Re: C aggregate passing (Rust kernel policy)

[Paul E. McKenney]

On Wed, Feb 26, 2025 at 10:54:12PM +0000, David Laight wrote:
> On Wed, 26 Feb 2025 13:21:41 -0800
> Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> 
> > On Wed, 26 Feb 2025 at 13:14, Linus Torvalds
> > <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > That "single read done as multiple reads" is sadly still accepted by
> > > the C standard, as far as I can tell. Because the standard still
> > > considers it "unobservable" unless I've missed some update.  
> > 
> > I want to clarify that I'm talking about perfectly normal and entirely
> > unannotated variable accesses.
> > 
> > Don't say "programmers should annotate their special accesses with
> > volatile if they want to avoid compiler-introduced TOCTOU issues".
> > 
> > Having humans have to work around failures in the language is not the way to go.
> > 
> > Particularly when there isn't even any advantage to it. I'm pretty
> > sure neither clang nor gcc actually rematerialize reads from memory,
> 
> I thought some of the very early READ_ONCE() were added because there
> was an actual problem with the generated code.
> But it has got entirely silly.
> In many cases gcc will generate an extra register-register transfer
> for a volatile read - I've seen it do a byte read, register move and
> then and with 0xff.
> I think adding a separate memory barrier would stop the read being
> rematerialized - but you also need to stop it doing (for example)
> two byte accesses for a 16bit variable - arm32 has a limited offset
> for 16bit memory accesses, so the compiler might be tempted to do
> two byte writes.

Perhaps some day GCC __atomic_load_n(__ATOMIC_RELAXED) will do what we
want for READ_ONCE().  Not holding my breath, though.  ;-)

							Thanx, Paul