Re: regression: tethering fails in 3.5 with iwlwifi

OK, but which netif_receive_skb(), as I count 5 occurrences in
net/mac80211/rx.c ?

Instead of skb_linearize() calls

you could try several values of XXX in

pskb_may_pull(skb, XXX)

So that you make sure XXX bytes are available in skb->head, and not
the whole frame.

One you know the limit for XXX, it might give a clue where a
pskb_may_pull(skb, XXX) is missing.

On Thu, Sep 20, 2012 at 2:35 PM, Johannes Berg
[off-list ref] wrote:

Hi,

quoted

quoted

            56138f5 iwlwifi: dont pull too much payload in skb head
            3edaf3e mac80211: manage AP netdev carrier state

The second patch (AP carrier state) actually exposed a connman issue
which I fixed and submitted a connman patch:
http://lists.connman.net/pipermail/connman/2012-September/011232.html

However, Eric's patch still causes tethering problems to me.

Let me recap a bit. Artem is using connman, which sets up the wifi
interface as part of a bridge. It runs wpa_supplicant to create an AP
(only AP and mesh mode interfaces can be bridged anyway).


Eric, you said:

quoted

I would say some part of the stack assumes a header (I dont know wich
one ?) is pulled in skb->head/data, and thats a bug.

Always use pskb_may_pull(skb, XXX) to make sure you can access XXX
bytes in skb->data

I thought we'd figure out which part of the stack assumes a header, so I
asked Artem to test a one-line patch which adds "skb_linearize()" before
"netif_receive_skb()" in mac80211. This makes it work, but I'm not sure
where after that the bad assumption might be.

johannes

--
To unsubscribe from this list: send the line "unsubscribe linux-wireless" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

I guess you only need to make sure 14 bytes of ethernet header are
available before eth_type_trans(skb, dev);

diff --git a/net/mac80211/rx.c b/net/mac80211/rx.c
index 61c621e..ffe5f84 100644
--- a/net/mac80211/rx.c
+++ b/net/mac80211/rx.c

@@ -1795,9 +1795,13 @@ ieee80211_deliver_skb(struct ieee80211_rx_data *rx)

                if (skb) {
                        /* deliver to local stack */
-                       skb->protocol = eth_type_trans(skb, dev);
-                       memset(skb->cb, 0, sizeof(skb->cb));
-                       netif_receive_skb(skb);
+                       if (pskb_may_pull(skb, sizeof(struct ethhdr))) {
+                               skb->protocol = eth_type_trans(skb, dev);
+                               memset(skb->cb, 0, sizeof(skb->cb));
+                               netif_receive_skb(skb);
+                       } else {
+                               kfree_skb(skb);
+                       }
                }
        }

On Thu, Sep 20, 2012 at 2:40 PM, Eric Dumazet [off-list ref] wrote:

OK, but which netif_receive_skb(), as I count 5 occurrences in
net/mac80211/rx.c ?

Instead of skb_linearize() calls

you could try several values of XXX in

pskb_may_pull(skb, XXX)

So that you make sure XXX bytes are available in skb->head, and not
the whole frame.

One you know the limit for XXX, it might give a clue where a
pskb_may_pull(skb, XXX) is missing.

On Thu, Sep 20, 2012 at 2:35 PM, Johannes Berg
[off-list ref] wrote:

quoted

Hi,

quoted

quoted

            56138f5 iwlwifi: dont pull too much payload in skb head
            3edaf3e mac80211: manage AP netdev carrier state

The second patch (AP carrier state) actually exposed a connman issue
which I fixed and submitted a connman patch:
http://lists.connman.net/pipermail/connman/2012-September/011232.html

However, Eric's patch still causes tethering problems to me.

Let me recap a bit. Artem is using connman, which sets up the wifi
interface as part of a bridge. It runs wpa_supplicant to create an AP
(only AP and mesh mode interfaces can be bridged anyway).


Eric, you said:

quoted

I would say some part of the stack assumes a header (I dont know wich
one ?) is pulled in skb->head/data, and thats a bug.

Always use pskb_may_pull(skb, XXX) to make sure you can access XXX
bytes in skb->data

I thought we'd figure out which part of the stack assumes a header, so I
asked Artem to test a one-line patch which adds "skb_linearize()" before
"netif_receive_skb()" in mac80211. This makes it work, but I'm not sure
where after that the bad assumption might be.

johannes

Or its the lines with CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS ?

What arch is it ?


On Thu, Sep 20, 2012 at 2:45 PM, Eric Dumazet [off-list ref] wrote:

On Thu, 2012-09-20 at 14:45 +0200, Eric Dumazet wrote:

Yeah I was looking at the same code just now. However, we had actually
inserted the skb_linearize() *after* eth_type_trans(), so I'm confused.
Maybe it still works, more or less by accident?

johannes

On Thu, 2012-09-20 at 14:47 +0200, Johannes Berg wrote:

quoted

                if (skb) {
                        /* deliver to local stack */
-                       skb->protocol = eth_type_trans(skb, dev);
-                       memset(skb->cb, 0, sizeof(skb->cb));
-                       netif_receive_skb(skb);
+                       if (pskb_may_pull(skb, sizeof(struct ethhdr))) {
+                               skb->protocol = eth_type_trans(skb, dev);
+                               memset(skb->cb, 0, sizeof(skb->cb));
+                               netif_receive_skb(skb);
+                       } else {
+                               kfree_skb(skb);
+                       }
                }
        }

Yeah I was looking at the same code just now. However, we had actually
inserted the skb_linearize() *after* eth_type_trans(), so I'm confused.

Ok actually, by the time we get here the ethernet header must be in the
skb head because we construct it from the 802.11 and llc/snap header.

johannes