Re: [PATCH 14/16] freezer: make freezing() test freeze conditions in effect instead of TIF_FREEZE

[Oleg Nesterov <oleg@redhat.com>]

On 08/19, Tejun Heo wrote:
>
> @@ -311,9 +315,11 @@ static int freezer_change_state(struct cgroup *cgroup,
>  
>  	switch (goal_state) {
>  	case CGROUP_THAWED:
> +		atomic_dec(&system_freezing_cnt);
>  		unfreeze_cgroup(cgroup, freezer);
>  		break;
>  	case CGROUP_FROZEN:
> +		atomic_inc(&system_freezing_cnt);

This is harmless, but afaics is not exactly right. CGROUP_FROZEN doesn't
need system_freezing_cnt != 0, everything is already frozen and we just
provoke freezing_slow_path() without any reason. Right?


> @@ -120,13 +120,18 @@ int freeze_processes(void)
>  {
>  	int error;
>
> +	if (!pm_freezing)
> +		atomic_inc(&system_freezing_cnt);
> +
>  	printk("Freezing user space processes ... ");
> +	pm_freezing = true;

and

> @@ -146,6 +151,11 @@ void thaw_processes(void)
>  {
>  	struct task_struct *g, *p;
>
> +	if (pm_freezing)
> +		atomic_dec(&system_freezing_cnt);
> +	pm_freezing = false;

I simply can't understand this... Why freeze_processes/thaw_processes
check pm_freezing?

IIUC, the calls to freeze/thaw should be serialized anyway (probably
pm_mutex ?). Otherwise this check can't help anyway. Say, _if_ it
is possible to call freeze_processes() with pm_freezing == T, then
the failure path or subsequent thaw_processes() will do the unbalanced
atomic_dec().

Oleg.