Re: [PATCH] KVM: arm64: Fix error checking for FFA_VERSION

On Sat, Nov 22, 2025 at 12:36 PM Marc Zyngier [off-list ref] wrote:

On Fri, 14 Nov 2025 11:11:53 +0000,
"=?utf-8?q?Kornel_Dul=C4=99ba?=" [off-list ref] wrote:

quoted

According to section 13.2 of the DEN0077 FF-A specification, when
firmware does not support the requested version, it should reply with
FFA_RET_NOT_SUPPORTED(-1). Table 13.6 specifies the type of the error
code as int32.
Currently, the error checking logic compares the unsigned long return
value it got from the SMC layer, against a "-1" literal. This fails due
to a type mismatch: the literal is extended to 64 bits, whereas the
register contains only 32 bits of ones(0x00000000ffffffff).
Consequently, hyp_ffa_init misinterprets the "-1" return value as an
invalid FF-A version. This prevents pKVM initialization on devices where
FF-A is not supported in firmware.

Is this statement accurate? I regularly boot KVM in protected mode in
environments that really cannot be suspected of implementing FF-A
(there is no EL3 to start with). And yet I don't see any failure of
the sort.

Please clarify the circumstances this is triggered.

I do have EL3 enabled, but the FF-A itself is not implemented.

Without this change kvm initialization fails with the following:

[    0.946776][    T1] kvm [1]: nv: 554 coarse grained trap handlers
[    0.952880][    T1] kvm [1]: nv: 669 fine grained trap handlers
[    0.958813][    T1] kvm [1]: IPA Size Limit: 44 bits
(...)
[    1.034089][    T1] kvm [1]: Failed to init hyp memory protection
[    1.041213][    T1] kvm [1]: error initializing Hyp mode: -95

I managed to narrow this down to the FFA version check by examining
all of the places in kvm initialization logic where -EOPNOTSUPP is
returned. Since printing anything in this part of the code is somewhat
problematic I replaced “return -EOPNOTSUPP” with “return res.s0” to
examine the problematic register value:

[1.041229][    T1] kvm [1]: error initializing Hyp mode: -1

Note that the return code itself is cast to int before being printed.
Then a colleague of mine recommended looking into the arm_ffa driver.
(“drivers/firmware/arm_ffa/driver.c”) There I found that in the
ffa_version_check function, the return value from the SMC call is cast
to s32 before being checked for errors.
I did the same in the kvm initialization logic, which is how this
patch was created. Furthermore I also examined the FF-A
specification(DEN0077), where the error code value type is specified
as int32.
With this change applied I can now see that kvm is up and running:

[    0.946839][    T1] kvm [1]: nv: 554 coarse grained trap handlers
[    0.952940][    T1] kvm [1]: nv: 669 fine grained trap handlers
[    0.958867][    T1] kvm [1]: IPA Size Limit: 44 bits
(...)
[    1.061717][    T1] kvm [1]: Protected hVHE mode initialized successfully

The /dev/kvm file is also there.