IB/cm: Do not queue work to a device that's going away
authorErez Shitrit <erezsh@mellanox.com>
Thu, 25 Jun 2015 14:13:22 +0000 (17:13 +0300)
committerDoug Ledford <dledford@redhat.com>
Tue, 14 Jul 2015 17:20:09 +0000 (13:20 -0400)
Whenever ib_cm gets remove_one call, like when there is a hot-unplug
event, the driver should mark itself as going_down and confirm that no
new works are going to be queued for that device.
so, the order of the actions are:
1. mark the going_down bit.
2. flush the wq.
3. [make sure no new works for that device.]
4. unregister mad agent.

otherwise, works that are already queued can be scheduled after the mad
agent was freed.

Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/core/cm.c

index dbddddd..3a972eb 100644 (file)
@@ -169,6 +169,7 @@ struct cm_device {
        struct ib_device *ib_device;
        struct device *device;
        u8 ack_delay;
+       int going_down;
        struct cm_port *port[0];
 };
 
@@ -805,6 +806,11 @@ static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
 {
        int wait_time;
        unsigned long flags;
+       struct cm_device *cm_dev;
+
+       cm_dev = ib_get_client_data(cm_id_priv->id.device, &cm_client);
+       if (!cm_dev)
+               return;
 
        spin_lock_irqsave(&cm.lock, flags);
        cm_cleanup_timewait(cm_id_priv->timewait_info);
@@ -818,8 +824,14 @@ static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
         */
        cm_id_priv->id.state = IB_CM_TIMEWAIT;
        wait_time = cm_convert_to_ms(cm_id_priv->av.timeout);
-       queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
-                          msecs_to_jiffies(wait_time));
+
+       /* Check if the device started its remove_one */
+       spin_lock_irq(&cm.lock);
+       if (!cm_dev->going_down)
+               queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
+                                  msecs_to_jiffies(wait_time));
+       spin_unlock_irq(&cm.lock);
+
        cm_id_priv->timewait_info = NULL;
 }
 
@@ -3305,6 +3317,11 @@ static int cm_establish(struct ib_cm_id *cm_id)
        struct cm_work *work;
        unsigned long flags;
        int ret = 0;
+       struct cm_device *cm_dev;
+
+       cm_dev = ib_get_client_data(cm_id->device, &cm_client);
+       if (!cm_dev)
+               return -ENODEV;
 
        work = kmalloc(sizeof *work, GFP_ATOMIC);
        if (!work)
@@ -3343,7 +3360,17 @@ static int cm_establish(struct ib_cm_id *cm_id)
        work->remote_id = cm_id->remote_id;
        work->mad_recv_wc = NULL;
        work->cm_event.event = IB_CM_USER_ESTABLISHED;
-       queue_delayed_work(cm.wq, &work->work, 0);
+
+       /* Check if the device started its remove_one */
+       spin_lock_irq(&cm.lock);
+       if (!cm_dev->going_down) {
+               queue_delayed_work(cm.wq, &work->work, 0);
+       } else {
+               kfree(work);
+               ret = -ENODEV;
+       }
+       spin_unlock_irq(&cm.lock);
+
 out:
        return ret;
 }
@@ -3394,6 +3421,7 @@ static void cm_recv_handler(struct ib_mad_agent *mad_agent,
        enum ib_cm_event_type event;
        u16 attr_id;
        int paths = 0;
+       int going_down = 0;
 
        switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
        case CM_REQ_ATTR_ID:
@@ -3452,7 +3480,19 @@ static void cm_recv_handler(struct ib_mad_agent *mad_agent,
        work->cm_event.event = event;
        work->mad_recv_wc = mad_recv_wc;
        work->port = port;
-       queue_delayed_work(cm.wq, &work->work, 0);
+
+       /* Check if the device started its remove_one */
+       spin_lock_irq(&cm.lock);
+       if (!port->cm_dev->going_down)
+               queue_delayed_work(cm.wq, &work->work, 0);
+       else
+               going_down = 1;
+       spin_unlock_irq(&cm.lock);
+
+       if (going_down) {
+               kfree(work);
+               ib_free_recv_mad(mad_recv_wc);
+       }
 }
 
 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
@@ -3771,7 +3811,7 @@ static void cm_add_one(struct ib_device *ib_device)
 
        cm_dev->ib_device = ib_device;
        cm_get_ack_delay(cm_dev);
-
+       cm_dev->going_down = 0;
        cm_dev->device = device_create(&cm_class, &ib_device->dev,
                                       MKDEV(0, 0), NULL,
                                       "%s", ib_device->name);
@@ -3864,14 +3904,23 @@ static void cm_remove_one(struct ib_device *ib_device)
        list_del(&cm_dev->list);
        write_unlock_irqrestore(&cm.device_lock, flags);
 
+       spin_lock_irq(&cm.lock);
+       cm_dev->going_down = 1;
+       spin_unlock_irq(&cm.lock);
+
        for (i = 1; i <= ib_device->phys_port_cnt; i++) {
                if (!rdma_cap_ib_cm(ib_device, i))
                        continue;
 
                port = cm_dev->port[i-1];
                ib_modify_port(ib_device, port->port_num, 0, &port_modify);
-               ib_unregister_mad_agent(port->mad_agent);
+               /*
+                * We flush the queue here after the going_down set, this
+                * verify that no new works will be queued in the recv handler,
+                * after that we can call the unregister_mad_agent
+                */
                flush_workqueue(cm.wq);
+               ib_unregister_mad_agent(port->mad_agent);
                cm_remove_port_fs(port);
        }
        device_unregister(cm_dev->device);