Re: [libmicrohttpd] Problem with Keep-Alive and MHD_OPTION_CONNECTION_TIMEOUT in MHD_USE_THREAD_PER_CONNECTION mode

[Christian Grothoff]

On 12/06/2015 01:49 AM, Evgeny Grin wrote:
> Christian,
> 
> You solved the problem. I was able to reproduce the problem: if you
> make a single request from chrome for example and wait more than
> timeout seconds, than all following requests will be put on long
> waiting. But it you make a new request from another client or tab,
> than all requests (including waiting ones) are served immediately.
> Debugger show that MHD select thread are waiting for return from
> MHD_sys_select_ when client are in long waiting. On latest SVN the
> problem is gone. Do we need any refactoring/renaming in MHD code as
> we are using MHD_EPOLL_TURBO in MHD_start_daemon_va to prevent
> sockets shutdown on W32?

Ah, that hack explains why shutdown() wasn't done on W32, which created
this issue in the first place.


Btw, I don't know if the shutdown() on line 1033 that I introduced (just
before close()) hurts in this respect. I just kept it as it was present
in the logic I moved to run earlier. There it was  #ifdef WINDOWS, so I
figured it was, eh, desired. But maybe it's OK and it's just that the
#ifdef WINDOWS vs. #ifdef CYGWIN vs. #ifdef MHD_WINSOCK_SOCKETS is a bit
confusing...


With respect to your last question, I don't think the W32-logic should
abuse the EPOLL_TURBO flag like that. It would seem simpler to just
#ifndef the shutdown() call in MHD_connection_close() and the
set-to-nonblocking logic in daemon.c. That may even result in a smaller
W32 binary (not sure the CC will understand that the flag is set
unconditionally and do DCE otherwise).

signature.asc
Description: OpenPGP digital signature