xref: /linux/arch/arm64/lib/csum.c (revision ae99fb8baafc881b35aa0b79d7ac0178a7c40c89)
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 /* Looks dumb, but generates nice-ish code */
11 static u64 accumulate(u64 sum, u64 data)
12 {
13 	__uint128_t tmp = (__uint128_t)sum + data;
14 	return tmp + (tmp >> 64);
15 }
16 
17 unsigned int do_csum(const unsigned char *buff, int len)
18 {
19 	unsigned int offset, shift, sum;
20 	const u64 *ptr;
21 	u64 data, sum64 = 0;
22 
23 	if (unlikely(len == 0))
24 		return 0;
25 
26 	offset = (unsigned long)buff & 7;
27 	/*
28 	 * This is to all intents and purposes safe, since rounding down cannot
29 	 * result in a different page or cache line being accessed, and @buff
30 	 * should absolutely not be pointing to anything read-sensitive. We do,
31 	 * however, have to be careful not to piss off KASAN, which means using
32 	 * unchecked reads to accommodate the head and tail, for which we'll
33 	 * compensate with an explicit check up-front.
34 	 */
35 	kasan_check_read(buff, len);
36 	ptr = (u64 *)(buff - offset);
37 	len = len + offset - 8;
38 
39 	/*
40 	 * Head: zero out any excess leading bytes. Shifting back by the same
41 	 * amount should be at least as fast as any other way of handling the
42 	 * odd/even alignment, and means we can ignore it until the very end.
43 	 */
44 	shift = offset * 8;
45 	data = READ_ONCE_NOCHECK(*ptr++);
46 #ifdef __LITTLE_ENDIAN
47 	data = (data >> shift) << shift;
48 #else
49 	data = (data << shift) >> shift;
50 #endif
51 
52 	/*
53 	 * Body: straightforward aligned loads from here on (the paired loads
54 	 * underlying the quadword type still only need dword alignment). The
55 	 * main loop strictly excludes the tail, so the second loop will always
56 	 * run at least once.
57 	 */
58 	while (unlikely(len > 64)) {
59 		__uint128_t tmp1, tmp2, tmp3, tmp4;
60 
61 		tmp1 = READ_ONCE_NOCHECK(*(__uint128_t *)ptr);
62 		tmp2 = READ_ONCE_NOCHECK(*(__uint128_t *)(ptr + 2));
63 		tmp3 = READ_ONCE_NOCHECK(*(__uint128_t *)(ptr + 4));
64 		tmp4 = READ_ONCE_NOCHECK(*(__uint128_t *)(ptr + 6));
65 
66 		len -= 64;
67 		ptr += 8;
68 
69 		/* This is the "don't dump the carry flag into a GPR" idiom */
70 		tmp1 += (tmp1 >> 64) | (tmp1 << 64);
71 		tmp2 += (tmp2 >> 64) | (tmp2 << 64);
72 		tmp3 += (tmp3 >> 64) | (tmp3 << 64);
73 		tmp4 += (tmp4 >> 64) | (tmp4 << 64);
74 		tmp1 = ((tmp1 >> 64) << 64) | (tmp2 >> 64);
75 		tmp1 += (tmp1 >> 64) | (tmp1 << 64);
76 		tmp3 = ((tmp3 >> 64) << 64) | (tmp4 >> 64);
77 		tmp3 += (tmp3 >> 64) | (tmp3 << 64);
78 		tmp1 = ((tmp1 >> 64) << 64) | (tmp3 >> 64);
79 		tmp1 += (tmp1 >> 64) | (tmp1 << 64);
80 		tmp1 = ((tmp1 >> 64) << 64) | sum64;
81 		tmp1 += (tmp1 >> 64) | (tmp1 << 64);
82 		sum64 = tmp1 >> 64;
83 	}
84 	while (len > 8) {
85 		__uint128_t tmp;
86 
87 		sum64 = accumulate(sum64, data);
88 		tmp = READ_ONCE_NOCHECK(*(__uint128_t *)ptr);
89 
90 		len -= 16;
91 		ptr += 2;
92 
93 #ifdef __LITTLE_ENDIAN
94 		data = tmp >> 64;
95 		sum64 = accumulate(sum64, tmp);
96 #else
97 		data = tmp;
98 		sum64 = accumulate(sum64, tmp >> 64);
99 #endif
100 	}
101 	if (len > 0) {
102 		sum64 = accumulate(sum64, data);
103 		data = READ_ONCE_NOCHECK(*ptr);
104 		len -= 8;
105 	}
106 	/*
107 	 * Tail: zero any over-read bytes similarly to the head, again
108 	 * preserving odd/even alignment.
109 	 */
110 	shift = len * -8;
111 #ifdef __LITTLE_ENDIAN
112 	data = (data << shift) >> shift;
113 #else
114 	data = (data >> shift) << shift;
115 #endif
116 	sum64 = accumulate(sum64, data);
117 
118 	/* Finally, folding */
119 	sum64 += (sum64 >> 32) | (sum64 << 32);
120 	sum = sum64 >> 32;
121 	sum += (sum >> 16) | (sum << 16);
122 	if (offset & 1)
123 		return (u16)swab32(sum);
124 
125 	return sum >> 16;
126 }
127