Re: [PATCH v3] mm/oom_kill: show oom eligibility when displaying the current memory state of all tasks

[Andrew Morton]

On Fri, 30 Jul 2021 17:20:02 +0100 Aaron Tomlin <atomlin@xxxxxxxxxx> wrote:

> Changes since v2:
>  - Use single character (e.g. 'R' for MMF_OOM_SKIP) as suggested
>    by Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx>
>  - Add new header to oom_dump_tasks documentation
>  - Provide further justification
> 
> 
> The output generated by dump_tasks() can be helpful to determine why
> there was an OOM condition and which rogue task potentially caused it.
> Please note that this is only provided when sysctl oom_dump_tasks is
> enabled.
> 
> At the present time, when showing potential OOM victims, we do not
> exclude any task that are not OOM eligible e.g. those that have
> MMF_OOM_SKIP set; it is possible that the last OOM killable victim was
> already OOM killed, yet the OOM reaper failed to reclaim memory and set
> MMF_OOM_SKIP. This can be confusing (or perhaps even be misleading) to the
> viewer. Now, we already unconditionally display a task's oom_score_adj_min
> value that can be set to OOM_SCORE_ADJ_MIN which is indicative of an
> "unkillable" task.
> 
> This patch provides a clear indication with regard to the OOM ineligibility
> (and why) of each displayed task with the addition of a new column namely
> "oom_skipped". An example is provided below:
> 
>     [ 5084.524970] [ pid ]   uid  tgid total_vm      rss pgtables_bytes swapents oom_score_adj oom_skipped name
>     [ 5084.526397] [660417]     0 660417    35869      683   167936        0         -1000 M conmon
>     [ 5084.526400] [660452]     0 660452   175834      472    86016        0          -998  pod
>     [ 5084.527460] [752415]     0 752415    35869      650   172032        0         -1000 M conmon
>     [ 5084.527462] [752575] 1001050000 752575   184205    11158   700416        0           999  npm
>     [ 5084.527467] [753606] 1001050000 753606   183380    46843  2134016        0           999  node
>     [ 5084.527581] Memory cgroup out of memory: Killed process 753606 (node) total-vm:733520kB, anon-rss:161228kB, file-rss:26144kB, shmem-rss:0kB, UID:1001050000
> 
> So, a single character 'M' is for OOM_SCORE_ADJ_MIN, 'R' MMF_OOM_SKIP and
> 'V' for in_vfork().
> 
> index 003d5cc3751b..4c79fa00ddb3 100644
> --- a/Documentation/admin-guide/sysctl/vm.rst
> +++ b/Documentation/admin-guide/sysctl/vm.rst
> @@ -650,8 +650,9 @@ oom_dump_tasks
>  Enables a system-wide task dump (excluding kernel threads) to be produced
>  when the kernel performs an OOM-killing and includes such information as
>  pid, uid, tgid, vm size, rss, pgtables_bytes, swapents, oom_score_adj
> -score, and name.  This is helpful to determine why the OOM killer was
> -invoked, to identify the rogue task that caused it, and to determine why
> +score, oom eligibility status and name.  This is helpful to determine why
> +the OOM killer was invoked, to identify the rogue task that caused it, and
> +to determine why

It would be better if the meaning of 'M', 'R' and 'V' were described here.

>  the OOM killer chose the task it did to kill.
>  
> +/**
> + * is_task_eligible_oom - determine if and why a task cannot be OOM killed
> + * @tsk: task to check
> + *
> + * Needs to be called with task_lock().
> + */
> +static const char * const is_task_oom_eligible(struct task_struct *p)

Name seems inappropriate.  task_oom_eligibility()?

> +{
> +	long adj;
> +
> +	adj = (long)p->signal->oom_score_adj;
> +	if (adj == OOM_SCORE_ADJ_MIN)
> +		return "M";
> +	else if (test_bit(MMF_OOM_SKIP, &p->mm->flags)
> +		return "R";
> +	else if (in_vfork(p))
> +		return "V";
> +	else
> +		return "";
> +}