On Jul 13, 2017, at 3:58 PM, Eric Biggers <ebigge...@gmail.com> wrote: > > Hi Michael, > > On Thu, Jul 13, 2017 at 03:29:44PM -0700, Michael Halcrow wrote: >> On Wed, Jul 12, 2017 at 02:00:30PM -0700, Eric Biggers wrote: >>> From: Eric Biggers <ebigg...@google.com> >>> >>> Currently, the fscrypt_context (i.e. the encryption xattr) does not >>> contain a cryptographically secure identifier for the master key's >>> payload. Therefore it's not possible to verify that the correct key was >>> supplied, which is problematic in multi-user scenarios. To make this >>> possible, define a new fscrypt_context version (v2) which includes a >>> key_hash field, and allow userspace to opt-in to it when setting an >>> encryption policy by setting fscrypt_policy.version to 2. For now just >>> zero the new field; a later patch will start setting it for real. >> >> The main concern that comes to mind is potentially blowing past the >> inline xattr size limit and allocating a new inode block. The >> security benefit probably outweighs that concern in this case. >> > > The way it adds up now for ext4 is: > > 128 bytes for base inode > + 32 bytes for i_extra fields > + 4 bytes for in-inode xattrs header > + 20 bytes for encryption xattr header + name > + 28 bytes for encryption xattr value > ---------------------------------- > = 212 bytes total. > > By adding the 16-byte 'key_hash' field it grows to 228 bytes total. So it > still > fits in a 256-byte inode, though it's getting closer to the limit. We could > save 8 bytes by instead using the design where master_key_descriptor is > extended > to 16 bytes and redefined as a cryptographically secure hash. But as noted, > that has some significant disadvantages. > > Also note that we don't really have to worry about leaving space for a SELinux > xattr anymore because with 256-byte inodes + encryption the SELinux xattr is > already being written to an external block, given that it requires about 52-62 > bytes (at least when using Android's SELinux policy; different SELinux > policies > may use different values), and 212 + 52 > 256. So if someone wants both > xattrs > in-inode they need to use 512-byte inodes already.
It is probably time to consider changing to a default of 512-byte inodes for larger filesystems anyway. In our testing, this affected performance only by a couple of percent under normal usage, and avoided a significant performance drop if the xattrs ever fall out of the inode. Cheers, Andreas
Description: Message signed with OpenPGP