[Devel] [PATCH v2 04/11] memcg: disable softirq in lock_page_cgroup()

[Greg Thelen]

If pages are being migrated from a memcg, then updates to that
memcg's page statistics are protected by grabbing a bit spin lock
using lock_page_cgroup().  In an upcoming commit memcg dirty page
accounting will be updating memcg page accounting (specifically:
num writeback pages) from softirq.  Avoid a deadlocking nested
spin lock attempt by disabling softirq on the local processor
when grabbing the page_cgroup bit_spin_lock in lock_page_cgroup().
This avoids the following deadlock:
statistic
      CPU 0             CPU 1
                    inc_file_mapped
                    rcu_read_lock
  start move
  synchronize_rcu
                    lock_page_cgroup
                      softirq
                      test_clear_page_writeback
                      mem_cgroup_dec_page_stat(NR_WRITEBACK)
                      rcu_read_lock
                      lock_page_cgroup   /* deadlock */
                      unlock_page_cgroup
                      rcu_read_unlock
                    unlock_page_cgroup
                    rcu_read_unlock

By disabling softirq in lock_page_cgroup, nested calls are avoided.
The softirq would be delayed until after inc_file_mapped enables
softirq when calling unlock_page_cgroup().

The normal, fast path, of memcg page stat updates typically
does not need to call lock_page_cgroup(), so this change does
not affect the performance of the common case page accounting.

Signed-off-by: Andrea Righi <arighi at develer.com>
Signed-off-by: Greg Thelen <gthelen at google.com>
---
 include/linux/page_cgroup.h |    6 ++++++
 1 files changed, 6 insertions(+), 0 deletions(-)

diff --git a/include/linux/page_cgroup.h b/include/linux/page_cgroup.h
index b59c298..0585546 100644
--- a/include/linux/page_cgroup.h
+++ b/include/linux/page_cgroup.h
@@ -3,6 +3,8 @@
 
 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
 #include <linux/bit_spinlock.h>
+#include <linux/hardirq.h>
+
 /*
  * Page Cgroup can be considered as an extended mem_map.
  * A page_cgroup page is associated with every page descriptor. The
@@ -119,12 +121,16 @@ static inline enum zone_type page_cgroup_zid(struct page_cgroup *pc)
 
 static inline void lock_page_cgroup(struct page_cgroup *pc)
 {
+	/* This routine is only deadlock safe from softirq or lower. */
+	VM_BUG_ON(in_irq());
+	local_bh_disable();
 	bit_spin_lock(PCG_LOCK, &pc->flags);
 }
 
 static inline void unlock_page_cgroup(struct page_cgroup *pc)
 {
 	bit_spin_unlock(PCG_LOCK, &pc->flags);
+	local_bh_enable();
 }
 
 #else /* CONFIG_CGROUP_MEM_RES_CTLR */
-- 
1.7.1

_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers