diff mbox series

[v4,7/8] crypto: arm64/polyval: Add PMULL accelerated implementation of POLYVAL

Message ID 20220412172816.917723-8-nhuck@google.com
State New
Headers show
Series crypto: HCTR2 support | expand

Commit Message

Nathan Huckleberry April 12, 2022, 5:28 p.m. UTC
Add hardware accelerated version of POLYVAL for ARM64 CPUs with
Crypto Extension support.

This implementation is accelerated using PMULL instructions to perform
the finite field computations.  For added efficiency, 8 blocks of the
message are processed simultaneously by precomputing the first 8
powers of the key.

Karatsuba multiplication is used instead of Schoolbook multiplication
because it was found to be slightly faster on ARM64 CPUs.  Montgomery
reduction must be used instead of Barrett reduction due to the
difference in modulus between POLYVAL's field and other finite fields.

Signed-off-by: Nathan Huckleberry <nhuck@google.com>
---
 arch/arm64/crypto/Kconfig           |   6 +
 arch/arm64/crypto/Makefile          |   3 +
 arch/arm64/crypto/polyval-ce-core.S | 366 ++++++++++++++++++++++++++++
 arch/arm64/crypto/polyval-ce-glue.c | 240 ++++++++++++++++++
 4 files changed, 615 insertions(+)
 create mode 100644 arch/arm64/crypto/polyval-ce-core.S
 create mode 100644 arch/arm64/crypto/polyval-ce-glue.c

Comments

Eric Biggers April 13, 2022, 5:53 a.m. UTC | #1
Many of the comments I made on the x86 version apply to this too, but a few
unique comments below:

On Tue, Apr 12, 2022 at 05:28:15PM +0000, Nathan Huckleberry wrote:
> Add hardware accelerated version of POLYVAL for ARM64 CPUs with
> Crypto Extension support.

As I've mentioned, it seems to be "Crypto Extensions", not "Crypto Extension".
That's what everything else in the kernel says, at least.  It's just a nit, but
it's unclear whether you intentionally decided to not change this, or missed it.

> diff --git a/arch/arm64/crypto/polyval-ce-core.S b/arch/arm64/crypto/polyval-ce-core.S
[..]
> +	.text
> +	.align	4
> +
> +	.arch	armv8-a+crypto
> +	.align	4

Having '.align' twice is redundant, right?

> +
> +.Lgstar:
> +	.quad	0xc200000000000000, 0xc200000000000000
> +
> +/*
> + * Computes the product of two 128-bit polynomials in X and Y and XORs the
> + * components of the 256-bit product into LO, MI, HI.
> + *
> + * X = [X_1 : X_0]
> + * Y = [Y_1 : Y_0]
> + *
> + * The multiplication produces four parts:
> + *   LOW: The polynomial given by performing carryless multiplication of X_0 and
> + *   Y_0
> + *   MID: The polynomial given by performing carryless multiplication of (X_0 +
> + *   X_1) and (Y_0 + Y_1)
> + *   HIGH: The polynomial given by performing carryless multiplication of X_1
> + *   and Y_1
> + *
> + * We compute:
> + *  LO += LOW
> + *  MI += MID
> + *  HI += HIGH
> + *
> + * Later, the 256-bit result can be extracted as:
> + *   [HI_1 : HI_0 + HI_1 + MI_1 + LO_1 :: LO_1 + HI_0 + MI_0 + LO_0 : LO_0]

What does the double colon mean?

> + * This step is done when computing the polynomial reduction for efficiency
> + * reasons.
> + */
> +.macro karatsuba1 X Y

A super brief explanation of why Karatsuba multiplication is used instead of
schoolbook multiplication would be helpful.

> +	X .req \X
> +	Y .req \Y
> +	ext	v25.16b, X.16b, X.16b, #8
> +	ext	v26.16b, Y.16b, Y.16b, #8
> +	eor	v25.16b, v25.16b, X.16b
> +	eor	v26.16b, v26.16b, Y.16b
> +	pmull	v27.1q, v25.1d, v26.1d
> +	pmull2	v28.1q, X.2d, Y.2d
> +	pmull	v29.1q, X.1d, Y.1d
> +	eor	MI.16b, MI.16b, v27.16b
> +	eor	HI.16b, HI.16b, v28.16b
> +	eor	LO.16b, LO.16b, v29.16b
> +	.unreq X
> +	.unreq Y
> +.endm

Maybe move the pmull into v27 and the eor into MI down a bit, since they have
more dependencies?  I.e.

	ext	v25.16b, X.16b, X.16b, #8
	ext	v26.16b, Y.16b, Y.16b, #8
	eor	v25.16b, v25.16b, X.16b
	eor	v26.16b, v26.16b, Y.16b
	pmull2	v28.1q, X.2d, Y.2d
	pmull	v29.1q, X.1d, Y.1d
	pmull	v27.1q, v25.1d, v26.1d
	eor	HI.16b, HI.16b, v28.16b
	eor	LO.16b, LO.16b, v29.16b
	eor	MI.16b, MI.16b, v27.16b

I don't know whether it will actually matter on arm64, but on lower-powered
arm32 CPUs this sort of thing definitely can matter.

> +.macro karatsuba1_store X Y
> +	X .req \X
> +	Y .req \Y
> +	ext	v25.16b, X.16b, X.16b, #8
> +	ext	v26.16b, Y.16b, Y.16b, #8
> +	eor	v25.16b, v25.16b, X.16b
> +	eor	v26.16b, v26.16b, Y.16b
> +	pmull	MI.1q, v25.1d, v26.1d
> +	pmull2	HI.1q, X.2d, Y.2d
> +	pmull	LO.1q, X.1d, Y.1d
> +	.unreq X
> +	.unreq Y
> +.endm

Likewise above.

> diff --git a/arch/arm64/crypto/polyval-ce-glue.c b/arch/arm64/crypto/polyval-ce-glue.c
[...]
> +struct polyval_async_ctx {
> +	struct cryptd_ahash *cryptd_tfm;
> +};

struct polyval_async_ctx is no longer used.

> +static int polyval_setkey(struct crypto_shash *tfm,
> +			const u8 *key, unsigned int keylen)
> +{
> +	struct polyval_ctx *ctx = crypto_shash_ctx(tfm);
> +	int i;
> +
> +	if (keylen != POLYVAL_BLOCK_SIZE)
> +		return -EINVAL;
> +
> +	BUILD_BUG_ON(sizeof(u128) != POLYVAL_BLOCK_SIZE);

Where does 'sizeof(u128)' come from?  This file doesn't use u128 at all.

- Eric
diff mbox series

Patch

diff --git a/arch/arm64/crypto/Kconfig b/arch/arm64/crypto/Kconfig
index 897f9a4b5b67..6457fd7ec7af 100644
--- a/arch/arm64/crypto/Kconfig
+++ b/arch/arm64/crypto/Kconfig
@@ -60,6 +60,12 @@  config CRYPTO_GHASH_ARM64_CE
 	select CRYPTO_GF128MUL
 	select CRYPTO_LIB_AES
 
+config CRYPTO_POLYVAL_ARM64_CE
+	tristate "POLYVAL using ARMv8 Crypto Extensions (for HCTR2)"
+	depends on KERNEL_MODE_NEON
+	select CRYPTO_CRYPTD
+	select CRYPTO_POLYVAL
+
 config CRYPTO_CRCT10DIF_ARM64_CE
 	tristate "CRCT10DIF digest algorithm using PMULL instructions"
 	depends on KERNEL_MODE_NEON && CRC_T10DIF
diff --git a/arch/arm64/crypto/Makefile b/arch/arm64/crypto/Makefile
index 09a805cc32d7..53f9af962b86 100644
--- a/arch/arm64/crypto/Makefile
+++ b/arch/arm64/crypto/Makefile
@@ -26,6 +26,9 @@  sm4-ce-y := sm4-ce-glue.o sm4-ce-core.o
 obj-$(CONFIG_CRYPTO_GHASH_ARM64_CE) += ghash-ce.o
 ghash-ce-y := ghash-ce-glue.o ghash-ce-core.o
 
+obj-$(CONFIG_CRYPTO_POLYVAL_ARM64_CE) += polyval-ce.o
+polyval-ce-y := polyval-ce-glue.o polyval-ce-core.o
+
 obj-$(CONFIG_CRYPTO_CRCT10DIF_ARM64_CE) += crct10dif-ce.o
 crct10dif-ce-y := crct10dif-ce-core.o crct10dif-ce-glue.o
 
diff --git a/arch/arm64/crypto/polyval-ce-core.S b/arch/arm64/crypto/polyval-ce-core.S
new file mode 100644
index 000000000000..c6882ddf839f
--- /dev/null
+++ b/arch/arm64/crypto/polyval-ce-core.S
@@ -0,0 +1,366 @@ 
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Implementation of POLYVAL using ARMv8 Crypto Extensions.
+ *
+ * Copyright 2021 Google LLC
+ */
+/*
+ * This is an efficient implementation of POLYVAL using ARMv8 Crypto Extensions
+ * It works on 8 blocks at a time, by precomputing the first 8 keys powers h^8,
+ * ..., h^1 in the POLYVAL finite field. This precomputation allows us to split
+ * finite field multiplication into two steps.
+ *
+ * In the first step, we consider h^i, m_i as normal polynomials of degree less
+ * than 128. We then compute p(x) = h^8m_0 + ... + h^1m_7 where multiplication
+ * is simply polynomial multiplication.
+ *
+ * In the second step, we compute the reduction of p(x) modulo the finite field
+ * modulus g(x) = x^128 + x^127 + x^126 + x^121 + 1.
+ *
+ * This two step process is equivalent to computing h^8m_0 + ... + h^1m_7 where
+ * multiplication is finite field multiplication. The advantage is that the
+ * two-step process  only requires 1 finite field reduction for every 8
+ * polynomial multiplications. Further parallelism is gained by interleaving the
+ * multiplications and polynomial reductions.
+ */
+
+#include <linux/linkage.h>
+#define STRIDE_BLOCKS 8
+
+BLOCKS_LEFT	.req	x2
+KEY_START	.req	x10
+EXTRA_BYTES	.req	x11
+TMP	.req	x13
+
+M0	.req	v0
+M1	.req	v1
+M2	.req	v2
+M3	.req	v3
+M4	.req	v4
+M5	.req	v5
+M6	.req	v6
+M7	.req	v7
+KEY8	.req	v8
+KEY7	.req	v9
+KEY6	.req	v10
+KEY5	.req	v11
+KEY4	.req	v12
+KEY3	.req	v13
+KEY2	.req	v14
+KEY1	.req	v15
+PL	.req	v16
+PH	.req	v17
+TMP_V	.req	v18
+LO	.req	v20
+MI	.req	v21
+HI	.req	v22
+SUM	.req	v23
+GSTAR	.req	v24
+
+	.text
+	.align	4
+
+	.arch	armv8-a+crypto
+	.align	4
+
+.Lgstar:
+	.quad	0xc200000000000000, 0xc200000000000000
+
+/*
+ * Computes the product of two 128-bit polynomials in X and Y and XORs the
+ * components of the 256-bit product into LO, MI, HI.
+ *
+ * X = [X_1 : X_0]
+ * Y = [Y_1 : Y_0]
+ *
+ * The multiplication produces four parts:
+ *   LOW: The polynomial given by performing carryless multiplication of X_0 and
+ *   Y_0
+ *   MID: The polynomial given by performing carryless multiplication of (X_0 +
+ *   X_1) and (Y_0 + Y_1)
+ *   HIGH: The polynomial given by performing carryless multiplication of X_1
+ *   and Y_1
+ *
+ * We compute:
+ *  LO += LOW
+ *  MI += MID
+ *  HI += HIGH
+ *
+ * Later, the 256-bit result can be extracted as:
+ *   [HI_1 : HI_0 + HI_1 + MI_1 + LO_1 :: LO_1 + HI_0 + MI_0 + LO_0 : LO_0]
+ * This step is done when computing the polynomial reduction for efficiency
+ * reasons.
+ */
+.macro karatsuba1 X Y
+	X .req \X
+	Y .req \Y
+	ext	v25.16b, X.16b, X.16b, #8
+	ext	v26.16b, Y.16b, Y.16b, #8
+	eor	v25.16b, v25.16b, X.16b
+	eor	v26.16b, v26.16b, Y.16b
+	pmull	v27.1q, v25.1d, v26.1d
+	pmull2	v28.1q, X.2d, Y.2d
+	pmull	v29.1q, X.1d, Y.1d
+	eor	MI.16b, MI.16b, v27.16b
+	eor	HI.16b, HI.16b, v28.16b
+	eor	LO.16b, LO.16b, v29.16b
+	.unreq X
+	.unreq Y
+.endm
+
+/*
+ * Same as karatsuba1, except overwrites HI, LO, MI rather than XORing into
+ * them.
+ */
+.macro karatsuba1_store X Y
+	X .req \X
+	Y .req \Y
+	ext	v25.16b, X.16b, X.16b, #8
+	ext	v26.16b, Y.16b, Y.16b, #8
+	eor	v25.16b, v25.16b, X.16b
+	eor	v26.16b, v26.16b, Y.16b
+	pmull	MI.1q, v25.1d, v26.1d
+	pmull2	HI.1q, X.2d, Y.2d
+	pmull	LO.1q, X.1d, Y.1d
+	.unreq X
+	.unreq Y
+.endm
+
+/*
+ * Computes the 256-bit polynomial represented by LO, HI, MI. Stores
+ * the result in PL, PH.
+ * [PH : PL] =
+ *   [HI_1 : HI_0 + HI_1 + MI_1 + LO_1 : LO_1 + LO_0 + MI_0 + HI_0 : LO_0]
+ */
+.macro karatsuba2
+	// v4 = [HI_1 + MI_1 : HI_0 + MI_0]
+	eor	v4.16b, HI.16b, MI.16b
+	// v4 = [HI_1 + MI_1 + LO_1 : HI_0 + MI_0 + LO_0]
+	eor	v4.16b, v4.16b, LO.16b
+	// v5 = [HI_0 : LO_1]
+	ext	v5.16b, LO.16b, HI.16b, #8
+	// v4 = [HI_0 + HI_1 + MI_1 + LO_1 : LO_1 + LO_0 + MI_0 + HI_0]
+	eor	v4.16b, v4.16b, v5.16b
+	// HI = [HI_0 : HI_1]
+	ext	HI.16b, HI.16b, HI.16b, #8
+	// LO = [LO_0 : LO_1]
+	ext	LO.16b, LO.16b, LO.16b, #8
+	// PH = [HI_1 : HI_0 + HI_1 + MI_1 + LO_1]
+	ext	PH.16b, v4.16b, HI.16b, #8
+	// PL = [LO_1 + LO_0 + MI_0 + HI_0 : LO_0]
+	ext	PL.16b, LO.16b, v4.16b, #8
+.endm
+
+/*
+ * Computes the 128-bit reduction of PL, PH. Stores the result in dest.
+ *
+ * This macro computes p(x) mod g(x) where p(x) is in montgomery form and g(x) =
+ * x^128 + x^127 + x^126 + x^121 + 1.
+ *
+ * We have a 256-bit polynomial P_H : P_L = P_3 : P_2 : P_1 : P_0 that is the
+ * product of two 128-bit polynomials in Montgomery form.  We need to reduce it
+ * mod g(x).  Also, since polynomials in Montgomery form have an "extra" factor
+ * of x^128, this product has two extra factors of x^128.  To get it back into
+ * Montgomery form, we need to remove one of these factors by dividing by x^128.
+ *
+ * To accomplish both of these goals, we add multiples of g(x) that cancel out
+ * the low 128 bits P_1 : P_0, leaving just the high 128 bits. Since the low
+ * bits are zero, the polynomial division by x^128 can be done by right shifting.
+ *
+ * Since the only nonzero term in the low 64 bits of g(x) is the constant term,
+ * the multiple of g(x) needed to cancel out P_0 is P_0 * g(x).  The CPU can
+ * only do 64x64 bit multiplications, so split P_0 * g(x) into x^128 * P_0 +
+ * x^64 * g*(x) * P_0 + P_0, where g*(x) is bits 64-127 of g(x).  Adding this to
+ * the original polynomial gives P_3 : P_2 + P_0 + T_1 : P_1 + T_0 : 0, where T
+ * = T_1 : T_0 = g*(x) * P_0.  Thus, bits 0-63 got "folded" into bits 64-191.
+ *
+ * Repeating this same process on the next 64 bits "folds" bits 64-127 into bits
+ * 128-255, giving the answer in bits 128-255. This time, we need to cancel P_1
+ * + T_0 in bits 64-127. The multiple of g(x) required is (P_1 + T_0) * g(x) *
+ * x^64. Adding this to our previous computation gives P_3 + P_1 + T_0 + V_1 :
+ * P_2 + P_0 + T_1 + V_0 : 0 : 0, where V = V_1 : V_0 = g*(x) * (P_1 + T_0).
+ *
+ * So our final computation is:
+ *   T = T_1 : T_0 = g*(x) * P_0
+ *   V = V_1 : V_0 = g*(x) * (T_0 + P_1)
+ *   p(x) / x^{128} mod g(x) = P_3 + P_1 + T_0 + V_1 : P_2 + P_0 + T_1 + V_0
+ *
+ * The implementation below saves a XOR instruction by computing P_1 + T_0 : P_0
+ * + T_1 and XORing it into dest, rather than directly XORing P_1 : P_0, T_0 :
+ * T1 into dest.  This allows us to reuse P_1 + T_0 when computing V.
+ */
+.macro montgomery_reduction dest
+	DEST .req \dest
+	// TMP_V = T_1 : T_0 = P_0 * g*(x)
+	pmull	TMP_V.1q, GSTAR.1d, PL.1d
+	// TMP_V = T_0 : T_1
+	ext	TMP_V.16b, TMP_V.16b, TMP_V.16b, #8
+	// TMP_V = P_1 + T_0 : P_0 + T_1
+	eor	TMP_V.16b, PL.16b, TMP_V.16b
+	// PH = P_3 + P_1 + T_0 : P_2 + P_0 + T_1
+	eor	PH.16b, PH.16b, TMP_V.16b
+	// TMP_V = V_1 : V_0 = (P_1 + T_0) * g*(x)
+	pmull2	TMP_V.1q, GSTAR.2d, TMP_V.2d
+	eor	DEST.16b, TMP_V.16b, PH.16b
+	.unreq DEST
+.endm
+
+/*
+ * Compute Polyval on 8 blocks.
+ *
+ * If reduce is set, also computes the montgomery reduction of the
+ * previous full_stride call and XORs with the first message block.
+ * (m_0 + REDUCE(PL, PH))h^8 + ... + m_7h^1.
+ * I.e., the first multiplication uses m_0 + REDUCE(PL, PH) instead of m_0.
+ *
+ * Sets PL, PH.
+ */
+.macro full_stride reduce
+	eor		LO.16b, LO.16b, LO.16b
+	eor		MI.16b, MI.16b, MI.16b
+	eor		HI.16b, HI.16b, HI.16b
+
+	ld1		{M0.16b, M1.16b, M2.16b, M3.16b}, [x0], #64
+	ld1		{M4.16b, M5.16b, M6.16b, M7.16b}, [x0], #64
+
+	karatsuba1 M7 KEY1
+	.if (\reduce)
+	pmull	TMP_V.1q, GSTAR.1d, PL.1d
+	.endif
+
+	karatsuba1 M6 KEY2
+	.if (\reduce)
+	ext	TMP_V.16b, TMP_V.16b, TMP_V.16b, #8
+	.endif
+
+	karatsuba1 M5 KEY3
+	.if (\reduce)
+	eor	TMP_V.16b, PL.16b, TMP_V.16b
+	.endif
+
+	karatsuba1 M4 KEY4
+	.if (\reduce)
+	eor	PH.16b, PH.16b, TMP_V.16b
+	.endif
+
+	karatsuba1 M3 KEY5
+	.if (\reduce)
+	pmull2	TMP_V.1q, GSTAR.2d, TMP_V.2d
+	.endif
+
+	karatsuba1 M2 KEY6
+	.if (\reduce)
+	eor	SUM.16b, TMP_V.16b, PH.16b
+	.endif
+
+	karatsuba1 M1 KEY7
+	eor	M0.16b, M0.16b, SUM.16b
+
+	karatsuba1 M0 KEY8
+	karatsuba2
+.endm
+
+/*
+ * Handle any extra blocks before
+ * full_stride loop.
+ */
+.macro partial_stride
+	add	x1, KEY_START, #(STRIDE_BLOCKS << 4)
+	sub	x1, x1, BLOCKS_LEFT, lsl #4
+	// Clobber key register
+	ld1	{KEY1.16b}, [x1], #16
+
+	ld1	{TMP_V.16b}, [x0], #16
+	eor	SUM.16b, SUM.16b, TMP_V.16b
+	karatsuba1_store KEY1 SUM
+	sub	BLOCKS_LEFT, BLOCKS_LEFT, #1
+
+	tst	BLOCKS_LEFT, #4
+	beq	.Lpartial4BlocksDone
+	ld1	{M0.16b, M1.16b,  M2.16b, M3.16b}, [x0], #64
+	// Clobber key registers
+	ld1	{KEY8.16b, KEY7.16b, KEY6.16b,  KEY5.16b}, [x1], #64
+	karatsuba1 M0 KEY8
+	karatsuba1 M1 KEY7
+	karatsuba1 M2 KEY6
+	karatsuba1 M3 KEY5
+.Lpartial4BlocksDone:
+	tst	BLOCKS_LEFT, #2
+	beq	.Lpartial2BlocksDone
+	ld1	{M0.16b, M1.16b}, [x0], #32
+	// Clobber key registers
+	ld1	{KEY8.16b, KEY7.16b}, [x1], #32
+	karatsuba1 M0 KEY8
+	karatsuba1 M1 KEY7
+.Lpartial2BlocksDone:
+	tst	BLOCKS_LEFT, #1
+	beq	.LpartialDone
+	ld1	{M0.16b}, [x0], #16
+	// Clobber key registers
+	ld1	{KEY8.16b}, [x1], #16
+	karatsuba1 M0 KEY8
+.LpartialDone:
+	karatsuba2
+	montgomery_reduction SUM
+.endm
+
+/*
+ * Perform montgomery multiplication in GF(2^128) and store result in op1.
+ *
+ * Computes op1*op2*x^{-128} mod x^128 + x^127 + x^126 + x^121 + 1
+ * If op1, op2 are in montgomery form, this computes the montgomery
+ * form of op1*op2.
+ *
+ * void pmull_polyval_mul(u8 *op1, const u8 *op2);
+ */
+SYM_FUNC_START(pmull_polyval_mul)
+	adr	TMP, .Lgstar
+	ld1	{GSTAR.2d}, [TMP]
+	ld1	{v0.16b}, [x0]
+	ld1	{v1.16b}, [x1]
+	karatsuba1_store v0 v1
+	karatsuba2
+	montgomery_reduction SUM
+	st1	{SUM.16b}, [x0]
+	ret
+SYM_FUNC_END(pmull_polyval_mul)
+
+/*
+ * Perform polynomial evaluation as specified by POLYVAL.  This computes:
+ * 	h^n * accumulator + h^n * m_0 + ... + h^1 * m_{n-1}
+ * where n=nblocks, h is the hash key, and m_i are the message blocks.
+ *
+ * x0 - pointer to message blocks
+ * x1 - pointer to precomputed key powers h^8 ... h^1
+ * x2 - number of blocks to hash
+ * x3 - pointer to accumulator
+ *
+ * void pmull_polyval_update(const u8 *in, const struct polyval_ctx *ctx,
+ *			     size_t nblocks, u8 *accumulator);
+ */
+SYM_FUNC_START(pmull_polyval_update)
+	adr	TMP, .Lgstar
+	mov	KEY_START, x1
+	ld1	{GSTAR.2d}, [TMP]
+	ld1	{SUM.16b}, [x3]
+	subs	BLOCKS_LEFT, BLOCKS_LEFT, #STRIDE_BLOCKS
+	blt .LstrideLoopExit
+	ld1	{KEY8.16b, KEY7.16b, KEY6.16b, KEY5.16b}, [x1], #64
+	ld1	{KEY4.16b, KEY3.16b, KEY2.16b, KEY1.16b}, [x1], #64
+	full_stride 0
+	subs	BLOCKS_LEFT, BLOCKS_LEFT, #STRIDE_BLOCKS
+	blt .LstrideLoopExitReduce
+.LstrideLoop:
+	full_stride 1
+	subs	BLOCKS_LEFT, BLOCKS_LEFT, #STRIDE_BLOCKS
+	bge	.LstrideLoop
+.LstrideLoopExitReduce:
+	montgomery_reduction SUM
+.LstrideLoopExit:
+	adds	BLOCKS_LEFT, BLOCKS_LEFT, #STRIDE_BLOCKS
+	beq	.LskipPartial
+	partial_stride
+.LskipPartial:
+	st1	{SUM.16b}, [x3]
+	ret
+SYM_FUNC_END(pmull_polyval_update)
diff --git a/arch/arm64/crypto/polyval-ce-glue.c b/arch/arm64/crypto/polyval-ce-glue.c
new file mode 100644
index 000000000000..7c4633f034db
--- /dev/null
+++ b/arch/arm64/crypto/polyval-ce-glue.c
@@ -0,0 +1,240 @@ 
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Accelerated POLYVAL implementation with ARMv8 Crypto Extension
+ * instructions. This file contains glue code.
+ *
+ * Copyright (c) 2007 Nokia Siemens Networks - Mikko Herranen <mh1@iki.fi>
+ * Copyright (c) 2009 Intel Corp.
+ *   Author: Huang Ying <ying.huang@intel.com>
+ * Copyright 2021 Google LLC
+ */
+/*
+ * Glue code based on ghash-clmulni-intel_glue.c.
+ *
+ * This implementation of POLYVAL uses montgomery multiplication accelerated by
+ * ARMv8 Crypto Extension instructions to implement the finite field operations.
+ *
+ */
+
+#include <crypto/algapi.h>
+#include <crypto/cryptd.h>
+#include <crypto/gf128mul.h>
+#include <crypto/internal/hash.h>
+#include <crypto/internal/simd.h>
+#include <crypto/polyval.h>
+#include <linux/crypto.h>
+#include <linux/init.h>
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/cpufeature.h>
+#include <asm/neon.h>
+#include <asm/simd.h>
+#include <asm/unaligned.h>
+
+#define NUM_PRECOMPUTE_POWERS	8
+
+struct polyval_async_ctx {
+	struct cryptd_ahash *cryptd_tfm;
+};
+
+struct polyval_ctx {
+	u8 key_powers[NUM_PRECOMPUTE_POWERS][POLYVAL_BLOCK_SIZE];
+};
+
+struct polyval_desc_ctx {
+	u8 buffer[POLYVAL_BLOCK_SIZE];
+	u32 bytes;
+};
+
+asmlinkage void pmull_polyval_update(const u8 *in, const struct polyval_ctx
+				     *ctx, size_t nblocks, u8 *accumulator);
+asmlinkage void pmull_polyval_mul(u8 *op1, const u8 *op2);
+
+static void reverse_be128(be128 *x)
+{
+	__be64 a = x->a;
+	__be64 b = x->b;
+
+	x->a = swab64(b);
+	x->b = swab64(a);
+}
+
+static void generic_polyval_mul(u8 *op1, const u8 *op2)
+{
+	be128 a, b;
+
+	// Assume one argument is in Montgomery form and one is not.
+	memcpy(&a, op1, sizeof(a));
+	memcpy(&b, op2, sizeof(b));
+	reverse_be128(&a);
+	reverse_be128(&b);
+	gf128mul_x_lle(&a, &a);
+	gf128mul_lle(&a, &b);
+	reverse_be128(&a);
+	memcpy(op1, &a, sizeof(a));
+}
+
+static void generic_polyval_update(const u8 *in, struct polyval_ctx *keys,
+			  size_t nblocks, u8 *accumulator)
+{
+	while (nblocks--) {
+		crypto_xor(accumulator, in, POLYVAL_BLOCK_SIZE);
+		generic_polyval_mul(accumulator, keys->key_powers[7]);
+		in += POLYVAL_BLOCK_SIZE;
+	}
+}
+
+static void internal_polyval_update(const u8 *in, struct polyval_ctx *keys,
+			  size_t nblocks, u8 *accumulator)
+{
+	if (likely(crypto_simd_usable())) {
+		kernel_neon_begin();
+		pmull_polyval_update(in, keys, nblocks, accumulator);
+		kernel_neon_end();
+	} else {
+		generic_polyval_update(in, keys, nblocks, accumulator);
+	}
+}
+
+static void internal_polyval_mul(u8 *op1, const u8 *op2)
+{
+	if (likely(crypto_simd_usable())) {
+		kernel_neon_begin();
+		pmull_polyval_mul(op1, op2);
+		kernel_neon_end();
+	} else {
+		generic_polyval_mul(op1, op2);
+	}
+}
+
+static int polyval_init(struct shash_desc *desc)
+{
+	struct polyval_desc_ctx *dctx = shash_desc_ctx(desc);
+
+	memset(dctx, 0, sizeof(*dctx));
+
+	return 0;
+}
+
+static int polyval_setkey(struct crypto_shash *tfm,
+			const u8 *key, unsigned int keylen)
+{
+	struct polyval_ctx *ctx = crypto_shash_ctx(tfm);
+	int i;
+
+	if (keylen != POLYVAL_BLOCK_SIZE)
+		return -EINVAL;
+
+	BUILD_BUG_ON(sizeof(u128) != POLYVAL_BLOCK_SIZE);
+
+	memcpy(ctx->key_powers[NUM_PRECOMPUTE_POWERS-1], key,
+	       POLYVAL_BLOCK_SIZE);
+
+	for (i = NUM_PRECOMPUTE_POWERS-2; i >= 0; i--) {
+		memcpy(ctx->key_powers[i], key, POLYVAL_BLOCK_SIZE);
+		internal_polyval_mul(ctx->key_powers[i], ctx->key_powers[i+1]);
+	}
+
+	return 0;
+}
+
+static int polyval_update(struct shash_desc *desc,
+			 const u8 *src, unsigned int srclen)
+{
+	struct polyval_desc_ctx *dctx = shash_desc_ctx(desc);
+	struct polyval_ctx *ctx = crypto_shash_ctx(desc->tfm);
+	u8 *pos;
+	unsigned int nblocks;
+	unsigned int n;
+
+	if (dctx->bytes) {
+		n = min(srclen, dctx->bytes);
+		pos = dctx->buffer + POLYVAL_BLOCK_SIZE - dctx->bytes;
+
+		dctx->bytes -= n;
+		srclen -= n;
+
+		while (n--)
+			*pos++ ^= *src++;
+
+		if (!dctx->bytes)
+			internal_polyval_mul(dctx->buffer,
+				ctx->key_powers[NUM_PRECOMPUTE_POWERS-1]);
+	}
+
+	nblocks = srclen/POLYVAL_BLOCK_SIZE;
+	internal_polyval_update(src, ctx, nblocks, dctx->buffer);
+	srclen -= nblocks*POLYVAL_BLOCK_SIZE;
+
+	if (srclen) {
+		dctx->bytes = POLYVAL_BLOCK_SIZE - srclen;
+		src += nblocks*POLYVAL_BLOCK_SIZE;
+		pos = dctx->buffer;
+		while (srclen--)
+			*pos++ ^= *src++;
+	}
+
+	return 0;
+}
+
+static int polyval_final(struct shash_desc *desc, u8 *dst)
+{
+	struct polyval_desc_ctx *dctx = shash_desc_ctx(desc);
+	struct polyval_ctx *ctx = crypto_shash_ctx(desc->tfm);
+
+	if (dctx->bytes) {
+		internal_polyval_mul(dctx->buffer,
+			ctx->key_powers[NUM_PRECOMPUTE_POWERS-1]);
+	}
+
+	dctx->bytes = 0;
+	memcpy(dst, dctx->buffer, POLYVAL_BLOCK_SIZE);
+
+	return 0;
+}
+
+static struct shash_alg polyval_alg = {
+	.digestsize	= POLYVAL_DIGEST_SIZE,
+	.init		= polyval_init,
+	.update		= polyval_update,
+	.final		= polyval_final,
+	.setkey		= polyval_setkey,
+	.descsize	= sizeof(struct polyval_desc_ctx),
+	.base		= {
+		.cra_name		= "polyval",
+		.cra_driver_name	= "polyval-ce",
+		.cra_priority		= 200,
+		.cra_blocksize		= POLYVAL_BLOCK_SIZE,
+		.cra_ctxsize		= sizeof(struct polyval_ctx),
+		.cra_module		= THIS_MODULE,
+	},
+};
+
+static int __init polyval_ce_mod_init(void)
+{
+	if (!cpu_have_named_feature(ASIMD))
+		return -ENODEV;
+
+	if (!cpu_have_named_feature(PMULL))
+		return -ENODEV;
+
+	return crypto_register_shash(&polyval_alg);
+}
+
+static void __exit polyval_ce_mod_exit(void)
+{
+	crypto_unregister_shash(&polyval_alg);
+}
+
+static const struct cpu_feature polyval_cpu_feature[] = {
+	{ cpu_feature(PMULL) }, { }
+};
+MODULE_DEVICE_TABLE(cpu, polyval_cpu_feature);
+
+module_init(polyval_ce_mod_init);
+module_exit(polyval_ce_mod_exit);
+
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("POLYVAL hash function accelerated by ARMv8 Crypto Extension");
+MODULE_ALIAS_CRYPTO("polyval");
+MODULE_ALIAS_CRYPTO("polyval-ce");