iommu/of: Handle PCI aliases properly
authorRobin Murphy <robin.murphy@arm.com>
Wed, 31 May 2017 17:52:29 +0000 (18:52 +0100)
committerJoerg Roedel <jroedel@suse.de>
Wed, 26 Jul 2017 09:11:57 +0000 (11:11 +0200)
When a PCI device has DMA quirks, we need to ensure that an upstream
IOMMU knows about all possible aliases, since the presence of a DMA
quirk does not preclude the device still also emitting transactions
(e.g. MSIs) on its 'real' RID. Similarly, the rules for bridge aliasing
are relatively complex, and some bridges may only take ownership of
transactions under particular transient circumstances, leading again to
multiple RIDs potentially being seen at the IOMMU for the given device.

Take all this into account in the OF code by translating every RID
produced by the alias walk, not just whichever one comes out last.
Happily, this also makes things tidy enough that we can reduce the
number of both total lines of code, and confusing levels of indirection,
by pulling the "iommus"/"iommu-map" parsing helpers back in-line again.

Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
drivers/iommu/of_iommu.c

index 8cb60829a7a1d8754017e100a010ed85f8e568c9..be8ac1ddec0656494fb2e5f9aa25bed67ce32f7b 100644 (file)
@@ -140,75 +140,39 @@ static const struct iommu_ops
        return ops;
 }
 
-static int __get_pci_rid(struct pci_dev *pdev, u16 alias, void *data)
-{
-       struct of_phandle_args *iommu_spec = data;
-
-       iommu_spec->args[0] = alias;
-       return iommu_spec->np == pdev->bus->dev.of_node;
-}
+struct of_pci_iommu_alias_info {
+       struct device *dev;
+       struct device_node *np;
+};
 
-static const struct iommu_ops
-*of_pci_iommu_init(struct pci_dev *pdev, struct device_node *bridge_np)
+static int of_pci_iommu_init(struct pci_dev *pdev, u16 alias, void *data)
 {
+       struct of_pci_iommu_alias_info *info = data;
        const struct iommu_ops *ops;
-       struct of_phandle_args iommu_spec;
+       struct of_phandle_args iommu_spec = { .args_count = 1 };
        int err;
 
-       /*
-        * Start by tracing the RID alias down the PCI topology as
-        * far as the host bridge whose OF node we have...
-        * (we're not even attempting to handle multi-alias devices yet)
-        */
-       iommu_spec.args_count = 1;
-       iommu_spec.np = bridge_np;
-       pci_for_each_dma_alias(pdev, __get_pci_rid, &iommu_spec);
-       /*
-        * ...then find out what that becomes once it escapes the PCI
-        * bus into the system beyond, and which IOMMU it ends up at.
-        */
-       iommu_spec.np = NULL;
-       err = of_pci_map_rid(bridge_np, iommu_spec.args[0], "iommu-map",
+       err = of_pci_map_rid(info->np, alias, "iommu-map",
                             "iommu-map-mask", &iommu_spec.np,
                             iommu_spec.args);
        if (err)
-               return err == -ENODEV ? NULL : ERR_PTR(err);
-
-       ops = of_iommu_xlate(&pdev->dev, &iommu_spec);
+               return err == -ENODEV ? 1 : err;
 
+       ops = of_iommu_xlate(info->dev, &iommu_spec);
        of_node_put(iommu_spec.np);
-       return ops;
-}
-
-static const struct iommu_ops
-*of_platform_iommu_init(struct device *dev, struct device_node *np)
-{
-       struct of_phandle_args iommu_spec;
-       const struct iommu_ops *ops = NULL;
-       int idx = 0;
 
-       /*
-        * We don't currently walk up the tree looking for a parent IOMMU.
-        * See the `Notes:' section of
-        * Documentation/devicetree/bindings/iommu/iommu.txt
-        */
-       while (!of_parse_phandle_with_args(np, "iommus", "#iommu-cells",
-                                          idx, &iommu_spec)) {
-               ops = of_iommu_xlate(dev, &iommu_spec);
-               of_node_put(iommu_spec.np);
-               idx++;
-               if (IS_ERR_OR_NULL(ops))
-                       break;
-       }
+       if (IS_ERR(ops))
+               return PTR_ERR(ops);
 
-       return ops;
+       return info->np == pdev->bus->dev.of_node;
 }
 
 const struct iommu_ops *of_iommu_configure(struct device *dev,
                                           struct device_node *master_np)
 {
-       const struct iommu_ops *ops;
+       const struct iommu_ops *ops = NULL;
        struct iommu_fwspec *fwspec = dev->iommu_fwspec;
+       int err;
 
        if (!master_np)
                return NULL;
@@ -221,18 +185,44 @@ const struct iommu_ops *of_iommu_configure(struct device *dev,
                iommu_fwspec_free(dev);
        }
 
-       if (dev_is_pci(dev))
-               ops = of_pci_iommu_init(to_pci_dev(dev), master_np);
-       else
-               ops = of_platform_iommu_init(dev, master_np);
+       /*
+        * We don't currently walk up the tree looking for a parent IOMMU.
+        * See the `Notes:' section of
+        * Documentation/devicetree/bindings/iommu/iommu.txt
+        */
+       if (dev_is_pci(dev)) {
+               struct of_pci_iommu_alias_info info = {
+                       .dev = dev,
+                       .np = master_np,
+               };
+
+               err = pci_for_each_dma_alias(to_pci_dev(dev),
+                                            of_pci_iommu_init, &info);
+               if (err) /* err > 0 means the walk stopped, but non-fatally */
+                       ops = ERR_PTR(min(err, 0));
+               else /* success implies both fwspec and ops are now valid */
+                       ops = dev->iommu_fwspec->ops;
+       } else {
+               struct of_phandle_args iommu_spec;
+               int idx = 0;
+
+               while (!of_parse_phandle_with_args(master_np, "iommus",
+                                                  "#iommu-cells",
+                                                  idx, &iommu_spec)) {
+                       ops = of_iommu_xlate(dev, &iommu_spec);
+                       of_node_put(iommu_spec.np);
+                       idx++;
+                       if (IS_ERR_OR_NULL(ops))
+                               break;
+               }
+       }
        /*
         * If we have reason to believe the IOMMU driver missed the initial
         * add_device callback for dev, replay it to get things in order.
         */
        if (!IS_ERR_OR_NULL(ops) && ops->add_device &&
            dev->bus && !dev->iommu_group) {
-               int err = ops->add_device(dev);
-
+               err = ops->add_device(dev);
                if (err)
                        ops = ERR_PTR(err);
        }