Re: [PATCH v2 1/5] iommu/amd: Simplify build_inv_address()

[Jason Gunthorpe <jgg@nvidia.com>]

On Tue, Apr 07, 2026 at 09:15:40AM +0800, Wei Wang wrote:
> On 4/6/26 9:30 PM, Jason Gunthorpe wrote:
> > On Mon, Apr 06, 2026 at 09:18:19PM +0800, Wei Wang wrote:
> > 
> > > > -	msb_diff = fls64(end ^ address) - 1;
> > > > +	if (unlikely(sz_lg2 > 52))
> > > > +		return (CMD_INV_IOMMU_ALL_PAGES_ADDRESS & PAGE_MASK) |
> > > > +		       CMD_INV_IOMMU_PAGES_SIZE_MASK;
> > 
> > I was looking at Ankit's series and thinking this is slightly better as:
> > 
> > +	if (unlikely(sz_lg2 > min(52, BITS_PER_LONG - 1)))
> > 
> > So a ULONG_MAX gather will trigger full invalidation on 32 bit as
> > well, though it shouldn't be necessary it seems a tish cleaner and
> > matches how vt-d ended up
> > 
> 
> I see the value in matching how VT‑d handles this. I’m slightly concerned
> that adding the BITS_PER_LONG check might imply that AMD‑IOMMU supports
> 32‑bit kernels.
> 
> Unlike VT‑d, the AMD IOMMU driver doesn’t build on 32‑bit at all. From
> Kconfig:
> 
> config AMD_IOMMU
>         depends on X86_64 && PCI && ACPI && HAVE_CMPXCHG_DOUBLE
> 
> config INTEL_IOMMU
>         depends on PCI_MSI && ACPI && X86
> 
> Given that the 32‑bit case is unreachable for AMD‑IOMMU, keeping the
> existing check might actually be clearer?

Ah, I didn't remember that, it makes sense. Let's leave it then

Jason