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
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.
11 #include <linux/kernel.h>
12 #include <linux/sched.h>
13 #include <linux/smp.h>
15 #include <linux/proc_fs.h>
18 #include <asm/cacheops.h>
19 #include <asm/cpu-type.h>
23 #include <asm/prefetch.h>
24 #include <asm/bootinfo.h>
25 #include <asm/mipsregs.h>
26 #include <asm/mmu_context.h>
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>
38 /* Registers used in the assembled routines. */
51 /* Handle labels (which must be positive integers). */
53 label_clear_nopref = 1,
57 label_copy_pref_store,
60 UASM_L_LA(_clear_nopref)
61 UASM_L_LA(_clear_pref)
62 UASM_L_LA(_copy_nopref)
63 UASM_L_LA(_copy_pref_both)
64 UASM_L_LA(_copy_pref_store)
66 /* We need one branch and therefore one relocation per target label. */
67 static struct uasm_label labels[5];
68 static struct uasm_reloc relocs[5];
70 #define cpu_is_r4600_v1_x() ((read_c0_prid() & 0xfffffff0) == 0x00002010)
71 #define cpu_is_r4600_v2_x() ((read_c0_prid() & 0xfffffff0) == 0x00002020)
74 * R6 has a limited offset of the pref instruction.
75 * Skip it if the offset is more than 9 bits.
77 #define _uasm_i_pref(a, b, c, d) \
79 if (cpu_has_mips_r6) { \
80 if (c <= 0xff && c >= -0x100) \
81 uasm_i_pref(a, b, c, d);\
83 uasm_i_pref(a, b, c, d); \
87 static int pref_bias_clear_store;
88 static int pref_bias_copy_load;
89 static int pref_bias_copy_store;
91 static u32 pref_src_mode;
92 static u32 pref_dst_mode;
94 static int clear_word_size;
95 static int copy_word_size;
97 static int half_clear_loop_size;
98 static int half_copy_loop_size;
100 static int cache_line_size;
101 #define cache_line_mask() (cache_line_size - 1)
104 pg_addiu(u32 **buf, unsigned int reg1, unsigned int reg2, unsigned int off)
106 if (cpu_has_64bit_gp_regs && DADDI_WAR && r4k_daddiu_bug()) {
108 uasm_i_lui(buf, T9, uasm_rel_hi(off));
109 uasm_i_addiu(buf, T9, T9, uasm_rel_lo(off));
111 uasm_i_addiu(buf, T9, ZERO, off);
112 uasm_i_daddu(buf, reg1, reg2, T9);
115 uasm_i_lui(buf, T9, uasm_rel_hi(off));
116 uasm_i_addiu(buf, T9, T9, uasm_rel_lo(off));
117 UASM_i_ADDU(buf, reg1, reg2, T9);
119 UASM_i_ADDIU(buf, reg1, reg2, off);
123 static void set_prefetch_parameters(void)
125 if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg)
130 if (cpu_has_64bit_gp_regs)
136 * The pref's used here are using "streaming" hints, which cause the
137 * copied data to be kicked out of the cache sooner. A page copy often
138 * ends up copying a lot more data than is commonly used, so this seems
139 * to make sense in terms of reducing cache pollution, but I've no real
140 * performance data to back this up.
142 if (cpu_has_prefetch) {
144 * XXX: Most prefetch bias values in here are based on
147 cache_line_size = cpu_dcache_line_size();
148 switch (current_cpu_type()) {
151 /* These processors only support the Pref_Load. */
152 pref_bias_copy_load = 256;
160 * Those values have been experimentally tuned for an
163 pref_bias_clear_store = 512;
164 pref_bias_copy_load = 256;
165 pref_bias_copy_store = 256;
166 pref_src_mode = Pref_LoadStreamed;
167 pref_dst_mode = Pref_StoreStreamed;
172 pref_bias_clear_store = 128;
173 pref_bias_copy_load = 128;
174 pref_bias_copy_store = 128;
176 * SB1 pass1 Pref_LoadStreamed/Pref_StoreStreamed
179 if (current_cpu_type() == CPU_SB1 &&
180 (current_cpu_data.processor_id & 0xff) < 0x02) {
181 pref_src_mode = Pref_Load;
182 pref_dst_mode = Pref_Store;
184 pref_src_mode = Pref_LoadStreamed;
185 pref_dst_mode = Pref_StoreStreamed;
190 /* Loongson-3 only support the Pref_Load/Pref_Store. */
191 pref_bias_clear_store = 128;
192 pref_bias_copy_load = 128;
193 pref_bias_copy_store = 128;
194 pref_src_mode = Pref_Load;
195 pref_dst_mode = Pref_Store;
199 pref_bias_clear_store = 128;
200 pref_bias_copy_load = 256;
201 pref_bias_copy_store = 128;
202 pref_src_mode = Pref_LoadStreamed;
205 * Bit 30 (Pref_PrepareForStore) has been
206 * removed from MIPS R6. Use bit 5
207 * (Pref_StoreStreamed).
209 pref_dst_mode = Pref_StoreStreamed;
211 pref_dst_mode = Pref_PrepareForStore;
215 if (cpu_has_cache_cdex_s)
216 cache_line_size = cpu_scache_line_size();
217 else if (cpu_has_cache_cdex_p)
218 cache_line_size = cpu_dcache_line_size();
221 * Too much unrolling will overflow the available space in
222 * clear_space_array / copy_page_array.
224 half_clear_loop_size = min(16 * clear_word_size,
225 max(cache_line_size >> 1,
226 4 * clear_word_size));
227 half_copy_loop_size = min(16 * copy_word_size,
228 max(cache_line_size >> 1,
229 4 * copy_word_size));
232 static void build_clear_store(u32 **buf, int off)
234 if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg) {
235 uasm_i_sd(buf, ZERO, off, A0);
237 uasm_i_sw(buf, ZERO, off, A0);
241 static inline void build_clear_pref(u32 **buf, int off)
243 if (off & cache_line_mask())
246 if (pref_bias_clear_store) {
247 _uasm_i_pref(buf, pref_dst_mode, pref_bias_clear_store + off,
249 } else if (cache_line_size == (half_clear_loop_size << 1)) {
250 if (cpu_has_cache_cdex_s) {
251 uasm_i_cache(buf, Create_Dirty_Excl_SD, off, A0);
252 } else if (cpu_has_cache_cdex_p) {
253 if (R4600_V1_HIT_CACHEOP_WAR && cpu_is_r4600_v1_x()) {
260 if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
261 uasm_i_lw(buf, ZERO, ZERO, AT);
263 uasm_i_cache(buf, Create_Dirty_Excl_D, off, A0);
268 extern u32 __clear_page_start;
269 extern u32 __clear_page_end;
270 extern u32 __copy_page_start;
271 extern u32 __copy_page_end;
273 void build_clear_page(void)
276 u32 *buf = &__clear_page_start;
277 struct uasm_label *l = labels;
278 struct uasm_reloc *r = relocs;
280 static atomic_t run_once = ATOMIC_INIT(0);
282 if (atomic_xchg(&run_once, 1)) {
286 memset(labels, 0, sizeof(labels));
287 memset(relocs, 0, sizeof(relocs));
289 set_prefetch_parameters();
292 * This algorithm makes the following assumptions:
293 * - The prefetch bias is a multiple of 2 words.
294 * - The prefetch bias is less than one page.
296 BUG_ON(pref_bias_clear_store % (2 * clear_word_size));
297 BUG_ON(PAGE_SIZE < pref_bias_clear_store);
299 off = PAGE_SIZE - pref_bias_clear_store;
300 if (off > 0xffff || !pref_bias_clear_store)
301 pg_addiu(&buf, A2, A0, off);
303 uasm_i_ori(&buf, A2, A0, off);
305 if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
306 uasm_i_lui(&buf, AT, uasm_rel_hi(0xa0000000));
308 off = cache_line_size ? min(8, pref_bias_clear_store / cache_line_size)
309 * cache_line_size : 0;
311 build_clear_pref(&buf, -off);
312 off -= cache_line_size;
314 uasm_l_clear_pref(&l, buf);
316 build_clear_pref(&buf, off);
317 build_clear_store(&buf, off);
318 off += clear_word_size;
319 } while (off < half_clear_loop_size);
320 pg_addiu(&buf, A0, A0, 2 * off);
323 build_clear_pref(&buf, off);
324 if (off == -clear_word_size)
325 uasm_il_bne(&buf, &r, A0, A2, label_clear_pref);
326 build_clear_store(&buf, off);
327 off += clear_word_size;
330 if (pref_bias_clear_store) {
331 pg_addiu(&buf, A2, A0, pref_bias_clear_store);
332 uasm_l_clear_nopref(&l, buf);
335 build_clear_store(&buf, off);
336 off += clear_word_size;
337 } while (off < half_clear_loop_size);
338 pg_addiu(&buf, A0, A0, 2 * off);
341 if (off == -clear_word_size)
342 uasm_il_bne(&buf, &r, A0, A2,
344 build_clear_store(&buf, off);
345 off += clear_word_size;
352 BUG_ON(buf > &__clear_page_end);
354 uasm_resolve_relocs(relocs, labels);
356 pr_debug("Synthesized clear page handler (%u instructions).\n",
357 (u32)(buf - &__clear_page_start));
359 pr_debug("\t.set push\n");
360 pr_debug("\t.set noreorder\n");
361 for (i = 0; i < (buf - &__clear_page_start); i++)
362 pr_debug("\t.word 0x%08x\n", (&__clear_page_start)[i]);
363 pr_debug("\t.set pop\n");
366 static void build_copy_load(u32 **buf, int reg, int off)
368 if (cpu_has_64bit_gp_regs) {
369 uasm_i_ld(buf, reg, off, A1);
371 uasm_i_lw(buf, reg, off, A1);
375 static void build_copy_store(u32 **buf, int reg, int off)
377 if (cpu_has_64bit_gp_regs) {
378 uasm_i_sd(buf, reg, off, A0);
380 uasm_i_sw(buf, reg, off, A0);
384 static inline void build_copy_load_pref(u32 **buf, int off)
386 if (off & cache_line_mask())
389 if (pref_bias_copy_load)
390 _uasm_i_pref(buf, pref_src_mode, pref_bias_copy_load + off, A1);
393 static inline void build_copy_store_pref(u32 **buf, int off)
395 if (off & cache_line_mask())
398 if (pref_bias_copy_store) {
399 _uasm_i_pref(buf, pref_dst_mode, pref_bias_copy_store + off,
401 } else if (cache_line_size == (half_copy_loop_size << 1)) {
402 if (cpu_has_cache_cdex_s) {
403 uasm_i_cache(buf, Create_Dirty_Excl_SD, off, A0);
404 } else if (cpu_has_cache_cdex_p) {
405 if (R4600_V1_HIT_CACHEOP_WAR && cpu_is_r4600_v1_x()) {
412 if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
413 uasm_i_lw(buf, ZERO, ZERO, AT);
415 uasm_i_cache(buf, Create_Dirty_Excl_D, off, A0);
420 void build_copy_page(void)
423 u32 *buf = &__copy_page_start;
424 struct uasm_label *l = labels;
425 struct uasm_reloc *r = relocs;
427 static atomic_t run_once = ATOMIC_INIT(0);
429 if (atomic_xchg(&run_once, 1)) {
433 memset(labels, 0, sizeof(labels));
434 memset(relocs, 0, sizeof(relocs));
436 set_prefetch_parameters();
439 * This algorithm makes the following assumptions:
440 * - All prefetch biases are multiples of 8 words.
441 * - The prefetch biases are less than one page.
442 * - The store prefetch bias isn't greater than the load
445 BUG_ON(pref_bias_copy_load % (8 * copy_word_size));
446 BUG_ON(pref_bias_copy_store % (8 * copy_word_size));
447 BUG_ON(PAGE_SIZE < pref_bias_copy_load);
448 BUG_ON(pref_bias_copy_store > pref_bias_copy_load);
450 off = PAGE_SIZE - pref_bias_copy_load;
451 if (off > 0xffff || !pref_bias_copy_load)
452 pg_addiu(&buf, A2, A0, off);
454 uasm_i_ori(&buf, A2, A0, off);
456 if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())
457 uasm_i_lui(&buf, AT, uasm_rel_hi(0xa0000000));
459 off = cache_line_size ? min(8, pref_bias_copy_load / cache_line_size) *
462 build_copy_load_pref(&buf, -off);
463 off -= cache_line_size;
465 off = cache_line_size ? min(8, pref_bias_copy_store / cache_line_size) *
468 build_copy_store_pref(&buf, -off);
469 off -= cache_line_size;
471 uasm_l_copy_pref_both(&l, buf);
473 build_copy_load_pref(&buf, off);
474 build_copy_load(&buf, T0, off);
475 build_copy_load_pref(&buf, off + copy_word_size);
476 build_copy_load(&buf, T1, off + copy_word_size);
477 build_copy_load_pref(&buf, off + 2 * copy_word_size);
478 build_copy_load(&buf, T2, off + 2 * copy_word_size);
479 build_copy_load_pref(&buf, off + 3 * copy_word_size);
480 build_copy_load(&buf, T3, off + 3 * copy_word_size);
481 build_copy_store_pref(&buf, off);
482 build_copy_store(&buf, T0, off);
483 build_copy_store_pref(&buf, off + copy_word_size);
484 build_copy_store(&buf, T1, off + copy_word_size);
485 build_copy_store_pref(&buf, off + 2 * copy_word_size);
486 build_copy_store(&buf, T2, off + 2 * copy_word_size);
487 build_copy_store_pref(&buf, off + 3 * copy_word_size);
488 build_copy_store(&buf, T3, off + 3 * copy_word_size);
489 off += 4 * copy_word_size;
490 } while (off < half_copy_loop_size);
491 pg_addiu(&buf, A1, A1, 2 * off);
492 pg_addiu(&buf, A0, A0, 2 * off);
495 build_copy_load_pref(&buf, off);
496 build_copy_load(&buf, T0, off);
497 build_copy_load_pref(&buf, off + copy_word_size);
498 build_copy_load(&buf, T1, off + copy_word_size);
499 build_copy_load_pref(&buf, off + 2 * copy_word_size);
500 build_copy_load(&buf, T2, off + 2 * copy_word_size);
501 build_copy_load_pref(&buf, off + 3 * copy_word_size);
502 build_copy_load(&buf, T3, off + 3 * copy_word_size);
503 build_copy_store_pref(&buf, off);
504 build_copy_store(&buf, T0, off);
505 build_copy_store_pref(&buf, off + copy_word_size);
506 build_copy_store(&buf, T1, off + copy_word_size);
507 build_copy_store_pref(&buf, off + 2 * copy_word_size);
508 build_copy_store(&buf, T2, off + 2 * copy_word_size);
509 build_copy_store_pref(&buf, off + 3 * copy_word_size);
510 if (off == -(4 * copy_word_size))
511 uasm_il_bne(&buf, &r, A2, A0, label_copy_pref_both);
512 build_copy_store(&buf, T3, off + 3 * copy_word_size);
513 off += 4 * copy_word_size;
516 if (pref_bias_copy_load - pref_bias_copy_store) {
517 pg_addiu(&buf, A2, A0,
518 pref_bias_copy_load - pref_bias_copy_store);
519 uasm_l_copy_pref_store(&l, buf);
522 build_copy_load(&buf, T0, off);
523 build_copy_load(&buf, T1, off + copy_word_size);
524 build_copy_load(&buf, T2, off + 2 * copy_word_size);
525 build_copy_load(&buf, T3, off + 3 * copy_word_size);
526 build_copy_store_pref(&buf, off);
527 build_copy_store(&buf, T0, off);
528 build_copy_store_pref(&buf, off + copy_word_size);
529 build_copy_store(&buf, T1, off + copy_word_size);
530 build_copy_store_pref(&buf, off + 2 * copy_word_size);
531 build_copy_store(&buf, T2, off + 2 * copy_word_size);
532 build_copy_store_pref(&buf, off + 3 * copy_word_size);
533 build_copy_store(&buf, T3, off + 3 * copy_word_size);
534 off += 4 * copy_word_size;
535 } while (off < half_copy_loop_size);
536 pg_addiu(&buf, A1, A1, 2 * off);
537 pg_addiu(&buf, A0, A0, 2 * off);
540 build_copy_load(&buf, T0, off);
541 build_copy_load(&buf, T1, off + copy_word_size);
542 build_copy_load(&buf, T2, off + 2 * copy_word_size);
543 build_copy_load(&buf, T3, off + 3 * copy_word_size);
544 build_copy_store_pref(&buf, off);
545 build_copy_store(&buf, T0, off);
546 build_copy_store_pref(&buf, off + copy_word_size);
547 build_copy_store(&buf, T1, off + copy_word_size);
548 build_copy_store_pref(&buf, off + 2 * copy_word_size);
549 build_copy_store(&buf, T2, off + 2 * copy_word_size);
550 build_copy_store_pref(&buf, off + 3 * copy_word_size);
551 if (off == -(4 * copy_word_size))
552 uasm_il_bne(&buf, &r, A2, A0,
553 label_copy_pref_store);
554 build_copy_store(&buf, T3, off + 3 * copy_word_size);
555 off += 4 * copy_word_size;
559 if (pref_bias_copy_store) {
560 pg_addiu(&buf, A2, A0, pref_bias_copy_store);
561 uasm_l_copy_nopref(&l, buf);
564 build_copy_load(&buf, T0, off);
565 build_copy_load(&buf, T1, off + copy_word_size);
566 build_copy_load(&buf, T2, off + 2 * copy_word_size);
567 build_copy_load(&buf, T3, off + 3 * copy_word_size);
568 build_copy_store(&buf, T0, off);
569 build_copy_store(&buf, T1, off + copy_word_size);
570 build_copy_store(&buf, T2, off + 2 * copy_word_size);
571 build_copy_store(&buf, T3, off + 3 * copy_word_size);
572 off += 4 * copy_word_size;
573 } while (off < half_copy_loop_size);
574 pg_addiu(&buf, A1, A1, 2 * off);
575 pg_addiu(&buf, A0, A0, 2 * off);
578 build_copy_load(&buf, T0, off);
579 build_copy_load(&buf, T1, off + copy_word_size);
580 build_copy_load(&buf, T2, off + 2 * copy_word_size);
581 build_copy_load(&buf, T3, off + 3 * copy_word_size);
582 build_copy_store(&buf, T0, off);
583 build_copy_store(&buf, T1, off + copy_word_size);
584 build_copy_store(&buf, T2, off + 2 * copy_word_size);
585 if (off == -(4 * copy_word_size))
586 uasm_il_bne(&buf, &r, A2, A0,
588 build_copy_store(&buf, T3, off + 3 * copy_word_size);
589 off += 4 * copy_word_size;
596 BUG_ON(buf > &__copy_page_end);
598 uasm_resolve_relocs(relocs, labels);
600 pr_debug("Synthesized copy page handler (%u instructions).\n",
601 (u32)(buf - &__copy_page_start));
603 pr_debug("\t.set push\n");
604 pr_debug("\t.set noreorder\n");
605 for (i = 0; i < (buf - &__copy_page_start); i++)
606 pr_debug("\t.word 0x%08x\n", (&__copy_page_start)[i]);
607 pr_debug("\t.set pop\n");
610 #ifdef CONFIG_SIBYTE_DMA_PAGEOPS
611 extern void clear_page_cpu(void *page);
612 extern void copy_page_cpu(void *to, void *from);
615 * Pad descriptors to cacheline, since each is exclusively owned by a
623 } ____cacheline_aligned_in_smp page_descr[DM_NUM_CHANNELS];
625 void clear_page(void *page)
627 u64 to_phys = CPHYSADDR((unsigned long)page);
628 unsigned int cpu = smp_processor_id();
630 /* if the page is not in KSEG0, use old way */
631 if ((long)KSEGX((unsigned long)page) != (long)CKSEG0)
632 return clear_page_cpu(page);
634 page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_ZERO_MEM |
635 M_DM_DSCRA_L2C_DEST | M_DM_DSCRA_INTERRUPT;
636 page_descr[cpu].dscr_b = V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE);
637 __raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT)));
640 * Don't really want to do it this way, but there's no
641 * reliable way to delay completion detection.
643 while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG)))
644 & M_DM_DSCR_BASE_INTERRUPT))
646 __raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE)));
648 EXPORT_SYMBOL(clear_page);
650 void copy_page(void *to, void *from)
652 u64 from_phys = CPHYSADDR((unsigned long)from);
653 u64 to_phys = CPHYSADDR((unsigned long)to);
654 unsigned int cpu = smp_processor_id();
656 /* if any page is not in KSEG0, use old way */
657 if ((long)KSEGX((unsigned long)to) != (long)CKSEG0
658 || (long)KSEGX((unsigned long)from) != (long)CKSEG0)
659 return copy_page_cpu(to, from);
661 page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_L2C_DEST |
662 M_DM_DSCRA_INTERRUPT;
663 page_descr[cpu].dscr_b = from_phys | V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE);
664 __raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT)));
667 * Don't really want to do it this way, but there's no
668 * reliable way to delay completion detection.
670 while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG)))
671 & M_DM_DSCR_BASE_INTERRUPT))
673 __raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE)));
675 EXPORT_SYMBOL(copy_page);
677 #endif /* CONFIG_SIBYTE_DMA_PAGEOPS */