lkml.org 
[lkml]   [2018]   [Jun]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [RFC PATCH 1/2] media: docs-rst: Add decoder UAPI specification to Codec Interfaces
From
Date
On 06/07/2018 07:53 PM, Nicolas Dufresne wrote:
> Le jeudi 07 juin 2018 à 16:30 +0900, Tomasz Figa a écrit :
>>>> v4l2-compliance (so probably one for Hans).
>>>> testUnlimitedOpens tries opening the device 100 times. On a normal
>>>> device this isn't a significant overhead, but when you're allocating
>>>> resources on a per instance basis it quickly adds up.
>>>> Internally I have state that has a limit of 64 codec instances (either
>>>> encode or decode), so either I allocate at start_streaming and fail on
>>>> the 65th one, or I fail on open. I generally take the view that
>>>> failing early is a good thing.
>>>> Opinions? Is 100 instances of an M2M device really sensible?
>>>
>>> Resources should not be allocated by the driver until needed (i.e. the
>>> queue_setup op is a good place for that).
>>>
>>> It is perfectly legal to open a video node just to call QUERYCAP to
>>> see what it is, and I don't expect that to allocate any hardware resources.
>>> And if I want to open it 100 times, then that should just work.
>>>
>>> It is *always* wrong to limit the number of open arbitrarily.
>>
>> That's a valid point indeed. Besides the querying use case, userspace
>> might just want to pre-open a bigger number of instances, but it
>> doesn't mean that they would be streaming all at the same time indeed.
>
> We have used in GStreamer the open() failure to be able to fallback to
> software when the instances are exhausted. The pros was it fails really
> early, so falling back is easy. If you remove this, it might not fail
> before STREAMON. At least in GStreamer, it too late to fallback to
> software. So I don't have better idea then limiting on Open calls.

It should fail when you call REQBUFS. That's the point at which you commit
to allocating resources. Everything before that is just querying things.

STREAMON is way too late, but REQBUFS/CREATE_BUFS (i.e. when queue_setup
is called) is a good point. You already allocate memory there, you can
also claim the m2m hw resource(s) you need.

Regards,

Hans

\
 
 \ /
  Last update: 2018-06-07 21:36    [W:1.901 / U:0.468 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site