From: Kinney, Michael D
Sent: Friday, August 16, 2019 12:23 AM
To: devel@edk2.groups.io; vit9696@protonmail.com; Kuo, Donald <donald.kuo@intel.com>; Kinney, Michael D <michael.d.kinney@intel.com>
Subject: RE: [edk2-devel] [PATCH] UefiCpuPkg: Adding a new TSC library by using CPUID(0x15) TSC leaf
Vitaly,
When implementing a UEFI Application, if you want maximum compatibility, you should use UEFI Services/Protocols and minimize as many HW assumptions as possible.
I understand, especially for accurate time and clock related services, the that the UEFI Specification defined Services/Protocols can be challenging to use.
When adding content that uses HW such as TSC or CPUID the UEFI App should be implemented with good detection logic to know it is safe to use that HW, and contain the fallback logic to use an alternate mechanism if the required HW is not present. This is a very different approach than some early FW initialization code that can safely make some HW assumptions that helps keep the FW init code small and efficient. This does imply that different libraries may be required for FW init and UEFI Applications.
Thanks,
Mike
Hi Donald,
Glad to hear it helped a little, and sorry for some outdated quotes.
Your clarification regarding model range is very helpful. Xeon W being client and thus having client clock makes sense, though I must say it was quite not obvious. I was referring to the same SDM table, and it would have been great to have vertical range binding instead of the misleading CPUID.
With that, however, I still do not see a way to dynamically determine the difference between Xeon client and server.
For us it is needed as we use TimerLib differently. For you TimerLib is a part of BSP, so you face no issue statically setting the clock frequency as a PCD. For us TimerLib is used as a part of an UEFI application compatible with different hardware, and in fact not just Intel. We have such "generic TimerLib" library, and to determine CPU frequency, as a fallback to CPUID 15H, it relies on ACPI PM timer. This is not great for two reasons:
- we have to maintain it ourselves, while we would have liked that be part of EDK II with different vendors contributing to the project.
- we still cannot find an obvious way to determine crystal clock when it is not reported, in particular for Xeon W and Xeon Scalable case.
I guess this is now out of the scope of this patch and perhaps even this exact library, but it will be helpful to hear relevant technical details on the issue and opinions on such "generic TimerLib" library to later appear in EDK II.
Best regards,
Vitaly
Appreciated for reviewing very detail. And looks your captured some piece of code is older version. And should be ok, I do got your point.
- I do missed RegEax error handling in case for CpuidCoreClockCalculateTscFrequency () should return 0 and EFI_UNSUPPORTED. AR: Will update it.
- Actually the information is from SDM Table 18-85
o As we know CPUID signature could be hard to identify processor XTAL frequency is? So we only check CPUID Leaf 0x15
o And the PCD will be defined depends on platform specific and during project early development. Based on SDM, Intel processor for CPUID.15h EAX and EBX is enumerated, but ECX could be possible not enumerated. So that is why we based on SDM Table 18-85 to create PCD as below:
§ Intel Xeon Server family: 25MHz
§ Intel Core and Intel Xeon W family: 24MHz
§ So regards the “06_55h” might cause the confused, we could review it whether remove it??
- Again, the XTAL frequency is optional and should be define in early phase of new project. Default should still use AcpiTimer.
o Platform / developer owner should well know the CPU generation XTAL frequency is? Server / Client / Atom ?
o For those CPU not supported should still use AcpiTimerLib (default)
Hello,
Thank you for the patch! I reviewed the code and noticed select issues explained below.
1. The following construction:
if (RegEbx == 0) {
DEBUG ((DEBUG_ERROR, "The CPU is not capble for Core Crystal Clock Frequency !!\n"));
ASSERT (RegEbx != 0);
}
Does not look good to me, and in my opinion, should be written differently:
if (RegEbx == 0 || RegEax == 0) {
DEBUG ((DEBUG_ERROR, "The CPU is not capble for Core Crystal Clock Frequency !!\n"));
ASSERT (RegEbx != 0);
ASSERT (RegEax != 0);
return 0;
}
The reason for the above code being wrong is potential division by zero below, which behaviour is undefined (and in fact unknown due to unspecified interrupt table state). Even though the intent is to not permit the use of this library on an unsupported platform, runtime checks and assertions are essentially NO-OPs, should be separate and not confused. For this to work properly the call to CpuidCoreClockCalculateTscFrequency should happen in library constructor with EFI_UNSUPPORTED return on CpuidCoreClockCalculateTscFrequency returning 0.
2. The notes about crystal clock frequency for 06_55H CPU signature:
"25000000 - Intel Xeon Processor Scalable Family with CPUID signature 06_55H(25MHz).<BR>\n"
# Intel Xeon Processor Scalable Family with CPUID signature 06_55H = 25000000 (25MHz)
are misleading in both this library and Intel SDM. Intel Xeon W processors have the same CPUID signature (06_55H), they report 0 crystal clock frequency, and their crystal clock frequency is 24 MHz. This should at least be mentioned in the lower part mentioning Intel Xeon W Processor Family(24MHz).
Actually, given that many Intel guys are here, I wonder whether anybody knows if there is any better approach to distinguish Xeon Scalable CPUs and Xeon W CPUs besides using brand string or using marketing frequency from CPUID 16H to determine crystal clock frequency (this is what Linux does at the moment)?
3. Intel Atom Denverton with CPUID signature (06_5FH), also known as Goldmont X, reports 0 crystal clock frequency as well, and has 25 MHz frequency. This is missing from SDM, but once again I believe it should be included in the two places from the above to avoid confusion.
Besides these 3 points, honestly, the library itself appears to be very limited for anything but embedding it into the firmware with known hardware, which already works just fine by defining the PCD. Missing 0 crystal clock frequency handling in runtime with hardcoding the PCD value looks very bad, because the number of modern Intel CPU models reporting 0 crystal clock frequency and having non 24 MHz frequency is quite far from 0. This makes the library essentially impossible to use in any UEFI application or third-party product even when targeting Skylake+ generation. To resolute this I vote for additional detection functionality to be added to the library to obtain crystal clock frequency when the CPUID reports 0. The PCD could be the last resort when no other method works. My opinion is that this should be resolved prior to merging the patch.
Best regards,
Vitaly