Re: [PATCH 0/7] preempt_count rework -v2

[Linus Torvalds]

On Tue, Sep 10, 2013 at 9:45 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> The usage site:
>
> #define preempt_enable() \
> do { \
>         barrier(); \
>         if (unlikely(preempt_count_dec_and_test())) \
>                 __preempt_schedule(); \
> } while (0)
>
> Already includes the barrier explicitly, so do we still need the memory
> clobber in that asm goto thing?

Yeah, you do to be safe, just to let gcc know that it may be changing
the memory location.

The issue is that because an "asm goto" cannot have outputs, I had to
change the (correct) "+m" input/output into just a "m" (input).

So without the memory clobber, gcc might decide that the thing doesn't
actually change the preempt count, and perhaps move a load of that
across the "asm goto".

Admittedly, that does sound almost impossibly unlikely, but I'd be
happier being careful.

> That said, your change results in:
>
> * ffffffff8106f45a:       65 ff 0c 25 e0 b7 00    decl   %gs:0xb7e0
>   ffffffff8106f461:       00
> * ffffffff8106f462:       74 0c                   je     ffffffff8106f470 <kick_process+0x50>
...
> Which is indeed perfect. So should I go 'fix' the other _and_test()
> functions we have to do this same thing?

It would be a good thing to test. There might be downsides with "asm
goto" (maybe it limits gcc some way), but it does in general save us
not only two instructions but also a register.

                 Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/