epoll_wait - 35 minutes equals infinity?

epoll_wait - 35 minutes equals infinity?

[Fredrik Arnerup]

Hi,

I was doing epoll_wait calls with timeout values of over an hour and I was
wondering why the call never returned. Then I read the code in
fs/eventpoll.c:

#define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX -
999ULL) / HZ)
...
jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
		MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;

I.e., with LONG_MAX of 2^31-1 and HZ=1000, any timeouts larger than 35
minutes are treated the same as -1 ("infinity").
Although I'm using 2.6.24, the code doesn't seem to have changed in 2.6.35.
I'm apparently not the first to have this problem
(http://www.provos.org/index.php?/archives/36-Libevent-1.4.4-stable-released
.html).

I suppose it is too late to change the interface to return an error when
setting too high a timeout and too tricky to fix epoll_wait to accept all
positive ints as timeouts.
Still, shouldn't the man page mention that there is a maximum allowable
value to the timeout and that it depends on whether you have a 64- or 32-bit
machine and what the timer frequency is?

/Fredrik Arnerup

Re: epoll_wait - 35 minutes equals infinity?

[Davide Libenzi]

On Mon, 18 Oct 2010, Fredrik Arnerup wrote:

> Hi,
> 
> I was doing epoll_wait calls with timeout values of over an hour and I was
> wondering why the call never returned. Then I read the code in
> fs/eventpoll.c:
> 
> #define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX -
> 999ULL) / HZ)
> ...
> jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
> 		MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
> 
> I.e., with LONG_MAX of 2^31-1 and HZ=1000, any timeouts larger than 35
> minutes are treated the same as -1 ("infinity").
> Although I'm using 2.6.24, the code doesn't seem to have changed in 2.6.35.
> I'm apparently not the first to have this problem
> (http://www.provos.org/index.php?/archives/36-Libevent-1.4.4-stable-released
> .html).
> 
> I suppose it is too late to change the interface to return an error when
> setting too high a timeout and too tricky to fix epoll_wait to accept all
> positive ints as timeouts.
> Still, shouldn't the man page mention that there is a maximum allowable
> value to the timeout and that it depends on whether you have a 64- or 32-bit
> machine and what the timer frequency is?

This is already queued somewhere for fixing:

http://lkml.org/lkml/2010/8/8/144



- Davide