Message-ID: <CAMbhsRTFWoEaWXQi7vKc3XVqDv3qVBVx8Ax5TfncJF8A4Txj_w@mail.gmail.com>
Date:	Tue, 25 Oct 2011 10:08:58 -0700
From:	Colin Cross <ccross@...roid.com>
To:	Mel Gorman <mgorman@...e.de>
Cc:	linux-kernel@...r.kernel.org,
	Andrew Morton <akpm@...ux-foundation.org>,
	KAMEZAWA Hiroyuki <kamezawa.hiroyu@...fujitsu.com>,
	Andrea Arcangeli <aarcange@...hat.com>,
	David Rientjes <rientjes@...gle.com>, linux-mm@...ck.org
Subject: Re: [PATCH] mm: avoid livelock on !__GFP_FS allocations

On Tue, Oct 25, 2011 at 4:23 AM, Mel Gorman <mgorman@...e.de> wrote:
> On Tue, Oct 25, 2011 at 02:26:56AM -0700, Colin Cross wrote:
>> On Tue, Oct 25, 2011 at 2:09 AM, Mel Gorman <mgorman@...e.de> wrote:
>> > On Mon, Oct 24, 2011 at 11:39:49PM -0700, Colin Cross wrote:
>> >> Under the following conditions, __alloc_pages_slowpath can loop
>> >> forever:
>> >> gfp_mask & __GFP_WAIT is true
>> >> gfp_mask & __GFP_FS is false
>> >> reclaim and compaction make no progress
>> >> order <= PAGE_ALLOC_COSTLY_ORDER
>> >>
>> >> These conditions happen very often during suspend and resume,
>> >> when pm_restrict_gfp_mask() effectively converts all GFP_KERNEL
>> >> allocations into __GFP_WAIT.
>> > b>
>> >> The oom killer is not run because gfp_mask & __GFP_FS is false,
>> >> but should_alloc_retry will always return true when order is less
>> >> than PAGE_ALLOC_COSTLY_ORDER.
>> >>
>> >> Fix __alloc_pages_slowpath to skip retrying when oom killer is
>> >> not allowed by the GFP flags, the same way it would skip if the
>> >> oom killer was allowed but disabled.
>> >>
>> >> Signed-off-by: Colin Cross <ccross@...roid.com>
>> >
>> > Hi Colin,
>> >
>> > Your patch functionally seems fine. I see the problem and we certainly
>> > do not want to have the OOM killer firing during suspend. I would prefer
>> > that the IO devices would not be suspended until reclaim was completed
>> > but I imagine that would be a lot harder.
>> >
>> > That said, it will be difficult to remember why checking __GFP_NOFAIL in
>> > this case is necessary and someone might "optimitise" it away later. It
>> > would be preferable if it was self-documenting. Maybe something like
>> > this? (This is totally untested)
>>
>> This issue is not limited to suspend, any GFP_NOIO allocation could
>> end up in the same loop.  Suspend is the most likely case, because it
>> effectively converts all GFP_KERNEL allocations into GFP_NOIO.
>>
>
> I see what you mean with GFP_NOIO but there is an important difference
> between GFP_NOIO and suspend.  A GFP_NOIO low-order allocation currently
> implies __GFP_NOFAIL as commented on in should_alloc_retry(). If no progress
> is made, we call wait_iff_congested() and sleep for a bit. As the system
> is running, kswapd and other process activity will proceed and eventually
> reclaim enough pages for the GFP_NOIO allocation to succeed. In a running
> system, GFP_NOIO can stall for a period of time but your patch will cause
> the allocation to fail. While I expect callers return ENOMEM or handle
> the situation properly with a wait-and-retry loop, there will be
> operations that fail that used to succeed. This is why I'd prefer it was
> a suspend-specific fix unless we know there is a case where a machine
> livelocks due to a GFP_NOIO allocation looping forever and even then I'd
> wonder why kswapd was not helping.

OK, I see the change in behavior you are trying to avoid.  With your
patch GFP_NOIO allocations can still fail during suspend, is that OK?
I'm also worried about GFP_NOIO allocations looping forever when swap
is not enabled, but I've never seen it happen, and it would probably
recover eventually when another tried tried a GFP_KERNEL allocation
and oom killed something.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists