[RFC PATCH v2 0/3] io_uring: add IORING_OP_UTIMENSAT

[RFC PATCH v2 0/3] io_uring: add IORING_OP_UTIMENSAT

[Méven Car]

From: Méven Car <meven@kde.org>

A file manager or cp -p that copies many files sets the times of every
copy after its last write. KDE's KIO does not use io_uring yet, but
could in the future. In a KIO prototype that copies batches of small
files through io_uring, open, create, read and write run in the ring,
but the metadata of each copy is set outside the ring, with futimens()
and, when needed, fchmod() and fchown(). On NFS and SMB each of these
calls is a round trip to the server.

In that prototype, setting the metadata from a pool of 16 threads
instead of one file at a time cut the copy of 200 files of 4 KiB by 22%
on NFS and 45% on SMB. The test copied from ext4 to an NFS 4.2 and an
SMB mount of the same machine, over loopback with a netem delay of 2 ms
each way. An opcode would give the same parallelism from the ring, and
would let a copy be one linked chain per file: WRITE, UTIMENSAT, CLOSE.

Patch 1 splits filename_utimensat() out of do_utimes_path(), so that
io_uring can pass a name imported at submission time, as it does for
mkdirat and the other path ops.

Patch 2 adds IORING_OP_UTIMENSAT with the semantics of utimensat(2).
Like the other path ops it runs from io-wq.

Patch 3 is the RFC part. On a local filesystem the handoff to io-wq
costs more than the setattr: in a qemu guest on v7.2.3, setting the
times of 2000 files on tmpfs took 6-8 ms with utimensat(2) and 19-29 ms
through the ring with patch 2 alone, 32 requests per submit. The patch
tries the request inline with LOOKUP_CACHED and trylocks, on filesystems
that declare FS_SETATTR_TIMES_NOWAIT, which only tmpfs sets. The ring
then matches the syscall (6.4-6.8 ms). ext4, xfs and btrfs cannot set
the flag, because their ->setattr() starts a journal transaction. Is a
per-filesystem flag acceptable here, or would you rather keep patch 2
alone?

IORING_OP_FCHMOD was proposed in 2024 [1] and not merged. The same
questions apply to fchmod and fchown, which a copy also needs, and a
single setattr opcode could set the mode, owner and times in one
->setattr() call. I started with the times because a copy always sets
them.

Testing: a userspace test of 30 cases (paths, AT_EMPTY_PATH on O_PATH
fds, NULL path, UTIME_OMIT, permission and read-only errors, links, copy
at submission, inline on tmpfs and punt on ramfs) passes in a qemu
guest on for-7.4/io_uring. v1 also passed on v7.2.3. A liburing helper,
test and man page will follow if the interface is acceptable.

[1] https://lore.kernel.org/io-uring/e291085644e14b3eb4d1c3995098bf4e@huawei.com/

Changes in v2:
- A NULL path now sets the times through the fd itself, like
  do_utimes_fd(), instead of a lookup of an empty name. As with
  utimensat(2), it takes no flags (-EINVAL), an O_PATH fd gives -EBADF,
  and an open file that was unlinked needs no revalidation. (Sashiko)
- Patch 3: the nowait attempt passes no delegated_inode to
  notify_change(), so a delegation gives -EWOULDBLOCK. v1 read di_inode,
  which does not build with CONFIG_FILE_LOCKING=n. (Sashiko)
- Patch 2 message: drop the wrong claim that AT_EMPTY_PATH on an O_PATH
  fd differs from the syscall.
- Not changed: the inline lookup can still call ->d_weak_revalidate()
  and audit_inode() before __vfs_utimes() returns -EAGAIN for a
  filesystem without FS_SETATTR_TIMES_NOWAIT. IORING_OP_OPENAT does the
  same lookup inline, but rarely fails after it, while here every
  request on ext4, xfs or btrfs looks the path up twice. (Sashiko)

v1: https://lore.kernel.org/io-uring/20261030123945.1704118-1-meven@kde.org/

Méven Car (3):
  fs: add filename_utimensat()
  io_uring: add support for utimensat
  io_uring: set times inline when the filesystem cannot sleep

 fs/internal.h                 |   5 ++
 fs/namespace.c                |  20 +++++++
 fs/utimes.c                   |  53 +++++++++++++----
 include/linux/fs.h            |   1 +
 include/uapi/linux/io_uring.h |   2 +
 io_uring/fs.c                 | 103 ++++++++++++++++++++++++++++++++++
 io_uring/fs.h                 |   4 ++
 io_uring/opdef.c              |   8 +++
 mm/shmem.c                    |   3 +-
 9 files changed, 187 insertions(+), 12 deletions(-)


base-commit: ecb6817b769fd7d71af2c9f446188649020b2e53
-- 
2.56.0

[RFC PATCH v2 1/3] fs: add filename_utimensat()

[Méven Car]

From: Méven Car <meven@kde.org>

Copy utilities that preserve timestamps, like cp -p or a file manager,
set the times of every copy after its last write. To do that through
io_uring, the request needs the utimensat() lookup on a name that was
imported at submission time.

io_uring takes the path of a request when it is submitted and looks it
up later, from a worker thread. The path ops it already has call the
filename_*() helpers of fs/namei.c with a struct filename. utimensat(2)
had no such helper: do_utimes_path() took the user pointer and imported
the name itself.

Split the lookup and the vfs_utimes() call out of do_utimes_path() into
filename_utimensat(), declared in fs/internal.h next to the other
filename_*() helpers. do_utimes_path() now only imports the name.

Signed-off-by: Méven Car <meven@kde.org>
---
 fs/internal.h |  2 ++
 fs/utimes.c   | 12 +++++++++---
 2 files changed, 11 insertions(+), 3 deletions(-)

diff --git a/fs/internal.h b/fs/internal.h
index c658c8a5ebd5..985b2bc5895e 100644
--- a/fs/internal.h
+++ b/fs/internal.h
@@ -63,6 +63,8 @@ int filename_mknodat(int dfd, struct filename *name, umode_t mode, unsigned int
 int filename_symlinkat(struct filename *from, int newdfd, struct filename *to);
 int filename_linkat(int olddfd, struct filename *old, int newdfd,
 			struct filename *new, int flags);
+int filename_utimensat(int dfd, struct filename *name,
+		       struct timespec64 *times, int flags);
 int vfs_tmpfile(struct mnt_idmap *idmap,
 		const struct path *parentpath,
 		struct file *file, umode_t mode);
diff --git a/fs/utimes.c b/fs/utimes.c
index e22664e4115f..8072f05ac8dc 100644
--- a/fs/utimes.c
+++ b/fs/utimes.c
@@ -79,8 +79,8 @@ int vfs_utimes(const struct path *path, struct timespec64 *times)
 }
 EXPORT_SYMBOL_GPL(vfs_utimes);
 
-static int do_utimes_path(int dfd, const char __user *filename,
-		struct timespec64 *times, int flags)
+int filename_utimensat(int dfd, struct filename *name,
+		       struct timespec64 *times, int flags)
 {
 	struct path path;
 	int lookup_flags = 0, error;
@@ -91,7 +91,6 @@ static int do_utimes_path(int dfd, const char __user *filename,
 	if (!(flags & AT_SYMLINK_NOFOLLOW))
 		lookup_flags |= LOOKUP_FOLLOW;
 
-	CLASS(filename_uflags, name)(filename, flags);
 retry:
 	error = filename_lookup(dfd, name, lookup_flags, &path, NULL);
 	if (error)
@@ -105,6 +104,13 @@ static int do_utimes_path(int dfd, const char __user *filename,
 	return error;
 }
 
+static int do_utimes_path(int dfd, const char __user *filename,
+		struct timespec64 *times, int flags)
+{
+	CLASS(filename_uflags, name)(filename, flags);
+	return filename_utimensat(dfd, name, times, flags);
+}
+
 static int do_utimes_fd(int fd, struct timespec64 *times, int flags)
 {
 	if (flags)
-- 
2.56.0

[RFC PATCH v2 2/3] io_uring: add support for utimensat

[Méven Car]

From: Méven Car <meven@kde.org>

A program that copies many files sets the times of every copy after
its last write. On NFS and SMB each of those calls is a round trip to
the server, and with io_uring it is one of the steps of the copy that
still have to leave the ring and run one file at a time.

IORING_OP_UTIMENSAT behaves like utimensat(2). sqe->fd is the directory
fd, sqe->addr the path, sqe->addr2 a pointer to two struct
__kernel_timespec, and sqe->utimes_flags takes AT_SYMLINK_NOFOLLOW and
AT_EMPTY_PATH. An addr2 of zero sets both times to the current time, as
a NULL times argument does. The times and the path are copied at
submission.

As with utimensat(2), a NULL path sets the times of the file that
sqe->fd refers to, without a lookup, and takes no flags. Two UTIME_OMIT
times complete with 0 without a lookup.

Like the other path ops, the request is issued from io-wq, apart from
the no-op, because the lookup and notify_change() can sleep. A fixed
file is not accepted as the directory fd. A failed request fails the
rest of its link.

Signed-off-by: Méven Car <meven@kde.org>
---
 include/uapi/linux/io_uring.h |  2 +
 io_uring/fs.c                 | 97 +++++++++++++++++++++++++++++++++++
 io_uring/fs.h                 |  4 ++
 io_uring/opdef.c              |  8 +++
 4 files changed, 111 insertions(+)

diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
index b42b66b424a7..6a98c9282e0b 100644
--- a/include/uapi/linux/io_uring.h
+++ b/include/uapi/linux/io_uring.h
@@ -76,6 +76,7 @@ struct io_uring_sqe {
 		__u32		install_fd_flags;
 		__u32		nop_flags;
 		__u32		pipe_flags;
+		__u32		utimes_flags;
 	};
 	__u64	user_data;	/* data to be passed back at completion time */
 	/* pack this to avoid bogus arm OABI complaints */
@@ -319,6 +320,7 @@ enum io_uring_op {
 	IORING_OP_PIPE,
 	IORING_OP_NOP128,
 	IORING_OP_URING_CMD128,
+	IORING_OP_UTIMENSAT,
 
 	/* this goes last, obviously */
 	IORING_OP_LAST,
diff --git a/io_uring/fs.c b/io_uring/fs.c
index 26ea841a22e7..71164bddd958 100644
--- a/io_uring/fs.c
+++ b/io_uring/fs.c
@@ -304,3 +304,100 @@ void io_link_cleanup(struct io_kiocb *req)
 	dismiss_delayed_filename(&sl->oldpath);
 	dismiss_delayed_filename(&sl->newpath);
 }
+
+/* Internal flags, above the AT_ flags of the SQE */
+#define IO_UTIMENSAT_NOOP	BIT(30)	/* both times UTIME_OMIT */
+#define IO_UTIMENSAT_NO_PATH	BIT(31)	/* NULL path: the dfd itself */
+
+struct io_utimensat {
+	struct file			*file;
+	int				dfd;
+	unsigned int			flags;
+	struct delayed_filename		filename;
+	/* NULL sets both times to the current time */
+	struct timespec64		*times;
+	struct timespec64		ts[2];
+};
+
+int io_utimensat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
+{
+	struct io_utimensat *ut = io_kiocb_to_cmd(req, struct io_utimensat);
+	struct __kernel_timespec __user *uts;
+	const char __user *fname;
+	int err;
+
+	if (sqe->len || sqe->buf_index || sqe->splice_fd_in || sqe->addr3 ||
+	    sqe->__pad2[0])
+		return -EINVAL;
+	if (unlikely(req->flags & REQ_F_FIXED_FILE))
+		return -EBADF;
+
+	ut->dfd = READ_ONCE(sqe->fd);
+	ut->flags = READ_ONCE(sqe->utimes_flags);
+	if (ut->flags & ~(AT_SYMLINK_NOFOLLOW | AT_EMPTY_PATH))
+		return -EINVAL;
+
+	uts = u64_to_user_ptr(READ_ONCE(sqe->addr2));
+	ut->times = NULL;
+	if (uts) {
+		if (get_timespec64(&ut->ts[0], &uts[0]) ||
+		    get_timespec64(&ut->ts[1], &uts[1]))
+			return -EFAULT;
+		ut->times = ut->ts;
+		if (ut->ts[0].tv_nsec == UTIME_OMIT &&
+		    ut->ts[1].tv_nsec == UTIME_OMIT) {
+			ut->flags |= IO_UTIMENSAT_NOOP;
+			return 0;
+		}
+	}
+
+	fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
+	if (!fname) {
+		if (ut->dfd == AT_FDCWD)
+			return -EFAULT;
+		if (ut->flags)
+			return -EINVAL;
+		ut->flags |= IO_UTIMENSAT_NO_PATH;
+		goto out;
+	}
+	err = delayed_getname_uflags(&ut->filename, fname, ut->flags);
+	if (unlikely(err))
+		return err;
+	req->flags |= REQ_F_NEED_CLEANUP;
+out:
+	req->flags |= REQ_F_FORCE_ASYNC;
+	return 0;
+}
+
+int io_utimensat(struct io_kiocb *req, unsigned int issue_flags)
+{
+	struct io_utimensat *ut = io_kiocb_to_cmd(req, struct io_utimensat);
+	int ret = 0;
+
+	WARN_ON_ONCE((issue_flags & IO_URING_F_NONBLOCK) &&
+		     !(ut->flags & IO_UTIMENSAT_NOOP));
+
+	if (ut->flags & IO_UTIMENSAT_NO_PATH) {
+		CLASS(fd, f)(ut->dfd);
+
+		ret = -EBADF;
+		if (!fd_empty(f))
+			ret = vfs_utimes(&fd_file(f)->f_path, ut->times);
+	} else if (!(ut->flags & IO_UTIMENSAT_NOOP)) {
+		CLASS(filename_complete_delayed, name)(&ut->filename);
+		ret = filename_utimensat(ut->dfd, name, ut->times, ut->flags);
+	}
+
+	req->flags &= ~REQ_F_NEED_CLEANUP;
+	if (ret < 0)
+		req_set_fail(req);
+	io_req_set_res(req, ret, 0);
+	return IOU_COMPLETE;
+}
+
+void io_utimensat_cleanup(struct io_kiocb *req)
+{
+	struct io_utimensat *ut = io_kiocb_to_cmd(req, struct io_utimensat);
+
+	dismiss_delayed_filename(&ut->filename);
+}
diff --git a/io_uring/fs.h b/io_uring/fs.h
index 0bb5efe3d6bb..2812dc3b4b2f 100644
--- a/io_uring/fs.h
+++ b/io_uring/fs.h
@@ -18,3 +18,7 @@ int io_symlinkat(struct io_kiocb *req, unsigned int issue_flags);
 int io_linkat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe);
 int io_linkat(struct io_kiocb *req, unsigned int issue_flags);
 void io_link_cleanup(struct io_kiocb *req);
+
+int io_utimensat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe);
+int io_utimensat(struct io_kiocb *req, unsigned int issue_flags);
+void io_utimensat_cleanup(struct io_kiocb *req);
diff --git a/io_uring/opdef.c b/io_uring/opdef.c
index cf3aa2242cd7..383a9b9d1df4 100644
--- a/io_uring/opdef.c
+++ b/io_uring/opdef.c
@@ -591,6 +591,10 @@ const struct io_issue_def io_issue_defs[] = {
 		.prep			= io_uring_cmd_prep,
 		.issue			= io_uring_cmd,
 	},
+	[IORING_OP_UTIMENSAT] = {
+		.prep			= io_utimensat_prep,
+		.issue			= io_utimensat,
+	},
 };
 
 const struct io_cold_def io_cold_defs[] = {
@@ -849,6 +853,10 @@ const struct io_cold_def io_cold_defs[] = {
 		.sqe_copy		= io_uring_cmd_sqe_copy,
 		.cleanup		= io_uring_cmd_cleanup,
 	},
+	[IORING_OP_UTIMENSAT] = {
+		.name			= "UTIMENSAT",
+		.cleanup		= io_utimensat_cleanup,
+	},
 };
 
 const char *io_uring_get_opcode(u8 opcode)
-- 
2.56.0

[RFC PATCH v2 3/3] io_uring: set times inline when the filesystem cannot sleep

[Méven Car]

From: Méven Car <meven@kde.org>

IORING_OP_UTIMENSAT always went to io-wq. On tmpfs the setattr only
changes the inode in memory, so the handoff to a worker cost more than
the work: 2000 files took 7 ms with utimensat(2) and 20 ms through the
ring.

Try the request inline first. The lookup uses LOOKUP_CACHED, and the
freeze protection and the inode lock are only tried, the former with
the new mnt_want_write_trylock(). A filesystem whose ->setattr() never
sleeps for a change of times alone says so with
FS_SETATTR_TIMES_NOWAIT, which tmpfs sets. Any other filesystem, an
uncached path, a contended lock or a delegation returns -EAGAIN, and
io_uring sends the request to io-wq as before.

Signed-off-by: Méven Car <meven@kde.org>
---
 fs/internal.h      |  5 ++++-
 fs/namespace.c     | 20 ++++++++++++++++++++
 fs/utimes.c        | 45 +++++++++++++++++++++++++++++++++++----------
 include/linux/fs.h |  1 +
 io_uring/fs.c      | 22 ++++++++++++++--------
 mm/shmem.c         |  3 ++-
 6 files changed, 76 insertions(+), 20 deletions(-)

diff --git a/fs/internal.h b/fs/internal.h
index 985b2bc5895e..2370faec54f5 100644
--- a/fs/internal.h
+++ b/fs/internal.h
@@ -64,7 +64,9 @@ int filename_symlinkat(struct filename *from, int newdfd, struct filename *to);
 int filename_linkat(int olddfd, struct filename *old, int newdfd,
 			struct filename *new, int flags);
 int filename_utimensat(int dfd, struct filename *name,
-		       struct timespec64 *times, int flags);
+		       struct timespec64 *times, int flags, bool nowait);
+int __vfs_utimes(const struct path *path, struct timespec64 *times,
+		 bool nowait);
 int vfs_tmpfile(struct mnt_idmap *idmap,
 		const struct path *parentpath,
 		struct file *file, umode_t mode);
@@ -86,6 +88,7 @@ extern int sb_prepare_remount_readonly(struct super_block *);
 extern void __init mnt_init(void);
 
 int mnt_get_write_access_file(struct file *file);
+int mnt_want_write_trylock(struct vfsmount *m);
 void mnt_put_write_access_file(struct file *file);
 
 extern void dissolve_on_fput(struct vfsmount *);
diff --git a/fs/namespace.c b/fs/namespace.c
index ae5dc64f8b45..f4aadd4e1958 100644
--- a/fs/namespace.c
+++ b/fs/namespace.c
@@ -499,6 +499,26 @@ int mnt_want_write(struct vfsmount *m)
 }
 EXPORT_SYMBOL_GPL(mnt_want_write);
 
+/**
+ * mnt_want_write_trylock - get write access to a mount without waiting
+ * @m: the mount on which to take a write
+ *
+ * Like mnt_want_write(), but returns -EAGAIN instead of waiting when the
+ * filesystem is frozen.  When the write operation is finished,
+ * mnt_drop_write() must be called.
+ */
+int mnt_want_write_trylock(struct vfsmount *m)
+{
+	int ret;
+
+	if (!sb_start_write_trylock(m->mnt_sb))
+		return -EAGAIN;
+	ret = mnt_get_write_access(m);
+	if (ret)
+		sb_end_write(m->mnt_sb);
+	return ret;
+}
+
 /**
  * mnt_get_write_access_file - get write access to a file's mount
  * @file: the file who's mount on which to take a write
diff --git a/fs/utimes.c b/fs/utimes.c
index 8072f05ac8dc..a0d67b8e96a1 100644
--- a/fs/utimes.c
+++ b/fs/utimes.c
@@ -18,7 +18,9 @@ static bool nsec_valid(long nsec)
 	return nsec >= 0 && nsec <= 999999999;
 }
 
-int vfs_utimes(const struct path *path, struct timespec64 *times)
+/* With @nowait, returns -EAGAIN instead of sleeping */
+int __vfs_utimes(const struct path *path, struct timespec64 *times,
+		 bool nowait)
 {
 	int error;
 	struct iattr newattrs;
@@ -34,9 +36,17 @@ int vfs_utimes(const struct path *path, struct timespec64 *times)
 			times = NULL;
 	}
 
-	error = mnt_want_write(path->mnt);
-	if (error)
-		goto out;
+	if (nowait) {
+		if (!(inode->i_sb->s_type->fs_flags & FS_SETATTR_TIMES_NOWAIT))
+			return -EAGAIN;
+		error = mnt_want_write_trylock(path->mnt);
+		if (error)
+			goto out;
+	} else {
+		error = mnt_want_write(path->mnt);
+		if (error)
+			goto out;
+	}
 
 	newattrs.ia_valid = ATTR_CTIME | ATTR_MTIME | ATTR_ATIME;
 	if (times) {
@@ -63,24 +73,35 @@ int vfs_utimes(const struct path *path, struct timespec64 *times)
 		newattrs.ia_valid |= ATTR_TOUCH;
 	}
 retry_deleg:
-	inode_lock(inode);
+	if (!nowait) {
+		inode_lock(inode);
+	} else if (!inode_trylock(inode)) {
+		error = -EAGAIN;
+		goto drop;
+	}
+	/* Without a delegated_inode, a delegation gives -EWOULDBLOCK */
 	error = notify_change(mnt_idmap(path->mnt), path->dentry, &newattrs,
-			      &delegated_inode);
+			      nowait ? NULL : &delegated_inode);
 	inode_unlock(inode);
 	if (is_delegated(&delegated_inode)) {
 		error = break_deleg_wait(&delegated_inode);
 		if (!error)
 			goto retry_deleg;
 	}
-
+drop:
 	mnt_drop_write(path->mnt);
 out:
 	return error;
 }
+
+int vfs_utimes(const struct path *path, struct timespec64 *times)
+{
+	return __vfs_utimes(path, times, false);
+}
 EXPORT_SYMBOL_GPL(vfs_utimes);
 
 int filename_utimensat(int dfd, struct filename *name,
-		       struct timespec64 *times, int flags)
+		       struct timespec64 *times, int flags, bool nowait)
 {
 	struct path path;
 	int lookup_flags = 0, error;
@@ -90,13 +111,17 @@ int filename_utimensat(int dfd, struct filename *name,
 
 	if (!(flags & AT_SYMLINK_NOFOLLOW))
 		lookup_flags |= LOOKUP_FOLLOW;
+	if (nowait)
+		lookup_flags |= LOOKUP_CACHED;
 
 retry:
 	error = filename_lookup(dfd, name, lookup_flags, &path, NULL);
 	if (error)
 		return error;
-	error = vfs_utimes(&path, times);
+	error = __vfs_utimes(&path, times, nowait);
 	path_put(&path);
+	if (nowait && error == -ESTALE)
+		return -EAGAIN;
 	if (retry_estale(error, lookup_flags)) {
 		lookup_flags |= LOOKUP_REVAL;
 		goto retry;
@@ -108,7 +133,7 @@ static int do_utimes_path(int dfd, const char __user *filename,
 		struct timespec64 *times, int flags)
 {
 	CLASS(filename_uflags, name)(filename, flags);
-	return filename_utimensat(dfd, name, times, flags);
+	return filename_utimensat(dfd, name, times, flags, false);
 }
 
 static int do_utimes_fd(int fd, struct timespec64 *times, int flags)
diff --git a/include/linux/fs.h b/include/linux/fs.h
index f9d1e05e8ae6..d1800fba15d8 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -2296,6 +2296,7 @@ struct file_system_type {
 #define FS_POWER_FREEZE		256	/* Always freeze on suspend/hibernate */
 #define FS_USERNS_MOUNT_RESTRICTED 512	/* Restrict mount in userns if not already visible */
 #define FS_USERNS_DELEGATABLE	1024	/* Can be mounted inside userns from outside */
+#define FS_SETATTR_TIMES_NOWAIT	2048	/* ->setattr() of times alone never sleeps */
 #define FS_RENAME_DOES_D_MOVE	32768	/* FS will handle d_move() during rename() internally. */
 	int (*init_fs_context)(struct fs_context *);
 	const struct fs_parameter_spec *parameters;
diff --git a/io_uring/fs.c b/io_uring/fs.c
index 71164bddd958..664ea918ae3b 100644
--- a/io_uring/fs.c
+++ b/io_uring/fs.c
@@ -358,34 +358,40 @@ int io_utimensat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
 		if (ut->flags)
 			return -EINVAL;
 		ut->flags |= IO_UTIMENSAT_NO_PATH;
-		goto out;
+		return 0;
 	}
 	err = delayed_getname_uflags(&ut->filename, fname, ut->flags);
 	if (unlikely(err))
 		return err;
 	req->flags |= REQ_F_NEED_CLEANUP;
-out:
-	req->flags |= REQ_F_FORCE_ASYNC;
 	return 0;
 }
 
 int io_utimensat(struct io_kiocb *req, unsigned int issue_flags)
 {
 	struct io_utimensat *ut = io_kiocb_to_cmd(req, struct io_utimensat);
+	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
 	int ret = 0;
 
-	WARN_ON_ONCE((issue_flags & IO_URING_F_NONBLOCK) &&
-		     !(ut->flags & IO_UTIMENSAT_NOOP));
-
 	if (ut->flags & IO_UTIMENSAT_NO_PATH) {
 		CLASS(fd, f)(ut->dfd);
 
 		ret = -EBADF;
 		if (!fd_empty(f))
-			ret = vfs_utimes(&fd_file(f)->f_path, ut->times);
+			ret = __vfs_utimes(&fd_file(f)->f_path, ut->times,
+					   force_nonblock);
+		if (ret == -EAGAIN && force_nonblock)
+			return -EAGAIN;
 	} else if (!(ut->flags & IO_UTIMENSAT_NOOP)) {
 		CLASS(filename_complete_delayed, name)(&ut->filename);
-		ret = filename_utimensat(ut->dfd, name, ut->times, ut->flags);
+		ret = filename_utimensat(ut->dfd, name, ut->times, ut->flags,
+					 force_nonblock);
+		if (ret == -EAGAIN && force_nonblock) {
+			ret = putname_to_delayed(&ut->filename,
+						 no_free_ptr(name));
+			if (likely(!ret))
+				return -EAGAIN;
+		}
 	}
 
 	req->flags &= ~REQ_F_NEED_CLEANUP;
diff --git a/mm/shmem.c b/mm/shmem.c
index 897fa2b61346..405b39fc5d0e 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -5310,7 +5310,8 @@ static struct file_system_type shmem_fs_type = {
 	.parameters	= shmem_fs_parameters,
 #endif
 	.kill_sb	= kill_anon_super,
-	.fs_flags	= FS_USERNS_MOUNT | FS_ALLOW_IDMAP | FS_MGTIME,
+	.fs_flags	= FS_USERNS_MOUNT | FS_ALLOW_IDMAP | FS_MGTIME |
+			  FS_SETATTR_TIMES_NOWAIT,
 };
 
 #if defined(CONFIG_SYSFS) && defined(CONFIG_TMPFS)
-- 
2.56.0

Re: [RFC PATCH v2 1/3] fs: add filename_utimensat()

[Jan Kara]

On Wed 04-11-26 12:06:02, Méven Car wrote:
> From: Méven Car <meven@kde.org>
> 
> Copy utilities that preserve timestamps, like cp -p or a file manager,
> set the times of every copy after its last write. To do that through
> io_uring, the request needs the utimensat() lookup on a name that was
> imported at submission time.
> 
> io_uring takes the path of a request when it is submitted and looks it
> up later, from a worker thread. The path ops it already has call the
> filename_*() helpers of fs/namei.c with a struct filename. utimensat(2)
> had no such helper: do_utimes_path() took the user pointer and imported
> the name itself.
> 
> Split the lookup and the vfs_utimes() call out of do_utimes_path() into
> filename_utimensat(), declared in fs/internal.h next to the other
> filename_*() helpers. do_utimes_path() now only imports the name.
> 
> Signed-off-by: Méven Car <meven@kde.org>

Looks ok. Feel free to add:

Reviewed-by: Jan Kara <jack@suse.cz>

								Honza

> ---
>  fs/internal.h |  2 ++
>  fs/utimes.c   | 12 +++++++++---
>  2 files changed, 11 insertions(+), 3 deletions(-)
> 
> diff --git a/fs/internal.h b/fs/internal.h
> index c658c8a5ebd5..985b2bc5895e 100644
> --- a/fs/internal.h
> +++ b/fs/internal.h
> @@ -63,6 +63,8 @@ int filename_mknodat(int dfd, struct filename *name, umode_t mode, unsigned int
>  int filename_symlinkat(struct filename *from, int newdfd, struct filename *to);
>  int filename_linkat(int olddfd, struct filename *old, int newdfd,
>  			struct filename *new, int flags);
> +int filename_utimensat(int dfd, struct filename *name,
> +		       struct timespec64 *times, int flags);
>  int vfs_tmpfile(struct mnt_idmap *idmap,
>  		const struct path *parentpath,
>  		struct file *file, umode_t mode);
> diff --git a/fs/utimes.c b/fs/utimes.c
> index e22664e4115f..8072f05ac8dc 100644
> --- a/fs/utimes.c
> +++ b/fs/utimes.c
> @@ -79,8 +79,8 @@ int vfs_utimes(const struct path *path, struct timespec64 *times)
>  }
>  EXPORT_SYMBOL_GPL(vfs_utimes);
>  
> -static int do_utimes_path(int dfd, const char __user *filename,
> -		struct timespec64 *times, int flags)
> +int filename_utimensat(int dfd, struct filename *name,
> +		       struct timespec64 *times, int flags)
>  {
>  	struct path path;
>  	int lookup_flags = 0, error;
> @@ -91,7 +91,6 @@ static int do_utimes_path(int dfd, const char __user *filename,
>  	if (!(flags & AT_SYMLINK_NOFOLLOW))
>  		lookup_flags |= LOOKUP_FOLLOW;
>  
> -	CLASS(filename_uflags, name)(filename, flags);
>  retry:
>  	error = filename_lookup(dfd, name, lookup_flags, &path, NULL);
>  	if (error)
> @@ -105,6 +104,13 @@ static int do_utimes_path(int dfd, const char __user *filename,
>  	return error;
>  }
>  
> +static int do_utimes_path(int dfd, const char __user *filename,
> +		struct timespec64 *times, int flags)
> +{
> +	CLASS(filename_uflags, name)(filename, flags);
> +	return filename_utimensat(dfd, name, times, flags);
> +}
> +
>  static int do_utimes_fd(int fd, struct timespec64 *times, int flags)
>  {
>  	if (flags)
> -- 
> 2.56.0
> 
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR

Re: [RFC PATCH v2 1/3] fs: add filename_utimensat()

[Christian Brauner]

On Mon, Nov 09, 2026 at 12:05:18PM +0100, Jan Kara wrote:
> On Wed 04-11-26 12:06:02, Méven Car wrote:
> > From: Méven Car <meven@kde.org>
> > 
> > Copy utilities that preserve timestamps, like cp -p or a file manager,
> > set the times of every copy after its last write. To do that through
> > io_uring, the request needs the utimensat() lookup on a name that was
> > imported at submission time.
> > 
> > io_uring takes the path of a request when it is submitted and looks it
> > up later, from a worker thread. The path ops it already has call the
> > filename_*() helpers of fs/namei.c with a struct filename. utimensat(2)
> > had no such helper: do_utimes_path() took the user pointer and imported
> > the name itself.
> > 
> > Split the lookup and the vfs_utimes() call out of do_utimes_path() into
> > filename_utimensat(), declared in fs/internal.h next to the other
> > filename_*() helpers. do_utimes_path() now only imports the name.
> > 
> > Signed-off-by: Méven Car <meven@kde.org>
> 
> Looks ok. Feel free to add:
> 
> Reviewed-by: Jan Kara <jack@suse.cz>

Seems fine.

Re: [RFC PATCH v2 3/3] io_uring: set times inline when the filesystem cannot sleep

[Christian Brauner]

On Wed, Nov 04, 2026 at 12:06:04PM +0100, Méven Car wrote:
> From: Méven Car <meven@kde.org>
> 
> IORING_OP_UTIMENSAT always went to io-wq. On tmpfs the setattr only
> changes the inode in memory, so the handoff to a worker cost more than
> the work: 2000 files took 7 ms with utimensat(2) and 20 ms through the
> ring.
> 
> Try the request inline first. The lookup uses LOOKUP_CACHED, and the
> freeze protection and the inode lock are only tried, the former with
> the new mnt_want_write_trylock(). A filesystem whose ->setattr() never
> sleeps for a change of times alone says so with
> FS_SETATTR_TIMES_NOWAIT, which tmpfs sets. Any other filesystem, an
> uncached path, a contended lock or a delegation returns -EAGAIN, and
> io_uring sends the request to io-wq as before.
> 
> Signed-off-by: Méven Car <meven@kde.org>
> ---
>  fs/internal.h      |  5 ++++-
>  fs/namespace.c     | 20 ++++++++++++++++++++

We do not add a bunch of generic APIs in a io_uring patch.

What exactly is the the idea here? for tmpfs io-wq handoff costs three
times the setattr.

So it tries the request inline with LOOKUP_CACHED, inode_trylock(), a
new mnt_want_write_trylock(), no delegation breaking, and bails with
EAGAIN unless the filesystem type sets a new FS_SETATTR_TIMES_NOWAIT
fs_flags bit. Bah.. and then only tmpfs. ext4, xfs and btrfs can't
because ->setattr() starts a transaction.

The inline path still runs the whole notify_change()
security_inode_setattr(), fsnotify_change(), and
security_inode_post_setattr(). That's all blocking material.

Every request on a filesystem without the bit does a cached lookup with
d_weak_revalidate() and audit_inode() side effects, fails in __vfs_utimes(),
and then io-wq looks the path up again. Eew.

Not acceptable in this form and for sure not with all this new
machinery.