Discussion:
[PATCH 0/4] fs: fcntl/fadvice fixes v2
(too old to reply)
Dmitry Monakhov
2014-10-18 15:21:24 UTC
Permalink
fcntl(F_SETFL) and fadvise performs direct manipulation with file's internals.
w/o notifying to fs layer. This behavior be not be suitable for some filesystems
(mostly stack-fs like ecryptfs, unionfs, etc). Let's introduce new ->set_flags()
callback for that purpose. This callback is responsible for flags check so
->check_flags() no longer required.

TOC:
fs: fcntl add set_flags wrapper -v2
fs: add fadvise file_operation
ecryptfs: add fadvise/set_flags calbacks
cifs: add set_flag callback

*OPEN ISSUE REMAINS*
This series does not fix all issues related with set_flags.
Race between fcntl(toggling O_DIRECT) vs write() is still possible
Usually O_DIRECT checked twice during call chain:
->xxx_file_write_iter
--->__generic_file_write_iter
So we may end-up up with two different values. Some filesystems (btrfs/xfs)
avoid this issue by copy-pasting __generic_file_write_iter.
One of possible way to fix this issue it to save flags in kiocb->ki_flags
as we already do with ->ki_pos. And fixup all places accordingly.
I've calculated numbers of direct access to ->f_flags it is close to 150,
half of that number is ->open() methods. So patch would not be gigantic.
And finally here is my question to AlViro and Christoph and other VFS-people:
*Are you agree with that approach?* Please say your word.

Loading...