[cfe-dev] Why does integer overflow sanitizer exits the program with zero exit code

Mon Nov 9 12:14:11 PST 2015

On Mon, Nov 9, 2015 at 11:52 AM, Kostya Serebryany <kcc at google.com> wrote:

> Intuitively, one may indeed think that some signed integer overflows are
> more dangerous than others.
> But scientifically, they are all bugs.
>

I agree, they are all bugs unless someone using overflows for optimizing
code. But from security perspective not all bugs are security
vulnerabilities. Right now I'm interested into security vulnerabilities
only.

My favorite real-life example:
> http://stackoverflow.com/questions/7682477/why-does-integer-overflow-on-x86-with-gcc-cause-an-infinite-loop
>
> It is totally unclear if we can come up with a reliable heuristic that
> will distinguish between less harmful and more harmful int overflows.
> E.g. what if the computation of an array index happens in a separate
> function?
>
>
In academic security research, what I've seen so far is that dynamic taint
analysis is done with conjunction with overflow checks. If the operands in
overflows are tainted by user input and that arithmetic overflow leads to
crashes etc. i.e., an user has the ability to crash the program by simply
providing different kinds of input, then it's more interesting. But again,
things are not that simple. Overflows can change the intended semantics of
the program which in turn can lead to security vulnerabilities.

Coming back to the original question, is there any API or option to enable
overflow checks for certain translation units?

--kcc
>
>
> On Mon, Nov 9, 2015 at 7:07 AM, Riyad Parvez via cfe-dev <
> cfe-dev at lists.llvm.org> wrote:
>
>> On Sun, Nov 8, 2015 at 11:07 PM, Alexander Potapenko <glider at google.com>
>> wrote:
>>
>>> Can you please give an example of a benign integer overflow?
>>>
>> Here 'benign' from security perspective is usually defined as arithmetic
>> overflows that don't trigger low-level security vulnerabilities e.g. buffer
>> overflows. Many arithmetic overflows don't trigger buffer overflows
>> eventually. Copy pasted from [1]:
>>
>> For example, if x is an int type variable, the statement if(x >= -2 &&
>> x<= 0x7ffffffd) will be translated into such a piece of assembly code by
>> GCC-4.2.0 compiler: mov eax, x; // eax = x add eax, 2; // eax = eax+2 js
>> target In this case, a large x such as 0x7fffffff causes an overflow in
>> above add instruction, but it is harmless as GCC actually uses this
>> overflow to reduce a comparison instruction.
>>
>> Note that by disabling the checks for non-pointer arithmetic you can
>>> easily throw the baby out with the bathwater. The compiler can easily take
>>> advantage of signed integer overflow being undefined irrespective of it
>>> being called "benign" or not.
>>>
>> There's a good paper on this, you can have a look if you are interested
>> [2].
>>
>> [1] https://www.utdallas.edu/~zxl111930/file/IntScope_NDSS09.pdf
>> [2] https://people.csail.mit.edu/nickolai/papers/wang-stack.pdf
>>
>>> sent from phone
>>> On Nov 7, 2015 1:19 PM, "Riyad Parvez via cfe-dev" <
>>> cfe-dev at lists.llvm.org> wrote:
>>>
>>>> Is it possible to enable integer overflow sanitizer for certain cases
>>>> as in many cases integer overflows are benign? I'm interested in mainly
>>>> integer overflow during pointer arithmetic or memory allocations. Or
>>>> enabling sanitizer for certain course files?
>>>>
>>>> Thanks,
>>>> Riyad
>>>>
>>>> On Tue, Nov 3, 2015 at 7:43 PM, Alexey Samsonov <vonosmas at gmail.com>
>>>> wrote:
>>>>
>>>>>
>>>>> On Tue, Nov 3, 2015 at 3:26 PM, Riyad Parvez via cfe-dev <
>>>>> cfe-dev at lists.llvm.org> wrote:
>>>>>
>>>>>> Thanks a lot.
>>>>>>
>>>>>> I had decided to detect overflow by SIGABRT. The issue SIGABRT is
>>>>>> also thorwn by when assertion violation occurs. So SIGABRT is not the
>>>>>> perfect indicator of overflow.
>>>>>>
>>>>>> I see all the arithmetic operations are performed by
>>>>>> "__ubsan_handle_*_overflow" built-in functions. If I catch SIGABRT,
>>>>>> "__ubsan_handle_*_overflow" functions in the stack, can I safely assume the
>>>>>> program is terminated in overflow?
>>>>>>
>>>>>
>>>>> Probably yes: if one of these function was called by the program, it
>>>>> will certainly report an overflow. In theory, we don't guarantee that the
>>>>> names of these functions will stay the same, or that we always report
>>>>> overflows through them (i.e. it's an implementation detail), but in
>>>>> practice it would unlikely change soon.
>>>>>
>>>>>
>>>>>>
>>>>>> Thanks,
>>>>>> Riyad
>>>>>>
>>>>>> On Fri, Oct 23, 2015 at 3:01 PM, Kostya Serebryany <kcc at google.com>
>>>>>> wrote:
>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> On Fri, Oct 23, 2015 at 11:35 AM, Riyad Parvez <
>>>>>>> riyad.parvez at uwaterloo.ca> wrote:
>>>>>>>
>>>>>>>> Interesting! "UBSAN_OPTIONS=halt_on_error=1:abort_on_error=1" works
>>>>>>>> fine, but just "UBSAN_OPTIONS=abort_on_error=1" doesn't work. Is it by
>>>>>>>> design?
>>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> sort of.
>>>>>>> by default, ubsan simply does not exit, so it does not care about
>>>>>>> abort_on_error.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>>
>>>>>>>> On Fri, Oct 23, 2015 at 2:30 PM, Kostya Serebryany <kcc at google.com>
>>>>>>>> wrote:
>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On Fri, Oct 23, 2015 at 11:27 AM, Riyad Parvez <
>>>>>>>>> riyad.parvez at uwaterloo.ca> wrote:
>>>>>>>>>
>>>>>>>>>> Thanks! It works on the latest version of clang, but I was using
>>>>>>>>>> clang 3.4 where it didn't work.
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>> 3.4 was loooong ago.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> On a more related note, SIGABRT better indicator of bugs than
>>>>>>>>>> non-zero exit status. So I've tried "ASAN_OPTIONS=abort_on_error=1" option
>>>>>>>>>> which works for address sanitizer where program exists with SIGABRT. But
>>>>>>>>>> "UBSAN_OPTIONS=abort_on_error=1" doesn't exit with SIGABRT. Is there any
>>>>>>>>>> way to make overflow sanitizer exit with SIGABRT similar to address
>>>>>>>>>> sanitizer?
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Just checked:
>>>>>>>>>
>>>>>>>>> % clang -fsanitize=address,signed-integer-overflow int-overflow.c
>>>>>>>>> && UBSAN_OPTIONS=halt_on_error=1:abort_on_error=1 ./a.out ; echo EXIT
>>>>>>>>> STATUS: $?
>>>>>>>>> int-overflow.c:5:5: runtime error: signed integer overflow:
>>>>>>>>> 1073741824 + 1073741824 cannot be represented in type 'int'
>>>>>>>>> SUMMARY: AddressSanitizer: undefined-behavior int-overflow.c:5:5
>>>>>>>>> in
>>>>>>>>> Aborted (core dumped)
>>>>>>>>> EXIT STATUS: 134
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> On Fri, Oct 23, 2015 at 1:06 PM, Kostya Serebryany <
>>>>>>>>>> kcc at google.com> wrote:
>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> On Fri, Oct 23, 2015 at 7:08 AM, Riyad Parvez <
>>>>>>>>>>> riyad.parvez at uwaterloo.ca> wrote:
>>>>>>>>>>>
>>>>>>>>>>>> Thanks for the reply.
>>>>>>>>>>>>
>>>>>>>>>>>> I am developing a fuzzer and interested to find overflows.
>>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> So am I :)
>>>>>>>>>>> llvm.org/docs/LibFuzzer.html
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>> Is there any way to detect when errors are detected? To explain
>>>>>>>>>>>> more, I have instrumented the binary to see which functions are called in
>>>>>>>>>>>> run-time. What are the functions that are called when overflow is detected?
>>>>>>>>>>>> When these functions are called I will know overflow is detected.
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>> On Thu, Oct 22, 2015 at 8:08 PM, Kostya Serebryany <
>>>>>>>>>>>> kcc at google.com> wrote:
>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> On Thu, Oct 22, 2015 at 11:52 AM, Riyad Parvez via cfe-dev <
>>>>>>>>>>>>> cfe-dev at lists.llvm.org> wrote:
>>>>>>>>>>>>>
>>>>>>>>>>>>>> Hi All,
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> With "-fsanitize=integer" flag, when an overflow is detected
>>>>>>>>>>>>>> the program is terminated with zero exit code.
>>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> You can change this behavior by using env. var.
>>>>>>>>>>>>> UBSAN_OPTIONS=halt_on_error=1
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>> I've tried this; didn't work.
>>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> What exactly do you run?
>>>>>>>>>>>
>>>>>>>>>>> Here is what works for me:
>>>>>>>>>>>
>>>>>>>>>>> % cat int-overflow.c
>>>>>>>>>>> #include <stdio.h>
>>>>>>>>>>> int x;
>>>>>>>>>>> int main(int argc, char **argv) {
>>>>>>>>>>>   x += 1 << 30;
>>>>>>>>>>>   x += 1 << 30;
>>>>>>>>>>>   printf("%d\n", x);
>>>>>>>>>>>   return 0;
>>>>>>>>>>> }
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> % clang -fsanitize=signed-integer-overflow int-overflow.c &&
>>>>>>>>>>> ./a.out ; echo EXIT STATUS: $?
>>>>>>>>>>> int-overflow.c:5:5: runtime error: signed integer overflow:
>>>>>>>>>>> 1073741824 + 1073741824 cannot be represented in type 'int'
>>>>>>>>>>> -2147483648
>>>>>>>>>>> EXIT STATUS: 0
>>>>>>>>>>> % clang -fsanitize=signed-integer-overflow int-overflow.c &&
>>>>>>>>>>> UBSAN_OPTIONS=halt_on_error=1 ./a.out ; echo EXIT STATUS: $?
>>>>>>>>>>> int-overflow.c:5:5: runtime error: signed integer overflow:
>>>>>>>>>>> 1073741824 + 1073741824 cannot be represented in type 'int'
>>>>>>>>>>> EXIT STATUS: 1
>>>>>>>>>>> %
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>> (Not sure if this is properly documented anywhere. Alexey? )
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>>> But with "-fsanitize=address" flag, the program terminates
>>>>>>>>>>>>>> with non-zero exit code. I think the address sanitizer behavior of non-zero
>>>>>>>>>>>>>> exit code is more intuitive since the program did exit in error. Is there
>>>>>>>>>>>>>> any reason integer overflow sanitizer exits the program with zero exit code?
>>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> One of the reasons, maybe:
>>>>>>>>>>>>> Programs are more often ubsan-unclean than asan-unclean, and
>>>>>>>>>>>>> halting on every ubsan message makes it harder to deploy the tool.
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> Thanks,
>>>>>>>>>>>>>> Riyad
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> _______________________________________________
>>>>>>>>>>>>>> cfe-dev mailing list
>>>>>>>>>>>>>> cfe-dev at lists.llvm.org
>>>>>>>>>>>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> cfe-dev mailing list
>>>>>> cfe-dev at lists.llvm.org
>>>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> Alexey Samsonov
>>>>> vonosmas at gmail.com
>>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> cfe-dev mailing list
>>>> cfe-dev at lists.llvm.org
>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>>>>
>>>>
>>
>> _______________________________________________
>> cfe-dev mailing list
>> cfe-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20151109/0c1e1725/attachment.html>