Re: Tearing down DMA transfer setup after DMA client has finished

From: Måns Rullgård
Date: Fri Nov 25 2016 - 10:21:38 EST

Russell King - ARM Linux <linux@xxxxxxxxxxxxxxx> writes:

> On Fri, Nov 25, 2016 at 02:40:21PM +0000, Måns Rullgård wrote:
>> Russell King - ARM Linux <linux@xxxxxxxxxxxxxxx> writes:
>> > On Fri, Nov 25, 2016 at 02:03:20PM +0000, Måns Rullgård wrote:
>> >> Russell King - ARM Linux <linux@xxxxxxxxxxxxxxx> writes:
>> >>
>> >> > On Fri, Nov 25, 2016 at 01:50:35PM +0000, Måns Rullgård wrote:
>> >> >> Russell King - ARM Linux <linux@xxxxxxxxxxxxxxx> writes:
>> >> >> > It would be unfair to augment the API and add the burden on everyone
>> >> >> > for the new API when 99.999% of the world doesn't require it.
>> >> >>
>> >> >> I don't think making this particular dma driver wait for the descriptor
>> >> >> callback to return before reusing a channel quite amounts to a horrid
>> >> >> hack. It certainly wouldn't burden anyone other than the poor drivers
>> >> >> for devices connected to it, all of which are specific to Sigma AFAIK.
>> >> >
>> >> > Except when you stop to think that delaying in a tasklet is exactly
>> >> > the same as randomly delaying in an interrupt handler - the tasklet
>> >> > runs on the return path back to the parent context of an interrupt
>> >> > handler. Even if you sleep in the tasklet, you're sleeping on behalf
>> >> > of the currently executing thread - if it's a RT thread, you effectively
>> >> > destroy the RT-ness of the thread. Let's hope no one cares about RT
>> >> > performance on that hardware...
>> >>
>> >> That's why I suggested to do this only if the needed delay is known to
>> >> be no more than a few bus cycles. The completion callback is currently
>> >> the only post-transfer interaction we have between the dma and device
>> >> drivers. To handle an arbitrarily long delay, some new interface will
>> >> be required.
>> >
>> > And now we're back at the point I made a few emails ago about undue
>> > burden which is just about quoted above...
>> So what do you suggest? Stick our heads in the sand and pretend
>> everything is perfect?
> Look, if you're going to be arsey, don't be surprised if I start getting
> the urge to repeat previous comments.
> Let's try and keep this on a technical basis for once, rather than
> decending into insults.

You're the one who constantly insults people. I'd be happy for you to

> So, wind back to my original email where I started talking about PL08x
> already doing something along these lines. Before a DMA user can make
> use of a DMA channel, it has to be requested. Once a DMA user has
> finished, it can free up the channel.
> What this means is that there's already a solution here - but it depends
> how many DMA channels and how many active DMA users there are. It's
> entirely possible to set the mapping up when a DMA user requests a
> DMA channel, leave it setup, and only tear it down when the channel
> is eventually freed.
> At that point, there's no need to spin-wait or sleep to delay the
> tear-down of the channel - and I'd suggest that approach _until_
> such time that there are more users than there are DMA channels. This
> has minimal overhead, it doesn't screw up RT threads (which include
> IRQ threads), and it doesn't spread the maintanence burden across
> drivers with a new custom API just for one SoC.

I never suggested a custom API for one SoC.

> If (or when) the number of active users exceeds the number of hardware
> DMA channels, then there's a decision to be made:
> 1) either limit the number of peripherals that we support DMA on for
> the SoC.

I don't think people would like being forced to choose between, say,
SATA and NAND flash.

> 2) add the delay or API as necessary and switch to dynamic channel
> allocation to incoming requests.

A fixed delay doesn't seem right. Since we don't know the exact amount
required, we'll need to make a guess and make it conservative enough
that it never ends up being too short. This will most likely end up
delaying things far more than is actually necessary.

The reality of the situation is that the current dmaengine api doesn't
adequately cover all real hardware situations. You seem to be of the
opinion that fixing this is an "undue burden."

> Until that point is reached, there's no point inventing new APIs for
> something that isn't actually a problem yet.

We're already at that point. The hardware has many more devices than
physical channels.

Måns Rullgård