Date:   Tue, 31 May 2022 18:15:36 +0200
From:   nicolas saenz julienne <nsaenz@...nel.org>
To:     Frederic Weisbecker <frederic@...nel.org>
Cc:     LKML <linux-kernel@...r.kernel.org>,
        Peter Zijlstra <peterz@...radead.org>,
        Phil Auld <pauld@...hat.com>,
        Alex Belits <abelits@...vell.com>,
        Xiongfeng Wang <wangxiongfeng2@...wei.com>,
        Neeraj Upadhyay <quic_neeraju@...cinc.com>,
        Thomas Gleixner <tglx@...utronix.de>,
        Yu Liao <liaoyu15@...wei.com>,
        Boqun Feng <boqun.feng@...il.com>,
        "Paul E . McKenney" <paulmck@...nel.org>,
        Marcelo Tosatti <mtosatti@...hat.com>,
        Paul Gortmaker <paul.gortmaker@...driver.com>,
        Uladzislau Rezki <uladzislau.rezki@...y.com>,
        Joel Fernandes <joel@...lfernandes.org>,
        Mark Rutland <mark.rutland@....com>
Subject: Re: [PATCH 20/21] rcu/context_tracking: Merge dynticks counter and
 context tracking states

On Tue, 2022-05-31 at 16:23 +0200, Frederic Weisbecker wrote:
> On Mon, May 30, 2022 at 08:02:57PM +0200, nicolas saenz julienne wrote:
> > Hi Frederic,
> > 
> > On Thu, 2022-05-19 at 16:58 +0200, Frederic Weisbecker wrote:
> > > Updating the context tracking state and the RCU dynticks counter
> > > atomically in a single operation is a first step towards improving CPU
> > > isolation. This makes the context tracking state updates fully ordered
> > > and therefore allow for later enhancements such as postponing some work
> > > while a task is running isolated in userspace until it ever comes back
> > > to the kernel.
> > > 
> > > The state field becomes divided in two parts:
> > > 
> > > 1) Two Lower bits for context tracking state:
> > > 
> > > 	CONTEXT_KERNEL = 0
> > >    	CONTEXT_IDLE = 1,
> > > 	CONTEXT_USER = 2,
> > > 	CONTEXT_GUEST = 3,
> > > 
> > > 2) Higher bits for RCU eqs dynticks counting:
> > > 
> > >     RCU_DYNTICKS_IDX = 4
> > > 
> > >    The dynticks counting is always incremented by this value.
> > >    (state & RCU_DYNTICKS_IDX) means we are NOT in an extended quiescent
> > >    state. This makes the chance for a collision more likely between two
> > >    RCU dynticks snapshots but wrapping up 28 bits of eqs dynticks
> > >    increments still takes some bad luck (also rdp.dynticks_snap could be
> > >    converted from int to long?)
> > > 
> > > Some RCU eqs functions have been renamed to better reflect their broader
> > > scope that now include context tracking state.
> > > 
> > > Signed-off-by: Frederic Weisbecker <frederic@...nel.org>
> > > Cc: Paul E. McKenney <paulmck@...nel.org>
> > > Cc: Peter Zijlstra <peterz@...radead.org>
> > > Cc: Thomas Gleixner <tglx@...utronix.de>
> > > Cc: Neeraj Upadhyay <quic_neeraju@...cinc.com>
> > > Cc: Uladzislau Rezki <uladzislau.rezki@...y.com>
> > > Cc: Joel Fernandes <joel@...lfernandes.org>
> > > Cc: Boqun Feng <boqun.feng@...il.com>
> > > Cc: Nicolas Saenz Julienne <nsaenz@...nel.org>
> > > Cc: Marcelo Tosatti <mtosatti@...hat.com>
> > > Cc: Xiongfeng Wang <wangxiongfeng2@...wei.com>
> > > Cc: Yu Liao<liaoyu15@...wei.com>
> > > Cc: Phil Auld <pauld@...hat.com>
> > > Cc: Paul Gortmaker<paul.gortmaker@...driver.com>
> > > Cc: Alex Belits <abelits@...vell.com>
> > > ---
> > 
> > While working on a feature on top of this series (IPI deferral stuff) I believe
> > I've found a discrepancy on how context state is being updated:
> > 
> >  - When servicing an IRQ from user-space, we increment dynticks, and clear the
> >    ct state to show we're in-kernel.
> > 
> >  - When servicing an IRQ from idle/guest or an NMI from any context we only
> >    increment the dynticks counter. The ct state remains unchanged.
> 
> Hmm, an IRQ from userspace does:
> 
>     ct_user_enter()
>     //run in user
>         //-----IRQ
>         ct_user_exit()
>         ct_irq_enter()
>         ct_irq_exit()
>         ct_user_enter()
>     //run in user
> 
> An IRQ from guest does:
> 
>     for (;;) {
>          context_tracking_guest_enter()
>         //vmrun
> 	//IRQ pending
>         #VMEXIT
>         context_tracking_guest_exit()
> 	local_irq_enable()
>         ct_irq_enter()
>         ct_irq_exit()
> 	local_irq_disable()
>     }
> 
> 
>     (although I see there is an "sti" right before "vmrun" so it looks
>     possible to have ct_irq_enter() after context_tracking_guest_enter()
>     if a host IRQ fires between the sti and the vmrun though I might be
>     missing some kvm subtelty).
> 
> An IRQ from idle does just:
> 
>     ct_idle_enter()
>         //IRQ
>         ct_irq_enter()
>         ct_irq_exit()
>     ct_idle_exit()
> 
> So guest looks mostly ok to me (except for the little sti before vmrun for
> which I have a doubt).

Yes, shouldn't have mentioned guests. I got carried away.

> But idle at least is an exception and CONTEXT_IDLE will remain during the
> interrupt handling. It's not that trivial to handle the idle case because
> ct_irq_exit() needs to know that it is called between ct_idle_enter() and
> ct_idle_exit().

Just for the record, this behaviour was already here regardless of this series,
so it's not something it needs to fix.

Something like this should work, right?

     ct_idle_enter()
         //IRQ or NMI
	 if (__ct_state() == CONTEXT_IDLE)
	     ct_idle_exit()
         ct_irq_enter()
	    ...
         ct_irq_exit()
	 if (needs_update_state()) //using irqentry_state_t for ex.
	     ct_idle_entry()
     ct_idle_exit()

Note that it's not a big issue as we can work around this behaviour by checking
through dynticks whether a CPU is really idle.

Do you think it's worth fixing nonetheless?

Regards,
Nicolas

Powered by blists - more mailing lists