Re: [PATCH RFC/RFT net-next 00/17] net: Convert neighbor tables to per-namespace

[Cong Wang]

On Tue, Jul 17, 2018 at 10:43 AM David Ahern <dsahern@xxxxxxxxx> wrote:
>
> On 7/17/18 11:40 AM, Cong Wang wrote:
> > On Tue, Jul 17, 2018 at 5:11 AM <dsahern@xxxxxxxxxx> wrote:
> >>
> >> From: David Ahern <dsahern@xxxxxxxxx>
> >>
> >> Nikita Leshenko reported that neighbor entries in one namespace can
> >> evict neighbor entries in another. The problem is that the neighbor
> >> tables have entries across all namespaces without separate accounting
> >> and with global limits on when to scan for entries to evict.
> >
> > It is nothing new, people including me already noticed this before.
> >
> >
> >>
> >> Resolve by making the neighbor tables for ipv4, ipv6 and decnet per
> >> namespace and making the accounting and threshold limits per namespace.
> >
> >
> > The last discussion about this a long time ago concluded that neigh
> > table entries are controllable by remote, so after moving it to per netns,
> > it would be easier to DOS the host.
> >
>
> There are still limits on the total number of entries and with
> per-namespace limits an admin has better control.

Per-netns limit is *exactly* the problem here.

Quote from David Miller:
"
From: ebiederm@xxxxxxxxxxxx (Eric W. Biederman)
Date: Wed, 25 Jun 2014 18:17:08 -0700

> I disagree that removing a global DOS prevention check is a benefit.
> Certainly large semantics changes like that should not happen without
> being discussed in the patch description.

Agreed, this is the most important core issue.

If we just make these things per netns, then as a result if you create
N namespaces we will allow N times more neighbour entries to be
sitting in the system at once.

Actually, I'm really surprised the limits get hit and this actually
causes problems.
"

You can see the original discussion here:
https://marc.info/?l=linux-netdev&m=140356141019653&w=2