Dave Miller <davem@davemloft.net> provided a detailed description of
why the way IPoIB is using neighbours for its own ipoib_neigh struct
is buggy:

    Any time an ipoib_neigh is changed, a sequence like the following is made:

    			spin_lock_irqsave(&priv->lock, flags);
    			/*
    			 * It's safe to call ipoib_put_ah() inside
    			 * priv->lock here, because we know that
    			 * path->ah will always hold one more reference,
    			 * so ipoib_put_ah() will never do more than
    			 * decrement the ref count.
    			 */
    			if (neigh->ah)
    				ipoib_put_ah(neigh->ah);
    			list_del(&neigh->list);
    			ipoib_neigh_free(dev, neigh);
    			spin_unlock_irqrestore(&priv->lock, flags);
    			ipoib_path_lookup(skb, n, dev);

    This doesn't work, because you're leaving a stale pointer to the freed up
    ipoib_neigh in the special neigh->ha pointer cookie.  Yes, it even fails
    with all the locking done to protect _changes_ to *ipoib_neigh(n), and
    with the code in ipoib_neigh_free() that NULLs out the pointer.

    The core issue is that read side calls to *to_ipoib_neigh(n) are not
    being synchronized at all, they are performed without any locking.  So
    whether we hold the lock or not when making changes to *ipoib_neigh(n)
    you still can have threads see references to freed up ipoib_neigh
    objects.

    	cpu 1			cpu 2
    	n = *ipoib_neigh()
    				*ipoib_neigh() = NULL
    				kfree(n)
    	n->foo == OOPS

    [..]

    Perhaps the ipoib code can have a private path database it manages
    entirely itself, which holds all the necessary information and is
    looked up by some generic key which is available easily at transmit
    time and does not involve generic neighbour entries.

See <http://marc.info/?l=linux-rdma&m=132812793105624&w=2> and
<http://marc.info/?l=linux-rdma&w=2&r=1&s=allows+references+to+freed+memory&q=b

>
for the full discussion.

This patch aims to solve the race conditions found in the IPoIB driver.

The patch removes the connection between the core networking neighbour
structure and the ipoib_neigh structure.  In addition to avoiding the
race described above, it allows us to handle SKBs carrying IP packets
that don't have any associated neighbour.

We add an ipoib_neigh hash table with N buckets where the key is the
destination hardware address.  The ipoib_neigh is fetched from the
hash table and instead of the stashed location in the neighbour
structure. The hash table uses both RCU and reference counting to
guarantee that no ipoib_neigh instance is ever deleted while in use.

Fetching the ipoib_neigh structure instance from the hash also makes
the special code in ipoib_start_xmit that handles remote and local
bonding failover redundant.

Aged ipoib_neigh instances are deleted by a garbage collection task
that runs every M seconds and deletes every ipoib_neigh instance that
was idle for at least 2*M seconds. The deletion is safe since the
ipoib_neigh instances are protected using RCU and reference count
mechanisms.

The number of buckets (N) and frequency of running the GC thread (M),
are taken from the exported arb_tbl.

Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>