From d37e428dd8222a07a9438839c190bb1aa6e5101a Mon Sep 17 00:00:00 2001
From: "Paul E. McKenney" <paulmck@kernel.org>
Date: Wed, 24 Nov 2021 15:12:15 -0800
Subject: [PATCH] rcu-tasks: Use more callback queues if contention encountered

References: bnc#1189998 (PREEMPT_RT prerequisite backports)
Patch-mainline: v5.17-rc0
Git-commit: ab97152f88a4d580b89f0b7cc3028ffac438216f

The rcupdate.rcu_task_enqueue_lim module parameter allows system
administrators to tune the number of callback queues used by the RCU
Tasks flavors.  However if callback storms are infrequent, it would
be better to operate with a single queue on a given system unless and
until that system actually needed more queues.  Systems not needing
more queues can then avoid the overhead of checking the extra queues
and especially avoid the overhead of fanning workqueue handlers out to
all CPUs to invoke callbacks.

This commit therefore switches to using all the CPUs' callback queues if
call_rcu_tasks_generic() encounters too much lock contention.  The amount
of lock contention to tolerate defaults to 100 contended lock acquisitions
per jiffy, and can be adjusted using the new rcupdate.rcu_task_contend_lim
module parameter.

Such switching is undertaken only if the rcupdate.rcu_task_enqueue_lim
module parameter is negative, which is its default value (-1).
This allows savvy systems administrators to set the number of queues
to some known good value and to not have to worry about the kernel doing
any second guessing.

[ paulmck: Apply feedback from Guillaume Tucker and kernelci. ]

Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Mel Gorman <mgorman@suse.de>
---
 Documentation/admin-guide/kernel-parameters.txt |  8 ++++++++
 kernel/rcu/tasks.h                              | 27 +++++++++++++++++++++----
 2 files changed, 31 insertions(+), 4 deletions(-)

diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
index b50a553a093c..72b4d08932a6 100644
--- a/Documentation/admin-guide/kernel-parameters.txt
+++ b/Documentation/admin-guide/kernel-parameters.txt
@@ -4758,6 +4758,14 @@
 			period to instead use normal non-expedited
 			grace-period processing.
 
+	rcupdate.rcu_task_contend_lim= [KNL]
+			Set the minimum number of callback-queuing-time
+			lock-contention events per jiffy required to
+			cause the RCU Tasks flavors to switch to per-CPU
+			callback queuing.  This switching only occurs
+			when rcupdate.rcu_task_enqueue_lim is set to
+			the default value of -1.
+
 	rcupdate.rcu_task_enqueue_lim= [KNL]
 			Set the number of callback queues to use for the
 			RCU Tasks family of RCU flavors.  The default
diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
index 6e4ab5b6bcfa..3d6b7c549fe3 100644
--- a/kernel/rcu/tasks.h
+++ b/kernel/rcu/tasks.h
@@ -142,6 +142,10 @@ module_param(rcu_task_stall_timeout, int, 0644);
 static int rcu_task_enqueue_lim __read_mostly = -1;
 module_param(rcu_task_enqueue_lim, int, 0444);
 
+static bool rcu_task_cb_adjust;
+static int rcu_task_contend_lim __read_mostly = 100;
+module_param(rcu_task_contend_lim, int, 0444);
+
 /* RCU tasks grace-period state for debugging. */
 #define RTGS_INIT		 0
 #define RTGS_WAIT_WAIT_CBS	 1
@@ -207,10 +211,13 @@ static void cblist_init_generic(struct rcu_tasks *rtp)
 	int lim;
 
 	raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
-	if (rcu_task_enqueue_lim < 0)
-		rcu_task_enqueue_lim = nr_cpu_ids;
-	else if (rcu_task_enqueue_lim == 0)
+	if (rcu_task_enqueue_lim < 0) {
+		rcu_task_enqueue_lim = 1;
+		rcu_task_cb_adjust = true;
+		pr_info("%s: Setting adjustable number of callback queues.\n", __func__);
+	} else if (rcu_task_enqueue_lim == 0) {
 		rcu_task_enqueue_lim = 1;
+	}
 	lim = rcu_task_enqueue_lim;
 
 	if (lim > nr_cpu_ids)
@@ -251,6 +258,7 @@ static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
 {
 	unsigned long flags;
 	unsigned long j;
+	bool needadjust = false;
 	bool needwake;
 	struct rcu_tasks_percpu *rtpcp;
 
@@ -266,7 +274,9 @@ static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
 			rtpcp->rtp_jiffies = j;
 			rtpcp->rtp_n_lock_retries = 0;
 		}
-		rtpcp->rtp_n_lock_retries++;
+		if (rcu_task_cb_adjust && ++rtpcp->rtp_n_lock_retries > rcu_task_contend_lim &&
+		    READ_ONCE(rtp->percpu_enqueue_lim) != nr_cpu_ids)
+			needadjust = true;  // Defer adjustment to avoid deadlock.
 	}
 	if (!rcu_segcblist_is_enabled(&rtpcp->cblist)) {
 		raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled.
@@ -276,6 +286,15 @@ static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
 	needwake = rcu_segcblist_empty(&rtpcp->cblist);
 	rcu_segcblist_enqueue(&rtpcp->cblist, rhp);
 	raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
+	if (unlikely(needadjust)) {
+		raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
+		if (rtp->percpu_enqueue_lim != nr_cpu_ids) {
+			WRITE_ONCE(rtp->percpu_enqueue_shift, ilog2(nr_cpu_ids));
+			smp_store_release(&rtp->percpu_enqueue_lim, nr_cpu_ids);
+			pr_info("Switching %s to per-CPU callback queuing.\n", rtp->name);
+		}
+		raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
+	}
 	/* We can't create the thread unless interrupts are enabled. */
 	if (needwake && READ_ONCE(rtp->kthread_ptr))
 		irq_work_queue(&rtpcp->rtp_irq_work);