Re: [PATCH] UefiCpuPkg/MtrrLib: Fix MtrrDebugPrintAllMtrrsWorker to avoid hang

[Laszlo Ersek <lersek@redhat.com>]

On 10/17/17 03:46, Ruiyu Ni wrote:
> ARRAY_SIZE(Mtrrs->Variables.Mtrr) was used in
> MtrrDebugPrintAllMtrrsWorker() to parse the MTRR registers.
> Instead, the actual variable MTRR count should be used.
> Otherwise, the uninitialized random data in MtrrSetting may cause
> MtrrLibSetMemoryType() hang.
> 
> Steven Shi found this bug in QEMU when using Q35 chip.
> 
> Contributed-under: TianoCore Contribution Agreement 1.1
> Signed-off-by: Ruiyu Ni <ruiyu.ni@intel.com>
> Cc: Steven Shi <steven.shi@intel.com>
> Cc: Laszlo Ersek <lersek@redhat.com>
> ---
>  UefiCpuPkg/Library/MtrrLib/MtrrLib.c | 9 ++++++---
>  1 file changed, 6 insertions(+), 3 deletions(-)
> 
> diff --git a/UefiCpuPkg/Library/MtrrLib/MtrrLib.c b/UefiCpuPkg/Library/MtrrLib/MtrrLib.c
> index 2fd1d0153e..cb22558103 100644
> --- a/UefiCpuPkg/Library/MtrrLib/MtrrLib.c
> +++ b/UefiCpuPkg/Library/MtrrLib/MtrrLib.c
> @@ -2776,6 +2776,7 @@ MtrrDebugPrintAllMtrrsWorker (
>      UINTN             RangeCount;
>      UINT64            MtrrValidBitsMask;
>      UINT64            MtrrValidAddressMask;
> +    UINT32            VariableMtrrCount;
>      MTRR_MEMORY_RANGE Ranges[
>        ARRAY_SIZE (mMtrrLibFixedMtrrTable) * sizeof (UINT64) + 2 * ARRAY_SIZE (Mtrrs->Variables.Mtrr) + 1
>        ];
> @@ -2785,6 +2786,8 @@ MtrrDebugPrintAllMtrrsWorker (
>        return;
>      }
>  
> +    VariableMtrrCount = GetVariableMtrrCountWorker ();
> +
>      if (MtrrSetting != NULL) {
>        Mtrrs = MtrrSetting;
>      } else {
> @@ -2802,7 +2805,7 @@ MtrrDebugPrintAllMtrrsWorker (
>        DEBUG((DEBUG_CACHE, "Fixed MTRR[%02d]   : %016lx\n", Index, Mtrrs->Fixed.Mtrr[Index]));
>      }
>  
> -    for (Index = 0; Index < ARRAY_SIZE (Mtrrs->Variables.Mtrr); Index++) {
> +    for (Index = 0; Index < VariableMtrrCount; Index++) {
>        if (((MSR_IA32_MTRR_PHYSMASK_REGISTER *)&Mtrrs->Variables.Mtrr[Index].Mask)->Bits.V == 0) {
>          //
>          // If mask is not valid, then do not display range
> @@ -2829,11 +2832,11 @@ MtrrDebugPrintAllMtrrsWorker (
>      RangeCount = 1;
>  
>      MtrrLibGetRawVariableRanges (
> -      &Mtrrs->Variables, ARRAY_SIZE (Mtrrs->Variables.Mtrr),
> +      &Mtrrs->Variables, VariableMtrrCount,
>        MtrrValidBitsMask, MtrrValidAddressMask, RawVariableRanges
>        );
>      MtrrLibApplyVariableMtrrs (
> -      RawVariableRanges, ARRAY_SIZE (RawVariableRanges),
> +      RawVariableRanges, VariableMtrrCount,
>        Ranges, ARRAY_SIZE (Ranges), &RangeCount
>        );
>  
> 

Assuming this patch is not committed yet:

Reviewed-by: Laszlo Ersek <lersek@redhat.com>


I have another (independent) comment:

This function is currently named "MtrrDebugPrintAllMtrrsWorker", and its
leading comment says "Worker function prints all MTRRs for debugging."

Because of this name and the documentation, I didn't understand
initially how the problem could cause a hang, given that none of the
printing loops would actually access anything out-of-bounds. Some of the
information printed would be garbage, but it should not cause a hang.

That's when I noticed that the function actually *applies* MTRR
settings, by calling MtrrLibApplyVariableMtrrs(). Even worse, the
MtrrLibApplyVariableMtrrs() and MtrrLibApplyFixedMtrrs() function calls
are wrapped by DEBUG_CODE(). This means that in a DEBUG/NOOPT build, the
function will apply MTRR settings, and in a RELEASE build, it won't.

I think this is wrong and should be fixed. A debug function (esp. one
that behaves differently in DEBUG/NOOPT vs. RELEASE) should have no side
effects.

The current situation is similar to:

  ASSERT (FunctionWithSideEffects () == EXPECTED_RETURN_VALUE);

which we all know is wrong.

Thanks
Laszlo