On Sat, Jul 11, 2015 at 10:35 AM, Robert Xiao <b...@cs.cmu.edu> wrote:
> On LP64 systems, reading a sysctl file containing an INT_MIN (-2147483648)
> could incorrectly show -18446744071562067968 due to an incorrect conversion
> in do_proc_dointvec_conv. This patch fixes the edge case by converting to
> unsigned int first to avoid sign extending INT_MIN to unsigned long.
>
> Test:
>
> root:/proc/sys/kernel# echo -2147483648 0 0 0 > printk
> root:/proc/sys/kernel# cat printk
>
> Without patch, produces -18446744071562067968 0 0 0.
> With patch, should produce -2147483648 0 0 0.
>
> Signed-off-by: Robert Xiao <b...@cs.cmu.edu>
> ---
> kernel/sysctl.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/sysctl.c b/kernel/sysctl.c
> index 19b62b5..464df36 100644
> --- a/kernel/sysctl.c
> +++ b/kernel/sysctl.c
> @@ -1995,10 +1995,10 @@ static int do_proc_dointvec_conv(bool *negp, unsigned
> long *lvalp,
> int val = *valp;
> if (val < 0) {
> *negp = true;
> - *lvalp = (unsigned long)-val;
> + *lvalp = (unsigned int)-val;
> } else {
> *negp = false;
> - *lvalp = (unsigned long)val;
> + *lvalp = (unsigned int)val;
> }
> }
> return 0;
I don't know why am I CC'ed on this - CC'ing Andrew along with Eric and
Kees who seem to have worked directly on sysctl.c not too long ago.
That said, I took a look at this and I think this patch is wrong.
Casting to unsigned int instead of unsigned long *after* the negation
is bogus, because we have
if (val < 0)
...
*lvalp = (unsigned long)-val;
and the compiler is free to assume -val to be positive and use the
sign-extend instruction. On gcc (GCC) 4.8.3 20140911 (Red Hat 4.8.3-7)
that I have here the cast to unsigned int works only with -O1, with -O2
it goes to town and uses cltq which sign-extends:
neg %eax
movb $0x1,(%rdi)
cltq
IMO the right way to do this would be to first cast to unsigned long
and then negate - that way we will first sign-extend and then negate an
unsigned, which is well defined. Also, this needs to be done not just
for do_proc_dointvec_conv(), but for do_proc_dointvec_minmax_conv() and
jiffies functions as well (although it's probably virtually impossible
to set val to exactly INT_MIN through jiffies write branches).
Speaking of write branches, only do_proc_dointvec_conv() does check
it's input properly, so that's something to look at.
Thanks,
Ilya
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
