On Fri, Nov 09, 2012 at 09:54:01AM +0000, Will Deacon wrote:
> On Fri, Nov 09, 2012 at 04:01:52AM +0000, Shiraz Hashim wrote:
> > From: Catalin Marinas <[email protected]>
> >
> > Clearing bit 22 in the PL310 Auxiliary Control register (shared
> > attribute override enable) has the side effect of transforming Normal
> > Shared Non-cacheable reads into Cacheable no-allocate reads.
> >
> > Coherent DMA buffers in Linux always have a Cacheable alias via the
> > kernel linear mapping and the processor can speculatively load cache
> > lines into the PL310 controller. With bit 22 cleared, Non-cacheable
> > reads would unexpectedly hit such cache lines leading to buffer
> > corruption.
>
> Is this still the case with recent kernels? I thought the dma-mapping/cma
> work avoided the cacheable alias, but perhaps I'm mistaken.
I haven't used CMA but DMA mappings are still normal memory
non-cacheable.
--
regards
Shiraz
On Mon, Nov 12, 2012 at 06:45:47AM +0000, Shiraz Hashim wrote:
> On Fri, Nov 09, 2012 at 09:54:01AM +0000, Will Deacon wrote:
> > On Fri, Nov 09, 2012 at 04:01:52AM +0000, Shiraz Hashim wrote:
> > > From: Catalin Marinas <[email protected]>
> > >
> > > Clearing bit 22 in the PL310 Auxiliary Control register (shared
> > > attribute override enable) has the side effect of transforming Normal
> > > Shared Non-cacheable reads into Cacheable no-allocate reads.
> > >
> > > Coherent DMA buffers in Linux always have a Cacheable alias via the
> > > kernel linear mapping and the processor can speculatively load cache
> > > lines into the PL310 controller. With bit 22 cleared, Non-cacheable
> > > reads would unexpectedly hit such cache lines leading to buffer
> > > corruption.
> >
> > Is this still the case with recent kernels? I thought the dma-mapping/cma
> > work avoided the cacheable alias, but perhaps I'm mistaken.
>
> I haven't used CMA but DMA mappings are still normal memory
> non-cacheable.
Ok, so trawling through the list reveals we only have this issue for normal
DMA mappings and not with CMA:
http://lists.infradead.org/pipermail/linux-arm-kernel/2012-October/124276.html
I wonder whether we shouldn't just fix that, rather than work around it with
a PL310-specific hack?
Will
Hi Catalin,
On Mon, Nov 12, 2012 at 10:56:41AM +0000, Will Deacon wrote:
> On Mon, Nov 12, 2012 at 06:45:47AM +0000, Shiraz Hashim wrote:
> > On Fri, Nov 09, 2012 at 09:54:01AM +0000, Will Deacon wrote:
> > > On Fri, Nov 09, 2012 at 04:01:52AM +0000, Shiraz Hashim wrote:
> > > > From: Catalin Marinas <[email protected]>
> > > >
> > > > Clearing bit 22 in the PL310 Auxiliary Control register (shared
> > > > attribute override enable) has the side effect of transforming Normal
> > > > Shared Non-cacheable reads into Cacheable no-allocate reads.
> > > >
> > > > Coherent DMA buffers in Linux always have a Cacheable alias via the
> > > > kernel linear mapping and the processor can speculatively load cache
> > > > lines into the PL310 controller. With bit 22 cleared, Non-cacheable
> > > > reads would unexpectedly hit such cache lines leading to buffer
> > > > corruption.
> > >
> > > Is this still the case with recent kernels? I thought the dma-mapping/cma
> > > work avoided the cacheable alias, but perhaps I'm mistaken.
> >
> > I haven't used CMA but DMA mappings are still normal memory
> > non-cacheable.
>
> Ok, so trawling through the list reveals we only have this issue for normal
> DMA mappings and not with CMA:
>
> http://lists.infradead.org/pipermail/linux-arm-kernel/2012-October/124276.html
>
> I wonder whether we shouldn't just fix that, rather than work around it with
> a PL310-specific hack?
What do you say?
--
regards
Shiraz