On 7/7/21 5:25 PM, Stefan Eissing wrote:
> In order to reproduce the logs I see on this weird behaviour, I'll attach the
> patch I made:
>
>
>
>
> With this, I see in a hanging process:
>
> [Wed Jul 07 15:20:23.014980 2021] [core:trace1] [pid 42471:tid
> 123145448968192] core_filters.c(429): (32)Broken pipe: [client
> 127.0.0.1:49824] core_output_filter: writing data to the network
> [Wed Jul 07 15:20:23.044541 2021] [mpm_event:debug] [pid 42471:tid
> 4525448704] event.c(599): AH: wakeup_listener: start
> [Wed Jul 07 15:20:23.044601 2021] [mpm_event:debug] [pid 42471:tid
> 4525448704] event.c(603): AH: wakeup_listener: apr_pollset_wakeup
> [Wed Jul 07 15:20:23.044715 2021] [mpm_event:debug] [pid 42471:tid
> 4525448704] event.c(610): AH: wakeup_listener: ap_queue_info_term
> [Wed Jul 07 15:20:23.044861 2021] [mpm_event:debug] [pid 42471:tid
> 123145461309440] event.c(1985): (4)Interrupted system call: AH: pollset
> returned listener_may_exit=0
Hm. The ap_log_error statically writes listener_may_exit=0. Can you put the
actual value of listener_may_exit in the log message?
> [Wed Jul 07 15:20:23.044817 2021] [mpm_event:debug] [pid 42471:tid
> 4525448704] event.c(629): AH: wakeup_listener: pthread_kill
> [Wed Jul 07 15:20:23.044954 2021] [mpm_event:trace1] [pid 42471:tid
> 123145461309440] event.c(1809): All workers are busy or dying, will close 1
> keep-alive connections
> [Wed Jul 07 15:20:23.044987 2021] [mpm_event:debug] [pid 42471:tid
> 4525448704] event.c(635): AH: wakeup_listener: end
> [Wed Jul 07 15:20:23.045154 2021] [mpm_event:trace1] [pid 42471:tid
> 4525448704] event.c(2940): graceful termination received, join workers
> [Wed Jul 07 15:20:25.927437 2021] [mpm_event:debug] [pid 42471:tid
> 123145461309440] event.c(1985): (70007)The timeout specified has expired: AH:
> pollset returned listener_may_exit=0
Looks like we get back in the poll and have to wait for the timeout.
> [Wed Jul 07 15:20:25.927468 2021] [mpm_event:trace1] [pid 42471:tid
> 123145461309440] event.c(1809): All workers are busy or dying, will close 0
> keep-alive connections
>
> Which says (to my eyes) that the pollset_wakeup happened, but the listener
> saw "listener_may_exit == 0". Is the volatile maybe not enough?
Regards
Rüdiger
