Date:   Sat, 14 Aug 2021 12:34:15 -0700
From:   Fāng-ruì Sòng <maskray@...gle.com>
To:     Segher Boessenkool <segher@...nel.crashing.org>
Cc:     Daniel Axtens <dja@...ens.net>,
        Nick Desaulniers <ndesaulniers@...gle.com>,
        linux-kernel@...r.kernel.org,
        Nathan Chancellor <nathan@...nel.org>,
        clang-built-linux@...glegroups.com,
        Paul Mackerras <paulus@...ba.org>,
        Bill Wendling <morbo@...gle.com>, linuxppc-dev@...ts.ozlabs.org
Subject: Re: [PATCH] ppc: add "-z notext" flag to disable diagnostic

On Sat, Aug 14, 2021 at 5:59 AM Segher Boessenkool
<segher@...nel.crashing.org> wrote:
>
> On Fri, Aug 13, 2021 at 01:05:08PM -0700, Fangrui Song wrote:
> > Text relocations are considered very awful by linker developers.
>
> By very few linker developers.

https://groups.google.com/g/generic-abi/c/Ckq19PfLxyk/m/uW29sgkoAgAJ
Ali Bahrami: "My opinion is that no one wants text relocations, but
not labeling them if they do exist doesn't seem right. I find the
presence of DF_TEXTREL very interesting when diagnosing various
issues."

https://gcc.gnu.org/legacy-ml/gcc/2016-04/msg00138.html
( "So why not simply create 'dynamic text relocations' then?  Is that
possible with a pure linker change?" )
Cary Coutant: "Ugh. Besides being a bad idea from a performance point
of view, it's not even always possible to do. Depending on the
architecture, a direct reference from an executable to a variable in a
shared library may not have the necessary reach."

binutils-gdb commit "Add linker option: --warn-shared-textrel to
produce warnings when adding a DT_TEXTREL to a shared object."
Nick Clifton

https://www.openwall.com/lists/musl/2020/09/26/3
Szabolcs Nagy: "nice.  and gcc passes -z text for static pie code so
that case should not end up with text rels."

Someone wrote "Overall, the overhead of processing text relocations
can cause serious performance degradation." in Solaris' Linker and
Libraries Guide.

Me :)
(I wrote lld/ELF 9.0.0~13.0.0 release notes and filed dozen of GNU
ld/gold bugs/feature requests)

> > binutils 2.35 added --enable-textrel-check={no,warn,error}
> > https://sourceware.org/bugzilla/show_bug.cgi?id=20824
>
> Yes, some people wanted the default to be configurable.  So now we have
> a default default that is sane, so most people get to reap the benefits
> of having defaults at all, but we also allow other people to shoot
> themselves (and people who have to deal with them) in the foot.
> "Progress".  Changing the defaults should be a one-time event, only done
> when the benefits strongly outweigh the costs.  Defaults should never be
> configurable (by the user).

ld.lld has such a non-configurable model and thus caught the issue
(which the patch intends to address).

Future --enable-textrel-check={yes,error} configured GNU ld will be similar.

> > I can imagine that in the future some Linux distributions (especially those
> > focusing on security) will default their binutils to use
> > --enable-textrel-check={no,warn,error}.
>
> How would this be a benefit to security?

https://flameeyes.blog/2016/01/16/textrels-text-relocations-and-their-impact-on-hardening-techniques/

https://github.com/golang/go/issues/9210
Android: "libexample.so has text relocations. This is wasting memory
and prevents security hardening. Please fix."

FWIW I contributed a glibc patch allowing TEXTREL to co-exist with ifunc.
It requires temporary mapping the text segment W^X.

> > In -no-pie links, R_PPC64_ADDR64 relocations are link-time constants.
>
> Where "link" includes dynamic links as well.  There are no constants.

"Link-time" usually refers to the processing of the static linker.

Dynamic links can use load-time or run-time.

> > There are no text relocations, therefore no need for -z notext.
>
> This is a choice by the compiler, nothing more.  It saves some process
> startup time, and allows slightly more maps to be shared by processes
> that run the same images.  But it is a tradeoff, so it might change; and
> of course it is not an ABI requirement.
>
>
> Segher

Text relocations are generally awful.
GNU ld and gold's traditional "add DF_TEXTREL on-demand" behavior made
such user errors easy to make.
I understand that kernels are special applications where we apply
relocations once and many user-space objection can be less of a
concern/ignored.
However, the Linux kernel is already in a position where many linker
options are controlled and thus should specify -z notext to make
the intention explicit, or fix the problems (I think x86-64 is good;
that said, powerpc
has a higher cost using PC-relative instructions so pay the oneshot relocation
time cost probably isn't a bad choice)



--
宋方睿

Powered by blists - more mailing lists