Re: [RFC PATCH 1/2] perf_events: add support for per-cpu per-cgroup monitoring (v3)

[Eric Dumazet <eric.dumazet@gmail.com>]

Le jeudi 09 septembre 2010 à 06:09 -0700, Stephane Eranian a écrit :
> This kernel patch adds the ability to filter monitoring based on
> container groups (cgroups). This is for use in per-cpu mode only.
>     
> The patch adds perf_event_attr.cgroup, a boolean, to activate
> this new mode. The cgroup is designated by passing in
> perf_event_attr.cgroup_fd, an opened file descriptor to
> the <mnt>/<cgroup>/perf_event.perf file.
>     
> This is the second version of this patch. It corrects the way
> time_enabled is accounted for. In cgroup mode, time_enabled reflects
> the time the cgroup was active, i.e., threads from the cgroup executed
> on the monitored CPU.  This is a more useful metric than just
> wall-clock. The meaning of time_enabled without cgroup is unaffected.
> 
> Signed-off-by: Stephane Eranian <eranian@google.com>


> +#ifdef CONFIG_CGROUPS
> +struct perf_cgroup_time {
> +	u64 time;
> +	u64 timestamp;
> +};
> +
> +struct perf_cgroup {
> +	struct cgroup_subsys_state css;
> +	struct perf_cgroup_time *time;

	struct perf_cgroup_time __percpu *time;

Please run sparse  after this "__percpu" change.



> +};
> +#endif
> +
>  /*
> +
> +	jc->time = alloc_percpu(struct perf_cgroup_time);
> +	if (!jc->time) {
> +		vfree(jc);
> +		return ERR_PTR(-ENOMEM);
> +	}
> +
> +	for_each_possible_cpu(c) {
> +		t = per_cpu_ptr(jc->time, c);
> +		t->time = 0;
> +		t->timestamp = 0;
> +	}

alloc_percpu() is zalloc_percpu() in fact, memory is already cleared.