On Tue, 20 Oct 2026 at 11:10, Linus Torvalds
<[email protected]> wrote:
>
> And it all really does look obviously correct, but ...
I think I'm just gping to merge this. I don't love all the changes,
but I love __get_user and __put_user less.
And the one change that makes me worry most is the i915 one, and I
have at least minimally tested that one by virtue of now actually
running the code in question on my laptop that does indeed have i915
integrated graphics.
It does change semantics, but it does so in a situation where I really
do think that the old code was actively buggy and would have triggered
a lockdep splat if it hit that situation.
The i915 code also was actively wrong in its comments about this all,
with some of it going back to some stone age code with
'verify_area(VERIFY_WRITE)' that used to actually trigger page faults
on i386 in order to do early COW.
We haven't done that since the 2.5.x days, and even then we only ever
did it on 32-bit kernels and only if CONFIG_X86_WP_WORKS_OK wasn't
enabled.
Put another way: I'm not claiming my patches are perfect, but I think
they are an improvement, and I'd rather merge this early and find any
issues than keep known-broken code around.
And yes, I might have missed some situation where __get/put_user() was
used for some reason I overlooked.
The reasons I at least thought about:
- the legacy case of "one single access_ok and multiple __get_users
is faster" case
This is not faster any more, and hasn't been for years and years.
But the cases I thought might care about performance - or the cases
that were trivial to convert - got converted with
"scoped_user_read/write_begin() {}" and unsafe_get/put_user() calls,
so now they *really* are faster.
- The vhost use-case that wants to use get_user() in atomic context,
but used __get_user() in order to avoid a warning
The warning was *correct*, and using __get_user() was a bug waiting
to happen, because some atomic situations aren't visible at runtime.
The fix is to actually use pagefault_disable/enable() and get_user()
to make this all explicit, avoid the warning, and avoid the issue that
some atomic sections aren't even visible in the preempt count.
- The "either kernel or user pointer" situation. That was the x86
coco/SEV case, and the alpha "is_prefetch" cases.
- A couple of "kernel pointer only" cases that should just use
"get_kernel_nofault()", but the code pre-dates that helper existing.
but it is possible that there is some case I didn't catch. Still, this
"remove __get/put_user" is way over-due and now that I finally got it
done, I don't want to leave it hanging.
Linus
