diff mbox

[net-next] net/mlx4_en: fix ethtool -x

Message ID 1465359858.24873.35.camel@edumazet-glaptop3.roam.corp.google.com
State Accepted, archived
Delegated to: David Miller
Headers show

Commit Message

Eric Dumazet June 8, 2016, 4:24 a.m. UTC 
From: Eric Dumazet <edumazet@google.com>

mlx4 RSS is limited to spread incoming packets to a power of two number
of queues.

An uniformly distibuted traffic would be split on queues 0 to N-1, N
being a power of two, each queue having a 1/N weight.

If number of RX queues is not a power of two, upper RX queues do not
receive traffic.

ethtool -x is lying, because it pretends some queues have higher weight.

Before patch:

lpaa24:~# ethtool -L eth1 rx 24
lpaa24:~# ethtool -x eth1
RX flow hash indirection table for eth1 with 24 RX ring(s):
    0:      0     1     2     3     4     5     6     7
    8:      8     9    10    11    12    13    14    15
   16:      0     1     2     3     4     5     6     7
RSS hash key:
e0:7c:3a:89:07:55:b6:58:69:cc:f4:e5:24:62:e3:25:88:6c:42:5b:d2:cb:9a:d2:e0:06:e1:dc:f9:09:a1:89:0f:a0:30:43:73:6f:0c:b6

If this information was correct, user space tools could expect queues 0
to 7 to receive twice more traffic than queues 8 to 15

After patch :

lpaa24:~# ethtool -L eth1 rx 24
lpaa24:~# ethtool -x eth1      
RX flow hash indirection table for eth1 with 24 RX ring(s):
    0:      0     1     2     3     4     5     6     7
    8:      8     9    10    11    12    13    14    15
RSS hash key:
da:7b:09:60:f1:ac:67:b4:d0:72:d4:ec:a2:e5:80:0a:ad:50:22:1a:f8:f9:66:54:5f:22:45:c3:88:f4:57:82:c1:c1:90:ed:70:cb:40:ce
lpaa24:~# ethtool -X eth1 equal 8
lpaa24:~# ethtool -x eth1 
RX flow hash indirection table for eth1 with 24 RX ring(s):
    0:      0     1     2     3     4     5     6     7
    8:      0     1     2     3     4     5     6     7
RSS hash key:
da:7b:09:60:f1:ac:67:b4:d0:72:d4:ec:a2:e5:80:0a:ad:50:22:1a:f8:f9:66:54:5f:22:45:c3:88:f4:57:82:c1:c1:90:ed:70:cb:40:ce

Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Maciej Żenczykowski <maze@google.com>
Cc: Eugenia Emantayev <eugenia@mellanox.com>
Cc: Wei Wang <weiwan@google.com>
Cc: Willem de Bruijn <willemb@google.com>
---
 drivers/net/ethernet/mellanox/mlx4/en_ethtool.c |   25 ++++++--------
 1 file changed, 12 insertions(+), 13 deletions(-)

Comments

Tariq Toukan June 9, 2016, 8 a.m. UTC | #1 
Thanks, looks good.

Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
David Miller June 10, 2016, 6:40 a.m. UTC | #2 
From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Tue, 07 Jun 2016 21:24:18 -0700

> From: Eric Dumazet <edumazet@google.com>
> 
> mlx4 RSS is limited to spread incoming packets to a power of two number
> of queues.
> 
> An uniformly distibuted traffic would be split on queues 0 to N-1, N
> being a power of two, each queue having a 1/N weight.
> 
> If number of RX queues is not a power of two, upper RX queues do not
> receive traffic.
> 
> ethtool -x is lying, because it pretends some queues have higher weight.
 ...
> Signed-off-by: Eric Dumazet <edumazet@google.com>
> Reported-by: Maciej Żenczykowski <maze@google.com>

Applied.
diff mbox

Patch

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_ethtool.c b/drivers/net/ethernet/mellanox/mlx4/en_ethtool.c
index fc95affaf76b..51a2e8252b82 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_ethtool.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_ethtool.c
@@ -1107,7 +1107,7 @@  static u32 mlx4_en_get_rxfh_indir_size(struct net_device *dev)
 {
 	struct mlx4_en_priv *priv = netdev_priv(dev);
 
-	return priv->rx_ring_num;
+	return rounddown_pow_of_two(priv->rx_ring_num);
 }
 
 static u32 mlx4_en_get_rxfh_key_size(struct net_device *netdev)
@@ -1141,19 +1141,17 @@  static int mlx4_en_get_rxfh(struct net_device *dev, u32 *ring_index, u8 *key,
 			    u8 *hfunc)
 {
 	struct mlx4_en_priv *priv = netdev_priv(dev);
-	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
-	int rss_rings;
-	size_t n = priv->rx_ring_num;
+	u32 n = mlx4_en_get_rxfh_indir_size(dev);
+	u32 i, rss_rings;
 	int err = 0;
 
-	rss_rings = priv->prof->rss_rings ?: priv->rx_ring_num;
-	rss_rings = 1 << ilog2(rss_rings);
+	rss_rings = priv->prof->rss_rings ?: n;
+	rss_rings = rounddown_pow_of_two(rss_rings);
 
-	while (n--) {
+	for (i = 0; i < n; i++) {
 		if (!ring_index)
 			break;
-		ring_index[n] = rss_map->qps[n % rss_rings].qpn -
-			rss_map->base_qpn;
+		ring_index[i] = i % rss_rings;
 	}
 	if (key)
 		memcpy(key, priv->rss_key, MLX4_EN_RSS_KEY_SIZE);
@@ -1166,6 +1164,7 @@  static int mlx4_en_set_rxfh(struct net_device *dev, const u32 *ring_index,
 			    const u8 *key, const u8 hfunc)
 {
 	struct mlx4_en_priv *priv = netdev_priv(dev);
+	u32 n = mlx4_en_get_rxfh_indir_size(dev);
 	struct mlx4_en_dev *mdev = priv->mdev;
 	int port_up = 0;
 	int err = 0;
@@ -1175,18 +1174,18 @@  static int mlx4_en_set_rxfh(struct net_device *dev, const u32 *ring_index,
 	/* Calculate RSS table size and make sure flows are spread evenly
 	 * between rings
 	 */
-	for (i = 0; i < priv->rx_ring_num; i++) {
+	for (i = 0; i < n; i++) {
 		if (!ring_index)
-			continue;
+			break;
 		if (i > 0 && !ring_index[i] && !rss_rings)
 			rss_rings = i;
 
-		if (ring_index[i] != (i % (rss_rings ?: priv->rx_ring_num)))
+		if (ring_index[i] != (i % (rss_rings ?: n)))
 			return -EINVAL;
 	}
 
 	if (!rss_rings)
-		rss_rings = priv->rx_ring_num;
+		rss_rings = n;
 
 	/* RSS table size must be an order of 2 */
 	if (!is_power_of_2(rss_rings))