[PATCH 0/2] lib/crypto: chacha: Handle lengths of 2 GiB or more on 32-bit arm and MIPS

[PATCH 0/2] lib/crypto: chacha: Handle lengths of 2 GiB or more on 32-bit arm and MIPS

[Vladimir Kukharenko via B4 Relay]

On 32-bit arm (the scalar code) and 32-bit MIPS, the ChaCha assembly
treats the length as signed, so lengths of 2 GiB or more are still
mishandled after "lib/crypto: chacha: Use size_t length", since size_t
is 32 bits there.

Patch 1 makes the arm assembly use unsigned conditions.  The MIPS
assembly relies on the sign of the remaining length, so patch 2 instead
limits each call to 1 GiB from C.

This applies on top of
https://lore.kernel.org/r/20261014152604.323310-1-ebiggers@kernel.org
Feel free to fold these into your series instead.

Tested with the lib/crypto KUnit tests in QEMU on arm (Cortex-A7 and
Cortex-A15) and 32-bit MIPS, plus a temporary 2 GiB test on the
Cortex-A7 and QEMU user mode runs of the MIPS assembly.

An LLM review found these bugs; I used an LLM-based assistant to write
the patches and reviewed its work.

---
Vladimir Kukharenko (2):
      lib/crypto: arm/chacha: Handle lengths of 2 GiB or more in scalar code
      lib/crypto: mips/chacha: Handle lengths of 2 GiB or more

 lib/crypto/arm/chacha-scalar-core.S |  8 ++++----
 lib/crypto/arm/chacha.h             |  2 +-
 lib/crypto/mips/chacha-core.S       |  8 ++++----
 lib/crypto/mips/chacha.h            | 20 +++++++++++++++++++-
 4 files changed, 28 insertions(+), 10 deletions(-)
---
base-commit: 40f47fa10a1d19c74f48a13d97b2eb19d5fb938d
change-id: 20261014-libcrypto-chacha-32bit-2gib-5a3e5e36a061
prerequisite-message-id: <20261014152604.323310-1-ebiggers@kernel.org>
prerequisite-patch-id: f66d78884d3d0cf68e1efc9635c084c6b700352f
prerequisite-patch-id: 64c39ea52fe4e25501701ffa262fb3525af29f0b

Best regards,
--  
Vladimir Kukharenko <vladimir@genproof.com>

[PATCH 1/2] lib/crypto: arm/chacha: Handle lengths of 2 GiB or more in scalar code

[Vladimir Kukharenko via B4 Relay]

From: Vladimir Kukharenko <vladimir@genproof.com>

chacha_doarm() compares the remaining length with signed conditions
(blt, movle/movgt, bgt), and DIV_ROUND_UP(bytes, CHACHA_BLOCK_SIZE) in
chacha_crypt_arch() wraps above 0xffffffc0, so lengths of 2 GiB or
more are mishandled.  size_t is 32 bits on arm, so declaring the length
as size_t does not avoid this.  On a Cortex-A7, an in-place
chacha20_crypt() of 2 GiB + 4101 bytes changed only one byte.

Use unsigned conditions (blo, movls/movhi, bhi) and
DIV_ROUND_UP_POW2().

This affects the scalar code, which is used when NEON isn't usable and
always on Cortex-A5 and Cortex-A7.  No in-tree caller passes such
lengths.  Tested in QEMU on a Cortex-A7 with CONFIG_VMSPLIT_1G.  An LLM
review of this code found the bug.

Assisted-by: LLM
Signed-off-by: Vladimir Kukharenko <vladimir@genproof.com>
---
 lib/crypto/arm/chacha-scalar-core.S | 8 ++++----
 lib/crypto/arm/chacha.h             | 2 +-
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/lib/crypto/arm/chacha-scalar-core.S b/lib/crypto/arm/chacha-scalar-core.S
index 509b2f277b57..d9d0a8cee941 100644
--- a/lib/crypto/arm/chacha-scalar-core.S
+++ b/lib/crypto/arm/chacha-scalar-core.S
@@ -163,7 +163,7 @@
 
 	// Use slow path if fewer than 64 bytes remain.
 	cmp		r11, #64
-	blt		.Lxor_slowpath\@
+	blo		.Lxor_slowpath\@
 
 	// Use slow path if IN and/or OUT isn't 4-byte aligned.  Needed even on
 	// ARMv6+, since ldmia and stmia (used below) still require alignment.
@@ -324,8 +324,8 @@
 	ldr		r14, [sp, #160]		// OUT
 	cmp		r9, #64
 	  mov		r0, sp
-	movle		r1, r9
-	movgt		r1, #64
+	movls		r1, r9
+	movhi		r1, #64
 	// r1 is number of bytes to XOR, in range [1, 64]
 
 .if __LINUX_ARM_ARCH__ < 6
@@ -360,7 +360,7 @@
 .Lxor_slowpath_done\@:
 	subs		r9, #64
 	add		sp, #96
-	bgt		.Lprepare_for_next_block\@
+	bhi		.Lprepare_for_next_block\@
 
 .Ldone\@:
 .endm	// _chacha
diff --git a/lib/crypto/arm/chacha.h b/lib/crypto/arm/chacha.h
index 8ce5375f58aa..5b366f4de78d 100644
--- a/lib/crypto/arm/chacha.h
+++ b/lib/crypto/arm/chacha.h
@@ -78,7 +78,7 @@ static void chacha_crypt_arch(struct chacha_state *state, u8 *dst,
 	if (!IS_ENABLED(CONFIG_KERNEL_MODE_NEON) || !neon_usable() ||
 	    bytes <= CHACHA_BLOCK_SIZE) {
 		chacha_doarm(dst, src, bytes, state, nrounds);
-		state->x[12] += DIV_ROUND_UP(bytes, CHACHA_BLOCK_SIZE);
+		state->x[12] += DIV_ROUND_UP_POW2(bytes, CHACHA_BLOCK_SIZE);
 		return;
 	}
 

-- 
2.43.0

[PATCH 2/2] lib/crypto: mips/chacha: Handle lengths of 2 GiB or more

[Vladimir Kukharenko via B4 Relay]

From: Vladimir Kukharenko <vladimir@genproof.com>

The MIPS ChaCha assembly, which is chacha_crypt_arch() itself, tests
the remaining length with signed branches, so for 2 GiB + 64 bytes or
more it takes the final partial block path too early and processes less
than requested.  It is only used on 32-bit kernels, where size_t is 32
bits, so declaring the length as size_t does not avoid this.  No
in-tree caller passes such lengths.

The partial block code also keeps the remaining length as a negative
value, so instead of changing the assembly, rename it to
chacha_crypt_mips() and add a C chacha_crypt_arch() that passes it at
most 1 GiB per call.

Tested in QEMU user mode: with a length of 0xfffff045 the assembly
processed about 4 KiB and advanced the block counter by 64; with the
same 1 GiB loop in C, a 1.25 GiB + 4101 byte buffer matches a C
reference.  An LLM review of this code found the bug.

Assisted-by: LLM
Signed-off-by: Vladimir Kukharenko <vladimir@genproof.com>
---
 lib/crypto/mips/chacha-core.S |  8 ++++----
 lib/crypto/mips/chacha.h      | 20 +++++++++++++++++++-
 2 files changed, 23 insertions(+), 5 deletions(-)

diff --git a/lib/crypto/mips/chacha-core.S b/lib/crypto/mips/chacha-core.S
index 706aeb850fb0..1a2253d91af7 100644
--- a/lib/crypto/mips/chacha-core.S
+++ b/lib/crypto/mips/chacha-core.S
@@ -196,9 +196,9 @@ CONCAT3(.Lchacha_mips_xor_aligned_, PLUS_ONE(x), _b: ;) \
 .text
 .set	reorder
 .set	noat
-.globl	chacha_crypt_arch
-.ent	chacha_crypt_arch
-chacha_crypt_arch:
+.globl	chacha_crypt_mips
+.ent	chacha_crypt_mips
+chacha_crypt_mips:
 	.frame	$sp, STACK_SIZE, $ra
 
 	/* Load number of rounds */
@@ -414,7 +414,7 @@ chacha_crypt_arch:
 
 	/* Jump table */
 	FOR_EACH_WORD(JMPTBL_UNALIGNED)
-.end chacha_crypt_arch
+.end chacha_crypt_mips
 .set at
 
 /* Input arguments
diff --git a/lib/crypto/mips/chacha.h b/lib/crypto/mips/chacha.h
index 5a565d0a8c26..15b39c744c28 100644
--- a/lib/crypto/mips/chacha.h
+++ b/lib/crypto/mips/chacha.h
@@ -6,8 +6,26 @@
  */
 
 #include <linux/kernel.h>
+#include <linux/sizes.h>
 
-asmlinkage void chacha_crypt_arch(struct chacha_state *state, u8 *dst,
+asmlinkage void chacha_crypt_mips(struct chacha_state *state, u8 *dst,
 				  const u8 *src, size_t bytes, int nrounds);
 asmlinkage void hchacha_block_arch(const struct chacha_state *state,
 				   u32 out[HCHACHA_OUT_WORDS], int nrounds);
+
+static void chacha_crypt_arch(struct chacha_state *state, u8 *dst,
+			      const u8 *src, size_t bytes, int nrounds)
+{
+	/*
+	 * chacha_crypt_mips() compares the length as a signed value, so pass
+	 * it at most 1 GiB at a time.
+	 */
+	do {
+		size_t todo = min(bytes, SZ_1G);
+
+		chacha_crypt_mips(state, dst, src, todo, nrounds);
+		bytes -= todo;
+		src += todo;
+		dst += todo;
+	} while (bytes);
+}

-- 
2.43.0