]> git.karo-electronics.de Git - karo-tx-linux.git/commitdiff
nvmet-rdma: Correctly handle RDMA device hot removal
authorSagi Grimberg <sagi@grimberg.me>
Thu, 19 May 2016 12:24:55 +0000 (15:24 +0300)
committerSagi Grimberg <sagi@grimberg.me>
Thu, 4 Aug 2016 14:43:06 +0000 (17:43 +0300)
When configuring a device attached listener, we may
see device removal events. In this case we return a
non-zero return code from the cm event handler which
implicitly destroys the cm_id. It is possible that in
the future the user will remove this listener and by
that trigger a second call to rdma_destroy_id on an
already destroyed cm_id -> BUG.

In addition, when a queue bound (active session) cm_id
generates a DEVICE_REMOVAL event we must guarantee all
resources are cleaned up by the time we return from the
event handler.

Introduce nvmet_rdma_device_removal which addresses
(or at least attempts to) both scenarios.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Christoph Hellwig <hch@lst.de>
drivers/nvme/target/rdma.c

index e06d504bdf0c81aad35c2643c299abec105e7337..48c811850c29b285b5e9a526086b0a749b04108e 100644 (file)
@@ -77,6 +77,7 @@ enum nvmet_rdma_queue_state {
        NVMET_RDMA_Q_CONNECTING,
        NVMET_RDMA_Q_LIVE,
        NVMET_RDMA_Q_DISCONNECTING,
+       NVMET_RDMA_IN_DEVICE_REMOVAL,
 };
 
 struct nvmet_rdma_queue {
@@ -984,7 +985,10 @@ static void nvmet_rdma_release_queue_work(struct work_struct *w)
        struct nvmet_rdma_device *dev = queue->dev;
 
        nvmet_rdma_free_queue(queue);
-       rdma_destroy_id(cm_id);
+
+       if (queue->state != NVMET_RDMA_IN_DEVICE_REMOVAL)
+               rdma_destroy_id(cm_id);
+
        kref_put(&dev->ref, nvmet_rdma_free_dev);
 }
 
@@ -1233,8 +1237,9 @@ static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
        switch (queue->state) {
        case NVMET_RDMA_Q_CONNECTING:
        case NVMET_RDMA_Q_LIVE:
-               disconnect = true;
                queue->state = NVMET_RDMA_Q_DISCONNECTING;
+       case NVMET_RDMA_IN_DEVICE_REMOVAL:
+               disconnect = true;
                break;
        case NVMET_RDMA_Q_DISCONNECTING:
                break;
@@ -1272,6 +1277,62 @@ static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
        schedule_work(&queue->release_work);
 }
 
+/**
+ * nvme_rdma_device_removal() - Handle RDMA device removal
+ * @queue:      nvmet rdma queue (cm id qp_context)
+ * @addr:      nvmet address (cm_id context)
+ *
+ * DEVICE_REMOVAL event notifies us that the RDMA device is about
+ * to unplug so we should take care of destroying our RDMA resources.
+ * This event will be generated for each allocated cm_id.
+ *
+ * Note that this event can be generated on a normal queue cm_id
+ * and/or a device bound listener cm_id (where in this case
+ * queue will be null).
+ *
+ * we claim ownership on destroying the cm_id. For queues we move
+ * the queue state to NVMET_RDMA_IN_DEVICE_REMOVAL and for port
+ * we nullify the priv to prevent double cm_id destruction and destroying
+ * the cm_id implicitely by returning a non-zero rc to the callout.
+ */
+static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
+               struct nvmet_rdma_queue *queue)
+{
+       unsigned long flags;
+
+       if (!queue) {
+               struct nvmet_port *port = cm_id->context;
+
+               /*
+                * This is a listener cm_id. Make sure that
+                * future remove_port won't invoke a double
+                * cm_id destroy. use atomic xchg to make sure
+                * we don't compete with remove_port.
+                */
+               if (xchg(&port->priv, NULL) != cm_id)
+                       return 0;
+       } else {
+               /*
+                * This is a queue cm_id. Make sure that
+                * release queue will not destroy the cm_id
+                * and schedule all ctrl queues removal (only
+                * if the queue is not disconnecting already).
+                */
+               spin_lock_irqsave(&queue->state_lock, flags);
+               if (queue->state != NVMET_RDMA_Q_DISCONNECTING)
+                       queue->state = NVMET_RDMA_IN_DEVICE_REMOVAL;
+               spin_unlock_irqrestore(&queue->state_lock, flags);
+               nvmet_rdma_queue_disconnect(queue);
+               flush_scheduled_work();
+       }
+
+       /*
+        * We need to return 1 so that the core will destroy
+        * it's own ID.  What a great API design..
+        */
+       return 1;
+}
+
 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
                struct rdma_cm_event *event)
 {
@@ -1294,20 +1355,11 @@ static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
                break;
        case RDMA_CM_EVENT_ADDR_CHANGE:
        case RDMA_CM_EVENT_DISCONNECTED:
-       case RDMA_CM_EVENT_DEVICE_REMOVAL:
        case RDMA_CM_EVENT_TIMEWAIT_EXIT:
-               /*
-                * We can get the device removal callback even for a
-                * CM ID that we aren't actually using.  In that case
-                * the context pointer is NULL, so we shouldn't try
-                * to disconnect a non-existing queue.  But we also
-                * need to return 1 so that the core will destroy
-                * it's own ID.  What a great API design..
-                */
-               if (queue)
-                       nvmet_rdma_queue_disconnect(queue);
-               else
-                       ret = 1;
+               nvmet_rdma_queue_disconnect(queue);
+               break;
+       case RDMA_CM_EVENT_DEVICE_REMOVAL:
+               ret = nvmet_rdma_device_removal(cm_id, queue);
                break;
        case RDMA_CM_EVENT_REJECTED:
        case RDMA_CM_EVENT_UNREACHABLE:
@@ -1396,9 +1448,10 @@ out_destroy_id:
 
 static void nvmet_rdma_remove_port(struct nvmet_port *port)
 {
-       struct rdma_cm_id *cm_id = port->priv;
+       struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
 
-       rdma_destroy_id(cm_id);
+       if (cm_id)
+               rdma_destroy_id(cm_id);
 }
 
 static struct nvmet_fabrics_ops nvmet_rdma_ops = {