We use jump label to enable pv-spinlock. With the changes in
(442e0973e927 Merge branch 'x86/jumplabel'), the jump label behaviour has changed
that would result in evntual hang of the VM since we would end up in a situation
where slowpath locks would halt the vcpus but we will not be able to wakeup the
vcpu by lock releaser using unlock kick.
Similar problem in Xen and more detailed description is available in
a945928ea270 (xen: Do not enable spinlocks before jump_label_init() has executed)
This patch splits kvm_spinlock_init to separate jump label changes with pvops
patching and also make jump label enabling after jump_label_init().
Signed-off-by: Raghavendra K T <[email protected]>
---
Thanks to Andrew Theurer who reported weird behaviour of pvspinlock
in 3.12-rc that led to my git bisection and investigation and Konrad
for his jump label findings for Xen.
arch/x86/kernel/kvm.c | 17 ++++++++++++++---
1 file changed, 14 insertions(+), 3 deletions(-)
diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
index 697b93a..a0e2a8a 100644
--- a/arch/x86/kernel/kvm.c
+++ b/arch/x86/kernel/kvm.c
@@ -775,11 +775,22 @@ void __init kvm_spinlock_init(void)
if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
return;
- printk(KERN_INFO "KVM setup paravirtual spinlock\n");
+ pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(kvm_lock_spinning);
+ pv_lock_ops.unlock_kick = kvm_unlock_kick;
+}
+
+static __init int kvm_spinlock_init_jump(void)
+{
+ if (!kvm_para_available())
+ return 0;
+ if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
+ return 0;
static_key_slow_inc(¶virt_ticketlocks_enabled);
+ printk(KERN_INFO "KVM setup paravirtual spinlock\n");
- pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(kvm_lock_spinning);
- pv_lock_ops.unlock_kick = kvm_unlock_kick;
+ return 0;
}
+early_initcall(kvm_spinlock_init_jump);
+
#endif /* CONFIG_PARAVIRT_SPINLOCKS */
--
1.7.11.7
Il 09/10/2013 11:03, Raghavendra K T ha scritto:
> We use jump label to enable pv-spinlock. With the changes in
> (442e0973e927 Merge branch 'x86/jumplabel'), the jump label behaviour has changed
> that would result in evntual hang of the VM since we would end up in a situation
> where slowpath locks would halt the vcpus but we will not be able to wakeup the
> vcpu by lock releaser using unlock kick.
>
> Similar problem in Xen and more detailed description is available in
> a945928ea270 (xen: Do not enable spinlocks before jump_label_init() has executed)
>
> This patch splits kvm_spinlock_init to separate jump label changes with pvops
> patching and also make jump label enabling after jump_label_init().
>
> Signed-off-by: Raghavendra K T <[email protected]>
> ---
> Thanks to Andrew Theurer who reported weird behaviour of pvspinlock
> in 3.12-rc that led to my git bisection and investigation and Konrad
> for his jump label findings for Xen.
>
> arch/x86/kernel/kvm.c | 17 ++++++++++++++---
> 1 file changed, 14 insertions(+), 3 deletions(-)
>
> diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
> index 697b93a..a0e2a8a 100644
> --- a/arch/x86/kernel/kvm.c
> +++ b/arch/x86/kernel/kvm.c
> @@ -775,11 +775,22 @@ void __init kvm_spinlock_init(void)
> if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
> return;
>
> - printk(KERN_INFO "KVM setup paravirtual spinlock\n");
> + pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(kvm_lock_spinning);
> + pv_lock_ops.unlock_kick = kvm_unlock_kick;
> +}
> +
> +static __init int kvm_spinlock_init_jump(void)
> +{
> + if (!kvm_para_available())
> + return 0;
> + if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
> + return 0;
>
> static_key_slow_inc(¶virt_ticketlocks_enabled);
> + printk(KERN_INFO "KVM setup paravirtual spinlock\n");
>
> - pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(kvm_lock_spinning);
> - pv_lock_ops.unlock_kick = kvm_unlock_kick;
> + return 0;
> }
> +early_initcall(kvm_spinlock_init_jump);
> +
> #endif /* CONFIG_PARAVIRT_SPINLOCKS */
>
Reviewed-by: Paolo Bonzini <[email protected]>
On Wed, 9 Oct 2013 14:33:21 +0530
Raghavendra K T <[email protected]> wrote:
> We use jump label to enable pv-spinlock. With the changes in
> (442e0973e927 Merge branch 'x86/jumplabel'), the jump label behaviour has changed
> that would result in evntual hang of the VM since we would end up in a situation
> where slowpath locks would halt the vcpus but we will not be able to wakeup the
> vcpu by lock releaser using unlock kick.
>
> Similar problem in Xen and more detailed description is available in
> a945928ea270 (xen: Do not enable spinlocks before jump_label_init() has executed)
>
> This patch splits kvm_spinlock_init to separate jump label changes with pvops
> patching and also make jump label enabling after jump_label_init().
>
> Signed-off-by: Raghavendra K T <[email protected]>
Reviewed-by: Steven Rostedt <[email protected]>
-- Steve
> ---
> Thanks to Andrew Theurer who reported weird behaviour of pvspinlock
> in 3.12-rc that led to my git bisection and investigation and Konrad
> for his jump label findings for Xen.
>
> arch/x86/kernel/kvm.c | 17 ++++++++++++++---
> 1 file changed, 14 insertions(+), 3 deletions(-)
>
On 10/09/2013 02:33 PM, Raghavendra K T wrote:
> We use jump label to enable pv-spinlock. With the changes in
> (442e0973e927 Merge branch 'x86/jumplabel'), the jump label behaviour has changed
> that would result in evntual hang of the VM since we would end up in a situation
> where slowpath locks would halt the vcpus but we will not be able to wakeup the
> vcpu by lock releaser using unlock kick.
>
> Similar problem in Xen and more detailed description is available in
> a945928ea270 (xen: Do not enable spinlocks before jump_label_init() has executed)
>
> This patch splits kvm_spinlock_init to separate jump label changes with pvops
> patching and also make jump label enabling after jump_label_init().
>
> Signed-off-by: Raghavendra K T <[email protected]>
Ingo,
Could you please take this fix?
This is reviewed by Paolo and Steven.
On Sun, Oct 13, 2013 at 01:38:57PM +0530, Raghavendra K T wrote:
> On 10/09/2013 02:33 PM, Raghavendra K T wrote:
> >We use jump label to enable pv-spinlock. With the changes in
> >(442e0973e927 Merge branch 'x86/jumplabel'), the jump label behaviour has changed
> >that would result in evntual hang of the VM since we would end up in a situation
> >where slowpath locks would halt the vcpus but we will not be able to wakeup the
> >vcpu by lock releaser using unlock kick.
> >
> >Similar problem in Xen and more detailed description is available in
> >a945928ea270 (xen: Do not enable spinlocks before jump_label_init() has executed)
> >
> >This patch splits kvm_spinlock_init to separate jump label changes with pvops
> >patching and also make jump label enabling after jump_label_init().
> >
> >Signed-off-by: Raghavendra K T <[email protected]>
>
> Ingo,
> Could you please take this fix?
> This is reviewed by Paolo and Steven.
It's KVM, I'll take it.
--
Gleb.