Re: workqueue deadlock

[Andrew Morton <akpm@osdl.org>]

On Sun, 10 Dec 2006 09:26:16 +0100
Ingo Molnar <mingo@elte.hu> wrote:

> 
> * Andrew Morton <akpm@osdl.org> wrote:
> 
> > > > Could do, not sure.  I'm planning on converting all the locking 
> > > > around here to preempt_disable() though.
> > > 
> > > please at least use an owner-recursive per-CPU lock,
> > 
> > a wot?
> 
> something like the pseudocode further below - when applied to a data 
> structure it has semantics and scalability close to that of 
> preempt_disable(), but it is still preemptible and the lock is specific.
> 
> > > not a naked preempt_disable()! The concurrency rules for data 
> > > structures changed via preempt_disable() are quite hard to sort out 
> > > after the fact. (preempt_disable() is too opaque,
> > 
> > preempt_disable() is the preferred way of holding off cpu hotplug.
> 
> well, preempt_disable() is the scheduler's internal mechanism to keep 
> tasks from being preempted. It is fast but it also has non-nice 
> side-effect:
> 
> 1) nothing tells us what the connection between preempt-disable and data 
>    structure is. If we let preempt_disable() spread then we'll end up 
>    with a situation like the BKL: all preempt_disable() sections become 
>    one big blob of code with hard-to-define specifications, and if we 
>    take out code from that blob stuff mysteriously breaks.

Well we can add some suitably-named wrapper around preempt_disable() to make
it obvious why we're calling it.  But I haven't noticed any such problem with
existing usages.

> 	void cpu_hotplug_lock(void)
> 	{
> 		int cpu = raw_smp_processor_id();
> 		/*
> 		 * Interrupts/softirqs are hotplug-safe:
> 		 */
> 		if (in_interrupt())
> 			return;
> 		if (current->hotplug_depth++)
> 			return;
> 		current->hotplug_lock = &per_cpu(hotplug_lock, cpu);
> 		mutex_lock(current->hotplug_lock);
> 	}

That's functionally equivalent to what we have now, and it isn't working
too well.