[pulseaudio-discuss] [PATCH] Do not use tsched watermark if tsched is disabled

[David Henningsson]

On 2010-09-08 02:52, pl bossart wrote:
>>> So what does this test mean? pavucontrol obviously affects the latency
>>> of the sink due to it's VI meters. This obviously increases the
>>> likelihood of a rewind being triggered.So, with this in mind what
>>> values do you suggest we pick?
>>
>> Pierre, is it your understanding that it is when DMA transfer collides with
>> cpu->RAM transfer that makes the DMA stream to become broken permanently? Or
>> exactly what is it that makes it break?
>
> The low-latency setting has nothing to do with rewinds.

Indirectly it is - if the rewind margin is greater than the actual 
latency there won't be any rewind.

> It's when you
> actually change the volume that you will rewind and remix with the new
> stream volume. You could set a super low latency and rewind exactly
> once when the stream starts.
>
> There are two issues with the rewind. One is a race condition between
> the DMA and the ring buffer update. If you rewind and update the ring
> buffer, the DMA may actually have already fetched older data.

To fetch the old data is not a problem IMO, that gives us just a ms more 
of the old data and we might miss a ms of the new data. That's not 
optimal, but not a complete disaster either.

What is the big problem here, is when we get a persistent crackling 
(even when everything has straightened out after the rewind!), and if I 
understand you correctly, that has something to do with the DMA 
controller...? Could you elaborate on what conditions it is that cause 
the persistent crackling?

> The
> second is that you could experience underflows if you don't do the
> update fast enough. By enabling logs you should be able to find out if
> there are real underflows.

Exactly.

-- 
David Henningsson, Canonical Ltd.
http://launchpad.net/~diwic