From: Christian Brauner <brauner@kernel.org>
Date: Thu, 24 Nov 2022 08:30:19 +0000 (+0100)
Subject: conf: create separate peer group for container's root
X-Git-Tag: v6.0.0~81^2~2
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=81d94a4eec65fecc244e2d2def457e1b510d8777;p=thirdparty%2Flxc.git

conf: create separate peer group for container's root

Finally, we turn the rootfs into a shared mount. Note, that this
doesn't reestablish mount propagation with the hosts mount
namespace. Instead we'll create a new peer group.

We're doing this because most workloads do rely on the rootfs being
a shared mount. For example, systemd daemon like sytemd-udevd run in
their own mount namespace. Their mount namespace has been made a
dependent mount (MS_SLAVE) with the host rootfs as it's dominating
mount. This means new mounts on the host propagate into the
respective services.

This is broken if we leave the container's rootfs a dependent mount.
In which case both the container's rootfs and the service's rootfs
will be dependent mounts with the host's rootfs as their dominating
mount. So if you were to mount over the rootfs from the host it
would not just propagate into the container's mount namespace it
would also propagate into the service. That's nonsense semantics for
nearly all relevant use-cases. Instead, establish the container's
rootfs as a separate peer group mirroring the behavior on the host.

Signed-off-by: Christian Brauner (Microsoft) <christian.brauner@ubuntu.com>
---

diff --git a/src/lxc/conf.c b/src/lxc/conf.c
index bc1b25464..63b697f9f 100644
--- a/src/lxc/conf.c
+++ b/src/lxc/conf.c
@@ -1606,8 +1606,11 @@ static int lxc_pivot_root(const struct lxc_rootfs *rootfs)
 		return log_error_errno(-errno, errno, "Failed to enter old root directory");
 
 	/*
-	 * Make fd_oldroot a depedent mount to make sure our umounts don't
-	 * propagate to the host.
+	 * Unprivileged containers will have had all their mounts turned into
+	 * dependent mounts when the container was created. But for privileged
+	 * containers we need to turn the old root mount tree into a dependent
+	 * mount tree to prevent propagating mounts and umounts into the host
+	 * mount namespace.
 	 */
 	ret = mount("", ".", "", MS_SLAVE | MS_REC, NULL);
 	if (ret < 0)
@@ -1621,6 +1624,31 @@ static int lxc_pivot_root(const struct lxc_rootfs *rootfs)
 	if (ret < 0)
 		return log_error_errno(-errno, errno, "Failed to re-enter new root directory \"%s\"", rootfs->mount);
 
+	/*
+	 * Finally, we turn the rootfs into a shared mount. Note, that this
+	 * doesn't reestablish mount propagation with the hosts mount
+	 * namespace. Instead we'll create a new peer group.
+	 *
+	 * We're doing this because most workloads do rely on the rootfs being
+	 * a shared mount. For example, systemd daemon like sytemd-udevd run in
+	 * their own mount namespace. Their mount namespace has been made a
+	 * dependent mount (MS_SLAVE) with the host rootfs as it's dominating
+	 * mount. This means new mounts on the host propagate into the
+	 * respective services.
+	 *
+	 * This is broken if we leave the container's rootfs a dependent mount.
+	 * In which case both the container's rootfs and the service's rootfs
+	 * will be dependent mounts with the host's rootfs as their dominating
+	 * mount. So if you were to mount over the rootfs from the host it
+	 * would not just propagate into the container's mount namespace it
+	 * would also propagate into the service. That's nonsense semantics for
+	 * nearly all relevant use-cases. Instead, establish the container's
+	 * rootfs as a separate peer group mirroring the behavior on the host.
+	 */
+	ret = mount("", ".", "", MS_SHARED | MS_REC, NULL);
+	if (ret < 0)
+		return log_error_errno(-errno, errno, "Failed to turn new root mount tree into shared mount tree");
+
 	TRACE("Changed into new rootfs \"%s\"", rootfs->mount);
 	return 0;
 }