From 1473c91ed85ae9ccc6605345b64e262931886bc8 Mon Sep 17 00:00:00 2001 From: Thomas Gleixner Date: Mon, 6 Jun 2011 12:20:33 +0200 Subject: [PATCH 131/352] sched: Move mmdrop to RCU on RT Takes sleeping locks and calls into the memory allocator, so nothing we want to do in task switch and oder atomic contexts. Signed-off-by: Thomas Gleixner --- include/linux/mm_types.h | 4 ++++ include/linux/sched.h | 11 +++++++++++ kernel/fork.c | 13 +++++++++++++ kernel/sched/core.c | 19 +++++++++++++++++-- 4 files changed, 45 insertions(+), 2 deletions(-) diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h index 54c704a..e8c44d3 100644 --- a/include/linux/mm_types.h +++ b/include/linux/mm_types.h @@ -11,6 +11,7 @@ #include #include #include +#include #include #include #include @@ -521,6 +522,9 @@ struct mm_struct { bool tlb_flush_batched; #endif struct uprobes_state uprobes_state; +#ifdef CONFIG_PREEMPT_RT_BASE + struct rcu_head delayed_drop; +#endif #ifdef CONFIG_X86_INTEL_MPX /* address of the bounds directory */ void __user *bd_addr; diff --git a/include/linux/sched.h b/include/linux/sched.h index 3ca69a0..1f416c1 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -3099,6 +3099,17 @@ static inline void mmdrop(struct mm_struct *mm) __mmdrop(mm); } +#ifdef CONFIG_PREEMPT_RT_BASE +extern void __mmdrop_delayed(struct rcu_head *rhp); +static inline void mmdrop_delayed(struct mm_struct *mm) +{ + if (atomic_dec_and_test(&mm->mm_count)) + call_rcu(&mm->delayed_drop, __mmdrop_delayed); +} +#else +# define mmdrop_delayed(mm) mmdrop(mm) +#endif + static inline void mmdrop_async_fn(struct work_struct *work) { struct mm_struct *mm = container_of(work, struct mm_struct, async_put_work); diff --git a/kernel/fork.c b/kernel/fork.c index 2a04c19..b0f2926 100644 --- a/kernel/fork.c +++ b/kernel/fork.c @@ -890,6 +890,19 @@ void __mmdrop(struct mm_struct *mm) } EXPORT_SYMBOL_GPL(__mmdrop); +#ifdef CONFIG_PREEMPT_RT_BASE +/* + * RCU callback for delayed mm drop. Not strictly rcu, but we don't + * want another facility to make this work. + */ +void __mmdrop_delayed(struct rcu_head *rhp) +{ + struct mm_struct *mm = container_of(rhp, struct mm_struct, delayed_drop); + + __mmdrop(mm); +} +#endif + static inline void __mmput(struct mm_struct *mm) { VM_BUG_ON(atomic_read(&mm->mm_users)); diff --git a/kernel/sched/core.c b/kernel/sched/core.c index 412542f..4403d29 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -2840,8 +2840,12 @@ static struct rq *finish_task_switch(struct task_struct *prev) finish_arch_post_lock_switch(); fire_sched_in_preempt_notifiers(current); + /* + * We use mmdrop_delayed() here so we don't have to do the + * full __mmdrop() when we are the last user. + */ if (mm) - mmdrop(mm); + mmdrop_delayed(mm); if (unlikely(prev_state == TASK_DEAD)) { if (prev->sched_class->task_dead) prev->sched_class->task_dead(prev); @@ -5690,6 +5694,8 @@ void sched_setnuma(struct task_struct *p, int nid) #endif /* CONFIG_NUMA_BALANCING */ #ifdef CONFIG_HOTPLUG_CPU +static DEFINE_PER_CPU(struct mm_struct *, idle_last_mm); + /* * Ensures that the idle task is using init_mm right before its cpu goes * offline. @@ -5704,7 +5710,12 @@ void idle_task_exit(void) switch_mm(mm, &init_mm, current); finish_arch_post_lock_switch(); } - mmdrop(mm); + /* + * Defer the cleanup to an alive cpu. On RT we can neither + * call mmdrop() nor mmdrop_delayed() from here. + */ + per_cpu(idle_last_mm, smp_processor_id()) = mm; + } /* @@ -7778,6 +7789,10 @@ int sched_cpu_dying(unsigned int cpu) update_max_interval(); nohz_balance_exit_idle(cpu); hrtick_clear(rq); + if (per_cpu(idle_last_mm, cpu)) { + mmdrop_delayed(per_cpu(idle_last_mm, cpu)); + per_cpu(idle_last_mm, cpu) = NULL; + } return 0; } #endif -- 2.7.4