[openib-general] Re: user_mad.c: deadlock?

[Roland Dreier]

    Michael> Makes sense?

No, I think you have to make a copy of the full list of agents and
clean out the file's list while holding the mutex.  Otherwise
ib_umad_close() could run while ib_umad_kill_port() is dealing with
the same file.  And also we need ib_umad_kill_port() to wait for any
in-progress ib_umad_close() calls, since we don't want to call
ib_unregister_mad_agent() after we've returned from the device removal
call.

The locking is a little tricky, I'll work something out.

 - R.