Re: [RFCv4,19/21] media: vim2m: add request support

[Dmitry Osipenko <digetx@gmail.com>]

On 12.03.2018 11:29, Tomasz Figa wrote:
> On Mon, Mar 12, 2018 at 5:25 PM, Paul Kocialkowski
> <paul.kocialkowski@bootlin.com> wrote:
>> Hi,
>>
>> On Mon, 2018-03-12 at 17:15 +0900, Tomasz Figa wrote:
>>> Hi Paul, Dmitry,
>>>
>>> On Mon, Mar 12, 2018 at 5:10 PM, Paul Kocialkowski
>>> <paul.kocialkowski@bootlin.com> wrote:
>>>> Hi,
>>>>
>>>> On Sun, 2018-03-11 at 22:42 +0300, Dmitry Osipenko wrote:
>>>>> Hello,
>>>>>
>>>>> On 07.03.2018 19:37, Paul Kocialkowski wrote:
>>>>>> Hi,
>>>>>>
>>>>>> First off, I'd like to take the occasion to say thank-you for
>>>>>> your
>>>>>> work.
>>>>>> This is a major piece of plumbing that is required for me to add
>>>>>> support
>>>>>> for the Allwinner CedarX VPU hardware in upstream Linux. Other
>>>>>> drivers,
>>>>>> such as tegra-vde (that was recently merged in staging) are also
>>>>>> badly
>>>>>> in need of this API.
>>>>>
>>>>> Certainly it would be good to have a common UAPI. Yet I haven't
>>>>> got my
>>>>> hands on
>>>>> trying to implement the V4L interface for the tegra-vde driver,
>>>>> but
>>>>> I've taken a
>>>>> look at Cedrus driver and for now I've one question:
>>>>>
>>>>> Would it be possible (or maybe already is) to have a single IOCTL
>>>>> that
>>>>> takes input/output buffers with codec parameters, processes the
>>>>> request(s) and returns to userspace when everything is done?
>>>>> Having 5
>>>>> context switches for a single frame decode (like Cedrus VAAPI
>>>>> driver
>>>>> does) looks like a bit of overhead.
>>>>
>>>> The V4L2 interface exposes ioctls for differents actions and I don't
>>>> think there's a combined ioctl for this. The request API was
>>>> introduced
>>>> precisely because we need to have consistency between the various
>>>> ioctls
>>>> needed for each frame. Maybe one single (atomic) ioctl would have
>>>> worked
>>>> too, but that's apparently not how the V4L2 API was designed.
>>>>
>>>> I don't think there is any particular overhead caused by having n
>>>> ioctls
>>>> instead of a single one. At least that would be very surprising
>>>> IMHO.
>>>
>>> Well, there is small syscall overhead, which normally shouldn't be
>>> very painful, although with all the speculative execution hardening,
>>> can't be sure of anything anymore. :)
>>
>> Oh, my mistake then, I had it in mind that it is not really something
>> noticeable. Hopefully, it won't be a limiting factor in our cases.
> 
> With typical frame rates achievable by hardware codecs, I doubt that
> it would be a limiting factor. We're using a similar API (a WiP
> version of pre-Request API prototype from long ago) in Chrome OS
> already without any performance issues.

Thank you very much for the answers!

The syscalls overhead is miserable in comparison to the rest of decoding, though
I wanted to clarify whether there is a way to avoid it. Atomic API sounds like
something that would suit well for that.