On 23.08.2017 10:57, Qu Wenruo wrote:

The patchset introduce a new framework to do more comprehensive (if not
the most) sanity check when reading out a leaf.

The new sanity checker will include:

1) Key order
   Existing code

2) Item boundary
   Existing code with enhanced checker to ensure item pointer doesn't
   overlap with item itself.

3) Key type based sanity checker
   Only EXTENT_DATA and EXTENT_CSUM checker is implemented yet.
   As each checker should go through review and tests, or it can easily
   make a valid btrfs failed to be mounted.
   So only two checkers are implemented as an example.

   Existing checker like INODE_REF checker can be moved to this
   framework easily, and we can centralize all existing checkers, make
   the rest of codes more clean.

Performance wise, it's just iterating a leaf.
And it will only get triggered when reading a leaf, cached leaf will
not go through such checker.
So it won't be a performance breaker.

I tested with the patchset applied on v4.13-rc6 with fstests, no
regression is detected.

Changelog:
v2:
  Use EUCLEAN to replace EIO as error code. (Suggested by Nikolay Borisov)
  Pass the whole key from check_leaf() to reduce btrfs_item_key_to_cpu()
  call. (Suggested by Nikolay Borisov)
  Add EXTENT_CSUM checker

Qu Wenruo (4):
  btrfs: Refactor check_leaf function for later expansion.
  btrfs: Check if item pointer overlap with item itself
  btrfs: Add sanity check for EXTENT_DATA when reading out leaf
  btrfs: Add checker for EXTENT_CSUM

 fs/btrfs/disk-io.c              | 189 +++++++++++++++++++++++++++++++++++-----
 include/uapi/linux/btrfs_tree.h |   1 +
 2 files changed, 167 insertions(+), 23 deletions(-)

For the whole series:

Reviewed-by: Nikolay Borisov <redacted>