On Wed, Nov 28, 2018 at 2:59 PM Jan Kara <jack@xxxxxxx> wrote:
>
> On Sun 25-11-18 15:43:40, Amir Goldstein wrote:
> > "dirent" events are referring to events that modify directory entries,
> > such as create,delete,rename. Those events should always be reported
> > on a watched directory, regardless if FS_EVENT_ON_CHILD is set
> > on the watch mask.
> >
> > fsnotify_nameremove() and fsnotify_move() were modified to no longer
> > set the FS_EVENT_ON_CHILD event bit. This is a semantic change to
> > align with the "dirent" event definition. It has no effect on any
> > existing backend, because dnotify, inotify and audit always requets the
> > child events and fanotify does not get the delete,rename events.
> >
> > The fsnotify_dirent() helper is used instead of fsnotify_parent() to
> > report a dirent event to dentry->d_parent without FS_EVENT_ON_CHILD
> > and regardless if parent has the FS_EVENT_ON_CHILD bit set.
> >
> > ALL_FSNOTIFY_DIRENT_EVENTS defines all the dirent event types and
> > those event types have been extracted out of FS_EVENTS_POSS_ON_CHILD.
> >
> > That means for a directory with an inotify watch and only dirent
> > events in the mask (i.e. create,delete,move), all children dentries
> > will no longer have the DCACHE_FSNOTIFY_PARENT_WATCHED flag set.
> > This will allow all events that happen on children to be optimized
> > away in __fsnotify_parent() without the need to dereference
> > child->d_parent->d_inode->i_fsnotify_mask.
> >
> > Since the dirent events are never repoted via __fsnotify_parent(),
> > this results in no change of logic, but only an optimization.
> >
> > Signed-off-by: Amir Goldstein <amir73il@xxxxxxxxx>
> > ---
> > include/linux/fsnotify.h | 43 +++++++++++++++++++++++++-------
> > include/linux/fsnotify_backend.h | 36 +++++++++++++++-----------
> > 2 files changed, 55 insertions(+), 24 deletions(-)
>
> The patch looks good. Just one question and two nits below.
>
> > +static inline int fsnotify_dirent(struct inode *dir, struct dentry *dentry,
> > + __u32 mask)
> > +{
> > + struct dentry *parent = NULL;
> > + int ret;
> > +
> > + if (!dir) {
> > + parent = dget_parent(dentry);
> > + dir = d_inode(parent);
> > + }
> > +
> > + ret = fsnotify(dir, mask, d_inode(dentry), FSNOTIFY_EVENT_INODE,
> > + dentry->d_name.name, 0);
> > +
> > + dput(parent);
> > + return ret;
>
> There's one more feature fsnotify_parent() provides - it calls
> take_dentry_name_snapshot() before calling into fsnotify and uses
> snapshotted name. I think you need to do the same here to provide the same
> protection for fsnotify_nameremove, don't you? Or maybe you don't since
> generally directory i_rwsem is held when unlinking and so dentry name cannot
> change but then it would be good to assert that? Who knows what some weird
> fs is doing...
>
Right.. I will add that assert and at the same time, for the same reason, this
helper doesn't need to dget_parent(), because d_parent should also be stable.
> > +/*
> > + * This is a list of all events that may get sent to a parent based on fs event
>
> ^^^ Line too long.
Heh?? checkpatch did not complain. It's 79 chars.
>
> > +#define ALL_FSNOTIFY_EVENTS (ALL_FSNOTIFY_DIRENT_EVENTS| \
> ^^^ space here
>
My patch did not change indentation AFAIK.
The new definition added is with tab (ALL_FSNOTIFY_DIRENT_EVENTS)
but some of the existing definitions, even ones moved around are with space(s).
Do you want me to change that?
Thanks,
Amir.
