[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [libmicrohttpd] Failed to Remove Listen FD from epoll set

From: Christian Grothoff
Subject: Re: [libmicrohttpd] Failed to Remove Listen FD from epoll set
Date: Wed, 10 Jun 2020 21:15:24 +0200
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:68.0) Gecko/20100101 Thunderbird/68.8.0

Hi Damon,

MHD _internally_ removes the listen socket from the epoll set if it is
at the connection limit. So if you have say near 1024 connections (or
whatever your limit is) open at the time of quiesce, the listen socket
might not be in the epoll-set. However, it _seems_ that the code is
handling this (see 'at_limit' and 'listen_socket_in_epoll' flags).

Also, even if you were to call quiesce twice, that should not cause the
failure you describe, because the code use the 'listen_socket_in_epoll'
flag to guard the call to EPOLL_CTL_DEL, so you should not even get a
panic in that case.

Sorry, not very helpful, I know.

Happy hacking!


On 6/8/20 11:30 PM, Damon Earp wrote:
> I'm running into an abort while trying to perform a graceful shutdown
> and am looking for some guidance for what to look for.
> Environment:
> Debian 9 (stretch), libmicrohttpd 0.9.62
> Daemon Options
> What's Happening
> Upon receiving a SIGINT our server tries to quiesce the daemon, and
> shutdown(2) the listen socket so no more new connections can be made. We
> then use a condition variable and a request count to block until a
> timeout (5 mins) or all requests are complete before stopping the daemon
> and closing the socket.
> In some cases, still trying to isolate when, the call MHD_quiesce_daemon
> leads to an abort. The external logger callback provides this line
> "Fatal error in GNU libmicrohttpd daemon.c:4740: Failed to remove listen
> FD from epoll set".
> I'm sure we are not calling quiesce multiple times: MHD_quiesce_daemon
> is wrapped with a mutex synchronized field to keep track if it has been
> called, and I've reviewed the shutdown call stack. As always I could be
> wrong.
> Are there other conditions where a listen socket is removed from the
> epoll set? We suspend and resume requests constantly, could a failure to
> resume or multiple calls of suspend to a connection cause such behavior?
> Any guidance for what to look for would be appreciated. I will be
> continuing to try to recreate in a more controlled environment with gdb.
> Thanks for your time!
> Damon

Attachment: signature.asc
Description: OpenPGP digital signature

reply via email to

[Prev in Thread] Current Thread [Next in Thread]