Re: [PATCH] Linux Kernel Markers 0.13 for 2.6.17

[Jeremy Fitzhardinge <jeremy@goop.org>]

Mathieu Desnoyers wrote:
> To protect code from being preempted, the macros preempt_disable and
> preempt_enable must normally be used. Logically, this macro must make sure gcc
> doesn't interleave preemptible code and non-preemptible code.
>   

No, it only needs to prevent globally visible side-effects from being 
moved into/out of preemptable blocks.  In practice that means memory 
updates (including the implicit ones that calls to external functions 
are assumed to make).

> Which makes me think that if I put barriers around my asm, call, asm trio, no
> other code will be interleaved. Is it right ?
>   

No global side effects, but code with local side effects could be moved 
around without changing the meaning of preempt.

For example:

	int foo;
	extern int global;

	foo = some_function();

	foo += 42;

	preempt_disable();
	// stuff
	preempt_enable();

	global = foo;
	foo += other_thing();

Assume here that some_function and other_function are extern, and so gcc 
has no insight into their behaviour and therefore conservatively assumes 
they have global side-effects.

The memory barriers in preempt_disable/enable will prevent gcc from 
moving any of the function calls into the non-preemptable region. But 
because "foo" is local and isn't visible to any other code, there's no 
reason why the "foo += 42" couldn't move into the preempt region.  
Likewise, the assignment to "global" can't move out of the range between 
the preempt_enable and the call to other_thing().

So in your case, if your equivalent of the non-preemptable block is the 
call to the marker function, then there's a good chance that the 
compiler might decide to move some other code in there.

Now it might be possible to take the addresses of labels to inhibit code 
motion into a particular range:

	{
		__label__ before, after;
		asm volatile("" : : "m" (*&&before), "m" (*&&after));	// gcc can't know what we're doing with the labels

	before:	;
		// stuff
	after:	;
	}

but that might be risky for several reasons: I don't know of any 
particular promises gcc makes in this circumstance; I suspect taking the 
address of a label will have a pretty severe inhibition on what 
optimisations gcc's is willing to use (it may prevent inlining 
altogether); and this looks pretty unusual, so there could be bugs.

    J