Re: [PATCH] net: macb: fix race condition in TX queue initialization

[Jakub Kicinski <kuba@kernel.org>]

On Wed, 18 Feb 2026 11:43:25 +0100 Ingo Rohloff wrote:
> Commit bf9cf80cab81e ("net: macb: Fix tx/rx malfunction after phy link
> down and up") initializes queue->tx_head and queue->tx_tail to zero
> without holding queue->tx_ptr_lock.

That commit does look quite buggy, but I'm not sure whether the fix is
effective. netif_tx_start_all_queues() and netif_tx_stop_all_queues()
do not take locks or wait for anything either. I don't see how they can
provide us with synchronization guarantees. Of course if we start with
queues stopped - the _chances_ of concurrent xmit and link_up are much
lower, but if the link happens to flap rapidly while xmit is happening
we're back to the same race.

Should we take queue->tx_ptr_lock in link_up, appropriately?
The patch you posted looks like a common sense improvement, so we can
apply it as well. But the fix should be locking and fixing whatever
questionable dark arts the link_up handler is performing on the queues
(we are resetting the pointers without freeing outstanding skbs!?).

When you repost please make sure to include a Fixes tag on the fix.
-- 
pw-bot: cr