Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Fri, 7 Oct 2022 15:09:18 +0000
From:   Joel Fernandes <joel@joelfernandes.org>
To:     "Paul E. McKenney" <paulmck@kernel.org>
Cc:     rcu@vger.kernel.org, linux-kernel@vger.kernel.org,
        rushikesh.s.kadam@intel.com, urezki@gmail.com,
        neeraj.iitr10@gmail.com, frederic@kernel.org, rostedt@goodmis.org,
        youssefesmat@google.com, surenb@google.com
Subject: Re: [PATCH v7 00/11] rcu: call_rcu() power improvements
Message-ID: <Y0BBHiqI/5JcSdLX@google.com>
References: <20221004024157.2470238-1-joel@joelfernandes.org>
 <20221006185543.GB4196@paulmck-ThinkPad-P17-Gen-1>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <20221006185543.GB4196@paulmck-ThinkPad-P17-Gen-1>
Precedence: bulk

On Thu, Oct 06, 2022 at 11:55:43AM -0700, Paul E. McKenney wrote:
> On Tue, Oct 04, 2022 at 02:41:46AM +0000, Joel Fernandes (Google) wrote:
> > v7 version of RCU lazy patches based on rcu/next branch.
> > 
> > To facilitate easier merge, I dropped tracing and other patches and just
> > implemented the new changes. I will post the tracing patches later along with
> > rcutop as I need to add new tracepoints that Frederic suggested.
> > 
> > Main recent changes:
> > 1. rcu_barrier() wake up only for lazy bypass list.
> > 2. Make all call_rcu() default-lazy and add call_rcu_flush() API.
> > 3. Take care of some callers using call_rcu_flush() API.
> > 4. Several refactorings suggested by Paul/Frederic.
> > 5. New call_rcu() to call_rcu_flush() conversions by Joel/Vlad/Paul.
> > 
> > I am seeing good performance and power with these patches on real ChromeOS x86
> > asymmetric hardware.
> > 
> > Earlier cover letter with lots of details is here:
> > https://lore.kernel.org/all/20220901221720.1105021-1-joel@joelfernandes.org/
> > 
> > List of recent changes:
> >     
> >     [ Frederic Weisbec: Program the lazy timer only if WAKE_NOT, since other
> >       deferral levels wake much earlier so for those it is not needed. ]
> >     
> >     [ Frederic Weisbec: Use flush flags to keep bypass API code clean. ]
> >     
> >     [ Frederic Weisbec: Make rcu_barrier() wake up only if main list empty. ]
> >     
> >     [ Frederic Weisbec: Remove extra 'else if' branch in rcu_nocb_try_bypass(). ]
> >     
> >     [ Joel: Fix issue where I was not resetting lazy_len after moving it to rdp ]
> >     
> >     [ Paul/Thomas/Joel: Make call_rcu() default lazy so users don't mess up. ]
> >     
> >     [ Paul/Frederic : Cosmetic changes, split out wakeup of nocb thread. ]
> >     
> >     [ Vlad/Joel : More call_rcu -> flush conversions ]
> 
> Thank you for your continued work on this!
> 
> I pulled these into an experimental branch, applied Uladzislau's
> Tested-by and ran a quick round of rcutorture.
> 
> From TREE02, TREE03, and TREE09 I got this:
> 
> In file included from kernel/rcu/tree.c:68:
> kernel/rcu/tree.h:449:13: error: ‘wake_nocb_gp’ used but never defined [-Werror]
>   449 | static bool wake_nocb_gp(struct rcu_data *rdp, bool force);
>       |             ^~~~~~~~~~~~
> 
> One could argue that this is not a big deal, except that Linus gets a
> bit tetchy when this sort of thing shows up in mainline.
> 
> Nevetheless, it looks like we might be getting quite close!

Yes, sorry this is because a hunk got misplaced as I mentioned in the other
thread. If you replace the second patch with the below, it will fix it,
thanks!

I pushed all the latest patches to:
https://github.com/joelagnel/linux-kernel.git
(branch  rcu-lazy-latest-oct722)

If you want to try those.

---8<-----------------------

From 978bf585cbd933048733d9ad9a80ed8de265c07e Mon Sep 17 00:00:00 2001
From: "Joel Fernandes (Google)" <joel@joelfernandes.org>
Date: Fri, 7 Oct 2022 02:41:00 +0000
Subject: [PATCH] rcu: Make call_rcu() lazy to save power

Implement timer-based RCU callback batching (also known as lazy
callbacks). With this we save about 5-10% of power consumed due
to RCU requests that happen when system is lightly loaded or idle.

By default, all async callbacks (queued via call_rcu) are marked
lazy. An alternate API call_rcu_flush() is provided for the few users,
for example synchronize_rcu(), that need the old behavior.

The batch is flushed whenever a certain amount of time has passed, or
the batch on a particular CPU grows too big. Also memory pressure will
flush it in a future patch.

To handle several corner cases automagically (such as rcu_barrier() and
hotplug), we re-use bypass lists which were originally introduced to
address lock contention, to handle lazy CBs as well. The bypass list
length has the lazy CB length included in it. A separate lazy CB length
counter is also introduced to keep track of the number of lazy CBs.

Suggested-by: Paul McKenney <paulmck@kernel.org>
Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
---
 include/linux/rcupdate.h |   7 ++
 kernel/rcu/Kconfig       |   8 ++
 kernel/rcu/rcu.h         |   8 ++
 kernel/rcu/tiny.c        |   2 +-
 kernel/rcu/tree.c        | 129 ++++++++++++++++++++-----------
 kernel/rcu/tree.h        |  11 ++-
 kernel/rcu/tree_exp.h    |   2 +-
 kernel/rcu/tree_nocb.h   | 161 +++++++++++++++++++++++++++++++--------
 8 files changed, 246 insertions(+), 82 deletions(-)

diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h
index 08605ce7379d..40ae36904825 100644
--- a/include/linux/rcupdate.h
+++ b/include/linux/rcupdate.h
@@ -108,6 +108,13 @@ static inline int rcu_preempt_depth(void)
 
 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
 
+#ifdef CONFIG_RCU_LAZY
+void call_rcu_flush(struct rcu_head *head, rcu_callback_t func);
+#else
+static inline void call_rcu_flush(struct rcu_head *head,
+		rcu_callback_t func) {  call_rcu(head, func); }
+#endif
+
 /* Internal to kernel */
 void rcu_init(void);
 extern int rcu_scheduler_active;
diff --git a/kernel/rcu/Kconfig b/kernel/rcu/Kconfig
index f53ad63b2bc6..edd632e68497 100644
--- a/kernel/rcu/Kconfig
+++ b/kernel/rcu/Kconfig
@@ -314,4 +314,12 @@ config TASKS_TRACE_RCU_READ_MB
 	  Say N here if you hate read-side memory barriers.
 	  Take the default if you are unsure.
 
+config RCU_LAZY
+	bool "RCU callback lazy invocation functionality"
+	depends on RCU_NOCB_CPU
+	default n
+	help
+	  To save power, batch RCU callbacks and flush after delay, memory
+	  pressure or callback list growing too big.
+
 endmenu # "RCU Subsystem"
diff --git a/kernel/rcu/rcu.h b/kernel/rcu/rcu.h
index be5979da07f5..65704cbc9df7 100644
--- a/kernel/rcu/rcu.h
+++ b/kernel/rcu/rcu.h
@@ -474,6 +474,14 @@ enum rcutorture_type {
 	INVALID_RCU_FLAVOR
 };
 
+#if defined(CONFIG_RCU_LAZY)
+unsigned long rcu_lazy_get_jiffies_till_flush(void);
+void rcu_lazy_set_jiffies_till_flush(unsigned long j);
+#else
+static inline unsigned long rcu_lazy_get_jiffies_till_flush(void) { return 0; }
+static inline void rcu_lazy_set_jiffies_till_flush(unsigned long j) { }
+#endif
+
 #if defined(CONFIG_TREE_RCU)
 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
 			    unsigned long *gp_seq);
diff --git a/kernel/rcu/tiny.c b/kernel/rcu/tiny.c
index a33a8d4942c3..810479cf17ba 100644
--- a/kernel/rcu/tiny.c
+++ b/kernel/rcu/tiny.c
@@ -44,7 +44,7 @@ static struct rcu_ctrlblk rcu_ctrlblk = {
 
 void rcu_barrier(void)
 {
-	wait_rcu_gp(call_rcu);
+	wait_rcu_gp(call_rcu_flush);
 }
 EXPORT_SYMBOL(rcu_barrier);
 
diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index 04f33191e5ed..c20544c4aa29 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -2728,47 +2728,8 @@ static void check_cb_ovld(struct rcu_data *rdp)
 	raw_spin_unlock_rcu_node(rnp);
 }
 
-/**
- * call_rcu() - Queue an RCU callback for invocation after a grace period.
- * @head: structure to be used for queueing the RCU updates.
- * @func: actual callback function to be invoked after the grace period
- *
- * The callback function will be invoked some time after a full grace
- * period elapses, in other words after all pre-existing RCU read-side
- * critical sections have completed.  However, the callback function
- * might well execute concurrently with RCU read-side critical sections
- * that started after call_rcu() was invoked.
- *
- * RCU read-side critical sections are delimited by rcu_read_lock()
- * and rcu_read_unlock(), and may be nested.  In addition, but only in
- * v5.0 and later, regions of code across which interrupts, preemption,
- * or softirqs have been disabled also serve as RCU read-side critical
- * sections.  This includes hardware interrupt handlers, softirq handlers,
- * and NMI handlers.
- *
- * Note that all CPUs must agree that the grace period extended beyond
- * all pre-existing RCU read-side critical section.  On systems with more
- * than one CPU, this means that when "func()" is invoked, each CPU is
- * guaranteed to have executed a full memory barrier since the end of its
- * last RCU read-side critical section whose beginning preceded the call
- * to call_rcu().  It also means that each CPU executing an RCU read-side
- * critical section that continues beyond the start of "func()" must have
- * executed a memory barrier after the call_rcu() but before the beginning
- * of that RCU read-side critical section.  Note that these guarantees
- * include CPUs that are offline, idle, or executing in user mode, as
- * well as CPUs that are executing in the kernel.
- *
- * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
- * resulting RCU callback function "func()", then both CPU A and CPU B are
- * guaranteed to execute a full memory barrier during the time interval
- * between the call to call_rcu() and the invocation of "func()" -- even
- * if CPU A and CPU B are the same CPU (but again only if the system has
- * more than one CPU).
- *
- * Implementation of these memory-ordering guarantees is described here:
- * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
- */
-void call_rcu(struct rcu_head *head, rcu_callback_t func)
+static void
+__call_rcu_common(struct rcu_head *head, rcu_callback_t func, bool lazy)
 {
 	static atomic_t doublefrees;
 	unsigned long flags;
@@ -2809,7 +2770,7 @@ void call_rcu(struct rcu_head *head, rcu_callback_t func)
 	}
 
 	check_cb_ovld(rdp);
-	if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags))
+	if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags, lazy))
 		return; // Enqueued onto ->nocb_bypass, so just leave.
 	// If no-CBs CPU gets here, rcu_nocb_try_bypass() acquired ->nocb_lock.
 	rcu_segcblist_enqueue(&rdp->cblist, head);
@@ -2831,8 +2792,84 @@ void call_rcu(struct rcu_head *head, rcu_callback_t func)
 		local_irq_restore(flags);
 	}
 }
-EXPORT_SYMBOL_GPL(call_rcu);
 
+#ifdef CONFIG_RCU_LAZY
+/**
+ * call_rcu_flush() - Queue RCU callback for invocation after grace period, and
+ * flush all lazy callbacks (including the new one) to the main ->cblist while
+ * doing so.
+ *
+ * @head: structure to be used for queueing the RCU updates.
+ * @func: actual callback function to be invoked after the grace period
+ *
+ * The callback function will be invoked some time after a full grace
+ * period elapses, in other words after all pre-existing RCU read-side
+ * critical sections have completed.
+ *
+ * Use this API instead of call_rcu() if you don't want the callback to be
+ * invoked after very long periods of time, which can happen on systems without
+ * memory pressure and on systems which are lightly loaded or mostly idle.
+ * This function will cause callbacks to be invoked sooner than later at the
+ * expense of extra power. Other than that, this function is identical to, and
+ * reuses call_rcu()'s logic. Refer to call_rcu() for more details about memory
+ * ordering and other functionality.
+ */
+void call_rcu_flush(struct rcu_head *head, rcu_callback_t func)
+{
+	return __call_rcu_common(head, func, false);
+}
+EXPORT_SYMBOL_GPL(call_rcu_flush);
+#endif
+
+/**
+ * call_rcu() - Queue an RCU callback for invocation after a grace period.
+ * By default the callbacks are 'lazy' and are kept hidden from the main
+ * ->cblist to prevent starting of grace periods too soon.
+ * If you desire grace periods to start very soon, use call_rcu_flush().
+ *
+ * @head: structure to be used for queueing the RCU updates.
+ * @func: actual callback function to be invoked after the grace period
+ *
+ * The callback function will be invoked some time after a full grace
+ * period elapses, in other words after all pre-existing RCU read-side
+ * critical sections have completed.  However, the callback function
+ * might well execute concurrently with RCU read-side critical sections
+ * that started after call_rcu() was invoked.
+ *
+ * RCU read-side critical sections are delimited by rcu_read_lock()
+ * and rcu_read_unlock(), and may be nested.  In addition, but only in
+ * v5.0 and later, regions of code across which interrupts, preemption,
+ * or softirqs have been disabled also serve as RCU read-side critical
+ * sections.  This includes hardware interrupt handlers, softirq handlers,
+ * and NMI handlers.
+ *
+ * Note that all CPUs must agree that the grace period extended beyond
+ * all pre-existing RCU read-side critical section.  On systems with more
+ * than one CPU, this means that when "func()" is invoked, each CPU is
+ * guaranteed to have executed a full memory barrier since the end of its
+ * last RCU read-side critical section whose beginning preceded the call
+ * to call_rcu().  It also means that each CPU executing an RCU read-side
+ * critical section that continues beyond the start of "func()" must have
+ * executed a memory barrier after the call_rcu() but before the beginning
+ * of that RCU read-side critical section.  Note that these guarantees
+ * include CPUs that are offline, idle, or executing in user mode, as
+ * well as CPUs that are executing in the kernel.
+ *
+ * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
+ * resulting RCU callback function "func()", then both CPU A and CPU B are
+ * guaranteed to execute a full memory barrier during the time interval
+ * between the call to call_rcu() and the invocation of "func()" -- even
+ * if CPU A and CPU B are the same CPU (but again only if the system has
+ * more than one CPU).
+ *
+ * Implementation of these memory-ordering guarantees is described here:
+ * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
+ */
+void call_rcu(struct rcu_head *head, rcu_callback_t func)
+{
+	return __call_rcu_common(head, func, true);
+}
+EXPORT_SYMBOL_GPL(call_rcu);
 
 /* Maximum number of jiffies to wait before draining a batch. */
 #define KFREE_DRAIN_JIFFIES (5 * HZ)
@@ -3507,7 +3544,7 @@ void synchronize_rcu(void)
 		if (rcu_gp_is_expedited())
 			synchronize_rcu_expedited();
 		else
-			wait_rcu_gp(call_rcu);
+			wait_rcu_gp(call_rcu_flush);
 		return;
 	}
 
@@ -3905,7 +3942,7 @@ static void rcu_barrier_entrain(struct rcu_data *rdp)
 	debug_rcu_head_queue(&rdp->barrier_head);
 	rcu_nocb_lock(rdp);
 	was_done = rcu_rdp_is_offloaded(rdp) && !rcu_segcblist_pend_cbs(&rdp->cblist);
-	WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
+	WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies, false));
 	if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head)) {
 		atomic_inc(&rcu_state.barrier_cpu_count);
 	} else {
@@ -4334,7 +4371,7 @@ void rcutree_migrate_callbacks(int cpu)
 	my_rdp = this_cpu_ptr(&rcu_data);
 	my_rnp = my_rdp->mynode;
 	rcu_nocb_lock(my_rdp); /* irqs already disabled. */
-	WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies));
+	WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies, false));
 	raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
 	/* Leverage recent GPs and set GP for new callbacks. */
 	needwake = rcu_advance_cbs(my_rnp, rdp) ||
diff --git a/kernel/rcu/tree.h b/kernel/rcu/tree.h
index 41faaf8eb872..1d803d39f0d1 100644
--- a/kernel/rcu/tree.h
+++ b/kernel/rcu/tree.h
@@ -263,14 +263,16 @@ struct rcu_data {
 	unsigned long last_fqs_resched;	/* Time of last rcu_resched(). */
 	unsigned long last_sched_clock;	/* Jiffies of last rcu_sched_clock_irq(). */
 
+	long lazy_len;			/* Length of buffered lazy callbacks. */
 	int cpu;
 };
 
 /* Values for nocb_defer_wakeup field in struct rcu_data. */
 #define RCU_NOCB_WAKE_NOT	0
 #define RCU_NOCB_WAKE_BYPASS	1
-#define RCU_NOCB_WAKE		2
-#define RCU_NOCB_WAKE_FORCE	3
+#define RCU_NOCB_WAKE_LAZY	2
+#define RCU_NOCB_WAKE		3
+#define RCU_NOCB_WAKE_FORCE	4
 
 #define RCU_JIFFIES_TILL_FORCE_QS (1 + (HZ > 250) + (HZ > 500))
 					/* For jiffies_till_first_fqs and */
@@ -440,9 +442,10 @@ static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp);
 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq);
 static void rcu_init_one_nocb(struct rcu_node *rnp);
 static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
-				  unsigned long j);
+				  unsigned long j, bool lazy);
 static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
-				bool *was_alldone, unsigned long flags);
+				bool *was_alldone, unsigned long flags,
+				bool lazy);
 static bool wake_nocb_gp(struct rcu_data *rdp, bool force);
 static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_empty,
 				 unsigned long flags);
diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h
index 18e9b4cd78ef..5cac05600798 100644
--- a/kernel/rcu/tree_exp.h
+++ b/kernel/rcu/tree_exp.h
@@ -937,7 +937,7 @@ void synchronize_rcu_expedited(void)
 
 	/* If expedited grace periods are prohibited, fall back to normal. */
 	if (rcu_gp_is_normal()) {
-		wait_rcu_gp(call_rcu);
+		wait_rcu_gp(call_rcu_flush);
 		return;
 	}
 
diff --git a/kernel/rcu/tree_nocb.h b/kernel/rcu/tree_nocb.h
index 6caade0683dd..ccd05ded6bb4 100644
--- a/kernel/rcu/tree_nocb.h
+++ b/kernel/rcu/tree_nocb.h
@@ -256,6 +256,31 @@ static bool wake_nocb_gp(struct rcu_data *rdp, bool force)
 	return __wake_nocb_gp(rdp_gp, rdp, force, flags);
 }
 
+/*
+ * LAZY_FLUSH_JIFFIES decides the maximum amount of time that
+ * can elapse before lazy callbacks are flushed. Lazy callbacks
+ * could be flushed much earlier for a number of other reasons
+ * however, LAZY_FLUSH_JIFFIES will ensure no lazy callbacks are
+ * left unsubmitted to RCU after those many jiffies.
+ */
+#define LAZY_FLUSH_JIFFIES (10 * HZ)
+static unsigned long jiffies_till_flush = LAZY_FLUSH_JIFFIES;
+
+#ifdef CONFIG_RCU_LAZY
+// To be called only from test code.
+void rcu_lazy_set_jiffies_till_flush(unsigned long jif)
+{
+	jiffies_till_flush = jif;
+}
+EXPORT_SYMBOL(rcu_lazy_set_jiffies_till_flush);
+
+unsigned long rcu_lazy_get_jiffies_till_flush(void)
+{
+	return jiffies_till_flush;
+}
+EXPORT_SYMBOL(rcu_lazy_get_jiffies_till_flush);
+#endif
+
 /*
  * Arrange to wake the GP kthread for this NOCB group at some future
  * time when it is safe to do so.
@@ -269,10 +294,14 @@ static void wake_nocb_gp_defer(struct rcu_data *rdp, int waketype,
 	raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
 
 	/*
-	 * Bypass wakeup overrides previous deferments. In case
-	 * of callback storm, no need to wake up too early.
+	 * Bypass wakeup overrides previous deferments. In case of
+	 * callback storm, no need to wake up too early.
 	 */
-	if (waketype == RCU_NOCB_WAKE_BYPASS) {
+	if (waketype == RCU_NOCB_WAKE_LAZY &&
+	    READ_ONCE(rdp->nocb_defer_wakeup) == RCU_NOCB_WAKE_NOT) {
+		mod_timer(&rdp_gp->nocb_timer, jiffies + jiffies_till_flush);
+		WRITE_ONCE(rdp_gp->nocb_defer_wakeup, waketype);
+	} else if (waketype == RCU_NOCB_WAKE_BYPASS) {
 		mod_timer(&rdp_gp->nocb_timer, jiffies + 2);
 		WRITE_ONCE(rdp_gp->nocb_defer_wakeup, waketype);
 	} else {
@@ -293,10 +322,13 @@ static void wake_nocb_gp_defer(struct rcu_data *rdp, int waketype,
  * proves to be initially empty, just return false because the no-CB GP
  * kthread may need to be awakened in this case.
  *
+ * Return true if there was something to be flushed and it succeeded, otherwise
+ * false.
+ *
  * Note that this function always returns true if rhp is NULL.
  */
 static bool rcu_nocb_do_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
-				     unsigned long j)
+				     unsigned long j, bool lazy)
 {
 	struct rcu_cblist rcl;
 
@@ -310,7 +342,20 @@ static bool rcu_nocb_do_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
 	/* Note: ->cblist.len already accounts for ->nocb_bypass contents. */
 	if (rhp)
 		rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
-	rcu_cblist_flush_enqueue(&rcl, &rdp->nocb_bypass, rhp);
+
+	/*
+	 * If the new CB requested was a lazy one, queue it onto the main
+	 * ->cblist so we can take advantage of a sooner grade period.
+	 */
+	if (lazy && rhp) {
+		rcu_cblist_flush_enqueue(&rcl, &rdp->nocb_bypass, NULL);
+		rcu_cblist_enqueue(&rcl, rhp);
+		WRITE_ONCE(rdp->lazy_len, 0);
+	} else {
+		rcu_cblist_flush_enqueue(&rcl, &rdp->nocb_bypass, rhp);
+		WRITE_ONCE(rdp->lazy_len, 0);
+	}
+
 	rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rcl);
 	WRITE_ONCE(rdp->nocb_bypass_first, j);
 	rcu_nocb_bypass_unlock(rdp);
@@ -326,13 +371,13 @@ static bool rcu_nocb_do_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
  * Note that this function always returns true if rhp is NULL.
  */
 static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
-				  unsigned long j)
+				  unsigned long j, bool lazy)
 {
 	if (!rcu_rdp_is_offloaded(rdp))
 		return true;
 	rcu_lockdep_assert_cblist_protected(rdp);
 	rcu_nocb_bypass_lock(rdp);
-	return rcu_nocb_do_flush_bypass(rdp, rhp, j);
+	return rcu_nocb_do_flush_bypass(rdp, rhp, j, lazy);
 }
 
 /*
@@ -345,7 +390,7 @@ static void rcu_nocb_try_flush_bypass(struct rcu_data *rdp, unsigned long j)
 	if (!rcu_rdp_is_offloaded(rdp) ||
 	    !rcu_nocb_bypass_trylock(rdp))
 		return;
-	WARN_ON_ONCE(!rcu_nocb_do_flush_bypass(rdp, NULL, j));
+	WARN_ON_ONCE(!rcu_nocb_do_flush_bypass(rdp, NULL, j, false));
 }
 
 /*
@@ -367,12 +412,14 @@ static void rcu_nocb_try_flush_bypass(struct rcu_data *rdp, unsigned long j)
  * there is only one CPU in operation.
  */
 static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
-				bool *was_alldone, unsigned long flags)
+				bool *was_alldone, unsigned long flags,
+				bool lazy)
 {
 	unsigned long c;
 	unsigned long cur_gp_seq;
 	unsigned long j = jiffies;
 	long ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
+	bool bypass_is_lazy = (ncbs == READ_ONCE(rdp->lazy_len));
 
 	lockdep_assert_irqs_disabled();
 
@@ -417,25 +464,29 @@ static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
 	// If there hasn't yet been all that many ->cblist enqueues
 	// this jiffy, tell the caller to enqueue onto ->cblist.  But flush
 	// ->nocb_bypass first.
-	if (rdp->nocb_nobypass_count < nocb_nobypass_lim_per_jiffy) {
+	// Lazy CBs throttle this back and do immediate bypass queuing.
+	if (rdp->nocb_nobypass_count < nocb_nobypass_lim_per_jiffy && !lazy) {
 		rcu_nocb_lock(rdp);
 		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
 		if (*was_alldone)
 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
 					    TPS("FirstQ"));
-		WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, j));
+
+		WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, j, false));
 		WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
 		return false; // Caller must enqueue the callback.
 	}
 
 	// If ->nocb_bypass has been used too long or is too full,
 	// flush ->nocb_bypass to ->cblist.
-	if ((ncbs && j != READ_ONCE(rdp->nocb_bypass_first)) ||
+	if ((ncbs && !bypass_is_lazy && j != READ_ONCE(rdp->nocb_bypass_first)) ||
+	    (ncbs &&  bypass_is_lazy &&
+		(time_after(j, READ_ONCE(rdp->nocb_bypass_first) + jiffies_till_flush))) ||
 	    ncbs >= qhimark) {
 		rcu_nocb_lock(rdp);
 		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
 
-		if (!rcu_nocb_flush_bypass(rdp, rhp, j)) {
+		if (!rcu_nocb_flush_bypass(rdp, rhp, j, lazy)) {
 			if (*was_alldone)
 				trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
 						    TPS("FirstQ"));
@@ -460,16 +511,29 @@ static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
 	// We need to use the bypass.
 	rcu_nocb_wait_contended(rdp);
 	rcu_nocb_bypass_lock(rdp);
+
 	ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
 	rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
 	rcu_cblist_enqueue(&rdp->nocb_bypass, rhp);
+
+	if (lazy)
+		WRITE_ONCE(rdp->lazy_len, rdp->lazy_len + 1);
+
 	if (!ncbs) {
 		WRITE_ONCE(rdp->nocb_bypass_first, j);
 		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("FirstBQ"));
 	}
+
 	rcu_nocb_bypass_unlock(rdp);
 	smp_mb(); /* Order enqueue before wake. */
-	if (ncbs) {
+
+	// A wake up of the grace period kthread or timer adjustment needs to
+	// be done only if:
+	// 1. Bypass list was fully empty before (this is the first bypass list entry).
+	//	Or, both the below conditions are met:
+	// 1. Bypass list had only lazy CBs before.
+	// 2. The new CB is non-lazy.
+	if (ncbs && (!bypass_is_lazy || lazy)) {
 		local_irq_restore(flags);
 	} else {
 		// No-CBs GP kthread might be indefinitely asleep, if so, wake.
@@ -497,9 +561,10 @@ static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_alldone,
 				 unsigned long flags)
 				 __releases(rdp->nocb_lock)
 {
+	long bypass_len;
 	unsigned long cur_gp_seq;
 	unsigned long j;
-	long len;
+	long len, lazy_len;
 	struct task_struct *t;
 
 	// If we are being polled or there is no kthread, just leave.
@@ -512,9 +577,16 @@ static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_alldone,
 	}
 	// Need to actually to a wakeup.
 	len = rcu_segcblist_n_cbs(&rdp->cblist);
+	bypass_len = rcu_cblist_n_cbs(&rdp->nocb_bypass);
+	lazy_len = READ_ONCE(rdp->lazy_len);
 	if (was_alldone) {
 		rdp->qlen_last_fqs_check = len;
-		if (!irqs_disabled_flags(flags)) {
+		// Only lazy CBs in bypass list
+		if (lazy_len && bypass_len == lazy_len) {
+			rcu_nocb_unlock_irqrestore(rdp, flags);
+			wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE_LAZY,
+					   TPS("WakeLazy"));
+		} else if (!irqs_disabled_flags(flags)) {
 			/* ... if queue was empty ... */
 			rcu_nocb_unlock_irqrestore(rdp, flags);
 			wake_nocb_gp(rdp, false);
@@ -611,6 +683,8 @@ static void nocb_gp_wait(struct rcu_data *my_rdp)
 	unsigned long flags;
 	bool gotcbs = false;
 	unsigned long j = jiffies;
+	bool lazy = false;
+	long lazy_ncbs;
 	bool needwait_gp = false; // This prevents actual uninitialized use.
 	bool needwake;
 	bool needwake_gp;
@@ -640,24 +714,41 @@ static void nocb_gp_wait(struct rcu_data *my_rdp)
 	 * won't be ignored for long.
 	 */
 	list_for_each_entry(rdp, &my_rdp->nocb_head_rdp, nocb_entry_rdp) {
+		bool flush_bypass = false;
+
 		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Check"));
 		rcu_nocb_lock_irqsave(rdp, flags);
 		lockdep_assert_held(&rdp->nocb_lock);
 		bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
-		if (bypass_ncbs &&
+		lazy_ncbs = READ_ONCE(rdp->lazy_len);
+
+		if (bypass_ncbs && (lazy_ncbs == bypass_ncbs) &&
+		    (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + jiffies_till_flush) ||
+		     bypass_ncbs > 2 * qhimark)) {
+			flush_bypass = true;
+		} else if (bypass_ncbs && (lazy_ncbs != bypass_ncbs) &&
 		    (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + 1) ||
 		     bypass_ncbs > 2 * qhimark)) {
-			// Bypass full or old, so flush it.
-			(void)rcu_nocb_try_flush_bypass(rdp, j);
-			bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
+			flush_bypass = true;
 		} else if (!bypass_ncbs && rcu_segcblist_empty(&rdp->cblist)) {
 			rcu_nocb_unlock_irqrestore(rdp, flags);
 			continue; /* No callbacks here, try next. */
 		}
+
+		if (flush_bypass) {
+			// Bypass full or old, so flush it.
+			(void)rcu_nocb_try_flush_bypass(rdp, j);
+			bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
+			lazy_ncbs = READ_ONCE(rdp->lazy_len);
+		}
+
 		if (bypass_ncbs) {
 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
-					    TPS("Bypass"));
-			bypass = true;
+				    bypass_ncbs == lazy_ncbs ? TPS("Lazy") : TPS("Bypass"));
+			if (bypass_ncbs == lazy_ncbs)
+				lazy = true;
+			else
+				bypass = true;
 		}
 		rnp = rdp->mynode;
 
@@ -705,12 +796,20 @@ static void nocb_gp_wait(struct rcu_data *my_rdp)
 	my_rdp->nocb_gp_gp = needwait_gp;
 	my_rdp->nocb_gp_seq = needwait_gp ? wait_gp_seq : 0;
 
-	if (bypass && !rcu_nocb_poll) {
-		// At least one child with non-empty ->nocb_bypass, so set
-		// timer in order to avoid stranding its callbacks.
-		wake_nocb_gp_defer(my_rdp, RCU_NOCB_WAKE_BYPASS,
-				   TPS("WakeBypassIsDeferred"));
+	// At least one child with non-empty ->nocb_bypass, so set
+	// timer in order to avoid stranding its callbacks.
+	if (!rcu_nocb_poll) {
+		// If bypass list only has lazy CBs. Add a deferred lazy wake up.
+		if (lazy && !bypass) {
+			wake_nocb_gp_defer(my_rdp, RCU_NOCB_WAKE_LAZY,
+					TPS("WakeLazyIsDeferred"));
+		// Otherwise add a deferred bypass wake up.
+		} else if (bypass) {
+			wake_nocb_gp_defer(my_rdp, RCU_NOCB_WAKE_BYPASS,
+					TPS("WakeBypassIsDeferred"));
+		}
 	}
+
 	if (rcu_nocb_poll) {
 		/* Polling, so trace if first poll in the series. */
 		if (gotcbs)
@@ -1036,7 +1135,7 @@ static long rcu_nocb_rdp_deoffload(void *arg)
 	 * return false, which means that future calls to rcu_nocb_try_bypass()
 	 * will refuse to put anything into the bypass.
 	 */
-	WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
+	WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies, false));
 	/*
 	 * Start with invoking rcu_core() early. This way if the current thread
 	 * happens to preempt an ongoing call to rcu_core() in the middle,
@@ -1278,6 +1377,7 @@ static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
 	raw_spin_lock_init(&rdp->nocb_gp_lock);
 	timer_setup(&rdp->nocb_timer, do_nocb_deferred_wakeup_timer, 0);
 	rcu_cblist_init(&rdp->nocb_bypass);
+	WRITE_ONCE(rdp->lazy_len, 0);
 	mutex_init(&rdp->nocb_gp_kthread_mutex);
 }
 
@@ -1560,16 +1660,17 @@ static void rcu_init_one_nocb(struct rcu_node *rnp)
 
 static bool wake_nocb_gp(struct rcu_data *rdp, bool force)
 {
+	return false;
 }
 
 static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
-				  unsigned long j)
+				  unsigned long j, bool lazy)
 {
 	return true;
 }
 
 static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
-				bool *was_alldone, unsigned long flags)
+				bool *was_alldone, unsigned long flags, bool lazy)
 {
 	return false;
 }
-- 
2.38.0.rc1.362.ged0d419d3c-goog