]> git.kernelconcepts.de Git - karo-tx-linux.git/commitdiff
sfc: handle failure to allocate TSOv2 contexts
authorEdward Cree <ecree@solarflare.com>
Thu, 17 Nov 2016 10:52:07 +0000 (10:52 +0000)
committerDavid S. Miller <davem@davemloft.net>
Fri, 18 Nov 2016 16:55:38 +0000 (11:55 -0500)
If we fail to init the TXQ because of insufficient TSOv2 contexts,
try again with TSOv2 disabled.

Signed-off-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/sfc/ef10.c

index 4b69ed9e63362c624b86fc9717f20358459f54f5..9cbe573563d75738bc35aecff5a905a2c7ba3c70 100644 (file)
@@ -2178,9 +2178,6 @@ static void efx_ef10_tx_init(struct efx_tx_queue *tx_queue)
        /* TSOv2 is a limited resource that can only be configured on a limited
         * number of queues. TSO without checksum offload is not really a thing,
         * so we only enable it for those queues.
-        *
-        * TODO: handle failure to allocate this in the case where we've used
-        * all the queues.
         */
        if (csum_offload && (nic_data->datapath_caps2 &
                        (1 << MC_CMD_GET_CAPABILITIES_V2_OUT_TX_TSO_V2_LBN))) {
@@ -2193,15 +2190,6 @@ static void efx_ef10_tx_init(struct efx_tx_queue *tx_queue)
        MCDI_SET_DWORD(inbuf, INIT_TXQ_IN_TARGET_EVQ, channel->channel);
        MCDI_SET_DWORD(inbuf, INIT_TXQ_IN_LABEL, tx_queue->queue);
        MCDI_SET_DWORD(inbuf, INIT_TXQ_IN_INSTANCE, tx_queue->queue);
-       MCDI_POPULATE_DWORD_3(inbuf, INIT_TXQ_IN_FLAGS,
-                             /* This flag was removed from mcdi_pcol.h for
-                              * the non-_EXT version of INIT_TXQ.  However,
-                              * firmware still honours it.
-                              */
-                             INIT_TXQ_EXT_IN_FLAG_TSOV2_EN, tso_v2,
-                             INIT_TXQ_IN_FLAG_IP_CSUM_DIS, !csum_offload,
-                             INIT_TXQ_IN_FLAG_TCP_CSUM_DIS, !csum_offload);
-
        MCDI_SET_DWORD(inbuf, INIT_TXQ_IN_OWNER_ID, 0);
        MCDI_SET_DWORD(inbuf, INIT_TXQ_IN_PORT_ID, nic_data->vport_id);
 
@@ -2217,10 +2205,30 @@ static void efx_ef10_tx_init(struct efx_tx_queue *tx_queue)
 
        inlen = MC_CMD_INIT_TXQ_IN_LEN(entries);
 
-       rc = efx_mcdi_rpc(efx, MC_CMD_INIT_TXQ, inbuf, inlen,
-                         NULL, 0, NULL);
-       if (rc)
-               goto fail;
+       do {
+               MCDI_POPULATE_DWORD_3(inbuf, INIT_TXQ_IN_FLAGS,
+                               /* This flag was removed from mcdi_pcol.h for
+                                * the non-_EXT version of INIT_TXQ.  However,
+                                * firmware still honours it.
+                                */
+                               INIT_TXQ_EXT_IN_FLAG_TSOV2_EN, tso_v2,
+                               INIT_TXQ_IN_FLAG_IP_CSUM_DIS, !csum_offload,
+                               INIT_TXQ_IN_FLAG_TCP_CSUM_DIS, !csum_offload);
+
+               rc = efx_mcdi_rpc_quiet(efx, MC_CMD_INIT_TXQ, inbuf, inlen,
+                                       NULL, 0, NULL);
+               if (rc == -ENOSPC && tso_v2) {
+                       /* Retry without TSOv2 if we're short on contexts. */
+                       tso_v2 = false;
+                       netif_warn(efx, probe, efx->net_dev,
+                                  "TSOv2 context not available to segment in hardware. TCP performance may be reduced.\n");
+               } else if (rc) {
+                       efx_mcdi_display_error(efx, MC_CMD_INIT_TXQ,
+                                              MC_CMD_INIT_TXQ_EXT_IN_LEN,
+                                              NULL, 0, rc);
+                       goto fail;
+               }
+       } while (rc);
 
        /* A previous user of this TX queue might have set us up the
         * bomb by writing a descriptor to the TX push collector but