ramips: Rework ramips_eth to not require irqsave locking anymore
authorJohn Crispin <john@openwrt.org>
Tue, 17 Jan 2012 11:23:11 +0000 (11:23 +0000)
committerJohn Crispin <john@openwrt.org>
Tue, 17 Jan 2012 11:23:11 +0000 (11:23 +0000)
Previously the tx housekeeping was done in a spin_lock_irqsave critical
section which causes irqs to be disabled during that time. Since the
housekeeping is already prepared to be scheduled as a tasklet process
the housekeeping only in softirq context and revise the locking between
the tx path and the housekeeping tasklet by using a normal spin_lock
which in most situations will be a NOP anyway.

This makes sure that interrupts are only disabled for a short time
since in the worst case the housekeeping might have to free up to 256
skbs.

Signed-off-by: Helmut Schaa <helmut.schaa@googlemail.com>
SVN-Revision: 29762

target/linux/ramips/files/drivers/net/ramips.c

index 63570c7914b6839e141a0bb1c20e077647296f0c..0bc6c0603c45d35a8c9f4a722b6b743107a8fd68 100644 (file)
@@ -215,7 +215,6 @@ ramips_eth_hard_start_xmit(struct sk_buff *skb, struct net_device *dev)
        unsigned long tx;
        unsigned int tx_next;
        unsigned int mapped_addr;
-       unsigned long flags;
 
        if (priv->plat->min_pkt_len) {
                if (skb->len < priv->plat->min_pkt_len) {
@@ -233,7 +232,7 @@ ramips_eth_hard_start_xmit(struct sk_buff *skb, struct net_device *dev)
        mapped_addr = (unsigned int) dma_map_single(NULL, skb->data, skb->len,
                                                    DMA_TO_DEVICE);
        dma_sync_single_for_device(NULL, mapped_addr, skb->len, DMA_TO_DEVICE);
-       spin_lock_irqsave(&priv->page_lock, flags);
+       spin_lock(&priv->page_lock);
        tx = ramips_fe_rr(RAMIPS_TX_CTX_IDX0);
        tx_next = (tx + 1) % NUM_TX_DESC;
 
@@ -250,11 +249,11 @@ ramips_eth_hard_start_xmit(struct sk_buff *skb, struct net_device *dev)
        priv->tx_skb[tx] = skb;
        wmb();
        ramips_fe_wr(tx_next, RAMIPS_TX_CTX_IDX0);
-       spin_unlock_irqrestore(&priv->page_lock, flags);
+       spin_unlock(&priv->page_lock);
        return NETDEV_TX_OK;
 
  out:
-       spin_unlock_irqrestore(&priv->page_lock, flags);
+       spin_unlock(&priv->page_lock);
        dev->stats.tx_dropped++;
        kfree_skb(skb);
        return NETDEV_TX_OK;
@@ -313,6 +312,7 @@ ramips_eth_tx_housekeeping(unsigned long ptr)
        struct net_device *dev = (struct net_device*)ptr;
        struct raeth_priv *priv = netdev_priv(dev);
 
+       spin_lock(&priv->page_lock);
        while ((priv->tx[priv->skb_free_idx].txd2 & TX_DMA_DONE) &&
               (priv->tx_skb[priv->skb_free_idx])) {
                dev_kfree_skb_irq(priv->tx_skb[priv->skb_free_idx]);
@@ -321,6 +321,7 @@ ramips_eth_tx_housekeeping(unsigned long ptr)
                if (priv->skb_free_idx >= NUM_TX_DESC)
                        priv->skb_free_idx = 0;
        }
+       spin_unlock(&priv->page_lock);
 
        ramips_fe_int_enable(RAMIPS_TX_DLY_INT);
 }
@@ -346,8 +347,10 @@ ramips_eth_irq(int irq, void *dev)
                tasklet_schedule(&priv->rx_tasklet);
        }
 
-       if (fe_int & RAMIPS_TX_DLY_INT)
-               ramips_eth_tx_housekeeping((unsigned long)dev);
+       if (fe_int & RAMIPS_TX_DLY_INT) {
+               ramips_fe_int_disable(RAMIPS_TX_DLY_INT);
+               tasklet_schedule(&priv->tx_housekeeping_tasklet);
+       }
 
        return IRQ_HANDLED;
 }