Re: [PATCH net v3 3/3] net: stmmac: document oversized AF_XDP frame handling

On 08/21, Maciej Fijalkowski wrote:
> On Thu, Aug 20, 2026 at 06:29:16PM -0700, Stanislav Fomichev wrote:
> > On 08/20, Maciej Fijalkowski wrote:
> > > On Wed, Aug 19, 2026 at 09:05:35AM -0700, Stanislav Fomichev wrote:
> > > > stmmac drops AF_XDP zero-copy frames that exceed taprio's queueMaxSDU
> > > > after xsk_tx_peek_desc() has reserved their completion entries.
> > > > 
> > > > Completing a rejected descriptor is unsafe because AF_XDP completions are
> > > > ordered: xsk_tx_completed(pool, 1) would complete the oldest outstanding
> > > > descriptor, which may still be owned by hardware. Instead, leave the
> > > > completion pending so the ring eventually wedges and increment the drop
> > > > counter to expose the application error without risking hardware
> > > > misbehavior.
> > > > 
> > > > Document this intentional ring imbalance at the check.
> > > > 
> > > > Signed-off-by: Stanislav Fomichev <[email protected]>
> > > > ---
> > > >  drivers/net/ethernet/stmicro/stmmac/stmmac_main.c | 4 ++++
> > > >  1 file changed, 4 insertions(+)
> > > > 
> > > > diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> > > > index 62de03e65a90..6a532747c039 100644
> > > > --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> > > > +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> > > > @@ -2713,6 +2713,10 @@ static bool stmmac_xdp_xmit_zc(struct stmmac_priv *priv, u32 queue, u32 budget)
> > > >  		if (priv->est && priv->est->enable &&
> > > >  		    priv->est->max_sdu[queue] &&
> > > >  		    xdp_desc.len > priv->est->max_sdu[queue]) {
> > > > +			/* Completions are ordered, so this descriptor cannot
> > > > +			 * be completed safely. Wedge the ring to expose the
> > > > +			 * application error instead.
> > > > +			 */
> > > >  			priv->xstats.max_sdu_txq_drop[queue]++;
> > > >  			continue;
> > > 
> > > Hmm. I read the discussion on v2. Maybe we could cancel cq entry here in
> > > this branch? Also it feels like something achievable at bind time when
> > > taprio is configured and vice versa?
> > > 
> > > Otherwise we over-commit cq entries.
> > 
> > What do you want to achieve with the cancel here? IIUC it will make it look
> > as if some (if the user has posted many) tx descriptor has not been consumed
> > by the kernel?
> 
> Oof. My bad. I meant completely different thing :D
> 
> Right now the semantics are that we post invalid/dropped addrs to cq (the
> rationale was that dropped descs are gone and unreachable which might
> eventually lead to dying traffic).
> 
> We should submit xdp_desc's addr to cq.
> 
> Regarding the comment included in code I must disagree. CQ entries no
> longer imply that 'this particular descriptor has been successfully sent
> by HW'.

But then we need to support some sort of out-of-order completions, no?
This looks similar to
https://lore.kernel.org/netdev/[email protected]/

We write desc to cq at xsk_tx_peek_desc, so when we get an error here
we might have already "queued" a bunch of cq entries (which will be
xsk_tx_completed(num) from sirq). So unless we rewrite the way
we do completions, there is no easy way to put that desc on cq without
breaking the order and racing with real completions from the HW.
Am I missing something here?

> > I do agree that a better idea is to probably do these checks during control
> > paths, but it's a bit more involved (and not sure if it's possible? if we
> > have a bunch of xsk sockets and we change that max_sdu, do we go over all
> > sockets on the system somehow?). My main motivation with this patch was
> > to make our LLM reviewers less chatty about preexisting issues.
> 
> I hear you, however I feel like we do not know this driver too much and
> probably we don't have a HW to test out such changes, so maybe let us try
> to fix existing behavior?

Let's definitely fix it properly if you have better ideas. But since we
don't have HW, I'm not super confident doing anything sophisticated
myself :-D