1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* -*- linux-c -*- ------------------------------------------------------- *
4 * Copyright 2002 H. Peter Anvin - All Rights Reserved
6 * ----------------------------------------------------------------------- */
11 * SSE-2 implementation of RAID-6 syndrome functions
15 #include <linux/raid/pq.h>
18 static const struct raid6_sse_constants {
20 } raid6_sse_constants __attribute__((aligned(16))) = {
21 { 0x1d1d1d1d1d1d1d1dULL, 0x1d1d1d1d1d1d1d1dULL },
24 static int raid6_have_sse2(void)
26 /* Not really boot_cpu but "all_cpus" */
27 return boot_cpu_has(X86_FEATURE_MMX) &&
28 boot_cpu_has(X86_FEATURE_FXSR) &&
29 boot_cpu_has(X86_FEATURE_XMM) &&
30 boot_cpu_has(X86_FEATURE_XMM2);
34 * Plain SSE2 implementation
36 static void raid6_sse21_gen_syndrome(int disks, size_t bytes, void **ptrs)
38 u8 **dptr = (u8 **)ptrs;
42 z0 = disks - 3; /* Highest data disk */
43 p = dptr[z0+1]; /* XOR parity */
44 q = dptr[z0+2]; /* RS syndrome */
48 asm volatile("movdqa %0,%%xmm0" : : "m" (raid6_sse_constants.x1d[0]));
49 asm volatile("pxor %xmm5,%xmm5"); /* Zero temp */
51 for ( d = 0 ; d < bytes ; d += 16 ) {
52 asm volatile("prefetchnta %0" : : "m" (dptr[z0][d]));
53 asm volatile("movdqa %0,%%xmm2" : : "m" (dptr[z0][d])); /* P[0] */
54 asm volatile("prefetchnta %0" : : "m" (dptr[z0-1][d]));
55 asm volatile("movdqa %xmm2,%xmm4"); /* Q[0] */
56 asm volatile("movdqa %0,%%xmm6" : : "m" (dptr[z0-1][d]));
57 for ( z = z0-2 ; z >= 0 ; z-- ) {
58 asm volatile("prefetchnta %0" : : "m" (dptr[z][d]));
59 asm volatile("pcmpgtb %xmm4,%xmm5");
60 asm volatile("paddb %xmm4,%xmm4");
61 asm volatile("pand %xmm0,%xmm5");
62 asm volatile("pxor %xmm5,%xmm4");
63 asm volatile("pxor %xmm5,%xmm5");
64 asm volatile("pxor %xmm6,%xmm2");
65 asm volatile("pxor %xmm6,%xmm4");
66 asm volatile("movdqa %0,%%xmm6" : : "m" (dptr[z][d]));
68 asm volatile("pcmpgtb %xmm4,%xmm5");
69 asm volatile("paddb %xmm4,%xmm4");
70 asm volatile("pand %xmm0,%xmm5");
71 asm volatile("pxor %xmm5,%xmm4");
72 asm volatile("pxor %xmm5,%xmm5");
73 asm volatile("pxor %xmm6,%xmm2");
74 asm volatile("pxor %xmm6,%xmm4");
76 asm volatile("movntdq %%xmm2,%0" : "=m" (p[d]));
77 asm volatile("pxor %xmm2,%xmm2");
78 asm volatile("movntdq %%xmm4,%0" : "=m" (q[d]));
79 asm volatile("pxor %xmm4,%xmm4");
82 asm volatile("sfence" : : : "memory");
87 static void raid6_sse21_xor_syndrome(int disks, int start, int stop,
88 size_t bytes, void **ptrs)
90 u8 **dptr = (u8 **)ptrs;
94 z0 = stop; /* P/Q right side optimization */
95 p = dptr[disks-2]; /* XOR parity */
96 q = dptr[disks-1]; /* RS syndrome */
100 asm volatile("movdqa %0,%%xmm0" : : "m" (raid6_sse_constants.x1d[0]));
102 for ( d = 0 ; d < bytes ; d += 16 ) {
103 asm volatile("movdqa %0,%%xmm4" :: "m" (dptr[z0][d]));
104 asm volatile("movdqa %0,%%xmm2" : : "m" (p[d]));
105 asm volatile("pxor %xmm4,%xmm2");
107 for ( z = z0-1 ; z >= start ; z-- ) {
108 asm volatile("pxor %xmm5,%xmm5");
109 asm volatile("pcmpgtb %xmm4,%xmm5");
110 asm volatile("paddb %xmm4,%xmm4");
111 asm volatile("pand %xmm0,%xmm5");
112 asm volatile("pxor %xmm5,%xmm4");
113 asm volatile("movdqa %0,%%xmm5" :: "m" (dptr[z][d]));
114 asm volatile("pxor %xmm5,%xmm2");
115 asm volatile("pxor %xmm5,%xmm4");
117 /* P/Q left side optimization */
118 for ( z = start-1 ; z >= 0 ; z-- ) {
119 asm volatile("pxor %xmm5,%xmm5");
120 asm volatile("pcmpgtb %xmm4,%xmm5");
121 asm volatile("paddb %xmm4,%xmm4");
122 asm volatile("pand %xmm0,%xmm5");
123 asm volatile("pxor %xmm5,%xmm4");
125 asm volatile("pxor %0,%%xmm4" : : "m" (q[d]));
126 /* Don't use movntdq for r/w memory area < cache line */
127 asm volatile("movdqa %%xmm4,%0" : "=m" (q[d]));
128 asm volatile("movdqa %%xmm2,%0" : "=m" (p[d]));
131 asm volatile("sfence" : : : "memory");
135 const struct raid6_calls raid6_sse2x1 = {
136 raid6_sse21_gen_syndrome,
137 raid6_sse21_xor_syndrome,
140 1 /* Has cache hints */
144 * Unrolled-by-2 SSE2 implementation
146 static void raid6_sse22_gen_syndrome(int disks, size_t bytes, void **ptrs)
148 u8 **dptr = (u8 **)ptrs;
152 z0 = disks - 3; /* Highest data disk */
153 p = dptr[z0+1]; /* XOR parity */
154 q = dptr[z0+2]; /* RS syndrome */
158 asm volatile("movdqa %0,%%xmm0" : : "m" (raid6_sse_constants.x1d[0]));
159 asm volatile("pxor %xmm5,%xmm5"); /* Zero temp */
160 asm volatile("pxor %xmm7,%xmm7"); /* Zero temp */
162 /* We uniformly assume a single prefetch covers at least 32 bytes */
163 for ( d = 0 ; d < bytes ; d += 32 ) {
164 asm volatile("prefetchnta %0" : : "m" (dptr[z0][d]));
165 asm volatile("movdqa %0,%%xmm2" : : "m" (dptr[z0][d])); /* P[0] */
166 asm volatile("movdqa %0,%%xmm3" : : "m" (dptr[z0][d+16])); /* P[1] */
167 asm volatile("movdqa %xmm2,%xmm4"); /* Q[0] */
168 asm volatile("movdqa %xmm3,%xmm6"); /* Q[1] */
169 for ( z = z0-1 ; z >= 0 ; z-- ) {
170 asm volatile("prefetchnta %0" : : "m" (dptr[z][d]));
171 asm volatile("pcmpgtb %xmm4,%xmm5");
172 asm volatile("pcmpgtb %xmm6,%xmm7");
173 asm volatile("paddb %xmm4,%xmm4");
174 asm volatile("paddb %xmm6,%xmm6");
175 asm volatile("pand %xmm0,%xmm5");
176 asm volatile("pand %xmm0,%xmm7");
177 asm volatile("pxor %xmm5,%xmm4");
178 asm volatile("pxor %xmm7,%xmm6");
179 asm volatile("movdqa %0,%%xmm5" : : "m" (dptr[z][d]));
180 asm volatile("movdqa %0,%%xmm7" : : "m" (dptr[z][d+16]));
181 asm volatile("pxor %xmm5,%xmm2");
182 asm volatile("pxor %xmm7,%xmm3");
183 asm volatile("pxor %xmm5,%xmm4");
184 asm volatile("pxor %xmm7,%xmm6");
185 asm volatile("pxor %xmm5,%xmm5");
186 asm volatile("pxor %xmm7,%xmm7");
188 asm volatile("movntdq %%xmm2,%0" : "=m" (p[d]));
189 asm volatile("movntdq %%xmm3,%0" : "=m" (p[d+16]));
190 asm volatile("movntdq %%xmm4,%0" : "=m" (q[d]));
191 asm volatile("movntdq %%xmm6,%0" : "=m" (q[d+16]));
194 asm volatile("sfence" : : : "memory");
198 static void raid6_sse22_xor_syndrome(int disks, int start, int stop,
199 size_t bytes, void **ptrs)
201 u8 **dptr = (u8 **)ptrs;
205 z0 = stop; /* P/Q right side optimization */
206 p = dptr[disks-2]; /* XOR parity */
207 q = dptr[disks-1]; /* RS syndrome */
211 asm volatile("movdqa %0,%%xmm0" : : "m" (raid6_sse_constants.x1d[0]));
213 for ( d = 0 ; d < bytes ; d += 32 ) {
214 asm volatile("movdqa %0,%%xmm4" :: "m" (dptr[z0][d]));
215 asm volatile("movdqa %0,%%xmm6" :: "m" (dptr[z0][d+16]));
216 asm volatile("movdqa %0,%%xmm2" : : "m" (p[d]));
217 asm volatile("movdqa %0,%%xmm3" : : "m" (p[d+16]));
218 asm volatile("pxor %xmm4,%xmm2");
219 asm volatile("pxor %xmm6,%xmm3");
221 for ( z = z0-1 ; z >= start ; z-- ) {
222 asm volatile("pxor %xmm5,%xmm5");
223 asm volatile("pxor %xmm7,%xmm7");
224 asm volatile("pcmpgtb %xmm4,%xmm5");
225 asm volatile("pcmpgtb %xmm6,%xmm7");
226 asm volatile("paddb %xmm4,%xmm4");
227 asm volatile("paddb %xmm6,%xmm6");
228 asm volatile("pand %xmm0,%xmm5");
229 asm volatile("pand %xmm0,%xmm7");
230 asm volatile("pxor %xmm5,%xmm4");
231 asm volatile("pxor %xmm7,%xmm6");
232 asm volatile("movdqa %0,%%xmm5" :: "m" (dptr[z][d]));
233 asm volatile("movdqa %0,%%xmm7" :: "m" (dptr[z][d+16]));
234 asm volatile("pxor %xmm5,%xmm2");
235 asm volatile("pxor %xmm7,%xmm3");
236 asm volatile("pxor %xmm5,%xmm4");
237 asm volatile("pxor %xmm7,%xmm6");
239 /* P/Q left side optimization */
240 for ( z = start-1 ; z >= 0 ; z-- ) {
241 asm volatile("pxor %xmm5,%xmm5");
242 asm volatile("pxor %xmm7,%xmm7");
243 asm volatile("pcmpgtb %xmm4,%xmm5");
244 asm volatile("pcmpgtb %xmm6,%xmm7");
245 asm volatile("paddb %xmm4,%xmm4");
246 asm volatile("paddb %xmm6,%xmm6");
247 asm volatile("pand %xmm0,%xmm5");
248 asm volatile("pand %xmm0,%xmm7");
249 asm volatile("pxor %xmm5,%xmm4");
250 asm volatile("pxor %xmm7,%xmm6");
252 asm volatile("pxor %0,%%xmm4" : : "m" (q[d]));
253 asm volatile("pxor %0,%%xmm6" : : "m" (q[d+16]));
254 /* Don't use movntdq for r/w memory area < cache line */
255 asm volatile("movdqa %%xmm4,%0" : "=m" (q[d]));
256 asm volatile("movdqa %%xmm6,%0" : "=m" (q[d+16]));
257 asm volatile("movdqa %%xmm2,%0" : "=m" (p[d]));
258 asm volatile("movdqa %%xmm3,%0" : "=m" (p[d+16]));
261 asm volatile("sfence" : : : "memory");
265 const struct raid6_calls raid6_sse2x2 = {
266 raid6_sse22_gen_syndrome,
267 raid6_sse22_xor_syndrome,
270 1 /* Has cache hints */
276 * Unrolled-by-4 SSE2 implementation
278 static void raid6_sse24_gen_syndrome(int disks, size_t bytes, void **ptrs)
280 u8 **dptr = (u8 **)ptrs;
284 z0 = disks - 3; /* Highest data disk */
285 p = dptr[z0+1]; /* XOR parity */
286 q = dptr[z0+2]; /* RS syndrome */
290 asm volatile("movdqa %0,%%xmm0" :: "m" (raid6_sse_constants.x1d[0]));
291 asm volatile("pxor %xmm2,%xmm2"); /* P[0] */
292 asm volatile("pxor %xmm3,%xmm3"); /* P[1] */
293 asm volatile("pxor %xmm4,%xmm4"); /* Q[0] */
294 asm volatile("pxor %xmm5,%xmm5"); /* Zero temp */
295 asm volatile("pxor %xmm6,%xmm6"); /* Q[1] */
296 asm volatile("pxor %xmm7,%xmm7"); /* Zero temp */
297 asm volatile("pxor %xmm10,%xmm10"); /* P[2] */
298 asm volatile("pxor %xmm11,%xmm11"); /* P[3] */
299 asm volatile("pxor %xmm12,%xmm12"); /* Q[2] */
300 asm volatile("pxor %xmm13,%xmm13"); /* Zero temp */
301 asm volatile("pxor %xmm14,%xmm14"); /* Q[3] */
302 asm volatile("pxor %xmm15,%xmm15"); /* Zero temp */
304 for ( d = 0 ; d < bytes ; d += 64 ) {
305 for ( z = z0 ; z >= 0 ; z-- ) {
306 /* The second prefetch seems to improve performance... */
307 asm volatile("prefetchnta %0" :: "m" (dptr[z][d]));
308 asm volatile("prefetchnta %0" :: "m" (dptr[z][d+32]));
309 asm volatile("pcmpgtb %xmm4,%xmm5");
310 asm volatile("pcmpgtb %xmm6,%xmm7");
311 asm volatile("pcmpgtb %xmm12,%xmm13");
312 asm volatile("pcmpgtb %xmm14,%xmm15");
313 asm volatile("paddb %xmm4,%xmm4");
314 asm volatile("paddb %xmm6,%xmm6");
315 asm volatile("paddb %xmm12,%xmm12");
316 asm volatile("paddb %xmm14,%xmm14");
317 asm volatile("pand %xmm0,%xmm5");
318 asm volatile("pand %xmm0,%xmm7");
319 asm volatile("pand %xmm0,%xmm13");
320 asm volatile("pand %xmm0,%xmm15");
321 asm volatile("pxor %xmm5,%xmm4");
322 asm volatile("pxor %xmm7,%xmm6");
323 asm volatile("pxor %xmm13,%xmm12");
324 asm volatile("pxor %xmm15,%xmm14");
325 asm volatile("movdqa %0,%%xmm5" :: "m" (dptr[z][d]));
326 asm volatile("movdqa %0,%%xmm7" :: "m" (dptr[z][d+16]));
327 asm volatile("movdqa %0,%%xmm13" :: "m" (dptr[z][d+32]));
328 asm volatile("movdqa %0,%%xmm15" :: "m" (dptr[z][d+48]));
329 asm volatile("pxor %xmm5,%xmm2");
330 asm volatile("pxor %xmm7,%xmm3");
331 asm volatile("pxor %xmm13,%xmm10");
332 asm volatile("pxor %xmm15,%xmm11");
333 asm volatile("pxor %xmm5,%xmm4");
334 asm volatile("pxor %xmm7,%xmm6");
335 asm volatile("pxor %xmm13,%xmm12");
336 asm volatile("pxor %xmm15,%xmm14");
337 asm volatile("pxor %xmm5,%xmm5");
338 asm volatile("pxor %xmm7,%xmm7");
339 asm volatile("pxor %xmm13,%xmm13");
340 asm volatile("pxor %xmm15,%xmm15");
342 asm volatile("movntdq %%xmm2,%0" : "=m" (p[d]));
343 asm volatile("pxor %xmm2,%xmm2");
344 asm volatile("movntdq %%xmm3,%0" : "=m" (p[d+16]));
345 asm volatile("pxor %xmm3,%xmm3");
346 asm volatile("movntdq %%xmm10,%0" : "=m" (p[d+32]));
347 asm volatile("pxor %xmm10,%xmm10");
348 asm volatile("movntdq %%xmm11,%0" : "=m" (p[d+48]));
349 asm volatile("pxor %xmm11,%xmm11");
350 asm volatile("movntdq %%xmm4,%0" : "=m" (q[d]));
351 asm volatile("pxor %xmm4,%xmm4");
352 asm volatile("movntdq %%xmm6,%0" : "=m" (q[d+16]));
353 asm volatile("pxor %xmm6,%xmm6");
354 asm volatile("movntdq %%xmm12,%0" : "=m" (q[d+32]));
355 asm volatile("pxor %xmm12,%xmm12");
356 asm volatile("movntdq %%xmm14,%0" : "=m" (q[d+48]));
357 asm volatile("pxor %xmm14,%xmm14");
360 asm volatile("sfence" : : : "memory");
364 static void raid6_sse24_xor_syndrome(int disks, int start, int stop,
365 size_t bytes, void **ptrs)
367 u8 **dptr = (u8 **)ptrs;
371 z0 = stop; /* P/Q right side optimization */
372 p = dptr[disks-2]; /* XOR parity */
373 q = dptr[disks-1]; /* RS syndrome */
377 asm volatile("movdqa %0,%%xmm0" :: "m" (raid6_sse_constants.x1d[0]));
379 for ( d = 0 ; d < bytes ; d += 64 ) {
380 asm volatile("movdqa %0,%%xmm4" :: "m" (dptr[z0][d]));
381 asm volatile("movdqa %0,%%xmm6" :: "m" (dptr[z0][d+16]));
382 asm volatile("movdqa %0,%%xmm12" :: "m" (dptr[z0][d+32]));
383 asm volatile("movdqa %0,%%xmm14" :: "m" (dptr[z0][d+48]));
384 asm volatile("movdqa %0,%%xmm2" : : "m" (p[d]));
385 asm volatile("movdqa %0,%%xmm3" : : "m" (p[d+16]));
386 asm volatile("movdqa %0,%%xmm10" : : "m" (p[d+32]));
387 asm volatile("movdqa %0,%%xmm11" : : "m" (p[d+48]));
388 asm volatile("pxor %xmm4,%xmm2");
389 asm volatile("pxor %xmm6,%xmm3");
390 asm volatile("pxor %xmm12,%xmm10");
391 asm volatile("pxor %xmm14,%xmm11");
393 for ( z = z0-1 ; z >= start ; z-- ) {
394 asm volatile("prefetchnta %0" :: "m" (dptr[z][d]));
395 asm volatile("prefetchnta %0" :: "m" (dptr[z][d+32]));
396 asm volatile("pxor %xmm5,%xmm5");
397 asm volatile("pxor %xmm7,%xmm7");
398 asm volatile("pxor %xmm13,%xmm13");
399 asm volatile("pxor %xmm15,%xmm15");
400 asm volatile("pcmpgtb %xmm4,%xmm5");
401 asm volatile("pcmpgtb %xmm6,%xmm7");
402 asm volatile("pcmpgtb %xmm12,%xmm13");
403 asm volatile("pcmpgtb %xmm14,%xmm15");
404 asm volatile("paddb %xmm4,%xmm4");
405 asm volatile("paddb %xmm6,%xmm6");
406 asm volatile("paddb %xmm12,%xmm12");
407 asm volatile("paddb %xmm14,%xmm14");
408 asm volatile("pand %xmm0,%xmm5");
409 asm volatile("pand %xmm0,%xmm7");
410 asm volatile("pand %xmm0,%xmm13");
411 asm volatile("pand %xmm0,%xmm15");
412 asm volatile("pxor %xmm5,%xmm4");
413 asm volatile("pxor %xmm7,%xmm6");
414 asm volatile("pxor %xmm13,%xmm12");
415 asm volatile("pxor %xmm15,%xmm14");
416 asm volatile("movdqa %0,%%xmm5" :: "m" (dptr[z][d]));
417 asm volatile("movdqa %0,%%xmm7" :: "m" (dptr[z][d+16]));
418 asm volatile("movdqa %0,%%xmm13" :: "m" (dptr[z][d+32]));
419 asm volatile("movdqa %0,%%xmm15" :: "m" (dptr[z][d+48]));
420 asm volatile("pxor %xmm5,%xmm2");
421 asm volatile("pxor %xmm7,%xmm3");
422 asm volatile("pxor %xmm13,%xmm10");
423 asm volatile("pxor %xmm15,%xmm11");
424 asm volatile("pxor %xmm5,%xmm4");
425 asm volatile("pxor %xmm7,%xmm6");
426 asm volatile("pxor %xmm13,%xmm12");
427 asm volatile("pxor %xmm15,%xmm14");
429 asm volatile("prefetchnta %0" :: "m" (q[d]));
430 asm volatile("prefetchnta %0" :: "m" (q[d+32]));
431 /* P/Q left side optimization */
432 for ( z = start-1 ; z >= 0 ; z-- ) {
433 asm volatile("pxor %xmm5,%xmm5");
434 asm volatile("pxor %xmm7,%xmm7");
435 asm volatile("pxor %xmm13,%xmm13");
436 asm volatile("pxor %xmm15,%xmm15");
437 asm volatile("pcmpgtb %xmm4,%xmm5");
438 asm volatile("pcmpgtb %xmm6,%xmm7");
439 asm volatile("pcmpgtb %xmm12,%xmm13");
440 asm volatile("pcmpgtb %xmm14,%xmm15");
441 asm volatile("paddb %xmm4,%xmm4");
442 asm volatile("paddb %xmm6,%xmm6");
443 asm volatile("paddb %xmm12,%xmm12");
444 asm volatile("paddb %xmm14,%xmm14");
445 asm volatile("pand %xmm0,%xmm5");
446 asm volatile("pand %xmm0,%xmm7");
447 asm volatile("pand %xmm0,%xmm13");
448 asm volatile("pand %xmm0,%xmm15");
449 asm volatile("pxor %xmm5,%xmm4");
450 asm volatile("pxor %xmm7,%xmm6");
451 asm volatile("pxor %xmm13,%xmm12");
452 asm volatile("pxor %xmm15,%xmm14");
454 asm volatile("movntdq %%xmm2,%0" : "=m" (p[d]));
455 asm volatile("movntdq %%xmm3,%0" : "=m" (p[d+16]));
456 asm volatile("movntdq %%xmm10,%0" : "=m" (p[d+32]));
457 asm volatile("movntdq %%xmm11,%0" : "=m" (p[d+48]));
458 asm volatile("pxor %0,%%xmm4" : : "m" (q[d]));
459 asm volatile("pxor %0,%%xmm6" : : "m" (q[d+16]));
460 asm volatile("pxor %0,%%xmm12" : : "m" (q[d+32]));
461 asm volatile("pxor %0,%%xmm14" : : "m" (q[d+48]));
462 asm volatile("movntdq %%xmm4,%0" : "=m" (q[d]));
463 asm volatile("movntdq %%xmm6,%0" : "=m" (q[d+16]));
464 asm volatile("movntdq %%xmm12,%0" : "=m" (q[d+32]));
465 asm volatile("movntdq %%xmm14,%0" : "=m" (q[d+48]));
467 asm volatile("sfence" : : : "memory");
472 const struct raid6_calls raid6_sse2x4 = {
473 raid6_sse24_gen_syndrome,
474 raid6_sse24_xor_syndrome,
477 1 /* Has cache hints */
480 #endif /* CONFIG_X86_64 */