Message-Id: <1437347852-24921-1-git-send-email-linux@rasmusvillemoes.dk>
Date:	Mon, 20 Jul 2015 01:17:30 +0200
From:	Rasmus Villemoes <linux@...musvillemoes.dk>
To:	mingo@...nel.org, akpm@...ux-foundation.org,
	Sasha Levin <sasha.levin@...cle.com>
Cc:	Rasmus Villemoes <linux@...musvillemoes.dk>,
	linux-kernel@...r.kernel.org
Subject: [RFC 1/3] compiler.h: enable builtin overflow checkers and add fallback code

Last year, Sasha Levin suggested adding wrappers for the
__builtin_*_overflow functions introduced with gcc 5.1 (based on
similar, but type-specific, functions in clang). This is another
attempt at providing such wrappers and fallback code for older compilers.

There are a few problems with the 'a+b < a' idiom for checking for
overflow: For signed types, it relies on undefined behaviour and is
not actually complete (it doesn't check underflow;
e.g. INT_MIN+INT_MIN == 0 isn't caught), and due to type promotion it
is wrong for all types narrower than int.

The new overflow.h is somewhat bulky, but that's mostly a result of
trying to be type-generic, complete (e.g. catching not only overflow
but also signed underflow) and not relying on undefined behaviour.

So is it worth it? I think it is, if nothing else for the documentation
value of seeing

  if (check_add_overflow(a, b, &d))
    return -EGOAWAY;
  do_stuff_with(d);

instead of the open-coded (and possibly wrong and/or incomplete and/or
UBsan-tickling)

  if (a+b < a)
    return -EGOAWAY;
  do_stuff_with(a+b);

While gcc does recognize the 'a+b < a' idiom for testing unsigned
overflow, it doesn't do nearly as good for unsigned multiplication
(there's also no single well-established idiom). So using
check_mul_overflow in kcalloc and friends may also make gcc generate
slightly better code.

To, admittedly rather artificially, illustrate the potential benefit,
a trivial program which verifies the correctness for all possible u8,
s8, u16 and s16 values runs in 68 seconds when using the fallback
code, and in 50 seconds when using the builtins.

Signed-off-by: Rasmus Villemoes <linux@...musvillemoes.dk>
---
 include/linux/compiler-clang.h |   6 ++
 include/linux/compiler-gcc.h   |   4 +
 include/linux/compiler-intel.h |   4 +
 include/linux/overflow.h       | 204 +++++++++++++++++++++++++++++++++++++++++
 4 files changed, 218 insertions(+)
 create mode 100644 include/linux/overflow.h

diff --git a/include/linux/compiler-clang.h b/include/linux/compiler-clang.h
index d1e49d52b640..f08003824080 100644
--- a/include/linux/compiler-clang.h
+++ b/include/linux/compiler-clang.h
@@ -10,3 +10,9 @@
 #undef uninitialized_var
 #define uninitialized_var(x) x = *(&(x))
 #endif
+
+/*
+ * clang defines __GNUC__, but does not implement the type-generic
+ * version of the builtin overflow checkers.
+ */
+#undef COMPILER_HAS_GENERIC_BUILTIN_OVERFLOW
diff --git a/include/linux/compiler-gcc.h b/include/linux/compiler-gcc.h
index dfaa7b3e9ae9..827ca2af785e 100644
--- a/include/linux/compiler-gcc.h
+++ b/include/linux/compiler-gcc.h
@@ -248,3 +248,7 @@
  * code
  */
 #define uninitialized_var(x) x = x
+
+#if GCC_VERSION >= 50100
+#define COMPILER_HAS_GENERIC_BUILTIN_OVERFLOW 1
+#endif
diff --git a/include/linux/compiler-intel.h b/include/linux/compiler-intel.h
index d4c71132d07f..8c9897b1b953 100644
--- a/include/linux/compiler-intel.h
+++ b/include/linux/compiler-intel.h
@@ -43,3 +43,7 @@
 #define __builtin_bswap16 _bswap16
 #endif
 
+/*
+ * icc defines __GNUC__, but does not implement the builtin overflow checkers.
+ */
+#undef COMPILER_HAS_GENERIC_BUILTIN_OVERFLOW
diff --git a/include/linux/overflow.h b/include/linux/overflow.h
new file mode 100644
index 000000000000..4e499ceac7ce
--- /dev/null
+++ b/include/linux/overflow.h
@@ -0,0 +1,204 @@
+#ifndef __LINUX_OVERFLOW_H
+#define __LINUX_OVERFLOW_H
+
+#include <linux/compiler.h>
+
+/*
+ * In the fallback code below, we need to compute the minimum and
+ * maximum values representable in a given type. These macros may also
+ * be useful elsewhere, so we provide them outside the
+ * COMPILER_HAS_GENERIC_BUILTIN_OVERFLOW block.
+ *
+ * It would seem more obvious to do something like
+ *
+ * #define type_min(T) (T)(is_signed_type(T) ? (T)1 << (8*sizeof(T)-1) : 0)
+ * #define type_max(T) (T)(is_signed_type(T) ? ((T)1 << (8*sizeof(T)-1)) - 1 : ~(T)0)
+ *
+ * Unfortunately, the middle expressions, strictly speaking, have
+ * undefined behaviour, and at least some versions of gcc warn about
+ * the type_max expression (but not if -fsanitize=undefined is in
+ * effect; in that case, the warning is deferred to runtime...).
+ *
+ * The slightly excessive casting in type_min is to make sure the
+ * macros also produce sensible values for the exotic type _Bool. [The
+ * overflow checkers only almost work for _Bool, but that's
+ * a-feature-not-a-bug, since people shouldn't be doing arithmetic on
+ * _Bools. Besides, the gcc builtins don't allow _Bool* as third
+ * argument.]
+ *
+ * Idea stolen from
+ * https://mail-index.netbsd.org/tech-misc/2007/02/05/0000.html -
+ * credit to Christian Biere.
+ */
+#define is_signed_type(type)       (((type)(-1)) < (type)1)
+#define __type_half_max(type) ((type)1 << (8*sizeof(type) - 1 - is_signed_type(type)))
+#define type_max(T) ((T)((__type_half_max(T) - 1) + __type_half_max(T)))
+#define type_min(T) ((T)((T)-type_max(T)-(T)1))
+
+
+#ifdef COMPILER_HAS_GENERIC_BUILTIN_OVERFLOW
+/*
+ * For simplicity and code hygiene, the fallback code below insists on
+ * a, b and *d having the same type (similar to the min() and max()
+ * macros), whereas gcc's type-generic overflow checkers accept
+ * different types. Hence we don't just make check_add_overflow an
+ * alias for __builtin_add_overflow, but add type checks similar to
+ * below.
+ */
+#define check_add_overflow(a, b, d) ({		\
+	typeof(a) __a = (a);			\
+	typeof(b) __b = (b);			\
+	typeof(d) __d = (d);			\
+	(void) (&__a == &__b);			\
+	(void) (&__a == __d);			\
+	__builtin_add_overflow(__a, __b, __d);	\
+})
+
+#define check_sub_overflow(a, b, d) ({		\
+	typeof(a) __a = (a);			\
+	typeof(b) __b = (b);			\
+	typeof(d) __d = (d);			\
+	(void) (&__a == &__b);			\
+	(void) (&__a == __d);			\
+	__builtin_sub_overflow(__a, __b, __d);	\
+})
+
+#define check_mul_overflow(a, b, d) ({		\
+	typeof(a) __a = (a);			\
+	typeof(b) __b = (b);			\
+	typeof(d) __d = (d);			\
+	(void) (&__a == &__b);			\
+	(void) (&__a == __d);			\
+	__builtin_mul_overflow(__a, __b, __d);	\
+})
+
+#else
+
+
+/* Checking for unsigned overflow is relatively easy without causing UB. */
+#define __unsigned_add_overflow(a, b, d) ({	\
+	typeof(a) __a = (a);			\
+	typeof(b) __b = (b);			\
+	typeof(d) __d = (d);			\
+	(void) (&__a == &__b);			\
+	(void) (&__a == __d);			\
+	*__d = __a + __b;			\
+	*__d < __a;				\
+})
+#define __unsigned_sub_overflow(a, b, d) ({	\
+	typeof(a) __a = (a);			\
+	typeof(b) __b = (b);			\
+	typeof(d) __d = (d);			\
+	(void) (&__a == &__b);			\
+	(void) (&__a == __d);			\
+	*__d = __a - __b;			\
+	__a < __b;				\
+})
+/*
+ * If one of a or b is a compile-time constant, one should pass that
+ * in b. This corresponds to the parameters to calloc(), where the
+ * second argument is often sizeof(something).
+ */
+#define __unsigned_mul_overflow(a, b, d) ({		\
+	typeof(a) __a = (a);				\
+	typeof(b) __b = (b);				\
+	typeof(d) __d = (d);				\
+	(void) (&__a == &__b);				\
+	(void) (&__a == __d);				\
+	*__d = __a * __b;				\
+	__b > 0 && __a > type_max(typeof(__a)) / __b;	\
+})
+
+/*
+ * For signed types, detecting overflow is much harder, especially if
+ * we want to avoid UB. But the interface of these macros is such that
+ * we must provide a result in *d, and in fact we must produce the
+ * result promised by gcc's builtins, which is simply the possibly
+ * wrapped-around value. Fortunately, we can just formally do the
+ * operations in the widest relevant unsigned type (u64) and then
+ * truncate the result - gcc is smart enough to generate the same code
+ * with and without the (u64) casts.
+ */
+
+/*
+ * Adding two signed integers can overflow only if they have the same
+ * sign, and overflow has happened iff the result has the opposite
+ * sign.
+ */
+#define __signed_add_overflow(a, b, d) ({	\
+	typeof(a) __a = (a);			\
+	typeof(b) __b = (b);			\
+	typeof(d) __d = (d);			\
+	(void) (&__a == &__b);			\
+	(void) (&__a == __d);			\
+	*__d = (u64)__a + (u64)__b;		\
+	(((~(__a ^ __b)) & (*__d ^ __a))	\
+		& type_min(typeof(__a))) != 0;	\
+})
+
+/*
+ * Subtraction is similar, except that overflow can now happen only
+ * when the signs are opposite. In this case, overflow has happened if
+ * the result has the opposite sign of a.
+ */
+#define __signed_sub_overflow(a, b, d) ({	\
+	typeof(a) __a = (a);			\
+	typeof(b) __b = (b);			\
+	typeof(d) __d = (d);			\
+	(void) (&__a == &__b);			\
+	(void) (&__a == __d);			\
+	*__d = (u64)__a - (u64)__b;		\
+	((((__a ^ __b)) & (*__d ^ __a))		\
+		& type_min(typeof(__a))) != 0;	\
+})
+
+/*
+ * Signed multiplication is rather hard. gcc always follows C99, so
+ * division is truncated towards 0. This means that we can write the
+ * overflow check like this:
+ *
+ * (a > 0 && (b > MAX/a || b < MIN/a)) ||
+ * (a < -1 && (b > MIN/a || b < MAX/a) ||
+ * (a == -1 && b == MIN)
+ *
+ * The redundant casts of -1 are to silence an annoying -Wtype-limits
+ * (included in -Wextra) warning: When the type is u8 or u16, the
+ * __b_c_e in check_mul_overflow obviously selects
+ * __unsigned_mul_overflow, but unfortunately gcc still parses this
+ * code and warns about the limited range of __b.
+ */
+
+#define __signed_mul_overflow(a, b, d) ({				\
+	typeof(a) __a = (a);						\
+	typeof(b) __b = (b);						\
+	typeof(d) __d = (d);						\
+	typeof(a) __tmax = type_max(typeof(a));				\
+	typeof(a) __tmin = type_min(typeof(a));				\
+	(void) (&__a == &__b);						\
+	(void) (&__a == __d);						\
+	*__d = (u64)__a * (u64)__b;					\
+	(__b > 0   && (__a > __tmax/__b || __a < __tmin/__b)) ||	\
+	(__b < (typeof(__b))-1  && (__a > __tmin/__b || __a < __tmax/__b)) || \
+	(__b == (typeof(__b))-1 && __a == __tmin);			\
+})
+
+
+#define check_add_overflow(a, b, d)					\
+	__builtin_choose_expr(is_signed_type(typeof(a)),		\
+			__signed_add_overflow(a, b, d),			\
+			__unsigned_add_overflow(a, b, d))
+
+#define check_sub_overflow(a, b, d)					\
+	__builtin_choose_expr(is_signed_type(typeof(a)),		\
+			__signed_sub_overflow(a, b, d),			\
+			__unsigned_sub_overflow(a, b, d))
+
+#define check_mul_overflow(a, b, d)					\
+	__builtin_choose_expr(is_signed_type(typeof(a)),		\
+			__signed_mul_overflow(a, b, d),			\
+			__unsigned_mul_overflow(a, b, d))
+
+
+#endif /* COMPILER_HAS_GENERIC_BUILTIN_OVERFLOW */
+
+#endif /* __LINUX_OVERFLOW_H */
-- 
2.1.3

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists