Hi,

first, some questions about the GART on your platform.

1) How many GARTs are usually implemented? One GART per device or is
   there a single GART covering all devices, or a mix of both?
2) Are devices allowed to DMA outside of the remappable range or will
   this fail?

Besides that I think IOMMU-API is not yet fully ready for GART-like
drivers like this one. But I will merge it anyway to get things moving.
But please answer or fix my objections first.

On Thu, Jan 05, 2012 at 09:11:48AM +0200, Hiroshi DOYU wrote:

+static void gart_iommu_domain_destroy(struct iommu_domain *domain)
+{
+	struct gart_device *gart = domain->priv;
+
+	spin_lock(&gart->client_lock);
+	if (!list_empty(&gart->client)) {
+		struct gart_client *c;
+
+		list_for_each_entry(c, &gart->client, list)
+			dev_err(gart->dev,
+				"%s is still attached\n", dev_name(c->dev));
+	}
+	spin_unlock(&gart->client_lock);

gart needs a NULL check. When you create a domain and immediatly destroy
it without ever attaching a device this code will dereference a NULL
pointer.
As a general improvement I suggest that you introduce a gart_domain
structure and store it in domain->priv (assigned in domain_init) instead
of using the hardware descriptor.

+static int gart_iommu_attach_dev(struct iommu_domain *domain,
+				 struct device *dev)
+{
+	struct gart_device *gart;
+	struct gart_client *client, *c;
+	int err = 0;
+
+	gart = dev_get_drvdata(dev->parent);
+	if (!gart)
+		return -EINVAL;
+	domain->priv = gart;

What happens when devices behind different GARTs are assigned to the
same domain? domain->priv can only hold a pointer to one hardware GART.
This can be solved by a 'struct gart_domain' holding a linked list to
all gart_devices in this domain.


	Joerg