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/*
2 * arch/x86_64/lib/csum-partial.c
3 *
4 * This file contains network checksum routines that are better done
5 * in an architecture-specific manner due to speed.
6 */
7
8#include <linux/compiler.h>
9#include <linux/module.h>
10#include <asm/checksum.h>
11
12static inline unsigned short from32to16(unsigned a)
13{
14 unsigned short b = a >> 16;
15 asm("addw %w2,%w0\n\t"
16 "adcw $0,%w0\n"
17 : "=r" (b)
18 : "0" (b), "r" (a));
19 return b;
20}
21
22/*
23 * Do a 64-bit checksum on an arbitrary memory area.
24 * Returns a 32bit checksum.
25 *
26 * This isn't as time critical as it used to be because many NICs
27 * do hardware checksumming these days.
28 *
29 * Things tried and found to not make it faster:
30 * Manual Prefetching
31 * Unrolling to an 128 bytes inner loop.
32 * Using interleaving with more registers to break the carry chains.
33 */
34static unsigned do_csum(const unsigned char *buff, unsigned len)
35{
36 unsigned odd, count;
37 unsigned long result = 0;
38
39 if (unlikely(len == 0))
40 return result;
41 odd = 1 & (unsigned long) buff;
42 if (unlikely(odd)) {
43 result = *buff << 8;
44 len--;
45 buff++;
46 }
47 count = len >> 1; /* nr of 16-bit words.. */
48 if (count) {
49 if (2 & (unsigned long) buff) {
50 result += *(unsigned short *)buff;
51 count--;
52 len -= 2;
53 buff += 2;
54 }
55 count >>= 1; /* nr of 32-bit words.. */
56 if (count) {
57 unsigned long zero;
58 unsigned count64;
59 if (4 & (unsigned long) buff) {
60 result += *(unsigned int *) buff;
61 count--;
62 len -= 4;
63 buff += 4;
64 }
65 count >>= 1; /* nr of 64-bit words.. */
66
67 /* main loop using 64byte blocks */
68 zero = 0;
69 count64 = count >> 3;
70 while (count64) {
71 asm("addq 0*8(%[src]),%[res]\n\t"
72 "adcq 1*8(%[src]),%[res]\n\t"
73 "adcq 2*8(%[src]),%[res]\n\t"
74 "adcq 3*8(%[src]),%[res]\n\t"
75 "adcq 4*8(%[src]),%[res]\n\t"
76 "adcq 5*8(%[src]),%[res]\n\t"
77 "adcq 6*8(%[src]),%[res]\n\t"
78 "adcq 7*8(%[src]),%[res]\n\t"
79 "adcq %[zero],%[res]"
80 : [res] "=r" (result)
81 : [src] "r" (buff), [zero] "r" (zero),
82 "[res]" (result));
83 buff += 64;
84 count64--;
85 }
86
87 /* last up to 7 8byte blocks */
88 count %= 8;
89 while (count) {
90 asm("addq %1,%0\n\t"
91 "adcq %2,%0\n"
92 : "=r" (result)
93 : "m" (*(unsigned long *)buff),
94 "r" (zero), "0" (result));
95 --count;
96 buff += 8;
97 }
98 result = add32_with_carry(result>>32,
99 result&0xffffffff);
100
101 if (len & 4) {
102 result += *(unsigned int *) buff;
103 buff += 4;
104 }
105 }
106 if (len & 2) {
107 result += *(unsigned short *) buff;
108 buff += 2;
109 }
110 }
111 if (len & 1)
112 result += *buff;
113 result = add32_with_carry(result>>32, result & 0xffffffff);
114 if (unlikely(odd)) {
115 result = from32to16(result);
116 result = ((result >> 8) & 0xff) | ((result & 0xff) << 8);
117 }
118 return result;
119}
120
121/*
122 * computes the checksum of a memory block at buff, length len,
123 * and adds in "sum" (32-bit)
124 *
125 * returns a 32-bit number suitable for feeding into itself
126 * or csum_tcpudp_magic
127 *
128 * this function must be called with even lengths, except
129 * for the last fragment, which may be odd
130 *
131 * it's best to have buff aligned on a 64-bit boundary
132 */
133__wsum csum_partial(const void *buff, int len, __wsum sum)
134{
135 return (__force __wsum)add32_with_carry(do_csum(buff, len),
136 (__force u32)sum);
137}
138
139/*
140 * this routine is used for miscellaneous IP-like checksums, mainly
141 * in icmp.c
142 */
143__sum16 ip_compute_csum(const void *buff, int len)
144{
145 return csum_fold(csum_partial(buff,len,0));
146}
147EXPORT_SYMBOL(ip_compute_csum);
148