net: fix races in netdev_tx_sent_queue()/dev_watchdog()

[ Upstream commit 95ecba62 ]

Some workloads hit the infamous dev_watchdog() message:

"NETDEV WATCHDOG: eth0 (xxxx): transmit queue XX timed out"

It seems possible to hit this even for perfectly normal
BQL enabled drivers:

1) Assume a TX queue was idle for more than dev->watchdog_timeo
   (5 seconds unless changed by the driver)

2) Assume a big packet is sent, exceeding current BQL limit.

3) Driver ndo_start_xmit() puts the packet in TX ring,
   and netdev_tx_sent_queue() is called.

4) QUEUE_STATE_STACK_XOFF could be set from netdev_tx_sent_queue()
   before txq->trans_start has been written.

5) txq->trans_start is written later, from netdev_start_xmit()

    if (rc == NETDEV_TX_OK)
          txq_trans_update(txq)

dev_watchdog() running on another cpu could read the old
txq->trans_start, and then see QUEUE_STATE_STACK_XOFF, because 5)
did not happen yet.

To solve the issue, write txq->trans_start right before one XOFF bit
is set :

- _QUEUE_STATE_DRV_XOFF from netif_tx_stop_queue()
- __QUEUE_STATE_STACK_XOFF from netdev_tx_sent_queue()

From dev_watchdog(), we have to read txq->state before txq->trans_start.

Add memory barriers to enforce correct ordering.

In the future, we could avoid writing over txq->trans_start for normal
operations, and rename this field to txq->xoff_start_time.

Fixes: bec251bc ("net: no longer stop all TX queues in dev_watchdog()")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://patch.msgid.link/20241015194118.3951657-1-edumazet@google.com


Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>