mirror of https://lore.kernel.org/lkml/
 help / color / mirror / Atom feed
* [PATCH] KVM: x86: Reevaluate pending interrupts when enabling ExtINT through LVT0
@ 2026-10-08 12:49 Matthew Leach
  2026-11-04 20:09 ` Sean Christopherson
  0 siblings, 1 reply; 3+ messages in thread
From: Matthew Leach @ 2026-10-08 12:49 UTC (permalink / raw)
  To: Sean Christopherson, Paolo Bonzini, Thomas Gleixner, Ingo Molnar,
	Borislav Petkov, Dave Hansen, x86, H. Peter Anvin
  Cc: kvm, linux-kernel, kernel, Matthew Leach

If the PIC output becomes asserted while LVT0 is masked, unmasking LVT0
in ExtINT mode makes that interrupt injectable without setting
KVM_REQ_EVENT. Without another event triggering reevaluation, the
interrupt remains pending and a guest waiting for timer ticks can hang.

Fix by setting KVM_REQ_EVENT on an LVT0 write where pic interrupts
become acceptable.

Signed-off-by: Matthew Leach <matthew.leach@collabora.com>
---
During calibrate_delay_converge() we spin waiting for jiffies to increment on a
tick interrupt. However, if KVM fails to deliver the interrupt this causes the
guest to hang, spinning on the while loop.

Reproduce with:

qemu-system-x86_64 \
  -cpu host \
  -accel kvm \
  -smp 1 \
  -machine pc,hpet=off \
  -kernel <bzimage> \
  -nographic \
  -append "console=ttyS0,115200 noapic apic=debug no-kvmclock clearcpuid=tsc"

Resulting in a hang at:

printk: legacy console [ttyS0] enabled
printk: legacy console [ttyS0] enabled
printk: legacy bootconsole [uart8250] disabled
printk: legacy bootconsole [uart8250] disabled
ACPI: Core revision 20260408
ACPI: setting ELCR to 0200 (from 0c00)
APIC: Switch to symmetric I/O mode setup
Not enabling interrupt remapping due to skipped IO-APIC setup
Enabled ExtINT on CPU#0

Note that the bug manifests itself only if a PIC interrupt is asserting during
the mask window, so it is timing based. Using nested virt helps, or apply the
following patch to the guest to provide enough time for a tick after masking:

--8<---------------cut here---------------start------------->8---
diff --git a/arch/x86/kernel/apic/apic.c b/arch/x86/kernel/apic/apic.c
index aa1e19979aa8..ddee09360dd0 100644
--- a/arch/x86/kernel/apic/apic.c
+++ b/arch/x86/kernel/apic/apic.c
@@ -1509,7 +1509,7 @@ static void apic_clear_isr(void)
 static void setup_local_APIC(void)
 {
 	int cpu = smp_processor_id();
-	unsigned int value;
+	unsigned int value, i;
 
 	if (apic_is_disabled) {
 		disable_ioapic_support();
@@ -1527,6 +1527,9 @@ static void setup_local_APIC(void)
 	value &= ~APIC_SPIV_APIC_ENABLED;
 	apic_write(APIC_SPIV, value);
 
+	for (i = 0; i < 10000000UL; i++)
+		cpu_relax();
+
 #ifdef CONFIG_X86_32
 	/* Pound the ESR really hard over the head with a big hammer - mbligh */
 	if (lapic_is_integrated() && apic->disable_esr) {
--8<---------------cut here---------------end--------------->8---
---
 arch/x86/kvm/lapic.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
index e1f3cea14765..ce7fbd6c45c2 100644
--- a/arch/x86/kvm/lapic.c
+++ b/arch/x86/kvm/lapic.c
@@ -2520,6 +2520,14 @@ static int kvm_lapic_reg_write(struct kvm_lapic *apic, u32 reg, u32 val)
 			val |= APIC_LVT_MASKED;
 		val &= apic_lvt_mask[index];
 		kvm_lapic_set_reg(apic, reg, val);
+
+		/*
+		 * Enabling ExtINT can make an already-asserted PIC interrupt
+		 * deliverable without another transition on the PIC output.
+		 */
+		if (reg == APIC_LVT0 && kvm_apic_accept_pic_intr(apic->vcpu))
+			kvm_make_request(KVM_REQ_EVENT, apic->vcpu);
+
 		break;
 	}
 

---
base-commit: a90ee4305c4a5df72c11b31dacfdc76e00fcf78a
change-id: 20261008-lapic-extint-lvt0-unmask-fix-52d2328af682

Best regards,
--  
Matt


^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: [PATCH] KVM: x86: Reevaluate pending interrupts when enabling ExtINT through LVT0
  2026-10-08 12:49 [PATCH] KVM: x86: Reevaluate pending interrupts when enabling ExtINT through LVT0 Matthew Leach
@ 2026-11-04 20:09 ` Sean Christopherson
  2026-11-16 14:17   ` Jiaming Fan
  0 siblings, 1 reply; 3+ messages in thread
From: Sean Christopherson @ 2026-11-04 20:09 UTC (permalink / raw)
  To: Matthew Leach
  Cc: Paolo Bonzini, Thomas Gleixner, Ingo Molnar, Borislav Petkov,
	Dave Hansen, x86, H. Peter Anvin, kvm, linux-kernel, kernel

On Thu, Oct 08, 2026, Matthew Leach wrote:
> If the PIC output becomes asserted while LVT0 is masked, unmasking LVT0
> in ExtINT mode makes that interrupt injectable without setting
> KVM_REQ_EVENT. Without another event triggering reevaluation, the
> interrupt remains pending and a guest waiting for timer ticks can hang.
> 
> Fix by setting KVM_REQ_EVENT on an LVT0 write where pic interrupts
> become acceptable.

Why just for LVT0?  Isn't this a bug in general?  I.e. can't we simply do:

diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
index d6e5fbc3c58c..ffdc67ca8fa3 100644
--- a/arch/x86/kvm/lapic.c
+++ b/arch/x86/kvm/lapic.c
@@ -2554,6 +2554,8 @@ static int kvm_lapic_reg_write(struct kvm_lapic *apic, u32 reg, u32 val)
 			val |= APIC_LVT_MASKED;
 		val &= apic_lvt_mask[index];
 		kvm_lapic_set_reg(apic, reg, val);
+		if (!(val & APIC_LVT_MASKED))
+			kvm_make_request(KVM_REQ_EVENT, apic->vcpu);
 		break;
 	}
 
Ahh, but there's nowhere for edge-triggered events to be latched, so this only
affects level-triggered events.  Which means, this?

diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
index d6e5fbc3c58c..b72930a8516d 100644
--- a/arch/x86/kvm/lapic.c
+++ b/arch/x86/kvm/lapic.c
@@ -2554,6 +2554,8 @@ static int kvm_lapic_reg_write(struct kvm_lapic *apic, u32 reg, u32 val)
 			val |= APIC_LVT_MASKED;
 		val &= apic_lvt_mask[index];
 		kvm_lapic_set_reg(apic, reg, val);
+		if (!(val & APIC_LVT_MASKED) && (val & APIC_LVT_LEVEL_TRIGGER))
+			kvm_make_request(KVM_REQ_EVENT, apic->vcpu);
 		break;
 	}
 
> Signed-off-by: Matthew Leach <matthew.leach@collabora.com>
> ---
> During calibrate_delay_converge() we spin waiting for jiffies to increment on a
> tick interrupt. However, if KVM fails to deliver the interrupt this causes the
> guest to hang, spinning on the while loop.
> 
> Reproduce with:
> 
> qemu-system-x86_64 \
>   -cpu host \
>   -accel kvm \
>   -smp 1 \
>   -machine pc,hpet=off \
>   -kernel <bzimage> \
>   -nographic \
>   -append "console=ttyS0,115200 noapic apic=debug no-kvmclock clearcpuid=tsc"
> 
> Resulting in a hang at:
> 
> printk: legacy console [ttyS0] enabled
> printk: legacy console [ttyS0] enabled
> printk: legacy bootconsole [uart8250] disabled
> printk: legacy bootconsole [uart8250] disabled
> ACPI: Core revision 20260408
> ACPI: setting ELCR to 0200 (from 0c00)
> APIC: Switch to symmetric I/O mode setup
> Not enabling interrupt remapping due to skipped IO-APIC setup
> Enabled ExtINT on CPU#0
> 
> Note that the bug manifests itself only if a PIC interrupt is asserting during
> the mask window, so it is timing based. Using nested virt helps, or apply the
> following patch to the guest to provide enough time for a tick after masking:
> 
> --8<---------------cut here---------------start------------->8---
> diff --git a/arch/x86/kernel/apic/apic.c b/arch/x86/kernel/apic/apic.c
> index aa1e19979aa8..ddee09360dd0 100644
> --- a/arch/x86/kernel/apic/apic.c
> +++ b/arch/x86/kernel/apic/apic.c
> @@ -1509,7 +1509,7 @@ static void apic_clear_isr(void)
>  static void setup_local_APIC(void)
>  {
>  	int cpu = smp_processor_id();
> -	unsigned int value;
> +	unsigned int value, i;
>  
>  	if (apic_is_disabled) {
>  		disable_ioapic_support();
> @@ -1527,6 +1527,9 @@ static void setup_local_APIC(void)
>  	value &= ~APIC_SPIV_APIC_ENABLED;
>  	apic_write(APIC_SPIV, value);
>  
> +	for (i = 0; i < 10000000UL; i++)
> +		cpu_relax();
> +
>  #ifdef CONFIG_X86_32
>  	/* Pound the ESR really hard over the head with a big hammer - mbligh */
>  	if (lapic_is_integrated() && apic->disable_esr) {
> --8<---------------cut here---------------end--------------->8---
> ---
>  arch/x86/kvm/lapic.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
> 
> diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
> index e1f3cea14765..ce7fbd6c45c2 100644
> --- a/arch/x86/kvm/lapic.c
> +++ b/arch/x86/kvm/lapic.c
> @@ -2520,6 +2520,14 @@ static int kvm_lapic_reg_write(struct kvm_lapic *apic, u32 reg, u32 val)
>  			val |= APIC_LVT_MASKED;
>  		val &= apic_lvt_mask[index];
>  		kvm_lapic_set_reg(apic, reg, val);
> +
> +		/*
> +		 * Enabling ExtINT can make an already-asserted PIC interrupt
> +		 * deliverable without another transition on the PIC output.
> +		 */
> +		if (reg == APIC_LVT0 && kvm_apic_accept_pic_intr(apic->vcpu))
> +			kvm_make_request(KVM_REQ_EVENT, apic->vcpu);
> +
>  		break;
>  	}
>  
> 
> ---
> base-commit: a90ee4305c4a5df72c11b31dacfdc76e00fcf78a
> change-id: 20261008-lapic-extint-lvt0-unmask-fix-52d2328af682
> 
> Best regards,
> --  
> Matt
> 
> 

^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: [PATCH] KVM: x86: Reevaluate pending interrupts when enabling ExtINT through LVT0
  2026-11-04 20:09 ` Sean Christopherson
@ 2026-11-16 14:17   ` Jiaming Fan
  0 siblings, 0 replies; 3+ messages in thread
From: Jiaming Fan @ 2026-11-16 14:17 UTC (permalink / raw)
  To: Sean Christopherson
  Cc: Matthew Leach, Paolo Bonzini, kvm, linux-kernel, kernel

On Wed, Nov 04, 2026, Sean Christopherson wrote:
> Ahh, but there's nowhere for edge-triggered events to be latched, so this only
> affects level-triggered events.  Which means, this?
>
> diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
> index d6e5fbc3c58c..b72930a8516d 100644
> --- a/arch/x86/kvm/lapic.c
> +++ b/arch/x86/kvm/lapic.c
> @@ -2554,6 +2554,8 @@ static int kvm_lapic_reg_write(struct kvm_lapic *apic, u32 reg, u32 val)
>  			val |= APIC_LVT_MASKED;
>  		val &= apic_lvt_mask[index];
>  		kvm_lapic_set_reg(apic, reg, val);
> +		if (!(val & APIC_LVT_MASKED) && (val & APIC_LVT_LEVEL_TRIGGER))
> +			kvm_make_request(KVM_REQ_EVENT, apic->vcpu);
>  		break;
>  	}

The trigger mode check doesn't catch the case Matthew hit.  Linux's
setup_local_APIC() writes LVT0 as plain APIC_DM_EXTINT, without
APIC_LVT_LEVEL_TRIGGER, so no request is made and the "noapic" hang
remains.  For ExtINT the pending interrupt is latched in the 8259A,
whose output stays asserted until the guest acks it, not in the LVT
entry, so the trigger mode bit doesn't tell whether anything is
pending.  Your first version (any unmasked LVT) does work for LVT0.

There is a second path with the same problem that neither version
covers: hardware-disabling the APIC through IA32_APIC_BASE.  With the
APIC disabled, kvm_apic_accept_pic_intr() returns true, so an ExtINT
latched while LVT0 was masked becomes deliverable, but
__kvm_apic_set_base() doesn't request KVM_REQ_EVENT either:

diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
--- a/arch/x86/kvm/lapic.c
+++ b/arch/x86/kvm/lapic.c
@@ -2825,6 +2825,7 @@ static void __kvm_apic_set_base(struct kvm_vcpu *vcpu, u64 value)
 		} else {
 			static_branch_inc(&apic_hw_disabled.key);
 			atomic_set_release(&apic->vcpu->kvm->arch.apic_map_dirty, DIRTY);
+			kvm_make_request(KVM_REQ_EVENT, vcpu);
 		}
 	}

I ran into the same hang independently (a small VMM booting Linux with
"noapic" on the in-kernel irqchip, roughly 1 in 5 boots when nested) and
wrote a selftest that reproduces both paths deterministically.  The
guest programs the master PIC and masks LVT0, the host raises IRQ4 with
KVM_IRQ_LINE and checks that the PIC latched it, then the guest unmasks
LVT0 (directly, or after an SPIV soft-disable as setup_local_APIC()
does) or hardware-disables the APIC, and opens a single interrupt window
(sti; nop; cli) to count deliveries.  It runs with xAPIC and x2APIC.

Results on kvm-x86/next (cbe54fc3462b), 3+ runs each, all deterministic:

  kernel                                 selftest result
  -------------------------------------  ----------------------------------
  unpatched                              "LVT0 unmasked after masking:
                                          got 1 IRQs, expected 2"
  Matthew's v1 condition                 "APIC hardware-disabled:
                                          got 3 IRQs, expected 4"
  any unmasked LVT (your 1st version)    "APIC hardware-disabled:
                                          got 3 IRQs, expected 4"
  unmasked + level (your 2nd version)    "LVT0 unmasked after masking:
                                          got 1 IRQs, expected 2"
  any unmasked LVT + APIC_BASE fix       pass

Tested on AMD only, nested under QEMU with AVIC disabled.  Not tested on
Intel or with APICv/AVIC enabled.  With Matthew's condition plus the
APIC_BASE fix, the full KVM selftests differ from the unpatched kernel
only in the new test, and kvm-unit-tests show no differences beyond the
"MOV SS + watchpoint" cases of the debug test, which fail intermittently
on both kernels in this nested setup.  I haven't run the full suites
with your first version yet.

Matthew, are you planning a v2?  If so, I'm happy to hand over the
selftest and the APIC_BASE fix.  Otherwise I can post a series with your
patch (reworked per Sean's first suggestion) followed by the APIC_BASE
fix and the selftest.

Thanks,
Jiaming


^ permalink raw reply	[flat|nested] 3+ messages in thread

end of thread, other threads:[~2026-11-16 14:17 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-10-08 12:49 [PATCH] KVM: x86: Reevaluate pending interrupts when enabling ExtINT through LVT0 Matthew Leach
2026-11-04 20:09 ` Sean Christopherson
2026-11-16 14:17   ` Jiaming Fan

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox

all inboxes | Powered by JetHome®