Re: [PATCH v13 10/13] x86/sgx: Add sgx_einit() for initializing enclaves

[Sean Christopherson]

On Wed, Aug 29, 2018 at 12:33:54AM -0700, Huang, Kai wrote:
> [snip..]
> 
> > > >
> > > > @@ -38,6 +39,18 @@ static LIST_HEAD(sgx_active_page_list);  static
> > > > DEFINE_SPINLOCK(sgx_active_page_list_lock);
> > > >  static struct task_struct *ksgxswapd_tsk;  static
> > > > DECLARE_WAIT_QUEUE_HEAD(ksgxswapd_waitq);
> > > > +static struct notifier_block sgx_pm_notifier; static u64
> > > > +sgx_pm_cnt;
> > > > +
> > > > +/* The cache for the last known values of IA32_SGXLEPUBKEYHASHx
> > > > +MSRs
> > > > for each
> > > > + * CPU. The entries are initialized when they are first used by
> > > > sgx_einit().
> > > > + */
> > > > +struct sgx_lepubkeyhash {
> > > > +	u64 msrs[4];
> > > > +	u64 pm_cnt;
> > >
> > > May I ask why do we need pm_cnt here? In fact why do we need suspend
> > > staff (namely, sgx_pm_cnt above, and related code in this patch) here
> > > in this patch? From the patch commit message I don't see why we need
> > > PM staff here. Please give comment why you need PM staff, or you may
> > > consider to split the PM staff to another patch.
> > 
> > Refining the commit message probably makes more sense because without PM
> > code sgx_einit() would be broken. The MSRs have been reset after waking up.
> > 
> > Some kind of counter is required to keep track of the power cycle. When going
> > to sleep the sgx_pm_cnt is increased. sgx_einit() compares the current value of
> > the global count to the value in the cache entry to see whether we are in a new
> > power cycle.
> 
> You mean reset to Intel default? I think we can also just reset the cached MSR
> values on each power cycle, which would be simpler, IMHO?

Refresh my brain, does hardware reset the MSRs on a transition to S3 or lower?

> I think we definitely need some code to handle S3-S5, but should be in
> separate patches, since I think the major impact of S3-S5 is entire EPC
> being destroyed. I think keeping pm_cnt is not sufficient enough to handle
> such case?
> > 
> > This brings up one question though: how do we deal with VM host going to sleep?
> > VM guest would not be aware of this.
> 
> IMO VM just gets "sudden loss of EPC" after suspend & resume in host. SGX driver
> and SDK should be able to handle "sudden loss of EPC", ie, co-working together to
> re-establish the missing enclaves.
> 
> Actually supporting "sudden loss of EPC" is a requirement to support live migration
> of VM w/ SGX. Internally long time ago we had a discussion and the decision was we
> should support SGX live migration given two facts:
> 
> 1) losing platform-dependent is not important. For example, losing sealing key is
> not a problem, as we could get secrets provisioned again from remote. 2) Both windows
> & linux driver commit to support "sudden loss of EPC".
> 
> I don't think we have to support in very first upstream driver, but I think we need
> to support someday.

Actually, we can easily support this in the driver, at least for SGX1
hardware.  SGX2 isn't difficult to handle, but we've intentionally
postponed those patches until SGX1 support is in mainline[1].
Accesses to the EPC after it is lost will cause faults.  Userspace
EPC accesses, e.g. ERESUME, will get a SIGSEGV that the process
should interpret as an "I should restart my enclave" event.  The
SDK already does this.  In the driver, we just need to be aware of
this potential behavior and not freak out.  Specifically, SGX_INVD
needs to not WARN on faults that may have been due to a the EPC
being nuked.  I think we can even remove the sgx_encl_pm_notifier()
code altogether.

[1] SGX1 hardware signals a #GP on an access to an invalid EPC page.
    SGX2 signals a #PF with the PF_SGX error code bit set.  This is
    problematic because the kernel looks at the PTEs for CR2 and sees
    nothing wrong, so it thinks it should just restart the
    instruction, leading to an infinite fault loop.  Resolving this
    is fairly straightforward, but a complete fix requires propagating
    PF_SGX down to the ENCLS fixup handler, which means plumbing the
    error code through the fixup handlers or smushing PF_SGX into
    trapnr.  Since there is a parallel effort to plumb the error code
    through the handlers, https://lkml.org/lkml/2018/8/6/924, we opted
    to do this in a separate series.

> Sean, 
> 
> Would you be able to comment here?
> 
> > 
> > I think the best measure would be to add a new parameter to sgx_einit() that
> > enforces update of the MSRs. The driver can then set this parameter in the case
> > when sgx_einit() returns SGX_INVALID_LICENSE. This is coherent because the
> > driver requires writable MSRs. It would not be coherent to do it directly in the
> > core because KVM does not require writable MSRs.
> 
> IMHO this is not required, as I mentioned above.
> 
> And 
> [snip...]
> 
> Thanks,
> -Kai