On Mon, Apr 18, 2011 at 10:30:14PM +1000, NeilBrown wrote:
> On Thu, 14 Apr 2011 11:41:38 +0100 Mel Gorman <mgorman@xxxxxxx> wrote:
>
> > If swap is backed by network storage such as NBD, there is a risk that a
> > large number of reclaimers can hang the system by consuming all
> > PF_MEMALLOC reserves. To avoid these hangs, the administrator must tune
> > min_free_kbytes in advance. This patch will throttle direct reclaimers
> > if half the PF_MEMALLOC reserves are in use as the system is at risk of
> > hanging. A message will be displayed so the administrator knows that
> > min_free_kbytes should be tuned to a higher value to avoid the
> > throttling in the future.
>
> This sounds like a much simpler approach than all the pre-allocation.
> Is it certain to work?
It should - at least I haven't conceived of a situation where it would
fail yet nor have I triggered the throttling logic during tests. The
logic was tested with a debugging patch that set the throttling
level higher.
> Are PF_MEMALLOC reserved only used from direct
> reclaim?
>
No. They are also used by kswapd and by a task that is being OOM killed.
> Is printing a message for the admin really a good idea?
Ordinarily no but initially I wanted to make it brutually obvious
when throttling is hit and what got hit. Ultimately it's more likely
to be a tracepoint.
> Auto-tuning is much
> better than requiring the sysadmin to tune.
That requires the memory reservation and pre-allocation patches. To
keep complexity down, I wanted to treat that as a separate series.
> Is throttling when we are low on memory really a problem that needs to be
> tuned away? Presumably we would get over the memory shortage fairly soon and
> the throttling would stop (??).
>
It depends on what the administrator wants really. If they don't care
about the stall, they can simply ignore the problem because as you say,
it should get resolved quickly and the throttled processes continue.
> > + if (printk_ratelimit())
> > + printk(KERN_INFO "Throttling %s due to reclaim pressure on "
> > + "network storage\n",
> > + current->comm);
> > + do {
> > + prepare_to_wait(&zone->zone_pgdat->pfmemalloc_wait, &wait,
> > + TASK_INTERRUPTIBLE);
> > + schedule();
> > + finish_wait(&zone->zone_pgdat->pfmemalloc_wait, &wait);
> > + } while (!pfmemalloc_watermark_ok(zone->zone_pgdat, high_zoneidx) &&
> > + !fatal_signal_pending(current));
> > +}
> > +
>
> This looks racing. It is my understanding that you should always perform the
> test between the 'prepare_to_wait' and the 'schedule'. Otherwise the wakeup
> could happen just before the prepare_to_wait and you never wake from the
> schedule.
> If that doesn't apply in this case I would appreciate a comment explaining
> why.
>
You're right, it's racy. Well spotted.
>
>
> > unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
> > gfp_t gfp_mask, nodemask_t *nodemask)
> > {
> > @@ -2131,6 +2188,8 @@ unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
> > .nodemask = nodemask,
> > };
> >
> > + throttle_direct_reclaim(gfp_mask, zonelist, nodemask);
> > +
> > trace_mm_vmscan_direct_reclaim_begin(order,
> > sc.may_writepage,
> > gfp_mask);
> > @@ -2482,6 +2541,13 @@ loop_again:
> > }
> >
> > }
> > +
> > + /* Wake throttled direct reclaimers if low watermark is met */
> > + if (sk_memalloc_socks() &&
> > + waitqueue_active(&pgdat->pfmemalloc_wait) &&
> > + pfmemalloc_watermark_ok(pgdat, MAX_NR_ZONES - 1))
> > + wake_up_interruptible(&pgdat->pfmemalloc_wait);
> > +
>
> This test on sk_memalloc_socks looks ugly to me.
> The VM shouldn't be checking on some networking state.
> Do we really need the test? It is not reasonable to always throttle direct
> reclaim when mem gets really low?
It's a micro-optimisation. Throttling is not currently necessary
as backing storage such as local block devices have mempools in
place that avoid dipping into the PF_MEMALLOC reserves. On a normal
configuration, that waitqueue will simply never be active so I can
remove the sk_memalloc_socks() test.
What about in slab though? A function call in the fast path is avoided
by using the sk_memalloc_socks tests which is nice.
> If we do need the test - could networking set some global flag in the VM
> which the VM can then test.
I'd like to keep the test in slab at least but adding a new global flag
feels like a waste. I could add a VM wrapper around sk_memalloc_socks()
that would effectively be a rename but that doesn't seem much better.
> Maybe one day we will have something other than network which needs special
> care with the last dregs of memory - then it could set the global flag too
> (in which case it should probably be a global counter).
>
When/if that happens, the naming would become more obvious. Right now,
it's network-related so doesn't seem unreasonable to have a
network-related check.
--
Mel Gorman
SUSE Labs
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>