Re: [PATCH] fscrypt: add support for IV_INO_LBLK_32 policies

From: Eric Biggers <ebiggers@kernel.org>
To: "Theodore Y. Ts'o" <tytso@mit.edu>
Cc: linux-fscrypt@vger.kernel.org,
	linux-f2fs-devel@lists.sourceforge.net,
	linux-ext4@vger.kernel.org, linux-fsdevel@vger.kernel.org,
	linux-crypto@vger.kernel.org, linux-mmc@vger.kernel.org,
	Satya Tangirala <satyat@google.com>,
	Jaegeuk Kim <jaegeuk@kernel.org>,
	Paul Crowley <paulcrowley@google.com>
Subject: Re: [PATCH] fscrypt: add support for IV_INO_LBLK_32 policies
Date: Tue, 19 May 2020 07:45:01 -0700	[thread overview]
Message-ID: <20200519144501.GA857@sol.localdomain> (raw)
In-Reply-To: <20200519111321.GE2396055@mit.edu>

On Tue, May 19, 2020 at 07:13:21AM -0400, Theodore Y. Ts'o wrote:
> On Fri, May 15, 2020 at 01:41:41PM -0700, Eric Biggers wrote:
> > From: Eric Biggers <ebiggers@google.com>
> > 
> > The eMMC inline crypto standard will only specify 32 DUN bits (a.k.a. IV
> > bits), unlike UFS's 64.  IV_INO_LBLK_64 is therefore not applicable, but
> > an encryption format which uses one key per policy and permits the
> > moving of encrypted file contents (as f2fs's garbage collector requires)
> > is still desirable.
> > 
> > To support such hardware, add a new encryption format IV_INO_LBLK_32
> > that makes the best use of the 32 bits: the IV is set to
> > 'SipHash-2-4(inode_number) + file_logical_block_number mod 2^32', where
> > the SipHash key is derived from the fscrypt master key.  We hash only
> > the inode number and not also the block number, because we need to
> > maintain contiguity of DUNs to merge bios.
> > 
> > Unlike with IV_INO_LBLK_64, with this format IV reuse is possible; this
> > is unavoidable given the size of the DUN.  This means this format should
> > only be used where the requirements of the first paragraph apply.
> > However, the hash spreads out the IVs in the whole usable range, and the
> > use of a keyed hash makes it difficult for an attacker to determine
> > which files use which IVs.
> > 
> > Besides the above differences, this flag works like IV_INO_LBLK_64 in
> > that on ext4 it is only allowed if the stable_inodes feature has been
> > enabled to prevent inode numbers and the filesystem UUID from changing.
> > 
> > Signed-off-by: Eric Biggers <ebiggers@google.com>
> 
> Reviewed-by: Theodore Ts'o <tytso@mit.edu>
> 
> I kind of wish we had Kunit tests with test vectors, but that's for
> another commit I think.
> 

We do have ciphertext verification tests in xfstests for all the existing
fscrypt options.  Actually, I had hacked one together for IV_INO_LBLK_32 before
sending this patch
(https://git.kernel.org/pub/scm/linux/kernel/git/ebiggers/xfstests-dev.git/commit/?id=55153ceee2948269c0359bd97fc0d58a26139c87).
I'll be sending it for review after I've looked over it again.

Similarly, since earlier this year, we now also have ciphertext verification
tests in Android's VTS (Vendor Test Suite)
(https://android.googlesource.com/platform/test/vts-testcase/kernel/+/refs/heads/master/encryption/).
I'll be adding one for this new flag.

These ciphertext verification tests test the round-trip from the key added by
userspace to the data on-disk -- even if the data is encrypted by inline crypto
hardware rather than the kernel itself.  So they're better than Kunit tests.

The thing I'm struggling with a bit is actually that when inline crypto is used,
IV_INO_LBLK_32 introduces a case where the DUN can wrap from 0xffffffff to 0,
and that case is new/special in that blocks can't be merged over that boundary
even if they are both logically and physically contiguous.  So, we could also
use a test that tests doing I/O around this boundary where the DUN wraps around.

- Eric