[RFC PATCH] rcu: call kvm_check_and_clear_guest_paused unconditionally

[RFC PATCH] rcu: call kvm_check_and_clear_guest_paused unconditionally

[Sergey Senozhatsky]

Do not call kvm_check_and_clear_guest_paused() only from the
stall branch (which requires an active grace period in the
first place), but instead handle PVCLOCK_GUEST_STOPPED as
early as possible.

pvclock_touch_watchdogs() touches various watchdogs, which
have different timeouts, so the earlier we handle stopped
VCPU the better (lockup watchdog does the same).

Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
---
diff --git a/kernel/rcu/tree_stall.h b/kernel/rcu/tree_stall.h
index 677ee3d8671b..5dd6ea2ead0c 100644
--- a/kernel/rcu/tree_stall.h
+++ b/kernel/rcu/tree_stall.h
@@ -657,6 +657,13 @@ static void check_cpu_stall(struct rcu_data *rdp)
 	unsigned long js;
 	struct rcu_node *rnp;
 
+	/*
+	 * If a virtual machine is stopped by the host it can look to
+	 * the watchdog like an RCU stall. Check to see if the host
+	 * stopped the vm.
+	 */
+	kvm_check_and_clear_guest_paused();
+
 	lockdep_assert_irqs_disabled();
 	if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
 	    !rcu_gp_in_progress())
@@ -699,14 +706,6 @@ static void check_cpu_stall(struct rcu_data *rdp)
 	    (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
 	    cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
 
-		/*
-		 * If a virtual machine is stopped by the host it can look to
-		 * the watchdog like an RCU stall. Check to see if the host
-		 * stopped the vm.
-		 */
-		if (kvm_check_and_clear_guest_paused())
-			return;
-
 		/* We haven't checked in, so go dump stack. */
 		print_cpu_stall(gps);
 		if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
@@ -717,14 +716,6 @@ static void check_cpu_stall(struct rcu_data *rdp)
 		   ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
 		   cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
 
-		/*
-		 * If a virtual machine is stopped by the host it can look to
-		 * the watchdog like an RCU stall. Check to see if the host
-		 * stopped the vm.
-		 */
-		if (kvm_check_and_clear_guest_paused())
-			return;
-
 		/* They had a few time units to dump stack, so complain. */
 		print_other_cpu_stall(gs2, gps);
 		if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
-- 
2.32.0.402.g57bb445576-goog

Re: [RFC PATCH] rcu: call kvm_check_and_clear_guest_paused unconditionally

[Sergey Senozhatsky]

On (21/07/16 14:41), Sergey Senozhatsky wrote:
> @@ -657,6 +657,13 @@ static void check_cpu_stall(struct rcu_data *rdp)
>  	unsigned long js;
>  	struct rcu_node *rnp;
>  
> +	/*
> +	 * If a virtual machine is stopped by the host it can look to
> +	 * the watchdog like an RCU stall. Check to see if the host
> +	 * stopped the vm.
> +	 */
> +	kvm_check_and_clear_guest_paused();
> +
>  	lockdep_assert_irqs_disabled();
>  	if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
>  	    !rcu_gp_in_progress())
> @@ -699,14 +706,6 @@ static void check_cpu_stall(struct rcu_data *rdp)
>  	    (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
>  	    cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
>  
> -		/*
> -		 * If a virtual machine is stopped by the host it can look to
> -		 * the watchdog like an RCU stall. Check to see if the host
> -		 * stopped the vm.
> -		 */
> -		if (kvm_check_and_clear_guest_paused())
> -			return;
> -
>  		/* We haven't checked in, so go dump stack. */
>  		print_cpu_stall(gps);
>  		if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
> @@ -717,14 +716,6 @@ static void check_cpu_stall(struct rcu_data *rdp)
>  		   ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
>  		   cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
>  
> -		/*
> -		 * If a virtual machine is stopped by the host it can look to
> -		 * the watchdog like an RCU stall. Check to see if the host
> -		 * stopped the vm.
> -		 */
> -		if (kvm_check_and_clear_guest_paused())
> -			return;
> -
>  		/* They had a few time units to dump stack, so complain. */
>  		print_other_cpu_stall(gs2, gps);
>  		if (READ_ONCE(rcu_cpu_stall_ftrace_dump))

This patch depends on
https://lore.kernel.org/lkml/20210716053405.1243239-1-senozhatsky@chromium.org/

If that x86/kvm patch lands, then we need to handle
PVCLOCK_GUEST_STOPPED in watchdogs.


In theory, this patch opens a small race window, if the VCPU gets preempted
after kvm_check_and_clear_guest_paused() (external interrupt, etc.)
But it's hard to tell how likely the problem is.

Re: [RFC PATCH] rcu: call kvm_check_and_clear_guest_paused unconditionally

[Paul E. McKenney]

On Fri, Jul 16, 2021 at 03:23:07PM +0900, Sergey Senozhatsky wrote:
> On (21/07/16 14:41), Sergey Senozhatsky wrote:
> > @@ -657,6 +657,13 @@ static void check_cpu_stall(struct rcu_data *rdp)
> >  	unsigned long js;
> >  	struct rcu_node *rnp;
> >  
> > +	/*
> > +	 * If a virtual machine is stopped by the host it can look to
> > +	 * the watchdog like an RCU stall. Check to see if the host
> > +	 * stopped the vm.
> > +	 */
> > +	kvm_check_and_clear_guest_paused();
> > +
> >  	lockdep_assert_irqs_disabled();
> >  	if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
> >  	    !rcu_gp_in_progress())
> > @@ -699,14 +706,6 @@ static void check_cpu_stall(struct rcu_data *rdp)
> >  	    (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
> >  	    cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
> >  
> > -		/*
> > -		 * If a virtual machine is stopped by the host it can look to
> > -		 * the watchdog like an RCU stall. Check to see if the host
> > -		 * stopped the vm.
> > -		 */
> > -		if (kvm_check_and_clear_guest_paused())
> > -			return;
> > -
> >  		/* We haven't checked in, so go dump stack. */
> >  		print_cpu_stall(gps);
> >  		if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
> > @@ -717,14 +716,6 @@ static void check_cpu_stall(struct rcu_data *rdp)
> >  		   ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
> >  		   cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
> >  
> > -		/*
> > -		 * If a virtual machine is stopped by the host it can look to
> > -		 * the watchdog like an RCU stall. Check to see if the host
> > -		 * stopped the vm.
> > -		 */
> > -		if (kvm_check_and_clear_guest_paused())
> > -			return;
> > -
> >  		/* They had a few time units to dump stack, so complain. */
> >  		print_other_cpu_stall(gs2, gps);
> >  		if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
> 
> This patch depends on
> https://lore.kernel.org/lkml/20210716053405.1243239-1-senozhatsky@chromium.org/
> 
> If that x86/kvm patch lands, then we need to handle
> PVCLOCK_GUEST_STOPPED in watchdogs.

OK, please let me know how and when you would like to proceed.

> In theory, this patch opens a small race window, if the VCPU gets preempted
> after kvm_check_and_clear_guest_paused() (external interrupt, etc.)
> But it's hard to tell how likely the problem is.

There is always attempting to provoke it, possibly accompanied by
artificially widening the race window.

							Thanx, Paul