bpf: tcp: Allow bpf-tcp-cc to call bpf_(get|set)sockopt

This patch allows the bpf-tcp-cc to call bpf_setsockopt. One use
case is to allow a bpf-tcp-cc switching to another cc during init().
For example, when the tcp flow is not ecn ready, the bpf_dctcp
can switch to another cc by calling setsockopt(TCP_CONGESTION).

During setsockopt(TCP_CONGESTION), the new tcp-cc's init() will be
called and this could cause a recursion but it is stopped by the
current trampoline's logic (in the prog->active counter).

While retiring a bpf-tcp-cc (e.g. in tcp_v[46]_destroy_sock()),
the tcp stack calls bpf-tcp-cc's release(). To avoid the retiring
bpf-tcp-cc making further changes to the sk, bpf_setsockopt is not
available to the bpf-tcp-cc's release(). This will avoid release()
making setsockopt() call that will potentially allocate new resources.

Although the bpf-tcp-cc already has a more powerful way to read tcp_sock
from the PTR_TO_BTF_ID, it is usually expected that bpf_getsockopt and
bpf_setsockopt are available together. Thus, bpf_getsockopt() is also
added to all tcp_congestion_ops except release().

When the old bpf-tcp-cc is calling setsockopt(TCP_CONGESTION)
to switch to a new cc, the old bpf-tcp-cc will be released by
bpf_struct_ops_put(). Thus, this patch also puts the bpf_struct_ops_map
after a rcu grace period because the trampoline's image cannot be freed
while the old bpf-tcp-cc is still running.

bpf-tcp-cc can only access icsk_ca_priv as SCALAR. All kernel's
tcp-cc is also accessing the icsk_ca_priv as SCALAR. The size
of icsk_ca_priv has already been raised a few times to avoid
extra kmalloc and memory referencing. The only exception is the
kernel's tcp_cdg.c that stores a kmalloc()-ed pointer in icsk_ca_priv.
To avoid the old bpf-tcp-cc accidentally overriding this tcp_cdg's pointer
value stored in icsk_ca_priv after switching and without over-complicating
the bpf's verifier for this one exception in tcp_cdg, this patch does not
allow switching to tcp_cdg. If there is a need, bpf_tcp_cdg can be
implemented and then use the bpf_sk_storage as the extended storage.

bpf_sk_setsockopt proto has only been recently added and used
in bpf-sockopt and bpf-iter-tcp, so impose the tcp_cdg limitation in the
same proto instead of adding a new proto specifically for bpf-tcp-cc.

Signed-off-by: Martin KaFai Lau <[e-mail]>
Signed-off-by: Alexei Starovoitov <[e-mail]>
Link: https://lore.kernel.org/bpf/[e-mail]

kernel/bpf/bpf_struct_ops.c

···
 
 struct bpf_struct_ops_map {
 	struct bpf_map map;
+	struct rcu_head rcu;
 	const struct bpf_struct_ops *st_ops;
 	/* protect map_update */
 	struct mutex lock;
···
 	return refcount_inc_not_zero(&kvalue->refcnt);
 }
 
+static void bpf_struct_ops_put_rcu(struct rcu_head *head)
+{
+	struct bpf_struct_ops_map *st_map;
+
+	st_map = container_of(head, struct bpf_struct_ops_map, rcu);
+	bpf_map_put(&st_map->map);
+}
+
 void bpf_struct_ops_put(const void *kdata)
 {
 	struct bpf_struct_ops_value *kvalue;
···
 
 		st_map = container_of(kvalue, struct bpf_struct_ops_map,
 				      kvalue);
-		bpf_map_put(&st_map->map);
+		/* The struct_ops's function may switch to another struct_ops.
+		 *
+		 * For example, bpf_tcp_cc_x->init() may switch to
+		 * another tcp_cc_y by calling
+		 * setsockopt(TCP_CONGESTION, "tcp_cc_y").
+		 * During the switch,  bpf_struct_ops_put(tcp_cc_x) is called
+		 * and its map->refcnt may reach 0 which then free its
+		 * trampoline image while tcp_cc_x is still running.
+		 *
+		 * Thus, a rcu grace period is needed here.
+		 */
+		call_rcu(&st_map->rcu, bpf_struct_ops_put_rcu);
 	}
 }

net/core/filter.c

···
 BPF_CALL_5(bpf_sk_setsockopt, struct sock *, sk, int, level,
 	   int, optname, char *, optval, int, optlen)
 {
+	if (level == SOL_TCP && optname == TCP_CONGESTION) {
+		if (optlen >= sizeof("cdg") - 1 &&
+		    !strncmp("cdg", optval, optlen))
+			return -ENOTSUPP;
+	}
+
 	return _bpf_setsockopt(sk, level, optname, optval, optlen);
 }
 

net/ipv4/bpf_tcp_ca.c

···
 #include <net/tcp.h>
 #include <net/bpf_sk_storage.h>
 
+/* "extern" is to avoid sparse warning.  It is only used in bpf_struct_ops.c. */
+extern struct bpf_struct_ops bpf_tcp_congestion_ops;
+
 static u32 optional_ops[] = {
 	offsetof(struct tcp_congestion_ops, init),
 	offsetof(struct tcp_congestion_ops, release),
···
 	.arg2_type	= ARG_ANYTHING,
 };
 
+static u32 prog_ops_moff(const struct bpf_prog *prog)
+{
+	const struct btf_member *m;
+	const struct btf_type *t;
+	u32 midx;
+
+	midx = prog->expected_attach_type;
+	t = bpf_tcp_congestion_ops.type;
+	m = &btf_type_member(t)[midx];
+
+	return btf_member_bit_offset(t, m) / 8;
+}
+
 static const struct bpf_func_proto *
 bpf_tcp_ca_get_func_proto(enum bpf_func_id func_id,
 			  const struct bpf_prog *prog)
···
 		return &bpf_sk_storage_get_proto;
 	case BPF_FUNC_sk_storage_delete:
 		return &bpf_sk_storage_delete_proto;
+	case BPF_FUNC_setsockopt:
+		/* Does not allow release() to call setsockopt.
+		 * release() is called when the current bpf-tcp-cc
+		 * is retiring.  It is not allowed to call
+		 * setsockopt() to make further changes which
+		 * may potentially allocate new resources.
+		 */
+		if (prog_ops_moff(prog) !=
+		    offsetof(struct tcp_congestion_ops, release))
+			return &bpf_sk_setsockopt_proto;
+		return NULL;
+	case BPF_FUNC_getsockopt:
+		/* Since get/setsockopt is usually expected to
+		 * be available together, disable getsockopt for
+		 * release also to avoid usage surprise.
+		 * The bpf-tcp-cc already has a more powerful way
+		 * to read tcp_sock from the PTR_TO_BTF_ID.
+		 */
+		if (prog_ops_moff(prog) !=
+		    offsetof(struct tcp_congestion_ops, release))
+			return &bpf_sk_getsockopt_proto;
+		return NULL;
 	default:
 		return bpf_base_func_proto(func_id);
 	}
···
 {
 	tcp_unregister_congestion_control(kdata);
 }
-
-/* Avoid sparse warning.  It is only used in bpf_struct_ops.c. */
-extern struct bpf_struct_ops bpf_tcp_congestion_ops;
 
 struct bpf_struct_ops bpf_tcp_congestion_ops = {
 	.verifier_ops = &bpf_tcp_ca_verifier_ops,