net: sch_generic: aviod concurrent reset and enqueue op for lockless qdisc
authorYunsheng Lin <linyunsheng@huawei.com>
Tue, 8 Sep 2020 11:02:34 +0000 (19:02 +0800)
committerDavid S. Miller <davem@davemloft.net>
Thu, 10 Sep 2020 19:38:26 +0000 (12:38 -0700)
Currently there is concurrent reset and enqueue operation for the
same lockless qdisc when there is no lock to synchronize the
q->enqueue() in __dev_xmit_skb() with the qdisc reset operation in
qdisc_deactivate() called by dev_deactivate_queue(), which may cause
out-of-bounds access for priv->ring[] in hns3 driver if user has
requested a smaller queue num when __dev_xmit_skb() still enqueue a
skb with a larger queue_mapping after the corresponding qdisc is
reset, and call hns3_nic_net_xmit() with that skb later.

Reused the existing synchronize_net() in dev_deactivate_many() to
make sure skb with larger queue_mapping enqueued to old qdisc(which
is saved in dev_queue->qdisc_sleeping) will always be reset when
dev_reset_queue() is called.

Fixes: 6b3ba9146fe6 ("net: sched: allow qdiscs to handle locking")
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/sched/sch_generic.c

index 265a61d011dfaa7ec0f8fb8aaede920784f690c9..54c417244642a1445c618e6adf0e38b2d4f84565 100644 (file)
@@ -1131,24 +1131,10 @@ EXPORT_SYMBOL(dev_activate);
 
 static void qdisc_deactivate(struct Qdisc *qdisc)
 {
-       bool nolock = qdisc->flags & TCQ_F_NOLOCK;
-
        if (qdisc->flags & TCQ_F_BUILTIN)
                return;
-       if (test_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state))
-               return;
-
-       if (nolock)
-               spin_lock_bh(&qdisc->seqlock);
-       spin_lock_bh(qdisc_lock(qdisc));
 
        set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
-
-       qdisc_reset(qdisc);
-
-       spin_unlock_bh(qdisc_lock(qdisc));
-       if (nolock)
-               spin_unlock_bh(&qdisc->seqlock);
 }
 
 static void dev_deactivate_queue(struct net_device *dev,
@@ -1165,6 +1151,30 @@ static void dev_deactivate_queue(struct net_device *dev,
        }
 }
 
+static void dev_reset_queue(struct net_device *dev,
+                           struct netdev_queue *dev_queue,
+                           void *_unused)
+{
+       struct Qdisc *qdisc;
+       bool nolock;
+
+       qdisc = dev_queue->qdisc_sleeping;
+       if (!qdisc)
+               return;
+
+       nolock = qdisc->flags & TCQ_F_NOLOCK;
+
+       if (nolock)
+               spin_lock_bh(&qdisc->seqlock);
+       spin_lock_bh(qdisc_lock(qdisc));
+
+       qdisc_reset(qdisc);
+
+       spin_unlock_bh(qdisc_lock(qdisc));
+       if (nolock)
+               spin_unlock_bh(&qdisc->seqlock);
+}
+
 static bool some_qdisc_is_busy(struct net_device *dev)
 {
        unsigned int i;
@@ -1213,12 +1223,20 @@ void dev_deactivate_many(struct list_head *head)
                dev_watchdog_down(dev);
        }
 
-       /* Wait for outstanding qdisc-less dev_queue_xmit calls.
+       /* Wait for outstanding qdisc-less dev_queue_xmit calls or
+        * outstanding qdisc enqueuing calls.
         * This is avoided if all devices are in dismantle phase :
         * Caller will call synchronize_net() for us
         */
        synchronize_net();
 
+       list_for_each_entry(dev, head, close_list) {
+               netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
+
+               if (dev_ingress_queue(dev))
+                       dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
+       }
+
        /* Wait for outstanding qdisc_run calls. */
        list_for_each_entry(dev, head, close_list) {
                while (some_qdisc_is_busy(dev)) {