sched: Add support for lazy preemption

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

@@ -174,6 +174,20 @@ 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() \
@@ -182,6 +196,12 @@ do { \
 	barrier(); \
 } while (0)
 
+#define preempt_lazy_disable() \
+do { \
+	inc_preempt_lazy_count(); \
+	barrier(); \
+} while (0)
+
 #define sched_preempt_enable_no_resched() \
 do { \
 	barrier(); \
@@ -219,13 +239,31 @@ do { \
 		__preempt_schedule(); \
 } while (0)
 
+/*
+ * open code preempt_check_resched() because it is not exported to modules and
+ * used by local_unlock() or bpf_enable_instrumentation().
+ */
+#define preempt_lazy_enable() \
+do { \
+	dec_preempt_lazy_count(); \
+	barrier(); \
+	if (should_resched(0)) \
+		__preempt_schedule(); \
+} while (0)
+
 #else /* !CONFIG_PREEMPTION */
 #define preempt_enable() \
 do { \
 	barrier(); \
 	preempt_count_dec(); \
 } while (0)
 
+#define preempt_lazy_enable() \
+do { \
+	dec_preempt_lazy_count(); \
+	barrier(); \
+} while (0)
+
 #define preempt_enable_notrace() \
 do { \
 	barrier(); \
@@ -267,6 +305,9 @@ do { \
 #define preempt_check_resched_rt()		barrier()
 #define preemptible()				0
 
+#define preempt_lazy_disable()			barrier()
+#define preempt_lazy_enable()			barrier()
+
 #endif /* CONFIG_PREEMPT_COUNT */
 
 #ifdef MODULE
@@ -285,7 +326,7 @@ do { \
 } while (0)
 #define preempt_fold_need_resched() \
 do { \
-	if (tif_need_resched()) \
+	if (tif_need_resched_now()) \
 		set_preempt_need_resched(); \
 } while (0)
 
@@ -413,8 +454,15 @@ extern void migrate_enable(void);
 
 #else
 
-static inline void migrate_disable(void) { }
-static inline void migrate_enable(void) { }
+static inline void migrate_disable(void)
+{
+	preempt_lazy_disable();
+}
+
+static inline void migrate_enable(void)
+{
+	preempt_lazy_enable();
+}
 
 #endif /* CONFIG_SMP */
 

include/linux/sched.h

@@ -2014,6 +2014,43 @@ static inline int test_tsk_need_resched(struct task_struct *tsk)
 	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
+
 #ifdef CONFIG_PREEMPT_RT
 static inline bool task_match_saved_state(struct task_struct *p, long match_state)
 {

include/linux/thread_info.h

@@ -163,7 +163,17 @@ static inline int test_ti_thread_flag(struct thread_info *ti, int flag)
 	clear_ti_thread_flag(task_thread_info(t), TIF_##fl)
 #endif /* !CONFIG_GENERIC_ENTRY */
 
-#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,

include/linux/trace_events.h

@@ -69,6 +69,7 @@ struct trace_entry {
 	unsigned char		flags;
 	unsigned char		preempt_count;
 	int			pid;
+	unsigned char		preempt_lazy_count;
 };
 
 #define TRACE_EVENT_TYPE_MAX						\
@@ -157,9 +158,10 @@ static inline void tracing_generic_entry_update(struct trace_entry *entry,
 						unsigned int trace_ctx)
 {
 	entry->preempt_count		= trace_ctx & 0xff;
+	entry->preempt_lazy_count	= (trace_ctx >> 16) & 0xff;
 	entry->pid			= current->pid;
 	entry->type			= type;
-	entry->flags =			trace_ctx >> 16;
+	entry->flags			= trace_ctx >> 24;
 }
 
 unsigned int tracing_gen_ctx_irq_test(unsigned int irqs_status);
@@ -172,6 +174,7 @@ enum trace_flag_type {
 	TRACE_FLAG_SOFTIRQ		= 0x10,
 	TRACE_FLAG_PREEMPT_RESCHED	= 0x20,
 	TRACE_FLAG_NMI			= 0x40,
+	TRACE_FLAG_NEED_RESCHED_LAZY	= 0x80,
 };
 
 #ifdef CONFIG_TRACE_IRQFLAGS_SUPPORT

kernel/Kconfig.preempt

@@ -1,5 +1,11 @@
 # SPDX-License-Identifier: GPL-2.0-only
 
+config HAVE_PREEMPT_LAZY
+	bool
+
+config PREEMPT_LAZY
+	def_bool y if HAVE_PREEMPT_LAZY && PREEMPT_RT
+
 choice
 	prompt "Preemption Model"
 	default PREEMPT_NONE

kernel/sched/core.c

@@ -986,6 +986,46 @@ void resched_curr(struct rq *rq)
 		trace_sched_wake_idle_without_ipi(cpu);
 }
 
+#ifdef CONFIG_PREEMPT_LAZY
+
+static int tsk_is_polling(struct task_struct *p)
+{
+#ifdef TIF_POLLING_NRFLAG
+	return test_tsk_thread_flag(p, TIF_POLLING_NRFLAG);
+#else
+	return 0;
+#endif
+}
+
+void resched_curr_lazy(struct rq *rq)
+{
+	struct task_struct *curr = rq->curr;
+	int cpu;
+
+	if (!sched_feat(PREEMPT_LAZY)) {
+		resched_curr(rq);
+		return;
+	}
+
+	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);
@@ -2141,6 +2181,7 @@ void migrate_disable(void)
 	preempt_disable();
 	this_rq()->nr_pinned++;
 	p->migration_disabled = 1;
+	preempt_lazy_disable();
 	preempt_enable();
 }
 EXPORT_SYMBOL_GPL(migrate_disable);
@@ -2171,6 +2212,7 @@ void migrate_enable(void)
 	barrier();
 	p->migration_disabled = 0;
 	this_rq()->nr_pinned--;
+	preempt_lazy_enable();
 	preempt_enable();
 }
 EXPORT_SYMBOL_GPL(migrate_enable);
@@ -4406,6 +4448,9 @@ int sched_fork(unsigned long clone_flags, struct task_struct *p)
 	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);
@@ -6250,6 +6295,7 @@ static void __sched notrace __schedule(unsigned int sched_mode)
 
 	next = pick_next_task(rq, prev, &rf);
 	clear_tsk_need_resched(prev);
+	clear_tsk_need_resched_lazy(prev);
 	clear_preempt_need_resched();
 #ifdef CONFIG_SCHED_DEBUG
 	rq->last_seen_need_resched_ns = 0;
@@ -6467,6 +6513,30 @@ static void __sched notrace preempt_schedule_common(void)
 	} 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_PREEMPTION
 /*
  * This is the entry point to schedule() from in-kernel preemption
@@ -6480,7 +6550,8 @@ asmlinkage __visible void __sched notrace preempt_schedule(void)
 	 */
 	if (likely(!preemptible()))
 		return;
-
+	if (!preemptible_lazy())
+		return;
 	preempt_schedule_common();
 }
 NOKPROBE_SYMBOL(preempt_schedule);
@@ -6513,6 +6584,9 @@ asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
 	if (likely(!preemptible()))
 		return;
 
+	if (!preemptible_lazy())
+		return;
+
 	do {
 		/*
 		 * Because the function tracer can trace preempt_count_sub()
@@ -8674,7 +8748,9 @@ void __init init_idle(struct task_struct *idle, int cpu)
 
 	/* 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:
 	 */