Re: [PATCH kernel 3/3] powerpc/iommu: Add iommu_ops to report capabilities and allow blocking domains

On 19/07/2022 04:09, Jason Gunthorpe wrote:

On Thu, Jul 14, 2022 at 06:18:22PM +1000, Alexey Kardashevskiy wrote:

quoted

+/*
+ * A simple iommu_ops to allow less cruft in generic VFIO code.
+ */
+static bool spapr_tce_iommu_capable(enum iommu_cap cap)
+{
+	switch (cap) {
+	case IOMMU_CAP_CACHE_COHERENCY:

I would add a remark here that it is because vfio is going to use
SPAPR mode but still checks that the iommu driver support coherency -
with out that detail it looks very strange to have caps without
implementing unmanaged domains

quoted

+static struct iommu_domain *spapr_tce_iommu_domain_alloc(unsigned int type)
+{
+	struct iommu_domain *dom;
+
+	if (type != IOMMU_DOMAIN_BLOCKED)
+		return NULL;
+
+	dom = kzalloc(sizeof(*dom), GFP_KERNEL);
+	if (!dom)
+		return NULL;
+
+	dom->geometry.aperture_start = 0;
+	dom->geometry.aperture_end = ~0ULL;
+	dom->geometry.force_aperture = true;

A blocked domain doesn't really have an aperture, all DMA is rejected,
so I think these can just be deleted and left at zero.

Generally I'm suggesting drivers just use a static singleton instance
for the blocked domain instead of the allocation like this, but that
is a very minor nit.

quoted

+static struct iommu_device *spapr_tce_iommu_probe_device(struct device *dev)
+{
+	struct pci_dev *pdev;
+	struct pci_controller *hose;
+
+	/* Weirdly iommu_device_register() assigns the same ops to all buses */
+	if (!dev_is_pci(dev))
+		return ERR_PTR(-EPERM);

Less "weirdly", more by design. The iommu driver should check if the
given struct device is supported or not, it isn't really a bus
specific operation.

quoted

+static struct iommu_group *spapr_tce_iommu_device_group(struct device *dev)
+{
+	struct pci_controller *hose;
+	struct pci_dev *pdev;
+
+	/* Weirdly iommu_device_register() assigns the same ops to all buses */
+	if (!dev_is_pci(dev))
+		return ERR_PTR(-EPERM);

This doesn't need repeating, if probe_device() fails then this will
never be called.

quoted

+static int spapr_tce_iommu_attach_dev(struct iommu_domain *dom,
+				      struct device *dev)
+{
+	struct iommu_group *grp = iommu_group_get(dev);
+	struct iommu_table_group *table_group;
+	int ret = -EINVAL;
+
+	if (!grp)
+		return -ENODEV;
+
+	table_group = iommu_group_get_iommudata(grp);
+
+	if (dom->type == IOMMU_DOMAIN_BLOCKED)
+		ret = table_group->ops->take_ownership(table_group);

Ideally there shouldn't be dom->type checks like this.


The blocking domain should have its own iommu_domain_ops that only
process the blocking operation. Ie call this like
spapr_tce_iommu_blocking_attach_dev()

Instead of having a "default_domain_ops" leave it NULL and create a
spapr_tce_blocking_domain_ops with these two functions and assign it
to domain->ops when creating. Then it is really clear these functions
are only called for the DOMAIN_BLOCKED type and you don't need to
check it.

quoted

+static void spapr_tce_iommu_detach_dev(struct iommu_domain *dom,
+				       struct device *dev)
+{
+	struct iommu_group *grp = iommu_group_get(dev);
+	struct iommu_table_group *table_group;
+
+	table_group = iommu_group_get_iommudata(grp);
+	WARN_ON(dom->type != IOMMU_DOMAIN_BLOCKED);
+	table_group->ops->release_ownership(table_group);
+}

Ditto

quoted

+struct iommu_group *pSeries_pci_device_group(struct pci_controller *hose,
+					     struct pci_dev *pdev)
+{
+	struct device_node *pdn, *dn = pdev->dev.of_node;
+	struct iommu_group *grp;
+	struct pci_dn *pci;
+
+	pdn = pci_dma_find(dn, NULL);
+	if (!pdn || !PCI_DN(pdn))
+		return ERR_PTR(-ENODEV);
+
+	pci = PCI_DN(pdn);
+	if (!pci->table_group)
+		return ERR_PTR(-ENODEV);
+
+	grp = pci->table_group->group;
+	if (!grp)
+		return ERR_PTR(-ENODEV);
+
+	return iommu_group_ref_get(grp);

Not for this series, but this is kind of backwards, the driver
specific data (ie the table_group) should be in
iommu_group_get_iommudata()...

It is there but here we are getting from a device to a group - a device 
is not added to a group yet when iommu_probe_device() works and tries 
adding a device via iommu_group_get_for_dev().

quoted

diff --git a/drivers/vfio/vfio_iommu_spapr_tce.c b/drivers/vfio/vfio_iommu_spapr_tce.c
index 8a65ea61744c..3b53b466e49b 100644
--- a/drivers/vfio/vfio_iommu_spapr_tce.c
+++ b/drivers/vfio/vfio_iommu_spapr_tce.c

@@ -1152,8 +1152,6 @@ static void tce_iommu_release_ownership(struct tce_container *container,
  	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
  		if (container->tables[i])
  			table_group->ops->unset_window(table_group, i);
-
-	table_group->ops->release_ownership(table_group);
  }
  
  static long tce_iommu_take_ownership(struct tce_container *container,

@@ -1161,10 +1159,6 @@ static long tce_iommu_take_ownership(struct tce_container *container,
  {
  	long i, ret = 0;
  
-	ret = table_group->ops->take_ownership(table_group);
-	if (ret)
-		return ret;
-
  	/* Set all windows to the new group */
  	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
  		struct iommu_table *tbl = container->tables[i];

@@ -1183,8 +1177,6 @@ static long tce_iommu_take_ownership(struct tce_container *container,
  	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
  		table_group->ops->unset_window(table_group, i);
  
-	table_group->ops->release_ownership(table_group);
-

This is great, makes alot of sense.

Anyhow, it all looks fine to me as is even:

Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>

Thanks. I'll try now to find an interested party to test this :)

Jason

-- 
Alexey