Re: [RFC PATCH] UBSan: run-time undefined behavior sanity checker

[Andrey Ryabinin]

On 10/22/2014 01:58 PM, Rasmus Villemoes wrote:
> On Mon, Oct 20 2014, Andrey Ryabinin <a.ryabinin@xxxxxxxxxxx> wrote:
> 
>> UBSan uses compile-time instrumentation to catch undefined behavior (UB).
>> Compiler inserts code that perform certain kinds of
>> checks before operations that could cause UB.
>> If check fails (i.e. UB detected) __ubsan_handle_* function called.
>> to print error message.
>>
>> So the most of the work is done by compiler.
>> This patch just implements ubsan handlers printing errors.
>>
>> GCC supports this since 4.9, however upcoming GCC 5.0 has
>> more checkers implemented.
> 
> [...]
> 
>> +
>> +#define REPORTED_BIT 31
>> +#define COLUMN_MASK (~(1U << REPORTED_BIT))
>> +
>> +static bool is_disabled(struct source_location *location)
>> +{
>> +	return test_and_set_bit(REPORTED_BIT,
>> +				(unsigned long *)&location->column);
>> +}
> 
> [...]
> 
>> +struct source_location {
>> +	const char *file_name;
>> +	u32 line;
>> +	u32 column;
>> +};
> 
> 
> AFAICT, this introduces UB and/or memory corruption on big-endian
> systems with BITS_PER_LONG==64. (Also, on both LE and BE 64 bit systems,
> there's the issue of the alignment of location->column, which is likely
> to be 4-but-not-8 byte aligned).
> 

You are right. This should fix it:

diff --git a/lib/ubsan.c b/lib/ubsan.c
index 7788f47..cfdf017 100644
--- a/lib/ubsan.c
+++ b/lib/ubsan.c
@@ -53,18 +53,24 @@ static bool handler_enabled(unsigned int handler)
 }

 #define REPORTED_BIT 31
-#define COLUMN_MASK (~(1U << REPORTED_BIT))
+
+#if (BITS_PER_LONG == 64) && defined(__BIG_ENDIAN)
+#define COLUMN_MASK (~(1U << REPORTED_BIT)
+#define LINE_MASK   (~0U)
+#else
+#define COLUMN_MASK (~0U)
+#define LINE_MASK   (~(1U << REPORTED_BIT))
+#endif

 static bool is_disabled(struct source_location *location)
 {
-	return test_and_set_bit(REPORTED_BIT,
-				(unsigned long *)&location->column);
+	return test_and_set_bit(REPORTED_BIT, &location->reported);
 }

 static void print_source_location(const char *prefix, struct source_location *loc)
 {
 	pr_err("%s %s:%d:%d\n", prefix, loc->file_name,
-		loc->line, loc->column & COLUMN_MASK);
+		loc->line & LINE_MASK, loc->column & COLUMN_MASK);
 }

 static bool type_is_int(struct type_descriptor *type)
diff --git a/lib/ubsan.h b/lib/ubsan.h
index e2d8634..8965591 100644
--- a/lib/ubsan.h
+++ b/lib/ubsan.h
@@ -15,8 +15,13 @@ struct type_descriptor {

 struct source_location {
 	const char *file_name;
-	u32 line;
-	u32 column;
+	union {
+		unsigned long reported;
+		struct {
+			u32 line;
+			u32 column;
+		};
+	};
 };

 struct overflow_data {


> Is the layout of struct source_location dictated by gcc? 
> 

Yes.

> Rasmus
> 

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