[PATCH 5.3 05/21] net/sched: fix race between deactivation and dequeue for NOLOCK qdisc

[Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>]

From: Paolo Abeni <pabeni@xxxxxxxxxx>

[ Upstream commit d518d2ed8640c1cbbbb6f63939e3e65471817367 ]

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 <shuali@xxxxxxxxxx>
Reported-and-tested-by: Davide Caratti <dcaratti@xxxxxxxxxx>
Signed-off-by: Paolo Abeni <pabeni@xxxxxxxxxx>
Signed-off-by: David S. Miller <davem@xxxxxxxxxxxxx>
Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
---
 include/net/pkt_sched.h |    7 ++++++-
 net/core/dev.c          |   16 ++++++++++------
 2 files changed, 16 insertions(+), 7 deletions(-)

--- a/include/net/pkt_sched.h
+++ b/include/net/pkt_sched.h
@@ -118,7 +118,12 @@ void __qdisc_run(struct Qdisc *q);
 static inline void qdisc_run(struct Qdisc *q)
 {
 	if (qdisc_run_begin(q)) {
-		__qdisc_run(q);
+		/* NOLOCK qdisc must check 'state' under the qdisc seqlock
+		 * to avoid racing with dev_qdisc_reset()
+		 */
+		if (!(q->flags & TCQ_F_NOLOCK) ||
+		    likely(!test_bit(__QDISC_STATE_DEACTIVATED, &q->state)))
+			__qdisc_run(q);
 		qdisc_run_end(q);
 	}
 }
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -3467,18 +3467,22 @@ static inline int __dev_xmit_skb(struct
 	qdisc_calculate_pkt_len(skb, q);
 
 	if (q->flags & TCQ_F_NOLOCK) {
-		if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
-			__qdisc_drop(skb, &to_free);
-			rc = NET_XMIT_DROP;
-		} else if ((q->flags & TCQ_F_CAN_BYPASS) && q->empty &&
-			   qdisc_run_begin(q)) {
+		if ((q->flags & TCQ_F_CAN_BYPASS) && q->empty &&
+		    qdisc_run_begin(q)) {
+			if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED,
+					      &q->state))) {
+				__qdisc_drop(skb, &to_free);
+				rc = NET_XMIT_DROP;
+				goto end_run;
+			}
 			qdisc_bstats_cpu_update(q, skb);
 
+			rc = NET_XMIT_SUCCESS;
 			if (sch_direct_xmit(skb, q, dev, txq, NULL, true))
 				__qdisc_run(q);
 
+end_run:
 			qdisc_run_end(q);
-			rc = NET_XMIT_SUCCESS;
 		} else {
 			rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
 			qdisc_run(q);