Outstanding CI issue

[mrc0mmand]

(I'm opening this mainly as a reminder to myself, so I don't debug the same stuff twice.)

In #433 CI started failing with:

 120/298 test-fs-util                                       FAIL              0.05s   killed by signal 6 SIGABRT
>>> SYSTEMD_KBD_MODEL_MAP=/build/src/locale/kbd-model-map SYSTEMD_LANGUAGE_FALLBACK_MAP=/build/src/locale/language-fallback-map MALLOC_PERTURB_=236 PATH=/build/build:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin /build/build/test-fs-util
 ✀  
stderr:
Assertion 'touch_file(a, false, test_mtime, test_uid, test_gid, 0640) >= 0' failed at ../src/test/test-fs-util.c:549, function test_touch_file(). Aborting.

The fail itself is caused by touch_file() calling chmod() on a symlink, which then returns EOPNOTSUPP. I suspect this is something that has changed in later kernels (and in this case it affects CI, because it runs in a docker container on Ubuntu Jammy). On C8S kernel this seems to work, but only on tmpfs:

# cat main.c 
#include <assert.h>
#include <errno.h>
#include <fcntl.h>
#include <stdio.h>
#include <sys/stat.h>
#include <sys/types.h>

int main(int argc, char *argv[]) {
        int fd, r;
        char fdpath[64];

        if (argc < 2)
                return 1;

        fd = open(argv[1], O_PATH|O_CLOEXEC|O_NOFOLLOW);
        if (fd < 0) {
                fprintf(stderr, "open() failed: %m\n");
                return 1;
        }

        sprintf(fdpath, "/proc/self/fd/%i", fd);
        r = chmod(fdpath, 0640);
        printf("chmod(%s)=%d (%d)\n", fdpath, r, errno);

        return r;
}

# mount -t tmpfs tmpfs /tmp
# (cd /tmp; ln -s symlink dangling; ln -s /etc/os-release symlink)
# (cd /var/tmp; ln -s symlink dangling; ln -s /etc/os-release symlink)
# gcc -o main main.c -D_GNU_SOURCE
# ./main /tmp/dangling
chmod(/proc/self/fd/3)=0 (0)
# ./main /tmp/symlink
chmod(/proc/self/fd/3)=0 (0)
# ./main /var/tmp/dangling 
chmod(/proc/self/fd/3)=-1 (95)
# ./main /var/tmp/symlink 
chmod(/proc/self/fd/3)=-1 (95)

And since the test uses /dev/shm/ for the test files (which is either tmpfs or some other form of tmpfs), it works.

On newer kernels all attempts to call chmod() on a symlink fail:

# uname -r
6.7.4-200.fc39.x86_64
# ./main /tmp/dangling 
chmod(/proc/self/fd/3)=-1 (95)
# ./main /tmp/symlink 
chmod(/proc/self/fd/3)=-1 (95)
# ./main /var/tmp/dangling 
chmod(/proc/self/fd/3)=-1 (95)
# ./main /var/tmp/symlink 
chmod(/proc/self/fd/3)=-1 (95)

The fix on the touch_file() side would be to call fchmod_and_chown() instead. Unfortunately, our fchmod_and_chown() in RHEL 8 is not clever enough and still does chmod() on a symlink, suffering from the same issue. For that we'd need 2dbb7e9 (+ commits before and after that change) to make it work properly, and that seems way too risky given how far in we're in RHEL 8 cycle.

So, when the need arises, I think it would be safer to just skip the offending test case, i.e.:

systemd-rhel8/src/test/test-fs-util.c

Lines 547 to 555 in ccaa361

	a = strjoina(p, "/lnk");
	assert_se(symlink("target", a) >= 0);
	assert_se(touch_file(a, false, test_mtime, test_uid, test_gid, 0640) >= 0);
	assert_se(lstat(a, &st) >= 0);
	assert_se(st.st_uid == test_uid);
	assert_se(st.st_gid == test_gid);
	assert_se(S_ISLNK(st.st_mode));
	assert_se((st.st_mode & 0777) == 0640);
	assert_se(timespec_load(&st.st_mtim) == test_mtime);

when running in GH Actions (or docker, whatever).