On 01/06/2020 17.37, Hillf Danton wrote:
>
> After updating the lru drain sequence, new comers avoid waiting for
> the current drainer, because he is flushing works on each online CPU,
> by trying to lock the mutex; the drainer OTOH tries to do works for
> those who fail to acquire the lock by checking the lru drain sequence
> after releasing lock.
>
> See eef1a429f234 ("mm/swap.c: piggyback lru_add_drain_all() calls")
> for reasons why we can skip waiting for the lock.
That patch tells nothing about such change in behaviour.
Callers like invalidate_bdev() really need synchronous drain to be sure
that pages have no extra reference from per-cpu vectors.
>
> The memory barriers around the sequence and the lock come together
> to remove waiters without their drain works bandoned.
>
> Cc: Sebastian Andrzej Siewior <[email protected]>
> Cc: Konstantin Khlebnikov <[email protected]>
> Signed-off-by: Hillf Danton <[email protected]>
> ---
> This is inspired by one of the works from Sebastian.
>
> --- a/mm/swap.c
> +++ b/mm/swap.c
> @@ -714,10 +714,11 @@ static void lru_add_drain_per_cpu(struct
> */
> void lru_add_drain_all(void)
> {
> - static seqcount_t seqcount = SEQCNT_ZERO(seqcount);
> + static unsigned int lru_drain_seq;
> static DEFINE_MUTEX(lock);
> static struct cpumask has_work;
> - int cpu, seq;
> + int cpu;
> + unsigned int seq;
>
> /*
> * Make sure nobody triggers this path before mm_percpu_wq is fully
> @@ -726,18 +727,16 @@ void lru_add_drain_all(void)
> if (WARN_ON(!mm_percpu_wq))
> return;
>
> - seq = raw_read_seqcount_latch(&seqcount);
> + lru_drain_seq++;
> + smp_mb();
>
> - mutex_lock(&lock);
> +more_work:
>
> - /*
> - * Piggyback on drain started and finished while we waited for lock:
> - * all pages pended at the time of our enter were drained from vectors.
> - */
> - if (__read_seqcount_retry(&seqcount, seq))
> - goto done;
> + if (!mutex_trylock(&lock))
> + return;
>
> - raw_write_seqcount_latch(&seqcount);
> + smp_mb();
> + seq = lru_drain_seq;
>
> cpumask_clear(&has_work);
>
> @@ -759,8 +758,11 @@ void lru_add_drain_all(void)
> for_each_cpu(cpu, &has_work)
> flush_work(&per_cpu(lru_add_drain_work, cpu));
>
> -done:
> mutex_unlock(&lock);
> +
> + smp_mb();
> + if (seq != lru_drain_seq)
> + goto more_work;
> }
> #else
> void lru_add_drain_all(void)
> --
>