Re: [edk2-devel] [Patch] MdeModulePkg DxeCore: Fix for missing MAT update

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

On 08/10/19 16:10, Liming Gao wrote:
> From: Mike Turner <miketur@microsoft.com>
> 
> The Fpdt driver (FirmwarePerformanceDxe) saves a memory address across
> reboots, and then does an AllocatePage for that memory address.
> If, on this boot, that memory comes from a Runtime memory bucket,
> the MAT table is not updated. This causes Windows to boot into Recovery.

(1) What is "MAT"?

> This patch blocks the memory manager from changing the page
> from a special bucket to a different memory type.  Once the buckets are
> allocated, we freeze the memory ranges for the OS, and fragmenting
> the special buckets will cause errors resuming from hibernate.

(2) My understanding is that CoreConvertPages() will only hand out the
requested pages if those pages are currently free. I suggest clarifying
the commit message that the intent is to prevent the allocation of
otherwise *free* pages, if the allocation would fragment special buckets.

(3) I don't understand the conjunction "and". I would understand if the
statement were:

    Once the buckets are allocated, we freeze the memory ranges for the
    OS, *because* fragmenting the special buckets *would* cause errors
    resuming from hibernate.

Is this the intent?

> 
> This patch is cherry pick from Project Mu:
> https://github.com/microsoft/mu_basecore/commit/a9be767d9be96af94016ebd391ea6f340920735a
> With the minor change,
> 1. Update commit message format to keep the message in 80 characters one line.
> 2. Remove // MU_CHANGE comments in source code.
> 
> Cc: Jian J Wang <jian.j.wang@intel.com>
> Cc: Hao A Wu <hao.a.wu@intel.com>
> Cc: Dandan Bi <dandan.bi@intel.com>
> Signed-off-by: Liming Gao <liming.gao@intel.com>
> ---
>  MdeModulePkg/Core/Dxe/Mem/Page.c | 43 ++++++++++++++++++++++++++++++++++------
>  1 file changed, 37 insertions(+), 6 deletions(-)
> 
> diff --git a/MdeModulePkg/Core/Dxe/Mem/Page.c b/MdeModulePkg/Core/Dxe/Mem/Page.c
> index bd9e116aa5..637518889d 100644
> --- a/MdeModulePkg/Core/Dxe/Mem/Page.c
> +++ b/MdeModulePkg/Core/Dxe/Mem/Page.c
> @@ -1265,12 +1265,13 @@ CoreInternalAllocatePages (
>    IN BOOLEAN                NeedGuard
>    )
>  {
> -  EFI_STATUS      Status;
> -  UINT64          Start;
> -  UINT64          NumberOfBytes;
> -  UINT64          End;
> -  UINT64          MaxAddress;
> -  UINTN           Alignment;
> +  EFI_STATUS       Status;
> +  UINT64           Start;
> +  UINT64           NumberOfBytes;
> +  UINT64           End;
> +  UINT64           MaxAddress;
> +  UINTN            Alignment;
> +  EFI_MEMORY_TYPE  CheckType;
>  
>    if ((UINT32)Type >= MaxAllocateType) {
>      return EFI_INVALID_PARAMETER;
> @@ -1321,6 +1322,7 @@ CoreInternalAllocatePages (
>    // if (Start + NumberOfBytes) rolls over 0 or
>    // if Start is above MAX_ALLOC_ADDRESS or
>    // if End is above MAX_ALLOC_ADDRESS,
> +  // if Start..End overlaps any tracked MemoryTypeStatistics range
>    // return EFI_NOT_FOUND.
>    //
>    if (Type == AllocateAddress) {
> @@ -1336,6 +1338,35 @@ CoreInternalAllocatePages (
>          (End > MaxAddress)) {
>        return EFI_NOT_FOUND;
>      }
> +
> +    // Problem summary
> +
> +    /*
> +    A driver is allowed to call AllocatePages using an AllocateAddress type.  This type of
> +    AllocatePage request the exact physical address if it is not used.  The existing code
> +    will allow this request even in 'special' pages.  The problem with this is that the
> +    reason to have 'special' pages for OS hibernate/resume is defeated as memory is
> +    fragmented.
> +    */

(4) This comment style is not native to edk2.

I think the "problem summary" line should be removed, and the actual
problem statement should use the following comment style:

  //
  // blah
  //


> +
> +    for (CheckType = (EFI_MEMORY_TYPE) 0; CheckType < EfiMaxMemoryType; CheckType++) {
> +      if (MemoryType != CheckType &&
> +          mMemoryTypeStatistics[CheckType].Special &&
> +          mMemoryTypeStatistics[CheckType].NumberOfPages > 0) {
> +        if (Start >= mMemoryTypeStatistics[CheckType].BaseAddress &&
> +            Start <= mMemoryTypeStatistics[CheckType].MaximumAddress) {
> +          return EFI_NOT_FOUND;
> +        }
> +        if (End >= mMemoryTypeStatistics[CheckType].BaseAddress &&
> +            End <= mMemoryTypeStatistics[CheckType].MaximumAddress) {
> +          return EFI_NOT_FOUND;
> +        }
> +        if (Start < mMemoryTypeStatistics[CheckType].BaseAddress &&
> +            End   > mMemoryTypeStatistics[CheckType].MaximumAddress) {
> +          return EFI_NOT_FOUND;
> +        }
> +      }
> +    }
>    }

(5) Checking for overlap (i.e., whether the intersection is non-empty)
can be done more simply (i.e., with fewer comparisons in the worst case,
and with less code):

  if (MAX (Start, mMemoryTypeStatistics[CheckType].BaseAddress) <=
      MIN (End, mMemoryTypeStatistics[CheckType].MaximumAddress)) {
    return EFI_NOT_FOUND;
  }

but the proposed intersection check is technically right already, IMO,
so there's no strong need to update it.

(Somewhat unusually for this kind of comparison, all four boundaries are
inclusive here.)

(6) Both the commit message and the code comment state that this problem
is specific to S4. Therefore, we can distinguish three cases:

(6a) Platform doesn't support (or doesn't enable) S4 at all.

(6b) Platform supports & enables S4, and this is a normal boot.

(6c) Platform supports & enables S4, and this is actually an S4 resume.

The code being proposed applies to all three cases. Is that the intent?
Shouldn't the new check be made conditional on (6c) -- from the boot
mode HOB --, or at least on (6b)||(6c) -- i.e. the check should be
disabled if S4 is absent entirely?

Thanks,
Laszlo


>  
>    if (Type == AllocateMaxAddress) {
>