Re: [PATCH v7 3/5] irqdomain: Add irq_domain_{push,pop}_irq() functions.

[Marc Zyngier]

Hi David,

On 09/08/17 23:51, David Daney wrote:
> For an already existing irqdomain hierarchy, as might be obtained via
> a call to pci_enable_msix_range(), a PCI driver wishing to add an
> additional irqdomain to the hierarchy needs to be able to insert the
> irqdomain to that already initialized hierarchy.  Calling
> irq_domain_create_hierarchy() allows the new irqdomain to be created,
> but no existing code allows for initializing the associated irq_data.
> 
> Add a couple of helper functions (irq_domain_push_irq() and
> irq_domain_pop_irq()) to initialize the irq_data for the new
> irqdomain added to an existing hierarchy.
> 
> Signed-off-by: David Daney <david.daney@xxxxxxxxxx>
> ---
>  include/linux/irqdomain.h |   3 +
>  kernel/irq/irqdomain.c    | 178 ++++++++++++++++++++++++++++++++++++++++++++++
>  2 files changed, 181 insertions(+)
> 
> diff --git a/include/linux/irqdomain.h b/include/linux/irqdomain.h
> index cac77a5..2318f29 100644
> --- a/include/linux/irqdomain.h
> +++ b/include/linux/irqdomain.h
> @@ -460,6 +460,9 @@ extern void irq_domain_free_irqs_common(struct irq_domain *domain,
>  extern void irq_domain_free_irqs_top(struct irq_domain *domain,
>  				     unsigned int virq, unsigned int nr_irqs);
>  
> +extern int irq_domain_push_irq(struct irq_domain *domain, int virq, void *arg);
> +extern int irq_domain_pop_irq(struct irq_domain *domain, int virq);
> +
>  extern int irq_domain_alloc_irqs_parent(struct irq_domain *domain,
>  					unsigned int irq_base,
>  					unsigned int nr_irqs, void *arg);
> diff --git a/kernel/irq/irqdomain.c b/kernel/irq/irqdomain.c
> index f1f2514..629f770 100644
> --- a/kernel/irq/irqdomain.c
> +++ b/kernel/irq/irqdomain.c
> @@ -1448,6 +1448,184 @@ int __irq_domain_alloc_irqs(struct irq_domain *domain, int irq_base,
>  	return ret;
>  }
>  
> +/* The irq_data was moved, fix the revmap to refer to the new location */
> +static void irq_domain_fix_revmap(struct irq_data *d)
> +{
> +	void **slot;
> +
> +	if (d->hwirq < d->domain->revmap_size)
> +		return; /* Not using radix tree. */
> +
> +	/* Fix up the revmap. */
> +	mutex_lock(&revmap_trees_mutex);
> +	slot = radix_tree_lookup_slot(&d->domain->revmap_tree, d->hwirq);
> +	if (slot)
> +		radix_tree_replace_slot(&d->domain->revmap_tree, slot, d);

radix_tree_replace_slot already deals with non-existing entries, so the
initial radix_tree_lookup_slot call is superfluous.

> +	mutex_unlock(&revmap_trees_mutex);
> +}
> +
> +/**
> + * irq_domain_push_irq() - Push a domain in to the top of a hierarchy.
> + * @domain:	Domain to push.
> + * @virq:	Irq to push the domain in to.
> + * @arg:	Passed to the irq_domain_ops alloc() function.
> + *
> + * For an already existing irqdomain hierarchy, as might be obtained
> + * via a call to pci_enable_msix(), add an additional domain to the
> + * head of the processing chain.  Must be called before request_irq()
> + * has been called.
> + */
> +int irq_domain_push_irq(struct irq_domain *domain, int virq, void *arg)
> +{
> +	struct irq_data *child_irq_data;
> +	struct irq_data *root_irq_data = irq_get_irq_data(virq);
> +	struct irq_desc *desc;
> +	int rv = 0;
> +
> +	/*
> +	 * Check that no action has been set, which indicates the virq
> +	 * is in a state where this function doesn't have to deal with
> +	 * races between interrupt handling and maintaining the
> +	 * hierarchy.  This will catch gross misuse.  Attempting to
> +	 * make the check race free would require holding locks across
> +	 * calls to struct irq_domain_ops->alloc(), which could lead
> +	 * to deadlock, so we just do a simple check before starting.
> +	 */
> +	desc = irq_to_desc(virq);
> +	if (!desc)
> +		return -EINVAL;
> +	if (WARN_ON(desc->action))
> +		return -EBUSY;
> +
> +	if (domain == NULL)
> +		return -EINVAL;
> +
> +	if (WARN_ON(!domain->ops->alloc))
> +		return -EINVAL;

I'd rather you use irq_domain_is_hierarchy() instead. Same effect, but
less likely to break in the long run.

> +
> +	if (!root_irq_data)
> +		return -EINVAL;
> +
> +	child_irq_data = kzalloc_node(sizeof(*child_irq_data), GFP_KERNEL,
> +				      irq_data_get_node(root_irq_data));
> +	if (!child_irq_data)
> +		return -ENOMEM;
> +
> +	mutex_lock(&irq_domain_mutex);
> +
> +	/* Copy the original irq_data. */
> +	*child_irq_data = *root_irq_data;
> +
> +	irq_domain_fix_revmap(child_irq_data);

What is the benefit of updating the revmap early? We don't do that in
the pop case. Can't we do it in one go once the allocation has succeeded?

> +
> +	/*
> +	 * Overwrite the root_irq_data, which is embedded in struct
> +	 * irq_desc, with values for this domain.
> +	 */
> +	root_irq_data->parent_data = child_irq_data;
> +	root_irq_data->domain = domain;
> +	root_irq_data->mask = 0;
> +	root_irq_data->hwirq = 0;
> +	root_irq_data->chip = NULL;
> +	root_irq_data->chip_data = NULL;
> +	rv = domain->ops->alloc(domain, virq, 1, arg);

That'd be irq_domain_alloc_irqs_hierarchy().

Overall, I'm a bit concerned that alloc() is allowed to be recursive
itself. Hopefully nobody will do that, but you never know. A possible
way of trapping this would be to only set parent_data *after* the
allocation has been done.

Another concern is that I never see domain->parent being checked. It
should match child_irq_data->domain, so that you can never push a domain
on an interrupt that is not part of the parent domain.

> +	if (rv) {
> +		/* Restore the original irq_data. */
> +		*root_irq_data = *child_irq_data;
> +		irq_domain_fix_revmap(root_irq_data);
> +		goto error;
> +	}
> +
> +	if (root_irq_data->hwirq < domain->revmap_size) {
> +		domain->linear_revmap[root_irq_data->hwirq] = virq;
> +	} else {
> +		mutex_lock(&revmap_trees_mutex);
> +		radix_tree_insert(&domain->revmap_tree,
> +				  root_irq_data->hwirq, root_irq_data);
> +		mutex_unlock(&revmap_trees_mutex);
> +	}

We already have this exact code twice (in irq_domain_insert_irq and
irq_domain_associate). How about making it a helper?

> +error:
> +	mutex_unlock(&irq_domain_mutex);
> +
> +	return rv;
> +}
> +EXPORT_SYMBOL_GPL(irq_domain_push_irq);
> +
> +/**
> + * irq_domain_pop_irq() - Remove a domain from the top of a hierarchy.
> + * @domain:	Domain to remove.
> + * @virq:	Irq to remove the domain from.
> + *
> + * Undo the effects of a call to irq_domain_push_irq().  Must be
> + * called either before request_irq() or after free_irq().
> + */
> +int irq_domain_pop_irq(struct irq_domain *domain, int virq)
> +{
> +	struct irq_data *root_irq_data = irq_get_irq_data(virq);
> +	struct irq_data *child_irq_data;
> +	struct irq_data *tmp_irq_data;
> +	struct irq_desc *desc;
> +
> +	/*
> +	 * Check that no action is set, which indicates the virq is in
> +	 * a state where this function doesn't have to deal with races
> +	 * between interrupt handling and maintaining the hierarchy.
> +	 * This will catch gross misuse.  Attempting to make the check
> +	 * race free would require holding locks across calls to
> +	 * struct irq_domain_ops->free(), which could lead to
> +	 * deadlock, so we just do a simple check before starting.
> +	 */
> +	desc = irq_to_desc(virq);
> +	if (!desc)
> +		return -EINVAL;
> +	if (WARN_ON(desc->action))
> +		return -EBUSY;
> +
> +	if (domain == NULL)
> +		return -EINVAL;
> +
> +	if (!root_irq_data)
> +		return -EINVAL;
> +
> +	tmp_irq_data = irq_domain_get_irq_data(domain, virq);
> +
> +	/* We can only "pop" if this domain is at the top of the list */
> +	if (WARN_ON(root_irq_data != tmp_irq_data))
> +		return -EINVAL;
> +
> +	if (WARN_ON(root_irq_data->domain != domain))
> +		return -EINVAL;
> +
> +	child_irq_data = root_irq_data->parent_data;
> +	if (WARN_ON(!child_irq_data))
> +		return -EINVAL;
> +
> +	mutex_lock(&irq_domain_mutex);
> +
> +	root_irq_data->parent_data = NULL;
> +
> +	if (root_irq_data->hwirq >= domain->revmap_size) {
> +		mutex_lock(&revmap_trees_mutex);
> +		radix_tree_delete(&domain->revmap_tree, root_irq_data->hwirq);
> +		mutex_unlock(&revmap_trees_mutex);
> +	}

What about clearing it from the revmap if it fits there? Also, this code
already exists in irq_domain_disassociate and irq_domain_remove_irq, and
making that a helper is overdue.

> +
> +	if (domain->ops->free)
> +		domain->ops->free(domain, virq, 1);

Use irq_domain_free_irqs_hierarchy(), making it conditional in that helper.

> +
> +	/* Restore the original irq_data. */
> +	*root_irq_data = *child_irq_data;
> +
> +	irq_domain_fix_revmap(root_irq_data);
> +
> +	mutex_unlock(&irq_domain_mutex);
> +
> +	kfree(child_irq_data);
> +
> +	return 0;
> +}
> +EXPORT_SYMBOL_GPL(irq_domain_pop_irq);
> +
>  /**
>   * irq_domain_free_irqs - Free IRQ number and associated data structures
>   * @virq:	base IRQ number
> 

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...