I got an IOMMU IO page fault. What to do now?

[Alex Deucher]

On Wed, Oct 27, 2021 at 1:24 PM Alex Deucher <alexdeucher at gmail.com> wrote:
>
> On Wed, Oct 27, 2021 at 1:20 PM Robin Murphy <robin.murphy at arm.com> wrote:
> >
> > On 27/10/2021 5:45 pm, Paul Menzel wrote:
> > > Dear Robin,
> > >
> > >
> > > On 25.10.21 18:01, Robin Murphy wrote:
> > >> On 2021-10-25 12:23, Christian König wrote:
> > >
> > >>> not sure how the IOMMU gives out addresses, but the printed ones look
> > >>> suspicious to me. Something like we are using an invalid address like
> > >>> -1 or similar.
> > >>
> > >> FWIW those look like believable DMA addresses to me, assuming that the
> > >> DMA mapping APIs are being backed iommu_dma_ops and the device has a
> > >> 40-bit DMA mask, since the IOVA allocator works top-down.
> > >>
> > >> Likely causes are either a race where the dma_unmap_*() call happens
> > >> before the hardware has really stopped accessing the relevant
> > >> addresses, or the device's DMA mask has been set larger than it should
> > >> be, and thus the upper bits have been truncated in the round-trip
> > >> through the hardware.
> > >>
> > >> Given the addresses involved, my suspicions would initially lean
> > >> towards the latter case - the faults are in the very topmost pages
> > >> which imply they're the first things mapped in that range. The other
> > >> contributing factor being the trick that the IOVA allocator plays for
> > >> PCI devices, where it tries to prefer 32-bit addresses. Thus you're
> > >> only likely to see this happen once you already have ~3.5-4GB of live
> > >> DMA-mapped memory to exhaust the 32-bit IOVA space (minus some
> > >> reserved areas) and start allocating from the full DMA mask. You
> > >> should be able to check that with a 5.13 or newer kernel by booting
> > >> with "iommu.forcedac=1" and seeing if it breaks immediately
> > >> (unfortunately with an older kernel you'd have to manually hack
> > >> iommu_dma_alloc_iova() to the same effect).
> > >
> > > I booted Linux 5.15-rc7 with `iommu.forcedac=1` and the system booted,
> > > and I could log in remotely over SSH. Please find the Linux kernel
> > > messages attached. (The system logs say lightdm failed to start, but it
> > > might be some other issue due to a change in the operating system.)
> >
> > OK, that looks like it's made the GPU blow up straight away, which is
> > what I was hoping for (and also appears to reveal another bug where it's
> > not handling probe failure very well - possibly trying to remove a
> > non-existent audio device?). Lightdm presumably fails to start because
> > it doesn't find any display devices, since amdgpu failed to probe.
> >
> > If you can boot the same kernel without "iommu.forcedac" and get a
> > successful probe and working display, that will imply that it is
> > managing to work OK with 32-bit DMA addresses, at which point I'd have
> > to leave it to Christian and Alex to figure out exactly where DMA
> > addresses are getting mangled. The only thing that stands out to me is
> > the reference to "gfx_v6_0", which makes me wonder whether it's related
> > to gmc_v6_0_sw_init() where a 44-bit DMA mask gets set. If so, that
> > would suggest that either this particular model of GPU is more limited
> > than expected, or that SoC only has 40 bits of address wired up between
> > the PCI host bridge and the IOMMU.
>
> That device only has a 40 bit DMA mask.  It looks like the code is wrong there.

The attached patch should fix it.

Alex
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-drm-amdgpu-gmc6-fix-DMA-mask.patch
Type: text/x-patch
Size: 974 bytes
Desc: not available
URL: <https://lists.freedesktop.org/archives/amd-gfx/attachments/20211027/f06837fc/attachment.bin>