IB/rdmavt: Don't wait for resources in QP reset

Per the IBTA spec, QP destroy shall fail if the QP is attached
to multicast groups, although the spec is silent on modify_qp
to reset state. It implies that ULP must deregister QP from
all mcast groups for destroy to succeed.
The faulty patch "IB/ipoib: Update broadcast object if PKey value
was changed in index 0" exposed two issues in rdmavt:
1. Rvt QP reset waits for qp references to go to zero.
This will hang if QP is attached to multicast groups.
2. The mcast group detach will fail for a QP in reset state
therefore preventing ULP from correcting the issue.
This patch moves the reference count wait to the the destroy QP
path and allows a QP mcast detach to work in the reset state.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Alex Estrin <alex.estrin@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

drivers/infiniband/sw/rdmavt/mcast.c

@@ -351,7 +351,7 @@ int rvt_detach_mcast(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 	int last = 0;
 	int ret = 0;
 
-	if (ibqp->qp_num <= 1 || qp->state == IB_QPS_RESET)
+	if (ibqp->qp_num <= 1)
 		return -EINVAL;
 
 	spin_lock_irq(&ibp->lock);

drivers/infiniband/sw/rdmavt/qp.c

@@ -717,7 +717,6 @@ static void rvt_reset_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp,
 
 		/* take qp out the hash and wait for it to be unused */
 		rvt_remove_qp(rdi, qp);
-		wait_event(qp->wait, !atomic_read(&qp->refcount));
 
 		/* grab the lock b/c it was locked at call time */
 		spin_lock_irq(&qp->r_lock);
@@ -1443,6 +1442,7 @@ int rvt_destroy_qp(struct ib_qp *ibqp)
 	spin_unlock(&qp->s_hlock);
 	spin_unlock_irq(&qp->r_lock);
 
+	wait_event(qp->wait, !atomic_read(&qp->refcount));
 	/* qpn is now available for use again */
 	rvt_free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num);