R4300 delay slot.
[powerpc.git] / arch / mips / mm / c-r4k.c
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) 1996 David S. Miller (dm@engr.sgi.com)
7  * Copyright (C) 1997, 1998, 1999, 2000, 2001, 2002 Ralf Baechle (ralf@gnu.org)
8  * Copyright (C) 1999, 2000 Silicon Graphics, Inc.
9  */
10 #include <linux/config.h>
11 #include <linux/init.h>
12 #include <linux/kernel.h>
13 #include <linux/sched.h>
14 #include <linux/mm.h>
15 #include <linux/bitops.h>
16
17 #include <asm/bcache.h>
18 #include <asm/bootinfo.h>
19 #include <asm/cacheops.h>
20 #include <asm/cpu.h>
21 #include <asm/cpu-features.h>
22 #include <asm/io.h>
23 #include <asm/page.h>
24 #include <asm/pgtable.h>
25 #include <asm/r4kcache.h>
26 #include <asm/system.h>
27 #include <asm/mmu_context.h>
28 #include <asm/war.h>
29
30 static unsigned long icache_size, dcache_size, scache_size;
31
32 /*
33  * Dummy cache handling routines for machines without boardcaches
34  */
35 static void no_sc_noop(void) {}
36
37 static struct bcache_ops no_sc_ops = {
38         .bc_enable = (void *)no_sc_noop,
39         .bc_disable = (void *)no_sc_noop,
40         .bc_wback_inv = (void *)no_sc_noop,
41         .bc_inv = (void *)no_sc_noop
42 };
43
44 struct bcache_ops *bcops = &no_sc_ops;
45
46 #define cpu_is_r4600_v1_x()     ((read_c0_prid() & 0xfffffff0) == 0x2010)
47 #define cpu_is_r4600_v2_x()     ((read_c0_prid() & 0xfffffff0) == 0x2020)
48
49 #define R4600_HIT_CACHEOP_WAR_IMPL                                      \
50 do {                                                                    \
51         if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())            \
52                 *(volatile unsigned long *)CKSEG1;                      \
53         if (R4600_V1_HIT_CACHEOP_WAR)                                   \
54                 __asm__ __volatile__("nop;nop;nop;nop");                \
55 } while (0)
56
57 static void (*r4k_blast_dcache_page)(unsigned long addr);
58
59 static inline void r4k_blast_dcache_page_dc32(unsigned long addr)
60 {
61         R4600_HIT_CACHEOP_WAR_IMPL;
62         blast_dcache32_page(addr);
63 }
64
65 static inline void r4k_blast_dcache_page_setup(void)
66 {
67         unsigned long  dc_lsize = cpu_dcache_line_size();
68
69         if (dc_lsize == 16)
70                 r4k_blast_dcache_page = blast_dcache16_page;
71         else if (dc_lsize == 32)
72                 r4k_blast_dcache_page = r4k_blast_dcache_page_dc32;
73 }
74
75 static void (* r4k_blast_dcache_page_indexed)(unsigned long addr);
76
77 static inline void r4k_blast_dcache_page_indexed_setup(void)
78 {
79         unsigned long dc_lsize = cpu_dcache_line_size();
80
81         if (dc_lsize == 16)
82                 r4k_blast_dcache_page_indexed = blast_dcache16_page_indexed;
83         else if (dc_lsize == 32)
84                 r4k_blast_dcache_page_indexed = blast_dcache32_page_indexed;
85 }
86
87 static void (* r4k_blast_dcache)(void);
88
89 static inline void r4k_blast_dcache_setup(void)
90 {
91         unsigned long dc_lsize = cpu_dcache_line_size();
92
93         if (dc_lsize == 16)
94                 r4k_blast_dcache = blast_dcache16;
95         else if (dc_lsize == 32)
96                 r4k_blast_dcache = blast_dcache32;
97 }
98
99 /* force code alignment (used for TX49XX_ICACHE_INDEX_INV_WAR) */
100 #define JUMP_TO_ALIGN(order) \
101         __asm__ __volatile__( \
102                 "b\t1f\n\t" \
103                 ".align\t" #order "\n\t" \
104                 "1:\n\t" \
105                 )
106 #define CACHE32_UNROLL32_ALIGN  JUMP_TO_ALIGN(10) /* 32 * 32 = 1024 */
107 #define CACHE32_UNROLL32_ALIGN2 JUMP_TO_ALIGN(11)
108
109 static inline void blast_r4600_v1_icache32(void)
110 {
111         unsigned long flags;
112
113         local_irq_save(flags);
114         blast_icache32();
115         local_irq_restore(flags);
116 }
117
118 static inline void tx49_blast_icache32(void)
119 {
120         unsigned long start = INDEX_BASE;
121         unsigned long end = start + current_cpu_data.icache.waysize;
122         unsigned long ws_inc = 1UL << current_cpu_data.icache.waybit;
123         unsigned long ws_end = current_cpu_data.icache.ways <<
124                                current_cpu_data.icache.waybit;
125         unsigned long ws, addr;
126
127         CACHE32_UNROLL32_ALIGN2;
128         /* I'm in even chunk.  blast odd chunks */
129         for (ws = 0; ws < ws_end; ws += ws_inc)
130                 for (addr = start + 0x400; addr < end; addr += 0x400 * 2)
131                         cache32_unroll32(addr|ws,Index_Invalidate_I);
132         CACHE32_UNROLL32_ALIGN;
133         /* I'm in odd chunk.  blast even chunks */
134         for (ws = 0; ws < ws_end; ws += ws_inc)
135                 for (addr = start; addr < end; addr += 0x400 * 2)
136                         cache32_unroll32(addr|ws,Index_Invalidate_I);
137 }
138
139 static inline void blast_icache32_r4600_v1_page_indexed(unsigned long page)
140 {
141         unsigned long flags;
142
143         local_irq_save(flags);
144         blast_icache32_page_indexed(page);
145         local_irq_restore(flags);
146 }
147
148 static inline void tx49_blast_icache32_page_indexed(unsigned long page)
149 {
150         unsigned long start = page;
151         unsigned long end = start + PAGE_SIZE;
152         unsigned long ws_inc = 1UL << current_cpu_data.icache.waybit;
153         unsigned long ws_end = current_cpu_data.icache.ways <<
154                                current_cpu_data.icache.waybit;
155         unsigned long ws, addr;
156
157         CACHE32_UNROLL32_ALIGN2;
158         /* I'm in even chunk.  blast odd chunks */
159         for (ws = 0; ws < ws_end; ws += ws_inc)
160                 for (addr = start + 0x400; addr < end; addr += 0x400 * 2)
161                         cache32_unroll32(addr|ws,Index_Invalidate_I);
162         CACHE32_UNROLL32_ALIGN;
163         /* I'm in odd chunk.  blast even chunks */
164         for (ws = 0; ws < ws_end; ws += ws_inc)
165                 for (addr = start; addr < end; addr += 0x400 * 2)
166                         cache32_unroll32(addr|ws,Index_Invalidate_I);
167 }
168
169 static void (* r4k_blast_icache_page)(unsigned long addr);
170
171 static inline void r4k_blast_icache_page_setup(void)
172 {
173         unsigned long ic_lsize = cpu_icache_line_size();
174
175         if (ic_lsize == 16)
176                 r4k_blast_icache_page = blast_icache16_page;
177         else if (ic_lsize == 32)
178                 r4k_blast_icache_page = blast_icache32_page;
179         else if (ic_lsize == 64)
180                 r4k_blast_icache_page = blast_icache64_page;
181 }
182
183
184 static void (* r4k_blast_icache_page_indexed)(unsigned long addr);
185
186 static inline void r4k_blast_icache_page_indexed_setup(void)
187 {
188         unsigned long ic_lsize = cpu_icache_line_size();
189
190         if (ic_lsize == 16)
191                 r4k_blast_icache_page_indexed = blast_icache16_page_indexed;
192         else if (ic_lsize == 32) {
193                 if (TX49XX_ICACHE_INDEX_INV_WAR)
194                         r4k_blast_icache_page_indexed =
195                                 tx49_blast_icache32_page_indexed;
196                 else if (R4600_V1_INDEX_ICACHEOP_WAR && cpu_is_r4600_v1_x())
197                         r4k_blast_icache_page_indexed =
198                                 blast_icache32_r4600_v1_page_indexed;
199                 else
200                         r4k_blast_icache_page_indexed =
201                                 blast_icache32_page_indexed;
202         } else if (ic_lsize == 64)
203                 r4k_blast_icache_page_indexed = blast_icache64_page_indexed;
204 }
205
206 static void (* r4k_blast_icache)(void);
207
208 static inline void r4k_blast_icache_setup(void)
209 {
210         unsigned long ic_lsize = cpu_icache_line_size();
211
212         if (ic_lsize == 16)
213                 r4k_blast_icache = blast_icache16;
214         else if (ic_lsize == 32) {
215                 if (R4600_V1_INDEX_ICACHEOP_WAR && cpu_is_r4600_v1_x())
216                         r4k_blast_icache = blast_r4600_v1_icache32;
217                 else if (TX49XX_ICACHE_INDEX_INV_WAR)
218                         r4k_blast_icache = tx49_blast_icache32;
219                 else
220                         r4k_blast_icache = blast_icache32;
221         } else if (ic_lsize == 64)
222                 r4k_blast_icache = blast_icache64;
223 }
224
225 static void (* r4k_blast_scache_page)(unsigned long addr);
226
227 static inline void r4k_blast_scache_page_setup(void)
228 {
229         unsigned long sc_lsize = cpu_scache_line_size();
230
231         if (sc_lsize == 16)
232                 r4k_blast_scache_page = blast_scache16_page;
233         else if (sc_lsize == 32)
234                 r4k_blast_scache_page = blast_scache32_page;
235         else if (sc_lsize == 64)
236                 r4k_blast_scache_page = blast_scache64_page;
237         else if (sc_lsize == 128)
238                 r4k_blast_scache_page = blast_scache128_page;
239 }
240
241 static void (* r4k_blast_scache_page_indexed)(unsigned long addr);
242
243 static inline void r4k_blast_scache_page_indexed_setup(void)
244 {
245         unsigned long sc_lsize = cpu_scache_line_size();
246
247         if (sc_lsize == 16)
248                 r4k_blast_scache_page_indexed = blast_scache16_page_indexed;
249         else if (sc_lsize == 32)
250                 r4k_blast_scache_page_indexed = blast_scache32_page_indexed;
251         else if (sc_lsize == 64)
252                 r4k_blast_scache_page_indexed = blast_scache64_page_indexed;
253         else if (sc_lsize == 128)
254                 r4k_blast_scache_page_indexed = blast_scache128_page_indexed;
255 }
256
257 static void (* r4k_blast_scache)(void);
258
259 static inline void r4k_blast_scache_setup(void)
260 {
261         unsigned long sc_lsize = cpu_scache_line_size();
262
263         if (sc_lsize == 16)
264                 r4k_blast_scache = blast_scache16;
265         else if (sc_lsize == 32)
266                 r4k_blast_scache = blast_scache32;
267         else if (sc_lsize == 64)
268                 r4k_blast_scache = blast_scache64;
269         else if (sc_lsize == 128)
270                 r4k_blast_scache = blast_scache128;
271 }
272
273 /*
274  * This is former mm's flush_cache_all() which really should be
275  * flush_cache_vunmap these days ...
276  */
277 static inline void local_r4k_flush_cache_all(void * args)
278 {
279         r4k_blast_dcache();
280         r4k_blast_icache();
281 }
282
283 static void r4k_flush_cache_all(void)
284 {
285         if (!cpu_has_dc_aliases)
286                 return;
287
288         on_each_cpu(local_r4k_flush_cache_all, NULL, 1, 1);
289 }
290
291 static inline void local_r4k___flush_cache_all(void * args)
292 {
293         r4k_blast_dcache();
294         r4k_blast_icache();
295
296         switch (current_cpu_data.cputype) {
297         case CPU_R4000SC:
298         case CPU_R4000MC:
299         case CPU_R4400SC:
300         case CPU_R4400MC:
301         case CPU_R10000:
302         case CPU_R12000:
303                 r4k_blast_scache();
304         }
305 }
306
307 static void r4k___flush_cache_all(void)
308 {
309         on_each_cpu(local_r4k___flush_cache_all, NULL, 1, 1);
310 }
311
312 static inline void local_r4k_flush_cache_range(void * args)
313 {
314         struct vm_area_struct *vma = args;
315         int exec;
316
317         if (!(cpu_context(smp_processor_id(), vma->vm_mm)))
318                 return;
319
320         exec = vma->vm_flags & VM_EXEC;
321         if (cpu_has_dc_aliases || exec)
322                 r4k_blast_dcache();
323         if (exec)
324                 r4k_blast_icache();
325 }
326
327 static void r4k_flush_cache_range(struct vm_area_struct *vma,
328         unsigned long start, unsigned long end)
329 {
330         on_each_cpu(local_r4k_flush_cache_range, vma, 1, 1);
331 }
332
333 static inline void local_r4k_flush_cache_mm(void * args)
334 {
335         struct mm_struct *mm = args;
336
337         if (!cpu_context(smp_processor_id(), mm))
338                 return;
339
340         r4k_blast_dcache();
341         r4k_blast_icache();
342
343         /*
344          * Kludge alert.  For obscure reasons R4000SC and R4400SC go nuts if we
345          * only flush the primary caches but R10000 and R12000 behave sane ...
346          */
347         if (current_cpu_data.cputype == CPU_R4000SC ||
348             current_cpu_data.cputype == CPU_R4000MC ||
349             current_cpu_data.cputype == CPU_R4400SC ||
350             current_cpu_data.cputype == CPU_R4400MC)
351                 r4k_blast_scache();
352 }
353
354 static void r4k_flush_cache_mm(struct mm_struct *mm)
355 {
356         if (!cpu_has_dc_aliases)
357                 return;
358
359         on_each_cpu(local_r4k_flush_cache_mm, mm, 1, 1);
360 }
361
362 struct flush_cache_page_args {
363         struct vm_area_struct *vma;
364         unsigned long page;
365 };
366
367 static inline void local_r4k_flush_cache_page(void *args)
368 {
369         struct flush_cache_page_args *fcp_args = args;
370         struct vm_area_struct *vma = fcp_args->vma;
371         unsigned long page = fcp_args->page;
372         int exec = vma->vm_flags & VM_EXEC;
373         struct mm_struct *mm = vma->vm_mm;
374         pgd_t *pgdp;
375         pud_t *pudp;
376         pmd_t *pmdp;
377         pte_t *ptep;
378
379         /*
380          * If ownes no valid ASID yet, cannot possibly have gotten
381          * this page into the cache.
382          */
383         if (cpu_context(smp_processor_id(), mm) == 0)
384                 return;
385
386         page &= PAGE_MASK;
387         pgdp = pgd_offset(mm, page);
388         pudp = pud_offset(pgdp, page);
389         pmdp = pmd_offset(pudp, page);
390         ptep = pte_offset(pmdp, page);
391
392         /*
393          * If the page isn't marked valid, the page cannot possibly be
394          * in the cache.
395          */
396         if (!(pte_val(*ptep) & _PAGE_PRESENT))
397                 return;
398
399         /*
400          * Doing flushes for another ASID than the current one is
401          * too difficult since stupid R4k caches do a TLB translation
402          * for every cache flush operation.  So we do indexed flushes
403          * in that case, which doesn't overly flush the cache too much.
404          */
405         if ((mm == current->active_mm) && (pte_val(*ptep) & _PAGE_VALID)) {
406                 if (cpu_has_dc_aliases || (exec && !cpu_has_ic_fills_f_dc)) {
407                         r4k_blast_dcache_page(page);
408                         if (exec && !cpu_icache_snoops_remote_store)
409                                 r4k_blast_scache_page(page);
410                 }
411                 if (exec)
412                         r4k_blast_icache_page(page);
413
414                 return;
415         }
416
417         /*
418          * Do indexed flush, too much work to get the (possible) TLB refills
419          * to work correctly.
420          */
421         page = INDEX_BASE + (page & (dcache_size - 1));
422         if (cpu_has_dc_aliases || (exec && !cpu_has_ic_fills_f_dc)) {
423                 r4k_blast_dcache_page_indexed(page);
424                 if (exec && !cpu_icache_snoops_remote_store)
425                         r4k_blast_scache_page_indexed(page);
426         }
427         if (exec) {
428                 if (cpu_has_vtag_icache) {
429                         int cpu = smp_processor_id();
430
431                         if (cpu_context(cpu, mm) != 0)
432                                 drop_mmu_context(mm, cpu);
433                 } else
434                         r4k_blast_icache_page_indexed(page);
435         }
436 }
437
438 static void r4k_flush_cache_page(struct vm_area_struct *vma, unsigned long page, unsigned long pfn)
439 {
440         struct flush_cache_page_args args;
441
442         args.vma = vma;
443         args.page = page;
444
445         on_each_cpu(local_r4k_flush_cache_page, &args, 1, 1);
446 }
447
448 static inline void local_r4k_flush_data_cache_page(void * addr)
449 {
450         r4k_blast_dcache_page((unsigned long) addr);
451 }
452
453 static void r4k_flush_data_cache_page(unsigned long addr)
454 {
455         on_each_cpu(local_r4k_flush_data_cache_page, (void *) addr, 1, 1);
456 }
457
458 struct flush_icache_range_args {
459         unsigned long __user start;
460         unsigned long __user end;
461 };
462
463 static inline void local_r4k_flush_icache_range(void *args)
464 {
465         struct flush_icache_range_args *fir_args = args;
466         unsigned long dc_lsize = current_cpu_data.dcache.linesz;
467         unsigned long ic_lsize = current_cpu_data.icache.linesz;
468         unsigned long sc_lsize = current_cpu_data.scache.linesz;
469         unsigned long start = fir_args->start;
470         unsigned long end = fir_args->end;
471         unsigned long addr, aend;
472
473         if (!cpu_has_ic_fills_f_dc) {
474                 if (end - start > dcache_size) {
475                         r4k_blast_dcache();
476                 } else {
477                         addr = start & ~(dc_lsize - 1);
478                         aend = (end - 1) & ~(dc_lsize - 1);
479
480                         while (1) {
481                                 /* Hit_Writeback_Inv_D */
482                                 protected_writeback_dcache_line(addr);
483                                 if (addr == aend)
484                                         break;
485                                 addr += dc_lsize;
486                         }
487                 }
488
489                 if (!cpu_icache_snoops_remote_store) {
490                         if (end - start > scache_size) {
491                                 r4k_blast_scache();
492                         } else {
493                                 addr = start & ~(sc_lsize - 1);
494                                 aend = (end - 1) & ~(sc_lsize - 1);
495
496                                 while (1) {
497                                         /* Hit_Writeback_Inv_D */
498                                         protected_writeback_scache_line(addr);
499                                         if (addr == aend)
500                                                 break;
501                                         addr += sc_lsize;
502                                 }
503                         }
504                 }
505         }
506
507         if (end - start > icache_size)
508                 r4k_blast_icache();
509         else {
510                 addr = start & ~(ic_lsize - 1);
511                 aend = (end - 1) & ~(ic_lsize - 1);
512                 while (1) {
513                         /* Hit_Invalidate_I */
514                         protected_flush_icache_line(addr);
515                         if (addr == aend)
516                                 break;
517                         addr += ic_lsize;
518                 }
519         }
520 }
521
522 static void r4k_flush_icache_range(unsigned long __user start,
523         unsigned long __user end)
524 {
525         struct flush_icache_range_args args;
526
527         args.start = start;
528         args.end = end;
529
530         on_each_cpu(local_r4k_flush_icache_range, &args, 1, 1);
531 }
532
533 /*
534  * Ok, this seriously sucks.  We use them to flush a user page but don't
535  * know the virtual address, so we have to blast away the whole icache
536  * which is significantly more expensive than the real thing.  Otoh we at
537  * least know the kernel address of the page so we can flush it
538  * selectivly.
539  */
540
541 struct flush_icache_page_args {
542         struct vm_area_struct *vma;
543         struct page *page;
544 };
545
546 static inline void local_r4k_flush_icache_page(void *args)
547 {
548         struct flush_icache_page_args *fip_args = args;
549         struct vm_area_struct *vma = fip_args->vma;
550         struct page *page = fip_args->page;
551
552         /*
553          * Tricky ...  Because we don't know the virtual address we've got the
554          * choice of either invalidating the entire primary and secondary
555          * caches or invalidating the secondary caches also.  With the subset
556          * enforcment on R4000SC, R4400SC, R10000 and R12000 invalidating the
557          * secondary cache will result in any entries in the primary caches
558          * also getting invalidated which hopefully is a bit more economical.
559          */
560         if (cpu_has_subset_pcaches) {
561                 unsigned long addr = (unsigned long) page_address(page);
562
563                 r4k_blast_scache_page(addr);
564                 ClearPageDcacheDirty(page);
565
566                 return;
567         }
568
569         if (!cpu_has_ic_fills_f_dc) {
570                 unsigned long addr = (unsigned long) page_address(page);
571                 r4k_blast_dcache_page(addr);
572                 if (!cpu_icache_snoops_remote_store)
573                         r4k_blast_scache_page(addr);
574                 ClearPageDcacheDirty(page);
575         }
576
577         /*
578          * We're not sure of the virtual address(es) involved here, so
579          * we have to flush the entire I-cache.
580          */
581         if (cpu_has_vtag_icache) {
582                 int cpu = smp_processor_id();
583
584                 if (cpu_context(cpu, vma->vm_mm) != 0)
585                         drop_mmu_context(vma->vm_mm, cpu);
586         } else
587                 r4k_blast_icache();
588 }
589
590 static void r4k_flush_icache_page(struct vm_area_struct *vma,
591         struct page *page)
592 {
593         struct flush_icache_page_args args;
594
595         /*
596          * If there's no context yet, or the page isn't executable, no I-cache
597          * flush is needed.
598          */
599         if (!(vma->vm_flags & VM_EXEC))
600                 return;
601
602         args.vma = vma;
603         args.page = page;
604
605         on_each_cpu(local_r4k_flush_icache_page, &args, 1, 1);
606 }
607
608
609 #ifdef CONFIG_DMA_NONCOHERENT
610
611 static void r4k_dma_cache_wback_inv(unsigned long addr, unsigned long size)
612 {
613         unsigned long end, a;
614
615         /* Catch bad driver code */
616         BUG_ON(size == 0);
617
618         if (cpu_has_subset_pcaches) {
619                 unsigned long sc_lsize = current_cpu_data.scache.linesz;
620
621                 if (size >= scache_size) {
622                         r4k_blast_scache();
623                         return;
624                 }
625
626                 a = addr & ~(sc_lsize - 1);
627                 end = (addr + size - 1) & ~(sc_lsize - 1);
628                 while (1) {
629                         flush_scache_line(a);   /* Hit_Writeback_Inv_SD */
630                         if (a == end)
631                                 break;
632                         a += sc_lsize;
633                 }
634                 return;
635         }
636
637         /*
638          * Either no secondary cache or the available caches don't have the
639          * subset property so we have to flush the primary caches
640          * explicitly
641          */
642         if (size >= dcache_size) {
643                 r4k_blast_dcache();
644         } else {
645                 unsigned long dc_lsize = current_cpu_data.dcache.linesz;
646
647                 R4600_HIT_CACHEOP_WAR_IMPL;
648                 a = addr & ~(dc_lsize - 1);
649                 end = (addr + size - 1) & ~(dc_lsize - 1);
650                 while (1) {
651                         flush_dcache_line(a);   /* Hit_Writeback_Inv_D */
652                         if (a == end)
653                                 break;
654                         a += dc_lsize;
655                 }
656         }
657
658         bc_wback_inv(addr, size);
659 }
660
661 static void r4k_dma_cache_inv(unsigned long addr, unsigned long size)
662 {
663         unsigned long end, a;
664
665         /* Catch bad driver code */
666         BUG_ON(size == 0);
667
668         if (cpu_has_subset_pcaches) {
669                 unsigned long sc_lsize = current_cpu_data.scache.linesz;
670
671                 if (size >= scache_size) {
672                         r4k_blast_scache();
673                         return;
674                 }
675
676                 a = addr & ~(sc_lsize - 1);
677                 end = (addr + size - 1) & ~(sc_lsize - 1);
678                 while (1) {
679                         flush_scache_line(a);   /* Hit_Writeback_Inv_SD */
680                         if (a == end)
681                                 break;
682                         a += sc_lsize;
683                 }
684                 return;
685         }
686
687         if (size >= dcache_size) {
688                 r4k_blast_dcache();
689         } else {
690                 unsigned long dc_lsize = current_cpu_data.dcache.linesz;
691
692                 R4600_HIT_CACHEOP_WAR_IMPL;
693                 a = addr & ~(dc_lsize - 1);
694                 end = (addr + size - 1) & ~(dc_lsize - 1);
695                 while (1) {
696                         flush_dcache_line(a);   /* Hit_Writeback_Inv_D */
697                         if (a == end)
698                                 break;
699                         a += dc_lsize;
700                 }
701         }
702
703         bc_inv(addr, size);
704 }
705 #endif /* CONFIG_DMA_NONCOHERENT */
706
707 /*
708  * While we're protected against bad userland addresses we don't care
709  * very much about what happens in that case.  Usually a segmentation
710  * fault will dump the process later on anyway ...
711  */
712 static void local_r4k_flush_cache_sigtramp(void * arg)
713 {
714         unsigned long ic_lsize = current_cpu_data.icache.linesz;
715         unsigned long dc_lsize = current_cpu_data.dcache.linesz;
716         unsigned long sc_lsize = current_cpu_data.scache.linesz;
717         unsigned long addr = (unsigned long) arg;
718
719         R4600_HIT_CACHEOP_WAR_IMPL;
720         protected_writeback_dcache_line(addr & ~(dc_lsize - 1));
721         if (!cpu_icache_snoops_remote_store)
722                 protected_writeback_scache_line(addr & ~(sc_lsize - 1));
723         protected_flush_icache_line(addr & ~(ic_lsize - 1));
724         if (MIPS4K_ICACHE_REFILL_WAR) {
725                 __asm__ __volatile__ (
726                         ".set push\n\t"
727                         ".set noat\n\t"
728                         ".set mips3\n\t"
729 #ifdef CONFIG_32BIT
730                         "la     $at,1f\n\t"
731 #endif
732 #ifdef CONFIG_64BIT
733                         "dla    $at,1f\n\t"
734 #endif
735                         "cache  %0,($at)\n\t"
736                         "nop; nop; nop\n"
737                         "1:\n\t"
738                         ".set pop"
739                         :
740                         : "i" (Hit_Invalidate_I));
741         }
742         if (MIPS_CACHE_SYNC_WAR)
743                 __asm__ __volatile__ ("sync");
744 }
745
746 static void r4k_flush_cache_sigtramp(unsigned long addr)
747 {
748         on_each_cpu(local_r4k_flush_cache_sigtramp, (void *) addr, 1, 1);
749 }
750
751 static void r4k_flush_icache_all(void)
752 {
753         if (cpu_has_vtag_icache)
754                 r4k_blast_icache();
755 }
756
757 static inline void rm7k_erratum31(void)
758 {
759         const unsigned long ic_lsize = 32;
760         unsigned long addr;
761
762         /* RM7000 erratum #31. The icache is screwed at startup. */
763         write_c0_taglo(0);
764         write_c0_taghi(0);
765
766         for (addr = INDEX_BASE; addr <= INDEX_BASE + 4096; addr += ic_lsize) {
767                 __asm__ __volatile__ (
768                         ".set noreorder\n\t"
769                         ".set mips3\n\t"
770                         "cache\t%1, 0(%0)\n\t"
771                         "cache\t%1, 0x1000(%0)\n\t"
772                         "cache\t%1, 0x2000(%0)\n\t"
773                         "cache\t%1, 0x3000(%0)\n\t"
774                         "cache\t%2, 0(%0)\n\t"
775                         "cache\t%2, 0x1000(%0)\n\t"
776                         "cache\t%2, 0x2000(%0)\n\t"
777                         "cache\t%2, 0x3000(%0)\n\t"
778                         "cache\t%1, 0(%0)\n\t"
779                         "cache\t%1, 0x1000(%0)\n\t"
780                         "cache\t%1, 0x2000(%0)\n\t"
781                         "cache\t%1, 0x3000(%0)\n\t"
782                         ".set\tmips0\n\t"
783                         ".set\treorder\n\t"
784                         :
785                         : "r" (addr), "i" (Index_Store_Tag_I), "i" (Fill));
786         }
787 }
788
789 static char *way_string[] __initdata = { NULL, "direct mapped", "2-way",
790         "3-way", "4-way", "5-way", "6-way", "7-way", "8-way"
791 };
792
793 static void __init probe_pcache(void)
794 {
795         struct cpuinfo_mips *c = &current_cpu_data;
796         unsigned int config = read_c0_config();
797         unsigned int prid = read_c0_prid();
798         unsigned long config1;
799         unsigned int lsize;
800
801         switch (c->cputype) {
802         case CPU_R4600:                 /* QED style two way caches? */
803         case CPU_R4700:
804         case CPU_R5000:
805         case CPU_NEVADA:
806                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
807                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
808                 c->icache.ways = 2;
809                 c->icache.waybit = ffs(icache_size/2) - 1;
810
811                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
812                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
813                 c->dcache.ways = 2;
814                 c->dcache.waybit= ffs(dcache_size/2) - 1;
815
816                 c->options |= MIPS_CPU_CACHE_CDEX_P;
817                 break;
818
819         case CPU_R5432:
820         case CPU_R5500:
821                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
822                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
823                 c->icache.ways = 2;
824                 c->icache.waybit= 0;
825
826                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
827                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
828                 c->dcache.ways = 2;
829                 c->dcache.waybit = 0;
830
831                 c->options |= MIPS_CPU_CACHE_CDEX_P;
832                 break;
833
834         case CPU_TX49XX:
835                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
836                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
837                 c->icache.ways = 4;
838                 c->icache.waybit= 0;
839
840                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
841                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
842                 c->dcache.ways = 4;
843                 c->dcache.waybit = 0;
844
845                 c->options |= MIPS_CPU_CACHE_CDEX_P;
846                 break;
847
848         case CPU_R4000PC:
849         case CPU_R4000SC:
850         case CPU_R4000MC:
851         case CPU_R4400PC:
852         case CPU_R4400SC:
853         case CPU_R4400MC:
854         case CPU_R4300:
855                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
856                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
857                 c->icache.ways = 1;
858                 c->icache.waybit = 0;   /* doesn't matter */
859
860                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
861                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
862                 c->dcache.ways = 1;
863                 c->dcache.waybit = 0;   /* does not matter */
864
865                 c->options |= MIPS_CPU_CACHE_CDEX_P;
866                 break;
867
868         case CPU_R10000:
869         case CPU_R12000:
870                 icache_size = 1 << (12 + ((config & R10K_CONF_IC) >> 29));
871                 c->icache.linesz = 64;
872                 c->icache.ways = 2;
873                 c->icache.waybit = 0;
874
875                 dcache_size = 1 << (12 + ((config & R10K_CONF_DC) >> 26));
876                 c->dcache.linesz = 32;
877                 c->dcache.ways = 2;
878                 c->dcache.waybit = 0;
879
880                 c->options |= MIPS_CPU_PREFETCH;
881                 break;
882
883         case CPU_VR4133:
884                 write_c0_config(config & ~CONF_EB);
885         case CPU_VR4131:
886                 /* Workaround for cache instruction bug of VR4131 */
887                 if (c->processor_id == 0x0c80U || c->processor_id == 0x0c81U ||
888                     c->processor_id == 0x0c82U) {
889                         config &= ~0x00000030U;
890                         config |= 0x00410000U;
891                         write_c0_config(config);
892                 }
893                 icache_size = 1 << (10 + ((config & CONF_IC) >> 9));
894                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
895                 c->icache.ways = 2;
896                 c->icache.waybit = ffs(icache_size/2) - 1;
897
898                 dcache_size = 1 << (10 + ((config & CONF_DC) >> 6));
899                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
900                 c->dcache.ways = 2;
901                 c->dcache.waybit = ffs(dcache_size/2) - 1;
902
903                 c->options |= MIPS_CPU_CACHE_CDEX_P;
904                 break;
905
906         case CPU_VR41XX:
907         case CPU_VR4111:
908         case CPU_VR4121:
909         case CPU_VR4122:
910         case CPU_VR4181:
911         case CPU_VR4181A:
912                 icache_size = 1 << (10 + ((config & CONF_IC) >> 9));
913                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
914                 c->icache.ways = 1;
915                 c->icache.waybit = 0;   /* doesn't matter */
916
917                 dcache_size = 1 << (10 + ((config & CONF_DC) >> 6));
918                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
919                 c->dcache.ways = 1;
920                 c->dcache.waybit = 0;   /* does not matter */
921
922                 c->options |= MIPS_CPU_CACHE_CDEX_P;
923                 break;
924
925         case CPU_RM7000:
926                 rm7k_erratum31();
927
928         case CPU_RM9000:
929                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
930                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
931                 c->icache.ways = 4;
932                 c->icache.waybit = ffs(icache_size / c->icache.ways) - 1;
933
934                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
935                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
936                 c->dcache.ways = 4;
937                 c->dcache.waybit = ffs(dcache_size / c->dcache.ways) - 1;
938
939 #if !defined(CONFIG_SMP) || !defined(RM9000_CDEX_SMP_WAR)
940                 c->options |= MIPS_CPU_CACHE_CDEX_P;
941 #endif
942                 c->options |= MIPS_CPU_PREFETCH;
943                 break;
944
945         default:
946                 if (!(config & MIPS_CONF_M))
947                         panic("Don't know how to probe P-caches on this cpu.");
948
949                 /*
950                  * So we seem to be a MIPS32 or MIPS64 CPU
951                  * So let's probe the I-cache ...
952                  */
953                 config1 = read_c0_config1();
954
955                 if ((lsize = ((config1 >> 19) & 7)))
956                         c->icache.linesz = 2 << lsize;
957                 else
958                         c->icache.linesz = lsize;
959                 c->icache.sets = 64 << ((config1 >> 22) & 7);
960                 c->icache.ways = 1 + ((config1 >> 16) & 7);
961
962                 icache_size = c->icache.sets *
963                               c->icache.ways *
964                               c->icache.linesz;
965                 c->icache.waybit = ffs(icache_size/c->icache.ways) - 1;
966
967                 if (config & 0x8)               /* VI bit */
968                         c->icache.flags |= MIPS_CACHE_VTAG;
969
970                 /*
971                  * Now probe the MIPS32 / MIPS64 data cache.
972                  */
973                 c->dcache.flags = 0;
974
975                 if ((lsize = ((config1 >> 10) & 7)))
976                         c->dcache.linesz = 2 << lsize;
977                 else
978                         c->dcache.linesz= lsize;
979                 c->dcache.sets = 64 << ((config1 >> 13) & 7);
980                 c->dcache.ways = 1 + ((config1 >> 7) & 7);
981
982                 dcache_size = c->dcache.sets *
983                               c->dcache.ways *
984                               c->dcache.linesz;
985                 c->dcache.waybit = ffs(dcache_size/c->dcache.ways) - 1;
986
987                 c->options |= MIPS_CPU_PREFETCH;
988                 break;
989         }
990
991         /*
992          * Processor configuration sanity check for the R4000SC erratum
993          * #5.  With page sizes larger than 32kB there is no possibility
994          * to get a VCE exception anymore so we don't care about this
995          * misconfiguration.  The case is rather theoretical anyway;
996          * presumably no vendor is shipping his hardware in the "bad"
997          * configuration.
998          */
999         if ((prid & 0xff00) == PRID_IMP_R4000 && (prid & 0xff) < 0x40 &&
1000             !(config & CONF_SC) && c->icache.linesz != 16 &&
1001             PAGE_SIZE <= 0x8000)
1002                 panic("Improper R4000SC processor configuration detected");
1003
1004         /* compute a couple of other cache variables */
1005         c->icache.waysize = icache_size / c->icache.ways;
1006         c->dcache.waysize = dcache_size / c->dcache.ways;
1007
1008         c->icache.sets = icache_size / (c->icache.linesz * c->icache.ways);
1009         c->dcache.sets = dcache_size / (c->dcache.linesz * c->dcache.ways);
1010
1011         /*
1012          * R10000 and R12000 P-caches are odd in a positive way.  They're 32kB
1013          * 2-way virtually indexed so normally would suffer from aliases.  So
1014          * normally they'd suffer from aliases but magic in the hardware deals
1015          * with that for us so we don't need to take care ourselves.
1016          */
1017         switch (c->cputype) {
1018         case CPU_20KC:
1019         case CPU_25KF:
1020         case CPU_R10000:
1021         case CPU_R12000:
1022         case CPU_SB1:
1023                 break;
1024         case CPU_24K:
1025                 if (!(read_c0_config7() & (1 << 16)))
1026         default:
1027                         if (c->dcache.waysize > PAGE_SIZE)
1028                                 c->dcache.flags |= MIPS_CACHE_ALIASES;
1029         }
1030
1031         switch (c->cputype) {
1032         case CPU_20KC:
1033                 /*
1034                  * Some older 20Kc chips doesn't have the 'VI' bit in
1035                  * the config register.
1036                  */
1037                 c->icache.flags |= MIPS_CACHE_VTAG;
1038                 break;
1039
1040         case CPU_AU1000:
1041         case CPU_AU1500:
1042         case CPU_AU1100:
1043         case CPU_AU1550:
1044         case CPU_AU1200:
1045                 c->icache.flags |= MIPS_CACHE_IC_F_DC;
1046                 break;
1047         }
1048
1049         printk("Primary instruction cache %ldkB, %s, %s, linesize %d bytes.\n",
1050                icache_size >> 10,
1051                cpu_has_vtag_icache ? "virtually tagged" : "physically tagged",
1052                way_string[c->icache.ways], c->icache.linesz);
1053
1054         printk("Primary data cache %ldkB, %s, linesize %d bytes.\n",
1055                dcache_size >> 10, way_string[c->dcache.ways], c->dcache.linesz);
1056 }
1057
1058 /*
1059  * If you even _breathe_ on this function, look at the gcc output and make sure
1060  * it does not pop things on and off the stack for the cache sizing loop that
1061  * executes in KSEG1 space or else you will crash and burn badly.  You have
1062  * been warned.
1063  */
1064 static int __init probe_scache(void)
1065 {
1066         extern unsigned long stext;
1067         unsigned long flags, addr, begin, end, pow2;
1068         unsigned int config = read_c0_config();
1069         struct cpuinfo_mips *c = &current_cpu_data;
1070         int tmp;
1071
1072         if (config & CONF_SC)
1073                 return 0;
1074
1075         begin = (unsigned long) &stext;
1076         begin &= ~((4 * 1024 * 1024) - 1);
1077         end = begin + (4 * 1024 * 1024);
1078
1079         /*
1080          * This is such a bitch, you'd think they would make it easy to do
1081          * this.  Away you daemons of stupidity!
1082          */
1083         local_irq_save(flags);
1084
1085         /* Fill each size-multiple cache line with a valid tag. */
1086         pow2 = (64 * 1024);
1087         for (addr = begin; addr < end; addr = (begin + pow2)) {
1088                 unsigned long *p = (unsigned long *) addr;
1089                 __asm__ __volatile__("nop" : : "r" (*p)); /* whee... */
1090                 pow2 <<= 1;
1091         }
1092
1093         /* Load first line with zero (therefore invalid) tag. */
1094         write_c0_taglo(0);
1095         write_c0_taghi(0);
1096         __asm__ __volatile__("nop; nop; nop; nop;"); /* avoid the hazard */
1097         cache_op(Index_Store_Tag_I, begin);
1098         cache_op(Index_Store_Tag_D, begin);
1099         cache_op(Index_Store_Tag_SD, begin);
1100
1101         /* Now search for the wrap around point. */
1102         pow2 = (128 * 1024);
1103         tmp = 0;
1104         for (addr = begin + (128 * 1024); addr < end; addr = begin + pow2) {
1105                 cache_op(Index_Load_Tag_SD, addr);
1106                 __asm__ __volatile__("nop; nop; nop; nop;"); /* hazard... */
1107                 if (!read_c0_taglo())
1108                         break;
1109                 pow2 <<= 1;
1110         }
1111         local_irq_restore(flags);
1112         addr -= begin;
1113
1114         scache_size = addr;
1115         c->scache.linesz = 16 << ((config & R4K_CONF_SB) >> 22);
1116         c->scache.ways = 1;
1117         c->dcache.waybit = 0;           /* does not matter */
1118
1119         return 1;
1120 }
1121
1122 typedef int (*probe_func_t)(unsigned long);
1123 extern int r5k_sc_init(void);
1124 extern int rm7k_sc_init(void);
1125
1126 static void __init setup_scache(void)
1127 {
1128         struct cpuinfo_mips *c = &current_cpu_data;
1129         unsigned int config = read_c0_config();
1130         probe_func_t probe_scache_kseg1;
1131         int sc_present = 0;
1132
1133         /*
1134          * Do the probing thing on R4000SC and R4400SC processors.  Other
1135          * processors don't have a S-cache that would be relevant to the
1136          * Linux memory managment.
1137          */
1138         switch (c->cputype) {
1139         case CPU_R4000SC:
1140         case CPU_R4000MC:
1141         case CPU_R4400SC:
1142         case CPU_R4400MC:
1143                 probe_scache_kseg1 = (probe_func_t) (CKSEG1ADDR(&probe_scache));
1144                 sc_present = probe_scache_kseg1(config);
1145                 if (sc_present)
1146                         c->options |= MIPS_CPU_CACHE_CDEX_S;
1147                 break;
1148
1149         case CPU_R10000:
1150         case CPU_R12000:
1151                 scache_size = 0x80000 << ((config & R10K_CONF_SS) >> 16);
1152                 c->scache.linesz = 64 << ((config >> 13) & 1);
1153                 c->scache.ways = 2;
1154                 c->scache.waybit= 0;
1155                 sc_present = 1;
1156                 break;
1157
1158         case CPU_R5000:
1159         case CPU_NEVADA:
1160 #ifdef CONFIG_R5000_CPU_SCACHE
1161                 r5k_sc_init();
1162 #endif
1163                 return;
1164
1165         case CPU_RM7000:
1166         case CPU_RM9000:
1167 #ifdef CONFIG_RM7000_CPU_SCACHE
1168                 rm7k_sc_init();
1169 #endif
1170                 return;
1171
1172         default:
1173                 sc_present = 0;
1174         }
1175
1176         if (!sc_present)
1177                 return;
1178
1179         if ((c->isa_level == MIPS_CPU_ISA_M32 ||
1180              c->isa_level == MIPS_CPU_ISA_M64) &&
1181             !(c->scache.flags & MIPS_CACHE_NOT_PRESENT))
1182                 panic("Dunno how to handle MIPS32 / MIPS64 second level cache");
1183
1184         /* compute a couple of other cache variables */
1185         c->scache.waysize = scache_size / c->scache.ways;
1186
1187         c->scache.sets = scache_size / (c->scache.linesz * c->scache.ways);
1188
1189         printk("Unified secondary cache %ldkB %s, linesize %d bytes.\n",
1190                scache_size >> 10, way_string[c->scache.ways], c->scache.linesz);
1191
1192         c->options |= MIPS_CPU_SUBSET_CACHES;
1193 }
1194
1195 static inline void coherency_setup(void)
1196 {
1197         change_c0_config(CONF_CM_CMASK, CONF_CM_DEFAULT);
1198
1199         /*
1200          * c0_status.cu=0 specifies that updates by the sc instruction use
1201          * the coherency mode specified by the TLB; 1 means cachable
1202          * coherent update on write will be used.  Not all processors have
1203          * this bit and; some wire it to zero, others like Toshiba had the
1204          * silly idea of putting something else there ...
1205          */
1206         switch (current_cpu_data.cputype) {
1207         case CPU_R4000PC:
1208         case CPU_R4000SC:
1209         case CPU_R4000MC:
1210         case CPU_R4400PC:
1211         case CPU_R4400SC:
1212         case CPU_R4400MC:
1213                 clear_c0_config(CONF_CU);
1214                 break;
1215         }
1216 }
1217
1218 void __init ld_mmu_r4xx0(void)
1219 {
1220         extern void build_clear_page(void);
1221         extern void build_copy_page(void);
1222         extern char except_vec2_generic;
1223         struct cpuinfo_mips *c = &current_cpu_data;
1224
1225         /* Default cache error handler for R4000 and R5000 family */
1226         memcpy((void *)(CAC_BASE   + 0x100), &except_vec2_generic, 0x80);
1227         memcpy((void *)(UNCAC_BASE + 0x100), &except_vec2_generic, 0x80);
1228
1229         probe_pcache();
1230         setup_scache();
1231
1232         r4k_blast_dcache_page_setup();
1233         r4k_blast_dcache_page_indexed_setup();
1234         r4k_blast_dcache_setup();
1235         r4k_blast_icache_page_setup();
1236         r4k_blast_icache_page_indexed_setup();
1237         r4k_blast_icache_setup();
1238         r4k_blast_scache_page_setup();
1239         r4k_blast_scache_page_indexed_setup();
1240         r4k_blast_scache_setup();
1241
1242         /*
1243          * Some MIPS32 and MIPS64 processors have physically indexed caches.
1244          * This code supports virtually indexed processors and will be
1245          * unnecessarily inefficient on physically indexed processors.
1246          */
1247         shm_align_mask = max_t( unsigned long,
1248                                 c->dcache.sets * c->dcache.linesz - 1,
1249                                 PAGE_SIZE - 1);
1250
1251         flush_cache_all         = r4k_flush_cache_all;
1252         __flush_cache_all       = r4k___flush_cache_all;
1253         flush_cache_mm          = r4k_flush_cache_mm;
1254         flush_cache_page        = r4k_flush_cache_page;
1255         flush_icache_page       = r4k_flush_icache_page;
1256         flush_cache_range       = r4k_flush_cache_range;
1257
1258         flush_cache_sigtramp    = r4k_flush_cache_sigtramp;
1259         flush_icache_all        = r4k_flush_icache_all;
1260         flush_data_cache_page   = r4k_flush_data_cache_page;
1261         flush_icache_range      = r4k_flush_icache_range;
1262
1263 #ifdef CONFIG_DMA_NONCOHERENT
1264         _dma_cache_wback_inv    = r4k_dma_cache_wback_inv;
1265         _dma_cache_wback        = r4k_dma_cache_wback_inv;
1266         _dma_cache_inv          = r4k_dma_cache_inv;
1267 #endif
1268
1269         __flush_cache_all();
1270         coherency_setup();
1271
1272         build_clear_page();
1273         build_copy_page();
1274 }