Re: [PATCH v2] x86: bring back rep movsq for user access on CPUs without ERMS

[Ingo Molnar <mingo@kernel.org>]

* Linus Torvalds <torvalds@linux-foundation.org> wrote:

> On Sun, 3 Sept 2023 at 13:49, Mateusz Guzik <mjguzik@gmail.com> wrote:
> >
> > "real fstat" is syscall(5, fd, &sb).
> >
> > Sapphire Rapids, will-it-scale, ops/s
> >
> > stock fstat     5088199
> > patched fstat   7625244 (+49%)
> > real fstat      8540383 (+67% / +12%)
> >
> > It dodges lockref et al, but it does not dodge SMAP which accounts for
> > the difference.
> 
> Side note, since I was looking at this, I hacked up a quick way for
> architectures to do their own optimized cp_new_stat() that avoids the
> double-buffering.
> 
> Sadly it *is* architecture-specific due to padding and
> architecture-specific field sizes (and thus EOVERFLOW rules), but it
> is what it is.
> 
> I don't know how much it matters, but it might make a difference. And
> 'stat()' is most certainly worth optimizing for, even if glibc has
> made our life more difficult.
> 
> Want to try out another entirely untested patch? Attached.
> 
>                 Linus

>  arch/x86/kernel/sys_x86_64.c | 44 ++++++++++++++++++++++++++++++++++++++++++++
>  fs/stat.c                    |  2 +-
>  include/linux/stat.h         |  2 ++
>  3 files changed, 47 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/x86/kernel/sys_x86_64.c b/arch/x86/kernel/sys_x86_64.c
> index c783aeb37dce..fca647f61bc1 100644
> --- a/arch/x86/kernel/sys_x86_64.c
> +++ b/arch/x86/kernel/sys_x86_64.c
> @@ -22,6 +22,50 @@
>  #include <asm/elf.h>
>  #include <asm/ia32.h>
>  
> +int cp_new_stat(struct kstat *stat, struct stat __user *ubuf)
> +{
> +	typeof(ubuf->st_uid) uid;
> +	typeof(ubuf->st_gid) gid;
> +	typeof(ubuf->st_dev) dev = new_encode_dev(stat->dev);
> +	typeof(ubuf->st_rdev) rdev = new_encode_dev(stat->rdev);
> +
> +	SET_UID(uid, from_kuid_munged(current_user_ns(), stat->uid));
> +	SET_GID(gid, from_kgid_munged(current_user_ns(), stat->gid));
> +
> +	if (!user_write_access_begin(ubuf, sizeof(struct stat)))
> +		return -EFAULT;
> +
> +	unsafe_put_user(dev,			&ubuf->st_dev,		Efault);
> +	unsafe_put_user(stat->ino,		&ubuf->st_ino,		Efault);
> +	unsafe_put_user(stat->nlink,		&ubuf->st_nlink,	Efault);
> +
> +	unsafe_put_user(stat->mode,		&ubuf->st_mode,		Efault);
> +	unsafe_put_user(uid,			&ubuf->st_uid,		Efault);
> +	unsafe_put_user(gid,			&ubuf->st_gid,		Efault);
> +	unsafe_put_user(0,			&ubuf->__pad0,		Efault);
> +	unsafe_put_user(rdev,			&ubuf->st_rdev,		Efault);
> +	unsafe_put_user(stat->size,		&ubuf->st_size,		Efault);
> +	unsafe_put_user(stat->blksize,		&ubuf->st_blksize,	Efault);
> +	unsafe_put_user(stat->blocks,		&ubuf->st_blocks,	Efault);
> +
> +	unsafe_put_user(stat->atime.tv_sec,	&ubuf->st_atime,	Efault);
> +	unsafe_put_user(stat->atime.tv_nsec,	&ubuf->st_atime_nsec,	Efault);
> +	unsafe_put_user(stat->mtime.tv_sec,	&ubuf->st_mtime,	Efault);
> +	unsafe_put_user(stat->mtime.tv_nsec,	&ubuf->st_mtime_nsec,	Efault);
> +	unsafe_put_user(stat->ctime.tv_sec,	&ubuf->st_ctime,	Efault);
> +	unsafe_put_user(stat->ctime.tv_nsec,	&ubuf->st_ctime_nsec,	Efault);
> +	unsafe_put_user(0,			&ubuf->__unused[0],	Efault);
> +	unsafe_put_user(0,			&ubuf->__unused[1],	Efault);
> +	unsafe_put_user(0,			&ubuf->__unused[2],	Efault);

/me performs happy dance at seeing proper use of vertical alignment in 
bulk-assignments.

If measurements support it then this looks like a nice optimization.

Thanks,

	Ingo