Pull ext4 updates from Ted Ts'o:
 "The first major feature for ext4 this merge window is the largedir
  feature, which allows ext4 directories to support over 2 billion
  directory entries (assuming ~64 byte file names; in practice, users
  will run into practical performance limits first.) This feature was
  originally written by the Lustre team, and credit goes to Artem
  Blagodarenko from Seagate for getting this feature upstream.

  The second major major feature allows ext4 to support extended
  attribute values up to 64k. This feature was also originally from
  Lustre, and has been enhanced by Tahsin Erdogan from Google with a
  deduplication feature so that if multiple files have the same xattr
  value (for example, Windows ACL's stored by Samba), only one copy will
  be stored on disk for encoding and caching efficiency.

  We also have the usual set of bug fixes, cleanups, and optimizations"

* tag 'ext4_for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4: (47 commits)
  ext4: fix spelling mistake: "prellocated" -> "preallocated"
  ext4: fix __ext4_new_inode() journal credits calculation
  ext4: skip ext4_init_security() and encryption on ea_inodes
  fs: generic_block_bmap(): initialize all of the fields in the temp bh
  ext4: change fast symlink test to not rely on i_blocks
  ext4: require key for truncate(2) of encrypted file
  ext4: don't bother checking for encryption key in ->mmap()
  ext4: check return value of kstrtoull correctly in reserved_clusters_store
  ext4: fix off-by-one fsmap error on 1k block filesystems
  ext4: return EFSBADCRC if a bad checksum error is found in ext4_find_entry()
  ext4: return EIO on read error in ext4_find_entry
  ext4: forbid encrypting root directory
  ext4: send parallel discards on commit completions
  ext4: avoid unnecessary stalls in ext4_evict_inode()
  ext4: add nombcache mount option
  ext4: strong binding of xattr inode references
  ext4: eliminate xattr entry e_hash recalculation for removes
  ext4: reserve space for xattr entries/names
  quota: add get_inode_usage callback to transfer multi-inode charges
  ext4: xattr inode deduplication
  ...

Pull fscrypt updates from Ted Ts'o:
 "Add support for 128-bit AES and some cleanups to fscrypt"

* tag 'fscrypt_for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/fscrypt:
  fscrypt: make ->dummy_context() return bool
  fscrypt: add support for AES-128-CBC
  fscrypt: inline fscrypt_free_filename()

ea_inode feature allows creating extended attributes that are up to
64k in size. Update __ext4_new_inode() to pick increased credit limits.

To avoid overallocating too many journal credits, update
__ext4_xattr_set_credits() to make a distinction between xattr create
vs update. This helps __ext4_new_inode() because all attributes are
known to be new, so we can save credits that are normally needed to
delete old values.

Also, have fscrypt specify its maximum context size so that we don't
end up allocating credits for 64k size.

Signed-off-by: Tahsin Erdogan <tahsin@google.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>

fscrypt provides facilities to use different encryption algorithms which
are selectable by userspace when setting the encryption policy. Currently,
only AES-256-XTS for file contents and AES-256-CBC-CTS for file names are
implemented. This is a clear case of kernel offers the mechanism and
userspace selects a policy. Similar to what dm-crypt and ecryptfs have.

This patch adds support for using AES-128-CBC for file contents and
AES-128-CBC-CTS for file name encryption. To mitigate watermarking
attacks, IVs are generated using the ESSIV algorithm. While AES-CBC is
actually slightly less secure than AES-XTS from a security point of view,
there is more widespread hardware support. Using AES-CBC gives us the
acceptable performance while still providing a moderate level of security
for persistent storage.

Especially low-powered embedded devices with crypto accelerators such as
CAAM or CESA often only support AES-CBC. Since using AES-CBC over AES-XTS
is basically thought of a last resort, we use AES-128-CBC over AES-256-CBC
since it has less encryption rounds and yields noticeable better
performance starting from a file size of just a few kB.

Signed-off-by: Daniel Walter <dwalter@sigma-star.at>
[david@sigma-star.at: addressed review comments]
Signed-off-by: David Gstir <david@sigma-star.at>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>

fscrypt_free_filename() only needs to do a kfree() of crypto_buf.name,
which works well as an inline function.  We can skip setting the various
pointers to NULL, since no user cares about it (the name is always freed
just before it goes out of scope).

Signed-off-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: David Gstir <david@sigma-star.at>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>

Replace bi_error with a new bi_status to allow for a clear conversion.
Note that device mapper overloaded bi_error with a private value, which
we'll have to keep arround at least for now and thus propagate to a
proper blk_status_t value.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>

Introduce a helper function fscrypt_match_name() which tests whether a
fscrypt_name matches a directory entry.  Also clean up the magic numbers
and document things properly.

Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>

When accessing an encrypted directory without the key, userspace must
operate on filenames derived from the ciphertext names, which contain
arbitrary bytes.  Since we must support filenames as long as NAME_MAX,
we can't always just base64-encode the ciphertext, since that may make
it too long.  Currently, this is solved by presenting long names in an
abbreviated form containing any needed filesystem-specific hashes (e.g.
to identify a directory block), then the last 16 bytes of ciphertext.
This needs to be sufficient to identify the actual name on lookup.

However, there is a bug.  It seems to have been assumed that due to the
use of a CBC (ciphertext block chaining)-based encryption mode, the last
16 bytes (i.e. the AES block size) of ciphertext would depend on the
full plaintext, preventing collisions.  However, we actually use CBC
with ciphertext stealing (CTS), which handles the last two blocks
specially, causing them to appear "flipped".  Thus, it's actually the
second-to-last block which depends on the full plaintext.

This caused long filenames that differ only near the end of their
plaintexts to, when observed without the key, point to the wrong inode
and be undeletable.  For example, with ext4:

    # echo pass | e4crypt add_key -p 16 edir/
    # seq -f "edir/abcdefghijklmnopqrstuvwxyz012345%.0f" 100000 | xargs touch
    # find edir/ -type f | xargs stat -c %i | sort | uniq | wc -l
    100000
    # sync
    # echo 3 > /proc/sys/vm/drop_caches
    # keyctl new_session
    # find edir/ -type f | xargs stat -c %i | sort | uniq | wc -l
    2004
    # rm -rf edir/
    rm: cannot remove 'edir/_A7nNFi3rhkEQlJ6P,hdzluhODKOeWx5V': Structure needs cleaning
    ...

To fix this, when presenting long encrypted filenames, encode the
second-to-last block of ciphertext rather than the last 16 bytes.

Although it would be nice to solve this without depending on a specific
encryption mode, that would mean doing a cryptographic hash like SHA-256
which would be much less efficient.  This way is sufficient for now, and
it's still compatible with encryption modes like HEH which are strong
pseudorandom permutations.  Also, changing the presented names is still
allowed at any time because they are only provided to allow applications
to do things like delete encrypted directories.  They're not designed to
be used to persistently identify files --- which would be hard to do
anyway, given that they're encrypted after all.

For ease of backports, this patch only makes the minimal fix to both
ext4 and f2fs.  It leaves ubifs as-is, since ubifs doesn't compare the
ciphertext block yet.  Follow-on patches will clean things up properly
and make the filesystems use a shared helper function.

Fixes: 5de0b4d0cd15 ("ext4 crypto: simplify and speed up filename encryption")
Reported-by: Gwendal Grignou <gwendal@chromium.org>
Cc: stable@vger.kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>

To mitigate some types of offline attacks, filesystem encryption is
designed to enforce that all files in an encrypted directory tree use
the same encryption policy (i.e. the same encryption context excluding
the nonce).  However, the fscrypt_has_permitted_context() function which
enforces this relies on comparing struct fscrypt_info's, which are only
available when we have the encryption keys.  This can cause two
incorrect behaviors:

1. If we have the parent directory's key but not the child's key, or
   vice versa, then fscrypt_has_permitted_context() returned false,
   causing applications to see EPERM or ENOKEY.  This is incorrect if
   the encryption contexts are in fact consistent.  Although we'd
   normally have either both keys or neither key in that case since the
   master_key_descriptors would be the same, this is not guaranteed
   because keys can be added or removed from keyrings at any time.

2. If we have neither the parent's key nor the child's key, then
   fscrypt_has_permitted_context() returned true, causing applications
   to see no error (or else an error for some other reason).  This is
   incorrect if the encryption contexts are in fact inconsistent, since
   in that case we should deny access.

To fix this, retrieve and compare the fscrypt_contexts if we are unable
to set up both fscrypt_infos.

While this slightly hurts performance when accessing an encrypted
directory tree without the key, this isn't a case we really need to be
optimizing for; access *with* the key is much more important.
Furthermore, the performance hit is barely noticeable given that we are
already retrieving the fscrypt_context and doing two keyring searches in
fscrypt_get_encryption_info().  If we ever actually wanted to optimize
this case we might start by caching the fscrypt_contexts.

Cc: stable@vger.kernel.org # 4.0+
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>

This commit exposes the necessary constants and structures for a
userspace program to pass filesystem encryption keys into the keyring.
The fscrypt_key structure was already part of the kernel ABI, this
change just makes it so programs no longer have to redeclare these
structures (like e4crypt in e2fsprogs currently does).

Note that we do not expose the other FS_*_KEY_SIZE constants as they are
not necessary. Only XTS is supported for contents_encryption_mode, so
currently FS_MAX_KEY_SIZE bytes of key material must always be passed to
the kernel.

This commit also removes __packed from fscrypt_key as it does not
contain any implicit padding and does not refer to an on-disk structure.

Signed-off-by: Joe Richey <joerichey@google.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>