Re: [PATCH] KVM: Fix kvm_vcpu_map[_readonly]() function prototypes

[Sean Christopherson]

On Tue, Mar 31, 2026, Peter Fang wrote:
> On Tue, Mar 31, 2026 at 02:22:47AM +0000, Yosry Ahmed wrote:
> > 
> > Most callers are converting a GPA to a GFN, I wonder if we should make

Not most, all.  The two outliers just do "gpa >> PAGE_SHIFT" instead of
using gpa_to_gfn().

> > the function take in a GPA instead? But then we'll need to the GPA not
> > being aligned to a page boundary (either do gpa_to_gfn() in
> > __kvm_vcpu_map() or fail if it's not aligned).

Just do gpa_to_pfn().  If someone gets confused, we can add a comment explaining
that kvm_vcpu_map() maps the entire page containing the gpa, but that should really
go without saying...

> Thanks for the feedback!
> 
> Mapping guest memory into the host feels more like a GFN-based operation
> to me. struct kvm_host_map is also designed around GFNs/PFNs so I think
> using gfn_t in the function prototypes seems more natural. The caller
> can handle the offset-in-page cases without creating a lot of complexity
> in the APIs. But I'm happy to rework this if there's a desire to make
> them more GPA-friendly.

I vote to rework the APIs (after first fixing the prototypes) to take a GPA.
I agree that mapping a page at a given gfn is conceptually more natural, but
as Yosry points out, requiring literally every caller to convert to a gfn doesn't
make a whole lot of sense from a code maintenance perspective.