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 replaced with ref 7,
which is _not_ release_on_unlock, and loop from the back, we'd see:
state->refs = [2 4 7] (start of idx=2)
state->refs = [2 4 7] (start of idx=1)
state->refs = [2 7] (start of idx=0, refs 7 and 4 swapped)
state->refs = [7] (after idx=0, 7 and 2 swapped, loop terminates)
Signed-off-by: Dave Marchevsky <davemarchevsky@xxxxxx>
Acked-by: Yonghong Song <yhs@xxxxxx>
cc: Kumar Kartikeya Dwivedi <memxor@xxxxxxxxx>
Fixes: 534e86bc6c66 ("bpf: Add 'release on unlock' logic for bpf_list_push_{front,back}")
---
v1 -> v2 lore.kernel.org/bpf/b5d46fd5-2693-cd46-9515-700fef1a110b@xxxxxxxx:
* Update second example in patch summary to use a new ref_id for the
non-release_on_unlock ref. (Yonghong)
* Add Yonghong's ack
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 d0ecc0b18b20..b0db9c10567b 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -5738,7 +5738,7 @@ 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
