Re: epoll_wait inaccurate timeout

From: Dmitry Banschikov
Date: Tue Dec 13 2016 - 02:47:56 EST


On Mon, Dec 12, 2016 at 2:34 PM, Eric Wong <e@xxxxxxxxx> wrote:
> +Cc folks who may know about timer stuff on epoll.
>
> Dmitry Banschikov <me@xxxxxxxxxxxxx> wrote:
>> Hi!
>>
>> I have a problem caused by inaccurate timeouts in epoll_wait(2).
>> Here are some parts of strace -tt output:
>
> Which kernel version are you using?

I was able to reproduce this problem in 4.4.0(Ubuntu 16.04) and
3.10.0(CentOS 7.2.1511) on
two different x86_64 machines.

>
>> 22578 09:33:46.959791 epoll_wait(5, <unfinished ...>
>> 22578 09:33:50.010794 <... epoll_wait resumed> [], 128, 1498) = 0
>> ...
>> 22034 09:35:07.686896 epoll_wait(5, <unfinished ...>
>> 22034 09:35:09.482526 <... epoll_wait resumed> [{EPOLLIN,
>> {u32=151458248, u64=151458248}}], 128, 362) = 1
>> ...
>> 22036 09:35:41.433241 epoll_wait(5, <unfinished ...>
>> 22036 09:35:43.176881 <... epoll_wait resumed> [], 128, 97) = 0
>>
>> In each example epoll_wait is blocked for too longer then asked in timeout.
>>
>> Is it normal?
>
> I don't think so, unless you have a huge /proc/<pid>/timerslack_ns
> set. But I mainly use -1 or 0 as the timeout value.

/proc/<pid>/timerslack_ns interface was added in 4.6.
For sure I can try to reproduce problem on fresh kernel if it can help
debugging.

I observe such epoll_wait behavior quite rarely - usually
one/two/three times per
hours of attempts to reproduce. And from strace output I can see that
system is not in state of resource starvation - because other threads
do some work
between call and return to/from epoll_wait.

Such timeout values for epoll_wait are generated by Boost ASIO library.
Internally it uses priority queue for storing timer events and use timeout to
nearest event from queue in epoll_wait.

What information can help to debug this issue?


>
>> Please CC me.
>
> That's standard procedure, here :)



--

Dmitry Banschikov