Special schedulers, one CPU only kernel, one only userland

[Brooks Davis]

On Thu, Aug 18, 2005 at 01:12:20PM -0400, John Baldwin wrote:
> On Thursday 18 August 2005 12:55 pm, Luigi Rizzo wrote:
> > On Thu, Aug 18, 2005 at 10:23:04AM -0400, John Baldwin wrote:
> > ...
> >
> > > > However I am still unclear on what happens if a detach() is racing with
> > > > the output path (leading to fxp_start()).
> > >
> > > Note that we first down the interface via fxp_stop() and then we unhook
> > > it from the network stack using ether_ifdetach().  Once we've done
> > > ether_ifdetach() the network stack can't get to the fxp device anymore.
> >
> > It might have gotten there before, then this sequence might occur:
> >
> > 	thread 'fxp_detach'	thread 'fxp_start'
> >
> > 	acquire fxp lock	wait for lock (goes to sleep maybe ?)
> > 	fxp_stop
> > 	release fxp lock
> > 	destroy everything
> > 	including the lock
> > 				resume, mtx_lock -> boom
> >
> > hmmm... it's really tricky to follow. Maybe this does not happen,
> > but i wouldn't know why as fxp_detach() is under giant but the
> > path leading to fxp_start is not...
> 
> ether_ifdetach() should be handling this for us I think by blocking until any 
> known top-half threads are out of the driver.  It may not be doing that yet, 
> but I think that is where it should happen similar to how we use 
> bus_teardown_intr() and callout_drain() in detach to block until other 
> threads are out of our driver if necessary.

Certainly we need assurance that nothing is going to try and touch
the driver before the driver's detach function calls if_free.  After
if_free returns, no aspect of the driver should touch the ifnet.  Other
parts of the networking system might still have refrence to it should we
end up with a refcouting scheme where the final free of ifnets comes
some time later.  All refrences to the softc need to be pruged prior to
exit from the detach routine.

-- Brooks

-- 
Any statement of the form "X is the one, true Y" is FALSE.
PGP fingerprint 655D 519C 26A7 82E7 2529  9BF0 5D8E 8BE9 F238 1AD4
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-arch/attachments/20050818/8fbfdc8f/attachment.bin