Hi Bjorn,

Am 06.04.2016 um 16:01 schrieb David Engraf:
> Hi Bjorn,
>
> Am 06.04.2016 um 15:35 schrieb Bjorn Helgaas:
>> Hi David,
>>
>> On Wed, Apr 06, 2016 at 01:51:55PM +0200, David Engraf wrote:
>>> Hi,
>>>
>>> I have an exception in __pci_bus_size_bridges() when pci_is_root_bus
>>> returns false but bus->self == NULL. My driver registers a virtual
>>> bus, like virtfn_add_bus(). pci_add_new_bus() is called with a
>>> parent but without a pci_dev. Thus bus->parent is set but bus->self
>>> is NULL. When __pci_bus_size_bridges() is called I get an exception
>>> at:
>>>
>>> switch (bus->self->class >> 8)
>>>
>>> The previous version of the code, checking for bus->self != NULL
>>> worked for me. I think an additional check is required to make sure
>>> we're not accessing a NULL pointer.
>>
>> When you say "previous version of the code," do you mean a previous
>> version of Linux worked correctly but a newer version does not?  What
>> versions are they?  Did you identify a commit that changed the
>> behavior?  I assume you're talking about an out-of-tree driver that
>> registers a virtual bus?  Can you point us to the code?  Can you share
>> the dmesg log, including the backtrace?
>
> With previous version, I mean reverting commit
> 2ba29e270e977b213a7d58ae1152c23a1c3074a3. This will check for bus->self
> instead of using pci_is_root_bus().
> I'm working on a driver which is basically based on drivers/pci/iov.c.
>
> The code looks like this:
>
> child = pci_add_new_bus(parent, NULL, busnr);
>
> This will create child with child->parent = parent and child->self =
> NULL. When the kernel calls __pci_bus_size_bridges(), accessing
> bus->self->class crashes:
>
> /* The root bus? */
> if (pci_is_root_bus(bus))
>      return;
>
> switch (bus->self->class >> 8) { <- crash here because of bus->self == NULL

With this patch the system no longer crashes and the virtual PCI bus 
works. It returns on bus->self == NULL, thus the switch instruction does 
not access a NULL pointer.

Best regards
- David