On 2/5/2021 10:41 AM, Yu, Yu-cheng wrote:

On 2/4/2021 12:27 PM, Dave Hansen wrote:

quoted

On 2/4/21 12:19 PM, Kees Cook wrote:

quoted

quoted

(e) A page where the processor observed a Write=1 PTE, started a 
write, set
     Dirty=1, but then observed a Write=0 PTE.  That's possible 
today, but
     will not happen on processors that support shadow stack.

What happens for "e" with/without CET? It sounds like direct writes to
such pages will be (correctly) rejected by the MMU?

A page fault would be generated regardless of CET support.

If CET were not around, the fault would be reported as a present, write
fault.

If this happened and CET were around (which shouldn't happen in
practice, it means we have a hardware issue) a page fault exception is
generated. 

Thanks for the clarification.  With or without CET, direct write to 
Write=0, Dirty=1 PTE triggers page fault.

quoted

Yu-cheng, I'm not sure there's enough debugging around to
tell us if this happens.  Would we even notice?

That potential hardware issue is, on a CET-capable system, a processor 
writes to a Write=1, Dirty=0 page, and then observes the PTE is Write=0, 
Dirty=1.  Let me think about it...

One way to detect the potential issue is adding a check when a 
non-shadow stack page's PTE goes from RW=0 to RW=1, like the following...

diff --git a/arch/x86/include/asm/pgtable.h b/arch/x86/include/asm/pgtable.h
index 29aa6f07e3c9..241b94a0fa77 100644
--- a/arch/x86/include/asm/pgtable.h
+++ b/arch/x86/include/asm/pgtable.h

@@ -405,6 +405,8 @@ static inline pte_t pte_mkyoung(pte_t pte)
  static inline pte_t pte_mkwrite(pte_t pte)
  {
  	if (cpu_feature_enabled(X86_FEATURE_SHSTK)) {
+		WARN_ONCE((pte_flags(pte) & (_PAGE_RW | _PAGE_DIRTY)) ==
+			  _PAGE_DIRTY, "Found transient shadow stack PTE\n");
  		if (pte_flags(pte) & _PAGE_COW) {
  			pte = pte_clear_flags(pte, _PAGE_COW);
  			pte = pte_set_flags(pte, _PAGE_DIRTY);

I run all my routine stress tests with the changes, and do not see any 
warning triggered.  If this change is desirable, we can probably add 
#ifdef CONFIG_DEBUG_VM around it and make it a separate patch.

--
Yu-cheng