On Thu, Jul 07, 2016 at 10:20:39AM +0900, Joonsoo Kim wrote:
> > @@ -3249,9 +3249,19 @@ static void kswapd_try_to_sleep(pg_data_t *pgdat,
> > int order,
> >
> > prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
> >
> > + /*
> > + * If kswapd has not been woken recently, then kswapd goes fully
> > + * to sleep. kcompactd may still need to wake if the original
> > + * request was high-order.
> > + */
> > + if (classzone_idx == -1) {
> > + wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
> > + classzone_idx = MAX_NR_ZONES - 1;
> > + goto full_sleep;
> > + }
>
> Passing -1 to kcompactd would cause the problem?
>
No, it ends up doing a wakeup and then going back to sleep which is not
what is required. I'll fix it.
> > @@ -3390,12 +3386,24 @@ static int kswapd(void *p)
> > * We can speed up thawing tasks if we don't call balance_pgdat
> > * after returning from the refrigerator
> > */
> > - if (!ret) {
> > - trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
> > + if (ret)
> > + continue;
> >
> > - /* return value ignored until next patch */
> > - balance_pgdat(pgdat, order, classzone_idx);
> > - }
> > + /*
> > + * Reclaim begins at the requested order but if a high-order
> > + * reclaim fails then kswapd falls back to reclaiming for
> > + * order-0. If that happens, kswapd will consider sleeping
> > + * for the order it finished reclaiming at (reclaim_order)
> > + * but kcompactd is woken to compact for the original
> > + * request (alloc_order).
> > + */
> > + trace_mm_vmscan_kswapd_wake(pgdat->node_id, alloc_order);
> > + reclaim_order = balance_pgdat(pgdat, alloc_order,
> > classzone_idx);
> > + if (reclaim_order < alloc_order)
> > + goto kswapd_try_sleep;
>
> This 'goto' would cause kswapd to sleep prematurely. We need to check
> *new* pgdat->kswapd_order and classzone_idx even in this case.
>
It only matters if the next request coming is also high-order requests but
one thing that needs to be avoided is kswapd staying awake periods of time
constantly reclaiming for high-order pages. This is why the check means
"If we reclaimed for high-order and failed, then consider sleeping now".
If allocations still require it, they direct reclaim instead.
"Fixing" this potentially causes reclaim storms from kswapd.
> > @@ -3418,10 +3426,10 @@ void wakeup_kswapd(struct zone *zone, int order,
> > enum zone_type classzone_idx)
> > if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
> > return;
> > pgdat = zone->zone_pgdat;
> > - if (pgdat->kswapd_max_order < order) {
> > - pgdat->kswapd_max_order = order;
> > - pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
> > - }
> > + if (pgdat->kswapd_classzone_idx == -1)
> > + pgdat->kswapd_classzone_idx = classzone_idx;
> > + pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
> > classzone_idx);
> > + pgdat->kswapd_order = max(pgdat->kswapd_order, order);
>
> Now, updating pgdat->skwapd_max_order and classzone_idx happens
> unconditionally. Before your patch, it is only updated toward hard
> constraint (e.g. higher order).
>
So? It's updating the request to suit the requirements of all pending
allocation requests that woke kswapd.
> And, I'd like to know why max() is used for classzone_idx rather than
> min()? I think that kswapd should balance the lowest zone requested.
>
If there are two allocation requests -- one zone-constraned and the other
zone-unconstrained, it does not make sense to have kswapd skip the pages
usable for the zone-unconstrained and waste a load of CPU. You could
argue that using min would satisfy the zone-constrained allocation faster
but that's at the cost of delaying the zone-unconstrained allocation and
wasting CPU. Bear in mind that using max may mean some lowmem pages get
freed anyway due to LRU order.
--
Mel Gorman
SUSE Labs
