Temporary hack to work around the "TX overflow" problem in the interim

period before updating to the new driver API (which can cope with
having TX packets in progress).

diff --git a/src/drivers/net/etherfabric.c b/src/drivers/net/etherfabric.c
index 68d2a2a..125dfa5 100644 (file)
--- a/src/drivers/net/etherfabric.c
+++ b/src/drivers/net/etherfabric.c
@@ -173,7 +173,6 @@ struct efab_nic {
        unsigned int eventq_read_ptr;   /* Falcon only */
        unsigned int tx_write_ptr;
        unsigned int rx_write_ptr;
        unsigned int eventq_read_ptr;   /* Falcon only */
        unsigned int tx_write_ptr;
        unsigned int rx_write_ptr;

-       int tx_in_progress;

 
        /** Port 0/1 on the NIC */
        int port;
 
        /** Port 0/1 on the NIC */
        int port;


@@ -2706,7 +2705,6 @@ static int etherfabric_poll ( struct nic *nic, int retrieve ) {
                        /* TX completed - mark as done */
                        DBG ( "TX id %x complete\n",
                              efab->tx_buf.id );
                        /* TX completed - mark as done */
                        DBG ( "TX id %x complete\n",
                              efab->tx_buf.id );

-                       efab->tx_in_progress = 0;

                } else if ( event.type == EFAB_EV_RX ) {
                        /* RX - find corresponding buffer */
                        for ( i = 0 ; i < EFAB_RX_BUFS ; i++ ) {
                } else if ( event.type == EFAB_EV_RX ) {
                        /* RX - find corresponding buffer */
                        for ( i = 0 ; i < EFAB_RX_BUFS ; i++ ) {


@@ -2766,15 +2764,6 @@ static void etherfabric_transmit ( struct nic *nic, const char *dest,
        struct efab_nic *efab = nic->priv_data;
        unsigned int nstype = htons ( type );
 
        struct efab_nic *efab = nic->priv_data;
        unsigned int nstype = htons ( type );
 

-       /* We can only transmit one packet at a time; a TX completion
-        * event must be received before we can transmit the next
-        * packet.  Since there is only one static TX buffer, we don't
-        * worry unduly about overflow, but we report it anyway.
-        */
-       if ( efab->tx_in_progress ) {
-               printf ( "TX overflow!\n" );
-       }
-

        /* Fill TX buffer, pad to ETH_ZLEN */
        memcpy ( efab->tx_buf.addr, dest, ETH_ALEN );
        memcpy ( efab->tx_buf.addr + ETH_ALEN, nic->node_addr, ETH_ALEN );
        /* Fill TX buffer, pad to ETH_ZLEN */
        memcpy ( efab->tx_buf.addr, dest, ETH_ALEN );
        memcpy ( efab->tx_buf.addr + ETH_ALEN, nic->node_addr, ETH_ALEN );


@@ -2789,13 +2778,10 @@ static void etherfabric_transmit ( struct nic *nic, const char *dest,
        /* Push TX descriptor */
        efab_push_tx_buffer ( efab, &efab->tx_buf );
 
        /* Push TX descriptor */
        efab_push_tx_buffer ( efab, &efab->tx_buf );
 

-       /* There is no way to wait for TX complete (i.e. TX buffer
-        * available to re-use for the next transmit) without reading
-        * from the event queue.  We therefore simply leave the TX
-        * buffer marked as "in use" until a TX completion event
-        * happens to be picked up by a call to etherfabric_poll().
+       /* Allow enough time for the packet to be transmitted.  This
+        * is a temporary hack until we update to the new driver API.

         */
         */

-       efab->tx_in_progress = 1;
+       udelay ( 20 );

 
        return;
 }
 
        return;
 }


@@ -2892,9 +2878,6 @@ static int etherfabric_probe ( struct nic *nic, struct pci_device *pci ) {
                return 0;
        memcpy ( nic->node_addr, efab.mac_addr, ETH_ALEN );
 
                return 0;
        memcpy ( nic->node_addr, efab.mac_addr, ETH_ALEN );
 

-       /* hello world */
-       printf ( "Found EtherFabric %s NIC %!\n", pci->name, nic->node_addr );
-

        /* point to NIC specific routines */
        nic->nic_op = &etherfabric_operations;
 
        /* point to NIC specific routines */
        nic->nic_op = &etherfabric_operations;