| Message ID | 20230502133847.14570-3-mkoutny@suse.com (mailing list archive) |
|---|---|
| State | Mainlined, archived |
| Headers | show |
| Series | Rework locking when rendering mountinfo cgroup paths | expand |
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
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.
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.
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(-)