Re: Regression :-) Re: [GIT PULL RESEND] x86/jumpmplabel changes for v3.12-rc1

[Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>]

On Wed, Sep 11, 2013 at 03:14:52PM -0400, Steven Rostedt wrote:
> On Wed, 11 Sep 2013 14:56:54 -0400
> Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> wrote:
> 
> 
> > > I'm looking to NAK your patch because it is obvious that the jump label
> > > code isn't doing what you expect it to be doing. And it wasn't until my
> > 
> > Actually it is OK. They need to be enabled before the SMP code kicks in.
> > 
> > > checks were in place for you to notice.
> > 
> > Any suggestion on how to resolve the crash?
> > 
> > The PV spinlock code is OK (I think, I need to think hard about this) until
> > the spinlocks start being used by multiple CPUs. At that point the
> > jump_lables have to be in place - otherwise you will end with a spinlock
> > going in a slowpath (patched over) and an kicker not using the slowpath
> > and never kicking the waiter. Which ends with a hanged system.
> 
> Note, a simple early_initcall() could do the trick. SMP isn't set up
> until much further in the boot process.
> 
> > 
> > Or simple said - jump labels have to be setup before we boot
> > the other CPUs. 
> 
> Right, and initcalls() can easily serve that purpose.
> 
> > 
> > This would affect the KVM guests as well, I think if the slowpath
> > waiter was blocking on the VCPU (which I think it is doing now, but
> > not entirely sure?)
> > 
> > P.S.
> > I am out on vacation tomorrow for a week. Boris (CC-ed here) can help.
> 
> Your patch isn't wrong per say, but I'm hesitant to apply it because it
> the result is different depending on whether JUMP_LABEL is configured
> or not. Using any jump_label() calls before jump_label_init() is
> called, is entering a gray area, and I think it should be avoided.
> 
> This patch should solve it for you:

And also the pv_lock_ops need to be set before alternative_asm
code is called :-) (Called from check_bugs()).

Otherwise you end up with some code still using the native slowpath
kicker/waiter while the modules might be using the Xen variant.

I knew that I forgot to mention something ..

With that in mind and your patch I made this one:

diff --git a/arch/x86/xen/spinlock.c b/arch/x86/xen/spinlock.c
index 253f63f..d90628d 100644
--- a/arch/x86/xen/spinlock.c
+++ b/arch/x86/xen/spinlock.c
@@ -267,11 +267,18 @@ void __init xen_init_spinlocks(void)
 		return;
 	}
 
-	static_key_slow_inc(&paravirt_ticketlocks_enabled);
-
 	pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(xen_lock_spinning);
 	pv_lock_ops.unlock_kick = xen_unlock_kick;
 }
+static __init int xen_init_spinlocks_jump(void)
+{
+	if (!xen_pvspin)
+		return 0;
+
+	static_key_slow_inc(&paravirt_ticketlocks_enabled);
+	return 0;
+}
+early_initcall(xen_init_spinlocks_jump);
 
 static __init int xen_parse_nopvspin(char *arg)
 {

which seem to work.