From: Paolo Abeni <pab...@redhat.com>
Date: Thu, 12 Sep 2019 12:02:42 +0200
> The test implemented by some_qdisc_is_busy() is somewhat loosy for
> NOLOCK qdisc, as we may hit the following scenario:
>
> CPU1 CPU2
> // in net_tx_action()
> clear_bit(__QDISC_STATE_SCHED...);
> // in some_qdisc_is_busy()
> val = (qdisc_is_running(q) ||
>
> test_bit(__QDISC_STATE_SCHED,
> &q->state));
> // here val is 0 but...
> qdisc_run(q)
> // ... CPU1 is going to run the qdisc next
>
> As a conseguence qdisc_run() in net_tx_action() can race with qdisc_reset()
> in dev_qdisc_reset(). Such race is not possible for !NOLOCK qdisc as
> both the above bit operations are under the root qdisc lock().
>
> After commit 021a17ed796b ("pfifo_fast: drop unneeded additional lock on
> dequeue")
> the race can cause use after free and/or null ptr dereference, but the root
> cause is likely older.
>
> This patch addresses the issue explicitly checking for deactivation under
> the seqlock for NOLOCK qdisc, so that the qdisc_run() in the critical
> scenario becomes a no-op.
>
> Note that the enqueue() op can still execute concurrently with
> dev_qdisc_reset(),
> but that is safe due to the skb_array() locking, and we can't avoid that
> for NOLOCK qdiscs.
>
> Fixes: 021a17ed796b ("pfifo_fast: drop unneeded additional lock on dequeue")
> Reported-by: Li Shuang <shu...@redhat.com>
> Reported-and-tested-by: Davide Caratti <dcara...@redhat.com>
> Signed-off-by: Paolo Abeni <pab...@redhat.com>
Applied and queued up for -stable, thanks.
