net: call rcu_read_lock early in process_backlog
authorJulian Anastasov <ja@ssi.bg>
Thu, 9 Jul 2015 06:59:10 +0000 (09:59 +0300)
committerBen Hutchings <ben@decadent.org.uk>
Wed, 12 Aug 2015 14:33:21 +0000 (16:33 +0200)
commit 2c17d27c36dcce2b6bf689f41a46b9e909877c21 upstream.

Incoming packet should be either in backlog queue or
in RCU read-side section. Otherwise, the final sequence of
flush_backlog() and synchronize_net() may miss packets
that can run without device reference:

CPU 1                  CPU 2
                       skb->dev: no reference
                       process_backlog:__skb_dequeue
                       process_backlog:local_irq_enable

on_each_cpu for
flush_backlog =>       IPI(hardirq): flush_backlog
                       - packet not found in backlog

                       CPU delayed ...
synchronize_net
- no ongoing RCU
read-side sections

netdev_run_todo,
rcu_barrier: no
ongoing callbacks
                       __netif_receive_skb_core:rcu_read_lock
                       - too late
free dev
                       process packet for freed dev

Fixes: 6e583ce5242f ("net: eliminate refcounting in backlog queue")
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: Julian Anastasov <ja@ssi.bg>
Signed-off-by: David S. Miller <davem@davemloft.net>
[bwh: Backported to 3.2:
 - Adjust context
 - No need to rename the label in __netif_receive_skb()]
Signed-off-by: Ben Hutchings <ben@decadent.org.uk>
net/core/dev.c

index 42895aa..7f43202 100644 (file)
@@ -3261,8 +3261,6 @@ static int __netif_receive_skb(struct sk_buff *skb)
 
        pt_prev = NULL;
 
-       rcu_read_lock();
-
 another_round:
 
        __this_cpu_inc(softnet_data.processed);
@@ -3357,7 +3355,6 @@ ncls:
        }
 
 out:
-       rcu_read_unlock();
        return ret;
 }
 
@@ -3378,34 +3375,31 @@ out:
  */
 int netif_receive_skb(struct sk_buff *skb)
 {
+       int ret;
+
        if (netdev_tstamp_prequeue)
                net_timestamp_check(skb);
 
        if (skb_defer_rx_timestamp(skb))
                return NET_RX_SUCCESS;
 
+       rcu_read_lock();
+
 #ifdef CONFIG_RPS
        {
                struct rps_dev_flow voidflow, *rflow = &voidflow;
-               int cpu, ret;
-
-               rcu_read_lock();
-
-               cpu = get_rps_cpu(skb->dev, skb, &rflow);
+               int cpu = get_rps_cpu(skb->dev, skb, &rflow);
 
                if (cpu >= 0) {
                        ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
                        rcu_read_unlock();
-               } else {
-                       rcu_read_unlock();
-                       ret = __netif_receive_skb(skb);
+                       return ret;
                }
-
-               return ret;
        }
-#else
-       return __netif_receive_skb(skb);
 #endif
+       ret = __netif_receive_skb(skb);
+       rcu_read_unlock();
+       return ret;
 }
 EXPORT_SYMBOL(netif_receive_skb);
 
@@ -3796,8 +3790,10 @@ static int process_backlog(struct napi_struct *napi, int quota)
                unsigned int qlen;
 
                while ((skb = __skb_dequeue(&sd->process_queue))) {
+                       rcu_read_lock();
                        local_irq_enable();
                        __netif_receive_skb(skb);
+                       rcu_read_unlock();
                        local_irq_disable();
                        input_queue_head_incr(sd);
                        if (++work >= quota) {