fs/btrfs/check-integrity.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) STRATO AG 2011.  All rights reserved.
   4  */
   5
   6 /*
   7  * This module can be used to catch cases when the btrfs kernel
   8  * code executes write requests to the disk that bring the file
   9  * system in an inconsistent state. In such a state, a power-loss
  10  * or kernel panic event would cause that the data on disk is
  11  * lost or at least damaged.
  12  *
  13  * Code is added that examines all block write requests during
  14  * runtime (including writes of the super block). Three rules
  15  * are verified and an error is printed on violation of the
  16  * rules:
  17  * 1. It is not allowed to write a disk block which is
  18  *    currently referenced by the super block (either directly
  19  *    or indirectly).
  20  * 2. When a super block is written, it is verified that all
  21  *    referenced (directly or indirectly) blocks fulfill the
  22  *    following requirements:
  23  *    2a. All referenced blocks have either been present when
  24  *        the file system was mounted, (i.e., they have been
  25  *        referenced by the super block) or they have been
  26  *        written since then and the write completion callback
  27  *        was called and no write error was indicated and a
  28  *        FLUSH request to the device where these blocks are
  29  *        located was received and completed.
  30  *    2b. All referenced blocks need to have a generation
  31  *        number which is equal to the parent's number.
  32  *
  33  * One issue that was found using this module was that the log
  34  * tree on disk became temporarily corrupted because disk blocks
  35  * that had been in use for the log tree had been freed and
  36  * reused too early, while being referenced by the written super
  37  * block.
  38  *
  39  * The search term in the kernel log that can be used to filter
  40  * on the existence of detected integrity issues is
  41  * "btrfs: attempt".
  42  *
  43  * The integrity check is enabled via mount options. These
  44  * mount options are only supported if the integrity check
  45  * tool is compiled by defining BTRFS_FS_CHECK_INTEGRITY.
  46  *
  47  * Example #1, apply integrity checks to all metadata:
  48  * mount /dev/sdb1 /mnt -o check_int
  49  *
  50  * Example #2, apply integrity checks to all metadata and
  51  * to data extents:
  52  * mount /dev/sdb1 /mnt -o check_int_data
  53  *
  54  * Example #3, apply integrity checks to all metadata and dump
  55  * the tree that the super block references to kernel messages
  56  * each time after a super block was written:
  57  * mount /dev/sdb1 /mnt -o check_int,check_int_print_mask=263
  58  *
  59  * If the integrity check tool is included and activated in
  60  * the mount options, plenty of kernel memory is used, and
  61  * plenty of additional CPU cycles are spent. Enabling this
  62  * functionality is not intended for normal use. In most
  63  * cases, unless you are a btrfs developer who needs to verify
  64  * the integrity of (super)-block write requests, do not
  65  * enable the config option BTRFS_FS_CHECK_INTEGRITY to
  66  * include and compile the integrity check tool.
  67  *
  68  * Expect millions of lines of information in the kernel log with an
  69  * enabled check_int_print_mask. Therefore set LOG_BUF_SHIFT in the
  70  * kernel config to at least 26 (which is 64MB). Usually the value is
  71  * limited to 21 (which is 2MB) in init/Kconfig. The file needs to be
  72  * changed like this before LOG_BUF_SHIFT can be set to a high value:
  73  * config LOG_BUF_SHIFT
  74  *       int "Kernel log buffer size (16 => 64KB, 17 => 128KB)"
  75  *       range 12 30
  76  */
  77
  78 #include <linux/sched.h>
  79 #include <linux/slab.h>
  80 #include <linux/mutex.h>
  81 #include <linux/blkdev.h>
  82 #include <linux/mm.h>
  83 #include <linux/string.h>
  84 #include <crypto/hash.h>
  85 #include "messages.h"
  86 #include "ctree.h"
  87 #include "disk-io.h"
  88 #include "transaction.h"
  89 #include "extent_io.h"
  90 #include "volumes.h"
  91 #include "print-tree.h"
  92 #include "locking.h"
  93 #include "check-integrity.h"
  94 #include "rcu-string.h"
  95 #include "compression.h"
  96
  97 #define BTRFSIC_BLOCK_HASHTABLE_SIZE 0x10000
  98 #define BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE 0x10000
  99 #define BTRFSIC_DEV2STATE_HASHTABLE_SIZE 0x100
 100 #define BTRFSIC_BLOCK_MAGIC_NUMBER 0x14491051
 101 #define BTRFSIC_BLOCK_LINK_MAGIC_NUMBER 0x11070807
 102 #define BTRFSIC_DEV2STATE_MAGIC_NUMBER 0x20111530
 103 #define BTRFSIC_BLOCK_STACK_FRAME_MAGIC_NUMBER 20111300
 104 #define BTRFSIC_TREE_DUMP_MAX_INDENT_LEVEL (200 - 6)    /* in characters,
 105                                                          * excluding " [...]" */
 106 #define BTRFSIC_GENERATION_UNKNOWN ((u64)-1)
 107
 108 /*
 109  * The definition of the bitmask fields for the print_mask.
 110  * They are specified with the mount option check_integrity_print_mask.
 111  */
 112 #define BTRFSIC_PRINT_MASK_SUPERBLOCK_WRITE                     0x00000001
 113 #define BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION         0x00000002
 114 #define BTRFSIC_PRINT_MASK_TREE_AFTER_SB_WRITE                  0x00000004
 115 #define BTRFSIC_PRINT_MASK_TREE_BEFORE_SB_WRITE                 0x00000008
 116 #define BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH                        0x00000010
 117 #define BTRFSIC_PRINT_MASK_END_IO_BIO_BH                        0x00000020
 118 #define BTRFSIC_PRINT_MASK_VERBOSE                              0x00000040
 119 #define BTRFSIC_PRINT_MASK_VERY_VERBOSE                         0x00000080
 120 #define BTRFSIC_PRINT_MASK_INITIAL_TREE                         0x00000100
 121 #define BTRFSIC_PRINT_MASK_INITIAL_ALL_TREES                    0x00000200
 122 #define BTRFSIC_PRINT_MASK_INITIAL_DATABASE                     0x00000400
 123 #define BTRFSIC_PRINT_MASK_NUM_COPIES                           0x00000800
 124 #define BTRFSIC_PRINT_MASK_TREE_WITH_ALL_MIRRORS                0x00001000
 125 #define BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH_VERBOSE                0x00002000
 126
 127 struct btrfsic_dev_state;
 128 struct btrfsic_state;
 129
 130 struct btrfsic_block {
 131         u32 magic_num;          /* only used for debug purposes */
 132         unsigned int is_metadata:1;     /* if it is meta-data, not data-data */
 133         unsigned int is_superblock:1;   /* if it is one of the superblocks */
 134         unsigned int is_iodone:1;       /* if is done by lower subsystem */
 135         unsigned int iodone_w_error:1;  /* error was indicated to endio */
 136         unsigned int never_written:1;   /* block was added because it was
 137                                          * referenced, not because it was
 138                                          * written */
 139         unsigned int mirror_num;        /* large enough to hold
 140                                          * BTRFS_SUPER_MIRROR_MAX */
 141         struct btrfsic_dev_state *dev_state;
 142         u64 dev_bytenr;         /* key, physical byte num on disk */
 143         u64 logical_bytenr;     /* logical byte num on disk */
 144         u64 generation;
 145         struct btrfs_disk_key disk_key; /* extra info to print in case of
 146                                          * issues, will not always be correct */
 147         struct list_head collision_resolving_node;      /* list node */
 148         struct list_head all_blocks_node;       /* list node */
 149
 150         /* the following two lists contain block_link items */
 151         struct list_head ref_to_list;   /* list */
 152         struct list_head ref_from_list; /* list */
 153         struct btrfsic_block *next_in_same_bio;
 154         void *orig_bio_private;
 155         bio_end_io_t *orig_bio_end_io;
 156         blk_opf_t submit_bio_bh_rw;
 157         u64 flush_gen; /* only valid if !never_written */
 158 };
 159
 160 /*
 161  * Elements of this type are allocated dynamically and required because
 162  * each block object can refer to and can be ref from multiple blocks.
 163  * The key to lookup them in the hashtable is the dev_bytenr of
 164  * the block ref to plus the one from the block referred from.
 165  * The fact that they are searchable via a hashtable and that a
 166  * ref_cnt is maintained is not required for the btrfs integrity
 167  * check algorithm itself, it is only used to make the output more
 168  * beautiful in case that an error is detected (an error is defined
 169  * as a write operation to a block while that block is still referenced).
 170  */
 171 struct btrfsic_block_link {
 172         u32 magic_num;          /* only used for debug purposes */
 173         u32 ref_cnt;
 174         struct list_head node_ref_to;   /* list node */
 175         struct list_head node_ref_from; /* list node */
 176         struct list_head collision_resolving_node;      /* list node */
 177         struct btrfsic_block *block_ref_to;
 178         struct btrfsic_block *block_ref_from;
 179         u64 parent_generation;
 180 };
 181
 182 struct btrfsic_dev_state {
 183         u32 magic_num;          /* only used for debug purposes */
 184         struct block_device *bdev;
 185         struct btrfsic_state *state;
 186         struct list_head collision_resolving_node;      /* list node */
 187         struct btrfsic_block dummy_block_for_bio_bh_flush;
 188         u64 last_flush_gen;
 189 };
 190
 191 struct btrfsic_block_hashtable {
 192         struct list_head table[BTRFSIC_BLOCK_HASHTABLE_SIZE];
 193 };
 194
 195 struct btrfsic_block_link_hashtable {
 196         struct list_head table[BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE];
 197 };
 198
 199 struct btrfsic_dev_state_hashtable {
 200         struct list_head table[BTRFSIC_DEV2STATE_HASHTABLE_SIZE];
 201 };
 202
 203 struct btrfsic_block_data_ctx {
 204         u64 start;              /* virtual bytenr */
 205         u64 dev_bytenr;         /* physical bytenr on device */
 206         u32 len;
 207         struct btrfsic_dev_state *dev;
 208         char **datav;
 209         struct page **pagev;
 210         void *mem_to_free;
 211 };
 212
 213 /* This structure is used to implement recursion without occupying
 214  * any stack space, refer to btrfsic_process_metablock() */
 215 struct btrfsic_stack_frame {
 216         u32 magic;
 217         u32 nr;
 218         int error;
 219         int i;
 220         int limit_nesting;
 221         int num_copies;
 222         int mirror_num;
 223         struct btrfsic_block *block;
 224         struct btrfsic_block_data_ctx *block_ctx;
 225         struct btrfsic_block *next_block;
 226         struct btrfsic_block_data_ctx next_block_ctx;
 227         struct btrfs_header *hdr;
 228         struct btrfsic_stack_frame *prev;
 229 };
 230
 231 /* Some state per mounted filesystem */
 232 struct btrfsic_state {
 233         u32 print_mask;
 234         int include_extent_data;
 235         struct list_head all_blocks_list;
 236         struct btrfsic_block_hashtable block_hashtable;
 237         struct btrfsic_block_link_hashtable block_link_hashtable;
 238         struct btrfs_fs_info *fs_info;
 239         u64 max_superblock_generation;
 240         struct btrfsic_block *latest_superblock;
 241         u32 metablock_size;
 242         u32 datablock_size;
 243 };
 244
 245 static int btrfsic_process_metablock(struct btrfsic_state *state,
 246                                      struct btrfsic_block *block,
 247                                      struct btrfsic_block_data_ctx *block_ctx,
 248                                      int limit_nesting, int force_iodone_flag);
 249 static void btrfsic_read_from_block_data(
 250         struct btrfsic_block_data_ctx *block_ctx,
 251         void *dst, u32 offset, size_t len);
 252 static int btrfsic_create_link_to_next_block(
 253                 struct btrfsic_state *state,
 254                 struct btrfsic_block *block,
 255                 struct btrfsic_block_data_ctx
 256                 *block_ctx, u64 next_bytenr,
 257                 int limit_nesting,
 258                 struct btrfsic_block_data_ctx *next_block_ctx,
 259                 struct btrfsic_block **next_blockp,
 260                 int force_iodone_flag,
 261                 int *num_copiesp, int *mirror_nump,
 262                 struct btrfs_disk_key *disk_key,
 263                 u64 parent_generation);
 264 static int btrfsic_handle_extent_data(struct btrfsic_state *state,
 265                                       struct btrfsic_block *block,
 266                                       struct btrfsic_block_data_ctx *block_ctx,
 267                                       u32 item_offset, int force_iodone_flag);
 268 static int btrfsic_map_block(struct btrfsic_state *state, u64 bytenr, u32 len,
 269                              struct btrfsic_block_data_ctx *block_ctx_out,
 270                              int mirror_num);
 271 static void btrfsic_release_block_ctx(struct btrfsic_block_data_ctx *block_ctx);
 272 static int btrfsic_read_block(struct btrfsic_state *state,
 273                               struct btrfsic_block_data_ctx *block_ctx);
 274 static int btrfsic_process_written_superblock(
 275                 struct btrfsic_state *state,
 276                 struct btrfsic_block *const block,
 277                 struct btrfs_super_block *const super_hdr);
 278 static void btrfsic_bio_end_io(struct bio *bp);
 279 static int btrfsic_is_block_ref_by_superblock(const struct btrfsic_state *state,
 280                                               const struct btrfsic_block *block,
 281                                               int recursion_level);
 282 static int btrfsic_check_all_ref_blocks(struct btrfsic_state *state,
 283                                         struct btrfsic_block *const block,
 284                                         int recursion_level);
 285 static void btrfsic_print_add_link(const struct btrfsic_state *state,
 286                                    const struct btrfsic_block_link *l);
 287 static void btrfsic_print_rem_link(const struct btrfsic_state *state,
 288                                    const struct btrfsic_block_link *l);
 289 static char btrfsic_get_block_type(const struct btrfsic_state *state,
 290                                    const struct btrfsic_block *block);
 291 static void btrfsic_dump_tree(const struct btrfsic_state *state);
 292 static void btrfsic_dump_tree_sub(const struct btrfsic_state *state,
 293                                   const struct btrfsic_block *block,
 294                                   int indent_level);
 295 static struct btrfsic_block_link *btrfsic_block_link_lookup_or_add(
 296                 struct btrfsic_state *state,
 297                 struct btrfsic_block_data_ctx *next_block_ctx,
 298                 struct btrfsic_block *next_block,
 299                 struct btrfsic_block *from_block,
 300                 u64 parent_generation);
 301 static struct btrfsic_block *btrfsic_block_lookup_or_add(
 302                 struct btrfsic_state *state,
 303                 struct btrfsic_block_data_ctx *block_ctx,
 304                 const char *additional_string,
 305                 int is_metadata,
 306                 int is_iodone,
 307                 int never_written,
 308                 int mirror_num,
 309                 int *was_created);
 310 static int btrfsic_process_superblock_dev_mirror(
 311                 struct btrfsic_state *state,
 312                 struct btrfsic_dev_state *dev_state,
 313                 struct btrfs_device *device,
 314                 int superblock_mirror_num,
 315                 struct btrfsic_dev_state **selected_dev_state,
 316                 struct btrfs_super_block *selected_super);
 317 static struct btrfsic_dev_state *btrfsic_dev_state_lookup(dev_t dev);
 318 static void btrfsic_cmp_log_and_dev_bytenr(struct btrfsic_state *state,
 319                                            u64 bytenr,
 320                                            struct btrfsic_dev_state *dev_state,
 321                                            u64 dev_bytenr);
 322
 323 static struct mutex btrfsic_mutex;
 324 static int btrfsic_is_initialized;
 325 static struct btrfsic_dev_state_hashtable btrfsic_dev_state_hashtable;
 326
 327
 328 static void btrfsic_block_init(struct btrfsic_block *b)
 329 {
 330         b->magic_num = BTRFSIC_BLOCK_MAGIC_NUMBER;
 331         b->dev_state = NULL;
 332         b->dev_bytenr = 0;
 333         b->logical_bytenr = 0;
 334         b->generation = BTRFSIC_GENERATION_UNKNOWN;
 335         b->disk_key.objectid = 0;
 336         b->disk_key.type = 0;
 337         b->disk_key.offset = 0;
 338         b->is_metadata = 0;
 339         b->is_superblock = 0;
 340         b->is_iodone = 0;
 341         b->iodone_w_error = 0;
 342         b->never_written = 0;
 343         b->mirror_num = 0;
 344         b->next_in_same_bio = NULL;
 345         b->orig_bio_private = NULL;
 346         b->orig_bio_end_io = NULL;
 347         INIT_LIST_HEAD(&b->collision_resolving_node);
 348         INIT_LIST_HEAD(&b->all_blocks_node);
 349         INIT_LIST_HEAD(&b->ref_to_list);
 350         INIT_LIST_HEAD(&b->ref_from_list);
 351         b->submit_bio_bh_rw = 0;
 352         b->flush_gen = 0;
 353 }
 354
 355 static struct btrfsic_block *btrfsic_block_alloc(void)
 356 {
 357         struct btrfsic_block *b;
 358
 359         b = kzalloc(sizeof(*b), GFP_NOFS);
 360         if (NULL != b)
 361                 btrfsic_block_init(b);
 362
 363         return b;
 364 }
 365
 366 static void btrfsic_block_free(struct btrfsic_block *b)
 367 {
 368         BUG_ON(!(NULL == b || BTRFSIC_BLOCK_MAGIC_NUMBER == b->magic_num));
 369         kfree(b);
 370 }
 371
 372 static void btrfsic_block_link_init(struct btrfsic_block_link *l)
 373 {
 374         l->magic_num = BTRFSIC_BLOCK_LINK_MAGIC_NUMBER;
 375         l->ref_cnt = 1;
 376         INIT_LIST_HEAD(&l->node_ref_to);
 377         INIT_LIST_HEAD(&l->node_ref_from);
 378         INIT_LIST_HEAD(&l->collision_resolving_node);
 379         l->block_ref_to = NULL;
 380         l->block_ref_from = NULL;
 381 }
 382
 383 static struct btrfsic_block_link *btrfsic_block_link_alloc(void)
 384 {
 385         struct btrfsic_block_link *l;
 386
 387         l = kzalloc(sizeof(*l), GFP_NOFS);
 388         if (NULL != l)
 389                 btrfsic_block_link_init(l);
 390
 391         return l;
 392 }
 393
 394 static void btrfsic_block_link_free(struct btrfsic_block_link *l)
 395 {
 396         BUG_ON(!(NULL == l || BTRFSIC_BLOCK_LINK_MAGIC_NUMBER == l->magic_num));
 397         kfree(l);
 398 }
 399
 400 static void btrfsic_dev_state_init(struct btrfsic_dev_state *ds)
 401 {
 402         ds->magic_num = BTRFSIC_DEV2STATE_MAGIC_NUMBER;
 403         ds->bdev = NULL;
 404         ds->state = NULL;
 405         INIT_LIST_HEAD(&ds->collision_resolving_node);
 406         ds->last_flush_gen = 0;
 407         btrfsic_block_init(&ds->dummy_block_for_bio_bh_flush);
 408         ds->dummy_block_for_bio_bh_flush.is_iodone = 1;
 409         ds->dummy_block_for_bio_bh_flush.dev_state = ds;
 410 }
 411
 412 static struct btrfsic_dev_state *btrfsic_dev_state_alloc(void)
 413 {
 414         struct btrfsic_dev_state *ds;
 415
 416         ds = kzalloc(sizeof(*ds), GFP_NOFS);
 417         if (NULL != ds)
 418                 btrfsic_dev_state_init(ds);
 419
 420         return ds;
 421 }
 422
 423 static void btrfsic_dev_state_free(struct btrfsic_dev_state *ds)
 424 {
 425         BUG_ON(!(NULL == ds ||
 426                  BTRFSIC_DEV2STATE_MAGIC_NUMBER == ds->magic_num));
 427         kfree(ds);
 428 }
 429
 430 static void btrfsic_block_hashtable_init(struct btrfsic_block_hashtable *h)
 431 {
 432         int i;
 433
 434         for (i = 0; i < BTRFSIC_BLOCK_HASHTABLE_SIZE; i++)
 435                 INIT_LIST_HEAD(h->table + i);
 436 }
 437
 438 static void btrfsic_block_hashtable_add(struct btrfsic_block *b,
 439                                         struct btrfsic_block_hashtable *h)
 440 {
 441         const unsigned int hashval =
 442             (((unsigned int)(b->dev_bytenr >> 16)) ^
 443              ((unsigned int)((uintptr_t)b->dev_state->bdev))) &
 444              (BTRFSIC_BLOCK_HASHTABLE_SIZE - 1);
 445
 446         list_add(&b->collision_resolving_node, h->table + hashval);
 447 }
 448
 449 static void btrfsic_block_hashtable_remove(struct btrfsic_block *b)
 450 {
 451         list_del(&b->collision_resolving_node);
 452 }
 453
 454 static struct btrfsic_block *btrfsic_block_hashtable_lookup(
 455                 struct block_device *bdev,
 456                 u64 dev_bytenr,
 457                 struct btrfsic_block_hashtable *h)
 458 {
 459         const unsigned int hashval =
 460             (((unsigned int)(dev_bytenr >> 16)) ^
 461              ((unsigned int)((uintptr_t)bdev))) &
 462              (BTRFSIC_BLOCK_HASHTABLE_SIZE - 1);
 463         struct btrfsic_block *b;
 464
 465         list_for_each_entry(b, h->table + hashval, collision_resolving_node) {
 466                 if (b->dev_state->bdev == bdev && b->dev_bytenr == dev_bytenr)
 467                         return b;
 468         }
 469
 470         return NULL;
 471 }
 472
 473 static void btrfsic_block_link_hashtable_init(
 474                 struct btrfsic_block_link_hashtable *h)
 475 {
 476         int i;
 477
 478         for (i = 0; i < BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE; i++)
 479                 INIT_LIST_HEAD(h->table + i);
 480 }
 481
 482 static void btrfsic_block_link_hashtable_add(
 483                 struct btrfsic_block_link *l,
 484                 struct btrfsic_block_link_hashtable *h)
 485 {
 486         const unsigned int hashval =
 487             (((unsigned int)(l->block_ref_to->dev_bytenr >> 16)) ^
 488              ((unsigned int)(l->block_ref_from->dev_bytenr >> 16)) ^
 489              ((unsigned int)((uintptr_t)l->block_ref_to->dev_state->bdev)) ^
 490              ((unsigned int)((uintptr_t)l->block_ref_from->dev_state->bdev)))
 491              & (BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE - 1);
 492
 493         BUG_ON(NULL == l->block_ref_to);
 494         BUG_ON(NULL == l->block_ref_from);
 495         list_add(&l->collision_resolving_node, h->table + hashval);
 496 }
 497
 498 static void btrfsic_block_link_hashtable_remove(struct btrfsic_block_link *l)
 499 {
 500         list_del(&l->collision_resolving_node);
 501 }
 502
 503 static struct btrfsic_block_link *btrfsic_block_link_hashtable_lookup(
 504                 struct block_device *bdev_ref_to,
 505                 u64 dev_bytenr_ref_to,
 506                 struct block_device *bdev_ref_from,
 507                 u64 dev_bytenr_ref_from,
 508                 struct btrfsic_block_link_hashtable *h)
 509 {
 510         const unsigned int hashval =
 511             (((unsigned int)(dev_bytenr_ref_to >> 16)) ^
 512              ((unsigned int)(dev_bytenr_ref_from >> 16)) ^
 513              ((unsigned int)((uintptr_t)bdev_ref_to)) ^
 514              ((unsigned int)((uintptr_t)bdev_ref_from))) &
 515              (BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE - 1);
 516         struct btrfsic_block_link *l;
 517
 518         list_for_each_entry(l, h->table + hashval, collision_resolving_node) {
 519                 BUG_ON(NULL == l->block_ref_to);
 520                 BUG_ON(NULL == l->block_ref_from);
 521                 if (l->block_ref_to->dev_state->bdev == bdev_ref_to &&
 522                     l->block_ref_to->dev_bytenr == dev_bytenr_ref_to &&
 523                     l->block_ref_from->dev_state->bdev == bdev_ref_from &&
 524                     l->block_ref_from->dev_bytenr == dev_bytenr_ref_from)
 525                         return l;
 526         }
 527
 528         return NULL;
 529 }
 530
 531 static void btrfsic_dev_state_hashtable_init(
 532                 struct btrfsic_dev_state_hashtable *h)
 533 {
 534         int i;
 535
 536         for (i = 0; i < BTRFSIC_DEV2STATE_HASHTABLE_SIZE; i++)
 537                 INIT_LIST_HEAD(h->table + i);
 538 }
 539
 540 static void btrfsic_dev_state_hashtable_add(
 541                 struct btrfsic_dev_state *ds,
 542                 struct btrfsic_dev_state_hashtable *h)
 543 {
 544         const unsigned int hashval =
 545             (((unsigned int)((uintptr_t)ds->bdev->bd_dev)) &
 546              (BTRFSIC_DEV2STATE_HASHTABLE_SIZE - 1));
 547
 548         list_add(&ds->collision_resolving_node, h->table + hashval);
 549 }
 550
 551 static void btrfsic_dev_state_hashtable_remove(struct btrfsic_dev_state *ds)
 552 {
 553         list_del(&ds->collision_resolving_node);
 554 }
 555
 556 static struct btrfsic_dev_state *btrfsic_dev_state_hashtable_lookup(dev_t dev,
 557                 struct btrfsic_dev_state_hashtable *h)
 558 {
 559         const unsigned int hashval =
 560                 dev & (BTRFSIC_DEV2STATE_HASHTABLE_SIZE - 1);
 561         struct btrfsic_dev_state *ds;
 562
 563         list_for_each_entry(ds, h->table + hashval, collision_resolving_node) {
 564                 if (ds->bdev->bd_dev == dev)
 565                         return ds;
 566         }
 567
 568         return NULL;
 569 }
 570
 571 static int btrfsic_process_superblock(struct btrfsic_state *state,
 572                                       struct btrfs_fs_devices *fs_devices)
 573 {
 574         struct btrfs_super_block *selected_super;
 575         struct list_head *dev_head = &fs_devices->devices;
 576         struct btrfs_device *device;
 577         struct btrfsic_dev_state *selected_dev_state = NULL;
 578         int ret = 0;
 579         int pass;
 580
 581         selected_super = kzalloc(sizeof(*selected_super), GFP_NOFS);
 582         if (!selected_super)
 583                 return -ENOMEM;
 584
 585         list_for_each_entry(device, dev_head, dev_list) {
 586                 int i;
 587                 struct btrfsic_dev_state *dev_state;
 588
 589                 if (!device->bdev || !device->name)
 590                         continue;
 591
 592                 dev_state = btrfsic_dev_state_lookup(device->bdev->bd_dev);
 593                 BUG_ON(NULL == dev_state);
 594                 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 595                         ret = btrfsic_process_superblock_dev_mirror(
 596                                         state, dev_state, device, i,
 597                                         &selected_dev_state, selected_super);
 598                         if (0 != ret && 0 == i) {
 599                                 kfree(selected_super);
 600                                 return ret;
 601                         }
 602                 }
 603         }
 604
 605         if (NULL == state->latest_superblock) {
 606                 pr_info("btrfsic: no superblock found!\n");
 607                 kfree(selected_super);
 608                 return -1;
 609         }
 610
 611         for (pass = 0; pass < 3; pass++) {
 612                 int num_copies;
 613                 int mirror_num;
 614                 u64 next_bytenr;
 615
 616                 switch (pass) {
 617                 case 0:
 618                         next_bytenr = btrfs_super_root(selected_super);
 619                         if (state->print_mask &
 620                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
 621                                 pr_info("root@%llu\n", next_bytenr);
 622                         break;
 623                 case 1:
 624                         next_bytenr = btrfs_super_chunk_root(selected_super);
 625                         if (state->print_mask &
 626                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
 627                                 pr_info("chunk@%llu\n", next_bytenr);
 628                         break;
 629                 case 2:
 630                         next_bytenr = btrfs_super_log_root(selected_super);
 631                         if (0 == next_bytenr)
 632                                 continue;
 633                         if (state->print_mask &
 634                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
 635                                 pr_info("log@%llu\n", next_bytenr);
 636                         break;
 637                 }
 638
 639                 num_copies = btrfs_num_copies(state->fs_info, next_bytenr,
 640                                               state->metablock_size);
 641                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
 642                         pr_info("num_copies(log_bytenr=%llu) = %d\n",
 643                                next_bytenr, num_copies);
 644
 645                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
 646                         struct btrfsic_block *next_block;
 647                         struct btrfsic_block_data_ctx tmp_next_block_ctx;
 648                         struct btrfsic_block_link *l;
 649
 650                         ret = btrfsic_map_block(state, next_bytenr,
 651                                                 state->metablock_size,
 652                                                 &tmp_next_block_ctx,
 653                                                 mirror_num);
 654                         if (ret) {
 655                                 pr_info("btrfsic: btrfsic_map_block(root @%llu, mirror %d) failed!\n",
 656                                        next_bytenr, mirror_num);
 657                                 kfree(selected_super);
 658                                 return -1;
 659                         }
 660
 661                         next_block = btrfsic_block_hashtable_lookup(
 662                                         tmp_next_block_ctx.dev->bdev,
 663                                         tmp_next_block_ctx.dev_bytenr,
 664                                         &state->block_hashtable);
 665                         BUG_ON(NULL == next_block);
 666
 667                         l = btrfsic_block_link_hashtable_lookup(
 668                                         tmp_next_block_ctx.dev->bdev,
 669                                         tmp_next_block_ctx.dev_bytenr,
 670                                         state->latest_superblock->dev_state->
 671                                         bdev,
 672                                         state->latest_superblock->dev_bytenr,
 673                                         &state->block_link_hashtable);
 674                         BUG_ON(NULL == l);
 675
 676                         ret = btrfsic_read_block(state, &tmp_next_block_ctx);
 677                         if (ret < (int)PAGE_SIZE) {
 678                                 pr_info("btrfsic: read @logical %llu failed!\n",
 679                                        tmp_next_block_ctx.start);
 680                                 btrfsic_release_block_ctx(&tmp_next_block_ctx);
 681                                 kfree(selected_super);
 682                                 return -1;
 683                         }
 684
 685                         ret = btrfsic_process_metablock(state,
 686                                                         next_block,
 687                                                         &tmp_next_block_ctx,
 688                                                         BTRFS_MAX_LEVEL + 3, 1);
 689                         btrfsic_release_block_ctx(&tmp_next_block_ctx);
 690                 }
 691         }
 692
 693         kfree(selected_super);
 694         return ret;
 695 }
 696
 697 static int btrfsic_process_superblock_dev_mirror(
 698                 struct btrfsic_state *state,
 699                 struct btrfsic_dev_state *dev_state,
 700                 struct btrfs_device *device,
 701                 int superblock_mirror_num,
 702                 struct btrfsic_dev_state **selected_dev_state,
 703                 struct btrfs_super_block *selected_super)
 704 {
 705         struct btrfs_fs_info *fs_info = state->fs_info;
 706         struct btrfs_super_block *super_tmp;
 707         u64 dev_bytenr;
 708         struct btrfsic_block *superblock_tmp;
 709         int pass;
 710         struct block_device *const superblock_bdev = device->bdev;
 711         struct page *page;
 712         struct address_space *mapping = superblock_bdev->bd_inode->i_mapping;
 713         int ret = 0;
 714
 715         /* super block bytenr is always the unmapped device bytenr */
 716         dev_bytenr = btrfs_sb_offset(superblock_mirror_num);
 717         if (dev_bytenr + BTRFS_SUPER_INFO_SIZE > device->commit_total_bytes)
 718                 return -1;
 719
 720         page = read_cache_page_gfp(mapping, dev_bytenr >> PAGE_SHIFT, GFP_NOFS);
 721         if (IS_ERR(page))
 722                 return -1;
 723
 724         super_tmp = page_address(page);
 725
 726         if (btrfs_super_bytenr(super_tmp) != dev_bytenr ||
 727             btrfs_super_magic(super_tmp) != BTRFS_MAGIC ||
 728             memcmp(device->uuid, super_tmp->dev_item.uuid, BTRFS_UUID_SIZE) ||
 729             btrfs_super_nodesize(super_tmp) != state->metablock_size ||
 730             btrfs_super_sectorsize(super_tmp) != state->datablock_size) {
 731                 ret = 0;
 732                 goto out;
 733         }
 734
 735         superblock_tmp =
 736             btrfsic_block_hashtable_lookup(superblock_bdev,
 737                                            dev_bytenr,
 738                                            &state->block_hashtable);
 739         if (NULL == superblock_tmp) {
 740                 superblock_tmp = btrfsic_block_alloc();
 741                 if (NULL == superblock_tmp) {
 742                         ret = -1;
 743                         goto out;
 744                 }
 745                 /* for superblock, only the dev_bytenr makes sense */
 746                 superblock_tmp->dev_bytenr = dev_bytenr;
 747                 superblock_tmp->dev_state = dev_state;
 748                 superblock_tmp->logical_bytenr = dev_bytenr;
 749                 superblock_tmp->generation = btrfs_super_generation(super_tmp);
 750                 superblock_tmp->is_metadata = 1;
 751                 superblock_tmp->is_superblock = 1;
 752                 superblock_tmp->is_iodone = 1;
 753                 superblock_tmp->never_written = 0;
 754                 superblock_tmp->mirror_num = 1 + superblock_mirror_num;
 755                 if (state->print_mask & BTRFSIC_PRINT_MASK_SUPERBLOCK_WRITE)
 756                         btrfs_info_in_rcu(fs_info,
 757                         "new initial S-block (bdev %p, %s) @%llu (%pg/%llu/%d)",
 758                                      superblock_bdev,
 759                                      rcu_str_deref(device->name), dev_bytenr,
 760                                      dev_state->bdev, dev_bytenr,
 761                                      superblock_mirror_num);
 762                 list_add(&superblock_tmp->all_blocks_node,
 763                          &state->all_blocks_list);
 764                 btrfsic_block_hashtable_add(superblock_tmp,
 765                                             &state->block_hashtable);
 766         }
 767
 768         /* select the one with the highest generation field */
 769         if (btrfs_super_generation(super_tmp) >
 770             state->max_superblock_generation ||
 771             0 == state->max_superblock_generation) {
 772                 memcpy(selected_super, super_tmp, sizeof(*selected_super));
 773                 *selected_dev_state = dev_state;
 774                 state->max_superblock_generation =
 775                     btrfs_super_generation(super_tmp);
 776                 state->latest_superblock = superblock_tmp;
 777         }
 778
 779         for (pass = 0; pass < 3; pass++) {
 780                 u64 next_bytenr;
 781                 int num_copies;
 782                 int mirror_num;
 783                 const char *additional_string = NULL;
 784                 struct btrfs_disk_key tmp_disk_key;
 785
 786                 tmp_disk_key.type = BTRFS_ROOT_ITEM_KEY;
 787                 tmp_disk_key.offset = 0;
 788                 switch (pass) {
 789                 case 0:
 790                         btrfs_set_disk_key_objectid(&tmp_disk_key,
 791                                                     BTRFS_ROOT_TREE_OBJECTID);
 792                         additional_string = "initial root ";
 793                         next_bytenr = btrfs_super_root(super_tmp);
 794                         break;
 795                 case 1:
 796                         btrfs_set_disk_key_objectid(&tmp_disk_key,
 797                                                     BTRFS_CHUNK_TREE_OBJECTID);
 798                         additional_string = "initial chunk ";
 799                         next_bytenr = btrfs_super_chunk_root(super_tmp);
 800                         break;
 801                 case 2:
 802                         btrfs_set_disk_key_objectid(&tmp_disk_key,
 803                                                     BTRFS_TREE_LOG_OBJECTID);
 804                         additional_string = "initial log ";
 805                         next_bytenr = btrfs_super_log_root(super_tmp);
 806                         if (0 == next_bytenr)
 807                                 continue;
 808                         break;
 809                 }
 810
 811                 num_copies = btrfs_num_copies(fs_info, next_bytenr,
 812                                               state->metablock_size);
 813                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
 814                         pr_info("num_copies(log_bytenr=%llu) = %d\n",
 815                                next_bytenr, num_copies);
 816                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
 817                         struct btrfsic_block *next_block;
 818                         struct btrfsic_block_data_ctx tmp_next_block_ctx;
 819                         struct btrfsic_block_link *l;
 820
 821                         if (btrfsic_map_block(state, next_bytenr,
 822                                               state->metablock_size,
 823                                               &tmp_next_block_ctx,
 824                                               mirror_num)) {
 825                                 pr_info("btrfsic: btrfsic_map_block(bytenr @%llu, mirror %d) failed!\n",
 826                                        next_bytenr, mirror_num);
 827                                 ret = -1;
 828                                 goto out;
 829                         }
 830
 831                         next_block = btrfsic_block_lookup_or_add(
 832                                         state, &tmp_next_block_ctx,
 833                                         additional_string, 1, 1, 0,
 834                                         mirror_num, NULL);
 835                         if (NULL == next_block) {
 836                                 btrfsic_release_block_ctx(&tmp_next_block_ctx);
 837                                 ret = -1;
 838                                 goto out;
 839                         }
 840
 841                         next_block->disk_key = tmp_disk_key;
 842                         next_block->generation = BTRFSIC_GENERATION_UNKNOWN;
 843                         l = btrfsic_block_link_lookup_or_add(
 844                                         state, &tmp_next_block_ctx,
 845                                         next_block, superblock_tmp,
 846                                         BTRFSIC_GENERATION_UNKNOWN);
 847                         btrfsic_release_block_ctx(&tmp_next_block_ctx);
 848                         if (NULL == l) {
 849                                 ret = -1;
 850                                 goto out;
 851                         }
 852                 }
 853         }
 854         if (state->print_mask & BTRFSIC_PRINT_MASK_INITIAL_ALL_TREES)
 855                 btrfsic_dump_tree_sub(state, superblock_tmp, 0);
 856
 857 out:
 858         put_page(page);
 859         return ret;
 860 }
 861
 862 static struct btrfsic_stack_frame *btrfsic_stack_frame_alloc(void)
 863 {
 864         struct btrfsic_stack_frame *sf;
 865
 866         sf = kzalloc(sizeof(*sf), GFP_NOFS);
 867         if (sf)
 868                 sf->magic = BTRFSIC_BLOCK_STACK_FRAME_MAGIC_NUMBER;
 869         return sf;
 870 }
 871
 872 static void btrfsic_stack_frame_free(struct btrfsic_stack_frame *sf)
 873 {
 874         BUG_ON(!(NULL == sf ||
 875                  BTRFSIC_BLOCK_STACK_FRAME_MAGIC_NUMBER == sf->magic));
 876         kfree(sf);
 877 }
 878
 879 static noinline_for_stack int btrfsic_process_metablock(
 880                 struct btrfsic_state *state,
 881                 struct btrfsic_block *const first_block,
 882                 struct btrfsic_block_data_ctx *const first_block_ctx,
 883                 int first_limit_nesting, int force_iodone_flag)
 884 {
 885         struct btrfsic_stack_frame initial_stack_frame = { 0 };
 886         struct btrfsic_stack_frame *sf;
 887         struct btrfsic_stack_frame *next_stack;
 888         struct btrfs_header *const first_hdr =
 889                 (struct btrfs_header *)first_block_ctx->datav[0];
 890
 891         BUG_ON(!first_hdr);
 892         sf = &initial_stack_frame;
 893         sf->error = 0;
 894         sf->i = -1;
 895         sf->limit_nesting = first_limit_nesting;
 896         sf->block = first_block;
 897         sf->block_ctx = first_block_ctx;
 898         sf->next_block = NULL;
 899         sf->hdr = first_hdr;
 900         sf->prev = NULL;
 901
 902 continue_with_new_stack_frame:
 903         sf->block->generation = btrfs_stack_header_generation(sf->hdr);
 904         if (0 == sf->hdr->level) {
 905                 struct btrfs_leaf *const leafhdr =
 906                     (struct btrfs_leaf *)sf->hdr;
 907
 908                 if (-1 == sf->i) {
 909                         sf->nr = btrfs_stack_header_nritems(&leafhdr->header);
 910
 911                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
 912                                 pr_info("leaf %llu items %d generation %llu owner %llu\n",
 913                                        sf->block_ctx->start, sf->nr,
 914                                        btrfs_stack_header_generation(
 915                                                &leafhdr->header),
 916                                        btrfs_stack_header_owner(
 917                                                &leafhdr->header));
 918                 }
 919
 920 continue_with_current_leaf_stack_frame:
 921                 if (0 == sf->num_copies || sf->mirror_num > sf->num_copies) {
 922                         sf->i++;
 923                         sf->num_copies = 0;
 924                 }
 925
 926                 if (sf->i < sf->nr) {
 927                         struct btrfs_item disk_item;
 928                         u32 disk_item_offset =
 929                                 (uintptr_t)(leafhdr->items + sf->i) -
 930                                 (uintptr_t)leafhdr;
 931                         struct btrfs_disk_key *disk_key;
 932                         u8 type;
 933                         u32 item_offset;
 934                         u32 item_size;
 935
 936                         if (disk_item_offset + sizeof(struct btrfs_item) >
 937                             sf->block_ctx->len) {
 938 leaf_item_out_of_bounce_error:
 939                                 pr_info(
 940                 "btrfsic: leaf item out of bounce at logical %llu, dev %pg\n",
 941                                        sf->block_ctx->start,
 942                                        sf->block_ctx->dev->bdev);
 943                                 goto one_stack_frame_backwards;
 944                         }
 945                         btrfsic_read_from_block_data(sf->block_ctx,
 946                                                      &disk_item,
 947                                                      disk_item_offset,
 948                                                      sizeof(struct btrfs_item));
 949                         item_offset = btrfs_stack_item_offset(&disk_item);
 950                         item_size = btrfs_stack_item_size(&disk_item);
 951                         disk_key = &disk_item.key;
 952                         type = btrfs_disk_key_type(disk_key);
 953
 954                         if (BTRFS_ROOT_ITEM_KEY == type) {
 955                                 struct btrfs_root_item root_item;
 956                                 u32 root_item_offset;
 957                                 u64 next_bytenr;
 958
 959                                 root_item_offset = item_offset +
 960                                         offsetof(struct btrfs_leaf, items);
 961                                 if (root_item_offset + item_size >
 962                                     sf->block_ctx->len)
 963                                         goto leaf_item_out_of_bounce_error;
 964                                 btrfsic_read_from_block_data(
 965                                         sf->block_ctx, &root_item,
 966                                         root_item_offset,
 967                                         item_size);
 968                                 next_bytenr = btrfs_root_bytenr(&root_item);
 969
 970                                 sf->error =
 971                                     btrfsic_create_link_to_next_block(
 972                                                 state,
 973                                                 sf->block,
 974                                                 sf->block_ctx,
 975                                                 next_bytenr,
 976                                                 sf->limit_nesting,
 977                                                 &sf->next_block_ctx,
 978                                                 &sf->next_block,
 979                                                 force_iodone_flag,
 980                                                 &sf->num_copies,
 981                                                 &sf->mirror_num,
 982                                                 disk_key,
 983                                                 btrfs_root_generation(
 984                                                 &root_item));
 985                                 if (sf->error)
 986                                         goto one_stack_frame_backwards;
 987
 988                                 if (NULL != sf->next_block) {
 989                                         struct btrfs_header *const next_hdr =
 990                                             (struct btrfs_header *)
 991                                             sf->next_block_ctx.datav[0];
 992
 993                                         next_stack =
 994                                             btrfsic_stack_frame_alloc();
 995                                         if (NULL == next_stack) {
 996                                                 sf->error = -1;
 997                                                 btrfsic_release_block_ctx(
 998                                                                 &sf->
 999                                                                 next_block_ctx);
1000                                                 goto one_stack_frame_backwards;
1001                                         }
1002
1003                                         next_stack->i = -1;
1004                                         next_stack->block = sf->next_block;
1005                                         next_stack->block_ctx =
1006                                             &sf->next_block_ctx;
1007                                         next_stack->next_block = NULL;
1008                                         next_stack->hdr = next_hdr;
1009                                         next_stack->limit_nesting =
1010                                             sf->limit_nesting - 1;
1011                                         next_stack->prev = sf;
1012                                         sf = next_stack;
1013                                         goto continue_with_new_stack_frame;
1014                                 }
1015                         } else if (BTRFS_EXTENT_DATA_KEY == type &&
1016                                    state->include_extent_data) {
1017                                 sf->error = btrfsic_handle_extent_data(
1018                                                 state,
1019                                                 sf->block,
1020                                                 sf->block_ctx,
1021                                                 item_offset,
1022                                                 force_iodone_flag);
1023                                 if (sf->error)
1024                                         goto one_stack_frame_backwards;
1025                         }
1026
1027                         goto continue_with_current_leaf_stack_frame;
1028                 }
1029         } else {
1030                 struct btrfs_node *const nodehdr = (struct btrfs_node *)sf->hdr;
1031
1032                 if (-1 == sf->i) {
1033                         sf->nr = btrfs_stack_header_nritems(&nodehdr->header);
1034
1035                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1036                                 pr_info("node %llu level %d items %d generation %llu owner %llu\n",
1037                                        sf->block_ctx->start,
1038                                        nodehdr->header.level, sf->nr,
1039                                        btrfs_stack_header_generation(
1040                                        &nodehdr->header),
1041                                        btrfs_stack_header_owner(
1042                                        &nodehdr->header));
1043                 }
1044
1045 continue_with_current_node_stack_frame:
1046                 if (0 == sf->num_copies || sf->mirror_num > sf->num_copies) {
1047                         sf->i++;
1048                         sf->num_copies = 0;
1049                 }
1050
1051                 if (sf->i < sf->nr) {
1052                         struct btrfs_key_ptr key_ptr;
1053                         u32 key_ptr_offset;
1054                         u64 next_bytenr;
1055
1056                         key_ptr_offset = (uintptr_t)(nodehdr->ptrs + sf->i) -
1057                                           (uintptr_t)nodehdr;
1058                         if (key_ptr_offset + sizeof(struct btrfs_key_ptr) >
1059                             sf->block_ctx->len) {
1060                                 pr_info(
1061                 "btrfsic: node item out of bounce at logical %llu, dev %pg\n",
1062                                        sf->block_ctx->start,
1063                                        sf->block_ctx->dev->bdev);
1064                                 goto one_stack_frame_backwards;
1065                         }
1066                         btrfsic_read_from_block_data(
1067                                 sf->block_ctx, &key_ptr, key_ptr_offset,
1068                                 sizeof(struct btrfs_key_ptr));
1069                         next_bytenr = btrfs_stack_key_blockptr(&key_ptr);
1070
1071                         sf->error = btrfsic_create_link_to_next_block(
1072                                         state,
1073                                         sf->block,
1074                                         sf->block_ctx,
1075                                         next_bytenr,
1076                                         sf->limit_nesting,
1077                                         &sf->next_block_ctx,
1078                                         &sf->next_block,
1079                                         force_iodone_flag,
1080                                         &sf->num_copies,
1081                                         &sf->mirror_num,
1082                                         &key_ptr.key,
1083                                         btrfs_stack_key_generation(&key_ptr));
1084                         if (sf->error)
1085                                 goto one_stack_frame_backwards;
1086
1087                         if (NULL != sf->next_block) {
1088                                 struct btrfs_header *const next_hdr =
1089                                     (struct btrfs_header *)
1090                                     sf->next_block_ctx.datav[0];
1091
1092                                 next_stack = btrfsic_stack_frame_alloc();
1093                                 if (NULL == next_stack) {
1094                                         sf->error = -1;
1095                                         goto one_stack_frame_backwards;
1096                                 }
1097
1098                                 next_stack->i = -1;
1099                                 next_stack->block = sf->next_block;
1100                                 next_stack->block_ctx = &sf->next_block_ctx;
1101                                 next_stack->next_block = NULL;
1102                                 next_stack->hdr = next_hdr;
1103                                 next_stack->limit_nesting =
1104                                     sf->limit_nesting - 1;
1105                                 next_stack->prev = sf;
1106                                 sf = next_stack;
1107                                 goto continue_with_new_stack_frame;
1108                         }
1109
1110                         goto continue_with_current_node_stack_frame;
1111                 }
1112         }
1113
1114 one_stack_frame_backwards:
1115         if (NULL != sf->prev) {
1116                 struct btrfsic_stack_frame *const prev = sf->prev;
1117
1118                 /* the one for the initial block is freed in the caller */
1119                 btrfsic_release_block_ctx(sf->block_ctx);
1120
1121                 if (sf->error) {
1122                         prev->error = sf->error;
1123                         btrfsic_stack_frame_free(sf);
1124                         sf = prev;
1125                         goto one_stack_frame_backwards;
1126                 }
1127
1128                 btrfsic_stack_frame_free(sf);
1129                 sf = prev;
1130                 goto continue_with_new_stack_frame;
1131         } else {
1132                 BUG_ON(&initial_stack_frame != sf);
1133         }
1134
1135         return sf->error;
1136 }
1137
1138 static void btrfsic_read_from_block_data(
1139         struct btrfsic_block_data_ctx *block_ctx,
1140         void *dstv, u32 offset, size_t len)
1141 {
1142         size_t cur;
1143         size_t pgoff;
1144         char *kaddr;
1145         char *dst = (char *)dstv;
1146         size_t start_offset = offset_in_page(block_ctx->start);
1147         unsigned long i = (start_offset + offset) >> PAGE_SHIFT;
1148
1149         WARN_ON(offset + len > block_ctx->len);
1150         pgoff = offset_in_page(start_offset + offset);
1151
1152         while (len > 0) {
1153                 cur = min(len, ((size_t)PAGE_SIZE - pgoff));
1154                 BUG_ON(i >= DIV_ROUND_UP(block_ctx->len, PAGE_SIZE));
1155                 kaddr = block_ctx->datav[i];
1156                 memcpy(dst, kaddr + pgoff, cur);
1157
1158                 dst += cur;
1159                 len -= cur;
1160                 pgoff = 0;
1161                 i++;
1162         }
1163 }
1164
1165 static int btrfsic_create_link_to_next_block(
1166                 struct btrfsic_state *state,
1167                 struct btrfsic_block *block,
1168                 struct btrfsic_block_data_ctx *block_ctx,
1169                 u64 next_bytenr,
1170                 int limit_nesting,
1171                 struct btrfsic_block_data_ctx *next_block_ctx,
1172                 struct btrfsic_block **next_blockp,
1173                 int force_iodone_flag,
1174                 int *num_copiesp, int *mirror_nump,
1175                 struct btrfs_disk_key *disk_key,
1176                 u64 parent_generation)
1177 {
1178         struct btrfs_fs_info *fs_info = state->fs_info;
1179         struct btrfsic_block *next_block = NULL;
1180         int ret;
1181         struct btrfsic_block_link *l;
1182         int did_alloc_block_link;
1183         int block_was_created;
1184
1185         *next_blockp = NULL;
1186         if (0 == *num_copiesp) {
1187                 *num_copiesp = btrfs_num_copies(fs_info, next_bytenr,
1188                                                 state->metablock_size);
1189                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
1190                         pr_info("num_copies(log_bytenr=%llu) = %d\n",
1191                                next_bytenr, *num_copiesp);
1192                 *mirror_nump = 1;
1193         }
1194
1195         if (*mirror_nump > *num_copiesp)
1196                 return 0;
1197
1198         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1199                 pr_info("btrfsic_create_link_to_next_block(mirror_num=%d)\n",
1200                        *mirror_nump);
1201         ret = btrfsic_map_block(state, next_bytenr,
1202                                 state->metablock_size,
1203                                 next_block_ctx, *mirror_nump);
1204         if (ret) {
1205                 pr_info("btrfsic: btrfsic_map_block(@%llu, mirror=%d) failed!\n",
1206                        next_bytenr, *mirror_nump);
1207                 btrfsic_release_block_ctx(next_block_ctx);
1208                 *next_blockp = NULL;
1209                 return -1;
1210         }
1211
1212         next_block = btrfsic_block_lookup_or_add(state,
1213                                                  next_block_ctx, "referenced ",
1214                                                  1, force_iodone_flag,
1215                                                  !force_iodone_flag,
1216                                                  *mirror_nump,
1217                                                  &block_was_created);
1218         if (NULL == next_block) {
1219                 btrfsic_release_block_ctx(next_block_ctx);
1220                 *next_blockp = NULL;
1221                 return -1;
1222         }
1223         if (block_was_created) {
1224                 l = NULL;
1225                 next_block->generation = BTRFSIC_GENERATION_UNKNOWN;
1226         } else {
1227                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE) {
1228                         if (next_block->logical_bytenr != next_bytenr &&
1229                             !(!next_block->is_metadata &&
1230                               0 == next_block->logical_bytenr))
1231                                 pr_info(
1232 "referenced block @%llu (%pg/%llu/%d) found in hash table, %c, bytenr mismatch (!= stored %llu)\n",
1233                                        next_bytenr, next_block_ctx->dev->bdev,
1234                                        next_block_ctx->dev_bytenr, *mirror_nump,
1235                                        btrfsic_get_block_type(state,
1236                                                               next_block),
1237                                        next_block->logical_bytenr);
1238                         else
1239                                 pr_info(
1240                 "referenced block @%llu (%pg/%llu/%d) found in hash table, %c\n",
1241                                        next_bytenr, next_block_ctx->dev->bdev,
1242                                        next_block_ctx->dev_bytenr, *mirror_nump,
1243                                        btrfsic_get_block_type(state,
1244                                                               next_block));
1245                 }
1246                 next_block->logical_bytenr = next_bytenr;
1247
1248                 next_block->mirror_num = *mirror_nump;
1249                 l = btrfsic_block_link_hashtable_lookup(
1250                                 next_block_ctx->dev->bdev,
1251                                 next_block_ctx->dev_bytenr,
1252                                 block_ctx->dev->bdev,
1253                                 block_ctx->dev_bytenr,
1254                                 &state->block_link_hashtable);
1255         }
1256
1257         next_block->disk_key = *disk_key;
1258         if (NULL == l) {
1259                 l = btrfsic_block_link_alloc();
1260                 if (NULL == l) {
1261                         btrfsic_release_block_ctx(next_block_ctx);
1262                         *next_blockp = NULL;
1263                         return -1;
1264                 }
1265
1266                 did_alloc_block_link = 1;
1267                 l->block_ref_to = next_block;
1268                 l->block_ref_from = block;
1269                 l->ref_cnt = 1;
1270                 l->parent_generation = parent_generation;
1271
1272                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1273                         btrfsic_print_add_link(state, l);
1274
1275                 list_add(&l->node_ref_to, &block->ref_to_list);
1276                 list_add(&l->node_ref_from, &next_block->ref_from_list);
1277
1278                 btrfsic_block_link_hashtable_add(l,
1279                                                  &state->block_link_hashtable);
1280         } else {
1281                 did_alloc_block_link = 0;
1282                 if (0 == limit_nesting) {
1283                         l->ref_cnt++;
1284                         l->parent_generation = parent_generation;
1285                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1286                                 btrfsic_print_add_link(state, l);
1287                 }
1288         }
1289
1290         if (limit_nesting > 0 && did_alloc_block_link) {
1291                 ret = btrfsic_read_block(state, next_block_ctx);
1292                 if (ret < (int)next_block_ctx->len) {
1293                         pr_info("btrfsic: read block @logical %llu failed!\n",
1294                                next_bytenr);
1295                         btrfsic_release_block_ctx(next_block_ctx);
1296                         *next_blockp = NULL;
1297                         return -1;
1298                 }
1299
1300                 *next_blockp = next_block;
1301         } else {
1302                 *next_blockp = NULL;
1303         }
1304         (*mirror_nump)++;
1305
1306         return 0;
1307 }
1308
1309 static int btrfsic_handle_extent_data(
1310                 struct btrfsic_state *state,
1311                 struct btrfsic_block *block,
1312                 struct btrfsic_block_data_ctx *block_ctx,
1313                 u32 item_offset, int force_iodone_flag)
1314 {
1315         struct btrfs_fs_info *fs_info = state->fs_info;
1316         struct btrfs_file_extent_item file_extent_item;
1317         u64 file_extent_item_offset;
1318         u64 next_bytenr;
1319         u64 num_bytes;
1320         u64 generation;
1321         struct btrfsic_block_link *l;
1322         int ret;
1323
1324         file_extent_item_offset = offsetof(struct btrfs_leaf, items) +
1325                                   item_offset;
1326         if (file_extent_item_offset +
1327             offsetof(struct btrfs_file_extent_item, disk_num_bytes) >
1328             block_ctx->len) {
1329                 pr_info("btrfsic: file item out of bounce at logical %llu, dev %pg\n",
1330                        block_ctx->start, block_ctx->dev->bdev);
1331                 return -1;
1332         }
1333
1334         btrfsic_read_from_block_data(block_ctx, &file_extent_item,
1335                 file_extent_item_offset,
1336                 offsetof(struct btrfs_file_extent_item, disk_num_bytes));
1337         if (BTRFS_FILE_EXTENT_REG != file_extent_item.type ||
1338             btrfs_stack_file_extent_disk_bytenr(&file_extent_item) == 0) {
1339                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERY_VERBOSE)
1340                         pr_info("extent_data: type %u, disk_bytenr = %llu\n",
1341                                file_extent_item.type,
1342                                btrfs_stack_file_extent_disk_bytenr(
1343                                &file_extent_item));
1344                 return 0;
1345         }
1346
1347         if (file_extent_item_offset + sizeof(struct btrfs_file_extent_item) >
1348             block_ctx->len) {
1349                 pr_info("btrfsic: file item out of bounce at logical %llu, dev %pg\n",
1350                        block_ctx->start, block_ctx->dev->bdev);
1351                 return -1;
1352         }
1353         btrfsic_read_from_block_data(block_ctx, &file_extent_item,
1354                                      file_extent_item_offset,
1355                                      sizeof(struct btrfs_file_extent_item));
1356         next_bytenr = btrfs_stack_file_extent_disk_bytenr(&file_extent_item);
1357         if (btrfs_stack_file_extent_compression(&file_extent_item) ==
1358             BTRFS_COMPRESS_NONE) {
1359                 next_bytenr += btrfs_stack_file_extent_offset(&file_extent_item);
1360                 num_bytes = btrfs_stack_file_extent_num_bytes(&file_extent_item);
1361         } else {
1362                 num_bytes = btrfs_stack_file_extent_disk_num_bytes(&file_extent_item);
1363         }
1364         generation = btrfs_stack_file_extent_generation(&file_extent_item);
1365
1366         if (state->print_mask & BTRFSIC_PRINT_MASK_VERY_VERBOSE)
1367                 pr_info("extent_data: type %u, disk_bytenr = %llu, offset = %llu, num_bytes = %llu\n",
1368                        file_extent_item.type,
1369                        btrfs_stack_file_extent_disk_bytenr(&file_extent_item),
1370                        btrfs_stack_file_extent_offset(&file_extent_item),
1371                        num_bytes);
1372         while (num_bytes > 0) {
1373                 u32 chunk_len;
1374                 int num_copies;
1375                 int mirror_num;
1376
1377                 if (num_bytes > state->datablock_size)
1378                         chunk_len = state->datablock_size;
1379                 else
1380                         chunk_len = num_bytes;
1381
1382                 num_copies = btrfs_num_copies(fs_info, next_bytenr,
1383                                               state->datablock_size);
1384                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
1385                         pr_info("num_copies(log_bytenr=%llu) = %d\n",
1386                                next_bytenr, num_copies);
1387                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
1388                         struct btrfsic_block_data_ctx next_block_ctx;
1389                         struct btrfsic_block *next_block;
1390                         int block_was_created;
1391
1392                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1393                                 pr_info("btrfsic_handle_extent_data(mirror_num=%d)\n",
1394                                         mirror_num);
1395                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERY_VERBOSE)
1396                                 pr_info("\tdisk_bytenr = %llu, num_bytes %u\n",
1397                                        next_bytenr, chunk_len);
1398                         ret = btrfsic_map_block(state, next_bytenr,
1399                                                 chunk_len, &next_block_ctx,
1400                                                 mirror_num);
1401                         if (ret) {
1402                                 pr_info("btrfsic: btrfsic_map_block(@%llu, mirror=%d) failed!\n",
1403                                        next_bytenr, mirror_num);
1404                                 return -1;
1405                         }
1406
1407                         next_block = btrfsic_block_lookup_or_add(
1408                                         state,
1409                                         &next_block_ctx,
1410                                         "referenced ",
1411                                         0,
1412                                         force_iodone_flag,
1413                                         !force_iodone_flag,
1414                                         mirror_num,
1415                                         &block_was_created);
1416                         if (NULL == next_block) {
1417                                 btrfsic_release_block_ctx(&next_block_ctx);
1418                                 return -1;
1419                         }
1420                         if (!block_was_created) {
1421                                 if ((state->print_mask &
1422                                      BTRFSIC_PRINT_MASK_VERBOSE) &&
1423                                     next_block->logical_bytenr != next_bytenr &&
1424                                     !(!next_block->is_metadata &&
1425                                       0 == next_block->logical_bytenr)) {
1426                                         pr_info(
1427 "referenced block @%llu (%pg/%llu/%d) found in hash table, D, bytenr mismatch (!= stored %llu)\n",
1428                                                next_bytenr,
1429                                                next_block_ctx.dev->bdev,
1430                                                next_block_ctx.dev_bytenr,
1431                                                mirror_num,
1432                                                next_block->logical_bytenr);
1433                                 }
1434                                 next_block->logical_bytenr = next_bytenr;
1435                                 next_block->mirror_num = mirror_num;
1436                         }
1437
1438                         l = btrfsic_block_link_lookup_or_add(state,
1439                                                              &next_block_ctx,
1440                                                              next_block, block,
1441                                                              generation);
1442                         btrfsic_release_block_ctx(&next_block_ctx);
1443                         if (NULL == l)
1444                                 return -1;
1445                 }
1446
1447                 next_bytenr += chunk_len;
1448                 num_bytes -= chunk_len;
1449         }
1450
1451         return 0;
1452 }
1453
1454 static int btrfsic_map_block(struct btrfsic_state *state, u64 bytenr, u32 len,
1455                              struct btrfsic_block_data_ctx *block_ctx_out,
1456                              int mirror_num)
1457 {
1458         struct btrfs_fs_info *fs_info = state->fs_info;
1459         int ret;
1460         u64 length;
1461         struct btrfs_io_context *multi = NULL;
1462         struct btrfs_device *device;
1463
1464         length = len;
1465         ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
1466                               bytenr, &length, &multi, mirror_num);
1467
1468         if (ret) {
1469                 block_ctx_out->start = 0;
1470                 block_ctx_out->dev_bytenr = 0;
1471                 block_ctx_out->len = 0;
1472                 block_ctx_out->dev = NULL;
1473                 block_ctx_out->datav = NULL;
1474                 block_ctx_out->pagev = NULL;
1475                 block_ctx_out->mem_to_free = NULL;
1476
1477                 return ret;
1478         }
1479
1480         device = multi->stripes[0].dev;
1481         if (test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state) ||
1482             !device->bdev || !device->name)
1483                 block_ctx_out->dev = NULL;
1484         else
1485                 block_ctx_out->dev = btrfsic_dev_state_lookup(
1486                                                         device->bdev->bd_dev);
1487         block_ctx_out->dev_bytenr = multi->stripes[0].physical;
1488         block_ctx_out->start = bytenr;
1489         block_ctx_out->len = len;
1490         block_ctx_out->datav = NULL;
1491         block_ctx_out->pagev = NULL;
1492         block_ctx_out->mem_to_free = NULL;
1493
1494         kfree(multi);
1495         if (NULL == block_ctx_out->dev) {
1496                 ret = -ENXIO;
1497                 pr_info("btrfsic: error, cannot lookup dev (#1)!\n");
1498         }
1499
1500         return ret;
1501 }
1502
1503 static void btrfsic_release_block_ctx(struct btrfsic_block_data_ctx *block_ctx)
1504 {
1505         if (block_ctx->mem_to_free) {
1506                 unsigned int num_pages;
1507
1508                 BUG_ON(!block_ctx->datav);
1509                 BUG_ON(!block_ctx->pagev);
1510                 num_pages = (block_ctx->len + (u64)PAGE_SIZE - 1) >>
1511                             PAGE_SHIFT;
1512                 /* Pages must be unmapped in reverse order */
1513                 while (num_pages > 0) {
1514                         num_pages--;
1515                         if (block_ctx->datav[num_pages])
1516                                 block_ctx->datav[num_pages] = NULL;
1517                         if (block_ctx->pagev[num_pages]) {
1518                                 __free_page(block_ctx->pagev[num_pages]);
1519                                 block_ctx->pagev[num_pages] = NULL;
1520                         }
1521                 }
1522
1523                 kfree(block_ctx->mem_to_free);
1524                 block_ctx->mem_to_free = NULL;
1525                 block_ctx->pagev = NULL;
1526                 block_ctx->datav = NULL;
1527         }
1528 }
1529
1530 static int btrfsic_read_block(struct btrfsic_state *state,
1531                               struct btrfsic_block_data_ctx *block_ctx)
1532 {
1533         unsigned int num_pages;
1534         unsigned int i;
1535         size_t size;
1536         u64 dev_bytenr;
1537         int ret;
1538
1539         BUG_ON(block_ctx->datav);
1540         BUG_ON(block_ctx->pagev);
1541         BUG_ON(block_ctx->mem_to_free);
1542         if (!PAGE_ALIGNED(block_ctx->dev_bytenr)) {
1543                 pr_info("btrfsic: read_block() with unaligned bytenr %llu\n",
1544                        block_ctx->dev_bytenr);
1545                 return -1;
1546         }
1547
1548         num_pages = (block_ctx->len + (u64)PAGE_SIZE - 1) >>
1549                     PAGE_SHIFT;
1550         size = sizeof(*block_ctx->datav) + sizeof(*block_ctx->pagev);
1551         block_ctx->mem_to_free = kcalloc(num_pages, size, GFP_NOFS);
1552         if (!block_ctx->mem_to_free)
1553                 return -ENOMEM;
1554         block_ctx->datav = block_ctx->mem_to_free;
1555         block_ctx->pagev = (struct page **)(block_ctx->datav + num_pages);
1556         ret = btrfs_alloc_page_array(num_pages, block_ctx->pagev);
1557         if (ret)
1558                 return ret;
1559
1560         dev_bytenr = block_ctx->dev_bytenr;
1561         for (i = 0; i < num_pages;) {
1562                 struct bio *bio;
1563                 unsigned int j;
1564
1565                 bio = bio_alloc(block_ctx->dev->bdev, num_pages - i,
1566                                 REQ_OP_READ, GFP_NOFS);
1567                 bio->bi_iter.bi_sector = dev_bytenr >> 9;
1568
1569                 for (j = i; j < num_pages; j++) {
1570                         ret = bio_add_page(bio, block_ctx->pagev[j],
1571                                            PAGE_SIZE, 0);
1572                         if (PAGE_SIZE != ret)
1573                                 break;
1574                 }
1575                 if (j == i) {
1576                         pr_info("btrfsic: error, failed to add a single page!\n");
1577                         return -1;
1578                 }
1579                 if (submit_bio_wait(bio)) {
1580                         pr_info("btrfsic: read error at logical %llu dev %pg!\n",
1581                                block_ctx->start, block_ctx->dev->bdev);
1582                         bio_put(bio);
1583                         return -1;
1584                 }
1585                 bio_put(bio);
1586                 dev_bytenr += (j - i) * PAGE_SIZE;
1587                 i = j;
1588         }
1589         for (i = 0; i < num_pages; i++)
1590                 block_ctx->datav[i] = page_address(block_ctx->pagev[i]);
1591
1592         return block_ctx->len;
1593 }
1594
1595 static void btrfsic_dump_database(struct btrfsic_state *state)
1596 {
1597         const struct btrfsic_block *b_all;
1598
1599         BUG_ON(NULL == state);
1600
1601         pr_info("all_blocks_list:\n");
1602         list_for_each_entry(b_all, &state->all_blocks_list, all_blocks_node) {
1603                 const struct btrfsic_block_link *l;
1604
1605                 pr_info("%c-block @%llu (%pg/%llu/%d)\n",
1606                        btrfsic_get_block_type(state, b_all),
1607                        b_all->logical_bytenr, b_all->dev_state->bdev,
1608                        b_all->dev_bytenr, b_all->mirror_num);
1609
1610                 list_for_each_entry(l, &b_all->ref_to_list, node_ref_to) {
1611                         pr_info(
1612                 " %c @%llu (%pg/%llu/%d) refers %u* to %c @%llu (%pg/%llu/%d)\n",
1613                                btrfsic_get_block_type(state, b_all),
1614                                b_all->logical_bytenr, b_all->dev_state->bdev,
1615                                b_all->dev_bytenr, b_all->mirror_num,
1616                                l->ref_cnt,
1617                                btrfsic_get_block_type(state, l->block_ref_to),
1618                                l->block_ref_to->logical_bytenr,
1619                                l->block_ref_to->dev_state->bdev,
1620                                l->block_ref_to->dev_bytenr,
1621                                l->block_ref_to->mirror_num);
1622                 }
1623
1624                 list_for_each_entry(l, &b_all->ref_from_list, node_ref_from) {
1625                         pr_info(
1626                 " %c @%llu (%pg/%llu/%d) is ref %u* from %c @%llu (%pg/%llu/%d)\n",
1627                                btrfsic_get_block_type(state, b_all),
1628                                b_all->logical_bytenr, b_all->dev_state->bdev,
1629                                b_all->dev_bytenr, b_all->mirror_num,
1630                                l->ref_cnt,
1631                                btrfsic_get_block_type(state, l->block_ref_from),
1632                                l->block_ref_from->logical_bytenr,
1633                                l->block_ref_from->dev_state->bdev,
1634                                l->block_ref_from->dev_bytenr,
1635                                l->block_ref_from->mirror_num);
1636                 }
1637
1638                 pr_info("\n");
1639         }
1640 }
1641
1642 /*
1643  * Test whether the disk block contains a tree block (leaf or node)
1644  * (note that this test fails for the super block)
1645  */
1646 static noinline_for_stack int btrfsic_test_for_metadata(
1647                 struct btrfsic_state *state,
1648                 char **datav, unsigned int num_pages)
1649 {
1650         struct btrfs_fs_info *fs_info = state->fs_info;
1651         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1652         struct btrfs_header *h;
1653         u8 csum[BTRFS_CSUM_SIZE];
1654         unsigned int i;
1655
1656         if (num_pages * PAGE_SIZE < state->metablock_size)
1657                 return 1; /* not metadata */
1658         num_pages = state->metablock_size >> PAGE_SHIFT;
1659         h = (struct btrfs_header *)datav[0];
1660
1661         if (memcmp(h->fsid, fs_info->fs_devices->fsid, BTRFS_FSID_SIZE))
1662                 return 1;
1663
1664         shash->tfm = fs_info->csum_shash;
1665         crypto_shash_init(shash);
1666
1667         for (i = 0; i < num_pages; i++) {
1668                 u8 *data = i ? datav[i] : (datav[i] + BTRFS_CSUM_SIZE);
1669                 size_t sublen = i ? PAGE_SIZE :
1670                                     (PAGE_SIZE - BTRFS_CSUM_SIZE);
1671
1672                 crypto_shash_update(shash, data, sublen);
1673         }
1674         crypto_shash_final(shash, csum);
1675         if (memcmp(csum, h->csum, fs_info->csum_size))
1676                 return 1;
1677
1678         return 0; /* is metadata */
1679 }
1680
1681 static void btrfsic_process_written_block(struct btrfsic_dev_state *dev_state,
1682                                           u64 dev_bytenr, char **mapped_datav,
1683                                           unsigned int num_pages,
1684                                           struct bio *bio, int *bio_is_patched,
1685                                           blk_opf_t submit_bio_bh_rw)
1686 {
1687         int is_metadata;
1688         struct btrfsic_block *block;
1689         struct btrfsic_block_data_ctx block_ctx;
1690         int ret;
1691         struct btrfsic_state *state = dev_state->state;
1692         struct block_device *bdev = dev_state->bdev;
1693         unsigned int processed_len;
1694
1695         if (NULL != bio_is_patched)
1696                 *bio_is_patched = 0;
1697
1698 again:
1699         if (num_pages == 0)
1700                 return;
1701
1702         processed_len = 0;
1703         is_metadata = (0 == btrfsic_test_for_metadata(state, mapped_datav,
1704                                                       num_pages));
1705
1706         block = btrfsic_block_hashtable_lookup(bdev, dev_bytenr,
1707                                                &state->block_hashtable);
1708         if (NULL != block) {
1709                 u64 bytenr = 0;
1710                 struct btrfsic_block_link *l, *tmp;
1711
1712                 if (block->is_superblock) {
1713                         bytenr = btrfs_super_bytenr((struct btrfs_super_block *)
1714                                                     mapped_datav[0]);
1715                         if (num_pages * PAGE_SIZE <
1716                             BTRFS_SUPER_INFO_SIZE) {
1717                                 pr_info("btrfsic: cannot work with too short bios!\n");
1718                                 return;
1719                         }
1720                         is_metadata = 1;
1721                         BUG_ON(!PAGE_ALIGNED(BTRFS_SUPER_INFO_SIZE));
1722                         processed_len = BTRFS_SUPER_INFO_SIZE;
1723                         if (state->print_mask &
1724                             BTRFSIC_PRINT_MASK_TREE_BEFORE_SB_WRITE) {
1725                                 pr_info("[before new superblock is written]:\n");
1726                                 btrfsic_dump_tree_sub(state, block, 0);
1727                         }
1728                 }
1729                 if (is_metadata) {
1730                         if (!block->is_superblock) {
1731                                 if (num_pages * PAGE_SIZE <
1732                                     state->metablock_size) {
1733                                         pr_info("btrfsic: cannot work with too short bios!\n");
1734                                         return;
1735                                 }
1736                                 processed_len = state->metablock_size;
1737                                 bytenr = btrfs_stack_header_bytenr(
1738                                                 (struct btrfs_header *)
1739                                                 mapped_datav[0]);
1740                                 btrfsic_cmp_log_and_dev_bytenr(state, bytenr,
1741                                                                dev_state,
1742                                                                dev_bytenr);
1743                         }
1744                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE) {
1745                                 if (block->logical_bytenr != bytenr &&
1746                                     !(!block->is_metadata &&
1747                                       block->logical_bytenr == 0))
1748                                         pr_info(
1749 "written block @%llu (%pg/%llu/%d) found in hash table, %c, bytenr mismatch (!= stored %llu)\n",
1750                                                bytenr, dev_state->bdev,
1751                                                dev_bytenr,
1752                                                block->mirror_num,
1753                                                btrfsic_get_block_type(state,
1754                                                                       block),
1755                                                block->logical_bytenr);
1756                                 else
1757                                         pr_info(
1758                 "written block @%llu (%pg/%llu/%d) found in hash table, %c\n",
1759                                                bytenr, dev_state->bdev,
1760                                                dev_bytenr, block->mirror_num,
1761                                                btrfsic_get_block_type(state,
1762                                                                       block));
1763                         }
1764                         block->logical_bytenr = bytenr;
1765                 } else {
1766                         if (num_pages * PAGE_SIZE <
1767                             state->datablock_size) {
1768                                 pr_info("btrfsic: cannot work with too short bios!\n");
1769                                 return;
1770                         }
1771                         processed_len = state->datablock_size;
1772                         bytenr = block->logical_bytenr;
1773                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1774                                 pr_info(
1775                 "written block @%llu (%pg/%llu/%d) found in hash table, %c\n",
1776                                        bytenr, dev_state->bdev, dev_bytenr,
1777                                        block->mirror_num,
1778                                        btrfsic_get_block_type(state, block));
1779                 }
1780
1781                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1782                         pr_info("ref_to_list: %cE, ref_from_list: %cE\n",
1783                                list_empty(&block->ref_to_list) ? ' ' : '!',
1784                                list_empty(&block->ref_from_list) ? ' ' : '!');
1785                 if (btrfsic_is_block_ref_by_superblock(state, block, 0)) {
1786                         pr_info(
1787 "btrfs: attempt to overwrite %c-block @%llu (%pg/%llu/%d), old(gen=%llu, objectid=%llu, type=%d, offset=%llu), new(gen=%llu), which is referenced by most recent superblock (superblockgen=%llu)!\n",
1788                                btrfsic_get_block_type(state, block), bytenr,
1789                                dev_state->bdev, dev_bytenr, block->mirror_num,
1790                                block->generation,
1791                                btrfs_disk_key_objectid(&block->disk_key),
1792                                block->disk_key.type,
1793                                btrfs_disk_key_offset(&block->disk_key),
1794                                btrfs_stack_header_generation(
1795                                        (struct btrfs_header *) mapped_datav[0]),
1796                                state->max_superblock_generation);
1797                         btrfsic_dump_tree(state);
1798                 }
1799
1800                 if (!block->is_iodone && !block->never_written) {
1801                         pr_info(
1802 "btrfs: attempt to overwrite %c-block @%llu (%pg/%llu/%d), oldgen=%llu, newgen=%llu, which is not yet iodone!\n",
1803                                btrfsic_get_block_type(state, block), bytenr,
1804                                dev_state->bdev, dev_bytenr, block->mirror_num,
1805                                block->generation,
1806                                btrfs_stack_header_generation(
1807                                        (struct btrfs_header *)
1808                                        mapped_datav[0]));
1809                         /* it would not be safe to go on */
1810                         btrfsic_dump_tree(state);
1811                         goto continue_loop;
1812                 }
1813
1814                 /*
1815                  * Clear all references of this block. Do not free
1816                  * the block itself even if is not referenced anymore
1817                  * because it still carries valuable information
1818                  * like whether it was ever written and IO completed.
1819                  */
1820                 list_for_each_entry_safe(l, tmp, &block->ref_to_list,
1821                                          node_ref_to) {
1822                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1823                                 btrfsic_print_rem_link(state, l);
1824                         l->ref_cnt--;
1825                         if (0 == l->ref_cnt) {
1826                                 list_del(&l->node_ref_to);
1827                                 list_del(&l->node_ref_from);
1828                                 btrfsic_block_link_hashtable_remove(l);
1829                                 btrfsic_block_link_free(l);
1830                         }
1831                 }
1832
1833                 block_ctx.dev = dev_state;
1834                 block_ctx.dev_bytenr = dev_bytenr;
1835                 block_ctx.start = bytenr;
1836                 block_ctx.len = processed_len;
1837                 block_ctx.pagev = NULL;
1838                 block_ctx.mem_to_free = NULL;
1839                 block_ctx.datav = mapped_datav;
1840
1841                 if (is_metadata || state->include_extent_data) {
1842                         block->never_written = 0;
1843                         block->iodone_w_error = 0;
1844                         if (NULL != bio) {
1845                                 block->is_iodone = 0;
1846                                 BUG_ON(NULL == bio_is_patched);
1847                                 if (!*bio_is_patched) {
1848                                         block->orig_bio_private =
1849                                             bio->bi_private;
1850                                         block->orig_bio_end_io =
1851                                             bio->bi_end_io;
1852                                         block->next_in_same_bio = NULL;
1853                                         bio->bi_private = block;
1854                                         bio->bi_end_io = btrfsic_bio_end_io;
1855                                         *bio_is_patched = 1;
1856                                 } else {
1857                                         struct btrfsic_block *chained_block =
1858                                             (struct btrfsic_block *)
1859                                             bio->bi_private;
1860
1861                                         BUG_ON(NULL == chained_block);
1862                                         block->orig_bio_private =
1863                                             chained_block->orig_bio_private;
1864                                         block->orig_bio_end_io =
1865                                             chained_block->orig_bio_end_io;
1866                                         block->next_in_same_bio = chained_block;
1867                                         bio->bi_private = block;
1868                                 }
1869                         } else {
1870                                 block->is_iodone = 1;
1871                                 block->orig_bio_private = NULL;
1872                                 block->orig_bio_end_io = NULL;
1873                                 block->next_in_same_bio = NULL;
1874                         }
1875                 }
1876
1877                 block->flush_gen = dev_state->last_flush_gen + 1;
1878                 block->submit_bio_bh_rw = submit_bio_bh_rw;
1879                 if (is_metadata) {
1880                         block->logical_bytenr = bytenr;
1881                         block->is_metadata = 1;
1882                         if (block->is_superblock) {
1883                                 BUG_ON(PAGE_SIZE !=
1884                                        BTRFS_SUPER_INFO_SIZE);
1885                                 ret = btrfsic_process_written_superblock(
1886                                                 state,
1887                                                 block,
1888                                                 (struct btrfs_super_block *)
1889                                                 mapped_datav[0]);
1890                                 if (state->print_mask &
1891                                     BTRFSIC_PRINT_MASK_TREE_AFTER_SB_WRITE) {
1892                                         pr_info("[after new superblock is written]:\n");
1893                                         btrfsic_dump_tree_sub(state, block, 0);
1894                                 }
1895                         } else {
1896                                 block->mirror_num = 0;  /* unknown */
1897                                 ret = btrfsic_process_metablock(
1898                                                 state,
1899                                                 block,
1900                                                 &block_ctx,
1901                                                 0, 0);
1902                         }
1903                         if (ret)
1904                                 pr_info("btrfsic: btrfsic_process_metablock(root @%llu) failed!\n",
1905                                        dev_bytenr);
1906                 } else {
1907                         block->is_metadata = 0;
1908                         block->mirror_num = 0;  /* unknown */
1909                         block->generation = BTRFSIC_GENERATION_UNKNOWN;
1910                         if (!state->include_extent_data
1911                             && list_empty(&block->ref_from_list)) {
1912                                 /*
1913                                  * disk block is overwritten with extent
1914                                  * data (not meta data) and we are configured
1915                                  * to not include extent data: take the
1916                                  * chance and free the block's memory
1917                                  */
1918                                 btrfsic_block_hashtable_remove(block);
1919                                 list_del(&block->all_blocks_node);
1920                                 btrfsic_block_free(block);
1921                         }
1922                 }
1923                 btrfsic_release_block_ctx(&block_ctx);
1924         } else {
1925                 /* block has not been found in hash table */
1926                 u64 bytenr;
1927
1928                 if (!is_metadata) {
1929                         processed_len = state->datablock_size;
1930                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1931                                 pr_info(
1932                         "written block (%pg/%llu/?) !found in hash table, D\n",
1933                                        dev_state->bdev, dev_bytenr);
1934                         if (!state->include_extent_data) {
1935                                 /* ignore that written D block */
1936                                 goto continue_loop;
1937                         }
1938
1939                         /* this is getting ugly for the
1940                          * include_extent_data case... */
1941                         bytenr = 0;     /* unknown */
1942                 } else {
1943                         processed_len = state->metablock_size;
1944                         bytenr = btrfs_stack_header_bytenr(
1945                                         (struct btrfs_header *)
1946                                         mapped_datav[0]);
1947                         btrfsic_cmp_log_and_dev_bytenr(state, bytenr, dev_state,
1948                                                        dev_bytenr);
1949                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1950                                 pr_info(
1951                         "written block @%llu (%pg/%llu/?) !found in hash table, M\n",
1952                                        bytenr, dev_state->bdev, dev_bytenr);
1953                 }
1954
1955                 block_ctx.dev = dev_state;
1956                 block_ctx.dev_bytenr = dev_bytenr;
1957                 block_ctx.start = bytenr;
1958                 block_ctx.len = processed_len;
1959                 block_ctx.pagev = NULL;
1960                 block_ctx.mem_to_free = NULL;
1961                 block_ctx.datav = mapped_datav;
1962
1963                 block = btrfsic_block_alloc();
1964                 if (NULL == block) {
1965                         btrfsic_release_block_ctx(&block_ctx);
1966                         goto continue_loop;
1967                 }
1968                 block->dev_state = dev_state;
1969                 block->dev_bytenr = dev_bytenr;
1970                 block->logical_bytenr = bytenr;
1971                 block->is_metadata = is_metadata;
1972                 block->never_written = 0;
1973                 block->iodone_w_error = 0;
1974                 block->mirror_num = 0;  /* unknown */
1975                 block->flush_gen = dev_state->last_flush_gen + 1;
1976                 block->submit_bio_bh_rw = submit_bio_bh_rw;
1977                 if (NULL != bio) {
1978                         block->is_iodone = 0;
1979                         BUG_ON(NULL == bio_is_patched);
1980                         if (!*bio_is_patched) {
1981                                 block->orig_bio_private = bio->bi_private;
1982                                 block->orig_bio_end_io = bio->bi_end_io;
1983                                 block->next_in_same_bio = NULL;
1984                                 bio->bi_private = block;
1985                                 bio->bi_end_io = btrfsic_bio_end_io;
1986                                 *bio_is_patched = 1;
1987                         } else {
1988                                 struct btrfsic_block *chained_block =
1989                                     (struct btrfsic_block *)
1990                                     bio->bi_private;
1991
1992                                 BUG_ON(NULL == chained_block);
1993                                 block->orig_bio_private =
1994                                     chained_block->orig_bio_private;
1995                                 block->orig_bio_end_io =
1996                                     chained_block->orig_bio_end_io;
1997                                 block->next_in_same_bio = chained_block;
1998                                 bio->bi_private = block;
1999                         }
2000                 } else {
2001                         block->is_iodone = 1;
2002                         block->orig_bio_private = NULL;
2003                         block->orig_bio_end_io = NULL;
2004                         block->next_in_same_bio = NULL;
2005                 }
2006                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2007                         pr_info("new written %c-block @%llu (%pg/%llu/%d)\n",
2008                                is_metadata ? 'M' : 'D',
2009                                block->logical_bytenr, block->dev_state->bdev,
2010                                block->dev_bytenr, block->mirror_num);
2011                 list_add(&block->all_blocks_node, &state->all_blocks_list);
2012                 btrfsic_block_hashtable_add(block, &state->block_hashtable);
2013
2014                 if (is_metadata) {
2015                         ret = btrfsic_process_metablock(state, block,
2016                                                         &block_ctx, 0, 0);
2017                         if (ret)
2018                                 pr_info("btrfsic: process_metablock(root @%llu) failed!\n",
2019                                        dev_bytenr);
2020                 }
2021                 btrfsic_release_block_ctx(&block_ctx);
2022         }
2023
2024 continue_loop:
2025         BUG_ON(!processed_len);
2026         dev_bytenr += processed_len;
2027         mapped_datav += processed_len >> PAGE_SHIFT;
2028         num_pages -= processed_len >> PAGE_SHIFT;
2029         goto again;
2030 }
2031
2032 static void btrfsic_bio_end_io(struct bio *bp)
2033 {
2034         struct btrfsic_block *block = bp->bi_private;
2035         int iodone_w_error;
2036
2037         /* mutex is not held! This is not save if IO is not yet completed
2038          * on umount */
2039         iodone_w_error = 0;
2040         if (bp->bi_status)
2041                 iodone_w_error = 1;
2042
2043         BUG_ON(NULL == block);
2044         bp->bi_private = block->orig_bio_private;
2045         bp->bi_end_io = block->orig_bio_end_io;
2046
2047         do {
2048                 struct btrfsic_block *next_block;
2049                 struct btrfsic_dev_state *const dev_state = block->dev_state;
2050
2051                 if ((dev_state->state->print_mask &
2052                      BTRFSIC_PRINT_MASK_END_IO_BIO_BH))
2053                         pr_info("bio_end_io(err=%d) for %c @%llu (%pg/%llu/%d)\n",
2054                                bp->bi_status,
2055                                btrfsic_get_block_type(dev_state->state, block),
2056                                block->logical_bytenr, dev_state->bdev,
2057                                block->dev_bytenr, block->mirror_num);
2058                 next_block = block->next_in_same_bio;
2059                 block->iodone_w_error = iodone_w_error;
2060                 if (block->submit_bio_bh_rw & REQ_PREFLUSH) {
2061                         dev_state->last_flush_gen++;
2062                         if ((dev_state->state->print_mask &
2063                              BTRFSIC_PRINT_MASK_END_IO_BIO_BH))
2064                                 pr_info("bio_end_io() new %pg flush_gen=%llu\n",
2065                                        dev_state->bdev,
2066                                        dev_state->last_flush_gen);
2067                 }
2068                 if (block->submit_bio_bh_rw & REQ_FUA)
2069                         block->flush_gen = 0; /* FUA completed means block is
2070                                                * on disk */
2071                 block->is_iodone = 1; /* for FLUSH, this releases the block */
2072                 block = next_block;
2073         } while (NULL != block);
2074
2075         bp->bi_end_io(bp);
2076 }
2077
2078 static int btrfsic_process_written_superblock(
2079                 struct btrfsic_state *state,
2080                 struct btrfsic_block *const superblock,
2081                 struct btrfs_super_block *const super_hdr)
2082 {
2083         struct btrfs_fs_info *fs_info = state->fs_info;
2084         int pass;
2085
2086         superblock->generation = btrfs_super_generation(super_hdr);
2087         if (!(superblock->generation > state->max_superblock_generation ||
2088               0 == state->max_superblock_generation)) {
2089                 if (state->print_mask & BTRFSIC_PRINT_MASK_SUPERBLOCK_WRITE)
2090                         pr_info(
2091         "btrfsic: superblock @%llu (%pg/%llu/%d) with old gen %llu <= %llu\n",
2092                                superblock->logical_bytenr,
2093                                superblock->dev_state->bdev,
2094                                superblock->dev_bytenr, superblock->mirror_num,
2095                                btrfs_super_generation(super_hdr),
2096                                state->max_superblock_generation);
2097         } else {
2098                 if (state->print_mask & BTRFSIC_PRINT_MASK_SUPERBLOCK_WRITE)
2099                         pr_info(
2100         "btrfsic: got new superblock @%llu (%pg/%llu/%d) with new gen %llu > %llu\n",
2101                                superblock->logical_bytenr,
2102                                superblock->dev_state->bdev,
2103                                superblock->dev_bytenr, superblock->mirror_num,
2104                                btrfs_super_generation(super_hdr),
2105                                state->max_superblock_generation);
2106
2107                 state->max_superblock_generation =
2108                     btrfs_super_generation(super_hdr);
2109                 state->latest_superblock = superblock;
2110         }
2111
2112         for (pass = 0; pass < 3; pass++) {
2113                 int ret;
2114                 u64 next_bytenr;
2115                 struct btrfsic_block *next_block;
2116                 struct btrfsic_block_data_ctx tmp_next_block_ctx;
2117                 struct btrfsic_block_link *l;
2118                 int num_copies;
2119                 int mirror_num;
2120                 const char *additional_string = NULL;
2121                 struct btrfs_disk_key tmp_disk_key = {0};
2122
2123                 btrfs_set_disk_key_objectid(&tmp_disk_key,
2124                                             BTRFS_ROOT_ITEM_KEY);
2125                 btrfs_set_disk_key_objectid(&tmp_disk_key, 0);
2126
2127                 switch (pass) {
2128                 case 0:
2129                         btrfs_set_disk_key_objectid(&tmp_disk_key,
2130                                                     BTRFS_ROOT_TREE_OBJECTID);
2131                         additional_string = "root ";
2132                         next_bytenr = btrfs_super_root(super_hdr);
2133                         if (state->print_mask &
2134                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
2135                                 pr_info("root@%llu\n", next_bytenr);
2136                         break;
2137                 case 1:
2138                         btrfs_set_disk_key_objectid(&tmp_disk_key,
2139                                                     BTRFS_CHUNK_TREE_OBJECTID);
2140                         additional_string = "chunk ";
2141                         next_bytenr = btrfs_super_chunk_root(super_hdr);
2142                         if (state->print_mask &
2143                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
2144                                 pr_info("chunk@%llu\n", next_bytenr);
2145                         break;
2146                 case 2:
2147                         btrfs_set_disk_key_objectid(&tmp_disk_key,
2148                                                     BTRFS_TREE_LOG_OBJECTID);
2149                         additional_string = "log ";
2150                         next_bytenr = btrfs_super_log_root(super_hdr);
2151                         if (0 == next_bytenr)
2152                                 continue;
2153                         if (state->print_mask &
2154                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
2155                                 pr_info("log@%llu\n", next_bytenr);
2156                         break;
2157                 }
2158
2159                 num_copies = btrfs_num_copies(fs_info, next_bytenr,
2160                                               BTRFS_SUPER_INFO_SIZE);
2161                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
2162                         pr_info("num_copies(log_bytenr=%llu) = %d\n",
2163                                next_bytenr, num_copies);
2164                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
2165                         int was_created;
2166
2167                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2168                                 pr_info("btrfsic_process_written_superblock(mirror_num=%d)\n", mirror_num);
2169                         ret = btrfsic_map_block(state, next_bytenr,
2170                                                 BTRFS_SUPER_INFO_SIZE,
2171                                                 &tmp_next_block_ctx,
2172                                                 mirror_num);
2173                         if (ret) {
2174                                 pr_info("btrfsic: btrfsic_map_block(@%llu, mirror=%d) failed!\n",
2175                                        next_bytenr, mirror_num);
2176                                 return -1;
2177                         }
2178
2179                         next_block = btrfsic_block_lookup_or_add(
2180                                         state,
2181                                         &tmp_next_block_ctx,
2182                                         additional_string,
2183                                         1, 0, 1,
2184                                         mirror_num,
2185                                         &was_created);
2186                         if (NULL == next_block) {
2187                                 btrfsic_release_block_ctx(&tmp_next_block_ctx);
2188                                 return -1;
2189                         }
2190
2191                         next_block->disk_key = tmp_disk_key;
2192                         if (was_created)
2193                                 next_block->generation =
2194                                     BTRFSIC_GENERATION_UNKNOWN;
2195                         l = btrfsic_block_link_lookup_or_add(
2196                                         state,
2197                                         &tmp_next_block_ctx,
2198                                         next_block,
2199                                         superblock,
2200                                         BTRFSIC_GENERATION_UNKNOWN);
2201                         btrfsic_release_block_ctx(&tmp_next_block_ctx);
2202                         if (NULL == l)
2203                                 return -1;
2204                 }
2205         }
2206
2207         if (WARN_ON(-1 == btrfsic_check_all_ref_blocks(state, superblock, 0)))
2208                 btrfsic_dump_tree(state);
2209
2210         return 0;
2211 }
2212
2213 static int btrfsic_check_all_ref_blocks(struct btrfsic_state *state,
2214                                         struct btrfsic_block *const block,
2215                                         int recursion_level)
2216 {
2217         const struct btrfsic_block_link *l;
2218         int ret = 0;
2219
2220         if (recursion_level >= 3 + BTRFS_MAX_LEVEL) {
2221                 /*
2222                  * Note that this situation can happen and does not
2223                  * indicate an error in regular cases. It happens
2224                  * when disk blocks are freed and later reused.
2225                  * The check-integrity module is not aware of any
2226                  * block free operations, it just recognizes block
2227                  * write operations. Therefore it keeps the linkage
2228                  * information for a block until a block is
2229                  * rewritten. This can temporarily cause incorrect
2230                  * and even circular linkage information. This
2231                  * causes no harm unless such blocks are referenced
2232                  * by the most recent super block.
2233                  */
2234                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2235                         pr_info("btrfsic: abort cyclic linkage (case 1).\n");
2236
2237                 return ret;
2238         }
2239
2240         /*
2241          * This algorithm is recursive because the amount of used stack
2242          * space is very small and the max recursion depth is limited.
2243          */
2244         list_for_each_entry(l, &block->ref_to_list, node_ref_to) {
2245                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2246                         pr_info(
2247                 "rl=%d, %c @%llu (%pg/%llu/%d) %u* refers to %c @%llu (%pg/%llu/%d)\n",
2248                                recursion_level,
2249                                btrfsic_get_block_type(state, block),
2250                                block->logical_bytenr, block->dev_state->bdev,
2251                                block->dev_bytenr, block->mirror_num,
2252                                l->ref_cnt,
2253                                btrfsic_get_block_type(state, l->block_ref_to),
2254                                l->block_ref_to->logical_bytenr,
2255                                l->block_ref_to->dev_state->bdev,
2256                                l->block_ref_to->dev_bytenr,
2257                                l->block_ref_to->mirror_num);
2258                 if (l->block_ref_to->never_written) {
2259                         pr_info(
2260 "btrfs: attempt to write superblock which references block %c @%llu (%pg/%llu/%d) which is never written!\n",
2261                                btrfsic_get_block_type(state, l->block_ref_to),
2262                                l->block_ref_to->logical_bytenr,
2263                                l->block_ref_to->dev_state->bdev,
2264                                l->block_ref_to->dev_bytenr,
2265                                l->block_ref_to->mirror_num);
2266                         ret = -1;
2267                 } else if (!l->block_ref_to->is_iodone) {
2268                         pr_info(
2269 "btrfs: attempt to write superblock which references block %c @%llu (%pg/%llu/%d) which is not yet iodone!\n",
2270                                btrfsic_get_block_type(state, l->block_ref_to),
2271                                l->block_ref_to->logical_bytenr,
2272                                l->block_ref_to->dev_state->bdev,
2273                                l->block_ref_to->dev_bytenr,
2274                                l->block_ref_to->mirror_num);
2275                         ret = -1;
2276                 } else if (l->block_ref_to->iodone_w_error) {
2277                         pr_info(
2278 "btrfs: attempt to write superblock which references block %c @%llu (%pg/%llu/%d) which has write error!\n",
2279                                btrfsic_get_block_type(state, l->block_ref_to),
2280                                l->block_ref_to->logical_bytenr,
2281                                l->block_ref_to->dev_state->bdev,
2282                                l->block_ref_to->dev_bytenr,
2283                                l->block_ref_to->mirror_num);
2284                         ret = -1;
2285                 } else if (l->parent_generation !=
2286                            l->block_ref_to->generation &&
2287                            BTRFSIC_GENERATION_UNKNOWN !=
2288                            l->parent_generation &&
2289                            BTRFSIC_GENERATION_UNKNOWN !=
2290                            l->block_ref_to->generation) {
2291                         pr_info(
2292 "btrfs: attempt to write superblock which references block %c @%llu (%pg/%llu/%d) with generation %llu != parent generation %llu!\n",
2293                                btrfsic_get_block_type(state, l->block_ref_to),
2294                                l->block_ref_to->logical_bytenr,
2295                                l->block_ref_to->dev_state->bdev,
2296                                l->block_ref_to->dev_bytenr,
2297                                l->block_ref_to->mirror_num,
2298                                l->block_ref_to->generation,
2299                                l->parent_generation);
2300                         ret = -1;
2301                 } else if (l->block_ref_to->flush_gen >
2302                            l->block_ref_to->dev_state->last_flush_gen) {
2303                         pr_info(
2304 "btrfs: attempt to write superblock which references block %c @%llu (%pg/%llu/%d) which is not flushed out of disk's write cache (block flush_gen=%llu, dev->flush_gen=%llu)!\n",
2305                                btrfsic_get_block_type(state, l->block_ref_to),
2306                                l->block_ref_to->logical_bytenr,
2307                                l->block_ref_to->dev_state->bdev,
2308                                l->block_ref_to->dev_bytenr,
2309                                l->block_ref_to->mirror_num, block->flush_gen,
2310                                l->block_ref_to->dev_state->last_flush_gen);
2311                         ret = -1;
2312                 } else if (-1 == btrfsic_check_all_ref_blocks(state,
2313                                                               l->block_ref_to,
2314                                                               recursion_level +
2315                                                               1)) {
2316                         ret = -1;
2317                 }
2318         }
2319
2320         return ret;
2321 }
2322
2323 static int btrfsic_is_block_ref_by_superblock(
2324                 const struct btrfsic_state *state,
2325                 const struct btrfsic_block *block,
2326                 int recursion_level)
2327 {
2328         const struct btrfsic_block_link *l;
2329
2330         if (recursion_level >= 3 + BTRFS_MAX_LEVEL) {
2331                 /* refer to comment at "abort cyclic linkage (case 1)" */
2332                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2333                         pr_info("btrfsic: abort cyclic linkage (case 2).\n");
2334
2335                 return 0;
2336         }
2337
2338         /*
2339          * This algorithm is recursive because the amount of used stack space
2340          * is very small and the max recursion depth is limited.
2341          */
2342         list_for_each_entry(l, &block->ref_from_list, node_ref_from) {
2343                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2344                         pr_info(
2345         "rl=%d, %c @%llu (%pg/%llu/%d) is ref %u* from %c @%llu (%pg/%llu/%d)\n",
2346                                recursion_level,
2347                                btrfsic_get_block_type(state, block),
2348                                block->logical_bytenr, block->dev_state->bdev,
2349                                block->dev_bytenr, block->mirror_num,
2350                                l->ref_cnt,
2351                                btrfsic_get_block_type(state, l->block_ref_from),
2352                                l->block_ref_from->logical_bytenr,
2353                                l->block_ref_from->dev_state->bdev,
2354                                l->block_ref_from->dev_bytenr,
2355                                l->block_ref_from->mirror_num);
2356                 if (l->block_ref_from->is_superblock &&
2357                     state->latest_superblock->dev_bytenr ==
2358                     l->block_ref_from->dev_bytenr &&
2359                     state->latest_superblock->dev_state->bdev ==
2360                     l->block_ref_from->dev_state->bdev)
2361                         return 1;
2362                 else if (btrfsic_is_block_ref_by_superblock(state,
2363                                                             l->block_ref_from,
2364                                                             recursion_level +
2365                                                             1))
2366                         return 1;
2367         }
2368
2369         return 0;
2370 }
2371
2372 static void btrfsic_print_add_link(const struct btrfsic_state *state,
2373                                    const struct btrfsic_block_link *l)
2374 {
2375         pr_info("add %u* link from %c @%llu (%pg/%llu/%d) to %c @%llu (%pg/%llu/%d)\n",
2376                l->ref_cnt,
2377                btrfsic_get_block_type(state, l->block_ref_from),
2378                l->block_ref_from->logical_bytenr,
2379                l->block_ref_from->dev_state->bdev,
2380                l->block_ref_from->dev_bytenr, l->block_ref_from->mirror_num,
2381                btrfsic_get_block_type(state, l->block_ref_to),
2382                l->block_ref_to->logical_bytenr,
2383                l->block_ref_to->dev_state->bdev, l->block_ref_to->dev_bytenr,
2384                l->block_ref_to->mirror_num);
2385 }
2386
2387 static void btrfsic_print_rem_link(const struct btrfsic_state *state,
2388                                    const struct btrfsic_block_link *l)
2389 {
2390         pr_info("rem %u* link from %c @%llu (%pg/%llu/%d) to %c @%llu (%pg/%llu/%d)\n",
2391                l->ref_cnt,
2392                btrfsic_get_block_type(state, l->block_ref_from),
2393                l->block_ref_from->logical_bytenr,
2394                l->block_ref_from->dev_state->bdev,
2395                l->block_ref_from->dev_bytenr, l->block_ref_from->mirror_num,
2396                btrfsic_get_block_type(state, l->block_ref_to),
2397                l->block_ref_to->logical_bytenr,
2398                l->block_ref_to->dev_state->bdev, l->block_ref_to->dev_bytenr,
2399                l->block_ref_to->mirror_num);
2400 }
2401
2402 static char btrfsic_get_block_type(const struct btrfsic_state *state,
2403                                    const struct btrfsic_block *block)
2404 {
2405         if (block->is_superblock &&
2406             state->latest_superblock->dev_bytenr == block->dev_bytenr &&
2407             state->latest_superblock->dev_state->bdev == block->dev_state->bdev)
2408                 return 'S';
2409         else if (block->is_superblock)
2410                 return 's';
2411         else if (block->is_metadata)
2412                 return 'M';
2413         else
2414                 return 'D';
2415 }
2416
2417 static void btrfsic_dump_tree(const struct btrfsic_state *state)
2418 {
2419         btrfsic_dump_tree_sub(state, state->latest_superblock, 0);
2420 }
2421
2422 static void btrfsic_dump_tree_sub(const struct btrfsic_state *state,
2423                                   const struct btrfsic_block *block,
2424                                   int indent_level)
2425 {
2426         const struct btrfsic_block_link *l;
2427         int indent_add;
2428         static char buf[80];
2429         int cursor_position;
2430
2431         /*
2432          * Should better fill an on-stack buffer with a complete line and
2433          * dump it at once when it is time to print a newline character.
2434          */
2435
2436         /*
2437          * This algorithm is recursive because the amount of used stack space
2438          * is very small and the max recursion depth is limited.
2439          */
2440         indent_add = sprintf(buf, "%c-%llu(%pg/%llu/%u)",
2441                              btrfsic_get_block_type(state, block),
2442                              block->logical_bytenr, block->dev_state->bdev,
2443                              block->dev_bytenr, block->mirror_num);
2444         if (indent_level + indent_add > BTRFSIC_TREE_DUMP_MAX_INDENT_LEVEL) {
2445                 printk("[...]\n");
2446                 return;
2447         }
2448         printk(buf);
2449         indent_level += indent_add;
2450         if (list_empty(&block->ref_to_list)) {
2451                 printk("\n");
2452                 return;
2453         }
2454         if (block->mirror_num > 1 &&
2455             !(state->print_mask & BTRFSIC_PRINT_MASK_TREE_WITH_ALL_MIRRORS)) {
2456                 printk(" [...]\n");
2457                 return;
2458         }
2459
2460         cursor_position = indent_level;
2461         list_for_each_entry(l, &block->ref_to_list, node_ref_to) {
2462                 while (cursor_position < indent_level) {
2463                         printk(" ");
2464                         cursor_position++;
2465                 }
2466                 if (l->ref_cnt > 1)
2467                         indent_add = sprintf(buf, " %d*--> ", l->ref_cnt);
2468                 else
2469                         indent_add = sprintf(buf, " --> ");
2470                 if (indent_level + indent_add >
2471                     BTRFSIC_TREE_DUMP_MAX_INDENT_LEVEL) {
2472                         printk("[...]\n");
2473                         cursor_position = 0;
2474                         continue;
2475                 }
2476
2477                 printk(buf);
2478
2479                 btrfsic_dump_tree_sub(state, l->block_ref_to,
2480                                       indent_level + indent_add);
2481                 cursor_position = 0;
2482         }
2483 }
2484
2485 static struct btrfsic_block_link *btrfsic_block_link_lookup_or_add(
2486                 struct btrfsic_state *state,
2487                 struct btrfsic_block_data_ctx *next_block_ctx,
2488                 struct btrfsic_block *next_block,
2489                 struct btrfsic_block *from_block,
2490                 u64 parent_generation)
2491 {
2492         struct btrfsic_block_link *l;
2493
2494         l = btrfsic_block_link_hashtable_lookup(next_block_ctx->dev->bdev,
2495                                                 next_block_ctx->dev_bytenr,
2496                                                 from_block->dev_state->bdev,
2497                                                 from_block->dev_bytenr,
2498                                                 &state->block_link_hashtable);
2499         if (NULL == l) {
2500                 l = btrfsic_block_link_alloc();
2501                 if (!l)
2502                         return NULL;
2503
2504                 l->block_ref_to = next_block;
2505                 l->block_ref_from = from_block;
2506                 l->ref_cnt = 1;
2507                 l->parent_generation = parent_generation;
2508
2509                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2510                         btrfsic_print_add_link(state, l);
2511
2512                 list_add(&l->node_ref_to, &from_block->ref_to_list);
2513                 list_add(&l->node_ref_from, &next_block->ref_from_list);
2514
2515                 btrfsic_block_link_hashtable_add(l,
2516                                                  &state->block_link_hashtable);
2517         } else {
2518                 l->ref_cnt++;
2519                 l->parent_generation = parent_generation;
2520                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2521                         btrfsic_print_add_link(state, l);
2522         }
2523
2524         return l;
2525 }
2526
2527 static struct btrfsic_block *btrfsic_block_lookup_or_add(
2528                 struct btrfsic_state *state,
2529                 struct btrfsic_block_data_ctx *block_ctx,
2530                 const char *additional_string,
2531                 int is_metadata,
2532                 int is_iodone,
2533                 int never_written,
2534                 int mirror_num,
2535                 int *was_created)
2536 {
2537         struct btrfsic_block *block;
2538
2539         block = btrfsic_block_hashtable_lookup(block_ctx->dev->bdev,
2540                                                block_ctx->dev_bytenr,
2541                                                &state->block_hashtable);
2542         if (NULL == block) {
2543                 struct btrfsic_dev_state *dev_state;
2544
2545                 block = btrfsic_block_alloc();
2546                 if (!block)
2547                         return NULL;
2548
2549                 dev_state = btrfsic_dev_state_lookup(block_ctx->dev->bdev->bd_dev);
2550                 if (NULL == dev_state) {
2551                         pr_info("btrfsic: error, lookup dev_state failed!\n");
2552                         btrfsic_block_free(block);
2553                         return NULL;
2554                 }
2555                 block->dev_state = dev_state;
2556                 block->dev_bytenr = block_ctx->dev_bytenr;
2557                 block->logical_bytenr = block_ctx->start;
2558                 block->is_metadata = is_metadata;
2559                 block->is_iodone = is_iodone;
2560                 block->never_written = never_written;
2561                 block->mirror_num = mirror_num;
2562                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2563                         pr_info("New %s%c-block @%llu (%pg/%llu/%d)\n",
2564                                additional_string,
2565                                btrfsic_get_block_type(state, block),
2566                                block->logical_bytenr, dev_state->bdev,
2567                                block->dev_bytenr, mirror_num);
2568                 list_add(&block->all_blocks_node, &state->all_blocks_list);
2569                 btrfsic_block_hashtable_add(block, &state->block_hashtable);
2570                 if (NULL != was_created)
2571                         *was_created = 1;
2572         } else {
2573                 if (NULL != was_created)
2574                         *was_created = 0;
2575         }
2576
2577         return block;
2578 }
2579
2580 static void btrfsic_cmp_log_and_dev_bytenr(struct btrfsic_state *state,
2581                                            u64 bytenr,
2582                                            struct btrfsic_dev_state *dev_state,
2583                                            u64 dev_bytenr)
2584 {
2585         struct btrfs_fs_info *fs_info = state->fs_info;
2586         struct btrfsic_block_data_ctx block_ctx;
2587         int num_copies;
2588         int mirror_num;
2589         int match = 0;
2590         int ret;
2591
2592         num_copies = btrfs_num_copies(fs_info, bytenr, state->metablock_size);
2593
2594         for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
2595                 ret = btrfsic_map_block(state, bytenr, state->metablock_size,
2596                                         &block_ctx, mirror_num);
2597                 if (ret) {
2598                         pr_info("btrfsic: btrfsic_map_block(logical @%llu, mirror %d) failed!\n",
2599                                bytenr, mirror_num);
2600                         continue;
2601                 }
2602
2603                 if (dev_state->bdev == block_ctx.dev->bdev &&
2604                     dev_bytenr == block_ctx.dev_bytenr) {
2605                         match++;
2606                         btrfsic_release_block_ctx(&block_ctx);
2607                         break;
2608                 }
2609                 btrfsic_release_block_ctx(&block_ctx);
2610         }
2611
2612         if (WARN_ON(!match)) {
2613                 pr_info(
2614 "btrfs: attempt to write M-block which contains logical bytenr that doesn't map to dev+physical bytenr of submit_bio, buffer->log_bytenr=%llu, submit_bio(bdev=%pg, phys_bytenr=%llu)!\n",
2615                        bytenr, dev_state->bdev, dev_bytenr);
2616                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
2617                         ret = btrfsic_map_block(state, bytenr,
2618                                                 state->metablock_size,
2619                                                 &block_ctx, mirror_num);
2620                         if (ret)
2621                                 continue;
2622
2623                         pr_info("read logical bytenr @%llu maps to (%pg/%llu/%d)\n",
2624                                bytenr, block_ctx.dev->bdev,
2625                                block_ctx.dev_bytenr, mirror_num);
2626                 }
2627         }
2628 }
2629
2630 static struct btrfsic_dev_state *btrfsic_dev_state_lookup(dev_t dev)
2631 {
2632         return btrfsic_dev_state_hashtable_lookup(dev,
2633                                                   &btrfsic_dev_state_hashtable);
2634 }
2635
2636 static void btrfsic_check_write_bio(struct bio *bio, struct btrfsic_dev_state *dev_state)
2637 {
2638         unsigned int segs = bio_segments(bio);
2639         u64 dev_bytenr = 512 * bio->bi_iter.bi_sector;
2640         u64 cur_bytenr = dev_bytenr;
2641         struct bvec_iter iter;
2642         struct bio_vec bvec;
2643         char **mapped_datav;
2644         int bio_is_patched = 0;
2645         int i = 0;
2646
2647         if (dev_state->state->print_mask & BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH)
2648                 pr_info(
2649 "submit_bio(rw=%d,0x%x, bi_vcnt=%u, bi_sector=%llu (bytenr %llu), bi_bdev=%p)\n",
2650                        bio_op(bio), bio->bi_opf, segs,
2651                        bio->bi_iter.bi_sector, dev_bytenr, bio->bi_bdev);
2652
2653         mapped_datav = kmalloc_array(segs, sizeof(*mapped_datav), GFP_NOFS);
2654         if (!mapped_datav)
2655                 return;
2656
2657         bio_for_each_segment(bvec, bio, iter) {
2658                 BUG_ON(bvec.bv_len != PAGE_SIZE);
2659                 mapped_datav[i] = page_address(bvec.bv_page);
2660                 i++;
2661
2662                 if (dev_state->state->print_mask &
2663                     BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH_VERBOSE)
2664                         pr_info("#%u: bytenr=%llu, len=%u, offset=%u\n",
2665                                i, cur_bytenr, bvec.bv_len, bvec.bv_offset);
2666                 cur_bytenr += bvec.bv_len;
2667         }
2668
2669         btrfsic_process_written_block(dev_state, dev_bytenr, mapped_datav, segs,
2670                                       bio, &bio_is_patched, bio->bi_opf);
2671         kfree(mapped_datav);
2672 }
2673
2674 static void btrfsic_check_flush_bio(struct bio *bio, struct btrfsic_dev_state *dev_state)
2675 {
2676         if (dev_state->state->print_mask & BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH)
2677                 pr_info("submit_bio(rw=%d,0x%x FLUSH, bdev=%p)\n",
2678                        bio_op(bio), bio->bi_opf, bio->bi_bdev);
2679
2680         if (dev_state->dummy_block_for_bio_bh_flush.is_iodone) {
2681                 struct btrfsic_block *const block =
2682                         &dev_state->dummy_block_for_bio_bh_flush;
2683
2684                 block->is_iodone = 0;
2685                 block->never_written = 0;
2686                 block->iodone_w_error = 0;
2687                 block->flush_gen = dev_state->last_flush_gen + 1;
2688                 block->submit_bio_bh_rw = bio->bi_opf;
2689                 block->orig_bio_private = bio->bi_private;
2690                 block->orig_bio_end_io = bio->bi_end_io;
2691                 block->next_in_same_bio = NULL;
2692                 bio->bi_private = block;
2693                 bio->bi_end_io = btrfsic_bio_end_io;
2694         } else if ((dev_state->state->print_mask &
2695                    (BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH |
2696                     BTRFSIC_PRINT_MASK_VERBOSE))) {
2697                 pr_info(
2698 "btrfsic_submit_bio(%pg) with FLUSH but dummy block already in use (ignored)!\n",
2699                        dev_state->bdev);
2700         }
2701 }
2702
2703 void btrfsic_check_bio(struct bio *bio)
2704 {
2705         struct btrfsic_dev_state *dev_state;
2706
2707         if (!btrfsic_is_initialized)
2708                 return;
2709
2710         /*
2711          * We can be called before btrfsic_mount, so there might not be a
2712          * dev_state.
2713          */
2714         dev_state = btrfsic_dev_state_lookup(bio->bi_bdev->bd_dev);
2715         mutex_lock(&btrfsic_mutex);
2716         if (dev_state) {
2717                 if (bio_op(bio) == REQ_OP_WRITE && bio_has_data(bio))
2718                         btrfsic_check_write_bio(bio, dev_state);
2719                 else if (bio->bi_opf & REQ_PREFLUSH)
2720                         btrfsic_check_flush_bio(bio, dev_state);
2721         }
2722         mutex_unlock(&btrfsic_mutex);
2723 }
2724
2725 int btrfsic_mount(struct btrfs_fs_info *fs_info,
2726                   struct btrfs_fs_devices *fs_devices,
2727                   int including_extent_data, u32 print_mask)
2728 {
2729         int ret;
2730         struct btrfsic_state *state;
2731         struct list_head *dev_head = &fs_devices->devices;
2732         struct btrfs_device *device;
2733
2734         if (!PAGE_ALIGNED(fs_info->nodesize)) {
2735                 pr_info("btrfsic: cannot handle nodesize %d not being a multiple of PAGE_SIZE %ld!\n",
2736                        fs_info->nodesize, PAGE_SIZE);
2737                 return -1;
2738         }
2739         if (!PAGE_ALIGNED(fs_info->sectorsize)) {
2740                 pr_info("btrfsic: cannot handle sectorsize %d not being a multiple of PAGE_SIZE %ld!\n",
2741                        fs_info->sectorsize, PAGE_SIZE);
2742                 return -1;
2743         }
2744         state = kvzalloc(sizeof(*state), GFP_KERNEL);
2745         if (!state)
2746                 return -ENOMEM;
2747
2748         if (!btrfsic_is_initialized) {
2749                 mutex_init(&btrfsic_mutex);
2750                 btrfsic_dev_state_hashtable_init(&btrfsic_dev_state_hashtable);
2751                 btrfsic_is_initialized = 1;
2752         }
2753         mutex_lock(&btrfsic_mutex);
2754         state->fs_info = fs_info;
2755         state->print_mask = print_mask;
2756         state->include_extent_data = including_extent_data;
2757         state->metablock_size = fs_info->nodesize;
2758         state->datablock_size = fs_info->sectorsize;
2759         INIT_LIST_HEAD(&state->all_blocks_list);
2760         btrfsic_block_hashtable_init(&state->block_hashtable);
2761         btrfsic_block_link_hashtable_init(&state->block_link_hashtable);
2762         state->max_superblock_generation = 0;
2763         state->latest_superblock = NULL;
2764
2765         list_for_each_entry(device, dev_head, dev_list) {
2766                 struct btrfsic_dev_state *ds;
2767
2768                 if (!device->bdev || !device->name)
2769                         continue;
2770
2771                 ds = btrfsic_dev_state_alloc();
2772                 if (NULL == ds) {
2773                         mutex_unlock(&btrfsic_mutex);
2774                         return -ENOMEM;
2775                 }
2776                 ds->bdev = device->bdev;
2777                 ds->state = state;
2778                 btrfsic_dev_state_hashtable_add(ds,
2779                                                 &btrfsic_dev_state_hashtable);
2780         }
2781
2782         ret = btrfsic_process_superblock(state, fs_devices);
2783         if (0 != ret) {
2784                 mutex_unlock(&btrfsic_mutex);
2785                 btrfsic_unmount(fs_devices);
2786                 return ret;
2787         }
2788
2789         if (state->print_mask & BTRFSIC_PRINT_MASK_INITIAL_DATABASE)
2790                 btrfsic_dump_database(state);
2791         if (state->print_mask & BTRFSIC_PRINT_MASK_INITIAL_TREE)
2792                 btrfsic_dump_tree(state);
2793
2794         mutex_unlock(&btrfsic_mutex);
2795         return 0;
2796 }
2797
2798 void btrfsic_unmount(struct btrfs_fs_devices *fs_devices)
2799 {
2800         struct btrfsic_block *b_all, *tmp_all;
2801         struct btrfsic_state *state;
2802         struct list_head *dev_head = &fs_devices->devices;
2803         struct btrfs_device *device;
2804
2805         if (!btrfsic_is_initialized)
2806                 return;
2807
2808         mutex_lock(&btrfsic_mutex);
2809
2810         state = NULL;
2811         list_for_each_entry(device, dev_head, dev_list) {
2812                 struct btrfsic_dev_state *ds;
2813
2814                 if (!device->bdev || !device->name)
2815                         continue;
2816
2817                 ds = btrfsic_dev_state_hashtable_lookup(
2818                                 device->bdev->bd_dev,
2819                                 &btrfsic_dev_state_hashtable);
2820                 if (NULL != ds) {
2821                         state = ds->state;
2822                         btrfsic_dev_state_hashtable_remove(ds);
2823                         btrfsic_dev_state_free(ds);
2824                 }
2825         }
2826
2827         if (NULL == state) {
2828                 pr_info("btrfsic: error, cannot find state information on umount!\n");
2829                 mutex_unlock(&btrfsic_mutex);
2830                 return;
2831         }
2832
2833         /*
2834          * Don't care about keeping the lists' state up to date,
2835          * just free all memory that was allocated dynamically.
2836          * Free the blocks and the block_links.
2837          */
2838         list_for_each_entry_safe(b_all, tmp_all, &state->all_blocks_list,
2839                                  all_blocks_node) {
2840                 struct btrfsic_block_link *l, *tmp;
2841
2842                 list_for_each_entry_safe(l, tmp, &b_all->ref_to_list,
2843                                          node_ref_to) {
2844                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2845                                 btrfsic_print_rem_link(state, l);
2846
2847                         l->ref_cnt--;
2848                         if (0 == l->ref_cnt)
2849                                 btrfsic_block_link_free(l);
2850                 }
2851
2852                 if (b_all->is_iodone || b_all->never_written)
2853                         btrfsic_block_free(b_all);
2854                 else
2855                         pr_info(
2856 "btrfs: attempt to free %c-block @%llu (%pg/%llu/%d) on umount which is not yet iodone!\n",
2857                                btrfsic_get_block_type(state, b_all),
2858                                b_all->logical_bytenr, b_all->dev_state->bdev,
2859                                b_all->dev_bytenr, b_all->mirror_num);
2860         }
2861
2862         mutex_unlock(&btrfsic_mutex);
2863
2864         kvfree(state);
2865 }