Skip to contentSkip to conversation

pci: crash on pci_unregister_driver after pci_register_driver fails

RFC In review Stephen Hemminger 2 participants 3 messages Sep 3, 2013 – Sep 4, 2013 PCI

lore
Share & tools
lore
#0 lore raw

RFC 1/1 pci: crash on pci_unregister_driver after pci_register_driver fails

While debugging another problem with a PCI driver, I noticed that if device probe routine returns an error, the kernel will crash when module is unloaded. It looks like pci_register_driver() sets drv->bus to be PCI then in the module unload.

module_unload
my_device_exit_module
pci_unregister_driver
bus_remove_driver
OOPS

One way to fix this would be to have pci_register_driver clear the bus flag (it has no reference) if an error was detected.

drivers/pci/pci-driver.c +7 -1
@@ -1120,6 +1120,8 @@ const struct dev_pm_ops pci_dev_pm_ops =
 int __pci_register_driver(struct pci_driver *drv, struct module *owner,
 			  const char *mod_name)
 {
+	int err;
+
 	/* initialize common driver fields */
 	drv->driver.name = drv->name;
 	drv->driver.bus = &pci_bus_type;
@@ -1130,7 +1132,11 @@ int __pci_register_driver(struct pci_dri
 	INIT_LIST_HEAD(&drv->dynids.list);
 
 	/* register with core */
-	return driver_register(&drv->driver);
+	err = driver_register(&drv->driver);
+	if (err)
+		drv->driver.bus = NULL;
+
+	return err;
 }
 
 /**
#1 lore raw
Quoted 5 lines · Stephen Hemminger

While debugging another problem with a PCI driver, I noticed that if device probe routine returns an error, the kernel will crash when module is unloaded. It looks like pci_register_driver() sets drv->bus to be PCI then in the module unload.

Quoted 8 lines

module_unload
my_device_exit_module
pci_unregister_driver
bus_remove_driver
OOPS

One way to fix this would be to have pci_register_driver clear the bus flag (it has no reference) if an error was detected.

Odd, no other buses do this, why would it matter for PCI? Maybe they just never fail their registering calls.

How about putting this in the driver core instead so that all busses get fixed?

thanks,

greg k-h

#2 lore raw
Quoted 18 lines · On Tue, 3 Sep 2013 16:06:19 -0700 Greg KH <[email protected]> wrote:
Quoted 5 lines · Stephen Hemminger

While debugging another problem with a PCI driver, I noticed that if device probe routine returns an error, the kernel will crash when module is unloaded. It looks like pci_register_driver() sets drv->bus to be PCI then in the module unload.

Quoted 8 lines

module_unload
my_device_exit_module
pci_unregister_driver
bus_remove_driver
OOPS

One way to fix this would be to have pci_register_driver clear the bus flag (it has no reference) if an error was detected.

Odd, no other buses do this, why would it matter for PCI? Maybe they just never fail their registering calls.

Quoted 6 lines

How about putting this in the driver core instead so that all busses get fixed?

thanks,

greg k-h

Never mind, buggy driver