| Message ID | 20210701125430.836308-1-minyard@acm.org (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | oom_kill: oom_score_adj broken for processes with small memory usage | expand |
On Thu 01-07-21 07:54:30, minyard@acm.org wrote: > From: Corey Minyard <cminyard@mvista.com> > > If you have a process with less than 1000 totalpages, the calculation: > > adj = (long)p->signal->oom_score_adj; > ... > adj *= totalpages / 1000; > > will always result in adj being zero no matter what oom_score_adj is, > which could result in the wrong process being picked for killing. > > Fix by adding 1000 to totalpages before dividing. Yes, this is a known limitation of the oom_score_adj and its scale. Is this a practical problem to be solved though? I mean 0-1000 pages is not really that much different from imprecision at a larger scale where tasks are effectively considered equal. I have to say I do not really like the proposed workaround. It doesn't really solve the problem yet it adds another special case.
On Fri, Jul 16, 2021 at 07:19:24AM +0200, Michal Hocko wrote: > On Thu 01-07-21 07:54:30, minyard@acm.org wrote: > > From: Corey Minyard <cminyard@mvista.com> > > > > If you have a process with less than 1000 totalpages, the calculation: > > > > adj = (long)p->signal->oom_score_adj; > > ... > > adj *= totalpages / 1000; > > > > will always result in adj being zero no matter what oom_score_adj is, > > which could result in the wrong process being picked for killing. > > > > Fix by adding 1000 to totalpages before dividing. > > Yes, this is a known limitation of the oom_score_adj and its scale. > Is this a practical problem to be solved though? I mean 0-1000 pages is > not really that much different from imprecision at a larger scale where > tasks are effectively considered equal. Known limitation? Is this documented? I couldn't find anything that said "oom_score_adj doesn't work at all with programs with <1000 pages besides setting the value to -1000". > > I have to say I do not really like the proposed workaround. It doesn't > really solve the problem yet it adds another special case. The problem is that if you have a small program, there is no way to set it's priority besides completely disablling the OOM killer for it. I don't understand the special case comment. How is this adding a special case? This patch removes a special case. Small programs working different than big programs is a special case. Making them all work the same is removing an element of surprise from someone expecting things to work as documented. -corey
On Fri, 16 Jul 2021 07:25:47 -0500 Corey Minyard <cminyard@mvista.com> wrote: > On Fri, Jul 16, 2021 at 07:19:24AM +0200, Michal Hocko wrote: > > On Thu 01-07-21 07:54:30, minyard@acm.org wrote: > > > From: Corey Minyard <cminyard@mvista.com> > > > > > > If you have a process with less than 1000 totalpages, the calculation: > > > > > > adj = (long)p->signal->oom_score_adj; > > > ... > > > adj *= totalpages / 1000; > > > > > > will always result in adj being zero no matter what oom_score_adj is, > > > which could result in the wrong process being picked for killing. > > > > > > Fix by adding 1000 to totalpages before dividing. > > > > Yes, this is a known limitation of the oom_score_adj and its scale. > > Is this a practical problem to be solved though? I mean 0-1000 pages is > > not really that much different from imprecision at a larger scale where > > tasks are effectively considered equal. > > Known limitation? Is this documented? I couldn't find anything that > said "oom_score_adj doesn't work at all with programs with <1000 pages > besides setting the value to -1000". > > > > > I have to say I do not really like the proposed workaround. It doesn't > > really solve the problem yet it adds another special case. > > The problem is that if you have a small program, there is no way to > set it's priority besides completely disablling the OOM killer for > it. > > I don't understand the special case comment. How is this adding a > special case? This patch removes a special case. Small programs > working different than big programs is a special case. Making them all > work the same is removing an element of surprise from someone expecting > things to work as documented. > Can we please get this resolved one way or the other?
On Thu, Sep 02, 2021 at 12:55:01PM -0700, Andrew Morton wrote: > On Fri, 16 Jul 2021 07:25:47 -0500 Corey Minyard <cminyard@mvista.com> wrote: > > > On Fri, Jul 16, 2021 at 07:19:24AM +0200, Michal Hocko wrote: > > > On Thu 01-07-21 07:54:30, minyard@acm.org wrote: > > > > From: Corey Minyard <cminyard@mvista.com> > > > > > > > > If you have a process with less than 1000 totalpages, the calculation: > > > > > > > > adj = (long)p->signal->oom_score_adj; > > > > ... > > > > adj *= totalpages / 1000; > > > > > > > > will always result in adj being zero no matter what oom_score_adj is, > > > > which could result in the wrong process being picked for killing. > > > > > > > > Fix by adding 1000 to totalpages before dividing. > > > > > > Yes, this is a known limitation of the oom_score_adj and its scale. > > > Is this a practical problem to be solved though? I mean 0-1000 pages is > > > not really that much different from imprecision at a larger scale where > > > tasks are effectively considered equal. > > > > Known limitation? Is this documented? I couldn't find anything that > > said "oom_score_adj doesn't work at all with programs with <1000 pages > > besides setting the value to -1000". > > > > > > > > I have to say I do not really like the proposed workaround. It doesn't > > > really solve the problem yet it adds another special case. > > > > The problem is that if you have a small program, there is no way to > > set it's priority besides completely disablling the OOM killer for > > it. > > > > I don't understand the special case comment. How is this adding a > > special case? This patch removes a special case. Small programs > > working different than big programs is a special case. Making them all > > work the same is removing an element of surprise from someone expecting > > things to work as documented. > > > > Can we please get this resolved one way or the other? My goal in submitting this is to avoid someone having to go through what I went through. I know it now, so it's not going to affect me again. We could document this, but to me it seems silly when something can just be made consistent to avoid having to document it. I got no response to my questions above, so I don't know what to make of it. Thanks Andrew, -corey
On Thu 02-09-21 12:55:01, Andrew Morton wrote: > On Fri, 16 Jul 2021 07:25:47 -0500 Corey Minyard <cminyard@mvista.com> wrote: > > > On Fri, Jul 16, 2021 at 07:19:24AM +0200, Michal Hocko wrote: > > > On Thu 01-07-21 07:54:30, minyard@acm.org wrote: > > > > From: Corey Minyard <cminyard@mvista.com> > > > > > > > > If you have a process with less than 1000 totalpages, the calculation: > > > > > > > > adj = (long)p->signal->oom_score_adj; > > > > ... > > > > adj *= totalpages / 1000; > > > > > > > > will always result in adj being zero no matter what oom_score_adj is, > > > > which could result in the wrong process being picked for killing. > > > > > > > > Fix by adding 1000 to totalpages before dividing. > > > > > > Yes, this is a known limitation of the oom_score_adj and its scale. > > > Is this a practical problem to be solved though? I mean 0-1000 pages is > > > not really that much different from imprecision at a larger scale where > > > tasks are effectively considered equal. > > > > Known limitation? Is this documented? I couldn't find anything that > > said "oom_score_adj doesn't work at all with programs with <1000 pages > > besides setting the value to -1000". > > > > > > > > I have to say I do not really like the proposed workaround. It doesn't > > > really solve the problem yet it adds another special case. > > > > The problem is that if you have a small program, there is no way to > > set it's priority besides completely disablling the OOM killer for > > it. > > > > I don't understand the special case comment. How is this adding a > > special case? This patch removes a special case. Small programs > > working different than big programs is a special case. Making them all > > work the same is removing an element of surprise from someone expecting > > things to work as documented. > > > > Can we please get this resolved one way or the other? As I've already said, I do not see this practical enough problem to warrant special treatment. Do we really care about controlling the oom behavior for tasks with <4MB of memory? I fully agree that the current situation is not ideal. The whole oom_score* API sucks but here we are with an user API that is effectivelly impossible to fix properly.
diff --git a/mm/oom_kill.c b/mm/oom_kill.c index eefd3f5fde46..4ae0b6193bcd 100644 --- a/mm/oom_kill.c +++ b/mm/oom_kill.c @@ -233,8 +233,11 @@ long oom_badness(struct task_struct *p, unsigned long totalpages) mm_pgtables_bytes(p->mm) / PAGE_SIZE; task_unlock(p); - /* Normalize to oom_score_adj units */ - adj *= totalpages / 1000; + /* + * Normalize to oom_score_adj units. You should never + * multiply by zero here, or oom_score_adj will not work. + */ + adj *= (totalpages + 1000) / 1000; points += adj; return points;