mlx4: avoid unnecessary dirtying of critical fields
authorEric Dumazet <edumazet@google.com>
Sun, 20 Nov 2016 17:24:36 +0000 (09:24 -0800)
committerDavid S. Miller <davem@davemloft.net>
Mon, 21 Nov 2016 16:33:31 +0000 (11:33 -0500)
While stressing a 40Gbit mlx4 NIC with busy polling, I found false
sharing in mlx4 driver that can be easily avoided.

This patch brings an additional 7 % performance improvement in UDP_RR
workload.

1) If we received no frame during one mlx4_en_process_rx_cq()
   invocation, no need to call mlx4_cq_set_ci() and/or dirty ring->cons

2) Do not refill rx buffers if we have plenty of them.
   This avoids false sharing and allows some bulk/batch optimizations.
   Page allocator and its locks will thank us.

Finally, mlx4_en_poll_rx_cq() should not return 0 if it determined
cpu handling NIC IRQ should be changed. We should return budget-1
instead, to not fool net_rx_action() and its netdev_budget.

v2: keep AVG_PERF_COUNTER(... polled) even if polled is 0

Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/mellanox/mlx4/en_rx.c

index 22f08f9ef4645869359783823127c0432fc7a591..6562f78b07f4370b5c1ea2c5e3a4221d7ebaeba8 100644 (file)
@@ -688,18 +688,23 @@ out_loopback:
        dev_kfree_skb_any(skb);
 }
 
-static void mlx4_en_refill_rx_buffers(struct mlx4_en_priv *priv,
-                                    struct mlx4_en_rx_ring *ring)
+static bool mlx4_en_refill_rx_buffers(struct mlx4_en_priv *priv,
+                                     struct mlx4_en_rx_ring *ring)
 {
-       int index = ring->prod & ring->size_mask;
+       u32 missing = ring->actual_size - (ring->prod - ring->cons);
 
-       while ((u32) (ring->prod - ring->cons) < ring->actual_size) {
-               if (mlx4_en_prepare_rx_desc(priv, ring, index,
+       /* Try to batch allocations, but not too much. */
+       if (missing < 8)
+               return false;
+       do {
+               if (mlx4_en_prepare_rx_desc(priv, ring,
+                                           ring->prod & ring->size_mask,
                                            GFP_ATOMIC | __GFP_COLD))
                        break;
                ring->prod++;
-               index = ring->prod & ring->size_mask;
-       }
+       } while (--missing);
+
+       return true;
 }
 
 /* When hardware doesn't strip the vlan, we need to calculate the checksum
@@ -1081,15 +1086,20 @@ consumed:
 
 out:
        rcu_read_unlock();
-       if (doorbell_pending)
-               mlx4_en_xmit_doorbell(priv->tx_ring[TX_XDP][cq->ring]);
 
+       if (polled) {
+               if (doorbell_pending)
+                       mlx4_en_xmit_doorbell(priv->tx_ring[TX_XDP][cq->ring]);
+
+               mlx4_cq_set_ci(&cq->mcq);
+               wmb(); /* ensure HW sees CQ consumer before we post new buffers */
+               ring->cons = cq->mcq.cons_index;
+       }
        AVG_PERF_COUNTER(priv->pstats.rx_coal_avg, polled);
-       mlx4_cq_set_ci(&cq->mcq);
-       wmb(); /* ensure HW sees CQ consumer before we post new buffers */
-       ring->cons = cq->mcq.cons_index;
-       mlx4_en_refill_rx_buffers(priv, ring);
-       mlx4_en_update_rx_prod_db(ring);
+
+       if (mlx4_en_refill_rx_buffers(priv, ring))
+               mlx4_en_update_rx_prod_db(ring);
+
        return polled;
 }
 
@@ -1131,10 +1141,13 @@ int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
                        return budget;
 
                /* Current cpu is not according to smp_irq_affinity -
-                * probably affinity changed. need to stop this NAPI
-                * poll, and restart it on the right CPU
+                * probably affinity changed. Need to stop this NAPI
+                * poll, and restart it on the right CPU.
+                * Try to avoid returning a too small value (like 0),
+                * to not fool net_rx_action() and its netdev_budget
                 */
-               done = 0;
+               if (done)
+                       done--;
        }
        /* Done for now */
        if (napi_complete_done(napi, done))