_rb_tree, &dedup_tree, de);
560 MemoryUse += sizeof(*de);
565 * Found entry in CRC tree
567 if (de->flags & HAMMER_DEDUP_ENTRY_FICTITIOUS) {
569 * Optimize the case where a CRC failure results in multiple
570 * SHA entries. If we unconditionally issue a data-read a
571 * degenerate situation where a colliding CRC's second SHA
572 * entry contains the lion's share of the deduplication
573 * candidates will result in excessive data block reads.
575 * Deal with the degenerate case by looking for a matching
576 * data_offset/data_len in the SHA elements we already have
577 * before reading the data block and generating a new SHA.
579 RB_FOREACH(sha_de, sha_dedup_entry_rb_tree, &de->u.fict_root) {
580 if (sha_de->leaf.data_offset ==
581 scan_leaf->data_offset &&
582 sha_de->leaf.data_len == scan_leaf->data_len) {
583 memcpy(temp.sha_hash, sha_de->sha_hash,
584 SHA256_DIGEST_LENGTH);
590 * Entry in CRC tree is fictitious, so we already had problems
591 * with this CRC. Upgrade (compute SHA) the candidate and
592 * dive into SHA subtree. If upgrade fails insert the candidate
593 * into Pass2 list (it will be processed later).
595 if (sha_de == NULL) {
596 if (upgrade_chksum(scan_leaf, temp.sha_hash))
599 sha_de = RB_FIND(sha_dedup_entry_rb_tree,
600 &de->u.fict_root, &temp);
604 * Nothing in SHA subtree so far, so this is a new
605 * 'dataset'. Insert new entry into SHA subtree.
607 if (sha_de == NULL) {
608 sha_de = calloc(1, sizeof(*sha_de));
609 sha_de->leaf = *scan_leaf;
610 memcpy(sha_de->sha_hash, temp.sha_hash,
611 SHA256_DIGEST_LENGTH);
612 RB_INSERT(sha_dedup_entry_rb_tree, &de->u.fict_root,
614 MemoryUse += sizeof(*sha_de);
615 goto upgrade_stats_sha;
619 * Found entry in SHA subtree, it means we have a potential
620 * dedup pair. Validate it (zones have to match and data_len
621 * field have to be the same too. If validation fails, treat
622 * it as a SHA collision (jump to sha256_failure).
624 if (validate_dedup_pair(&sha_de->leaf, scan_leaf))
628 * We have a valid dedup pair (SHA match, validated).
630 * In case of technical failure (dedup pair was good, but
631 * ioctl failed anyways) insert the candidate into Pass2 list
632 * (we will try to dedup it after we are done with the rest of
635 * If ioctl fails because either of blocks is in the non-dedup
636 * zone (we can dedup only in LARGE_DATA and SMALL_DATA) don't
637 * bother with the candidate and terminate early.
639 * If ioctl fails because of big-block underflow replace the
640 * leaf node that found dedup entry represents with scan_leaf.
642 error = deduplicate(&sha_de->leaf, scan_leaf);
645 goto upgrade_stats_sha;
646 case DEDUP_TECH_FAILURE:
648 case DEDUP_CMP_FAILURE:
650 case DEDUP_INVALID_ZONE:
651 goto terminate_early;
652 case DEDUP_UNDERFLOW:
654 sha_de->leaf = *scan_leaf;
655 memcpy(sha_de->sha_hash, temp.sha_hash,
656 SHA256_DIGEST_LENGTH);
657 goto upgrade_stats_sha;
658 case DEDUP_VERS_FAILURE:
659 errx(1, "HAMMER filesystem must be at least "
660 "version 5 to dedup");
663 fprintf(stderr, "Unknown error\n");
664 goto terminate_early;
668 * Ooh la la.. SHA-256 collision. Terminate early, there's
669 * nothing we can do here.
672 ++dedup_sha_failures;
673 goto terminate_early;
676 * Candidate CRC is good for now (we found an entry in CRC
677 * tree and it's not fictitious). This means we have a
678 * potential dedup pair.
680 if (validate_dedup_pair(&de->leaf, scan_leaf))
684 * We have a valid dedup pair (CRC match, validated)
686 error = deduplicate(&de->leaf, scan_leaf);
690 case DEDUP_TECH_FAILURE:
692 case DEDUP_CMP_FAILURE:
694 case DEDUP_INVALID_ZONE:
695 goto terminate_early;
696 case DEDUP_UNDERFLOW:
698 de->leaf = *scan_leaf;
700 case DEDUP_VERS_FAILURE:
701 errx(1, "HAMMER filesystem must be at least "
702 "version 5 to dedup");
705 fprintf(stderr, "Unknown error\n");
706 goto terminate_early;
711 * We got a CRC collision - either ioctl failed because of
712 * the comparison failure or validation of the potential
713 * dedup pair went bad. In all cases insert both blocks
714 * into SHA subtree (this requires checksum upgrade) and mark
715 * entry that corresponds to this CRC in the CRC tree
716 * fictitious, so that all futher operations with this CRC go
717 * through SHA subtree.
719 ++dedup_crc_failures;
722 * Insert block that was represented by now fictitious dedup
723 * entry (create a new SHA entry and preserve stats of the
724 * old CRC one). If checksum upgrade fails insert the
725 * candidate into Pass2 list and return - keep both trees
728 sha_de = calloc(1, sizeof(*sha_de));
729 sha_de->leaf = de->leaf;
730 sha_de->ref_blks = de->u.de.ref_blks;
731 sha_de->ref_size = de->u.de.ref_size;
732 if (upgrade_chksum(&sha_de->leaf, sha_de->sha_hash)) {
736 MemoryUse += sizeof(*sha_de);
738 RB_INIT(&de->u.fict_root);
740 * Here we can insert without prior checking because the tree
741 * is empty at this point
743 RB_INSERT(sha_dedup_entry_rb_tree, &de->u.fict_root, sha_de);
746 * Mark entry in CRC tree fictitious
748 de->flags |= HAMMER_DEDUP_ENTRY_FICTITIOUS;
751 * Upgrade checksum of the candidate and insert it into
752 * SHA subtree. If upgrade fails insert the candidate into
755 if (upgrade_chksum(scan_leaf, temp.sha_hash))
757 sha_de = RB_FIND(sha_dedup_entry_rb_tree, &de->u.fict_root,
760 /* There is an entry with this SHA already, but the only
761 * RB-tree element at this point is that entry we just
762 * added. We know for sure these blocks are different
763 * (this is crc_failure branch) so treat it as SHA
768 sha_de = calloc(1, sizeof(*sha_de));
769 sha_de->leaf = *scan_leaf;
770 memcpy(sha_de->sha_hash, temp.sha_hash, SHA256_DIGEST_LENGTH);
771 RB_INSERT(sha_dedup_entry_rb_tree, &de->u.fict_root, sha_de);
772 MemoryUse += sizeof(*sha_de);
773 goto upgrade_stats_sha;
777 de->u.de.ref_blks += 1;
778 de->u.de.ref_size += scan_leaf->data_len;
782 sha_de->ref_blks += 1;
783 sha_de->ref_size += scan_leaf->data_len;
788 * If in pass2 mode don't insert anything, fall through to
791 if ((flags & DEDUP_PASS2) == 0) {
792 pass2_de = calloc(1, sizeof(*pass2_de));
793 pass2_de->leaf = *scan_leaf;
794 STAILQ_INSERT_TAIL(&pass2_dedup_queue, pass2_de, sq_entry);
795 dedup_skipped_size += scan_leaf->data_len;
801 * Early termination path. Fixup stats.
803 dedup_alloc_size += scan_leaf->data_len;
804 dedup_ref_size += scan_leaf->data_len;
810 upgrade_chksum(hammer_btree_leaf_elm_t leaf, uint8_t *sha_hash)
812 struct hammer_ioc_data data;
813 char *buf = malloc(DEDUP_BUF);
817 bzero(&data, sizeof(data));
818 data.elm = leaf->base;
820 data.size = DEDUP_BUF;
823 if (ioctl(glob_fd, HAMMERIOC_GET_DATA, &data) < 0) {
824 fprintf(stderr, "Get-data failed: %s\n", strerror(errno));
828 DedupDataReads += leaf->data_len;
830 if (data.leaf.data_len != leaf->data_len) {
835 if (data.leaf.base.btype == HAMMER_BTREE_TYPE_RECORD &&
836 data.leaf.base.rec_type == HAMMER_RECTYPE_DATA) {
838 SHA256_Update(&ctx, (void *)buf, data.leaf.data_len);
839 SHA256_Final(sha_hash, &ctx);
849 sigAlrm(int signo __unused)
856 sigInfo(int signo __unused)
863 scan_pfs(char *filesystem, scan_pfs_cb_t func, const char *id)
865 struct hammer_ioc_mirror_rw mirror;
866 hammer_ioc_mrecord_any_t mrec;
867 struct hammer_btree_leaf_elm elm;
868 char *buf = malloc(DEDUP_BUF);
875 DedupCurrentRecords = 0;
876 signal(SIGINFO, sigInfo);
877 signal(SIGALRM, sigAlrm);
880 * Deduplication happens per element so hammer(8) is in full
881 * control of the ioctl()s to actually perform it. SIGALRM
882 * needs to be handled within hammer(8) but a checkpoint
883 * is needed for resuming. Use cycle file for that.
885 * Try to obtain the previous obj_id from the cycle file and
886 * if not available just start from the beginning.
888 bzero(&mirror, sizeof(mirror));
889 hammer_key_beg_init(&mirror.key_beg);
890 hammer_get_cycle(&mirror.key_beg, &mirror.tid_beg);
892 if (mirror.key_beg.obj_id != (int64_t)HAMMER_MIN_OBJID) {
894 fprintf(stderr, "%s: mirror-read: Resuming at object %016jx\n",
895 id, (uintmax_t)mirror.key_beg.obj_id);
898 hammer_key_end_init(&mirror.key_end);
900 mirror.tid_beg = glob_pfs.ondisk->sync_beg_tid;
901 mirror.tid_end = glob_pfs.ondisk->sync_end_tid;
902 mirror.head.flags |= HAMMER_IOC_MIRROR_NODATA; /* we want only keys */
904 mirror.size = DEDUP_BUF;
905 mirror.pfs_id = glob_pfs.pfs_id;
906 mirror.shared_uuid = glob_pfs.ondisk->shared_uuid;
908 if (VerboseOpt && DedupCrcStart == 0) {
909 printf("%s %s: objspace %016jx:%04x %016jx:%04x\n",
911 (uintmax_t)mirror.key_beg.obj_id,
912 mirror.key_beg.localization,
913 (uintmax_t)mirror.key_end.obj_id,
914 mirror.key_end.localization);
915 printf("%s %s: pfs_id %d\n",
916 id, filesystem, glob_pfs.pfs_id);
923 mirror.pfs_id = glob_pfs.pfs_id;
924 mirror.shared_uuid = glob_pfs.ondisk->shared_uuid;
925 if (ioctl(glob_fd, HAMMERIOC_MIRROR_READ, &mirror) < 0) {
926 err(1, "Mirror-read %s failed", filesystem);
929 if (mirror.head.flags & HAMMER_IOC_HEAD_ERROR) {
930 errx(1, "Mirror-read %s fatal error %d",
931 filesystem, mirror.head.error);
936 while (offset < mirror.count) {
937 mrec = (void *)((char *)buf + offset);
938 bytes = HAMMER_HEAD_DOALIGN(mrec->head.rec_size);
939 if (offset + bytes > mirror.count) {
940 errx(1, "Misaligned record");
943 assert((mrec->head.type &
944 HAMMER_MRECF_TYPE_MASK) ==
945 HAMMER_MREC_TYPE_REC);
947 elm = mrec->rec.leaf;
948 if (elm.base.btype != HAMMER_BTREE_TYPE_RECORD)
950 if (elm.base.rec_type != HAMMER_RECTYPE_DATA)
952 ++DedupCurrentRecords;
953 if (DedupCrcStart != DedupCrcEnd) {
954 if (elm.data_crc < DedupCrcStart)
957 elm.data_crc >= DedupCrcEnd) {
964 mirror.key_beg = mirror.key_cur;
965 if (DidInterrupt || SigAlrmFlag) {
967 fprintf(stderr, "%s\n",
968 (DidInterrupt ? "Interrupted" : "Timeout"));
969 hammer_set_cycle(&mirror.key_cur, mirror.tid_beg);
971 fprintf(stderr, "Cyclefile %s updated for "
972 "continuation\n", CyclePath);
976 if (DedupTotalRecords) {
977 humanize_unsigned(buf1, sizeof(buf1),
980 humanize_unsigned(buf2, sizeof(buf2),
981 dedup_successes_bytes,
983 fprintf(stderr, "%s count %7jd/%jd "
985 "ioread %s newddup %s\n",
987 (intmax_t)DedupCurrentRecords,
988 (intmax_t)DedupTotalRecords,
989 (int)(DedupCurrentRecords * 100 /
991 (int)(DedupCurrentRecords * 10000 /
992 DedupTotalRecords % 100),
995 fprintf(stderr, "%s count %-7jd\n",
997 (intmax_t)DedupCurrentRecords);
1001 } while (mirror.count != 0);
1003 signal(SIGINFO, SIG_IGN);
1004 signal(SIGALRM, SIG_IGN);
1011 dump_simulated_dedup(void)
1013 struct sim_dedup_entry *sim_de;
1015 printf("=== Dumping simulated dedup entries:\n");
1016 RB_FOREACH(sim_de, sim_dedup_entry_rb_tree, &sim_dedup_tree) {
1017 printf("\tcrc=%08x cnt=%ju size=%ju\n",
1019 (intmax_t)sim_de->ref_blks,
1020 (intmax_t)sim_de->ref_size);
1022 printf("end of dump ===\n");
1027 dump_real_dedup(void)
1029 struct dedup_entry *de;
1030 struct sha_dedup_entry *sha_de;
1033 printf("=== Dumping dedup entries:\n");
1034 RB_FOREACH(de, dedup_entry_rb_tree, &dedup_tree) {
1035 if (de->flags & HAMMER_DEDUP_ENTRY_FICTITIOUS) {
1036 printf("\tcrc=%08x fictitious\n", de->leaf.data_crc);
1038 RB_FOREACH(sha_de, sha_dedup_entry_rb_tree, &de->u.fict_root) {
1039 printf("\t\tcrc=%08x cnt=%ju size=%ju\n\t"
1041 sha_de->leaf.data_crc,
1042 (intmax_t)sha_de->ref_blks,
1043 (intmax_t)sha_de->ref_size);
1044 for (i = 0; i < SHA256_DIGEST_LENGTH; ++i)
1045 printf("%02x", sha_de->sha_hash[i]);
1049 printf("\tcrc=%08x cnt=%ju size=%ju\n",
1051 (intmax_t)de->u.de.ref_blks,
1052 (intmax_t)de->u.de.ref_size);
1055 printf("end of dump ===\n");
1060 dedup_usage(int code)
1063 "hammer dedup-simulate <filesystem>\n"
1064 "hammer dedup <filesystem>\n"