On 08/10/2013 09:43 AM, Linus Torvalds wrote:
> On Sat, Aug 10, 2013 at 9:09 AM, H. Peter Anvin <hpa@zytor.com> wrote:
>>
>> Do you have any quantification of "munches throughput?"  It seems odd
>> that it would be worse than polling for preempt all over the kernel, but
>> perhaps the additional locking is what costs.
> 
> Actually, the big thing for true preemption is not so much the preempt
> count itself, but the fact that when the preempt count goes back to
> zero we have that "check if we should have been preempted" thing.
> 
> And in particular, the conditional function call that goes along with it.
> 
> The thing is, even if that is almost never taken, just the fact that
> there is a conditional function call very often makes code generation
> *much* worse. A function that is a leaf function with no stack frame
> with no preemption often turns into a non-leaf function with stackcheck
> frames when you enable preemption, just because it had a RCU read
> region which disabled preemption.
> 
> It's similar to the kind of code generation issue that Andi's patches
> are trying to work on.
> 
> Andi did the "test and jump to a different section to call the
> scheduler with registers saved" as an assembly stub in one of his
> patches in this series exactly to avoid the cost of this for the
> might_sleep() case, and generated that GET_THREAD_AND_SCHEDULE asm
> macro for it. But look at that asm macro, and compare it to
> "preempt_check_resched()"..
> 
> I have often wanted to have access to that kind of thing from C code.
> It's not unusual. Think lock failure paths, not Tom Jones.
> 

Hmm... if that is really the big issue then I'm wondering if
preempt_enable() &c shouldn't be rewritten in assembly... if nothing
else to get the outbound call out of view of the C compiler; it could
even be turned into an exception instruction.

There are a few other things that one have to wonder about: the
preempt_count is currently located in the thread_info structure, but
since by definition we can't switch a thread that is preemption-locked
it should work in a percpu variable as well.

We could then play a really ugly stunt by marking NEED_RESCHED by adding
0x7fffffff to the counter.  Then the whole sequence becomes something like:

	subl $1,%fs:preempt_count
	jno 1f
	call __naked_preempt_schedule	/* Or a trap */
1:

For architectures with conditional traps the trapping option becomes
even more attractive.

	-hpa