Re: [PATCH v2] mm/slab_common: Deleting kobject in kmem_cache_destroy() without holding slab_mutex/cpu_hotplug_lock

From: Waiman Long
Date: Wed Aug 10 2022 - 14:46:05 EST


On 8/10/22 14:27, Waiman Long wrote:
On 8/10/22 14:10, Roman Gushchin wrote:
On Wed, Aug 10, 2022 at 12:49:46PM -0400, Waiman Long wrote:
A circular locking problem is reported by lockdep due to the following
circular locking dependency.

   +--> cpu_hotplug_lock --> slab_mutex --> kn->active --+
   |                                                     |
   +-----------------------------------------------------+

The forward cpu_hotplug_lock ==> slab_mutex ==> kn->active dependency
happens in

   kmem_cache_destroy():    cpus_read_lock(); mutex_lock(&slab_mutex);
   ==> sysfs_slab_unlink()
       ==> kobject_del()
           ==> kernfs_remove()
          ==> __kernfs_remove()
              ==> kernfs_drain(): rwsem_acquire(&kn->dep_map, ...);

The backward kn->active ==> cpu_hotplug_lock dependency happens in

   kernfs_fop_write_iter(): kernfs_get_active();
   ==> slab_attr_store()
       ==> cpu_partial_store()
           ==> flush_all(): cpus_read_lock()

One way to break this circular locking chain is to avoid holding
cpu_hotplug_lock and slab_mutex while deleting the kobject in
sysfs_slab_unlink() which should be equivalent to doing a write_lock
and write_unlock pair of the kn->active virtual lock.

Since the kobject structures are not protected by slab_mutex or the
cpu_hotplug_lock, we can certainly release those locks before doing
the delete operation.

Move sysfs_slab_unlink() and sysfs_slab_release() to the newly
created kmem_cache_release() and call it outside the slab_mutex &
cpu_hotplug_lock critical sections.

Signed-off-by: Waiman Long <longman@xxxxxxxxxx>
---
  [v2] Break kmem_cache_release() helper into 2 separate ones.

  mm/slab_common.c | 54 +++++++++++++++++++++++++++++++++---------------
  1 file changed, 37 insertions(+), 17 deletions(-)

diff --git a/mm/slab_common.c b/mm/slab_common.c
index 17996649cfe3..7742d0446d8b 100644
--- a/mm/slab_common.c
+++ b/mm/slab_common.c
@@ -392,6 +392,36 @@ kmem_cache_create(const char *name, unsigned int size, unsigned int align,
  }
  EXPORT_SYMBOL(kmem_cache_create);
  +#ifdef SLAB_SUPPORTS_SYSFS
+static void kmem_cache_workfn_release(struct kmem_cache *s)
+{
+    sysfs_slab_release(s);
+}
+#else
+static void kmem_cache_workfn_release(struct kmem_cache *s)
+{
+    slab_kmem_cache_release(s);
+}
+#endif
+
+/*
+ * For a given kmem_cache, kmem_cache_destroy() should only be called
+ * once or there will be a use-after-free problem. The actual deletion
+ * and release of the kobject does not need slab_mutex or cpu_hotplug_lock
+ * protection. So they are now done without holding those locks.
+ */
+static void kmem_cache_release(struct kmem_cache *s)
+{
+#ifdef SLAB_SUPPORTS_SYSFS
+    sysfs_slab_unlink(s);
+#endif
+
+    if (s->flags & SLAB_TYPESAFE_BY_RCU)
+        schedule_work(&slab_caches_to_rcu_destroy_work);
+    else
+        kmem_cache_workfn_release(s);
+}
+
  static void slab_caches_to_rcu_destroy_workfn(struct work_struct *work)
  {
      LIST_HEAD(to_destroy);
@@ -418,11 +448,7 @@ static void slab_caches_to_rcu_destroy_workfn(struct work_struct *work)
      list_for_each_entry_safe(s, s2, &to_destroy, list) {
          debugfs_slab_release(s);
          kfence_shutdown_cache(s);
-#ifdef SLAB_SUPPORTS_SYSFS
-        sysfs_slab_release(s);
-#else
-        slab_kmem_cache_release(s);
-#endif
+        kmem_cache_workfn_release(s);
      }
  }
  @@ -437,20 +463,10 @@ static int shutdown_cache(struct kmem_cache *s)
      list_del(&s->list);
        if (s->flags & SLAB_TYPESAFE_BY_RCU) {
-#ifdef SLAB_SUPPORTS_SYSFS
-        sysfs_slab_unlink(s);
-#endif
          list_add_tail(&s->list, &slab_caches_to_rcu_destroy);
-        schedule_work(&slab_caches_to_rcu_destroy_work);
Hi Waiman!

This version is much more readable, thank you!

But can we, please, leave this schedule_work(&slab_caches_to_rcu_destroy_work)
call here? I don't see a good reason to move it, do I miss something?
It's nice to have list_add_tail() and schedule_work() calls nearby, so
it's obvious we can't miss the latter.

The reason that I need to move out schedule_work() as well is to make sure that sysfs_slab_unlink() is called before sysfs_slab_release(). I can't guarantee that if I do schedule_work() first. On the other hand, moving sysfs_slab_unlink() into kmem_cache_workfn_release() introduces unknown delay of when the sysfs file will be removed. I can add some comment to make it more clear.

OK, I just realize that the current patch doesn't have the ordering guarantee either if another kmem_cache_destroy() is happening in parallel. I will have to push sysfs_slab_unlink() into kmem_cache_workfn_release() and tolerate some delay in the disappearance of the sysfs files. Now I can move schedule_work() back to after list_add_tail().

Cheers,
Longman