patches/preempt-lazy-support.patch

Subject: sched: Add support for lazy preemption
From: Thomas Gleixner <tglx@linutronix.de>
Date: Fri, 26 Oct 2012 18:50:54 +0100

It has become an obsession to mitigate the determinism vs. throughput
loss of RT. Looking at the mainline semantics of preemption points
gives a hint why RT sucks throughput wise for ordinary SCHED_OTHER
tasks. One major issue is the wakeup of tasks which are right away
preempting the waking task while the waking task holds a lock on which
the woken task will block right after having preempted the wakee. In
mainline this is prevented due to the implicit preemption disable of
spin/rw_lock held regions. On RT this is not possible due to the fully
preemptible nature of sleeping spinlocks.

Though for a SCHED_OTHER task preempting another SCHED_OTHER task this
is really not a correctness issue. RT folks are concerned about
SCHED_FIFO/RR tasks preemption and not about the purely fairness
driven SCHED_OTHER preemption latencies.

So I introduced a lazy preemption mechanism which only applies to
SCHED_OTHER tasks preempting another SCHED_OTHER task. Aside of the
existing preempt_count each tasks sports now a preempt_lazy_count
which is manipulated on lock acquiry and release. This is slightly
incorrect as for lazyness reasons I coupled this on
migrate_disable/enable so some other mechanisms get the same treatment
(e.g. get_cpu_light).

Now on the scheduler side instead of setting NEED_RESCHED this sets
NEED_RESCHED_LAZY in case of a SCHED_OTHER/SCHED_OTHER preemption and
therefor allows to exit the waking task the lock held region before
the woken task preempts. That also works better for cross CPU wakeups
as the other side can stay in the adaptive spinning loop.

For RT class preemption there is no change. This simply sets
NEED_RESCHED and forgoes the lazy preemption counter.

 Initial test do not expose any observable latency increasement, but
history shows that I've been proven wrong before :)

The lazy preemption mode is per default on, but with
CONFIG_SCHED_DEBUG enabled it can be disabled via:

 # echo NO_PREEMPT_LAZY >/sys/kernel/debug/sched_features

and reenabled via

 # echo PREEMPT_LAZY >/sys/kernel/debug/sched_features

The test results so far are very machine and workload dependent, but
there is a clear trend that it enhances the non RT workload
performance.

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
---
 include/linux/preempt.h      |   29 ++++++++++++++++-
 include/linux/sched.h        |   37 ++++++++++++++++++++++
 include/linux/thread_info.h  |   12 ++++++-
 include/linux/trace_events.h |    1 
 kernel/Kconfig.preempt       |    6 +++
 kernel/sched/core.c          |   72 +++++++++++++++++++++++++++++++++++++++++--
 kernel/sched/fair.c          |   16 ++++-----
 kernel/sched/features.h      |    3 +
 kernel/sched/sched.h         |    9 +++++
 kernel/trace/trace.c         |   37 +++++++++++++---------
 kernel/trace/trace.h         |    2 +
 kernel/trace/trace_output.c  |   14 +++++++-
 12 files changed, 209 insertions(+), 29 deletions(-)

--- a/include/linux/preempt.h
+++ b/include/linux/preempt.h
@@ -153,6 +153,20 @@ extern void preempt_count_sub(int val);
 #define preempt_count_inc() preempt_count_add(1)
 #define preempt_count_dec() preempt_count_sub(1)
 
+#ifdef CONFIG_PREEMPT_LAZY
+#define add_preempt_lazy_count(val)	do { preempt_lazy_count() += (val); } while (0)
+#define sub_preempt_lazy_count(val)	do { preempt_lazy_count() -= (val); } while (0)
+#define inc_preempt_lazy_count()	add_preempt_lazy_count(1)
+#define dec_preempt_lazy_count()	sub_preempt_lazy_count(1)
+#define preempt_lazy_count()		(current_thread_info()->preempt_lazy_count)
+#else
+#define add_preempt_lazy_count(val)	do { } while (0)
+#define sub_preempt_lazy_count(val)	do { } while (0)
+#define inc_preempt_lazy_count()	do { } while (0)
+#define dec_preempt_lazy_count()	do { } while (0)
+#define preempt_lazy_count()		(0)
+#endif
+
 #ifdef CONFIG_PREEMPT_COUNT
 
 #define preempt_disable() \
@@ -161,6 +175,12 @@ do { \
 	barrier(); \
 } while (0)
 
+#define preempt_lazy_disable() \
+do { \
+	inc_preempt_lazy_count(); \
+	barrier(); \
+} while (0)
+
 #define sched_preempt_enable_no_resched() \
 do { \
 	barrier(); \
@@ -198,6 +218,13 @@ do { \
 		__preempt_schedule(); \
 } while (0)
 
+#define preempt_lazy_enable() \
+do { \
+	dec_preempt_lazy_count(); \
+	barrier(); \
+	preempt_check_resched(); \
+} while (0)
+
 #else /* !CONFIG_PREEMPT */
 #define preempt_enable() \
 do { \
@@ -264,7 +291,7 @@ do { \
 } while (0)
 #define preempt_fold_need_resched() \
 do { \
-	if (tif_need_resched()) \
+	if (tif_need_resched_now()) \
 		set_preempt_need_resched(); \
 } while (0)
 
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -3342,6 +3342,43 @@ static inline int test_tsk_need_resched(
 	return unlikely(test_tsk_thread_flag(tsk,TIF_NEED_RESCHED));
 }
 
+#ifdef CONFIG_PREEMPT_LAZY
+static inline void set_tsk_need_resched_lazy(struct task_struct *tsk)
+{
+	set_tsk_thread_flag(tsk,TIF_NEED_RESCHED_LAZY);
+}
+
+static inline void clear_tsk_need_resched_lazy(struct task_struct *tsk)
+{
+	clear_tsk_thread_flag(tsk,TIF_NEED_RESCHED_LAZY);
+}
+
+static inline int test_tsk_need_resched_lazy(struct task_struct *tsk)
+{
+	return unlikely(test_tsk_thread_flag(tsk,TIF_NEED_RESCHED_LAZY));
+}
+
+static inline int need_resched_lazy(void)
+{
+	return test_thread_flag(TIF_NEED_RESCHED_LAZY);
+}
+
+static inline int need_resched_now(void)
+{
+	return test_thread_flag(TIF_NEED_RESCHED);
+}
+
+#else
+static inline void clear_tsk_need_resched_lazy(struct task_struct *tsk) { }
+static inline int need_resched_lazy(void) { return 0; }
+
+static inline int need_resched_now(void)
+{
+	return test_thread_flag(TIF_NEED_RESCHED);
+}
+
+#endif
+
 static inline int restart_syscall(void)
 {
 	set_tsk_thread_flag(current, TIF_SIGPENDING);
--- a/include/linux/thread_info.h
+++ b/include/linux/thread_info.h
@@ -107,7 +107,17 @@ static inline int test_ti_thread_flag(st
 #define test_thread_flag(flag) \
 	test_ti_thread_flag(current_thread_info(), flag)
 
-#define tif_need_resched() test_thread_flag(TIF_NEED_RESCHED)
+#ifdef CONFIG_PREEMPT_LAZY
+#define tif_need_resched()	(test_thread_flag(TIF_NEED_RESCHED) || \
+				 test_thread_flag(TIF_NEED_RESCHED_LAZY))
+#define tif_need_resched_now()	(test_thread_flag(TIF_NEED_RESCHED))
+#define tif_need_resched_lazy()	test_thread_flag(TIF_NEED_RESCHED_LAZY))
+
+#else
+#define tif_need_resched()	test_thread_flag(TIF_NEED_RESCHED)
+#define tif_need_resched_now()	test_thread_flag(TIF_NEED_RESCHED)
+#define tif_need_resched_lazy()	0
+#endif
 
 #ifndef CONFIG_HAVE_ARCH_WITHIN_STACK_FRAMES
 static inline int arch_within_stack_frames(const void * const stack,
--- a/include/linux/trace_events.h
+++ b/include/linux/trace_events.h
@@ -58,6 +58,7 @@ struct trace_entry {
 	int			pid;
 	unsigned short		migrate_disable;
 	unsigned short		padding;
+	unsigned char		preempt_lazy_count;
 };
 
 #define TRACE_EVENT_TYPE_MAX						\
--- a/kernel/Kconfig.preempt
+++ b/kernel/Kconfig.preempt
@@ -6,6 +6,12 @@ config PREEMPT_RT_BASE
 	bool
 	select PREEMPT
 
+config HAVE_PREEMPT_LAZY
+	bool
+
+config PREEMPT_LAZY
+	def_bool y if HAVE_PREEMPT_LAZY && PREEMPT_RT_FULL
+
 choice
 	prompt "Preemption Model"
 	default PREEMPT_NONE
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -510,6 +510,38 @@ void resched_curr(struct rq *rq)
 		trace_sched_wake_idle_without_ipi(cpu);
 }
 
+#ifdef CONFIG_PREEMPT_LAZY
+void resched_curr_lazy(struct rq *rq)
+{
+	struct task_struct *curr = rq->curr;
+	int cpu;
+
+	if (!sched_feat(PREEMPT_LAZY)) {
+		resched_curr(rq);
+		return;
+	}
+
+	lockdep_assert_held(&rq->lock);
+
+	if (test_tsk_need_resched(curr))
+		return;
+
+	if (test_tsk_need_resched_lazy(curr))
+		return;
+
+	set_tsk_need_resched_lazy(curr);
+
+	cpu = cpu_of(rq);
+	if (cpu == smp_processor_id())
+		return;
+
+	/* NEED_RESCHED_LAZY must be visible before we test polling */
+	smp_mb();
+	if (!tsk_is_polling(curr))
+		smp_send_reschedule(cpu);
+}
+#endif
+
 void resched_cpu(int cpu)
 {
 	struct rq *rq = cpu_rq(cpu);
@@ -2531,6 +2563,9 @@ int sched_fork(unsigned long clone_flags
 	p->on_cpu = 0;
 #endif
 	init_task_preempt_count(p);
+#ifdef CONFIG_HAVE_PREEMPT_LAZY
+	task_thread_info(p)->preempt_lazy_count = 0;
+#endif
 #ifdef CONFIG_SMP
 	plist_node_init(&p->pushable_tasks, MAX_PRIO);
 	RB_CLEAR_NODE(&p->pushable_dl_tasks);
@@ -3362,6 +3397,7 @@ void migrate_disable(void)
 	}
 
 	preempt_disable();
+	preempt_lazy_disable();
 	pin_current_cpu();
 	p->migrate_disable = 1;
 	preempt_enable();
@@ -3401,6 +3437,7 @@ void migrate_enable(void)
 
 	unpin_current_cpu();
 	preempt_enable();
+	preempt_lazy_enable();
 }
 EXPORT_SYMBOL(migrate_enable);
 #endif
@@ -3530,6 +3567,7 @@ static void __sched notrace __schedule(b
 
 	next = pick_next_task(rq, prev, cookie);
 	clear_tsk_need_resched(prev);
+	clear_tsk_need_resched_lazy(prev);
 	clear_preempt_need_resched();
 	rq->clock_skip_update = 0;
 
@@ -3675,6 +3713,30 @@ static void __sched notrace preempt_sche
 	} while (need_resched());
 }
 
+#ifdef CONFIG_PREEMPT_LAZY
+/*
+ * If TIF_NEED_RESCHED is then we allow to be scheduled away since this is
+ * set by a RT task. Oterwise we try to avoid beeing scheduled out as long as
+ * preempt_lazy_count counter >0.
+ */
+static __always_inline int preemptible_lazy(void)
+{
+	if (test_thread_flag(TIF_NEED_RESCHED))
+		return 1;
+	if (current_thread_info()->preempt_lazy_count)
+		return 0;
+	return 1;
+}
+
+#else
+
+static inline int preemptible_lazy(void)
+{
+	return 1;
+}
+
+#endif
+
 #ifdef CONFIG_PREEMPT
 /*
  * this is the entry point to schedule() from in-kernel preemption
@@ -3689,7 +3751,8 @@ asmlinkage __visible void __sched notrac
 	 */
 	if (likely(!preemptible()))
 		return;
-
+	if (!preemptible_lazy())
+		return;
 	preempt_schedule_common();
 }
 NOKPROBE_SYMBOL(preempt_schedule);
@@ -3716,6 +3779,9 @@ asmlinkage __visible void __sched notrac
 	if (likely(!preemptible()))
 		return;
 
+	if (!preemptible_lazy())
+		return;
+
 	do {
 		/*
 		 * Because the function tracer can trace preempt_count_sub()
@@ -5481,7 +5547,9 @@ void init_idle(struct task_struct *idle,
 
 	/* Set the preempt count _outside_ the spinlocks! */
 	init_idle_preempt_count(idle, cpu);
-
+#ifdef CONFIG_HAVE_PREEMPT_LAZY
+	task_thread_info(idle)->preempt_lazy_count = 0;
+#endif
 	/*
 	 * The idle tasks have their own, simple scheduling class:
 	 */
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -3518,7 +3518,7 @@ check_preempt_tick(struct cfs_rq *cfs_rq
 	ideal_runtime = sched_slice(cfs_rq, curr);
 	delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
 	if (delta_exec > ideal_runtime) {
-		resched_curr(rq_of(cfs_rq));
+		resched_curr_lazy(rq_of(cfs_rq));
 		/*
 		 * The current task ran long enough, ensure it doesn't get
 		 * re-elected due to buddy favours.
@@ -3542,7 +3542,7 @@ check_preempt_tick(struct cfs_rq *cfs_rq
 		return;
 
 	if (delta > ideal_runtime)
-		resched_curr(rq_of(cfs_rq));
+		resched_curr_lazy(rq_of(cfs_rq));
 }
 
 static void
@@ -3684,7 +3684,7 @@ entity_tick(struct cfs_rq *cfs_rq, struc
 	 * validating it and just reschedule.
 	 */
 	if (queued) {
-		resched_curr(rq_of(cfs_rq));
+		resched_curr_lazy(rq_of(cfs_rq));
 		return;
 	}
 	/*
@@ -3866,7 +3866,7 @@ static void __account_cfs_rq_runtime(str
 	 * hierarchy can be throttled
 	 */
 	if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
-		resched_curr(rq_of(cfs_rq));
+		resched_curr_lazy(rq_of(cfs_rq));
 }
 
 static __always_inline
@@ -4494,7 +4494,7 @@ static void hrtick_start_fair(struct rq
 
 		if (delta < 0) {
 			if (rq->curr == p)
-				resched_curr(rq);
+				resched_curr_lazy(rq);
 			return;
 		}
 		hrtick_start(rq, delta);
@@ -5905,7 +5905,7 @@ static void check_preempt_wakeup(struct
 	return;
 
 preempt:
-	resched_curr(rq);
+	resched_curr_lazy(rq);
 	/*
 	 * Only set the backward buddy when the current task is still
 	 * on the rq. This can happen when a wakeup gets interleaved
@@ -8631,7 +8631,7 @@ static void task_fork_fair(struct task_s
 		 * 'current' within the tree based on its new key value.
 		 */
 		swap(curr->vruntime, se->vruntime);
-		resched_curr(rq);
+		resched_curr_lazy(rq);
 	}
 
 	se->vruntime -= cfs_rq->min_vruntime;
@@ -8655,7 +8655,7 @@ prio_changed_fair(struct rq *rq, struct
 	 */
 	if (rq->curr == p) {
 		if (p->prio > oldprio)
-			resched_curr(rq);
+			resched_curr_lazy(rq);
 	} else
 		check_preempt_curr(rq, p, 0);
 }
--- a/kernel/sched/features.h
+++ b/kernel/sched/features.h
@@ -47,6 +47,9 @@ SCHED_FEAT(NONTASK_CAPACITY, true)
 
 #ifdef CONFIG_PREEMPT_RT_FULL
 SCHED_FEAT(TTWU_QUEUE, false)
+# ifdef CONFIG_PREEMPT_LAZY
+SCHED_FEAT(PREEMPT_LAZY, true)
+# endif
 #else
 
 /*
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -1347,6 +1347,15 @@ extern void init_sched_fair_class(void);
 extern void resched_curr(struct rq *rq);
 extern void resched_cpu(int cpu);
 
+#ifdef CONFIG_PREEMPT_LAZY
+extern void resched_curr_lazy(struct rq *rq);
+#else
+static inline void resched_curr_lazy(struct rq *rq)
+{
+	resched_curr(rq);
+}
+#endif
+
 extern struct rt_bandwidth def_rt_bandwidth;
 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
 
--- a/kernel/trace/trace.c
+++ b/kernel/trace/trace.c
@@ -1897,6 +1897,7 @@ tracing_generic_entry_update(struct trac
 	struct task_struct *tsk = current;
 
 	entry->preempt_count		= pc & 0xff;
+	entry->preempt_lazy_count	= preempt_lazy_count();
 	entry->pid			= (tsk) ? tsk->pid : 0;
 	entry->flags =
 #ifdef CONFIG_TRACE_IRQFLAGS_SUPPORT
@@ -1907,7 +1908,8 @@ tracing_generic_entry_update(struct trac
 		((pc & NMI_MASK    ) ? TRACE_FLAG_NMI     : 0) |
 		((pc & HARDIRQ_MASK) ? TRACE_FLAG_HARDIRQ : 0) |
 		((pc & SOFTIRQ_MASK) ? TRACE_FLAG_SOFTIRQ : 0) |
-		(tif_need_resched() ? TRACE_FLAG_NEED_RESCHED : 0) |
+		(tif_need_resched_now() ? TRACE_FLAG_NEED_RESCHED : 0) |
+		(need_resched_lazy() ? TRACE_FLAG_NEED_RESCHED_LAZY : 0) |
 		(test_preempt_need_resched() ? TRACE_FLAG_PREEMPT_RESCHED : 0);
 
 	entry->migrate_disable = (tsk) ? __migrate_disabled(tsk) & 0xFF : 0;
@@ -2894,15 +2896,17 @@ get_total_entries(struct trace_buffer *b
 
 static void print_lat_help_header(struct seq_file *m)
 {
-	seq_puts(m, "#                  _------=> CPU#            \n"
-		    "#                 / _-----=> irqs-off        \n"
-		    "#                | / _----=> need-resched    \n"
-		    "#                || / _---=> hardirq/softirq \n"
-		    "#                ||| / _--=> preempt-depth   \n"
-		    "#                |||| / _--=> migrate-disable\n"
-		    "#                ||||| /     delay           \n"
-		    "#  cmd     pid   |||||| time  |   caller     \n"
-		    "#     \\   /      |||||  \\   |   /          \n");
+	seq_puts(m, "#                   _--------=> CPU#              \n"
+		    "#                  / _-------=> irqs-off          \n"
+		    "#                 | / _------=> need-resched      \n"
+		    "#                 || / _-----=> need-resched_lazy \n"
+		    "#                 ||| / _----=> hardirq/softirq   \n"
+		    "#                 |||| / _---=> preempt-depth     \n"
+		    "#                 ||||| / _--=> preempt-lazy-depth\n"
+		    "#                 |||||| / _-=> migrate-disable   \n"
+		    "#                 ||||||| /     delay             \n"
+		    "#  cmd     pid    |||||||| time  |   caller       \n"
+		    "#     \\   /      ||||||||  \\   |   /            \n");
 }
 
 static void print_event_info(struct trace_buffer *buf, struct seq_file *m)
@@ -2928,11 +2932,14 @@ static void print_func_help_header_irq(s
 	print_event_info(buf, m);
 	seq_puts(m, "#                              _-----=> irqs-off\n"
 		    "#                             / _----=> need-resched\n"
-		    "#                            | / _---=> hardirq/softirq\n"
-		    "#                            || / _--=> preempt-depth\n"
-		    "#                            ||| /     delay\n"
-		    "#           TASK-PID   CPU#  ||||    TIMESTAMP  FUNCTION\n"
-		    "#              | |       |   ||||       |         |\n");
+		    "#                            |/  _-----=> need-resched_lazy\n"
+		    "#                            || / _---=> hardirq/softirq\n"
+		    "#                            ||| / _--=> preempt-depth\n"
+		    "#                            |||| /_--=> preempt-lazy-depth\n"
+		    "#                            |||||  _-=> migrate-disable   \n"
+		    "#                            ||||| /    delay\n"
+		    "#           TASK-PID   CPU#  ||||||    TIMESTAMP  FUNCTION\n"
+		    "#              | |       |   ||||||       |         |\n");
 }
 
 void
--- a/kernel/trace/trace.h
+++ b/kernel/trace/trace.h
@@ -124,6 +124,7 @@ struct kretprobe_trace_entry_head {
  *  NEED_RESCHED	- reschedule is requested
  *  HARDIRQ		- inside an interrupt handler
  *  SOFTIRQ		- inside a softirq handler
+ *  NEED_RESCHED_LAZY	- lazy reschedule is requested
  */
 enum trace_flag_type {
 	TRACE_FLAG_IRQS_OFF		= 0x01,
@@ -133,6 +134,7 @@ enum trace_flag_type {
 	TRACE_FLAG_SOFTIRQ		= 0x10,
 	TRACE_FLAG_PREEMPT_RESCHED	= 0x20,
 	TRACE_FLAG_NMI			= 0x40,
+	TRACE_FLAG_NEED_RESCHED_LAZY	= 0x80,
 };
 
 #define TRACE_BUF_SIZE		1024
--- a/kernel/trace/trace_output.c
+++ b/kernel/trace/trace_output.c
@@ -386,6 +386,7 @@ int trace_print_lat_fmt(struct trace_seq
 {
 	char hardsoft_irq;
 	char need_resched;
+	char need_resched_lazy;
 	char irqs_off;
 	int hardirq;
 	int softirq;
@@ -416,6 +417,9 @@ int trace_print_lat_fmt(struct trace_seq
 		break;
 	}
 
+	need_resched_lazy =
+		(entry->flags & TRACE_FLAG_NEED_RESCHED_LAZY) ? 'L' : '.';
+
 	hardsoft_irq =
 		(nmi && hardirq)     ? 'Z' :
 		nmi                  ? 'z' :
@@ -424,14 +428,20 @@ int trace_print_lat_fmt(struct trace_seq
 		softirq              ? 's' :
 		                       '.' ;
 
-	trace_seq_printf(s, "%c%c%c",
-			 irqs_off, need_resched, hardsoft_irq);
+	trace_seq_printf(s, "%c%c%c%c",
+			 irqs_off, need_resched, need_resched_lazy,
+			 hardsoft_irq);
 
 	if (entry->preempt_count)
 		trace_seq_printf(s, "%x", entry->preempt_count);
 	else
 		trace_seq_putc(s, '.');
 
+	if (entry->preempt_lazy_count)
+		trace_seq_printf(s, "%x", entry->preempt_lazy_count);
+	else
+		trace_seq_putc(s, '.');
+
 	if (entry->migrate_disable)
 		trace_seq_printf(s, "%x", entry->migrate_disable);
 	else