fscrypt: use smp_load_acquire() for fscrypt_prepared_key

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_prepared_key includes a pointer to a crypto_skcipher object,
which is internal to and is allocated by the crypto subsystem.  By using
READ_ONCE() for it, we're relying on internal implementation details of
the crypto subsystem.

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: 5fee36095c ("fscrypt: add inline encryption support")
Cc: Satya Tangirala <satyat@google.com>
Link: https://lore.kernel.org/r/20200721225920.114347-3-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>

fs/crypto/inline_crypt.c

@@ -176,8 +176,10 @@ int fscrypt_prepare_inline_crypt_key(struct fscrypt_prepared_key *prep_key,
 		}
 	}
 	/*
-	 * Pairs with READ_ONCE() in fscrypt_is_key_prepared().  (Only matters
-	 * for the per-mode keys, which are shared by multiple inodes.)
+	 * Pairs with the smp_load_acquire() in fscrypt_is_key_prepared().
+	 * I.e., here we publish ->blk_key with a RELEASE barrier so that
+	 * concurrent tasks can ACQUIRE it.  Note that this concurrency is only
+	 * possible for per-mode keys, not for per-file keys.
 	 */
 	smp_store_release(&prep_key->blk_key, blk_key);
 	return 0;