[hdiethelm]

Fallback in case isolcpus is not set: After #3903, no CPU was assigned to the real-time threads and a warning is shown. This assumed all users set at least isolcpus as mentioned in https://linuxcnc.org/docs/html/install/latency-test.html#_tuning_preempt_rt_for_latency.

However, it turned out, not assigning a CPU when isolcpus is not set results in worse latency. And it fails tests/threads.0 sporadically. This restores the behavior from before, assigning the highest number CPU if isolcpus is not set.

The reason sched_getaffinity() is used and not long ncpus = sysconf(_SC_NPROCESSORS_CONF) is that on linux, you can offline CPU's or disable SMT on the kernel command line which could result in an issue. sched_getaffinity() returns the processor's the thread can run on and uses the highest of these. However, it will exclude isolated CPU's if parse_cpu_list() fails to detect the isolated CPU's.

Fixes #4734
Partially fixes #4525

[rene-dev]

Highest number could choose an efficiency core: https://stackoverflow.com/a/79238548

[hdiethelm]

Highest number could choose an efficiency core: https://stackoverflow.com/a/79238548

Not could, it will for sure. This also came to my mind while working on the kernel command line generator. However, I have no way to check it if I fix it. Do you have such a PC so you can test this?

[hdiethelm]

Hmm, on the other hand, using the efficiency cores could be better or worse for latency. More GHz does not mean automatically less latency. This needs testing. But I can't, I don't have such a PC.

isolcpus should anyway always be set. This PR is mostly to not brake cases where it is not done and restores the old behavior.

However, for the kernel command line tool I am working on, it would be good to take the big/little issue into account. Right now, it will fail in this case.

[grandixximo]

Merged this, and the follow up can continue on #4719

[rmu75]

Intel 12th generation+ efficiency cores should have performance comparable to skylake (intel core 6th generation), so are more than powerful enough. I don't think that the microarchitecture will have noticeable impacts on latency, but things like numa-architecture, powersaving, speed governor etc... could behave differently than on more uniform CPUs and have an impact.

I can test on laptop hardware with Alder Lake 12th generation i5-1235U with 2 performance cores (4 threads) and 8 efficiency cores (8 threads).