Re: [PATCH 8/9] per-cgroup boot time

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Wed, 2011-09-14 at 17:04 -0300, Glauber Costa wrote:
> Record the time in which the cgroup was created. This can be
> used to provide a more accurate boottime information in
> cpuacct.proc.stat.
> 
> Signed-off-by: Glauber Costa <glommer@parallels.com>
> ---
>  kernel/sched.c |    8 +++++++-
>  1 files changed, 7 insertions(+), 1 deletions(-)
> 
> diff --git a/kernel/sched.c b/kernel/sched.c
> index 4611c54..8f254d0 100644
> --- a/kernel/sched.c
> +++ b/kernel/sched.c
> @@ -9120,6 +9120,7 @@ struct cpuacct {
>  	/* cpuusage holds pointer to a u64-type object on every cpu */
>  	u64 __percpu *cpuusage;
>  	u64 __percpu *cpustat;
> +	struct timespec start_time;
>  };
>  
>  struct cgroup_subsys cpuacct_subsys;
> @@ -9172,6 +9173,8 @@ static struct cgroup_subsys_state *cpuacct_create(
>  		acct[CPUACCT_STAT_IDLE] = kstat_cpu(i).cpustat.idle;
>  		acct[CPUACCT_STAT_IOWAIT] = kstat_cpu(i).cpustat.iowait;
>  	}
> +
> +	get_monotonic_boottime(&ca->start_time);
>  	return &ca->css;
>  
>  out_free_usage:
> @@ -9316,13 +9319,16 @@ static int cpuacct_proc_stat(struct cgroup *cgrp, struct cftype *cft,
>  	unsigned int per_softirq_sums[NR_SOFTIRQS] = {0};
>  	struct timespec boottime;
>  	struct cpuacct *ca = cgroup_ca(cgrp);
> +	struct timespec ts;
>  	u64 *cpustat;
>  
>  	user = nice = system = idle = iowait =
>  		irq = softirq = steal = cputime64_zero;
>  	guest = guest_nice = cputime64_zero;
>  	getboottime(&boottime);
> -	jif = boottime.tv_sec;
> +	getboottime(&boottime);
> +	ts = timespec_add(boottime, ca->start_time);
> +	jif = ts.tv_sec;
>  
>  	for_each_possible_cpu(i) {
>  		cpustat = per_cpu_ptr(ca->cpustat, i);


I'm confused, what does it do? You take a boot time timestamp at cgroup
creation, add that to all boot-time readings and print the result. How
does that make sense? Subtracting the start_time, maybe, that would make
the cgroup creation time 0, adding, not so much.