On Fri, Dec 01, 2023 at 05:57:35PM +0100, Christian Brauner wrote:

On Wed, Nov 29, 2023 at 03:50:24PM -0600, Seth Forshee (DigitalOcean) wrote:

quoted

cap_inode_getsecurity() implements a handful of policies for capability
xattrs read by userspace:

 - It returns EINVAL if the on-disk capability is in v1 format.

 - It masks off all bits in magic_etc except for the version and
   VFS_CAP_FLAGS_EFFECTIVE.

 - v3 capabilities are converted to v2 format if the rootid returned to
   userspace would be 0 or if the rootid corresponds to root in an
   ancestor user namespace.

 - It returns EOVERFLOW for a v3 capability whose rootid does not map to
   a valid id in current_user_ns() or to root in an ancestor namespace.

Nice. Precise and clear, please just drop these bullet points into the
kernel-doc of that function.

Will do.

quoted

+/**
+ * vfs_caps_to_user_xattr - convert vfs_caps to caps xattr for userspace
+ *
+ * @idmap:       idmap of the mount the inode was found from
+ * @dest_userns: user namespace for ids in xattr data
+ * @vfs_caps:    source vfs_caps data
+ * @data:        destination buffer for rax xattr caps data
+ * @size:        size of the @data buffer
+ *
+ * Converts a kernel-interrnal capability into the raw security.capability
+ * xattr format. Includes permission checking and v2->v3 conversion as
+ * appropriate.
+ *
+ * If the xattr is being read or written through an idmapped mount the
+ * idmap of the vfsmount must be passed through @idmap. This function
+ * will then take care to map the rootid according to @idmap.
+ *
+ * Return: On success, return 0; on error, return < 0.
+ */
+int vfs_caps_to_user_xattr(struct mnt_idmap *idmap,
+			   struct user_namespace *dest_userns,
+			   const struct vfs_caps *vfs_caps,
+			   void *data, int size)
+{
+	struct vfs_ns_cap_data *ns_caps = data;
+	bool is_v3;
+	u32 magic;
+
+	/* Preserve previous behavior of returning EINVAL for v1 caps */
+	if ((vfs_caps->magic_etc & VFS_CAP_REVISION_MASK) == VFS_CAP_REVISION_1)
+		return -EINVAL;
+
+	size = __vfs_caps_to_xattr(idmap, dest_userns, vfs_caps, data, size);
+	if (size < 0)
+		return size;
+
+	magic = vfs_caps->magic_etc &
+		(VFS_CAP_REVISION_MASK | VFS_CAP_FLAGS_EFFECTIVE);
+	ns_caps->magic_etc = cpu_to_le32(magic);
+
+	/*
+	 * If this is a v3 capability with a valid, non-zero rootid, return
+	 * the v3 capability to userspace. A v3 capability with a rootid of
+	 * 0 will be converted to a v2 capability below for compatibility
+	 * with old userspace.
+	 */
+	is_v3 = (vfs_caps->magic_etc & VFS_CAP_REVISION_MASK) == VFS_CAP_REVISION_3;
+	if (is_v3) {
+		uid_t rootid = le32_to_cpu(ns_caps->rootid);
+		if (rootid != (uid_t)-1 && rootid != (uid_t)0)
+			return size;
+	}
+
+	if (!rootid_owns_currentns(vfs_caps->rootid))
+		return -EOVERFLOW;

For a v2 cap that we read vfs_caps->rootid will be vfsuid 0, right?
So that means we're guaranteed to resolve that in the initial user
namespace. IOW, rootid_owns_currentns() will indeed work with a pure v2
cap. Ok. Just making sure that I understand that this won't cause
EOVERFLOW for v2. But you would've likely seen that in tests right away.

Yes, that's all correct.