On Tue, Oct 27, 2020 at 09:15:17PM +0100, Ard Biesheuvel wrote:
> On Tue, 27 Oct 2020 at 21:12, Nick Desaulniers <ndesaulniers@xxxxxxxxxx> wrote:
> >
> > On Tue, Oct 27, 2020 at 12:25 PM Geert Uytterhoeven
> > <geert@xxxxxxxxxxxxxx> wrote:
> > >
> > > Hi Nick,
> > >
> > > CC Josh
> > >
> > > On Mon, Oct 26, 2020 at 6:49 PM Nick Desaulniers
> > > <ndesaulniers@xxxxxxxxxx> wrote:
> > > > On Mon, Oct 26, 2020 at 10:44 AM Geert Uytterhoeven
> > > > <geert@xxxxxxxxxxxxxx> wrote:
> > > > > On Mon, Oct 26, 2020 at 6:39 PM Ard Biesheuvel <ardb@xxxxxxxxxx> wrote:
> > > > > > On Mon, 26 Oct 2020 at 17:01, Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
> > > > > > > On Mon, Oct 26, 2020 at 2:29 PM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
> > > > > > > > On Mon, Oct 26, 2020 at 1:29 PM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
> > > > > > > > > I.e. including the ".eh_frame" warning. I have tried bisecting that
> > > > > > > > > warning (i.e. with be2881824ae9eb92 reverted), but that leads me to
> > > > > > > > > commit b3e5d80d0c48c0cc ("arm64/build: Warn on orphan section
> > > > > > > > > placement"), which is another red herring.
> > > > > > > >
> > > > > > > > kernel/bpf/core.o is the only file containing an eh_frame section,
> > > > > > > > causing the warning.
> > > >
> > > > When I see .eh_frame, I think -fno-asynchronous-unwind-tables is
> > > > missing from someone's KBUILD_CFLAGS.
> > > > But I don't see anything curious in kernel/bpf/Makefile, unless
> > > > cc-disable-warning is somehow broken.
> > >
> > > I tracked it down to kernel/bpf/core.c:___bpf_prog_run() being tagged
> > > with __no_fgcse aka __attribute__((optimize("-fno-gcse"))).
> > >
> > > Even if the function is trivially empty ("return 0;"), a ".eh_frame" section
> > > is generated. Removing the __no_fgcse tag fixes that.
> >
> > That's weird. I feel pretty strongly that unless we're working around
> > a well understood compiler bug with a comment that links to a
> > submitted bug report, turning off rando compiler optimizations is a
> > terrible hack for which one must proceed straight to jail; do not pass
> > go; do not collect $200. But maybe I'd feel differently for this case
> > given the context of the change that added it. (Ard mentions
> > retpolines+orc+objtool; can someone share the relevant SHA if you have
> > it handy so I don't have to go digging?)
>
> commit 3193c0836f203a91bef96d88c64cccf0be090d9c
> Author: Josh Poimboeuf <jpoimboe@xxxxxxxxxx>
> Date: Wed Jul 17 20:36:45 2019 -0500
>
> bpf: Disable GCC -fgcse optimization for ___bpf_prog_run()
>
> has
>
> Fixes: e55a73251da3 ("bpf: Fix ORC unwinding in non-JIT BPF code")
That commit is broken.
I had this patch in my queue:
-#define __no_fgcse __attribute__((optimize("-fno-gcse")))
+#define __no_fgcse __attribute__((optimize("-fno-gcse,-fno-omit-frame-pointer")))
Sounds like you want to add -fno-asynchronous-unwind-tables to the above list?
> and mentions objtool and CONFIG_RETPOLINE.
>
> > (I feel the same about there
> > being an empty asm(); statement in the definition of asm_volatile_goto
> > for compiler-gcc.h). Might be time to "fix the compiler."
> >
> > (It sounds like Arvind is both in agreement with my sentiment, and has
> > the root cause).
> >
>
> I agree that the __no_fgcse hack is terrible. Does Clang support the
> following pragmas?
>
> #pragma GCC push_options
> #pragma GCC optimize ("-fno-gcse")
> #pragma GCC pop_options
That will work too, but optimize("-fno...,-fno..,-fno..") is imo cleaner.
