Linux Ethernet Bridging - Aug 2020 - [Bridge] [PATCH net v2] net: bridge: clear skb private space on bridge dev xmit

On 31/07/2020 20:37, Nikolay Aleksandrov wrote:
> On 31/07/2020 20:27, David Ahern wrote:
>> On 7/31/20 10:26 AM, Nikolay Aleksandrov wrote:
>>> We need to clear all of the bridge private skb variables as they
can be
>>> stale due to the packet being recirculated through the stack and
then
>>> transmitted through the bridge device. Similar memset is already
done on
>>> bridge's input. We've seen cases where proxyarp_replied was
1 on routed
>>> multicast packets transmitted through the bridge to ports with
neigh
>>> suppress which were getting dropped. Same thing can in theory
happen with
>>> the port isolation bit as well.
>>>
>>> Fixes: 821f1b21cabb ("bridge: add new BR_NEIGH_SUPPRESS port
flag to suppress arp and nd flood")
>>> Signed-off-by: Nikolay Aleksandrov <nikolay at
cumulusnetworks.com>
>>> ---
>>>  net/bridge/br_device.c | 2 ++
>>>  1 file changed, 2 insertions(+)
>>>
>>> diff --git a/net/bridge/br_device.c b/net/bridge/br_device.c
>>> index 8c7b78f8bc23..9a2fb4aa1a10 100644
>>> --- a/net/bridge/br_device.c
>>> +++ b/net/bridge/br_device.c
>>> @@ -36,6 +36,8 @@ netdev_tx_t br_dev_xmit(struct sk_buff *skb,
struct net_device *dev)
>>>  	const unsigned char *dest;
>>>  	u16 vid = 0;
>>>  
>>> +	memset(skb->cb, 0, sizeof(struct br_input_skb_cb));
>>> +
>>>  	rcu_read_lock();
>>>  	nf_ops = rcu_dereference(nf_br_ops);
>>>  	if (nf_ops && nf_ops->br_dev_xmit_hook(skb)) {
>>>
>>
>> What's the performance hit of doing this on every packet?
>>
>> Can you just set a flag that tells the code to reset on recirculation?
>> Seems like br_input_skb_cb has space for that.
>>
> 
> Virtually non-existent, we had a patch that turned that field into a 16
byte
> field so that is really 2 8 byte stores. It is already cache hot, we could
> initialize each individual field separately as br_input does.
> 
> I don't want to waste flags on such thing, this makes it future-proof 
> and I'll remove the individual field zeroing later which will alleviate
> the cost further.
> 
Also note that we already do this on input for each packet since the
struct was reduced to 16 bytes. It's the safest way since every different
sub-part of the bridge uses some set of these private variables and
we've had many similar bugs where they were used stale or unintentionally
were not initialized for some path.

On 31/07/2020 20:51, Nikolay Aleksandrov wrote:
> On 31/07/2020 20:37, Nikolay Aleksandrov wrote:
>> On 31/07/2020 20:27, David Ahern wrote:
>>> On 7/31/20 10:26 AM, Nikolay Aleksandrov wrote:
>>>> We need to clear all of the bridge private skb variables as
they can be
>>>> stale due to the packet being recirculated through the stack
and then
>>>> transmitted through the bridge device. Similar memset is
already done on
>>>> bridge's input. We've seen cases where proxyarp_replied
was 1 on routed
>>>> multicast packets transmitted through the bridge to ports with
neigh
>>>> suppress which were getting dropped. Same thing can in theory
happen with
>>>> the port isolation bit as well.
>>>>
>>>> Fixes: 821f1b21cabb ("bridge: add new BR_NEIGH_SUPPRESS
port flag to suppress arp and nd flood")
>>>> Signed-off-by: Nikolay Aleksandrov <nikolay at
cumulusnetworks.com>
>>>> ---
>>>>  net/bridge/br_device.c | 2 ++
>>>>  1 file changed, 2 insertions(+)
>>>>
>>>> diff --git a/net/bridge/br_device.c b/net/bridge/br_device.c
>>>> index 8c7b78f8bc23..9a2fb4aa1a10 100644
>>>> --- a/net/bridge/br_device.c
>>>> +++ b/net/bridge/br_device.c
>>>> @@ -36,6 +36,8 @@ netdev_tx_t br_dev_xmit(struct sk_buff *skb,
struct net_device *dev)
>>>>  	const unsigned char *dest;
>>>>  	u16 vid = 0;
>>>>  
>>>> +	memset(skb->cb, 0, sizeof(struct br_input_skb_cb));
>>>> +
>>>>  	rcu_read_lock();
>>>>  	nf_ops = rcu_dereference(nf_br_ops);
>>>>  	if (nf_ops && nf_ops->br_dev_xmit_hook(skb)) {
>>>>
>>>
>>> What's the performance hit of doing this on every packet?
>>>
>>> Can you just set a flag that tells the code to reset on
recirculation?
>>> Seems like br_input_skb_cb has space for that.
>>>
>>
>> Virtually non-existent, we had a patch that turned that field into a 16
byte
>> field so that is really 2 8 byte stores. It is already cache hot, we
could
>> initialize each individual field separately as br_input does.
>>
>> I don't want to waste flags on such thing, this makes it
future-proof
>> and I'll remove the individual field zeroing later which will
alleviate
>> the cost further.
>>
> 
> Also note that we already do this on input for each packet since the
> struct was reduced to 16 bytes. It's the safest way since every
different
> sub-part of the bridge uses some set of these private variables and
> we've had many similar bugs where they were used stale or
unintentionally
> were not initialized for some path.
> 
In addition this doesn't need to be a recirculation, in theory it could
happen
by a routed packet to svi on the bridge which got its skb->cb initialized
before
hitting the bridge's xmit function. So a flag can't catch all possible
cases.

We need to clear all of the bridge private skb variables as they can be
stale due to the packet having skb->cb initialized earlier and then
transmitted through the bridge device. Similar memset is already done on
bridge's input. We've seen cases where proxyarp_replied was 1 on routed
multicast packets transmitted through the bridge to ports with neigh
suppress and were getting dropped. Same thing can in theory happen with the
port isolation bit as well. We clear only the struct part after the bridge
pointer (currently 8 bytes) since the pointer is always set later.
We can now remove the redundant zeroing of frag_max_size.
Also add a BUILD_BUG_ON to make sure we catch any movement of the bridge
dev pointer.

Fixes: 821f1b21cabb ("bridge: add new BR_NEIGH_SUPPRESS port flag to
suppress arp and nd flood")
Signed-off-by: Nikolay Aleksandrov <nikolay at cumulusnetworks.com>
---
v2: clear only the second half of the struct which contains the
fields that are used in various bridge parts, this replaced the rep stos
instruction with a single movq on my x86 and in general reduces
the clear area to 8 bytes, and in addition we can remove the now
redundant zeroing of frag_max_size as it will be already cleared,
add a build_bug_on to make sure we catch any movement of the bridge
dev pointer

 net/bridge/br_device.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/net/bridge/br_device.c b/net/bridge/br_device.c
index 8c7b78f8bc23..4f7880c99d3c 100644
--- a/net/bridge/br_device.c
+++ b/net/bridge/br_device.c
@@ -36,6 +36,12 @@ netdev_tx_t br_dev_xmit(struct sk_buff *skb, struct
net_device *dev)
 	const unsigned char *dest;
 	u16 vid = 0;
 
+	/* clear all private fields after the bridge dev pointer */
+	BUILD_BUG_ON(offsetof(struct br_input_skb_cb, brdev) > 0);
+	memset(skb->cb + sizeof(struct net_device *),
+	       0,
+	       sizeof(struct br_input_skb_cb) - sizeof(struct net_device *));
+
 	rcu_read_lock();
 	nf_ops = rcu_dereference(nf_br_ops);
 	if (nf_ops && nf_ops->br_dev_xmit_hook(skb)) {
@@ -50,7 +56,6 @@ netdev_tx_t br_dev_xmit(struct sk_buff *skb, struct net_device
*dev)
 
 	br_switchdev_frame_unmark(skb);
 	BR_INPUT_SKB_CB(skb)->brdev = dev;
-	BR_INPUT_SKB_CB(skb)->frag_max_size = 0;
 
 	skb_reset_mac_header(skb);
 	skb_pull(skb, ETH_HLEN);
-- 
2.25.4