Skip to contentSkip to conversation

sched/deadline: Fix DL server initialization for initially offline CPUs

PATCH In review Hui Su 2 participants 4 messages Aug 11 – Aug 27

Subsystems (2)
lore
Share & tools
lore

Hui Su

author
#0 lore raw

PATCH 1/1 sched/deadline: Fix DL server initialization for initially offline CPUs

Commit 9f239df55546 ("sched/deadline: Initialize dl_servers after SMP") moved DL server initialization to sched_init_smp(), assuming that all CPUs are online after SMP initialization.

That assumption does not hold when CPUs are intentionally left offline at boot, for example with maxcpus=. sched_init_dl_servers() only walks online CPUs, so the fair and sched_ext DL servers of an initially offline CPU never get their default DL parameters configured. Bringing the CPU online later does not run this initialization again, leaving the server parameters at zero.

For example, boot a two-vCPU QEMU guest with maxcpus=1 and then bring CPU1 online:

  # echo 1 > /sys/devices/system/cpu/cpu1/online
  # cat /sys/kernel/debug/sched/fair_server/cpu1/runtime
  0
  # cat /sys/kernel/debug/sched/fair_server/cpu1/period
  0

With a zero runtime the fair server cannot be activated, so fair tasks on the late-online CPU can be starved when higher-priority tasks monopolize the CPU, potentially causing user-space workloads to stall.

Initialize DL servers for all possible CPUs instead. For CPUs that are already online, preserve the existing initialization behavior. For an initially offline CPU, initialize the server parameters and keep its bandwidth reservation local to the runqueue, but do not account it to the root domain until the CPU becomes active. A subsequent root-domain rebuild then publishes the reservation.

Avoid updating the runqueue clock and setting up the current CBS period for offline CPUs. The configured runtime and period are initialized at boot, while the dynamic runtime and absolute deadline are initialized through the normal CBS wakeup path when the server is first started.

After the fix, the same test reports:

  # cat /sys/kernel/debug/sched/fair_server/cpu1/runtime
  50000000
  # cat /sys/kernel/debug/sched/fair_server/cpu1/period
  1000000000
Fixes 9f239df55546 ("sched/deadline: Initialize dl_servers after SMP")
---
 kernel/sched/deadline.c | 24 ++++++++++++++++--------
 1 file changed, 16 insertions(+), 8 deletions(-)
kernel/sched/deadline.c +16 -8
index 200300043fa5..0445565ed9bb 100644
@@ -1845,14 +1845,16 @@ void sched_init_dl_servers(void)
 	struct rq *rq;
 	struct sched_dl_entity *dl_se;
 
-	for_each_online_cpu(cpu) {
+	for_each_possible_cpu(cpu) {
 		u64 runtime =  50 * NSEC_PER_MSEC;
 		u64 period = 1000 * NSEC_PER_MSEC;
+		bool online = cpu_online(cpu);
 
 		rq = cpu_rq(cpu);
 
 		guard(rq_lock_irq)(rq);
-		update_rq_clock(rq);
+		if (online)
+			update_rq_clock(rq);
 
 		dl_se = &rq->fair_server;
 
@@ -1862,7 +1864,8 @@ void sched_init_dl_servers(void)
 
 		dl_se->dl_server = 1;
 		dl_se->dl_defer = 1;
-		setup_new_dl_entity(dl_se);
+		if (online)
+			setup_new_dl_entity(dl_se);
 
 #ifdef CONFIG_SCHED_CLASS_EXT
 		dl_se = &rq->ext_server;
@@ -1873,7 +1876,8 @@ void sched_init_dl_servers(void)
 
 		dl_se->dl_server = 1;
 		dl_se->dl_defer = 1;
-		setup_new_dl_entity(dl_se);
+		if (online)
+			setup_new_dl_entity(dl_se);
 
 		/*
 		 * No BPF scheduler is loaded at boot, so the ext_server has no
@@ -1918,14 +1922,18 @@ int dl_server_apply_params(struct sched_dl_entity *dl_se, u64 runtime, u64 perio
 	guard(raw_spinlock)(&dl_b->lock);
 
 	cpus = dl_bw_cpus(cpu);
-	cap = dl_bw_capacity(cpu);
 
-	if (__dl_overflow(dl_b, cap, old_bw, new_bw))
-		return -EBUSY;
+	if (!init || cpu_active(cpu)) {
+		cap = dl_bw_capacity(cpu);
+
+		if (__dl_overflow(dl_b, cap, old_bw, new_bw))
+			return -EBUSY;
+	}
 
 	if (init) {
 		__add_rq_bw(new_bw, &rq->dl);
-		__dl_add(dl_b, new_bw, cpus);
+		if (cpu_active(cpu))
+			__dl_add(dl_b, new_bw, cpus);
 		dl_se->dl_bw_attached = 1;
 	} else if (dl_se->dl_bw_attached) {
 		__dl_sub(dl_b, dl_se->dl_bw, cpus);
Signature · 1 line
2.54.0
maintainer Acked-by on v1 1/1
#1 lore raw

Hi Hui Su,

Quoted 43 lines · Hui Su

Commit 9f239df55546 ("sched/deadline: Initialize dl_servers after SMP") moved DL server initialization to sched_init_smp(), assuming that all CPUs are online after SMP initialization.

That assumption does not hold when CPUs are intentionally left offline at boot, for example with maxcpus=. sched_init_dl_servers() only walks online CPUs, so the fair and sched_ext DL servers of an initially offline CPU never get their default DL parameters configured. Bringing the CPU online later does not run this initialization again, leaving the server parameters at zero.

For example, boot a two-vCPU QEMU guest with maxcpus=1 and then bring CPU1 online:

  # echo 1 > /sys/devices/system/cpu/cpu1/online
  # cat /sys/kernel/debug/sched/fair_server/cpu1/runtime
  0
  # cat /sys/kernel/debug/sched/fair_server/cpu1/period
  0

With a zero runtime the fair server cannot be activated, so fair tasks on the late-online CPU can be starved when higher-priority tasks monopolize the CPU, potentially causing user-space workloads to stall.

Initialize DL servers for all possible CPUs instead. For CPUs that are already online, preserve the existing initialization behavior. For an initially offline CPU, initialize the server parameters and keep its bandwidth reservation local to the runqueue, but do not account it to the root domain until the CPU becomes active. A subsequent root-domain rebuild then publishes the reservation.

Avoid updating the runqueue clock and setting up the current CBS period for offline CPUs. The configured runtime and period are initialized at boot, while the dynamic runtime and absolute deadline are initialized through the normal CBS wakeup path when the server is first started.

After the fix, the same test reports:

  # cat /sys/kernel/debug/sched/fair_server/cpu1/runtime
  50000000
  # cat /sys/kernel/debug/sched/fair_server/cpu1/period
  1000000000
Quoted 4 lines
Fixes 9f239df55546 ("sched/deadline: Initialize dl_servers after SMP")

---

Thanks for the fix, makes sense to me.

BTW, while reviewing, sashiko noted a related pre-existing issue in the init=false path of dl_server_apply_params(): __dl_sub() and __dl_add() are called unconditionally with cpus that could be 0 during CPU hot-unplug, when the CPU is no longer active but still online. sched_server_write_common() only checks cpu_online(), so a debugfs write during that window could hit a divide-by-zero.

Could you take a look and consider posting an additional patch to address it?

Thanks,
Juri

author
#2 lore raw
Quoted 62 lines

Hi Hui Su,

Quoted 43 lines · Hui Su

Commit 9f239df55546 ("sched/deadline: Initialize dl_servers after SMP") moved DL server initialization to sched_init_smp(), assuming that all CPUs are online after SMP initialization.

That assumption does not hold when CPUs are intentionally left offline at boot, for example with maxcpus=. sched_init_dl_servers() only walks online CPUs, so the fair and sched_ext DL servers of an initially offline CPU never get their default DL parameters configured. Bringing the CPU online later does not run this initialization again, leaving the server parameters at zero.

For example, boot a two-vCPU QEMU guest with maxcpus=1 and then bring CPU1 online:

  # echo 1 > /sys/devices/system/cpu/cpu1/online
  # cat /sys/kernel/debug/sched/fair_server/cpu1/runtime
  0
  # cat /sys/kernel/debug/sched/fair_server/cpu1/period
  0

With a zero runtime the fair server cannot be activated, so fair tasks on the late-online CPU can be starved when higher-priority tasks monopolize the CPU, potentially causing user-space workloads to stall.

Initialize DL servers for all possible CPUs instead. For CPUs that are already online, preserve the existing initialization behavior. For an initially offline CPU, initialize the server parameters and keep its bandwidth reservation local to the runqueue, but do not account it to the root domain until the CPU becomes active. A subsequent root-domain rebuild then publishes the reservation.

Avoid updating the runqueue clock and setting up the current CBS period for offline CPUs. The configured runtime and period are initialized at boot, while the dynamic runtime and absolute deadline are initialized through the normal CBS wakeup path when the server is first started.

After the fix, the same test reports:

  # cat /sys/kernel/debug/sched/fair_server/cpu1/runtime
  50000000
  # cat /sys/kernel/debug/sched/fair_server/cpu1/period
  1000000000
Quoted 4 lines
Fixes 9f239df55546 ("sched/deadline: Initialize dl_servers after SMP")

---

Thanks for the fix, makes sense to me.

BTW, while reviewing, sashiko noted a related pre-existing issue in the init=false path of dl_server_apply_params(): __dl_sub() and __dl_add() are called unconditionally with cpus that could be 0 during CPU hot-unplug, when the CPU is no longer active but still online. sched_server_write_common() only checks cpu_online(), so a debugfs write during that window could hit a divide-by-zero.

Quoted 2 lines

Could you take a look and consider posting an additional patch to address it?
---

Thanks for pointing this out. I will take a look at the issue and post a separate patch to fix it.

Thanks.