[PATCH net-next] stmmac: align RX buffers

[Marc Zyngier]

On Fri, 20 Aug 2021 18:35:45 +0100,
Matteo Croce <mcroce at linux.microsoft.com> wrote:
> 
> On Fri, Aug 20, 2021 at 7:24 PM Marc Zyngier <maz at kernel.org> wrote:
> >
> > On Fri, 20 Aug 2021 18:14:30 +0100,
> > Matteo Croce <mcroce at linux.microsoft.com> wrote:
> > >
> > > On Fri, Aug 20, 2021 at 7:09 PM Marc Zyngier <maz at kernel.org> wrote:
> > > >
> > > > On Fri, 20 Aug 2021 17:38:14 +0100,
> > > > Matteo Croce <mcroce at linux.microsoft.com> wrote:
> > > > >
> > > > > On Fri, Aug 20, 2021 at 6:26 PM Marc Zyngier <maz at kernel.org> wrote:
> > > > > >
> > > > > > On Fri, 20 Aug 2021 11:37:03 +0100,
> > > > > > Matteo Croce <mcroce at linux.microsoft.com> wrote:
> > > > > > >
> > > > > > > On Thu, Aug 19, 2021 at 6:29 PM Marc Zyngier <maz at kernel.org> wrote:
> > > > > >
> > > > > > [...]
> > > > > >
> > > > > > > > diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac.h b/drivers/net/ethernet/stmicro/stmmac/stmmac.h
> > > > > > > > index fcdb1d20389b..244aa6579ef4 100644
> > > > > > > > --- a/drivers/net/ethernet/stmicro/stmmac/stmmac.h
> > > > > > > > +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac.h
> > > > > > > > @@ -341,7 +341,7 @@ static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
> > > > > > > >         if (stmmac_xdp_is_enabled(priv))
> > > > > > > >                 return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
> > > > > > > >
> > > > > > > > -       return NET_SKB_PAD + NET_IP_ALIGN;
> > > > > > > > +       return 8 + NET_IP_ALIGN;
> > > > > > > >  }
> > > > > > > >
> > > > > > > >  void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);
> > > > > > > >
> > > > > > > > I don't see the system corrupting packets anymore. Is that exactly
> > > > > > > > what you had in mind? This really seems to point to a basic buffer
> > > > > > > > overflow.
> > > > > >
> > > > > > [...]
> > > > > >
> > > > > > > Sorry, I meant something like:
> > > > > > >
> > > > > > > -       return NET_SKB_PAD + NET_IP_ALIGN;
> > > > > > > +       return 8;
> > > > > > >
> > > > > > > I had some hardware which DMA fails if the receive buffer was not word
> > > > > > > aligned, but this seems not the case, as 8 + NET_IP_ALIGN = 10, and
> > > > > > > it's not aligned too.
> > > > > >
> > > > > > No error in that case either, as expected. Given that NET_SKB_PAD is
> > > > > > likely to expand to 64, it is likely a DMA buffer overflow which
> > > > > > probably only triggers for large-ish packets.
> > > > > >
> > > > > > Now, we're almost at -rc7, and we don't have a solution in sight.
> > > > > >
> > > > > > Can we please revert this until we have an understanding of what is
> > > > > > happening? I'll hopefully have more cycles to work on the issue once
> > > > > > 5.14 is out, and hopefully the maintainers of this driver can chime in
> > > > > > (they have been pretty quiet so far).
> > > > > >
> > > > > > Thanks,
> > > > > >
> > > > > >         M.
> > > > > >
> > > > > > --
> > > > > > Without deviation from the norm, progress is not possible.
> > > > >
> > > > > Last try, what about adding only NET_IP_ALIGN and leaving NET_SKB_PAD?
> > > > >
> > > > > -       return NET_SKB_PAD + NET_IP_ALIGN;
> > > > > +       return NET_IP_ALIGN;
> > > > >
> > > > > I think that alloc_skb adds another NET_SKB_PAD anyway.
> > > >
> > > > I don't see any packet corruption with this. However, this doesn't
> > > > prove that this is correct either. What was the rational for adding
> > > > NET_SKB_PAD the first place?
> > > >
> > >
> > > I think it's wrong. The original offset was 0, and to align it to the
> > > boundary we need to add just NET_IP_ALIGN, which is two.
> > > NET_SKB_PAD is a much bigger value, (I think 64), which is used to
> > > reserve space to prepend an header, e.g. with tunnels.
> >
> > How about the other adjustments that Eric mentioned regarding the size
> > of the buffer? Aren't they required?
> >
> 
> I guess that if stmmac_rx_buf1_len() needed such adjustment, it would
> be already broken when XDP is in use.
> When you use XDP, stmmac_rx_offset() adds a pretty big headroom of 256
> byte, which would easily trigger an overflow if not accounted.
> Did you try attaching a simple XDP program on a stock 5.13 kernel?

Yes, as mentioned in [1], to which you replied...

	M.

[1] https://lore.kernel.org/r/87wnohqty1.wl-maz@kernel.org

-- 
Without deviation from the norm, progress is not possible.