svn commit: r246614 - head/sys/dev/usb/wlan
John Baldwin
jhb at freebsd.org
Wed Feb 13 16:00:06 UTC 2013
On Tuesday, February 12, 2013 2:43:32 am Hans Petter Selasky wrote:
> On Monday 11 February 2013 17:33:02 John Baldwin wrote:
> > On Sunday, February 10, 2013 5:36:17 am Hans Petter Selasky wrote:
> > > Author: hselasky
> > > Date: Sun Feb 10 10:36:16 2013
> > > New Revision: 246614
> > > URL: http://svnweb.freebsd.org/changeset/base/246614
> > >
> > > Log:
> > > - Streamline detach logic in wlan drivers, so that
> > >
> > > freed memory cannot be used during detach.
> > >
> > > - Remove all panic() calls from the urtw driver because
> > >
> > > panic() is not appropriate here.
> > >
> > > - Remove redundant checks for device detached in
> > >
> > > device detach callbacks.
> > >
> > > - Use DEVMETHOD_END to mark end of device methods.
> >
> > Using a detached flag to bail from ioctl generally means you are doing
> > things wrong in detach. The correct solution is to always detach your
> > ifnet first, then start tearing down other state. In general with device
> > detach routines the first order of business is removing external
> > references such as character devices, ifnets, etc. and only start shutting
> > down the hardware and releasing state once those steps have completed.
>
> Hi,
>
> What I can do to solve the problem is to lock a mutex while detaching. Is
the
> ifnet detach routine non-blocking?
>
> Why do I say that? It is because we are in a chicken-egg situation. USB is
> feeding data into ifnet and ifnet is feeding data into USB. Each stack is
> running under its own lock. The current approach is:
>
> 1) Stop data traffic in both directions
> 2) Make sure no more init happens (ioctl fix)
> 3) Free USB and IFNET.
>
> I see currently no way I can atomically stop all traffic and prevent futher
> init at the same time, while holding a single mutex. Do you?
You shouldn't call routines that can drain like if_detach() or destroy_dev()
or the like while holding any mutexes period. I think you need a 0) step
which is "detach external consumers" including cdev's (destroy_dev()) and
ifnet's (if_detach). Once both of those routines have finished, you can then
proceed with actually stopping device operation, and 2) is not needed as it
can't happen once 0) has finished.
When a device has both an ifnet and a cdev you may need something to handle
the case of one part being dead but not the other. For now you can call
if_detach() first as it doesn't sleep (though that is a bug). The real fix
for this case (and things like tun(4) can need this as well) is to have a way
to call just the non-blocking parts of things like if_detach() and
destroy_dev() first to mark the relevant portions as dead and then follow
those with blocking calls that do the drain. This is all part of step 0
though.
--
John Baldwin
More information about the svn-src-head
mailing list