[PATCH v2 0/1] ext4: Fix stale data exposure caused with dioread_nolock

[Ojaswin Mujoo]

** Changes in v2 ** 

- Moved the logic to mark buffer as new in ext4_get_block_unwritten, as
  it is actually specific to this function, rather than _ext4_get_block
  which is a more generic function called from codepaths that dont need
  this logic

- Tested with -g rw,quick as well as custom replication test (present in
  commit message). No issues with ps == bs as well as sub page size
  scenarios.

- TODO: Add logic in ext4_block_zero_page_range() to ignore zeroing if
  the buffer is unwritten. This needs some review of the involved code paths 
  hence sending the fix to stale data first right now and will get to
  this in a new patch.

- v1:
  https://lore.kernel.org/linux-ext4/cover.1693909504.git.ojaswin@xxxxxxxxxxxxx/

** Original Cover **

The detailed report on the issues faced and the root cause can be found
in the commit message. I've intentionally added all the details to
commit message so that it can be tracked in the future, let me know if
its too long and I can try stripping some info.

For this particular fix, I've tested these patches with xfstests -g
quick with:

- 64k block size, 64k pagesize 
- 4k blocksize 64k pagesize
- both with and without nodelalloc 

and I don't see any regressions. I'll plan to run more tests on this and
report back if I notice anything. Suggestions or ideas are welcome.

Regards,
ojaswin

Ojaswin Mujoo (1):
  ext4: Mark buffer new if it is unwritten to avoid stale data exposure

 fs/ext4/inode.c | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)

-- 
2.39.3