[PATCH 1/5] eventpoll: Convert epoll_put_uevent() to scoped user access

[PATCH 1/5] eventpoll: Convert epoll_put_uevent() to scoped user access

[renpanpan]

From: Eric Dumazet <edumazet@google.com>

Saves two function calls, and one stac/clac pair.

stac/clac is rather expensive on older cpus like Zen 2.

A synthetic network stress test gives a ~1.5% increase of pps
on AMD Zen 2.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Kuniyuki Iwashima <kuniyu@google.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
---
 include/linux/eventpoll.h | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/include/linux/eventpoll.h b/include/linux/eventpoll.h
index ccb478eb174b..ea9ca0e4172a 100644
--- a/include/linux/eventpoll.h
+++ b/include/linux/eventpoll.h
@@ -82,11 +82,14 @@ 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))
-		return NULL;
-
+	scoped_user_write_access_size(uevent, sizeof(*uevent), efault) {
+		unsafe_put_user(revents, &uevent->events, efault);
+		unsafe_put_user(data, &uevent->data, efault);
+	}
 	return uevent+1;
+
+efault:
+	return NULL;
 }
 #endif
 
-- 
2.25.1

Re: [PATCH 1/5] eventpoll: Convert epoll_put_uevent() to scoped user access

[Christophe Leroy (CS GROUP)]

Le 10/03/2026 à 08:54, renpanpan a écrit :
> [Vous ne recevez pas souvent de courriers de renpanpan@kylinos.cn. Découvrez pourquoi ceci est important à https://aka.ms/LearnAboutSenderIdentification ]
> 
> From: Eric Dumazet <edumazet@google.com>
> 
> Saves two function calls, and one stac/clac pair.
> 
> stac/clac is rather expensive on older cpus like Zen 2.
> 
> A synthetic network stress test gives a ~1.5% increase of pps
> on AMD Zen 2.
> 
> Signed-off-by: Eric Dumazet <edumazet@google.com>
> Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
> Cc: Dave Hansen <dave.hansen@intel.com>
> Cc: Kuniyuki Iwashima <kuniyu@google.com>
> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
> ---
>   include/linux/eventpoll.h | 11 +++++++----
>   1 file changed, 7 insertions(+), 4 deletions(-)
> 
> diff --git a/include/linux/eventpoll.h b/include/linux/eventpoll.h
> index ccb478eb174b..ea9ca0e4172a 100644
> --- a/include/linux/eventpoll.h
> +++ b/include/linux/eventpoll.h
> @@ -82,11 +82,14 @@ 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))
> -               return NULL;
> -
> +       scoped_user_write_access_size(uevent, sizeof(*uevent), efault) {

As already mentionned this could be simplified:

	scoped_user_write_access(uevent, efault) {

> +               unsafe_put_user(revents, &uevent->events, efault);
> +               unsafe_put_user(data, &uevent->data, efault);
> +       }
>          return uevent+1;
> +
> +efault:
> +       return NULL;
>   }
>   #endif
> 
> --
> 2.25.1
> 

Re: [PATCH 1/5] eventpoll: Convert epoll_put_uevent() to scoped user access

[Eric Dumazet]

On Tue, Mar 10, 2026 at 9:29 AM Christophe Leroy (CS GROUP)
<chleroy@kernel.org> wrote:
>
>
>
> Le 10/03/2026 à 08:54, renpanpan a écrit :
> > [Vous ne recevez pas souvent de courriers de renpanpan@kylinos.cn. Découvrez pourquoi ceci est important à https://aka.ms/LearnAboutSenderIdentification ]
> >
> > From: Eric Dumazet <edumazet@google.com>
> >
> > Saves two function calls, and one stac/clac pair.
> >
> > stac/clac is rather expensive on older cpus like Zen 2.
> >
> > A synthetic network stress test gives a ~1.5% increase of pps
> > on AMD Zen 2.
> >
> > Signed-off-by: Eric Dumazet <edumazet@google.com>
> > Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
> > Cc: Dave Hansen <dave.hansen@intel.com>
> > Cc: Kuniyuki Iwashima <kuniyu@google.com>
> > Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
> > ---
> >   include/linux/eventpoll.h | 11 +++++++----
> >   1 file changed, 7 insertions(+), 4 deletions(-)
> >
> > diff --git a/include/linux/eventpoll.h b/include/linux/eventpoll.h
> > index ccb478eb174b..ea9ca0e4172a 100644
> > --- a/include/linux/eventpoll.h
> > +++ b/include/linux/eventpoll.h
> > @@ -82,11 +82,14 @@ 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))
> > -               return NULL;
> > -
> > +       scoped_user_write_access_size(uevent, sizeof(*uevent), efault) {
>
> As already mentionned this could be simplified:

Note the patch was already merged in Linus tree.

Honestly having two different macros while we have 4 users for both
of them seems a bit overkill to me.

Re: [PATCH 1/5] eventpoll: Convert epoll_put_uevent() to scoped user access

[Christophe Leroy (CS GROUP)]

Le 10/03/2026 à 09:32, Eric Dumazet a écrit :
> On Tue, Mar 10, 2026 at 9:29 AM Christophe Leroy (CS GROUP)
> <chleroy@kernel.org> wrote:
>>
>>
>>
>> Le 10/03/2026 à 08:54, renpanpan a écrit :
>>> [Vous ne recevez pas souvent de courriers de renpanpan@kylinos.cn. Découvrez pourquoi ceci est important à https://aka.ms/LearnAboutSenderIdentification ]
>>>
>>> From: Eric Dumazet <edumazet@google.com>
>>>
>>> Saves two function calls, and one stac/clac pair.
>>>
>>> stac/clac is rather expensive on older cpus like Zen 2.
>>>
>>> A synthetic network stress test gives a ~1.5% increase of pps
>>> on AMD Zen 2.
>>>
>>> Signed-off-by: Eric Dumazet <edumazet@google.com>
>>> Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
>>> Cc: Dave Hansen <dave.hansen@intel.com>
>>> Cc: Kuniyuki Iwashima <kuniyu@google.com>
>>> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
>>> ---
>>>    include/linux/eventpoll.h | 11 +++++++----
>>>    1 file changed, 7 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/include/linux/eventpoll.h b/include/linux/eventpoll.h
>>> index ccb478eb174b..ea9ca0e4172a 100644
>>> --- a/include/linux/eventpoll.h
>>> +++ b/include/linux/eventpoll.h
>>> @@ -82,11 +82,14 @@ 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))
>>> -               return NULL;
>>> -
>>> +       scoped_user_write_access_size(uevent, sizeof(*uevent), efault) {
>>
>> As already mentionned this could be simplified:
> 
> Note the patch was already merged in Linus tree.
> 
> Honestly having two different macros while we have 4 users for both
> of them seems a bit overkill to me.

There have at the time beeing (v7.0-rc3) 66 caller of 
user_...access_begin(), I expect most of them being converted to scoped 
user access over time. In addition the simplicity of scoped user access 
should help grow the numbre of users of user access by block.

I find it quite convenient to not have to add the size when it can be 
extracted from the pointer type.

Christophe

Re: [PATCH 1/5] eventpoll: Convert epoll_put_uevent() to scoped user access

[David Laight]

On Tue, 10 Mar 2026 11:25:48 +0100
"Christophe Leroy (CS GROUP)" <chleroy@kernel.org> wrote:

> Le 10/03/2026 à 09:32, Eric Dumazet a écrit :
> > On Tue, Mar 10, 2026 at 9:29 AM Christophe Leroy (CS GROUP)
> > <chleroy@kernel.org> wrote:  
> >>
> >>
> >>
> >> Le 10/03/2026 à 08:54, renpanpan a écrit :  
> >>> [Vous ne recevez pas souvent de courriers de renpanpan@kylinos.cn. Découvrez pourquoi ceci est important à https://aka.ms/LearnAboutSenderIdentification ]
> >>>
> >>> From: Eric Dumazet <edumazet@google.com>
> >>>
> >>> Saves two function calls, and one stac/clac pair.
> >>>
> >>> stac/clac is rather expensive on older cpus like Zen 2.
> >>>
> >>> A synthetic network stress test gives a ~1.5% increase of pps
> >>> on AMD Zen 2.
> >>>
> >>> Signed-off-by: Eric Dumazet <edumazet@google.com>
> >>> Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
> >>> Cc: Dave Hansen <dave.hansen@intel.com>
> >>> Cc: Kuniyuki Iwashima <kuniyu@google.com>
> >>> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
> >>> ---
> >>>    include/linux/eventpoll.h | 11 +++++++----
> >>>    1 file changed, 7 insertions(+), 4 deletions(-)
> >>>
> >>> diff --git a/include/linux/eventpoll.h b/include/linux/eventpoll.h
> >>> index ccb478eb174b..ea9ca0e4172a 100644
> >>> --- a/include/linux/eventpoll.h
> >>> +++ b/include/linux/eventpoll.h
> >>> @@ -82,11 +82,14 @@ 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))
> >>> -               return NULL;
> >>> -
> >>> +       scoped_user_write_access_size(uevent, sizeof(*uevent), efault) {  
> >>
> >> As already mentionned this could be simplified:  
> > 
> > Note the patch was already merged in Linus tree.
> > 
> > Honestly having two different macros while we have 4 users for both
> > of them seems a bit overkill to me.  
> 
> There have at the time beeing (v7.0-rc3) 66 caller of 
> user_...access_begin(), I expect most of them being converted to scoped 
> user access over time. In addition the simplicity of scoped user access 
> should help grow the numbre of users of user access by block.
> 
> I find it quite convenient to not have to add the size when it can be 
> extracted from the pointer type.

Lets see if it possible to support both:
	foo(ptr, label)
and:
	foo(ptr, size, label)

#define foo(p, ...) foo_1(p, __VA_ARGS__, sizeof *p)
#define foo_1(p, s_or_l, l_or_ps, ...) \
	foo_2(p, __VA_OPT__(s_or_l, l_or_ps, foo_x(__VA_ARGS__)) l_or_ps, s_or_pl)
#define foo_x(one_arg)
#define foo_2(p, s, l, ...) foo_real(p, s, l)

untested - but might work!
(foo_x() is there to error foo(a, b, c, d))

	David

> 
> Christophe
> 

Re: [PATCH 1/5] eventpoll: Convert epoll_put_uevent() to scoped user access

[David Laight]

On Tue, 10 Mar 2026 09:32:53 +0100
Eric Dumazet <edumazet@google.com> wrote:

> On Tue, Mar 10, 2026 at 9:29 AM Christophe Leroy (CS GROUP)
> <chleroy@kernel.org> wrote:
> >
> >
> >
> > Le 10/03/2026 à 08:54, renpanpan a écrit :  
> > > [Vous ne recevez pas souvent de courriers de renpanpan@kylinos.cn. Découvrez pourquoi ceci est important à https://aka.ms/LearnAboutSenderIdentification ]
> > >
> > > From: Eric Dumazet <edumazet@google.com>
> > >
> > > Saves two function calls, and one stac/clac pair.
> > >
> > > stac/clac is rather expensive on older cpus like Zen 2.
> > >
> > > A synthetic network stress test gives a ~1.5% increase of pps
> > > on AMD Zen 2.
> > >
> > > Signed-off-by: Eric Dumazet <edumazet@google.com>
> > > Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
> > > Cc: Dave Hansen <dave.hansen@intel.com>
> > > Cc: Kuniyuki Iwashima <kuniyu@google.com>
> > > Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
> > > ---
> > >   include/linux/eventpoll.h | 11 +++++++----
> > >   1 file changed, 7 insertions(+), 4 deletions(-)
> > >
> > > diff --git a/include/linux/eventpoll.h b/include/linux/eventpoll.h
> > > index ccb478eb174b..ea9ca0e4172a 100644
> > > --- a/include/linux/eventpoll.h
> > > +++ b/include/linux/eventpoll.h
> > > @@ -82,11 +82,14 @@ 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))
> > > -               return NULL;
> > > -
> > > +       scoped_user_write_access_size(uevent, sizeof(*uevent), efault) {  
> >
> > As already mentionned this could be simplified:  
> 
> Note the patch was already merged in Linus tree.
> 
> Honestly having two different macros while we have 4 users for both
> of them seems a bit overkill to me.
> 

I'm also not sure all the 'hidden size' macros really help readability.
(especially of the header file that defines them all.)
It isn't as though the source lines would get overlong - especially
if the macro name was shorter.

But now is the time to change anything - before there are too many users.

	David