tg3: Add tx BD budgeting code

As the driver breaks large skb fragments into smaller submissions to the
hardware, there is a new danger that BDs might get exhausted before all
fragments have been mapped.  This patch adds code to make sure tx BDs
aren't oversubscribed and flag the condition if it happens.

Signed-off-by: Matt Carlson <mcarlson@broadcom.com>
Reviewed-by: Michael Chan <mchan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/tg3.c

@@ -5924,7 +5924,7 @@ static inline void tg3_tx_set_bd(struct tg3_tx_buffer_desc *txbd,
 	txbd->vlan_tag = (mss << TXD_MSS_SHIFT) | (vlan << TXD_VLAN_TAG_SHIFT);
 }
 
-static bool tg3_tx_frag_set(struct tg3_napi *tnapi, u32 entry,
+static bool tg3_tx_frag_set(struct tg3_napi *tnapi, u32 *entry, u32 *budget,
 			    dma_addr_t map, u32 len, u32 flags,
 			    u32 mss, u32 vlan)
 {
@@ -5940,7 +5940,14 @@ static bool tg3_tx_frag_set(struct tg3_napi *tnapi, u32 entry,
 	if (tg3_40bit_overflow_test(tp, map, len))
 		hwbug = 1;
 
-	tg3_tx_set_bd(&tnapi->tx_ring[entry], map, len, flags, mss, vlan);
+	if (*budget) {
+		tg3_tx_set_bd(&tnapi->tx_ring[*entry], map,
+			      len, flags, mss, vlan);
+		(*budget)--;
+	} else
+		hwbug = 1;
+
+	*entry = NEXT_TX(*entry);
 
 	return hwbug;
 }
@@ -5986,12 +5993,12 @@ static void tg3_tx_skb_unmap(struct tg3_napi *tnapi, u32 entry, int last)
 /* Workaround 4GB and 40-bit hardware DMA bugs. */
 static int tigon3_dma_hwbug_workaround(struct tg3_napi *tnapi,
 				       struct sk_buff *skb,
+				       u32 *entry, u32 *budget,
 				       u32 base_flags, u32 mss, u32 vlan)
 {
 	struct tg3 *tp = tnapi->tp;
 	struct sk_buff *new_skb;
 	dma_addr_t new_addr = 0;
-	u32 entry = tnapi->tx_prod;
 	int ret = 0;
 
 	if (GET_ASIC_REV(tp->pci_chip_rev_id) != ASIC_REV_5701)
@@ -6017,14 +6024,14 @@ static int tigon3_dma_hwbug_workaround(struct tg3_napi *tnapi,
 		} else {
 			base_flags |= TXD_FLAG_END;
 
-			tnapi->tx_buffers[entry].skb = new_skb;
-			dma_unmap_addr_set(&tnapi->tx_buffers[entry],
+			tnapi->tx_buffers[*entry].skb = new_skb;
+			dma_unmap_addr_set(&tnapi->tx_buffers[*entry],
 					   mapping, new_addr);
 
-			if (tg3_tx_frag_set(tnapi, entry, new_addr,
+			if (tg3_tx_frag_set(tnapi, entry, budget, new_addr,
 					    new_skb->len, base_flags,
 					    mss, vlan)) {
-				tg3_tx_skb_unmap(tnapi, entry, 0);
+				tg3_tx_skb_unmap(tnapi, *entry, 0);
 				dev_kfree_skb(new_skb);
 				ret = -1;
 			}
@@ -6086,6 +6093,7 @@ static netdev_tx_t tg3_start_xmit(struct sk_buff *skb, struct net_device *dev)
 {
 	struct tg3 *tp = netdev_priv(dev);
 	u32 len, entry, base_flags, mss, vlan = 0;
+	u32 budget;
 	int i = -1, would_hit_hwbug;
 	dma_addr_t mapping;
 	struct tg3_napi *tnapi;
@@ -6097,12 +6105,14 @@ static netdev_tx_t tg3_start_xmit(struct sk_buff *skb, struct net_device *dev)
 	if (tg3_flag(tp, ENABLE_TSS))
 		tnapi++;
 
+	budget = tg3_tx_avail(tnapi);
+
 	/* We are running in BH disabled context with netif_tx_lock
 	 * and TX reclaim runs via tp->napi.poll inside of a software
 	 * interrupt.  Furthermore, IRQ processing runs lockless so we have
 	 * no IRQ context deadlocks to worry about either.  Rejoice!
 	 */
-	if (unlikely(tg3_tx_avail(tnapi) <= (skb_shinfo(skb)->nr_frags + 1))) {
+	if (unlikely(budget <= (skb_shinfo(skb)->nr_frags + 1))) {
 		if (!netif_tx_queue_stopped(txq)) {
 			netif_tx_stop_queue(txq);
 
@@ -6214,13 +6224,11 @@ static netdev_tx_t tg3_start_xmit(struct sk_buff *skb, struct net_device *dev)
 	if (tg3_flag(tp, 5701_DMA_BUG))
 		would_hit_hwbug = 1;
 
-	if (tg3_tx_frag_set(tnapi, entry, mapping, len, base_flags |
+	if (tg3_tx_frag_set(tnapi, &entry, &budget, mapping, len, base_flags |
 			  ((skb_shinfo(skb)->nr_frags == 0) ? TXD_FLAG_END : 0),
 			    mss, vlan))
 		would_hit_hwbug = 1;
 
-	entry = NEXT_TX(entry);
-
 	/* Now loop through additional data fragments, and queue them. */
 	if (skb_shinfo(skb)->nr_frags > 0) {
 		u32 tmp_mss = mss;
@@ -6246,12 +6254,11 @@ static netdev_tx_t tg3_start_xmit(struct sk_buff *skb, struct net_device *dev)
 			if (pci_dma_mapping_error(tp->pdev, mapping))
 				goto dma_error;
 
-			if (tg3_tx_frag_set(tnapi, entry, mapping, len,
-				  base_flags | ((i == last) ? TXD_FLAG_END : 0),
+			if (tg3_tx_frag_set(tnapi, &entry, &budget, mapping,
+					    len, base_flags |
+					    ((i == last) ? TXD_FLAG_END : 0),
 					    tmp_mss, vlan))
 				would_hit_hwbug = 1;
-
-			entry = NEXT_TX(entry);
 		}
 	}
 
@@ -6261,11 +6268,11 @@ static netdev_tx_t tg3_start_xmit(struct sk_buff *skb, struct net_device *dev)
 		/* If the workaround fails due to memory/mapping
 		 * failure, silently drop this packet.
 		 */
-		if (tigon3_dma_hwbug_workaround(tnapi, skb, base_flags,
-						mss, vlan))
+		entry = tnapi->tx_prod;
+		budget = tg3_tx_avail(tnapi);
+		if (tigon3_dma_hwbug_workaround(tnapi, skb, &entry, &budget,
+						base_flags, mss, vlan))
 			goto out_unlock;
-
-		entry = NEXT_TX(tnapi->tx_prod);
 	}
 
 	skb_tx_timestamp(skb);
@@ -11206,6 +11213,7 @@ static int tg3_run_loopback(struct tg3 *tp, u32 pktsz, int loopback_mode)
 {
 	u32 mac_mode, rx_start_idx, rx_idx, tx_idx, opaque_key;
 	u32 base_flags = 0, mss = 0, desc_idx, coal_now, data_off, val;
+	u32 budget;
 	struct sk_buff *skb, *rx_skb;
 	u8 *tx_data;
 	dma_addr_t map;
@@ -11376,7 +11384,8 @@ static int tg3_run_loopback(struct tg3 *tp, u32 pktsz, int loopback_mode)
 
 	rx_start_idx = rnapi->hw_status->idx[0].rx_producer;
 
-	if (tg3_tx_frag_set(tnapi, tnapi->tx_prod, map, tx_len,
+	budget = tg3_tx_avail(tnapi);
+	if (tg3_tx_frag_set(tnapi, &val, &budget, map, tx_len,
 			    base_flags | TXD_FLAG_END, mss, 0)) {
 		tnapi->tx_buffers[val].skb = NULL;
 		dev_kfree_skb(skb);