[PATCH] mm/cgroup.c add error check

[PATCH] mm/cgroup.c add error check

[KOSAKI Motohiro]

on heavy workload, call_usermodehelper() may failure
because it use kzmalloc(GFP_ATOMIC).

but userland want receive release notificcation even heavy workload.

thus, We should retry if -ENOMEM happend.


Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
CC: "Paul Menage" <menage@google.com>
CC: Li Zefan <lizf@cn.fujitsu.com>

---
 kernel/cgroup.c |   10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

Index: b/kernel/cgroup.c
===================================================================
--- a/kernel/cgroup.c   2008-04-29 18:00:53.000000000 +0900
+++ b/kernel/cgroup.c   2008-05-06 20:28:23.000000000 +0900
@@ -3072,6 +3072,8 @@ void __css_put(struct cgroup_subsys_stat
  */
 static void cgroup_release_agent(struct work_struct *work)
 {
+       int err;
+
        BUG_ON(work != &release_agent_work);
        mutex_lock(&cgroup_mutex);
        spin_lock(&release_list_lock);
@@ -3111,7 +3113,13 @@ static void cgroup_release_agent(struct
                 * since the exec could involve hitting disk and hence
                 * be a slow process */
                mutex_unlock(&cgroup_mutex);
-               call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
+
+retry:
+               err = call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
+               if (err == -ENOMEM) {
+                       schedule();
+                       goto retry;
+               }
                kfree(pathbuf);
                mutex_lock(&cgroup_mutex);
                spin_lock(&release_list_lock);



--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] mm/cgroup.c add error check

[Andi Kleen]

KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> writes:

> on heavy workload, call_usermodehelper() may failure
> because it use kzmalloc(GFP_ATOMIC).

Better just fix it to not use GFP_ATOMIC in the first place.

-Andi

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] mm/cgroup.c add error check

[KOSAKI Motohiro]

>  > on heavy workload, call_usermodehelper() may failure
>  > because it use kzmalloc(GFP_ATOMIC).
>
>  Better just fix it to not use GFP_ATOMIC in the first place.

Ah, makes sense.
I'll try toi create that patch.

but if GFP_KERNEL is used, We still need error check, IMHO.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] mm/cgroup.c add error check

[Andi Kleen]

KOSAKI Motohiro wrote:
>>  > on heavy workload, call_usermodehelper() may failure
>>  > because it use kzmalloc(GFP_ATOMIC).
>>
>>  Better just fix it to not use GFP_ATOMIC in the first place.
> 
> Ah, makes sense.
> I'll try toi create that patch.

Thanks.

> 
> but if GFP_KERNEL is used, We still need error check, IMHO.

Yes, but no retry (or if you're sure you cannot fail use __GFP_NOFAIL
too, but that is nasty because it has some risk of deadlock under severe
oom conditions)

-Andi


--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] mm/cgroup.c add error check

[KOSAKI Motohiro]

>  > but if GFP_KERNEL is used, We still need error check, IMHO.
>
>  Yes, but no retry (or if you're sure you cannot fail use __GFP_NOFAIL
>  too, but that is nasty because it has some risk of deadlock under severe
>  oom conditions)

in general coding style, you are right.

but not down-to-earth idea in that case.
call_usermodehelper() is just wrapper of fork-exec.

I don't hope change largely exec() code patch.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] mm/cgroup.c add error check

[Paul Menage]

On Tue, May 6, 2008 at 4:02 AM, KOSAKI Motohiro
<kosaki.motohiro@jp.fujitsu.com> wrote:
>
>  on heavy workload, call_usermodehelper() may failure
>  because it use kzmalloc(GFP_ATOMIC).
>
>  but userland want receive release notificcation even heavy workload.
>
>  thus, We should retry if -ENOMEM happend.
>
>
>  Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
>  CC: "Paul Menage" <menage@google.com>
>  CC: Li Zefan <lizf@cn.fujitsu.com>
>
>  ---
>   kernel/cgroup.c |   10 +++++++++-
>   1 file changed, 9 insertions(+), 1 deletion(-)
>
>  Index: b/kernel/cgroup.c
>  ===================================================================
>  --- a/kernel/cgroup.c   2008-04-29 18:00:53.000000000 +0900
>  +++ b/kernel/cgroup.c   2008-05-06 20:28:23.000000000 +0900
>  @@ -3072,6 +3072,8 @@ void __css_put(struct cgroup_subsys_stat
>   */
>   static void cgroup_release_agent(struct work_struct *work)
>   {
>  +       int err;
>  +
>         BUG_ON(work != &release_agent_work);
>         mutex_lock(&cgroup_mutex);
>         spin_lock(&release_list_lock);
>  @@ -3111,7 +3113,13 @@ static void cgroup_release_agent(struct
>                  * since the exec could involve hitting disk and hence
>                  * be a slow process */
>                 mutex_unlock(&cgroup_mutex);
>  -               call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
>  +
>  +retry:
>  +               err = call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
>  +               if (err == -ENOMEM) {
>  +                       schedule();
>  +                       goto retry;
>  +               }

I'm not sure that an infinite loop retry is a great idea. Assuming
that call_usermodehelper() is changed to not use GFP_ATOMIC (which
should be safe at least in the case when UMG_WAIT_EXEC is set, since
we can clearly sleep in that case) then I'd be inclined not to check.
If we're so low on memory that we can't fork/exec, then the helper
binary itself could just as easily OOM as soon as we've launched it.
So there's still no guarantee. Userspace should just have to check for
empty cgroups occasionally even if they are using notify_on_release.

Paul

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] mm/cgroup.c add error check

[KAMEZAWA Hiroyuki]

On Tue, 6 May 2008 22:08:08 -0700
"Paul Menage" <menage@google.com> wrote:

> On Tue, May 6, 2008 at 4:02 AM, KOSAKI Motohiro
> <kosaki.motohiro@jp.fujitsu.com> wrote:
> >
> >  on heavy workload, call_usermodehelper() may failure
> >  because it use kzmalloc(GFP_ATOMIC).
> >
> >  but userland want receive release notificcation even heavy workload.
> >
> >  thus, We should retry if -ENOMEM happend.
> >
> >
> >  Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
> >  CC: "Paul Menage" <menage@google.com>
> >  CC: Li Zefan <lizf@cn.fujitsu.com>
> >
> >  ---
> >   kernel/cgroup.c |   10 +++++++++-
> >   1 file changed, 9 insertions(+), 1 deletion(-)
> >
> >  Index: b/kernel/cgroup.c
> >  ===================================================================
> >  --- a/kernel/cgroup.c   2008-04-29 18:00:53.000000000 +0900
> >  +++ b/kernel/cgroup.c   2008-05-06 20:28:23.000000000 +0900
> >  @@ -3072,6 +3072,8 @@ void __css_put(struct cgroup_subsys_stat
> >   */
> >   static void cgroup_release_agent(struct work_struct *work)
> >   {
> >  +       int err;
> >  +
> >         BUG_ON(work != &release_agent_work);
> >         mutex_lock(&cgroup_mutex);
> >         spin_lock(&release_list_lock);
> >  @@ -3111,7 +3113,13 @@ static void cgroup_release_agent(struct
> >                  * since the exec could involve hitting disk and hence
> >                  * be a slow process */
> >                 mutex_unlock(&cgroup_mutex);
> >  -               call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
> >  +
> >  +retry:
> >  +               err = call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
> >  +               if (err == -ENOMEM) {
> >  +                       schedule();
> >  +                       goto retry;
> >  +               }
> 
> I'm not sure that an infinite loop retry is a great idea. Assuming
> that call_usermodehelper() is changed to not use GFP_ATOMIC (which
> should be safe at least in the case when UMG_WAIT_EXEC is set, since
> we can clearly sleep in that case) then I'd be inclined not to check.
> If we're so low on memory that we can't fork/exec, then the helper
> binary itself could just as easily OOM as soon as we've launched it.
> So there's still no guarantee. Userspace should just have to check for
> empty cgroups occasionally even if they are using notify_on_release.
> 
Hmm, but It seems call_usermodehelper()'s fails in silent, no messages.
A notify which can be silently dropped is useless.
Can we add 'printk("notify_on_release: ... is failed")' or some workaround ?

Thanks,
-Kame



--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] mm/cgroup.c add error check

[Paul Menage]

On Wed, May 7, 2008 at 12:45 AM, KAMEZAWA Hiroyuki
<kamezawa.hiroyu@jp.fujitsu.com> wrote:
>  Hmm, but It seems call_usermodehelper()'s fails in silent, no messages.
>  A notify which can be silently dropped is useless.

True, but if we're so short of memory that we can't fork, userspace
probably won't be able to do much about the notification even if it
gets it.

>  Can we add 'printk("notify_on_release: ... is failed")' or some workaround ?
>

Sounds reasonable for debugging, but don't expect any automated
middleware to notice it.

Paul

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>