Re: [PATCH] UefiCpuPkg/MtrrLib: Add flag to skip disabling MTRRs prior to MTRR change.

["Duran, Leo" <leo.duran@amd.com>]

> -----Original Message-----
> From: Duran, Leo
> Sent: Friday, September 21, 2018 11:53 AM
> To: 'Ni, Ruiyu' <ruiyu.ni@Intel.com>; Laszlo Ersek <lersek@redhat.com>;
> edk2-devel@lists.01.org
> Cc: Dong, Eric <eric.dong@intel.com>
> Subject: RE: [edk2] [PATCH] UefiCpuPkg/MtrrLib: Add flag to skip disabling
> MTRRs prior to MTRR change.
> 
> 
> 
> > -----Original Message-----
> > From: Ni, Ruiyu <ruiyu.ni@Intel.com>
> > Sent: Wednesday, September 19, 2018 3:59 AM
> > To: Duran, Leo <leo.duran@amd.com>; Laszlo Ersek <lersek@redhat.com>;
> > edk2-devel@lists.01.org
> > Cc: Dong, Eric <eric.dong@intel.com>
> > Subject: Re: [edk2] [PATCH] UefiCpuPkg/MtrrLib: Add flag to skip
> > disabling MTRRs prior to MTRR change.
> >
> > On 9/18/2018 10:57 PM, Duran, Leo wrote:
> > >
> > >
> > >> -----Original Message-----
> > >> From: Ni, Ruiyu [mailto:ruiyu.ni@Intel.com]
> > >> Sent: Tuesday, September 18, 2018 3:34 AM
> > >> To: Laszlo Ersek <lersek@redhat.com>; Duran, Leo
> > <leo.duran@amd.com>;
> > >> edk2-devel@lists.01.org
> > >> Cc: Dong, Eric <eric.dong@intel.com>
> > >> Subject: Re: [edk2] [PATCH] UefiCpuPkg/MtrrLib: Add flag to skip
> > >> disabling MTRRs prior to MTRR change.
> > >>
> > >> On 9/18/2018 12:38 AM, Laszlo Ersek wrote:
> > >>> On 09/17/18 18:20, Duran, Leo wrote:
> > >>>>
> > >>>>> -----Original Message-----
> > >>>>> From: Ni, Ruiyu <ruiyu.ni@Intel.com>
> > >>>>> Sent: Thursday, September 13, 2018 11:44 PM
> > >>>>> To: Duran, Leo <leo.duran@amd.com>; Laszlo Ersek
> > >>>>> <lersek@redhat.com>; edk2-devel@lists.01.org
> > >>>>> Cc: Dong, Eric <eric.dong@intel.com>
> > >>>>> Subject: Re: [edk2] [PATCH] UefiCpuPkg/MtrrLib: Add flag to skip
> > >>>>> disabling MTRRs prior to MTRR change.
> > >>>>>
> > >>>>> On 9/14/2018 3:31 AM, Duran, Leo wrote:
> > >>>>>>
> > >>>>>>
> > >>>>>>> -----Original Message-----
> > >>>>>>> From: Ni, Ruiyu <ruiyu.ni@intel.com>
> > >>>>>>> Sent: Wednesday, September 12, 2018 9:39 PM
> > >>>>>>> To: Duran, Leo <leo.duran@amd.com>; Laszlo Ersek
> > >>>>> <lersek@redhat.com>;
> > >>>>>>> edk2-devel@lists.01.org
> > >>>>>>> Cc: Dong, Eric <eric.dong@intel.com>
> > >>>>>>> Subject: RE: [PATCH] UefiCpuPkg/MtrrLib: Add flag to skip
> > >>>>>>> disabling MTRRs prior to MTRR change.
> > >>>>>>>
> > >>>>>>> Leo,
> > >>>>>>> Sorry I was in leave yesterday so didn't see the mail.
> > >>>>>>> Which MSRs are shared? Only the
> > >>>>> MSR_IA32_MTRR_DEF_TYPE_REGISTER?
> > >>>>>>> Or all the MSRs that configures the CPU MTRR setting?
> > >>>>>>>
> > >>>>>>
> > >>>>>> Hi Ray,
> > >>>>>> The MTTR config MSRs are also shared by threads within a core.
> > >>>>>>
> > >>>>>
> > >>>>> Hi Leo,
> > >>>>> Do you think that fixing the caller is more proper?
> > >>>>
> > >>>> Hi Ray,
> > >>>> Actually,
> > >>>> The proposed PCD is the simplest solution, as that works for us
> > >>>> and does
> > >> not change the existing (default) flow.
> > >>>>
> > >>>> That is,
> > >>>> I'd prefer making a decision about the PCD in platform-specific
> > >>>> code,
> > >> rather than introducing complex detection and heuristics at the
> > >> caller level in
> > >> EDK2 (just for AMD).
> > >>>>
> > >>>> So, please approve the PCD.
> > >>
> > >> Leo,
> > >> I agree with you on the first part "the PCD is the simplest solution".
> > >> But this really looks like a workaround of the real issue.
> > >> For a multiple-socket system, it may contain S sockets, each socket
> > >> contains C cores and each core contains T threads. In summary the
> > >> system contains S * C * T threads.
> > >> As you said all threads inside a core share the MTRR setting.
> > >> Do all cores inside a socket share the MTRR setting?
> > >> Do all sockets share the MTRR setting?
> > >>
> > >> If one of the answer of above questions is "no", how can we
> > >> configure the PCD?
> > >>
> > > [Duran, Leo]
> > > Hi Ray,
> > > The MTTR settings are share by threads within a core (but each core
> > > has its own, etc.) The PCD would be set in our platform-specific
> > > code (e.g.,
> > it can be set at build-time in the .DSC file).
> > >
> > > As I mentioned,
> > > We don't need (Mtrr.Enable=0) to change MTRR settings, so having the
> > PCD to skip (Mtrr.Enable=0) is reasonable for us.
> > >
> > > Leo.
> > >
> >
> > If the PCD is false, no thread disables the MTRR before programming it.
> > Is it safe? Per Intel's SDM, it's not.
> >
> > Maybe it works in AMD's case. But I still suggest we change the
> > caller, which is more natural.
> > At least I'd like to see how potential-ugly the change can be.
> > We can then discuss how to make the ugly change better looking.
> >
> 
> Hi Ray,
> Please pardon the late reply.
> The main problem with changes to "caller" code is that dependencies are
> SoC-specific, so the detection code would not scale over time.
> Again, the proposed PCD does not alter existing flow (so existing code will
> continue to work as-is), and would give us a lever we can use in platform-
> specific code  (without requiring surgery in EDK2 "caller" code).
> 
BTW,
If you're concerned that someone may inadvertently set the PCD in their platform, I can ensure the PCD only applies on AMD (similar to changes I introduced in the APIC library).
For example, something like this:
  //
  // Disable MTRRs
  //
  if (!StandardSignatureIsAuthenticAMD () || !PcdGetBool (PcdSkipDisableMtrrsOnPreMtrrChangeOnAmd)) {
    DefType.Uint64 = AsmReadMsr64 (MSR_IA32_MTRR_DEF_TYPE);
    DefType.Bits.E = 0;
    AsmWriteMsr64 (MSR_IA32_MTRR_DEF_TYPE, DefType.Uint64);
 }

Please let me know if that's better, and will submit an updated patch.

> Thanks,
> Leo.
> 
> > >>>
> > >>> - From my side, if it works for you, it works for me. (The general
> > >>> trend has been to avoid adding more PCDs to the "core" package DEC
> > >>> files, but I'm 100% neutral on that.)
> > >>>
> > >>> Laszlo
> > >>>
> > >>
> > >> Laszlo,
> > >> Thanks for pointing out the general trend. Yes less PCDs are very
> > welcomed.
> > >> To me, PCD is no different from protocol. And even worse, because
> > >> it's very easily to be over-used.
> > >> But I am not sure whether a PCD has to be introduced for this issue.
> > >> Maybe even we choose to fix the caller, the PCD is still needed. I
> > >> am not sure.
> > >>
> > >> --
> > >> Thanks,
> > >> Ray
> > > _______________________________________________
> > > edk2-devel mailing list
> > > edk2-devel@lists.01.org
> > > https://lists.01.org/mailman/listinfo/edk2-devel
> > >
> >
> >
> > --
> > Thanks,
> > Ray