Re: [tip:sched/core] sched/numa: Move task_numa_free() to __put_task_struct()

[Mike Galbraith <bitbucket@online.de>]

On Mon, 2014-04-07 at 07:29 +0200, Mike Galbraith wrote:

> I'm not getting it.
> 
> I moved task_numa_free() from one interrupts enabled spot to another.
> But, with numa=fake=4 and lockdep enabled, not only does lockdep gripe,
> my little box locks up on splat.  Saying spin_lock/unlock_irq() did the
> expected, just moved lockdep gripe to task_numa_fault().
> 
> 
> > [ 2590.270067]  Possible interrupt unsafe locking scenario:
> > [ 2590.270067]
> > [ 2590.270067]        CPU0                    CPU1
> > [ 2590.270067]        ----                    ----
> > [ 2590.270067]   lock(&(&grp->lock)->rlock);
> > [ 2590.270067]                                local_irq_disable();
> > [ 2590.270067]                                lock(&(&new_timer->it_lock)->rlock);
> > [ 2590.270067]                                lock(&(&grp->lock)->rlock);
> > [ 2590.270067]   <Interrupt>
> > [ 2590.270067]     lock(&(&new_timer->it_lock)->rlock);
> > [ 2590.270067]
> > [ 2590.270067]  *** DEADLOCK ***
> 
> Ok, so how did I manage that HARDIRQ-safe -> HARDIRQ-unsafe?

Think I'll turn lockdep off, and make context switches take a good long
while after finish_lock_switch(), but meanwhile, this made it happy.

Sasha reports that lockdep claims 156654f491dd8d52687a5fbe1637f472a52ce75b made
numa_group.lock interrupt unsafe.  While I don't see how that could be given the
commit in question moved task_numa_free() from one irq enabled region to another,
the below does make both gripes and lockup upon gripe with numa=fake=4 go away.

Reported-by: Sasha Levin <sasha.levin@oracle.com>
Not-signed-off-by: Mike Galbraith <bitbucket@online.de>
---
 kernel/sched/fair.c  |   12 +++++++-----
 kernel/sched/sched.h |    9 +++++++++
 2 files changed, 16 insertions(+), 5 deletions(-)

--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -1497,7 +1497,7 @@ static void task_numa_placement(struct t
 	/* If the task is part of a group prevent parallel updates to group stats */
 	if (p->numa_group) {
 		group_lock = &p->numa_group->lock;
-		spin_lock(group_lock);
+		spin_lock_irq(group_lock);
 	}
 
 	/* Find the node with the highest number of faults */
@@ -1572,7 +1572,7 @@ static void task_numa_placement(struct t
 			}
 		}
 
-		spin_unlock(group_lock);
+		spin_unlock_irq(group_lock);
 	}
 
 	/* Preferred node as the node with the most faults */
@@ -1677,7 +1677,8 @@ static void task_numa_group(struct task_
 	if (!join)
 		return;
 
-	double_lock(&my_grp->lock, &grp->lock);
+	BUG_ON(irqs_disabled());
+	double_lock_irq(&my_grp->lock, &grp->lock);
 
 	for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
 		my_grp->faults[i] -= p->numa_faults_memory[i];
@@ -1692,6 +1693,7 @@ static void task_numa_group(struct task_
 
 	spin_unlock(&my_grp->lock);
 	spin_unlock(&grp->lock);
+	local_irq_enable();
 
 	rcu_assign_pointer(p->numa_group, grp);
 
@@ -1710,14 +1712,14 @@ void task_numa_free(struct task_struct *
 	void *numa_faults = p->numa_faults_memory;
 
 	if (grp) {
-		spin_lock(&grp->lock);
+		spin_lock_irq(&grp->lock);
 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
 			grp->faults[i] -= p->numa_faults_memory[i];
 		grp->total_faults -= p->total_numa_faults;
 
 		list_del(&p->numa_entry);
 		grp->nr_tasks--;
-		spin_unlock(&grp->lock);
+		spin_unlock_irq(&grp->lock);
 		rcu_assign_pointer(p->numa_group, NULL);
 		put_numa_group(grp);
 	}
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -1388,6 +1388,15 @@ static inline void double_lock(spinlock_
 	spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
 }
 
+static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
+{
+	if (l1 > l2)
+		swap(l1, l2);
+
+	spin_lock_irq(l1);
+	spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
+}
+
 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
 {
 	if (l1 > l2)