[RFC] Failed neighbors attached to routes are not released

On Mon, Aug 22, 2011 at 7:44 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> Le vendredi 19 août 2011 à 21:53 +0300, Guy Yur a écrit :
>> Hi,
>>
>> The issue I am seeing is with a neighbor used as a gateway in a route,
>> when the neighbor gets nud FAILED it will not be removed from the
>> neighbor cache.
>> The reference count for the neighbor remains > 1
>> when neigh_periodic_work() is called.
>>
>> Issue noticed with IPv6 neighbors on Arch Linux with kernel 3.0.3
>> kernel config includes CONFIG_IPV6_ROUTER_PREF
>>
>> The problem affects routing when the neighbor loses connectivity and returns.
>>
>> Scenario: Using a default route and a static route through different interfaces.
>> When the neighbor gateway of the static route goes down the traffic
>> will move to the default gateway as expected.
>> Once the static route neighbor comes back up it is not asked for
>> neighbor solicitation
>> because the route is marked as FAILED and the traffic will continue to
>> pass through the default gateway.
>>
>> Steps to reproduce the route not being removed:
>> 1. add an IPv6 address on an interface
>> 2. add a route to a network through a gateway on the interface's network
>> 3. make sure the gateway address is not reachable
>> 4. ping6 a host in the route network
>> 5. "ip -6 nei" will show the gateway neighbor as FAILED and it won't be released
>>
>> Steps to reproduce the routing problem:
>> 1. client and two gateway machines (A and B)
>> 2. on the client define a static route through A and a default route through B
>> 3. disconnect eth on A
>> 4. ping6 a host in the network that should go through A
>>    after a while the traffic will move through B which is the default route
>> 5. reconnect eth on A
>> 6. ping6 a host in the network again, the traffic will still go through B
>>    "ip -6 nei" on the client will still show A as FAILED
>>
>>
>> Patch to change the nud state to NONE if the reference count > 1
>> allowing the neighbor to be released in a future pass.
>
> I wonder why a 'future pass' is needed at all.
>
> Shouldnt we immediately detect link becomes alive and force an immediate
> flush at this point, before waiting a garbage collect timer ?
>

I tested removing the neighbor as done in neigh_flush_dev() instead of
just setting to NUD_NONE,
it seems that the neighbor shouldn't be removed if it has routes attached to it.
If the neighbor is removed the route won't be considered at all.

The problem of the state remaining FAILED does need to be handled.
Like the case of no neighbor, a FAILED state means find_match() won't
try to send neighbor solicitation probes.
The neighbor will remain in a state of NUD_FAILED and the route won't
be used, only a direct communication with the neighbor will change its
state.

Do you think it is better to change the state from NUD_FAILED to
NUD_NONE in neigh_timer_handler() instead?

Updated patch to set the state to NUD_NOARP if the state was valid,
same as done in neigh_flush_dev().
Not sure if it is needed or the state can be set to NUD_NONE in all cases.

the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html