Re: ssb: Fix DMA-API compilation for non-PCI systems

Re: ssb: Fix DMA-API compilation for non-PCI systems

[David Miller]

I applied this since it's a build buster and we should
get it fixed quickly.

But as I look at this, the thing to really do here is get
rid of these stupid switch statements.  None of it is needed.

The layer that sets up the SSB devices should simply save the
generic device pointer for the DMA entity into some new
->dma_dev struct member.

Then you just use generic dma_*() interfaces unilaterally.  For PCI
devices it will "just work".

Re: ssb: Fix DMA-API compilation for non-PCI systems

[Michael Buesch]

On Monday 10 November 2008 22:52:31 David Miller wrote:
> 
> I applied this since it's a build buster and we should
> get it fixed quickly.
> 
> But as I look at this, the thing to really do here is get
> rid of these stupid switch statements.  None of it is needed.
> 
> The layer that sets up the SSB devices should simply save the
> generic device pointer for the DMA entity into some new
> ->dma_dev struct member.
>
> Then you just use generic dma_*() interfaces unilaterally.  For PCI
> devices it will "just work".

We tried that. It doesn't work. It breaks on several architectures
that do special things in the pci_dma... functions.
See the git history and mailinglists, please.

-- 
Greetings Michael.

Re: ssb: Fix DMA-API compilation for non-PCI systems

[David Miller]

From: Michael Buesch <mb@bu3sch.de>
Date: Mon, 10 Nov 2008 23:07:37 +0100

> On Monday 10 November 2008 22:52:31 David Miller wrote:
> > 
> > I applied this since it's a build buster and we should
> > get it fixed quickly.
> > 
> > But as I look at this, the thing to really do here is get
> > rid of these stupid switch statements.  None of it is needed.
> > 
> > The layer that sets up the SSB devices should simply save the
> > generic device pointer for the DMA entity into some new
> > ->dma_dev struct member.
> >
> > Then you just use generic dma_*() interfaces unilaterally.  For PCI
> > devices it will "just work".
> 
> We tried that. It doesn't work. It breaks on several architectures
> that do special things in the pci_dma... functions.
> See the git history and mailinglists, please.

Then those architectures need to fix their implementations,
simply.  They aren't implementing the generic DMA interfaces
properly.  Other subsystems do make use of the generic DMA
interfaces in this way.

Even something as simple as checking for &pci_bus_type in the
dma_*() routines (like powerpc did for quite some time) is
enough to fix the problem.

Re: ssb: Fix DMA-API compilation for non-PCI systems

[Michael Buesch]

On Monday 10 November 2008 23:53:43 David Miller wrote:
> Then those architectures need to fix their implementations,
> simply.  They aren't implementing the generic DMA interfaces
> properly.  Other subsystems do make use of the generic DMA
> interfaces in this way.
> 
> Even something as simple as checking for &pci_bus_type in the
> dma_*() routines (like powerpc did for quite some time) is
> enough to fix the problem.

Yeah, sure. The architectures are broken.
But the bugreports hit _my_ code. I got pretty tired of forwarding stuff
to the arch maintainers, so I implemented something that does work in
every situation. The bugreports stopped immediately (Except this tiny build
failure that does only hit under rare circumstances).

DMA is pretty much a mess, if you look at how ever arch does implement the API
in a slightly different way.
If somebody wants to fix this, I'm certainly OK with it. But for now we will
have to live with SSB implementing a wrapper around the ssb-host-device DMA APIs.

-- 
Greetings Michael.

Re: ssb: Fix DMA-API compilation for non-PCI systems

[Christoph Hellwig]

On Tue, Nov 11, 2008 at 12:32:00AM +0100, Michael Buesch wrote:
> On Monday 10 November 2008 23:53:43 David Miller wrote:
> > Then those architectures need to fix their implementations,
> > simply.  They aren't implementing the generic DMA interfaces
> > properly.  Other subsystems do make use of the generic DMA
> > interfaces in this way.
> > 
> > Even something as simple as checking for &pci_bus_type in the
> > dma_*() routines (like powerpc did for quite some time) is
> > enough to fix the problem.
> 
> Yeah, sure. The architectures are broken.
> But the bugreports hit _my_ code. I got pretty tired of forwarding stuff
> to the arch maintainers, so I implemented something that does work in
> every situation. The bugreports stopped immediately (Except this tiny build
> failure that does only hit under rare circumstances).
> 
> DMA is pretty much a mess, if you look at how ever arch does implement the API
> in a slightly different way.

We have tons of PCI drivers using the dma_ routines.  So please report
the broken architecture to get it fixes instead of whining.