sched_ext: Add a DL server for sched_ext tasks

sched_ext currently suffers starvation due to RT. The same workload when
converted to EXT can get zero runtime if RT is 100% running, causing EXT
processes to stall. Fix it by adding a DL server for EXT.

A kselftest is also included later to confirm that both DL servers are
functioning correctly:

 # ./runner -t rt_stall
 ===== START =====
 TEST: rt_stall
 DESCRIPTION: Verify that RT tasks cannot stall SCHED_EXT tasks
 OUTPUT:
 TAP version 13
 1..1
 # Runtime of FAIR task (PID 1511) is 0.250000 seconds
 # Runtime of RT task (PID 1512) is 4.750000 seconds
 # FAIR task got 5.00% of total runtime
 ok 1 PASS: FAIR task got more than 4.00% of runtime
 TAP version 13
 1..1
 # Runtime of EXT task (PID 1514) is 0.250000 seconds
 # Runtime of RT task (PID 1515) is 4.750000 seconds
 # EXT task got 5.00% of total runtime
 ok 2 PASS: EXT task got more than 4.00% of runtime
 TAP version 13
 1..1
 # Runtime of FAIR task (PID 1517) is 0.250000 seconds
 # Runtime of RT task (PID 1518) is 4.750000 seconds
 # FAIR task got 5.00% of total runtime
 ok 3 PASS: FAIR task got more than 4.00% of runtime
 TAP version 13
 1..1
 # Runtime of EXT task (PID 1521) is 0.250000 seconds
 # Runtime of RT task (PID 1522) is 4.750000 seconds
 # EXT task got 5.00% of total runtime
 ok 4 PASS: EXT task got more than 4.00% of runtime
 ok 1 rt_stall #
 =====  END  =====

Co-developed-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Juri Lelli <juri.lelli@redhat.com>
Tested-by: Christian Loehle <christian.loehle@arm.com>
Link: https://patch.msgid.link/20260126100050.3854740-5-arighi@nvidia.com

Author: arighi

kernel/sched/core.c

@@ -8484,6 +8484,9 @@ int sched_cpu_dying(unsigned int cpu)
 		dump_rq_tasks(rq, KERN_WARNING);
 	}
 	dl_server_stop(&rq->fair_server);
+#ifdef CONFIG_SCHED_CLASS_EXT
+	dl_server_stop(&rq->ext_server);
+#endif
 	rq_unlock_irqrestore(rq, &rf);
 
 	calc_load_migrate(rq);
@@ -8689,6 +8692,9 @@ void __init sched_init(void)
 		hrtick_rq_init(rq);
 		atomic_set(&rq->nr_iowait, 0);
 		fair_server_init(rq);
+#ifdef CONFIG_SCHED_CLASS_EXT
+		ext_server_init(rq);
+#endif
 
 #ifdef CONFIG_SCHED_CORE
 		rq->core = rq;

kernel/sched/deadline.c

@@ -1449,8 +1449,8 @@ static void update_curr_dl_se(struct rq *rq, struct sched_dl_entity *dl_se, s64
 		dl_se->dl_defer_idle = 0;
 
 	/*
-	 * The fair server can consume its runtime while throttled (not queued/
-	 * running as regular CFS).
+	 * The DL server can consume its runtime while throttled (not
+	 * queued / running as regular CFS).
 	 *
 	 * If the server consumes its entire runtime in this state. The server
 	 * is not required for the current period. Thus, reset the server by
@@ -1535,10 +1535,10 @@ static void update_curr_dl_se(struct rq *rq, struct sched_dl_entity *dl_se, s64
 	}
 
 	/*
-	 * The fair server (sole dl_server) does not account for real-time
-	 * workload because it is running fair work.
+	 * The dl_server does not account for real-time workload because it
+	 * is running fair work.
 	 */
-	if (dl_se == &rq->fair_server)
+	if (dl_se->dl_server)
 		return;
 
 #ifdef CONFIG_RT_GROUP_SCHED
@@ -1573,9 +1573,9 @@ static void update_curr_dl_se(struct rq *rq, struct sched_dl_entity *dl_se, s64
  * In the non-defer mode, the idle time is not accounted, as the
  * server provides a guarantee.
  *
- * If the dl_server is in defer mode, the idle time is also considered
- * as time available for the fair server, avoiding a penalty for the
- * rt scheduler that did not consumed that time.
+ * If the dl_server is in defer mode, the idle time is also considered as
+ * time available for the dl_server, avoiding a penalty for the rt
+ * scheduler that did not consumed that time.
  */
 void dl_server_update_idle(struct sched_dl_entity *dl_se, s64 delta_exec)
 {
@@ -1860,6 +1860,18 @@ void sched_init_dl_servers(void)
 		dl_se->dl_server = 1;
 		dl_se->dl_defer = 1;
 		setup_new_dl_entity(dl_se);
+
+#ifdef CONFIG_SCHED_CLASS_EXT
+		dl_se = &rq->ext_server;
+
+		WARN_ON(dl_server(dl_se));
+
+		dl_server_apply_params(dl_se, runtime, period, 1);
+
+		dl_se->dl_server = 1;
+		dl_se->dl_defer = 1;
+		setup_new_dl_entity(dl_se);
+#endif
 	}
 }
 
@@ -3198,6 +3210,36 @@ void dl_add_task_root_domain(struct task_struct *p)
 	raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
 }
 
+static void dl_server_add_bw(struct root_domain *rd, int cpu)
+{
+	struct sched_dl_entity *dl_se;
+
+	dl_se = &cpu_rq(cpu)->fair_server;
+	if (dl_server(dl_se) && cpu_active(cpu))
+		__dl_add(&rd->dl_bw, dl_se->dl_bw, dl_bw_cpus(cpu));
+
+#ifdef CONFIG_SCHED_CLASS_EXT
+	dl_se = &cpu_rq(cpu)->ext_server;
+	if (dl_server(dl_se) && cpu_active(cpu))
+		__dl_add(&rd->dl_bw, dl_se->dl_bw, dl_bw_cpus(cpu));
+#endif
+}
+
+static u64 dl_server_read_bw(int cpu)
+{
+	u64 dl_bw = 0;
+
+	if (cpu_rq(cpu)->fair_server.dl_server)
+		dl_bw += cpu_rq(cpu)->fair_server.dl_bw;
+
+#ifdef CONFIG_SCHED_CLASS_EXT
+	if (cpu_rq(cpu)->ext_server.dl_server)
+		dl_bw += cpu_rq(cpu)->ext_server.dl_bw;
+#endif
+
+	return dl_bw;
+}
+
 void dl_clear_root_domain(struct root_domain *rd)
 {
 	int i;
@@ -3216,12 +3258,8 @@ void dl_clear_root_domain(struct root_domain *rd)
 	 * dl_servers are not tasks. Since dl_add_task_root_domain ignores
 	 * them, we need to account for them here explicitly.
 	 */
-	for_each_cpu(i, rd->span) {
-		struct sched_dl_entity *dl_se = &cpu_rq(i)->fair_server;
-
-		if (dl_server(dl_se) && cpu_active(i))
-			__dl_add(&rd->dl_bw, dl_se->dl_bw, dl_bw_cpus(i));
-	}
+	for_each_cpu(i, rd->span)
+		dl_server_add_bw(rd, i);
 }
 
 void dl_clear_root_domain_cpu(int cpu)
@@ -3720,7 +3758,7 @@ static int dl_bw_manage(enum dl_bw_request req, int cpu, u64 dl_bw)
 	unsigned long flags, cap;
 	struct dl_bw *dl_b;
 	bool overflow = 0;
-	u64 fair_server_bw = 0;
+	u64 dl_server_bw = 0;
 
 	rcu_read_lock_sched();
 	dl_b = dl_bw_of(cpu);
@@ -3753,27 +3791,26 @@ static int dl_bw_manage(enum dl_bw_request req, int cpu, u64 dl_bw)
 		cap -= arch_scale_cpu_capacity(cpu);
 
 		/*
-		 * cpu is going offline and NORMAL tasks will be moved away
-		 * from it. We can thus discount dl_server bandwidth
-		 * contribution as it won't need to be servicing tasks after
-		 * the cpu is off.
+		 * cpu is going offline and NORMAL and EXT tasks will be
+		 * moved away from it. We can thus discount dl_server
+		 * bandwidth contribution as it won't need to be servicing
+		 * tasks after the cpu is off.
 		 */
-		if (cpu_rq(cpu)->fair_server.dl_server)
-			fair_server_bw = cpu_rq(cpu)->fair_server.dl_bw;
+		dl_server_bw = dl_server_read_bw(cpu);
 
 		/*
 		 * Not much to check if no DEADLINE bandwidth is present.
 		 * dl_servers we can discount, as tasks will be moved out the
 		 * offlined CPUs anyway.
 		 */
-		if (dl_b->total_bw - fair_server_bw > 0) {
+		if (dl_b->total_bw - dl_server_bw > 0) {
 			/*
 			 * Leaving at least one CPU for DEADLINE tasks seems a
 			 * wise thing to do. As said above, cpu is not offline
 			 * yet, so account for that.
 			 */
 			if (dl_bw_cpus(cpu) - 1)
-				overflow = __dl_overflow(dl_b, cap, fair_server_bw, 0);
+				overflow = __dl_overflow(dl_b, cap, dl_server_bw, 0);
 			else
 				overflow = 1;
 		}

kernel/sched/ext.c

@@ -958,6 +958,8 @@ static void update_curr_scx(struct rq *rq)
 		if (!curr->scx.slice)
 			touch_core_sched(rq, curr);
 	}
+
+	dl_server_update(&rq->ext_server, delta_exec);
 }
 
 static bool scx_dsq_priq_less(struct rb_node *node_a,
@@ -1501,6 +1503,10 @@ static void enqueue_task_scx(struct rq *rq, struct task_struct *p, int enq_flags
 	if (enq_flags & SCX_ENQ_WAKEUP)
 		touch_core_sched(rq, p);
 
+	/* Start dl_server if this is the first task being enqueued */
+	if (rq->scx.nr_running == 1)
+		dl_server_start(&rq->ext_server);
+
 	do_enqueue_task(rq, p, enq_flags, sticky_cpu);
 out:
 	rq->scx.flags &= ~SCX_RQ_IN_WAKEUP;
@@ -2512,6 +2518,33 @@ static struct task_struct *pick_task_scx(struct rq *rq, struct rq_flags *rf)
 	return do_pick_task_scx(rq, rf, false);
 }
 
+/*
+ * Select the next task to run from the ext scheduling class.
+ *
+ * Use do_pick_task_scx() directly with @force_scx enabled, since the
+ * dl_server must always select a sched_ext task.
+ */
+static struct task_struct *
+ext_server_pick_task(struct sched_dl_entity *dl_se, struct rq_flags *rf)
+{
+	if (!scx_enabled())
+		return NULL;
+
+	return do_pick_task_scx(dl_se->rq, rf, true);
+}
+
+/*
+ * Initialize the ext server deadline entity.
+ */
+void ext_server_init(struct rq *rq)
+{
+	struct sched_dl_entity *dl_se = &rq->ext_server;
+
+	init_dl_entity(dl_se);
+
+	dl_server_init(dl_se, rq, ext_server_pick_task);
+}
+
 #ifdef CONFIG_SCHED_CORE
 /**
  * scx_prio_less - Task ordering for core-sched

kernel/sched/idle.c

@@ -537,6 +537,9 @@ static void update_curr_idle(struct rq *rq)
 	se->exec_start = now;
 
 	dl_server_update_idle(&rq->fair_server, delta_exec);
+#ifdef CONFIG_SCHED_CLASS_EXT
+	dl_server_update_idle(&rq->ext_server, delta_exec);
+#endif
 }
 
 /*

kernel/sched/sched.h

@@ -414,6 +414,7 @@ extern void dl_server_init(struct sched_dl_entity *dl_se, struct rq *rq,
 extern void sched_init_dl_servers(void);
 
 extern void fair_server_init(struct rq *rq);
+extern void ext_server_init(struct rq *rq);
 extern void __dl_server_attach_root(struct sched_dl_entity *dl_se, struct rq *rq);
 extern int dl_server_apply_params(struct sched_dl_entity *dl_se,
 		    u64 runtime, u64 period, bool init);
@@ -1171,6 +1172,7 @@ struct rq {
 	struct dl_rq		dl;
 #ifdef CONFIG_SCHED_CLASS_EXT
 	struct scx_rq		scx;
+	struct sched_dl_entity	ext_server;
 #endif
 
 	struct sched_dl_entity	fair_server;

kernel/sched/topology.c

@@ -508,6 +508,11 @@ void rq_attach_root(struct rq *rq, struct root_domain *rd)
 	if (rq->fair_server.dl_server)
 		__dl_server_attach_root(&rq->fair_server, rq);
 
+#ifdef CONFIG_SCHED_CLASS_EXT
+	if (rq->ext_server.dl_server)
+		__dl_server_attach_root(&rq->ext_server, rq);
+#endif
+
 	rq_unlock_irqrestore(rq, &rf);
 
 	if (old_rd)