Re: [PATCH v2 4/5] iommu: Use dma_addr_t for IOVA arguments

[Bjorn Helgaas]

On Thu, May 08, 2014 at 12:44:12PM +0200, Arnd Bergmann wrote:
> On Wednesday 07 May 2014 18:18:14 Bjorn Helgaas wrote:
> > On Wed, May 07, 2014 at 09:58:58AM +0200, Arnd Bergmann wrote:
> > > On Tuesday 06 May 2014 16:48:40 Bjorn Helgaas wrote:
> > > > Convert the "iova" arguments of iommu_map(), iommu_unmap(), etc., from
> > > > "unsigned long" to dma_addr_t.
> > > > 
> > > > bb5547acfcd8 ("iommu/fsl: Make iova dma_addr_t in the iommu_iova_to_phys
> > > > API") did this for iommu_iova_to_phys(), but didn't fix the rest of the
> > > > IOMMU API.
> > > 
> > > This patch looks 100% correct, but I'm not convinced it's a good idea:
> > > On 32-bit platforms (i.e. most of the ones you change), doing 64-bit
> > > arithmetic has a noticeable overhead. I am not aware of an IOMMU that
> > > actually uses 64-bit DMA addresses on its slave side, usually they
> > > are used to translate addresses from 32-bit masters into 64-bit
> > > memory addresses, so using 'unsigned long' seems better from a practical
> > > point of view as opposed to the strict type correctness.
> > 
> > The current x86 IOMMUs support DMA addresses larger than 32 bits, but
> > obviously those platforms usually run 64-bit kernels so "unsigned
> > long" is already 64 bits.
> > 
> > I guess you're thinking about cases where "unsigned long" is 32 bits,
> > the IOMMU only supports 32 bit DMA addresses, and dma_addr_t is 64
> > bits.  If the IOMMU only supports 32-bit DMA addresses, is there any
> > value in having a 64-bit dma_addr_t?
> 
> Two cases:
> 
> a) You can have a system with some DMA masters that are 64-bit capable,
>    and other masters that can only do 32-bit DMA and for this reason
>    use an IOMMU. I expect to see more of these in the future.
> 
> b) We build kernels that run on a multitude of ARM32 platforms these
>    days. We have some that require a 32-bit dma_addr_t and some that
>    don't (because they always use swiotlb or an IOMMU).

Those both make sense (I assume you meant some ARM32 platforms require
a *64-bit* dma_addr_t, i.e., you might want to build an ILP32 kernel
with a 64-bit dma_addr_t).

I doubt there would be a noticeable performance effect since these are
relatively low-frequency interfaces (map, unmap, report_fault), but I
don't have any numbers, so I'll drop this for now.

Bjorn
--
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/