2006-08-27 14:01:26

by Oleg Nesterov

Subject: [PATCH -mm] select_bad_process: cleanup 'releasing' check

On top of "select_bad_process: kill a bogus PF_DEAD/TASK_DEAD check"

No logic changes, but imho easier to read.

Signed-off-by: Oleg Nesterov <[email protected]>

--- 2.6.18-rc4/mm/oom_kill.c~ 2006-08-27 20:56:23.000000000 +0400
+++ 2.6.18-rc4/mm/oom_kill.c 2006-08-27 21:58:32.000000000 +0400
@@ -205,7 +205,6 @@ static struct task_struct *select_bad_pr
do_posix_clock_monotonic_gettime(&uptime);
do_each_thread(g, p) {
unsigned long points;
- int releasing;

/*
* skip kernel threads and tasks which have already released
@@ -227,16 +226,15 @@ static struct task_struct *select_bad_pr
* the process of exiting and releasing its resources.
* Otherwise we could get an OOM deadlock.
*/
- releasing = test_tsk_thread_flag(p, TIF_MEMDIE) ||
- p->flags & PF_EXITING;
- if (releasing) {
- if (p->flags & PF_EXITING && p == current) {
- chosen = p;
- *ppoints = ULONG_MAX;
- break;
- }
- return ERR_PTR(-1UL);
- }
+ if ((p->flags & PF_EXITING) && p == current) {
+ chosen = p;
+ *ppoints = ULONG_MAX;
+ break;
+ }
+ if ((p->flags & PF_EXITING) ||
+ test_tsk_thread_flag(p, TIF_MEMDIE))
+ return ERR_PTR(-1UL);
+
if (p->oomkilladj == OOM_DISABLE)
continue;

@@ -246,6 +244,7 @@ static struct task_struct *select_bad_pr
*ppoints = points;
}
} while_each_thread(g, p);
+
return chosen;
}

2006-08-28 10:45:04

by Nick Piggin

Subject: Re: [PATCH -mm] select_bad_process: cleanup 'releasing' check

On Sun, Aug 27, 2006 at 10:25:38PM +0400, Oleg Nesterov wrote:
> On top of "select_bad_process: kill a bogus PF_DEAD/TASK_DEAD check"
>
> No logic changes, but imho easier to read.
>
> Signed-off-by: Oleg Nesterov <[email protected]>
>
> --- 2.6.18-rc4/mm/oom_kill.c~ 2006-08-27 20:56:23.000000000 +0400
> +++ 2.6.18-rc4/mm/oom_kill.c 2006-08-27 21:58:32.000000000 +0400
> @@ -205,7 +205,6 @@ static struct task_struct *select_bad_pr
> do_posix_clock_monotonic_gettime(&uptime);
> do_each_thread(g, p) {
> unsigned long points;
> - int releasing;
>
> /*
> * skip kernel threads and tasks which have already released
> @@ -227,16 +226,15 @@ static struct task_struct *select_bad_pr
> * the process of exiting and releasing its resources.
> * Otherwise we could get an OOM deadlock.
> */
> - releasing = test_tsk_thread_flag(p, TIF_MEMDIE) ||
> - p->flags & PF_EXITING;
> - if (releasing) {
> - if (p->flags & PF_EXITING && p == current) {
> - chosen = p;
> - *ppoints = ULONG_MAX;
> - break;
> - }
> - return ERR_PTR(-1UL);
> - }
> + if ((p->flags & PF_EXITING) && p == current) {
> + chosen = p;
> + *ppoints = ULONG_MAX;
> + break;
> + }
> + if ((p->flags & PF_EXITING) ||
> + test_tsk_thread_flag(p, TIF_MEMDIE))
> + return ERR_PTR(-1UL);
> +

Hmm, actually I think I spot a bug in the original logic: we don't want
to have more than 1 task with TIF_MEMDIE at once, becaues that gives it
access to memory reserves (but I saw it first in the new formulation, so
maybe that does suggest it is more readable ;)

What I think should be done is the check for TIF_MEMDIE (and return -1)
first, and then the PF_EXITING test. At which point, if current is found to
be exiting, it should be chosen but not break... that way a subsequent MEMDIE
or EXITING task has the chance to trigger the -1 return.

Anyway, if you don't want to do all that, I will when my hand gets better.
Otherwise the 3 patches you sent look good, they could all have an

Acked-by: Nick Piggin <[email protected]>

Thanks,
Nick

2006-08-28 11:00:19

by Oleg Nesterov

Subject: Re: [PATCH -mm] select_bad_process: cleanup 'releasing' check

On 08/28, Nick Piggin wrote:
>
> On Sun, Aug 27, 2006 at 10:25:38PM +0400, Oleg Nesterov wrote:
> >
> > - releasing = test_tsk_thread_flag(p, TIF_MEMDIE) ||
> > - p->flags & PF_EXITING;
> > - if (releasing) {
> > - if (p->flags & PF_EXITING && p == current) {
> > - chosen = p;
> > - *ppoints = ULONG_MAX;
> > - break;
> > - }
> > - return ERR_PTR(-1UL);
> > - }
> > + if ((p->flags & PF_EXITING) && p == current) {
> > + chosen = p;
> > + *ppoints = ULONG_MAX;
> > + break;
> > + }
> > + if ((p->flags & PF_EXITING) ||
> > + test_tsk_thread_flag(p, TIF_MEMDIE))
> > + return ERR_PTR(-1UL);
> > +
>
> Hmm, actually I think I spot a bug in the original logic: we don't want
> to have more than 1 task with TIF_MEMDIE at once, becaues that gives it
> access to memory reserves (but I saw it first in the new formulation, so
> maybe that does suggest it is more readable ;)
>
> What I think should be done is the check for TIF_MEMDIE (and return -1)
> first, and then the PF_EXITING test. At which point, if current is found to
> be exiting, it should be chosen but not break... that way a subsequent MEMDIE
> or EXITING task has the chance to trigger the -1 return.

Aha! The logic looked somewhat strange to me, but ...

> Anyway, if you don't want to do all that, I will when my hand gets better.

I have little understanding of this magic, i'd better not to try to fix it.

> Otherwise the 3 patches you sent look good, they could all have an
>
> Acked-by: Nick Piggin <[email protected]>

Thanks!

Oleg.