* [PATCH 2/2] bpf: Invalidate dynptr clones in every call frame on release
[not found] <apD0GtLd8lXgXD2e@1wt.eu>
@ 2026-08-28 4:46 ` Xu Yunxiang
2026-08-28 5:24 ` Greg Kroah-Hartman
0 siblings, 1 reply; 2+ messages in thread
From: Xu Yunxiang @ 2026-08-28 4:46 UTC (permalink / raw)
To: Willy Tarreau
Cc: Alexei Starovoitov, Daniel Borkmann, security, andrii, memxor,
eddyz87, Amery Hung, Greg Kroah-Hartman, stable
unmark_stack_slots_dynptr() walks clones only in the submitting
pointer's frame. A program that clones a ringbuf dynptr in a subprog,
submits the clone, then submits the original in the caller is therefore
accepted: the caller's STACK_DYNPTR slot is never cleared. The second
bpf_ringbuf_commit() flips BUSY on an already-committed header and a
userspace consumer stalls.
Walk every frame and invalidate matching STACK_DYNPTR slots, the same
per-slot work destroy_if_dynptr_stack_slot() already does in one frame.
Leave the existing WARN_ON_ONCE(release_reference()) as a swallow: a
hard -EINVAL would reject legitimate programs that release once in a
different frame than they acquired (6.6/6.12/6.13), and on 6.14+ the
second lookup fails because the ref was already deleted.
Depends on "bpf: Invalidate dynptr slices by dynptr_id on release" only
for the slice helper call inside the new loop; the frame walk is
independent.
This is the minimal stable equivalent of the cross-frame half of
308c7a0ae8859. That refactor is not a viable cherry-pick for stable.
Cc: stable@vger.kernel.org # 6.6.y 6.12.y 6.18.y
Signed-off-by: Xu Yunxiang <xyx2021@mail.ustc.edu.cn>
---
kernel/bpf/verifier.c | 43 +++++++++++++++++++++
1 file changed
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -811,7 +811,7 @@
static int unmark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
{
struct bpf_func_state *state = func(env, reg);
- int spi, ref_obj_id, i;
+ int spi, ref_obj_id, i, frm;
spi = dynptr_get_spi(env, reg);
if (spi < 0)
@@ -843,22 +843,33 @@
WARN_ON_ONCE(release_reference(env, ref_obj_id));
invalidate_slices_of_dynptr(env, state->stack[spi].spilled_ptr.id);
- /* Invalidate any dynptr clones */
- for (i = 1; i < state->allocated_stack / BPF_REG_SIZE; i++) {
- if (state->stack[i].spilled_ptr.ref_obj_id != ref_obj_id)
- continue;
+ /* Invalidate any dynptr clones, including those in other frames.
+ * acquired_refs is per-frame on 6.6/6.12/6.13; a clone submitted in a
+ * callee must still clear the caller's STACK_DYNPTR slot. Do not turn
+ * a failing release_reference() into -EINVAL: that rejects legitimate
+ * cross-frame single releases on those trees.
+ */
+ for (frm = 0; frm <= env->cur_state->curframe; frm++) {
+ struct bpf_func_state *f = env->cur_state->frame[frm];
- /* it should always be the case that if the ref obj id
- * matches then the stack slot also belongs to a
- * dynptr
- */
- if (state->stack[i].slot_type[0] != STACK_DYNPTR) {
- verbose(env, "verifier internal error: misconfigured ref_obj_id\n");
- return -EFAULT;
- }
- if (state->stack[i].spilled_ptr.dynptr.first_slot) {
- invalidate_slices_of_dynptr(env, state->stack[i].spilled_ptr.id);
- invalidate_dynptr(env, state, i);
+ if (!f)
+ continue;
+ for (i = 1; i < f->allocated_stack / BPF_REG_SIZE; i++) {
+ if (f->stack[i].spilled_ptr.ref_obj_id != ref_obj_id)
+ continue;
+
+ /* it should always be the case that if the ref obj id
+ * matches then the stack slot also belongs to a
+ * dynptr
+ */
+ if (f->stack[i].slot_type[0] != STACK_DYNPTR) {
+ verbose(env, "verifier internal error: misconfigured ref_obj_id\n");
+ return -EFAULT;
+ }
+ if (f->stack[i].spilled_ptr.dynptr.first_slot) {
+ invalidate_slices_of_dynptr(env, f->stack[i].spilled_ptr.id);
+ invalidate_dynptr(env, f, i);
+ }
}
}
--
BPF-SemSan disclosure follow-up
^ permalink raw reply [flat|nested] 2+ messages in thread