]> git.karo-electronics.de Git - linux-beck.git/commit
net/mlx5e: XDP TX xmit more
authorSaeed Mahameed <saeedm@mellanox.com>
Wed, 21 Sep 2016 09:19:49 +0000 (12:19 +0300)
committerDavid S. Miller <davem@davemloft.net>
Thu, 22 Sep 2016 06:51:41 +0000 (02:51 -0400)
commit35b510e257f7516546a0a3f725f71dfbccc3f733
treeeae31e97372200fe825241bba2481174730b3764
parentb5503b994ed5ed8dbfe821317e7b5b38acb065c5
net/mlx5e: XDP TX xmit more

Previously we rang XDP SQ doorbell on every forwarded XDP packet.

Here we introduce a xmit more like mechanism that will queue up more
than one packet into SQ (up to RX napi budget) w/o notifying the hardware.

Once RX napi budget is consumed and we exit napi RX loop, we will
flush (doorbell) all XDP looped packets in case there are such.

XDP forward packet rate:

Comparing XDP with and w/o xmit more (bulk transmit):

RX Cores    XDP TX       XDP TX (xmit more)
---------------------------------------------------
1           6.5Mpps      12.4Mpps
2          13.2Mpps      24.2Mpps
4          25.2Mpps      36.3Mpps*
8          36.3Mpps*     36.3Mpps*

*My xmitter was limited to 36.3Mpps, so it is the bottleneck.
It seems that receive side can handle more.

Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/mellanox/mlx5/core/en.h
drivers/net/ethernet/mellanox/mlx5/core/en_rx.c