Re: [Cluster-devel] [PATCH v7 05/13] fat: make fat_update_time get its own timestamp

On Thu, 2023-08-10 at 02:44 +0900, OGAWA Hirofumi wrote:
> Jeff Layton <[email protected]> writes:
> 
> > On Thu, 2023-08-10 at 00:17 +0900, OGAWA Hirofumi wrote:
> > > Jan Kara <[email protected]> writes:
> 
> [...]
> 
> > My mistake re: lazytime vs. relatime, but Jan is correct that this
> > shouldn't break anything there.
> 
> Actually breaks ("break" means not corrupt fs, means it breaks lazytime
> optimization). It is just not always, but it should be always for some
> userspaces.
> 
> > The logic in the revised generic_update_time is different because FAT is
> > is a bit strange. fat_update_time does extra truncation on the timestamp
> > that it is handed beyond what timestamp_truncate() does.
> > fat_truncate_time is called in many different places too, so I don't
> > feel comfortable making big changes to how that works.
> > 
> > In the case of generic_update_time, it calls inode_update_timestamps
> > which returns a mask that shows which timestamps got updated. It then
> > marks the dirty_flags appropriately for what was actually changed.
> > 
> > generic_update_time is used across many filesystems so we need to ensure
> > that it's OK to use even when multigrain timestamps are enabled. Those
> > haven't been enabled in FAT though, so I didn't bother, and left it to
> > dirtying the inode in the same way it was before, even though it now
> > fetches its own timestamps from the clock. Given the way that the mtime
> > and ctime are smooshed together in FAT, that seemed reasonable.
> > 
> > Is there a particular case or flag combination you're concerned about
> > here?
> 
> Yes. Because FAT has strange timestamps that different granularity on
> disk . This is why generic time truncation doesn't work for FAT.
> 
> Well anyway, my concern is the only following part. In
> generic_update_time(), S_[CM]TIME are not the cause of I_DIRTY_SYNC if
> lazytime mode.
> 
> -	if ((flags & S_VERSION) && inode_maybe_inc_iversion(inode, false))
> +	if ((flags & (S_VERSION|S_CTIME|S_MTIME)) && inode_maybe_inc_iversion(inode, false))
> 		dirty_flags |= I_DIRTY_SYNC;
> 

That would be wrong. The problem is that we're changing how update_time
works:

Previously, update_time was given a timestamp and a set of S_* flags to
indicate which fields should be updated. Now, update_time is not given a
timestamp. It needs to fetch it itself, but that subtly changes the
meaning of the flags field.

It now means "these fields needed to be updated when I last checked".
The timestamp and i_version may now be different from when the flags
field was set. This means that if any of S_CTIME/S_MTIME/S_VERSION were
set that we need to attempt to update all 3 of them. They may now be
different from the timestamp or version that we ultimately end up with.

The above may look to you like it would always cause I_DIRTY_SYNC to be
set on any ctime or mtime update, but inode_maybe_inc_iversion only
returns true if it actually updated i_version, and it only does that if
someone issued a ->getattr against the file since the last time it was
updated.

So, this shouldn't generate any more DIRTY_SYNC updates than it did
before.
-- 
Jeff Layton <[email protected]>