Re: [PATCH v2 1/3] arm64: mm: Fix rodata=full block mapping support for realm guests

[Catalin Marinas]

On Tue, Apr 07, 2026 at 02:06:10PM +0100, Ryan Roberts wrote:
> On 07/04/2026 11:52, Catalin Marinas wrote:
> > As the name implies, linear_map_requires_bbml2 tracks only this -
> > BBML2_NOABORT is required because the linear map uses large blocks.
> > Prior to your patches, that's only used as far as
> > linear_map_maybe_split_to_ptes() and if splitting took place, this
> > variable is no longer relevant (should be turned to false but since it's
> > not used, it doesn't matter).
> > 
> > With your patches, its use was extended to runtime and I think it
> > remains true even if linear_map_maybe_split_to_ptes() changed the block
> > mappings. Do we need this:
> 
> I'll admit it is ugly but it's not a bug; the system capabilitites are finalized
> by the time we call linear_map_maybe_split_to_ptes().
> 
> The "if (!linear_map_requires_bbml2 || is_kfence_address((void *)start))" check
> in split_kernel_leaf_mapping() would ideally be "if (!force_pte_mapping() ||
> is_kfence_address((void *)start))", but it is not safe to call
> force_pte_mapping() from a secondary cpu prior to finalizing the system caps.
> I'm reusing the flag that I already had available to work around that.

The confusing part is that the flag may be false incorrectly due to the
is_realm_world() evaluation. Nothing to do with this patch though and
the subject even mentions the rodata=full case. We should fix it
separately.

We could have set the flag to zero in linear_map_maybe_split_to_ptes(),
though split_kernel_leaf_mapping() already exits early due to the
!system_supports_bbml2_noabort() && system_capabilities_finalized(), so
not a correctness issue.

> But regardless, I think we are talking about the pre-existing
> is_real_world() bug, so I'm not personally planning to do anything further here
> unless you shout.

Not for this series. Steven or Suzuki should address the other problem
with is_realm_world().

-- 
Catalin