RDS: IB: fix panic due to handlers running post teardown
authorSantosh Shilimkar <santosh.shilimkar@oracle.com>
Thu, 29 Sep 2016 18:07:11 +0000 (11:07 -0700)
committerSantosh Shilimkar <santosh.shilimkar@oracle.com>
Mon, 2 Jan 2017 22:02:55 +0000 (14:02 -0800)
Shutdown code reaping loop takes care of emptying the
CQ's before they being destroyed. And once tasklets are
killed, the hanlders are not expected to run.

But because of core tasklet code issues, tasklet handler could
still run even after tasklet_kill,
RDS IB shutdown code already reaps the CQs before freeing
cq/qp resources so as such the handlers have nothing left
to do post shutdown.

On other hand any handler running after teardown and trying
to access already freed qp/cq resources causes issues
Patch fixes this race by  makes sure that handlers returns
without any action post teardown.

Reviewed-by: Wengang <wen.gang.wang@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
net/rds/ib.h
net/rds/ib_cm.c

index 1fe9f79fead52dd8cc00569d6955872ca86ad825..540458928f3c8eab4529b4072826b882563c941e 100644 (file)
@@ -185,6 +185,7 @@ struct rds_ib_connection {
 
        /* Endpoint role in connection */
        bool                    i_active_side;
+       atomic_t                i_cq_quiesce;
 
        /* Send/Recv vectors */
        int                     i_scq_vector;
index 33c8584ada1fdeddf66a312b8d41fbcb458daff2..ce3775abc6e7a1d30e335aaea749a9840e949786 100644 (file)
@@ -128,6 +128,8 @@ void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_even
                          ic->i_flowctl ? ", flow control" : "");
        }
 
+       atomic_set(&ic->i_cq_quiesce, 0);
+
        /* Init rings and fill recv. this needs to wait until protocol
         * negotiation is complete, since ring layout is different
         * from 3.1 to 4.1.
@@ -267,6 +269,10 @@ static void rds_ib_tasklet_fn_send(unsigned long data)
 
        rds_ib_stats_inc(s_ib_tasklet_call);
 
+       /* if cq has been already reaped, ignore incoming cq event */
+       if (atomic_read(&ic->i_cq_quiesce))
+               return;
+
        poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
        ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
        poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
@@ -308,6 +314,10 @@ static void rds_ib_tasklet_fn_recv(unsigned long data)
 
        rds_ib_stats_inc(s_ib_tasklet_call);
 
+       /* if cq has been already reaped, ignore incoming cq event */
+       if (atomic_read(&ic->i_cq_quiesce))
+               return;
+
        memset(&state, 0, sizeof(state));
        poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
        ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
@@ -804,6 +814,8 @@ void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
                tasklet_kill(&ic->i_send_tasklet);
                tasklet_kill(&ic->i_recv_tasklet);
 
+               atomic_set(&ic->i_cq_quiesce, 1);
+
                /* first destroy the ib state that generates callbacks */
                if (ic->i_cm_id->qp)
                        rdma_destroy_qp(ic->i_cm_id);