On Tue, Oct 6, 2015 at 12:33 AM, Dave Hansen
[off-list ref] wrote:

Did you mean that as a cleanup, or to fix the regression?

Purely as a cleanup to try to avoid the (already existing) special
case in at least ext4 - and possibly others.

So that patch was meant just for discussion - it's not really fixing
any existing bugs, and I didn't actually keep it live in my tree.

I'm planning on just doing the revert for now, but I'll wait a bit to
see how this thread pans out first.

Since the page isn't faulted in yet, iov_iter_copy_from_user_atomic()
had already set copied=0

Not necessarily. For your case that only does one-byte writes, yes.
But in the generic case you may well have a page-crossing source area
in user space, and get a partial success from the user copy. It's that
partial success case (when the rest of the missing data isn't
necessarily already up-to-date) that I'd like to have low-level
filesystems not have to worry about.

           Linus