Tasklet usage in the DRM

Tasklet usage in the DRM

[Michel Dänzer]

I just read an article on LWN's kernel page about plans to remove
tasklets, and I thought I'd explain what the DRM is using tasklets for.
Maybe there's other ways to satisfy the requirements equally well or
even better.


The i915 driver uses a tasklet to make sure a GL buffer swap blit or
flip takes effect (or at least starts in the case of a blit) during the
vertical blank period, to avoid tearing. I chose a tasklet for this
purpose because:

      * The traditional method of the vertical blank interrupt waking up
        the user process, which would then emit the buffer swap
        commands, didn't even come close to avoiding tearing. So I
        suspect a workqueue wouldn't cut it either.
      * The processing of scheduled buffer swaps could potentially take
        a long time, so doing it in hardirq context could cause high IRQ
        latency.
      * It requires holding the DRM lock, so the 'each tasklet can only
        run once at a time' restriction is fine.


I'm looking forward to any suggestions what to do with this in case
tasklets disappear, and I'll gladly provide further clarification on the
requirements.


-- 
Earthling Michel Dänzer           |          http://tungstengraphics.com
Libre software enthusiast         |          Debian, X and DRI developer

Re: Tasklet usage in the DRM

[Daniel Walker]

On Fri, 2007-06-29 at 09:09 +0200, Michel Dänzer wrote:
> I just read an article on LWN's kernel page about plans to remove
> tasklets, and I thought I'd explain what the DRM is using tasklets for.
> Maybe there's other ways to satisfy the requirements equally well or
> even better.
> 
> 
> The i915 driver uses a tasklet to make sure a GL buffer swap blit or
> flip takes effect (or at least starts in the case of a blit) during the
> vertical blank period, to avoid tearing. I chose a tasklet for this
> purpose because:
> 
>       * The traditional method of the vertical blank interrupt waking up
>         the user process, which would then emit the buffer swap
>         commands, didn't even come close to avoiding tearing. So I
>         suspect a workqueue wouldn't cut it either.
>       * The processing of scheduled buffer swaps could potentially take
>         a long time, so doing it in hardirq context could cause high IRQ
>         latency.
>       * It requires holding the DRM lock, so the 'each tasklet can only
>         run once at a time' restriction is fine.
> 
> 
> I'm looking forward to any suggestions what to do with this in case
> tasklets disappear, and I'll gladly provide further clarification on the
> requirements.

If it hasn't been done someone with this hardware could test Steven's
patches to see if they do satisfy the requirements..

I think the most recent version was posted in this message,
http://lkml.org/lkml/2007/6/27/289

Subject was "[RFC PATCH 0/8] Convert all tasklets to workqueues V3" if
people are subscribed to LKML ..

(Added Steven to the CC list)

Daniel