IB/ipoib: Avoid a race condition between start_xmit and cm_rep_handler
authorAaron Knister <aaron.s.knister@nasa.gov>
Fri, 24 Aug 2018 12:42:46 +0000 (08:42 -0400)
committerJason Gunthorpe <jgg@mellanox.com>
Wed, 5 Sep 2018 21:32:06 +0000 (15:32 -0600)
Inside of start_xmit() the call to check if the connection is up and the
queueing of the packets for later transmission is not atomic which leaves
a window where cm_rep_handler can run, set the connection up, dequeue
pending packets and leave the subsequently queued packets by start_xmit()
sitting on neigh->queue until they're dropped when the connection is torn
down. This only applies to connected mode. These dropped packets can
really upset TCP, for example, and cause multi-minute delays in
transmission for open connections.

Here's the code in start_xmit where we check to see if the connection is
up:

       if (ipoib_cm_get(neigh)) {
               if (ipoib_cm_up(neigh)) {
                       ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
                       goto unref;
               }
       }

The race occurs if cm_rep_handler execution occurs after the above
connection check (specifically if it gets to the point where it acquires
priv->lock to dequeue pending skb's) but before the below code snippet in
start_xmit where packets are queued.

       if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
               push_pseudo_header(skb, phdr->hwaddr);
               spin_lock_irqsave(&priv->lock, flags);
               __skb_queue_tail(&neigh->queue, skb);
               spin_unlock_irqrestore(&priv->lock, flags);
       } else {
               ++dev->stats.tx_dropped;
               dev_kfree_skb_any(skb);
       }

The patch acquires the netif tx lock in cm_rep_handler for the section
where it sets the connection up and dequeues and retransmits deferred
skb's.

Fixes: 839fcaba355a ("IPoIB: Connected mode experimental support")
Cc: stable@vger.kernel.org
Signed-off-by: Aaron Knister <aaron.s.knister@nasa.gov>
Tested-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
drivers/infiniband/ulp/ipoib/ipoib_cm.c

index ea01b8dd2be606206193408ca7bf9c42e4ea4e39..3d5424f335cb06e1dcaad16a4f4c3c7910d3295e 100644 (file)
@@ -1027,12 +1027,14 @@ static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id,
 
        skb_queue_head_init(&skqueue);
 
+       netif_tx_lock_bh(p->dev);
        spin_lock_irq(&priv->lock);
        set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
        if (p->neigh)
                while ((skb = __skb_dequeue(&p->neigh->queue)))
                        __skb_queue_tail(&skqueue, skb);
        spin_unlock_irq(&priv->lock);
+       netif_tx_unlock_bh(p->dev);
 
        while ((skb = __skb_dequeue(&skqueue))) {
                skb->dev = p->dev;