balance_dirty_pages() includes pages under writeback in gdtc->wb_dirty when checking the per-wb background threshold for strictlimit BDIs. However, wb_over_bg_thresh() only considers reclaimable dirty pages when deciding whether to start more background writeback. This can repeatedly wake the background worker even when there are too few dirty pages to write. In a 15-second FUSE writeback test, background wakeup requests dropped from 2657 to 682 (74% drop), while writeback_start events remained nearly unchanged at 99 and 98. Check wb_over_bg_thresh() before starting background writeback through the strictlimit condition. Leave the global background threshold check and the wakeup required for dirty throttling unchanged. Fixes: 0437906841d0 ("mm: start background writeback based on per-wb threshold for strictlimit BDIs") Reviewed-by: Joanne Koong Reviewed-by: Jan Kara Signed-off-by: Avi Weiss --- v2: - explain the preliminary strictlimit threshold check v1: https://lore.kernel.org/r/20261031070515.3042848-2-thnkslprpt@gmail.com mm/page-writeback.c | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/mm/page-writeback.c b/mm/page-writeback.c index 499a35473e4f..0529b4a7e99c 100644 --- a/mm/page-writeback.c +++ b/mm/page-writeback.c @@ -1824,9 +1824,15 @@ static int balance_dirty_pages(struct bdi_writeback *wb, balance_domain_limits(mdtc, strictlimit); } + /* + * gdtc->wb_dirty includes pages under writeback, so use it as + * a cheap preliminary check before wb_over_bg_thresh() decides + * whether background writeback is needed. + */ if (!writeback_in_progress(wb) && (nr_dirty > gdtc->bg_thresh || - (strictlimit && gdtc->wb_dirty > gdtc->wb_bg_thresh))) + (strictlimit && gdtc->wb_dirty > gdtc->wb_bg_thresh && + wb_over_bg_thresh(wb)))) wb_start_background_writeback(wb); /* base-commit: 36aad21d50d1af3f6954bb0c988257a908202477 -- 2.43.0