Re: IPv6 regression introduced by commit 3b6761d18bc11f2af2a6fc494e9026d39593f22c

[David Ahern <dsahern@gmail.com>]

On 9/27/20 11:48 PM, Baptiste Jonglez wrote:
> On 27-09-20, David Ahern wrote:
>> On 9/27/20 9:10 AM, Baptiste Jonglez wrote:
>>> On 27-09-20, Baptiste Jonglez wrote:
>>>> 1) failing IPv6 neighbours, what Alarig reported.  We are seeing this
>>>>    on a full-view BGP router with rather low amount of IPv6 traffic
>>>>    (around 10-20 Mbps)
>>>
>>> Ok, I found a quick way to reproduce this issue:
>>>
>>>     # for net in {1..9999}; do ip -6 route add 2001:db8:ffff:${net}::/64 via fe80::4242 dev lo; done
>>>
>>> and then:
>>>
>>>     # for net in {1..9999}; do ping -c1 2001:db8:ffff:${net}::1; done
>>>
>>> This quickly gets to a situation where ping fails early with:
>>>
>>>     ping: connect: Network is unreachable
>>>
>>> At this point, IPv6 connectivity is broken.  The kernel is no longer
>>> replying to IPv6 neighbor solicitation from other hosts on local
>>> networks.
>>>
>>> When this happens, the "fib_rt_alloc" field from /proc/net/rt6_stats
>>> is roughly equal to net.ipv6.route.max_size (a bit more in my tests).
>>>
>>> Interestingly, the system appears to stay in this broken state
>>> indefinitely, even without trying to send new IPv6 traffic.  The
>>> fib_rt_alloc statistics does not decrease.
>>>
>>
>> fib_rt_alloc is incremented by calls to ip6_dst_alloc. Each of your
>> 9,999 pings is to a unique address and hence causes a dst to be
>> allocated and the counter to be incremented. It is never decremented.
>> That is standard operating procedure.
> 
> Ok, then this is a change in behaviour.  Here is a graph of fib_rt_alloc
> on a busy router (IPv6 full view, moderate IPv6 traffic) with 4.9 kernel:
> 
>   https://files.polyno.me/tmp/rt6_stats_fib_rt_alloc_4.9.png
> 
> It varies quite a lot and stays around 50, so clearly it can be
> decremented in regular operation.
> 
> On 4.19 and later, it does seem to be decremented only when a route is
> removed (ip -6 route delete).  Here is the same graph on a router with a
> 4.19 kernel and a large net.ipv6.route.max_size:
> 
>    https://files.polyno.me/tmp/rt6_stats_fib_rt_alloc_4.19.png
> 
> Overall, do you mean that fib_rt_alloc is a red herring and is not a good
> marker of the issue?
> 

$ git checkout v4.9
$ egrep -r fib_rt_alloc include/ net/
include//net/ip6_fib.h:	__u32		fib_rt_alloc;		/* permanent routes	*/
net//ipv6/route.c:		   net->ipv6.rt6_stats->fib_rt_alloc,

The first declares it; the second prints it. That's it, no other users
so I have no idea why it shows any changes in your v4.9 graph.

Looking git history shows that Wei actually wired up the stats with
commit 81eb8447daae3b62247aa66bb17b82f8fef68249

Author: Wei Wang <weiwan@google.com>
Date:   Fri Oct 6 12:06:11 2017 -0700

    ipv6: take care of rt6_stats

That patch adds an inc but no dec for this stat which is what you show
in your 4.19 graph

Coming back to the bigger problem, fib_rt_alloc has *no* bearing on the
ability to create dst entries which is what the max_route_size sysctl
affects (not FIB entries which are now unbounded, but dst_entry
instances which is when a FIB entry has been hit and used in the
datapath to move packets).

Eric investigated a similar problem recently which resulted in

commit d8882935fcae28bceb5f6f56f09cded8d36d85e6
Author: Eric Dumazet <edumazet@google.com>
Date:   Fri May 8 07:34:14 2020 -0700

    ipv6: use DST_NOCOUNT in ip6_rt_pcpu_alloc()

and I believe is released in v5.8.