Re: [PATCH] mm/vmscan: Do not block forever at shrink_inactive_list().

[Tetsuo Handa]

David Rientjes wrote:
> On Mon, 26 May 2014, Tetsuo Handa wrote:
> 
> > In shrink_inactive_list(), we do not insert delay at
> > 
> >   if (!sc->hibernation_mode && !current_is_kswapd())
> >     wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
> > 
> > if sc->hibernation_mode != 0.
> > Follow the same reason, we should not insert delay at
> > 
> >   while (unlikely(too_many_isolated(zone, file, sc))) {
> >     congestion_wait(BLK_RW_ASYNC, HZ/10);
> > 
> >     /* We are about to die and free our memory. Return now. */
> >     if (fatal_signal_pending(current))
> >       return SWAP_CLUSTER_MAX;
> >   }
> > 
> > if sc->hibernation_mode != 0.
> > 
> > Signed-off-by: Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx>
> > ---
> >  mm/vmscan.c |    3 +++
> >  1 files changed, 3 insertions(+), 0 deletions(-)
> > 
> > diff --git a/mm/vmscan.c b/mm/vmscan.c
> > index 32c661d..89c42ca 100644
> > --- a/mm/vmscan.c
> > +++ b/mm/vmscan.c
> > @@ -1362,6 +1362,9 @@ static int too_many_isolated(struct zone *zone, int file,
> >  	if (current_is_kswapd())
> >  		return 0;
> >  
> > +	if (sc->hibernation_mode)
> > +		return 0;
> > +
> >  	if (!global_reclaim(sc))
> >  		return 0;
> >  
> 
> This isn't the only too_many_isolated() functions that do a delay, how is 
> the too_many_isolated() in mm/compaction.c different?
> 

I don't know. But today I realized that this patch is not sufficient.

I'm trying to find why __alloc_pages_slowpath() cannot return for many minutes
when a certain type of memory pressure is given on a RHEL7 environment with
4 CPU / 2GB RAM. Today I tried to use ftrace for examining the breakdown of
time-consuming functions inside __alloc_pages_slowpath(). But on the first run,
all processes are trapped into this too_many_isolated()/congestion_wait() loop
while kswapd is not running; stalling forever because nobody can perform
operations for making too_many_isolated() to return 0.

This means that, under rare circumstances, it is possible that all processes
other than kswapd are trapped into too_many_isolated()/congestion_wait() loop
while kswapd is sleeping because this loop assumes that somebody else shall
wake up kswapd and kswapd shall perform operations for making
too_many_isolated() to return 0. However, we cannot guarantee that kswapd is
waken by somebody nor kswapd is not blocked by blocking operations inside
shrinker functions (e.g. mutex_lock()).

We need some more changes. I'm thinking memory allocation watchdog thread.
Add an "unsigned long" field to "struct task_struct", set jiffies to the field
upon entry of GFP_WAIT-able memory allocation attempts, and clear the field
upon returning from GFP_WAIT-able memory allocation attempts. A kernel thread
periodically scans task list and compares the field and jiffies, and (at least)
print warning messages (maybe optionally trigger OOM-killer or kernel panic)
if single memory allocation attempt is taking too long (e.g. 60 seconds).
What do you think?
--
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/