svcrdma: set XPT_CONG_CTRL flag for bc xprt
diff mbox

Message ID 20170326231254.1319.26075.stgit@manet.1015granger.net
State New
Headers show

Commit Message

Chuck Lever March 26, 2017, 11:27 p.m. UTC
Same change as Kinglong Mee's fix for the TCP backchannel service.

Fixes: 5283b03ee5cd ("nfs/nfsd/sunrpc: enforce transport...")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
---
Some (perhaps late) review comments on 5283b03ee5cd:

I have reservations about returning RPC_PROG_MISMATCH in this case.
RPC_PROG_UNAVAIL is more sensible. But the use of UDP with NFSv4 is
not an RPC-level error, thus reporting the problem here seems like a
layering violation.

I'm not sure why an explicit check is needed: if the server isn't
listening on UDP, wouldn't clients see a transport-level rejection
(like ECONNREFUSED)?

Are we certain that all client implementations (including
backchannel clients) will do something useful when presented with
such a rejection? At least in the backchannel case, the Linux server
had no idea what to do with RPC_PROG_MISMATCH on the backchannel.
The workload stopped dead, no error report anywhere.

 net/sunrpc/xprtrdma/svc_rdma_transport.c |    1 +
 1 file changed, 1 insertion(+)


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

Comments

Jeff Layton March 27, 2017, 1:21 a.m. UTC | #1
On Sun, 2017-03-26 at 19:27 -0400, Chuck Lever wrote:
> Same change as Kinglong Mee's fix for the TCP backchannel service.
> 

Good catch. I guess I didn't do a good job of hunting down all of the
transports where this needed to be set. I'll give them another pass
again tomorrow to make sure I didn't miss any others.

> Fixes: 5283b03ee5cd ("nfs/nfsd/sunrpc: enforce transport...")
> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
> ---
> Some (perhaps late) review comments on 5283b03ee5cd:
> 
> I have reservations about returning RPC_PROG_MISMATCH in this case.
> RPC_PROG_UNAVAIL is more sensible. But the use of UDP with NFSv4 is
> not an RPC-level error, thus reporting the problem here seems like a
> layering violation.
> 
> I'm not sure why an explicit check is needed: if the server isn't
> listening on UDP, wouldn't clients see a transport-level rejection
> (like ECONNREFUSED)?
> 

Sure, if the server isn't listening on UDP...

The point of that patch is to enforce not allowing v4 over UDP when the
server is listening on UDP to serve earlier versions.

As far as the error...From RFC 5531:

            PROG_UNAVAIL  = 1, /* remote hasn't exported program  */
            PROG_MISMATCH = 2, /* remote can't support version #  */

Consider the case where the server is listening on both TCP and UDP,
and is serving both v3 and v4. Someone tries to send a v4 RPC over UDP.

The RPC program in that case (nfs) is supported over UDP, but the
version (v4) is not. So I disagree here. PROG_MISMATCH seems like the
better fit to me.

> Are we certain that all client implementations (including
> backchannel clients) will do something useful when presented with
> such a rejection? At least in the backchannel case, the Linux server
> had no idea what to do with RPC_PROG_MISMATCH on the backchannel.
> The workload stopped dead, no error report anywhere.
> 

Ouch. I think this would get translated into EPROTONOSUPPORT in the
client code. That should have ended up with nfsd4_mark_cb_down being
called with that error?...but I think that function may be effectively
neutered:

static void warn_no_callback_path(struct nfs4_client *clp, int reason)
{
        dprintk("NFSD: warning: no callback path to client %.*s: error %d\n",
                (int)clp->cl_name.len, clp->cl_name.data, reason);
}

Note that it emits a dprintk instead of a printk. Should we promote
that to something more visible?

>  net/sunrpc/xprtrdma/svc_rdma_transport.c |    1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/net/sunrpc/xprtrdma/svc_rdma_transport.c b/net/sunrpc/xprtrdma/svc_rdma_transport.c
> index c13a5c3..fc8f14c 100644
> --- a/net/sunrpc/xprtrdma/svc_rdma_transport.c
> +++ b/net/sunrpc/xprtrdma/svc_rdma_transport.c
> @@ -127,6 +127,7 @@ static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *serv,
>  	xprt = &cma_xprt->sc_xprt;
>  
>  	svc_xprt_init(net, &svc_rdma_bc_class, xprt, serv);
> +	set_bit(XPT_CONG_CTRL, &xprt->xpt_flags);
>  	serv->sv_bc_xprt = xprt;
>  
>  	dprintk("svcrdma: %s(%p)\n", __func__, xprt);
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
> the body of a message to majordomo@vger.kernel.org
> 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 majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Chuck Lever March 27, 2017, 2:38 a.m. UTC | #2
Hey Jeff-


> On Mar 26, 2017, at 9:21 PM, Jeff Layton <jlayton@poochiereds.net> wrote:
> 
>> On Sun, 2017-03-26 at 19:27 -0400, Chuck Lever wrote:
>> Same change as Kinglong Mee's fix for the TCP backchannel service.
>> 
> 
> Good catch. I guess I didn't do a good job of hunting down all of the
> transports where this needed to be set. I'll give them another pass
> again tomorrow to make sure I didn't miss any others.
> 
>> Fixes: 5283b03ee5cd ("nfs/nfsd/sunrpc: enforce transport...")
>> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
>> ---
>> Some (perhaps late) review comments on 5283b03ee5cd:
>> 
>> I have reservations about returning RPC_PROG_MISMATCH in this case.
>> RPC_PROG_UNAVAIL is more sensible. But the use of UDP with NFSv4 is
>> not an RPC-level error, thus reporting the problem here seems like a
>> layering violation.
>> 
>> I'm not sure why an explicit check is needed: if the server isn't
>> listening on UDP, wouldn't clients see a transport-level rejection
>> (like ECONNREFUSED)?
>> 
> 
> Sure, if the server isn't listening on UDP...
> 
> The point of that patch is to enforce not allowing v4 over UDP when the
> server is listening on UDP to serve earlier versions.
> 
> As far as the error...From RFC 5531:
> 
>             PROG_UNAVAIL  = 1, /* remote hasn't exported program  */
>             PROG_MISMATCH = 2, /* remote can't support version #  */
> 
> Consider the case where the server is listening on both TCP and UDP,
> and is serving both v3 and v4. Someone tries to send a v4 RPC over UDP.
> 
> The RPC program in that case (nfs) is supported over UDP, but the
> version (v4) is not. So I disagree here. PROG_MISMATCH seems like the
> better fit to me.

Then the server should report the correct version range in the
rejection. The RPC response I saw on the wire claimed that 4
was the maximum supported version.


>> Are we certain that all client implementations (including
>> backchannel clients) will do something useful when presented with
>> such a rejection? At least in the backchannel case, the Linux server
>> had no idea what to do with RPC_PROG_MISMATCH on the backchannel.
>> The workload stopped dead, no error report anywhere.
>> 
> 
> Ouch. I think this would get translated into EPROTONOSUPPORT in the
> client code. That should have ended up with nfsd4_mark_cb_down being
> called with that error?...but I think that function may be effectively
> neutered:
> 
> static void warn_no_callback_path(struct nfs4_client *clp, int reason)
> {
>        dprintk("NFSD: warning: no callback path to client %.*s: error %d\n",
>                (int)clp->cl_name.len, clp->cl_name.data, reason);
> }
> 
> Note that it emits a dprintk instead of a printk. Should we promote
> that to something more visible?

You don't want a warning if the client never provided a
callback path. But if one was provided, and it disappears,
that might be useful to know.

OTOH some might blanch at the log flood, should something
else go wrong.

An error counter might be the least we can do, if not a
one-shot pr_warn.


>> net/sunrpc/xprtrdma/svc_rdma_transport.c |    1 +
>> 1 file changed, 1 insertion(+)
>> 
>> diff --git a/net/sunrpc/xprtrdma/svc_rdma_transport.c b/net/sunrpc/xprtrdma/svc_rdma_transport.c
>> index c13a5c3..fc8f14c 100644
>> --- a/net/sunrpc/xprtrdma/svc_rdma_transport.c
>> +++ b/net/sunrpc/xprtrdma/svc_rdma_transport.c
>> @@ -127,6 +127,7 @@ static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *serv,
>>    xprt = &cma_xprt->sc_xprt;
>> 
>>    svc_xprt_init(net, &svc_rdma_bc_class, xprt, serv);
>> +    set_bit(XPT_CONG_CTRL, &xprt->xpt_flags);
>>    serv->sv_bc_xprt = xprt;
>> 
>>    dprintk("svcrdma: %s(%p)\n", __func__, xprt);
>> 
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
>> the body of a message to majordomo@vger.kernel.org
>> 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 majordomo@vger.kernel.org
> 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 majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Chuck Lever March 27, 2017, 2:41 a.m. UTC | #3
> On Mar 26, 2017, at 10:38 PM, Chuck Lever <chucklever@gmail.com> wrote:
> 
> Hey Jeff-
> 
> 
>>> On Mar 26, 2017, at 9:21 PM, Jeff Layton <jlayton@poochiereds.net> wrote:
>>> 
>>> On Sun, 2017-03-26 at 19:27 -0400, Chuck Lever wrote:
>>> Same change as Kinglong Mee's fix for the TCP backchannel service.
>>> 
>> 
>> Good catch. I guess I didn't do a good job of hunting down all of the
>> transports where this needed to be set. I'll give them another pass
>> again tomorrow to make sure I didn't miss any others.
>> 
>>> Fixes: 5283b03ee5cd ("nfs/nfsd/sunrpc: enforce transport...")
>>> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
>>> ---
>>> Some (perhaps late) review comments on 5283b03ee5cd:
>>> 
>>> I have reservations about returning RPC_PROG_MISMATCH in this case.
>>> RPC_PROG_UNAVAIL is more sensible. But the use of UDP with NFSv4 is
>>> not an RPC-level error, thus reporting the problem here seems like a
>>> layering violation.
>>> 
>>> I'm not sure why an explicit check is needed: if the server isn't
>>> listening on UDP, wouldn't clients see a transport-level rejection
>>> (like ECONNREFUSED)?
>>> 
>> 
>> Sure, if the server isn't listening on UDP...
>> 
>> The point of that patch is to enforce not allowing v4 over UDP when the
>> server is listening on UDP to serve earlier versions.
>> 
>> As far as the error...From RFC 5531:
>> 
>>            PROG_UNAVAIL  = 1, /* remote hasn't exported program  */
>>            PROG_MISMATCH = 2, /* remote can't support version #  */
>> 
>> Consider the case where the server is listening on both TCP and UDP,
>> and is serving both v3 and v4. Someone tries to send a v4 RPC over UDP.
>> 
>> The RPC program in that case (nfs) is supported over UDP, but the
>> version (v4) is not. So I disagree here. PROG_MISMATCH seems like the
>> better fit to me.
> 
> Then the server should report the correct version range in the
> rejection. The RPC response I saw on the wire claimed that 4
> was the maximum supported version.

Of course, versions 2 and 3 do not make sense for
the backchannel. So I'm not sure what you would report
in that case.


>>> Are we certain that all client implementations (including
>>> backchannel clients) will do something useful when presented with
>>> such a rejection? At least in the backchannel case, the Linux server
>>> had no idea what to do with RPC_PROG_MISMATCH on the backchannel.
>>> The workload stopped dead, no error report anywhere.
>>> 
>> 
>> Ouch. I think this would get translated into EPROTONOSUPPORT in the
>> client code. That should have ended up with nfsd4_mark_cb_down being
>> called with that error?...but I think that function may be effectively
>> neutered:
>> 
>> static void warn_no_callback_path(struct nfs4_client *clp, int reason)
>> {
>>       dprintk("NFSD: warning: no callback path to client %.*s: error %d\n",
>>               (int)clp->cl_name.len, clp->cl_name.data, reason);
>> }
>> 
>> Note that it emits a dprintk instead of a printk. Should we promote
>> that to something more visible?
> 
> You don't want a warning if the client never provided a
> callback path. But if one was provided, and it disappears,
> that might be useful to know.
> 
> OTOH some might blanch at the log flood, should something
> else go wrong.
> 
> An error counter might be the least we can do, if not a
> one-shot pr_warn.
> 
> 
>>> net/sunrpc/xprtrdma/svc_rdma_transport.c |    1 +
>>> 1 file changed, 1 insertion(+)
>>> 
>>> diff --git a/net/sunrpc/xprtrdma/svc_rdma_transport.c b/net/sunrpc/xprtrdma/svc_rdma_transport.c
>>> index c13a5c3..fc8f14c 100644
>>> --- a/net/sunrpc/xprtrdma/svc_rdma_transport.c
>>> +++ b/net/sunrpc/xprtrdma/svc_rdma_transport.c
>>> @@ -127,6 +127,7 @@ static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *serv,
>>>   xprt = &cma_xprt->sc_xprt;
>>> 
>>>   svc_xprt_init(net, &svc_rdma_bc_class, xprt, serv);
>>> +    set_bit(XPT_CONG_CTRL, &xprt->xpt_flags);
>>>   serv->sv_bc_xprt = xprt;
>>> 
>>>   dprintk("svcrdma: %s(%p)\n", __func__, xprt);
>>> 
>>> --
>>> To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
>>> the body of a message to majordomo@vger.kernel.org
>>> 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 majordomo@vger.kernel.org
>> 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 majordomo@vger.kernel.org
> 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 majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Jeff Layton March 27, 2017, 11:07 a.m. UTC | #4
On Sun, 2017-03-26 at 22:41 -0400, Chuck Lever wrote:
> > On Mar 26, 2017, at 10:38 PM, Chuck Lever <chucklever@gmail.com> wrote:
> > 
> > Hey Jeff-
> > 
> > 
> > > > On Mar 26, 2017, at 9:21 PM, Jeff Layton <jlayton@poochiereds.net> wrote:
> > > > 
> > > > On Sun, 2017-03-26 at 19:27 -0400, Chuck Lever wrote:
> > > > Same change as Kinglong Mee's fix for the TCP backchannel service.
> > > > 
> > > 
> > > Good catch. I guess I didn't do a good job of hunting down all of the
> > > transports where this needed to be set. I'll give them another pass
> > > again tomorrow to make sure I didn't miss any others.
> > > 
> > > > Fixes: 5283b03ee5cd ("nfs/nfsd/sunrpc: enforce transport...")
> > > > Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
> > > > ---
> > > > Some (perhaps late) review comments on 5283b03ee5cd:
> > > > 
> > > > I have reservations about returning RPC_PROG_MISMATCH in this case.
> > > > RPC_PROG_UNAVAIL is more sensible. But the use of UDP with NFSv4 is
> > > > not an RPC-level error, thus reporting the problem here seems like a
> > > > layering violation.
> > > > 
> > > > I'm not sure why an explicit check is needed: if the server isn't
> > > > listening on UDP, wouldn't clients see a transport-level rejection
> > > > (like ECONNREFUSED)?
> > > > 
> > > 
> > > Sure, if the server isn't listening on UDP...
> > > 
> > > The point of that patch is to enforce not allowing v4 over UDP when the
> > > server is listening on UDP to serve earlier versions.
> > > 
> > > As far as the error...From RFC 5531:
> > > 
> > >            PROG_UNAVAIL  = 1, /* remote hasn't exported program  */
> > >            PROG_MISMATCH = 2, /* remote can't support version #  */
> > > 
> > > Consider the case where the server is listening on both TCP and UDP,
> > > and is serving both v3 and v4. Someone tries to send a v4 RPC over UDP.
> > > 
> > > The RPC program in that case (nfs) is supported over UDP, but the
> > > version (v4) is not. So I disagree here. PROG_MISMATCH seems like the
> > > better fit to me.
> > 
> > Then the server should report the correct version range in the
> > rejection. The RPC response I saw on the wire claimed that 4
> > was the maximum supported version.
> 
> Of course, versions 2 and 3 do not make sense for
> the backchannel. So I'm not sure what you would report
> in that case.
> 

Yeah, that's clearly a bug. The problem is that we currently track
min/max versions on a per-program basis, but really we need to track
them per-program + per-transport.

Another way to fix it would be to set that info more dynamically at the
time of the error. Walk the pg_vers array and if we're on a non-
congestion control transport, skip any entries that require it.

That said, I'm not aware of anything that actually uses that version
info, aside from people poking around at it with rpcinfo. Is there
anything that actually does? If not, then I'm not terribly concerned
about getting it right, though it would be nice to have.
J . Bruce Fields March 29, 2017, 1:22 a.m. UTC | #5
On Sun, Mar 26, 2017 at 09:21:39PM -0400, Jeff Layton wrote:
> On Sun, 2017-03-26 at 19:27 -0400, Chuck Lever wrote:
> > Are we certain that all client implementations (including
> > backchannel clients) will do something useful when presented with
> > such a rejection? At least in the backchannel case, the Linux server
> > had no idea what to do with RPC_PROG_MISMATCH on the backchannel.
> > The workload stopped dead, no error report anywhere.
> > 
> 
> Ouch. I think this would get translated into EPROTONOSUPPORT in the
> client code. That should have ended up with nfsd4_mark_cb_down being
> called with that error?...but I think that function may be effectively
> neutered:

Are we worrying now about a server that tries to open an NFSv4.0
callback connection using UDP?

That would be a very broken server.  And broken in a way that I think is
pretty unlikely to actually happen in practice.

Maybe I'm missing something.

> static void warn_no_callback_path(struct nfs4_client *clp, int reason)
> {
>         dprintk("NFSD: warning: no callback path to client %.*s: error %d\n",
>                 (int)clp->cl_name.len, clp->cl_name.data, reason);
> }

In NFSv4.0 a failing callback connection is absolutely normal (e.g. if
the client is behind a firewall).  We might want to provide some better
diagnostics to help people figure out why a given client isn't getting
delegations, but we don't want to log this by default.

Even in the 4.1 case I wonder if some pretty common failures (e.g.
losing contact with the client) might get noticed by the callback code
first.

So, dprintk is right here.

--b.

> 
> Note that it emits a dprintk instead of a printk. Should we promote
> that to something more visible?
> 
> >  net/sunrpc/xprtrdma/svc_rdma_transport.c |    1 +
> >  1 file changed, 1 insertion(+)
> > 
> > diff --git a/net/sunrpc/xprtrdma/svc_rdma_transport.c b/net/sunrpc/xprtrdma/svc_rdma_transport.c
> > index c13a5c3..fc8f14c 100644
> > --- a/net/sunrpc/xprtrdma/svc_rdma_transport.c
> > +++ b/net/sunrpc/xprtrdma/svc_rdma_transport.c
> > @@ -127,6 +127,7 @@ static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *serv,
> >  	xprt = &cma_xprt->sc_xprt;
> >  
> >  	svc_xprt_init(net, &svc_rdma_bc_class, xprt, serv);
> > +	set_bit(XPT_CONG_CTRL, &xprt->xpt_flags);
> >  	serv->sv_bc_xprt = xprt;
> >  
> >  	dprintk("svcrdma: %s(%p)\n", __func__, xprt);
> > 
> > --
> > To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
> > the body of a message to majordomo@vger.kernel.org
> > 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 majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
J . Bruce Fields March 29, 2017, 1:27 a.m. UTC | #6
On Sun, Mar 26, 2017 at 07:27:35PM -0400, Chuck Lever wrote:
> Same change as Kinglong Mee's fix for the TCP backchannel service.

Thanks, applying!

And, apologies to all, I've let a few patches pile up the last week,
hoping to catch up and pass along a bugfix pull to Linus before next
week (when I'll need to be mostly offline again).

--b.

> 
> Fixes: 5283b03ee5cd ("nfs/nfsd/sunrpc: enforce transport...")
> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
> ---
> Some (perhaps late) review comments on 5283b03ee5cd:
> 
> I have reservations about returning RPC_PROG_MISMATCH in this case.
> RPC_PROG_UNAVAIL is more sensible. But the use of UDP with NFSv4 is
> not an RPC-level error, thus reporting the problem here seems like a
> layering violation.
> 
> I'm not sure why an explicit check is needed: if the server isn't
> listening on UDP, wouldn't clients see a transport-level rejection
> (like ECONNREFUSED)?
> 
> Are we certain that all client implementations (including
> backchannel clients) will do something useful when presented with
> such a rejection? At least in the backchannel case, the Linux server
> had no idea what to do with RPC_PROG_MISMATCH on the backchannel.
> The workload stopped dead, no error report anywhere.
> 
>  net/sunrpc/xprtrdma/svc_rdma_transport.c |    1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/net/sunrpc/xprtrdma/svc_rdma_transport.c b/net/sunrpc/xprtrdma/svc_rdma_transport.c
> index c13a5c3..fc8f14c 100644
> --- a/net/sunrpc/xprtrdma/svc_rdma_transport.c
> +++ b/net/sunrpc/xprtrdma/svc_rdma_transport.c
> @@ -127,6 +127,7 @@ static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *serv,
>  	xprt = &cma_xprt->sc_xprt;
>  
>  	svc_xprt_init(net, &svc_rdma_bc_class, xprt, serv);
> +	set_bit(XPT_CONG_CTRL, &xprt->xpt_flags);
>  	serv->sv_bc_xprt = xprt;
>  
>  	dprintk("svcrdma: %s(%p)\n", __func__, xprt);
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Jeff Layton March 29, 2017, 11:02 a.m. UTC | #7
On Tue, 2017-03-28 at 21:22 -0400, J. Bruce Fields wrote:
> On Sun, Mar 26, 2017 at 09:21:39PM -0400, Jeff Layton wrote:
> > On Sun, 2017-03-26 at 19:27 -0400, Chuck Lever wrote:
> > > Are we certain that all client implementations (including
> > > backchannel clients) will do something useful when presented with
> > > such a rejection? At least in the backchannel case, the Linux server
> > > had no idea what to do with RPC_PROG_MISMATCH on the backchannel.
> > > The workload stopped dead, no error report anywhere.
> > > 
> > 
> > Ouch. I think this would get translated into EPROTONOSUPPORT in the
> > client code. That should have ended up with nfsd4_mark_cb_down being
> > called with that error?...but I think that function may be effectively
> > neutered:
> 
> Are we worrying now about a server that tries to open an NFSv4.0
> callback connection using UDP?
> 
> That would be a very broken server.  And broken in a way that I think is
> pretty unlikely to actually happen in practice.
> 
> Maybe I'm missing something.
> 

The client is what was broken here. Chuck's initial patch is correct --
we do need to flag the RDMA backchannel xprt with XPT_CONG_CTRL. When
that wasn't there, the RDMA bc was rejecting callbacks from the sever.

The question at this point is whether we need to do anything further. It
sounds like the answer to that question is "no" and I'm fine with that.


> > static void warn_no_callback_path(struct nfs4_client *clp, int reason)
> > {
> >         dprintk("NFSD: warning: no callback path to client %.*s: error %d\n",
> >                 (int)clp->cl_name.len, clp->cl_name.data, reason);
> > }
> 
> In NFSv4.0 a failing callback connection is absolutely normal (e.g. if
> the client is behind a firewall).  We might want to provide some better
> diagnostics to help people figure out why a given client isn't getting
> delegations, but we don't want to log this by default.
> 
> Even in the 4.1 case I wonder if some pretty common failures (e.g.
> losing contact with the client) might get noticed by the callback code
> first.
> 
> So, dprintk is right here.
> 
> --b.

> > 
> > Note that it emits a dprintk instead of a printk. Should we promote
> > that to something more visible?
> > 
> > >  net/sunrpc/xprtrdma/svc_rdma_transport.c |    1 +
> > >  1 file changed, 1 insertion(+)
> > > 
> > > diff --git a/net/sunrpc/xprtrdma/svc_rdma_transport.c b/net/sunrpc/xprtrdma/svc_rdma_transport.c
> > > index c13a5c3..fc8f14c 100644
> > > --- a/net/sunrpc/xprtrdma/svc_rdma_transport.c
> > > +++ b/net/sunrpc/xprtrdma/svc_rdma_transport.c
> > > @@ -127,6 +127,7 @@ static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *serv,
> > >  	xprt = &cma_xprt->sc_xprt;
> > >  
> > >  	svc_xprt_init(net, &svc_rdma_bc_class, xprt, serv);
> > > +	set_bit(XPT_CONG_CTRL, &xprt->xpt_flags);
> > >  	serv->sv_bc_xprt = xprt;
> > >  
> > >  	dprintk("svcrdma: %s(%p)\n", __func__, xprt);
> > > 
> > > --
> > > To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
> > > the body of a message to majordomo@vger.kernel.org
> > > More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch
diff mbox

diff --git a/net/sunrpc/xprtrdma/svc_rdma_transport.c b/net/sunrpc/xprtrdma/svc_rdma_transport.c
index c13a5c3..fc8f14c 100644
--- a/net/sunrpc/xprtrdma/svc_rdma_transport.c
+++ b/net/sunrpc/xprtrdma/svc_rdma_transport.c
@@ -127,6 +127,7 @@  static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *serv,
 	xprt = &cma_xprt->sc_xprt;
 
 	svc_xprt_init(net, &svc_rdma_bc_class, xprt, serv);
+	set_bit(XPT_CONG_CTRL, &xprt->xpt_flags);
 	serv->sv_bc_xprt = xprt;
 
 	dprintk("svcrdma: %s(%p)\n", __func__, xprt);