Re: [PATCH 1/4 v2] dmaengine: add a tx_free method to structdma_async_tx_descriptor

[Guennadi Liakhovetski]

On Wed, 10 Dec 2008, Dan Williams wrote:

> On Wed, Dec 10, 2008 at 5:30 PM, Guennadi Liakhovetski
> <g.liakhovetski@xxxxxx> wrote:
> > Hi Dan,
> >
> > On Wed, 10 Dec 2008, Dan Williams wrote:
> >
> >> On Wed, Dec 10, 2008 at 3:36 AM, Guennadi Liakhovetski
> >> <g.liakhovetski@xxxxxx> wrote:
> >> > From: Guennadi Liakhovetski <lg@xxxxxxx>
> >> >
> >> > Some users reuse DMA transaction descriptors multiple times and need an
> >> > explicit call to release them. An example of such a user is Video4Linux, which
> >> > has to be able to release descriptors on ioctl(VIDIOC_DQBUF).
> >> >
> >> > Signed-off-by: Guennadi Liakhovetski <lg@xxxxxxx>
> >>
> >>
> >> Hi Guennadi,
> >>
> >> Other dmaengine drivers have tasklets that scan the list of completed
> >> descriptors and free the "acked" ones.  This happens in the cleanup
> >> tasklet...
> >> /me looks
> >> ...hmm this driver does not have a cleanup routine?  Ideally support
> >> for ioctl(VIDIOC_DQBUF) could be achieved through this mechanism
> >> without needing to increase the size of dma_async_tx_descriptor (which
> >> has cache utilization impacts on other drivers).
> >
> > You mean there are drivers, that have struct dma_async_tx_descriptor on
> > stack?
> 
> huh?

How does the struct's size impact their cache utilisation then?

> > And you mean, that increasing the size
> > of the struct by one pointer and letting users explicitly free those
> > descriptors when they want is worse than introducing a tasklet that will
> > have to periodically scan the list of descriptors while other hot paths
> > will move elements to and from this list, look for acked elements, lock
> > the list and free those elements? Periodically, because although we have
> > an event when to free them - on ioctl - there is no API to trigger that
> > tasklet.
> 
> There are a few events that trigger this: completion interrupt,
> someone polls is_tx_complete, we run out of descriptors.

Sorry, actually, it's not VIDIOC_DQBUF where we have to free buffer(s), 
it's VIDIOC_STREAMOFF. And in a normal case there are no more completion 
interrupts, no allocation requests, and noone is interested in 
is_tx_complete. Normally you would just get a dma_release_channel after 
that. Or, of course, the application may decide to restart capture, then 
we get requests, completions, etc. again. So, yes, I could do this, it 
just seems a bit unnatural to me.

> > Or am I missing something? I can do this, it just sounds strange
> > to me.
> 
> You are missing that existing drivers need to do this anyway to to
> handle operation completion actions.  So, while they are at it they
> also free the descriptor which relieves client code from needing to
> track what it has in-flight versus completed, especially since clients
> may not get a handle to each descriptor a driver creates on its
> behalf.
> 
> What I am suggesting is that other dmaengine drivers would handle this
> implicitly after one of the above events...  can this happen in the
> ipu_idmac case as well?

Sure it can... Ok, so, the decision is - a garbage-collection tasklet? Ok, 
will do this.

Thanks
Guennadi
---
Guennadi Liakhovetski, Ph.D.
Freelance Open-Source Software Developer
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/