rc_config_directory and rc_config_set_value introduced in c09eeca49145b0 fail to pass test on hidden functions on ppc64el[1] and mips64el[2]. The reason is unclear. But some clue is found by Dmitry Yu Okunev > On "ppc64el" static function compiled with "gcc" is visible by "runtests.sh" > (that uses "readelf"), but it's not on amd64. 1. https://buildd.debian.org/status/fetch.php?pkg=openrc&arch=ppc64el&ver=0.20.4-1&stamp=1456247204 2. https://buildd.debian.org/status/fetch.php?pkg=openrc&arch=mips64el&ver=0.20.4-1&stamp=1456251406 Reproducible: Always
As shown in this bug, this is causing OpenRC's test suite to fail for these platforms. Do you have any suggestions? Thanks, William
i can't reproduce w/mips64. would help if someone can attach a tarball of the failing objects.
Created attachment 506796 [details] Build log (ppc64) We're hitting this on ppc64 (*not* ppc64le) in Adélie, which is causing the aforementioned test failure. Attached is the build log. Next attachment will be the build tree. We use the Alpine build tools instead of Portage, and this is the contents of the /etc/abuild.conf on the ppc64 builder (like /etc/portage/make.conf): export CFLAGS="-O2 -ggdb -mcpu=970 -mtune=power9 -maltivec -mlong-double-64 -fno-inline-small-functions" export CXXFLAGS="$CFLAGS" export CPPFLAGS="$CFLAGS" export LDFLAGS="-Wl,--as-needed" export JOBS=2 export MAKEFLAGS=-j$JOBS USE_COLORS=1 SRCDEST=/var/cache/distfiles REPODEST=/srv/packages/ export OPTIONS="!strip" export DEFAULT_DBG="YesPlease" export DISTRO_NAME="Adélie Linux" export DISTRO_BUG_URL="https://bts.adelielinux.org/"
As requested by vapier@, a tarball of the build directory with object files is available at http://foxkit.us/linux/openrc-0.24.1-build.txz – unfortunately, it is over the 1,000 KB limit for Bugzilla attachments. This should resolve the reason that this bug was marked RESO NEEDINFO, so I am changing the status back to CONFIRMED.
(In reply to A. Wilcox (awilfox) from comment #4) > As requested by vapier@, a tarball of the build directory with object files > is available at http://foxkit.us/linux/openrc-0.24.1-build.txz – 404 Not Found.
Created attachment 565258 [details] Build objects The file has been re-uploaded to my host. I also used more extreme xz options to make it fit under the 1000 KB limit so it can be hosted on BZ as well. Please reopen this bug. It is still very much a bug.
Please paste 'emerge --info' from the affected system as well.
(In reply to A. Wilcox (awilfox) from comment #6) > Created attachment 565258 [details] > Build objects Thanks, that helps. You have two types of test failures: 1. "exported" libc symbols 2. "exported" rc_* symbols TL;DR: Both are openrc bugs and not toolchain bugs. openrc should fix (or better yet remove) tests that sed through readelf output. More details: I'll focus on "exported" rc_* symbols here as they are reported in original bug for openrc-0.24.1: > * Checking hidden functions in librc.so ... > * Missing hidden defs: > rc_config_directory > rc_config_set_value > rc_parse_service_state Here is a test definition: https://github.com/OpenRC/openrc/blob/0.24.x/src/test/runtests.sh#L55 ```sh ebegin "Checking hidden functions in librc.so" sed -n '/^librc_hidden_proto/s:.*(\(.*\))$:\1:p' ${librc_srcdir}/librc.h \ | LC_ALL=C sort -u \ > librc.funcs.hidden.list readelf -Wr $(grep -l '#include[[:space:]]"librc\.h"' ${librc_srcdir}/*.c | sed 's:\.c$:.o:') \ | egrep -v -e 'R_PARISC_(DP|SEG)REL' \ | awk '$5 ~ /^rc_/ {print $5}' \ | LC_ALL=C sort -u \ | egrep -v '^rc_environ_fd$' \ > librc.funcs.hidden.out syms=$(diff -u librc.funcs.hidden.list librc.funcs.hidden.out | sed -n '/^+[^+]/s:^+::p') [ -z "${syms}" ] eend $? "Missing hidden defs:"$'\n'"${syms}" ``` A few things to note here: 1. the proxy of exported symbols is output of readelf on object files, not a final .so library. 2. the proxy of exported symbol is a presence of relocation for a symbol that has 'rc_*' name. 3. this script does not really check if 'rc_*' symbols are local or not. To double-check it's not a hidden but local symbol you can looks at symbol type: This is supposed to be local: openrc-0.24.1/src/test $ LANG=C readelf -a ../librc/librc-misc.o | fgrep rc_config_set 000000000dd0 000a0000000a R_PPC64_REL24 00000000000001b0 rc_config_set_value + 0 000000000ec8 000a0000000a R_PPC64_REL24 00000000000001b0 rc_config_set_value + 0 10: 00000000000001b0 648 FUNC LOCAL DEFAULT [<localentry>: 8] 1 rc_config_set_value This is supposed to be hidden: openrc-0.24.1/src/test $ LANG=C readelf -a ../librc/librc-misc.o | fgrep rc_config_list 000000000d94 003b0000000a R_PPC64_REL24 0000000000000a80 __RC_rc_config_list + 0 000000000ea0 003b0000000a R_PPC64_REL24 0000000000000a80 __RC_rc_config_list + 0 59: 0000000000000a80 496 FUNC GLOBAL HIDDEN [<localentry>: 8] 1 __RC_rc_config_list 62: 0000000000000a80 496 FUNC GLOBAL DEFAULT [<localentry>: 8] 1 rc_config_list There is no discrepancy here. Test script should be fixed. Now, why it's not failing for everyone then? I don't know. My guess it's because of gcc-6.4.0 applying different inlining heuristics than gcc-7.3.0/gcc-8.2.0 and a combination of passed CFLAGS: -O2 -ggdb -mcpu=970 -mtune=power9 -maltivec -mlong-double-64 -fno-inline-small-functions Or maybe your gcc is ubilt differently WRT pie, stack protection, etc. Hard to say without looking at 'gcc -v'. Normally 'rc_config_set_value' can get inlined in compiler frontend as it's a static local function https://github.com/OpenRC/openrc/blob/0.24.x/src/librc/librc-misc.c#L207
Sent test removal for review as https://github.com/OpenRC/openrc/pull/292
These tests have been removed since they do not work well cross platform. William
