Hi,

29.02.2024 02:51, Pablo Neira Ayuso wrote:

On Wed, Feb 28, 2024 at 02:47:03PM +0300, Alexander Ofitserov wrote:

quoted

The gtp_link_ops operations structure for the subsystem must be
registered after registering the gtp_net_ops pernet operations structure.

A fix for this was already applied, see:

commit 136cfaca22567a03bbb3bf53a43d8cb5748b80ec
Author: Vasiliy Kovalev [off-list ref]
Date:   Wed Feb 14 19:27:33 2024 +0300

     gtp: fix use-after-free and null-ptr-deref in gtp_genl_dump_pdp()

quoted

diff --git a/drivers/net/gtp.c b/drivers/net/gtp.c
index 2129ae42c7030..0ddec4cc84093 100644
--- a/drivers/net/gtp.c
+++ b/drivers/net/gtp.c

@@ -1903,26 +1903,26 @@ static int __init gtp_init(void)
  
  	get_random_bytes(&gtp_h_initval, sizeof(gtp_h_initval));
  
-	err = rtnl_link_register(&gtp_link_ops);
+	err = register_pernet_subsys(&gtp_net_ops);
  	if (err < 0)
  		goto error_out;

BTW, I like that this calls register_pernet_subsys() before
rtnl_link_register(), where a rtnetlink request could come before
pernet is set up.

quoted

-	err = register_pernet_subsys(&gtp_net_ops);
+	err = rtnl_link_register(&gtp_link_ops);
  	if (err < 0)
-		goto unreg_rtnl_link;
+		goto unreg_pernet_subsys;
  
  	err = genl_register_family(&gtp_genl_family);
  	if (err < 0)
-		goto unreg_pernet_subsys;
+		goto unreg_rtnl_link;
  
  	pr_info("GTP module loaded (pdp ctx size %zd bytes)\n",
  		sizeof(struct pdp_ctx));
  	return 0;
  
-unreg_pernet_subsys:
-	unregister_pernet_subsys(&gtp_net_ops);
  unreg_rtnl_link:
  	rtnl_link_unregister(&gtp_link_ops);
+unreg_pernet_subsys:
+	unregister_pernet_subsys(&gtp_net_ops);
  error_out:
  	pr_err("error loading GTP module loaded\n");
  	return err;
-- 
2.42.1

This patch fixes another problem, but a similar one, since the sequence 
is incorrect when registering subsystems.

Initially, the registration sequence in the gtp module was as follows:

1) rtnl_link_register();

2) genl_register_family();

3) register_pernet_subsys();

During debugging of the module, when starting the syzkaller reproducer, 
it turned out that after genl_register_family() (2),

without waiting for register_pernet_subsys()(3), the .dumpit  event is 
triggered, in which the data of the unregistered pernet subsystem is 
accessed.

That is, the bug was fixed by the commit

136cfaca2256 ("gtp: fix use-after-free and null-ptr-deref in 
gtp_genl_dump_pdp()") [1]

and the registration sequence became as follows:

1) rtnl_link_register();

2) register_pernet_subsys();

3) genl_register_family();

However, syzkaller has discovered another problem:

after registering rtnl_link_register, the .newlink event is triggered, 
in which the data of the unregistered pernet subsystem is accessed.

This problem is reproducible on current stable kernels and the latest 
upstream kernel 6.8-rc6, in which the patch 136cfaca2256 [1] is applied.

Therefore, the correct sequence should be as follows:

1) register_pernet_subsys();

2) rtnl_link_register();

3) genl_register_family();

The proposed patch is developed on top of the commit changes [1], does 
not conflict with it and fixes the described bug.

[1] 
https://lore.kernel.org/lkml/20240220160434.29bcaf43@kernel.org/T/#mb1f72c2ad57b7ea6d47333e8616beccf8bce0e23 (local)

-- 
Regards,
Vasiliy Kovalev