Skip to contentSkip to conversation

x86: fix idle notifier not being called in CONFIG_X86_32

PATCH In review Illyas Mansoor 3 participants 6 messages Mar 4, 2013 – Mar 25, 2013 x86

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86: fix idle notifier not being called in CONFIG_X86_32

Idle notifier not registered if CONFIG_X86_32 is defined, those callbacks are empty for X86_32 platform.

ifdef CONFIG_X86_64
void enter_idle(void);
void exit_idle(void);
else
static inline void enter_idle(void) { }
static inline void exit_idle(void) { }
static inline void __exit_idle(void) { }
endif

Make this work on X86_32 platforms by
removing the restriction for X86_64

---
 arch/x86/include/asm/idle.h |    6 ------
 arch/x86/kernel/process.c   |    4 ----
 2 files changed, 10 deletions(-)
arch/x86/include/asm/idle.h +0 -6
index c5d1785..489ac46 100644
@@ -8,14 +8,8 @@ struct notifier_block;
 void idle_notifier_register(struct notifier_block *n);
 void idle_notifier_unregister(struct notifier_block *n);
 
-#ifdef CONFIG_X86_64
 void enter_idle(void);
 void exit_idle(void);
-#else /* !CONFIG_X86_64 */
-static inline void enter_idle(void) { }
-static inline void exit_idle(void) { }
-static inline void __exit_idle(void) { }
-#endif /* CONFIG_X86_64 */
 
 void amd_e400_remove_cpu(int cpu);
 
arch/x86/kernel/process.c +0 -4
index 14ae100..043e553 100644
@@ -38,7 +38,6 @@
  */
 DEFINE_PER_CPU_SHARED_ALIGNED(struct tss_struct, init_tss) = INIT_TSS;
 
-#ifdef CONFIG_X86_64
 static DEFINE_PER_CPU(unsigned char, is_idle);
 static ATOMIC_NOTIFIER_HEAD(idle_notifier);
 
@@ -53,7 +52,6 @@ void idle_notifier_unregister(struct notifier_block *n)
 	atomic_notifier_chain_unregister(&idle_notifier, n);
 }
 EXPORT_SYMBOL_GPL(idle_notifier_unregister);
-#endif
 
 struct kmem_cache *task_xstate_cachep;
 EXPORT_SYMBOL_GPL(task_xstate_cachep);
@@ -277,7 +275,6 @@ static inline void play_dead(void)
 }
 #endif
 
-#ifdef CONFIG_X86_64
 void enter_idle(void)
 {
 	this_cpu_write(is_idle, 1);
@@ -299,7 +296,6 @@ void exit_idle(void)
 		return;
 	__exit_idle();
 }
-#endif
 
 /*
  * The idle thread. There's no useful work to be
Signature · 1 line
1.7.9.5
#1 lore raw
Quoted 12 lines · Illyas Mansoor

Idle notifier not registered if CONFIG_X86_32 is defined, those callbacks are empty for X86_32 platform.

ifdef CONFIG_X86_64
void enter_idle(void);
void exit_idle(void);
else
static inline void enter_idle(void) { }
static inline void exit_idle(void) { }
static inline void __exit_idle(void) { }
endif

Quoted 2 lines

Make this work on X86_32 platforms by
removing the restriction for X86_64

What will they be used for?

Thanks,

Ingo

#2 lore raw
Quoted 12 lines · Illyas Mansoor
Quoted 11 lines

Idle notifier not registered if CONFIG_X86_32 is defined, those callbacks are empty for X86_32 platform.

ifdef CONFIG_X86_64
void enter_idle(void);
void exit_idle(void);
else
static inline void enter_idle(void) { } static inline void exit_idle(void) { } static inline void __exit_idle(void) { } endif

Quoted 4 lines
Quoted 2 lines

Make this work on X86_32 platforms by
removing the restriction for X86_64

What will they be used for?

It's being used by interactive governor, and since the idle notifications are not received
It breaks the governor functionality on X86_32

Thanks,
-Illyas

#3 lore raw
Quoted 12 lines · Mansoor, Illyas
Quoted 12 lines · Illyas Mansoor
Quoted 11 lines

Idle notifier not registered if CONFIG_X86_32 is defined, those callbacks are empty for X86_32 platform.

ifdef CONFIG_X86_64
void enter_idle(void);
void exit_idle(void);
else
static inline void enter_idle(void) { } static inline void exit_idle(void) { } static inline void __exit_idle(void) { } endif

Quoted 7 lines
Quoted 4 lines
Quoted 2 lines

Make this work on X86_32 platforms by
removing the restriction for X86_64

What will they be used for?

It's being used by interactive governor, and since the idle notifications are not received It breaks the governor functionality on X86_32

But we never allowed idle notifiers on 32-bit and wanted to phase them out even on x86-64 as well.

There's ongoing work to improve power saving in the scheduler - see Alex Shi's patchset on lkml: I think the two pieces of code should cooperate within the scheduler instead of going in two directions, duplicating effort and getting in each other's way ...

Thanks,

Ingo

#4 lore raw
Quoted 29 lines

-----Original Message-----
From: Ingo Molnar [mailto:[email protected]] On Behalf Of Ingo Molnar
Sent: Wednesday, March 06, 2013 4:06 PM
To: Mansoor, Illyas; Peter Zijlstra
Cc: Linux Kernel; Linux PM; Frederic Weisbecker; Ingo Molnar; X86; Len Brown; Thomas Gleixner; Matthew Garrett; Tejun Heo; Paul E. McKenney; Rudramuni, Vishwesh M; [email protected]; [email protected]; Kumar P, Mahesh; Sil, Dyut K; Arjan van de Ven
Subject: Re: [PATCH] x86: fix idle notifier not being called in CONFIG_X86_32

Quoted 15 lines · Mansoor, Illyas
Quoted 15 lines · Illyas Mansoor
Quoted 13 lines

Idle notifier not registered if CONFIG_X86_32 is defined, those callbacks are empty for X86_32 platform.

ifdef CONFIG_X86_64
void enter_idle(void);
void exit_idle(void);
else
static inline void enter_idle(void) { } static inline void exit_idle(void) { } static inline void __exit_idle(void) { } endif

Make this work on X86_32 platforms by removing the restriction for X86_64

Quoted 7 lines
Quoted 5 lines
Quoted 1 line

What will they be used for?

It's being used by interactive governor, and since the idle notifications are not received It breaks the governor functionality on X86_32

But we never allowed idle notifiers on 32-bit and wanted to phase them out

even

Quoted 2 lines

on
x86-64 as well.

Support for Idle notifiers on 32-bit got broken from commit 90e240142bd31ff10aeda5a280a53153f4eff004
Before that they were getting registered from process_64.c since it was not guarded using
CONFIG_X86_64

Commit 90e240142bd31ff10aeda5a280a53153f4eff004 merged x8_64 and x86_32 into process.c and put
Restriction x86_64 on idle notifier which was correct since the notifier calls were taken from process_64.c

Quoted 4 lines

There's ongoing work to improve power saving in the scheduler - see Alex Shi's patchset on lkml: I think the two pieces of code should cooperate within the scheduler instead of going in two directions, duplicating effort and getting

in each

Quoted 1 line

other's way ...

Agree, it should not be conflicting each other.
Is there an alternative to idle notifications once those are deprecated. Sorry couldn't find that from Alex Shi's patches on LKML

Thanks,
Illyas

Acked-by
#5 lore raw

I'm okay with this patch since it makes 32-bit and 64-bit consistent, and technically, it fixes a regression.

If/when we get rid of the notifier, we should do it for both 32-and 64-bit at the same time.

It isn't clear to me that Alex's work will solve this particular problem, or that there is a code or logical conflict -- though maybe you are referring to some code I've not read yet.

So for this trivial patch...

But I think the bigger problem here is the expectation that out of tree drivers will not break. The interactive governor is not upstream.

Len Brown, Intel Open Source Technology Center