Analysis from Atin:

<analysis>

I am seeing glusterd is receiving RPC_CLNT_CONNECT in glusterd_svc_common_rpc_notify () even though shd service is not up (note - this is applicable for all daemon services including brick process). In further debugging what I found is socket_connect_finish () calls __socket_connect_finish () which returns 0 which means its able to connect to the socket and hence RPC_CLNT_CONNECT event is sent to the transport. Is this right? My understanding is __socket_connect_finish () shouldn't be returning 0 here.

</analysis>

(In reply to Raghavendra G from comment #1)
> Analysis from Atin:
> 
> <analysis>
> 
> I am seeing glusterd is receiving RPC_CLNT_CONNECT in
> glusterd_svc_common_rpc_notify () even though shd service is not up (note -
> this is applicable for all daemon services including brick process). In
> further debugging what I found is socket_connect_finish () calls
> __socket_connect_finish () which returns 0 which means its able to connect
> to the socket and hence RPC_CLNT_CONNECT event is sent to the transport. Is
> this right? My understanding is __socket_connect_finish () shouldn't be
> returning 0 here.

Yes. __socket_connect_finish (or getsockopt (socket, SOL_SOCKET, SO_ERROR)  specifically) should've returned ECONNREFUSED as shd is down. I am not sure why this is not the case. Wondering whether there is an issue in TCP/IP (socket) layer or the way we are using getsockopt.

> 
> </analysis>

> 
> Yes. __socket_connect_finish (or getsockopt (socket, SOL_SOCKET, SO_ERROR) 
> specifically) should've returned ECONNREFUSED as shd is down. I am not sure
> why this is not the case. Wondering whether there is an issue in TCP/IP
> (socket) layer or the way we are using getsockopt.

Is it possible to confirm that no other process is listening on same address (IP:PORT) as shd?

Looked into the issue with Du, saw that the problem is in gluster rpc code (no other processes listening on same ip/port).

Following is the control flow we observed:

* connect(2) returned -1 with ECONNREFUSED.
* A shutdown is called on socket, with the intention of processing the failure in poller thread (why error has to be processed in poller thread when connect has failed with an error other than EAGAIN/EBUSY/EINPROGRESS?).
* socket is added for polling.
* socket_event_handle is called with all three - pollin, pollout and pollerr - events set.
* As this is the pollin event received when priv->connected is 0, socket_connect_finish is called, which in turn doesn't find any error set on socket. getsockopt not returning any error might be valid as connect had _failed_ with errno other than EINPROGRESS. connection is deemed successful and higher layers are notified with RPC_CONNECT event.
* On further execution of socket_event_handle, pollerr is processed and a RPC_DISCONNECT event is sent to higher layer.

Some things to ponder about:
* why did we receive a pollin/pollout event on socket when connect has failed clearly?
* why error has to be processed in poller thread when connect has failed with an error other than EAGAIN/EBUSY/EINPROGRESS?

(In reply to Raghavendra G from comment #5)
> Following is the control flow we observed:
> 
> * connect(2) returned -1 with ECONNREFUSED.
> * A shutdown is called on socket, with the intention of processing the
> failure in poller thread (why error has to be processed in poller thread
> when connect has failed with an error other than EAGAIN/EBUSY/EINPROGRESS?).
> * socket is added for polling.
> * socket_event_handle is called with all three - pollin, pollout and pollerr
> - events set.
> * As this is the pollin event received when priv->connected is 0,
> socket_connect_finish is called, which in turn doesn't find any error set on
> socket. getsockopt not returning any error might be valid as connect had
> _failed_ with errno other than EINPROGRESS. connection is deemed successful
> and higher layers are notified with RPC_CONNECT event.
> * On further execution of socket_event_handle, pollerr is processed and a
> RPC_DISCONNECT event is sent to higher layer.
> 
> Some things to ponder about:
> * why did we receive a pollin/pollout event on socket when connect has
> failed clearly?

This is what is surprising to me as well. Is it possible that we got these events because someone is using this socket already?

> * why error has to be processed in poller thread when connect has failed
> with an error other than EAGAIN/EBUSY/EINPROGRESS?

May be reading through http://www.faqs.org/faqs/unix-faq/socket/ might help us answer some of the questions raised above.

REVIEW: http://review.gluster.org/15440 (socket: pollerr event shouldn't trigger socket_connnect_finish) posted (#1) for review on master by Atin Mukherjee (amukherj)

REVIEW: http://review.gluster.org/15440 (socket: pollerr event shouldn't trigger socket_connnect_finish) posted (#2) for review on master by Atin Mukherjee (amukherj)

REVIEW: http://review.gluster.org/15440 (socket: pollerr event shouldn't trigger socket_connnect_finish) posted (#3) for review on master by Atin Mukherjee (amukherj)

How about if we pretend that SO_ERROR doesn't exist?  (As a general rule, something that can't be trusted might as well not exist.)  There must be some other way we can tell whether the socket is actually connected, and generate the correct events or state changes.  Does POLLIN|POLLERR occur in any other cases?  Would an immediate non-blocking read give us ENOTCONN vs. EWOULDBLOCK to distinguish these cases?

REVIEW: http://review.gluster.org/15440 (socket: pollerr event shouldn't trigger socket_connnect_finish) posted (#4) for review on master by Shyamsundar Ranganathan (srangana)

REVIEW: http://review.gluster.org/15440 (socket: pollerr event shouldn't trigger socket_connnect_finish) posted (#5) for review on master by Atin Mukherjee (amukherj)

COMMIT: http://review.gluster.org/15440 committed in master by Shyamsundar Ranganathan (srangana) 
------
commit 12ae3c448c8435decdff27643d5785698ac19dff
Author: Atin Mukherjee <amukherj>
Date:   Thu Sep 8 11:33:59 2016 +0530

    socket: pollerr event shouldn't trigger socket_connnect_finish
    
    If connect fails with any other error than EINPROGRESS we cannot get
    the error status using getsockopt (... SO_ERROR ... ). Hence we need
    to remember the state of connect and take appropriate action in the
    event_handler for the same.
    
    As an added note, a event can come where poll_err is HUP and we have
    poll_in as well (i.e some status was written to the socket), so for
    such cases we need to finish the connect, process the data and then
    the poll_err as is the case in the current code.
    
    Special thanks to Kaushal M & Raghavendra G for figuring out the issue.
    
    Change-Id: Ic45ad59ff8ab1d0a9d2cab2c924ad940b9d38528
    BUG: 1372356
    Signed-off-by: Atin Mukherjee <amukherj>
    Signed-off-by: Shyam <srangana>
    Reviewed-on: http://review.gluster.org/15440
    Smoke: Gluster Build System <jenkins.org>
    NetBSD-regression: NetBSD Build System <jenkins.org>
    CentOS-regression: Gluster Build System <jenkins.org>
    Reviewed-by: Raghavendra G <rgowdapp>

This bug is getting closed because a release has been made available that should address the reported issue. In case the problem is still not fixed with glusterfs-3.10.0, please open a new bug report.

glusterfs-3.10.0 has been announced on the Gluster mailinglists [1], packages for several distributions should become available in the near future. Keep an eye on the Gluster Users mailinglist [2] and the update infrastructure for your distribution.

[1] http://lists.gluster.org/pipermail/gluster-users/2017-February/030119.html
[2] https://www.gluster.org/pipermail/gluster-users/