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 "hammer.h"
  36 #include <libutil.h>
  37 #include <crypto/sha2/sha2.h>
  38
  39 #define DEDUP_BUF (64 * 1024)
  40
  41 /* Sorted list of block CRCs - light version for dedup-simulate */
  42 struct sim_dedup_entry_rb_tree;
  43 RB_HEAD(sim_dedup_entry_rb_tree, sim_dedup_entry) sim_dedup_tree =
  44                                         RB_INITIALIZER(&sim_dedup_tree);
  45 RB_PROTOTYPE2(sim_dedup_entry_rb_tree, sim_dedup_entry, rb_entry,
  46                 rb_sim_dedup_entry_compare, hammer_crc_t);
  47
  48 struct sim_dedup_entry {
  49         hammer_crc_t    crc;
  50         u_int64_t       ref_blks; /* number of blocks referenced */
  51         u_int64_t       ref_size; /* size of data referenced */
  52         RB_ENTRY(sim_dedup_entry) rb_entry;
  53 };
  54
  55 /* Sorted list of HAMMER B-Tree keys */
  56 struct dedup_entry_rb_tree;
  57 struct sha_dedup_entry_rb_tree;
  58
  59 RB_HEAD(dedup_entry_rb_tree, dedup_entry) dedup_tree =
  60                                         RB_INITIALIZER(&dedup_tree);
  61 RB_PROTOTYPE2(dedup_entry_rb_tree, dedup_entry, rb_entry,
  62                 rb_dedup_entry_compare, hammer_crc_t);
  63
  64 RB_PROTOTYPE(sha_dedup_entry_rb_tree, sha_dedup_entry, fict_entry,
  65                 rb_sha_dedup_entry_compare);
  66
  67 struct dedup_entry {
  68         struct hammer_btree_leaf_elm leaf;
  69         union {
  70                 struct {
  71                         u_int64_t ref_blks;
  72                         u_int64_t ref_size;
  73                 } de;
  74                 RB_HEAD(sha_dedup_entry_rb_tree, sha_dedup_entry) fict_root;
  75         } u;
  76         u_int8_t flags;
  77         RB_ENTRY(dedup_entry) rb_entry;
  78 };
  79
  80 #define HAMMER_DEDUP_ENTRY_FICTITIOUS   0x0001
  81
  82 struct sha_dedup_entry {
  83         struct hammer_btree_leaf_elm    leaf;
  84         u_int64_t                       ref_blks;
  85         u_int64_t                       ref_size;
  86         u_int8_t                        sha_hash[SHA256_DIGEST_LENGTH];
  87         RB_ENTRY(sha_dedup_entry)       fict_entry;
  88 };
  89
  90 /*
  91  * Pass2 list - contains entries that were not dedup'ed because ioctl failed
  92  */
  93 STAILQ_HEAD(, pass2_dedup_entry) pass2_dedup_queue =
  94                                 STAILQ_HEAD_INITIALIZER(pass2_dedup_queue);
  95
  96 struct pass2_dedup_entry {
  97         struct hammer_btree_leaf_elm    leaf;
  98         STAILQ_ENTRY(pass2_dedup_entry) sq_entry;
  99 };
 100
 101 #define DEDUP_PASS2     0x0001 /* process_btree_elm() mode */
 102
 103 /* PFS global ids - we deal with just one PFS at a run */
 104 int glob_fd;
 105 struct hammer_ioc_pseudofs_rw glob_pfs;
 106
 107 /*
 108  * Global accounting variables
 109  *
 110  * Last three don't have to be 64-bit, just to be safe..
 111  */
 112 u_int64_t dedup_alloc_size = 0;
 113 u_int64_t dedup_ref_size = 0;
 114 u_int64_t dedup_skipped_size = 0;
 115 u_int64_t dedup_crc_failures = 0;
 116 u_int64_t dedup_sha_failures = 0;
 117 u_int64_t dedup_underflows = 0;
 118
 119 static int rb_sim_dedup_entry_compare(struct sim_dedup_entry *sim_de1,
 120                                 struct sim_dedup_entry *sim_de2);
 121 static int rb_dedup_entry_compare(struct dedup_entry *de1,
 122                                 struct dedup_entry *de2);
 123 static int rb_sha_dedup_entry_compare(struct sha_dedup_entry *sha_de1,
 124                                 struct sha_dedup_entry *sha_de2);
 125 typedef int (*scan_pfs_cb_t)(hammer_btree_leaf_elm_t scan_leaf, int flags);
 126 static void scan_pfs(char *filesystem, scan_pfs_cb_t func);
 127 static int collect_btree_elm(hammer_btree_leaf_elm_t scan_leaf, int flags);
 128 static int process_btree_elm(hammer_btree_leaf_elm_t scan_leaf, int flags);
 129 static int upgrade_chksum(hammer_btree_leaf_elm_t leaf, u_int8_t *sha_hash);
 130 static void dump_simulated_dedup(void);
 131 static void dump_real_dedup(void);
 132 static void dedup_usage(int code);
 133
 134 RB_GENERATE2(sim_dedup_entry_rb_tree, sim_dedup_entry, rb_entry,
 135                 rb_sim_dedup_entry_compare, hammer_crc_t, crc);
 136 RB_GENERATE2(dedup_entry_rb_tree, dedup_entry, rb_entry,
 137                 rb_dedup_entry_compare, hammer_crc_t, leaf.data_crc);
 138 RB_GENERATE(sha_dedup_entry_rb_tree, sha_dedup_entry, fict_entry,
 139                 rb_sha_dedup_entry_compare);
 140
 141 static int
 142 rb_sim_dedup_entry_compare(struct sim_dedup_entry *sim_de1,
 143                         struct sim_dedup_entry *sim_de2)
 144 {
 145         if (sim_de1->crc < sim_de2->crc)
 146                 return (-1);
 147         if (sim_de1->crc > sim_de2->crc)
 148                 return (1);
 149
 150         return (0);
 151 }
 152
 153 static int
 154 rb_dedup_entry_compare(struct dedup_entry *de1, struct dedup_entry *de2)
 155 {
 156         if (de1->leaf.data_crc < de2->leaf.data_crc)
 157                 return (-1);
 158         if (de1->leaf.data_crc > de2->leaf.data_crc)
 159                 return (1);
 160
 161         return (0);
 162 }
 163
 164 static int
 165 rb_sha_dedup_entry_compare(struct sha_dedup_entry *sha_de1,
 166                         struct sha_dedup_entry *sha_de2)
 167 {
 168         unsigned long *h1 = (unsigned long *)&sha_de1->sha_hash;
 169         unsigned long *h2 = (unsigned long *)&sha_de2->sha_hash;
 170         int i;
 171
 172         for (i = 0; i < SHA256_DIGEST_LENGTH / (int)sizeof(unsigned long); ++i) {
 173                 if (h1[i] < h2[i])
 174                         return (-1);
 175                 if (h1[i] > h2[i])
 176                         return (1);
 177         }
 178
 179         return (0);
 180 }
 181
 182 /*
 183  * dedup-simulate <filesystem>
 184  */
 185 void
 186 hammer_cmd_dedup_simulate(char **av, int ac)
 187 {
 188         struct sim_dedup_entry *sim_de;
 189
 190         if (ac != 1)
 191                 dedup_usage(1);
 192
 193         glob_fd = getpfs(&glob_pfs, av[0]);
 194
 195         printf("Dedup-simulate ");
 196         scan_pfs(av[0], &collect_btree_elm);
 197         printf("Dedup-simulate %s succeeded\n", av[0]);
 198
 199         relpfs(glob_fd, &glob_pfs);
 200
 201         if (VerboseOpt >= 2)
 202                 dump_simulated_dedup();
 203
 204         /*
 205          * Calculate simulated dedup ratio and get rid of the tree
 206          */
 207         while ((sim_de = RB_ROOT(&sim_dedup_tree)) != NULL) {
 208                 assert(sim_de->ref_blks != 0);
 209                 dedup_ref_size += sim_de->ref_size;
 210                 dedup_alloc_size += sim_de->ref_size / sim_de->ref_blks;
 211
 212                 RB_REMOVE(sim_dedup_entry_rb_tree, &sim_dedup_tree, sim_de);
 213                 free(sim_de);
 214         }
 215
 216         printf("Simulated dedup ratio = %.2f\n",
 217             (double)dedup_ref_size / dedup_alloc_size);
 218 }
 219
 220 /*
 221  * dedup <filesystem>
 222  */
 223 void
 224 hammer_cmd_dedup(char **av, int ac)
 225 {
 226         struct dedup_entry *de;
 227         struct sha_dedup_entry *sha_de;
 228         struct pass2_dedup_entry *pass2_de;
 229         char buf[8];
 230
 231         if (TimeoutOpt > 0)
 232                 alarm(TimeoutOpt);
 233
 234         if (ac != 1)
 235                 dedup_usage(1);
 236
 237         STAILQ_INIT(&pass2_dedup_queue);
 238
 239         glob_fd = getpfs(&glob_pfs, av[0]);
 240
 241         printf("Dedup ");
 242         scan_pfs(av[0], &process_btree_elm);
 243
 244         while ((pass2_de = STAILQ_FIRST(&pass2_dedup_queue)) != NULL) {
 245                 if (process_btree_elm(&pass2_de->leaf, DEDUP_PASS2))
 246                         dedup_skipped_size -= pass2_de->leaf.data_len;
 247
 248                 STAILQ_REMOVE_HEAD(&pass2_dedup_queue, sq_entry);
 249                 free(pass2_de);
 250         }
 251         assert(STAILQ_EMPTY(&pass2_dedup_queue));
 252
 253         printf("Dedup %s succeeded\n", av[0]);
 254
 255         relpfs(glob_fd, &glob_pfs);
 256
 257         if (VerboseOpt >= 2)
 258                 dump_real_dedup();
 259
 260         /*
 261          * Calculate dedup ratio and get rid of the trees
 262          */
 263         while ((de = RB_ROOT(&dedup_tree)) != NULL) {
 264                 if (de->flags & HAMMER_DEDUP_ENTRY_FICTITIOUS) {
 265                         while ((sha_de = RB_ROOT(&de->u.fict_root)) != NULL) {
 266                                 assert(sha_de->ref_blks != 0);
 267                                 dedup_ref_size += sha_de->ref_size;
 268                                 dedup_alloc_size += sha_de->ref_size / sha_de->ref_blks;
 269
 270                                 RB_REMOVE(sha_dedup_entry_rb_tree,
 271                                                 &de->u.fict_root, sha_de);
 272                                 free(sha_de);
 273                         }
 274                         assert(RB_EMPTY(&de->u.fict_root));
 275                 } else {
 276                         assert(de->u.de.ref_blks != 0);
 277                         dedup_ref_size += de->u.de.ref_size;
 278                         dedup_alloc_size += de->u.de.ref_size / de->u.de.ref_blks;
 279                 }
 280
 281                 RB_REMOVE(dedup_entry_rb_tree, &dedup_tree, de);
 282                 free(de);
 283         }
 284         assert(RB_EMPTY(&dedup_tree));
 285
 286         assert(dedup_alloc_size != 0);
 287         humanize_unsigned(buf, sizeof(buf), dedup_ref_size, "B", 1024);
 288         printf("Dedup ratio = %.2f\n"
 289                "    %8s referenced\n",
 290                (double)dedup_ref_size / dedup_alloc_size,
 291                buf
 292         );
 293         humanize_unsigned(buf, sizeof(buf), dedup_alloc_size, "B", 1024);
 294         printf("    %8s allocated\n", buf);
 295         humanize_unsigned(buf, sizeof(buf), dedup_skipped_size, "B", 1024);
 296         printf("    %8s skipped\n", buf);
 297         printf("    %8jd CRC collisions\n"
 298                "    %8jd SHA collisions\n"
 299                "    %8jd bigblock underflows\n",
 300                (intmax_t)dedup_crc_failures,
 301                (intmax_t)dedup_sha_failures,
 302                (intmax_t)dedup_underflows
 303         );
 304 }
 305
 306 static int
 307 collect_btree_elm(hammer_btree_leaf_elm_t scan_leaf, int flags __unused)
 308 {
 309         struct sim_dedup_entry *sim_de;
 310
 311         sim_de = RB_LOOKUP(sim_dedup_entry_rb_tree, &sim_dedup_tree, scan_leaf->data_crc);
 312
 313         if (sim_de == NULL) {
 314                 sim_de = calloc(sizeof(*sim_de), 1);
 315                 sim_de->crc = scan_leaf->data_crc;
 316                 RB_INSERT(sim_dedup_entry_rb_tree, &sim_dedup_tree, sim_de);
 317         }
 318
 319         sim_de->ref_blks += 1;
 320         sim_de->ref_size += scan_leaf->data_len;
 321         return (1);
 322 }
 323
 324 static __inline int
 325 validate_dedup_pair(hammer_btree_leaf_elm_t p, hammer_btree_leaf_elm_t q)
 326 {
 327         if ((p->data_offset & HAMMER_OFF_ZONE_MASK) !=
 328             (q->data_offset & HAMMER_OFF_ZONE_MASK)) {
 329                 return (1);
 330         }
 331         if (p->data_len != q->data_len) {
 332                 return (1);
 333         }
 334
 335         return (0);
 336 }
 337
 338 #define DEDUP_TECH_FAILURE      1
 339 #define DEDUP_CMP_FAILURE       2
 340 #define DEDUP_INVALID_ZONE      3
 341 #define DEDUP_UNDERFLOW         4
 342
 343 static __inline int
 344 deduplicate(hammer_btree_leaf_elm_t p, hammer_btree_leaf_elm_t q)
 345 {
 346         struct hammer_ioc_dedup dedup;
 347
 348         bzero(&dedup, sizeof(dedup));
 349
 350         /*
 351          * If data_offset fields are the same there is no need to run ioctl,
 352          * candidate is already dedup'ed.
 353          */
 354         if (p->data_offset == q->data_offset) {
 355                 return (0);
 356         }
 357
 358         dedup.elm1 = p->base;
 359         dedup.elm2 = q->base;
 360         RunningIoctl = 1;
 361         if (ioctl(glob_fd, HAMMERIOC_DEDUP, &dedup) < 0) {
 362                 /* Technical failure - locking or w/e */
 363                 return (DEDUP_TECH_FAILURE);
 364         }
 365         if (dedup.head.flags & HAMMER_IOC_DEDUP_CMP_FAILURE) {
 366                 return (DEDUP_CMP_FAILURE);
 367         }
 368         if (dedup.head.flags & HAMMER_IOC_DEDUP_INVALID_ZONE) {
 369                 return (DEDUP_INVALID_ZONE);
 370         }
 371         if (dedup.head.flags & HAMMER_IOC_DEDUP_UNDERFLOW) {
 372                 return (DEDUP_UNDERFLOW);
 373         }
 374         RunningIoctl = 0;
 375
 376         return (0);
 377 }
 378
 379 static int
 380 process_btree_elm(hammer_btree_leaf_elm_t scan_leaf, int flags)
 381 {
 382         struct dedup_entry *de;
 383         struct sha_dedup_entry *sha_de, temp;
 384         struct pass2_dedup_entry *pass2_de;
 385         int error;
 386
 387         de = RB_LOOKUP(dedup_entry_rb_tree, &dedup_tree, scan_leaf->data_crc);
 388         if (de == NULL) {
 389                 /*
 390                  * Insert new entry into CRC tree
 391                  */
 392                 de = calloc(sizeof(*de), 1);
 393                 de->leaf = *scan_leaf;
 394                 RB_INSERT(dedup_entry_rb_tree, &dedup_tree, de);
 395                 goto upgrade_stats;
 396         }
 397
 398         /*
 399          * Found entry in CRC tree
 400          */
 401         if (de->flags & HAMMER_DEDUP_ENTRY_FICTITIOUS) {
 402                 /*
 403                  * Entry in CRC tree is fictious, so we already had problems
 404                  * with this CRC. Upgrade (compute SHA) the candidate and
 405                  * dive into SHA subtree. If upgrade fails insert the candidate
 406                  * into Pass2 list (it will be processed later).
 407                  */
 408                 if (upgrade_chksum(scan_leaf, temp.sha_hash))
 409                         goto pass2_insert;
 410
 411                 sha_de = RB_FIND(sha_dedup_entry_rb_tree, &de->u.fict_root, &temp);
 412                 if (sha_de == NULL) {
 413                         /*
 414                          * Nothing in SHA subtree so far, so this is a new
 415                          * 'dataset'. Insert new entry into SHA subtree.
 416                          */
 417                         sha_de = calloc(sizeof(*sha_de), 1);
 418                         sha_de->leaf = *scan_leaf;
 419                         memcpy(sha_de->sha_hash, temp.sha_hash, SHA256_DIGEST_LENGTH);
 420                         RB_INSERT(sha_dedup_entry_rb_tree, &de->u.fict_root, sha_de);
 421                         goto upgrade_stats_sha;
 422                 }
 423
 424                 /*
 425                  * Found entry in SHA subtree, it means we have a potential
 426                  * dedup pair. Validate it (zones have to match and data_len
 427                  * field have to be the same too. If validation fails, treat
 428                  * it as a SHA collision (jump to sha256_failure).
 429                  */
 430                 if (validate_dedup_pair(&sha_de->leaf, scan_leaf))
 431                         goto sha256_failure;
 432
 433                 /*
 434                  * We have a valid dedup pair (SHA match, validated).
 435                  *
 436                  * In case of technical failure (dedup pair was good, but
 437                  * ioctl failed anyways) insert the candidate into Pass2 list
 438                  * (we will try to dedup it after we are done with the rest of
 439                  * the tree).
 440                  *
 441                  * If ioctl fails because either of blocks is in the non-dedup
 442                  * zone (we can dedup only in LARGE_DATA and SMALL_DATA) don't
 443                  * bother with the candidate and terminate early.
 444                  *
 445                  * If ioctl fails because of bigblock underflow replace the
 446                  * leaf node that found dedup entry represents with scan_leaf.
 447                  */
 448                 error = deduplicate(&sha_de->leaf, scan_leaf);
 449                 switch(error) {
 450                 case DEDUP_TECH_FAILURE:
 451                         goto pass2_insert;
 452                 case DEDUP_CMP_FAILURE:
 453                         goto sha256_failure;
 454                 case DEDUP_INVALID_ZONE:
 455                         goto terminate_early;
 456                 case DEDUP_UNDERFLOW:
 457                         ++dedup_underflows;
 458                         sha_de->leaf = *scan_leaf;
 459                         memcpy(sha_de->sha_hash, temp.sha_hash, SHA256_DIGEST_LENGTH);
 460                         goto upgrade_stats_sha;
 461                 default:
 462                         goto upgrade_stats_sha;
 463                 }
 464
 465                 /*
 466                  * Ooh la la.. SHA-256 collision. Terminate early, there's
 467                  * nothing we can do here.
 468                  */
 469 sha256_failure:
 470                 ++dedup_sha_failures;
 471                 goto terminate_early;
 472         } else {
 473                 /*
 474                  * Candidate CRC is good for now (we found an entry in CRC
 475                  * tree and it's not fictious). This means we have a
 476                  * potential dedup pair.
 477                  */
 478                 if (validate_dedup_pair(&de->leaf, scan_leaf))
 479                         goto crc_failure;
 480
 481                 /*
 482                  * We have a valid dedup pair (CRC match, validated)
 483                  */
 484                 error = deduplicate(&de->leaf, scan_leaf);
 485                 switch(error) {
 486                 case DEDUP_TECH_FAILURE:
 487                         goto pass2_insert;
 488                 case DEDUP_CMP_FAILURE:
 489                         goto crc_failure;
 490                 case DEDUP_INVALID_ZONE:
 491                         goto terminate_early;
 492                 case DEDUP_UNDERFLOW:
 493                         ++dedup_underflows;
 494                         de->leaf = *scan_leaf;
 495                         goto upgrade_stats;
 496                 default:
 497                         goto upgrade_stats;
 498                 }
 499
 500 crc_failure:
 501                 /*
 502                  * We got a CRC collision - either ioctl failed because of
 503                  * the comparison failure or validation of the potential
 504                  * dedup pair went bad. In all cases insert both blocks
 505                  * into SHA subtree (this requires checksum upgrade) and mark
 506                  * entry that corresponds to this CRC in the CRC tree
 507                  * fictious, so that all futher operations with this CRC go
 508                  * through SHA subtree.
 509                  */
 510                 ++dedup_crc_failures;
 511                 /*
 512                  * Insert block that was represented by now fictious dedup
 513                  * entry (create a new SHA entry and preserve stats of the
 514                  * old CRC one). If checksum upgrade fails insert the
 515                  * candidate into Pass2 list and return - keep both trees
 516                  * unmodified.
 517                  */
 518                 sha_de = calloc(sizeof(*sha_de), 1);
 519                 sha_de->leaf = de->leaf;
 520                 sha_de->ref_blks = de->u.de.ref_blks;
 521                 sha_de->ref_size = de->u.de.ref_size;
 522                 if (upgrade_chksum(&sha_de->leaf, sha_de->sha_hash)) {
 523                         free(sha_de);
 524                         goto pass2_insert;
 525                 }
 526
 527                 RB_INIT(&de->u.fict_root);
 528                 /*
 529                  * Here we can insert without prior checking because the tree
 530                  * is empty at this point
 531                  */
 532                 RB_INSERT(sha_dedup_entry_rb_tree, &de->u.fict_root, sha_de);
 533
 534                 /*
 535                  * Mark entry in CRC tree fictious
 536                  */
 537                 de->flags |= HAMMER_DEDUP_ENTRY_FICTITIOUS;
 538
 539                 /*
 540                  * Upgrade checksum of the candidate and insert it into
 541                  * SHA subtree. If upgrade fails insert the candidate into
 542                  * Pass2 list.
 543                  */
 544                 if (upgrade_chksum(scan_leaf, temp.sha_hash)) {
 545                         goto pass2_insert;
 546                 }
 547                 sha_de = RB_FIND(sha_dedup_entry_rb_tree, &de->u.fict_root, &temp);
 548                 if (sha_de != NULL)
 549                         /* There is an entry with this SHA already, but the only
 550                          * RB-tree element at this point is that entry we just
 551                          * added. We know for sure these blocks are different
 552                          * (this is crc_failure branch) so treat it as SHA
 553                          * collision.
 554                          */
 555                         goto sha256_failure;
 556
 557                 sha_de = calloc(sizeof(*sha_de), 1);
 558                 sha_de->leaf = *scan_leaf;
 559                 memcpy(sha_de->sha_hash, temp.sha_hash, SHA256_DIGEST_LENGTH);
 560                 RB_INSERT(sha_dedup_entry_rb_tree, &de->u.fict_root, sha_de);
 561                 goto upgrade_stats_sha;
 562         }
 563
 564 upgrade_stats:
 565         de->u.de.ref_blks += 1;
 566         de->u.de.ref_size += scan_leaf->data_len;
 567         return (1);
 568
 569 upgrade_stats_sha:
 570         sha_de->ref_blks += 1;
 571         sha_de->ref_size += scan_leaf->data_len;
 572         return (1);
 573
 574 pass2_insert:
 575         /*
 576          * If in pass2 mode don't insert anything, fall through to
 577          * terminate_early
 578          */
 579         if ((flags & DEDUP_PASS2) == 0) {
 580                 pass2_de = calloc(sizeof(*pass2_de), 1);
 581                 pass2_de->leaf = *scan_leaf;
 582                 STAILQ_INSERT_TAIL(&pass2_dedup_queue, pass2_de, sq_entry);
 583                 dedup_skipped_size += scan_leaf->data_len;
 584                 return (1);
 585         }
 586
 587 terminate_early:
 588         /*
 589          * Early termination path. Fixup stats.
 590          */
 591         dedup_alloc_size += scan_leaf->data_len;
 592         dedup_ref_size += scan_leaf->data_len;
 593         return (0);
 594 }
 595
 596 static int
 597 upgrade_chksum(hammer_btree_leaf_elm_t leaf, u_int8_t *sha_hash)
 598 {
 599         struct hammer_ioc_data data;
 600         char *buf = malloc(DEDUP_BUF);
 601         SHA256_CTX ctx;
 602         int error;
 603
 604         bzero(&data, sizeof(data));
 605         data.elm = leaf->base;
 606         data.ubuf = buf;
 607         data.size = DEDUP_BUF;
 608
 609         error = 0;
 610         if (ioctl(glob_fd, HAMMERIOC_GET_DATA, &data) < 0) {
 611                 fprintf(stderr, "Get-data failed: %s\n", strerror(errno));
 612                 error = 1;
 613                 goto done;
 614         }
 615
 616         if (data.leaf.data_len != leaf->data_len) {
 617                 error = 1;
 618                 goto done;
 619         }
 620
 621         if (data.leaf.base.btype == HAMMER_BTREE_TYPE_RECORD &&
 622             data.leaf.base.rec_type == HAMMER_RECTYPE_DATA) {
 623                 SHA256_Init(&ctx);
 624                 SHA256_Update(&ctx, (void *)buf, data.leaf.data_len);
 625                 SHA256_Final(sha_hash, &ctx);
 626         }
 627
 628 done:
 629         free(buf);
 630         return (error);
 631 }
 632
 633 static void
 634 scan_pfs(char *filesystem, scan_pfs_cb_t func)
 635 {
 636         struct hammer_ioc_mirror_rw mirror;
 637         hammer_ioc_mrecord_any_t mrec;
 638         struct hammer_btree_leaf_elm elm;
 639         char *buf = malloc(DEDUP_BUF);
 640         int offset, bytes;
 641
 642         bzero(&mirror, sizeof(mirror));
 643         hammer_key_beg_init(&mirror.key_beg);
 644         hammer_key_end_init(&mirror.key_end);
 645
 646         mirror.tid_beg = glob_pfs.ondisk->sync_beg_tid;
 647         mirror.tid_end = glob_pfs.ondisk->sync_end_tid;
 648         mirror.head.flags |= HAMMER_IOC_MIRROR_NODATA; /* we want only keys */
 649         mirror.ubuf = buf;
 650         mirror.size = DEDUP_BUF;
 651         mirror.pfs_id = glob_pfs.pfs_id;
 652         mirror.shared_uuid = glob_pfs.ondisk->shared_uuid;
 653
 654         printf("%s: objspace %016jx:%04x %016jx:%04x pfs_id %d\n",
 655                 filesystem,
 656                 (uintmax_t)mirror.key_beg.obj_id,
 657                 mirror.key_beg.localization,
 658                 (uintmax_t)mirror.key_end.obj_id,
 659                 mirror.key_end.localization,
 660                 glob_pfs.pfs_id);
 661         fflush(stdout);
 662
 663         do {
 664                 mirror.count = 0;
 665                 mirror.pfs_id = glob_pfs.pfs_id;
 666                 mirror.shared_uuid = glob_pfs.ondisk->shared_uuid;
 667                 if (ioctl(glob_fd, HAMMERIOC_MIRROR_READ, &mirror) < 0) {
 668                         fprintf(stderr, "Mirror-read %s failed: %s\n",
 669                                 filesystem, strerror(errno));
 670                         exit(1);
 671                 }
 672                 if (mirror.head.flags & HAMMER_IOC_HEAD_ERROR) {
 673                         fprintf(stderr, "Mirror-read %s fatal error %d\n",
 674                                 filesystem, mirror.head.error);
 675                         exit(1);
 676                 }
 677                 if (mirror.count) {
 678                         offset = 0;
 679                         while (offset < mirror.count) {
 680                                 mrec = (void *)((char *)buf + offset);
 681                                 bytes = HAMMER_HEAD_DOALIGN(mrec->head.rec_size);
 682                                 if (offset + bytes > mirror.count) {
 683                                         fprintf(stderr, "Misaligned record\n");
 684                                         exit(1);
 685                                 }
 686                                 assert((mrec->head.type &
 687                                     HAMMER_MRECF_TYPE_MASK) == HAMMER_MREC_TYPE_REC);
 688
 689                                 elm = mrec->rec.leaf;
 690                                 if (elm.base.btype == HAMMER_BTREE_TYPE_RECORD &&
 691                                     elm.base.rec_type == HAMMER_RECTYPE_DATA) {
 692                                         func(&elm, 0);
 693                                 }
 694                                 offset += bytes;
 695                         }
 696                 }
 697                 mirror.key_beg = mirror.key_cur;
 698         } while (mirror.count != 0);
 699
 700         free(buf);
 701 }
 702
 703 static void
 704 dump_simulated_dedup(void)
 705 {
 706         struct sim_dedup_entry *sim_de;
 707
 708         printf("=== Dumping simulated dedup entries:\n");
 709         RB_FOREACH(sim_de, sim_dedup_entry_rb_tree, &sim_dedup_tree) {
 710                 printf("\tcrc=%08x cnt=%llu size=%llu\n", sim_de->crc,
 711                     sim_de->ref_blks, sim_de->ref_size);
 712         }
 713         printf("end of dump ===\n");
 714 }
 715
 716 static void
 717 dump_real_dedup(void)
 718 {
 719         struct dedup_entry *de;
 720         struct sha_dedup_entry *sha_de;
 721         int i;
 722
 723         printf("=== Dumping dedup entries:\n");
 724         RB_FOREACH(de, dedup_entry_rb_tree, &dedup_tree) {
 725                 if (de->flags & HAMMER_DEDUP_ENTRY_FICTITIOUS) {
 726                         printf("\tcrc=%08x fictious\n", de->leaf.data_crc);
 727
 728                         RB_FOREACH(sha_de, sha_dedup_entry_rb_tree, &de->u.fict_root) {
 729                                 printf("\t\tcrc=%08x cnt=%llu size=%llu\n\t\t\tsha=",
 730                                         sha_de->leaf.data_crc,
 731                                         sha_de->ref_blks,
 732                                         sha_de->ref_size);
 733                                 for (i = 0; i < SHA256_DIGEST_LENGTH; ++i)
 734                                         printf("%02x", sha_de->sha_hash[i]);
 735                                 printf("\n");
 736                         }
 737                 } else {
 738                         printf("\tcrc=%08x cnt=%llu size=%llu\n",
 739                                 de->leaf.data_crc,
 740                                 de->u.de.ref_blks,
 741                                 de->u.de.ref_size);
 742                 }
 743         }
 744         printf("end of dump ===\n");
 745 }
 746
 747 static void
 748 dedup_usage(int code)
 749 {
 750         fprintf(stderr,
 751                 "hammer dedup-simulate <filesystem>\n"
 752                 "hammer dedup <filesystem>\n"
 753         );
 754         exit(code);
 755 }