lkml.org 
[lkml]   [2011]   [Jun]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [GIT PULL] Re: REGRESSION: Performance regressions from switching anon_vma->lock to mutex
On Fri, Jun 17, 2011 at 11:43:33AM +0200, Ingo Molnar wrote:
>
> * Paul E. McKenney <paulmck@linux.vnet.ibm.com> wrote:
>
> > On Fri, Jun 17, 2011 at 12:58:03AM +0200, Ingo Molnar wrote:
> > >
> > > * Andi Kleen <ak@linux.intel.com> wrote:
> > >
> > > > > There's a crazy solution for that: the idle thread could process
> > > > > RCU callbacks carefully, as if it was running user-space code.
> > > >
> > > > In Ben's kernel NFS server case the system may not be idle.
> > >
> > > An always-100%-busy NFS server is very unlikely, but even in the
> > > hypothetical case a kernel NFS server is really performing system
> > > calls from a kernel thread in essence. If it doesn't do it explicitly
> > > then its main loop can easily include a "check RCU callbacks" call.
> >
> > As long as they make sure to call it in a clean environment: no
> > locks held and so on. But I am a bit worried about the possibility
> > of someone forgetting to put one of these where it is needed -- it
> > would work just fine for most workloads, but could fail only for
> > rare workloads.
>
> Yeah, some sort of worst-case-tick mechanism would guarantee that we
> wont remain without RCU GC.

Agreed!

> > That said, invoking RCU core/callback processing from the scheduler
> > context certainly sounds like an interesting way to speed up grace
> > periods.
>
> It also moves whatever priority logic is needed closer to the
> scheduler that has to touch those data structures anyway.
>
> RCU, at least partially, is a scheduler driven garbage collector even
> today: beyond context switch quiescent states the main practical role
> of the per CPU timer tick itself is scheduling. So having it close to
> when we do context-switches anyway looks pretty natural - worth
> trying.
>
> It might not work out in practice, but at first sight it would
> simplify a few things i think.

OK, please see below for a patch that not only builds, but actually
passes minimal testing. ;-)

Possible expectations and outcomes:

1. Reduced grace-period latencies on !NO_HZ systems and
on NO_HZ systems where each CPU goes non-idle frequently.
(My unscientific testing shows little or no benefit, but
then again, I was running rcutorture, which specializes
in long read-side critical sections. And I was running
NO_HZ, which is less likely to show benefit.)

I would not expect direct call to have any benefit over
softirq invocation -- sub-microsecond softirq overhead
won't matter to multi-millisecond

2. Better cache locality. I am a bit skeptical, but there is
some chance that this might reduce cache misses on task_struct.
Direct call might well do better than softirq here.

3. Easier conversion to user-space operation. I was figuring
on using POSIX signals for scheduler_tick() and for softirq,
so wasn't worried about it, but might well be simpler.

Anything else?

On eliminating softirq, I must admit that I am a bit worried about
invoking the RCU core code from scheduler_tick(), but there are some
changes that I could make that would reduce force_quiescent_state()
worst-case latency. At the expense of lengthening grace periods,
unfortunately, but the reduction will be needed for RT on larger systems
anyway, so might as well try it.

Thanx, Paul

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

rcu: Experimental change driving RCU core from scheduler

This change causes RCU's context-switch code to check to see if the RCU
core needs anything from the current CPU, and, if so, invoke the RCU
core code. One possible improvement from this experiment is reduced
grace-period latency on systems that either have NO_HZ=n or that have
all CPUs frequently going non-idle. The invocation of the RCU core code
is currently via softirq, though a plausible next step in the experiment
might be to instead use a direct function call.

Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>

diff --git a/kernel/rcutree.c b/kernel/rcutree.c
index 405a5fd..6b7a43a 100644
--- a/kernel/rcutree.c
+++ b/kernel/rcutree.c
@@ -87,6 +87,10 @@ static struct rcu_state *rcu_state;
int rcu_scheduler_active __read_mostly;
EXPORT_SYMBOL_GPL(rcu_scheduler_active);

+static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
+static int rcu_pending(int cpu);
+static void rcu_process_callbacks(struct softirq_action *unused);
+
#ifdef CONFIG_RCU_BOOST

/*
@@ -159,8 +163,14 @@ void rcu_bh_qs(int cpu)
*/
void rcu_note_context_switch(int cpu)
{
+ unsigned long flags;
+
rcu_sched_qs(cpu);
rcu_preempt_note_context_switch(cpu);
+ local_irq_save(flags);
+ if (rcu_pending(cpu))
+ invoke_rcu_core();
+ local_irq_restore(flags);
}
EXPORT_SYMBOL_GPL(rcu_note_context_switch);

@@ -182,9 +192,6 @@ module_param(qlowmark, int, 0);
int rcu_cpu_stall_suppress __read_mostly;
module_param(rcu_cpu_stall_suppress, int, 0644);

-static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
-static int rcu_pending(int cpu);
-
/*
* Return the number of RCU-sched batches processed thus far for debug & stats.
*/

\
 
 \ /
  Last update: 2011-06-17 18:51    [W:0.453 / U:0.020 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site