x86/mm: Ensure forced page table splitting

There are a few kernel users like kfence that require 4k pages to work
correctly and do not support large mappings.  They use set_memory_4k()
to break down those large mappings.

That, in turn relies on cpa_data->force_split option to indicate to
set_memory code that it should split page tables regardless of whether
the need to be.

But, a recent change added an optimization which would return early
if a set_memory request came in that did not change permissions.  It
did not consult ->force_split and would mistakenly optimize away the
splitting that set_memory_4k() needs.  This broke kfence.

Skip the same-permission optimization when ->force_split is set.

Fixes: 127960a ("x86/mm: Inhibit _PAGE_NX changes from cpa_process_alias()")
Signed-off-by: Dave Hansen <[email protected]>
Tested-by: Marco Elver <[email protected]>
Cc: Peter Zijlstra (Intel) <[email protected]>
Link: https://lore.kernel.org/all/CA+G9fYuFxZTxkeS35VTZMXwQvohu73W3xbZ5NtjebsVvH6hCuA@mail.gmail.com/

arch/x86/mm/pat/set_memory.c

@@ -1727,7 +1727,8 @@ static int __change_page_attr_set_clr(struct cpa_data *cpa, int primary)
 	/*
 	 * No changes, easy!
 	 */
-	if (!(pgprot_val(cpa->mask_set) | pgprot_val(cpa->mask_clr)))
+	if (!(pgprot_val(cpa->mask_set) | pgprot_val(cpa->mask_clr)) &&
+	    !cpa->force_split)
 		return ret;
 
 	while (rempages) {