| Message ID | 20190701201847.251028-1-shakeelb@google.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [v2] mm, vmscan: prevent useless kswapd loops | expand |
On 7/1/19 1:18 PM, Shakeel Butt wrote: > On production we have noticed hard lockups on large machines running > large jobs due to kswaps hoarding lru lock within isolate_lru_pages when > sc->reclaim_idx is 0 which is a small zone. The lru was couple hundred > GiBs and the condition (page_zonenum(page) > sc->reclaim_idx) in > isolate_lru_pages was basically skipping GiBs of pages while holding the > LRU spinlock with interrupt disabled. > > On further inspection, it seems like there are two issues: > > 1) If the kswapd on the return from balance_pgdat() could not sleep > (i.e. node is still unbalanced), the classzone_idx is unintentionally > set to 0 and the whole reclaim cycle of kswapd will try to reclaim > only the lowest and smallest zone while traversing the whole memory. > > 2) Fundamentally isolate_lru_pages() is really bad when the allocation > has woken kswapd for a smaller zone on a very large machine running very > large jobs. It can hoard the LRU spinlock while skipping over 100s of > GiBs of pages. > > This patch only fixes the (1). The (2) needs a more fundamental solution. > To fix (1), in the kswapd context, if pgdat->kswapd_classzone_idx is > invalid use the classzone_idx of the previous kswapd loop otherwise use > the one the waker has requested. > > Fixes: e716f2eb24de ("mm, vmscan: prevent kswapd sleeping prematurely > due to mismatched classzone_idx") > > Signed-off-by: Shakeel Butt <shakeelb@google.com> > --- > Changelog since v1: > - fixed the patch based on Yang Shi's comment. > > mm/vmscan.c | 27 +++++++++++++++------------ > 1 file changed, 15 insertions(+), 12 deletions(-) > > diff --git a/mm/vmscan.c b/mm/vmscan.c > index 9e3292ee5c7c..eacf87f07afe 100644 > --- a/mm/vmscan.c > +++ b/mm/vmscan.c > @@ -3760,19 +3760,18 @@ static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx) > } > > /* > - * pgdat->kswapd_classzone_idx is the highest zone index that a recent > - * allocation request woke kswapd for. When kswapd has not woken recently, > - * the value is MAX_NR_ZONES which is not a valid index. This compares a > - * given classzone and returns it or the highest classzone index kswapd > - * was recently woke for. > + * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be > + * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not > + * a valid index then either kswapd runs for first time or kswapd couldn't sleep > + * after previous reclaim attempt (node is still unbalanced). In that case > + * return the zone index of the previous kswapd reclaim cycle. > */ > static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat, > - enum zone_type classzone_idx) > + enum zone_type prev_classzone_idx) > { > if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) > - return classzone_idx; > - > - return max(pgdat->kswapd_classzone_idx, classzone_idx); > + return prev_classzone_idx; > + return pgdat->kswapd_classzone_idx; > } > > static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, > @@ -3908,7 +3907,7 @@ static int kswapd(void *p) > > /* Read the new order and classzone_idx */ > alloc_order = reclaim_order = pgdat->kswapd_order; > - classzone_idx = kswapd_classzone_idx(pgdat, 0); > + classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); > pgdat->kswapd_order = 0; > pgdat->kswapd_classzone_idx = MAX_NR_ZONES; > > @@ -3961,8 +3960,12 @@ void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order, > if (!cpuset_zone_allowed(zone, gfp_flags)) > return; > pgdat = zone->zone_pgdat; > - pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, > - classzone_idx); > + > + if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) > + pgdat->kswapd_classzone_idx = classzone_idx; > + else > + pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, > + classzone_idx); > pgdat->kswapd_order = max(pgdat->kswapd_order, order); > if (!waitqueue_active(&pgdat->kswapd_wait)) > return; I agree the manipulation to classzone_idx looks convoluted. This version looks correct to me. You could add: Reviewed-by: Yang Shi <yang.shi@linux.alibaba.com>
On Mon, Jul 01, 2019 at 01:18:47PM -0700, Shakeel Butt wrote: > On production we have noticed hard lockups on large machines running > large jobs due to kswaps hoarding lru lock within isolate_lru_pages when > sc->reclaim_idx is 0 which is a small zone. The lru was couple hundred > GiBs and the condition (page_zonenum(page) > sc->reclaim_idx) in > isolate_lru_pages was basically skipping GiBs of pages while holding the > LRU spinlock with interrupt disabled. > > On further inspection, it seems like there are two issues: > > 1) If the kswapd on the return from balance_pgdat() could not sleep > (i.e. node is still unbalanced), the classzone_idx is unintentionally > set to 0 and the whole reclaim cycle of kswapd will try to reclaim > only the lowest and smallest zone while traversing the whole memory. > > 2) Fundamentally isolate_lru_pages() is really bad when the allocation > has woken kswapd for a smaller zone on a very large machine running very > large jobs. It can hoard the LRU spinlock while skipping over 100s of > GiBs of pages. > > This patch only fixes the (1). The (2) needs a more fundamental solution. > To fix (1), in the kswapd context, if pgdat->kswapd_classzone_idx is > invalid use the classzone_idx of the previous kswapd loop otherwise use > the one the waker has requested. > > Fixes: e716f2eb24de ("mm, vmscan: prevent kswapd sleeping prematurely > due to mismatched classzone_idx") > > Signed-off-by: Shakeel Butt <shakeelb@google.com> Acked-by: Mel Gorman <mgorman@techsingularity.net>
diff --git a/mm/vmscan.c b/mm/vmscan.c index 9e3292ee5c7c..eacf87f07afe 100644 --- a/mm/vmscan.c +++ b/mm/vmscan.c @@ -3760,19 +3760,18 @@ static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx) } /* - * pgdat->kswapd_classzone_idx is the highest zone index that a recent - * allocation request woke kswapd for. When kswapd has not woken recently, - * the value is MAX_NR_ZONES which is not a valid index. This compares a - * given classzone and returns it or the highest classzone index kswapd - * was recently woke for. + * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be + * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not + * a valid index then either kswapd runs for first time or kswapd couldn't sleep + * after previous reclaim attempt (node is still unbalanced). In that case + * return the zone index of the previous kswapd reclaim cycle. */ static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat, - enum zone_type classzone_idx) + enum zone_type prev_classzone_idx) { if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) - return classzone_idx; - - return max(pgdat->kswapd_classzone_idx, classzone_idx); + return prev_classzone_idx; + return pgdat->kswapd_classzone_idx; } static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, @@ -3908,7 +3907,7 @@ static int kswapd(void *p) /* Read the new order and classzone_idx */ alloc_order = reclaim_order = pgdat->kswapd_order; - classzone_idx = kswapd_classzone_idx(pgdat, 0); + classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); pgdat->kswapd_order = 0; pgdat->kswapd_classzone_idx = MAX_NR_ZONES; @@ -3961,8 +3960,12 @@ void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order, if (!cpuset_zone_allowed(zone, gfp_flags)) return; pgdat = zone->zone_pgdat; - pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, - classzone_idx); + + if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) + pgdat->kswapd_classzone_idx = classzone_idx; + else + pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, + classzone_idx); pgdat->kswapd_order = max(pgdat->kswapd_order, order); if (!waitqueue_active(&pgdat->kswapd_wait)) return;
On production we have noticed hard lockups on large machines running large jobs due to kswaps hoarding lru lock within isolate_lru_pages when sc->reclaim_idx is 0 which is a small zone. The lru was couple hundred GiBs and the condition (page_zonenum(page) > sc->reclaim_idx) in isolate_lru_pages was basically skipping GiBs of pages while holding the LRU spinlock with interrupt disabled. On further inspection, it seems like there are two issues: 1) If the kswapd on the return from balance_pgdat() could not sleep (i.e. node is still unbalanced), the classzone_idx is unintentionally set to 0 and the whole reclaim cycle of kswapd will try to reclaim only the lowest and smallest zone while traversing the whole memory. 2) Fundamentally isolate_lru_pages() is really bad when the allocation has woken kswapd for a smaller zone on a very large machine running very large jobs. It can hoard the LRU spinlock while skipping over 100s of GiBs of pages. This patch only fixes the (1). The (2) needs a more fundamental solution. To fix (1), in the kswapd context, if pgdat->kswapd_classzone_idx is invalid use the classzone_idx of the previous kswapd loop otherwise use the one the waker has requested. Fixes: e716f2eb24de ("mm, vmscan: prevent kswapd sleeping prematurely due to mismatched classzone_idx") Signed-off-by: Shakeel Butt <shakeelb@google.com> --- Changelog since v1: - fixed the patch based on Yang Shi's comment. mm/vmscan.c | 27 +++++++++++++++------------ 1 file changed, 15 insertions(+), 12 deletions(-)