On Wed, Mar 20, 2024 at 10:51:34AM +0100, Jan Kara wrote:
> On Wed 20-03-24 11:37:57, Amir Goldstein wrote:
> > On Wed, Mar 20, 2024 at 10:47 AM Christian Brauner <brauner@xxxxxxxxxx> wrote:
> > >
> > > On Sun, Mar 17, 2024 at 08:41:51PM +0200, Amir Goldstein wrote:
> > > > Define a container struct fsnotify_sb_info to hold per-sb state,
> > > > including the reference to sb marks connector.
> > > >
> > > > Allocate the fsnotify_sb_info state before attaching connector to any
> > > > object on the sb and free it only when killing sb.
> > > >
> > > > This state is going to be used for storing per priority watched objects
> > > > counters.
> > > >
> > > > Suggested-by: Jan Kara <jack@xxxxxxx>
> > > > Signed-off-by: Amir Goldstein <amir73il@xxxxxxxxx>
> > > > ---
> > > > fs/notify/fsnotify.c | 16 +++++++++++++---
> > > > fs/notify/fsnotify.h | 9 ++++++++-
> > > > fs/notify/mark.c | 32 +++++++++++++++++++++++++++++++-
> > > > include/linux/fs.h | 8 ++++----
> > > > include/linux/fsnotify_backend.h | 17 +++++++++++++++++
> > > > 5 files changed, 73 insertions(+), 9 deletions(-)
> > > >
> > > > diff --git a/fs/notify/fsnotify.c b/fs/notify/fsnotify.c
> > > > index 503e7c75e777..fb3f36bc6ea9 100644
> > > > --- a/fs/notify/fsnotify.c
> > > > +++ b/fs/notify/fsnotify.c
> > > > @@ -89,11 +89,18 @@ static void fsnotify_unmount_inodes(struct super_block *sb)
> > > >
> > > > void fsnotify_sb_delete(struct super_block *sb)
> > > > {
> > > > + struct fsnotify_sb_info *sbinfo = fsnotify_sb_info(sb);
> > > > +
> > > > + /* Were any marks ever added to any object on this sb? */
> > > > + if (!sbinfo)
> > > > + return;
> > > > +
> > > > fsnotify_unmount_inodes(sb);
> > > > fsnotify_clear_marks_by_sb(sb);
> > > > /* Wait for outstanding object references from connectors */
> > > > wait_var_event(fsnotify_sb_watched_objects(sb),
> > > > !atomic_long_read(fsnotify_sb_watched_objects(sb)));
> > > > + kfree(sbinfo);
> > > > }
> > > >
> > > > /*
> > > > @@ -489,6 +496,7 @@ int fsnotify(__u32 mask, const void *data, int data_type, struct inode *dir,
> > > > {
> > > > const struct path *path = fsnotify_data_path(data, data_type);
> > > > struct super_block *sb = fsnotify_data_sb(data, data_type);
> > > > + struct fsnotify_sb_info *sbinfo = fsnotify_sb_info(sb);
> > > > struct fsnotify_iter_info iter_info = {};
> > > > struct mount *mnt = NULL;
> > > > struct inode *inode2 = NULL;
> > > > @@ -525,7 +533,7 @@ int fsnotify(__u32 mask, const void *data, int data_type, struct inode *dir,
> > > > * SRCU because we have no references to any objects and do not
> > > > * need SRCU to keep them "alive".
> > > > */
> > > > - if (!sb->s_fsnotify_marks &&
> > > > + if ((!sbinfo || !sbinfo->sb_marks) &&
> > > > (!mnt || !mnt->mnt_fsnotify_marks) &&
> > > > (!inode || !inode->i_fsnotify_marks) &&
> > > > (!inode2 || !inode2->i_fsnotify_marks))
> > > > @@ -552,8 +560,10 @@ int fsnotify(__u32 mask, const void *data, int data_type, struct inode *dir,
> > > >
> > > > iter_info.srcu_idx = srcu_read_lock(&fsnotify_mark_srcu);
> > > >
> > > > - iter_info.marks[FSNOTIFY_ITER_TYPE_SB] =
> > > > - fsnotify_first_mark(&sb->s_fsnotify_marks);
> > > > + if (sbinfo) {
> > > > + iter_info.marks[FSNOTIFY_ITER_TYPE_SB] =
> > > > + fsnotify_first_mark(&sbinfo->sb_marks);
> > > > + }
> > > > if (mnt) {
> > > > iter_info.marks[FSNOTIFY_ITER_TYPE_VFSMOUNT] =
> > > > fsnotify_first_mark(&mnt->mnt_fsnotify_marks);
> > > > diff --git a/fs/notify/fsnotify.h b/fs/notify/fsnotify.h
> > > > index 8b73ad45cc71..378f9ec6d64b 100644
> > > > --- a/fs/notify/fsnotify.h
> > > > +++ b/fs/notify/fsnotify.h
> > > > @@ -53,6 +53,13 @@ static inline struct super_block *fsnotify_connector_sb(
> > > > return fsnotify_object_sb(conn->obj, conn->type);
> > > > }
> > > >
> > > > +static inline fsnotify_connp_t *fsnotify_sb_marks(struct super_block *sb)
> > > > +{
> > > > + struct fsnotify_sb_info *sbinfo = fsnotify_sb_info(sb);
> > > > +
> > > > + return sbinfo ? &sbinfo->sb_marks : NULL;
> > > > +}
> > > > +
> > > > /* destroy all events sitting in this groups notification queue */
> > > > extern void fsnotify_flush_notify(struct fsnotify_group *group);
> > > >
> > > > @@ -78,7 +85,7 @@ static inline void fsnotify_clear_marks_by_mount(struct vfsmount *mnt)
> > > > /* run the list of all marks associated with sb and destroy them */
> > > > static inline void fsnotify_clear_marks_by_sb(struct super_block *sb)
> > > > {
> > > > - fsnotify_destroy_marks(&sb->s_fsnotify_marks);
> > > > + fsnotify_destroy_marks(fsnotify_sb_marks(sb));
> > > > }
> > > >
> > > > /*
> > > > diff --git a/fs/notify/mark.c b/fs/notify/mark.c
> > > > index 0b703f9e6344..db053e0e218d 100644
> > > > --- a/fs/notify/mark.c
> > > > +++ b/fs/notify/mark.c
> > > > @@ -105,7 +105,7 @@ static fsnotify_connp_t *fsnotify_object_connp(void *obj, int obj_type)
> > > > case FSNOTIFY_OBJ_TYPE_VFSMOUNT:
> > > > return &real_mount(obj)->mnt_fsnotify_marks;
> > > > case FSNOTIFY_OBJ_TYPE_SB:
> > > > - return &((struct super_block *)obj)->s_fsnotify_marks;
> > > > + return fsnotify_sb_marks(obj);
> > > > default:
> > > > return NULL;
> > > > }
> > > > @@ -568,6 +568,26 @@ int fsnotify_compare_groups(struct fsnotify_group *a, struct fsnotify_group *b)
> > > > return -1;
> > > > }
> > > >
> > > > +static int fsnotify_attach_info_to_sb(struct super_block *sb)
> > > > +{
> > > > + struct fsnotify_sb_info *sbinfo;
> > > > +
> > > > + /* sb info is freed on fsnotify_sb_delete() */
> > > > + sbinfo = kzalloc(sizeof(*sbinfo), GFP_KERNEL);
> > > > + if (!sbinfo)
> > > > + return -ENOMEM;
> > > > +
> > > > + /*
> > > > + * cmpxchg() provides the barrier so that callers of fsnotify_sb_info()
> > > > + * will observe an initialized structure
> > > > + */
> > > > + if (cmpxchg(&sb->s_fsnotify_info, NULL, sbinfo)) {
> > > > + /* Someone else created sbinfo for us */
> > > > + kfree(sbinfo);
> > > > + }
> > >
> > > Alternatively, you could consider using wait_var_event() to let
> > > concurrent attachers wait for s_fsnotify_info to be initialized using a
> > > sentinel value to indicate that the caller should wait. But not sure if
> > > it's worth it.
> >
> > Not worth it IMO. Adding watches is an extremely rare event
> > in the grand picture.
>
> Agreed. The cmpxchg() scheme has generally proven to be good enough in
> similar situations and simple enough to understand...
Thanks, sounds good to me.
