Re: [PATCH v2 2/3] fsnotify: Protect i_fsnotify_mask and child flags with inode rwsem

[Amir Goldstein]

On Fri, Oct 21, 2022 at 12:17 PM Christian Brauner <brauner@xxxxxxxxxx> wrote:
>
> On Thu, Oct 20, 2022 at 06:03:09PM -0700, Stephen Brennan wrote:
> > When an inode is interested in events on its children, it must set
> > DCACHE_FSNOTIFY_PARENT_WATCHED flag on all its children. Currently, when
> > the fsnotify connector is removed and i_fsnotify_mask becomes zero, we
> > lazily allow __fsnotify_parent() to do this the next time we see an
> > event on a child.
> >
> > However, if the list of children is very long (e.g., in the millions),
> > and lots of activity is occurring on the directory, then it's possible
> > for many CPUs to end up blocked on the inode spinlock in
> > __fsnotify_update_child_flags(). Each CPU will then redundantly iterate
> > over the very long list of children. This situation can cause soft
> > lockups.
> >
> > To avoid this, stop lazily updating child flags in __fsnotify_parent().
> > Instead, update flags when we disconnect a mark connector. Remember the
> > state of the children flags in the fsnotify_mark_connector flags.
> > Provide mutual exclusion by holding i_rwsem exclusive while we update
> > children, and use the cached state to avoid updating flags
> > unnecessarily.
> >
> > Signed-off-by: Stephen Brennan <stephen.s.brennan@xxxxxxxxxx>
> > ---
> >
> >  fs/notify/fsnotify.c             |  22 ++++++-
> >  fs/notify/fsnotify.h             |  31 ++++++++-
> >  fs/notify/mark.c                 | 106 ++++++++++++++++++++-----------
> >  include/linux/fsnotify_backend.h |   8 +++
> >  4 files changed, 127 insertions(+), 40 deletions(-)
> >
> > diff --git a/fs/notify/fsnotify.c b/fs/notify/fsnotify.c
> > index 6c338322f0c3..f83eca4fb841 100644
> > --- a/fs/notify/fsnotify.c
> > +++ b/fs/notify/fsnotify.c
> > @@ -103,13 +103,15 @@ void fsnotify_sb_delete(struct super_block *sb)
> >   * parent cares.  Thus when an event happens on a child it can quickly tell
> >   * if there is a need to find a parent and send the event to the parent.
> >   */
> > -void __fsnotify_update_child_dentry_flags(struct inode *inode)
> > +bool __fsnotify_update_children_dentry_flags(struct inode *inode)
> >  {
> >       struct dentry *alias, *child;
> >       int watched;
> >
> >       if (!S_ISDIR(inode->i_mode))
> > -             return;
> > +             return false;
> > +
> > +     lockdep_assert_held_write(&inode->i_rwsem);
> >
> >       /* determine if the children should tell inode about their events */
> >       watched = fsnotify_inode_watches_children(inode);
> > @@ -133,6 +135,20 @@ void __fsnotify_update_child_dentry_flags(struct inode *inode)
> >               spin_unlock(&child->d_lock);
> >       }
> >       spin_unlock(&alias->d_lock);
> > +     return watched;
> > +}
> > +
> > +void __fsnotify_update_child_dentry_flags(struct inode *inode, struct dentry *dentry)
> > +{
> > +     /*
> > +      * Flag would be cleared soon by
> > +      * __fsnotify_update_child_dentry_flags(), but as an
> > +      * optimization, clear it now.
> > +      */
> > +     spin_lock(&dentry->d_lock);
> > +     if (!fsnotify_inode_watches_children(inode))
> > +             dentry->d_flags &= ~DCACHE_FSNOTIFY_PARENT_WATCHED;
> > +     spin_unlock(&dentry->d_lock);
> >  }
> >
> >  /* Are inode/sb/mount interested in parent and name info with this event? */
> > @@ -203,7 +219,7 @@ int __fsnotify_parent(struct dentry *dentry, __u32 mask, const void *data,
> >       p_inode = parent->d_inode;
> >       p_mask = fsnotify_inode_watches_children(p_inode);
> >       if (unlikely(parent_watched && !p_mask))
> > -             __fsnotify_update_child_dentry_flags(p_inode);
> > +             __fsnotify_update_child_dentry_flags(p_inode, dentry);
> >
> >       /*
> >        * Include parent/name in notification either if some notification
> > diff --git a/fs/notify/fsnotify.h b/fs/notify/fsnotify.h
> > index fde74eb333cc..182d93014c6b 100644
> > --- a/fs/notify/fsnotify.h
> > +++ b/fs/notify/fsnotify.h
> > @@ -70,11 +70,40 @@ static inline void fsnotify_clear_marks_by_sb(struct super_block *sb)
> >       fsnotify_destroy_marks(&sb->s_fsnotify_marks);
> >  }
> >
> > +static inline bool fsnotify_children_need_update(struct fsnotify_mark_connector *conn,
> > +                                                 struct inode *inode)
> > +{
> > +     bool watched, flags_set;
> > +     watched = fsnotify_inode_watches_children(inode);
>
> nit: I'd leave a blank line after the variable declarations. Same for
> fsnotify_update_children_dentry_flags() below.
>
> > +     flags_set = conn->flags & FSNOTIFY_CONN_FLAG_WATCHES_CHILDREN;
> > +     return (watched && !flags_set) || (!watched && flags_set);
> > +}
> > +
> >  /*
> >   * update the dentry->d_flags of all of inode's children to indicate if inode cares
> >   * about events that happen to its children.
> >   */
> > -extern void __fsnotify_update_child_dentry_flags(struct inode *inode);
> > +extern bool __fsnotify_update_children_dentry_flags(struct inode *inode);
> > +
> > +static inline void fsnotify_update_children_dentry_flags(struct fsnotify_mark_connector *conn,
> > +                                                         struct inode *inode)
>
> Should that be a static inline function in a header seems a bit big. :)

I agree.
This helper has exactly one caller and should be placed right below it.

Thanks for spotting that,
Amir.