Re: [PATCH net-next] stmmac: align RX buffers

[Matteo Croce]

On Thu, Aug 12, 2021 at 1:05 PM Marc Zyngier <maz@xxxxxxxxxx> wrote:
>
> On Thu, 12 Aug 2021 11:18:35 +0100,
> Matteo Croce <mcroce@xxxxxxxxxxxxxxxxxxx> wrote:
> >
> > On Thu, 12 Aug 2021 10:48:03 +0200
> > Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote:
> >
> > >
> > >
> > > On 8/11/21 4:16 PM, Marc Zyngier wrote:
> > > > On Wed, 11 Aug 2021 13:53:59 +0100,
> > > > Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote:
> > > >
> > > >> Are you sure you do not need to adjust stmmac_set_bfsize(),
> > > >> stmmac_rx_buf1_len() and stmmac_rx_buf2_len() ?
> > > >>
> > > >> Presumably DEFAULT_BUFSIZE also want to be increased by NET_SKB_PAD
> > > >>
> > > >> Patch for stmmac_rx_buf1_len() :
> > > >>
> > > >> diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> > > >> b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c index
> > > >> 7b8404a21544cf29668e8a14240c3971e6bce0c3..041a74e7efca3436bfe3e17f972dd156173957a9
> > > >> 100644 --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +++
> > > >> b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c @@ -4508,12
> > > >> +4508,12 @@ static unsigned int stmmac_rx_buf1_len(struct
> > > >> stmmac_priv *priv, /* First descriptor, not last descriptor and
> > > >> not split header */ if (status & rx_not_ls)
> > > >> -               return priv->dma_buf_sz;
> > > >> +               return priv->dma_buf_sz - NET_SKB_PAD -
> > > >> NET_IP_ALIGN;
> > > >>         plen = stmmac_get_rx_frame_len(priv, p, coe);
> > > >>
> > > >>         /* First descriptor and last descriptor and not split
> > > >> header */
> > > >> -       return min_t(unsigned int, priv->dma_buf_sz, plen);
> > > >> +       return min_t(unsigned int, priv->dma_buf_sz - NET_SKB_PAD
> > > >> - NET_IP_ALIGN, plen); }
> > > >>
> > > >>  static unsigned int stmmac_rx_buf2_len(struct stmmac_priv *priv,
> > > >
> > > > Feels like a major deficiency of the original patch. Happy to test a
> > > > more complete patch if/when you have one.
> > >
> > > I wont have time in the immediate future.
> > >
> > > Matteo, if you do not work on a fix, I suggest we revert
> > >  a955318fe67ec0d962760b5ee58e74bffaf649b8 stmmac: align RX buffers
> > >
> > > before a more polished version can be submitted.
> > >
> >
> > Better to use stmmac_rx_offset() so to have the correct length when
> > using XDP. Also, when XDP is enabled, the offset was
> > XDP_PACKET_HEADROOM (i.e. 256 bytes) even before the change, so it
> > could be already broken. Mark, can you try on the Jetson TX2 by
> > attaching an XDP program and see if it works without my change?
>
> Sorry, you'll have to hold my hand here, as I know exactly nothing
> about XDP....
>

Attach the attached object with:

ip link set eth0 xdp object passall.o

This is an empty XDP program, its source:

__attribute__((section("prog"), used))
int xdp_main(struct xdp_md *ctx)
{
       return XDP_PASS;
}

Every packet will pass untouched, but the offset will be shifted from
0 to 256 bytes, which could trigger the problem anyway:

> > A possible fix, which takes in account also the XDP headroom for
> > stmmac_rx_buf1_len() only could be (only compile tested, I don't have
> > the hardware now):
>
> However, this doesn't fix my issue. I still get all sort of
> corruption. Probably stmmac_rx_buf2_len() also need adjusting (it has
> a similar logic as its buf1 counterpart...)
>
> Unless you can fix it very quickly, and given that we're towards the
> end of the cycle, I'd be more comfortable if we reverted this patch.
>

Can it be that the HW can't do DMA on an address which is not word aligned?
What if you replace NET_SKB_PAD with, let's say, 8?

Regards,
-- 
per aspera ad upstream

Attachment: passall.o
Description: application/object