On Mon, Dec 21, 2020 at 03:33:30PM -0800, Linus Torvalds wrote:
> On Mon, Dec 21, 2020 at 3:12 PM Yu Zhao <yuzhao@xxxxxxxxxx> wrote:
> >
> > I can't say I disagree with you but the man has made the call and I
> > think we should just move on.
>
> "The man" can always be convinced by numbers.
>
> So if somebody comes up with an alternate patch, and explains it, and
> shows that it is better - go for it.
>
> I just think that if mprotect() can take the mmap lock for writing,
> then userfaultfd sure as hell can. What odd load does people have
> where userfaultfd is more important than mprotect?
>
> So as far as the man is concerned, I think "just fix userfaultfd" is
> simply the default obvious operation.
>
> Not necessarily a final endpoint.
>
> Linus
My first instinct is to be conservative and revert 09854ba94c6a ("mm:
do_wp_page() simplification") so people are less likely to come back
and complain about performance issues from holding mmap lock for
write when clearing pte_write.
That being said, I do like the simplicity of 09854ba94c6a as well as
having one simple rule that dictates what we should do when clearing
pte_write(). And "userfaultfd is not the most important part of the
system" is a fair point.
