Re: [PATCH v9 09/27] rust: add `compiler_builtins` crate

From: Gary Guo <gary@garyguo.net>
To: Nick Desaulniers <ndesaulniers@google.com>
Cc: "Miguel Ojeda" <ojeda@kernel.org>,
	"Linus Torvalds" <torvalds@linux-foundation.org>,
	"Greg Kroah-Hartman" <gregkh@linuxfoundation.org>,
	rust-for-linux@vger.kernel.org, linux-kernel@vger.kernel.org,
	linux-fsdevel@vger.kernel.org, patches@lists.linux.dev,
	"Jarkko Sakkinen" <jarkko@kernel.org>,
	"Alex Gaynor" <alex.gaynor@gmail.com>,
	"Wedson Almeida Filho" <wedsonaf@google.com>,
	"Sven Van Asbroeck" <thesven73@gmail.com>,
	"Boqun Feng" <boqun.feng@gmail.com>,
	"Björn Roy Baron" <bjorn3_gh@protonmail.com>,
	"Masahiro Yamada" <masahiroy@kernel.org>,
	"Ard Biesheuvel" <ardb@kernel.org>
Subject: Re: [PATCH v9 09/27] rust: add `compiler_builtins` crate
Date: Mon, 29 Aug 2022 18:11:33 +0100	[thread overview]
Message-ID: <20220829181133.00005f01@garyguo.net> (raw)
In-Reply-To: <CAKwvOdkc0Qhwu=gfe1+H23TnAa6jnO6A3ZCO687dH6mSrATmDA@mail.gmail.com>

On Wed, 24 Aug 2022 11:38:46 -0700
Nick Desaulniers <ndesaulniers@google.com> wrote:

> Another idea I had, and Ard mentioned to me this morning that efistub
> does something similar:
> 
> I think objcopy can rename symbol references.  So instead of calling
> foo, I think you can use objcopy to call bar instead.
> 
> To avoid the case of Rust core refering to symbols typically provided
> by the --rtlib={libgcc|compiler-rt}, what if we:
> 1. build core.o (with reference to the unfavorable symbols)
> 2. use objcopy to rename references in core.o to the symbols listed in
> this file; maybe prepend them with `rust_core_` or something.
> 3. provide this crate with panic'ing definitions, but for the renamed
> symbols.
> 
> That way, enabling CONFIG_RUST=y won't hide the case where C code is
> triggering such libcalls from the C compiler?
> 
> To restate the concern, I don't want CONFIG_RUST=y to hide that fact
> that someone wrote `float x, y; if (x != y) ...` in C code, where the
> C compiler will likely lower that to a libcall to __nesf2 since with
> the current approach in this patch, linkage would proceed. The current
> behavior is to intentionally fail to link.  I think what I describe
> above may work to keep this intended behavior of the kernel's build
> system.
> 
> Ah, looks like there's objcopy flags:
>        --redefine-sym old=new
>            Change the name of a symbol old, to new.  This can be
> useful when one is trying link
>            two things together for which you have no source, and there
> are name collisions.
> 
>        --redefine-syms=filename
>            Apply --redefine-sym to each symbol pair "old new" listed
> in the file filename.
>            filename is simply a flat file, with one symbol pair per
> line.  Line comments may be
>            introduced by the hash character.  This option may be given
> more than once.
> 
> So probably could start with my diff from above, but use
> --redefine-syms=filename instead of --strip-symbols=.
> 
> --
> Thanks,
> ~Nick Desaulniers

This is the approach we are about to take, see
https://github.com/Rust-for-Linux/linux/pull/779.

It's easy for just one symbol that is known to be not defined, but it
can be more complex as some of these symbols can be defined on a
platform but not another, so we would have to generate a dynamic list
of what symbols to "hijack" depending on the config. Currently we
avoid this issue by having all symbols in compiler_builtins
crate weak, but we couldn't weakly redefine a symbol.

It's certainly doable, though, but it require some effort. I am
not quite available recently so hadn't further extend and polish my
patch.

Best,
Gary