Re: [RFC PATCH 0/3] initalise ff-a after finalising pKVM

[Ben Horgan <ben.horgan@arm.com>]

Hi Levi,

On 5/5/26 10:54, Yeoreum Yun wrote:
> This patch is split out from the patchset [0] --
> fix FF-A call failure with pKVM when the FF-A driver is built-in,
> specifically the IMA-related part.
> 
> When pKVM is enabled, the FF-A driver must be initialised after pKVM.
> Otherwise, pKVM cannot negotiate the FF-A version or obtain the RX/TX
> buffer information, leading to failures in FF-A calls.
> 
> Currently, pKVM initialisation completes at device_initcall_sync,
> while ffa_init() runs at the device_initcall level.
> 
> So far, linker deployes kvm_arm_init() before ffa_init(), and SMCs can
> still be trapped even before finalise_pkvm() is invoked.
> As a result, this issue has not been observed.
> 
> However, relying on above stuff is fragile.
> Therefore, when pKVM is enabled, the FF-A infrastructure should be
> initialised only after pKVM initialisation has been fully finalised.
> 
> To achieve this, introduce an ffa_root_dev ("arm-ffa") and
> a corresponding driver to defer initialisation of the FF-A infrastructure
> until pKVM initialisation is complete, and to defer probing of all FF-A devices until then
> when pKVM is enabled.
> 
> This patch is based on v7.1-rc2
> 
> Question:
> 
> FF-A initialisation can occur at late_initcall. Because it may be deferred,
> some FF-A requests cannot be serviced at that stage.
> A typical example is the EFI runtime variable service using DIRECT_MSG_REQ.
> 
> Depending on the platform, the EFI runtime variable service runs with StandaloneMm
> and uses FF-A DIRECT_REQ. However, when pKVM is enabled, FF-A initialisation
> may be deferred to late_initcall. In this case, load_uefi_certs()
> can fail if it is invoked before the FF-A driver is initialised
> via deferred_probe_initcall().
> 
> Moving load_uefi_certs() to late_initcall_sync, as in the third patch,
> seems not to have any problem since late_initcall and
> late_initcall_sync are both of do_basic_setup() and it's before loading
> init process. However, it is still unclear whether
> it would be better to allow DIRECT_MSG_REQ in kvm_host_ffa_handler()

The spec doesn't allow this. Looking at DEN0077A 1.2 REL0:

Section 13.2.2 says:

"If they are compatible, it enables them to determine which Framework functionalities can be used. Hence, negotiation of
the version must happen before an invocation of any other FF-A ABI."

and a bit further down

"Once the caller invokes any FF-A ABI other than FFA_VERSION, the version negotiation phase is complete."

I would have thought that an SP would only go into the waiting state once the version negotiation is done.

Thanks,

Ben

> even before FF-A version negotiation since handler’s purpose seems to hook
> certain memory operations, and DIRECT_MSG_REQ has been available
> since FF-A specification v1.0.
> 
> Any feedback or alternative suggestions would be appreciated!
> 
> Link: https://lore.kernel.org/all/20260422162449.1814615-1-yeoreum.yun@arm.com/ [0]
> 
> Yeoreum Yun (3):
>   arm64: KVM: defer kvm_init() to finalise_pkvm() when pKVM is enabled
>   firmware: arm_ffa: initialise ff-a after finalising pKVM
>     initialisation
>   security: integrity: call load_uefi_certs() at late_initcall_sync
> 
>  arch/arm64/kvm/arm.c                          |   8 +-
>  arch/arm64/kvm/pkvm.c                         |  15 ++-
>  drivers/firmware/arm_ffa/bus.c                | 125 +++++++++++++++++-
>  drivers/firmware/arm_ffa/common.h             |  13 +-
>  drivers/firmware/arm_ffa/driver.c             |  21 ++-
>  drivers/firmware/arm_ffa/smccc.c              |   2 +-
>  security/integrity/platform_certs/load_uefi.c |   2 +-
>  7 files changed, 166 insertions(+), 20 deletions(-)
> 
> 
> base-commit: 7fd2df204f342fc17d1a0bfcd474b24232fb0f32