On 08/15/2017 06:50 AM, Marc Zyngier wrote:
Hi David,
On 09/08/17 23:51, David Daney wrote:
[...]
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.
This comment I don't understand. To replace an element in the tree, you
must know the slot. I see no alternative to calling
radix_tree_lookup_slot(). If I am mistaken, it would be helpful to know
in a little more detail how you think it should be done.
+ 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.
Good idea, I will do it.
+
+ 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?
The idea was to have it be valid for the ops->alloc() call, but ...
We don't do that in
the pop case. Can't we do it in one go once the allocation has succeeded?
... the code in irqdomain.c that calls ops->alloc() doesn't have revmap
established at the call site, so I agree that this can go after the
allocation so we don't have to undo it on failure.
+
+ /*
+ * 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().
Yes.
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.
I will try it.
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.
I will add a check for this.
+ 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?
OK.
+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);ops->alloc()
+ }
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.
I will investigate doing something like that.
+
+ if (domain->ops->free)
+ domain->ops->free(domain, virq, 1);
Use irq_domain_free_irqs_hierarchy(), making it conditional in that helper.
OK
+
+ /* 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.
