Re: __get_user slower than get_user (was Re: [RFC PATCH V3 0/5] Hi:)

[Linus Torvalds]

On Tue, Jan 8, 2019 at 8:31 PM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
>
> Linus, given that you just changed all users of access_ok anyway, do
> you still think that the access_ok() conversion to return a speculation
> sanitized pointer or NULL is too big a conversion?

I didn't actually change a single access_ok().

I changed the (very few) users of "user_access_begin()" to do an
access_ok() in them. There were 8 of them total.

It turns out that two of those cases (the strn*_user() ones) found
bugs in the implementation of access_ok() of two architectures, and
then looking at the others found that six more architectures also had
problems, but those weren't actually because of any access_ok()
changes, they were pre-existing issues. So we definitely had
unfortunate bugs in access_ok(), but they were mostly the benign kind
(ir the "use arguments twice - a real potential bug, but not one that
actually likely makes any difference to existing users)

Changing all 600+ users of access_ok() would be painful.

That said, one thing I *would* like to do is to just get rid of
__get_user() and __put_user() entirely. Or rather, just make them do
exactly the same thing that the normal "get_user()"/"put_user()"
functions do.

And then, _within_ the case of get_user()/put_user(), doing the
access_ok() as a data dependency rather than a lfence should be easy
enough.

                     Linus