Re: [PATCH 8/8] i386: bitops: smp_mb__{before, after}_clear_bit() definitions

[Jeremy Fitzhardinge <jeremy@goop.org>]

Satyam Sharma wrote:
> Consider this (the above two functions exist only for clear_bit(),
> the atomic variant, as you already know), the _only_ memory reference
> we care about is that of the address of the passed bit-string:
>
> (1) The compiler must not optimize / elid it (i.e. we need to disallow
>     compiler optimization for that reference) -- but we've already taken
>     care of that with the __asm__ __volatile__ and the constraints on
>     the memory "addr" operand there, and,
> (2) For the i386, it also includes an implicit memory (CPU) barrier
>     already.
>
> So I /think/ it makes sense to let the compiler optimize _other_ memory
> references across the call to clear_bit(). There's a difference. I think
> we'd be safe even if we do this, because the synchronization in callers
> must be based upon the _passed bit-string_, otherwise _they_ are the
> ones who're buggy.
>
> [ However, elsewhere Jeremy Fitzhardinge mentioned the case of
>   some callers, for instance, doing a memset() on an alias of
>   the same bit-string. But again, I think that is dodgy/buggy/
>   extremely border-line usage on the caller's side itself ...
>   *unless* the caller is doing that inside a higher-level lock
>   anyway, in which case he wouldn't be needing to use the
>   locked variants either ... ]
>   

You miss my point.  If you have:

	memset(&my_bitmask, 0, sizeof(my_bitmask));
	set_bit(my_bitmask, 44);

Then unless the set_bit has a constraint argument which covers the whole
of the (multiword) bitmask, the compiler may see fit to interleave the
memset writes with the set_bit in bad ways.  In other words, plain "+m"
(*(long *)ptr) won't cut it.  You'd need "+m" (my_bitmask), I think.

    J