Re: [PATCH net-next] net: mvneta: try to use in-irq pp cache in mvneta_txq_bufs_free

[Jesper Dangaard Brouer <brouer@redhat.com>]

On Fri, 25 Sep 2020 13:29:00 +0200
Lorenzo Bianconi <lorenzo.bianconi@redhat.com> wrote:

> >
> > On Fri, 25 Sep 2020 12:01:32 +0200
> > Lorenzo Bianconi <lorenzo@kernel.org> wrote:
> >  
> > > Try to recycle the xdp tx buffer into the in-irq page_pool cache if
> > > mvneta_txq_bufs_free is executed in the NAPI context.  
> >
> > NACK - I don't think this is safe.  That is also why I named the
> > function postfix rx_napi.  The page pool->alloc.cache is associated
> > with the drivers RX-queue.  The xdp_frame's that gets freed could be
> > coming from a remote driver that use page_pool. This remote drivers
> > RX-queue processing can run concurrently on a different CPU, than this
> > drivers TXQ-cleanup.  
> 
> ack, right. What about if we do it just XDP_TX use case? Like:
> 
> if (napi && buf->type == MVNETA_TYPE_XDP_TX)
>    xdp_return_frame_rx_napi(buf->xdpf);
> else
>    xdp_return_frame(buf->xdpf);
> 
> In this way we are sure the packet is coming from local page_pool.

Yes, that case XDP_TX should be safe.

> >
> > If you want to speedup this, I instead suggest that you add a
> > xdp_return_frame_bulk API.  
> 
> I will look at it

Great!

Notice that bulk return should be easy/obvious in most drivers, as they
(like mvneta in mvneta_txq_bufs_free()) have a loop that process
several TXQ completions.

I did a quick tests on mlx5 with xdp_redirect_map and perf report shows
__xdp_return calls at the top#1 overhead.

# Overhead  CPU  Symbol                              
# ........  ...  ....................................
#
     8.46%  003  [k] __xdp_return                    
     6.41%  003  [k] dma_direct_map_page             
     4.65%  003  [k] bpf_xdp_redirect_map            
     4.58%  003  [k] dma_direct_unmap_page           
     4.04%  003  [k] xdp_do_redirect                 
     3.53%  003  [k] __page_pool_put_page            
     3.27%  003  [k] dma_direct_sync_single_for_cpu  
     2.63%  003  [k] dev_map_enqueue                 
     2.28%  003  [k] page_pool_refill_alloc_cache    
     1.69%  003  [k] bq_enqueue.isra.0               
     1.15%  003  [k] _raw_spin_lock                  
     0.92%  003  [k] xdp_return_frame                

Thus, there will be a benefit from implementing a bulk return.  Also
for your XDP_TX case, as the overhead in __xdp_return also exist for
rx_napi variant.


> > > Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
> > > ---
> > >  drivers/net/ethernet/marvell/mvneta.c | 11 +++++++----
> > >  1 file changed, 7 insertions(+), 4 deletions(-)
> > >
> > > diff --git a/drivers/net/ethernet/marvell/mvneta.c b/drivers/net/ethernet/marvell/mvneta.c
> > > index 14df3aec285d..646fbf4ed638 100644
> > > --- a/drivers/net/ethernet/marvell/mvneta.c
> > > +++ b/drivers/net/ethernet/marvell/mvneta.c
> > > @@ -1831,7 +1831,7 @@ static struct mvneta_tx_queue *mvneta_tx_done_policy(struct mvneta_port *pp,
> > >  /* Free tx queue skbuffs */
> > >  static void mvneta_txq_bufs_free(struct mvneta_port *pp,
> > >                                struct mvneta_tx_queue *txq, int num,
> > > -                              struct netdev_queue *nq)
> > > +                              struct netdev_queue *nq, bool napi)
> > >  {
> > >       unsigned int bytes_compl = 0, pkts_compl = 0;
> > >       int i;
> > > @@ -1854,7 +1854,10 @@ static void mvneta_txq_bufs_free(struct mvneta_port *pp,
> > >                       dev_kfree_skb_any(buf->skb);
> > >               } else if (buf->type == MVNETA_TYPE_XDP_TX ||
> > >                          buf->type == MVNETA_TYPE_XDP_NDO) {
> > > -                     xdp_return_frame(buf->xdpf);
> > > +                     if (napi)
> > > +                             xdp_return_frame_rx_napi(buf->xdpf);
> > > +                     else
> > > +                             xdp_return_frame(buf->xdpf);
> > >               }
> > >       }
> > >
> > > @@ -1872,7 +1875,7 @@ static void mvneta_txq_done(struct mvneta_port *pp,
> > >       if (!tx_done)
> > >               return;
> > >
> > > -     mvneta_txq_bufs_free(pp, txq, tx_done, nq);
> > > +     mvneta_txq_bufs_free(pp, txq, tx_done, nq, true);
> > >
> > >       txq->count -= tx_done;
> > >
> > > @@ -2859,7 +2862,7 @@ static void mvneta_txq_done_force(struct mvneta_port *pp,
> > >       struct netdev_queue *nq = netdev_get_tx_queue(pp->dev, txq->id);
> > >       int tx_done = txq->count;
> > >
> > > -     mvneta_txq_bufs_free(pp, txq, tx_done, nq);
> > > +     mvneta_txq_bufs_free(pp, txq, tx_done, nq, false);
> > >
> > >       /* reset txq */
> > >       txq->count = 0;  

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer