On 10/08/16 10:44, Mauricio Faria de Oliveira wrote:
This patch employs the refcount in struct pci_controller to track
the references from PCI devices and buses (struct pci_dev/pci_bus).
In order to do that without modifying any PCI scan/probe approach
(e.g., PCI_PROBE_DEVTREE and PCI_PROBE_NORMAL), it leverages some
of the PCI arch-specific callback: pci_(add|release)_device() and
pci_(add|remove)_bus().
(a small change is required for PCI_PROBE_DEVTREE, which makes it
consistent with PCI_PROBE_NORMAL - the pci_dev should inherit the
parent pci_bus's phb pointer - see pci_setup_device() in probe.c)
This also has the advantage that locking for kref_(get|put)() is
satisfied by the 'pci_rescan_remove_lock' mutex, which is normal
practice for usage of the PCI subsystem - thus already in place.
More details added in comment on pcibios_release_device().
Signed-off-by: Mauricio Faria de Oliveira <mauri...@linux.vnet.ibm.com>
void pcibios_release_device(struct pci_dev *dev)
{
struct pci_controller *phb = pci_bus_to_host(dev->bus);
+ pr_debug("PCI %s, pci_dev %p, phb %p\n", dev_name(&dev->dev), dev, phb);
+
eeh_remove_device(dev);
if (phb->controller_ops.release_device)
phb->controller_ops.release_device(dev);
+
+ if (unlikely(!phb))
+ pr_warn("%s: PCI device %s has null PHB; refcount bug!",
+ __func__, dev_name(&dev->dev)); /* WARN_ON ahead */
This should probably go before trying to dereference phb->controller_ops
above?
--
Andrew Donnellan OzLabs, ADL Canberra
andrew.donnel...@au1.ibm.com IBM Australia Limited
