Re: [3.11.4] Thunderbolt/PCI unplug oops in pci_pme_list_scan

[Yijing Wang <wangyijing@huawei.com>]

On 2013/11/20 9:20, Bjorn Helgaas wrote:
> On Tue, Nov 19, 2013 at 6:14 PM, Yijing Wang <wangyijing@huawei.com> wrote:
>>>>> [bhelgaas: changelog, tag for stable]
>>>>> Reported-by: David Bulkow <David.Bulkow@stratus.com>
>>>>> Reported-by: Mika Westerberg <mika.westerberg@linux.intel.com>
>>>>> Signed-off-by: Yinghai Lu <yinghai@kernel.org>
>>>>> Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
>>>>> CC: stable@vger.kernel.org  # v2.6.32+
>>>>
>>>> Hi Bjorn,
>>>>    This issue in X86 seems to be introduced after commit 928bea9 "PCI: Delay enabling bridges until they're needed"
>>>> So this patch needs to back port to 2.6.32+ ?
>>>
>>> 928bea9 might have made it more visible, but the underlying problem is that
>>> we enable the device once in the probe path, and disable it twice in the
>>> remove path.  That problem exists in 2.6.32.61:
>>>
>>>   pcie_portdrv_probe                # .probe() method
>>>     pcie_port_device_register
>>>       pci_enable_device             <-- enable
>>>
>>>   pcie_portdrv_remove               # .remove() method
>>>     pcie_port_device_remove
>>>       pci_disable_device            <-- disable #1
>>>     pci_disable_device              <-- disable #2
>>
>> During assign unassigned resources, we also enable the port device,
>>
>> fs_initcall(pcibios_assign_resources);
>>   pci_assign_unassigned_resources;
>>     pci_enable_bridges()
>>       pci_enable_device()
>>
>>
>> So I think before the commit 928bea9 , the pci bridge device enable and disable is symmetrical.
>> After the commit 928bea9, we only enable bridge once, but still remove twice.
> 
> The port driver should be symmetrical, regardless of what happens
> outside it.  We have to be able to bind/unbind/bind/unbind
> indefinitely.
> 
> Do you think the patch is a problem for current upstream, or are you
> just saying it doesn't need to be backported as far as 2.6.32?  I
> frankly don't care that much if those old kernels pick it up or not.
> All I'm saying is that the problem this fixes is present that far
> back.
> 
> I'm not really interested in doing a lot more digging about ancient
> kernels, unless you think it's going to break something if applied to
> them.

As you said, this is not a big problem, it's ok for current upstream,
and for 3.4 3.5 3.10 stable tree, it seems just a enable and disable symmetry problem.
I don't find anything unsafe about this even though when we unbind pcie port driver,
the pcie port maybe still enable. I have no objection to backport it to ancient kernels. :)

Thanks!
Yijing.


> 
> Bjorn
> 
> 


-- 
Thanks!
Yijing