sched_ext: Allocate scx_kick_cpus_pnt_seqs lazily using kvzalloc()

On systems with >4096 CPUs, scx_kick_cpus_pnt_seqs allocation fails during
boot because it exceeds the 32,768 byte percpu allocator limit.

Restructure to use DEFINE_PER_CPU() for the per-CPU pointers, with each CPU
pointing to its own kvzalloc'd array. Move allocation from boot time to
scx_enable() and free in scx_disable(), so the O(nr_cpu_ids^2) memory is only
consumed when sched_ext is active.

Use RCU to guard against racing with free. Arrays are freed via call_rcu()
and kick_cpus_irq_workfn() uses rcu_dereference_bh() with a NULL check.

While at it, rename to scx_kick_pseqs for brevity and update comments to
clarify these are pick_task sequence numbers.

v2: RCU protect scx_kick_seqs to manage kick_cpus_irq_workfn() racing
    against disable as per Andrea.

v3: Fix bugs notcied by Andrea.

Reported-by: Phil Auld <pauld@redhat.com>
Link: http://lkml.kernel.org/r/20251007133523.GA93086@pauld.westford.csb
Cc: Andrea Righi <arighi@nvidia.com>
Reviewed-by: Emil Tsalapatis <emil@etsalapatis.com>
Reviewed-by: Phil Auld <pauld@redhat.com>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>

1 files changed, 79 insertions, 10 deletions

diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
index 1352e6a5b089..c645d47124e7 100644
--- a/kernel/sched/ext.c
+++ b/kernel/sched/ext.c
@@ -67,8 +67,19 @@ static unsigned long scx_watchdog_timestamp = INITIAL_JIFFIES;
 
 static struct delayed_work scx_watchdog_work;
 
-/* for %SCX_KICK_WAIT */
-static unsigned long __percpu *scx_kick_cpus_pnt_seqs;
+/*
+ * For %SCX_KICK_WAIT: Each CPU has a pointer to an array of pick_task sequence
+ * numbers. The arrays are allocated with kvzalloc() as size can exceed percpu
+ * allocator limits on large machines. O(nr_cpu_ids^2) allocation, allocated
+ * lazily when enabling and freed when disabling to avoid waste when sched_ext
+ * isn't active.
+ */
+struct scx_kick_pseqs {
+	struct rcu_head		rcu;
+	unsigned long		seqs[];
+};
+
+static DEFINE_PER_CPU(struct scx_kick_pseqs __rcu *, scx_kick_pseqs);
 
 /*
  * Direct dispatch marker.
@@ -3877,6 +3888,27 @@ static const char *scx_exit_reason(enum scx_exit_kind kind)
 	}
 }
 
+static void free_kick_pseqs_rcu(struct rcu_head *rcu)
+{
+	struct scx_kick_pseqs *pseqs = container_of(rcu, struct scx_kick_pseqs, rcu);
+
+	kvfree(pseqs);
+}
+
+static void free_kick_pseqs(void)
+{
+	int cpu;
+
+	for_each_possible_cpu(cpu) {
+		struct scx_kick_pseqs **pseqs = per_cpu_ptr(&scx_kick_pseqs, cpu);
+		struct scx_kick_pseqs *to_free;
+
+		to_free = rcu_replace_pointer(*pseqs, NULL, true);
+		if (to_free)
+			call_rcu(&to_free->rcu, free_kick_pseqs_rcu);
+	}
+}
+
 static void scx_disable_workfn(struct kthread_work *work)
 {
 	struct scx_sched *sch = container_of(work, struct scx_sched, disable_work);
@@ -4013,6 +4045,7 @@ static void scx_disable_workfn(struct kthread_work *work)
 	free_percpu(scx_dsp_ctx);
 	scx_dsp_ctx = NULL;
 	scx_dsp_max_batch = 0;
+	free_kick_pseqs();
 
 	mutex_unlock(&scx_enable_mutex);
 
@@ -4375,6 +4408,33 @@ static void scx_vexit(struct scx_sched *sch,
 	irq_work_queue(&sch->error_irq_work);
 }
 
+static int alloc_kick_pseqs(void)
+{
+	int cpu;
+
+	/*
+	 * Allocate per-CPU arrays sized by nr_cpu_ids. Use kvzalloc as size
+	 * can exceed percpu allocator limits on large machines.
+	 */
+	for_each_possible_cpu(cpu) {
+		struct scx_kick_pseqs **pseqs = per_cpu_ptr(&scx_kick_pseqs, cpu);
+		struct scx_kick_pseqs *new_pseqs;
+
+		WARN_ON_ONCE(rcu_access_pointer(*pseqs));
+
+		new_pseqs = kvzalloc_node(struct_size(new_pseqs, seqs, nr_cpu_ids),
+					  GFP_KERNEL, cpu_to_node(cpu));
+		if (!new_pseqs) {
+			free_kick_pseqs();
+			return -ENOMEM;
+		}
+
+		rcu_assign_pointer(*pseqs, new_pseqs);
+	}
+
+	return 0;
+}
+
 static struct scx_sched *scx_alloc_and_add_sched(struct sched_ext_ops *ops)
 {
 	struct scx_sched *sch;
@@ -4517,15 +4577,19 @@ static int scx_enable(struct sched_ext_ops *ops, struct bpf_link *link)
 
 	mutex_lock(&scx_enable_mutex);
 
+	ret = alloc_kick_pseqs();
+	if (ret)
+		goto err_unlock;
+
 	if (scx_enable_state() != SCX_DISABLED) {
 		ret = -EBUSY;
-		goto err_unlock;
+		goto err_free_pseqs;
 	}
 
 	sch = scx_alloc_and_add_sched(ops);
 	if (IS_ERR(sch)) {
 		ret = PTR_ERR(sch);
-		goto err_unlock;
+		goto err_free_pseqs;
 	}
 
 	/*
@@ -4728,6 +4792,8 @@ static int scx_enable(struct sched_ext_ops *ops, struct bpf_link *link)
 
 	return 0;
 
+err_free_pseqs:
+	free_kick_pseqs();
 err_unlock:
 	mutex_unlock(&scx_enable_mutex);
 	return ret;
@@ -5109,10 +5175,18 @@ static void kick_cpus_irq_workfn(struct irq_work *irq_work)
 {
 	struct rq *this_rq = this_rq();
 	struct scx_rq *this_scx = &this_rq->scx;
-	unsigned long *pseqs = this_cpu_ptr(scx_kick_cpus_pnt_seqs);
+	struct scx_kick_pseqs __rcu *pseqs_pcpu = __this_cpu_read(scx_kick_pseqs);
 	bool should_wait = false;
+	unsigned long *pseqs;
 	s32 cpu;
 
+	if (unlikely(!pseqs_pcpu)) {
+		pr_warn_once("kick_cpus_irq_workfn() called with NULL scx_kick_pseqs");
+		return;
+	}
+
+	pseqs = rcu_dereference_bh(pseqs_pcpu)->seqs;
+
 	for_each_cpu(cpu, this_scx->cpus_to_kick) {
 		should_wait |= kick_one_cpu(cpu, this_rq, pseqs);
 		cpumask_clear_cpu(cpu, this_scx->cpus_to_kick);
@@ -5235,11 +5309,6 @@ void __init init_sched_ext_class(void)
 
 	scx_idle_init_masks();
 
-	scx_kick_cpus_pnt_seqs =
-		__alloc_percpu(sizeof(scx_kick_cpus_pnt_seqs[0]) * nr_cpu_ids,
-			       __alignof__(scx_kick_cpus_pnt_seqs[0]));
-	BUG_ON(!scx_kick_cpus_pnt_seqs);
-
 	for_each_possible_cpu(cpu) {
 		struct rq *rq = cpu_rq(cpu);
 		int  n = cpu_to_node(cpu);