fs/btrfs/check-integrity.c

   1 /*
   2  * Copyright (C) STRATO AG 2011.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 /*
  20  * This module can be used to catch cases when the btrfs kernel
  21  * code executes write requests to the disk that bring the file
  22  * system in an inconsistent state. In such a state, a power-loss
  23  * or kernel panic event would cause that the data on disk is
  24  * lost or at least damaged.
  25  *
  26  * Code is added that examines all block write requests during
  27  * runtime (including writes of the super block). Three rules
  28  * are verified and an error is printed on violation of the
  29  * rules:
  30  * 1. It is not allowed to write a disk block which is
  31  *    currently referenced by the super block (either directly
  32  *    or indirectly).
  33  * 2. When a super block is written, it is verified that all
  34  *    referenced (directly or indirectly) blocks fulfill the
  35  *    following requirements:
  36  *    2a. All referenced blocks have either been present when
  37  *        the file system was mounted, (i.e., they have been
  38  *        referenced by the super block) or they have been
  39  *        written since then and the write completion callback
  40  *        was called and a FLUSH request to the device where
  41  *        these blocks are located was received and completed.
  42  *    2b. All referenced blocks need to have a generation
  43  *        number which is equal to the parent's number.
  44  *
  45  * One issue that was found using this module was that the log
  46  * tree on disk became temporarily corrupted because disk blocks
  47  * that had been in use for the log tree had been freed and
  48  * reused too early, while being referenced by the written super
  49  * block.
  50  *
  51  * The search term in the kernel log that can be used to filter
  52  * on the existence of detected integrity issues is
  53  * "btrfs: attempt".
  54  *
  55  * The integrity check is enabled via mount options. These
  56  * mount options are only supported if the integrity check
  57  * tool is compiled by defining BTRFS_FS_CHECK_INTEGRITY.
  58  *
  59  * Example #1, apply integrity checks to all metadata:
  60  * mount /dev/sdb1 /mnt -o check_int
  61  *
  62  * Example #2, apply integrity checks to all metadata and
  63  * to data extents:
  64  * mount /dev/sdb1 /mnt -o check_int_data
  65  *
  66  * Example #3, apply integrity checks to all metadata and dump
  67  * the tree that the super block references to kernel messages
  68  * each time after a super block was written:
  69  * mount /dev/sdb1 /mnt -o check_int,check_int_print_mask=263
  70  *
  71  * If the integrity check tool is included and activated in
  72  * the mount options, plenty of kernel memory is used, and
  73  * plenty of additional CPU cycles are spent. Enabling this
  74  * functionality is not intended for normal use. In most
  75  * cases, unless you are a btrfs developer who needs to verify
  76  * the integrity of (super)-block write requests, do not
  77  * enable the config option BTRFS_FS_CHECK_INTEGRITY to
  78  * include and compile the integrity check tool.
  79  */
  80
  81 #include <linux/sched.h>
  82 #include <linux/slab.h>
  83 #include <linux/buffer_head.h>
  84 #include <linux/mutex.h>
  85 #include <linux/crc32c.h>
  86 #include <linux/genhd.h>
  87 #include <linux/blkdev.h>
  88 #include "ctree.h"
  89 #include "disk-io.h"
  90 #include "transaction.h"
  91 #include "extent_io.h"
  92 #include "volumes.h"
  93 #include "print-tree.h"
  94 #include "locking.h"
  95 #include "check-integrity.h"
  96
  97 #define BTRFSIC_BLOCK_HASHTABLE_SIZE 0x10000
  98 #define BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE 0x10000
  99 #define BTRFSIC_DEV2STATE_HASHTABLE_SIZE 0x100
 100 #define BTRFSIC_BLOCK_MAGIC_NUMBER 0x14491051
 101 #define BTRFSIC_BLOCK_LINK_MAGIC_NUMBER 0x11070807
 102 #define BTRFSIC_DEV2STATE_MAGIC_NUMBER 0x20111530
 103 #define BTRFSIC_BLOCK_STACK_FRAME_MAGIC_NUMBER 20111300
 104 #define BTRFSIC_TREE_DUMP_MAX_INDENT_LEVEL (200 - 6)    /* in characters,
 105                                                          * excluding " [...]" */
 106 #define BTRFSIC_BLOCK_SIZE PAGE_SIZE
 107
 108 #define BTRFSIC_GENERATION_UNKNOWN ((u64)-1)
 109
 110 /*
 111  * The definition of the bitmask fields for the print_mask.
 112  * They are specified with the mount option check_integrity_print_mask.
 113  */
 114 #define BTRFSIC_PRINT_MASK_SUPERBLOCK_WRITE                     0x00000001
 115 #define BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION         0x00000002
 116 #define BTRFSIC_PRINT_MASK_TREE_AFTER_SB_WRITE                  0x00000004
 117 #define BTRFSIC_PRINT_MASK_TREE_BEFORE_SB_WRITE                 0x00000008
 118 #define BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH                        0x00000010
 119 #define BTRFSIC_PRINT_MASK_END_IO_BIO_BH                        0x00000020
 120 #define BTRFSIC_PRINT_MASK_VERBOSE                              0x00000040
 121 #define BTRFSIC_PRINT_MASK_VERY_VERBOSE                         0x00000080
 122 #define BTRFSIC_PRINT_MASK_INITIAL_TREE                         0x00000100
 123 #define BTRFSIC_PRINT_MASK_INITIAL_ALL_TREES                    0x00000200
 124 #define BTRFSIC_PRINT_MASK_INITIAL_DATABASE                     0x00000400
 125 #define BTRFSIC_PRINT_MASK_NUM_COPIES                           0x00000800
 126 #define BTRFSIC_PRINT_MASK_TREE_WITH_ALL_MIRRORS                0x00001000
 127
 128 struct btrfsic_dev_state;
 129 struct btrfsic_state;
 130
 131 struct btrfsic_block {
 132         u32 magic_num;          /* only used for debug purposes */
 133         unsigned int is_metadata:1;     /* if it is meta-data, not data-data */
 134         unsigned int is_superblock:1;   /* if it is one of the superblocks */
 135         unsigned int is_iodone:1;       /* if is done by lower subsystem */
 136         unsigned int iodone_w_error:1;  /* error was indicated to endio */
 137         unsigned int never_written:1;   /* block was added because it was
 138                                          * referenced, not because it was
 139                                          * written */
 140         unsigned int mirror_num:2;      /* large enough to hold
 141                                          * BTRFS_SUPER_MIRROR_MAX */
 142         struct btrfsic_dev_state *dev_state;
 143         u64 dev_bytenr;         /* key, physical byte num on disk */
 144         u64 logical_bytenr;     /* logical byte num on disk */
 145         u64 generation;
 146         struct btrfs_disk_key disk_key; /* extra info to print in case of
 147                                          * issues, will not always be correct */
 148         struct list_head collision_resolving_node;      /* list node */
 149         struct list_head all_blocks_node;       /* list node */
 150
 151         /* the following two lists contain block_link items */
 152         struct list_head ref_to_list;   /* list */
 153         struct list_head ref_from_list; /* list */
 154         struct btrfsic_block *next_in_same_bio;
 155         void *orig_bio_bh_private;
 156         union {
 157                 bio_end_io_t *bio;
 158                 bh_end_io_t *bh;
 159         } orig_bio_bh_end_io;
 160         int submit_bio_bh_rw;
 161         u64 flush_gen; /* only valid if !never_written */
 162 };
 163
 164 /*
 165  * Elements of this type are allocated dynamically and required because
 166  * each block object can refer to and can be ref from multiple blocks.
 167  * The key to lookup them in the hashtable is the dev_bytenr of
 168  * the block ref to plus the one from the block refered from.
 169  * The fact that they are searchable via a hashtable and that a
 170  * ref_cnt is maintained is not required for the btrfs integrity
 171  * check algorithm itself, it is only used to make the output more
 172  * beautiful in case that an error is detected (an error is defined
 173  * as a write operation to a block while that block is still referenced).
 174  */
 175 struct btrfsic_block_link {
 176         u32 magic_num;          /* only used for debug purposes */
 177         u32 ref_cnt;
 178         struct list_head node_ref_to;   /* list node */
 179         struct list_head node_ref_from; /* list node */
 180         struct list_head collision_resolving_node;      /* list node */
 181         struct btrfsic_block *block_ref_to;
 182         struct btrfsic_block *block_ref_from;
 183         u64 parent_generation;
 184 };
 185
 186 struct btrfsic_dev_state {
 187         u32 magic_num;          /* only used for debug purposes */
 188         struct block_device *bdev;
 189         struct btrfsic_state *state;
 190         struct list_head collision_resolving_node;      /* list node */
 191         struct btrfsic_block dummy_block_for_bio_bh_flush;
 192         u64 last_flush_gen;
 193         char name[BDEVNAME_SIZE];
 194 };
 195
 196 struct btrfsic_block_hashtable {
 197         struct list_head table[BTRFSIC_BLOCK_HASHTABLE_SIZE];
 198 };
 199
 200 struct btrfsic_block_link_hashtable {
 201         struct list_head table[BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE];
 202 };
 203
 204 struct btrfsic_dev_state_hashtable {
 205         struct list_head table[BTRFSIC_DEV2STATE_HASHTABLE_SIZE];
 206 };
 207
 208 struct btrfsic_block_data_ctx {
 209         u64 start;              /* virtual bytenr */
 210         u64 dev_bytenr;         /* physical bytenr on device */
 211         u32 len;
 212         struct btrfsic_dev_state *dev;
 213         char *data;
 214         struct buffer_head *bh; /* do not use if set to NULL */
 215 };
 216
 217 /* This structure is used to implement recursion without occupying
 218  * any stack space, refer to btrfsic_process_metablock() */
 219 struct btrfsic_stack_frame {
 220         u32 magic;
 221         u32 nr;
 222         int error;
 223         int i;
 224         int limit_nesting;
 225         int num_copies;
 226         int mirror_num;
 227         struct btrfsic_block *block;
 228         struct btrfsic_block_data_ctx *block_ctx;
 229         struct btrfsic_block *next_block;
 230         struct btrfsic_block_data_ctx next_block_ctx;
 231         struct btrfs_header *hdr;
 232         struct btrfsic_stack_frame *prev;
 233 };
 234
 235 /* Some state per mounted filesystem */
 236 struct btrfsic_state {
 237         u32 print_mask;
 238         int include_extent_data;
 239         int csum_size;
 240         struct list_head all_blocks_list;
 241         struct btrfsic_block_hashtable block_hashtable;
 242         struct btrfsic_block_link_hashtable block_link_hashtable;
 243         struct btrfs_root *root;
 244         u64 max_superblock_generation;
 245         struct btrfsic_block *latest_superblock;
 246 };
 247
 248 static void btrfsic_block_init(struct btrfsic_block *b);
 249 static struct btrfsic_block *btrfsic_block_alloc(void);
 250 static void btrfsic_block_free(struct btrfsic_block *b);
 251 static void btrfsic_block_link_init(struct btrfsic_block_link *n);
 252 static struct btrfsic_block_link *btrfsic_block_link_alloc(void);
 253 static void btrfsic_block_link_free(struct btrfsic_block_link *n);
 254 static void btrfsic_dev_state_init(struct btrfsic_dev_state *ds);
 255 static struct btrfsic_dev_state *btrfsic_dev_state_alloc(void);
 256 static void btrfsic_dev_state_free(struct btrfsic_dev_state *ds);
 257 static void btrfsic_block_hashtable_init(struct btrfsic_block_hashtable *h);
 258 static void btrfsic_block_hashtable_add(struct btrfsic_block *b,
 259                                         struct btrfsic_block_hashtable *h);
 260 static void btrfsic_block_hashtable_remove(struct btrfsic_block *b);
 261 static struct btrfsic_block *btrfsic_block_hashtable_lookup(
 262                 struct block_device *bdev,
 263                 u64 dev_bytenr,
 264                 struct btrfsic_block_hashtable *h);
 265 static void btrfsic_block_link_hashtable_init(
 266                 struct btrfsic_block_link_hashtable *h);
 267 static void btrfsic_block_link_hashtable_add(
 268                 struct btrfsic_block_link *l,
 269                 struct btrfsic_block_link_hashtable *h);
 270 static void btrfsic_block_link_hashtable_remove(struct btrfsic_block_link *l);
 271 static struct btrfsic_block_link *btrfsic_block_link_hashtable_lookup(
 272                 struct block_device *bdev_ref_to,
 273                 u64 dev_bytenr_ref_to,
 274                 struct block_device *bdev_ref_from,
 275                 u64 dev_bytenr_ref_from,
 276                 struct btrfsic_block_link_hashtable *h);
 277 static void btrfsic_dev_state_hashtable_init(
 278                 struct btrfsic_dev_state_hashtable *h);
 279 static void btrfsic_dev_state_hashtable_add(
 280                 struct btrfsic_dev_state *ds,
 281                 struct btrfsic_dev_state_hashtable *h);
 282 static void btrfsic_dev_state_hashtable_remove(struct btrfsic_dev_state *ds);
 283 static struct btrfsic_dev_state *btrfsic_dev_state_hashtable_lookup(
 284                 struct block_device *bdev,
 285                 struct btrfsic_dev_state_hashtable *h);
 286 static struct btrfsic_stack_frame *btrfsic_stack_frame_alloc(void);
 287 static void btrfsic_stack_frame_free(struct btrfsic_stack_frame *sf);
 288 static int btrfsic_process_superblock(struct btrfsic_state *state,
 289                                       struct btrfs_fs_devices *fs_devices);
 290 static int btrfsic_process_metablock(struct btrfsic_state *state,
 291                                      struct btrfsic_block *block,
 292                                      struct btrfsic_block_data_ctx *block_ctx,
 293                                      struct btrfs_header *hdr,
 294                                      int limit_nesting, int force_iodone_flag);
 295 static int btrfsic_create_link_to_next_block(
 296                 struct btrfsic_state *state,
 297                 struct btrfsic_block *block,
 298                 struct btrfsic_block_data_ctx
 299                 *block_ctx, u64 next_bytenr,
 300                 int limit_nesting,
 301                 struct btrfsic_block_data_ctx *next_block_ctx,
 302                 struct btrfsic_block **next_blockp,
 303                 int force_iodone_flag,
 304                 int *num_copiesp, int *mirror_nump,
 305                 struct btrfs_disk_key *disk_key,
 306                 u64 parent_generation);
 307 static int btrfsic_handle_extent_data(struct btrfsic_state *state,
 308                                       struct btrfsic_block *block,
 309                                       struct btrfsic_block_data_ctx *block_ctx,
 310                                       u32 item_offset, int force_iodone_flag);
 311 static int btrfsic_map_block(struct btrfsic_state *state, u64 bytenr, u32 len,
 312                              struct btrfsic_block_data_ctx *block_ctx_out,
 313                              int mirror_num);
 314 static int btrfsic_map_superblock(struct btrfsic_state *state, u64 bytenr,
 315                                   u32 len, struct block_device *bdev,
 316                                   struct btrfsic_block_data_ctx *block_ctx_out);
 317 static void btrfsic_release_block_ctx(struct btrfsic_block_data_ctx *block_ctx);
 318 static int btrfsic_read_block(struct btrfsic_state *state,
 319                               struct btrfsic_block_data_ctx *block_ctx);
 320 static void btrfsic_dump_database(struct btrfsic_state *state);
 321 static int btrfsic_test_for_metadata(struct btrfsic_state *state,
 322                                      const u8 *data, unsigned int size);
 323 static void btrfsic_process_written_block(struct btrfsic_dev_state *dev_state,
 324                                           u64 dev_bytenr, u8 *mapped_data,
 325                                           unsigned int len, struct bio *bio,
 326                                           int *bio_is_patched,
 327                                           struct buffer_head *bh,
 328                                           int submit_bio_bh_rw);
 329 static int btrfsic_process_written_superblock(
 330                 struct btrfsic_state *state,
 331                 struct btrfsic_block *const block,
 332                 struct btrfs_super_block *const super_hdr);
 333 static void btrfsic_bio_end_io(struct bio *bp, int bio_error_status);
 334 static void btrfsic_bh_end_io(struct buffer_head *bh, int uptodate);
 335 static int btrfsic_is_block_ref_by_superblock(const struct btrfsic_state *state,
 336                                               const struct btrfsic_block *block,
 337                                               int recursion_level);
 338 static int btrfsic_check_all_ref_blocks(struct btrfsic_state *state,
 339                                         struct btrfsic_block *const block,
 340                                         int recursion_level);
 341 static void btrfsic_print_add_link(const struct btrfsic_state *state,
 342                                    const struct btrfsic_block_link *l);
 343 static void btrfsic_print_rem_link(const struct btrfsic_state *state,
 344                                    const struct btrfsic_block_link *l);
 345 static char btrfsic_get_block_type(const struct btrfsic_state *state,
 346                                    const struct btrfsic_block *block);
 347 static void btrfsic_dump_tree(const struct btrfsic_state *state);
 348 static void btrfsic_dump_tree_sub(const struct btrfsic_state *state,
 349                                   const struct btrfsic_block *block,
 350                                   int indent_level);
 351 static struct btrfsic_block_link *btrfsic_block_link_lookup_or_add(
 352                 struct btrfsic_state *state,
 353                 struct btrfsic_block_data_ctx *next_block_ctx,
 354                 struct btrfsic_block *next_block,
 355                 struct btrfsic_block *from_block,
 356                 u64 parent_generation);
 357 static struct btrfsic_block *btrfsic_block_lookup_or_add(
 358                 struct btrfsic_state *state,
 359                 struct btrfsic_block_data_ctx *block_ctx,
 360                 const char *additional_string,
 361                 int is_metadata,
 362                 int is_iodone,
 363                 int never_written,
 364                 int mirror_num,
 365                 int *was_created);
 366 static int btrfsic_process_superblock_dev_mirror(
 367                 struct btrfsic_state *state,
 368                 struct btrfsic_dev_state *dev_state,
 369                 struct btrfs_device *device,
 370                 int superblock_mirror_num,
 371                 struct btrfsic_dev_state **selected_dev_state,
 372                 struct btrfs_super_block *selected_super);
 373 static struct btrfsic_dev_state *btrfsic_dev_state_lookup(
 374                 struct block_device *bdev);
 375 static void btrfsic_cmp_log_and_dev_bytenr(struct btrfsic_state *state,
 376                                            u64 bytenr,
 377                                            struct btrfsic_dev_state *dev_state,
 378                                            u64 dev_bytenr, char *data);
 379
 380 static struct mutex btrfsic_mutex;
 381 static int btrfsic_is_initialized;
 382 static struct btrfsic_dev_state_hashtable btrfsic_dev_state_hashtable;
 383
 384
 385 static void btrfsic_block_init(struct btrfsic_block *b)
 386 {
 387         b->magic_num = BTRFSIC_BLOCK_MAGIC_NUMBER;
 388         b->dev_state = NULL;
 389         b->dev_bytenr = 0;
 390         b->logical_bytenr = 0;
 391         b->generation = BTRFSIC_GENERATION_UNKNOWN;
 392         b->disk_key.objectid = 0;
 393         b->disk_key.type = 0;
 394         b->disk_key.offset = 0;
 395         b->is_metadata = 0;
 396         b->is_superblock = 0;
 397         b->is_iodone = 0;
 398         b->iodone_w_error = 0;
 399         b->never_written = 0;
 400         b->mirror_num = 0;
 401         b->next_in_same_bio = NULL;
 402         b->orig_bio_bh_private = NULL;
 403         b->orig_bio_bh_end_io.bio = NULL;
 404         INIT_LIST_HEAD(&b->collision_resolving_node);
 405         INIT_LIST_HEAD(&b->all_blocks_node);
 406         INIT_LIST_HEAD(&b->ref_to_list);
 407         INIT_LIST_HEAD(&b->ref_from_list);
 408         b->submit_bio_bh_rw = 0;
 409         b->flush_gen = 0;
 410 }
 411
 412 static struct btrfsic_block *btrfsic_block_alloc(void)
 413 {
 414         struct btrfsic_block *b;
 415
 416         b = kzalloc(sizeof(*b), GFP_NOFS);
 417         if (NULL != b)
 418                 btrfsic_block_init(b);
 419
 420         return b;
 421 }
 422
 423 static void btrfsic_block_free(struct btrfsic_block *b)
 424 {
 425         BUG_ON(!(NULL == b || BTRFSIC_BLOCK_MAGIC_NUMBER == b->magic_num));
 426         kfree(b);
 427 }
 428
 429 static void btrfsic_block_link_init(struct btrfsic_block_link *l)
 430 {
 431         l->magic_num = BTRFSIC_BLOCK_LINK_MAGIC_NUMBER;
 432         l->ref_cnt = 1;
 433         INIT_LIST_HEAD(&l->node_ref_to);
 434         INIT_LIST_HEAD(&l->node_ref_from);
 435         INIT_LIST_HEAD(&l->collision_resolving_node);
 436         l->block_ref_to = NULL;
 437         l->block_ref_from = NULL;
 438 }
 439
 440 static struct btrfsic_block_link *btrfsic_block_link_alloc(void)
 441 {
 442         struct btrfsic_block_link *l;
 443
 444         l = kzalloc(sizeof(*l), GFP_NOFS);
 445         if (NULL != l)
 446                 btrfsic_block_link_init(l);
 447
 448         return l;
 449 }
 450
 451 static void btrfsic_block_link_free(struct btrfsic_block_link *l)
 452 {
 453         BUG_ON(!(NULL == l || BTRFSIC_BLOCK_LINK_MAGIC_NUMBER == l->magic_num));
 454         kfree(l);
 455 }
 456
 457 static void btrfsic_dev_state_init(struct btrfsic_dev_state *ds)
 458 {
 459         ds->magic_num = BTRFSIC_DEV2STATE_MAGIC_NUMBER;
 460         ds->bdev = NULL;
 461         ds->state = NULL;
 462         ds->name[0] = '\0';
 463         INIT_LIST_HEAD(&ds->collision_resolving_node);
 464         ds->last_flush_gen = 0;
 465         btrfsic_block_init(&ds->dummy_block_for_bio_bh_flush);
 466         ds->dummy_block_for_bio_bh_flush.is_iodone = 1;
 467         ds->dummy_block_for_bio_bh_flush.dev_state = ds;
 468 }
 469
 470 static struct btrfsic_dev_state *btrfsic_dev_state_alloc(void)
 471 {
 472         struct btrfsic_dev_state *ds;
 473
 474         ds = kzalloc(sizeof(*ds), GFP_NOFS);
 475         if (NULL != ds)
 476                 btrfsic_dev_state_init(ds);
 477
 478         return ds;
 479 }
 480
 481 static void btrfsic_dev_state_free(struct btrfsic_dev_state *ds)
 482 {
 483         BUG_ON(!(NULL == ds ||
 484                  BTRFSIC_DEV2STATE_MAGIC_NUMBER == ds->magic_num));
 485         kfree(ds);
 486 }
 487
 488 static void btrfsic_block_hashtable_init(struct btrfsic_block_hashtable *h)
 489 {
 490         int i;
 491
 492         for (i = 0; i < BTRFSIC_BLOCK_HASHTABLE_SIZE; i++)
 493                 INIT_LIST_HEAD(h->table + i);
 494 }
 495
 496 static void btrfsic_block_hashtable_add(struct btrfsic_block *b,
 497                                         struct btrfsic_block_hashtable *h)
 498 {
 499         const unsigned int hashval =
 500             (((unsigned int)(b->dev_bytenr >> 16)) ^
 501              ((unsigned int)((uintptr_t)b->dev_state->bdev))) &
 502              (BTRFSIC_BLOCK_HASHTABLE_SIZE - 1);
 503
 504         list_add(&b->collision_resolving_node, h->table + hashval);
 505 }
 506
 507 static void btrfsic_block_hashtable_remove(struct btrfsic_block *b)
 508 {
 509         list_del(&b->collision_resolving_node);
 510 }
 511
 512 static struct btrfsic_block *btrfsic_block_hashtable_lookup(
 513                 struct block_device *bdev,
 514                 u64 dev_bytenr,
 515                 struct btrfsic_block_hashtable *h)
 516 {
 517         const unsigned int hashval =
 518             (((unsigned int)(dev_bytenr >> 16)) ^
 519              ((unsigned int)((uintptr_t)bdev))) &
 520              (BTRFSIC_BLOCK_HASHTABLE_SIZE - 1);
 521         struct list_head *elem;
 522
 523         list_for_each(elem, h->table + hashval) {
 524                 struct btrfsic_block *const b =
 525                     list_entry(elem, struct btrfsic_block,
 526                                collision_resolving_node);
 527
 528                 if (b->dev_state->bdev == bdev && b->dev_bytenr == dev_bytenr)
 529                         return b;
 530         }
 531
 532         return NULL;
 533 }
 534
 535 static void btrfsic_block_link_hashtable_init(
 536                 struct btrfsic_block_link_hashtable *h)
 537 {
 538         int i;
 539
 540         for (i = 0; i < BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE; i++)
 541                 INIT_LIST_HEAD(h->table + i);
 542 }
 543
 544 static void btrfsic_block_link_hashtable_add(
 545                 struct btrfsic_block_link *l,
 546                 struct btrfsic_block_link_hashtable *h)
 547 {
 548         const unsigned int hashval =
 549             (((unsigned int)(l->block_ref_to->dev_bytenr >> 16)) ^
 550              ((unsigned int)(l->block_ref_from->dev_bytenr >> 16)) ^
 551              ((unsigned int)((uintptr_t)l->block_ref_to->dev_state->bdev)) ^
 552              ((unsigned int)((uintptr_t)l->block_ref_from->dev_state->bdev)))
 553              & (BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE - 1);
 554
 555         BUG_ON(NULL == l->block_ref_to);
 556         BUG_ON(NULL == l->block_ref_from);
 557         list_add(&l->collision_resolving_node, h->table + hashval);
 558 }
 559
 560 static void btrfsic_block_link_hashtable_remove(struct btrfsic_block_link *l)
 561 {
 562         list_del(&l->collision_resolving_node);
 563 }
 564
 565 static struct btrfsic_block_link *btrfsic_block_link_hashtable_lookup(
 566                 struct block_device *bdev_ref_to,
 567                 u64 dev_bytenr_ref_to,
 568                 struct block_device *bdev_ref_from,
 569                 u64 dev_bytenr_ref_from,
 570                 struct btrfsic_block_link_hashtable *h)
 571 {
 572         const unsigned int hashval =
 573             (((unsigned int)(dev_bytenr_ref_to >> 16)) ^
 574              ((unsigned int)(dev_bytenr_ref_from >> 16)) ^
 575              ((unsigned int)((uintptr_t)bdev_ref_to)) ^
 576              ((unsigned int)((uintptr_t)bdev_ref_from))) &
 577              (BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE - 1);
 578         struct list_head *elem;
 579
 580         list_for_each(elem, h->table + hashval) {
 581                 struct btrfsic_block_link *const l =
 582                     list_entry(elem, struct btrfsic_block_link,
 583                                collision_resolving_node);
 584
 585                 BUG_ON(NULL == l->block_ref_to);
 586                 BUG_ON(NULL == l->block_ref_from);
 587                 if (l->block_ref_to->dev_state->bdev == bdev_ref_to &&
 588                     l->block_ref_to->dev_bytenr == dev_bytenr_ref_to &&
 589                     l->block_ref_from->dev_state->bdev == bdev_ref_from &&
 590                     l->block_ref_from->dev_bytenr == dev_bytenr_ref_from)
 591                         return l;
 592         }
 593
 594         return NULL;
 595 }
 596
 597 static void btrfsic_dev_state_hashtable_init(
 598                 struct btrfsic_dev_state_hashtable *h)
 599 {
 600         int i;
 601
 602         for (i = 0; i < BTRFSIC_DEV2STATE_HASHTABLE_SIZE; i++)
 603                 INIT_LIST_HEAD(h->table + i);
 604 }
 605
 606 static void btrfsic_dev_state_hashtable_add(
 607                 struct btrfsic_dev_state *ds,
 608                 struct btrfsic_dev_state_hashtable *h)
 609 {
 610         const unsigned int hashval =
 611             (((unsigned int)((uintptr_t)ds->bdev)) &
 612              (BTRFSIC_DEV2STATE_HASHTABLE_SIZE - 1));
 613
 614         list_add(&ds->collision_resolving_node, h->table + hashval);
 615 }
 616
 617 static void btrfsic_dev_state_hashtable_remove(struct btrfsic_dev_state *ds)
 618 {
 619         list_del(&ds->collision_resolving_node);
 620 }
 621
 622 static struct btrfsic_dev_state *btrfsic_dev_state_hashtable_lookup(
 623                 struct block_device *bdev,
 624                 struct btrfsic_dev_state_hashtable *h)
 625 {
 626         const unsigned int hashval =
 627             (((unsigned int)((uintptr_t)bdev)) &
 628              (BTRFSIC_DEV2STATE_HASHTABLE_SIZE - 1));
 629         struct list_head *elem;
 630
 631         list_for_each(elem, h->table + hashval) {
 632                 struct btrfsic_dev_state *const ds =
 633                     list_entry(elem, struct btrfsic_dev_state,
 634                                collision_resolving_node);
 635
 636                 if (ds->bdev == bdev)
 637                         return ds;
 638         }
 639
 640         return NULL;
 641 }
 642
 643 static int btrfsic_process_superblock(struct btrfsic_state *state,
 644                                       struct btrfs_fs_devices *fs_devices)
 645 {
 646         int ret = 0;
 647         struct btrfs_super_block *selected_super;
 648         struct list_head *dev_head = &fs_devices->devices;
 649         struct btrfs_device *device;
 650         struct btrfsic_dev_state *selected_dev_state = NULL;
 651         int pass;
 652
 653         BUG_ON(NULL == state);
 654         selected_super = kmalloc(sizeof(*selected_super), GFP_NOFS);
 655         if (NULL == selected_super) {
 656                 printk(KERN_INFO "btrfsic: error, kmalloc failed!\n");
 657                 return -1;
 658         }
 659
 660         list_for_each_entry(device, dev_head, dev_list) {
 661                 int i;
 662                 struct btrfsic_dev_state *dev_state;
 663
 664                 if (!device->bdev || !device->name)
 665                         continue;
 666
 667                 dev_state = btrfsic_dev_state_lookup(device->bdev);
 668                 BUG_ON(NULL == dev_state);
 669                 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 670                         ret = btrfsic_process_superblock_dev_mirror(
 671                                         state, dev_state, device, i,
 672                                         &selected_dev_state, selected_super);
 673                         if (0 != ret && 0 == i) {
 674                                 kfree(selected_super);
 675                                 return ret;
 676                         }
 677                 }
 678         }
 679
 680         if (NULL == state->latest_superblock) {
 681                 printk(KERN_INFO "btrfsic: no superblock found!\n");
 682                 kfree(selected_super);
 683                 return -1;
 684         }
 685
 686         state->csum_size = btrfs_super_csum_size(selected_super);
 687
 688         for (pass = 0; pass < 3; pass++) {
 689                 int num_copies;
 690                 int mirror_num;
 691                 u64 next_bytenr;
 692
 693                 switch (pass) {
 694                 case 0:
 695                         next_bytenr = btrfs_super_root(selected_super);
 696                         if (state->print_mask &
 697                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
 698                                 printk(KERN_INFO "root@%llu\n",
 699                                        (unsigned long long)next_bytenr);
 700                         break;
 701                 case 1:
 702                         next_bytenr = btrfs_super_chunk_root(selected_super);
 703                         if (state->print_mask &
 704                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
 705                                 printk(KERN_INFO "chunk@%llu\n",
 706                                        (unsigned long long)next_bytenr);
 707                         break;
 708                 case 2:
 709                         next_bytenr = btrfs_super_log_root(selected_super);
 710                         if (0 == next_bytenr)
 711                                 continue;
 712                         if (state->print_mask &
 713                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
 714                                 printk(KERN_INFO "log@%llu\n",
 715                                        (unsigned long long)next_bytenr);
 716                         break;
 717                 }
 718
 719                 num_copies =
 720                     btrfs_num_copies(&state->root->fs_info->mapping_tree,
 721                                      next_bytenr, PAGE_SIZE);
 722                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
 723                         printk(KERN_INFO "num_copies(log_bytenr=%llu) = %d\n",
 724                                (unsigned long long)next_bytenr, num_copies);
 725
 726                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
 727                         struct btrfsic_block *next_block;
 728                         struct btrfsic_block_data_ctx tmp_next_block_ctx;
 729                         struct btrfsic_block_link *l;
 730                         struct btrfs_header *hdr;
 731
 732                         ret = btrfsic_map_block(state, next_bytenr, PAGE_SIZE,
 733                                                 &tmp_next_block_ctx,
 734                                                 mirror_num);
 735                         if (ret) {
 736                                 printk(KERN_INFO "btrfsic:"
 737                                        " btrfsic_map_block(root @%llu,"
 738                                        " mirror %d) failed!\n",
 739                                        (unsigned long long)next_bytenr,
 740                                        mirror_num);
 741                                 kfree(selected_super);
 742                                 return -1;
 743                         }
 744
 745                         next_block = btrfsic_block_hashtable_lookup(
 746                                         tmp_next_block_ctx.dev->bdev,
 747                                         tmp_next_block_ctx.dev_bytenr,
 748                                         &state->block_hashtable);
 749                         BUG_ON(NULL == next_block);
 750
 751                         l = btrfsic_block_link_hashtable_lookup(
 752                                         tmp_next_block_ctx.dev->bdev,
 753                                         tmp_next_block_ctx.dev_bytenr,
 754                                         state->latest_superblock->dev_state->
 755                                         bdev,
 756                                         state->latest_superblock->dev_bytenr,
 757                                         &state->block_link_hashtable);
 758                         BUG_ON(NULL == l);
 759
 760                         ret = btrfsic_read_block(state, &tmp_next_block_ctx);
 761                         if (ret < (int)BTRFSIC_BLOCK_SIZE) {
 762                                 printk(KERN_INFO
 763                                        "btrfsic: read @logical %llu failed!\n",
 764                                        (unsigned long long)
 765                                        tmp_next_block_ctx.start);
 766                                 btrfsic_release_block_ctx(&tmp_next_block_ctx);
 767                                 kfree(selected_super);
 768                                 return -1;
 769                         }
 770
 771                         hdr = (struct btrfs_header *)tmp_next_block_ctx.data;
 772                         ret = btrfsic_process_metablock(state,
 773                                                         next_block,
 774                                                         &tmp_next_block_ctx,
 775                                                         hdr,
 776                                                         BTRFS_MAX_LEVEL + 3, 1);
 777                         btrfsic_release_block_ctx(&tmp_next_block_ctx);
 778                 }
 779         }
 780
 781         kfree(selected_super);
 782         return ret;
 783 }
 784
 785 static int btrfsic_process_superblock_dev_mirror(
 786                 struct btrfsic_state *state,
 787                 struct btrfsic_dev_state *dev_state,
 788                 struct btrfs_device *device,
 789                 int superblock_mirror_num,
 790                 struct btrfsic_dev_state **selected_dev_state,
 791                 struct btrfs_super_block *selected_super)
 792 {
 793         struct btrfs_super_block *super_tmp;
 794         u64 dev_bytenr;
 795         struct buffer_head *bh;
 796         struct btrfsic_block *superblock_tmp;
 797         int pass;
 798         struct block_device *const superblock_bdev = device->bdev;
 799
 800         /* super block bytenr is always the unmapped device bytenr */
 801         dev_bytenr = btrfs_sb_offset(superblock_mirror_num);
 802         bh = __bread(superblock_bdev, dev_bytenr / 4096, 4096);
 803         if (NULL == bh)
 804                 return -1;
 805         super_tmp = (struct btrfs_super_block *)
 806             (bh->b_data + (dev_bytenr & 4095));
 807
 808         if (btrfs_super_bytenr(super_tmp) != dev_bytenr ||
 809             strncmp((char *)(&(super_tmp->magic)), BTRFS_MAGIC,
 810                     sizeof(super_tmp->magic)) ||
 811             memcmp(device->uuid, super_tmp->dev_item.uuid, BTRFS_UUID_SIZE)) {
 812                 brelse(bh);
 813                 return 0;
 814         }
 815
 816         superblock_tmp =
 817             btrfsic_block_hashtable_lookup(superblock_bdev,
 818                                            dev_bytenr,
 819                                            &state->block_hashtable);
 820         if (NULL == superblock_tmp) {
 821                 superblock_tmp = btrfsic_block_alloc();
 822                 if (NULL == superblock_tmp) {
 823                         printk(KERN_INFO "btrfsic: error, kmalloc failed!\n");
 824                         brelse(bh);
 825                         return -1;
 826                 }
 827                 /* for superblock, only the dev_bytenr makes sense */
 828                 superblock_tmp->dev_bytenr = dev_bytenr;
 829                 superblock_tmp->dev_state = dev_state;
 830                 superblock_tmp->logical_bytenr = dev_bytenr;
 831                 superblock_tmp->generation = btrfs_super_generation(super_tmp);
 832                 superblock_tmp->is_metadata = 1;
 833                 superblock_tmp->is_superblock = 1;
 834                 superblock_tmp->is_iodone = 1;
 835                 superblock_tmp->never_written = 0;
 836                 superblock_tmp->mirror_num = 1 + superblock_mirror_num;
 837                 if (state->print_mask & BTRFSIC_PRINT_MASK_SUPERBLOCK_WRITE)
 838                         printk(KERN_INFO "New initial S-block (bdev %p, %s)"
 839                                " @%llu (%s/%llu/%d)\n",
 840                                superblock_bdev, device->name,
 841                                (unsigned long long)dev_bytenr,
 842                                dev_state->name,
 843                                (unsigned long long)dev_bytenr,
 844                                superblock_mirror_num);
 845                 list_add(&superblock_tmp->all_blocks_node,
 846                          &state->all_blocks_list);
 847                 btrfsic_block_hashtable_add(superblock_tmp,
 848                                             &state->block_hashtable);
 849         }
 850
 851         /* select the one with the highest generation field */
 852         if (btrfs_super_generation(super_tmp) >
 853             state->max_superblock_generation ||
 854             0 == state->max_superblock_generation) {
 855                 memcpy(selected_super, super_tmp, sizeof(*selected_super));
 856                 *selected_dev_state = dev_state;
 857                 state->max_superblock_generation =
 858                     btrfs_super_generation(super_tmp);
 859                 state->latest_superblock = superblock_tmp;
 860         }
 861
 862         for (pass = 0; pass < 3; pass++) {
 863                 u64 next_bytenr;
 864                 int num_copies;
 865                 int mirror_num;
 866                 const char *additional_string = NULL;
 867                 struct btrfs_disk_key tmp_disk_key;
 868
 869                 tmp_disk_key.type = BTRFS_ROOT_ITEM_KEY;
 870                 tmp_disk_key.offset = 0;
 871                 switch (pass) {
 872                 case 0:
 873                         tmp_disk_key.objectid =
 874                             cpu_to_le64(BTRFS_ROOT_TREE_OBJECTID);
 875                         additional_string = "initial root ";
 876                         next_bytenr = btrfs_super_root(super_tmp);
 877                         break;
 878                 case 1:
 879                         tmp_disk_key.objectid =
 880                             cpu_to_le64(BTRFS_CHUNK_TREE_OBJECTID);
 881                         additional_string = "initial chunk ";
 882                         next_bytenr = btrfs_super_chunk_root(super_tmp);
 883                         break;
 884                 case 2:
 885                         tmp_disk_key.objectid =
 886                             cpu_to_le64(BTRFS_TREE_LOG_OBJECTID);
 887                         additional_string = "initial log ";
 888                         next_bytenr = btrfs_super_log_root(super_tmp);
 889                         if (0 == next_bytenr)
 890                                 continue;
 891                         break;
 892                 }
 893
 894                 num_copies =
 895                     btrfs_num_copies(&state->root->fs_info->mapping_tree,
 896                                      next_bytenr, PAGE_SIZE);
 897                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
 898                         printk(KERN_INFO "num_copies(log_bytenr=%llu) = %d\n",
 899                                (unsigned long long)next_bytenr, num_copies);
 900                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
 901                         struct btrfsic_block *next_block;
 902                         struct btrfsic_block_data_ctx tmp_next_block_ctx;
 903                         struct btrfsic_block_link *l;
 904
 905                         if (btrfsic_map_block(state, next_bytenr, PAGE_SIZE,
 906                                               &tmp_next_block_ctx,
 907                                               mirror_num)) {
 908                                 printk(KERN_INFO "btrfsic: btrfsic_map_block("
 909                                        "bytenr @%llu, mirror %d) failed!\n",
 910                                        (unsigned long long)next_bytenr,
 911                                        mirror_num);
 912                                 brelse(bh);
 913                                 return -1;
 914                         }
 915
 916                         next_block = btrfsic_block_lookup_or_add(
 917                                         state, &tmp_next_block_ctx,
 918                                         additional_string, 1, 1, 0,
 919                                         mirror_num, NULL);
 920                         if (NULL == next_block) {
 921                                 btrfsic_release_block_ctx(&tmp_next_block_ctx);
 922                                 brelse(bh);
 923                                 return -1;
 924                         }
 925
 926                         next_block->disk_key = tmp_disk_key;
 927                         next_block->generation = BTRFSIC_GENERATION_UNKNOWN;
 928                         l = btrfsic_block_link_lookup_or_add(
 929                                         state, &tmp_next_block_ctx,
 930                                         next_block, superblock_tmp,
 931                                         BTRFSIC_GENERATION_UNKNOWN);
 932                         btrfsic_release_block_ctx(&tmp_next_block_ctx);
 933                         if (NULL == l) {
 934                                 brelse(bh);
 935                                 return -1;
 936                         }
 937                 }
 938         }
 939         if (state->print_mask & BTRFSIC_PRINT_MASK_INITIAL_ALL_TREES)
 940                 btrfsic_dump_tree_sub(state, superblock_tmp, 0);
 941
 942         brelse(bh);
 943         return 0;
 944 }
 945
 946 static struct btrfsic_stack_frame *btrfsic_stack_frame_alloc(void)
 947 {
 948         struct btrfsic_stack_frame *sf;
 949
 950         sf = kzalloc(sizeof(*sf), GFP_NOFS);
 951         if (NULL == sf)
 952                 printk(KERN_INFO "btrfsic: alloc memory failed!\n");
 953         else
 954                 sf->magic = BTRFSIC_BLOCK_STACK_FRAME_MAGIC_NUMBER;
 955         return sf;
 956 }
 957
 958 static void btrfsic_stack_frame_free(struct btrfsic_stack_frame *sf)
 959 {
 960         BUG_ON(!(NULL == sf ||
 961                  BTRFSIC_BLOCK_STACK_FRAME_MAGIC_NUMBER == sf->magic));
 962         kfree(sf);
 963 }
 964
 965 static int btrfsic_process_metablock(
 966                 struct btrfsic_state *state,
 967                 struct btrfsic_block *const first_block,
 968                 struct btrfsic_block_data_ctx *const first_block_ctx,
 969                 struct btrfs_header *const first_hdr,
 970                 int first_limit_nesting, int force_iodone_flag)
 971 {
 972         struct btrfsic_stack_frame initial_stack_frame = { 0 };
 973         struct btrfsic_stack_frame *sf;
 974         struct btrfsic_stack_frame *next_stack;
 975
 976         sf = &initial_stack_frame;
 977         sf->error = 0;
 978         sf->i = -1;
 979         sf->limit_nesting = first_limit_nesting;
 980         sf->block = first_block;
 981         sf->block_ctx = first_block_ctx;
 982         sf->next_block = NULL;
 983         sf->hdr = first_hdr;
 984         sf->prev = NULL;
 985
 986 continue_with_new_stack_frame:
 987         sf->block->generation = le64_to_cpu(sf->hdr->generation);
 988         if (0 == sf->hdr->level) {
 989                 struct btrfs_leaf *const leafhdr =
 990                     (struct btrfs_leaf *)sf->hdr;
 991
 992                 if (-1 == sf->i) {
 993                         sf->nr = le32_to_cpu(leafhdr->header.nritems);
 994
 995                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
 996                                 printk(KERN_INFO
 997                                        "leaf %llu items %d generation %llu"
 998                                        " owner %llu\n",
 999                                        (unsigned long long)
1000                                        sf->block_ctx->start,
1001                                        sf->nr,
1002                                        (unsigned long long)
1003                                        le64_to_cpu(leafhdr->header.generation),
1004                                        (unsigned long long)
1005                                        le64_to_cpu(leafhdr->header.owner));
1006                 }
1007
1008 continue_with_current_leaf_stack_frame:
1009                 if (0 == sf->num_copies || sf->mirror_num > sf->num_copies) {
1010                         sf->i++;
1011                         sf->num_copies = 0;
1012                 }
1013
1014                 if (sf->i < sf->nr) {
1015                         struct btrfs_item *disk_item = leafhdr->items + sf->i;
1016                         struct btrfs_disk_key *disk_key = &disk_item->key;
1017                         u8 type;
1018                         const u32 item_offset = le32_to_cpu(disk_item->offset);
1019
1020                         type = disk_key->type;
1021
1022                         if (BTRFS_ROOT_ITEM_KEY == type) {
1023                                 const struct btrfs_root_item *const root_item =
1024                                     (struct btrfs_root_item *)
1025                                     (sf->block_ctx->data +
1026                                      offsetof(struct btrfs_leaf, items) +
1027                                      item_offset);
1028                                 const u64 next_bytenr =
1029                                     le64_to_cpu(root_item->bytenr);
1030
1031                                 sf->error =
1032                                     btrfsic_create_link_to_next_block(
1033                                                 state,
1034                                                 sf->block,
1035                                                 sf->block_ctx,
1036                                                 next_bytenr,
1037                                                 sf->limit_nesting,
1038                                                 &sf->next_block_ctx,
1039                                                 &sf->next_block,
1040                                                 force_iodone_flag,
1041                                                 &sf->num_copies,
1042                                                 &sf->mirror_num,
1043                                                 disk_key,
1044                                                 le64_to_cpu(root_item->
1045                                                 generation));
1046                                 if (sf->error)
1047                                         goto one_stack_frame_backwards;
1048
1049                                 if (NULL != sf->next_block) {
1050                                         struct btrfs_header *const next_hdr =
1051                                             (struct btrfs_header *)
1052                                             sf->next_block_ctx.data;
1053
1054                                         next_stack =
1055                                             btrfsic_stack_frame_alloc();
1056                                         if (NULL == next_stack) {
1057                                                 btrfsic_release_block_ctx(
1058                                                                 &sf->
1059                                                                 next_block_ctx);
1060                                                 goto one_stack_frame_backwards;
1061                                         }
1062
1063                                         next_stack->i = -1;
1064                                         next_stack->block = sf->next_block;
1065                                         next_stack->block_ctx =
1066                                             &sf->next_block_ctx;
1067                                         next_stack->next_block = NULL;
1068                                         next_stack->hdr = next_hdr;
1069                                         next_stack->limit_nesting =
1070                                             sf->limit_nesting - 1;
1071                                         next_stack->prev = sf;
1072                                         sf = next_stack;
1073                                         goto continue_with_new_stack_frame;
1074                                 }
1075                         } else if (BTRFS_EXTENT_DATA_KEY == type &&
1076                                    state->include_extent_data) {
1077                                 sf->error = btrfsic_handle_extent_data(
1078                                                 state,
1079                                                 sf->block,
1080                                                 sf->block_ctx,
1081                                                 item_offset,
1082                                                 force_iodone_flag);
1083                                 if (sf->error)
1084                                         goto one_stack_frame_backwards;
1085                         }
1086
1087                         goto continue_with_current_leaf_stack_frame;
1088                 }
1089         } else {
1090                 struct btrfs_node *const nodehdr = (struct btrfs_node *)sf->hdr;
1091
1092                 if (-1 == sf->i) {
1093                         sf->nr = le32_to_cpu(nodehdr->header.nritems);
1094
1095                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1096                                 printk(KERN_INFO "node %llu level %d items %d"
1097                                        " generation %llu owner %llu\n",
1098                                        (unsigned long long)
1099                                        sf->block_ctx->start,
1100                                        nodehdr->header.level, sf->nr,
1101                                        (unsigned long long)
1102                                        le64_to_cpu(nodehdr->header.generation),
1103                                        (unsigned long long)
1104                                        le64_to_cpu(nodehdr->header.owner));
1105                 }
1106
1107 continue_with_current_node_stack_frame:
1108                 if (0 == sf->num_copies || sf->mirror_num > sf->num_copies) {
1109                         sf->i++;
1110                         sf->num_copies = 0;
1111                 }
1112
1113                 if (sf->i < sf->nr) {
1114                         struct btrfs_key_ptr *disk_key_ptr =
1115                             nodehdr->ptrs + sf->i;
1116                         const u64 next_bytenr =
1117                             le64_to_cpu(disk_key_ptr->blockptr);
1118
1119                         sf->error = btrfsic_create_link_to_next_block(
1120                                         state,
1121                                         sf->block,
1122                                         sf->block_ctx,
1123                                         next_bytenr,
1124                                         sf->limit_nesting,
1125                                         &sf->next_block_ctx,
1126                                         &sf->next_block,
1127                                         force_iodone_flag,
1128                                         &sf->num_copies,
1129                                         &sf->mirror_num,
1130                                         &disk_key_ptr->key,
1131                                         le64_to_cpu(disk_key_ptr->generation));
1132                         if (sf->error)
1133                                 goto one_stack_frame_backwards;
1134
1135                         if (NULL != sf->next_block) {
1136                                 struct btrfs_header *const next_hdr =
1137                                     (struct btrfs_header *)
1138                                     sf->next_block_ctx.data;
1139
1140                                 next_stack = btrfsic_stack_frame_alloc();
1141                                 if (NULL == next_stack)
1142                                         goto one_stack_frame_backwards;
1143
1144                                 next_stack->i = -1;
1145                                 next_stack->block = sf->next_block;
1146                                 next_stack->block_ctx = &sf->next_block_ctx;
1147                                 next_stack->next_block = NULL;
1148                                 next_stack->hdr = next_hdr;
1149                                 next_stack->limit_nesting =
1150                                     sf->limit_nesting - 1;
1151                                 next_stack->prev = sf;
1152                                 sf = next_stack;
1153                                 goto continue_with_new_stack_frame;
1154                         }
1155
1156                         goto continue_with_current_node_stack_frame;
1157                 }
1158         }
1159
1160 one_stack_frame_backwards:
1161         if (NULL != sf->prev) {
1162                 struct btrfsic_stack_frame *const prev = sf->prev;
1163
1164                 /* the one for the initial block is freed in the caller */
1165                 btrfsic_release_block_ctx(sf->block_ctx);
1166
1167                 if (sf->error) {
1168                         prev->error = sf->error;
1169                         btrfsic_stack_frame_free(sf);
1170                         sf = prev;
1171                         goto one_stack_frame_backwards;
1172                 }
1173
1174                 btrfsic_stack_frame_free(sf);
1175                 sf = prev;
1176                 goto continue_with_new_stack_frame;
1177         } else {
1178                 BUG_ON(&initial_stack_frame != sf);
1179         }
1180
1181         return sf->error;
1182 }
1183
1184 static int btrfsic_create_link_to_next_block(
1185                 struct btrfsic_state *state,
1186                 struct btrfsic_block *block,
1187                 struct btrfsic_block_data_ctx *block_ctx,
1188                 u64 next_bytenr,
1189                 int limit_nesting,
1190                 struct btrfsic_block_data_ctx *next_block_ctx,
1191                 struct btrfsic_block **next_blockp,
1192                 int force_iodone_flag,
1193                 int *num_copiesp, int *mirror_nump,
1194                 struct btrfs_disk_key *disk_key,
1195                 u64 parent_generation)
1196 {
1197         struct btrfsic_block *next_block = NULL;
1198         int ret;
1199         struct btrfsic_block_link *l;
1200         int did_alloc_block_link;
1201         int block_was_created;
1202
1203         *next_blockp = NULL;
1204         if (0 == *num_copiesp) {
1205                 *num_copiesp =
1206                     btrfs_num_copies(&state->root->fs_info->mapping_tree,
1207                                      next_bytenr, PAGE_SIZE);
1208                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
1209                         printk(KERN_INFO "num_copies(log_bytenr=%llu) = %d\n",
1210                                (unsigned long long)next_bytenr, *num_copiesp);
1211                 *mirror_nump = 1;
1212         }
1213
1214         if (*mirror_nump > *num_copiesp)
1215                 return 0;
1216
1217         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1218                 printk(KERN_INFO
1219                        "btrfsic_create_link_to_next_block(mirror_num=%d)\n",
1220                        *mirror_nump);
1221         ret = btrfsic_map_block(state, next_bytenr,
1222                                 BTRFSIC_BLOCK_SIZE,
1223                                 next_block_ctx, *mirror_nump);
1224         if (ret) {
1225                 printk(KERN_INFO
1226                        "btrfsic: btrfsic_map_block(@%llu, mirror=%d) failed!\n",
1227                        (unsigned long long)next_bytenr, *mirror_nump);
1228                 btrfsic_release_block_ctx(next_block_ctx);
1229                 *next_blockp = NULL;
1230                 return -1;
1231         }
1232
1233         next_block = btrfsic_block_lookup_or_add(state,
1234                                                  next_block_ctx, "referenced ",
1235                                                  1, force_iodone_flag,
1236                                                  !force_iodone_flag,
1237                                                  *mirror_nump,
1238                                                  &block_was_created);
1239         if (NULL == next_block) {
1240                 btrfsic_release_block_ctx(next_block_ctx);
1241                 *next_blockp = NULL;
1242                 return -1;
1243         }
1244         if (block_was_created) {
1245                 l = NULL;
1246                 next_block->generation = BTRFSIC_GENERATION_UNKNOWN;
1247         } else {
1248                 if (next_block->logical_bytenr != next_bytenr &&
1249                     !(!next_block->is_metadata &&
1250                       0 == next_block->logical_bytenr)) {
1251                         printk(KERN_INFO
1252                                "Referenced block @%llu (%s/%llu/%d)"
1253                                " found in hash table, %c,"
1254                                " bytenr mismatch (!= stored %llu).\n",
1255                                (unsigned long long)next_bytenr,
1256                                next_block_ctx->dev->name,
1257                                (unsigned long long)next_block_ctx->dev_bytenr,
1258                                *mirror_nump,
1259                                btrfsic_get_block_type(state, next_block),
1260                                (unsigned long long)next_block->logical_bytenr);
1261                 } else if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1262                         printk(KERN_INFO
1263                                "Referenced block @%llu (%s/%llu/%d)"
1264                                " found in hash table, %c.\n",
1265                                (unsigned long long)next_bytenr,
1266                                next_block_ctx->dev->name,
1267                                (unsigned long long)next_block_ctx->dev_bytenr,
1268                                *mirror_nump,
1269                                btrfsic_get_block_type(state, next_block));
1270                 next_block->logical_bytenr = next_bytenr;
1271
1272                 next_block->mirror_num = *mirror_nump;
1273                 l = btrfsic_block_link_hashtable_lookup(
1274                                 next_block_ctx->dev->bdev,
1275                                 next_block_ctx->dev_bytenr,
1276                                 block_ctx->dev->bdev,
1277                                 block_ctx->dev_bytenr,
1278                                 &state->block_link_hashtable);
1279         }
1280
1281         next_block->disk_key = *disk_key;
1282         if (NULL == l) {
1283                 l = btrfsic_block_link_alloc();
1284                 if (NULL == l) {
1285                         printk(KERN_INFO "btrfsic: error, kmalloc failed!\n");
1286                         btrfsic_release_block_ctx(next_block_ctx);
1287                         *next_blockp = NULL;
1288                         return -1;
1289                 }
1290
1291                 did_alloc_block_link = 1;
1292                 l->block_ref_to = next_block;
1293                 l->block_ref_from = block;
1294                 l->ref_cnt = 1;
1295                 l->parent_generation = parent_generation;
1296
1297                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1298                         btrfsic_print_add_link(state, l);
1299
1300                 list_add(&l->node_ref_to, &block->ref_to_list);
1301                 list_add(&l->node_ref_from, &next_block->ref_from_list);
1302
1303                 btrfsic_block_link_hashtable_add(l,
1304                                                  &state->block_link_hashtable);
1305         } else {
1306                 did_alloc_block_link = 0;
1307                 if (0 == limit_nesting) {
1308                         l->ref_cnt++;
1309                         l->parent_generation = parent_generation;
1310                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1311                                 btrfsic_print_add_link(state, l);
1312                 }
1313         }
1314
1315         if (limit_nesting > 0 && did_alloc_block_link) {
1316                 ret = btrfsic_read_block(state, next_block_ctx);
1317                 if (ret < (int)BTRFSIC_BLOCK_SIZE) {
1318                         printk(KERN_INFO
1319                                "btrfsic: read block @logical %llu failed!\n",
1320                                (unsigned long long)next_bytenr);
1321                         btrfsic_release_block_ctx(next_block_ctx);
1322                         *next_blockp = NULL;
1323                         return -1;
1324                 }
1325
1326                 *next_blockp = next_block;
1327         } else {
1328                 *next_blockp = NULL;
1329         }
1330         (*mirror_nump)++;
1331
1332         return 0;
1333 }
1334
1335 static int btrfsic_handle_extent_data(
1336                 struct btrfsic_state *state,
1337                 struct btrfsic_block *block,
1338                 struct btrfsic_block_data_ctx *block_ctx,
1339                 u32 item_offset, int force_iodone_flag)
1340 {
1341         int ret;
1342         struct btrfs_file_extent_item *file_extent_item =
1343             (struct btrfs_file_extent_item *)(block_ctx->data +
1344                                               offsetof(struct btrfs_leaf,
1345                                                        items) + item_offset);
1346         u64 next_bytenr =
1347             le64_to_cpu(file_extent_item->disk_bytenr) +
1348             le64_to_cpu(file_extent_item->offset);
1349         u64 num_bytes = le64_to_cpu(file_extent_item->num_bytes);
1350         u64 generation = le64_to_cpu(file_extent_item->generation);
1351         struct btrfsic_block_link *l;
1352
1353         if (state->print_mask & BTRFSIC_PRINT_MASK_VERY_VERBOSE)
1354                 printk(KERN_INFO "extent_data: type %u, disk_bytenr = %llu,"
1355                        " offset = %llu, num_bytes = %llu\n",
1356                        file_extent_item->type,
1357                        (unsigned long long)
1358                        le64_to_cpu(file_extent_item->disk_bytenr),
1359                        (unsigned long long)
1360                        le64_to_cpu(file_extent_item->offset),
1361                        (unsigned long long)
1362                        le64_to_cpu(file_extent_item->num_bytes));
1363         if (BTRFS_FILE_EXTENT_REG != file_extent_item->type ||
1364             ((u64)0) == le64_to_cpu(file_extent_item->disk_bytenr))
1365                 return 0;
1366         while (num_bytes > 0) {
1367                 u32 chunk_len;
1368                 int num_copies;
1369                 int mirror_num;
1370
1371                 if (num_bytes > BTRFSIC_BLOCK_SIZE)
1372                         chunk_len = BTRFSIC_BLOCK_SIZE;
1373                 else
1374                         chunk_len = num_bytes;
1375
1376                 num_copies =
1377                     btrfs_num_copies(&state->root->fs_info->mapping_tree,
1378                                      next_bytenr, PAGE_SIZE);
1379                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
1380                         printk(KERN_INFO "num_copies(log_bytenr=%llu) = %d\n",
1381                                (unsigned long long)next_bytenr, num_copies);
1382                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
1383                         struct btrfsic_block_data_ctx next_block_ctx;
1384                         struct btrfsic_block *next_block;
1385                         int block_was_created;
1386
1387                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1388                                 printk(KERN_INFO "btrfsic_handle_extent_data("
1389                                        "mirror_num=%d)\n", mirror_num);
1390                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERY_VERBOSE)
1391                                 printk(KERN_INFO
1392                                        "\tdisk_bytenr = %llu, num_bytes %u\n",
1393                                        (unsigned long long)next_bytenr,
1394                                        chunk_len);
1395                         ret = btrfsic_map_block(state, next_bytenr,
1396                                                 chunk_len, &next_block_ctx,
1397                                                 mirror_num);
1398                         if (ret) {
1399                                 printk(KERN_INFO
1400                                        "btrfsic: btrfsic_map_block(@%llu,"
1401                                        " mirror=%d) failed!\n",
1402                                        (unsigned long long)next_bytenr,
1403                                        mirror_num);
1404                                 return -1;
1405                         }
1406
1407                         next_block = btrfsic_block_lookup_or_add(
1408                                         state,
1409                                         &next_block_ctx,
1410                                         "referenced ",
1411                                         0,
1412                                         force_iodone_flag,
1413                                         !force_iodone_flag,
1414                                         mirror_num,
1415                                         &block_was_created);
1416                         if (NULL == next_block) {
1417                                 printk(KERN_INFO
1418                                        "btrfsic: error, kmalloc failed!\n");
1419                                 btrfsic_release_block_ctx(&next_block_ctx);
1420                                 return -1;
1421                         }
1422                         if (!block_was_created) {
1423                                 if (next_block->logical_bytenr != next_bytenr &&
1424                                     !(!next_block->is_metadata &&
1425                                       0 == next_block->logical_bytenr)) {
1426                                         printk(KERN_INFO
1427                                                "Referenced block"
1428                                                " @%llu (%s/%llu/%d)"
1429                                                " found in hash table, D,"
1430                                                " bytenr mismatch"
1431                                                " (!= stored %llu).\n",
1432                                                (unsigned long long)next_bytenr,
1433                                                next_block_ctx.dev->name,
1434                                                (unsigned long long)
1435                                                next_block_ctx.dev_bytenr,
1436                                                mirror_num,
1437                                                (unsigned long long)
1438                                                next_block->logical_bytenr);
1439                                 }
1440                                 next_block->logical_bytenr = next_bytenr;
1441                                 next_block->mirror_num = mirror_num;
1442                         }
1443
1444                         l = btrfsic_block_link_lookup_or_add(state,
1445                                                              &next_block_ctx,
1446                                                              next_block, block,
1447                                                              generation);
1448                         btrfsic_release_block_ctx(&next_block_ctx);
1449                         if (NULL == l)
1450                                 return -1;
1451                 }
1452
1453                 next_bytenr += chunk_len;
1454                 num_bytes -= chunk_len;
1455         }
1456
1457         return 0;
1458 }
1459
1460 static int btrfsic_map_block(struct btrfsic_state *state, u64 bytenr, u32 len,
1461                              struct btrfsic_block_data_ctx *block_ctx_out,
1462                              int mirror_num)
1463 {
1464         int ret;
1465         u64 length;
1466         struct btrfs_bio *multi = NULL;
1467         struct btrfs_device *device;
1468
1469         length = len;
1470         ret = btrfs_map_block(&state->root->fs_info->mapping_tree, READ,
1471                               bytenr, &length, &multi, mirror_num);
1472
1473         device = multi->stripes[0].dev;
1474         block_ctx_out->dev = btrfsic_dev_state_lookup(device->bdev);
1475         block_ctx_out->dev_bytenr = multi->stripes[0].physical;
1476         block_ctx_out->start = bytenr;
1477         block_ctx_out->len = len;
1478         block_ctx_out->data = NULL;
1479         block_ctx_out->bh = NULL;
1480
1481         if (0 == ret)
1482                 kfree(multi);
1483         if (NULL == block_ctx_out->dev) {
1484                 ret = -ENXIO;
1485                 printk(KERN_INFO "btrfsic: error, cannot lookup dev (#1)!\n");
1486         }
1487
1488         return ret;
1489 }
1490
1491 static int btrfsic_map_superblock(struct btrfsic_state *state, u64 bytenr,
1492                                   u32 len, struct block_device *bdev,
1493                                   struct btrfsic_block_data_ctx *block_ctx_out)
1494 {
1495         block_ctx_out->dev = btrfsic_dev_state_lookup(bdev);
1496         block_ctx_out->dev_bytenr = bytenr;
1497         block_ctx_out->start = bytenr;
1498         block_ctx_out->len = len;
1499         block_ctx_out->data = NULL;
1500         block_ctx_out->bh = NULL;
1501         if (NULL != block_ctx_out->dev) {
1502                 return 0;
1503         } else {
1504                 printk(KERN_INFO "btrfsic: error, cannot lookup dev (#2)!\n");
1505                 return -ENXIO;
1506         }
1507 }
1508
1509 static void btrfsic_release_block_ctx(struct btrfsic_block_data_ctx *block_ctx)
1510 {
1511         if (NULL != block_ctx->bh) {
1512                 brelse(block_ctx->bh);
1513                 block_ctx->bh = NULL;
1514         }
1515 }
1516
1517 static int btrfsic_read_block(struct btrfsic_state *state,
1518                               struct btrfsic_block_data_ctx *block_ctx)
1519 {
1520         block_ctx->bh = NULL;
1521         if (block_ctx->dev_bytenr & 4095) {
1522                 printk(KERN_INFO
1523                        "btrfsic: read_block() with unaligned bytenr %llu\n",
1524                        (unsigned long long)block_ctx->dev_bytenr);
1525                 return -1;
1526         }
1527         if (block_ctx->len > 4096) {
1528                 printk(KERN_INFO
1529                        "btrfsic: read_block() with too huge size %d\n",
1530                        block_ctx->len);
1531                 return -1;
1532         }
1533
1534         block_ctx->bh = __bread(block_ctx->dev->bdev,
1535                                 block_ctx->dev_bytenr >> 12, 4096);
1536         if (NULL == block_ctx->bh)
1537                 return -1;
1538         block_ctx->data = block_ctx->bh->b_data;
1539
1540         return block_ctx->len;
1541 }
1542
1543 static void btrfsic_dump_database(struct btrfsic_state *state)
1544 {
1545         struct list_head *elem_all;
1546
1547         BUG_ON(NULL == state);
1548
1549         printk(KERN_INFO "all_blocks_list:\n");
1550         list_for_each(elem_all, &state->all_blocks_list) {
1551                 const struct btrfsic_block *const b_all =
1552                     list_entry(elem_all, struct btrfsic_block,
1553                                all_blocks_node);
1554                 struct list_head *elem_ref_to;
1555                 struct list_head *elem_ref_from;
1556
1557                 printk(KERN_INFO "%c-block @%llu (%s/%llu/%d)\n",
1558                        btrfsic_get_block_type(state, b_all),
1559                        (unsigned long long)b_all->logical_bytenr,
1560                        b_all->dev_state->name,
1561                        (unsigned long long)b_all->dev_bytenr,
1562                        b_all->mirror_num);
1563
1564                 list_for_each(elem_ref_to, &b_all->ref_to_list) {
1565                         const struct btrfsic_block_link *const l =
1566                             list_entry(elem_ref_to,
1567                                        struct btrfsic_block_link,
1568                                        node_ref_to);
1569
1570                         printk(KERN_INFO " %c @%llu (%s/%llu/%d)"
1571                                " refers %u* to"
1572                                " %c @%llu (%s/%llu/%d)\n",
1573                                btrfsic_get_block_type(state, b_all),
1574                                (unsigned long long)b_all->logical_bytenr,
1575                                b_all->dev_state->name,
1576                                (unsigned long long)b_all->dev_bytenr,
1577                                b_all->mirror_num,
1578                                l->ref_cnt,
1579                                btrfsic_get_block_type(state, l->block_ref_to),
1580                                (unsigned long long)
1581                                l->block_ref_to->logical_bytenr,
1582                                l->block_ref_to->dev_state->name,
1583                                (unsigned long long)l->block_ref_to->dev_bytenr,
1584                                l->block_ref_to->mirror_num);
1585                 }
1586
1587                 list_for_each(elem_ref_from, &b_all->ref_from_list) {
1588                         const struct btrfsic_block_link *const l =
1589                             list_entry(elem_ref_from,
1590                                        struct btrfsic_block_link,
1591                                        node_ref_from);
1592
1593                         printk(KERN_INFO " %c @%llu (%s/%llu/%d)"
1594                                " is ref %u* from"
1595                                " %c @%llu (%s/%llu/%d)\n",
1596                                btrfsic_get_block_type(state, b_all),
1597                                (unsigned long long)b_all->logical_bytenr,
1598                                b_all->dev_state->name,
1599                                (unsigned long long)b_all->dev_bytenr,
1600                                b_all->mirror_num,
1601                                l->ref_cnt,
1602                                btrfsic_get_block_type(state, l->block_ref_from),
1603                                (unsigned long long)
1604                                l->block_ref_from->logical_bytenr,
1605                                l->block_ref_from->dev_state->name,
1606                                (unsigned long long)
1607                                l->block_ref_from->dev_bytenr,
1608                                l->block_ref_from->mirror_num);
1609                 }
1610
1611                 printk(KERN_INFO "\n");
1612         }
1613 }
1614
1615 /*
1616  * Test whether the disk block contains a tree block (leaf or node)
1617  * (note that this test fails for the super block)
1618  */
1619 static int btrfsic_test_for_metadata(struct btrfsic_state *state,
1620                                      const u8 *data, unsigned int size)
1621 {
1622         struct btrfs_header *h;
1623         u8 csum[BTRFS_CSUM_SIZE];
1624         u32 crc = ~(u32)0;
1625         int fail = 0;
1626         int crc_fail = 0;
1627
1628         h = (struct btrfs_header *)data;
1629
1630         if (memcmp(h->fsid, state->root->fs_info->fsid, BTRFS_UUID_SIZE))
1631                 fail++;
1632
1633         crc = crc32c(crc, data + BTRFS_CSUM_SIZE, PAGE_SIZE - BTRFS_CSUM_SIZE);
1634         btrfs_csum_final(crc, csum);
1635         if (memcmp(csum, h->csum, state->csum_size))
1636                 crc_fail++;
1637
1638         return fail || crc_fail;
1639 }
1640
1641 static void btrfsic_process_written_block(struct btrfsic_dev_state *dev_state,
1642                                           u64 dev_bytenr,
1643                                           u8 *mapped_data, unsigned int len,
1644                                           struct bio *bio,
1645                                           int *bio_is_patched,
1646                                           struct buffer_head *bh,
1647                                           int submit_bio_bh_rw)
1648 {
1649         int is_metadata;
1650         struct btrfsic_block *block;
1651         struct btrfsic_block_data_ctx block_ctx;
1652         int ret;
1653         struct btrfsic_state *state = dev_state->state;
1654         struct block_device *bdev = dev_state->bdev;
1655
1656         WARN_ON(len > PAGE_SIZE);
1657         is_metadata = (0 == btrfsic_test_for_metadata(state, mapped_data, len));
1658         if (NULL != bio_is_patched)
1659                 *bio_is_patched = 0;
1660
1661         block = btrfsic_block_hashtable_lookup(bdev, dev_bytenr,
1662                                                &state->block_hashtable);
1663         if (NULL != block) {
1664                 u64 bytenr = 0;
1665                 struct list_head *elem_ref_to;
1666                 struct list_head *tmp_ref_to;
1667
1668                 if (block->is_superblock) {
1669                         bytenr = le64_to_cpu(((struct btrfs_super_block *)
1670                                               mapped_data)->bytenr);
1671                         is_metadata = 1;
1672                         if (state->print_mask &
1673                             BTRFSIC_PRINT_MASK_TREE_BEFORE_SB_WRITE) {
1674                                 printk(KERN_INFO
1675                                        "[before new superblock is written]:\n");
1676                                 btrfsic_dump_tree_sub(state, block, 0);
1677                         }
1678                 }
1679                 if (is_metadata) {
1680                         if (!block->is_superblock) {
1681                                 bytenr = le64_to_cpu(((struct btrfs_header *)
1682                                                       mapped_data)->bytenr);
1683                                 btrfsic_cmp_log_and_dev_bytenr(state, bytenr,
1684                                                                dev_state,
1685                                                                dev_bytenr,
1686                                                                mapped_data);
1687                         }
1688                         if (block->logical_bytenr != bytenr) {
1689                                 printk(KERN_INFO
1690                                        "Written block @%llu (%s/%llu/%d)"
1691                                        " found in hash table, %c,"
1692                                        " bytenr mismatch"
1693                                        " (!= stored %llu).\n",
1694                                        (unsigned long long)bytenr,
1695                                        dev_state->name,
1696                                        (unsigned long long)dev_bytenr,
1697                                        block->mirror_num,
1698                                        btrfsic_get_block_type(state, block),
1699                                        (unsigned long long)
1700                                        block->logical_bytenr);
1701                                 block->logical_bytenr = bytenr;
1702                         } else if (state->print_mask &
1703                                    BTRFSIC_PRINT_MASK_VERBOSE)
1704                                 printk(KERN_INFO
1705                                        "Written block @%llu (%s/%llu/%d)"
1706                                        " found in hash table, %c.\n",
1707                                        (unsigned long long)bytenr,
1708                                        dev_state->name,
1709                                        (unsigned long long)dev_bytenr,
1710                                        block->mirror_num,
1711                                        btrfsic_get_block_type(state, block));
1712                 } else {
1713                         bytenr = block->logical_bytenr;
1714                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1715                                 printk(KERN_INFO
1716                                        "Written block @%llu (%s/%llu/%d)"
1717                                        " found in hash table, %c.\n",
1718                                        (unsigned long long)bytenr,
1719                                        dev_state->name,
1720                                        (unsigned long long)dev_bytenr,
1721                                        block->mirror_num,
1722                                        btrfsic_get_block_type(state, block));
1723                 }
1724
1725                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1726                         printk(KERN_INFO
1727                                "ref_to_list: %cE, ref_from_list: %cE\n",
1728                                list_empty(&block->ref_to_list) ? ' ' : '!',
1729                                list_empty(&block->ref_from_list) ? ' ' : '!');
1730                 if (btrfsic_is_block_ref_by_superblock(state, block, 0)) {
1731                         printk(KERN_INFO "btrfs: attempt to overwrite %c-block"
1732                                " @%llu (%s/%llu/%d), old(gen=%llu,"
1733                                " objectid=%llu, type=%d, offset=%llu),"
1734                                " new(gen=%llu),"
1735                                " which is referenced by most recent superblock"
1736                                " (superblockgen=%llu)!\n",
1737                                btrfsic_get_block_type(state, block),
1738                                (unsigned long long)bytenr,
1739                                dev_state->name,
1740                                (unsigned long long)dev_bytenr,
1741                                block->mirror_num,
1742                                (unsigned long long)block->generation,
1743                                (unsigned long long)
1744                                le64_to_cpu(block->disk_key.objectid),
1745                                block->disk_key.type,
1746                                (unsigned long long)
1747                                le64_to_cpu(block->disk_key.offset),
1748                                (unsigned long long)
1749                                le64_to_cpu(((struct btrfs_header *)
1750                                             mapped_data)->generation),
1751                                (unsigned long long)
1752                                state->max_superblock_generation);
1753                         btrfsic_dump_tree(state);
1754                 }
1755
1756                 if (!block->is_iodone && !block->never_written) {
1757                         printk(KERN_INFO "btrfs: attempt to overwrite %c-block"
1758                                " @%llu (%s/%llu/%d), oldgen=%llu, newgen=%llu,"
1759                                " which is not yet iodone!\n",
1760                                btrfsic_get_block_type(state, block),
1761                                (unsigned long long)bytenr,
1762                                dev_state->name,
1763                                (unsigned long long)dev_bytenr,
1764                                block->mirror_num,
1765                                (unsigned long long)block->generation,
1766                                (unsigned long long)
1767                                le64_to_cpu(((struct btrfs_header *)
1768                                             mapped_data)->generation));
1769                         /* it would not be safe to go on */
1770                         btrfsic_dump_tree(state);
1771                         return;
1772                 }
1773
1774                 /*
1775                  * Clear all references of this block. Do not free
1776                  * the block itself even if is not referenced anymore
1777                  * because it still carries valueable information
1778                  * like whether it was ever written and IO completed.
1779                  */
1780                 list_for_each_safe(elem_ref_to, tmp_ref_to,
1781                                    &block->ref_to_list) {
1782                         struct btrfsic_block_link *const l =
1783                             list_entry(elem_ref_to,
1784                                        struct btrfsic_block_link,
1785                                        node_ref_to);
1786
1787                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1788                                 btrfsic_print_rem_link(state, l);
1789                         l->ref_cnt--;
1790                         if (0 == l->ref_cnt) {
1791                                 list_del(&l->node_ref_to);
1792                                 list_del(&l->node_ref_from);
1793                                 btrfsic_block_link_hashtable_remove(l);
1794                                 btrfsic_block_link_free(l);
1795                         }
1796                 }
1797
1798                 if (block->is_superblock)
1799                         ret = btrfsic_map_superblock(state, bytenr, len,
1800                                                      bdev, &block_ctx);
1801                 else
1802                         ret = btrfsic_map_block(state, bytenr, len,
1803                                                 &block_ctx, 0);
1804                 if (ret) {
1805                         printk(KERN_INFO
1806                                "btrfsic: btrfsic_map_block(root @%llu)"
1807                                " failed!\n", (unsigned long long)bytenr);
1808                         return;
1809                 }
1810                 block_ctx.data = mapped_data;
1811                 /* the following is required in case of writes to mirrors,
1812                  * use the same that was used for the lookup */
1813                 block_ctx.dev = dev_state;
1814                 block_ctx.dev_bytenr = dev_bytenr;
1815
1816                 if (is_metadata || state->include_extent_data) {
1817                         block->never_written = 0;
1818                         block->iodone_w_error = 0;
1819                         if (NULL != bio) {
1820                                 block->is_iodone = 0;
1821                                 BUG_ON(NULL == bio_is_patched);
1822                                 if (!*bio_is_patched) {
1823                                         block->orig_bio_bh_private =
1824                                             bio->bi_private;
1825                                         block->orig_bio_bh_end_io.bio =
1826                                             bio->bi_end_io;
1827                                         block->next_in_same_bio = NULL;
1828                                         bio->bi_private = block;
1829                                         bio->bi_end_io = btrfsic_bio_end_io;
1830                                         *bio_is_patched = 1;
1831                                 } else {
1832                                         struct btrfsic_block *chained_block =
1833                                             (struct btrfsic_block *)
1834                                             bio->bi_private;
1835
1836                                         BUG_ON(NULL == chained_block);
1837                                         block->orig_bio_bh_private =
1838                                             chained_block->orig_bio_bh_private;
1839                                         block->orig_bio_bh_end_io.bio =
1840                                             chained_block->orig_bio_bh_end_io.
1841                                             bio;
1842                                         block->next_in_same_bio = chained_block;
1843                                         bio->bi_private = block;
1844                                 }
1845                         } else if (NULL != bh) {
1846                                 block->is_iodone = 0;
1847                                 block->orig_bio_bh_private = bh->b_private;
1848                                 block->orig_bio_bh_end_io.bh = bh->b_end_io;
1849                                 block->next_in_same_bio = NULL;
1850                                 bh->b_private = block;
1851                                 bh->b_end_io = btrfsic_bh_end_io;
1852                         } else {
1853                                 block->is_iodone = 1;
1854                                 block->orig_bio_bh_private = NULL;
1855                                 block->orig_bio_bh_end_io.bio = NULL;
1856                                 block->next_in_same_bio = NULL;
1857                         }
1858                 }
1859
1860                 block->flush_gen = dev_state->last_flush_gen + 1;
1861                 block->submit_bio_bh_rw = submit_bio_bh_rw;
1862                 if (is_metadata) {
1863                         block->logical_bytenr = bytenr;
1864                         block->is_metadata = 1;
1865                         if (block->is_superblock) {
1866                                 ret = btrfsic_process_written_superblock(
1867                                                 state,
1868                                                 block,
1869                                                 (struct btrfs_super_block *)
1870                                                 mapped_data);
1871                                 if (state->print_mask &
1872                                     BTRFSIC_PRINT_MASK_TREE_AFTER_SB_WRITE) {
1873                                         printk(KERN_INFO
1874                                         "[after new superblock is written]:\n");
1875                                         btrfsic_dump_tree_sub(state, block, 0);
1876                                 }
1877                         } else {
1878                                 block->mirror_num = 0;  /* unknown */
1879                                 ret = btrfsic_process_metablock(
1880                                                 state,
1881                                                 block,
1882                                                 &block_ctx,
1883                                                 (struct btrfs_header *)
1884                                                 block_ctx.data,
1885                                                 0, 0);
1886                         }
1887                         if (ret)
1888                                 printk(KERN_INFO
1889                                        "btrfsic: btrfsic_process_metablock"
1890                                        "(root @%llu) failed!\n",
1891                                        (unsigned long long)dev_bytenr);
1892                 } else {
1893                         block->is_metadata = 0;
1894                         block->mirror_num = 0;  /* unknown */
1895                         block->generation = BTRFSIC_GENERATION_UNKNOWN;
1896                         if (!state->include_extent_data
1897                             && list_empty(&block->ref_from_list)) {
1898                                 /*
1899                                  * disk block is overwritten with extent
1900                                  * data (not meta data) and we are configured
1901                                  * to not include extent data: take the
1902                                  * chance and free the block's memory
1903                                  */
1904                                 btrfsic_block_hashtable_remove(block);
1905                                 list_del(&block->all_blocks_node);
1906                                 btrfsic_block_free(block);
1907                         }
1908                 }
1909                 btrfsic_release_block_ctx(&block_ctx);
1910         } else {
1911                 /* block has not been found in hash table */
1912                 u64 bytenr;
1913
1914                 if (!is_metadata) {
1915                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1916                                 printk(KERN_INFO "Written block (%s/%llu/?)"
1917                                        " !found in hash table, D.\n",
1918                                        dev_state->name,
1919                                        (unsigned long long)dev_bytenr);
1920                         if (!state->include_extent_data)
1921                                 return; /* ignore that written D block */
1922
1923                         /* this is getting ugly for the
1924                          * include_extent_data case... */
1925                         bytenr = 0;     /* unknown */
1926                         block_ctx.start = bytenr;
1927                         block_ctx.len = len;
1928                         block_ctx.bh = NULL;
1929                 } else {
1930                         bytenr = le64_to_cpu(((struct btrfs_header *)
1931                                               mapped_data)->bytenr);
1932                         btrfsic_cmp_log_and_dev_bytenr(state, bytenr, dev_state,
1933                                                        dev_bytenr,
1934                                                        mapped_data);
1935                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1936                                 printk(KERN_INFO
1937                                        "Written block @%llu (%s/%llu/?)"
1938                                        " !found in hash table, M.\n",
1939                                        (unsigned long long)bytenr,
1940                                        dev_state->name,
1941                                        (unsigned long long)dev_bytenr);
1942
1943                         ret = btrfsic_map_block(state, bytenr, len, &block_ctx,
1944                                                 0);
1945                         if (ret) {
1946                                 printk(KERN_INFO
1947                                        "btrfsic: btrfsic_map_block(root @%llu)"
1948                                        " failed!\n",
1949                                        (unsigned long long)dev_bytenr);
1950                                 return;
1951                         }
1952                 }
1953                 block_ctx.data = mapped_data;
1954                 /* the following is required in case of writes to mirrors,
1955                  * use the same that was used for the lookup */
1956                 block_ctx.dev = dev_state;
1957                 block_ctx.dev_bytenr = dev_bytenr;
1958
1959                 block = btrfsic_block_alloc();
1960                 if (NULL == block) {
1961                         printk(KERN_INFO "btrfsic: error, kmalloc failed!\n");
1962                         btrfsic_release_block_ctx(&block_ctx);
1963                         return;
1964                 }
1965                 block->dev_state = dev_state;
1966                 block->dev_bytenr = dev_bytenr;
1967                 block->logical_bytenr = bytenr;
1968                 block->is_metadata = is_metadata;
1969                 block->never_written = 0;
1970                 block->iodone_w_error = 0;
1971                 block->mirror_num = 0;  /* unknown */
1972                 block->flush_gen = dev_state->last_flush_gen + 1;
1973                 block->submit_bio_bh_rw = submit_bio_bh_rw;
1974                 if (NULL != bio) {
1975                         block->is_iodone = 0;
1976                         BUG_ON(NULL == bio_is_patched);
1977                         if (!*bio_is_patched) {
1978                                 block->orig_bio_bh_private = bio->bi_private;
1979                                 block->orig_bio_bh_end_io.bio = bio->bi_end_io;
1980                                 block->next_in_same_bio = NULL;
1981                                 bio->bi_private = block;
1982                                 bio->bi_end_io = btrfsic_bio_end_io;
1983                                 *bio_is_patched = 1;
1984                         } else {
1985                                 struct btrfsic_block *chained_block =
1986                                     (struct btrfsic_block *)
1987                                     bio->bi_private;
1988
1989                                 BUG_ON(NULL == chained_block);
1990                                 block->orig_bio_bh_private =
1991                                     chained_block->orig_bio_bh_private;
1992                                 block->orig_bio_bh_end_io.bio =
1993                                     chained_block->orig_bio_bh_end_io.bio;
1994                                 block->next_in_same_bio = chained_block;
1995                                 bio->bi_private = block;
1996                         }
1997                 } else if (NULL != bh) {
1998                         block->is_iodone = 0;
1999                         block->orig_bio_bh_private = bh->b_private;
2000                         block->orig_bio_bh_end_io.bh = bh->b_end_io;
2001                         block->next_in_same_bio = NULL;
2002                         bh->b_private = block;
2003                         bh->b_end_io = btrfsic_bh_end_io;
2004                 } else {
2005                         block->is_iodone = 1;
2006                         block->orig_bio_bh_private = NULL;
2007                         block->orig_bio_bh_end_io.bio = NULL;
2008                         block->next_in_same_bio = NULL;
2009                 }
2010                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2011                         printk(KERN_INFO
2012                                "New written %c-block @%llu (%s/%llu/%d)\n",
2013                                is_metadata ? 'M' : 'D',
2014                                (unsigned long long)block->logical_bytenr,
2015                                block->dev_state->name,
2016                                (unsigned long long)block->dev_bytenr,
2017                                block->mirror_num);
2018                 list_add(&block->all_blocks_node, &state->all_blocks_list);
2019                 btrfsic_block_hashtable_add(block, &state->block_hashtable);
2020
2021                 if (is_metadata) {
2022                         ret = btrfsic_process_metablock(state, block,
2023                                                         &block_ctx,
2024                                                         (struct btrfs_header *)
2025                                                         block_ctx.data, 0, 0);
2026                         if (ret)
2027                                 printk(KERN_INFO
2028                                        "btrfsic: process_metablock(root @%llu)"
2029                                        " failed!\n",
2030                                        (unsigned long long)dev_bytenr);
2031                 }
2032                 btrfsic_release_block_ctx(&block_ctx);
2033         }
2034 }
2035
2036 static void btrfsic_bio_end_io(struct bio *bp, int bio_error_status)
2037 {
2038         struct btrfsic_block *block = (struct btrfsic_block *)bp->bi_private;
2039         int iodone_w_error;
2040
2041         /* mutex is not held! This is not save if IO is not yet completed
2042          * on umount */
2043         iodone_w_error = 0;
2044         if (bio_error_status)
2045                 iodone_w_error = 1;
2046
2047         BUG_ON(NULL == block);
2048         bp->bi_private = block->orig_bio_bh_private;
2049         bp->bi_end_io = block->orig_bio_bh_end_io.bio;
2050
2051         do {
2052                 struct btrfsic_block *next_block;
2053                 struct btrfsic_dev_state *const dev_state = block->dev_state;
2054
2055                 if ((dev_state->state->print_mask &
2056                      BTRFSIC_PRINT_MASK_END_IO_BIO_BH))
2057                         printk(KERN_INFO
2058                                "bio_end_io(err=%d) for %c @%llu (%s/%llu/%d)\n",
2059                                bio_error_status,
2060                                btrfsic_get_block_type(dev_state->state, block),
2061                                (unsigned long long)block->logical_bytenr,
2062                                dev_state->name,
2063                                (unsigned long long)block->dev_bytenr,
2064                                block->mirror_num);
2065                 next_block = block->next_in_same_bio;
2066                 block->iodone_w_error = iodone_w_error;
2067                 if (block->submit_bio_bh_rw & REQ_FLUSH) {
2068                         dev_state->last_flush_gen++;
2069                         if ((dev_state->state->print_mask &
2070                              BTRFSIC_PRINT_MASK_END_IO_BIO_BH))
2071                                 printk(KERN_INFO
2072                                        "bio_end_io() new %s flush_gen=%llu\n",
2073                                        dev_state->name,
2074                                        (unsigned long long)
2075                                        dev_state->last_flush_gen);
2076                 }
2077                 if (block->submit_bio_bh_rw & REQ_FUA)
2078                         block->flush_gen = 0; /* FUA completed means block is
2079                                                * on disk */
2080                 block->is_iodone = 1; /* for FLUSH, this releases the block */
2081                 block = next_block;
2082         } while (NULL != block);
2083
2084         bp->bi_end_io(bp, bio_error_status);
2085 }
2086
2087 static void btrfsic_bh_end_io(struct buffer_head *bh, int uptodate)
2088 {
2089         struct btrfsic_block *block = (struct btrfsic_block *)bh->b_private;
2090         int iodone_w_error = !uptodate;
2091         struct btrfsic_dev_state *dev_state;
2092
2093         BUG_ON(NULL == block);
2094         dev_state = block->dev_state;
2095         if ((dev_state->state->print_mask & BTRFSIC_PRINT_MASK_END_IO_BIO_BH))
2096                 printk(KERN_INFO
2097                        "bh_end_io(error=%d) for %c @%llu (%s/%llu/%d)\n",
2098                        iodone_w_error,
2099                        btrfsic_get_block_type(dev_state->state, block),
2100                        (unsigned long long)block->logical_bytenr,
2101                        block->dev_state->name,
2102                        (unsigned long long)block->dev_bytenr,
2103                        block->mirror_num);
2104
2105         block->iodone_w_error = iodone_w_error;
2106         if (block->submit_bio_bh_rw & REQ_FLUSH) {
2107                 dev_state->last_flush_gen++;
2108                 if ((dev_state->state->print_mask &
2109                      BTRFSIC_PRINT_MASK_END_IO_BIO_BH))
2110                         printk(KERN_INFO
2111                                "bh_end_io() new %s flush_gen=%llu\n",
2112                                dev_state->name,
2113                                (unsigned long long)dev_state->last_flush_gen);
2114         }
2115         if (block->submit_bio_bh_rw & REQ_FUA)
2116                 block->flush_gen = 0; /* FUA completed means block is on disk */
2117
2118         bh->b_private = block->orig_bio_bh_private;
2119         bh->b_end_io = block->orig_bio_bh_end_io.bh;
2120         block->is_iodone = 1; /* for FLUSH, this releases the block */
2121         bh->b_end_io(bh, uptodate);
2122 }
2123
2124 static int btrfsic_process_written_superblock(
2125                 struct btrfsic_state *state,
2126                 struct btrfsic_block *const superblock,
2127                 struct btrfs_super_block *const super_hdr)
2128 {
2129         int pass;
2130
2131         superblock->generation = btrfs_super_generation(super_hdr);
2132         if (!(superblock->generation > state->max_superblock_generation ||
2133               0 == state->max_superblock_generation)) {
2134                 if (state->print_mask & BTRFSIC_PRINT_MASK_SUPERBLOCK_WRITE)
2135                         printk(KERN_INFO
2136                                "btrfsic: superblock @%llu (%s/%llu/%d)"
2137                                " with old gen %llu <= %llu\n",
2138                                (unsigned long long)superblock->logical_bytenr,
2139                                superblock->dev_state->name,
2140                                (unsigned long long)superblock->dev_bytenr,
2141                                superblock->mirror_num,
2142                                (unsigned long long)
2143                                btrfs_super_generation(super_hdr),
2144                                (unsigned long long)
2145                                state->max_superblock_generation);
2146         } else {
2147                 if (state->print_mask & BTRFSIC_PRINT_MASK_SUPERBLOCK_WRITE)
2148                         printk(KERN_INFO
2149                                "btrfsic: got new superblock @%llu (%s/%llu/%d)"
2150                                " with new gen %llu > %llu\n",
2151                                (unsigned long long)superblock->logical_bytenr,
2152                                superblock->dev_state->name,
2153                                (unsigned long long)superblock->dev_bytenr,
2154                                superblock->mirror_num,
2155                                (unsigned long long)
2156                                btrfs_super_generation(super_hdr),
2157                                (unsigned long long)
2158                                state->max_superblock_generation);
2159
2160                 state->max_superblock_generation =
2161                     btrfs_super_generation(super_hdr);
2162                 state->latest_superblock = superblock;
2163         }
2164
2165         for (pass = 0; pass < 3; pass++) {
2166                 int ret;
2167                 u64 next_bytenr;
2168                 struct btrfsic_block *next_block;
2169                 struct btrfsic_block_data_ctx tmp_next_block_ctx;
2170                 struct btrfsic_block_link *l;
2171                 int num_copies;
2172                 int mirror_num;
2173                 const char *additional_string = NULL;
2174                 struct btrfs_disk_key tmp_disk_key;
2175
2176                 tmp_disk_key.type = BTRFS_ROOT_ITEM_KEY;
2177                 tmp_disk_key.offset = 0;
2178
2179                 switch (pass) {
2180                 case 0:
2181                         tmp_disk_key.objectid =
2182                             cpu_to_le64(BTRFS_ROOT_TREE_OBJECTID);
2183                         additional_string = "root ";
2184                         next_bytenr = btrfs_super_root(super_hdr);
2185                         if (state->print_mask &
2186                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
2187                                 printk(KERN_INFO "root@%llu\n",
2188                                        (unsigned long long)next_bytenr);
2189                         break;
2190                 case 1:
2191                         tmp_disk_key.objectid =
2192                             cpu_to_le64(BTRFS_CHUNK_TREE_OBJECTID);
2193                         additional_string = "chunk ";
2194                         next_bytenr = btrfs_super_chunk_root(super_hdr);
2195                         if (state->print_mask &
2196                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
2197                                 printk(KERN_INFO "chunk@%llu\n",
2198                                        (unsigned long long)next_bytenr);
2199                         break;
2200                 case 2:
2201                         tmp_disk_key.objectid =
2202                             cpu_to_le64(BTRFS_TREE_LOG_OBJECTID);
2203                         additional_string = "log ";
2204                         next_bytenr = btrfs_super_log_root(super_hdr);
2205                         if (0 == next_bytenr)
2206                                 continue;
2207                         if (state->print_mask &
2208                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
2209                                 printk(KERN_INFO "log@%llu\n",
2210                                        (unsigned long long)next_bytenr);
2211                         break;
2212                 }
2213
2214                 num_copies =
2215                     btrfs_num_copies(&state->root->fs_info->mapping_tree,
2216                                      next_bytenr, PAGE_SIZE);
2217                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
2218                         printk(KERN_INFO "num_copies(log_bytenr=%llu) = %d\n",
2219                                (unsigned long long)next_bytenr, num_copies);
2220                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
2221                         int was_created;
2222
2223                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2224                                 printk(KERN_INFO
2225                                        "btrfsic_process_written_superblock("
2226                                        "mirror_num=%d)\n", mirror_num);
2227                         ret = btrfsic_map_block(state, next_bytenr, PAGE_SIZE,
2228                                                 &tmp_next_block_ctx,
2229                                                 mirror_num);
2230                         if (ret) {
2231                                 printk(KERN_INFO
2232                                        "btrfsic: btrfsic_map_block(@%llu,"
2233                                        " mirror=%d) failed!\n",
2234                                        (unsigned long long)next_bytenr,
2235                                        mirror_num);
2236                                 return -1;
2237                         }
2238
2239                         next_block = btrfsic_block_lookup_or_add(
2240                                         state,
2241                                         &tmp_next_block_ctx,
2242                                         additional_string,
2243                                         1, 0, 1,
2244                                         mirror_num,
2245                                         &was_created);
2246                         if (NULL == next_block) {
2247                                 printk(KERN_INFO
2248                                        "btrfsic: error, kmalloc failed!\n");
2249                                 btrfsic_release_block_ctx(&tmp_next_block_ctx);
2250                                 return -1;
2251                         }
2252
2253                         next_block->disk_key = tmp_disk_key;
2254                         if (was_created)
2255                                 next_block->generation =
2256                                     BTRFSIC_GENERATION_UNKNOWN;
2257                         l = btrfsic_block_link_lookup_or_add(
2258                                         state,
2259                                         &tmp_next_block_ctx,
2260                                         next_block,
2261                                         superblock,
2262                                         BTRFSIC_GENERATION_UNKNOWN);
2263                         btrfsic_release_block_ctx(&tmp_next_block_ctx);
2264                         if (NULL == l)
2265                                 return -1;
2266                 }
2267         }
2268
2269         if (-1 == btrfsic_check_all_ref_blocks(state, superblock, 0)) {
2270                 WARN_ON(1);
2271                 btrfsic_dump_tree(state);
2272         }
2273
2274         return 0;
2275 }
2276
2277 static int btrfsic_check_all_ref_blocks(struct btrfsic_state *state,
2278                                         struct btrfsic_block *const block,
2279                                         int recursion_level)
2280 {
2281         struct list_head *elem_ref_to;
2282         int ret = 0;
2283
2284         if (recursion_level >= 3 + BTRFS_MAX_LEVEL) {
2285                 /*
2286                  * Note that this situation can happen and does not
2287                  * indicate an error in regular cases. It happens
2288                  * when disk blocks are freed and later reused.
2289                  * The check-integrity module is not aware of any
2290                  * block free operations, it just recognizes block
2291                  * write operations. Therefore it keeps the linkage
2292                  * information for a block until a block is
2293                  * rewritten. This can temporarily cause incorrect
2294                  * and even circular linkage informations. This
2295                  * causes no harm unless such blocks are referenced
2296                  * by the most recent super block.
2297                  */
2298                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2299                         printk(KERN_INFO
2300                                "btrfsic: abort cyclic linkage (case 1).\n");
2301
2302                 return ret;
2303         }
2304
2305         /*
2306          * This algorithm is recursive because the amount of used stack
2307          * space is very small and the max recursion depth is limited.
2308          */
2309         list_for_each(elem_ref_to, &block->ref_to_list) {
2310                 const struct btrfsic_block_link *const l =
2311                     list_entry(elem_ref_to, struct btrfsic_block_link,
2312                                node_ref_to);
2313
2314                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2315                         printk(KERN_INFO
2316                                "rl=%d, %c @%llu (%s/%llu/%d)"
2317                                " %u* refers to %c @%llu (%s/%llu/%d)\n",
2318                                recursion_level,
2319                                btrfsic_get_block_type(state, block),
2320                                (unsigned long long)block->logical_bytenr,
2321                                block->dev_state->name,
2322                                (unsigned long long)block->dev_bytenr,
2323                                block->mirror_num,
2324                                l->ref_cnt,
2325                                btrfsic_get_block_type(state, l->block_ref_to),
2326                                (unsigned long long)
2327                                l->block_ref_to->logical_bytenr,
2328                                l->block_ref_to->dev_state->name,
2329                                (unsigned long long)l->block_ref_to->dev_bytenr,
2330                                l->block_ref_to->mirror_num);
2331                 if (l->block_ref_to->never_written) {
2332                         printk(KERN_INFO "btrfs: attempt to write superblock"
2333                                " which references block %c @%llu (%s/%llu/%d)"
2334                                " which is never written!\n",
2335                                btrfsic_get_block_type(state, l->block_ref_to),
2336                                (unsigned long long)
2337                                l->block_ref_to->logical_bytenr,
2338                                l->block_ref_to->dev_state->name,
2339                                (unsigned long long)l->block_ref_to->dev_bytenr,
2340                                l->block_ref_to->mirror_num);
2341                         ret = -1;
2342                 } else if (!l->block_ref_to->is_iodone) {
2343                         printk(KERN_INFO "btrfs: attempt to write superblock"
2344                                " which references block %c @%llu (%s/%llu/%d)"
2345                                " which is not yet iodone!\n",
2346                                btrfsic_get_block_type(state, l->block_ref_to),
2347                                (unsigned long long)
2348                                l->block_ref_to->logical_bytenr,
2349                                l->block_ref_to->dev_state->name,
2350                                (unsigned long long)l->block_ref_to->dev_bytenr,
2351                                l->block_ref_to->mirror_num);
2352                         ret = -1;
2353                 } else if (l->parent_generation !=
2354                            l->block_ref_to->generation &&
2355                            BTRFSIC_GENERATION_UNKNOWN !=
2356                            l->parent_generation &&
2357                            BTRFSIC_GENERATION_UNKNOWN !=
2358                            l->block_ref_to->generation) {
2359                         printk(KERN_INFO "btrfs: attempt to write superblock"
2360                                " which references block %c @%llu (%s/%llu/%d)"
2361                                " with generation %llu !="
2362                                " parent generation %llu!\n",
2363                                btrfsic_get_block_type(state, l->block_ref_to),
2364                                (unsigned long long)
2365                                l->block_ref_to->logical_bytenr,
2366                                l->block_ref_to->dev_state->name,
2367                                (unsigned long long)l->block_ref_to->dev_bytenr,
2368                                l->block_ref_to->mirror_num,
2369                                (unsigned long long)l->block_ref_to->generation,
2370                                (unsigned long long)l->parent_generation);
2371                         ret = -1;
2372                 } else if (l->block_ref_to->flush_gen >
2373                            l->block_ref_to->dev_state->last_flush_gen) {
2374                         printk(KERN_INFO "btrfs: attempt to write superblock"
2375                                " which references block %c @%llu (%s/%llu/%d)"
2376                                " which is not flushed out of disk's write cache"
2377                                " (block flush_gen=%llu,"
2378                                " dev->flush_gen=%llu)!\n",
2379                                btrfsic_get_block_type(state, l->block_ref_to),
2380                                (unsigned long long)
2381                                l->block_ref_to->logical_bytenr,
2382                                l->block_ref_to->dev_state->name,
2383                                (unsigned long long)l->block_ref_to->dev_bytenr,
2384                                l->block_ref_to->mirror_num,
2385                                (unsigned long long)block->flush_gen,
2386                                (unsigned long long)
2387                                l->block_ref_to->dev_state->last_flush_gen);
2388                         ret = -1;
2389                 } else if (-1 == btrfsic_check_all_ref_blocks(state,
2390                                                               l->block_ref_to,
2391                                                               recursion_level +
2392                                                               1)) {
2393                         ret = -1;
2394                 }
2395         }
2396
2397         return ret;
2398 }
2399
2400 static int btrfsic_is_block_ref_by_superblock(
2401                 const struct btrfsic_state *state,
2402                 const struct btrfsic_block *block,
2403                 int recursion_level)
2404 {
2405         struct list_head *elem_ref_from;
2406
2407         if (recursion_level >= 3 + BTRFS_MAX_LEVEL) {
2408                 /* refer to comment at "abort cyclic linkage (case 1)" */
2409                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2410                         printk(KERN_INFO
2411                                "btrfsic: abort cyclic linkage (case 2).\n");
2412
2413                 return 0;
2414         }
2415
2416         /*
2417          * This algorithm is recursive because the amount of used stack space
2418          * is very small and the max recursion depth is limited.
2419          */
2420         list_for_each(elem_ref_from, &block->ref_from_list) {
2421                 const struct btrfsic_block_link *const l =
2422                     list_entry(elem_ref_from, struct btrfsic_block_link,
2423                                node_ref_from);
2424
2425                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2426                         printk(KERN_INFO
2427                                "rl=%d, %c @%llu (%s/%llu/%d)"
2428                                " is ref %u* from %c @%llu (%s/%llu/%d)\n",
2429                                recursion_level,
2430                                btrfsic_get_block_type(state, block),
2431                                (unsigned long long)block->logical_bytenr,
2432                                block->dev_state->name,
2433                                (unsigned long long)block->dev_bytenr,
2434                                block->mirror_num,
2435                                l->ref_cnt,
2436                                btrfsic_get_block_type(state, l->block_ref_from),
2437                                (unsigned long long)
2438                                l->block_ref_from->logical_bytenr,
2439                                l->block_ref_from->dev_state->name,
2440                                (unsigned long long)
2441                                l->block_ref_from->dev_bytenr,
2442                                l->block_ref_from->mirror_num);
2443                 if (l->block_ref_from->is_superblock &&
2444                     state->latest_superblock->dev_bytenr ==
2445                     l->block_ref_from->dev_bytenr &&
2446                     state->latest_superblock->dev_state->bdev ==
2447                     l->block_ref_from->dev_state->bdev)
2448                         return 1;
2449                 else if (btrfsic_is_block_ref_by_superblock(state,
2450                                                             l->block_ref_from,
2451                                                             recursion_level +
2452                                                             1))
2453                         return 1;
2454         }
2455
2456         return 0;
2457 }
2458
2459 static void btrfsic_print_add_link(const struct btrfsic_state *state,
2460                                    const struct btrfsic_block_link *l)
2461 {
2462         printk(KERN_INFO
2463                "Add %u* link from %c @%llu (%s/%llu/%d)"
2464                " to %c @%llu (%s/%llu/%d).\n",
2465                l->ref_cnt,
2466                btrfsic_get_block_type(state, l->block_ref_from),
2467                (unsigned long long)l->block_ref_from->logical_bytenr,
2468                l->block_ref_from->dev_state->name,
2469                (unsigned long long)l->block_ref_from->dev_bytenr,
2470                l->block_ref_from->mirror_num,
2471                btrfsic_get_block_type(state, l->block_ref_to),
2472                (unsigned long long)l->block_ref_to->logical_bytenr,
2473                l->block_ref_to->dev_state->name,
2474                (unsigned long long)l->block_ref_to->dev_bytenr,
2475                l->block_ref_to->mirror_num);
2476 }
2477
2478 static void btrfsic_print_rem_link(const struct btrfsic_state *state,
2479                                    const struct btrfsic_block_link *l)
2480 {
2481         printk(KERN_INFO
2482                "Rem %u* link from %c @%llu (%s/%llu/%d)"
2483                " to %c @%llu (%s/%llu/%d).\n",
2484                l->ref_cnt,
2485                btrfsic_get_block_type(state, l->block_ref_from),
2486                (unsigned long long)l->block_ref_from->logical_bytenr,
2487                l->block_ref_from->dev_state->name,
2488                (unsigned long long)l->block_ref_from->dev_bytenr,
2489                l->block_ref_from->mirror_num,
2490                btrfsic_get_block_type(state, l->block_ref_to),
2491                (unsigned long long)l->block_ref_to->logical_bytenr,
2492                l->block_ref_to->dev_state->name,
2493                (unsigned long long)l->block_ref_to->dev_bytenr,
2494                l->block_ref_to->mirror_num);
2495 }
2496
2497 static char btrfsic_get_block_type(const struct btrfsic_state *state,
2498                                    const struct btrfsic_block *block)
2499 {
2500         if (block->is_superblock &&
2501             state->latest_superblock->dev_bytenr == block->dev_bytenr &&
2502             state->latest_superblock->dev_state->bdev == block->dev_state->bdev)
2503                 return 'S';
2504         else if (block->is_superblock)
2505                 return 's';
2506         else if (block->is_metadata)
2507                 return 'M';
2508         else
2509                 return 'D';
2510 }
2511
2512 static void btrfsic_dump_tree(const struct btrfsic_state *state)
2513 {
2514         btrfsic_dump_tree_sub(state, state->latest_superblock, 0);
2515 }
2516
2517 static void btrfsic_dump_tree_sub(const struct btrfsic_state *state,
2518                                   const struct btrfsic_block *block,
2519                                   int indent_level)
2520 {
2521         struct list_head *elem_ref_to;
2522         int indent_add;
2523         static char buf[80];
2524         int cursor_position;
2525
2526         /*
2527          * Should better fill an on-stack buffer with a complete line and
2528          * dump it at once when it is time to print a newline character.
2529          */
2530
2531         /*
2532          * This algorithm is recursive because the amount of used stack space
2533          * is very small and the max recursion depth is limited.
2534          */
2535         indent_add = sprintf(buf, "%c-%llu(%s/%llu/%d)",
2536                              btrfsic_get_block_type(state, block),
2537                              (unsigned long long)block->logical_bytenr,
2538                              block->dev_state->name,
2539                              (unsigned long long)block->dev_bytenr,
2540                              block->mirror_num);
2541         if (indent_level + indent_add > BTRFSIC_TREE_DUMP_MAX_INDENT_LEVEL) {
2542                 printk("[...]\n");
2543                 return;
2544         }
2545         printk(buf);
2546         indent_level += indent_add;
2547         if (list_empty(&block->ref_to_list)) {
2548                 printk("\n");
2549                 return;
2550         }
2551         if (block->mirror_num > 1 &&
2552             !(state->print_mask & BTRFSIC_PRINT_MASK_TREE_WITH_ALL_MIRRORS)) {
2553                 printk(" [...]\n");
2554                 return;
2555         }
2556
2557         cursor_position = indent_level;
2558         list_for_each(elem_ref_to, &block->ref_to_list) {
2559                 const struct btrfsic_block_link *const l =
2560                     list_entry(elem_ref_to, struct btrfsic_block_link,
2561                                node_ref_to);
2562
2563                 while (cursor_position < indent_level) {
2564                         printk(" ");
2565                         cursor_position++;
2566                 }
2567                 if (l->ref_cnt > 1)
2568                         indent_add = sprintf(buf, " %d*--> ", l->ref_cnt);
2569                 else
2570                         indent_add = sprintf(buf, " --> ");
2571                 if (indent_level + indent_add >
2572                     BTRFSIC_TREE_DUMP_MAX_INDENT_LEVEL) {
2573                         printk("[...]\n");
2574                         cursor_position = 0;
2575                         continue;
2576                 }
2577
2578                 printk(buf);
2579
2580                 btrfsic_dump_tree_sub(state, l->block_ref_to,
2581                                       indent_level + indent_add);
2582                 cursor_position = 0;
2583         }
2584 }
2585
2586 static struct btrfsic_block_link *btrfsic_block_link_lookup_or_add(
2587                 struct btrfsic_state *state,
2588                 struct btrfsic_block_data_ctx *next_block_ctx,
2589                 struct btrfsic_block *next_block,
2590                 struct btrfsic_block *from_block,
2591                 u64 parent_generation)
2592 {
2593         struct btrfsic_block_link *l;
2594
2595         l = btrfsic_block_link_hashtable_lookup(next_block_ctx->dev->bdev,
2596                                                 next_block_ctx->dev_bytenr,
2597                                                 from_block->dev_state->bdev,
2598                                                 from_block->dev_bytenr,
2599                                                 &state->block_link_hashtable);
2600         if (NULL == l) {
2601                 l = btrfsic_block_link_alloc();
2602                 if (NULL == l) {
2603                         printk(KERN_INFO
2604                                "btrfsic: error, kmalloc" " failed!\n");
2605                         return NULL;
2606                 }
2607
2608                 l->block_ref_to = next_block;
2609                 l->block_ref_from = from_block;
2610                 l->ref_cnt = 1;
2611                 l->parent_generation = parent_generation;
2612
2613                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2614                         btrfsic_print_add_link(state, l);
2615
2616                 list_add(&l->node_ref_to, &from_block->ref_to_list);
2617                 list_add(&l->node_ref_from, &next_block->ref_from_list);
2618
2619                 btrfsic_block_link_hashtable_add(l,
2620                                                  &state->block_link_hashtable);
2621         } else {
2622                 l->ref_cnt++;
2623                 l->parent_generation = parent_generation;
2624                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2625                         btrfsic_print_add_link(state, l);
2626         }
2627
2628         return l;
2629 }
2630
2631 static struct btrfsic_block *btrfsic_block_lookup_or_add(
2632                 struct btrfsic_state *state,
2633                 struct btrfsic_block_data_ctx *block_ctx,
2634                 const char *additional_string,
2635                 int is_metadata,
2636                 int is_iodone,
2637                 int never_written,
2638                 int mirror_num,
2639                 int *was_created)
2640 {
2641         struct btrfsic_block *block;
2642
2643         block = btrfsic_block_hashtable_lookup(block_ctx->dev->bdev,
2644                                                block_ctx->dev_bytenr,
2645                                                &state->block_hashtable);
2646         if (NULL == block) {
2647                 struct btrfsic_dev_state *dev_state;
2648
2649                 block = btrfsic_block_alloc();
2650                 if (NULL == block) {
2651                         printk(KERN_INFO "btrfsic: error, kmalloc failed!\n");
2652                         return NULL;
2653                 }
2654                 dev_state = btrfsic_dev_state_lookup(block_ctx->dev->bdev);
2655                 if (NULL == dev_state) {
2656                         printk(KERN_INFO
2657                                "btrfsic: error, lookup dev_state failed!\n");
2658                         btrfsic_block_free(block);
2659                         return NULL;
2660                 }
2661                 block->dev_state = dev_state;
2662                 block->dev_bytenr = block_ctx->dev_bytenr;
2663                 block->logical_bytenr = block_ctx->start;
2664                 block->is_metadata = is_metadata;
2665                 block->is_iodone = is_iodone;
2666                 block->never_written = never_written;
2667                 block->mirror_num = mirror_num;
2668                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2669                         printk(KERN_INFO
2670                                "New %s%c-block @%llu (%s/%llu/%d)\n",
2671                                additional_string,
2672                                btrfsic_get_block_type(state, block),
2673                                (unsigned long long)block->logical_bytenr,
2674                                dev_state->name,
2675                                (unsigned long long)block->dev_bytenr,
2676                                mirror_num);
2677                 list_add(&block->all_blocks_node, &state->all_blocks_list);
2678                 btrfsic_block_hashtable_add(block, &state->block_hashtable);
2679                 if (NULL != was_created)
2680                         *was_created = 1;
2681         } else {
2682                 if (NULL != was_created)
2683                         *was_created = 0;
2684         }
2685
2686         return block;
2687 }
2688
2689 static void btrfsic_cmp_log_and_dev_bytenr(struct btrfsic_state *state,
2690                                            u64 bytenr,
2691                                            struct btrfsic_dev_state *dev_state,
2692                                            u64 dev_bytenr, char *data)
2693 {
2694         int num_copies;
2695         int mirror_num;
2696         int ret;
2697         struct btrfsic_block_data_ctx block_ctx;
2698         int match = 0;
2699
2700         num_copies = btrfs_num_copies(&state->root->fs_info->mapping_tree,
2701                                       bytenr, PAGE_SIZE);
2702
2703         for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
2704                 ret = btrfsic_map_block(state, bytenr, PAGE_SIZE,
2705                                         &block_ctx, mirror_num);
2706                 if (ret) {
2707                         printk(KERN_INFO "btrfsic:"
2708                                " btrfsic_map_block(logical @%llu,"
2709                                " mirror %d) failed!\n",
2710                                (unsigned long long)bytenr, mirror_num);
2711                         continue;
2712                 }
2713
2714                 if (dev_state->bdev == block_ctx.dev->bdev &&
2715                     dev_bytenr == block_ctx.dev_bytenr) {
2716                         match++;
2717                         btrfsic_release_block_ctx(&block_ctx);
2718                         break;
2719                 }
2720                 btrfsic_release_block_ctx(&block_ctx);
2721         }
2722
2723         if (!match) {
2724                 printk(KERN_INFO "btrfs: attempt to write M-block which contains logical bytenr that doesn't map to dev+physical bytenr of submit_bio,"
2725                        " buffer->log_bytenr=%llu, submit_bio(bdev=%s,"
2726                        " phys_bytenr=%llu)!\n",
2727                        (unsigned long long)bytenr, dev_state->name,
2728                        (unsigned long long)dev_bytenr);
2729                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
2730                         ret = btrfsic_map_block(state, bytenr, PAGE_SIZE,
2731                                                 &block_ctx, mirror_num);
2732                         if (ret)
2733                                 continue;
2734
2735                         printk(KERN_INFO "Read logical bytenr @%llu maps to"
2736                                " (%s/%llu/%d)\n",
2737                                (unsigned long long)bytenr,
2738                                block_ctx.dev->name,
2739                                (unsigned long long)block_ctx.dev_bytenr,
2740                                mirror_num);
2741                 }
2742                 WARN_ON(1);
2743         }
2744 }
2745
2746 static struct btrfsic_dev_state *btrfsic_dev_state_lookup(
2747                 struct block_device *bdev)
2748 {
2749         struct btrfsic_dev_state *ds;
2750
2751         ds = btrfsic_dev_state_hashtable_lookup(bdev,
2752                                                 &btrfsic_dev_state_hashtable);
2753         return ds;
2754 }
2755
2756 int btrfsic_submit_bh(int rw, struct buffer_head *bh)
2757 {
2758         struct btrfsic_dev_state *dev_state;
2759
2760         if (!btrfsic_is_initialized)
2761                 return submit_bh(rw, bh);
2762
2763         mutex_lock(&btrfsic_mutex);
2764         /* since btrfsic_submit_bh() might also be called before
2765          * btrfsic_mount(), this might return NULL */
2766         dev_state = btrfsic_dev_state_lookup(bh->b_bdev);
2767
2768         /* Only called to write the superblock (incl. FLUSH/FUA) */
2769         if (NULL != dev_state &&
2770             (rw & WRITE) && bh->b_size > 0) {
2771                 u64 dev_bytenr;
2772
2773                 dev_bytenr = 4096 * bh->b_blocknr;
2774                 if (dev_state->state->print_mask &
2775                     BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH)
2776                         printk(KERN_INFO
2777                                "submit_bh(rw=0x%x, blocknr=%lu (bytenr %llu),"
2778                                " size=%lu, data=%p, bdev=%p)\n",
2779                                rw, (unsigned long)bh->b_blocknr,
2780                                (unsigned long long)dev_bytenr,
2781                                (unsigned long)bh->b_size, bh->b_data,
2782                                bh->b_bdev);
2783                 btrfsic_process_written_block(dev_state, dev_bytenr,
2784                                               bh->b_data, bh->b_size, NULL,
2785                                               NULL, bh, rw);
2786         } else if (NULL != dev_state && (rw & REQ_FLUSH)) {
2787                 if (dev_state->state->print_mask &
2788                     BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH)
2789                         printk(KERN_INFO
2790                                "submit_bh(rw=0x%x) FLUSH, bdev=%p)\n",
2791                                rw, bh->b_bdev);
2792                 if (!dev_state->dummy_block_for_bio_bh_flush.is_iodone) {
2793                         if ((dev_state->state->print_mask &
2794                              (BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH |
2795                               BTRFSIC_PRINT_MASK_VERBOSE)))
2796                                 printk(KERN_INFO
2797                                        "btrfsic_submit_bh(%s) with FLUSH"
2798                                        " but dummy block already in use"
2799                                        " (ignored)!\n",
2800                                        dev_state->name);
2801                 } else {
2802                         struct btrfsic_block *const block =
2803                                 &dev_state->dummy_block_for_bio_bh_flush;
2804
2805                         block->is_iodone = 0;
2806                         block->never_written = 0;
2807                         block->iodone_w_error = 0;
2808                         block->flush_gen = dev_state->last_flush_gen + 1;
2809                         block->submit_bio_bh_rw = rw;
2810                         block->orig_bio_bh_private = bh->b_private;
2811                         block->orig_bio_bh_end_io.bh = bh->b_end_io;
2812                         block->next_in_same_bio = NULL;
2813                         bh->b_private = block;
2814                         bh->b_end_io = btrfsic_bh_end_io;
2815                 }
2816         }
2817         mutex_unlock(&btrfsic_mutex);
2818         return submit_bh(rw, bh);
2819 }
2820
2821 void btrfsic_submit_bio(int rw, struct bio *bio)
2822 {
2823         struct btrfsic_dev_state *dev_state;
2824
2825         if (!btrfsic_is_initialized) {
2826                 submit_bio(rw, bio);
2827                 return;
2828         }
2829
2830         mutex_lock(&btrfsic_mutex);
2831         /* since btrfsic_submit_bio() is also called before
2832          * btrfsic_mount(), this might return NULL */
2833         dev_state = btrfsic_dev_state_lookup(bio->bi_bdev);
2834         if (NULL != dev_state &&
2835             (rw & WRITE) && NULL != bio->bi_io_vec) {
2836                 unsigned int i;
2837                 u64 dev_bytenr;
2838                 int bio_is_patched;
2839
2840                 dev_bytenr = 512 * bio->bi_sector;
2841                 bio_is_patched = 0;
2842                 if (dev_state->state->print_mask &
2843                     BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH)
2844                         printk(KERN_INFO
2845                                "submit_bio(rw=0x%x, bi_vcnt=%u,"
2846                                " bi_sector=%lu (bytenr %llu), bi_bdev=%p)\n",
2847                                rw, bio->bi_vcnt, (unsigned long)bio->bi_sector,
2848                                (unsigned long long)dev_bytenr,
2849                                bio->bi_bdev);
2850
2851                 for (i = 0; i < bio->bi_vcnt; i++) {
2852                         u8 *mapped_data;
2853
2854                         mapped_data = kmap(bio->bi_io_vec[i].bv_page);
2855                         if ((BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH |
2856                              BTRFSIC_PRINT_MASK_VERBOSE) ==
2857                             (dev_state->state->print_mask &
2858                              (BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH |
2859                               BTRFSIC_PRINT_MASK_VERBOSE)))
2860                                 printk(KERN_INFO
2861                                        "#%u: page=%p, mapped=%p, len=%u,"
2862                                        " offset=%u\n",
2863                                        i, bio->bi_io_vec[i].bv_page,
2864                                        mapped_data,
2865                                        bio->bi_io_vec[i].bv_len,
2866                                        bio->bi_io_vec[i].bv_offset);
2867                         btrfsic_process_written_block(dev_state, dev_bytenr,
2868                                                       mapped_data,
2869                                                       bio->bi_io_vec[i].bv_len,
2870                                                       bio, &bio_is_patched,
2871                                                       NULL, rw);
2872                         kunmap(bio->bi_io_vec[i].bv_page);
2873                         dev_bytenr += bio->bi_io_vec[i].bv_len;
2874                 }
2875         } else if (NULL != dev_state && (rw & REQ_FLUSH)) {
2876                 if (dev_state->state->print_mask &
2877                     BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH)
2878                         printk(KERN_INFO
2879                                "submit_bio(rw=0x%x) FLUSH, bdev=%p)\n",
2880                                rw, bio->bi_bdev);
2881                 if (!dev_state->dummy_block_for_bio_bh_flush.is_iodone) {
2882                         if ((dev_state->state->print_mask &
2883                              (BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH |
2884                               BTRFSIC_PRINT_MASK_VERBOSE)))
2885                                 printk(KERN_INFO
2886                                        "btrfsic_submit_bio(%s) with FLUSH"
2887                                        " but dummy block already in use"
2888                                        " (ignored)!\n",
2889                                        dev_state->name);
2890                 } else {
2891                         struct btrfsic_block *const block =
2892                                 &dev_state->dummy_block_for_bio_bh_flush;
2893
2894                         block->is_iodone = 0;
2895                         block->never_written = 0;
2896                         block->iodone_w_error = 0;
2897                         block->flush_gen = dev_state->last_flush_gen + 1;
2898                         block->submit_bio_bh_rw = rw;
2899                         block->orig_bio_bh_private = bio->bi_private;
2900                         block->orig_bio_bh_end_io.bio = bio->bi_end_io;
2901                         block->next_in_same_bio = NULL;
2902                         bio->bi_private = block;
2903                         bio->bi_end_io = btrfsic_bio_end_io;
2904                 }
2905         }
2906         mutex_unlock(&btrfsic_mutex);
2907
2908         submit_bio(rw, bio);
2909 }
2910
2911 int btrfsic_mount(struct btrfs_root *root,
2912                   struct btrfs_fs_devices *fs_devices,
2913                   int including_extent_data, u32 print_mask)
2914 {
2915         int ret;
2916         struct btrfsic_state *state;
2917         struct list_head *dev_head = &fs_devices->devices;
2918         struct btrfs_device *device;
2919
2920         state = kzalloc(sizeof(*state), GFP_NOFS);
2921         if (NULL == state) {
2922                 printk(KERN_INFO "btrfs check-integrity: kmalloc() failed!\n");
2923                 return -1;
2924         }
2925
2926         if (!btrfsic_is_initialized) {
2927                 mutex_init(&btrfsic_mutex);
2928                 btrfsic_dev_state_hashtable_init(&btrfsic_dev_state_hashtable);
2929                 btrfsic_is_initialized = 1;
2930         }
2931         mutex_lock(&btrfsic_mutex);
2932         state->root = root;
2933         state->print_mask = print_mask;
2934         state->include_extent_data = including_extent_data;
2935         state->csum_size = 0;
2936         INIT_LIST_HEAD(&state->all_blocks_list);
2937         btrfsic_block_hashtable_init(&state->block_hashtable);
2938         btrfsic_block_link_hashtable_init(&state->block_link_hashtable);
2939         state->max_superblock_generation = 0;
2940         state->latest_superblock = NULL;
2941
2942         list_for_each_entry(device, dev_head, dev_list) {
2943                 struct btrfsic_dev_state *ds;
2944                 char *p;
2945
2946                 if (!device->bdev || !device->name)
2947                         continue;
2948
2949                 ds = btrfsic_dev_state_alloc();
2950                 if (NULL == ds) {
2951                         printk(KERN_INFO
2952                                "btrfs check-integrity: kmalloc() failed!\n");
2953                         mutex_unlock(&btrfsic_mutex);
2954                         return -1;
2955                 }
2956                 ds->bdev = device->bdev;
2957                 ds->state = state;
2958                 bdevname(ds->bdev, ds->name);
2959                 ds->name[BDEVNAME_SIZE - 1] = '\0';
2960                 for (p = ds->name; *p != '\0'; p++);
2961                 while (p > ds->name && *p != '/')
2962                         p--;
2963                 if (*p == '/')
2964                         p++;
2965                 strlcpy(ds->name, p, sizeof(ds->name));
2966                 btrfsic_dev_state_hashtable_add(ds,
2967                                                 &btrfsic_dev_state_hashtable);
2968         }
2969
2970         ret = btrfsic_process_superblock(state, fs_devices);
2971         if (0 != ret) {
2972                 mutex_unlock(&btrfsic_mutex);
2973                 btrfsic_unmount(root, fs_devices);
2974                 return ret;
2975         }
2976
2977         if (state->print_mask & BTRFSIC_PRINT_MASK_INITIAL_DATABASE)
2978                 btrfsic_dump_database(state);
2979         if (state->print_mask & BTRFSIC_PRINT_MASK_INITIAL_TREE)
2980                 btrfsic_dump_tree(state);
2981
2982         mutex_unlock(&btrfsic_mutex);
2983         return 0;
2984 }
2985
2986 void btrfsic_unmount(struct btrfs_root *root,
2987                      struct btrfs_fs_devices *fs_devices)
2988 {
2989         struct list_head *elem_all;
2990         struct list_head *tmp_all;
2991         struct btrfsic_state *state;
2992         struct list_head *dev_head = &fs_devices->devices;
2993         struct btrfs_device *device;
2994
2995         if (!btrfsic_is_initialized)
2996                 return;
2997
2998         mutex_lock(&btrfsic_mutex);
2999
3000         state = NULL;
3001         list_for_each_entry(device, dev_head, dev_list) {
3002                 struct btrfsic_dev_state *ds;
3003
3004                 if (!device->bdev || !device->name)
3005                         continue;
3006
3007                 ds = btrfsic_dev_state_hashtable_lookup(
3008                                 device->bdev,
3009                                 &btrfsic_dev_state_hashtable);
3010                 if (NULL != ds) {
3011                         state = ds->state;
3012                         btrfsic_dev_state_hashtable_remove(ds);
3013                         btrfsic_dev_state_free(ds);
3014                 }
3015         }
3016
3017         if (NULL == state) {
3018                 printk(KERN_INFO
3019                        "btrfsic: error, cannot find state information"
3020                        " on umount!\n");
3021                 mutex_unlock(&btrfsic_mutex);
3022                 return;
3023         }
3024
3025         /*
3026          * Don't care about keeping the lists' state up to date,
3027          * just free all memory that was allocated dynamically.
3028          * Free the blocks and the block_links.
3029          */
3030         list_for_each_safe(elem_all, tmp_all, &state->all_blocks_list) {
3031                 struct btrfsic_block *const b_all =
3032                     list_entry(elem_all, struct btrfsic_block,
3033                                all_blocks_node);
3034                 struct list_head *elem_ref_to;
3035                 struct list_head *tmp_ref_to;
3036
3037                 list_for_each_safe(elem_ref_to, tmp_ref_to,
3038                                    &b_all->ref_to_list) {
3039                         struct btrfsic_block_link *const l =
3040                             list_entry(elem_ref_to,
3041                                        struct btrfsic_block_link,
3042                                        node_ref_to);
3043
3044                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
3045                                 btrfsic_print_rem_link(state, l);
3046
3047                         l->ref_cnt--;
3048                         if (0 == l->ref_cnt)
3049                                 btrfsic_block_link_free(l);
3050                 }
3051
3052                 if (b_all->is_iodone)
3053                         btrfsic_block_free(b_all);
3054                 else
3055                         printk(KERN_INFO "btrfs: attempt to free %c-block"
3056                                " @%llu (%s/%llu/%d) on umount which is"
3057                                " not yet iodone!\n",
3058                                btrfsic_get_block_type(state, b_all),
3059                                (unsigned long long)b_all->logical_bytenr,
3060                                b_all->dev_state->name,
3061                                (unsigned long long)b_all->dev_bytenr,
3062                                b_all->mirror_num);
3063         }
3064
3065         mutex_unlock(&btrfsic_mutex);
3066
3067         kfree(state);
3068 }