Merge branch 'lockless-qdisc-packet-stuck'

Yunsheng Lin says:

====================
ix packet stuck problem for lockless qdisc

This patchset fixes the packet stuck problem mentioned in [1].

Patch 1: Add STATE_MISSED flag to fix packet stuck problem.
Patch 2: Fix a tx_action rescheduling problem after STATE_MISSED
         flag is added in patch 1.
Patch 3: Fix the significantly higher CPU consumption problem when
         multiple threads are competing on a saturated outgoing
         device.

V8: Change function name as suggested by Jakub and fix some typo
    in patch 3, adjust commit log in patch 2, and add Acked-by
    from Jakub.
V7: Fix netif_tx_wake_queue() data race noted by Jakub.
V6: Some performance optimization in patch 1 suggested by Jakub
    and drop NET_XMIT_DROP checking in patch 3.
V5: add patch 3 to fix the problem reported by Michal Kubecek.
V4: Change STATE_NEED_RESCHEDULE to STATE_MISSED and add patch 2.

[1]. https://lkml.org/lkml/2019/10/9/42
====================

Signed-off-by: David S. Miller <davem@davemloft.net>

1 files changed, 24 insertions, 5 deletions

diff --git a/net/core/dev.c b/net/core/dev.c
index 222b1d322c96..ef8cf7619baf 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -3853,7 +3853,8 @@ static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
 
 	if (q->flags & TCQ_F_NOLOCK) {
 		rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
-		qdisc_run(q);
+		if (likely(!netif_xmit_frozen_or_stopped(txq)))
+			qdisc_run(q);
 
 		if (unlikely(to_free))
 			kfree_skb_list(to_free);
@@ -5025,25 +5026,43 @@ static __latent_entropy void net_tx_action(struct softirq_action *h)
 		sd->output_queue_tailp = &sd->output_queue;
 		local_irq_enable();
 
+		rcu_read_lock();
+
 		while (head) {
 			struct Qdisc *q = head;
 			spinlock_t *root_lock = NULL;
 
 			head = head->next_sched;
 
-			if (!(q->flags & TCQ_F_NOLOCK)) {
-				root_lock = qdisc_lock(q);
-				spin_lock(root_lock);
-			}
 			/* We need to make sure head->next_sched is read
 			 * before clearing __QDISC_STATE_SCHED
 			 */
 			smp_mb__before_atomic();
+
+			if (!(q->flags & TCQ_F_NOLOCK)) {
+				root_lock = qdisc_lock(q);
+				spin_lock(root_lock);
+			} else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED,
+						     &q->state))) {
+				/* There is a synchronize_net() between
+				 * STATE_DEACTIVATED flag being set and
+				 * qdisc_reset()/some_qdisc_is_busy() in
+				 * dev_deactivate(), so we can safely bail out
+				 * early here to avoid data race between
+				 * qdisc_deactivate() and some_qdisc_is_busy()
+				 * for lockless qdisc.
+				 */
+				clear_bit(__QDISC_STATE_SCHED, &q->state);
+				continue;
+			}
+
 			clear_bit(__QDISC_STATE_SCHED, &q->state);
 			qdisc_run(q);
 			if (root_lock)
 				spin_unlock(root_lock);
 		}
+
+		rcu_read_unlock();
 	}
 
 	xfrm_dev_backlog(sd);