xref: /linux/arch/riscv/lib/crc-clmul-template.h (revision 4f9786035f9e519db41375818e1d0b5f20da2f10)
1 /* SPDX-License-Identifier: GPL-2.0-or-later */
2 /* Copyright 2025 Google LLC */
3 
4 /*
5  * This file is a "template" that generates a CRC function optimized using the
6  * RISC-V Zbc (scalar carryless multiplication) extension.  The includer of this
7  * file must define the following parameters to specify the type of CRC:
8  *
9  *	crc_t: the data type of the CRC, e.g. u32 for a 32-bit CRC
10  *	LSB_CRC: 0 for a msb (most-significant-bit) first CRC, i.e. natural
11  *		 mapping between bits and polynomial coefficients
12  *	         1 for a lsb (least-significant-bit) first CRC, i.e. reflected
13  *	         mapping between bits and polynomial coefficients
14  */
15 
16 #include <asm/byteorder.h>
17 #include <linux/minmax.h>
18 
19 #define CRC_BITS	(8 * sizeof(crc_t))	/* a.k.a. 'n' */
20 
21 static inline unsigned long clmul(unsigned long a, unsigned long b)
22 {
23 	unsigned long res;
24 
25 	asm(".option push\n"
26 	    ".option arch,+zbc\n"
27 	    "clmul %0, %1, %2\n"
28 	    ".option pop\n"
29 	    : "=r" (res) : "r" (a), "r" (b));
30 	return res;
31 }
32 
33 static inline unsigned long clmulh(unsigned long a, unsigned long b)
34 {
35 	unsigned long res;
36 
37 	asm(".option push\n"
38 	    ".option arch,+zbc\n"
39 	    "clmulh %0, %1, %2\n"
40 	    ".option pop\n"
41 	    : "=r" (res) : "r" (a), "r" (b));
42 	return res;
43 }
44 
45 static inline unsigned long clmulr(unsigned long a, unsigned long b)
46 {
47 	unsigned long res;
48 
49 	asm(".option push\n"
50 	    ".option arch,+zbc\n"
51 	    "clmulr %0, %1, %2\n"
52 	    ".option pop\n"
53 	    : "=r" (res) : "r" (a), "r" (b));
54 	return res;
55 }
56 
57 /*
58  * crc_load_long() loads one "unsigned long" of aligned data bytes, producing a
59  * polynomial whose bit order matches the CRC's bit order.
60  */
61 #ifdef CONFIG_64BIT
62 #  if LSB_CRC
63 #    define crc_load_long(x)	le64_to_cpup(x)
64 #  else
65 #    define crc_load_long(x)	be64_to_cpup(x)
66 #  endif
67 #else
68 #  if LSB_CRC
69 #    define crc_load_long(x)	le32_to_cpup(x)
70 #  else
71 #    define crc_load_long(x)	be32_to_cpup(x)
72 #  endif
73 #endif
74 
75 /* XOR @crc into the end of @msgpoly that represents the high-order terms. */
76 static inline unsigned long
77 crc_clmul_prep(crc_t crc, unsigned long msgpoly)
78 {
79 #if LSB_CRC
80 	return msgpoly ^ crc;
81 #else
82 	return msgpoly ^ ((unsigned long)crc << (BITS_PER_LONG - CRC_BITS));
83 #endif
84 }
85 
86 /*
87  * Multiply the long-sized @msgpoly by x^n (a.k.a. x^CRC_BITS) and reduce it
88  * modulo the generator polynomial G.  This gives the CRC of @msgpoly.
89  */
90 static inline crc_t
91 crc_clmul_long(unsigned long msgpoly, const struct crc_clmul_consts *consts)
92 {
93 	unsigned long tmp;
94 
95 	/*
96 	 * First step of Barrett reduction with integrated multiplication by
97 	 * x^n: calculate floor((msgpoly * x^n) / G).  This is the value by
98 	 * which G needs to be multiplied to cancel out the x^n and higher terms
99 	 * of msgpoly * x^n.  Do it using the following formula:
100 	 *
101 	 * msb-first:
102 	 *    floor((msgpoly * floor(x^(BITS_PER_LONG-1+n) / G)) / x^(BITS_PER_LONG-1))
103 	 * lsb-first:
104 	 *    floor((msgpoly * floor(x^(BITS_PER_LONG-1+n) / G) * x) / x^BITS_PER_LONG)
105 	 *
106 	 * barrett_reduction_const_1 contains floor(x^(BITS_PER_LONG-1+n) / G),
107 	 * which fits a long exactly.  Using any lower power of x there would
108 	 * not carry enough precision through the calculation, while using any
109 	 * higher power of x would require extra instructions to handle a wider
110 	 * multiplication.  In the msb-first case, using this power of x results
111 	 * in needing a floored division by x^(BITS_PER_LONG-1), which matches
112 	 * what clmulr produces.  In the lsb-first case, a factor of x gets
113 	 * implicitly introduced by each carryless multiplication (shown as
114 	 * '* x' above), and the floored division instead needs to be by
115 	 * x^BITS_PER_LONG which matches what clmul produces.
116 	 */
117 #if LSB_CRC
118 	tmp = clmul(msgpoly, consts->barrett_reduction_const_1);
119 #else
120 	tmp = clmulr(msgpoly, consts->barrett_reduction_const_1);
121 #endif
122 
123 	/*
124 	 * Second step of Barrett reduction:
125 	 *
126 	 *    crc := (msgpoly * x^n) + (G * floor((msgpoly * x^n) / G))
127 	 *
128 	 * This reduces (msgpoly * x^n) modulo G by adding the appropriate
129 	 * multiple of G to it.  The result uses only the x^0..x^(n-1) terms.
130 	 * HOWEVER, since the unreduced value (msgpoly * x^n) is zero in those
131 	 * terms in the first place, it is more efficient to do the equivalent:
132 	 *
133 	 *    crc := ((G - x^n) * floor((msgpoly * x^n) / G)) mod x^n
134 	 *
135 	 * In the lsb-first case further modify it to the following which avoids
136 	 * a shift, as the crc ends up in the physically low n bits from clmulr:
137 	 *
138 	 *    product := ((G - x^n) * x^(BITS_PER_LONG - n)) * floor((msgpoly * x^n) / G) * x
139 	 *    crc := floor(product / x^(BITS_PER_LONG + 1 - n)) mod x^n
140 	 *
141 	 * barrett_reduction_const_2 contains the constant multiplier (G - x^n)
142 	 * or (G - x^n) * x^(BITS_PER_LONG - n) from the formulas above.  The
143 	 * cast of the result to crc_t is essential, as it applies the mod x^n!
144 	 */
145 #if LSB_CRC
146 	return clmulr(tmp, consts->barrett_reduction_const_2);
147 #else
148 	return clmul(tmp, consts->barrett_reduction_const_2);
149 #endif
150 }
151 
152 /* Update @crc with the data from @msgpoly. */
153 static inline crc_t
154 crc_clmul_update_long(crc_t crc, unsigned long msgpoly,
155 		      const struct crc_clmul_consts *consts)
156 {
157 	return crc_clmul_long(crc_clmul_prep(crc, msgpoly), consts);
158 }
159 
160 /* Update @crc with 1 <= @len < sizeof(unsigned long) bytes of data. */
161 static inline crc_t
162 crc_clmul_update_partial(crc_t crc, const u8 *p, size_t len,
163 			 const struct crc_clmul_consts *consts)
164 {
165 	unsigned long msgpoly;
166 	size_t i;
167 
168 #if LSB_CRC
169 	msgpoly = (unsigned long)p[0] << (BITS_PER_LONG - 8);
170 	for (i = 1; i < len; i++)
171 		msgpoly = (msgpoly >> 8) ^ ((unsigned long)p[i] << (BITS_PER_LONG - 8));
172 #else
173 	msgpoly = p[0];
174 	for (i = 1; i < len; i++)
175 		msgpoly = (msgpoly << 8) ^ p[i];
176 #endif
177 
178 	if (len >= sizeof(crc_t)) {
179 	#if LSB_CRC
180 		msgpoly ^= (unsigned long)crc << (BITS_PER_LONG - 8*len);
181 	#else
182 		msgpoly ^= (unsigned long)crc << (8*len - CRC_BITS);
183 	#endif
184 		return crc_clmul_long(msgpoly, consts);
185 	}
186 #if LSB_CRC
187 	msgpoly ^= (unsigned long)crc << (BITS_PER_LONG - 8*len);
188 	return crc_clmul_long(msgpoly, consts) ^ (crc >> (8*len));
189 #else
190 	msgpoly ^= crc >> (CRC_BITS - 8*len);
191 	return crc_clmul_long(msgpoly, consts) ^ (crc << (8*len));
192 #endif
193 }
194 
195 static inline crc_t
196 crc_clmul(crc_t crc, const void *p, size_t len,
197 	  const struct crc_clmul_consts *consts)
198 {
199 	size_t align;
200 
201 	/* This implementation assumes that the CRC fits in an unsigned long. */
202 	BUILD_BUG_ON(sizeof(crc_t) > sizeof(unsigned long));
203 
204 	/* If the buffer is not long-aligned, align it. */
205 	align = (unsigned long)p % sizeof(unsigned long);
206 	if (align && len) {
207 		align = min(sizeof(unsigned long) - align, len);
208 		crc = crc_clmul_update_partial(crc, p, align, consts);
209 		p += align;
210 		len -= align;
211 	}
212 
213 	if (len >= 4 * sizeof(unsigned long)) {
214 		unsigned long m0, m1;
215 
216 		m0 = crc_clmul_prep(crc, crc_load_long(p));
217 		m1 = crc_load_long(p + sizeof(unsigned long));
218 		p += 2 * sizeof(unsigned long);
219 		len -= 2 * sizeof(unsigned long);
220 		/*
221 		 * Main loop.  Each iteration starts with a message polynomial
222 		 * (x^BITS_PER_LONG)*m0 + m1, then logically extends it by two
223 		 * more longs of data to form x^(3*BITS_PER_LONG)*m0 +
224 		 * x^(2*BITS_PER_LONG)*m1 + x^BITS_PER_LONG*m2 + m3, then
225 		 * "folds" that back into a congruent (modulo G) value that uses
226 		 * just m0 and m1 again.  This is done by multiplying m0 by the
227 		 * precomputed constant (x^(3*BITS_PER_LONG) mod G) and m1 by
228 		 * the precomputed constant (x^(2*BITS_PER_LONG) mod G), then
229 		 * adding the results to m2 and m3 as appropriate.  Each such
230 		 * multiplication produces a result twice the length of a long,
231 		 * which in RISC-V is two instructions clmul and clmulh.
232 		 *
233 		 * This could be changed to fold across more than 2 longs at a
234 		 * time if there is a CPU that can take advantage of it.
235 		 */
236 		do {
237 			unsigned long p0, p1, p2, p3;
238 
239 			p0 = clmulh(m0, consts->fold_across_2_longs_const_hi);
240 			p1 = clmul(m0, consts->fold_across_2_longs_const_hi);
241 			p2 = clmulh(m1, consts->fold_across_2_longs_const_lo);
242 			p3 = clmul(m1, consts->fold_across_2_longs_const_lo);
243 			m0 = (LSB_CRC ? p1 ^ p3 : p0 ^ p2) ^ crc_load_long(p);
244 			m1 = (LSB_CRC ? p0 ^ p2 : p1 ^ p3) ^
245 			     crc_load_long(p + sizeof(unsigned long));
246 
247 			p += 2 * sizeof(unsigned long);
248 			len -= 2 * sizeof(unsigned long);
249 		} while (len >= 2 * sizeof(unsigned long));
250 
251 		crc = crc_clmul_long(m0, consts);
252 		crc = crc_clmul_update_long(crc, m1, consts);
253 	}
254 
255 	while (len >= sizeof(unsigned long)) {
256 		crc = crc_clmul_update_long(crc, crc_load_long(p), consts);
257 		p += sizeof(unsigned long);
258 		len -= sizeof(unsigned long);
259 	}
260 
261 	if (len)
262 		crc = crc_clmul_update_partial(crc, p, len, consts);
263 
264 	return crc;
265 }
266