Hi Kees,

thanks for the review!

On Fri, May 13, 2022 at 03:36:26PM -0700, Kees Cook wrote:

On May 4, 2022 12:54:18 PM PDT, Matthias Kaehlcke [off-list ref] wrote:

quoted

Extend LoadPin to allow loading of kernel files from trusted dm-verity [1]
devices.

This change adds the concept of trusted verity devices to LoadPin. LoadPin
maintains a list of root digests of verity devices it considers trusted.
Userspace can populate this list through an ioctl on the new LoadPin
securityfs entry 'dm-verity'. The ioctl receives a file descriptor of
a file with verity digests as parameter. Verity reads the digests from
this file after confirming that the file is located on the pinned root.
The list of trusted digests can only be set up once, which is typically
done at boot time.

When a kernel file is read LoadPin first checks (as usual) whether the file
is located on the pinned root, if so the file can be loaded. Otherwise, if
the verity extension is enabled, LoadPin determines whether the file is
located on a verity backed device and whether the root digest of that

I think this should be "... on an already trusted device ..."

It's not entirely clear which part you want me to substitute. 'an already
trusted device' makes me wonder whether you are thinking about reading the
list of digests, and not the general case of reading a kernel file, which
this paragraph intends to describe.

quoted

device is in the list of trusted digests. The file can be loaded if the
verity device has a trusted root digest.

Background:

As of now LoadPin restricts loading of kernel files to a single pinned
filesystem, typically the rootfs. This works for many systems, however it
can result in a bloated rootfs (and OTA updates) on platforms where
multiple boards with different hardware configurations use the same rootfs
image. Especially when 'optional' files are large it may be preferable to
download/install them only when they are actually needed by a given board.
Chrome OS uses Downloadable Content (DLC) [2] to deploy certain 'packages'
at runtime. As an example a DLC package could contain firmware for a
peripheral that is not present on all boards. DLCs use dm-verity to verify
the integrity of the DLC content.

[1] https://www.kernel.org/doc/html/latest/admin-guide/device-mapper/verity.html
[2] https://chromium.googlesource.com/chromiumos/platform2/+/HEAD/dlcservice/docs/developer.md

Signed-off-by: Matthias Kaehlcke <mka@chromium.org>
---

Changes in v3:
- added securityfs for LoadPin (currently only populated when
 CONFIG_SECURITY_LOADPIN_VERITY=y)
- added uapi include for LoadPin
- changed the interface for setting up the list of trusted
 digests from sysctl to ioctl on securityfs entry
- added stub for loadpin_is_fs_trusted() to be used
 CONFIG_SECURITY_LOADPIN_VERITY is not select
- depend on CONFIG_SECURITYFS instead of CONFIG_SYSTCL
- updated Kconfig help
- minor changes in read_trusted_verity_root_digests()
- updated commit message

Changes in v2:
- userspace now passes the path of the file with the verity digests
 via systcl, instead of the digests themselves
- renamed sysctl file to 'trusted_verity_root_digests_path'
- have CONFIG_SECURITY_LOADPIN_VERITY depend on CONFIG_SYSCTL
- updated Kconfig doc
- updated commit message

include/uapi/linux/loadpin.h |  19 ++++
security/loadpin/Kconfig     |  16 +++
security/loadpin/loadpin.c   | 184 ++++++++++++++++++++++++++++++++++-
3 files changed, 218 insertions(+), 1 deletion(-)
create mode 100644 include/uapi/linux/loadpin.h

diff --git a/security/loadpin/loadpin.c b/security/loadpin/loadpin.c
index b12f7d986b1e..c29ce562a366 100644
--- a/security/loadpin/loadpin.c
+++ b/security/loadpin/loadpin.c

...

quoted

+static int read_trusted_verity_root_digests(unsigned int fd)
+{
+	struct fd f;
+	void *data;

Probably easier if this is u8 *?

Maybe slightly, it would then require a cast when passing it to
kernel_read_file()

quoted

+	int rc;
+	char *p, *d;
+
+	/* The list of trusted root digests can only be set up once */
+	if (!list_empty(&trusted_verity_root_digests))
+		return -EPERM;
+
+	f = fdget(fd);
+	if (!f.file)
+		return -EINVAL;
+
+	data = kzalloc(SZ_4K, GFP_KERNEL);
+	if (!data) {
+		rc = -ENOMEM;
+		goto err;
+	}
+
+	rc = kernel_read_file(f.file, 0, &data, SZ_4K - 1, NULL, READING_POLICY);
+	if (rc < 0)
+		goto err;
+
+	((char *)data)[rc] = '\0';
+
+	p = strim(data);
+	while ((d = strsep(&p, ",")) != NULL) {

Maybe be flexible and add newline as a separator too?

Sure, I can add that. I'd also be fine with just allowing a newline as
separator, which seems a reasonable format for a sysfs file.

quoted

+		int len = strlen(d);
+		struct trusted_root_digest *trd;
+
+		if (len % 2) {
+			rc = -EPROTO;
+			goto err;
+		}
+
+		len /= 2;
+
+		trd = kzalloc(sizeof(*trd), GFP_KERNEL);

With the struct change, this could be:

kzalloc(struct_size(trd, data, len), ...)

Will change

quoted

+		if (!trd) {
+			rc = -ENOMEM;
+			goto err;
+		}
+
+		trd->data = kzalloc(len, GFP_KERNEL);
+		if (!trd->data) {
+			kfree(trd);
+			rc = -ENOMEM;
+			goto err;
+		}
+
+		if (hex2bin(trd->data, d, len)) {
+			kfree(trd);
+			kfree(trd->data);
+			rc = -EPROTO;
+			goto err;
+		}
+
+		trd->len = len;
+
+		list_add_tail(&trd->node, &trusted_verity_root_digests);
+	}
+
+	kfree(data);
+	fdput(f);
+
+	if (!list_empty(&trusted_verity_root_digests))
+		dm_verity_loadpin_set_trusted_root_digests(&trusted_verity_root_digests);
+
+	return 0;
+
+err:
+	kfree(data);
+

Maybe add a comment that any load failure will invalidate the entire list?

ok

Otherwise looks good! The only other thing I can think of is pondering more
about more carefully failing closed. E.g. instead of just throwing away all
the other hashes on a file load failure, maybe lock out future attempts to
set it too? I'm not sure this is actually useful, though. :P it shouldn't be
possible to corrupt the file, etc. But in the universe where things like
DirtyCOW happens, I've gotten even more paranoid. ;)

Sure, we can do that