Skip to contentSkip to conversation

powercap: dtpm: Guard against missing energy model in dtpm callbacks

PATCH In review Elazar Leibovich 3 participants 7 messages Jun 11 – Jun 24

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover powercap: dtpm: Guard against missing energy model in dtpm callbacks

The dtpm_cpu and dtpm_devfreq powercap drivers dereference the perf domain returned by em_cpu_get()/em_pd_get() without checking it for NULL. When the energy model is absent - for instance when a CPU becomes impossible at runtime so get_cpu_device() returns NULL, or when the device EM was never registered or has already been unregistered - the dtpm callbacks, several of which are reachable directly from sysfs via the powercap zone attributes, crash on a NULL pointer dereference.

Add the missing NULL checks, mirroring the guard that already exists in get_pd_power_uw() since commit 46dc57406887 ("powercap: dtpm_cpu: Fix NULL pointer dereference in get_pd_power_uw()").

Note these patches only handle a perf domain that is already gone when the callback starts. A callback racing with em_dev_unregister_perf_domain() can still observe a non-NULL perf domain that is freed underneath it, because only the EM perf state table is RCU-protected, not struct em_perf_domain itself. Closing that race needs RCU lifetime guarantees from the EM core plus holding rcu_read_lock() across the perf domain access in the readers; this will be addressed in future patches.

Sivan Zohar-Kotzer (2):
powercap: dtpm_cpu: Guard em_cpu_get() against NULL return in sysfs
callbacks
powercap: dtpm_devfreq: Guard em_pd_get() against NULL return in
callbacks

 drivers/powercap/dtpm_cpu.c     | 6 ++++++
 drivers/powercap/dtpm_devfreq.c | 9 +++++++++
 2 files changed, 15 insertions(+)
Signature · 1 line
2.50.1 (Apple Git-155)
#1 lore raw

PATCH 1/2 powercap: dtpm_cpu: Guard em_cpu_get() against NULL return in sysfs callbacks

From: Sivan Zohar-Kotzer <[email protected]>

em_cpu_get() can return NULL when the CPU has no energy model registered (e.g. during hotplug races or initialization failures). Two call sites miss the NULL test:

1. set_pd_power_limit() — reachable from sysfs via the powercap constraint power_limit_uw store path. If NULL, return the current power limit unchanged (no-op) instead of crashing.

2. update_pd_power_uw() — called from CPU hotplug handlers via dtpm_update_power(). Return -EINVAL if the EM is absent, since this is an internal consistency failure that should not be swallowed.

The other sysfs-reachable path, get_pd_power_uw(), already has a NULL guard returning 0.

Fixes 0e8f68d7f048 ("powercap/drivers/dtpm: Add CPU energy model based support")
---
 drivers/powercap/dtpm_cpu.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/powercap/dtpm_cpu.c +6 -0
index 99390ec1481f..0a460f97bf15 100644
@@ -47,6 +47,9 @@ static u64 set_pd_power_limit(struct dtpm *dtpm, u64 power_limit)
 	u64 power;
 	int i, nr_cpus;
 
+	if (!pd)
+		return dtpm->power_limit;
+
 	nr_cpus = cpumask_weight_and(cpu_online_mask, to_cpumask(pd->cpus));
 
 	rcu_read_lock();
@@ -125,6 +128,9 @@ static int update_pd_power_uw(struct dtpm *dtpm)
 	struct em_perf_state *table;
 	int nr_cpus;
 
+	if (!em)
+		return -EINVAL;
+
 	nr_cpus = cpumask_weight_and(cpu_online_mask, to_cpumask(em->cpus));
 
 	rcu_read_lock();
Signature · 1 line
2.50.1 (Apple Git-155)
#2 lore raw

Re: [PATCH 1/2] powercap: dtpm_cpu: Guard em_cpu_get() against NULL return in sysfs callbacks

Hi,

Quoted 9 lines · Elazar Leibovich

From: Sivan Zohar-Kotzer <[email protected]>

em_cpu_get() can return NULL when the CPU has no energy model registered (e.g. during hotplug races or initialization failures). Two call sites miss the NULL test:

1. set_pd_power_limit() — reachable from sysfs via the powercap constraint power_limit_uw store path. If NULL, return the current power limit unchanged (no-op) instead of crashing.

If there is no energy model, then __dtpm_cpu_setup() fails and dtpm_ops is not set. Consequently, set_pd_power_limit() can not be called.

Quoted 3 lines

2. update_pd_power_uw() — called from CPU hotplug handlers via dtpm_update_power(). Return -EINVAL if the EM is absent, since this is an internal consistency failure that should not be swallowed.

If there is no energy model, then __dtpm_cpu_setup() fails and dtpm_per_cpu is not set and dtpm_update_power() is not called.

online/offline:

        if (dtpm_cpu)
                dtpm_update_power(&dtpm_cpu->dtpm);
Quoted 35 lines

The other sysfs-reachable path, get_pd_power_uw(), already has a NULL guard returning 0.

Fixes 0e8f68d7f048 ("powercap/drivers/dtpm: Add CPU energy model based support")
---
  drivers/powercap/dtpm_cpu.c | 6 ++++++
  1 file changed, 6 insertions(+)
drivers/powercap/dtpm_cpu.c +6 -0
index 99390ec1481f..0a460f97bf15 100644
@@ -47,6 +47,9 @@ static u64 set_pd_power_limit(struct dtpm *dtpm, u64 power_limit)
  	u64 power;
  	int i, nr_cpus;
 
+	if (!pd)
+		return dtpm->power_limit;
+
  	nr_cpus = cpumask_weight_and(cpu_online_mask, to_cpumask(pd->cpus));
 
  	rcu_read_lock();
@@ -125,6 +128,9 @@ static int update_pd_power_uw(struct dtpm *dtpm)
  	struct em_perf_state *table;
  	int nr_cpus;
 
+	if (!em)
+		return -EINVAL;
+
  	nr_cpus = cpumask_weight_and(cpu_online_mask, to_cpumask(em->cpus));
 
  	rcu_read_lock();
#3 lore raw

PATCH 2/2 powercap: dtpm_devfreq: Guard em_pd_get() against NULL return in callbacks

In reply to Elazar Leibovich

From: Sivan Zohar-Kotzer <[email protected]>

em_pd_get() returns NULL when the device has no energy model registered. All three dtpm callbacks dereference the perf domain without checking it:

1. set_pd_power_limit() — reachable from sysfs via the powercap constraint power_limit_uw store path. If NULL, return the current power limit unchanged (no-op) instead of crashing.

2. get_pd_power_uw() — reachable from sysfs via the power_uw read path. Return 0 if the EM is absent.

3. update_pd_power_uw() — called via dtpm_update_power(). Return -EINVAL if the EM is absent.

This mirrors the equivalent NULL checks in dtpm_cpu.c.

Note this only handles the case where the EM is already gone (or was never registered) when the callback starts. A callback racing with em_dev_unregister_perf_domain() can still observe a non-NULL perf domain that is freed underneath it; closing that race requires lifetime guarantees from the EM core and is addressed separately.

Fixes e44655617317 ("powercap/drivers/dtpm: Add dtpm devfreq with energy model support")
---
 drivers/powercap/dtpm_devfreq.c | 9 +++++++++
 1 file changed, 9 insertions(+)
drivers/powercap/dtpm_devfreq.c +9 -0
index d1dff6ccab12..cf16e2756481 100644
@@ -39,6 +39,9 @@ static int update_pd_power_uw(struct dtpm *dtpm)
 	struct em_perf_domain *pd = em_pd_get(dev);
 	struct em_perf_state *table;
 
+	if (!pd)
+		return -EINVAL;
+
 	rcu_read_lock();
 	table = em_perf_state_from_pd(pd);
 
@@ -60,6 +63,9 @@ static u64 set_pd_power_limit(struct dtpm *dtpm, u64 power_limit)
 	unsigned long freq;
 	int i;
 
+	if (!pd)
+		return dtpm->power_limit;
+
 	rcu_read_lock();
 	table = em_perf_state_from_pd(pd);
 	for (i = 0; i < pd->nr_perf_states; i++) {
@@ -102,6 +108,9 @@ static u64 get_pd_power_uw(struct dtpm *dtpm)
 	u64 power = 0;
 	int i;
 
+	if (!pd)
+		return 0;
+
 	mutex_lock(&devfreq->lock);
 	status = devfreq->last_status;
 	mutex_unlock(&devfreq->lock);
Signature · 1 line
2.50.1 (Apple Git-155)
#4 lore raw

PATCH v2 1/2 powercap: dtpm_cpu: Guard em_cpu_get() against NULL return in sysfs callbacks

In reply to Elazar Leibovich

From: Sivan Zohar-Kotzer <[email protected]>

em_cpu_get() can return NULL when a CPU becomes impossible. Two call sites miss the NULL test:

1. set_pd_power_limit() — reachable from sysfs via the powercap constraint power_limit_uw store path. If the CPU has become impossible then pd will be NULL, return the current power limit unchanged (no-op) instead of crashing.

2. update_pd_power_uw() — called from CPU hotplug handlers via dtpm_update_power(). While reaching this with a NULL em may be technically impossible today, the check is kept for defense in depth against future refactoring or subtle races.

The other sysfs-reachable path, get_pd_power_uw(), already has a NULL guard returning 0.

Fixes 0e8f68d7f048 ("powercap/drivers/dtpm: Add CPU energy model based support")
---
 drivers/powercap/dtpm_cpu.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/powercap/dtpm_cpu.c +6 -0
index 21355db64..886cbe922 100644
@@ -47,6 +47,9 @@ static u64 set_pd_power_limit(struct dtpm *dtpm, u64 power_limit)
 	u64 power;
 	int i, nr_cpus;
 
+	if (!pd)
+		return dtpm->power_limit;
+
 	nr_cpus = cpumask_weight_and(cpu_online_mask, to_cpumask(pd->cpus));
 
 	rcu_read_lock();
@@ -125,6 +128,9 @@ static int update_pd_power_uw(struct dtpm *dtpm)
 	struct em_perf_state *table;
 	int nr_cpus;
 
+	if (!em)
+		return -EINVAL;
+
 	nr_cpus = cpumask_weight_and(cpu_online_mask, to_cpumask(em->cpus));
 
 	rcu_read_lock();
Signature · 1 line
2.39.5 (Apple Git-154)
#5 lore raw

Re: [PATCH v2 1/2] powercap: dtpm_cpu: Guard em_cpu_get() against NULL return in sysfs callbacks

Quoted 9 lines · On 6/24/26 22:31, [email protected] wrote:

From: Sivan Zohar-Kotzer <[email protected]>

em_cpu_get() can return NULL when a CPU becomes impossible. Two call sites miss the NULL test:

1. set_pd_power_limit() — reachable from sysfs via the powercap constraint power_limit_uw store path. If the CPU has become impossible then pd will be NULL, return the current power limit unchanged (no-op) instead of crashing.

Can you provide a script or a recipe spotting these issues ?

Quoted 40 lines

2. update_pd_power_uw() — called from CPU hotplug handlers via dtpm_update_power(). While reaching this with a NULL em may be technically impossible today, the check is kept for defense in depth against future refactoring or subtle races.

The other sysfs-reachable path, get_pd_power_uw(), already has a NULL guard returning 0.

Fixes 0e8f68d7f048 ("powercap/drivers/dtpm: Add CPU energy model based support")
---
  drivers/powercap/dtpm_cpu.c | 6 ++++++
  1 file changed, 6 insertions(+)
drivers/powercap/dtpm_cpu.c +6 -0
index 21355db64..886cbe922 100644
@@ -47,6 +47,9 @@ static u64 set_pd_power_limit(struct dtpm *dtpm, u64 power_limit)
  	u64 power;
  	int i, nr_cpus;
 
+	if (!pd)
+		return dtpm->power_limit;
+
  	nr_cpus = cpumask_weight_and(cpu_online_mask, to_cpumask(pd->cpus));
 
  	rcu_read_lock();
@@ -125,6 +128,9 @@ static int update_pd_power_uw(struct dtpm *dtpm)
  	struct em_perf_state *table;
  	int nr_cpus;
 
+	if (!em)
+		return -EINVAL;
+
  	nr_cpus = cpumask_weight_and(cpu_online_mask, to_cpumask(em->cpus));
 
  	rcu_read_lock();
#6 lore raw

Re: [PATCH v2 1/2] powercap: dtpm_cpu: Guard em_cpu_get() against NULL return in sysfs callbacks

Quoted 4 lines · Daniel Lezcano

Can you provide a script or a recipe spotting these issues ?

Thanks!

I'll describe the thought process.

In 46dc57406887 we found a single instance of em_cpu_get() without a NULL check even though it can return a NULL. I think the catalyst was seeing that this function could return NULL and not seeing NULL test on a call site.

After some discussion we concluded that when a CPU becomes impossible em_cpu_get can race and dereference a NULL pointer theoretically.

While we acknowledge this isn't a huge issue, a NULL test in a non-critical path doesn't seem like a huge issue either.

We grep'ed the code for other usages of em_cpu_get, and traced the calling process to see if they can race a CPU becoming impossible.

(While looking at the code, we noticed the rcu does not guard the PD pointer, but that's for another WIP patch)

I'm not sure if that's was the question, but we used AI to:

1. Trace the possible call stack of a function. It did a tremendous job there. Faster and more precise than myself, seeing through "vtables" and describing the callstack correctly.
2. Find similar cases of RCU used in the kernel, and try to imitate the behavior (for the not-yet-sent RCU WIP patch)
3. General kernel chores (find maintainers list, running sparse, etc.)
4. Create a reproducer for the RCU race to test it on a VM.

Let me know if you want more details!