sched_ext: Implement SCX_OPS_ALLOW_QUEUED_WAKEUP

A task wakeup can be either processed on the waker's CPU or bounced to the
wakee's previous CPU using an IPI (ttwu_queue). Bouncing to the wakee's CPU
avoids the waker's CPU locking and accessing the wakee's rq which can be
expensive across cache and node boundaries.

When ttwu_queue path is taken, select_task_rq() and thus ops.select_cpu()
may be skipped in some cases (racing against the wakee switching out). As
this confused some BPF schedulers, there wasn't a good way for a BPF
scheduler to tell whether idle CPU selection has been skipped, ops.enqueue()
couldn't insert tasks into foreign local DSQs, and the performance
difference on machines with simple toplogies were minimal, sched_ext
disabled ttwu_queue.

However, this optimization makes noticeable difference on more complex
topologies and a BPF scheduler now has an easy way tell whether
ops.select_cpu() was skipped since 9b671793c7d9 ("sched_ext, scx_qmap: Add
and use SCX_ENQ_CPU_SELECTED") and can insert tasks into foreign local DSQs
since 5b26f7b920f7 ("sched_ext: Allow SCX_DSQ_LOCAL_ON for direct
dispatches").

Implement SCX_OPS_ALLOW_QUEUED_WAKEUP which allows BPF schedulers to choose
to enable ttwu_queue optimization.

v2: Update the patch description and comment re. ops.select_cpu() being
    skipped in some cases as opposed to always as per Neel.

Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Neel Natu <neelnatu@google.com>
Reported-by: Barret Rhoden <brho@google.com>
Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Andrea Righi <arighi@nvidia.com>

1 files changed, 26 insertions, 6 deletions

diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
index 98d5f2f68f38..2e1a1e4fc304 100644
--- a/kernel/sched/ext.c
+++ b/kernel/sched/ext.c
@@ -96,7 +96,7 @@ enum scx_ops_flags {
 	/*
 	 * Keep built-in idle tracking even if ops.update_idle() is implemented.
 	 */
-	SCX_OPS_KEEP_BUILTIN_IDLE = 1LLU << 0,
+	SCX_OPS_KEEP_BUILTIN_IDLE	= 1LLU << 0,
 
 	/*
 	 * By default, if there are no other task to run on the CPU, ext core
@@ -104,7 +104,7 @@ enum scx_ops_flags {
 	 * flag is specified, such tasks are passed to ops.enqueue() with
 	 * %SCX_ENQ_LAST. See the comment above %SCX_ENQ_LAST for more info.
 	 */
-	SCX_OPS_ENQ_LAST	= 1LLU << 1,
+	SCX_OPS_ENQ_LAST		= 1LLU << 1,
 
 	/*
 	 * An exiting task may schedule after PF_EXITING is set. In such cases,
@@ -117,13 +117,13 @@ enum scx_ops_flags {
 	 * depend on pid lookups and wants to handle these tasks directly, the
 	 * following flag can be used.
 	 */
-	SCX_OPS_ENQ_EXITING	= 1LLU << 2,
+	SCX_OPS_ENQ_EXITING		= 1LLU << 2,
 
 	/*
 	 * If set, only tasks with policy set to SCHED_EXT are attached to
 	 * sched_ext. If clear, SCHED_NORMAL tasks are also included.
 	 */
-	SCX_OPS_SWITCH_PARTIAL	= 1LLU << 3,
+	SCX_OPS_SWITCH_PARTIAL		= 1LLU << 3,
 
 	/*
 	 * A migration disabled task can only execute on its current CPU. By
@@ -136,7 +136,23 @@ enum scx_ops_flags {
 	 * current CPU while p->nr_cpus_allowed keeps tracking p->user_cpus_ptr
 	 * and thus may disagree with cpumask_weight(p->cpus_ptr).
 	 */
-	SCX_OPS_ENQ_MIGRATION_DISABLED = 1LLU << 4,
+	SCX_OPS_ENQ_MIGRATION_DISABLED	= 1LLU << 4,
+
+	/*
+	 * Queued wakeup (ttwu_queue) is a wakeup optimization that invokes
+	 * ops.enqueue() on the ops.select_cpu() selected or the wakee's
+	 * previous CPU via IPI (inter-processor interrupt) to reduce cacheline
+	 * transfers. When this optimization is enabled, ops.select_cpu() is
+	 * skipped in some cases (when racing against the wakee switching out).
+	 * As the BPF scheduler may depend on ops.select_cpu() being invoked
+	 * during wakeups, queued wakeup is disabled by default.
+	 *
+	 * If this ops flag is set, queued wakeup optimization is enabled and
+	 * the BPF scheduler must be able to handle ops.enqueue() invoked on the
+	 * wakee's CPU without preceding ops.select_cpu() even for tasks which
+	 * may be executed on multiple CPUs.
+	 */
+	SCX_OPS_ALLOW_QUEUED_WAKEUP	= 1LLU << 5,
 
 	/*
 	 * CPU cgroup support flags
@@ -147,6 +163,7 @@ enum scx_ops_flags {
 				  SCX_OPS_ENQ_LAST |
 				  SCX_OPS_ENQ_EXITING |
 				  SCX_OPS_ENQ_MIGRATION_DISABLED |
+				  SCX_OPS_ALLOW_QUEUED_WAKEUP |
 				  SCX_OPS_SWITCH_PARTIAL |
 				  SCX_OPS_HAS_CGROUP_WEIGHT,
 };
@@ -897,6 +914,7 @@ DEFINE_STATIC_KEY_FALSE(__scx_switched_all);
 static struct sched_ext_ops scx_ops;
 static bool scx_warned_zero_slice;
 
+DEFINE_STATIC_KEY_FALSE(scx_ops_allow_queued_wakeup);
 static DEFINE_STATIC_KEY_FALSE(scx_ops_enq_last);
 static DEFINE_STATIC_KEY_FALSE(scx_ops_enq_exiting);
 static DEFINE_STATIC_KEY_FALSE(scx_ops_enq_migration_disabled);
@@ -4717,6 +4735,7 @@ static void scx_ops_disable_workfn(struct kthread_work *work)
 	static_branch_disable(&__scx_ops_enabled);
 	for (i = SCX_OPI_BEGIN; i < SCX_OPI_END; i++)
 		static_branch_disable(&scx_has_op[i]);
+	static_branch_disable(&scx_ops_allow_queued_wakeup);
 	static_branch_disable(&scx_ops_enq_last);
 	static_branch_disable(&scx_ops_enq_exiting);
 	static_branch_disable(&scx_ops_enq_migration_disabled);
@@ -5348,9 +5367,10 @@ static int scx_ops_enable(struct sched_ext_ops *ops, struct bpf_link *link)
 		if (((void (**)(void))ops)[i])
 			static_branch_enable(&scx_has_op[i]);
 
+	if (ops->flags & SCX_OPS_ALLOW_QUEUED_WAKEUP)
+		static_branch_enable(&scx_ops_allow_queued_wakeup);
 	if (ops->flags & SCX_OPS_ENQ_LAST)
 		static_branch_enable(&scx_ops_enq_last);
-
 	if (ops->flags & SCX_OPS_ENQ_EXITING)
 		static_branch_enable(&scx_ops_enq_exiting);
 	if (ops->flags & SCX_OPS_ENQ_MIGRATION_DISABLED)