sched: Restructure task_mm_cid_work for predictability

This patchset moves the task_mm_cid_work to a preemptible and migratable context. This reduces the impact of this task to the scheduling latency of real time tasks.
The change makes the recurrence of the task a bit more predictable. We also add optimisation and fixes to make sure the task_mm_cid_work works as intended.

The behaviour causing latency was introduced in commit 223baf9d17f2 ("sched: Fix performance regression introduced by mm_cid") which introduced a task work tied to the scheduler tick.
That approach presents two possible issues:
* the task work runs before returning to user and causes, in fact, a scheduling latency (with order of magnitude significant in PREEMPT_RT)
* periodic tasks with short runtime are less likely to run during the tick, hence they might not run the task work at all

Patch 1 allows the mm_cids to be actually compacted when a process reduces its number of threads, which was not the case since the same mm_cids were reused to improve cache locality, more details in [3].

Patch 2 contains the main changes, removing the task_work on the scheduler tick and using a delayed_work instead. Additionally, we terminate the call immediately if we see that no mm_cid is actually active, which could happen on processes sleeping for long time or which exited but whose mm has not been freed yet.

Patch 3 adds a selftest to validate the functionality of the task_mm_cid_work (i.e. to compact the mm_cids). The test fails if patch 1 is not applied and is flaky without patch 2. We expect it to always pass with the entire patchset applied.

Changes since V3 [1]:
* Fixes on the selftest
* Minor style issues in comments and indentation
* Use of perror where possible
* Add a barrier to align threads execution
* Improve test failure and error handling

Changes since V2 [2]:
* Change the order of the patches
* Merge patches changing the main delayed_work logic
* Improved self-test to spawn 1 less thread and use the main one instead

Changes since V1 [3]:
* Re-arm the delayed_work at each invocation
* Cancel the work synchronously at mmdrop
* Remove next scan fields and completely rely on the delayed_work
* Shrink mm_cid allocation with nr thread/affinity (Mathieu Desnoyers)
* Add self test

Overhead comparison in [3]

[1] - https://lore.kernel.org/linux-kernel/[email protected]/
[2] - https://lore.kernel.org/linux-kernel/[email protected]/
[3] - https://lore.kernel.org/linux-kernel/[email protected]/

Gabriele Monaco (2):
sched: Move task_mm_cid_work to mm delayed work
rseq/selftests: Add test for mm_cid compaction

Mathieu Desnoyers (1):
sched: Compact RSEQ concurrency IDs with reduced threads and affinity

 include/linux/mm_types.h                      |  23 ++-
 include/linux/sched.h                         |   1 -
 kernel/sched/core.c                           |  66 +------
 kernel/sched/sched.h                          |  32 ++-
 tools/testing/selftests/rseq/.gitignore       |   1 +
 tools/testing/selftests/rseq/Makefile         |   2 +-
 .../selftests/rseq/mm_cid_compaction_test.c   | 185 ++++++++++++++++++
 7 files changed, 231 insertions(+), 79 deletions(-)
 create mode 100644 tools/testing/selftests/rseq/mm_cid_compaction_test.c

base-commit: 5bc55a333a2f7316b58edc7573e8e893f7acb532

2.47.1

From: Mathieu Desnoyers <[email protected]>

When a process reduces its number of threads or clears bits in its CPU affinity mask, the mm_cid allocation should eventually converge towards smaller values.

However, the change introduced by:

commit 7e019dcc470f ("sched: Improve cache locality of RSEQ concurrency IDs for intermittent workloads")

adds a per-mm/CPU recent_cid which is never unset unless a thread migrates.

This is a tradeoff between:

A) Preserving cache locality after a transition from many threads to few threads, or after reducing the hamming weight of the allowed CPU mask.

B) Making the mm_cid upper bounds wrt nr threads and allowed CPU mask easy to document and understand.

C) Allowing applications to eventually react to mm_cid compaction after reduction of the nr threads or allowed CPU mask, making the tracking of mm_cid compaction easier by shrinking it back towards 0 or not.

D) Making sure applications that periodically reduce and then increase again the nr threads or allowed CPU mask still benefit from good cache locality with mm_cid.

Introduce the following changes:

* After shrinking the number of threads or reducing the number of allowed CPUs, reduce the value of max_nr_cid so expansion of CID allocation will preserve cache locality if the number of threads or allowed CPUs increase again.

* Only re-use a recent_cid if it is within the max_nr_cid upper bound, else find the first available CID.

---
 include/linux/mm_types.h |  7 ++++---
 kernel/sched/sched.h     | 25 ++++++++++++++++++++++---
 2 files changed, 26 insertions(+), 6 deletions(-)

index 332cee2856620..14fc1b39c0cf3 100644
@@ -873,10 +873,11 @@ struct mm_struct {
 		 */
 		unsigned int nr_cpus_allowed;
 		/**
-		 * @max_nr_cid: Maximum number of concurrency IDs allocated.
+		 * @max_nr_cid: Maximum number of allowed concurrency
+		 *              IDs allocated.
 		 *
-		 * Track the highest number of concurrency IDs allocated for the
-		 * mm.
+		 * Track the highest number of allowed concurrency IDs
+		 * allocated for the mm.
 		 */
 		atomic_t max_nr_cid;
 		/**

index c5d67a43fe524..13db32370a381 100644
@@ -3666,10 +3666,28 @@ static inline int __mm_cid_try_get(struct task_struct *t, struct mm_struct *mm)
 {
 	struct cpumask *cidmask = mm_cidmask(mm);
 	struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
-	int cid = __this_cpu_read(pcpu_cid->recent_cid);
+	int cid, max_nr_cid, allowed_max_nr_cid;
 
+	/*
+	 * After shrinking the number of threads or reducing the number
+	 * of allowed cpus, reduce the value of max_nr_cid so expansion
+	 * of cid allocation will preserve cache locality if the number
+	 * of threads or allowed cpus increase again.
+	 */
+	max_nr_cid = atomic_read(&mm->max_nr_cid);
+	while ((allowed_max_nr_cid = min_t(int, READ_ONCE(mm->nr_cpus_allowed),
+					   atomic_read(&mm->mm_users))),
+	       max_nr_cid > allowed_max_nr_cid) {
+		/* atomic_try_cmpxchg loads previous mm->max_nr_cid into max_nr_cid. */
+		if (atomic_try_cmpxchg(&mm->max_nr_cid, &max_nr_cid, allowed_max_nr_cid)) {
+			max_nr_cid = allowed_max_nr_cid;
+			break;
+		}
+	}
 	/* Try to re-use recent cid. This improves cache locality. */
-	if (!mm_cid_is_unset(cid) && !cpumask_test_and_set_cpu(cid, cidmask))
+	cid = __this_cpu_read(pcpu_cid->recent_cid);
+	if (!mm_cid_is_unset(cid) && cid < max_nr_cid &&
+	    !cpumask_test_and_set_cpu(cid, cidmask))
 		return cid;
 	/*
 	 * Expand cid allocation if the maximum number of concurrency

@@ -3677,8 +3695,9 @@ static inline int __mm_cid_try_get(struct task_struct *t, struct mm_struct *mm)
 	 * and number of threads. Expanding cid allocation as much as
 	 * possible improves cache locality.
 	 */
-	cid = atomic_read(&mm->max_nr_cid);
+	cid = max_nr_cid;
 	while (cid < READ_ONCE(mm->nr_cpus_allowed) && cid < atomic_read(&mm->mm_users)) {
+		/* atomic_try_cmpxchg loads previous mm->max_nr_cid into cid. */
 		if (!atomic_try_cmpxchg(&mm->max_nr_cid, &cid, cid + 1))
 			continue;
 		if (!cpumask_test_and_set_cpu(cid, cidmask))

2.47.1

Currently, the task_mm_cid_work function is called in a task work triggered by a scheduler tick to frequently compact the mm_cids of each process. 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 a new delayed work connected to the mm_struct rather than in the task context before returning to userspace.

This delayed work is initialised while allocating the mm and disabled before freeing it, its execution is no longer triggered by scheduler ticks but run periodically based on the defined MM_CID_SCAN_DELAY.

The main advantage of this change is that the function can be offloaded to a different CPU and even preempted by RT tasks.

Moreover, this new behaviour could be more predictable with periodic tasks with short runtime, which may rarely run during a scheduler tick. Now, the work is always scheduled with the same periodicity for each mm (though the periodicity is not guaranteed due to interference from other tasks, but mm_cid compaction is mostly best effort).

To avoid excessively increased runtime, we quickly return from the function if we have no work to be done (i.e. no mm_cid is allocated). This is helpful for tasks that sleep for a long time, but also for terminated task. We are no longer following the process' state, hence the function continues to run after a process terminates but before its mm is freed.

---
 include/linux/mm_types.h | 16 ++++++----
 include/linux/sched.h    |  1 -
 kernel/sched/core.c      | 66 +++++-----------------------------------
 kernel/sched/sched.h     |  7 -----
 4 files changed, 18 insertions(+), 72 deletions(-)

index 14fc1b39c0cf3..26d631de6fe1b 100644
@@ -859,12 +859,6 @@ struct mm_struct {
 		 * runqueue locks.
 		 */
 		struct mm_cid __percpu *pcpu_cid;
-		/*
-		 * @mm_cid_next_scan: Next mm_cid scan (in jiffies).
-		 *
-		 * When the next mm_cid scan is due (in jiffies).
-		 */
-		unsigned long mm_cid_next_scan;
 		/**
 		 * @nr_cpus_allowed: Number of CPUs allowed for mm.
 		 *

@@ -887,6 +881,7 @@ struct mm_struct {
 		 * mm nr_cpus_allowed updates.
 		 */
 		raw_spinlock_t cpus_allowed_lock;
+		struct delayed_work mm_cid_work;
 #endif
 #ifdef CONFIG_MMU
 		atomic_long_t pgtables_bytes;	/* size of all page tables */

@@ -1175,11 +1170,16 @@ static inline void vma_iter_init(struct vma_iterator *vmi,
 
 #ifdef CONFIG_SCHED_MM_CID
 
+#define SCHED_MM_CID_PERIOD_NS	(100ULL * 1000000)	/* 100ms */
+#define MM_CID_SCAN_DELAY	100			/* 100ms */
+
 enum mm_cid_state {
 	MM_CID_UNSET = -1U,		/* Unset state has lazy_put flag set. */
 	MM_CID_LAZY_PUT = (1U << 31),
 };
 
+extern void task_mm_cid_work(struct work_struct *work);
+
 static inline bool mm_cid_is_unset(int cid)
 {
 	return cid == MM_CID_UNSET;

@@ -1252,12 +1252,16 @@ static inline int mm_alloc_cid_noprof(struct mm_struct *mm, struct task_struct *
 	if (!mm->pcpu_cid)
 		return -ENOMEM;
 	mm_init_cid(mm, p);
+	INIT_DELAYED_WORK(&mm->mm_cid_work, task_mm_cid_work);
+	schedule_delayed_work(&mm->mm_cid_work,
+			      msecs_to_jiffies(MM_CID_SCAN_DELAY));
 	return 0;
 }
 #define mm_alloc_cid(...)	alloc_hooks(mm_alloc_cid_noprof(__VA_ARGS__))
 
 static inline void mm_destroy_cid(struct mm_struct *mm)
 {
+	disable_delayed_work_sync(&mm->mm_cid_work);
 	free_percpu(mm->pcpu_cid);
 	mm->pcpu_cid = NULL;
 }

index 64934e0830af3..901dadeba3e80 100644
@@ -1381,7 +1381,6 @@ 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 */
-	struct callback_head		cid_work;
 #endif
 
 	struct tlbflush_unmap_batch	tlb_ubc;

index 13db32370a381..4b3396d4c642c 100644
@@ -3590,16 +3590,11 @@ extern void sched_dynamic_update(int mode);
 
 #ifdef CONFIG_SCHED_MM_CID
 
-#define SCHED_MM_CID_PERIOD_NS	(100ULL * 1000000)	/* 100ms */
-#define MM_CID_SCAN_DELAY	100			/* 100ms */
-
 extern raw_spinlock_t cid_lock;
 extern int use_cid_lock;
 
 extern void sched_mm_cid_migrate_from(struct task_struct *t);
 extern void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t);
-extern void task_tick_mm_cid(struct rq *rq, struct task_struct *curr);
-extern void init_sched_mm_cid(struct task_struct *t);
 
 static inline void __mm_cid_put(struct mm_struct *mm, int cid)
 {

@@ -3867,8 +3862,6 @@ static inline void switch_mm_cid(struct rq *rq,
 static inline void switch_mm_cid(struct rq *rq, struct task_struct *prev, struct task_struct *next) { }
 static inline void sched_mm_cid_migrate_from(struct task_struct *t) { }
 static inline void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t) { }
-static inline void task_tick_mm_cid(struct rq *rq, struct task_struct *curr) { }
-static inline void init_sched_mm_cid(struct task_struct *t) { }
 #endif /* !CONFIG_SCHED_MM_CID */
 
 extern u64 avg_vruntime(struct cfs_rq *cfs_rq);

2.47.1

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.

To: Mathieu Desnoyers <[email protected]>

---
 tools/testing/selftests/rseq/.gitignore       |   1 +
 tools/testing/selftests/rseq/Makefile         |   2 +-
 .../selftests/rseq/mm_cid_compaction_test.c   | 185 ++++++++++++++++++
 3 files changed, 187 insertions(+), 1 deletion(-)
 create mode 100644 tools/testing/selftests/rseq/mm_cid_compaction_test.c

index 16496de5f6ce4..2c89f97e4f737 100644
@@ -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

index 5a3432fceb586..ce1b38f46a355 100644
@@ -16,7 +16,7 @@ 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
+		param_test_mm_cid_benchmark param_test_mm_cid_compare_twice mm_cid_compaction_test
 
 TEST_GEN_PROGS_EXTENDED = librseq.so
 

2.47.1