kernel - Expand p_nice effect more, fix %cpu calculation bug
[dragonfly.git] / sbin / hammer / cmd_dedup.c
1 /*
2  * Copyright (c) 2010 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Ilya Dryomov <idryomov@gmail.com>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
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
16  *    distribution.
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.
20  *
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
32  * SUCH DAMAGE.
33  */
34
35 #include <libutil.h>
36 #include <crypto/sha2/sha2.h>
37
38 #include "hammer.h"
39
40 #define DEDUP_BUF (64 * 1024)
41
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);
48
49 struct sim_dedup_entry {
50         hammer_crc_t    crc;
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;
54 };
55
56 struct dedup_entry {
57         struct hammer_btree_leaf_elm leaf;
58         union {
59                 struct {
60                         uint64_t ref_blks;
61                         uint64_t ref_size;
62                 } de;
63                 RB_HEAD(sha_dedup_entry_rb_tree, sha_dedup_entry) fict_root;
64         } u;
65         uint8_t flags;
66         RB_ENTRY(dedup_entry) rb_entry;
67 };
68
69 #define HAMMER_DEDUP_ENTRY_FICTITIOUS   0x0001
70
71 struct sha_dedup_entry {
72         struct hammer_btree_leaf_elm    leaf;
73         uint64_t                        ref_blks;
74         uint64_t                        ref_size;
75         uint8_t                         sha_hash[SHA256_DIGEST_LENGTH];
76         RB_ENTRY(sha_dedup_entry)       fict_entry;
77 };
78
79 /* Sorted list of HAMMER B-Tree keys */
80 struct dedup_entry_rb_tree;
81 struct sha_dedup_entry_rb_tree;
82
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);
87
88 RB_PROTOTYPE(sha_dedup_entry_rb_tree, sha_dedup_entry, fict_entry,
89                 rb_sha_dedup_entry_compare);
90
91 /*
92  * Pass2 list - contains entries that were not dedup'ed because ioctl failed
93  */
94 STAILQ_HEAD(, pass2_dedup_entry) pass2_dedup_queue =
95                                 STAILQ_HEAD_INITIALIZER(pass2_dedup_queue);
96
97 struct pass2_dedup_entry {
98         struct hammer_btree_leaf_elm    leaf;
99         STAILQ_ENTRY(pass2_dedup_entry) sq_entry;
100 };
101
102 #define DEDUP_PASS2     0x0001 /* process_btree_elm() mode */
103
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;
112
113 /* PFS global ids - we deal with just one PFS at a run */
114 static int glob_fd;
115 static struct hammer_ioc_pseudofs_rw glob_pfs;
116
117 /*
118  * Global accounting variables
119  *
120  * Last three don't have to be 64-bit, just to be safe..
121  */
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;
130
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);
146
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);
153
154 static int
155 rb_sim_dedup_entry_compare(struct sim_dedup_entry *sim_de1,
156                         struct sim_dedup_entry *sim_de2)
157 {
158         if (sim_de1->crc < sim_de2->crc)
159                 return (-1);
160         if (sim_de1->crc > sim_de2->crc)
161                 return (1);
162
163         return (0);
164 }
165
166 static int
167 rb_dedup_entry_compare(struct dedup_entry *de1, struct dedup_entry *de2)
168 {
169         if (de1->leaf.data_crc < de2->leaf.data_crc)
170                 return (-1);
171         if (de1->leaf.data_crc > de2->leaf.data_crc)
172                 return (1);
173
174         return (0);
175 }
176
177 static int
178 rb_sha_dedup_entry_compare(struct sha_dedup_entry *sha_de1,
179                         struct sha_dedup_entry *sha_de2)
180 {
181         unsigned long *h1 = (unsigned long *)&sha_de1->sha_hash;
182         unsigned long *h2 = (unsigned long *)&sha_de2->sha_hash;
183         int i;
184
185         for (i = 0; i < SHA256_DIGEST_LENGTH / (int)sizeof(unsigned long); ++i) {
186                 if (h1[i] < h2[i])
187                         return (-1);
188                 if (h1[i] > h2[i])
189                         return (1);
190         }
191
192         return (0);
193 }
194
195 /*
196  * dedup-simulate <filesystem>
197  */
198 void
199 hammer_cmd_dedup_simulate(char **av, int ac)
200 {
201         struct sim_dedup_entry *sim_de;
202
203         if (ac != 1)
204                 dedup_usage(1);
205
206         glob_fd = getpfs(&glob_pfs, av[0]);
207
208         /*
209          * Collection passes (memory limited)
210          */
211         printf("Dedup-simulate running\n");
212         do {
213                 DedupCrcStart = DedupCrcEnd;
214                 DedupCrcEnd = 0;
215                 MemoryUse = 0;
216
217                 if (VerboseOpt) {
218                         printf("B-Tree pass  crc-range %08x-max\n",
219                                 DedupCrcStart);
220                         fflush(stdout);
221                 }
222                 scan_pfs(av[0], collect_btree_elm, "simu-pass");
223
224                 if (VerboseOpt >= 2)
225                         dump_simulated_dedup();
226
227                 /*
228                  * Calculate simulated dedup ratio and get rid of the tree
229                  */
230                 while ((sim_de = RB_ROOT(&sim_dedup_tree)) != NULL) {
231                         assert(sim_de->ref_blks != 0);
232                         dedup_ref_size += sim_de->ref_size;
233                         dedup_alloc_size += sim_de->ref_size / sim_de->ref_blks;
234
235                         RB_REMOVE(sim_dedup_entry_rb_tree, &sim_dedup_tree, sim_de);
236                         free(sim_de);
237                 }
238                 if (DedupCrcEnd && VerboseOpt == 0)
239                         printf(".");
240         } while (DedupCrcEnd);
241
242         printf("Dedup-simulate %s succeeded\n", av[0]);
243         relpfs(glob_fd, &glob_pfs);
244
245         printf("Simulated dedup ratio = %.2f\n",
246             (dedup_alloc_size != 0) ?
247                 (double)dedup_ref_size / dedup_alloc_size : 0);
248 }
249
250 /*
251  * dedup <filesystem>
252  */
253 void
254 hammer_cmd_dedup(char **av, int ac)
255 {
256         struct dedup_entry *de;
257         struct sha_dedup_entry *sha_de;
258         struct pass2_dedup_entry *pass2_de;
259         char *tmp;
260         char buf[8];
261         int needfree = 0;
262
263         if (TimeoutOpt > 0)
264                 alarm(TimeoutOpt);
265
266         if (ac != 1)
267                 dedup_usage(1);
268
269         STAILQ_INIT(&pass2_dedup_queue);
270
271         glob_fd = getpfs(&glob_pfs, av[0]);
272
273         /*
274          * A cycle file is _required_ for resuming dedup after the timeout
275          * specified with -t has expired. If no -c option, then place a
276          * .dedup.cycle file either in the PFS snapshots directory or in
277          * the default snapshots directory.
278          */
279         if (!CyclePath) {
280                 if (glob_pfs.ondisk->snapshots[0] != '/')
281                         asprintf(&tmp, "%s/%s/.dedup.cycle",
282                             SNAPSHOTS_BASE, av[0]);
283                 else
284                         asprintf(&tmp, "%s/.dedup.cycle",
285                             glob_pfs.ondisk->snapshots);
286                 CyclePath = tmp;
287                 needfree = 1;
288         }
289
290         /*
291          * Pre-pass to cache the btree
292          */
293         scan_pfs(av[0], count_btree_elm, "pre-pass ");
294         DedupTotalRecords = DedupCurrentRecords;
295
296         /*
297          * Collection passes (memory limited)
298          */
299         printf("Dedup running\n");
300         do {
301                 DedupCrcStart = DedupCrcEnd;
302                 DedupCrcEnd = 0;
303                 MemoryUse = 0;
304
305                 if (VerboseOpt) {
306                         printf("B-Tree pass  crc-range %08x-max\n",
307                                 DedupCrcStart);
308                         fflush(stdout);
309                 }
310                 scan_pfs(av[0], process_btree_elm, "main-pass");
311
312                 while ((pass2_de = STAILQ_FIRST(&pass2_dedup_queue)) != NULL) {
313                         if (process_btree_elm(&pass2_de->leaf, DEDUP_PASS2))
314                                 dedup_skipped_size -= pass2_de->leaf.data_len;
315
316                         STAILQ_REMOVE_HEAD(&pass2_dedup_queue, sq_entry);
317                         free(pass2_de);
318                 }
319                 assert(STAILQ_EMPTY(&pass2_dedup_queue));
320
321                 if (VerboseOpt >= 2)
322                         dump_real_dedup();
323
324                 /*
325                  * Calculate dedup ratio and get rid of the trees
326                  */
327                 while ((de = RB_ROOT(&dedup_tree)) != NULL) {
328                         if (de->flags & HAMMER_DEDUP_ENTRY_FICTITIOUS) {
329                                 while ((sha_de = RB_ROOT(&de->u.fict_root)) != NULL) {
330                                         assert(sha_de->ref_blks != 0);
331                                         dedup_ref_size += sha_de->ref_size;
332                                         dedup_alloc_size += sha_de->ref_size / sha_de->ref_blks;
333
334                                         RB_REMOVE(sha_dedup_entry_rb_tree,
335                                                         &de->u.fict_root, sha_de);
336                                         free(sha_de);
337                                 }
338                                 assert(RB_EMPTY(&de->u.fict_root));
339                         } else {
340                                 assert(de->u.de.ref_blks != 0);
341                                 dedup_ref_size += de->u.de.ref_size;
342                                 dedup_alloc_size += de->u.de.ref_size / de->u.de.ref_blks;
343                         }
344
345                         RB_REMOVE(dedup_entry_rb_tree, &dedup_tree, de);
346                         free(de);
347                 }
348                 assert(RB_EMPTY(&dedup_tree));
349                 if (DedupCrcEnd && VerboseOpt == 0)
350                         printf(".");
351         } while (DedupCrcEnd);
352
353         printf("Dedup %s succeeded\n", av[0]);
354         relpfs(glob_fd, &glob_pfs);
355
356         humanize_unsigned(buf, sizeof(buf), dedup_ref_size, "B", 1024);
357         printf("Dedup ratio = %.2f (in this run)\n"
358                "    %8s referenced\n",
359                ((dedup_alloc_size != 0) ?
360                         (double)dedup_ref_size / dedup_alloc_size : 0),
361                buf
362         );
363         humanize_unsigned(buf, sizeof(buf), dedup_alloc_size, "B", 1024);
364         printf("    %8s allocated\n", buf);
365         humanize_unsigned(buf, sizeof(buf), dedup_skipped_size, "B", 1024);
366         printf("    %8s skipped\n", buf);
367         printf("    %8jd CRC collisions\n"
368                "    %8jd SHA collisions\n"
369                "    %8jd big-block underflows\n"
370                "    %8jd new dedup records\n"
371                "    %8jd new dedup bytes\n",
372                (intmax_t)dedup_crc_failures,
373                (intmax_t)dedup_sha_failures,
374                (intmax_t)dedup_underflows,
375                (intmax_t)dedup_successes_count,
376                (intmax_t)dedup_successes_bytes
377         );
378
379         /* Once completed remove cycle file */
380         hammer_reset_cycle();
381
382         /* We don't want to mess up with other directives */
383         if (needfree) {
384                 free(tmp);
385                 CyclePath = NULL;
386         }
387 }
388
389 static int
390 count_btree_elm(hammer_btree_leaf_elm_t scan_leaf __unused, int flags __unused)
391 {
392         return(1);
393 }
394
395 static int
396 collect_btree_elm(hammer_btree_leaf_elm_t scan_leaf, int flags __unused)
397 {
398         struct sim_dedup_entry *sim_de;
399
400         /*
401          * If we are using too much memory we have to clean some out, which
402          * will cause the run to use multiple passes.  Be careful of integer
403          * overflows!
404          */
405         if (MemoryUse > MemoryLimit) {
406                 DedupCrcEnd = DedupCrcStart +
407                               (uint32_t)(DedupCrcEnd - DedupCrcStart - 1) / 2;
408                 if (VerboseOpt) {
409                         printf("memory limit  crc-range %08x-%08x\n",
410                                 DedupCrcStart, DedupCrcEnd);
411                         fflush(stdout);
412                 }
413                 for (;;) {
414                         sim_de = RB_MAX(sim_dedup_entry_rb_tree,
415                                         &sim_dedup_tree);
416                         if (sim_de == NULL || sim_de->crc < DedupCrcEnd)
417                                 break;
418                         RB_REMOVE(sim_dedup_entry_rb_tree,
419                                   &sim_dedup_tree, sim_de);
420                         MemoryUse -= sizeof(*sim_de);
421                         free(sim_de);
422                 }
423         }
424
425         /*
426          * Collect statistics based on the CRC only, do not try to read
427          * any data blocks or run SHA hashes.
428          */
429         sim_de = RB_LOOKUP(sim_dedup_entry_rb_tree, &sim_dedup_tree,
430                            scan_leaf->data_crc);
431
432         if (sim_de == NULL) {
433                 sim_de = calloc(1, sizeof(*sim_de));
434                 sim_de->crc = scan_leaf->data_crc;
435                 RB_INSERT(sim_dedup_entry_rb_tree, &sim_dedup_tree, sim_de);
436                 MemoryUse += sizeof(*sim_de);
437         }
438
439         sim_de->ref_blks += 1;
440         sim_de->ref_size += scan_leaf->data_len;
441         return (1);
442 }
443
444 static __inline int
445 validate_dedup_pair(hammer_btree_leaf_elm_t p, hammer_btree_leaf_elm_t q)
446 {
447         if (HAMMER_ZONE(p->data_offset) != HAMMER_ZONE(q->data_offset)) {
448                 return (1);
449         }
450         if (p->data_len != q->data_len) {
451                 return (1);
452         }
453
454         return (0);
455 }
456
457 #define DEDUP_TECH_FAILURE      1
458 #define DEDUP_CMP_FAILURE       2
459 #define DEDUP_INVALID_ZONE      3
460 #define DEDUP_UNDERFLOW         4
461 #define DEDUP_VERS_FAILURE      5
462
463 static __inline int
464 deduplicate(hammer_btree_leaf_elm_t p, hammer_btree_leaf_elm_t q)
465 {
466         struct hammer_ioc_dedup dedup;
467
468         bzero(&dedup, sizeof(dedup));
469
470         /*
471          * If data_offset fields are the same there is no need to run ioctl,
472          * candidate is already dedup'ed.
473          */
474         if (p->data_offset == q->data_offset) {
475                 return (0);
476         }
477
478         dedup.elm1 = p->base;
479         dedup.elm2 = q->base;
480         RunningIoctl = 1;
481         if (ioctl(glob_fd, HAMMERIOC_DEDUP, &dedup) < 0) {
482                 if (errno == EOPNOTSUPP) {
483                         /* must be at least version 5 */
484                         return (DEDUP_VERS_FAILURE);
485                 }
486                 /* Technical failure - locking or w/e */
487                 return (DEDUP_TECH_FAILURE);
488         }
489         if (dedup.head.flags & HAMMER_IOC_DEDUP_CMP_FAILURE) {
490                 return (DEDUP_CMP_FAILURE);
491         }
492         if (dedup.head.flags & HAMMER_IOC_DEDUP_INVALID_ZONE) {
493                 return (DEDUP_INVALID_ZONE);
494         }
495         if (dedup.head.flags & HAMMER_IOC_DEDUP_UNDERFLOW) {
496                 return (DEDUP_UNDERFLOW);
497         }
498         RunningIoctl = 0;
499         ++dedup_successes_count;
500         dedup_successes_bytes += p->data_len;
501         return (0);
502 }
503
504 static int
505 process_btree_elm(hammer_btree_leaf_elm_t scan_leaf, int flags)
506 {
507         struct dedup_entry *de;
508         struct sha_dedup_entry *sha_de, temp;
509         struct pass2_dedup_entry *pass2_de;
510         int error;
511
512         /*
513          * If we are using too much memory we have to clean some out, which
514          * will cause the run to use multiple passes.  Be careful of integer
515          * overflows!
516          */
517         while (MemoryUse > MemoryLimit) {
518                 DedupCrcEnd = DedupCrcStart +
519                               (uint32_t)(DedupCrcEnd - DedupCrcStart - 1) / 2;
520                 if (VerboseOpt) {
521                         printf("memory limit  crc-range %08x-%08x\n",
522                                 DedupCrcStart, DedupCrcEnd);
523                         fflush(stdout);
524                 }
525
526                 for (;;) {
527                         de = RB_MAX(dedup_entry_rb_tree, &dedup_tree);
528                         if (de == NULL || de->leaf.data_crc < DedupCrcEnd)
529                                 break;
530                         if (de->flags & HAMMER_DEDUP_ENTRY_FICTITIOUS) {
531                                 while ((sha_de = RB_ROOT(&de->u.fict_root)) !=
532                                        NULL) {
533                                         RB_REMOVE(sha_dedup_entry_rb_tree,
534                                                   &de->u.fict_root, sha_de);
535                                         MemoryUse -= sizeof(*sha_de);
536                                         free(sha_de);
537                                 }
538                         }
539                         RB_REMOVE(dedup_entry_rb_tree, &dedup_tree, de);
540                         MemoryUse -= sizeof(*de);
541                         free(de);
542                 }
543         }
544
545         /*
546          * Collect statistics based on the CRC.  Colliding CRCs usually
547          * cause a SHA sub-tree to be created under the de.
548          *
549          * Trivial case if de not found.
550          */
551         de = RB_LOOKUP(dedup_entry_rb_tree, &dedup_tree, scan_leaf->data_crc);
552         if (de == NULL) {
553                 de = calloc(1, sizeof(*de));
554                 de->leaf = *scan_leaf;
555                 RB_INSERT(dedup_entry_rb_tree, &dedup_tree, de);
556                 MemoryUse += sizeof(*de);
557                 goto upgrade_stats;
558         }
559
560         /*
561          * Found entry in CRC tree
562          */
563         if (de->flags & HAMMER_DEDUP_ENTRY_FICTITIOUS) {
564                 /*
565                  * Optimize the case where a CRC failure results in multiple
566                  * SHA entries.  If we unconditionally issue a data-read a
567                  * degenerate situation where a colliding CRC's second SHA
568                  * entry contains the lion's share of the deduplication
569                  * candidates will result in excessive data block reads.
570                  *
571                  * Deal with the degenerate case by looking for a matching
572                  * data_offset/data_len in the SHA elements we already have
573                  * before reading the data block and generating a new SHA.
574                  */
575                 RB_FOREACH(sha_de, sha_dedup_entry_rb_tree, &de->u.fict_root) {
576                         if (sha_de->leaf.data_offset ==
577                                                 scan_leaf->data_offset &&
578                             sha_de->leaf.data_len == scan_leaf->data_len) {
579                                 memcpy(temp.sha_hash, sha_de->sha_hash,
580                                         SHA256_DIGEST_LENGTH);
581                                 break;
582                         }
583                 }
584
585                 /*
586                  * Entry in CRC tree is fictitious, so we already had problems
587                  * with this CRC. Upgrade (compute SHA) the candidate and
588                  * dive into SHA subtree. If upgrade fails insert the candidate
589                  * into Pass2 list (it will be processed later).
590                  */
591                 if (sha_de == NULL) {
592                         if (upgrade_chksum(scan_leaf, temp.sha_hash))
593                                 goto pass2_insert;
594
595                         sha_de = RB_FIND(sha_dedup_entry_rb_tree,
596                                          &de->u.fict_root, &temp);
597                 }
598
599                 /*
600                  * Nothing in SHA subtree so far, so this is a new
601                  * 'dataset'. Insert new entry into SHA subtree.
602                  */
603                 if (sha_de == NULL) {
604                         sha_de = calloc(1, sizeof(*sha_de));
605                         sha_de->leaf = *scan_leaf;
606                         memcpy(sha_de->sha_hash, temp.sha_hash,
607                                SHA256_DIGEST_LENGTH);
608                         RB_INSERT(sha_dedup_entry_rb_tree, &de->u.fict_root,
609                                   sha_de);
610                         MemoryUse += sizeof(*sha_de);
611                         goto upgrade_stats_sha;
612                 }
613
614                 /*
615                  * Found entry in SHA subtree, it means we have a potential
616                  * dedup pair. Validate it (zones have to match and data_len
617                  * field have to be the same too. If validation fails, treat
618                  * it as a SHA collision (jump to sha256_failure).
619                  */
620                 if (validate_dedup_pair(&sha_de->leaf, scan_leaf))
621                         goto sha256_failure;
622
623                 /*
624                  * We have a valid dedup pair (SHA match, validated).
625                  *
626                  * In case of technical failure (dedup pair was good, but
627                  * ioctl failed anyways) insert the candidate into Pass2 list
628                  * (we will try to dedup it after we are done with the rest of
629                  * the tree).
630                  *
631                  * If ioctl fails because either of blocks is in the non-dedup
632                  * zone (we can dedup only in LARGE_DATA and SMALL_DATA) don't
633                  * bother with the candidate and terminate early.
634                  *
635                  * If ioctl fails because of big-block underflow replace the
636                  * leaf node that found dedup entry represents with scan_leaf.
637                  */
638                 error = deduplicate(&sha_de->leaf, scan_leaf);
639                 switch(error) {
640                 case 0:
641                         goto upgrade_stats_sha;
642                 case DEDUP_TECH_FAILURE:
643                         goto pass2_insert;
644                 case DEDUP_CMP_FAILURE:
645                         goto sha256_failure;
646                 case DEDUP_INVALID_ZONE:
647                         goto terminate_early;
648                 case DEDUP_UNDERFLOW:
649                         ++dedup_underflows;
650                         sha_de->leaf = *scan_leaf;
651                         memcpy(sha_de->sha_hash, temp.sha_hash,
652                                 SHA256_DIGEST_LENGTH);
653                         goto upgrade_stats_sha;
654                 case DEDUP_VERS_FAILURE:
655                         errx(1, "HAMMER filesystem must be at least "
656                                 "version 5 to dedup");
657                 default:
658                         fprintf(stderr, "Unknown error\n");
659                         goto terminate_early;
660                 }
661
662                 /*
663                  * Ooh la la.. SHA-256 collision. Terminate early, there's
664                  * nothing we can do here.
665                  */
666 sha256_failure:
667                 ++dedup_sha_failures;
668                 goto terminate_early;
669         } else {
670                 /*
671                  * Candidate CRC is good for now (we found an entry in CRC
672                  * tree and it's not fictitious). This means we have a
673                  * potential dedup pair.
674                  */
675                 if (validate_dedup_pair(&de->leaf, scan_leaf))
676                         goto crc_failure;
677
678                 /*
679                  * We have a valid dedup pair (CRC match, validated)
680                  */
681                 error = deduplicate(&de->leaf, scan_leaf);
682                 switch(error) {
683                 case 0:
684                         goto upgrade_stats;
685                 case DEDUP_TECH_FAILURE:
686                         goto pass2_insert;
687                 case DEDUP_CMP_FAILURE:
688                         goto crc_failure;
689                 case DEDUP_INVALID_ZONE:
690                         goto terminate_early;
691                 case DEDUP_UNDERFLOW:
692                         ++dedup_underflows;
693                         de->leaf = *scan_leaf;
694                         goto upgrade_stats;
695                 case DEDUP_VERS_FAILURE:
696                         errx(1, "HAMMER filesystem must be at least "
697                                 "version 5 to dedup");
698                 default:
699                         fprintf(stderr, "Unknown error\n");
700                         goto terminate_early;
701                 }
702
703 crc_failure:
704                 /*
705                  * We got a CRC collision - either ioctl failed because of
706                  * the comparison failure or validation of the potential
707                  * dedup pair went bad. In all cases insert both blocks
708                  * into SHA subtree (this requires checksum upgrade) and mark
709                  * entry that corresponds to this CRC in the CRC tree
710                  * fictitious, so that all futher operations with this CRC go
711                  * through SHA subtree.
712                  */
713                 ++dedup_crc_failures;
714
715                 /*
716                  * Insert block that was represented by now fictitious dedup
717                  * entry (create a new SHA entry and preserve stats of the
718                  * old CRC one). If checksum upgrade fails insert the
719                  * candidate into Pass2 list and return - keep both trees
720                  * unmodified.
721                  */
722                 sha_de = calloc(1, sizeof(*sha_de));
723                 sha_de->leaf = de->leaf;
724                 sha_de->ref_blks = de->u.de.ref_blks;
725                 sha_de->ref_size = de->u.de.ref_size;
726                 if (upgrade_chksum(&sha_de->leaf, sha_de->sha_hash)) {
727                         free(sha_de);
728                         goto pass2_insert;
729                 }
730                 MemoryUse += sizeof(*sha_de);
731
732                 RB_INIT(&de->u.fict_root);
733                 /*
734                  * Here we can insert without prior checking because the tree
735                  * is empty at this point
736                  */
737                 RB_INSERT(sha_dedup_entry_rb_tree, &de->u.fict_root, sha_de);
738
739                 /*
740                  * Mark entry in CRC tree fictitious
741                  */
742                 de->flags |= HAMMER_DEDUP_ENTRY_FICTITIOUS;
743
744                 /*
745                  * Upgrade checksum of the candidate and insert it into
746                  * SHA subtree. If upgrade fails insert the candidate into
747                  * Pass2 list.
748                  */
749                 if (upgrade_chksum(scan_leaf, temp.sha_hash)) {
750                         goto pass2_insert;
751                 }
752                 sha_de = RB_FIND(sha_dedup_entry_rb_tree, &de->u.fict_root,
753                                  &temp);
754                 if (sha_de != NULL)
755                         /* There is an entry with this SHA already, but the only
756                          * RB-tree element at this point is that entry we just
757                          * added. We know for sure these blocks are different
758                          * (this is crc_failure branch) so treat it as SHA
759                          * collision.
760                          */
761                         goto sha256_failure;
762
763                 sha_de = calloc(1, sizeof(*sha_de));
764                 sha_de->leaf = *scan_leaf;
765                 memcpy(sha_de->sha_hash, temp.sha_hash, SHA256_DIGEST_LENGTH);
766                 RB_INSERT(sha_dedup_entry_rb_tree, &de->u.fict_root, sha_de);
767                 MemoryUse += sizeof(*sha_de);
768                 goto upgrade_stats_sha;
769         }
770
771 upgrade_stats:
772         de->u.de.ref_blks += 1;
773         de->u.de.ref_size += scan_leaf->data_len;
774         return (1);
775
776 upgrade_stats_sha:
777         sha_de->ref_blks += 1;
778         sha_de->ref_size += scan_leaf->data_len;
779         return (1);
780
781 pass2_insert:
782         /*
783          * If in pass2 mode don't insert anything, fall through to
784          * terminate_early
785          */
786         if ((flags & DEDUP_PASS2) == 0) {
787                 pass2_de = calloc(1, sizeof(*pass2_de));
788                 pass2_de->leaf = *scan_leaf;
789                 STAILQ_INSERT_TAIL(&pass2_dedup_queue, pass2_de, sq_entry);
790                 dedup_skipped_size += scan_leaf->data_len;
791                 return (1);
792         }
793
794 terminate_early:
795         /*
796          * Early termination path. Fixup stats.
797          */
798         dedup_alloc_size += scan_leaf->data_len;
799         dedup_ref_size += scan_leaf->data_len;
800         return (0);
801 }
802
803 static int
804 upgrade_chksum(hammer_btree_leaf_elm_t leaf, uint8_t *sha_hash)
805 {
806         struct hammer_ioc_data data;
807         char *buf = malloc(DEDUP_BUF);
808         SHA256_CTX ctx;
809         int error;
810
811         bzero(&data, sizeof(data));
812         data.elm = leaf->base;
813         data.ubuf = buf;
814         data.size = DEDUP_BUF;
815
816         error = 0;
817         if (ioctl(glob_fd, HAMMERIOC_GET_DATA, &data) < 0) {
818                 fprintf(stderr, "Get-data failed: %s\n", strerror(errno));
819                 error = 1;
820                 goto done;
821         }
822         DedupDataReads += leaf->data_len;
823
824         if (data.leaf.data_len != leaf->data_len) {
825                 error = 1;
826                 goto done;
827         }
828
829         if (data.leaf.base.btype == HAMMER_BTREE_TYPE_RECORD &&
830             data.leaf.base.rec_type == HAMMER_RECTYPE_DATA) {
831                 SHA256_Init(&ctx);
832                 SHA256_Update(&ctx, (void *)buf, data.leaf.data_len);
833                 SHA256_Final(sha_hash, &ctx);
834         }
835
836 done:
837         free(buf);
838         return (error);
839 }
840
841 static void
842 sigAlrm(int signo __unused)
843 {
844         SigAlrmFlag = 1;
845 }
846
847 static void
848 sigInfo(int signo __unused)
849 {
850         SigInfoFlag = 1;
851 }
852
853 static void
854 scan_pfs(char *filesystem, scan_pfs_cb_t func, const char *id)
855 {
856         struct hammer_ioc_mirror_rw mirror;
857         hammer_ioc_mrecord_any_t mrec;
858         struct hammer_btree_leaf_elm elm;
859         char *buf = malloc(DEDUP_BUF);
860         char buf1[8];
861         char buf2[8];
862         int offset, bytes;
863
864         SigInfoFlag = 0;
865         DedupDataReads = 0;
866         DedupCurrentRecords = 0;
867         signal(SIGINFO, sigInfo);
868         signal(SIGALRM, sigAlrm);
869
870         /*
871          * Deduplication happens per element so hammer(8) is in full
872          * control of the ioctl()s to actually perform it. SIGALRM
873          * needs to be handled within hammer(8) but a checkpoint
874          * is needed for resuming. Use cycle file for that.
875          *
876          * Try to obtain the previous obj_id from the cycle file and
877          * if not available just start from the beginning.
878          */
879         bzero(&mirror, sizeof(mirror));
880         hammer_key_beg_init(&mirror.key_beg);
881         hammer_get_cycle(&mirror.key_beg, &mirror.tid_beg);
882
883         if (mirror.key_beg.obj_id != (int64_t)HAMMER_MIN_OBJID) {
884                 if (VerboseOpt)
885                         fprintf(stderr, "%s: mirror-read: Resuming at object %016jx\n",
886                             id, (uintmax_t)mirror.key_beg.obj_id);
887         }
888
889         hammer_key_end_init(&mirror.key_end);
890
891         mirror.tid_beg = glob_pfs.ondisk->sync_beg_tid;
892         mirror.tid_end = glob_pfs.ondisk->sync_end_tid;
893         mirror.head.flags |= HAMMER_IOC_MIRROR_NODATA; /* we want only keys */
894         mirror.ubuf = buf;
895         mirror.size = DEDUP_BUF;
896         mirror.pfs_id = glob_pfs.pfs_id;
897         mirror.shared_uuid = glob_pfs.ondisk->shared_uuid;
898
899         if (VerboseOpt && DedupCrcStart == 0) {
900                 printf("%s %s: objspace %016jx:%04x %016jx:%04x\n",
901                         id, filesystem,
902                         (uintmax_t)mirror.key_beg.obj_id,
903                         mirror.key_beg.localization,
904                         (uintmax_t)mirror.key_end.obj_id,
905                         mirror.key_end.localization);
906                 printf("%s %s: pfs_id %d\n",
907                         id, filesystem, glob_pfs.pfs_id);
908         }
909         fflush(stdout);
910         fflush(stderr);
911
912         do {
913                 mirror.count = 0;
914                 mirror.pfs_id = glob_pfs.pfs_id;
915                 mirror.shared_uuid = glob_pfs.ondisk->shared_uuid;
916                 if (ioctl(glob_fd, HAMMERIOC_MIRROR_READ, &mirror) < 0)
917                         err(1, "Mirror-read %s failed", filesystem);
918                 if (mirror.head.flags & HAMMER_IOC_HEAD_ERROR)
919                         errx(1, "Mirror-read %s fatal error %d",
920                                 filesystem, mirror.head.error);
921                 if (mirror.count) {
922                         offset = 0;
923                         while (offset < mirror.count) {
924                                 mrec = (void *)((char *)buf + offset);
925                                 bytes = HAMMER_HEAD_DOALIGN(mrec->head.rec_size);
926                                 if (offset + bytes > mirror.count)
927                                         errx(1, "Misaligned record");
928                                 assert((mrec->head.type &
929                                        HAMMER_MRECF_TYPE_MASK) ==
930                                        HAMMER_MREC_TYPE_REC);
931                                 offset += bytes;
932                                 elm = mrec->rec.leaf;
933                                 if (elm.base.btype != HAMMER_BTREE_TYPE_RECORD)
934                                         continue;
935                                 if (elm.base.rec_type != HAMMER_RECTYPE_DATA)
936                                         continue;
937                                 ++DedupCurrentRecords;
938                                 if (DedupCrcStart != DedupCrcEnd) {
939                                         if (elm.data_crc < DedupCrcStart)
940                                                 continue;
941                                         if (DedupCrcEnd &&
942                                             elm.data_crc >= DedupCrcEnd) {
943                                                 continue;
944                                         }
945                                 }
946                                 func(&elm, 0);
947                         }
948                 }
949                 mirror.key_beg = mirror.key_cur;
950                 if (DidInterrupt || SigAlrmFlag) {
951                         if (VerboseOpt)
952                                 fprintf(stderr, "%s\n",
953                                     (DidInterrupt ? "Interrupted" : "Timeout"));
954                         hammer_set_cycle(&mirror.key_cur, mirror.tid_beg);
955                         if (VerboseOpt)
956                                 fprintf(stderr, "Cyclefile %s updated for "
957                                     "continuation\n", CyclePath);
958                         exit(1);
959                 }
960                 if (SigInfoFlag) {
961                         if (DedupTotalRecords) {
962                                 humanize_unsigned(buf1, sizeof(buf1),
963                                                   DedupDataReads,
964                                                   "B", 1024);
965                                 humanize_unsigned(buf2, sizeof(buf2),
966                                                   dedup_successes_bytes,
967                                                   "B", 1024);
968                                 fprintf(stderr, "%s count %7jd/%jd "
969                                                 "(%02d.%02d%%) "
970                                                 "ioread %s newddup %s\n",
971                                         id,
972                                         (intmax_t)DedupCurrentRecords,
973                                         (intmax_t)DedupTotalRecords,
974                                         (int)(DedupCurrentRecords * 100 /
975                                                 DedupTotalRecords),
976                                         (int)(DedupCurrentRecords * 10000 /
977                                                 DedupTotalRecords % 100),
978                                         buf1, buf2);
979                         } else {
980                                 fprintf(stderr, "%s count %-7jd\n",
981                                         id,
982                                         (intmax_t)DedupCurrentRecords);
983                         }
984                         SigInfoFlag = 0;
985                 }
986         } while (mirror.count != 0);
987
988         signal(SIGINFO, SIG_IGN);
989         signal(SIGALRM, SIG_IGN);
990
991         free(buf);
992 }
993
994 static void
995 dump_simulated_dedup(void)
996 {
997         struct sim_dedup_entry *sim_de;
998
999         printf("=== Dumping simulated dedup entries:\n");
1000         RB_FOREACH(sim_de, sim_dedup_entry_rb_tree, &sim_dedup_tree) {
1001                 printf("\tcrc=%08x cnt=%ju size=%ju\n",
1002                         sim_de->crc,
1003                         (intmax_t)sim_de->ref_blks,
1004                         (intmax_t)sim_de->ref_size);
1005         }
1006         printf("end of dump ===\n");
1007 }
1008
1009 static void
1010 dump_real_dedup(void)
1011 {
1012         struct dedup_entry *de;
1013         struct sha_dedup_entry *sha_de;
1014         int i;
1015
1016         printf("=== Dumping dedup entries:\n");
1017         RB_FOREACH(de, dedup_entry_rb_tree, &dedup_tree) {
1018                 if (de->flags & HAMMER_DEDUP_ENTRY_FICTITIOUS) {
1019                         printf("\tcrc=%08x fictitious\n", de->leaf.data_crc);
1020
1021                         RB_FOREACH(sha_de, sha_dedup_entry_rb_tree, &de->u.fict_root) {
1022                                 printf("\t\tcrc=%08x cnt=%ju size=%ju\n\t"
1023                                        "\t\tsha=",
1024                                        sha_de->leaf.data_crc,
1025                                        (intmax_t)sha_de->ref_blks,
1026                                        (intmax_t)sha_de->ref_size);
1027                                 for (i = 0; i < SHA256_DIGEST_LENGTH; ++i)
1028                                         printf("%02x", sha_de->sha_hash[i]);
1029                                 printf("\n");
1030                         }
1031                 } else {
1032                         printf("\tcrc=%08x cnt=%ju size=%ju\n",
1033                                de->leaf.data_crc,
1034                                (intmax_t)de->u.de.ref_blks,
1035                                (intmax_t)de->u.de.ref_size);
1036                 }
1037         }
1038         printf("end of dump ===\n");
1039 }
1040
1041 static void
1042 dedup_usage(int code)
1043 {
1044         fprintf(stderr,
1045                 "hammer dedup-simulate <filesystem>\n"
1046                 "hammer dedup <filesystem>\n"
1047         );
1048         exit(code);
1049 }