[PATCH 0/2] softirq: fix idle accounting for an interrupted idle task

[PATCH 0/2] softirq: fix idle accounting for an interrupted idle task

[Radu Toader]

tick_irq_exit() skips tick_nohz_irq_exit() whenever the CPU's run queue
has tasks on it, even though the CPU is sitting in the idle loop. Under
sched_ext that is an ordinary steady state - a BPF scheduler holding a
runnable task back, e.g. a cgroup at its CPU limit - and with an
unstable sched_clock the CPU's clock is then never resynchronized on
the way out of idle, so the idle stretch is charged to the next task as
runtime.

Patch 1 adds scx_tickbug, an example scheduler that holds one tagged
task runnable on an idling CPU for a configurable stretch. It exists
because the in-tree example schedulers do reach that state but leave it
within microseconds, so nothing accumulates and the bug is invisible.
Patch 2 is the fix, one line of behaviour: ask is_idle_task(current)
instead of sched_core_idle_cpu().

The patches are independent in the files they touch, but patch 1 comes
first because without it nothing in this tree can produce the state
patch 2 is about.

Measurements
------------

Two kernels built from this tree, differing only in patch 2, booted in
turn on one 16-CPU x86_64 KVM guest with "no-kvmclock tsc=unstable".
Both are needed: tsc=unstable alone only marks the clocksource,
because mark_tsc_unstable() clears sched_clock stability only when
using_native_sched_clock(), which is false under kvm-clock.

Load: one thread spinning 10 ms in every 100 ms in a cgroup, pinned to
the CPU the scheduler holds it on. 20 s windows, three repetitions per
cell, arms interleaved. The figure is the cgroup's cpu.stat usage_usec
over the thread's own CLOCK_MONOTONIC measurement of its busy time,
which is an upper bound on the CPU it can have used, so correct is
<= 1.00:

  hold (ms)   without patch 2   with patch 2
      7            1.32            0.964
     13            1.21            0.963
     23            2.00            0.963
     37            1.69            0.964
     53            1.33            0.964

  control, fair class, no BPF scheduler loaded:
                   1.005           1.006

Nothing in the fixed kernel's figures varies with the hold, and none of
the unfixed kernel's is within 20% of correct. At the 23 ms hold the
matched pair is 0.087 CPU of real work charged as 0.173 unfixed against
0.084 fixed, with 173 and 172 releases respectively - the same work,
accounted twice over.

The unfixed kernel also understates how long it held the task, which is
the same bug from the other side: the scheduler measures the hold with
scx_bpf_now(), the sched_clock that is not being resynchronized. For a
23 ms configured hold it reported a 5.2 ms mean hold where the fixed
kernel reported 15.2 ms for the identical configuration. So the fixed
kernel accounts correctly across holds three times longer than the ones
the unfixed kernel gets wrong.

An earlier measurement, with an out-of-tree sched_ext scheduler that
caps how many threads of a cgroup run at once: one thread spinning in a
cgroup limited to 0.1 CPU in an 8-vCPU guest was charged 0.27 CPUs, and
four such threads 1.12 CPUs, while a sched_switch trace and the work
completed both showed 0.100. With nohz=off, idle=poll or a stable
sched_clock the three agree. That scheduler is not part of this series,
so that one cannot be reproduced from this tree; patch 1 is what makes
everything above reproducible.

What is not shown, and what was not done
----------------------------------------

No loss in /proc/stat. The sum of that CPU's fields 2 to 9 over the
window stayed within 0.6% of the window on both kernels, so in this
configuration the kcpustat_irq_exit() half of what tick_nohz_irq_exit()
does has no visible effect, unlike the forced idle case in 548796e2e70b.
A kprobe on tick_nohz_irq_exit() filtered to the idle task counted 998
calls per 20 s window on the held CPU without patch 2 against 1321 with
it, n=3 each, but those ranges overlap and the two kernels run the load
on different timings, so that bounds the mechanism rather than
measuring it.

The ttwu_pending widening described in patch 2 was reached by reading
the code, not measured on its own. The control arm above bounds any
effect on an ordinary workload (1.005 against 1.006) without isolating
that path.

There is no automated test, and I do not see how to write a useful one.
The wrong number only appears with an unstable sched_clock, which takes
a command line, so a selftest on a machine with a stable TSC would pass
on an unfixed kernel and prove nothing. scx_tickbug reproduces the
state and reports counters, but it asserts nothing and is not a
selftest.

All of the figures come from one 16-CPU x86_64 KVM guest on one host.
No other architecture, topology or guest clocksource was tried.

Open question
-------------

The Fixes: tag on patch 2 names f0e1a0643a59, the commit that made "the
idle task runs while nr_running > 0" reachable without
CONFIG_SCHED_CORE. 548796e2e70b is arguable instead, as the commit that
fixed this same gate for core scheduling with a predicate narrower than
the problem; and the ttwu_pending case predates both. Happy to retag.

Tooling
-------

The bug was found, and both patches, both changelogs and this cover
letter were written, by Claude Opus 5 (Claude Code), in a session
measuring cgroup CPU limit enforcement under sched_ext. The prompts
were about why a cgroup's measured CPU usage disagreed with a
sched_switch trace by a factor of 2.7; the model localized that to this
gate, wrote the fix and the reproducer, ran the measurements and wrote
the text. The content is tool-generated throughout, per
Documentation/process/generated-content.rst. I have reviewed it, I am
signing off on it under the DCO, and I can answer for it in review.

Radu Toader (2):
  sched_ext: Add scx_tickbug, a scheduler that holds a task on an idle
    CPU
  softirq: Treat an interrupted idle task as idle in tick_irq_exit()

 include/linux/sched.h             |   2 -
 kernel/sched/syscalls.c           |  12 -
 kernel/softirq.c                  |  12 +-
 tools/sched_ext/Makefile          |   2 +-
 tools/sched_ext/README.md         |  13 ++
 tools/sched_ext/scx_tickbug.bpf.c | 364 ++++++++++++++++++++++++++++++
 tools/sched_ext/scx_tickbug.c     | 166 ++++++++++++++
 7 files changed, 554 insertions(+), 17 deletions(-)
 create mode 100644 tools/sched_ext/scx_tickbug.bpf.c
 create mode 100644 tools/sched_ext/scx_tickbug.c


base-commit: 05441d697f7f9a927264f4058cbbc6a03dfcac9d
-- 
2.43.0

[PATCH 1/2] sched_ext: Add scx_tickbug, a scheduler that holds a task on an idle CPU

[Radu Toader]

The state behind the tick_irq_exit() gate - the idle task running on a
CPU whose run queue is not empty - is reachable with the existing
example schedulers but not holdable with them, which makes everything
that depends on how long it lasts impossible to measure.

enqueue_task_scx() raises rq->nr_running before it calls ops.enqueue(),
so a scheduler that stashes the task in a map of its own leaves it
runnable and accounted on that CPU with nothing for the CPU to pick.
scx_central already does exactly that, and then calls
scx_bpf_kick_cpu() immediately, so the window is microseconds wide.

This scheduler is the same thing with the kick removed: a tagged task
is parked on the target CPU's run queue and released only by a timer
pinned to another CPU, so the window is as wide as -H says. Untagged
tasks go through a shared DSQ and per-CPU kthreads keep their local DSQ
and preemption, so the machine stays usable and the timer's softirq
cannot be starved.

Two things it has to get right, both of which cost a measurement round
before they were found. The release timer must kick @target_cpu only
when something is actually held: an unconditional SCX_KICK_PREEMPT
lands on the task released a moment earlier and throws it off the CPU
mid-run, whereupon ops.enqueue() holds it again, and the load then
reports using half the CPU it asked for. And a tagged task that is
already running goes straight back into the local DSQ instead of being
held, so the hold only ever delays the start of a run - the state the
gate under test mis-judges - and never interrupts one. nr_requeue
counts the second case so a run can show whether it happened.

It reports the mean and worst hold it actually achieved, which is the
independent variable of any experiment using it and is not otherwise
observable from userspace. Do not use a hold that divides the load's
period: with a 100 ms spinner, holds of 5/10/20 ms phase-lock and the
task is only ever held for microseconds.

Tooling: written by Claude Opus 5 (Claude Code) while measuring what
the tick_irq_exit() idle gate changes. All of this patch is
tool-generated.

Signed-off-by: Radu Toader <radu.m.toader@gmail.com>
---
 tools/sched_ext/Makefile          |   2 +-
 tools/sched_ext/README.md         |  13 ++
 tools/sched_ext/scx_tickbug.bpf.c | 364 ++++++++++++++++++++++++++++++
 tools/sched_ext/scx_tickbug.c     | 166 ++++++++++++++
 4 files changed, 544 insertions(+), 1 deletion(-)
 create mode 100644 tools/sched_ext/scx_tickbug.bpf.c
 create mode 100644 tools/sched_ext/scx_tickbug.c

diff --git a/tools/sched_ext/Makefile b/tools/sched_ext/Makefile
index 70d2503b6b33..02e1500bf402 100644
--- a/tools/sched_ext/Makefile
+++ b/tools/sched_ext/Makefile
@@ -191,7 +191,7 @@ $(INCLUDE_DIR)/%.bpf.skel.h: $(SCXOBJ_DIR)/%.bpf.o $(INCLUDE_DIR)/vmlinux.h $(BP
 
 SCX_COMMON_DEPS := include/scx/common.h include/scx/user_exit_info.h | $(BINDIR)
 
-c-sched-targets = scx_simple scx_cpu0 scx_qmap scx_central scx_flatcg scx_userland scx_pair scx_sdt
+c-sched-targets = scx_simple scx_cpu0 scx_qmap scx_central scx_tickbug scx_flatcg scx_userland scx_pair scx_sdt
 
 $(addprefix $(BINDIR)/,$(c-sched-targets)): \
 	$(BINDIR)/%: \
diff --git a/tools/sched_ext/README.md b/tools/sched_ext/README.md
index 0ee5a3d997e5..4addc7c2ed23 100644
--- a/tools/sched_ext/README.md
+++ b/tools/sched_ext/README.md
@@ -185,6 +185,19 @@ where this could be particularly useful is running VMs, where running with
 infinite slices and no timer ticks allows the VM to avoid unnecessary expensive
 vmexits.
 
+## scx_tickbug
+
+A scheduler whose only job is to keep one task runnable on a CPU that is in the
+idle loop, for a length of time given on the command line. That state - the
+idle task running while the run queue is not empty - is reachable by several of
+the schedulers here, but none of them holds it long enough to observe what
+depends on it. Tasks named "spin" are parked on the target CPU's run queue and
+released only by a periodic timer; everything else runs through a shared DSQ so
+the machine stays usable.
+
+This is a test scheduler, not a policy. It exists so that the idle-time
+bookkeeping skipped by an interrupt taken in that state can be measured.
+
 ## scx_flatcg
 
 A flattened cgroup hierarchy scheduler. This scheduler implements hierarchical
diff --git a/tools/sched_ext/scx_tickbug.bpf.c b/tools/sched_ext/scx_tickbug.bpf.c
new file mode 100644
index 000000000000..e6a6d627f042
--- /dev/null
+++ b/tools/sched_ext/scx_tickbug.bpf.c
@@ -0,0 +1,364 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * scx_tickbug: keep a task runnable on a CPU that is in the idle loop.
+ *
+ * This scheduler exists to hold one state for a controllable length of time:
+ * the idle task running on a CPU whose run queue is not empty.
+ *
+ * enqueue_task_scx() raises rq->nr_running before it calls ops.enqueue(), so a
+ * BPF scheduler that stashes the task somewhere of its own - here a BPF queue
+ * map - leaves it runnable and accounted on that CPU with nothing for the CPU
+ * to pick. The CPU enters the idle loop and stays there. In that state
+ * idle_cpu() is false, because idle_rq() tests rq->nr_running, while
+ * is_idle_task(current) is true.
+ *
+ * tick_irq_exit() used to ask idle_cpu(), so every interrupt taken in that
+ * idle loop skipped tick_nohz_irq_exit(). Each skip leaves TS_FLAG_IDLE_ACTIVE
+ * clear, and the next exit from the idle loop then also skips
+ * sched_clock_idle_wakeup_event(), so with an unstable sched_clock the CPU's
+ * clock is never resynchronized: the task picked next starts with a stale
+ * exec_start and the next tick charges it the whole idle stretch.
+ *
+ * The in-tree example schedulers can reach this state but not hold it.
+ * scx_central stashes a task in a BPF queue exactly like this, then
+ * immediately calls scx_bpf_kick_cpu(), so the window is microseconds wide and
+ * nothing measurable accumulates. Here the only thing that releases a held
+ * task is a periodic timer, so the window is hold_ns wide and can be swept.
+ *
+ * Only tasks named by @tag_comm are held; everything else runs through a
+ * shared DSQ so the machine stays usable while the test runs.
+ *
+ * Copyright (c) 2026 Radu Toader
+ */
+#include <scx/common.bpf.h>
+
+char _license[] SEC("license") = "GPL";
+
+enum {
+	SHARED_DSQ	= 0,
+	MS_TO_NS	= 1000LLU * 1000,
+};
+
+/* CPU whose run queue holds the tagged task while the CPU idles. */
+const volatile s32 target_cpu;
+/* CPU the release timer is pinned to. Must not be @target_cpu. */
+const volatile s32 timer_cpu = 1;
+/* How long a tagged task is held before it is allowed to run. */
+const volatile u64 hold_ns = 10 * MS_TO_NS;
+const volatile u64 slice_ns;
+const volatile u32 nr_cpu_ids = 1;	/* !0 for veristat, set during init */
+
+bool timer_pinned = true;
+bool timer_started;
+
+/*
+ * Set by the release timer, consumed by ops.dispatch() on @target_cpu. Without
+ * it ops.dispatch() would hand the held task straight back out, because the
+ * core calls ops.dispatch() whenever the CPU is about to go idle.
+ */
+bool release_now;
+
+u64 nr_enq, nr_held, nr_released, nr_timers;
+u64 nr_passthru, nr_locals, nr_lost, nr_other_cpu, nr_overflow, nr_requeue;
+/*
+ * How many tagged tasks are held right now. The release timer only kicks when
+ * this is non-zero: a kick with nothing held would preempt whatever is running
+ * on @target_cpu, and since that is normally the task just released, the task
+ * would lose most of its CPU to its own release mechanism. The first version of
+ * this scheduler did exactly that and reported half the usage it should.
+ */
+u64 nr_held_now;
+/* Proof that the hold actually happened, and how wide it was. */
+u64 held_at, total_held_ns, max_held_ns;
+
+UEI_DEFINE(uei);
+
+struct {
+	__uint(type, BPF_MAP_TYPE_QUEUE);
+	__uint(max_entries, 1024);
+	__type(value, s32);
+} hold_q SEC(".maps");
+
+struct tickbug_timer {
+	struct bpf_timer timer;
+};
+
+struct {
+	__uint(type, BPF_MAP_TYPE_ARRAY);
+	__uint(max_entries, 1);
+	__type(key, u32);
+	__type(value, struct tickbug_timer);
+} release_timer SEC(".maps");
+
+/*
+ * The load generator is called "spin". Compare 5 bytes so the terminating NUL
+ * has to match too and a longer name starting with "spin" is not held.
+ */
+static bool is_tagged(struct task_struct *p)
+{
+	return !bpf_strncmp(p->comm, 5, "spin");
+}
+
+s32 BPF_STRUCT_OPS(tickbug_select_cpu, struct task_struct *p,
+		   s32 prev_cpu, u64 wake_flags)
+{
+	if (is_tagged(p))
+		return target_cpu;
+	return prev_cpu;
+}
+
+void BPF_STRUCT_OPS(tickbug_enqueue, struct task_struct *p, u64 enq_flags)
+{
+	s32 pid = p->pid;
+
+	__sync_fetch_and_add(&nr_enq, 1);
+
+	/*
+	 * Per-CPU kthreads go to the head of their own CPU's local DSQ with
+	 * preemption. The release timer runs from a softirq, so if a kthread
+	 * such as ksoftirqd were starved behind a user task the held task
+	 * would never be released.
+	 */
+	if ((p->flags & PF_KTHREAD) && p->nr_cpus_allowed == 1) {
+		__sync_fetch_and_add(&nr_locals, 1);
+		scx_bpf_dsq_insert(p, SCX_DSQ_LOCAL, slice_ns,
+				   enq_flags | SCX_ENQ_PREEMPT);
+		return;
+	}
+
+	if (is_tagged(p)) {
+		if (scx_bpf_task_cpu(p) != target_cpu) {
+			/*
+			 * ops.select_cpu() is only a hint and is not called on
+			 * every path. Counted rather than forced: a run with a
+			 * non-zero count here did not test what it meant to.
+			 */
+			__sync_fetch_and_add(&nr_other_cpu, 1);
+		} else if (scx_bpf_task_running(p)) {
+			/*
+			 * Already on the CPU and merely re-enqueued, e.g. after
+			 * a kthread preempted it. Holding it here would take
+			 * CPU away from work already in progress, which is not
+			 * the state under test: the hold belongs at the start
+			 * of a run, not in the middle of one. Put it straight
+			 * back so it carries on.
+			 */
+			__sync_fetch_and_add(&nr_requeue, 1);
+			scx_bpf_dsq_insert(p, SCX_DSQ_LOCAL, slice_ns, enq_flags);
+			return;
+		} else if (bpf_map_push_elem(&hold_q, &pid, 0)) {
+			__sync_fetch_and_add(&nr_overflow, 1);
+		} else {
+			/*
+			 * This is the whole scheduler. The task is already in
+			 * target_cpu's rq->nr_running; parking its pid here
+			 * and deliberately not kicking the CPU leaves it in
+			 * the idle loop with a runnable task queued.
+			 */
+			held_at = scx_bpf_now() ?: 1;
+			__sync_fetch_and_add(&nr_held, 1);
+			__sync_fetch_and_add(&nr_held_now, 1);
+			return;
+		}
+	}
+
+	__sync_fetch_and_add(&nr_passthru, 1);
+	scx_bpf_dsq_insert(p, SHARED_DSQ, slice_ns, enq_flags);
+}
+
+/* Let one held task run on the CPU we are dispatching for. */
+static bool release_one(void)
+{
+	struct task_struct *p;
+	s32 pid;
+
+	bpf_repeat(BPF_MAX_LOOPS) {
+		u64 now, at;
+
+		if (bpf_map_pop_elem(&hold_q, &pid))
+			return false;
+
+		p = bpf_task_from_pid(pid);
+		if (!p) {
+			__sync_fetch_and_add(&nr_lost, 1);
+			__sync_fetch_and_sub(&nr_held_now, 1);
+			continue;
+		}
+
+		/*
+		 * Affinity may have changed while the task sat in the queue.
+		 * Such a task goes to the shared DSQ instead of being pinned
+		 * to a CPU it is no longer allowed on.
+		 */
+		if (!bpf_cpumask_test_cpu(target_cpu, p->cpus_ptr) ||
+		    is_migration_disabled(p)) {
+			scx_bpf_dsq_insert(p, SHARED_DSQ, slice_ns, 0);
+			bpf_task_release(p);
+			__sync_fetch_and_sub(&nr_held_now, 1);
+			if (!scx_bpf_dispatch_nr_slots())
+				return false;
+			continue;
+		}
+
+		scx_bpf_dsq_insert(p, SCX_DSQ_LOCAL, slice_ns, 0);
+		bpf_task_release(p);
+		__sync_fetch_and_sub(&nr_held_now, 1);
+
+		at = held_at;
+		now = scx_bpf_now();
+		if (at && now > at) {
+			u64 d = now - at;
+
+			__sync_fetch_and_add(&total_held_ns, d);
+			if (d > max_held_ns)
+				max_held_ns = d;
+		}
+		held_at = 0;
+
+		__sync_fetch_and_add(&nr_released, 1);
+		return true;
+	}
+
+	return false;
+}
+
+static void start_release_timer(void)
+{
+	struct bpf_timer *timer;
+	u32 key = 0;
+	int ret;
+
+	if (likely(timer_started))
+		return;
+
+	timer = bpf_map_lookup_elem(&release_timer, &key);
+	if (!timer) {
+		scx_bpf_error("failed to look up the release timer");
+		return;
+	}
+
+	ret = bpf_timer_start(timer, hold_ns, BPF_F_TIMER_CPU_PIN);
+	if (ret == -EINVAL) {
+		/* BPF_F_TIMER_CPU_PIN needs >= 6.7 */
+		timer_pinned = false;
+		ret = bpf_timer_start(timer, hold_ns, 0);
+	}
+	if (ret) {
+		scx_bpf_error("bpf_timer_start failed (%d)", ret);
+		return;
+	}
+
+	timer_started = true;
+}
+
+void BPF_STRUCT_OPS(tickbug_dispatch, s32 cpu, struct task_struct *prev)
+{
+	/*
+	 * Arm the timer from the CPU it must be pinned to. ops.init() runs in
+	 * whatever context loaded the scheduler, which is the wrong CPU.
+	 */
+	if (cpu == timer_cpu)
+		start_release_timer();
+
+	if (cpu == target_cpu && release_now) {
+		release_now = false;
+		if (release_one())
+			return;
+	}
+
+	scx_bpf_dsq_move_to_local(SHARED_DSQ, 0);
+}
+
+static int release_timerfn(void *map, int *key, struct bpf_timer *timer)
+{
+	s32 curr_cpu = bpf_get_smp_processor_id();
+	int ret;
+
+	if (timer_pinned && curr_cpu == target_cpu) {
+		scx_bpf_error("release timer ran on the target CPU %d", curr_cpu);
+		return 0;
+	}
+
+	/*
+	 * Nothing held means @target_cpu is either idle with an empty run
+	 * queue or busy with the task released last time. Kicking in either
+	 * case is at best pointless and at worst preempts that task.
+	 */
+	if (!nr_held_now)
+		goto rearm;
+
+	/*
+	 * Allow one held task through and make @target_cpu notice.
+	 * SCX_KICK_PREEMPT sends a reschedule IPI, which does not go through
+	 * irq_enter() and so does not run tick_irq_exit() itself: it only ends
+	 * the idle stretch whose bookkeeping the mis-gated interrupts dropped.
+	 */
+	release_now = true;
+	scx_bpf_kick_cpu(target_cpu, SCX_KICK_PREEMPT);
+
+rearm:
+
+	__sync_fetch_and_add(&nr_timers, 1);
+
+	ret = bpf_timer_start(timer, hold_ns,
+			      timer_pinned ? BPF_F_TIMER_CPU_PIN : 0);
+	if (ret)
+		scx_bpf_error("bpf_timer_start failed (%d)", ret);
+	return 0;
+}
+
+int BPF_STRUCT_OPS_SLEEPABLE(tickbug_init)
+{
+	struct bpf_timer *timer;
+	u32 key = 0;
+	int ret;
+
+	if (target_cpu == timer_cpu) {
+		scx_bpf_error("target and timer CPU must differ (both %d)",
+			      target_cpu);
+		return -EINVAL;
+	}
+
+	ret = scx_bpf_create_dsq(SHARED_DSQ, -1);
+	if (ret) {
+		scx_bpf_error("scx_bpf_create_dsq failed (%d)", ret);
+		return ret;
+	}
+
+	timer = bpf_map_lookup_elem(&release_timer, &key);
+	if (!timer)
+		return -ESRCH;
+
+	bpf_timer_init(timer, &release_timer, CLOCK_MONOTONIC);
+	bpf_timer_set_callback(timer, release_timerfn);
+
+	/* Get ops.dispatch() to run on @timer_cpu so it arms the timer. */
+	scx_bpf_kick_cpu(timer_cpu, 0);
+
+	return 0;
+}
+
+void BPF_STRUCT_OPS(tickbug_exit, struct scx_exit_info *ei)
+{
+	UEI_RECORD(uei, ei);
+}
+
+SCX_OPS_DEFINE(tickbug_ops,
+	       /*
+		* The last runnable task on a CPU has to come through
+		* ops.enqueue() too, otherwise the tagged task would simply
+		* keep running on @target_cpu and never be held.
+		*/
+	       .flags			= SCX_OPS_ENQ_LAST,
+
+	       .select_cpu		= (void *)tickbug_select_cpu,
+	       .enqueue			= (void *)tickbug_enqueue,
+	       .dispatch		= (void *)tickbug_dispatch,
+	       .init			= (void *)tickbug_init,
+	       .exit			= (void *)tickbug_exit,
+	       /*
+		* A held task is runnable and not running, which is what the
+		* watchdog looks for. Keep the timeout well above any hold.
+		*/
+	       .timeout_ms		= 5000U,
+	       .name			= "tickbug");
diff --git a/tools/sched_ext/scx_tickbug.c b/tools/sched_ext/scx_tickbug.c
new file mode 100644
index 000000000000..8acca5c1ff6a
--- /dev/null
+++ b/tools/sched_ext/scx_tickbug.c
@@ -0,0 +1,166 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright (c) 2026 Radu Toader
+ */
+#define _GNU_SOURCE
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <inttypes.h>
+#include <signal.h>
+#include <assert.h>
+#include <libgen.h>
+#include <bpf/bpf.h>
+#include <scx/common.h>
+#include "scx_tickbug.bpf.skel.h"
+
+const char help_fmt[] =
+"Hold a task runnable on a CPU that is in the idle loop.\n"
+"\n"
+"Tasks named \"spin\" are queued on the target CPU's run queue and left\n"
+"there, so the CPU idles with a runnable task queued, until a periodic\n"
+"timer releases one. Every other task runs through a shared DSQ.\n"
+"\n"
+"See the top-level comment in .bpf.c for more details.\n"
+"\n"
+"Usage: %s [-c CPU] [-t CPU] [-H HOLD_MS] [-s SLICE_US] [-v]\n"
+"\n"
+"  -c CPU        Target CPU, the one that idles with a task queued (default: 0)\n"
+"  -t CPU        CPU the release timer is pinned to (default: 1)\n"
+"  -H HOLD_MS    How long a task is held before release (default: 10)\n"
+"  -s SLICE_US   Override slice duration\n"
+"  -v            Print libbpf debug messages\n"
+"  -h            Display this help and exit\n";
+
+static bool verbose;
+static volatile int exit_req;
+
+static int libbpf_print_fn(enum libbpf_print_level level, const char *format, va_list args)
+{
+	if (level == LIBBPF_DEBUG && !verbose)
+		return 0;
+	return vfprintf(stderr, format, args);
+}
+
+static void sigint_handler(int dummy)
+{
+	exit_req = 1;
+}
+
+int main(int argc, char **argv)
+{
+	struct scx_tickbug *skel;
+	struct bpf_link *link;
+	__u64 seq = 0, ecode;
+	__u32 nr_cpu_ids;
+	__s32 opt;
+
+	libbpf_set_print(libbpf_print_fn);
+	signal(SIGINT, sigint_handler);
+	signal(SIGTERM, sigint_handler);
+restart:
+	optind = 1;
+	skel = SCX_OPS_OPEN(tickbug_ops, scx_tickbug);
+
+	nr_cpu_ids = libbpf_num_possible_cpus();
+	assert(nr_cpu_ids > 0);
+	assert(nr_cpu_ids <= INT32_MAX);
+
+	skel->rodata->nr_cpu_ids = nr_cpu_ids;
+	skel->rodata->target_cpu = 0;
+	skel->rodata->timer_cpu = 1;
+	skel->rodata->hold_ns = 10ULL * 1000 * 1000;
+	skel->rodata->slice_ns = __COMPAT_ENUM_OR_ZERO("scx_public_consts", "SCX_SLICE_DFL");
+
+	while ((opt = getopt(argc, argv, "c:t:H:s:vh")) != -1) {
+		switch (opt) {
+		case 'c':
+		case 't': {
+			unsigned long cpu = strtoul(optarg, NULL, 0);
+
+			if (cpu >= nr_cpu_ids) {
+				fprintf(stderr, "invalid CPU id %lu (%u max)\n",
+					cpu, nr_cpu_ids - 1);
+				scx_tickbug__destroy(skel);
+				return 1;
+			}
+			if (opt == 'c')
+				skel->rodata->target_cpu = (s32)cpu;
+			else
+				skel->rodata->timer_cpu = (s32)cpu;
+			break;
+		}
+		case 'H':
+			skel->rodata->hold_ns = strtoull(optarg, NULL, 0) * 1000 * 1000;
+			break;
+		case 's':
+			skel->rodata->slice_ns = strtoull(optarg, NULL, 0) * 1000;
+			break;
+		case 'v':
+			verbose = true;
+			break;
+		default:
+			fprintf(stderr, help_fmt, basename(argv[0]));
+			return opt != 'h';
+		}
+	}
+
+	if (skel->rodata->target_cpu == skel->rodata->timer_cpu) {
+		fprintf(stderr, "the target CPU and the timer CPU must differ (both %d)\n",
+			skel->rodata->target_cpu);
+		scx_tickbug__destroy(skel);
+		return 1;
+	}
+	if (!skel->rodata->hold_ns) {
+		fprintf(stderr, "a hold of 0 would not hold anything\n");
+		scx_tickbug__destroy(skel);
+		return 1;
+	}
+
+	SCX_OPS_LOAD(skel, tickbug_ops, scx_tickbug, uei);
+	link = SCX_OPS_ATTACH(skel, tickbug_ops, scx_tickbug);
+
+	printf("target_cpu=%d timer_cpu=%d hold_ms=%llu slice_us=%llu\n",
+	       skel->rodata->target_cpu, skel->rodata->timer_cpu,
+	       (unsigned long long)(skel->rodata->hold_ns / 1000000),
+	       (unsigned long long)(skel->rodata->slice_ns / 1000));
+
+	if (!skel->data->timer_pinned)
+		printf("WARNING: BPF_F_TIMER_CPU_PIN unavailable, release timer is not pinned\n");
+
+	while (!exit_req && !UEI_EXITED(skel, uei)) {
+		__u64 released = skel->bss->nr_released;
+
+		printf("[SEQ %llu]\n", seq++);
+		printf("enq    :%10" PRIu64 "   held:%10" PRIu64
+		       " released:%10" PRIu64 " timers:%10" PRIu64 "\n",
+		       skel->bss->nr_enq, skel->bss->nr_held,
+		       skel->bss->nr_released, skel->bss->nr_timers);
+		printf("pass   :%10" PRIu64 "  local:%10" PRIu64
+		       "     lost:%10" PRIu64 " othrcpu:%10" PRIu64
+		       " overflow:%10" PRIu64 " requeue:%10" PRIu64
+		       " now:%10" PRIu64 "\n",
+		       skel->bss->nr_passthru, skel->bss->nr_locals,
+		       skel->bss->nr_lost, skel->bss->nr_other_cpu,
+		       skel->bss->nr_overflow, skel->bss->nr_requeue,
+		       skel->bss->nr_held_now);
+		/*
+		 * The mean and the worst hold are the independent variable of
+		 * the whole experiment: without them the run proves nothing.
+		 */
+		printf("hold_us: mean %8llu   max %8llu\n",
+		       (unsigned long long)(released ?
+			       (skel->bss->total_held_ns / released) / 1000 : 0),
+		       (unsigned long long)(skel->bss->max_held_ns / 1000));
+		fflush(stdout);
+		sleep(1);
+	}
+
+	bpf_link__destroy(link);
+	ecode = UEI_REPORT(skel, uei);
+	scx_tickbug__destroy(skel);
+
+	if (!exit_req && UEI_ECODE_RESTART(ecode))
+		goto restart;
+	return 0;
+}
-- 
2.43.0

[PATCH 2/2] softirq: Treat an interrupted idle task as idle in tick_irq_exit()

[Radu Toader]

tick_irq_exit() only runs tick_nohz_irq_exit() if sched_core_idle_cpu(),
which is idle_cpu() unless core scheduling is built in. idle_cpu() is
false when the run queue has tasks while the idle task runs.

That is the normal state of a CPU under sched_ext whose BPF scheduler
holds back a runnable task of that CPU in one of its queues, e.g. a task
of a cgroup at its CPU limit: the task stays queued on the CPU's run
queue and counts in rq->nr_running, while the CPU sits in the idle loop.

tick_nohz_irq_exit() is what marks an idle CPU as idle again for the
scheduler clock after an interrupt in the idle loop
(TS_FLAG_IDLE_ACTIVE, sched_clock_idle_sleep_event()). Without it, the
next exit from the idle loop doesn't resynchronize the CPU's sched_clock
with sched_clock_idle_wakeup_event(). With an unstable sched_clock, e.g.
under kvm-clock without a stable TSC, the clock of a CPU in tick-stopped
idle is clamped to its last stamp plus a tick, so a task picked after a
reschedule IPI (which doesn't go through irq_enter()) starts with a
stale exec_start, and the next tick charges the whole idle stretch to
it. A cgroup holding one such task was charged twice the CPU time it
could possibly have used; numbers are in the cover letter.

Commit 548796e2e70b ("sched/core: introduce sched_core_idle_cpu()")
fixed this same gate for the other case of "the idle task runs while
nr_running > 0", forced idle under core scheduling, where the symptom
was too little idle time in /proc/stat. sched_ext made that state
reachable without CONFIG_SCHED_CORE, and there the symptom is idle time
charged to a task as runtime.

Ask whether the idle task is running, which covers both cases. CPUs
whose run queue is really empty are unaffected; a CPU interrupted
between leaving the idle loop and switching to a task has need_resched()
set or has already left the idle accounting, see tick_nohz_irq_exit().
sched_core_idle_cpu() has no users left and goes.

The new gate is strictly wider than the old one, and deliberately so.
idle_cpu() is idle_rq(), rq->curr == rq->idle && !rq->nr_running &&
!rq->ttwu_pending, so besides the two cases above it also skipped a CPU
idling with a wakeup queued by TTWU_QUEUE whose IPI has not run yet.
That case is now covered too, on any kernel and not just under
sched_ext. It is the same latch failure for a shorter window, so
covering it looks right rather than merely harmless, but it was reached
by reading the code and not measured on its own. What keeps the wider
gate from giving a CPU idle accounting it should not have is that
tick_nohz_irq_exit() tests TS_FLAG_INIDLE itself and only does the idle
bookkeeping inside the idle loop.

Compile-tested in four x86_64 defconfig variants - SCHED_CORE=y with
NO_HZ_FULL=y, SCHED_CORE=n, NO_HZ_COMMON=n, and SCHED_CLASS_EXT=y with
SCHED_CORE=y and NO_HZ_FULL=y - with no new warnings at W=1.

Tooling: the bug was found and this patch and changelog were written by
Claude Opus 5 (Claude Code), in a session measuring cgroup CPU limit
enforcement under sched_ext. All of the patch and all of the changelog
are tool-generated; the cover letter says what was and was not measured.
By the threat model in Documentation/process/threat-model.rst this is a
regular bug and not a vulnerability: it misreports CPU time under a
legitimate configuration and crosses no privilege or isolation boundary.

Fixes: f0e1a0643a59 ("sched_ext: Implement BPF extensible scheduler class")
Signed-off-by: Radu Toader <radu.m.toader@gmail.com>
---
 include/linux/sched.h   |  2 --
 kernel/sched/syscalls.c | 12 ------------
 kernel/softirq.c        | 12 ++++++++++--
 3 files changed, 10 insertions(+), 16 deletions(-)

diff --git a/include/linux/sched.h b/include/linux/sched.h
index 1c93da75226e..0e704ea9416f 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -2326,11 +2326,9 @@ extern void sched_core_free(struct task_struct *tsk);
 extern void sched_core_fork(struct task_struct *p);
 extern int sched_core_share_pid(unsigned int cmd, pid_t pid, enum pid_type type,
 				unsigned long uaddr);
-extern int sched_core_idle_cpu(int cpu);
 #else
 static inline void sched_core_free(struct task_struct *tsk) { }
 static inline void sched_core_fork(struct task_struct *p) { }
-static inline int sched_core_idle_cpu(int cpu) { return idle_cpu(cpu); }
 #endif
 
 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
diff --git a/kernel/sched/syscalls.c b/kernel/sched/syscalls.c
index 45a127deeadd..22e316680a86 100644
--- a/kernel/sched/syscalls.c
+++ b/kernel/sched/syscalls.c
@@ -194,18 +194,6 @@ struct task_struct *idle_task(int cpu)
 	return cpu_rq(cpu)->idle;
 }
 
-#ifdef CONFIG_SCHED_CORE
-int sched_core_idle_cpu(int cpu)
-{
-	struct rq *rq = cpu_rq(cpu);
-
-	if (sched_core_enabled(rq) && rq->curr == rq->idle)
-		return 1;
-
-	return idle_cpu(cpu);
-}
-#endif /* CONFIG_SCHED_CORE */
-
 /**
  * find_process_by_pid - find a process with a matching PID value.
  * @pid: the pid in question.
diff --git a/kernel/softirq.c b/kernel/softirq.c
index 7fc2200392cd..e7d649c29c0c 100644
--- a/kernel/softirq.c
+++ b/kernel/softirq.c
@@ -738,8 +738,16 @@ static inline void tick_irq_exit(void)
 #ifdef CONFIG_NO_HZ_COMMON
 	int cpu = smp_processor_id();
 
-	/* Make sure that timer wheel updates are propagated */
-	if ((sched_core_idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
+	/*
+	 * Make sure that timer wheel updates are propagated. An interrupted
+	 * idle loop counts as idle even with runnable tasks queued on the
+	 * CPU: forced idle under core scheduling, and the tasks sched_ext
+	 * keeps on a CPU's run queue while its BPF scheduler holds them back.
+	 * idle_cpu() would skip the idle bookkeeping of tick_nohz_irq_exit()
+	 * for those, and with an unstable sched_clock the CPU's clock then
+	 * isn't resynchronized when it leaves the idle loop.
+	 */
+	if ((is_idle_task(current) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
 		if (hardirq_count() == HARDIRQ_OFFSET)
 			tick_nohz_irq_exit();
 	}
-- 
2.43.0