Re: [dpdk-dev] [PATCH] net/mlx5: fix transmit descriptor with VLAN insertions

[Matan Azrad <matan@mellanox.com>]

Hi

From: Slava Ovsiienko <viacheslavo@mellanox.com>
> Hi, Matan
> 
> Please, see below.
> 
> > -----Original Message-----
> > From: Matan Azrad <matan@mellanox.com>
> > Sent: Tuesday, September 3, 2019 16:05
> > To: Slava Ovsiienko <viacheslavo@mellanox.com>; dev@dpdk.org
> > Cc: stable@dpdk.org
> > Subject: RE: [PATCH] net/mlx5: fix transmit descriptor with VLAN
> > insertions
> >
> > Hi Slava
> >
> > From: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
> > > If VLAN tag insertion transmit offload is engaged
> > > (DEV_TX_OFFLOAD_VLAN_INSERT in tx queue configuration is set) the
> > > transmit descriptor may be built with wrong format, due to packet
> > > length is not adjusted. Also, the ring buffer wrapup is not handled
> correctly.
> > >
> > > Fixes: 18a1c20044c0 ("net/mlx5: implement Tx burst template")
> > > Cc: stable@dpdk.org
> > >
> > > Signed-off-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
> > > ---
> > >  drivers/net/mlx5/mlx5_rxtx.c | 5 +++--
> > >  1 file changed, 3 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/drivers/net/mlx5/mlx5_rxtx.c
> > > b/drivers/net/mlx5/mlx5_rxtx.c index 8ec90c3..f540977 100644
> > > --- a/drivers/net/mlx5/mlx5_rxtx.c
> > > +++ b/drivers/net/mlx5/mlx5_rxtx.c
> > > @@ -2861,13 +2861,14 @@ enum mlx5_txcmp_code {
> > >  	memcpy(pdst, buf, MLX5_DSEG_MIN_INLINE_SIZE);
> > >  	buf += MLX5_DSEG_MIN_INLINE_SIZE;
> > >  	pdst += MLX5_DSEG_MIN_INLINE_SIZE;
> > > +	len -= MLX5_DSEG_MIN_INLINE_SIZE;
> > >  	/* Insert VLAN ethertype + VLAN tag. Pointer is aligned. */
> > >  	assert(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE));
> > > +	if (unlikely(pdst >= (uint8_t *)txq->wqes_end))
> >
> > Why unlikely?
> > Wraparound is expected even in good cases, no?
> 
> It is good case but supposed to be rare. Wraparound happens once on entire
> ring buffer fill.

1/ring size - event - it is part of the good flow.
I think it is better to let the branch prediction to do the work in this case and in all other places when good flow may happen for both true or false.


> At this codepoint it ever has the less chance to happen - many of
> wraparounds happen on data copying (the code below in this function).
> 
> With best regards,
> Slava
> 
> 
> >
> >
> >
> > > +		pdst = (uint8_t *)txq->wqes;
> > >  	*(uint32_t *)pdst = rte_cpu_to_be_32((RTE_ETHER_TYPE_VLAN <<
> > > 16) |
> > >  					      loc->mbuf->vlan_tci);
> > >  	pdst += sizeof(struct rte_vlan_hdr);
> > > -	if (unlikely(pdst >= (uint8_t *)txq->wqes_end))
> > > -		pdst = (uint8_t *)txq->wqes;
> > >  	/*
> > >  	 * The WQEBB space availability is checked by caller.
> > >  	 * Here we should be aware of WQE ring buffer wraparound only.
> > > --
> > > 1.8.3.1