Re: LLD and glibc

[Fangrui Song <maskray@google.com>]

On 2021-10-25, Adhemerval Zanella wrote:
>I tracked down the issues with lld on glibc and it seems slight better
>than my previous attempt.
>
>To use lld without having to mess with my toolchain installation, I was
>trying to use -fuse-ld=lld plus -B option.  However it does not work
>correctly with the LIBC_PROG_BINUTILS macro, since it issues the
>passed CC with -print-prog-name=ld.  In this case gcc will always dump
>the ld path, not the lld (it does work if I use -print-prog-name=lld).
>I think this is a gcc issue since it should be agnostic whether
>-fuse-ld is used, but I think we can try to workaround it on glibc
>configure (use the -fuse-ld argument on -print-prog-name if the former
>is presented).
>
>So to actually use lld I had to adjust the links on my build-many-glibcs
>built toolchain to explicit point to lld (both bin/x86_64-glibc-linux-gnu-ld
>and x86_64-glibc-linux-gnu/bin/ld).
>
>The new issue is the same H.J has reported some time earlier:
>
>ld: error: cannot preempt symbol: foo_protected
>>>> defined in /home/azanella/Projects/glibc/build/x86_64-linux-gnu-lld/elf/ifuncmod5.so
>>>> referenced by ifuncmain5.c:22 (/home/azanella/Projects/glibc/glibc-git/elf/ifuncmain5.c:22)
>>>>               /home/azanella/Projects/glibc/build/x86_64-linux-gnu-lld/elf/ifuncmain5.o:(get_foo_protected)
>
>I think we should just disable the tests ldd does not support (with the
>patch below) instead of get back of the discussion whether it should be
>supported or not.
>
>The good news is with release clang+llvm 13 and with the testcase patch
>I could build and run make check successfully without any regression.
>
>I am writing this because we need to fix both issues (configure with -fuse-ld
>and the failure on make check) so we can start review the DT_RELR patchset [1]
>
>[1] https://patchwork.sourceware.org/project/glibc/list/?series=3895
>
>---
>
>diff --git a/elf/Makefile b/elf/Makefile
>index bdcf4cb885..7ddc4b39a3 100644
>--- a/elf/Makefile
>+++ b/elf/Makefile
>@@ -427,11 +427,14 @@ ifeq (yes,$(build-shared))
> tests += tst-ifunc-fault-lazy tst-ifunc-fault-bindnow
> # Note: sysdeps/x86_64/ifuncmain8.c uses ifuncmain8.
> tests-internal += \
>-        ifuncmain1 ifuncmain1pic ifuncmain1vis ifuncmain1vispic \
>+        ifuncmain1pic ifuncmain1vis ifuncmain1vispic \
>         ifuncmain1staticpic \
>         ifuncmain2 ifuncmain2pic ifuncmain3 ifuncmain4 \
>-        ifuncmain5 ifuncmain5pic ifuncmain5staticpic \
>+        ifuncmain5pic ifuncmain5staticpic \
>         ifuncmain7 ifuncmain7pic
>+ifeq (no,$(with-lld))
>+test-internal += ifuncmain1 ifuncmain5
>+endif
> ifeq (yes,$(have-gcc-ifunc))
> tests-internal += ifuncmain9 ifuncmain9pic
> endif
>@@ -451,7 +454,10 @@ endif
> tests-internal += $(ifunc-pie-tests)
> tests-pie += $(ifunc-pie-tests)
> endif
>-modules-names += ifuncmod1 ifuncmod3 ifuncmod5 ifuncmod6
>+modules-names += ifuncmod1 ifuncmod3 ifuncmod6
>+ifeq (no,$(with-lld))
>+modules-names += ifuncmod5 >+endif
> endif
> endif
>
>
>

ifuncmain1.c and ifuncmain5.c take the address of the function `foo_protected`.
If compiled with -fno-pic, the produced relocation is R_X86_64_PC32.

In gold and LLD, it is an error to reference an undefined function symbol
with a non-GOT-generating non-PLT-generating relocation from the main
executable (LLD call it "cannot preempt symbol: foo_protected").
This is to guard against risky canonical PLT entry which cannot work
with a protected definition
(https://maskray.me/blog/2021-01-09-copy-relocations-canonical-plt-entries-and-protected).

(I wonder why the test can link with GNU ld. We know it has some limited
support for "copy relocations on protected data symbols" "on x86 but I
don't know it does anything with function symbols. Anyway, protected
function symbols are quite broken in GCC that is why ifuncmod5.c uses
inline assembly instead of proper visibility attribute.)

I could not reproduce the link error because on most Linux
distributions, GCC is configured with enable-default-pie and
ifuncmain5.c is compiled with the implicit -fpie. If I manually compile
ifuncmain5.c with -fno-pic and re-link ifuncmain5 with LLD, I can see
the intended diagnostic "cannot preempt symbol: foo_protected".

You may want to take some of the description to make up a better commit
message.

Reviewed-by: Fangrui Song <maskray@google.com>