New talloc feature: memlimits

[simo]

On Sat, 2012-09-29 at 08:27 +1000, Andrew Bartlett wrote:
> On Sun, 2012-09-23 at 00:10 -0600, idra at samba.org wrote:
> > Hello list,
> > 
> > during the recent SDC Conference we had the Samba4 LDAP server hammered by
> > the Codenomicon guys. A few bugs were found where we ended up allocating huge
> > amounts of memory.
> > 
> > These bugs will need fixing, but the situation reminded me that we still have
> > little or no control on what users can do over ldap. In particular we have no
> > good way to limit resources, and it is relatively easy to DoS the LDAP server
> > by making it allocate huge amounts of memory.
> > 
> > So I had the idea of limiting memory allocation to arbitrarily settable sizes
> > based on talloc contextes.
> > 
> > Attached you can find an initial implementation of this feature with basic
> > tests.
> > 
> > By using talloc_set_memlimit() on a context we can decide the maximum amount
> > of memory that can be used by any alloction on that context or any of its
> > children. Attempting to allocate more memory than allowed results in a failed
> > allocation.
> > Stealing memory under a memlimited hierarchy does not fail even if the new
> > total use exceed the limit, but any further allocation on the context will
> > fail. This means we'll need to be careful on how we create temporary contexts
> > and then steal data.
> > 
> > Memory limits can nest and any allocation will reflect in the parents memory
> > limits as well. This allows for a context to have larger limits and then
> > have individual smaller limits for childrens down the hierarchy.
> > 
> > Well, enough said, if there are any objections on committin gthis change please
> > speak up, otherwise I will push by the end of the week.
> 
> This certainly fits well with the memory model used in the AD DC, where
> most memory is allocated on a parent context, that eventually ends up at
> the connecting socket.  It also helps that in most cases we prefer to
> use this pattern:
> 
> TALLOC_CTX *tmp_ctx = talloc_new(mem_ctx);
> 
> (do work)
> 
> talloc_steal(mem_ctx, ret)
> TALLOC_FREE(tmp_ctx)
> 
> (rather than steal from a talloc_stackframe() or from a context built on
> NULL). 
> 
> It also helps that this isn't a new idea - I remember a discussion with
> tridge about this early in the new talloc.  (This was in a context of
> discussions about if we should gracefully handle out of memory at all
> under a modern unix VM system).
> 
> The challenge of course is that finding out what code doesn't deal with
> memory limits cleanly, and working out what the runtime cost is.  It
> seems that it would make talloc_steal() a much more intensive operation
> than it currently is.
> 
> Finally, when we do a talloc_steal(), it seems the limit pointer is not
> updated on the child chunks.  How do we know these limit pointers will
> remain valid? 
> 
> I understand your rationale, but I think this needs more work and some
> very careful positive review rather than a 'push at the end of the
> week. 

Indeed I went through it with metze and I have a much improved version
in one of my trees. It passed tests last night but haven't pushed yet.

You can find it here:
https://git.samba.org/idra/samba.git/?p=idra/samba.git;a=shortlog;h=refs/heads/talloc_memlimit

This version deals with realloc too which was left out (oops :), it also
properly propagates limit contexts to children when needed.

I am quite confident we can start pushing this version, but if you find
any bug it would be awesome to be able to push an even cleaner patchset.

Simo.

-- 
Simo Sorce
Samba Team GPL Compliance Officer <simo at samba.org>
Principal Software Engineer at Red Hat, Inc. <simo at redhat.com>