On Sun, 2021-11-28 at 04:26 +0900, Alexey Avramov wrote: > I will present the results of the new tests here. > > TLDR; > ===== > No one Mel's patch doesn't prevent stalls in my tests. Seems there may be a problem with the THROTTLE_WRITEBACK bits.. > $ for i in {1..10}; do tail /dev/zero; done > -- 1. with noswap ..because the bandaid below (made of 8cd7c588 shards) on top of Mel's last pulled that one-liner's very pointy fangs. --- mm/backing-dev.c | 5 +++++ mm/vmscan.c | 8 +++++++- 2 files changed, 12 insertions(+), 1 deletion(-) --- a/mm/backing-dev.c +++ b/mm/backing-dev.c @@ -1055,3 +1055,8 @@ long congestion_wait(int sync, long time return ret; } EXPORT_SYMBOL(congestion_wait); + +int async_bdi_congested(void) +{ + return atomic_read(&nr_wb_congested[BLK_RW_ASYNC]) != 0; +} --- a/mm/vmscan.c +++ b/mm/vmscan.c @@ -1021,6 +1021,8 @@ static void handle_write_error(struct ad unlock_page(page); } +extern int async_bdi_congested(void); + void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason) { wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason]; @@ -1048,6 +1050,10 @@ void reclaim_throttle(pg_data_t *pgdat, */ switch(reason) { case VMSCAN_THROTTLE_WRITEBACK: + if (!async_bdi_congested()) { + cond_resched(); + return; + } timeout = HZ/10; if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) { @@ -1079,7 +1085,7 @@ void reclaim_throttle(pg_data_t *pgdat, } prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE); - ret = schedule_timeout(timeout); + ret = io_schedule_timeout(timeout); finish_wait(wqh, &wait); if (reason == VMSCAN_THROTTLE_WRITEBACK)