Re: [PATCH net v2 1/9] bridge: Fix the way to find old local fdb entries in br_fdb_changeaddr

[Vlad Yasevich <vyasevic@redhat.com>]

On 01/07/2014 11:33 AM, Toshiaki Makita wrote:
> On Tue, 2014-01-07 at 09:44 -0500, Vlad Yasevich wrote:
>> On 01/07/2014 07:42 AM, Toshiaki Makita wrote:
>>> On Mon, 2014-01-06 at 06:29 -0500, Vlad Yasevich wrote:
>>>> On 01/05/2014 10:26 AM, Toshiaki Makita wrote:
>>>>> On Fri, 2014-01-03 at 15:46 -0500, Vlad Yasevich wrote:
>>>>>> On 01/03/2014 02:28 PM, Vlad Yasevich wrote:
>>>>>>> On 12/17/2013 07:03 AM, Toshiaki Makita wrote:
>>>>>>>> br_fdb_changeaddr() assumes that there is at most one local entry per port
>>>>>>>> per vlan. It used to be true, but since commit 36fd2b63e3b4 ("bridge: allow
>>>>>>>> creating/deleting fdb entries via netlink"), it has not been so.
>>>>>>>> Therefore, the function might fail to search a correct previous address
>>>>>>>> to be deleted and delete an arbitrary local entry if user has added local
>>>>>>>> entries manually.
>>>>>>>>
>>>>>>>> Example of problematic case:
>>>>>>>>   ip link set eth0 address ee:ff:12:34:56:78
>>>>>>>>   brctl addif br0 eth0
>>>>>>>>   bridge fdb add 12:34:56:78:90:ab dev eth0 master
>>>>>>>>   ip link set eth0 address aa:bb:cc:dd:ee:ff
>>>>>>>> Then, the address 12:34:56:78:90:ab might be deleted instead of
>>>>>>>> ee:ff:12:34:56:78, the original mac address of eth0.
>>>>>>>>
>>>>>>>> Address this issue by introducing a new flag, added_by_user, to struct
>>>>>>>> net_bridge_fdb_entry.
>>>>>>>>
>>>>>>>> Note that br_fdb_delete_by_port() has to set added_by_user to 0 in case
>>>>>>>> like:
>>>>>>>>   ip link set eth0 address 12:34:56:78:90:ab
>>>>>>>>   ip link set eth1 address aa:bb:cc:dd:ee:ff
>>>>>>>>   brctl addif br0 eth0
>>>>>>>>   bridge fdb add aa:bb:cc:dd:ee:ff dev eth0 master
>>>>>>>>   brctl addif br0 eth1
>>>>>>>>   brctl delif br0 eth0
>>>>>>>> In this case, kernel should delete the user-added entry aa:bb:cc:dd:ee:ff,
>>>>>>>> but it also should have been added by "brctl addif br0 eth1" originally,
>>>>>>>> so we don't delete it and treat it a new kernel-created entry.
>>>>>>>>
>>>>>>>
>>>>>>> I was looking over my patch series that adds something similar to this
>>>>>>> and noticed that you are not handing the NTF_USE case.  That case was
>>>>>>> always troublesome for me as it allows for 2 different way to create
>>>>>>> the same FDB: one through br_fdb_update() and one through fdb_add_entry().
>>>>>>>
>>>>>>> It is possible, though I haven't found any users yet, that NTF_USE
>>>>>>> may be used and in that case, bridge will create a dynamic fdb and
>>>>>>> disregard all NUD flags.  In case case, add_by_user will not be set
>>>>>>> either.
>>>>>>>
>>>>>>> I think that the above is broken and plan to submit a fix shortly.
>>>>>>
>>>>>> Just looked again at my NTF_USE patch and while it seems ok, the whole
>>>>>> NTF_USE usage is racy to begin with and I am really starting to question
>>>>>> it's validity.
>>>>>>
>>>>>> Presently, br_fdb_update() will not update local fdb entries.   Instead
>>>>>> it will log a misleading warning...  It will only let you update
>>>>>> non-local entries.  This is fine for user-created entries, but any
>>>>>> operation on dynamically created entries will only persist until
>>>>>> the next packet.  It also races against the packet, so there is
>>>>>> absolutely no guarantee that the values of fdb->dst and fdb->updated
>>>>>> will be consistent..
>>>>>>
>>>>>> It seems to me that the update capability of NTF_USE would actually be
>>>>>> of more value on local or user-created fdb entries.
>>>>>>
>>>>>> The fdb creation capability of NTF_USE should be disabled.
>>>>>>
>>>>>> Thoughts?
>>>>>
>>>>> I ignored NTF_USE in this patch because I regard it as emulating kernel
>>>>> creating entries after investigating git log.
>>>>>
>>>>> http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=0c5c2d3089068d4aa378f7a40d2b5ad9d4f52ce8
>>>>> http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=292d1398983f3514a0eab13b7606df7f4730b498
>>>>>
>>>>> So I think NTF_USE shouldn't set added_by_user.
>>>>> And to emulate kernel creating entries, simply calling br_fdb_update()
>>>>> is the right way, isn't it?
>>>>
>>>> You can create dynamic entries (emulating the kernel) without NTF_USE.
>>>> Just set the NUD_REACHABLE.  Notice that arp cache only uses NTF_USE
>>>> to trigger and arp notification.  The creation is still triggered via
>>>> other netlink flags.
>>>>
>>>> The more I look at this the more I think NTF_USE should not create
>>>> an entry all by itself.
>>>
>>> I haven't fully understood you yet.
>>> Currently NTF_USE behaves as if the port receives a frame and it seems
>>> to work, though the ability to create entries is different from neigh
>>> subsystem.
>>> Why do you want to change the behavior?
>>> Are you worried about inconsistency of NLM-flags/NUD-state with NTF_USE
>>> between neigh and bridge?
>>
>> No, it is inconsistent with other NLM/NUD-state within bridge.  As
>> an fdb creation flag NTF_USE is confusing.  It will create an entry
>> without NLM_F_CREATE being set.  It will ignore NLM_F_EXCL flag as
>> well.  It will additionally ignore any NUD-state flags that may be set
>> in the netlink message.  So it may not be doing what the user wishes.
> 
> I don't know which of NTF-flags and NLM-flags/NUD-state should be given
> high priority on setting. For now, in bridge, NTF_USE masks any other
> flags. If this is not proper way for netlink/neighbour, I will agree
> with you. If not sure, I have no motivation to change existing behavior
> that might be expected by some users.
> 
>>
>> It also provides duplicate functionality.  The same results are achieved
>> by setting NLM_F_CREATE flag and NUD_REACHABLE state in the message.
> 
> br_fdb_update() never updates fdb->used, which is visible by user,
> unlike fdb_add_entry().

Thanks for pointing this out.  It looks like there are some
inconsistencies in the fdb->used markings as well.

> 
> If it is duplicate functionality, isn't NTF_USE itself no use?
> What can be achieved by changing capability of creation and update of
> local entries?

Update of local entries gives you port redirection, but doing it under
rcu.  Not sure if it really makes much sense though...

NTF_USE makes since for neighbor cache as it triggers an ARP and a
refresh of the entry.  Suppose, NTF_USE on the fdb entry should
trigger a refresh as well, but causing a create has to be explicit.

I think I'll just send my patch and we can continue this discussion
there.

-vlad
> 
> Thanks,
> Toshiaki Makita
>