svn commit: r296909 - head/sys/ofed/drivers/infiniband/ulp/ipoib

[Gleb Smirnoff]

On Tue, Mar 15, 2016 at 10:43:53AM -0700, John Baldwin wrote:
J> > Log:
J> >   Fix witness panic in the ipoib_ioctl() function when unloading the
J> >   ipoib module.
J> >   
J> >   The bpfdetach() function is trying to turn off promiscious mode on the
J> >   network interface it is attached to while holding a mutex. The fix
J> >   consists of ignoring any further calls to the ipoib_ioctl() function
J> >   when the network interface is going to be detached. The ipoib_ioctl()
J> >   function might sleep.
J> >   
J> >   Sponsored by:	Mellanox Technologies
J> >   MFC after:	1 week
J> 
J> In all of the other NIC drivers I have worked on the fix has been to call
J> if_detach (or ether_ifdetach) as the first step in your detach method before
J> you try to stop the interface.  The idea is to remove external connections
J> from your driver first, and the only after those have drained do you actually
J> shut down the hardware.  Given you are calling if_detach() just before
J> if_free() below, ipoib just seems to be broken here.
J> 
J> For example, detach in a typical NIC driver does this:
J> 
J> 	struct foo_softc *sc;
J> 
J> 	sc = device_get_softc(dev);
J> 	ether_ifdetach(sc->sc_ifp);
J> 	FOO_LOCK(sc);
J> 	foo_stop(sc);
J> 	FOO_UNLOCK(sc);
J> 	callout_drain(&sc->timer);
J> 	bus_teardown_intr(...);
J> 	bus_release_resource(...);
J> 	if_free(sc->sc_ifp);
J> 
J> Similarly, devices with a character device in /dev should be calling
J> destroy_dev() first before shutting down hardware, etc.

In the new KPI in the projects/ifnet branch, I am doing the following:

- there is only one if_detach(), that does both detach and free
- if_detach() is called at the very last stage, after stopping hardware (if hw is still present)
- driver is responsible to be ready to be called from upper half of the stack,
  regardless of the state of hardware. Ususally it means that it should lock internal mutex,
  and check internal RUNNING flag.

Note, that hardware can go away asynchronously, so demand to call if_detach()
before stopping hardware can't be fulfiled.

-- 
Totus tuus, Glebius.