sched-mmdrop-delayed.patch

Needs thread context (pgd_lock) -> ifdeffed. workqueues wont work with
RT

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>

4 files changed, 49 insertions, 3 deletions

diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index 1ec126f23eaa..c303a27981fe 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -12,6 +12,7 @@
 #include <linux/completion.h>
 #include <linux/cpumask.h>
 #include <linux/page-debug-flags.h>
+#include <linux/rcupdate.h>
 #include <asm/page.h>
 #include <asm/mmu.h>
 
@@ -393,6 +394,9 @@ struct mm_struct {
 #ifdef CONFIG_CPUMASK_OFFSTACK
 	struct cpumask cpumask_allocation;
 #endif
+#ifdef CONFIG_PREEMPT_RT_BASE
+	struct rcu_head delayed_drop;
+#endif
 };
 
 static inline void mm_init_cpumask(struct mm_struct *mm)
diff --git a/include/linux/sched.h b/include/linux/sched.h
index bf637e27d129..866ddae3a98d 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -2264,12 +2264,24 @@ extern struct mm_struct * mm_alloc(void);
 
 /* mmdrop drops the mm and the page tables */
 extern void __mmdrop(struct mm_struct *);
+
 static inline void mmdrop(struct mm_struct * mm)
 {
 	if (unlikely(atomic_dec_and_test(&mm->mm_count)))
 		__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
+
 /* mmput gets rid of the mappings and all user-space */
 extern void mmput(struct mm_struct *);
 /* Grab a reference to a task's mm, if it is not already going away */
diff --git a/kernel/fork.c b/kernel/fork.c
index 6679093b31e2..56188b0c9d21 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -201,7 +201,7 @@ EXPORT_SYMBOL_GPL(__put_task_struct);
 #else
 void __put_task_struct_cb(struct rcu_head *rhp)
 {
-	struct task_struct *tsk = container_of(rhp, struct task_struct, rcu);
+	struct task_struct *tsk = container_of(rhp, struct task_struct, put_rcu);
 
 	__put_task_struct(tsk);
 
@@ -552,6 +552,19 @@ 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
+
 /*
  * Decrement the use count and release all resources for an mm.
  */
diff --git a/kernel/sched.c b/kernel/sched.c
index 6e2bf9e29ca5..3ff2d0f1fcfe 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -3174,8 +3174,12 @@ static void finish_task_switch(struct rq *rq, struct task_struct *prev)
 	finish_lock_switch(rq, prev);
 
 	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)) {
 		/*
 		 * Remove function-return probe instances associated with this
@@ -6303,6 +6307,8 @@ static int migration_cpu_stop(void *data)
 
 #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.
@@ -6315,7 +6321,12 @@ void idle_task_exit(void)
 
 	if (mm != &init_mm)
 		switch_mm(mm, &init_mm, current);
-	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;
 }
 
 /*
@@ -6660,6 +6671,12 @@ migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
 		migrate_nr_uninterruptible(rq);
 		calc_global_load_remove(rq);
 		break;
+	case CPU_DEAD:
+		if (per_cpu(idle_last_mm, cpu)) {
+			mmdrop(per_cpu(idle_last_mm, cpu));
+			per_cpu(idle_last_mm, cpu) = NULL;
+		}
+		break;
 #endif
 	}