Re: [PATCH] PCI AER: pci_get_domain_bus_and_slot() call missing required pci_dev_put()

[Alex Williamson <alex.williamson@redhat.com>]

On Mon, 2012-12-03 at 10:40 -0700, Shuah Khan wrote:
> On Mon, Dec 3, 2012 at 9:37 AM, Betty Dall <betty.dall@hp.com> wrote:
> > The function aer_recover_queue() makes a call to pci_get_domain_bus_and_slot().
> > That function is documented to require that the caller decrement the
> > reference count by calling pci_dev_put(). This patch adds the missing
> > call to pci_dev_put().
> >
> > Signed-off-by: Betty Dall <betty.dall@hp.com>
> > ---
> >  drivers/pci/pcie/aer/aerdrv_core.c |    1 +
> >  1 files changed, 1 insertions(+), 0 deletions(-)
> >
> > diff --git a/drivers/pci/pcie/aer/aerdrv_core.c b/drivers/pci/pcie/aer/aerdrv_core.c
> > index af4e31c..43caf53 100644
> > --- a/drivers/pci/pcie/aer/aerdrv_core.c
> > +++ b/drivers/pci/pcie/aer/aerdrv_core.c
> > @@ -616,6 +616,7 @@ static void aer_recover_work_func(struct work_struct *work)
> >                         continue;
> >                 }
> >                 do_recovery(pdev, entry.severity);
> > +               pci_dev_put(pdev);
> 
> Are you sure pci_get_domain_bus_and_slot() does indeed increment the
> reference count? I see that in the comments right above the
> definition.
> 
> pci_dev_get() does that for sure. I see a call trace of
> 
> pci_dev_get() -> get_device() -> kobj_to_dev(kobject_get(&dev->kobj)  etc.
> 
> and kref_get() finally incrementing the
> 
> static inline void kref_get(struct kref *kref)
> {
>         WARN_ON(!atomic_read(&kref->refcount));
>         atomic_inc(&kref->refcount);
> }
> 
> However, I could not find any evidence that
> pci_get_domain_bus_and_slot() does that. I see
> pci_get_domain_bus_and_slot() calling pci_domain_nr() and there are
> several architecture specific defines for pci_domain_nr(). It is not
> clear to me where does refcount get incremented from
> pci_get_domain_bus_and_slot() is called.
> 
> What am I missing?

Trace back the for_each_pci_dev()

for_each_pci_dev
 -> pci_get_device
  -> pci_get_subsys
   -> pci_get_dev_by_id
    -> bus_find_device
     -> get_device()

The real question is how many places do we call for_each_pci_dev()
without a put.  It seems like pci_get_domain_bus_and_slot() doesn't even
get it right since it should include a put of the non-matched devices.
Thanks,

Alex