Skip to contentSkip to conversation

x86/mm: Refuse W^X violations

PATCH Superseded Peter Zijlstra 4 participants 5 messages Aug 25, 2022

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Peter Zijlstra

author maintainer
#0 lore raw

PATCH 1/1 x86/mm: Refuse W^X violations

x86 has STRICT_*_RWX, but not even a warning when someone violates it.

Add this warning and fully refuse the transition.

---

arch/x86/mm/pat/set_memory.c +29 -0
index 1abd5438f126..9e9bef3f36b3 100644
@@ -579,6 +579,30 @@ static inline pgprot_t static_protections(pgprot_t prot, unsigned long start,
 	return __pgprot(pgprot_val(prot) & ~forbidden);
 }
 
+/*
+ * Validate and enforce strict W^X semantics.
+ */
+static inline pgprot_t verify_rwx(pgprot_t old, pgprot_t new, unsigned long start,
+				  unsigned long pfn, unsigned long npg)
+{
+	unsigned long end;
+
+	if (!((pgprot_val(old) ^ pgprot_val(new)) & (_PAGE_RW | _PAGE_NX)))
+		return new;
+
+	if ((pgprot_val(new) & (_PAGE_RW | _PAGE_NX)) != _PAGE_RW)
+		return new;
+
+	end = start + npg * PAGE_SIZE - 1;
+	WARN(1, "CPA refuse W^X violation: %016llx -> %016llx range: 0x%016lx - 0x%016lx PFN %lx\n",
+	     (unsigned long long)pgprot_val(old),
+	     (unsigned long long)pgprot_val(new),
+	     start, end, pfn);
+
+	/* refuse the transition into WX */
+	return old;
+}
+
 /*
  * Lookup the page table entry for a virtual address in a specific pgd.
  * Return a pointer to the entry and the level of the mapping.
@@ -885,6 +909,8 @@ static int __should_split_large_page(pte_t *kpte, unsigned long address,
 	new_prot = static_protections(req_prot, lpaddr, old_pfn, numpages,
 				      psize, CPA_DETECT);
 
+	new_prot = verify_rwx(old_prot, new_prot, lpaddr, old_pfn, numpages);
+
 	/*
 	 * If there is a conflict, split the large page.
 	 *
@@ -1525,6 +1551,7 @@ static int __change_page_attr(struct cpa_data *cpa, int primary)
 
 	if (level == PG_LEVEL_4K) {
 		pte_t new_pte;
+		pgprot_t old_prot = pte_pgprot(old_pte);
 		pgprot_t new_prot = pte_pgprot(old_pte);
 		unsigned long pfn = pte_pfn(old_pte);
 
@@ -1536,6 +1563,8 @@ static int __change_page_attr(struct cpa_data *cpa, int primary)
 		new_prot = static_protections(new_prot, address, pfn, 1, 0,
 					      CPA_PROTECT);
 
+		new_prot = verify_rwx(old_prot, new_prot, address, pfn, 1);
+
 		new_prot = pgprot_clear_protnone_bits(new_prot);
 
 		/*
#1 lore raw
Quoted 13 lines · Peter Zijlstra

x86 has STRICT_*_RWX, but not even a warning when someone violates it.

Add this warning and fully refuse the transition.

---

arch/x86/mm/pat/set_memory.c +0 -0
index 1abd5438f126..9e9bef3f36b3 100644
@@ -579,6 +579,30 @@ static inline pgprot_t static_protections(pgprot_t prot, unsigned long start,
 	return __pgprot(pgprot_val(prot) & ~forbidden);

arch/x86/mm/pat/set_memory.c static inline pgprot_t static_protections(pgprot_t prot, unsigned long start, · line 589

Quoted 10 lines
arch/x86/mm/pat/set_memory.c +8 -0
 }
 
+/*
+ * Validate and enforce strict W^X semantics.
+ */
+static inline pgprot_t verify_rwx(pgprot_t old, pgprot_t new, unsigned long start,
+				  unsigned long pfn, unsigned long npg)
+{
+	unsigned long end;
+

I think this needs

	if (!(__supported_pte_mask & _PAGE_NX))
		return new;

to play nice with non-PAE 32-bit kernels.

arch/x86/mm/pat/set_memory.c static inline pgprot_t static_protections(pgprot_t prot, unsigned long start, · line 597

Quoted 8 lines
arch/x86/mm/pat/set_memory.c +8 -0
+	if (!((pgprot_val(old) ^ pgprot_val(new)) & (_PAGE_RW | _PAGE_NX)))
+		return new;
+
+	if ((pgprot_val(new) & (_PAGE_RW | _PAGE_NX)) != _PAGE_RW)
+		return new;
+
+	end = start + npg * PAGE_SIZE - 1;
+	WARN(1, "CPA refuse W^X violation: %016llx -> %016llx range: 0x%016lx - 0x%016lx PFN %lx\n",

WARN_ONCE() to avoid eternal spam if something does go sideways?

Quoted 7 lines
arch/x86/mm/pat/set_memory.c +7 -0
+	     (unsigned long long)pgprot_val(old),
+	     (unsigned long long)pgprot_val(new),
+	     start, end, pfn);
+
+	/* refuse the transition into WX */
+	return old;
+}
#2 lore raw
Quoted 10 lines · Sean Christopherson

arch/x86/mm/pat/set_memory.c static inline pgprot_t static_protections(pgprot_t prot, unsigned long start, · line 589

Quoted 8 lines
arch/x86/mm/pat/set_memory.c +8 -0
+/*
+ * Validate and enforce strict W^X semantics.
+ */
+static inline pgprot_t verify_rwx(pgprot_t old, pgprot_t new, unsigned long start,
+				  unsigned long pfn, unsigned long npg)
+{
+	unsigned long end;
+

I think this needs

Quoted 4 lines
	if (!(__supported_pte_mask & _PAGE_NX))
		return new;

to play nice with non-PAE 32-bit kernels.

Good catch. Nit: I'd probably write this up as:

	if (!cpu_feature_enabled(X86_FEATURE_NX))
		return new;

That gets us our fancy static branches and is a bit easier on the eyes.
I checked and don't see a way for __supported_pte_mask to have _PAGE_NX
clear when X86_FEATURE_NX==1.

author maintainer
#4 lore raw
Quoted 5 lines · Kees Cook
Quoted 1 line · Peter Zijlstra

x86 has STRICT_*_RWX, but not even a warning when someone violates it.

Yes please. I assume this is only kernel pages? Doing this globally is nice too, but runs into annoying problems[1].

Yeah, this interface should only be used on kernel pages.