linux/arch/x86/crypto/polyval-clmulni_asm.S
Nathan Huckleberry 34f7f6c301 crypto: x86/polyval - Add PCLMULQDQ accelerated implementation of POLYVAL
Add hardware accelerated version of POLYVAL for x86-64 CPUs with
PCLMULQDQ support.

This implementation is accelerated using PCLMULQDQ 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.

Schoolbook multiplication is used instead of Karatsuba multiplication
because it was found to be slightly faster on x86-64 machines.
Montgomery reduction must be used instead of Barrett reduction due to
the difference in modulus between POLYVAL's field and other finite
fields.

More information on POLYVAL can be found in the HCTR2 paper:
"Length-preserving encryption with HCTR2":
https://eprint.iacr.org/2021/1441.pdf

Signed-off-by: Nathan Huckleberry <nhuck@google.com>
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2022-06-10 16:40:17 +08:00

322 lines
9.1 KiB
ArmAsm

/* SPDX-License-Identifier: GPL-2.0 */
/*
* Copyright 2021 Google LLC
*/
/*
* This is an efficient implementation of POLYVAL using intel PCLMULQDQ-NI
* instructions. 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>
#include <asm/frame.h>
#define STRIDE_BLOCKS 8
#define GSTAR %xmm7
#define PL %xmm8
#define PH %xmm9
#define TMP_XMM %xmm11
#define LO %xmm12
#define HI %xmm13
#define MI %xmm14
#define SUM %xmm15
#define KEY_POWERS %rdi
#define MSG %rsi
#define BLOCKS_LEFT %rdx
#define ACCUMULATOR %rcx
#define TMP %rax
.section .rodata.cst16.gstar, "aM", @progbits, 16
.align 16
.Lgstar:
.quad 0xc200000000000000, 0xc200000000000000
.text
/*
* Performs schoolbook1_iteration on two lists of 128-bit polynomials of length
* count pointed to by MSG and KEY_POWERS.
*/
.macro schoolbook1 count
.set i, 0
.rept (\count)
schoolbook1_iteration i 0
.set i, (i +1)
.endr
.endm
/*
* Computes the product of two 128-bit polynomials at the memory locations
* specified by (MSG + 16*i) and (KEY_POWERS + 16*i) and XORs the components of
* the 256-bit product into LO, MI, HI.
*
* Given:
* X = [X_1 : X_0]
* Y = [Y_1 : Y_0]
*
* We compute:
* LO += X_0 * Y_0
* MI += X_0 * Y_1 + X_1 * Y_0
* HI += X_1 * Y_1
*
* Later, the 256-bit result can be extracted as:
* [HI_1 : HI_0 + MI_1 : LO_1 + MI_0 : LO_0]
* This step is done when computing the polynomial reduction for efficiency
* reasons.
*
* If xor_sum == 1, then also XOR the value of SUM into m_0. This avoids an
* extra multiplication of SUM and h^8.
*/
.macro schoolbook1_iteration i xor_sum
movups (16*\i)(MSG), %xmm0
.if (\i == 0 && \xor_sum == 1)
pxor SUM, %xmm0
.endif
vpclmulqdq $0x01, (16*\i)(KEY_POWERS), %xmm0, %xmm2
vpclmulqdq $0x00, (16*\i)(KEY_POWERS), %xmm0, %xmm1
vpclmulqdq $0x10, (16*\i)(KEY_POWERS), %xmm0, %xmm3
vpclmulqdq $0x11, (16*\i)(KEY_POWERS), %xmm0, %xmm4
vpxor %xmm2, MI, MI
vpxor %xmm1, LO, LO
vpxor %xmm4, HI, HI
vpxor %xmm3, MI, MI
.endm
/*
* Performs the same computation as schoolbook1_iteration, except we expect the
* arguments to already be loaded into xmm0 and xmm1 and we set the result
* registers LO, MI, and HI directly rather than XOR'ing into them.
*/
.macro schoolbook1_noload
vpclmulqdq $0x01, %xmm0, %xmm1, MI
vpclmulqdq $0x10, %xmm0, %xmm1, %xmm2
vpclmulqdq $0x00, %xmm0, %xmm1, LO
vpclmulqdq $0x11, %xmm0, %xmm1, HI
vpxor %xmm2, MI, MI
.endm
/*
* Computes the 256-bit polynomial represented by LO, HI, MI. Stores
* the result in PL, PH.
* [PH : PL] = [HI_1 : HI_0 + MI_1 : LO_1 + MI_0 : LO_0]
*/
.macro schoolbook2
vpslldq $8, MI, PL
vpsrldq $8, MI, PH
pxor LO, PL
pxor HI, PH
.endm
/*
* Computes the 128-bit reduction of PH : PL. 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 PH : PL = 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) * (P_1 + T_0)
* 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 into dest, rather than separately XORing P_1 : P_0 and T_0 :
* T_1 into dest. This allows us to reuse P_1 + T_0 when computing V.
*/
.macro montgomery_reduction dest
vpclmulqdq $0x00, PL, GSTAR, TMP_XMM # TMP_XMM = T_1 : T_0 = P_0 * g*(x)
pshufd $0b01001110, TMP_XMM, TMP_XMM # TMP_XMM = T_0 : T_1
pxor PL, TMP_XMM # TMP_XMM = P_1 + T_0 : P_0 + T_1
pxor TMP_XMM, PH # PH = P_3 + P_1 + T_0 : P_2 + P_0 + T_1
pclmulqdq $0x11, GSTAR, TMP_XMM # TMP_XMM = V_1 : V_0 = V = [(P_1 + T_0) * g*(x)]
vpxor TMP_XMM, PH, \dest
.endm
/*
* Compute schoolbook multiplication for 8 blocks
* m_0h^8 + ... + m_7h^1
*
* 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.
*/
.macro full_stride reduce
pxor LO, LO
pxor HI, HI
pxor MI, MI
schoolbook1_iteration 7 0
.if \reduce
vpclmulqdq $0x00, PL, GSTAR, TMP_XMM
.endif
schoolbook1_iteration 6 0
.if \reduce
pshufd $0b01001110, TMP_XMM, TMP_XMM
.endif
schoolbook1_iteration 5 0
.if \reduce
pxor PL, TMP_XMM
.endif
schoolbook1_iteration 4 0
.if \reduce
pxor TMP_XMM, PH
.endif
schoolbook1_iteration 3 0
.if \reduce
pclmulqdq $0x11, GSTAR, TMP_XMM
.endif
schoolbook1_iteration 2 0
.if \reduce
vpxor TMP_XMM, PH, SUM
.endif
schoolbook1_iteration 1 0
schoolbook1_iteration 0 1
addq $(8*16), MSG
schoolbook2
.endm
/*
* Process BLOCKS_LEFT blocks, where 0 < BLOCKS_LEFT < STRIDE_BLOCKS
*/
.macro partial_stride
mov BLOCKS_LEFT, TMP
shlq $4, TMP
addq $(16*STRIDE_BLOCKS), KEY_POWERS
subq TMP, KEY_POWERS
movups (MSG), %xmm0
pxor SUM, %xmm0
movaps (KEY_POWERS), %xmm1
schoolbook1_noload
dec BLOCKS_LEFT
addq $16, MSG
addq $16, KEY_POWERS
test $4, BLOCKS_LEFT
jz .Lpartial4BlocksDone
schoolbook1 4
addq $(4*16), MSG
addq $(4*16), KEY_POWERS
.Lpartial4BlocksDone:
test $2, BLOCKS_LEFT
jz .Lpartial2BlocksDone
schoolbook1 2
addq $(2*16), MSG
addq $(2*16), KEY_POWERS
.Lpartial2BlocksDone:
test $1, BLOCKS_LEFT
jz .LpartialDone
schoolbook1 1
.LpartialDone:
schoolbook2
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 clmul_polyval_mul(u8 *op1, const u8 *op2);
*/
SYM_FUNC_START(clmul_polyval_mul)
FRAME_BEGIN
vmovdqa .Lgstar(%rip), GSTAR
movups (%rdi), %xmm0
movups (%rsi), %xmm1
schoolbook1_noload
schoolbook2
montgomery_reduction SUM
movups SUM, (%rdi)
FRAME_END
RET
SYM_FUNC_END(clmul_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.
*
* rdi - pointer to precomputed key powers h^8 ... h^1
* rsi - pointer to message blocks
* rdx - number of blocks to hash
* rcx - pointer to the accumulator
*
* void clmul_polyval_update(const struct polyval_tfm_ctx *keys,
* const u8 *in, size_t nblocks, u8 *accumulator);
*/
SYM_FUNC_START(clmul_polyval_update)
FRAME_BEGIN
vmovdqa .Lgstar(%rip), GSTAR
movups (ACCUMULATOR), SUM
subq $STRIDE_BLOCKS, BLOCKS_LEFT
js .LstrideLoopExit
full_stride 0
subq $STRIDE_BLOCKS, BLOCKS_LEFT
js .LstrideLoopExitReduce
.LstrideLoop:
full_stride 1
subq $STRIDE_BLOCKS, BLOCKS_LEFT
jns .LstrideLoop
.LstrideLoopExitReduce:
montgomery_reduction SUM
.LstrideLoopExit:
add $STRIDE_BLOCKS, BLOCKS_LEFT
jz .LskipPartial
partial_stride
.LskipPartial:
movups SUM, (ACCUMULATOR)
FRAME_END
RET
SYM_FUNC_END(clmul_polyval_update)