2015-04-14 16:49:53

by Michal Hocko

[permalink] [raw]
Subject: Re: [patch 09/12] mm: page_alloc: private memory reserves for OOM-killing allocations

[Sorry for the late reply]

On Wed 25-03-15 02:17:13, Johannes Weiner wrote:
> The OOM killer connects random tasks in the system with unknown
> dependencies between them, and the OOM victim might well get blocked
> behind the task that is trying to allocate. That means that while
> allocations can issue OOM kills to improve the low memory situation,
> which generally frees more than they are going to take out, they can
> not rely on their *own* OOM kills to make forward progress for them.
>
> Secondly, we want to avoid a racing allocation swooping in to steal
> the work of the OOM killing allocation, causing spurious allocation
> failures. The one that put in the work must have priority - if its
> efforts are enough to serve both allocations that's fine, otherwise
> concurrent allocations should be forced to issue their own OOM kills.
>
> Keep some pages below the min watermark reserved for OOM-killing
> allocations to protect them from blocking victims and concurrent
> allocations not pulling their weight.

Yes, this makes a lot of sense. I am just not sure I am happy about some
details.

[...]
> @@ -3274,6 +3290,7 @@ void show_free_areas(unsigned int filter)
> show_node(zone);
> printk("%s"
> " free:%lukB"
> + " oom:%lukB"
> " min:%lukB"
> " low:%lukB"
> " high:%lukB"
> @@ -3306,6 +3323,7 @@ void show_free_areas(unsigned int filter)
> "\n",
> zone->name,
> K(zone_page_state(zone, NR_FREE_PAGES)),
> + K(oom_wmark_pages(zone)),
> K(min_wmark_pages(zone)),
> K(low_wmark_pages(zone)),
> K(high_wmark_pages(zone)),

Do we really need to export the new watermark into the userspace?
How would it help user/admin? OK, maybe show_free_areas could be helpful
for oom reports but why to export it in /proc/zoneinfo which is done
further down?

> @@ -5747,17 +5765,18 @@ static void __setup_per_zone_wmarks(void)
>
> min_pages = zone->managed_pages / 1024;
> min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
> - zone->watermark[WMARK_MIN] = min_pages;
> + zone->watermark[WMARK_OOM] = min_pages;
> } else {
> /*
> * If it's a lowmem zone, reserve a number of pages
> * proportionate to the zone's size.
> */
> - zone->watermark[WMARK_MIN] = tmp;
> + zone->watermark[WMARK_OOM] = tmp;
> }
>
> - zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
> - zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
> + zone->watermark[WMARK_MIN] = oom_wmark_pages(zone) + (tmp >> 3);
> + zone->watermark[WMARK_LOW] = oom_wmark_pages(zone) + (tmp >> 2);
> + zone->watermark[WMARK_HIGH] = oom_wmark_pages(zone) + (tmp >> 1);

This will basically elevate the min watermark, right? And that might lead
to subtle performance differences even when OOM killer is not invoked
because the direct reclaim will start sooner.
Shouldn't we rather give WMARK_OOM half of WMARK_MIN instead?

>
> __mod_zone_page_state(zone, NR_ALLOC_BATCH,
> high_wmark_pages(zone) - low_wmark_pages(zone) -
> diff --git a/mm/vmstat.c b/mm/vmstat.c
> index 1fd0886a389f..a62f16ef524c 100644
> --- a/mm/vmstat.c
> +++ b/mm/vmstat.c
> @@ -1188,6 +1188,7 @@ static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
> seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
> seq_printf(m,
> "\n pages free %lu"
> + "\n oom %lu"
> "\n min %lu"
> "\n low %lu"
> "\n high %lu"
> @@ -1196,6 +1197,7 @@ static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
> "\n present %lu"
> "\n managed %lu",
> zone_page_state(zone, NR_FREE_PAGES),
> + oom_wmark_pages(zone),
> min_wmark_pages(zone),
> low_wmark_pages(zone),
> high_wmark_pages(zone),
> --
> 2.3.3
>

--
Michal Hocko
SUSE Labs


2015-04-24 19:14:08

by Johannes Weiner

[permalink] [raw]
Subject: Re: [patch 09/12] mm: page_alloc: private memory reserves for OOM-killing allocations

On Tue, Apr 14, 2015 at 06:49:40PM +0200, Michal Hocko wrote:
> On Wed 25-03-15 02:17:13, Johannes Weiner wrote:
> > @@ -5747,17 +5765,18 @@ static void __setup_per_zone_wmarks(void)
> >
> > min_pages = zone->managed_pages / 1024;
> > min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
> > - zone->watermark[WMARK_MIN] = min_pages;
> > + zone->watermark[WMARK_OOM] = min_pages;
> > } else {
> > /*
> > * If it's a lowmem zone, reserve a number of pages
> > * proportionate to the zone's size.
> > */
> > - zone->watermark[WMARK_MIN] = tmp;
> > + zone->watermark[WMARK_OOM] = tmp;
> > }
> >
> > - zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
> > - zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
> > + zone->watermark[WMARK_MIN] = oom_wmark_pages(zone) + (tmp >> 3);
> > + zone->watermark[WMARK_LOW] = oom_wmark_pages(zone) + (tmp >> 2);
> > + zone->watermark[WMARK_HIGH] = oom_wmark_pages(zone) + (tmp >> 1);
>
> This will basically elevate the min watermark, right? And that might lead
> to subtle performance differences even when OOM killer is not invoked
> because the direct reclaim will start sooner.

It will move the min watermark a bit closer to the kswapd watermarks,
so I guess the risk of entering direct reclaim when kswapd won't wake
up fast enough before concurrent allocator slowpaths deplete the zone
from low to min is marginally increased. That seems like a farfetched
worry, especially given that waking up a sleeping kswapd is not a high
frequency event in the first place.

> Shouldn't we rather give WMARK_OOM half of WMARK_MIN instead?

I guess conceptually that would work as well, since an OOM killing
task is technically reclaiming memory and this reserve is meant to
help reclaiming tasks make forward progress.

That being said, the separate OOM reserve was designed for when the
allocation can actually fail: deplete our own little reserve before
returning failure. But it looks like neither the low-order nor the
GFP_NOFS deadlock fixes got any traction, and so right now all OOM
killing allocations still have the potential to deadlock. Is there a
reason we shouldn't just let them do an ALLOC_NO_WATERMARK allocation
after the OOM victim exited (or timed out)?

Otherwise, I'll just do that in the next iteration.