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:
> > 
> > Here is the detailed execution scenario showing the race.
> 
> No. You've added random smp_mb() calls, but you don't actually show what 
> the f*ck they are protecting against.
> 
> For example
> 
> > First sys_membarrier smp_mb():
> 
> I'm not AT ALL interested in the sys_membarrier() parts. You can hav ea 
> million memory barriers there, and I won't care. I'm interested in what 
> you think the memory barriers elsewhere protect against. It's a barrier 
> between _which_ two operations? 
> 
> You can't say it's a barrier "around" the
> 
> 	cpumask_clear(mm_cpumask, cpu);
> 
> because a barrier is between two things. So if you want to add two 
> barriers around that mm_cpumask acces, you need to describe the _three_ 
> events you're barriers between in that call-path (with mm_cpumask being 
> just one of them)
> 
> And then, once you've described _those_ three events, you describe what 
> the sys_membarrier interaction is, and how mm_cpumask is involved there.
> 
> I'm not interested in the user-space code. Don't even quote it. It's 
> irrelevant apart from the actual semantics you want to guarantee for the 
> new membarrier() system call. So don't quote the code, just explain what 
> the actual barriers are.
> 

The two event pairs we are looking at are:

Pair 1)

* memory accesses (load/stores) performed by user-space thread before
  context switch.
* cpumask_clear_cpu(cpu, mm_cpumask(prev));

Pair 2)

* cpumask_set_cpu(cpu, mm_cpumask(next));
* memory accessses (load/stores) performed by user-space thread after
  context switch.

I can see two ways to add memory barriers in switch_mm that would
provide ordering for these two memory access pairs:

Either A)

switch_mm()
  smp_mb__before_clear_bit();
  cpumask_clear_cpu(cpu, mm_cpumask(prev));
  cpumask_set_cpu(cpu, mm_cpumask(next));
  smp_mb__after_set_bit();

or B)

switch_mm()
  cpumask_set_cpu(cpu, mm_cpumask(next));
  smp_mb__before_clear_bit();
  cpumask_clear_cpu(cpu, mm_cpumask(prev));

(B) seems like a clear win, as we get the ordering right for both pairs
with a single memory barrier, but I don't know if changing the set/clear
bit order could have nasty side-effects on other mm_cpumask users.

sys_membarrier uses the mm_cpumask to iterate on all CPUs on which the
current process's mm is in use, so it can issue a smp_mb() through an
IPI on all CPUs that need it. Without appropriate ordering of pairs 1-2
detailed above, we could miss a CPU that actually needs a memory
barrier.

Thanks,

Mathieu

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