[v4,05/10] fsnotify: send MOVE_SELF event with parent/name info
diff mbox series

Message ID 20200702125744.10535-6-amir73il@gmail.com
State New
Headers show
Series
  • fanotify events with name info
Related show

Commit Message

Amir Goldstein July 2, 2020, 12:57 p.m. UTC
MOVE_SELF event does not get reported to a parent watching children
when a child is moved, but it can be reported to sb/mount mark with
parent/name info if group is interested in parent/name info.

Use the fsnotify_parent() helper to send a MOVE_SELF event and adjust
fsnotify() to handle the case of an event "on child" that should not
be sent to the watching parent's inode mark.

Signed-off-by: Amir Goldstein <amir73il@gmail.com>
---
 fs/notify/fsnotify.c             | 21 +++++++++++++++++----
 include/linux/fsnotify.h         |  5 +----
 include/linux/fsnotify_backend.h |  2 +-
 3 files changed, 19 insertions(+), 9 deletions(-)

Comments

Jan Kara July 14, 2020, 12:13 p.m. UTC | #1
On Thu 02-07-20 15:57:39, Amir Goldstein wrote:
> MOVE_SELF event does not get reported to a parent watching children
> when a child is moved, but it can be reported to sb/mount mark with
> parent/name info if group is interested in parent/name info.
> 
> Use the fsnotify_parent() helper to send a MOVE_SELF event and adjust
> fsnotify() to handle the case of an event "on child" that should not
> be sent to the watching parent's inode mark.
> 
> Signed-off-by: Amir Goldstein <amir73il@gmail.com>
> ---
>  fs/notify/fsnotify.c             | 21 +++++++++++++++++----
>  include/linux/fsnotify.h         |  5 +----
>  include/linux/fsnotify_backend.h |  2 +-
>  3 files changed, 19 insertions(+), 9 deletions(-)
> 
> diff --git a/fs/notify/fsnotify.c b/fs/notify/fsnotify.c
> index 6683c77a5b13..0faf5b09a73e 100644
> --- a/fs/notify/fsnotify.c
> +++ b/fs/notify/fsnotify.c
> @@ -352,6 +352,7 @@ int fsnotify(struct inode *to_tell, __u32 mask, const void *data, int data_type,
>  	struct super_block *sb = to_tell->i_sb;
>  	struct inode *dir = S_ISDIR(to_tell->i_mode) ? to_tell : NULL;
>  	struct mount *mnt = NULL;
> +	struct inode *inode = NULL;
>  	struct inode *child = NULL;
>  	int ret = 0;
>  	__u32 test_mask, marks_mask;
> @@ -362,6 +363,14 @@ int fsnotify(struct inode *to_tell, __u32 mask, const void *data, int data_type,
>  	if (mask & FS_EVENT_ON_CHILD)
>  		child = fsnotify_data_inode(data, data_type);
>  
> +	/*
> +	 * If event is "on child" then to_tell is a watching parent.
> +	 * An event "on child" may be sent to mount/sb mark with parent/name
> +	 * info, but not appropriate for watching parent (e.g. FS_MOVE_SELF).
> +	 */
> +	if (!child || (mask & FS_EVENTS_POSS_ON_CHILD))
> +		inode = to_tell;

I'm now confused. Don't you want to fill in FSNOTIFY_OBJ_TYPE_INODE below
for FS_MOVE_SELF event? But this condition is false for it so you won't do
it?

> +
>  	/*
>  	 * Optimization: srcu_read_lock() has a memory barrier which can
>  	 * be expensive.  It protects walking the *_fsnotify_marks lists.
> @@ -369,14 +378,17 @@ int fsnotify(struct inode *to_tell, __u32 mask, const void *data, int data_type,
>  	 * SRCU because we have no references to any objects and do not
>  	 * need SRCU to keep them "alive".
>  	 */
> -	if (!to_tell->i_fsnotify_marks && !sb->s_fsnotify_marks &&
> +	if (!sb->s_fsnotify_marks &&
>  	    (!mnt || !mnt->mnt_fsnotify_marks) &&
> +	    (!inode || !inode->i_fsnotify_marks) &&
>  	    (!child || !child->i_fsnotify_marks))
>  		return 0;
>  
> -	marks_mask = to_tell->i_fsnotify_mask | sb->s_fsnotify_mask;
> +	marks_mask = sb->s_fsnotify_mask;
>  	if (mnt)
>  		marks_mask |= mnt->mnt_fsnotify_mask;
> +	if (inode)
> +		marks_mask |= inode->i_fsnotify_mask;
>  	if (child)
>  		marks_mask |= child->i_fsnotify_mask;
>  
> @@ -390,14 +402,15 @@ int fsnotify(struct inode *to_tell, __u32 mask, const void *data, int data_type,
>  
>  	iter_info.srcu_idx = srcu_read_lock(&fsnotify_mark_srcu);
>  
> -	iter_info.marks[FSNOTIFY_OBJ_TYPE_INODE] =
> -		fsnotify_first_mark(&to_tell->i_fsnotify_marks);
>  	iter_info.marks[FSNOTIFY_OBJ_TYPE_SB] =
>  		fsnotify_first_mark(&sb->s_fsnotify_marks);
>  	if (mnt) {
>  		iter_info.marks[FSNOTIFY_OBJ_TYPE_VFSMOUNT] =
>  			fsnotify_first_mark(&mnt->mnt_fsnotify_marks);
>  	}
> +	if (inode)
> +		iter_info.marks[FSNOTIFY_OBJ_TYPE_INODE] =
> +			fsnotify_first_mark(&inode->i_fsnotify_marks);
>  	if (child) {
>  		iter_info.marks[FSNOTIFY_OBJ_TYPE_CHILD] =
>  			fsnotify_first_mark(&child->i_fsnotify_marks);
> diff --git a/include/linux/fsnotify.h b/include/linux/fsnotify.h
> index 044cae3a0628..61dccaf21e7b 100644
> --- a/include/linux/fsnotify.h
> +++ b/include/linux/fsnotify.h
> @@ -131,7 +131,6 @@ static inline void fsnotify_move(struct inode *old_dir, struct inode *new_dir,
>  	u32 fs_cookie = fsnotify_get_cookie();
>  	__u32 old_dir_mask = FS_MOVED_FROM;
>  	__u32 new_dir_mask = FS_MOVED_TO;
> -	__u32 mask = FS_MOVE_SELF;
>  	const struct qstr *new_name = &moved->d_name;
>  
>  	if (old_dir == new_dir)
> @@ -140,7 +139,6 @@ static inline void fsnotify_move(struct inode *old_dir, struct inode *new_dir,
>  	if (isdir) {
>  		old_dir_mask |= FS_ISDIR;
>  		new_dir_mask |= FS_ISDIR;
> -		mask |= FS_ISDIR;
>  	}
>  
>  	fsnotify_name(old_dir, old_dir_mask, source, old_name, fs_cookie);
> @@ -149,8 +147,7 @@ static inline void fsnotify_move(struct inode *old_dir, struct inode *new_dir,
>  	if (target)
>  		fsnotify_link_count(target);
>  
> -	if (source)
> -		fsnotify(source, mask, source, FSNOTIFY_EVENT_INODE, NULL, 0);
> +	fsnotify_dentry(moved, FS_MOVE_SELF);

I'm somewhat unsure about this. Does this mean that 'moved' is guaranteed
to be positive or that you've made sure that all the code below
fsnotify_dentry() is actually fine with a negative dentry? I don't find
either trivial to verify so some note in a changelog or maybe even a
separate patch for this would be useful.

								Honza
Amir Goldstein July 14, 2020, 12:44 p.m. UTC | #2
On Tue, Jul 14, 2020 at 3:13 PM Jan Kara <jack@suse.cz> wrote:
>
> On Thu 02-07-20 15:57:39, Amir Goldstein wrote:
> > MOVE_SELF event does not get reported to a parent watching children
> > when a child is moved, but it can be reported to sb/mount mark with
> > parent/name info if group is interested in parent/name info.
> >
> > Use the fsnotify_parent() helper to send a MOVE_SELF event and adjust
> > fsnotify() to handle the case of an event "on child" that should not
> > be sent to the watching parent's inode mark.
> >
> > Signed-off-by: Amir Goldstein <amir73il@gmail.com>
> > ---
> >  fs/notify/fsnotify.c             | 21 +++++++++++++++++----
> >  include/linux/fsnotify.h         |  5 +----
> >  include/linux/fsnotify_backend.h |  2 +-
> >  3 files changed, 19 insertions(+), 9 deletions(-)
> >
> > diff --git a/fs/notify/fsnotify.c b/fs/notify/fsnotify.c
> > index 6683c77a5b13..0faf5b09a73e 100644
> > --- a/fs/notify/fsnotify.c
> > +++ b/fs/notify/fsnotify.c
> > @@ -352,6 +352,7 @@ int fsnotify(struct inode *to_tell, __u32 mask, const void *data, int data_type,
> >       struct super_block *sb = to_tell->i_sb;
> >       struct inode *dir = S_ISDIR(to_tell->i_mode) ? to_tell : NULL;
> >       struct mount *mnt = NULL;
> > +     struct inode *inode = NULL;
> >       struct inode *child = NULL;
> >       int ret = 0;
> >       __u32 test_mask, marks_mask;
> > @@ -362,6 +363,14 @@ int fsnotify(struct inode *to_tell, __u32 mask, const void *data, int data_type,
> >       if (mask & FS_EVENT_ON_CHILD)
> >               child = fsnotify_data_inode(data, data_type);
> >
> > +     /*
> > +      * If event is "on child" then to_tell is a watching parent.
> > +      * An event "on child" may be sent to mount/sb mark with parent/name
> > +      * info, but not appropriate for watching parent (e.g. FS_MOVE_SELF).
> > +      */
> > +     if (!child || (mask & FS_EVENTS_POSS_ON_CHILD))
> > +             inode = to_tell;
>
> I'm now confused. Don't you want to fill in FSNOTIFY_OBJ_TYPE_INODE below
> for FS_MOVE_SELF event? But this condition is false for it so you won't do
> it?
>

I do not.
For events with the flag FS_EVENT_ON_CHILD, the inode in
FSNOTIFY_OBJ_TYPE_INODE is always the parent and the inode in
FSNOTIFY_OBJ_TYPE_CHILD is always the child.
So FS_MOVE_SELF will be reported if sb/mount are watching
or if child inode is watching, but NOT if only parent inode is watching.

I realize I may have been able to make other choices, but seemed like
the most consistent choice to me.
If you see a better option, let me know.

> > +
> >       /*
> >        * Optimization: srcu_read_lock() has a memory barrier which can
> >        * be expensive.  It protects walking the *_fsnotify_marks lists.
> > @@ -369,14 +378,17 @@ int fsnotify(struct inode *to_tell, __u32 mask, const void *data, int data_type,
> >        * SRCU because we have no references to any objects and do not
> >        * need SRCU to keep them "alive".
> >        */
> > -     if (!to_tell->i_fsnotify_marks && !sb->s_fsnotify_marks &&
> > +     if (!sb->s_fsnotify_marks &&
> >           (!mnt || !mnt->mnt_fsnotify_marks) &&
> > +         (!inode || !inode->i_fsnotify_marks) &&
> >           (!child || !child->i_fsnotify_marks))
> >               return 0;
> >
> > -     marks_mask = to_tell->i_fsnotify_mask | sb->s_fsnotify_mask;
> > +     marks_mask = sb->s_fsnotify_mask;
> >       if (mnt)
> >               marks_mask |= mnt->mnt_fsnotify_mask;
> > +     if (inode)
> > +             marks_mask |= inode->i_fsnotify_mask;
> >       if (child)
> >               marks_mask |= child->i_fsnotify_mask;
> >
> > @@ -390,14 +402,15 @@ int fsnotify(struct inode *to_tell, __u32 mask, const void *data, int data_type,
> >
> >       iter_info.srcu_idx = srcu_read_lock(&fsnotify_mark_srcu);
> >
> > -     iter_info.marks[FSNOTIFY_OBJ_TYPE_INODE] =
> > -             fsnotify_first_mark(&to_tell->i_fsnotify_marks);
> >       iter_info.marks[FSNOTIFY_OBJ_TYPE_SB] =
> >               fsnotify_first_mark(&sb->s_fsnotify_marks);
> >       if (mnt) {
> >               iter_info.marks[FSNOTIFY_OBJ_TYPE_VFSMOUNT] =
> >                       fsnotify_first_mark(&mnt->mnt_fsnotify_marks);
> >       }
> > +     if (inode)
> > +             iter_info.marks[FSNOTIFY_OBJ_TYPE_INODE] =
> > +                     fsnotify_first_mark(&inode->i_fsnotify_marks);
> >       if (child) {
> >               iter_info.marks[FSNOTIFY_OBJ_TYPE_CHILD] =
> >                       fsnotify_first_mark(&child->i_fsnotify_marks);
> > diff --git a/include/linux/fsnotify.h b/include/linux/fsnotify.h
> > index 044cae3a0628..61dccaf21e7b 100644
> > --- a/include/linux/fsnotify.h
> > +++ b/include/linux/fsnotify.h
> > @@ -131,7 +131,6 @@ static inline void fsnotify_move(struct inode *old_dir, struct inode *new_dir,
> >       u32 fs_cookie = fsnotify_get_cookie();
> >       __u32 old_dir_mask = FS_MOVED_FROM;
> >       __u32 new_dir_mask = FS_MOVED_TO;
> > -     __u32 mask = FS_MOVE_SELF;
> >       const struct qstr *new_name = &moved->d_name;
> >
> >       if (old_dir == new_dir)
> > @@ -140,7 +139,6 @@ static inline void fsnotify_move(struct inode *old_dir, struct inode *new_dir,
> >       if (isdir) {
> >               old_dir_mask |= FS_ISDIR;
> >               new_dir_mask |= FS_ISDIR;
> > -             mask |= FS_ISDIR;
> >       }
> >
> >       fsnotify_name(old_dir, old_dir_mask, source, old_name, fs_cookie);
> > @@ -149,8 +147,7 @@ static inline void fsnotify_move(struct inode *old_dir, struct inode *new_dir,
> >       if (target)
> >               fsnotify_link_count(target);
> >
> > -     if (source)
> > -             fsnotify(source, mask, source, FSNOTIFY_EVENT_INODE, NULL, 0);
> > +     fsnotify_dentry(moved, FS_MOVE_SELF);
>
> I'm somewhat unsure about this. Does this mean that 'moved' is guaranteed
> to be positive or that you've made sure that all the code below
> fsnotify_dentry() is actually fine with a negative dentry? I don't find
> either trivial to verify so some note in a changelog or maybe even a
> separate patch for this would be useful.
>

Oh, it's true. I should've mentioned it or separate this change.
I guess my reaction was the opposite of yours - it seemed obvious to
me that moved
dentry is positive - vfs_rename() is called under lock_rename() and it seemed
obvious to me that callers verified positive source, but in any case,
vfs_rename()
starts with may_delete() that verifies positive rename victim and
debugfs_rename(), the other caller of fsnotify_move() also verifies
positive victim.

I will add this information to the commit message.

Thanks,
Amir.

Patch
diff mbox series

diff --git a/fs/notify/fsnotify.c b/fs/notify/fsnotify.c
index 6683c77a5b13..0faf5b09a73e 100644
--- a/fs/notify/fsnotify.c
+++ b/fs/notify/fsnotify.c
@@ -352,6 +352,7 @@  int fsnotify(struct inode *to_tell, __u32 mask, const void *data, int data_type,
 	struct super_block *sb = to_tell->i_sb;
 	struct inode *dir = S_ISDIR(to_tell->i_mode) ? to_tell : NULL;
 	struct mount *mnt = NULL;
+	struct inode *inode = NULL;
 	struct inode *child = NULL;
 	int ret = 0;
 	__u32 test_mask, marks_mask;
@@ -362,6 +363,14 @@  int fsnotify(struct inode *to_tell, __u32 mask, const void *data, int data_type,
 	if (mask & FS_EVENT_ON_CHILD)
 		child = fsnotify_data_inode(data, data_type);
 
+	/*
+	 * If event is "on child" then to_tell is a watching parent.
+	 * An event "on child" may be sent to mount/sb mark with parent/name
+	 * info, but not appropriate for watching parent (e.g. FS_MOVE_SELF).
+	 */
+	if (!child || (mask & FS_EVENTS_POSS_ON_CHILD))
+		inode = to_tell;
+
 	/*
 	 * Optimization: srcu_read_lock() has a memory barrier which can
 	 * be expensive.  It protects walking the *_fsnotify_marks lists.
@@ -369,14 +378,17 @@  int fsnotify(struct inode *to_tell, __u32 mask, const void *data, int data_type,
 	 * SRCU because we have no references to any objects and do not
 	 * need SRCU to keep them "alive".
 	 */
-	if (!to_tell->i_fsnotify_marks && !sb->s_fsnotify_marks &&
+	if (!sb->s_fsnotify_marks &&
 	    (!mnt || !mnt->mnt_fsnotify_marks) &&
+	    (!inode || !inode->i_fsnotify_marks) &&
 	    (!child || !child->i_fsnotify_marks))
 		return 0;
 
-	marks_mask = to_tell->i_fsnotify_mask | sb->s_fsnotify_mask;
+	marks_mask = sb->s_fsnotify_mask;
 	if (mnt)
 		marks_mask |= mnt->mnt_fsnotify_mask;
+	if (inode)
+		marks_mask |= inode->i_fsnotify_mask;
 	if (child)
 		marks_mask |= child->i_fsnotify_mask;
 
@@ -390,14 +402,15 @@  int fsnotify(struct inode *to_tell, __u32 mask, const void *data, int data_type,
 
 	iter_info.srcu_idx = srcu_read_lock(&fsnotify_mark_srcu);
 
-	iter_info.marks[FSNOTIFY_OBJ_TYPE_INODE] =
-		fsnotify_first_mark(&to_tell->i_fsnotify_marks);
 	iter_info.marks[FSNOTIFY_OBJ_TYPE_SB] =
 		fsnotify_first_mark(&sb->s_fsnotify_marks);
 	if (mnt) {
 		iter_info.marks[FSNOTIFY_OBJ_TYPE_VFSMOUNT] =
 			fsnotify_first_mark(&mnt->mnt_fsnotify_marks);
 	}
+	if (inode)
+		iter_info.marks[FSNOTIFY_OBJ_TYPE_INODE] =
+			fsnotify_first_mark(&inode->i_fsnotify_marks);
 	if (child) {
 		iter_info.marks[FSNOTIFY_OBJ_TYPE_CHILD] =
 			fsnotify_first_mark(&child->i_fsnotify_marks);
diff --git a/include/linux/fsnotify.h b/include/linux/fsnotify.h
index 044cae3a0628..61dccaf21e7b 100644
--- a/include/linux/fsnotify.h
+++ b/include/linux/fsnotify.h
@@ -131,7 +131,6 @@  static inline void fsnotify_move(struct inode *old_dir, struct inode *new_dir,
 	u32 fs_cookie = fsnotify_get_cookie();
 	__u32 old_dir_mask = FS_MOVED_FROM;
 	__u32 new_dir_mask = FS_MOVED_TO;
-	__u32 mask = FS_MOVE_SELF;
 	const struct qstr *new_name = &moved->d_name;
 
 	if (old_dir == new_dir)
@@ -140,7 +139,6 @@  static inline void fsnotify_move(struct inode *old_dir, struct inode *new_dir,
 	if (isdir) {
 		old_dir_mask |= FS_ISDIR;
 		new_dir_mask |= FS_ISDIR;
-		mask |= FS_ISDIR;
 	}
 
 	fsnotify_name(old_dir, old_dir_mask, source, old_name, fs_cookie);
@@ -149,8 +147,7 @@  static inline void fsnotify_move(struct inode *old_dir, struct inode *new_dir,
 	if (target)
 		fsnotify_link_count(target);
 
-	if (source)
-		fsnotify(source, mask, source, FSNOTIFY_EVENT_INODE, NULL, 0);
+	fsnotify_dentry(moved, FS_MOVE_SELF);
 	audit_inode_child(new_dir, moved, AUDIT_TYPE_CHILD_CREATE);
 }
 
diff --git a/include/linux/fsnotify_backend.h b/include/linux/fsnotify_backend.h
index a7363c33211e..28f6cf704875 100644
--- a/include/linux/fsnotify_backend.h
+++ b/include/linux/fsnotify_backend.h
@@ -89,7 +89,7 @@ 
  * It may include events that can be sent to an inode/sb/mount mark, but cannot
  * be sent to a parent watching children.
  */
-#define FS_EVENTS_POSS_TO_PARENT (FS_EVENTS_POSS_ON_CHILD)
+#define FS_EVENTS_POSS_TO_PARENT (FS_EVENTS_POSS_ON_CHILD | FS_MOVE_SELF)
 
 /* Events that can be reported to backends */
 #define ALL_FSNOTIFY_EVENTS (ALL_FSNOTIFY_DIRENT_EVENTS | \