Re: [PATCH v2 2/2] x86/sgx: Implement EUPDATESVN and opportunistically call it during first EPC page alloc

[Jarkko Sakkinen]

On Tue, Apr 08, 2025 at 09:40:14AM +0300, Jarkko Sakkinen wrote:
> On Tue, Apr 08, 2025 at 12:06:32AM +0000, Huang, Kai wrote:
> > On Mon, 2025-04-07 at 08:23 +0000, Reshetova, Elena wrote:
> > > > On Fri, Apr 04, 2025 at 06:53:17AM +0000, Reshetova, Elena wrote:
> > > > > > On Wed, Apr 02, 2025 at 01:11:25PM +0000, Reshetova, Elena wrote:
> > > > > > > > > current SGX kernel code does not handle such errors in any other
> > > > way
> > > > > > > > > than notifying that operation failed for other ENCLS leaves. So, I don't
> > > > > > > > > see why ENCLS[EUPDATESVN] should be different from existing
> > > > > > behaviour?
> > > > > > > > 
> > > > > > > > While not disagreeing fully (it depends on call site), in some
> > > > > > > > situations it is more difficult to take more preventive actions.
> > > > > > > > 
> > > > > > > > This is a situation where we know that there are *zero* EPC pages in
> > > > > > > > traffic so it is relatively easy to stop the madness, isn't it?
> > > > > > > > 
> > > > > > > > I guess the best action would be make sgx_alloc_epc_page() return
> > > > > > > > consistently -ENOMEM, if the unexpected happens.
> > > > > > > 
> > > > > > > But this would be very misleading imo. We do have memory, even page
> > > > > > > allocation might function as normal in EPC, the only thing that is broken
> > > > > > > can be EUPDATESVN functionality. Returning -ENOMEM in this case
> > > > seems
> > > > > > > wrong.
> > > > > > 
> > > > > > This makes it not misleading at all:
> > > > > > 
> > > > > > 	pr_err("EUPDATESVN: unknown error %d\n", ret);
> > > > > > 
> > > > > > Since hardware should never return this, it indicates a kernel bug.
> > > > > 
> > > > > OK, so you propose in this case to print the above message, sgx_updatesvn
> > > > > returning an error, and then NULL from __sgx_alloc_epc_page_from_node
> > > > and
> > > > > the __sgx_alloc_epc_page returning -ENOMEM after an iteration over
> > > > > a whole set of numa nodes given that we will keep getting the unknown
> > > > error
> > > > > on each node upon trying to do an allocation from each one?
> > > > 
> > > > I'd disable ioctl's in this case and return -ENOMEM. It's a cheap sanity
> > > > check. Should not ever happen, but if e.g., a new kernel patch breaks
> > > > anything, it could help catching issues.
> > > > 
> > > > We are talking here about situation that is never expected to happen so I
> > > > don't think it is too heavy hammer here. Here it makes sense because not
> > > > much effort is required to implement the counter-measures.
> > > 
> > > OK, but does it really make sense to explicitly disable ioctls? 
> > > Note that everything *in practice* will be disabled simply because not a single page
> > > anymore can be allocated from EPC since we are getting -ENOMEM on EPC
> > > page allocation. Also, note that any approach we chose should be symmetrical
> > > to SGX virtualization side also, which doesn´t use ioctls at all. Simply returning
> > > -ENOMEM for page allocation in EPC seems like a correct symmetrical solution
> > > that would work for both nativel enclaves and EPC pages allocated for VMs.
> > > And nothing would  be able to proceed creating/managing enclaves at this point. 
> > > 
> > 
> > Right, failing ioctls() doesn't cover SGX virtualization.  If we ever want to
> > fail, we should fail the EPC allocation.
> 
> "I guess the best action would be make sgx_alloc_epc_page() return
>  consistently -ENOMEM, if the unexpected happens." -me
> 
> > 
> > Btw, for the unknown error, and any other errors which should not happen,
> > couldn't we use the ENCLS_WARN()?  AFAICT there are already cases that we are
> > using ENCLS_WARN() for those "impossible-to-happen-errors".

Sorry forgot to response this. I don't have anything against this but at
minimum disabling allocation should be combined with it (in case kernel
command-line does not have oops_on_warn or whatever the option was
called).

> > 
> > E.g., in __sgx_encl_extend():
> > 
> > 	        ret = __eextend(sgx_get_epc_virt_addr(encl->secs.epc_page),
> >                                 sgx_get_epc_virt_addr(epc_page) + offset);
> >                 if (ret) {
> >                         if (encls_failed(ret))
> >                                 ENCLS_WARN(ret, "EEXTEND");
> >    
> >                         return -EIO;
> >                 }
> 
> BR, Jarkko
> 

BR, Jarkko