1 // SPDX-License-Identifier: GPL-2.0-only 1
2 /*
3 * Copyright (C) 2016 Intel Corporation
4 *
5 * Author: Gayatri Kammela <gayatri.kammela@in
6 * Author: Megha Dey <megha.dey@linux.intel.co
7 */
8
9 #ifdef CONFIG_AS_AVX512
10
11 #include <linux/raid/pq.h>
12 #include "x86.h"
13
14 static int raid6_has_avx512(void)
15 {
16 return boot_cpu_has(X86_FEATURE_AVX2)
17 boot_cpu_has(X86_FEATURE_AVX)
18 boot_cpu_has(X86_FEATURE_AVX51
19 boot_cpu_has(X86_FEATURE_AVX51
20 boot_cpu_has(X86_FEATURE_AVX51
21 boot_cpu_has(X86_FEATURE_AVX51
22 }
23
24 static void raid6_2data_recov_avx512(int disks
25 int failb
26 {
27 u8 *p, *q, *dp, *dq;
28 const u8 *pbmul; /* P multiplie
29 const u8 *qmul; /* Q multiplie
30 const u8 x0f = 0x0f;
31
32 p = (u8 *)ptrs[disks-2];
33 q = (u8 *)ptrs[disks-1];
34
35 /*
36 * Compute syndrome with zero for the
37 * Use the dead data pages as temporar
38 * delta p and delta q
39 */
40
41 dp = (u8 *)ptrs[faila];
42 ptrs[faila] = (void *)raid6_empty_zero
43 ptrs[disks-2] = dp;
44 dq = (u8 *)ptrs[failb];
45 ptrs[failb] = (void *)raid6_empty_zero
46 ptrs[disks-1] = dq;
47
48 raid6_call.gen_syndrome(disks, bytes,
49
50 /* Restore pointer table */
51 ptrs[faila] = dp;
52 ptrs[failb] = dq;
53 ptrs[disks-2] = p;
54 ptrs[disks-1] = q;
55
56 /* Now, pick the proper data tables */
57 pbmul = raid6_vgfmul[raid6_gfexi[failb
58 qmul = raid6_vgfmul[raid6_gfinv[raid6
59 raid6_gfexp[failb]]];
60
61 kernel_fpu_begin();
62
63 /* zmm0 = x0f[16] */
64 asm volatile("vpbroadcastb %0, %%zmm7"
65
66 while (bytes) {
67 #ifdef CONFIG_X86_64
68 asm volatile("vmovdqa64 %0, %%
69 "vmovdqa64 %1, %%
70 "vmovdqa64 %2, %%
71 "vmovdqa64 %3, %%
72 "vpxorq %4, %%zmm
73 "vpxorq %5, %%zmm
74 "vpxorq %6, %%zmm
75 "vpxorq %7, %%zmm
76 :
77 : "m" (q[0]), "m"
78 "m" (p[64]), "m
79 "m" (dp[0]), "m
80
81 /*
82 * 1 = dq[0] ^ q[0]
83 * 9 = dq[64] ^ q[64]
84 * 0 = dp[0] ^ p[0]
85 * 8 = dp[64] ^ p[64]
86 */
87
88 asm volatile("vbroadcasti64x2
89 "vbroadcasti64x2
90 :
91 : "m" (qmul[0]),
92
93 asm volatile("vpsraw $4, %%zmm
94 "vpsraw $4, %%zmm
95 "vpandq %%zmm7, %
96 "vpandq %%zmm7, %
97 "vpandq %%zmm7, %
98 "vpandq %%zmm7, %
99 "vpshufb %%zmm9,
100 "vpshufb %%zmm1,
101 "vpshufb %%zmm12,
102 "vpshufb %%zmm3,
103 "vpxorq %%zmm14,
104 "vpxorq %%zmm4, %
105 :
106 : );
107
108 /*
109 * 5 = qx[0]
110 * 15 = qx[64]
111 */
112
113 asm volatile("vbroadcasti64x2
114 "vbroadcasti64x2
115 "vpsraw $4, %%zmm
116 "vpsraw $4, %%zmm
117 "vpandq %%zmm7, %
118 "vpandq %%zmm7, %
119 "vpandq %%zmm7, %
120 "vpandq %%zmm7, %
121 "vpshufb %%zmm14,
122 "vpshufb %%zmm3,
123 "vpshufb %%zmm6,
124 "vpshufb %%zmm2,
125 "vpxorq %%zmm4, %
126 "vpxorq %%zmm12,
127 :
128 : "m" (pbmul[0]),
129
130 /*
131 * 1 = pbmul[px[0]]
132 * 13 = pbmul[px[64]]
133 */
134 asm volatile("vpxorq %%zmm5, %
135 "vpxorq %%zmm15,
136 :
137 : );
138
139 /*
140 * 1 = db = DQ
141 * 13 = db[64] = DQ[64]
142 */
143 asm volatile("vmovdqa64 %%zmm1
144 "vmovdqa64 %%zmm1
145 "vpxorq %%zmm1, %
146 "vpxorq %%zmm13,
147 :
148 : "m" (dq[0]), "m
149
150 asm volatile("vmovdqa64 %%zmm0
151 "vmovdqa64 %%zmm8
152 :
153 : "m" (dp[0]), "m
154
155 bytes -= 128;
156 p += 128;
157 q += 128;
158 dp += 128;
159 dq += 128;
160 #else
161 asm volatile("vmovdqa64 %0, %%
162 "vmovdqa64 %1, %%
163 "vpxorq %2, %%zmm
164 "vpxorq %3, %%zmm
165 :
166 : "m" (*q), "m" (
167
168 /* 1 = dq ^ q; 0 = dp ^ p */
169
170 asm volatile("vbroadcasti64x2
171 "vbroadcasti64x2
172 :
173 : "m" (qmul[0]),
174
175 /*
176 * 1 = dq ^ q
177 * 3 = dq ^ p >> 4
178 */
179 asm volatile("vpsraw $4, %%zmm
180 "vpandq %%zmm7, %
181 "vpandq %%zmm7, %
182 "vpshufb %%zmm1,
183 "vpshufb %%zmm3,
184 "vpxorq %%zmm4, %
185 :
186 : );
187
188 /* 5 = qx */
189
190 asm volatile("vbroadcasti64x2
191 "vbroadcasti64x2
192 :
193 : "m" (pbmul[0]),
194
195 asm volatile("vpsraw $4, %%zmm
196 "vpandq %%zmm7, %
197 "vpandq %%zmm7, %
198 "vpshufb %%zmm3,
199 "vpshufb %%zmm2,
200 "vpxorq %%zmm4, %
201 :
202 : );
203
204 /* 1 = pbmul[px] */
205 asm volatile("vpxorq %%zmm5, %
206 /* 1 = db = DQ */
207 "vmovdqa64 %%zmm1
208 :
209 : "m" (dq[0]));
210
211 asm volatile("vpxorq %%zmm1, %
212 "vmovdqa64 %%zmm0
213 :
214 : "m" (dp[0]));
215
216 bytes -= 64;
217 p += 64;
218 q += 64;
219 dp += 64;
220 dq += 64;
221 #endif
222 }
223
224 kernel_fpu_end();
225 }
226
227 static void raid6_datap_recov_avx512(int disks
228 void **pt
229 {
230 u8 *p, *q, *dq;
231 const u8 *qmul; /* Q multiplie
232 const u8 x0f = 0x0f;
233
234 p = (u8 *)ptrs[disks-2];
235 q = (u8 *)ptrs[disks-1];
236
237 /*
238 * Compute syndrome with zero for the
239 * Use the dead data page as temporary
240 */
241
242 dq = (u8 *)ptrs[faila];
243 ptrs[faila] = (void *)raid6_empty_zero
244 ptrs[disks-1] = dq;
245
246 raid6_call.gen_syndrome(disks, bytes,
247
248 /* Restore pointer table */
249 ptrs[faila] = dq;
250 ptrs[disks-1] = q;
251
252 /* Now, pick the proper data tables */
253 qmul = raid6_vgfmul[raid6_gfinv[raid6
254
255 kernel_fpu_begin();
256
257 asm volatile("vpbroadcastb %0, %%zmm7"
258
259 while (bytes) {
260 #ifdef CONFIG_X86_64
261 asm volatile("vmovdqa64 %0, %%
262 "vmovdqa64 %1, %%
263 "vpxorq %2, %%zmm
264 "vpxorq %3, %%zmm
265 :
266 : "m" (dq[0]), "m
267 "m" (q[64]));
268
269 /*
270 * 3 = q[0] ^ dq[0]
271 * 8 = q[64] ^ dq[64]
272 */
273 asm volatile("vbroadcasti64x2
274 "vmovapd %%zmm0,
275 "vbroadcasti64x2
276 "vmovapd %%zmm1,
277 :
278 : "m" (qmul[0]),
279
280 asm volatile("vpsraw $4, %%zmm
281 "vpsraw $4, %%zmm
282 "vpandq %%zmm7, %
283 "vpandq %%zmm7, %
284 "vpandq %%zmm7, %
285 "vpandq %%zmm7, %
286 "vpshufb %%zmm3,
287 "vpshufb %%zmm8,
288 "vpshufb %%zmm6,
289 "vpshufb %%zmm12,
290 "vpxorq %%zmm0, %
291 "vpxorq %%zmm13,
292 :
293 : );
294
295 /*
296 * 1 = qmul[q[0] ^ dq[0]]
297 * 14 = qmul[q[64] ^ dq[64]]
298 */
299 asm volatile("vmovdqa64 %0, %%
300 "vmovdqa64 %1, %%
301 "vpxorq %%zmm1, %
302 "vpxorq %%zmm14,
303 :
304 : "m" (p[0]), "m"
305
306 /*
307 * 2 = p[0] ^ qmul[q[0] ^ d
308 * 12 = p[64] ^ qmul[q[64] ^ d
309 */
310
311 asm volatile("vmovdqa64 %%zmm1
312 "vmovdqa64 %%zmm1
313 "vmovdqa64 %%zmm2
314 "vmovdqa64 %%zmm1
315 :
316 : "m" (dq[0]), "m
317 "m" (p[64]));
318
319 bytes -= 128;
320 p += 128;
321 q += 128;
322 dq += 128;
323 #else
324 asm volatile("vmovdqa64 %0, %%
325 "vpxorq %1, %%zmm
326 :
327 : "m" (dq[0]), "m
328
329 /* 3 = q ^ dq */
330
331 asm volatile("vbroadcasti64x2
332 "vbroadcasti64x2
333 :
334 : "m" (qmul[0]),
335
336 asm volatile("vpsraw $4, %%zmm
337 "vpandq %%zmm7, %
338 "vpandq %%zmm7, %
339 "vpshufb %%zmm3,
340 "vpshufb %%zmm6,
341 "vpxorq %%zmm0, %
342 :
343 : );
344
345 /* 1 = qmul[q ^ dq] */
346
347 asm volatile("vmovdqa64 %0, %%
348 "vpxorq %%zmm1, %
349 :
350 : "m" (p[0]));
351
352 /* 2 = p ^ qmul[q ^ dq] */
353
354 asm volatile("vmovdqa64 %%zmm1
355 "vmovdqa64 %%zmm2
356 :
357 : "m" (dq[0]), "m
358
359 bytes -= 64;
360 p += 64;
361 q += 64;
362 dq += 64;
363 #endif
364 }
365
366 kernel_fpu_end();
367 }
368
369 const struct raid6_recov_calls raid6_recov_avx
370 .data2 = raid6_2data_recov_avx512,
371 .datap = raid6_datap_recov_avx512,
372 .valid = raid6_has_avx512,
373 #ifdef CONFIG_X86_64
374 .name = "avx512x2",
375 #else
376 .name = "avx512x1",
377 #endif
378 .priority = 3,
379 };
380
381 #else
382 #warning "your version of binutils lacks AVX51
383 #endif
384
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.