igbvf: Make next_to_watch a pointer and adjust memory barriers to avoid races

This change is meant to address several race issues that become possible
because next_to_watch could possibly be set to a value that shows that the
descriptor is done when it is not.  In order to correct that we instead make
next_to_watch a pointer that is set to NULL during cleanup, and set to the
eop_desc after the descriptor rings have been written.

To enforce proper ordering the next_to_watch pointer is not set until after
a wmb writing the values to the last descriptor in a transmit.  In order to
guarantee that the descriptor is not read until after the eop_desc we use the
read_barrier_depends which is only really necessary on the alpha architecture.

Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Acked-by: Greg Rose <gregory.v.rose@intel.com>
Tested-by: Sibai Li <sibai.li@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

2 files changed, 31 insertions, 23 deletions

diff --git a/drivers/net/ethernet/intel/igbvf/igbvf.h b/drivers/net/ethernet/intel/igbvf/igbvf.h
index fdca7b672776..a1463e3d14c0 100644
--- a/drivers/net/ethernet/intel/igbvf/igbvf.h
+++ b/drivers/net/ethernet/intel/igbvf/igbvf.h
@@ -127,8 +127,8 @@ struct igbvf_buffer {
 		/* Tx */
 		struct {
 			unsigned long time_stamp;
+			union e1000_adv_tx_desc *next_to_watch;
 			u16 length;
-			u16 next_to_watch;
 			u16 mapped_as_page;
 		};
 		/* Rx */
diff --git a/drivers/net/ethernet/intel/igbvf/netdev.c b/drivers/net/ethernet/intel/igbvf/netdev.c
index f53f7136e508..d60cd4393415 100644
--- a/drivers/net/ethernet/intel/igbvf/netdev.c
+++ b/drivers/net/ethernet/intel/igbvf/netdev.c
@@ -797,20 +797,31 @@ static bool igbvf_clean_tx_irq(struct igbvf_ring *tx_ring)
 	struct sk_buff *skb;
 	union e1000_adv_tx_desc *tx_desc, *eop_desc;
 	unsigned int total_bytes = 0, total_packets = 0;
-	unsigned int i, eop, count = 0;
+	unsigned int i, count = 0;
 	bool cleaned = false;
 
 	i = tx_ring->next_to_clean;
-	eop = tx_ring->buffer_info[i].next_to_watch;
-	eop_desc = IGBVF_TX_DESC_ADV(*tx_ring, eop);
+	buffer_info = &tx_ring->buffer_info[i];
+	eop_desc = buffer_info->next_to_watch;
+
+	do {
+		/* if next_to_watch is not set then there is no work pending */
+		if (!eop_desc)
+			break;
+
+		/* prevent any other reads prior to eop_desc */
+		read_barrier_depends();
+
+		/* if DD is not set pending work has not been completed */
+		if (!(eop_desc->wb.status & cpu_to_le32(E1000_TXD_STAT_DD)))
+			break;
+
+		/* clear next_to_watch to prevent false hangs */
+		buffer_info->next_to_watch = NULL;
 
-	while ((eop_desc->wb.status & cpu_to_le32(E1000_TXD_STAT_DD)) &&
-	       (count < tx_ring->count)) {
-		rmb();	/* read buffer_info after eop_desc status */
 		for (cleaned = false; !cleaned; count++) {
 			tx_desc = IGBVF_TX_DESC_ADV(*tx_ring, i);
-			buffer_info = &tx_ring->buffer_info[i];
-			cleaned = (i == eop);
+			cleaned = (tx_desc == eop_desc);
 			skb = buffer_info->skb;
 
 			if (skb) {
@@ -831,10 +842,12 @@ static bool igbvf_clean_tx_irq(struct igbvf_ring *tx_ring)
 			i++;
 			if (i == tx_ring->count)
 				i = 0;
+
+			buffer_info = &tx_ring->buffer_info[i];
 		}
-		eop = tx_ring->buffer_info[i].next_to_watch;
-		eop_desc = IGBVF_TX_DESC_ADV(*tx_ring, eop);
-	}
+
+		eop_desc = buffer_info->next_to_watch;
+	} while (count < tx_ring->count);
 
 	tx_ring->next_to_clean = i;
 
@@ -1961,7 +1974,6 @@ static int igbvf_tso(struct igbvf_adapter *adapter,
 	context_desc->seqnum_seed = 0;
 
 	buffer_info->time_stamp = jiffies;
-	buffer_info->next_to_watch = i;
 	buffer_info->dma = 0;
 	i++;
 	if (i == tx_ring->count)
@@ -2021,7 +2033,6 @@ static inline bool igbvf_tx_csum(struct igbvf_adapter *adapter,
 		context_desc->mss_l4len_idx = 0;
 
 		buffer_info->time_stamp = jiffies;
-		buffer_info->next_to_watch = i;
 		buffer_info->dma = 0;
 		i++;
 		if (i == tx_ring->count)
@@ -2061,8 +2072,7 @@ static int igbvf_maybe_stop_tx(struct net_device *netdev, int size)
 
 static inline int igbvf_tx_map_adv(struct igbvf_adapter *adapter,
                                    struct igbvf_ring *tx_ring,
-                                   struct sk_buff *skb,
-                                   unsigned int first)
+				   struct sk_buff *skb)
 {
 	struct igbvf_buffer *buffer_info;
 	struct pci_dev *pdev = adapter->pdev;
@@ -2077,7 +2087,6 @@ static inline int igbvf_tx_map_adv(struct igbvf_adapter *adapter,
 	buffer_info->length = len;
 	/* set time_stamp *before* dma to help avoid a possible race */
 	buffer_info->time_stamp = jiffies;
-	buffer_info->next_to_watch = i;
 	buffer_info->mapped_as_page = false;
 	buffer_info->dma = dma_map_single(&pdev->dev, skb->data, len,
 					  DMA_TO_DEVICE);
@@ -2100,7 +2109,6 @@ static inline int igbvf_tx_map_adv(struct igbvf_adapter *adapter,
 		BUG_ON(len >= IGBVF_MAX_DATA_PER_TXD);
 		buffer_info->length = len;
 		buffer_info->time_stamp = jiffies;
-		buffer_info->next_to_watch = i;
 		buffer_info->mapped_as_page = true;
 		buffer_info->dma = skb_frag_dma_map(&pdev->dev, frag, 0, len,
 						DMA_TO_DEVICE);
@@ -2109,7 +2117,6 @@ static inline int igbvf_tx_map_adv(struct igbvf_adapter *adapter,
 	}
 
 	tx_ring->buffer_info[i].skb = skb;
-	tx_ring->buffer_info[first].next_to_watch = i;
 
 	return ++count;
 
@@ -2120,7 +2127,6 @@ dma_error:
 	buffer_info->dma = 0;
 	buffer_info->time_stamp = 0;
 	buffer_info->length = 0;
-	buffer_info->next_to_watch = 0;
 	buffer_info->mapped_as_page = false;
 	if (count)
 		count--;
@@ -2139,7 +2145,8 @@ dma_error:
 
 static inline void igbvf_tx_queue_adv(struct igbvf_adapter *adapter,
                                       struct igbvf_ring *tx_ring,
-                                      int tx_flags, int count, u32 paylen,
+				      int tx_flags, int count,
+				      unsigned int first, u32 paylen,
                                       u8 hdr_len)
 {
 	union e1000_adv_tx_desc *tx_desc = NULL;
@@ -2189,6 +2196,7 @@ static inline void igbvf_tx_queue_adv(struct igbvf_adapter *adapter,
 	 * such as IA-64). */
 	wmb();
 
+	tx_ring->buffer_info[first].next_to_watch = tx_desc;
 	tx_ring->next_to_use = i;
 	writel(i, adapter->hw.hw_addr + tx_ring->tail);
 	/* we need this if more than one processor can write to our tail
@@ -2255,11 +2263,11 @@ static netdev_tx_t igbvf_xmit_frame_ring_adv(struct sk_buff *skb,
 	 * count reflects descriptors mapped, if 0 then mapping error
 	 * has occurred and we need to rewind the descriptor queue
 	 */
-	count = igbvf_tx_map_adv(adapter, tx_ring, skb, first);
+	count = igbvf_tx_map_adv(adapter, tx_ring, skb);
 
 	if (count) {
 		igbvf_tx_queue_adv(adapter, tx_ring, tx_flags, count,
-		                   skb->len, hdr_len);
+				   first, skb->len, hdr_len);
 		/* Make sure there is space in the ring for the next send. */
 		igbvf_maybe_stop_tx(netdev, MAX_SKB_FRAGS + 4);
 	} else {