LWIP buffer alignment

[Tim Harvey]

On Tue, Sep 9, 2025 at 12:04 AM Jerome Forissier
<jerome.forissier at linaro.org> wrote:
>
>
>
> On 9/8/25 18:44, Tim Harvey wrote:
> > On Mon, Sep 8, 2025 at 7:21 AM Jerome Forissier
> > <jerome.forissier at linaro.org> wrote:
> >>
> >> Hi Tim,
> >>
> >> On 9/5/25 19:45, Tim Harvey wrote:
> >>> Hi Jerome,
> >>>
> >>> I'm working with a board that has an IMX ENETC MAC on it
> >>> drivers/net/fsl_enetc.c and when configured for LWIP it fails and
> >>> warns about RX buffers not being 64B aligned:
> >>> WARNING at drivers/net/fsl_enetc.c:776/enetc_setup_rx_bdr()!
> >>>
> >>> I tried changing MEM_ALIGNMENT to 64 in lib/lwip/u-boot/lwipopts.h but
> >>> that did not align the buffers to 64B. Where are the lwip buffers
> >>> created and do you have any suggestions for working around this?
> >>
> >> The LwIP RX buffers are created in net-lwip.c. See eth_init_rings().
> >> You might have PKTSIZE_ALIGN and/or PKTALIGN set wrongly.
> >>
> >
> > Hi Jerome,
> >
> > Thanks for pointing me to eth_init_rings(). This function uses
> > net_pkt_buf as the base for the net_rx_packets array which is a static
> > global and not aligned.
> >
> > Looks like we need:
> > diff --git a/net/lwip/net-lwip.c b/net/lwip/net-lwip.c
> > index 660ceb10cbe..aec05d55de6 100644
> > --- a/net/lwip/net-lwip.c
> > +++ b/net/lwip/net-lwip.c
> > @@ -30,7 +30,7 @@ void (*push_packet)(void *, int len) = 0;
> >  static int net_try_count;
> >  static int net_restarted;
> >  int net_restart_wrap;
> > -static uchar net_pkt_buf[(PKTBUFSRX) * PKTSIZE_ALIGN + PKTALIGN];
> > +static uchar net_pkt_buf[(PKTBUFSRX) * PKTSIZE_ALIGN + PKTALIGN]
> > __attribute__((aligned(PKTALIGN)));
> >  uchar *net_rx_packets[PKTBUFSRX];
> >  uchar *net_rx_packet;
> >  const u8 net_bcast_ethaddr[6] = { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff };
> >
> > If you agree I'll submit that. This does resolve the issue I see with
> > the enetc driver using LwIP.
>
> Please do. But then I wonder how come the legacy NET works OK? Or does it?
> net/net.c has the same initialization sequence.
>

Hi Jerome,

Sorry for the late reply here. I can't explain how legacy NET aligns
Ok - just lucky I guess as net_pkt_buf does end up with the right
alignment there but definitely does not with lwip.

Best Regards,

Tim