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;
Content-Type: text/plain;
        charset=utf-8
Mime-Version: 1.0 (Mac OS X Mail 14.0 \(3654.40.0.2.32\))
Subject: Re: Default to HW mSBC on capable controllers ?
From:   Marcel Holtmann <marcel@holtmann.org>
In-Reply-To: <20201222131749.vhatruzz3ztnw6k5@pali>
Date:   Tue, 22 Dec 2020 22:11:40 +0100
Cc:     Luiz Augusto von Dentz <luiz.dentz@gmail.com>,
        Joakim Tjernlund <Joakim.Tjernlund@infinera.com>,
        "linux-bluetooth@vger.kernel.org" <linux-bluetooth@vger.kernel.org>
Content-Transfer-Encoding: 8BIT
Message-Id: <D4997390-D306-420E-9A4F-6A241EAB0A1E@holtmann.org>
References: <CY4PR1001MB2389075CC44E480B446535E5F4C30@CY4PR1001MB2389.namprd10.prod.outlook.com>
 <CABBYNZKy4KXWqLdZu7C49jJ_nMbmBOdMjtM2_5OQg2ruUHUh_w@mail.gmail.com>
 <20201221211437.4s27cl6t4v27sugh@pali>
 <CABBYNZ+tGt4Duf=aYzWPG0OSATj0ZN6oQeFmvw=Un_JVK9C_rQ@mail.gmail.com>
 <20201222092100.ru5inf45v55qoa4m@pali>
 <BC13598F-C074-4CB3-B24E-10A6E0CF84DB@holtmann.org>
 <20201222101930.awcxtsgltrrde2d3@pali>
 <F698F31D-063A-4A9E-B2BD-D61FA937F561@holtmann.org>
 <20201222114628.pq3lmrvzepel3dj4@pali>
 <FEEDD760-B78B-44F5-9561-A0CFF08AF515@holtmann.org>
 <20201222131749.vhatruzz3ztnw6k5@pali>
To:     =?utf-8?Q?Pali_Roh=C3=A1r?= <pali@kernel.org>
Precedence: bulk

Hi Pali,

>>>>>>>>>>> There seems to be quite a few USB controllers gaining the BTUSB_WIDEBAND_SPEECH which I guess means HW mSBC but currently there is no way to select this mode.
>>>>>>>>>>> Any idea if one could patch the kernel to default to HW mSBC and user apps like bluealsa/pulseaudio would just use it automatically?
>>>>>>>>>> 
>>>>>>>>>> It is in our plan to support HW offloading, but that doesn't mean all
>>>>>>>>>> platforms will be supported since that depends on the PCM lines being
>>>>>>>>>> connected to BT controller in the first place.
>>>>>>>>> 
>>>>>>>>> Dedicated PCM lines are used in embedded world and maybe also still in
>>>>>>>>> some mobile segment. I remember that e.g. Nokia N900 had this setup. And
>>>>>>>>> it was quite crazy how it was finally configured... but it worked!
>>>>>>>>> 
>>>>>>>>> But this is nothing for classic x86 laptops with USB bluetooth
>>>>>>>>> controllers on classic intel bluetooth+wifi mPCIe cards where SCO
>>>>>>>>> traffic is routed via HCI (over USB). And not via dedicated PCM pins.
>>>>>>>>> Moreover I think there are not any mainstream laptop which have PCM pins
>>>>>>>>> on mPCIe slots usable for such bluetooth mPCIe cards.
>>>>>>>>> 
>>>>>>>>> For classic desktop / laptop it is needed to deal with fact that SCO
>>>>>>>>> audio is routed via HCI (like A2DP) and therefore support for Enhanced
>>>>>>>>> Setup Synchronous Connection HCI command.
>>>>>>>>> 
>>>>>>>>> AFAIK even for routing SCO over PCM when mSBC hw encoder is used,
>>>>>>>>> Enhanced Setup Synchronous Connection HCI command is required.
>>>>>>>> 
>>>>>>>> So you are saying that we should do PCM over HCI and that would
>>>>>>>> actually work (meaning we have enough bandwidth)?
>>>>>>> 
>>>>>>> This is something which needs to be tested. And without full
>>>>>>> implementation (with control of all parameters) we cannot say YES or NO.
>>>>>>> 
>>>>>>> And if you are aware of bandwidth, Enhanced Setup Synchronous Connection
>>>>>>> HCI command allows you to use also software based CVSD codec. Meaning
>>>>>>> that CVSD encoding/decoding can be done by application and therefore
>>>>>>> decreasing amount of data to transfer to bluetooth adapter.
>>>>>>> 
>>>>>>> As I said this command is needed also if you want to use mSBC hw encoder
>>>>>>> over PCM, so I think usage of Enhanced Setup Synchronous Connection HCI
>>>>>>> command always have benefits to implement it (I have unfinished and
>>>>>>> untested implementation).
>>>>>> 
>>>>>> CVSD is an air codec only. Controller<->Host is PCM in that case.
>>>>> 
>>>>> Hello!
>>>>> 
>>>>> This is not fully truth. You can use also use CVSD between Controller
>>>>> and Host. Enhanced Setup Synchronous Connection HCI allows you to do it.
>>>> 
>>>> so the spec is unclear in case of HCI transport. It could be stipulated that CVSD is also possible. However it is not practical since the controller would have to re-encode in case a different air codec is used. The reality is that all “legacy” devices decided for CVSD air codec and PCM host codec.
>>>> 
>>>> Going into this direction of allowing CVSD as host codec is just a waste of time.
>>> 
>>> Well, people are complaining about CVSD quality and there are still lot
>>> of headsets which support only CVSD. Therefore I see a good reason to
>>> try encode/decode of CVSD in software and check if software encoder or
>>> decoder cannot increase this quality. I'm not saying that it can really
>>> improve it but without testing we do not know. And as I know how poor
>>> are some hardware implementations of codecs (but not all) I would not be
>>> surprised if for some bluetooth adapters this can really increase
>>> quality for CVSD transports... But without real tests we do not know.
>> 
>> I really don’t see how this is going to help. I have no interest in this. I think it is a waste of time.
> 
> It is simple. If hw encoder is poor you can write your own in software.
> Then you can compare audio quality of hw encoder in adapter and software
> encoder in userspace.
> 
> I'm sure that people are interested even in increasing quality of CVSD
> transport and they can spend time on own userspace codec.
> 
> I understand that it does not bring any value for you and it can be a
> waste of time for you.
> 
> But please do not block research in this field which can be done by
> other people independently of you.
> 
> The worst what can happen is that (independent) researcher just confirms
> what you have wrote. But even this would be very positive conclusion!

but I am not adding a new kernel-userspace API so that you can tinker with it only to realize that we then get an unneeded API that I have to support for years to come.

If you want this then do it with your own kernel patches and provide numbers that would show this works and improves certain devices. Unless I see that, I am sticking with my point of view.

>>>>>> Only with mSBC you have both air codec and Controller<->Host running with mSBC.
>>>>> 
>>>>> No, you can use it also with other codecs. With all transparent codecs
>>>>> and CVSD, alaw and ulaw based on capabilities of your bluetooth
>>>>> adapters.
>>>> 
>>>> The case when it is not CVSD, alaw and ulaw is pretty clear written in the spec. Both have to be transparent.
>>> 
>>> Yes, for air part it has to be either: CVSD, ALAW, ULAW or TRANSPARENT.
>>> So all other additional codecs use transparent on AIR part.
>>> 
>>> But on local part it can be also any other codec defined in some
>>> assigned number document.
>> 
>> That is not clearly defined. The Codec_Format is an enumeration that is used in multiple places. I think that table is missing an indication what values are valid for certain HCI commands and protocols. We did this for other assigned numbers as well.
>> 
>> As I said, I see zero benefit of using CVSD as input/output codec over HCI.
> 
> Yes, it is unclear and therefore research in this field by other people
> can be interesting. To know what is reality.
> 
> I'm not saying that you have to use it. My argument is to allow other
> people to do it and do not block them.

I am not blocking you writing kernel patches. I am just not putting them upstream just for sake of it.

>>>> I think that the spec and the assigned numbers document is not really clear in specifying which of the Coding_Format fields apply when HCI transport is used vs external channel.
>>> 
>>> It is months ago when I read specification last time, but I think that
>>> at the time when I read it I understood that also CVSD codec may be used
>>> by local part if is supported.
>>> 
>>>> I would really assume that when the host is running the codec, then it is transparent.
>>> 
>>> This is truth for "Setup Synchronous Connection". But not for "Enhanced
>>> Setup Synchronous Connection". In "Enhanced" version you can have local
>>> codec for host also "CVSD", also "mSBC" and also any other which adapter
>>> returns in the "local codec" list or in "vendor codec" list. And for
>>> local part in "Enhanced" version this is *not* transparent anymore.
>>> 
>>> Of course "Enhanced" version support also "transparent" code and behave
>>> like in classic non-enhanced version.
>> 
>> Read my comment again since I did check the spec. It is super unclear on what values are support for input/output codec format when using HCI transport. I bet you that once you try this, you will see limitation in the controllers and they are just telling you Invalid Parameters.
>> 
>>>> When the codec is offload to the controller, then the host transport is Linear PCM.
>>>> 
>>>> So you I would plan for CVSD<->CVSD<->PCM, Transparent<->{mSBC,LC3}<->PCM, Transparent<->Transparent<->{mSBC,LC3} for the SCO/eSCO cases.
>>> 
>>> You forgot AuriStream codec which is used and heavily supported by lot
>>> of headsets. Just people do not know that is is supported.
>> 
>> But then again, that is Transparent. So Transparent<->Transparent<->{mSBC,LC3,AuriStream,etc}.
> 
> Yes, it is transparent, but with different parameters which needs to be
> passed to Setup Synchronous Connection parameters. And because this
> codec is SW, encoding and decoding is done in software, userspace must
> be able to "tune" parameters for Setup Synchronous Connection command.
> 
> Without it, it is not possible to use AuriStream codec, even it is
> Transparent.
> 
>>> I would really suggest to look at my API for specifying codec parameters
>>> on socket and allow usage of "Enhanced Setup Synchronous Connection".
>>> 
>>> I do not think it is waste of time. At least it allows doing all kind of
>>> test what adapters support and what is the quality of transport.
>> 
>> I am not going to expose the full crazy that is that HCI command to userspace.
> 
> Ok, I understand that you do not like alaw and ulaw codecs. But do not
> block usage of AuriStream. This codec is supported by many headsets and
> has better quality than CVSD (which is currently supported by kernel).
> 
> Also, please keep in mind that alaw and ulaw codecs are
> telecommunication standards and these codecs are still used.
> 
> People leaving in Europe and working in telecommunication sector would
> know that alaw is codec main voice codec for audio PSTN calls and is
> used also in VOIP lines.
> 
> So even if you marked it as legacy, it does not mean that it is not
> used. Yes, it is legacy, but it is still there and tons of voice calls
> in Europe are encoded in alaw.

This argument is mood. They are not used as air codecs with Bluetooth. If you don’t use Wideband speech, it will be CVSD as air codec and nothing else.

> But at least, please do not blok usage of AuriStream codec by Linux
> kernel. People are complaining about bad quality, AuriStream can improve
> it, but this decision just say "we do not want improvement".
> 
> I was open for any discussion how to bring it, but for one year I have
> received any positive reaction how to achieve it. I designed API which
> would allow it but there was no discussion about it, how to either
> improve it or something else.

We can design something to allow tuning transparent setup, but if you want to play with CVSD as host codec, then you are on your own.

Regards

Marcel