x86/speculation/l1tf: Exempt zeroed PTEs from inversion

It turns out that we should *not* invert all not-present mappings,
because the all zeroes case is obviously special.

clear_page() does not undergo the XOR logic to invert the address bits,
i.e. PTE, PMD and PUD entries that have not been individually written
will have val=0 and so will trigger __pte_needs_invert(). As a result,
{pte,pmd,pud}_pfn() will return the wrong PFN value, i.e. all ones
(adjusted by the max PFN mask) instead of zero. A zeroed entry is ok
because the page at physical address 0 is reserved early in boot
specifically to mitigate L1TF, so explicitly exempt them from the
inversion when reading the PFN.

Manifested as an unexpected mprotect(..., PROT_NONE) failure when called
on a VMA that has VM_PFNMAP and was mmap'd to as something other than
PROT_NONE but never used. mprotect() sends the PROT_NONE request down
prot_none_walk(), which walks the PTEs to check the PFNs.
prot_none_pte_entry() gets the bogus PFN from pte_pfn() and returns
-EACCES because it thinks mprotect() is trying to adjust a high MMIO
address.

[ This is a very modified version of Sean's original patch, but all
  credit goes to Sean for doing this and also pointing out that
  sometimes the __pte_needs_invert() function only gets the protection
  bits, not the full eventual pte.  But zero remains special even in
  just protection bits, so that's ok.   - Linus ]

Fixes: f22cc87 ("x86/speculation/l1tf: Invert all not present mappings")
Signed-off-by: Sean Christopherson <[email protected]>
Acked-by: Andi Kleen <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Cc: Josh Poimboeuf <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Vlastimil Babka <[email protected]>
Cc: Dave Hansen <[email protected]>
Cc: Greg Kroah-Hartman <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>

arch/x86/include/asm/pgtable-invert.h

@@ -4,9 +4,18 @@
 
 #ifndef __ASSEMBLY__
 
+/*
+ * A clear pte value is special, and doesn't get inverted.
+ *
+ * Note that even users that only pass a pgprot_t (rather
+ * than a full pte) won't trigger the special zero case,
+ * because even PAGE_NONE has _PAGE_PROTNONE | _PAGE_ACCESSED
+ * set. So the all zero case really is limited to just the
+ * cleared page table entry case.
+ */
 static inline bool __pte_needs_invert(u64 val)
 {
-	return !(val & _PAGE_PRESENT);
+	return val && !(val & _PAGE_PRESENT);
 }
 
 /* Get a mask to xor with the page table entry to get the correct pfn. */