Re: [PATCH] rwonce: fix crash by removing READ_ONCE() for unaligned read

[Jann Horn]

On Wed, Mar 26, 2025 at 11:41 PM Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> On Wed, 26 Mar 2025 at 14:24, Arnd Bergmann <arnd@xxxxxxxx> wrote:
> >
> > I've applied this on top of the asm-generic branch, but I just sent
> > the pull request with the regression to Linus an hour ago.
> >
> > I'll try to get a new pull request out tomorrow.
>
> I will ignore that pull request, and wait for your updated one.
>
> That said, this whole thing worries me. The fact that the compiler
> magically makes READ_ONCE() require alignment that it normally doesn't
> require seems like a bug waiting to happen somewhere else.

To be clear, this is not the compiler's doing. The arch-specific arm64
code defines a custom version of the __READ_ONCE() macro used by
READ_ONCE() in LTO builds:

https://web.git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/arm64/include/asm/rwonce.h

/*
 * When building with LTO, there is an increased risk of the compiler
 * converting an address dependency headed by a READ_ONCE() invocation
 * into a control dependency and consequently allowing for harmful
 * reordering by the CPU.
 *
 * Ensure that such transformations are harmless by overriding the generic
 * READ_ONCE() definition with one that provides RCpc acquire semantics
 * when building with LTO.
 */