]> asedeno.scripts.mit.edu Git - linux.git/commitdiff
vfio/type1: Give hardware MSI regions precedence
authorRobin Murphy <robin.murphy@arm.com>
Thu, 10 Aug 2017 19:11:50 +0000 (13:11 -0600)
committerAlex Williamson <alex.williamson@redhat.com>
Thu, 10 Aug 2017 19:11:50 +0000 (13:11 -0600)
If the IOMMU driver advertises 'real' reserved regions for MSIs, but
still includes the software-managed region as well, we are currently
blind to the former and will configure the IOMMU domain to map MSIs into
the latter, which is unlikely to work as expected.

Since it would take a ridiculous hardware topology for both regions to
be valid (which would be rather difficult to support in general), we
should be safe to assume that the presence of any hardware regions makes
the software region irrelevant. However, the IOMMU driver might still
advertise the software region by default, particularly if the hardware
regions are filled in elsewhere by generic code, so it might not be fair
for VFIO to be super-strict about not mixing them. To that end, make
vfio_iommu_has_sw_msi() robust against the presence of both region types
at once, so that we end up doing what is almost certainly right, rather
than what is almost certainly wrong.

Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Tested-by: Shameer Kolothum <shameerali.kolothum.thodi@huawei.com>
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
drivers/vfio/vfio_iommu_type1.c

index 2328be628f21e24fe4153ac67a877568031b7a39..92155cce926d1c38bdfc5fb415b43ade481d1591 100644 (file)
@@ -1169,13 +1169,21 @@ static bool vfio_iommu_has_sw_msi(struct iommu_group *group, phys_addr_t *base)
        INIT_LIST_HEAD(&group_resv_regions);
        iommu_get_group_resv_regions(group, &group_resv_regions);
        list_for_each_entry(region, &group_resv_regions, list) {
+               /*
+                * The presence of any 'real' MSI regions should take
+                * precedence over the software-managed one if the
+                * IOMMU driver happens to advertise both types.
+                */
+               if (region->type == IOMMU_RESV_MSI) {
+                       ret = false;
+                       break;
+               }
+
                if (region->type == IOMMU_RESV_SW_MSI) {
                        *base = region->start;
                        ret = true;
-                       goto out;
                }
        }
-out:
        list_for_each_entry_safe(region, next, &group_resv_regions, list)
                kfree(region);
        return ret;