Re: [PATCH] ArmPkg/GicV3Dxe: Don't signal EOI on arbitrary interrupts

["Leif Lindholm" <leif@nuviainc.com>]

On Tue, Aug 24, 2021 at 17:31:32 +0200, Ard Biesheuvel wrote:
> Currently, at ExitBootServices() time, the GICv3 driver signals
> End-Of-Interrupt (EOI) on all interrupt lines that are supported by the
> interrupt controller. This appears to have been carried over from the
> GICv2 version, but has been turned into something that violates the GIC
> spec, and may trigger SError exceptions on some implementations.
> 
> Marc puts it as follows:
> 
>   The GIC interrupt state machine is pretty strict. An interrupt can
>   only be deactivated (with or without prior priority drop) if it has
>   been acknowledged first. In GIC speak, this means that only the
>   following sequences are valid:
> 
>   With EOImode==0:
> 	  x = ICC_IAR{0,1}_EL1;
> 	  ICC_EOIR{0,1}_EL1 = x;
> 
>   With EOImode==1:
> 	  x = ICC_IAR{0,1}_EL1;
> 	  ICC_EOIR{0,1}_EL1 = x;
> 	  ICC_DIR_EL1 = x;
> 
>   Any write to ICC_EOIR{0,1}_EL1 that isn't the direct consequence of
>   the same value being read from ICC_IAR{0,1}_EL1, and with the correct
>   nesting, breaks the state machine and leads to unpredictable results
>   that affects *all* interrupts in the system (most likely, the priority
>   system is dead). See Figure 4-3 ("Interrupt handling state machine")
>   in Arm IHI 0069F for a description of the acceptable transitions.
> 
>   Additionally, on implementations that have ICC_CTLR_EL1.SEIS==1, a
>   SError may be generated to signal the error. See the various
> 
>   <quote>
> 	  IMPLEMENTATION_DEFINED "SError ....";
>   </quote>
> 
>   that are all over the pseudocode contained in the same architecture
>   spec. Needless to say, this is pretty final for any SW that would do
>   silly things on such implementations (which do exist).
> 
> Given that in our implementation, every signalled interrupt is acked,
> handled and EOId in sequence, there is no reason to EOI all interrupts
> at ExitBootServices() time in the first place, so let's just drop this
> code. This fixes an issue reported by Marc where an SError is triggered
> by this code, bringing down the system.
> 
> Reported-by: Marc Zyngier <maz@kernel.org>
> Signed-off-by: Ard Biesheuvel <ardb@kernel.org>

Yikes.
Pretty sure that wasn't the right thing to do on gicv2 either.

Reviewed-by: Leif Lindholm <leif@nuviainc.com>

> ---
> This is a clear bugfix, but given how late we are in the cycle, I will
> leave it up to Liming to decide whether we can still take this for the
> upcoming stable tag.

I can see hypothetical situations where this could break existing
(broken) firmware ... so I would prefer to leave it until after the
stable tag unless there is substantial benefit to including it here.

/
    Leif

>  ArmPkg/Drivers/ArmGic/GicV3/ArmGicV3Dxe.c | 4 ----
>  1 file changed, 4 deletions(-)
> 
> diff --git a/ArmPkg/Drivers/ArmGic/GicV3/ArmGicV3Dxe.c b/ArmPkg/Drivers/ArmGic/GicV3/ArmGicV3Dxe.c
> index 85ee4c87b6d1..fa515d1a01ba 100644
> --- a/ArmPkg/Drivers/ArmGic/GicV3/ArmGicV3Dxe.c
> +++ b/ArmPkg/Drivers/ArmGic/GicV3/ArmGicV3Dxe.c
> @@ -344,10 +344,6 @@ GicV3ExitBootServicesEvent (
>      GicV3DisableInterruptSource (&gHardwareInterruptV3Protocol, Index);
>    }
>  
> -  for (Index = 0; Index < mGicNumInterrupts; Index++) {
> -    GicV3EndOfInterrupt (&gHardwareInterruptV3Protocol, Index);
> -  }
> -
>    // Disable Gic Interface
>    ArmGicV3DisableInterruptInterface ();
>  
> -- 
> 2.30.2
>