On Fri, 2023-09-29 at 11:44 +0200, Christian Brauner wrote:

quoted

It is a lot of churn though.

I think that i_{a,c,m}time shouldn't be accessed directly by
filesystems same as no filesystem should really access i_{g,u}id which
we also provide i_{g,u}id_{read,write}() accessors for. The mode is
another example where really most often should use helpers because of all
the set*id stripping that we need to do (and the bugs that we had
because of this...).

The interdependency between ctime and mtime is enough to hide this in
accessors. The other big advantage is simply grepability. So really I
would like to see this change even without the type switch.

In other words, there's no need to lump the two changes together. Do the
conversion part and we can argue about the switch to discrete integers
separately.

The other adavantage is that we have a cycle to see any possible
regression from the conversion.

Thoughts anyone?

That works for me, and sort of what I was planning anyway. I mostly just
did the change to timestamp storage to see what it would look like
afterward.

FWIW, I'm planning to do a v2 patchbomb early next week, with the
changes that Chuck suggested (specific helpers for fetching the _sec and
_nsec fields). For now, I'll drop the change from timespec64 to discrete
fields. We can do that in a separate follow-on set.
-- 
Jeff Layton [off-list ref]