Re: dyntick-idle CPU and node's qsmask

[Paul E. McKenney]

On Tue, Nov 20, 2018 at 12:42:43PM -0800, Joel Fernandes wrote:
> On Sun, Nov 11, 2018 at 10:36:18AM -0800, Paul E. McKenney wrote:
> > On Sun, Nov 11, 2018 at 10:09:16AM -0800, Joel Fernandes wrote:
> > > On Sat, Nov 10, 2018 at 08:22:10PM -0800, Paul E. McKenney wrote:
> > > > On Sat, Nov 10, 2018 at 07:09:25PM -0800, Joel Fernandes wrote:
> > > > > On Sat, Nov 10, 2018 at 03:04:36PM -0800, Paul E. McKenney wrote:
> > > > > > On Sat, Nov 10, 2018 at 01:46:59PM -0800, Joel Fernandes wrote:
> > > > > > > Hi Paul and everyone,
> > > > > > > 
> > > > > > > I was tracing/studying the RCU code today in paul/dev branch and noticed that
> > > > > > > for dyntick-idle CPUs, the RCU GP thread is clearing the rnp->qsmask
> > > > > > > corresponding to the leaf node for the idle CPU, and reporting a QS on their
> > > > > > > behalf.
> > > > > > > 
> > > > > > > rcu_sched-10    [003]    40.008039: rcu_fqs:              rcu_sched 792 0 dti
> > > > > > > rcu_sched-10    [003]    40.008039: rcu_fqs:              rcu_sched 801 2 dti
> > > > > > > rcu_sched-10    [003]    40.008041: rcu_quiescent_state_report: rcu_sched 805 5>0 0 0 3 0
> > > > > > > 
> > > > > > > That's all good but I was wondering if we can do better for the idle CPUs if
> > > > > > > we can some how not set the qsmask of the node in the first place. Then no
> > > > > > > reporting would be needed of quiescent state is needed for idle CPUs right?
> > > > > > > And we would also not need to acquire the rnp lock I think.
> > > > > > > 
> > > > > > > At least for a single node tree RCU system, it seems that would avoid needing
> > > > > > > to acquire the lock without complications. Anyway let me know your thoughts
> > > > > > > and happy to discuss this at the hallways of the LPC as well for folks
> > > > > > > attending :)
> > > > > > 
> > > > > > We could, but that would require consulting the rcu_data structure for
> > > > > > each CPU while initializing the grace period, thus increasing the number
> > > > > > of cache misses during grace-period initialization and also shortly after
> > > > > > for any non-idle CPUs.  This seems backwards on busy systems where each
> > > > > 
> > > > > When I traced, it appears to me that rcu_data structure of a remote CPU was
> > > > > being consulted anyway by the rcu_sched thread. So it seems like such cache
> > > > > miss would happen anyway whether it is during grace-period initialization or
> > > > > during the fqs stage? I guess I'm trying to say, the consultation of remote
> > > > > CPU's rcu_data happens anyway.
> > > > 
> > > > Hmmm...
> > > > 
> > > > The rcu_gp_init() function does access an rcu_data structure, but it is
> > > > that of the current CPU, so shouldn't involve a communications cache miss,
> > > > at least not in the common case.
> > > > 
> > > > Or are you seeing these cross-CPU rcu_data accesses in rcu_gp_fqs() or
> > > > functions that it calls?  In that case, please see below.
> > > 
> > > Yes, it was rcu_implicit_dynticks_qs called from rcu_gp_fqs.
> > > 
> > > > > > CPU will with high probability report its own quiescent state before three
> > > > > > jiffies pass, in which case the cache misses on the rcu_data structures
> > > > > > would be wasted motion.
> > > > > 
> > > > > If all the CPUs are busy and reporting their QS themselves, then I think the
> > > > > qsmask is likely 0 so then rcu_implicit_dynticks_qs (called from
> > > > > force_qs_rnp) wouldn't be called and so there would no cache misses on
> > > > > rcu_data right?
> > > > 
> > > > Yes, but assuming that all CPUs report their quiescent states before
> > > > the first call to rcu_gp_fqs().  One exception is when some CPU is
> > > > looping in the kernel for many milliseconds without passing through a
> > > > quiescent state.  This is because for recent kernels, cond_resched()
> > > > is not a quiescent state until the grace period is something like 100
> > > > milliseconds old.  (For older kernels, cond_resched() was never an RCU
> > > > quiescent state unless it actually scheduled.)
> > > > 
> > > > Why wait 100 milliseconds?  Because otherwise the increase in
> > > > cond_resched() overhead shows up all too well, causing 0day test robot
> > > > to complain bitterly.  Besides, I would expect that in the common case,
> > > > CPUs would be executing usermode code.
> > > 
> > > Makes sense. I was also wondering about this other thing you mentioned about
> > > waiting for 3 jiffies before reporting the idle CPU's quiescent state. Does
> > > that mean that even if a single CPU is dyntick-idle for a long period of
> > > time, then the minimum grace period duration would be atleast 3 jiffies? In
> > > our mobile embedded devices, jiffies is set to 3.33ms (HZ=300) to keep power
> > > consumption low. Not that I'm saying its an issue or anything (since IIUC if
> > > someone wants shorter grace periods, they should just use expedited GPs), but
> > > it sounds like it would be shorter GP if we just set the qsmask early on some
> > > how and we can manage the overhead of doing so.
> > 
> > First, there is some autotuning of the delay based on HZ:
> > 
> > #define RCU_JIFFIES_TILL_FORCE_QS (1 + (HZ > 250) + (HZ > 500))
> > 
> > So at HZ=300, you should be seeing a two-jiffy delay rather than the
> > usual HZ=1000 three-jiffy delay.  Of course, this means that the delay
> > is 6.67ms rather than the usual 3ms, but the theory is that lower HZ
> > rates often mean slower instruction execution and thus a desire for
> > lower RCU overhead.  There is further autotuning based on number of
> > CPUs, but this does not kick in until you have 256 CPUs on your system,
> > and I bet that smartphones aren't there yet.  Nevertheless, check out
> > RCU_JIFFIES_FQS_DIV for more info on this.
> > 
> > But you can always override this autotuning using the following kernel
> > boot paramters:
> > 
> > rcutree.jiffies_till_first_fqs
> > rcutree.jiffies_till_next_fqs
> 
> Slightly related, I was just going through your patch in the dev branch "doc:
> Now jiffies_till_sched_qs solicits from cond_resched()".
> 
> If I understand correctly, what you're trying to do is set
> rcu_data.rcu_urgent_qs if you've not heard from the CPU long enough from
> rcu_implicit_dynticks_qs.
> 
> Then in the other paths, you are reading this value and similuating a dyntick
> idle transition even though you may not be really going into dyntick-idle.
> Actually in the scheduler-tick, you are also using it to set NEED_RESCHED
> appropriately.
> 
> Did I get it right so far?

Partially.

The simulated dyntick-idle transition happens if the grace period extends
for even longer, so that ->rcu_need_heavy_qs gets set.  Up to that point,
all that is asked for is a local-to-the-CPU report of a quiescent state.

> I was thinking if we could simplify rcu_note_context_switch (the parts that
> call rcu_momentary_dyntick_idle), if we did the following in
> rcu_implicit_dynticks_qs.
> 
> Since we already call rcu_qs in rcu_note_context_switch, that would clear the
> rdp->cpu_no_qs flag. Then there should be no need to call
> rcu_momentary_dyntick_idle from rcu_note_context switch.

But does this also work for the rcu_all_qs() code path?

> I think this would simplify cond_resched as well.  Could this avoid the need
> for having an rcu_all_qs at all? Hopefully I didn't some Tasks-RCU corner cases..

There is also the code path from cond_resched() in PREEMPT=n kernels.
This needs rcu_all_qs().  Though it is quite possible that some additional
code collapsing is possible.

> Basically for some background, I was thinking can we simplify the code that
> calls "rcu_momentary_dyntick_idle" since we already register a qs in other
> ways (like by resetting cpu_no_qs).

One complication is that rcu_all_qs() is invoked with interrupts
and preemption enabled, while rcu_note_context_switch() is
invoked with interrupts disabled.  Also, as you say, Tasks RCU.
Plus rcu_all_qs() wants to exit immediately if there is nothing to
do, while rcu_note_context_switch() must unconditionally do rcu_qs()
-- yes, it could check, but that would be redundant with the checks
within rcu_qs().  The one function traces and the other one doesn't,
but it would be OK if both traced.  (I hope, anyway:  The cond_resched()
performance requirements are surprisingly severe.)  Aside from that,
the two functions are quite similar.

It would of course be possible to create a common helper function that
rcu_all_qs() and rcu_note_context_switch() both became simple wrappers
for, but it is not clear that this would actually be shorter or simpler.

> I should probably start drawing some pictures to make sense of everything,
> but do let me know if I have a point ;-) Thanks for your time.

This stuff is admittedly a bit fiddly.  Again, it took some serious
work to avoid cond_resched() performance regressions.

> - Joel
> 
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index c818e0c91a81..5aa0259c014d 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -1063,7 +1063,7 @@ static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
>  	 * read-side critical section that started before the beginning
>  	 * of the current RCU grace period.
>  	 */
> -	if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
> +	if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap) || !rdp->cpu_no_qs.b.norm) {

If I am not too confused, this change could cause trouble for
nohz_full CPUs looping in the kernel.  Such CPUs don't necessarily take
scheduler-clock interrupts, last I checked, and this could prevent the
CPU from reporting its quiescent state to core RCU.

Or am I missing something here?

							Thanx, Paul

>  		trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
>  		rcu_gpnum_ovf(rnp, rdp);
>  		return 1;
>