[PATCH v7 05/13] fat: make fat_update_time get its own timestamp
Jeff Layton
jlayton at kernel.org
Wed Aug 9 07:22:54 PDT 2023
On Wed, 2023-08-09 at 22:36 +0900, OGAWA Hirofumi wrote:
> Jeff Layton <jlayton at kernel.org> writes:
>
> > On Wed, 2023-08-09 at 17:37 +0900, OGAWA Hirofumi wrote:
> > > Jeff Layton <jlayton at kernel.org> writes:
> > >
> > > > Also, it may be that things have changed by the time we get to calling
> > > > fat_update_time after checking inode_needs_update_time. Ensure that we
> > > > attempt the i_version bump if any of the S_* flags besides S_ATIME are
> > > > set.
> > >
> > > I'm not sure what it meaning though, this is from
> > > generic_update_time(). Are you going to change generic_update_time()
> > > too? If so, it doesn't break lazytime feature?
> > >
> >
> > Yes. generic_update_time is also being changed in a similar fashion.
> > This shouldn't break the lazytime feature: lazytime is all about how and
> > when timestamps get written to disk. This work is all about which
> > clocksource the timestamps originally come from.
>
> I can only find the following update in this series, another series
> updates generic_update_time()? The patch updates only if S_VERSION is
> set.
>
> Your fat patch sets I_DIRTY_SYNC always instead of I_DIRTY_TIME. When I
> last time checked lazytime, and it was depending on I_DIRTY_TIME.
>
> Are you sure it doesn't break lazytime? I'm totally confusing, and
> really similar with generic_update_time()?
>
I'm a little confused too. Why do you believe this will break
-o relatime handling? This patch changes two things:
1/ it has fat_update_time fetch its own timestamp (and ignore the "now"
parameter). This is in line with the changes in patch #3 of this series,
which explains the rationale for this in more detail.
2/ it changes fat_update_time to also update the i_version if any of
S_CTIME|S_MTIME|S_VERSION are set. relatime is all about the S_ATIME,
and it is specifically excluded from that set.
The rationale for the second change is is also in patch #3, but
basically, we can't guarantee that current_time hasn't changed since we
last checked for inode_needs_update_time, so if any of
S_CTIME/S_MTIME/S_VERSION have changed, then we need to assume that any
of them may need to be changed and attempt to update all 3.
That said, I think the logic in fat_update_time isn't quite right. I
think want something like this on top of this patch to ensure that the
S_CTIME and S_MTIME get updated, even if the flags only have S_VERSION
set.
Thoughts?
---------------------8<-----------------------
diff --git a/fs/fat/misc.c b/fs/fat/misc.c
index 080a5035483f..313eef02f45c 100644
--- a/fs/fat/misc.c
+++ b/fs/fat/misc.c
@@ -346,15 +346,21 @@ int fat_update_time(struct inode *inode, int flags)
if (inode->i_ino == MSDOS_ROOT_INO)
return 0;
- if (flags & (S_ATIME | S_CTIME | S_MTIME)) {
- fat_truncate_time(inode, NULL, flags);
- if (inode->i_sb->s_flags & SB_LAZYTIME)
- dirty_flags |= I_DIRTY_TIME;
- else
- dirty_flags |= I_DIRTY_SYNC;
- }
+ /*
+ * If any of the flags indicate an expicit change to the file, then we
+ * need to ensure that we attempt to update all of 3. We do not do
+ * this in the case of an S_ATIME-only update.
+ */
+ if (flags & (S_CTIME | S_MTIME | S_VERSION))
+ flags |= S_CTIME | S_MTIME | S_VERSION;
+
+ fat_truncate_time(inode, NULL, flags);
+ if (inode->i_sb->s_flags & SB_LAZYTIME)
+ dirty_flags |= I_DIRTY_TIME;
+ else
+ dirty_flags |= I_DIRTY_SYNC;
- if ((flags & (S_VERSION|S_CTIME|S_MTIME)) && inode_maybe_inc_iversion(inode, false))
+ if ((flags & S_VERSION) && inode_maybe_inc_iversion(inode, false))
dirty_flags |= I_DIRTY_SYNC;
More information about the linux-mtd
mailing list