On Mon, 06 Aug 2007 19:43:46 +0200
Miklos Szeredi [off-list ref] wrote:

quoted

Separate the handling of the local ia_valid bitmask from the one in
attr->ia_valid. This allows us to hand off the actual handling of the
ATTR_KILL_* flags to the .setattr i_op when one is defined.

notify_change still needs to process those flags for the local ia_valid
variable, since it uses that to decide whether to return early, and to pass
a (hopefully) appropriate bitmask to fsnotify_change.

I agree with this change and fuse will make use of it as well.

Maybe instead of unconditionally moving attr_kill_to_mode() inside
->setattr() it could be made conditional based on an inode flag
similarly to S_NOCMTIME.  Advantages:

 - no need to modify a lot of in-tree filesystems
 - no silent breakage of out-of-tree fs

Actually I think the new flag would be used by exacly the same
filesystems as S_NOCMTIME, so maybe it would make sense to rename
S_NOCMTIME to something more generic (S_NOATTRUPDATE or whatever) and
use that.

But that could still break out-of-tree fs, so a separate flag is
probably better.

In the past I've been told that adding new flags is something of a
"last resort". Since it's not strictly necessary to fix this then
it may be best to avoid that.

That said, if the concensus is that we need a transition mechanism,
then I'd be open to such a suggestion.

-- 
Jeff Layton [off-list ref]