Loading...
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);
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 unsigned short from32to16(unsigned a)
15{
16 unsigned short b = a >> 16;
17 asm("addw %w2,%w0\n\t"
18 "adcw $0,%w0\n"
19 : "=r" (b)
20 : "0" (b), "r" (a));
21 return b;
22}
23
24/*
25 * Do a checksum on an arbitrary memory area.
26 * Returns a 32bit checksum.
27 *
28 * This isn't as time critical as it used to be because many NICs
29 * do hardware checksumming these days.
30 *
31 * Still, with CHECKSUM_COMPLETE this is called to compute
32 * checksums on IPv6 headers (40 bytes) and other small parts.
33 * it's best to have buff aligned on a 64-bit boundary
34 */
35__wsum csum_partial(const void *buff, int len, __wsum sum)
36{
37 u64 temp64 = (__force u64)sum;
38 unsigned odd, result;
39
40 odd = 1 & (unsigned long) buff;
41 if (unlikely(odd)) {
42 if (unlikely(len == 0))
43 return sum;
44 temp64 = ror32((__force u32)sum, 8);
45 temp64 += (*(unsigned char *)buff << 8);
46 len--;
47 buff++;
48 }
49
50 while (unlikely(len >= 64)) {
51 asm("addq 0*8(%[src]),%[res]\n\t"
52 "adcq 1*8(%[src]),%[res]\n\t"
53 "adcq 2*8(%[src]),%[res]\n\t"
54 "adcq 3*8(%[src]),%[res]\n\t"
55 "adcq 4*8(%[src]),%[res]\n\t"
56 "adcq 5*8(%[src]),%[res]\n\t"
57 "adcq 6*8(%[src]),%[res]\n\t"
58 "adcq 7*8(%[src]),%[res]\n\t"
59 "adcq $0,%[res]"
60 : [res] "+r" (temp64)
61 : [src] "r" (buff)
62 : "memory");
63 buff += 64;
64 len -= 64;
65 }
66
67 if (len & 32) {
68 asm("addq 0*8(%[src]),%[res]\n\t"
69 "adcq 1*8(%[src]),%[res]\n\t"
70 "adcq 2*8(%[src]),%[res]\n\t"
71 "adcq 3*8(%[src]),%[res]\n\t"
72 "adcq $0,%[res]"
73 : [res] "+r" (temp64)
74 : [src] "r" (buff)
75 : "memory");
76 buff += 32;
77 }
78 if (len & 16) {
79 asm("addq 0*8(%[src]),%[res]\n\t"
80 "adcq 1*8(%[src]),%[res]\n\t"
81 "adcq $0,%[res]"
82 : [res] "+r" (temp64)
83 : [src] "r" (buff)
84 : "memory");
85 buff += 16;
86 }
87 if (len & 8) {
88 asm("addq 0*8(%[src]),%[res]\n\t"
89 "adcq $0,%[res]"
90 : [res] "+r" (temp64)
91 : [src] "r" (buff)
92 : "memory");
93 buff += 8;
94 }
95 if (len & 7) {
96 unsigned int shift = (8 - (len & 7)) * 8;
97 unsigned long trail;
98
99 trail = (load_unaligned_zeropad(buff) << shift) >> shift;
100
101 asm("addq %[trail],%[res]\n\t"
102 "adcq $0,%[res]"
103 : [res] "+r" (temp64)
104 : [trail] "r" (trail));
105 }
106 result = add32_with_carry(temp64 >> 32, temp64 & 0xffffffff);
107 if (unlikely(odd)) {
108 result = from32to16(result);
109 result = ((result >> 8) & 0xff) | ((result & 0xff) << 8);
110 }
111 return (__force __wsum)result;
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);