Re: [patch 2/3] scheduler: add full memory barriers upon task switch at runqueue lock/unlock

[Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>]

* Linus Torvalds (torvalds@linux-foundation.org) wrote:
> 
> 
> On Mon, 1 Feb 2010, Mathieu Desnoyers wrote:
> > 
> > However, this does not deal with mm_cpumask update, and we cannot use
> > the per-cpu rq lock, as it's a process-wide data structure updated with
> > clear_bit/set_bit in switch_mm(). So at the very least, we would have to
> > add memory barriers in switch_mm() on some architectures to deal with
> > this.
> 
> I'd much rather have a "switch_mm()" is a guaranteed memory barrier logic, 
> because quite frankly, I don't see how it ever couldn't be one anyway. It 
> fundamentally needs to do at least a TLB context switch (which may be just 
> switching an ASI around, not flushing the whole TLB, of course), and I bet 
> that for 99% of all architectures, that is already pretty much guaranteed 
> to be equivalent to a memory barrier.
> 
> It certainly is for x86. "mov to cr0" is serializing (setting any control 
> register except cr8 is serializing). And I strongly suspect other 
> architectures will be too.

What we have to be careful about here is that it's not enough to just
rely on switch_mm() containing a memory barrier. What we really need to
enforce is that switch_mm() issues memory barriers both _before_ and
_after_ mm_cpumask modification. The "after" part is usually dealt with
by the TLB context switch, but the "before" part usually isn't.

> 
> Btw, one reason to strongly prefer "switch_mm()" over any random context 
> switch is that at least it won't affect inter-thread (kernel or user-land) 
> switching, including switching to/from the idle thread.
> 
> So I'd be _much_ more open to a "let's guarantee that 'switch_mm()' always 
> implies a memory barrier" model than to playing clever games with 
> spinlocks.

If we really want to make this patch less intrusive, we can consider
iterating on each online cpu in sys_membarrier() rather than on the
mm_cpumask. But it comes at the cost of useless cache-line bouncing on
large machines with few threads running in the process, as we would grab
the rq locks one by one for all cpus.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68