[PATCH] x86: dovetail: Harden __send_ipi_mask for KVM guests

[PATCH] x86: dovetail: Harden __send_ipi_mask for KVM guests

[Jan Kiszka]

From: Jan Kiszka <jan.kiszka@siemens.com>

This function is called in case paravirtual IPIs of a KVM host are used.
Just like for hard submission handlers, we need to harden the path to
avoid splashes like these:

 IRQ pipeline: some code running in oob context 'Xenomai'
               called an in-band only routine
 CPU: 0 PID: 410 Comm: (udev-worker) Not tainted 5.10.229-cip54+ #265
 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.16.2-3-gd478f380-rebuilt.opensuse.org 04/01/2014
 IRQ stage: Xenomai
 Call Trace:
  dump_stack+0x84/0x9c
  inband_irq_save+0xa/0x30
  __send_ipi_mask+0x81/0x260
  ? preempt_count_sub+0x96/0xe0
  ? _raw_spin_unlock_irq+0x33/0x50
  ? xnsched_pick_next+0xe0/0xe0
  ? ___xnsched_run+0x491/0x5b0
  ? ktime_get_mono_fast_ns+0x50/0xb0
  ___xnsched_run+0x491/0x5b0
  run_oob_call+0x97/0x170
  xnthread_start+0x23f/0x330
  ? rt_stack_deliver+0x340/0x340 [rtnet]
  rtdm_task_init.part.14+0x7e/0xc0

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
---
 arch/x86/kernel/kvm.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
index 5df1061fc6ee1..b183caf41aa44 100644
--- a/arch/x86/kernel/kvm.c
+++ b/arch/x86/kernel/kvm.c
@@ -525,7 +525,7 @@ static void __send_ipi_mask(const struct cpumask *mask, int vector)
 	if (cpumask_empty(mask))
 		return;
 
-	local_irq_save(flags);
+	flags = hard_local_irq_save();
 
 	switch (vector) {
 	default:
@@ -563,7 +563,7 @@ static void __send_ipi_mask(const struct cpumask *mask, int vector)
 			  ret);
 	}
 
-	local_irq_restore(flags);
+	hard_local_irq_restore(flags);
 }
 
 static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
-- 
2.43.0

Re: [PATCH] x86: dovetail: Harden __send_ipi_mask for KVM guests

[Philippe Gerum]

Jan Kiszka <jan.kiszka@siemens.com> writes:

> From: Jan Kiszka <jan.kiszka@siemens.com>
>
> This function is called in case paravirtual IPIs of a KVM host are used.
> Just like for hard submission handlers, we need to harden the path to
> avoid splashes like these:
>
>  IRQ pipeline: some code running in oob context 'Xenomai'
>                called an in-band only routine
>  CPU: 0 PID: 410 Comm: (udev-worker) Not tainted 5.10.229-cip54+ #265
>  Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.16.2-3-gd478f380-rebuilt.opensuse.org 04/01/2014
>  IRQ stage: Xenomai
>  Call Trace:
>   dump_stack+0x84/0x9c
>   inband_irq_save+0xa/0x30
>   __send_ipi_mask+0x81/0x260
>   ? preempt_count_sub+0x96/0xe0
>   ? _raw_spin_unlock_irq+0x33/0x50
>   ? xnsched_pick_next+0xe0/0xe0
>   ? ___xnsched_run+0x491/0x5b0
>   ? ktime_get_mono_fast_ns+0x50/0xb0
>   ___xnsched_run+0x491/0x5b0
>   run_oob_call+0x97/0x170
>   xnthread_start+0x23f/0x330
>   ? rt_stack_deliver+0x340/0x340 [rtnet]
>   rtdm_task_init.part.14+0x7e/0xc0
>
> Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
> ---
>  arch/x86/kernel/kvm.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
> index 5df1061fc6ee1..b183caf41aa44 100644
> --- a/arch/x86/kernel/kvm.c
> +++ b/arch/x86/kernel/kvm.c
> @@ -525,7 +525,7 @@ static void __send_ipi_mask(const struct cpumask *mask, int vector)
>  	if (cpumask_empty(mask))
>  		return;
>  
> -	local_irq_save(flags);
> +	flags = hard_local_irq_save();
>  
>  	switch (vector) {
>  	default:
> @@ -563,7 +563,7 @@ static void __send_ipi_mask(const struct cpumask *mask, int vector)
>  			  ret);
>  	}
>  
> -	local_irq_restore(flags);
> +	hard_local_irq_restore(flags);
>  }
>  
>  static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)

Ack.

-- 
Philippe.

Re: [PATCH] x86: dovetail: Harden __send_ipi_mask for KVM guests

[Florian Bezdeka]

On Mon, 2025-02-24 at 18:15 +0100, Philippe Gerum wrote:
> Jan Kiszka <jan.kiszka@siemens.com> writes:
> 
> > From: Jan Kiszka <jan.kiszka@siemens.com>
> > 
> > This function is called in case paravirtual IPIs of a KVM host are used.
> > Just like for hard submission handlers, we need to harden the path to
> > avoid splashes like these:
> > 
> >  IRQ pipeline: some code running in oob context 'Xenomai'
> >                called an in-band only routine
> >  CPU: 0 PID: 410 Comm: (udev-worker) Not tainted 5.10.229-cip54+ #265
> >  Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.16.2-3-gd478f380-rebuilt.opensuse.org 04/01/2014
> >  IRQ stage: Xenomai
> >  Call Trace:
> >   dump_stack+0x84/0x9c
> >   inband_irq_save+0xa/0x30
> >   __send_ipi_mask+0x81/0x260
> >   ? preempt_count_sub+0x96/0xe0
> >   ? _raw_spin_unlock_irq+0x33/0x50
> >   ? xnsched_pick_next+0xe0/0xe0
> >   ? ___xnsched_run+0x491/0x5b0
> >   ? ktime_get_mono_fast_ns+0x50/0xb0
> >   ___xnsched_run+0x491/0x5b0
> >   run_oob_call+0x97/0x170
> >   xnthread_start+0x23f/0x330
> >   ? rt_stack_deliver+0x340/0x340 [rtnet]
> >   rtdm_task_init.part.14+0x7e/0xc0
> > 
> > Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
> > ---
> >  arch/x86/kernel/kvm.c | 4 ++--
> >  1 file changed, 2 insertions(+), 2 deletions(-)
> > 
> > diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
> > index 5df1061fc6ee1..b183caf41aa44 100644
> > --- a/arch/x86/kernel/kvm.c
> > +++ b/arch/x86/kernel/kvm.c
> > @@ -525,7 +525,7 @@ static void __send_ipi_mask(const struct cpumask *mask, int vector)
> >  	if (cpumask_empty(mask))
> >  		return;
> >  
> > -	local_irq_save(flags);
> > +	flags = hard_local_irq_save();
> >  
> >  	switch (vector) {
> >  	default:
> > @@ -563,7 +563,7 @@ static void __send_ipi_mask(const struct cpumask *mask, int vector)
> >  			  ret);
> >  	}
> >  
> > -	local_irq_restore(flags);
> > +	hard_local_irq_restore(flags);
> >  }
> >  
> >  static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
> 
> Ack.

Ack. The hyperv PV code is also affected (__send_ipi_mask_ex).

> 
> -- 
> Philippe.