Re: [EXTERNAL] [PATCH net] net: marvell: prestera: Fix skb leak on error in prestera_rxtx_xmit()

From: Ian Rogers

Date: Fri Nov 20 2026 - 12:37:41 EST


On Thu, Nov 19, 2026 at 9:39 AM Elad Nachman <enachman@xxxxxxxxxxx> wrote:
>
> > prestera_rxtx_xmit() is invoked directly from prestera_port_xmit(), the
> > driver's .ndo_start_xmit() callback, whose return type is netdev_tx_t.
> > When skb_cow_head() or prestera_dsa_build() fails, prestera_rxtx_xmit()
> > returns NET_XMIT_DROP (0x01, a qdisc congestion code rather than a
> > netdev_tx_t value) without freeing skb.
> >
> > In dev_hard_start_xmit(), any return code other than NETDEV_TX_OK (0x00)
> > is treated as a busy/requeue condition unless dev_xmit_complete() holds;
> > for an unrecognized positive value like 0x01, the core neither frees nor
> > requeues the skb (or hits the unrecognized-return WARN path), leaking
> > the packet buffer.
>
> Did you check it yourself or was this sentence AI generated?
>
> Please check here:
> https://elixir.bootlin.com/linux/v7.4-rc1/source/include/linux/netdevice.h#L122
>
> and here:
> https://elixir.bootlin.com/linux/v7.4-rc1/source/include/linux/netdevice.h#L144
>
> You indeed need to free the SKB, but as long as rc < 0xf (which is true for both the
> NET_XMIT_DROP and the NETDEV_TX_OK cases), the behavior of dev_hard_start_xmit() is the same:
> https://elixir.bootlin.com/linux/v7.4-rc1/source/net/core/dev.c#L4074

My mistake on that paragraph - you are completely right. Because
NET_XMIT_DROP (0x01) is < NET_XMIT_MASK (0x0f), dev_xmit_complete()
returns true and dev_hard_start_xmit() assumes the driver already
consumed the skb, which is why returning without freeing skb leaks it.

> >
> > Free the skb with dev_kfree_skb_any(), increment the core tx_dropped
> > counter via dev_core_stats_tx_dropped_inc() (so .ndo_get_stats64 reports
> > the software drop), and return NETDEV_TX_OK.
> >
> > Found by enabling Clang's -Wassign-enum on an x86_64 allyesconfig build,
> > with AI coding assistant help to audit the warning and prepare the fix.
> > Build-tested with Clang; not runtime-tested on Prestera switch hardware.
> >
> > Fixes: 501ef3066c89 ("net: marvell: prestera: Add driver for Prestera family ASIC devices")
> > Signed-off-by: Ian Rogers <mailto:irogers@xxxxxxxxxx>
> > ---
> > drivers/net/ethernet/marvell/prestera/prestera_rxtx.c | 9 +++++++--
> > 1 file changed, 7 insertions(+), 2 deletions(-)
> >
> > diff --git a/drivers/net/ethernet/marvell/prestera/prestera_rxtx.c b/drivers/net/ethernet/marvell/prestera/prestera_rxtx.c
> > index c4ea247d2470..7ac9220d33c5 100644
> > --- a/drivers/net/ethernet/marvell/prestera/prestera_rxtx.c
> > +++ b/drivers/net/ethernet/marvell/prestera/prestera_rxtx.c
> > @@ -817,13 +817,18 @@ netdev_tx_t prestera_rxtx_xmit(struct prestera_port *port, struct sk_buff *skb)
> > dsa.port_num = port->hw_id;
> >
> > if (skb_cow_head(skb, PRESTERA_DSA_HLEN) < 0)
> > - return NET_XMIT_DROP;
> > + goto drop_skb;
> >
> > skb_push(skb, PRESTERA_DSA_HLEN);
> > memmove(skb->data, skb->data + PRESTERA_DSA_HLEN, 2 * ETH_ALEN);
> >
> > if (prestera_dsa_build(&dsa, skb->data + 2 * ETH_ALEN) != 0)
> > - return NET_XMIT_DROP;
> > + goto drop_skb;
> >
> > return prestera_sdma_xmit(&port->sw->rxtx->sdma, skb);
> > +
> > +drop_skb:
> > + dev_kfree_skb_any(skb);
> > + dev_core_stats_tx_dropped_inc(port->dev);
> > + return NETDEV_TX_OK;
>
> For the sake of clarity I would return here NET_XMIT_DROP . The behavior in dev_hard_start_xmit() would be the same.
> There are 53 drivers which currently return this value, this is not an isolated case of just the prestera switchdev one:
> https://elixir.bootlin.com/linux/v7.4-rc1/C/ident/NET_XMIT_DROP
>
> The rest of the patch is fine.

Sounds good (I had used NETDEV_TX_OK to match prestera_sdma_xmit()
just above and avoid Clang's -Wassign-enum on netdev_tx_t, which is
how I spotted the missing free). I'll send a [PATCH net v2] keeping
'return NET_XMIT_DROP;' and fixing the commit message.

Thanks,
Ian

> > }
> > --
> > 2.56.1.708.g77570a6a1d-goog
> >
>
> Elad.