Re: [PATCH bpf v2 1/2] bpf, lpm_trie: Allow access from sleepable BPF programs

["Emil Tsalapatis" <emil@etsalapatis.com>]

On Tue Jun 9, 2026 at 9:55 AM EDT, Vlad Poenaru wrote:
> trie_lookup_elem() annotates its rcu_dereference_check() walks with
> only rcu_read_lock_bh_held().  Because rcu_dereference_check(p, c)
> resolves to "c || rcu_read_lock_held()", this passes for XDP/NAPI and
> classic RCU readers but fails for sleepable BPF programs, which enter
> via __bpf_prog_enter_sleepable() and hold only rcu_read_lock_trace().
>
> trie_update_elem() and trie_delete_elem() have the same problem in a
> different form: they walk the trie with plain rcu_dereference(), which
> asserts rcu_read_lock_held() unconditionally.  Both are reachable from
> sleepable BPF programs via the bpf_map_update_elem / bpf_map_delete_elem
> helpers, and from the syscall path under classic rcu_read_lock().  In
> the writer paths the trie is actually protected by trie->lock (an
> rqspinlock taken across the walk); we never relied on the RCU read-side
> lock to keep nodes alive there.
>
> A sleepable LSM hook that ends up touching an LPM trie therefore
> triggers lockdep on debug kernels:
>
>   =============================
>   WARNING: suspicious RCU usage
>   7.1.0-... Tainted: G            E
>   -----------------------------
>   kernel/bpf/lpm_trie.c:249 suspicious rcu_dereference_check() usage!
>   1 lock held by net_tests/540:
>    #0: (rcu_tasks_trace_srcu_struct){....}-{0:0},
>        at: __bpf_prog_enter_sleepable+0x26/0x280
>   Call Trace:
>    dump_stack_lvl
>    lockdep_rcu_suspicious
>    trie_lookup_elem
>    bpf_prog_..._enforce_security_socket_connect
>    bpf_trampoline_...
>    security_socket_connect
>    __sys_connect
>    do_syscall_64
>
> This is lockdep-only -- no UAF, since Tasks Trace RCU does serialize
> against the trie's reclaim path -- but it spams the console once per
> distinct callsite on every debug kernel running a sleepable BPF LSM
> that touches an LPM trie, which is increasingly common.
>
> For the lookup path, switch the rcu_dereference_check() annotation
> from rcu_read_lock_bh_held() to bpf_rcu_lock_held(), which accepts all
> three contexts (classic, BH, Tasks Trace).  Other map types already
> follow this convention.
>
> For trie_update_elem() and trie_delete_elem(), annotate the walks as
> rcu_dereference_protected(*p, 1) -- matching trie_free() in the same
> file -- since trie->lock is held across the walk.  rqspinlock has no
> lockdep_map, so the predicate degenerates to '1' rather than
> lockdep_is_held(&trie->lock); the protection is real but not
> machine-verifiable.  trie_get_next_key() also uses bare
> rcu_dereference() but is reachable only from the BPF syscall, which
> holds classic rcu_read_lock() before dispatching, so it is left
> untouched.
>
> Fixes: 694cea395fde ("bpf: Allow RCU-protected lookups to happen from bh context")
> Cc: stable@vger.kernel.org
> Signed-off-by: Vlad Poenaru <vlad.wing@gmail.com>

Reviewed-by: Emil Tsalapatis <emil@etsalapatis.com>

> ---
>  kernel/bpf/lpm_trie.c | 8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
>
> diff --git a/kernel/bpf/lpm_trie.c b/kernel/bpf/lpm_trie.c
> index 0f57608b385d..4d6f25db9ba1 100644
> --- a/kernel/bpf/lpm_trie.c
> +++ b/kernel/bpf/lpm_trie.c
> @@ -246,7 +246,7 @@ static void *trie_lookup_elem(struct bpf_map *map, void *_key)
>  
>  	/* Start walking the trie from the root node ... */
>  
> -	for (node = rcu_dereference_check(trie->root, rcu_read_lock_bh_held());
> +	for (node = rcu_dereference_check(trie->root, bpf_rcu_lock_held());
>  	     node;) {
>  		unsigned int next_bit;
>  		size_t matchlen;
> @@ -280,7 +280,7 @@ static void *trie_lookup_elem(struct bpf_map *map, void *_key)
>  		 */
>  		next_bit = extract_bit(key->data, node->prefixlen);
>  		node = rcu_dereference_check(node->child[next_bit],
> -					     rcu_read_lock_bh_held());
> +					     bpf_rcu_lock_held());
>  	}
>  
>  	if (!found)
> @@ -359,7 +359,7 @@ static long trie_update_elem(struct bpf_map *map,
>  	 */
>  	slot = &trie->root;
>  
> -	while ((node = rcu_dereference(*slot))) {
> +	while ((node = rcu_dereference_protected(*slot, 1))) {
>  		matchlen = longest_prefix_match(trie, node, key);
>  
>  		if (node->prefixlen != matchlen ||
> @@ -482,7 +482,7 @@ static long trie_delete_elem(struct bpf_map *map, void *_key)
>  	trim = &trie->root;
>  	trim2 = trim;
>  	parent = NULL;
> -	while ((node = rcu_dereference(*trim))) {
> +	while ((node = rcu_dereference_protected(*trim, 1))) {
>  		matchlen = longest_prefix_match(trie, node, key);
>  
>  		if (node->prefixlen != matchlen ||