Re: [PATCH] pci: Account for virtual buses in pci_acs_path_enabled

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Sun, Aug 5, 2012 at 11:55 PM, Alex Williamson
<alex.williamson@xxxxxxxxxx> wrote:
> On Sun, 2012-08-05 at 23:30 -0600, Bjorn Helgaas wrote:
>> On Sat, Aug 4, 2012 at 12:19 PM, Alex Williamson
>> <alex.williamson@xxxxxxxxxx> wrote:
>> > It's possible to have buses without an associated bridge
>> > (bus->self == NULL).  SR-IOV can generate such buses.  When
>> > we find these, skip to the parent bus to look for the next
>> > ACS test.
>>
>> To make sure I understand the problem here, I think you're referring
>> to the situation where an SR-IOV device can span several bus numbers,
>> e.g., the "VFs Spanning Multiple Bus Numbers" implementation note in
>> the SR-IOV 1.1 spec, sec. 2.1.2.
>>
>> It says "All PFs must be located on the Device's captured Bus Number"
>> -- I think that means every PF will be directly on a bridge's
>> secondary bus and hence will have a valid dev->bus->self pointer.
>>
>> However, VFs need not be on the same bus number.  If a VF is on
>> (captured Bus Number plus 1), I think we allocate a new struct pci_bus
>> for it, but there's no P2P bridge that leads to that bus, so the
>> bus->self pointer is probably NULL.
>
> Yes, exactly.  virtfn_add_bus() is where we're creating this new bus.
>
>> This makes me quite nervous, because I bet there are many places that
>> assume every non-root bus has a valid bus->self pointer  -- I know I
>> certainly had that assumption.
>>
>> I looked at callers of pci_is_root_bus(), and at first glance, it seems like
>> iommu_init_device(), intel_iommu_add_device(), pci_acs_path_enabled(),
>
>
> These 3 are handled by this patch, plus the intel and amd iommu patches
> I sent.
>
>> pci_get_interrupt_pin(), pci_common_swizzle(),
>
> If sr-iov is the only source of these virtual buses, these are probably
> ok since VFs don't support INTx.
>
>> pci_find_upstream_pcie_bridge(), and
>
> Here the pci_is_root_bus() is after a pci_is_pcie() check, so again if
> sr-iov only (and assuming VFs properly report PCIe capability), we
> shouldn't stumble on it.
>
>> pci_bus_release_bridge_resources() all might have similar problems.
>
> This one might deserve further investigation.  Thanks,

We can fix all these places piecemeal, but that doesn't feel like a
very satisfying solution.  It makes it much harder to know that each
place is correct, and this oddity of a bus with no upstream bridge is
still lying around, waiting to bite us again later.

What other possible ways of fixing this do we have?  Could we set
bus->self (multiple buses would then point to the same bridge, and I
don't know if that would break something)?  Add something like a
pci_upstream_p2p_bridge() interface that would encapsulate traversing
the bus->parent and bus->self links?

Since these fake VF buses don't have a bridge that points to them, I
think the only place we keep a pointer to them is in the parent bus's
"children" list (updated in pci_add_new_bus()).  And now I'm confused
about when we should use bus->children and when we should use
bus->devices and why we should have both.

Does pci_walk_bus() work correctly with these VFs on fake buses?  It
doesn't use "children", so I can't see how it would ever find them.

Aren't you sorry you opened this can of worms?  :)

>> > Signed-off-by: Alex Williamson <alex.williamson@xxxxxxxxxx>
>> > ---
>> >
>> > David Ahern reported an oops from iommu drivers passing NULL into
>> > this function for the same mistake.  Harden this function against
>> > assuming bus->self is valid as well.  David, please include this
>> > patch as well as the iommu patches in your testing.
>> >
>> >  drivers/pci/pci.c |   22 +++++++++++++++++-----
>> >  1 file changed, 17 insertions(+), 5 deletions(-)
>> >
>> > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
>> > index f3ea977..e11a49c 100644
>> > --- a/drivers/pci/pci.c
>> > +++ b/drivers/pci/pci.c
>> > @@ -2486,18 +2486,30 @@ bool pci_acs_enabled(struct pci_dev *pdev, u16 acs_flags)
>> >  bool pci_acs_path_enabled(struct pci_dev *start,
>> >                           struct pci_dev *end, u16 acs_flags)
>> >  {
>> > -       struct pci_dev *pdev, *parent = start;
>> > +       struct pci_dev *pdev = start;
>> > +       struct pci_bus *bus;
>> >
>> >         do {
>> > -               pdev = parent;
>> > -
>> >                 if (!pci_acs_enabled(pdev, acs_flags))
>> >                         return false;
>> >
>> > -               if (pci_is_root_bus(pdev->bus))
>> > +               bus = pdev->bus;
>> > +
>> > +               if (pci_is_root_bus(bus))
>> >                         return (end == NULL);
>> >
>> > -               parent = pdev->bus->self;
>> > +               /*
>> > +                * Skip buses without an associated bridge.  In this
>> > +                * case move to the parent and continue.
>> > +                */
>> > +               while (!bus->self) {
>> > +                       if (!pci_is_root_bus(bus))
>> > +                               bus = bus->parent;
>> > +                       else
>> > +                               return (end == NULL);
>> > +               }
>> > +
>> > +               pdev = bus->self;
>> >         } while (pdev != end);
>> >
>> >         return true;
>> >
>
>
>
--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html


[Index of Archives]     [DMA Engine]     [Linux Coverity]     [Linux USB]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [Greybus]

  Powered by Linux