fsnotify: assert that file_start_write() is not held in permission hooks

filesystem may be modified in the context of fanotify permission events
(e.g. by HSM service), so assert that sb freeze protection is not held.

If the assertion fails, then the following deadlock would be possible:

CPU0				CPU1			CPU2
-------------------------------------------------------------------------
file_start_write()#0
...
  fsnotify_perm()
    fanotify_get_response() =>	(read event and fill file)
				...
				...			freeze_super()
				...			  sb_wait_write()
				...
				vfs_write()
				  file_start_write()#1

This example demonstrates a use case of an hierarchical storage management
(HSM) service that uses fanotify permission events to fill the content of
a file before access, while a 3rd process starts fsfreeze.

This creates a circular dependeny:
  file_start_write()#0 => fanotify_get_response =>
    file_start_write()#1 =>
      sb_wait_write() =>
        file_end_write()#0

Where file_end_write()#0 can never be called and none of the threads can
make progress.

The assertion is checked for both MAY_READ and MAY_WRITE permission
hooks in preparation for a pre-modify permission event.

The assertion is not checked for an open permission event, because
do_open() takes mnt_want_write() in O_TRUNC case, meaning that it is not
safe to write to filesystem in the content of an open permission event.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Link: https://lore.kernel.org/r/20231212094440.250945-5-amir73il@gmail.com
Signed-off-by: Christian Brauner <brauner@kernel.org>

0 files changed, 0 insertions, 0 deletions