[PATCH] smp: Suppress false DEBUG_PREEMPT warning in smp_call_on_cpu()

[PATCH] smp: Suppress false DEBUG_PREEMPT warning in smp_call_on_cpu()

[Nam Cao]

While booting UP (uniprocessor) kernel with CONFIG_DEBUG_PREEMPT=y, the
following warning is observed:

BUG: using smp_processor_id() in preemptible [00000000] code: swapper/0/1
caller is debug_smp_processor_id+0x1c/0x28
CPU: 0 UID: 0 PID: 1 Comm: swapper/0 Not tainted 6.18.0-rc2-00012-g002733e953a7 #111 PREEMPT_RT
Hardware name: MangoPi MQ Pro (DT)
Call Trace:
    dump_backtrace
    show_stack
    dump_stack_lvl
    dump_stack
    check_preemption_disabled.isra.0
    debug_smp_processor_id
    check_unaligned_access
    smp_call_on_cpu
    check_unaligned_access_all_cpus
    do_one_initcall
    kernel_init_freeable
    kernel_init
    ret_from_fork_kernel
    ret_from_fork_kernel_asm

This is a false warning. The UP-variant of smp_call_on_cpu() simply calls
the callback, and thus debug_smp_processor_id() thinks the context is
unsafe for smp_processor_id(), which is obviously false because this is UP
kernel.

This appears after commit 06ddd17521bf ("sched/smp: Always define
is_percpu_thread() and scheduler_ipi()"). Before this commit,
is_percpu_thread() always returns true on UP kernel and thus
debug_smp_processor_id() always sees a per-cpu thread and never warns. But
now is_percpu_thread() returns false for this case.

Suppress this warning with a migrate_disable()+migrate_enable() pair.

Fixes: 06ddd17521bf ("sched/smp: Always define is_percpu_thread() and scheduler_ipi()")
Signed-off-by: Nam Cao <namcao@linutronix.de>
---
grep shows me that the followings also have the same false warning:

    arch/arm64/kernel/watchdog_hld.c
    drivers/hwmon/dell-smm-hwmon.c
    drivers/platform/x86/dell/dcdbas.c
    kernel/watchdog.c
---
 kernel/up.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/kernel/up.c b/kernel/up.c
index df50828cc2f0..37c9f0d39b36 100644
--- a/kernel/up.c
+++ b/kernel/up.c
@@ -64,7 +64,14 @@ int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys)
 
 	if (phys)
 		hypervisor_pin_vcpu(0);
+
+	/* suppress warnings from debug_smp_processor_id() */
+	migrate_disable();
+
 	ret = func(par);
+
+	migrate_enable();
+
 	if (phys)
 		hypervisor_pin_vcpu(-1);
 
-- 
2.39.5

Re: [PATCH] smp: Suppress false DEBUG_PREEMPT warning in smp_call_on_cpu()

[Sebastian Andrzej Siewior]

On 2025-10-24 09:07:14 [+0200], Nam Cao wrote:
> While booting UP (uniprocessor) kernel with CONFIG_DEBUG_PREEMPT=y, the
> following warning is observed:
> 
> BUG: using smp_processor_id() in preemptible [00000000] code: swapper/0/1
> caller is debug_smp_processor_id+0x1c/0x28
> CPU: 0 UID: 0 PID: 1 Comm: swapper/0 Not tainted 6.18.0-rc2-00012-g002733e953a7 #111 PREEMPT_RT
> Hardware name: MangoPi MQ Pro (DT)
> Call Trace:
>     dump_backtrace
>     show_stack
>     dump_stack_lvl
>     dump_stack
>     check_preemption_disabled.isra.0
>     debug_smp_processor_id
>     check_unaligned_access
>     smp_call_on_cpu
>     check_unaligned_access_all_cpus
>     do_one_initcall
>     kernel_init_freeable
>     kernel_init
>     ret_from_fork_kernel
>     ret_from_fork_kernel_asm

Can be compressed to:

   Using smp_processor_id() in a smp_call_on_cpu() callback triggers a
   "using ... in preemptible" warning in UP builds.

> This is a false warning. The UP-variant of smp_call_on_cpu() simply calls
> the callback, and thus debug_smp_processor_id() thinks the context is
> unsafe for smp_processor_id(), which is obviously false because this is UP
> kernel.

"Only the UP implementation is affected by this"

> This appears after commit 06ddd17521bf ("sched/smp: Always define
> is_percpu_thread() and scheduler_ipi()"). Before this commit,
> is_percpu_thread() always returns true on UP kernel and thus
> debug_smp_processor_id() always sees a per-cpu thread and never warns. But
> now is_percpu_thread() returns false for this case.
> 
> Suppress this warning with a migrate_disable()+migrate_enable() pair.

Right. This is one of the possibilities. The other one would be to also
workqueue on UP and preserve the same semantic.
I don't mind this.

> Fixes: 06ddd17521bf ("sched/smp: Always define is_percpu_thread() and scheduler_ipi()")
> Signed-off-by: Nam Cao <namcao@linutronix.de>
> ---
> grep shows me that the followings also have the same false warning:
> 
>     arch/arm64/kernel/watchdog_hld.c
>     drivers/hwmon/dell-smm-hwmon.c
>     drivers/platform/x86/dell/dcdbas.c
>     kernel/watchdog.c
> ---
>  kernel/up.c | 7 +++++++
>  1 file changed, 7 insertions(+)
> 
> diff --git a/kernel/up.c b/kernel/up.c
> index df50828cc2f0..37c9f0d39b36 100644
> --- a/kernel/up.c
> +++ b/kernel/up.c
> @@ -64,7 +64,14 @@ int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys)
>  
>  	if (phys)
>  		hypervisor_pin_vcpu(0);
> +
> +	/* suppress warnings from debug_smp_processor_id() */

If you want to add a comment, what about something like

  /* Preserve not being migratable such as SMP variant does */ 

> +	migrate_disable();
> +
>  	ret = func(par);
> +
> +	migrate_enable();
> +
>  	if (phys)
>  		hypervisor_pin_vcpu(-1);
>  

Sebastian

Re: [PATCH] smp: Suppress false DEBUG_PREEMPT warning in smp_call_on_cpu()

[Nam Cao]

Sebastian Andrzej Siewior <bigeasy@linutronix.de> writes:
>> This appears after commit 06ddd17521bf ("sched/smp: Always define
>> is_percpu_thread() and scheduler_ipi()"). Before this commit,
>> is_percpu_thread() always returns true on UP kernel and thus
>> debug_smp_processor_id() always sees a per-cpu thread and never warns. But
>> now is_percpu_thread() returns false for this case.
>> 
>> Suppress this warning with a migrate_disable()+migrate_enable() pair.
>
> Right. This is one of the possibilities. The other one would be to also
> workqueue on UP and preserve the same semantic.
> I don't mind this.

Yeah it's a tradeoff. I'm not sure if someone will be bothered by the
overhead that workqueue introduces, so I use this cheap solution.

>> +
>> +	/* suppress warnings from debug_smp_processor_id() */
>
> If you want to add a comment, what about something like
>
>   /* Preserve not being migratable such as SMP variant does */ 

Wouldn't it be a bit misleading? It's true that technically SMP variant
is not migratable, but that's because it is a per-cpu thread, not
because of migrate_disable(). To me, the comment sounds like that there
is also a migrate_disable() in SMP variant.

Nam

Re: [PATCH] smp: Suppress false DEBUG_PREEMPT warning in smp_call_on_cpu()

[Sebastian Andrzej Siewior]

On 2025-11-05 09:42:42 [+0100], Nam Cao wrote:
> >> +
> >> +	/* suppress warnings from debug_smp_processor_id() */
> >
> > If you want to add a comment, what about something like
> >
> >   /* Preserve not being migratable such as SMP variant does */ 
> 
> Wouldn't it be a bit misleading? It's true that technically SMP variant
> is not migratable, but that's because it is a per-cpu thread, not
> because of migrate_disable(). To me, the comment sounds like that there
> is also a migrate_disable() in SMP variant.

A per-cpu thread ensures that the control flow can not migrate to
another CPU. This is also what stops debug_smp_processor_id() warnings
from work invoked from queue_work_on() vs queue_work().

> Nam

Sebastian