Re: [edk2-devel] [PATCH v1 2/2] UefiCpuPkg/PiSmmCpuDxeSmm: Check BspIndex first before lock cmpxchg

["Laszlo Ersek" <lersek@redhat.com>]

On 2/1/24 12:20, Wu, Jiaxin wrote:
> This patch is to check BspIndex first before lock cmpxchg operation.
> It's the optimization to lower the resource contention caused by the
> atomic compare exchange operation, so as to improve the SMI
> performance for BSP election.
> 
> Cc: Ray Ni <ray.ni@intel.com>
> Cc: Laszlo Ersek <lersek@redhat.com>
> Cc: Eric Dong <eric.dong@intel.com>
> Cc: Zeng Star <star.zeng@intel.com>
> Cc: Gerd Hoffmann <kraxel@redhat.com>
> Cc: Rahul Kumar <rahul1.kumar@intel.com>
> Signed-off-by: Jiaxin Wu <jiaxin.wu@intel.com>
> ---
>  UefiCpuPkg/PiSmmCpuDxeSmm/MpService.c | 12 +++++++-----
>  1 file changed, 7 insertions(+), 5 deletions(-)
> 
> diff --git a/UefiCpuPkg/PiSmmCpuDxeSmm/MpService.c b/UefiCpuPkg/PiSmmCpuDxeSmm/MpService.c
> index e988ce0542..479024d294 100644
> --- a/UefiCpuPkg/PiSmmCpuDxeSmm/MpService.c
> +++ b/UefiCpuPkg/PiSmmCpuDxeSmm/MpService.c
> @@ -1652,15 +1652,17 @@ SmiRendezvous (
>              }
>            } else {
>              //
>              // Platform hook fails to determine, use default BSP election method
>              //
> -            InterlockedCompareExchange32 (
> -              (UINT32 *)&mSmmMpSyncData->BspIndex,
> -              (UINT32)-1,
> -              (UINT32)CpuIndex
> -              );
> +            if (mSmmMpSyncData->BspIndex == (UINT32)-1) {
> +              InterlockedCompareExchange32 (
> +                (UINT32 *)&mSmmMpSyncData->BspIndex,
> +                (UINT32)-1,
> +                (UINT32)CpuIndex
> +                );
> +            }
>            }
>          }
>        }
>  
>        //

This patch makes me uncomfortable. I understand what it intends to do,
and the intent is not wrong, but we're again treating "volatile UINT32"
as atomic, and non-reorderable against the
InterlockedCompareExchange32(). This kind of "optimization" is what
people write cautionary tales about, later. It would be nice to see a
semi-formal *proof* that this cannot backfire.

Either way: I'm not trying to block the patch. If Ray is happy with it,
I don't object. OVMF implements PlatformSmmBspElection() in
"OvmfPkg/Library/SmmCpuPlatformHookLibQemu/SmmCpuPlatformHookLibQemu.c",
always returning EFI_SUCCESS [*], so the code that is being modified
here cannot be reached in OVMF.

[*] commit 43df61878d94 ("OvmfPkg: enable SMM Monarch Election in
PiSmmCpuDxeSmm", 2020-03-04)

Laszlo



-=-=-=-=-=-=-=-=-=-=-=-
Groups.io Links: You receive all messages sent to this group.
View/Reply Online (#115036): https://edk2.groups.io/g/devel/message/115036
Mute This Topic: https://groups.io/mt/104094808/7686176
Group Owner: devel+owner@edk2.groups.io
Unsubscribe: https://edk2.groups.io/g/devel/unsub [rebecca@openfw.io]
-=-=-=-=-=-=-=-=-=-=-=-