Hi Jo?o Paulo,
On Mon, Apr 29, 2013 at 5:51 PM, Jo?o Paulo Rechi Vita
<jprvita at openbossa.org> wrote:
> On Mon, Apr 29, 2013 at 11:59 AM, Tanu Kaskinen <tanu.kaskinen at intel.com> wrote:
>> On Mon, 2013-04-29 at 11:49 -0300, Jo?o Paulo Rechi Vita wrote:
>>> On Mon, Apr 29, 2013 at 11:38 AM, Tanu Kaskinen <tanu.kaskinen at intel.com> wrote:
>>> > On Mon, 2013-04-29 at 16:20 +0200, Mikel Astiz wrote:
>>> >> On Mon, Apr 29, 2013 at 4:01 PM, Tanu Kaskinen <tanu.kaskinen at intel.com> wrote:
>>> >> > On Fri, 2013-04-26 at 12:30 -0300, jprvita at gmail.com wrote:
>>> >> >> src/modules/bluetooth/module-bluetooth-device.c | 1 +
>>> >> >> 1 file changed, 1 insertion(+)
>>> >> >>
>>> >> >> diff --git a/src/modules/bluetooth/module-bluetooth-device.c b/src/modules/bluetooth/module-bluetooth-device.c
>>> >> >> index c877df2..8695c80 100644
>>> >> >> --- a/src/modules/bluetooth/module-bluetooth-device.c
>>> >> >> +++ b/src/modules/bluetooth/module-bluetooth-device.c
>>> >> >> @@ -2243,6 +2243,7 @@ static int add_card(struct userdata *u) {
>>> >> >> pa_proplist_sets(data.proplist, "bluez.path", device->path);
>>> >> >> pa_proplist_setf(data.proplist, "bluez.class", "0x%06x", (unsigned) device->class);
>>> >> >> pa_proplist_sets(data.proplist, "bluez.name", device->name);
>>> >> >> + pa_proplist_sets(data.proplist, "bluez.alias", device->alias);
>>> >> >
>>> >> > device->alias is not guaranteed to be non-NULL, so this code can crash.
>>> >> > If the Alias property is non-optional in Bluez, bluetooth-util should
>>> >> > ensure that it is indeed always set.
>>> >>
>>> >> BlueZ 4 API guarantees the property will be present, and therefore
>>> >> device->alias should always be set. The obvious exception is BlueZ
>>> >> misbehaving, but this condition might hold true for device->name as
>>> >> well, which we're not checking either.
>>> >>
>>> >> I don't think there's any other scenario where the alias could be NULL
>>> >> and the name be set outside bluetooth-util.
>>> >>
>>> >> So while your point is valid, I suggest we fix this in a later patch,
>>> >> in a similar way that the BlueZ 5 patchset does (set device_info_valid
>>> >> to -1 if any non-optional property is missing).
>>> >
>>> > Ah, so this is fixed at least for BlueZ 5. Will the same property
>>> > parsing code be used for BlueZ 4? If not, does somebody promise to write
>>> > the fix for BlueZ 4?
>>> >
>>>
>>> Both in BlueZ 5 and BlueZ 4 the 'Alias' property is not optional, if
>>> it's not present it's a bug in BlueZ. I don't think we should protect
>>> ourselfs more than the usual asserts before dereferencing pointers.
>>
>> Yes, but our policy is to use assertions only for guarding against bugs
>> in our own process. Assertions are not used for handling bugs in other
>> processes, such as bluetoothd.
>>
>>> If
>>> the property is not set but everything else is fine the only thing
>>> that will be missing is a value for PA_PROP_DEVICE_DESCRIPTION.
>>> Depending on how PulseAudio deals with description-less devices we can
>>> either not care or set a default description. Setting the device_info
>>> to invalid makes the device unusable.
>>
>> I'm not sure if you're saying that it's fine to call pa_proplist_sets()
>> with a NULL value? It's not fine, it will cause a crash (due to an
>> assertion in pa_proplist_sets()).
>>
>
> No, I'm asking if it's fine not to call
> 'pa_proplist_sets(data.proplist, PA_PROP_DEVICE_DESCRIPTION, "Some
> string here")' before calling 'pa_card_new(core, &data)', that is,
> never setting the value of PA_PROP_DEVICE_DESCRIPTION. If it's not
> fine (that is, we cannot leave PA_PROP_DEVICE_DESCRIPTION empty) I
> suggest we set it to a default value, or an empty string, instead of
> making the device unusable.
>
>> Making the device unusable is an acceptable (and IMHO the best) way of
>> handling BlueZ bugs. But if everybody else thinks that doing it for
>> BlueZ 4 is waste of time, I won't insist on doing it.
>>
>
> I don't think is a waste of time, and this discussion is also valid
> for how this case should be handled in BlueZ 5. I agree the bug is in
> BlueZ in this case, but if we can recover without any operational loss
> from it, I think it's better to do so.
As Tanu already pointed out, assertions should be limited to detecting
internal issues in PulseAudio, regardless of external misbehaviors
(with the exception of libraries IIRC). There have been similar
discussions in the mailing list before.
In the particular case of D-Bus misbehaviors, I believe we should
theoretically reset the connection completely, since we can't make any
assumptions about a component that is not behaving as expected.
However, being more pragmatic, having a fallback mechanism in some
cases seems reasonable, like in this example, where we can ignore
objects (i.e. devices) that lack some of the non-optional properties.
We should IMO avoid trying to set default values for missing
non-optional properties, since this obfuscates the codebase and might
hide the root issue.
Cheers,
Mikel
