2005-04-01 22:32:54

by Andrea Arcangeli

Subject: oom-killer disable for iscsi/lvm2/multipath userland critical sections

Hello,

some private discussion (that was continuing some kernel-summit-discuss
thread) ended in the below patch. I also liked a textual "disable"
instead of value "-17" (internally to the kernel it could be represented
the same way, but the /proc parsing would be more complicated). If you
prefer textual "disable" we can change this of course.

Comments welcome.

From: Andrea Arcangeli <[email protected]>
Subject: oom killer protection

iscsi/lvm2/multipath needs guaranteed protection from the oom-killer.

Signed-off-by: Andrea Arcangeli <[email protected]>

--- 2.6.12-seccomp/fs/proc/base.c.~1~ 2005-03-25 05:13:28.000000000 +0100
+++ 2.6.12-seccomp/fs/proc/base.c 2005-04-01 23:47:22.000000000 +0200
@@ -751,7 +751,7 @@ static ssize_t oom_adjust_write(struct f
if (copy_from_user(buffer, buf, count))
return -EFAULT;
oom_adjust = simple_strtol(buffer, &end, 0);
- if (oom_adjust < -16 || oom_adjust > 15)
+ if ((oom_adjust < -16 || oom_adjust > 15) && oom_adjust != OOM_DISABLE)
return -EINVAL;
if (*end == '\n')
end++;
--- 2.6.12-seccomp/include/linux/mm.h.~1~ 2005-03-25 05:13:28.000000000 +0100
+++ 2.6.12-seccomp/include/linux/mm.h 2005-04-01 23:53:11.000000000 +0200
@@ -856,5 +856,8 @@ int in_gate_area_no_task(unsigned long a
#define in_gate_area(task, addr) ({(void)task; in_gate_area_no_task(addr);})
#endif /* __HAVE_ARCH_GATE_AREA */

+/* /proc/<pid>/oom_adj set to -17 protects from the oom-killer */
+#define OOM_DISABLE -17
+
#endif /* __KERNEL__ */
#endif /* _LINUX_MM_H */
--- 2.6.12-seccomp/mm/oom_kill.c.~1~ 2005-03-08 01:02:30.000000000 +0100
+++ 2.6.12-seccomp/mm/oom_kill.c 2005-04-01 23:46:18.000000000 +0200
@@ -145,7 +145,7 @@ static struct task_struct * select_bad_p
do_posix_clock_monotonic_gettime(&uptime);
do_each_thread(g, p)
/* skip the init task with pid == 1 */
- if (p->pid > 1) {
+ if (p->pid > 1 && p->oomkilladj != OOM_DISABLE) {
unsigned long points;

/*

2005-04-02 00:20:34

by Dmitry Yusupov

Subject: Re: oom-killer disable for iscsi/lvm2/multipath userland critical sections

Andrea,

I just successfully tested the patch on my environment. It actually
resolved OOM-killer problem for my iscsid.

Important note: daemon's parent must be init.

In my test, OOM-killer killed everything around but iscsid, and iscsid
successfully finished registration of new SCSI host in the middle of
crazy OOM-killer :)

Thanks!
Dima

On Sat, 2005-04-02 at 00:14 +0200, Andrea Arcangeli wrote:
> Hello,
>
> some private discussion (that was continuing some kernel-summit-discuss
> thread) ended in the below patch. I also liked a textual "disable"
> instead of value "-17" (internally to the kernel it could be represented
> the same way, but the /proc parsing would be more complicated). If you
> prefer textual "disable" we can change this of course.
>
> Comments welcome.
>
> From: Andrea Arcangeli <[email protected]>
> Subject: oom killer protection
>
> iscsi/lvm2/multipath needs guaranteed protection from the oom-killer.
>
> Signed-off-by: Andrea Arcangeli <[email protected]>
>
> --- 2.6.12-seccomp/fs/proc/base.c.~1~ 2005-03-25 05:13:28.000000000 +0100
> +++ 2.6.12-seccomp/fs/proc/base.c 2005-04-01 23:47:22.000000000 +0200
> @@ -751,7 +751,7 @@ static ssize_t oom_adjust_write(struct f
> if (copy_from_user(buffer, buf, count))
> return -EFAULT;
> oom_adjust = simple_strtol(buffer, &end, 0);
> - if (oom_adjust < -16 || oom_adjust > 15)
> + if ((oom_adjust < -16 || oom_adjust > 15) && oom_adjust != OOM_DISABLE)
> return -EINVAL;
> if (*end == '\n')
> end++;
> --- 2.6.12-seccomp/include/linux/mm.h.~1~ 2005-03-25 05:13:28.000000000 +0100
> +++ 2.6.12-seccomp/include/linux/mm.h 2005-04-01 23:53:11.000000000 +0200
> @@ -856,5 +856,8 @@ int in_gate_area_no_task(unsigned long a
> #define in_gate_area(task, addr) ({(void)task; in_gate_area_no_task(addr);})
> #endif /* __HAVE_ARCH_GATE_AREA */
>
> +/* /proc/<pid>/oom_adj set to -17 protects from the oom-killer */
> +#define OOM_DISABLE -17
> +
> #endif /* __KERNEL__ */
> #endif /* _LINUX_MM_H */
> --- 2.6.12-seccomp/mm/oom_kill.c.~1~ 2005-03-08 01:02:30.000000000 +0100
> +++ 2.6.12-seccomp/mm/oom_kill.c 2005-04-01 23:46:18.000000000 +0200
> @@ -145,7 +145,7 @@ static struct task_struct * select_bad_p
> do_posix_clock_monotonic_gettime(&uptime);
> do_each_thread(g, p)
> /* skip the init task with pid == 1 */
> - if (p->pid > 1) {
> + if (p->pid > 1 && p->oomkilladj != OOM_DISABLE) {
> unsigned long points;
>
> /*
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/