[RFC PATCH v3 2/3] fs: Make vfs_mknod() to check CAP_MKNOD in user namespace of sb

[Michael Weiß]

Check CAP_MKNOD for user namespace of sb with ns_cabable() in
fs/namei.c. This will allow lsm-based guarding of device node
creation in non-initial user namespace by stripping out SB_I_NODEV
for mounts in its own namespace.

Currently, device access is blocked unconditionally in
may_open_dev() and mounts inside unprivileged user namespaces get
SB_I_NODEV set in sb->s_iflags causing open() to fail with -EACCES.

Device access by cgroups is mediated in the following places
1) fs/namei.c:
	inode_permission() -> devcgroup_inode_permission
	vfs_mknod() and -> devcgroup_inode_mknod

2) block/bdev.c:
	blkdev_get_by_dev() -> devcgroup_check_permission

3) drivers/gpu/drm/amd/amdkfd/kfd_priv.h:
	kfd_devcgroup_check_permission -> devcgroup_check_permission

We leave this all in place. However, a lsm now can implement the
security hook security_inode_mknod() which is called directly after
the devcgroup_inode_mknod() in vfs_mknod() and remove the SB_I_NODEV.
This will let the call to may_open_dev() during open() succeed.

Turning the check form capable(CAP_MKNOD) to ns_capable(sb->s_userns,
CAP_MKNOD) is inherently save due to SB_I_NODEV. However, this may
allow to create device nodes which then could not be opened.

To give user space some time to adopt, we introduce a sysctl knob
which must be explicitly set to "1" to activate the use of
ns_capable(). Otherwise, we just check the global capability for the
current task as before.

I tested this approach in a GyroidOS container using the small
devguard LSM of the followup commit.

Signed-off-by: Michael Weiß <michael.weiss@xxxxxxxxxxxxxxxxxxx>
---
 fs/namei.c | 30 +++++++++++++++++++++++++++++-
 1 file changed, 29 insertions(+), 1 deletion(-)

diff --git a/fs/namei.c b/fs/namei.c
index 71c13b2990b4..cc61545e02ce 100644
--- a/fs/namei.c
+++ b/fs/namei.c
@@ -1032,6 +1032,7 @@ static int sysctl_protected_symlinks __read_mostly;
 static int sysctl_protected_hardlinks __read_mostly;
 static int sysctl_protected_fifos __read_mostly;
 static int sysctl_protected_regular __read_mostly;
+static int sysctl_nscap_mknod __read_mostly;
 
 #ifdef CONFIG_SYSCTL
 static struct ctl_table namei_sysctls[] = {
@@ -1071,6 +1072,15 @@ static struct ctl_table namei_sysctls[] = {
 		.extra1		= SYSCTL_ZERO,
 		.extra2		= SYSCTL_TWO,
 	},
+	{
+		.procname	= "nscap_mknod",
+		.data		= &sysctl_nscap_mknod,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.proc_handler	= proc_dointvec_minmax,
+		.extra1		= SYSCTL_ZERO,
+		.extra2		= SYSCTL_ONE,
+	},
 	{ }
 };
 
@@ -3940,6 +3950,24 @@ inline struct dentry *user_path_create(int dfd, const char __user *pathname,
 }
 EXPORT_SYMBOL(user_path_create);
 
+/**
+ * sb_mknod_capable - check userns of sb for CAP_MKNOD
+ * @sb:	super block to which userns CAP_MKNOD should be checked
+ *
+ * Check userns of sb for CAP_MKNOD
+ *
+ * Check CAP_MKNOD for owning user namespace of sb if corresponding sysctl is set.
+ * Otherwise just check global capability for current task. This allows
+ * lsm-based guarding of device node creation in non-initial user namespace.
+ */
+static bool sb_mknod_capable(struct super_block *sb)
+{
+	struct user_namespace *user_ns;
+
+	user_ns = sysctl_nscap_mknod ? sb->s_user_ns : &init_user_ns;
+	return ns_capable(user_ns, CAP_MKNOD);
+}
+
 /**
  * vfs_mknod - create device node or file
  * @idmap:	idmap of the mount the inode was found from
@@ -3966,7 +3994,7 @@ int vfs_mknod(struct mnt_idmap *idmap, struct inode *dir,
 		return error;
 
 	if ((S_ISCHR(mode) || S_ISBLK(mode)) && !is_whiteout &&
-	    !capable(CAP_MKNOD))
+	    !sb_mknod_capable(dentry->d_sb))
 		return -EPERM;
 
 	if (!dir->i_op->mknod)
-- 
2.30.2