Re: [PATCH] srcu: switch work func to allow concurrent gp

[Pingfan Liu <kernelfans@gmail.com>]

On Tue, Dec 13, 2022 at 11:51:21AM -0800, Boqun Feng wrote:
> On Tue, Dec 13, 2022 at 09:00:30PM +0800, Pingfan Liu wrote:
> > On Thu, Dec 01, 2022 at 09:53:36PM +0800, Pingfan Liu wrote:
> > > On Wed, Nov 30, 2022 at 08:53:43AM -0800, Boqun Feng wrote:
> > > > On Wed, Nov 30, 2022 at 04:39:02PM +0800, Pingfan Liu wrote:
> > > > > ssp->srcu_cb_mutex is introduced to allow the other srcu state machine
> > > > > to advance as soon as possible. But according to the implement of
> > > > > workqueue, the same work_struct is serialized and can not run
> > > > > concurrently in fact.
> > > > > 
> > > > > Quoting from Documentation/core-api/workqueue.rst
> > > > > "
> > > > > Non-reentrance Conditions
> > > > > =========================
> > > > > 
> > > > > Workqueue guarantees that a work item cannot be re-entrant if the following
> > > > > conditions hold after a work item gets queued:
> > > > > 
> > > > >         1. The work function hasn't been changed.
> > > > >         2. No one queues the work item to another workqueue.
> > > > >         3. The work item hasn't been reinitiated.
> > > > > "
> > > > > 
> > > > > To allow the concurrence to some extent, it can be achieved by changing
> > > > > the work function to break the conditions. As a result, when
> > > > > srcu_gp_end() releases srcu_gp_mutex, a new state machine can begin.
> > > > > 
> > > > > Signed-off-by: Pingfan Liu <kernelfans@gmail.com>
> > > > > Cc: Lai Jiangshan <jiangshanlai@gmail.com>
> > > > > Cc: "Paul E. McKenney" <paulmck@kernel.org>
> > > > > Cc: Frederic Weisbecker <frederic@kernel.org>
> > > > > Cc: Josh Triplett <josh@joshtriplett.org>
> > > > > Cc: Steven Rostedt <rostedt@goodmis.org>
> > > > > Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
> > > > > Cc: "Zhang, Qiang1" <qiang1.zhang@intel.com>
> > > > > To: rcu@vger.kernel.org
> > > > > ---
> > > > >  kernel/rcu/srcutree.c | 19 +++++++++++++++++++
> > > > >  1 file changed, 19 insertions(+)
> > > > > 
> > > > > diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
> > > > > index 1c304fec89c0..56dd9bb2c8b8 100644
> > > > > --- a/kernel/rcu/srcutree.c
> > > > > +++ b/kernel/rcu/srcutree.c
> > > > > @@ -75,6 +75,7 @@ static bool __read_mostly srcu_init_done;
> > > > >  static void srcu_invoke_callbacks(struct work_struct *work);
> > > > >  static void srcu_reschedule(struct srcu_struct *ssp, unsigned long delay);
> > > > >  static void process_srcu(struct work_struct *work);
> > > > > +static void process_srcu_wrap(struct work_struct *work);
> > > > >  static void srcu_delay_timer(struct timer_list *t);
> > > > >  
> > > > >  /* Wrappers for lock acquisition and release, see raw_spin_lock_rcu_node(). */
> > > > > @@ -763,6 +764,11 @@ static void srcu_gp_end(struct srcu_struct *ssp)
> > > > >  		cbdelay = 0;
> > > > >  
> > > > >  	WRITE_ONCE(ssp->srcu_last_gp_end, ktime_get_mono_fast_ns());
> > > > > +	/* Change work func so work can be concurrent */
> > > > > +	if (ssp->work.work.func == process_srcu_wrap)
> > > > > +		ssp->work.work.func = process_srcu;
> > > > > +	else
> > > > > +		ssp->work.work.func = process_srcu_wrap;
> > > > 
> > > > This looks really hacky ;-) It would be good that workqueue has an API
> > > > to allow "resetting" a work.
> > > > 
> > > 
> > > Indeed, it is hacky and it can be done by using alternative work_struct
> > > so that from the workqueue API, it is intact.
> > > 
> > > > Do you have any number of the potential performance improvement?
> > > > 
> > > 
> > > I will try to bring out one. But the result may heavily depend on the
> > > test case.
> > > 
> > 
> > Sorry to update late, I was interrupted by something else.
> 
> No worries, thanks for looking into it.
> 
> > 
> > I used two machine to test
> > 
> > -1. on 144 cpus hpe-dl560gen10
> > 
> > modprobe rcutorture   torture_type=srcud  fwd_progress=144 fwd_progress_holdoff=1 shutdown_secs=36000 stat_interval=60 verbose=1
> > 
> > # no patch
> > [37587.632155] srcud:  End-test grace-period state: g28287244 f0x0 total-gps=28287244
> > #my patch
> > [36443.468017] srcud:  End-test grace-period state: g29026056 f0x0 total-gps=29026056
> > 
> > 
> > -2. on 256 cpus amd-milan
> > modprobe rcutorture   torture_type=srcud  fwd_progress=256 fwd_progress_holdoff=1 shutdown_secs=36000 stat_interval=60 verbose=1
> > 
> > # no patch
> > [36093.605732] srcud:  End-test grace-period state: g10850284 f0x0 total-gps=10850284
> > #my patch
> > [36093.856713] srcud:  End-test grace-period state: g10672632 f0x0 total-gps=10672632
> > 
> > 
> > The first test shows that it has about 2.6% improvement, while the
> > second test shows it has no significant effect.
> > 
> > 
> > I wonder if any test case can be in flavour of my patch. Any suggestion?
> > 
> 
> I actually wait you to tell me that ;-) Your trick clearly can improve
> the degree of parallel a bit, but IIUC you need that srcu_end_gp() runs
> at the same time with the srcu work function (i.e. pending bit is
> cleared) to archieve that. And that's really a small time window ;-)
> 

Yes, it is smaller than I expected.

> And there's always the Amdahl's law: do we know whether the state
> machine processing of SRCU is the bottleneck in any workload?
> 
> And are we willing to give CPU time to SRCU state machine processing
> instead of anything else?
> 
> My suggestion is to keep this patch somewhere, and whenever you see a
> related problem, try the patch and see whether it helps ;-)
> 

Yes. Thanks for your suggestion.

Again, appreciate for your help and advice.

Regards,

	Pingfan

> Regards,
> Boqun
> 
> > 
> > Thanks,
> > 
> > 	Pingfan
> > 
> > > Thanks,
> > > 
> > > 	Pingfan
> > > 
> > > > Regards,
> > > > Boqun
> > > > 
> > > > >  	rcu_seq_end(&ssp->srcu_gp_seq);
> > > > >  	gpseq = rcu_seq_current(&ssp->srcu_gp_seq);
> > > > >  	if (ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, gpseq))
> > > > > @@ -1637,6 +1643,19 @@ static void process_srcu(struct work_struct *work)
> > > > >  	srcu_reschedule(ssp, curdelay);
> > > > >  }
> > > > >  
> > > > > +/*
> > > > > + * The ssp->work is expected to be concurrent to some extent, but the current
> > > > > + * workqueue does not support the concurrence on the same work. (Refer to the
> > > > > + * section "Non-reentrance Conditions" in the file workqueue.rst)
> > > > > + * Resolving it by changing the work func.
> > > > > + *
> > > > > + * Prevent compilering from optimizing out it.
> > > > > + */
> > > > > +static __used void process_srcu_wrap(struct work_struct *work)
> > > > > +{
> > > > > +	process_srcu(work);
> > > > > +}
> > > > > +
> > > > >  void srcutorture_get_gp_data(enum rcutorture_type test_type,
> > > > >  			     struct srcu_struct *ssp, int *flags,
> > > > >  			     unsigned long *gp_seq)
> > > > > -- 
> > > > > 2.31.1
> > > > >