Hi,

On Wed, Oct 15, 2014 at 06:43:40PM +0300, Tomi Valkeinen wrote:

quoted

quoted

Somehow omapfb device is still unbound from the driver, as I can then
bind it again, causing probe to be called. Which breaks everything.

I would've thought that unbinding is not possible if remove is missing,
but that doesn't seem to be the case. I guess it just means that remove
is not called when the driver & device are unbound.

if no remove it provided on platform_driver structure, platform bus
assumes you have nothing to do on your ->remove(), so you end up leaking
all resources you allocated on ->probe() (unless you *really* don't need
to do anything on ->remove).

Yep. That's quite odd, still. grep shows quite many uses of __exit_p(),
and all for remove callback. So, if you have something to release in
remove(), you should set it always, for both module and built-in. And if
you don't have anything to release, you would always just set .release
to NULL.

I mean, what's the use case for __exit_p()? With a quick glance, at
least some of the other users also use __exit_p() the same way omapdss
does (i.e. in the wrong way).

__exit_p() meant something else a few years back, perhaps those were
left over from some tree-wide cleanups.

quoted

quoted

We have 18 __exit_p()s in omapdss and related drivers. I guess they are
all broken the same way.

yup, I should've grepped.

quoted

Note that omapfb unbind & bind does not work even with this patch, but
results in a crash as some old state is left into omapdss. The same
happens also with unloading and loading omapfb module (but keeping
omapdss module loaded).

It worked fine for me. I unbound and bound omapfb multiple times.

Hmm, ok. Odd, the bug was quite clear and I think it should happen every
time. Well, I was using omap4. If you used AM4xx, that's basically omap3
DSS. Maybe there's a diff there.

could very well be :-)

quoted

quoted

So there seems to be more issues around this.

quite a few more, I'd say

Yep, I'll have a look at this.

alright

-- 
balbi