Typically, the cryptographic APIs that fscrypt uses take keys as byte
arrays, which avoids endianness issues.  However, siphash_key_t is an
exception.  It is defined as 'u64 key[2];', i.e. the 128-bit key is
expected to be given directly as two 64-bit words in CPU endianness.

fscrypt_derive_dirhash_key() and fscrypt_setup_iv_ino_lblk_32_key()
forgot to take this into account.  Therefore, the SipHash keys used to
index encrypted+casefolded directories differ on big endian vs. little
endian platforms, as do the SipHash keys used to hash inode numbers for
IV_INO_LBLK_32-encrypted directories.  This makes such directories
non-portable between these platforms.

Fix this by always using the little endian order.  This is a breaking
change for big endian platforms, but this should be fine in practice
since these features (encrypt+casefold support, and the IV_INO_LBLK_32
flag) aren't known to actually be used on any big endian platforms yet.

Fixes: aa408f835d02 ("fscrypt: derive dirhash key for casefolded directories")
Fixes: e3b1078bedd3 ("fscrypt: add support for IV_INO_LBLK_32 policies")
Cc: <stable@vger.kernel.org> # v5.6+
Link: https://lore.kernel.org/r/20210605075033.54424-1-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>

Currently it's impossible to delete files that use an unsupported
encryption policy, as the kernel will just return an error when
performing any operation on the top-level encrypted directory, even just
a path lookup into the directory or opening the directory for readdir.

More specifically, this occurs in any of the following cases:

- The encryption context has an unrecognized version number.  Current
  kernels know about v1 and v2, but there could be more versions in the
  future.

- The encryption context has unrecognized encryption modes
  (FSCRYPT_MODE_*) or flags (FSCRYPT_POLICY_FLAG_*), an unrecognized
  combination of modes, or reserved bits set.

- The encryption key has been added and the encryption modes are
  recognized but aren't available in the crypto API -- for example, a
  directory is encrypted with FSCRYPT_MODE_ADIANTUM but the kernel
  doesn't have CONFIG_CRYPTO_ADIANTUM enabled.

It's desirable to return errors for most operations on files that use an
unsupported encryption policy, but the current behavior is too strict.
We need to allow enough to delete files, so that people can't be stuck
with undeletable files when downgrading kernel versions.  That includes
allowing directories to be listed and allowing dentries to be looked up.

Fix this by modifying the key setup logic to treat an unsupported
encryption policy in the same way as "key unavailable" in the cases that
are required for a recursive delete to work: preparing for a readdir or
a dentry lookup, revalidating a dentry, or checking whether an inode has
the same encryption policy as its parent directory.

Reviewed-by: Andreas Dilger <adilger@dilger.ca>
Link: https://lore.kernel.org/r/20201203022041.230976-10-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>

Now that fscrypt_get_encryption_info() is only called from files in
fs/crypto/ (due to all key setup now being handled by higher-level
helper functions instead of directly by filesystems), unexport it and
move its declaration to fscrypt_private.h.

Reviewed-by: Andreas Dilger <adilger@dilger.ca>
Link: https://lore.kernel.org/r/20201203022041.230976-9-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>

The stated reasons for separating fscrypt_master_key::mk_secret_sem from
the standard semaphore contained in every 'struct key' no longer apply.

First, due to commit a992b20cd4ee ("fscrypt: add
fscrypt_prepare_new_inode() and fscrypt_set_context()"),
fscrypt_get_encryption_info() is no longer called from within a
filesystem transaction.

Second, due to commit d3ec10aa9581 ("KEYS: Don't write out to userspace
while holding key semaphore"), the semaphore for the "keyring" key type
no longer ranks above page faults.

That leaves performance as the only possible reason to keep the separate
mk_secret_sem.  Specifically, having mk_secret_sem reduces the
contention between setup_file_encryption_key() and
FS_IOC_{ADD,REMOVE}_ENCRYPTION_KEY.  However, these ioctls aren't
executed often, so this doesn't seem to be worth the extra complexity.

Therefore, simplify the locking design by just using key->sem instead of
mk_secret_sem.

Link: https://lore.kernel.org/r/20201117032626.320275-1-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>

There isn't really any valid reason to use __FSCRYPT_MODE_MAX or
FSCRYPT_POLICY_FLAGS_VALID in a userspace program.  These constants are
only meant to be used by the kernel internally, and they are defined in
the UAPI header next to the mode numbers and flags only so that kernel
developers don't forget to update them when adding new modes or flags.

In https://lkml.kernel.org/r/20201005074133.1958633-2-satyat@google.com
there was an example of someone wanting to use __FSCRYPT_MODE_MAX in a
user program, and it was wrong because the program would have broken if
__FSCRYPT_MODE_MAX were ever increased.  So having this definition
available is harmful.  FSCRYPT_POLICY_FLAGS_VALID has the same problem.

So, remove these definitions from the UAPI header.  Replace
FSCRYPT_POLICY_FLAGS_VALID with just listing the valid flags explicitly
in the one kernel function that needs it.  Move __FSCRYPT_MODE_MAX to
fscrypt_private.h, remove the double underscores (which were only
present to discourage use by userspace), and add a BUILD_BUG_ON() and
comments to (hopefully) ensure it is kept in sync.

Keep the old name FS_POLICY_FLAGS_VALID, since it's been around for
longer and there's a greater chance that removing it would break source
compatibility with some program.  Indeed, mtd-utils is using it in
an #ifdef, and removing it would introduce compiler warnings (about
FS_POLICY_FLAGS_PAD_* being redefined) into the mtd-utils build.
However, reduce its value to 0x07 so that it only includes the flags
with old names (the ones present before Linux 5.4), and try to make it
clear that it's now "frozen" and no new flags should be added to it.

Fixes: 2336d0deb2d4 ("fscrypt: use FSCRYPT_ prefix for uapi constants")
Cc: <stable@vger.kernel.org> # v5.4+
Link: https://lore.kernel.org/r/20201024005132.495952-1-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>

I_CREATING isn't actually set until the inode has been assigned an inode
number and inserted into the inode hash table.  So the WARN_ON() in
fscrypt_setup_iv_ino_lblk_32_key() is wrong, and it can trigger when
creating an encrypted file on ext4.  Remove it.

This was sometimes causing xfstest generic/602 to fail on ext4.  I
didn't notice it before because due to a separate oversight, new inodes
that haven't been assigned an inode number yet don't necessarily have
i_ino == 0 as I had thought, so by chance I never saw the test fail.

Fixes: a992b20cd4ee ("fscrypt: add fscrypt_prepare_new_inode() and fscrypt_set_context()")
Reported-by: Theodore Y. Ts'o <tytso@mit.edu>
Link: https://lore.kernel.org/r/20201031004556.87862-1-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>

The behavior of the test_dummy_encryption mount option is that when a
new file (or directory or symlink) is created in an unencrypted
directory, it's automatically encrypted using a dummy encryption policy.
That's it; in particular, the encryption (or lack thereof) of existing
files (or directories or symlinks) doesn't change.

Unfortunately the implementation of test_dummy_encryption is a bit weird
and confusing.  When test_dummy_encryption is enabled and a file is
being created in an unencrypted directory, we set up an encryption key
(->i_crypt_info) for the directory.  This isn't actually used to do any
encryption, however, since the directory is still unencrypted!  Instead,
->i_crypt_info is only used for inheriting the encryption policy.

One consequence of this is that the filesystem ends up providing a
"dummy context" (policy + nonce) instead of a "dummy policy".  In
commit ed318a6cc0b6 ("fscrypt: support test_dummy_encryption=v2"), I
mistakenly thought this was required.  However, actually the nonce only
ends up being used to derive a key that is never used.

Another consequence of this implementation is that it allows for
'inode->i_crypt_info != NULL && !IS_ENCRYPTED(inode)', which is an edge
case that can be forgotten about.  For example, currently
FS_IOC_GET_ENCRYPTION_POLICY on an unencrypted directory may return the
dummy encryption policy when the filesystem is mounted with
test_dummy_encryption.  That seems like the wrong thing to do, since
again, the directory itself is not actually encrypted.

Therefore, switch to a more logical and maintainable implementation
where the dummy encryption policy inheritance is done without setting up
keys for unencrypted directories.  This involves:

- Adding a function fscrypt_policy_to_inherit() which returns the
  encryption policy to inherit from a directory.  This can be a real
  policy, a dummy policy, or no policy.

- Replacing struct fscrypt_dummy_context, ->get_dummy_context(), etc.
  with struct fscrypt_dummy_policy, ->get_dummy_policy(), etc.

- Making fscrypt_fname_encrypted_size() take an fscrypt_policy instead
  of an inode.

Acked-by: Jaegeuk Kim <jaegeuk@kernel.org>
Acked-by: Jeff Layton <jlayton@kernel.org>
Link: https://lore.kernel.org/r/20200917041136.178600-13-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>

fscrypt_get_encryption_info() has never actually been safe to call in a
context that needs GFP_NOFS, since it calls crypto_alloc_skcipher().

crypto_alloc_skcipher() isn't GFP_NOFS-safe, even if called under
memalloc_nofs_save().  This is because it may load kernel modules, and
also because it internally takes crypto_alg_sem.  Other tasks can do
GFP_KERNEL allocations while holding crypto_alg_sem for write.

The use of fscrypt_init_mutex isn't GFP_NOFS-safe either.

So, stop pretending that fscrypt_get_encryption_info() is nofs-safe.
I.e., when it allocates memory, just use GFP_KERNEL instead of GFP_NOFS.

Note, another reason to do this is that GFP_NOFS is deprecated in favor
of using memalloc_nofs_save() in the proper places.

Acked-by: Jeff Layton <jlayton@kernel.org>
Link: https://lore.kernel.org/r/20200917041136.178600-10-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>

fscrypt_get_encryption_info() is intended to be GFP_NOFS-safe.  But
actually it isn't, since it uses functions like crypto_alloc_skcipher()
which aren't GFP_NOFS-safe, even when called under memalloc_nofs_save().
Therefore it can deadlock when called from a context that needs
GFP_NOFS, e.g. during an ext4 transaction or between f2fs_lock_op() and
f2fs_unlock_op().  This happens when creating a new encrypted file.

We can't fix this by just not setting up the key for new inodes right
away, since new symlinks need their key to encrypt the symlink target.

So we need to set up the new inode's key before starting the
transaction.  But just calling fscrypt_get_encryption_info() earlier
doesn't work, since it assumes the encryption context is already set,
and the encryption context can't be set until the transaction.

The recently proposed fscrypt support for the ceph filesystem
(https://lkml.kernel.org/linux-fscrypt/20200821182813.52570-1-jlayton@kernel.org/T/#u)
will have this same ordering problem too, since ceph will need to
encrypt new symlinks before setting their encryption context.

Finally, f2fs can deadlock when the filesystem is mounted with
'-o test_dummy_encryption' and a new file is created in an existing
unencrypted directory.  Similarly, this is caused by holding too many
locks when calling fscrypt_get_encryption_info().

To solve all these problems, add new helper functions:

- fscrypt_prepare_new_inode() sets up a new inode's encryption key
  (fscrypt_info), using the parent directory's encryption policy and a
  new random nonce.  It neither reads nor writes the encryption context.

- fscrypt_set_context() persists the encryption context of a new inode,
  using the information from the fscrypt_info already in memory.  This
  replaces fscrypt_inherit_context().

Temporarily keep fscrypt_inherit_context() around until all filesystems
have been converted to use fscrypt_set_context().

Acked-by: Jeff Layton <jlayton@kernel.org>
Link: https://lore.kernel.org/r/20200917041136.178600-2-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>

Normally smp_store_release() or cmpxchg_release() is paired with
smp_load_acquire().  Sometimes smp_load_acquire() can be replaced with
the more lightweight READ_ONCE().  However, for this to be safe, all the
published memory must only be accessed in a way that involves the
pointer itself.  This may not be the case if allocating the object also
involves initializing a static or global variable, for example.

fscrypt_info includes various sub-objects which are internal to and are
allocated by other kernel subsystems such as keyrings and crypto.  So by
using READ_ONCE() for ->i_crypt_info, we're relying on internal
implementation details of these other kernel subsystems.

Remove this fragile assumption by using smp_load_acquire() instead.

(Note: I haven't seen any real-world problems here.  This change is just
fixing the code to be guaranteed correct and less fragile.)

Fixes: e37a784d8b6a ("fscrypt: use READ_ONCE() to access ->i_crypt_info")
Link: https://lore.kernel.org/r/20200721225920.114347-5-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>