linux-kernel - Re: [PATCH] mm, vmscan: add cond_resched into shrink_node

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-Id: <20161205161629.GD3924@linux.vnet.ibm.com>
Date:   Mon, 5 Dec 2016 08:16:29 -0800
From:   "Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
To:     Michal Hocko <mhocko@...nel.org>
Cc:     Balbir Singh <bsingharora@...il.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Mel Gorman <mgorman@...e.de>,
        Johannes Weiner <hannes@...xchg.org>,
        Vlastimil Babka <vbabka@...e.cz>,
        linux-mm <linux-mm@...ck.org>,
        LKML <linux-kernel@...r.kernel.org>,
        Boris Zhmurov <bb@...nelpanic.ru>,
        "Christopher S. Aker" <caker@...shore.net>,
        Donald Buczek <buczek@...gen.mpg.de>,
        Paul Menzel <pmenzel@...gen.mpg.de>
Subject: Re: [PATCH] mm, vmscan: add cond_resched into shrink_node_memcg

On Mon, Dec 05, 2016 at 01:49:55PM +0100, Michal Hocko wrote:
> [CC Paul - sorry I've tried to save you from more emails...]
> 
> On Mon 05-12-16 23:44:27, Balbir Singh wrote:
> > >
> > > Hi,
> > > there were multiple reportes of the similar RCU stalls. Only Boris has
> > > confirmed that this patch helps in his workload. Others might see a
> > > slightly different issue and that should be investigated if it is the
> > > case. As pointed out by Paul [1] cond_resched might be not sufficient
> > > to silence RCU stalls because that would require a real scheduling.
> > > This is a separate problem, though, and Paul is working with Peter [2]
> > > to resolve it.
> > >
> > > Anyway, I believe that this patch should be a good start because it
> > > really seems that nr_taken=0 during the LRU isolation can be triggered
> > > in the real life. All reporters are agreeing to start seeing this issue
> > > when moving on to 4.8 kernel which might be just a coincidence or a
> > > different behavior of some subsystem. Well, MM has moved from zone to
> > > node reclaim but I couldn't have found any direct relation to that
> > > change.
> > >
> > > [1] http://lkml.kernel.org/r/20161130142955.GS3924@linux.vnet.ibm.com
> > > [2] http://lkml.kernel.org/r/20161201124024.GB3924@linux.vnet.ibm.com
> > >
> > >  mm/vmscan.c | 2 ++
> > >  1 file changed, 2 insertions(+)
> > >
> > > diff --git a/mm/vmscan.c b/mm/vmscan.c
> > > index c05f00042430..c4abf08861d2 100644
> > > --- a/mm/vmscan.c
> > > +++ b/mm/vmscan.c
> > > @@ -2362,6 +2362,8 @@ static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memc
> > >                         }
> > >                 }
> > >
> > > +               cond_resched();
> > > +
> > 
> > I see a cond_resched_rcu_qs() as a part of linux next inside the while
> > (nr[..]) loop.
> 
> This is a left over from Paul's initial attempt to fix this issue. I
> expect him to drop his patch from his tree. He has considered it
> experimental anyway.

To prevent further confusion, I am dropping these patches from my tree:

80c099e11c19 ("mm: Prevent shrink_node() RCU CPU stall warnings")
34c53f5cd399 ("mm: Prevent shrink_node_memcg() RCU CPU stall warnings")

If you need them, please feel free to pull them in.

Given that I don't have those, I am dropping this one as well:

f2a471ffc8a8 ("rcu: Allow boot-time use of cond_resched_rcu_qs()")

If you need it, please let me know.

> > Do we need this as well?
> 
> Paul is working with Peter to make cond_resched general and cover RCU
> stalls even when cond_resched doesn't schedule because there is no
> runnable task.

And 0day just told me that my current attempt gets a 227% increase in
context switches on the unlink tests in LTP, so back to the drawing
board...

						Thanx, Paul