Re: [PATCH] driver core: Make it safe to use get_device() if the reference count is zero

[Jason Yan]

On 2017/12/14 16:10, Greg KH wrote:
> On Thu, Dec 14, 2017 at 03:56:46PM +0800, Jason Yan wrote:
> > On 2017/12/14 15:42, Greg KH wrote:
> > > On Thu, Dec 14, 2017 at 11:39:36AM +0800, Jason Yan wrote:
> > > > Some driviers may have the chance to increase a reference count that
> > > > has dropped to zero when using get_device() because of their design.
> > > Then those drivers are broken :)
> > >
> > > > We have met such a issue with scsi:
> > > > https://www.spinics.net/lists/linux-scsi/msg115295.html
> > > >
> > > > The scsi core will keep the scsi device object in the host list after
> > > > it has been deleted and the iterator can still find it. All of the
> > > > places where need iterating have to check the state of the scsi device
> > > > and this makes a lot of code redundancy and complexity.
> > > >
> > > > Provide a safe mechanism in get_device() by using
> > > > kobject_get_unless_zero().
> > > >
> > > > Suggested-by: Bart Van Assche <bart.vanassche@xxxxxxx>
> > > > Signed-off-by: Jason Yan <yanaijie@xxxxxxxxxx>
> > > > CC: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
> > > > CC: Bart Van Assche <bart.vanassche@xxxxxxx>
> > > > CC: Ewan D. Milne <emilne@xxxxxxxxxx>
> > > > CC: James E.J. Bottomley <jejb@xxxxxxxxxxxxxxxxxx>
> > > > CC: Christoph Hellwig <hch@xxxxxx>
> > > > ---
> > > >    drivers/base/core.c | 2 +-
> > > >    1 file changed, 1 insertion(+), 1 deletion(-)
> > > >
> > > > diff --git a/drivers/base/core.c b/drivers/base/core.c
> > > > index 12ebd05..cc74810 100644
> > > > --- a/drivers/base/core.c
> > > > +++ b/drivers/base/core.c
> > > > @@ -1916,7 +1916,7 @@ EXPORT_SYMBOL_GPL(device_register);
> > > >     */
> > > >    struct device *get_device(struct device *dev)
> > > >    {
> > > > -	return dev ? kobj_to_dev(kobject_get(&dev->kobj)) : NULL;
> > > > +	return dev && kobject_get_unless_zero(&dev->kobj) ? dev : NULL;
> > > I really don't want to do this, the bus the device is on should prevent
> > > this from happening.
> > >
> > > Also, once that reference count drops to zero, the memory should be
> > > freed, so you really have a stale pointer here, and this code would fail
> > > if you had slab debugging enabled anyway.
> >
> > Actually I don't want this either. But the design of scsi core will leave
> > the scsi device on the host list after it is deleted, and it can  be
> > found later and the refcount have a very big chance to increase from
> > zero again. And after a lot of discussion it seems that the scsi layer
> > is difficult to change the situation in the near future.
>
> Keeping a 'struct device' reference counted chunk of memory on a list
> that has a different lifetime rule from that device itself, is crazy.

The lifetime rule is the same. That device itself will delete from the
host list in the destructor, scsi_device_dev_release_usercontext(), and
the memory will be freed after that. That's why this issue came out.

> And yes, I remember how all of this came about, but I really don't have
> the time to work on it myself...
>
> > > So I don't even think this fixes the issue you think it fixes :)
> >
> > This issue is very easy to reproduce on my machine and I have tested the
> > patch and it really fixes the issue.
>
> Even with slab debugging enabled?  If so, what is keeping that memory
> from being freed once the reference count drops to 0?

As above, the memory is not freed yet when we increasing the refconunt 
from zero, so it's nothing to do with slab debugging enabled or not. If
we want to free it, we have to grab host lock first to delete it from
the list, so if we are grabing the host lock, we can increase the
refcount safely from 0 to 1.

> I think you are just papering over the real issue here, which is one
> reason I really do not like the get_unless_zero() function at all.
>
> thanks,
>
> greg k-h
>
> .