Re: [PATCH bpf-next v2 6/6] selftests/bpf: make sure bpf_testmod handling racing link destroying well.

[Martin KaFai Lau <martin.lau@linux.dev>]

On 5/6/24 10:56 PM, Kui-Feng Lee wrote:
> Subsystems that manage struct_ops objects may attempt to detach a link when
> the link has been released or is about to be released. The test in
> this patch demonstrate to developers the correct way to handle this
> situation using a locking mechanism and atomic64_inc_not_zero().
> 
> A subsystem must ensure that a link is valid when detaching the link. In
> order to achieve that, the subsystem may need to obtain a lock to safeguard
> a table that holds the pointer to the link being detached. However, the
> subsystem cannot invoke link->ops->detach() while holding the lock because
> other tasks may be in the process of unregistering, which could lead to a
> deadlock. This is why atomic64_inc_not_zero() is used to maintain the

Other tasks un-registering in parallel is not the reason for deadlock. The 
deadlock is because the link detach will call unreg() which usually will acquire 
the same lock (the detach_mutex here) and there is lock ordering with the 
update_mutex also. Hence, the link detach must be done after releasing the 
detach_mutex. After releasing the detach_mutex, the link is protected by its refcnt.

I think the above should be put as comments in bpf_dummy_do_link_detach for the 
subsystem to reference later.

> link's validity. (Refer to bpf_dummy_do_link_detach() in the previous patch
> for more details.)
> 
> This test make sure the pattern mentioned above work correctly.
> 
> Signed-off-by: Kui-Feng Lee <thinker.li@gmail.com>
> ---
>   .../bpf/prog_tests/test_struct_ops_module.c   | 44 +++++++++++++++++++
>   1 file changed, 44 insertions(+)
> 
> diff --git a/tools/testing/selftests/bpf/prog_tests/test_struct_ops_module.c b/tools/testing/selftests/bpf/prog_tests/test_struct_ops_module.c
> index 9f6657b53a93..1e37037cfd8a 100644
> --- a/tools/testing/selftests/bpf/prog_tests/test_struct_ops_module.c
> +++ b/tools/testing/selftests/bpf/prog_tests/test_struct_ops_module.c
> @@ -292,6 +292,48 @@ static void test_subsystem_detach(void)
>   	struct_ops_detach__destroy(skel);
>   }
>   
> +/* A subsystem detachs a link while the link is going to be free. */
> +static void test_subsystem_detach_free(void)
> +{
> +	LIBBPF_OPTS(bpf_test_run_opts, topts,
> +		    .data_in = &pkt_v4,
> +		    .data_size_in = sizeof(pkt_v4));
> +	struct struct_ops_detach *skel;
> +	struct bpf_link *link = NULL;
> +	int prog_fd;
> +	int err;
> +
> +	skel = struct_ops_detach__open_and_load();
> +	if (!ASSERT_OK_PTR(skel, "struct_ops_detach_open_and_load"))
> +		return;
> +
> +	link = bpf_map__attach_struct_ops(skel->maps.testmod_do_detach);
> +	if (!ASSERT_OK_PTR(link, "attach_struct_ops"))
> +		goto cleanup;
> +
> +	bpf_link__destroy(link);
> +
> +	prog_fd = bpf_program__fd(skel->progs.start_detach);
> +	if (!ASSERT_GE(prog_fd, 0, "start_detach_fd"))
> +		goto cleanup;
> +
> +	/* Do detachment from the registered subsystem */
> +	err = bpf_prog_test_run_opts(prog_fd, &topts);
> +	if (!ASSERT_OK(err, "start_detach_run"))
> +		goto cleanup;
> +
> +	/* The link may have zero refcount value and may have been
> +	 * unregistered, so the detachment from the subsystem should fail.
> +	 */
> +	ASSERT_EQ(topts.retval, (u32)-ENOENT, "start_detach_run retval");
> +
> +	/* Sync RCU to make sure the link is freed without any crash */
> +	ASSERT_OK(kern_sync_rcu(), "sync rcu");
> +
> +cleanup:
> +	struct_ops_detach__destroy(skel);
> +}
> +
>   void serial_test_struct_ops_module(void)
>   {
>   	if (test__start_subtest("test_struct_ops_load"))
> @@ -304,5 +346,7 @@ void serial_test_struct_ops_module(void)
>   		test_detach_link();
>   	if (test__start_subtest("test_subsystem_detach"))
>   		test_subsystem_detach();
> +	if (test__start_subtest("test_subsystem_detach_free"))
> +		test_subsystem_detach_free();
>   }
>