[net-next,04/11] ixgbe: delay tail write to every 'n' packets

Submitted by Jeff Kirsher on April 21, 2017, 1:50 a.m.

Details

Message ID 20170421015029.18994-5-jeffrey.t.kirsher@intel.com
State Changes Requested
Delegated to: David Miller
Headers show

Commit Message

Jeff Kirsher April 21, 2017, 1:50 a.m.
From: John Fastabend <john.r.fastabend@intel.com>

Current XDP implementation hits the tail on every XDP_TX return
code. This patch changes driver behavior to only hit the tail after
packet processing is complete.

With this patch I can run XDP drop programs @ 14+Mpps and XDP_TX
programs are at ~13.5Mpps.

Signed-off-by: John Fastabend <john.r.fastabend@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
---
 drivers/net/ethernet/intel/ixgbe/ixgbe_main.c | 28 ++++++++++++++++-----------
 1 file changed, 17 insertions(+), 11 deletions(-)

Patch hide | download patch | download mbox

diff --git a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
index 92375eb4622c..8a2b8e871da7 100644
--- a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
+++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
@@ -2281,6 +2281,7 @@  static int ixgbe_clean_rx_irq(struct ixgbe_q_vector *q_vector,
 	unsigned int mss = 0;
 #endif /* IXGBE_FCOE */
 	u16 cleaned_count = ixgbe_desc_unused(rx_ring);
+	bool xdp_xmit = false;
 
 	while (likely(total_rx_packets < budget)) {
 		union ixgbe_adv_rx_desc *rx_desc;
@@ -2320,10 +2321,12 @@  static int ixgbe_clean_rx_irq(struct ixgbe_q_vector *q_vector,
 		}
 
 		if (IS_ERR(skb)) {
-			if (PTR_ERR(skb) == -IXGBE_XDP_TX)
+			if (PTR_ERR(skb) == -IXGBE_XDP_TX) {
+				xdp_xmit = true;
 				ixgbe_rx_buffer_flip(rx_ring, rx_buffer, size);
-			else
+			} else {
 				rx_buffer->pagecnt_bias++;
+			}
 			total_rx_packets++;
 			total_rx_bytes += size;
 		} else if (skb) {
@@ -2391,6 +2394,16 @@  static int ixgbe_clean_rx_irq(struct ixgbe_q_vector *q_vector,
 		total_rx_packets++;
 	}
 
+	if (xdp_xmit) {
+		struct ixgbe_ring *ring = adapter->xdp_ring[smp_processor_id()];
+
+		/* Force memory writes to complete before letting h/w
+		 * know there are new descriptors to fetch.
+		 */
+		wmb();
+		writel(ring->next_to_use, ring->tail);
+	}
+
 	u64_stats_update_begin(&rx_ring->syncp);
 	rx_ring->stats.packets += total_rx_packets;
 	rx_ring->stats.bytes += total_rx_bytes;
@@ -8237,14 +8250,8 @@  static int ixgbe_xmit_xdp_ring(struct ixgbe_adapter *adapter,
 	tx_desc->read.olinfo_status =
 		cpu_to_le32(len << IXGBE_ADVTXD_PAYLEN_SHIFT);
 
-	/* Force memory writes to complete before letting h/w know there
-	 * are new descriptors to fetch.  (Only applicable for weak-ordered
-	 * memory model archs, such as IA-64).
-	 *
-	 * We also need this memory barrier to make certain all of the
-	 * status bits have been updated before next_to_watch is written.
-	 */
-	wmb();
+	/* Avoid any potential race with xdp_xmit and cleanup */
+	smp_wmb();
 
 	/* set next_to_watch value indicating a packet is present */
 	i++;
@@ -8254,7 +8261,6 @@  static int ixgbe_xmit_xdp_ring(struct ixgbe_adapter *adapter,
 	tx_buffer->next_to_watch = tx_desc;
 	ring->next_to_use = i;
 
-	writel(i, ring->tail);
 	return IXGBE_XDP_TX;
 }