[PATCH 1/2] sched, livepatch: Fix transition stall for CPU-bound tasks
From: Dylan Hatch
Date: Mon Nov 16 2026 - 23:53:15 EST
Commit 676e8cf70cb0 ("sched,livepatch: Untangle cond_resched() and
live-patching") moved klp_sched_try_switch() from cond_resched() into
__schedule() and gated it on READ_ONCE(curr->__state) & TASK_FREEZABLE.
This causes livepatch transitions to stall when a CPU-bound kthread or
PF_USER_WORKER thread (such as vhost-<pid>) spins in a kernel loop with
cond_resched() for two reasons:
1. A sole runnable task on a CPU (cfs_rq->nr_queued == 1) never has
TIF_NEED_RESCHED set during sched_tick(), and wake_up_state(task,
TASK_INTERRUPTIBLE) in klp_send_signals() is a no-op for a
TASK_RUNNING kthread. Because _cond_resched() is a no-op under
CONFIG_PREEMPTION=y, the task only enters __schedule(SM_PREEMPT) on
return from interrupts via irqentry_exit_cond_resched() when
TIF_NEED_RESCHED is set.
2. Even when __schedule() is entered via preemption or an unfreezable
sleep (such as PF_NOFREEZE kthreads or PF_USER_WORKER threads
sleeping in TASK_INTERRUPTIBLE), curr->__state does not have
TASK_FREEZABLE set, so klp_sched_try_switch() skips the task.
Livepatch safety in __schedule() is already guaranteed by
stack_trace_save_tsk_reliable() inside klp_check_stack().
Fix both issues by:
- Removing the TASK_FREEZABLE check from klp_sched_try_switch() and
checking TIF_PATCH_PENDING inline via klp_sched_need_resched().
- Calling resched_curr(rq) in sched_tick() when klp_sched_need_resched()
is true for the current task, ensuring CPU-bound tasks with a pending
livepatch enter __schedule() at the next timer tick.
Note that when a CPU-bound task is the sole runnable task on its CPU,
entering __schedule(SM_PREEMPT) from preempt_schedule_irq() leaves an
in-kernel interrupt pt_regs frame on the stack during
klp_sched_try_switch(). Completing the transition in that context
requires arch_stack_walk_reliable() to support unwinding across
in-kernel exception frames (as CONFIG_UNWINDER_ORC=y does on x86_64);
configs and architectures whose reliable unwinder rejects in-kernel
exception frames (x86 with CONFIG_FRAME_POINTER=y, arm64, powerpc, and
s390) still require another runnable task to context-switch the spinning
task out until exception-boundary reliable unwinding (e.g. via SFrame)
is available.
Fixes: 676e8cf70cb0 ("sched,livepatch: Untangle cond_resched() and live-patching")
Signed-off-by: Dylan Hatch <dylanbhatch@xxxxxxxxxx>
---
include/linux/livepatch_sched.h | 13 +++++++++++--
kernel/sched/core.c | 3 ++-
2 files changed, 13 insertions(+), 3 deletions(-)
diff --git a/include/linux/livepatch_sched.h b/include/linux/livepatch_sched.h
index 065c185f27638..748b6012e2e06 100644
--- a/include/linux/livepatch_sched.h
+++ b/include/linux/livepatch_sched.h
@@ -11,15 +11,24 @@ void __klp_sched_try_switch(void);
DECLARE_STATIC_KEY_FALSE(klp_sched_try_switch_key);
+static __always_inline bool klp_sched_need_resched(struct task_struct *curr)
+{
+ return static_branch_unlikely(&klp_sched_try_switch_key) &&
+ unlikely(test_tsk_thread_flag(curr, TIF_PATCH_PENDING));
+}
+
static __always_inline void klp_sched_try_switch(struct task_struct *curr)
{
- if (static_branch_unlikely(&klp_sched_try_switch_key) &&
- READ_ONCE(curr->__state) & TASK_FREEZABLE)
+ if (klp_sched_need_resched(curr))
__klp_sched_try_switch();
}
#else /* !CONFIG_LIVEPATCH */
static inline void klp_sched_try_switch(struct task_struct *curr) {}
+static inline bool klp_sched_need_resched(struct task_struct *curr)
+{
+ return false;
+}
#endif /* CONFIG_LIVEPATCH */
#endif /* _LINUX_LIVEPATCH_SCHED_H_ */
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 04b62c16fa8ca..fcea9cea3edf1 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -5854,7 +5854,8 @@ void sched_tick(void)
hw_pressure = arch_scale_hw_pressure(cpu_of(rq));
update_hw_load_avg(rq_clock_task(rq), rq, hw_pressure);
- if (dynamic_preempt_lazy() && tif_test_bit(TIF_NEED_RESCHED_LAZY))
+ if ((dynamic_preempt_lazy() && tif_test_bit(TIF_NEED_RESCHED_LAZY)) ||
+ klp_sched_need_resched(curr))
resched_curr(rq);
donor->sched_class->task_tick(rq, donor, 0);
--
2.56.0.679.gc527c65306-goog