Re: [PATCH]: Use stop_machine_run in the Intel RNG driver

[Andrew Morton <akpm@linux-foundation.org>]

On Tue, 27 Feb 2007 07:22:00 -0500 Prarit Bhargava <prarit@redhat.com> wrote:

> Replace call_smp_function with stop_machine_run in the Intel RNG driver.
> 
> CPU A has done read_lock(&lock)
> CPU B has done write_lock_irq(&lock) and is waiting for A to release the lock.
> 
> A third CPU calls call_smp_function and issues the IPI.  CPU A takes CPU C's
> IPI.  CPU B is waiting with interrupts disabled and does not see the IPI.
> CPU C is stuck waiting for CPU B to respond to the IPI.
> 
> Deadlock.

I think what you're describing here is just the standard old
smp_call_function() deadlock, rather than anything which is specific to
intel-rng, yes?

It is "well known" that you can't call smp_call_function() with local
interrupts disabled.  In fact i386 will spit a warning if you try it.


intel-rng doesn't do that, but what it _does_ do is:

	smp_call_function(..., wait = 0);
	local_irq_disable();

so some CPUs will still be entering the IPI while this CPU has gone and
disabled interrupts, thus exposing us to the deadlock, yes?

In which case a suitable fix might be to make intel-rng spin until all the
other CPUs have entered intel_init_wait().

> The solution is to use stop_machine_run instead of call_smp_function
> (call_smp_function should not be called in situations where the CPUs may
> be suspended).

But that seems to be a nice change anyway.  It took rather a lot of code
churn to do it, and it does find it necessary to export stop_machine_run()
to modules, but that seems OK too.