Re: Attempting to unify some user space copying APIs

[Vasily Gorbik <gor@linux.ibm.com>]

On Fri, Mar 27, 2026 at 05:20:57PM -0700, Linus Torvalds wrote:
> I'm sending this out to the architecture maintainer list (and a random
> collection of arch maintainers that I just took from "recent pull
> requests") in the hope that people can test this on platforms that I'm
> too incompetent to check myself (read: I haven't installed all the
> cross-compiler environments).
...
>  arch/arm64/include/asm/uaccess.h   | 36 ++++++++++++++----------------
>  arch/powerpc/include/asm/uaccess.h |  4 ++--
>  arch/powerpc/lib/pmem.c            |  4 ++--
>  arch/x86/include/asm/uaccess_64.h  | 13 +++++------
>  arch/x86/lib/usercopy_64.c         |  2 +-
>  lib/iov_iter.c                     | 19 +++++++++++-----
>  6 files changed, 42 insertions(+), 36 deletions(-)

On s390 this builds and runs fine.

One thing I noticed: copy_from_user_iter is clean - the boilerplate just
wraps raw_copy_from_user which has no instrumentation of its own.

copy_from_user_iter_nocache is a bit different: the boilerplate wraps
the actor with instrument_copy_from_user_before/after, but the nocache
actors already instrument internally.

On s390 (and all arches without ARCH_HAS_NOCACHE_UACCESS) the generic
__copy_from_user_inatomic_nocache calls __copy_from_user_inatomic
which does its own instrument_copy_from_user_before/after around
raw_copy_from_user, so we end up with double kasan/kcsan/kmsan calls.

On x86 __copy_from_user_inatomic_nocache only does a bare
kasan_check_write(), so that one gets doubled, while kcsan and kmsan are
newly added through the boilerplate.

I only looked at s390 and compared to x86.