sched/fair: WARN and refuse to set buddy when !se->on_rq

If we set a next or last buddy for a se that is not on_rq, we will end up taking a NULL pointer dereference in wakeup_preempt_entity via pick_next_task_fair.

Detect when we would be about to do that, throw a warning and then refuse to actually set it.

This has been suggested at least twice[0][1]: just do it.

[0] https://marc.info/?l=linux-kernel&m=146651668921468&w=2
[1] https://lkml.org/lkml/2016/6/16/663

Cc Konstantin Khlebnikov <[email protected]>

Cc Ben Segall <[email protected]>

Cc Peter Zijlstra <[email protected]>

Cc Thomas Gleixner <[email protected]>

Signed-off-by Daniel Axtens <[email protected]>

---

I recently had to debug a problem with these (we hadn't backported Konstantin's patches in this area) and this would have saved a lot of time/pain.

v2: use SCHED_WARN_ON to restrict when the test is run. This is a
macro for WARN_ON_ONCE, which is convenient.

---
 kernel/sched/fair.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

kernel/sched/fair.c +8 -2

index d71109321841..44b94cfe02cb 100644
@@ -6168,8 +6168,11 @@ static void set_last_buddy(struct sched_entity *se)
 	if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
 		return;
 
-	for_each_sched_entity(se)
+	for_each_sched_entity(se) {
+		if (SCHED_WARN_ON(!se->on_rq))
+			return;
 		cfs_rq_of(se)->last = se;
+	}
 }
 
 static void set_next_buddy(struct sched_entity *se)

@@ -6177,8 +6180,11 @@ static void set_next_buddy(struct sched_entity *se)
 	if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
 		return;
 
-	for_each_sched_entity(se)
+	for_each_sched_entity(se) {
+		if (SCHED_WARN_ON(!se->on_rq))
+			return;
 		cfs_rq_of(se)->next = se;
+	}
 }
 
 static void set_skip_buddy(struct sched_entity *se)

Signature · 1 line

2.11.0

Hi Konstantin and Peter,

Just checking if this version was OK with you - I hadn't heard anything and I noticed it's not in -next so I just wanted to check to see if there were any other changes you wanted.

Regards,
Daniel

Quoted 62 lines · Daniel Axtens

If we set a next or last buddy for a se that is not on_rq, we will end up taking a NULL pointer dereference in wakeup_preempt_entity via pick_next_task_fair.

Detect when we would be about to do that, throw a warning and then refuse to actually set it.

This has been suggested at least twice[0][1]: just do it.

[0] https://marc.info/?l=linux-kernel&m=146651668921468&w=2
[1] https://lkml.org/lkml/2016/6/16/663

Cc Konstantin Khlebnikov <[email protected]>

Cc Ben Segall <[email protected]>

Cc Peter Zijlstra <[email protected]>

Cc Thomas Gleixner <[email protected]>

Signed-off-by Daniel Axtens <[email protected]>

---

I recently had to debug a problem with these (we hadn't backported Konstantin's patches in this area) and this would have saved a lot of time/pain.

v2: use SCHED_WARN_ON to restrict when the test is run. This is a
macro for WARN_ON_ONCE, which is convenient.

---
  kernel/sched/fair.c | 10 ++++++++--
  1 file changed, 8 insertions(+), 2 deletions(-)

kernel/sched/fair.c +8 -2

index d71109321841..44b94cfe02cb 100644
@@ -6168,8 +6168,11 @@ static void set_last_buddy(struct sched_entity *se)
  	if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
  		return;
 
-	for_each_sched_entity(se)
+	for_each_sched_entity(se) {
+		if (SCHED_WARN_ON(!se->on_rq))
+			return;
  		cfs_rq_of(se)->last = se;
+	}
  }
 
  static void set_next_buddy(struct sched_entity *se)

@@ -6177,8 +6180,11 @@ static void set_next_buddy(struct sched_entity *se)
  	if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
  		return;
 
-	for_each_sched_entity(se)
+	for_each_sched_entity(se) {
+		if (SCHED_WARN_ON(!se->on_rq))
+			return;
  		cfs_rq_of(se)->next = se;
+	}
  }
 
  static void set_skip_buddy(struct sched_entity *se)

Signature · 1 line

2.11.0

If we set a next or last buddy for a se that is not on_rq, we will end up taking a NULL pointer dereference in wakeup_preempt_entity via pick_next_task_fair.

Detect when we would be about to do that, throw a warning and then refuse to actually set it.

This has been suggested at least twice[0][1]: just do it.

[0] https://marc.info/?l=linux-kernel&m=146651668921468&w=2
[1] https://lkml.org/lkml/2016/6/16/663

Cc Konstantin Khlebnikov <[email protected]>

Cc Ben Segall <[email protected]>

Cc Peter Zijlstra <[email protected]>

Cc Thomas Gleixner <[email protected]>

Signed-off-by Daniel Axtens <[email protected]>

---

I recently had to debug a problem with these (we hadn't backported Konstantin's patches in this area) and this would have saved a lot of time/pain.

v2: use SCHED_WARN_ON to restrict when the test is run. This is a
macro for WARN_ON_ONCE, which is convenient.

---
 kernel/sched/fair.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

kernel/sched/fair.c +0 -0

index d71109321841..44b94cfe02cb 100644

From 6d3aed3d8a0573d0a6eb1160ccd0a0713f4dbc2f Mon Sep 17 00:00:00 2001

Quoted 36 lines

If we set a next or last buddy for a se that is not on_rq, we will end up taking a NULL pointer dereference in wakeup_preempt_entity via pick_next_task_fair.

Detect when we would be about to do that, throw a warning and then refuse to actually set it.

This has been suggested at least twice[0][1]: just do it.

[0] https://marc.info/?l=linux-kernel&m=146651668921468&w=2
[1] https://lkml.org/lkml/2016/6/16/663

Cc Konstantin Khlebnikov <[email protected]>

Cc Ben Segall <[email protected]>

Cc Peter Zijlstra <[email protected]>

Cc Thomas Gleixner <[email protected]>

Signed-off-by Daniel Axtens <[email protected]>

---

I recently had to debug a problem with these (we hadn't backported Konstantin's patches in this area) and this would have saved a lot of time/pain.

v2: use SCHED_WARN_ON to restrict when the test is run. This is a
macro for WARN_ON_ONCE, which is convenient.

---
 kernel/sched/fair.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

kernel/sched/fair.c +0 -0

index d71109321841..44b94cfe02cb 100644

kernel/sched/fair.c static void set_last_buddy(struct sched_entity *se) · line 6175

Quoted 9 lines

kernel/sched/fair.c +4 -1

@@ -6168,8 +6168,11 @@ static void set_last_buddy(struct sched_entity *se)
 	if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
 		return;
 
-	for_each_sched_entity(se)
+	for_each_sched_entity(se) {
+		if (SCHED_WARN_ON(!se->on_rq))
+			return;
 		cfs_rq_of(se)->last = se;
+	}

This won't build in the !CONFIG_SCHED_DEBUG case, because of the naive definition in sched.h:

#define SCHED_WARN_ON(x) ((void)(x))

That should be changed to something like:

#define SCHED_WARN_ON(x) ((void)(x))

I've applied the fix below. (untested at the moment)

Thanks,

Ingo

========================>
From 6d3aed3d8a0573d0a6eb1160ccd0a0713f4dbc2f Mon Sep 17 00:00:00 2001
From: Ingo Molnar <[email protected]>
Date: Tue, 20 Jun 2017 12:24:42 +0200
Subject: [PATCH] sched/debug: Fix SCHED_WARN_ON() to return a value on !CONFIG_SCHED_DEBUG as well

This definition of SCHED_WARN_ON():

#define SCHED_WARN_ON(x) ((void)(x))

is not fully compatible with the 'real' WARN_ON_ONCE() primitive, as it has no return value, so it cannot be used in conditionals.

Fix it.

Cc Daniel Axtens <[email protected]>

Cc Konstantin Khlebnikov <[email protected]>

Cc Peter Zijlstra <[email protected]>

Cc Mike Galbraith <[email protected]>

Cc Thomas Gleixner <[email protected]>

Cc Linus Torvalds <[email protected]>

Cc [email protected]

Signed-off-by Ingo Molnar <[email protected]>

---
 kernel/sched/sched.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

kernel/sched/sched.h +2 -2

index f2ef759a4cb6..e0329d10bdb8 100644
@@ -39,9 +39,9 @@
 #include "cpuacct.h"
 
 #ifdef CONFIG_SCHED_DEBUG
-#define SCHED_WARN_ON(x)	WARN_ONCE(x, #x)
+# define SCHED_WARN_ON(x)	WARN_ONCE(x, #x)
 #else
-#define SCHED_WARN_ON(x)	((void)(x))
+# define SCHED_WARN_ON(x)	({ (void)(x), 0; })
 #endif
 
 struct rq;