On Fri, Dec 14, 2018 at 9:18 PM William Allen Simpson <william.allen.simpson@gmail.com> wrote:

On 12/11/18 12:30 AM, gaurav gangalwar wrote:
> As we can see in the log highlighted we are still left with one ref even after SVC_DESTROY in UMNT and all SVC_RELEASE.
>
[...]

> I tried with SVC_RELEASE instead of SVC_DESTROY in same patch, it worked at least for connection on which UMNT came, we are releasing the xprt for it.
>
Then it would seem that there already is another SVC_DESTROY somewhere else. The whole
point of SVC_DESTROY is once-only. So your SVC_DESTROY didn't do anything.

So do we need any fix in Ganesha code for this issue?

We do SVC_DESTROY in svc_vc_recv if there is connection close from client, so I don't think we need to do it on UMNT in Ganesha.

[...]

> As we can see in the logs refcnt is going to zero for the connection on which UMNT came.
> But there are logs for other connections also, which we are not cleaning up as there is no UMNT on them, but we are polling on them and they are getting closed,
> I just have single client and running same script, looks like client is opening multiple connections and closing them on UMNT.
>
What client is running multiple connections on the same fd?

There are no multiple connections on same fd, there are multiple fds/connections from client which are getting closed without UMNT call on them so they will not get cleaned from server. I pasted the logs for it.

Point is doing SVC_RELEASE in UMNT is not fixing the issue, we are still left with uncleaned connections.

As per discussion here

https://github.com/nfs-ganesha/ntirpc/pull/160

Looks like we need fix in tirpc only.

Regards,

Gaurav