Re: [PATCH] cpu/hotplug: Cache number of online CPUs

[Ingo Molnar]

* Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx> wrote:

> ----- On Jul 4, 2019, at 6:33 PM, Thomas Gleixner tglx@xxxxxxxxxxxxx wrote:
> 
> > On Thu, 4 Jul 2019, Mathieu Desnoyers wrote:
> >> ----- On Jul 4, 2019, at 5:10 PM, Thomas Gleixner tglx@xxxxxxxxxxxxx wrote:
> >> >
> >> > num_online_cpus() is racy today vs. CPU hotplug operations as
> >> > long as you don't hold the hotplug lock.
> >> 
> >> Fair point, AFAIU none of the loads performed within num_online_cpus()
> >> seem to rely on atomic nor volatile accesses. So not using a volatile
> >> access to load the cached value should not introduce any regression.
> >> 
> >> I'm concerned that some code may rely on re-fetching of the cached
> >> value between iterations of a loop. The lack of READ_ONCE() would
> >> let the compiler keep a lifted load within a register and never
> >> re-fetch, unless there is a cpu_relax() or a barrier() within the
> >> loop.
> > 
> > If someone really wants to write code which can handle concurrent CPU
> > hotplug operations and rely on that information, then it's probably better
> > to write out:
> > 
> >     ncpus = READ_ONCE(__num_online_cpus);
> > 
> > explicitely along with a big fat comment.
> > 
> > I can't figure out why one wants to do that and how it is supposed to work,
> > but my brain is in shutdown mode already :)
> > 
> > I'd rather write a proper kernel doc comment for num_online_cpus() which
> > explains what the constraints are instead of pretending that the READ_ONCE
> > in the inline has any meaning.
> 
> The other aspect I am concerned about is freedom given to the compiler 
> to perform the store to __num_online_cpus non-atomically, or the load 
> non-atomically due to memory pressure.

What connection does "memory pressure" have to what the compiler does? 

Did you confuse it with "register pressure"?

> Is that something we should be concerned about ?

Once I understand it :)

> I thought we had WRITE_ONCE and READ_ONCE to take care of that kind of 
> situation.

Store and load tearing is one of the minor properties of READ_ONCE() and 
WRITE_ONCE() - the main properties are the ordering guarantees.

Since __num_online_cpus is neither weirdly aligned nor is it written via 
constants I don't see how load/store tearing could occur. Can you outline 
such a scenario?

> The semantic I am looking for here is C11's relaxed atomics.

What does this mean?

Thanks,

	Ingo