[Intel-gfx] [PATCH 00/33] fbcon notifier begone!

[Daniel Vetter]

On Mon, May 27, 2019 at 9:17 AM Daniel Vetter <daniel at ffwll.ch> wrote:
>
> On Sat, May 25, 2019 at 07:19:28PM +0200, Sam Ravnborg wrote:
> > Hi Daniel.
> >
> > Good work, nice cleanup all over.
> >
> > A few comments to a few patches - not something that warrant a
> > new series to be posted as long as it is fixed before the patches are
> > applied.
>
> Hm yeah good idea, I'll add that to the next version.

Actually I thought a bit more about the locking story, and it's a bit
worse than my simple plan here. I think better to just have that
floating around, and not make it look like it's an easy simple
cleanup.

The case I forgot about is that any places that has a printk can
recurse through the console_lock, which means we need a lot more
try_lock than I originally thought we'd need.
-Daniel

>
> > > btw for future plans: I think this is tricky enough (it's old code and all
> > > that) that we should let this soak for 2-3 kernel releases. I think the
> > > following would be nice subsequent cleanup/fixes:
> > >
> > > - push the console lock completely from fbmem.c to fbcon.c. I think we're
> > >   mostly there with prep, but needs to pondering of corner cases.
> > I wonder - should this code consistently use __acquire() etc so we could
> > get a little static analysis help for the locking?
> >
> > I have not played with this for several years and I do not know the
> > maturity of this today.
>
> Ime these sparse annotations are pretty useless because too inflexible. I
> tend to use runtime locking checks based on lockdep. Those are more
> accurate, and work even when the place the lock is taken is very far away
> from where you're checking, without having to annoate all functions in
> between. We need that for something like console_lock which is a very big
> lock. Only downside is that only paths you hit at runtime are checked.
>
> > > - move fbcon.c from using indices for tracking fb_info (and accessing
> > >   registered_fbs without proper locking all the time) to real fb_info
> > >   pointers with the right amount of refcounting. Mostly motivated by the
> > >   fun I had trying to simplify fbcon_exit().
> > >
> > > - make sure that fbcon call lock/unlock_fb when it calls fbmem.c
> > >   functions, and sprinkle assert_lockdep_held around in fbmem.c. This
> > >   needs the console_lock cleanups first.
> > >
> > > But I think that's material for maybe next year or so.
> > Or maybe after next kernel release.
> > Could we put this nice plan into todo.rst or similar so we do not have
> > to hunt it down by asking google?
> >
> > For the whole series you can add my:
> > Reviewed-by: Sam Ravnborg <sam at ravnborg.org>
> >
> > Some parts are reviewed as "this looks entirely correct", other parts
> > I would claim that I actually understood.
> > And after having spend some hours on this a r-b seems in order.
>
> Thanks, Daniel
> --
> Daniel Vetter
> Software Engineer, Intel Corporation
> http://blog.ffwll.ch



-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch