On 13.06.2017 18:07, Thierry Reding wrote:

On Tue, Jun 13, 2017 at 05:00:28PM +0300, Dmitry Osipenko wrote:

quoted

On 13.06.2017 16:43, Thierry Reding wrote:

quoted

On Tue, May 23, 2017 at 03:14:22AM +0300, Dmitry Osipenko wrote:

quoted

The framebuffers console fbcon_startup() increments the tegra_drm module
'use' refcount via try_module_get(), causing an interlock of the DRM subsys
and the tegra_drm modules. In result, the tegra_drm module can't be unloaded
using rmmod.

Signed-off-by: Dmitry Osipenko <redacted>
---
 drivers/gpu/drm/tegra/fb.c | 1 -
 1 file changed, 1 deletion(-)

That's done on purpose because otherwise you could just rip out the
driver from under the framebuffer emulation and things would crash.

My understanding is that the right way to unload a module is to unbind
the driver first (which will cause the framebuffer to be removed and
hence the reference to be dropped) before the rmmod.

Thierry

Aha, interesting. I'll try the unbinding and will drop this patch from the
series, thank you for the clarification. I haven't observed any crashes on a
module reloading (framebuffer detached/attached just fine), maybe I was lucky then.

It's possible that it works by accident. The driver removes the
framebuffer as part of the driver removal process, so technically
nothing should crash. However, if, for any reason, anyone was holding
on to a reference to the framebuffer (not sure if that is even possible)
the module needs to stay around long enough as well, otherwise the
function pointers would become dangling. The module reference makes sure
that this doesn't happen (as long as a framebuffer exists, the ops will
stick around).

So even if there isn't a way to make it crash today, the code is still
correct in taking the reference.

Thank you very much for a more detailed explanation. I've successfully unloaded
the tegra_drm module after unbinding the VT's, the patch is dropped from the series.

echo 0 | tee /sys/devices/virtual/vtconsole/vtcon*/bind

-- 
Dmitry