possible unintended integer truncation in fs/ext4/extents.c:get_implied_cluster_alloc

possible unintended integer truncation in fs/ext4/extents.c:get_implied_cluster_alloc

[PaX Team]

Hello folks,

while running a simple analyzer plugin on linux 3.12.5 written by Emese Revfy
we found a case in ext4 that looks like a potential problem. the code looks
like this:

4082 		map->m_pblk = (ee_start & ~(sbi->s_cluster_ratio - 1)) +
4083 			c_offset;

here the expression ~(sbi->s_cluster_ratio - 1) will first do the negation
on an unsigned int then extend the result to unsigned long long (i.e, there's
a 32->64 bit conversion on both 32 and 64 bit archs) and stores it as such.
now this will obviously lose the higher 32 bits of ee_start and the question
is: is this intended behaviour or a bug? later the code compares map->m_pblk
against ee_block which is as unsigned int only so there's some mixture of
integer types here that may warrant further review.

cheers,
 PaX Team

Re: possible unintended integer truncation in fs/ext4/extents.c:get_implied_cluster_alloc

[Theodore Ts'o]

On Sat, Dec 14, 2013 at 11:29:18PM +0100, PaX Team wrote:
> Hello folks,
> 
> while running a simple analyzer plugin on linux 3.12.5 written by Emese Revfy
> we found a case in ext4 that looks like a potential problem. the code looks
> like this:
> 
> 4082 		map->m_pblk = (ee_start & ~(sbi->s_cluster_ratio - 1)) +
> 4083 			c_offset;
> 
> here the expression ~(sbi->s_cluster_ratio - 1) will first do the negation
> on an unsigned int then extend the result to unsigned long long (i.e, there's
> a 32->64 bit conversion on both 32 and 64 bit archs) and stores it as such.
> now this will obviously lose the higher 32 bits of ee_start and the question
> is: is this intended behaviour or a bug? later the code compares map->m_pblk
> against ee_block which is as unsigned int only so there's some mixture of
> integer types here that may warrant further review.

So C's integer promotion and sign extension rules are very obscure and
confusing, and that may be a reason why we should put in an explicit
cast, but it looks like the right thing is happening here.  Here's a
test program --- what am I missing?

					- Ted

#include <stdio.h>

typedef unsigned long long ext4_fsblk_t;


/* Mask out the the low */
#define EXT4_PHYS_CMASK(cr, pblk) (pblk & \
				    ~((ext4_fsblk_t) cr - 1))

int main(int argc, char **argv)
{

	ext4_fsblk_t	b, pblk;
	unsigned short	cr = 32;
	

	pblk = 0x123456789ABC;
	printf("%llx\n", pblk);

	b = ~(cr - 1);
	printf("%llx\n", b);

	b = (pblk & ~(cr - 1));
	printf("%llx\n", b);

	b = EXT4_PHYS_CMASK(cr, pblk);
	printf("%llx\n", b);

	return 0;
}

Re: possible unintended integer truncation in fs/ext4/extents.c:get_implied_cluster_alloc

[Theodore Ts'o]

On Thu, Dec 19, 2013 at 11:59:20PM -0500, Theodore Ts'o wrote:
> So C's integer promotion and sign extension rules are very obscure and
> confusing, and that may be a reason why we should put in an explicit
> cast, but it looks like the right thing is happening here.  Here's a
> test program --- what am I missing?

Never mind, I see the problem.  It works if cr is a short, but not if
it is an unsigned int.  Sigh...

 					- Ted