Loading...
1/*
2 * This file is subject to the terms and conditions of the GNU General Public
3 * License. See the file "COPYING" in the main directory of this archive
4 * for more details.
5 *
6 * Copyright (C) 2003, 04, 05 Ralf Baechle (ralf@linux-mips.org)
7 * Copyright (C) 2007 Maciej W. Rozycki
8 * Copyright (C) 2008 Thiemo Seufer
9 * Copyright (C) 2012 MIPS Technologies, Inc.
10 */
11#include <linux/kernel.h>
12#include <linux/sched.h>
13#include <linux/smp.h>
14#include <linux/mm.h>
15#include <linux/proc_fs.h>
16
17#include <asm/bugs.h>
18#include <asm/cacheops.h>
19#include <asm/cpu-type.h>
20#include <asm/inst.h>
21#include <asm/io.h>
22#include <asm/page.h>
23#include <asm/pgtable.h>
24#include <asm/prefetch.h>
25#include <asm/bootinfo.h>
26#include <asm/mipsregs.h>
27#include <asm/mmu_context.h>
28#include <asm/cpu.h>
29#include <asm/war.h>
30
31#ifdef CONFIG_SIBYTE_DMA_PAGEOPS
32#include <asm/sibyte/sb1250.h>
33#include <asm/sibyte/sb1250_regs.h>
34#include <asm/sibyte/sb1250_dma.h>
35#endif
36
37#include <asm/uasm.h>
38
39/* Registers used in the assembled routines. */
40#define ZERO 0
41#define AT 2
42#define A0 4
43#define A1 5
44#define A2 6
45#define T0 8
46#define T1 9
47#define T2 10
48#define T3 11
49#define T9 25
50#define RA 31
51
52/* Handle labels (which must be positive integers). */
53enum label_id {
54 label_clear_nopref = 1,
55 label_clear_pref,
56 label_copy_nopref,
57 label_copy_pref_both,
58 label_copy_pref_store,
59};
60
61UASM_L_LA(_clear_nopref)
62UASM_L_LA(_clear_pref)
63UASM_L_LA(_copy_nopref)
64UASM_L_LA(_copy_pref_both)
65UASM_L_LA(_copy_pref_store)
66
67/* We need one branch and therefore one relocation per target label. */
68static struct uasm_label labels[5];
69static struct uasm_reloc relocs[5];
70
71#define cpu_is_r4600_v1_x() ((read_c0_prid() & 0xfffffff0) == 0x00002010)
72#define cpu_is_r4600_v2_x() ((read_c0_prid() & 0xfffffff0) == 0x00002020)
73
74/*
75 * R6 has a limited offset of the pref instruction.
76 * Skip it if the offset is more than 9 bits.
77 */
78#define _uasm_i_pref(a, b, c, d) \
79do { \
80 if (cpu_has_mips_r6) { \
81 if (c <= 0xff && c >= -0x100) \
82 uasm_i_pref(a, b, c, d);\
83 } else { \
84 uasm_i_pref(a, b, c, d); \
85 } \
86} while(0)
87
88static int pref_bias_clear_store;
89static int pref_bias_copy_load;
90static int pref_bias_copy_store;
91
92static u32 pref_src_mode;
93static u32 pref_dst_mode;
94
95static int clear_word_size;
96static int copy_word_size;
97
98static int half_clear_loop_size;
99static int half_copy_loop_size;
100
101static int cache_line_size;
102#define cache_line_mask() (cache_line_size - 1)
103
104static inline void
105pg_addiu(u32 **buf, unsigned int reg1, unsigned int reg2, unsigned int off)
106{
107 if (cpu_has_64bit_gp_regs && DADDI_WAR && r4k_daddiu_bug()) {
108 if (off > 0x7fff) {
109 uasm_i_lui(buf, T9, uasm_rel_hi(off));
110 uasm_i_addiu(buf, T9, T9, uasm_rel_lo(off));
111 } else
112 uasm_i_addiu(buf, T9, ZERO, off);
113 uasm_i_daddu(buf, reg1, reg2, T9);
114 } else {
115 if (off > 0x7fff) {
116 uasm_i_lui(buf, T9, uasm_rel_hi(off));
117 uasm_i_addiu(buf, T9, T9, uasm_rel_lo(off));
118 UASM_i_ADDU(buf, reg1, reg2, T9);
119 } else
120 UASM_i_ADDIU(buf, reg1, reg2, off);
121 }
122}
123
124static void set_prefetch_parameters(void)
125{
126 if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg)
127 clear_word_size = 8;
128 else
129 clear_word_size = 4;
130
131 if (cpu_has_64bit_gp_regs)
132 copy_word_size = 8;
133 else
134 copy_word_size = 4;
135
136 /*
137 * The pref's used here are using "streaming" hints, which cause the
138 * copied data to be kicked out of the cache sooner. A page copy often
139 * ends up copying a lot more data than is commonly used, so this seems
140 * to make sense in terms of reducing cache pollution, but I've no real
141 * performance data to back this up.
142 */
143 if (cpu_has_prefetch) {
144 /*
145 * XXX: Most prefetch bias values in here are based on
146 * guesswork.
147 */
148 cache_line_size = cpu_dcache_line_size();
149 switch (current_cpu_type()) {
150 case CPU_R5500:
151 case CPU_TX49XX:
152 /* These processors only support the Pref_Load. */
153 pref_bias_copy_load = 256;
154 break;
155
156 case CPU_R10000:
157 case CPU_R12000:
158 case CPU_R14000:
159 case CPU_R16000:
160 /*
161 * Those values have been experimentally tuned for an
162 * Origin 200.
163 */
164 pref_bias_clear_store = 512;
165 pref_bias_copy_load = 256;
166 pref_bias_copy_store = 256;
167 pref_src_mode = Pref_LoadStreamed;
168 pref_dst_mode = Pref_StoreStreamed;
169 break;
170
171 case CPU_SB1:
172 case CPU_SB1A:
173 pref_bias_clear_store = 128;
174 pref_bias_copy_load = 128;
175 pref_bias_copy_store = 128;
176 /*
177 * SB1 pass1 Pref_LoadStreamed/Pref_StoreStreamed
178 * hints are broken.
179 */
180 if (current_cpu_type() == CPU_SB1 &&
181 (current_cpu_data.processor_id & 0xff) < 0x02) {
182 pref_src_mode = Pref_Load;
183 pref_dst_mode = Pref_Store;
184 } else {
185 pref_src_mode = Pref_LoadStreamed;
186 pref_dst_mode = Pref_StoreStreamed;
187 }
188 break;
189
190 case CPU_LOONGSON3:
191 /* Loongson-3 only support the Pref_Load/Pref_Store. */
192 pref_bias_clear_store = 128;
193 pref_bias_copy_load = 128;
194 pref_bias_copy_store = 128;
195 pref_src_mode = Pref_Load;
196 pref_dst_mode = Pref_Store;
197 break;
198
199 default:
200 pref_bias_clear_store = 128;
201 pref_bias_copy_load = 256;
202 pref_bias_copy_store = 128;
203 pref_src_mode = Pref_LoadStreamed;
204 if (cpu_has_mips_r6)
205 /*
206 * Bit 30 (Pref_PrepareForStore) has been
207 * removed from MIPS R6. Use bit 5
208 * (Pref_StoreStreamed).
209 */
210 pref_dst_mode = Pref_StoreStreamed;
211 else
212 pref_dst_mode = Pref_PrepareForStore;
213 break;
214 }
215 } else {
216 if (cpu_has_cache_cdex_s)
217 cache_line_size = cpu_scache_line_size();
218 else if (cpu_has_cache_cdex_p)
219 cache_line_size = cpu_dcache_line_size();
220 }
221 /*
222 * Too much unrolling will overflow the available space in
223 * clear_space_array / copy_page_array.
224 */
225 half_clear_loop_size = min(16 * clear_word_size,
226 max(cache_line_size >> 1,
227 4 * clear_word_size));
228 half_copy_loop_size = min(16 * copy_word_size,
229 max(cache_line_size >> 1,
230 4 * copy_word_size));
231}
232
233static void build_clear_store(u32 **buf, int off)
234{
235 if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg) {
236 uasm_i_sd(buf, ZERO, off, A0);
237 } else {
238 uasm_i_sw(buf, ZERO, off, A0);
239 }
240}
241
242static inline void build_clear_pref(u32 **buf, int off)
243{
244 if (off & cache_line_mask())
245 return;
246
247 if (pref_bias_clear_store) {
248 _uasm_i_pref(buf, pref_dst_mode, pref_bias_clear_store + off,
249 A0);
250 } else if (cache_line_size == (half_clear_loop_size << 1)) {
251 if (cpu_has_cache_cdex_s) {
252 uasm_i_cache(buf, Create_Dirty_Excl_SD, off, A0);
253 } else if (cpu_has_cache_cdex_p) {
254 if (R4600_V1_HIT_CACHEOP_WAR && cpu_is_r4600_v1_x()) {
255 uasm_i_nop(buf);
256 uasm_i_nop(buf);
257 uasm_i_nop(buf);
258 uasm_i_nop(buf);
259 }
260
261 if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
262 uasm_i_lw(buf, ZERO, ZERO, AT);
263
264 uasm_i_cache(buf, Create_Dirty_Excl_D, off, A0);
265 }
266 }
267}
268
269extern u32 __clear_page_start;
270extern u32 __clear_page_end;
271extern u32 __copy_page_start;
272extern u32 __copy_page_end;
273
274void build_clear_page(void)
275{
276 int off;
277 u32 *buf = &__clear_page_start;
278 struct uasm_label *l = labels;
279 struct uasm_reloc *r = relocs;
280 int i;
281 static atomic_t run_once = ATOMIC_INIT(0);
282
283 if (atomic_xchg(&run_once, 1)) {
284 return;
285 }
286
287 memset(labels, 0, sizeof(labels));
288 memset(relocs, 0, sizeof(relocs));
289
290 set_prefetch_parameters();
291
292 /*
293 * This algorithm makes the following assumptions:
294 * - The prefetch bias is a multiple of 2 words.
295 * - The prefetch bias is less than one page.
296 */
297 BUG_ON(pref_bias_clear_store % (2 * clear_word_size));
298 BUG_ON(PAGE_SIZE < pref_bias_clear_store);
299
300 off = PAGE_SIZE - pref_bias_clear_store;
301 if (off > 0xffff || !pref_bias_clear_store)
302 pg_addiu(&buf, A2, A0, off);
303 else
304 uasm_i_ori(&buf, A2, A0, off);
305
306 if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
307 uasm_i_lui(&buf, AT, uasm_rel_hi(0xa0000000));
308
309 off = cache_line_size ? min(8, pref_bias_clear_store / cache_line_size)
310 * cache_line_size : 0;
311 while (off) {
312 build_clear_pref(&buf, -off);
313 off -= cache_line_size;
314 }
315 uasm_l_clear_pref(&l, buf);
316 do {
317 build_clear_pref(&buf, off);
318 build_clear_store(&buf, off);
319 off += clear_word_size;
320 } while (off < half_clear_loop_size);
321 pg_addiu(&buf, A0, A0, 2 * off);
322 off = -off;
323 do {
324 build_clear_pref(&buf, off);
325 if (off == -clear_word_size)
326 uasm_il_bne(&buf, &r, A0, A2, label_clear_pref);
327 build_clear_store(&buf, off);
328 off += clear_word_size;
329 } while (off < 0);
330
331 if (pref_bias_clear_store) {
332 pg_addiu(&buf, A2, A0, pref_bias_clear_store);
333 uasm_l_clear_nopref(&l, buf);
334 off = 0;
335 do {
336 build_clear_store(&buf, off);
337 off += clear_word_size;
338 } while (off < half_clear_loop_size);
339 pg_addiu(&buf, A0, A0, 2 * off);
340 off = -off;
341 do {
342 if (off == -clear_word_size)
343 uasm_il_bne(&buf, &r, A0, A2,
344 label_clear_nopref);
345 build_clear_store(&buf, off);
346 off += clear_word_size;
347 } while (off < 0);
348 }
349
350 uasm_i_jr(&buf, RA);
351 uasm_i_nop(&buf);
352
353 BUG_ON(buf > &__clear_page_end);
354
355 uasm_resolve_relocs(relocs, labels);
356
357 pr_debug("Synthesized clear page handler (%u instructions).\n",
358 (u32)(buf - &__clear_page_start));
359
360 pr_debug("\t.set push\n");
361 pr_debug("\t.set noreorder\n");
362 for (i = 0; i < (buf - &__clear_page_start); i++)
363 pr_debug("\t.word 0x%08x\n", (&__clear_page_start)[i]);
364 pr_debug("\t.set pop\n");
365}
366
367static void build_copy_load(u32 **buf, int reg, int off)
368{
369 if (cpu_has_64bit_gp_regs) {
370 uasm_i_ld(buf, reg, off, A1);
371 } else {
372 uasm_i_lw(buf, reg, off, A1);
373 }
374}
375
376static void build_copy_store(u32 **buf, int reg, int off)
377{
378 if (cpu_has_64bit_gp_regs) {
379 uasm_i_sd(buf, reg, off, A0);
380 } else {
381 uasm_i_sw(buf, reg, off, A0);
382 }
383}
384
385static inline void build_copy_load_pref(u32 **buf, int off)
386{
387 if (off & cache_line_mask())
388 return;
389
390 if (pref_bias_copy_load)
391 _uasm_i_pref(buf, pref_src_mode, pref_bias_copy_load + off, A1);
392}
393
394static inline void build_copy_store_pref(u32 **buf, int off)
395{
396 if (off & cache_line_mask())
397 return;
398
399 if (pref_bias_copy_store) {
400 _uasm_i_pref(buf, pref_dst_mode, pref_bias_copy_store + off,
401 A0);
402 } else if (cache_line_size == (half_copy_loop_size << 1)) {
403 if (cpu_has_cache_cdex_s) {
404 uasm_i_cache(buf, Create_Dirty_Excl_SD, off, A0);
405 } else if (cpu_has_cache_cdex_p) {
406 if (R4600_V1_HIT_CACHEOP_WAR && cpu_is_r4600_v1_x()) {
407 uasm_i_nop(buf);
408 uasm_i_nop(buf);
409 uasm_i_nop(buf);
410 uasm_i_nop(buf);
411 }
412
413 if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
414 uasm_i_lw(buf, ZERO, ZERO, AT);
415
416 uasm_i_cache(buf, Create_Dirty_Excl_D, off, A0);
417 }
418 }
419}
420
421void build_copy_page(void)
422{
423 int off;
424 u32 *buf = &__copy_page_start;
425 struct uasm_label *l = labels;
426 struct uasm_reloc *r = relocs;
427 int i;
428 static atomic_t run_once = ATOMIC_INIT(0);
429
430 if (atomic_xchg(&run_once, 1)) {
431 return;
432 }
433
434 memset(labels, 0, sizeof(labels));
435 memset(relocs, 0, sizeof(relocs));
436
437 set_prefetch_parameters();
438
439 /*
440 * This algorithm makes the following assumptions:
441 * - All prefetch biases are multiples of 8 words.
442 * - The prefetch biases are less than one page.
443 * - The store prefetch bias isn't greater than the load
444 * prefetch bias.
445 */
446 BUG_ON(pref_bias_copy_load % (8 * copy_word_size));
447 BUG_ON(pref_bias_copy_store % (8 * copy_word_size));
448 BUG_ON(PAGE_SIZE < pref_bias_copy_load);
449 BUG_ON(pref_bias_copy_store > pref_bias_copy_load);
450
451 off = PAGE_SIZE - pref_bias_copy_load;
452 if (off > 0xffff || !pref_bias_copy_load)
453 pg_addiu(&buf, A2, A0, off);
454 else
455 uasm_i_ori(&buf, A2, A0, off);
456
457 if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
458 uasm_i_lui(&buf, AT, uasm_rel_hi(0xa0000000));
459
460 off = cache_line_size ? min(8, pref_bias_copy_load / cache_line_size) *
461 cache_line_size : 0;
462 while (off) {
463 build_copy_load_pref(&buf, -off);
464 off -= cache_line_size;
465 }
466 off = cache_line_size ? min(8, pref_bias_copy_store / cache_line_size) *
467 cache_line_size : 0;
468 while (off) {
469 build_copy_store_pref(&buf, -off);
470 off -= cache_line_size;
471 }
472 uasm_l_copy_pref_both(&l, buf);
473 do {
474 build_copy_load_pref(&buf, off);
475 build_copy_load(&buf, T0, off);
476 build_copy_load_pref(&buf, off + copy_word_size);
477 build_copy_load(&buf, T1, off + copy_word_size);
478 build_copy_load_pref(&buf, off + 2 * copy_word_size);
479 build_copy_load(&buf, T2, off + 2 * copy_word_size);
480 build_copy_load_pref(&buf, off + 3 * copy_word_size);
481 build_copy_load(&buf, T3, off + 3 * copy_word_size);
482 build_copy_store_pref(&buf, off);
483 build_copy_store(&buf, T0, off);
484 build_copy_store_pref(&buf, off + copy_word_size);
485 build_copy_store(&buf, T1, off + copy_word_size);
486 build_copy_store_pref(&buf, off + 2 * copy_word_size);
487 build_copy_store(&buf, T2, off + 2 * copy_word_size);
488 build_copy_store_pref(&buf, off + 3 * copy_word_size);
489 build_copy_store(&buf, T3, off + 3 * copy_word_size);
490 off += 4 * copy_word_size;
491 } while (off < half_copy_loop_size);
492 pg_addiu(&buf, A1, A1, 2 * off);
493 pg_addiu(&buf, A0, A0, 2 * off);
494 off = -off;
495 do {
496 build_copy_load_pref(&buf, off);
497 build_copy_load(&buf, T0, off);
498 build_copy_load_pref(&buf, off + copy_word_size);
499 build_copy_load(&buf, T1, off + copy_word_size);
500 build_copy_load_pref(&buf, off + 2 * copy_word_size);
501 build_copy_load(&buf, T2, off + 2 * copy_word_size);
502 build_copy_load_pref(&buf, off + 3 * copy_word_size);
503 build_copy_load(&buf, T3, off + 3 * copy_word_size);
504 build_copy_store_pref(&buf, off);
505 build_copy_store(&buf, T0, off);
506 build_copy_store_pref(&buf, off + copy_word_size);
507 build_copy_store(&buf, T1, off + copy_word_size);
508 build_copy_store_pref(&buf, off + 2 * copy_word_size);
509 build_copy_store(&buf, T2, off + 2 * copy_word_size);
510 build_copy_store_pref(&buf, off + 3 * copy_word_size);
511 if (off == -(4 * copy_word_size))
512 uasm_il_bne(&buf, &r, A2, A0, label_copy_pref_both);
513 build_copy_store(&buf, T3, off + 3 * copy_word_size);
514 off += 4 * copy_word_size;
515 } while (off < 0);
516
517 if (pref_bias_copy_load - pref_bias_copy_store) {
518 pg_addiu(&buf, A2, A0,
519 pref_bias_copy_load - pref_bias_copy_store);
520 uasm_l_copy_pref_store(&l, buf);
521 off = 0;
522 do {
523 build_copy_load(&buf, T0, off);
524 build_copy_load(&buf, T1, off + copy_word_size);
525 build_copy_load(&buf, T2, off + 2 * copy_word_size);
526 build_copy_load(&buf, T3, off + 3 * copy_word_size);
527 build_copy_store_pref(&buf, off);
528 build_copy_store(&buf, T0, off);
529 build_copy_store_pref(&buf, off + copy_word_size);
530 build_copy_store(&buf, T1, off + copy_word_size);
531 build_copy_store_pref(&buf, off + 2 * copy_word_size);
532 build_copy_store(&buf, T2, off + 2 * copy_word_size);
533 build_copy_store_pref(&buf, off + 3 * copy_word_size);
534 build_copy_store(&buf, T3, off + 3 * copy_word_size);
535 off += 4 * copy_word_size;
536 } while (off < half_copy_loop_size);
537 pg_addiu(&buf, A1, A1, 2 * off);
538 pg_addiu(&buf, A0, A0, 2 * off);
539 off = -off;
540 do {
541 build_copy_load(&buf, T0, off);
542 build_copy_load(&buf, T1, off + copy_word_size);
543 build_copy_load(&buf, T2, off + 2 * copy_word_size);
544 build_copy_load(&buf, T3, off + 3 * copy_word_size);
545 build_copy_store_pref(&buf, off);
546 build_copy_store(&buf, T0, off);
547 build_copy_store_pref(&buf, off + copy_word_size);
548 build_copy_store(&buf, T1, off + copy_word_size);
549 build_copy_store_pref(&buf, off + 2 * copy_word_size);
550 build_copy_store(&buf, T2, off + 2 * copy_word_size);
551 build_copy_store_pref(&buf, off + 3 * copy_word_size);
552 if (off == -(4 * copy_word_size))
553 uasm_il_bne(&buf, &r, A2, A0,
554 label_copy_pref_store);
555 build_copy_store(&buf, T3, off + 3 * copy_word_size);
556 off += 4 * copy_word_size;
557 } while (off < 0);
558 }
559
560 if (pref_bias_copy_store) {
561 pg_addiu(&buf, A2, A0, pref_bias_copy_store);
562 uasm_l_copy_nopref(&l, buf);
563 off = 0;
564 do {
565 build_copy_load(&buf, T0, off);
566 build_copy_load(&buf, T1, off + copy_word_size);
567 build_copy_load(&buf, T2, off + 2 * copy_word_size);
568 build_copy_load(&buf, T3, off + 3 * copy_word_size);
569 build_copy_store(&buf, T0, off);
570 build_copy_store(&buf, T1, off + copy_word_size);
571 build_copy_store(&buf, T2, off + 2 * copy_word_size);
572 build_copy_store(&buf, T3, off + 3 * copy_word_size);
573 off += 4 * copy_word_size;
574 } while (off < half_copy_loop_size);
575 pg_addiu(&buf, A1, A1, 2 * off);
576 pg_addiu(&buf, A0, A0, 2 * off);
577 off = -off;
578 do {
579 build_copy_load(&buf, T0, off);
580 build_copy_load(&buf, T1, off + copy_word_size);
581 build_copy_load(&buf, T2, off + 2 * copy_word_size);
582 build_copy_load(&buf, T3, off + 3 * copy_word_size);
583 build_copy_store(&buf, T0, off);
584 build_copy_store(&buf, T1, off + copy_word_size);
585 build_copy_store(&buf, T2, off + 2 * copy_word_size);
586 if (off == -(4 * copy_word_size))
587 uasm_il_bne(&buf, &r, A2, A0,
588 label_copy_nopref);
589 build_copy_store(&buf, T3, off + 3 * copy_word_size);
590 off += 4 * copy_word_size;
591 } while (off < 0);
592 }
593
594 uasm_i_jr(&buf, RA);
595 uasm_i_nop(&buf);
596
597 BUG_ON(buf > &__copy_page_end);
598
599 uasm_resolve_relocs(relocs, labels);
600
601 pr_debug("Synthesized copy page handler (%u instructions).\n",
602 (u32)(buf - &__copy_page_start));
603
604 pr_debug("\t.set push\n");
605 pr_debug("\t.set noreorder\n");
606 for (i = 0; i < (buf - &__copy_page_start); i++)
607 pr_debug("\t.word 0x%08x\n", (&__copy_page_start)[i]);
608 pr_debug("\t.set pop\n");
609}
610
611#ifdef CONFIG_SIBYTE_DMA_PAGEOPS
612extern void clear_page_cpu(void *page);
613extern void copy_page_cpu(void *to, void *from);
614
615/*
616 * Pad descriptors to cacheline, since each is exclusively owned by a
617 * particular CPU.
618 */
619struct dmadscr {
620 u64 dscr_a;
621 u64 dscr_b;
622 u64 pad_a;
623 u64 pad_b;
624} ____cacheline_aligned_in_smp page_descr[DM_NUM_CHANNELS];
625
626void sb1_dma_init(void)
627{
628 int i;
629
630 for (i = 0; i < DM_NUM_CHANNELS; i++) {
631 const u64 base_val = CPHYSADDR((unsigned long)&page_descr[i]) |
632 V_DM_DSCR_BASE_RINGSZ(1);
633 void *base_reg = IOADDR(A_DM_REGISTER(i, R_DM_DSCR_BASE));
634
635 __raw_writeq(base_val, base_reg);
636 __raw_writeq(base_val | M_DM_DSCR_BASE_RESET, base_reg);
637 __raw_writeq(base_val | M_DM_DSCR_BASE_ENABL, base_reg);
638 }
639}
640
641void clear_page(void *page)
642{
643 u64 to_phys = CPHYSADDR((unsigned long)page);
644 unsigned int cpu = smp_processor_id();
645
646 /* if the page is not in KSEG0, use old way */
647 if ((long)KSEGX((unsigned long)page) != (long)CKSEG0)
648 return clear_page_cpu(page);
649
650 page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_ZERO_MEM |
651 M_DM_DSCRA_L2C_DEST | M_DM_DSCRA_INTERRUPT;
652 page_descr[cpu].dscr_b = V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE);
653 __raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT)));
654
655 /*
656 * Don't really want to do it this way, but there's no
657 * reliable way to delay completion detection.
658 */
659 while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG)))
660 & M_DM_DSCR_BASE_INTERRUPT))
661 ;
662 __raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE)));
663}
664
665void copy_page(void *to, void *from)
666{
667 u64 from_phys = CPHYSADDR((unsigned long)from);
668 u64 to_phys = CPHYSADDR((unsigned long)to);
669 unsigned int cpu = smp_processor_id();
670
671 /* if any page is not in KSEG0, use old way */
672 if ((long)KSEGX((unsigned long)to) != (long)CKSEG0
673 || (long)KSEGX((unsigned long)from) != (long)CKSEG0)
674 return copy_page_cpu(to, from);
675
676 page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_L2C_DEST |
677 M_DM_DSCRA_INTERRUPT;
678 page_descr[cpu].dscr_b = from_phys | V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE);
679 __raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT)));
680
681 /*
682 * Don't really want to do it this way, but there's no
683 * reliable way to delay completion detection.
684 */
685 while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG)))
686 & M_DM_DSCR_BASE_INTERRUPT))
687 ;
688 __raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE)));
689}
690
691#endif /* CONFIG_SIBYTE_DMA_PAGEOPS */
1/*
2 * This file is subject to the terms and conditions of the GNU General Public
3 * License. See the file "COPYING" in the main directory of this archive
4 * for more details.
5 *
6 * Copyright (C) 2003, 04, 05 Ralf Baechle (ralf@linux-mips.org)
7 * Copyright (C) 2007 Maciej W. Rozycki
8 * Copyright (C) 2008 Thiemo Seufer
9 * Copyright (C) 2012 MIPS Technologies, Inc.
10 */
11#include <linux/kernel.h>
12#include <linux/sched.h>
13#include <linux/smp.h>
14#include <linux/mm.h>
15#include <linux/proc_fs.h>
16
17#include <asm/bugs.h>
18#include <asm/cacheops.h>
19#include <asm/cpu-type.h>
20#include <asm/inst.h>
21#include <asm/io.h>
22#include <asm/page.h>
23#include <asm/prefetch.h>
24#include <asm/bootinfo.h>
25#include <asm/mipsregs.h>
26#include <asm/mmu_context.h>
27#include <asm/regdef.h>
28#include <asm/cpu.h>
29
30#ifdef CONFIG_SIBYTE_DMA_PAGEOPS
31#include <asm/sibyte/sb1250.h>
32#include <asm/sibyte/sb1250_regs.h>
33#include <asm/sibyte/sb1250_dma.h>
34#endif
35
36#include <asm/uasm.h>
37
38/* Handle labels (which must be positive integers). */
39enum label_id {
40 label_clear_nopref = 1,
41 label_clear_pref,
42 label_copy_nopref,
43 label_copy_pref_both,
44 label_copy_pref_store,
45};
46
47UASM_L_LA(_clear_nopref)
48UASM_L_LA(_clear_pref)
49UASM_L_LA(_copy_nopref)
50UASM_L_LA(_copy_pref_both)
51UASM_L_LA(_copy_pref_store)
52
53/* We need one branch and therefore one relocation per target label. */
54static struct uasm_label labels[5];
55static struct uasm_reloc relocs[5];
56
57#define cpu_is_r4600_v1_x() ((read_c0_prid() & 0xfffffff0) == 0x00002010)
58#define cpu_is_r4600_v2_x() ((read_c0_prid() & 0xfffffff0) == 0x00002020)
59
60/*
61 * R6 has a limited offset of the pref instruction.
62 * Skip it if the offset is more than 9 bits.
63 */
64#define _uasm_i_pref(a, b, c, d) \
65do { \
66 if (cpu_has_mips_r6) { \
67 if (c <= 0xff && c >= -0x100) \
68 uasm_i_pref(a, b, c, d);\
69 } else { \
70 uasm_i_pref(a, b, c, d); \
71 } \
72} while(0)
73
74static int pref_bias_clear_store;
75static int pref_bias_copy_load;
76static int pref_bias_copy_store;
77
78static u32 pref_src_mode;
79static u32 pref_dst_mode;
80
81static int clear_word_size;
82static int copy_word_size;
83
84static int half_clear_loop_size;
85static int half_copy_loop_size;
86
87static int cache_line_size;
88#define cache_line_mask() (cache_line_size - 1)
89
90static inline void
91pg_addiu(u32 **buf, unsigned int reg1, unsigned int reg2, unsigned int off)
92{
93 if (cpu_has_64bit_gp_regs &&
94 IS_ENABLED(CONFIG_CPU_DADDI_WORKAROUNDS) &&
95 r4k_daddiu_bug()) {
96 if (off > 0x7fff) {
97 uasm_i_lui(buf, GPR_T9, uasm_rel_hi(off));
98 uasm_i_addiu(buf, GPR_T9, GPR_T9, uasm_rel_lo(off));
99 } else
100 uasm_i_addiu(buf, GPR_T9, GPR_ZERO, off);
101 uasm_i_daddu(buf, reg1, reg2, GPR_T9);
102 } else {
103 if (off > 0x7fff) {
104 uasm_i_lui(buf, GPR_T9, uasm_rel_hi(off));
105 uasm_i_addiu(buf, GPR_T9, GPR_T9, uasm_rel_lo(off));
106 UASM_i_ADDU(buf, reg1, reg2, GPR_T9);
107 } else
108 UASM_i_ADDIU(buf, reg1, reg2, off);
109 }
110}
111
112static void set_prefetch_parameters(void)
113{
114 if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg)
115 clear_word_size = 8;
116 else
117 clear_word_size = 4;
118
119 if (cpu_has_64bit_gp_regs)
120 copy_word_size = 8;
121 else
122 copy_word_size = 4;
123
124 /*
125 * The pref's used here are using "streaming" hints, which cause the
126 * copied data to be kicked out of the cache sooner. A page copy often
127 * ends up copying a lot more data than is commonly used, so this seems
128 * to make sense in terms of reducing cache pollution, but I've no real
129 * performance data to back this up.
130 */
131 if (cpu_has_prefetch) {
132 /*
133 * XXX: Most prefetch bias values in here are based on
134 * guesswork.
135 */
136 cache_line_size = cpu_dcache_line_size();
137 switch (current_cpu_type()) {
138 case CPU_R5500:
139 case CPU_TX49XX:
140 /* These processors only support the Pref_Load. */
141 pref_bias_copy_load = 256;
142 break;
143
144 case CPU_R10000:
145 case CPU_R12000:
146 case CPU_R14000:
147 case CPU_R16000:
148 /*
149 * Those values have been experimentally tuned for an
150 * Origin 200.
151 */
152 pref_bias_clear_store = 512;
153 pref_bias_copy_load = 256;
154 pref_bias_copy_store = 256;
155 pref_src_mode = Pref_LoadStreamed;
156 pref_dst_mode = Pref_StoreStreamed;
157 break;
158
159 case CPU_SB1:
160 case CPU_SB1A:
161 pref_bias_clear_store = 128;
162 pref_bias_copy_load = 128;
163 pref_bias_copy_store = 128;
164 /*
165 * SB1 pass1 Pref_LoadStreamed/Pref_StoreStreamed
166 * hints are broken.
167 */
168 if (current_cpu_type() == CPU_SB1 &&
169 (current_cpu_data.processor_id & 0xff) < 0x02) {
170 pref_src_mode = Pref_Load;
171 pref_dst_mode = Pref_Store;
172 } else {
173 pref_src_mode = Pref_LoadStreamed;
174 pref_dst_mode = Pref_StoreStreamed;
175 }
176 break;
177
178 case CPU_LOONGSON64:
179 /* Loongson-3 only support the Pref_Load/Pref_Store. */
180 pref_bias_clear_store = 128;
181 pref_bias_copy_load = 128;
182 pref_bias_copy_store = 128;
183 pref_src_mode = Pref_Load;
184 pref_dst_mode = Pref_Store;
185 break;
186
187 default:
188 pref_bias_clear_store = 128;
189 pref_bias_copy_load = 256;
190 pref_bias_copy_store = 128;
191 pref_src_mode = Pref_LoadStreamed;
192 if (cpu_has_mips_r6)
193 /*
194 * Bit 30 (Pref_PrepareForStore) has been
195 * removed from MIPS R6. Use bit 5
196 * (Pref_StoreStreamed).
197 */
198 pref_dst_mode = Pref_StoreStreamed;
199 else
200 pref_dst_mode = Pref_PrepareForStore;
201 break;
202 }
203 } else {
204 if (cpu_has_cache_cdex_s)
205 cache_line_size = cpu_scache_line_size();
206 else if (cpu_has_cache_cdex_p)
207 cache_line_size = cpu_dcache_line_size();
208 }
209 /*
210 * Too much unrolling will overflow the available space in
211 * clear_space_array / copy_page_array.
212 */
213 half_clear_loop_size = min(16 * clear_word_size,
214 max(cache_line_size >> 1,
215 4 * clear_word_size));
216 half_copy_loop_size = min(16 * copy_word_size,
217 max(cache_line_size >> 1,
218 4 * copy_word_size));
219}
220
221static void build_clear_store(u32 **buf, int off)
222{
223 if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg) {
224 uasm_i_sd(buf, GPR_ZERO, off, GPR_A0);
225 } else {
226 uasm_i_sw(buf, GPR_ZERO, off, GPR_A0);
227 }
228}
229
230static inline void build_clear_pref(u32 **buf, int off)
231{
232 if (off & cache_line_mask())
233 return;
234
235 if (pref_bias_clear_store) {
236 _uasm_i_pref(buf, pref_dst_mode, pref_bias_clear_store + off,
237 GPR_A0);
238 } else if (cache_line_size == (half_clear_loop_size << 1)) {
239 if (cpu_has_cache_cdex_s) {
240 uasm_i_cache(buf, Create_Dirty_Excl_SD, off, GPR_A0);
241 } else if (cpu_has_cache_cdex_p) {
242 if (IS_ENABLED(CONFIG_WAR_R4600_V1_HIT_CACHEOP) &&
243 cpu_is_r4600_v1_x()) {
244 uasm_i_nop(buf);
245 uasm_i_nop(buf);
246 uasm_i_nop(buf);
247 uasm_i_nop(buf);
248 }
249
250 if (IS_ENABLED(CONFIG_WAR_R4600_V2_HIT_CACHEOP) &&
251 cpu_is_r4600_v2_x())
252 uasm_i_lw(buf, GPR_ZERO, GPR_ZERO, GPR_AT);
253
254 uasm_i_cache(buf, Create_Dirty_Excl_D, off, GPR_A0);
255 }
256 }
257}
258
259extern u32 __clear_page_start;
260extern u32 __clear_page_end;
261extern u32 __copy_page_start;
262extern u32 __copy_page_end;
263
264void build_clear_page(void)
265{
266 int off;
267 u32 *buf = &__clear_page_start;
268 struct uasm_label *l = labels;
269 struct uasm_reloc *r = relocs;
270 int i;
271 static atomic_t run_once = ATOMIC_INIT(0);
272
273 if (atomic_xchg(&run_once, 1)) {
274 return;
275 }
276
277 memset(labels, 0, sizeof(labels));
278 memset(relocs, 0, sizeof(relocs));
279
280 set_prefetch_parameters();
281
282 /*
283 * This algorithm makes the following assumptions:
284 * - The prefetch bias is a multiple of 2 words.
285 * - The prefetch bias is less than one page.
286 */
287 BUG_ON(pref_bias_clear_store % (2 * clear_word_size));
288 BUG_ON(PAGE_SIZE < pref_bias_clear_store);
289
290 off = PAGE_SIZE - pref_bias_clear_store;
291 if (off > 0xffff || !pref_bias_clear_store)
292 pg_addiu(&buf, GPR_A2, GPR_A0, off);
293 else
294 uasm_i_ori(&buf, GPR_A2, GPR_A0, off);
295
296 if (IS_ENABLED(CONFIG_WAR_R4600_V2_HIT_CACHEOP) && cpu_is_r4600_v2_x())
297 uasm_i_lui(&buf, GPR_AT, uasm_rel_hi(0xa0000000));
298
299 off = cache_line_size ? min(8, pref_bias_clear_store / cache_line_size)
300 * cache_line_size : 0;
301 while (off) {
302 build_clear_pref(&buf, -off);
303 off -= cache_line_size;
304 }
305 uasm_l_clear_pref(&l, buf);
306 do {
307 build_clear_pref(&buf, off);
308 build_clear_store(&buf, off);
309 off += clear_word_size;
310 } while (off < half_clear_loop_size);
311 pg_addiu(&buf, GPR_A0, GPR_A0, 2 * off);
312 off = -off;
313 do {
314 build_clear_pref(&buf, off);
315 if (off == -clear_word_size)
316 uasm_il_bne(&buf, &r, GPR_A0, GPR_A2, label_clear_pref);
317 build_clear_store(&buf, off);
318 off += clear_word_size;
319 } while (off < 0);
320
321 if (pref_bias_clear_store) {
322 pg_addiu(&buf, GPR_A2, GPR_A0, pref_bias_clear_store);
323 uasm_l_clear_nopref(&l, buf);
324 off = 0;
325 do {
326 build_clear_store(&buf, off);
327 off += clear_word_size;
328 } while (off < half_clear_loop_size);
329 pg_addiu(&buf, GPR_A0, GPR_A0, 2 * off);
330 off = -off;
331 do {
332 if (off == -clear_word_size)
333 uasm_il_bne(&buf, &r, GPR_A0, GPR_A2,
334 label_clear_nopref);
335 build_clear_store(&buf, off);
336 off += clear_word_size;
337 } while (off < 0);
338 }
339
340 uasm_i_jr(&buf, GPR_RA);
341 uasm_i_nop(&buf);
342
343 BUG_ON(buf > &__clear_page_end);
344
345 uasm_resolve_relocs(relocs, labels);
346
347 pr_debug("Synthesized clear page handler (%u instructions).\n",
348 (u32)(buf - &__clear_page_start));
349
350 pr_debug("\t.set push\n");
351 pr_debug("\t.set noreorder\n");
352 for (i = 0; i < (buf - &__clear_page_start); i++)
353 pr_debug("\t.word 0x%08x\n", (&__clear_page_start)[i]);
354 pr_debug("\t.set pop\n");
355}
356
357static void build_copy_load(u32 **buf, int reg, int off)
358{
359 if (cpu_has_64bit_gp_regs) {
360 uasm_i_ld(buf, reg, off, GPR_A1);
361 } else {
362 uasm_i_lw(buf, reg, off, GPR_A1);
363 }
364}
365
366static void build_copy_store(u32 **buf, int reg, int off)
367{
368 if (cpu_has_64bit_gp_regs) {
369 uasm_i_sd(buf, reg, off, GPR_A0);
370 } else {
371 uasm_i_sw(buf, reg, off, GPR_A0);
372 }
373}
374
375static inline void build_copy_load_pref(u32 **buf, int off)
376{
377 if (off & cache_line_mask())
378 return;
379
380 if (pref_bias_copy_load)
381 _uasm_i_pref(buf, pref_src_mode, pref_bias_copy_load + off, GPR_A1);
382}
383
384static inline void build_copy_store_pref(u32 **buf, int off)
385{
386 if (off & cache_line_mask())
387 return;
388
389 if (pref_bias_copy_store) {
390 _uasm_i_pref(buf, pref_dst_mode, pref_bias_copy_store + off,
391 GPR_A0);
392 } else if (cache_line_size == (half_copy_loop_size << 1)) {
393 if (cpu_has_cache_cdex_s) {
394 uasm_i_cache(buf, Create_Dirty_Excl_SD, off, GPR_A0);
395 } else if (cpu_has_cache_cdex_p) {
396 if (IS_ENABLED(CONFIG_WAR_R4600_V1_HIT_CACHEOP) &&
397 cpu_is_r4600_v1_x()) {
398 uasm_i_nop(buf);
399 uasm_i_nop(buf);
400 uasm_i_nop(buf);
401 uasm_i_nop(buf);
402 }
403
404 if (IS_ENABLED(CONFIG_WAR_R4600_V2_HIT_CACHEOP) &&
405 cpu_is_r4600_v2_x())
406 uasm_i_lw(buf, GPR_ZERO, GPR_ZERO, GPR_AT);
407
408 uasm_i_cache(buf, Create_Dirty_Excl_D, off, GPR_A0);
409 }
410 }
411}
412
413void build_copy_page(void)
414{
415 int off;
416 u32 *buf = &__copy_page_start;
417 struct uasm_label *l = labels;
418 struct uasm_reloc *r = relocs;
419 int i;
420 static atomic_t run_once = ATOMIC_INIT(0);
421
422 if (atomic_xchg(&run_once, 1)) {
423 return;
424 }
425
426 memset(labels, 0, sizeof(labels));
427 memset(relocs, 0, sizeof(relocs));
428
429 set_prefetch_parameters();
430
431 /*
432 * This algorithm makes the following assumptions:
433 * - All prefetch biases are multiples of 8 words.
434 * - The prefetch biases are less than one page.
435 * - The store prefetch bias isn't greater than the load
436 * prefetch bias.
437 */
438 BUG_ON(pref_bias_copy_load % (8 * copy_word_size));
439 BUG_ON(pref_bias_copy_store % (8 * copy_word_size));
440 BUG_ON(PAGE_SIZE < pref_bias_copy_load);
441 BUG_ON(pref_bias_copy_store > pref_bias_copy_load);
442
443 off = PAGE_SIZE - pref_bias_copy_load;
444 if (off > 0xffff || !pref_bias_copy_load)
445 pg_addiu(&buf, GPR_A2, GPR_A0, off);
446 else
447 uasm_i_ori(&buf, GPR_A2, GPR_A0, off);
448
449 if (IS_ENABLED(CONFIG_WAR_R4600_V2_HIT_CACHEOP) && cpu_is_r4600_v2_x())
450 uasm_i_lui(&buf, GPR_AT, uasm_rel_hi(0xa0000000));
451
452 off = cache_line_size ? min(8, pref_bias_copy_load / cache_line_size) *
453 cache_line_size : 0;
454 while (off) {
455 build_copy_load_pref(&buf, -off);
456 off -= cache_line_size;
457 }
458 off = cache_line_size ? min(8, pref_bias_copy_store / cache_line_size) *
459 cache_line_size : 0;
460 while (off) {
461 build_copy_store_pref(&buf, -off);
462 off -= cache_line_size;
463 }
464 uasm_l_copy_pref_both(&l, buf);
465 do {
466 build_copy_load_pref(&buf, off);
467 build_copy_load(&buf, GPR_T0, off);
468 build_copy_load_pref(&buf, off + copy_word_size);
469 build_copy_load(&buf, GPR_T1, off + copy_word_size);
470 build_copy_load_pref(&buf, off + 2 * copy_word_size);
471 build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size);
472 build_copy_load_pref(&buf, off + 3 * copy_word_size);
473 build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size);
474 build_copy_store_pref(&buf, off);
475 build_copy_store(&buf, GPR_T0, off);
476 build_copy_store_pref(&buf, off + copy_word_size);
477 build_copy_store(&buf, GPR_T1, off + copy_word_size);
478 build_copy_store_pref(&buf, off + 2 * copy_word_size);
479 build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size);
480 build_copy_store_pref(&buf, off + 3 * copy_word_size);
481 build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size);
482 off += 4 * copy_word_size;
483 } while (off < half_copy_loop_size);
484 pg_addiu(&buf, GPR_A1, GPR_A1, 2 * off);
485 pg_addiu(&buf, GPR_A0, GPR_A0, 2 * off);
486 off = -off;
487 do {
488 build_copy_load_pref(&buf, off);
489 build_copy_load(&buf, GPR_T0, off);
490 build_copy_load_pref(&buf, off + copy_word_size);
491 build_copy_load(&buf, GPR_T1, off + copy_word_size);
492 build_copy_load_pref(&buf, off + 2 * copy_word_size);
493 build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size);
494 build_copy_load_pref(&buf, off + 3 * copy_word_size);
495 build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size);
496 build_copy_store_pref(&buf, off);
497 build_copy_store(&buf, GPR_T0, off);
498 build_copy_store_pref(&buf, off + copy_word_size);
499 build_copy_store(&buf, GPR_T1, off + copy_word_size);
500 build_copy_store_pref(&buf, off + 2 * copy_word_size);
501 build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size);
502 build_copy_store_pref(&buf, off + 3 * copy_word_size);
503 if (off == -(4 * copy_word_size))
504 uasm_il_bne(&buf, &r, GPR_A2, GPR_A0, label_copy_pref_both);
505 build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size);
506 off += 4 * copy_word_size;
507 } while (off < 0);
508
509 if (pref_bias_copy_load - pref_bias_copy_store) {
510 pg_addiu(&buf, GPR_A2, GPR_A0,
511 pref_bias_copy_load - pref_bias_copy_store);
512 uasm_l_copy_pref_store(&l, buf);
513 off = 0;
514 do {
515 build_copy_load(&buf, GPR_T0, off);
516 build_copy_load(&buf, GPR_T1, off + copy_word_size);
517 build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size);
518 build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size);
519 build_copy_store_pref(&buf, off);
520 build_copy_store(&buf, GPR_T0, off);
521 build_copy_store_pref(&buf, off + copy_word_size);
522 build_copy_store(&buf, GPR_T1, off + copy_word_size);
523 build_copy_store_pref(&buf, off + 2 * copy_word_size);
524 build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size);
525 build_copy_store_pref(&buf, off + 3 * copy_word_size);
526 build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size);
527 off += 4 * copy_word_size;
528 } while (off < half_copy_loop_size);
529 pg_addiu(&buf, GPR_A1, GPR_A1, 2 * off);
530 pg_addiu(&buf, GPR_A0, GPR_A0, 2 * off);
531 off = -off;
532 do {
533 build_copy_load(&buf, GPR_T0, off);
534 build_copy_load(&buf, GPR_T1, off + copy_word_size);
535 build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size);
536 build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size);
537 build_copy_store_pref(&buf, off);
538 build_copy_store(&buf, GPR_T0, off);
539 build_copy_store_pref(&buf, off + copy_word_size);
540 build_copy_store(&buf, GPR_T1, off + copy_word_size);
541 build_copy_store_pref(&buf, off + 2 * copy_word_size);
542 build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size);
543 build_copy_store_pref(&buf, off + 3 * copy_word_size);
544 if (off == -(4 * copy_word_size))
545 uasm_il_bne(&buf, &r, GPR_A2, GPR_A0,
546 label_copy_pref_store);
547 build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size);
548 off += 4 * copy_word_size;
549 } while (off < 0);
550 }
551
552 if (pref_bias_copy_store) {
553 pg_addiu(&buf, GPR_A2, GPR_A0, pref_bias_copy_store);
554 uasm_l_copy_nopref(&l, buf);
555 off = 0;
556 do {
557 build_copy_load(&buf, GPR_T0, off);
558 build_copy_load(&buf, GPR_T1, off + copy_word_size);
559 build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size);
560 build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size);
561 build_copy_store(&buf, GPR_T0, off);
562 build_copy_store(&buf, GPR_T1, off + copy_word_size);
563 build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size);
564 build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size);
565 off += 4 * copy_word_size;
566 } while (off < half_copy_loop_size);
567 pg_addiu(&buf, GPR_A1, GPR_A1, 2 * off);
568 pg_addiu(&buf, GPR_A0, GPR_A0, 2 * off);
569 off = -off;
570 do {
571 build_copy_load(&buf, GPR_T0, off);
572 build_copy_load(&buf, GPR_T1, off + copy_word_size);
573 build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size);
574 build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size);
575 build_copy_store(&buf, GPR_T0, off);
576 build_copy_store(&buf, GPR_T1, off + copy_word_size);
577 build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size);
578 if (off == -(4 * copy_word_size))
579 uasm_il_bne(&buf, &r, GPR_A2, GPR_A0,
580 label_copy_nopref);
581 build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size);
582 off += 4 * copy_word_size;
583 } while (off < 0);
584 }
585
586 uasm_i_jr(&buf, GPR_RA);
587 uasm_i_nop(&buf);
588
589 BUG_ON(buf > &__copy_page_end);
590
591 uasm_resolve_relocs(relocs, labels);
592
593 pr_debug("Synthesized copy page handler (%u instructions).\n",
594 (u32)(buf - &__copy_page_start));
595
596 pr_debug("\t.set push\n");
597 pr_debug("\t.set noreorder\n");
598 for (i = 0; i < (buf - &__copy_page_start); i++)
599 pr_debug("\t.word 0x%08x\n", (&__copy_page_start)[i]);
600 pr_debug("\t.set pop\n");
601}
602
603#ifdef CONFIG_SIBYTE_DMA_PAGEOPS
604extern void clear_page_cpu(void *page);
605extern void copy_page_cpu(void *to, void *from);
606
607/*
608 * Pad descriptors to cacheline, since each is exclusively owned by a
609 * particular CPU.
610 */
611struct dmadscr {
612 u64 dscr_a;
613 u64 dscr_b;
614 u64 pad_a;
615 u64 pad_b;
616} ____cacheline_aligned_in_smp page_descr[DM_NUM_CHANNELS];
617
618void clear_page(void *page)
619{
620 u64 to_phys = CPHYSADDR((unsigned long)page);
621 unsigned int cpu = smp_processor_id();
622
623 /* if the page is not in KSEG0, use old way */
624 if ((long)KSEGX((unsigned long)page) != (long)CKSEG0)
625 return clear_page_cpu(page);
626
627 page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_ZERO_MEM |
628 M_DM_DSCRA_L2C_DEST | M_DM_DSCRA_INTERRUPT;
629 page_descr[cpu].dscr_b = V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE);
630 __raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT)));
631
632 /*
633 * Don't really want to do it this way, but there's no
634 * reliable way to delay completion detection.
635 */
636 while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG)))
637 & M_DM_DSCR_BASE_INTERRUPT))
638 ;
639 __raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE)));
640}
641EXPORT_SYMBOL(clear_page);
642
643void copy_page(void *to, void *from)
644{
645 u64 from_phys = CPHYSADDR((unsigned long)from);
646 u64 to_phys = CPHYSADDR((unsigned long)to);
647 unsigned int cpu = smp_processor_id();
648
649 /* if any page is not in KSEG0, use old way */
650 if ((long)KSEGX((unsigned long)to) != (long)CKSEG0
651 || (long)KSEGX((unsigned long)from) != (long)CKSEG0)
652 return copy_page_cpu(to, from);
653
654 page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_L2C_DEST |
655 M_DM_DSCRA_INTERRUPT;
656 page_descr[cpu].dscr_b = from_phys | V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE);
657 __raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT)));
658
659 /*
660 * Don't really want to do it this way, but there's no
661 * reliable way to delay completion detection.
662 */
663 while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG)))
664 & M_DM_DSCR_BASE_INTERRUPT))
665 ;
666 __raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE)));
667}
668EXPORT_SYMBOL(copy_page);
669
670#endif /* CONFIG_SIBYTE_DMA_PAGEOPS */