** Description changed:
+ [Impact]
+
+ *
+
+ [Test Plan]
+
+ *
+
+ [Where problems could occur]
+
+ *
+
+ [Other Info]
+
+ *
+
+ ---original---
SUMMARY
Since core24 revision 2124 (version 20260824, in latest/stable since
2026-09-10), the base snap
ships tar 1.35+dfsg-3ubuntu0.4. Its fix for CVE-2025-45582
(debian/patches/CVE-2025-45582-*.patch,
first added in 1.35+dfsg-3ubuntu0.2) extracts archive members through
gnulib's openat2() wrapper
with RESOLVE_BENEATH.
snapd's seccomp template denies openat2 with EPERM
(interfaces/seccomp/template.go):
# Deny openat2 from being used as struct open_how is behind a pointer and
it
# may be used to pass suid/sgid mode.
# openat2 <dirfd> <path> <how> <size>
~openat2
gnulib only falls back to its openat()-based emulation when openat2 fails
with ENOSYS
(CVE-2025-45582-gnulib-1.patch, gnu/openat2.c):
r = syscall (SYS_openat2, dfd, filename, how, howsize);
if (! (r < 0 && errno == ENOSYS))
return r;
So under strict confinement tar can no longer extract any member below the
top-level directory.
It fails with "Cannot mkdir: Operation not permitted" / "Cannot open:
Operation not permitted",
or with "Permission denied" in devmode. This affects any strictly confined
core24 snap that runs
tar, including dpkg-source (dpkg-buildpackage -S for 3.0 (quilt) packages
always extracts the
orig tarball).
STEPS TO REPRODUCE
1. Use any strictly confined snap with base: core24 and the home plug
(core24 rev 2124 installed).
2. Create a tarball with a nested path:
mkdir -p ~/t/src/a/b && echo x > ~/t/src/a/b/f && echo y > ~/t/src/top
tar -C ~/t/src -cf ~/t/t.tar . && mkdir ~/t/out
3. Extract it with core24's tar inside the snap:
snap run --shell <snap> -c '/usr/bin/tar -xf ~/t/t.tar -C ~/t/out'
Expected: extraction succeeds (it does with core24 rev 1643, tar
1.35+dfsg-3build1).
Actual:
/usr/bin/tar: ./a/b: Cannot mkdir: Operation not permitted
/usr/bin/tar: ./a/b/f: Cannot open: Operation not permitted
/usr/bin/tar: Exiting with failure status due to previous errors
Minimal syscall reproducer (static binary run inside the snap):
struct open_how h = {.flags = O_WRONLY|O_CREAT, .mode = 0644, .resolve =
RESOLVE_BENEATH};
syscall(SYS_openat2, AT_FDCWD, "d/f", &h, sizeof h);
host: openat2 -> 3 (ok)
inside snap: openat2 -> -1 (Operation not permitted)
openat -> 3 (ok)
IMPACT
The main affected snap is flamenco (https://github.com/canonical/flamenco), a
strictly confined
core24 snap used to build Ubuntu .NET source packages. "flamenco pack"
extracts the orig tarball
with tar and then runs "dpkg-buildpackage -S", whose dpkg-source extracts the
orig tarball again.
Both now fail inside the snap.
History:
1. flamenco originally bundled its own tar via "stage-packages: tar". That
tar, taken from the
noble archive, failed inside the snap with the same "Cannot mkdir/open:
Permission denied"
errors.
2. This was worked around in
https://github.com/canonical/flamenco/pull/9/changes/c28072bf358749629242c63b5f38b7291d52a4e9
by removing the bundled tar and calling core24's /usr/bin/tar instead. At
the time core24
latest/stable was rev 1643, which ships tar 1.35+dfsg-3build1. That
version predates the
CVE-2025-45582 fix, never calls openat2, and extracted correctly under
confinement.
3. On 2026-09-10 core24 rev 2124 (20260824) was promoted to latest/stable. It
ships
tar 1.35+dfsg-3ubuntu0.4, which includes the CVE-2025-45582 fix and
extracts through openat2.
Machines picked it up through normal base snap auto-refresh, with no
change to flamenco and no
reinstall:
- fresh CI runners got it immediately, and flamenco-based CI jobs have
failed since 2026-09-11;
- a developer machine was only offered the update on 2026-09-28 (snapd
reported
"no updates available" for core24 until then) and started failing right
after the refresh.
The same flamenco build therefore went from working to broken purely
because the base snap
changed underneath it. The flamenco-side fix from step 2 cannot help,
because the tar that now
calls openat2 is the one in the base snap.
4. core24 20260928 (rev 2236, currently in beta/edge) ships the same tar
(1.35+dfsg-3ubuntu0.4), so this will not go away with the next core24
release.
Current workaround in flamenco: an LD_PRELOAD shim that makes
syscall(SYS_openat2, ...) return
ENOSYS, so gnulib falls back to its openat()-based emulation. That emulation
still enforces
RESOLVE_BENEATH, so the CVE-2025-45582 protection is kept. The shim is only
needed because
snapd returns EPERM instead of ENOSYS.
SUGGESTED FIX
Return ENOSYS instead of EPERM for openat2 in the default seccomp template
(for example with
snapd's errno action syntax, if supported for this rule). Userspace written
against newer kernels
treats ENOSYS as "fall back to openat()". That keeps the original concern
(suid/sgid modes hidden
behind the open_how pointer) addressed without breaking tar, dpkg-source or
other gnulib-based
tools.
ENVIRONMENT
- snapd 2.76.3+ubuntu24.04 (snap rev 27738)
- core24 20260824 (rev 2124), tar 1.35+dfsg-3ubuntu0.4
- Ubuntu 24.04.5 LTS, kernel 7.0.0-34-generic
- Also reproduced on GitHub Actions ubuntu-24.04 runners (kernel
6.17.0-1022-azure)
- Confinement: strict (EPERM) and devmode (EACCES)
--
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2168888
Title:
seccomp: denying openat2 with EPERM breaks GNU tar (CVE-2025-45582
fix) in core24 snaps
To manage notifications about this bug go to:
https://bugs.launchpad.net/snapd/+bug/2168888/+subscriptions
--
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs