Message-ID: <6adba3bb-ec71-2dac-f6ea-2f93ee0de5ef@lab.ntt.co.jp>
Date:   Tue, 22 Jan 2019 17:45:38 +0900
From:   Toshiaki Makita <makita.toshiaki@....ntt.co.jp>
To:     Zahari Doychev <zahari.doychev@...ux.com>
Cc:     Cong Wang <xiyou.wangcong@...il.com>,
        Linux Kernel Network Developers <netdev@...r.kernel.org>,
        bridge@...ts.linux-foundation.org,
        Nikolay Aleksandrov <nikolay@...ulusnetworks.com>,
        Roopa Prabhu <roopa@...ulusnetworks.com>,
        Johannes Berg <johannes@...solutions.net>,
        Jamal Hadi Salim <jhs@...atatu.com>,
        Jiri Pirko <jiri@...nulli.us>
Subject: Re: [Bridge] [PATCH 1/2] net: bridge: fix tc added QinQ forwarding

On 2019/01/22 6:11, Zahari Doychev wrote:
> On Fri, Jan 18, 2019 at 11:29:52AM +0900, Toshiaki Makita wrote:
>> On 2019/01/18 4:19, Cong Wang wrote:
>>> On Thu, Jan 17, 2019 at 12:59 AM Toshiaki Makita
>>> <makita.toshiaki@....ntt.co.jp> wrote:
>>>> On 2019/01/17 17:17, Zahari Doychev wrote:
>>>>> On Tue, Jan 15, 2019 at 03:11:28PM +0900, Toshiaki Makita wrote:
>>>>>> On 2019/01/13 22:59, Zahari Doychev wrote:
>>>>>>> Use the skb->mac_len instead of using the ETH_HLEN when pushing the skb
>>>>>>> data pointer. This fixes sending incorrect packets when more than one
>>>>>>> vlan tags are pushed by tc-vlan and the mac header length is bigger than
>>>>>>> ETH_HLEN. In this way the vlan tagged contained in the skb is inserted at
>>>>>>> right offset in the packet payload before sending the packet.
>>>>>>>
>>>>>>> Signed-off-by: Zahari Doychev <zahari.doychev@...ux.com>
>>>>>>> ---
>>>>>>>  net/bridge/br_forward.c | 2 +-
>>>>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>>>
>>>>>>> diff --git a/net/bridge/br_forward.c b/net/bridge/br_forward.c
>>>>>>> index 5372e2042adf..55f928043f77 100644
>>>>>>> --- a/net/bridge/br_forward.c
>>>>>>> +++ b/net/bridge/br_forward.c
>>>>>>> @@ -39,7 +39,7 @@ int br_dev_queue_push_xmit(struct net *net, struct sock *sk, struct sk_buff *skb
>>>>>>>     if (!is_skb_forwardable(skb->dev, skb))
>>>>>>>             goto drop;
>>>>>>>
>>>>>>> -   skb_push(skb, ETH_HLEN);
>>>>>>> +   skb_push(skb, skb->mac_len);
>>>>>>>     br_drop_fake_rtable(skb);
>>>>>>>
>>>>>>>     if (skb->ip_summed == CHECKSUM_PARTIAL &&
>>>>>>>
>>>>>>
>>>>>> I guess you mean skb->data points to mac_header + ETH_HLEN + VLAN_HLEN
>>>>>> when bridge receives skbs in br_handle_frame()?
>>>>>
>>>>> yes, this is what I see.
>>>>>
>>>>>> If so, the behavior of act_vlan is odd. Normal double tagged skbs from
>>>>>> hardware devices should have skb->data pointing to mac_header + ETH_HLEN
>>>>>> because they just call eth_type_trans() before entering
>>>>>> netif_receive_skb()...
>>>>>> I think act_vlan needs some fix.
>>>>>
>>>>> The act_valn is using the skb_vlan_push(...) to add the vlan tags and in this
>>>>> way increasing the skb->data and mac_len. So I think I can add a fix there to
>>>>> set the skb->data to point to mac_header + ETH_HLEN when more tags are added.
>>>>
>>>> As skb->data always points to mac_header after calling skb_vlan_push(),
>>>> we probably need to remember mac_len before invocation of it?
>>>>
>>>> The problem should be this part in tcf_vlan_act():
>>>>
>>>>> out:
>>>>>       if (skb_at_tc_ingress(skb))
>>>>>               skb_pull_rcsum(skb, skb->mac_len);
>>>>
>>>> skb->mac_len should not be used here.
>>>
>>> I am confused. This code is to push skb->data back to network header.
>>> If skb_vlan_push() pushes skb->data to mac header, then this code
>>> is correct for pulling it back to network header, as skb->mac_len is
>>> updated accordingly inside skb_vlan_push() too.
>>>
>>> What goes wrong here? skb->mac_len isn't correct for double tagging?
>>
>> Bridge and VLAN code (skb_vlan_untag in __netif_receive_skb_core)
>> expects skb->data to point to the start of VLAN header, not the next
>> (network) header. After calling tcf_vlan_act() on ingress filter,
>> skb->data points to the next of VLAN header (network header), while
>> non-hwaccel VLAN packets (including double tagged ones) from NIC drivers
>> have skb->data pointing to the start of VLAN header as expected.
>>
>> I'm not sure if mac_len should or should not be updated in
>> skb_vlan_push(). Anyway IIRC skb_vlan_untag() and bridge code do not
>> rely on mac_len so mac_len should not cause problems there.
> 
> Replacing the usage of the skb->maclen in act vlan with ETH_HLEN in the
> skb_push/pull calls fixes the problem. I have to do some more testing then
> I can send a new patch.

Using the original mac_len (before entering skb_vlan_push()) would be
safer, since mac_len is not necessarily be ETH_HLEN (with reorder_hdr
off, setting vlan_protocol in tun_pi, ...).

Actually a number of VLAN-related code assumes ETH_HLEN, though...

> I still see the problem in br_vlan __allowed_ingress when the skb->vlan_proto 
> and br->vlan_proto does not match. If the network header is not reset then the 
> flow dissector does not correctly detect the vlan tags. The the network header
> points to the second tag where as the skb->data points to the first one and the
> egress rule in my example cannot match. Is this the correct way to fix this
> and is __allowed_ingress the correct place?

network_header needs to point to L3 (IP) header, at least with
CHECKSUM_PARTIAL. network_header offset is fixed up in
br_dev_queue_push_xmit() for such packets, so the fix at
__allowed_ingress() would not be a complete fix.
Requiring network_header to be at VLAN header is not compatible with
drivers that support CHECKSUM_PARTIAL. I guess something needs to be
fixed in flower, if it has such a requirement.

-- 
Toshiaki Makita

Powered by blists - more mailing lists