On Fri, Jan 9, 2015 at 1:50 PM, Jani Nikula jani.nikula@intel.com wrote:
I have a slightly uneasy feeling about handing out drm_panel pointers (both from here and of_drm_find_panel) without refcounting. If the panel driver gets removed, whoever called the find functions will have a dangling pointer. I supposed this will be discussed on drm-devel.
There's been some discussion already about exactly this problem (but with drm bridges) with Thierry and some other people. Cc'ed them all hopefully. Especially when the panel/bridge is a separate driver there's imo indeed an issue. -Daniel
On 01/13/2015 12:08 AM, Daniel Vetter wrote:
On Fri, Jan 9, 2015 at 1:50 PM, Jani Nikula jani.nikula@intel.com wrote:
I have a slightly uneasy feeling about handing out drm_panel pointers (both from here and of_drm_find_panel) without refcounting. If the panel driver gets removed, whoever called the find functions will have a dangling pointer. I supposed this will be discussed on drm-devel.
refcounting does not seems to me a good solution, drm_panel is exposed by device driver and device driver can be unbound unconditionally at any time. This problem affects many frameworks not only drm_panel.
I work on resource tracking framework which tries to solve the problem in a generic way[1].
[1]: https://lkml.org/lkml/2014/12/10/342
Regards Andrzej
There's been some discussion already about exactly this problem (but with drm bridges) with Thierry and some other people. Cc'ed them all hopefully. Especially when the panel/bridge is a separate driver there's imo indeed an issue. -Daniel
On Tue, Jan 13, 2015 at 12:08:11AM +0100, Daniel Vetter wrote:
On Fri, Jan 9, 2015 at 1:50 PM, Jani Nikula jani.nikula@intel.com wrote:
I have a slightly uneasy feeling about handing out drm_panel pointers (both from here and of_drm_find_panel) without refcounting. If the panel driver gets removed, whoever called the find functions will have a dangling pointer. I supposed this will be discussed on drm-devel.
There's been some discussion already about exactly this problem (but with drm bridges) with Thierry and some other people. Cc'ed them all hopefully. Especially when the panel/bridge is a separate driver there's imo indeed an issue.
I posted patches some time ago to create a generic registry to do the actual ref-counting[0]. It didn't seem to be very well received by Greg for the core, so perhaps we could test-drive it in DRM first for panels and bridges and once it's matured a bit it could still be promoted to the driver core, or maybe lib/.
The difficult part about it is that while reference counting gives you the benefit of keeping a valid pointer around, you may still want to have a method of getting notified of the panel going away. I've thought a bit about that and I think we could probably integrate that into the registry, since that will notice anyway.
Thierry
[0]: http://www.spinics.net/linux/lists/kernel/msg1859333.html
dri-devel@lists.freedesktop.org
-
Andrzej Hajda -
Daniel Vetter -
Thierry Reding