The timeout argument has an upper limit. Any values above that limit are treated the same as -1, i.e. to wait indefinitely. The limit is given by:
#define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX - 999ULL) / HZ)
That is, the limit depends on the size of a long and the timer frequency. Assuming the a long is never smaller than 32 bits and HZ never larger than 1000, the worst case is 35 minutes. I think this should be mentioned under "BUGS".
Although this is likely to be fixed in the future (http://lkml.org/lkml/2010/8/8/144), the problem exists in at least 2.6.14 - 2.6.35. I don't know if select(2) and poll(2) are affected.
I aplied the following patch:
@@ -198,6 +198,18 @@ If the new file descriptor becomes ready,
it will cause the
.BR epoll_wait ()
call to unblock.
+In kernels before 2.6.37, a
+value larger than approximately
+.I LONG_MAX / HZ
+milliseconds is treated as \-1 (i.e., infinity).
+Thus, for example, on a system where the
+is 4 and the kernel
+value is 1000,
+this means that timeouts greater than 35.79 minutes are treated as infinity.
.SH "SEE ALSO"
.BR epoll_create (2),
.BR epoll_ctl (2),