Re: [PATCH v2] epoll: Use user_write_access_begin() and unsafe_put_user() in epoll_put_uevent().

[Linus Torvalds]

On Tue, 28 Oct 2025 at 10:56, Kuniyuki Iwashima <kuniyu@xxxxxxxxxx> wrote:
>
> Let's use user_write_access_begin() and unsafe_put_user() in
> epoll_put_uevent().
>
> We saw 2% more pps with udp_rr by saving a stac/clac pair.

This patch looks fine to me. Simple and targeted.

> Another option would be to use can_do_masked_user_access()
> and masked_user_access_begin(), but we saw 3% regression. (See Link)

So I find this intriguing, because generally,
masked_user_access_begin() should _never_ be slower than
user_write_access_begin().

user_write_access_begin() ends up doing a __uaccess_begin_nospec() on
x86, which is not just the STAC instruction, but also a barrier.

In contrast, masked_user_access_begin() obviously also has the STAC,
but it avoids the barrier and only uses a simple conditional mask.

So I get the feeling that you did something wrong. In particular,
following your link, I see you describe that case (2) as

  2) masked_user_access_begin() + masked_user_access_begin()
  97% pps compared to 1).
  96% calls of ep_try_send_events().

and you mention masked_user_access_begin() *twice*.

Which would certainly explain why it's slower.

Can you show what the patch you used is?

Because I think the proper patch should look something like the
attached.. For me, that generates


        movabs $0x123456789abcdef,%rcx
        cmp    %rcx,%r15
        cmova  %rcx,%r15
        stac
        mov    %r12d,(%r15)
        mov    %rax,0x4(%r15)
        clac

which honestly should be pretty much optimal.

(That 0x123456789abcdef is just a placeholder for the USER_PTR_MAX
value - it gets rewritten at boot to the right value).

NOTE! The attached patch has absolutely not been tested. I only used
it to verify the code generation visually, so you should definitely
check it yourself.

               Linus
 include/linux/eventpoll.h | 13 +++++++++++--
 1 file changed, 11 insertions(+), 2 deletions(-)

diff --git a/include/linux/eventpoll.h b/include/linux/eventpoll.h
index ccb478eb174b..6bceed34bb21 100644
--- a/include/linux/eventpoll.h
+++ b/include/linux/eventpoll.h
@@ -82,11 +82,20 @@ static inline struct epoll_event __user *
 epoll_put_uevent(__poll_t revents, __u64 data,
 		 struct epoll_event __user *uevent)
 {
-	if (__put_user(revents, &uevent->events) ||
-	    __put_user(data, &uevent->data))
+	if (can_do_masked_user_access())
+		uevent = masked_user_access_begin(uevent);
+	else if (!user_write_access_begin(uevent, sizeof(*uevent)))
 		return NULL;
 
+	unsafe_put_user(revents, &uevent->events, Efault);
+	unsafe_put_user(data, &uevent->data, Efault);
+
+	user_write_access_end();
 	return uevent+1;
+
+Efault:
+	user_access_end();
+	return NULL;
 }
 #endif