The pvrusb2 driver was getting had by this scenario:
1. Task A calls kthread_stop() for task B.
2. Before exiting, then Task B calls kthread_stop() for task C.
The problem is, kthread_stop() wants to allocate an internal resource
to itself (i.e. acquire a lock), which won't be released until
kthread_stop() returns. But kthread_stop() won't return until task B
is dead. But task B won't die until it finishes its call to
kthread_stop() for task C, and that will block waiting on the resource
already allocated inside task A. Deadlock.
With the pvrusb2 driver, task A is the caller to pvr_exit(), task B is
the control thread run inside of pvrusb2-context.c, and task C is any
worker thread run inside of pvrusb2-hdw.c.
This problem got introduced by the previous threading setup change,
which was itself an attempt to fix a module tear-down race (which it
actually did fix). The lesson here is that a task being waited on as
part of a kthread_stop() simply cannot be allow to also issue a
kthread_stop() - or we make sure not to issue the enclosing
kthread_stop() until we know that the inner kthread_stop() has
completed first. The solution for the pvrusb2 driver is some hackish
code which changes the main control thread tear down into a two step
process. This then makes it possible to delay issuing the
kthread_stop() on the control thread until after we know that
everything has been torn down first. (And yes, we really need that
kthread_stop() because it's the only way to safely guarantee that a
module-referencing kernel thread has safely returned back out of the
module before we finally remove the module.)
Signed-off-by: Mike Isely <isely@pobox.com>
Signed-off-by: Mauro Carvalho Chehab <mchehab@infradead.org>
18ecbb4771