rootless: make /sys/fs/cgroup/* read-only #1869
Conversation
AkihiroSuda
force-pushed
the
bind-sys-2
branch
from
e4d76aa
to
6de6a84
Compare
| { | ||
| Source: "/sys", | ||
| Destination: "/sys", | ||
| Type: "none", |
There was a problem hiding this comment.
this is the only place where I see it used. In the runtime-specs example bind is used as the type for a bind mount (https://github.com/opencontainers/runtime-spec/blob/master/config.md#example-linux), both Docker and Podman generate "bind".
Anyway since it is only in the example and not in the specs, and you are just moving this code snippet, it is probably fine to leave it as it is.
There was a problem hiding this comment.
type: "none" is entirely valid. There was a bug previously in runc when we didn't handle this properly, but the type field (when it comes to mount(2)) doesn't affect bindmounts -- what matters is that the bind option is set. Docker set bind presumably by accident, and Podman probably copied the mistake (to be clear -- it doesn't matter either way, but it does lead to confusion).
But this code is used for more than example -- it's used by at least a few projects to convert a configuration to the rootless version of said configuration.
libcontainer/specconv/example.go
Outdated
| @@ -162,6 +162,10 @@ func ToRootless(spec *specs.Spec) { | |||
| var namespaces []specs.LinuxNamespace | |||
|
|
|||
| // Remove networkns from the spec. | |||
| // | |||
| // TODO: removal of networkns should be optional, | |||
There was a problem hiding this comment.
FYI. When someone puts TODOs in PRs, it makes reviewers nervous.
todo when?
by who?
what is the issue now?
what breaks while this isn't done?
AkihiroSuda
force-pushed
the
bind-sys-2
branch
from
6de6a84
to
6450010
Compare
|
addressed comments |
|
Thanks, LGTM |
libcontainer/specconv/example.go
Outdated
| var mounts []specs.Mount | ||
| for _, mount := range spec.Mounts { | ||
| // Ignore all mounts that are under /sys. | ||
| if strings.HasPrefix(mount.Destination, "/sys") { |
There was a problem hiding this comment.
I'm not sure if this is always fine -- if someone wants to mount a fake filesystem over parts of /sys then I don't see why we would block it (let alone ignore it).
Ultimately this is used during config generation so it's not the end of the world, but it is something to keep in mind.
libcontainer/specconv/example.go
Outdated
| // Without this hack, `df` fails with "df: /sys/firmware/efi/efivars" | ||
| // because the efivars entry exists in mtab (because /sys is rbind-mounted) | ||
| // but /sys/firmware is masked. | ||
| if strings.HasPrefix(masked, "/sys") { |
There was a problem hiding this comment.
I'm not sure this is fine for rootless mode -- if you unmask paths you could be leaking information to containers (so while they might not be able to write to the masked paths, they are able to read them). I mean, you probably can't read most of /sys that is masked anyway, but I'm not sure it's a great idea.
The default rootless spec bind-mounts /sys with "rbind,ro" (because mounting sysfs requires netns to be unshared), however, it does not make subfilesystems mounted under /sys read-only. So, files under /sys/fs/cgroup were unexpectedly writable when they are chmod/chowned via privileged helpers such as pam_cgfs. This patch fixes the issue by mounting /sys/fs/cgroup/* as read-only explicitly. Signed-off-by: Akihiro Suda <suda.akihiro@lab.ntt.co.jp>
AkihiroSuda
force-pushed
the
bind-sys-2
branch
from
6450010
to
28dc2e6
Compare
|
addressed comments, although it became more complicated |
|
@giuseppe latest revision still LGTY? |
|
@AkihiroSuda yes, latest version LGTM |
|
@cyphar PTAL |
|
ping @cyphar |
|
On second thought maybe we should rather suggest users not mounting Rootless BuildKit already removed For Rootless Moby, I'm going to close this PR but I may change my mind later. |
The default rootless spec bind-mounts /sys with "rbind,ro" (because mounting sysfs
requires netns to be unshared), however, it does not make subfilesystems mounted
under /sys read-only.
So, files under /sys/fs/cgroup were unexpectedly writable when they are chmod/chowned
via privileged helpers such as pam_cgfs.
This patch fixes the issue by mounting /sys/fs/cgroup/* as read-only explicitly.
Signed-off-by: Akihiro Suda suda.akihiro@lab.ntt.co.jp