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 no write error was indicated and a
  41  *        FLUSH request to the device where these blocks are
  42  *        located was received and completed.
  43  *    2b. All referenced blocks need to have a generation
  44  *        number which is equal to the parent's number.
  45  *
  46  * One issue that was found using this module was that the log
  47  * tree on disk became temporarily corrupted because disk blocks
  48  * that had been in use for the log tree had been freed and
  49  * reused too early, while being referenced by the written super
  50  * block.
  51  *
  52  * The search term in the kernel log that can be used to filter
  53  * on the existence of detected integrity issues is
  54  * "btrfs: attempt".
  55  *
  56  * The integrity check is enabled via mount options. These
  57  * mount options are only supported if the integrity check
  58  * tool is compiled by defining BTRFS_FS_CHECK_INTEGRITY.
  59  *
  60  * Example #1, apply integrity checks to all metadata:
  61  * mount /dev/sdb1 /mnt -o check_int
  62  *
  63  * Example #2, apply integrity checks to all metadata and
  64  * to data extents:
  65  * mount /dev/sdb1 /mnt -o check_int_data
  66  *
  67  * Example #3, apply integrity checks to all metadata and dump
  68  * the tree that the super block references to kernel messages
  69  * each time after a super block was written:
  70  * mount /dev/sdb1 /mnt -o check_int,check_int_print_mask=263
  71  *
  72  * If the integrity check tool is included and activated in
  73  * the mount options, plenty of kernel memory is used, and
  74  * plenty of additional CPU cycles are spent. Enabling this
  75  * functionality is not intended for normal use. In most
  76  * cases, unless you are a btrfs developer who needs to verify
  77  * the integrity of (super)-block write requests, do not
  78  * enable the config option BTRFS_FS_CHECK_INTEGRITY to
  79  * include and compile the integrity check tool.
  80  */
  81
  82 #include <linux/sched.h>
  83 #include <linux/slab.h>
  84 #include <linux/buffer_head.h>
  85 #include <linux/mutex.h>
  86 #include <linux/crc32c.h>
  87 #include <linux/genhd.h>
  88 #include <linux/blkdev.h>
  89 #include "ctree.h"
  90 #include "disk-io.h"
  91 #include "transaction.h"
  92 #include "extent_io.h"
  93 #include "volumes.h"
  94 #include "print-tree.h"
  95 #include "locking.h"
  96 #include "check-integrity.h"
  97 #include "rcu-string.h"
  98
  99 #define BTRFSIC_BLOCK_HASHTABLE_SIZE 0x10000
 100 #define BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE 0x10000
 101 #define BTRFSIC_DEV2STATE_HASHTABLE_SIZE 0x100
 102 #define BTRFSIC_BLOCK_MAGIC_NUMBER 0x14491051
 103 #define BTRFSIC_BLOCK_LINK_MAGIC_NUMBER 0x11070807
 104 #define BTRFSIC_DEV2STATE_MAGIC_NUMBER 0x20111530
 105 #define BTRFSIC_BLOCK_STACK_FRAME_MAGIC_NUMBER 20111300
 106 #define BTRFSIC_TREE_DUMP_MAX_INDENT_LEVEL (200 - 6)    /* in characters,
 107                                                          * excluding " [...]" */
 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;        /* 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 **datav;
 214         struct page **pagev;
 215         void *mem_to_free;
 216 };
 217
 218 /* This structure is used to implement recursion without occupying
 219  * any stack space, refer to btrfsic_process_metablock() */
 220 struct btrfsic_stack_frame {
 221         u32 magic;
 222         u32 nr;
 223         int error;
 224         int i;
 225         int limit_nesting;
 226         int num_copies;
 227         int mirror_num;
 228         struct btrfsic_block *block;
 229         struct btrfsic_block_data_ctx *block_ctx;
 230         struct btrfsic_block *next_block;
 231         struct btrfsic_block_data_ctx next_block_ctx;
 232         struct btrfs_header *hdr;
 233         struct btrfsic_stack_frame *prev;
 234 };
 235
 236 /* Some state per mounted filesystem */
 237 struct btrfsic_state {
 238         u32 print_mask;
 239         int include_extent_data;
 240         int csum_size;
 241         struct list_head all_blocks_list;
 242         struct btrfsic_block_hashtable block_hashtable;
 243         struct btrfsic_block_link_hashtable block_link_hashtable;
 244         struct btrfs_root *root;
 245         u64 max_superblock_generation;
 246         struct btrfsic_block *latest_superblock;
 247         u32 metablock_size;
 248         u32 datablock_size;
 249 };
 250
 251 static void btrfsic_block_init(struct btrfsic_block *b);
 252 static struct btrfsic_block *btrfsic_block_alloc(void);
 253 static void btrfsic_block_free(struct btrfsic_block *b);
 254 static void btrfsic_block_link_init(struct btrfsic_block_link *n);
 255 static struct btrfsic_block_link *btrfsic_block_link_alloc(void);
 256 static void btrfsic_block_link_free(struct btrfsic_block_link *n);
 257 static void btrfsic_dev_state_init(struct btrfsic_dev_state *ds);
 258 static struct btrfsic_dev_state *btrfsic_dev_state_alloc(void);
 259 static void btrfsic_dev_state_free(struct btrfsic_dev_state *ds);
 260 static void btrfsic_block_hashtable_init(struct btrfsic_block_hashtable *h);
 261 static void btrfsic_block_hashtable_add(struct btrfsic_block *b,
 262                                         struct btrfsic_block_hashtable *h);
 263 static void btrfsic_block_hashtable_remove(struct btrfsic_block *b);
 264 static struct btrfsic_block *btrfsic_block_hashtable_lookup(
 265                 struct block_device *bdev,
 266                 u64 dev_bytenr,
 267                 struct btrfsic_block_hashtable *h);
 268 static void btrfsic_block_link_hashtable_init(
 269                 struct btrfsic_block_link_hashtable *h);
 270 static void btrfsic_block_link_hashtable_add(
 271                 struct btrfsic_block_link *l,
 272                 struct btrfsic_block_link_hashtable *h);
 273 static void btrfsic_block_link_hashtable_remove(struct btrfsic_block_link *l);
 274 static struct btrfsic_block_link *btrfsic_block_link_hashtable_lookup(
 275                 struct block_device *bdev_ref_to,
 276                 u64 dev_bytenr_ref_to,
 277                 struct block_device *bdev_ref_from,
 278                 u64 dev_bytenr_ref_from,
 279                 struct btrfsic_block_link_hashtable *h);
 280 static void btrfsic_dev_state_hashtable_init(
 281                 struct btrfsic_dev_state_hashtable *h);
 282 static void btrfsic_dev_state_hashtable_add(
 283                 struct btrfsic_dev_state *ds,
 284                 struct btrfsic_dev_state_hashtable *h);
 285 static void btrfsic_dev_state_hashtable_remove(struct btrfsic_dev_state *ds);
 286 static struct btrfsic_dev_state *btrfsic_dev_state_hashtable_lookup(
 287                 struct block_device *bdev,
 288                 struct btrfsic_dev_state_hashtable *h);
 289 static struct btrfsic_stack_frame *btrfsic_stack_frame_alloc(void);
 290 static void btrfsic_stack_frame_free(struct btrfsic_stack_frame *sf);
 291 static int btrfsic_process_superblock(struct btrfsic_state *state,
 292                                       struct btrfs_fs_devices *fs_devices);
 293 static int btrfsic_process_metablock(struct btrfsic_state *state,
 294                                      struct btrfsic_block *block,
 295                                      struct btrfsic_block_data_ctx *block_ctx,
 296                                      int limit_nesting, int force_iodone_flag);
 297 static void btrfsic_read_from_block_data(
 298         struct btrfsic_block_data_ctx *block_ctx,
 299         void *dst, u32 offset, size_t len);
 300 static int btrfsic_create_link_to_next_block(
 301                 struct btrfsic_state *state,
 302                 struct btrfsic_block *block,
 303                 struct btrfsic_block_data_ctx
 304                 *block_ctx, u64 next_bytenr,
 305                 int limit_nesting,
 306                 struct btrfsic_block_data_ctx *next_block_ctx,
 307                 struct btrfsic_block **next_blockp,
 308                 int force_iodone_flag,
 309                 int *num_copiesp, int *mirror_nump,
 310                 struct btrfs_disk_key *disk_key,
 311                 u64 parent_generation);
 312 static int btrfsic_handle_extent_data(struct btrfsic_state *state,
 313                                       struct btrfsic_block *block,
 314                                       struct btrfsic_block_data_ctx *block_ctx,
 315                                       u32 item_offset, int force_iodone_flag);
 316 static int btrfsic_map_block(struct btrfsic_state *state, u64 bytenr, u32 len,
 317                              struct btrfsic_block_data_ctx *block_ctx_out,
 318                              int mirror_num);
 319 static int btrfsic_map_superblock(struct btrfsic_state *state, u64 bytenr,
 320                                   u32 len, struct block_device *bdev,
 321                                   struct btrfsic_block_data_ctx *block_ctx_out);
 322 static void btrfsic_release_block_ctx(struct btrfsic_block_data_ctx *block_ctx);
 323 static int btrfsic_read_block(struct btrfsic_state *state,
 324                               struct btrfsic_block_data_ctx *block_ctx);
 325 static void btrfsic_dump_database(struct btrfsic_state *state);
 326 static void btrfsic_complete_bio_end_io(struct bio *bio, int err);
 327 static int btrfsic_test_for_metadata(struct btrfsic_state *state,
 328                                      char **datav, unsigned int num_pages);
 329 static void btrfsic_process_written_block(struct btrfsic_dev_state *dev_state,
 330                                           u64 dev_bytenr, char **mapped_datav,
 331                                           unsigned int num_pages,
 332                                           struct bio *bio, int *bio_is_patched,
 333                                           struct buffer_head *bh,
 334                                           int submit_bio_bh_rw);
 335 static int btrfsic_process_written_superblock(
 336                 struct btrfsic_state *state,
 337                 struct btrfsic_block *const block,
 338                 struct btrfs_super_block *const super_hdr);
 339 static void btrfsic_bio_end_io(struct bio *bp, int bio_error_status);
 340 static void btrfsic_bh_end_io(struct buffer_head *bh, int uptodate);
 341 static int btrfsic_is_block_ref_by_superblock(const struct btrfsic_state *state,
 342                                               const struct btrfsic_block *block,
 343                                               int recursion_level);
 344 static int btrfsic_check_all_ref_blocks(struct btrfsic_state *state,
 345                                         struct btrfsic_block *const block,
 346                                         int recursion_level);
 347 static void btrfsic_print_add_link(const struct btrfsic_state *state,
 348                                    const struct btrfsic_block_link *l);
 349 static void btrfsic_print_rem_link(const struct btrfsic_state *state,
 350                                    const struct btrfsic_block_link *l);
 351 static char btrfsic_get_block_type(const struct btrfsic_state *state,
 352                                    const struct btrfsic_block *block);
 353 static void btrfsic_dump_tree(const struct btrfsic_state *state);
 354 static void btrfsic_dump_tree_sub(const struct btrfsic_state *state,
 355                                   const struct btrfsic_block *block,
 356                                   int indent_level);
 357 static struct btrfsic_block_link *btrfsic_block_link_lookup_or_add(
 358                 struct btrfsic_state *state,
 359                 struct btrfsic_block_data_ctx *next_block_ctx,
 360                 struct btrfsic_block *next_block,
 361                 struct btrfsic_block *from_block,
 362                 u64 parent_generation);
 363 static struct btrfsic_block *btrfsic_block_lookup_or_add(
 364                 struct btrfsic_state *state,
 365                 struct btrfsic_block_data_ctx *block_ctx,
 366                 const char *additional_string,
 367                 int is_metadata,
 368                 int is_iodone,
 369                 int never_written,
 370                 int mirror_num,
 371                 int *was_created);
 372 static int btrfsic_process_superblock_dev_mirror(
 373                 struct btrfsic_state *state,
 374                 struct btrfsic_dev_state *dev_state,
 375                 struct btrfs_device *device,
 376                 int superblock_mirror_num,
 377                 struct btrfsic_dev_state **selected_dev_state,
 378                 struct btrfs_super_block *selected_super);
 379 static struct btrfsic_dev_state *btrfsic_dev_state_lookup(
 380                 struct block_device *bdev);
 381 static void btrfsic_cmp_log_and_dev_bytenr(struct btrfsic_state *state,
 382                                            u64 bytenr,
 383                                            struct btrfsic_dev_state *dev_state,
 384                                            u64 dev_bytenr);
 385
 386 static struct mutex btrfsic_mutex;
 387 static int btrfsic_is_initialized;
 388 static struct btrfsic_dev_state_hashtable btrfsic_dev_state_hashtable;
 389
 390
 391 static void btrfsic_block_init(struct btrfsic_block *b)
 392 {
 393         b->magic_num = BTRFSIC_BLOCK_MAGIC_NUMBER;
 394         b->dev_state = NULL;
 395         b->dev_bytenr = 0;
 396         b->logical_bytenr = 0;
 397         b->generation = BTRFSIC_GENERATION_UNKNOWN;
 398         b->disk_key.objectid = 0;
 399         b->disk_key.type = 0;
 400         b->disk_key.offset = 0;
 401         b->is_metadata = 0;
 402         b->is_superblock = 0;
 403         b->is_iodone = 0;
 404         b->iodone_w_error = 0;
 405         b->never_written = 0;
 406         b->mirror_num = 0;
 407         b->next_in_same_bio = NULL;
 408         b->orig_bio_bh_private = NULL;
 409         b->orig_bio_bh_end_io.bio = NULL;
 410         INIT_LIST_HEAD(&b->collision_resolving_node);
 411         INIT_LIST_HEAD(&b->all_blocks_node);
 412         INIT_LIST_HEAD(&b->ref_to_list);
 413         INIT_LIST_HEAD(&b->ref_from_list);
 414         b->submit_bio_bh_rw = 0;
 415         b->flush_gen = 0;
 416 }
 417
 418 static struct btrfsic_block *btrfsic_block_alloc(void)
 419 {
 420         struct btrfsic_block *b;
 421
 422         b = kzalloc(sizeof(*b), GFP_NOFS);
 423         if (NULL != b)
 424                 btrfsic_block_init(b);
 425
 426         return b;
 427 }
 428
 429 static void btrfsic_block_free(struct btrfsic_block *b)
 430 {
 431         BUG_ON(!(NULL == b || BTRFSIC_BLOCK_MAGIC_NUMBER == b->magic_num));
 432         kfree(b);
 433 }
 434
 435 static void btrfsic_block_link_init(struct btrfsic_block_link *l)
 436 {
 437         l->magic_num = BTRFSIC_BLOCK_LINK_MAGIC_NUMBER;
 438         l->ref_cnt = 1;
 439         INIT_LIST_HEAD(&l->node_ref_to);
 440         INIT_LIST_HEAD(&l->node_ref_from);
 441         INIT_LIST_HEAD(&l->collision_resolving_node);
 442         l->block_ref_to = NULL;
 443         l->block_ref_from = NULL;
 444 }
 445
 446 static struct btrfsic_block_link *btrfsic_block_link_alloc(void)
 447 {
 448         struct btrfsic_block_link *l;
 449
 450         l = kzalloc(sizeof(*l), GFP_NOFS);
 451         if (NULL != l)
 452                 btrfsic_block_link_init(l);
 453
 454         return l;
 455 }
 456
 457 static void btrfsic_block_link_free(struct btrfsic_block_link *l)
 458 {
 459         BUG_ON(!(NULL == l || BTRFSIC_BLOCK_LINK_MAGIC_NUMBER == l->magic_num));
 460         kfree(l);
 461 }
 462
 463 static void btrfsic_dev_state_init(struct btrfsic_dev_state *ds)
 464 {
 465         ds->magic_num = BTRFSIC_DEV2STATE_MAGIC_NUMBER;
 466         ds->bdev = NULL;
 467         ds->state = NULL;
 468         ds->name[0] = '\0';
 469         INIT_LIST_HEAD(&ds->collision_resolving_node);
 470         ds->last_flush_gen = 0;
 471         btrfsic_block_init(&ds->dummy_block_for_bio_bh_flush);
 472         ds->dummy_block_for_bio_bh_flush.is_iodone = 1;
 473         ds->dummy_block_for_bio_bh_flush.dev_state = ds;
 474 }
 475
 476 static struct btrfsic_dev_state *btrfsic_dev_state_alloc(void)
 477 {
 478         struct btrfsic_dev_state *ds;
 479
 480         ds = kzalloc(sizeof(*ds), GFP_NOFS);
 481         if (NULL != ds)
 482                 btrfsic_dev_state_init(ds);
 483
 484         return ds;
 485 }
 486
 487 static void btrfsic_dev_state_free(struct btrfsic_dev_state *ds)
 488 {
 489         BUG_ON(!(NULL == ds ||
 490                  BTRFSIC_DEV2STATE_MAGIC_NUMBER == ds->magic_num));
 491         kfree(ds);
 492 }
 493
 494 static void btrfsic_block_hashtable_init(struct btrfsic_block_hashtable *h)
 495 {
 496         int i;
 497
 498         for (i = 0; i < BTRFSIC_BLOCK_HASHTABLE_SIZE; i++)
 499                 INIT_LIST_HEAD(h->table + i);
 500 }
 501
 502 static void btrfsic_block_hashtable_add(struct btrfsic_block *b,
 503                                         struct btrfsic_block_hashtable *h)
 504 {
 505         const unsigned int hashval =
 506             (((unsigned int)(b->dev_bytenr >> 16)) ^
 507              ((unsigned int)((uintptr_t)b->dev_state->bdev))) &
 508              (BTRFSIC_BLOCK_HASHTABLE_SIZE - 1);
 509
 510         list_add(&b->collision_resolving_node, h->table + hashval);
 511 }
 512
 513 static void btrfsic_block_hashtable_remove(struct btrfsic_block *b)
 514 {
 515         list_del(&b->collision_resolving_node);
 516 }
 517
 518 static struct btrfsic_block *btrfsic_block_hashtable_lookup(
 519                 struct block_device *bdev,
 520                 u64 dev_bytenr,
 521                 struct btrfsic_block_hashtable *h)
 522 {
 523         const unsigned int hashval =
 524             (((unsigned int)(dev_bytenr >> 16)) ^
 525              ((unsigned int)((uintptr_t)bdev))) &
 526              (BTRFSIC_BLOCK_HASHTABLE_SIZE - 1);
 527         struct list_head *elem;
 528
 529         list_for_each(elem, h->table + hashval) {
 530                 struct btrfsic_block *const b =
 531                     list_entry(elem, struct btrfsic_block,
 532                                collision_resolving_node);
 533
 534                 if (b->dev_state->bdev == bdev && b->dev_bytenr == dev_bytenr)
 535                         return b;
 536         }
 537
 538         return NULL;
 539 }
 540
 541 static void btrfsic_block_link_hashtable_init(
 542                 struct btrfsic_block_link_hashtable *h)
 543 {
 544         int i;
 545
 546         for (i = 0; i < BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE; i++)
 547                 INIT_LIST_HEAD(h->table + i);
 548 }
 549
 550 static void btrfsic_block_link_hashtable_add(
 551                 struct btrfsic_block_link *l,
 552                 struct btrfsic_block_link_hashtable *h)
 553 {
 554         const unsigned int hashval =
 555             (((unsigned int)(l->block_ref_to->dev_bytenr >> 16)) ^
 556              ((unsigned int)(l->block_ref_from->dev_bytenr >> 16)) ^
 557              ((unsigned int)((uintptr_t)l->block_ref_to->dev_state->bdev)) ^
 558              ((unsigned int)((uintptr_t)l->block_ref_from->dev_state->bdev)))
 559              & (BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE - 1);
 560
 561         BUG_ON(NULL == l->block_ref_to);
 562         BUG_ON(NULL == l->block_ref_from);
 563         list_add(&l->collision_resolving_node, h->table + hashval);
 564 }
 565
 566 static void btrfsic_block_link_hashtable_remove(struct btrfsic_block_link *l)
 567 {
 568         list_del(&l->collision_resolving_node);
 569 }
 570
 571 static struct btrfsic_block_link *btrfsic_block_link_hashtable_lookup(
 572                 struct block_device *bdev_ref_to,
 573                 u64 dev_bytenr_ref_to,
 574                 struct block_device *bdev_ref_from,
 575                 u64 dev_bytenr_ref_from,
 576                 struct btrfsic_block_link_hashtable *h)
 577 {
 578         const unsigned int hashval =
 579             (((unsigned int)(dev_bytenr_ref_to >> 16)) ^
 580              ((unsigned int)(dev_bytenr_ref_from >> 16)) ^
 581              ((unsigned int)((uintptr_t)bdev_ref_to)) ^
 582              ((unsigned int)((uintptr_t)bdev_ref_from))) &
 583              (BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE - 1);
 584         struct list_head *elem;
 585
 586         list_for_each(elem, h->table + hashval) {
 587                 struct btrfsic_block_link *const l =
 588                     list_entry(elem, struct btrfsic_block_link,
 589                                collision_resolving_node);
 590
 591                 BUG_ON(NULL == l->block_ref_to);
 592                 BUG_ON(NULL == l->block_ref_from);
 593                 if (l->block_ref_to->dev_state->bdev == bdev_ref_to &&
 594                     l->block_ref_to->dev_bytenr == dev_bytenr_ref_to &&
 595                     l->block_ref_from->dev_state->bdev == bdev_ref_from &&
 596                     l->block_ref_from->dev_bytenr == dev_bytenr_ref_from)
 597                         return l;
 598         }
 599
 600         return NULL;
 601 }
 602
 603 static void btrfsic_dev_state_hashtable_init(
 604                 struct btrfsic_dev_state_hashtable *h)
 605 {
 606         int i;
 607
 608         for (i = 0; i < BTRFSIC_DEV2STATE_HASHTABLE_SIZE; i++)
 609                 INIT_LIST_HEAD(h->table + i);
 610 }
 611
 612 static void btrfsic_dev_state_hashtable_add(
 613                 struct btrfsic_dev_state *ds,
 614                 struct btrfsic_dev_state_hashtable *h)
 615 {
 616         const unsigned int hashval =
 617             (((unsigned int)((uintptr_t)ds->bdev)) &
 618              (BTRFSIC_DEV2STATE_HASHTABLE_SIZE - 1));
 619
 620         list_add(&ds->collision_resolving_node, h->table + hashval);
 621 }
 622
 623 static void btrfsic_dev_state_hashtable_remove(struct btrfsic_dev_state *ds)
 624 {
 625         list_del(&ds->collision_resolving_node);
 626 }
 627
 628 static struct btrfsic_dev_state *btrfsic_dev_state_hashtable_lookup(
 629                 struct block_device *bdev,
 630                 struct btrfsic_dev_state_hashtable *h)
 631 {
 632         const unsigned int hashval =
 633             (((unsigned int)((uintptr_t)bdev)) &
 634              (BTRFSIC_DEV2STATE_HASHTABLE_SIZE - 1));
 635         struct list_head *elem;
 636
 637         list_for_each(elem, h->table + hashval) {
 638                 struct btrfsic_dev_state *const ds =
 639                     list_entry(elem, struct btrfsic_dev_state,
 640                                collision_resolving_node);
 641
 642                 if (ds->bdev == bdev)
 643                         return ds;
 644         }
 645
 646         return NULL;
 647 }
 648
 649 static int btrfsic_process_superblock(struct btrfsic_state *state,
 650                                       struct btrfs_fs_devices *fs_devices)
 651 {
 652         int ret = 0;
 653         struct btrfs_super_block *selected_super;
 654         struct list_head *dev_head = &fs_devices->devices;
 655         struct btrfs_device *device;
 656         struct btrfsic_dev_state *selected_dev_state = NULL;
 657         int pass;
 658
 659         BUG_ON(NULL == state);
 660         selected_super = kzalloc(sizeof(*selected_super), GFP_NOFS);
 661         if (NULL == selected_super) {
 662                 printk(KERN_INFO "btrfsic: error, kmalloc failed!\n");
 663                 return -1;
 664         }
 665
 666         list_for_each_entry(device, dev_head, dev_list) {
 667                 int i;
 668                 struct btrfsic_dev_state *dev_state;
 669
 670                 if (!device->bdev || !device->name)
 671                         continue;
 672
 673                 dev_state = btrfsic_dev_state_lookup(device->bdev);
 674                 BUG_ON(NULL == dev_state);
 675                 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 676                         ret = btrfsic_process_superblock_dev_mirror(
 677                                         state, dev_state, device, i,
 678                                         &selected_dev_state, selected_super);
 679                         if (0 != ret && 0 == i) {
 680                                 kfree(selected_super);
 681                                 return ret;
 682                         }
 683                 }
 684         }
 685
 686         if (NULL == state->latest_superblock) {
 687                 printk(KERN_INFO "btrfsic: no superblock found!\n");
 688                 kfree(selected_super);
 689                 return -1;
 690         }
 691
 692         state->csum_size = btrfs_super_csum_size(selected_super);
 693
 694         for (pass = 0; pass < 3; pass++) {
 695                 int num_copies;
 696                 int mirror_num;
 697                 u64 next_bytenr;
 698
 699                 switch (pass) {
 700                 case 0:
 701                         next_bytenr = btrfs_super_root(selected_super);
 702                         if (state->print_mask &
 703                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
 704                                 printk(KERN_INFO "root@%llu\n", next_bytenr);
 705                         break;
 706                 case 1:
 707                         next_bytenr = btrfs_super_chunk_root(selected_super);
 708                         if (state->print_mask &
 709                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
 710                                 printk(KERN_INFO "chunk@%llu\n", next_bytenr);
 711                         break;
 712                 case 2:
 713                         next_bytenr = btrfs_super_log_root(selected_super);
 714                         if (0 == next_bytenr)
 715                                 continue;
 716                         if (state->print_mask &
 717                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
 718                                 printk(KERN_INFO "log@%llu\n", next_bytenr);
 719                         break;
 720                 }
 721
 722                 num_copies =
 723                     btrfs_num_copies(state->root->fs_info,
 724                                      next_bytenr, state->metablock_size);
 725                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
 726                         printk(KERN_INFO "num_copies(log_bytenr=%llu) = %d\n",
 727                                next_bytenr, num_copies);
 728
 729                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
 730                         struct btrfsic_block *next_block;
 731                         struct btrfsic_block_data_ctx tmp_next_block_ctx;
 732                         struct btrfsic_block_link *l;
 733
 734                         ret = btrfsic_map_block(state, next_bytenr,
 735                                                 state->metablock_size,
 736                                                 &tmp_next_block_ctx,
 737                                                 mirror_num);
 738                         if (ret) {
 739                                 printk(KERN_INFO "btrfsic:"
 740                                        " btrfsic_map_block(root @%llu,"
 741                                        " mirror %d) failed!\n",
 742                                        next_bytenr, mirror_num);
 743                                 kfree(selected_super);
 744                                 return -1;
 745                         }
 746
 747                         next_block = btrfsic_block_hashtable_lookup(
 748                                         tmp_next_block_ctx.dev->bdev,
 749                                         tmp_next_block_ctx.dev_bytenr,
 750                                         &state->block_hashtable);
 751                         BUG_ON(NULL == next_block);
 752
 753                         l = btrfsic_block_link_hashtable_lookup(
 754                                         tmp_next_block_ctx.dev->bdev,
 755                                         tmp_next_block_ctx.dev_bytenr,
 756                                         state->latest_superblock->dev_state->
 757                                         bdev,
 758                                         state->latest_superblock->dev_bytenr,
 759                                         &state->block_link_hashtable);
 760                         BUG_ON(NULL == l);
 761
 762                         ret = btrfsic_read_block(state, &tmp_next_block_ctx);
 763                         if (ret < (int)PAGE_CACHE_SIZE) {
 764                                 printk(KERN_INFO
 765                                        "btrfsic: read @logical %llu failed!\n",
 766                                        tmp_next_block_ctx.start);
 767                                 btrfsic_release_block_ctx(&tmp_next_block_ctx);
 768                                 kfree(selected_super);
 769                                 return -1;
 770                         }
 771
 772                         ret = btrfsic_process_metablock(state,
 773                                                         next_block,
 774                                                         &tmp_next_block_ctx,
 775                                                         BTRFS_MAX_LEVEL + 3, 1);
 776                         btrfsic_release_block_ctx(&tmp_next_block_ctx);
 777                 }
 778         }
 779
 780         kfree(selected_super);
 781         return ret;
 782 }
 783
 784 static int btrfsic_process_superblock_dev_mirror(
 785                 struct btrfsic_state *state,
 786                 struct btrfsic_dev_state *dev_state,
 787                 struct btrfs_device *device,
 788                 int superblock_mirror_num,
 789                 struct btrfsic_dev_state **selected_dev_state,
 790                 struct btrfs_super_block *selected_super)
 791 {
 792         struct btrfs_super_block *super_tmp;
 793         u64 dev_bytenr;
 794         struct buffer_head *bh;
 795         struct btrfsic_block *superblock_tmp;
 796         int pass;
 797         struct block_device *const superblock_bdev = device->bdev;
 798
 799         /* super block bytenr is always the unmapped device bytenr */
 800         dev_bytenr = btrfs_sb_offset(superblock_mirror_num);
 801         if (dev_bytenr + BTRFS_SUPER_INFO_SIZE > device->total_bytes)
 802                 return -1;
 803         bh = __bread(superblock_bdev, dev_bytenr / 4096,
 804                      BTRFS_SUPER_INFO_SIZE);
 805         if (NULL == bh)
 806                 return -1;
 807         super_tmp = (struct btrfs_super_block *)
 808             (bh->b_data + (dev_bytenr & 4095));
 809
 810         if (btrfs_super_bytenr(super_tmp) != dev_bytenr ||
 811             btrfs_super_magic(super_tmp) != BTRFS_MAGIC ||
 812             memcmp(device->uuid, super_tmp->dev_item.uuid, BTRFS_UUID_SIZE) ||
 813             btrfs_super_nodesize(super_tmp) != state->metablock_size ||
 814             btrfs_super_leafsize(super_tmp) != state->metablock_size ||
 815             btrfs_super_sectorsize(super_tmp) != state->datablock_size) {
 816                 brelse(bh);
 817                 return 0;
 818         }
 819
 820         superblock_tmp =
 821             btrfsic_block_hashtable_lookup(superblock_bdev,
 822                                            dev_bytenr,
 823                                            &state->block_hashtable);
 824         if (NULL == superblock_tmp) {
 825                 superblock_tmp = btrfsic_block_alloc();
 826                 if (NULL == superblock_tmp) {
 827                         printk(KERN_INFO "btrfsic: error, kmalloc failed!\n");
 828                         brelse(bh);
 829                         return -1;
 830                 }
 831                 /* for superblock, only the dev_bytenr makes sense */
 832                 superblock_tmp->dev_bytenr = dev_bytenr;
 833                 superblock_tmp->dev_state = dev_state;
 834                 superblock_tmp->logical_bytenr = dev_bytenr;
 835                 superblock_tmp->generation = btrfs_super_generation(super_tmp);
 836                 superblock_tmp->is_metadata = 1;
 837                 superblock_tmp->is_superblock = 1;
 838                 superblock_tmp->is_iodone = 1;
 839                 superblock_tmp->never_written = 0;
 840                 superblock_tmp->mirror_num = 1 + superblock_mirror_num;
 841                 if (state->print_mask & BTRFSIC_PRINT_MASK_SUPERBLOCK_WRITE)
 842                         printk_in_rcu(KERN_INFO "New initial S-block (bdev %p, %s)"
 843                                      " @%llu (%s/%llu/%d)\n",
 844                                      superblock_bdev,
 845                                      rcu_str_deref(device->name), dev_bytenr,
 846                                      dev_state->name, dev_bytenr,
 847                                      superblock_mirror_num);
 848                 list_add(&superblock_tmp->all_blocks_node,
 849                          &state->all_blocks_list);
 850                 btrfsic_block_hashtable_add(superblock_tmp,
 851                                             &state->block_hashtable);
 852         }
 853
 854         /* select the one with the highest generation field */
 855         if (btrfs_super_generation(super_tmp) >
 856             state->max_superblock_generation ||
 857             0 == state->max_superblock_generation) {
 858                 memcpy(selected_super, super_tmp, sizeof(*selected_super));
 859                 *selected_dev_state = dev_state;
 860                 state->max_superblock_generation =
 861                     btrfs_super_generation(super_tmp);
 862                 state->latest_superblock = superblock_tmp;
 863         }
 864
 865         for (pass = 0; pass < 3; pass++) {
 866                 u64 next_bytenr;
 867                 int num_copies;
 868                 int mirror_num;
 869                 const char *additional_string = NULL;
 870                 struct btrfs_disk_key tmp_disk_key;
 871
 872                 tmp_disk_key.type = BTRFS_ROOT_ITEM_KEY;
 873                 tmp_disk_key.offset = 0;
 874                 switch (pass) {
 875                 case 0:
 876                         btrfs_set_disk_key_objectid(&tmp_disk_key,
 877                                                     BTRFS_ROOT_TREE_OBJECTID);
 878                         additional_string = "initial root ";
 879                         next_bytenr = btrfs_super_root(super_tmp);
 880                         break;
 881                 case 1:
 882                         btrfs_set_disk_key_objectid(&tmp_disk_key,
 883                                                     BTRFS_CHUNK_TREE_OBJECTID);
 884                         additional_string = "initial chunk ";
 885                         next_bytenr = btrfs_super_chunk_root(super_tmp);
 886                         break;
 887                 case 2:
 888                         btrfs_set_disk_key_objectid(&tmp_disk_key,
 889                                                     BTRFS_TREE_LOG_OBJECTID);
 890                         additional_string = "initial log ";
 891                         next_bytenr = btrfs_super_log_root(super_tmp);
 892                         if (0 == next_bytenr)
 893                                 continue;
 894                         break;
 895                 }
 896
 897                 num_copies =
 898                     btrfs_num_copies(state->root->fs_info,
 899                                      next_bytenr, state->metablock_size);
 900                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
 901                         printk(KERN_INFO "num_copies(log_bytenr=%llu) = %d\n",
 902                                next_bytenr, num_copies);
 903                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
 904                         struct btrfsic_block *next_block;
 905                         struct btrfsic_block_data_ctx tmp_next_block_ctx;
 906                         struct btrfsic_block_link *l;
 907
 908                         if (btrfsic_map_block(state, next_bytenr,
 909                                               state->metablock_size,
 910                                               &tmp_next_block_ctx,
 911                                               mirror_num)) {
 912                                 printk(KERN_INFO "btrfsic: btrfsic_map_block("
 913                                        "bytenr @%llu, mirror %d) failed!\n",
 914                                        next_bytenr, mirror_num);
 915                                 brelse(bh);
 916                                 return -1;
 917                         }
 918
 919                         next_block = btrfsic_block_lookup_or_add(
 920                                         state, &tmp_next_block_ctx,
 921                                         additional_string, 1, 1, 0,
 922                                         mirror_num, NULL);
 923                         if (NULL == next_block) {
 924                                 btrfsic_release_block_ctx(&tmp_next_block_ctx);
 925                                 brelse(bh);
 926                                 return -1;
 927                         }
 928
 929                         next_block->disk_key = tmp_disk_key;
 930                         next_block->generation = BTRFSIC_GENERATION_UNKNOWN;
 931                         l = btrfsic_block_link_lookup_or_add(
 932                                         state, &tmp_next_block_ctx,
 933                                         next_block, superblock_tmp,
 934                                         BTRFSIC_GENERATION_UNKNOWN);
 935                         btrfsic_release_block_ctx(&tmp_next_block_ctx);
 936                         if (NULL == l) {
 937                                 brelse(bh);
 938                                 return -1;
 939                         }
 940                 }
 941         }
 942         if (state->print_mask & BTRFSIC_PRINT_MASK_INITIAL_ALL_TREES)
 943                 btrfsic_dump_tree_sub(state, superblock_tmp, 0);
 944
 945         brelse(bh);
 946         return 0;
 947 }
 948
 949 static struct btrfsic_stack_frame *btrfsic_stack_frame_alloc(void)
 950 {
 951         struct btrfsic_stack_frame *sf;
 952
 953         sf = kzalloc(sizeof(*sf), GFP_NOFS);
 954         if (NULL == sf)
 955                 printk(KERN_INFO "btrfsic: alloc memory failed!\n");
 956         else
 957                 sf->magic = BTRFSIC_BLOCK_STACK_FRAME_MAGIC_NUMBER;
 958         return sf;
 959 }
 960
 961 static void btrfsic_stack_frame_free(struct btrfsic_stack_frame *sf)
 962 {
 963         BUG_ON(!(NULL == sf ||
 964                  BTRFSIC_BLOCK_STACK_FRAME_MAGIC_NUMBER == sf->magic));
 965         kfree(sf);
 966 }
 967
 968 static int btrfsic_process_metablock(
 969                 struct btrfsic_state *state,
 970                 struct btrfsic_block *const first_block,
 971                 struct btrfsic_block_data_ctx *const first_block_ctx,
 972                 int first_limit_nesting, int force_iodone_flag)
 973 {
 974         struct btrfsic_stack_frame initial_stack_frame = { 0 };
 975         struct btrfsic_stack_frame *sf;
 976         struct btrfsic_stack_frame *next_stack;
 977         struct btrfs_header *const first_hdr =
 978                 (struct btrfs_header *)first_block_ctx->datav[0];
 979
 980         BUG_ON(!first_hdr);
 981         sf = &initial_stack_frame;
 982         sf->error = 0;
 983         sf->i = -1;
 984         sf->limit_nesting = first_limit_nesting;
 985         sf->block = first_block;
 986         sf->block_ctx = first_block_ctx;
 987         sf->next_block = NULL;
 988         sf->hdr = first_hdr;
 989         sf->prev = NULL;
 990
 991 continue_with_new_stack_frame:
 992         sf->block->generation = le64_to_cpu(sf->hdr->generation);
 993         if (0 == sf->hdr->level) {
 994                 struct btrfs_leaf *const leafhdr =
 995                     (struct btrfs_leaf *)sf->hdr;
 996
 997                 if (-1 == sf->i) {
 998                         sf->nr = btrfs_stack_header_nritems(&leafhdr->header);
 999
1000                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1001                                 printk(KERN_INFO
1002                                        "leaf %llu items %d generation %llu"
1003                                        " owner %llu\n",
1004                                        sf->block_ctx->start, sf->nr,
1005                                        btrfs_stack_header_generation(
1006                                                &leafhdr->header),
1007                                        btrfs_stack_header_owner(
1008                                                &leafhdr->header));
1009                 }
1010
1011 continue_with_current_leaf_stack_frame:
1012                 if (0 == sf->num_copies || sf->mirror_num > sf->num_copies) {
1013                         sf->i++;
1014                         sf->num_copies = 0;
1015                 }
1016
1017                 if (sf->i < sf->nr) {
1018                         struct btrfs_item disk_item;
1019                         u32 disk_item_offset =
1020                                 (uintptr_t)(leafhdr->items + sf->i) -
1021                                 (uintptr_t)leafhdr;
1022                         struct btrfs_disk_key *disk_key;
1023                         u8 type;
1024                         u32 item_offset;
1025                         u32 item_size;
1026
1027                         if (disk_item_offset + sizeof(struct btrfs_item) >
1028                             sf->block_ctx->len) {
1029 leaf_item_out_of_bounce_error:
1030                                 printk(KERN_INFO
1031                                        "btrfsic: leaf item out of bounce at logical %llu, dev %s\n",
1032                                        sf->block_ctx->start,
1033                                        sf->block_ctx->dev->name);
1034                                 goto one_stack_frame_backwards;
1035                         }
1036                         btrfsic_read_from_block_data(sf->block_ctx,
1037                                                      &disk_item,
1038                                                      disk_item_offset,
1039                                                      sizeof(struct btrfs_item));
1040                         item_offset = btrfs_stack_item_offset(&disk_item);
1041                         item_size = btrfs_stack_item_offset(&disk_item);
1042                         disk_key = &disk_item.key;
1043                         type = btrfs_disk_key_type(disk_key);
1044
1045                         if (BTRFS_ROOT_ITEM_KEY == type) {
1046                                 struct btrfs_root_item root_item;
1047                                 u32 root_item_offset;
1048                                 u64 next_bytenr;
1049
1050                                 root_item_offset = item_offset +
1051                                         offsetof(struct btrfs_leaf, items);
1052                                 if (root_item_offset + item_size >
1053                                     sf->block_ctx->len)
1054                                         goto leaf_item_out_of_bounce_error;
1055                                 btrfsic_read_from_block_data(
1056                                         sf->block_ctx, &root_item,
1057                                         root_item_offset,
1058                                         item_size);
1059                                 next_bytenr = btrfs_root_bytenr(&root_item);
1060
1061                                 sf->error =
1062                                     btrfsic_create_link_to_next_block(
1063                                                 state,
1064                                                 sf->block,
1065                                                 sf->block_ctx,
1066                                                 next_bytenr,
1067                                                 sf->limit_nesting,
1068                                                 &sf->next_block_ctx,
1069                                                 &sf->next_block,
1070                                                 force_iodone_flag,
1071                                                 &sf->num_copies,
1072                                                 &sf->mirror_num,
1073                                                 disk_key,
1074                                                 btrfs_root_generation(
1075                                                 &root_item));
1076                                 if (sf->error)
1077                                         goto one_stack_frame_backwards;
1078
1079                                 if (NULL != sf->next_block) {
1080                                         struct btrfs_header *const next_hdr =
1081                                             (struct btrfs_header *)
1082                                             sf->next_block_ctx.datav[0];
1083
1084                                         next_stack =
1085                                             btrfsic_stack_frame_alloc();
1086                                         if (NULL == next_stack) {
1087                                                 btrfsic_release_block_ctx(
1088                                                                 &sf->
1089                                                                 next_block_ctx);
1090                                                 goto one_stack_frame_backwards;
1091                                         }
1092
1093                                         next_stack->i = -1;
1094                                         next_stack->block = sf->next_block;
1095                                         next_stack->block_ctx =
1096                                             &sf->next_block_ctx;
1097                                         next_stack->next_block = NULL;
1098                                         next_stack->hdr = next_hdr;
1099                                         next_stack->limit_nesting =
1100                                             sf->limit_nesting - 1;
1101                                         next_stack->prev = sf;
1102                                         sf = next_stack;
1103                                         goto continue_with_new_stack_frame;
1104                                 }
1105                         } else if (BTRFS_EXTENT_DATA_KEY == type &&
1106                                    state->include_extent_data) {
1107                                 sf->error = btrfsic_handle_extent_data(
1108                                                 state,
1109                                                 sf->block,
1110                                                 sf->block_ctx,
1111                                                 item_offset,
1112                                                 force_iodone_flag);
1113                                 if (sf->error)
1114                                         goto one_stack_frame_backwards;
1115                         }
1116
1117                         goto continue_with_current_leaf_stack_frame;
1118                 }
1119         } else {
1120                 struct btrfs_node *const nodehdr = (struct btrfs_node *)sf->hdr;
1121
1122                 if (-1 == sf->i) {
1123                         sf->nr = btrfs_stack_header_nritems(&nodehdr->header);
1124
1125                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1126                                 printk(KERN_INFO "node %llu level %d items %d"
1127                                        " generation %llu owner %llu\n",
1128                                        sf->block_ctx->start,
1129                                        nodehdr->header.level, sf->nr,
1130                                        btrfs_stack_header_generation(
1131                                        &nodehdr->header),
1132                                        btrfs_stack_header_owner(
1133                                        &nodehdr->header));
1134                 }
1135
1136 continue_with_current_node_stack_frame:
1137                 if (0 == sf->num_copies || sf->mirror_num > sf->num_copies) {
1138                         sf->i++;
1139                         sf->num_copies = 0;
1140                 }
1141
1142                 if (sf->i < sf->nr) {
1143                         struct btrfs_key_ptr key_ptr;
1144                         u32 key_ptr_offset;
1145                         u64 next_bytenr;
1146
1147                         key_ptr_offset = (uintptr_t)(nodehdr->ptrs + sf->i) -
1148                                           (uintptr_t)nodehdr;
1149                         if (key_ptr_offset + sizeof(struct btrfs_key_ptr) >
1150                             sf->block_ctx->len) {
1151                                 printk(KERN_INFO
1152                                        "btrfsic: node item out of bounce at logical %llu, dev %s\n",
1153                                        sf->block_ctx->start,
1154                                        sf->block_ctx->dev->name);
1155                                 goto one_stack_frame_backwards;
1156                         }
1157                         btrfsic_read_from_block_data(
1158                                 sf->block_ctx, &key_ptr, key_ptr_offset,
1159                                 sizeof(struct btrfs_key_ptr));
1160                         next_bytenr = btrfs_stack_key_blockptr(&key_ptr);
1161
1162                         sf->error = btrfsic_create_link_to_next_block(
1163                                         state,
1164                                         sf->block,
1165                                         sf->block_ctx,
1166                                         next_bytenr,
1167                                         sf->limit_nesting,
1168                                         &sf->next_block_ctx,
1169                                         &sf->next_block,
1170                                         force_iodone_flag,
1171                                         &sf->num_copies,
1172                                         &sf->mirror_num,
1173                                         &key_ptr.key,
1174                                         btrfs_stack_key_generation(&key_ptr));
1175                         if (sf->error)
1176                                 goto one_stack_frame_backwards;
1177
1178                         if (NULL != sf->next_block) {
1179                                 struct btrfs_header *const next_hdr =
1180                                     (struct btrfs_header *)
1181                                     sf->next_block_ctx.datav[0];
1182
1183                                 next_stack = btrfsic_stack_frame_alloc();
1184                                 if (NULL == next_stack)
1185                                         goto one_stack_frame_backwards;
1186
1187                                 next_stack->i = -1;
1188                                 next_stack->block = sf->next_block;
1189                                 next_stack->block_ctx = &sf->next_block_ctx;
1190                                 next_stack->next_block = NULL;
1191                                 next_stack->hdr = next_hdr;
1192                                 next_stack->limit_nesting =
1193                                     sf->limit_nesting - 1;
1194                                 next_stack->prev = sf;
1195                                 sf = next_stack;
1196                                 goto continue_with_new_stack_frame;
1197                         }
1198
1199                         goto continue_with_current_node_stack_frame;
1200                 }
1201         }
1202
1203 one_stack_frame_backwards:
1204         if (NULL != sf->prev) {
1205                 struct btrfsic_stack_frame *const prev = sf->prev;
1206
1207                 /* the one for the initial block is freed in the caller */
1208                 btrfsic_release_block_ctx(sf->block_ctx);
1209
1210                 if (sf->error) {
1211                         prev->error = sf->error;
1212                         btrfsic_stack_frame_free(sf);
1213                         sf = prev;
1214                         goto one_stack_frame_backwards;
1215                 }
1216
1217                 btrfsic_stack_frame_free(sf);
1218                 sf = prev;
1219                 goto continue_with_new_stack_frame;
1220         } else {
1221                 BUG_ON(&initial_stack_frame != sf);
1222         }
1223
1224         return sf->error;
1225 }
1226
1227 static void btrfsic_read_from_block_data(
1228         struct btrfsic_block_data_ctx *block_ctx,
1229         void *dstv, u32 offset, size_t len)
1230 {
1231         size_t cur;
1232         size_t offset_in_page;
1233         char *kaddr;
1234         char *dst = (char *)dstv;
1235         size_t start_offset = block_ctx->start & ((u64)PAGE_CACHE_SIZE - 1);
1236         unsigned long i = (start_offset + offset) >> PAGE_CACHE_SHIFT;
1237
1238         WARN_ON(offset + len > block_ctx->len);
1239         offset_in_page = (start_offset + offset) &
1240                          ((unsigned long)PAGE_CACHE_SIZE - 1);
1241
1242         while (len > 0) {
1243                 cur = min(len, ((size_t)PAGE_CACHE_SIZE - offset_in_page));
1244                 BUG_ON(i >= (block_ctx->len + PAGE_CACHE_SIZE - 1) >>
1245                             PAGE_CACHE_SHIFT);
1246                 kaddr = block_ctx->datav[i];
1247                 memcpy(dst, kaddr + offset_in_page, cur);
1248
1249                 dst += cur;
1250                 len -= cur;
1251                 offset_in_page = 0;
1252                 i++;
1253         }
1254 }
1255
1256 static int btrfsic_create_link_to_next_block(
1257                 struct btrfsic_state *state,
1258                 struct btrfsic_block *block,
1259                 struct btrfsic_block_data_ctx *block_ctx,
1260                 u64 next_bytenr,
1261                 int limit_nesting,
1262                 struct btrfsic_block_data_ctx *next_block_ctx,
1263                 struct btrfsic_block **next_blockp,
1264                 int force_iodone_flag,
1265                 int *num_copiesp, int *mirror_nump,
1266                 struct btrfs_disk_key *disk_key,
1267                 u64 parent_generation)
1268 {
1269         struct btrfsic_block *next_block = NULL;
1270         int ret;
1271         struct btrfsic_block_link *l;
1272         int did_alloc_block_link;
1273         int block_was_created;
1274
1275         *next_blockp = NULL;
1276         if (0 == *num_copiesp) {
1277                 *num_copiesp =
1278                     btrfs_num_copies(state->root->fs_info,
1279                                      next_bytenr, state->metablock_size);
1280                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
1281                         printk(KERN_INFO "num_copies(log_bytenr=%llu) = %d\n",
1282                                next_bytenr, *num_copiesp);
1283                 *mirror_nump = 1;
1284         }
1285
1286         if (*mirror_nump > *num_copiesp)
1287                 return 0;
1288
1289         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1290                 printk(KERN_INFO
1291                        "btrfsic_create_link_to_next_block(mirror_num=%d)\n",
1292                        *mirror_nump);
1293         ret = btrfsic_map_block(state, next_bytenr,
1294                                 state->metablock_size,
1295                                 next_block_ctx, *mirror_nump);
1296         if (ret) {
1297                 printk(KERN_INFO
1298                        "btrfsic: btrfsic_map_block(@%llu, mirror=%d) failed!\n",
1299                        next_bytenr, *mirror_nump);
1300                 btrfsic_release_block_ctx(next_block_ctx);
1301                 *next_blockp = NULL;
1302                 return -1;
1303         }
1304
1305         next_block = btrfsic_block_lookup_or_add(state,
1306                                                  next_block_ctx, "referenced ",
1307                                                  1, force_iodone_flag,
1308                                                  !force_iodone_flag,
1309                                                  *mirror_nump,
1310                                                  &block_was_created);
1311         if (NULL == next_block) {
1312                 btrfsic_release_block_ctx(next_block_ctx);
1313                 *next_blockp = NULL;
1314                 return -1;
1315         }
1316         if (block_was_created) {
1317                 l = NULL;
1318                 next_block->generation = BTRFSIC_GENERATION_UNKNOWN;
1319         } else {
1320                 if (next_block->logical_bytenr != next_bytenr &&
1321                     !(!next_block->is_metadata &&
1322                       0 == next_block->logical_bytenr)) {
1323                         printk(KERN_INFO
1324                                "Referenced block @%llu (%s/%llu/%d)"
1325                                " found in hash table, %c,"
1326                                " bytenr mismatch (!= stored %llu).\n",
1327                                next_bytenr, next_block_ctx->dev->name,
1328                                next_block_ctx->dev_bytenr, *mirror_nump,
1329                                btrfsic_get_block_type(state, next_block),
1330                                next_block->logical_bytenr);
1331                 } else if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1332                         printk(KERN_INFO
1333                                "Referenced block @%llu (%s/%llu/%d)"
1334                                " found in hash table, %c.\n",
1335                                next_bytenr, next_block_ctx->dev->name,
1336                                next_block_ctx->dev_bytenr, *mirror_nump,
1337                                btrfsic_get_block_type(state, next_block));
1338                 next_block->logical_bytenr = next_bytenr;
1339
1340                 next_block->mirror_num = *mirror_nump;
1341                 l = btrfsic_block_link_hashtable_lookup(
1342                                 next_block_ctx->dev->bdev,
1343                                 next_block_ctx->dev_bytenr,
1344                                 block_ctx->dev->bdev,
1345                                 block_ctx->dev_bytenr,
1346                                 &state->block_link_hashtable);
1347         }
1348
1349         next_block->disk_key = *disk_key;
1350         if (NULL == l) {
1351                 l = btrfsic_block_link_alloc();
1352                 if (NULL == l) {
1353                         printk(KERN_INFO "btrfsic: error, kmalloc failed!\n");
1354                         btrfsic_release_block_ctx(next_block_ctx);
1355                         *next_blockp = NULL;
1356                         return -1;
1357                 }
1358
1359                 did_alloc_block_link = 1;
1360                 l->block_ref_to = next_block;
1361                 l->block_ref_from = block;
1362                 l->ref_cnt = 1;
1363                 l->parent_generation = parent_generation;
1364
1365                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1366                         btrfsic_print_add_link(state, l);
1367
1368                 list_add(&l->node_ref_to, &block->ref_to_list);
1369                 list_add(&l->node_ref_from, &next_block->ref_from_list);
1370
1371                 btrfsic_block_link_hashtable_add(l,
1372                                                  &state->block_link_hashtable);
1373         } else {
1374                 did_alloc_block_link = 0;
1375                 if (0 == limit_nesting) {
1376                         l->ref_cnt++;
1377                         l->parent_generation = parent_generation;
1378                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1379                                 btrfsic_print_add_link(state, l);
1380                 }
1381         }
1382
1383         if (limit_nesting > 0 && did_alloc_block_link) {
1384                 ret = btrfsic_read_block(state, next_block_ctx);
1385                 if (ret < (int)next_block_ctx->len) {
1386                         printk(KERN_INFO
1387                                "btrfsic: read block @logical %llu failed!\n",
1388                                next_bytenr);
1389                         btrfsic_release_block_ctx(next_block_ctx);
1390                         *next_blockp = NULL;
1391                         return -1;
1392                 }
1393
1394                 *next_blockp = next_block;
1395         } else {
1396                 *next_blockp = NULL;
1397         }
1398         (*mirror_nump)++;
1399
1400         return 0;
1401 }
1402
1403 static int btrfsic_handle_extent_data(
1404                 struct btrfsic_state *state,
1405                 struct btrfsic_block *block,
1406                 struct btrfsic_block_data_ctx *block_ctx,
1407                 u32 item_offset, int force_iodone_flag)
1408 {
1409         int ret;
1410         struct btrfs_file_extent_item file_extent_item;
1411         u64 file_extent_item_offset;
1412         u64 next_bytenr;
1413         u64 num_bytes;
1414         u64 generation;
1415         struct btrfsic_block_link *l;
1416
1417         file_extent_item_offset = offsetof(struct btrfs_leaf, items) +
1418                                   item_offset;
1419         if (file_extent_item_offset +
1420             offsetof(struct btrfs_file_extent_item, disk_num_bytes) >
1421             block_ctx->len) {
1422                 printk(KERN_INFO
1423                        "btrfsic: file item out of bounce at logical %llu, dev %s\n",
1424                        block_ctx->start, block_ctx->dev->name);
1425                 return -1;
1426         }
1427
1428         btrfsic_read_from_block_data(block_ctx, &file_extent_item,
1429                 file_extent_item_offset,
1430                 offsetof(struct btrfs_file_extent_item, disk_num_bytes));
1431         if (BTRFS_FILE_EXTENT_REG != file_extent_item.type ||
1432             btrfs_stack_file_extent_disk_bytenr(&file_extent_item) == 0) {
1433                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERY_VERBOSE)
1434                         printk(KERN_INFO "extent_data: type %u, disk_bytenr = %llu\n",
1435                                file_extent_item.type,
1436                                btrfs_stack_file_extent_disk_bytenr(
1437                                &file_extent_item));
1438                 return 0;
1439         }
1440
1441         if (file_extent_item_offset + sizeof(struct btrfs_file_extent_item) >
1442             block_ctx->len) {
1443                 printk(KERN_INFO
1444                        "btrfsic: file item out of bounce at logical %llu, dev %s\n",
1445                        block_ctx->start, block_ctx->dev->name);
1446                 return -1;
1447         }
1448         btrfsic_read_from_block_data(block_ctx, &file_extent_item,
1449                                      file_extent_item_offset,
1450                                      sizeof(struct btrfs_file_extent_item));
1451         next_bytenr = btrfs_stack_file_extent_disk_bytenr(&file_extent_item) +
1452                       btrfs_stack_file_extent_offset(&file_extent_item);
1453         generation = btrfs_stack_file_extent_generation(&file_extent_item);
1454         num_bytes = btrfs_stack_file_extent_num_bytes(&file_extent_item);
1455         generation = btrfs_stack_file_extent_generation(&file_extent_item);
1456
1457         if (state->print_mask & BTRFSIC_PRINT_MASK_VERY_VERBOSE)
1458                 printk(KERN_INFO "extent_data: type %u, disk_bytenr = %llu,"
1459                        " offset = %llu, num_bytes = %llu\n",
1460                        file_extent_item.type,
1461                        btrfs_stack_file_extent_disk_bytenr(&file_extent_item),
1462                        btrfs_stack_file_extent_offset(&file_extent_item),
1463                        num_bytes);
1464         while (num_bytes > 0) {
1465                 u32 chunk_len;
1466                 int num_copies;
1467                 int mirror_num;
1468
1469                 if (num_bytes > state->datablock_size)
1470                         chunk_len = state->datablock_size;
1471                 else
1472                         chunk_len = num_bytes;
1473
1474                 num_copies =
1475                     btrfs_num_copies(state->root->fs_info,
1476                                      next_bytenr, state->datablock_size);
1477                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
1478                         printk(KERN_INFO "num_copies(log_bytenr=%llu) = %d\n",
1479                                next_bytenr, num_copies);
1480                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
1481                         struct btrfsic_block_data_ctx next_block_ctx;
1482                         struct btrfsic_block *next_block;
1483                         int block_was_created;
1484
1485                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1486                                 printk(KERN_INFO "btrfsic_handle_extent_data("
1487                                        "mirror_num=%d)\n", mirror_num);
1488                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERY_VERBOSE)
1489                                 printk(KERN_INFO
1490                                        "\tdisk_bytenr = %llu, num_bytes %u\n",
1491                                        next_bytenr, chunk_len);
1492                         ret = btrfsic_map_block(state, next_bytenr,
1493                                                 chunk_len, &next_block_ctx,
1494                                                 mirror_num);
1495                         if (ret) {
1496                                 printk(KERN_INFO
1497                                        "btrfsic: btrfsic_map_block(@%llu,"
1498                                        " mirror=%d) failed!\n",
1499                                        next_bytenr, mirror_num);
1500                                 return -1;
1501                         }
1502
1503                         next_block = btrfsic_block_lookup_or_add(
1504                                         state,
1505                                         &next_block_ctx,
1506                                         "referenced ",
1507                                         0,
1508                                         force_iodone_flag,
1509                                         !force_iodone_flag,
1510                                         mirror_num,
1511                                         &block_was_created);
1512                         if (NULL == next_block) {
1513                                 printk(KERN_INFO
1514                                        "btrfsic: error, kmalloc failed!\n");
1515                                 btrfsic_release_block_ctx(&next_block_ctx);
1516                                 return -1;
1517                         }
1518                         if (!block_was_created) {
1519                                 if (next_block->logical_bytenr != next_bytenr &&
1520                                     !(!next_block->is_metadata &&
1521                                       0 == next_block->logical_bytenr)) {
1522                                         printk(KERN_INFO
1523                                                "Referenced block"
1524                                                " @%llu (%s/%llu/%d)"
1525                                                " found in hash table, D,"
1526                                                " bytenr mismatch"
1527                                                " (!= stored %llu).\n",
1528                                                next_bytenr,
1529                                                next_block_ctx.dev->name,
1530                                                next_block_ctx.dev_bytenr,
1531                                                mirror_num,
1532                                                next_block->logical_bytenr);
1533                                 }
1534                                 next_block->logical_bytenr = next_bytenr;
1535                                 next_block->mirror_num = mirror_num;
1536                         }
1537
1538                         l = btrfsic_block_link_lookup_or_add(state,
1539                                                              &next_block_ctx,
1540                                                              next_block, block,
1541                                                              generation);
1542                         btrfsic_release_block_ctx(&next_block_ctx);
1543                         if (NULL == l)
1544                                 return -1;
1545                 }
1546
1547                 next_bytenr += chunk_len;
1548                 num_bytes -= chunk_len;
1549         }
1550
1551         return 0;
1552 }
1553
1554 static int btrfsic_map_block(struct btrfsic_state *state, u64 bytenr, u32 len,
1555                              struct btrfsic_block_data_ctx *block_ctx_out,
1556                              int mirror_num)
1557 {
1558         int ret;
1559         u64 length;
1560         struct btrfs_bio *multi = NULL;
1561         struct btrfs_device *device;
1562
1563         length = len;
1564         ret = btrfs_map_block(state->root->fs_info, READ,
1565                               bytenr, &length, &multi, mirror_num);
1566
1567         if (ret) {
1568                 block_ctx_out->start = 0;
1569                 block_ctx_out->dev_bytenr = 0;
1570                 block_ctx_out->len = 0;
1571                 block_ctx_out->dev = NULL;
1572                 block_ctx_out->datav = NULL;
1573                 block_ctx_out->pagev = NULL;
1574                 block_ctx_out->mem_to_free = NULL;
1575
1576                 return ret;
1577         }
1578
1579         device = multi->stripes[0].dev;
1580         block_ctx_out->dev = btrfsic_dev_state_lookup(device->bdev);
1581         block_ctx_out->dev_bytenr = multi->stripes[0].physical;
1582         block_ctx_out->start = bytenr;
1583         block_ctx_out->len = len;
1584         block_ctx_out->datav = NULL;
1585         block_ctx_out->pagev = NULL;
1586         block_ctx_out->mem_to_free = NULL;
1587
1588         kfree(multi);
1589         if (NULL == block_ctx_out->dev) {
1590                 ret = -ENXIO;
1591                 printk(KERN_INFO "btrfsic: error, cannot lookup dev (#1)!\n");
1592         }
1593
1594         return ret;
1595 }
1596
1597 static int btrfsic_map_superblock(struct btrfsic_state *state, u64 bytenr,
1598                                   u32 len, struct block_device *bdev,
1599                                   struct btrfsic_block_data_ctx *block_ctx_out)
1600 {
1601         block_ctx_out->dev = btrfsic_dev_state_lookup(bdev);
1602         block_ctx_out->dev_bytenr = bytenr;
1603         block_ctx_out->start = bytenr;
1604         block_ctx_out->len = len;
1605         block_ctx_out->datav = NULL;
1606         block_ctx_out->pagev = NULL;
1607         block_ctx_out->mem_to_free = NULL;
1608         if (NULL != block_ctx_out->dev) {
1609                 return 0;
1610         } else {
1611                 printk(KERN_INFO "btrfsic: error, cannot lookup dev (#2)!\n");
1612                 return -ENXIO;
1613         }
1614 }
1615
1616 static void btrfsic_release_block_ctx(struct btrfsic_block_data_ctx *block_ctx)
1617 {
1618         if (block_ctx->mem_to_free) {
1619                 unsigned int num_pages;
1620
1621                 BUG_ON(!block_ctx->datav);
1622                 BUG_ON(!block_ctx->pagev);
1623                 num_pages = (block_ctx->len + (u64)PAGE_CACHE_SIZE - 1) >>
1624                             PAGE_CACHE_SHIFT;
1625                 while (num_pages > 0) {
1626                         num_pages--;
1627                         if (block_ctx->datav[num_pages]) {
1628                                 kunmap(block_ctx->pagev[num_pages]);
1629                                 block_ctx->datav[num_pages] = NULL;
1630                         }
1631                         if (block_ctx->pagev[num_pages]) {
1632                                 __free_page(block_ctx->pagev[num_pages]);
1633                                 block_ctx->pagev[num_pages] = NULL;
1634                         }
1635                 }
1636
1637                 kfree(block_ctx->mem_to_free);
1638                 block_ctx->mem_to_free = NULL;
1639                 block_ctx->pagev = NULL;
1640                 block_ctx->datav = NULL;
1641         }
1642 }
1643
1644 static int btrfsic_read_block(struct btrfsic_state *state,
1645                               struct btrfsic_block_data_ctx *block_ctx)
1646 {
1647         unsigned int num_pages;
1648         unsigned int i;
1649         u64 dev_bytenr;
1650         int ret;
1651
1652         BUG_ON(block_ctx->datav);
1653         BUG_ON(block_ctx->pagev);
1654         BUG_ON(block_ctx->mem_to_free);
1655         if (block_ctx->dev_bytenr & ((u64)PAGE_CACHE_SIZE - 1)) {
1656                 printk(KERN_INFO
1657                        "btrfsic: read_block() with unaligned bytenr %llu\n",
1658                        block_ctx->dev_bytenr);
1659                 return -1;
1660         }
1661
1662         num_pages = (block_ctx->len + (u64)PAGE_CACHE_SIZE - 1) >>
1663                     PAGE_CACHE_SHIFT;
1664         block_ctx->mem_to_free = kzalloc((sizeof(*block_ctx->datav) +
1665                                           sizeof(*block_ctx->pagev)) *
1666                                          num_pages, GFP_NOFS);
1667         if (!block_ctx->mem_to_free)
1668                 return -1;
1669         block_ctx->datav = block_ctx->mem_to_free;
1670         block_ctx->pagev = (struct page **)(block_ctx->datav + num_pages);
1671         for (i = 0; i < num_pages; i++) {
1672                 block_ctx->pagev[i] = alloc_page(GFP_NOFS);
1673                 if (!block_ctx->pagev[i])
1674                         return -1;
1675         }
1676
1677         dev_bytenr = block_ctx->dev_bytenr;
1678         for (i = 0; i < num_pages;) {
1679                 struct bio *bio;
1680                 unsigned int j;
1681                 DECLARE_COMPLETION_ONSTACK(complete);
1682
1683                 bio = btrfs_io_bio_alloc(GFP_NOFS, num_pages - i);
1684                 if (!bio) {
1685                         printk(KERN_INFO
1686                                "btrfsic: bio_alloc() for %u pages failed!\n",
1687                                num_pages - i);
1688                         return -1;
1689                 }
1690                 bio->bi_bdev = block_ctx->dev->bdev;
1691                 bio->bi_sector = dev_bytenr >> 9;
1692                 bio->bi_end_io = btrfsic_complete_bio_end_io;
1693                 bio->bi_private = &complete;
1694
1695                 for (j = i; j < num_pages; j++) {
1696                         ret = bio_add_page(bio, block_ctx->pagev[j],
1697                                            PAGE_CACHE_SIZE, 0);
1698                         if (PAGE_CACHE_SIZE != ret)
1699                                 break;
1700                 }
1701                 if (j == i) {
1702                         printk(KERN_INFO
1703                                "btrfsic: error, failed to add a single page!\n");
1704                         return -1;
1705                 }
1706                 submit_bio(READ, bio);
1707
1708                 /* this will also unplug the queue */
1709                 wait_for_completion(&complete);
1710
1711                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) {
1712                         printk(KERN_INFO
1713                                "btrfsic: read error at logical %llu dev %s!\n",
1714                                block_ctx->start, block_ctx->dev->name);
1715                         bio_put(bio);
1716                         return -1;
1717                 }
1718                 bio_put(bio);
1719                 dev_bytenr += (j - i) * PAGE_CACHE_SIZE;
1720                 i = j;
1721         }
1722         for (i = 0; i < num_pages; i++) {
1723                 block_ctx->datav[i] = kmap(block_ctx->pagev[i]);
1724                 if (!block_ctx->datav[i]) {
1725                         printk(KERN_INFO "btrfsic: kmap() failed (dev %s)!\n",
1726                                block_ctx->dev->name);
1727                         return -1;
1728                 }
1729         }
1730
1731         return block_ctx->len;
1732 }
1733
1734 static void btrfsic_complete_bio_end_io(struct bio *bio, int err)
1735 {
1736         complete((struct completion *)bio->bi_private);
1737 }
1738
1739 static void btrfsic_dump_database(struct btrfsic_state *state)
1740 {
1741         struct list_head *elem_all;
1742
1743         BUG_ON(NULL == state);
1744
1745         printk(KERN_INFO "all_blocks_list:\n");
1746         list_for_each(elem_all, &state->all_blocks_list) {
1747                 const struct btrfsic_block *const b_all =
1748                     list_entry(elem_all, struct btrfsic_block,
1749                                all_blocks_node);
1750                 struct list_head *elem_ref_to;
1751                 struct list_head *elem_ref_from;
1752
1753                 printk(KERN_INFO "%c-block @%llu (%s/%llu/%d)\n",
1754                        btrfsic_get_block_type(state, b_all),
1755                        b_all->logical_bytenr, b_all->dev_state->name,
1756                        b_all->dev_bytenr, b_all->mirror_num);
1757
1758                 list_for_each(elem_ref_to, &b_all->ref_to_list) {
1759                         const struct btrfsic_block_link *const l =
1760                             list_entry(elem_ref_to,
1761                                        struct btrfsic_block_link,
1762                                        node_ref_to);
1763
1764                         printk(KERN_INFO " %c @%llu (%s/%llu/%d)"
1765                                " refers %u* to"
1766                                " %c @%llu (%s/%llu/%d)\n",
1767                                btrfsic_get_block_type(state, b_all),
1768                                b_all->logical_bytenr, b_all->dev_state->name,
1769                                b_all->dev_bytenr, b_all->mirror_num,
1770                                l->ref_cnt,
1771                                btrfsic_get_block_type(state, l->block_ref_to),
1772                                l->block_ref_to->logical_bytenr,
1773                                l->block_ref_to->dev_state->name,
1774                                l->block_ref_to->dev_bytenr,
1775                                l->block_ref_to->mirror_num);
1776                 }
1777
1778                 list_for_each(elem_ref_from, &b_all->ref_from_list) {
1779                         const struct btrfsic_block_link *const l =
1780                             list_entry(elem_ref_from,
1781                                        struct btrfsic_block_link,
1782                                        node_ref_from);
1783
1784                         printk(KERN_INFO " %c @%llu (%s/%llu/%d)"
1785                                " is ref %u* from"
1786                                " %c @%llu (%s/%llu/%d)\n",
1787                                btrfsic_get_block_type(state, b_all),
1788                                b_all->logical_bytenr, b_all->dev_state->name,
1789                                b_all->dev_bytenr, b_all->mirror_num,
1790                                l->ref_cnt,
1791                                btrfsic_get_block_type(state, l->block_ref_from),
1792                                l->block_ref_from->logical_bytenr,
1793                                l->block_ref_from->dev_state->name,
1794                                l->block_ref_from->dev_bytenr,
1795                                l->block_ref_from->mirror_num);
1796                 }
1797
1798                 printk(KERN_INFO "\n");
1799         }
1800 }
1801
1802 /*
1803  * Test whether the disk block contains a tree block (leaf or node)
1804  * (note that this test fails for the super block)
1805  */
1806 static int btrfsic_test_for_metadata(struct btrfsic_state *state,
1807                                      char **datav, unsigned int num_pages)
1808 {
1809         struct btrfs_header *h;
1810         u8 csum[BTRFS_CSUM_SIZE];
1811         u32 crc = ~(u32)0;
1812         unsigned int i;
1813
1814         if (num_pages * PAGE_CACHE_SIZE < state->metablock_size)
1815                 return 1; /* not metadata */
1816         num_pages = state->metablock_size >> PAGE_CACHE_SHIFT;
1817         h = (struct btrfs_header *)datav[0];
1818
1819         if (memcmp(h->fsid, state->root->fs_info->fsid, BTRFS_UUID_SIZE))
1820                 return 1;
1821
1822         for (i = 0; i < num_pages; i++) {
1823                 u8 *data = i ? datav[i] : (datav[i] + BTRFS_CSUM_SIZE);
1824                 size_t sublen = i ? PAGE_CACHE_SIZE :
1825                                     (PAGE_CACHE_SIZE - BTRFS_CSUM_SIZE);
1826
1827                 crc = crc32c(crc, data, sublen);
1828         }
1829         btrfs_csum_final(crc, csum);
1830         if (memcmp(csum, h->csum, state->csum_size))
1831                 return 1;
1832
1833         return 0; /* is metadata */
1834 }
1835
1836 static void btrfsic_process_written_block(struct btrfsic_dev_state *dev_state,
1837                                           u64 dev_bytenr, char **mapped_datav,
1838                                           unsigned int num_pages,
1839                                           struct bio *bio, int *bio_is_patched,
1840                                           struct buffer_head *bh,
1841                                           int submit_bio_bh_rw)
1842 {
1843         int is_metadata;
1844         struct btrfsic_block *block;
1845         struct btrfsic_block_data_ctx block_ctx;
1846         int ret;
1847         struct btrfsic_state *state = dev_state->state;
1848         struct block_device *bdev = dev_state->bdev;
1849         unsigned int processed_len;
1850
1851         if (NULL != bio_is_patched)
1852                 *bio_is_patched = 0;
1853
1854 again:
1855         if (num_pages == 0)
1856                 return;
1857
1858         processed_len = 0;
1859         is_metadata = (0 == btrfsic_test_for_metadata(state, mapped_datav,
1860                                                       num_pages));
1861
1862         block = btrfsic_block_hashtable_lookup(bdev, dev_bytenr,
1863                                                &state->block_hashtable);
1864         if (NULL != block) {
1865                 u64 bytenr = 0;
1866                 struct list_head *elem_ref_to;
1867                 struct list_head *tmp_ref_to;
1868
1869                 if (block->is_superblock) {
1870                         bytenr = btrfs_super_bytenr((struct btrfs_super_block *)
1871                                                     mapped_datav[0]);
1872                         if (num_pages * PAGE_CACHE_SIZE <
1873                             BTRFS_SUPER_INFO_SIZE) {
1874                                 printk(KERN_INFO
1875                                        "btrfsic: cannot work with too short bios!\n");
1876                                 return;
1877                         }
1878                         is_metadata = 1;
1879                         BUG_ON(BTRFS_SUPER_INFO_SIZE & (PAGE_CACHE_SIZE - 1));
1880                         processed_len = BTRFS_SUPER_INFO_SIZE;
1881                         if (state->print_mask &
1882                             BTRFSIC_PRINT_MASK_TREE_BEFORE_SB_WRITE) {
1883                                 printk(KERN_INFO
1884                                        "[before new superblock is written]:\n");
1885                                 btrfsic_dump_tree_sub(state, block, 0);
1886                         }
1887                 }
1888                 if (is_metadata) {
1889                         if (!block->is_superblock) {
1890                                 if (num_pages * PAGE_CACHE_SIZE <
1891                                     state->metablock_size) {
1892                                         printk(KERN_INFO
1893                                                "btrfsic: cannot work with too short bios!\n");
1894                                         return;
1895                                 }
1896                                 processed_len = state->metablock_size;
1897                                 bytenr = btrfs_stack_header_bytenr(
1898                                                 (struct btrfs_header *)
1899                                                 mapped_datav[0]);
1900                                 btrfsic_cmp_log_and_dev_bytenr(state, bytenr,
1901                                                                dev_state,
1902                                                                dev_bytenr);
1903                         }
1904                         if (block->logical_bytenr != bytenr) {
1905                                 printk(KERN_INFO
1906                                        "Written block @%llu (%s/%llu/%d)"
1907                                        " found in hash table, %c,"
1908                                        " bytenr mismatch"
1909                                        " (!= stored %llu).\n",
1910                                        bytenr, dev_state->name, dev_bytenr,
1911                                        block->mirror_num,
1912                                        btrfsic_get_block_type(state, block),
1913                                        block->logical_bytenr);
1914                                 block->logical_bytenr = bytenr;
1915                         } else if (state->print_mask &
1916                                    BTRFSIC_PRINT_MASK_VERBOSE)
1917                                 printk(KERN_INFO
1918                                        "Written block @%llu (%s/%llu/%d)"
1919                                        " found in hash table, %c.\n",
1920                                        bytenr, dev_state->name, dev_bytenr,
1921                                        block->mirror_num,
1922                                        btrfsic_get_block_type(state, block));
1923                 } else {
1924                         if (num_pages * PAGE_CACHE_SIZE <
1925                             state->datablock_size) {
1926                                 printk(KERN_INFO
1927                                        "btrfsic: cannot work with too short bios!\n");
1928                                 return;
1929                         }
1930                         processed_len = state->datablock_size;
1931                         bytenr = block->logical_bytenr;
1932                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1933                                 printk(KERN_INFO
1934                                        "Written block @%llu (%s/%llu/%d)"
1935                                        " found in hash table, %c.\n",
1936                                        bytenr, dev_state->name, dev_bytenr,
1937                                        block->mirror_num,
1938                                        btrfsic_get_block_type(state, block));
1939                 }
1940
1941                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1942                         printk(KERN_INFO
1943                                "ref_to_list: %cE, ref_from_list: %cE\n",
1944                                list_empty(&block->ref_to_list) ? ' ' : '!',
1945                                list_empty(&block->ref_from_list) ? ' ' : '!');
1946                 if (btrfsic_is_block_ref_by_superblock(state, block, 0)) {
1947                         printk(KERN_INFO "btrfs: attempt to overwrite %c-block"
1948                                " @%llu (%s/%llu/%d), old(gen=%llu,"
1949                                " objectid=%llu, type=%d, offset=%llu),"
1950                                " new(gen=%llu),"
1951                                " which is referenced by most recent superblock"
1952                                " (superblockgen=%llu)!\n",
1953                                btrfsic_get_block_type(state, block), bytenr,
1954                                dev_state->name, dev_bytenr, block->mirror_num,
1955                                block->generation,
1956                                btrfs_disk_key_objectid(&block->disk_key),
1957                                block->disk_key.type,
1958                                btrfs_disk_key_offset(&block->disk_key),
1959                                btrfs_stack_header_generation(
1960                                        (struct btrfs_header *) mapped_datav[0]),
1961                                state->max_superblock_generation);
1962                         btrfsic_dump_tree(state);
1963                 }
1964
1965                 if (!block->is_iodone && !block->never_written) {
1966                         printk(KERN_INFO "btrfs: attempt to overwrite %c-block"
1967                                " @%llu (%s/%llu/%d), oldgen=%llu, newgen=%llu,"
1968                                " which is not yet iodone!\n",
1969                                btrfsic_get_block_type(state, block), bytenr,
1970                                dev_state->name, dev_bytenr, block->mirror_num,
1971                                block->generation,
1972                                btrfs_stack_header_generation(
1973                                        (struct btrfs_header *)
1974                                        mapped_datav[0]));
1975                         /* it would not be safe to go on */
1976                         btrfsic_dump_tree(state);
1977                         goto continue_loop;
1978                 }
1979
1980                 /*
1981                  * Clear all references of this block. Do not free
1982                  * the block itself even if is not referenced anymore
1983                  * because it still carries valueable information
1984                  * like whether it was ever written and IO completed.
1985                  */
1986                 list_for_each_safe(elem_ref_to, tmp_ref_to,
1987                                    &block->ref_to_list) {
1988                         struct btrfsic_block_link *const l =
1989                             list_entry(elem_ref_to,
1990                                        struct btrfsic_block_link,
1991                                        node_ref_to);
1992
1993                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1994                                 btrfsic_print_rem_link(state, l);
1995                         l->ref_cnt--;
1996                         if (0 == l->ref_cnt) {
1997                                 list_del(&l->node_ref_to);
1998                                 list_del(&l->node_ref_from);
1999                                 btrfsic_block_link_hashtable_remove(l);
2000                                 btrfsic_block_link_free(l);
2001                         }
2002                 }
2003
2004                 if (block->is_superblock)
2005                         ret = btrfsic_map_superblock(state, bytenr,
2006                                                      processed_len,
2007                                                      bdev, &block_ctx);
2008                 else
2009                         ret = btrfsic_map_block(state, bytenr, processed_len,
2010                                                 &block_ctx, 0);
2011                 if (ret) {
2012                         printk(KERN_INFO
2013                                "btrfsic: btrfsic_map_block(root @%llu)"
2014                                " failed!\n", bytenr);
2015                         goto continue_loop;
2016                 }
2017                 block_ctx.datav = mapped_datav;
2018                 /* the following is required in case of writes to mirrors,
2019                  * use the same that was used for the lookup */
2020                 block_ctx.dev = dev_state;
2021                 block_ctx.dev_bytenr = dev_bytenr;
2022
2023                 if (is_metadata || state->include_extent_data) {
2024                         block->never_written = 0;
2025                         block->iodone_w_error = 0;
2026                         if (NULL != bio) {
2027                                 block->is_iodone = 0;
2028                                 BUG_ON(NULL == bio_is_patched);
2029                                 if (!*bio_is_patched) {
2030                                         block->orig_bio_bh_private =
2031                                             bio->bi_private;
2032                                         block->orig_bio_bh_end_io.bio =
2033                                             bio->bi_end_io;
2034                                         block->next_in_same_bio = NULL;
2035                                         bio->bi_private = block;
2036                                         bio->bi_end_io = btrfsic_bio_end_io;
2037                                         *bio_is_patched = 1;
2038                                 } else {
2039                                         struct btrfsic_block *chained_block =
2040                                             (struct btrfsic_block *)
2041                                             bio->bi_private;
2042
2043                                         BUG_ON(NULL == chained_block);
2044                                         block->orig_bio_bh_private =
2045                                             chained_block->orig_bio_bh_private;
2046                                         block->orig_bio_bh_end_io.bio =
2047                                             chained_block->orig_bio_bh_end_io.
2048                                             bio;
2049                                         block->next_in_same_bio = chained_block;
2050                                         bio->bi_private = block;
2051                                 }
2052                         } else if (NULL != bh) {
2053                                 block->is_iodone = 0;
2054                                 block->orig_bio_bh_private = bh->b_private;
2055                                 block->orig_bio_bh_end_io.bh = bh->b_end_io;
2056                                 block->next_in_same_bio = NULL;
2057                                 bh->b_private = block;
2058                                 bh->b_end_io = btrfsic_bh_end_io;
2059                         } else {
2060                                 block->is_iodone = 1;
2061                                 block->orig_bio_bh_private = NULL;
2062                                 block->orig_bio_bh_end_io.bio = NULL;
2063                                 block->next_in_same_bio = NULL;
2064                         }
2065                 }
2066
2067                 block->flush_gen = dev_state->last_flush_gen + 1;
2068                 block->submit_bio_bh_rw = submit_bio_bh_rw;
2069                 if (is_metadata) {
2070                         block->logical_bytenr = bytenr;
2071                         block->is_metadata = 1;
2072                         if (block->is_superblock) {
2073                                 BUG_ON(PAGE_CACHE_SIZE !=
2074                                        BTRFS_SUPER_INFO_SIZE);
2075                                 ret = btrfsic_process_written_superblock(
2076                                                 state,
2077                                                 block,
2078                                                 (struct btrfs_super_block *)
2079                                                 mapped_datav[0]);
2080                                 if (state->print_mask &
2081                                     BTRFSIC_PRINT_MASK_TREE_AFTER_SB_WRITE) {
2082                                         printk(KERN_INFO
2083                                         "[after new superblock is written]:\n");
2084                                         btrfsic_dump_tree_sub(state, block, 0);
2085                                 }
2086                         } else {
2087                                 block->mirror_num = 0;  /* unknown */
2088                                 ret = btrfsic_process_metablock(
2089                                                 state,
2090                                                 block,
2091                                                 &block_ctx,
2092                                                 0, 0);
2093                         }
2094                         if (ret)
2095                                 printk(KERN_INFO
2096                                        "btrfsic: btrfsic_process_metablock"
2097                                        "(root @%llu) failed!\n",
2098                                        dev_bytenr);
2099                 } else {
2100                         block->is_metadata = 0;
2101                         block->mirror_num = 0;  /* unknown */
2102                         block->generation = BTRFSIC_GENERATION_UNKNOWN;
2103                         if (!state->include_extent_data
2104                             && list_empty(&block->ref_from_list)) {
2105                                 /*
2106                                  * disk block is overwritten with extent
2107                                  * data (not meta data) and we are configured
2108                                  * to not include extent data: take the
2109                                  * chance and free the block's memory
2110                                  */
2111                                 btrfsic_block_hashtable_remove(block);
2112                                 list_del(&block->all_blocks_node);
2113                                 btrfsic_block_free(block);
2114                         }
2115                 }
2116                 btrfsic_release_block_ctx(&block_ctx);
2117         } else {
2118                 /* block has not been found in hash table */
2119                 u64 bytenr;
2120
2121                 if (!is_metadata) {
2122                         processed_len = state->datablock_size;
2123                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2124                                 printk(KERN_INFO "Written block (%s/%llu/?)"
2125                                        " !found in hash table, D.\n",
2126                                        dev_state->name, dev_bytenr);
2127                         if (!state->include_extent_data) {
2128                                 /* ignore that written D block */
2129                                 goto continue_loop;
2130                         }
2131
2132                         /* this is getting ugly for the
2133                          * include_extent_data case... */
2134                         bytenr = 0;     /* unknown */
2135                         block_ctx.start = bytenr;
2136                         block_ctx.len = processed_len;
2137                         block_ctx.mem_to_free = NULL;
2138                         block_ctx.pagev = NULL;
2139                 } else {
2140                         processed_len = state->metablock_size;
2141                         bytenr = btrfs_stack_header_bytenr(
2142                                         (struct btrfs_header *)
2143                                         mapped_datav[0]);
2144                         btrfsic_cmp_log_and_dev_bytenr(state, bytenr, dev_state,
2145                                                        dev_bytenr);
2146                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2147                                 printk(KERN_INFO
2148                                        "Written block @%llu (%s/%llu/?)"
2149                                        " !found in hash table, M.\n",
2150                                        bytenr, dev_state->name, dev_bytenr);
2151
2152                         ret = btrfsic_map_block(state, bytenr, processed_len,
2153                                                 &block_ctx, 0);
2154                         if (ret) {
2155                                 printk(KERN_INFO
2156                                        "btrfsic: btrfsic_map_block(root @%llu)"
2157                                        " failed!\n",
2158                                        dev_bytenr);
2159                                 goto continue_loop;
2160                         }
2161                 }
2162                 block_ctx.datav = mapped_datav;
2163                 /* the following is required in case of writes to mirrors,
2164                  * use the same that was used for the lookup */
2165                 block_ctx.dev = dev_state;
2166                 block_ctx.dev_bytenr = dev_bytenr;
2167
2168                 block = btrfsic_block_alloc();
2169                 if (NULL == block) {
2170                         printk(KERN_INFO "btrfsic: error, kmalloc failed!\n");
2171                         btrfsic_release_block_ctx(&block_ctx);
2172                         goto continue_loop;
2173                 }
2174                 block->dev_state = dev_state;
2175                 block->dev_bytenr = dev_bytenr;
2176                 block->logical_bytenr = bytenr;
2177                 block->is_metadata = is_metadata;
2178                 block->never_written = 0;
2179                 block->iodone_w_error = 0;
2180                 block->mirror_num = 0;  /* unknown */
2181                 block->flush_gen = dev_state->last_flush_gen + 1;
2182                 block->submit_bio_bh_rw = submit_bio_bh_rw;
2183                 if (NULL != bio) {
2184                         block->is_iodone = 0;
2185                         BUG_ON(NULL == bio_is_patched);
2186                         if (!*bio_is_patched) {
2187                                 block->orig_bio_bh_private = bio->bi_private;
2188                                 block->orig_bio_bh_end_io.bio = bio->bi_end_io;
2189                                 block->next_in_same_bio = NULL;
2190                                 bio->bi_private = block;
2191                                 bio->bi_end_io = btrfsic_bio_end_io;
2192                                 *bio_is_patched = 1;
2193                         } else {
2194                                 struct btrfsic_block *chained_block =
2195                                     (struct btrfsic_block *)
2196                                     bio->bi_private;
2197
2198                                 BUG_ON(NULL == chained_block);
2199                                 block->orig_bio_bh_private =
2200                                     chained_block->orig_bio_bh_private;
2201                                 block->orig_bio_bh_end_io.bio =
2202                                     chained_block->orig_bio_bh_end_io.bio;
2203                                 block->next_in_same_bio = chained_block;
2204                                 bio->bi_private = block;
2205                         }
2206                 } else if (NULL != bh) {
2207                         block->is_iodone = 0;
2208                         block->orig_bio_bh_private = bh->b_private;
2209                         block->orig_bio_bh_end_io.bh = bh->b_end_io;
2210                         block->next_in_same_bio = NULL;
2211                         bh->b_private = block;
2212                         bh->b_end_io = btrfsic_bh_end_io;
2213                 } else {
2214                         block->is_iodone = 1;
2215                         block->orig_bio_bh_private = NULL;
2216                         block->orig_bio_bh_end_io.bio = NULL;
2217                         block->next_in_same_bio = NULL;
2218                 }
2219                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2220                         printk(KERN_INFO
2221                                "New written %c-block @%llu (%s/%llu/%d)\n",
2222                                is_metadata ? 'M' : 'D',
2223                                block->logical_bytenr, block->dev_state->name,
2224                                block->dev_bytenr, block->mirror_num);
2225                 list_add(&block->all_blocks_node, &state->all_blocks_list);
2226                 btrfsic_block_hashtable_add(block, &state->block_hashtable);
2227
2228                 if (is_metadata) {
2229                         ret = btrfsic_process_metablock(state, block,
2230                                                         &block_ctx, 0, 0);
2231                         if (ret)
2232                                 printk(KERN_INFO
2233                                        "btrfsic: process_metablock(root @%llu)"
2234                                        " failed!\n",
2235                                        dev_bytenr);
2236                 }
2237                 btrfsic_release_block_ctx(&block_ctx);
2238         }
2239
2240 continue_loop:
2241         BUG_ON(!processed_len);
2242         dev_bytenr += processed_len;
2243         mapped_datav += processed_len >> PAGE_CACHE_SHIFT;
2244         num_pages -= processed_len >> PAGE_CACHE_SHIFT;
2245         goto again;
2246 }
2247
2248 static void btrfsic_bio_end_io(struct bio *bp, int bio_error_status)
2249 {
2250         struct btrfsic_block *block = (struct btrfsic_block *)bp->bi_private;
2251         int iodone_w_error;
2252
2253         /* mutex is not held! This is not save if IO is not yet completed
2254          * on umount */
2255         iodone_w_error = 0;
2256         if (bio_error_status)
2257                 iodone_w_error = 1;
2258
2259         BUG_ON(NULL == block);
2260         bp->bi_private = block->orig_bio_bh_private;
2261         bp->bi_end_io = block->orig_bio_bh_end_io.bio;
2262
2263         do {
2264                 struct btrfsic_block *next_block;
2265                 struct btrfsic_dev_state *const dev_state = block->dev_state;
2266
2267                 if ((dev_state->state->print_mask &
2268                      BTRFSIC_PRINT_MASK_END_IO_BIO_BH))
2269                         printk(KERN_INFO
2270                                "bio_end_io(err=%d) for %c @%llu (%s/%llu/%d)\n",
2271                                bio_error_status,
2272                                btrfsic_get_block_type(dev_state->state, block),
2273                                block->logical_bytenr, dev_state->name,
2274                                block->dev_bytenr, block->mirror_num);
2275                 next_block = block->next_in_same_bio;
2276                 block->iodone_w_error = iodone_w_error;
2277                 if (block->submit_bio_bh_rw & REQ_FLUSH) {
2278                         dev_state->last_flush_gen++;
2279                         if ((dev_state->state->print_mask &
2280                              BTRFSIC_PRINT_MASK_END_IO_BIO_BH))
2281                                 printk(KERN_INFO
2282                                        "bio_end_io() new %s flush_gen=%llu\n",
2283                                        dev_state->name,
2284                                        dev_state->last_flush_gen);
2285                 }
2286                 if (block->submit_bio_bh_rw & REQ_FUA)
2287                         block->flush_gen = 0; /* FUA completed means block is
2288                                                * on disk */
2289                 block->is_iodone = 1; /* for FLUSH, this releases the block */
2290                 block = next_block;
2291         } while (NULL != block);
2292
2293         bp->bi_end_io(bp, bio_error_status);
2294 }
2295
2296 static void btrfsic_bh_end_io(struct buffer_head *bh, int uptodate)
2297 {
2298         struct btrfsic_block *block = (struct btrfsic_block *)bh->b_private;
2299         int iodone_w_error = !uptodate;
2300         struct btrfsic_dev_state *dev_state;
2301
2302         BUG_ON(NULL == block);
2303         dev_state = block->dev_state;
2304         if ((dev_state->state->print_mask & BTRFSIC_PRINT_MASK_END_IO_BIO_BH))
2305                 printk(KERN_INFO
2306                        "bh_end_io(error=%d) for %c @%llu (%s/%llu/%d)\n",
2307                        iodone_w_error,
2308                        btrfsic_get_block_type(dev_state->state, block),
2309                        block->logical_bytenr, block->dev_state->name,
2310                        block->dev_bytenr, block->mirror_num);
2311
2312         block->iodone_w_error = iodone_w_error;
2313         if (block->submit_bio_bh_rw & REQ_FLUSH) {
2314                 dev_state->last_flush_gen++;
2315                 if ((dev_state->state->print_mask &
2316                      BTRFSIC_PRINT_MASK_END_IO_BIO_BH))
2317                         printk(KERN_INFO
2318                                "bh_end_io() new %s flush_gen=%llu\n",
2319                                dev_state->name, dev_state->last_flush_gen);
2320         }
2321         if (block->submit_bio_bh_rw & REQ_FUA)
2322                 block->flush_gen = 0; /* FUA completed means block is on disk */
2323
2324         bh->b_private = block->orig_bio_bh_private;
2325         bh->b_end_io = block->orig_bio_bh_end_io.bh;
2326         block->is_iodone = 1; /* for FLUSH, this releases the block */
2327         bh->b_end_io(bh, uptodate);
2328 }
2329
2330 static int btrfsic_process_written_superblock(
2331                 struct btrfsic_state *state,
2332                 struct btrfsic_block *const superblock,
2333                 struct btrfs_super_block *const super_hdr)
2334 {
2335         int pass;
2336
2337         superblock->generation = btrfs_super_generation(super_hdr);
2338         if (!(superblock->generation > state->max_superblock_generation ||
2339               0 == state->max_superblock_generation)) {
2340                 if (state->print_mask & BTRFSIC_PRINT_MASK_SUPERBLOCK_WRITE)
2341                         printk(KERN_INFO
2342                                "btrfsic: superblock @%llu (%s/%llu/%d)"
2343                                " with old gen %llu <= %llu\n",
2344                                superblock->logical_bytenr,
2345                                superblock->dev_state->name,
2346                                superblock->dev_bytenr, superblock->mirror_num,
2347                                btrfs_super_generation(super_hdr),
2348                                state->max_superblock_generation);
2349         } else {
2350                 if (state->print_mask & BTRFSIC_PRINT_MASK_SUPERBLOCK_WRITE)
2351                         printk(KERN_INFO
2352                                "btrfsic: got new superblock @%llu (%s/%llu/%d)"
2353                                " with new gen %llu > %llu\n",
2354                                superblock->logical_bytenr,
2355                                superblock->dev_state->name,
2356                                superblock->dev_bytenr, superblock->mirror_num,
2357                                btrfs_super_generation(super_hdr),
2358                                state->max_superblock_generation);
2359
2360                 state->max_superblock_generation =
2361                     btrfs_super_generation(super_hdr);
2362                 state->latest_superblock = superblock;
2363         }
2364
2365         for (pass = 0; pass < 3; pass++) {
2366                 int ret;
2367                 u64 next_bytenr;
2368                 struct btrfsic_block *next_block;
2369                 struct btrfsic_block_data_ctx tmp_next_block_ctx;
2370                 struct btrfsic_block_link *l;
2371                 int num_copies;
2372                 int mirror_num;
2373                 const char *additional_string = NULL;
2374                 struct btrfs_disk_key tmp_disk_key = {0};
2375
2376                 btrfs_set_disk_key_objectid(&tmp_disk_key,
2377                                             BTRFS_ROOT_ITEM_KEY);
2378                 btrfs_set_disk_key_objectid(&tmp_disk_key, 0);
2379
2380                 switch (pass) {
2381                 case 0:
2382                         btrfs_set_disk_key_objectid(&tmp_disk_key,
2383                                                     BTRFS_ROOT_TREE_OBJECTID);
2384                         additional_string = "root ";
2385                         next_bytenr = btrfs_super_root(super_hdr);
2386                         if (state->print_mask &
2387                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
2388                                 printk(KERN_INFO "root@%llu\n", next_bytenr);
2389                         break;
2390                 case 1:
2391                         btrfs_set_disk_key_objectid(&tmp_disk_key,
2392                                                     BTRFS_CHUNK_TREE_OBJECTID);
2393                         additional_string = "chunk ";
2394                         next_bytenr = btrfs_super_chunk_root(super_hdr);
2395                         if (state->print_mask &
2396                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
2397                                 printk(KERN_INFO "chunk@%llu\n", next_bytenr);
2398                         break;
2399                 case 2:
2400                         btrfs_set_disk_key_objectid(&tmp_disk_key,
2401                                                     BTRFS_TREE_LOG_OBJECTID);
2402                         additional_string = "log ";
2403                         next_bytenr = btrfs_super_log_root(super_hdr);
2404                         if (0 == next_bytenr)
2405                                 continue;
2406                         if (state->print_mask &
2407                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
2408                                 printk(KERN_INFO "log@%llu\n", next_bytenr);
2409                         break;
2410                 }
2411
2412                 num_copies =
2413                     btrfs_num_copies(state->root->fs_info,
2414                                      next_bytenr, BTRFS_SUPER_INFO_SIZE);
2415                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
2416                         printk(KERN_INFO "num_copies(log_bytenr=%llu) = %d\n",
2417                                next_bytenr, num_copies);
2418                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
2419                         int was_created;
2420
2421                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2422                                 printk(KERN_INFO
2423                                        "btrfsic_process_written_superblock("
2424                                        "mirror_num=%d)\n", mirror_num);
2425                         ret = btrfsic_map_block(state, next_bytenr,
2426                                                 BTRFS_SUPER_INFO_SIZE,
2427                                                 &tmp_next_block_ctx,
2428                                                 mirror_num);
2429                         if (ret) {
2430                                 printk(KERN_INFO
2431                                        "btrfsic: btrfsic_map_block(@%llu,"
2432                                        " mirror=%d) failed!\n",
2433                                        next_bytenr, mirror_num);
2434                                 return -1;
2435                         }
2436
2437                         next_block = btrfsic_block_lookup_or_add(
2438                                         state,
2439                                         &tmp_next_block_ctx,
2440                                         additional_string,
2441                                         1, 0, 1,
2442                                         mirror_num,
2443                                         &was_created);
2444                         if (NULL == next_block) {
2445                                 printk(KERN_INFO
2446                                        "btrfsic: error, kmalloc failed!\n");
2447                                 btrfsic_release_block_ctx(&tmp_next_block_ctx);
2448                                 return -1;
2449                         }
2450
2451                         next_block->disk_key = tmp_disk_key;
2452                         if (was_created)
2453                                 next_block->generation =
2454                                     BTRFSIC_GENERATION_UNKNOWN;
2455                         l = btrfsic_block_link_lookup_or_add(
2456                                         state,
2457                                         &tmp_next_block_ctx,
2458                                         next_block,
2459                                         superblock,
2460                                         BTRFSIC_GENERATION_UNKNOWN);
2461                         btrfsic_release_block_ctx(&tmp_next_block_ctx);
2462                         if (NULL == l)
2463                                 return -1;
2464                 }
2465         }
2466
2467         if (-1 == btrfsic_check_all_ref_blocks(state, superblock, 0)) {
2468                 WARN_ON(1);
2469                 btrfsic_dump_tree(state);
2470         }
2471
2472         return 0;
2473 }
2474
2475 static int btrfsic_check_all_ref_blocks(struct btrfsic_state *state,
2476                                         struct btrfsic_block *const block,
2477                                         int recursion_level)
2478 {
2479         struct list_head *elem_ref_to;
2480         int ret = 0;
2481
2482         if (recursion_level >= 3 + BTRFS_MAX_LEVEL) {
2483                 /*
2484                  * Note that this situation can happen and does not
2485                  * indicate an error in regular cases. It happens
2486                  * when disk blocks are freed and later reused.
2487                  * The check-integrity module is not aware of any
2488                  * block free operations, it just recognizes block
2489                  * write operations. Therefore it keeps the linkage
2490                  * information for a block until a block is
2491                  * rewritten. This can temporarily cause incorrect
2492                  * and even circular linkage informations. This
2493                  * causes no harm unless such blocks are referenced
2494                  * by the most recent super block.
2495                  */
2496                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2497                         printk(KERN_INFO
2498                                "btrfsic: abort cyclic linkage (case 1).\n");
2499
2500                 return ret;
2501         }
2502
2503         /*
2504          * This algorithm is recursive because the amount of used stack
2505          * space is very small and the max recursion depth is limited.
2506          */
2507         list_for_each(elem_ref_to, &block->ref_to_list) {
2508                 const struct btrfsic_block_link *const l =
2509                     list_entry(elem_ref_to, struct btrfsic_block_link,
2510                                node_ref_to);
2511
2512                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2513                         printk(KERN_INFO
2514                                "rl=%d, %c @%llu (%s/%llu/%d)"
2515                                " %u* refers to %c @%llu (%s/%llu/%d)\n",
2516                                recursion_level,
2517                                btrfsic_get_block_type(state, block),
2518                                block->logical_bytenr, block->dev_state->name,
2519                                block->dev_bytenr, block->mirror_num,
2520                                l->ref_cnt,
2521                                btrfsic_get_block_type(state, l->block_ref_to),
2522                                l->block_ref_to->logical_bytenr,
2523                                l->block_ref_to->dev_state->name,
2524                                l->block_ref_to->dev_bytenr,
2525                                l->block_ref_to->mirror_num);
2526                 if (l->block_ref_to->never_written) {
2527                         printk(KERN_INFO "btrfs: attempt to write superblock"
2528                                " which references block %c @%llu (%s/%llu/%d)"
2529                                " which is never written!\n",
2530                                btrfsic_get_block_type(state, l->block_ref_to),
2531                                l->block_ref_to->logical_bytenr,
2532                                l->block_ref_to->dev_state->name,
2533                                l->block_ref_to->dev_bytenr,
2534                                l->block_ref_to->mirror_num);
2535                         ret = -1;
2536                 } else if (!l->block_ref_to->is_iodone) {
2537                         printk(KERN_INFO "btrfs: attempt to write superblock"
2538                                " which references block %c @%llu (%s/%llu/%d)"
2539                                " which is not yet iodone!\n",
2540                                btrfsic_get_block_type(state, l->block_ref_to),
2541                                l->block_ref_to->logical_bytenr,
2542                                l->block_ref_to->dev_state->name,
2543                                l->block_ref_to->dev_bytenr,
2544                                l->block_ref_to->mirror_num);
2545                         ret = -1;
2546                 } else if (l->block_ref_to->iodone_w_error) {
2547                         printk(KERN_INFO "btrfs: attempt to write superblock"
2548                                " which references block %c @%llu (%s/%llu/%d)"
2549                                " which has write error!\n",
2550                                btrfsic_get_block_type(state, l->block_ref_to),
2551                                l->block_ref_to->logical_bytenr,
2552                                l->block_ref_to->dev_state->name,
2553                                l->block_ref_to->dev_bytenr,
2554                                l->block_ref_to->mirror_num);
2555                         ret = -1;
2556                 } else if (l->parent_generation !=
2557                            l->block_ref_to->generation &&
2558                            BTRFSIC_GENERATION_UNKNOWN !=
2559                            l->parent_generation &&
2560                            BTRFSIC_GENERATION_UNKNOWN !=
2561                            l->block_ref_to->generation) {
2562                         printk(KERN_INFO "btrfs: attempt to write superblock"
2563                                " which references block %c @%llu (%s/%llu/%d)"
2564                                " with generation %llu !="
2565                                " parent generation %llu!\n",
2566                                btrfsic_get_block_type(state, l->block_ref_to),
2567                                l->block_ref_to->logical_bytenr,
2568                                l->block_ref_to->dev_state->name,
2569                                l->block_ref_to->dev_bytenr,
2570                                l->block_ref_to->mirror_num,
2571                                l->block_ref_to->generation,
2572                                l->parent_generation);
2573                         ret = -1;
2574                 } else if (l->block_ref_to->flush_gen >
2575                            l->block_ref_to->dev_state->last_flush_gen) {
2576                         printk(KERN_INFO "btrfs: attempt to write superblock"
2577                                " which references block %c @%llu (%s/%llu/%d)"
2578                                " which is not flushed out of disk's write cache"
2579                                " (block flush_gen=%llu,"
2580                                " dev->flush_gen=%llu)!\n",
2581                                btrfsic_get_block_type(state, l->block_ref_to),
2582                                l->block_ref_to->logical_bytenr,
2583                                l->block_ref_to->dev_state->name,
2584                                l->block_ref_to->dev_bytenr,
2585                                l->block_ref_to->mirror_num, block->flush_gen,
2586                                l->block_ref_to->dev_state->last_flush_gen);
2587                         ret = -1;
2588                 } else if (-1 == btrfsic_check_all_ref_blocks(state,
2589                                                               l->block_ref_to,
2590                                                               recursion_level +
2591                                                               1)) {
2592                         ret = -1;
2593                 }
2594         }
2595
2596         return ret;
2597 }
2598
2599 static int btrfsic_is_block_ref_by_superblock(
2600                 const struct btrfsic_state *state,
2601                 const struct btrfsic_block *block,
2602                 int recursion_level)
2603 {
2604         struct list_head *elem_ref_from;
2605
2606         if (recursion_level >= 3 + BTRFS_MAX_LEVEL) {
2607                 /* refer to comment at "abort cyclic linkage (case 1)" */
2608                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2609                         printk(KERN_INFO
2610                                "btrfsic: abort cyclic linkage (case 2).\n");
2611
2612                 return 0;
2613         }
2614
2615         /*
2616          * This algorithm is recursive because the amount of used stack space
2617          * is very small and the max recursion depth is limited.
2618          */
2619         list_for_each(elem_ref_from, &block->ref_from_list) {
2620                 const struct btrfsic_block_link *const l =
2621                     list_entry(elem_ref_from, struct btrfsic_block_link,
2622                                node_ref_from);
2623
2624                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2625                         printk(KERN_INFO
2626                                "rl=%d, %c @%llu (%s/%llu/%d)"
2627                                " is ref %u* from %c @%llu (%s/%llu/%d)\n",
2628                                recursion_level,
2629                                btrfsic_get_block_type(state, block),
2630                                block->logical_bytenr, block->dev_state->name,
2631                                block->dev_bytenr, block->mirror_num,
2632                                l->ref_cnt,
2633                                btrfsic_get_block_type(state, l->block_ref_from),
2634                                l->block_ref_from->logical_bytenr,
2635                                l->block_ref_from->dev_state->name,
2636                                l->block_ref_from->dev_bytenr,
2637                                l->block_ref_from->mirror_num);
2638                 if (l->block_ref_from->is_superblock &&
2639                     state->latest_superblock->dev_bytenr ==
2640                     l->block_ref_from->dev_bytenr &&
2641                     state->latest_superblock->dev_state->bdev ==
2642                     l->block_ref_from->dev_state->bdev)
2643                         return 1;
2644                 else if (btrfsic_is_block_ref_by_superblock(state,
2645                                                             l->block_ref_from,
2646                                                             recursion_level +
2647                                                             1))
2648                         return 1;
2649         }
2650
2651         return 0;
2652 }
2653
2654 static void btrfsic_print_add_link(const struct btrfsic_state *state,
2655                                    const struct btrfsic_block_link *l)
2656 {
2657         printk(KERN_INFO
2658                "Add %u* link from %c @%llu (%s/%llu/%d)"
2659                " to %c @%llu (%s/%llu/%d).\n",
2660                l->ref_cnt,
2661                btrfsic_get_block_type(state, l->block_ref_from),
2662                l->block_ref_from->logical_bytenr,
2663                l->block_ref_from->dev_state->name,
2664                l->block_ref_from->dev_bytenr, l->block_ref_from->mirror_num,
2665                btrfsic_get_block_type(state, l->block_ref_to),
2666                l->block_ref_to->logical_bytenr,
2667                l->block_ref_to->dev_state->name, l->block_ref_to->dev_bytenr,
2668                l->block_ref_to->mirror_num);
2669 }
2670
2671 static void btrfsic_print_rem_link(const struct btrfsic_state *state,
2672                                    const struct btrfsic_block_link *l)
2673 {
2674         printk(KERN_INFO
2675                "Rem %u* link from %c @%llu (%s/%llu/%d)"
2676                " to %c @%llu (%s/%llu/%d).\n",
2677                l->ref_cnt,
2678                btrfsic_get_block_type(state, l->block_ref_from),
2679                l->block_ref_from->logical_bytenr,
2680                l->block_ref_from->dev_state->name,
2681                l->block_ref_from->dev_bytenr, l->block_ref_from->mirror_num,
2682                btrfsic_get_block_type(state, l->block_ref_to),
2683                l->block_ref_to->logical_bytenr,
2684                l->block_ref_to->dev_state->name, l->block_ref_to->dev_bytenr,
2685                l->block_ref_to->mirror_num);
2686 }
2687
2688 static char btrfsic_get_block_type(const struct btrfsic_state *state,
2689                                    const struct btrfsic_block *block)
2690 {
2691         if (block->is_superblock &&
2692             state->latest_superblock->dev_bytenr == block->dev_bytenr &&
2693             state->latest_superblock->dev_state->bdev == block->dev_state->bdev)
2694                 return 'S';
2695         else if (block->is_superblock)
2696                 return 's';
2697         else if (block->is_metadata)
2698                 return 'M';
2699         else
2700                 return 'D';
2701 }
2702
2703 static void btrfsic_dump_tree(const struct btrfsic_state *state)
2704 {
2705         btrfsic_dump_tree_sub(state, state->latest_superblock, 0);
2706 }
2707
2708 static void btrfsic_dump_tree_sub(const struct btrfsic_state *state,
2709                                   const struct btrfsic_block *block,
2710                                   int indent_level)
2711 {
2712         struct list_head *elem_ref_to;
2713         int indent_add;
2714         static char buf[80];
2715         int cursor_position;
2716
2717         /*
2718          * Should better fill an on-stack buffer with a complete line and
2719          * dump it at once when it is time to print a newline character.
2720          */
2721
2722         /*
2723          * This algorithm is recursive because the amount of used stack space
2724          * is very small and the max recursion depth is limited.
2725          */
2726         indent_add = sprintf(buf, "%c-%llu(%s/%llu/%d)",
2727                              btrfsic_get_block_type(state, block),
2728                              block->logical_bytenr, block->dev_state->name,
2729                              block->dev_bytenr, block->mirror_num);
2730         if (indent_level + indent_add > BTRFSIC_TREE_DUMP_MAX_INDENT_LEVEL) {
2731                 printk("[...]\n");
2732                 return;
2733         }
2734         printk(buf);
2735         indent_level += indent_add;
2736         if (list_empty(&block->ref_to_list)) {
2737                 printk("\n");
2738                 return;
2739         }
2740         if (block->mirror_num > 1 &&
2741             !(state->print_mask & BTRFSIC_PRINT_MASK_TREE_WITH_ALL_MIRRORS)) {
2742                 printk(" [...]\n");
2743                 return;
2744         }
2745
2746         cursor_position = indent_level;
2747         list_for_each(elem_ref_to, &block->ref_to_list) {
2748                 const struct btrfsic_block_link *const l =
2749                     list_entry(elem_ref_to, struct btrfsic_block_link,
2750                                node_ref_to);
2751
2752                 while (cursor_position < indent_level) {
2753                         printk(" ");
2754                         cursor_position++;
2755                 }
2756                 if (l->ref_cnt > 1)
2757                         indent_add = sprintf(buf, " %d*--> ", l->ref_cnt);
2758                 else
2759                         indent_add = sprintf(buf, " --> ");
2760                 if (indent_level + indent_add >
2761                     BTRFSIC_TREE_DUMP_MAX_INDENT_LEVEL) {
2762                         printk("[...]\n");
2763                         cursor_position = 0;
2764                         continue;
2765                 }
2766
2767                 printk(buf);
2768
2769                 btrfsic_dump_tree_sub(state, l->block_ref_to,
2770                                       indent_level + indent_add);
2771                 cursor_position = 0;
2772         }
2773 }
2774
2775 static struct btrfsic_block_link *btrfsic_block_link_lookup_or_add(
2776                 struct btrfsic_state *state,
2777                 struct btrfsic_block_data_ctx *next_block_ctx,
2778                 struct btrfsic_block *next_block,
2779                 struct btrfsic_block *from_block,
2780                 u64 parent_generation)
2781 {
2782         struct btrfsic_block_link *l;
2783
2784         l = btrfsic_block_link_hashtable_lookup(next_block_ctx->dev->bdev,
2785                                                 next_block_ctx->dev_bytenr,
2786                                                 from_block->dev_state->bdev,
2787                                                 from_block->dev_bytenr,
2788                                                 &state->block_link_hashtable);
2789         if (NULL == l) {
2790                 l = btrfsic_block_link_alloc();
2791                 if (NULL == l) {
2792                         printk(KERN_INFO
2793                                "btrfsic: error, kmalloc" " failed!\n");
2794                         return NULL;
2795                 }
2796
2797                 l->block_ref_to = next_block;
2798                 l->block_ref_from = from_block;
2799                 l->ref_cnt = 1;
2800                 l->parent_generation = parent_generation;
2801
2802                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2803                         btrfsic_print_add_link(state, l);
2804
2805                 list_add(&l->node_ref_to, &from_block->ref_to_list);
2806                 list_add(&l->node_ref_from, &next_block->ref_from_list);
2807
2808                 btrfsic_block_link_hashtable_add(l,
2809                                                  &state->block_link_hashtable);
2810         } else {
2811                 l->ref_cnt++;
2812                 l->parent_generation = parent_generation;
2813                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2814                         btrfsic_print_add_link(state, l);
2815         }
2816
2817         return l;
2818 }
2819
2820 static struct btrfsic_block *btrfsic_block_lookup_or_add(
2821                 struct btrfsic_state *state,
2822                 struct btrfsic_block_data_ctx *block_ctx,
2823                 const char *additional_string,
2824                 int is_metadata,
2825                 int is_iodone,
2826                 int never_written,
2827                 int mirror_num,
2828                 int *was_created)
2829 {
2830         struct btrfsic_block *block;
2831
2832         block = btrfsic_block_hashtable_lookup(block_ctx->dev->bdev,
2833                                                block_ctx->dev_bytenr,
2834                                                &state->block_hashtable);
2835         if (NULL == block) {
2836                 struct btrfsic_dev_state *dev_state;
2837
2838                 block = btrfsic_block_alloc();
2839                 if (NULL == block) {
2840                         printk(KERN_INFO "btrfsic: error, kmalloc failed!\n");
2841                         return NULL;
2842                 }
2843                 dev_state = btrfsic_dev_state_lookup(block_ctx->dev->bdev);
2844                 if (NULL == dev_state) {
2845                         printk(KERN_INFO
2846                                "btrfsic: error, lookup dev_state failed!\n");
2847                         btrfsic_block_free(block);
2848                         return NULL;
2849                 }
2850                 block->dev_state = dev_state;
2851                 block->dev_bytenr = block_ctx->dev_bytenr;
2852                 block->logical_bytenr = block_ctx->start;
2853                 block->is_metadata = is_metadata;
2854                 block->is_iodone = is_iodone;
2855                 block->never_written = never_written;
2856                 block->mirror_num = mirror_num;
2857                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2858                         printk(KERN_INFO
2859                                "New %s%c-block @%llu (%s/%llu/%d)\n",
2860                                additional_string,
2861                                btrfsic_get_block_type(state, block),
2862                                block->logical_bytenr, dev_state->name,
2863                                block->dev_bytenr, mirror_num);
2864                 list_add(&block->all_blocks_node, &state->all_blocks_list);
2865                 btrfsic_block_hashtable_add(block, &state->block_hashtable);
2866                 if (NULL != was_created)
2867                         *was_created = 1;
2868         } else {
2869                 if (NULL != was_created)
2870                         *was_created = 0;
2871         }
2872
2873         return block;
2874 }
2875
2876 static void btrfsic_cmp_log_and_dev_bytenr(struct btrfsic_state *state,
2877                                            u64 bytenr,
2878                                            struct btrfsic_dev_state *dev_state,
2879                                            u64 dev_bytenr)
2880 {
2881         int num_copies;
2882         int mirror_num;
2883         int ret;
2884         struct btrfsic_block_data_ctx block_ctx;
2885         int match = 0;
2886
2887         num_copies = btrfs_num_copies(state->root->fs_info,
2888                                       bytenr, state->metablock_size);
2889
2890         for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
2891                 ret = btrfsic_map_block(state, bytenr, state->metablock_size,
2892                                         &block_ctx, mirror_num);
2893                 if (ret) {
2894                         printk(KERN_INFO "btrfsic:"
2895                                " btrfsic_map_block(logical @%llu,"
2896                                " mirror %d) failed!\n",
2897                                bytenr, mirror_num);
2898                         continue;
2899                 }
2900
2901                 if (dev_state->bdev == block_ctx.dev->bdev &&
2902                     dev_bytenr == block_ctx.dev_bytenr) {
2903                         match++;
2904                         btrfsic_release_block_ctx(&block_ctx);
2905                         break;
2906                 }
2907                 btrfsic_release_block_ctx(&block_ctx);
2908         }
2909
2910         if (!match) {
2911                 printk(KERN_INFO "btrfs: attempt to write M-block which contains logical bytenr that doesn't map to dev+physical bytenr of submit_bio,"
2912                        " buffer->log_bytenr=%llu, submit_bio(bdev=%s,"
2913                        " phys_bytenr=%llu)!\n",
2914                        bytenr, dev_state->name, dev_bytenr);
2915                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
2916                         ret = btrfsic_map_block(state, bytenr,
2917                                                 state->metablock_size,
2918                                                 &block_ctx, mirror_num);
2919                         if (ret)
2920                                 continue;
2921
2922                         printk(KERN_INFO "Read logical bytenr @%llu maps to"
2923                                " (%s/%llu/%d)\n",
2924                                bytenr, block_ctx.dev->name,
2925                                block_ctx.dev_bytenr, mirror_num);
2926                 }
2927                 WARN_ON(1);
2928         }
2929 }
2930
2931 static struct btrfsic_dev_state *btrfsic_dev_state_lookup(
2932                 struct block_device *bdev)
2933 {
2934         struct btrfsic_dev_state *ds;
2935
2936         ds = btrfsic_dev_state_hashtable_lookup(bdev,
2937                                                 &btrfsic_dev_state_hashtable);
2938         return ds;
2939 }
2940
2941 int btrfsic_submit_bh(int rw, struct buffer_head *bh)
2942 {
2943         struct btrfsic_dev_state *dev_state;
2944
2945         if (!btrfsic_is_initialized)
2946                 return submit_bh(rw, bh);
2947
2948         mutex_lock(&btrfsic_mutex);
2949         /* since btrfsic_submit_bh() might also be called before
2950          * btrfsic_mount(), this might return NULL */
2951         dev_state = btrfsic_dev_state_lookup(bh->b_bdev);
2952
2953         /* Only called to write the superblock (incl. FLUSH/FUA) */
2954         if (NULL != dev_state &&
2955             (rw & WRITE) && bh->b_size > 0) {
2956                 u64 dev_bytenr;
2957
2958                 dev_bytenr = 4096 * bh->b_blocknr;
2959                 if (dev_state->state->print_mask &
2960                     BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH)
2961                         printk(KERN_INFO
2962                                "submit_bh(rw=0x%x, blocknr=%lu (bytenr %llu),"
2963                                " size=%lu, data=%p, bdev=%p)\n",
2964                                rw, (unsigned long)bh->b_blocknr, dev_bytenr,
2965                                (unsigned long)bh->b_size, bh->b_data,
2966                                bh->b_bdev);
2967                 btrfsic_process_written_block(dev_state, dev_bytenr,
2968                                               &bh->b_data, 1, NULL,
2969                                               NULL, bh, rw);
2970         } else if (NULL != dev_state && (rw & REQ_FLUSH)) {
2971                 if (dev_state->state->print_mask &
2972                     BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH)
2973                         printk(KERN_INFO
2974                                "submit_bh(rw=0x%x FLUSH, bdev=%p)\n",
2975                                rw, bh->b_bdev);
2976                 if (!dev_state->dummy_block_for_bio_bh_flush.is_iodone) {
2977                         if ((dev_state->state->print_mask &
2978                              (BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH |
2979                               BTRFSIC_PRINT_MASK_VERBOSE)))
2980                                 printk(KERN_INFO
2981                                        "btrfsic_submit_bh(%s) with FLUSH"
2982                                        " but dummy block already in use"
2983                                        " (ignored)!\n",
2984                                        dev_state->name);
2985                 } else {
2986                         struct btrfsic_block *const block =
2987                                 &dev_state->dummy_block_for_bio_bh_flush;
2988
2989                         block->is_iodone = 0;
2990                         block->never_written = 0;
2991                         block->iodone_w_error = 0;
2992                         block->flush_gen = dev_state->last_flush_gen + 1;
2993                         block->submit_bio_bh_rw = rw;
2994                         block->orig_bio_bh_private = bh->b_private;
2995                         block->orig_bio_bh_end_io.bh = bh->b_end_io;
2996                         block->next_in_same_bio = NULL;
2997                         bh->b_private = block;
2998                         bh->b_end_io = btrfsic_bh_end_io;
2999                 }
3000         }
3001         mutex_unlock(&btrfsic_mutex);
3002         return submit_bh(rw, bh);
3003 }
3004
3005 void btrfsic_submit_bio(int rw, struct bio *bio)
3006 {
3007         struct btrfsic_dev_state *dev_state;
3008
3009         if (!btrfsic_is_initialized) {
3010                 submit_bio(rw, bio);
3011                 return;
3012         }
3013
3014         mutex_lock(&btrfsic_mutex);
3015         /* since btrfsic_submit_bio() is also called before
3016          * btrfsic_mount(), this might return NULL */
3017         dev_state = btrfsic_dev_state_lookup(bio->bi_bdev);
3018         if (NULL != dev_state &&
3019             (rw & WRITE) && NULL != bio->bi_io_vec) {
3020                 unsigned int i;
3021                 u64 dev_bytenr;
3022                 int bio_is_patched;
3023                 char **mapped_datav;
3024
3025                 dev_bytenr = 512 * bio->bi_sector;
3026                 bio_is_patched = 0;
3027                 if (dev_state->state->print_mask &
3028                     BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH)
3029                         printk(KERN_INFO
3030                                "submit_bio(rw=0x%x, bi_vcnt=%u,"
3031                                " bi_sector=%lu (bytenr %llu), bi_bdev=%p)\n",
3032                                rw, bio->bi_vcnt, (unsigned long)bio->bi_sector,
3033                                dev_bytenr, bio->bi_bdev);
3034
3035                 mapped_datav = kmalloc(sizeof(*mapped_datav) * bio->bi_vcnt,
3036                                        GFP_NOFS);
3037                 if (!mapped_datav)
3038                         goto leave;
3039                 for (i = 0; i < bio->bi_vcnt; i++) {
3040                         BUG_ON(bio->bi_io_vec[i].bv_len != PAGE_CACHE_SIZE);
3041                         mapped_datav[i] = kmap(bio->bi_io_vec[i].bv_page);
3042                         if (!mapped_datav[i]) {
3043                                 while (i > 0) {
3044                                         i--;
3045                                         kunmap(bio->bi_io_vec[i].bv_page);
3046                                 }
3047                                 kfree(mapped_datav);
3048                                 goto leave;
3049                         }
3050                         if ((BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH |
3051                              BTRFSIC_PRINT_MASK_VERBOSE) ==
3052                             (dev_state->state->print_mask &
3053                              (BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH |
3054                               BTRFSIC_PRINT_MASK_VERBOSE)))
3055                                 printk(KERN_INFO
3056                                        "#%u: page=%p, len=%u, offset=%u\n",
3057                                        i, bio->bi_io_vec[i].bv_page,
3058                                        bio->bi_io_vec[i].bv_len,
3059                                        bio->bi_io_vec[i].bv_offset);
3060                 }
3061                 btrfsic_process_written_block(dev_state, dev_bytenr,
3062                                               mapped_datav, bio->bi_vcnt,
3063                                               bio, &bio_is_patched,
3064                                               NULL, rw);
3065                 while (i > 0) {
3066                         i--;
3067                         kunmap(bio->bi_io_vec[i].bv_page);
3068                 }
3069                 kfree(mapped_datav);
3070         } else if (NULL != dev_state && (rw & REQ_FLUSH)) {
3071                 if (dev_state->state->print_mask &
3072                     BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH)
3073                         printk(KERN_INFO
3074                                "submit_bio(rw=0x%x FLUSH, bdev=%p)\n",
3075                                rw, bio->bi_bdev);
3076                 if (!dev_state->dummy_block_for_bio_bh_flush.is_iodone) {
3077                         if ((dev_state->state->print_mask &
3078                              (BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH |
3079                               BTRFSIC_PRINT_MASK_VERBOSE)))
3080                                 printk(KERN_INFO
3081                                        "btrfsic_submit_bio(%s) with FLUSH"
3082                                        " but dummy block already in use"
3083                                        " (ignored)!\n",
3084                                        dev_state->name);
3085                 } else {
3086                         struct btrfsic_block *const block =
3087                                 &dev_state->dummy_block_for_bio_bh_flush;
3088
3089                         block->is_iodone = 0;
3090                         block->never_written = 0;
3091                         block->iodone_w_error = 0;
3092                         block->flush_gen = dev_state->last_flush_gen + 1;
3093                         block->submit_bio_bh_rw = rw;
3094                         block->orig_bio_bh_private = bio->bi_private;
3095                         block->orig_bio_bh_end_io.bio = bio->bi_end_io;
3096                         block->next_in_same_bio = NULL;
3097                         bio->bi_private = block;
3098                         bio->bi_end_io = btrfsic_bio_end_io;
3099                 }
3100         }
3101 leave:
3102         mutex_unlock(&btrfsic_mutex);
3103
3104         submit_bio(rw, bio);
3105 }
3106
3107 int btrfsic_mount(struct btrfs_root *root,
3108                   struct btrfs_fs_devices *fs_devices,
3109                   int including_extent_data, u32 print_mask)
3110 {
3111         int ret;
3112         struct btrfsic_state *state;
3113         struct list_head *dev_head = &fs_devices->devices;
3114         struct btrfs_device *device;
3115
3116         if (root->nodesize != root->leafsize) {
3117                 printk(KERN_INFO
3118                        "btrfsic: cannot handle nodesize %d != leafsize %d!\n",
3119                        root->nodesize, root->leafsize);
3120                 return -1;
3121         }
3122         if (root->nodesize & ((u64)PAGE_CACHE_SIZE - 1)) {
3123                 printk(KERN_INFO
3124                        "btrfsic: cannot handle nodesize %d not being a multiple of PAGE_CACHE_SIZE %ld!\n",
3125                        root->nodesize, (unsigned long)PAGE_CACHE_SIZE);
3126                 return -1;
3127         }
3128         if (root->leafsize & ((u64)PAGE_CACHE_SIZE - 1)) {
3129                 printk(KERN_INFO
3130                        "btrfsic: cannot handle leafsize %d not being a multiple of PAGE_CACHE_SIZE %ld!\n",
3131                        root->leafsize, (unsigned long)PAGE_CACHE_SIZE);
3132                 return -1;
3133         }
3134         if (root->sectorsize & ((u64)PAGE_CACHE_SIZE - 1)) {
3135                 printk(KERN_INFO
3136                        "btrfsic: cannot handle sectorsize %d not being a multiple of PAGE_CACHE_SIZE %ld!\n",
3137                        root->sectorsize, (unsigned long)PAGE_CACHE_SIZE);
3138                 return -1;
3139         }
3140         state = kzalloc(sizeof(*state), GFP_NOFS);
3141         if (NULL == state) {
3142                 printk(KERN_INFO "btrfs check-integrity: kmalloc() failed!\n");
3143                 return -1;
3144         }
3145
3146         if (!btrfsic_is_initialized) {
3147                 mutex_init(&btrfsic_mutex);
3148                 btrfsic_dev_state_hashtable_init(&btrfsic_dev_state_hashtable);
3149                 btrfsic_is_initialized = 1;
3150         }
3151         mutex_lock(&btrfsic_mutex);
3152         state->root = root;
3153         state->print_mask = print_mask;
3154         state->include_extent_data = including_extent_data;
3155         state->csum_size = 0;
3156         state->metablock_size = root->nodesize;
3157         state->datablock_size = root->sectorsize;
3158         INIT_LIST_HEAD(&state->all_blocks_list);
3159         btrfsic_block_hashtable_init(&state->block_hashtable);
3160         btrfsic_block_link_hashtable_init(&state->block_link_hashtable);
3161         state->max_superblock_generation = 0;
3162         state->latest_superblock = NULL;
3163
3164         list_for_each_entry(device, dev_head, dev_list) {
3165                 struct btrfsic_dev_state *ds;
3166                 char *p;
3167
3168                 if (!device->bdev || !device->name)
3169                         continue;
3170
3171                 ds = btrfsic_dev_state_alloc();
3172                 if (NULL == ds) {
3173                         printk(KERN_INFO
3174                                "btrfs check-integrity: kmalloc() failed!\n");
3175                         mutex_unlock(&btrfsic_mutex);
3176                         return -1;
3177                 }
3178                 ds->bdev = device->bdev;
3179                 ds->state = state;
3180                 bdevname(ds->bdev, ds->name);
3181                 ds->name[BDEVNAME_SIZE - 1] = '\0';
3182                 for (p = ds->name; *p != '\0'; p++);
3183                 while (p > ds->name && *p != '/')
3184                         p--;
3185                 if (*p == '/')
3186                         p++;
3187                 strlcpy(ds->name, p, sizeof(ds->name));
3188                 btrfsic_dev_state_hashtable_add(ds,
3189                                                 &btrfsic_dev_state_hashtable);
3190         }
3191
3192         ret = btrfsic_process_superblock(state, fs_devices);
3193         if (0 != ret) {
3194                 mutex_unlock(&btrfsic_mutex);
3195                 btrfsic_unmount(root, fs_devices);
3196                 return ret;
3197         }
3198
3199         if (state->print_mask & BTRFSIC_PRINT_MASK_INITIAL_DATABASE)
3200                 btrfsic_dump_database(state);
3201         if (state->print_mask & BTRFSIC_PRINT_MASK_INITIAL_TREE)
3202                 btrfsic_dump_tree(state);
3203
3204         mutex_unlock(&btrfsic_mutex);
3205         return 0;
3206 }
3207
3208 void btrfsic_unmount(struct btrfs_root *root,
3209                      struct btrfs_fs_devices *fs_devices)
3210 {
3211         struct list_head *elem_all;
3212         struct list_head *tmp_all;
3213         struct btrfsic_state *state;
3214         struct list_head *dev_head = &fs_devices->devices;
3215         struct btrfs_device *device;
3216
3217         if (!btrfsic_is_initialized)
3218                 return;
3219
3220         mutex_lock(&btrfsic_mutex);
3221
3222         state = NULL;
3223         list_for_each_entry(device, dev_head, dev_list) {
3224                 struct btrfsic_dev_state *ds;
3225
3226                 if (!device->bdev || !device->name)
3227                         continue;
3228
3229                 ds = btrfsic_dev_state_hashtable_lookup(
3230                                 device->bdev,
3231                                 &btrfsic_dev_state_hashtable);
3232                 if (NULL != ds) {
3233                         state = ds->state;
3234                         btrfsic_dev_state_hashtable_remove(ds);
3235                         btrfsic_dev_state_free(ds);
3236                 }
3237         }
3238
3239         if (NULL == state) {
3240                 printk(KERN_INFO
3241                        "btrfsic: error, cannot find state information"
3242                        " on umount!\n");
3243                 mutex_unlock(&btrfsic_mutex);
3244                 return;
3245         }
3246
3247         /*
3248          * Don't care about keeping the lists' state up to date,
3249          * just free all memory that was allocated dynamically.
3250          * Free the blocks and the block_links.
3251          */
3252         list_for_each_safe(elem_all, tmp_all, &state->all_blocks_list) {
3253                 struct btrfsic_block *const b_all =
3254                     list_entry(elem_all, struct btrfsic_block,
3255                                all_blocks_node);
3256                 struct list_head *elem_ref_to;
3257                 struct list_head *tmp_ref_to;
3258
3259                 list_for_each_safe(elem_ref_to, tmp_ref_to,
3260                                    &b_all->ref_to_list) {
3261                         struct btrfsic_block_link *const l =
3262                             list_entry(elem_ref_to,
3263                                        struct btrfsic_block_link,
3264                                        node_ref_to);
3265
3266                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
3267                                 btrfsic_print_rem_link(state, l);
3268
3269                         l->ref_cnt--;
3270                         if (0 == l->ref_cnt)
3271                                 btrfsic_block_link_free(l);
3272                 }
3273
3274                 if (b_all->is_iodone || b_all->never_written)
3275                         btrfsic_block_free(b_all);
3276                 else
3277                         printk(KERN_INFO "btrfs: attempt to free %c-block"
3278                                " @%llu (%s/%llu/%d) on umount which is"
3279                                " not yet iodone!\n",
3280                                btrfsic_get_block_type(state, b_all),
3281                                b_all->logical_bytenr, b_all->dev_state->name,
3282                                b_all->dev_bytenr, b_all->mirror_num);
3283         }
3284
3285         mutex_unlock(&btrfsic_mutex);
3286
3287         kfree(state);
3288 }