On Fri, Jul 11, 2025 at 07:02:17AM +0200, Nam Cao wrote:
> On Thu, Jul 10, 2025 at 05:47:57PM +0800, Xi Ruoyao wrote:
> > It didn't work :(.
>
> Argh :(
>
> Another possibility is that you are running into event starvation problem.
>
> Can you give the below patch a try? It is not the real fix, the patch hurts
> performance badly. But if starvation is really your problem, it should
> ameliorate the situation:
>
> diff --git a/fs/eventpoll.c b/fs/eventpoll.c
> index 895256cd2786..0dcf8e18de0d 100644
> --- a/fs/eventpoll.c
> +++ b/fs/eventpoll.c
> @@ -1764,6 +1764,8 @@ static int ep_send_events(struct eventpoll *ep,
> __llist_add(n, &txlist);
> }
>
> + struct llist_node *shuffle = llist_del_all(&ep->rdllist);
> +
> llist_for_each_entry_safe(epi, tmp, txlist.first, rdllink) {
> init_llist_node(&epi->rdllink);
>
> @@ -1778,6 +1780,13 @@ static int ep_send_events(struct eventpoll *ep,
> }
> }
>
> + if (shuffle) {
> + struct llist_node *last = shuffle;
> + while (last->next)
> + last = last->next;
> + llist_add_batch(shuffle, last, &ep->rdllist);
> + }
> +
> __pm_relax(ep->ws);
> mutex_unlock(&ep->mtx);
>
I think we should revert the fix so we have time to fix it properly
during v6.17+. This patch was a bit too adventurous for a fix in the
first place tbh.
