[PATCH bpf-next 1/2] bpf: Fix release_on_unlock release logic for multiple refs

[Dave Marchevsky <davemarchevsky@fb.com>]

Consider a verifier state with three acquired references, all with
release_on_unlock = true:

            idx  0 1 2
  state->refs = [2 4 6]

(with 2, 4, and 6 being the ref ids).

When bpf_spin_unlock is called, process_spin_lock will loop through all
acquired_refs and, for each ref, if it's release_on_unlock, calls
release_reference on it. That function in turn calls
release_reference_state, which removes the reference from state->refs by
swapping the reference state with the last reference state in
refs array and decrements acquired_refs count.

process_spin_lock's loop logic, which is essentially:

  for (i = 0; i < state->acquired_refs; i++) {
    if (!state->refs[i].release_on_unlock)
      continue;
    release_reference(state->refs[i].id);
  }

will fail to release release_on_unlock references which are swapped from
the end. Running this logic on our example demonstrates:

  state->refs = [2 4 6] (start of idx=0 iter)
    release state->refs[0] by swapping w/ state->refs[2]

  state->refs = [6 4]   (start of idx=1)
    release state->refs[1], no need to swap as it's the last idx

  state->refs = [6]     (start of idx=2, loop terminates)

ref_id 6 should have been removed but was skipped.

Fix this by looping from back-to-front, which results in refs that are
candidates for removal being swapped with refs which have already been
examined and kept. If we modify our initial example such that ref 6 is
not release_on_unlock and loop from the back, we'd see:

  state->refs = [2 4 6] (start of idx=2)

  state->refs = [2 4 6] (start of idx=1)

  state->refs = [2 6]   (start of idx=0)

  state->refs = [6]     (after idx=0, loop terminates)

Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
cc: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Fixes: 534e86bc6c66 ("bpf: Add 'release on unlock' logic for bpf_list_push_{front,back}")
---

I noticed this while testing ng_ds version of rbtree. Submitting
separately so that this fix can be applied before the rest of rbtree
work, as the latter will likely need a few respins.

An alternative to this fix would be to modify or add new helper
functions which enable safe release_reference in a loop. The additional
complexity of this alternative seems unnecessary to me for now as this
is currently the only place in verifier where release_reference in a
loop is used.

 kernel/bpf/verifier.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index 4e7f1d085e53..ac3e1219a7a5 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -5726,7 +5726,7 @@ static int process_spin_lock(struct bpf_verifier_env *env, int regno,
 		cur->active_lock.ptr = NULL;
 		cur->active_lock.id = 0;
 
-		for (i = 0; i < fstate->acquired_refs; i++) {
+		for (i = fstate->acquired_refs - 1; i >= 0; i--) {
 			int err;
 
 			/* Complain on error because this reference state cannot
-- 
2.30.2