Re: [RFC 3/3] preempt_tracer: Use a percpu variable to control traceble calls

From: Peter Zijlstra
Date: Wed May 29 2019 - 04:45:20 EST


On Tue, May 28, 2019 at 05:16:24PM +0200, Daniel Bristot de Oliveira wrote:
> #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
> defined(CONFIG_TRACE_PREEMPT_TOGGLE))
> +
> +DEFINE_PER_CPU(int, __traced_preempt_count) = 0;
> /*
> * If the value passed in is equal to the current preempt count
> * then we just disabled preemption. Start timing the latency.
> */
> void preempt_latency_start(int val)
> {
> - if (preempt_count() == val) {
> + int curr = this_cpu_read(__traced_preempt_count);

We actually have this_cpu_add_return();

> +
> + if (!curr) {
> unsigned long ip = get_lock_parent_ip();
> #ifdef CONFIG_DEBUG_PREEMPT
> current->preempt_disable_ip = ip;
> #endif
> trace_preempt_off(CALLER_ADDR0, ip);
> }
> +
> + this_cpu_write(__traced_preempt_count, curr + val);
> }
>
> static inline void preempt_add_start_latency(int val)
> @@ -3200,8 +3206,12 @@ NOKPROBE_SYMBOL(preempt_count_add);
> */
> void preempt_latency_stop(int val)
> {
> - if (preempt_count() == val)
> + int curr = this_cpu_read(__traced_preempt_count) - val;

this_cpu_sub_return();

> +
> + if (!curr)
> trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
> +
> + this_cpu_write(__traced_preempt_count, curr);
> }

Can't say I love this, but it is miles better than the last patch.