Re: [PATCH 1/1] IOMMU: Save pci device id instead of pci_dev* pointer for DMAR devices

[David Woodhouse <dwmw2@infradead.org>]

[-- Attachment #1: Type: text/plain, Size: 2118 bytes --]

On Fri, 2013-11-08 at 08:46 -0700, Bjorn Helgaas wrote:
> 
> I don't know the IOMMU drivers well either, but it seems like they
> rely on notifications of device addition and removal (see
> iommu_bus_notifier()).  It doesn't seem right for them to also use the
> generic PCI interfaces like pci_get_domain_bus_and_slot() because the
> IOMMU driver should already know what devices exist and their
> lifetimes.  It seems like confusion to mix the two.  But I don't have
> a concrete suggestion.

The generic IOMMU code has a notifier, and calls through to an
->add_device() method in the specific IOMMU driver's iommu_ops.

The Intel IOMMU driver predates that, and its scheme for mapping devices
to the correct DMAR unit is different. It happens entirely within the
get_domain_for_dev() function, which happens when we're first asked to
set up a mapping for a given device (when we don't already have the
answer stashed in dev->archdata).

I think we should add an ->add_device() method to the Intel IOMMU
driver, and make it do much of what's in get_domain_for_dev() right now
— finding the "proxy" device (the upstream PCIe bridge or whatever), and
then looking through the ACPI DMAR table to find which DMAR unit that's
attached to. Then we stash that information (dmar, devfn) in
dev->archdata, and get_domain_for_dev() still has *some* work to do,
actually allocating a logical domain on the IOMMU in question, but not
as much. And refcount the damn domain instead of playing the horrid
tricks we currently do to hang it off the upstream proxy device *too*.

My main concern here is that the DMAR table contains the PCI bus numbers
at boot time. Doing the lookup later will only work if we don't renumber
busses. Or if we have a way to look things up based on the *original*
bus number.

The Intel IOMMU also has a bus notifier of its own which it only uses to
know when a driver is *detached*, so it can tear down the logical domain
for the corresponding device. Would be nice to have the generic IOMMU
notifier call a callback for us then too, perhaps.


-- 
dwmw2


[-- Attachment #2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5745 bytes --]