1// SPDX-License-Identifier: GPL-2.0
2/*
3 * arch/x86_64/lib/csum-partial.c
4 *
5 * This file contains network checksum routines that are better done
6 * in an architecture-specific manner due to speed.
7 */
8
9#include <linux/compiler.h>
10#include <linux/export.h>
11#include <asm/checksum.h>
12#include <asm/word-at-a-time.h>
13
14static inline __wsum csum_finalize_sum(u64 temp64)
15{
16	return (__force __wsum)((temp64 + ror64(temp64, 32)) >> 32);
17}
18
19static inline unsigned long update_csum_40b(unsigned long sum, const unsigned long m[5])
20{
21	asm("addq %1,%0\n\t"
22	     "adcq %2,%0\n\t"
23	     "adcq %3,%0\n\t"
24	     "adcq %4,%0\n\t"
25	     "adcq %5,%0\n\t"
26	     "adcq $0,%0"
27		:"+r" (sum)
28		:"m" (m[0]), "m" (m[1]), "m" (m[2]),
29		 "m" (m[3]), "m" (m[4]));
30	return sum;
31}
32
33/*
34 * Do a checksum on an arbitrary memory area.
35 * Returns a 32bit checksum.
36 *
37 * This isn't as time critical as it used to be because many NICs
38 * do hardware checksumming these days.
39 *
40 * Still, with CHECKSUM_COMPLETE this is called to compute
41 * checksums on IPv6 headers (40 bytes) and other small parts.
42 * it's best to have buff aligned on a 64-bit boundary
43 */
44__wsum csum_partial(const void *buff, int len, __wsum sum)
45{
46	u64 temp64 = (__force u64)sum;
47
48	/* Do two 40-byte chunks in parallel to get better ILP */
49	if (likely(len >= 80)) {
50		u64 temp64_2 = 0;
51		do {
52			temp64 = update_csum_40b(temp64, buff);
53			temp64_2 = update_csum_40b(temp64_2, buff + 40);
54			buff += 80;
55			len -= 80;
56		} while (len >= 80);
57
58		asm("addq %1,%0\n\t"
59		    "adcq $0,%0"
60		    :"+r" (temp64): "r" (temp64_2));
61	}
62
63	/*
64	 * len == 40 is the hot case due to IPv6 headers, so return
65	 * early for that exact case without checking the tail bytes.
66	 */
67	if (len >= 40) {
68		temp64 = update_csum_40b(temp64, buff);
69		len -= 40;
70		if (!len)
71			return csum_finalize_sum(temp64);
72		buff += 40;
73	}
74
75	if (len & 32) {
76		asm("addq 0*8(%[src]),%[res]\n\t"
77		    "adcq 1*8(%[src]),%[res]\n\t"
78		    "adcq 2*8(%[src]),%[res]\n\t"
79		    "adcq 3*8(%[src]),%[res]\n\t"
80		    "adcq $0,%[res]"
81		    : [res] "+r"(temp64)
82		    : [src] "r"(buff), "m"(*(const char(*)[32])buff));
83		buff += 32;
84	}
85	if (len & 16) {
86		asm("addq 0*8(%[src]),%[res]\n\t"
87		    "adcq 1*8(%[src]),%[res]\n\t"
88		    "adcq $0,%[res]"
89		    : [res] "+r"(temp64)
90		    : [src] "r"(buff), "m"(*(const char(*)[16])buff));
91		buff += 16;
92	}
93	if (len & 8) {
94		asm("addq 0*8(%[src]),%[res]\n\t"
95		    "adcq $0,%[res]"
96		    : [res] "+r"(temp64)
97		    : [src] "r"(buff), "m"(*(const char(*)[8])buff));
98		buff += 8;
99	}
100	if (len & 7) {
101		unsigned int shift = (-len << 3) & 63;
102		unsigned long trail;
103
104		trail = (load_unaligned_zeropad(buff) << shift) >> shift;
105
106		asm("addq %[trail],%[res]\n\t"
107		    "adcq $0,%[res]"
108		    : [res] "+r"(temp64)
109		    : [trail] "r"(trail));
110	}
111	return csum_finalize_sum(temp64);
112}
113EXPORT_SYMBOL(csum_partial);
114
115/*
116 * this routine is used for miscellaneous IP-like checksums, mainly
117 * in icmp.c
118 */
119__sum16 ip_compute_csum(const void *buff, int len)
120{
121	return csum_fold(csum_partial(buff, len, 0));
122}
123EXPORT_SYMBOL(ip_compute_csum);
124