| 1 | /* SPDX-License-Identifier: GPL-2.0-or-later */ |
| 2 | /* |
| 3 | * POLYVAL library functions, x86_64 optimized |
| 4 | * |
| 5 | * Copyright 2025 Google LLC |
| 6 | */ |
| 7 | #include <asm/fpu/api.h> |
| 8 | #include <linux/cpufeature.h> |
| 9 | |
| 10 | #define NUM_H_POWERS 8 |
| 11 | |
| 12 | static __ro_after_init DEFINE_STATIC_KEY_FALSE(have_pclmul_avx); |
| 13 | |
| 14 | asmlinkage void polyval_mul_pclmul_avx(struct polyval_elem *a, |
| 15 | const struct polyval_elem *b); |
| 16 | asmlinkage void polyval_blocks_pclmul_avx(struct polyval_elem *acc, |
| 17 | const struct polyval_key *key, |
| 18 | const u8 *data, size_t nblocks); |
| 19 | |
| 20 | static void polyval_preparekey_arch(struct polyval_key *key, |
| 21 | const u8 raw_key[POLYVAL_BLOCK_SIZE]) |
| 22 | { |
| 23 | static_assert(ARRAY_SIZE(key->h_powers) == NUM_H_POWERS); |
| 24 | memcpy(&key->h_powers[NUM_H_POWERS - 1], raw_key, POLYVAL_BLOCK_SIZE); |
| 25 | if (static_branch_likely(&have_pclmul_avx) && irq_fpu_usable()) { |
| 26 | kernel_fpu_begin(); |
| 27 | for (int i = NUM_H_POWERS - 2; i >= 0; i--) { |
| 28 | key->h_powers[i] = key->h_powers[i + 1]; |
| 29 | polyval_mul_pclmul_avx( |
| 30 | a: &key->h_powers[i], |
| 31 | b: &key->h_powers[NUM_H_POWERS - 1]); |
| 32 | } |
| 33 | kernel_fpu_end(); |
| 34 | } else { |
| 35 | for (int i = NUM_H_POWERS - 2; i >= 0; i--) { |
| 36 | key->h_powers[i] = key->h_powers[i + 1]; |
| 37 | polyval_mul_generic(a: &key->h_powers[i], |
| 38 | b: &key->h_powers[NUM_H_POWERS - 1]); |
| 39 | } |
| 40 | } |
| 41 | } |
| 42 | |
| 43 | static void polyval_mul_arch(struct polyval_elem *acc, |
| 44 | const struct polyval_key *key) |
| 45 | { |
| 46 | if (static_branch_likely(&have_pclmul_avx) && irq_fpu_usable()) { |
| 47 | kernel_fpu_begin(); |
| 48 | polyval_mul_pclmul_avx(a: acc, b: &key->h_powers[NUM_H_POWERS - 1]); |
| 49 | kernel_fpu_end(); |
| 50 | } else { |
| 51 | polyval_mul_generic(a: acc, b: &key->h_powers[NUM_H_POWERS - 1]); |
| 52 | } |
| 53 | } |
| 54 | |
| 55 | static void polyval_blocks_arch(struct polyval_elem *acc, |
| 56 | const struct polyval_key *key, |
| 57 | const u8 *data, size_t nblocks) |
| 58 | { |
| 59 | if (static_branch_likely(&have_pclmul_avx) && irq_fpu_usable()) { |
| 60 | do { |
| 61 | /* Allow rescheduling every 4 KiB. */ |
| 62 | size_t n = min_t(size_t, nblocks, |
| 63 | 4096 / POLYVAL_BLOCK_SIZE); |
| 64 | |
| 65 | kernel_fpu_begin(); |
| 66 | polyval_blocks_pclmul_avx(acc, key, data, nblocks: n); |
| 67 | kernel_fpu_end(); |
| 68 | data += n * POLYVAL_BLOCK_SIZE; |
| 69 | nblocks -= n; |
| 70 | } while (nblocks); |
| 71 | } else { |
| 72 | polyval_blocks_generic(acc, key: &key->h_powers[NUM_H_POWERS - 1], |
| 73 | data, nblocks); |
| 74 | } |
| 75 | } |
| 76 | |
| 77 | #define polyval_mod_init_arch polyval_mod_init_arch |
| 78 | static void polyval_mod_init_arch(void) |
| 79 | { |
| 80 | if (boot_cpu_has(X86_FEATURE_PCLMULQDQ) && |
| 81 | boot_cpu_has(X86_FEATURE_AVX)) |
| 82 | static_branch_enable(&have_pclmul_avx); |
| 83 | } |
| 84 | |