Re: [PATCH, RESEND 1/2] dvb-frontends: fix i2c access helpers for KASAN

From: Arnd Bergmann
Date: Thu Nov 30 2017 - 09:06:22 EST


On Thu, Nov 30, 2017 at 1:49 PM, Mauro Carvalho Chehab
<mchehab@xxxxxxxxxx> wrote:
>> Link: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81715
>> Signed-off-by: Arnd Bergmann <arnd@xxxxxxxx>
>> ---
>> I'm undecided here whether there should be a comment pointing
>> to PR81715 for each file that the bogus local variable workaround
>> to prevent it from being cleaned up again. It's probably not
>> necessary since anything that causes actual problems would also
>> trigger a build warning.

>
> This kind of sucks, and it is completely unexpected... why val is
> so special that it would require this kind of hack?

It's explained in the gcc bug report: basically gcc always skipped
one optimization on inline function arguments that it does on
normal variables. Without KASAN and asan-stack, we didn't
notice because the impact was fairly small, but I ended up finally
getting to the bottom of it in September, and it finally got fixed.

I had an older version of the patch that was much more invasive
before we understood what exactly is happening, see
https://lkml.org/lkml/2017/3/2/484

> Also, there's always a risk of someone see it and decide to
> simplify the code, returning it to the previous state.
>
> So, if we're willing to do something like that, IMHO, we should have
> some macro that would document it, and fall back to the direct
> code if the compiler is not gcc 5, 6 or 7.

Older compilers are also affected and will produce better code
with my change, the difference is just smaller without asan-stack
(added ion gcc-5) is disabled, since that increases the stack
space used by each variable to (IIRC) 32 bytes.

The fixed gcc-8 produces identical code with and without my
change.

I don't think that a macro would help here at all, but if you
prefer, I could add a link to that gcc bug in each function that
has the problem.

Arnd