On Sat, Apr 25, 2020 at 9:17 AM Jason Baron jbaron@akamai.com wrote:

On 4/24/20 3:00 PM, Khazhismel Kumykov wrote:

...

In the event that we add to ovflist, before 339ddb53d373 we would be woken up by ep_scan_ready_list, and did no wakeup in ep_poll_callback. With that wakeup removed, if we add to ovflist here, we may never wake up. Rather than adding back the ep_scan_ready_list wakeup - which was resulting in unnecessary wakeups, trigger a wake-up in ep_poll_callback.

I'm just curious which 'wakeup' we are talking about here? There is: wake_up(&ep->wq) for the 'ep' and then there is the nested one via: ep_poll_safewake(ep, epi). It seems to me that its only about the later one being missing not both? Is your workload using nested epoll?

If so, it might make sense to just do the later, since the point of the original patch was to minimize unnecessary wakeups.

The missing wake-ups were when we added to ovflist instead of rdllist. Both are "the ready list" together - so I'd think we'd want the same wakeups regardless of which specific list we added to. ep_poll_callback isn't nested specific?

Thanks,

-Jason

...

We noticed that one of our workloads was missing wakeups starting with 339ddb53d373 and upon manual inspection, this wakeup seemed missing to me. With this patch added, we no longer see missing wakeups. I haven't yet tried to make a small reproducer, but the existing kselftests in filesystem/epoll passed for me with this patch.

Fixes: 339ddb53d373 ("fs/epoll: remove unnecessary wakeups of nested epoll")

Signed-off-by: Khazhismel Kumykov khazhy@google.com Reviewed-by: Roman Penyaev rpenyaev@suse.de Cc: Alexander Viro viro@zeniv.linux.org.uk Cc: Heiher r@hev.cc Cc: Jason Baron jbaron@akamai.com Cc: stable@vger.kernel.org

---

v2: use if/elif instead of goto + cleanup suggested by Roman fs/eventpoll.c | 18 +++++++++--------- 1 file changed, 9 insertions(+), 9 deletions(-)

diff --git a/fs/eventpoll.c b/fs/eventpoll.c index 8c596641a72b..d6ba0e52439b 100644 --- a/fs/eventpoll.c +++ b/fs/eventpoll.c @@ -1171,6 +1171,10 @@ static inline bool chain_epi_lockless(struct epitem *epi) { struct eventpoll *ep = epi->ep;

• /* Fast preliminary check */
  

• if (epi->next != EP_UNACTIVE_PTR)
  

•         return false;
  

• /* Check that the same epi has not been just chained from another CPU */
  if (cmpxchg(&epi->next, EP_UNACTIVE_PTR, NULL) != EP_UNACTIVE_PTR)
          return false;
  

@@ -1237,16 +1241,12 @@ static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, v * chained in ep->ovflist and requeued later on. */ if (READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR) {

•         if (epi->next == EP_UNACTIVE_PTR &&
  

•             chain_epi_lockless(epi))
  

•         if (chain_epi_lockless(epi))
  

•                 ep_pm_stay_awake_rcu(epi);
  

• } else if (!ep_is_linked(epi)) {
  

•         /* In the usual case, add event to ready list. */
  

•         if (list_add_tail_lockless(&epi->rdllink, &ep->rdllist))
                  ep_pm_stay_awake_rcu(epi);
  

•         goto out_unlock;
  

• }
  

• /* If this file is already in the ready list we exit soon */
  

• if (!ep_is_linked(epi) &&
  

•     list_add_tail_lockless(&epi->rdllink, &ep->rdllist)) {
  

•         ep_pm_stay_awake_rcu(epi);
  }
  
  /*
  
  

On 2020-04-29 06:12, Jason Baron wrote:

On 4/28/20 2:10 PM, Roman Penyaev wrote:

...

On 2020-04-27 22:38, Jason Baron wrote:

...

On 4/25/20 4:59 PM, Khazhismel Kumykov wrote:

...

On Sat, Apr 25, 2020 at 9:17 AM Jason Baron jbaron@akamai.com wrote:

...

On 4/24/20 3:00 PM, Khazhismel Kumykov wrote:

...

In the event that we add to ovflist, before 339ddb53d373 we would be woken up by ep_scan_ready_list, and did no wakeup in ep_poll_callback. With that wakeup removed, if we add to ovflist here, we may never wake up. Rather than adding back the ep_scan_ready_list wakeup - which was resulting in unnecessary wakeups, trigger a wake-up in ep_poll_callback.

I'm just curious which 'wakeup' we are talking about here? There is: wake_up(&ep->wq) for the 'ep' and then there is the nested one via: ep_poll_safewake(ep, epi). It seems to me that its only about the later one being missing not both? Is your workload using nested epoll?

If so, it might make sense to just do the later, since the point of the original patch was to minimize unnecessary wakeups.

The missing wake-ups were when we added to ovflist instead of rdllist. Both are "the ready list" together - so I'd think we'd want the same wakeups regardless of which specific list we added to. ep_poll_callback isn't nested specific?

So I was thinking that ep_poll() would see these events on the ovflist without an explicit wakeup, b/c the overflow list being active implies that the ep_poll() path would add them to the rdllist in ep_scan_read_list(). Thus, it will see the events either in the current ep_poll() context or via a subsequent syscall to epoll_wait().

However, there are other paths that can call ep_scan_ready_list() thus I agree with you that both wakeups here are necessary.

I do think are are still (smaller) potential races in ep_scan_ready_list() where we have:

write_lock_irq(&ep->lock);         list_splice_init(&ep->rdllist, &txlist);         WRITE_ONCE(ep->ovflist, NULL);         write_unlock_irq(&ep->lock);

And in the ep_poll path we have:

static inline int ep_events_available(struct eventpoll *ep) {         return !list_empty_careful(&ep->rdllist) ||                 READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR; }

Seems to me that first bit needs to be the following, since ep_events_available() is now checked in a lockless way:

write_lock_irq(&ep->lock);     WRITE_ONCE(ep->ovflist, NULL);     smp_wmb();         list_splice_init(&ep->rdllist, &txlist);         write_unlock_irq(&ep->lock);

Hi Jason,

For the first chunk you refer the order seems irrelevant. Either you see something not empty, you go take the lock and then check lists under the lock, either you see all lists are empty.

Hi Roman,

It does matter. Let's say we have:

epfd1->epfd2->socket. And thread a is doing an epoll_wait() on epfd1, and thread b is doing epoll_wait on epfd2. then:

1. data comes in on socket

ep_poll_callback() wakes up threads a and b

2. thread a runs

ep_poll() ep_scan_ready_list() ep_send_events_proc() ep_item_poll() ep_scan_ready_list() list_splice_init(&ep->rdllist, &txlist);

3. now thread b is running

ep_poll() ep_events_available() returns false schedule_hrtimeout_range()

Thus, thread c has missed a wakeup and will never get it.

Similarly, for the second chunk I referenced.

Hi Jason,

Yes, that makes sense.

...

...

And also this bit:

WRITE_ONCE(ep->ovflist, EP_UNACTIVE_PTR)>>         /*          * Quickly re-inject items left on "txlist".          */         list_splice(&txlist, &ep->rdllist);

Should I think better be reversed as well to:

list_splice(&txlist, &ep->rdllist); smp_wmb(); WRITE_ONCE(ep->ovflist, EP_UNACTIVE_PTR);

But this one is much more interesting.  I understand what you are trying to achieve: we can't leave both lists empty for the short period of time, if there is something left the caller of ep_events_available() should be able to see one of the lists is not empty, otherwise it can be too late.

But the problem you've spotted is much worse. Some remains can be in the txlist (this can happen if caller of epoll_wait wants to harvest only 1 event, but there are more in the ->rdlist). And we again get the lost wakeup.

Problem is reproduced by the test below.  The idea is simple: we have 10 threads and 10 event fds. Each thread can harvest only 1 event. 1 producer fires all 10 events at once and waits that all 10 events will be observed by 10 threads.

The fix is basically a revert of 339ddb53d373 with 1 major exception: we do wakeups from ep_scan_ready_list() but if txlist is not empty && if ep_scan_ready_list() is called from the routine, which sends events, not reads it (thus we protect ourselves from repeated wake ups)

I will send the code a bit later.

This was discussed as part of the original discussion around 339ddb53d373: https://lkml.org/lkml/2019/10/7/905

True! This is the exact scenario which is covered by the test from my previous email. And the test fails. I forgot about this discussion.

The context was more a performance difference rather than a semantic difference, but as I pointed out I believe that behavior pre-dates the above commit and goes back to: 86c0517 fs/epoll: deal with wait_queue only once

There, since the thread is left on the waitqueue over the ep_scan_ready_list() thus the ep_wakeup() (that was removed in 339ddb53d373), would no longer wakeup other potential waiters.

So since I think this behavior change goes back to 5.0 and there really haven't been any reports, I don't think there are too many apps relying on these semantics that your test case is showing. It would be interesting to confirm that your test does indeed succeed/fail before/after that patch.

The main problem is that two sequential wakeups can hit the same thread, while the wait queue entry is still in the list. You also described this in https://lkml.org/lkml/2019/10/7/905

If we wakeup only from ep_poll_callback (current code state) we have to be sure each wakeup hits new wait queue entry. This behavior can be achieved with autoremove_wake_function, which should be also good performance-wise, since after an explicit wakeup you don't need to remove the wait entry from the list under the lock.

At the end it turns out like this:

1. either we need to wakeup additionally from ep_scan_ready_list(), this is what we had prior 339ddb53d373.

2. either we wait with autoremove_wake_function, thus we guarantee, that each new wakeup hits new thread.

I'm testing both variants using the last test I sent yesterday, works so far.

I personally tend to the second variant, I really don't like "we-wakeup-from-all-the-places" solution.

Could you please take a look on the patch below? It probably needs some tweaks, but the idea should be clear.

Also, as part of that original discussion, you had a patch that I think addresses this. I would be ok with that, in addition to a patch to address the ordering issue I pointed out. I can post a patch for the former, if you think this plan makes sense?

Go ahead with you reordering findings! That is correct for a single wakeup, which we have in the ep_poll_callback().

-- Roman

diff --git a/fs/eventpoll.c b/fs/eventpoll.c index d6ba0e52439b..1057598cd299 100644 --- a/fs/eventpoll.c +++ b/fs/eventpoll.c @@ -1822,7 +1822,6 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events, { int res = 0, eavail, timed_out = 0; u64 slack = 0; - bool waiter = false; wait_queue_entry_t wait; ktime_t expires, *to = NULL;

@@ -1867,21 +1866,12 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events, */ ep_reset_busy_poll_napi_id(ep);

- /* - * We don't have any available event to return to the caller. We need - * to sleep here, and we will be woken by ep_poll_callback() when events - * become available. - */ - if (!waiter) { - waiter = true; - init_waitqueue_entry(&wait, current); - + do { + init_wait(&wait); write_lock_irq(&ep->lock); __add_wait_queue_exclusive(&ep->wq, &wait); write_unlock_irq(&ep->lock); - }

- for (;;) { /* * We don't want to sleep if the ep_poll_callback() sends us * a wakeup in between. That's why we set the task state @@ -1911,6 +1901,16 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events, timed_out = 1; break; } + + /* We were woken up, thus go and try to harvest some events */ + eavail = 1; + + } while (0); + + if (!list_empty_careful(&wait.entry)) { + write_lock_irq(&ep->lock); + __remove_wait_queue(&ep->wq, &wait); + write_unlock_irq(&ep->lock); } __set_current_state(TASK_RUNNING); @@ -1925,12 +1925,6 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events, !(res = ep_send_events(ep, events, maxevents)) && !timed_out) goto fetch_events;

- if (waiter) { - write_lock_irq(&ep->lock); - __remove_wait_queue(&ep->wq, &wait); - write_unlock_irq(&ep->lock); - } - return res; }