net/mlx5: fix vectorized Rx burst termination
authorAlexander Kozyrev <akozyrev@mellanox.com>
Tue, 2 Jun 2020 03:50:41 +0000 (03:50 +0000)
committerFerruh Yigit <ferruh.yigit@intel.com>
Wed, 3 Jun 2020 15:20:32 +0000 (17:20 +0200)
commitc9cc554ba423641d3515671269c5648dab3bb9ad
tree9a05bb10a0fe349005ddf359226a94538b6aef02
parenta1da6f624c9b42035c231a8f44269f9530f55c86
net/mlx5: fix vectorized Rx burst termination

Maximum burst size of Vectorized Rx burst routine is set to
MLX5_VPMD_RX_MAX_BURST(64). This limits the performance of any
application that would like to gather more than 64 packets from
the single Rx burst for batch processing (i.e. VPP).

The situation gets worse with a mix of zipped and unzipped CQEs.
They are processed separately and the Rx burst function returns
small number of packets every call.

Repeat the cycle of gathering packets from the vectorized Rx routine
until a requested number of packets are collected or there are no
more CQEs left to process.

Fixes: 6cb559d67b83 ("net/mlx5: add vectorized Rx/Tx burst for x86")
Cc: stable@dpdk.org
Signed-off-by: Alexander Kozyrev <akozyrev@mellanox.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
Acked-by: Matan Azrad <matan@mellanox.com>
drivers/net/mlx5/mlx5_rxtx_vec.c
drivers/net/mlx5/mlx5_rxtx_vec_altivec.h
drivers/net/mlx5/mlx5_rxtx_vec_neon.h
drivers/net/mlx5/mlx5_rxtx_vec_sse.h