lkml.org 
[lkml]   [2013]   [Nov]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v4] mm, oom: Fix race when selecting process to kill
On Mon 11-11-13 16:21:57, Sameer Nanda wrote:
> The selection of the process to be killed happens in two spots:
> first in select_bad_process and then a further refinement by
> looking for child processes in oom_kill_process. Since this is
> a two step process, it is possible that the process selected by
> select_bad_process may get a SIGKILL just before oom_kill_process
> executes. If this were to happen, __unhash_process deletes this
> process from the thread_group list. This results in oom_kill_process
> getting stuck in an infinite loop when traversing the thread_group
> list of the selected process.
>
> Fix this race by adding a pid_alive check for the selected process
> with tasklist_lock held in oom_kill_process.
>
> Signed-off-by: Sameer Nanda <snanda@chromium.org>
> ---
> mm/oom_kill.c | 24 +++++++++++++++++++-----
> 1 file changed, 19 insertions(+), 5 deletions(-)
>
> diff --git a/mm/oom_kill.c b/mm/oom_kill.c
> index 6738c47..57638ef 100644
> --- a/mm/oom_kill.c
> +++ b/mm/oom_kill.c
> @@ -413,12 +413,20 @@ void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
> DEFAULT_RATELIMIT_BURST);
>
> /*
> + * while_each_thread is currently not RCU safe. Lets hold the
> + * tasklist_lock across all invocations of while_each_thread (including
> + * the one in find_lock_task_mm) in this function.
> + */
> + read_lock(&tasklist_lock);
> +
> + /*
> * If the task is already exiting, don't alarm the sysadmin or kill
> * its children or threads, just set TIF_MEMDIE so it can die quickly
> */
> - if (p->flags & PF_EXITING) {
> + if (p->flags & PF_EXITING || !pid_alive(p)) {
> set_tsk_thread_flag(p, TIF_MEMDIE);
> put_task_struct(p);
> + read_unlock(&tasklist_lock);
> return;
> }

show_mem used to be one of a bottleneck but now that we have Mel's "mm:
do not walk all of system memory during show_mem" it shouldn't be a big
deal anymore.
The real trouble is with dump_tasks which might be zillions of tasks and
we do not want to hold tasklist_lock for that long.

So no this would regress on the huge machines and yes we have seen
reports like that and explicit requests to backport 6b0c81b3be114 (mm,
oom: reduce dependency on tasklist_lock) so this would be a step
backwards although I see there is a real problem that it tries to fix.

--
Michal Hocko
SUSE Labs


\
 
 \ /
  Last update: 2013-11-12 18:21    [W:0.125 / U:1.744 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site