On Tuesday July 10, akpm@linux-foundation.org wrote:

Yes, thanks.  It doesn't actually tell us why we want to implement
this attribute and it doesn't tell us what the implications of failing
to do so are, but I guess we can take that on trust from the NFS guys.

You would like to think so, but remember NFSv4 was designed by a
committee :-)

The 'change' number is used for cache consistency, and as the spec
makes very strong statements about the 'change' number, it is very
hard (or impossible) to implement a server correctly without storing a
change number in stable storage (just one of my grips about V4).

But I suspect the ext4 implementation doesn't actually do this.  afaict we
won't update i_version for file overwrites (especially if s_time_gran can
indeed be 1,000,000,000) and of course for MAP_SHARED modifications.  What
would be the implications of this?

The first part sounds like a bug - i_version should really be updated
by every call to ->commit_write (if that is still what it is called).

The MAP_SHARED thing is less obvious.  I guess every time we notice
that the page might have been changed, we need to increment i_version.

And how does the NFS server know that the filesystem implements i_version? 
Will a zero-value of i_version have special significance, telling the
server to not send this attribute, perhaps?

That is a very important question.  Zero probably makes sense, but
what ever it is needs to be agreed and documented.
And just by-the-way, the server doesn't really have the option of not
sending the attribute.  If i_version isn't defined, it has to fake
something using mtime, and hope that is good enough.

Alternately we could mandate that i_version is always kept up-to-date
and if a filesystem doesn't have anything to load from storage, it
just sets it to the current time in nanoseconds.

That would mean that a client would need to flush it's cache whenever
the inode fell out of cache on the server, but I don't think we can
reliably do better than that.

I think I like that approach.

So my vote is to increment i_version in common code every time any
change is made to the file, and alloc_inode should initialise it to
current time, which might be changed by the filesystem before it calls
unlock_new_inode. 
... but doesn't lustre want to control its i_version... so maybe not :-(

NeilBrown