I've pushed changes to umr (master) and our linux (amd-staging-4.12
should be on
soonish) trees that allow in limited circumstances to VM decode through
IOMMU which has been tested for both iommu active and inactive.
umr previously could not fully resolve addresses when IOMMU is active
since the GPU is programmed with dma mapped addresses. As such reading
IBs or shaders or VM decoding in general would fail.
The change to the kernel adds a tracepoint which logs all dma map/unmap
operations (via the TTM) which umr can read and use to access physical
memory behind the mappings.
To enable this you need to enable two trace points:
echo 1 >
echo 1 >
You need to enable the traces before running your test case/application.
If the trace file can't hold enough for your application you can tail it
to ~/trace which umr will read preferentially (make sure you delete this
file when done or you will be reading old mappings).
cat /sys/kernel/debug/tracing/trace_pipe > ~/trace
umr can process the traces fairly quickly (at about 1,500,000+ lines/sec
on my Carrizo...) so theoretically heavy workloads should be manageable.
The traces contain PCI information which umr uses to match mappings so
multiple GPU systems should work fine as well. Though testing would be
amd-gfx mailing list