On Mon, 18 Jun 2007, Ingo Molnar wrote:
>
> ok. Do we have an guarantee that cpu_relax() is also an smp_rmb()?
The common use for cpu_relax() is basically for code that does
while (*ptr != val)
cpu_relax();
so yes, an architecture that doesn't notice writes by other CPU's on its
own had *better* have an implied read barrier in its "cpu_relax()"
implementation. For example, the irq handling code does
while (desc->status & IRQ_INPROGRESS)
cpu_relax();
which is explicitly about waiting for another CPU to get out of their
interrupt handler. And one classic use for it in drivers is obviously the
while (time_before (jiffies, next))
cpu_relax();
kind of setup (and "jiffies" may well be updated on another CPU: the fact
that it is "volatile" is just a *compiler* barrier just like cpu_relax()
itself will also be, not a "smp_rmb()" kind of hardware barrier).
So we could certainly add the smp_rmb() to make it more explicit, and it
wouldn't be *wrong*.
But quite frankly, I'd personally rather not - if it were to make a
difference in some situation, it would just be papering over a bug in
cpu_relax() itself.
The whole point of cpu_relax() is about busy-looping, after all. And the
only thing you really *can* busy-loop on in a CPU is basically a memory
value.
So the smp_rmb() would I think distract from the issue, and at best paper
over some totally separate bug.
> hm, this might still go into a non-nice busy loop on SMP: one cpu runs
> the strace, another one runs two tasks, one of which is runnable but not
> on the runqueue (the one we are waiting for). In that case we'd call
> yield() on this CPU in a loop
Sure. I agree - we can get into a loop that calls yield(). But I think a
loop that calls yield() had better be ok - we're explicitly giving the
scheduler the ability to to schedule anything else that is relevant.
So I think yield()'ing is fundamentally different from busy-looping any
other way.
Would it be better to be able to have a wait-queue, and actually *sleep*
on it, and not even busy-loop using yield? Yeah, possibly. I cannot
personally bring myself to care about that kind of corner-case situation,
though.
Linus
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
