net/cpsw: redo rx skb allocation in rx path
authorSebastian Siewior <bigeasy@linutronix.de>
Tue, 23 Apr 2013 07:31:39 +0000 (07:31 +0000)
committerDavid S. Miller <davem@davemloft.net>
Thu, 25 Apr 2013 08:11:50 +0000 (04:11 -0400)
In case that we run into OOM during the allocation of the new rx-skb we
don't get one and we have one skb less than we used to have. If this
continues to happen then we end up with no rx-skbs at all.
This patch changes the following:
- if we fail to allocate the new skb, then we treat the currently
  completed skb as the new one and so drop the currently received data.
- instead of testing multiple times if the device is gone we rely one
  the status field which is set to -ENOSYS in case the channel is going
  down and incomplete requests are purged.
  cpdma_chan_stop() removes most of the packages with -ENOSYS. The
  currently active packet which is removed has the "tear down" bit set.
  So if that bit is set, we send ENOSYS as well otherwise we pass the
  status bits which are required to figure out which of the two possible
  just finished.

Acked-by: Mugunthan V N <mugunthanvnm@ti.com>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/ti/cpsw.c
drivers/net/ethernet/ti/davinci_cpdma.c

index 05f11b854d0aaecc8c377a01ee5e2d27e47302d5..a066c41beac74b6e3302fe2e014264b097eeeda2 100644 (file)
@@ -468,43 +468,36 @@ void cpsw_tx_handler(void *token, int len, int status)
 void cpsw_rx_handler(void *token, int len, int status)
 {
        struct sk_buff          *skb = token;
+       struct sk_buff          *new_skb;
        struct net_device       *ndev = skb->dev;
        struct cpsw_priv        *priv = netdev_priv(ndev);
        int                     ret = 0;
 
        cpsw_dual_emac_src_port_detect(status, priv, ndev, skb);
 
-       /* free and bail if we are shutting down */
-       if (unlikely(!netif_running(ndev)) ||
-                       unlikely(!netif_carrier_ok(ndev))) {
+       if (unlikely(status < 0)) {
+               /* the interface is going down, skbs are purged */
                dev_kfree_skb_any(skb);
                return;
        }
-       if (likely(status >= 0)) {
+
+       new_skb = netdev_alloc_skb_ip_align(ndev, priv->rx_packet_max);
+       if (new_skb) {
                skb_put(skb, len);
                cpts_rx_timestamp(priv->cpts, skb);
                skb->protocol = eth_type_trans(skb, ndev);
                netif_receive_skb(skb);
                priv->stats.rx_bytes += len;
                priv->stats.rx_packets++;
-               skb = NULL;
-       }
-
-       if (unlikely(!netif_running(ndev))) {
-               if (skb)
-                       dev_kfree_skb_any(skb);
-               return;
+       } else {
+               priv->stats.rx_dropped++;
+               new_skb = skb;
        }
 
-       if (likely(!skb)) {
-               skb = netdev_alloc_skb_ip_align(ndev, priv->rx_packet_max);
-               if (WARN_ON(!skb))
-                       return;
-
-               ret = cpdma_chan_submit(priv->rxch, skb, skb->data,
-                                       skb_tailroom(skb), 0);
-       }
-       WARN_ON(ret < 0);
+       ret = cpdma_chan_submit(priv->rxch, new_skb, new_skb->data,
+                       skb_tailroom(new_skb), 0);
+       if (WARN_ON(ret < 0))
+               dev_kfree_skb_any(new_skb);
 }
 
 static irqreturn_t cpsw_interrupt(int irq, void *dev_id)
index 3cc20e7da01ba001e4bad8ddca3994c79b15dec2..6b0a89f6f6a59384355dbe3a8efbd8f2589c3098 100644 (file)
@@ -776,6 +776,7 @@ static int __cpdma_chan_process(struct cpdma_chan *chan)
        struct cpdma_ctlr               *ctlr = chan->ctlr;
        struct cpdma_desc __iomem       *desc;
        int                             status, outlen;
+       int                             cb_status = 0;
        struct cpdma_desc_pool          *pool = ctlr->pool;
        dma_addr_t                      desc_dma;
        unsigned long                   flags;
@@ -811,8 +812,12 @@ static int __cpdma_chan_process(struct cpdma_chan *chan)
        }
 
        spin_unlock_irqrestore(&chan->lock, flags);
+       if (unlikely(status & CPDMA_DESC_TD_COMPLETE))
+               cb_status = -ENOSYS;
+       else
+               cb_status = status;
 
-       __cpdma_chan_free(chan, desc, outlen, status);
+       __cpdma_chan_free(chan, desc, outlen, cb_status);
        return status;
 
 unlock_ret: