The function current_cgns_cgroup_from_root() expects a stable
cgroup_root, which is currently ensured with RCU read side paired with
cgroup_destroy_root() called after RCU period.

The particular current_cgns_cgroup_from_root() is called from VFS code
and cgroup_root stability can be also ensured by namespace_sem. Mark it
explicitly as a preparation for further rework.

Signed-off-by: Michal Koutný <mkoutny@suse.com>
---
 fs/namespace.c         | 5 ++++-
 include/linux/mount.h  | 4 ++++
 kernel/cgroup/cgroup.c | 7 +++----
 3 files changed, 11 insertions(+), 5 deletions(-)

diff --git a/fs/namespace.c b/fs/namespace.c
index 54847db5b819..0d2333832064 100644
--- a/fs/namespace.c
+++ b/fs/namespace.c
@@ -71,7 +71,10 @@ static DEFINE_IDA(mnt_group_ida);
 static struct hlist_head *mount_hashtable __read_mostly;
 static struct hlist_head *mountpoint_hashtable __read_mostly;
 static struct kmem_cache *mnt_cache __read_mostly;
-static DECLARE_RWSEM(namespace_sem);
+DECLARE_RWSEM(namespace_sem);
+#ifdef CONFIG_LOCKDEP
+EXPORT_SYMBOL_GPL(namespace_sem);
+#endif
 static HLIST_HEAD(unmounted);	/* protected by namespace_sem */
 static LIST_HEAD(ex_mountpoints); /* protected by namespace_sem */
 
diff --git a/include/linux/mount.h b/include/linux/mount.h
index 1ea326c368f7..6277435f6748 100644
--- a/include/linux/mount.h
+++ b/include/linux/mount.h
@@ -80,6 +80,10 @@ static inline struct mnt_idmap *mnt_idmap(const struct vfsmount *mnt)
 	return smp_load_acquire(&mnt->mnt_idmap);
 }
 
+#ifdef CONFIG_LOCKDEP
+extern struct rw_semaphore namespace_sem;
+#endif
+
 extern int mnt_want_write(struct vfsmount *mnt);
 extern int mnt_want_write_file(struct file *file);
 extern void mnt_drop_write(struct vfsmount *mnt);
diff --git a/kernel/cgroup/cgroup.c b/kernel/cgroup/cgroup.c
index 55e5f0110e3b..32d693a797b9 100644
--- a/kernel/cgroup/cgroup.c
+++ b/kernel/cgroup/cgroup.c
@@ -1440,13 +1440,12 @@ current_cgns_cgroup_from_root(struct cgroup_root *root)
 
 	lockdep_assert_held(&css_set_lock);
 
-	rcu_read_lock();
+	/* namespace_sem ensures `root` stability on unmount */
+	lockdep_assert(lockdep_is_held_type(&namespace_sem, -1));
 
 	cset = current->nsproxy->cgroup_ns->root_cset;
 	res = __cset_cgroup_from_root(cset, root);
 
-	rcu_read_unlock();
-
 	return res;
 }
 
@@ -1454,7 +1453,7 @@ current_cgns_cgroup_from_root(struct cgroup_root *root)
  * Look up cgroup associated with current task's cgroup namespace on the default
  * hierarchy.
  *
- * Unlike current_cgns_cgroup_from_root(), this doesn't need locks:
+ * Relaxed locking requirements:
  * - Internal rcu_read_lock is unnecessary because we don't dereference any rcu
  *   pointers.
  * - css_set_lock is not needed because we just read cset->dfl_cgrp.
-- 
2.40.1

On Tue, May 02, 2023 at 03:38:46PM +0200, Michal Koutný wrote:
> The function current_cgns_cgroup_from_root() expects a stable
> cgroup_root, which is currently ensured with RCU read side paired with
> cgroup_destroy_root() called after RCU period.
> 
> The particular current_cgns_cgroup_from_root() is called from VFS code
> and cgroup_root stability can be also ensured by namespace_sem. Mark it
> explicitly as a preparation for further rework.
> 
> Signed-off-by: Michal Koutný <mkoutny@suse.com>
> ---
>  fs/namespace.c         | 5 ++++-
>  include/linux/mount.h  | 4 ++++
>  kernel/cgroup/cgroup.c | 7 +++----
>  3 files changed, 11 insertions(+), 5 deletions(-)
> 
> diff --git a/fs/namespace.c b/fs/namespace.c
> index 54847db5b819..0d2333832064 100644
> --- a/fs/namespace.c
> +++ b/fs/namespace.c
> @@ -71,7 +71,10 @@ static DEFINE_IDA(mnt_group_ida);
>  static struct hlist_head *mount_hashtable __read_mostly;
>  static struct hlist_head *mountpoint_hashtable __read_mostly;
>  static struct kmem_cache *mnt_cache __read_mostly;
> -static DECLARE_RWSEM(namespace_sem);
> +DECLARE_RWSEM(namespace_sem);
> +#ifdef CONFIG_LOCKDEP
> +EXPORT_SYMBOL_GPL(namespace_sem);
> +#endif
>  static HLIST_HEAD(unmounted);	/* protected by namespace_sem */
>  static LIST_HEAD(ex_mountpoints); /* protected by namespace_sem */
>  
> diff --git a/include/linux/mount.h b/include/linux/mount.h
> index 1ea326c368f7..6277435f6748 100644
> --- a/include/linux/mount.h
> +++ b/include/linux/mount.h
> @@ -80,6 +80,10 @@ static inline struct mnt_idmap *mnt_idmap(const struct vfsmount *mnt)
>  	return smp_load_acquire(&mnt->mnt_idmap);
>  }
>  
> +#ifdef CONFIG_LOCKDEP
> +extern struct rw_semaphore namespace_sem;
> +#endif

Nope, we're not putting namespace_sem in a header. The code it protects
is massively sensitive and it interacts with mount_lock and other locks.
This stays private to fs/namespace.c as far as I'm concerned.

Hello,

On Tue, May 23, 2023 at 12:42:46PM +0200, Christian Brauner wrote:
...
> Nope, we're not putting namespace_sem in a header. The code it protects
> is massively sensitive and it interacts with mount_lock and other locks.
> This stays private to fs/namespace.c as far as I'm concerned.

Michal, would it make sense to add a separate locking in cgroup.c? It'll add
a bit more overhead but not massively so and we should be able to get
similar gain without entangling with namespace locking.

Thanks.

-- 
tejun

On 5/2/23 09:38, Michal Koutný wrote:
> The function current_cgns_cgroup_from_root() expects a stable
> cgroup_root, which is currently ensured with RCU read side paired with
> cgroup_destroy_root() called after RCU period.
>
> The particular current_cgns_cgroup_from_root() is called from VFS code
> and cgroup_root stability can be also ensured by namespace_sem. Mark it
> explicitly as a preparation for further rework.
>
> Signed-off-by: Michal Koutný <mkoutny@suse.com>
> ---
>   fs/namespace.c         | 5 ++++-
>   include/linux/mount.h  | 4 ++++
>   kernel/cgroup/cgroup.c | 7 +++----
>   3 files changed, 11 insertions(+), 5 deletions(-)
>
> diff --git a/fs/namespace.c b/fs/namespace.c
> index 54847db5b819..0d2333832064 100644
> --- a/fs/namespace.c
> +++ b/fs/namespace.c
> @@ -71,7 +71,10 @@ static DEFINE_IDA(mnt_group_ida);
>   static struct hlist_head *mount_hashtable __read_mostly;
>   static struct hlist_head *mountpoint_hashtable __read_mostly;
>   static struct kmem_cache *mnt_cache __read_mostly;
> -static DECLARE_RWSEM(namespace_sem);
> +DECLARE_RWSEM(namespace_sem);
> +#ifdef CONFIG_LOCKDEP
> +EXPORT_SYMBOL_GPL(namespace_sem);

I don't think fs/namespace.o and kernel/cgroup/cgroup.o can't be built 
into a kernel module. I doubt we need to export it.


> +#endif
>   static HLIST_HEAD(unmounted);	/* protected by namespace_sem */
>   static LIST_HEAD(ex_mountpoints); /* protected by namespace_sem */
>   
> diff --git a/include/linux/mount.h b/include/linux/mount.h
> index 1ea326c368f7..6277435f6748 100644
> --- a/include/linux/mount.h
> +++ b/include/linux/mount.h
> @@ -80,6 +80,10 @@ static inline struct mnt_idmap *mnt_idmap(const struct vfsmount *mnt)
>   	return smp_load_acquire(&mnt->mnt_idmap);
>   }
>   
> +#ifdef CONFIG_LOCKDEP
> +extern struct rw_semaphore namespace_sem;
> +#endif
> +
>   extern int mnt_want_write(struct vfsmount *mnt);
>   extern int mnt_want_write_file(struct file *file);
>   extern void mnt_drop_write(struct vfsmount *mnt);
> diff --git a/kernel/cgroup/cgroup.c b/kernel/cgroup/cgroup.c
> index 55e5f0110e3b..32d693a797b9 100644
> --- a/kernel/cgroup/cgroup.c
> +++ b/kernel/cgroup/cgroup.c
> @@ -1440,13 +1440,12 @@ current_cgns_cgroup_from_root(struct cgroup_root *root)
>   
>   	lockdep_assert_held(&css_set_lock);
>   
> -	rcu_read_lock();
> +	/* namespace_sem ensures `root` stability on unmount */
> +	lockdep_assert(lockdep_is_held_type(&namespace_sem, -1));
It will be easier if you just use lockdep_is_held() without the 2nd argment.
>   
>   	cset = current->nsproxy->cgroup_ns->root_cset;
>   	res = __cset_cgroup_from_root(cset, root);
>   
> -	rcu_read_unlock();
> -
>   	return res;
>   }
>   
> @@ -1454,7 +1453,7 @@ current_cgns_cgroup_from_root(struct cgroup_root *root)
>    * Look up cgroup associated with current task's cgroup namespace on the default
>    * hierarchy.
>    *
> - * Unlike current_cgns_cgroup_from_root(), this doesn't need locks:
> + * Relaxed locking requirements:
>    * - Internal rcu_read_lock is unnecessary because we don't dereference any rcu
>    *   pointers.
>    * - css_set_lock is not needed because we just read cset->dfl_cgrp.
Cheers,
Longman