RDMA/cma: Teach lockdep about the order of rtnl and lock

commit 32ac9e4399b12d3e54d312a0e0e30ed5cd19bd4e upstream.

This lock ordering only happens when bonding is enabled and a certain
bonding related event fires. However, since it can happen this is a global
restriction on lock ordering.

Teach lockdep about the order directly and unconditionally so bugs here
are found quickly.

See https://syzkaller.appspot.com/bug?extid=55de90ab5f44172b0c90

Link: https://lore.kernel.org/r/20200227203651.GA27185@ziepe.ca
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c
index 0b530646f1e515bc89b8432f5cf384a9ee11be96..bb2f2807cd58aeff6d8893ea0f97645a0c6c2747 100644 (file)
--- a/drivers/infiniband/core/cma.c
+++ b/drivers/infiniband/core/cma.c
@@ -4746,6 +4746,19 @@ static int __init cma_init(void)
 {
        int ret;
 
+       /*
+        * There is a rare lock ordering dependency in cma_netdev_callback()
+        * that only happens when bonding is enabled. Teach lockdep that rtnl
+        * must never be nested under lock so it can find these without having
+        * to test with bonding.
+        */
+       if (IS_ENABLED(CONFIG_LOCKDEP)) {
+               rtnl_lock();
+               mutex_lock(&lock);
+               mutex_unlock(&lock);
+               rtnl_unlock();
+       }
+
        cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
        if (!cma_wq)
                return -ENOMEM;