Hi,
from time to time, when I close my application using libmicrohttpd (SIGINT
is intercepted for a cleanup), the code gets stuck on MHD_stop_daemon and
only leaves if I wait long enough (last time 5 minutes, I don't know if
there's an explicit timeout involved here), unless I terminate it.
Using gdb to attach to the main pid, I can see that the cause is the
pthread_join called by close_all_connections that apparently never returns.
I'm attaching the whole stack trace for your interest:
#0 0x00000035c1008e60 in pthread_join (threadid=139782655043328,
thread_return=thread_return@entry=0x7ffffb6c0908) at pthread_join.c:92
#1 0x00007f21db491c89 in close_all_connections (daemon=daemon@entry
=0x7e5300)
at daemon.c:3916
#2 0x00007f21db494e90 in MHD_stop_daemon (daemon=0x7e5300) at daemon.c:4064
#3 0x000000000040efb9 in main (argc=1, argv=0x7ffffb6c1008) at janus.c:1844
All the times it happened, there was actually no connection pending on the
server, as the cleanup was done when the server wasn't serving any request
anymore. That said, attaching to the other pids still up, I could find
another one related to MHD, where the backtrace seemed to indicate there
still was a thread serving a connection that wasn't there anymore, and that
was waiting on a poll():
#0 0x00000035c0ce9a5d in poll () at ../sysdeps/unix/syscall-template.S:81
#1 0x00007f91faf19d49 in poll (__timeout=<optimized out>, __nfds=1, __fds=
0x7f91c3ffec60) at /usr/include/bits/poll2.h:46
#2 MHD_handle_connection (data=0x7f91c4031c40) at daemon.c:803
#3 0x00000035c1007d15 in start_thread (arg=0x7f91c3fff700)
at pthread_create.c:308
#4 0x00000035c0cf253d in clone ()
at ../sysdeps/unix/sysv/linux/x86_64/clone.S:114
What could be the cause? I guess that one of the threads serving the
connections (MHD_USE_THREAD_PER_CONNECTION is used in my case) was still
alive, but as I anticipated there's actually no request pending anymore,
and all the tabs in the browser were closed as well, which should have
closed the Ajax requests I use to contact the MHD-based server. What may be
causing the poll to get stuck all that time, and how can I force it to
unblock? EINTR can't be used as it is intercepted by the code, and a
continue is issued without checking whether the webserver is closing or
not. I'm using 0.9.33 by the way, not sure whether this was already fixed
in 0.9.34 or not.
Thanks,
Lorenzo
