sfc: Poll for MCDI completion once before timeout occurs
authorRobert Stonehouse <rstonehouse@solarflare.com>
Wed, 9 Oct 2013 10:52:48 +0000 (11:52 +0100)
committerBen Hutchings <bhutchings@solarflare.com>
Fri, 6 Dec 2013 22:27:55 +0000 (22:27 +0000)
There is an as-yet unexplained bug that sometimes prevents (or delays)
the driver seeing the completion event for a completed MCDI request on
the SFC9120.  The requested configuration change will have happened
but the driver assumes it to have failed, and this can result in
further failures.  We can mitigate this by polling for completion
after unsuccessfully waiting for an event.

Fixes: 8127d661e77f ('sfc: Add support for Solarflare SFC9100 family')
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
drivers/net/ethernet/sfc/mcdi.c

index 9f26e46ee283227275c92c4808351350f5134ddb..4b0bd8a1514dbb7035c7c0562cefc072901e585c 100644 (file)
@@ -630,6 +630,16 @@ int efx_mcdi_rpc_finish(struct efx_nic *efx, unsigned cmd, size_t inlen,
                rc = efx_mcdi_await_completion(efx);
 
        if (rc != 0) {
+               netif_err(efx, hw, efx->net_dev,
+                         "MC command 0x%x inlen %d mode %d timed out\n",
+                         cmd, (int)inlen, mcdi->mode);
+
+               if (mcdi->mode == MCDI_MODE_EVENTS && efx_mcdi_poll_once(efx)) {
+                       netif_err(efx, hw, efx->net_dev,
+                                 "MCDI request was completed without an event\n");
+                       rc = 0;
+               }
+
                /* Close the race with efx_mcdi_ev_cpl() executing just too late
                 * and completing a request we've just cancelled, by ensuring
                 * that the seqno check therein fails.
@@ -638,11 +648,9 @@ int efx_mcdi_rpc_finish(struct efx_nic *efx, unsigned cmd, size_t inlen,
                ++mcdi->seqno;
                ++mcdi->credits;
                spin_unlock_bh(&mcdi->iface_lock);
+       }
 
-               netif_err(efx, hw, efx->net_dev,
-                         "MC command 0x%x inlen %d mode %d timed out\n",
-                         cmd, (int)inlen, mcdi->mode);
-       } else {
+       if (rc == 0) {
                size_t hdr_len, data_len;
 
                /* At the very least we need a memory barrier here to ensure