On 2022-09-04 23:16:12 [+0800], Guo Ren wrote: > > diff --git a/arch/riscv/kernel/entry.S b/arch/riscv/kernel/entry.S > > index b9eda3fcbd6d..595100a4c2c7 100644 > > --- a/arch/riscv/kernel/entry.S > > +++ b/arch/riscv/kernel/entry.S > > @@ -361,9 +361,14 @@ restore_all: > > resume_kernel: > > REG_L s0, TASK_TI_PREEMPT_COUNT(tp) > > bnez s0, restore_all > > - REG_L s0, TASK_TI_FLAGS(tp) > > - andi s0, s0, _TIF_NEED_RESCHED > > + REG_L s1, TASK_TI_FLAGS(tp) > > + andi s0, s1, _TIF_NEED_RESCHED > > + bnez s0, 1f > > + REG_L s0, TASK_TI_PREEMPT_LAZY_COUNT(tp) > > + bnez s0, restore_all > > + andi s0, s1, _TIF_NEED_RESCHED_LAZY > Can you tell me, who increased/decreased the PREEMPT_LAZY_COUNT? And > who set NEED_RESCHED_LAZY? There is "generic" code in the PREEMPT_RT patch doing that. The counter is incremented/ decremented via preempt_lazy_enable()/disable() and one of the user is migrate_disable()/enable(). Basically if a task is task_is_realtime() then NEED_RESCHED is set for the wakeup. For the remaining states (SCHED_OTHER, …) NEED_RESCHED_LAZY is set for the wakeup. This can be delayed if the task is in a "preempt disable lazy" section (similar to a preempt_disable() section) but a task_is_realtime() can still be scheduled if needed. See details at https://git.kernel.org/pub/scm/linux/kernel/git/rt/linux-rt-devel.git/plain/patches/sched__Add_support_for_lazy_preemption.patch?h=linux-6.0.y-rt-patches Sebastian
