bpf: Fix bpf_tcp_sock and bpf_sk_fullsock issue related to bpf_sk_release

Lorenz Bauer [thanks!] reported that a ptr returned by bpf_tcp_sock(sk)
can still be accessed after bpf_sk_release(sk).
Both bpf_tcp_sock() and bpf_sk_fullsock() have the same issue.
This patch addresses them together.

A simple reproducer looks like this:

	sk = bpf_sk_lookup_tcp();
	/* if (!sk) ... */
	tp = bpf_tcp_sock(sk);
	/* if (!tp) ... */
	bpf_sk_release(sk);
	snd_cwnd = tp->snd_cwnd; /* oops! The verifier does not complain. */

The problem is the verifier did not scrub the register's states of
the tcp_sock ptr (tp) after bpf_sk_release(sk).

[ Note that when calling bpf_tcp_sock(sk), the sk is not always
  refcount-acquired. e.g. bpf_tcp_sock(skb->sk). The verifier works
  fine for this case. ]

Currently, the verifier does not track if a helper's return ptr (in REG_0)
is "carry"-ing one of its argument's refcount status. To carry this info,
the reg1->id needs to be stored in reg0.

One approach was tried, like "reg0->id = reg1->id", when calling
"bpf_tcp_sock()".  The main idea was to avoid adding another "ref_obj_id"
for the same reg.  However, overlapping the NULL marking and ref
tracking purpose in one "id" does not work well:

	ref_sk = bpf_sk_lookup_tcp();
	fullsock = bpf_sk_fullsock(ref_sk);
	tp = bpf_tcp_sock(ref_sk);
	if (!fullsock) {
	     bpf_sk_release(ref_sk);
	     return 0;
	}
	/* fullsock_reg->id is marked for NOT-NULL.
	 * Same for tp_reg->id because they have the same id.
	 */

	/* oops. verifier did not complain about the missing !tp check */
	snd_cwnd = tp->snd_cwnd;

Hence, a new "ref_obj_id" is needed in "struct bpf_reg_state".
With a new ref_obj_id, when bpf_sk_release(sk) is called, the verifier can
scrub all reg states which has a ref_obj_id match.  It is done with the
changes in release_reg_references() in this patch.

While fixing it, sk_to_full_sk() is removed from bpf_tcp_sock() and
bpf_sk_fullsock() to avoid these helpers from returning
another ptr. It will make bpf_sk_release(tp) possible:

	sk = bpf_sk_lookup_tcp();
	/* if (!sk) ... */
	tp = bpf_tcp_sock(sk);
	/* if (!tp) ... */
	bpf_sk_release(tp);

A separate helper "bpf_get_listener_sock()" will be added in a later
patch to do sk_to_full_sk().

Misc change notes:
- To allow bpf_sk_release(tp), the arg of bpf_sk_release() is changed
  from ARG_PTR_TO_SOCKET to ARG_PTR_TO_SOCK_COMMON.  ARG_PTR_TO_SOCKET
  is removed from bpf.h since no helper is using it.

- arg_type_is_refcounted() is renamed to arg_type_may_be_refcounted()
  because ARG_PTR_TO_SOCK_COMMON is the only one and skb->sk is not
  refcounted.  All bpf_sk_release(), bpf_sk_fullsock() and bpf_tcp_sock()
  take ARG_PTR_TO_SOCK_COMMON.

- check_refcount_ok() ensures is_acquire_function() cannot take
  arg_type_may_be_refcounted() as its argument.

- The check_func_arg() can only allow one refcount-ed arg.  It is
  guaranteed by check_refcount_ok() which ensures at most one arg can be
  refcounted.  Hence, it is a verifier internal error if >1 refcount arg
  found in check_func_arg().

- In release_reference(), release_reference_state() is called
  first to ensure a match on "reg->ref_obj_id" can be found before
  scrubbing the reg states with release_reg_references().

- reg_is_refcounted() is no longer needed.
  1. In mark_ptr_or_null_regs(), its usage is replaced by
     "ref_obj_id && ref_obj_id == id" because,
     when is_null == true, release_reference_state() should only be
     called on the ref_obj_id obtained by a acquire helper (i.e.
     is_acquire_function() == true).  Otherwise, the following
     would happen:

	sk = bpf_sk_lookup_tcp();
	/* if (!sk) { ... } */
	fullsock = bpf_sk_fullsock(sk);
	if (!fullsock) {
		/*
		 * release_reference_state(fullsock_reg->ref_obj_id)
		 * where fullsock_reg->ref_obj_id == sk_reg->ref_obj_id.
		 *
		 * Hence, the following bpf_sk_release(sk) will fail
		 * because the ref state has already been released in the
		 * earlier release_reference_state(fullsock_reg->ref_obj_id).
		 */
		bpf_sk_release(sk);
	}

  2. In release_reg_references(), the current reg_is_refcounted() call
     is unnecessary because the id check is enough.

- The type_is_refcounted() and type_is_refcounted_or_null()
  are no longer needed also because reg_is_refcounted() is removed.

Fixes: 655a51e5 ("bpf: Add struct bpf_tcp_sock and BPF_FUNC_tcp_sock")
Reported-by: Lorenz Bauer <lmb@cloudflare.com>
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

include/linux/bpf.h

@@ -193,7 +193,6 @@ enum bpf_arg_type {
 
 	ARG_PTR_TO_CTX,		/* pointer to context */
 	ARG_ANYTHING,		/* any (initialized) argument is ok */
-	ARG_PTR_TO_SOCKET,	/* pointer to bpf_sock */
 	ARG_PTR_TO_SPIN_LOCK,	/* pointer to bpf_spin_lock */
 	ARG_PTR_TO_SOCK_COMMON,	/* pointer to sock_common */
 };

include/linux/bpf_verifier.h

@@ -66,6 +66,46 @@ struct bpf_reg_state {
 	 * same reference to the socket, to determine proper reference freeing.
 	 */
 	u32 id;
+	/* PTR_TO_SOCKET and PTR_TO_TCP_SOCK could be a ptr returned
+	 * from a pointer-cast helper, bpf_sk_fullsock() and
+	 * bpf_tcp_sock().
+	 *
+	 * Consider the following where "sk" is a reference counted
+	 * pointer returned from "sk = bpf_sk_lookup_tcp();":
+	 *
+	 * 1: sk = bpf_sk_lookup_tcp();
+	 * 2: if (!sk) { return 0; }
+	 * 3: fullsock = bpf_sk_fullsock(sk);
+	 * 4: if (!fullsock) { bpf_sk_release(sk); return 0; }
+	 * 5: tp = bpf_tcp_sock(fullsock);
+	 * 6: if (!tp) { bpf_sk_release(sk); return 0; }
+	 * 7: bpf_sk_release(sk);
+	 * 8: snd_cwnd = tp->snd_cwnd;  // verifier will complain
+	 *
+	 * After bpf_sk_release(sk) at line 7, both "fullsock" ptr and
+	 * "tp" ptr should be invalidated also.  In order to do that,
+	 * the reg holding "fullsock" and "sk" need to remember
+	 * the original refcounted ptr id (i.e. sk_reg->id) in ref_obj_id
+	 * such that the verifier can reset all regs which have
+	 * ref_obj_id matching the sk_reg->id.
+	 *
+	 * sk_reg->ref_obj_id is set to sk_reg->id at line 1.
+	 * sk_reg->id will stay as NULL-marking purpose only.
+	 * After NULL-marking is done, sk_reg->id can be reset to 0.
+	 *
+	 * After "fullsock = bpf_sk_fullsock(sk);" at line 3,
+	 * fullsock_reg->ref_obj_id is set to sk_reg->ref_obj_id.
+	 *
+	 * After "tp = bpf_tcp_sock(fullsock);" at line 5,
+	 * tp_reg->ref_obj_id is set to fullsock_reg->ref_obj_id
+	 * which is the same as sk_reg->ref_obj_id.
+	 *
+	 * From the verifier perspective, if sk, fullsock and tp
+	 * are not NULL, they are the same ptr with different
+	 * reg->type.  In particular, bpf_sk_release(tp) is also
+	 * allowed and has the same effect as bpf_sk_release(sk).
+	 */
+	u32 ref_obj_id;
 	/* For scalar types (SCALAR_VALUE), this represents our knowledge of
 	 * the actual value.
 	 * For pointer types, this represents the variable part of the offset

kernel/bpf/verifier.c

@@ -212,7 +212,7 @@ struct bpf_call_arg_meta {
 	int access_size;
 	s64 msize_smax_value;
 	u64 msize_umax_value;
-	int ptr_id;
+	int ref_obj_id;
 	int func_id;
 };
 
@@ -346,35 +346,15 @@ static bool reg_type_may_be_null(enum bpf_reg_type type)
 	       type == PTR_TO_TCP_SOCK_OR_NULL;
 }
 
-static bool type_is_refcounted(enum bpf_reg_type type)
-{
-	return type == PTR_TO_SOCKET;
-}
-
-static bool type_is_refcounted_or_null(enum bpf_reg_type type)
-{
-	return type == PTR_TO_SOCKET || type == PTR_TO_SOCKET_OR_NULL;
-}
-
-static bool reg_is_refcounted(const struct bpf_reg_state *reg)
-{
-	return type_is_refcounted(reg->type);
-}
-
 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
 {
 	return reg->type == PTR_TO_MAP_VALUE &&
 		map_value_has_spin_lock(reg->map_ptr);
 }
 
-static bool reg_is_refcounted_or_null(const struct bpf_reg_state *reg)
+static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
 {
-	return type_is_refcounted_or_null(reg->type);
-}
-
-static bool arg_type_is_refcounted(enum bpf_arg_type type)
-{
-	return type == ARG_PTR_TO_SOCKET;
+	return type == ARG_PTR_TO_SOCK_COMMON;
 }
 
 /* Determine whether the function releases some resources allocated by another
@@ -392,6 +372,12 @@ static bool is_acquire_function(enum bpf_func_id func_id)
 		func_id == BPF_FUNC_sk_lookup_udp;
 }
 
+static bool is_ptr_cast_function(enum bpf_func_id func_id)
+{
+	return func_id == BPF_FUNC_tcp_sock ||
+		func_id == BPF_FUNC_sk_fullsock;
+}
+
 /* string representation of 'enum bpf_reg_type' */
 static const char * const reg_type_str[] = {
 	[NOT_INIT]		= "?",
@@ -465,7 +451,8 @@ static void print_verifier_state(struct bpf_verifier_env *env,
 			if (t == PTR_TO_STACK)
 				verbose(env, ",call_%d", func(env, reg)->callsite);
 		} else {
-			verbose(env, "(id=%d", reg->id);
+			verbose(env, "(id=%d ref_obj_id=%d", reg->id,
+				reg->ref_obj_id);
 			if (t != SCALAR_VALUE)
 				verbose(env, ",off=%d", reg->off);
 			if (type_is_pkt_pointer(t))
@@ -2414,16 +2401,15 @@ static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
 		/* Any sk pointer can be ARG_PTR_TO_SOCK_COMMON */
 		if (!type_is_sk_pointer(type))
 			goto err_type;
-	} else if (arg_type == ARG_PTR_TO_SOCKET) {
-		expected_type = PTR_TO_SOCKET;
-		if (type != expected_type)
-			goto err_type;
-		if (meta->ptr_id || !reg->id) {
-			verbose(env, "verifier internal error: mismatched references meta=%d, reg=%d\n",
-				meta->ptr_id, reg->id);
-			return -EFAULT;
+		if (reg->ref_obj_id) {
+			if (meta->ref_obj_id) {
+				verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
+					regno, reg->ref_obj_id,
+					meta->ref_obj_id);
+				return -EFAULT;
+			}
+			meta->ref_obj_id = reg->ref_obj_id;
 		}
-		meta->ptr_id = reg->id;
 	} else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
 		if (meta->func_id == BPF_FUNC_spin_lock) {
 			if (process_spin_lock(env, regno, true))
@@ -2740,32 +2726,38 @@ static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
 	return true;
 }
 
-static bool check_refcount_ok(const struct bpf_func_proto *fn)
+static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
 {
 	int count = 0;
 
-	if (arg_type_is_refcounted(fn->arg1_type))
+	if (arg_type_may_be_refcounted(fn->arg1_type))
 		count++;
-	if (arg_type_is_refcounted(fn->arg2_type))
+	if (arg_type_may_be_refcounted(fn->arg2_type))
 		count++;
-	if (arg_type_is_refcounted(fn->arg3_type))
+	if (arg_type_may_be_refcounted(fn->arg3_type))
 		count++;
-	if (arg_type_is_refcounted(fn->arg4_type))
+	if (arg_type_may_be_refcounted(fn->arg4_type))
 		count++;
-	if (arg_type_is_refcounted(fn->arg5_type))
+	if (arg_type_may_be_refcounted(fn->arg5_type))
 		count++;
 
+	/* A reference acquiring function cannot acquire
+	 * another refcounted ptr.
+	 */
+	if (is_acquire_function(func_id) && count)
+		return false;
+
 	/* We only support one arg being unreferenced at the moment,
 	 * which is sufficient for the helper functions we have right now.
 	 */
 	return count <= 1;
 }
 
-static int check_func_proto(const struct bpf_func_proto *fn)
+static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
 {
 	return check_raw_mode_ok(fn) &&
 	       check_arg_pair_ok(fn) &&
-	       check_refcount_ok(fn) ? 0 : -EINVAL;
+	       check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
 }
 
 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
@@ -2799,19 +2791,20 @@ static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
 }
 
 static void release_reg_references(struct bpf_verifier_env *env,
-				   struct bpf_func_state *state, int id)
+				   struct bpf_func_state *state,
+				   int ref_obj_id)
 {
 	struct bpf_reg_state *regs = state->regs, *reg;
 	int i;
 
 	for (i = 0; i < MAX_BPF_REG; i++)
-		if (regs[i].id == id)
+		if (regs[i].ref_obj_id == ref_obj_id)
 			mark_reg_unknown(env, regs, i);
 
 	bpf_for_each_spilled_reg(i, state, reg) {
 		if (!reg)
 			continue;
-		if (reg_is_refcounted(reg) && reg->id == id)
+		if (reg->ref_obj_id == ref_obj_id)
 			__mark_reg_unknown(reg);
 	}
 }
@@ -2820,15 +2813,20 @@ static void release_reg_references(struct bpf_verifier_env *env,
  * resources. Identify all copies of the same pointer and clear the reference.
  */
 static int release_reference(struct bpf_verifier_env *env,
-			     struct bpf_call_arg_meta *meta)
+			     int ref_obj_id)
 {
 	struct bpf_verifier_state *vstate = env->cur_state;
+	int err;
 	int i;
 
+	err = release_reference_state(cur_func(env), ref_obj_id);
+	if (err)
+		return err;
+
 	for (i = 0; i <= vstate->curframe; i++)
-		release_reg_references(env, vstate->frame[i], meta->ptr_id);
+		release_reg_references(env, vstate->frame[i], ref_obj_id);
 
-	return release_reference_state(cur_func(env), meta->ptr_id);
+	return 0;
 }
 
 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
@@ -3047,7 +3045,7 @@ static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn
 	memset(&meta, 0, sizeof(meta));
 	meta.pkt_access = fn->pkt_access;
 
-	err = check_func_proto(fn);
+	err = check_func_proto(fn, func_id);
 	if (err) {
 		verbose(env, "kernel subsystem misconfigured func %s#%d\n",
 			func_id_name(func_id), func_id);
@@ -3093,7 +3091,7 @@ static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn
 			return err;
 		}
 	} else if (is_release_function(func_id)) {
-		err = release_reference(env, &meta);
+		err = release_reference(env, meta.ref_obj_id);
 		if (err) {
 			verbose(env, "func %s#%d reference has not been acquired before\n",
 				func_id_name(func_id), func_id);
@@ -3154,8 +3152,10 @@ static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn
 
 			if (id < 0)
 				return id;
-			/* For release_reference() */
+			/* For mark_ptr_or_null_reg() */
 			regs[BPF_REG_0].id = id;
+			/* For release_reference() */
+			regs[BPF_REG_0].ref_obj_id = id;
 		} else {
 			/* For mark_ptr_or_null_reg() */
 			regs[BPF_REG_0].id = ++env->id_gen;
@@ -3170,6 +3170,10 @@ static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn
 		return -EINVAL;
 	}
 
+	if (is_ptr_cast_function(func_id))
+		/* For release_reference() */
+		regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
+
 	do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
 
 	err = check_map_func_compatibility(env, meta.map_ptr, func_id);
@@ -4665,11 +4669,19 @@ static void mark_ptr_or_null_reg(struct bpf_func_state *state,
 		} else if (reg->type == PTR_TO_TCP_SOCK_OR_NULL) {
 			reg->type = PTR_TO_TCP_SOCK;
 		}
-		if (is_null || !(reg_is_refcounted(reg) ||
-				 reg_may_point_to_spin_lock(reg))) {
-			/* We don't need id from this point onwards anymore,
-			 * thus we should better reset it, so that state
-			 * pruning has chances to take effect.
+		if (is_null) {
+			/* We don't need id and ref_obj_id from this point
+			 * onwards anymore, thus we should better reset it,
+			 * so that state pruning has chances to take effect.
+			 */
+			reg->id = 0;
+			reg->ref_obj_id = 0;
+		} else if (!reg_may_point_to_spin_lock(reg)) {
+			/* For not-NULL ptr, reg->ref_obj_id will be reset
+			 * in release_reg_references().
+			 *
+			 * reg->id is still used by spin_lock ptr. Other
+			 * than spin_lock ptr type, reg->id can be reset.
 			 */
 			reg->id = 0;
 		}
@@ -4684,11 +4696,16 @@ static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
 {
 	struct bpf_func_state *state = vstate->frame[vstate->curframe];
 	struct bpf_reg_state *reg, *regs = state->regs;
+	u32 ref_obj_id = regs[regno].ref_obj_id;
 	u32 id = regs[regno].id;
 	int i, j;
 
-	if (reg_is_refcounted_or_null(&regs[regno]) && is_null)
-		release_reference_state(state, id);
+	if (ref_obj_id && ref_obj_id == id && is_null)
+		/* regs[regno] is in the " == NULL" branch.
+		 * No one could have freed the reference state before
+		 * doing the NULL check.
+		 */
+		WARN_ON_ONCE(release_reference_state(state, id));
 
 	for (i = 0; i < MAX_BPF_REG; i++)
 		mark_ptr_or_null_reg(state, &regs[i], id, is_null);

net/core/filter.c

@@ -1796,8 +1796,6 @@ static const struct bpf_func_proto bpf_skb_pull_data_proto = {
 
 BPF_CALL_1(bpf_sk_fullsock, struct sock *, sk)
 {
-	sk = sk_to_full_sk(sk);
-
 	return sk_fullsock(sk) ? (unsigned long)sk : (unsigned long)NULL;
 }
 
@@ -5266,7 +5264,7 @@ static const struct bpf_func_proto bpf_sk_release_proto = {
 	.func		= bpf_sk_release,
 	.gpl_only	= false,
 	.ret_type	= RET_INTEGER,
-	.arg1_type	= ARG_PTR_TO_SOCKET,
+	.arg1_type	= ARG_PTR_TO_SOCK_COMMON,
 };
 
 BPF_CALL_5(bpf_xdp_sk_lookup_udp, struct xdp_buff *, ctx,
@@ -5407,8 +5405,6 @@ u32 bpf_tcp_sock_convert_ctx_access(enum bpf_access_type type,
 
 BPF_CALL_1(bpf_tcp_sock, struct sock *, sk)
 {
-	sk = sk_to_full_sk(sk);
-
 	if (sk_fullsock(sk) && sk->sk_protocol == IPPROTO_TCP)
 		return (unsigned long)sk;