[PATCH] iommu/arm-smmu-v3: Maintain valid access attributes for non-coherent SMMU

[Jason Gunthorpe]

On Mon, Jan 05, 2026 at 04:02:34PM +0000, Robin Murphy wrote:

> > It is reasonable that Linux will set the attributes properly based on
> > what it is doing. Setting the wrong attributes and expecting the HW to
> > ignore them seems like a hacky direction.
> 
> Oh, I'm not saying that we *shouldn't* set our attributes more exactly -
> this would still be needed for doing things the "right" way too - I just
> want to be very clear on the reasons why. 

At least I know of HW where the SMMU fetches covered by COHACC:

 * Translation table walks.
 * Fetches of L1STD, STE, L1CD and CD.
 * Command queue, Event queue and PRI queue access.
 * GERROR, CMD_SYNC, Event queue and PRI queue MSIs, if supported.

Have a mixture of coherency support. The SOC has multiple fabrics, one
non-coherent one specifically for isochronous traffic.  In HW some
SMMU sub-units (like the table walk) been wired to the isochronous
fabric, while others are using the normal coherent fabric.

So when it comes to this statement:

 If either the SMMU or system cannot *fully* support IO-coherent
 access to SMMU structures/queues/translations, this reads as 0.

The HW is "partially" IO-coherent, so COHACC is 0.

It has been lucky that so far the incorrect attributes haven't caused a
problem, but the next spin of this HW may have issue here. I'd like to
see it fixed.

> bug per se, and although it's indeed not 100% robust, the cases where it
> doesn't hold are more often than not for the wrong reason. Therefore I would
> say doing this purely for the sake of working around bad firmware - and
> especially errata - is just as hacky if not more so.

Yeah, maybe, I am also curious what is motivating Dawei to do this work..

> the DMA API aspect I mean is that in
> general we need some sort of DMA_ATTR_NO_SNOOP when mapping/allocating such
> isochronous buffers/pagetables etc., to make the DMA layer still do the
> cache maintenance/non-cacheable remaps despite dev_is_dma_coherent() being
> true.

I have a feeling this missing support underlies some of the reasons
why FW might lie and set COHACC=0 as it resolves "how does the GPU
driver cache flush things" with no effort..

Jason