1/*
2 * Copyright (C) 2012 Intel Corporation
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; version 2
7 * of the License.
8 */
9
10#ifdef CONFIG_AS_SSSE3
11
12#include <linux/raid/pq.h>
13#include "x86.h"
14
15static int raid6_has_ssse3(void)
16{
17	return boot_cpu_has(X86_FEATURE_XMM) &&
18		boot_cpu_has(X86_FEATURE_XMM2) &&
19		boot_cpu_has(X86_FEATURE_SSSE3);
20}
21
22static void raid6_2data_recov_ssse3(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	static const u8 __aligned(16) x0f[16] = {
29		 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f,
30		 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f};
31
32	p = (u8 *)ptrs[disks-2];
33	q = (u8 *)ptrs[disks-1];
34
35	/* Compute syndrome with zero for the missing data pages
36	   Use the dead data pages as temporary storage for
37	   delta p and delta q */
38	dp = (u8 *)ptrs[faila];
39	ptrs[faila] = (void *)raid6_empty_zero_page;
40	ptrs[disks-2] = dp;
41	dq = (u8 *)ptrs[failb];
42	ptrs[failb] = (void *)raid6_empty_zero_page;
43	ptrs[disks-1] = dq;
44
45	raid6_call.gen_syndrome(disks, bytes, ptrs);
46
47	/* Restore pointer table */
48	ptrs[faila]   = dp;
49	ptrs[failb]   = dq;
50	ptrs[disks-2] = p;
51	ptrs[disks-1] = q;
52
53	/* Now, pick the proper data tables */
54	pbmul = raid6_vgfmul[raid6_gfexi[failb-faila]];
55	qmul  = raid6_vgfmul[raid6_gfinv[raid6_gfexp[faila] ^
56		raid6_gfexp[failb]]];
57
58	kernel_fpu_begin();
59
60	asm volatile("movdqa %0,%%xmm7" : : "m" (x0f[0]));
61
62#ifdef CONFIG_X86_64
63	asm volatile("movdqa %0,%%xmm6" : : "m" (qmul[0]));
64	asm volatile("movdqa %0,%%xmm14" : : "m" (pbmul[0]));
65	asm volatile("movdqa %0,%%xmm15" : : "m" (pbmul[16]));
66#endif
67
68	/* Now do it... */
69	while (bytes) {
70#ifdef CONFIG_X86_64
71		/* xmm6, xmm14, xmm15 */
72
73		asm volatile("movdqa %0,%%xmm1" : : "m" (q[0]));
74		asm volatile("movdqa %0,%%xmm9" : : "m" (q[16]));
75		asm volatile("movdqa %0,%%xmm0" : : "m" (p[0]));
76		asm volatile("movdqa %0,%%xmm8" : : "m" (p[16]));
77		asm volatile("pxor   %0,%%xmm1" : : "m" (dq[0]));
78		asm volatile("pxor   %0,%%xmm9" : : "m" (dq[16]));
79		asm volatile("pxor   %0,%%xmm0" : : "m" (dp[0]));
80		asm volatile("pxor   %0,%%xmm8" : : "m" (dp[16]));
81
82		/* xmm0/8 = px */
83
84		asm volatile("movdqa %xmm6,%xmm4");
85		asm volatile("movdqa %0,%%xmm5" : : "m" (qmul[16]));
86		asm volatile("movdqa %xmm6,%xmm12");
87		asm volatile("movdqa %xmm5,%xmm13");
88		asm volatile("movdqa %xmm1,%xmm3");
89		asm volatile("movdqa %xmm9,%xmm11");
90		asm volatile("movdqa %xmm0,%xmm2"); /* xmm2/10 = px */
91		asm volatile("movdqa %xmm8,%xmm10");
92		asm volatile("psraw  $4,%xmm1");
93		asm volatile("psraw  $4,%xmm9");
94		asm volatile("pand   %xmm7,%xmm3");
95		asm volatile("pand   %xmm7,%xmm11");
96		asm volatile("pand   %xmm7,%xmm1");
97		asm volatile("pand   %xmm7,%xmm9");
98		asm volatile("pshufb %xmm3,%xmm4");
99		asm volatile("pshufb %xmm11,%xmm12");
100		asm volatile("pshufb %xmm1,%xmm5");
101		asm volatile("pshufb %xmm9,%xmm13");
102		asm volatile("pxor   %xmm4,%xmm5");
103		asm volatile("pxor   %xmm12,%xmm13");
104
105		/* xmm5/13 = qx */
106
107		asm volatile("movdqa %xmm14,%xmm4");
108		asm volatile("movdqa %xmm15,%xmm1");
109		asm volatile("movdqa %xmm14,%xmm12");
110		asm volatile("movdqa %xmm15,%xmm9");
111		asm volatile("movdqa %xmm2,%xmm3");
112		asm volatile("movdqa %xmm10,%xmm11");
113		asm volatile("psraw  $4,%xmm2");
114		asm volatile("psraw  $4,%xmm10");
115		asm volatile("pand   %xmm7,%xmm3");
116		asm volatile("pand   %xmm7,%xmm11");
117		asm volatile("pand   %xmm7,%xmm2");
118		asm volatile("pand   %xmm7,%xmm10");
119		asm volatile("pshufb %xmm3,%xmm4");
120		asm volatile("pshufb %xmm11,%xmm12");
121		asm volatile("pshufb %xmm2,%xmm1");
122		asm volatile("pshufb %xmm10,%xmm9");
123		asm volatile("pxor   %xmm4,%xmm1");
124		asm volatile("pxor   %xmm12,%xmm9");
125
126		/* xmm1/9 = pbmul[px] */
127		asm volatile("pxor   %xmm5,%xmm1");
128		asm volatile("pxor   %xmm13,%xmm9");
129		/* xmm1/9 = db = DQ */
130		asm volatile("movdqa %%xmm1,%0" : "=m" (dq[0]));
131		asm volatile("movdqa %%xmm9,%0" : "=m" (dq[16]));
132
133		asm volatile("pxor   %xmm1,%xmm0");
134		asm volatile("pxor   %xmm9,%xmm8");
135		asm volatile("movdqa %%xmm0,%0" : "=m" (dp[0]));
136		asm volatile("movdqa %%xmm8,%0" : "=m" (dp[16]));
137
138		bytes -= 32;
139		p += 32;
140		q += 32;
141		dp += 32;
142		dq += 32;
143#else
144		asm volatile("movdqa %0,%%xmm1" : : "m" (*q));
145		asm volatile("movdqa %0,%%xmm0" : : "m" (*p));
146		asm volatile("pxor   %0,%%xmm1" : : "m" (*dq));
147		asm volatile("pxor   %0,%%xmm0" : : "m" (*dp));
148
149		/* 1 = dq ^ q
150		 * 0 = dp ^ p
151		 */
152		asm volatile("movdqa %0,%%xmm4" : : "m" (qmul[0]));
153		asm volatile("movdqa %0,%%xmm5" : : "m" (qmul[16]));
154
155		asm volatile("movdqa %xmm1,%xmm3");
156		asm volatile("psraw  $4,%xmm1");
157		asm volatile("pand   %xmm7,%xmm3");
158		asm volatile("pand   %xmm7,%xmm1");
159		asm volatile("pshufb %xmm3,%xmm4");
160		asm volatile("pshufb %xmm1,%xmm5");
161		asm volatile("pxor   %xmm4,%xmm5");
162
163		asm volatile("movdqa %xmm0,%xmm2"); /* xmm2 = px */
164
165		/* xmm5 = qx */
166
167		asm volatile("movdqa %0,%%xmm4" : : "m" (pbmul[0]));
168		asm volatile("movdqa %0,%%xmm1" : : "m" (pbmul[16]));
169		asm volatile("movdqa %xmm2,%xmm3");
170		asm volatile("psraw  $4,%xmm2");
171		asm volatile("pand   %xmm7,%xmm3");
172		asm volatile("pand   %xmm7,%xmm2");
173		asm volatile("pshufb %xmm3,%xmm4");
174		asm volatile("pshufb %xmm2,%xmm1");
175		asm volatile("pxor   %xmm4,%xmm1");
176
177		/* xmm1 = pbmul[px] */
178		asm volatile("pxor   %xmm5,%xmm1");
179		/* xmm1 = db = DQ */
180		asm volatile("movdqa %%xmm1,%0" : "=m" (*dq));
181
182		asm volatile("pxor   %xmm1,%xmm0");
183		asm volatile("movdqa %%xmm0,%0" : "=m" (*dp));
184
185		bytes -= 16;
186		p += 16;
187		q += 16;
188		dp += 16;
189		dq += 16;
190#endif
191	}
192
193	kernel_fpu_end();
194}
195
196
197static void raid6_datap_recov_ssse3(int disks, size_t bytes, int faila,
198		void **ptrs)
199{
200	u8 *p, *q, *dq;
201	const u8 *qmul;		/* Q multiplier table */
202	static const u8 __aligned(16) x0f[16] = {
203		 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f,
204		 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f};
205
206	p = (u8 *)ptrs[disks-2];
207	q = (u8 *)ptrs[disks-1];
208
209	/* Compute syndrome with zero for the missing data page
210	   Use the dead data page as temporary storage for delta q */
211	dq = (u8 *)ptrs[faila];
212	ptrs[faila] = (void *)raid6_empty_zero_page;
213	ptrs[disks-1] = dq;
214
215	raid6_call.gen_syndrome(disks, bytes, ptrs);
216
217	/* Restore pointer table */
218	ptrs[faila]   = dq;
219	ptrs[disks-1] = q;
220
221	/* Now, pick the proper data tables */
222	qmul  = raid6_vgfmul[raid6_gfinv[raid6_gfexp[faila]]];
223
224	kernel_fpu_begin();
225
226	asm volatile("movdqa %0, %%xmm7" : : "m" (x0f[0]));
227
228	while (bytes) {
229#ifdef CONFIG_X86_64
230		asm volatile("movdqa %0, %%xmm3" : : "m" (dq[0]));
231		asm volatile("movdqa %0, %%xmm4" : : "m" (dq[16]));
232		asm volatile("pxor %0, %%xmm3" : : "m" (q[0]));
233		asm volatile("movdqa %0, %%xmm0" : : "m" (qmul[0]));
234
235		/* xmm3 = q[0] ^ dq[0] */
236
237		asm volatile("pxor %0, %%xmm4" : : "m" (q[16]));
238		asm volatile("movdqa %0, %%xmm1" : : "m" (qmul[16]));
239
240		/* xmm4 = q[16] ^ dq[16] */
241
242		asm volatile("movdqa %xmm3, %xmm6");
243		asm volatile("movdqa %xmm4, %xmm8");
244
245		/* xmm4 = xmm8 = q[16] ^ dq[16] */
246
247		asm volatile("psraw $4, %xmm3");
248		asm volatile("pand %xmm7, %xmm6");
249		asm volatile("pand %xmm7, %xmm3");
250		asm volatile("pshufb %xmm6, %xmm0");
251		asm volatile("pshufb %xmm3, %xmm1");
252		asm volatile("movdqa %0, %%xmm10" : : "m" (qmul[0]));
253		asm volatile("pxor %xmm0, %xmm1");
254		asm volatile("movdqa %0, %%xmm11" : : "m" (qmul[16]));
255
256		/* xmm1 = qmul[q[0] ^ dq[0]] */
257
258		asm volatile("psraw $4, %xmm4");
259		asm volatile("pand %xmm7, %xmm8");
260		asm volatile("pand %xmm7, %xmm4");
261		asm volatile("pshufb %xmm8, %xmm10");
262		asm volatile("pshufb %xmm4, %xmm11");
263		asm volatile("movdqa %0, %%xmm2" : : "m" (p[0]));
264		asm volatile("pxor %xmm10, %xmm11");
265		asm volatile("movdqa %0, %%xmm12" : : "m" (p[16]));
266
267		/* xmm11 = qmul[q[16] ^ dq[16]] */
268
269		asm volatile("pxor %xmm1, %xmm2");
270
271		/* xmm2 = p[0] ^ qmul[q[0] ^ dq[0]] */
272
273		asm volatile("pxor %xmm11, %xmm12");
274
275		/* xmm12 = p[16] ^ qmul[q[16] ^ dq[16]] */
276
277		asm volatile("movdqa %%xmm1, %0" : "=m" (dq[0]));
278		asm volatile("movdqa %%xmm11, %0" : "=m" (dq[16]));
279
280		asm volatile("movdqa %%xmm2, %0" : "=m" (p[0]));
281		asm volatile("movdqa %%xmm12, %0" : "=m" (p[16]));
282
283		bytes -= 32;
284		p += 32;
285		q += 32;
286		dq += 32;
287
288#else
289		asm volatile("movdqa %0, %%xmm3" : : "m" (dq[0]));
290		asm volatile("movdqa %0, %%xmm0" : : "m" (qmul[0]));
291		asm volatile("pxor %0, %%xmm3" : : "m" (q[0]));
292		asm volatile("movdqa %0, %%xmm1" : : "m" (qmul[16]));
293
294		/* xmm3 = *q ^ *dq */
295
296		asm volatile("movdqa %xmm3, %xmm6");
297		asm volatile("movdqa %0, %%xmm2" : : "m" (p[0]));
298		asm volatile("psraw $4, %xmm3");
299		asm volatile("pand %xmm7, %xmm6");
300		asm volatile("pand %xmm7, %xmm3");
301		asm volatile("pshufb %xmm6, %xmm0");
302		asm volatile("pshufb %xmm3, %xmm1");
303		asm volatile("pxor %xmm0, %xmm1");
304
305		/* xmm1 = qmul[*q ^ *dq */
306
307		asm volatile("pxor %xmm1, %xmm2");
308
309		/* xmm2 = *p ^ qmul[*q ^ *dq] */
310
311		asm volatile("movdqa %%xmm1, %0" : "=m" (dq[0]));
312		asm volatile("movdqa %%xmm2, %0" : "=m" (p[0]));
313
314		bytes -= 16;
315		p += 16;
316		q += 16;
317		dq += 16;
318#endif
319	}
320
321	kernel_fpu_end();
322}
323
324const struct raid6_recov_calls raid6_recov_ssse3 = {
325	.data2 = raid6_2data_recov_ssse3,
326	.datap = raid6_datap_recov_ssse3,
327	.valid = raid6_has_ssse3,
328#ifdef CONFIG_X86_64
329	.name = "ssse3x2",
330#else
331	.name = "ssse3x1",
332#endif
333	.priority = 1,
334};
335
336#else
337#warning "your version of binutils lacks SSSE3 support"
338#endif
339