Merge branch 'Avoid dummy bpf_offload_netdev in __bpf_prog_dev_bound_init'

Eduard Zingerman says:

====================
For a device bound BPF program with flag BPF_F_XDP_DEV_BOUND_ONLY,
in case if device does not support offload, __bpf_prog_dev_bound_init()
creates a dummy bpf_offload_netdev struct with .offdev field set to NULL.

This dummy struct might be reused for programs without this flag
bound to the same device. However, bpf_prog_offload_verifier_prep()
that uses bpf_offload_netdev assumes that .offdev field cannot be NULL.

This bug was reported by syzbot in [1].

[1] https://lore.kernel.org/bpf/000000000000d97f3c060479c4f8@google.com/
====================

Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>

Author: Martin KaFai Lau

kernel/bpf/offload.c

@@ -199,12 +199,14 @@ static int __bpf_prog_dev_bound_init(struct bpf_prog *prog, struct net_device *n
 	offload->netdev = netdev;
 
 	ondev = bpf_offload_find_netdev(offload->netdev);
+	/* When program is offloaded require presence of "true"
+	 * bpf_offload_netdev, avoid the one created for !ondev case below.
+	 */
+	if (bpf_prog_is_offloaded(prog->aux) && (!ondev || !ondev->offdev)) {
+		err = -EINVAL;
+		goto err_free;
+	}
 	if (!ondev) {
-		if (bpf_prog_is_offloaded(prog->aux)) {
-			err = -EINVAL;
-			goto err_free;
-		}
-
 		/* When only binding to the device, explicitly
 		 * create an entry in the hashtable.
 		 */

tools/testing/selftests/bpf/prog_tests/xdp_dev_bound_only.c

@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: GPL-2.0
+#include <net/if.h>
+#include <test_progs.h>
+#include <network_helpers.h>
+
+#define LOCAL_NETNS "xdp_dev_bound_only_netns"
+
+static int load_dummy_prog(char *name, __u32 ifindex, __u32 flags)
+{
+	struct bpf_insn insns[] = { BPF_MOV64_IMM(BPF_REG_0, 0), BPF_EXIT_INSN() };
+	LIBBPF_OPTS(bpf_prog_load_opts, opts);
+
+	opts.prog_flags = flags;
+	opts.prog_ifindex = ifindex;
+	return bpf_prog_load(BPF_PROG_TYPE_XDP, name, "GPL", insns, ARRAY_SIZE(insns), &opts);
+}
+
+/* A test case for bpf_offload_netdev->offload handling bug:
+ * - create a veth device (does not support offload);
+ * - create a device bound XDP program with BPF_F_XDP_DEV_BOUND_ONLY flag
+ *   (such programs are not offloaded);
+ * - create a device bound XDP program without flags (such programs are offloaded).
+ * This might lead to 'BUG: kernel NULL pointer dereference'.
+ */
+void test_xdp_dev_bound_only_offdev(void)
+{
+	struct nstoken *tok = NULL;
+	__u32 ifindex;
+	int fd1 = -1;
+	int fd2 = -1;
+
+	SYS(out, "ip netns add " LOCAL_NETNS);
+	tok = open_netns(LOCAL_NETNS);
+	if (!ASSERT_OK_PTR(tok, "open_netns"))
+		goto out;
+	SYS(out, "ip link add eth42 type veth");
+	ifindex = if_nametoindex("eth42");
+	if (!ASSERT_NEQ(ifindex, 0, "if_nametoindex")) {
+		perror("if_nametoindex");
+		goto out;
+	}
+	fd1 = load_dummy_prog("dummy1", ifindex, BPF_F_XDP_DEV_BOUND_ONLY);
+	if (!ASSERT_GE(fd1, 0, "load_dummy_prog #1")) {
+		perror("load_dummy_prog #1");
+		goto out;
+	}
+	/* Program with ifindex is considered offloaded, however veth
+	 * does not support offload => error should be reported.
+	 */
+	fd2 = load_dummy_prog("dummy2", ifindex, 0);
+	ASSERT_EQ(fd2, -EINVAL, "load_dummy_prog #2 (offloaded)");
+
+out:
+	close(fd1);
+	close(fd2);
+	close_netns(tok);
+	/* eth42 was added inside netns, removing the netns will
+	 * also remove eth42 veth pair.
+	 */
+	SYS_NOFAIL("ip netns del " LOCAL_NETNS);
+}