[PATCH v2 00/10] Make RCU Tasks scan idle tasks

[PATCH v2 00/10] Make RCU Tasks scan idle tasks

[neeraj.upadhyay]

From: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>

Architectures for which all deep-idle and entry-exit functions are
marked noinstr, synchronize_rcu_tasks_rude() is not required. However,
disabling synchronize_rcu_tasks_rude() for such architectures require
RCU-tasks to pay attention to idle tasks until they enter the context
where RCU is not watching. In addition, NMI context need to track RCU-tasks
entry/exit for NO_HZ_FULL CPUs. This series adds changes for both of
these requirements.

The series is available in git tree at:

https://git.kernel.org/pub/scm/linux/kernel/git/neeraj.upadhyay/linux-rcu.git/log/?h=rcu.tasks.idle

Changes since v1: https://lore.kernel.org/lkml/20240807163506.434885-1-neeraj.upadhyay@kernel.org/

- Split into multiple patches (Paul)
- Various restructuring updates (Frederic)
- Fix race with early boot code (Frederic)
- Check RCU watching state for idle injection tasks.
- Include patch to turn synchronize_rcu_tasks_rude() into noop
  for architectures with CONFIG_ARCH_WANTS_NO_INSTR enabled.

Neeraj Upadhyay (9):
  rcu: Change rdp arg to cpu number for
    rcu_watching_snap_stopped_since()
  rcu: Make some rcu_watching_* functions global
  rcu/tasks: Move holdout checks for idle task to a separate function
  rcu/tasks: Create rcu_idle_task_is_holdout() definition for !SMP
  rcu/tasks: Consider idle tasks not running on CPU as non-holdouts
  rcu/tasks: Check RCU watching state for holdout idle tasks
  rcu/tasks: Check RCU watching state for holdout idle injection tasks
  rcu/tasks: Make RCU-tasks pay attention to idle tasks
  context_tracking: Invoke RCU-tasks enter/exit for NMI context

Paul E. McKenney (1):
  rcu: Allow short-circuiting of synchronize_rcu_tasks_rude()

 .../RCU/Design/Requirements/Requirements.rst  |  12 +-
 kernel/context_tracking.c                     |  11 +-
 kernel/rcu/rcu.h                              |   4 +
 kernel/rcu/tasks.h                            | 148 ++++++++++++++----
 kernel/rcu/tree.c                             |  16 +-
 kernel/rcu/tree_exp.h                         |   2 +-
 6 files changed, 140 insertions(+), 53 deletions(-)

-- 
2.40.1

[PATCH v2 01/10] rcu: Change rdp arg to cpu number for rcu_watching_snap_stopped_since()

[neeraj.upadhyay]

From: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>

In preparation of making rcu_watching_snap_stopped_since() available
for RCU code outside of tree.c, change the rdp argument to cpu number.

Signed-off-by: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>
---
 kernel/rcu/tree.c     | 14 +++++++-------
 kernel/rcu/tree_exp.h |  2 +-
 2 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index a60616e69b66..ea17dd2d0344 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -312,16 +312,16 @@ static bool rcu_watching_snap_in_eqs(int snap)
  * rcu_watching_snap_stopped_since() - Has RCU stopped watching a given CPU
  * since the specified @snap?
  *
- * @rdp: The rcu_data corresponding to the CPU for which to check EQS.
+ * @cpu: The CPU for which to check EQS.
  * @snap: rcu_watching snapshot taken when the CPU wasn't in an EQS.
  *
- * Returns true if the CPU corresponding to @rdp has spent some time in an
- * extended quiescent state since @snap. Note that this doesn't check if it
- * /still/ is in an EQS, just that it went through one since @snap.
+ * Returns true if the CPU has spent some time in an extended quiescent state
+ * since @snap. Note that this doesn't check if it /still/ is in an EQS, just
+ * that it went through one since @snap.
  *
  * This is meant to be used in a loop waiting for a CPU to go through an EQS.
  */
-static bool rcu_watching_snap_stopped_since(struct rcu_data *rdp, int snap)
+static bool rcu_watching_snap_stopped_since(int cpu, int snap)
 {
 	/*
 	 * The first failing snapshot is already ordered against the accesses
@@ -334,7 +334,7 @@ static bool rcu_watching_snap_stopped_since(struct rcu_data *rdp, int snap)
 	if (WARN_ON_ONCE(rcu_watching_snap_in_eqs(snap)))
 		return true;
 
-	return snap != ct_rcu_watching_cpu_acquire(rdp->cpu);
+	return snap != ct_rcu_watching_cpu_acquire(cpu);
 }
 
 /*
@@ -826,7 +826,7 @@ static int rcu_watching_snap_recheck(struct rcu_data *rdp)
 	 * read-side critical section that started before the beginning
 	 * of the current RCU grace period.
 	 */
-	if (rcu_watching_snap_stopped_since(rdp, rdp->watching_snap)) {
+	if (rcu_watching_snap_stopped_since(rdp->cpu, rdp->watching_snap)) {
 		trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
 		rcu_gpnum_ovf(rnp, rdp);
 		return 1;
diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h
index fb664d3a01c9..e3bd4c18a852 100644
--- a/kernel/rcu/tree_exp.h
+++ b/kernel/rcu/tree_exp.h
@@ -401,7 +401,7 @@ static void __sync_rcu_exp_select_node_cpus(struct rcu_exp_work *rewp)
 		unsigned long mask = rdp->grpmask;
 
 retry_ipi:
-		if (rcu_watching_snap_stopped_since(rdp, rdp->exp_watching_snap)) {
+		if (rcu_watching_snap_stopped_since(rdp->cpu, rdp->exp_watching_snap)) {
 			mask_ofl_test |= mask;
 			continue;
 		}
-- 
2.40.1

[PATCH v2 02/10] rcu: Make some rcu_watching_* functions global

[neeraj.upadhyay]

From: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>

rcu_watching_snap_in_eqs() and rcu_watching_snap_stopped_since()
will be used in subsequent commits by RCU-tasks to check rcu_watching
state of idle tasks. So, make these functions global.

Signed-off-by: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>
---
 kernel/rcu/rcu.h  | 4 ++++
 kernel/rcu/tree.c | 4 ++--
 2 files changed, 6 insertions(+), 2 deletions(-)

diff --git a/kernel/rcu/rcu.h b/kernel/rcu/rcu.h
index feb3ac1dc5d5..5fec1b52039c 100644
--- a/kernel/rcu/rcu.h
+++ b/kernel/rcu/rcu.h
@@ -609,6 +609,8 @@ void srcutorture_get_gp_data(struct srcu_struct *sp, int *flags,
 
 #ifdef CONFIG_TINY_RCU
 static inline bool rcu_watching_zero_in_eqs(int cpu, int *vp) { return false; }
+static inline bool rcu_watching_snap_in_eqs(int snap) { return false; }
+static inline bool rcu_watching_snap_stopped_since(int cpu, int snap) { return false; }
 static inline unsigned long rcu_get_gp_seq(void) { return 0; }
 static inline unsigned long rcu_exp_batches_completed(void) { return 0; }
 static inline unsigned long
@@ -622,6 +624,8 @@ static inline void rcu_gp_slow_register(atomic_t *rgssp) { }
 static inline void rcu_gp_slow_unregister(atomic_t *rgssp) { }
 #else /* #ifdef CONFIG_TINY_RCU */
 bool rcu_watching_zero_in_eqs(int cpu, int *vp);
+bool rcu_watching_snap_in_eqs(int snap);
+bool rcu_watching_snap_stopped_since(int cpu, int snap);
 unsigned long rcu_get_gp_seq(void);
 unsigned long rcu_exp_batches_completed(void);
 unsigned long srcu_batches_completed(struct srcu_struct *sp);
diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index ea17dd2d0344..5ecbf85f157d 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -303,7 +303,7 @@ static void rcu_watching_online(void)
  * Return true if the snapshot returned from ct_rcu_watching()
  * indicates that RCU is in an extended quiescent state.
  */
-static bool rcu_watching_snap_in_eqs(int snap)
+bool rcu_watching_snap_in_eqs(int snap)
 {
 	return !(snap & CT_RCU_WATCHING);
 }
@@ -321,7 +321,7 @@ static bool rcu_watching_snap_in_eqs(int snap)
  *
  * This is meant to be used in a loop waiting for a CPU to go through an EQS.
  */
-static bool rcu_watching_snap_stopped_since(int cpu, int snap)
+bool rcu_watching_snap_stopped_since(int cpu, int snap)
 {
 	/*
 	 * The first failing snapshot is already ordered against the accesses
-- 
2.40.1

[PATCH v2 03/10] rcu/tasks: Move holdout checks for idle task to a separate function

[neeraj.upadhyay]

From: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>

Move checks for an idle task being a holdout task for RCU-tasks
to a separate function - rcu_idle_task_is_holdout(). This function
will be used in subsequent commits to add additional checks for
idle task. No functional change intended.

Suggested-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>
---
 kernel/rcu/tasks.h | 14 +++++++++++---
 1 file changed, 11 insertions(+), 3 deletions(-)

diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
index 6333f4ccf024..56015ced3f37 100644
--- a/kernel/rcu/tasks.h
+++ b/kernel/rcu/tasks.h
@@ -976,6 +976,15 @@ static void rcu_tasks_pregp_step(struct list_head *hop)
 	synchronize_rcu();
 }
 
+static bool rcu_idle_task_is_holdout(struct task_struct *t, int cpu)
+{
+	/* Idle tasks on offline CPUs are RCU-tasks quiescent states. */
+	if (!rcu_cpu_online(cpu))
+		return false;
+
+	return true;
+}
+
 /* Check for quiescent states since the pregp's synchronize_rcu() */
 static bool rcu_tasks_is_holdout(struct task_struct *t)
 {
@@ -995,9 +1004,8 @@ static bool rcu_tasks_is_holdout(struct task_struct *t)
 
 	cpu = task_cpu(t);
 
-	/* Idle tasks on offline CPUs are RCU-tasks quiescent states. */
-	if (t == idle_task(cpu) && !rcu_cpu_online(cpu))
-		return false;
+	if (t == idle_task(cpu))
+		return rcu_idle_task_is_holdout(t, cpu);
 
 	return true;
 }
-- 
2.40.1

[PATCH v2 04/10] rcu/tasks: Create rcu_idle_task_is_holdout() definition for !SMP

[neeraj.upadhyay]

From: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>

rcu_idle_task_is_holdout() is called in rcu_tasks_kthread() context.
As idle tasks cannot be non-voluntary preempted, non-running idle tasks
are not in RCU-tasks critical section. So, idle task is not a RCU-tasks
holdout task on !SMP (which also covers TINY_RCU).

Signed-off-by: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>
---
 kernel/rcu/tasks.h | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
index 56015ced3f37..b794deeaf6d8 100644
--- a/kernel/rcu/tasks.h
+++ b/kernel/rcu/tasks.h
@@ -976,6 +976,7 @@ static void rcu_tasks_pregp_step(struct list_head *hop)
 	synchronize_rcu();
 }
 
+#ifdef CONFIG_SMP
 static bool rcu_idle_task_is_holdout(struct task_struct *t, int cpu)
 {
 	/* Idle tasks on offline CPUs are RCU-tasks quiescent states. */
@@ -984,6 +985,17 @@ static bool rcu_idle_task_is_holdout(struct task_struct *t, int cpu)
 
 	return true;
 }
+#else /* #ifdef CONFIG_SMP */
+static inline bool rcu_idle_task_is_holdout(struct task_struct *t, int cpu)
+{
+	/*
+	 * rcu_idle_task_is_holdout() is called in rcu_tasks_kthread()
+	 * context. Idle thread would have done a voluntary context
+	 * switch.
+	 */
+	return false;
+}
+#endif
 
 /* Check for quiescent states since the pregp's synchronize_rcu() */
 static bool rcu_tasks_is_holdout(struct task_struct *t)
-- 
2.40.1

[PATCH v2 05/10] rcu/tasks: Consider idle tasks not running on CPU as non-holdouts

[neeraj.upadhyay]

From: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>

As idle tasks cannot be non-voluntary preempted, idle tasks which
are not running on CPU are not in RCU-tasks read side critical
section. So, remove them for holdout tasks for RCU-tasks.

Signed-off-by: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>
---
 kernel/rcu/tasks.h | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
index b794deeaf6d8..9523aff6cdae 100644
--- a/kernel/rcu/tasks.h
+++ b/kernel/rcu/tasks.h
@@ -983,6 +983,15 @@ static bool rcu_idle_task_is_holdout(struct task_struct *t, int cpu)
 	if (!rcu_cpu_online(cpu))
 		return false;
 
+	/*
+	 * As idle tasks cannot be involuntary preempted, non-running idle tasks
+	 * are not in RCU-tasks critical section.
+	 * synchronize_rcu() calls in rcu_tasks_pregp_step() and rcu_tasks_postgp()
+	 * ensure that all ->on_cpu transitions are complete.
+	 */
+	if (!t->on_cpu)
+		return false;
+
 	return true;
 }
 #else /* #ifdef CONFIG_SMP */
-- 
2.40.1

[PATCH v2 06/10] rcu/tasks: Check RCU watching state for holdout idle tasks

[neeraj.upadhyay]

From: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>

Use RCU watching state of a CPU to check whether RCU-tasks GP
need to wait for idle task on that CPU. Idle tasks which are
in deep-idle states where RCU is not watching or which have
transitioned to/from deep-idle state do not block RCU-tasks
grace period.

Signed-off-by: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>
---
 kernel/rcu/tasks.h | 36 +++++++++++++++++++++++++++++++++---
 1 file changed, 33 insertions(+), 3 deletions(-)

diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
index 9523aff6cdae..d8506d2e6f54 100644
--- a/kernel/rcu/tasks.h
+++ b/kernel/rcu/tasks.h
@@ -36,6 +36,8 @@ typedef void (*postgp_func_t)(struct rcu_tasks *rtp);
  * @cpu: CPU number corresponding to this entry.
  * @index: Index of this CPU in rtpcp_array of the rcu_tasks structure.
  * @rtpp: Pointer to the rcu_tasks structure.
+ * @rcu_watching_snap: Per-GP RCU-watching snapshot for idle tasks.
+ * @rcu_watching_snap_rec: RCU-watching snapshot recorded for idle task.
  */
 struct rcu_tasks_percpu {
 	struct rcu_segcblist cblist;
@@ -52,6 +54,8 @@ struct rcu_tasks_percpu {
 	int cpu;
 	int index;
 	struct rcu_tasks *rtpp;
+	int rcu_watching_snap;
+	bool rcu_watching_snap_rec;
 };
 
 /**
@@ -957,9 +961,14 @@ static void rcu_tasks_wait_gp(struct rcu_tasks *rtp)
 // rcu_tasks_pregp_step() and by the scheduler's locks and interrupt
 // disabling.
 
+void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
+DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
+
 /* Pre-grace-period preparation. */
 static void rcu_tasks_pregp_step(struct list_head *hop)
 {
+	int cpu;
+
 	/*
 	 * Wait for all pre-existing t->on_rq and t->nvcsw transitions
 	 * to complete.  Invoking synchronize_rcu() suffices because all
@@ -974,11 +983,20 @@ static void rcu_tasks_pregp_step(struct list_head *hop)
 	 * grace period.
 	 */
 	synchronize_rcu();
+
+	/* Initialize watching snapshots for this GP */
+	for_each_possible_cpu(cpu) {
+		struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rcu_tasks.rtpcpu, cpu);
+
+		rtpcp->rcu_watching_snap_rec = false;
+	}
 }
 
 #ifdef CONFIG_SMP
 static bool rcu_idle_task_is_holdout(struct task_struct *t, int cpu)
 {
+	struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rcu_tasks.rtpcpu, cpu);
+
 	/* Idle tasks on offline CPUs are RCU-tasks quiescent states. */
 	if (!rcu_cpu_online(cpu))
 		return false;
@@ -992,6 +1010,21 @@ static bool rcu_idle_task_is_holdout(struct task_struct *t, int cpu)
 	if (!t->on_cpu)
 		return false;
 
+	if (!rtpcp->rcu_watching_snap_rec) {
+		/*
+		 * Do plain access. Ordering between remote CPU's pre idle accesses
+		 * and post rcu-tasks grace period is provided by synchronize_rcu()
+		 * in rcu_tasks_postgp().
+		 */
+		rtpcp->rcu_watching_snap = ct_rcu_watching_cpu(cpu);
+		rtpcp->rcu_watching_snap_rec = true;
+		/* RCU-idle contexts are RCU-tasks quiescent state for idle tasks. */
+		if (rcu_watching_snap_in_eqs(rtpcp->rcu_watching_snap))
+			return false;
+	} else if (rcu_watching_snap_stopped_since(cpu, rtpcp->rcu_watching_snap)) {
+		return false;
+	}
+
 	return true;
 }
 #else /* #ifdef CONFIG_SMP */
@@ -1042,9 +1075,6 @@ static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop)
 	}
 }
 
-void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
-DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
-
 /* Processing between scanning taskslist and draining the holdout list. */
 static void rcu_tasks_postscan(struct list_head *hop)
 {
-- 
2.40.1

[PATCH v2 07/10] rcu/tasks: Check RCU watching state for holdout idle injection tasks

[neeraj.upadhyay]

From: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>

Use RCU watching state of a CPU to check whether RCU-tasks GP
need to wait for idle injection task on that CPU. Idle injection
tasks which are in deep-idle states where RCU is not watching or
which have transitioned to/from deep-idle state do not block
RCU-tasks grace period.

Signed-off-by: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>
---
 kernel/rcu/tasks.h | 63 +++++++++++++++++++++++++++++++++++-----------
 1 file changed, 48 insertions(+), 15 deletions(-)

diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
index d8506d2e6f54..1947f9b6346d 100644
--- a/kernel/rcu/tasks.h
+++ b/kernel/rcu/tasks.h
@@ -38,6 +38,8 @@ typedef void (*postgp_func_t)(struct rcu_tasks *rtp);
  * @rtpp: Pointer to the rcu_tasks structure.
  * @rcu_watching_snap: Per-GP RCU-watching snapshot for idle tasks.
  * @rcu_watching_snap_rec: RCU-watching snapshot recorded for idle task.
+ * @rcu_watching_idle_inj_snap: Per-GP RCU-watching snapshot for idle inject task.
+ * @rcu_watching_idle_inj_rec: RCU-watching snapshot recorded for idle inject task.
  */
 struct rcu_tasks_percpu {
 	struct rcu_segcblist cblist;
@@ -56,6 +58,8 @@ struct rcu_tasks_percpu {
 	struct rcu_tasks *rtpp;
 	int rcu_watching_snap;
 	bool rcu_watching_snap_rec;
+	int rcu_watching_idle_inj_snap;
+	bool rcu_watching_idle_inj_rec;
 };
 
 /**
@@ -989,10 +993,34 @@ static void rcu_tasks_pregp_step(struct list_head *hop)
 		struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rcu_tasks.rtpcpu, cpu);
 
 		rtpcp->rcu_watching_snap_rec = false;
+		rtpcp->rcu_watching_idle_inj_rec = false;
 	}
 }
 
 #ifdef CONFIG_SMP
+static bool rcu_idle_check_rcu_watching(int *rcu_watching_snap, bool *rcu_watching_rec, int cpu)
+{
+	if (!*rcu_watching_rec) {
+		/*
+		 * Do plain access. Ordering between remote CPU's pre idle accesses
+		 * and post rcu-tasks grace period is provided by synchronize_rcu()
+		 * in rcu_tasks_postgp().
+		 */
+		*rcu_watching_snap = ct_rcu_watching_cpu(cpu);
+		*rcu_watching_rec = true;
+		if (rcu_watching_snap_in_eqs(*rcu_watching_snap))
+			/*
+			 * RCU-idle contexts are RCU-tasks quiescent state for idle
+			 * (and idle injection) tasks.
+			 */
+			return false;
+	} else if (rcu_watching_snap_stopped_since(cpu, *rcu_watching_snap)) {
+		return false;
+	}
+
+	return true;
+}
+
 static bool rcu_idle_task_is_holdout(struct task_struct *t, int cpu)
 {
 	struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rcu_tasks.rtpcpu, cpu);
@@ -1010,22 +1038,16 @@ static bool rcu_idle_task_is_holdout(struct task_struct *t, int cpu)
 	if (!t->on_cpu)
 		return false;
 
-	if (!rtpcp->rcu_watching_snap_rec) {
-		/*
-		 * Do plain access. Ordering between remote CPU's pre idle accesses
-		 * and post rcu-tasks grace period is provided by synchronize_rcu()
-		 * in rcu_tasks_postgp().
-		 */
-		rtpcp->rcu_watching_snap = ct_rcu_watching_cpu(cpu);
-		rtpcp->rcu_watching_snap_rec = true;
-		/* RCU-idle contexts are RCU-tasks quiescent state for idle tasks. */
-		if (rcu_watching_snap_in_eqs(rtpcp->rcu_watching_snap))
-			return false;
-	} else if (rcu_watching_snap_stopped_since(cpu, rtpcp->rcu_watching_snap)) {
-		return false;
-	}
+	return rcu_idle_check_rcu_watching(&rtpcp->rcu_watching_snap,
+			&rtpcp->rcu_watching_snap_rec, cpu);
+}
 
-	return true;
+static bool rcu_idle_inj_is_holdout(struct task_struct *t, int cpu)
+{
+	struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rcu_tasks.rtpcpu, cpu);
+
+	return rcu_idle_check_rcu_watching(&rtpcp->rcu_watching_idle_inj_snap,
+			&rtpcp->rcu_watching_idle_inj_rec, cpu);
 }
 #else /* #ifdef CONFIG_SMP */
 static inline bool rcu_idle_task_is_holdout(struct task_struct *t, int cpu)
@@ -1037,6 +1059,15 @@ static inline bool rcu_idle_task_is_holdout(struct task_struct *t, int cpu)
 	 */
 	return false;
 }
+
+static inline bool rcu_idle_inj_is_holdout(struct task_struct *t, int cpu)
+{
+	/*
+	 * Idle injection tasks are PF_IDLE within preempt disabled
+	 * region. So, we should not enter this call for !SMP.
+	 */
+	return false;
+}
 #endif
 
 /* Check for quiescent states since the pregp's synchronize_rcu() */
@@ -1060,6 +1091,8 @@ static bool rcu_tasks_is_holdout(struct task_struct *t)
 
 	if (t == idle_task(cpu))
 		return rcu_idle_task_is_holdout(t, cpu);
+	else if (is_idle_task(t))
+		return rcu_idle_inj_is_holdout(t, cpu);
 
 	return true;
 }
-- 
2.40.1

[PATCH v2 08/10] rcu/tasks: Make RCU-tasks pay attention to idle tasks

[neeraj.upadhyay]

From: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>

Currently, idle tasks are ignored by RCU-tasks. Change this to
start paying attention to idle tasks except in deep-idle functions
where RCU is not watching. With this, for architectures where
kernel entry/exit and deep-idle functions have been properly tagged
noinstr, Tasks Rude RCU can be disabled.

[ neeraj.upadhyay: Frederic Weisbecker and Paul E. McKenney feedback. ]

Signed-off-by: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>
---
 .../RCU/Design/Requirements/Requirements.rst  | 12 +++---
 kernel/rcu/tasks.h                            | 41 ++++++++-----------
 2 files changed, 24 insertions(+), 29 deletions(-)

diff --git a/Documentation/RCU/Design/Requirements/Requirements.rst b/Documentation/RCU/Design/Requirements/Requirements.rst
index 6125e7068d2c..5016b85d53d7 100644
--- a/Documentation/RCU/Design/Requirements/Requirements.rst
+++ b/Documentation/RCU/Design/Requirements/Requirements.rst
@@ -2611,8 +2611,8 @@ critical sections that are delimited by voluntary context switches, that
 is, calls to schedule(), cond_resched(), and
 synchronize_rcu_tasks(). In addition, transitions to and from
 userspace execution also delimit tasks-RCU read-side critical sections.
-Idle tasks are ignored by Tasks RCU, and Tasks Rude RCU may be used to
-interact with them.
+Idle tasks which are idle from RCU's perspective are ignored by Tasks RCU,
+and Tasks Rude RCU may be used to interact with them.
 
 Note well that involuntary context switches are *not* Tasks-RCU quiescent
 states.  After all, in preemptible kernels, a task executing code in a
@@ -2643,10 +2643,10 @@ moniker.  And this operation is considered to be quite rude by real-time
 workloads that don't want their ``nohz_full`` CPUs receiving IPIs and
 by battery-powered systems that don't want their idle CPUs to be awakened.
 
-Once kernel entry/exit and deep-idle functions have been properly tagged
-``noinstr``, Tasks RCU can start paying attention to idle tasks (except
-those that are idle from RCU's perspective) and then Tasks Rude RCU can
-be removed from the kernel.
+As Tasks RCU now pays attention to idle tasks (except those that are idle
+from RCU's perspective), once kernel entry/exit and deep-idle functions have
+been properly tagged ``noinstr``, Tasks Rude RCU can be removed from the
+kernel.
 
 The tasks-rude-RCU API is also reader-marking-free and thus quite compact,
 consisting solely of synchronize_rcu_tasks_rude().
diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
index 1947f9b6346d..72dc0d0a4a8f 100644
--- a/kernel/rcu/tasks.h
+++ b/kernel/rcu/tasks.h
@@ -912,14 +912,15 @@ static void rcu_tasks_wait_gp(struct rcu_tasks *rtp)
 ////////////////////////////////////////////////////////////////////////
 //
 // Simple variant of RCU whose quiescent states are voluntary context
-// switch, cond_resched_tasks_rcu_qs(), user-space execution, and idle.
-// As such, grace periods can take one good long time.  There are no
-// read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
-// because this implementation is intended to get the system into a safe
-// state for some of the manipulations involved in tracing and the like.
-// Finally, this implementation does not support high call_rcu_tasks()
-// rates from multiple CPUs.  If this is required, per-CPU callback lists
-// will be needed.
+// switch, cond_resched_tasks_rcu_qs(), user-space execution, and idle
+// tasks which are in RCU-idle context. As such, grace periods can take
+// one good long time.  There are no read-side primitives similar to
+// rcu_read_lock() and rcu_read_unlock() because this implementation is
+// intended to get the system into a safe state for some of the
+// manipulations involved in tracing and the like. Finally, this
+// implementation does not support high call_rcu_tasks() rates from
+// multiple CPUs.  If this is required, per-CPU callback lists will be
+// needed.
 //
 // The implementation uses rcu_tasks_wait_gp(), which relies on function
 // pointers in the rcu_tasks structure.  The rcu_spawn_tasks_kthread()
@@ -1079,14 +1080,6 @@ static bool rcu_tasks_is_holdout(struct task_struct *t)
 	if (!READ_ONCE(t->on_rq))
 		return false;
 
-	/*
-	 * Idle tasks (or idle injection) within the idle loop are RCU-tasks
-	 * quiescent states. But CPU boot code performed by the idle task
-	 * isn't a quiescent state.
-	 */
-	if (is_idle_task(t))
-		return false;
-
 	cpu = task_cpu(t);
 
 	if (t == idle_task(cpu))
@@ -1265,11 +1258,12 @@ static void tasks_rcu_exit_srcu_stall(struct timer_list *unused)
  * period elapses, in other words after all currently executing RCU
  * read-side critical sections have completed. call_rcu_tasks() assumes
  * that the read-side critical sections end at a voluntary context
- * switch (not a preemption!), cond_resched_tasks_rcu_qs(), entry into idle,
- * or transition to usermode execution.  As such, there are no read-side
- * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
- * this primitive is intended to determine that all tasks have passed
- * through a safe state, not so much for data-structure synchronization.
+ * switch (not a preemption!), cond_resched_tasks_rcu_qs(), entry into
+ * RCU-idle context or transition to usermode execution. As such, there
+ * are no read-side primitives analogous to rcu_read_lock() and
+ * rcu_read_unlock() because this primitive is intended to determine
+ * that all tasks have passed through a safe state, not so much for
+ * data-structure synchronization.
  *
  * See the description of call_rcu() for more detailed information on
  * memory ordering guarantees.
@@ -1287,8 +1281,9 @@ EXPORT_SYMBOL_GPL(call_rcu_tasks);
  * grace period has elapsed, in other words after all currently
  * executing rcu-tasks read-side critical sections have elapsed.  These
  * read-side critical sections are delimited by calls to schedule(),
- * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
- * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
+ * cond_resched_tasks_rcu_qs(), idle execution within RCU-idle context,
+ * userspace execution, calls to synchronize_rcu_tasks(), and (in theory,
+ * anyway) cond_resched().
  *
  * This is a very specialized primitive, intended only for a few uses in
  * tracing and other situations requiring manipulation of function
-- 
2.40.1

[PATCH v2 09/10] context_tracking: Invoke RCU-tasks enter/exit for NMI context

[neeraj.upadhyay]

From: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>

rcu_task_enter() and rcu_task_exit() are not called on NMI
entry and exit. So, Tasks-RCU-Rude grace period wait is required to
ensure that NMI handlers have entered/exited into Tasks-RCU eqs.
For architectures which do not require Tasks-RCU-Rude (as the code
sections where RCU is not watching are marked as noinstr), when
those architectures switch to not using Tasks-RCU-Rude, NMI handlers
task exit to eqs will need to be handled correctly for Tasks-RCU holdout
tasks running on nohz_full CPUs. As it is safe to call these two
functions from NMI context, remove the in_nmi() check. This ensures
that RCU-tasks entry/exit is marked correctly for NMI handlers.
With this check removed, all callers of ct_kernel_exit_state() and
ct_kernel_enter_state() now also call rcu_task_exit() and
rcu_task_enter() respectively. So, fold rcu_task_exit() and
rcu_task_entry() calls into ct_kernel_exit_state() and
ct_kernel_enter_state().

Reported-by: Frederic Weisbecker <frederic@kernel.org>
Suggested-by: Frederic Weisbecker <frederic@kernel.org>
Suggested-by: "Paul E. McKenney" <paulmck@kernel.org>
Reviewed-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>
---
 kernel/context_tracking.c | 11 +++--------
 1 file changed, 3 insertions(+), 8 deletions(-)

diff --git a/kernel/context_tracking.c b/kernel/context_tracking.c
index 938c48952d26..85ced563af23 100644
--- a/kernel/context_tracking.c
+++ b/kernel/context_tracking.c
@@ -91,6 +91,7 @@ static noinstr void ct_kernel_exit_state(int offset)
 	seq = ct_state_inc(offset);
 	// RCU is no longer watching.  Better be in extended quiescent state!
 	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && (seq & CT_RCU_WATCHING));
+	rcu_task_exit();
 }
 
 /*
@@ -102,6 +103,8 @@ static noinstr void ct_kernel_enter_state(int offset)
 {
 	int seq;
 
+	rcu_task_enter();
+
 	/*
 	 * CPUs seeing atomic_add_return() must see prior idle sojourns,
 	 * and we also must force ordering with the next RCU read-side
@@ -149,7 +152,6 @@ static void noinstr ct_kernel_exit(bool user, int offset)
 	// RCU is watching here ...
 	ct_kernel_exit_state(offset);
 	// ... but is no longer watching here.
-	rcu_task_exit();
 }
 
 /*
@@ -173,7 +175,6 @@ static void noinstr ct_kernel_enter(bool user, int offset)
 		ct->nesting++;
 		return;
 	}
-	rcu_task_enter();
 	// RCU is not watching here ...
 	ct_kernel_enter_state(offset);
 	// ... but is watching here.
@@ -238,9 +239,6 @@ void noinstr ct_nmi_exit(void)
 	// RCU is watching here ...
 	ct_kernel_exit_state(CT_RCU_WATCHING);
 	// ... but is no longer watching here.
-
-	if (!in_nmi())
-		rcu_task_exit();
 }
 
 /**
@@ -273,9 +271,6 @@ void noinstr ct_nmi_enter(void)
 	 */
 	if (!rcu_is_watching_curr_cpu()) {
 
-		if (!in_nmi())
-			rcu_task_enter();
-
 		// RCU is not watching here ...
 		ct_kernel_enter_state(CT_RCU_WATCHING);
 		// ... but is watching here.
-- 
2.40.1

[PATCH v2 10/10] rcu: Allow short-circuiting of synchronize_rcu_tasks_rude()

[neeraj.upadhyay]

From: "Paul E. McKenney" <paulmck@kernel.org>

There are now architectures for which all deep-idle and entry-exit
functions are properly inlined or marked noinstr.  Such architectures do
not need synchronize_rcu_tasks_rude(), or will not once RCU Tasks has
been modified to pay attention to idle tasks.  This commit therefore
allows a CONFIG_ARCH_HAS_NOINSTR_MARKINGS Kconfig option to turn
synchronize_rcu_tasks_rude() into a no-op.

To facilitate testing, kernels built by rcutorture scripting will enable
RCU Tasks Trace even on systems that do not need it.

[ paulmck: Apply Peter Zijlstra feedback. ]

Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>
---
 kernel/rcu/tasks.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
index 72dc0d0a4a8f..ef9de6b91a3d 100644
--- a/kernel/rcu/tasks.h
+++ b/kernel/rcu/tasks.h
@@ -1485,7 +1485,8 @@ static void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
  */
 void synchronize_rcu_tasks_rude(void)
 {
-	synchronize_rcu_tasks_generic(&rcu_tasks_rude);
+	if (!IS_ENABLED(CONFIG_ARCH_WANTS_NO_INSTR) || IS_ENABLED(CONFIG_FORCE_TASKS_RUDE_RCU))
+		synchronize_rcu_tasks_generic(&rcu_tasks_rude);
 }
 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
 
-- 
2.40.1

Re: [PATCH v2 07/10] rcu/tasks: Check RCU watching state for holdout idle injection tasks

[Frederic Weisbecker]

Le Wed, Oct 09, 2024 at 06:21:24PM +0530, neeraj.upadhyay@kernel.org a écrit :
> From: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>
> 
> Use RCU watching state of a CPU to check whether RCU-tasks GP
> need to wait for idle injection task on that CPU. Idle injection
> tasks which are in deep-idle states where RCU is not watching or
> which have transitioned to/from deep-idle state do not block
> RCU-tasks grace period.
> 
> Signed-off-by: Neeraj Upadhyay <neeraj.upadhyay@kernel.org>

For now this should work because there is a single user that is
a per-cpu kthread, therefore no RCU-watching writer can race
against another (real idle VS idle injection or idle_injection VS
idle injection) without going first through a voluntary context
switch. But who knows about the future? If an idle injection kthread
is preempted by another idle injection right after clearing PF_IDLE,
there could be some spurious QS accounted for the preempted
kthread.

So perhaps we can consider idle injection as any normal task and
wait for it to voluntary schedule?

Well I see DEFAULT_DURATION_JIFFIES = 6, which is 60 ms on HZ=100.

Yeah that's a lot...so perhaps this patch is needed after all...

> ---
>  kernel/rcu/tasks.h | 63 +++++++++++++++++++++++++++++++++++-----------
>  1 file changed, 48 insertions(+), 15 deletions(-)
> 
> diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
> index d8506d2e6f54..1947f9b6346d 100644
> --- a/kernel/rcu/tasks.h
> +++ b/kernel/rcu/tasks.h
> @@ -38,6 +38,8 @@ typedef void (*postgp_func_t)(struct rcu_tasks *rtp);
>   * @rtpp: Pointer to the rcu_tasks structure.
>   * @rcu_watching_snap: Per-GP RCU-watching snapshot for idle tasks.
>   * @rcu_watching_snap_rec: RCU-watching snapshot recorded for idle task.
> + * @rcu_watching_idle_inj_snap: Per-GP RCU-watching snapshot for idle inject task.
> + * @rcu_watching_idle_inj_rec: RCU-watching snapshot recorded for idle inject task.
>   */
>  struct rcu_tasks_percpu {
>  	struct rcu_segcblist cblist;
> @@ -56,6 +58,8 @@ struct rcu_tasks_percpu {
>  	struct rcu_tasks *rtpp;
>  	int rcu_watching_snap;
>  	bool rcu_watching_snap_rec;
> +	int rcu_watching_idle_inj_snap;
> +	bool rcu_watching_idle_inj_rec;

So how about:

struct rcu_watching_task {
    int snap;
    bool rec;
}
...
struct rcu_tasks_percpu {
       ...
       struct rcu_watching_task idle_task;
       struct rcu_watching_task idle_inject;
}

Thanks.