[RFCv2 PATCH bpf-next 12/18] bpf: Add OBJ_NON_OWNING_REF type flag

[Dave Marchevsky <davemarchevsky@fb.com>]

Consider a pointer to a type that would normally need acquire / release
semantics to be safely held. There may be scenarios where such a pointer
can be safely held without the need to acquire a reference.

For example, although a PTR_TO_BTF_ID for a rbtree_map node is released
via bpf_rbtree_add helper, the helper doesn't change the address of the
node and must be called with the rbtree_map's spinlock held. Since the
only way to remove a node from the rbtree - bpf_rbtree_remove helper -
requires the same lock, the newly-added node cannot be removed by a
concurrently-running program until the lock is released. Therefore it is
safe to hold a reference to this node until bpf_rbtree_unlock is called.

This patch introduces a new type flag and associated verifier logic to
handle such "non-owning" references.

Currently the only usecase I have is the rbtree example above, so the
verifier logic is straightforward:
  * Tag return types of bpf_rbtree_{add,find} with OBJ_NON_OWNING_REF
    * These both require the rbtree lock to be held to return anything
    non-NULL
    * Since ret type for both is PTR_TO_BTF_ID_OR_NULL, if lock is not
    held and NULL is returned, existing mark_ptr_or_null_reg logic
    will clear reg type.
    * So if mark_ptr_or_null_reg logic turns the returned reg into a
    PTR_TO_BTF_ID | OBJ_NON_OWNING_REF, verifier knows lock is held.

  * When the lock is released the verifier invalidates any regs holding
  non owning refs similarly to existing release_reference logic - but no
  need to clear ref_obj_id as an 'owning' reference was never acquired.

[ TODO: Currently the invalidation logic in
clear_rbtree_node_non_owning_refs is not parametrized by map so
unlocking any rbtree lock will invalidate all non-owning refs ]

Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
---
 include/linux/bpf.h   |  1 +
 kernel/bpf/rbtree.c   |  4 +--
 kernel/bpf/verifier.c | 65 +++++++++++++++++++++++++++++++++++++++----
 3 files changed, 62 insertions(+), 8 deletions(-)

diff --git a/include/linux/bpf.h b/include/linux/bpf.h
index b762c6b3dcfb..f164bd6e2f3a 100644
--- a/include/linux/bpf.h
+++ b/include/linux/bpf.h
@@ -415,6 +415,7 @@ enum bpf_type_flag {
 	/* Size is known at compile time. */
 	MEM_FIXED_SIZE		= BIT(10 + BPF_BASE_TYPE_BITS),
 
+	OBJ_NON_OWNING_REF	= BIT(11 + BPF_BASE_TYPE_BITS),
 	__BPF_TYPE_FLAG_MAX,
 	__BPF_TYPE_LAST_FLAG	= __BPF_TYPE_FLAG_MAX - 1,
 };
diff --git a/kernel/bpf/rbtree.c b/kernel/bpf/rbtree.c
index b5d158254de6..cc89639df8a2 100644
--- a/kernel/bpf/rbtree.c
+++ b/kernel/bpf/rbtree.c
@@ -144,7 +144,7 @@ BPF_CALL_3(bpf_rbtree_add, struct bpf_map *, map, void *, value, void *, cb)
 const struct bpf_func_proto bpf_rbtree_add_proto = {
 	.func = bpf_rbtree_add,
 	.gpl_only = true,
-	.ret_type = RET_PTR_TO_BTF_ID_OR_NULL,
+	.ret_type = RET_PTR_TO_BTF_ID_OR_NULL | OBJ_NON_OWNING_REF,
 	.ret_btf_id = BPF_PTR_POISON,
 	.arg1_type = ARG_CONST_MAP_PTR,
 	.arg2_type = ARG_PTR_TO_BTF_ID | OBJ_RELEASE,
@@ -167,7 +167,7 @@ BPF_CALL_3(bpf_rbtree_find, struct bpf_map *, map, void *, key, void *, cb)
 const struct bpf_func_proto bpf_rbtree_find_proto = {
 	.func = bpf_rbtree_find,
 	.gpl_only = true,
-	.ret_type = RET_PTR_TO_BTF_ID_OR_NULL,
+	.ret_type = RET_PTR_TO_BTF_ID_OR_NULL | OBJ_NON_OWNING_REF,
 	.ret_btf_id = BPF_PTR_POISON,
 	.arg1_type = ARG_CONST_MAP_PTR,
 	.arg2_type = ARG_ANYTHING,
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index 3c9af1047d80..26aa228fa860 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -469,6 +469,11 @@ static bool type_is_rdonly_mem(u32 type)
 	return type & MEM_RDONLY;
 }
 
+static bool type_is_non_owning_ref(u32 type)
+{
+	return type & OBJ_NON_OWNING_REF;
+}
+
 static bool type_may_be_null(u32 type)
 {
 	return type & PTR_MAYBE_NULL;
@@ -595,7 +600,9 @@ static bool function_returns_rbtree_node(enum bpf_func_id func_id)
 static const char *reg_type_str(struct bpf_verifier_env *env,
 				enum bpf_reg_type type)
 {
-	char postfix[16] = {0}, prefix[32] = {0};
+	char postfix[32] = {0}, prefix[32] = {0};
+	unsigned int postfix_idx = 0;
+
 	static const char * const str[] = {
 		[NOT_INIT]		= "?",
 		[SCALAR_VALUE]		= "scalar",
@@ -620,11 +627,18 @@ static const char *reg_type_str(struct bpf_verifier_env *env,
 		[PTR_TO_SPIN_LOCK]	= "spin_lock",
 	};
 
-	if (type & PTR_MAYBE_NULL) {
+	if (type_may_be_null(type)) {
 		if (base_type(type) == PTR_TO_BTF_ID)
-			strncpy(postfix, "or_null_", 16);
+			postfix_idx += strlcpy(postfix + postfix_idx, "or_null_", 32 - postfix_idx);
 		else
-			strncpy(postfix, "_or_null", 16);
+			postfix_idx += strlcpy(postfix + postfix_idx, "_or_null", 32 - postfix_idx);
+	}
+
+	if (type_is_non_owning_ref(type)) {
+		if (base_type(type) == PTR_TO_BTF_ID)
+			postfix_idx += strlcpy(postfix + postfix_idx, "non_own_", 32 - postfix_idx);
+		else
+			postfix_idx += strlcpy(postfix + postfix_idx, "_non_own", 32 - postfix_idx);
 	}
 
 	if (type & MEM_RDONLY)
@@ -5758,7 +5772,14 @@ static const struct bpf_reg_types int_ptr_types = {
 static const struct bpf_reg_types spin_lock_types = {
 	.types = {
 		PTR_TO_MAP_VALUE,
-		PTR_TO_SPIN_LOCK
+		PTR_TO_SPIN_LOCK,
+	},
+};
+
+static const struct bpf_reg_types btf_ptr_types = {
+	.types = {
+		PTR_TO_BTF_ID,
+		PTR_TO_BTF_ID | OBJ_NON_OWNING_REF,
 	},
 };
 
@@ -5767,7 +5788,6 @@ static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
 static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM | MEM_ALLOC } };
 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
-static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
 static const struct bpf_reg_types percpu_btf_ptr_types = { .types = { PTR_TO_BTF_ID | MEM_PERCPU } };
 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
@@ -6723,6 +6743,33 @@ static int release_reference(struct bpf_verifier_env *env,
 	return 0;
 }
 
+static void clear_non_owning_ref_regs(struct bpf_verifier_env *env,
+				      struct bpf_func_state *state)
+{
+	struct bpf_reg_state *regs = state->regs, *reg;
+	int i;
+
+	for (i = 0; i < MAX_BPF_REG; i++)
+		if (type_is_non_owning_ref(regs[i].type))
+			mark_reg_unknown(env, regs, i);
+
+	bpf_for_each_spilled_reg(i, state, reg) {
+		if (!reg)
+			continue;
+		if (type_is_non_owning_ref(reg->type))
+			__mark_reg_unknown(env, reg);
+	}
+}
+
+static void clear_rbtree_node_non_owning_refs(struct bpf_verifier_env *env)
+{
+	struct bpf_verifier_state *vstate = env->cur_state;
+	int i;
+
+	for (i = 0; i <= vstate->curframe; i++)
+		clear_non_owning_ref_regs(env, vstate->frame[i]);
+}
+
 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
 				    struct bpf_reg_state *regs)
 {
@@ -7584,6 +7631,12 @@ static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn
 			return -EFAULT;
 		}
 		break;
+	case BPF_FUNC_rbtree_unlock:
+		/* TODO clear_rbtree_node_non_owning_refs calls should be
+		 * parametrized by base_type or ideally by owning map
+		 */
+		clear_rbtree_node_non_owning_refs(env);
+		break;
 	}
 
 	if (err)
-- 
2.30.2