Re: [PATCH v5 12/18] x86/sgx: Add EPC OOM path to forcefully reclaim EPC

[Haitao Huang]

On Tue, 10 Oct 2023 19:31:19 -0500, Huang, Kai <kai.huang@xxxxxxxxx> wrote:

> On Tue, 2023-10-10 at 12:05 -0500, Haitao Huang wrote:
> > On Mon, 09 Oct 2023 21:12:27 -0500, Huang, Kai <kai.huang@xxxxxxxxx>  
> > wrote:
> >
> > >
> > > > > > >
> > > > > > Later the hosting process could migrated/reassigned to another
> > > > cgroup?
> > > > > > What to do when the new cgroup is OOM?
> > > > > >
> > > > >
> > > > > You addressed in the documentation, no?
> > > > >
> > > > > +Migration
> > > > > +---------
> > > > > +
> > > > > +Once an EPC page is charged to a cgroup (during allocation), it
> > > > > +remains charged to the original cgroup until the page is released
> > > > > +or reclaimed.  Migrating a process to a different cgroup doesn't
> > > > > +move the EPC charges that it incurred while in the previous  
> > cgroup
> > > > > +to its new cgroup.
> > > >
> > > > Should we kill the enclave though because some VA pages may be in  
> > the
> > > > new
> > > > group?
> > > >
> > >
> > > I guess acceptable?
> > >
> > > And any difference if you keep VA/SECS to unreclaimabe list?
> >
> > Tracking VA/SECS allows all cgroups, in which an enclave has allocation,
> > to identify the enclave following the back pointer and kill it as  
> > needed.
> >
> > > If you migrate one
> > > enclave to another cgroup, the old EPC pages stay in the old cgroup
> > > while the
> > > new one is charged to the new group IIUC.
> > >
> > > I am not cgroup expert, but by searching some old thread it appears  
> > this
> > > isn't a
> > > supported model:
> > >
> > > https://lore.kernel.org/lkml/YEyR9181Qgzt+Ps9@xxxxxxxxxxxxxxx/
> > >
> >
> > IIUC it's a different problem here. If we don't track the allocated VAs  
> > in
> > the new group, then the enclave that spans the two groups can't be  
> > killed
> > by the new group. If so, some enclave could just hide in some small  
> > group
> > and never gets killed but keeps allocating in a different group?
>
> I mean from the link above IIUC migrating enclave among different  
> cgroups simply
> isn't a supported model, thus any bad behaviour isn't a big concern in  
> terms of
> decision making.

If we leave some pages in a cgroup unkillable, we are in the same  
situation of not able to enforce a cgroup limit as that we are are in if  
we don't kill VMs for lower limits.

I think not supporting migration of pages between cgroups should not leave  
a gap for enforcement just like we don't want to have an enforcement gap  
if we let VMs to hold pages once it is launched.

Haitao