In order to support the 9K MTU the MPRQ feature should be updated to allow a packet to take more than one stride (single linear buffer). Receiving a packet into multiple adjacent strides should be implemented. The reason preventing the packet to be received into multiple strides is that the data buffer must be preceded with some HEAD_ROOM space. In the current implementation the HEAD_ROOM space is borrowed by the PMD from the tail of the preceding stride. If packet takes multiple strides the tail of stride may be overwritten with a packet data and the memory can't be borrowed to provide the HEAD_ROOM space for the next packet. Special care is needed to prevent the HEAD_ROOM corruption as such: - copy a whole packet into a separate memory buffer if scatter is off - copy an overlapping data only and craft a multi-segment mbuf otherwise After multi-stride support for packets receiving is in place it is possible to reduce the stride size for more efficient memory utilization. Introduce the mprq_log_stride_size device parameter to configure a stride size for MPRQ. Default stride size is set to 2048 bytes.
Signed-off-by: Alexander Kozyrev <akozy...@mellanox.com> --- v1: https://patchwork.dpdk.org/cover/67558/ v2: https://patchwork.dpdk.org/cover/67670/ merge documentation and implementation in one commit v3: rollback to the simple burst Rx in case the packet size is too big to fit into the stride and the mprq_log_stride_size is not configured Alexander Kozyrev (3): net/mlx5: add a devarg to specify MPRQ stride size net/mlx5: enable MPRQ multi-stride operations net/mlx5: add multi-segment packets in MPRQ mode doc/guides/nics/mlx5.rst | 17 ++++- doc/guides/rel_notes/release_20_05.rst | 1 + drivers/net/mlx5/mlx5.c | 34 ++++++++-- drivers/net/mlx5/mlx5.h | 1 + drivers/net/mlx5/mlx5_defs.h | 3 + drivers/net/mlx5/mlx5_rxq.c | 70 +++++++++++--------- drivers/net/mlx5/mlx5_rxtx.c | 113 +++++++++++++++++++-------------- drivers/net/mlx5/mlx5_rxtx.h | 2 +- 8 files changed, 154 insertions(+), 87 deletions(-) -- 1.8.3.1
