Skip to contentSkip to conversation

bpf: allow using bpf_kptr_xchg even if the MEM_RCU flag is set

PATCH Superseded Chengkaitao 2 participants 3 messages Feb 14

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH v7 4/5 bpf: allow using bpf_kptr_xchg even if the MEM_RCU flag is set

From: Kaitao Cheng <[email protected]>

For the following scenario:

    struct tree_node {
	struct bpf_refcount ref;
	struct bpf_rb_node node;
	struct node_data __kptr * node_data;
	u64 key;
    };

This means node_data would have the type PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF | MEM_RCU.

When traversing an rbtree using bpf_rbtree_left/right, if we need to use bpf_kptr_xchg to read the __kptr pointer, we still need to follow the remove-read-add sequence.

This patch allows us to use bpf_kptr_xchg to directly read the __kptr pointer without any prior operations.

---
 kernel/bpf/verifier.c | 2 ++
 1 file changed, 2 insertions(+)
kernel/bpf/verifier.c +2 -0
index 3536a91ff8c7..5baa12af3a4e 100644
@@ -9320,6 +9320,7 @@ static const struct bpf_reg_types percpu_btf_ptr_types = {
 		PTR_TO_BTF_ID | MEM_PERCPU,
 		PTR_TO_BTF_ID | MEM_PERCPU | MEM_RCU,
 		PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED,
+		PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF | MEM_RCU,
 	}
 };
 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
@@ -9491,6 +9492,7 @@ static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
 	case PTR_TO_BTF_ID | MEM_ALLOC:
 	case PTR_TO_BTF_ID | MEM_PERCPU | MEM_ALLOC:
 	case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF:
+	case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF | MEM_RCU:
 		if (meta->func_id != BPF_FUNC_spin_lock && meta->func_id != BPF_FUNC_spin_unlock &&
 		    meta->func_id != BPF_FUNC_kptr_xchg) {
 			verifier_bug(env, "unimplemented handling of MEM_ALLOC");
Signature · 1 line
2.50.1 (Apple Git-155)
#1 lore raw

PATCH v7 5/5 selftests/bpf: Add test case for rbtree nodes that contain both bpf_refcount and kptr fields.

From: Kaitao Cheng <[email protected]>

Allow bpf_kptr_xchg to directly operate on pointers marked with NON_OWN_REF | MEM_RCU.

In the example demonstrated in this patch, as long as "struct bpf_refcount ref" exists, the __kptr pointer is guaranteed to carry the MEM_RCU flag. The ref member itself does not need to be explicitly used.

---
 .../selftests/bpf/progs/rbtree_search_kptr.c  | 123 ++++++++++++++++++
 1 file changed, 123 insertions(+)
tools/testing/selftests/bpf/progs/rbtree_search_kptr.c +123 -0
index 069fc64b0167..610aae45e2dc 100644
@@ -18,11 +18,21 @@ struct tree_node {
 	struct node_data __kptr * node_data;
 };
 
+struct tree_node_ref {
+	struct bpf_refcount ref;
+	struct bpf_rb_node node;
+	u64 key;
+	struct node_data __kptr * node_data;
+};
+
 #define private(name) SEC(".data." #name) __hidden __aligned(8)
 
 private(A) struct bpf_rb_root root __contains(tree_node, node);
 private(A) struct bpf_spin_lock lock;
 
+private(B) struct bpf_rb_root root_r __contains(tree_node_ref, node);
+private(B) struct bpf_spin_lock lock_r;
+
 static bool less(struct bpf_rb_node *a, const struct bpf_rb_node *b)
 {
 	struct tree_node *node_a, *node_b;
@@ -130,6 +140,119 @@ long rbtree_search_kptr(void *ctx)
 	return ret;
 }
 
+static bool less_r(struct bpf_rb_node *a, const struct bpf_rb_node *b)
+{
+	struct tree_node_ref *node_a, *node_b;
+
+	node_a = container_of(a, struct tree_node_ref, node);
+	node_b = container_of(b, struct tree_node_ref, node);
+
+	return node_a->key < node_b->key;
+}
+
+SEC("syscall")
+__retval(0)
+long rbtree_search_kptr_ref(void *ctx)
+{
+	struct tree_node_ref *tnode_r, *tnode_m;
+	struct bpf_rb_node *rb_n;
+	struct node_data __kptr * node_data;
+	int lookup_key  = NR_NODES / 2;
+	int lookup_data = NR_NODES / 2;
+	int i, data, ret = 0;
+
+	for (i = 0; i < NR_NODES && can_loop; i++) {
+		tnode_r = bpf_obj_new(typeof(*tnode_r));
+		if (!tnode_r)
+			return __LINE__;
+
+		node_data = bpf_obj_new(typeof(*node_data));
+		if (!node_data) {
+			bpf_obj_drop(tnode_r);
+			return __LINE__;
+		}
+
+		tnode_r->key = i;
+		node_data->data = i;
+
+		node_data = bpf_kptr_xchg(&tnode_r->node_data, node_data);
+		if (node_data)
+			bpf_obj_drop(node_data);
+
+		/* Unused reference */
+		tnode_m = bpf_refcount_acquire(tnode_r);
+		if (!tnode_m)
+			return __LINE__;
+
+		bpf_spin_lock(&lock_r);
+		bpf_rbtree_add(&root_r, &tnode_r->node, less_r);
+		bpf_spin_unlock(&lock_r);
+
+		bpf_obj_drop(tnode_m);
+	}
+
+	bpf_spin_lock(&lock_r);
+	rb_n = bpf_rbtree_root(&root_r);
+	while (rb_n && can_loop) {
+		tnode_r = container_of(rb_n, struct tree_node_ref, node);
+		node_data = bpf_kptr_xchg(&tnode_r->node_data, NULL);
+		if (!node_data) {
+			ret = __LINE__;
+			goto fail;
+		}
+
+		data = node_data->data;
+		node_data = bpf_kptr_xchg(&tnode_r->node_data, node_data);
+		if (node_data) {
+			bpf_spin_unlock(&lock_r);
+			bpf_obj_drop(node_data);
+			return __LINE__;
+		}
+
+		if (lookup_key == tnode_r->key) {
+			if (data == lookup_data)
+				break;
+
+			ret = __LINE__;
+			goto fail;
+		}
+
+		if (lookup_key < tnode_r->key)
+			rb_n = bpf_rbtree_left(&root_r, rb_n);
+		else
+			rb_n = bpf_rbtree_right(&root_r, rb_n);
+	}
+	bpf_spin_unlock(&lock_r);
+
+	while (can_loop) {
+		bpf_spin_lock(&lock_r);
+		rb_n = bpf_rbtree_first(&root_r);
+		if (!rb_n) {
+			bpf_spin_unlock(&lock_r);
+			return 0;
+		}
+
+		rb_n = bpf_rbtree_remove(&root_r, rb_n);
+		if (!rb_n) {
+			ret = __LINE__;
+			goto fail;
+		}
+		bpf_spin_unlock(&lock_r);
+
+		tnode_r = container_of(rb_n, struct tree_node_ref, node);
+
+		node_data = bpf_kptr_xchg(&tnode_r->node_data, NULL);
+		if (node_data)
+			bpf_obj_drop(node_data);
+
+		bpf_obj_drop(tnode_r);
+	}
+
+	return 0;
+fail:
+	bpf_spin_unlock(&lock_r);
+	return ret;
+}
 
 SEC("syscall")
 __failure __msg("R1 type=scalar expected=map_value, ptr_, ptr_")
Signature · 1 line
2.50.1 (Apple Git-155)
#2 lore raw
In reply to Chengkaitao
Was the new type added to the correct array? This hunk adds the entry to percpu_btf_ptr_types, which backs ARG_PTR_TO_PERCPU_BTF_ID in compatible_reg_types[]. That array is used by bpf_per_cpu_ptr…

kernel/bpf/verifier.c line 9305

Quoted 10 lines
kernel/bpf/verifier.c +1 -0
@@ -9300,6 +9300,7 @@
 		PTR_TO_BTF_ID | MEM_PERCPU,
 		PTR_TO_BTF_ID | MEM_PERCPU | MEM_RCU,
 		PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED,
+		PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF | MEM_RCU,
 	}
 };

Was the new type added to the correct array? This hunk adds the entry to percpu_btf_ptr_types, which backs ARG_PTR_TO_PERCPU_BTF_ID in compatible_reg_types[]. That array is used by bpf_per_cpu_ptr and bpf_this_cpu_ptr, not by bpf_kptr_xchg.

For bpf_kptr_xchg arg1 (ARG_KPTR_XCHG_DEST), the verifier looks up kptr_xchg_dest_types instead:

kernel/bpf/verifier.c:check_reg_type() {
    ...
    compatible = compatible_reg_types[base_type(arg_type)];
    ...
    for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
        expected = compatible->types[i];
        ...
        if (type == expected)
            goto found;
    }
    ...
    return -EACCES;

kptr_xchg_dest_types currently contains only:

static const struct bpf_reg_types kptr_xchg_dest_types = {
.types = {
PTR_TO_MAP_VALUE,
PTR_TO_BTF_ID | MEM_ALLOC,
PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF,
}
};

A register with PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF | MEM_RCU will not match any of these entries, so check_reg_type() returns -EACCES before ever reaching the found: label and the switch statement below.

It looks like the entry should be added to kptr_xchg_dest_types rather than percpu_btf_ptr_types. Alexei made the same suggestion on v6 of this series:

https://lore.kernel.org/bpf/CAADnVQJdwqeoqi+mPjQjmNbUqz2RxAjq-dVsGb7c_MCqpqmwhA@mail.gmail.com/

kernel/bpf/verifier.c line 9481

Quoted 7 lines
kernel/bpf/verifier.c +1 -0
@@ -9475,6 +9476,7 @@
 	case PTR_TO_BTF_ID | MEM_ALLOC:
 	case PTR_TO_BTF_ID | MEM_PERCPU | MEM_ALLOC:
 	case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF:
+	case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF | MEM_RCU:
 		if (meta->func_id != BPF_FUNC_spin_lock && meta->func_id != BPF_FUNC_spin_unlock &&
 		    meta->func_id != BPF_FUNC_kptr_xchg) {

This switch case is correct, but it can only be reached via the found:
label after the type-matching loop succeeds. Since the first hunk adds the type to the wrong array, the loop never matches for bpf_kptr_xchg and this case is unreachable for its intended purpose.

---
AI reviewed your patch. Please fix the bug or email reply why it's not a bug.
See: https://github.com/kernel-patches/vmtest/blob/master/ci/claude/README.md

CI run summary: https://github.com/kernel-patches/bpf/actions/runs/22011467184

AI-authorship-score: low
AI-authorship-explanation: The patch went through 7 revisions on the mailing list with human reviewer interaction, and the type of mistake (adding to the wrong array) is characteristic of human error in a codebase with many similarly structured type arrays.
issues-found: 1
issue-severity-score: medium
issue-severity-explanation: The new type was added to percpu_btf_ptr_types instead of kptr_xchg_dest_types, so the patch does not achieve its stated goal of allowing bpf_kptr_xchg on MEM_RCU pointers; there is no user-visible crash or security issue, but the feature is silently broken.