On Fri 02-02-24 07:47:50, Kent Overstreet wrote: > On Fri, Feb 02, 2024 at 01:03:57PM +0100, Jan Kara wrote: > > On Fri 26-01-24 21:08:28, Kent Overstreet wrote: > > > *_lock_nested() is fundamentally broken; lockdep needs to check lock > > > ordering, but we cannot device a total ordering on an unbounded number > > > of elements with only a few subclasses. > > > > > > the replacement is to define lock ordering with a proper comparison > > > function. > > > > > > fs/pipe.c was already doing everything correctly otherwise, nothing > > > much changes here. > > > > > > Cc: Alexander Viro <viro@xxxxxxxxxxxxxxxxxx> > > > Cc: Christian Brauner <brauner@xxxxxxxxxx> > > > Cc: Jan Kara <jack@xxxxxxx> > > > Signed-off-by: Kent Overstreet <kent.overstreet@xxxxxxxxx> > > > > I had to digest for a while what this new lockdep lock ordering feature is > > about. I have one pending question - what is the motivation of this > > conversion of pipe code? AFAIU we don't have any problems with lockdep > > annotations on pipe->mutex because there are always only two subclasses? > > It's one of the easier conversions to do, and ideally /all/ users of > subclasses would go away. > > Start with the easier ones, figure out those patterns, then the > harder... I see, thanks for explanation. So in the pipes case I actually like that the patch makes the locking less obfuscated with lockdep details (to which I'm mostly used to but still ;)). So feel free to add: Reviewed-by: Jan Kara <jack@xxxxxxx> for this patch. I'm not 100% convinced it will be always possible to replace subclasses with the new ordering mechanism but I guess time will show. Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
