Re: [RFC PATCH 2/2] Add per-cgroup CPU controller statistics

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

On Wed, 2009-02-25 at 16:29 +0530, Bharata B Rao wrote:
> From: Balaji Rao <balajirrao@gmail.com>
> 
> sched: Add cpu controller statistics
> 
> Add per-cgroup cpu controller statistics like system time and user time
> consumed by groups of tasks.

Do we want this unconditionally?

> Signed-off-by: Balaji Rao <balajirrao@gmail.com>
> Signed-off-by: Bharata B Rao <bharata@linux.vnet.ibm.com>
> ---
>  kernel/sched.c |  106 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 106 insertions(+)
> 
> --- a/kernel/sched.c
> +++ b/kernel/sched.c
> @@ -263,10 +263,25 @@ struct cfs_rq;
>  
>  static LIST_HEAD(task_groups);
>  
> +#ifdef CONFIG_CGROUP_SCHED
> +enum cpu_cgroup_stat_index {
> +	CPU_CGROUP_STAT_UTIME, /* User time of the task group */
> +	CPU_CGROUP_STAT_STIME, /* Kernel time of the task group */
> +	CPU_CGROUP_STAT_GTIME, /* Guest time of the task group */
> +
> +	CPU_CGROUP_STAT_NSTATS,
> +};
> +
> +struct cpu_cgroup_stat {
> +	struct percpu_counter cpustat[CPU_CGROUP_STAT_NSTATS];
> +};
> +#endif
> +
>  /* task group related information */
>  struct task_group {
>  #ifdef CONFIG_CGROUP_SCHED
>  	struct cgroup_subsys_state css;
> +	struct cpu_cgroup_stat *stat;
>  #endif
>  
>  #ifdef CONFIG_USER_SCHED
> @@ -4159,6 +4174,19 @@ unsigned long long task_delta_exec(struc
>  	return ns;
>  }
>  
> +#ifdef CONFIG_CGROUP_SCHED
> +static void account_task_group_time(struct task_struct *p,
> +			enum cpu_cgroup_stat_index idx, int val)
> +{
> +	struct task_group *tg = task_group(p);
> +
> +	if (likely(tg->stat))
> +		percpu_counter_add(&tg->stat->cpustat[idx], val);
> +}
> +#else
> +#define account_task_group_time(x, y, z)	{ 0; }

inline please, so we get argument validation.

> +#endif
> +
>  /*
>   * Account user cpu time to a process.
>   * @p: the process that the cpu time gets accounted to
> @@ -4182,6 +4210,10 @@ void account_user_time(struct task_struc
>  		cpustat->nice = cputime64_add(cpustat->nice, tmp);
>  	else
>  		cpustat->user = cputime64_add(cpustat->user, tmp);
> +
> +	account_task_group_time(p, CPU_CGROUP_STAT_UTIME,
> +			cputime_to_msecs(cputime));
> +
>  	/* Account for user time used */
>  	acct_update_integrals(p);
>  }
> @@ -4206,6 +4238,9 @@ static void account_guest_time(struct ta
>  	account_group_user_time(p, cputime);
>  	p->gtime = cputime_add(p->gtime, cputime);
>  
> +	account_task_group_time(p, CPU_CGROUP_STAT_GTIME,
> +			cputime_to_msecs(cputime));
> +
>  	/* Add guest time to cpustat. */
>  	cpustat->user = cputime64_add(cpustat->user, tmp);
>  	cpustat->guest = cputime64_add(cpustat->guest, tmp);
> @@ -4243,6 +4278,8 @@ void account_system_time(struct task_str
>  	else
>  		cpustat->system = cputime64_add(cpustat->system, tmp);
>  
> +	account_task_group_time(p, CPU_CGROUP_STAT_STIME,
> +				cputime_to_msecs(cputime));
>  	/* Account for system time used */
>  	acct_update_integrals(p);
>  }
> @@ -9290,6 +9327,7 @@ static struct cgroup_subsys_state *
>  cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
>  {
>  	struct task_group *tg, *parent;
> +	int i;
>  
>  	if (!cgrp->parent) {
>  		/* This is early initialization for the top cgroup */
> @@ -9301,6 +9339,15 @@ cpu_cgroup_create(struct cgroup_subsys *
>  	if (IS_ERR(tg))
>  		return ERR_PTR(-ENOMEM);
>  
> +	tg->stat = kmalloc(sizeof(struct cpu_cgroup_stat), GFP_KERNEL);
> +	if (!tg->stat) {
> +		sched_destroy_group(tg);
> +		return ERR_PTR(-ENOMEM);
> +	}
> +
> +	for (i = 0; i < CPU_CGROUP_STAT_NSTATS; i++)
> +		percpu_counter_init(&tg->stat->cpustat[i], 0);

percpu_counter_init() can fail with -ENOMEM.

>  	return &tg->css;
>  }
>  
> @@ -9308,6 +9355,13 @@ static void
>  cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
>  {
>  	struct task_group *tg = cgroup_tg(cgrp);
> +	int i;
> +
> +	if (tg->stat) {
> +		for (i = 0; i < CPU_CGROUP_STAT_NSTATS; i++)
> +			percpu_counter_destroy(&tg->stat->cpustat[i]);
> +		kfree(tg->stat);
> +	}
>  
>  	sched_destroy_group(tg);
>  }
> @@ -9336,6 +9390,53 @@ cpu_cgroup_attach(struct cgroup_subsys *
>  	sched_move_task(tsk);
>  }
>  
> +static void cpu_cgroup_initialize(void)
> +{
> +	int i;
> +	struct cpu_cgroup_stat *stat;
> +
> +	stat = kmalloc(sizeof(struct cpu_cgroup_stat) , GFP_KERNEL);
> +	if (!stat)
> +		return;
> +
> +	for (i = 0; i < CPU_CGROUP_STAT_NSTATS; i++)
> +		percpu_counter_init(&stat->cpustat[i], 0);
> +	init_task_group.stat = stat;
> +}
> +
> +static s64 cpu_cgroup_read_stat(struct cpu_cgroup_stat *stat,
> +		enum cpu_cgroup_stat_index idx)
> +{
> +	if (stat)
> +		return percpu_counter_read(&stat->cpustat[idx]);
> +	return 0;
> +}
> +
> +static const struct cpu_cgroup_stat_desc {
> +	const char *msg;
> +	u64 unit;
> +} cpu_cgroup_stat_desc[] = {
> +	[CPU_CGROUP_STAT_UTIME] = { "utime", 1, },
> +	[CPU_CGROUP_STAT_STIME] = { "stime", 1, },
> +	[CPU_CGROUP_STAT_GTIME] = { "gtime", 1, },
> +};
> +
> +static int cpu_cgroup_stats_show(struct cgroup *cgrp, struct cftype *cft,
> +		struct cgroup_map_cb *cb)
> +{
> +	struct task_group *tg = cgroup_tg(cgrp);
> +	struct cpu_cgroup_stat *stat = tg->stat;
> +	int i;
> +
> +	for (i = 0; i < CPU_CGROUP_STAT_NSTATS; i++) {
> +		s64 val;
> +		val = cpu_cgroup_read_stat(stat, i);
> +		val *= cpu_cgroup_stat_desc[i].unit;
> +		cb->fill(cb, cpu_cgroup_stat_desc[i].msg, val);
> +	}
> +	return 0;
> +}
> +
>  #ifdef CONFIG_FAIR_GROUP_SCHED
>  static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
>  				u64 shareval)
> @@ -9395,6 +9496,10 @@ static struct cftype cpu_files[] = {
>  		.write_u64 = cpu_rt_period_write_uint,
>  	},
>  #endif
> +	{
> +		.name = "stat",
> +		.read_map = cpu_cgroup_stats_show,
> +	},
>  };
>  
>  static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
> @@ -9409,6 +9514,7 @@ struct cgroup_subsys cpu_cgroup_subsys =
>  	.can_attach	= cpu_cgroup_can_attach,
>  	.attach		= cpu_cgroup_attach,
>  	.populate	= cpu_cgroup_populate,
> +	.initialize	= cpu_cgroup_initialize,
>  	.subsys_id	= cpu_cgroup_subsys_id,
>  	.early_init	= 1,
>  };