Re: assertion "lp_max >= core_max" failed

On Sun, Sep 22, 2024 at 01:08:21PM -0000, Michael van Elst wrote:
> [email protected] (Christof Meerwald) writes:
> 
> >Turns out in my case the CPUID_HTT flag is not set (so lp_max = 1),
> >but the maximum number of cores per package is set to 16 - 1, so
> >core_max = 16.
> 
> That's a confusing number, the chip has 14 cores and 28 threads.
> If HTT is disabled, that would still be only 14 threads. So
> I guess, the VM software fakes these values.

Some of it is definitely faked, but I am not sure what the real
meaning of these is anyway.

BTW, this particular VM has 2 CPU cores assigned to it.

The one real Intel CPU I currently have access to is a "Intel(R)
Pentium(R) CPU N3540" - here the cpuid results would set lp_max=16 and
core_max=8, although the CPU only has 4 cores/threads.

It seems to me that most VMs set the CPUID_HTT to false.

Anyway, another Intel-based VM I have access to has CPUID_HTT false and

  maximum IDs for cores in pkg       = 0x0 (0)

so it would have lp_max=1 and core_max=1, but also has 2 CPU cores
assigned.


> >but we don't check for that condition in the Intel case - should we?
> >Adding that case for Intel seems to fix it for me, e.g.
> 
> >	/* Check for leaf 4 support. */
> >	if (ci->ci_max_cpuid >= 4 &&
> >	    (ci->ci_feat_val[0] & CPUID_HTT) != 0) {
> >		/* Maximum number of Cores per package (eax[31:26]). */
> >		x86_cpuid2(4, 0, descs);
> >		core_max = __SHIFTOUT(descs[0], CPUID_DCP_CORE_P_PKG)
> >		    + 1;
> >	} else {
> >		core_max = 1;
> >	}
> 
> Wouldn't that tell that we only have 1 core / 1 thread ?

I am not sure - I am currently running NetBSD with that change and it
has 2 CPU cores assigned and the 2 cores show up in /proc/cpuinfo and
in dmesg:

[     1.000004] cpu0 at mainbus0 apid 0
[     1.000004] cpu0: Use lfence to serialize rdtsc
[     1.000004] cpu0: Intel(R) Xeon(R) CPU E5-2690 v4 @ 2.60GHz, id 0x406f1
[     1.000004] cpu0: node 0, package 0, core 0, smt 0
[     1.000004] cpu1 at mainbus0 apid 1
[     1.000004] cpu1: Intel(R) Xeon(R) CPU E5-2690 v4 @ 2.60GHz, id 0x406f1
[     1.000004] cpu1: node 0, package 1, core 0, smt 0


> Maybe instead of failing when lp_max < core_max, we should just
> gracefully handle this case as lp_max == core_max.

That was actually my initial attempt (which also worked), but then I
saw that the AMD code checked for CPUID_HTT, so thought maybe the
Intel code should do the same?

But now I am thinking that maybe for CPUID_HTT false, both AMD and
Intel should essentially set lp_max to core_max (instead of setting
core_max to 1).


Christof

-- 

https://cmeerw.org                             sip:cmeerw at cmeerw.org
mailto:cmeerw at cmeerw.org                   xmpp:cmeerw at cmeerw.org