Received-SPF: pass (google.com: domain of linux-bluetooth-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <CAOOgLvqcSoaM7hgdzUA+JK0qgHfjhKKcLS2ScWNvsEiEj4xkoQ@mail.gmail.com>
 <20201019141608.401899-1-marijns95@gmail.com> <CAOOgLvoowRsJbdK-nyFi9V8hDAaBBA5j3ohwDoD0QVSdUF3h=A@mail.gmail.com>
 <CABBYNZKxrfR2rKuZt-T2_u2KXcS7d=nE-BuNQm+hZ7zuPb-y5A@mail.gmail.com>
 <CANX-K3s2eRaJeLScJpaUp=idkdjMhLHSAKTkyAhXTBWyK2Chaw@mail.gmail.com> <CAOOgLvo321AHC1aEv7N8EC4vzwTNUAohfWd02=Z1g2mt+ia7ew@mail.gmail.com>
In-Reply-To: <CAOOgLvo321AHC1aEv7N8EC4vzwTNUAohfWd02=Z1g2mt+ia7ew@mail.gmail.com>
From:   Marijn Suijten <marijns95@gmail.com>
Date:   Tue, 20 Oct 2020 11:11:03 +0200
Message-ID: <CANX-K3sX25OwpJSoDkA9=8bZcSD0jME0NsrRj0nk62Dh+NghtQ@mail.gmail.com>
Subject: Re: avrcp: possible race condition during connection establishment
To:     Marek Czerski <ma.czerski@gmail.com>
Cc:     Luiz Augusto von Dentz <luiz.dentz@gmail.com>,
        "linux-bluetooth@vger.kernel.org" <linux-bluetooth@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

Hi Marek,

(Apologies for sending this to you twice; missed the reply-all button
so this didn't end up on the mailing list)

On Tue, 20 Oct 2020 at 09:31, Marek Czerski <ma.czerski@gmail.com> wrote:
>
> Hi Marijn,
>
> pon., 19 pa=C5=BA 2020 o 21:53 Marijn Suijten <marijns95@gmail.com> napis=
a=C5=82(a):
> >
> > Hi Luiz, Marek,
> >
> > On Mon, 19 Oct 2020 at 18:47, Luiz Augusto von Dentz
> > <luiz.dentz@gmail.com> wrote:
> > >
> > > Hi Marek, Marijn,
> > >
> > > On Mon, Oct 19, 2020 at 9:06 AM Marek Czerski <ma.czerski@gmail.com> =
wrote:
> > > >
> > > > Hi Marijn,
> > > >
> > > > pon., 19 pa=C5=BA 2020 o 16:16 Marijn Suijten <marijns95@gmail.com>=
 napisa=C5=82(a):
> > > > >
> > > > > Hi Marek,
> > > > >
> > > > > > Hi all,
> > > > > >
> > > > > > I was looking into, so called, absolute volume control that was
> > > > > > introduced in AVRCP v1.4. What I want to achieve is to send aud=
io from
> > > > > > android smartphone to linux based device running bluez and make=
 the
> > > > > > volume control on the smartphone side to control the volume on =
the
> > > > > > device. So the device is the a2dp sink + avrcp CT/TG and the ph=
one is
> > > > > > a2dp source + avrcp CT/TG.
> > > > > >
> > > > > > I assume that if all is working correctly then on the dbus the =
Volume
> > > > > > property of the org.bluez.MediaTransport1 will be changed by th=
e
> > > > > > volume control of the phone and changes made to this property f=
rom the
> > > > > > device would propagate to the phone volume slider.
> > > > > >
> > > > > > This is not happening and what I believe is the cause is that
> > > > > > AVRCP_EVENT_VOLUME_CHANGED event registration request sent from=
 the
> > > > > > phone is rejected by the bluez. I can see that on the wireshark=
 snoop
> > > > > > from the device's bluetooth adapter. And on wireshark I see tha=
t
> > > > > > AVRCP_EVENT_VOLUME_CHANGED event registration is sent by the ph=
one
> > > > > > before bluez initializes session->supported_events variable (no=
t
> > > > > > really sure about that). I think that this rejection makes the =
phone
> > > > > > to not send SetAbsoluteVolume commands to the device on volume =
change.
> > > > >
> > > > > I looked into the same issue earlier this year, see
> > > > > 20200118194841.439036-1-marijns95@gmail.com [1].  The gist of it =
is that BlueZ
> > > > > bases supported_events solely on the remote AVRCP controller vers=
ion, which
> > > > > Android sets to 1.3 when it is a media source [2].  This version =
is not
> > > > > relevant in your use-case because the Android phone is the AVRCP =
target while
> > > > > blueZ is the controller.
> > > > >
> > > >
> > > > I didn't tested Your patch but after looking at the code it seems t=
hat
> > > > just applying Your patch would solve my problem.
> >
> > It should, even though I don't immediately see how swapping
> > target/controller_init as per your initial email solves the problem.
> >
>
> The sequence of events is important here. The
> session->supported_events must be initialised (which is done in
> target_init) before first AVRCP_REGISTER_NOTIFICATION command from the
> remote side controller. If supported_events is not initialised and
> AVRCP_REGISTER_NOTIFICATION arrives the
> avrcp_handle_register_notification function rejects the event
> registration. And this is what I see in wireshark.

That's interesting and means we are dealing with two distinct problems
here. No matter what order I flip controller/target_init in
AVRCP_REGISTER_NOTIFICATION is always received after both functions
complete. handle_vendordep_pdu is registered before these
initialization functions making this possible though. I wonder if PDUs
are dropped when the callback is registered later, after init. Perhaps
controller/target_init need to be split in two, a static
initialization part based on the remote profile version followed by
registering this callback and finally functions that send commands to
the remote like avrcp_get_capabilities.

> And in my case, calling target_init earlier (the is before
> controller_init) solves the problem because supported_events is
> already initialised at the time the AVRCP_REGISTER_NOTIFICATION
> arrives.

That means your phone reports an AVRCP remote version of at least
0x104, otherwise supported_events would not contain
AVRCP_EVENT_VOLUME_CHANGED. Can you validate this in wireshark or
DBG("%p version 0x%04x", ...) in the logs? I'm seeing:

    profiles/audio/avrcp.c:target_init() 0x5614a2cbab70 version 0x0103

Either way the split Luiz and I mentioned mean one variant of
supported_events (the one relevant for registering
AVRCP_EVENT_VOLUME_CHANGED) is initialized in controller_init.

> But I think that the problem is more general. There is some gap
> between AVCTP connection establishment and BlueZ readiness to handle
> commands from the remote side. There could be devices that send
> commands during this gap and those devices would not work correctly
> with BlueZ despite the fact that they are consistent with the
> specification (or not ?).
>
> > > > Regarding avrcp version, in android there is developer option to se=
t
> > > > avrcp version. For example my Xiaomi redmi 8 (android 10) reports
> > > > version according to this setting, but samsung galaxy s7 (android 8=
)
> > > > always report version 1.4 regardless of this setting.
> >
> > I think this is the version used for the AVRCP Remote Target [1]. The
> > version linked above, the one that is causing problems here is that of
> > the AVRCP controller.
> >
> > There have always been problems changing these settings in developer
> > options - sometimes they work, sometimes they don't. The system has
> > been overhauled for Android 11 and is now much more consistent though
> > avrcp still runs through a property. If I remember correctly
> > restarting bluetooth or re-adding a device (in case of SDP caching) is
> > the only solution.
> >
> > > We need to rework a little bit how the controller/target works, this
> > > roles are actually supposed to be interpreted as client/server and
> > > much like GATT they can be used simultaneously, so we need a target
> > > versions and a controller version and independent supported events.
> >
> > Yes, I think that's the change we proposed last time and are proposing
> > now. Unfortunately relevant functions need to know if the local
> > (BlueZ) end is operating as target or controller for that particular
> > operation.
> >
> > > That said if the remote side controller version does not indicate 1.4
> > > or later we obviously can't support absolute volume control as that i=
s
> > > reserved in earlier versions.
> >
> > That depends on the way audio is flowing. When the local (BlueZ) end
> > is the sink (rendering the audio) meaning it takes the role of AVRCP
> > controller (the situation in the mail from Marek), the remote must be
> > the AVRCP target and hence its controller version is irrelevant.
> > Likewise when BlueZ is the audio source (AVRCP target), it should only
> > care about the remote controller profile and version.
> > But that is exactly what you described above: BlueZ needs to attain
> > separation between handling notification registrations (and perhaps
> > other code) as a controller and target separately.
> >
>
> I don't know if I'm interpreting your words correctly but I think that
> You make an assumption that a2dp sink must be avrcp CT and a2dp source
> is avrcp TG.
> In case of the absolute volume control it is not true. For absolute
> volume control to work it is the other way around which makes both
> sides be CT and TG at the same time.

Quite the contrary, in my previous mail I attempted to explain
multiple times that the CT/TG designation in the spec is used for
initiator and target respectively, not the AVRCP controller and target
profile. Indeed, depending on the transaction either side (controller
or target) can be the initiator (CT). In other words CT and TG are
_not_ abbreviations for the controller and target profile.

> a2dp sink is avrcp CT for
> controlling playback and a2dp source is avrcp CT for setting the
> volume. It is the a2dp source that is sending
> AVRCP_SET_ABSOLUTE_VOLUME commands.

Correct, yet the A2DP sink will always take on the AVRCP controller
profile where the phone being the A2DP source takes the AVRCP target
profile.

> > > > > It was decided in that mail thread to split supported_events in t=
wo; one based
> > > > > on the external controller version (when BlueZ operates as target=
 it'll
> > > > > validate incoming notification registrations) and the other based=
 on what BlueZ
> > > > > currently supports as controller.
> > > > >
> > > > > The second check might not be all too relevant and is already cov=
ered by the
> > > > > switch-case; perhaps it makes more sense to base this check on th=
e external
> > > > > target version, and again validate whether we expect to receive t=
hat particular
> > > > > notification registration?
> > > > >
> > > > > Both checks together implicitly validate what BlueZ supports loca=
lly in its
> > > > > role of controller or target, as remote_{target,controller}_suppo=
rted_events
> > > > > (anticipated names of the new members replacing supported_events)=
 will only be
> > > > > set to events that BlueZ is able to emit.
> > > > >
> > > >
> > > > One thing is not clear for me, what is the purpose of the
> > > > supported_events ?  It is used in two places:
> > > > First is the avrcp_handle_register_notification function. If the
> > > > remote side want to register itself for specific event notification=
 it
> > > > does not matter what version of avrcp that remote side supports. If=
 it
> > > > ask for specific event it clearly support that event.
> >
> > Looking at commit 4ae6135 that introduced this check it was apparently
> > causing crashes without. I can only guess that devices were sending
> > vendor-specific notification requests, assuming implementations are
> > supposed to filter out anything beyond what their reported version
> > supports? Either way input validation is still a sane thing to do.
> >
> > > > Second is in avrcp_handle_get_capabilities in CAP_EVENTS_SUPPORTED
> > > > case. Does it matter if local side reply with events that are not
> > > > supported in the version of avrcp supported by the remote side ?
> >
> > I guess it is sane here to not report capabilities the CT should -
> > given it's version - not know about. I could not find anything in the
> > 1.6.2 spec mandating this though, perhaps Luiz knows.
> >
> > > As I said we need to split the supported events to
> > > ct(client)/tg(server) to avoid interpreting them as the same, it is
> > > very odd that the remote would have different versions for each role
> >
> > Yeah, Android exposes different versions for its AVRCP Remote and
> > AVRCP Remote Target.
> > This is slightly confusing though: we discussed earlier that CT and TG
> > is about the initiator and receiver respectively, with no direct
> > mapping to AVRCP remote and target as either side can be the initiator
> > (sender) of a command. Correct?
> > In this specific case a notification registration will always be
> > initiated by a CT and fulfilled by the TG, no matter whether it is an
> > AVRCP remote or target.
> >
> > > but it looks like this is happening in this case although it is work
> > > confirming if the CT version if in fact 1.3 as well we cannot enable
> > > absolute volume control as that is not supported by that version, wha=
t
> > > we can perhaps is to detect if SetAbsoluteVolume control is used then
> > > update the events for the session.
> >
> > Again, in the specific case of this issue it is fine if the CT
> > (initiator, the Android phone, supposed to behave like the target as
> > it is the audio source) reports a controller version of 1.3 as we are
> > only concerned about the target version, which will be the one
> > registering for EVENT_VOLUME_CHANGED notifications from the
> > sink/controller (BlueZ). Android always reports the AVRCP Remote
> > Target as 1.4 or higher based on developer settings [1].
> >
> > > > > Unfortunately my ramblings in that mail shadowed an important que=
stion: how to
> > > > > determine in avrcp_handle_register_notification whether BlueZ is =
running as
> > > > > controller or target?  set_volume in transport.c derives this fro=
m
> > > > > transport->source_watch but there seems to be no easy access to t=
he
> > > > > accompanying transport in avrcp_handle_register_notification.  Wi=
th this
> > > > > question answered I'll be able to update and resubmit the origina=
l patch.
> > > > >
> > > > > > To test my theory i changed the session_init_control function i=
n the
> > > > > > profiles/audio/avrcp.c to call first target_init and then
> > > > > > controller_init. This caused  the AVRCP_EVENT_VOLUME_CHANGED ev=
ent not
> > > > > > been rejected and the volume control from the phone works as ex=
pected.
> > > > > >
> > > > > > After reading AVRCP specification I did not find any reason for=
 the CT
> > > > > > on the phone side not to send event registration immediately af=
ter the
> > > > > > AVCTP connection establishment. So I believe that bluez should =
not
> > > > > > reject event registration in this case.
> > > > > >
> > > > > > Best Regards,
> > > > > > Marek Czerski
> > > > >
> > > > > Best regards,
> > > > > Marijn Suijten
> > > > >
> > > > > [1]: https://marc.info/?l=3Dlinux-bluetooth&m=3D157937699001093
> > > > > [2]: https://android.googlesource.com/platform/system/bt/+/androi=
d-11.0.0_r4/bta/av/bta_av_main.cc#761
> > > >
> > > >
> > > > Best regards,
> > > > Marek Czerski
> > >
> > >
> > >
> > > --
> > > Luiz Augusto von Dentz
> >
> > P.S.: I hope it is fine to respond to two emails in one, seems like
> > that will get confusing real quick if depth increases.
> >
> > - Marijn
> >
> > [1]: https://android.googlesource.com/platform/system/bt/+/android-11.0=
.0_r4/bta/av/bta_av_main.cc#612
>
> Best regards,
> Marek

- Marijn