On 02/03/2016 12:44 PM, Vladimir Davydov wrote:
On Wed, Feb 03, 2016 at 10:52:13AM +0300, Dmitry Safonov wrote: ...diff --git a/mm/slab_common.c b/mm/slab_common.c index b50aef0..2bfc0b1 100644 --- a/mm/slab_common.c +++ b/mm/slab_common.c @@ -451,6 +451,8 @@ EXPORT_SYMBOL(kmem_cache_create); static int shutdown_cache(struct kmem_cache *s, struct list_head *release, bool *need_rcu_barrier) { + sysfs_slab_remove(s); +shutdown_cache is called with slab_mutex held. slab_attr_store may take the mutex. So placing sysfs_slab_remove here introduces a potential deadlock.if (__kmem_cache_shutdown(s) != 0) return -EBUSY;@@ -468,13 +470,8 @@ static void release_caches(struct list_head *release, bool need_rcu_barrier)if (need_rcu_barrier) rcu_barrier();- list_for_each_entry_safe(s, s2, release, list) {-#ifdef SLAB_SUPPORTS_SYSFS - sysfs_slab_remove(s); -#else + list_for_each_entry_safe(s, s2, release, list) slab_kmem_cache_release(s); -#endif - } }#if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)@@ -614,6 +611,7 @@ void memcg_destroy_kmem_caches(struct mem_cgroup *memcg) list_for_each_entry_safe(s, s2, &slab_caches, list) { if (is_root_cache(s) || s->memcg_params.memcg != memcg) continue; +Please remove this hunk./* * The cgroup is about to be freed and therefore has no charges * left. Hence, all its caches must be empty by now. diff --git a/mm/slub.c b/mm/slub.c index 2e1355a..b6a68b7 100644 --- a/mm/slub.c +++ b/mm/slub.c @@ -5296,11 +5296,6 @@ static void memcg_propagate_slab_attrs(struct kmem_cache *s) #endif }-static void kmem_cache_release(struct kobject *k)-{ - slab_kmem_cache_release(to_slab(k)); -} - static const struct sysfs_ops slab_sysfs_ops = { .show = slab_attr_show, .store = slab_attr_store, @@ -5308,7 +5303,6 @@ static const struct sysfs_ops slab_sysfs_ops = {static struct kobj_type slab_ktype = {.sysfs_ops = &slab_sysfs_ops, - .release = kmem_cache_release,I surmise this will resurrect the bug that was fixed by 41a212859a4dd ("slub: use sysfs'es release mechanism for kmem_cache").
So, I move __kmem_cache_shutdown into slab_kmem_cache_release, release list will still be collected under slab_mutex. Seems like no other way to do this. Will resend mended version.
Thanks, Vladimir
-- Regards, Dmitry Safonov
