Staging: octeon-ethernet: Fix race freeing transmit buffers.

The existing code had the following race:

Thread-1                       Thread-2

inc/read in_use
                               inc/read in_use
inc tx_free_list[qos].len
                               inc tx_free_list[qos].len

The actual in_use value was incremented twice, but thread-1 is going
to free memory based on its stale value, and will free one too many
times.  The result is that memory is freed back to the kernel while
its packet is still in the transmit buffer.  If the memory is
overwritten before it is transmitted, the hardware will put a valid
checksum on it and send it out (just like it does with good packets).
If by chance the TCP flags are clobbered but not the addresses or
ports, the result can be a broken TCP stream.

The fix is to track the number of freed packets in a single location
(a Fetch-and-Add Unit register).  That way it can never get out of sync
with itself.

We try to free up to MAX_SKB_TO_FREE (currently 10) buffers at a time.
If fewer are available we adjust the free count with the difference.
The action of claiming buffers to free is atomic so two threads cannot
claim the same buffers.

Signed-off-by: David Daney <ddaney@caviumnetworks.com>
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>

1 files changed, 33 insertions, 23 deletions

diff --git a/drivers/staging/octeon/ethernet-tx.c b/drivers/staging/octeon/ethernet-tx.c
index bfd3dd2..81a8513 100644
--- a/drivers/staging/octeon/ethernet-tx.c
+++ b/drivers/staging/octeon/ethernet-tx.c
@@ -47,6 +47,7 @@
 
 #include "ethernet-defines.h"
 #include "octeon-ethernet.h"
+#include "ethernet-tx.h"
 #include "ethernet-util.h"
 
 #include "cvmx-wqe.h"
@@ -82,8 +83,10 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 	uint64_t old_scratch2;
 	int dropped;
 	int qos;
+	int queue_it_up;
 	struct octeon_ethernet *priv = netdev_priv(dev);
-	int32_t in_use;
+	int32_t skb_to_free;
+	int32_t undo;
 	int32_t buffers_to_free;
 #if REUSE_SKBUFFS_WITHOUT_FREE
 	unsigned char *fpa_head;
@@ -120,15 +123,15 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 		old_scratch2 = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
 
 		/*
-		 * Assume we're going to be able t osend this
-		 * packet. Fetch and increment the number of pending
-		 * packets for output.
+		 * Fetch and increment the number of packets to be
+		 * freed.
 		 */
 		cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH + 8,
 					       FAU_NUM_PACKET_BUFFERS_TO_FREE,
 					       0);
 		cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH,
-					       priv->fau + qos * 4, 1);
+					       priv->fau + qos * 4,
+					       MAX_SKB_TO_FREE);
 	}
 
 	/*
@@ -286,16 +289,30 @@ dont_put_skbuff_in_hw:
 	if (USE_ASYNC_IOBDMA) {
 		/* Get the number of skbuffs in use by the hardware */
 		CVMX_SYNCIOBDMA;
-		in_use = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
+		skb_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 		buffers_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
 	} else {
 		/* Get the number of skbuffs in use by the hardware */
-		in_use = cvmx_fau_fetch_and_add32(priv->fau + qos * 4, 1);
+		skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
+						       MAX_SKB_TO_FREE);
 		buffers_to_free =
 		    cvmx_fau_fetch_and_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
 	}
 
 	/*
+	 * We try to claim MAX_SKB_TO_FREE buffers.  If there were not
+	 * that many available, we have to un-claim (undo) any that
+	 * were in excess.  If skb_to_free is positive we will free
+	 * that many buffers.
+	 */
+	undo = skb_to_free > 0 ?
+		MAX_SKB_TO_FREE : skb_to_free + MAX_SKB_TO_FREE;
+	if (undo > 0)
+		cvmx_fau_atomic_add32(priv->fau+qos*4, -undo);
+	skb_to_free = -skb_to_free > MAX_SKB_TO_FREE ?
+		MAX_SKB_TO_FREE : -skb_to_free;
+
+	/*
 	 * If we're sending faster than the receive can free them then
 	 * don't do the HW free.
 	 */
@@ -330,38 +347,31 @@ dont_put_skbuff_in_hw:
 		cvmx_scratch_write64(CVMX_SCR_SCRATCH + 8, old_scratch2);
 	}
 
+	queue_it_up = 0;
 	if (unlikely(dropped)) {
 		dev_kfree_skb_any(skb);
-		cvmx_fau_atomic_add32(priv->fau + qos * 4, -1);
 		priv->stats.tx_dropped++;
 	} else {
 		if (USE_SKBUFFS_IN_HW) {
 			/* Put this packet on the queue to be freed later */
 			if (pko_command.s.dontfree)
-				skb_queue_tail(&priv->tx_free_list[qos], skb);
-			else {
+				queue_it_up = 1;
+			else
 				cvmx_fau_atomic_add32
 				    (FAU_NUM_PACKET_BUFFERS_TO_FREE, -1);
-				cvmx_fau_atomic_add32(priv->fau + qos * 4, -1);
-			}
 		} else {
 			/* Put this packet on the queue to be freed later */
-			skb_queue_tail(&priv->tx_free_list[qos], skb);
+			queue_it_up = 1;
 		}
 	}
 
-	/* Free skbuffs not in use by the hardware, possibly two at a time */
-	if (skb_queue_len(&priv->tx_free_list[qos]) > in_use) {
+	if (queue_it_up) {
 		spin_lock(&priv->tx_free_list[qos].lock);
-		/*
-		 * Check again now that we have the lock. It might
-		 * have changed.
-		 */
-		if (skb_queue_len(&priv->tx_free_list[qos]) > in_use)
-			dev_kfree_skb(__skb_dequeue(&priv->tx_free_list[qos]));
-		if (skb_queue_len(&priv->tx_free_list[qos]) > in_use)
-			dev_kfree_skb(__skb_dequeue(&priv->tx_free_list[qos]));
+		__skb_queue_tail(&priv->tx_free_list[qos], skb);
+		cvm_oct_free_tx_skbs(priv, skb_to_free, qos, 0);
 		spin_unlock(&priv->tx_free_list[qos].lock);
+	} else {
+		cvm_oct_free_tx_skbs(priv, skb_to_free, qos, 1);
 	}
 
 	return 0;