Book a Demo!
CoCalc Logo Icon
StoreFeaturesDocsShareSupportNewsAboutPoliciesSign UpSign In
torvalds
GitHub Repository: torvalds/linux
 Path: blob/master/lib/crc/x86/crc-pclmul-template.h
29267 views
1
/* SPDX-License-Identifier: GPL-2.0-or-later */

2
/*

3
 * Macros for accessing the [V]PCLMULQDQ-based CRC functions that are

4
 * instantiated by crc-pclmul-template.S

5
 *

6
 * Copyright 2025 Google LLC

7
 *

8
 * Author: Eric Biggers <[email protected]>

9
 */

10
#ifndef _CRC_PCLMUL_TEMPLATE_H

11
#define _CRC_PCLMUL_TEMPLATE_H

12


13
#include <asm/cpufeatures.h>

14
#include <asm/simd.h>

15
#include <linux/static_call.h>

16
#include "crc-pclmul-consts.h"

17


18
#define DECLARE_CRC_PCLMUL_FUNCS(prefix, crc_t)				\

19
crc_t prefix##_pclmul_sse(crc_t crc, const u8 *p, size_t len,		\

20
			  const void *consts_ptr);			\

21
crc_t prefix##_vpclmul_avx2(crc_t crc, const u8 *p, size_t len,		\

22
			    const void *consts_ptr);			\

23
crc_t prefix##_vpclmul_avx512(crc_t crc, const u8 *p, size_t len,	\

24
			      const void *consts_ptr);			\

25
DEFINE_STATIC_CALL(prefix##_pclmul, prefix##_pclmul_sse)

26


27
static inline bool have_vpclmul(void)

28
{

29
	return boot_cpu_has(X86_FEATURE_VPCLMULQDQ) &&

30
	       boot_cpu_has(X86_FEATURE_AVX2) &&

31
	       cpu_has_xfeatures(XFEATURE_MASK_YMM, NULL);

32
}

33


34
static inline bool have_avx512(void)

35
{

36
	return boot_cpu_has(X86_FEATURE_AVX512BW) &&

37
	       boot_cpu_has(X86_FEATURE_AVX512VL) &&

38
	       !boot_cpu_has(X86_FEATURE_PREFER_YMM) &&

39
	       cpu_has_xfeatures(XFEATURE_MASK_AVX512, NULL);

40
}

41


42
/*

43
 * Call a [V]PCLMULQDQ optimized CRC function if the data length is at least 16

44
 * bytes, the CPU has PCLMULQDQ support, and the current context may use SIMD.

45
 *

46
 * 16 bytes is the minimum length supported by the [V]PCLMULQDQ functions.

47
 * There is overhead associated with kernel_fpu_begin() and kernel_fpu_end(),

48
 * varying by CPU and factors such as which parts of the "FPU" state userspace

49
 * has touched, which could result in a larger cutoff being better.  Indeed, a

50
 * larger cutoff is usually better for a *single* message.  However, the

51
 * overhead of the FPU section gets amortized if multiple FPU sections get

52
 * executed before returning to userspace, since the XSAVE and XRSTOR occur only

53
 * once.  Considering that and the fact that the [V]PCLMULQDQ code is lighter on

54
 * the dcache than the table-based code is, a 16-byte cutoff seems to work well.

55
 */

56
#define CRC_PCLMUL(crc, p, len, prefix, consts, have_pclmulqdq)		\

57
do {									\

58
	if ((len) >= 16 && static_branch_likely(&(have_pclmulqdq)) &&	\

59
	    likely(irq_fpu_usable())) {					\

60
		const void *consts_ptr;					\

61
									\

62
		consts_ptr = (consts).fold_across_128_bits_consts;	\

63
		kernel_fpu_begin();					\

64
		crc = static_call(prefix##_pclmul)((crc), (p), (len),	\

65
						   consts_ptr);		\

66
		kernel_fpu_end();					\

67
		return crc;						\

68
	}								\

69
} while (0)

70


71
#endif /* _CRC_PCLMUL_TEMPLATE_H */

72


73