commit 334f3f6d7a16 ("powerpc/entry: Disable interrupts before irqentry_exit")
fixed a BUG in preempt_schedule_irq() by calling local_irq_disable()
unconditionally in arch_interrupt_exit_prepare() before irqentry_exit().
The fix is correct in intent but uses the wrong primitive for PPC64.
On PPC64, local_irq_disable() only sets the soft-mask
(irq_soft_mask = IRQS_DISABLED). It does not set PACA_IRQ_HARD_DIS in
irq_happened and does not clear MSR[EE].
This causes a continuous WARN_ON boot hang on Power11 pSeries LPARs
using the dedicated-cede cpuidle path. When the CPU wakes from H_CEDE,
interrupt_exit_kernel_prepare() calls replay_soft_interrupts(), which
dispatches pending async handlers (timer_interrupt, do_IRQ) using the
DEFINE_INTERRUPT_HANDLER_ASYNC macro. That macro calls
arch_interrupt_async_exit_prepare() -> arch_interrupt_exit_prepare()
before irqentry_exit(). With local_irq_disable(), PACA_IRQ_HARD_DIS
is not set, but next_interrupt() unconditionally asserts it:
WARN_ON(!(local_paca->irq_happened & PACA_IRQ_HARD_DIS));
This fires on every replayed interrupt, looping indefinitely and
preventing boot completion.
Fix this by replacing local_irq_disable() with hard_irq_disable().
On PPC64, hard_irq_disable() sets irq_soft_mask to IRQS_ALL_DISABLED,
sets PACA_IRQ_HARD_DIS in irq_happened, and clears MSR[EE] — satisfying
all of:
- lockdep_assert_irqs_disabled() in irqentry_exit_to_kernel_mode()
- next_interrupt()'s WARN_ON(!(irq_happened & PACA_IRQ_HARD_DIS))
- preempt_schedule_irq()'s BUG_ON(!irqs_disabled())
On PPC32/non-64, hard_irq_disable() is equivalent to local_irq_disable(),
so there is no regression on those platforms.
Fixes: 334f3f6d7a16 ("powerpc/entry: Disable interrupts before irqentry_exit")
Reported-by: Venkat Rao Bagalkote <redacted>
Closes: https://lore.kernel.org/all/6f9bfb0f-b14c-468e-bb9f-c157d120d0dc@linux.ibm.com/
Signed-off-by: Mukesh Kumar Chaurasiya (IBM) <redacted>
---
arch/powerpc/include/asm/entry-common.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
@@ -270,7 +270,7 @@ static inline void arch_interrupt_exit_prepare(struct pt_regs *regs)}/* irqentry_exit expects to be called with interrupts disabled */-local_irq_disable();+hard_irq_disable();}staticinlinevoidarch_interrupt_async_enter_prepare(structpt_regs*regs)
On 12/08/26 8:50 pm, Mukesh Kumar Chaurasiya (IBM) wrote:
commit 334f3f6d7a16 ("powerpc/entry: Disable interrupts before irqentry_exit")
fixed a BUG in preempt_schedule_irq() by calling local_irq_disable()
unconditionally in arch_interrupt_exit_prepare() before irqentry_exit().
The fix is correct in intent but uses the wrong primitive for PPC64.
On PPC64, local_irq_disable() only sets the soft-mask
(irq_soft_mask = IRQS_DISABLED). It does not set PACA_IRQ_HARD_DIS in
irq_happened and does not clear MSR[EE].
This causes a continuous WARN_ON boot hang on Power11 pSeries LPARs
using the dedicated-cede cpuidle path. When the CPU wakes from H_CEDE,
interrupt_exit_kernel_prepare() calls replay_soft_interrupts(), which
dispatches pending async handlers (timer_interrupt, do_IRQ) using the
DEFINE_INTERRUPT_HANDLER_ASYNC macro. That macro calls
arch_interrupt_async_exit_prepare() -> arch_interrupt_exit_prepare()
before irqentry_exit(). With local_irq_disable(), PACA_IRQ_HARD_DIS
is not set, but next_interrupt() unconditionally asserts it:
WARN_ON(!(local_paca->irq_happened & PACA_IRQ_HARD_DIS));
This fires on every replayed interrupt, looping indefinitely and
preventing boot completion.
Fix this by replacing local_irq_disable() with hard_irq_disable().
On PPC64, hard_irq_disable() sets irq_soft_mask to IRQS_ALL_DISABLED,
sets PACA_IRQ_HARD_DIS in irq_happened, and clears MSR[EE] — satisfying
all of:
- lockdep_assert_irqs_disabled() in irqentry_exit_to_kernel_mode()
- next_interrupt()'s WARN_ON(!(irq_happened & PACA_IRQ_HARD_DIS))
- preempt_schedule_irq()'s BUG_ON(!irqs_disabled())
On PPC32/non-64, hard_irq_disable() is equivalent to local_irq_disable(),
so there is no regression on those platforms.
Fixes: 334f3f6d7a16 ("powerpc/entry: Disable interrupts before irqentry_exit")
Reported-by: Venkat Rao Bagalkote <redacted>
Closes: https://lore.kernel.org/all/6f9bfb0f-b14c-468e-bb9f-c157d120d0dc@linux.ibm.com/
Signed-off-by: Mukesh Kumar Chaurasiya (IBM) <redacted>
---
Tested this, and it fixes the reported issue.
Tested-by: Venkat Rao Bagalkote <redacted>
Regards,
Venkat.
@@ -270,7 +270,7 @@ static inline void arch_interrupt_exit_prepare(struct pt_regs *regs)}/* irqentry_exit expects to be called with interrupts disabled */-local_irq_disable();+hard_irq_disable();}staticinlinevoidarch_interrupt_async_enter_prepare(structpt_regs*regs)
Le 13/08/2026 à 08:36, Venkat Rao Bagalkote a écrit :
On 12/08/26 8:50 pm, Mukesh Kumar Chaurasiya (IBM) wrote:
quoted
commit 334f3f6d7a16 ("powerpc/entry: Disable interrupts before
irqentry_exit")
fixed a BUG in preempt_schedule_irq() by calling local_irq_disable()
unconditionally in arch_interrupt_exit_prepare() before irqentry_exit().
The fix is correct in intent but uses the wrong primitive for PPC64.
On PPC64, local_irq_disable() only sets the soft-mask
(irq_soft_mask = IRQS_DISABLED). It does not set PACA_IRQ_HARD_DIS in
irq_happened and does not clear MSR[EE].
This causes a continuous WARN_ON boot hang on Power11 pSeries LPARs
using the dedicated-cede cpuidle path. When the CPU wakes from H_CEDE,
interrupt_exit_kernel_prepare() calls replay_soft_interrupts(), which
dispatches pending async handlers (timer_interrupt, do_IRQ) using the
DEFINE_INTERRUPT_HANDLER_ASYNC macro. That macro calls
arch_interrupt_async_exit_prepare() -> arch_interrupt_exit_prepare()
before irqentry_exit(). With local_irq_disable(), PACA_IRQ_HARD_DIS
is not set, but next_interrupt() unconditionally asserts it:
WARN_ON(!(local_paca->irq_happened & PACA_IRQ_HARD_DIS));
This fires on every replayed interrupt, looping indefinitely and
preventing boot completion.
Fix this by replacing local_irq_disable() with hard_irq_disable().
On PPC64, hard_irq_disable() sets irq_soft_mask to IRQS_ALL_DISABLED,
sets PACA_IRQ_HARD_DIS in irq_happened, and clears MSR[EE] — satisfying
all of:
- lockdep_assert_irqs_disabled() in irqentry_exit_to_kernel_mode()
- next_interrupt()'s WARN_ON(!(irq_happened & PACA_IRQ_HARD_DIS))
- preempt_schedule_irq()'s BUG_ON(!irqs_disabled())
On PPC32/non-64, hard_irq_disable() is equivalent to local_irq_disable(),
so there is no regression on those platforms.
Fixes: 334f3f6d7a16 ("powerpc/entry: Disable interrupts before
irqentry_exit")
Reported-by: Venkat Rao Bagalkote <redacted>
Closes: https://eur01.safelinks.protection.outlook.com/?
url=https%3A%2F%2Flore.kernel.org%2Fall%2F6f9bfb0f-b14c-468e-bb9f-
c157d120d0dc%40linux.ibm.com%2F&data=05%7C02%7Cchristophe.leroy%40csgroup.eu%7Cf780f6a2abcf425c875808def90533e4%7C8b87af7d86474dc78df45f69a2011bb5%7C0%7C0%7C639221997929524168%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=kOer%2B610LkQs0ZKBn%2FutIB6j1bMoFasK4DmX4AEWOXA%3D&reserved=0
Signed-off-by: Mukesh Kumar Chaurasiya (IBM) <redacted>
---
Tested this, and it fixes the reported issue.
Tested-by: Venkat Rao Bagalkote <redacted>
On 13/08/26 12:12 pm, Christophe Leroy (CS GROUP) wrote:
Le 13/08/2026 à 08:36, Venkat Rao Bagalkote a écrit :
quoted
On 12/08/26 8:50 pm, Mukesh Kumar Chaurasiya (IBM) wrote:
quoted
commit 334f3f6d7a16 ("powerpc/entry: Disable interrupts before
irqentry_exit")
fixed a BUG in preempt_schedule_irq() by calling local_irq_disable()
unconditionally in arch_interrupt_exit_prepare() before
irqentry_exit().
The fix is correct in intent but uses the wrong primitive for PPC64.
On PPC64, local_irq_disable() only sets the soft-mask
(irq_soft_mask = IRQS_DISABLED). It does not set PACA_IRQ_HARD_DIS in
irq_happened and does not clear MSR[EE].
This causes a continuous WARN_ON boot hang on Power11 pSeries LPARs
using the dedicated-cede cpuidle path. When the CPU wakes from H_CEDE,
interrupt_exit_kernel_prepare() calls replay_soft_interrupts(), which
dispatches pending async handlers (timer_interrupt, do_IRQ) using the
DEFINE_INTERRUPT_HANDLER_ASYNC macro. That macro calls
arch_interrupt_async_exit_prepare() -> arch_interrupt_exit_prepare()
before irqentry_exit(). With local_irq_disable(), PACA_IRQ_HARD_DIS
is not set, but next_interrupt() unconditionally asserts it:
WARN_ON(!(local_paca->irq_happened & PACA_IRQ_HARD_DIS));
This fires on every replayed interrupt, looping indefinitely and
preventing boot completion.
Fix this by replacing local_irq_disable() with hard_irq_disable().
On PPC64, hard_irq_disable() sets irq_soft_mask to IRQS_ALL_DISABLED,
sets PACA_IRQ_HARD_DIS in irq_happened, and clears MSR[EE] — satisfying
all of:
- lockdep_assert_irqs_disabled() in irqentry_exit_to_kernel_mode()
- next_interrupt()'s WARN_ON(!(irq_happened & PACA_IRQ_HARD_DIS))
- preempt_schedule_irq()'s BUG_ON(!irqs_disabled())
On PPC32/non-64, hard_irq_disable() is equivalent to
local_irq_disable(),
so there is no regression on those platforms.
Fixes: 334f3f6d7a16 ("powerpc/entry: Disable interrupts before
irqentry_exit")
Reported-by: Venkat Rao Bagalkote <redacted>
Closes: https://eur01.safelinks.protection.outlook.com/?
url=https%3A%2F%2Flore.kernel.org%2Fall%2F6f9bfb0f-b14c-468e-bb9f-
c157d120d0dc%40linux.ibm.com%2F&data=05%7C02%7Cchristophe.leroy%40csgroup.eu%7Cf780f6a2abcf425c875808def90533e4%7C8b87af7d86474dc78df45f69a2011bb5%7C0%7C0%7C639221997929524168%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=kOer%2B610LkQs0ZKBn%2FutIB6j1bMoFasK4DmX4AEWOXA%3D&reserved=0
Signed-off-by: Mukesh Kumar Chaurasiya (IBM) <redacted>
---
Tested this, and it fixes the reported issue.
Tested-by: Venkat Rao Bagalkote <redacted>
Hi Christophe,
I think there are two separate issues being mixed together here, both of
which were reported by me.
The first issue is the early boot hang. I tested the patch in that
context and reported that the boot hang still persisted.
The second issue is the WARN in __replay_soft_interrupts(). For that
issue, the same patch fixes the problem, which is why I provided the
Tested-by tag.
I believe the confusion comes from the fact that the same patch ended up
being discussed in relation to both reports. My earlier comment that the
patch did not fix the issue was referring to the boot hang, while my
Tested-by was for the WARN in __replay_soft_interrupts().
So, to clarify:
Boot hang issue: not fixed by the patch.
WARN in __replay_soft_interrupts(): fixed by the patch.
Regards,
Venkat
On 8/12/26 8:50 PM, Mukesh Kumar Chaurasiya (IBM) wrote:
commit 334f3f6d7a16 ("powerpc/entry: Disable interrupts before irqentry_exit")
fixed a BUG in preempt_schedule_irq() by calling local_irq_disable()
unconditionally in arch_interrupt_exit_prepare() before irqentry_exit().
The fix is correct in intent but uses the wrong primitive for PPC64.
On PPC64, local_irq_disable() only sets the soft-mask
(irq_soft_mask = IRQS_DISABLED). It does not set PACA_IRQ_HARD_DIS in
irq_happened and does not clear MSR[EE].
This causes a continuous WARN_ON boot hang on Power11 pSeries LPARs
using the dedicated-cede cpuidle path. When the CPU wakes from H_CEDE,
interrupt_exit_kernel_prepare() calls replay_soft_interrupts(), which
dispatches pending async handlers (timer_interrupt, do_IRQ) using the
DEFINE_INTERRUPT_HANDLER_ASYNC macro. That macro calls
arch_interrupt_async_exit_prepare() -> arch_interrupt_exit_prepare()
before irqentry_exit(). With local_irq_disable(), PACA_IRQ_HARD_DIS
is not set, but next_interrupt() unconditionally asserts it:
WARN_ON(!(local_paca->irq_happened & PACA_IRQ_HARD_DIS));
This fires on every replayed interrupt, looping indefinitely and
preventing boot completion.
Fix this by replacing local_irq_disable() with hard_irq_disable().
On PPC64, hard_irq_disable() sets irq_soft_mask to IRQS_ALL_DISABLED,
sets PACA_IRQ_HARD_DIS in irq_happened, and clears MSR[EE] — satisfying
all of:
- lockdep_assert_irqs_disabled() in irqentry_exit_to_kernel_mode()
- next_interrupt()'s WARN_ON(!(irq_happened & PACA_IRQ_HARD_DIS))
- preempt_schedule_irq()'s BUG_ON(!irqs_disabled())
On PPC32/non-64, hard_irq_disable() is equivalent to local_irq_disable(),
so there is no regression on those platforms.
This is good explanation, but what i not understanding is,
why pattern of setting PACA_IRQ_HARD_DIS changed?
Previous code at interrupt_exit_kernel_prepare which did local_irq_disable too.
Please check where was PACA_IRQ_HARD_DIS set without GENERIC_ENTRY which was preventing
this from happening?
@@ -270,7 +270,7 @@ static inline void arch_interrupt_exit_prepare(struct pt_regs *regs)}/* irqentry_exit expects to be called with interrupts disabled */-local_irq_disable();+hard_irq_disable();}staticinlinevoidarch_interrupt_async_enter_prepare(structpt_regs*regs)
On Fri, Aug 14, 2026 at 11:06:43AM +0530, Shrikanth Hegde wrote:
[...]
This is good explanation, but what i not understanding is,
why pattern of setting PACA_IRQ_HARD_DIS changed?
Previous code at interrupt_exit_kernel_prepare which did local_irq_disable too.
Please check where was PACA_IRQ_HARD_DIS set without GENERIC_ENTRY which was preventing
this from happening?
That's an excellent point. I think we are looking at this incorrectly,
If this is related to the nested soft replay interrupt than this whole
dynamics changes, local_irq_disable will work for the actual interrupt
but when we go nested something is telling that the nested interrupt
that interrupts are enabled. Let me look at this more deeply.
Regards,
Mukesh
On Fri, Aug 14, 2026 at 11:42:22AM +0530, Mukesh Kumar Chaurasiya wrote:
On Fri, Aug 14, 2026 at 11:06:43AM +0530, Shrikanth Hegde wrote:
[...]
quoted
This is good explanation, but what i not understanding is,
why pattern of setting PACA_IRQ_HARD_DIS changed?
Previous code at interrupt_exit_kernel_prepare which did local_irq_disable too.
Please check where was PACA_IRQ_HARD_DIS set without GENERIC_ENTRY which was preventing
this from happening?
That's an excellent point. I think we are looking at this incorrectly,
If this is related to the nested soft replay interrupt than this whole
dynamics changes, local_irq_disable will work for the actual interrupt
but when we go nested something is telling that the nested interrupt
that interrupts are enabled. Let me look at this more deeply.
Regards,
Mukesh
On 15/08/26 10:27 am, Mukesh Kumar Chaurasiya wrote:
quoted hunk
On Fri, Aug 14, 2026 at 11:42:22AM +0530, Mukesh Kumar Chaurasiya wrote:
quoted
On Fri, Aug 14, 2026 at 11:06:43AM +0530, Shrikanth Hegde wrote:
[...]
quoted
This is good explanation, but what i not understanding is,
why pattern of setting PACA_IRQ_HARD_DIS changed?
Previous code at interrupt_exit_kernel_prepare which did local_irq_disable too.
Please check where was PACA_IRQ_HARD_DIS set without GENERIC_ENTRY which was preventing
this from happening?
That's an excellent point. I think we are looking at this incorrectly,
If this is related to the nested soft replay interrupt than this whole
dynamics changes, local_irq_disable will work for the actual interrupt
but when we go nested something is telling that the nested interrupt
that interrupts are enabled. Let me look at this more deeply.
Regards,
Mukesh
Hey Venkat,
Thanks for verifying.
The regs.softe = IRQS_DISABLED change only prevents
irqentry_exit_cond_resched() preemption path from running — it does nothing
about PACA_IRQ_HARD_DIS being cleared by the handler and not restored.
hard_irq_disable() fixes both: it atomically sets PACA_IRQ_HARD_DIS, sets
irq_soft_mask = IRQS_ALL_DISABLED, and clears MSR[EE], leaving a fully
consistent disabled state that next_interrupt() expects, regardless of
which entry path (idle wakeup or irqrestore) triggered the replay.
To answer's shrikanth question
it worked before because there was literally no interrupt-disabling code
on the handler exit path during replay. The new GENERIC_ENTRY
infrastructure brought a real exit sequence with it, and that sequence
needs hard_irq_disable() — not local_irq_disable() — to keep PPC64's
two-level interrupt state (irq_soft_mask + PACA_IRQ_HARD_DIS) consistent.
Hey Madhavan,
I think we are ok with this patch if there are no more objections from
anyone.
Regards,
Mukesh
Hey Venkat,
Thanks for verifying.
The regs.softe = IRQS_DISABLED change only prevents
irqentry_exit_cond_resched() preemption path from running — it does nothing
about PACA_IRQ_HARD_DIS being cleared by the handler and not restored.
hard_irq_disable() fixes both: it atomically sets PACA_IRQ_HARD_DIS, sets
irq_soft_mask = IRQS_ALL_DISABLED, and clears MSR[EE], leaving a fully
consistent disabled state that next_interrupt() expects, regardless of
which entry path (idle wakeup or irqrestore) triggered the replay.
To answer's shrikanth question
it worked before because there was literally no interrupt-disabling code
on the handler exit path during replay. The new GENERIC_ENTRY
infrastructure brought a real exit sequence with it, and that sequence
needs hard_irq_disable() — not local_irq_disable() — to keep PPC64's
two-level interrupt state (irq_soft_mask + PACA_IRQ_HARD_DIS) consistent.
Hey Madhavan,
I think we are ok with this patch if there are no more objections from
anyone.
Venkat,
Can you collect vmcore with panic on warning enabled?
It maybe nested replay of soft interrupts which maybe causing problems
here. We should see the regs state.
On Wed, Aug 19, 2026 at 04:22:31PM +0530, Shrikanth Hegde wrote:
Can you collect vmcore with panic on warning enabled?
It maybe nested replay of soft interrupts which maybe causing problems here.
We should see the regs state.
quoted
Regards,
Mukesh
Hey Venkat,
I have identified the root cause for this and discussed it with
Shrikanth also. We have agreed that the fix we have now is a proper fix
and why we got this issue now and not earlier. I'll send out a new
revision for this.
Regards,
Mukesh