Re: [RFC PATCH 1/8] fscrypt: add fscrypt_prepare_new_inode() and fscrypt_set_context()

[Eric Biggers <ebiggers@kernel.org>]

On Mon, Aug 24, 2020 at 03:42:59PM -0400, Jeff Layton wrote:
> On Mon, 2020-08-24 at 12:02 -0700, Eric Biggers wrote:
> > On Mon, Aug 24, 2020 at 02:47:07PM -0400, Jeff Layton wrote:
> > > On Mon, 2020-08-24 at 11:21 -0700, Eric Biggers wrote:
> > > > On Mon, Aug 24, 2020 at 12:48:48PM -0400, Jeff Layton wrote:
> > > > > > +void fscrypt_hash_inode_number(struct fscrypt_info *ci,
> > > > > > +			       const struct fscrypt_master_key *mk)
> > > > > > +{
> > > > > > +	WARN_ON(ci->ci_inode->i_ino == 0);
> > > > > > +	WARN_ON(!mk->mk_ino_hash_key_initialized);
> > > > > > +
> > > > > > +	ci->ci_hashed_ino = (u32)siphash_1u64(ci->ci_inode->i_ino,
> > > > > > +					      &mk->mk_ino_hash_key);
> > > > > 
> > > > > i_ino is an unsigned long. Will this produce a consistent results on
> > > > > arches with 32 and 64 bit long values? I think it'd be nice to ensure
> > > > > that we can access an encrypted directory created on a 32-bit host from
> > > > > (e.g.) a 64-bit host.
> > > > 
> > > > The result is the same regardless of word size and endianness.
> > > > siphash_1u64(v, k) is equivalent to:
> > > > 
> > > > 	__le64 x = cpu_to_le64(v);
> > > > 	siphash(&x, 8, k);
> > > > 
> > > 
> > > In the case where you have an (on-storage) inode number that is larger
> > > than 2^32, x will almost certainly be different on a 32 vs. 64-bit
> > > wordsize.
> > > 
> > > On the box with the 32-bit wordsize, you'll end up promoting i_ino to a
> > > 64-bit word and the upper 32 bits will be zeroed out. So it seems like
> > > this means that if you're using inline hardware you're going to end up
> > > with a result that won't work correctly across different wordsizes.
> > 
> > That's only possible if the VFS is truncating the inode number, which would also
> > break userspace in lots of ways like making applications think that files are
> > hard-linked together when they aren't.  Also, IV_INO_LBLK_64 would break.
> > 
> > The correct fix for that would be to make inode::i_ino 64-bit.
> > 
> 
> ...or just ask the filesystem for the 64-bit inode number via ->getattr
> or a new op. You could also just truncate it down to 32 bits or xor the
> top and bottom bits together first, etc...
> 
> > Note that ext4 and f2fs (currently the only filesystems that support the
> > IV_INO_LBLK_* flags) only support 32-bit inode numbers.
> > 
> 
> Ahh, ok. That explains why it's not been an issue so far. Still, if
> you're reworking this code anyway, you might want to consider avoiding
> i_ino here.

Let's just enforce ino_bits <= 32 for IV_INO_LBLK_32 for now,
like is done for IV_INO_LBLK_64:
https://lkml.kernel.org/r/20200824203841.1707847-1-ebiggers@kernel.org

There's no need to add extra complexity for something that no one wants yet.

(And as mentioned, this won't prevent ceph or other filesystems with 64-bit
inode numbers from adding support for fscrypt, as IV_INO_LBLK_32 support is
optional and has a pretty specific use case.)

- Eric