>
> On Wed, Dec 17, 2025 at 03:18:42PM +0000, Loftus, Ciara wrote:
> > > Subject: Re: [PATCH 1/1] net/i40e: unset vector flag when scalar path is
> > > chosen
> > >
> > > On Wed, Dec 17, 2025 at 02:54:36PM +0000, Ciara Loftus wrote:
> > > > The tx_vec_allowed flag should be set to false if a scalar Tx path is
> > > > chosen.
> > > >
> > > > Fixes: 1ff08bb7ad90 ("net/i40e: use common Tx path selection
> > > infrastructure")
> > > >
> > > > Signed-off-by: Ciara Loftus <[email protected]>
> > > > ---
> > > > drivers/net/intel/i40e/i40e_rxtx.c | 3 +++
> > > > 1 file changed, 3 insertions(+)
> > > >
> > > > diff --git a/drivers/net/intel/i40e/i40e_rxtx.c
> > > b/drivers/net/intel/i40e/i40e_rxtx.c
> > > > index 2db58c6b24..dd859bda4e 100644
> > > > --- a/drivers/net/intel/i40e/i40e_rxtx.c
> > > > +++ b/drivers/net/intel/i40e/i40e_rxtx.c
> > > > @@ -3631,6 +3631,9 @@ i40e_set_tx_function(struct rte_eth_dev
> *dev)
> > > > ad->tx_func_type == I40E_TX_ALTIVEC ||
> > > > ad->tx_func_type == I40E_TX_AVX2)
> > > > dev->recycle_tx_mbufs_reuse =
> > > i40e_recycle_tx_mbufs_reuse_vec;
> > > > +
> > > > + if (i40e_tx_path_infos[ad->tx_func_type].features.simd_width <
> > > RTE_VECT_SIMD_128)
> > > > + ad->tx_vec_allowed = false;
> > > > }
> > > >
> > > Under what circumstances would this be a problem, or under what
> > > circumstances would we have this situaion?
> >
> > The circumstances we would have this situation is when the driver
> > determines that tx vectorisation is allowed but we end up selecting
> > a scalar path. The result would be "vector_tx" being set for the txq
> > during tx_queue_start and later during tx_queue_stop, the wrong
> > path being taken in ci_txq_release_all_mbufs
> >
> Would this be better fixed by some refactoring to chose the release
> function based on ad->tx_func_type, rather than having a separate vector
> flag?
>
> If not, may I suggest that you remove the if condition, and just do:
> ad->tx_vec_allowed =
> (i40e_tx_path_infos[ad->tx_func_type].features.simd_width <
> RTE_VECT_SIMD_128);
The release function is in the common code so it must use the
txq->vector_tx flag.
I think we need to also keep the adapter->tx_vec_allowed flag as it's
used elsewhere in the driver beyond setting the appropriate value to
txq->vector_tx. eg. in the i40e_dev_tx_queue_setup_runtime function.
I think there might scope for some further refactoring here but it would
end up touching a number of places in the code. In the interest of time
and simplicity I propose implementing the fix you suggested above
for now and maybe revisiting further refactoring later.
