Re: [PATCH 14/19] mm: Introduce a cgroup for pinned memory

[Michal Hocko <mhocko@suse.com>]

On Wed 15-02-23 15:07:05, Jason Gunthorpe wrote:
> On Wed, Feb 15, 2023 at 08:00:22PM +0100, Michal Hocko wrote:
> > On Mon 06-02-23 14:32:37, Tejun Heo wrote:
> > > Hello,
> > > 
> > > On Mon, Feb 06, 2023 at 07:40:55PM -0400, Jason Gunthorpe wrote:
> > > > (a) kind of destroys the point of this as a sandboxing tool
> > > > 
> > > > It is not so harmful to use memory that someone else has been charged
> > > > with allocating.
> > > > 
> > > > But it is harmful to pin memory if someone else is charged for the
> > > > pin. It means it is unpredictable how much memory a sandbox can
> > > > actually lock down.
> > > > 
> > > > Plus we have the double accounting problem, if 1000 processes in
> > > > different cgroups open the tmpfs and all pin the memory then cgroup A
> > > > will be charged 1000x for the memory and hit its limit, possibly
> > > > creating a DOS from less priv to more priv
> > > 
> > > Let's hear what memcg people think about it. I'm not a fan of disassociating
> > > the ownership and locker of the same page but it is true that actively
> > > increasing locked consumption on a remote cgroup is awkward too.
> > 
> > One thing that is not really clear to me is whether those pins do
> > actually have any "ownership".
> 
> In most cases the ownship traces back to a file descriptor. When the
> file is closed the pin goes away.

This assumes a specific use of {un}pin_user_page*, right? IIUC the
cgroup charging is meant to be used from vm_account but that doesn't
really tell anything about the lifetime nor the ownership. Maybe this is
just a matter of documentation update...

> > The interface itself doesn't talk about
> > anything like that and so it seems perfectly fine to unpin from a
> > completely different context then pinning. 
> 
> Yes, concievably the close of the FD can be in a totally different
> process with a different cgroup.

Wouldn't you get an unbalanced charges then? How can admin recover that
situation?

> > If there is no enforcement then Tejun is right and relying on memcg
> > ownership is likely the only reliable way to use for tracking. The
> > downside is sharing obviously but this is the same problem we
> > already do deal with with shared pages.
> 
> I think this does not work well because the owner in a memcg sense is
> unrelated to the file descriptor which is the true owner.
> 
> So we can get cases where the pin is charged to the wrong cgroup which
> is effectively fatal for sandboxing, IMHO.

OK, I see. This makes it really much more complicated then.
 
> > Another thing that is not really clear to me is how the limit is
> > actually going to be used in practice. As there is no concept of a
> > reclaim for pins then I can imagine that it would be quite easy to
> > reach the hard limit and essentially DoS any further use of pins. 
> 
> Yes, that is the purpose. It is to sandbox pin users to put some limit
> on the effect they have on the full machine.
> 
> It replaces the rlimit mess that was doing the same thing.

arguably rlimit has a concept of the owner at least AFAICS. I do realize
this is not really great wrt a high level resource control though.

> > Cross cgroup pinning would make it even worse because it could
> > become a DoS vector very easily. Practically speaking what tends to
> > be a corner case in the memcg limit world would be norm for pin
> > based limit.
> 
> This is why the cgroup charged for the pin must be tightly linked to
> some cgroup that is obviously connected to the creator of the FD
> owning the pin.

The problem I can see is that the fd is just too fluid for tracking. You
can pass fd over to a different cgroup context and then all the tracking
just loses any trail to an owner.

I can see how the underlying memcg tracking information is not really
feasible for your usecases but I am really worried that it is just too
easy to misaccount without any other proper ownership tracking.
-- 
Michal Hocko
SUSE Labs