On Mon, Jan 12, 2026 at 03:53:29PM +0000, Will Deacon wrote:

On Wed, Jan 07, 2026 at 08:36:46PM -0400, Jason Gunthorpe wrote:

quoted

On Wed, Jan 07, 2026 at 09:20:06PM +0000, Will Deacon wrote:

quoted

quoted

 drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.h   |  2 ++
 .../iommu/arm/arm-smmu-v3/arm-smmu-v3-test.c  | 18 ++++++++++---
 drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c   | 27 ++++++++++++++-----
 3 files changed, 37 insertions(+), 10 deletions(-)

Hmm. So this appears to ignore the safe bits entirely, whereas the
rationale for the change is that going from {MEV,EATS} disabled to
enabled is safe (which I agree with). 

The argument was it doesn't matter for either direction be it disabled
to enabled or vice versa, see my reply to Mustfa in the v4 posting:

https://lore.kernel.org/all/20251218180129.GA254720@nvidia.com/ (local)

It would be good to include some of that rationale in the comment and
commit message for patch 3, as at the moment it only talks about the
change in one direction.

Sure, I can help Nicolin with that.

I'm also still not convinced that this is generally safe, even if it
works within what Linux currently does. For example, if somebody tries
to disable S2S and enable ATS at the same time, couldn't you transiently
get an illegal STE?

I would argue that the driver will never concurrently support S2S and
ATS together for the same device, it doesn't make sense as far as I
can understand.

You are correct that there is an illegal STE issue here in this case.

However, keep in mind, if there is concurrent DMA while the driver is
trying to do such a thing there must be a STE error, and we should try
to make it be a non-valid STE error.

Still, it seems easy enough to improve, do not add EATS to the safe
bits if either the current or new STE has S2S set. That will force a
V=0 and avoid the illegal STE. Nicolin?

Thanks,
Jason