Previously the clients_lock was protecting the clients array against
concurrent addition/removal of clients but was also accessed from IRQ
context. This meant that it had to be a spinlock and that the add() and
remove() callbacks in which clients need to do allocation and take
mutexes can't be called under the clients_lock. To work around this these
callbacks were moved to workqueues. This not only introduced significant
complexity but is also subtly broken in at least one way.

In ism_dev_init() and ism_dev_exit() clients[i]->tgt_ism is used to
communicate the added/removed ISM device to the work function. While
write access to client[i]->tgt_ism is protected by the clients_lock and
the code waits that there is no pending add/remove work before and after
setting clients[i]->tgt_ism this is not enough. The problem is that the
wait happens based on per ISM device counters. Thus a concurrent
ism_dev_init()/ism_dev_exit() for a different ISM device may overwrite
a clients[i]->tgt_ism between unlocking the clients_lock and the
subsequent wait for the work to finnish.

Thankfully with the clients_lock no longer held in IRQ context it can be
turned into a mutex which can be held during the calls to add()/remove()
completely removing the need for the workqueues and the associated
broken housekeeping including the per ISM device counters and the
clients[i]->tgt_ism.

Fixes: 89e7d2ba ("net/ism: Add new API for client registration")
Signed-off-by: Niklas Schnelle <schnelle@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>