GCC 10 in Gentoo

March 23, 2020

New gcc releases are always great fun to have: new optimisations, new CPUs and platforms supported, new diagnostic warnings. But also occasional new bugs or fixes that break backwards compatibility. I was wondering the other day: what does it take to get live gcc right from git tree into gentoo. That would allow people to try gcc in VMs or chroots in a full gentoo environment. We would be able to find out breakages before the release comes out, would allow users to test early fixes and (most important) have another way to break our systems.

I cleaned up toolchain.eclass a bit and made live gcc-10 ebuild. It should be as easy to create live ebuild for stable gcc branches. You just need to copy the ebuild and drop EGIT_BRANCH=master line. toolchain.eclass should do the right thing. To test that ebuild actually works I decided to switch my main desktop to live gcc form master to see how good/bad it is.

gcc-config bug

First thing to break was sys-devel/gcc-config tool. gcc-config allows you to switch compilers at runtime by juggling /usr/bin/gcc symlink (and a few nearby symlinks). gcc-config previously assumed that gcc versions could be lexicographically sorted (used sort tool). As a result it always picked gcc-9 as the latest even when gcc-10 was present.

gcc-config is also a low-dependency tool written in bash mostly to allow you to recover system from broken state with inactive gcc. It has no direct access to advanced sorting functions, but it needs to solve a simple problem of ordering gcc versions (like 9.2.0, 10.0.1, 11.0.0) to pick the most recent gcc version as a primary libstdc++.so provider.

$ printf '%s\n' 9.2.0 10.0.1 11.0.0 | sort
10.0.1
11.0.0
9.2.0

Here is a quiz question 1 for you: how would you implement simple version-aware sorting assuming <number>.<number>.<number> versioning?

Quiz answer 1: one of the solutions is (spoiler) fix in gcc-config.

Looks like an obscure bug. People should not rely on lexicographical number sorting generally, right?

Quiz question 2: guess how many more bugs I encountered related to the fact that ‘10’ < ‘9’ as a string.

ebuild bugs

The next bug was in gcc ebuild itself (well, in toolchain.eclass). Ebuilds also happen to be written in bash and you need to be careful with number arithmetic there:

$ [[ 10 < 9 ]] && echo yes || echo no
yes
$ [[ 10 -lt 9 ]] && echo yes || echo no
no

toolchain.eclass did not get it right. I had to tweak it with a patch. It happened to work for all previous gcc versions. I wondered how widespread this problem was in ebuilds:

$ git grep -E '\[\[.*[<>]\s*[0-9]+.*\]\]' | cat
eclass/kernel-2.eclass:if [[ -n ${KV_MINOR} &&  ${KV_MAJOR}.${KV_MINOR}.${KV_PATCH} < 2.6.27 ]] ; then
net-libs/webkit-gtk/webkit-gtk-2.24.4.ebuild:if tc-is-gcc && [[ $(gcc-version) < 4.9 ]] ; then
net-mail/mailgraph/mailgraph-1.14-r2.ebuild:if [[ ${REPLACING_VERSIONS} < 1.13 ]] ; then
net-wireless/cpyrit-cuda/cpyrit-cuda-0.5.0.ebuild:if tc-is-gcc && [[ $(gcc-version) > 4.8 ]]; then
profiles/prefix/windows/winnt/profile.bashrc:[[ ${#mysrcs[@]} < 2 ]] && exit 0
sys-cluster/corosync/corosync-2.3.5.ebuild:if [[ ${REPLACING_VERSIONS} < 2.0 ]]; then
sys-cluster/corosync/corosync-2.4.2.ebuild:if [[ ${REPLACING_VERSIONS} < 2.0 ]]; then
sys-cluster/torque/torque-4.1.7-r1.ebuild:if [[ -z "${REPLACING_VERSIONS}" ]] || [[ ${REPLACING_VERSIONS} < 4 ]]; then
sys-cluster/torque/torque-4.2.10-r1.ebuild:if [[ ${showmessage} > 0 ]]; then
sys-libs/libcxx/libcxx-10.0.0.9999.ebuild:if tc-is-gcc && [[ $(gcc-version) < 4.7 ]] ; then
sys-libs/libcxx/libcxx-10.0.0_rc3.ebuild:if tc-is-gcc && [[ $(gcc-version) < 4.7 ]] ; then
sys-libs/libcxx/libcxx-11.0.0.9999.ebuild:if tc-is-gcc && [[ $(gcc-version) < 4.7 ]] ; then
sys-libs/libcxx/libcxx-7.1.0.ebuild:if tc-is-gcc && [[ $(gcc-version) < 4.7 ]] ; then
sys-libs/libcxx/libcxx-8.0.1.ebuild:if tc-is-gcc && [[ $(gcc-version) < 4.7 ]] ; then
sys-libs/libcxx/libcxx-9.0.1.ebuild:if tc-is-gcc && [[ $(gcc-version) < 4.7 ]] ; then
...

These all are wrong. Note: it’s a very crude grep. I’m sure there are a lot more bugs like that. https://bugs.gentoo.org/705240 tracks a few cases I managed to grep out. If you have found more of these please pile your bugs on! After blockers fixed I got gcc-10 installed and selected properly.

-fno-common change

Then I attempted to build newly released version of dev-lang/erlang as part of usual ebuild maintenance. The build failed with duplicate symbol definition. I assumed it was a bug in the newly written code, and hacked up a fix for upstream. Slightly later I tried to build older erlang versions in repository and discovered they were also broken. Searching around I encountered https://gcc.gnu.org/PR85678 which talks about switching -fcommon default to -fno-common to effectively forbid code similar to below:

// a.c
int a;
// b.c
int a;
$ gcc-9.3.0 a.c b.c -o libab.so -shared -fPIC
<ok>
$ gcc-10.0.1 a.c b.c -o libab.so -shared -fPIC
ld: /tmp/ccWIg1Zj.o:(.bss+0x0): multiple definition of `a'; /tmp/ccdJsLyo.o:(.bss+0x0): first defined here
collect2: error: ld returned 1 exit status

Now you have to specify explicit extern to convert one definition site to declaration and avoid duplicate (mergeable) definitions.

Quiz question 3: how many packages do you think are broken like that? One? Ten? A hundred? Guess a number.

Luckily it’s easy to find most of the packages using gcc-9 before gcc-10 is released just by trying to build packages with CFLAGS="$CFLAGS -fno-common". Toralf did just that using his magic tinderbox setup and built an almost complete list of affected packages. See blockers in the blocker bug. If you got a new fancy failure please pile your new bug onto blocker above. Also feel free to pick a bug from there and work on a patch for gentoo and/or upstream. We will need many hands to fix those leftovers. Luckily the fixes are very mechanical and can be done without too deep understanding of projects’ internals. https://wiki.gentoo.org/wiki/Gcc_10_porting_notes/fno_common page has more hints. I spent some time fixing individual packages broken on my system and then sent out wider announcement: https://archives.gentoo.org/gentoo-dev/message/086ce3c09dda598aa3bdee3fe55a3dca

Quiz answer 3: https://bugs.gentoo.org/705764 reports 585 packages broken so far.

Aside from -fno-common bugs other things started popping up.

vim crash

In https://bugs.gentoo.org/706324 gentoo user <lekto@o2.pl> reported vim crash on gcc-10. I was glad that someone else tried it and found a subtle real issue. I tried to build-and-run vim on my machine and managed to reproduce the failure.

Quiz question 4: guess what caused the crash! Is it a compiler bug or not? How picky you think vim is to a c compiler and it’s properties? The crash backtrace looked like that:

#7  0x00007f43b3ee8359 in __libc_message (action=<optimized out>,
  fmt=fmt@entry=0x7f43b3fffd4c "*** %s ***: %s terminated\n")
    at ../sysdeps/posix/libc_fatal.c:181
#8  0x00007f43b3f81545 in __GI___fortify_fail_abort (need_backtrace=need_backtrace@entry=true,
  msg=msg@entry=0x7f43b3fffcd8 "buffer overflow detected")
    at fortify_fail.c:28
#9  0x00007f43b3f81581 in __GI___fortify_fail (
  msg=msg@entry=0x7f43b3fffcd8 "buffer overflow detected")
    at fortify_fail.c:44
#10 0x00007f43b3f7f720 in __GI___chk_fail () at chk_fail.c:28
#11 0x0000563edb430ad9 in strcpy (__src=0x563edb48b7a3 "0", __dest=0x563edc345bd1 "") at /usr/include/bits/string_fortified.h:90
#12 add_nr_var (nr=<optimized out>, name=0x563edb48b7a3 "0", v=<optimizedout>, dp=0x563edc345f68) at userfunc.c:625

Buffer overflow. Uh-oh, that should never happen, right? gcc build log even reported the line as potentially having a buffer overflow at the same userfunc.c:625 line gdb pointed me to:

x86_64-pc-linux-gnu-gcc -c -I. -Iproto -DHAVE_CONFIG_H     -march=sandybridge -mtune=sandybridge -maes --param=l1-cache-size=32 --param=l1-cache-line-size=64 --param=l2-cache-size=8192 -O2 -pipe -fdiagnostics-show-option -frecord-gcc-switches -Wall -Wextra -Wstack-protector -g        -o objects/userfunc.o userfunc.c
In file included from /usr/include/string.h:494,
                 from os_unix.h:465,
                 from vim.h:234,
                 from userfunc.c:14:
In function 'strcpy',
    inlined from 'add_nr_var' at userfunc.c:625:5,
    inlined from 'call_user_func' at userfunc.c:858:5,
    inlined from 'call_func' at userfunc.c:1626:7:
/usr/include/bits/string_fortified.h:90:10: warning: '__builtin___memcpy_chk' writing 2 bytes into a region of size 1 overflows the destination [-Wstringop-overflow=]
   90 |   return __builtin___strcpy_chk (__dest, __src, __bos (__dest));
      |          ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Turns out internally vim uses the following hack to implement key/value store with variable length keys:

struct dictitem_S
 {
     typval_T di_tv;     // type and value of the variable
     char_u   di_flags;  // flags (only used for variable)
     char_u   di_key[1]; // key (actually longer!)
 };
//...
#define STRCPY(d, s)    strcpy((char *)(d), (char *)(s))
STRCPY(v->di_key, name);

And any string function like strcpy() or memcpy() is statically known to gcc as a buffer overflow: di_key is always 1 byte long. Runtime buffer overflow checking is enabled by passing -D_FORTIFY_SOURCE=2 to gcc. Many distributions enable overflow checking by default. gentoo is no exception. The workaround vim uses to avoid these failures is to disable buffer overflow checks from being emitted by using -D_FORTIFY_SOURCE=1 define. Except that in this case -D_FORTIFY_SOURCE=1 was not applied. To see why let’s look at the configure.ac around _FORTIFY_SOURCE handling:

gccversion=`$CC -dumpversion`
dnl ...
gccmajor=`echo "$gccversion" | sed -e 's/^\([[1-9]]\)\..*$/\1/g'`
dnl ...
AC_MSG_CHECKING(whether we need -D_FORTIFY_SOURCE=1)
if test "$gccmajor" -gt "3"; then
  dnl slightly simplified cimparing to actual code
  CFLAGS="$CFLAGS -U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=1"
  CPPFLAGS="$CPPFLAGS -U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=1"
  AC_MSG_RESULT(yes)
else
  AC_MSG_RESULT(no)
fi

Do you see the bug here? vim assumes gcc -dumpversion has format of <digit>.<anything>. But gcc-10.0.1 is <digit><digit>.<anything>. As a result -D_FORTIFY_SOURCE=1 was not applied and we got broken binary. The fix is trivial: patch.

-gccmajor=`echo "$gccversion" | sed -e 's/^\([[1-9]]\)\..*$/\1/g'`
+gccmajor=`echo "$gccversion" | sed -e 's/^\([[0-9]]\+\)\..*$/\1/g'`

Arguably vim should not use known-broken c constructs and use something else instead. Be it manually managed void * memory chunks or flexible arrays on modern compilers. I would not be surprised if gcc already generates invalid code for vim assuming that out-of-bounds array access is not supposed to happen in the code. That would allow gcc to delete most of code working with 1-byte arrays as dead. From discussion in https://github.com/vim/vim/issues/5581 it looks like single-byte-sized array are staying for longer though.

Quiz answer 4: gcc version parsing did not expect two digits.

perl crash

Somehow perl was also broken by gcc-10:

x86_64-pc-linux-gnu-gcc -c -DPERL_CORE -fwrapv -fpcc-struct-return -pipe -D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64 -march=sandybridge -mtune=sandybridge -maes --param=l1-cache-size=32 --param=l1-cache-line-size=64 --param=l2-cache-size=8192 -O2 -pipe -fdiagnostics-show-option -frecord-gcc-switches -Wall -Wextra -Wstack-protector -g -Wall -fPIC gv.c
...
LD_LIBRARY_PATH=/tmp/portage/dev-lang/perl-5.30.1/work/perl-5.30.1 /tmp/portage/dev-lang/perl-5.30.1/work/perl-5.30.1/preload /tmp/portage/dev-lang/perl-5.30.1/work/perl-5.30.1/libperl.so.5.30.1 ./miniperl -w -Ilib -Idist/Exporter/lib -MExporter -e '<?>' || sh -c 'echo >&2 Failed to build miniperl.  Please run make minitest; exit 1'
Attempt to free unreferenced scalar: SV 0x5555ed3e1378.
/bin/sh: line 1: 4057907 Segmentation fault      (core dumped)

This one is hard. Can you quickly guess what is suspiciously wrong here? In this case the runtime (perl build-time) failure happens due to use of -fpcc-struct-return flag that changes compiler’s ABI:

-fpcc-struct-return:
    Return "short" "struct" and "union" values in memory
    like longer ones, rather than in registers.

Looking at the upstream fix this flag is a result of configure script thinking it deals with gcc-1:

-1*) dflt="$dflt -fpcc-struct-return" ;;
+1.*) dflt="$dflt -fpcc-struct-return" ;;

Once again version parsing did not expect two digits.

linux crash

The next test for a new compiler is to try to boot into kernel built by gcc-10. Rebuilding and reinstalling grub2 caused no problems. But rebuilding the kernel made it unbootable on a real machine. Worst thing was that I got no screen output at all after a boot loader prompt. For some reason qemu-system-x86_64 was able to boot kernel just fine. Not easy to debug. I needed some indication how far the boot process got. I managed to get it in a few ways: via efifb earlycon and via xdbc (USB-3 debug capability). The simplest one that does not require second machine was efifb earlycon.

efifb earlycon

On EFI systems you can emit text output almost instantly at kernel boot. As EFI is already initialized it provides kernel a graphical frame buffer: a memory range to write your pixels in. EFI frame buffer is slightly different from VGA text mode but not too much. Kernel only needs to find out where frame buffer memory resides to render glyphs right there. Enabling early frame buffer appeared to be a bit tricky though. We need two things:

  1. a few unusual features built into kernel
  2. kernel parameters to enable early console

Kernel config:

Kernel parameters: "earlycon=efifb keep_bootcon". This allowed me to get an early boot failure on screen:

Kernel panic — not syncing: stack-protector: Kernel stack is corrupted in: start_secondary+0x191/0x1a0
CPU: 1 PID: 0 Comm: swapper/1 Not tainted 5.6.0-rc5—00235—gfffb08b37df9 #139
Hardware name: Gigabyte Technology Co., Ltd. To be filled by O.E.M./H77M—D3H, BIOS F12 11/14/2013
Call Trace:
  dump_stack+0x71/0xa0
  panic+0x107/0x2b8
  ? start_secondary+0x191/0x1a0
  __stack_chk_fail+0x15/0x20
  start_secondary+0x191/0x1a0
  secondary_startup_64+0xa4/0xb0
-—-[ end Kernel panic — not syncing: stack—protector: Kernel stack is corrupted in: start_secondary+0x191

Woohoo! That I was able to work with. I looked at start_secondary() definition:

/*
 * Activate a secondary processor.
 */
static void notrace start_secondary(void *unused)
{
    /*
     * Don't put *anything* except direct CPU state initialization
     * before cpu_init(), SMP booting is too fragile that we want to
     * limit the things done here to the most necessary things.
     */
    cr4_init();
    // ...
    cpu_init();
    // ...
    check_tsc_sync_target();
    // ...
    set_cpu_online(smp_processor_id(), true);
    // ...

    /* to prevent fake stack check failure in clock setup */
    boot_init_stack_canary();

    // ...
    cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
}

/*
 * Initialize the stackprotector canary value.
 *
 * NOTE: this must only be called from functions that never return,
 * and it must always be inlined.
 */
static __always_inline void boot_init_stack_canary(void)
{
    u64 canary;
    u64 tsc;

    BUILD_BUG_ON(offsetof(struct fixed_percpu_data, stack_canary) != 40);
    /*
     * We both use the random pool and the current TSC as a source
     * of randomness. The TSC only matters for very early init,
     * there it already has some randomness on most systems. Later
     * on during the bootup the random pool has true entropy too.
     */
    get_random_bytes(&canary, sizeof(canary));
    tsc = rdtsc();
    canary += tsc + (tsc << 32UL);
    canary &= CANARY_MASK;

    current->stack_canary = canary;
    this_cpu_write(fixed_percpu_data.stack_canary, canary);
}

Here start_secondary() detected a stack corruption failure and reported it with __stack_chk_fail(). Note: start_secondary() is itself responsible for initial stack canary setup. The workaround to make a kernel boot was to avoid stack protection of start_secondary():

--- a/arch/x86/kernel/Makefile
+++ b/arch/x86/kernel/Makefile
@@ -11,6 +11,12 @@ extra-y+= vmlinux.lds

 CPPFLAGS_vmlinux.lds += -U$(UTS_MACHINE)

+# smpboot's init_secondary initializes stack canary.
+# Make sure we don't emit stack checks before it's
+# initialized.
+nostackp := $(call cc-option, -fno-stack-protector)
+CFLAGS_smpboot.o := $(nostackp)
+
 ifdef CONFIG_FUNCTION_TRACER
 # Do not profile debug and lowlevel utilities
 CFLAGS_REMOVE_tsc.o = -pg

Kernel stack protection itself is enabled by CONFIG_STACKPROTECTOR_STRONG=y option. The real fix is discussed in https://lkml.org/lkml/2020/3/14/186 and will involve marking only start_secondary() as exempted from protection.

Parting words

After this short exercise I think gcc-10 is somewhat usable in gentoo. Now to the real bugs like https://gcc.gnu.org/PR94185 and https://gcc.gnu.org/PR93763.

Have fun!