On 06/14/16 14:02, Borislav Petkov wrote:
> On Tue, Jun 14, 2016 at 01:54:25PM -0700, H. Peter Anvin wrote:
>> There was that. It is still possible that we end up with NOP a JMP
>> right before another JMP; we could perhaps make the patching code
>> smarter and see if we have a JMP immediately after.
>
> Yeah, I still can't get reproduce that reliably - I remember seeing it
> at some point but then dismissing it for another, higher-prio thing. And
> now the whole memory is hazy at best.
>
> But, you're giving me a great idea right now - I have this kernel
> disassembler tool which dumps alternative sections already and I could
> teach it to look for pathological cases around the patching sites and
> scream.
>
> Something for my TODO list when I get a quiet moment.
>
It's not really pathological; the issue is that asm goto() with an
unreachable clause after it doesn't tell gcc that a certain code path
ought to be linear, so we tell it to fall through. However, if gcc then
wants to have a jump there for whatever reason (perhaps it is part of a
loop) we end up with a redundant jump, so a patch site followed by a JMP
is entirely reasonable.
-hpa
