[patch 1/6] preempt_lazy: remove redundant test_thread_flag() call

[patch 1/6] preempt_lazy: remove redundant test_thread_flag() call

[Mike Galbraith]

__preempt_count_dec_and_test() already does it.

Signed-off-by: Mike Galbraith <umgwanakikbuti@gmail.com>
---
 include/linux/preempt.h |    3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

--- a/include/linux/preempt.h
+++ b/include/linux/preempt.h
@@ -126,8 +126,7 @@ do { \
 #define preempt_enable_notrace() \
 do { \
 	barrier(); \
-	if (unlikely(__preempt_count_dec_and_test() || \
-				test_thread_flag(TIF_NEED_RESCHED_LAZY))) \
+	if (unlikely(__preempt_count_dec_and_test())) \
 		__preempt_schedule_context(); \
 } while (0)
 #else

Re: [patch 1/6] preempt_lazy: remove redundant test_thread_flag() call

[Steven Rostedt]

On Fri, 02 May 2014 13:12:46 +0200
Mike Galbraith <umgwanakikbuti@gmail.com> wrote:

> __preempt_count_dec_and_test() already does it.

Already does what? :-)

Hmm, do all archs check? Perhaps we need to also update
include/asm-generic/preempt.h?

-- Steve

> 
> Signed-off-by: Mike Galbraith <umgwanakikbuti@gmail.com>
> ---
>  include/linux/preempt.h |    3 +--
>  1 file changed, 1 insertion(+), 2 deletions(-)
> 
> --- a/include/linux/preempt.h
> +++ b/include/linux/preempt.h
> @@ -126,8 +126,7 @@ do { \
>  #define preempt_enable_notrace() \
>  do { \
>  	barrier(); \
> -	if (unlikely(__preempt_count_dec_and_test() || \
> -				test_thread_flag(TIF_NEED_RESCHED_LAZY))) \
> +	if (unlikely(__preempt_count_dec_and_test())) \
>  		__preempt_schedule_context(); \
>  } while (0)
>  #else
> 
> 

Re: [patch 1/6] preempt_lazy: remove redundant test_thread_flag() call

[Mike Galbraith]

On Mon, 2014-05-05 at 22:37 -0400, Steven Rostedt wrote: 
> On Fri, 02 May 2014 13:12:46 +0200
> Mike Galbraith <umgwanakikbuti@gmail.com> wrote:
> 
> > __preempt_count_dec_and_test() already does it.
> 
> Already does what? :-)

Poo.. yeah that.

> Hmm, do all archs check? Perhaps we need to also update
> include/asm-generic/preempt.h?

Per cpu preempt_count is x86, and tif_need_resched() handles laziness,
so all should be well.  Did I just manage to overlook the obvious for
the second time?

-Mike

Re: [patch 1/6] preempt_lazy: remove redundant test_thread_flag() call

[Sebastian Andrzej Siewior]

* Mike Galbraith | 2014-05-06 07:33:51 [+0200]:

>> Hmm, do all archs check? Perhaps we need to also update
>> include/asm-generic/preempt.h?
>
>Per cpu preempt_count is x86, and tif_need_resched() handles laziness,
>so all should be well.  Did I just manage to overlook the obvious for
>the second time?

In include/asm-generic/preempt.h we have
| static __always_inline bool __preempt_count_dec_and_test(void)
| {
|         return !--*preempt_count_ptr() && tif_need_resched();
| }

and tif_need_resched() is in include/linux/thread_info.h
| #ifdef CONFIG_PREEMPT_LAZY
| #define tif_need_resched() (test_thread_flag(TIF_NEED_RESCHED) || \
|                 test_thread_flag(TIF_NEED_RESCHED_LAZY))
| #else
| #define tif_need_resched() (test_thread_flag(TIF_NEED_RESCHED))
| #endif  

So I guess we are fine.

>-Mike

Sebastian

[patch 2/6] x86/preempt-lazy adapt implementation to percpu __preempt_count

[Mike Galbraith]

Signed-off-by: Mike Galbraith <umgwanakikbuti@gmail.com>
---
 arch/x86/include/asm/preempt.h |    8 ++++++++
 arch/x86/kernel/asm-offsets.c  |    1 +
 arch/x86/kernel/entry_32.S     |    9 ++++++---
 arch/x86/kernel/entry_64.S     |    7 +++++--
 4 files changed, 20 insertions(+), 5 deletions(-)

--- a/arch/x86/include/asm/preempt.h
+++ b/arch/x86/include/asm/preempt.h
@@ -94,7 +94,11 @@ static __always_inline bool __preempt_co
 {
 	if (____preempt_count_dec_and_test())
 		return true;
+#ifdef CONFIG_PREEMPT_LAZY
 	return test_thread_flag(TIF_NEED_RESCHED_LAZY);
+#else
+	return false;
+#endif
 }
 
 /*
@@ -102,8 +106,12 @@ static __always_inline bool __preempt_co
  */
 static __always_inline bool should_resched(void)
 {
+#ifdef CONFIG_PREEMPT_LAZY
 	return unlikely(!__this_cpu_read_4(__preempt_count) || \
 			test_thread_flag(TIF_NEED_RESCHED_LAZY));
+#else
+	return unlikely(!__this_cpu_read_4(__preempt_count));
+#endif
 }
 
 #ifdef CONFIG_PREEMPT
--- a/arch/x86/kernel/asm-offsets.c
+++ b/arch/x86/kernel/asm-offsets.c
@@ -72,4 +72,5 @@ void common(void) {
 
 	BLANK();
 	DEFINE(PTREGS_SIZE, sizeof(struct pt_regs));
+	DEFINE(_PREEMPT_ENABLED, PREEMPT_ENABLED);
 }
--- a/arch/x86/kernel/entry_32.S
+++ b/arch/x86/kernel/entry_32.S
@@ -365,19 +365,22 @@ ENTRY(resume_kernel)
 need_resched:
 	# preempt count == 0 + NEED_RS set?
 	cmpl $0,PER_CPU_VAR(__preempt_count)
+#ifndef CONFIG_PREEMPT_LAZY
+	jnz restore_all
+#else
 	jz test_int_off
 
 	# atleast preempt count == 0 ?
-	cmpl $_TIF_NEED_RESCHED,PER_CPU_VAR(__preempt_count)
+	cmpl $_PREEMPT_ENABLED,PER_CPU_VAR(__preempt_count)
 	jne restore_all
 
 	cmpl $0,TI_preempt_lazy_count(%ebp)	# non-zero preempt_lazy_count ?
 	jnz restore_all
 
-	testl $_TIF_NEED_RESCHED_LAZY, %ecx
+	testl $_TIF_NEED_RESCHED_LAZY, TI_flags(%ebp)
 	jz restore_all
-
 test_int_off:
+#endif
 	testl $X86_EFLAGS_IF,PT_EFLAGS(%esp)	# interrupts off (exception path) ?
 	jz restore_all
 	call preempt_schedule_irq
--- a/arch/x86/kernel/entry_64.S
+++ b/arch/x86/kernel/entry_64.S
@@ -1104,10 +1104,13 @@ ENTRY(native_iret)
 	/* rcx:	 threadinfo. interrupts off. */
 ENTRY(retint_kernel)
 	cmpl $0,PER_CPU_VAR(__preempt_count)
+#ifndef CONFIG_PREEMPT_LAZY
+	jnz  retint_restore_args
+#else
 	jz  check_int_off
 
 	# atleast preempt count == 0 ?
-	cmpl $_TIF_NEED_RESCHED,PER_CPU_VAR(__preempt_count)
+	cmpl $_PREEMPT_ENABLED,PER_CPU_VAR(__preempt_count)
 	jnz retint_restore_args
 
 	cmpl $0, TI_preempt_lazy_count(%rcx)
@@ -1115,8 +1118,8 @@ ENTRY(retint_kernel)
 
 	bt $TIF_NEED_RESCHED_LAZY,TI_flags(%rcx)
 	jnc  retint_restore_args

[patch 3/6] stomp-machine: fix typo in kernel/cpu.c

[Mike Galbraith]

Signed-off-by: Mike Galbraith <umgwanakikbuti@gmail.com>
---
 kernel/cpu.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

--- a/kernel/cpu.c
+++ b/kernel/cpu.c
@@ -649,7 +649,7 @@ static int __ref _cpu_down(unsigned int
 		/* CPU didn't die: tell everyone.  Can't complain. */
 		smpboot_unpark_threads(cpu);
 		cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
-		goto out_cancel;
+		goto out_release;
 	}
 	BUG_ON(cpu_online(cpu));
 

[patch 4/6] stomp-machine: fix typo in kernel/stop_machine.c

[Mike Galbraith]

Signed-off-by: Mike Galbraith <umgwanakikbuti@gmail.com>
---
 kernel/stop_machine.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

--- a/kernel/stop_machine.c
+++ b/kernel/stop_machine.c
@@ -671,7 +671,7 @@ int stop_machine_from_inactive_cpu(int (
 	ret = multi_cpu_stop(&msdata);
 
 	/* Busy wait for completion. */
-	while (!atomic_read(&done.nr_todo))
+	while (atomic_read(&done.nr_todo))
 		cpu_relax();
 
 	mutex_unlock(&stop_cpus_mutex);

[patch 5/6] stomp-machine: create lg_global_trylock_relax() primitive

[Mike Galbraith]

Create lg_global_trylock_relax() for use by stopper thread when it cannot
schedule, to deal with stop_cpus_lock, which is now an lglock.

Signed-off-by: Mike Galbraith <umgwanakikbuti@gmail.com>
---
 include/linux/lglock.h      |    6 ++++++
 include/linux/spinlock_rt.h |    1 +
 kernel/locking/lglock.c     |   25 +++++++++++++++++++++++++
 kernel/locking/rtmutex.c    |    5 +++++
 4 files changed, 37 insertions(+)

--- a/include/linux/lglock.h
+++ b/include/linux/lglock.h
@@ -74,4 +74,10 @@ void lg_local_unlock_cpu(struct lglock *
 void lg_global_lock(struct lglock *lg);
 void lg_global_unlock(struct lglock *lg);
 
+#ifndef CONFIG_PREEMPT_RT_FULL
+#define lg_global_trylock_relax(name)	lg_global_lock(name)
+#else
+void lg_global_trylock_relax(struct lglock *lg);
+#endif
+
 #endif
--- a/include/linux/spinlock_rt.h
+++ b/include/linux/spinlock_rt.h
@@ -35,6 +35,7 @@ extern int atomic_dec_and_spin_lock(atom
  */
 extern void __lockfunc __rt_spin_lock(struct rt_mutex *lock);
 extern void __lockfunc __rt_spin_unlock(struct rt_mutex *lock);
+extern int __lockfunc __rt_spin_trylock(struct rt_mutex *lock);
 
 #define spin_lock(lock)				\
 	do {					\
--- a/kernel/locking/lglock.c
+++ b/kernel/locking/lglock.c
@@ -105,3 +105,28 @@ void lg_global_unlock(struct lglock *lg)
 	preempt_enable_nort();
 }
 EXPORT_SYMBOL(lg_global_unlock);
+
+#ifdef CONFIG_PREEMPT_RT_FULL
+/*
+ * HACK: If you use this, you get to keep the pieces.
+ * Used in queue_stop_cpus_work() when stop machinery
+ * is called from inactive CPU, so we can't schedule.
+ */
+# define lg_do_trylock_relax(l)			\
+	do {					\
+		while (!__rt_spin_trylock(l))	\
+			cpu_relax();		\
+	} while (0)
+
+void lg_global_trylock_relax(struct lglock *lg)
+{
+	int i;
+
+	lock_acquire_exclusive(&lg->lock_dep_map, 0, 0, NULL, _RET_IP_);
+	for_each_possible_cpu(i) {
+		lg_lock_ptr *lock;
+		lock = per_cpu_ptr(lg->lock, i);
+		lg_do_trylock_relax(lock);
+	}
+}
+#endif
--- a/kernel/locking/rtmutex.c
+++ b/kernel/locking/rtmutex.c
@@ -1001,6 +1001,11 @@ void __lockfunc rt_spin_unlock_wait(spin
 }
 EXPORT_SYMBOL(rt_spin_unlock_wait);
 
+int __lockfunc __rt_spin_trylock(struct rt_mutex *lock)
+{
+	return rt_mutex_trylock(lock);
+}
+
 int __lockfunc rt_spin_trylock(spinlock_t *lock)
 {
 	int ret = rt_mutex_trylock(&lock->lock);

[patch 6/6] stomp-machine: use lg_global_trylock_relax() to dead with stop_cpus_lock lglock

[Mike Galbraith]

If the stop machinery is called from inactive CPU we cannot use
lg_global_lock(), because some other stomp machine invocation might be
in progress and the lock can be contended.  We cannot schedule from this
context, so use the lovely new lg_global_trylock_relax() primitive to
do what we used to do via one mutex_trylock()/cpu_relax() loop.  We
now do that trylock()/relax() across an entire herd of locks. Joy.

Signed-off-by: Mike Galbraith <umgwanakikbuti@gmail.com>
---
 kernel/stop_machine.c |   24 ++++++++++++++----------
 1 file changed, 14 insertions(+), 10 deletions(-)

--- a/kernel/stop_machine.c
+++ b/kernel/stop_machine.c
@@ -266,7 +266,7 @@ int stop_two_cpus(unsigned int cpu1, uns
 	struct irq_cpu_stop_queue_work_info call_args;
 	struct multi_stop_data msdata;
 
-	preempt_disable();
+	preempt_disable_nort();
 	msdata = (struct multi_stop_data){
 		.fn = fn,
 		.data = arg,
@@ -299,7 +299,7 @@ int stop_two_cpus(unsigned int cpu1, uns
 	 * This relies on the stopper workqueues to be FIFO.
 	 */
 	if (!cpu_active(cpu1) || !cpu_active(cpu2)) {
-		preempt_enable();
+		preempt_enable_nort();
 		return -ENOENT;
 	}
 
@@ -313,7 +313,7 @@ int stop_two_cpus(unsigned int cpu1, uns
 				 &irq_cpu_stop_queue_work,
 				 &call_args, 1);
 	lg_local_unlock(&stop_cpus_lock);
-	preempt_enable();
+	preempt_enable_nort();
 
 	wait_for_stop_done(&done);
 
@@ -346,7 +346,7 @@ static DEFINE_PER_CPU(struct cpu_stop_wo
 
 static void queue_stop_cpus_work(const struct cpumask *cpumask,
 				 cpu_stop_fn_t fn, void *arg,
-				 struct cpu_stop_done *done)
+				 struct cpu_stop_done *done, bool inactive)
 {
 	struct cpu_stop_work *work;
 	unsigned int cpu;
@@ -360,11 +360,13 @@ static void queue_stop_cpus_work(const s
 	}
 
 	/*
-	 * Disable preemption while queueing to avoid getting
-	 * preempted by a stopper which might wait for other stoppers
-	 * to enter @fn which can lead to deadlock.
+	 * Make sure that all work is queued on all cpus before
+	 * any of the cpus can execute it.
 	 */
-	lg_global_lock(&stop_cpus_lock);
+	if (!inactive)
+		lg_global_lock(&stop_cpus_lock);
+	else
+		lg_global_trylock_relax(&stop_cpus_lock);
 	for_each_cpu(cpu, cpumask)
 		cpu_stop_queue_work(cpu, &per_cpu(stop_cpus_work, cpu));
 	lg_global_unlock(&stop_cpus_lock);
@@ -376,7 +378,7 @@ static int __stop_cpus(const struct cpum
 	struct cpu_stop_done done;
 
 	cpu_stop_init_done(&done, cpumask_weight(cpumask));
-	queue_stop_cpus_work(cpumask, fn, arg, &done);
+	queue_stop_cpus_work(cpumask, fn, arg, &done, false);
 	wait_for_stop_done(&done);
 	return done.executed ? done.ret : -ENOENT;
 }
@@ -572,6 +574,8 @@ static int __init cpu_stop_init(void)
 		INIT_LIST_HEAD(&stopper->works);
 	}
 
+	lg_lock_init(&stop_cpus_lock, "stop_cpus_lock");
+
 	BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
 	stop_machine_initialized = true;
 	return 0;
@@ -667,7 +671,7 @@ int stop_machine_from_inactive_cpu(int (
 	set_state(&msdata, MULTI_STOP_PREPARE);
 	cpu_stop_init_done(&done, num_active_cpus());
 	queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
-			     &done);
+			     &done, true);
 	ret = multi_cpu_stop(&msdata);
 
 	/* Busy wait for completion. */