[PATCH v2 3/3] debugobjects: Reduce contention on the global pool_lock

From: Waiman Long
Date: Fri Dec 02 2016 - 16:16:42 EST


On a large SMP system with many CPUs, the global pool_lock may become
a performance bottleneck as all the CPUs that need to allocate or
free debug objects have to take the lock. That can sometimes cause
soft lockups like:

NMI watchdog: BUG: soft lockup - CPU#35 stuck for 22s! [rcuos/1:21]
...
RIP: 0010:[<ffffffff817c216b>] [<ffffffff817c216b>]
_raw_spin_unlock_irqrestore+0x3b/0x60
...
Call Trace:
[<ffffffff813f40d1>] free_object+0x81/0xb0
[<ffffffff813f4f33>] debug_check_no_obj_freed+0x193/0x220
[<ffffffff81101a59>] ? trace_hardirqs_on_caller+0xf9/0x1c0
[<ffffffff81284996>] ? file_free_rcu+0x36/0x60
[<ffffffff81251712>] kmem_cache_free+0xd2/0x380
[<ffffffff81284960>] ? fput+0x90/0x90
[<ffffffff81284996>] file_free_rcu+0x36/0x60
[<ffffffff81124c23>] rcu_nocb_kthread+0x1b3/0x550
[<ffffffff81124b71>] ? rcu_nocb_kthread+0x101/0x550
[<ffffffff81124a70>] ? sync_exp_work_done.constprop.63+0x50/0x50
[<ffffffff810c59d1>] kthread+0x101/0x120
[<ffffffff81101a59>] ? trace_hardirqs_on_caller+0xf9/0x1c0
[<ffffffff817c2d32>] ret_from_fork+0x22/0x50

To reduce the amount of contention on the pool_lock, the actual
kmem_cache_free() of the debug objects will be delayed if the pool_lock
is busy. This will temporarily increase the amount of free objects
available at the free pool when the system is busy. As a result,
the number of kmem_cache allocation and freeing should be reduced.

This patch also groups the freeing of debug objects in a batch of 4
to reduce the total number of lock/unlock cycles.

Signed-off-by: Waiman Long <longman@xxxxxxxxxx>
---
lib/debugobjects.c | 31 ++++++++++++++++++++++---------
1 file changed, 22 insertions(+), 9 deletions(-)

diff --git a/lib/debugobjects.c b/lib/debugobjects.c
index dc78217..5476bbe 100644
--- a/lib/debugobjects.c
+++ b/lib/debugobjects.c
@@ -172,25 +172,38 @@ static struct debug_obj *lookup_object(void *addr, struct debug_bucket *b)

/*
* workqueue function to free objects.
+ *
+ * To reduce contention on the global pool_lock, the actual freeing of
+ * debug objects will be delayed if the pool_lock is busy. We also free
+ * the objects in a batch of 4 for each lock/unlock cycle.
*/
+#define ODEBUG_FREE_BATCH 4
static void free_obj_work(struct work_struct *work)
{
- struct debug_obj *obj;
+ struct debug_obj *objs[ODEBUG_FREE_BATCH];
unsigned long flags;
+ int i;

- raw_spin_lock_irqsave(&pool_lock, flags);
- while (obj_pool_free > debug_objects_pool_size) {
- obj = hlist_entry(obj_pool.first, typeof(*obj), node);
- hlist_del(&obj->node);
- obj_pool_free--;
- debug_objects_freed++;
+ if (!raw_spin_trylock_irqsave(&pool_lock, flags))
+ return;
+ while (obj_pool_free >= debug_objects_pool_size + ODEBUG_FREE_BATCH) {
+ for (i = 0; i < ODEBUG_FREE_BATCH; i++) {
+ objs[i] = hlist_entry(obj_pool.first,
+ typeof(*objs[0]), node);
+ hlist_del(&objs[i]->node);
+ }
+
+ obj_pool_free -= ODEBUG_FREE_BATCH;
+ debug_objects_freed += ODEBUG_FREE_BATCH;
/*
* We release pool_lock across kmem_cache_free() to
* avoid contention on pool_lock.
*/
raw_spin_unlock_irqrestore(&pool_lock, flags);
- kmem_cache_free(obj_cache, obj);
- raw_spin_lock_irqsave(&pool_lock, flags);
+ for (i = 0; i < ODEBUG_FREE_BATCH; i++)
+ kmem_cache_free(obj_cache, objs[i]);
+ if (!raw_spin_trylock_irqsave(&pool_lock, flags))
+ return;
}
raw_spin_unlock_irqrestore(&pool_lock, flags);
}
--
1.8.3.1