On 1/27/20 8:36 PM, Sebastian Andrzej Siewior wrote:
I've been looking at usage of per-CPU variable cpu_hw_events in arch/powerpc/perf/core-book3s.c. power_pmu_enable() and power_pmu_disable() (pmu::pmu_enable() and pmu::pmu_disable()) are accessing the variable and the callbacks are invoked always with disabled interrupts. power_pmu_event_init() (pmu::event_init()) is invoked from preemptible context and uses get_cpu_var() to obtain a stable pointer (by disabling preemption). pmu::pmu_enable() and pmu::pmu_disable() can be invoked via a hrtimer (perf_mux_hrtimer_handler()) and it invokes pmu::pmu_enable() and pmu::pmu_disable() as part of the callback. Is there anything that prevents the timer callback to interrupt pmu::event_init() while it is accessing per-CPU data?
Sorry for the delayed response. Yes, currently we dont have anything that prevents the timer callback to interrupt pmu::event_init. Nice catch. Thanks for pointing this out. Looking at the code, per-cpu variable access are made to check for constraints and for Branch Stack (BHRB). So could wrap this block of pmu::event_init with local_irq_save/restore. Will send a patch to fix it. Maddy
Sebastian
