Re: [Cluster-devel] [PATCH v7 12/13] ext4: switch to multigrain timestamps

On Wed, 2023-09-20 at 10:41 +0200, Christian Brauner wrote:
> > > f1 was last written to *after* f2 was last written to. If the timesta=
mp of f1
> > > is then lower than the timestamp of f2, timestamps are fundamentally =
broken.
> > >=20
> > > Many things in user-space depend on timestamps, such as build system
> > > centered around 'make', but also 'find ... -newer ...'.
> > >=20
> >=20
> >=20
> > What does breakage with make look like in this situation? The "fuzz"
> > here is going to be on the order of a jiffy. The typical case for make
> > timestamp comparisons is comparing source files vs. a build target. If
> > those are being written nearly simultaneously, then that could be an
> > issue, but is that a typical behavior? It seems like it would be hard t=
o
> > rely on that anyway, esp. given filesystems like NFS that can do lazy
> > writeback.
> >=20
> > One of the operating principles with this series is that timestamps can
> > be of varying granularity between different files. Note that Linux
> > already violates this assumption when you're working across filesystems
> > of different types.
> >=20
> > As to potential fixes if this is a real problem:
> >=20
> > I don't really want to put this behind a mount or mkfs option (a'la
> > relatime, etc.), but that is one possibility.
> >=20
> > I wonder if it would be feasible to just advance the coarse-grained
> > current_time whenever we end up updating a ctime with a fine-grained
> > timestamp? It might produce some inode write amplification. Files that
>=20
> Less than ideal imho.
>=20
> If this risks breaking existing workloads by enabling it unconditionally
> and there isn't a clear way to detect and handle these situations
> without risk of regression then we should move this behind a mount
> option.
>=20
> So how about the following:
>=20
> From cb14add421967f6e374eb77c36cc4a0526b10d17 Mon Sep 17 00:00:00 2001
> From: Christian Brauner <[email protected]>
> Date: Wed, 20 Sep 2023 10:00:08 +0200
> Subject: [PATCH] vfs: move multi-grain timestamps behind a mount option
>=20
> While we initially thought we can do this unconditionally it turns out
> that this might break existing workloads that rely on timestamps in very
> specific ways and we always knew this was a possibility. Move
> multi-grain timestamps behind a vfs mount option.
>=20
> Signed-off-by: Christian Brauner <[email protected]>
> ---
> =A0fs/fs_context.c     | 18 ++++++++++++++++++
> =A0fs/inode.c          |  4 ++--
> =A0fs/proc_namespace.c |  1 +
> =A0fs/stat.c           |  2 +-
> =A0include/linux/fs.h  |  4 +++-
> =A05 files changed, 25 insertions(+), 4 deletions(-)
>=20
> diff --git a/fs/fs_context.c b/fs/fs_context.c
> index a0ad7a0c4680..dd4dade0bb9e 100644
> --- a/fs/fs_context.c
> +++ b/fs/fs_context.c
> @@ -44,6 +44,7 @@ static const struct constant_table common_set_sb_flag[]=
 =3D {
> =A0=09{ "mand",=09SB_MANDLOCK },
> =A0=09{ "ro",=09=09SB_RDONLY },
> =A0=09{ "sync",=09SB_SYNCHRONOUS },
> +=09{ "mgtime",=09SB_MGTIME },
> =A0=09{ },
> =A0};
> =A0
>=20
> @@ -52,18 +53,32 @@ static const struct constant_table common_clear_sb_fl=
ag[] =3D {
> =A0=09{ "nolazytime",=09SB_LAZYTIME },
> =A0=09{ "nomand",=09SB_MANDLOCK },
> =A0=09{ "rw",=09=09SB_RDONLY },
> +=09{ "nomgtime",=09SB_MGTIME },
> =A0=09{ },
> =A0};
> =A0
>=20
> +static inline int check_mgtime(unsigned int token, const struct fs_conte=
xt *fc)
> +{
> +=09if (token !=3D SB_MGTIME)
> +=09=09return 0;
> +=09if (!(fc->fs_type->fs_flags & FS_MGTIME))
> +=09=09return invalf(fc, "Filesystem doesn't support multi-grain timestam=
ps");
> +=09return 0;
> +}
> +
> =A0/*
> =A0=A0* Check for a common mount option that manipulates s_flags.
> =A0=A0*/
> =A0static int vfs_parse_sb_flag(struct fs_context *fc, const char *key)
> =A0{
> =A0=09unsigned int token;
> +=09int ret;
> =A0
>=20
> =A0=09token =3D lookup_constant(common_set_sb_flag, key, 0);
> =A0=09if (token) {
> +=09=09ret =3D check_mgtime(token, fc);
> +=09=09if (ret)
> +=09=09=09return ret;
> =A0=09=09fc->sb_flags |=3D token;
> =A0=09=09fc->sb_flags_mask |=3D token;
> =A0=09=09return 0;
> @@ -71,6 +86,9 @@ static int vfs_parse_sb_flag(struct fs_context *fc, con=
st char *key)
> =A0
>=20
> =A0=09token =3D lookup_constant(common_clear_sb_flag, key, 0);
> =A0=09if (token) {
> +=09=09ret =3D check_mgtime(token, fc);
> +=09=09if (ret)
> +=09=09=09return ret;
> =A0=09=09fc->sb_flags &=3D ~token;
> =A0=09=09fc->sb_flags_mask |=3D token;
> =A0=09=09return 0;
> diff --git a/fs/inode.c b/fs/inode.c
> index 54237f4242ff..fd1a2390aaa3 100644
> --- a/fs/inode.c
> +++ b/fs/inode.c
> @@ -2141,7 +2141,7 @@ EXPORT_SYMBOL(current_mgtime);
> =A0
>=20
> =A0static struct timespec64 current_ctime(struct inode *inode)
> =A0{
> -=09if (is_mgtime(inode))
> +=09if (IS_MGTIME(inode))
> =A0=09=09return current_mgtime(inode);
> =A0=09return current_time(inode);
> =A0}
> @@ -2588,7 +2588,7 @@ struct timespec64 inode_set_ctime_current(struct in=
ode *inode)
> =A0=09=09now =3D current_time(inode);
> =A0
>=20
> =A0=09=09/* Just copy it into place if it's not multigrain */
> -=09=09if (!is_mgtime(inode)) {
> +=09=09if (!IS_MGTIME(inode)) {
> =A0=09=09=09inode_set_ctime_to_ts(inode, now);
> =A0=09=09=09return now;
> =A0=09=09}
> diff --git a/fs/proc_namespace.c b/fs/proc_namespace.c
> index 250eb5bf7b52..08f5bf4d2c6c 100644
> --- a/fs/proc_namespace.c
> +++ b/fs/proc_namespace.c
> @@ -49,6 +49,7 @@ static int show_sb_opts(struct seq_file *m, struct supe=
r_block *sb)
> =A0=09=09{ SB_DIRSYNC, ",dirsync" },
> =A0=09=09{ SB_MANDLOCK, ",mand" },
> =A0=09=09{ SB_LAZYTIME, ",lazytime" },
> +=09=09{ SB_MGTIME, ",mgtime" },
> =A0=09=09{ 0, NULL }
> =A0=09};
> =A0=09const struct proc_fs_opts *fs_infop;
> diff --git a/fs/stat.c b/fs/stat.c
> index 6e60389d6a15..2f18dd5de18b 100644
> --- a/fs/stat.c
> +++ b/fs/stat.c
> @@ -90,7 +90,7 @@ void generic_fillattr(struct mnt_idmap *idmap, u32 requ=
est_mask,
> =A0=09stat->size =3D i_size_read(inode);
> =A0=09stat->atime =3D inode->i_atime;
> =A0
>=20
> -=09if (is_mgtime(inode)) {
> +=09if (IS_MGTIME(inode)) {
> =A0=09=09fill_mg_cmtime(stat, request_mask, inode);
> =A0=09} else {
> =A0=09=09stat->mtime =3D inode->i_mtime;
> diff --git a/include/linux/fs.h b/include/linux/fs.h
> index 4aeb3fa11927..03e415fb3a7c 100644
> --- a/include/linux/fs.h
> +++ b/include/linux/fs.h
> @@ -1114,6 +1114,7 @@ extern int send_sigurg(struct fown_struct *fown);
> =A0#define SB_NODEV        BIT(2)=09/* Disallow access to device special =
files */
> =A0#define SB_NOEXEC       BIT(3)=09/* Disallow program execution */
> =A0#define SB_SYNCHRONOUS  BIT(4)=09/* Writes are synced at once */
> +#define SB_MGTIME=09BIT(5)=09/* Use multi-grain timestamps */
> =A0#define SB_MANDLOCK     BIT(6)=09/* Allow mandatory locks on an FS */
> =A0#define SB_DIRSYNC      BIT(7)=09/* Directory modifications are synchr=
onous */
> =A0#define SB_NOATIME      BIT(10)=09/* Do not update access times. */
> @@ -2105,6 +2106,7 @@ static inline bool sb_rdonly(const struct super_blo=
ck *sb) { return sb->s_flags
> =A0=09=09=09=09=09((inode)->i_flags & (S_SYNC|S_DIRSYNC)))
> =A0#define IS_MANDLOCK(inode)=09__IS_FLG(inode, SB_MANDLOCK)
> =A0#define IS_NOATIME(inode)=09__IS_FLG(inode, SB_RDONLY|SB_NOATIME)
> +#define IS_MGTIME(inode)=09__IS_FLG(inode, SB_MGTIME)
> =A0#define IS_I_VERSION(inode)=09__IS_FLG(inode, SB_I_VERSION)
> =A0
>=20
> =A0#define IS_NOQUOTA(inode)=09((inode)->i_flags & S_NOQUOTA)
> @@ -2366,7 +2368,7 @@ struct file_system_type {
> =A0=A0*/
> =A0static inline bool is_mgtime(const struct inode *inode)
> =A0{
> -=09return inode->i_sb->s_type->fs_flags & FS_MGTIME;
> +=09return inode->i_sb->s_flags & SB_MGTIME;
> =A0}
> =A0
>=20
> =A0extern struct dentry *mount_bdev(struct file_system_type *fs_type,

The mount option looks reasonable. Thanks for throwing together the
patch. Maybe in the future we can come up with a way to mitigate the
problems and do this unconditionally?

Reviewed-by: Jeff Layton <[email protected]>