casaroli commented on PR #21424: URL: https://github.com/apache/nuttx/pull/21424#issuecomment-6795722414
On SMP with arm64, the run-time accounting of critmonitor subtracts a PMU cycle count of one CPU from a stamp that another CPU took (nxsched_update_critmon(), nxsched_critmon_runtime(); master does the same in nxsched_update_critmon()). The PMU cycle counters of the CPUs are not synchronized: on qemu-armv8a:nsh_smp they differ by about 3 ms (each secondary CPU resets PMCCNTR when it starts), and CLOCK_THREAD_CPUTIME_ID goes back by up to 3 ms, thousands of times in 10 s; the CPU load of `ps` is wrong too. With CNTVCT_EL0 as the time base, the same test gives 0 steps back. Which fix do you prefer? 1. With SMP, every CPU accounts only its own thread: other CPUs account through an async SMP call, and a reader adds the running slice only for a thread on its own CPU (a lag of up to one sample period). 2. An arch declares a per-CPU perf counter, and with SMP critmonitor uses a counter that is the same on all CPUs for the run time and the CPU load (CNTVCT_EL0 on arm64). I will fix it in a separate PR before this one. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
