On Mon, Jan 19, 2015 at 12:50:52PM +0000, Will Deacon wrote:
> On Mon, Jan 19, 2015 at 12:43:06PM +0000, Thierry Reding wrote:
> > On Sun, Jan 18, 2015 at 01:18:51PM +0200, Laurent Pinchart wrote:
> > > On Sunday 18 January 2015 15:54:34 Alexandre Courbot wrote:
> > > > On 01/16/2015 08:18 AM, Laurent Pinchart wrote:
> > [...]
> > > > > The second way is to implement a mechanism to let drivers signal that they
> > > > > want to handle DMA mappings themselves. As the mappings need in the
> > > > > general case to be created before the probe function  is called
> > > > 
> > > > Sorry for being ignorant here, but why is that?
> > > 
> > > Because a driver can call the DMA mapping API in its probe function, to 
> > > allocate DMA coherent memory for instance. We need to ensure that the DMA 
> > > mapping IOMMU has set up the required IOMMU ops by that time. As explained 
> > > above I don't like the idea of sprinkling explicit calls to initialize IOMMU 
> > > support in the vast majority of drivers, especially when they shouldn't be 
> > > IOMMU-aware, so we then need to initialize everything that is needed before 
> > > the first call to the DMA mapping API.
> > 
> > The original patch that Hiroshi posted based on my comments was to have
> > the driver core call iommu_attach(), which would then set up everything
> > needed right before calling into the driver's ->probe(). This works
> > quite nicely because by definition the driver can't allocate any DMA
> > before ->probe(). And, like you said, this allows deferred probe to be
> > used.
> > 
> > To me it's so obviously the right solution that I remain flabbergasted
> > with how much resistance it's received (or how much it's being ignored).
> 
> Have you considered reposting the patches based on what we currently have
> (which has the advantage of identifying a specific IOMMU instance)?

No, I hadn't. Initially my patches included a solution for identifying
individual IOMMU instances, too. There was a small registry with a list
of struct iommus. That was supposed to get used to store per-instance
data (by being embedded in a driver-specific structure). I'd need to
look in more detail how that could be done with the infrastructure that
your patchset creates. I'm somewhat burried below other tasks right now
so don't expect to have any time to look into this anytime before -rc6
or -rc7 at the earliest.

Thierry