On Thu, Jun 15, 2023 at 01:27:47PM +0200, Simon Horman wrote:
> On Thu, Jun 15, 2023 at 11:55:29AM +0200, Maciej Fijalkowski wrote:
> > On Thu, Jun 15, 2023 at 11:45:36AM +0200, Simon Horman wrote:
>
> Hi Marciej,
>
> > Hi Simon,
> >
> > > In igc_clean_rx_irq() the result of a call to igc_xdp_run_prog() is assigned
> > > to the skb local variable. This may be an ERR_PTR.
> > >
> > > A little later the following is executed, which seems to be a
> > > possible dereference of an ERR_PTR.
> > >
> > > total_bytes += skb->len;
> > >
> > > Avoid this problem by continuing the loop in which all of the
> > > above occurs once the handling of the NULL case completes.
> > >
> > > This proposed fix is speculative - I do not have deep knowledge of this
> > > driver. And I am concerned about the effect of skipping the following
> > > logic:
> > >
> > > igc_put_rx_buffer(rx_ring, rx_buffer, rx_buffer_pgcnt);
> > > cleaned_count++;
> >
> > this will break - you have to recycle the buffer to have it going.
>
> Thanks. As I said I wasn't sure about the fix: it was a strawman.
>
> > > Flagged by Smatch as:
> > >
> > > .../igc_main.c:2467 igc_xdp_run_prog() warn: passing zero to 'ERR_PTR'
> >
> > how about PTR_ERR_OR_ZERO() ? this would silence smatch and is not an
> > intrusive change. another way is to get rid of ERR_PTR() around skb/xdp
> > run result but i think the former would be just fine.
>
> Sorry, there were two warnings. And I accidently trimmed the one
> that is more relevant instead of the one that is less relevant.
> I do agree the one above does not appear to be a bug.
>
> But I am concerned abut this one:
>
> .../igc_main.c:2618 igc_clean_rx_irq() error: 'skb' dereferencing possible ERR_PTR()
>
> If skb is an error pointer, e.g. ERR_PTR(-IGC_XDP_PASS), and
> it is dereferenced, that would be a problem, right?
IGC_XDP_PASS is 0. -0 is still 0 right?
this means skb is NULL and igc_{build,construct}_skb() will init it. For
ERR_PTR, igc_cleanup_headers() does IS_ERR() against it and continues. So
you will get to line 2618 only for valid skb, it just happens that logic
is written in a way that skb is supposed to carry XDP return code. We
removed this in ice for example but i40e works like that for many years
without issues, AFAICT...
>
> Perhaps I'm missing something obvious and this can't occur.
> But it does seem possible to me.
>
> >
> > >
> > > Compile tested only.
> > >
> > > Fixes: 26575105d6ed ("igc: Add initial XDP support")
> > > Signed-off-by: Simon Horman <horms@xxxxxxxxxx>
> > > ---
> > > drivers/net/ethernet/intel/igc/igc_main.c | 1 +
> > > 1 file changed, 1 insertion(+)
> > >
> > > diff --git a/drivers/net/ethernet/intel/igc/igc_main.c b/drivers/net/ethernet/intel/igc/igc_main.c
> > > index 88145c30c919..b58c8a674bd1 100644
> > > --- a/drivers/net/ethernet/intel/igc/igc_main.c
> > > +++ b/drivers/net/ethernet/intel/igc/igc_main.c
> > > @@ -2586,6 +2586,7 @@ static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget)
> > >
> > > total_packets++;
> > > total_bytes += size;
> > > + continue;
> > > } else if (skb)
> > > igc_add_rx_frag(rx_ring, rx_buffer, skb, size);
> > > else if (ring_uses_build_skb(rx_ring))
> > >
> > >
> >
