Re: [PATCH v3 2/2] x86, cpuid: allow cpuid_read() to schedule

From: Eric Dumazet
Date: Fri Mar 23 2018 - 21:01:16 EST




On 03/23/2018 03:17 PM, H. Peter Anvin wrote:
> On 03/23/18 14:58, Eric Dumazet wrote:
>> I noticed high latencies caused by a daemon periodically reading various
>> MSR and cpuid on all cpus. KASAN kernels would see ~10ms latencies
>> simply reading one cpuid. Even without KASAN, sending IPI to CPU
>> in deep sleep state or blocking hard IRQ in a a long section,
>> then waiting for the answer can consume hundreds of usec or more.
>>
>> Switching to smp_call_function_single_async() and a completion
>> allows to reschedule and not burn cpu cycles.
>
> That being said, the Right Way for a daemon to read multiple MSRs and
> CPUIDs on multiple CPUs is to spawn a thread for each CPU and use CPU
> affinity to lock them down. No IPI is needed to access MSRs on the
> current CPU, and CPUID doesn't even need kernel entry.

Indeed, assuming a daemon can have threads running on all cpus :/

Some environments like to partition cpus for different jobs/containers.

Yes, we can avoid IPI by carefully re-designing these user programs.