On Thu, 2008-07-31 at 15:26 +0200, Sebastien Dugue wrote:

On Thu, 31 Jul 2008 23:01:39 +1000 Michael Ellerman [off-list ref] wrote:

quoted

On Thu, 2008-07-31 at 22:58 +1000, Michael Ellerman wrote:

quoted

On Thu, 2008-07-31 at 14:00 +0200, Sebastien Dugue wrote:

quoted

On Thu, 31 Jul 2008 21:40:56 +1000 Michael Ellerman [off-list ref] wrote:

quoted

This boot ordering stuff is pretty hairy, so I might have missed
something, but this is how the code is ordered AFAICT:

start_kernel()
	init_IRQ()
	...
	local_irq_enable()
	...
	rest_init()
		kernel_thread()
			kernel_init()
				smp_prepare_cpus()
					smp_xics_probe()	(via smp_ops->probe())


What's stopping us from taking an irq between local_irq_enable() and
smp_xics_probe() ?  Is it just that no one's request_irq()'ed them yet?

  It's hairy, I agree, but as you've mentioned no one has done a request_irq()
at that point. The first one to do it is smp_xics_probe() for the IPI.

Hmm, I don't think that's strong enough. I can trivially cause irqs to
fire during a kexec reboot just by mashing the keyboard.

And during a kdump boot all sorts of stuff could be firing. Even during
a clean boot, from firmware, I don't think we can guarantee that
nothing's going to fire.

.. after a bit of testing ..

It seems it actually works (sort of). 

xics_remap_irq() calls irq_radix_revmap_lookup(), which calls:

ptr = radix_tree_lookup(&host->revmap_data.tree, hwirq);

And because host->revmap_data.tree was zalloc'ed we trip on the first
check here:

@#$% ctrl-enter == send!

Continuing ...

void *radix_tree_lookup(struct radix_tree_root *root, unsigned long index)
{
        unsigned int height, shift;
        struct radix_tree_node *node, **slot;

        node = rcu_dereference(root->rnode);
        if (node == NULL)
                return NULL;

Which means irq_radix_revmap_lookup() will return NO_IRQ, which is cool.

  Which is what I intended so that as long as no IRQ is registered we
return NO_IRQ.

quoted

So I think it can fly, as long as we're happy that we can't reverse map
anything until smp_xics_probe() - and I think that's true, as any irq we
take will be invalid.

  That's true as no IRQs are registered before smp_xics_probe() and for any
interrupt we might get before that, irq_radix_revmap_lookup() will return
NO_IRQ.

Cool, we agree :) 

My only worry is that we might be relying on on the particular radix
tree implementation a bit too much. Is it documented somewhere that
the /very/ first check is for root->rnode != NULL, and the rest of the
root may be unintialised?

And I think it needs a big fat comment in the irq code saying that it's
safe because revmap_data is zalloc'ed, and that means the radix lookup
will fail (safely).

cheers

-- 
Michael Ellerman
OzLabs, IBM Australia Development Lab

wwweb: http://michael.ellerman.id.au
phone: +61 2 6212 1183 (tie line 70 21183)

We do not inherit the earth from our ancestors,
we borrow it from our children. - S.M.A.R.T Person