fs/btrfs/ref-verify.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2014 Facebook.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/stacktrace.h>
   8 #include "ctree.h"
   9 #include "disk-io.h"
  10 #include "locking.h"
  11 #include "delayed-ref.h"
  12 #include "ref-verify.h"
  13
  14 /*
  15  * Used to keep track the roots and number of refs each root has for a given
  16  * bytenr.  This just tracks the number of direct references, no shared
  17  * references.
  18  */
  19 struct root_entry {
  20         u64 root_objectid;
  21         u64 num_refs;
  22         struct rb_node node;
  23 };
  24
  25 /*
  26  * These are meant to represent what should exist in the extent tree, these can
  27  * be used to verify the extent tree is consistent as these should all match
  28  * what the extent tree says.
  29  */
  30 struct ref_entry {
  31         u64 root_objectid;
  32         u64 parent;
  33         u64 owner;
  34         u64 offset;
  35         u64 num_refs;
  36         struct rb_node node;
  37 };
  38
  39 #define MAX_TRACE       16
  40
  41 /*
  42  * Whenever we add/remove a reference we record the action.  The action maps
  43  * back to the delayed ref action.  We hold the ref we are changing in the
  44  * action so we can account for the history properly, and we record the root we
  45  * were called with since it could be different from ref_root.  We also store
  46  * stack traces because that's how I roll.
  47  */
  48 struct ref_action {
  49         int action;
  50         u64 root;
  51         struct ref_entry ref;
  52         struct list_head list;
  53         unsigned long trace[MAX_TRACE];
  54         unsigned int trace_len;
  55 };
  56
  57 /*
  58  * One of these for every block we reference, it holds the roots and references
  59  * to it as well as all of the ref actions that have occurred to it.  We never
  60  * free it until we unmount the file system in order to make sure re-allocations
  61  * are happening properly.
  62  */
  63 struct block_entry {
  64         u64 bytenr;
  65         u64 len;
  66         u64 num_refs;
  67         int metadata;
  68         int from_disk;
  69         struct rb_root roots;
  70         struct rb_root refs;
  71         struct rb_node node;
  72         struct list_head actions;
  73 };
  74
  75 static struct block_entry *insert_block_entry(struct rb_root *root,
  76                                               struct block_entry *be)
  77 {
  78         struct rb_node **p = &root->rb_node;
  79         struct rb_node *parent_node = NULL;
  80         struct block_entry *entry;
  81
  82         while (*p) {
  83                 parent_node = *p;
  84                 entry = rb_entry(parent_node, struct block_entry, node);
  85                 if (entry->bytenr > be->bytenr)
  86                         p = &(*p)->rb_left;
  87                 else if (entry->bytenr < be->bytenr)
  88                         p = &(*p)->rb_right;
  89                 else
  90                         return entry;
  91         }
  92
  93         rb_link_node(&be->node, parent_node, p);
  94         rb_insert_color(&be->node, root);
  95         return NULL;
  96 }
  97
  98 static struct block_entry *lookup_block_entry(struct rb_root *root, u64 bytenr)
  99 {
 100         struct rb_node *n;
 101         struct block_entry *entry = NULL;
 102
 103         n = root->rb_node;
 104         while (n) {
 105                 entry = rb_entry(n, struct block_entry, node);
 106                 if (entry->bytenr < bytenr)
 107                         n = n->rb_right;
 108                 else if (entry->bytenr > bytenr)
 109                         n = n->rb_left;
 110                 else
 111                         return entry;
 112         }
 113         return NULL;
 114 }
 115
 116 static struct root_entry *insert_root_entry(struct rb_root *root,
 117                                             struct root_entry *re)
 118 {
 119         struct rb_node **p = &root->rb_node;
 120         struct rb_node *parent_node = NULL;
 121         struct root_entry *entry;
 122
 123         while (*p) {
 124                 parent_node = *p;
 125                 entry = rb_entry(parent_node, struct root_entry, node);
 126                 if (entry->root_objectid > re->root_objectid)
 127                         p = &(*p)->rb_left;
 128                 else if (entry->root_objectid < re->root_objectid)
 129                         p = &(*p)->rb_right;
 130                 else
 131                         return entry;
 132         }
 133
 134         rb_link_node(&re->node, parent_node, p);
 135         rb_insert_color(&re->node, root);
 136         return NULL;
 137
 138 }
 139
 140 static int comp_refs(struct ref_entry *ref1, struct ref_entry *ref2)
 141 {
 142         if (ref1->root_objectid < ref2->root_objectid)
 143                 return -1;
 144         if (ref1->root_objectid > ref2->root_objectid)
 145                 return 1;
 146         if (ref1->parent < ref2->parent)
 147                 return -1;
 148         if (ref1->parent > ref2->parent)
 149                 return 1;
 150         if (ref1->owner < ref2->owner)
 151                 return -1;
 152         if (ref1->owner > ref2->owner)
 153                 return 1;
 154         if (ref1->offset < ref2->offset)
 155                 return -1;
 156         if (ref1->offset > ref2->offset)
 157                 return 1;
 158         return 0;
 159 }
 160
 161 static struct ref_entry *insert_ref_entry(struct rb_root *root,
 162                                           struct ref_entry *ref)
 163 {
 164         struct rb_node **p = &root->rb_node;
 165         struct rb_node *parent_node = NULL;
 166         struct ref_entry *entry;
 167         int cmp;
 168
 169         while (*p) {
 170                 parent_node = *p;
 171                 entry = rb_entry(parent_node, struct ref_entry, node);
 172                 cmp = comp_refs(entry, ref);
 173                 if (cmp > 0)
 174                         p = &(*p)->rb_left;
 175                 else if (cmp < 0)
 176                         p = &(*p)->rb_right;
 177                 else
 178                         return entry;
 179         }
 180
 181         rb_link_node(&ref->node, parent_node, p);
 182         rb_insert_color(&ref->node, root);
 183         return NULL;
 184
 185 }
 186
 187 static struct root_entry *lookup_root_entry(struct rb_root *root, u64 objectid)
 188 {
 189         struct rb_node *n;
 190         struct root_entry *entry = NULL;
 191
 192         n = root->rb_node;
 193         while (n) {
 194                 entry = rb_entry(n, struct root_entry, node);
 195                 if (entry->root_objectid < objectid)
 196                         n = n->rb_right;
 197                 else if (entry->root_objectid > objectid)
 198                         n = n->rb_left;
 199                 else
 200                         return entry;
 201         }
 202         return NULL;
 203 }
 204
 205 #ifdef CONFIG_STACKTRACE
 206 static void __save_stack_trace(struct ref_action *ra)
 207 {
 208         ra->trace_len = stack_trace_save(ra->trace, MAX_TRACE, 2);
 209 }
 210
 211 static void __print_stack_trace(struct btrfs_fs_info *fs_info,
 212                                 struct ref_action *ra)
 213 {
 214         if (ra->trace_len == 0) {
 215                 btrfs_err(fs_info, "  ref-verify: no stacktrace");
 216                 return;
 217         }
 218         stack_trace_print(ra->trace, ra->trace_len, 2);
 219 }
 220 #else
 221 static void inline __save_stack_trace(struct ref_action *ra)
 222 {
 223 }
 224
 225 static void inline __print_stack_trace(struct btrfs_fs_info *fs_info,
 226                                        struct ref_action *ra)
 227 {
 228         btrfs_err(fs_info, "  ref-verify: no stacktrace support");
 229 }
 230 #endif
 231
 232 static void free_block_entry(struct block_entry *be)
 233 {
 234         struct root_entry *re;
 235         struct ref_entry *ref;
 236         struct ref_action *ra;
 237         struct rb_node *n;
 238
 239         while ((n = rb_first(&be->roots))) {
 240                 re = rb_entry(n, struct root_entry, node);
 241                 rb_erase(&re->node, &be->roots);
 242                 kfree(re);
 243         }
 244
 245         while((n = rb_first(&be->refs))) {
 246                 ref = rb_entry(n, struct ref_entry, node);
 247                 rb_erase(&ref->node, &be->refs);
 248                 kfree(ref);
 249         }
 250
 251         while (!list_empty(&be->actions)) {
 252                 ra = list_first_entry(&be->actions, struct ref_action,
 253                                       list);
 254                 list_del(&ra->list);
 255                 kfree(ra);
 256         }
 257         kfree(be);
 258 }
 259
 260 static struct block_entry *add_block_entry(struct btrfs_fs_info *fs_info,
 261                                            u64 bytenr, u64 len,
 262                                            u64 root_objectid)
 263 {
 264         struct block_entry *be = NULL, *exist;
 265         struct root_entry *re = NULL;
 266
 267         re = kzalloc(sizeof(struct root_entry), GFP_KERNEL);
 268         be = kzalloc(sizeof(struct block_entry), GFP_KERNEL);
 269         if (!be || !re) {
 270                 kfree(re);
 271                 kfree(be);
 272                 return ERR_PTR(-ENOMEM);
 273         }
 274         be->bytenr = bytenr;
 275         be->len = len;
 276
 277         re->root_objectid = root_objectid;
 278         re->num_refs = 0;
 279
 280         spin_lock(&fs_info->ref_verify_lock);
 281         exist = insert_block_entry(&fs_info->block_tree, be);
 282         if (exist) {
 283                 if (root_objectid) {
 284                         struct root_entry *exist_re;
 285
 286                         exist_re = insert_root_entry(&exist->roots, re);
 287                         if (exist_re)
 288                                 kfree(re);
 289                 }
 290                 kfree(be);
 291                 return exist;
 292         }
 293
 294         be->num_refs = 0;
 295         be->metadata = 0;
 296         be->from_disk = 0;
 297         be->roots = RB_ROOT;
 298         be->refs = RB_ROOT;
 299         INIT_LIST_HEAD(&be->actions);
 300         if (root_objectid)
 301                 insert_root_entry(&be->roots, re);
 302         else
 303                 kfree(re);
 304         return be;
 305 }
 306
 307 static int add_tree_block(struct btrfs_fs_info *fs_info, u64 ref_root,
 308                           u64 parent, u64 bytenr, int level)
 309 {
 310         struct block_entry *be;
 311         struct root_entry *re;
 312         struct ref_entry *ref = NULL, *exist;
 313
 314         ref = kmalloc(sizeof(struct ref_entry), GFP_KERNEL);
 315         if (!ref)
 316                 return -ENOMEM;
 317
 318         if (parent)
 319                 ref->root_objectid = 0;
 320         else
 321                 ref->root_objectid = ref_root;
 322         ref->parent = parent;
 323         ref->owner = level;
 324         ref->offset = 0;
 325         ref->num_refs = 1;
 326
 327         be = add_block_entry(fs_info, bytenr, fs_info->nodesize, ref_root);
 328         if (IS_ERR(be)) {
 329                 kfree(ref);
 330                 return PTR_ERR(be);
 331         }
 332         be->num_refs++;
 333         be->from_disk = 1;
 334         be->metadata = 1;
 335
 336         if (!parent) {
 337                 ASSERT(ref_root);
 338                 re = lookup_root_entry(&be->roots, ref_root);
 339                 ASSERT(re);
 340                 re->num_refs++;
 341         }
 342         exist = insert_ref_entry(&be->refs, ref);
 343         if (exist) {
 344                 exist->num_refs++;
 345                 kfree(ref);
 346         }
 347         spin_unlock(&fs_info->ref_verify_lock);
 348
 349         return 0;
 350 }
 351
 352 static int add_shared_data_ref(struct btrfs_fs_info *fs_info,
 353                                u64 parent, u32 num_refs, u64 bytenr,
 354                                u64 num_bytes)
 355 {
 356         struct block_entry *be;
 357         struct ref_entry *ref;
 358
 359         ref = kzalloc(sizeof(struct ref_entry), GFP_KERNEL);
 360         if (!ref)
 361                 return -ENOMEM;
 362         be = add_block_entry(fs_info, bytenr, num_bytes, 0);
 363         if (IS_ERR(be)) {
 364                 kfree(ref);
 365                 return PTR_ERR(be);
 366         }
 367         be->num_refs += num_refs;
 368
 369         ref->parent = parent;
 370         ref->num_refs = num_refs;
 371         if (insert_ref_entry(&be->refs, ref)) {
 372                 spin_unlock(&fs_info->ref_verify_lock);
 373                 btrfs_err(fs_info, "existing shared ref when reading from disk?");
 374                 kfree(ref);
 375                 return -EINVAL;
 376         }
 377         spin_unlock(&fs_info->ref_verify_lock);
 378         return 0;
 379 }
 380
 381 static int add_extent_data_ref(struct btrfs_fs_info *fs_info,
 382                                struct extent_buffer *leaf,
 383                                struct btrfs_extent_data_ref *dref,
 384                                u64 bytenr, u64 num_bytes)
 385 {
 386         struct block_entry *be;
 387         struct ref_entry *ref;
 388         struct root_entry *re;
 389         u64 ref_root = btrfs_extent_data_ref_root(leaf, dref);
 390         u64 owner = btrfs_extent_data_ref_objectid(leaf, dref);
 391         u64 offset = btrfs_extent_data_ref_offset(leaf, dref);
 392         u32 num_refs = btrfs_extent_data_ref_count(leaf, dref);
 393
 394         ref = kzalloc(sizeof(struct ref_entry), GFP_KERNEL);
 395         if (!ref)
 396                 return -ENOMEM;
 397         be = add_block_entry(fs_info, bytenr, num_bytes, ref_root);
 398         if (IS_ERR(be)) {
 399                 kfree(ref);
 400                 return PTR_ERR(be);
 401         }
 402         be->num_refs += num_refs;
 403
 404         ref->parent = 0;
 405         ref->owner = owner;
 406         ref->root_objectid = ref_root;
 407         ref->offset = offset;
 408         ref->num_refs = num_refs;
 409         if (insert_ref_entry(&be->refs, ref)) {
 410                 spin_unlock(&fs_info->ref_verify_lock);
 411                 btrfs_err(fs_info, "existing ref when reading from disk?");
 412                 kfree(ref);
 413                 return -EINVAL;
 414         }
 415
 416         re = lookup_root_entry(&be->roots, ref_root);
 417         if (!re) {
 418                 spin_unlock(&fs_info->ref_verify_lock);
 419                 btrfs_err(fs_info, "missing root in new block entry?");
 420                 return -EINVAL;
 421         }
 422         re->num_refs += num_refs;
 423         spin_unlock(&fs_info->ref_verify_lock);
 424         return 0;
 425 }
 426
 427 static int process_extent_item(struct btrfs_fs_info *fs_info,
 428                                struct btrfs_path *path, struct btrfs_key *key,
 429                                int slot, int *tree_block_level)
 430 {
 431         struct btrfs_extent_item *ei;
 432         struct btrfs_extent_inline_ref *iref;
 433         struct btrfs_extent_data_ref *dref;
 434         struct btrfs_shared_data_ref *sref;
 435         struct extent_buffer *leaf = path->nodes[0];
 436         u32 item_size = btrfs_item_size_nr(leaf, slot);
 437         unsigned long end, ptr;
 438         u64 offset, flags, count;
 439         int type, ret;
 440
 441         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 442         flags = btrfs_extent_flags(leaf, ei);
 443
 444         if ((key->type == BTRFS_EXTENT_ITEM_KEY) &&
 445             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 446                 struct btrfs_tree_block_info *info;
 447
 448                 info = (struct btrfs_tree_block_info *)(ei + 1);
 449                 *tree_block_level = btrfs_tree_block_level(leaf, info);
 450                 iref = (struct btrfs_extent_inline_ref *)(info + 1);
 451         } else {
 452                 if (key->type == BTRFS_METADATA_ITEM_KEY)
 453                         *tree_block_level = key->offset;
 454                 iref = (struct btrfs_extent_inline_ref *)(ei + 1);
 455         }
 456
 457         ptr = (unsigned long)iref;
 458         end = (unsigned long)ei + item_size;
 459         while (ptr < end) {
 460                 iref = (struct btrfs_extent_inline_ref *)ptr;
 461                 type = btrfs_extent_inline_ref_type(leaf, iref);
 462                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 463                 switch (type) {
 464                 case BTRFS_TREE_BLOCK_REF_KEY:
 465                         ret = add_tree_block(fs_info, offset, 0, key->objectid,
 466                                              *tree_block_level);
 467                         break;
 468                 case BTRFS_SHARED_BLOCK_REF_KEY:
 469                         ret = add_tree_block(fs_info, 0, offset, key->objectid,
 470                                              *tree_block_level);
 471                         break;
 472                 case BTRFS_EXTENT_DATA_REF_KEY:
 473                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 474                         ret = add_extent_data_ref(fs_info, leaf, dref,
 475                                                   key->objectid, key->offset);
 476                         break;
 477                 case BTRFS_SHARED_DATA_REF_KEY:
 478                         sref = (struct btrfs_shared_data_ref *)(iref + 1);
 479                         count = btrfs_shared_data_ref_count(leaf, sref);
 480                         ret = add_shared_data_ref(fs_info, offset, count,
 481                                                   key->objectid, key->offset);
 482                         break;
 483                 default:
 484                         btrfs_err(fs_info, "invalid key type in iref");
 485                         ret = -EINVAL;
 486                         break;
 487                 }
 488                 if (ret)
 489                         break;
 490                 ptr += btrfs_extent_inline_ref_size(type);
 491         }
 492         return ret;
 493 }
 494
 495 static int process_leaf(struct btrfs_root *root,
 496                         struct btrfs_path *path, u64 *bytenr, u64 *num_bytes)
 497 {
 498         struct btrfs_fs_info *fs_info = root->fs_info;
 499         struct extent_buffer *leaf = path->nodes[0];
 500         struct btrfs_extent_data_ref *dref;
 501         struct btrfs_shared_data_ref *sref;
 502         u32 count;
 503         int i = 0, tree_block_level = 0, ret;
 504         struct btrfs_key key;
 505         int nritems = btrfs_header_nritems(leaf);
 506
 507         for (i = 0; i < nritems; i++) {
 508                 btrfs_item_key_to_cpu(leaf, &key, i);
 509                 switch (key.type) {
 510                 case BTRFS_EXTENT_ITEM_KEY:
 511                         *num_bytes = key.offset;
 512                 case BTRFS_METADATA_ITEM_KEY:
 513                         *bytenr = key.objectid;
 514                         ret = process_extent_item(fs_info, path, &key, i,
 515                                                   &tree_block_level);
 516                         break;
 517                 case BTRFS_TREE_BLOCK_REF_KEY:
 518                         ret = add_tree_block(fs_info, key.offset, 0,
 519                                              key.objectid, tree_block_level);
 520                         break;
 521                 case BTRFS_SHARED_BLOCK_REF_KEY:
 522                         ret = add_tree_block(fs_info, 0, key.offset,
 523                                              key.objectid, tree_block_level);
 524                         break;
 525                 case BTRFS_EXTENT_DATA_REF_KEY:
 526                         dref = btrfs_item_ptr(leaf, i,
 527                                               struct btrfs_extent_data_ref);
 528                         ret = add_extent_data_ref(fs_info, leaf, dref, *bytenr,
 529                                                   *num_bytes);
 530                         break;
 531                 case BTRFS_SHARED_DATA_REF_KEY:
 532                         sref = btrfs_item_ptr(leaf, i,
 533                                               struct btrfs_shared_data_ref);
 534                         count = btrfs_shared_data_ref_count(leaf, sref);
 535                         ret = add_shared_data_ref(fs_info, key.offset, count,
 536                                                   *bytenr, *num_bytes);
 537                         break;
 538                 default:
 539                         break;
 540                 }
 541                 if (ret)
 542                         break;
 543         }
 544         return ret;
 545 }
 546
 547 /* Walk down to the leaf from the given level */
 548 static int walk_down_tree(struct btrfs_root *root, struct btrfs_path *path,
 549                           int level, u64 *bytenr, u64 *num_bytes)
 550 {
 551         struct btrfs_fs_info *fs_info = root->fs_info;
 552         struct extent_buffer *eb;
 553         u64 block_bytenr, gen;
 554         int ret = 0;
 555
 556         while (level >= 0) {
 557                 if (level) {
 558                         struct btrfs_key first_key;
 559
 560                         block_bytenr = btrfs_node_blockptr(path->nodes[level],
 561                                                            path->slots[level]);
 562                         gen = btrfs_node_ptr_generation(path->nodes[level],
 563                                                         path->slots[level]);
 564                         btrfs_node_key_to_cpu(path->nodes[level], &first_key,
 565                                               path->slots[level]);
 566                         eb = read_tree_block(fs_info, block_bytenr, gen,
 567                                              level - 1, &first_key);
 568                         if (IS_ERR(eb))
 569                                 return PTR_ERR(eb);
 570                         if (!extent_buffer_uptodate(eb)) {
 571                                 free_extent_buffer(eb);
 572                                 return -EIO;
 573                         }
 574                         btrfs_tree_read_lock(eb);
 575                         btrfs_set_lock_blocking_read(eb);
 576                         path->nodes[level-1] = eb;
 577                         path->slots[level-1] = 0;
 578                         path->locks[level-1] = BTRFS_READ_LOCK_BLOCKING;
 579                 } else {
 580                         ret = process_leaf(root, path, bytenr, num_bytes);
 581                         if (ret)
 582                                 break;
 583                 }
 584                 level--;
 585         }
 586         return ret;
 587 }
 588
 589 /* Walk up to the next node that needs to be processed */
 590 static int walk_up_tree(struct btrfs_path *path, int *level)
 591 {
 592         int l;
 593
 594         for (l = 0; l < BTRFS_MAX_LEVEL; l++) {
 595                 if (!path->nodes[l])
 596                         continue;
 597                 if (l) {
 598                         path->slots[l]++;
 599                         if (path->slots[l] <
 600                             btrfs_header_nritems(path->nodes[l])) {
 601                                 *level = l;
 602                                 return 0;
 603                         }
 604                 }
 605                 btrfs_tree_unlock_rw(path->nodes[l], path->locks[l]);
 606                 free_extent_buffer(path->nodes[l]);
 607                 path->nodes[l] = NULL;
 608                 path->slots[l] = 0;
 609                 path->locks[l] = 0;
 610         }
 611
 612         return 1;
 613 }
 614
 615 static void dump_ref_action(struct btrfs_fs_info *fs_info,
 616                             struct ref_action *ra)
 617 {
 618         btrfs_err(fs_info,
 619 "  Ref action %d, root %llu, ref_root %llu, parent %llu, owner %llu, offset %llu, num_refs %llu",
 620                   ra->action, ra->root, ra->ref.root_objectid, ra->ref.parent,
 621                   ra->ref.owner, ra->ref.offset, ra->ref.num_refs);
 622         __print_stack_trace(fs_info, ra);
 623 }
 624
 625 /*
 626  * Dumps all the information from the block entry to printk, it's going to be
 627  * awesome.
 628  */
 629 static void dump_block_entry(struct btrfs_fs_info *fs_info,
 630                              struct block_entry *be)
 631 {
 632         struct ref_entry *ref;
 633         struct root_entry *re;
 634         struct ref_action *ra;
 635         struct rb_node *n;
 636
 637         btrfs_err(fs_info,
 638 "dumping block entry [%llu %llu], num_refs %llu, metadata %d, from disk %d",
 639                   be->bytenr, be->len, be->num_refs, be->metadata,
 640                   be->from_disk);
 641
 642         for (n = rb_first(&be->refs); n; n = rb_next(n)) {
 643                 ref = rb_entry(n, struct ref_entry, node);
 644                 btrfs_err(fs_info,
 645 "  ref root %llu, parent %llu, owner %llu, offset %llu, num_refs %llu",
 646                           ref->root_objectid, ref->parent, ref->owner,
 647                           ref->offset, ref->num_refs);
 648         }
 649
 650         for (n = rb_first(&be->roots); n; n = rb_next(n)) {
 651                 re = rb_entry(n, struct root_entry, node);
 652                 btrfs_err(fs_info, "  root entry %llu, num_refs %llu",
 653                           re->root_objectid, re->num_refs);
 654         }
 655
 656         list_for_each_entry(ra, &be->actions, list)
 657                 dump_ref_action(fs_info, ra);
 658 }
 659
 660 /*
 661  * btrfs_ref_tree_mod: called when we modify a ref for a bytenr
 662  * @root: the root we are making this modification from.
 663  * @bytenr: the bytenr we are modifying.
 664  * @num_bytes: number of bytes.
 665  * @parent: the parent bytenr.
 666  * @ref_root: the original root owner of the bytenr.
 667  * @owner: level in the case of metadata, inode in the case of data.
 668  * @offset: 0 for metadata, file offset for data.
 669  * @action: the action that we are doing, this is the same as the delayed ref
 670  *      action.
 671  *
 672  * This will add an action item to the given bytenr and do sanity checks to make
 673  * sure we haven't messed something up.  If we are making a new allocation and
 674  * this block entry has history we will delete all previous actions as long as
 675  * our sanity checks pass as they are no longer needed.
 676  */
 677 int btrfs_ref_tree_mod(struct btrfs_root *root, u64 bytenr, u64 num_bytes,
 678                        u64 parent, u64 ref_root, u64 owner, u64 offset,
 679                        int action)
 680 {
 681         struct btrfs_fs_info *fs_info = root->fs_info;
 682         struct ref_entry *ref = NULL, *exist;
 683         struct ref_action *ra = NULL;
 684         struct block_entry *be = NULL;
 685         struct root_entry *re = NULL;
 686         int ret = 0;
 687         bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
 688
 689         if (!btrfs_test_opt(root->fs_info, REF_VERIFY))
 690                 return 0;
 691
 692         ref = kzalloc(sizeof(struct ref_entry), GFP_NOFS);
 693         ra = kmalloc(sizeof(struct ref_action), GFP_NOFS);
 694         if (!ra || !ref) {
 695                 kfree(ref);
 696                 kfree(ra);
 697                 ret = -ENOMEM;
 698                 goto out;
 699         }
 700
 701         if (parent) {
 702                 ref->parent = parent;
 703         } else {
 704                 ref->root_objectid = ref_root;
 705                 ref->owner = owner;
 706                 ref->offset = offset;
 707         }
 708         ref->num_refs = (action == BTRFS_DROP_DELAYED_REF) ? -1 : 1;
 709
 710         memcpy(&ra->ref, ref, sizeof(struct ref_entry));
 711         /*
 712          * Save the extra info from the delayed ref in the ref action to make it
 713          * easier to figure out what is happening.  The real ref's we add to the
 714          * ref tree need to reflect what we save on disk so it matches any
 715          * on-disk refs we pre-loaded.
 716          */
 717         ra->ref.owner = owner;
 718         ra->ref.offset = offset;
 719         ra->ref.root_objectid = ref_root;
 720         __save_stack_trace(ra);
 721
 722         INIT_LIST_HEAD(&ra->list);
 723         ra->action = action;
 724         ra->root = root->root_key.objectid;
 725
 726         /*
 727          * This is an allocation, preallocate the block_entry in case we haven't
 728          * used it before.
 729          */
 730         ret = -EINVAL;
 731         if (action == BTRFS_ADD_DELAYED_EXTENT) {
 732                 /*
 733                  * For subvol_create we'll just pass in whatever the parent root
 734                  * is and the new root objectid, so let's not treat the passed
 735                  * in root as if it really has a ref for this bytenr.
 736                  */
 737                 be = add_block_entry(root->fs_info, bytenr, num_bytes, ref_root);
 738                 if (IS_ERR(be)) {
 739                         kfree(ra);
 740                         ret = PTR_ERR(be);
 741                         goto out;
 742                 }
 743                 be->num_refs++;
 744                 if (metadata)
 745                         be->metadata = 1;
 746
 747                 if (be->num_refs != 1) {
 748                         btrfs_err(fs_info,
 749                         "re-allocated a block that still has references to it!");
 750                         dump_block_entry(fs_info, be);
 751                         dump_ref_action(fs_info, ra);
 752                         goto out_unlock;
 753                 }
 754
 755                 while (!list_empty(&be->actions)) {
 756                         struct ref_action *tmp;
 757
 758                         tmp = list_first_entry(&be->actions, struct ref_action,
 759                                                list);
 760                         list_del(&tmp->list);
 761                         kfree(tmp);
 762                 }
 763         } else {
 764                 struct root_entry *tmp;
 765
 766                 if (!parent) {
 767                         re = kmalloc(sizeof(struct root_entry), GFP_NOFS);
 768                         if (!re) {
 769                                 kfree(ref);
 770                                 kfree(ra);
 771                                 ret = -ENOMEM;
 772                                 goto out;
 773                         }
 774                         /*
 775                          * This is the root that is modifying us, so it's the
 776                          * one we want to lookup below when we modify the
 777                          * re->num_refs.
 778                          */
 779                         ref_root = root->root_key.objectid;
 780                         re->root_objectid = root->root_key.objectid;
 781                         re->num_refs = 0;
 782                 }
 783
 784                 spin_lock(&root->fs_info->ref_verify_lock);
 785                 be = lookup_block_entry(&root->fs_info->block_tree, bytenr);
 786                 if (!be) {
 787                         btrfs_err(fs_info,
 788 "trying to do action %d to bytenr %llu num_bytes %llu but there is no existing entry!",
 789                                   action, (unsigned long long)bytenr,
 790                                   (unsigned long long)num_bytes);
 791                         dump_ref_action(fs_info, ra);
 792                         kfree(ref);
 793                         kfree(ra);
 794                         goto out_unlock;
 795                 }
 796
 797                 if (!parent) {
 798                         tmp = insert_root_entry(&be->roots, re);
 799                         if (tmp) {
 800                                 kfree(re);
 801                                 re = tmp;
 802                         }
 803                 }
 804         }
 805
 806         exist = insert_ref_entry(&be->refs, ref);
 807         if (exist) {
 808                 if (action == BTRFS_DROP_DELAYED_REF) {
 809                         if (exist->num_refs == 0) {
 810                                 btrfs_err(fs_info,
 811 "dropping a ref for a existing root that doesn't have a ref on the block");
 812                                 dump_block_entry(fs_info, be);
 813                                 dump_ref_action(fs_info, ra);
 814                                 kfree(ra);
 815                                 goto out_unlock;
 816                         }
 817                         exist->num_refs--;
 818                         if (exist->num_refs == 0) {
 819                                 rb_erase(&exist->node, &be->refs);
 820                                 kfree(exist);
 821                         }
 822                 } else if (!be->metadata) {
 823                         exist->num_refs++;
 824                 } else {
 825                         btrfs_err(fs_info,
 826 "attempting to add another ref for an existing ref on a tree block");
 827                         dump_block_entry(fs_info, be);
 828                         dump_ref_action(fs_info, ra);
 829                         kfree(ra);
 830                         goto out_unlock;
 831                 }
 832                 kfree(ref);
 833         } else {
 834                 if (action == BTRFS_DROP_DELAYED_REF) {
 835                         btrfs_err(fs_info,
 836 "dropping a ref for a root that doesn't have a ref on the block");
 837                         dump_block_entry(fs_info, be);
 838                         dump_ref_action(fs_info, ra);
 839                         kfree(ra);
 840                         goto out_unlock;
 841                 }
 842         }
 843
 844         if (!parent && !re) {
 845                 re = lookup_root_entry(&be->roots, ref_root);
 846                 if (!re) {
 847                         /*
 848                          * This shouldn't happen because we will add our re
 849                          * above when we lookup the be with !parent, but just in
 850                          * case catch this case so we don't panic because I
 851                          * didn't think of some other corner case.
 852                          */
 853                         btrfs_err(fs_info, "failed to find root %llu for %llu",
 854                                   root->root_key.objectid, be->bytenr);
 855                         dump_block_entry(fs_info, be);
 856                         dump_ref_action(fs_info, ra);
 857                         kfree(ra);
 858                         goto out_unlock;
 859                 }
 860         }
 861         if (action == BTRFS_DROP_DELAYED_REF) {
 862                 if (re)
 863                         re->num_refs--;
 864                 be->num_refs--;
 865         } else if (action == BTRFS_ADD_DELAYED_REF) {
 866                 be->num_refs++;
 867                 if (re)
 868                         re->num_refs++;
 869         }
 870         list_add_tail(&ra->list, &be->actions);
 871         ret = 0;
 872 out_unlock:
 873         spin_unlock(&root->fs_info->ref_verify_lock);
 874 out:
 875         if (ret)
 876                 btrfs_clear_opt(fs_info->mount_opt, REF_VERIFY);
 877         return ret;
 878 }
 879
 880 /* Free up the ref cache */
 881 void btrfs_free_ref_cache(struct btrfs_fs_info *fs_info)
 882 {
 883         struct block_entry *be;
 884         struct rb_node *n;
 885
 886         if (!btrfs_test_opt(fs_info, REF_VERIFY))
 887                 return;
 888
 889         spin_lock(&fs_info->ref_verify_lock);
 890         while ((n = rb_first(&fs_info->block_tree))) {
 891                 be = rb_entry(n, struct block_entry, node);
 892                 rb_erase(&be->node, &fs_info->block_tree);
 893                 free_block_entry(be);
 894                 cond_resched_lock(&fs_info->ref_verify_lock);
 895         }
 896         spin_unlock(&fs_info->ref_verify_lock);
 897 }
 898
 899 void btrfs_free_ref_tree_range(struct btrfs_fs_info *fs_info, u64 start,
 900                                u64 len)
 901 {
 902         struct block_entry *be = NULL, *entry;
 903         struct rb_node *n;
 904
 905         if (!btrfs_test_opt(fs_info, REF_VERIFY))
 906                 return;
 907
 908         spin_lock(&fs_info->ref_verify_lock);
 909         n = fs_info->block_tree.rb_node;
 910         while (n) {
 911                 entry = rb_entry(n, struct block_entry, node);
 912                 if (entry->bytenr < start) {
 913                         n = n->rb_right;
 914                 } else if (entry->bytenr > start) {
 915                         n = n->rb_left;
 916                 } else {
 917                         be = entry;
 918                         break;
 919                 }
 920                 /* We want to get as close to start as possible */
 921                 if (be == NULL ||
 922                     (entry->bytenr < start && be->bytenr > start) ||
 923                     (entry->bytenr < start && entry->bytenr > be->bytenr))
 924                         be = entry;
 925         }
 926
 927         /*
 928          * Could have an empty block group, maybe have something to check for
 929          * this case to verify we were actually empty?
 930          */
 931         if (!be) {
 932                 spin_unlock(&fs_info->ref_verify_lock);
 933                 return;
 934         }
 935
 936         n = &be->node;
 937         while (n) {
 938                 be = rb_entry(n, struct block_entry, node);
 939                 n = rb_next(n);
 940                 if (be->bytenr < start && be->bytenr + be->len > start) {
 941                         btrfs_err(fs_info,
 942                                 "block entry overlaps a block group [%llu,%llu]!",
 943                                 start, len);
 944                         dump_block_entry(fs_info, be);
 945                         continue;
 946                 }
 947                 if (be->bytenr < start)
 948                         continue;
 949                 if (be->bytenr >= start + len)
 950                         break;
 951                 if (be->bytenr + be->len > start + len) {
 952                         btrfs_err(fs_info,
 953                                 "block entry overlaps a block group [%llu,%llu]!",
 954                                 start, len);
 955                         dump_block_entry(fs_info, be);
 956                 }
 957                 rb_erase(&be->node, &fs_info->block_tree);
 958                 free_block_entry(be);
 959         }
 960         spin_unlock(&fs_info->ref_verify_lock);
 961 }
 962
 963 /* Walk down all roots and build the ref tree, meant to be called at mount */
 964 int btrfs_build_ref_tree(struct btrfs_fs_info *fs_info)
 965 {
 966         struct btrfs_path *path;
 967         struct extent_buffer *eb;
 968         u64 bytenr = 0, num_bytes = 0;
 969         int ret, level;
 970
 971         if (!btrfs_test_opt(fs_info, REF_VERIFY))
 972                 return 0;
 973
 974         path = btrfs_alloc_path();
 975         if (!path)
 976                 return -ENOMEM;
 977
 978         eb = btrfs_read_lock_root_node(fs_info->extent_root);
 979         btrfs_set_lock_blocking_read(eb);
 980         level = btrfs_header_level(eb);
 981         path->nodes[level] = eb;
 982         path->slots[level] = 0;
 983         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
 984
 985         while (1) {
 986                 /*
 987                  * We have to keep track of the bytenr/num_bytes we last hit
 988                  * because we could have run out of space for an inline ref, and
 989                  * would have had to added a ref key item which may appear on a
 990                  * different leaf from the original extent item.
 991                  */
 992                 ret = walk_down_tree(fs_info->extent_root, path, level,
 993                                      &bytenr, &num_bytes);
 994                 if (ret)
 995                         break;
 996                 ret = walk_up_tree(path, &level);
 997                 if (ret < 0)
 998                         break;
 999                 if (ret > 0) {
1000                         ret = 0;
1001                         break;
1002                 }
1003         }
1004         if (ret) {
1005                 btrfs_clear_opt(fs_info->mount_opt, REF_VERIFY);
1006                 btrfs_free_ref_cache(fs_info);
1007         }
1008         btrfs_free_path(path);
1009         return ret;
1010 }