xref: /linux/lib/raid6/recov_avx2.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2012 Intel Corporation
4  * Author: Jim Kukunas <james.t.kukunas@linux.intel.com>
5  */
6 
7 #include <linux/raid/pq.h>
8 #include "x86.h"
9 
10 static int raid6_has_avx2(void)
11 {
12 	return boot_cpu_has(X86_FEATURE_AVX2) &&
13 		boot_cpu_has(X86_FEATURE_AVX);
14 }
15 
16 static void raid6_2data_recov_avx2(int disks, size_t bytes, int faila,
17 		int failb, void **ptrs)
18 {
19 	u8 *p, *q, *dp, *dq;
20 	const u8 *pbmul;	/* P multiplier table for B data */
21 	const u8 *qmul;		/* Q multiplier table (for both) */
22 	const u8 x0f = 0x0f;
23 
24 	p = (u8 *)ptrs[disks-2];
25 	q = (u8 *)ptrs[disks-1];
26 
27 	/* Compute syndrome with zero for the missing data pages
28 	   Use the dead data pages as temporary storage for
29 	   delta p and delta q */
30 	dp = (u8 *)ptrs[faila];
31 	ptrs[faila] = (void *)raid6_empty_zero_page;
32 	ptrs[disks-2] = dp;
33 	dq = (u8 *)ptrs[failb];
34 	ptrs[failb] = (void *)raid6_empty_zero_page;
35 	ptrs[disks-1] = dq;
36 
37 	raid6_call.gen_syndrome(disks, bytes, ptrs);
38 
39 	/* Restore pointer table */
40 	ptrs[faila]   = dp;
41 	ptrs[failb]   = dq;
42 	ptrs[disks-2] = p;
43 	ptrs[disks-1] = q;
44 
45 	/* Now, pick the proper data tables */
46 	pbmul = raid6_vgfmul[raid6_gfexi[failb-faila]];
47 	qmul  = raid6_vgfmul[raid6_gfinv[raid6_gfexp[faila] ^
48 		raid6_gfexp[failb]]];
49 
50 	kernel_fpu_begin();
51 
52 	/* ymm0 = x0f[16] */
53 	asm volatile("vpbroadcastb %0, %%ymm7" : : "m" (x0f));
54 
55 	while (bytes) {
56 #ifdef CONFIG_X86_64
57 		asm volatile("vmovdqa %0, %%ymm1" : : "m" (q[0]));
58 		asm volatile("vmovdqa %0, %%ymm9" : : "m" (q[32]));
59 		asm volatile("vmovdqa %0, %%ymm0" : : "m" (p[0]));
60 		asm volatile("vmovdqa %0, %%ymm8" : : "m" (p[32]));
61 		asm volatile("vpxor %0, %%ymm1, %%ymm1" : : "m" (dq[0]));
62 		asm volatile("vpxor %0, %%ymm9, %%ymm9" : : "m" (dq[32]));
63 		asm volatile("vpxor %0, %%ymm0, %%ymm0" : : "m" (dp[0]));
64 		asm volatile("vpxor %0, %%ymm8, %%ymm8" : : "m" (dp[32]));
65 
66 		/*
67 		 * 1 = dq[0]  ^ q[0]
68 		 * 9 = dq[32] ^ q[32]
69 		 * 0 = dp[0]  ^ p[0]
70 		 * 8 = dp[32] ^ p[32]
71 		 */
72 
73 		asm volatile("vbroadcasti128 %0, %%ymm4" : : "m" (qmul[0]));
74 		asm volatile("vbroadcasti128 %0, %%ymm5" : : "m" (qmul[16]));
75 
76 		asm volatile("vpsraw $4, %ymm1, %ymm3");
77 		asm volatile("vpsraw $4, %ymm9, %ymm12");
78 		asm volatile("vpand %ymm7, %ymm1, %ymm1");
79 		asm volatile("vpand %ymm7, %ymm9, %ymm9");
80 		asm volatile("vpand %ymm7, %ymm3, %ymm3");
81 		asm volatile("vpand %ymm7, %ymm12, %ymm12");
82 		asm volatile("vpshufb %ymm9, %ymm4, %ymm14");
83 		asm volatile("vpshufb %ymm1, %ymm4, %ymm4");
84 		asm volatile("vpshufb %ymm12, %ymm5, %ymm15");
85 		asm volatile("vpshufb %ymm3, %ymm5, %ymm5");
86 		asm volatile("vpxor %ymm14, %ymm15, %ymm15");
87 		asm volatile("vpxor %ymm4, %ymm5, %ymm5");
88 
89 		/*
90 		 * 5 = qx[0]
91 		 * 15 = qx[32]
92 		 */
93 
94 		asm volatile("vbroadcasti128 %0, %%ymm4" : : "m" (pbmul[0]));
95 		asm volatile("vbroadcasti128 %0, %%ymm1" : : "m" (pbmul[16]));
96 		asm volatile("vpsraw $4, %ymm0, %ymm2");
97 		asm volatile("vpsraw $4, %ymm8, %ymm6");
98 		asm volatile("vpand %ymm7, %ymm0, %ymm3");
99 		asm volatile("vpand %ymm7, %ymm8, %ymm14");
100 		asm volatile("vpand %ymm7, %ymm2, %ymm2");
101 		asm volatile("vpand %ymm7, %ymm6, %ymm6");
102 		asm volatile("vpshufb %ymm14, %ymm4, %ymm12");
103 		asm volatile("vpshufb %ymm3, %ymm4, %ymm4");
104 		asm volatile("vpshufb %ymm6, %ymm1, %ymm13");
105 		asm volatile("vpshufb %ymm2, %ymm1, %ymm1");
106 		asm volatile("vpxor %ymm4, %ymm1, %ymm1");
107 		asm volatile("vpxor %ymm12, %ymm13, %ymm13");
108 
109 		/*
110 		 * 1  = pbmul[px[0]]
111 		 * 13 = pbmul[px[32]]
112 		 */
113 		asm volatile("vpxor %ymm5, %ymm1, %ymm1");
114 		asm volatile("vpxor %ymm15, %ymm13, %ymm13");
115 
116 		/*
117 		 * 1 = db = DQ
118 		 * 13 = db[32] = DQ[32]
119 		 */
120 		asm volatile("vmovdqa %%ymm1, %0" : "=m" (dq[0]));
121 		asm volatile("vmovdqa %%ymm13,%0" : "=m" (dq[32]));
122 		asm volatile("vpxor %ymm1, %ymm0, %ymm0");
123 		asm volatile("vpxor %ymm13, %ymm8, %ymm8");
124 
125 		asm volatile("vmovdqa %%ymm0, %0" : "=m" (dp[0]));
126 		asm volatile("vmovdqa %%ymm8, %0" : "=m" (dp[32]));
127 
128 		bytes -= 64;
129 		p += 64;
130 		q += 64;
131 		dp += 64;
132 		dq += 64;
133 #else
134 		asm volatile("vmovdqa %0, %%ymm1" : : "m" (*q));
135 		asm volatile("vmovdqa %0, %%ymm0" : : "m" (*p));
136 		asm volatile("vpxor %0, %%ymm1, %%ymm1" : : "m" (*dq));
137 		asm volatile("vpxor %0, %%ymm0, %%ymm0" : : "m" (*dp));
138 
139 		/* 1 = dq ^ q;  0 = dp ^ p */
140 
141 		asm volatile("vbroadcasti128 %0, %%ymm4" : : "m" (qmul[0]));
142 		asm volatile("vbroadcasti128 %0, %%ymm5" : : "m" (qmul[16]));
143 
144 		/*
145 		 * 1 = dq ^ q
146 		 * 3 = dq ^ p >> 4
147 		 */
148 		asm volatile("vpsraw $4, %ymm1, %ymm3");
149 		asm volatile("vpand %ymm7, %ymm1, %ymm1");
150 		asm volatile("vpand %ymm7, %ymm3, %ymm3");
151 		asm volatile("vpshufb %ymm1, %ymm4, %ymm4");
152 		asm volatile("vpshufb %ymm3, %ymm5, %ymm5");
153 		asm volatile("vpxor %ymm4, %ymm5, %ymm5");
154 
155 		/* 5 = qx */
156 
157 		asm volatile("vbroadcasti128 %0, %%ymm4" : : "m" (pbmul[0]));
158 		asm volatile("vbroadcasti128 %0, %%ymm1" : : "m" (pbmul[16]));
159 
160 		asm volatile("vpsraw $4, %ymm0, %ymm2");
161 		asm volatile("vpand %ymm7, %ymm0, %ymm3");
162 		asm volatile("vpand %ymm7, %ymm2, %ymm2");
163 		asm volatile("vpshufb %ymm3, %ymm4, %ymm4");
164 		asm volatile("vpshufb %ymm2, %ymm1, %ymm1");
165 		asm volatile("vpxor %ymm4, %ymm1, %ymm1");
166 
167 		/* 1 = pbmul[px] */
168 		asm volatile("vpxor %ymm5, %ymm1, %ymm1");
169 		/* 1 = db = DQ */
170 		asm volatile("vmovdqa %%ymm1, %0" : "=m" (dq[0]));
171 
172 		asm volatile("vpxor %ymm1, %ymm0, %ymm0");
173 		asm volatile("vmovdqa %%ymm0, %0" : "=m" (dp[0]));
174 
175 		bytes -= 32;
176 		p += 32;
177 		q += 32;
178 		dp += 32;
179 		dq += 32;
180 #endif
181 	}
182 
183 	kernel_fpu_end();
184 }
185 
186 static void raid6_datap_recov_avx2(int disks, size_t bytes, int faila,
187 		void **ptrs)
188 {
189 	u8 *p, *q, *dq;
190 	const u8 *qmul;		/* Q multiplier table */
191 	const u8 x0f = 0x0f;
192 
193 	p = (u8 *)ptrs[disks-2];
194 	q = (u8 *)ptrs[disks-1];
195 
196 	/* Compute syndrome with zero for the missing data page
197 	   Use the dead data page as temporary storage for delta q */
198 	dq = (u8 *)ptrs[faila];
199 	ptrs[faila] = (void *)raid6_empty_zero_page;
200 	ptrs[disks-1] = dq;
201 
202 	raid6_call.gen_syndrome(disks, bytes, ptrs);
203 
204 	/* Restore pointer table */
205 	ptrs[faila]   = dq;
206 	ptrs[disks-1] = q;
207 
208 	/* Now, pick the proper data tables */
209 	qmul  = raid6_vgfmul[raid6_gfinv[raid6_gfexp[faila]]];
210 
211 	kernel_fpu_begin();
212 
213 	asm volatile("vpbroadcastb %0, %%ymm7" : : "m" (x0f));
214 
215 	while (bytes) {
216 #ifdef CONFIG_X86_64
217 		asm volatile("vmovdqa %0, %%ymm3" : : "m" (dq[0]));
218 		asm volatile("vmovdqa %0, %%ymm8" : : "m" (dq[32]));
219 		asm volatile("vpxor %0, %%ymm3, %%ymm3" : : "m" (q[0]));
220 		asm volatile("vpxor %0, %%ymm8, %%ymm8" : : "m" (q[32]));
221 
222 		/*
223 		 * 3 = q[0] ^ dq[0]
224 		 * 8 = q[32] ^ dq[32]
225 		 */
226 		asm volatile("vbroadcasti128 %0, %%ymm0" : : "m" (qmul[0]));
227 		asm volatile("vmovapd %ymm0, %ymm13");
228 		asm volatile("vbroadcasti128 %0, %%ymm1" : : "m" (qmul[16]));
229 		asm volatile("vmovapd %ymm1, %ymm14");
230 
231 		asm volatile("vpsraw $4, %ymm3, %ymm6");
232 		asm volatile("vpsraw $4, %ymm8, %ymm12");
233 		asm volatile("vpand %ymm7, %ymm3, %ymm3");
234 		asm volatile("vpand %ymm7, %ymm8, %ymm8");
235 		asm volatile("vpand %ymm7, %ymm6, %ymm6");
236 		asm volatile("vpand %ymm7, %ymm12, %ymm12");
237 		asm volatile("vpshufb %ymm3, %ymm0, %ymm0");
238 		asm volatile("vpshufb %ymm8, %ymm13, %ymm13");
239 		asm volatile("vpshufb %ymm6, %ymm1, %ymm1");
240 		asm volatile("vpshufb %ymm12, %ymm14, %ymm14");
241 		asm volatile("vpxor %ymm0, %ymm1, %ymm1");
242 		asm volatile("vpxor %ymm13, %ymm14, %ymm14");
243 
244 		/*
245 		 * 1  = qmul[q[0]  ^ dq[0]]
246 		 * 14 = qmul[q[32] ^ dq[32]]
247 		 */
248 		asm volatile("vmovdqa %0, %%ymm2" : : "m" (p[0]));
249 		asm volatile("vmovdqa %0, %%ymm12" : : "m" (p[32]));
250 		asm volatile("vpxor %ymm1, %ymm2, %ymm2");
251 		asm volatile("vpxor %ymm14, %ymm12, %ymm12");
252 
253 		/*
254 		 * 2  = p[0]  ^ qmul[q[0]  ^ dq[0]]
255 		 * 12 = p[32] ^ qmul[q[32] ^ dq[32]]
256 		 */
257 
258 		asm volatile("vmovdqa %%ymm1, %0" : "=m" (dq[0]));
259 		asm volatile("vmovdqa %%ymm14, %0" : "=m" (dq[32]));
260 		asm volatile("vmovdqa %%ymm2, %0" : "=m" (p[0]));
261 		asm volatile("vmovdqa %%ymm12,%0" : "=m" (p[32]));
262 
263 		bytes -= 64;
264 		p += 64;
265 		q += 64;
266 		dq += 64;
267 #else
268 		asm volatile("vmovdqa %0, %%ymm3" : : "m" (dq[0]));
269 		asm volatile("vpxor %0, %%ymm3, %%ymm3" : : "m" (q[0]));
270 
271 		/* 3 = q ^ dq */
272 
273 		asm volatile("vbroadcasti128 %0, %%ymm0" : : "m" (qmul[0]));
274 		asm volatile("vbroadcasti128 %0, %%ymm1" : : "m" (qmul[16]));
275 
276 		asm volatile("vpsraw $4, %ymm3, %ymm6");
277 		asm volatile("vpand %ymm7, %ymm3, %ymm3");
278 		asm volatile("vpand %ymm7, %ymm6, %ymm6");
279 		asm volatile("vpshufb %ymm3, %ymm0, %ymm0");
280 		asm volatile("vpshufb %ymm6, %ymm1, %ymm1");
281 		asm volatile("vpxor %ymm0, %ymm1, %ymm1");
282 
283 		/* 1 = qmul[q ^ dq] */
284 
285 		asm volatile("vmovdqa %0, %%ymm2" : : "m" (p[0]));
286 		asm volatile("vpxor %ymm1, %ymm2, %ymm2");
287 
288 		/* 2 = p ^ qmul[q ^ dq] */
289 
290 		asm volatile("vmovdqa %%ymm1, %0" : "=m" (dq[0]));
291 		asm volatile("vmovdqa %%ymm2, %0" : "=m" (p[0]));
292 
293 		bytes -= 32;
294 		p += 32;
295 		q += 32;
296 		dq += 32;
297 #endif
298 	}
299 
300 	kernel_fpu_end();
301 }
302 
303 const struct raid6_recov_calls raid6_recov_avx2 = {
304 	.data2 = raid6_2data_recov_avx2,
305 	.datap = raid6_datap_recov_avx2,
306 	.valid = raid6_has_avx2,
307 #ifdef CONFIG_X86_64
308 	.name = "avx2x2",
309 #else
310 	.name = "avx2x1",
311 #endif
312 	.priority = 2,
313 };
314