[PATCH RFC] exec: avoid possible undefined behavior in count()

[PATCH RFC] exec: avoid possible undefined behavior in count()

[Xi Wang]

The tricky problem is this check:

	if (i++ >= max)

icc (mis)optimizes this check as:

	if (++i > max)

The check now becomes a no-op since max is MAX_ARG_STRINGS (0x7FFFFFFF).

This is "allowed" by the C standard, assuming i++ never overflows,
because signed integer overflow is undefined behavior.  This optimization
effectively reverts the previous commit 362e6663ef ("exec.c, compat.c:
fix count(), compat_count() bounds checking") that tries to fix the check.

This patch simply moves ++ after the check.

Signed-off-by: Xi Wang <xi.wang@gmail.com>
---
Not sure how many people are using Intel's icc to compiled the kernel.
Some projects like LinuxDNA did.

The kernel uses gcc's -fno-strict-overflow to disable this optimization.
icc probably doesn't recognize the option.

To illustrate the problem, try this simple program:

int count(int i, int max)
{
        if (i++ >= max) {
                __builtin_trap();
                return -1;
        }
        return i;
}

#include <stdio.h>
#include <stdlib.h>

int main(int argc, char **argv)
{
        int x = atoi(argv[1]);
        int max = atoi(argv[2]);
        printf("%d %d %d\n", x, max, count(x, max));
}

$ gcc -O2 t.c
$ ./a.out 2147483647 2147483647
Illegal instruction (core dumped)

$ icc -O2 t.c
$ ./a.out 2147483647 2147483647
2147483647 2147483647 -2147483648

There's no difference whether we add -fno-strict-overflow or not.
---
 fs/exec.c |    3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/fs/exec.c b/fs/exec.c
index 18c45ca..20df02c 100644
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -434,8 +434,9 @@ static int count(struct user_arg_ptr argv, int max)
 			if (IS_ERR(p))
 				return -EFAULT;
 
-			if (i++ >= max)
+			if (i >= max)
 				return -E2BIG;
+			++i;
 
 			if (fatal_signal_pending(current))
 				return -ERESTARTNOHAND;
-- 
1.7.10.4

Re: [PATCH RFC] exec: avoid possible undefined behavior in count()

[Andrew Morton]

On Sun,  6 Jan 2013 00:29:05 -0500
Xi Wang <xi.wang@gmail.com> wrote:

> The tricky problem is this check:
> 
> 	if (i++ >= max)
> 
> icc (mis)optimizes this check as:
> 
> 	if (++i > max)
> 
> The check now becomes a no-op since max is MAX_ARG_STRINGS (0x7FFFFFFF).
> 
> This is "allowed" by the C standard, assuming i++ never overflows,
> because signed integer overflow is undefined behavior.  This optimization
> effectively reverts the previous commit 362e6663ef ("exec.c, compat.c:
> fix count(), compat_count() bounds checking") that tries to fix the check.
> 
> This patch simply moves ++ after the check.
> 
> ...
>
> --- a/fs/exec.c
> +++ b/fs/exec.c
> @@ -434,8 +434,9 @@ static int count(struct user_arg_ptr argv, int max)
>  			if (IS_ERR(p))
>  				return -EFAULT;
>  
> -			if (i++ >= max)
> +			if (i >= max)
>  				return -E2BIG;
> +			++i;
>  
>  			if (fatal_signal_pending(current))
>  				return -ERESTARTNOHAND;

I have no problem working around a compiler bug when the workaround is
so small and simple.  For clarity and accuracy I renamed the patch to
"fs/exec.c: work around icc miscompilation".  

However I'd also like to be able to add "this bug has been reported to
the icc developers and will be fixed in version X.Y"?

Re: [PATCH RFC] exec: avoid possible undefined behavior in count()

[Xi Wang]

On 1/7/13 4:44 PM, Andrew Morton wrote:
> I have no problem working around a compiler bug when the workaround is
> so small and simple.  For clarity and accuracy I renamed the patch to
> "fs/exec.c: work around icc miscompilation".

Thanks!

> However I'd also like to be able to add "this bug has been reported to
> the icc developers and will be fixed in version X.Y"?

The icc developers have confirmed this bug and filed a defect ticket.
I'll let you know if there's any further update.

- xi