Re: iotop: khugepaged at 99.99% (2.6.38.X)

[Andrea Arcangeli <aarcange@redhat.com>]

On Fri, May 06, 2011 at 04:24:04PM +0200, Thomas Sattler wrote:
> > Aaarg, wrong kernel tree. I patched and compiled 2.6.38.5.
> > Do you think it is important to stay with 2.6.38.2, after
> > we know 2.6.38.4 is also affected?
> 
> I bootet 2.6.38.5.aa1 ("aa1" for the "make-it-worse-patch")

Sorry, unfortunately the make-it-worse-patch had a misplaced #if 0
which resulted in the VM not being able to reclaim, it should have
been around __alloc_pages_direct_compact and instead it was around
__alloc_pages_direct_reclaim (I noticed the hard way too).

The second patch (hotfix, not the make-it-worse) I sent should work
just fine instead.

Other ways we could fix it (if my vmstat per-cpu theory is right)
would be to call the equivalent of start_cpu_timer() to
schedule_delayed_work_on every CPU after congestion_wait returns
before re-evaluating too_many_isolated (however that would still add a
100msec latency here and there plus doing some overscheduling in
possibly no VM-congested situations where just one task quit releasing
all anon memory in the inactive list), or probably to always return
false from too_many_isolated if nr_isolated_anon <
threshold*CONFIG_NR_CPUS would be enough to sort the per-cpu
accounting error.. but personally I prefer to nuke the function for
all reasons mentioned in the prev email and go ahead and drop the
isolated counter too. However a more strict fix would give more
confirmation that we're not hiding a stat accounting error and confirm
my theory, but for the long run (after having spent a day reading that
function) I don't really like to keep it.

The correct make-it-worse patch would be this (and this time I tested
it before sending ;). This should speedup the time it takes to
reproduce as it'll always enter reclaim with __GFP_NO_KSWAPD
allocations (while previously it'd enter reclaim only if compaction
failed). And entering reclaim without kswapd running and churning over
the per-cpu stats and adding stuff from active to the inactive list
even when the inactive list gets trimmed to zero by an exit(), would
screw things up.

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 9f8a97b..3dcd442 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -2093,6 +2093,7 @@ rebalance:
 	if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
 		goto nopage;
 
+#if 0
 	/*
 	 * Try direct compaction. The first pass is asynchronous. Subsequent
 	 * attempts after direct reclaim are synchronous
@@ -2105,7 +2106,8 @@ rebalance:
 					sync_migration);
 	if (page)
 		goto got_pg;
-	sync_migration = !(gfp_mask & __GFP_NO_KSWAPD);
+#endif
+	sync_migration = true;
 
 	/* Try direct reclaim and then allocating */
 	page = __alloc_pages_direct_reclaim(gfp_mask, order,