RDMA/core: Fix race when resolving IP address
authorDag Moxnes <dag.moxnes@oracle.com>
Tue, 9 Jul 2019 11:50:26 +0000 (13:50 +0200)
committerJason Gunthorpe <jgg@mellanox.com>
Tue, 9 Jul 2019 19:27:04 +0000 (16:27 -0300)
Use the neighbour lock when copying the MAC address from the neighbour
data struct in dst_fetch_ha.

When not using the lock, it is possible for the function to race with
neigh_update(), causing it to copy an torn MAC address:

rdma_resolve_addr()
  rdma_resolve_ip()
    addr_resolve()
      addr_resolve_neigh()
        fetch_ha()
          dst_fetch_ha()
     memcpy(dev_addr->dst_dev_addr, n->ha, MAX_ADDR_LEN)

and

net_ioctl()
  arp_ioctl()
    arp_rec_delete()
      arp_invalidate()
        neigh_update()
          __neigh_update()
    memcpy(&neigh->ha, lladdr, dev->addr_len)

It is possible to provoke this error by calling rdma_resolve_addr() in a
tight loop, while deleting the corresponding ARP entry in another tight
loop.

Fixes: 51d45974515c ("infiniband: addr: Consolidate code to fetch neighbour hardware address from dst.")
Signed-off-by: Dag Moxnes <dag.moxnes@oracle.com>
Signed-off-by: HÃ¥kon Bugge <haakon.bugge@oracle.com>
Reviewed-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
drivers/infiniband/core/addr.c

index 2f7d141..9b76a8f 100644 (file)
@@ -337,7 +337,7 @@ static int dst_fetch_ha(const struct dst_entry *dst,
                neigh_event_send(n, NULL);
                ret = -ENODATA;
        } else {
-               memcpy(dev_addr->dst_dev_addr, n->ha, MAX_ADDR_LEN);
+               neigh_ha_snapshot(dev_addr->dst_dev_addr, n, dst->dev);
        }
 
        neigh_release(n);