lkml.org 
[lkml]   [2019]   [Jul]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH ghak90 V6 02/10] audit: add container id
On Tue, Jul 16, 2019 at 11:37 AM Richard Guy Briggs <rgb@redhat.com> wrote:
> On 2019-07-15 17:09, Paul Moore wrote:
> > On Mon, Jul 8, 2019 at 2:12 PM Richard Guy Briggs <rgb@redhat.com> wrote:
> > > On 2019-05-30 19:26, Paul Moore wrote:
> >
> > ...
> >
> > > > I like the creativity, but I worry that at some point these
> > > > limitations are going to be raised (limits have a funny way of doing
> > > > that over time) and we will be in trouble. I say "trouble" because I
> > > > want to be able to quickly do an audit container ID comparison and
> > > > we're going to pay a penalty for these larger values (we'll need this
> > > > when we add multiple auditd support and the requisite record routing).
> > > >
> > > > Thinking about this makes me also realize we probably need to think a
> > > > bit longer about audit container ID conflicts between orchestrators.
> > > > Right now we just take the value that is given to us by the
> > > > orchestrator, but if we want to allow multiple container orchestrators
> > > > to work without some form of cooperation in userspace (I think we have
> > > > to assume the orchestrators will not talk to each other) we likely
> > > > need to have some way to block reuse of an audit container ID. We
> > > > would either need to prevent the orchestrator from explicitly setting
> > > > an audit container ID to a currently in use value, or instead generate
> > > > the audit container ID in the kernel upon an event triggered by the
> > > > orchestrator (e.g. a write to a /proc file). I suspect we should
> > > > start looking at the idr code, I think we will need to make use of it.
> > >
> > > To address this, I'd suggest that it is enforced to only allow the
> > > setting of descendants and to maintain a master list of audit container
> > > identifiers (with a hash table if necessary later) that includes the
> > > container owner.
> >
> > We're discussing the audit container ID management policy elsewhere in
> > this thread so I won't comment on that here, but I did want to say
> > that we will likely need something better than a simple list of audit
> > container IDs from the start. It's common for systems to have
> > thousands of containers now (or multiple thousands), which tells me
> > that a list is a poor choice. You mentioned a hash table, so I would
> > suggest starting with that over the list for the initial patchset.
>
> I saw that as an internal incremental improvement that did not affect
> the API, so I wanted to keep things a bit simpler (as you've requested
> in the past) to get this going, and add that enhancement later.

In general a simple approach is a good way to start when the
problem/use-case is not very well understood; in other words, don't
spend a lot of time/effort optimizing something you don't yet
understand. In this case we know that people want to deploy a *lot*
of containers on a single system so we should design the data
structures appropriately. A list is simply not a good fit here, I
believe/hope you know that too.

> I'll start working on it now. The hash table would simply point to
> lists anyways unless you can recommend a better approach.

I assume when you say "point to lists" you are talking about using
lists for the hash buckets? If so, yes that should be fine at this
point. In general if the per-bucket lists become a bottleneck we can
look at the size of the table (or make it tunable) or even use a
different approach entirely. Ultimately the data store is an
implementation detail private to the audit subsystem in the kernel so
we should be able to change it as necessary without breaking anything.

--
paul moore
www.paul-moore.com

\
 
 \ /
  Last update: 2019-07-16 18:09    [W:0.092 / U:12.772 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site