[PATCH v1 1/2] uaccess: Add __user_write_access_begin().

[Kuniyuki Iwashima]

On Thu, Oct 23, 2025 at 1:29 AM David Laight
<david.laight.linux at gmail.com> wrote:
>
> On Wed, 22 Oct 2025 19:37:27 -1000
> Linus Torvalds <torvalds at linux-foundation.org> wrote:
>
> > On Wed, 22 Oct 2025 at 14:05, Kuniyuki Iwashima <kuniyu at google.com> wrote:
> > >
> > > unsafe_put_user() can be used to save a stac/clac pair, but
> > > masked_user_access_begin() or user_access_begin() introduces
> > > an unnecessary address masking or access_ok().
> > >
> > > Add a low-level helper for such a use case.
> >
> > I really suspect that you cannot actually measure the cost of the
> > extra masking, and would be much happier if you just used a regular
> > "user_access_begin()" (perhaps the "user_write_access_begin()"
> > variant).
>
> Or wait for scoped_user_write_access() to get committed and then use that.

IIUC, scoped_user_write_access() is simply inlined to
masked_user_access_begin() or user_access_begin(), and this
is the case where I saw no improvement or even worse performance.

>
>         David
>
> >
> > The masking is very cheap - literally just a couple of ALU
> > instructions. And unless you can actually measure some real advantage
> > of avoiding it, let's not add another helper to this area.

Yes, it's only 3 instructions on x86_64, but by saving them
I saw better performance constantly.  Please see the numbers here.
https://lore.kernel.org/lkml/20251024051653.66329-1-kuniyu@google.com/


> >
> > We spent a fair amount of time undoing years of "__get_user()" and
> > "__put_user()" cases that didn't actually help, and sometimes only
> > made it hard to see where the actual user pointer validation was done.
> >
> >                Linus
> >
>