On 13.02.2018 03:48, Adrian Gonzalez wrote:
> Hello
> I've been seeing these panics with dovecot 2.2.31 (and older versions)
> under FreeBSD about once every 2-3 months or so.  I know it's been
> discussed here in the list, but I didn't see a definitive fix.  If
> it's been fixed, please ignore this post (sorry!).
> I finally managed to observe the bug on a server that has the patch
> that prints out the kevent() timespec values and got:
> Nov 29 16:46:27 mail dovecot: master: Panic: kevent(events=74,
> ts=59.1000000000) failed: Invalid argument
> Looking through the code, the panic is triggered in
> io_loop_handler_run_internal().  The timeout values for the kevent()
> call are obtained by calling io_loop_get_wait_time(), which in turn
> calls timeout_get_wait_time() (both in ioloop.c).
> The timeout is computed by subtracting the value returned by
> gettimeofday() from timeout->next_run, and it looks like in very rare
> cases the result in tv_r->tv_usec is 1000000uS (1 second).
> So, it seems that if gettimeofday() returns exactly 0 microseconds,
> and timeout->next_run has exactly 1000000 microseconds, the kevent()
> call will fail (1 in 1e12 chance?).
> I'm not entirely sure where timeout->next_run is computed, but looking
> at timeout_update_next(), it adjusts the seconds only if the
> microseconds are greater than 1000000:  (ioloop.c, line 235)
> if (timeout->next_run.tv_usec > 1000000) {
>                 timeout->next_run.tv_sec++;
>                 timeout->next_run.tv_usec -= 1000000;
> }
> It seems like the comparison here should be greater or equal, but I'm
> not familiar enough with the code to say with absolute certainty.
> I added more debug error messages to the code in
> timeout_update_next(), and although I do see next_run->tv_usec at
> 1000000 several times a day, I haven't seen a panic since early
> December on any of our servers, so I decided to post this to see if
> this analysis makes sense.


I noticed that further down we actually use >= instead of >, so I'm
guessing it was just missed here. Thanks.


Reply via email to