fix potential wild pointer when NIC is dying

net/core/dev.c

Quoted 4 lines · Changli Gao

net/core/dev.c +0 -0

index a10a216..fe4a821 100644

[]

net/core/dev.c int netif_receive_skb(struct sk_buff *skb) · line 2795

Quoted 2 lines

net/core/dev.c +1 -1

-static void flush_backlog(void *arg)
+static int flush_backlog(void *arg)

Why change this to return int?

Quoted 1 line

+     return 0;

It seems to always return 0.

fix potential wild pointer when NIC is dying.

flush_backlog() works with the assumption: the NIC doesn't enqueue packets to linux kernel, so there are two places, which packets are in, softnet queue or being processed in net-rx softirq. flush_backlog() is used to drop the first kind of packets, and for the later, a grace period is used to wait the finishing of the packets processing.

It always works without RPS. If RPS is used, although the NIC doesn't enqueue packets to linux kernel, RPS may do. There may be condition, a grace period has passed due to softirq running time limit, there are still packets, which refer to the died NIC, and are enqueued by RPS after flush_backlog() returns.

Signed-off-by Changli Gao <[email protected]>

----

 net/core/dev.c |   24 +++++++++++++++---------
 1 file changed, 15 insertions(+), 9 deletions(-)

net/core/dev.c +15 -10

index a10a216..fe4a821 100644
@@ -131,6 +131,7 @@
 #include <linux/random.h>
 #include <trace/events/napi.h>
 #include <linux/pci.h>
+#include <linux/stop_machine.h>
 
 #include "net-sysfs.h"
 

@@ -2791,19 +2792,24 @@ int netif_receive_skb(struct sk_buff *skb)
 EXPORT_SYMBOL(netif_receive_skb);
 
 /* Network device is going away, flush any packets still pending  */
-static void flush_backlog(void *arg)
+static int flush_backlog(void *arg)
 {
 	struct net_device *dev = arg;
-	struct softnet_data *queue = &__get_cpu_var(softnet_data);
 	struct sk_buff *skb, *tmp;
+	struct softnet_data *queue;
+	int cpu;
 
-	rps_lock(queue);
-	skb_queue_walk_safe(&queue->input_pkt_queue, skb, tmp)
-		if (skb->dev == dev) {
-			__skb_unlink(skb, &queue->input_pkt_queue);
-			kfree_skb(skb);
+	for_each_online_cpu(cpu) {
+		queue = &per_cpu(softnet_data, cpu);
+		skb_queue_walk_safe(&queue->input_pkt_queue, skb, tmp) {
+			if (skb->dev == dev) {
+				__skb_unlink(skb, &queue->input_pkt_queue);
+				kfree_skb(skb);
+			}
 		}
-	rps_unlock(queue);
+	}
+
+	return 0;
 }
 
 static int napi_gro_complete(struct sk_buff *skb)

@@ -5027,7 +5033,7 @@ void netdev_run_todo(void)
 
 		dev->reg_state = NETREG_UNREGISTERED;
 
-		on_each_cpu(flush_backlog, dev, 1);
+		stop_machine(flush_backlog, dev, NULL);
 
 		netdev_wait_allrefs(dev);
 
--

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

Le mercredi 14 avril 2010 à 20:18 +0800, Changli Gao a écrit :

I dont see how the problem can happens, and how RPS is involved.

Did you got a single panic, could you provide us a stack trace ?

Maybe are you referring to NAPI ?

NAPI process packets delivered by NIC, and through RPS deliver it to a (possibly) remote CPU queue.

But at device dismantle time, we should stop NAPI on this device and packet delivery machinery. RPS being on or not, NAPI wont deliver new packets. The fact that NAPI can be throtled doesnt change the napi instance being disabled at this point. No more packet will be delivered (RPS or not)

Only after this point we call flush_backlog() to make sure we dont have any queued packet in each cpu input_pkt_queue pointing to the device we dismantle.

RPS doesnt change this at all.

Hmm ???

Thanks, I got it.