Merge branch 'upstream'
[pandora-kernel.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/cache.h>
20 #include <asm/cacheops.h>
21 #include <asm/cpu.h>
22 #include <asm/cpu-features.h>
23 #include <asm/io.h>
24 #include <asm/page.h>
25 #include <asm/pgtable.h>
26 #include <asm/r4kcache.h>
27 #include <asm/system.h>
28 #include <asm/mmu_context.h>
29 #include <asm/war.h>
30 #include <asm/cacheflush.h> /* for run_uncached() */
31
32 /*
33  * Must die.
34  */
35 static unsigned long icache_size __read_mostly;
36 static unsigned long dcache_size __read_mostly;
37 static unsigned long scache_size __read_mostly;
38
39 /*
40  * Dummy cache handling routines for machines without boardcaches
41  */
42 static void no_sc_noop(void) {}
43
44 static struct bcache_ops no_sc_ops = {
45         .bc_enable = (void *)no_sc_noop,
46         .bc_disable = (void *)no_sc_noop,
47         .bc_wback_inv = (void *)no_sc_noop,
48         .bc_inv = (void *)no_sc_noop
49 };
50
51 struct bcache_ops *bcops = &no_sc_ops;
52
53 #define cpu_is_r4600_v1_x()     ((read_c0_prid() & 0xfffffff0) == 0x00002010)
54 #define cpu_is_r4600_v2_x()     ((read_c0_prid() & 0xfffffff0) == 0x00002020)
55
56 #define R4600_HIT_CACHEOP_WAR_IMPL                                      \
57 do {                                                                    \
58         if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())            \
59                 *(volatile unsigned long *)CKSEG1;                      \
60         if (R4600_V1_HIT_CACHEOP_WAR)                                   \
61                 __asm__ __volatile__("nop;nop;nop;nop");                \
62 } while (0)
63
64 static void (*r4k_blast_dcache_page)(unsigned long addr);
65
66 static inline void r4k_blast_dcache_page_dc32(unsigned long addr)
67 {
68         R4600_HIT_CACHEOP_WAR_IMPL;
69         blast_dcache32_page(addr);
70 }
71
72 static inline void r4k_blast_dcache_page_setup(void)
73 {
74         unsigned long  dc_lsize = cpu_dcache_line_size();
75
76         if (dc_lsize == 16)
77                 r4k_blast_dcache_page = blast_dcache16_page;
78         else if (dc_lsize == 32)
79                 r4k_blast_dcache_page = r4k_blast_dcache_page_dc32;
80 }
81
82 static void (* r4k_blast_dcache_page_indexed)(unsigned long addr);
83
84 static inline void r4k_blast_dcache_page_indexed_setup(void)
85 {
86         unsigned long dc_lsize = cpu_dcache_line_size();
87
88         if (dc_lsize == 16)
89                 r4k_blast_dcache_page_indexed = blast_dcache16_page_indexed;
90         else if (dc_lsize == 32)
91                 r4k_blast_dcache_page_indexed = blast_dcache32_page_indexed;
92 }
93
94 static void (* r4k_blast_dcache)(void);
95
96 static inline void r4k_blast_dcache_setup(void)
97 {
98         unsigned long dc_lsize = cpu_dcache_line_size();
99
100         if (dc_lsize == 16)
101                 r4k_blast_dcache = blast_dcache16;
102         else if (dc_lsize == 32)
103                 r4k_blast_dcache = blast_dcache32;
104 }
105
106 /* force code alignment (used for TX49XX_ICACHE_INDEX_INV_WAR) */
107 #define JUMP_TO_ALIGN(order) \
108         __asm__ __volatile__( \
109                 "b\t1f\n\t" \
110                 ".align\t" #order "\n\t" \
111                 "1:\n\t" \
112                 )
113 #define CACHE32_UNROLL32_ALIGN  JUMP_TO_ALIGN(10) /* 32 * 32 = 1024 */
114 #define CACHE32_UNROLL32_ALIGN2 JUMP_TO_ALIGN(11)
115
116 static inline void blast_r4600_v1_icache32(void)
117 {
118         unsigned long flags;
119
120         local_irq_save(flags);
121         blast_icache32();
122         local_irq_restore(flags);
123 }
124
125 static inline void tx49_blast_icache32(void)
126 {
127         unsigned long start = INDEX_BASE;
128         unsigned long end = start + current_cpu_data.icache.waysize;
129         unsigned long ws_inc = 1UL << current_cpu_data.icache.waybit;
130         unsigned long ws_end = current_cpu_data.icache.ways <<
131                                current_cpu_data.icache.waybit;
132         unsigned long ws, addr;
133
134         CACHE32_UNROLL32_ALIGN2;
135         /* I'm in even chunk.  blast odd chunks */
136         for (ws = 0; ws < ws_end; ws += ws_inc)
137                 for (addr = start + 0x400; addr < end; addr += 0x400 * 2)
138                         cache32_unroll32(addr|ws,Index_Invalidate_I);
139         CACHE32_UNROLL32_ALIGN;
140         /* I'm in odd chunk.  blast even chunks */
141         for (ws = 0; ws < ws_end; ws += ws_inc)
142                 for (addr = start; addr < end; addr += 0x400 * 2)
143                         cache32_unroll32(addr|ws,Index_Invalidate_I);
144 }
145
146 static inline void blast_icache32_r4600_v1_page_indexed(unsigned long page)
147 {
148         unsigned long flags;
149
150         local_irq_save(flags);
151         blast_icache32_page_indexed(page);
152         local_irq_restore(flags);
153 }
154
155 static inline void tx49_blast_icache32_page_indexed(unsigned long page)
156 {
157         unsigned long start = page;
158         unsigned long end = start + PAGE_SIZE;
159         unsigned long ws_inc = 1UL << current_cpu_data.icache.waybit;
160         unsigned long ws_end = current_cpu_data.icache.ways <<
161                                current_cpu_data.icache.waybit;
162         unsigned long ws, addr;
163
164         CACHE32_UNROLL32_ALIGN2;
165         /* I'm in even chunk.  blast odd chunks */
166         for (ws = 0; ws < ws_end; ws += ws_inc)
167                 for (addr = start + 0x400; addr < end; addr += 0x400 * 2)
168                         cache32_unroll32(addr|ws,Index_Invalidate_I);
169         CACHE32_UNROLL32_ALIGN;
170         /* I'm in odd chunk.  blast even chunks */
171         for (ws = 0; ws < ws_end; ws += ws_inc)
172                 for (addr = start; addr < end; addr += 0x400 * 2)
173                         cache32_unroll32(addr|ws,Index_Invalidate_I);
174 }
175
176 static void (* r4k_blast_icache_page)(unsigned long addr);
177
178 static inline void r4k_blast_icache_page_setup(void)
179 {
180         unsigned long ic_lsize = cpu_icache_line_size();
181
182         if (ic_lsize == 16)
183                 r4k_blast_icache_page = blast_icache16_page;
184         else if (ic_lsize == 32)
185                 r4k_blast_icache_page = blast_icache32_page;
186         else if (ic_lsize == 64)
187                 r4k_blast_icache_page = blast_icache64_page;
188 }
189
190
191 static void (* r4k_blast_icache_page_indexed)(unsigned long addr);
192
193 static inline void r4k_blast_icache_page_indexed_setup(void)
194 {
195         unsigned long ic_lsize = cpu_icache_line_size();
196
197         if (ic_lsize == 16)
198                 r4k_blast_icache_page_indexed = blast_icache16_page_indexed;
199         else if (ic_lsize == 32) {
200                 if (R4600_V1_INDEX_ICACHEOP_WAR && cpu_is_r4600_v1_x())
201                         r4k_blast_icache_page_indexed =
202                                 blast_icache32_r4600_v1_page_indexed;
203                 else if (TX49XX_ICACHE_INDEX_INV_WAR)
204                         r4k_blast_icache_page_indexed =
205                                 tx49_blast_icache32_page_indexed;
206                 else
207                         r4k_blast_icache_page_indexed =
208                                 blast_icache32_page_indexed;
209         } else if (ic_lsize == 64)
210                 r4k_blast_icache_page_indexed = blast_icache64_page_indexed;
211 }
212
213 static void (* r4k_blast_icache)(void);
214
215 static inline void r4k_blast_icache_setup(void)
216 {
217         unsigned long ic_lsize = cpu_icache_line_size();
218
219         if (ic_lsize == 16)
220                 r4k_blast_icache = blast_icache16;
221         else if (ic_lsize == 32) {
222                 if (R4600_V1_INDEX_ICACHEOP_WAR && cpu_is_r4600_v1_x())
223                         r4k_blast_icache = blast_r4600_v1_icache32;
224                 else if (TX49XX_ICACHE_INDEX_INV_WAR)
225                         r4k_blast_icache = tx49_blast_icache32;
226                 else
227                         r4k_blast_icache = blast_icache32;
228         } else if (ic_lsize == 64)
229                 r4k_blast_icache = blast_icache64;
230 }
231
232 static void (* r4k_blast_scache_page)(unsigned long addr);
233
234 static inline void r4k_blast_scache_page_setup(void)
235 {
236         unsigned long sc_lsize = cpu_scache_line_size();
237
238         if (sc_lsize == 16)
239                 r4k_blast_scache_page = blast_scache16_page;
240         else if (sc_lsize == 32)
241                 r4k_blast_scache_page = blast_scache32_page;
242         else if (sc_lsize == 64)
243                 r4k_blast_scache_page = blast_scache64_page;
244         else if (sc_lsize == 128)
245                 r4k_blast_scache_page = blast_scache128_page;
246 }
247
248 static void (* r4k_blast_scache_page_indexed)(unsigned long addr);
249
250 static inline void r4k_blast_scache_page_indexed_setup(void)
251 {
252         unsigned long sc_lsize = cpu_scache_line_size();
253
254         if (sc_lsize == 16)
255                 r4k_blast_scache_page_indexed = blast_scache16_page_indexed;
256         else if (sc_lsize == 32)
257                 r4k_blast_scache_page_indexed = blast_scache32_page_indexed;
258         else if (sc_lsize == 64)
259                 r4k_blast_scache_page_indexed = blast_scache64_page_indexed;
260         else if (sc_lsize == 128)
261                 r4k_blast_scache_page_indexed = blast_scache128_page_indexed;
262 }
263
264 static void (* r4k_blast_scache)(void);
265
266 static inline void r4k_blast_scache_setup(void)
267 {
268         unsigned long sc_lsize = cpu_scache_line_size();
269
270         if (sc_lsize == 16)
271                 r4k_blast_scache = blast_scache16;
272         else if (sc_lsize == 32)
273                 r4k_blast_scache = blast_scache32;
274         else if (sc_lsize == 64)
275                 r4k_blast_scache = blast_scache64;
276         else if (sc_lsize == 128)
277                 r4k_blast_scache = blast_scache128;
278 }
279
280 /*
281  * This is former mm's flush_cache_all() which really should be
282  * flush_cache_vunmap these days ...
283  */
284 static inline void local_r4k_flush_cache_all(void * args)
285 {
286         r4k_blast_dcache();
287         r4k_blast_icache();
288 }
289
290 static void r4k_flush_cache_all(void)
291 {
292         if (!cpu_has_dc_aliases)
293                 return;
294
295         on_each_cpu(local_r4k_flush_cache_all, NULL, 1, 1);
296 }
297
298 static inline void local_r4k___flush_cache_all(void * args)
299 {
300         r4k_blast_dcache();
301         r4k_blast_icache();
302
303         switch (current_cpu_data.cputype) {
304         case CPU_R4000SC:
305         case CPU_R4000MC:
306         case CPU_R4400SC:
307         case CPU_R4400MC:
308         case CPU_R10000:
309         case CPU_R12000:
310                 r4k_blast_scache();
311         }
312 }
313
314 static void r4k___flush_cache_all(void)
315 {
316         on_each_cpu(local_r4k___flush_cache_all, NULL, 1, 1);
317 }
318
319 static inline void local_r4k_flush_cache_range(void * args)
320 {
321         struct vm_area_struct *vma = args;
322         int exec;
323
324         if (!(cpu_context(smp_processor_id(), vma->vm_mm)))
325                 return;
326
327         exec = vma->vm_flags & VM_EXEC;
328         if (cpu_has_dc_aliases || exec)
329                 r4k_blast_dcache();
330         if (exec)
331                 r4k_blast_icache();
332 }
333
334 static void r4k_flush_cache_range(struct vm_area_struct *vma,
335         unsigned long start, unsigned long end)
336 {
337         on_each_cpu(local_r4k_flush_cache_range, vma, 1, 1);
338 }
339
340 static inline void local_r4k_flush_cache_mm(void * args)
341 {
342         struct mm_struct *mm = args;
343
344         if (!cpu_context(smp_processor_id(), mm))
345                 return;
346
347         r4k_blast_dcache();
348         r4k_blast_icache();
349
350         /*
351          * Kludge alert.  For obscure reasons R4000SC and R4400SC go nuts if we
352          * only flush the primary caches but R10000 and R12000 behave sane ...
353          */
354         if (current_cpu_data.cputype == CPU_R4000SC ||
355             current_cpu_data.cputype == CPU_R4000MC ||
356             current_cpu_data.cputype == CPU_R4400SC ||
357             current_cpu_data.cputype == CPU_R4400MC)
358                 r4k_blast_scache();
359 }
360
361 static void r4k_flush_cache_mm(struct mm_struct *mm)
362 {
363         if (!cpu_has_dc_aliases)
364                 return;
365
366         on_each_cpu(local_r4k_flush_cache_mm, mm, 1, 1);
367 }
368
369 struct flush_cache_page_args {
370         struct vm_area_struct *vma;
371         unsigned long addr;
372 };
373
374 static inline void local_r4k_flush_cache_page(void *args)
375 {
376         struct flush_cache_page_args *fcp_args = args;
377         struct vm_area_struct *vma = fcp_args->vma;
378         unsigned long addr = fcp_args->addr;
379         int exec = vma->vm_flags & VM_EXEC;
380         struct mm_struct *mm = vma->vm_mm;
381         pgd_t *pgdp;
382         pud_t *pudp;
383         pmd_t *pmdp;
384         pte_t *ptep;
385
386         /*
387          * If ownes no valid ASID yet, cannot possibly have gotten
388          * this page into the cache.
389          */
390         if (cpu_context(smp_processor_id(), mm) == 0)
391                 return;
392
393         addr &= PAGE_MASK;
394         pgdp = pgd_offset(mm, addr);
395         pudp = pud_offset(pgdp, addr);
396         pmdp = pmd_offset(pudp, addr);
397         ptep = pte_offset(pmdp, addr);
398
399         /*
400          * If the page isn't marked valid, the page cannot possibly be
401          * in the cache.
402          */
403         if (!(pte_val(*ptep) & _PAGE_PRESENT))
404                 return;
405
406         /*
407          * Doing flushes for another ASID than the current one is
408          * too difficult since stupid R4k caches do a TLB translation
409          * for every cache flush operation.  So we do indexed flushes
410          * in that case, which doesn't overly flush the cache too much.
411          */
412         if ((mm == current->active_mm) && (pte_val(*ptep) & _PAGE_VALID)) {
413                 if (cpu_has_dc_aliases || (exec && !cpu_has_ic_fills_f_dc)) {
414                         r4k_blast_dcache_page(addr);
415                         if (exec && !cpu_icache_snoops_remote_store)
416                                 r4k_blast_scache_page(addr);
417                 }
418                 if (exec)
419                         r4k_blast_icache_page(addr);
420
421                 return;
422         }
423
424         /*
425          * Do indexed flush, too much work to get the (possible) TLB refills
426          * to work correctly.
427          */
428         addr = INDEX_BASE + (addr & (dcache_size - 1));
429         if (cpu_has_dc_aliases || (exec && !cpu_has_ic_fills_f_dc)) {
430                 r4k_blast_dcache_page_indexed(addr);
431                 if (exec && !cpu_icache_snoops_remote_store)
432                         r4k_blast_scache_page_indexed(addr);
433         }
434         if (exec) {
435                 if (cpu_has_vtag_icache) {
436                         int cpu = smp_processor_id();
437
438                         if (cpu_context(cpu, mm) != 0)
439                                 drop_mmu_context(mm, cpu);
440                 } else
441                         r4k_blast_icache_page_indexed(addr);
442         }
443 }
444
445 static void r4k_flush_cache_page(struct vm_area_struct *vma,
446         unsigned long addr, unsigned long pfn)
447 {
448         struct flush_cache_page_args args;
449
450         args.vma = vma;
451         args.addr = addr;
452
453         on_each_cpu(local_r4k_flush_cache_page, &args, 1, 1);
454 }
455
456 static inline void local_r4k_flush_data_cache_page(void * addr)
457 {
458         r4k_blast_dcache_page((unsigned long) addr);
459 }
460
461 static void r4k_flush_data_cache_page(unsigned long addr)
462 {
463         on_each_cpu(local_r4k_flush_data_cache_page, (void *) addr, 1, 1);
464 }
465
466 struct flush_icache_range_args {
467         unsigned long __user start;
468         unsigned long __user end;
469 };
470
471 static inline void local_r4k_flush_icache_range(void *args)
472 {
473         struct flush_icache_range_args *fir_args = args;
474         unsigned long dc_lsize = cpu_dcache_line_size();
475         unsigned long ic_lsize = cpu_icache_line_size();
476         unsigned long sc_lsize = cpu_scache_line_size();
477         unsigned long start = fir_args->start;
478         unsigned long end = fir_args->end;
479         unsigned long addr, aend;
480
481         if (!cpu_has_ic_fills_f_dc) {
482                 if (end - start > dcache_size) {
483                         r4k_blast_dcache();
484                 } else {
485                         R4600_HIT_CACHEOP_WAR_IMPL;
486                         addr = start & ~(dc_lsize - 1);
487                         aend = (end - 1) & ~(dc_lsize - 1);
488
489                         while (1) {
490                                 /* Hit_Writeback_Inv_D */
491                                 protected_writeback_dcache_line(addr);
492                                 if (addr == aend)
493                                         break;
494                                 addr += dc_lsize;
495                         }
496                 }
497
498                 if (!cpu_icache_snoops_remote_store) {
499                         if (end - start > scache_size) {
500                                 r4k_blast_scache();
501                         } else {
502                                 addr = start & ~(sc_lsize - 1);
503                                 aend = (end - 1) & ~(sc_lsize - 1);
504
505                                 while (1) {
506                                         /* Hit_Writeback_Inv_SD */
507                                         protected_writeback_scache_line(addr);
508                                         if (addr == aend)
509                                                 break;
510                                         addr += sc_lsize;
511                                 }
512                         }
513                 }
514         }
515
516         if (end - start > icache_size)
517                 r4k_blast_icache();
518         else {
519                 addr = start & ~(ic_lsize - 1);
520                 aend = (end - 1) & ~(ic_lsize - 1);
521                 while (1) {
522                         /* Hit_Invalidate_I */
523                         protected_flush_icache_line(addr);
524                         if (addr == aend)
525                                 break;
526                         addr += ic_lsize;
527                 }
528         }
529 }
530
531 static void r4k_flush_icache_range(unsigned long __user start,
532         unsigned long __user end)
533 {
534         struct flush_icache_range_args args;
535
536         args.start = start;
537         args.end = end;
538
539         on_each_cpu(local_r4k_flush_icache_range, &args, 1, 1);
540         instruction_hazard();
541 }
542
543 /*
544  * Ok, this seriously sucks.  We use them to flush a user page but don't
545  * know the virtual address, so we have to blast away the whole icache
546  * which is significantly more expensive than the real thing.  Otoh we at
547  * least know the kernel address of the page so we can flush it
548  * selectivly.
549  */
550
551 struct flush_icache_page_args {
552         struct vm_area_struct *vma;
553         struct page *page;
554 };
555
556 static inline void local_r4k_flush_icache_page(void *args)
557 {
558         struct flush_icache_page_args *fip_args = args;
559         struct vm_area_struct *vma = fip_args->vma;
560         struct page *page = fip_args->page;
561
562         /*
563          * Tricky ...  Because we don't know the virtual address we've got the
564          * choice of either invalidating the entire primary and secondary
565          * caches or invalidating the secondary caches also.  With the subset
566          * enforcment on R4000SC, R4400SC, R10000 and R12000 invalidating the
567          * secondary cache will result in any entries in the primary caches
568          * also getting invalidated which hopefully is a bit more economical.
569          */
570         if (cpu_has_subset_pcaches) {
571                 unsigned long addr = (unsigned long) page_address(page);
572
573                 r4k_blast_scache_page(addr);
574                 ClearPageDcacheDirty(page);
575
576                 return;
577         }
578
579         if (!cpu_has_ic_fills_f_dc) {
580                 unsigned long addr = (unsigned long) page_address(page);
581                 r4k_blast_dcache_page(addr);
582                 if (!cpu_icache_snoops_remote_store)
583                         r4k_blast_scache_page(addr);
584                 ClearPageDcacheDirty(page);
585         }
586
587         /*
588          * We're not sure of the virtual address(es) involved here, so
589          * we have to flush the entire I-cache.
590          */
591         if (cpu_has_vtag_icache) {
592                 int cpu = smp_processor_id();
593
594                 if (cpu_context(cpu, vma->vm_mm) != 0)
595                         drop_mmu_context(vma->vm_mm, cpu);
596         } else
597                 r4k_blast_icache();
598 }
599
600 static void r4k_flush_icache_page(struct vm_area_struct *vma,
601         struct page *page)
602 {
603         struct flush_icache_page_args args;
604
605         /*
606          * If there's no context yet, or the page isn't executable, no I-cache
607          * flush is needed.
608          */
609         if (!(vma->vm_flags & VM_EXEC))
610                 return;
611
612         args.vma = vma;
613         args.page = page;
614
615         on_each_cpu(local_r4k_flush_icache_page, &args, 1, 1);
616 }
617
618
619 #ifdef CONFIG_DMA_NONCOHERENT
620
621 static void r4k_dma_cache_wback_inv(unsigned long addr, unsigned long size)
622 {
623         unsigned long end, a;
624
625         /* Catch bad driver code */
626         BUG_ON(size == 0);
627
628         if (cpu_has_subset_pcaches) {
629                 unsigned long sc_lsize = cpu_scache_line_size();
630
631                 if (size >= scache_size) {
632                         r4k_blast_scache();
633                         return;
634                 }
635
636                 a = addr & ~(sc_lsize - 1);
637                 end = (addr + size - 1) & ~(sc_lsize - 1);
638                 while (1) {
639                         flush_scache_line(a);   /* Hit_Writeback_Inv_SD */
640                         if (a == end)
641                                 break;
642                         a += sc_lsize;
643                 }
644                 return;
645         }
646
647         /*
648          * Either no secondary cache or the available caches don't have the
649          * subset property so we have to flush the primary caches
650          * explicitly
651          */
652         if (size >= dcache_size) {
653                 r4k_blast_dcache();
654         } else {
655                 unsigned long dc_lsize = cpu_dcache_line_size();
656
657                 R4600_HIT_CACHEOP_WAR_IMPL;
658                 a = addr & ~(dc_lsize - 1);
659                 end = (addr + size - 1) & ~(dc_lsize - 1);
660                 while (1) {
661                         flush_dcache_line(a);   /* Hit_Writeback_Inv_D */
662                         if (a == end)
663                                 break;
664                         a += dc_lsize;
665                 }
666         }
667
668         bc_wback_inv(addr, size);
669 }
670
671 static void r4k_dma_cache_inv(unsigned long addr, unsigned long size)
672 {
673         unsigned long end, a;
674
675         /* Catch bad driver code */
676         BUG_ON(size == 0);
677
678         if (cpu_has_subset_pcaches) {
679                 unsigned long sc_lsize = cpu_scache_line_size();
680
681                 if (size >= scache_size) {
682                         r4k_blast_scache();
683                         return;
684                 }
685
686                 a = addr & ~(sc_lsize - 1);
687                 end = (addr + size - 1) & ~(sc_lsize - 1);
688                 while (1) {
689                         flush_scache_line(a);   /* Hit_Writeback_Inv_SD */
690                         if (a == end)
691                                 break;
692                         a += sc_lsize;
693                 }
694                 return;
695         }
696
697         if (size >= dcache_size) {
698                 r4k_blast_dcache();
699         } else {
700                 unsigned long dc_lsize = cpu_dcache_line_size();
701
702                 R4600_HIT_CACHEOP_WAR_IMPL;
703                 a = addr & ~(dc_lsize - 1);
704                 end = (addr + size - 1) & ~(dc_lsize - 1);
705                 while (1) {
706                         flush_dcache_line(a);   /* Hit_Writeback_Inv_D */
707                         if (a == end)
708                                 break;
709                         a += dc_lsize;
710                 }
711         }
712
713         bc_inv(addr, size);
714 }
715 #endif /* CONFIG_DMA_NONCOHERENT */
716
717 /*
718  * While we're protected against bad userland addresses we don't care
719  * very much about what happens in that case.  Usually a segmentation
720  * fault will dump the process later on anyway ...
721  */
722 static void local_r4k_flush_cache_sigtramp(void * arg)
723 {
724         unsigned long ic_lsize = cpu_icache_line_size();
725         unsigned long dc_lsize = cpu_dcache_line_size();
726         unsigned long sc_lsize = cpu_scache_line_size();
727         unsigned long addr = (unsigned long) arg;
728
729         R4600_HIT_CACHEOP_WAR_IMPL;
730         protected_writeback_dcache_line(addr & ~(dc_lsize - 1));
731         if (!cpu_icache_snoops_remote_store)
732                 protected_writeback_scache_line(addr & ~(sc_lsize - 1));
733         protected_flush_icache_line(addr & ~(ic_lsize - 1));
734         if (MIPS4K_ICACHE_REFILL_WAR) {
735                 __asm__ __volatile__ (
736                         ".set push\n\t"
737                         ".set noat\n\t"
738                         ".set mips3\n\t"
739 #ifdef CONFIG_32BIT
740                         "la     $at,1f\n\t"
741 #endif
742 #ifdef CONFIG_64BIT
743                         "dla    $at,1f\n\t"
744 #endif
745                         "cache  %0,($at)\n\t"
746                         "nop; nop; nop\n"
747                         "1:\n\t"
748                         ".set pop"
749                         :
750                         : "i" (Hit_Invalidate_I));
751         }
752         if (MIPS_CACHE_SYNC_WAR)
753                 __asm__ __volatile__ ("sync");
754 }
755
756 static void r4k_flush_cache_sigtramp(unsigned long addr)
757 {
758         on_each_cpu(local_r4k_flush_cache_sigtramp, (void *) addr, 1, 1);
759 }
760
761 static void r4k_flush_icache_all(void)
762 {
763         if (cpu_has_vtag_icache)
764                 r4k_blast_icache();
765 }
766
767 static inline void rm7k_erratum31(void)
768 {
769         const unsigned long ic_lsize = 32;
770         unsigned long addr;
771
772         /* RM7000 erratum #31. The icache is screwed at startup. */
773         write_c0_taglo(0);
774         write_c0_taghi(0);
775
776         for (addr = INDEX_BASE; addr <= INDEX_BASE + 4096; addr += ic_lsize) {
777                 __asm__ __volatile__ (
778                         ".set push\n\t"
779                         ".set noreorder\n\t"
780                         ".set mips3\n\t"
781                         "cache\t%1, 0(%0)\n\t"
782                         "cache\t%1, 0x1000(%0)\n\t"
783                         "cache\t%1, 0x2000(%0)\n\t"
784                         "cache\t%1, 0x3000(%0)\n\t"
785                         "cache\t%2, 0(%0)\n\t"
786                         "cache\t%2, 0x1000(%0)\n\t"
787                         "cache\t%2, 0x2000(%0)\n\t"
788                         "cache\t%2, 0x3000(%0)\n\t"
789                         "cache\t%1, 0(%0)\n\t"
790                         "cache\t%1, 0x1000(%0)\n\t"
791                         "cache\t%1, 0x2000(%0)\n\t"
792                         "cache\t%1, 0x3000(%0)\n\t"
793                         ".set pop\n"
794                         :
795                         : "r" (addr), "i" (Index_Store_Tag_I), "i" (Fill));
796         }
797 }
798
799 static char *way_string[] __initdata = { NULL, "direct mapped", "2-way",
800         "3-way", "4-way", "5-way", "6-way", "7-way", "8-way"
801 };
802
803 static void __init probe_pcache(void)
804 {
805         struct cpuinfo_mips *c = &current_cpu_data;
806         unsigned int config = read_c0_config();
807         unsigned int prid = read_c0_prid();
808         unsigned long config1;
809         unsigned int lsize;
810
811         switch (c->cputype) {
812         case CPU_R4600:                 /* QED style two way caches? */
813         case CPU_R4700:
814         case CPU_R5000:
815         case CPU_NEVADA:
816                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
817                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
818                 c->icache.ways = 2;
819                 c->icache.waybit = ffs(icache_size/2) - 1;
820
821                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
822                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
823                 c->dcache.ways = 2;
824                 c->dcache.waybit= ffs(dcache_size/2) - 1;
825
826                 c->options |= MIPS_CPU_CACHE_CDEX_P;
827                 break;
828
829         case CPU_R5432:
830         case CPU_R5500:
831                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
832                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
833                 c->icache.ways = 2;
834                 c->icache.waybit= 0;
835
836                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
837                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
838                 c->dcache.ways = 2;
839                 c->dcache.waybit = 0;
840
841                 c->options |= MIPS_CPU_CACHE_CDEX_P;
842                 break;
843
844         case CPU_TX49XX:
845                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
846                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
847                 c->icache.ways = 4;
848                 c->icache.waybit= 0;
849
850                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
851                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
852                 c->dcache.ways = 4;
853                 c->dcache.waybit = 0;
854
855                 c->options |= MIPS_CPU_CACHE_CDEX_P;
856                 break;
857
858         case CPU_R4000PC:
859         case CPU_R4000SC:
860         case CPU_R4000MC:
861         case CPU_R4400PC:
862         case CPU_R4400SC:
863         case CPU_R4400MC:
864         case CPU_R4300:
865                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
866                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
867                 c->icache.ways = 1;
868                 c->icache.waybit = 0;   /* doesn't matter */
869
870                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
871                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
872                 c->dcache.ways = 1;
873                 c->dcache.waybit = 0;   /* does not matter */
874
875                 c->options |= MIPS_CPU_CACHE_CDEX_P;
876                 break;
877
878         case CPU_R10000:
879         case CPU_R12000:
880                 icache_size = 1 << (12 + ((config & R10K_CONF_IC) >> 29));
881                 c->icache.linesz = 64;
882                 c->icache.ways = 2;
883                 c->icache.waybit = 0;
884
885                 dcache_size = 1 << (12 + ((config & R10K_CONF_DC) >> 26));
886                 c->dcache.linesz = 32;
887                 c->dcache.ways = 2;
888                 c->dcache.waybit = 0;
889
890                 c->options |= MIPS_CPU_PREFETCH;
891                 break;
892
893         case CPU_VR4133:
894                 write_c0_config(config & ~CONF_EB);
895         case CPU_VR4131:
896                 /* Workaround for cache instruction bug of VR4131 */
897                 if (c->processor_id == 0x0c80U || c->processor_id == 0x0c81U ||
898                     c->processor_id == 0x0c82U) {
899                         config &= ~0x00000030U;
900                         config |= 0x00410000U;
901                         write_c0_config(config);
902                 }
903                 icache_size = 1 << (10 + ((config & CONF_IC) >> 9));
904                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
905                 c->icache.ways = 2;
906                 c->icache.waybit = ffs(icache_size/2) - 1;
907
908                 dcache_size = 1 << (10 + ((config & CONF_DC) >> 6));
909                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
910                 c->dcache.ways = 2;
911                 c->dcache.waybit = ffs(dcache_size/2) - 1;
912
913                 c->options |= MIPS_CPU_CACHE_CDEX_P;
914                 break;
915
916         case CPU_VR41XX:
917         case CPU_VR4111:
918         case CPU_VR4121:
919         case CPU_VR4122:
920         case CPU_VR4181:
921         case CPU_VR4181A:
922                 icache_size = 1 << (10 + ((config & CONF_IC) >> 9));
923                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
924                 c->icache.ways = 1;
925                 c->icache.waybit = 0;   /* doesn't matter */
926
927                 dcache_size = 1 << (10 + ((config & CONF_DC) >> 6));
928                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
929                 c->dcache.ways = 1;
930                 c->dcache.waybit = 0;   /* does not matter */
931
932                 c->options |= MIPS_CPU_CACHE_CDEX_P;
933                 break;
934
935         case CPU_RM7000:
936                 rm7k_erratum31();
937
938         case CPU_RM9000:
939                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
940                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
941                 c->icache.ways = 4;
942                 c->icache.waybit = ffs(icache_size / c->icache.ways) - 1;
943
944                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
945                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
946                 c->dcache.ways = 4;
947                 c->dcache.waybit = ffs(dcache_size / c->dcache.ways) - 1;
948
949 #if !defined(CONFIG_SMP) || !defined(RM9000_CDEX_SMP_WAR)
950                 c->options |= MIPS_CPU_CACHE_CDEX_P;
951 #endif
952                 c->options |= MIPS_CPU_PREFETCH;
953                 break;
954
955         default:
956                 if (!(config & MIPS_CONF_M))
957                         panic("Don't know how to probe P-caches on this cpu.");
958
959                 /*
960                  * So we seem to be a MIPS32 or MIPS64 CPU
961                  * So let's probe the I-cache ...
962                  */
963                 config1 = read_c0_config1();
964
965                 if ((lsize = ((config1 >> 19) & 7)))
966                         c->icache.linesz = 2 << lsize;
967                 else
968                         c->icache.linesz = lsize;
969                 c->icache.sets = 64 << ((config1 >> 22) & 7);
970                 c->icache.ways = 1 + ((config1 >> 16) & 7);
971
972                 icache_size = c->icache.sets *
973                               c->icache.ways *
974                               c->icache.linesz;
975                 c->icache.waybit = ffs(icache_size/c->icache.ways) - 1;
976
977                 if (config & 0x8)               /* VI bit */
978                         c->icache.flags |= MIPS_CACHE_VTAG;
979
980                 /*
981                  * Now probe the MIPS32 / MIPS64 data cache.
982                  */
983                 c->dcache.flags = 0;
984
985                 if ((lsize = ((config1 >> 10) & 7)))
986                         c->dcache.linesz = 2 << lsize;
987                 else
988                         c->dcache.linesz= lsize;
989                 c->dcache.sets = 64 << ((config1 >> 13) & 7);
990                 c->dcache.ways = 1 + ((config1 >> 7) & 7);
991
992                 dcache_size = c->dcache.sets *
993                               c->dcache.ways *
994                               c->dcache.linesz;
995                 c->dcache.waybit = ffs(dcache_size/c->dcache.ways) - 1;
996
997                 c->options |= MIPS_CPU_PREFETCH;
998                 break;
999         }
1000
1001         /*
1002          * Processor configuration sanity check for the R4000SC erratum
1003          * #5.  With page sizes larger than 32kB there is no possibility
1004          * to get a VCE exception anymore so we don't care about this
1005          * misconfiguration.  The case is rather theoretical anyway;
1006          * presumably no vendor is shipping his hardware in the "bad"
1007          * configuration.
1008          */
1009         if ((prid & 0xff00) == PRID_IMP_R4000 && (prid & 0xff) < 0x40 &&
1010             !(config & CONF_SC) && c->icache.linesz != 16 &&
1011             PAGE_SIZE <= 0x8000)
1012                 panic("Improper R4000SC processor configuration detected");
1013
1014         /* compute a couple of other cache variables */
1015         c->icache.waysize = icache_size / c->icache.ways;
1016         c->dcache.waysize = dcache_size / c->dcache.ways;
1017
1018         c->icache.sets = icache_size / (c->icache.linesz * c->icache.ways);
1019         c->dcache.sets = dcache_size / (c->dcache.linesz * c->dcache.ways);
1020
1021         /*
1022          * R10000 and R12000 P-caches are odd in a positive way.  They're 32kB
1023          * 2-way virtually indexed so normally would suffer from aliases.  So
1024          * normally they'd suffer from aliases but magic in the hardware deals
1025          * with that for us so we don't need to take care ourselves.
1026          */
1027         switch (c->cputype) {
1028         case CPU_20KC:
1029         case CPU_25KF:
1030         case CPU_R10000:
1031         case CPU_R12000:
1032         case CPU_SB1:
1033                 break;
1034         case CPU_24K:
1035                 if (!(read_c0_config7() & (1 << 16)))
1036         default:
1037                         if (c->dcache.waysize > PAGE_SIZE)
1038                                 c->dcache.flags |= MIPS_CACHE_ALIASES;
1039         }
1040
1041         switch (c->cputype) {
1042         case CPU_20KC:
1043                 /*
1044                  * Some older 20Kc chips doesn't have the 'VI' bit in
1045                  * the config register.
1046                  */
1047                 c->icache.flags |= MIPS_CACHE_VTAG;
1048                 break;
1049
1050         case CPU_AU1000:
1051         case CPU_AU1500:
1052         case CPU_AU1100:
1053         case CPU_AU1550:
1054         case CPU_AU1200:
1055                 c->icache.flags |= MIPS_CACHE_IC_F_DC;
1056                 break;
1057         }
1058
1059         printk("Primary instruction cache %ldkB, %s, %s, linesize %d bytes.\n",
1060                icache_size >> 10,
1061                cpu_has_vtag_icache ? "virtually tagged" : "physically tagged",
1062                way_string[c->icache.ways], c->icache.linesz);
1063
1064         printk("Primary data cache %ldkB, %s, linesize %d bytes.\n",
1065                dcache_size >> 10, way_string[c->dcache.ways], c->dcache.linesz);
1066 }
1067
1068 /*
1069  * If you even _breathe_ on this function, look at the gcc output and make sure
1070  * it does not pop things on and off the stack for the cache sizing loop that
1071  * executes in KSEG1 space or else you will crash and burn badly.  You have
1072  * been warned.
1073  */
1074 static int __init probe_scache(void)
1075 {
1076         extern unsigned long stext;
1077         unsigned long flags, addr, begin, end, pow2;
1078         unsigned int config = read_c0_config();
1079         struct cpuinfo_mips *c = &current_cpu_data;
1080         int tmp;
1081
1082         if (config & CONF_SC)
1083                 return 0;
1084
1085         begin = (unsigned long) &stext;
1086         begin &= ~((4 * 1024 * 1024) - 1);
1087         end = begin + (4 * 1024 * 1024);
1088
1089         /*
1090          * This is such a bitch, you'd think they would make it easy to do
1091          * this.  Away you daemons of stupidity!
1092          */
1093         local_irq_save(flags);
1094
1095         /* Fill each size-multiple cache line with a valid tag. */
1096         pow2 = (64 * 1024);
1097         for (addr = begin; addr < end; addr = (begin + pow2)) {
1098                 unsigned long *p = (unsigned long *) addr;
1099                 __asm__ __volatile__("nop" : : "r" (*p)); /* whee... */
1100                 pow2 <<= 1;
1101         }
1102
1103         /* Load first line with zero (therefore invalid) tag. */
1104         write_c0_taglo(0);
1105         write_c0_taghi(0);
1106         __asm__ __volatile__("nop; nop; nop; nop;"); /* avoid the hazard */
1107         cache_op(Index_Store_Tag_I, begin);
1108         cache_op(Index_Store_Tag_D, begin);
1109         cache_op(Index_Store_Tag_SD, begin);
1110
1111         /* Now search for the wrap around point. */
1112         pow2 = (128 * 1024);
1113         tmp = 0;
1114         for (addr = begin + (128 * 1024); addr < end; addr = begin + pow2) {
1115                 cache_op(Index_Load_Tag_SD, addr);
1116                 __asm__ __volatile__("nop; nop; nop; nop;"); /* hazard... */
1117                 if (!read_c0_taglo())
1118                         break;
1119                 pow2 <<= 1;
1120         }
1121         local_irq_restore(flags);
1122         addr -= begin;
1123
1124         scache_size = addr;
1125         c->scache.linesz = 16 << ((config & R4K_CONF_SB) >> 22);
1126         c->scache.ways = 1;
1127         c->dcache.waybit = 0;           /* does not matter */
1128
1129         return 1;
1130 }
1131
1132 extern int r5k_sc_init(void);
1133 extern int rm7k_sc_init(void);
1134
1135 static void __init setup_scache(void)
1136 {
1137         struct cpuinfo_mips *c = &current_cpu_data;
1138         unsigned int config = read_c0_config();
1139         int sc_present = 0;
1140
1141         /*
1142          * Do the probing thing on R4000SC and R4400SC processors.  Other
1143          * processors don't have a S-cache that would be relevant to the
1144          * Linux memory managment.
1145          */
1146         switch (c->cputype) {
1147         case CPU_R4000SC:
1148         case CPU_R4000MC:
1149         case CPU_R4400SC:
1150         case CPU_R4400MC:
1151                 sc_present = run_uncached(probe_scache);
1152                 if (sc_present)
1153                         c->options |= MIPS_CPU_CACHE_CDEX_S;
1154                 break;
1155
1156         case CPU_R10000:
1157         case CPU_R12000:
1158                 scache_size = 0x80000 << ((config & R10K_CONF_SS) >> 16);
1159                 c->scache.linesz = 64 << ((config >> 13) & 1);
1160                 c->scache.ways = 2;
1161                 c->scache.waybit= 0;
1162                 sc_present = 1;
1163                 break;
1164
1165         case CPU_R5000:
1166         case CPU_NEVADA:
1167 #ifdef CONFIG_R5000_CPU_SCACHE
1168                 r5k_sc_init();
1169 #endif
1170                 return;
1171
1172         case CPU_RM7000:
1173         case CPU_RM9000:
1174 #ifdef CONFIG_RM7000_CPU_SCACHE
1175                 rm7k_sc_init();
1176 #endif
1177                 return;
1178
1179         default:
1180                 sc_present = 0;
1181         }
1182
1183         if (!sc_present)
1184                 return;
1185
1186         if ((c->isa_level == MIPS_CPU_ISA_M32R1 ||
1187              c->isa_level == MIPS_CPU_ISA_M64R1) &&
1188             !(c->scache.flags & MIPS_CACHE_NOT_PRESENT))
1189                 panic("Dunno how to handle MIPS32 / MIPS64 second level cache");
1190
1191         /* compute a couple of other cache variables */
1192         c->scache.waysize = scache_size / c->scache.ways;
1193
1194         c->scache.sets = scache_size / (c->scache.linesz * c->scache.ways);
1195
1196         printk("Unified secondary cache %ldkB %s, linesize %d bytes.\n",
1197                scache_size >> 10, way_string[c->scache.ways], c->scache.linesz);
1198
1199         c->options |= MIPS_CPU_SUBSET_CACHES;
1200 }
1201
1202 static inline void coherency_setup(void)
1203 {
1204         change_c0_config(CONF_CM_CMASK, CONF_CM_DEFAULT);
1205
1206         /*
1207          * c0_status.cu=0 specifies that updates by the sc instruction use
1208          * the coherency mode specified by the TLB; 1 means cachable
1209          * coherent update on write will be used.  Not all processors have
1210          * this bit and; some wire it to zero, others like Toshiba had the
1211          * silly idea of putting something else there ...
1212          */
1213         switch (current_cpu_data.cputype) {
1214         case CPU_R4000PC:
1215         case CPU_R4000SC:
1216         case CPU_R4000MC:
1217         case CPU_R4400PC:
1218         case CPU_R4400SC:
1219         case CPU_R4400MC:
1220                 clear_c0_config(CONF_CU);
1221                 break;
1222         }
1223 }
1224
1225 void __init r4k_cache_init(void)
1226 {
1227         extern void build_clear_page(void);
1228         extern void build_copy_page(void);
1229         extern char except_vec2_generic;
1230         struct cpuinfo_mips *c = &current_cpu_data;
1231
1232         /* Default cache error handler for R4000 and R5000 family */
1233         set_uncached_handler (0x100, &except_vec2_generic, 0x80);
1234
1235         probe_pcache();
1236         setup_scache();
1237
1238         r4k_blast_dcache_page_setup();
1239         r4k_blast_dcache_page_indexed_setup();
1240         r4k_blast_dcache_setup();
1241         r4k_blast_icache_page_setup();
1242         r4k_blast_icache_page_indexed_setup();
1243         r4k_blast_icache_setup();
1244         r4k_blast_scache_page_setup();
1245         r4k_blast_scache_page_indexed_setup();
1246         r4k_blast_scache_setup();
1247
1248         /*
1249          * Some MIPS32 and MIPS64 processors have physically indexed caches.
1250          * This code supports virtually indexed processors and will be
1251          * unnecessarily inefficient on physically indexed processors.
1252          */
1253         shm_align_mask = max_t( unsigned long,
1254                                 c->dcache.sets * c->dcache.linesz - 1,
1255                                 PAGE_SIZE - 1);
1256
1257         flush_cache_all         = r4k_flush_cache_all;
1258         __flush_cache_all       = r4k___flush_cache_all;
1259         flush_cache_mm          = r4k_flush_cache_mm;
1260         flush_cache_page        = r4k_flush_cache_page;
1261         flush_icache_page       = r4k_flush_icache_page;
1262         flush_cache_range       = r4k_flush_cache_range;
1263
1264         flush_cache_sigtramp    = r4k_flush_cache_sigtramp;
1265         flush_icache_all        = r4k_flush_icache_all;
1266         flush_data_cache_page   = r4k_flush_data_cache_page;
1267         flush_icache_range      = r4k_flush_icache_range;
1268
1269 #ifdef CONFIG_DMA_NONCOHERENT
1270         _dma_cache_wback_inv    = r4k_dma_cache_wback_inv;
1271         _dma_cache_wback        = r4k_dma_cache_wback_inv;
1272         _dma_cache_inv          = r4k_dma_cache_inv;
1273 #endif
1274
1275         build_clear_page();
1276         build_copy_page();
1277         local_r4k___flush_cache_all(NULL);
1278         coherency_setup();
1279 }