Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1534484 > unrolled thread

[PATCH 0/3] debugobjects: Reduce global pool_lock contention

Back to article view | Back to linux.kernel

Contents

  [PATCH 0/3] debugobjects: Reduce global pool_lock contention Waiman Long <longman@redhat.com> - 2016-12-01 23:00 +0100
    [PATCH 1/3] debugobjects: Scale thresholds with # of CPUs Waiman Long <longman@redhat.com> - 2016-12-01 23:00 +0100

#1534484 — [PATCH 0/3] debugobjects: Reduce global pool_lock contention

This patchset aims to reduce contention of the global pool_lock
while improving performance at the same time. It is done to resolve
the following soft lockup problem with a debug kernel in some of the
large SMP systems:

 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

On a 8-socket IvyBridge-EX system (120 cores, 240 threads), the
elapsed time of a 4.9-rc7 kernel parallel build (make -j 240) was
reduced from 7m57s to 7m19s with a patched 4.9-rc7 kernel. There
was also about a 10X reduction in the number of debug objects being
allocated from or freed to the kmemcache during the kernel build.

Waiman Long (3):
  debugobjects: Scale thresholds with # of CPUs
  debugobjects: Track number of kmem_cache_alloc/kmem_cache_free done
  debugobjects: Reduce contention on the global pool_lock

 lib/debugobjects.c | 57 ++++++++++++++++++++++++++++++++++++++++++------------
 1 file changed, 45 insertions(+), 12 deletions(-)

-- 
1.8.3.1

[toc] | [next] | [standalone]

#1534485 — [PATCH 1/3] debugobjects: Scale thresholds with # of CPUs

On a large SMP systems with hundreds of CPUs, the current thresholds
for allocating and freeing debug objects (256 and 1024 respectively)
may not work well. This can cause a lot of needless calls to
kmem_aloc() and kmem_free() on those systems.

To alleviate this thrashing problem, the object freeing threshold
is now increased to "1024 + # of CPUs * 32". Whereas the object
allocation threshold is increased to "256 + # of CPUs * 4". That
should make the debug objects subsystem scale better with the number
of CPUs available in the system.

Signed-off-by: Waiman Long <longman@redhat.com>
---
 lib/debugobjects.c | 20 +++++++++++++++-----
 1 file changed, 15 insertions(+), 5 deletions(-)

diff --git a/lib/debugobjects.c b/lib/debugobjects.c
index a8e1260..bb415a7 100644
--- a/lib/debugobjects.c
+++ b/lib/debugobjects.c
@@ -52,7 +52,10 @@ struct debug_bucket {
 static int			debug_objects_warnings __read_mostly;
 static int			debug_objects_enabled __read_mostly
 				= CONFIG_DEBUG_OBJECTS_ENABLE_DEFAULT;
-
+static int			debug_objects_pool_size __read_mostly
+				= ODEBUG_POOL_SIZE;
+static int			debug_objects_pool_min_level __read_mostly
+				= ODEBUG_POOL_MIN_LEVEL;
 static struct debug_obj_descr	*descr_test  __read_mostly;
 
 static void free_obj_work(struct work_struct *work);
@@ -88,13 +91,13 @@ static void fill_pool(void)
 	struct debug_obj *new;
 	unsigned long flags;
 
-	if (likely(obj_pool_free >= ODEBUG_POOL_MIN_LEVEL))
+	if (likely(obj_pool_free >= debug_objects_pool_min_level))
 		return;
 
 	if (unlikely(!obj_cache))
 		return;
 
-	while (obj_pool_free < ODEBUG_POOL_MIN_LEVEL) {
+	while (obj_pool_free < debug_objects_pool_min_level) {
 
 		new = kmem_cache_zalloc(obj_cache, gfp);
 		if (!new)
@@ -169,7 +172,7 @@ static void free_obj_work(struct work_struct *work)
 	unsigned long flags;
 
 	raw_spin_lock_irqsave(&pool_lock, flags);
-	while (obj_pool_free > ODEBUG_POOL_SIZE) {
+	while (obj_pool_free > debug_objects_pool_size) {
 		obj = hlist_entry(obj_pool.first, typeof(*obj), node);
 		hlist_del(&obj->node);
 		obj_pool_free--;
@@ -198,7 +201,7 @@ static void free_object(struct debug_obj *obj)
 	 * schedule work when the pool is filled and the cache is
 	 * initialized:
 	 */
-	if (obj_pool_free > ODEBUG_POOL_SIZE && obj_cache)
+	if (obj_pool_free > debug_objects_pool_size && obj_cache)
 		sched = keventd_up();
 	hlist_add_head(&obj->node, &obj_pool);
 	obj_pool_free++;
@@ -1108,4 +1111,11 @@ void __init debug_objects_mem_init(void)
 		pr_warn("out of memory.\n");
 	} else
 		debug_objects_selftest();
+
+	/*
+	 * Increase the thresholds for allocating and freeing objects
+	 * according to the number of possible CPUs available in the system.
+	 */
+	debug_objects_pool_size += num_possible_cpus() * 32;
+	debug_objects_pool_min_level += num_possible_cpus() * 4;
 }
-- 
1.8.3.1

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web