Re: [PATCH kernel v4 6/8] genirq/irqdomain: Move hierarchical IRQ cleanup to kobject_release

From: Thomas Gleixner
Date: Mon Nov 30 2020 - 18:14:43 EST


Alexey,

On Tue, Nov 24 2020 at 17:17, Alexey Kardashevskiy wrote:
> This moves hierarchical domain's irqs cleanup into the kobject release
> hook to make irq_domain_free_irqs() as simple as kobject_put.

Truly simple: Simply broken in multiple ways.

CONFIG_SPARSE_IRQ=n is now completely buggered. It does not even compile
anymore. Running core code changes through a larger set of cross
compilers is neither rocket science nor optional.

For CONFIG_SPARSE_IRQ=y, see below.

> @@ -1675,14 +1679,11 @@ void irq_domain_free_irqs(unsigned int virq, unsigned int nr_irqs)
> "NULL pointer, cannot free irq\n"))
> return;
>
> - mutex_lock(&irq_domain_mutex);
> - for (i = 0; i < nr_irqs; i++)
> - irq_domain_remove_irq(virq + i);
> - irq_domain_free_irqs_hierarchy(data->domain, virq, nr_irqs);
> - mutex_unlock(&irq_domain_mutex);
> + for (i = 0; i < nr_irqs; i++) {
> + struct irq_desc *desc = irq_to_desc(virq + i);
>
> - irq_domain_free_irq_data(virq, nr_irqs);
> - irq_free_descs(virq, nr_irqs);
> + kobject_put(&desc->kobj);

So up to this point both irq_dispose_mapping() _and_
irq_domain_free_irqs() invoked irq_free_descs().

Let's look at the call chains:

irq_domain_free_irqs()
irq_free_descs()
mutex_lock(&sparse_irq_lock);
for (i...)
free_desc(from + i)
irq_remove_debugfs_entry();
unregister_irq_proc();
irq_sysfs_del();
delete_irq_desc();
call_rcu();
bitmap_clear(allocated_irqs, ...);
mutex_unlock(&sparse_irq_lock);

with your modifications it does:

irq_domain_free_irqs()
for (i...)
kobject_put(&desc->kobj)
irq_kobj_release()
if (desc->free_irq)
desc->free_irq(desc);
irq_remove_debugfs_entry();
unregister_irq_proc();
delete_irq_desc();
call_rcu();

Can you spot the wreckage? It's not even subtle, it's more than obvious.

1) None of the operations in irq_kobj_release() is protected by
sparse_irq_lock anymore. There was a comment in free_desc() which
explained what is protected. You removed parts of that comment
and just left the sysfs portion of it above delete_irq_desc()
which is completely bogus because you removed the irq_sysfs_del()
call.

2) Nothing removes the freed interrupts from the allocation
bitmap. Run this often enough and you exhausted the interrupt
space.

And no, you cannot just go and invoke irq_free_descs() instead of
kobject_put(), simply because you'd create lock order inversion vs. the
free_irq() callback.

So no, it's not that simple and I'm not at all interested in another
respin of this with some more duct tape applied.

It can be done, but that needs way more thought, a proper design which
preserves the existing semantics completely and wants to be a fine
grained series where each patch does exactly ONE small thing which is
reviewable and testable on _ALL_ users of this code, i.e. _ALL_
architectures and irq chip implementations.

Thanks,

tglx