Skip to contentSkip to conversation

p54: enforce strict tx_queue limits

PATCH In review Christian Lamparter 2 participants 3 messages Dec 9, 2008 – Dec 10, 2008 Networking drivers (wireless)

lore
Share & tools
lore
#0 lore raw

PATCH 2/2 p54: enforce strict tx_queue limits

The patch fixes an old FIXME in p54pci.c by moving the "queue full" check into the common library, where we can deal with it properly.

---

drivers/net/wireless/p54/p54common.c +21 -1
diff -Nurp a/drivers/net/wireless/p54/p54common.c b/drivers/net/wireless/p54/p54common.c
@@ -889,7 +889,27 @@ static int p54_assign_address(struct iee
 		return -EINVAL;
 
 	spin_lock_irqsave(&priv->tx_queue.lock, flags);
+
 	left = skb_queue_len(&priv->tx_queue);
+	if (unlikely(left >= 28)) {
+		/*
+		 * The tx_queue is nearly full!
+		 * We have throttle normal data traffic, because we must
+		 * have a few spare slots for control frames left.
+		 */
+		ieee80211_stop_queues(dev);
+
+		if (unlikely(left == 32)) {
+			/*
+			 * The tx_queue is now really full.
+			 *
+			 * TODO: check if the device has crashed and reset it.
+			 */
+			spin_unlock_irqrestore(&priv->tx_queue.lock, flags);
+			return -ENOSPC;
+		}
+	}
+
 	while (left--) {
 		u32 hole_size;
 		info = IEEE80211_SKB_CB(entry);
@@ -918,7 +938,7 @@ static int p54_assign_address(struct iee
 	if (!target_skb) {
 		spin_unlock_irqrestore(&priv->tx_queue.lock, flags);
 		ieee80211_stop_queues(dev);
-		return -ENOMEM;
+		return -ENOSPC;
 	}
 
 	info = IEEE80211_SKB_CB(skb);
drivers/net/wireless/p54/p54pci.c +0 -7
diff -Nurp a/drivers/net/wireless/p54/p54pci.c b/drivers/net/wireless/p54/p54pci.c
@@ -332,13 +332,6 @@ static void p54p_tx(struct ieee80211_hw 
 
 	P54P_WRITE(dev_int, cpu_to_le32(ISL38XX_DEV_INT_UPDATE));
 	P54P_READ(dev_int);
-
-	/* FIXME: unlikely to happen because the device usually runs out of
-	   memory before we fill the ring up, but we can make it impossible */
-	if (idx - device_idx > ARRAY_SIZE(ring_control->tx_data) - 2) {
-		p54_free_skb(dev, skb);
-		printk(KERN_INFO "%s: tx overflow.\n", wiphy_name(dev->wiphy));
-	}
 }
 
 static void p54p_stop(struct ieee80211_hw *dev)
#1 lore raw
Quoted 11 lines · Christian Lamparter

The patch fixes an old FIXME in p54pci.c by moving the "queue full" check into the common library, where we can deal with it properly.

---
diff -Nurp a/drivers/net/wireless/p54/p54common.c b/drivers/net/wireless/p54/p54common.c

drivers/net/wireless/p54/p54common.c +0 -0
@@ -889,7 +889,27 @@ static int p54_assign_address(struct iee
 		return -EINVAL;
 

drivers/net/wireless/p54/p54common.c static int p54_assign_address(struct iee · line 900

Quoted 10 lines
drivers/net/wireless/p54/p54common.c +8 -0
 	spin_lock_irqsave(&priv->tx_queue.lock, flags);
+
 	left = skb_queue_len(&priv->tx_queue);
+	if (unlikely(left >= 28)) {
+		/*
+		 * The tx_queue is nearly full!
+		 * We have throttle normal data traffic, because we must
+		 * have a few spare slots for control frames left.
+		 */
+		ieee80211_stop_queues(dev);

I do not intend this as a criticism of the patch, but how do the queues get restarted after this?

Larry

#2 lore raw
Quoted 12 lines · On Wednesday 10 December 2008 03:39:00 Larry Finger wrote:
Quoted 11 lines · Christian Lamparter

The patch fixes an old FIXME in p54pci.c by moving the "queue full" check into the common library, where we can deal with it properly.

---
diff -Nurp a/drivers/net/wireless/p54/p54common.c b/drivers/net/wireless/p54/p54common.c

drivers/net/wireless/p54/p54common.c +0 -0
@@ -889,7 +889,27 @@ static int p54_assign_address(struct iee
 		return -EINVAL;
 
Quoted 14 lines

drivers/net/wireless/p54/p54common.c static int p54_assign_address(struct iee · line 900

Quoted 10 lines
drivers/net/wireless/p54/p54common.c +8 -0
 	spin_lock_irqsave(&priv->tx_queue.lock, flags);
+
 	left = skb_queue_len(&priv->tx_queue);
+	if (unlikely(left >= 28)) {
+		/*
+		 * The tx_queue is nearly full!
+		 * We have throttle normal data traffic, because we must
+		 * have a few spare slots for control frames left.
+		 */
+		ieee80211_stop_queues(dev);

I do not intend this as a criticism of the patch, but how do the queues get restarted after this?

The idea is that p54_free_skb & p54_rx_frame_sent callbacks, which are called whenever the firmware has finished processing a "device control" or wifi frames. (And at that point there're atleast 28 pending frames, so one should eventually get freed unless the firmware crashed!)

These routines unlink the frame form the tx_queue and call p54_wake_free_queues if there's enough space for a new data frame.

Regards,
Chr