korg/linux
0
0
Fork 0
mirror of https://mirrors.bfsu.edu.cn/git/linux.git synced 2024-11-30 07:34:12 +08:00
Code Issues Packages Projects Releases Wiki Activity
cc9cf350d1
linux/arch/arm/crypto/chacha-neon-core.S
Ard Biesheuvel 86cd97ec4b crypto: arm/chacha-neon - optimize for non-block size multiples 
The current NEON based ChaCha implementation for ARM is optimized for
multiples of 4x the ChaCha block size (64 bytes). This makes sense for
block encryption, but given that ChaCha is also often used in the
context of networking, it makes sense to consider arbitrary length
inputs as well.

For example, WireGuard typically uses 1420 byte packets, and performing
ChaCha encryption involves 5 invocations of chacha_4block_xor_neon()
and 3 invocations of chacha_block_xor_neon(), where the last one also
involves a memcpy() using a buffer on the stack to process the final
chunk of 1420 % 64 == 12 bytes.

Let's optimize for this case as well, by letting chacha_4block_xor_neon()
deal with any input size between 64 and 256 bytes, using NEON permutation
instructions and overlapping loads and stores. This way, the 140 byte
tail of a 1420 byte input buffer can simply be processed in one go.

This results in the following performance improvements for 1420 byte
blocks, without significant impact on power-of-2 input sizes. (Note
that Raspberry Pi is widely used in combination with a 32-bit kernel,
even though the core is 64-bit capable)

   Cortex-A8  (BeagleBone)       :   7%
   Cortex-A15 (Calxeda Midway)   :  21%
   Cortex-A53 (Raspberry Pi 3)   :   3%
   Cortex-A72 (Raspberry Pi 4)   :  19%

Cc: Eric Biggers <ebiggers@google.com>
Cc: "Jason A . Donenfeld" <Jason@zx2c4.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-11-13 20:38:44 +11:00

644 lines
15 KiB
ArmAsm
Raw Blame History

/*
* ChaCha/XChaCha NEON helper functions
*
* Copyright (C) 2016 Linaro, Ltd. <ard.biesheuvel@linaro.org>
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License version 2 as
* published by the Free Software Foundation.
*
* Based on:
* ChaCha20 256-bit cipher algorithm, RFC7539, x64 SSE3 functions
*
* Copyright (C) 2015 Martin Willi
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*/
/*
* NEON doesn't have a rotate instruction. The alternatives are, more or less:
*
* (a) vshl.u32 + vsri.u32 (needs temporary register)
* (b) vshl.u32 + vshr.u32 + vorr (needs temporary register)
* (c) vrev32.16 (16-bit rotations only)
* (d) vtbl.8 + vtbl.8 (multiple of 8 bits rotations only,
* needs index vector)
*
* ChaCha has 16, 12, 8, and 7-bit rotations. For the 12 and 7-bit rotations,
* the only choices are (a) and (b). We use (a) since it takes two-thirds the
* cycles of (b) on both Cortex-A7 and Cortex-A53.
*
* For the 16-bit rotation, we use vrev32.16 since it's consistently fastest
* and doesn't need a temporary register.
*
* For the 8-bit rotation, we use vtbl.8 + vtbl.8. On Cortex-A7, this sequence
* is twice as fast as (a), even when doing (a) on multiple registers
* simultaneously to eliminate the stall between vshl and vsri. Also, it
* parallelizes better when temporary registers are scarce.
*
* A disadvantage is that on Cortex-A53, the vtbl sequence is the same speed as
* (a), so the need to load the rotation table actually makes the vtbl method
* slightly slower overall on that CPU (~1.3% slower ChaCha20). Still, it
* seems to be a good compromise to get a more significant speed boost on some
* CPUs, e.g. ~4.8% faster ChaCha20 on Cortex-A7.
*/
#include <linux/linkage.h>
#include <asm/cache.h>
.text
.fpu neon
.align 5
/*
* chacha_permute - permute one block
*
* Permute one 64-byte block where the state matrix is stored in the four NEON
* registers q0-q3. It performs matrix operations on four words in parallel,
* but requires shuffling to rearrange the words after each round.
*
* The round count is given in r3.
*
* Clobbers: r3, ip, q4-q5
*/
chacha_permute:
adr ip, .Lrol8_table
vld1.8 {d10}, [ip, :64]
.Ldoubleround:
// x0 += x1, x3 = rotl32(x3 ^ x0, 16)
vadd.i32 q0, q0, q1
veor q3, q3, q0
vrev32.16 q3, q3
// x2 += x3, x1 = rotl32(x1 ^ x2, 12)
vadd.i32 q2, q2, q3
veor q4, q1, q2
vshl.u32 q1, q4, #12
vsri.u32 q1, q4, #20
// x0 += x1, x3 = rotl32(x3 ^ x0, 8)
vadd.i32 q0, q0, q1
veor q3, q3, q0
vtbl.8 d6, {d6}, d10
vtbl.8 d7, {d7}, d10
// x2 += x3, x1 = rotl32(x1 ^ x2, 7)
vadd.i32 q2, q2, q3
veor q4, q1, q2
vshl.u32 q1, q4, #7
vsri.u32 q1, q4, #25
// x1 = shuffle32(x1, MASK(0, 3, 2, 1))
vext.8 q1, q1, q1, #4
// x2 = shuffle32(x2, MASK(1, 0, 3, 2))
vext.8 q2, q2, q2, #8
// x3 = shuffle32(x3, MASK(2, 1, 0, 3))
vext.8 q3, q3, q3, #12
// x0 += x1, x3 = rotl32(x3 ^ x0, 16)
vadd.i32 q0, q0, q1
veor q3, q3, q0
vrev32.16 q3, q3
// x2 += x3, x1 = rotl32(x1 ^ x2, 12)
vadd.i32 q2, q2, q3
veor q4, q1, q2
vshl.u32 q1, q4, #12
vsri.u32 q1, q4, #20
// x0 += x1, x3 = rotl32(x3 ^ x0, 8)
vadd.i32 q0, q0, q1
veor q3, q3, q0
vtbl.8 d6, {d6}, d10
vtbl.8 d7, {d7}, d10
// x2 += x3, x1 = rotl32(x1 ^ x2, 7)
vadd.i32 q2, q2, q3
veor q4, q1, q2
vshl.u32 q1, q4, #7
vsri.u32 q1, q4, #25
// x1 = shuffle32(x1, MASK(2, 1, 0, 3))
vext.8 q1, q1, q1, #12
// x2 = shuffle32(x2, MASK(1, 0, 3, 2))
vext.8 q2, q2, q2, #8
// x3 = shuffle32(x3, MASK(0, 3, 2, 1))
vext.8 q3, q3, q3, #4
subs r3, r3, #2
bne .Ldoubleround
bx lr
ENDPROC(chacha_permute)
ENTRY(chacha_block_xor_neon)
// r0: Input state matrix, s
// r1: 1 data block output, o
// r2: 1 data block input, i
// r3: nrounds
push {lr}
// x0..3 = s0..3
add ip, r0, #0x20
vld1.32 {q0-q1}, [r0]
vld1.32 {q2-q3}, [ip]
vmov q8, q0
vmov q9, q1
vmov q10, q2
vmov q11, q3
bl chacha_permute
add ip, r2, #0x20
vld1.8 {q4-q5}, [r2]
vld1.8 {q6-q7}, [ip]
// o0 = i0 ^ (x0 + s0)
vadd.i32 q0, q0, q8
veor q0, q0, q4
// o1 = i1 ^ (x1 + s1)
vadd.i32 q1, q1, q9
veor q1, q1, q5
// o2 = i2 ^ (x2 + s2)
vadd.i32 q2, q2, q10
veor q2, q2, q6
// o3 = i3 ^ (x3 + s3)
vadd.i32 q3, q3, q11
veor q3, q3, q7
add ip, r1, #0x20
vst1.8 {q0-q1}, [r1]
vst1.8 {q2-q3}, [ip]
pop {pc}
ENDPROC(chacha_block_xor_neon)
ENTRY(hchacha_block_neon)
// r0: Input state matrix, s
// r1: output (8 32-bit words)
// r2: nrounds
push {lr}
vld1.32 {q0-q1}, [r0]!
vld1.32 {q2-q3}, [r0]
mov r3, r2
bl chacha_permute
vst1.32 {q0}, [r1]!
vst1.32 {q3}, [r1]
pop {pc}
ENDPROC(hchacha_block_neon)
.align 4
.Lctrinc: .word 0, 1, 2, 3
.Lrol8_table: .byte 3, 0, 1, 2, 7, 4, 5, 6
.align 5
ENTRY(chacha_4block_xor_neon)
push {r4, lr}
mov r4, sp // preserve the stack pointer
sub ip, sp, #0x20 // allocate a 32 byte buffer
bic ip, ip, #0x1f // aligned to 32 bytes
mov sp, ip
// r0: Input state matrix, s
// r1: 4 data blocks output, o
// r2: 4 data blocks input, i
// r3: nrounds
//
// This function encrypts four consecutive ChaCha blocks by loading
// the state matrix in NEON registers four times. The algorithm performs
// each operation on the corresponding word of each state matrix, hence
// requires no word shuffling. The words are re-interleaved before the
// final addition of the original state and the XORing step.
//
// x0..15[0-3] = s0..15[0-3]
add ip, r0, #0x20
vld1.32 {q0-q1}, [r0]
vld1.32 {q2-q3}, [ip]
adr lr, .Lctrinc
vdup.32 q15, d7[1]
vdup.32 q14, d7[0]
vld1.32 {q4}, [lr, :128]
vdup.32 q13, d6[1]
vdup.32 q12, d6[0]
vdup.32 q11, d5[1]
vdup.32 q10, d5[0]
vadd.u32 q12, q12, q4 // x12 += counter values 0-3
vdup.32 q9, d4[1]
vdup.32 q8, d4[0]
vdup.32 q7, d3[1]
vdup.32 q6, d3[0]
vdup.32 q5, d2[1]
vdup.32 q4, d2[0]
vdup.32 q3, d1[1]
vdup.32 q2, d1[0]
vdup.32 q1, d0[1]
vdup.32 q0, d0[0]
adr ip, .Lrol8_table
b 1f
.Ldoubleround4:
vld1.32 {q8-q9}, [sp, :256]
1:
// x0 += x4, x12 = rotl32(x12 ^ x0, 16)
// x1 += x5, x13 = rotl32(x13 ^ x1, 16)
// x2 += x6, x14 = rotl32(x14 ^ x2, 16)
// x3 += x7, x15 = rotl32(x15 ^ x3, 16)
vadd.i32 q0, q0, q4
vadd.i32 q1, q1, q5
vadd.i32 q2, q2, q6
vadd.i32 q3, q3, q7
veor q12, q12, q0
veor q13, q13, q1
veor q14, q14, q2
veor q15, q15, q3
vrev32.16 q12, q12
vrev32.16 q13, q13
vrev32.16 q14, q14
vrev32.16 q15, q15
// x8 += x12, x4 = rotl32(x4 ^ x8, 12)
// x9 += x13, x5 = rotl32(x5 ^ x9, 12)
// x10 += x14, x6 = rotl32(x6 ^ x10, 12)
// x11 += x15, x7 = rotl32(x7 ^ x11, 12)
vadd.i32 q8, q8, q12
vadd.i32 q9, q9, q13
vadd.i32 q10, q10, q14
vadd.i32 q11, q11, q15
vst1.32 {q8-q9}, [sp, :256]
veor q8, q4, q8
veor q9, q5, q9
vshl.u32 q4, q8, #12
vshl.u32 q5, q9, #12
vsri.u32 q4, q8, #20
vsri.u32 q5, q9, #20
veor q8, q6, q10
veor q9, q7, q11
vshl.u32 q6, q8, #12
vshl.u32 q7, q9, #12
vsri.u32 q6, q8, #20
vsri.u32 q7, q9, #20
// x0 += x4, x12 = rotl32(x12 ^ x0, 8)
// x1 += x5, x13 = rotl32(x13 ^ x1, 8)
// x2 += x6, x14 = rotl32(x14 ^ x2, 8)
// x3 += x7, x15 = rotl32(x15 ^ x3, 8)
vld1.8 {d16}, [ip, :64]
vadd.i32 q0, q0, q4
vadd.i32 q1, q1, q5
vadd.i32 q2, q2, q6
vadd.i32 q3, q3, q7
veor q12, q12, q0
veor q13, q13, q1
veor q14, q14, q2
veor q15, q15, q3
vtbl.8 d24, {d24}, d16
vtbl.8 d25, {d25}, d16
vtbl.8 d26, {d26}, d16
vtbl.8 d27, {d27}, d16
vtbl.8 d28, {d28}, d16
vtbl.8 d29, {d29}, d16
vtbl.8 d30, {d30}, d16
vtbl.8 d31, {d31}, d16
vld1.32 {q8-q9}, [sp, :256]
// x8 += x12, x4 = rotl32(x4 ^ x8, 7)
// x9 += x13, x5 = rotl32(x5 ^ x9, 7)
// x10 += x14, x6 = rotl32(x6 ^ x10, 7)
// x11 += x15, x7 = rotl32(x7 ^ x11, 7)
vadd.i32 q8, q8, q12
vadd.i32 q9, q9, q13
vadd.i32 q10, q10, q14
vadd.i32 q11, q11, q15
vst1.32 {q8-q9}, [sp, :256]
veor q8, q4, q8
veor q9, q5, q9
vshl.u32 q4, q8, #7
vshl.u32 q5, q9, #7
vsri.u32 q4, q8, #25
vsri.u32 q5, q9, #25
veor q8, q6, q10
veor q9, q7, q11
vshl.u32 q6, q8, #7
vshl.u32 q7, q9, #7
vsri.u32 q6, q8, #25
vsri.u32 q7, q9, #25
vld1.32 {q8-q9}, [sp, :256]
// x0 += x5, x15 = rotl32(x15 ^ x0, 16)
// x1 += x6, x12 = rotl32(x12 ^ x1, 16)
// x2 += x7, x13 = rotl32(x13 ^ x2, 16)
// x3 += x4, x14 = rotl32(x14 ^ x3, 16)
vadd.i32 q0, q0, q5
vadd.i32 q1, q1, q6
vadd.i32 q2, q2, q7
vadd.i32 q3, q3, q4
veor q15, q15, q0
veor q12, q12, q1
veor q13, q13, q2
veor q14, q14, q3
vrev32.16 q15, q15
vrev32.16 q12, q12
vrev32.16 q13, q13
vrev32.16 q14, q14
// x10 += x15, x5 = rotl32(x5 ^ x10, 12)
// x11 += x12, x6 = rotl32(x6 ^ x11, 12)
// x8 += x13, x7 = rotl32(x7 ^ x8, 12)
// x9 += x14, x4 = rotl32(x4 ^ x9, 12)
vadd.i32 q10, q10, q15
vadd.i32 q11, q11, q12
vadd.i32 q8, q8, q13
vadd.i32 q9, q9, q14
vst1.32 {q8-q9}, [sp, :256]
veor q8, q7, q8
veor q9, q4, q9
vshl.u32 q7, q8, #12
vshl.u32 q4, q9, #12
vsri.u32 q7, q8, #20
vsri.u32 q4, q9, #20
veor q8, q5, q10
veor q9, q6, q11
vshl.u32 q5, q8, #12
vshl.u32 q6, q9, #12
vsri.u32 q5, q8, #20
vsri.u32 q6, q9, #20
// x0 += x5, x15 = rotl32(x15 ^ x0, 8)
// x1 += x6, x12 = rotl32(x12 ^ x1, 8)
// x2 += x7, x13 = rotl32(x13 ^ x2, 8)
// x3 += x4, x14 = rotl32(x14 ^ x3, 8)
vld1.8 {d16}, [ip, :64]
vadd.i32 q0, q0, q5
vadd.i32 q1, q1, q6
vadd.i32 q2, q2, q7
vadd.i32 q3, q3, q4
veor q15, q15, q0
veor q12, q12, q1
veor q13, q13, q2
veor q14, q14, q3
vtbl.8 d30, {d30}, d16
vtbl.8 d31, {d31}, d16
vtbl.8 d24, {d24}, d16
vtbl.8 d25, {d25}, d16
vtbl.8 d26, {d26}, d16
vtbl.8 d27, {d27}, d16
vtbl.8 d28, {d28}, d16
vtbl.8 d29, {d29}, d16
vld1.32 {q8-q9}, [sp, :256]
// x10 += x15, x5 = rotl32(x5 ^ x10, 7)
// x11 += x12, x6 = rotl32(x6 ^ x11, 7)
// x8 += x13, x7 = rotl32(x7 ^ x8, 7)
// x9 += x14, x4 = rotl32(x4 ^ x9, 7)
vadd.i32 q10, q10, q15
vadd.i32 q11, q11, q12
vadd.i32 q8, q8, q13
vadd.i32 q9, q9, q14
vst1.32 {q8-q9}, [sp, :256]
veor q8, q7, q8
veor q9, q4, q9
vshl.u32 q7, q8, #7
vshl.u32 q4, q9, #7
vsri.u32 q7, q8, #25
vsri.u32 q4, q9, #25
veor q8, q5, q10
veor q9, q6, q11
vshl.u32 q5, q8, #7
vshl.u32 q6, q9, #7
vsri.u32 q5, q8, #25
vsri.u32 q6, q9, #25
subs r3, r3, #2
bne .Ldoubleround4
// x0..7[0-3] are in q0-q7, x10..15[0-3] are in q10-q15.
// x8..9[0-3] are on the stack.
// Re-interleave the words in the first two rows of each block (x0..7).
// Also add the counter values 0-3 to x12[0-3].
vld1.32 {q8}, [lr, :128] // load counter values 0-3
vzip.32 q0, q1 // => (0 1 0 1) (0 1 0 1)
vzip.32 q2, q3 // => (2 3 2 3) (2 3 2 3)
vzip.32 q4, q5 // => (4 5 4 5) (4 5 4 5)
vzip.32 q6, q7 // => (6 7 6 7) (6 7 6 7)
vadd.u32 q12, q8 // x12 += counter values 0-3
vswp d1, d4
vswp d3, d6
vld1.32 {q8-q9}, [r0]! // load s0..7
vswp d9, d12
vswp d11, d14
// Swap q1 and q4 so that we'll free up consecutive registers (q0-q1)
// after XORing the first 32 bytes.
vswp q1, q4
// First two rows of each block are (q0 q1) (q2 q6) (q4 q5) (q3 q7)
// x0..3[0-3] += s0..3[0-3] (add orig state to 1st row of each block)
vadd.u32 q0, q0, q8
vadd.u32 q2, q2, q8
vadd.u32 q4, q4, q8
vadd.u32 q3, q3, q8
// x4..7[0-3] += s4..7[0-3] (add orig state to 2nd row of each block)
vadd.u32 q1, q1, q9
vadd.u32 q6, q6, q9
vadd.u32 q5, q5, q9
vadd.u32 q7, q7, q9
// XOR first 32 bytes using keystream from first two rows of first block
vld1.8 {q8-q9}, [r2]!
veor q8, q8, q0
veor q9, q9, q1
vst1.8 {q8-q9}, [r1]!
// Re-interleave the words in the last two rows of each block (x8..15).
vld1.32 {q8-q9}, [sp, :256]
mov sp, r4 // restore original stack pointer
ldr r4, [r4, #8] // load number of bytes
vzip.32 q12, q13 // => (12 13 12 13) (12 13 12 13)
vzip.32 q14, q15 // => (14 15 14 15) (14 15 14 15)
vzip.32 q8, q9 // => (8 9 8 9) (8 9 8 9)
vzip.32 q10, q11 // => (10 11 10 11) (10 11 10 11)
vld1.32 {q0-q1}, [r0] // load s8..15
vswp d25, d28
vswp d27, d30
vswp d17, d20
vswp d19, d22
// Last two rows of each block are (q8 q12) (q10 q14) (q9 q13) (q11 q15)
// x8..11[0-3] += s8..11[0-3] (add orig state to 3rd row of each block)
vadd.u32 q8, q8, q0
vadd.u32 q10, q10, q0
vadd.u32 q9, q9, q0
vadd.u32 q11, q11, q0
// x12..15[0-3] += s12..15[0-3] (add orig state to 4th row of each block)
vadd.u32 q12, q12, q1
vadd.u32 q14, q14, q1
vadd.u32 q13, q13, q1
vadd.u32 q15, q15, q1
// XOR the rest of the data with the keystream
vld1.8 {q0-q1}, [r2]!
subs r4, r4, #96
veor q0, q0, q8
veor q1, q1, q12
ble .Lle96
vst1.8 {q0-q1}, [r1]!
vld1.8 {q0-q1}, [r2]!
subs r4, r4, #32
veor q0, q0, q2
veor q1, q1, q6
ble .Lle128
vst1.8 {q0-q1}, [r1]!
vld1.8 {q0-q1}, [r2]!
subs r4, r4, #32
veor q0, q0, q10
veor q1, q1, q14
ble .Lle160
vst1.8 {q0-q1}, [r1]!
vld1.8 {q0-q1}, [r2]!
subs r4, r4, #32
veor q0, q0, q4
veor q1, q1, q5
ble .Lle192
vst1.8 {q0-q1}, [r1]!
vld1.8 {q0-q1}, [r2]!
subs r4, r4, #32
veor q0, q0, q9
veor q1, q1, q13
ble .Lle224
vst1.8 {q0-q1}, [r1]!
vld1.8 {q0-q1}, [r2]!
subs r4, r4, #32
veor q0, q0, q3
veor q1, q1, q7
blt .Llt256
.Lout:
vst1.8 {q0-q1}, [r1]!
vld1.8 {q0-q1}, [r2]
veor q0, q0, q11
veor q1, q1, q15
vst1.8 {q0-q1}, [r1]
pop {r4, pc}
.Lle192:
vmov q4, q9
vmov q5, q13
.Lle160:
// nothing to do
.Lfinalblock:
// Process the final block if processing less than 4 full blocks.
// Entered with 32 bytes of ChaCha cipher stream in q4-q5, and the
// previous 32 byte output block that still needs to be written at
// [r1] in q0-q1.
beq .Lfullblock
.Lpartialblock:
adr lr, .Lpermute + 32
add r2, r2, r4
add lr, lr, r4
add r4, r4, r1
vld1.8 {q2-q3}, [lr]
vld1.8 {q6-q7}, [r2]
add r4, r4, #32
vtbl.8 d4, {q4-q5}, d4
vtbl.8 d5, {q4-q5}, d5
vtbl.8 d6, {q4-q5}, d6
vtbl.8 d7, {q4-q5}, d7
veor q6, q6, q2
veor q7, q7, q3
vst1.8 {q6-q7}, [r4] // overlapping stores
vst1.8 {q0-q1}, [r1]
pop {r4, pc}
.Lfullblock:
vmov q11, q4
vmov q15, q5
b .Lout
.Lle96:
vmov q4, q2
vmov q5, q6
b .Lfinalblock
.Lle128:
vmov q4, q10
vmov q5, q14
b .Lfinalblock
.Lle224:
vmov q4, q3
vmov q5, q7
b .Lfinalblock
.Llt256:
vmov q4, q11
vmov q5, q15
b .Lpartialblock
ENDPROC(chacha_4block_xor_neon)
.align L1_CACHE_SHIFT
.Lpermute:
.byte 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07
.byte 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f
.byte 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17
.byte 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f
.byte 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07
.byte 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f
.byte 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17
.byte 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f
Reference in New Issue View Git Blame Copy Permalink