[Intel-gfx] [RFC PATCH] drm/nouveau: fix nested locking in mmap handler

[Ingo Molnar]

* Maarten Lankhorst <maarten.lankhorst at canonical.com> wrote:

> > I think the Nouveau guys need to comment further on this, but 
> > returning -EFAULT might break existing user-space, and that's not 
> > allowed, but IIRC the return value of "presumed" is only a hint, and 
> > if it's incorrect will only trigger future command stream patching.
> >
> > Otherwise reviewing mostly the TTM stuff. FWIW, from wat I can tell 
> > the vmwgfx driver doesn't need any fixups.
>
> Well because we read the list of buffer objects the presumed offsets are 
> at least read-mapped. Although I guess in the worst case the mapping 
> might disappear before the syscall copies back the data. So if -EFAULT 
> happens here then userspace messed up in some way, either by forgetting 
> to map the offsets read-write, which cannot happen with libdrm or 
> free'ing the bo list before the syscall returns, which would probably 
> result in libdrm crashing shortly afterwards anyway.
> 
> So I don't know whether to swallow that -EFAULT or not, which is what I 
> asked.

In such a case returning -EFAULT is very strongly preferred.

If there's a user-space bug, such as a context life time race between 
graphics context creation/destruction and user threads making use of the 
graphics context, then getting the -EFAULT would be very helpful to 
user-space debugging as well.

Thanks,

	Ingo