Re: WARN_ON added to rpc_create()

["J. Bruce Fields" <bfields@fieldses.org>]

On Wed, Aug 03, 2016 at 03:40:11PM -0400, Chuck Lever wrote:
> 
> > On Aug 3, 2016, at 1:47 PM, bfields@fieldses.org wrote:
> > 
> > On Wed, Aug 03, 2016 at 11:27:47AM -0400, Chuck Lever wrote:
> >> Hi Bruce-
> >> 
> >> I see that commit 39a9beab5acb83176e8b9a4f0778749a09341f1f
> >> Author:     J. Bruce Fields <bfields@redhat.com>
> >> AuthorDate: Tue May 17 12:38:21 2016 -0400
> >> 
> >>    rpc: share one xps between all backchannels
> >> 
> >> has added this piece of code:
> >> 
> >> @@ -452,10 +452,20 @@ static struct rpc_clnt *rpc_create_xprt(struct rpc_create_args *args,
> >>        struct rpc_clnt *clnt = NULL;
> >>        struct rpc_xprt_switch *xps;
> >> 
> >> -       xps = xprt_switch_alloc(xprt, GFP_KERNEL);
> >> -       if (xps == NULL) {
> >> -               xprt_put(xprt);
> >> -               return ERR_PTR(-ENOMEM);
> >> +       if (args->bc_xprt && args->bc_xprt->xpt_bc_xps) {
> >> +               WARN_ON(args->protocol != XPRT_TRANSPORT_BC_TCP);
> >> +               xps = args->bc_xprt->xpt_bc_xps;
> >> +               xprt_switch_get(xps);
> >> +       } else {
> >> 
> >> 
> >> the WARN_ON here fires on the server whenever I use NFSv4.1 on RDMA.
> >> 
> >> Can you say why it was added? Is there something RPC/RDMA needs to
> >> do to make the code safe?
> > 
> > What is args->protocol in this case?
> > 
> > Digging around...  OK, I missed that BC_TCP and BC_RDMA were defined as
> > OR's of an XPRT_TRANSPORT_BC bit with the identifier of the underlying
> > transport.  That makes sense.
> > 
> > So, I should have just used XPRT_TRANSPORT_BC there--I think all I meant
> > was "is this a backchannel".
> > 
> > Does that fix the problem?
> 
> This simple fix eliminates the log noise:
> 
> diff --git a/net/sunrpc/clnt.c b/net/sunrpc/clnt.c
> index 2808d55..f94caf7 100644
> --- a/net/sunrpc/clnt.c
> +++ b/net/sunrpc/clnt.c
> @@ -520,7 +520,7 @@ struct rpc_clnt *rpc_create(struct rpc_create_args *args)
>         char servername[48];
>  
>         if (args->bc_xprt) {
> -               WARN_ON(args->protocol != XPRT_TRANSPORT_BC_TCP);
> +               WARN_ON(!(args->protocol & XPRT_TRANSPORT_BC));
>                 xprt = args->bc_xprt->xpt_bc_xprt;
>                 if (xprt) {
>                         xprt_get(xprt);
> 
> 
> This code seems to come from:
> 
> commit d50039ea5ee63c589b0434baa5ecf6e5075bb6f9
> Author:     J. Bruce Fields <bfields@redhat.com>
> AuthorDate: Mon May 16 17:03:42 2016 -0400
> 
>     nfsd4/rpc: move backchannel create logic into rpc code
> 
> 
> Where it may have been copied from:
> 
> -static struct rpc_clnt *create_backchannel_client(struct rpc_create_args *args)
> -{
> -       struct rpc_xprt *xprt;
> -
> -       if (args->protocol != XPRT_TRANSPORT_BC_TCP)
> -               return rpc_create(args);
> -
> -       xprt = args->bc_xprt->xpt_bc_xprt;
> -       if (xprt) {
> -               xprt_get(xprt);
> -               return rpc_create_xprt(args, xprt);
> -       }
> -
> -       return rpc_create(args);
> -}
> 
> There's no warning here. In fact, protocol != BC_TCP seems to
> be expected.

The protocol should be BC_TCP (OK, actually just BC) if and only if
bc_xprt is set.

(The BC_TCP case is the 4.1+ case, the other is the 4.0 case.  In the
4.1+ case, the new client uses an existing (client-initiated)
connection, in the 4.0 case, the new client must also have a new
connection.

In the 4.0 case we'll always create a new xprt, in the 4.1 case we might
or might not--depends on whether that particular connection has been
used for a backchannel previously.)

--b.