On Thu, 27 Jan 2022 at 20:26, Eric Biggers [off-list ref] wrote:

On Thu, Jan 27, 2022 at 10:42:49AM +0100, Ard Biesheuvel wrote:

quoted

quoted

diff --git a/include/crypto/xctr.h b/include/crypto/xctr.h
new file mode 100644
index 000000000000..0d025e08ca26
--- /dev/null
+++ b/include/crypto/xctr.h

@@ -0,0 +1,19 @@
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+/*
+ * XCTR: XOR Counter mode
+ *
+ * Copyright 2021 Google LLC
+ */
+
+#include <asm/unaligned.h>
+
+#ifndef _CRYPTO_XCTR_H
+#define _CRYPTO_XCTR_H
+
+static inline void u32_to_le_block(u8 *a, u32 x, unsigned int size)
+{
+       memset(a, 0, size);
+       put_unaligned(cpu_to_le32(x), (u32 *)a);

Please use put_unaligned_le32() here.

And casting 'a' to (u32 *) is invalid C, so just pass 'a' directly.
Otherwise, the compiler might infer that 'a' is guaranteed to be
aligned after all, and use an aligned access instead.

I agree that put_unaligned_le32() is more suitable here, but I don't think
casting 'a' to 'u32 *' is undefined; it's only dereferencing it that would be
undefined.  If such casts were undefined, then get_unaligned() and
put_unaligned() would be unusable under any circumstance.  Here's an example of
code that would be incorrect in that case:
https://lore.kernel.org/linux-crypto/20220119093109.1567314-1-ardb@kernel.org (local)

Good point :-)