Re: [PATCH v4 12/15] iommu/vt-d: Cleanup get_valid_domain_for_dev()

[Alex Williamson <alex.williamson@redhat.com>]

On Fri, 2 Aug 2019 15:17:45 +0800
Lu Baolu <baolu.lu@linux.intel.com> wrote:

> Hi Alex,
> 
> Thanks for reporting this. I will try to find a machine with a
> pcie-to-pci bridge and get this issue fixed. I will update you
> later.

Further debug below...

> On 8/2/19 9:30 AM, Alex Williamson wrote:
> > DMAR: No ATSR found
> > DMAR: dmar0: Using Queued invalidation
> > DMAR: dmar1: Using Queued invalidation
> > pci 0000:00:00.0: DMAR: Software identity mapping
> > pci 0000:00:01.0: DMAR: Software identity mapping
> > pci 0000:00:02.0: DMAR: Software identity mapping
> > pci 0000:00:16.0: DMAR: Software identity mapping
> > pci 0000:00:1a.0: DMAR: Software identity mapping
> > pci 0000:00:1b.0: DMAR: Software identity mapping
> > pci 0000:00:1c.0: DMAR: Software identity mapping
> > pci 0000:00:1c.5: DMAR: Software identity mapping
> > pci 0000:00:1c.6: DMAR: Software identity mapping
> > pci 0000:00:1c.7: DMAR: Software identity mapping
> > pci 0000:00:1d.0: DMAR: Software identity mapping
> > pci 0000:00:1f.0: DMAR: Software identity mapping
> > pci 0000:00:1f.2: DMAR: Software identity mapping
> > pci 0000:00:1f.3: DMAR: Software identity mapping
> > pci 0000:01:00.0: DMAR: Software identity mapping
> > pci 0000:01:00.1: DMAR: Software identity mapping
> > pci 0000:03:00.0: DMAR: Software identity mapping
> > pci 0000:04:00.0: DMAR: Software identity mapping
> > DMAR: Setting RMRR:
> > pci 0000:00:02.0: DMAR: Setting identity map [0xbf800000 - 0xcf9fffff]
> > pci 0000:00:1a.0: DMAR: Setting identity map [0xbe8d1000 - 0xbe8dffff]
> > pci 0000:00:1d.0: DMAR: Setting identity map [0xbe8d1000 - 0xbe8dffff]
> > DMAR: Prepare 0-16MiB unity mapping for LPC
> > pci 0000:00:1f.0: DMAR: Setting identity map [0x0 - 0xffffff]
> > pci 0000:00:00.0: Adding to iommu group 0
> > pci 0000:00:00.0: Using iommu direct mapping
> > pci 0000:00:01.0: Adding to iommu group 1
> > pci 0000:00:01.0: Using iommu direct mapping
> > pci 0000:00:02.0: Adding to iommu group 2
> > pci 0000:00:02.0: Using iommu direct mapping
> > pci 0000:00:16.0: Adding to iommu group 3
> > pci 0000:00:16.0: Using iommu direct mapping
> > pci 0000:00:1a.0: Adding to iommu group 4
> > pci 0000:00:1a.0: Using iommu direct mapping
> > pci 0000:00:1b.0: Adding to iommu group 5
> > pci 0000:00:1b.0: Using iommu direct mapping
> > pci 0000:00:1c.0: Adding to iommu group 6
> > pci 0000:00:1c.0: Using iommu direct mapping
> > pci 0000:00:1c.5: Adding to iommu group 7
> > pci 0000:00:1c.5: Using iommu direct mapping
> > pci 0000:00:1c.6: Adding to iommu group 8
> > pci 0000:00:1c.6: Using iommu direct mapping
> > pci 0000:00:1c.7: Adding to iommu group 9

Note that group 9 contains device 00:1c.7

> > pci 0000:00:1c.7: Using iommu direct mapping

I'm booted with iommu=pt, so the domain type is IOMMU_DOMAIN_PT

> > pci 0000:00:1d.0: Adding to iommu group 10
> > pci 0000:00:1d.0: Using iommu direct mapping
> > pci 0000:00:1f.0: Adding to iommu group 11
> > pci 0000:00:1f.0: Using iommu direct mapping
> > pci 0000:00:1f.2: Adding to iommu group 11
> > pci 0000:00:1f.3: Adding to iommu group 11
> > pci 0000:01:00.0: Adding to iommu group 1
> > pci 0000:01:00.1: Adding to iommu group 1
> > pci 0000:03:00.0: Adding to iommu group 12
> > pci 0000:03:00.0: Using iommu direct mapping
> > pci 0000:04:00.0: Adding to iommu group 13
> > pci 0000:04:00.0: Using iommu direct mapping
> > pci 0000:05:00.0: Adding to iommu group 9

05:00.0 is downstream of 00:1c.7 and in the same group.  As above, the
domain is type IOMMU_DOMAIN_IDENTITY, so we take the following branch:

        } else {
                if (device_def_domain_type(dev) == IOMMU_DOMAIN_DMA) {

Default domain type is IOMMU_DOMAIN_DMA because of the code block in
device_def_domain_type() handling bridges to conventional PCI and
conventional PCI endpoints.

                        ret = iommu_request_dma_domain_for_dev(dev);

This fails in request_default_domain_for_dev() because there's more
than one device in the group.

                        if (ret) {
                                dmar_domain->flags |= DOMAIN_FLAG_LOSE_CHILDREN;
                                if (!get_private_domain_for_dev(dev)) {

With this commit, this now returns NULL because find_domain() does find
a domain, the same one we found before this code block.

                                        dev_warn(dev,
                                                 "Failed to get a private domain.\n");
                                        return -ENOMEM;
                                }

So the key factors are that I'm booting with iommu=pt and I have a
PCIe-to-PCI bridge grouped with its parent root port.  The bridge
wants an IOMMU_DOMAIN_DMA, but the group domain is already of type
IOMMU_DOMAIN_IDENTITY.  A temporary workaround is to not use
passthrough mode, but this is a regression versus previous kernels.
Thanks,

Alex