After FS_IOC_REMOVE_ENCRYPTION_KEY removes a key, it syncs the
filesystem and tries to get and put all inodes that were unlocked by the
key so that unused inodes get evicted via fscrypt_drop_inode().
Normally, the inodes are all clean due to the sync.
However, after the filesystem is sync'ed, userspace can modify and close
one of the files. (Userspace is *supposed* to close the files before
removing the key. But it doesn't always happen, and the kernel can't
assume it.) This causes the inode to be dirtied and have i_count == 0.
Then, fscrypt_drop_inode() failed to consider this case and indicated
that the inode can be dropped, causing the write to be lost.
On f2fs, other problems such as a filesystem freeze could occur due to
the inode being freed while still on f2fs's dirty inode list.
Fix this bug by making fscrypt_drop_inode() only drop clean inodes.
I've written an xfstest which detects this bug on ext4, f2fs, and ubifs.
Fixes: b1c0ec3599 ("fscrypt: add FS_IOC_REMOVE_ENCRYPTION_KEY ioctl")
Cc: <stable@vger.kernel.org> # v5.4+
Link: https://lore.kernel.org/r/20200305084138.653498-1-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>
Now that there's sometimes a second type of per-file key (the dirhash
key), clarify some function names, macros, and documentation that
specifically deal with per-file *encryption* keys.
Link: https://lore.kernel.org/r/20200120223201.241390-4-ebiggers@kernel.org
Reviewed-by: Daniel Rosenberg <drosen@google.com>
Signed-off-by: Eric Biggers <ebiggers@google.com>
When we allow indexed directories to use both encryption and
casefolding, for the dirhash we can't just hash the ciphertext filenames
that are stored on-disk (as is done currently) because the dirhash must
be case insensitive, but the stored names are case-preserving. Nor can
we hash the plaintext names with an unkeyed hash (or a hash keyed with a
value stored on-disk like ext4's s_hash_seed), since that would leak
information about the names that encryption is meant to protect.
Instead, if we can accept a dirhash that's only computable when the
fscrypt key is available, we can hash the plaintext names with a keyed
hash using a secret key derived from the directory's fscrypt master key.
We'll use SipHash-2-4 for this purpose.
Prepare for this by deriving a SipHash key for each casefolded encrypted
directory. Make sure to handle deriving the key not only when setting
up the directory's fscrypt_info, but also in the case where the casefold
flag is enabled after the fscrypt_info was already set up. (We could
just always derive the key regardless of casefolding, but that would
introduce unnecessary overhead for people not using casefolding.)
Signed-off-by: Daniel Rosenberg <drosen@google.com>
[EB: improved commit message, updated fscrypt.rst, squashed with change
that avoids unnecessarily deriving the key, and many other cleanups]
Link: https://lore.kernel.org/r/20200120223201.241390-3-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>
FSCRYPT_POLICY_FLAG_DIRECT_KEY is currently only allowed with Adiantum
encryption. But FS_IOC_SET_ENCRYPTION_POLICY allowed it in combination
with other encryption modes, and an error wasn't reported until later
when the encrypted directory was actually used.
Fix it to report the error earlier by validating the correct use of the
DIRECT_KEY flag in fscrypt_supported_policy(), similar to how we
validate the IV_INO_LBLK_64 flag.
Link: https://lore.kernel.org/r/20191209211829.239800-3-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>
As a sanity check, verify that the allocated crypto_skcipher actually
has the ivsize that fscrypt is assuming it has. This will always be the
case unless there's a bug. But if there ever is such a bug (e.g. like
there was in earlier versions of the ESSIV conversion patch [1]) it's
preferable for it to be immediately obvious, and not rely on the
ciphertext verification tests failing due to uninitialized IV bytes.
[1] https://lkml.kernel.org/linux-crypto/20190702215517.GA69157@gmail.com/
Link: https://lore.kernel.org/r/20191209203918.225691-1-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>
Crypto API users shouldn't really be accessing struct skcipher_alg
directly. <crypto/skcipher.h> already has a function
crypto_skcipher_driver_name(), so use that instead.
No change in behavior.
Link: https://lore.kernel.org/r/20191209203810.225302-1-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>
Inline encryption hardware compliant with the UFS v2.1 standard or with
the upcoming version of the eMMC standard has the following properties:
(1) Per I/O request, the encryption key is specified by a previously
loaded keyslot. There might be only a small number of keyslots.
(2) Per I/O request, the starting IV is specified by a 64-bit "data unit
number" (DUN). IV bits 64-127 are assumed to be 0. The hardware
automatically increments the DUN for each "data unit" of
configurable size in the request, e.g. for each filesystem block.
Property (1) makes it inefficient to use the traditional fscrypt
per-file keys. Property (2) precludes the use of the existing
DIRECT_KEY fscrypt policy flag, which needs at least 192 IV bits.
Therefore, add a new fscrypt policy flag IV_INO_LBLK_64 which causes the
encryption to modified as follows:
- The encryption keys are derived from the master key, encryption mode
number, and filesystem UUID.
- The IVs are chosen as (inode_number << 32) | file_logical_block_num.
For filenames encryption, file_logical_block_num is 0.
Since the file nonces aren't used in the key derivation, many files may
share the same encryption key. This is much more efficient on the
target hardware. Including the inode number in the IVs and mixing the
filesystem UUID into the keys ensures that data in different files is
nevertheless still encrypted differently.
Additionally, limiting the inode and block numbers to 32 bits and
placing the block number in the low bits maintains compatibility with
the 64-bit DUN convention (property (2) above).
Since this scheme assumes that inode numbers are stable (which may
preclude filesystem shrinking) and that inode and file logical block
numbers are at most 32-bit, IV_INO_LBLK_64 will only be allowed on
filesystems that meet these constraints. These are acceptable
limitations for the cases where this format would actually be used.
Note that IV_INO_LBLK_64 is an on-disk format, not an implementation.
This patch just adds support for it using the existing filesystem layer
encryption. A later patch will add support for inline encryption.
Reviewed-by: Paul Crowley <paulcrowley@google.com>
Co-developed-by: Satya Tangirala <satyat@google.com>
Signed-off-by: Satya Tangirala <satyat@google.com>
Signed-off-by: Eric Biggers <ebiggers@google.com>
The access to logged_impl_name is technically a data race, which tools
like KCSAN could complain about in the future. See:
https://github.com/google/ktsan/wiki/READ_ONCE-and-WRITE_ONCE
Fix by using xchg(), which also ensures that only one thread does the
logging.
This also required switching from bool to int, to avoid a build error on
the RISC-V architecture which doesn't implement xchg on bytes.
Signed-off-by: Eric Biggers <ebiggers@google.com>
memset the struct fscrypt_info to zero before freeing. This isn't
really needed currently, since there's no secret key directly in the
fscrypt_info. But there's a decent chance that someone will add such a
field in the future, e.g. in order to use an API that takes a raw key
such as siphash(). So it's good to do this as a hardening measure.
Signed-off-by: Eric Biggers <ebiggers@google.com>
Instead of open-coding the calculations for ESSIV handling, use an ESSIV
skcipher which does all of this under the hood. ESSIV was added to the
crypto API in v5.4.
This is based on a patch from Ard Biesheuvel, but reworked to apply
after all the fscrypt changes that went into v5.4.
Tested with 'kvm-xfstests -c ext4,f2fs -g encrypt', including the
ciphertext verification tests for v1 and v2 encryption policies.
Originally-from: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Acked-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Allow the FS_IOC_ADD_ENCRYPTION_KEY and FS_IOC_REMOVE_ENCRYPTION_KEY
ioctls to be used by non-root users to add and remove encryption keys
from the filesystem-level crypto keyrings, subject to limitations.
Motivation: while privileged fscrypt key management is sufficient for
some users (e.g. Android and Chromium OS, where a privileged process
manages all keys), the old API by design also allows non-root users to
set up and use encrypted directories, and we don't want to regress on
that. Especially, we don't want to force users to continue using the
old API, running into the visibility mismatch between files and keyrings
and being unable to "lock" encrypted directories.
Intuitively, the ioctls have to be privileged since they manipulate
filesystem-level state. However, it's actually safe to make them
unprivileged if we very carefully enforce some specific limitations.
First, each key must be identified by a cryptographic hash so that a
user can't add the wrong key for another user's files. For v2
encryption policies, we use the key_identifier for this. v1 policies
don't have this, so managing keys for them remains privileged.
Second, each key a user adds is charged to their quota for the keyrings
service. Thus, a user can't exhaust memory by adding a huge number of
keys. By default each non-root user is allowed up to 200 keys; this can
be changed using the existing sysctl 'kernel.keys.maxkeys'.
Third, if multiple users add the same key, we keep track of those users
of the key (of which there remains a single copy), and won't really
remove the key, i.e. "lock" the encrypted files, until all those users
have removed it. This prevents denial of service attacks that would be
possible under simpler schemes, such allowing the first user who added a
key to remove it -- since that could be a malicious user who has
compromised the key. Of course, encryption keys should be kept secret,
but the idea is that using encryption should never be *less* secure than
not using encryption, even if your key was compromised.
We tolerate that a user will be unable to really remove a key, i.e.
unable to "lock" their encrypted files, if another user has added the
same key. But in a sense, this is actually a good thing because it will
avoid providing a false notion of security where a key appears to have
been removed when actually it's still in memory, available to any
attacker who compromises the operating system kernel.
Reviewed-by: Theodore Ts'o <tytso@mit.edu>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Add a new fscrypt policy version, "v2". It has the following changes
from the original policy version, which we call "v1" (*):
- Master keys (the user-provided encryption keys) are only ever used as
input to HKDF-SHA512. This is more flexible and less error-prone, and
it avoids the quirks and limitations of the AES-128-ECB based KDF.
Three classes of cryptographically isolated subkeys are defined:
- Per-file keys, like used in v1 policies except for the new KDF.
- Per-mode keys. These implement the semantics of the DIRECT_KEY
flag, which for v1 policies made the master key be used directly.
These are also planned to be used for inline encryption when
support for it is added.
- Key identifiers (see below).
- Each master key is identified by a 16-byte master_key_identifier,
which is derived from the key itself using HKDF-SHA512. This prevents
users from associating the wrong key with an encrypted file or
directory. This was easily possible with v1 policies, which
identified the key by an arbitrary 8-byte master_key_descriptor.
- The key must be provided in the filesystem-level keyring, not in a
process-subscribed keyring.
The following UAPI additions are made:
- The existing ioctl FS_IOC_SET_ENCRYPTION_POLICY can now be passed a
fscrypt_policy_v2 to set a v2 encryption policy. It's disambiguated
from fscrypt_policy/fscrypt_policy_v1 by the version code prefix.
- A new ioctl FS_IOC_GET_ENCRYPTION_POLICY_EX is added. It allows
getting the v1 or v2 encryption policy of an encrypted file or
directory. The existing FS_IOC_GET_ENCRYPTION_POLICY ioctl could not
be used because it did not have a way for userspace to indicate which
policy structure is expected. The new ioctl includes a size field, so
it is extensible to future fscrypt policy versions.
- The ioctls FS_IOC_ADD_ENCRYPTION_KEY, FS_IOC_REMOVE_ENCRYPTION_KEY,
and FS_IOC_GET_ENCRYPTION_KEY_STATUS now support managing keys for v2
encryption policies. Such keys are kept logically separate from keys
for v1 encryption policies, and are identified by 'identifier' rather
than by 'descriptor'. The 'identifier' need not be provided when
adding a key, since the kernel will calculate it anyway.
This patch temporarily keeps adding/removing v2 policy keys behind the
same permission check done for adding/removing v1 policy keys:
capable(CAP_SYS_ADMIN). However, the next patch will carefully take
advantage of the cryptographically secure master_key_identifier to allow
non-root users to add/remove v2 policy keys, thus providing a full
replacement for v1 policies.
(*) Actually, in the API fscrypt_policy::version is 0 while on-disk
fscrypt_context::format is 1. But I believe it makes the most sense
to advance both to '2' to have them be in sync, and to consider the
numbering to start at 1 except for the API quirk.
Reviewed-by: Paul Crowley <paulcrowley@google.com>
Reviewed-by: Theodore Ts'o <tytso@mit.edu>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Add a new fscrypt ioctl, FS_IOC_REMOVE_ENCRYPTION_KEY. This ioctl
removes an encryption key that was added by FS_IOC_ADD_ENCRYPTION_KEY.
It wipes the secret key itself, then "locks" the encrypted files and
directories that had been unlocked using that key -- implemented by
evicting the relevant dentries and inodes from the VFS caches.
The problem this solves is that many fscrypt users want the ability to
remove encryption keys, causing the corresponding encrypted directories
to appear "locked" (presented in ciphertext form) again. Moreover,
users want removing an encryption key to *really* remove it, in the
sense that the removed keys cannot be recovered even if kernel memory is
compromised, e.g. by the exploit of a kernel security vulnerability or
by a physical attack. This is desirable after a user logs out of the
system, for example. In many cases users even already assume this to be
the case and are surprised to hear when it's not.
It is not sufficient to simply unlink the master key from the keyring
(or to revoke or invalidate it), since the actual encryption transform
objects are still pinned in memory by their inodes. Therefore, to
really remove a key we must also evict the relevant inodes.
Currently one workaround is to run 'sync && echo 2 >
/proc/sys/vm/drop_caches'. But, that evicts all unused inodes in the
system rather than just the inodes associated with the key being
removed, causing severe performance problems. Moreover, it requires
root privileges, so regular users can't "lock" their encrypted files.
Another workaround, used in Chromium OS kernels, is to add a new
VFS-level ioctl FS_IOC_DROP_CACHE which is a more restricted version of
drop_caches that operates on a single super_block. It does:
shrink_dcache_sb(sb);
invalidate_inodes(sb, false);
But it's still a hack. Yet, the major users of filesystem encryption
want this feature badly enough that they are actually using these hacks.
To properly solve the problem, start maintaining a list of the inodes
which have been "unlocked" using each master key. Originally this
wasn't possible because the kernel didn't keep track of in-use master
keys at all. But, with the ->s_master_keys keyring it is now possible.
Then, add an ioctl FS_IOC_REMOVE_ENCRYPTION_KEY. It finds the specified
master key in ->s_master_keys, then wipes the secret key itself, which
prevents any additional inodes from being unlocked with the key. Then,
it syncs the filesystem and evicts the inodes in the key's list. The
normal inode eviction code will free and wipe the per-file keys (in
->i_crypt_info). Note that freeing ->i_crypt_info without evicting the
inodes was also considered, but would have been racy.
Some inodes may still be in use when a master key is removed, and we
can't simply revoke random file descriptors, mmap's, etc. Thus, the
ioctl simply skips in-use inodes, and returns -EBUSY to indicate that
some inodes weren't evicted. The master key *secret* is still removed,
but the fscrypt_master_key struct remains to keep track of the remaining
inodes. Userspace can then retry the ioctl to evict the remaining
inodes. Alternatively, if userspace adds the key again, the refreshed
secret will be associated with the existing list of inodes so they
remain correctly tracked for future key removals.
The ioctl doesn't wipe pagecache pages. Thus, we tolerate that after a
kernel compromise some portions of plaintext file contents may still be
recoverable from memory. This can be solved by enabling page poisoning
system-wide, which security conscious users may choose to do. But it's
very difficult to solve otherwise, e.g. note that plaintext file
contents may have been read in other places than pagecache pages.
Like FS_IOC_ADD_ENCRYPTION_KEY, FS_IOC_REMOVE_ENCRYPTION_KEY is
initially restricted to privileged users only. This is sufficient for
some use cases, but not all. A later patch will relax this restriction,
but it will require introducing key hashes, among other changes.
Reviewed-by: Theodore Ts'o <tytso@mit.edu>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Add a new fscrypt ioctl, FS_IOC_ADD_ENCRYPTION_KEY. This ioctl adds an
encryption key to the filesystem's fscrypt keyring ->s_master_keys,
making any files encrypted with that key appear "unlocked".
Why we need this
~~~~~~~~~~~~~~~~
The main problem is that the "locked/unlocked" (ciphertext/plaintext)
status of encrypted files is global, but the fscrypt keys are not.
fscrypt only looks for keys in the keyring(s) the process accessing the
filesystem is subscribed to: the thread keyring, process keyring, and
session keyring, where the session keyring may contain the user keyring.
Therefore, userspace has to put fscrypt keys in the keyrings for
individual users or sessions. But this means that when a process with a
different keyring tries to access encrypted files, whether they appear
"unlocked" or not is nondeterministic. This is because it depends on
whether the files are currently present in the inode cache.
Fixing this by consistently providing each process its own view of the
filesystem depending on whether it has the key or not isn't feasible due
to how the VFS caches work. Furthermore, while sometimes users expect
this behavior, it is misguided for two reasons. First, it would be an
OS-level access control mechanism largely redundant with existing access
control mechanisms such as UNIX file permissions, ACLs, LSMs, etc.
Encryption is actually for protecting the data at rest.
Second, almost all users of fscrypt actually do need the keys to be
global. The largest users of fscrypt, Android and Chromium OS, achieve
this by having PID 1 create a "session keyring" that is inherited by
every process. This works, but it isn't scalable because it prevents
session keyrings from being used for any other purpose.
On general-purpose Linux distros, the 'fscrypt' userspace tool [1] can't
similarly abuse the session keyring, so to make 'sudo' work on all
systems it has to link all the user keyrings into root's user keyring
[2]. This is ugly and raises security concerns. Moreover it can't make
the keys available to system services, such as sshd trying to access the
user's '~/.ssh' directory (see [3], [4]) or NetworkManager trying to
read certificates from the user's home directory (see [5]); or to Docker
containers (see [6], [7]).
By having an API to add a key to the *filesystem* we'll be able to fix
the above bugs, remove userspace workarounds, and clearly express the
intended semantics: the locked/unlocked status of an encrypted directory
is global, and encryption is orthogonal to OS-level access control.
Why not use the add_key() syscall
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We use an ioctl for this API rather than the existing add_key() system
call because the ioctl gives us the flexibility needed to implement
fscrypt-specific semantics that will be introduced in later patches:
- Supporting key removal with the semantics such that the secret is
removed immediately and any unused inodes using the key are evicted;
also, the eviction of any in-use inodes can be retried.
- Calculating a key-dependent cryptographic identifier and returning it
to userspace.
- Allowing keys to be added and removed by non-root users, but only keys
for v2 encryption policies; and to prevent denial-of-service attacks,
users can only remove keys they themselves have added, and a key is
only really removed after all users who added it have removed it.
Trying to shoehorn these semantics into the keyrings syscalls would be
very difficult, whereas the ioctls make things much easier.
However, to reuse code the implementation still uses the keyrings
service internally. Thus we get lockless RCU-mode key lookups without
having to re-implement it, and the keys automatically show up in
/proc/keys for debugging purposes.
References:
[1] https://github.com/google/fscrypt
[2] https://goo.gl/55cCrI#heading=h.vf09isp98isb
[3] https://github.com/google/fscrypt/issues/111#issuecomment-444347939
[4] https://github.com/google/fscrypt/issues/116
[5] https://bugs.launchpad.net/ubuntu/+source/fscrypt/+bug/1770715
[6] https://github.com/google/fscrypt/issues/128
[7] https://askubuntu.com/questions/1130306/cannot-run-docker-on-an-encrypted-filesystem
Reviewed-by: Theodore Ts'o <tytso@mit.edu>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Rename keyinfo.c to keysetup.c since this better describes what the file
does (sets up the key), and it matches the new file keysetup_v1.c.
Reviewed-by: Theodore Ts'o <tytso@mit.edu>
Signed-off-by: Eric Biggers <ebiggers@google.com>
