fs/btrfs/check-integrity.c

   1 /*
   2  * Copyright (C) STRATO AG 2011.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 /*
  20  * This module can be used to catch cases when the btrfs kernel
  21  * code executes write requests to the disk that bring the file
  22  * system in an inconsistent state. In such a state, a power-loss
  23  * or kernel panic event would cause that the data on disk is
  24  * lost or at least damaged.
  25  *
  26  * Code is added that examines all block write requests during
  27  * runtime (including writes of the super block). Three rules
  28  * are verified and an error is printed on violation of the
  29  * rules:
  30  * 1. It is not allowed to write a disk block which is
  31  *    currently referenced by the super block (either directly
  32  *    or indirectly).
  33  * 2. When a super block is written, it is verified that all
  34  *    referenced (directly or indirectly) blocks fulfill the
  35  *    following requirements:
  36  *    2a. All referenced blocks have either been present when
  37  *        the file system was mounted, (i.e., they have been
  38  *        referenced by the super block) or they have been
  39  *        written since then and the write completion callback
  40  *        was called and no write error was indicated and a
  41  *        FLUSH request to the device where these blocks are
  42  *        located was received and completed.
  43  *    2b. All referenced blocks need to have a generation
  44  *        number which is equal to the parent's number.
  45  *
  46  * One issue that was found using this module was that the log
  47  * tree on disk became temporarily corrupted because disk blocks
  48  * that had been in use for the log tree had been freed and
  49  * reused too early, while being referenced by the written super
  50  * block.
  51  *
  52  * The search term in the kernel log that can be used to filter
  53  * on the existence of detected integrity issues is
  54  * "btrfs: attempt".
  55  *
  56  * The integrity check is enabled via mount options. These
  57  * mount options are only supported if the integrity check
  58  * tool is compiled by defining BTRFS_FS_CHECK_INTEGRITY.
  59  *
  60  * Example #1, apply integrity checks to all metadata:
  61  * mount /dev/sdb1 /mnt -o check_int
  62  *
  63  * Example #2, apply integrity checks to all metadata and
  64  * to data extents:
  65  * mount /dev/sdb1 /mnt -o check_int_data
  66  *
  67  * Example #3, apply integrity checks to all metadata and dump
  68  * the tree that the super block references to kernel messages
  69  * each time after a super block was written:
  70  * mount /dev/sdb1 /mnt -o check_int,check_int_print_mask=263
  71  *
  72  * If the integrity check tool is included and activated in
  73  * the mount options, plenty of kernel memory is used, and
  74  * plenty of additional CPU cycles are spent. Enabling this
  75  * functionality is not intended for normal use. In most
  76  * cases, unless you are a btrfs developer who needs to verify
  77  * the integrity of (super)-block write requests, do not
  78  * enable the config option BTRFS_FS_CHECK_INTEGRITY to
  79  * include and compile the integrity check tool.
  80  *
  81  * Expect millions of lines of information in the kernel log with an
  82  * enabled check_int_print_mask. Therefore set LOG_BUF_SHIFT in the
  83  * kernel config to at least 26 (which is 64MB). Usually the value is
  84  * limited to 21 (which is 2MB) in init/Kconfig. The file needs to be
  85  * changed like this before LOG_BUF_SHIFT can be set to a high value:
  86  * config LOG_BUF_SHIFT
  87  *       int "Kernel log buffer size (16 => 64KB, 17 => 128KB)"
  88  *       range 12 30
  89  */
  90
  91 #include <linux/sched.h>
  92 #include <linux/slab.h>
  93 #include <linux/buffer_head.h>
  94 #include <linux/mutex.h>
  95 #include <linux/genhd.h>
  96 #include <linux/blkdev.h>
  97 #include <linux/vmalloc.h>
  98 #include <linux/string.h>
  99 #include "ctree.h"
 100 #include "disk-io.h"
 101 #include "hash.h"
 102 #include "transaction.h"
 103 #include "extent_io.h"
 104 #include "volumes.h"
 105 #include "print-tree.h"
 106 #include "locking.h"
 107 #include "check-integrity.h"
 108 #include "rcu-string.h"
 109 #include "compression.h"
 110
 111 #define BTRFSIC_BLOCK_HASHTABLE_SIZE 0x10000
 112 #define BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE 0x10000
 113 #define BTRFSIC_DEV2STATE_HASHTABLE_SIZE 0x100
 114 #define BTRFSIC_BLOCK_MAGIC_NUMBER 0x14491051
 115 #define BTRFSIC_BLOCK_LINK_MAGIC_NUMBER 0x11070807
 116 #define BTRFSIC_DEV2STATE_MAGIC_NUMBER 0x20111530
 117 #define BTRFSIC_BLOCK_STACK_FRAME_MAGIC_NUMBER 20111300
 118 #define BTRFSIC_TREE_DUMP_MAX_INDENT_LEVEL (200 - 6)    /* in characters,
 119                                                          * excluding " [...]" */
 120 #define BTRFSIC_GENERATION_UNKNOWN ((u64)-1)
 121
 122 /*
 123  * The definition of the bitmask fields for the print_mask.
 124  * They are specified with the mount option check_integrity_print_mask.
 125  */
 126 #define BTRFSIC_PRINT_MASK_SUPERBLOCK_WRITE                     0x00000001
 127 #define BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION         0x00000002
 128 #define BTRFSIC_PRINT_MASK_TREE_AFTER_SB_WRITE                  0x00000004
 129 #define BTRFSIC_PRINT_MASK_TREE_BEFORE_SB_WRITE                 0x00000008
 130 #define BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH                        0x00000010
 131 #define BTRFSIC_PRINT_MASK_END_IO_BIO_BH                        0x00000020
 132 #define BTRFSIC_PRINT_MASK_VERBOSE                              0x00000040
 133 #define BTRFSIC_PRINT_MASK_VERY_VERBOSE                         0x00000080
 134 #define BTRFSIC_PRINT_MASK_INITIAL_TREE                         0x00000100
 135 #define BTRFSIC_PRINT_MASK_INITIAL_ALL_TREES                    0x00000200
 136 #define BTRFSIC_PRINT_MASK_INITIAL_DATABASE                     0x00000400
 137 #define BTRFSIC_PRINT_MASK_NUM_COPIES                           0x00000800
 138 #define BTRFSIC_PRINT_MASK_TREE_WITH_ALL_MIRRORS                0x00001000
 139 #define BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH_VERBOSE                0x00002000
 140
 141 struct btrfsic_dev_state;
 142 struct btrfsic_state;
 143
 144 struct btrfsic_block {
 145         u32 magic_num;          /* only used for debug purposes */
 146         unsigned int is_metadata:1;     /* if it is meta-data, not data-data */
 147         unsigned int is_superblock:1;   /* if it is one of the superblocks */
 148         unsigned int is_iodone:1;       /* if is done by lower subsystem */
 149         unsigned int iodone_w_error:1;  /* error was indicated to endio */
 150         unsigned int never_written:1;   /* block was added because it was
 151                                          * referenced, not because it was
 152                                          * written */
 153         unsigned int mirror_num;        /* large enough to hold
 154                                          * BTRFS_SUPER_MIRROR_MAX */
 155         struct btrfsic_dev_state *dev_state;
 156         u64 dev_bytenr;         /* key, physical byte num on disk */
 157         u64 logical_bytenr;     /* logical byte num on disk */
 158         u64 generation;
 159         struct btrfs_disk_key disk_key; /* extra info to print in case of
 160                                          * issues, will not always be correct */
 161         struct list_head collision_resolving_node;      /* list node */
 162         struct list_head all_blocks_node;       /* list node */
 163
 164         /* the following two lists contain block_link items */
 165         struct list_head ref_to_list;   /* list */
 166         struct list_head ref_from_list; /* list */
 167         struct btrfsic_block *next_in_same_bio;
 168         void *orig_bio_bh_private;
 169         union {
 170                 bio_end_io_t *bio;
 171                 bh_end_io_t *bh;
 172         } orig_bio_bh_end_io;
 173         int submit_bio_bh_rw;
 174         u64 flush_gen; /* only valid if !never_written */
 175 };
 176
 177 /*
 178  * Elements of this type are allocated dynamically and required because
 179  * each block object can refer to and can be ref from multiple blocks.
 180  * The key to lookup them in the hashtable is the dev_bytenr of
 181  * the block ref to plus the one from the block referred from.
 182  * The fact that they are searchable via a hashtable and that a
 183  * ref_cnt is maintained is not required for the btrfs integrity
 184  * check algorithm itself, it is only used to make the output more
 185  * beautiful in case that an error is detected (an error is defined
 186  * as a write operation to a block while that block is still referenced).
 187  */
 188 struct btrfsic_block_link {
 189         u32 magic_num;          /* only used for debug purposes */
 190         u32 ref_cnt;
 191         struct list_head node_ref_to;   /* list node */
 192         struct list_head node_ref_from; /* list node */
 193         struct list_head collision_resolving_node;      /* list node */
 194         struct btrfsic_block *block_ref_to;
 195         struct btrfsic_block *block_ref_from;
 196         u64 parent_generation;
 197 };
 198
 199 struct btrfsic_dev_state {
 200         u32 magic_num;          /* only used for debug purposes */
 201         struct block_device *bdev;
 202         struct btrfsic_state *state;
 203         struct list_head collision_resolving_node;      /* list node */
 204         struct btrfsic_block dummy_block_for_bio_bh_flush;
 205         u64 last_flush_gen;
 206         char name[BDEVNAME_SIZE];
 207 };
 208
 209 struct btrfsic_block_hashtable {
 210         struct list_head table[BTRFSIC_BLOCK_HASHTABLE_SIZE];
 211 };
 212
 213 struct btrfsic_block_link_hashtable {
 214         struct list_head table[BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE];
 215 };
 216
 217 struct btrfsic_dev_state_hashtable {
 218         struct list_head table[BTRFSIC_DEV2STATE_HASHTABLE_SIZE];
 219 };
 220
 221 struct btrfsic_block_data_ctx {
 222         u64 start;              /* virtual bytenr */
 223         u64 dev_bytenr;         /* physical bytenr on device */
 224         u32 len;
 225         struct btrfsic_dev_state *dev;
 226         char **datav;
 227         struct page **pagev;
 228         void *mem_to_free;
 229 };
 230
 231 /* This structure is used to implement recursion without occupying
 232  * any stack space, refer to btrfsic_process_metablock() */
 233 struct btrfsic_stack_frame {
 234         u32 magic;
 235         u32 nr;
 236         int error;
 237         int i;
 238         int limit_nesting;
 239         int num_copies;
 240         int mirror_num;
 241         struct btrfsic_block *block;
 242         struct btrfsic_block_data_ctx *block_ctx;
 243         struct btrfsic_block *next_block;
 244         struct btrfsic_block_data_ctx next_block_ctx;
 245         struct btrfs_header *hdr;
 246         struct btrfsic_stack_frame *prev;
 247 };
 248
 249 /* Some state per mounted filesystem */
 250 struct btrfsic_state {
 251         u32 print_mask;
 252         int include_extent_data;
 253         int csum_size;
 254         struct list_head all_blocks_list;
 255         struct btrfsic_block_hashtable block_hashtable;
 256         struct btrfsic_block_link_hashtable block_link_hashtable;
 257         struct btrfs_root *root;
 258         u64 max_superblock_generation;
 259         struct btrfsic_block *latest_superblock;
 260         u32 metablock_size;
 261         u32 datablock_size;
 262 };
 263
 264 static void btrfsic_block_init(struct btrfsic_block *b);
 265 static struct btrfsic_block *btrfsic_block_alloc(void);
 266 static void btrfsic_block_free(struct btrfsic_block *b);
 267 static void btrfsic_block_link_init(struct btrfsic_block_link *n);
 268 static struct btrfsic_block_link *btrfsic_block_link_alloc(void);
 269 static void btrfsic_block_link_free(struct btrfsic_block_link *n);
 270 static void btrfsic_dev_state_init(struct btrfsic_dev_state *ds);
 271 static struct btrfsic_dev_state *btrfsic_dev_state_alloc(void);
 272 static void btrfsic_dev_state_free(struct btrfsic_dev_state *ds);
 273 static void btrfsic_block_hashtable_init(struct btrfsic_block_hashtable *h);
 274 static void btrfsic_block_hashtable_add(struct btrfsic_block *b,
 275                                         struct btrfsic_block_hashtable *h);
 276 static void btrfsic_block_hashtable_remove(struct btrfsic_block *b);
 277 static struct btrfsic_block *btrfsic_block_hashtable_lookup(
 278                 struct block_device *bdev,
 279                 u64 dev_bytenr,
 280                 struct btrfsic_block_hashtable *h);
 281 static void btrfsic_block_link_hashtable_init(
 282                 struct btrfsic_block_link_hashtable *h);
 283 static void btrfsic_block_link_hashtable_add(
 284                 struct btrfsic_block_link *l,
 285                 struct btrfsic_block_link_hashtable *h);
 286 static void btrfsic_block_link_hashtable_remove(struct btrfsic_block_link *l);
 287 static struct btrfsic_block_link *btrfsic_block_link_hashtable_lookup(
 288                 struct block_device *bdev_ref_to,
 289                 u64 dev_bytenr_ref_to,
 290                 struct block_device *bdev_ref_from,
 291                 u64 dev_bytenr_ref_from,
 292                 struct btrfsic_block_link_hashtable *h);
 293 static void btrfsic_dev_state_hashtable_init(
 294                 struct btrfsic_dev_state_hashtable *h);
 295 static void btrfsic_dev_state_hashtable_add(
 296                 struct btrfsic_dev_state *ds,
 297                 struct btrfsic_dev_state_hashtable *h);
 298 static void btrfsic_dev_state_hashtable_remove(struct btrfsic_dev_state *ds);
 299 static struct btrfsic_dev_state *btrfsic_dev_state_hashtable_lookup(
 300                 struct block_device *bdev,
 301                 struct btrfsic_dev_state_hashtable *h);
 302 static struct btrfsic_stack_frame *btrfsic_stack_frame_alloc(void);
 303 static void btrfsic_stack_frame_free(struct btrfsic_stack_frame *sf);
 304 static int btrfsic_process_superblock(struct btrfsic_state *state,
 305                                       struct btrfs_fs_devices *fs_devices);
 306 static int btrfsic_process_metablock(struct btrfsic_state *state,
 307                                      struct btrfsic_block *block,
 308                                      struct btrfsic_block_data_ctx *block_ctx,
 309                                      int limit_nesting, int force_iodone_flag);
 310 static void btrfsic_read_from_block_data(
 311         struct btrfsic_block_data_ctx *block_ctx,
 312         void *dst, u32 offset, size_t len);
 313 static int btrfsic_create_link_to_next_block(
 314                 struct btrfsic_state *state,
 315                 struct btrfsic_block *block,
 316                 struct btrfsic_block_data_ctx
 317                 *block_ctx, u64 next_bytenr,
 318                 int limit_nesting,
 319                 struct btrfsic_block_data_ctx *next_block_ctx,
 320                 struct btrfsic_block **next_blockp,
 321                 int force_iodone_flag,
 322                 int *num_copiesp, int *mirror_nump,
 323                 struct btrfs_disk_key *disk_key,
 324                 u64 parent_generation);
 325 static int btrfsic_handle_extent_data(struct btrfsic_state *state,
 326                                       struct btrfsic_block *block,
 327                                       struct btrfsic_block_data_ctx *block_ctx,
 328                                       u32 item_offset, int force_iodone_flag);
 329 static int btrfsic_map_block(struct btrfsic_state *state, u64 bytenr, u32 len,
 330                              struct btrfsic_block_data_ctx *block_ctx_out,
 331                              int mirror_num);
 332 static void btrfsic_release_block_ctx(struct btrfsic_block_data_ctx *block_ctx);
 333 static int btrfsic_read_block(struct btrfsic_state *state,
 334                               struct btrfsic_block_data_ctx *block_ctx);
 335 static void btrfsic_dump_database(struct btrfsic_state *state);
 336 static int btrfsic_test_for_metadata(struct btrfsic_state *state,
 337                                      char **datav, unsigned int num_pages);
 338 static void btrfsic_process_written_block(struct btrfsic_dev_state *dev_state,
 339                                           u64 dev_bytenr, char **mapped_datav,
 340                                           unsigned int num_pages,
 341                                           struct bio *bio, int *bio_is_patched,
 342                                           struct buffer_head *bh,
 343                                           int submit_bio_bh_rw);
 344 static int btrfsic_process_written_superblock(
 345                 struct btrfsic_state *state,
 346                 struct btrfsic_block *const block,
 347                 struct btrfs_super_block *const super_hdr);
 348 static void btrfsic_bio_end_io(struct bio *bp);
 349 static void btrfsic_bh_end_io(struct buffer_head *bh, int uptodate);
 350 static int btrfsic_is_block_ref_by_superblock(const struct btrfsic_state *state,
 351                                               const struct btrfsic_block *block,
 352                                               int recursion_level);
 353 static int btrfsic_check_all_ref_blocks(struct btrfsic_state *state,
 354                                         struct btrfsic_block *const block,
 355                                         int recursion_level);
 356 static void btrfsic_print_add_link(const struct btrfsic_state *state,
 357                                    const struct btrfsic_block_link *l);
 358 static void btrfsic_print_rem_link(const struct btrfsic_state *state,
 359                                    const struct btrfsic_block_link *l);
 360 static char btrfsic_get_block_type(const struct btrfsic_state *state,
 361                                    const struct btrfsic_block *block);
 362 static void btrfsic_dump_tree(const struct btrfsic_state *state);
 363 static void btrfsic_dump_tree_sub(const struct btrfsic_state *state,
 364                                   const struct btrfsic_block *block,
 365                                   int indent_level);
 366 static struct btrfsic_block_link *btrfsic_block_link_lookup_or_add(
 367                 struct btrfsic_state *state,
 368                 struct btrfsic_block_data_ctx *next_block_ctx,
 369                 struct btrfsic_block *next_block,
 370                 struct btrfsic_block *from_block,
 371                 u64 parent_generation);
 372 static struct btrfsic_block *btrfsic_block_lookup_or_add(
 373                 struct btrfsic_state *state,
 374                 struct btrfsic_block_data_ctx *block_ctx,
 375                 const char *additional_string,
 376                 int is_metadata,
 377                 int is_iodone,
 378                 int never_written,
 379                 int mirror_num,
 380                 int *was_created);
 381 static int btrfsic_process_superblock_dev_mirror(
 382                 struct btrfsic_state *state,
 383                 struct btrfsic_dev_state *dev_state,
 384                 struct btrfs_device *device,
 385                 int superblock_mirror_num,
 386                 struct btrfsic_dev_state **selected_dev_state,
 387                 struct btrfs_super_block *selected_super);
 388 static struct btrfsic_dev_state *btrfsic_dev_state_lookup(
 389                 struct block_device *bdev);
 390 static void btrfsic_cmp_log_and_dev_bytenr(struct btrfsic_state *state,
 391                                            u64 bytenr,
 392                                            struct btrfsic_dev_state *dev_state,
 393                                            u64 dev_bytenr);
 394
 395 static struct mutex btrfsic_mutex;
 396 static int btrfsic_is_initialized;
 397 static struct btrfsic_dev_state_hashtable btrfsic_dev_state_hashtable;
 398
 399
 400 static void btrfsic_block_init(struct btrfsic_block *b)
 401 {
 402         b->magic_num = BTRFSIC_BLOCK_MAGIC_NUMBER;
 403         b->dev_state = NULL;
 404         b->dev_bytenr = 0;
 405         b->logical_bytenr = 0;
 406         b->generation = BTRFSIC_GENERATION_UNKNOWN;
 407         b->disk_key.objectid = 0;
 408         b->disk_key.type = 0;
 409         b->disk_key.offset = 0;
 410         b->is_metadata = 0;
 411         b->is_superblock = 0;
 412         b->is_iodone = 0;
 413         b->iodone_w_error = 0;
 414         b->never_written = 0;
 415         b->mirror_num = 0;
 416         b->next_in_same_bio = NULL;
 417         b->orig_bio_bh_private = NULL;
 418         b->orig_bio_bh_end_io.bio = NULL;
 419         INIT_LIST_HEAD(&b->collision_resolving_node);
 420         INIT_LIST_HEAD(&b->all_blocks_node);
 421         INIT_LIST_HEAD(&b->ref_to_list);
 422         INIT_LIST_HEAD(&b->ref_from_list);
 423         b->submit_bio_bh_rw = 0;
 424         b->flush_gen = 0;
 425 }
 426
 427 static struct btrfsic_block *btrfsic_block_alloc(void)
 428 {
 429         struct btrfsic_block *b;
 430
 431         b = kzalloc(sizeof(*b), GFP_NOFS);
 432         if (NULL != b)
 433                 btrfsic_block_init(b);
 434
 435         return b;
 436 }
 437
 438 static void btrfsic_block_free(struct btrfsic_block *b)
 439 {
 440         BUG_ON(!(NULL == b || BTRFSIC_BLOCK_MAGIC_NUMBER == b->magic_num));
 441         kfree(b);
 442 }
 443
 444 static void btrfsic_block_link_init(struct btrfsic_block_link *l)
 445 {
 446         l->magic_num = BTRFSIC_BLOCK_LINK_MAGIC_NUMBER;
 447         l->ref_cnt = 1;
 448         INIT_LIST_HEAD(&l->node_ref_to);
 449         INIT_LIST_HEAD(&l->node_ref_from);
 450         INIT_LIST_HEAD(&l->collision_resolving_node);
 451         l->block_ref_to = NULL;
 452         l->block_ref_from = NULL;
 453 }
 454
 455 static struct btrfsic_block_link *btrfsic_block_link_alloc(void)
 456 {
 457         struct btrfsic_block_link *l;
 458
 459         l = kzalloc(sizeof(*l), GFP_NOFS);
 460         if (NULL != l)
 461                 btrfsic_block_link_init(l);
 462
 463         return l;
 464 }
 465
 466 static void btrfsic_block_link_free(struct btrfsic_block_link *l)
 467 {
 468         BUG_ON(!(NULL == l || BTRFSIC_BLOCK_LINK_MAGIC_NUMBER == l->magic_num));
 469         kfree(l);
 470 }
 471
 472 static void btrfsic_dev_state_init(struct btrfsic_dev_state *ds)
 473 {
 474         ds->magic_num = BTRFSIC_DEV2STATE_MAGIC_NUMBER;
 475         ds->bdev = NULL;
 476         ds->state = NULL;
 477         ds->name[0] = '\0';
 478         INIT_LIST_HEAD(&ds->collision_resolving_node);
 479         ds->last_flush_gen = 0;
 480         btrfsic_block_init(&ds->dummy_block_for_bio_bh_flush);
 481         ds->dummy_block_for_bio_bh_flush.is_iodone = 1;
 482         ds->dummy_block_for_bio_bh_flush.dev_state = ds;
 483 }
 484
 485 static struct btrfsic_dev_state *btrfsic_dev_state_alloc(void)
 486 {
 487         struct btrfsic_dev_state *ds;
 488
 489         ds = kzalloc(sizeof(*ds), GFP_NOFS);
 490         if (NULL != ds)
 491                 btrfsic_dev_state_init(ds);
 492
 493         return ds;
 494 }
 495
 496 static void btrfsic_dev_state_free(struct btrfsic_dev_state *ds)
 497 {
 498         BUG_ON(!(NULL == ds ||
 499                  BTRFSIC_DEV2STATE_MAGIC_NUMBER == ds->magic_num));
 500         kfree(ds);
 501 }
 502
 503 static void btrfsic_block_hashtable_init(struct btrfsic_block_hashtable *h)
 504 {
 505         int i;
 506
 507         for (i = 0; i < BTRFSIC_BLOCK_HASHTABLE_SIZE; i++)
 508                 INIT_LIST_HEAD(h->table + i);
 509 }
 510
 511 static void btrfsic_block_hashtable_add(struct btrfsic_block *b,
 512                                         struct btrfsic_block_hashtable *h)
 513 {
 514         const unsigned int hashval =
 515             (((unsigned int)(b->dev_bytenr >> 16)) ^
 516              ((unsigned int)((uintptr_t)b->dev_state->bdev))) &
 517              (BTRFSIC_BLOCK_HASHTABLE_SIZE - 1);
 518
 519         list_add(&b->collision_resolving_node, h->table + hashval);
 520 }
 521
 522 static void btrfsic_block_hashtable_remove(struct btrfsic_block *b)
 523 {
 524         list_del(&b->collision_resolving_node);
 525 }
 526
 527 static struct btrfsic_block *btrfsic_block_hashtable_lookup(
 528                 struct block_device *bdev,
 529                 u64 dev_bytenr,
 530                 struct btrfsic_block_hashtable *h)
 531 {
 532         const unsigned int hashval =
 533             (((unsigned int)(dev_bytenr >> 16)) ^
 534              ((unsigned int)((uintptr_t)bdev))) &
 535              (BTRFSIC_BLOCK_HASHTABLE_SIZE - 1);
 536         struct btrfsic_block *b;
 537
 538         list_for_each_entry(b, h->table + hashval, collision_resolving_node) {
 539                 if (b->dev_state->bdev == bdev && b->dev_bytenr == dev_bytenr)
 540                         return b;
 541         }
 542
 543         return NULL;
 544 }
 545
 546 static void btrfsic_block_link_hashtable_init(
 547                 struct btrfsic_block_link_hashtable *h)
 548 {
 549         int i;
 550
 551         for (i = 0; i < BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE; i++)
 552                 INIT_LIST_HEAD(h->table + i);
 553 }
 554
 555 static void btrfsic_block_link_hashtable_add(
 556                 struct btrfsic_block_link *l,
 557                 struct btrfsic_block_link_hashtable *h)
 558 {
 559         const unsigned int hashval =
 560             (((unsigned int)(l->block_ref_to->dev_bytenr >> 16)) ^
 561              ((unsigned int)(l->block_ref_from->dev_bytenr >> 16)) ^
 562              ((unsigned int)((uintptr_t)l->block_ref_to->dev_state->bdev)) ^
 563              ((unsigned int)((uintptr_t)l->block_ref_from->dev_state->bdev)))
 564              & (BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE - 1);
 565
 566         BUG_ON(NULL == l->block_ref_to);
 567         BUG_ON(NULL == l->block_ref_from);
 568         list_add(&l->collision_resolving_node, h->table + hashval);
 569 }
 570
 571 static void btrfsic_block_link_hashtable_remove(struct btrfsic_block_link *l)
 572 {
 573         list_del(&l->collision_resolving_node);
 574 }
 575
 576 static struct btrfsic_block_link *btrfsic_block_link_hashtable_lookup(
 577                 struct block_device *bdev_ref_to,
 578                 u64 dev_bytenr_ref_to,
 579                 struct block_device *bdev_ref_from,
 580                 u64 dev_bytenr_ref_from,
 581                 struct btrfsic_block_link_hashtable *h)
 582 {
 583         const unsigned int hashval =
 584             (((unsigned int)(dev_bytenr_ref_to >> 16)) ^
 585              ((unsigned int)(dev_bytenr_ref_from >> 16)) ^
 586              ((unsigned int)((uintptr_t)bdev_ref_to)) ^
 587              ((unsigned int)((uintptr_t)bdev_ref_from))) &
 588              (BTRFSIC_BLOCK_LINK_HASHTABLE_SIZE - 1);
 589         struct btrfsic_block_link *l;
 590
 591         list_for_each_entry(l, h->table + hashval, collision_resolving_node) {
 592                 BUG_ON(NULL == l->block_ref_to);
 593                 BUG_ON(NULL == l->block_ref_from);
 594                 if (l->block_ref_to->dev_state->bdev == bdev_ref_to &&
 595                     l->block_ref_to->dev_bytenr == dev_bytenr_ref_to &&
 596                     l->block_ref_from->dev_state->bdev == bdev_ref_from &&
 597                     l->block_ref_from->dev_bytenr == dev_bytenr_ref_from)
 598                         return l;
 599         }
 600
 601         return NULL;
 602 }
 603
 604 static void btrfsic_dev_state_hashtable_init(
 605                 struct btrfsic_dev_state_hashtable *h)
 606 {
 607         int i;
 608
 609         for (i = 0; i < BTRFSIC_DEV2STATE_HASHTABLE_SIZE; i++)
 610                 INIT_LIST_HEAD(h->table + i);
 611 }
 612
 613 static void btrfsic_dev_state_hashtable_add(
 614                 struct btrfsic_dev_state *ds,
 615                 struct btrfsic_dev_state_hashtable *h)
 616 {
 617         const unsigned int hashval =
 618             (((unsigned int)((uintptr_t)ds->bdev)) &
 619              (BTRFSIC_DEV2STATE_HASHTABLE_SIZE - 1));
 620
 621         list_add(&ds->collision_resolving_node, h->table + hashval);
 622 }
 623
 624 static void btrfsic_dev_state_hashtable_remove(struct btrfsic_dev_state *ds)
 625 {
 626         list_del(&ds->collision_resolving_node);
 627 }
 628
 629 static struct btrfsic_dev_state *btrfsic_dev_state_hashtable_lookup(
 630                 struct block_device *bdev,
 631                 struct btrfsic_dev_state_hashtable *h)
 632 {
 633         const unsigned int hashval =
 634             (((unsigned int)((uintptr_t)bdev)) &
 635              (BTRFSIC_DEV2STATE_HASHTABLE_SIZE - 1));
 636         struct btrfsic_dev_state *ds;
 637
 638         list_for_each_entry(ds, h->table + hashval, collision_resolving_node) {
 639                 if (ds->bdev == bdev)
 640                         return ds;
 641         }
 642
 643         return NULL;
 644 }
 645
 646 static int btrfsic_process_superblock(struct btrfsic_state *state,
 647                                       struct btrfs_fs_devices *fs_devices)
 648 {
 649         int ret = 0;
 650         struct btrfs_super_block *selected_super;
 651         struct list_head *dev_head = &fs_devices->devices;
 652         struct btrfs_device *device;
 653         struct btrfsic_dev_state *selected_dev_state = NULL;
 654         int pass;
 655
 656         BUG_ON(NULL == state);
 657         selected_super = kzalloc(sizeof(*selected_super), GFP_NOFS);
 658         if (NULL == selected_super) {
 659                 printk(KERN_INFO "btrfsic: error, kmalloc failed!\n");
 660                 return -ENOMEM;
 661         }
 662
 663         list_for_each_entry(device, dev_head, dev_list) {
 664                 int i;
 665                 struct btrfsic_dev_state *dev_state;
 666
 667                 if (!device->bdev || !device->name)
 668                         continue;
 669
 670                 dev_state = btrfsic_dev_state_lookup(device->bdev);
 671                 BUG_ON(NULL == dev_state);
 672                 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 673                         ret = btrfsic_process_superblock_dev_mirror(
 674                                         state, dev_state, device, i,
 675                                         &selected_dev_state, selected_super);
 676                         if (0 != ret && 0 == i) {
 677                                 kfree(selected_super);
 678                                 return ret;
 679                         }
 680                 }
 681         }
 682
 683         if (NULL == state->latest_superblock) {
 684                 printk(KERN_INFO "btrfsic: no superblock found!\n");
 685                 kfree(selected_super);
 686                 return -1;
 687         }
 688
 689         state->csum_size = btrfs_super_csum_size(selected_super);
 690
 691         for (pass = 0; pass < 3; pass++) {
 692                 int num_copies;
 693                 int mirror_num;
 694                 u64 next_bytenr;
 695
 696                 switch (pass) {
 697                 case 0:
 698                         next_bytenr = btrfs_super_root(selected_super);
 699                         if (state->print_mask &
 700                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
 701                                 printk(KERN_INFO "root@%llu\n", next_bytenr);
 702                         break;
 703                 case 1:
 704                         next_bytenr = btrfs_super_chunk_root(selected_super);
 705                         if (state->print_mask &
 706                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
 707                                 printk(KERN_INFO "chunk@%llu\n", next_bytenr);
 708                         break;
 709                 case 2:
 710                         next_bytenr = btrfs_super_log_root(selected_super);
 711                         if (0 == next_bytenr)
 712                                 continue;
 713                         if (state->print_mask &
 714                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
 715                                 printk(KERN_INFO "log@%llu\n", next_bytenr);
 716                         break;
 717                 }
 718
 719                 num_copies =
 720                     btrfs_num_copies(state->root->fs_info,
 721                                      next_bytenr, state->metablock_size);
 722                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
 723                         printk(KERN_INFO "num_copies(log_bytenr=%llu) = %d\n",
 724                                next_bytenr, num_copies);
 725
 726                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
 727                         struct btrfsic_block *next_block;
 728                         struct btrfsic_block_data_ctx tmp_next_block_ctx;
 729                         struct btrfsic_block_link *l;
 730
 731                         ret = btrfsic_map_block(state, next_bytenr,
 732                                                 state->metablock_size,
 733                                                 &tmp_next_block_ctx,
 734                                                 mirror_num);
 735                         if (ret) {
 736                                 printk(KERN_INFO "btrfsic:"
 737                                        " btrfsic_map_block(root @%llu,"
 738                                        " mirror %d) failed!\n",
 739                                        next_bytenr, mirror_num);
 740                                 kfree(selected_super);
 741                                 return -1;
 742                         }
 743
 744                         next_block = btrfsic_block_hashtable_lookup(
 745                                         tmp_next_block_ctx.dev->bdev,
 746                                         tmp_next_block_ctx.dev_bytenr,
 747                                         &state->block_hashtable);
 748                         BUG_ON(NULL == next_block);
 749
 750                         l = btrfsic_block_link_hashtable_lookup(
 751                                         tmp_next_block_ctx.dev->bdev,
 752                                         tmp_next_block_ctx.dev_bytenr,
 753                                         state->latest_superblock->dev_state->
 754                                         bdev,
 755                                         state->latest_superblock->dev_bytenr,
 756                                         &state->block_link_hashtable);
 757                         BUG_ON(NULL == l);
 758
 759                         ret = btrfsic_read_block(state, &tmp_next_block_ctx);
 760                         if (ret < (int)PAGE_SIZE) {
 761                                 printk(KERN_INFO
 762                                        "btrfsic: read @logical %llu failed!\n",
 763                                        tmp_next_block_ctx.start);
 764                                 btrfsic_release_block_ctx(&tmp_next_block_ctx);
 765                                 kfree(selected_super);
 766                                 return -1;
 767                         }
 768
 769                         ret = btrfsic_process_metablock(state,
 770                                                         next_block,
 771                                                         &tmp_next_block_ctx,
 772                                                         BTRFS_MAX_LEVEL + 3, 1);
 773                         btrfsic_release_block_ctx(&tmp_next_block_ctx);
 774                 }
 775         }
 776
 777         kfree(selected_super);
 778         return ret;
 779 }
 780
 781 static int btrfsic_process_superblock_dev_mirror(
 782                 struct btrfsic_state *state,
 783                 struct btrfsic_dev_state *dev_state,
 784                 struct btrfs_device *device,
 785                 int superblock_mirror_num,
 786                 struct btrfsic_dev_state **selected_dev_state,
 787                 struct btrfs_super_block *selected_super)
 788 {
 789         struct btrfs_super_block *super_tmp;
 790         u64 dev_bytenr;
 791         struct buffer_head *bh;
 792         struct btrfsic_block *superblock_tmp;
 793         int pass;
 794         struct block_device *const superblock_bdev = device->bdev;
 795
 796         /* super block bytenr is always the unmapped device bytenr */
 797         dev_bytenr = btrfs_sb_offset(superblock_mirror_num);
 798         if (dev_bytenr + BTRFS_SUPER_INFO_SIZE > device->commit_total_bytes)
 799                 return -1;
 800         bh = __bread(superblock_bdev, dev_bytenr / 4096,
 801                      BTRFS_SUPER_INFO_SIZE);
 802         if (NULL == bh)
 803                 return -1;
 804         super_tmp = (struct btrfs_super_block *)
 805             (bh->b_data + (dev_bytenr & 4095));
 806
 807         if (btrfs_super_bytenr(super_tmp) != dev_bytenr ||
 808             btrfs_super_magic(super_tmp) != BTRFS_MAGIC ||
 809             memcmp(device->uuid, super_tmp->dev_item.uuid, BTRFS_UUID_SIZE) ||
 810             btrfs_super_nodesize(super_tmp) != state->metablock_size ||
 811             btrfs_super_sectorsize(super_tmp) != state->datablock_size) {
 812                 brelse(bh);
 813                 return 0;
 814         }
 815
 816         superblock_tmp =
 817             btrfsic_block_hashtable_lookup(superblock_bdev,
 818                                            dev_bytenr,
 819                                            &state->block_hashtable);
 820         if (NULL == superblock_tmp) {
 821                 superblock_tmp = btrfsic_block_alloc();
 822                 if (NULL == superblock_tmp) {
 823                         printk(KERN_INFO "btrfsic: error, kmalloc failed!\n");
 824                         brelse(bh);
 825                         return -1;
 826                 }
 827                 /* for superblock, only the dev_bytenr makes sense */
 828                 superblock_tmp->dev_bytenr = dev_bytenr;
 829                 superblock_tmp->dev_state = dev_state;
 830                 superblock_tmp->logical_bytenr = dev_bytenr;
 831                 superblock_tmp->generation = btrfs_super_generation(super_tmp);
 832                 superblock_tmp->is_metadata = 1;
 833                 superblock_tmp->is_superblock = 1;
 834                 superblock_tmp->is_iodone = 1;
 835                 superblock_tmp->never_written = 0;
 836                 superblock_tmp->mirror_num = 1 + superblock_mirror_num;
 837                 if (state->print_mask & BTRFSIC_PRINT_MASK_SUPERBLOCK_WRITE)
 838                         btrfs_info_in_rcu(device->dev_root->fs_info,
 839                                 "new initial S-block (bdev %p, %s) @%llu (%s/%llu/%d)",
 840                                      superblock_bdev,
 841                                      rcu_str_deref(device->name), dev_bytenr,
 842                                      dev_state->name, dev_bytenr,
 843                                      superblock_mirror_num);
 844                 list_add(&superblock_tmp->all_blocks_node,
 845                          &state->all_blocks_list);
 846                 btrfsic_block_hashtable_add(superblock_tmp,
 847                                             &state->block_hashtable);
 848         }
 849
 850         /* select the one with the highest generation field */
 851         if (btrfs_super_generation(super_tmp) >
 852             state->max_superblock_generation ||
 853             0 == state->max_superblock_generation) {
 854                 memcpy(selected_super, super_tmp, sizeof(*selected_super));
 855                 *selected_dev_state = dev_state;
 856                 state->max_superblock_generation =
 857                     btrfs_super_generation(super_tmp);
 858                 state->latest_superblock = superblock_tmp;
 859         }
 860
 861         for (pass = 0; pass < 3; pass++) {
 862                 u64 next_bytenr;
 863                 int num_copies;
 864                 int mirror_num;
 865                 const char *additional_string = NULL;
 866                 struct btrfs_disk_key tmp_disk_key;
 867
 868                 tmp_disk_key.type = BTRFS_ROOT_ITEM_KEY;
 869                 tmp_disk_key.offset = 0;
 870                 switch (pass) {
 871                 case 0:
 872                         btrfs_set_disk_key_objectid(&tmp_disk_key,
 873                                                     BTRFS_ROOT_TREE_OBJECTID);
 874                         additional_string = "initial root ";
 875                         next_bytenr = btrfs_super_root(super_tmp);
 876                         break;
 877                 case 1:
 878                         btrfs_set_disk_key_objectid(&tmp_disk_key,
 879                                                     BTRFS_CHUNK_TREE_OBJECTID);
 880                         additional_string = "initial chunk ";
 881                         next_bytenr = btrfs_super_chunk_root(super_tmp);
 882                         break;
 883                 case 2:
 884                         btrfs_set_disk_key_objectid(&tmp_disk_key,
 885                                                     BTRFS_TREE_LOG_OBJECTID);
 886                         additional_string = "initial log ";
 887                         next_bytenr = btrfs_super_log_root(super_tmp);
 888                         if (0 == next_bytenr)
 889                                 continue;
 890                         break;
 891                 }
 892
 893                 num_copies =
 894                     btrfs_num_copies(state->root->fs_info,
 895                                      next_bytenr, state->metablock_size);
 896                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
 897                         printk(KERN_INFO "num_copies(log_bytenr=%llu) = %d\n",
 898                                next_bytenr, num_copies);
 899                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
 900                         struct btrfsic_block *next_block;
 901                         struct btrfsic_block_data_ctx tmp_next_block_ctx;
 902                         struct btrfsic_block_link *l;
 903
 904                         if (btrfsic_map_block(state, next_bytenr,
 905                                               state->metablock_size,
 906                                               &tmp_next_block_ctx,
 907                                               mirror_num)) {
 908                                 printk(KERN_INFO "btrfsic: btrfsic_map_block("
 909                                        "bytenr @%llu, mirror %d) failed!\n",
 910                                        next_bytenr, mirror_num);
 911                                 brelse(bh);
 912                                 return -1;
 913                         }
 914
 915                         next_block = btrfsic_block_lookup_or_add(
 916                                         state, &tmp_next_block_ctx,
 917                                         additional_string, 1, 1, 0,
 918                                         mirror_num, NULL);
 919                         if (NULL == next_block) {
 920                                 btrfsic_release_block_ctx(&tmp_next_block_ctx);
 921                                 brelse(bh);
 922                                 return -1;
 923                         }
 924
 925                         next_block->disk_key = tmp_disk_key;
 926                         next_block->generation = BTRFSIC_GENERATION_UNKNOWN;
 927                         l = btrfsic_block_link_lookup_or_add(
 928                                         state, &tmp_next_block_ctx,
 929                                         next_block, superblock_tmp,
 930                                         BTRFSIC_GENERATION_UNKNOWN);
 931                         btrfsic_release_block_ctx(&tmp_next_block_ctx);
 932                         if (NULL == l) {
 933                                 brelse(bh);
 934                                 return -1;
 935                         }
 936                 }
 937         }
 938         if (state->print_mask & BTRFSIC_PRINT_MASK_INITIAL_ALL_TREES)
 939                 btrfsic_dump_tree_sub(state, superblock_tmp, 0);
 940
 941         brelse(bh);
 942         return 0;
 943 }
 944
 945 static struct btrfsic_stack_frame *btrfsic_stack_frame_alloc(void)
 946 {
 947         struct btrfsic_stack_frame *sf;
 948
 949         sf = kzalloc(sizeof(*sf), GFP_NOFS);
 950         if (NULL == sf)
 951                 printk(KERN_INFO "btrfsic: alloc memory failed!\n");
 952         else
 953                 sf->magic = BTRFSIC_BLOCK_STACK_FRAME_MAGIC_NUMBER;
 954         return sf;
 955 }
 956
 957 static void btrfsic_stack_frame_free(struct btrfsic_stack_frame *sf)
 958 {
 959         BUG_ON(!(NULL == sf ||
 960                  BTRFSIC_BLOCK_STACK_FRAME_MAGIC_NUMBER == sf->magic));
 961         kfree(sf);
 962 }
 963
 964 static int btrfsic_process_metablock(
 965                 struct btrfsic_state *state,
 966                 struct btrfsic_block *const first_block,
 967                 struct btrfsic_block_data_ctx *const first_block_ctx,
 968                 int first_limit_nesting, int force_iodone_flag)
 969 {
 970         struct btrfsic_stack_frame initial_stack_frame = { 0 };
 971         struct btrfsic_stack_frame *sf;
 972         struct btrfsic_stack_frame *next_stack;
 973         struct btrfs_header *const first_hdr =
 974                 (struct btrfs_header *)first_block_ctx->datav[0];
 975
 976         BUG_ON(!first_hdr);
 977         sf = &initial_stack_frame;
 978         sf->error = 0;
 979         sf->i = -1;
 980         sf->limit_nesting = first_limit_nesting;
 981         sf->block = first_block;
 982         sf->block_ctx = first_block_ctx;
 983         sf->next_block = NULL;
 984         sf->hdr = first_hdr;
 985         sf->prev = NULL;
 986
 987 continue_with_new_stack_frame:
 988         sf->block->generation = le64_to_cpu(sf->hdr->generation);
 989         if (0 == sf->hdr->level) {
 990                 struct btrfs_leaf *const leafhdr =
 991                     (struct btrfs_leaf *)sf->hdr;
 992
 993                 if (-1 == sf->i) {
 994                         sf->nr = btrfs_stack_header_nritems(&leafhdr->header);
 995
 996                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
 997                                 printk(KERN_INFO
 998                                        "leaf %llu items %d generation %llu"
 999                                        " owner %llu\n",
1000                                        sf->block_ctx->start, sf->nr,
1001                                        btrfs_stack_header_generation(
1002                                                &leafhdr->header),
1003                                        btrfs_stack_header_owner(
1004                                                &leafhdr->header));
1005                 }
1006
1007 continue_with_current_leaf_stack_frame:
1008                 if (0 == sf->num_copies || sf->mirror_num > sf->num_copies) {
1009                         sf->i++;
1010                         sf->num_copies = 0;
1011                 }
1012
1013                 if (sf->i < sf->nr) {
1014                         struct btrfs_item disk_item;
1015                         u32 disk_item_offset =
1016                                 (uintptr_t)(leafhdr->items + sf->i) -
1017                                 (uintptr_t)leafhdr;
1018                         struct btrfs_disk_key *disk_key;
1019                         u8 type;
1020                         u32 item_offset;
1021                         u32 item_size;
1022
1023                         if (disk_item_offset + sizeof(struct btrfs_item) >
1024                             sf->block_ctx->len) {
1025 leaf_item_out_of_bounce_error:
1026                                 printk(KERN_INFO
1027                                        "btrfsic: leaf item out of bounce at logical %llu, dev %s\n",
1028                                        sf->block_ctx->start,
1029                                        sf->block_ctx->dev->name);
1030                                 goto one_stack_frame_backwards;
1031                         }
1032                         btrfsic_read_from_block_data(sf->block_ctx,
1033                                                      &disk_item,
1034                                                      disk_item_offset,
1035                                                      sizeof(struct btrfs_item));
1036                         item_offset = btrfs_stack_item_offset(&disk_item);
1037                         item_size = btrfs_stack_item_size(&disk_item);
1038                         disk_key = &disk_item.key;
1039                         type = btrfs_disk_key_type(disk_key);
1040
1041                         if (BTRFS_ROOT_ITEM_KEY == type) {
1042                                 struct btrfs_root_item root_item;
1043                                 u32 root_item_offset;
1044                                 u64 next_bytenr;
1045
1046                                 root_item_offset = item_offset +
1047                                         offsetof(struct btrfs_leaf, items);
1048                                 if (root_item_offset + item_size >
1049                                     sf->block_ctx->len)
1050                                         goto leaf_item_out_of_bounce_error;
1051                                 btrfsic_read_from_block_data(
1052                                         sf->block_ctx, &root_item,
1053                                         root_item_offset,
1054                                         item_size);
1055                                 next_bytenr = btrfs_root_bytenr(&root_item);
1056
1057                                 sf->error =
1058                                     btrfsic_create_link_to_next_block(
1059                                                 state,
1060                                                 sf->block,
1061                                                 sf->block_ctx,
1062                                                 next_bytenr,
1063                                                 sf->limit_nesting,
1064                                                 &sf->next_block_ctx,
1065                                                 &sf->next_block,
1066                                                 force_iodone_flag,
1067                                                 &sf->num_copies,
1068                                                 &sf->mirror_num,
1069                                                 disk_key,
1070                                                 btrfs_root_generation(
1071                                                 &root_item));
1072                                 if (sf->error)
1073                                         goto one_stack_frame_backwards;
1074
1075                                 if (NULL != sf->next_block) {
1076                                         struct btrfs_header *const next_hdr =
1077                                             (struct btrfs_header *)
1078                                             sf->next_block_ctx.datav[0];
1079
1080                                         next_stack =
1081                                             btrfsic_stack_frame_alloc();
1082                                         if (NULL == next_stack) {
1083                                                 sf->error = -1;
1084                                                 btrfsic_release_block_ctx(
1085                                                                 &sf->
1086                                                                 next_block_ctx);
1087                                                 goto one_stack_frame_backwards;
1088                                         }
1089
1090                                         next_stack->i = -1;
1091                                         next_stack->block = sf->next_block;
1092                                         next_stack->block_ctx =
1093                                             &sf->next_block_ctx;
1094                                         next_stack->next_block = NULL;
1095                                         next_stack->hdr = next_hdr;
1096                                         next_stack->limit_nesting =
1097                                             sf->limit_nesting - 1;
1098                                         next_stack->prev = sf;
1099                                         sf = next_stack;
1100                                         goto continue_with_new_stack_frame;
1101                                 }
1102                         } else if (BTRFS_EXTENT_DATA_KEY == type &&
1103                                    state->include_extent_data) {
1104                                 sf->error = btrfsic_handle_extent_data(
1105                                                 state,
1106                                                 sf->block,
1107                                                 sf->block_ctx,
1108                                                 item_offset,
1109                                                 force_iodone_flag);
1110                                 if (sf->error)
1111                                         goto one_stack_frame_backwards;
1112                         }
1113
1114                         goto continue_with_current_leaf_stack_frame;
1115                 }
1116         } else {
1117                 struct btrfs_node *const nodehdr = (struct btrfs_node *)sf->hdr;
1118
1119                 if (-1 == sf->i) {
1120                         sf->nr = btrfs_stack_header_nritems(&nodehdr->header);
1121
1122                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1123                                 printk(KERN_INFO "node %llu level %d items %d"
1124                                        " generation %llu owner %llu\n",
1125                                        sf->block_ctx->start,
1126                                        nodehdr->header.level, sf->nr,
1127                                        btrfs_stack_header_generation(
1128                                        &nodehdr->header),
1129                                        btrfs_stack_header_owner(
1130                                        &nodehdr->header));
1131                 }
1132
1133 continue_with_current_node_stack_frame:
1134                 if (0 == sf->num_copies || sf->mirror_num > sf->num_copies) {
1135                         sf->i++;
1136                         sf->num_copies = 0;
1137                 }
1138
1139                 if (sf->i < sf->nr) {
1140                         struct btrfs_key_ptr key_ptr;
1141                         u32 key_ptr_offset;
1142                         u64 next_bytenr;
1143
1144                         key_ptr_offset = (uintptr_t)(nodehdr->ptrs + sf->i) -
1145                                           (uintptr_t)nodehdr;
1146                         if (key_ptr_offset + sizeof(struct btrfs_key_ptr) >
1147                             sf->block_ctx->len) {
1148                                 printk(KERN_INFO
1149                                        "btrfsic: node item out of bounce at logical %llu, dev %s\n",
1150                                        sf->block_ctx->start,
1151                                        sf->block_ctx->dev->name);
1152                                 goto one_stack_frame_backwards;
1153                         }
1154                         btrfsic_read_from_block_data(
1155                                 sf->block_ctx, &key_ptr, key_ptr_offset,
1156                                 sizeof(struct btrfs_key_ptr));
1157                         next_bytenr = btrfs_stack_key_blockptr(&key_ptr);
1158
1159                         sf->error = btrfsic_create_link_to_next_block(
1160                                         state,
1161                                         sf->block,
1162                                         sf->block_ctx,
1163                                         next_bytenr,
1164                                         sf->limit_nesting,
1165                                         &sf->next_block_ctx,
1166                                         &sf->next_block,
1167                                         force_iodone_flag,
1168                                         &sf->num_copies,
1169                                         &sf->mirror_num,
1170                                         &key_ptr.key,
1171                                         btrfs_stack_key_generation(&key_ptr));
1172                         if (sf->error)
1173                                 goto one_stack_frame_backwards;
1174
1175                         if (NULL != sf->next_block) {
1176                                 struct btrfs_header *const next_hdr =
1177                                     (struct btrfs_header *)
1178                                     sf->next_block_ctx.datav[0];
1179
1180                                 next_stack = btrfsic_stack_frame_alloc();
1181                                 if (NULL == next_stack) {
1182                                         sf->error = -1;
1183                                         goto one_stack_frame_backwards;
1184                                 }
1185
1186                                 next_stack->i = -1;
1187                                 next_stack->block = sf->next_block;
1188                                 next_stack->block_ctx = &sf->next_block_ctx;
1189                                 next_stack->next_block = NULL;
1190                                 next_stack->hdr = next_hdr;
1191                                 next_stack->limit_nesting =
1192                                     sf->limit_nesting - 1;
1193                                 next_stack->prev = sf;
1194                                 sf = next_stack;
1195                                 goto continue_with_new_stack_frame;
1196                         }
1197
1198                         goto continue_with_current_node_stack_frame;
1199                 }
1200         }
1201
1202 one_stack_frame_backwards:
1203         if (NULL != sf->prev) {
1204                 struct btrfsic_stack_frame *const prev = sf->prev;
1205
1206                 /* the one for the initial block is freed in the caller */
1207                 btrfsic_release_block_ctx(sf->block_ctx);
1208
1209                 if (sf->error) {
1210                         prev->error = sf->error;
1211                         btrfsic_stack_frame_free(sf);
1212                         sf = prev;
1213                         goto one_stack_frame_backwards;
1214                 }
1215
1216                 btrfsic_stack_frame_free(sf);
1217                 sf = prev;
1218                 goto continue_with_new_stack_frame;
1219         } else {
1220                 BUG_ON(&initial_stack_frame != sf);
1221         }
1222
1223         return sf->error;
1224 }
1225
1226 static void btrfsic_read_from_block_data(
1227         struct btrfsic_block_data_ctx *block_ctx,
1228         void *dstv, u32 offset, size_t len)
1229 {
1230         size_t cur;
1231         size_t offset_in_page;
1232         char *kaddr;
1233         char *dst = (char *)dstv;
1234         size_t start_offset = block_ctx->start & ((u64)PAGE_SIZE - 1);
1235         unsigned long i = (start_offset + offset) >> PAGE_SHIFT;
1236
1237         WARN_ON(offset + len > block_ctx->len);
1238         offset_in_page = (start_offset + offset) & (PAGE_SIZE - 1);
1239
1240         while (len > 0) {
1241                 cur = min(len, ((size_t)PAGE_SIZE - offset_in_page));
1242                 BUG_ON(i >= DIV_ROUND_UP(block_ctx->len, PAGE_SIZE));
1243                 kaddr = block_ctx->datav[i];
1244                 memcpy(dst, kaddr + offset_in_page, cur);
1245
1246                 dst += cur;
1247                 len -= cur;
1248                 offset_in_page = 0;
1249                 i++;
1250         }
1251 }
1252
1253 static int btrfsic_create_link_to_next_block(
1254                 struct btrfsic_state *state,
1255                 struct btrfsic_block *block,
1256                 struct btrfsic_block_data_ctx *block_ctx,
1257                 u64 next_bytenr,
1258                 int limit_nesting,
1259                 struct btrfsic_block_data_ctx *next_block_ctx,
1260                 struct btrfsic_block **next_blockp,
1261                 int force_iodone_flag,
1262                 int *num_copiesp, int *mirror_nump,
1263                 struct btrfs_disk_key *disk_key,
1264                 u64 parent_generation)
1265 {
1266         struct btrfsic_block *next_block = NULL;
1267         int ret;
1268         struct btrfsic_block_link *l;
1269         int did_alloc_block_link;
1270         int block_was_created;
1271
1272         *next_blockp = NULL;
1273         if (0 == *num_copiesp) {
1274                 *num_copiesp =
1275                     btrfs_num_copies(state->root->fs_info,
1276                                      next_bytenr, state->metablock_size);
1277                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
1278                         printk(KERN_INFO "num_copies(log_bytenr=%llu) = %d\n",
1279                                next_bytenr, *num_copiesp);
1280                 *mirror_nump = 1;
1281         }
1282
1283         if (*mirror_nump > *num_copiesp)
1284                 return 0;
1285
1286         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1287                 printk(KERN_INFO
1288                        "btrfsic_create_link_to_next_block(mirror_num=%d)\n",
1289                        *mirror_nump);
1290         ret = btrfsic_map_block(state, next_bytenr,
1291                                 state->metablock_size,
1292                                 next_block_ctx, *mirror_nump);
1293         if (ret) {
1294                 printk(KERN_INFO
1295                        "btrfsic: btrfsic_map_block(@%llu, mirror=%d) failed!\n",
1296                        next_bytenr, *mirror_nump);
1297                 btrfsic_release_block_ctx(next_block_ctx);
1298                 *next_blockp = NULL;
1299                 return -1;
1300         }
1301
1302         next_block = btrfsic_block_lookup_or_add(state,
1303                                                  next_block_ctx, "referenced ",
1304                                                  1, force_iodone_flag,
1305                                                  !force_iodone_flag,
1306                                                  *mirror_nump,
1307                                                  &block_was_created);
1308         if (NULL == next_block) {
1309                 btrfsic_release_block_ctx(next_block_ctx);
1310                 *next_blockp = NULL;
1311                 return -1;
1312         }
1313         if (block_was_created) {
1314                 l = NULL;
1315                 next_block->generation = BTRFSIC_GENERATION_UNKNOWN;
1316         } else {
1317                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE) {
1318                         if (next_block->logical_bytenr != next_bytenr &&
1319                             !(!next_block->is_metadata &&
1320                               0 == next_block->logical_bytenr))
1321                                 printk(KERN_INFO
1322                                        "Referenced block @%llu (%s/%llu/%d) found in hash table, %c, bytenr mismatch (!= stored %llu).\n",
1323                                        next_bytenr, next_block_ctx->dev->name,
1324                                        next_block_ctx->dev_bytenr, *mirror_nump,
1325                                        btrfsic_get_block_type(state,
1326                                                               next_block),
1327                                        next_block->logical_bytenr);
1328                         else
1329                                 printk(KERN_INFO
1330                                        "Referenced block @%llu (%s/%llu/%d) found in hash table, %c.\n",
1331                                        next_bytenr, next_block_ctx->dev->name,
1332                                        next_block_ctx->dev_bytenr, *mirror_nump,
1333                                        btrfsic_get_block_type(state,
1334                                                               next_block));
1335                 }
1336                 next_block->logical_bytenr = next_bytenr;
1337
1338                 next_block->mirror_num = *mirror_nump;
1339                 l = btrfsic_block_link_hashtable_lookup(
1340                                 next_block_ctx->dev->bdev,
1341                                 next_block_ctx->dev_bytenr,
1342                                 block_ctx->dev->bdev,
1343                                 block_ctx->dev_bytenr,
1344                                 &state->block_link_hashtable);
1345         }
1346
1347         next_block->disk_key = *disk_key;
1348         if (NULL == l) {
1349                 l = btrfsic_block_link_alloc();
1350                 if (NULL == l) {
1351                         printk(KERN_INFO "btrfsic: error, kmalloc failed!\n");
1352                         btrfsic_release_block_ctx(next_block_ctx);
1353                         *next_blockp = NULL;
1354                         return -1;
1355                 }
1356
1357                 did_alloc_block_link = 1;
1358                 l->block_ref_to = next_block;
1359                 l->block_ref_from = block;
1360                 l->ref_cnt = 1;
1361                 l->parent_generation = parent_generation;
1362
1363                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1364                         btrfsic_print_add_link(state, l);
1365
1366                 list_add(&l->node_ref_to, &block->ref_to_list);
1367                 list_add(&l->node_ref_from, &next_block->ref_from_list);
1368
1369                 btrfsic_block_link_hashtable_add(l,
1370                                                  &state->block_link_hashtable);
1371         } else {
1372                 did_alloc_block_link = 0;
1373                 if (0 == limit_nesting) {
1374                         l->ref_cnt++;
1375                         l->parent_generation = parent_generation;
1376                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1377                                 btrfsic_print_add_link(state, l);
1378                 }
1379         }
1380
1381         if (limit_nesting > 0 && did_alloc_block_link) {
1382                 ret = btrfsic_read_block(state, next_block_ctx);
1383                 if (ret < (int)next_block_ctx->len) {
1384                         printk(KERN_INFO
1385                                "btrfsic: read block @logical %llu failed!\n",
1386                                next_bytenr);
1387                         btrfsic_release_block_ctx(next_block_ctx);
1388                         *next_blockp = NULL;
1389                         return -1;
1390                 }
1391
1392                 *next_blockp = next_block;
1393         } else {
1394                 *next_blockp = NULL;
1395         }
1396         (*mirror_nump)++;
1397
1398         return 0;
1399 }
1400
1401 static int btrfsic_handle_extent_data(
1402                 struct btrfsic_state *state,
1403                 struct btrfsic_block *block,
1404                 struct btrfsic_block_data_ctx *block_ctx,
1405                 u32 item_offset, int force_iodone_flag)
1406 {
1407         int ret;
1408         struct btrfs_file_extent_item file_extent_item;
1409         u64 file_extent_item_offset;
1410         u64 next_bytenr;
1411         u64 num_bytes;
1412         u64 generation;
1413         struct btrfsic_block_link *l;
1414
1415         file_extent_item_offset = offsetof(struct btrfs_leaf, items) +
1416                                   item_offset;
1417         if (file_extent_item_offset +
1418             offsetof(struct btrfs_file_extent_item, disk_num_bytes) >
1419             block_ctx->len) {
1420                 printk(KERN_INFO
1421                        "btrfsic: file item out of bounce at logical %llu, dev %s\n",
1422                        block_ctx->start, block_ctx->dev->name);
1423                 return -1;
1424         }
1425
1426         btrfsic_read_from_block_data(block_ctx, &file_extent_item,
1427                 file_extent_item_offset,
1428                 offsetof(struct btrfs_file_extent_item, disk_num_bytes));
1429         if (BTRFS_FILE_EXTENT_REG != file_extent_item.type ||
1430             btrfs_stack_file_extent_disk_bytenr(&file_extent_item) == 0) {
1431                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERY_VERBOSE)
1432                         printk(KERN_INFO "extent_data: type %u, disk_bytenr = %llu\n",
1433                                file_extent_item.type,
1434                                btrfs_stack_file_extent_disk_bytenr(
1435                                &file_extent_item));
1436                 return 0;
1437         }
1438
1439         if (file_extent_item_offset + sizeof(struct btrfs_file_extent_item) >
1440             block_ctx->len) {
1441                 printk(KERN_INFO
1442                        "btrfsic: file item out of bounce at logical %llu, dev %s\n",
1443                        block_ctx->start, block_ctx->dev->name);
1444                 return -1;
1445         }
1446         btrfsic_read_from_block_data(block_ctx, &file_extent_item,
1447                                      file_extent_item_offset,
1448                                      sizeof(struct btrfs_file_extent_item));
1449         next_bytenr = btrfs_stack_file_extent_disk_bytenr(&file_extent_item);
1450         if (btrfs_stack_file_extent_compression(&file_extent_item) ==
1451             BTRFS_COMPRESS_NONE) {
1452                 next_bytenr += btrfs_stack_file_extent_offset(&file_extent_item);
1453                 num_bytes = btrfs_stack_file_extent_num_bytes(&file_extent_item);
1454         } else {
1455                 num_bytes = btrfs_stack_file_extent_disk_num_bytes(&file_extent_item);
1456         }
1457         generation = btrfs_stack_file_extent_generation(&file_extent_item);
1458
1459         if (state->print_mask & BTRFSIC_PRINT_MASK_VERY_VERBOSE)
1460                 printk(KERN_INFO "extent_data: type %u, disk_bytenr = %llu,"
1461                        " offset = %llu, num_bytes = %llu\n",
1462                        file_extent_item.type,
1463                        btrfs_stack_file_extent_disk_bytenr(&file_extent_item),
1464                        btrfs_stack_file_extent_offset(&file_extent_item),
1465                        num_bytes);
1466         while (num_bytes > 0) {
1467                 u32 chunk_len;
1468                 int num_copies;
1469                 int mirror_num;
1470
1471                 if (num_bytes > state->datablock_size)
1472                         chunk_len = state->datablock_size;
1473                 else
1474                         chunk_len = num_bytes;
1475
1476                 num_copies =
1477                     btrfs_num_copies(state->root->fs_info,
1478                                      next_bytenr, state->datablock_size);
1479                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
1480                         printk(KERN_INFO "num_copies(log_bytenr=%llu) = %d\n",
1481                                next_bytenr, num_copies);
1482                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
1483                         struct btrfsic_block_data_ctx next_block_ctx;
1484                         struct btrfsic_block *next_block;
1485                         int block_was_created;
1486
1487                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1488                                 printk(KERN_INFO "btrfsic_handle_extent_data("
1489                                        "mirror_num=%d)\n", mirror_num);
1490                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERY_VERBOSE)
1491                                 printk(KERN_INFO
1492                                        "\tdisk_bytenr = %llu, num_bytes %u\n",
1493                                        next_bytenr, chunk_len);
1494                         ret = btrfsic_map_block(state, next_bytenr,
1495                                                 chunk_len, &next_block_ctx,
1496                                                 mirror_num);
1497                         if (ret) {
1498                                 printk(KERN_INFO
1499                                        "btrfsic: btrfsic_map_block(@%llu,"
1500                                        " mirror=%d) failed!\n",
1501                                        next_bytenr, mirror_num);
1502                                 return -1;
1503                         }
1504
1505                         next_block = btrfsic_block_lookup_or_add(
1506                                         state,
1507                                         &next_block_ctx,
1508                                         "referenced ",
1509                                         0,
1510                                         force_iodone_flag,
1511                                         !force_iodone_flag,
1512                                         mirror_num,
1513                                         &block_was_created);
1514                         if (NULL == next_block) {
1515                                 printk(KERN_INFO
1516                                        "btrfsic: error, kmalloc failed!\n");
1517                                 btrfsic_release_block_ctx(&next_block_ctx);
1518                                 return -1;
1519                         }
1520                         if (!block_was_created) {
1521                                 if ((state->print_mask &
1522                                      BTRFSIC_PRINT_MASK_VERBOSE) &&
1523                                     next_block->logical_bytenr != next_bytenr &&
1524                                     !(!next_block->is_metadata &&
1525                                       0 == next_block->logical_bytenr)) {
1526                                         printk(KERN_INFO
1527                                                "Referenced block"
1528                                                " @%llu (%s/%llu/%d)"
1529                                                " found in hash table, D,"
1530                                                " bytenr mismatch"
1531                                                " (!= stored %llu).\n",
1532                                                next_bytenr,
1533                                                next_block_ctx.dev->name,
1534                                                next_block_ctx.dev_bytenr,
1535                                                mirror_num,
1536                                                next_block->logical_bytenr);
1537                                 }
1538                                 next_block->logical_bytenr = next_bytenr;
1539                                 next_block->mirror_num = mirror_num;
1540                         }
1541
1542                         l = btrfsic_block_link_lookup_or_add(state,
1543                                                              &next_block_ctx,
1544                                                              next_block, block,
1545                                                              generation);
1546                         btrfsic_release_block_ctx(&next_block_ctx);
1547                         if (NULL == l)
1548                                 return -1;
1549                 }
1550
1551                 next_bytenr += chunk_len;
1552                 num_bytes -= chunk_len;
1553         }
1554
1555         return 0;
1556 }
1557
1558 static int btrfsic_map_block(struct btrfsic_state *state, u64 bytenr, u32 len,
1559                              struct btrfsic_block_data_ctx *block_ctx_out,
1560                              int mirror_num)
1561 {
1562         int ret;
1563         u64 length;
1564         struct btrfs_bio *multi = NULL;
1565         struct btrfs_device *device;
1566
1567         length = len;
1568         ret = btrfs_map_block(state->root->fs_info, READ,
1569                               bytenr, &length, &multi, mirror_num);
1570
1571         if (ret) {
1572                 block_ctx_out->start = 0;
1573                 block_ctx_out->dev_bytenr = 0;
1574                 block_ctx_out->len = 0;
1575                 block_ctx_out->dev = NULL;
1576                 block_ctx_out->datav = NULL;
1577                 block_ctx_out->pagev = NULL;
1578                 block_ctx_out->mem_to_free = NULL;
1579
1580                 return ret;
1581         }
1582
1583         device = multi->stripes[0].dev;
1584         block_ctx_out->dev = btrfsic_dev_state_lookup(device->bdev);
1585         block_ctx_out->dev_bytenr = multi->stripes[0].physical;
1586         block_ctx_out->start = bytenr;
1587         block_ctx_out->len = len;
1588         block_ctx_out->datav = NULL;
1589         block_ctx_out->pagev = NULL;
1590         block_ctx_out->mem_to_free = NULL;
1591
1592         kfree(multi);
1593         if (NULL == block_ctx_out->dev) {
1594                 ret = -ENXIO;
1595                 printk(KERN_INFO "btrfsic: error, cannot lookup dev (#1)!\n");
1596         }
1597
1598         return ret;
1599 }
1600
1601 static void btrfsic_release_block_ctx(struct btrfsic_block_data_ctx *block_ctx)
1602 {
1603         if (block_ctx->mem_to_free) {
1604                 unsigned int num_pages;
1605
1606                 BUG_ON(!block_ctx->datav);
1607                 BUG_ON(!block_ctx->pagev);
1608                 num_pages = (block_ctx->len + (u64)PAGE_SIZE - 1) >>
1609                             PAGE_SHIFT;
1610                 while (num_pages > 0) {
1611                         num_pages--;
1612                         if (block_ctx->datav[num_pages]) {
1613                                 kunmap(block_ctx->pagev[num_pages]);
1614                                 block_ctx->datav[num_pages] = NULL;
1615                         }
1616                         if (block_ctx->pagev[num_pages]) {
1617                                 __free_page(block_ctx->pagev[num_pages]);
1618                                 block_ctx->pagev[num_pages] = NULL;
1619                         }
1620                 }
1621
1622                 kfree(block_ctx->mem_to_free);
1623                 block_ctx->mem_to_free = NULL;
1624                 block_ctx->pagev = NULL;
1625                 block_ctx->datav = NULL;
1626         }
1627 }
1628
1629 static int btrfsic_read_block(struct btrfsic_state *state,
1630                               struct btrfsic_block_data_ctx *block_ctx)
1631 {
1632         unsigned int num_pages;
1633         unsigned int i;
1634         u64 dev_bytenr;
1635         int ret;
1636
1637         BUG_ON(block_ctx->datav);
1638         BUG_ON(block_ctx->pagev);
1639         BUG_ON(block_ctx->mem_to_free);
1640         if (block_ctx->dev_bytenr & ((u64)PAGE_SIZE - 1)) {
1641                 printk(KERN_INFO
1642                        "btrfsic: read_block() with unaligned bytenr %llu\n",
1643                        block_ctx->dev_bytenr);
1644                 return -1;
1645         }
1646
1647         num_pages = (block_ctx->len + (u64)PAGE_SIZE - 1) >>
1648                     PAGE_SHIFT;
1649         block_ctx->mem_to_free = kzalloc((sizeof(*block_ctx->datav) +
1650                                           sizeof(*block_ctx->pagev)) *
1651                                          num_pages, GFP_NOFS);
1652         if (!block_ctx->mem_to_free)
1653                 return -ENOMEM;
1654         block_ctx->datav = block_ctx->mem_to_free;
1655         block_ctx->pagev = (struct page **)(block_ctx->datav + num_pages);
1656         for (i = 0; i < num_pages; i++) {
1657                 block_ctx->pagev[i] = alloc_page(GFP_NOFS);
1658                 if (!block_ctx->pagev[i])
1659                         return -1;
1660         }
1661
1662         dev_bytenr = block_ctx->dev_bytenr;
1663         for (i = 0; i < num_pages;) {
1664                 struct bio *bio;
1665                 unsigned int j;
1666
1667                 bio = btrfs_io_bio_alloc(GFP_NOFS, num_pages - i);
1668                 if (!bio) {
1669                         printk(KERN_INFO
1670                                "btrfsic: bio_alloc() for %u pages failed!\n",
1671                                num_pages - i);
1672                         return -1;
1673                 }
1674                 bio->bi_bdev = block_ctx->dev->bdev;
1675                 bio->bi_iter.bi_sector = dev_bytenr >> 9;
1676                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
1677
1678                 for (j = i; j < num_pages; j++) {
1679                         ret = bio_add_page(bio, block_ctx->pagev[j],
1680                                            PAGE_SIZE, 0);
1681                         if (PAGE_SIZE != ret)
1682                                 break;
1683                 }
1684                 if (j == i) {
1685                         printk(KERN_INFO
1686                                "btrfsic: error, failed to add a single page!\n");
1687                         return -1;
1688                 }
1689                 if (submit_bio_wait(bio)) {
1690                         printk(KERN_INFO
1691                                "btrfsic: read error at logical %llu dev %s!\n",
1692                                block_ctx->start, block_ctx->dev->name);
1693                         bio_put(bio);
1694                         return -1;
1695                 }
1696                 bio_put(bio);
1697                 dev_bytenr += (j - i) * PAGE_SIZE;
1698                 i = j;
1699         }
1700         for (i = 0; i < num_pages; i++) {
1701                 block_ctx->datav[i] = kmap(block_ctx->pagev[i]);
1702                 if (!block_ctx->datav[i]) {
1703                         printk(KERN_INFO "btrfsic: kmap() failed (dev %s)!\n",
1704                                block_ctx->dev->name);
1705                         return -1;
1706                 }
1707         }
1708
1709         return block_ctx->len;
1710 }
1711
1712 static void btrfsic_dump_database(struct btrfsic_state *state)
1713 {
1714         const struct btrfsic_block *b_all;
1715
1716         BUG_ON(NULL == state);
1717
1718         printk(KERN_INFO "all_blocks_list:\n");
1719         list_for_each_entry(b_all, &state->all_blocks_list, all_blocks_node) {
1720                 const struct btrfsic_block_link *l;
1721
1722                 printk(KERN_INFO "%c-block @%llu (%s/%llu/%d)\n",
1723                        btrfsic_get_block_type(state, b_all),
1724                        b_all->logical_bytenr, b_all->dev_state->name,
1725                        b_all->dev_bytenr, b_all->mirror_num);
1726
1727                 list_for_each_entry(l, &b_all->ref_to_list, node_ref_to) {
1728                         printk(KERN_INFO " %c @%llu (%s/%llu/%d)"
1729                                " refers %u* to"
1730                                " %c @%llu (%s/%llu/%d)\n",
1731                                btrfsic_get_block_type(state, b_all),
1732                                b_all->logical_bytenr, b_all->dev_state->name,
1733                                b_all->dev_bytenr, b_all->mirror_num,
1734                                l->ref_cnt,
1735                                btrfsic_get_block_type(state, l->block_ref_to),
1736                                l->block_ref_to->logical_bytenr,
1737                                l->block_ref_to->dev_state->name,
1738                                l->block_ref_to->dev_bytenr,
1739                                l->block_ref_to->mirror_num);
1740                 }
1741
1742                 list_for_each_entry(l, &b_all->ref_from_list, node_ref_from) {
1743                         printk(KERN_INFO " %c @%llu (%s/%llu/%d)"
1744                                " is ref %u* from"
1745                                " %c @%llu (%s/%llu/%d)\n",
1746                                btrfsic_get_block_type(state, b_all),
1747                                b_all->logical_bytenr, b_all->dev_state->name,
1748                                b_all->dev_bytenr, b_all->mirror_num,
1749                                l->ref_cnt,
1750                                btrfsic_get_block_type(state, l->block_ref_from),
1751                                l->block_ref_from->logical_bytenr,
1752                                l->block_ref_from->dev_state->name,
1753                                l->block_ref_from->dev_bytenr,
1754                                l->block_ref_from->mirror_num);
1755                 }
1756
1757                 printk(KERN_INFO "\n");
1758         }
1759 }
1760
1761 /*
1762  * Test whether the disk block contains a tree block (leaf or node)
1763  * (note that this test fails for the super block)
1764  */
1765 static int btrfsic_test_for_metadata(struct btrfsic_state *state,
1766                                      char **datav, unsigned int num_pages)
1767 {
1768         struct btrfs_header *h;
1769         u8 csum[BTRFS_CSUM_SIZE];
1770         u32 crc = ~(u32)0;
1771         unsigned int i;
1772
1773         if (num_pages * PAGE_SIZE < state->metablock_size)
1774                 return 1; /* not metadata */
1775         num_pages = state->metablock_size >> PAGE_SHIFT;
1776         h = (struct btrfs_header *)datav[0];
1777
1778         if (memcmp(h->fsid, state->root->fs_info->fsid, BTRFS_UUID_SIZE))
1779                 return 1;
1780
1781         for (i = 0; i < num_pages; i++) {
1782                 u8 *data = i ? datav[i] : (datav[i] + BTRFS_CSUM_SIZE);
1783                 size_t sublen = i ? PAGE_SIZE :
1784                                     (PAGE_SIZE - BTRFS_CSUM_SIZE);
1785
1786                 crc = btrfs_crc32c(crc, data, sublen);
1787         }
1788         btrfs_csum_final(crc, csum);
1789         if (memcmp(csum, h->csum, state->csum_size))
1790                 return 1;
1791
1792         return 0; /* is metadata */
1793 }
1794
1795 static void btrfsic_process_written_block(struct btrfsic_dev_state *dev_state,
1796                                           u64 dev_bytenr, char **mapped_datav,
1797                                           unsigned int num_pages,
1798                                           struct bio *bio, int *bio_is_patched,
1799                                           struct buffer_head *bh,
1800                                           int submit_bio_bh_rw)
1801 {
1802         int is_metadata;
1803         struct btrfsic_block *block;
1804         struct btrfsic_block_data_ctx block_ctx;
1805         int ret;
1806         struct btrfsic_state *state = dev_state->state;
1807         struct block_device *bdev = dev_state->bdev;
1808         unsigned int processed_len;
1809
1810         if (NULL != bio_is_patched)
1811                 *bio_is_patched = 0;
1812
1813 again:
1814         if (num_pages == 0)
1815                 return;
1816
1817         processed_len = 0;
1818         is_metadata = (0 == btrfsic_test_for_metadata(state, mapped_datav,
1819                                                       num_pages));
1820
1821         block = btrfsic_block_hashtable_lookup(bdev, dev_bytenr,
1822                                                &state->block_hashtable);
1823         if (NULL != block) {
1824                 u64 bytenr = 0;
1825                 struct btrfsic_block_link *l, *tmp;
1826
1827                 if (block->is_superblock) {
1828                         bytenr = btrfs_super_bytenr((struct btrfs_super_block *)
1829                                                     mapped_datav[0]);
1830                         if (num_pages * PAGE_SIZE <
1831                             BTRFS_SUPER_INFO_SIZE) {
1832                                 printk(KERN_INFO
1833                                        "btrfsic: cannot work with too short bios!\n");
1834                                 return;
1835                         }
1836                         is_metadata = 1;
1837                         BUG_ON(BTRFS_SUPER_INFO_SIZE & (PAGE_SIZE - 1));
1838                         processed_len = BTRFS_SUPER_INFO_SIZE;
1839                         if (state->print_mask &
1840                             BTRFSIC_PRINT_MASK_TREE_BEFORE_SB_WRITE) {
1841                                 printk(KERN_INFO
1842                                        "[before new superblock is written]:\n");
1843                                 btrfsic_dump_tree_sub(state, block, 0);
1844                         }
1845                 }
1846                 if (is_metadata) {
1847                         if (!block->is_superblock) {
1848                                 if (num_pages * PAGE_SIZE <
1849                                     state->metablock_size) {
1850                                         printk(KERN_INFO
1851                                                "btrfsic: cannot work with too short bios!\n");
1852                                         return;
1853                                 }
1854                                 processed_len = state->metablock_size;
1855                                 bytenr = btrfs_stack_header_bytenr(
1856                                                 (struct btrfs_header *)
1857                                                 mapped_datav[0]);
1858                                 btrfsic_cmp_log_and_dev_bytenr(state, bytenr,
1859                                                                dev_state,
1860                                                                dev_bytenr);
1861                         }
1862                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE) {
1863                                 if (block->logical_bytenr != bytenr &&
1864                                     !(!block->is_metadata &&
1865                                       block->logical_bytenr == 0))
1866                                         printk(KERN_INFO
1867                                                "Written block @%llu (%s/%llu/%d) found in hash table, %c, bytenr mismatch (!= stored %llu).\n",
1868                                                bytenr, dev_state->name,
1869                                                dev_bytenr,
1870                                                block->mirror_num,
1871                                                btrfsic_get_block_type(state,
1872                                                                       block),
1873                                                block->logical_bytenr);
1874                                 else
1875                                         printk(KERN_INFO
1876                                                "Written block @%llu (%s/%llu/%d) found in hash table, %c.\n",
1877                                                bytenr, dev_state->name,
1878                                                dev_bytenr, block->mirror_num,
1879                                                btrfsic_get_block_type(state,
1880                                                                       block));
1881                         }
1882                         block->logical_bytenr = bytenr;
1883                 } else {
1884                         if (num_pages * PAGE_SIZE <
1885                             state->datablock_size) {
1886                                 printk(KERN_INFO
1887                                        "btrfsic: cannot work with too short bios!\n");
1888                                 return;
1889                         }
1890                         processed_len = state->datablock_size;
1891                         bytenr = block->logical_bytenr;
1892                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1893                                 printk(KERN_INFO
1894                                        "Written block @%llu (%s/%llu/%d)"
1895                                        " found in hash table, %c.\n",
1896                                        bytenr, dev_state->name, dev_bytenr,
1897                                        block->mirror_num,
1898                                        btrfsic_get_block_type(state, block));
1899                 }
1900
1901                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1902                         printk(KERN_INFO
1903                                "ref_to_list: %cE, ref_from_list: %cE\n",
1904                                list_empty(&block->ref_to_list) ? ' ' : '!',
1905                                list_empty(&block->ref_from_list) ? ' ' : '!');
1906                 if (btrfsic_is_block_ref_by_superblock(state, block, 0)) {
1907                         printk(KERN_INFO "btrfs: attempt to overwrite %c-block"
1908                                " @%llu (%s/%llu/%d), old(gen=%llu,"
1909                                " objectid=%llu, type=%d, offset=%llu),"
1910                                " new(gen=%llu),"
1911                                " which is referenced by most recent superblock"
1912                                " (superblockgen=%llu)!\n",
1913                                btrfsic_get_block_type(state, block), bytenr,
1914                                dev_state->name, dev_bytenr, block->mirror_num,
1915                                block->generation,
1916                                btrfs_disk_key_objectid(&block->disk_key),
1917                                block->disk_key.type,
1918                                btrfs_disk_key_offset(&block->disk_key),
1919                                btrfs_stack_header_generation(
1920                                        (struct btrfs_header *) mapped_datav[0]),
1921                                state->max_superblock_generation);
1922                         btrfsic_dump_tree(state);
1923                 }
1924
1925                 if (!block->is_iodone && !block->never_written) {
1926                         printk(KERN_INFO "btrfs: attempt to overwrite %c-block"
1927                                " @%llu (%s/%llu/%d), oldgen=%llu, newgen=%llu,"
1928                                " which is not yet iodone!\n",
1929                                btrfsic_get_block_type(state, block), bytenr,
1930                                dev_state->name, dev_bytenr, block->mirror_num,
1931                                block->generation,
1932                                btrfs_stack_header_generation(
1933                                        (struct btrfs_header *)
1934                                        mapped_datav[0]));
1935                         /* it would not be safe to go on */
1936                         btrfsic_dump_tree(state);
1937                         goto continue_loop;
1938                 }
1939
1940                 /*
1941                  * Clear all references of this block. Do not free
1942                  * the block itself even if is not referenced anymore
1943                  * because it still carries valuable information
1944                  * like whether it was ever written and IO completed.
1945                  */
1946                 list_for_each_entry_safe(l, tmp, &block->ref_to_list,
1947                                          node_ref_to) {
1948                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
1949                                 btrfsic_print_rem_link(state, l);
1950                         l->ref_cnt--;
1951                         if (0 == l->ref_cnt) {
1952                                 list_del(&l->node_ref_to);
1953                                 list_del(&l->node_ref_from);
1954                                 btrfsic_block_link_hashtable_remove(l);
1955                                 btrfsic_block_link_free(l);
1956                         }
1957                 }
1958
1959                 block_ctx.dev = dev_state;
1960                 block_ctx.dev_bytenr = dev_bytenr;
1961                 block_ctx.start = bytenr;
1962                 block_ctx.len = processed_len;
1963                 block_ctx.pagev = NULL;
1964                 block_ctx.mem_to_free = NULL;
1965                 block_ctx.datav = mapped_datav;
1966
1967                 if (is_metadata || state->include_extent_data) {
1968                         block->never_written = 0;
1969                         block->iodone_w_error = 0;
1970                         if (NULL != bio) {
1971                                 block->is_iodone = 0;
1972                                 BUG_ON(NULL == bio_is_patched);
1973                                 if (!*bio_is_patched) {
1974                                         block->orig_bio_bh_private =
1975                                             bio->bi_private;
1976                                         block->orig_bio_bh_end_io.bio =
1977                                             bio->bi_end_io;
1978                                         block->next_in_same_bio = NULL;
1979                                         bio->bi_private = block;
1980                                         bio->bi_end_io = btrfsic_bio_end_io;
1981                                         *bio_is_patched = 1;
1982                                 } else {
1983                                         struct btrfsic_block *chained_block =
1984                                             (struct btrfsic_block *)
1985                                             bio->bi_private;
1986
1987                                         BUG_ON(NULL == chained_block);
1988                                         block->orig_bio_bh_private =
1989                                             chained_block->orig_bio_bh_private;
1990                                         block->orig_bio_bh_end_io.bio =
1991                                             chained_block->orig_bio_bh_end_io.
1992                                             bio;
1993                                         block->next_in_same_bio = chained_block;
1994                                         bio->bi_private = block;
1995                                 }
1996                         } else if (NULL != bh) {
1997                                 block->is_iodone = 0;
1998                                 block->orig_bio_bh_private = bh->b_private;
1999                                 block->orig_bio_bh_end_io.bh = bh->b_end_io;
2000                                 block->next_in_same_bio = NULL;
2001                                 bh->b_private = block;
2002                                 bh->b_end_io = btrfsic_bh_end_io;
2003                         } else {
2004                                 block->is_iodone = 1;
2005                                 block->orig_bio_bh_private = NULL;
2006                                 block->orig_bio_bh_end_io.bio = NULL;
2007                                 block->next_in_same_bio = NULL;
2008                         }
2009                 }
2010
2011                 block->flush_gen = dev_state->last_flush_gen + 1;
2012                 block->submit_bio_bh_rw = submit_bio_bh_rw;
2013                 if (is_metadata) {
2014                         block->logical_bytenr = bytenr;
2015                         block->is_metadata = 1;
2016                         if (block->is_superblock) {
2017                                 BUG_ON(PAGE_SIZE !=
2018                                        BTRFS_SUPER_INFO_SIZE);
2019                                 ret = btrfsic_process_written_superblock(
2020                                                 state,
2021                                                 block,
2022                                                 (struct btrfs_super_block *)
2023                                                 mapped_datav[0]);
2024                                 if (state->print_mask &
2025                                     BTRFSIC_PRINT_MASK_TREE_AFTER_SB_WRITE) {
2026                                         printk(KERN_INFO
2027                                         "[after new superblock is written]:\n");
2028                                         btrfsic_dump_tree_sub(state, block, 0);
2029                                 }
2030                         } else {
2031                                 block->mirror_num = 0;  /* unknown */
2032                                 ret = btrfsic_process_metablock(
2033                                                 state,
2034                                                 block,
2035                                                 &block_ctx,
2036                                                 0, 0);
2037                         }
2038                         if (ret)
2039                                 printk(KERN_INFO
2040                                        "btrfsic: btrfsic_process_metablock"
2041                                        "(root @%llu) failed!\n",
2042                                        dev_bytenr);
2043                 } else {
2044                         block->is_metadata = 0;
2045                         block->mirror_num = 0;  /* unknown */
2046                         block->generation = BTRFSIC_GENERATION_UNKNOWN;
2047                         if (!state->include_extent_data
2048                             && list_empty(&block->ref_from_list)) {
2049                                 /*
2050                                  * disk block is overwritten with extent
2051                                  * data (not meta data) and we are configured
2052                                  * to not include extent data: take the
2053                                  * chance and free the block's memory
2054                                  */
2055                                 btrfsic_block_hashtable_remove(block);
2056                                 list_del(&block->all_blocks_node);
2057                                 btrfsic_block_free(block);
2058                         }
2059                 }
2060                 btrfsic_release_block_ctx(&block_ctx);
2061         } else {
2062                 /* block has not been found in hash table */
2063                 u64 bytenr;
2064
2065                 if (!is_metadata) {
2066                         processed_len = state->datablock_size;
2067                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2068                                 printk(KERN_INFO "Written block (%s/%llu/?)"
2069                                        " !found in hash table, D.\n",
2070                                        dev_state->name, dev_bytenr);
2071                         if (!state->include_extent_data) {
2072                                 /* ignore that written D block */
2073                                 goto continue_loop;
2074                         }
2075
2076                         /* this is getting ugly for the
2077                          * include_extent_data case... */
2078                         bytenr = 0;     /* unknown */
2079                 } else {
2080                         processed_len = state->metablock_size;
2081                         bytenr = btrfs_stack_header_bytenr(
2082                                         (struct btrfs_header *)
2083                                         mapped_datav[0]);
2084                         btrfsic_cmp_log_and_dev_bytenr(state, bytenr, dev_state,
2085                                                        dev_bytenr);
2086                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2087                                 printk(KERN_INFO
2088                                        "Written block @%llu (%s/%llu/?)"
2089                                        " !found in hash table, M.\n",
2090                                        bytenr, dev_state->name, dev_bytenr);
2091                 }
2092
2093                 block_ctx.dev = dev_state;
2094                 block_ctx.dev_bytenr = dev_bytenr;
2095                 block_ctx.start = bytenr;
2096                 block_ctx.len = processed_len;
2097                 block_ctx.pagev = NULL;
2098                 block_ctx.mem_to_free = NULL;
2099                 block_ctx.datav = mapped_datav;
2100
2101                 block = btrfsic_block_alloc();
2102                 if (NULL == block) {
2103                         printk(KERN_INFO "btrfsic: error, kmalloc failed!\n");
2104                         btrfsic_release_block_ctx(&block_ctx);
2105                         goto continue_loop;
2106                 }
2107                 block->dev_state = dev_state;
2108                 block->dev_bytenr = dev_bytenr;
2109                 block->logical_bytenr = bytenr;
2110                 block->is_metadata = is_metadata;
2111                 block->never_written = 0;
2112                 block->iodone_w_error = 0;
2113                 block->mirror_num = 0;  /* unknown */
2114                 block->flush_gen = dev_state->last_flush_gen + 1;
2115                 block->submit_bio_bh_rw = submit_bio_bh_rw;
2116                 if (NULL != bio) {
2117                         block->is_iodone = 0;
2118                         BUG_ON(NULL == bio_is_patched);
2119                         if (!*bio_is_patched) {
2120                                 block->orig_bio_bh_private = bio->bi_private;
2121                                 block->orig_bio_bh_end_io.bio = bio->bi_end_io;
2122                                 block->next_in_same_bio = NULL;
2123                                 bio->bi_private = block;
2124                                 bio->bi_end_io = btrfsic_bio_end_io;
2125                                 *bio_is_patched = 1;
2126                         } else {
2127                                 struct btrfsic_block *chained_block =
2128                                     (struct btrfsic_block *)
2129                                     bio->bi_private;
2130
2131                                 BUG_ON(NULL == chained_block);
2132                                 block->orig_bio_bh_private =
2133                                     chained_block->orig_bio_bh_private;
2134                                 block->orig_bio_bh_end_io.bio =
2135                                     chained_block->orig_bio_bh_end_io.bio;
2136                                 block->next_in_same_bio = chained_block;
2137                                 bio->bi_private = block;
2138                         }
2139                 } else if (NULL != bh) {
2140                         block->is_iodone = 0;
2141                         block->orig_bio_bh_private = bh->b_private;
2142                         block->orig_bio_bh_end_io.bh = bh->b_end_io;
2143                         block->next_in_same_bio = NULL;
2144                         bh->b_private = block;
2145                         bh->b_end_io = btrfsic_bh_end_io;
2146                 } else {
2147                         block->is_iodone = 1;
2148                         block->orig_bio_bh_private = NULL;
2149                         block->orig_bio_bh_end_io.bio = NULL;
2150                         block->next_in_same_bio = NULL;
2151                 }
2152                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2153                         printk(KERN_INFO
2154                                "New written %c-block @%llu (%s/%llu/%d)\n",
2155                                is_metadata ? 'M' : 'D',
2156                                block->logical_bytenr, block->dev_state->name,
2157                                block->dev_bytenr, block->mirror_num);
2158                 list_add(&block->all_blocks_node, &state->all_blocks_list);
2159                 btrfsic_block_hashtable_add(block, &state->block_hashtable);
2160
2161                 if (is_metadata) {
2162                         ret = btrfsic_process_metablock(state, block,
2163                                                         &block_ctx, 0, 0);
2164                         if (ret)
2165                                 printk(KERN_INFO
2166                                        "btrfsic: process_metablock(root @%llu)"
2167                                        " failed!\n",
2168                                        dev_bytenr);
2169                 }
2170                 btrfsic_release_block_ctx(&block_ctx);
2171         }
2172
2173 continue_loop:
2174         BUG_ON(!processed_len);
2175         dev_bytenr += processed_len;
2176         mapped_datav += processed_len >> PAGE_SHIFT;
2177         num_pages -= processed_len >> PAGE_SHIFT;
2178         goto again;
2179 }
2180
2181 static void btrfsic_bio_end_io(struct bio *bp)
2182 {
2183         struct btrfsic_block *block = (struct btrfsic_block *)bp->bi_private;
2184         int iodone_w_error;
2185
2186         /* mutex is not held! This is not save if IO is not yet completed
2187          * on umount */
2188         iodone_w_error = 0;
2189         if (bp->bi_error)
2190                 iodone_w_error = 1;
2191
2192         BUG_ON(NULL == block);
2193         bp->bi_private = block->orig_bio_bh_private;
2194         bp->bi_end_io = block->orig_bio_bh_end_io.bio;
2195
2196         do {
2197                 struct btrfsic_block *next_block;
2198                 struct btrfsic_dev_state *const dev_state = block->dev_state;
2199
2200                 if ((dev_state->state->print_mask &
2201                      BTRFSIC_PRINT_MASK_END_IO_BIO_BH))
2202                         printk(KERN_INFO
2203                                "bio_end_io(err=%d) for %c @%llu (%s/%llu/%d)\n",
2204                                bp->bi_error,
2205                                btrfsic_get_block_type(dev_state->state, block),
2206                                block->logical_bytenr, dev_state->name,
2207                                block->dev_bytenr, block->mirror_num);
2208                 next_block = block->next_in_same_bio;
2209                 block->iodone_w_error = iodone_w_error;
2210                 if (block->submit_bio_bh_rw & REQ_PREFLUSH) {
2211                         dev_state->last_flush_gen++;
2212                         if ((dev_state->state->print_mask &
2213                              BTRFSIC_PRINT_MASK_END_IO_BIO_BH))
2214                                 printk(KERN_INFO
2215                                        "bio_end_io() new %s flush_gen=%llu\n",
2216                                        dev_state->name,
2217                                        dev_state->last_flush_gen);
2218                 }
2219                 if (block->submit_bio_bh_rw & REQ_FUA)
2220                         block->flush_gen = 0; /* FUA completed means block is
2221                                                * on disk */
2222                 block->is_iodone = 1; /* for FLUSH, this releases the block */
2223                 block = next_block;
2224         } while (NULL != block);
2225
2226         bp->bi_end_io(bp);
2227 }
2228
2229 static void btrfsic_bh_end_io(struct buffer_head *bh, int uptodate)
2230 {
2231         struct btrfsic_block *block = (struct btrfsic_block *)bh->b_private;
2232         int iodone_w_error = !uptodate;
2233         struct btrfsic_dev_state *dev_state;
2234
2235         BUG_ON(NULL == block);
2236         dev_state = block->dev_state;
2237         if ((dev_state->state->print_mask & BTRFSIC_PRINT_MASK_END_IO_BIO_BH))
2238                 printk(KERN_INFO
2239                        "bh_end_io(error=%d) for %c @%llu (%s/%llu/%d)\n",
2240                        iodone_w_error,
2241                        btrfsic_get_block_type(dev_state->state, block),
2242                        block->logical_bytenr, block->dev_state->name,
2243                        block->dev_bytenr, block->mirror_num);
2244
2245         block->iodone_w_error = iodone_w_error;
2246         if (block->submit_bio_bh_rw & REQ_PREFLUSH) {
2247                 dev_state->last_flush_gen++;
2248                 if ((dev_state->state->print_mask &
2249                      BTRFSIC_PRINT_MASK_END_IO_BIO_BH))
2250                         printk(KERN_INFO
2251                                "bh_end_io() new %s flush_gen=%llu\n",
2252                                dev_state->name, dev_state->last_flush_gen);
2253         }
2254         if (block->submit_bio_bh_rw & REQ_FUA)
2255                 block->flush_gen = 0; /* FUA completed means block is on disk */
2256
2257         bh->b_private = block->orig_bio_bh_private;
2258         bh->b_end_io = block->orig_bio_bh_end_io.bh;
2259         block->is_iodone = 1; /* for FLUSH, this releases the block */
2260         bh->b_end_io(bh, uptodate);
2261 }
2262
2263 static int btrfsic_process_written_superblock(
2264                 struct btrfsic_state *state,
2265                 struct btrfsic_block *const superblock,
2266                 struct btrfs_super_block *const super_hdr)
2267 {
2268         int pass;
2269
2270         superblock->generation = btrfs_super_generation(super_hdr);
2271         if (!(superblock->generation > state->max_superblock_generation ||
2272               0 == state->max_superblock_generation)) {
2273                 if (state->print_mask & BTRFSIC_PRINT_MASK_SUPERBLOCK_WRITE)
2274                         printk(KERN_INFO
2275                                "btrfsic: superblock @%llu (%s/%llu/%d)"
2276                                " with old gen %llu <= %llu\n",
2277                                superblock->logical_bytenr,
2278                                superblock->dev_state->name,
2279                                superblock->dev_bytenr, superblock->mirror_num,
2280                                btrfs_super_generation(super_hdr),
2281                                state->max_superblock_generation);
2282         } else {
2283                 if (state->print_mask & BTRFSIC_PRINT_MASK_SUPERBLOCK_WRITE)
2284                         printk(KERN_INFO
2285                                "btrfsic: got new superblock @%llu (%s/%llu/%d)"
2286                                " with new gen %llu > %llu\n",
2287                                superblock->logical_bytenr,
2288                                superblock->dev_state->name,
2289                                superblock->dev_bytenr, superblock->mirror_num,
2290                                btrfs_super_generation(super_hdr),
2291                                state->max_superblock_generation);
2292
2293                 state->max_superblock_generation =
2294                     btrfs_super_generation(super_hdr);
2295                 state->latest_superblock = superblock;
2296         }
2297
2298         for (pass = 0; pass < 3; pass++) {
2299                 int ret;
2300                 u64 next_bytenr;
2301                 struct btrfsic_block *next_block;
2302                 struct btrfsic_block_data_ctx tmp_next_block_ctx;
2303                 struct btrfsic_block_link *l;
2304                 int num_copies;
2305                 int mirror_num;
2306                 const char *additional_string = NULL;
2307                 struct btrfs_disk_key tmp_disk_key = {0};
2308
2309                 btrfs_set_disk_key_objectid(&tmp_disk_key,
2310                                             BTRFS_ROOT_ITEM_KEY);
2311                 btrfs_set_disk_key_objectid(&tmp_disk_key, 0);
2312
2313                 switch (pass) {
2314                 case 0:
2315                         btrfs_set_disk_key_objectid(&tmp_disk_key,
2316                                                     BTRFS_ROOT_TREE_OBJECTID);
2317                         additional_string = "root ";
2318                         next_bytenr = btrfs_super_root(super_hdr);
2319                         if (state->print_mask &
2320                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
2321                                 printk(KERN_INFO "root@%llu\n", next_bytenr);
2322                         break;
2323                 case 1:
2324                         btrfs_set_disk_key_objectid(&tmp_disk_key,
2325                                                     BTRFS_CHUNK_TREE_OBJECTID);
2326                         additional_string = "chunk ";
2327                         next_bytenr = btrfs_super_chunk_root(super_hdr);
2328                         if (state->print_mask &
2329                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
2330                                 printk(KERN_INFO "chunk@%llu\n", next_bytenr);
2331                         break;
2332                 case 2:
2333                         btrfs_set_disk_key_objectid(&tmp_disk_key,
2334                                                     BTRFS_TREE_LOG_OBJECTID);
2335                         additional_string = "log ";
2336                         next_bytenr = btrfs_super_log_root(super_hdr);
2337                         if (0 == next_bytenr)
2338                                 continue;
2339                         if (state->print_mask &
2340                             BTRFSIC_PRINT_MASK_ROOT_CHUNK_LOG_TREE_LOCATION)
2341                                 printk(KERN_INFO "log@%llu\n", next_bytenr);
2342                         break;
2343                 }
2344
2345                 num_copies =
2346                     btrfs_num_copies(state->root->fs_info,
2347                                      next_bytenr, BTRFS_SUPER_INFO_SIZE);
2348                 if (state->print_mask & BTRFSIC_PRINT_MASK_NUM_COPIES)
2349                         printk(KERN_INFO "num_copies(log_bytenr=%llu) = %d\n",
2350                                next_bytenr, num_copies);
2351                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
2352                         int was_created;
2353
2354                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2355                                 printk(KERN_INFO
2356                                        "btrfsic_process_written_superblock("
2357                                        "mirror_num=%d)\n", mirror_num);
2358                         ret = btrfsic_map_block(state, next_bytenr,
2359                                                 BTRFS_SUPER_INFO_SIZE,
2360                                                 &tmp_next_block_ctx,
2361                                                 mirror_num);
2362                         if (ret) {
2363                                 printk(KERN_INFO
2364                                        "btrfsic: btrfsic_map_block(@%llu,"
2365                                        " mirror=%d) failed!\n",
2366                                        next_bytenr, mirror_num);
2367                                 return -1;
2368                         }
2369
2370                         next_block = btrfsic_block_lookup_or_add(
2371                                         state,
2372                                         &tmp_next_block_ctx,
2373                                         additional_string,
2374                                         1, 0, 1,
2375                                         mirror_num,
2376                                         &was_created);
2377                         if (NULL == next_block) {
2378                                 printk(KERN_INFO
2379                                        "btrfsic: error, kmalloc failed!\n");
2380                                 btrfsic_release_block_ctx(&tmp_next_block_ctx);
2381                                 return -1;
2382                         }
2383
2384                         next_block->disk_key = tmp_disk_key;
2385                         if (was_created)
2386                                 next_block->generation =
2387                                     BTRFSIC_GENERATION_UNKNOWN;
2388                         l = btrfsic_block_link_lookup_or_add(
2389                                         state,
2390                                         &tmp_next_block_ctx,
2391                                         next_block,
2392                                         superblock,
2393                                         BTRFSIC_GENERATION_UNKNOWN);
2394                         btrfsic_release_block_ctx(&tmp_next_block_ctx);
2395                         if (NULL == l)
2396                                 return -1;
2397                 }
2398         }
2399
2400         if (WARN_ON(-1 == btrfsic_check_all_ref_blocks(state, superblock, 0)))
2401                 btrfsic_dump_tree(state);
2402
2403         return 0;
2404 }
2405
2406 static int btrfsic_check_all_ref_blocks(struct btrfsic_state *state,
2407                                         struct btrfsic_block *const block,
2408                                         int recursion_level)
2409 {
2410         const struct btrfsic_block_link *l;
2411         int ret = 0;
2412
2413         if (recursion_level >= 3 + BTRFS_MAX_LEVEL) {
2414                 /*
2415                  * Note that this situation can happen and does not
2416                  * indicate an error in regular cases. It happens
2417                  * when disk blocks are freed and later reused.
2418                  * The check-integrity module is not aware of any
2419                  * block free operations, it just recognizes block
2420                  * write operations. Therefore it keeps the linkage
2421                  * information for a block until a block is
2422                  * rewritten. This can temporarily cause incorrect
2423                  * and even circular linkage informations. This
2424                  * causes no harm unless such blocks are referenced
2425                  * by the most recent super block.
2426                  */
2427                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2428                         printk(KERN_INFO
2429                                "btrfsic: abort cyclic linkage (case 1).\n");
2430
2431                 return ret;
2432         }
2433
2434         /*
2435          * This algorithm is recursive because the amount of used stack
2436          * space is very small and the max recursion depth is limited.
2437          */
2438         list_for_each_entry(l, &block->ref_to_list, node_ref_to) {
2439                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2440                         printk(KERN_INFO
2441                                "rl=%d, %c @%llu (%s/%llu/%d)"
2442                                " %u* refers to %c @%llu (%s/%llu/%d)\n",
2443                                recursion_level,
2444                                btrfsic_get_block_type(state, block),
2445                                block->logical_bytenr, block->dev_state->name,
2446                                block->dev_bytenr, block->mirror_num,
2447                                l->ref_cnt,
2448                                btrfsic_get_block_type(state, l->block_ref_to),
2449                                l->block_ref_to->logical_bytenr,
2450                                l->block_ref_to->dev_state->name,
2451                                l->block_ref_to->dev_bytenr,
2452                                l->block_ref_to->mirror_num);
2453                 if (l->block_ref_to->never_written) {
2454                         printk(KERN_INFO "btrfs: attempt to write superblock"
2455                                " which references block %c @%llu (%s/%llu/%d)"
2456                                " which is never written!\n",
2457                                btrfsic_get_block_type(state, l->block_ref_to),
2458                                l->block_ref_to->logical_bytenr,
2459                                l->block_ref_to->dev_state->name,
2460                                l->block_ref_to->dev_bytenr,
2461                                l->block_ref_to->mirror_num);
2462                         ret = -1;
2463                 } else if (!l->block_ref_to->is_iodone) {
2464                         printk(KERN_INFO "btrfs: attempt to write superblock"
2465                                " which references block %c @%llu (%s/%llu/%d)"
2466                                " which is not yet iodone!\n",
2467                                btrfsic_get_block_type(state, l->block_ref_to),
2468                                l->block_ref_to->logical_bytenr,
2469                                l->block_ref_to->dev_state->name,
2470                                l->block_ref_to->dev_bytenr,
2471                                l->block_ref_to->mirror_num);
2472                         ret = -1;
2473                 } else if (l->block_ref_to->iodone_w_error) {
2474                         printk(KERN_INFO "btrfs: attempt to write superblock"
2475                                " which references block %c @%llu (%s/%llu/%d)"
2476                                " which has write error!\n",
2477                                btrfsic_get_block_type(state, l->block_ref_to),
2478                                l->block_ref_to->logical_bytenr,
2479                                l->block_ref_to->dev_state->name,
2480                                l->block_ref_to->dev_bytenr,
2481                                l->block_ref_to->mirror_num);
2482                         ret = -1;
2483                 } else if (l->parent_generation !=
2484                            l->block_ref_to->generation &&
2485                            BTRFSIC_GENERATION_UNKNOWN !=
2486                            l->parent_generation &&
2487                            BTRFSIC_GENERATION_UNKNOWN !=
2488                            l->block_ref_to->generation) {
2489                         printk(KERN_INFO "btrfs: attempt to write superblock"
2490                                " which references block %c @%llu (%s/%llu/%d)"
2491                                " with generation %llu !="
2492                                " parent generation %llu!\n",
2493                                btrfsic_get_block_type(state, l->block_ref_to),
2494                                l->block_ref_to->logical_bytenr,
2495                                l->block_ref_to->dev_state->name,
2496                                l->block_ref_to->dev_bytenr,
2497                                l->block_ref_to->mirror_num,
2498                                l->block_ref_to->generation,
2499                                l->parent_generation);
2500                         ret = -1;
2501                 } else if (l->block_ref_to->flush_gen >
2502                            l->block_ref_to->dev_state->last_flush_gen) {
2503                         printk(KERN_INFO "btrfs: attempt to write superblock"
2504                                " which references block %c @%llu (%s/%llu/%d)"
2505                                " which is not flushed out of disk's write cache"
2506                                " (block flush_gen=%llu,"
2507                                " dev->flush_gen=%llu)!\n",
2508                                btrfsic_get_block_type(state, l->block_ref_to),
2509                                l->block_ref_to->logical_bytenr,
2510                                l->block_ref_to->dev_state->name,
2511                                l->block_ref_to->dev_bytenr,
2512                                l->block_ref_to->mirror_num, block->flush_gen,
2513                                l->block_ref_to->dev_state->last_flush_gen);
2514                         ret = -1;
2515                 } else if (-1 == btrfsic_check_all_ref_blocks(state,
2516                                                               l->block_ref_to,
2517                                                               recursion_level +
2518                                                               1)) {
2519                         ret = -1;
2520                 }
2521         }
2522
2523         return ret;
2524 }
2525
2526 static int btrfsic_is_block_ref_by_superblock(
2527                 const struct btrfsic_state *state,
2528                 const struct btrfsic_block *block,
2529                 int recursion_level)
2530 {
2531         const struct btrfsic_block_link *l;
2532
2533         if (recursion_level >= 3 + BTRFS_MAX_LEVEL) {
2534                 /* refer to comment at "abort cyclic linkage (case 1)" */
2535                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2536                         printk(KERN_INFO
2537                                "btrfsic: abort cyclic linkage (case 2).\n");
2538
2539                 return 0;
2540         }
2541
2542         /*
2543          * This algorithm is recursive because the amount of used stack space
2544          * is very small and the max recursion depth is limited.
2545          */
2546         list_for_each_entry(l, &block->ref_from_list, node_ref_from) {
2547                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2548                         printk(KERN_INFO
2549                                "rl=%d, %c @%llu (%s/%llu/%d)"
2550                                " is ref %u* from %c @%llu (%s/%llu/%d)\n",
2551                                recursion_level,
2552                                btrfsic_get_block_type(state, block),
2553                                block->logical_bytenr, block->dev_state->name,
2554                                block->dev_bytenr, block->mirror_num,
2555                                l->ref_cnt,
2556                                btrfsic_get_block_type(state, l->block_ref_from),
2557                                l->block_ref_from->logical_bytenr,
2558                                l->block_ref_from->dev_state->name,
2559                                l->block_ref_from->dev_bytenr,
2560                                l->block_ref_from->mirror_num);
2561                 if (l->block_ref_from->is_superblock &&
2562                     state->latest_superblock->dev_bytenr ==
2563                     l->block_ref_from->dev_bytenr &&
2564                     state->latest_superblock->dev_state->bdev ==
2565                     l->block_ref_from->dev_state->bdev)
2566                         return 1;
2567                 else if (btrfsic_is_block_ref_by_superblock(state,
2568                                                             l->block_ref_from,
2569                                                             recursion_level +
2570                                                             1))
2571                         return 1;
2572         }
2573
2574         return 0;
2575 }
2576
2577 static void btrfsic_print_add_link(const struct btrfsic_state *state,
2578                                    const struct btrfsic_block_link *l)
2579 {
2580         printk(KERN_INFO
2581                "Add %u* link from %c @%llu (%s/%llu/%d)"
2582                " to %c @%llu (%s/%llu/%d).\n",
2583                l->ref_cnt,
2584                btrfsic_get_block_type(state, l->block_ref_from),
2585                l->block_ref_from->logical_bytenr,
2586                l->block_ref_from->dev_state->name,
2587                l->block_ref_from->dev_bytenr, l->block_ref_from->mirror_num,
2588                btrfsic_get_block_type(state, l->block_ref_to),
2589                l->block_ref_to->logical_bytenr,
2590                l->block_ref_to->dev_state->name, l->block_ref_to->dev_bytenr,
2591                l->block_ref_to->mirror_num);
2592 }
2593
2594 static void btrfsic_print_rem_link(const struct btrfsic_state *state,
2595                                    const struct btrfsic_block_link *l)
2596 {
2597         printk(KERN_INFO
2598                "Rem %u* link from %c @%llu (%s/%llu/%d)"
2599                " to %c @%llu (%s/%llu/%d).\n",
2600                l->ref_cnt,
2601                btrfsic_get_block_type(state, l->block_ref_from),
2602                l->block_ref_from->logical_bytenr,
2603                l->block_ref_from->dev_state->name,
2604                l->block_ref_from->dev_bytenr, l->block_ref_from->mirror_num,
2605                btrfsic_get_block_type(state, l->block_ref_to),
2606                l->block_ref_to->logical_bytenr,
2607                l->block_ref_to->dev_state->name, l->block_ref_to->dev_bytenr,
2608                l->block_ref_to->mirror_num);
2609 }
2610
2611 static char btrfsic_get_block_type(const struct btrfsic_state *state,
2612                                    const struct btrfsic_block *block)
2613 {
2614         if (block->is_superblock &&
2615             state->latest_superblock->dev_bytenr == block->dev_bytenr &&
2616             state->latest_superblock->dev_state->bdev == block->dev_state->bdev)
2617                 return 'S';
2618         else if (block->is_superblock)
2619                 return 's';
2620         else if (block->is_metadata)
2621                 return 'M';
2622         else
2623                 return 'D';
2624 }
2625
2626 static void btrfsic_dump_tree(const struct btrfsic_state *state)
2627 {
2628         btrfsic_dump_tree_sub(state, state->latest_superblock, 0);
2629 }
2630
2631 static void btrfsic_dump_tree_sub(const struct btrfsic_state *state,
2632                                   const struct btrfsic_block *block,
2633                                   int indent_level)
2634 {
2635         const struct btrfsic_block_link *l;
2636         int indent_add;
2637         static char buf[80];
2638         int cursor_position;
2639
2640         /*
2641          * Should better fill an on-stack buffer with a complete line and
2642          * dump it at once when it is time to print a newline character.
2643          */
2644
2645         /*
2646          * This algorithm is recursive because the amount of used stack space
2647          * is very small and the max recursion depth is limited.
2648          */
2649         indent_add = sprintf(buf, "%c-%llu(%s/%llu/%u)",
2650                              btrfsic_get_block_type(state, block),
2651                              block->logical_bytenr, block->dev_state->name,
2652                              block->dev_bytenr, block->mirror_num);
2653         if (indent_level + indent_add > BTRFSIC_TREE_DUMP_MAX_INDENT_LEVEL) {
2654                 printk("[...]\n");
2655                 return;
2656         }
2657         printk(buf);
2658         indent_level += indent_add;
2659         if (list_empty(&block->ref_to_list)) {
2660                 printk("\n");
2661                 return;
2662         }
2663         if (block->mirror_num > 1 &&
2664             !(state->print_mask & BTRFSIC_PRINT_MASK_TREE_WITH_ALL_MIRRORS)) {
2665                 printk(" [...]\n");
2666                 return;
2667         }
2668
2669         cursor_position = indent_level;
2670         list_for_each_entry(l, &block->ref_to_list, node_ref_to) {
2671                 while (cursor_position < indent_level) {
2672                         printk(" ");
2673                         cursor_position++;
2674                 }
2675                 if (l->ref_cnt > 1)
2676                         indent_add = sprintf(buf, " %d*--> ", l->ref_cnt);
2677                 else
2678                         indent_add = sprintf(buf, " --> ");
2679                 if (indent_level + indent_add >
2680                     BTRFSIC_TREE_DUMP_MAX_INDENT_LEVEL) {
2681                         printk("[...]\n");
2682                         cursor_position = 0;
2683                         continue;
2684                 }
2685
2686                 printk(buf);
2687
2688                 btrfsic_dump_tree_sub(state, l->block_ref_to,
2689                                       indent_level + indent_add);
2690                 cursor_position = 0;
2691         }
2692 }
2693
2694 static struct btrfsic_block_link *btrfsic_block_link_lookup_or_add(
2695                 struct btrfsic_state *state,
2696                 struct btrfsic_block_data_ctx *next_block_ctx,
2697                 struct btrfsic_block *next_block,
2698                 struct btrfsic_block *from_block,
2699                 u64 parent_generation)
2700 {
2701         struct btrfsic_block_link *l;
2702
2703         l = btrfsic_block_link_hashtable_lookup(next_block_ctx->dev->bdev,
2704                                                 next_block_ctx->dev_bytenr,
2705                                                 from_block->dev_state->bdev,
2706                                                 from_block->dev_bytenr,
2707                                                 &state->block_link_hashtable);
2708         if (NULL == l) {
2709                 l = btrfsic_block_link_alloc();
2710                 if (NULL == l) {
2711                         printk(KERN_INFO
2712                                "btrfsic: error, kmalloc" " failed!\n");
2713                         return NULL;
2714                 }
2715
2716                 l->block_ref_to = next_block;
2717                 l->block_ref_from = from_block;
2718                 l->ref_cnt = 1;
2719                 l->parent_generation = parent_generation;
2720
2721                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2722                         btrfsic_print_add_link(state, l);
2723
2724                 list_add(&l->node_ref_to, &from_block->ref_to_list);
2725                 list_add(&l->node_ref_from, &next_block->ref_from_list);
2726
2727                 btrfsic_block_link_hashtable_add(l,
2728                                                  &state->block_link_hashtable);
2729         } else {
2730                 l->ref_cnt++;
2731                 l->parent_generation = parent_generation;
2732                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2733                         btrfsic_print_add_link(state, l);
2734         }
2735
2736         return l;
2737 }
2738
2739 static struct btrfsic_block *btrfsic_block_lookup_or_add(
2740                 struct btrfsic_state *state,
2741                 struct btrfsic_block_data_ctx *block_ctx,
2742                 const char *additional_string,
2743                 int is_metadata,
2744                 int is_iodone,
2745                 int never_written,
2746                 int mirror_num,
2747                 int *was_created)
2748 {
2749         struct btrfsic_block *block;
2750
2751         block = btrfsic_block_hashtable_lookup(block_ctx->dev->bdev,
2752                                                block_ctx->dev_bytenr,
2753                                                &state->block_hashtable);
2754         if (NULL == block) {
2755                 struct btrfsic_dev_state *dev_state;
2756
2757                 block = btrfsic_block_alloc();
2758                 if (NULL == block) {
2759                         printk(KERN_INFO "btrfsic: error, kmalloc failed!\n");
2760                         return NULL;
2761                 }
2762                 dev_state = btrfsic_dev_state_lookup(block_ctx->dev->bdev);
2763                 if (NULL == dev_state) {
2764                         printk(KERN_INFO
2765                                "btrfsic: error, lookup dev_state failed!\n");
2766                         btrfsic_block_free(block);
2767                         return NULL;
2768                 }
2769                 block->dev_state = dev_state;
2770                 block->dev_bytenr = block_ctx->dev_bytenr;
2771                 block->logical_bytenr = block_ctx->start;
2772                 block->is_metadata = is_metadata;
2773                 block->is_iodone = is_iodone;
2774                 block->never_written = never_written;
2775                 block->mirror_num = mirror_num;
2776                 if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
2777                         printk(KERN_INFO
2778                                "New %s%c-block @%llu (%s/%llu/%d)\n",
2779                                additional_string,
2780                                btrfsic_get_block_type(state, block),
2781                                block->logical_bytenr, dev_state->name,
2782                                block->dev_bytenr, mirror_num);
2783                 list_add(&block->all_blocks_node, &state->all_blocks_list);
2784                 btrfsic_block_hashtable_add(block, &state->block_hashtable);
2785                 if (NULL != was_created)
2786                         *was_created = 1;
2787         } else {
2788                 if (NULL != was_created)
2789                         *was_created = 0;
2790         }
2791
2792         return block;
2793 }
2794
2795 static void btrfsic_cmp_log_and_dev_bytenr(struct btrfsic_state *state,
2796                                            u64 bytenr,
2797                                            struct btrfsic_dev_state *dev_state,
2798                                            u64 dev_bytenr)
2799 {
2800         int num_copies;
2801         int mirror_num;
2802         int ret;
2803         struct btrfsic_block_data_ctx block_ctx;
2804         int match = 0;
2805
2806         num_copies = btrfs_num_copies(state->root->fs_info,
2807                                       bytenr, state->metablock_size);
2808
2809         for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
2810                 ret = btrfsic_map_block(state, bytenr, state->metablock_size,
2811                                         &block_ctx, mirror_num);
2812                 if (ret) {
2813                         printk(KERN_INFO "btrfsic:"
2814                                " btrfsic_map_block(logical @%llu,"
2815                                " mirror %d) failed!\n",
2816                                bytenr, mirror_num);
2817                         continue;
2818                 }
2819
2820                 if (dev_state->bdev == block_ctx.dev->bdev &&
2821                     dev_bytenr == block_ctx.dev_bytenr) {
2822                         match++;
2823                         btrfsic_release_block_ctx(&block_ctx);
2824                         break;
2825                 }
2826                 btrfsic_release_block_ctx(&block_ctx);
2827         }
2828
2829         if (WARN_ON(!match)) {
2830                 printk(KERN_INFO "btrfs: attempt to write M-block which contains logical bytenr that doesn't map to dev+physical bytenr of submit_bio,"
2831                        " buffer->log_bytenr=%llu, submit_bio(bdev=%s,"
2832                        " phys_bytenr=%llu)!\n",
2833                        bytenr, dev_state->name, dev_bytenr);
2834                 for (mirror_num = 1; mirror_num <= num_copies; mirror_num++) {
2835                         ret = btrfsic_map_block(state, bytenr,
2836                                                 state->metablock_size,
2837                                                 &block_ctx, mirror_num);
2838                         if (ret)
2839                                 continue;
2840
2841                         printk(KERN_INFO "Read logical bytenr @%llu maps to"
2842                                " (%s/%llu/%d)\n",
2843                                bytenr, block_ctx.dev->name,
2844                                block_ctx.dev_bytenr, mirror_num);
2845                 }
2846         }
2847 }
2848
2849 static struct btrfsic_dev_state *btrfsic_dev_state_lookup(
2850                 struct block_device *bdev)
2851 {
2852         struct btrfsic_dev_state *ds;
2853
2854         ds = btrfsic_dev_state_hashtable_lookup(bdev,
2855                                                 &btrfsic_dev_state_hashtable);
2856         return ds;
2857 }
2858
2859 int btrfsic_submit_bh(int op, int op_flags, struct buffer_head *bh)
2860 {
2861         struct btrfsic_dev_state *dev_state;
2862
2863         if (!btrfsic_is_initialized)
2864                 return submit_bh(op, op_flags, bh);
2865
2866         mutex_lock(&btrfsic_mutex);
2867         /* since btrfsic_submit_bh() might also be called before
2868          * btrfsic_mount(), this might return NULL */
2869         dev_state = btrfsic_dev_state_lookup(bh->b_bdev);
2870
2871         /* Only called to write the superblock (incl. FLUSH/FUA) */
2872         if (NULL != dev_state &&
2873             (op == REQ_OP_WRITE) && bh->b_size > 0) {
2874                 u64 dev_bytenr;
2875
2876                 dev_bytenr = 4096 * bh->b_blocknr;
2877                 if (dev_state->state->print_mask &
2878                     BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH)
2879                         printk(KERN_INFO
2880                                "submit_bh(op=0x%x,0x%x, blocknr=%llu "
2881                                "(bytenr %llu), size=%zu, data=%p, bdev=%p)\n",
2882                                op, op_flags, (unsigned long long)bh->b_blocknr,
2883                                dev_bytenr, bh->b_size, bh->b_data, bh->b_bdev);
2884                 btrfsic_process_written_block(dev_state, dev_bytenr,
2885                                               &bh->b_data, 1, NULL,
2886                                               NULL, bh, op_flags);
2887         } else if (NULL != dev_state && (op_flags & REQ_PREFLUSH)) {
2888                 if (dev_state->state->print_mask &
2889                     BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH)
2890                         printk(KERN_INFO
2891                                "submit_bh(op=0x%x,0x%x FLUSH, bdev=%p)\n",
2892                                op, op_flags, bh->b_bdev);
2893                 if (!dev_state->dummy_block_for_bio_bh_flush.is_iodone) {
2894                         if ((dev_state->state->print_mask &
2895                              (BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH |
2896                               BTRFSIC_PRINT_MASK_VERBOSE)))
2897                                 printk(KERN_INFO
2898                                        "btrfsic_submit_bh(%s) with FLUSH"
2899                                        " but dummy block already in use"
2900                                        " (ignored)!\n",
2901                                        dev_state->name);
2902                 } else {
2903                         struct btrfsic_block *const block =
2904                                 &dev_state->dummy_block_for_bio_bh_flush;
2905
2906                         block->is_iodone = 0;
2907                         block->never_written = 0;
2908                         block->iodone_w_error = 0;
2909                         block->flush_gen = dev_state->last_flush_gen + 1;
2910                         block->submit_bio_bh_rw = op_flags;
2911                         block->orig_bio_bh_private = bh->b_private;
2912                         block->orig_bio_bh_end_io.bh = bh->b_end_io;
2913                         block->next_in_same_bio = NULL;
2914                         bh->b_private = block;
2915                         bh->b_end_io = btrfsic_bh_end_io;
2916                 }
2917         }
2918         mutex_unlock(&btrfsic_mutex);
2919         return submit_bh(op, op_flags, bh);
2920 }
2921
2922 static void __btrfsic_submit_bio(struct bio *bio)
2923 {
2924         struct btrfsic_dev_state *dev_state;
2925
2926         if (!btrfsic_is_initialized)
2927                 return;
2928
2929         mutex_lock(&btrfsic_mutex);
2930         /* since btrfsic_submit_bio() is also called before
2931          * btrfsic_mount(), this might return NULL */
2932         dev_state = btrfsic_dev_state_lookup(bio->bi_bdev);
2933         if (NULL != dev_state &&
2934             (bio_op(bio) == REQ_OP_WRITE) && NULL != bio->bi_io_vec) {
2935                 unsigned int i;
2936                 u64 dev_bytenr;
2937                 u64 cur_bytenr;
2938                 int bio_is_patched;
2939                 char **mapped_datav;
2940
2941                 dev_bytenr = 512 * bio->bi_iter.bi_sector;
2942                 bio_is_patched = 0;
2943                 if (dev_state->state->print_mask &
2944                     BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH)
2945                         printk(KERN_INFO
2946                                "submit_bio(rw=%d,0x%x, bi_vcnt=%u,"
2947                                " bi_sector=%llu (bytenr %llu), bi_bdev=%p)\n",
2948                                bio_op(bio), bio->bi_opf, bio->bi_vcnt,
2949                                (unsigned long long)bio->bi_iter.bi_sector,
2950                                dev_bytenr, bio->bi_bdev);
2951
2952                 mapped_datav = kmalloc_array(bio->bi_vcnt,
2953                                              sizeof(*mapped_datav), GFP_NOFS);
2954                 if (!mapped_datav)
2955                         goto leave;
2956                 cur_bytenr = dev_bytenr;
2957                 for (i = 0; i < bio->bi_vcnt; i++) {
2958                         BUG_ON(bio->bi_io_vec[i].bv_len != PAGE_SIZE);
2959                         mapped_datav[i] = kmap(bio->bi_io_vec[i].bv_page);
2960                         if (!mapped_datav[i]) {
2961                                 while (i > 0) {
2962                                         i--;
2963                                         kunmap(bio->bi_io_vec[i].bv_page);
2964                                 }
2965                                 kfree(mapped_datav);
2966                                 goto leave;
2967                         }
2968                         if (dev_state->state->print_mask &
2969                             BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH_VERBOSE)
2970                                 printk(KERN_INFO
2971                                        "#%u: bytenr=%llu, len=%u, offset=%u\n",
2972                                        i, cur_bytenr, bio->bi_io_vec[i].bv_len,
2973                                        bio->bi_io_vec[i].bv_offset);
2974                         cur_bytenr += bio->bi_io_vec[i].bv_len;
2975                 }
2976                 btrfsic_process_written_block(dev_state, dev_bytenr,
2977                                               mapped_datav, bio->bi_vcnt,
2978                                               bio, &bio_is_patched,
2979                                               NULL, bio->bi_opf);
2980                 while (i > 0) {
2981                         i--;
2982                         kunmap(bio->bi_io_vec[i].bv_page);
2983                 }
2984                 kfree(mapped_datav);
2985         } else if (NULL != dev_state && (bio->bi_opf & REQ_PREFLUSH)) {
2986                 if (dev_state->state->print_mask &
2987                     BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH)
2988                         printk(KERN_INFO
2989                                "submit_bio(rw=%d,0x%x FLUSH, bdev=%p)\n",
2990                                bio_op(bio), bio->bi_opf, bio->bi_bdev);
2991                 if (!dev_state->dummy_block_for_bio_bh_flush.is_iodone) {
2992                         if ((dev_state->state->print_mask &
2993                              (BTRFSIC_PRINT_MASK_SUBMIT_BIO_BH |
2994                               BTRFSIC_PRINT_MASK_VERBOSE)))
2995                                 printk(KERN_INFO
2996                                        "btrfsic_submit_bio(%s) with FLUSH"
2997                                        " but dummy block already in use"
2998                                        " (ignored)!\n",
2999                                        dev_state->name);
3000                 } else {
3001                         struct btrfsic_block *const block =
3002                                 &dev_state->dummy_block_for_bio_bh_flush;
3003
3004                         block->is_iodone = 0;
3005                         block->never_written = 0;
3006                         block->iodone_w_error = 0;
3007                         block->flush_gen = dev_state->last_flush_gen + 1;
3008                         block->submit_bio_bh_rw = bio->bi_opf;
3009                         block->orig_bio_bh_private = bio->bi_private;
3010                         block->orig_bio_bh_end_io.bio = bio->bi_end_io;
3011                         block->next_in_same_bio = NULL;
3012                         bio->bi_private = block;
3013                         bio->bi_end_io = btrfsic_bio_end_io;
3014                 }
3015         }
3016 leave:
3017         mutex_unlock(&btrfsic_mutex);
3018 }
3019
3020 void btrfsic_submit_bio(struct bio *bio)
3021 {
3022         __btrfsic_submit_bio(bio);
3023         submit_bio(bio);
3024 }
3025
3026 int btrfsic_submit_bio_wait(struct bio *bio)
3027 {
3028         __btrfsic_submit_bio(bio);
3029         return submit_bio_wait(bio);
3030 }
3031
3032 int btrfsic_mount(struct btrfs_root *root,
3033                   struct btrfs_fs_devices *fs_devices,
3034                   int including_extent_data, u32 print_mask)
3035 {
3036         int ret;
3037         struct btrfsic_state *state;
3038         struct list_head *dev_head = &fs_devices->devices;
3039         struct btrfs_device *device;
3040
3041         if (root->nodesize & ((u64)PAGE_SIZE - 1)) {
3042                 printk(KERN_INFO
3043                        "btrfsic: cannot handle nodesize %d not being a multiple of PAGE_SIZE %ld!\n",
3044                        root->nodesize, PAGE_SIZE);
3045                 return -1;
3046         }
3047         if (root->sectorsize & ((u64)PAGE_SIZE - 1)) {
3048                 printk(KERN_INFO
3049                        "btrfsic: cannot handle sectorsize %d not being a multiple of PAGE_SIZE %ld!\n",
3050                        root->sectorsize, PAGE_SIZE);
3051                 return -1;
3052         }
3053         state = kzalloc(sizeof(*state), GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
3054         if (!state) {
3055                 state = vzalloc(sizeof(*state));
3056                 if (!state) {
3057                         printk(KERN_INFO "btrfs check-integrity: vzalloc() failed!\n");
3058                         return -1;
3059                 }
3060         }
3061
3062         if (!btrfsic_is_initialized) {
3063                 mutex_init(&btrfsic_mutex);
3064                 btrfsic_dev_state_hashtable_init(&btrfsic_dev_state_hashtable);
3065                 btrfsic_is_initialized = 1;
3066         }
3067         mutex_lock(&btrfsic_mutex);
3068         state->root = root;
3069         state->print_mask = print_mask;
3070         state->include_extent_data = including_extent_data;
3071         state->csum_size = 0;
3072         state->metablock_size = root->nodesize;
3073         state->datablock_size = root->sectorsize;
3074         INIT_LIST_HEAD(&state->all_blocks_list);
3075         btrfsic_block_hashtable_init(&state->block_hashtable);
3076         btrfsic_block_link_hashtable_init(&state->block_link_hashtable);
3077         state->max_superblock_generation = 0;
3078         state->latest_superblock = NULL;
3079
3080         list_for_each_entry(device, dev_head, dev_list) {
3081                 struct btrfsic_dev_state *ds;
3082                 const char *p;
3083
3084                 if (!device->bdev || !device->name)
3085                         continue;
3086
3087                 ds = btrfsic_dev_state_alloc();
3088                 if (NULL == ds) {
3089                         printk(KERN_INFO
3090                                "btrfs check-integrity: kmalloc() failed!\n");
3091                         mutex_unlock(&btrfsic_mutex);
3092                         return -1;
3093                 }
3094                 ds->bdev = device->bdev;
3095                 ds->state = state;
3096                 bdevname(ds->bdev, ds->name);
3097                 ds->name[BDEVNAME_SIZE - 1] = '\0';
3098                 p = kbasename(ds->name);
3099                 strlcpy(ds->name, p, sizeof(ds->name));
3100                 btrfsic_dev_state_hashtable_add(ds,
3101                                                 &btrfsic_dev_state_hashtable);
3102         }
3103
3104         ret = btrfsic_process_superblock(state, fs_devices);
3105         if (0 != ret) {
3106                 mutex_unlock(&btrfsic_mutex);
3107                 btrfsic_unmount(root, fs_devices);
3108                 return ret;
3109         }
3110
3111         if (state->print_mask & BTRFSIC_PRINT_MASK_INITIAL_DATABASE)
3112                 btrfsic_dump_database(state);
3113         if (state->print_mask & BTRFSIC_PRINT_MASK_INITIAL_TREE)
3114                 btrfsic_dump_tree(state);
3115
3116         mutex_unlock(&btrfsic_mutex);
3117         return 0;
3118 }
3119
3120 void btrfsic_unmount(struct btrfs_root *root,
3121                      struct btrfs_fs_devices *fs_devices)
3122 {
3123         struct btrfsic_block *b_all, *tmp_all;
3124         struct btrfsic_state *state;
3125         struct list_head *dev_head = &fs_devices->devices;
3126         struct btrfs_device *device;
3127
3128         if (!btrfsic_is_initialized)
3129                 return;
3130
3131         mutex_lock(&btrfsic_mutex);
3132
3133         state = NULL;
3134         list_for_each_entry(device, dev_head, dev_list) {
3135                 struct btrfsic_dev_state *ds;
3136
3137                 if (!device->bdev || !device->name)
3138                         continue;
3139
3140                 ds = btrfsic_dev_state_hashtable_lookup(
3141                                 device->bdev,
3142                                 &btrfsic_dev_state_hashtable);
3143                 if (NULL != ds) {
3144                         state = ds->state;
3145                         btrfsic_dev_state_hashtable_remove(ds);
3146                         btrfsic_dev_state_free(ds);
3147                 }
3148         }
3149
3150         if (NULL == state) {
3151                 printk(KERN_INFO
3152                        "btrfsic: error, cannot find state information"
3153                        " on umount!\n");
3154                 mutex_unlock(&btrfsic_mutex);
3155                 return;
3156         }
3157
3158         /*
3159          * Don't care about keeping the lists' state up to date,
3160          * just free all memory that was allocated dynamically.
3161          * Free the blocks and the block_links.
3162          */
3163         list_for_each_entry_safe(b_all, tmp_all, &state->all_blocks_list,
3164                                  all_blocks_node) {
3165                 struct btrfsic_block_link *l, *tmp;
3166
3167                 list_for_each_entry_safe(l, tmp, &b_all->ref_to_list,
3168                                          node_ref_to) {
3169                         if (state->print_mask & BTRFSIC_PRINT_MASK_VERBOSE)
3170                                 btrfsic_print_rem_link(state, l);
3171
3172                         l->ref_cnt--;
3173                         if (0 == l->ref_cnt)
3174                                 btrfsic_block_link_free(l);
3175                 }
3176
3177                 if (b_all->is_iodone || b_all->never_written)
3178                         btrfsic_block_free(b_all);
3179                 else
3180                         printk(KERN_INFO "btrfs: attempt to free %c-block"
3181                                " @%llu (%s/%llu/%d) on umount which is"
3182                                " not yet iodone!\n",
3183                                btrfsic_get_block_type(state, b_all),
3184                                b_all->logical_bytenr, b_all->dev_state->name,
3185                                b_all->dev_bytenr, b_all->mirror_num);
3186         }
3187
3188         mutex_unlock(&btrfsic_mutex);
3189
3190         kvfree(state);
3191 }