[PATCH net v2 2/2] netpoll: avoid blocking on the transmit lock in queue_process

[Karl Mehltretter <kmehltretter@gmail.com>]

queue_process() disables interrupts before taking the device transmit
lock with HARD_TX_LOCK(). The lock may sleep on PREEMPT_RT:

  BUG: sleeping function called from invalid context
  in_atomic(): 0, irqs_disabled(): 1, non_block: 0
  Workqueue: events queue_process
  rt_spin_lock
  queue_process
  process_one_work

Use HARD_TX_TRYLOCK() instead. If another CPU owns the lock, put the skb
back at the head of the deferred queue, restore interrupts and retry on
the next tick. Keep the longer HZ / 10 backoff for a stopped transmit
queue or a driver which returns busy.

Fixes: 3640543df26f ("[PATCH] netpoll: fix netpoll lockup")
Cc: stable@vger.kernel.org # 6.12+
Assisted-by: LLM
Signed-off-by: Karl Mehltretter <kmehltretter@gmail.com>
---
 net/core/netpoll.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/net/core/netpoll.c b/net/core/netpoll.c
index e0cfcb05468e2..8d8e3c33cfdfe 100644
--- a/net/core/netpoll.c
+++ b/net/core/netpoll.c
@@ -166,7 +166,13 @@ static void queue_process(struct work_struct *work)
 			skb_set_queue_mapping(skb, q_index);
 		}
 		txq = netdev_get_tx_queue(dev, q_index);
-		HARD_TX_LOCK(dev, txq, smp_processor_id());
+		if (!HARD_TX_TRYLOCK(dev, txq)) {
+			netpoll_txq_queue_head(npinfo, skb);
+			local_irq_restore(flags);
+
+			schedule_delayed_work(&npinfo->tx_work, 1);
+			return;
+		}
 		if (netif_xmit_frozen_or_stopped(txq) ||
 		    !dev_xmit_complete(netpoll_start_xmit(skb, dev, txq))) {
 			netpoll_txq_queue_head(npinfo, skb);
-- 
2.53.0