Skip to contentSkip to conversation

Extend mwait idle to optimize away IPIs when possible

RFC In review Venkatesh Pallipadi 7 participants 22 messages Feb 6, 2012 – Feb 15, 2012

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

RFC 1/1 Extend mwait idle to optimize away IPIs when possible

smp_call_function_single and ttwu_queue_remote sends unconditional IPI to target CPU. However, if the target CPU is in mwait based idle, we can do IPI-less wakeups using the magical powers of monitor-mwait. Doing this has certain advantages:
* Lower overhead on Async IPI send path. Measurements on Westmere based systems show savings on "no wait" smp_call_function_single with idle target CPU (as measured on the sender side).
local socket smp_call_func cost goes from ~1600 to ~1200 cycles remote socket smp_call_func cost goes from ~2000 to ~1800 cycles
* Avoiding actual interrupts shows a measurable reduction (10%) in system non-idle cycles and cache-references with micro-benchmark sending IPI from one CPU to all the other mostly idle CPUs in the system.
* On a mostly idle system, turbostat shows a tiny decrease in C0(active) time and a corresponding increase in C6 state (Each row being 10min avg)

          %c0   %c1   %c6
  Before
  Run 1  1.51  2.93 95.55
  Run 2  1.48  2.86 95.65
  Run 3  1.46  2.78 95.74
  After
  Run 1  1.35  2.63 96.00
  Run 2  1.46  2.78 95.74
  Run 3  1.37  2.63 95.98

* As a bonus, we can avoid sched/call IPI overhead altogether in a special case. When CPU Y has woken up CPU X (which can take 50-100us to actually wakeup from a deep idle state) and CPU Z wants to send IPI to CPU X in this period. It can get it for free.

We started looking at this with one of our workloads where system is partially busy and we noticed some kernel hotspots in find_next_bit and default_send_IPI_mask_sequence_phys coming from sched wakeup (futex wakeups) and networking call functions. So, this change addresses those two specific IPI types. This could be extended to nohz_kick, etc.

Note:
* This only helps when target CPU is idle. When it is busy we will still send IPI as before.
* Only for X86_64 and mwait_idle_with_hints for now, with limited testing.
* Will need some accounting for these wakeups exported for powertop and friends.

Comments?

---
 arch/x86/Kconfig                   |    5 ++++
 arch/x86/include/asm/thread_info.h |    1 +
 arch/x86/kernel/acpi/cstate.c      |   22 ++++++++++++++++
 arch/x86/kernel/process_64.c       |   12 +++++++++
 include/linux/sched.h              |   47 ++++++++++++++++++++++++++++++++++++
 include/linux/thread_info.h        |    2 +
 kernel/sched/core.c                |   40 +++++++++++++++++++++++++++++-
 kernel/smp.c                       |    4 +++
 8 files changed, 132 insertions(+), 1 deletions(-)
arch/x86/Kconfig +5 -0
index 5bed94e..6a56cb0 100644
@@ -170,6 +170,11 @@ config GENERIC_TIME_VSYSCALL
 	bool
 	default X86_64
 
+config AVOID_IPI_ON_IDLE
+	bool
+	depends on X86_64
+	default X86_64
+
 config ARCH_HAS_CPU_RELAX
 	def_bool y
 
arch/x86/include/asm/thread_info.h +1 -0
index bc817cd..24e4a56 100644
@@ -95,6 +95,7 @@ struct thread_info {
 #define TIF_BLOCKSTEP		25	/* set when we want DEBUGCTLMSR_BTF */
 #define TIF_LAZY_MMU_UPDATES	27	/* task is updating the mmu lazily */
 #define TIF_SYSCALL_TRACEPOINT	28	/* syscall tracepoint instrumentation */
+#define TIF_IPI_PENDING		29	/* pending IPI work on idle exit */
 
 #define _TIF_SYSCALL_TRACE	(1 << TIF_SYSCALL_TRACE)
 #define _TIF_NOTIFY_RESUME	(1 << TIF_NOTIFY_RESUME)
arch/x86/kernel/acpi/cstate.c +22 -0
index f50e7fb..340c4e9 100644
@@ -161,6 +161,14 @@ EXPORT_SYMBOL_GPL(acpi_processor_ffh_cstate_probe);
  */
 void mwait_idle_with_hints(unsigned long ax, unsigned long cx)
 {
+#ifdef CONFIG_AVOID_IPI_ON_IDLE
+	int oldval, cpu = smp_processor_id();
+
+	if (!system_using_cpu_idle_sync)
+		system_using_cpu_idle_sync = 1;
+
+	atomic_set(&per_cpu(cpu_idle_sync, cpu), CPU_STATE_IDLE);
+#endif
 	if (!need_resched()) {
 		if (this_cpu_has(X86_FEATURE_CLFLUSH_MONITOR))
 			clflush((void *)&current_thread_info()->flags);
@@ -170,6 +178,20 @@ void mwait_idle_with_hints(unsigned long ax, unsigned long cx)
 		if (!need_resched())
 			__mwait(ax, cx);
 	}
+#ifdef CONFIG_AVOID_IPI_ON_IDLE
+	oldval = atomic_cmpxchg(&per_cpu(cpu_idle_sync, cpu),
+				CPU_STATE_IDLE, CPU_STATE_RUNNING);
+
+	if (oldval != CPU_STATE_IDLE) {
+		oldval = atomic_cmpxchg(&per_cpu(cpu_idle_sync, cpu),
+					CPU_STATE_WOKENUP, CPU_STATE_RUNNING);
+		while (oldval != CPU_STATE_WOKENUP) {
+			rep_nop();
+			oldval = atomic_cmpxchg(&per_cpu(cpu_idle_sync, cpu),
+					CPU_STATE_WOKENUP, CPU_STATE_RUNNING);
+		}
+	}
+#endif
 }
 
 void acpi_processor_ffh_cstate_enter(struct acpi_processor_cx *cx)
arch/x86/kernel/process_64.c +12 -0
index 9b9fe4a..4c6bb4a 100644
@@ -153,6 +153,18 @@ void cpu_idle(void)
 			   has already called exit_idle. But some idle
 			   loops can be woken up without interrupt. */
 			__exit_idle();
+
+#ifdef CONFIG_AVOID_IPI_ON_IDLE
+			if (ipi_pending()) {
+				clear_ipi_pending();
+				local_bh_disable();
+				local_irq_disable();
+				generic_smp_call_function_single_interrupt();
+				scheduler_wakeup_self_check();
+				local_irq_enable();
+				local_bh_enable();
+			}
+#endif
 		}
 
 		tick_nohz_idle_exit();
include/linux/sched.h +47 -0

59 lines not on this page · raw message

include/linux/thread_info.h +2 -0
index 8d03f07..9a458e2 100644
@@ -98,6 +98,8 @@ static inline int test_ti_thread_flag(struct thread_info *ti, int flag)
 #define set_need_resched()	set_thread_flag(TIF_NEED_RESCHED)
 #define clear_need_resched()	clear_thread_flag(TIF_NEED_RESCHED)
 
+#define clear_ipi_pending()	clear_thread_flag(TIF_IPI_PENDING)
+
 #if defined TIF_RESTORE_SIGMASK && !defined HAVE_SET_RESTORE_SIGMASK
 /*
  * An arch can define its own version of set_restore_sigmask() to get the
kernel/sched/core.c +39 -1
index df00cb0..14c9fc8 100644
@@ -492,6 +492,9 @@ static inline void init_hrtick(void)
 }
 #endif	/* CONFIG_SCHED_HRTICK */
 
+DEFINE_PER_CPU(atomic_t, cpu_idle_sync);
+__read_mostly int system_using_cpu_idle_sync;
+
 /*
  * resched_task - mark a task 'to be rescheduled now'.
  *
@@ -505,6 +508,29 @@ static inline void init_hrtick(void)
 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
 #endif
 
+#ifdef CONFIG_AVOID_IPI_ON_IDLE
+int smp_need_ipi(int cpu)
+{
+	int oldval;
+
+	if (!system_using_cpu_idle_sync || cpu == smp_processor_id())
+		return 1;
+
+	oldval = atomic_cmpxchg(&per_cpu(cpu_idle_sync, cpu),
+				CPU_STATE_IDLE, CPU_STATE_WAKING);
+
+	if (oldval == CPU_STATE_RUNNING)
+		return 1;
+
+	if (oldval == CPU_STATE_IDLE) {
+		set_tsk_ipi_pending(idle_task(cpu));
+		atomic_set(&per_cpu(cpu_idle_sync, cpu), CPU_STATE_WOKENUP);
+	}
+
+	return 0;
+}
+#endif
+
 void resched_task(struct task_struct *p)
 {
 	int cpu;
@@ -1456,6 +1482,16 @@ static void sched_ttwu_pending(void)
 	raw_spin_unlock(&rq->lock);
 }
 
+#ifdef CONFIG_AVOID_IPI_ON_IDLE
+void scheduler_wakeup_self_check(void)
+{
+	if (llist_empty(&this_rq()->wake_list))
+		return;
+
+	sched_ttwu_pending();
+}
+#endif
+
 void scheduler_ipi(void)
 {
 	if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
@@ -1490,7 +1526,8 @@ void scheduler_ipi(void)
 static void ttwu_queue_remote(struct task_struct *p, int cpu)
 {
 	if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
-		smp_send_reschedule(cpu);
+		if (smp_need_ipi(cpu))
+			smp_send_reschedule(cpu);
 }
 
 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
@@ -6943,6 +6980,7 @@ void __init sched_init(void)
 #endif
 		init_rq_hrtick(rq);
 		atomic_set(&rq->nr_iowait, 0);
+		atomic_set(&per_cpu(cpu_idle_sync, i), CPU_STATE_RUNNING);
 	}
 
 	set_load_weight(&init_task);
kernel/smp.c +4 -0
index db197d6..48a9c4a 100644
@@ -8,6 +8,7 @@
 #include <linux/kernel.h>
 #include <linux/export.h>
 #include <linux/percpu.h>
+#include <linux/sched.h>
 #include <linux/init.h>
 #include <linux/gfp.h>
 #include <linux/smp.h>
@@ -145,6 +146,9 @@ void generic_exec_single(int cpu, struct call_single_data *data, int wait)
 	list_add_tail(&data->list, &dst->list);
 	raw_spin_unlock_irqrestore(&dst->lock, flags);
 
+	if (ipi)
+		ipi = smp_need_ipi(cpu);
+
 	/*
 	 * The list addition should be visible before sending the IPI
 	 * handler locks the list to pull the entry off it because of
Signature · 1 line
1.7.7.3
maintainer
#1 lore raw
Quoted 35 lines · Venkatesh Pallipadi

smp_call_function_single and ttwu_queue_remote sends unconditional IPI to target CPU. However, if the target CPU is in mwait based idle, we can do IPI-less wakeups using the magical powers of monitor-mwait. Doing this has certain advantages:
* Lower overhead on Async IPI send path. Measurements on Westmere based systems show savings on "no wait" smp_call_function_single with idle target CPU (as measured on the sender side).
local socket smp_call_func cost goes from ~1600 to ~1200 cycles remote socket smp_call_func cost goes from ~2000 to ~1800 cycles
* Avoiding actual interrupts shows a measurable reduction (10%) in system non-idle cycles and cache-references with micro-benchmark sending IPI from one CPU to all the other mostly idle CPUs in the system.
* On a mostly idle system, turbostat shows a tiny decrease in C0(active) time and a corresponding increase in C6 state (Each row being 10min avg)

          %c0   %c1   %c6
  Before
  Run 1  1.51  2.93 95.55
  Run 2  1.48  2.86 95.65
  Run 3  1.46  2.78 95.74
  After
  Run 1  1.35  2.63 96.00
  Run 2  1.46  2.78 95.74
  Run 3  1.37  2.63 95.98

* As a bonus, we can avoid sched/call IPI overhead altogether in a special case. When CPU Y has woken up CPU X (which can take 50-100us to actually wakeup from a deep idle state) and CPU Z wants to send IPI to CPU X in this period. It can get it for free.

We started looking at this with one of our workloads where system is partially busy and we noticed some kernel hotspots in find_next_bit and default_send_IPI_mask_sequence_phys coming from sched wakeup (futex wakeups) and networking call functions. So, this change addresses those two specific IPI types. This could be extended to nohz_kick, etc.

Quoted 7 lines

Note:
* This only helps when target CPU is idle. When it is busy we will still send IPI as before.
* Only for X86_64 and mwait_idle_with_hints for now, with limited testing.
* Will need some accounting for these wakeups exported for powertop and friends.

Comments?

Curiously you avoided the existing tsk_is_polling() magic, which IIRC is doing something similar for waking from the idle loop.

#2 lore raw
Quoted 42 lines · Peter Zijlstra
Quoted 41 lines · Venkatesh Pallipadi

smp_call_function_single and ttwu_queue_remote sends unconditional IPI to target CPU. However, if the target CPU is in mwait based idle, we can do IPI-less wakeups using the magical powers of monitor-mwait. Doing this has certain advantages:
* Lower overhead on Async IPI send path. Measurements on Westmere based systems show savings on "no wait" smp_call_function_single with idle target CPU (as measured on the sender side).
  local socket smp_call_func cost goes from ~1600 to ~1200 cycles remote socket smp_call_func cost goes from ~2000 to ~1800 cycles
* Avoiding actual interrupts shows a measurable reduction (10%) in system non-idle cycles and cache-references with micro-benchmark sending IPI from one CPU to all the other mostly idle CPUs in the system.
* On a mostly idle system, turbostat shows a tiny decrease in C0(active) time and a corresponding increase in C6 state (Each row being 10min avg) %c0   %c1   %c6
  Before
  Run 1  1.51  2.93 95.55
  Run 2  1.48  2.86 95.65
  Run 3  1.46  2.78 95.74
  After
  Run 1  1.35  2.63 96.00
  Run 2  1.46  2.78 95.74
  Run 3  1.37  2.63 95.98

* As a bonus, we can avoid sched/call IPI overhead altogether in a special case. When CPU Y has woken up CPU X (which can take 50-100us to actually wakeup from a deep idle state) and CPU Z wants to send IPI to CPU X in this period. It can get it for free.

We started looking at this with one of our workloads where system is partially busy and we noticed some kernel hotspots in find_next_bit and default_send_IPI_mask_sequence_phys coming from sched wakeup (futex wakeups) and networking call functions. So, this change addresses those two specific IPI types. This could be extended to nohz_kick, etc.

Note:
* This only helps when target CPU is idle. When it is busy we will still send IPI as before.
* Only for X86_64 and mwait_idle_with_hints for now, with limited testing.
* Will need some accounting for these wakeups exported for powertop and friends.

Quoted 5 lines
Quoted 1 line

Comments?

Curiously you avoided the existing tsk_is_polling() magic, which IIRC is doing something similar for waking from the idle loop.

Yes. That needs remote CPU's current task, which extends onto rq lock, which I was trying to avoid. So, I went with conditional waiting on idle exit for the small window of WAKING to WOKEN state change, as we know we are always polling in the mwait loop.

#3 lore raw
In reply to Venkatesh Pallipadi

On 02/06/2012 12:42 PM, Venkatesh Pallipadi wrote:
[...]

arch/x86/Kconfig config GENERIC_TIME_VSYSCALL · line 177

Quoted 12 lines

index 5bed94e..6a56cb0 100644

arch/x86/Kconfig +5 -0
@@ -170,6 +170,11 @@ config GENERIC_TIME_VSYSCALL
  	bool
  	default X86_64
 
+config AVOID_IPI_ON_IDLE
+	bool
+	depends on X86_64
+	default X86_64
+

Can you get rid of this part...

Quoted 3 lines
arch/x86/Kconfig +0 -0
  config ARCH_HAS_CPU_RELAX
  	def_bool y
 

[...]

arch/x86/kernel/acpi/cstate.c EXPORT_SYMBOL_GPL(acpi_processor_ffh_cstate_probe); · line 164

Quoted 9 lines
arch/x86/kernel/acpi/cstate.c +1 -0
index f50e7fb..340c4e9 100644
@@ -161,6 +161,14 @@ EXPORT_SYMBOL_GPL(acpi_processor_ffh_cstate_probe);
   */
  void mwait_idle_with_hints(unsigned long ax, unsigned long cx)
  {
+#ifdef CONFIG_AVOID_IPI_ON_IDLE

... and change all of these to #ifdef CONFIG_64BIT

arch/x86/kernel/acpi/cstate.c EXPORT_SYMBOL_GPL(acpi_processor_ffh_cstate_probe); · line 166

Quoted 2 lines
arch/x86/kernel/acpi/cstate.c +2 -0
+	int oldval, cpu = smp_processor_id();
+

.
.
.

Thanks,
David Daney

#4 lore raw
Quoted 21 lines · David Daney

On 02/06/2012 12:42 PM, Venkatesh Pallipadi wrote:
[...]

arch/x86/Kconfig config GENERIC_TIME_VSYSCALL · line 177

Quoted 12 lines

index 5bed94e..6a56cb0 100644

arch/x86/Kconfig +5 -0
@@ -170,6 +170,11 @@ config GENERIC_TIME_VSYSCALL
      bool
      default X86_64
 
+config AVOID_IPI_ON_IDLE
+    bool
+    depends on X86_64
+    default X86_64
+

Can you get rid of this part...

Quoted 3 lines
arch/x86/Kconfig +0 -0
  config ARCH_HAS_CPU_RELAX
      def_bool y
 

[...]

Quoted 16 lines

arch/x86/kernel/acpi/cstate.c EXPORT_SYMBOL_GPL(acpi_processor_ffh_cstate_probe); · line 164

Quoted 10 lines
arch/x86/kernel/acpi/cstate.c +1 -0
index f50e7fb..340c4e9 100644
@@ -161,6 +161,14 @@ EXPORT_SYMBOL_GPL(acpi_processor_ffh_cstate_probe);
   */
  void mwait_idle_with_hints(unsigned long ax, unsigned long cx)
  {
+#ifdef CONFIG_AVOID_IPI_ON_IDLE

... and change all of these to #ifdef CONFIG_64BIT

Quoted 2 lines
arch/x86/kernel/acpi/cstate.c +2 -0
+    int oldval, cpu = smp_processor_id();
+

Why should this be 64-bit specific??

-hpa

Signature · 2 lines
H. Peter Anvin, Intel Open Source Technology Center
I work for Intel.  I don't speak on their behalf.
#5 lore raw
Quoted 36 lines · H. Peter Anvin
Quoted 35 lines · David Daney

On 02/06/2012 12:42 PM, Venkatesh Pallipadi wrote:
[...]

arch/x86/Kconfig config GENERIC_TIME_VSYSCALL · line 177

Quoted 12 lines

index 5bed94e..6a56cb0 100644

arch/x86/Kconfig +5 -0
@@ -170,6 +170,11 @@ config GENERIC_TIME_VSYSCALL
       bool
       default X86_64
 
+config AVOID_IPI_ON_IDLE
+    bool
+    depends on X86_64
+    default X86_64
+

Can you get rid of this part...

Quoted 3 lines
arch/x86/Kconfig +0 -0
   config ARCH_HAS_CPU_RELAX
       def_bool y
 

[...]

arch/x86/kernel/acpi/cstate.c EXPORT_SYMBOL_GPL(acpi_processor_ffh_cstate_probe); · line 164

Quoted 10 lines
arch/x86/kernel/acpi/cstate.c +1 -0
index f50e7fb..340c4e9 100644
@@ -161,6 +161,14 @@ EXPORT_SYMBOL_GPL(acpi_processor_ffh_cstate_probe);
    */
   void mwait_idle_with_hints(unsigned long ax, unsigned long cx)
   {
+#ifdef CONFIG_AVOID_IPI_ON_IDLE

... and change all of these to #ifdef CONFIG_64BIT

Quoted 5 lines
Quoted 2 lines

arch/x86/kernel/acpi/cstate.c EXPORT_SYMBOL_GPL(acpi_processor_ffh_cstate_probe); · line 166

Quoted 2 lines
arch/x86/kernel/acpi/cstate.c +2 -0
+    int oldval, cpu = smp_processor_id();
+

Why should this be 64-bit specific??

Well I don't know that part. The gist of my comment was to get rid of an unneeded new config variable.

If this works just as well for 32-bit, then by all means do it there too.

In any event, I think you want the kernel to automatically do it if the CPU is capable, and not have a manual config setting you have to fiddle with.

David Daney

#6 lore raw
Quoted 10 lines · David Daney

Well I don't know that part. The gist of my comment was to get rid of an unneeded new config variable.

If this works just as well for 32-bit, then by all means do it there too.

In any event, I think you want the kernel to automatically do it if the CPU is capable, and not have a manual config setting you have to fiddle with.

If this is a transitional thing then it might justify it.

-hpa

#7 lore raw
Quoted 6 lines · H. Peter Anvin
Quoted 5 lines · David Daney

Well I don't know that part. The gist of my comment was to get rid of an unneeded new config variable.

Quoted 9 lines
Quoted 6 lines

If this works just as well for 32-bit, then by all means do it there too.

In any event, I think you want the kernel to automatically do it if the CPU is capable, and not have a manual config setting you have to fiddle with.

If this is a transitional thing then it might justify it.

There is nothing specific to 64 bit here. Its just that I had only tested this on 64 bit right now and wanted to get early feedback. So, in the spirit of RFC, I sent this out as x86_64 only for the time being. If and when this will be sent for inclusion, this will should work on both 32 bit and 64 bit on x86. Though we will need a config option to not add any run time overhead in generic code (and to not have #ifdef CONFIG_X86 in kernel/).

Thanks,
Venki

#8 lore raw
In reply to Venkatesh Pallipadi
Quoted 5 lines · Venkatesh Pallipadi

* Lower overhead on Async IPI send path. Measurements on Westmere based systems show savings on "no wait" smp_call_function_single with idle target CPU (as measured on the sender side).
local socket smp_call_func cost goes from ~1600 to ~1200 cycles remote socket smp_call_func cost goes from ~2000 to ~1800 cycles

Interesting that savings in the remote socket is less compared to the local socket.

kernel/sched/core.c static inline void init_hrtick(void) · line 520

Quoted 9 lines
kernel/sched/core.c +9 -0
+int smp_need_ipi(int cpu)
+{
+	int oldval;
+
+	if (!system_using_cpu_idle_sync || cpu == smp_processor_id())
+		return 1;
+
+	oldval = atomic_cmpxchg(&per_cpu(cpu_idle_sync, cpu),
+				CPU_STATE_IDLE, CPU_STATE_WAKING);

To avoid too many cache line bounces for the case when the cpu is in the running state, we should do a read to check if the state is in idle before going ahead with the locked operation?

kernel/sched/core.c static inline void init_hrtick(void) · line 530

Quoted 10 lines
kernel/sched/core.c +10 -0
+
+	if (oldval == CPU_STATE_RUNNING)
+		return 1;
+
+	if (oldval == CPU_STATE_IDLE) {
+		set_tsk_ipi_pending(idle_task(cpu));
+		atomic_set(&per_cpu(cpu_idle_sync, cpu), CPU_STATE_WOKENUP);
+	}
+
+	return 0;

We should probably disable interrupts around this, otherwise any delay in transitioning to wokenup from waking will cause the idle cpu to be stuck for similar amount of time.

thanks,
suresh

#9 lore raw
Quoted 9 lines · Suresh Siddha
Quoted 5 lines · Venkatesh Pallipadi

* Lower overhead on Async IPI send path. Measurements on Westmere based systems show savings on "no wait" smp_call_function_single with idle target CPU (as measured on the sender side).
  local socket smp_call_func cost goes from ~1600 to ~1200 cycles remote socket smp_call_func cost goes from ~2000 to ~1800 cycles

Interesting that savings in the remote socket is less compared to the local socket.

Yes. I was not sure whether it has something to do with mwait/IPI wakeups work in hardware or it has something to do with C-state the target CPU is woken out of.

Quoted 15 lines

kernel/sched/core.c static inline void init_hrtick(void) · line 520

Quoted 10 lines
kernel/sched/core.c +9 -0
+int smp_need_ipi(int cpu)
+{
+     int oldval;
+
+     if (!system_using_cpu_idle_sync || cpu == smp_processor_id())
+             return 1;
+
+     oldval = atomic_cmpxchg(&per_cpu(cpu_idle_sync, cpu),
+                             CPU_STATE_IDLE, CPU_STATE_WAKING);

To avoid too many cache line bounces for the case when the cpu is in the running state, we should do a read to check if the state is in idle before going ahead with the locked operation?

Agree. Will add it on patch refresh.

Quoted 6 lines
Quoted 6 lines
kernel/sched/core.c +6 -0
+
+     if (oldval == CPU_STATE_RUNNING)
+             return 1;
+
+     if (oldval == CPU_STATE_IDLE) {
+             set_tsk_ipi_pending(idle_task(cpu));
Quoted 9 lines
Quoted 4 lines

+             atomic_set(&per_cpu(cpu_idle_sync, cpu), CPU_STATE_WOKENUP);
+     }
+
+     return 0;

We should probably disable interrupts around this, otherwise any delay in transitioning to wokenup from waking will cause the idle cpu to be stuck for similar amount of time.

Makes sense. Will add that too.

Thanks,
Venki

Quoted 3 lines

thanks,
suresh

#10 lore raw
In reply to Venkatesh Pallipadi
Quoted 4 lines · Venkatesh Pallipadi

smp_call_function_single and ttwu_queue_remote sends unconditional IPI to target CPU. However, if the target CPU is in mwait based idle, we can do IPI-less wakeups using the magical powers of monitor-mwait. Doing this has certain advantages:

Actually I'm trying to do the similar thing on MIPS.

The difference is that I want task_is_polling() to do something. The basic idea is:

Review

Quoted 8 lines
(quoted diff) +8 -0
+			if (ipi_pending()) {
+				clear_ipi_pending();
+				local_bh_disable();
+				local_irq_disable();
+				generic_smp_call_function_single_interrupt();
+				scheduler_wakeup_self_check();
+				local_irq_enable();
+				local_bh_enable();

I let cpu_idle() check if there is anything to do as your above code.

And task_is_polling() handle the others with below patch:
---

kernel/sched/core.c +5 -3
index 5255c9d..09f633d 100644
@@ -527,15 +527,16 @@ void resched_task(struct task_struct *p)
 		smp_send_reschedule(cpu);
 }
 
-void resched_cpu(int cpu)
+int resched_cpu(int cpu)
 {
 	struct rq *rq = cpu_rq(cpu);
 	unsigned long flags;
 
 	if (!raw_spin_trylock_irqsave(&rq->lock, flags))
-		return;
+		return 0;
 	resched_task(cpu_curr(cpu));
 	raw_spin_unlock_irqrestore(&rq->lock, flags);
+	return 1;
 }
 
 #ifdef CONFIG_NO_HZ
@@ -1484,7 +1485,8 @@ void scheduler_ipi(void)
 
 static void ttwu_queue_remote(struct task_struct *p, int cpu)
 {
-	if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
+	if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list) &&
+						!resched_cpu(cpu))
 		smp_send_reschedule(cpu);
 }
 

Thought?

Thanks,
Yong

#11 lore raw
Quoted 35 lines · Yong Zhang
Quoted 4 lines · Venkatesh Pallipadi

smp_call_function_single and ttwu_queue_remote sends unconditional IPI to target CPU. However, if the target CPU is in mwait based idle, we can do IPI-less wakeups using the magical powers of monitor-mwait. Doing this has certain advantages:

Actually I'm trying to do the similar thing on MIPS.

The difference is that I want task_is_polling() to do something. The basic idea is:

Quoted 8 lines

+                     if (ipi_pending()) {
+                             clear_ipi_pending();
+                             local_bh_disable();
+                             local_irq_disable();
+                             generic_smp_call_function_single_interrupt();
+                             scheduler_wakeup_self_check();
+                             local_irq_enable();
+                             local_bh_enable();

I let cpu_idle() check if there is anything to do as your above code.

And task_is_polling() handle the others with below patch:
---

kernel/sched/core.c +1 -1
index 5255c9d..09f633d 100644
@@ -527,15 +527,16 @@ void resched_task(struct task_struct *p)
                smp_send_reschedule(cpu);
  }
 
-void resched_cpu(int cpu)
+int resched_cpu(int cpu)
  {

kernel/sched/core.c void resched_task(struct task_struct *p) · line 540

Quoted 10 lines
kernel/sched/core.c +2 -1
        struct rq *rq = cpu_rq(cpu);
        unsigned long flags;
 
        if (!raw_spin_trylock_irqsave(&rq->lock, flags))
-               return;
+               return 0;
        resched_task(cpu_curr(cpu));
        raw_spin_unlock_irqrestore(&rq->lock, flags);
+       return 1;
  }

Two points -
rq->lock: I tried something similar first. One hurdle with checking task_is_polling() is that you need rq->lock to check it. And adding lock+unlock (without wait) in wakeup path ended up being no net gain compared to IPI. And when we actually end up spinning on that lock, thats going to add overhead in the common path. That is the reason I switched to atomic compare exchange and moving any wait onto the target CPU coming out of idle.

resched_task: ttwu_queue_remote() does not imply that the remote CPU will do a resched. Today there is a IPI and IPI handler calls onto check_preempt_wakeup() and if the current task has higher precedence than the waking up task, then there will be just an activation of new task and no resched. Using resched_task above breaks check_preempt_wakeup() and always calls a resched on remote CPU after the IPI, which would be change in behavior.

Thanks,
Venki

Quoted 16 lines
kernel/sched/core.c +2 -1
  #ifdef CONFIG_NO_HZ
@@ -1484,7 +1485,8 @@ void scheduler_ipi(void)
 
  static void ttwu_queue_remote(struct task_struct *p, int cpu)
  {
-       if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
+       if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list) &&
+                                               !resched_cpu(cpu))
                smp_send_reschedule(cpu);
  }
 

Thought?

Thanks,
Yong

#12 lore raw
Quoted 39 lines · Venki Pallipadi
Quoted 38 lines · Yong Zhang
Quoted 4 lines · Venkatesh Pallipadi

smp_call_function_single and ttwu_queue_remote sends unconditional IPI to target CPU. However, if the target CPU is in mwait based idle, we can do IPI-less wakeups using the magical powers of monitor-mwait. Doing this has certain advantages:

Actually I'm trying to do the similar thing on MIPS.

The difference is that I want task_is_polling() to do something. The basic idea is:

Quoted 8 lines

+ ? ? ? ? ? ? ? ? ? ? if (ipi_pending()) {
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? clear_ipi_pending();
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? local_bh_disable();
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? local_irq_disable();
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? generic_smp_call_function_single_interrupt();
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? scheduler_wakeup_self_check();
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? local_irq_enable();
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? local_bh_enable();

I let cpu_idle() check if there is anything to do as your above code.

And task_is_polling() handle the others with below patch:
---

kernel/sched/core.c +1 -1
index 5255c9d..09f633d 100644
@@ -527,15 +527,16 @@ void resched_task(struct task_struct *p)
 ? ? ? ? ? ? ? ?smp_send_reschedule(cpu);
 ?}
 
-void resched_cpu(int cpu)
+int resched_cpu(int cpu)
 ?{
 ? ? ? ?struct rq *rq = cpu_rq(cpu);
 ? ? ? ?unsigned long flags;
 
Quoted 8 lines

kernel/sched/core.c void resched_task(struct task_struct *p) · line 540

Quoted 7 lines
kernel/sched/core.c +2 -1
 ? ? ? ?if (!raw_spin_trylock_irqsave(&rq->lock, flags))
- ? ? ? ? ? ? ? return;
+ ? ? ? ? ? ? ? return 0;
 ? ? ? ?resched_task(cpu_curr(cpu));
 ? ? ? ?raw_spin_unlock_irqrestore(&rq->lock, flags);
+ ? ? ? return 1;
 ?}

I assume we are talking about 'return from idle' but seems I don't make it clear.

Quoted 8 lines

Two points -
rq->lock: I tried something similar first. One hurdle with checking task_is_polling() is that you need rq->lock to check it. And adding lock+unlock (without wait) in wakeup path ended up being no net gain compared to IPI. And when we actually end up spinning on that lock, thats going to add overhead in the common path. That is the reason I switched to atomic compare exchange and moving any wait onto the target CPU coming out of idle.

I see. But actually we will not spinning on that lock because we use 'trylock' in resched_cpu(). And you are right there is indeed a little overhead (resched_task()) if we hold the lock but it can be tolerated IMHO.

BTW, mind showing you test case thus we can collect some common data?

Quoted 8 lines

resched_task: ttwu_queue_remote() does not imply that the remote CPU will do a resched. Today there is a IPI and IPI handler calls onto check_preempt_wakeup() and if the current task has higher precedence than the waking up task, then there will be just an activation of new task and no resched. Using resched_task above breaks check_preempt_wakeup() and always calls a resched on remote CPU after the IPI, which would be change in behavior.

Yeah, if the remote cpu is not idle, mine will change the behavior; but if the remote cpu is idle, it will always rescheduled, right?

So maybe we could introduce resched_idle_cpu() to make things more clear:

int resched_idle_cpu(int cpu)

{
        struct rq *rq = cpu_rq(cpu);
        unsigned long flags;
        int ret = 0; 

        if (!raw_spin_trylock_irqsave(&rq->lock, flags))
                goto out; 
        if (!idle_cpu(cpu))
                goto out_unlock;
        resched_task(cpu_curr(cpu));
                ret = 1; 
out_unlock:
        raw_spin_unlock_irqrestore(&rq->lock, flags);
out:
        return ret; 
}

Thanks,
Yong

Quoted 25 lines

Thanks,
Venki

Quoted 14 lines

?#ifdef CONFIG_NO_HZ

kernel/sched/core.c +2 -1
@@ -1484,7 +1485,8 @@ void scheduler_ipi(void)
 
 ?static void ttwu_queue_remote(struct task_struct *p, int cpu)
 ?{
- ? ? ? if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
+ ? ? ? if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list) &&
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? !resched_cpu(cpu))
 ? ? ? ? ? ? ? ?smp_send_reschedule(cpu);
 ?}
 

Thought?

Quoted 2 lines

Thanks,
Yong

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Signature · 1 line
Only stand for myself
#13 lore raw
Quoted 61 lines · Yong Zhang
Quoted 39 lines · Venki Pallipadi
Quoted 38 lines · Yong Zhang
Quoted 4 lines · Venkatesh Pallipadi

smp_call_function_single and ttwu_queue_remote sends unconditional IPI to target CPU. However, if the target CPU is in mwait based idle, we can do IPI-less wakeups using the magical powers of monitor-mwait. Doing this has certain advantages:

Actually I'm trying to do the similar thing on MIPS.

The difference is that I want task_is_polling() to do something. The basic idea is:

Quoted 8 lines

+ ? ? ? ? ? ? ? ? ? ? if (ipi_pending()) {
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? clear_ipi_pending();
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? local_bh_disable();
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? local_irq_disable();
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? generic_smp_call_function_single_interrupt();
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? scheduler_wakeup_self_check();
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? local_irq_enable();
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? local_bh_enable();

I let cpu_idle() check if there is anything to do as your above code.

And task_is_polling() handle the others with below patch:
---

kernel/sched/core.c +1 -1
index 5255c9d..09f633d 100644
@@ -527,15 +527,16 @@ void resched_task(struct task_struct *p)
 ? ? ? ? ? ? ? ?smp_send_reschedule(cpu);
 ?}
 
-void resched_cpu(int cpu)
+int resched_cpu(int cpu)
 ?{
 ? ? ? ?struct rq *rq = cpu_rq(cpu);
 ? ? ? ?unsigned long flags;
 
Quoted 8 lines

kernel/sched/core.c void resched_task(struct task_struct *p) · line 540

Quoted 7 lines
kernel/sched/core.c +2 -1
 ? ? ? ?if (!raw_spin_trylock_irqsave(&rq->lock, flags))
- ? ? ? ? ? ? ? return;
+ ? ? ? ? ? ? ? return 0;
 ? ? ? ?resched_task(cpu_curr(cpu));
 ? ? ? ?raw_spin_unlock_irqrestore(&rq->lock, flags);
+ ? ? ? return 1;
 ?}

I assume we are talking about 'return from idle' but seems I don't make it clear.

Quoted 8 lines

Two points -
rq->lock: I tried something similar first. One hurdle with checking task_is_polling() is that you need rq->lock to check it. And adding lock+unlock (without wait) in wakeup path ended up being no net gain compared to IPI. And when we actually end up spinning on that lock, thats going to add overhead in the common path. That is the reason I switched to atomic compare exchange and moving any wait onto the target CPU coming out of idle.

Quoted 2 lines

I see. But actually we will not spinning on that lock because we use 'trylock' in resched_cpu().

Ahh. Sorry I missed the trylock in there...

Quoted 3 lines

And you are right there is indeed a
little overhead (resched_task()) if we hold the lock but it can be tolerated IMHO.

One advantage I got by using atomic stuff instead of rq->lock was as I mentioned in the patch description, if 2 CPUs are trying to send IPI to same target CPU around same time (50-100 us if CPU is in deep C-state in x86).

Quoted 2 lines

BTW, mind showing you test case thus we can collect some common data?

Test case was a silly clock measure around __smp_call_function_single() with optimization I had in generic_exec_single(). Attaching the patch I had..

Quoted 20 lines
Quoted 8 lines

resched_task: ttwu_queue_remote() does not imply that the remote CPU will do a resched. Today there is a IPI and IPI handler calls onto check_preempt_wakeup() and if the current task has higher precedence than the waking up task, then there will be just an activation of new task and no resched. Using resched_task above breaks check_preempt_wakeup() and always calls a resched on remote CPU after the IPI, which would be change in behavior.

Yeah, if the remote cpu is not idle, mine will change the behavior; but if the remote cpu is idle, it will always rescheduled, right?

So maybe we could introduce resched_idle_cpu() to make things more clear:

int resched_idle_cpu(int cpu)
{
       struct rq *rq = cpu_rq(cpu);
       unsigned long flags;
       int ret = 0;

Quoted 12 lines

       if (!raw_spin_trylock_irqsave(&rq->lock, flags)) goto out;
       if (!idle_cpu(cpu))
               goto out_unlock; resched_task(cpu_curr(cpu));
               ret = 1;
out_unlock:
       raw_spin_unlock_irqrestore(&rq->lock, flags); out:
       return ret;
}

This should likely work. But, if you do want to use similar logic in smp_call_function() or idle load balance kick etc, you need additional bit other than need_resched() as there we only need irq+softirq and not necessarily a resched.
At this time I am not sure how poll wakeup logic works in MIPS. But, if it is something that is similar to x86 mwait and we can wakeup with a bit other than TIF_NEED_RESCHED, we can generalize most of the changes in my RFC and share it across archs.

-Venki

Quoted 23 lines
Quoted 15 lines
Quoted 15 lines

?#ifdef CONFIG_NO_HZ

kernel/sched/core.c +2 -1
@@ -1484,7 +1485,8 @@ void scheduler_ipi(void)
 
 ?static void ttwu_queue_remote(struct task_struct *p, int cpu)
 ?{
- ? ? ? if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
+ ? ? ? if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list) &&
+ ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? !resched_cpu(cpu))
 ? ? ? ? ? ? ? ?smp_send_reschedule(cpu);
 ?}
 

Thought?

Quoted 7 lines
Quoted 2 lines

Thanks,
Yong

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Quoted 2 lines

--
Only stand for myself

From fd0f349bffdf61fda8a8085b435ec40d9ddfba33 Mon Sep 17 00:00:00 2001
From: Venkatesh Pallipadi <[email protected]>
Date: Wed, 1 Feb 2012 17:02:59 -0800
Subject: [PATCH 3/5] test: ipicost test routine

Silly test to measure ipicost.

$ taskset 0x1 cat /proc/ipicost; dmesg | tail -$(($(grep processor /proc/cpuinfo | wc -l)*2))
---
 fs/proc/Makefile  |    2 +-
 fs/proc/ipicost.c |  107 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 108 insertions(+), 1 deletions(-)
 create mode 100644 fs/proc/ipicost.c
fs/proc/Makefile +1 -1
index c1c7293..4407c6f 100644
@@ -8,7 +8,7 @@ proc-y			:= nommu.o task_nommu.o
 proc-$(CONFIG_MMU)	:= mmu.o task_mmu.o
 
 proc-y       += inode.o root.o base.o generic.o array.o \
-		proc_tty.o
+		proc_tty.o ipicost.o
 proc-y	+= cmdline.o
 proc-y	+= consoles.o
 proc-y	+= cpuinfo.o
fs/proc/ipicost.c new file +107 -0

107 lines not on this page · raw message

Signature · 1 line
1.7.7.3
#14 lore raw
Quoted 5 lines · Venki Pallipadi

This should likely work. But, if you do want to use similar logic in smp_call_function() or idle load balance kick etc, you need additional bit other than need_resched() as there we only need irq+softirq and not necessarily a resched.

Yeah, putting resched_idle_cpu() into smp_call_function_*() is a bit ugly though resched could work in the scenario because you have generic_smp_call_function_single_interrupt() called in cpu_idle();

Quoted 1 line

At this time I am not sure how poll wakeup logic works in MIPS. But,

A typical example is busy loop:
void __noreturn cpu_idle(void)

{
	...

        while (1) {
                tick_nohz_idle_enter();
                rcu_idle_enter();
                while (!need_resched() && cpu_online(cpu)) {
                	;
		}   
                rcu_idle_exit();
                tick_nohz_idle_exit();
                preempt_enable_no_resched();
                schedule();
		preempt_disable();
	}

	...
}
Quoted 3 lines

if it is something that is similar to x86 mwait and we can wakeup with a bit other than TIF_NEED_RESCHED, we can generalize most of the changes in my RFC and share it across archs.

Yeah, there are many things we could share IMHO, especially the hook in cpu_idle().

Thanks,
Yong

maintainer
#15 lore raw
In reply to Venkatesh Pallipadi
Quoted 3 lines · Venkatesh Pallipadi

smp_call_function_single and ttwu_queue_remote sends unconditional IPI to target CPU. However, if the target CPU is in mwait based idle, we can do IPI-less wakeups using the magical powers of monitor-mwait.

So I was thinking, why not change native_smp_send_reschedule() and native_send_call_func_single_ipi() and keep the change entirely inside the arch?

Ideally its only APIC/idle that know about this detail, the scheduler (or other consumers) really don't care about how the other cpu comes to run the callback.

#16 lore raw
Quoted 5 lines · Peter Zijlstra
Quoted 3 lines · Venkatesh Pallipadi

smp_call_function_single and ttwu_queue_remote sends unconditional IPI to target CPU. However, if the target CPU is in mwait based idle, we can do IPI-less wakeups using the magical powers of monitor-mwait.

Quoted 8 lines

So I was thinking, why not change native_smp_send_reschedule() and native_send_call_func_single_ipi() and keep the change entirely inside the arch?

Ideally its only APIC/idle that know about this detail, the scheduler (or other consumers) really don't care about how the other cpu comes to run the callback.

OK. Moving most of this into arch code will be cleaner. But, Yong mentioned in this thread that he was looking to do something similar on MIPS. So, we may end up with some code duplication though..

maintainer
#17 lore raw
Quoted 3 lines · Venki Pallipadi

OK. Moving most of this into arch code will be cleaner. But, Yong mentioned in this thread that he was looking to do something similar on MIPS. So, we may end up with some code duplication though..

If there's significant duplication you can always move a few helpers into lib/, but I think we shouldn't let this get outside of the PIC and idle code (both clearly arch specific), all we're doing is optimizing the IPI delivery, conceptually we're still sending one.

#18 lore raw
In reply to Venki Pallipadi
Quoted 10 lines · Venki Pallipadi
Quoted 9 lines · Peter Zijlstra
Quoted 3 lines · Venkatesh Pallipadi

smp_call_function_single and ttwu_queue_remote sends unconditional IPI to target CPU. However, if the target CPU is in mwait based idle, we can do IPI-less wakeups using the magical powers of monitor-mwait.

So I was thinking, why not change native_smp_send_reschedule() and native_send_call_func_single_ipi() and keep the change entirely inside the arch?

Quoted 8 lines
Quoted 4 lines

Ideally its only APIC/idle that know about this detail, the scheduler (or other consumers) really don't care about how the other cpu comes to run the callback.

OK. Moving most of this into arch code will be cleaner. But, Yong mentioned in this thread that he was looking to do something similar on MIPS. So, we may end up with some code duplication though..

Yeah, most of the things could be shared. Such as the added logic when return from idle and the hooks into ttwu_queue_remote() and generic_exec_single().

BTW, I think we should also let the arch provide a wrapper for TIF_IPI_PENDING, such as tsk_ipi_pending().

Thanks,
Yong

Quoted 5 lines

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Signature · 1 line
Only stand for myself
maintainer
#19 lore raw
Quoted 4 lines · Yong Zhang

Yeah, most of the things could be shared. Such as the added logic when return from idle and the hooks into ttwu_queue_remote() and generic_exec_single().

The thing is, those last two should not have hooks. Nor am I convinced anything outside of the arch idle loop should have any.

#20 lore raw
Quoted 7 lines · Peter Zijlstra
Quoted 4 lines · Yong Zhang

Yeah, most of the things could be shared. Such as the added logic when return from idle and the hooks into ttwu_queue_remote() and generic_exec_single().

The thing is, those last two should not have hooks.

Yeah, it's general place; and smp_need_ipi() plays well.

Quoted 2 lines

Nor am I convinced
anything outside of the arch idle loop should have any.

'idle loop' is the key point :)

Thanks,
Yong

#21 lore raw
In reply to Peter Zijlstra
Quoted 9 lines · Peter Zijlstra
Quoted 4 lines · Yong Zhang

Yeah, most of the things could be shared. Such as the added logic when return from idle and the hooks into ttwu_queue_remote() and generic_exec_single().

The thing is, those last two should not have hooks. Nor am I convinced anything outside of the arch idle loop should have any.

OK. Will put stuff in arch and respin soon..

-Venki