Re: [PATCH v6 1/4] rcu: Make call_rcu() lazy to save power

[Joel Fernandes]

> On Sep 24, 2022, at 7:28 PM, Joel Fernandes <joel@xxxxxxxxxxxxxxxxx> wrote:
> 
> Hi Frederic, thanks for the response, replies
> below courtesy fruit company’s device:
> 
>>> On Sep 24, 2022, at 6:46 PM, Frederic Weisbecker <frederic@xxxxxxxxxx> wrote:
>>> 
>>> On Thu, Sep 22, 2022 at 10:01:01PM +0000, Joel Fernandes (Google) wrote:
>>> @@ -3902,7 +3939,11 @@ static void rcu_barrier_entrain(struct rcu_data *rdp)
>>>   rdp->barrier_head.func = rcu_barrier_callback;
>>>   debug_rcu_head_queue(&rdp->barrier_head);
>>>   rcu_nocb_lock(rdp);
>>> -    WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
>>> +    /*
>>> +     * Flush the bypass list, but also wake up the GP thread as otherwise
>>> +     * bypass/lazy CBs maynot be noticed, and can cause real long delays!
>>> +     */
>>> +    WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies, FLUSH_BP_WAKE));
>> 
>> This fixes an issue that goes beyond lazy implementation. It should be done
>> in a separate patch, handling rcu_segcblist_entrain() as well, with "Fixes: " tag.
> 
> I wanted to do that, however on discussion with
> Paul I thought of making this optimization only for
> all lazy bypass CBs. That makes it directly related
> this patch since the laziness notion is first
> introduced here. On the other hand I could make
> this change in a later patch since we are not
> super bisectable anyway courtesy of the last
> patch (which is not really an issue if the CONFIG
> is kept off during someone’s bisection.

Or are we saying it’s worth doing the wake up for rcu barrier even for regular bypass CB? That’d save 2 jiffies on rcu barrier. If we agree it’s needed, then yes splitting the patch makes sense.

Please let me know your opinions, thanks,

 - Joel




> 
>> And then FLUSH_BP_WAKE is probably not needed anymore. 
> 
> It is needed as the API is in tree_nocb.h and we
> have to have that handle the details of laziness
> there rather than tree.c. We could add new apis
> to get rid of flag but it’s cleaner (and Paul seemed
> to be ok with it).
> 
>>>   if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head)) {
>>>       atomic_inc(&rcu_state.barrier_cpu_count);
>>>   } else {
>>> @@ -269,10 +294,14 @@ static void wake_nocb_gp_defer(struct rcu_data *rdp, int waketype,
>>>   raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
>>> 
>>>   /*
>>> -     * Bypass wakeup overrides previous deferments. In case
>>> -     * of callback storm, no need to wake up too early.
>>> +     * Bypass wakeup overrides previous deferments. In case of
>>> +     * callback storm, no need to wake up too early.
>>>    */
>>> -    if (waketype == RCU_NOCB_WAKE_BYPASS) {
>>> +    if (waketype == RCU_NOCB_WAKE_LAZY
>>> +        && READ_ONCE(rdp->nocb_defer_wakeup) == RCU_NOCB_WAKE_NOT) {
>> 
>> This can be a plain READ since ->nocb_defer_wakeup is only written under ->nocb_gp_lock.
> 
> Yes makes sense, will do.
> 
>>> +        mod_timer(&rdp_gp->nocb_timer, jiffies + jiffies_till_flush);
>>> +        WRITE_ONCE(rdp_gp->nocb_defer_wakeup, waketype);
>>> +    } else if (waketype == RCU_NOCB_WAKE_BYPASS) {
>>>       mod_timer(&rdp_gp->nocb_timer, jiffies + 2);
>>>       WRITE_ONCE(rdp_gp->nocb_defer_wakeup, waketype);
>>>   } else {
>>> @@ -512,9 +598,16 @@ static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_alldone,
>>>   }
>>>   // Need to actually to a wakeup.
>>>   len = rcu_segcblist_n_cbs(&rdp->cblist);
>>> +    bypass_len = rcu_cblist_n_cbs(&rdp->nocb_bypass);
>>> +    lazy_len = READ_ONCE(rdp->lazy_len);
>>>   if (was_alldone) {
>>>       rdp->qlen_last_fqs_check = len;
>>> -        if (!irqs_disabled_flags(flags)) {
>>> +        // Only lazy CBs in bypass list
>>> +        if (lazy_len && bypass_len == lazy_len) {
>>> +            rcu_nocb_unlock_irqrestore(rdp, flags);
>>> +            wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE_LAZY,
>>> +                       TPS("WakeLazy"));
>> 
>> I'm trying to think of a case where rcu_nocb_try_bypass() returns false
>> (queue to regular list) but then call_rcu() -> __call_rcu_nocb_wake() ends up
>> seeing a lazy bypass queue even though we are queueing a non-lazy callback
>> (should have flushed in this case).
>> 
>> Looks like it shouldn't happen, even with concurrent (de-offloading) but just
>> in case, can we add:
> 
> Yes I also feel this couldn’t happen because irq is
> off and nocb lock is held throughout the calls to
> the above 2 functions. Unless I missed the race
> you’re describing?
> 
>> 
>>     WARN_ON_ONCE(lazy_len != len)
> 
> But this condition can be true even in normal
> circumstances? len also contains DONE CBs
> which are ready to be invoked. Or did I miss
> something?
> 
> Thanks,
> 
>  - Joel
> 
>> 
>>> +        } else if (!irqs_disabled_flags(flags)) {
>>>           /* ... if queue was empty ... */
>>>           rcu_nocb_unlock_irqrestore(rdp, flags);
>>>           wake_nocb_gp(rdp, false);
>> 
>> Thanks.