On Thu, Sep 7, 2023 at 5:02 AM Chuyi Zhou <zhouchuyi@xxxxxxxxxxxxx> wrote:
>
> Hello,
> 在 2023/9/7 01:17, Alexei Starovoitov 写道:
> [...cut...]
> >>> This iter can be used in all ctx-s which is nice, but let's
> >>> make the verifier enforce rcu_read_lock/unlock done by bpf prog
> >>> instead of doing in the ctor/dtor of iter, since
> >>> in sleepable progs the verifier won't recognize that body is RCU CS.
> >>> We'd need to teach the verifier to allow bpf_iter_process_new()
> >>> inside in_rcu_cs() and make sure there is no rcu_read_unlock
> >>> while BPF_ITER_STATE_ACTIVE.
> >>> bpf_iter_process_destroy() would become a nop.
> >>
> >> Thanks for your review!
> >>
> >> I think bpf_iter_process_{new, next, destroy} should be protected by
> >> bpf_rcu_read_lock/unlock explicitly whether the prog is sleepable or
> >> not, right?
> >
> > Correct. By explicit bpf_rcu_read_lock() in case of sleepable progs
> > or just by using them in normal bpf progs that have implicit rcu_read_lock()
> > done before calling into them.
> Thanks for your explanation, I missed the latter.
> >
> >> I'm not very familiar with the BPF verifier, but I believe
> >> there is still a risk in directly calling these kfuns even if
> >> in_rcu_cs() is true.
> >>
> >> Maby what we actually need here is to enforce BPF verifier to check
> >> env->cur_state->active_rcu_lock is true when we want to call these kfuncs.
> >
> > active_rcu_lock means explicit bpf_rcu_read_lock.
> > Currently we do allow bpf_rcu_read_lock in non-sleepable, but it's pointless.
> >
> > Technically we can extend the check:
> > if (in_rbtree_lock_required_cb(env) && (rcu_lock ||
> > rcu_unlock)) {
> > verbose(env, "Calling
> > bpf_rcu_read_{lock,unlock} in unnecessary rbtree callback\n");
> > return -EACCES;
> > }
> > to discourage their use in all non-sleepable, but it will break some progs.
> >
> > I think it's ok to check in_rcu_cs() to allow bpf_iter_process_*().
> > If bpf prog adds explicit and unnecessary bpf_rcu_read_lock() around
> > the iter ops it won't do any harm.
> > Just need to make sure that rcu unlock logic:
> > } else if (rcu_unlock) {
> > bpf_for_each_reg_in_vstate(env->cur_state,
> > state, reg, ({
> > if (reg->type & MEM_RCU) {
> > reg->type &= ~(MEM_RCU |
> > PTR_MAYBE_NULL);
> > reg->type |= PTR_UNTRUSTED;
> > }
> > }));
> > clears iter state that depends on rcu.
> >
> > I thought about changing mark_stack_slots_iter() to do
> > st->type = PTR_TO_STACK | MEM_RCU;
> > so that the above clearing logic kicks in,
> > but it might be better to have something iter specific.
> > is_iter_reg_valid_init() should probably be changed to
> > make sure reg->type is not UNTRUSTED.
> >
> Maybe it's something looks like the following?
>
> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> index bb78212fa5b2..9185c4a40a21 100644
> --- a/kernel/bpf/verifier.c
> +++ b/kernel/bpf/verifier.c
> @@ -1172,7 +1172,15 @@ static bool is_dynptr_type_expected(struct
> bpf_verifier_env *env, struct bpf_reg
>
> static void __mark_reg_known_zero(struct bpf_reg_state *reg);
>
> +static bool in_rcu_cs(struct bpf_verifier_env *env);
> +
> +/* check whether we are using bpf_iter_process_*() or bpf_iter_css_*() */
> +static bool is_iter_need_rcu(struct bpf_kfunc_call_arg_meta *meta)
> +{
> +
> +}
> static int mark_stack_slots_iter(struct bpf_verifier_env *env,
> + struct bpf_kfunc_call_arg_meta *meta,
> struct bpf_reg_state *reg, int insn_idx,
> struct btf *btf, u32 btf_id, int nr_slots)
> {
> @@ -1193,6 +1201,12 @@ static int mark_stack_slots_iter(struct
> bpf_verifier_env *env,
>
> __mark_reg_known_zero(st);
> st->type = PTR_TO_STACK; /* we don't have dedicated reg
> type */
> + if (is_iter_need_rcu(meta)) {
> + if (in_rcu_cs(env))
> + st->type |= MEM_RCU;
> + else
> + st->type |= PTR_UNTRUSTED;
> + }
> st->live |= REG_LIVE_WRITTEN;
> st->ref_obj_id = i == 0 ? id : 0;
> st->iter.btf = btf;
> @@ -1281,6 +1295,8 @@ static bool is_iter_reg_valid_init(struct
> bpf_verifier_env *env, struct bpf_reg_
> struct bpf_stack_state *slot = &state->stack[spi - i];
> struct bpf_reg_state *st = &slot->spilled_ptr;
>
> + if (st->type & PTR_UNTRUSTED)
> + return false;
Yep. All makes sense to me.
