Re: [EXT] Re: [v3,net-next 1/4] net: qos: introduce a gate control flow action

[Vlad Buslov]

On Thu 23 Apr 2020 at 12:15, Po Liu <po.liu@xxxxxxx> wrote:
> Hi Vlad Buslov,
>
>> > >> > +static enum hrtimer_restart gate_timer_func(struct hrtimer *timer)
>> {
>> > >> > +     struct gate_action *gact = container_of(timer, struct
>> gate_action,
>> > >> > +                                             hitimer);
>> > >> > +     struct tcf_gate_params *p = get_gate_param(gact);
>> > >> > +     struct tcfg_gate_entry *next;
>> > >> > +     ktime_t close_time, now;
>> > >> > +
>> > >> > +     spin_lock(&gact->entry_lock);
>> > >> > +
>> > >> > +     next = rcu_dereference_protected(gact->next_entry,
>> > >> > +
>> > >> > + lockdep_is_held(&gact->entry_lock));
>> > >> > +
>> > >> > +     /* cycle start, clear pending bit, clear total octets */
>> > >> > +     gact->current_gate_status = next->gate_state ?
>> > >> GATE_ACT_GATE_OPEN : 0;
>> > >> > +     gact->current_entry_octets = 0;
>> > >> > +     gact->current_max_octets = next->maxoctets;
>> > >> > +
>> > >> > +     gact->current_close_time = ktime_add_ns(gact-
>> > >current_close_time,
>> > >> > +                                             next->interval);
>> > >> > +
>> > >> > +     close_time = gact->current_close_time;
>> > >> > +
>> > >> > +     if (list_is_last(&next->list, &p->entries))
>> > >> > +             next = list_first_entry(&p->entries,
>> > >> > +                                     struct tcfg_gate_entry, list);
>> > >> > +     else
>> > >> > +             next = list_next_entry(next, list);
>> > >> > +
>> > >> > +     now = gate_get_time(gact);
>> > >> > +
>> > >> > +     if (ktime_after(now, close_time)) {
>> > >> > +             ktime_t cycle, base;
>> > >> > +             u64 n;
>> > >> > +
>> > >> > +             cycle = p->tcfg_cycletime;
>> > >> > +             base = ns_to_ktime(p->tcfg_basetime);
>> > >> > +             n = div64_u64(ktime_sub_ns(now, base), cycle);
>> > >> > +             close_time = ktime_add_ns(base, (n + 1) * cycle);
>> > >> > +     }
>> > >> > +
>> > >> > +     rcu_assign_pointer(gact->next_entry, next);
>> > >> > +     spin_unlock(&gact->entry_lock);
>> > >>
>> > >> I have couple of question about synchronization here:
>> > >>
>> > >> - Why do you need next_entry to be rcu pointer? It is only assigned
>> > >> here with entry_lock protection and in init code before action is
>> > >> visible to concurrent users. I don't see any unlocked rcu-protected
>> > >> readers here that could benefit from it.
>> > >>
>> > >> - Why create dedicated entry_lock instead of using already existing
>> > >> per- action tcf_lock?
>> > >
>> > > Will try to use the tcf_lock for verification.
>
> I think I added entry_lock was that I can't get the tc_action common parameter in this  timer function. If I insist to use the tcf_lock, I have to move the hrtimer to struct tcf_gate which has tc_action common.
> What do you think?

Well, if you use tcf_lock instead of rcu to sync with fastpath, the you
don't need to implement struct gate_action as standalone object pointed
to by rcu pointer from base structure that includes tc_action common.
All the necessary data can be included in tcf_gate structure directly
and used from both timer and action fastpath. See pedit for example of
action that doesn't use rcu for fastpath.

>
>> > > The thoughts came from that the timer period arrived then check
>> > > through the list and then update next time would take much more
>> time.
>> > > Action function would be busy when traffic. So use a separate lock
>> > > here for
>> > >
>> > >>
>> > >> > +
>> > >> > +     hrtimer_set_expires(&gact->hitimer, close_time);
>> > >> > +
>> > >> > +     return HRTIMER_RESTART;
>> > >> > +}
>> > >> > +
>> > >> > +static int tcf_gate_act(struct sk_buff *skb, const struct tc_action *a,
>> > >> > +                     struct tcf_result *res) {
>> > >> > +     struct tcf_gate *g = to_gate(a);
>> > >> > +     struct gate_action *gact;
>> > >> > +     int action;
>> > >> > +
>> > >> > +     tcf_lastuse_update(&g->tcf_tm);
>> > >> > +     bstats_cpu_update(this_cpu_ptr(g->common.cpu_bstats), skb);
>> > >> > +
>> > >> > +     action = READ_ONCE(g->tcf_action);
>> > >> > +     rcu_read_lock();
>> > >>
>> > >> Action fastpath is already rcu read lock protected, you don't need
>> > >> to manually obtain it.
>> > >
>> > > Will be removed.
>> > >
>> > >>
>> > >> > +     gact = rcu_dereference_bh(g->actg);
>> > >> > +     if (unlikely(gact->current_gate_status & GATE_ACT_PENDING))
>> > >> > + {
>> > >>
>> > >> Can't current_gate_status be concurrently modified by timer callback?
>> > >> This function doesn't use entry_lock to synchronize with timer.
>> > >
>> > > Will try tcf_lock either.
>> > >
>> > >>
>> > >> > +             rcu_read_unlock();
>> > >> > +             return action;
>> > >> > +     }
>> > >> > +
>> > >> > +     if (!(gact->current_gate_status & GATE_ACT_GATE_OPEN))
>> > >>
>> > >> ...and here
>> > >>
>> > >> > +             goto drop;
>> > >> > +
>> > >> > +     if (gact->current_max_octets >= 0) {
>> > >> > +             gact->current_entry_octets += qdisc_pkt_len(skb);
>> > >> > +             if (gact->current_entry_octets >
>> > >> > + gact->current_max_octets) {
>> > >>
>> > >> here also.
>> > >>
>> > >> > +
>> > >> > + qstats_overlimit_inc(this_cpu_ptr(g->common.cpu_qstats));
>> > >>
>> > >> Please use tcf_action_inc_overlimit_qstats() and other wrappers for
>> > stats.
>> > >> Otherwise it will crash if user passes
>> > TCA_ACT_FLAGS_NO_PERCPU_STATS
>> > >> flag.
>> > >
>> > > The tcf_action_inc_overlimit_qstats() can't show limit counts in tc
>> > > show
>> > command. Is there anything need to do?
>> >
>> > What do you mean? Internally tcf_action_inc_overlimit_qstats() just
>> > calls qstats_overlimit_inc, if cpu_qstats percpu counter is not NULL:
>> >
>> >
>> >         if (likely(a->cpu_qstats)) {
>> >                 qstats_overlimit_inc(this_cpu_ptr(a->cpu_qstats));
>> >                 return;
>> >         }
>> >
>> > Is there a subtle bug somewhere in this function?
>> 
>> Sorry, I updated using the tcf_action_*, and the counting is ok. I moved
>> back to the qstats_overlimit_inc() because tcf_action_* () include the
>> spin_lock(&a->tcfa_lock).
>> I would update to  tcf_action_* () increate.
>> 
>> >
>> > >
>> > > Br,
>> > > Po Liu
>> 
>> Thanks a lot.
>> 
>> Br,
>> Po Liu
>
> Thanks a lot.
>
> Br,
> Po Liu