Re: [PATCH] arm64: swiotlb: Keep the default size for protected guests

[Will Deacon]

On Mon, Aug 10, 2026 at 04:44:51PM +0100, Catalin Marinas wrote:
> On Mon, Aug 10, 2026 at 03:14:47PM +0100, Will Deacon wrote:
> > On Mon, Aug 10, 2026 at 10:08:18AM -0300, Jason Gunthorpe wrote:
> > > On Mon, Aug 10, 2026 at 12:46:51PM +0100, Will Deacon wrote:
> > > > On Mon, Aug 10, 2026 at 01:37:11PM +0200, Marek Szyprowski wrote:
> > > > > On 10.08.2026 12:20, Will Deacon wrote:
> > > > > > On Mon, Aug 10, 2026 at 02:59:42PM +0530, Aneesh Kumar K.V wrote:
> > > > > >> Marek Szyprowski <m.szyprowski@xxxxxxxxxxx> writes:
> > > > > >>
> > > > > >>> On 07.08.2026 20:20, Jason Gunthorpe wrote:
> > > > > >>>> On Fri, Aug 07, 2026 at 06:13:49PM +0000, Mostafa Saleh wrote:
> > > > > >>>>> But the whole thing is best effort anyway, the kernel picks
> > > > > >>>>> IO_TLB_DEFAULT_SIZE which does not depend on the system topology or
> > > > > >>>>> how many devices or how much DMA they do.
> > > > > >>>>> SWIOTLB memory is wasted if unused so we should be careful around
> > > > > >>>>> that as it would be the other way around and users would have to
> > > > > >>>>> decrease it manually.
> > > > > >>>> Yeah, it is why the arch code shouldn't really be sizing it directly,
> > > > > >>>> it should be done in common code and, yes, we are probably going to
> > > > > >>>> have to do something alot smarter to have the common code better
> > > > > >>>> auto-tune this for the CC case..
> > > > > >>> What about the $subject patch? I assume that it is still needed to
> > > > > >>>
> > > > > >>> restore the behavior that was altered by the "[PATCH v8 00/23]
> > > > > >>>
> > > > > >>> dma-mapping: Track shared DMA state through direct, pool and swiotlb
> > > > > >>>
> > > > > >>> paths?" patchset?
> > > > > >>>
> > > > > >> I would request that we pick this patch to fix the regression described
> > > > > >> in https://lore.kernel.org/all/yq5azeyxyfol.fsf@xxxxxxxxxx/.
> > > > > > I really don't think we need it. CCA hardware isn't exactly widespread
> > > > > > and the KVM host side patches don't appear close to being merged.
> > > > > 
> > > > > Does this mean that the branch for-next/coco [1] won't go to v7.3-rc1?
> > > 
> > > Let's not make progress on guest support contingent on KVM CCA host
> > > side patches please. I expect the CSPs will have VM instance types
> > > available based on CCA within quarters, and Linux as a Guest should
> > > work in those environments regardless of what KVM is doing. I don't
> > > really expect full KVM support for years, frankly, the patchset is
> > > massive. Even Intel and AMD don't have full KVM support yet.
> > 
> > To be clear: the only part I'm pushing back on is the realm-specific hack
> > to size the SWIOTLB area in arch/arm64/. Even with that hack applied, I
> > don't believe a one-size-fits-all value is going to work for everybody,
> > so it was really great to see that special-case removed in:
> > 
> > https://lore.kernel.org/all/20260717180442.110954-18-aneesh.kumar@xxxxxxxxxx/
> 
> We had Mostafa's patch here:
> 
> https://lore.kernel.org/all/20260603110522.3331819-4-smostafa@xxxxxxxxxx/
> 
> adding the same swiotlb buffer allocation for pKVM guests and realms.

Yes, and I think that's bad for pKVM guests which is why I'm happy to
see that part removed.

> Aneesh's cleanup of SWIOTLB_FORCE inadvertently removed the 'else'
> clause and limited_addressing is now always checked, meaning that we get
> size limiting irrespective of whether we still need a default swiotlb.

Inadvertently, perhaps, but it's got Tested-bys from three different
companies on it and I think it's a good change. I don't think we should
be special-casing these environments based on some random collection of
virtio devices in a testing setup.

> Normally I would consider this a regression on top of mainline and
> Mostafa's patches rather than a cleanup. However, if the default size of
> the bounce buffer was never sufficient for realms, as you said, we may
> need something better here anyway, so not worth fixing.
> 
> I can see x86 in mem_encrypt_setup_arch() adjusting the bounce buffer to
> 6% of the guest memory. We could make up similar logic for arm64 or we
> could just rely on CONFIG_SWIOTLB_DYNAMIC (default off currently for
> arm64 I think).

As Jason said, this should be done outside the arch code.

Will