Fix some incompatibilites between KASAN and FORTIFY_SOURCE

3 KASAN self-tests fail on a kernel with both KASAN and FORTIFY_SOURCE:
memchr, memcmp and strlen. I have observed this on x86 and powerpc.

When FORTIFY_SOURCE is on, a number of functions are replaced with fortified versions, which attempt to check the sizes of the operands. However, these functions often directly invoke __builtin_foo() once they have performed the fortify check.

This breaks things in 2 ways:

- the three function calls are technically dead code, and can be eliminated. When __builtin_ versions are used, the compiler can detect this.

- Using __builtins may bypass KASAN checks if the compiler decides to inline it's own implementation as sequence of instructions, rather than emit a function call that goes out to a KASAN-instrumented implementation.

The patches address each reason in turn. Finally, test_memcmp used a stack array without explicit initialisation, which can sometimes break too, so fix that up.

v3: resend with Reviewed-bys, hopefully for inclusion in 5.8.

v2: - some cleanups, don't mess with arch code as I missed some wrinkles.
- add stack array init (patch 3)

Daniel Axtens (3):
kasan: stop tests being eliminated as dead code with FORTIFY_SOURCE string.h: fix incompatibility between FORTIFY_SOURCE and KASAN
kasan: initialise array in kasan_memcmp test

 include/linux/string.h | 60 +++++++++++++++++++++++++++++++++---------
 lib/test_kasan.c       | 32 +++++++++++++---------
 2 files changed, 68 insertions(+), 24 deletions(-)

2.20.1

3 KASAN self-tests fail on a kernel with both KASAN and FORTIFY_SOURCE:
memchr, memcmp and strlen.

When FORTIFY_SOURCE is on, a number of functions are replaced with fortified versions, which attempt to check the sizes of the operands. However, these functions often directly invoke __builtin_foo() once they have performed the fortify check. The compiler can detect that the results of these functions are not used, and knows that they have no other side effects, and so can eliminate them as dead code.

Why are only memchr, memcmp and strlen affected?
================================================

Of string and string-like functions, kasan_test tests:

 * strchr  ->  not affected, no fortified version
 * strrchr ->  likewise
 * strcmp  ->  likewise
 * strncmp ->  likewise

* strnlen -> not affected, the fortify source implementation calls the underlying strnlen implementation which is instrumented, not a builtin

* strlen -> affected, the fortify souce implementation calls a __builtin version which the compiler can determine is dead.

 * memchr  ->  likewise
 * memcmp  ->  likewise

* memset -> not affected, the compiler knows that memset writes to its
first argument and therefore is not dead.

Why does this not affect the functions normally?
================================================

In string.h, these functions are not marked as __pure, so the compiler cannot know that they do not have side effects. If relevant functions are marked as __pure in string.h, we see the following warnings and the functions are elided:

lib/test_kasan.c: In function ‘kasan_memchr’:
lib/test_kasan.c:606:2: warning: statement with no effect [-Wunused-value]
memchr(ptr, '1', size + 1);
^~~~~~~~~~~~~~~~~~~~~~~~~~
lib/test_kasan.c: In function ‘kasan_memcmp’:
lib/test_kasan.c:622:2: warning: statement with no effect [-Wunused-value]
memcmp(ptr, arr, size+1);
^~~~~~~~~~~~~~~~~~~~~~~~
lib/test_kasan.c: In function ‘kasan_strings’:
lib/test_kasan.c:645:2: warning: statement with no effect [-Wunused-value]
strchr(ptr, '1');
^~~~~~~~~~~~~~~~
...

This annotation would make sense to add and could be added at any point, so the behaviour of test_kasan.c should change.

The fix
=======

Make all the functions that are pure write their results to a global, which makes them live. The strlen and memchr tests now pass.

The memcmp test still fails to trigger, which is addressed in the next patch.

---
 lib/test_kasan.c | 30 +++++++++++++++++++-----------
 1 file changed, 19 insertions(+), 11 deletions(-)

index e3087d90e00d..939f395a5392 100644
@@ -23,6 +23,14 @@
 
 #include <asm/page.h>
 
+/*
+ * We assign some test results to these globals to make sure the tests
+ * are not eliminated as dead code.
+ */
+
+int kasan_int_result;
+void *kasan_ptr_result;
+
 /*
  * Note: test functions are marked noinline so that their names appear in
  * reports.

@@ -622,7 +630,7 @@ static noinline void __init kasan_memchr(void)
 	if (!ptr)
 		return;
 
-	memchr(ptr, '1', size + 1);
+	kasan_ptr_result = memchr(ptr, '1', size + 1);
 	kfree(ptr);
 }
 

@@ -637,8 +645,7 @@ static noinline void __init kasan_memcmp(void)
 	if (!ptr)
 		return;
 
-	memset(arr, 0, sizeof(arr));
-	memcmp(ptr, arr, size+1);
+	kasan_int_result = memcmp(ptr, arr, size + 1);
 	kfree(ptr);
 }
 

@@ -661,22 +668,22 @@ static noinline void __init kasan_strings(void)
 	 * will likely point to zeroed byte.
 	 */
 	ptr += 16;
-	strchr(ptr, '1');
+	kasan_ptr_result = strchr(ptr, '1');
 
 	pr_info("use-after-free in strrchr\n");
-	strrchr(ptr, '1');
+	kasan_ptr_result = strrchr(ptr, '1');
 
 	pr_info("use-after-free in strcmp\n");
-	strcmp(ptr, "2");
+	kasan_int_result = strcmp(ptr, "2");
 
 	pr_info("use-after-free in strncmp\n");
-	strncmp(ptr, "2", 1);
+	kasan_int_result = strncmp(ptr, "2", 1);
 
 	pr_info("use-after-free in strlen\n");
-	strlen(ptr);
+	kasan_int_result = strlen(ptr);
 
 	pr_info("use-after-free in strnlen\n");
-	strnlen(ptr, 1);
+	kasan_int_result = strnlen(ptr, 1);
 }
 
 static noinline void __init kasan_bitops(void)

@@ -743,11 +750,12 @@ static noinline void __init kasan_bitops(void)
 	__test_and_change_bit(BITS_PER_LONG + BITS_PER_BYTE, bits);
 
 	pr_info("out-of-bounds in test_bit\n");
-	(void)test_bit(BITS_PER_LONG + BITS_PER_BYTE, bits);
+	kasan_int_result = test_bit(BITS_PER_LONG + BITS_PER_BYTE, bits);
 
 #if defined(clear_bit_unlock_is_negative_byte)
 	pr_info("out-of-bounds in clear_bit_unlock_is_negative_byte\n");
-	clear_bit_unlock_is_negative_byte(BITS_PER_LONG + BITS_PER_BYTE, bits);
+	kasan_int_result = clear_bit_unlock_is_negative_byte(BITS_PER_LONG +
+		BITS_PER_BYTE, bits);
 #endif
 	kfree(bits);
 }

2.20.1

The memcmp KASAN self-test fails on a kernel with both KASAN and FORTIFY_SOURCE.

When FORTIFY_SOURCE is on, a number of functions are replaced with fortified versions, which attempt to check the sizes of the operands. However, these functions often directly invoke __builtin_foo() once they have performed the fortify check. Using __builtins may bypass KASAN checks if the compiler decides to inline it's own implementation as sequence of instructions, rather than emit a function call that goes out to a KASAN-instrumented implementation.

Why is only memcmp affected?
============================

Of the string and string-like functions that kasan_test tests, only memcmp is replaced by an inline sequence of instructions in my testing on x86 with gcc version 9.2.1 20191008 (Ubuntu 9.2.1-9ubuntu2).

I believe this is due to compiler heuristics. For example, if I annotate kmalloc calls with the alloc_size annotation (and disable some fortify compile-time checking!), the compiler will replace every memset except the one in kmalloc_uaf_memset with inline instructions. (I have some WIP patches to add this annotation.)

Does this affect other functions in string.h?
=============================================

Yes. Anything that uses __builtin_* rather than __real_* could be affected. This looks like:

- strncpy
- strcat
- strlen
- strlcpy maybe, under some circumstances?
- strncat under some circumstances
- memset
- memcpy
- memmove
- memcmp (as noted)
- memchr
- strcpy

Whether a function call is emitted always depends on the compiler. Most bugs should get caught by FORTIFY_SOURCE, but the missed memcmp test shows that this is not always the case.

Isn't FORTIFY_SOURCE disabled with KASAN?
========================================-

The string headers on all arches supporting KASAN disable fortify with kasan, but only when address sanitisation is _also_ disabled. For example from x86:

#if defined(CONFIG_KASAN) && !defined(__SANITIZE_ADDRESS__) /*
* For files that are not instrumented (e.g. mm/slub.c) we
* should use not instrumented version of mem* functions. */
#define memcpy(dst, src, len) __memcpy(dst, src, len) #define memmove(dst, src, len) __memmove(dst, src, len) #define memset(s, c, n) __memset(s, c, n)

#ifndef __NO_FORTIFY
#define __NO_FORTIFY /* FORTIFY_SOURCE uses __builtin_memcpy, etc. */ #endif

#endif

This comes from commit 6974f0c4555e ("include/linux/string.h: add the option of fortified string.h functions"), and doesn't work when KASAN is enabled and the file is supposed to be sanitised - as with test_kasan.c

I'm pretty sure this is not wrong, but not as expansive it should be:

* we shouldn't use __builtin_memcpy etc in files where we don't have instrumentation - it could devolve into a function call to memcpy, which will be instrumented. Rather, we should use __memcpy which by convention is not instrumented.

* we also shouldn't be using __builtin_memcpy when we have a KASAN instrumented file, because it could be replaced with inline asm that will not be instrumented.

What is correct behaviour?
==========================

Firstly, there is some overlap between fortification and KASAN: both provide some level of _runtime_ checking. Only fortify provides compile-time checking.

KASAN and fortify can pick up different things at runtime:

- Some fortify functions, notably the string functions, could easily be modified to consider sub-object sizes (e.g. members within a struct), and I have some WIP patches to do this. KASAN cannot detect these because it cannot insert poision between members of a struct.

- KASAN can detect many over-reads/over-writes when the sizes of both operands are unknown, which fortify cannot.

So there are a couple of options:

1) Flip the test: disable fortify in santised files and enable it in unsanitised files. This at least stops us missing KASAN checking, but we lose the fortify checking.

2) Make the fortify code always call out to real versions. Do this only for KASAN, for fear of losing the inlining opportunities we get from __builtin_*.

(We can't use kasan_check_{read,write}: because the fortify functions are _extern inline_, you can't include _static_ inline functions without a compiler warning. kasan_check_{read,write} are static inline so we can't use them even when they would otherwise be suitable.)

Take approach 2 and call out to real versions when KASAN is enabled.

Use __underlying_foo to distinguish from __real_foo: __real_foo always refers to the kernel's implementation of foo, __underlying_foo could be either the kernel implementation or the __builtin_foo implementation.

This is sometimes enough to make the memcmp test succeed with FORTIFY_SOURCE enabled. It is at least enough to get the function call into the module. One more fix is needed to make it reliable: see the next patch.

---
 include/linux/string.h | 60 +++++++++++++++++++++++++++++++++---------
 1 file changed, 48 insertions(+), 12 deletions(-)

index 6dfbb2efa815..9b7a0632e87a 100644
@@ -272,6 +272,31 @@ void __read_overflow3(void) __compiletime_error("detected read beyond size of ob
 void __write_overflow(void) __compiletime_error("detected write beyond size of object passed as 1st parameter");
 
 #if !defined(__NO_FORTIFY) && defined(__OPTIMIZE__) && defined(CONFIG_FORTIFY_SOURCE)
+
+#ifdef CONFIG_KASAN
+extern void *__underlying_memchr(const void *p, int c, __kernel_size_t size) __RENAME(memchr);
+extern int __underlying_memcmp(const void *p, const void *q, __kernel_size_t size) __RENAME(memcmp);
+extern void *__underlying_memcpy(void *p, const void *q, __kernel_size_t size) __RENAME(memcpy);
+extern void *__underlying_memmove(void *p, const void *q, __kernel_size_t size) __RENAME(memmove);
+extern void *__underlying_memset(void *p, int c, __kernel_size_t size) __RENAME(memset);
+extern char *__underlying_strcat(char *p, const char *q) __RENAME(strcat);
+extern char *__underlying_strcpy(char *p, const char *q) __RENAME(strcpy);
+extern __kernel_size_t __underlying_strlen(const char *p) __RENAME(strlen);
+extern char *__underlying_strncat(char *p, const char *q, __kernel_size_t count) __RENAME(strncat);
+extern char *__underlying_strncpy(char *p, const char *q, __kernel_size_t size) __RENAME(strncpy);
+#else
+#define __underlying_memchr	__builtin_memchr
+#define __underlying_memcmp	__builtin_memcmp
+#define __underlying_memcpy	__builtin_memcpy
+#define __underlying_memmove	__builtin_memmove
+#define __underlying_memset	__builtin_memset
+#define __underlying_strcat	__builtin_strcat
+#define __underlying_strcpy	__builtin_strcpy
+#define __underlying_strlen	__builtin_strlen
+#define __underlying_strncat	__builtin_strncat
+#define __underlying_strncpy	__builtin_strncpy
+#endif
+
 __FORTIFY_INLINE char *strncpy(char *p, const char *q, __kernel_size_t size)
 {
 	size_t p_size = __builtin_object_size(p, 0);

@@ -279,14 +304,14 @@ __FORTIFY_INLINE char *strncpy(char *p, const char *q, __kernel_size_t size)
 		__write_overflow();
 	if (p_size < size)
 		fortify_panic(__func__);
-	return __builtin_strncpy(p, q, size);
+	return __underlying_strncpy(p, q, size);
 }
 
 __FORTIFY_INLINE char *strcat(char *p, const char *q)
 {
 	size_t p_size = __builtin_object_size(p, 0);
 	if (p_size == (size_t)-1)
-		return __builtin_strcat(p, q);
+		return __underlying_strcat(p, q);
 	if (strlcat(p, q, p_size) >= p_size)
 		fortify_panic(__func__);
 	return p;

@@ -300,7 +325,7 @@ __FORTIFY_INLINE __kernel_size_t strlen(const char *p)
 	/* Work around gcc excess stack consumption issue */
 	if (p_size == (size_t)-1 ||
 	    (__builtin_constant_p(p[p_size - 1]) && p[p_size - 1] == '\0'))
-		return __builtin_strlen(p);
+		return __underlying_strlen(p);
 	ret = strnlen(p, p_size);
 	if (p_size <= ret)
 		fortify_panic(__func__);

@@ -333,7 +358,7 @@ __FORTIFY_INLINE size_t strlcpy(char *p, const char *q, size_t size)
 			__write_overflow();
 		if (len >= p_size)
 			fortify_panic(__func__);
-		__builtin_memcpy(p, q, len);
+		__underlying_memcpy(p, q, len);
 		p[len] = '\0';
 	}
 	return ret;

@@ -346,12 +371,12 @@ __FORTIFY_INLINE char *strncat(char *p, const char *q, __kernel_size_t count)
 	size_t p_size = __builtin_object_size(p, 0);
 	size_t q_size = __builtin_object_size(q, 0);
 	if (p_size == (size_t)-1 && q_size == (size_t)-1)
-		return __builtin_strncat(p, q, count);
+		return __underlying_strncat(p, q, count);
 	p_len = strlen(p);
 	copy_len = strnlen(q, count);
 	if (p_size < p_len + copy_len + 1)
 		fortify_panic(__func__);
-	__builtin_memcpy(p + p_len, q, copy_len);
+	__underlying_memcpy(p + p_len, q, copy_len);
 	p[p_len + copy_len] = '\0';
 	return p;
 }

@@ -363,7 +388,7 @@ __FORTIFY_INLINE void *memset(void *p, int c, __kernel_size_t size)
 		__write_overflow();
 	if (p_size < size)
 		fortify_panic(__func__);
-	return __builtin_memset(p, c, size);
+	return __underlying_memset(p, c, size);
 }
 
 __FORTIFY_INLINE void *memcpy(void *p, const void *q, __kernel_size_t size)

@@ -378,7 +403,7 @@ __FORTIFY_INLINE void *memcpy(void *p, const void *q, __kernel_size_t size)
 	}
 	if (p_size < size || q_size < size)
 		fortify_panic(__func__);
-	return __builtin_memcpy(p, q, size);
+	return __underlying_memcpy(p, q, size);
 }
 
 __FORTIFY_INLINE void *memmove(void *p, const void *q, __kernel_size_t size)

@@ -393,7 +418,7 @@ __FORTIFY_INLINE void *memmove(void *p, const void *q, __kernel_size_t size)
 	}
 	if (p_size < size || q_size < size)
 		fortify_panic(__func__);
-	return __builtin_memmove(p, q, size);
+	return __underlying_memmove(p, q, size);
 }
 
 extern void *__real_memscan(void *, int, __kernel_size_t) __RENAME(memscan);

@@ -419,7 +444,7 @@ __FORTIFY_INLINE int memcmp(const void *p, const void *q, __kernel_size_t size)
 	}
 	if (p_size < size || q_size < size)
 		fortify_panic(__func__);
-	return __builtin_memcmp(p, q, size);
+	return __underlying_memcmp(p, q, size);
 }
 
 __FORTIFY_INLINE void *memchr(const void *p, int c, __kernel_size_t size)

@@ -429,7 +454,7 @@ __FORTIFY_INLINE void *memchr(const void *p, int c, __kernel_size_t size)
 		__read_overflow();
 	if (p_size < size)
 		fortify_panic(__func__);
-	return __builtin_memchr(p, c, size);
+	return __underlying_memchr(p, c, size);
 }
 
 void *__real_memchr_inv(const void *s, int c, size_t n) __RENAME(memchr_inv);

@@ -460,11 +485,22 @@ __FORTIFY_INLINE char *strcpy(char *p, const char *q)
 	size_t p_size = __builtin_object_size(p, 0);
 	size_t q_size = __builtin_object_size(q, 0);
 	if (p_size == (size_t)-1 && q_size == (size_t)-1)
-		return __builtin_strcpy(p, q);
+		return __underlying_strcpy(p, q);
 	memcpy(p, q, strlen(q) + 1);
 	return p;
 }
 
+/* Don't use these outside the FORITFY_SOURCE implementation */
+#undef __underlying_memchr
+#undef __underlying_memcmp
+#undef __underlying_memcpy
+#undef __underlying_memmove
+#undef __underlying_memset
+#undef __underlying_strcat
+#undef __underlying_strcpy
+#undef __underlying_strlen
+#undef __underlying_strncat
+#undef __underlying_strncpy
 #endif
 
 /**

2.20.1

memcmp may bail out before accessing all the memory if the buffers contain differing bytes. kasan_memcmp calls memcmp with a stack array. Stack variables are not necessarily initialised (in the absence of a compiler plugin, at least). Sometimes this causes the memcpy to bail early thus fail to trigger kasan.

Make sure the array initialised to zero in the code.

No other test is dependent on the contents of an array on the stack.

---
 lib/test_kasan.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

index 939f395a5392..7700097842c8 100644
@@ -638,7 +638,7 @@ static noinline void __init kasan_memcmp(void)
 {
 	char *ptr;
 	size_t size = 24;
-	int arr[9];
+	int arr[9] = {};
 
 	pr_info("out-of-bounds in memcmp\n");
 	ptr = kmalloc(size, GFP_KERNEL | __GFP_ZERO);

2.20.1

Ah! It turns out I accidentally removed the memset in patch 1. No idea why I did that. I'll fix up patch 1 to not remove the memset and drop this patch.

Daniel

Thanks, Daniel!

For the series:

(Though I will mirror Dmitry's comment[1] on patch 3 -- I also have a memset() already present in my branch...)

I'd been digging into what turns out to be this issue, which we were seeing sporadically[2] with the KUnit port of these tests. v7 of the KUnit port[3] includes your changes, and fixes the issues.

Cheers,
-- David

[1]: https://lkml.org/lkml/2020/4/23/838
[2]: https://lkml.org/lkml/2020/4/18/570
[3]: https://lkml.org/lkml/2020/4/24/80