kernel not recovering from statd port change

Commit Message

J. Bruce Fields Aug. 21, 2014, 9:34 p.m. UTC

While testing server restart somebody noticed that knfsd can't recover
from statd restarting with a new port.

From only a very quick skim of the code it looked like creating the nsm
client with RPC_CLNT_CREATE_AUTOBIND should cause us to call rpcbind
again on connection failures, but that doesn't seem to be working.

Any ideas?  I'll keep looking....

--b.

commit 2c9fb5570fe2
Author: J. Bruce Fields <[e-mail]>
Date:   Wed Aug 20 17:21:32 2014 -0400

    lockd: allow rebinding to statd
    
    During normal operation statd isn't restarted, but it may be if, for
    example, the server is shut down and restarted to simulate a shutdown or
    perform some kind of failover.  In that case the kernel may need to
    query rpcbind again to get statd's new port number.
    
    Symptoms were locking failures after a manual server restart (without
    rebooting the machine), and loopback network traces showing the new
    kernel nfsd attempting to contact statd at its old port number.
    
    This was probably introduced by cb7323fffa85, which first allowed
    reusing the statd rpc client, but it looks like a reference count may
    typically have prevented any symptoms until e498daa81295 "LOCKD: Clear
    ln->nsm_clnt only when ln->nsm_users is zero".
    
    Fixes: cb7323fffa85 "lockd: create and use per-net NSM RPC clients on MON/UNMON requests"
    Signed-off-by: J. Bruce Fields <[e-mail]>

--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to [e-mail]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Comments

Colin Hudler Sept. 4, 2014, 8:42 p.m. UTC | #1

I've been debugging the same thing on an Ubuntu 12.04 server running 
3.8.0-44, and ended up in the same place you are.  Did you find out 
anything more? I have carefully inserted rpc_force_rebind() near 
nlm_client_get, but I don't think it is a good fix for others.

In production servers, I am starting rpc.statd with "--port #####", 
which does seem to solve the problem. NLM (vs NSM) apparently doesn't 
suffer from it.

One thing that puzzles me is that of several hundred NFS clients only a 
handful have a problem getting a lock. The problem clients are running 
3.2 and 2.6.26. The not-problem clients are 3.8 mostly. NFSv3.

On 08/21/2014 04:34 PM, J. Bruce Fields wrote:
> While testing server restart somebody noticed that knfsd can't recover
> from statd restarting with a new port.
>
>  From only a very quick skim of the code it looked like creating the nsm
> client with RPC_CLNT_CREATE_AUTOBIND should cause us to call rpcbind
> again on connection failures, but that doesn't seem to be working.
>
> Any ideas?  I'll keep looking....
>
> --b.
>
> commit 2c9fb5570fe2
> Author: J. Bruce Fields <[e-mail]>
> Date:   Wed Aug 20 17:21:32 2014 -0400
>
>      lockd: allow rebinding to statd
>
>      During normal operation statd isn't restarted, but it may be if, for
>      example, the server is shut down and restarted to simulate a shutdown or
>      perform some kind of failover.  In that case the kernel may need to
>      query rpcbind again to get statd's new port number.
>
>      Symptoms were locking failures after a manual server restart (without
>      rebooting the machine), and loopback network traces showing the new
>      kernel nfsd attempting to contact statd at its old port number.
>
>      This was probably introduced by cb7323fffa85, which first allowed
>      reusing the statd rpc client, but it looks like a reference count may
>      typically have prevented any symptoms until e498daa81295 "LOCKD: Clear
>      ln->nsm_clnt only when ln->nsm_users is zero".
>
>      Fixes: cb7323fffa85 "lockd: create and use per-net NSM RPC clients on MON/UNMON requests"
>      Signed-off-by: J. Bruce Fields <[e-mail]>
>
> diff --git a/fs/lockd/mon.c b/fs/lockd/mon.c
> index 1812f026960c..3bce1d318435 100644
> --- a/fs/lockd/mon.c
> +++ b/fs/lockd/mon.c
> @@ -80,7 +80,8 @@ static struct rpc_clnt *nsm_create(struct net *net)
>   		.program		= &nsm_program,
>   		.version		= NSM_VERSION,
>   		.authflavor		= RPC_AUTH_NULL,
> -		.flags			= RPC_CLNT_CREATE_NOPING,
> +		.flags			= RPC_CLNT_CREATE_NOPING|
> +			                  RPC_CLNT_CREATE_AUTOBIND,
>   	};
>
>   	return rpc_create(&args);
> --
> To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
> the body of a message to [e-mail]
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to [e-mail]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Trond Myklebust Sept. 4, 2014, 9:01 p.m. UTC | #2

On Thu, Aug 21, 2014 at 5:34 PM, J. Bruce Fields <[e-mail]> wrote:
> While testing server restart somebody noticed that knfsd can't recover
> from statd restarting with a new port.
>
> From only a very quick skim of the code it looked like creating the nsm
> client with RPC_CLNT_CREATE_AUTOBIND should cause us to call rpcbind
> again on connection failures, but that doesn't seem to be working.
>
> Any ideas?  I'll keep looking....
>
> --b.
>
> commit 2c9fb5570fe2
> Author: J. Bruce Fields <[e-mail]>
> Date:   Wed Aug 20 17:21:32 2014 -0400
>
>     lockd: allow rebinding to statd
>
>     During normal operation statd isn't restarted, but it may be if, for
>     example, the server is shut down and restarted to simulate a shutdown or
>     perform some kind of failover.  In that case the kernel may need to
>     query rpcbind again to get statd's new port number.
>
>     Symptoms were locking failures after a manual server restart (without
>     rebooting the machine), and loopback network traces showing the new
>     kernel nfsd attempting to contact statd at its old port number.
>
>     This was probably introduced by cb7323fffa85, which first allowed
>     reusing the statd rpc client, but it looks like a reference count may
>     typically have prevented any symptoms until e498daa81295 "LOCKD: Clear
>     ln->nsm_clnt only when ln->nsm_users is zero".
>
>     Fixes: cb7323fffa85 "lockd: create and use per-net NSM RPC clients on MON/UNMON requests"
>     Signed-off-by: J. Bruce Fields <[e-mail]>
>
> diff --git a/fs/lockd/mon.c b/fs/lockd/mon.c
> index 1812f026960c..3bce1d318435 100644
> --- a/fs/lockd/mon.c
> +++ b/fs/lockd/mon.c
> @@ -80,7 +80,8 @@ static struct rpc_clnt *nsm_create(struct net *net)
>                 .program                = &nsm_program,
>                 .version                = NSM_VERSION,
>                 .authflavor             = RPC_AUTH_NULL,
> -               .flags                  = RPC_CLNT_CREATE_NOPING,
> +               .flags                  = RPC_CLNT_CREATE_NOPING|

RPC_CLNT_CREATE_HARDRTRY |

> +                                         RPC_CLNT_CREATE_AUTOBIND,
>         };
>
>         return rpc_create(&args);

Patch

diff --git a/fs/lockd/mon.c b/fs/lockd/mon.c
index 1812f026960c..3bce1d318435 100644
--- a/fs/lockd/mon.c
+++ b/fs/lockd/mon.c
@@ -80,7 +80,8 @@  static struct rpc_clnt *nsm_create(struct net *net)
 		.program		= &nsm_program,
 		.version		= NSM_VERSION,
 		.authflavor		= RPC_AUTH_NULL,
-		.flags			= RPC_CLNT_CREATE_NOPING,
+		.flags			= RPC_CLNT_CREATE_NOPING|
+			                  RPC_CLNT_CREATE_AUTOBIND,
 	};
 
 	return rpc_create(&args);