Re: [patch] cpufreq: mark cpufreq_tsc() as core_initcall_sync

[Oleg Nesterov <oleg@tv-sign.ru>]

On 11/29, Paul E. McKenney wrote:
>
> On Wed, Nov 29, 2006 at 11:16:46PM +0300, Oleg Nesterov wrote:
> > 
> > Hmm... SRCU can't be used from irq, yes. But I think that both versions
> > (spinlock needs _irqsave) can ?
> 
> I didn't think you could call wait_event() from irq.

Ah, sorry for confusion, I talked only about read lock/unlock of course.

Just in case, it is not safe to do srcu_read_{,un}lock() from irq,

	per_cpu_ptr(sp->per_cpu_ref, smp_processor_id())->c[idx]++
	                                                  ^^^^^^^^
we need local_t for that.

> For the locked version, you would also need spin_lock_irqsave() or some
> such to avoid self-deadlock.
> 
> For the atomic version, the fact that synchronize_qrcu() increments
> the new counter before decrmenting the old one should mean that calls
> to qrcu_read_lock() and qrcu_read_unlock() can be called from irq.

Yes, exactly! There is another reason, suppose we did

	qp->completed++;
	atomic_inc(qp->ctr + (idx ^ 0x1));

In that case the reader could be stalled if synchronize_qrcu() takes a
preemption in between.

> But synchronize_qrcu() must be called from process context, since it
> can block.

Surely.

Oleg.