arm64: topology: Fix false warning in counters_read_on_cpu() for same-CPU reads

[PATCH] arm64: topology: Fix false warning in counters_read_on_cpu() for same-CPU reads

Posted by Sumit Gupta 1 week, 3 days ago

The counters_read_on_cpu() function warns when called with IRQs disabled
to prevent deadlock in smp_call_function_single(). However, this warning
is spurious when reading counters on the current CPU since no IPI is
needed for same-CPU reads.

Commit 12eb8f4fff24 ("cpufreq: CPPC: Update FIE arch_freq_scale in ticks
for non-PCC regs") changed the CPPC Frequency Invariance Engine to read
AMU counters directly from the scheduler tick for non-PCC register
spaces (like FFH), instead of deferring to a kthread. This means
counters_read_on_cpu() is now called with IRQs disabled from the tick
handler, triggering the warning:

| WARNING: arch/arm64/kernel/topology.c:410 at counters_read_on_cpu
| ...
| Call trace:
|  counters_read_on_cpu+0x88/0xa8 (P)
|  cpc_read_ffh+0xdc/0x148
|  cpc_read+0x260/0x518
|  cppc_get_perf_ctrs+0xf0/0x398
|  __cppc_scale_freq_tick+0x4c/0x148 [cppc_cpufreq]
|  cppc_scale_freq_tick+0x44/0x88 [cppc_cpufreq]
|  topology_scale_freq_tick+0x34/0x58
|  sched_tick+0x58/0x300
|  update_process_times+0xcc/0x120
|  tick_nohz_handler+0xa8/0x260
|  __hrtimer_run_queues+0x154/0x360
|  hrtimer_interrupt+0xf4/0x2b0
|  arch_timer_handler_phys+0x4c/0x78
|  ....
|  CPPC Cpufreq:__cppc_scale_freq_tick: failed to read perf counters
|  ....

Fix this by calling the counter read function directly for same-CPU
case, bypassing smp_call_function_single() entirely. Use get_cpu() to
disable preemption as the counter read functions call this_cpu_has_cap()
which requires a non-preemptible context.

Fixes: 12eb8f4fff24 ("cpufreq: CPPC: Update FIE arch_freq_scale in ticks for non-PCC regs")
Signed-off-by: Sumit Gupta <sumitg@nvidia.com>
---
 arch/arm64/kernel/topology.c | 21 +++++++++++++++++++--
 1 file changed, 19 insertions(+), 2 deletions(-)

diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c
index 539b38935182..57b71f403007 100644
--- a/arch/arm64/kernel/topology.c
+++ b/arch/arm64/kernel/topology.c
@@ -401,12 +401,29 @@ static inline
 int counters_read_on_cpu(int cpu, smp_call_func_t func, u64 *val)
 {
 	/*
-	 * Abort call on counterless CPU or when interrupts are
-	 * disabled - can lead to deadlock in smp sync call.
+	 * Abort call on counterless CPU.
 	 */
 	if (!cpu_has_amu_feat(cpu))
 		return -EOPNOTSUPP;
 
+	/*
+	 * For same-CPU reads, call the function directly since no IPI
+	 * is needed and this is safe even with IRQs disabled.
+	 * Use get_cpu() to disable preemption as the counter read
+	 * functions call this_cpu_has_cap() which requires a
+	 * non-preemptible context.
+	 */
+	if (cpu == get_cpu()) {
+		func(val);
+		put_cpu();
+		return 0;
+	}
+	put_cpu();
+
+	/*
+	 * Reading from a remote CPU requires IRQs enabled to avoid
+	 * deadlock in smp_call_function_single().
+	 */
 	if (WARN_ON_ONCE(irqs_disabled()))
 		return -EPERM;
 
-- 
2.34.1

Re: [PATCH] arm64: topology: Fix false warning in counters_read_on_cpu() for same-CPU reads

Posted by Jie Zhan 1 week, 2 days ago

Hi Sumit,

On 1/27/2026 4:07 PM, Sumit Gupta wrote:
> The counters_read_on_cpu() function warns when called with IRQs disabled
> to prevent deadlock in smp_call_function_single(). However, this warning
> is spurious when reading counters on the current CPU since no IPI is
> needed for same-CPU reads.
> 
> Commit 12eb8f4fff24 ("cpufreq: CPPC: Update FIE arch_freq_scale in ticks
> for non-PCC regs") changed the CPPC Frequency Invariance Engine to read
> AMU counters directly from the scheduler tick for non-PCC register
> spaces (like FFH), instead of deferring to a kthread. This means
> counters_read_on_cpu() is now called with IRQs disabled from the tick
> handler, triggering the warning:
> 
> | WARNING: arch/arm64/kernel/topology.c:410 at counters_read_on_cpu
> | ...
> | Call trace:
> |  counters_read_on_cpu+0x88/0xa8 (P)
> |  cpc_read_ffh+0xdc/0x148
> |  cpc_read+0x260/0x518
> |  cppc_get_perf_ctrs+0xf0/0x398
> |  __cppc_scale_freq_tick+0x4c/0x148 [cppc_cpufreq]
> |  cppc_scale_freq_tick+0x44/0x88 [cppc_cpufreq]
> |  topology_scale_freq_tick+0x34/0x58
> |  sched_tick+0x58/0x300
> |  update_process_times+0xcc/0x120
> |  tick_nohz_handler+0xa8/0x260
> |  __hrtimer_run_queues+0x154/0x360
> |  hrtimer_interrupt+0xf4/0x2b0
> |  arch_timer_handler_phys+0x4c/0x78
> |  ....
> |  CPPC Cpufreq:__cppc_scale_freq_tick: failed to read perf counters
> |  ....
> 
> Fix this by calling the counter read function directly for same-CPU
> case, bypassing smp_call_function_single() entirely. Use get_cpu() to
> disable preemption as the counter read functions call this_cpu_has_cap()
> which requires a non-preemptible context.
> 
> Fixes: 12eb8f4fff24 ("cpufreq: CPPC: Update FIE arch_freq_scale in ticks for non-PCC regs")
> Signed-off-by: Sumit Gupta <sumitg@nvidia.com>

Reviewed-by: Jie Zhan <zhanjie9@hisilicon.com>

Looks fine for me except for the minor comment wrapping.

Thanks for spotting this.
I may have missed the warning log in the FFH test.

This happens during the short window in cpufreq_policy_online() between
driver->init() and the CREATE_POLICY notifier that gets AMU FIE ready.
After that, CPPC FIE will be stopped.

Ideally this can be merged together with Viresh's PR since the CPPC FIE
changes are there.
https://lore.kernel.org/all/j4qdid7iqmng4gzb5ozefemjkep3wx2b5z2yki5tnqc3vzvzf4@kvrnarvdod5p/

Jie

> ---
>  arch/arm64/kernel/topology.c | 21 +++++++++++++++++++--
>  1 file changed, 19 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c
> index 539b38935182..57b71f403007 100644
> --- a/arch/arm64/kernel/topology.c
> +++ b/arch/arm64/kernel/topology.c
> @@ -401,12 +401,29 @@ static inline
>  int counters_read_on_cpu(int cpu, smp_call_func_t func, u64 *val)
>  {
>  	/*
> -	 * Abort call on counterless CPU or when interrupts are
> -	 * disabled - can lead to deadlock in smp sync call.
> +	 * Abort call on counterless CPU.
>  	 */
>  	if (!cpu_has_amu_feat(cpu))
>  		return -EOPNOTSUPP;
>  
> +	/*
> +	 * For same-CPU reads, call the function directly since no IPI
> +	 * is needed and this is safe even with IRQs disabled.
> +	 * Use get_cpu() to disable preemption as the counter read
> +	 * functions call this_cpu_has_cap() which requires a
> +	 * non-preemptible context.
> +	 */
Wrap at 80 chars?
> +	if (cpu == get_cpu()) {
> +		func(val);
> +		put_cpu();
> +		return 0;
> +	}
> +	put_cpu();
> +
> +	/*
> +	 * Reading from a remote CPU requires IRQs enabled to avoid
> +	 * deadlock in smp_call_function_single().
> +	 */
>  	if (WARN_ON_ONCE(irqs_disabled()))
>  		return -EPERM;
>

Re: [PATCH] arm64: topology: Fix false warning in counters_read_on_cpu() for same-CPU reads

Posted by Will Deacon 1 week, 2 days ago

On Wed, Jan 28, 2026 at 06:50:42PM +0800, Jie Zhan wrote:
> On 1/27/2026 4:07 PM, Sumit Gupta wrote:
> > The counters_read_on_cpu() function warns when called with IRQs disabled
> > to prevent deadlock in smp_call_function_single(). However, this warning
> > is spurious when reading counters on the current CPU since no IPI is
> > needed for same-CPU reads.
> > 
> > Commit 12eb8f4fff24 ("cpufreq: CPPC: Update FIE arch_freq_scale in ticks
> > for non-PCC regs") changed the CPPC Frequency Invariance Engine to read
> > AMU counters directly from the scheduler tick for non-PCC register
> > spaces (like FFH), instead of deferring to a kthread. This means
> > counters_read_on_cpu() is now called with IRQs disabled from the tick
> > handler, triggering the warning:
> > 
> > | WARNING: arch/arm64/kernel/topology.c:410 at counters_read_on_cpu
> > | ...
> > | Call trace:
> > |  counters_read_on_cpu+0x88/0xa8 (P)
> > |  cpc_read_ffh+0xdc/0x148
> > |  cpc_read+0x260/0x518
> > |  cppc_get_perf_ctrs+0xf0/0x398
> > |  __cppc_scale_freq_tick+0x4c/0x148 [cppc_cpufreq]
> > |  cppc_scale_freq_tick+0x44/0x88 [cppc_cpufreq]
> > |  topology_scale_freq_tick+0x34/0x58
> > |  sched_tick+0x58/0x300
> > |  update_process_times+0xcc/0x120
> > |  tick_nohz_handler+0xa8/0x260
> > |  __hrtimer_run_queues+0x154/0x360
> > |  hrtimer_interrupt+0xf4/0x2b0
> > |  arch_timer_handler_phys+0x4c/0x78
> > |  ....
> > |  CPPC Cpufreq:__cppc_scale_freq_tick: failed to read perf counters
> > |  ....
> > 
> > Fix this by calling the counter read function directly for same-CPU
> > case, bypassing smp_call_function_single() entirely. Use get_cpu() to
> > disable preemption as the counter read functions call this_cpu_has_cap()
> > which requires a non-preemptible context.
> > 
> > Fixes: 12eb8f4fff24 ("cpufreq: CPPC: Update FIE arch_freq_scale in ticks for non-PCC regs")
> > Signed-off-by: Sumit Gupta <sumitg@nvidia.com>
> 
> Reviewed-by: Jie Zhan <zhanjie9@hisilicon.com>
> 
> Looks fine for me except for the minor comment wrapping.
> 
> Thanks for spotting this.
> I may have missed the warning log in the FFH test.
> 
> This happens during the short window in cpufreq_policy_online() between
> driver->init() and the CREATE_POLICY notifier that gets AMU FIE ready.
> After that, CPPC FIE will be stopped.
> 
> Ideally this can be merged together with Viresh's PR since the CPPC FIE
> changes are there.
> https://lore.kernel.org/all/j4qdid7iqmng4gzb5ozefemjkep3wx2b5z2yki5tnqc3vzvzf4@kvrnarvdod5p/

Right, looks like this should go via Rafael but if it doesn't make the merge
window then I can pick it up at -rc1 (please remind me :)

Will