The patch titled vmscan: make kswapd use a correct order has been added to the -mm tree. Its filename is vmscan-make-kswapd-use-a-correct-order.patch Before you just go and hit "reply", please: a) Consider who else should be cc'ed b) Prefer to cc a suitable mailing list as well c) Ideally: find the original patch on the mailing list and do a reply-to-all to that, adding suitable additional cc's *** Remember to use Documentation/SubmitChecklist when testing your code *** See http://userweb.kernel.org/~akpm/stuff/added-to-mm.txt to find out what to do about this The current -mm tree may be found at http://userweb.kernel.org/~akpm/mmotm/ ------------------------------------------------------ Subject: vmscan: make kswapd use a correct order From: Minchan Kim <minchan.kim@xxxxxxxxx> If we wake up prematurely, it means we should keep going on reclaiming not new order page but at old order page. Sometime new order can be smaller than old order by below race so it could make failure of old order page reclaiming. T0: Task 1 wakes up kswapd with order-3 T1: So, kswapd starts to reclaim pages using balance_pgdat T2: Task 2 wakes up kswapd with order-2 because pages reclaimed by T1 are consumed quickly. T3: kswapd exits balance_pgdat and will do following: T4-1: In beginning of kswapd's loop, pgdat->kswapd_max_order will be reset with zero. T4-2: 'order' will be set to pgdat->kswapd_max_order(0), since it enters the false branch of 'if (order (3) < new_order (2))' T4-3: If previous balance_pgdat can't meet requirement of order-2 free pages by high watermark, it will start reclaiming again. So balance_pgdat will use order-0 to do reclaim while it really should use order-2 at the moment. T4-4: At last, Task 1 can't get the any page if it wanted with GFP_ATOMIC. Reported-by: Shaohua Li <shaohua.li@xxxxxxxxx> Signed-off-by: Minchan Kim <minchan.kim@xxxxxxxxx> Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@xxxxxxxxxxxxxx> Reviewed-by: Shaohua Li <shaohua.li@xxxxxxxxx> Acked-by: Mel Gorman <mel@xxxxxxxxx> Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> --- mm/vmscan.c | 23 +++++++++++++++++++---- 1 file changed, 19 insertions(+), 4 deletions(-) diff -puN mm/vmscan.c~vmscan-make-kswapd-use-a-correct-order mm/vmscan.c --- a/mm/vmscan.c~vmscan-make-kswapd-use-a-correct-order +++ a/mm/vmscan.c @@ -2454,13 +2454,18 @@ out: return sc.nr_reclaimed; } -static void kswapd_try_to_sleep(pg_data_t *pgdat, int order) +/* + * Return true if we slept enough. Otherwise, return false + */ +static bool kswapd_try_to_sleep(pg_data_t *pgdat, int order) { long remaining = 0; + bool slept = false; + DEFINE_WAIT(wait); if (freezing(current) || kthread_should_stop()) - return; + return slept; prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); @@ -2489,6 +2494,7 @@ static void kswapd_try_to_sleep(pg_data_ set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); schedule(); set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); + slept = true; } else { if (remaining) count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); @@ -2496,6 +2502,8 @@ static void kswapd_try_to_sleep(pg_data_ count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); } finish_wait(&pgdat->kswapd_wait, &wait); + + return slept; } /* @@ -2557,8 +2565,15 @@ static int kswapd(void *p) */ order = new_order; } else { - kswapd_try_to_sleep(pgdat, order); - order = pgdat->kswapd_max_order; + /* + * If we wake up after enough sleeping, it means + * we reclaimed enough pages at that order. so + * we starts reclaim new order in this time. + * Otherwise, it was a premature sleep so we should + * keep going on reclaiming at that order pages. + */ + if (kswapd_try_to_sleep(pgdat, order)) + order = pgdat->kswapd_max_order; } ret = try_to_freeze(); _ Patches currently in -mm which might be from minchan.kim@xxxxxxxxx are linux-next.patch mm-compactionc-avoid-double-mem_cgroup_del_lru.patch mm-vmap-area-cache.patch mm-find_get_pages_contig-fixlet.patch mm-deactivate-invalidated-pages.patch mm-deactivate-invalidated-pages-fix.patch vmalloc-remove-redundant-unlikely.patch vmscan-make-kswapd-use-a-correct-order.patch memcg-add-page_cgroup-flags-for-dirty-page-tracking.patch memcg-document-cgroup-dirty-memory-interfaces.patch memcg-document-cgroup-dirty-memory-interfaces-fix.patch memcg-create-extensible-page-stat-update-routines.patch memcg-add-lock-to-synchronize-page-accounting-and-migration.patch memcg-remove-unnecessary-return-from-void-returning-mem_cgroup_del_lru_list.patch memcg-use-zalloc-rather-than-mallocmemset.patch -- To unsubscribe from this list: send the line "unsubscribe mm-commits" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html