[PATCH v2 0/2] More fsnotify hook optimizations

[PATCH v2 0/2] More fsnotify hook optimizations

[Amir Goldstein]

Jan,

Following v2 addresses your review comments on v1 [1].

Changes since v1:
- Raname macro to FMODE_NOTIFY_ACCESS_PERM()
- Remove unneeded and unused event set macros

[1] https://lore.kernel.org/linux-fsdevel/20250707170704.303772-1-amir73il@gmail.com/

Amir Goldstein (2):
  fsnotify: merge file_set_fsnotify_mode_from_watchers() with open perm
    hook
  fsnotify: optimize FMODE_NONOTIFY_PERM for the common cases

 fs/file_table.c          |  2 +-
 fs/notify/fsnotify.c     | 87 ++++++++++++++++++++++++----------------
 fs/open.c                |  6 +--
 include/linux/fs.h       | 12 +++---
 include/linux/fsnotify.h | 35 +++-------------
 5 files changed, 68 insertions(+), 74 deletions(-)

-- 
2.43.0

[PATCH v2 1/2] fsnotify: merge file_set_fsnotify_mode_from_watchers() with open perm hook

[Amir Goldstein]

Create helper fsnotify_open_perm_and_set_mode() that moves the
fsnotify_open_perm() hook into file_set_fsnotify_mode_from_watchers().

This will allow some more optimizations.

Signed-off-by: Amir Goldstein <amir73il@gmail.com>
---
 fs/file_table.c          |  2 +-
 fs/notify/fsnotify.c     | 22 +++++++++++++---------
 fs/open.c                |  6 +++---
 include/linux/fsnotify.h |  8 +++-----
 4 files changed, 20 insertions(+), 18 deletions(-)

diff --git a/fs/file_table.c b/fs/file_table.c
index f09d79a98111..81c72576e548 100644
--- a/fs/file_table.c
+++ b/fs/file_table.c
@@ -199,7 +199,7 @@ static int init_file(struct file *f, int flags, const struct cred *cred)
 	file_ref_init(&f->f_ref, 1);
 	/*
 	 * Disable permission and pre-content events for all files by default.
-	 * They may be enabled later by file_set_fsnotify_mode_from_watchers().
+	 * They may be enabled later by fsnotify_open_perm_and_set_mode().
 	 */
 	file_set_fsnotify_mode(f, FMODE_NONOTIFY_PERM);
 	return 0;
diff --git a/fs/notify/fsnotify.c b/fs/notify/fsnotify.c
index e2b4f17a48bb..de7e7425428b 100644
--- a/fs/notify/fsnotify.c
+++ b/fs/notify/fsnotify.c
@@ -656,12 +656,12 @@ EXPORT_SYMBOL_GPL(fsnotify);
 
 #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS
 /*
- * At open time we check fsnotify_sb_has_priority_watchers() and set the
- * FMODE_NONOTIFY_ mode bits accordignly.
+ * At open time we check fsnotify_sb_has_priority_watchers(), call the open perm
+ * hook and set the FMODE_NONOTIFY_ mode bits accordignly.
  * Later, fsnotify permission hooks do not check if there are permission event
  * watches, but that there were permission event watches at open time.
  */
-void file_set_fsnotify_mode_from_watchers(struct file *file)
+int fsnotify_open_perm_and_set_mode(struct file *file)
 {
 	struct dentry *dentry = file->f_path.dentry, *parent;
 	struct super_block *sb = dentry->d_sb;
@@ -669,7 +669,7 @@ void file_set_fsnotify_mode_from_watchers(struct file *file)
 
 	/* Is it a file opened by fanotify? */
 	if (FMODE_FSNOTIFY_NONE(file->f_mode))
-		return;
+		return 0;
 
 	/*
 	 * Permission events is a super set of pre-content events, so if there
@@ -679,7 +679,7 @@ void file_set_fsnotify_mode_from_watchers(struct file *file)
 	if (likely(!fsnotify_sb_has_priority_watchers(sb,
 						FSNOTIFY_PRIO_CONTENT))) {
 		file_set_fsnotify_mode(file, FMODE_NONOTIFY_PERM);
-		return;
+		return 0;
 	}
 
 	/*
@@ -689,8 +689,9 @@ void file_set_fsnotify_mode_from_watchers(struct file *file)
 	if ((!d_is_dir(dentry) && !d_is_reg(dentry)) ||
 	    likely(!fsnotify_sb_has_priority_watchers(sb,
 						FSNOTIFY_PRIO_PRE_CONTENT))) {
-		file_set_fsnotify_mode(file, FMODE_NONOTIFY | FMODE_NONOTIFY_PERM);
-		return;
+		file_set_fsnotify_mode(file, FMODE_NONOTIFY |
+				       FMODE_NONOTIFY_PERM);
+		goto open_perm;
 	}
 
 	/*
@@ -702,7 +703,7 @@ void file_set_fsnotify_mode_from_watchers(struct file *file)
 				     FSNOTIFY_PRE_CONTENT_EVENTS))) {
 		/* Enable pre-content events */
 		file_set_fsnotify_mode(file, 0);
-		return;
+		goto open_perm;
 	}
 
 	/* Is parent watching for pre-content events on this file? */
@@ -713,11 +714,14 @@ void file_set_fsnotify_mode_from_watchers(struct file *file)
 		if (p_mask & FSNOTIFY_PRE_CONTENT_EVENTS) {
 			/* Enable pre-content events */
 			file_set_fsnotify_mode(file, 0);
-			return;
+			goto open_perm;
 		}
 	}
 	/* Nobody watching for pre-content events from this file */
 	file_set_fsnotify_mode(file, FMODE_NONOTIFY | FMODE_NONOTIFY_PERM);
+
+open_perm:
+	return fsnotify_open_perm(file);
 }
 #endif
 
diff --git a/fs/open.c b/fs/open.c
index 7828234a7caa..f240b96ce586 100644
--- a/fs/open.c
+++ b/fs/open.c
@@ -943,12 +943,12 @@ static int do_dentry_open(struct file *f,
 		goto cleanup_all;
 
 	/*
-	 * Set FMODE_NONOTIFY_* bits according to existing permission watches.
+	 * Call fsnotify open permission hook and set FMODE_NONOTIFY_* bits
+	 * according to existing permission watches.
 	 * If FMODE_NONOTIFY mode was already set for an fanotify fd or for a
 	 * pseudo file, this call will not change the mode.
 	 */
-	file_set_fsnotify_mode_from_watchers(f);
-	error = fsnotify_open_perm(f);
+	error = fsnotify_open_perm_and_set_mode(f);
 	if (error)
 		goto cleanup_all;
 
diff --git a/include/linux/fsnotify.h b/include/linux/fsnotify.h
index 454d8e466958..8c1fa617d375 100644
--- a/include/linux/fsnotify.h
+++ b/include/linux/fsnotify.h
@@ -129,7 +129,7 @@ static inline int fsnotify_file(struct file *file, __u32 mask)
 
 #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS
 
-void file_set_fsnotify_mode_from_watchers(struct file *file);
+int fsnotify_open_perm_and_set_mode(struct file *file);
 
 /*
  * fsnotify_file_area_perm - permission hook before access to file range
@@ -215,9 +215,6 @@ static inline int fsnotify_open_perm(struct file *file)
 {
 	int ret;
 
-	if (likely(!FMODE_FSNOTIFY_PERM(file->f_mode)))
-		return 0;
-
 	if (file->f_flags & __FMODE_EXEC) {
 		ret = fsnotify_path(&file->f_path, FS_OPEN_EXEC_PERM);
 		if (ret)
@@ -228,8 +225,9 @@ static inline int fsnotify_open_perm(struct file *file)
 }
 
 #else
-static inline void file_set_fsnotify_mode_from_watchers(struct file *file)
+static inline int fsnotify_open_perm_and_set_mode(struct file *file)
 {
+	return 0;
 }
 
 static inline int fsnotify_file_area_perm(struct file *file, int perm_mask,
-- 
2.43.0

[PATCH v2 2/2] fsnotify: optimize FMODE_NONOTIFY_PERM for the common cases

[Amir Goldstein]

The most unlikely watched permission event is FAN_ACCESS_PERM, because
at the time that it was introduced there were no evictable ignore mark,
so subscribing to FAN_ACCESS_PERM would have incured a very high
overhead.

Yet, when we set the fmode to FMODE_NOTIFY_HSM(), we never skip trying
to send FAN_ACCESS_PERM, which is almost always a waste of cycles.

We got to this logic because of bundling open permisson events and access
permission events in the same category and because FAN_OPEN_PERM is a
commonly used event.

By open coding fsnotify_open_perm() in fsnotify_open_perm_and_set_mode(),
we no longer need to regard FAN_OPEN*_PERM when calculating fmode.

This leaves the case of having pre-content events and not having access
permission events in the object masks a more likely case than the other
way around.

Rework the fmode macros and code so that their meaning now refers only
to hooks on an already open file:

- FMODE_NOTIFY_NONE()		skip all events
- FMODE_NOTIFY_ACCESS_PERM()	send all access permission events
- FMODE_NOTIFY_HSM()		send pre-conent permission events

Signed-off-by: Amir Goldstein <amir73il@gmail.com>
---
 fs/notify/fsnotify.c     | 75 ++++++++++++++++++++++++----------------
 include/linux/fs.h       | 12 +++----
 include/linux/fsnotify.h | 27 ++-------------
 3 files changed, 53 insertions(+), 61 deletions(-)

diff --git a/fs/notify/fsnotify.c b/fs/notify/fsnotify.c
index de7e7425428b..079b868552c2 100644
--- a/fs/notify/fsnotify.c
+++ b/fs/notify/fsnotify.c
@@ -199,8 +199,8 @@ static bool fsnotify_event_needs_parent(struct inode *inode, __u32 mnt_mask,
 }
 
 /* Are there any inode/mount/sb objects that watch for these events? */
-static inline bool fsnotify_object_watched(struct inode *inode, __u32 mnt_mask,
-					   __u32 mask)
+static inline __u32 fsnotify_object_watched(struct inode *inode, __u32 mnt_mask,
+					    __u32 mask)
 {
 	__u32 marks_mask = READ_ONCE(inode->i_fsnotify_mask) | mnt_mask |
 			   READ_ONCE(inode->i_sb->s_fsnotify_mask);
@@ -665,7 +665,7 @@ int fsnotify_open_perm_and_set_mode(struct file *file)
 {
 	struct dentry *dentry = file->f_path.dentry, *parent;
 	struct super_block *sb = dentry->d_sb;
-	__u32 mnt_mask, p_mask;
+	__u32 mnt_mask, p_mask = 0;
 
 	/* Is it a file opened by fanotify? */
 	if (FMODE_FSNOTIFY_NONE(file->f_mode))
@@ -683,45 +683,60 @@ int fsnotify_open_perm_and_set_mode(struct file *file)
 	}
 
 	/*
-	 * If there are permission event watchers but no pre-content event
-	 * watchers, set FMODE_NONOTIFY | FMODE_NONOTIFY_PERM to indicate that.
+	 * OK, there are some permission event watchers. Check if anybody is
+	 * watching for permission events on *this* file.
 	 */
-	if ((!d_is_dir(dentry) && !d_is_reg(dentry)) ||
-	    likely(!fsnotify_sb_has_priority_watchers(sb,
-						FSNOTIFY_PRIO_PRE_CONTENT))) {
-		file_set_fsnotify_mode(file, FMODE_NONOTIFY |
-				       FMODE_NONOTIFY_PERM);
-		goto open_perm;
+	mnt_mask = READ_ONCE(real_mount(file->f_path.mnt)->mnt_fsnotify_mask);
+	p_mask = fsnotify_object_watched(d_inode(dentry), mnt_mask,
+					 ALL_FSNOTIFY_PERM_EVENTS);
+	if (dentry->d_flags & DCACHE_FSNOTIFY_PARENT_WATCHED) {
+		parent = dget_parent(dentry);
+		p_mask |= fsnotify_inode_watches_children(d_inode(parent));
+		dput(parent);
 	}
 
 	/*
-	 * OK, there are some pre-content watchers. Check if anybody is
-	 * watching for pre-content events on *this* file.
+	 * Legacy FAN_ACCESS_PERM events have very high performance overhead,
+	 * so unlikely to be used in the wild. If they are used there will be
+	 * no optimizations at all.
 	 */
-	mnt_mask = READ_ONCE(real_mount(file->f_path.mnt)->mnt_fsnotify_mask);
-	if (unlikely(fsnotify_object_watched(d_inode(dentry), mnt_mask,
-				     FSNOTIFY_PRE_CONTENT_EVENTS))) {
-		/* Enable pre-content events */
+	if (unlikely(p_mask & FS_ACCESS_PERM)) {
+		/* Enable all permission and pre-content events */
 		file_set_fsnotify_mode(file, 0);
 		goto open_perm;
 	}
 
-	/* Is parent watching for pre-content events on this file? */
-	if (dentry->d_flags & DCACHE_FSNOTIFY_PARENT_WATCHED) {
-		parent = dget_parent(dentry);
-		p_mask = fsnotify_inode_watches_children(d_inode(parent));
-		dput(parent);
-		if (p_mask & FSNOTIFY_PRE_CONTENT_EVENTS) {
-			/* Enable pre-content events */
-			file_set_fsnotify_mode(file, 0);
-			goto open_perm;
-		}
+	/*
+	 * Pre-content events are only supported on regular files.
+	 * If there are pre-content event watchers and no permission access
+	 * watchers, set FMODE_NONOTIFY | FMODE_NONOTIFY_PERM to indicate that.
+	 * That is the common case with HSM service.
+	 */
+	if (d_is_reg(dentry) && (p_mask & FSNOTIFY_PRE_CONTENT_EVENTS)) {
+		file_set_fsnotify_mode(file, FMODE_NONOTIFY |
+					     FMODE_NONOTIFY_PERM);
+		goto open_perm;
 	}
-	/* Nobody watching for pre-content events from this file */
-	file_set_fsnotify_mode(file, FMODE_NONOTIFY | FMODE_NONOTIFY_PERM);
+
+	/* Nobody watching permission and pre-content events on this file */
+	file_set_fsnotify_mode(file, FMODE_NONOTIFY_PERM);
 
 open_perm:
-	return fsnotify_open_perm(file);
+	/*
+	 * Send open perm events depending on object masks and regardless of
+	 * FMODE_NONOTIFY_PERM.
+	 */
+	if (file->f_flags & __FMODE_EXEC && p_mask & FS_OPEN_EXEC_PERM) {
+		int ret = fsnotify_path(&file->f_path, FS_OPEN_EXEC_PERM);
+
+		if (ret)
+			return ret;
+	}
+
+	if (p_mask & FS_OPEN_PERM)
+		return fsnotify_path(&file->f_path, FS_OPEN_PERM);
+
+	return 0;
 }
 #endif
 
diff --git a/include/linux/fs.h b/include/linux/fs.h
index 3b074fe4c8a9..81a6463bfd76 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -200,12 +200,12 @@ typedef int (dio_iodone_t)(struct kiocb *iocb, loff_t offset,
 
 /*
  * The two FMODE_NONOTIFY* define which fsnotify events should not be generated
- * for a file. These are the possible values of (f->f_mode &
- * FMODE_FSNOTIFY_MASK) and their meaning:
+ * for an open file. These are the possible values of
+ * (f->f_mode & FMODE_FSNOTIFY_MASK) and their meaning:
  *
  * FMODE_NONOTIFY - suppress all (incl. non-permission) events.
  * FMODE_NONOTIFY_PERM - suppress permission (incl. pre-content) events.
- * FMODE_NONOTIFY | FMODE_NONOTIFY_PERM - suppress only pre-content events.
+ * FMODE_NONOTIFY | FMODE_NONOTIFY_PERM - suppress only FAN_ACCESS_PERM.
  */
 #define FMODE_FSNOTIFY_MASK \
 	(FMODE_NONOTIFY | FMODE_NONOTIFY_PERM)
@@ -213,13 +213,13 @@ typedef int (dio_iodone_t)(struct kiocb *iocb, loff_t offset,
 #define FMODE_FSNOTIFY_NONE(mode) \
 	((mode & FMODE_FSNOTIFY_MASK) == FMODE_NONOTIFY)
 #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS
-#define FMODE_FSNOTIFY_PERM(mode) \
+#define FMODE_FSNOTIFY_HSM(mode) \
 	((mode & FMODE_FSNOTIFY_MASK) == 0 || \
 	 (mode & FMODE_FSNOTIFY_MASK) == (FMODE_NONOTIFY | FMODE_NONOTIFY_PERM))
-#define FMODE_FSNOTIFY_HSM(mode) \
+#define FMODE_FSNOTIFY_ACCESS_PERM(mode) \
 	((mode & FMODE_FSNOTIFY_MASK) == 0)
 #else
-#define FMODE_FSNOTIFY_PERM(mode)	0
+#define FMODE_FSNOTIFY_ACCESS_PERM(mode) 0
 #define FMODE_FSNOTIFY_HSM(mode)	0
 #endif
 
diff --git a/include/linux/fsnotify.h b/include/linux/fsnotify.h
index 8c1fa617d375..28a9cb13fbfa 100644
--- a/include/linux/fsnotify.h
+++ b/include/linux/fsnotify.h
@@ -147,9 +147,6 @@ static inline int fsnotify_file_area_perm(struct file *file, int perm_mask,
 	if (!(perm_mask & (MAY_READ | MAY_WRITE | MAY_ACCESS)))
 		return 0;
 
-	if (likely(!FMODE_FSNOTIFY_PERM(file->f_mode)))
-		return 0;
-
 	/*
 	 * read()/write() and other types of access generate pre-content events.
 	 */
@@ -160,7 +157,8 @@ static inline int fsnotify_file_area_perm(struct file *file, int perm_mask,
 			return ret;
 	}
 
-	if (!(perm_mask & MAY_READ))
+	if (!(perm_mask & MAY_READ) ||
+	    likely(!FMODE_FSNOTIFY_ACCESS_PERM(file->f_mode)))
 		return 0;
 
 	/*
@@ -208,22 +206,6 @@ static inline int fsnotify_file_perm(struct file *file, int perm_mask)
 	return fsnotify_file_area_perm(file, perm_mask, NULL, 0);
 }
 
-/*
- * fsnotify_open_perm - permission hook before file open
- */
-static inline int fsnotify_open_perm(struct file *file)
-{
-	int ret;
-
-	if (file->f_flags & __FMODE_EXEC) {
-		ret = fsnotify_path(&file->f_path, FS_OPEN_EXEC_PERM);
-		if (ret)
-			return ret;
-	}
-
-	return fsnotify_path(&file->f_path, FS_OPEN_PERM);
-}
-
 #else
 static inline int fsnotify_open_perm_and_set_mode(struct file *file)
 {
@@ -251,11 +233,6 @@ static inline int fsnotify_file_perm(struct file *file, int perm_mask)
 {
 	return 0;
 }
-
-static inline int fsnotify_open_perm(struct file *file)
-{
-	return 0;
-}
 #endif
 
 /*
-- 
2.43.0

Re: [PATCH v2 0/2] More fsnotify hook optimizations

[Jan Kara]

Hello!

On Tue 08-07-25 16:36:39, Amir Goldstein wrote:
> Following v2 addresses your review comments on v1 [1].
> 
> Changes since v1:
> - Raname macro to FMODE_NOTIFY_ACCESS_PERM()
> - Remove unneeded and unused event set macros
> 
> [1] https://lore.kernel.org/linux-fsdevel/20250707170704.303772-1-amir73il@gmail.com/
> 
> Amir Goldstein (2):
>   fsnotify: merge file_set_fsnotify_mode_from_watchers() with open perm
>     hook
>   fsnotify: optimize FMODE_NONOTIFY_PERM for the common cases

Thanks! I've merged the patches to my tree (after fixing up some typos).

								Honza

> 
>  fs/file_table.c          |  2 +-
>  fs/notify/fsnotify.c     | 87 ++++++++++++++++++++++++----------------
>  fs/open.c                |  6 +--
>  include/linux/fs.h       | 12 +++---
>  include/linux/fsnotify.h | 35 +++-------------
>  5 files changed, 68 insertions(+), 74 deletions(-)
> 
> -- 
> 2.43.0
> 
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR