epoll: autoremove wakers even more aggressively

If a process is killed or otherwise exits while having active network connections and many threads waiting on epoll_wait, the threads will all be woken immediately, but not removed from ep->wq. Then when network traffic scans ep->wq in wake_up, every wakeup attempt will fail, and will not remove the entries from the list.

This means that the cost of the wakeup attempt is far higher than usual, does not decrease, and this also competes with the dying threads trying to actually make progress and remove themselves from the wq.

Handle this by removing visited epoll wq entries unconditionally, rather than only when the wakeup succeeds - the structure of ep_poll means that the only potential loss is the timed_out->eavail heuristic, which now can race and result in a redundant ep_send_events attempt. (But only when incoming data and a timeout actually race, not on every timeout)

fs/eventpoll.c +22 -0

index e2daa940ebce..8b56b94e2f56 100644
@@ -1745,10 +1745,25 @@ static struct timespec64 *ep_timeout_to_timespec(struct timespec64 *to, long ms)
 	ktime_get_ts64(&now);
 	*to = timespec64_add_safe(now, *to);
 	return to;
 }
 
+/*
+ * autoremove_wake_function, but remove even on failure to wake up, because we
+ * know that default_wake_function/ttwu will only fail if the thread is already
+ * woken, and in that case the ep_poll loop will remove the entry anyways, not
+ * try to reuse it.
+ */
+static int ep_autoremove_wake_function(struct wait_queue_entry *wq_entry,
+				       unsigned int mode, int sync, void *key)
+{
+	int ret = default_wake_function(wq_entry, mode, sync, key);
+
+	list_del_init(&wq_entry->entry);
+	return ret;
+}
+
 /**
  * ep_poll - Retrieves ready events, and delivers them to the caller-supplied
  *           event buffer.
  *
  * @ep: Pointer to the eventpoll context.

@@ -1826,12 +1841,19 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
 		 * chance to harvest new event. Otherwise wakeup can be
 		 * lost. This is also good performance-wise, because on
 		 * normal wakeup path no need to call __remove_wait_queue()
 		 * explicitly, thus ep->lock is not taken, which halts the
 		 * event delivery.
+		 *
+		 * In fact, we now use an even more aggressive function that
+		 * unconditionally removes, because we don't reuse the wait
+		 * entry between loop iterations. This lets us also avoid the
+		 * performance issue if a process is killed, causing all of its
+		 * threads to wake up without being removed normally.
 		 */
 		init_wait(&wait);
+		wait.func = ep_autoremove_wake_function;
 
 		write_lock_irq(&ep->lock);
 		/*
 		 * Barrierless variant, waitqueue_active() is called under
 		 * the same lock on wakeup ep_poll_callback() side, so it

Signature · 1 line

2.36.1.476.g0c4daa206d-goog

Quoted 10 lines · On Wed, 15 Jun 2022 14:24:23 -0700 Benjamin Segall <[email protected]> wrote:

If a process is killed or otherwise exits while having active network connections and many threads waiting on epoll_wait, the threads will all be woken immediately, but not removed from ep->wq. Then when network traffic scans ep->wq in wake_up, every wakeup attempt will fail, and will not remove the entries from the list.

This means that the cost of the wakeup attempt is far higher than usual, does not decrease, and this also competes with the dying threads trying to actually make progress and remove themselves from the wq.

Quoted 6 lines

Handle this by removing visited epoll wq entries unconditionally, rather than only when the wakeup succeeds - the structure of ep_poll means that the only potential loss is the timed_out->eavail heuristic, which now can race and result in a redundant ep_send_events attempt. (But only when incoming data and a timeout actually race, not on every timeout)

Thanks. I added people from 412895f03cbf96 ("epoll: atomically remove wait entry on wake up") to cc. Hopefully someone there can help review and maybe test this.

Quoted 20 lines · Andrew Morton

Quoted 10 lines · On Wed, 15 Jun 2022 14:24:23 -0700 Benjamin Segall <[email protected]> wrote:

If a process is killed or otherwise exits while having active network connections and many threads waiting on epoll_wait, the threads will all be woken immediately, but not removed from ep->wq. Then when network traffic scans ep->wq in wake_up, every wakeup attempt will fail, and will not remove the entries from the list.

This means that the cost of the wakeup attempt is far higher than usual, does not decrease, and this also competes with the dying threads trying to actually make progress and remove themselves from the wq.

Quoted 6 lines

Handle this by removing visited epoll wq entries unconditionally, rather than only when the wakeup succeeds - the structure of ep_poll means that the only potential loss is the timed_out->eavail heuristic, which now can race and result in a redundant ep_send_events attempt. (But only when incoming data and a timeout actually race, not on every timeout)

Quoted 5 lines

Thanks. I added people from 412895f03cbf96 ("epoll: atomically remove wait entry on wake up") to cc. Hopefully someone there can help review and maybe test this.

Thanks Andrew. Just wanted to add that we are seeing this issue in production with real workloads and it has caused hard lockups. Particularly network heavy workloads with a lot of threads in epoll_wait() can easily trigger this issue if they get killed (oom-killed in our case).

Quoted 27 lines · On Wed, 29 Jun 2022 18:12:46 -0700 Shakeel Butt <[email protected]> wrote:

Quoted 25 lines · Andrew Morton

Quoted 10 lines · On Wed, 15 Jun 2022 14:24:23 -0700 Benjamin Segall <[email protected]> wrote:

If a process is killed or otherwise exits while having active network connections and many threads waiting on epoll_wait, the threads will all be woken immediately, but not removed from ep->wq. Then when network traffic scans ep->wq in wake_up, every wakeup attempt will fail, and will not remove the entries from the list.

This means that the cost of the wakeup attempt is far higher than usual, does not decrease, and this also competes with the dying threads trying to actually make progress and remove themselves from the wq.

Quoted 6 lines

Handle this by removing visited epoll wq entries unconditionally, rather than only when the wakeup succeeds - the structure of ep_poll means that the only potential loss is the timed_out->eavail heuristic, which now can race and result in a redundant ep_send_events attempt. (But only when incoming data and a timeout actually race, not on every timeout)

Thanks. I added people from 412895f03cbf96 ("epoll: atomically remove wait entry on wake up") to cc. Hopefully someone there can help review and maybe test this.

Quoted 5 lines

Thanks Andrew. Just wanted to add that we are seeing this issue in production with real workloads and it has caused hard lockups. Particularly network heavy workloads with a lot of threads in epoll_wait() can easily trigger this issue if they get killed (oom-killed in our case).

Hard lockups are undesirable. Is a cc:stable justified here?

Quoted 36 lines · Andrew Morton

Quoted 32 lines · On Wed, 29 Jun 2022 18:12:46 -0700 Shakeel Butt <[email protected]> wrote:

Quoted 20 lines · Andrew Morton

Quoted 10 lines · On Wed, 15 Jun 2022 14:24:23 -0700 Benjamin Segall <[email protected]> wrote:

If a process is killed or otherwise exits while having active network connections and many threads waiting on epoll_wait, the threads will all be woken immediately, but not removed from ep->wq. Then when network traffic scans ep->wq in wake_up, every wakeup attempt will fail, and will not remove the entries from the list.

This means that the cost of the wakeup attempt is far higher than usual, does not decrease, and this also competes with the dying threads trying to actually make progress and remove themselves from the wq.

Quoted 6 lines

Handle this by removing visited epoll wq entries unconditionally, rather than only when the wakeup succeeds - the structure of ep_poll means that the only potential loss is the timed_out->eavail heuristic, which now can race and result in a redundant ep_send_events attempt. (But only when incoming data and a timeout actually race, not on every timeout)

Quoted 5 lines

Thanks. I added people from 412895f03cbf96 ("epoll: atomically remove wait entry on wake up") to cc. Hopefully someone there can help review and maybe test this.

Thanks Andrew. Just wanted to add that we are seeing this issue in production with real workloads and it has caused hard lockups. Particularly network heavy workloads with a lot of threads in epoll_wait() can easily trigger this issue if they get killed (oom-killed in our case).

Quoted 1 line

Hard lockups are undesirable. Is a cc:stable justified here?

Not for now as I don't know if we can blame a patch which might be the source of this behavior.

...

Quoted 9 lines

Quoted 6 lines

Quoted 4 lines

production with real workloads and it has caused hard lockups. Particularly network heavy workloads with a lot of threads in epoll_wait() can easily trigger this issue if they get killed (oom-killed in our case).

Hard lockups are undesirable. Is a cc:stable justified here?

Not for now as I don't know if we can blame a patch which might be the source of this behavior.

I am able to repro the epoll hard lockup on next-20220715 with Ben's patch reverted. The repro is a simple TCP server and tens of clients communicating over loopback. Though to cause the hard lockup I have to create a couple thousand threads in epoll_wait() in server and also reduce the kernel.watchdog_thresh. With Ben's patch the repro does not cause the hard lockup even with kernel.watchdog.thresh=1.

Please add:

Tested-by Shakeel Butt <[email protected]>