Re: 2.6.18-rc5-mm1 + all hotfixes -- INFO: possible recursive locking detected

["Miles Lane" <miles.lane@gmail.com>]

On 9/6/06, Stefan Richter <stefanr@s5r6.in-berlin.de> wrote:
> I wrote:
> > I wrote:
> >> Or maybe it's older. Nodemgr takes class->subsys.rwsem and
> >> device.bus->subsys.rwsem. It always did. Could there be a change in
> >> driver core which makes this recursive? Or has it always been recursive?
> >> For example,
> >>
> >> static void nodemgr_update_pdrv(struct node_entry *ne)
> >> {
> >>      struct unit_directory *ud;
> >>      struct hpsb_protocol_driver *pdrv;
> >>      struct class *class = &nodemgr_ud_class;
> >>      struct class_device *cdev;
> >>
> >>      down_read(&class->subsys.rwsem);
> >>      list_for_each_entry(cdev, &class->children, node) {
>
> This may be wrong anyway. According to include/linux/device.h,
> class->sem should be used to protect access to class->children. There
> are more places in nodemgr of this sort.
>
> >>              ud = container_of(cdev, struct unit_directory, class_dev);
> >>              if (ud->ne != ne || !ud->device.driver)
> >>                      continue;
> >>
> >>              pdrv = container_of(ud->device.driver, struct hpsb_protocol_driver, driver);
> >>
> >>              if (pdrv->update && pdrv->update(ud)) {
> >>                      down_write(&ud->device.bus->subsys.rwsem);
> >>                      device_release_driver(&ud->device);
> >>                      up_write(&ud->device.bus->subsys.rwsem);
> >>              }
> >>      }
> >>      up_read(&class->subsys.rwsem);
> >> }
> >
> > Hi Greg,
> >
> > perhaps you could advise on this. It appears from grepping through the
> > sources that drivers/ieee1394/nodemgr.c is the only one with mixed
> > access to device.bus->subsys.rwsem and class->subsys.rwsem.
> >
> > Other usages of subsys.rwsem that I found are:
> > 1a.) dev->bus->subsys.rwsem
> > driver/ide/ide-proc.c and drivers/net/phy/phy_device.c take
> > dev->bus->subsys.rwsem. drivers/pnp/card.c takes dev.bus->subsys.rwsem.
> >
> > 1b.) driver.bus->subsys.rwsem
> > drivers/s390/net/qeth_proc.c takes driver.bus->subsys.rwsem.
> >
> > 2.) class->subsys.rwsem
> > drivers/scsi/hosts.c takes class->subsys.rwsem.
> >
> > 3.) bustype.subsys.rwsem
> > drivers/input/serio/serio.c takes serio_bus.subsys.rwsem.
> > drivers/input/gameport/gameport.c takes gameport_bus.subsys.rwsem.
> > drivers/base/power/shutdown.c takes devices_subsys.rwsem.
> > drivers/usb/core/devices.c and devio.c take usb_bus_type.subsys.rwsem.
> >
> > Do class->subsys.rwsem, bus->subsys.rwsem, and bus_type.subsys.rwsem
> > point to identical or different lock instances?
> >
> > Either way, could it hurt to convert nodemgr to uniformly use
> > ieee1394_bus_type.subsys.rwsem all over the place?

I don't have time to do the bisection testing.  If there is a patch
you'd like me to test against 2.6.18-rc5-mm1+all hotfixes, please let
me know.  I apologize for not being able to narrow this down further
for you.

        Miles