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

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: bec251b ("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>

include/linux/netdevice.h

@@ -3325,6 +3325,12 @@ static inline void netif_tx_wake_all_queues(struct net_device *dev)
 
 static __always_inline void netif_tx_stop_queue(struct netdev_queue *dev_queue)
 {
+	/* Paired with READ_ONCE() from dev_watchdog() */
+	WRITE_ONCE(dev_queue->trans_start, jiffies);
+
+	/* This barrier is paired with smp_mb() from dev_watchdog() */
+	smp_mb__before_atomic();
+
 	/* Must be an atomic op see netif_txq_try_stop() */
 	set_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state);
 }
@@ -3451,6 +3457,12 @@ static inline void netdev_tx_sent_queue(struct netdev_queue *dev_queue,
 	if (likely(dql_avail(&dev_queue->dql) >= 0))
 		return;
 
+	/* Paired with READ_ONCE() from dev_watchdog() */
+	WRITE_ONCE(dev_queue->trans_start, jiffies);
+
+	/* This barrier is paired with smp_mb() from dev_watchdog() */
+	smp_mb__before_atomic();
+
 	set_bit(__QUEUE_STATE_STACK_XOFF, &dev_queue->state);
 
 	/*

net/sched/sch_generic.c

@@ -512,9 +512,15 @@ static void dev_watchdog(struct timer_list *t)
 				struct netdev_queue *txq;
 
 				txq = netdev_get_tx_queue(dev, i);
-				trans_start = READ_ONCE(txq->trans_start);
 				if (!netif_xmit_stopped(txq))
 					continue;
+
+				/* Paired with WRITE_ONCE() + smp_mb...() in
+				 * netdev_tx_sent_queue() and netif_tx_stop_queue().
+				 */
+				smp_mb();
+				trans_start = READ_ONCE(txq->trans_start);
+
 				if (time_after(jiffies, trans_start + dev->watchdog_timeo)) {
 					timedout_ms = jiffies_to_msecs(jiffies - trans_start);
 					atomic_long_inc(&txq->trans_timeout);