Date:   Thu, 29 Dec 2016 21:54:49 +0800
From:   Chen Yu <yu.c.chen@...el.com>
To:     Marcelo Tosatti <mtosatti@...hat.com>
Cc:     kvm@...r.kernel.org, x86@...nel.org, linux-kernel@...r.kernel.org,
        "Rafael J. Wysocki" <rjw@...ysocki.net>,
        Len Brown <lenb@...nel.org>,
        Paolo Bonzini <pbonzini@...hat.com>,
        Radim Krcmar <rkrcmar@...hat.com>,
        Thomas Gleixner <tglx@...utronix.de>,
        Ingo Molnar <mingo@...hat.com>,
        "H. Peter Anvin" <hpa@...or.com>,
        Wanpeng Li <wanpeng.li@...mail.com>
Subject: Re: [PATCH][RFC] VM: x86: Return ealier if clocksource has not
 changed

On Thu, Dec 29, 2016 at 07:57:33AM -0200, Marcelo Tosatti wrote:
> On Thu, Dec 29, 2016 at 04:59:02PM +0800, Chen Yu wrote:
> > On Tue, Dec 27, 2016 at 01:32:47PM -0200, Marcelo Tosatti wrote:
> > > On Tue, Dec 27, 2016 at 04:06:44PM +0800, Chen Yu wrote:
> > > > Hi Marcelo,
> > > > On Mon, Dec 26, 2016 at 05:44:25PM -0200, Marcelo Tosatti wrote:
> > > > > On Fri, Dec 23, 2016 at 04:41:53PM +0800, Chen Yu wrote:
> > > > > > Currently the notifier of pvclock_gtod_notify() get invoked
> > > > > > frequently due to the periodic update_wall_time(). This might
> > > > > > slow down the system a little bit as there might be redundant
> > > > > > execution code path and unnecessary lock contention
> > > > > > in update_pvclock_gtod(), which was found when I was doing
> > > > > > suspend/resume speed testings. As pvclock_gtod_notify()
> > > > > > should be invoked only when clocksource has changed, according to
> > > > > > Commit 16e8d74d2da9 ("KVM: x86: notifier for clocksource changes")
> > > > > > , either we can add a new notifier for clocksource switch,
> > > > > > or we can simply bypass the following code in pvclock_gtod_notify()
> > > > > > earlier if there is no clocksource switch.
> > > > > > 
> > > > > > Cc: Paolo Bonzini <pbonzini@...hat.com>
> > > > > > Cc: "Radim Krcmar" <rkrcmar@...hat.com>
> > > > > > Cc: Marcelo Tosatti <mtosatti@...hat.com>
> > > > > > Cc: Thomas Gleixner <tglx@...utronix.de>
> > > > > > Cc: Ingo Molnar <mingo@...hat.com>
> > > > > > Cc: "H. Peter Anvin" <hpa@...or.com>
> > > > > > Cc: Wanpeng Li <wanpeng.li@...mail.com>
> > > > > > Signed-off-by: Chen Yu <yu.c.chen@...el.com>
> > > > > > ---
> > > > > >  arch/x86/kvm/x86.c | 5 +++--
> > > > > >  1 file changed, 3 insertions(+), 2 deletions(-)
> > > > > > 
> > > > > > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > > > > > index 445c51b..54aa32d 100644
> > > > > > --- a/arch/x86/kvm/x86.c
> > > > > > +++ b/arch/x86/kvm/x86.c
> > > > > > @@ -5961,13 +5961,14 @@ static int pvclock_gtod_notify(struct notifier_block *nb, unsigned long unused,
> > > > > >  	struct pvclock_gtod_data *gtod = &pvclock_gtod_data;
> > > > > >  	struct timekeeper *tk = priv;
> > > > > >  
> > > > > > +	if (likely(gtod->clock.vclock_mode == VCLOCK_TSC))
> > > > > > +		return 0;
> > > > > 
> > > > > I think this is only safe if any of the values in "struct
> > > > > pvclock_gtod_data" are unchanged. Otherwise the local (KVM) copy is 
> > > > > kept incorrect.
> > > > I missread the code previously and I thought only under the condition
> > > > the clocksource has been switched to another one will the KVM copy
> > > > be touched. Apparently it is not the case because the copy should
> > > > be updated on-time during normal tick, right?
> > > > thanks for your reply,
> > > 
> > > Yes, it is updated during the normal tick, and mult/freq values change.
> > > 
> > > However, if none of them change, its not necessary to call the callback.
> > > Perhaps you can check if any of the values changed and only 
> > > invoke the callback in that case?
> > >
> > Yes, this should be an optimization, but most of the callers(workload) come
> > from update_wall_time(), and in this code path the clock source's cycle
> > should already be updated in most cases, so this optimization should not take
> > much effect to reduce the burden I guess?
> > 
> > Thanks,
> > Yu
> 
> I don't understand your reasoning.
> 
> "If the clock source parameters are already updated then optimization
> does not make much effect".
> 
> If the clock source parameters are updated (that is there has been no
> change in any of the values in pvclock_gtod_data), then you can skip
> the callback. This case reduces the burden.
> 
> Right?
> 
> 
Yes, in general case we can improve the code logic.
Previously I was thinking of the case I encountered:

1. There are quite some invokes of pvclock_gtod_notify() caught by ftrace, and
   most of them should be triggered by update_wall_time()
2. If we optimize the code not to invoke pvclock_gtod_notify() if there is no
   modification of pvclock_gtod_data, it will reduce the burden for general use
   cases, but not for update_wall_time(), as it has already been modified
   in update_wall_time.

But yes, you are right, above is just my scenario, the optimization you mentioned
is a generic solution for most cases. Do you mean the following solution?

Index: linux/kernel/time/timekeeping.c
===================================================================
--- linux.orig/kernel/time/timekeeping.c
+++ linux/kernel/time/timekeeping.c
@@ -527,7 +527,13 @@ static RAW_NOTIFIER_HEAD(pvclock_gtod_ch
 
 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
 {
-	raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
+	static struct timekeeper prev_timekeeper;
+
+	/* Only notify if the clocksource has changed.*/
+	if (memcmp(tk, &prev_timekeeper, sizeof(struct timekeeper))) {
+		raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
+		memcpy(&prev_timekeeper, tk, sizeof(struct timekeeper));
+	}
 }
 
 /**

Powered by blists - more mailing lists