[FOR,STABLE,4.1,-v2] ext4 crypto: don't regenerate the per-inode encryption key unnecessarily

From 8912051518ca75a94670652722c72e412eafe674 Mon Sep 17 00:00:00 2001
From: Theodore Ts'o <tytso@mit.edu>
Date: Sun, 8 Oct 2017 13:32:21 -0400
Subject: [PATCH] ext4 crypto: don't regenerate the per-inode encryption key
 unnecessarily

[ Relevant upstream commit: 1b53cf9815bb4744958d41f3795d5d5a1d365e2d ]

This fixes the same problem as upstream commit 1b53cf9815bb: "fscrypt:
remove broken support for detecting keyring key revocation".
Specifically, key revocations racing with readpage operations will
cause the kernel to crash and burn with a BUG_ON or a NULL pointer
dereference in a block I/O callback stemming from an ext4_readpage()
operation.

This fix is needed to fix prevent xfstests test runs from crashing
while running the generic/421 test.

The root cause is different in the 4.1 kernel, however, since the
4.1's encryption handling is so _primitive_ compared to later kernels.
The code isn't actually explicitly checking for revoked keys.
Instead, the code is neededly regenerating the per-file encryption key
on every mmap() or open() or directory operation (in the case of a
directory inode).  Yelch!

If the file is already opened and actively being read, and there is a
racing open() after the user's master key has been revoked, there will
be the same net effect as the problem fixed by upstream commit
1b53cf9815bb --- the per-file key will be marked as invalid and this
will cause a BUG_ON.

In the AOSP 3.18 and 4.4 android-common kernels, the more modern
version of ext4 encryption have been backported, including a backport
of upstream commit 1b53cf9815bb.  This is a dozen plus commits, and
isn't really suitable for the Upstream LTS kernel.  So instead, this
is the simplest bug which fixes the same high-level issue as the
upstream commit, without dragging in all of the other non-bug fixes
improvements to the ext4 encryption code found in newer kernels.

Signed-off-by: Theodore Ts'o <tytso@mit.edu>
---
 fs/ext4/crypto_key.c | 15 ++++++++++++---
 fs/ext4/ext4.h       |  1 +
 fs/ext4/super.c      |  3 +++
 3 files changed, 16 insertions(+), 3 deletions(-)

On Sun, Oct 08, 2017 at 05:55:30PM -0400, Theodore Ts'o wrote:
> diff --git a/fs/ext4/crypto_key.c b/fs/ext4/crypto_key.c
> index 52170d0b7c40..4f9818719d61 100644
> --- a/fs/ext4/crypto_key.c
> +++ b/fs/ext4/crypto_key.c
> @@ -99,9 +99,17 @@ int ext4_generate_encryption_key(struct inode *inode)
>  	struct ext4_encryption_context ctx;
>  	struct user_key_payload *ukp;
>  	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
> -	int res = ext4_xattr_get(inode, EXT4_XATTR_INDEX_ENCRYPTION,
> -				 EXT4_XATTR_NAME_ENCRYPTION_CONTEXT,
> -				 &ctx, sizeof(ctx));
> +	int res;
> +
> +	mutex_lock(&ei->i_encryption_lock);
> +	if (ext4_has_encryption_key(inode)) {
> +		mutex_unlock(&ei->i_encryption_lock);
> +		return 0;
> +	}
> +
> +	res = ext4_xattr_get(inode, EXT4_XATTR_INDEX_ENCRYPTION,
> +			     EXT4_XATTR_NAME_ENCRYPTION_CONTEXT,
> +			     &ctx, sizeof(ctx));
>  
>  	if (res != sizeof(ctx)) {
>  		if (res > 0)
> @@ -154,6 +162,7 @@ out:
>  		key_put(keyring_key);
>  	if (res < 0)
>  		crypt_key->mode = EXT4_ENCRYPTION_MODE_INVALID;
> +	mutex_unlock(&ei->i_encryption_lock);
>  	return res;

This looks better, but maybe the call to ext4_has_encryption_key() should also
be removed from ext4_get_fname_crypto_ctx() --- so it calls
ext4_generate_encryption_key() unconditionally, like ext4_file_open() and
ext4_file_mmap() do?  Otherwise it will still be doing a racy check without the
lock.

Eric