Re: [PATCH 1/3] crypto: marvell/cesa - replace dma_to_phys with dma_map_single

From: Boris Brezillon
Date: Fri Mar 18 2016 - 10:20:53 EST


On Fri, 18 Mar 2016 09:51:37 -0400
Sinan Kaya <okaya@xxxxxxxxxxxxxx> wrote:

> On 3/18/2016 7:25 AM, Robin Murphy wrote:
> > On 18/03/16 09:30, Boris Brezillon wrote:
> >> On Thu, 17 Mar 2016 23:50:20 +0000
> >> Russell King - ARM Linux <linux@xxxxxxxxxxxxxxxx> wrote:
> >>
> >>> On Thu, Mar 17, 2016 at 07:17:24PM -0400, okaya@xxxxxxxxxxxxxx wrote:
> >>>> What is the correct way? I don't want to write engine->sram_dma = sram
> >>>
> >>> Well, what the driver _is_ wanting to do is to go from a CPU physical
> >>> address to a device DMA address. phys_to_dma() looks like the correct
> >>> thing there to me, but I guess that's just an offset and doesn't take
> >>> account of any IOMMU that may be in the way.
> >>>
> >>> If you have an IOMMU, then the whole phys_to_dma() thing is a total
> >>> failure as it only does a linear translation, and there are no
> >>> interfaces in the kernel to take account of an IOMMU in the way. So,
> >>> it needs something designed for the job, implemented and discussed by
> >>> the normal methods of proposing a new cross-arch interface for drivers
> >>> to use.
> >>>
> >>> What I'm certain of, though, is that the change proposed in this patch
> >>> will break current users of this driver: virt_to_page() on an address
> >>> returned by ioremap() is completely undefined, and will result in
> >>> either a kernel oops, or if not poking at memory which isn't a struct
> >>> page, ultimately resulting in something that isn't SRAM being pointed
> >>> to by "engine->sram_dma".
> >>>
> >>
> >> Or we could just do
> >>
> >> engine->sram_dma = res->start;
> >>
> >> which is pretty much what the SRAM/genalloc code is doing already.
> >
> > As Russell points out this is yet another type of "set up a DMA master to access something other than kernel RAM" - there's already discussion in progress over how to handle this for dmaengine slaves[1], so gathering more use-cases might help distil exactly what the design of not-strictly-DMA-but-so-closely-coupled-it-can't-really-live-anywhere-else needs to be.
> >
> > Robin.
> >
> > [1]:http://lists.infradead.org/pipermail/linux-arm-kernel/2016-March/414422.html
> >
>
> Thanks for the link.
>
> dma_map_resource looks like to be the correct way of doing things. Just from
> the purist point of view, a driver is not supposed to know the physical address
> of a DMA address. That kills the intent of using DMA API. When programming descriptors,
> the DMA addresses should be programmed not physical addresses so that the same
> driver can be used in a system with IOMMU. The IOMMU DMA ops will remap the DMA
> address to a bus address that is not physical address. All of this operation needs
> to be isolated from the device driver.
>
>
> I don't know the architecture or the driver enough to write this. This is not ideally
> right but I can do this if Boris you are OK with this.
>
> engine->sram_dma = res->start;

I don't know.

How about waiting for the 'dma_{map,unmap}_resource' discussion to
settle down before removing phy_to_dma()/dma_to_phys() APIs (as
suggested by Robin and Russell)?


--
Boris Brezillon, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com