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

[Huang, Kai]

On Mon, 2023-10-09 at 20:42 -0500, Haitao Huang wrote:
> Hi Sean
> 
> On Mon, 09 Oct 2023 19:23:04 -0500, Sean Christopherson  
> <seanjc@xxxxxxxxxx> wrote:
> 
> > On Mon, Oct 09, 2023, Kai Huang wrote:
> > > On Fri, 2023-09-22 at 20:06 -0700, Haitao Huang wrote:
> > > > +/**
> > > > + * sgx_epc_oom() - invoke EPC out-of-memory handling on target LRU
> > > > + * @lru:	LRU that is low
> > > > + *
> > > > + * Return:	%true if a victim was found and kicked.
> > > > + */
> > > > +bool sgx_epc_oom(struct sgx_epc_lru_lists *lru)
> > > > +{
> > > > +	struct sgx_epc_page *victim;
> > > > +
> > > > +	spin_lock(&lru->lock);
> > > > +	victim = sgx_oom_get_victim(lru);
> > > > +	spin_unlock(&lru->lock);
> > > > +
> > > > +	if (!victim)
> > > > +		return false;
> > > > +
> > > > +	if (victim->flags & SGX_EPC_OWNER_PAGE)
> > > > +		return sgx_oom_encl_page(victim->encl_page);
> > > > +
> > > > +	if (victim->flags & SGX_EPC_OWNER_ENCL)
> > > > +		return sgx_oom_encl(victim->encl);
> > > 
> > > I hate to bring this up, at least at this stage, but I am wondering why  
> > > we need
> > > to put VA and SECS pages to the unreclaimable list, but cannot keep an
> > > "enclave_list" instead?
> > 
> > The motivation for tracking EPC pages instead of enclaves was so that  
> > the EPC
> > OOM-killer could "kill" VMs as well as host-owned enclaves.  The virtual  
> > EPC code
> > didn't actually kill the VM process, it instead just freed all of the  
> > EPC pages
> > and abused the SGX architecture to effectively make the guest recreate  
> > all its
> > enclaves (IIRC, QEMU does the same thing to "support" live migration).
> > 
> > Looks like y'all punted on that with:
> > 
> >   The EPC pages allocated for KVM guests by the virtual EPC driver are  
> > not
> >   reclaimable by the host kernel [5]. Therefore they are not tracked by  
> > any
> >   LRU lists for reclaiming purposes in this implementation, but they are
> >   charged toward the cgroup of the user processs (e.g., QEMU) launching  
> > the
> >   guest.  And when the cgroup EPC usage reaches its limit, the virtual  
> > EPC
> >   driver will stop allocating more EPC for the VM, and return SIGBUS to  
> > the
> >   user process which would abort the VM launch.
> > 
> > which IMO is a hack, unless returning SIGBUS is actually enforced  
> > somehow.  Relying
> > on userspace to be kind enough to kill its VMs kinda defeats the purpose  
> > of cgroup
> > enforcement.  E.g. if the hard limit for a EPC cgroup is lowered,  
> > userspace running
> > encalves in a VM could continue on and refuse to give up its EPC, and  
> > thus run above
> > its limit in perpetuity.
> > 
> Cgroup would refuse to allocate more when limit is reached so VMs can not  
> run above limit.
> 
> IIRC VMs only support static EPC size right now, reaching limit at launch  
> means the EPC size given in command line for QEMU is not appropriate. So  
> VM should not launch, hence the current behavior.
> 
> [all EPC pages in guest are allocated on page fault caused by the  
> sensitization process in guest kernel during init, which is part of the VM  
> Launch process. So SIGNBUS will turn into failed VM launch.]
> 
> Once it is launched, guest kernel would have 'total capacity' given by the  
> static value from QEMU option. And it would start paging when it is used  
> up, never would ask for more from host.
> 
> For future with dynamic EPC for running guests, QEMU could handle  
> allocation failure and pass SIGBUS to the running guest kernel.  Is that  
> correct understanding?
> 
> 
> > I can see userspace wanting to explicitly terminate the VM instead of  
> > "silently"
> > the VM's enclaves, but that seems like it should be a knob in the  
> > virtual EPC
> > code.
> 
> If my understanding above is correct and understanding your statement  
> above correctly, then don't see we really need separate knob for vEPC  
> code. Reaching a cgroup limit by a running guest (assuming dynamic  
> allocation implemented) should not translate automatically killing the VM.  
> Instead, it's user space job to work with guest to handle allocation  
> failure. Guest could page and kill enclaves.
> 

IIUC Sean was talking about changing misc.max _after_ you launch SGX VMs:

1) misc.max = 100M
2) Launch VMs with total virtual EPC size = 100M	<- success
3) misc.max = 50M

3) will also succeed, but nothing will happen, the VMs will be still holding
100M EPC.

You need to somehow track virtual EPC and kill VM instead.

(or somehow fail to do 3) if it is also an acceptable option.)