Previously, system call filtering (to prevent builders from storing files with
setuid/setgid permission bits or extended attributes) was performed using a
blocklist. While this looks simple at first, it actually carries significant
security and maintainability risks: after all, the kernel may add new syscalls
to achieve the same functionality one is trying to block, and it can even be
hard to actually add the syscall to the blocklist when building against a C
library that doesn't know about it yet. For a recent demonstration of this
happening in practice to Nix, see the introduction of fchmodat2 [0] [1].
The allowlist approach does not share the same drawback. While it does require
a rather large list of harmless syscalls to be maintained in the codebase,
failing to update this list (and roll out the update to all users) in time has
rather benign effects; at worst, very recent programs that already rely on new
syscalls will fail with an error the same way they would on a slightly older
kernel that doesn't support them yet. Most importantly, no unintended new ways
of performing dangerous operations will be silently allowed.
Another possible drawback is reduced system call performance due to the larger
filter created by the allowlist requiring more computation [2]. However, this
issue has not convincingly been demonstrated yet in practice, for example in
systemd or various browsers. To the contrary, it has been measured that the the
actual filter constructed here has approximately the same overhead as a very
simple filter blocking only one system call.
This commit tries to keep the behavior as close to unchanged as possible. The
system call list is in line with libseccomp 2.5.5 and glibc 2.39, which are the
latest versions at the point of writing. Since libseccomp 2.5.5 is already a
requirement and the distributions shipping this together with older versions of
glibc are mostly not a thing any more, this should not lead to more build
failures any more.
[0] https://github.com/NixOS/nixpkgs/issues/300635
[1] https://github.com/NixOS/nix/issues/10424
[2] https://github.com/flatpak/flatpak/pull/4462#issuecomment-1061690607
Change-Id: I541be3ea9b249bcceddfed6a5a13ac10b11e16ad
With Linux kernel >=6.6 & glibc 2.39 a `fchmodat2(2)` is available that
isn't filtered away by the libseccomp sandbox.
Being able to use this to bypass that restriction has surprising results
for some builds such as lxc[1]:
> With kernel ≥6.6 and glibc 2.39, lxc's install phase uses fchmodat2,
> which slips through 9b88e52846/src/libstore/build/local-derivation-goal.cc (L1650-L1663).
> The fixupPhase then uses fchmodat, which fails.
> With older kernel or glibc, setting the suid bit fails in the
> install phase, which is not treated as fatal, and then the
> fixup phase does not try to set it again.
Please note that there are still ways to bypass this sandbox[2] and this is
mostly a fix for the breaking builds.
This change works by creating a syscall filter for the `fchmodat2`
syscall (number 452 on most systems). The problem is that glibc 2.39
is needed to have the correct syscall number available via
`__NR_fchmodat2` / `__SNR_fchmodat2`, but this flake is still on
nixpkgs 23.11. To have this change everywhere and not dependent on the
glibc this package is built against, I added a header
"fchmodat2-compat.hh" that sets the syscall number based on the
architecture. On most platforms its 452 according to glibc with a few
exceptions:
$ rg --pcre2 'define __NR_fchmodat2 (?!452)'
sysdeps/unix/sysv/linux/x86_64/x32/arch-syscall.h
58:#define __NR_fchmodat2 1073742276
sysdeps/unix/sysv/linux/mips/mips64/n32/arch-syscall.h
67:#define __NR_fchmodat2 6452
sysdeps/unix/sysv/linux/mips/mips64/n64/arch-syscall.h
62:#define __NR_fchmodat2 5452
sysdeps/unix/sysv/linux/mips/mips32/arch-syscall.h
70:#define __NR_fchmodat2 4452
sysdeps/unix/sysv/linux/alpha/arch-syscall.h
59:#define __NR_fchmodat2 562
I added a small regression-test to the setuid integration-test that
attempts to set the suid bit on a file using the fchmodat2 syscall.
I confirmed that the test fails without the change in
local-derivation-goal.
Additionally, we require libseccomp 2.5.5 or greater now: as it turns
out, libseccomp maintains an internal syscall table and
validates each rule against it. This means that when using libseccomp
2.5.4 or older, one may pass `452` as syscall number against it, but
since it doesn't exist in the internal structure, `libseccomp` will refuse
to create a filter for that. This happens with nixpkgs-23.11, i.e. on
stable NixOS and when building Lix against the project's flake.
To work around that
* a backport of libseccomp 2.5.5 on upstream nixpkgs has been
scheduled[3].
* the package now uses libseccomp 2.5.5 on its own already. This is to
provide a quick fix since the correct fix for 23.11 is still a staging cycle
away.
We still need the compat header though since `SCMP_SYS(fchmodat2)`
internally transforms this into `__SNR_fchmodat2` which points to
`__NR_fchmodat2` from glibc 2.39, so it wouldn't build on glibc 2.38.
The updated syscall table from libseccomp 2.5.5 is NOT used for that
step, but used later, so we need both, our compat header and their
syscall table 🤷
Relevant PRs in CppNix:
* https://github.com/NixOS/nix/pull/10591
* https://github.com/NixOS/nix/pull/10501
[1] https://github.com/NixOS/nixpkgs/issues/300635#issuecomment-2031073804
[2] https://github.com/NixOS/nixpkgs/issues/300635#issuecomment-2030844251
[3] https://github.com/NixOS/nixpkgs/pull/306070
(cherry picked from commit ba6804518772e6afb403dd55478365d4b863c854)
Change-Id: I6921ab5a363188c6bff617750d00bb517276b7fe
