x86/efi: Fix kernel panic when CONFIG_DEBUG_VIRTUAL is enabled

When CONFIG_DEBUG_VIRTUAL is enabled, all accesses to __pa(address) are
monitored to see whether address falls in direct mapping or kernel text
mapping (see Documentation/x86/x86_64/mm.txt for details), if it does
not, the kernel panics. During 1:1 mapping of EFI runtime services we access
virtual addresses which are == physical addresses, thus the 1:1 mapping
and these addresses do not fall in either of the above two regions and
hence when passed as arguments to __pa() kernel panics as reported by
Dave Hansen here https://lkml.kernel.org/r/[email protected].

So, before calling __pa() virtual addresses should be validated which
results in skipping call to split_page_count() and that should be fine
because it is used to keep track of everything *but* 1:1 mappings.

Signed-off-by: Sai Praneeth Prakhya <[email protected]>
Reported-by: Dave Hansen <[email protected]>
Reviewed-by: Borislav Petkov <[email protected]>
Cc: Ricardo Neri <[email protected]>
Cc: Glenn P Williamson <[email protected]>
Cc: Ravi Shankar <[email protected]>
Signed-off-by: Matt Fleming <[email protected]>

1 files changed, 6 insertions, 3 deletions

diff --git a/arch/x86/mm/pageattr.c b/arch/x86/mm/pageattr.c
index 727158cb3b3c..9abe0c9b1098 100644
--- a/arch/x86/mm/pageattr.c
+++ b/arch/x86/mm/pageattr.c
@@ -648,9 +648,12 @@ __split_large_page(struct cpa_data *cpa, pte_t *kpte, unsigned long address,
 	for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
 		set_pte(&pbase[i], pfn_pte(pfn, canon_pgprot(ref_prot)));
 
-	if (pfn_range_is_mapped(PFN_DOWN(__pa(address)),
-				PFN_DOWN(__pa(address)) + 1))
-		split_page_count(level);
+	if (virt_addr_valid(address)) {
+		unsigned long pfn = PFN_DOWN(__pa(address));
+
+		if (pfn_range_is_mapped(pfn, pfn + 1))
+			split_page_count(level);
+	}
 
 	/*
 	 * Install the new, split up pagetable.