xsk_init() previously registered the PF_XDP socket family before the
per-net subsystem and other prerequisites (netdevice notifier, caches)
were fully initialized.

This exposed .create = xsk_create() to user space while per-netns
state (net->xdp.lock/list) was still uninitialized. A task with
CAP_NET_RAW could trigger this during boot/module load by calling
socket(PF_XDP, SOCK_RAW, 0) concurrently with xsk_init(), leading
to a NULL pointer dereference or use-after-free in the list manipulation.

To fix this, move sock_register() to the end of the initialization
sequence, ensuring that all required kernel structures are ready before
exposing the AF_XDP interface to userspace.

Accordingly, reorder the error unwind path to ensure proper cleanup
in reverse order of initialization. Also, explicitly add
kmem_cache_destroy() in the error path to prevent leaking
xsk_tx_generic_cache if the registration fails.

Fixes: c0c77d8fb787 ("xsk: add user memory registration support sockopt")
Signed-off-by: Kery Qi <qikeyu2017@gmail.com>
---
 net/xdp/xsk.c | 19 ++++++++++---------
 1 file changed, 10 insertions(+), 9 deletions(-)

diff --git a/net/xdp/xsk.c b/net/xdp/xsk.c
index f093c3453f64..d402f23dfd8e 100644
--- a/net/xdp/xsk.c
+++ b/net/xdp/xsk.c
@@ -23,6 +23,7 @@
 #include <linux/netdevice.h>
 #include <linux/rculist.h>
 #include <linux/vmalloc.h>
+#include <linux/slab.h>
 #include <net/xdp_sock_drv.h>
 #include <net/busy_poll.h>
 #include <net/netdev_lock.h>
@@ -1922,13 +1923,9 @@ static int __init xsk_init(void)
 	if (err)
 		goto out;
 
-	err = sock_register(&xsk_family_ops);
-	if (err)
-		goto out_proto;
-
 	err = register_pernet_subsys(&xsk_net_ops);
 	if (err)
-		goto out_sk;
+		goto out_proto;
 
 	err = register_netdevice_notifier(&xsk_netdev_notifier);
 	if (err)
@@ -1939,17 +1936,21 @@ static int __init xsk_init(void)
 						 0, SLAB_HWCACHE_ALIGN, NULL);
 	if (!xsk_tx_generic_cache) {
 		err = -ENOMEM;
-		goto out_unreg_notif;
+		goto out_notifier;
 	}
 
+	err = sock_register(&xsk_family_ops);
+	if (err)
+		goto out_cache;
+
 	return 0;
 
-out_unreg_notif:
+out_cache:
+	kmem_cache_destroy(xsk_tx_generic_cache);
+out_notifier:
 	unregister_netdevice_notifier(&xsk_netdev_notifier);
 out_pernet:
 	unregister_pernet_subsys(&xsk_net_ops);
-out_sk:
-	sock_unregister(PF_XDP);
 out_proto:
 	proto_unregister(&xsk_proto);
 out:
-- 
2.34.1

On 01/09, Kery Qi wrote:
> xsk_init() previously registered the PF_XDP socket family before the
> per-net subsystem and other prerequisites (netdevice notifier, caches)
> were fully initialized.
> 
> This exposed .create = xsk_create() to user space while per-netns
> state (net->xdp.lock/list) was still uninitialized. A task with
> CAP_NET_RAW could trigger this during boot/module load by calling
> socket(PF_XDP, SOCK_RAW, 0) concurrently with xsk_init(), leading
> to a NULL pointer dereference or use-after-free in the list manipulation.
> 
> To fix this, move sock_register() to the end of the initialization
> sequence, ensuring that all required kernel structures are ready before
> exposing the AF_XDP interface to userspace.
> 
> Accordingly, reorder the error unwind path to ensure proper cleanup
> in reverse order of initialization. Also, explicitly add
> kmem_cache_destroy() in the error path to prevent leaking
> xsk_tx_generic_cache if the registration fails.

Is it something that you've hit in real life? xsk_init happens
so early during the init process (fs_init) that I don't understand
why the oder would matter.