Re: [PATCH V6 04/10] PCI/DOE: Introduce pci_doe_create_doe_devices

[Dan Williams]

On Fri, Feb 4, 2022 at 8:28 AM Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
>
> On Fri, Feb 04, 2022 at 02:51:16PM +0000, Jonathan Cameron wrote:
> > On Thu, 3 Feb 2022 16:44:37 -0600
> > Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
> > > On Mon, Jan 31, 2022 at 11:19:46PM -0800, ira.weiny@xxxxxxxxx wrote:
>
> > > > + * pci_doe_create_doe_devices - Create auxiliary DOE devices for all DOE
> > > > + *                              mailboxes found
> > > > + * @pci_dev: The PCI device to scan for DOE mailboxes
> > > > + *
> > > > + * There is no coresponding destroy of these devices.  This function associates
> > > > + * the DOE auxiliary devices created with the pci_dev passed in.  That
> > > > + * association is device managed (devm_*) such that the DOE auxiliary device
> > > > + * lifetime is always greater than or equal to the lifetime of the pci_dev.
> > >
> > > This seems backwards.  What does it mean if the DOE aux dev
> > > lifetime is *greater* than that of the pci_dev?  Surely you can't
> > > access a PCI DOE Capability if the pci_dev is gone?
> >
> > I think the description is inaccurate - the end of life is the same
> > as that of the PCI driver binding to the pci_dev.  It'll get cleared
> > up if that is unbound etc.
>
> I don't know much about devm, but I *think* the devm things get
> released by devres_release_all(), which is called by
> __device_release_driver() after it calls the bus or driver's .remove()
> method (pci_device_remove(), in this case).
>
> So in this case, I think the aux dev is created after the pci_dev and
> released after the PCI driver and the PCI core are done with the
> pci_dev.  I assume some refcounting prevents the pci_dev from actually
> being deallocated until the aux dev is done with it.
>
> I'm not confident that this is a robust situation.

devm is a replacement for hand coding driver ->remove() handlers.
Anything devm allocated at ->probe() will be freed in the proper
reverse order by the driver core after it calls ->remove(). Ideally
for pure devm usage the ->remove() handler can be elided altogether.
I'll go read this patch to make sure it follows the expected pattern
which is:

1/ Parent device driver performs kmalloc(), device_initialize(), and
device_add() of a child device.
2/ Parent registers a devm handler for that child device that will
trigger device_unregister() at remove

During parent device unregister or unbind the devm action will
complete device_unregister() for all children first.

That process is independent of the device lifetime that can be
arbitrarily extended by 3rd party get_device() or
CONFIG_DEBUG_KOBJECT_RELEASE. The device core / kobject hierarchy
guarantees that the parent device is pinned until after child-device
final put event. I.e. final put_device() on a child also triggers a
put_device() on the parent paired with the get_device() taken on the
parent at device_add() time.