Hi Bjorn,

On 23.11.2016 00:13, Bjorn Helgaas wrote:

Hi Tomasz,

On Fri, Jun 10, 2016 at 09:55:19PM +0200, Tomasz Nowicki wrote:

quoted

Implement pci_acpi_scan_root and other arch-specific call so that ARM64
can start using ACPI to setup and enumerate PCI buses.

Prior to buses enumeration the pci_acpi_scan_root() implementation looks
for configuration space start address (obtained through ACPI _CBA method or
MCFG interface). If succeed, it uses ECAM library to create new mapping.
Then it attaches generic ECAM ops (pci_generic_ecam_ops) which are used
for accessing configuration space later on.
...

quoted

+static struct acpi_pci_root_ops acpi_pci_root_ops = {
+	.release_info = pci_acpi_generic_release_info,
+};
+
+/* Interface called from ACPI code to setup PCI host controller */
 struct pci_bus *pci_acpi_scan_root(struct acpi_pci_root *root)
 {
-	/* TODO: Should be revisited when implementing PCI on ACPI */
-	return NULL;
+	int node = acpi_get_node(root->device->handle);
+	struct acpi_pci_generic_root_info *ri;
+	struct pci_bus *bus, *child;
+
+	ri = kzalloc_node(sizeof(*ri), GFP_KERNEL, node);
+	if (!ri)
+		return NULL;
+
+	ri->cfg = pci_acpi_setup_ecam_mapping(root);
+	if (!ri->cfg) {
+		kfree(ri);
+		return NULL;
+	}
+
+	acpi_pci_root_ops.pci_ops = &ri->cfg->ops->pci_ops;

This has already been merged, but this isn't right, is it?  We're
writing a host controller-specific pointer into the single system-wide
acpi_pci_root_ops, then passing it on to acpi_pci_root_create().

Today, I think ri->cfg->ops->pci_ops is always &pci_generic_ecam_ops,
from this path:

  ri->cfg = pci_acpi_setup_ecam_mapping
    cfg = pci_ecam_create(..., &pci_generic_ecam_ops)
      cfg = kzalloc(...)
      cfg->ops = ops             # &pci_generic_ecam_ops

But we're about to merge the ECAM quirks series, which will mean it
may not be &pci_generic_ecam_ops.  Even apart from the ECAM quirks, we
should avoid this pattern of putting device-specific info in a single
shared structure because it's too difficult to verify that it's
correct.

Well spotted. I agree, we need to fix this. How about this:

diff --git a/arch/arm64/kernel/pci.c b/arch/arm64/kernel/pci.c
index fb439c7..31c0e1c 100644
--- a/arch/arm64/kernel/pci.c
+++ b/arch/arm64/kernel/pci.c

@@ -152,33 +152,35 @@ static void pci_acpi_generic_release_info(struct 

acpi_pci_root_info *ci)

         ri = container_of(ci, struct acpi_pci_generic_root_info, common);
         pci_ecam_free(ri->cfg);
+       kfree(ci->ops);
         kfree(ri);
  }

-static struct acpi_pci_root_ops acpi_pci_root_ops = {
-       .release_info = pci_acpi_generic_release_info,
-};
-
  /* Interface called from ACPI code to setup PCI host controller */
  struct pci_bus *pci_acpi_scan_root(struct acpi_pci_root *root)
  {
         int node = acpi_get_node(root->device->handle);
         struct acpi_pci_generic_root_info *ri;
         struct pci_bus *bus, *child;
+       struct acpi_pci_root_ops *root_ops;

         ri = kzalloc_node(sizeof(*ri), GFP_KERNEL, node);
         if (!ri)
                 return NULL;

+       root_ops = kzalloc_node(sizeof(*root_ops), GFP_KERNEL, node);
+       if (!root_ops)
+               return NULL;
+
         ri->cfg = pci_acpi_setup_ecam_mapping(root);
         if (!ri->cfg) {
                 kfree(ri);
+               kfree(root_ops);
                 return NULL;
         }

-       acpi_pci_root_ops.pci_ops = &ri->cfg->ops->pci_ops;
-       bus = acpi_pci_root_create(root, &acpi_pci_root_ops, &ri->common,
-                                  ri->cfg);
+       root_ops->release_info = pci_acpi_generic_release_info;
+       root_ops->pci_ops = &ri->cfg->ops->pci_ops;
+       bus = acpi_pci_root_create(root, root_ops, &ri->common, ri->cfg);
         if (!bus)
                 return NULL;

Of course, this should be the part of ECAM quirks core patches.

The other option we have is to remove "struct pci_ops *pci_ops;" from 
acpi_pci_root_ops structure and pass struct pci_ops as an extra argument 
to acpi_pci_root_create(). What do you think?

Thanks,
Tomasz