Re: [PATCH 0/10] DMA-API debugging facility

From: FUJITA Tomonori
Date: Sat Nov 22 2008 - 05:17:28 EST


On Sat, 22 Nov 2008 10:33:18 +0100
Joerg Roedel <joro@xxxxxxxxxx> wrote:

> On Sat, Nov 22, 2008 at 12:27:43PM +0900, FUJITA Tomonori wrote:
> > On Fri, 21 Nov 2008 18:45:51 +0100
> > Ingo Molnar <mingo@xxxxxxx> wrote:
> >
> > >
> > > * Joerg Roedel <joerg.roedel@xxxxxxx> wrote:
> > >
> > > > On Fri, Nov 21, 2008 at 05:24:29PM +0000, David Woodhouse wrote:
> > > > > On Fri, 2008-11-21 at 18:20 +0100, Joerg Roedel wrote:
> > > > > > Ok, I will move the generic bits to lib/ and include/linux and let
> > > > > > architectures decide if they want to use it.
> > > > >
> > > > > Once you've done that, I'll try to hook it up on PowerPC to make
> > > > > sure it works there.
> > > >
> > > > Ok, cool. Thanks
> > >
> > > i'll give it a whirl on x86 once the allocation bug is resolved. x86
> > > testing will be the most interesting in practice, because most drivers
> > > there are developed with no dynamic DMA in mind. (many of the x86
> > > drivers were developed before IOMMUs were supported in Linux)
> >
> > Yeah, one of the problems due to this is that some drivers wrongly
> > assume that the dma mapping operations never fail (they do DMA with a
> > wrong address). With VT-d and AMD IOMMU, the dma mapping operations
> > fail just because of OOM.
>
> At least AMD IOMMU code will not fail because of memory shortage. All
> necessary data structures, including the pagetables, are preallocated.
> The only place were it might fail is dma_alloc_coherent. But that is not
> specific to that driver.

Oops, you are right. Somehow I wrongly thought that AMD IOMMU
allocates pte on the fly like VT-d.


> > Some time ago, I hooked the fault injection mechanism to the dma
> > mapping operations (I linked struct fault_attr to struct pci_dev so
> > you can make dma_map_single/sg fail with a particular pci device). It
> > might interest some people:
> >
> > git://git.kernel.org/pub/scm/linux/kernel/git/tomo/linux-2.6-misc.git dmafault
>
> This would also be helpful to find bugs in drivers together with this
> code. Do you plan to submit it?

Yeah, with your debug mechanism, this might be more helpful because
the debug mechanism can catch drivers that don't handle dma mapping
failure properly. I'll submit this after you finish the debug
mechanism.

Thanks,
--
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/