Re: [Intel Alderlake CPUID issue]

[Ian Rogers <irogers@google.com>]

On Tue, Jan 30, 2024 at 4:35 PM Tingrun Chen <tchen168@asu.edu> wrote:
>
> `perf test 7` fail on Alderlake virtual machine, The test is failing
> as #system_tsc_freq has a zero value.Below is the output of `./perf
> test 7 -vv`:
> ```
> Couldn't bump rlimit(MEMLOCK), failures may take place when creating
> BPF maps, etc
>   7: Simple expression parser                                        :
> --- start ---
> test child forked, pid 82095
> Using CPUID GenuineIntel-6-BA-2
> parsing metric: 1+1
> parsing metric: FOO+BAR
> lookup(FOO): val 1.000000
> lookup(BAR): val 2.000000
> parsing metric: (BAR/2)%2
> lookup(BAR): val 2.000000
> parsing metric: 1 - -4
> parsing metric: (FOO-1)*2 + (BAR/2)%2 - -4
> lookup(FOO): val 1.000000
> lookup(BAR): val 2.000000
> parsing metric: 1-1 | 1
> parsing metric: 1-1 & 1
> parsing metric: min(1,2) + 1
> parsing metric: max(1,2) + 1
> parsing metric: 1+1 if 3*4 else 0
> parsing metric: 100 if 1 else 200 if 1 else 300
> parsing metric: 100 if 0 else 200 if 1 else 300
> parsing metric: 100 if 1 else 200 if 0 else 300
> parsing metric: 100 if 0 else 200 if 0 else 300
> parsing metric: 1.1 + 2.1
> parsing metric: .1 + 2.
> parsing metric: d_ratio(1, 2)
> parsing metric: d_ratio(2.5, 0)
> parsing metric: 1.1 < 2.2
> parsing metric: 2.2 > 1.1
> parsing metric: 1.1 < 1.1
> parsing metric: 2.2 > 2.2
> parsing metric: 2.2 < 1.1
> parsing metric: 1.1 > 2.2
> parsing metric: 1.1e10 < 1.1e100
> parsing metric: 1.1e2 > 1.1e-2
> parsing metric: FOO/0
> lookup(FOO): val 1.000000
> division by zero
> parsing metric: BAR/
> lookup(BAR): val 2.000000
> syntax error
> parsing metric: FOO + BAR + BAZ + BOZO
> parsing metric: EVENT1\,param\=?@ + EVENT2\,param\=?@
> parsing metric: dash\-event1 - dash\-event2
> parsing metric: EVENT1 if #smt_on else EVENT2
> literal: #smt_on = 0.000000
> parsing metric: EVENT1 if #core_wide else EVENT2
> literal: #core_wide = 0.000000
> parsing metric: 1.0 if EVENT1 > 100.0 else 1.0
> parsing metric: 0 & EVENT1 > 0
> parsing metric: EVENT1 > 0 & 0
> parsing metric: 1 & EVENT1 > 0
> parsing metric: EVENT1 > 0 & 1
> parsing metric: 1 | EVENT1 > 0
> parsing metric: EVENT1 > 0 | 1
> parsing metric: 0 | EVENT1 > 0
> parsing metric: EVENT1 > 0 | 0
> parsing metric: #num_cpus_online
> literal: #num_cpus_online = 8.000000
> parsing metric: #num_cpus
> literal: #num_cpus = 8.000000
> parsing metric: #num_cores
> literal: #num_cores = 8.000000
> parsing metric: #num_dies
> literal: #num_dies = 1.000000
> parsing metric: #num_packages
> literal: #num_packages = 1.000000
> parsing metric: #system_tsc_freq
> Failed to find TSC frequency in /proc/cpuinfoliteral: #system_tsc_freq
> = 0.000000
> FAILED tests/expr.c:249 #system_tsc_freq > 0
> test child finished with -1
> ---- end ----
> Simple expression parser: FAILED!
> ```
> The test can pass in my AlderLakephysical machine. So, in my physical
> machine it seems that in the `arch_get_tsc_freq`
> (https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/arch/x86/util/tsc.c?h=perf-tools-next)
> function, the call to `get_cpuid_0` will read a valid value for lvl,
> then it will calculate the tsc using the information directly from the
> intel instruction `cpuid`; however, in the virtual machine, the lvl
> will be less than 0x15, then it will fall back to read `/proc/cpuinfo`
> and it does not contain something like `@ 3.0GHz`. Interestingly, the
> `model name` entry of `/proc/cpuinfo` in both my physical machine and
> the virtual machine does not contain the frequency information like `@
> 3.0GHz`.

+Kan Liang

Thanks Tingrun,

IIRC the "@ 3.0GHz" parsing approach was for older machines that
lacked the information in the cpuid. Could there be anything special
in your virtual machine set up that will cause the cpuid to look older
than it is? Kan, any thoughts and can we work around this?

Thanks,
Ian