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

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

[Betty Dall]

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);
 	}
 }
 #endif
-- 
1.7.7.6

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

[Shuah Khan]

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?

Thanks,
-- Shuah

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

[Alex Williamson]

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

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

[Alex Williamson]

On Mon, 2012-12-03 at 12:48 -0700, Alex Williamson wrote:
> 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.

My bad, pci_get_dev_by_id() does a put on the previous devices, keeping
the counts in order.  Thanks,

Alex

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

[Shuah Khan]

On Mon, Dec 3, 2012 at 1:00 PM, Alex Williamson
<alex.williamson@redhat.com> wrote:
> On Mon, 2012-12-03 at 12:48 -0700, Alex Williamson wrote:
>> 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.
>
> My bad, pci_get_dev_by_id() does a put on the previous devices, keeping
> the counts in order.  Thanks,
>
> Alex
>
Thanks Alex.

Betty,

You have my

Reviewed-by: Shuah Khan <shuah.khan@hp.com>

Also if you think this is applicable to stable as well, could you
please CC it for stable.

-- Shuah

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

[Bjorn Helgaas]

On Mon, Dec 3, 2012 at 1:59 PM, Shuah Khan <shuahkhan@gmail.com> wrote:
> On Mon, Dec 3, 2012 at 1:00 PM, Alex Williamson
> <alex.williamson@redhat.com> wrote:
>> On Mon, 2012-12-03 at 12:48 -0700, Alex Williamson wrote:
>>> 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.
>>
>> My bad, pci_get_dev_by_id() does a put on the previous devices, keeping
>> the counts in order.  Thanks,
>>
>> Alex
>>
> Thanks Alex.
>
> Betty,
>
> You have my
>
> Reviewed-by: Shuah Khan <shuah.khan@hp.com>
>
> Also if you think this is applicable to stable as well, could you
> please CC it for stable.

Thanks, Betty!  I'll merge this as a fix for v3.8.

Did you find this by inspection, or can you trigger some bad behavior?

Bjorn

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

[Shuah Khan]

On Mon, Dec 3, 2012 at 12:48 PM, Alex Williamson
<alex.williamson@redhat.com> wrote:
> 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()

Thanks. Didn't trace for_each_pci_dev()
>
> 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,
>

Probably several. I have seen only one or two cases pci_dev_put() is
called after pci_get_domain_bus_and_slot(). Did a quick search for
for_each_pci_dev() and found several instances that don't do
pci_dev_put()

-- Shuah