[PATCH v3 0/4] sched: Run task_mm_cid_work in batches to lower latency

[PATCH v3 0/4] sched: Run task_mm_cid_work in batches to lower latency

[Gabriele Monaco]

This V2 of [1] is a continuation of [2] but using a simpler approach.
The task_mm_cid_work runs as a task_work returning to userspace and
causes a non-negligible scheduling latency, mostly due to its iterations
over all cores.

Split the work into several batches, each call to task_mm_cid_work will
not run for all cpus but just for a configurable number of cpus. Next
runs will pick up where the previous left off.
The mechanism that avoids running too frequently (100ms) is enforced
only when finishing all cpus, that is when starting from 0.

Also improve the predictability of the scan on short running tasks by
scheduling it from rseq_sched_switch_event, which runs on every task
switch (similar behaviour to [2]), the same workaround on the tick for
long running tasks seen in [2] was ported also here.

The duration of a full scan depends now on the workload, where workloads
with less threads are more likely to take longer.
Tests with cyclictest (threads with 100us period) and hackbench
(processes) on a 128 CPUs machine measuring the time to complete the
scan as well as the time between non-complete scans showed the following
(batch size of 8: 16 iterations):

cyclictest: delay 0-400 us , complete scan 1.5-2 ms
hackbench: delay 5us - 3ms , complete scan 1.5-15 ms

With the observed worst case for hackbench, it would take more than 800
CPUs to reach the current 100ms limit.

The problematic latency observed on a full scan (128 CPUs), had a
duration of the call to task_mm_cid_scan around 35 us, where 20 us is
considered a relevant latency on this machine.

Measurements showed these durations for each call to task_mm_cid_scan:

batch size  8:  1-11 us (majority below 10)
batch size 16:  3-16 us (majority below 10)
batch size 32: 10-21 us (majority above 15)

This led to a choice of 16 as default batch size.

Patch 1 add support for prev_sum_exec_runtime to the RT, deadline and
sched_ext classes as it is supported by fair, this is required to avoid
calling rseq_preempt on tick if the runtime is below a threshold.

Patch 2 schedules the task_mm_cid_work from rseq_sched_switch_event().

Patch 3 splits the work into batches.

Patch 4 adds a selftest to validate the functionality of the
task_mm_cid_work (i.e. to compact the mm_cids).

Changes since V2 [3]:
* Rebase on rseq rework [4].
* Revert to using task_work.
* Start the work in rseq_sched_switch_event().

Changes since V1 [1]:
* Use cpu_possible_mask in scan.
* Make sure batches have the same number of CPUs also if mask is sparse.
* Run the task on rseq_handle_notify_resume as in [2] but call directly.
* Schedule the work and mm_cid update on tick for long running tasks.
* Fix condition for need_scan only on first batch.
* Change RSEQ_CID_SCAN_BATCH default to be a power of 2.
* Rebase selftest on [2].
* Increase the selftest timeout on large systems.

To: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
To: Peter Zijlstra <peterz@infradead.org>
To: Thomas Gleixner <tglx@linutronix.de>
To: Ingo Molnar <mingo@redhat.com>

[1] - https://lore.kernel.org/lkml/20250217112317.258716-1-gmonaco@redhat.com
[2] - https://lore.kernel.org/lkml/20250707144824.117014-1-gmonaco@redhat.com
[3] - https://lore.kernel.org/lkml/20250716160603.138385-6-gmonaco@redhat.com
[4] - https://lore.kernel.org/lkml/20250908212737.353775467@linutronix.de

Gabriele Monaco (4):
  sched: Add prev_sum_exec_runtime support for RT, DL and SCX classes
  rseq: Schedule the mm_cid_compaction from rseq_sched_switch_event()
  sched: Compact RSEQ concurrency IDs in batches
  selftests/rseq: Add test for mm_cid compaction

 include/linux/mm_types.h                      |  26 +++
 include/linux/rseq.h                          |   3 +
 include/linux/sched.h                         |   3 +
 init/Kconfig                                  |  12 ++
 kernel/sched/core.c                           |  79 ++++++-
 kernel/sched/deadline.c                       |   1 +
 kernel/sched/ext.c                            |   1 +
 kernel/sched/rt.c                             |   1 +
 kernel/sched/sched.h                          |   2 +
 tools/testing/selftests/rseq/.gitignore       |   1 +
 tools/testing/selftests/rseq/Makefile         |   2 +-
 .../selftests/rseq/mm_cid_compaction_test.c   | 204 ++++++++++++++++++
 12 files changed, 324 insertions(+), 11 deletions(-)
 create mode 100644 tools/testing/selftests/rseq/mm_cid_compaction_test.c


base-commit: 1822acbae2c9b8a6e6472809b42eab72cd7bf80c
-- 
2.51.0

[PATCH v3 1/4] sched: Add prev_sum_exec_runtime support for RT, DL and SCX classes

[Gabriele Monaco]

The fair scheduling class relies on prev_sum_exec_runtime to compute the
duration of the task's runtime since it was last scheduled. This value
is currently not required by other scheduling classes but can be useful
to understand long running tasks and take certain actions (e.g. during a
scheduler tick).

Add support for prev_sum_exec_runtime to the RT, deadline and sched_ext
classes by simply assigning the sum_exec_runtime at each set_next_task.

Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Gabriele Monaco <gmonaco@redhat.com>
---
 kernel/sched/deadline.c | 1 +
 kernel/sched/ext.c      | 1 +
 kernel/sched/rt.c       | 1 +
 3 files changed, 3 insertions(+)

diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
index e2d51f4306b3..212d6bf5a732 100644
--- a/kernel/sched/deadline.c
+++ b/kernel/sched/deadline.c
@@ -2342,6 +2342,7 @@ static void set_next_task_dl(struct rq *rq, struct task_struct *p, bool first)
 	p->se.exec_start = rq_clock_task(rq);
 	if (on_dl_rq(&p->dl))
 		update_stats_wait_end_dl(dl_rq, dl_se);
+	p->se.prev_sum_exec_runtime = p->se.sum_exec_runtime;
 
 	/* You can't push away the running task */
 	dequeue_pushable_dl_task(rq, p);
diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
index 7dedc9a16281..7c2d23e6d0df 100644
--- a/kernel/sched/ext.c
+++ b/kernel/sched/ext.c
@@ -3257,6 +3257,7 @@ static void set_next_task_scx(struct rq *rq, struct task_struct *p, bool first)
 	}
 
 	p->se.exec_start = rq_clock_task(rq);
+	p->se.prev_sum_exec_runtime = p->se.sum_exec_runtime;
 
 	/* see dequeue_task_scx() on why we skip when !QUEUED */
 	if (SCX_HAS_OP(sch, running) && (p->scx.flags & SCX_TASK_QUEUED))
diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
index 7936d4333731..8c713d74672a 100644
--- a/kernel/sched/rt.c
+++ b/kernel/sched/rt.c
@@ -1644,6 +1644,7 @@ static inline void set_next_task_rt(struct rq *rq, struct task_struct *p, bool f
 	p->se.exec_start = rq_clock_task(rq);
 	if (on_rt_rq(&p->rt))
 		update_stats_wait_end_rt(rt_rq, rt_se);
+	p->se.prev_sum_exec_runtime = p->se.sum_exec_runtime;
 
 	/* The running task is never eligible for pushing */
 	dequeue_pushable_task(rq, p);
-- 
2.51.0

[PATCH v3 2/4] rseq: Schedule the mm_cid_compaction from rseq_sched_switch_event()

[Gabriele Monaco]

Currently the mm_cid_compaction is triggered by the scheduler tick and
runs in a task_work, behaviour is more unpredictable with periodic tasks
with short runtime, which may rarely run during a tick.

Schedule the mm_cid_compaction from the rseq_sched_switch_event() call
only if the scan is required, that is when the pseudo-period of 100ms
elapsed.

Keep a tick handler used for long running tasks that are never preempted
(i.e. that never call rseq_sched_switch_event), which triggers a
compaction and mm_cid update only in that case.

Signed-off-by: Gabriele Monaco <gmonaco@redhat.com>
---
 include/linux/mm_types.h | 11 +++++++++
 include/linux/rseq.h     |  3 +++
 include/linux/sched.h    |  3 +++
 kernel/sched/core.c      | 48 ++++++++++++++++++++++++++++++++++------
 kernel/sched/sched.h     |  2 ++
 5 files changed, 60 insertions(+), 7 deletions(-)

diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index 08bc2442db93..5dab88707014 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -1424,6 +1424,13 @@ static inline void mm_set_cpus_allowed(struct mm_struct *mm, const struct cpumas
 	WRITE_ONCE(mm->nr_cpus_allowed, cpumask_weight(mm_allowed));
 	raw_spin_unlock(&mm->cpus_allowed_lock);
 }
+
+static inline bool mm_cid_needs_scan(struct mm_struct *mm)
+{
+	if (!mm)
+		return false;
+	return time_after(jiffies, READ_ONCE(mm->mm_cid_next_scan));
+}
 #else /* CONFIG_SCHED_MM_CID */
 static inline void mm_init_cid(struct mm_struct *mm, struct task_struct *p) { }
 static inline int mm_alloc_cid(struct mm_struct *mm, struct task_struct *p) { return 0; }
@@ -1434,6 +1441,10 @@ static inline unsigned int mm_cid_size(void)
 	return 0;
 }
 static inline void mm_set_cpus_allowed(struct mm_struct *mm, const struct cpumask *cpumask) { }
+static inline bool mm_cid_needs_scan(struct mm_struct *mm)
+{
+	return false;
+}
 #endif /* CONFIG_SCHED_MM_CID */
 
 struct mmu_gather;
diff --git a/include/linux/rseq.h b/include/linux/rseq.h
index b8ea95011ec3..12eecde46ff5 100644
--- a/include/linux/rseq.h
+++ b/include/linux/rseq.h
@@ -4,6 +4,7 @@
 
 #ifdef CONFIG_RSEQ
 #include <linux/sched.h>
+#include <linux/mm_types.h>
 
 void __rseq_handle_slowpath(struct pt_regs *regs);
 
@@ -68,6 +69,8 @@ static __always_inline void rseq_sched_switch_event(struct task_struct *t)
 			rseq_raise_notify_resume(t);
 		}
 	}
+	if (mm_cid_needs_scan(t->mm))
+		task_add_mm_cid(t);
 }
 
 /*
diff --git a/include/linux/sched.h b/include/linux/sched.h
index 857ed17d443b..80c1afb2087d 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -1407,6 +1407,7 @@ struct task_struct {
 	int				last_mm_cid;	/* Most recent cid in mm */
 	int				migrate_from_cpu;
 	int				mm_cid_active;	/* Whether cid bitmap is active */
+	unsigned long			last_cid_reset;	/* Time of last reset in jiffies */
 	struct callback_head		cid_work;
 #endif
 
@@ -2300,6 +2301,7 @@ void sched_mm_cid_before_execve(struct task_struct *t);
 void sched_mm_cid_after_execve(struct task_struct *t);
 void sched_mm_cid_fork(struct task_struct *t);
 void sched_mm_cid_exit_signals(struct task_struct *t);
+void task_add_mm_cid(struct task_struct *t);
 static inline int task_mm_cid(struct task_struct *t)
 {
 	return t->mm_cid;
@@ -2309,6 +2311,7 @@ static inline void sched_mm_cid_before_execve(struct task_struct *t) { }
 static inline void sched_mm_cid_after_execve(struct task_struct *t) { }
 static inline void sched_mm_cid_fork(struct task_struct *t) { }
 static inline void sched_mm_cid_exit_signals(struct task_struct *t) { }
+static inline void task_add_mm_cid(struct task_struct *t) { }
 static inline int task_mm_cid(struct task_struct *t)
 {
 	/*
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index e742a655c9a8..30652bb4a223 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -10840,19 +10840,53 @@ void init_sched_mm_cid(struct task_struct *t)
 	init_task_work(&t->cid_work, task_mm_cid_work);
 }
 
+void task_add_mm_cid(struct task_struct *t)
+{
+	struct callback_head *work = &t->cid_work;
+
+	if (work->next != work)
+		return;
+	/* No page allocation under rq lock */
+	task_work_add(t, work, TWA_RESUME);
+}
+
 void task_tick_mm_cid(struct rq *rq, struct task_struct *curr)
 {
-	struct callback_head *work = &curr->cid_work;
-	unsigned long now = jiffies;
+	u64 rtime = curr->se.sum_exec_runtime - curr->se.prev_sum_exec_runtime;
 
+	/*
+	 * If a task is running unpreempted for a long time, it won't get its
+	 * mm_cid compacted and won't update its mm_cid value after a
+	 * compaction occurs.
+	 * For such a task, this function does two things:
+	 * A) trigger the mm_cid recompaction,
+	 * B) trigger an update of the task's rseq->mm_cid field at some point
+	 * after recompaction, so it can get a mm_cid value closer to 0.
+	 * A change in the mm_cid triggers an rseq_preempt.
+	 *
+	 * B occurs once after the compaction work completes, neither A nor B
+	 * run as long as the compaction work is pending, the task is exiting
+	 * or is not a userspace task.
+	 */
 	if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) ||
-	    work->next != work)
+	    test_tsk_thread_flag(curr, TIF_NOTIFY_RESUME))
 		return;
-	if (time_before(now, READ_ONCE(curr->mm->mm_cid_next_scan)))
+	if (rtime < RSEQ_UNPREEMPTED_THRESHOLD)
 		return;
-
-	/* No page allocation under rq lock */
-	task_work_add(curr, work, TWA_RESUME);
+	if (mm_cid_needs_scan(curr->mm)) {
+		/* Trigger mm_cid recompaction */
+		task_add_mm_cid(curr);
+	} else if (time_after(jiffies, curr->last_cid_reset +
+			      msecs_to_jiffies(MM_CID_SCAN_DELAY))) {
+		/* Update mm_cid field */
+		if (!curr->mm_cid_active)
+			return;
+		mm_cid_snapshot_time(rq, curr->mm);
+		mm_cid_put_lazy(curr);
+		curr->last_mm_cid = curr->mm_cid = mm_cid_get(rq, curr, curr->mm);
+		rseq_sched_set_task_mm_cid(curr, curr->mm_cid);
+		rseq_sched_switch_event(curr);
+	}
 }
 
 void sched_mm_cid_exit_signals(struct task_struct *t)
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 8f14d231e7a7..8c0fb3b0fb35 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -3512,6 +3512,7 @@ extern const char *preempt_modes[];
 
 #define SCHED_MM_CID_PERIOD_NS	(100ULL * 1000000)	/* 100ms */
 #define MM_CID_SCAN_DELAY	100			/* 100ms */
+#define RSEQ_UNPREEMPTED_THRESHOLD	SCHED_MM_CID_PERIOD_NS
 
 extern raw_spinlock_t cid_lock;
 extern int use_cid_lock;
@@ -3715,6 +3716,7 @@ static inline int mm_cid_get(struct rq *rq, struct task_struct *t,
 	int cid;
 
 	lockdep_assert_rq_held(rq);
+	t->last_cid_reset = jiffies;
 	cpumask = mm_cidmask(mm);
 	cid = __this_cpu_read(pcpu_cid->cid);
 	if (mm_cid_is_valid(cid)) {
-- 
2.51.0

[PATCH v3 3/4] sched: Compact RSEQ concurrency IDs in batches

[Gabriele Monaco]

Currently, task_mm_cid_work() is called from resume_user_mode_work().
This can delay the execution of the corresponding thread for the entire
duration of the function, negatively affecting the response in case of
real time tasks.
In practice, we observe task_mm_cid_work increasing the latency of
30-35us on a 128 cores system, this order of magnitude is meaningful
under PREEMPT_RT.

Run the task_mm_cid_work in batches of up to CONFIG_RSEQ_CID_SCAN_BATCH
CPUs, this reduces the duration of the delay for each scan.

The task_mm_cid_work contains a mechanism to avoid running more
frequently than every 100ms. Keep this pseudo-periodicity only on
complete scans.
This means each call to task_mm_cid_work returns prematurely if the
period did not elapse and a scan is not ongoing (i.e. the next batch to
scan is not the first).
This way full scans are not excessively delayed while still keeping each
run, and introduced latency, short.

Fixes: 223baf9d17f2 ("sched: Fix performance regression introduced by mm_cid")
Signed-off-by: Gabriele Monaco <gmonaco@redhat.com>
---
 include/linux/mm_types.h | 15 +++++++++++++++
 init/Kconfig             | 12 ++++++++++++
 kernel/sched/core.c      | 31 ++++++++++++++++++++++++++++---
 3 files changed, 55 insertions(+), 3 deletions(-)

diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index 5dab88707014..83f6dc06b15f 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -994,6 +994,13 @@ struct mm_struct {
 		 * When the next mm_cid scan is due (in jiffies).
 		 */
 		unsigned long mm_cid_next_scan;
+		/*
+		 * @mm_cid_scan_batch: Counter for batch used in the next scan.
+		 *
+		 * Scan in batches of CONFIG_RSEQ_CID_SCAN_BATCH. This field
+		 * increments at each scan and reset when all batches are done.
+		 */
+		unsigned int mm_cid_scan_batch;
 		/**
 		 * @nr_cpus_allowed: Number of CPUs allowed for mm.
 		 *
@@ -1389,6 +1396,7 @@ static inline void mm_init_cid(struct mm_struct *mm, struct task_struct *p)
 	raw_spin_lock_init(&mm->cpus_allowed_lock);
 	cpumask_copy(mm_cpus_allowed(mm), &p->cpus_mask);
 	cpumask_clear(mm_cidmask(mm));
+	mm->mm_cid_scan_batch = 0;
 }
 
 static inline int mm_alloc_cid_noprof(struct mm_struct *mm, struct task_struct *p)
@@ -1427,8 +1435,15 @@ static inline void mm_set_cpus_allowed(struct mm_struct *mm, const struct cpumas
 
 static inline bool mm_cid_needs_scan(struct mm_struct *mm)
 {
+	unsigned int next_batch;
+
 	if (!mm)
 		return false;
+	next_batch = READ_ONCE(mm->mm_cid_scan_batch);
+	/* Always needs scan unless it's the first batch. */
+	if (CONFIG_RSEQ_CID_SCAN_BATCH * next_batch < num_possible_cpus() &&
+	    next_batch)
+		return true;
 	return time_after(jiffies, READ_ONCE(mm->mm_cid_next_scan));
 }
 #else /* CONFIG_SCHED_MM_CID */
diff --git a/init/Kconfig b/init/Kconfig
index 854b35e33318..8905d64c2598 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -1918,6 +1918,18 @@ config DEBUG_RSEQ
 
 	  If unsure, say N.
 
+config RSEQ_CID_SCAN_BATCH
+	int "Number of CPUs to scan at every mm_cid compaction attempt"
+	range 1 NR_CPUS
+	default 16
+	depends on SCHED_MM_CID
+	help
+	  CPUs are scanned pseudo-periodically to compact the CID of each task,
+	  this operation can take a longer amount of time on systems with many
+	  CPUs, resulting in higher scheduling latency for the current task.
+	  A higher value means the CID is compacted faster, but results in
+	  higher scheduling latency.
+
 config CACHESTAT_SYSCALL
 	bool "Enable cachestat() system call" if EXPERT
 	default y
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 30652bb4a223..14b79c143d26 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -10784,11 +10784,11 @@ static void sched_mm_cid_remote_clear_weight(struct mm_struct *mm, int cpu,
 
 static void task_mm_cid_work(struct callback_head *work)
 {
+	int weight, cpu, from_cpu, this_batch, next_batch, idx;
 	unsigned long now = jiffies, old_scan, next_scan;
 	struct task_struct *t = current;
 	struct cpumask *cidmask;
 	struct mm_struct *mm;
-	int weight, cpu;
 
 	WARN_ON_ONCE(t != container_of(work, struct task_struct, cid_work));
 
@@ -10798,6 +10798,17 @@ static void task_mm_cid_work(struct callback_head *work)
 	mm = t->mm;
 	if (!mm)
 		return;
+	this_batch = READ_ONCE(mm->mm_cid_scan_batch);
+	next_batch = this_batch + 1;
+	from_cpu = cpumask_nth(this_batch * CONFIG_RSEQ_CID_SCAN_BATCH,
+			       cpu_possible_mask);
+	if (from_cpu >= nr_cpu_ids) {
+		from_cpu = 0;
+		next_batch = 1;
+	}
+	/* Delay scan only if we are done with all cpus. */
+	if (from_cpu != 0)
+		goto cid_compact;
 	old_scan = READ_ONCE(mm->mm_cid_next_scan);
 	next_scan = now + msecs_to_jiffies(MM_CID_SCAN_DELAY);
 	if (!old_scan) {
@@ -10813,17 +10824,31 @@ static void task_mm_cid_work(struct callback_head *work)
 		return;
 	if (!try_cmpxchg(&mm->mm_cid_next_scan, &old_scan, next_scan))
 		return;
+
+cid_compact:
+	if (!try_cmpxchg(&mm->mm_cid_scan_batch, &this_batch, next_batch))
+		return;
 	cidmask = mm_cidmask(mm);
 	/* Clear cids that were not recently used. */
-	for_each_possible_cpu(cpu)
+	idx = 0;
+	cpu = from_cpu;
+	for_each_cpu_from(cpu, cpu_possible_mask) {
+		if (idx++ == CONFIG_RSEQ_CID_SCAN_BATCH)
+			break;
 		sched_mm_cid_remote_clear_old(mm, cpu);
+	}
 	weight = cpumask_weight(cidmask);
 	/*
 	 * Clear cids that are greater or equal to the cidmask weight to
 	 * recompact it.
 	 */
-	for_each_possible_cpu(cpu)
+	idx = 0;
+	cpu = from_cpu;
+	for_each_cpu_from(cpu, cpu_possible_mask) {
+		if (idx++ == CONFIG_RSEQ_CID_SCAN_BATCH)
+			break;
 		sched_mm_cid_remote_clear_weight(mm, cpu, weight);
+	}
 }
 
 void init_sched_mm_cid(struct task_struct *t)
-- 
2.51.0

[PATCH v3 4/4] selftests/rseq: Add test for mm_cid compaction

[Gabriele Monaco]

A task in the kernel (task_mm_cid_work) runs somewhat periodically to
compact the mm_cid for each process. Add a test to validate that it runs
correctly and timely.

The test spawns 1 thread pinned to each CPU, then each thread, including
the main one, runs in short bursts for some time. During this period, the
mm_cids should be spanning all numbers between 0 and nproc.

At the end of this phase, a thread with high enough mm_cid (>= nproc/2)
is selected to be the new leader, all other threads terminate.

After some time, the only running thread should see 0 as mm_cid, if that
doesn't happen, the compaction mechanism didn't work and the test fails.

The test never fails if only 1 core is available, in which case, we
cannot test anything as the only available mm_cid is 0.

Acked-by: Shuah Khan <skhan@linuxfoundation.org>
Signed-off-by: Gabriele Monaco <gmonaco@redhat.com>
---
 tools/testing/selftests/rseq/.gitignore       |   1 +
 tools/testing/selftests/rseq/Makefile         |   2 +-
 .../selftests/rseq/mm_cid_compaction_test.c   | 204 ++++++++++++++++++
 3 files changed, 206 insertions(+), 1 deletion(-)
 create mode 100644 tools/testing/selftests/rseq/mm_cid_compaction_test.c

diff --git a/tools/testing/selftests/rseq/.gitignore b/tools/testing/selftests/rseq/.gitignore
index 0fda241fa62b..b3920c59bf40 100644
--- a/tools/testing/selftests/rseq/.gitignore
+++ b/tools/testing/selftests/rseq/.gitignore
@@ -3,6 +3,7 @@ basic_percpu_ops_test
 basic_percpu_ops_mm_cid_test
 basic_test
 basic_rseq_op_test
+mm_cid_compaction_test
 param_test
 param_test_benchmark
 param_test_compare_twice
diff --git a/tools/testing/selftests/rseq/Makefile b/tools/testing/selftests/rseq/Makefile
index 0d0a5fae5954..bc4d940f66d4 100644
--- a/tools/testing/selftests/rseq/Makefile
+++ b/tools/testing/selftests/rseq/Makefile
@@ -17,7 +17,7 @@ OVERRIDE_TARGETS = 1
 TEST_GEN_PROGS = basic_test basic_percpu_ops_test basic_percpu_ops_mm_cid_test param_test \
 		param_test_benchmark param_test_compare_twice param_test_mm_cid \
 		param_test_mm_cid_benchmark param_test_mm_cid_compare_twice \
-		syscall_errors_test
+		syscall_errors_test mm_cid_compaction_test
 
 TEST_GEN_PROGS_EXTENDED = librseq.so
 
diff --git a/tools/testing/selftests/rseq/mm_cid_compaction_test.c b/tools/testing/selftests/rseq/mm_cid_compaction_test.c
new file mode 100644
index 000000000000..d13623625f5a
--- /dev/null
+++ b/tools/testing/selftests/rseq/mm_cid_compaction_test.c
@@ -0,0 +1,204 @@
+// SPDX-License-Identifier: LGPL-2.1
+#define _GNU_SOURCE
+#include <assert.h>
+#include <pthread.h>
+#include <sched.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <stddef.h>
+
+#include "../kselftest.h"
+#include "rseq.h"
+
+#define VERBOSE 0
+#define printf_verbose(fmt, ...)                    \
+	do {                                        \
+		if (VERBOSE)                        \
+			printf(fmt, ##__VA_ARGS__); \
+	} while (0)
+
+/* 50 ms */
+#define RUNNER_PERIOD 50000
+/*
+ * Number of runs before we terminate or get the token.
+ * The number is slowly increasing with the number of CPUs as the compaction
+ * process can take longer on larger systems. This is an arbitrary value.
+ */
+#define THREAD_RUNS (3 + args->num_cpus/8)
+
+/*
+ * Number of times we check that the mm_cid were compacted.
+ * Checks are repeated every RUNNER_PERIOD.
+ */
+#define MM_CID_COMPACT_TIMEOUT 10
+
+struct thread_args {
+	int cpu;
+	int num_cpus;
+	pthread_mutex_t *token;
+	pthread_barrier_t *barrier;
+	pthread_t *tinfo;
+	struct thread_args *args_head;
+};
+
+static void __noreturn *thread_runner(void *arg)
+{
+	struct thread_args *args = arg;
+	int i, ret, curr_mm_cid;
+	cpu_set_t cpumask;
+
+	CPU_ZERO(&cpumask);
+	CPU_SET(args->cpu, &cpumask);
+	ret = pthread_setaffinity_np(pthread_self(), sizeof(cpumask), &cpumask);
+	if (ret) {
+		errno = ret;
+		perror("Error: failed to set affinity");
+		abort();
+	}
+	pthread_barrier_wait(args->barrier);
+
+	for (i = 0; i < THREAD_RUNS; i++)
+		usleep(RUNNER_PERIOD);
+	curr_mm_cid = rseq_current_mm_cid();
+	/*
+	 * We select one thread with high enough mm_cid to be the new leader.
+	 * All other threads (including the main thread) will terminate.
+	 * After some time, the mm_cid of the only remaining thread should
+	 * converge to 0, if not, the test fails.
+	 */
+	if (curr_mm_cid >= args->num_cpus / 2 &&
+	    !pthread_mutex_trylock(args->token)) {
+		printf_verbose(
+			"cpu%d has mm_cid=%d and will be the new leader.\n",
+			sched_getcpu(), curr_mm_cid);
+		for (i = 0; i < args->num_cpus; i++) {
+			if (args->tinfo[i] == pthread_self())
+				continue;
+			ret = pthread_join(args->tinfo[i], NULL);
+			if (ret) {
+				errno = ret;
+				perror("Error: failed to join thread");
+				abort();
+			}
+		}
+		pthread_barrier_destroy(args->barrier);
+		free(args->tinfo);
+		free(args->token);
+		free(args->barrier);
+		free(args->args_head);
+
+		for (i = 0; i < MM_CID_COMPACT_TIMEOUT; i++) {
+			curr_mm_cid = rseq_current_mm_cid();
+			printf_verbose("run %d: mm_cid=%d on cpu%d.\n", i,
+				       curr_mm_cid, sched_getcpu());
+			if (curr_mm_cid == 0)
+				exit(EXIT_SUCCESS);
+			usleep(RUNNER_PERIOD);
+		}
+		exit(EXIT_FAILURE);
+	}
+	printf_verbose("cpu%d has mm_cid=%d and is going to terminate.\n",
+		       sched_getcpu(), curr_mm_cid);
+	pthread_exit(NULL);
+}
+
+int test_mm_cid_compaction(void)
+{
+	cpu_set_t affinity;
+	int i, j, ret = 0, num_threads;
+	pthread_t *tinfo;
+	pthread_mutex_t *token;
+	pthread_barrier_t *barrier;
+	struct thread_args *args;
+
+	sched_getaffinity(0, sizeof(affinity), &affinity);
+	num_threads = CPU_COUNT(&affinity);
+	tinfo = calloc(num_threads, sizeof(*tinfo));
+	if (!tinfo) {
+		perror("Error: failed to allocate tinfo");
+		return -1;
+	}
+	args = calloc(num_threads, sizeof(*args));
+	if (!args) {
+		perror("Error: failed to allocate args");
+		ret = -1;
+		goto out_free_tinfo;
+	}
+	token = malloc(sizeof(*token));
+	if (!token) {
+		perror("Error: failed to allocate token");
+		ret = -1;
+		goto out_free_args;
+	}
+	barrier = malloc(sizeof(*barrier));
+	if (!barrier) {
+		perror("Error: failed to allocate barrier");
+		ret = -1;
+		goto out_free_token;
+	}
+	if (num_threads == 1) {
+		fprintf(stderr, "Cannot test on a single cpu. "
+				"Skipping mm_cid_compaction test.\n");
+		/* only skipping the test, this is not a failure */
+		goto out_free_barrier;
+	}
+	pthread_mutex_init(token, NULL);
+	ret = pthread_barrier_init(barrier, NULL, num_threads);
+	if (ret) {
+		errno = ret;
+		perror("Error: failed to initialise barrier");
+		goto out_free_barrier;
+	}
+	for (i = 0, j = 0; i < CPU_SETSIZE && j < num_threads; i++) {
+		if (!CPU_ISSET(i, &affinity))
+			continue;
+		args[j].num_cpus = num_threads;
+		args[j].tinfo = tinfo;
+		args[j].token = token;
+		args[j].barrier = barrier;
+		args[j].cpu = i;
+		args[j].args_head = args;
+		if (!j) {
+			/* The first thread is the main one */
+			tinfo[0] = pthread_self();
+			++j;
+			continue;
+		}
+		ret = pthread_create(&tinfo[j], NULL, thread_runner, &args[j]);
+		if (ret) {
+			errno = ret;
+			perror("Error: failed to create thread");
+			abort();
+		}
+		++j;
+	}
+	printf_verbose("Started %d threads.\n", num_threads);
+
+	/* Also main thread will terminate if it is not selected as leader */
+	thread_runner(&args[0]);
+
+	/* only reached in case of errors */
+out_free_barrier:
+	free(barrier);
+out_free_token:
+	free(token);
+out_free_args:
+	free(args);
+out_free_tinfo:
+	free(tinfo);
+
+	return ret;
+}
+
+int main(int argc, char **argv)
+{
+	if (!rseq_mm_cid_available()) {
+		fprintf(stderr, "Error: rseq_mm_cid unavailable\n");
+		return -1;
+	}
+	if (test_mm_cid_compaction())
+		return -1;
+	return 0;
+}
-- 
2.51.0