[dpdk-dev] [PATCH v2 4/7] net/mlx4: merge Tx path functions

[Ophir Munk]

Hi,

On Tuesday, October 24, 2017 4:52 PM, Nélio Laranjeiro wrote:
> 
> On Mon, Oct 23, 2017 at 02:21:57PM +0000, Ophir Munk wrote:
> > From: Matan Azrad <matan at mellanox.com>
> >
> > Merge tx_burst and mlx4_post_send functions to prevent double asking
> > about WQ remain space.
> >
> > This should improve performance.
> >
> > Signed-off-by: Matan Azrad <matan at mellanox.com>
> > ---
> >  drivers/net/mlx4/mlx4_rxtx.c | 353
> > +++++++++++++++++++++----------------------
> >  1 file changed, 170 insertions(+), 183 deletions(-)
> 
> What are the real expectation you have on the remaining patches of the
> series?
> 
> According to the comment of this commit log "This should improve
> performance" there are too many barriers at each packet/segment level to
> improve something.
> 
> The point is, mlx4_burst_tx() should write all the WQE without any barrier as
> it is processing a burst of packets (whereas Verbs functions which may only
> process a single packet).  

> The lonely barrier which should be present is the
> one to ensure that all the host memory is flushed before triggering the Tx
> doorbell.
> 

There is a known ConnectX-3 HW limitation: the first 4 bytes of every TXWBB (64 bytes chunks) should be 
written in a reversed order (from last TXWBB to first TXWBB). 
The last 60 bytes of any TXWBB can be written in any order (before writing the first 4 bytes).
Is your last statement (using lonely barrier) is in accordance with this limitation? Please explain.

> There is also too many cases handled which are useless in bursts situation,
> this function needs to be re-written to its minimal use case i.e.  processing a
> valid burst of packets/segments and triggering at the end of the burst the Tx
> doorbell.
> 
> Regards,
> 
> --
> Nélio Laranjeiro
> 6WIND