Re: [PATCH 01/13] compiler.h: Split {READ, WRITE}_ONCE definitions out into rwonce.h

On Mon, Nov 11, 2019 at 9:10 AM Christian Borntraeger
[off-list ref] wrote:

On 08.11.19 20:57, Arnd Bergmann wrote:

quoted

On Fri, Nov 8, 2019 at 6:01 PM Will Deacon [off-list ref] wrote:

quoted

In preparation for allowing architectures to define their own
implementation of the 'READ_ONCE()' macro, move the generic
'{READ,WRITE}_ONCE()' definitions out of the unwieldy 'linux/compiler.h'
and into a new 'rwonce.h' header under 'asm-generic'.

Adding Christian Bornträger to Cc, he originally added the
READ_ONCE()/WRITE_ONCE()
code.

I wonder if it would be appropriate now to revert back to a much simpler version
of these helpers for any modern compiler. As I understand, only gcc-4.6 and
gcc4.7 actually need the song-and-dance version with the union and switch/case,
while for others, we can might be able back to a macro doing a volatile access.

As far as I know this particular issue with  volatile access on aggregate types
was fixed in gcc 4.8. On the other hand we know that the current construct will
work on all compilers. Not so sure about the orignal ACCESS_ONCE implementation.

I've seen problems with clang on the current version, leading to unnecessary
temporaries being spilled to the stack in some cases, so I think it would still
help to simplify it.

We probably don't want the exact ACCESS_ONCE() implementation back
that existed before, but rather something that implements the stricter
READ_ONCE() and WRITE_ONCE(). I'd probably also want to avoid the
__builtin_memcpy() exception for odd-sized accesses and instead have
a separate way to do those.

      Arnd
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization