Re: [PATCH] memcg, vmscan: Do not wait for writeback if killed

From: Michal Hocko
Date: Thu Dec 03 2015 - 04:08:46 EST


On Wed 02-12-15 14:25:03, Andrew Morton wrote:
> On Wed, 2 Dec 2015 15:26:18 +0100 Michal Hocko <mhocko@xxxxxxxxxx> wrote:
>
> > From: Michal Hocko <mhocko@xxxxxxxx>
> >
> > Legacy memcg reclaim waits for pages under writeback to prevent from a
> > premature oom killer invocation because there was no memcg dirty limit
> > throttling implemented back then.
> >
> > This heuristic might complicate situation when the writeback cannot make
> > forward progress because of the global OOM situation. E.g. filesystem
> > backed by the loop device relies on the underlying filesystem hosting
> > the image to make forward progress which cannot be guaranteed and so
> > we might end up triggering OOM killer to resolve the situation. If the
> > oom victim happens to be the task stuck in wait_on_page_writeback in the
> > memcg reclaim then we are basically deadlocked.
> >
> > Introduce wait_on_page_writeback_killable and use it in this path to
> > prevent from the issue. shrink_page_list will back off if the wait
> > was interrupted.
> >
> > ...
> >
> > --- a/mm/vmscan.c
> > +++ b/mm/vmscan.c
> > @@ -1021,10 +1021,19 @@ static unsigned long shrink_page_list(struct list_head *page_list,
> >
> > /* Case 3 above */
> > } else {
> > + int ret;
> > +
> > unlock_page(page);
> > - wait_on_page_writeback(page);
> > + ret = wait_on_page_writeback_killable(page);
> > /* then go back and try same page again */
> > list_add_tail(&page->lru, page_list);
> > +
> > + /*
> > + * We've got killed while waiting here so
> > + * expedite our way out from the reclaim
> > + */
> > + if (ret)
> > + break;
> > continue;
> > }
> > }
>
> This function is 350 lines long and it takes a bit of effort to work
> out what that `break' is breaking from and where it goes next. I think
> you want a "goto keep_killed" here for consistency and sanity.

Yeah, sounds better. See an update below:

> Also, there's high risk here of a pending signal causing the code to
> fall into some busy loop where it repeatedly tries to do something but
> then bales out without doing it. It's unobvious how this change avoids
> such things. (Maybe it *does* avoid such things, but it should be
> obvious!).

shrink_page_list is called from __alloc_contig_migrate_range and
shrink_inactive_list. Both of them handle fatal_signal_pending and bail
out. I was relying on this behavior. I realize this is far from optimal
wrt. readability but I do not have a great idea how to improve it
without sticking more fatal_signal_pending checks into the reclaim path.

So you think a comment would be sufficient?
---
diff --git a/mm/vmscan.c b/mm/vmscan.c
index 98a1934493af..2e8ee9e5fcb5 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -1031,9 +1031,12 @@ static unsigned long shrink_page_list(struct list_head *page_list,
/*
* We've got killed while waiting here so
* expedite our way out from the reclaim
+ *
+ * Our callers should make sure we do not
+ * get here with fatal signals again.
*/
if (ret)
- break;
+ goto keep_killed;
continue;
}
}
@@ -1227,6 +1230,7 @@ static unsigned long shrink_page_list(struct list_head *page_list,
VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
}

+keep_killed:
mem_cgroup_uncharge_list(&free_pages);
try_to_unmap_flush();
free_hot_cold_page_list(&free_pages, true);
--
Michal Hocko
SUSE Labs
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/