Re: PerformancePkg on multiple platform

[prabin ca <prabinca4u@gmail.com>]

Hi Laszlo,

Thank you for your valid input, Its really help full for me. I will look in
the " A Tour Beyond BIOS    Implementing Profiling in with EDK II" with
your points.

have a nice day.

On Fri, Aug 3, 2018 at 11:47 PM, Laszlo Ersek <lersek@redhat.com> wrote:

> Hi Prabin,
>
> On 08/03/18 09:29, prabin ca wrote:
> > Hi Team,
> >
> > I’m new to uefi and edk. Currently I’m trying to get performance of my
> dxe driver using PerformancePkg of EDK source code.
> >
> > I’m using perf_start and perf_end T respected check points, it’s hot
> build and tested well in 2/3 platform. It’s giving proper response.
> >
> > But when I’m Checking with multiple platform, in some of platforms it’s
> not working and got hang in uefi she’ll itself.
> >
> > Please help me to resolve asap, it will be really helpful.
>
> I can only give you some hints, because thus far I have also failed to
> figure out how performance measurements should be enabled for a platform
> from scratch.
>
> Earlier I thought that an interested platform should include modules
> from PerformancePkg. I got some good results that way; however:
>
> - it wouldn't work for AARCH64, and so I filed
>   <https://bugzilla.tianocore.org/show_bug.cgi?id=679>
>
> - ultimately PerformancePkg was removed in commit 922c1e94e4bb
>   completely. (And for that reason I now closed TianoCore#679 as
>   WONTFIX.)
>
> Commit 922c1e94e4bb doesn't really give pointers for enabling
> performance measurements in a platform -- it refers the user to the DP
> shell command instead of the standalone DP application, but that's only
> for displaying the stats, not for recording them.
>
> One document where I found information is the Intel whitepaper
>
>   A Tour Beyond BIOS
>   Implementing Profiling in with EDK II
>
> (just google it). "Part III – Performance Profile" is relevant.
>
> The first section of that talks about enabling ACPI FPDT (ACPI Firmware
> Performance Data Table) support in edk2, for exposing firmware
> performance data to the OS. Again, that's not about *recording* the stats.
>
> The second section of the same chapter seems to describe how stats can
> be recorded however. AIUI, the DXE Core can provide the PERFORMANCE[_EX]
> protocols, if the DxeCorePerformanceLib instance is linked into it by
> the platform DSC file. In turn DXE modules can send measurements to the
> protocol via the PerformanceLib APIs / DxePerformanceLib instance.
>
> In fact, the relevant library INF files have good documentation, we just
> have to know what to look at.
>
> Recording stats in the PEI phase, via HOBs:
> - MdeModulePkg/Library/PeiPerformanceLib/PeiPerformanceLib.inf
>
> Recording stats in the DXE phase (protocol provider and consumer):
> - MdeModulePkg/Library/DxeCorePerformanceLib/DxeCorePerformanceLib.inf
> - MdeModulePkg/Library/DxePerformanceLib/DxePerformanceLib.inf
>
> Recording stats in SMM (protocol provider and consumer):
> - MdeModulePkg/Library/SmmCorePerformanceLib/SmmCorePerformanceLib.inf
> - MdeModulePkg/Library/SmmPerformanceLib/SmmPerformanceLib.inf
>
> Fetching the collected stats (for display or otherwise):
> - MdeModulePkg/Library/DxeSmmPerformanceLib/DxeSmmPerformanceLib.inf
>
> "Do nothing" library instance for all module types:
> - MdePkg/Library/BasePerformanceLibNull/BasePerformanceLibNull.inf
>
>
> Now obviously this doesn't explain why it works for you on some
> platforms and why it doesn't on others. You haven't shared much
> information about the platforms in question. I can make one guess: if
> the performance protocol provided by the platform's DXE Core
> (DxeCorePerformanceLib) is incompatible with the DxePerformanceLib
> instance that is linked into your DXE driver, then communication between
> them will likely fail.
>
> To my understanding, the performance protocol is not standard (PI or
> UEFI), so it's likely best if you build your DXE driver together with
> the DXE Core, on all platforms that you want to check. It's probably
> unsafe to link a DXE driver against DxePerformanceLib (rather than
> BasePerformanceLibNull) and run it on a random platform.
>
> Hopefully this helps a little.
> Laszlo
>