diff mbox series

[v3] mm/slab.c: add node spinlock protect in __cache_free_alien

Message ID 20200730101934.38343-1-qiang.zhang@windriver.com (mailing list archive)
State New, archived
Headers show
Series [v3] mm/slab.c: add node spinlock protect in __cache_free_alien | expand

Commit Message

Zhang, Qiang July 30, 2020, 10:19 a.m. UTC 
From: Zhang Qiang <qiang.zhang@windriver.com>

for example:
			        node0
	cpu0				                cpu1
slab_dead_cpu
   >mutex_lock(&slab_mutex)
     >cpuup_canceled                            slab_dead_cpu
       >mask = cpumask_of_node(node)               >mutex_lock(&slab_mutex)
       >n = get_node(cachep0, node0)
       >spin_lock_irq(n&->list_lock)
       >if (!cpumask_empty(mask)) == true
       	>spin_unlock_irq(&n->list_lock)
	>goto free_slab
       ....
   >mutex_unlock(&slab_mutex)

....						   >cpuup_canceled
						     >mask = cpumask_of_node(node)
kmem_cache_free(cachep0 )			     >n = get_node(cachep0, node0)
 >__cache_free_alien(cachep0 )			     >spin_lock_irq(n&->list_lock)
   >n = get_node(cachep0, node0)		     >if (!cpumask_empty(mask)) == false
   >if (n->alien && n->alien[page_node])	     >alien = n->alien
     >alien = n->alien[page_node]	             >n->alien = NULL
     >....					     >spin_unlock_irq(&n->list_lock)
						     >....

Due to multiple cpu offline, The same cache in a node may be operated in
parallel,the "n->alien" should be protect.

Fixes: 6731d4f12315 ("slab: Convert to hotplug state machine")
Signed-off-by: Zhang Qiang <qiang.zhang@windriver.com>
---
 v1->v2->v3:
 change submission information and fixes tags.

 mm/slab.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

Comments

David Rientjes July 30, 2020, 11:45 p.m. UTC | #1 
On Thu, 30 Jul 2020, qiang.zhang@windriver.com wrote:

> From: Zhang Qiang <qiang.zhang@windriver.com>
> 
> for example:
> 			        node0
> 	cpu0				                cpu1
> slab_dead_cpu
>    >mutex_lock(&slab_mutex)
>      >cpuup_canceled                            slab_dead_cpu
>        >mask = cpumask_of_node(node)               >mutex_lock(&slab_mutex)
>        >n = get_node(cachep0, node0)
>        >spin_lock_irq(n&->list_lock)
>        >if (!cpumask_empty(mask)) == true
>        	>spin_unlock_irq(&n->list_lock)
> 	>goto free_slab
>        ....
>    >mutex_unlock(&slab_mutex)
> 
> ....						   >cpuup_canceled
> 						     >mask = cpumask_of_node(node)
> kmem_cache_free(cachep0 )			     >n = get_node(cachep0, node0)
>  >__cache_free_alien(cachep0 )			     >spin_lock_irq(n&->list_lock)
>    >n = get_node(cachep0, node0)		     >if (!cpumask_empty(mask)) == false
>    >if (n->alien && n->alien[page_node])	     >alien = n->alien
>      >alien = n->alien[page_node]	             >n->alien = NULL
>      >....					     >spin_unlock_irq(&n->list_lock)
> 						     >....
> 

As mentioned in the review of v1 of this patch, we likely want to do a fix 
for cpuup_canceled() instead.
diff mbox series

Patch

diff --git a/mm/slab.c b/mm/slab.c
index a89633603b2d..290523c90b4e 100644
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -759,8 +759,10 @@  static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
 
 	n = get_node(cachep, node);
 	STATS_INC_NODEFREES(cachep);
+	spin_lock(&n->list_lock);
 	if (n->alien && n->alien[page_node]) {
 		alien = n->alien[page_node];
+		spin_unlock(&n->list_lock);
 		ac = &alien->ac;
 		spin_lock(&alien->lock);
 		if (unlikely(ac->avail == ac->limit)) {
@@ -769,14 +771,15 @@  static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
 		}
 		ac->entry[ac->avail++] = objp;
 		spin_unlock(&alien->lock);
-		slabs_destroy(cachep, &list);
 	} else {
+		spin_unlock(&n->list_lock);
 		n = get_node(cachep, page_node);
 		spin_lock(&n->list_lock);
 		free_block(cachep, &objp, 1, page_node, &list);
 		spin_unlock(&n->list_lock);
-		slabs_destroy(cachep, &list);
 	}
+
+	slabs_destroy(cachep, &list);
 	return 1;
 }