This patchset introduces the new implementation of tx_burst
routine of mlx5 PMD. The existing implementation is based on the
several branches of tx_burst routines optimizied for most
common sets of Tx hardware offload, supported by Mellanox NICs.
It was not very easy to update, support and develop such kind
of code - multiple branches impose multiple points to process.
Also, many of frequently requested offload combinations are not
supported yet in the most efficient way. That leads to selecting
of not completely matching tx_burst routine and harms the performance.
The new Tx datapath:
- introduces the unified template for tx_burst routine to generate
on compile time the most efficient instances, tuned for specified
sets of hardware offloads
- the best matching instance is chosen in run-time at sending queue
configuration
- almost all possible sets of Tx are supported
- the best possible method to send each packet is chosen on runtime
- introduces new devargs for more accurate tuning of data inlining
Signed-off-by: Viacheslav Ovsiienko <viachesl...@mellanox.com>
Viacheslav Ovsiienko (7):
net/mlx5: remove Tx datapath implementation
net/mlx5: add Tx datapath related devargs
net/mlx5: update Tx datapath definitions
net/mlx5: add Tx datapath configuration and setup
net/mlx5: introduce Tx burst routine template
net/mlx5: implement Tx burst template
net/mlx5: add minimal required Tx data inline
doc/guides/nics/mlx5.rst | 168 +-
drivers/net/mlx5/mlx5.c | 155 +-
drivers/net/mlx5/mlx5.h | 13 +-
drivers/net/mlx5/mlx5_defs.h | 36 +-
drivers/net/mlx5/mlx5_devx_cmds.c | 100 +-
drivers/net/mlx5/mlx5_ethdev.c | 58 -
drivers/net/mlx5/mlx5_prm.h | 355 ++-
drivers/net/mlx5/mlx5_rxtx.c | 5260 ++++++++++++++++++++++-----------
drivers/net/mlx5/mlx5_rxtx.h | 333 +--
drivers/net/mlx5/mlx5_rxtx_vec.c | 175 --
drivers/net/mlx5/mlx5_rxtx_vec_neon.h | 289 --
drivers/net/mlx5/mlx5_rxtx_vec_sse.h | 284 --
drivers/net/mlx5/mlx5_txq.c | 242 +-
13 files changed, 4389 insertions(+), 3079 deletions(-)
--
1.8.3.1
