[replying to both of your messages here]
On Wed, Feb 21, 2024 at 3:59 AM Alexei Starovoitov alexei.starovoitov@gmail.com wrote:
On Fri, Feb 16, 2024 at 10:50:10AM +0100, Benjamin Tissoires wrote:
static bool is_rbtree_lock_required_kfunc(u32 btf_id) { return is_bpf_rbtree_api_kfunc(btf_id); @@ -12140,6 +12143,16 @@ static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn, } }
if (is_bpf_timer_set_sleepable_cb_kfunc(meta.func_id)) {err = push_callback_call(env, insn, insn_idx, meta.subprogno,set_timer_callback_state);if (err) {verbose(env, "kfunc %s#%d failed callback verification\n",func_name, meta.func_id);return err;}}All makes sense so far. Please squash all the fix and repost. It's hard to do a proper review in this shape of the patch.
Yeah, I was expecting a very quick "I know why you are crashing", not a full review here.
As far as rcu_read_lock/unlock that is done in callback... it feels buggy and unnecessary.
This rcu approach is indeed wrong, but there still needs to be some locking if bpf_timer_set_callback() or bpf_timer_set_sleepable_cb() is called while the work just started. I went with a semaphore in v3 as it seemed lightweight enough there. Please shout if you disagree :)
Anyway, I've also dropped the flags in bpf_timer_init() in v3 to only add BPF_F_TIMER_SLEEPABLE in bpf_timer_start().
V3 (not RFC) is coming.
Cheers, Benjamin
bpf prog and timer won't disappear while work is queued. array and hash map will call bpf_obj_free_timer() before going away.
And things like:
rcu_read_lock();callback_fn = rcu_dereference(t->sleepable_cb_fn);rcu_read_unlock();if (!callback_fn)return;is 99% broken. if (!callback_fn) line is UAF.