On 18/06/2018 12:22, Viresh Kumar wrote:
> On 15-06-18, 11:19, Daniel Lezcano wrote:
>> +/**
>> + * idle_injection_stop - stops the idle injections
>> + * @ii_dev: a pointer to an idle injection_device structure
>> + *
>> + * The function stops the idle injection and waits for the threads to
>> + * complete. If we are in the process of injecting an idle cycle, then
>> + * this will wait the end of the cycle.
>> + *
>> + * When the function returns there is no more idle injection
>> + * activity. The kthreads are scheduled out and the periodic timer is
>> + * off.
>> + */
>> +void idle_injection_stop(struct idle_injection_device *ii_dev)
>> +{
>> + struct idle_injection_thread *iit;
>> + unsigned int cpu;
>> +
>> + pr_debug("Stopping injecting idle cycles on CPUs '%*pbl'\n",
>> + cpumask_pr_args(to_cpumask(ii_dev->cpumask)));
>> +
>> + hrtimer_cancel(&ii_dev->timer);
>> +
>> + /*
>> + * We want the guarantee we have a quescient point where
>> + * parked threads stay in there state while we are stopping
>> + * the idle injection. After exiting the loop, if any CPU is
>> + * plugged in, the 'should_run' boolean being false, the
>> + * smpboot main loop schedules the task out.
>> + */
>> + cpu_hotplug_disable();
>> +
>> + for_each_cpu_and(cpu, to_cpumask(ii_dev->cpumask), cpu_online_mask) {
>
> Maybe you should do below for all CPUs in the mask. Is the below usecase
> possible ?
>
> - CPU0-4 are part of the mask and are all online.
> - hrtimer fires and sets should_run for all of them to 1.
^^
hrtimer_cancel gives you the guarantee, the timer is no longer active
and there is no execution in the timer handler. So the timer can no
longer fire after hrtimer_cancel() is called (which is a blocking call).
> - Right at this time CPU3 goes offline, so the thread gets parked with
> should_run == 1. Is there a reason why this can't happen ?
> - Now we unregister the stuff and CPU3 again comes online.
> - Because it had should_run as true, we again run the thread and Crash.
>
> makes sense ?
>> + iit = per_cpu_ptr(&idle_injection_thread, cpu);
>> + iit->should_run = 0;
>> +
>> + wait_task_inactive(iit->tsk, 0);
>
> I am not very sure of what guarantees this will provide.
We get the guarantee any idle injection cycle is ended.
> @Peter: Do you see any more race scenarios here ?
>
>> + }
>> +
>> + cpu_hotplug_enable();
>> +}
>
>> +struct idle_injection_device *idle_injection_register(struct cpumask
>> *cpumask)
>> +{
>> + struct idle_injection_device *ii_dev;
>> + int cpu;
>> +
>> + ii_dev = kzalloc(sizeof(*ii_dev) + cpumask_size(), GFP_KERNEL);
>> + if (!ii_dev)
>> + return NULL;
>> +
>> + cpumask_copy(to_cpumask(ii_dev->cpumask), cpumask);
>> + hrtimer_init(&ii_dev->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
>> + ii_dev->timer.function = idle_injection_wakeup_fn;
>> +
>> + for_each_cpu(cpu, to_cpumask(ii_dev->cpumask)) {
>> +
>> + if (per_cpu(idle_injection_device, cpu)) {
>> + pr_err("cpu%d is already registered\n", cpu);
>> + goto out_rollback_per_cpu;
>> + }
>> +
>> + per_cpu(idle_injection_device, cpu) = ii_dev;
>> + }
>> +
>> + return ii_dev;
>> +
>> +out_rollback_per_cpu:
>> + for_each_cpu(cpu, to_cpumask(ii_dev->cpumask))
>> + per_cpu(idle_injection_device, cpu) = NULL;
>
> So if two parts of the kernel call this routine with the same cpumask, then
> the
> second call will also overwrite the masks with NULL and return error. That
> will
> screw up things a bit here.
Apparently there is a misunderstanding :)
https://lkml.org/lkml/2018/5/29/209 (at the end)
--
<http://www.linaro.org/> Linaro.org │ Open source software for ARM SoCs
Follow Linaro: <http://www.facebook.com/pages/Linaro> Facebook |
<http://twitter.com/#!/linaroorg> Twitter |
<http://www.linaro.org/linaro-blog/> Blog
