Messages in this thread | | | Date | Thu, 11 Apr 2013 07:08:29 -0500 | From | Robin Holt <> | Subject | Re: [PATCH] Do not force shutdown/reboot to boot cpu. |
| |
On Thu, Apr 11, 2013 at 07:03:58AM -0500, Robin Holt wrote: > On Thu, Apr 11, 2013 at 02:00:27PM +0200, Ingo Molnar wrote: > > > > * Robin Holt <holt@sgi.com> wrote: > > > > > > Ok, so it looks profilable. > > > > > > > > The result above is not surprising: most CPUs sit in idle and don't do anything, > > > > while the loop goes on, right? > > > > > > > > The interesting thing to profile would be the parallel bring-down, with the > > > > simplest global lock solution you mentioned. In that case most CPUs should be > > > > doing 'something' all the time - maybe spinning on the lock, maybe something else, > > > > right? > > > > > > Again, mostly looks idle. > > > > Forgot to suggest: > > > > perf record -a /sbin/reboot > > I used perf record -a /sbin/reboot -f -d -n
OK. Looking at Russ' patch, I understand now why it is looking idle. We are still serially doing the DOWN_PREPARE, etc. All those other cpus are still sitting idle.
Can we call the __cpu_down functions from an smp_call_function()?
Robin > > Robin > > > > ... to capture remote CPU activity too. > > > > > Events: 5M cycles > > > 31.69% swapper [kernel.kallsyms] [k] update_cfs_rq_blocked_load > > > 14.22% swapper [kernel.kallsyms] [k] load_balance > > > 12.95% swapper [kernel.kallsyms] [k] ktime_get > > > 4.64% swapper [kernel.kallsyms] [k] idle_cpu > > > 3.46% swapper [kernel.kallsyms] [k] uv_read_rtc > > > 2.26% swapper [kernel.kallsyms] [k] ktime_get_update_offsets > > > 2.25% swapper [kernel.kallsyms] [k] rcu_check_callbacks > > > 1.72% swapper [kernel.kallsyms] [k] _raw_spin_lock_irqsave > > > 1.57% swapper [kernel.kallsyms] [k] native_write_msr_safe > > > 1.53% swapper [kernel.kallsyms] [k] native_safe_halt > > > 1.52% swapper [kernel.kallsyms] [k] apic_timer_interrupt > > > 1.52% swapper [kernel.kallsyms] [k] update_blocked_averages > > > 1.51% swapper [kernel.kallsyms] [k] __lock_text_start > > > 1.48% swapper [kernel.kallsyms] [k] rcu_process_gp_end > > > 1.40% swapper [kernel.kallsyms] [k] rcu_process_callbacks > > > 1.19% reboot [kernel.kallsyms] [k] kmem_cache_alloc_node > > > 0.63% swapper [kernel.kallsyms] [k] check_for_new_grace_period > > > 0.58% swapper [kernel.kallsyms] [k] rebalance_domains > > > 0.55% swapper [kernel.kallsyms] [k] cpumask_next_and > > > 0.54% swapper [kernel.kallsyms] [k] __tick_nohz_idle_enter > > > 0.53% swapper [kernel.kallsyms] [k] perf_adjust_freq_unthr_context > > > 0.49% swapper [kernel.kallsyms] [k] _raw_spin_lock > > > > If even perf record -a shows a mostly idle system, then the overhead must be in > > sleep/wakeup latencies - for that the next step would be to figure out where all > > the waiting happens, for example via call-graph context-switch profiling: > > > > perf stat --null perf record -a -g -e sched:sched_switch /sbin/reboot > > > > (the perf stat --null will tell us the runtime of the whole operation.) > > > > Thanks, > > > > Ingo
| |