Re: [libmicrohttpd] Busy waiting up to one second using connection timeout

[Evgeny Grin]

The report itself is correct, thanks for this!

However while the patch fixed one problem, it creates another one.
The problem connected to low granularity (rounded to one second) of the 
MHD's internal timer. If connection timeout is set to one second (for 
example) then connection may expire at any moment up to one second with 
the patch applied.
A full example: if activity is recorded at 50.999 seconds, which is 
rounded to 50 internally, then connection with 1 second timeout can be 
closed at 51.0, which is not what is expected by user.

The proper solution would be increase of the resolution of the internal 
MHD timer.
I'll handle it.

--
Evgeny

On 11.08.2021 23:39, Christian Grothoff wrote:
> Thanks for the detailed report, the patch looks good as-is, applied to
> master as
>
> f5d387df7e7fb8de1a5dd0739ddb83a8b19fe64b
>
> Happy hacking!
>
> Christian
>
> On 8/11/21 6:02 PM, Irion, Alexander wrote:
> > Hello!
> >
> > I discovered a problem with closing connections, when
> > MHD_OPTION_CONNECTION_TIMEOUT is used. When I set the timeout option to
> > 3 seconds and connect with a browser, I get a timeout value from
> > MHD_get_timeout() of 3000 ms returned. I then sleep the desired time and
> > call MHD_run again(). Then, MHD_get_timeout() returns repetitive a
> > timeout of 0 which causes thousands of calls to MHD_run(), taking up
> > several hundred milliseconds and burning CPU time.
> >
> > I use an external event loop with MHD_DAEMON_INFO_EPOLL_FD.
> >
> > As reason I found that the condition  timeout <
> > (MHD_monotonic_sec_counter () - connection->last_activity) will not be
> > true, after waiting the timeout time. The right side will be equal to
> > timeout, so the connection is not closed, MHD_get_timeout() continues to
> > return a timeout value of 0, and this runs repetitively in a loop for
> > some thousand iterations, until the time jumps to the next second.
> >
> > In the attachment please find a small patch for fixing this issue. It's
> > just turning the < into a <=.