On Wed, Oct 22, 2025 at 07:17:20PM +0200, Maciej Fijalkowski wrote:
> On Wed, Oct 22, 2025 at 12:32:00AM +0700, Alessandro Decina wrote:
> 
> Hi Alessandro,

Hey,

Thanks for the review!


> 
> > Whenever a status descriptor is received, i40e processes and skips over
> > it, correctly updating next_to_process but forgetting to update
> > next_to_clean. In the next iteration this accidentally causes the
> > creation of an invalid multi-buffer xdp_buff where the first fragment
> > is the status descriptor.
> > 
> > If then a skb is constructed from such an invalid buffer - because the
> > eBPF program returns XDP_PASS - a panic occurs:
> 
> can you elaborate on the test case that would reproduce this? I suppose
> AF_XDP ZC with jumbo frames, doing XDP_PASS, but what was FDIR setup that
> caused status descriptors?

Doesn't have to be jumbo or multi-frag, anything that does XDP_PASS
reproduces, as long as status descriptors are posted. 

See the scenarios here 
https://lore.kernel.org/netdev/aPkDtuVgbS4J-Og_@lima-default/

As for what's causing the status descriptors, I haven't been able to
figure that out. I just know that I periodically get
I40E_RX_PROG_STATUS_DESC_FD_FILTER_STATUS. Happy to dig deeper if you
have any ideas!

> > diff --git a/drivers/net/ethernet/intel/i40e/i40e_xsk.c 
> > b/drivers/net/ethernet/intel/i40e/i40e_xsk.c
> > index 9f47388eaba5..dbc19083bbb7 100644
> > --- a/drivers/net/ethernet/intel/i40e/i40e_xsk.c
> > +++ b/drivers/net/ethernet/intel/i40e/i40e_xsk.c
> > @@ -441,13 +441,18 @@ int i40e_clean_rx_irq_zc(struct i40e_ring *rx_ring, 
> > int budget)
> >             dma_rmb();
> >  
> >             if (i40e_rx_is_programming_status(qword)) {
> > +                   u16 ntp;
> > +
> >                     i40e_clean_programming_status(rx_ring,
> >                                                   rx_desc->raw.qword[0],
> >                                                   qword);
> >                     bi = *i40e_rx_bi(rx_ring, next_to_process);
> >                     xsk_buff_free(bi);
> > -                   if (++next_to_process == count)
> > +                   ntp = next_to_process++;
> > +                   if (next_to_process == count)
> >                             next_to_process = 0;
> > +                   if (next_to_clean == ntp)
> > +                           next_to_clean = next_to_process;
> 
> I wonder if this is more readable?
> 
> diff --git a/drivers/net/ethernet/intel/i40e/i40e_xsk.c 
> b/drivers/net/ethernet/intel/i40e/i40e_xsk.c
> index 9f47388eaba5..36f412a2d836 100644
> --- a/drivers/net/ethernet/intel/i40e/i40e_xsk.c
> +++ b/drivers/net/ethernet/intel/i40e/i40e_xsk.c
> @@ -446,6 +446,10 @@ int i40e_clean_rx_irq_zc(struct i40e_ring *rx_ring, int 
> budget)
>                                                     qword);
>                       bi = *i40e_rx_bi(rx_ring, next_to_process);
>                       xsk_buff_free(bi);
> +                     if (next_to_clean == next_to_process) {
> +                             if (++next_to_clean == count)
> +                                     next_to_clean = 0;
> +                     }
>                       if (++next_to_process == count)
>                               next_to_process = 0;
>                       continue;
> 
> >                     continue;
> >             }

Probably because I've looked at it for longer, I find my version clearer
(I think I copied it from another driver actually). But I don't really
mind, happy to switch to yours if you prefer!

Ciao
Alessandro

Reply via email to