[RFC PATCH 0/2] cpufreq/amd-pstate: Prevent scheduling when atomic on PREEMPT_RT

[RFC PATCH 0/2] cpufreq/amd-pstate: Prevent scheduling when atomic on PREEMPT_RT

[K Prateek Nayak]

Bert reported hitting "BUG: scheduling while atomic" when running
amd-pstate-ut on a PREEMPT_RT kernel [1].

Since reader-writer locks turn sleepable on PREEMPT_RT, they are not
suitable to be used in the scheduler hot-path under rq_lock to grab the
cpufreq policy object.

Unfortunately, the amd-pstate driver has a tight coupling between the
cpufreq_policy object and the cpudata stored in it as the driver_data.

Trying to grab a read reference on PREEMPT_RT can cause "scheduling
while atomic" if a concurrent writer is active, and trying to grab a
nested reference in presence of a writer can cause a deadlock (manifests
as lockup) since the reader fast-path is disabled on PREEMPT_RT to
prevent write-side starvation.

The two patches included removes cases of grabbing a nested read
reference to the cpufreq policy in amd-pstate, and modifies the
cpufreq_driver->adjust_perf() callback to take the raw policy reference
cached by the schedutil governor respectively.

The policy object outlives the governor and the driver making it safe to
use this cache reference from the sugov data. Any changes to the policy
will end up calling cpufreq_driver->set_policy() or
governor->set_limits() once the policy is modified which should ensure
eventual consistency despite not holding the read-side.

Series has been tested with amd-pstate-ut on PREEMPT_RT kernel which
successfully passes without any splats on LOCKDEP + DEBUG_ATOMIC_SLEEP
config. Additionally, the driver switch test from Gautham [2] was run
for 10min on the same config without observing any splats.

[1] https://lore.kernel.org/all/20250731092316.3191-1-spasswolf@web.de/
[2] https://lore.kernel.org/all/aJRN2wMLAnhDFykv@BLRRASHENOY1.amd.com/

Patches are based on:

  git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm.git bleeding-edge

at commit 8aa9e2868a8e ("Merge branch 'pm-runtime-cleanup' into
bleeding-edge").
---
K Prateek Nayak (2):
  cpufreq/amd-pstate: Pass the policy to amd_pstate_update()
  cpufreq: Pass the policy to cpufreq_driver->adjust_perf()

 drivers/cpufreq/amd-pstate.c     | 14 +++++---------
 drivers/cpufreq/cpufreq.c        |  4 ++--
 drivers/cpufreq/intel_pstate.c   |  4 ++--
 include/linux/cpufreq.h          |  4 ++--
 kernel/sched/cpufreq_schedutil.c |  5 +++--
 5 files changed, 14 insertions(+), 17 deletions(-)


base-commit: 8aa9e2868a8e38ccc5228399fc641d54aea444ed
-- 
2.34.1

[RFC PATCH 1/2] cpufreq/amd-pstate: Pass the policy to amd_pstate_update()

[K Prateek Nayak]

All callers of amd_pstate_update() already have a reference to the
cpufreq_policy object.

Pass the entire policy object and grab the cpudata using
"policy->driver_data" instead of passing the cpudata and unnecessarily
grabbing another read-side reference to the cpufreq policy object when
it is already available in the caller.

No functional changes intended.

Signed-off-by: K Prateek Nayak <kprateek.nayak@amd.com>
---
 drivers/cpufreq/amd-pstate.c | 11 ++++-------
 1 file changed, 4 insertions(+), 7 deletions(-)

diff --git a/drivers/cpufreq/amd-pstate.c b/drivers/cpufreq/amd-pstate.c
index c45bc98721d2..5818a92d96b9 100644
--- a/drivers/cpufreq/amd-pstate.c
+++ b/drivers/cpufreq/amd-pstate.c
@@ -565,15 +565,12 @@ static inline bool amd_pstate_sample(struct amd_cpudata *cpudata)
 	return true;
 }
 
-static void amd_pstate_update(struct amd_cpudata *cpudata, u8 min_perf,
+static void amd_pstate_update(struct cpufreq_policy *policy, u8 min_perf,
 			      u8 des_perf, u8 max_perf, bool fast_switch, int gov_flags)
 {
-	struct cpufreq_policy *policy __free(put_cpufreq_policy) = cpufreq_cpu_get(cpudata->cpu);
+	struct amd_cpudata *cpudata = policy->driver_data;
 	union perf_cached perf = READ_ONCE(cpudata->perf);
 
-	if (!policy)
-		return;
-
 	/* limit the max perf when core performance boost feature is disabled */
 	if (!cpudata->boost_supported)
 		max_perf = min_t(u8, perf.nominal_perf, max_perf);
@@ -675,7 +672,7 @@ static int amd_pstate_update_freq(struct cpufreq_policy *policy,
 	if (!fast_switch)
 		cpufreq_freq_transition_begin(policy, &freqs);
 
-	amd_pstate_update(cpudata, perf.min_limit_perf, des_perf,
+	amd_pstate_update(policy, perf.min_limit_perf, des_perf,
 			  perf.max_limit_perf, fast_switch,
 			  policy->governor->flags);
 
@@ -737,7 +734,7 @@ static void amd_pstate_adjust_perf(unsigned int cpu,
 	if (max_perf < min_perf)
 		max_perf = min_perf;
 
-	amd_pstate_update(cpudata, min_perf, des_perf, max_perf, true,
+	amd_pstate_update(policy, min_perf, des_perf, max_perf, true,
 			policy->governor->flags);
 }
 
-- 
2.34.1

[RFC PATCH 2/2] cpufreq: Pass the policy to cpufreq_driver->adjust_perf()

[K Prateek Nayak]

cpufreq_cpu_get() can sleep on PREEMPT_RT in presence of concurrent
writer(s), however amd-pstate depends on fetching the cpudata via the
policy's driver data which necessitates grabbing the reference.

Since schedutil governor can call "cpufreq_driver->update_perf()"
during sched_tick/enqueue/dequeue with rq_lock held and IRQs disabled,
fetching the policy object using the cpufreq_cpu_get() helper in the
scheduler fast-path leads to "BUG: scheduling while atomic" on
PREEMPT_RT [1].

Pass the cached cpufreq policy object in sg_policy to the update_perf()
instead of just the CPU. The CPU can be inferred using "policy->cpu".

The lifetime of cpufreq_policy object outlasts that of the governor and
the cpufreq driver (allocated when the CPU is onlined and only reclaimed
when the CPU is offlined / the CPU device is removed) which makes it
safe to be referenced throughout the governor's lifetime.

Link: https://lore.kernel.org/all/20250731092316.3191-1-spasswolf@web.de/ [1]
Signed-off-by: K Prateek Nayak <kprateek.nayak@amd.com>
---
 drivers/cpufreq/amd-pstate.c     | 3 +--
 drivers/cpufreq/cpufreq.c        | 4 ++--
 drivers/cpufreq/intel_pstate.c   | 4 ++--
 include/linux/cpufreq.h          | 4 ++--
 kernel/sched/cpufreq_schedutil.c | 5 +++--
 5 files changed, 10 insertions(+), 10 deletions(-)

diff --git a/drivers/cpufreq/amd-pstate.c b/drivers/cpufreq/amd-pstate.c
index 5818a92d96b9..455e58a9b738 100644
--- a/drivers/cpufreq/amd-pstate.c
+++ b/drivers/cpufreq/amd-pstate.c
@@ -697,13 +697,12 @@ static unsigned int amd_pstate_fast_switch(struct cpufreq_policy *policy,
 	return policy->cur;
 }
 
-static void amd_pstate_adjust_perf(unsigned int cpu,
+static void amd_pstate_adjust_perf(struct cpufreq_policy *policy,
 				   unsigned long _min_perf,
 				   unsigned long target_perf,
 				   unsigned long capacity)
 {
 	u8 max_perf, min_perf, des_perf, cap_perf;
-	struct cpufreq_policy *policy __free(put_cpufreq_policy) = cpufreq_cpu_get(cpu);
 	struct amd_cpudata *cpudata;
 	union perf_cached perf;
 
diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c
index 50dde2980f1b..8bdc8f9b8d86 100644
--- a/drivers/cpufreq/cpufreq.c
+++ b/drivers/cpufreq/cpufreq.c
@@ -2244,12 +2244,12 @@ EXPORT_SYMBOL_GPL(cpufreq_driver_fast_switch);
  * parallel with either ->target() or ->target_index() or ->fast_switch() for
  * the same CPU.
  */
-void cpufreq_driver_adjust_perf(unsigned int cpu,
+void cpufreq_driver_adjust_perf(struct cpufreq_policy *policy,
 				 unsigned long min_perf,
 				 unsigned long target_perf,
 				 unsigned long capacity)
 {
-	cpufreq_driver->adjust_perf(cpu, min_perf, target_perf, capacity);
+	cpufreq_driver->adjust_perf(policy, min_perf, target_perf, capacity);
 }
 
 /**
diff --git a/drivers/cpufreq/intel_pstate.c b/drivers/cpufreq/intel_pstate.c
index ec4abe374573..8d25f0f2925c 100644
--- a/drivers/cpufreq/intel_pstate.c
+++ b/drivers/cpufreq/intel_pstate.c
@@ -3237,12 +3237,12 @@ static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
 	return target_pstate * cpu->pstate.scaling;
 }
 
-static void intel_cpufreq_adjust_perf(unsigned int cpunum,
+static void intel_cpufreq_adjust_perf(struct cpufreq_policy *policy,
 				      unsigned long min_perf,
 				      unsigned long target_perf,
 				      unsigned long capacity)
 {
-	struct cpudata *cpu = all_cpu_data[cpunum];
+	struct cpudata *cpu = all_cpu_data[policy->cpu];
 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
 	int old_pstate = cpu->pstate.current_pstate;
 	int cap_pstate, min_pstate, max_pstate, target_pstate;
diff --git a/include/linux/cpufreq.h b/include/linux/cpufreq.h
index 0465d1e6f72a..fd26b3a4aa28 100644
--- a/include/linux/cpufreq.h
+++ b/include/linux/cpufreq.h
@@ -367,7 +367,7 @@ struct cpufreq_driver {
 	 * conditions) scale invariance can be disabled, which causes the
 	 * schedutil governor to fall back to the latter.
 	 */
-	void		(*adjust_perf)(unsigned int cpu,
+	void		(*adjust_perf)(struct cpufreq_policy *policy,
 				       unsigned long min_perf,
 				       unsigned long target_perf,
 				       unsigned long capacity);
@@ -612,7 +612,7 @@ struct cpufreq_governor {
 /* Pass a target to the cpufreq driver */
 unsigned int cpufreq_driver_fast_switch(struct cpufreq_policy *policy,
 					unsigned int target_freq);
-void cpufreq_driver_adjust_perf(unsigned int cpu,
+void cpufreq_driver_adjust_perf(struct cpufreq_policy *policy,
 				unsigned long min_perf,
 				unsigned long target_perf,
 				unsigned long capacity);
diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
index 0ab5f9d4bc59..307f3076635e 100644
--- a/kernel/sched/cpufreq_schedutil.c
+++ b/kernel/sched/cpufreq_schedutil.c
@@ -461,6 +461,7 @@ static void sugov_update_single_perf(struct update_util_data *hook, u64 time,
 				     unsigned int flags)
 {
 	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
+	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 	unsigned long prev_util = sg_cpu->util;
 	unsigned long max_cap;
 
@@ -482,10 +483,10 @@ static void sugov_update_single_perf(struct update_util_data *hook, u64 time,
 	if (sugov_hold_freq(sg_cpu) && sg_cpu->util < prev_util)
 		sg_cpu->util = prev_util;
 
-	cpufreq_driver_adjust_perf(sg_cpu->cpu, sg_cpu->bw_min,
+	cpufreq_driver_adjust_perf(sg_policy->policy, sg_cpu->bw_min,
 				   sg_cpu->util, max_cap);
 
-	sg_cpu->sg_policy->last_freq_update_time = time;
+	sg_policy->last_freq_update_time = time;
 }
 
 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
-- 
2.34.1

Re: [RFC PATCH 1/2] cpufreq/amd-pstate: Pass the policy to amd_pstate_update()

[Mario Limonciello (AMD) (kernel.org)]

On 1/6/2026 1:36 AM, K Prateek Nayak wrote:
> All callers of amd_pstate_update() already have a reference to the
> cpufreq_policy object.
> 
> Pass the entire policy object and grab the cpudata using
> "policy->driver_data" instead of passing the cpudata and unnecessarily
> grabbing another read-side reference to the cpufreq policy object when
> it is already available in the caller.
> 
> No functional changes intended.
> 
> Signed-off-by: K Prateek Nayak <kprateek.nayak@amd.com>
Reviewed-by: Mario Limonciello (AMD) <superm1@kernel.org>
> ---
>   drivers/cpufreq/amd-pstate.c | 11 ++++-------
>   1 file changed, 4 insertions(+), 7 deletions(-)
> 
> diff --git a/drivers/cpufreq/amd-pstate.c b/drivers/cpufreq/amd-pstate.c
> index c45bc98721d2..5818a92d96b9 100644
> --- a/drivers/cpufreq/amd-pstate.c
> +++ b/drivers/cpufreq/amd-pstate.c
> @@ -565,15 +565,12 @@ static inline bool amd_pstate_sample(struct amd_cpudata *cpudata)
>   	return true;
>   }
>   
> -static void amd_pstate_update(struct amd_cpudata *cpudata, u8 min_perf,
> +static void amd_pstate_update(struct cpufreq_policy *policy, u8 min_perf,
>   			      u8 des_perf, u8 max_perf, bool fast_switch, int gov_flags)
>   {
> -	struct cpufreq_policy *policy __free(put_cpufreq_policy) = cpufreq_cpu_get(cpudata->cpu);
> +	struct amd_cpudata *cpudata = policy->driver_data;
>   	union perf_cached perf = READ_ONCE(cpudata->perf);
>   
> -	if (!policy)
> -		return;
> -
>   	/* limit the max perf when core performance boost feature is disabled */
>   	if (!cpudata->boost_supported)
>   		max_perf = min_t(u8, perf.nominal_perf, max_perf);
> @@ -675,7 +672,7 @@ static int amd_pstate_update_freq(struct cpufreq_policy *policy,
>   	if (!fast_switch)
>   		cpufreq_freq_transition_begin(policy, &freqs);
>   
> -	amd_pstate_update(cpudata, perf.min_limit_perf, des_perf,
> +	amd_pstate_update(policy, perf.min_limit_perf, des_perf,
>   			  perf.max_limit_perf, fast_switch,
>   			  policy->governor->flags);
>   
> @@ -737,7 +734,7 @@ static void amd_pstate_adjust_perf(unsigned int cpu,
>   	if (max_perf < min_perf)
>   		max_perf = min_perf;
>   
> -	amd_pstate_update(cpudata, min_perf, des_perf, max_perf, true,
> +	amd_pstate_update(policy, min_perf, des_perf, max_perf, true,
>   			policy->governor->flags);
>   }
>   

Re: [RFC PATCH 2/2] cpufreq: Pass the policy to cpufreq_driver->adjust_perf()

[Mario Limonciello (AMD) (kernel.org)]

On 1/6/2026 1:36 AM, K Prateek Nayak wrote:
> cpufreq_cpu_get() can sleep on PREEMPT_RT in presence of concurrent
> writer(s), however amd-pstate depends on fetching the cpudata via the
> policy's driver data which necessitates grabbing the reference.
> 
> Since schedutil governor can call "cpufreq_driver->update_perf()"
> during sched_tick/enqueue/dequeue with rq_lock held and IRQs disabled,
> fetching the policy object using the cpufreq_cpu_get() helper in the
> scheduler fast-path leads to "BUG: scheduling while atomic" on
> PREEMPT_RT [1].
> 
> Pass the cached cpufreq policy object in sg_policy to the update_perf()
> instead of just the CPU. The CPU can be inferred using "policy->cpu".
> 
> The lifetime of cpufreq_policy object outlasts that of the governor and
> the cpufreq driver (allocated when the CPU is onlined and only reclaimed
> when the CPU is offlined / the CPU device is removed) which makes it
> safe to be referenced throughout the governor's lifetime.
> 
> Link: https://lore.kernel.org/all/20250731092316.3191-1-spasswolf@web.de/ [1]

I think you should have these tags instead:
Reported-by: Bert Karwatzki <spasswolf@web.de>
Closes:https://lore.kernel.org/all/20250731092316.3191-1-spasswolf@web.de/ 
[1]

> Signed-off-by: K Prateek Nayak <kprateek.nayak@amd.com>
> ---
>   drivers/cpufreq/amd-pstate.c     | 3 +--
>   drivers/cpufreq/cpufreq.c        | 4 ++--
>   drivers/cpufreq/intel_pstate.c   | 4 ++--
>   include/linux/cpufreq.h          | 4 ++--
>   kernel/sched/cpufreq_schedutil.c | 5 +++--
>   5 files changed, 10 insertions(+), 10 deletions(-)
> 
> diff --git a/drivers/cpufreq/amd-pstate.c b/drivers/cpufreq/amd-pstate.c
> index 5818a92d96b9..455e58a9b738 100644
> --- a/drivers/cpufreq/amd-pstate.c
> +++ b/drivers/cpufreq/amd-pstate.c
> @@ -697,13 +697,12 @@ static unsigned int amd_pstate_fast_switch(struct cpufreq_policy *policy,
>   	return policy->cur;
>   }
>   
> -static void amd_pstate_adjust_perf(unsigned int cpu,
> +static void amd_pstate_adjust_perf(struct cpufreq_policy *policy,
>   				   unsigned long _min_perf,
>   				   unsigned long target_perf,
>   				   unsigned long capacity)
>   {
>   	u8 max_perf, min_perf, des_perf, cap_perf;
> -	struct cpufreq_policy *policy __free(put_cpufreq_policy) = cpufreq_cpu_get(cpu);
>   	struct amd_cpudata *cpudata;
>   	union perf_cached perf;
>   
> diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c
> index 50dde2980f1b..8bdc8f9b8d86 100644
> --- a/drivers/cpufreq/cpufreq.c
> +++ b/drivers/cpufreq/cpufreq.c
> @@ -2244,12 +2244,12 @@ EXPORT_SYMBOL_GPL(cpufreq_driver_fast_switch);
>    * parallel with either ->target() or ->target_index() or ->fast_switch() for
>    * the same CPU.
>    */
> -void cpufreq_driver_adjust_perf(unsigned int cpu,
> +void cpufreq_driver_adjust_perf(struct cpufreq_policy *policy,
>   				 unsigned long min_perf,
>   				 unsigned long target_perf,
>   				 unsigned long capacity)
>   {
> -	cpufreq_driver->adjust_perf(cpu, min_perf, target_perf, capacity);
> +	cpufreq_driver->adjust_perf(policy, min_perf, target_perf, capacity);
>   }
>   
>   /**
> diff --git a/drivers/cpufreq/intel_pstate.c b/drivers/cpufreq/intel_pstate.c
> index ec4abe374573..8d25f0f2925c 100644
> --- a/drivers/cpufreq/intel_pstate.c
> +++ b/drivers/cpufreq/intel_pstate.c
> @@ -3237,12 +3237,12 @@ static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
>   	return target_pstate * cpu->pstate.scaling;
>   }
>   
> -static void intel_cpufreq_adjust_perf(unsigned int cpunum,
> +static void intel_cpufreq_adjust_perf(struct cpufreq_policy *policy,
>   				      unsigned long min_perf,
>   				      unsigned long target_perf,
>   				      unsigned long capacity)
>   {
> -	struct cpudata *cpu = all_cpu_data[cpunum];
> +	struct cpudata *cpu = all_cpu_data[policy->cpu];
>   	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
>   	int old_pstate = cpu->pstate.current_pstate;
>   	int cap_pstate, min_pstate, max_pstate, target_pstate;
> diff --git a/include/linux/cpufreq.h b/include/linux/cpufreq.h
> index 0465d1e6f72a..fd26b3a4aa28 100644
> --- a/include/linux/cpufreq.h
> +++ b/include/linux/cpufreq.h
> @@ -367,7 +367,7 @@ struct cpufreq_driver {
>   	 * conditions) scale invariance can be disabled, which causes the
>   	 * schedutil governor to fall back to the latter.
>   	 */
> -	void		(*adjust_perf)(unsigned int cpu,
> +	void		(*adjust_perf)(struct cpufreq_policy *policy,
>   				       unsigned long min_perf,
>   				       unsigned long target_perf,
>   				       unsigned long capacity);
> @@ -612,7 +612,7 @@ struct cpufreq_governor {
>   /* Pass a target to the cpufreq driver */
>   unsigned int cpufreq_driver_fast_switch(struct cpufreq_policy *policy,
>   					unsigned int target_freq);
> -void cpufreq_driver_adjust_perf(unsigned int cpu,
> +void cpufreq_driver_adjust_perf(struct cpufreq_policy *policy,
>   				unsigned long min_perf,
>   				unsigned long target_perf,
>   				unsigned long capacity);
> diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
> index 0ab5f9d4bc59..307f3076635e 100644
> --- a/kernel/sched/cpufreq_schedutil.c
> +++ b/kernel/sched/cpufreq_schedutil.c
> @@ -461,6 +461,7 @@ static void sugov_update_single_perf(struct update_util_data *hook, u64 time,
>   				     unsigned int flags)
>   {
>   	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
> +	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
>   	unsigned long prev_util = sg_cpu->util;
>   	unsigned long max_cap;
>   
> @@ -482,10 +483,10 @@ static void sugov_update_single_perf(struct update_util_data *hook, u64 time,
>   	if (sugov_hold_freq(sg_cpu) && sg_cpu->util < prev_util)
>   		sg_cpu->util = prev_util;
>   
> -	cpufreq_driver_adjust_perf(sg_cpu->cpu, sg_cpu->bw_min,
> +	cpufreq_driver_adjust_perf(sg_policy->policy, sg_cpu->bw_min,
>   				   sg_cpu->util, max_cap);
>   
> -	sg_cpu->sg_policy->last_freq_update_time = time;
> +	sg_policy->last_freq_update_time = time;
>   }
>   
>   static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)

Re: [RFC PATCH 2/2] cpufreq: Pass the policy to cpufreq_driver->adjust_perf()

[K Prateek Nayak]

Hello Mario,

On 1/7/2026 1:01 AM, Mario Limonciello (AMD) (kernel.org) wrote:
> 
> 
> On 1/6/2026 1:36 AM, K Prateek Nayak wrote:
>> cpufreq_cpu_get() can sleep on PREEMPT_RT in presence of concurrent
>> writer(s), however amd-pstate depends on fetching the cpudata via the
>> policy's driver data which necessitates grabbing the reference.
>>
>> Since schedutil governor can call "cpufreq_driver->update_perf()"
>> during sched_tick/enqueue/dequeue with rq_lock held and IRQs disabled,
>> fetching the policy object using the cpufreq_cpu_get() helper in the
>> scheduler fast-path leads to "BUG: scheduling while atomic" on
>> PREEMPT_RT [1].
>>
>> Pass the cached cpufreq policy object in sg_policy to the update_perf()
>> instead of just the CPU. The CPU can be inferred using "policy->cpu".
>>
>> The lifetime of cpufreq_policy object outlasts that of the governor and
>> the cpufreq driver (allocated when the CPU is onlined and only reclaimed
>> when the CPU is offlined / the CPU device is removed) which makes it
>> safe to be referenced throughout the governor's lifetime.
>>
>> Link: https://lore.kernel.org/all/20250731092316.3191-1-spasswolf@web.de/ [1]
> 
> I think you should have these tags instead:
> Reported-by: Bert Karwatzki <spasswolf@web.de>
> Closes:https://lore.kernel.org/all/20250731092316.3191-1-spasswolf@web.de/ [1]

Ack! I'll update it in the next version. Thank you for the review.

-- 
Thanks and Regards,
Prateek

Re: [RFC PATCH 0/2] cpufreq/amd-pstate: Prevent scheduling when atomic on PREEMPT_RT

[Viresh Kumar]

On 06-01-26, 07:36, K Prateek Nayak wrote:
> Bert reported hitting "BUG: scheduling while atomic" when running
> amd-pstate-ut on a PREEMPT_RT kernel [1].
> 
> Since reader-writer locks turn sleepable on PREEMPT_RT, they are not
> suitable to be used in the scheduler hot-path under rq_lock to grab the
> cpufreq policy object.
> 
> Unfortunately, the amd-pstate driver has a tight coupling between the
> cpufreq_policy object and the cpudata stored in it as the driver_data.
> 
> Trying to grab a read reference on PREEMPT_RT can cause "scheduling
> while atomic" if a concurrent writer is active, and trying to grab a
> nested reference in presence of a writer can cause a deadlock (manifests
> as lockup) since the reader fast-path is disabled on PREEMPT_RT to
> prevent write-side starvation.
> 
> The two patches included removes cases of grabbing a nested read
> reference to the cpufreq policy in amd-pstate, and modifies the
> cpufreq_driver->adjust_perf() callback to take the raw policy reference
> cached by the schedutil governor respectively.
> 
> The policy object outlives the governor and the driver making it safe to
> use this cache reference from the sugov data. Any changes to the policy
> will end up calling cpufreq_driver->set_policy() or
> governor->set_limits() once the policy is modified which should ensure
> eventual consistency despite not holding the read-side.
> 
> Series has been tested with amd-pstate-ut on PREEMPT_RT kernel which
> successfully passes without any splats on LOCKDEP + DEBUG_ATOMIC_SLEEP
> config. Additionally, the driver switch test from Gautham [2] was run
> for 10min on the same config without observing any splats.
> 
> [1] https://lore.kernel.org/all/20250731092316.3191-1-spasswolf@web.de/
> [2] https://lore.kernel.org/all/aJRN2wMLAnhDFykv@BLRRASHENOY1.amd.com/

Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

-- 
viresh