On Tue, Oct 27, 2020 at 1:32 PM Arvind Sankar [off-list ref] wrote:

On Tue, Oct 27, 2020 at 01:28:02PM -0700, Nick Desaulniers wrote:

quoted

quoted

commit 3193c0836f203a91bef96d88c64cccf0be090d9c
Author: Josh Poimboeuf [off-list ref]
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")

and mentions objtool and CONFIG_RETPOLINE.

Thanks for the context.  It might be time to revisit the above commit.
If I revert it (small conflict that's easy to fixup),
kernel/bpf/core.o builds cleanly with defconfig+GCC-9.3, so maybe
obtool did get smart enough to handle that case?  Probably regresses
the performance of that main dispatch loop for BPF, but not sure what
folks are expecting when retpolines are enabled.
--
Thanks,
~Nick Desaulniers

The objtool issue was with RETPOLINE disabled.

Ah, sorry, in that case default-CONFIG_RETPOLINE+gcc-9.3:
kernel/bpf/core.o: warning: objtool: ___bpf_prog_run()+0x8d4: sibling
call from callable instruction with modified stack frame
-- 
Thanks,
~Nick Desaulniers