Message-ID: <YzMVzpvvHbbkFMs8@google.com>
Date:   Tue, 27 Sep 2022 15:25:02 +0000
From:   Joel Fernandes <joel@...lfernandes.org>
To:     "Paul E. McKenney" <paulmck@...nel.org>
Cc:     Uladzislau Rezki <urezki@...il.com>, rcu@...r.kernel.org,
        linux-kernel@...r.kernel.org, rushikesh.s.kadam@...el.com,
        neeraj.iitr10@...il.com, frederic@...nel.org, rostedt@...dmis.org
Subject: Re: [PATCH v6 1/4] rcu: Make call_rcu() lazy to save power

On Tue, Sep 27, 2022 at 07:30:20AM -0700, Paul E. McKenney wrote:
> On Tue, Sep 27, 2022 at 02:22:56PM +0000, Joel Fernandes wrote:
> > On Tue, Sep 27, 2022 at 07:14:03AM -0700, Paul E. McKenney wrote:
> > > On Tue, Sep 27, 2022 at 01:05:41PM +0000, Joel Fernandes wrote:
> > > > On Mon, Sep 26, 2022 at 08:22:46PM -0700, Paul E. McKenney wrote:
> > > > [..]
> > > > > > > > >>> --- a/kernel/workqueue.c
> > > > > > > > >>> +++ b/kernel/workqueue.c
> > > > > > > > >>> @@ -1771,7 +1771,7 @@ bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
> > > > > > > > >>> 
> > > > > > > > >>>        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
> > > > > > > > >>>                rwork->wq = wq;
> > > > > > > > >>> -               call_rcu(&rwork->rcu, rcu_work_rcufn);
> > > > > > > > >>> +               call_rcu_flush(&rwork->rcu, rcu_work_rcufn);
> > > > > > > > >>>                return true;
> > > > > > > > >>>        }
> > > > > > > > >>> 
> > > > > > > > >>> <snip>
> > > > > > > > >>> 
> > > > > > > > >>> ?
> > > > > > > > >>> 
> > > > > > > > >>> But it does not fully solve my boot-up issue. Will debug tomorrow further.
> > > > > > > > >> 
> > > > > > > > >> Ah, but at least its progress, thanks. Could you send me a patch to include
> > > > > > > > >> in the next revision with details of this?
> > > > > > > > >> 
> > > > > > > > >>>> Might one more proactive approach be to use Coccinelle to locate such
> > > > > > > > >>>> callback functions?  We might not want -all- callbacks that do wakeups
> > > > > > > > >>>> to use call_rcu_flush(), but knowing which are which should speed up
> > > > > > > > >>>> slow-boot debugging by quite a bit.
> > > > > > > > >>>> 
> > > > > > > > >>>> Or is there a better way to do this?
> > > > > > > > >>>> 
> > > > > > > > >>> I am not sure what Coccinelle is. If we had something automated that measures
> > > > > > > > >>> a boot time and if needed does some profiling it would be good. Otherwise it
> > > > > > > > >>> is a manual debugging mainly, IMHO.
> > > > > > > > >> 
> > > > > > > > >> Paul, What about using a default-off kernel CONFIG that splats on all lazy
> > > > > > > > >> call_rcu() callbacks that do a wake up. We could use the trace hooks to do it
> > > > > > > > >> in kernel I think. I can talk to Steve to get ideas on how to do that but I
> > > > > > > > >> think it can be done purely from trace events (we might need a new
> > > > > > > > >> trace_end_invoke_callback to fire after the callback is invoked). Thoughts?
> > > > > > > > > 
> > > > > > > > > Could you look for wakeups invoked between trace_rcu_batch_start() and
> > > > > > > > > trace_rcu_batch_end() that are not from interrupt context?  This would
> > > > > > > > > of course need to be associated with a task rather than a CPU.
> > > > > > > > 
> > > > > > > > Yes this sounds good, but we also need to know if the callbacks are
> > > > > > > > lazy or not since wake-up is ok from a non lazy one. I think I’ll
> > > > > > > > need a table to track that at queuing time.
> > > > > > > 
> > > > > > > Agreed.
> > > > > > > 
> > > > > > > > > Note that you would need to check for wakeups from interrupt handlers
> > > > > > > > > even with the extra trace_end_invoke_callback().  The window where an
> > > > > > > > > interrupt handler could do a wakeup would be reduced, but not eliminated.
> > > > > > > > 
> > > > > > > > True! Since this is a  debugging option, can we not just disable interrupts across callback invocation?
> > > > > > > 
> > > > > > > Not without terminally annoying lockdep, at least for any RCU callbacks
> > > > > > > doing things like spin_lock_bh().
> > > > > > > 
> > > > > > 
> > > > > > Sorry if my last email bounced. Looks like my iPhone betrayed me this once ;)
> > > > > > 
> > > > > > I was thinking something like this:
> > > > > > 1. Put a flag in rcu_head to mark CBs as lazy.
> > > > > > 2. Add a trace_rcu_invoke_callback_end() trace point.
> > > > > > 
> > > > > > Both #1 and #2 can be a debug CONFIG option. #2 can be a tracepoint and not
> > > > > > exposed if needed.
> > > > > > 
> > > > > > 3. Put an in-kernel probe on both trace_rcu_invoke_callback_start() and
> > > > > > trace_rcu_invoke_callback_end(). In the start probe, set a per-task flag if
> > > > > > the current CB is lazy. In the end probe, clear it.
> > > > > > 
> > > > > > 4. Put an in-kernel probe on trace_rcu_sched_wakeup().
> > > > > > 
> > > > > > Splat in the wake up probe if:
> > > > > > 1. Hard IRQs are on.
> > > > > > 2. The per-cpu flag is set.
> > > > > > 
> > > > > > #3 actually does not even need probes if we can directly call the functions
> > > > > > from the rcu_do_batch() function.
> > > > > 
> > > > > This is fine for an experiment or a debugging session, but a solution
> > > > > based totally on instrumentation would be better for production use.
> > > > 
> > > > Maybe we can borrow the least-significant bit of rhp->func to mark laziness?
> > > > Then it can be production as long as we're ok with the trace_sched_wakeup
> > > > probe.
> > > 
> > > Last time I tried this, there were architectures that could have odd-valued
> > > function addresses.  Maybe this is no longer the case?
> > 
> > Oh ok! If this happens, maybe we can just make it depend on x86-64 assuming
> > x86-64 does not have pointer oddness. We can also add a warning for if the
> > function address is odd before setting the bit.
> 
> Let me rephrase this...  ;-)
> 
> Given that this used to not work and still might not work, let's see
> if we can find some other way to debug this.  Unless and until it can
> be demonstrated that there is no supported compiler that will generated
> odd-valued function addresses on any supported architecture.
> 
> Plus there was a time that x86 did odd-valued pointer addresses.
> The instruction set is plenty fine with this, so it would have to be a
> compiler and assembly-language convention to avoid it.

Ok, so then I am not sure how to make it work in production at the moment. I
could track the lazy callbacks in a hashtable but then that's overhead.

Or, I could focus on trying Vlad's config and figure out what's going on and
keep the auto-debug for later.

On another thought, this is the sort of thing that should be doable via Daniel
Bristot's runtime verification framework, as its a classical "see if these
traces look right" issue which should be teachable to a computer with a few rules.

thanks,

 - Joel

Powered by blists - more mailing lists