1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * x86-optimized CRC32 functions
4 *
5 * Copyright (C) 2008 Intel Corporation
6 * Copyright 2012 Xyratex Technology Limited
7 * Copyright 2024 Google LLC
8 */
9
10 #include "crc-pclmul-template.h"
11
12 static __ro_after_init DEFINE_STATIC_KEY_FALSE(have_crc32);
13 static __ro_after_init DEFINE_STATIC_KEY_FALSE(have_pclmulqdq);
14 static __ro_after_init DEFINE_STATIC_KEY_FALSE(have_vpclmul_avx512);
15
16 DECLARE_CRC_PCLMUL_FUNCS(crc32_lsb, u32);
17
crc32_le_arch(u32 crc,const u8 * p,size_t len)18 static inline u32 crc32_le_arch(u32 crc, const u8 *p, size_t len)
19 {
20 CRC_PCLMUL(crc, p, len, crc32_lsb, crc32_lsb_0xedb88320_consts,
21 have_pclmulqdq);
22 return crc32_le_base(crc, p, len);
23 }
24
25 #ifdef CONFIG_X86_64
26 #define CRC32_INST "crc32q %1, %q0"
27 #else
28 #define CRC32_INST "crc32l %1, %0"
29 #endif
30
31 /*
32 * Use carryless multiply version of crc32c when buffer size is >= 512 to
33 * account for FPU state save/restore overhead.
34 */
35 #define CRC32C_PCLMUL_BREAKEVEN 512
36
37 asmlinkage u32 crc32c_x86_3way(u32 crc, const u8 *buffer, size_t len);
38
crc32c_arch(u32 crc,const u8 * p,size_t len)39 static inline u32 crc32c_arch(u32 crc, const u8 *p, size_t len)
40 {
41 size_t num_longs;
42
43 if (!static_branch_likely(&have_crc32))
44 return crc32c_base(crc, p, len);
45
46 if (IS_ENABLED(CONFIG_X86_64) && len >= CRC32C_PCLMUL_BREAKEVEN &&
47 static_branch_likely(&have_pclmulqdq) && crypto_simd_usable()) {
48 /*
49 * Long length, the vector registers are usable, and the CPU is
50 * 64-bit and supports both CRC32 and PCLMULQDQ instructions.
51 * It is worthwhile to divide the data into multiple streams,
52 * CRC them independently, and combine them using PCLMULQDQ.
53 * crc32c_x86_3way() does this using 3 streams, which is the
54 * most that x86_64 CPUs have traditionally been capable of.
55 *
56 * However, due to improved VPCLMULQDQ performance on newer
57 * CPUs, use crc32_lsb_vpclmul_avx512() instead of
58 * crc32c_x86_3way() when the CPU supports VPCLMULQDQ and has a
59 * "good" implementation of AVX-512.
60 *
61 * Future work: the optimal strategy on Zen 3--5 is actually to
62 * use both crc32q and VPCLMULQDQ in parallel. Unfortunately,
63 * different numbers of streams and vector lengths are optimal
64 * on each CPU microarchitecture, making it challenging to take
65 * advantage of this. (Zen 5 even supports 7 parallel crc32q, a
66 * major upgrade.) For now, just choose between
67 * crc32c_x86_3way() and crc32_lsb_vpclmul_avx512(). The latter
68 * is needed anyway for crc32_le(), so we just reuse it here.
69 */
70 kernel_fpu_begin();
71 if (static_branch_likely(&have_vpclmul_avx512))
72 crc = crc32_lsb_vpclmul_avx512(crc, p, len,
73 crc32_lsb_0x82f63b78_consts.fold_across_128_bits_consts);
74 else
75 crc = crc32c_x86_3way(crc, p, len);
76 kernel_fpu_end();
77 return crc;
78 }
79
80 /*
81 * Short length, XMM registers unusable, or the CPU is 32-bit; but the
82 * CPU supports CRC32 instructions. Just issue a single stream of CRC32
83 * instructions inline. While this doesn't use the CPU's CRC32
84 * throughput very well, it avoids the need to combine streams. Stream
85 * combination would be inefficient here.
86 */
87
88 for (num_longs = len / sizeof(unsigned long);
89 num_longs != 0; num_longs--, p += sizeof(unsigned long))
90 asm(CRC32_INST : "+r" (crc) : ASM_INPUT_RM (*(unsigned long *)p));
91
92 if (sizeof(unsigned long) > 4 && (len & 4)) {
93 asm("crc32l %1, %0" : "+r" (crc) : ASM_INPUT_RM (*(u32 *)p));
94 p += 4;
95 }
96 if (len & 2) {
97 asm("crc32w %1, %0" : "+r" (crc) : ASM_INPUT_RM (*(u16 *)p));
98 p += 2;
99 }
100 if (len & 1)
101 asm("crc32b %1, %0" : "+r" (crc) : ASM_INPUT_RM (*p));
102
103 return crc;
104 }
105
106 #define crc32_be_arch crc32_be_base /* not implemented on this arch */
107
108 #define crc32_mod_init_arch crc32_mod_init_arch
crc32_mod_init_arch(void)109 static inline void crc32_mod_init_arch(void)
110 {
111 if (boot_cpu_has(X86_FEATURE_XMM4_2))
112 static_branch_enable(&have_crc32);
113 if (boot_cpu_has(X86_FEATURE_PCLMULQDQ)) {
114 static_branch_enable(&have_pclmulqdq);
115 if (have_vpclmul()) {
116 if (have_avx512()) {
117 static_call_update(crc32_lsb_pclmul,
118 crc32_lsb_vpclmul_avx512);
119 static_branch_enable(&have_vpclmul_avx512);
120 } else {
121 static_call_update(crc32_lsb_pclmul,
122 crc32_lsb_vpclmul_avx2);
123 }
124 }
125 }
126 }
127
crc32_optimizations_arch(void)128 static inline u32 crc32_optimizations_arch(void)
129 {
130 u32 optimizations = 0;
131
132 if (static_key_enabled(&have_crc32))
133 optimizations |= CRC32C_OPTIMIZATION;
134 if (static_key_enabled(&have_pclmulqdq))
135 optimizations |= CRC32_LE_OPTIMIZATION;
136 return optimizations;
137 }
138