x86: do not use TSC as cycle count

[dreamliner787-9]

See #87 for more detail.

Test with seL4/seL4_libs#116.

[dreamliner787-9]

I'm not quite sure on how to deal with this in the kernel though: https://github.com/seL4/seL4/blob/1326364bc9135d9445d936ebc01e38a402c1f4c6/include/arch/x86/arch/benchmark.h#L20

Changing that rdtsc to rdpmc means that the kernel will assume that userspace have set up the PMU accordingly.

[Indanz]

I'm not quite sure on how to deal with this in the kernel though: https://github.com/seL4/seL4/blob/1326364bc9135d9445d936ebc01e38a402c1f4c6/include/arch/x86/arch/benchmark.h#L20

Changing that rdtsc to rdpmc means that the kernel will assume that userspace have set up the PMU accordingly.

If it faults if it isn't set up properly, then it's not an option. Otherwise, the kernel can setup a sane init value at boot and use that. If PMC access is exported to user space, they can change the config later if they want.

But I think we should remove all kernel benchmarking code, is it uses by anyone? Most things you can measure from user space and it's easy to filter out high measurements caused by scheduling. If no one is actively using this, then high chances this isn't working anyway for one reason or the other.

[KurtWu10]

But I think we should remove all kernel benchmarking code, is it uses by anyone?

Yes, sDDF's echo server benchmarking relies on the kernel for kernel utilisation. Or are you referring to RFC-16 on the PMU cap?

[Indanz]

Yes, sDDF's echo server benchmarking relies on the kernel for kernel utilisation.

What is it trying to measure? I would expect that the overhead caused by the kernel tracking would make any measurements done there close to useless.

Or are you referring to RFC-16 on the PMU cap?

No, that's unrelated.

Btw, neither of the sel4bench_private_serialize_pmc() implementation have a memory clobber in the asm, so compiler could be free to re-order memory accesses in unexpected ways.

[KurtWu10]

What is it trying to measure?

Per-pd user/kernel cycle counts. For kernel cycle counts, it shows whether the kernel is running on the fastpath.