Re: [PATCH] media: don't drop front-end reference count for ->detach

From: Arnd Bergmann
Date: Wed Jan 03 2018 - 18:03:04 EST


On Wed, Jan 3, 2018 at 11:23 AM, Mauro Carvalho Chehab
<mchehab@xxxxxxxxxx> wrote:
> Em Tue, 2 Jan 2018 10:48:54 +0100
> Arnd Bergmann <arnd@xxxxxxxx> escreveu:

>> @@ -2965,7 +2968,6 @@ void dvb_frontend_detach(struct dvb_frontend* fe)
>> dvb_frontend_invoke_release(fe, fe->ops.release_sec);
>> dvb_frontend_invoke_release(fe, fe->ops.tuner_ops.release);
>> dvb_frontend_invoke_release(fe, fe->ops.analog_ops.release);
>> - dvb_frontend_invoke_release(fe, fe->ops.detach);
>> dvb_frontend_put(fe);
>
> Hmm... stb0899 is not the only driver using detach:
>
> drivers/media/dvb-frontends/stb0899_drv.c: .detach = stb0899_detach,
> drivers/media/pci/saa7146/hexium_gemini.c: .detach = hexium_detach,
> drivers/media/pci/saa7146/hexium_orion.c: .detach = hexium_detach,
> drivers/media/pci/saa7146/mxb.c: .detach = mxb_detach,
> drivers/media/pci/ttpci/av7110.c: .detach = av7110_detach,
> drivers/media/pci/ttpci/budget-av.c: .detach = budget_av_detach,
> drivers/media/pci/ttpci/budget-ci.c: .detach = budget_ci_detach,
> drivers/media/pci/ttpci/budget-patch.c: .detach = budget_patch_detach,
> drivers/media/pci/ttpci/budget.c: .detach = budget_detach,

I'm pretty sure I checked this before and found stb0899_detach to be the
only implementation of dvb_frontend_ops:detach.

The other eight you quoted are all setting the member in struct
saa7146_extension, not struct dvb_frontend_ops, so they are
not called using the same method.

> Unfortunately, I don't have any device that would be affected by
> this change, but it sounds risky to not call this code anymore:
>
> #ifdef CONFIG_MEDIA_ATTACH
> dvb_detach(release);
> #endif
>
> for .detach ops, as it has the potential of preventing unbind on
> those drivers.

The problem that Wolfgang reported originally was specifically that
this dvb_detach() call was one more than there should be,
leading to a negative reference count. As far as I can tell, this
is true for every user of the 'detach' callback.

>> diff --git a/drivers/media/usb/dvb-usb/pctv452e.c b/drivers/media/usb/dvb-usb/pctv452e.c
>> index 0af74383083d..ae793dac4964 100644
>> --- a/drivers/media/usb/dvb-usb/pctv452e.c
>> +++ b/drivers/media/usb/dvb-usb/pctv452e.c
>> @@ -913,14 +913,6 @@ static int pctv452e_frontend_attach(struct dvb_usb_adapter *a)
>> &a->dev->i2c_adap);
>> if (!a->fe_adap[0].fe)
>> return -ENODEV;
>> -
>> - /*
>> - * dvb_frontend will call dvb_detach for both stb0899_detach
>> - * and stb0899_release but we only do dvb_attach(stb0899_attach).
>> - * Increment the module refcount instead.
>> - */
>> - symbol_get(stb0899_attach);
>
>
> IMHO, the safest fix would be, instead, to do:
>
> #ifdef CONFIG_MEDIA_ATTACH
> symbol_get(stb0899_attach);
> #endif
>
> Btw, we have some code similar to that on other drivers
> with either symbol_get() or symbol_put().

This would work if we do it for every user of a driver that
has a detach callback in dvb_frontend_ops, but I still think
my suggested patch is correct and less error-prone as a
workaround here.

> Yeah, I agree that this sucks. The right fix here is to use i2c high level
> interfaces, binding it via i2c bus, instead of using the symbol_get()
> based dvb_attach() macro.
>
> We're (very) slowing doing such changes along the media subsystem.

Good to know there is a proper solution already. I had some ideas
for how to change it, but the i2c bus method is clearly better than
whatever I had in mind there.

Arnd