1 /*
2  * Copyright (C) 2012 Intel Corporation
3  * Author: Jim Kukunas <james.t.kukunas@linux.intel.com>
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License
7  * as published by the Free Software Foundation; version 2
8  * of the License.
9  */
10 
11 #ifdef CONFIG_AS_AVX2
12 
13 #include <linux/raid/pq.h>
14 #include "x86.h"
15 
raid6_has_avx2(void)16 static int raid6_has_avx2(void)
17 {
18 	return boot_cpu_has(X86_FEATURE_AVX2) &&
19 		boot_cpu_has(X86_FEATURE_AVX);
20 }
21 
raid6_2data_recov_avx2(int disks,size_t bytes,int faila,int failb,void ** ptrs)22 static void raid6_2data_recov_avx2(int disks, size_t bytes, int faila,
23 		int failb, void **ptrs)
24 {
25 	u8 *p, *q, *dp, *dq;
26 	const u8 *pbmul;	/* P multiplier table for B data */
27 	const u8 *qmul;		/* Q multiplier table (for both) */
28 	const u8 x0f = 0x0f;
29 
30 	p = (u8 *)ptrs[disks-2];
31 	q = (u8 *)ptrs[disks-1];
32 
33 	/* Compute syndrome with zero for the missing data pages
34 	   Use the dead data pages as temporary storage for
35 	   delta p and delta q */
36 	dp = (u8 *)ptrs[faila];
37 	ptrs[faila] = (void *)raid6_empty_zero_page;
38 	ptrs[disks-2] = dp;
39 	dq = (u8 *)ptrs[failb];
40 	ptrs[failb] = (void *)raid6_empty_zero_page;
41 	ptrs[disks-1] = dq;
42 
43 	raid6_call.gen_syndrome(disks, bytes, ptrs);
44 
45 	/* Restore pointer table */
46 	ptrs[faila]   = dp;
47 	ptrs[failb]   = dq;
48 	ptrs[disks-2] = p;
49 	ptrs[disks-1] = q;
50 
51 	/* Now, pick the proper data tables */
52 	pbmul = raid6_vgfmul[raid6_gfexi[failb-faila]];
53 	qmul  = raid6_vgfmul[raid6_gfinv[raid6_gfexp[faila] ^
54 		raid6_gfexp[failb]]];
55 
56 	kernel_fpu_begin();
57 
58 	/* ymm0 = x0f[16] */
59 	asm volatile("vpbroadcastb %0, %%ymm7" : : "m" (x0f));
60 
61 	while (bytes) {
62 #ifdef CONFIG_X86_64
63 		asm volatile("vmovdqa %0, %%ymm1" : : "m" (q[0]));
64 		asm volatile("vmovdqa %0, %%ymm9" : : "m" (q[32]));
65 		asm volatile("vmovdqa %0, %%ymm0" : : "m" (p[0]));
66 		asm volatile("vmovdqa %0, %%ymm8" : : "m" (p[32]));
67 		asm volatile("vpxor %0, %%ymm1, %%ymm1" : : "m" (dq[0]));
68 		asm volatile("vpxor %0, %%ymm9, %%ymm9" : : "m" (dq[32]));
69 		asm volatile("vpxor %0, %%ymm0, %%ymm0" : : "m" (dp[0]));
70 		asm volatile("vpxor %0, %%ymm8, %%ymm8" : : "m" (dp[32]));
71 
72 		/*
73 		 * 1 = dq[0]  ^ q[0]
74 		 * 9 = dq[32] ^ q[32]
75 		 * 0 = dp[0]  ^ p[0]
76 		 * 8 = dp[32] ^ p[32]
77 		 */
78 
79 		asm volatile("vbroadcasti128 %0, %%ymm4" : : "m" (qmul[0]));
80 		asm volatile("vbroadcasti128 %0, %%ymm5" : : "m" (qmul[16]));
81 
82 		asm volatile("vpsraw $4, %ymm1, %ymm3");
83 		asm volatile("vpsraw $4, %ymm9, %ymm12");
84 		asm volatile("vpand %ymm7, %ymm1, %ymm1");
85 		asm volatile("vpand %ymm7, %ymm9, %ymm9");
86 		asm volatile("vpand %ymm7, %ymm3, %ymm3");
87 		asm volatile("vpand %ymm7, %ymm12, %ymm12");
88 		asm volatile("vpshufb %ymm9, %ymm4, %ymm14");
89 		asm volatile("vpshufb %ymm1, %ymm4, %ymm4");
90 		asm volatile("vpshufb %ymm12, %ymm5, %ymm15");
91 		asm volatile("vpshufb %ymm3, %ymm5, %ymm5");
92 		asm volatile("vpxor %ymm14, %ymm15, %ymm15");
93 		asm volatile("vpxor %ymm4, %ymm5, %ymm5");
94 
95 		/*
96 		 * 5 = qx[0]
97 		 * 15 = qx[32]
98 		 */
99 
100 		asm volatile("vbroadcasti128 %0, %%ymm4" : : "m" (pbmul[0]));
101 		asm volatile("vbroadcasti128 %0, %%ymm1" : : "m" (pbmul[16]));
102 		asm volatile("vpsraw $4, %ymm0, %ymm2");
103 		asm volatile("vpsraw $4, %ymm8, %ymm6");
104 		asm volatile("vpand %ymm7, %ymm0, %ymm3");
105 		asm volatile("vpand %ymm7, %ymm8, %ymm14");
106 		asm volatile("vpand %ymm7, %ymm2, %ymm2");
107 		asm volatile("vpand %ymm7, %ymm6, %ymm6");
108 		asm volatile("vpshufb %ymm14, %ymm4, %ymm12");
109 		asm volatile("vpshufb %ymm3, %ymm4, %ymm4");
110 		asm volatile("vpshufb %ymm6, %ymm1, %ymm13");
111 		asm volatile("vpshufb %ymm2, %ymm1, %ymm1");
112 		asm volatile("vpxor %ymm4, %ymm1, %ymm1");
113 		asm volatile("vpxor %ymm12, %ymm13, %ymm13");
114 
115 		/*
116 		 * 1  = pbmul[px[0]]
117 		 * 13 = pbmul[px[32]]
118 		 */
119 		asm volatile("vpxor %ymm5, %ymm1, %ymm1");
120 		asm volatile("vpxor %ymm15, %ymm13, %ymm13");
121 
122 		/*
123 		 * 1 = db = DQ
124 		 * 13 = db[32] = DQ[32]
125 		 */
126 		asm volatile("vmovdqa %%ymm1, %0" : "=m" (dq[0]));
127 		asm volatile("vmovdqa %%ymm13,%0" : "=m" (dq[32]));
128 		asm volatile("vpxor %ymm1, %ymm0, %ymm0");
129 		asm volatile("vpxor %ymm13, %ymm8, %ymm8");
130 
131 		asm volatile("vmovdqa %%ymm0, %0" : "=m" (dp[0]));
132 		asm volatile("vmovdqa %%ymm8, %0" : "=m" (dp[32]));
133 
134 		bytes -= 64;
135 		p += 64;
136 		q += 64;
137 		dp += 64;
138 		dq += 64;
139 #else
140 		asm volatile("vmovdqa %0, %%ymm1" : : "m" (*q));
141 		asm volatile("vmovdqa %0, %%ymm0" : : "m" (*p));
142 		asm volatile("vpxor %0, %%ymm1, %%ymm1" : : "m" (*dq));
143 		asm volatile("vpxor %0, %%ymm0, %%ymm0" : : "m" (*dp));
144 
145 		/* 1 = dq ^ q;  0 = dp ^ p */
146 
147 		asm volatile("vbroadcasti128 %0, %%ymm4" : : "m" (qmul[0]));
148 		asm volatile("vbroadcasti128 %0, %%ymm5" : : "m" (qmul[16]));
149 
150 		/*
151 		 * 1 = dq ^ q
152 		 * 3 = dq ^ p >> 4
153 		 */
154 		asm volatile("vpsraw $4, %ymm1, %ymm3");
155 		asm volatile("vpand %ymm7, %ymm1, %ymm1");
156 		asm volatile("vpand %ymm7, %ymm3, %ymm3");
157 		asm volatile("vpshufb %ymm1, %ymm4, %ymm4");
158 		asm volatile("vpshufb %ymm3, %ymm5, %ymm5");
159 		asm volatile("vpxor %ymm4, %ymm5, %ymm5");
160 
161 		/* 5 = qx */
162 
163 		asm volatile("vbroadcasti128 %0, %%ymm4" : : "m" (pbmul[0]));
164 		asm volatile("vbroadcasti128 %0, %%ymm1" : : "m" (pbmul[16]));
165 
166 		asm volatile("vpsraw $4, %ymm0, %ymm2");
167 		asm volatile("vpand %ymm7, %ymm0, %ymm3");
168 		asm volatile("vpand %ymm7, %ymm2, %ymm2");
169 		asm volatile("vpshufb %ymm3, %ymm4, %ymm4");
170 		asm volatile("vpshufb %ymm2, %ymm1, %ymm1");
171 		asm volatile("vpxor %ymm4, %ymm1, %ymm1");
172 
173 		/* 1 = pbmul[px] */
174 		asm volatile("vpxor %ymm5, %ymm1, %ymm1");
175 		/* 1 = db = DQ */
176 		asm volatile("vmovdqa %%ymm1, %0" : "=m" (dq[0]));
177 
178 		asm volatile("vpxor %ymm1, %ymm0, %ymm0");
179 		asm volatile("vmovdqa %%ymm0, %0" : "=m" (dp[0]));
180 
181 		bytes -= 32;
182 		p += 32;
183 		q += 32;
184 		dp += 32;
185 		dq += 32;
186 #endif
187 	}
188 
189 	kernel_fpu_end();
190 }
191 
raid6_datap_recov_avx2(int disks,size_t bytes,int faila,void ** ptrs)192 static void raid6_datap_recov_avx2(int disks, size_t bytes, int faila,
193 		void **ptrs)
194 {
195 	u8 *p, *q, *dq;
196 	const u8 *qmul;		/* Q multiplier table */
197 	const u8 x0f = 0x0f;
198 
199 	p = (u8 *)ptrs[disks-2];
200 	q = (u8 *)ptrs[disks-1];
201 
202 	/* Compute syndrome with zero for the missing data page
203 	   Use the dead data page as temporary storage for delta q */
204 	dq = (u8 *)ptrs[faila];
205 	ptrs[faila] = (void *)raid6_empty_zero_page;
206 	ptrs[disks-1] = dq;
207 
208 	raid6_call.gen_syndrome(disks, bytes, ptrs);
209 
210 	/* Restore pointer table */
211 	ptrs[faila]   = dq;
212 	ptrs[disks-1] = q;
213 
214 	/* Now, pick the proper data tables */
215 	qmul  = raid6_vgfmul[raid6_gfinv[raid6_gfexp[faila]]];
216 
217 	kernel_fpu_begin();
218 
219 	asm volatile("vpbroadcastb %0, %%ymm7" : : "m" (x0f));
220 
221 	while (bytes) {
222 #ifdef CONFIG_X86_64
223 		asm volatile("vmovdqa %0, %%ymm3" : : "m" (dq[0]));
224 		asm volatile("vmovdqa %0, %%ymm8" : : "m" (dq[32]));
225 		asm volatile("vpxor %0, %%ymm3, %%ymm3" : : "m" (q[0]));
226 		asm volatile("vpxor %0, %%ymm8, %%ymm8" : : "m" (q[32]));
227 
228 		/*
229 		 * 3 = q[0] ^ dq[0]
230 		 * 8 = q[32] ^ dq[32]
231 		 */
232 		asm volatile("vbroadcasti128 %0, %%ymm0" : : "m" (qmul[0]));
233 		asm volatile("vmovapd %ymm0, %ymm13");
234 		asm volatile("vbroadcasti128 %0, %%ymm1" : : "m" (qmul[16]));
235 		asm volatile("vmovapd %ymm1, %ymm14");
236 
237 		asm volatile("vpsraw $4, %ymm3, %ymm6");
238 		asm volatile("vpsraw $4, %ymm8, %ymm12");
239 		asm volatile("vpand %ymm7, %ymm3, %ymm3");
240 		asm volatile("vpand %ymm7, %ymm8, %ymm8");
241 		asm volatile("vpand %ymm7, %ymm6, %ymm6");
242 		asm volatile("vpand %ymm7, %ymm12, %ymm12");
243 		asm volatile("vpshufb %ymm3, %ymm0, %ymm0");
244 		asm volatile("vpshufb %ymm8, %ymm13, %ymm13");
245 		asm volatile("vpshufb %ymm6, %ymm1, %ymm1");
246 		asm volatile("vpshufb %ymm12, %ymm14, %ymm14");
247 		asm volatile("vpxor %ymm0, %ymm1, %ymm1");
248 		asm volatile("vpxor %ymm13, %ymm14, %ymm14");
249 
250 		/*
251 		 * 1  = qmul[q[0]  ^ dq[0]]
252 		 * 14 = qmul[q[32] ^ dq[32]]
253 		 */
254 		asm volatile("vmovdqa %0, %%ymm2" : : "m" (p[0]));
255 		asm volatile("vmovdqa %0, %%ymm12" : : "m" (p[32]));
256 		asm volatile("vpxor %ymm1, %ymm2, %ymm2");
257 		asm volatile("vpxor %ymm14, %ymm12, %ymm12");
258 
259 		/*
260 		 * 2  = p[0]  ^ qmul[q[0]  ^ dq[0]]
261 		 * 12 = p[32] ^ qmul[q[32] ^ dq[32]]
262 		 */
263 
264 		asm volatile("vmovdqa %%ymm1, %0" : "=m" (dq[0]));
265 		asm volatile("vmovdqa %%ymm14, %0" : "=m" (dq[32]));
266 		asm volatile("vmovdqa %%ymm2, %0" : "=m" (p[0]));
267 		asm volatile("vmovdqa %%ymm12,%0" : "=m" (p[32]));
268 
269 		bytes -= 64;
270 		p += 64;
271 		q += 64;
272 		dq += 64;
273 #else
274 		asm volatile("vmovdqa %0, %%ymm3" : : "m" (dq[0]));
275 		asm volatile("vpxor %0, %%ymm3, %%ymm3" : : "m" (q[0]));
276 
277 		/* 3 = q ^ dq */
278 
279 		asm volatile("vbroadcasti128 %0, %%ymm0" : : "m" (qmul[0]));
280 		asm volatile("vbroadcasti128 %0, %%ymm1" : : "m" (qmul[16]));
281 
282 		asm volatile("vpsraw $4, %ymm3, %ymm6");
283 		asm volatile("vpand %ymm7, %ymm3, %ymm3");
284 		asm volatile("vpand %ymm7, %ymm6, %ymm6");
285 		asm volatile("vpshufb %ymm3, %ymm0, %ymm0");
286 		asm volatile("vpshufb %ymm6, %ymm1, %ymm1");
287 		asm volatile("vpxor %ymm0, %ymm1, %ymm1");
288 
289 		/* 1 = qmul[q ^ dq] */
290 
291 		asm volatile("vmovdqa %0, %%ymm2" : : "m" (p[0]));
292 		asm volatile("vpxor %ymm1, %ymm2, %ymm2");
293 
294 		/* 2 = p ^ qmul[q ^ dq] */
295 
296 		asm volatile("vmovdqa %%ymm1, %0" : "=m" (dq[0]));
297 		asm volatile("vmovdqa %%ymm2, %0" : "=m" (p[0]));
298 
299 		bytes -= 32;
300 		p += 32;
301 		q += 32;
302 		dq += 32;
303 #endif
304 	}
305 
306 	kernel_fpu_end();
307 }
308 
309 const struct raid6_recov_calls raid6_recov_avx2 = {
310 	.data2 = raid6_2data_recov_avx2,
311 	.datap = raid6_datap_recov_avx2,
312 	.valid = raid6_has_avx2,
313 #ifdef CONFIG_X86_64
314 	.name = "avx2x2",
315 #else
316 	.name = "avx2x1",
317 #endif
318 	.priority = 2,
319 };
320 
321 #else
322 #warning "your version of binutils lacks AVX2 support"
323 #endif
324