xref: /linux/arch/loongarch/lib/csum.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
1 // SPDX-License-Identifier: GPL-2.0-only
2 // Copyright (C) 2019-2020 Arm Ltd.
3 
4 #include <linux/compiler.h>
5 #include <linux/kasan-checks.h>
6 #include <linux/kernel.h>
7 
8 #include <net/checksum.h>
9 
10 static u64 accumulate(u64 sum, u64 data)
11 {
12 	sum += data;
13 	if (sum < data)
14 		sum += 1;
15 	return sum;
16 }
17 
18 /*
19  * We over-read the buffer and this makes KASAN unhappy. Instead, disable
20  * instrumentation and call kasan explicitly.
21  */
22 unsigned int __no_sanitize_address do_csum(const unsigned char *buff, int len)
23 {
24 	unsigned int offset, shift, sum;
25 	const u64 *ptr;
26 	u64 data, sum64 = 0;
27 
28 	if (unlikely(len == 0))
29 		return 0;
30 
31 	offset = (unsigned long)buff & 7;
32 	/*
33 	 * This is to all intents and purposes safe, since rounding down cannot
34 	 * result in a different page or cache line being accessed, and @buff
35 	 * should absolutely not be pointing to anything read-sensitive. We do,
36 	 * however, have to be careful not to piss off KASAN, which means using
37 	 * unchecked reads to accommodate the head and tail, for which we'll
38 	 * compensate with an explicit check up-front.
39 	 */
40 	kasan_check_read(buff, len);
41 	ptr = (u64 *)(buff - offset);
42 	len = len + offset - 8;
43 
44 	/*
45 	 * Head: zero out any excess leading bytes. Shifting back by the same
46 	 * amount should be at least as fast as any other way of handling the
47 	 * odd/even alignment, and means we can ignore it until the very end.
48 	 */
49 	shift = offset * 8;
50 	data = *ptr++;
51 	data = (data >> shift) << shift;
52 
53 	/*
54 	 * Body: straightforward aligned loads from here on (the paired loads
55 	 * underlying the quadword type still only need dword alignment). The
56 	 * main loop strictly excludes the tail, so the second loop will always
57 	 * run at least once.
58 	 */
59 	while (unlikely(len > 64)) {
60 		__uint128_t tmp1, tmp2, tmp3, tmp4;
61 
62 		tmp1 = *(__uint128_t *)ptr;
63 		tmp2 = *(__uint128_t *)(ptr + 2);
64 		tmp3 = *(__uint128_t *)(ptr + 4);
65 		tmp4 = *(__uint128_t *)(ptr + 6);
66 
67 		len -= 64;
68 		ptr += 8;
69 
70 		/* This is the "don't dump the carry flag into a GPR" idiom */
71 		tmp1 += (tmp1 >> 64) | (tmp1 << 64);
72 		tmp2 += (tmp2 >> 64) | (tmp2 << 64);
73 		tmp3 += (tmp3 >> 64) | (tmp3 << 64);
74 		tmp4 += (tmp4 >> 64) | (tmp4 << 64);
75 		tmp1 = ((tmp1 >> 64) << 64) | (tmp2 >> 64);
76 		tmp1 += (tmp1 >> 64) | (tmp1 << 64);
77 		tmp3 = ((tmp3 >> 64) << 64) | (tmp4 >> 64);
78 		tmp3 += (tmp3 >> 64) | (tmp3 << 64);
79 		tmp1 = ((tmp1 >> 64) << 64) | (tmp3 >> 64);
80 		tmp1 += (tmp1 >> 64) | (tmp1 << 64);
81 		tmp1 = ((tmp1 >> 64) << 64) | sum64;
82 		tmp1 += (tmp1 >> 64) | (tmp1 << 64);
83 		sum64 = tmp1 >> 64;
84 	}
85 	while (len > 8) {
86 		__uint128_t tmp;
87 
88 		sum64 = accumulate(sum64, data);
89 		tmp = *(__uint128_t *)ptr;
90 
91 		len -= 16;
92 		ptr += 2;
93 
94 		data = tmp >> 64;
95 		sum64 = accumulate(sum64, tmp);
96 	}
97 	if (len > 0) {
98 		sum64 = accumulate(sum64, data);
99 		data = *ptr;
100 		len -= 8;
101 	}
102 	/*
103 	 * Tail: zero any over-read bytes similarly to the head, again
104 	 * preserving odd/even alignment.
105 	 */
106 	shift = len * -8;
107 	data = (data << shift) >> shift;
108 	sum64 = accumulate(sum64, data);
109 
110 	/* Finally, folding */
111 	sum64 += (sum64 >> 32) | (sum64 << 32);
112 	sum = sum64 >> 32;
113 	sum += (sum >> 16) | (sum << 16);
114 	if (offset & 1)
115 		return (u16)swab32(sum);
116 
117 	return sum >> 16;
118 }
119 
120 __sum16 csum_ipv6_magic(const struct in6_addr *saddr,
121 			const struct in6_addr *daddr,
122 			__u32 len, __u8 proto, __wsum csum)
123 {
124 	__uint128_t src, dst;
125 	u64 sum = (__force u64)csum;
126 
127 	src = *(const __uint128_t *)saddr->s6_addr;
128 	dst = *(const __uint128_t *)daddr->s6_addr;
129 
130 	sum += (__force u32)htonl(len);
131 	sum += (u32)proto << 24;
132 	src += (src >> 64) | (src << 64);
133 	dst += (dst >> 64) | (dst << 64);
134 
135 	sum = accumulate(sum, src >> 64);
136 	sum = accumulate(sum, dst >> 64);
137 
138 	sum += ((sum >> 32) | (sum << 32));
139 	return csum_fold((__force __wsum)(sum >> 32));
140 }
141 EXPORT_SYMBOL(csum_ipv6_magic);
142