On Tue, Nov 17, 2020 at 08:09:22PM -0500, Steven Rostedt wrote:
> On Tue, 17 Nov 2020 16:42:44 -0800
> Matt Mullins <mmull...@mmlx.us> wrote:
>
>
> > > Indeed with a stub function, I don't see any need for
> > > READ_ONCE/WRITE_ONCE.
> >
> > I'm not sure if this is a practical issue, but without WRITE_ONCE, can't
> > the write be torn? A racing __traceiter_ could potentially see a
> > half-modified function pointer, which wouldn't work out too well.
>
> This has been discussed before, and Linus said:
>
> "We add READ_ONCE and WRITE_ONCE annotations when they make sense. Not
> because of some theoretical "compiler is free to do garbage"
> arguments. If such garbage happens, we need to fix the compiler"
>
> https://lore.kernel.org/lkml/CAHk-=wi_KeD1M-_-_SU_H92vJ-yNkDnAGhAS=rr1ynngwkw...@mail.gmail.com/
I have to ask... Did the ARM compilers get fixed? As of a few
months ago, they would tear stores of some constant values.
> > This was actually my gut instinct before I wrote the __GFP_NOFAIL
> > instead -- currently that whole array's memory ordering is provided by
> > RCU and I didn't dive deep enough to evaluate getting too clever with
> > atomic modifications to it.
>
> The pointers are always going to be the architecture word size (by
> definition), and any compiler that tears a write of a long is broken.
But yes, if the write is of a non-constant pointer, the compiler does
have less leverage.
Thanx, Paul
