On Wed, Jun 24, 2020 at 11:11 AM Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> So in *those* cases, we'd basically just do "oh, ok, we are supposed
> to use a kernel pointer" based on the setfs value.
The important part here si that we don't need to change any
interfaces, because we don't add that whole "carry the bit around with
the pointer".
I agree that that would be the nice clean interface _if_ we intended
for this to be something we care about. But we already _have_ that
interface in the "iter" code, I absolutely do not think we should
create a new one.
So that's why the (admittedly hacky) "just support the old model for
special cases" kind of approach. Make it really easy to convert some
very specific individual places that might care, and make it very
obvious what those places are.
Maybe in a year or two, there's only a couple such places, and we can
see if we can clean those up separately. But make it easy do the
transition to the new model by _not_ changing the basic logic for now.
See what I'm aiming for?
In particular, for architectures that haven't been modified, this
results in zero changes what-so-ever.
Linus
