xref: /linux/arch/riscv/lib/csum.c (revision 4f77c3462308c62ffe7129cc18b9ac937f44b5a5)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Checksum library
4  *
5  * Influenced by arch/arm64/lib/csum.c
6  * Copyright (C) 2023-2024 Rivos Inc.
7  */
8 #include <linux/bitops.h>
9 #include <linux/compiler.h>
10 #include <linux/jump_label.h>
11 #include <linux/kasan-checks.h>
12 #include <linux/kernel.h>
13 
14 #include <asm/cpufeature.h>
15 
16 #include <net/checksum.h>
17 
18 /* Default version is sufficient for 32 bit */
19 #ifndef CONFIG_32BIT
20 __sum16 csum_ipv6_magic(const struct in6_addr *saddr,
21 			const struct in6_addr *daddr,
22 			__u32 len, __u8 proto, __wsum csum)
23 {
24 	unsigned int ulen, uproto;
25 	unsigned long sum = (__force unsigned long)csum;
26 
27 	sum += (__force unsigned long)saddr->s6_addr32[0];
28 	sum += (__force unsigned long)saddr->s6_addr32[1];
29 	sum += (__force unsigned long)saddr->s6_addr32[2];
30 	sum += (__force unsigned long)saddr->s6_addr32[3];
31 
32 	sum += (__force unsigned long)daddr->s6_addr32[0];
33 	sum += (__force unsigned long)daddr->s6_addr32[1];
34 	sum += (__force unsigned long)daddr->s6_addr32[2];
35 	sum += (__force unsigned long)daddr->s6_addr32[3];
36 
37 	ulen = (__force unsigned int)htonl((unsigned int)len);
38 	sum += ulen;
39 
40 	uproto = (__force unsigned int)htonl(proto);
41 	sum += uproto;
42 
43 	/*
44 	 * Zbb support saves 4 instructions, so not worth checking without
45 	 * alternatives if supported
46 	 */
47 	if (IS_ENABLED(CONFIG_RISCV_ISA_ZBB) &&
48 	    IS_ENABLED(CONFIG_RISCV_ALTERNATIVE)) {
49 		unsigned long fold_temp;
50 
51 		/*
52 		 * Zbb is likely available when the kernel is compiled with Zbb
53 		 * support, so nop when Zbb is available and jump when Zbb is
54 		 * not available.
55 		 */
56 		asm goto(ALTERNATIVE("j %l[no_zbb]", "nop", 0,
57 					      RISCV_ISA_EXT_ZBB, 1)
58 				  :
59 				  :
60 				  :
61 				  : no_zbb);
62 		asm(".option push					\n\
63 		.option arch,+zbb					\n\
64 			rori	%[fold_temp], %[sum], 32		\n\
65 			add	%[sum], %[fold_temp], %[sum]		\n\
66 			srli	%[sum], %[sum], 32			\n\
67 			not	%[fold_temp], %[sum]			\n\
68 			roriw	%[sum], %[sum], 16			\n\
69 			subw	%[sum], %[fold_temp], %[sum]		\n\
70 		.option pop"
71 		: [sum] "+r" (sum), [fold_temp] "=&r" (fold_temp));
72 		return (__force __sum16)(sum >> 16);
73 	}
74 no_zbb:
75 	sum += ror64(sum, 32);
76 	sum >>= 32;
77 	return csum_fold((__force __wsum)sum);
78 }
79 EXPORT_SYMBOL(csum_ipv6_magic);
80 #endif /* !CONFIG_32BIT */
81 
82 #ifdef CONFIG_32BIT
83 #define OFFSET_MASK 3
84 #elif CONFIG_64BIT
85 #define OFFSET_MASK 7
86 #endif
87 
88 static inline __no_sanitize_address unsigned long
89 do_csum_common(const unsigned long *ptr, const unsigned long *end,
90 	       unsigned long data)
91 {
92 	unsigned int shift;
93 	unsigned long csum = 0, carry = 0;
94 
95 	/*
96 	 * Do 32-bit reads on RV32 and 64-bit reads otherwise. This should be
97 	 * faster than doing 32-bit reads on architectures that support larger
98 	 * reads.
99 	 */
100 	while (ptr < end) {
101 		csum += data;
102 		carry += csum < data;
103 		data = *(ptr++);
104 	}
105 
106 	/*
107 	 * Perform alignment (and over-read) bytes on the tail if any bytes
108 	 * leftover.
109 	 */
110 	shift = ((long)ptr - (long)end) * 8;
111 #ifdef __LITTLE_ENDIAN
112 	data = (data << shift) >> shift;
113 #else
114 	data = (data >> shift) << shift;
115 #endif
116 	csum += data;
117 	carry += csum < data;
118 	csum += carry;
119 	csum += csum < carry;
120 
121 	return csum;
122 }
123 
124 /*
125  * Algorithm accounts for buff being misaligned.
126  * If buff is not aligned, will over-read bytes but not use the bytes that it
127  * shouldn't. The same thing will occur on the tail-end of the read.
128  */
129 static inline __no_sanitize_address unsigned int
130 do_csum_with_alignment(const unsigned char *buff, int len)
131 {
132 	unsigned int offset, shift;
133 	unsigned long csum, data;
134 	const unsigned long *ptr, *end;
135 
136 	/*
137 	 * Align address to closest word (double word on rv64) that comes before
138 	 * buff. This should always be in the same page and cache line.
139 	 * Directly call KASAN with the alignment we will be using.
140 	 */
141 	offset = (unsigned long)buff & OFFSET_MASK;
142 	kasan_check_read(buff, len);
143 	ptr = (const unsigned long *)(buff - offset);
144 
145 	/*
146 	 * Clear the most significant bytes that were over-read if buff was not
147 	 * aligned.
148 	 */
149 	shift = offset * 8;
150 	data = *(ptr++);
151 #ifdef __LITTLE_ENDIAN
152 	data = (data >> shift) << shift;
153 #else
154 	data = (data << shift) >> shift;
155 #endif
156 	end = (const unsigned long *)(buff + len);
157 	csum = do_csum_common(ptr, end, data);
158 
159 #ifdef CC_HAS_ASM_GOTO_TIED_OUTPUT
160 	/*
161 	 * Zbb support saves 6 instructions, so not worth checking without
162 	 * alternatives if supported
163 	 */
164 	if (IS_ENABLED(CONFIG_RISCV_ISA_ZBB) &&
165 	    IS_ENABLED(CONFIG_RISCV_ALTERNATIVE)) {
166 		unsigned long fold_temp;
167 
168 		/*
169 		 * Zbb is likely available when the kernel is compiled with Zbb
170 		 * support, so nop when Zbb is available and jump when Zbb is
171 		 * not available.
172 		 */
173 		asm goto(ALTERNATIVE("j %l[no_zbb]", "nop", 0,
174 					      RISCV_ISA_EXT_ZBB, 1)
175 				  :
176 				  :
177 				  :
178 				  : no_zbb);
179 
180 #ifdef CONFIG_32BIT
181 		asm_goto_output(".option push			\n\
182 		.option arch,+zbb				\n\
183 			rori	%[fold_temp], %[csum], 16	\n\
184 			andi	%[offset], %[offset], 1		\n\
185 			add	%[csum], %[fold_temp], %[csum]	\n\
186 			beq	%[offset], zero, %l[end]	\n\
187 			rev8	%[csum], %[csum]		\n\
188 		.option pop"
189 			: [csum] "+r" (csum), [fold_temp] "=&r" (fold_temp)
190 			: [offset] "r" (offset)
191 			:
192 			: end);
193 
194 		return (unsigned short)csum;
195 #else /* !CONFIG_32BIT */
196 		asm_goto_output(".option push			\n\
197 		.option arch,+zbb				\n\
198 			rori	%[fold_temp], %[csum], 32	\n\
199 			add	%[csum], %[fold_temp], %[csum]	\n\
200 			srli	%[csum], %[csum], 32		\n\
201 			roriw	%[fold_temp], %[csum], 16	\n\
202 			addw	%[csum], %[fold_temp], %[csum]	\n\
203 			andi	%[offset], %[offset], 1		\n\
204 			beq	%[offset], zero, %l[end]	\n\
205 			rev8	%[csum], %[csum]		\n\
206 		.option pop"
207 			: [csum] "+r" (csum), [fold_temp] "=&r" (fold_temp)
208 			: [offset] "r" (offset)
209 			:
210 			: end);
211 
212 		return (csum << 16) >> 48;
213 #endif /* !CONFIG_32BIT */
214 end:
215 		return csum >> 16;
216 	}
217 no_zbb:
218 #endif /* CC_HAS_ASM_GOTO_TIED_OUTPUT */
219 #ifndef CONFIG_32BIT
220 	csum += ror64(csum, 32);
221 	csum >>= 32;
222 #endif
223 	csum = (u32)csum + ror32((u32)csum, 16);
224 	if (offset & 1)
225 		return (u16)swab32(csum);
226 	return csum >> 16;
227 }
228 
229 /*
230  * Does not perform alignment, should only be used if machine has fast
231  * misaligned accesses, or when buff is known to be aligned.
232  */
233 static inline __no_sanitize_address unsigned int
234 do_csum_no_alignment(const unsigned char *buff, int len)
235 {
236 	unsigned long csum, data;
237 	const unsigned long *ptr, *end;
238 
239 	ptr = (const unsigned long *)(buff);
240 	data = *(ptr++);
241 
242 	kasan_check_read(buff, len);
243 
244 	end = (const unsigned long *)(buff + len);
245 	csum = do_csum_common(ptr, end, data);
246 
247 	/*
248 	 * Zbb support saves 6 instructions, so not worth checking without
249 	 * alternatives if supported
250 	 */
251 	if (IS_ENABLED(CONFIG_RISCV_ISA_ZBB) &&
252 	    IS_ENABLED(CONFIG_RISCV_ALTERNATIVE)) {
253 		unsigned long fold_temp;
254 
255 		/*
256 		 * Zbb is likely available when the kernel is compiled with Zbb
257 		 * support, so nop when Zbb is available and jump when Zbb is
258 		 * not available.
259 		 */
260 		asm goto(ALTERNATIVE("j %l[no_zbb]", "nop", 0,
261 					      RISCV_ISA_EXT_ZBB, 1)
262 				  :
263 				  :
264 				  :
265 				  : no_zbb);
266 
267 #ifdef CONFIG_32BIT
268 		asm (".option push				\n\
269 		.option arch,+zbb				\n\
270 			rori	%[fold_temp], %[csum], 16	\n\
271 			add	%[csum], %[fold_temp], %[csum]	\n\
272 		.option pop"
273 			: [csum] "+r" (csum), [fold_temp] "=&r" (fold_temp)
274 			:
275 			: );
276 
277 #else /* !CONFIG_32BIT */
278 		asm (".option push				\n\
279 		.option arch,+zbb				\n\
280 			rori	%[fold_temp], %[csum], 32	\n\
281 			add	%[csum], %[fold_temp], %[csum]	\n\
282 			srli	%[csum], %[csum], 32		\n\
283 			roriw	%[fold_temp], %[csum], 16	\n\
284 			addw	%[csum], %[fold_temp], %[csum]	\n\
285 		.option pop"
286 			: [csum] "+r" (csum), [fold_temp] "=&r" (fold_temp)
287 			:
288 			: );
289 #endif /* !CONFIG_32BIT */
290 		return csum >> 16;
291 	}
292 no_zbb:
293 #ifndef CONFIG_32BIT
294 	csum += ror64(csum, 32);
295 	csum >>= 32;
296 #endif
297 	csum = (u32)csum + ror32((u32)csum, 16);
298 	return csum >> 16;
299 }
300 
301 /*
302  * Perform a checksum on an arbitrary memory address.
303  * Will do a light-weight address alignment if buff is misaligned, unless
304  * cpu supports fast misaligned accesses.
305  */
306 unsigned int do_csum(const unsigned char *buff, int len)
307 {
308 	if (unlikely(len <= 0))
309 		return 0;
310 
311 	/*
312 	 * Significant performance gains can be seen by not doing alignment
313 	 * on machines with fast misaligned accesses.
314 	 *
315 	 * There is some duplicate code between the "with_alignment" and
316 	 * "no_alignment" implmentations, but the overlap is too awkward to be
317 	 * able to fit in one function without introducing multiple static
318 	 * branches. The largest chunk of overlap was delegated into the
319 	 * do_csum_common function.
320 	 */
321 	if (has_fast_unaligned_accesses() || (((unsigned long)buff & OFFSET_MASK) == 0))
322 		return do_csum_no_alignment(buff, len);
323 
324 	return do_csum_with_alignment(buff, len);
325 }
326