Re: [patch] oom-killer sysrq-f fix

[Andrew Morton <akpm@osdl.org>]

Coywolf Qi Hunt <coywolf@lovecn.org> wrote:
>
> >>--- 2.6.12-rc1-mm2/mm/oom_kill.c	2005-03-03 17:12:18.000000000 +0800
> >>+++ 2.6.12-rc1-mm2-cy/mm/oom_kill.c	2005-03-25 08:07:19.000000000 +0800
> >>@@ -20,6 +20,7 @@
> >> #include <linux/swap.h>
> >> #include <linux/timex.h>
> >> #include <linux/jiffies.h>
> >>+#include <linux/hardirq.h>
> >> 
> >> /* #define DEBUG */
> >> 
> >>@@ -283,6 +284,9 @@ retry:
> >> 	if (mm)
> >> 		mmput(mm);
> >> 
> >>+	if (in_interrupt())
> >>+		return;
> > 
> > 
> > That'll make the whole feature a no-op, won't it?
> 
> It won't be a no-op. I have tested it. It works well.
> I pressed sysrq-f, loging bash got killed and I had to re-login.

(looks)

OK.  But the patch is still deadlocky because we do task_lock() from
interrupt context.

> > 
> > The thing needs to be moved into process context via schedule_work().  I
> > haven't got around to it yet.
> > 
> 
> I don't think schedule_work() is a good option here.  Since sysrq-f is emergent,
> we just let oom-killer send SIGKILL in interrupt context and return.
> 
> We needn't send SIGKILL in a process context. That'll be slow and [events] may got delayed.

There isn't much choice.  It should work OK - schedule_task doesn't
allocate any memory.

keventd could be off allocating some memory somewhere, in which case it
could take some time to respond, but it isn't worth running another kernel
thread for sysrq-f.