Re: [PATCH v2] ksoftirqd: Enable IRQs and call cond_resched() before poking RCU

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Wed, Jan 07, 2015 at 08:33:29PM -0800, Calvin Owens wrote:
> On Wednesday 01/07 at 08:52 -0800, Paul E. McKenney wrote:
> > On Tue, Jan 06, 2015 at 06:19:26PM -0800, Paul E. McKenney wrote:
> > > On Tue, Jan 06, 2015 at 05:49:06PM -0800, Calvin Owens wrote:
> > > > While debugging an issue with excessive softirq usage, I encountered the
> > > > following note in commit 3e339b5dae24a706 ("softirq: Use hotplug thread
> > > > infrastructure"):
> > > > 
> > > >     [ paulmck: Call rcu_note_context_switch() with interrupts enabled. ]
> > > > 
> > > > ...but despite this note, the patch still calls RCU with IRQs disabled.
> > > > 
> > > > This seemingly innocuous change caused a significant regression in softirq
> > > > CPU usage on the sending side of a large TCP transfer (~1 GB/s): when
> > > > introducing 0.01% packet loss, the softirq usage would jump to around 25%,
> > > > spiking as high as 50%. Before the change, the usage would never exceed 5%.
> > > > 
> > > > Moving the call to rcu_note_context_switch() after the cond_sched() call,
> > > > as it was originally before the hotplug patch, completely eliminated this
> > > > problem.
> > > > 
> > > > Signed-off-by: Calvin Owens <calvinowens@fb.com>
> > > > ---
> > > > Changes since v1:
> > > > 	I mixed up the kernel versions I was patching against, sorry!
> > > > 
> > > >  kernel/softirq.c | 6 +++++-
> > > >  1 file changed, 5 insertions(+), 1 deletion(-)
> > > > 
> > > > diff --git a/kernel/softirq.c b/kernel/softirq.c
> > > > index 501baa9..9e787d8 100644
> > > > --- a/kernel/softirq.c
> > > > +++ b/kernel/softirq.c
> > > > @@ -656,9 +656,13 @@ static void run_ksoftirqd(unsigned int cpu)
> > > >  		 * in the task stack here.
> > > >  		 */
> > > >  		__do_softirq();
> > > > -		rcu_note_context_switch();
> > > >  		local_irq_enable();
> > > >  		cond_resched();
> > > 
> > > If this is for 3.20, we can just replace cond_resched() with
> > > cond_resched_rcu_qs(), and get rid of the direct call to
> > > rcu_note_context_switch().  This has the benefit of avoiding
> > > needless rcu_note_context_switch() overhead if cond_resched()
> > > actually did a reschedule.
> > > 
> > > But don't try it in 3.19 or earlier.  ;-)
> > 
> > As in the following for 3.20.  Does this version work for you?
> 
> That's great, thanks :)
> 
> Should this go to stable as well? It is technically a regression, albeit
> a rather long-standing one.

Your original patch could go into stable, but my updated version depends
on functionality not present before 3.20.

> My original scenario was a bit contrived, but I tested this on some real
> loads and it makes a difference: on a heavily loaded Proxygen server,
> the aggregate softirq CPU usage decreases by roughly 10% (relative)
> given the same amount of traffic with the patch. It also produces
> statistically significant performance wins at higher loads on
> webservers: about a 1% reduction in overall CPU utilization and improved
> latency metrics.

OK, good to know.  I have added this information to the commit log, please
see below for updated commit.

							Thanx, Paul

------------------------------------------------------------------------

ksoftirqd: Enable IRQs and call cond_resched() before poking RCU

While debugging an issue with excessive softirq usage, I encountered the
following note in commit 3e339b5dae24a706 ("softirq: Use hotplug thread
infrastructure"):
    
        [ paulmck: Call rcu_note_context_switch() with interrupts enabled. ]
    
...but despite this note, the patch still calls RCU with IRQs disabled.

This seemingly innocuous change caused a significant regression in softirq
CPU usage on the sending side of a large TCP transfer (~1 GB/s): when
introducing 0.01% packet loss, the softirq usage would jump to around
25%, spiking as high as 50%. Before the change, the usage would never
exceed 5%.  On a heavily loaded Proxygen server, the aggregate softirq
CPU usage decreases by roughly 10% (relative) given the same amount
of traffic with the patch. It also produces statistically significant
performance wins at higher loads on webservers: about a 1% reduction in
overall CPU utilization and improved latency metrics.

Moving the call to rcu_note_context_switch() after the cond_sched() call,
as it was originally before the hotplug patch, completely eliminated this
problem, but the new cond_resched_rcu_qs() provides shorter code and
avoids double RCU notification in the case where cond_resched() really
did a context switch.

Signed-off-by: Calvin Owens <calvinowens@fb.com>
[ paulmck: Substituted shiny new cond_resched_rcu_qs() primitive. ]
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
[ paulmck: Added Calvin's measurements on Proxygen server and webservers. ]

diff --git a/kernel/softirq.c b/kernel/softirq.c
index 501baa9ac1be..8cdb98847c7b 100644
--- a/kernel/softirq.c
+++ b/kernel/softirq.c
@@ -656,9 +656,8 @@ static void run_ksoftirqd(unsigned int cpu)
 		 * in the task stack here.
 		 */
 		__do_softirq();
-		rcu_note_context_switch();
 		local_irq_enable();
-		cond_resched();
+		cond_resched_rcu_qs();
 		return;
 	}
 	local_irq_enable();