gianfar: Revive SKB recycling
authorAnton Vorontsov <avorontsov@ru.mvista.com>
Tue, 10 Nov 2009 14:11:10 +0000 (14:11 +0000)
committerDavid S. Miller <davem@davemloft.net>
Thu, 12 Nov 2009 03:03:36 +0000 (19:03 -0800)
Before calling gfar_clean_tx_ring() the driver grabs an irqsave
spinlock, and then tries to recycle skbs. But since
skb_recycle_check() returns 0 with IRQs disabled, we'll never
recycle any skbs.

It appears that gfar_clean_tx_ring() and gfar_start_xmit() are
mostly idependent and can work in parallel, except when they
modify num_txbdfree.

So we can drop the lock from most sections and thus fix the skb
recycling.

Signed-off-by: Anton Vorontsov <avorontsov@ru.mvista.com>
Acked-by: Kumar Gala <galak@kernel.crashing.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/gianfar.c

index fde430a0b84bd5c20a3ba52538daca04f2390925..16def131c390852e4dd5c30c30a936fc682513a2 100644 (file)
@@ -1928,14 +1928,11 @@ static int gfar_start_xmit(struct sk_buff *skb, struct net_device *dev)
        /* total number of fragments in the SKB */
        nr_frags = skb_shinfo(skb)->nr_frags;
 
-       spin_lock_irqsave(&tx_queue->txlock, flags);
-
        /* check if there is space to queue this packet */
        if ((nr_frags+1) > tx_queue->num_txbdfree) {
                /* no space, stop the queue */
                netif_tx_stop_queue(txq);
                dev->stats.tx_fifo_errors++;
-               spin_unlock_irqrestore(&tx_queue->txlock, flags);
                return NETDEV_TX_BUSY;
        }
 
@@ -1998,6 +1995,20 @@ static int gfar_start_xmit(struct sk_buff *skb, struct net_device *dev)
 
        lstatus |= BD_LFLAG(TXBD_CRC | TXBD_READY) | skb_headlen(skb);
 
+       /*
+        * We can work in parallel with gfar_clean_tx_ring(), except
+        * when modifying num_txbdfree. Note that we didn't grab the lock
+        * when we were reading the num_txbdfree and checking for available
+        * space, that's because outside of this function it can only grow,
+        * and once we've got needed space, it cannot suddenly disappear.
+        *
+        * The lock also protects us from gfar_error(), which can modify
+        * regs->tstat and thus retrigger the transfers, which is why we
+        * also must grab the lock before setting ready bit for the first
+        * to be transmitted BD.
+        */
+       spin_lock_irqsave(&tx_queue->txlock, flags);
+
        /*
         * The powerpc-specific eieio() is used, as wmb() has too strong
         * semantics (it requires synchronization between cacheable and
@@ -2225,6 +2236,8 @@ static int gfar_clean_tx_ring(struct gfar_priv_tx_q *tx_queue)
        skb_dirtytx = tx_queue->skb_dirtytx;
 
        while ((skb = tx_queue->tx_skbuff[skb_dirtytx])) {
+               unsigned long flags;
+
                frags = skb_shinfo(skb)->nr_frags;
                lbdp = skip_txbd(bdp, frags, base, tx_ring_size);
 
@@ -2269,7 +2282,9 @@ static int gfar_clean_tx_ring(struct gfar_priv_tx_q *tx_queue)
                        TX_RING_MOD_MASK(tx_ring_size);
 
                howmany++;
+               spin_lock_irqsave(&tx_queue->txlock, flags);
                tx_queue->num_txbdfree += frags + 1;
+               spin_unlock_irqrestore(&tx_queue->txlock, flags);
        }
 
        /* If we freed a buffer, we can restart transmission, if necessary */
@@ -2548,7 +2563,6 @@ static int gfar_poll(struct napi_struct *napi, int budget)
        int tx_cleaned = 0, i, left_over_budget = budget;
        unsigned long serviced_queues = 0;
        int num_queues = 0;
-       unsigned long flags;
 
        num_queues = gfargrp->num_rx_queues;
        budget_per_queue = budget/num_queues;
@@ -2568,14 +2582,7 @@ static int gfar_poll(struct napi_struct *napi, int budget)
                        rx_queue = priv->rx_queue[i];
                        tx_queue = priv->tx_queue[rx_queue->qindex];
 
-                       /* If we fail to get the lock,
-                        * don't bother with the TX BDs */
-                       if (spin_trylock_irqsave(&tx_queue->txlock, flags)) {
-                               tx_cleaned += gfar_clean_tx_ring(tx_queue);
-                               spin_unlock_irqrestore(&tx_queue->txlock,
-                                                       flags);
-                       }
-
+                       tx_cleaned += gfar_clean_tx_ring(tx_queue);
                        rx_cleaned_per_queue = gfar_clean_rx_ring(rx_queue,
                                                        budget_per_queue);
                        rx_cleaned += rx_cleaned_per_queue;