On Wed, Aug 15, 2018 at 08:37:22PM -0400, Aaron S. Knister wrote:
> 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.
>
> I've got a reproducer available if it's needed.
>
> 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;
> }
> }
Agree with Leon on the locking. Find a more elegant way to write
this.
> 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);
> }
Somehow I think the spinlock should be held across the skb_queue_len as
well. Right?
I wonder if the 'neigh->ah' has the same racing problem and needs the
same fixing:
} else if (neigh->ah) {
neigh_refresh_path(neigh, phdr->hwaddr, dev);
}
??
Have a feeling that needs a READ_ONCE to be correct...
Jason
