bpf: Allow refcounted bpf_rb_node used in bpf_rbtree_{remove,left,right}

The bpf_rbtree_{remove,left,right} requires the root's lock to be held.
They also check the node_internal->owner is still owned by that root
before proceeding, so it is safe to allow refcounted bpf_rb_node
pointer to be used in these kfuncs.

In a bpf fq implementation which is much closer to the kernel fq,
https://lore.kernel.org/bpf/20250418224652.105998-13-martin.lau@linux.dev/,
a networking flow (allocated by bpf_obj_new) can be added to two different
rbtrees. There are cases that the flow is searched from one rbtree,
held the refcount of the flow, and then removed from another rbtree:

struct fq_flow {
	struct bpf_rb_node	fq_node;
	struct bpf_rb_node	rate_node;
	struct bpf_refcount	refcount;
	unsigned long		sk_long;
};

int bpf_fq_enqueue(...)
{
	/* ... */

	bpf_spin_lock(&root->lock);
	while (can_loop) {
		/* ... */
		if (!p)
			break;
		gc_f = bpf_rb_entry(p, struct fq_flow, fq_node);
		if (gc_f->sk_long == sk_long) {
			f = bpf_refcount_acquire(gc_f);
			break;
		}
		/* ... */
	}
	bpf_spin_unlock(&root->lock);

	if (f) {
		bpf_spin_lock(&q->lock);
		bpf_rbtree_remove(&q->delayed, &f->rate_node);
		bpf_spin_unlock(&q->lock);
	}
}

bpf_rbtree_{left,right} do not need this change but are relaxed together
with bpf_rbtree_remove instead of adding extra verifier logic
to exclude these kfuncs.

To avoid bi-sect failure, this patch also changes the selftests together.

The "rbtree_api_remove_unadded_node" is not expecting verifier's error.
The test now expects bpf_rbtree_remove(&groot, &m->node) to return NULL.
The test uses __retval(0) to ensure this NULL return value.

Some of the "only take non-owning..." failure messages are changed also.

Acked-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>
Link: https://lore.kernel.org/r/20250506015857.817950-5-martin.lau@linux.dev
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

1 files changed, 15 insertions, 14 deletions

diff --git a/tools/testing/selftests/bpf/progs/rbtree_fail.c b/tools/testing/selftests/bpf/progs/rbtree_fail.c
index dbd5eee8e25e..4acb6af2dfe3 100644
--- a/tools/testing/selftests/bpf/progs/rbtree_fail.c
+++ b/tools/testing/selftests/bpf/progs/rbtree_fail.c
@@ -69,11 +69,11 @@ long rbtree_api_nolock_first(void *ctx)
 }
 
 SEC("?tc")
-__failure __msg("rbtree_remove node input must be non-owning ref")
+__retval(0)
 long rbtree_api_remove_unadded_node(void *ctx)
 {
 	struct node_data *n, *m;
-	struct bpf_rb_node *res;
+	struct bpf_rb_node *res_n, *res_m;
 
 	n = bpf_obj_new(typeof(*n));
 	if (!n)
@@ -88,19 +88,20 @@ long rbtree_api_remove_unadded_node(void *ctx)
 	bpf_spin_lock(&glock);
 	bpf_rbtree_add(&groot, &n->node, less);
 
-	/* This remove should pass verifier */
-	res = bpf_rbtree_remove(&groot, &n->node);
-	n = container_of(res, struct node_data, node);
+	res_n = bpf_rbtree_remove(&groot, &n->node);
 
-	/* This remove shouldn't, m isn't in an rbtree */
-	res = bpf_rbtree_remove(&groot, &m->node);
-	m = container_of(res, struct node_data, node);
+	res_m = bpf_rbtree_remove(&groot, &m->node);
 	bpf_spin_unlock(&glock);
 
-	if (n)
-		bpf_obj_drop(n);
-	if (m)
-		bpf_obj_drop(m);
+	bpf_obj_drop(m);
+	if (res_n)
+		bpf_obj_drop(container_of(res_n, struct node_data, node));
+	if (res_m) {
+		bpf_obj_drop(container_of(res_m, struct node_data, node));
+		/* m was not added to the rbtree */
+		return 2;
+	}
+
 	return 0;
 }
 
@@ -178,7 +179,7 @@ err_out:
 }
 
 SEC("?tc")
-__failure __msg("rbtree_remove node input must be non-owning ref")
+__failure __msg("bpf_rbtree_remove can only take non-owning or refcounted bpf_rb_node pointer")
 long rbtree_api_add_release_unlock_escape(void *ctx)
 {
 	struct node_data *n;
@@ -202,7 +203,7 @@ long rbtree_api_add_release_unlock_escape(void *ctx)
 }
 
 SEC("?tc")
-__failure __msg("rbtree_remove node input must be non-owning ref")
+__failure __msg("bpf_rbtree_remove can only take non-owning or refcounted bpf_rb_node pointer")
 long rbtree_api_first_release_unlock_escape(void *ctx)
 {
 	struct bpf_rb_node *res;