Re: [PATCH 1/3] msi: add forgotten pci_dev_put(pdev) to populate_msi_sysfs()

[Veaceslav Falico <vfalico@redhat.com>]

On Wed, Sep 25, 2013 at 03:30:14PM -0600, Bjorn Helgaas wrote:
>On Wed, Sep 25, 2013 at 3:08 PM, Bjorn Helgaas <bhelgaas@google.com> wrote:
>> Also, I think it is incorrect that free_msi_irqs() does this:
>>
>>                 if (entry->kobj.parent) {
>>                         kobject_del(&entry->kobj);
>>                         kobject_put(&entry->kobj);
>>                 }
>>
>>                 list_del(&entry->list);
>>                 kfree(entry);
>>
>> I think the "kfree(entry)" should be in msi_kobj_release() instead.
>
>Oh, I see you fixed this in patch 3/3.  I hadn't read that far yet :)
>
>Did you find these problems by inspection, or is there some easy way
>to trigger bad behavior?  Just wondering if this is some way I can
>reproduce the problem.

Hi,

I've first found it by building with CONFIG_DEBUG_KOBJECT_RELEASE and
CONFIG_DEBUG_OBJECTS - it shows that it's freeing an object in an active
state (I'm just running insmod/rmmod tg3 concurently, but I guess it's
reproducible with any driver that uses msi/x).

Without CONFIG_DEBUG_OBJECTS it's also reproducible, and without
CONFIG_DEBUG_KOBJECT_RELEASE it's really hard to reproduce, but still
reproducible (I've hit it with tg3 as a slave of bonding and concurently
running rmmod bonding/ifup bond0 - though it's really hard). It just panics
in kobject_put(), iirc.

So, with those CONFIG_DEBUG_* it's easily triggerable, without - quite
hard.

Hope that helps.

p.s. I'll adress your other comments tomorrow already, it's quite late here
and I don't want to do something stupid now :).

Thanks a lot!

>
>Bjorn