[Bridge] [PATCH net] bridge: shutdown bridge device before removing it

Xin Long lucien.xin at gmail.com
Mon Apr 24 15:21:19 UTC 2017


On Mon, Apr 24, 2017 at 10:53 PM, Nikolay Aleksandrov
<nikolay at cumulusnetworks.com> wrote:
> On 24/04/17 17:41, Xin Long wrote:
>> On Mon, Apr 24, 2017 at 8:07 PM, Nikolay Aleksandrov
>> <nikolay at cumulusnetworks.com> wrote:
>>> On 24/04/17 14:01, Nikolay Aleksandrov wrote:
>>>> On 24/04/17 10:25, Xin Long wrote:
>>>>> During removing a bridge device, if the bridge is still up, a new mdb entry
>>>>> still can be added in br_multicast_add_group() after all mdb entries are
>>>>> removed in br_multicast_dev_del(). Like the path:
>>>>>
>>>>>   mld_ifc_timer_expire ->
>>>>>     mld_sendpack -> ...
>>>>>       br_multicast_rcv ->
>>>>>         br_multicast_add_group
>>>>>
>>>>> The new mp's timer will be set up. If the timer expires after the bridge
>>>>> is freed, it may cause use-after-free panic in br_multicast_group_expired.
>>>>> This can happen when ip link remove a bridge or destroy a netns with a
>>>>> bridge device inside.
>>>>>
>>>>> As we can see in br_del_bridge, brctl is also supposed to remove a bridge
>>>>> device after it's shutdown.
>>>>>
>>>>> This patch is to call dev_close at the beginning of br_dev_delete so that
>>>>> netif_running check in br_multicast_add_group can avoid this issue. But
>>>>> to keep consistent with before, it will not remove the IFF_UP check in
>>>>> br_del_bridge for brctl.
>>>>>
>>>>> Reported-by: Jianwen Ji <jiji at redhat.com>
>>>>> Signed-off-by: Xin Long <lucien.xin at gmail.com>
>>>>> ---
>>>>>  net/bridge/br_if.c | 2 ++
>>>>>  1 file changed, 2 insertions(+)
>>>>>
>>>>
>>>> +CC bridge maintainers
>>>>
>>>> I can see how this could happen, could you also provide the traceback ?
>>>>
>>>> The patch looks good to me, actually I think it fixes another issue with
>>>> mcast stats where the percpu pointer can be accessed after it's freed if
>>>> an mcast packet can get sent via br->dev after the br_multicast_dev_del() call.
>>>> This is definitely stable material, if I'm not mistaken the issue is there since
>>>> the introduction of br_dev_delete:
>>>> commit e10177abf842
>>>> Author: Satish Ashok <sashok at cumulusnetworks.com>
>>>> Date:   Wed Jul 15 07:16:51 2015 -0700
>>>>
>>>>     bridge: multicast: fix handling of temp and perm entries
>>>>
>>>>
>>>>
>>>> Acked-by: Nikolay Aleksandrov <nikolay at cumulusnetworks.com>
>>>>
>>>
>>> Actually I have a better idea for a fix because dev_close() for a single device is rather heavy.
>>> Why don't you move the mdb flush logic in the bridge's ndo_uninit() callback ?
>>> That should have the same effect and be much faster.
>> Yes. But it seems that all cleanups for bridge should be done after
>> it's shutdown since beginning according to brctl. I'm not sure if there
>> are still other problems caused by this. maybe safer to use dev_close.
>> I need to check more to confirm this.
>>
>
> ndo_uninit() is after the device has been stopped, so it is the same as
> your fix as I said.
got that your suggestion can fix this issue. what I'm afraid of is there
are still other problems like this issue, like "the percpu pointer" one
you just mentioned above, though it's already fixed by ndo_uninit.
dev_close would just avoid ALL this kind of issues if there still are. :)

But if you can be sure no more issue like this one, I'm all for that,
will improve this patch with your suggestion.


>
>> I also have another question about mp->timer removing.
>> As we can see, now it removes this timer with del_timer, instead of
>> del_timer_sync. What if the timer is running when del_timer ?
>> How can we be sure that br_multicast_group_expired will be done
>> before the bridge dev is freed. synchronize_net ?
>>
>
> Yeah, I've been thinking about that and the only race is that the timer
> might have fired and waiting for the lock while the mdb is being flushed
> thus the cancel_timer() won't affect it and then it will enter and see
> that !netif_running(br->dev), but unfortunately there's a bug because we
> cannot guarantee that br->dev still exists at that point.
> This is a different bug though.
exactly, the bad thing is it's pretty hard to reproduce even if this bz exists,
since the timer process can not be preemptable. synchronize_net probably
could avoid it (not sure).

>
>>>
>>> By the way I just noticed that there's also a memory leak - the mdb hash is reallocated
>>> and not freed due to the mdb rehash, here's also kmemleak's object:
>>>
>> yeps, ;-)
>>
>>> unreferenced object 0xffff8800540ba800 (size 2048):
>>>   comm "softirq", pid 0, jiffies 4520588901 (age 5787.284s)
>>>   hex dump (first 32 bytes):
>>>     00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
>>>     00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
>>>   backtrace:
>>>     [<ffffffff816e2287>] kmemleak_alloc+0x67/0xc0
>>>     [<ffffffff81260bea>] __kmalloc+0x1ba/0x3e0
>>>     [<ffffffffa05c60ee>] br_mdb_rehash+0x5e/0x340 [bridge]
>>>     [<ffffffffa05c74af>] br_multicast_new_group+0x43f/0x6e0 [bridge]
>>>     [<ffffffffa05c7aa3>] br_multicast_add_group+0x203/0x260 [bridge]
>>>     [<ffffffffa05ca4b5>] br_multicast_rcv+0x945/0x11d0 [bridge]
>>>     [<ffffffffa05b6b10>] br_dev_xmit+0x180/0x470 [bridge]
>>>     [<ffffffff815c781b>] dev_hard_start_xmit+0xbb/0x3d0
>>>     [<ffffffff815c8743>] __dev_queue_xmit+0xb13/0xc10
>>>     [<ffffffff815c8850>] dev_queue_xmit+0x10/0x20
>>>     [<ffffffffa02f8d7a>] ip6_finish_output2+0x5ca/0xac0 [ipv6]
>>>     [<ffffffffa02fbfc6>] ip6_finish_output+0x126/0x2c0 [ipv6]
>>>     [<ffffffffa02fc245>] ip6_output+0xe5/0x390 [ipv6]
>>>     [<ffffffffa032b92c>] NF_HOOK.constprop.44+0x6c/0x240 [ipv6]
>>>     [<ffffffffa032bd16>] mld_sendpack+0x216/0x3e0 [ipv6]
>>>     [<ffffffffa032d5eb>] mld_ifc_timer_expire+0x18b/0x2b0 [ipv6]
>>>
>>>
>>>
>


More information about the Bridge mailing list