xref: /openbmc/linux/lib/raid6/recov_avx512.c (revision 1c2dd16a)
1 /*
2  * Copyright (C) 2016 Intel Corporation
3  *
4  * Author: Gayatri Kammela <gayatri.kammela@intel.com>
5  * Author: Megha Dey <megha.dey@linux.intel.com>
6  *
7  * This program is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU General Public License
9  * as published by the Free Software Foundation; version 2
10  * of the License.
11  *
12  */
13 
14 #ifdef CONFIG_AS_AVX512
15 
16 #include <linux/raid/pq.h>
17 #include "x86.h"
18 
19 static int raid6_has_avx512(void)
20 {
21 	return boot_cpu_has(X86_FEATURE_AVX2) &&
22 		boot_cpu_has(X86_FEATURE_AVX) &&
23 		boot_cpu_has(X86_FEATURE_AVX512F) &&
24 		boot_cpu_has(X86_FEATURE_AVX512BW) &&
25 		boot_cpu_has(X86_FEATURE_AVX512VL) &&
26 		boot_cpu_has(X86_FEATURE_AVX512DQ);
27 }
28 
29 static void raid6_2data_recov_avx512(int disks, size_t bytes, int faila,
30 				     int failb, void **ptrs)
31 {
32 	u8 *p, *q, *dp, *dq;
33 	const u8 *pbmul;	/* P multiplier table for B data */
34 	const u8 *qmul;		/* Q multiplier table (for both) */
35 	const u8 x0f = 0x0f;
36 
37 	p = (u8 *)ptrs[disks-2];
38 	q = (u8 *)ptrs[disks-1];
39 
40 	/*
41 	 * Compute syndrome with zero for the missing data pages
42 	 * Use the dead data pages as temporary storage for
43 	 * delta p and delta q
44 	 */
45 
46 	dp = (u8 *)ptrs[faila];
47 	ptrs[faila] = (void *)raid6_empty_zero_page;
48 	ptrs[disks-2] = dp;
49 	dq = (u8 *)ptrs[failb];
50 	ptrs[failb] = (void *)raid6_empty_zero_page;
51 	ptrs[disks-1] = dq;
52 
53 	raid6_call.gen_syndrome(disks, bytes, ptrs);
54 
55 	/* Restore pointer table */
56 	ptrs[faila]   = dp;
57 	ptrs[failb]   = dq;
58 	ptrs[disks-2] = p;
59 	ptrs[disks-1] = q;
60 
61 	/* Now, pick the proper data tables */
62 	pbmul = raid6_vgfmul[raid6_gfexi[failb-faila]];
63 	qmul  = raid6_vgfmul[raid6_gfinv[raid6_gfexp[faila] ^
64 		raid6_gfexp[failb]]];
65 
66 	kernel_fpu_begin();
67 
68 	/* zmm0 = x0f[16] */
69 	asm volatile("vpbroadcastb %0, %%zmm7" : : "m" (x0f));
70 
71 	while (bytes) {
72 #ifdef CONFIG_X86_64
73 		asm volatile("vmovdqa64 %0, %%zmm1\n\t"
74 			     "vmovdqa64 %1, %%zmm9\n\t"
75 			     "vmovdqa64 %2, %%zmm0\n\t"
76 			     "vmovdqa64 %3, %%zmm8\n\t"
77 			     "vpxorq %4, %%zmm1, %%zmm1\n\t"
78 			     "vpxorq %5, %%zmm9, %%zmm9\n\t"
79 			     "vpxorq %6, %%zmm0, %%zmm0\n\t"
80 			     "vpxorq %7, %%zmm8, %%zmm8"
81 			     :
82 			     : "m" (q[0]), "m" (q[64]), "m" (p[0]),
83 			       "m" (p[64]), "m" (dq[0]), "m" (dq[64]),
84 			       "m" (dp[0]), "m" (dp[64]));
85 
86 		/*
87 		 * 1 = dq[0]  ^ q[0]
88 		 * 9 = dq[64] ^ q[64]
89 		 * 0 = dp[0]  ^ p[0]
90 		 * 8 = dp[64] ^ p[64]
91 		 */
92 
93 		asm volatile("vbroadcasti64x2 %0, %%zmm4\n\t"
94 			     "vbroadcasti64x2 %1, %%zmm5"
95 			     :
96 			     : "m" (qmul[0]), "m" (qmul[16]));
97 
98 		asm volatile("vpsraw $4, %%zmm1, %%zmm3\n\t"
99 			     "vpsraw $4, %%zmm9, %%zmm12\n\t"
100 			     "vpandq %%zmm7, %%zmm1, %%zmm1\n\t"
101 			     "vpandq %%zmm7, %%zmm9, %%zmm9\n\t"
102 			     "vpandq %%zmm7, %%zmm3, %%zmm3\n\t"
103 			     "vpandq %%zmm7, %%zmm12, %%zmm12\n\t"
104 			     "vpshufb %%zmm9, %%zmm4, %%zmm14\n\t"
105 			     "vpshufb %%zmm1, %%zmm4, %%zmm4\n\t"
106 			     "vpshufb %%zmm12, %%zmm5, %%zmm15\n\t"
107 			     "vpshufb %%zmm3, %%zmm5, %%zmm5\n\t"
108 			     "vpxorq %%zmm14, %%zmm15, %%zmm15\n\t"
109 			     "vpxorq %%zmm4, %%zmm5, %%zmm5"
110 			     :
111 			     : );
112 
113 		/*
114 		 * 5 = qx[0]
115 		 * 15 = qx[64]
116 		 */
117 
118 		asm volatile("vbroadcasti64x2 %0, %%zmm4\n\t"
119 			     "vbroadcasti64x2 %1, %%zmm1\n\t"
120 			     "vpsraw $4, %%zmm0, %%zmm2\n\t"
121 			     "vpsraw $4, %%zmm8, %%zmm6\n\t"
122 			     "vpandq %%zmm7, %%zmm0, %%zmm3\n\t"
123 			     "vpandq %%zmm7, %%zmm8, %%zmm14\n\t"
124 			     "vpandq %%zmm7, %%zmm2, %%zmm2\n\t"
125 			     "vpandq %%zmm7, %%zmm6, %%zmm6\n\t"
126 			     "vpshufb %%zmm14, %%zmm4, %%zmm12\n\t"
127 			     "vpshufb %%zmm3, %%zmm4, %%zmm4\n\t"
128 			     "vpshufb %%zmm6, %%zmm1, %%zmm13\n\t"
129 			     "vpshufb %%zmm2, %%zmm1, %%zmm1\n\t"
130 			     "vpxorq %%zmm4, %%zmm1, %%zmm1\n\t"
131 			     "vpxorq %%zmm12, %%zmm13, %%zmm13"
132 			     :
133 			     : "m" (pbmul[0]), "m" (pbmul[16]));
134 
135 		/*
136 		 * 1  = pbmul[px[0]]
137 		 * 13 = pbmul[px[64]]
138 		 */
139 		asm volatile("vpxorq %%zmm5, %%zmm1, %%zmm1\n\t"
140 			     "vpxorq %%zmm15, %%zmm13, %%zmm13"
141 			     :
142 			     : );
143 
144 		/*
145 		 * 1 = db = DQ
146 		 * 13 = db[64] = DQ[64]
147 		 */
148 		asm volatile("vmovdqa64 %%zmm1, %0\n\t"
149 			     "vmovdqa64 %%zmm13,%1\n\t"
150 			     "vpxorq %%zmm1, %%zmm0, %%zmm0\n\t"
151 			     "vpxorq %%zmm13, %%zmm8, %%zmm8"
152 			     :
153 			     : "m" (dq[0]), "m" (dq[64]));
154 
155 		asm volatile("vmovdqa64 %%zmm0, %0\n\t"
156 			     "vmovdqa64 %%zmm8, %1"
157 			     :
158 			     : "m" (dp[0]), "m" (dp[64]));
159 
160 		bytes -= 128;
161 		p += 128;
162 		q += 128;
163 		dp += 128;
164 		dq += 128;
165 #else
166 		asm volatile("vmovdqa64 %0, %%zmm1\n\t"
167 			     "vmovdqa64 %1, %%zmm0\n\t"
168 			     "vpxorq %2, %%zmm1, %%zmm1\n\t"
169 			     "vpxorq %3, %%zmm0, %%zmm0"
170 			     :
171 			     : "m" (*q), "m" (*p), "m"(*dq), "m" (*dp));
172 
173 		/* 1 = dq ^ q;  0 = dp ^ p */
174 
175 		asm volatile("vbroadcasti64x2 %0, %%zmm4\n\t"
176 			     "vbroadcasti64x2 %1, %%zmm5"
177 			     :
178 			     : "m" (qmul[0]), "m" (qmul[16]));
179 
180 		/*
181 		 * 1 = dq ^ q
182 		 * 3 = dq ^ p >> 4
183 		 */
184 		asm volatile("vpsraw $4, %%zmm1, %%zmm3\n\t"
185 			     "vpandq %%zmm7, %%zmm1, %%zmm1\n\t"
186 			     "vpandq %%zmm7, %%zmm3, %%zmm3\n\t"
187 			     "vpshufb %%zmm1, %%zmm4, %%zmm4\n\t"
188 			     "vpshufb %%zmm3, %%zmm5, %%zmm5\n\t"
189 			     "vpxorq %%zmm4, %%zmm5, %%zmm5"
190 			     :
191 			     : );
192 
193 		/* 5 = qx */
194 
195 		asm volatile("vbroadcasti64x2 %0, %%zmm4\n\t"
196 			     "vbroadcasti64x2 %1, %%zmm1"
197 			     :
198 			     : "m" (pbmul[0]), "m" (pbmul[16]));
199 
200 		asm volatile("vpsraw $4, %%zmm0, %%zmm2\n\t"
201 			     "vpandq %%zmm7, %%zmm0, %%zmm3\n\t"
202 			     "vpandq %%zmm7, %%zmm2, %%zmm2\n\t"
203 			     "vpshufb %%zmm3, %%zmm4, %%zmm4\n\t"
204 			     "vpshufb %%zmm2, %%zmm1, %%zmm1\n\t"
205 			     "vpxorq %%zmm4, %%zmm1, %%zmm1"
206 			     :
207 			     : );
208 
209 		/* 1 = pbmul[px] */
210 		asm volatile("vpxorq %%zmm5, %%zmm1, %%zmm1\n\t"
211 			     /* 1 = db = DQ */
212 			     "vmovdqa64 %%zmm1, %0\n\t"
213 			     :
214 			     : "m" (dq[0]));
215 
216 		asm volatile("vpxorq %%zmm1, %%zmm0, %%zmm0\n\t"
217 			     "vmovdqa64 %%zmm0, %0"
218 			     :
219 			     : "m" (dp[0]));
220 
221 		bytes -= 64;
222 		p += 64;
223 		q += 64;
224 		dp += 64;
225 		dq += 64;
226 #endif
227 	}
228 
229 	kernel_fpu_end();
230 }
231 
232 static void raid6_datap_recov_avx512(int disks, size_t bytes, int faila,
233 				     void **ptrs)
234 {
235 	u8 *p, *q, *dq;
236 	const u8 *qmul;		/* Q multiplier table */
237 	const u8 x0f = 0x0f;
238 
239 	p = (u8 *)ptrs[disks-2];
240 	q = (u8 *)ptrs[disks-1];
241 
242 	/*
243 	 * Compute syndrome with zero for the missing data page
244 	 * Use the dead data page as temporary storage for delta q
245 	 */
246 
247 	dq = (u8 *)ptrs[faila];
248 	ptrs[faila] = (void *)raid6_empty_zero_page;
249 	ptrs[disks-1] = dq;
250 
251 	raid6_call.gen_syndrome(disks, bytes, ptrs);
252 
253 	/* Restore pointer table */
254 	ptrs[faila]   = dq;
255 	ptrs[disks-1] = q;
256 
257 	/* Now, pick the proper data tables */
258 	qmul  = raid6_vgfmul[raid6_gfinv[raid6_gfexp[faila]]];
259 
260 	kernel_fpu_begin();
261 
262 	asm volatile("vpbroadcastb %0, %%zmm7" : : "m" (x0f));
263 
264 	while (bytes) {
265 #ifdef CONFIG_X86_64
266 		asm volatile("vmovdqa64 %0, %%zmm3\n\t"
267 			     "vmovdqa64 %1, %%zmm8\n\t"
268 			     "vpxorq %2, %%zmm3, %%zmm3\n\t"
269 			     "vpxorq %3, %%zmm8, %%zmm8"
270 			     :
271 			     : "m" (dq[0]), "m" (dq[64]), "m" (q[0]),
272 			       "m" (q[64]));
273 
274 		/*
275 		 * 3 = q[0] ^ dq[0]
276 		 * 8 = q[64] ^ dq[64]
277 		 */
278 		asm volatile("vbroadcasti64x2 %0, %%zmm0\n\t"
279 			     "vmovapd %%zmm0, %%zmm13\n\t"
280 			     "vbroadcasti64x2 %1, %%zmm1\n\t"
281 			     "vmovapd %%zmm1, %%zmm14"
282 			     :
283 			     : "m" (qmul[0]), "m" (qmul[16]));
284 
285 		asm volatile("vpsraw $4, %%zmm3, %%zmm6\n\t"
286 			     "vpsraw $4, %%zmm8, %%zmm12\n\t"
287 			     "vpandq %%zmm7, %%zmm3, %%zmm3\n\t"
288 			     "vpandq %%zmm7, %%zmm8, %%zmm8\n\t"
289 			     "vpandq %%zmm7, %%zmm6, %%zmm6\n\t"
290 			     "vpandq %%zmm7, %%zmm12, %%zmm12\n\t"
291 			     "vpshufb %%zmm3, %%zmm0, %%zmm0\n\t"
292 			     "vpshufb %%zmm8, %%zmm13, %%zmm13\n\t"
293 			     "vpshufb %%zmm6, %%zmm1, %%zmm1\n\t"
294 			     "vpshufb %%zmm12, %%zmm14, %%zmm14\n\t"
295 			     "vpxorq %%zmm0, %%zmm1, %%zmm1\n\t"
296 			     "vpxorq %%zmm13, %%zmm14, %%zmm14"
297 			     :
298 			     : );
299 
300 		/*
301 		 * 1  = qmul[q[0]  ^ dq[0]]
302 		 * 14 = qmul[q[64] ^ dq[64]]
303 		 */
304 		asm volatile("vmovdqa64 %0, %%zmm2\n\t"
305 			     "vmovdqa64 %1, %%zmm12\n\t"
306 			     "vpxorq %%zmm1, %%zmm2, %%zmm2\n\t"
307 			     "vpxorq %%zmm14, %%zmm12, %%zmm12"
308 			     :
309 			     : "m" (p[0]), "m" (p[64]));
310 
311 		/*
312 		 * 2  = p[0]  ^ qmul[q[0]  ^ dq[0]]
313 		 * 12 = p[64] ^ qmul[q[64] ^ dq[64]]
314 		 */
315 
316 		asm volatile("vmovdqa64 %%zmm1, %0\n\t"
317 			     "vmovdqa64 %%zmm14, %1\n\t"
318 			     "vmovdqa64 %%zmm2, %2\n\t"
319 			     "vmovdqa64 %%zmm12,%3"
320 			     :
321 			     : "m" (dq[0]), "m" (dq[64]), "m" (p[0]),
322 			       "m" (p[64]));
323 
324 		bytes -= 128;
325 		p += 128;
326 		q += 128;
327 		dq += 128;
328 #else
329 		asm volatile("vmovdqa64 %0, %%zmm3\n\t"
330 			     "vpxorq %1, %%zmm3, %%zmm3"
331 			     :
332 			     : "m" (dq[0]), "m" (q[0]));
333 
334 		/* 3 = q ^ dq */
335 
336 		asm volatile("vbroadcasti64x2 %0, %%zmm0\n\t"
337 			     "vbroadcasti64x2 %1, %%zmm1"
338 			     :
339 			     : "m" (qmul[0]), "m" (qmul[16]));
340 
341 		asm volatile("vpsraw $4, %%zmm3, %%zmm6\n\t"
342 			     "vpandq %%zmm7, %%zmm3, %%zmm3\n\t"
343 			     "vpandq %%zmm7, %%zmm6, %%zmm6\n\t"
344 			     "vpshufb %%zmm3, %%zmm0, %%zmm0\n\t"
345 			     "vpshufb %%zmm6, %%zmm1, %%zmm1\n\t"
346 			     "vpxorq %%zmm0, %%zmm1, %%zmm1"
347 			     :
348 			     : );
349 
350 		/* 1 = qmul[q ^ dq] */
351 
352 		asm volatile("vmovdqa64 %0, %%zmm2\n\t"
353 			     "vpxorq %%zmm1, %%zmm2, %%zmm2"
354 			     :
355 			     : "m" (p[0]));
356 
357 		/* 2 = p ^ qmul[q ^ dq] */
358 
359 		asm volatile("vmovdqa64 %%zmm1, %0\n\t"
360 			     "vmovdqa64 %%zmm2, %1"
361 			     :
362 			     : "m" (dq[0]), "m" (p[0]));
363 
364 		bytes -= 64;
365 		p += 64;
366 		q += 64;
367 		dq += 64;
368 #endif
369 	}
370 
371 	kernel_fpu_end();
372 }
373 
374 const struct raid6_recov_calls raid6_recov_avx512 = {
375 	.data2 = raid6_2data_recov_avx512,
376 	.datap = raid6_datap_recov_avx512,
377 	.valid = raid6_has_avx512,
378 #ifdef CONFIG_X86_64
379 	.name = "avx512x2",
380 #else
381 	.name = "avx512x1",
382 #endif
383 	.priority = 3,
384 };
385 
386 #else
387 #warning "your version of binutils lacks AVX512 support"
388 #endif
389