On 12/20/2012 07:12 PM, Oleg Nesterov wrote:

On 12/20, Srivatsa S. Bhat wrote:

quoted

On 12/20/2012 12:44 AM, Oleg Nesterov wrote:

quoted

We need 2 helpers for writer, the 1st one does synchronize_sched() and the
2nd one takes rwlock. A generic percpu_write_lock() simply calls them both.

Ah, that's the problem no? Users of reader-writer locks expect to run in
atomic context (ie., they don't want to sleep).

Ah, I misunderstood.

Sure, percpu_write_lock() should be might_sleep(), and this is not
symmetric to percpu_read_lock().

quoted

We can't expose an API that
can make the task go to sleep under the covers!

Why? Just this should be documented. However I would not worry until we
find another user. Until then we do not even need to add percpu_write_lock
or try to generalize this code too much.

Hmm.. But considering the disable_nonboot_cpus() case you mentioned below, I'm
only getting farther away from using synchronize_sched() ;-) And that also makes
it easier to expose a generic percpu rwlock API, like Tejun was suggesting.
So I'll give it a shot.

quoted

quoted

To me, the main question is: can we use synchronize_sched() in cpu_down?
It is slow.

Haha :-) So we don't want smp_mb() in the reader,

We need mb() + rmb(). Plust cli/sti unless this arch has optimized
this_cpu_add() like x86 (as you pointed out).

quoted

*and* also don't want
synchronize_sched() in the writer! Sounds like saying we want to have the cake
and eat it too ;-) :P

Personally I'd vote for synchronize_sched() but I am not sure. And I do
not really understand the problem space.

quoted

And moreover, since I'm still not convinced about the writer API part if use
synchronize_sched(), I'd rather avoid synchronize_sched().)

Understand.

And yes, synchronize_sched() adds more problems. For example, where should
we call it? I do not this _cpu_down() should do this, in this case, say,
disable_nonboot_cpus() needs num_online_cpus() synchronize_sched's.

Ouch! I should have seen that coming!

So probably cpu_down() should call it before cpu_maps_update_begin(), this
makes the locking even less obvious.

True.

In short. What I am trying to say is, don't ask me I do not know ;)

OK then, I'll go with what I believe is a reasonably good way (not necessarily
the best way) to deal with this:

I'll avoid the use of synchronize_sched(), expose a decent-looking percpu
rwlock implementation, use it in CPU hotplug and get rid of stop_machine().
That would certainly be a good starting base, IMHO.

Regards,
Srivatsa S. Bhat