Re: [PATCH v2 1/3] kcov: support comparison operands collection

[Dmitry Vyukov]

On Mon, Oct 9, 2017 at 8:37 PM, Mark Rutland <mark.rutland@xxxxxxx> wrote:
> On Mon, Oct 09, 2017 at 08:15:10PM +0200, 'Dmitry Vyukov' via syzkaller wrote:
>> On Mon, Oct 9, 2017 at 5:46 PM, Mark Rutland <mark.rutland@xxxxxxx> wrote:
>> > On Mon, Oct 09, 2017 at 05:05:19PM +0200, Alexander Potapenko wrote:
>
>> > ... I note that a few places in the kernel use a 128-bit type. Are
>> > 128-bit comparisons not instrumented?
>>
>> Yes, they are not instrumented.
>> How many are there? Can you give some examples?
>
> From a quick scan, it doesn't looks like there are currently any
> comparisons.
>
> It's used as a data type in a few places under arm64:
>
> arch/arm64/include/asm/checksum.h:      __uint128_t tmp;
> arch/arm64/include/asm/checksum.h:      tmp = *(const __uint128_t *)iph;
> arch/arm64/include/asm/fpsimd.h:                        __uint128_t vregs[32];
> arch/arm64/include/uapi/asm/ptrace.h:   __uint128_t     vregs[32];
> arch/arm64/include/uapi/asm/sigcontext.h:       __uint128_t vregs[32];
> arch/arm64/kernel/signal32.c:   __uint128_t     raw;
> arch/arm64/kvm/guest.c: __uint128_t tmp;


Then I think we just continue ignoring them for now :)
In the future we can extend kcov to trace 128-bits values. We will
need to add a special flag and write 2 consecutive entries for them.
Or something along these lines.


>> >> +     area = t->kcov_area;
>> >> +     /* The first 64-bit word is the number of subsequent PCs. */
>> >> +     pos = READ_ONCE(area[0]) + 1;
>> >> +     if (likely(pos < t->kcov_size)) {
>> >> +             area[pos] = ip;
>> >> +             WRITE_ONCE(area[0], pos);
>> >
>> > Not a new problem, but if the area for one thread is mmap'd, and read by
>> > another thread, these two writes could be seen out-of-order, since we
>> > don't have an smp_wmb() between them.
>> >
>> > I guess Syzkaller doesn't read the mmap'd kcov file from another thread?
>>
>>
>> Yes, that's the intention. If you read coverage from another thread,
>> you can't know coverage from what exactly you read. So the usage
>> pattern is:
>>
>> reset coverage;
>> do something;
>> read coverage;
>
> Ok. I guess without a use-case for reading this from another thread it doesn't
> really matter.
>
> Thanks,
> Mark.