* [PATCH 1/2] sched_ext: Add lazy preemption support
2026-09-17 7:02 [PATCHSET v6 sched_ext/for-7.4] sched_ext: Add lazy preemption support Andrea Righi
@ 2026-09-17 7:02 ` Andrea Righi
2026-09-17 7:02 ` [PATCH 2/2] selftests/sched_ext: Add lazy preemption tests Andrea Righi
1 sibling, 0 replies; 3+ messages in thread
From: Andrea Righi @ 2026-09-17 7:02 UTC (permalink / raw)
To: Tejun Heo, David Vernet, Changwoo Min
Cc: Emil Tsalapatis, Cheng-Yang Chou, sched-ext, linux-kernel
The fair scheduling class can request lazy rescheduling, deferring an
in-kernel scheduling boundary until returning to user space or until the
next scheduler tick. sched_ext only exposes immediate preemption,
preventing BPF schedulers from making the same trade-off.
Add SCX_ENQ_PREEMPT_LAZY and SCX_KICK_PREEMPT_LAZY. Both expire the
current sched_ext task slice but request lazy rescheduling. Immediate
preemption, WAIT and plain kicks take precedence when requests are
combined, while a lazy enqueue to a non-local DSQ retains the
head-insertion semantics of SCX_ENQ_PREEMPT.
Add SCX_OPS_LAZY_SLICE_EXPIRY as the default expiry policy for newly
enabled tasks and initialize it before ops.enable(). Add
scx_bpf_task_set_slice_expiry() so the owning scheduler can override the
policy per task from any callback while preserving sub-scheduler task
ownership boundaries. Bypass continues to force immediate expiry.
For an infinite-slice task whose tick is stopped, set the lazy
reschedule request before restoring the scheduler tick dependency. This
lets the dependency update's IPI serve the request on return to user
space and guarantees that the next tick promotes it otherwise. Use
pinned rq locking for remote kick delivery so RQCF_UPDATED reflects
clock updates since lock acquisition.
Accumulate kick requests independently and resolve precedence while
holding the target rq lock. Reject invalid SCX_KICK_IDLE combinations
and unknown kick flags, which were previously silently ignored.
Signed-off-by: Andrea Righi <arighi@nvidia.com>
---
include/linux/sched/ext.h | 8 +
kernel/sched/ext/ext.c | 152 +++++++++++++++---
kernel/sched/ext/internal.h | 40 ++++-
kernel/sched/ext/sub.c | 15 +-
tools/sched_ext/include/scx/compat.bpf.h | 17 ++
.../sched_ext/include/scx/enum_defs.autogen.h | 3 +
.../sched_ext/include/scx/enums.autogen.bpf.h | 6 +
tools/sched_ext/include/scx/enums.autogen.h | 2 +
.../sched_ext/include/scx/enums_abi.autogen.h | 5 +-
9 files changed, 210 insertions(+), 38 deletions(-)
diff --git a/include/linux/sched/ext.h b/include/linux/sched/ext.h
index 8de69843c2150..7c759baf4813f 100644
--- a/include/linux/sched/ext.h
+++ b/include/linux/sched/ext.h
@@ -278,6 +278,14 @@ struct sched_ext_entity {
*/
bool disallow; /* reject switching into SCX */
+ /*
+ * If set, depletion of this task's slice at the scheduler tick requests
+ * lazy instead of immediate rescheduling. Initialized from
+ * %SCX_OPS_LAZY_SLICE_EXPIRY immediately before ops.enable() and may be
+ * modified afterwards with scx_bpf_task_set_slice_expiry().
+ */
+ bool slice_expires_lazy;
+
/* cold fields */
#ifdef CONFIG_EXT_GROUP_SCHED
struct cgroup *cgrp_moving_from;
diff --git a/kernel/sched/ext/ext.c b/kernel/sched/ext/ext.c
index 4080902a528cd..e1d6fddda1815 100644
--- a/kernel/sched/ext/ext.c
+++ b/kernel/sched/ext/ext.c
@@ -383,11 +383,11 @@ static bool rq_is_open(struct rq *rq, u64 enq_flags)
return true;
/*
- * %SCX_ENQ_PREEMPT clears $curr's slice if on SCX and kicks dispatch,
- * so allow it to avoid spuriously triggering reenq on a combined
+ * The preemption flags clear $curr's slice if on SCX and kick dispatch,
+ * so allow them to avoid spuriously triggering reenq on a combined
* PREEMPT|IMMED insertion.
*/
- if (enq_flags & SCX_ENQ_PREEMPT) {
+ if (enq_flags & (SCX_ENQ_PREEMPT | SCX_ENQ_PREEMPT_LAZY)) {
struct task_struct *curr = rq->curr;
/*
@@ -1515,6 +1515,26 @@ static void call_task_dequeue(struct scx_sched *sch, struct rq *rq,
p->scx.flags &= ~SCX_TASK_IN_CUSTODY;
}
+/*
+ * A task with an infinite slice may be running with its tick stopped. Lazy
+ * rescheduling doesn't send an IPI, so restore the tick dependency to
+ * guarantee that the lazy request is promoted by a real scheduler tick. Set
+ * the lazy request first so that the dependency update's IPI can also serve it
+ * on return to user space.
+ */
+static void scx_resched_curr_lazy(struct rq *rq)
+{
+ resched_curr_lazy(rq);
+
+ if (rq->scx.flags & SCX_RQ_CAN_STOP_TICK) {
+ rq->scx.flags &= ~SCX_RQ_CAN_STOP_TICK;
+ if (rq->clock_update_flags < RQCF_UPDATED)
+ update_rq_clock(rq);
+ update_other_load_avgs(rq);
+ sched_update_tick_dependency(rq);
+ }
+}
+
static void rq_owned_post_enq(struct scx_sched *sch, struct rq *rq,
struct scx_dispatch_q *dsq, struct task_struct *p,
u64 enq_flags)
@@ -1577,12 +1597,16 @@ static void rq_owned_post_enq(struct scx_sched *sch, struct rq *rq,
if (rq->scx.flags & SCX_RQ_IN_DISPATCH)
return;
- if ((enq_flags & SCX_ENQ_PREEMPT) && p != rq->curr &&
+ if ((enq_flags & (SCX_ENQ_PREEMPT | SCX_ENQ_PREEMPT_LAZY)) && p != rq->curr &&
rq->curr->sched_class == &ext_sched_class) {
- if (likely(scx_set_task_slice(rq->curr, 0)))
- resched_curr(rq);
- else
+ if (likely(scx_set_task_slice(rq->curr, 0))) {
+ if (enq_flags & SCX_ENQ_PREEMPT)
+ resched_curr(rq);
+ else
+ scx_resched_curr_lazy(rq);
+ } else {
__scx_add_event(sch, SCX_EV_SLICE_DENIED, 1);
+ }
}
}
@@ -1672,7 +1696,7 @@ static void scx_dispatch_enqueue(struct scx_sched *sch, struct rq *rq,
scx_error(sch, "DSQ ID 0x%016llx already had PRIQ-enqueued tasks",
dsq->id);
- if (enq_flags & (SCX_ENQ_HEAD | SCX_ENQ_PREEMPT)) {
+ if (enq_flags & (SCX_ENQ_HEAD | SCX_ENQ_PREEMPT | SCX_ENQ_PREEMPT_LAZY)) {
/* new task inserted at head - use fastpath */
if (dsq_insert_head(dsq, p) && !(dsq->id & SCX_DSQ_FLAG_BUILTIN))
rcu_assign_pointer(dsq->first_task, p);
@@ -2386,7 +2410,7 @@ void scx_move_local_task_to_local_dsq(struct scx_sched *sch, struct task_struct
WARN_ON_ONCE(p->scx.holding_cpu >= 0);
- if (enq_flags & (SCX_ENQ_HEAD | SCX_ENQ_PREEMPT))
+ if (enq_flags & (SCX_ENQ_HEAD | SCX_ENQ_PREEMPT | SCX_ENQ_PREEMPT_LAZY))
dsq_insert_head(dst_dsq, p);
else
list_add_tail(&p->scx.dsq_list.node, &dst_dsq->list);
@@ -3804,8 +3828,14 @@ static void task_tick_scx(struct rq *rq, struct task_struct *curr, int queued)
else if (SCX_HAS_OP(sch, tick))
SCX_CALL_OP_TASK(sch, tick, rq, curr);
- if (!curr->scx.slice)
- resched_curr(rq);
+ if (!curr->scx.slice) {
+ /* the slice can't be trusted while bypassing */
+ if (READ_ONCE(curr->scx.slice_expires_lazy) &&
+ !scx_bypassing(sch, cpu_of(rq)))
+ scx_resched_curr_lazy(rq);
+ else
+ resched_curr(rq);
+ }
}
#ifdef CONFIG_EXT_GROUP_SCHED
@@ -3921,6 +3951,7 @@ static void __scx_enable_task(struct scx_sched *sch, struct task_struct *p)
weight = sched_prio_to_weight[p->static_prio - MAX_RT_PRIO];
p->scx.weight = sched_weight_to_cgroup(weight);
+ p->scx.slice_expires_lazy = sch->ops.flags & SCX_OPS_LAZY_SLICE_EXPIRY;
if (SCX_HAS_OP(sch, enable))
SCX_CALL_OP_TASK(sch, enable, rq, p);
@@ -5371,6 +5402,7 @@ static void scx_sched_free_rcu_work(struct work_struct *work)
free_cpumask_var(pcpu->cpus_to_kick);
free_cpumask_var(pcpu->cpus_to_kick_if_idle);
free_cpumask_var(pcpu->cpus_to_preempt);
+ free_cpumask_var(pcpu->cpus_to_preempt_lazy);
free_cpumask_var(pcpu->cpus_to_wait);
exit_dsq(scx_bypass_dsq(sch, cpu));
@@ -6883,6 +6915,9 @@ static void scx_dump_cpu(struct scx_sched *sch, struct seq_buf *s,
if (!cpumask_empty(pcpu->cpus_to_preempt))
scx_dump_line(&ns, " cpus_to_preempt: %*pb",
cpumask_pr_args(pcpu->cpus_to_preempt));
+ if (!cpumask_empty(pcpu->cpus_to_preempt_lazy))
+ scx_dump_line(&ns, " preempt_lazy : %*pb",
+ cpumask_pr_args(pcpu->cpus_to_preempt_lazy));
if (!cpumask_empty(pcpu->cpus_to_wait))
scx_dump_line(&ns, " cpus_to_wait : %*pb",
cpumask_pr_args(pcpu->cpus_to_wait));
@@ -7200,6 +7235,7 @@ struct scx_sched *scx_alloc_and_add_sched(struct scx_enable_cmd *cmd,
if (!zalloc_cpumask_var_node(&pcpu->cpus_to_kick, GFP_KERNEL, node) ||
!zalloc_cpumask_var_node(&pcpu->cpus_to_kick_if_idle, GFP_KERNEL, node) ||
!zalloc_cpumask_var_node(&pcpu->cpus_to_preempt, GFP_KERNEL, node) ||
+ !zalloc_cpumask_var_node(&pcpu->cpus_to_preempt_lazy, GFP_KERNEL, node) ||
!zalloc_cpumask_var_node(&pcpu->cpus_to_wait, GFP_KERNEL, node)) {
ret = -ENOMEM;
goto err_free_pcpu;
@@ -7335,6 +7371,7 @@ struct scx_sched *scx_alloc_and_add_sched(struct scx_enable_cmd *cmd,
free_cpumask_var(pcpu->cpus_to_kick);
free_cpumask_var(pcpu->cpus_to_kick_if_idle);
free_cpumask_var(pcpu->cpus_to_preempt);
+ free_cpumask_var(pcpu->cpus_to_preempt_lazy);
free_cpumask_var(pcpu->cpus_to_wait);
}
for_each_possible_cpu(cpu) {
@@ -8459,13 +8496,23 @@ static bool kick_one_cpu(s32 cpu, struct scx_sched_pcpu *pcpu, struct rq *this_r
{
struct rq *rq = cpu_rq(cpu);
struct scx_rq *this_scx = &this_rq->scx;
+ struct rq_flags rf;
const struct sched_class *cur_class;
bool should_wait = false;
bool kickable;
- unsigned long flags;
+ bool preempt, preempt_lazy, wait, immediate;
- raw_spin_rq_lock_irqsave(rq, flags);
+ rq_lock_irqsave(rq, &rf);
cur_class = rq->curr->sched_class;
+ preempt = cpumask_test_cpu(cpu, pcpu->cpus_to_preempt);
+ preempt_lazy = cpumask_test_cpu(cpu, pcpu->cpus_to_preempt_lazy);
+ wait = cpumask_test_cpu(cpu, pcpu->cpus_to_wait);
+ /*
+ * Immediate preemption, waiting and a plain kick take precedence over
+ * lazy preemption. The lazy request still clears the slice, so all
+ * accumulated requests are served.
+ */
+ immediate = preempt || wait || cpumask_test_cpu(cpu, pcpu->cpus_to_kick);
/*
* During CPU hotplug, a CPU may depend on kicking itself to make
@@ -8479,19 +8526,23 @@ static bool kick_one_cpu(s32 cpu, struct scx_sched_pcpu *pcpu, struct rq *this_r
!sched_class_above(cur_class, &ext_sched_class);
if (kickable && !scx_missing_caps(pcpu->sch, cpu, SCX_CAP_BASE)) {
- if (cpumask_test_cpu(cpu, pcpu->cpus_to_preempt)) {
+ if (preempt || preempt_lazy) {
if (cur_class == &ext_sched_class) {
u64 caps = scx_caps_for_preempt(pcpu->sch, rq, 0);
- if (unlikely(scx_missing_caps(pcpu->sch, cpu, caps)))
+ if (unlikely(scx_missing_caps(pcpu->sch, cpu, caps))) {
__scx_add_event(pcpu->sch, SCX_EV_SUB_PREEMPT_DENIED, 1);
- else if (unlikely(!scx_set_task_slice(rq->curr, 0)))
+ /* degrade to a plain, immediate kick */
+ immediate = true;
+ } else if (unlikely(!scx_set_task_slice(rq->curr, 0))) {
__scx_add_event(pcpu->sch, SCX_EV_SLICE_DENIED, 1);
+ }
}
cpumask_clear_cpu(cpu, pcpu->cpus_to_preempt);
+ cpumask_clear_cpu(cpu, pcpu->cpus_to_preempt_lazy);
}
- if (cpumask_test_cpu(cpu, pcpu->cpus_to_wait)) {
+ if (wait) {
if (cur_class == &ext_sched_class) {
cpumask_set_cpu(cpu, this_scx->cpus_to_sync);
ksyncs[cpu] = rq->scx.kick_sync;
@@ -8500,17 +8551,21 @@ static bool kick_one_cpu(s32 cpu, struct scx_sched_pcpu *pcpu, struct rq *this_r
cpumask_clear_cpu(cpu, pcpu->cpus_to_wait);
}
- resched_curr(rq);
+ if (immediate)
+ resched_curr(rq);
+ else
+ scx_resched_curr_lazy(rq);
} else {
/* a kickable cpu was skipped solely for the missing caps */
if (kickable)
__scx_add_event(pcpu->sch, SCX_EV_SUB_KICK_DENIED, 1);
cpumask_clear_cpu(cpu, pcpu->cpus_to_preempt);
+ cpumask_clear_cpu(cpu, pcpu->cpus_to_preempt_lazy);
cpumask_clear_cpu(cpu, pcpu->cpus_to_wait);
}
scx_rq_lock_drop(rq);
- raw_spin_rq_unlock_irqrestore(rq, flags);
+ rq_unlock_irqrestore(rq, &rf);
return should_wait;
}
@@ -8566,6 +8621,14 @@ static void kick_cpus_irq_workfn(struct irq_work *irq_work)
cpumask_clear_cpu(cpu, pcpu->cpus_to_kick_if_idle);
}
+ /*
+ * kick_one_cpu() clears the lazy bit of every cpu it visited
+ * above; visit the remaining requests which contain lazy
+ * preemption, see scx_kick_cpu().
+ */
+ for_each_cpu(cpu, pcpu->cpus_to_preempt_lazy)
+ should_wait |= kick_one_cpu(cpu, pcpu, this_rq, ksyncs);
+
for_each_cpu(cpu, pcpu->cpus_to_kick_if_idle) {
kick_one_cpu_if_idle(cpu, pcpu, this_rq);
cpumask_clear_cpu(cpu, pcpu->cpus_to_kick_if_idle);
@@ -9531,6 +9594,31 @@ __bpf_kfunc bool scx_bpf_task_set_dsq_vtime(struct task_struct *p, u64 vtime,
return true;
}
+/**
+ * scx_bpf_task_set_slice_expiry - Set task's slice expiry policy
+ * @p: task of interest
+ * @lazy: whether slice expiry should request lazy rescheduling
+ * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
+ *
+ * Choose whether depletion of @p's slice at the scheduler tick requests lazy
+ * or immediate rescheduling. @p must be on the calling scheduler.
+ *
+ * Return %true on success, %false if @p is not on the calling scheduler.
+ */
+__bpf_kfunc bool scx_bpf_task_set_slice_expiry(struct task_struct *p, bool lazy,
+ const struct bpf_prog_aux *aux)
+{
+ struct scx_sched *sch;
+
+ guard(rcu)();
+ sch = scx_prog_sched(aux);
+ if (unlikely(!sch || !scx_task_on_sched(sch, p)))
+ return false;
+
+ WRITE_ONCE(p->scx.slice_expires_lazy, lazy);
+ return true;
+}
+
void scx_kick_cpu(struct scx_sched *sch, s32 cpu, u64 flags)
{
struct scx_sched_pcpu *pcpu;
@@ -9539,6 +9627,16 @@ void scx_kick_cpu(struct scx_sched *sch, s32 cpu, u64 flags)
if (!scx_kf_allowed_ctx(sch))
return;
+ if (unlikely(flags & ~(SCX_KICK_IDLE | SCX_KICK_PREEMPT | SCX_KICK_WAIT |
+ SCX_KICK_PREEMPT_LAZY))) {
+ scx_error(sch, "invalid kick flags 0x%llx", flags);
+ return;
+ }
+ if (unlikely((flags & SCX_KICK_IDLE) &&
+ (flags & (SCX_KICK_PREEMPT | SCX_KICK_PREEMPT_LAZY | SCX_KICK_WAIT)))) {
+ scx_error(sch, "PREEMPT/WAIT cannot be used with SCX_KICK_IDLE");
+ return;
+ }
local_irq_save(irq_flags);
@@ -9564,9 +9662,6 @@ void scx_kick_cpu(struct scx_sched *sch, s32 cpu, u64 flags)
if (flags & SCX_KICK_IDLE) {
struct rq *target_rq = cpu_rq(cpu);
- if (unlikely(flags & (SCX_KICK_PREEMPT | SCX_KICK_WAIT)))
- scx_error(sch, "PREEMPT/WAIT cannot be used with SCX_KICK_IDLE");
-
if (raw_spin_rq_trylock(target_rq)) {
if (can_skip_idle_kick(target_rq)) {
scx_rq_lock_drop(target_rq);
@@ -9578,12 +9673,15 @@ void scx_kick_cpu(struct scx_sched *sch, s32 cpu, u64 flags)
}
cpumask_set_cpu(cpu, pcpu->cpus_to_kick_if_idle);
} else {
- cpumask_set_cpu(cpu, pcpu->cpus_to_kick);
-
+ /* Accumulate requests and resolve their precedence at delivery. */
if (flags & SCX_KICK_PREEMPT)
cpumask_set_cpu(cpu, pcpu->cpus_to_preempt);
+ if (flags & SCX_KICK_PREEMPT_LAZY)
+ cpumask_set_cpu(cpu, pcpu->cpus_to_preempt_lazy);
if (flags & SCX_KICK_WAIT)
cpumask_set_cpu(cpu, pcpu->cpus_to_wait);
+ if (!(flags & SCX_KICK_PREEMPT_LAZY))
+ cpumask_set_cpu(cpu, pcpu->cpus_to_kick);
}
if (list_empty(&pcpu->to_kick_node))
@@ -9623,8 +9721,9 @@ __bpf_kfunc void scx_bpf_kick_cpu(s32 cpu, u64 flags, const struct bpf_prog_aux
* cid-addressed equivalent of scx_bpf_kick_cpu(). An invalid @cid aborts the
* scheduler via scx_cid_to_cpu(). Caps are enforced on the delivery path: a
* kick is dropped if the caller lacks baseline access on @cid, and a
- * %SCX_KICK_PREEMPT degrades to a plain reschedule if the caller lacks
- * %SCX_CAP_PREEMPT for a task outside its subtree.
+ * %SCX_KICK_PREEMPT or %SCX_KICK_PREEMPT_LAZY request degrades to a plain
+ * reschedule if the caller lacks %SCX_CAP_PREEMPT for a task outside its
+ * subtree.
*/
__bpf_kfunc void scx_bpf_kick_cid(s32 cid, u64 flags, const struct bpf_prog_aux *aux)
{
@@ -10691,6 +10790,7 @@ __bpf_kfunc_end_defs();
BTF_KFUNCS_START(scx_kfunc_ids_any)
BTF_ID_FLAGS(func, scx_bpf_task_set_slice, KF_IMPLICIT_ARGS | KF_RCU);
BTF_ID_FLAGS(func, scx_bpf_task_set_dsq_vtime, KF_IMPLICIT_ARGS | KF_RCU);
+BTF_ID_FLAGS(func, scx_bpf_task_set_slice_expiry, KF_IMPLICIT_ARGS | KF_RCU);
BTF_ID_FLAGS(func, scx_bpf_kick_cpu, KF_IMPLICIT_ARGS)
BTF_ID_FLAGS(func, scx_bpf_kick_cid, KF_IMPLICIT_ARGS)
BTF_ID_FLAGS(func, scx_bpf_dsq_nr_queued, KF_IMPLICIT_ARGS)
diff --git a/kernel/sched/ext/internal.h b/kernel/sched/ext/internal.h
index e1eb3a0d456cb..bb527e891c8cd 100644
--- a/kernel/sched/ext/internal.h
+++ b/kernel/sched/ext/internal.h
@@ -215,6 +215,20 @@ enum scx_ops_flags {
*/
SCX_OPS_TID_TO_TASK = 1LLU << 8,
+ /*
+ * If set, tasks default to requesting lazy rescheduling when their slice
+ * runs out at the tick, the way fair.c expires a slice from update_curr().
+ * The default is copied to p->scx.slice_expires_lazy immediately before
+ * ops.enable(), after which scx_bpf_task_set_slice_expiry() may override
+ * it per task.
+ * A task in user space still reschedules on the way back from the tick; a
+ * task in the kernel runs on to its next return to user space or to the
+ * next tick, which promotes the request. When lazy preemption is disabled
+ * at runtime, the request behaves like an immediate reschedule.
+ * Rescheduling while bypassing stays immediate.
+ */
+ SCX_OPS_LAZY_SLICE_EXPIRY = 1LLU << 9,
+
SCX_OPS_ALL_FLAGS = SCX_OPS_KEEP_BUILTIN_IDLE |
SCX_OPS_ENQ_LAST |
SCX_OPS_ENQ_EXITING |
@@ -223,7 +237,8 @@ enum scx_ops_flags {
SCX_OPS_SWITCH_PARTIAL |
SCX_OPS_BUILTIN_IDLE_PER_NODE |
SCX_OPS_ALWAYS_ENQ_IMMED |
- SCX_OPS_TID_TO_TASK,
+ SCX_OPS_TID_TO_TASK |
+ SCX_OPS_LAZY_SLICE_EXPIRY,
/* high 8 bits are internal, don't include in SCX_OPS_ALL_FLAGS */
__SCX_OPS_INTERNAL_MASK = 0xffLLU << 56,
@@ -1327,6 +1342,7 @@ struct scx_sched_pcpu {
cpumask_var_t cpus_to_kick;
cpumask_var_t cpus_to_kick_if_idle;
cpumask_var_t cpus_to_preempt;
+ cpumask_var_t cpus_to_preempt_lazy;
cpumask_var_t cpus_to_wait;
struct list_head to_kick_node;
@@ -1407,15 +1423,16 @@ struct scx_sched_pnode {
* the allocation pattern.
*
* ENQ_IMMED insert an IMMED task onto the cid's local DSQ
- * - kick the cid's cpu (except SCX_KICK_PREEMPT)
+ * - kick the cid's cpu (except SCX_KICK_PREEMPT and
+ * SCX_KICK_PREEMPT_LAZY)
*
* ENQ insert any task onto the cid's local DSQ (implies ENQ_IMMED)
*
* PREEMPT preempt any task running on the cid regardless of the owning
* sched (implies ENQ). Preempting a task in the sched's own subtree
* doesn't require any cap.
- * - SCX_ENQ_PREEMPT inserts
- * - SCX_KICK_PREEMPT kicks
+ * - SCX_ENQ_PREEMPT and SCX_ENQ_PREEMPT_LAZY insert
+ * - SCX_KICK_PREEMPT and SCX_KICK_PREEMPT_LAZY kick
*
* PERF control the cid's cpu power/perf management state, currently the
* cpufreq target set through scx_bpf_cidperf_set(). Hardware
@@ -1685,6 +1702,15 @@ enum scx_enq_flags {
*/
SCX_ENQ_PREEMPT = 1LLU << 32,
+ /*
+ * Like %SCX_ENQ_PREEMPT, but request lazy rescheduling. The current
+ * task's slice is still cleared immediately so that the next scheduling
+ * boundary observes the new ordering. %SCX_ENQ_PREEMPT takes precedence
+ * if both are specified. Implies %SCX_ENQ_HEAD, which is all it means on
+ * a non-local DSQ, as with %SCX_ENQ_PREEMPT.
+ */
+ SCX_ENQ_PREEMPT_LAZY = 1LLU << 35,
+
/*
* Only allowed on local DSQs. Guarantees that the task either gets
* on the CPU immediately and stays on it, or gets reenqueued back
@@ -1811,6 +1837,12 @@ enum scx_kick_flags {
* is not on SCX.
*/
SCX_KICK_WAIT = 1LLU << 2,
+
+ /*
+ * Like %SCX_KICK_PREEMPT, but request lazy rescheduling. If combined
+ * with %SCX_KICK_PREEMPT or %SCX_KICK_WAIT, rescheduling is immediate.
+ */
+ SCX_KICK_PREEMPT_LAZY = 1LLU << 3,
};
enum scx_tg_flags {
diff --git a/kernel/sched/ext/sub.c b/kernel/sched/ext/sub.c
index 380a5653dc529..a88b87614b55e 100644
--- a/kernel/sched/ext/sub.c
+++ b/kernel/sched/ext/sub.c
@@ -686,9 +686,10 @@ void scx_rescue_init(struct rq *rq)
* rescue is enabled, or @rq's reject DSQ after recording the reenq reason on
* @p.
*
- * %SCX_ENQ_IMMED, %SCX_ENQ_PREEMPT and %SCX_ENQ_HEAD are cleared when diverting
- * to rescue or reject. %SCX_ENQ_PREEMPT is also cleared on a fallback
- * migration-disabled admission.
+ * %SCX_ENQ_IMMED, %SCX_ENQ_PREEMPT, %SCX_ENQ_PREEMPT_LAZY and %SCX_ENQ_HEAD
+ * are cleared when diverting to rescue or reject. %SCX_ENQ_PREEMPT and
+ * %SCX_ENQ_PREEMPT_LAZY are also cleared on a fallback migration-disabled
+ * admission.
*
* Bypass doesn't need special-casing as a bypassing sched's tasks are enqueued
* to and run by its nearest non-bypassing ancestor. If root is bypassing, it
@@ -709,7 +710,7 @@ struct scx_dispatch_q *scx_resolve_local_dsq(struct scx_sched *sch, struct rq *r
* On a remote activation the scheduling sched (@asch) differs from
* @p's owner (@sch). Check caps against the scheduling sched.
*/
- if (*enq_flags & SCX_ENQ_PREEMPT)
+ if (*enq_flags & (SCX_ENQ_PREEMPT | SCX_ENQ_PREEMPT_LAZY))
needed |= scx_caps_for_preempt(asch, rq, *enq_flags);
missing = scx_missing_caps(asch, cpu_of(rq), needed);
@@ -726,7 +727,7 @@ struct scx_dispatch_q *scx_resolve_local_dsq(struct scx_sched *sch, struct rq *r
if (unlikely(!scx_rq_online(rq) || is_migration_disabled(p) ||
p->migration_pending)) {
__scx_add_event(sch, SCX_EV_SUB_FORCED_ADMIT, 1);
- *enq_flags &= ~SCX_ENQ_PREEMPT;
+ *enq_flags &= ~(SCX_ENQ_PREEMPT | SCX_ENQ_PREEMPT_LAZY);
return &rq->scx.local_dsq;
}
@@ -735,8 +736,8 @@ struct scx_dispatch_q *scx_resolve_local_dsq(struct scx_sched *sch, struct rq *r
* or HEAD - a diversion has no priority and IMMED is not allowed on
* non-local DSQs. Strip the enq and task flags along with the slice.
*/
- *enq_flags &= ~(SCX_ENQ_IMMED | SCX_ENQ_PREEMPT | SCX_ENQ_HEAD |
- SCX_ENQ_APPLY_SLICE | SCX_ENQ_SLICE_DFL);
+ *enq_flags &= ~(SCX_ENQ_IMMED | SCX_ENQ_PREEMPT | SCX_ENQ_PREEMPT_LAZY |
+ SCX_ENQ_HEAD | SCX_ENQ_APPLY_SLICE | SCX_ENQ_SLICE_DFL);
p->scx.flags &= ~SCX_TASK_IMMED;
/* the enqueuer opted for rescue instead of rejection and reenqueue */
diff --git a/tools/sched_ext/include/scx/compat.bpf.h b/tools/sched_ext/include/scx/compat.bpf.h
index 6944221f96cc0..9491acb8d059f 100644
--- a/tools/sched_ext/include/scx/compat.bpf.h
+++ b/tools/sched_ext/include/scx/compat.bpf.h
@@ -403,6 +403,23 @@ static inline void scx_bpf_task_set_dsq_vtime(struct task_struct *p, u64 vtime)
p->scx.dsq_vtime = vtime;
}
+/*
+ * v7.4: Add per-task control over whether slice expiry requests lazy or
+ * immediate rescheduling. Older kernels have no equivalent operation, so the
+ * compatibility wrapper is a no-op when the kfunc is unavailable. As with the
+ * setters above, discard the kfunc's authority result for consistency.
+ *
+ * Keep the wrapper until pre-v7.4 kernels fall out of the sched_ext scheduler
+ * support window.
+ */
+bool scx_bpf_task_set_slice_expiry___new(struct task_struct *p, bool lazy) __ksym __weak;
+
+static inline void scx_bpf_task_set_slice_expiry(struct task_struct *p, bool lazy)
+{
+ if (bpf_ksym_exists(scx_bpf_task_set_slice_expiry___new))
+ scx_bpf_task_set_slice_expiry___new(p, lazy);
+}
+
/*
* v7.1: New scx_bpf_dsq_reenq() that allows re-enqueues on more DSQs. This
* will eventually deprecate scx_bpf_reenqueue_local().
diff --git a/tools/sched_ext/include/scx/enum_defs.autogen.h b/tools/sched_ext/include/scx/enum_defs.autogen.h
index 63b6b14b19bd4..5d25a29ab9782 100644
--- a/tools/sched_ext/include/scx/enum_defs.autogen.h
+++ b/tools/sched_ext/include/scx/enum_defs.autogen.h
@@ -85,6 +85,7 @@
#define HAVE_SCX_ENQ_HEAD
#define HAVE_SCX_ENQ_CPU_SELECTED
#define HAVE_SCX_ENQ_PREEMPT
+#define HAVE_SCX_ENQ_PREEMPT_LAZY
#define HAVE_SCX_ENQ_IMMED
#define HAVE_SCX_ENQ_RESCUE
#define HAVE_SCX_ENQ_REENQ
@@ -149,6 +150,7 @@
#define HAVE_SCX_KICK_IDLE
#define HAVE_SCX_KICK_PREEMPT
#define HAVE_SCX_KICK_WAIT
+#define HAVE_SCX_KICK_PREEMPT_LAZY
#define HAVE_SCX_OPI_BEGIN
#define HAVE_SCX_OPI_NORMAL_BEGIN
#define HAVE_SCX_OPI_NORMAL_END
@@ -164,6 +166,7 @@
#define HAVE_SCX_OPS_BUILTIN_IDLE_PER_NODE
#define HAVE_SCX_OPS_ALWAYS_ENQ_IMMED
#define HAVE_SCX_OPS_TID_TO_TASK
+#define HAVE_SCX_OPS_LAZY_SLICE_EXPIRY
#define HAVE_SCX_OPS_ALL_FLAGS
#define HAVE___SCX_OPS_INTERNAL_MASK
#define HAVE_SCX_OPS_HAS_CPU_PREEMPT
diff --git a/tools/sched_ext/include/scx/enums.autogen.bpf.h b/tools/sched_ext/include/scx/enums.autogen.bpf.h
index 7268131010de3..6afd4179cce50 100644
--- a/tools/sched_ext/include/scx/enums.autogen.bpf.h
+++ b/tools/sched_ext/include/scx/enums.autogen.bpf.h
@@ -112,6 +112,9 @@ const volatile u64 __SCX_KICK_PREEMPT __weak;
const volatile u64 __SCX_KICK_WAIT __weak;
#define SCX_KICK_WAIT __SCX_KICK_WAIT
+const volatile u64 __SCX_KICK_PREEMPT_LAZY __weak;
+#define SCX_KICK_PREEMPT_LAZY __SCX_KICK_PREEMPT_LAZY
+
const volatile u64 __SCX_ENQ_WAKEUP __weak;
#define SCX_ENQ_WAKEUP __SCX_ENQ_WAKEUP
@@ -121,6 +124,9 @@ const volatile u64 __SCX_ENQ_HEAD __weak;
const volatile u64 __SCX_ENQ_PREEMPT __weak;
#define SCX_ENQ_PREEMPT __SCX_ENQ_PREEMPT
+const volatile u64 __SCX_ENQ_PREEMPT_LAZY __weak;
+#define SCX_ENQ_PREEMPT_LAZY __SCX_ENQ_PREEMPT_LAZY
+
const volatile u64 __SCX_ENQ_IMMED __weak;
#define SCX_ENQ_IMMED __SCX_ENQ_IMMED
diff --git a/tools/sched_ext/include/scx/enums.autogen.h b/tools/sched_ext/include/scx/enums.autogen.h
index e616326545172..16e500b10fc65 100644
--- a/tools/sched_ext/include/scx/enums.autogen.h
+++ b/tools/sched_ext/include/scx/enums.autogen.h
@@ -41,9 +41,11 @@
SCX_ENUM_SET(skel, scx_kick_flags, SCX_KICK_IDLE); \
SCX_ENUM_SET(skel, scx_kick_flags, SCX_KICK_PREEMPT); \
SCX_ENUM_SET(skel, scx_kick_flags, SCX_KICK_WAIT); \
+ SCX_ENUM_SET(skel, scx_kick_flags, SCX_KICK_PREEMPT_LAZY); \
SCX_ENUM_SET(skel, scx_enq_flags, SCX_ENQ_WAKEUP); \
SCX_ENUM_SET(skel, scx_enq_flags, SCX_ENQ_HEAD); \
SCX_ENUM_SET(skel, scx_enq_flags, SCX_ENQ_PREEMPT); \
+ SCX_ENUM_SET(skel, scx_enq_flags, SCX_ENQ_PREEMPT_LAZY); \
SCX_ENUM_SET(skel, scx_enq_flags, SCX_ENQ_IMMED); \
SCX_ENUM_SET(skel, scx_enq_flags, SCX_ENQ_RESCUE); \
SCX_ENUM_SET(skel, scx_enq_flags, SCX_ENQ_REENQ); \
diff --git a/tools/sched_ext/include/scx/enums_abi.autogen.h b/tools/sched_ext/include/scx/enums_abi.autogen.h
index d53899764f5ac..d6f38bebd8134 100644
--- a/tools/sched_ext/include/scx/enums_abi.autogen.h
+++ b/tools/sched_ext/include/scx/enums_abi.autogen.h
@@ -97,6 +97,7 @@ static const struct __scx_enum_abi_val __scx_enum_abi_vals[]
{ "scx_enq_flags", "SCX_ENQ_HEAD", 0x10000LLU },
{ "scx_enq_flags", "SCX_ENQ_CPU_SELECTED", 0x100000LLU },
{ "scx_enq_flags", "SCX_ENQ_PREEMPT", 0x100000000LLU },
+ { "scx_enq_flags", "SCX_ENQ_PREEMPT_LAZY", 0x800000000LLU },
{ "scx_enq_flags", "SCX_ENQ_IMMED", 0x200000000LLU },
{ "scx_enq_flags", "SCX_ENQ_RESCUE", 0x400000000LLU },
{ "scx_enq_flags", "SCX_ENQ_REENQ", 0x10000000000LLU },
@@ -161,6 +162,7 @@ static const struct __scx_enum_abi_val __scx_enum_abi_vals[]
{ "scx_kick_flags", "SCX_KICK_IDLE", 0x1LLU },
{ "scx_kick_flags", "SCX_KICK_PREEMPT", 0x2LLU },
{ "scx_kick_flags", "SCX_KICK_WAIT", 0x4LLU },
+ { "scx_kick_flags", "SCX_KICK_PREEMPT_LAZY", 0x8LLU },
{ "scx_opi", "SCX_OPI_BEGIN", 0x0LLU },
{ "scx_opi", "SCX_OPI_NORMAL_BEGIN", 0x0LLU },
{ "scx_opi", "SCX_OPI_NORMAL_END", 0x21LLU },
@@ -176,7 +178,8 @@ static const struct __scx_enum_abi_val __scx_enum_abi_vals[]
{ "scx_ops_flags", "SCX_OPS_BUILTIN_IDLE_PER_NODE", 0x40LLU },
{ "scx_ops_flags", "SCX_OPS_ALWAYS_ENQ_IMMED", 0x80LLU },
{ "scx_ops_flags", "SCX_OPS_TID_TO_TASK", 0x100LLU },
- { "scx_ops_flags", "SCX_OPS_ALL_FLAGS", 0x1ffLLU },
+ { "scx_ops_flags", "SCX_OPS_LAZY_SLICE_EXPIRY", 0x200LLU },
+ { "scx_ops_flags", "SCX_OPS_ALL_FLAGS", 0x3ffLLU },
{ "scx_ops_flags", "__SCX_OPS_INTERNAL_MASK", 0xff00000000000000LLU },
{ "scx_ops_flags", "SCX_OPS_HAS_CPU_PREEMPT", 0x100000000000000LLU },
{ "scx_ops_state", "SCX_OPSS_NONE", 0x0LLU },
--
2.55.0
^ permalink raw reply [flat|nested] 3+ messages in thread* [PATCH 2/2] selftests/sched_ext: Add lazy preemption tests
2026-09-17 7:02 [PATCHSET v6 sched_ext/for-7.4] sched_ext: Add lazy preemption support Andrea Righi
2026-09-17 7:02 ` [PATCH 1/2] " Andrea Righi
@ 2026-09-17 7:02 ` Andrea Righi
1 sibling, 0 replies; 3+ messages in thread
From: Andrea Righi @ 2026-09-17 7:02 UTC (permalink / raw)
To: Tejun Heo, David Vernet, Changwoo Min
Cc: Emil Tsalapatis, Cheng-Yang Chou, sched-ext, linux-kernel
Add trace-based coverage for immediate and lazy preemption. An fexit
probe on __resched_curr() records the requested TIF after the current
task's slice is cleared. SCX_OPS_ENQ_LAST drives the victim through
ops.stopping() without relying on an unrelated wakeup.
Exercise kick requests individually, in both orders and combined in one
call. Verify that immediate preemption and WAIT take precedence over
lazy preemption, and that combining immediate and lazy enqueue
preemption follows the same rule. Also test overriding the expiry
default in both directions through scx_bpf_task_set_slice_expiry(), and
add invalid kick flag coverage.
Extend the NO_HZ_FULL test with infinite-slice victims. Wait until the
scheduler tick is observed to stop, then verify that lazy enqueue and
kick requests restart it and make forward progress. Move the controller
off all NO_HZ_FULL CPUs and increase the phase and watchdog timeouts to
tolerate scheduling delays in oversubscribed test VMs.
Skip trace-based cases when their kernel BTF targets are unavailable,
while allowing the independent invalid-flag cases to run.
Signed-off-by: Andrea Righi <arighi@nvidia.com>
---
tools/testing/selftests/sched_ext/Makefile | 1 +
tools/testing/selftests/sched_ext/kick.bpf.c | 191 +++++
tools/testing/selftests/sched_ext/kick.c | 655 ++++++++++++++++++
.../selftests/sched_ext/nohz_tick.bpf.c | 62 +-
tools/testing/selftests/sched_ext/nohz_tick.c | 280 +++++++-
5 files changed, 1162 insertions(+), 27 deletions(-)
create mode 100644 tools/testing/selftests/sched_ext/kick.bpf.c
create mode 100644 tools/testing/selftests/sched_ext/kick.c
diff --git a/tools/testing/selftests/sched_ext/Makefile b/tools/testing/selftests/sched_ext/Makefile
index 5f5dd9ab903ae..c4ec9b21a4c2a 100644
--- a/tools/testing/selftests/sched_ext/Makefile
+++ b/tools/testing/selftests/sched_ext/Makefile
@@ -172,6 +172,7 @@ auto-test-targets := \
exit \
hotplug \
init_enable_count \
+ kick \
maximal \
maybe_null \
minimal \
diff --git a/tools/testing/selftests/sched_ext/kick.bpf.c b/tools/testing/selftests/sched_ext/kick.bpf.c
new file mode 100644
index 0000000000000..7d4eb0919f099
--- /dev/null
+++ b/tools/testing/selftests/sched_ext/kick.bpf.c
@@ -0,0 +1,191 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES
+ */
+#include <scx/common.bpf.h>
+
+char _license[] SEC("license") = "GPL";
+
+enum kick_scenario {
+ KICK_IMMEDIATE,
+ KICK_LAZY,
+ KICK_LAZY_THEN_IMMEDIATE,
+ KICK_IMMEDIATE_THEN_LAZY,
+ KICK_PLAIN_THEN_LAZY,
+ KICK_LAZY_THEN_PLAIN,
+ KICK_BOTH,
+ KICK_LAZY_WAIT,
+ ENQ_BOTH,
+ TICK_EXPIRY,
+ INVALID_KICK_IDLE,
+ INVALID_KICK_UNKNOWN,
+};
+
+enum kick_state {
+ KICK_STATE_IDLE,
+ KICK_STATE_ARMED,
+ KICK_STATE_QUEUED,
+ KICK_STATE_RESCHED,
+ KICK_STATE_DONE,
+};
+
+const volatile u32 scenario;
+const volatile s32 victim_pid;
+const volatile s32 challenger_pid;
+const volatile s32 target_cpu;
+const volatile s32 slice_expiry_override = -1;
+
+u32 state;
+u64 slice_before;
+u64 slice_at_resched;
+s32 resched_tif;
+u64 nr_wait_callbacks;
+bool enq_preempt_issued;
+
+UEI_DEFINE(uei);
+
+static bool is_trace_scenario(void)
+{
+ return scenario <= TICK_EXPIRY;
+}
+
+void BPF_STRUCT_OPS(kick_enqueue, struct task_struct *p, u64 enq_flags)
+{
+ switch (scenario) {
+ case ENQ_BOTH:
+ if (p->pid == victim_pid) {
+ scx_bpf_dsq_insert(p, SCX_DSQ_GLOBAL, SCX_SLICE_INF,
+ enq_flags);
+ } else if (p->pid == challenger_pid &&
+ state == KICK_STATE_QUEUED) {
+ enq_preempt_issued = true;
+ scx_bpf_dsq_insert(p, SCX_DSQ_LOCAL, SCX_SLICE_DFL,
+ enq_flags | SCX_ENQ_PREEMPT |
+ SCX_ENQ_PREEMPT_LAZY);
+ } else {
+ scx_bpf_dsq_insert(p, SCX_DSQ_GLOBAL, SCX_SLICE_DFL,
+ enq_flags);
+ }
+ goto kick_last;
+ case INVALID_KICK_IDLE:
+ scx_bpf_kick_cpu(scx_bpf_task_cpu(p), SCX_KICK_PREEMPT_LAZY | SCX_KICK_IDLE);
+ break;
+ case INVALID_KICK_UNKNOWN:
+ scx_bpf_kick_cpu(scx_bpf_task_cpu(p), 1LLU << 63);
+ break;
+ }
+
+ scx_bpf_dsq_insert(p, SCX_DSQ_GLOBAL, SCX_SLICE_DFL, enq_flags);
+kick_last:
+ if (enq_flags & SCX_ENQ_LAST)
+ scx_bpf_kick_cpu(target_cpu, SCX_KICK_IDLE);
+}
+
+static void set_slice_expiry_override(struct task_struct *p)
+{
+ if (p->pid == victim_pid && slice_expiry_override >= 0)
+ scx_bpf_task_set_slice_expiry(p, slice_expiry_override);
+}
+
+void BPF_STRUCT_OPS(kick_running, struct task_struct *p)
+{
+ if (!is_trace_scenario() || p->pid != victim_pid)
+ return;
+ set_slice_expiry_override(p);
+ if (__sync_val_compare_and_swap(&state, KICK_STATE_ARMED, KICK_STATE_QUEUED) !=
+ KICK_STATE_ARMED)
+ return;
+
+ slice_before = p->scx.slice;
+
+ switch (scenario) {
+ case KICK_IMMEDIATE:
+ scx_bpf_kick_cpu(target_cpu, SCX_KICK_PREEMPT);
+ break;
+ case KICK_LAZY:
+ scx_bpf_kick_cpu(target_cpu, SCX_KICK_PREEMPT_LAZY);
+ break;
+ case KICK_LAZY_THEN_IMMEDIATE:
+ scx_bpf_kick_cpu(target_cpu, SCX_KICK_PREEMPT_LAZY);
+ scx_bpf_kick_cpu(target_cpu, SCX_KICK_PREEMPT);
+ break;
+ case KICK_IMMEDIATE_THEN_LAZY:
+ scx_bpf_kick_cpu(target_cpu, SCX_KICK_PREEMPT);
+ scx_bpf_kick_cpu(target_cpu, SCX_KICK_PREEMPT_LAZY);
+ break;
+ case KICK_PLAIN_THEN_LAZY:
+ scx_bpf_kick_cpu(target_cpu, 0);
+ scx_bpf_kick_cpu(target_cpu, SCX_KICK_PREEMPT_LAZY);
+ break;
+ case KICK_LAZY_THEN_PLAIN:
+ scx_bpf_kick_cpu(target_cpu, SCX_KICK_PREEMPT_LAZY);
+ scx_bpf_kick_cpu(target_cpu, 0);
+ break;
+ case KICK_BOTH:
+ scx_bpf_kick_cpu(target_cpu, SCX_KICK_PREEMPT |
+ SCX_KICK_PREEMPT_LAZY);
+ break;
+ case KICK_LAZY_WAIT:
+ scx_bpf_kick_cpu(target_cpu, SCX_KICK_PREEMPT_LAZY |
+ SCX_KICK_WAIT);
+ break;
+ case ENQ_BOTH:
+ /* The userspace controller wakes a competing task. */
+ break;
+ case TICK_EXPIRY:
+ /* no kick: the slice runs out at the tick */
+ break;
+ }
+}
+
+SEC("fexit/__resched_curr")
+int BPF_PROG(kick_need_resched, struct rq *rq, int tif)
+{
+ struct task_struct *task = BPF_CORE_READ(rq, curr);
+ u64 slice;
+
+ if (!task || BPF_CORE_READ(task, pid) != victim_pid ||
+ BPF_CORE_READ(rq, cpu) != target_cpu || state != KICK_STATE_QUEUED)
+ return 0;
+ if (scenario == ENQ_BOTH && !enq_preempt_issued)
+ return 0;
+
+ slice = BPF_CORE_READ(task, scx.slice);
+ if (slice)
+ return 0;
+
+ slice_at_resched = slice;
+ resched_tif = tif;
+ state = KICK_STATE_RESCHED;
+ return 0;
+}
+
+SEC("fentry/kick_sync_wait_bal_cb")
+int BPF_PROG(kick_wait_callback, struct rq *rq)
+{
+ if (scenario == KICK_LAZY_WAIT &&
+ BPF_CORE_READ(rq, cpu) == target_cpu)
+ __sync_fetch_and_add(&nr_wait_callbacks, 1);
+ return 0;
+}
+
+void BPF_STRUCT_OPS(kick_stopping, struct task_struct *p, bool runnable)
+{
+ if (p->pid == victim_pid && state == KICK_STATE_RESCHED)
+ state = KICK_STATE_DONE;
+}
+
+void BPF_STRUCT_OPS(kick_exit, struct scx_exit_info *ei)
+{
+ UEI_RECORD(uei, ei);
+}
+
+SEC(".struct_ops.link")
+struct sched_ext_ops kick_ops = {
+ .enqueue = (void *)kick_enqueue,
+ .running = (void *)kick_running,
+ .stopping = (void *)kick_stopping,
+ .exit = (void *)kick_exit,
+ .flags = SCX_OPS_ENQ_LAST,
+ .name = "kick",
+};
diff --git a/tools/testing/selftests/sched_ext/kick.c b/tools/testing/selftests/sched_ext/kick.c
new file mode 100644
index 0000000000000..f64baa948d9d5
--- /dev/null
+++ b/tools/testing/selftests/sched_ext/kick.c
@@ -0,0 +1,655 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES
+ */
+#define _GNU_SOURCE
+#include <bpf/bpf.h>
+#include <errno.h>
+#include <linux/sched.h>
+#include <sched.h>
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/prctl.h>
+#include <sys/wait.h>
+#include <unistd.h>
+#include <scx/common.h>
+
+#ifdef HAVE_GENHDR
+#include "autoconf.h"
+#endif
+
+#include "kick.bpf.skel.h"
+#include "scx_test.h"
+
+#define WAIT_LOOPS 3000
+
+enum kick_scenario {
+ KICK_IMMEDIATE,
+ KICK_LAZY,
+ KICK_LAZY_THEN_IMMEDIATE,
+ KICK_IMMEDIATE_THEN_LAZY,
+ KICK_PLAIN_THEN_LAZY,
+ KICK_LAZY_THEN_PLAIN,
+ KICK_BOTH,
+ KICK_LAZY_WAIT,
+ ENQ_BOTH,
+ TICK_EXPIRY,
+ INVALID_KICK_IDLE,
+ INVALID_KICK_UNKNOWN,
+};
+
+enum kick_state {
+ KICK_STATE_IDLE,
+ KICK_STATE_ARMED,
+ KICK_STATE_QUEUED,
+ KICK_STATE_RESCHED,
+ KICK_STATE_DONE,
+};
+
+struct victim {
+ pid_t pid;
+ int start_fd;
+};
+
+struct observation {
+ u64 slice_before;
+ u64 slice_at_resched;
+ u64 nr_wait_callbacks;
+ s32 resched_tif;
+};
+
+struct kick_ctx {
+ cpu_set_t original_mask;
+ int target_cpu;
+};
+
+static bool enum_supported(const char *type, const char *name)
+{
+ u64 value;
+
+ return __COMPAT_read_enum(type, name, &value);
+}
+
+static enum scx_test_status setup_controller(void **ctx_ptr)
+{
+ struct kick_ctx *ctx;
+ cpu_set_t controller_mask;
+ int cpu, controller_cpu = -1;
+
+ ctx = calloc(1, sizeof(*ctx));
+ SCX_FAIL_IF(!ctx, "Failed to allocate context");
+ if (sched_getaffinity(0, sizeof(ctx->original_mask),
+ &ctx->original_mask)) {
+ free(ctx);
+ SCX_FAIL("Failed to get affinity (%d)", errno);
+ }
+
+ for (cpu = 0; cpu < CPU_SETSIZE; cpu++) {
+ if (!CPU_ISSET(cpu, &ctx->original_mask))
+ continue;
+ if (controller_cpu < 0) {
+ controller_cpu = cpu;
+ } else {
+ ctx->target_cpu = cpu;
+ break;
+ }
+ }
+ if (cpu == CPU_SETSIZE) {
+ printf("SKIP: two allowed CPUs are required\n");
+ free(ctx);
+ return SCX_TEST_SKIP;
+ }
+
+ CPU_ZERO(&controller_mask);
+ CPU_SET(controller_cpu, &controller_mask);
+ if (sched_setaffinity(0, sizeof(controller_mask), &controller_mask)) {
+ free(ctx);
+ SCX_FAIL("Failed to pin controller to CPU %d (%d)",
+ controller_cpu, errno);
+ }
+
+ *ctx_ptr = ctx;
+ return SCX_TEST_PASS;
+}
+
+static void cleanup_controller(void *ctx_ptr)
+{
+ struct kick_ctx *ctx = ctx_ptr;
+
+ sched_setaffinity(0, sizeof(ctx->original_mask), &ctx->original_mask);
+ free(ctx);
+}
+
+static struct victim spawn_victim(int cpu)
+{
+ struct victim victim = { .pid = -1, .start_fd = -1 };
+ int ready[2], start[2];
+ pid_t parent = getpid();
+ char byte = 1;
+
+ if (pipe(ready))
+ return victim;
+ if (pipe(start)) {
+ close(ready[0]);
+ close(ready[1]);
+ return victim;
+ }
+
+ victim.pid = fork();
+ if (!victim.pid) {
+ cpu_set_t mask;
+
+ close(ready[0]);
+ close(start[1]);
+ if (prctl(PR_SET_PDEATHSIG, SIGKILL) || getppid() != parent)
+ _exit(1);
+ CPU_ZERO(&mask);
+ CPU_SET(cpu, &mask);
+ if (sched_setaffinity(0, sizeof(mask), &mask))
+ _exit(1);
+ if (write(ready[1], &byte, 1) != 1)
+ _exit(1);
+ close(ready[1]);
+ if (read(start[0], &byte, 1) != 1)
+ _exit(1);
+ close(start[0]);
+ for (;;)
+ asm volatile("" ::: "memory");
+ }
+ if (victim.pid < 0) {
+ close(ready[0]);
+ close(ready[1]);
+ close(start[0]);
+ close(start[1]);
+ return victim;
+ }
+
+ close(ready[1]);
+ close(start[0]);
+ if (read(ready[0], &byte, 1) != 1) {
+ close(ready[0]);
+ close(start[1]);
+ kill(victim.pid, SIGKILL);
+ waitpid(victim.pid, NULL, 0);
+ victim.pid = -1;
+ return victim;
+ }
+ close(ready[0]);
+ victim.start_fd = start[1];
+ return victim;
+}
+
+static void stop_victim(struct victim *victim)
+{
+ if (victim->start_fd >= 0)
+ close(victim->start_fd);
+ if (victim->pid > 0) {
+ kill(victim->pid, SIGKILL);
+ waitpid(victim->pid, NULL, 0);
+ }
+}
+
+static bool start_victim(struct victim *victim)
+{
+ char byte = 1;
+
+ if (write(victim->start_fd, &byte, 1) != 1)
+ return false;
+ close(victim->start_fd);
+ victim->start_fd = -1;
+ return true;
+}
+
+static bool wait_for_state(struct kick *skel, u32 wanted)
+{
+ int i;
+
+ for (i = 0; i < WAIT_LOOPS; i++) {
+ if (__atomic_load_n(&skel->bss->state, __ATOMIC_ACQUIRE) == wanted)
+ return true;
+ if (skel->data->uei.kind != EXIT_KIND(SCX_EXIT_NONE))
+ return false;
+ usleep(1000);
+ }
+ return false;
+}
+
+static bool wait_for_counter(struct kick *skel, const u64 *counter, u64 wanted)
+{
+ int i;
+
+ for (i = 0; i < WAIT_LOOPS; i++) {
+ if (__atomic_load_n(counter, __ATOMIC_ACQUIRE) >= wanted)
+ return true;
+ if (skel->data->uei.kind != EXIT_KIND(SCX_EXIT_NONE))
+ return false;
+ usleep(1000);
+ }
+ return false;
+}
+
+static enum scx_test_status trace_one(struct kick_ctx *ctx, u32 scenario,
+ u64 ops_flags,
+ s32 slice_expiry_override,
+ struct observation *obs)
+{
+ struct bpf_link *ops_link = NULL;
+ struct kick *skel = NULL;
+ struct victim challenger = { .pid = -1, .start_fd = -1 };
+ struct victim victim;
+ enum scx_test_status ret = SCX_TEST_FAIL;
+ int cpu = ctx->target_cpu;
+
+ victim = spawn_victim(cpu);
+ if (victim.pid < 0) {
+ SCX_ERR("Failed to spawn victim");
+ return SCX_TEST_FAIL;
+ }
+ if (scenario == ENQ_BOTH) {
+ challenger = spawn_victim(cpu);
+ if (challenger.pid < 0) {
+ SCX_ERR("Failed to spawn enqueue challenger");
+ goto out;
+ }
+ }
+
+ skel = kick__open();
+ if (!skel) {
+ SCX_ERR("Failed to open scenario %u", scenario);
+ goto out;
+ }
+ SCX_ENUM_INIT(skel);
+ skel->rodata->scenario = scenario;
+ skel->rodata->victim_pid = victim.pid;
+ skel->rodata->challenger_pid = challenger.pid;
+ skel->rodata->target_cpu = cpu;
+ skel->rodata->slice_expiry_override = slice_expiry_override;
+ skel->struct_ops.kick_ops->flags |= ops_flags;
+ if (scenario != KICK_LAZY_WAIT)
+ bpf_program__set_autoload(skel->progs.kick_wait_callback,
+ false);
+ if (kick__load(skel)) {
+ SCX_ERR("Failed to load scenario %u", scenario);
+ goto out;
+ }
+
+ bpf_map__set_autoattach(skel->maps.kick_ops, false);
+ if (kick__attach(skel)) {
+ SCX_ERR("Failed to attach __resched_curr tracer");
+ goto out;
+ }
+ skel->bss->state = KICK_STATE_ARMED;
+ ops_link = bpf_map__attach_struct_ops(skel->maps.kick_ops);
+ if (!ops_link) {
+ SCX_ERR("Failed to attach scenario %u", scenario);
+ goto out;
+ }
+ if (!start_victim(&victim)) {
+ SCX_ERR("Failed to start victim");
+ goto out;
+ }
+ if (scenario == ENQ_BOTH) {
+ if (!wait_for_state(skel, KICK_STATE_QUEUED)) {
+ SCX_ERR("Enqueue scenario did not arm");
+ goto out;
+ }
+ if (!start_victim(&challenger)) {
+ SCX_ERR("Failed to start enqueue challenger");
+ goto out;
+ }
+ }
+ if (!wait_for_state(skel, KICK_STATE_DONE)) {
+ SCX_ERR("Scenario %u stopped in state %u, exit kind %d", scenario, skel->bss->state,
+ skel->data->uei.kind);
+ goto out;
+ }
+ if (scenario == KICK_LAZY_WAIT &&
+ !wait_for_counter(skel, &skel->bss->nr_wait_callbacks, 1)) {
+ SCX_ERR("WAIT callback did not run");
+ goto out;
+ }
+
+ obs->slice_before = skel->bss->slice_before;
+ obs->slice_at_resched = skel->bss->slice_at_resched;
+ obs->nr_wait_callbacks = skel->bss->nr_wait_callbacks;
+ obs->resched_tif = skel->bss->resched_tif;
+ ret = SCX_TEST_PASS;
+out:
+ stop_victim(&challenger);
+ stop_victim(&victim);
+ if (ops_link)
+ bpf_link__destroy(ops_link);
+ if (skel)
+ kick__destroy(skel);
+ return ret;
+}
+
+static bool observation_valid(const struct observation *obs)
+{
+ return obs->slice_before > 0 && !obs->slice_at_resched;
+}
+
+static int active_lazy_mode(void)
+{
+ char buf[128];
+ FILE *file;
+
+ file = fopen("/sys/kernel/debug/sched/preempt", "r");
+ if (!file)
+ return -1;
+ if (!fgets(buf, sizeof(buf), file)) {
+ fclose(file);
+ return -1;
+ }
+ fclose(file);
+
+ if (strstr(buf, "(lazy)"))
+ return 1;
+ if (strstr(buf, "(none)") || strstr(buf, "(voluntary)") ||
+ strstr(buf, "(full)"))
+ return 0;
+ return -1;
+}
+
+static bool trace_target_supported(const char *name,
+ enum bpf_attach_type attach_type)
+{
+ return libbpf_find_vmlinux_btf_id(name, attach_type) > 0;
+}
+
+static enum scx_test_status setup_immediate(void **ctx)
+{
+ if (!enum_supported("scx_kick_flags", "SCX_KICK_PREEMPT")) {
+ printf("SKIP: SCX_KICK_PREEMPT is not supported\n");
+ return SCX_TEST_SKIP;
+ }
+ if (!trace_target_supported("__resched_curr", BPF_TRACE_FEXIT)) {
+ printf("SKIP: __resched_curr is not available in BTF\n");
+ return SCX_TEST_SKIP;
+ }
+ return setup_controller(ctx);
+}
+
+static enum scx_test_status setup_lazy(void **ctx)
+{
+ if (!enum_supported("scx_kick_flags", "SCX_KICK_PREEMPT_LAZY")) {
+ printf("SKIP: SCX_KICK_PREEMPT_LAZY is not supported\n");
+ return SCX_TEST_SKIP;
+ }
+ return setup_immediate(ctx);
+}
+
+static enum scx_test_status setup_tick(void **ctx)
+{
+ if (!enum_supported("scx_ops_flags", "SCX_OPS_LAZY_SLICE_EXPIRY")) {
+ printf("SKIP: SCX_OPS_LAZY_SLICE_EXPIRY is not supported\n");
+ return SCX_TEST_SKIP;
+ }
+ return setup_lazy(ctx);
+}
+
+static enum scx_test_status setup_coalesce(void **ctx)
+{
+ if (!enum_supported("scx_enq_flags", "SCX_ENQ_PREEMPT_LAZY")) {
+ printf("SKIP: SCX_ENQ_PREEMPT_LAZY is not supported\n");
+ return SCX_TEST_SKIP;
+ }
+ if (!trace_target_supported("kick_sync_wait_bal_cb", BPF_TRACE_FENTRY)) {
+ printf("SKIP: kick_sync_wait_bal_cb is not available in BTF\n");
+ return SCX_TEST_SKIP;
+ }
+ return setup_lazy(ctx);
+}
+
+static enum scx_test_status setup_invalid(void **ctx)
+{
+ if (!enum_supported("scx_kick_flags", "SCX_KICK_PREEMPT_LAZY")) {
+ printf("SKIP: SCX_KICK_PREEMPT_LAZY is not supported\n");
+ return SCX_TEST_SKIP;
+ }
+ return setup_controller(ctx);
+}
+
+static enum scx_test_status run_immediate(void *ctx)
+{
+ struct observation obs;
+ enum scx_test_status status;
+
+ status = trace_one(ctx, KICK_IMMEDIATE, 0, -1, &obs);
+ SCX_EQ(status, SCX_TEST_PASS);
+ SCX_ASSERT(observation_valid(&obs));
+ return SCX_TEST_PASS;
+}
+
+static enum scx_test_status run_lazy(void *ctx)
+{
+ struct observation immediate, lazy;
+ enum scx_test_status status;
+ int lazy_mode;
+
+ status = trace_one(ctx, KICK_IMMEDIATE, 0, -1, &immediate);
+ SCX_EQ(status, SCX_TEST_PASS);
+ SCX_ASSERT(observation_valid(&immediate));
+ status = trace_one(ctx, KICK_LAZY, 0, -1, &lazy);
+ SCX_EQ(status, SCX_TEST_PASS);
+ SCX_ASSERT(observation_valid(&lazy));
+
+ lazy_mode = active_lazy_mode();
+ if (lazy_mode > 0)
+ SCX_FAIL_IF(lazy.resched_tif == immediate.resched_tif,
+ "Lazy mode used immediate TIF %d", lazy.resched_tif);
+ else if (!lazy_mode)
+ SCX_EQ(lazy.resched_tif, immediate.resched_tif);
+ else
+ printf("INFO: preemption mode unavailable; lazy TIF was %d, immediate TIF was %d\n",
+ lazy.resched_tif, immediate.resched_tif);
+
+ return SCX_TEST_PASS;
+}
+
+static enum scx_test_status run_coalesce(void *ctx)
+{
+ struct observation immediate, lazy_first, immediate_first;
+ struct observation plain_first, lazy_then_plain;
+ struct observation both, lazy_wait, enq_both;
+ enum scx_test_status status;
+
+ status = trace_one(ctx, KICK_IMMEDIATE, 0, -1, &immediate);
+ SCX_EQ(status, SCX_TEST_PASS);
+ SCX_ASSERT(observation_valid(&immediate));
+ status = trace_one(ctx, KICK_LAZY_THEN_IMMEDIATE, 0, -1,
+ &lazy_first);
+ SCX_EQ(status, SCX_TEST_PASS);
+ SCX_ASSERT(observation_valid(&lazy_first));
+ status = trace_one(ctx, KICK_IMMEDIATE_THEN_LAZY, 0, -1,
+ &immediate_first);
+ SCX_EQ(status, SCX_TEST_PASS);
+ SCX_ASSERT(observation_valid(&immediate_first));
+ SCX_EQ(lazy_first.resched_tif, immediate.resched_tif);
+ SCX_EQ(immediate_first.resched_tif, immediate.resched_tif);
+
+ /*
+ * A plain kick in the same batch doesn't clear the slice by itself.
+ * The lazy preemption must still expire it, and the plain kick must
+ * still reschedule immediately, whichever came first.
+ */
+ status = trace_one(ctx, KICK_PLAIN_THEN_LAZY, 0, -1,
+ &plain_first);
+ SCX_EQ(status, SCX_TEST_PASS);
+ SCX_ASSERT(observation_valid(&plain_first));
+ status = trace_one(ctx, KICK_LAZY_THEN_PLAIN, 0, -1,
+ &lazy_then_plain);
+ SCX_EQ(status, SCX_TEST_PASS);
+ SCX_ASSERT(observation_valid(&lazy_then_plain));
+ SCX_EQ(plain_first.resched_tif, immediate.resched_tif);
+ SCX_EQ(lazy_then_plain.resched_tif, immediate.resched_tif);
+
+ /* Immediate kick and WAIT both take precedence over lazy preemption. */
+ status = trace_one(ctx, KICK_BOTH, 0, -1, &both);
+ SCX_EQ(status, SCX_TEST_PASS);
+ SCX_ASSERT(observation_valid(&both));
+ status = trace_one(ctx, KICK_LAZY_WAIT, 0, -1, &lazy_wait);
+ SCX_EQ(status, SCX_TEST_PASS);
+ SCX_ASSERT(observation_valid(&lazy_wait));
+ SCX_EQ(both.resched_tif, immediate.resched_tif);
+ SCX_EQ(lazy_wait.resched_tif, immediate.resched_tif);
+ SCX_GT(lazy_wait.nr_wait_callbacks, 0);
+
+ /* Immediate enqueue preemption likewise takes precedence over lazy. */
+ status = trace_one(ctx, ENQ_BOTH, 0, -1, &enq_both);
+ SCX_EQ(status, SCX_TEST_PASS);
+ SCX_ASSERT(observation_valid(&enq_both));
+ SCX_EQ(enq_both.resched_tif, immediate.resched_tif);
+ return SCX_TEST_PASS;
+}
+
+/*
+ * A slice running out at the tick reschedules immediately by default and
+ * lazily with SCX_OPS_LAZY_SLICE_EXPIRY, the way fair.c expires a slice.
+ */
+static enum scx_test_status run_tick(void *ctx)
+{
+ struct observation immediate, lazy, force_lazy, force_immediate;
+ enum scx_test_status status;
+ u64 lazy_flag;
+ int lazy_mode;
+ bool found;
+
+ found = __COMPAT_read_enum("scx_ops_flags", "SCX_OPS_LAZY_SLICE_EXPIRY",
+ &lazy_flag);
+ SCX_ASSERT(found);
+ status = trace_one(ctx, TICK_EXPIRY, 0, -1, &immediate);
+ SCX_EQ(status, SCX_TEST_PASS);
+ SCX_ASSERT(observation_valid(&immediate));
+ status = trace_one(ctx, TICK_EXPIRY, lazy_flag, -1, &lazy);
+ SCX_EQ(status, SCX_TEST_PASS);
+ SCX_ASSERT(observation_valid(&lazy));
+ status = trace_one(ctx, TICK_EXPIRY, 0, 1, &force_lazy);
+ SCX_EQ(status, SCX_TEST_PASS);
+ SCX_ASSERT(observation_valid(&force_lazy));
+ status = trace_one(ctx, TICK_EXPIRY, lazy_flag, 0,
+ &force_immediate);
+ SCX_EQ(status, SCX_TEST_PASS);
+ SCX_ASSERT(observation_valid(&force_immediate));
+
+ lazy_mode = active_lazy_mode();
+ if (lazy_mode > 0)
+ SCX_FAIL_IF(lazy.resched_tif == immediate.resched_tif,
+ "Lazy slice expiry used immediate TIF %d", lazy.resched_tif);
+ else if (!lazy_mode)
+ SCX_EQ(lazy.resched_tif, immediate.resched_tif);
+ else
+ printf("INFO: preemption mode unavailable; lazy TIF was %d, immediate TIF was %d\n",
+ lazy.resched_tif, immediate.resched_tif);
+ SCX_EQ(force_lazy.resched_tif, lazy.resched_tif);
+ SCX_EQ(force_immediate.resched_tif, immediate.resched_tif);
+
+ return SCX_TEST_PASS;
+}
+
+static enum scx_test_status invalid_one(struct kick_ctx *ctx, u32 scenario)
+{
+ struct bpf_link *ops_link = NULL;
+ struct kick *skel = NULL;
+ struct victim victim;
+ enum scx_test_status ret = SCX_TEST_FAIL;
+ int cpu = ctx->target_cpu;
+ int i;
+
+ victim = spawn_victim(cpu);
+ if (victim.pid < 0)
+ return SCX_TEST_FAIL;
+
+ skel = kick__open();
+ if (!skel)
+ goto out;
+ SCX_ENUM_INIT(skel);
+ skel->rodata->scenario = scenario;
+ bpf_program__set_autoload(skel->progs.kick_need_resched, false);
+ bpf_program__set_autoload(skel->progs.kick_wait_callback, false);
+ if (kick__load(skel))
+ goto out;
+ ops_link = bpf_map__attach_struct_ops(skel->maps.kick_ops);
+ if (!ops_link || !start_victim(&victim))
+ goto out;
+
+ for (i = 0; i < WAIT_LOOPS; i++) {
+ if (skel->data->uei.kind == EXIT_KIND(SCX_EXIT_ERROR)) {
+ ret = SCX_TEST_PASS;
+ break;
+ }
+ usleep(1000);
+ }
+out:
+ stop_victim(&victim);
+ if (ops_link)
+ bpf_link__destroy(ops_link);
+ if (skel)
+ kick__destroy(skel);
+ return ret;
+}
+
+static enum scx_test_status run_invalid(void *ctx)
+{
+ enum scx_test_status status;
+ u32 scenario;
+
+ for (scenario = INVALID_KICK_IDLE;
+ scenario <= INVALID_KICK_UNKNOWN; scenario++) {
+ status = invalid_one(ctx, scenario);
+ SCX_EQ(status, SCX_TEST_PASS);
+ }
+ return SCX_TEST_PASS;
+}
+
+static struct scx_test kick_immediate = {
+ .name = "kick_immediate",
+ .description = "Trace immediate kick slice expiration and rescheduling",
+ .setup = setup_immediate,
+ .run = run_immediate,
+ .cleanup = cleanup_controller,
+};
+
+static struct scx_test kick_lazy = {
+ .name = "kick_lazy",
+ .description = "Trace lazy kick slice expiration and rescheduling",
+ .setup = setup_lazy,
+ .run = run_lazy,
+ .cleanup = cleanup_controller,
+};
+
+static struct scx_test kick_coalesce = {
+ .name = "kick_coalesce",
+ .description = "Verify lazy preemption coalesces with immediate requests",
+ .setup = setup_coalesce,
+ .run = run_coalesce,
+ .cleanup = cleanup_controller,
+};
+
+static struct scx_test kick_tick = {
+ .name = "kick_tick",
+ .description = "Trace slice expiry at the tick, immediate and lazy",
+ .setup = setup_tick,
+ .run = run_tick,
+ .cleanup = cleanup_controller,
+};
+
+static struct scx_test kick_invalid = {
+ .name = "kick_invalid",
+ .description = "Verify invalid kick flag combinations fail",
+ .setup = setup_invalid,
+ .run = run_invalid,
+ .cleanup = cleanup_controller,
+};
+
+__attribute__((constructor))
+static void register_kick_tests(void)
+{
+ scx_test_register(&kick_immediate);
+ scx_test_register(&kick_lazy);
+ scx_test_register(&kick_coalesce);
+ scx_test_register(&kick_tick);
+ scx_test_register(&kick_invalid);
+}
diff --git a/tools/testing/selftests/sched_ext/nohz_tick.bpf.c b/tools/testing/selftests/sched_ext/nohz_tick.bpf.c
index 6998c5dd6bcb9..3f45233acc464 100644
--- a/tools/testing/selftests/sched_ext/nohz_tick.bpf.c
+++ b/tools/testing/selftests/sched_ext/nohz_tick.bpf.c
@@ -9,10 +9,24 @@
char _license[] SEC("license") = "GPL";
const volatile s32 test_cpu;
-bool finite_phase;
+enum nohz_phase {
+ NOHZ_PHASE_INF,
+ NOHZ_PHASE_FINITE,
+ NOHZ_PHASE_LAZY_ENQ,
+ NOHZ_PHASE_LAZY_KICK,
+};
+
+u32 phase;
+s32 victim_pid;
+s32 challenger_pid;
+s32 trigger_pid;
u64 nr_inf_running;
u64 nr_finite_running;
u64 nr_finite_ticks;
+u64 nr_lazy_victim_running;
+u64 nr_lazy_enq_running;
+u64 nr_lazy_kick_running;
+u64 nr_lazy_ticks;
UEI_DEFINE(uei);
@@ -24,9 +38,31 @@ s32 BPF_STRUCT_OPS(nohz_tick_select_cpu, struct task_struct *p, s32 prev_cpu,
void BPF_STRUCT_OPS(nohz_tick_enqueue, struct task_struct *p, u64 enq_flags)
{
- u64 slice = finite_phase ? 1000000ULL : SCX_SLICE_INF;
+ u64 slice;
+ u64 dsq_id = SCX_DSQ_GLOBAL;
+
+ switch (phase) {
+ case NOHZ_PHASE_INF:
+ slice = SCX_SLICE_INF;
+ break;
+ case NOHZ_PHASE_FINITE:
+ slice = 1000000ULL;
+ break;
+ case NOHZ_PHASE_LAZY_ENQ:
+ case NOHZ_PHASE_LAZY_KICK:
+ dsq_id = SCX_DSQ_LOCAL;
+ slice = p->pid == victim_pid ? SCX_SLICE_INF : SCX_SLICE_DFL;
+ if (phase == NOHZ_PHASE_LAZY_ENQ && p->pid == challenger_pid)
+ enq_flags |= SCX_ENQ_PREEMPT_LAZY;
+ break;
+ default:
+ slice = SCX_SLICE_DFL;
+ break;
+ }
- scx_bpf_dsq_insert(p, SCX_DSQ_GLOBAL, slice, enq_flags);
+ scx_bpf_dsq_insert(p, dsq_id, slice, enq_flags);
+ if (phase == NOHZ_PHASE_LAZY_KICK && p->pid == trigger_pid)
+ scx_bpf_kick_cpu(test_cpu, SCX_KICK_PREEMPT_LAZY);
if (enq_flags & SCX_ENQ_LAST)
scx_bpf_kick_cpu(test_cpu, SCX_KICK_IDLE);
}
@@ -36,16 +72,28 @@ void BPF_STRUCT_OPS(nohz_tick_running, struct task_struct *p)
if (bpf_get_smp_processor_id() != test_cpu)
return;
- if (finite_phase)
+ if (phase == NOHZ_PHASE_FINITE)
__sync_fetch_and_add(&nr_finite_running, 1);
- else
+ else if (phase == NOHZ_PHASE_INF)
__sync_fetch_and_add(&nr_inf_running, 1);
+ else if (p->pid == victim_pid)
+ __sync_fetch_and_add(&nr_lazy_victim_running, 1);
+ else if (phase == NOHZ_PHASE_LAZY_ENQ && p->pid == challenger_pid)
+ __sync_fetch_and_add(&nr_lazy_enq_running, 1);
+ else if (phase == NOHZ_PHASE_LAZY_KICK && p->pid == challenger_pid)
+ __sync_fetch_and_add(&nr_lazy_kick_running, 1);
}
void BPF_STRUCT_OPS(nohz_tick_tick, struct task_struct *p)
{
- if (bpf_get_smp_processor_id() == test_cpu && finite_phase)
+ if (bpf_get_smp_processor_id() != test_cpu)
+ return;
+
+ if (phase == NOHZ_PHASE_FINITE)
__sync_fetch_and_add(&nr_finite_ticks, 1);
+ else if ((phase == NOHZ_PHASE_LAZY_ENQ ||
+ phase == NOHZ_PHASE_LAZY_KICK) && p->pid == victim_pid)
+ __sync_fetch_and_add(&nr_lazy_ticks, 1);
}
void BPF_STRUCT_OPS(nohz_tick_exit, struct scx_exit_info *ei)
@@ -61,5 +109,5 @@ struct sched_ext_ops nohz_tick_ops = {
.tick = (void *)nohz_tick_tick,
.exit = (void *)nohz_tick_exit,
.name = "nohz_tick",
- .timeout_ms = 1000U,
+ .timeout_ms = 5000U,
};
diff --git a/tools/testing/selftests/sched_ext/nohz_tick.c b/tools/testing/selftests/sched_ext/nohz_tick.c
index 028f54391c2ca..7b3042b4794d3 100644
--- a/tools/testing/selftests/sched_ext/nohz_tick.c
+++ b/tools/testing/selftests/sched_ext/nohz_tick.c
@@ -28,12 +28,27 @@
#endif
#define MIN_FINITE_TICKS 3
-#define PHASE_TIMEOUT_MS 1000
+#define PHASE_TIMEOUT_MS 5000
+#define TICK_STOP_STABLE_MS 100
struct nohz_tick_ctx {
struct nohz_tick *skel;
cpu_set_t original_mask;
int test_cpu;
+ int housekeeping_cpu;
+ bool lazy_supported;
+};
+
+enum nohz_phase {
+ NOHZ_PHASE_INF,
+ NOHZ_PHASE_FINITE,
+ NOHZ_PHASE_LAZY_ENQ,
+ NOHZ_PHASE_LAZY_KICK,
+};
+
+struct gated_worker {
+ pid_t pid;
+ int start_fd;
};
static int first_allowed_cpu(const cpu_set_t *mask, int first, int last)
@@ -47,20 +62,21 @@ static int first_allowed_cpu(const cpu_set_t *mask, int first, int last)
return -1;
}
-static int find_nohz_full_cpu(const cpu_set_t *allowed)
+static int read_nohz_full_mask(cpu_set_t *mask)
{
char buf[4096], *cur, *end;
FILE *file;
+ int ret = 0;
file = fopen("/sys/devices/system/cpu/nohz_full", "r");
if (!file)
- return -1;
+ return -errno;
if (!fgets(buf, sizeof(buf), file)) {
- fclose(file);
- return -1;
+ ret = ferror(file) ? -errno : -EINVAL;
+ goto out;
}
- fclose(file);
+ CPU_ZERO(mask);
cur = buf;
while (*cur) {
long first, last;
@@ -73,25 +89,30 @@ static int find_nohz_full_cpu(const cpu_set_t *allowed)
errno = 0;
first = strtol(cur, &end, 10);
- if (errno || end == cur || first < 0 || first >= CPU_SETSIZE)
- return -1;
+ if (errno || end == cur || first < 0 || first >= CPU_SETSIZE) {
+ ret = -EINVAL;
+ goto out;
+ }
cur = end;
last = first;
if (*cur == '-') {
cur++;
errno = 0;
last = strtol(cur, &end, 10);
- if (errno || end == cur || last < first)
- return -1;
+ if (errno || end == cur || last < first) {
+ ret = -EINVAL;
+ goto out;
+ }
cur = end;
}
- cpu = first_allowed_cpu(allowed, first, last);
- if (cpu >= 0)
- return cpu;
+ for (cpu = first; cpu <= last && cpu < CPU_SETSIZE; cpu++)
+ CPU_SET(cpu, mask);
}
- return -1;
+out:
+ fclose(file);
+ return ret;
}
static pid_t start_worker(int cpu)
@@ -132,6 +153,87 @@ static void stop_worker(pid_t pid)
waitpid(pid, NULL, 0);
}
+static struct gated_worker spawn_gated_worker(int cpu)
+{
+ struct gated_worker worker = { .pid = -1, .start_fd = -1 };
+ int ready[2], start[2];
+ pid_t parent = getpid();
+ char byte = 1;
+
+ if (pipe(ready))
+ return worker;
+ if (pipe(start)) {
+ close(ready[0]);
+ close(ready[1]);
+ return worker;
+ }
+
+ worker.pid = fork();
+ if (!worker.pid) {
+ struct sched_param param = {};
+ cpu_set_t mask;
+
+ close(ready[0]);
+ close(start[1]);
+ if (prctl(PR_SET_PDEATHSIG, SIGKILL) || getppid() != parent)
+ _exit(1);
+ CPU_ZERO(&mask);
+ CPU_SET(cpu, &mask);
+ if (sched_setaffinity(0, sizeof(mask), &mask))
+ _exit(1);
+ if (sched_setscheduler(0, SCHED_EXT, ¶m))
+ _exit(1);
+ if (write(ready[1], &byte, 1) != 1)
+ _exit(1);
+ close(ready[1]);
+ if (read(start[0], &byte, 1) != 1)
+ _exit(1);
+ close(start[0]);
+ for (;;)
+ asm volatile("" ::: "memory");
+ }
+ if (worker.pid < 0) {
+ close(ready[0]);
+ close(ready[1]);
+ close(start[0]);
+ close(start[1]);
+ return worker;
+ }
+
+ close(ready[1]);
+ close(start[0]);
+ if (read(ready[0], &byte, 1) != 1) {
+ close(ready[0]);
+ close(start[1]);
+ kill(worker.pid, SIGKILL);
+ waitpid(worker.pid, NULL, 0);
+ worker.pid = -1;
+ return worker;
+ }
+ close(ready[0]);
+ worker.start_fd = start[1];
+ return worker;
+}
+
+static bool start_gated_worker(struct gated_worker *worker)
+{
+ char byte = 1;
+
+ if (write(worker->start_fd, &byte, 1) != 1)
+ return false;
+ close(worker->start_fd);
+ worker->start_fd = -1;
+ return true;
+}
+
+static void stop_gated_worker(struct gated_worker *worker)
+{
+ if (worker->start_fd >= 0)
+ close(worker->start_fd);
+ stop_worker(worker->pid);
+ worker->pid = -1;
+}
+
static int pause_worker(pid_t pid)
{
int status;
@@ -159,11 +261,34 @@ static bool wait_for_counter(const u64 *counter, u64 value, int timeout_ms)
return false;
}
+static bool wait_for_tick_stop(const u64 *counter, int timeout_ms)
+{
+ u64 prev = __atomic_load_n(counter, __ATOMIC_RELAXED);
+ int elapsed, stable = 0;
+
+ for (elapsed = 0; elapsed < timeout_ms; elapsed++) {
+ u64 curr;
+
+ usleep(1000);
+ curr = __atomic_load_n(counter, __ATOMIC_RELAXED);
+ if (curr == prev) {
+ if (++stable >= TICK_STOP_STABLE_MS)
+ return true;
+ } else {
+ prev = curr;
+ stable = 0;
+ }
+ }
+
+ return false;
+}
+
static enum scx_test_status setup(void **ctx_ptr)
{
struct nohz_tick_ctx *ctx;
- cpu_set_t controller_mask;
- int cpu;
+ cpu_set_t controller_mask, nohz_full_mask;
+ u64 enum_value;
+ int cpu, i, ret;
ctx = calloc(1, sizeof(*ctx));
SCX_FAIL_IF(!ctx, "Failed to allocate context");
@@ -173,15 +298,27 @@ static enum scx_test_status setup(void **ctx_ptr)
SCX_FAIL("Failed to get affinity (%d)", errno);
}
- cpu = find_nohz_full_cpu(&ctx->original_mask);
- if (cpu < 0) {
+ ret = read_nohz_full_mask(&nohz_full_mask);
+ if (ret) {
+ fprintf(stderr, "SKIP: failed to read NOHZ_FULL mask (%d)\n", ret);
+ free(ctx);
+ return SCX_TEST_SKIP;
+ }
+
+ for (cpu = 0; cpu < CPU_SETSIZE; cpu++)
+ if (CPU_ISSET(cpu, &ctx->original_mask) &&
+ CPU_ISSET(cpu, &nohz_full_mask))
+ break;
+ if (cpu == CPU_SETSIZE) {
fprintf(stderr, "SKIP: no allowed NOHZ_FULL CPU\n");
free(ctx);
return SCX_TEST_SKIP;
}
controller_mask = ctx->original_mask;
- CPU_CLR(cpu, &controller_mask);
+ for (i = 0; i < CPU_SETSIZE; i++)
+ if (CPU_ISSET(i, &nohz_full_mask))
+ CPU_CLR(i, &controller_mask);
if (CPU_COUNT(&controller_mask) == 0) {
fprintf(stderr, "SKIP: no housekeeping CPU available\n");
free(ctx);
@@ -189,6 +326,13 @@ static enum scx_test_status setup(void **ctx_ptr)
}
ctx->test_cpu = cpu;
+ ctx->housekeeping_cpu = first_allowed_cpu(&controller_mask, 0,
+ CPU_SETSIZE - 1);
+ ctx->lazy_supported =
+ __COMPAT_read_enum("scx_enq_flags", "SCX_ENQ_PREEMPT_LAZY",
+ &enum_value) &&
+ __COMPAT_read_enum("scx_kick_flags", "SCX_KICK_PREEMPT_LAZY",
+ &enum_value);
ctx->skel = nohz_tick__open();
if (!ctx->skel) {
free(ctx);
@@ -220,11 +364,15 @@ static enum scx_test_status run(void *ctx_ptr)
struct nohz_tick_ctx *ctx = ctx_ptr;
struct nohz_tick *skel = ctx->skel;
struct bpf_link *link = NULL;
+ struct gated_worker victim = { .pid = -1, .start_fd = -1 };
+ struct gated_worker challenger = { .pid = -1, .start_fd = -1 };
+ struct gated_worker trigger = { .pid = -1, .start_fd = -1 };
enum scx_test_status status = SCX_TEST_FAIL;
pid_t finite_worker = -1;
pid_t inf_worker = -1;
u64 finite_running;
u64 finite_ticks;
+ u64 victim_running;
int ret;
link = bpf_map__attach_struct_ops(skel->maps.nohz_tick_ops);
@@ -260,7 +408,8 @@ static enum scx_test_status run(void *ctx_ptr)
/*
* The next EXT task receives a finite slice and must restart the tick.
*/
- __atomic_store_n(&skel->bss->finite_phase, true, __ATOMIC_RELEASE);
+ __atomic_store_n(&skel->bss->phase, NOHZ_PHASE_FINITE,
+ __ATOMIC_RELEASE);
finite_worker = start_worker(ctx->test_cpu);
if (finite_worker < 0) {
SCX_ERR("Failed to start finite-slice worker (%d)", errno);
@@ -308,7 +457,95 @@ static enum scx_test_status run(void *ctx_ptr)
finite_ticks));
goto out;
}
+ stop_worker(finite_worker);
+ finite_worker = -1;
+ stop_worker(inf_worker);
+ inf_worker = -1;
+ if (!ctx->lazy_supported)
+ goto check_exit;
+
+ /*
+ * A lazy local enqueue must restart the tick after clearing the slice of
+ * an infinite-slice task on a full-dynticks CPU.
+ */
+ __atomic_store_n(&skel->bss->phase, NOHZ_PHASE_LAZY_ENQ,
+ __ATOMIC_RELEASE);
+ victim = spawn_gated_worker(ctx->test_cpu);
+ challenger = spawn_gated_worker(ctx->test_cpu);
+ if (victim.pid < 0 || challenger.pid < 0) {
+ SCX_ERR("Failed to spawn lazy-enqueue workers");
+ goto out;
+ }
+ skel->bss->victim_pid = victim.pid;
+ skel->bss->challenger_pid = challenger.pid;
+ if (!start_gated_worker(&victim) ||
+ !wait_for_counter(&skel->bss->nr_lazy_victim_running, 1,
+ PHASE_TIMEOUT_MS)) {
+ SCX_ERR("Lazy-enqueue victim was not scheduled");
+ goto out;
+ }
+ if (!wait_for_tick_stop(&skel->bss->nr_lazy_ticks,
+ PHASE_TIMEOUT_MS)) {
+ SCX_ERR("Tick did not stop before lazy enqueue");
+ goto out;
+ }
+
+ if (!start_gated_worker(&challenger) ||
+ !wait_for_counter(&skel->bss->nr_lazy_enq_running, 1,
+ PHASE_TIMEOUT_MS)) {
+ SCX_ERR("Lazy enqueue made no progress on CPU %d", ctx->test_cpu);
+ goto out;
+ }
+ stop_gated_worker(&challenger);
+ stop_gated_worker(&victim);
+
+ /* Repeat with a lazy kick delivered from a housekeeping CPU. */
+ __atomic_store_n(&skel->bss->phase, NOHZ_PHASE_LAZY_KICK,
+ __ATOMIC_RELEASE);
+ victim = spawn_gated_worker(ctx->test_cpu);
+ challenger = spawn_gated_worker(ctx->test_cpu);
+ trigger = spawn_gated_worker(ctx->housekeeping_cpu);
+ if (victim.pid < 0 || challenger.pid < 0 || trigger.pid < 0) {
+ SCX_ERR("Failed to spawn lazy-kick workers");
+ goto out;
+ }
+ skel->bss->victim_pid = victim.pid;
+ skel->bss->challenger_pid = challenger.pid;
+ skel->bss->trigger_pid = trigger.pid;
+ victim_running = __atomic_load_n(&skel->bss->nr_lazy_victim_running,
+ __ATOMIC_RELAXED);
+ if (!start_gated_worker(&victim) ||
+ !wait_for_counter(&skel->bss->nr_lazy_victim_running,
+ victim_running + 1,
+ PHASE_TIMEOUT_MS)) {
+ SCX_ERR("Lazy-kick victim was not scheduled");
+ goto out;
+ }
+ if (!start_gated_worker(&challenger)) {
+ SCX_ERR("Failed to start lazy-kick challenger");
+ goto out;
+ }
+ if (!wait_for_tick_stop(&skel->bss->nr_lazy_ticks,
+ PHASE_TIMEOUT_MS)) {
+ SCX_ERR("Tick did not stop before lazy kick");
+ goto out;
+ }
+ if (__atomic_load_n(&skel->bss->nr_lazy_kick_running,
+ __ATOMIC_RELAXED)) {
+ SCX_ERR("Lazy-kick challenger ran before the kick");
+ goto out;
+ }
+ if (!start_gated_worker(&trigger) ||
+ !wait_for_counter(&skel->bss->nr_lazy_kick_running, 1,
+ PHASE_TIMEOUT_MS)) {
+ SCX_ERR("Lazy kick made no progress on CPU %d", ctx->test_cpu);
+ goto out;
+ }
+ stop_gated_worker(&trigger);
+ stop_gated_worker(&challenger);
+ stop_gated_worker(&victim);
+check_exit:
if (skel->data->uei.kind != EXIT_KIND(SCX_EXIT_NONE)) {
SCX_ERR("Scheduler exited unexpectedly (kind=%llu code=%lld)",
(unsigned long long)skel->data->uei.kind,
@@ -321,6 +558,9 @@ static enum scx_test_status run(void *ctx_ptr)
(unsigned long long)skel->bss->nr_finite_ticks);
status = SCX_TEST_PASS;
out:
+ stop_gated_worker(&trigger);
+ stop_gated_worker(&challenger);
+ stop_gated_worker(&victim);
stop_worker(finite_worker);
stop_worker(inf_worker);
if (link)
--
2.55.0
^ permalink raw reply [flat|nested] 3+ messages in thread