Date:   Thu, 7 Mar 2019 10:33:02 +0200
From:   Leon Romanovsky <leon@...nel.org>
To:     Steve Wise <swise@...ngridcomputing.com>
Cc:     dsahern@...il.com, stephen@...workplumber.org,
        netdev@...r.kernel.org, linux-rdma@...r.kernel.org
Subject: Re: [PATCH v1 iproute2-next 1/4] rdma: add helper rd_sendrecv_msg()

On Wed, Mar 06, 2019 at 03:50:13PM -0600, Steve Wise wrote:
>
> On 3/4/2019 8:13 AM, Steve Wise wrote:
> > Hey Leon, adding this to rd_recv_msg():
> >
> > @@ -693,10 +693,28 @@ int rd_recv_msg(struct rd *rd, mnl_cb_t callback, void
> > *data, unsigned int seq)
> >                 ret = mnl_cb_run(buf, ret, seq, portid, callback, data);
> >         } while (ret > 0);
> >
> > +       if (ret < 0)
> > +               perror(NULL);
> > +
> >         mnl_socket_close(rd->nl);
> >         return ret;
> >  }
> >
> > Results in unexpected errors being logged when doing a query such as:
> >
> > [root@...vo1 iproute2]# ./rdma/rdma res show qp lqpn 176
> > error: Invalid argument
> > link mlx5_0/1 lqpn 176 type UD state RTS sq-psn 0 comm [ib_core]
> > error: Invalid argument
> > error: No such file or directory
> > error: Invalid argument
> > error: No such file or directory
> >
> > It appears the "invalid argument" errors are due to rdmatool sending a
> > RDMA_NLDEV_CMD_RES_QP_GET command using the doit kernel method to allow
> > querying for just a QP with lqpn = 176.  However, rdmatool isn't passing a
> > port index in the messages that generate the "invalid argument" error from
> > the kernel.  IE you must provide a device index and port index when issuing
> > a doit command vs a dumpit command.  I think.

QPs are per-device and not per-port, so I can't issue "real" port on
multi-port devices.

> >
> > This error was not found because rd_recv_msg() never displayed any errors
> > previously.  Further, the RES_FUNC() massive macro has code that will retry
> > a failed doit call with a dumpit call.  I think _##name() should distinguish
> > between failures reported by the kernel doit function vs failures because no
> > doit function exists.  Not sure how to support that.

We can suppress prints from failures to find .doit, by adding extra
parameter to _res_send_idx_msg(): for example _res_send_idx_msg(..., no_error_print);
and provide this "no_error_print" to rd_recv_msg(), through "void
*data".

> >
> >
> >         static inline int _##name(struct rd *rd)
> > \
> >         {
> > \
> >                 uint32_t idx;
> > \
> >                 int ret;
> > \
> >                 if (id) {
> > \
> >                         ret = rd_doit_index(rd, &idx);
> > \
> >                         if (ret) {
> > \
> >                                 ret = _res_send_idx_msg(rd, command,
> > \
> >                                                         name##_idx_parse_cb,
> > \
> >                                                         idx, id);
> > \
> >                                 if (!ret)
> > \
> >                                         return ret;
> > \
> >                                 /* Fallback for old systems without .doit
> > callbacks */ \
> >                         }
> > \
> >                 }
> > \
> >                 return _res_send_msg(rd, command, name##_parse_cb);
> > \
> >         }
> > \
> >
> >
> >
> > The "no such file or dir" errors are being returned because, in my setup,
> > there are 2 other links that do not have lqpn 176.   So there are 2 issues
> > uncovered by adding generic printing of errors in rd_recv_msg()
> >
> > 1) the doit code in rdmatool is generating requests for a doit method in the
> > kernel w/o providing a port index.
> > 2) some paths in rdmatool should not print "benign" errors like filtering on
> > a GET command causing a "does not exist" error returned by the kernel doit
> > func.
> >
> > #1 is a bug, IMO.  Can you propose a fix?
> > #2 could be solved by adding an error callback func passed to rd_recv_msg().
> > Then the RES_FUNC() functions could parse errors like "no such file or dir"
> > when doing a filtered query and silently drop them.  And functions like
> > dev_set_name() would display all errors returned because there are no
> > expected errors other than "success".
> >
> > Steve.
> >
>
> Hey Leon, you've been quiet. :)   Thoughts?

I missed your email, sorry.

>
> Thanks,
>
> Steve.
>
>

Download attachment "signature.asc" of type "application/pgp-signature" (802 bytes)

Powered by blists - more mailing lists