mm: ignore the boosting of watermark under lowmemory

When low on free memory and a fragmentation event at the same time
will cause the zone watermarks to be offsetted by watermark_boost
which in default is 150% of high watermark. The side effects of this
is the time taken by the reclaim to reach that level and thus the
subsequent reset of the watermark_boost can be higher and hence
killing of the FG tasks can happen in that time frame despite
the free memory is above zone->_watermark[WMARK_HIGH]. Ignore
boosting of watermark under this condition and thus we can create
more memory headroom.

Change-Id: Ia1fa5d766eee851ad3456f41b4b2d623cf4eeee3
Signed-off-by: Charan Teja Reddy <charante@codeaurora.org>
This commit is contained in:
Charan Teja Reddy 2019-11-13 20:12:10 +05:30 committed by Gerrit - the friendly Code Review server
parent 4b832ee13f
commit feeb6c80da

View File

@ -2224,11 +2224,37 @@ static bool can_steal_fallback(unsigned int order, int start_mt)
return false;
}
static bool boost_eligible(struct zone *z)
{
unsigned long high_wmark, threshold;
unsigned long reclaim_eligible, free_pages;
high_wmark = z->_watermark[WMARK_HIGH];
reclaim_eligible = zone_page_state_snapshot(z, NR_ZONE_INACTIVE_FILE) +
zone_page_state_snapshot(z, NR_ZONE_ACTIVE_FILE);
free_pages = zone_page_state(z, NR_FREE_PAGES);
threshold = high_wmark + (2 * mult_frac(high_wmark,
watermark_boost_factor, 10000));
/*
* Don't boost watermark If we are already low on memory where the
* boosting can simply put the watermarks at higher levels for a
* longer duration of time and thus the other users relied on the
* watermarks are forced to choose unintended decissions. If memory
* is so low, kswapd in normal mode should help.
*/
if (reclaim_eligible < threshold && free_pages < threshold)
return false;
return true;
}
static inline void boost_watermark(struct zone *zone)
{
unsigned long max_boost;
if (!watermark_boost_factor)
if (!watermark_boost_factor || !boost_eligible(zone))
return;
max_boost = mult_frac(zone->_watermark[WMARK_HIGH],