Re: [RFC 2/2] kvm: eventfd: detect integer overflow using check_*_overflow

[Sasha Levin <sasha.levin@oracle.com>]

On 11/26/2014 01:23 PM, Linus Torvalds wrote:
> On Wed, Nov 26, 2014 at 9:55 AM, Sasha Levin <sasha.levin@oracle.com> wrote:
>> >
>> > Good point. Do you think there's an advantage in using GCC's overflow
>> > checker in this case?
> No. However, if your coccinelle script can be changed to verify that
> the type of the expression is unsigned, _that_ would be useful.

I'm pretty sure that this is something GCC will warn you about in the
compilation stage.

> And the "multiplication overflow" may actually be a way to generate
> better code. Possibly. I'm not entirely sure exactly what gcc actually
> does. How many multiplication overflow tests do we actually have,
> though?

Well, there are two straightforward checks in the kcalloc() family. They're
not the issue though. The problem is the *unchecked* *signed* integer overflows
lurking around.

kernel/time/ntp.c:process_adjtimex_modes():

	if (txc->modes & ADJ_FREQUENCY) {
                time_freq = txc->freq * PPM_SCALE;   <=== Undefined overflow
                time_freq = min(time_freq, MAXFREQ_SCALED);
                time_freq = max(time_freq, -MAXFREQ_SCALED);
                /* update pps_freq */
                pps_set_freq(time_freq);
        }

The multiplication is between signed integers, and it overflows (user triggerable).


Thanks,
Sasha