On Wed, Sep 03, 2025, Atish Kumar Patra wrote:

On Fri, Aug 29, 2025 at 1:47 PM Sean Christopherson [off-list ref] wrote:

quoted

On Fri, Aug 29, 2025, Atish Patra wrote:

quoted

+static inline bool kvm_is_gpa_in_writable_memslot(struct kvm *kvm, gpa_t gpa)
+{
+     bool writable;
+     unsigned long hva = gfn_to_hva_prot(kvm, gpa_to_gfn(gpa), &writable);
+
+     return !kvm_is_error_hva(hva) && writable;

I don't hate this API, but I don't love it either.  Because knowing that the
_memslot_ is writable doesn't mean all that much.  E.g. in this usage:

        hva = kvm_vcpu_gfn_to_hva_prot(vcpu, shmem >> PAGE_SHIFT, &writable);
        if (kvm_is_error_hva(hva) || !writable)
                return SBI_ERR_INVALID_ADDRESS;

        ret = kvm_vcpu_write_guest(vcpu, shmem, &zero_sta, sizeof(zero_sta));
        if (ret)
                return SBI_ERR_FAILURE;

the error code returned to the guest will be different if the memslot is read-only
versus if the VMA is read-only (or not even mapped!).  Unless every read-only
memslot is explicitly communicated as such to the guest, I don't see how the guest
can *know* that a memslot is read-only, so returning INVALID_ADDRESS in that case
but not when the underlying VMA isn't writable seems odd.

It's also entirely possible the memslot could be replaced with a read-only memslot
after the check, or vice versa, i.e. become writable after being rejected.  Is it
*really* a problem to return FAILURE if the guest attempts to setup steal-time in
a read-only memslot?  I.e. why not do this and call it good?

Reposting the response as gmail converted my previous response as
html. Sorry for the spam.

From a functionality pov, that should be fine. However, we have
explicit error conditions for read only memory defined in the SBI STA
specification[1].
Technically, we will violate the spec if we return FAILURE instead of
INVALID_ADDRESS for read only memslot.

But KVM is already violating the spec, as kvm_vcpu_write_guest() redoes the
memslot lookup and so could encounter a read-only memslot (if it races with
a memslot update), and because the underlying memory could be read-only even if
the memslot is writable.

Why not simply return SBI_ERR_INVALID_ADDRESS on kvm_vcpu_write_guest() failure?
The only downside of that is KVM will also return SBI_ERR_INVALID_ADDRESS if the
userspace mapping is completely missing, but AFAICT that doesn't seem to be an
outright spec violation.