linux 2.6.16.38 w/ vs2.0.3-rc1
[linux-2.6.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 indexmask = current_cpu_data.icache.waysize - 1;
158         unsigned long start = INDEX_BASE + (page & indexmask);
159         unsigned long end = start + PAGE_SIZE;
160         unsigned long ws_inc = 1UL << current_cpu_data.icache.waybit;
161         unsigned long ws_end = current_cpu_data.icache.ways <<
162                                current_cpu_data.icache.waybit;
163         unsigned long ws, addr;
164
165         CACHE32_UNROLL32_ALIGN2;
166         /* I'm in even chunk.  blast odd chunks */
167         for (ws = 0; ws < ws_end; ws += ws_inc)
168                 for (addr = start + 0x400; addr < end; addr += 0x400 * 2)
169                         cache32_unroll32(addr|ws,Index_Invalidate_I);
170         CACHE32_UNROLL32_ALIGN;
171         /* I'm in odd chunk.  blast even chunks */
172         for (ws = 0; ws < ws_end; ws += ws_inc)
173                 for (addr = start; addr < end; addr += 0x400 * 2)
174                         cache32_unroll32(addr|ws,Index_Invalidate_I);
175 }
176
177 static void (* r4k_blast_icache_page)(unsigned long addr);
178
179 static inline void r4k_blast_icache_page_setup(void)
180 {
181         unsigned long ic_lsize = cpu_icache_line_size();
182
183         if (ic_lsize == 16)
184                 r4k_blast_icache_page = blast_icache16_page;
185         else if (ic_lsize == 32)
186                 r4k_blast_icache_page = blast_icache32_page;
187         else if (ic_lsize == 64)
188                 r4k_blast_icache_page = blast_icache64_page;
189 }
190
191
192 static void (* r4k_blast_icache_page_indexed)(unsigned long addr);
193
194 static inline void r4k_blast_icache_page_indexed_setup(void)
195 {
196         unsigned long ic_lsize = cpu_icache_line_size();
197
198         if (ic_lsize == 16)
199                 r4k_blast_icache_page_indexed = blast_icache16_page_indexed;
200         else if (ic_lsize == 32) {
201                 if (R4600_V1_INDEX_ICACHEOP_WAR && cpu_is_r4600_v1_x())
202                         r4k_blast_icache_page_indexed =
203                                 blast_icache32_r4600_v1_page_indexed;
204                 else if (TX49XX_ICACHE_INDEX_INV_WAR)
205                         r4k_blast_icache_page_indexed =
206                                 tx49_blast_icache32_page_indexed;
207                 else
208                         r4k_blast_icache_page_indexed =
209                                 blast_icache32_page_indexed;
210         } else if (ic_lsize == 64)
211                 r4k_blast_icache_page_indexed = blast_icache64_page_indexed;
212 }
213
214 static void (* r4k_blast_icache)(void);
215
216 static inline void r4k_blast_icache_setup(void)
217 {
218         unsigned long ic_lsize = cpu_icache_line_size();
219
220         if (ic_lsize == 16)
221                 r4k_blast_icache = blast_icache16;
222         else if (ic_lsize == 32) {
223                 if (R4600_V1_INDEX_ICACHEOP_WAR && cpu_is_r4600_v1_x())
224                         r4k_blast_icache = blast_r4600_v1_icache32;
225                 else if (TX49XX_ICACHE_INDEX_INV_WAR)
226                         r4k_blast_icache = tx49_blast_icache32;
227                 else
228                         r4k_blast_icache = blast_icache32;
229         } else if (ic_lsize == 64)
230                 r4k_blast_icache = blast_icache64;
231 }
232
233 static void (* r4k_blast_scache_page)(unsigned long addr);
234
235 static inline void r4k_blast_scache_page_setup(void)
236 {
237         unsigned long sc_lsize = cpu_scache_line_size();
238
239         if (scache_size == 0)
240                 r4k_blast_scache_page = (void *)no_sc_noop;
241         else if (sc_lsize == 16)
242                 r4k_blast_scache_page = blast_scache16_page;
243         else if (sc_lsize == 32)
244                 r4k_blast_scache_page = blast_scache32_page;
245         else if (sc_lsize == 64)
246                 r4k_blast_scache_page = blast_scache64_page;
247         else if (sc_lsize == 128)
248                 r4k_blast_scache_page = blast_scache128_page;
249 }
250
251 static void (* r4k_blast_scache_page_indexed)(unsigned long addr);
252
253 static inline void r4k_blast_scache_page_indexed_setup(void)
254 {
255         unsigned long sc_lsize = cpu_scache_line_size();
256
257         if (scache_size == 0)
258                 r4k_blast_scache_page_indexed = (void *)no_sc_noop;
259         else if (sc_lsize == 16)
260                 r4k_blast_scache_page_indexed = blast_scache16_page_indexed;
261         else if (sc_lsize == 32)
262                 r4k_blast_scache_page_indexed = blast_scache32_page_indexed;
263         else if (sc_lsize == 64)
264                 r4k_blast_scache_page_indexed = blast_scache64_page_indexed;
265         else if (sc_lsize == 128)
266                 r4k_blast_scache_page_indexed = blast_scache128_page_indexed;
267 }
268
269 static void (* r4k_blast_scache)(void);
270
271 static inline void r4k_blast_scache_setup(void)
272 {
273         unsigned long sc_lsize = cpu_scache_line_size();
274
275         if (scache_size == 0)
276                 r4k_blast_scache = (void *)no_sc_noop;
277         else if (sc_lsize == 16)
278                 r4k_blast_scache = blast_scache16;
279         else if (sc_lsize == 32)
280                 r4k_blast_scache = blast_scache32;
281         else if (sc_lsize == 64)
282                 r4k_blast_scache = blast_scache64;
283         else if (sc_lsize == 128)
284                 r4k_blast_scache = blast_scache128;
285 }
286
287 /*
288  * This is former mm's flush_cache_all() which really should be
289  * flush_cache_vunmap these days ...
290  */
291 static inline void local_r4k_flush_cache_all(void * args)
292 {
293         r4k_blast_dcache();
294         r4k_blast_icache();
295 }
296
297 static void r4k_flush_cache_all(void)
298 {
299         if (!cpu_has_dc_aliases)
300                 return;
301
302         on_each_cpu(local_r4k_flush_cache_all, NULL, 1, 1);
303 }
304
305 static inline void local_r4k___flush_cache_all(void * args)
306 {
307         r4k_blast_dcache();
308         r4k_blast_icache();
309
310         switch (current_cpu_data.cputype) {
311         case CPU_R4000SC:
312         case CPU_R4000MC:
313         case CPU_R4400SC:
314         case CPU_R4400MC:
315         case CPU_R10000:
316         case CPU_R12000:
317                 r4k_blast_scache();
318         }
319 }
320
321 static void r4k___flush_cache_all(void)
322 {
323         on_each_cpu(local_r4k___flush_cache_all, NULL, 1, 1);
324 }
325
326 static inline void local_r4k_flush_cache_range(void * args)
327 {
328         struct vm_area_struct *vma = args;
329         int exec;
330
331         if (!(cpu_context(smp_processor_id(), vma->vm_mm)))
332                 return;
333
334         exec = vma->vm_flags & VM_EXEC;
335         if (cpu_has_dc_aliases || exec)
336                 r4k_blast_dcache();
337         if (exec)
338                 r4k_blast_icache();
339 }
340
341 static void r4k_flush_cache_range(struct vm_area_struct *vma,
342         unsigned long start, unsigned long end)
343 {
344         on_each_cpu(local_r4k_flush_cache_range, vma, 1, 1);
345 }
346
347 static inline void local_r4k_flush_cache_mm(void * args)
348 {
349         struct mm_struct *mm = args;
350
351         if (!cpu_context(smp_processor_id(), mm))
352                 return;
353
354         r4k_blast_dcache();
355         r4k_blast_icache();
356
357         /*
358          * Kludge alert.  For obscure reasons R4000SC and R4400SC go nuts if we
359          * only flush the primary caches but R10000 and R12000 behave sane ...
360          */
361         if (current_cpu_data.cputype == CPU_R4000SC ||
362             current_cpu_data.cputype == CPU_R4000MC ||
363             current_cpu_data.cputype == CPU_R4400SC ||
364             current_cpu_data.cputype == CPU_R4400MC)
365                 r4k_blast_scache();
366 }
367
368 static void r4k_flush_cache_mm(struct mm_struct *mm)
369 {
370         if (!cpu_has_dc_aliases)
371                 return;
372
373         on_each_cpu(local_r4k_flush_cache_mm, mm, 1, 1);
374 }
375
376 struct flush_cache_page_args {
377         struct vm_area_struct *vma;
378         unsigned long addr;
379         unsigned long pfn;
380 };
381
382 static inline void local_r4k_flush_cache_page(void *args)
383 {
384         struct flush_cache_page_args *fcp_args = args;
385         struct vm_area_struct *vma = fcp_args->vma;
386         unsigned long addr = fcp_args->addr;
387         unsigned long paddr = fcp_args->pfn << PAGE_SHIFT;
388         int exec = vma->vm_flags & VM_EXEC;
389         struct mm_struct *mm = vma->vm_mm;
390         pgd_t *pgdp;
391         pud_t *pudp;
392         pmd_t *pmdp;
393         pte_t *ptep;
394
395         /*
396          * If ownes no valid ASID yet, cannot possibly have gotten
397          * this page into the cache.
398          */
399         if (cpu_context(smp_processor_id(), mm) == 0)
400                 return;
401
402         addr &= PAGE_MASK;
403         pgdp = pgd_offset(mm, addr);
404         pudp = pud_offset(pgdp, addr);
405         pmdp = pmd_offset(pudp, addr);
406         ptep = pte_offset(pmdp, addr);
407
408         /*
409          * If the page isn't marked valid, the page cannot possibly be
410          * in the cache.
411          */
412         if (!(pte_val(*ptep) & _PAGE_PRESENT))
413                 return;
414
415         /*
416          * Doing flushes for another ASID than the current one is
417          * too difficult since stupid R4k caches do a TLB translation
418          * for every cache flush operation.  So we do indexed flushes
419          * in that case, which doesn't overly flush the cache too much.
420          */
421         if ((mm == current->active_mm) && (pte_val(*ptep) & _PAGE_VALID)) {
422                 if (cpu_has_dc_aliases || (exec && !cpu_has_ic_fills_f_dc)) {
423                         r4k_blast_dcache_page(addr);
424                         if (exec && !cpu_icache_snoops_remote_store)
425                                 r4k_blast_scache_page(addr);
426                 }
427                 if (exec)
428                         r4k_blast_icache_page(addr);
429
430                 return;
431         }
432
433         /*
434          * Do indexed flush, too much work to get the (possible) TLB refills
435          * to work correctly.
436          */
437         if (cpu_has_dc_aliases || (exec && !cpu_has_ic_fills_f_dc)) {
438                 r4k_blast_dcache_page_indexed(cpu_has_pindexed_dcache ?
439                                               paddr : addr);
440                 if (exec && !cpu_icache_snoops_remote_store) {
441                         r4k_blast_scache_page_indexed(paddr);
442                 }
443         }
444         if (exec) {
445                 if (cpu_has_vtag_icache) {
446                         int cpu = smp_processor_id();
447
448                         if (cpu_context(cpu, mm) != 0)
449                                 drop_mmu_context(mm, cpu);
450                 } else
451                         r4k_blast_icache_page_indexed(addr);
452         }
453 }
454
455 static void r4k_flush_cache_page(struct vm_area_struct *vma,
456         unsigned long addr, unsigned long pfn)
457 {
458         struct flush_cache_page_args args;
459
460         args.vma = vma;
461         args.addr = addr;
462         args.pfn = pfn;
463
464         on_each_cpu(local_r4k_flush_cache_page, &args, 1, 1);
465 }
466
467 static inline void local_r4k_flush_data_cache_page(void * addr)
468 {
469         r4k_blast_dcache_page((unsigned long) addr);
470 }
471
472 static void r4k_flush_data_cache_page(unsigned long addr)
473 {
474         on_each_cpu(local_r4k_flush_data_cache_page, (void *) addr, 1, 1);
475 }
476
477 struct flush_icache_range_args {
478         unsigned long start;
479         unsigned long end;
480 };
481
482 static inline void local_r4k_flush_icache_range(void *args)
483 {
484         struct flush_icache_range_args *fir_args = args;
485         unsigned long start = fir_args->start;
486         unsigned long end = fir_args->end;
487
488         if (!cpu_has_ic_fills_f_dc) {
489                 if (end - start > dcache_size) {
490                         r4k_blast_dcache();
491                 } else {
492                         R4600_HIT_CACHEOP_WAR_IMPL;
493                         protected_blast_dcache_range(start, end);
494                 }
495
496                 if (!cpu_icache_snoops_remote_store && scache_size) {
497                         if (end - start > scache_size)
498                                 r4k_blast_scache();
499                         else
500                                 protected_blast_scache_range(start, end);
501                 }
502         }
503
504         if (end - start > icache_size)
505                 r4k_blast_icache();
506         else
507                 protected_blast_icache_range(start, end);
508 }
509
510 static void r4k_flush_icache_range(unsigned long start, unsigned long end)
511 {
512         struct flush_icache_range_args args;
513
514         args.start = start;
515         args.end = end;
516
517         on_each_cpu(local_r4k_flush_icache_range, &args, 1, 1);
518         instruction_hazard();
519 }
520
521 /*
522  * Ok, this seriously sucks.  We use them to flush a user page but don't
523  * know the virtual address, so we have to blast away the whole icache
524  * which is significantly more expensive than the real thing.  Otoh we at
525  * least know the kernel address of the page so we can flush it
526  * selectivly.
527  */
528
529 struct flush_icache_page_args {
530         struct vm_area_struct *vma;
531         struct page *page;
532 };
533
534 static inline void local_r4k_flush_icache_page(void *args)
535 {
536         struct flush_icache_page_args *fip_args = args;
537         struct vm_area_struct *vma = fip_args->vma;
538         struct page *page = fip_args->page;
539
540         /*
541          * Tricky ...  Because we don't know the virtual address we've got the
542          * choice of either invalidating the entire primary and secondary
543          * caches or invalidating the secondary caches also.  With the subset
544          * enforcment on R4000SC, R4400SC, R10000 and R12000 invalidating the
545          * secondary cache will result in any entries in the primary caches
546          * also getting invalidated which hopefully is a bit more economical.
547          */
548         if (cpu_has_subset_pcaches) {
549                 unsigned long addr = (unsigned long) page_address(page);
550
551                 r4k_blast_scache_page(addr);
552                 ClearPageDcacheDirty(page);
553
554                 return;
555         }
556
557         if (!cpu_has_ic_fills_f_dc) {
558                 unsigned long addr = (unsigned long) page_address(page);
559                 r4k_blast_dcache_page(addr);
560                 if (!cpu_icache_snoops_remote_store)
561                         r4k_blast_scache_page(addr);
562                 ClearPageDcacheDirty(page);
563         }
564
565         /*
566          * We're not sure of the virtual address(es) involved here, so
567          * we have to flush the entire I-cache.
568          */
569         if (cpu_has_vtag_icache) {
570                 int cpu = smp_processor_id();
571
572                 if (cpu_context(cpu, vma->vm_mm) != 0)
573                         drop_mmu_context(vma->vm_mm, cpu);
574         } else
575                 r4k_blast_icache();
576 }
577
578 static void r4k_flush_icache_page(struct vm_area_struct *vma,
579         struct page *page)
580 {
581         struct flush_icache_page_args args;
582
583         /*
584          * If there's no context yet, or the page isn't executable, no I-cache
585          * flush is needed.
586          */
587         if (!(vma->vm_flags & VM_EXEC))
588                 return;
589
590         args.vma = vma;
591         args.page = page;
592
593         on_each_cpu(local_r4k_flush_icache_page, &args, 1, 1);
594 }
595
596
597 #ifdef CONFIG_DMA_NONCOHERENT
598
599 static void r4k_dma_cache_wback_inv(unsigned long addr, unsigned long size)
600 {
601         /* Catch bad driver code */
602         BUG_ON(size == 0);
603
604         if (cpu_has_subset_pcaches) {
605                 if (size >= scache_size)
606                         r4k_blast_scache();
607                 else
608                         blast_scache_range(addr, addr + size);
609                 return;
610         }
611
612         /*
613          * Either no secondary cache or the available caches don't have the
614          * subset property so we have to flush the primary caches
615          * explicitly
616          */
617         if (size >= dcache_size) {
618                 r4k_blast_dcache();
619         } else {
620                 R4600_HIT_CACHEOP_WAR_IMPL;
621                 blast_dcache_range(addr, addr + size);
622         }
623
624         bc_wback_inv(addr, size);
625 }
626
627 static void r4k_dma_cache_inv(unsigned long addr, unsigned long size)
628 {
629         /* Catch bad driver code */
630         BUG_ON(size == 0);
631
632         if (cpu_has_subset_pcaches) {
633                 if (size >= scache_size)
634                         r4k_blast_scache();
635                 else
636                         blast_scache_range(addr, addr + size);
637                 return;
638         }
639
640         if (size >= dcache_size) {
641                 r4k_blast_dcache();
642         } else {
643                 R4600_HIT_CACHEOP_WAR_IMPL;
644                 blast_dcache_range(addr, addr + size);
645         }
646
647         bc_inv(addr, size);
648 }
649 #endif /* CONFIG_DMA_NONCOHERENT */
650
651 /*
652  * While we're protected against bad userland addresses we don't care
653  * very much about what happens in that case.  Usually a segmentation
654  * fault will dump the process later on anyway ...
655  */
656 static void local_r4k_flush_cache_sigtramp(void * arg)
657 {
658         unsigned long ic_lsize = cpu_icache_line_size();
659         unsigned long dc_lsize = cpu_dcache_line_size();
660         unsigned long sc_lsize = cpu_scache_line_size();
661         unsigned long addr = (unsigned long) arg;
662
663         R4600_HIT_CACHEOP_WAR_IMPL;
664         protected_writeback_dcache_line(addr & ~(dc_lsize - 1));
665         if (!cpu_icache_snoops_remote_store && scache_size)
666                 protected_writeback_scache_line(addr & ~(sc_lsize - 1));
667         protected_flush_icache_line(addr & ~(ic_lsize - 1));
668         if (MIPS4K_ICACHE_REFILL_WAR) {
669                 __asm__ __volatile__ (
670                         ".set push\n\t"
671                         ".set noat\n\t"
672                         ".set mips3\n\t"
673 #ifdef CONFIG_32BIT
674                         "la     $at,1f\n\t"
675 #endif
676 #ifdef CONFIG_64BIT
677                         "dla    $at,1f\n\t"
678 #endif
679                         "cache  %0,($at)\n\t"
680                         "nop; nop; nop\n"
681                         "1:\n\t"
682                         ".set pop"
683                         :
684                         : "i" (Hit_Invalidate_I));
685         }
686         if (MIPS_CACHE_SYNC_WAR)
687                 __asm__ __volatile__ ("sync");
688 }
689
690 static void r4k_flush_cache_sigtramp(unsigned long addr)
691 {
692         on_each_cpu(local_r4k_flush_cache_sigtramp, (void *) addr, 1, 1);
693 }
694
695 static void r4k_flush_icache_all(void)
696 {
697         if (cpu_has_vtag_icache)
698                 r4k_blast_icache();
699 }
700
701 static inline void rm7k_erratum31(void)
702 {
703         const unsigned long ic_lsize = 32;
704         unsigned long addr;
705
706         /* RM7000 erratum #31. The icache is screwed at startup. */
707         write_c0_taglo(0);
708         write_c0_taghi(0);
709
710         for (addr = INDEX_BASE; addr <= INDEX_BASE + 4096; addr += ic_lsize) {
711                 __asm__ __volatile__ (
712                         ".set push\n\t"
713                         ".set noreorder\n\t"
714                         ".set mips3\n\t"
715                         "cache\t%1, 0(%0)\n\t"
716                         "cache\t%1, 0x1000(%0)\n\t"
717                         "cache\t%1, 0x2000(%0)\n\t"
718                         "cache\t%1, 0x3000(%0)\n\t"
719                         "cache\t%2, 0(%0)\n\t"
720                         "cache\t%2, 0x1000(%0)\n\t"
721                         "cache\t%2, 0x2000(%0)\n\t"
722                         "cache\t%2, 0x3000(%0)\n\t"
723                         "cache\t%1, 0(%0)\n\t"
724                         "cache\t%1, 0x1000(%0)\n\t"
725                         "cache\t%1, 0x2000(%0)\n\t"
726                         "cache\t%1, 0x3000(%0)\n\t"
727                         ".set pop\n"
728                         :
729                         : "r" (addr), "i" (Index_Store_Tag_I), "i" (Fill));
730         }
731 }
732
733 static char *way_string[] __initdata = { NULL, "direct mapped", "2-way",
734         "3-way", "4-way", "5-way", "6-way", "7-way", "8-way"
735 };
736
737 static void __init probe_pcache(void)
738 {
739         struct cpuinfo_mips *c = &current_cpu_data;
740         unsigned int config = read_c0_config();
741         unsigned int prid = read_c0_prid();
742         unsigned long config1;
743         unsigned int lsize;
744
745         switch (c->cputype) {
746         case CPU_R4600:                 /* QED style two way caches? */
747         case CPU_R4700:
748         case CPU_R5000:
749         case CPU_NEVADA:
750                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
751                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
752                 c->icache.ways = 2;
753                 c->icache.waybit = ffs(icache_size/2) - 1;
754
755                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
756                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
757                 c->dcache.ways = 2;
758                 c->dcache.waybit= ffs(dcache_size/2) - 1;
759
760                 c->options |= MIPS_CPU_CACHE_CDEX_P;
761                 break;
762
763         case CPU_R5432:
764         case CPU_R5500:
765                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
766                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
767                 c->icache.ways = 2;
768                 c->icache.waybit= 0;
769
770                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
771                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
772                 c->dcache.ways = 2;
773                 c->dcache.waybit = 0;
774
775                 c->options |= MIPS_CPU_CACHE_CDEX_P;
776                 break;
777
778         case CPU_TX49XX:
779                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
780                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
781                 c->icache.ways = 4;
782                 c->icache.waybit= 0;
783
784                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
785                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
786                 c->dcache.ways = 4;
787                 c->dcache.waybit = 0;
788
789                 c->options |= MIPS_CPU_CACHE_CDEX_P;
790                 break;
791
792         case CPU_R4000PC:
793         case CPU_R4000SC:
794         case CPU_R4000MC:
795         case CPU_R4400PC:
796         case CPU_R4400SC:
797         case CPU_R4400MC:
798         case CPU_R4300:
799                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
800                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
801                 c->icache.ways = 1;
802                 c->icache.waybit = 0;   /* doesn't matter */
803
804                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
805                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
806                 c->dcache.ways = 1;
807                 c->dcache.waybit = 0;   /* does not matter */
808
809                 c->options |= MIPS_CPU_CACHE_CDEX_P;
810                 break;
811
812         case CPU_R10000:
813         case CPU_R12000:
814                 icache_size = 1 << (12 + ((config & R10K_CONF_IC) >> 29));
815                 c->icache.linesz = 64;
816                 c->icache.ways = 2;
817                 c->icache.waybit = 0;
818
819                 dcache_size = 1 << (12 + ((config & R10K_CONF_DC) >> 26));
820                 c->dcache.linesz = 32;
821                 c->dcache.ways = 2;
822                 c->dcache.waybit = 0;
823
824                 c->options |= MIPS_CPU_PREFETCH;
825                 break;
826
827         case CPU_VR4133:
828                 write_c0_config(config & ~CONF_EB);
829         case CPU_VR4131:
830                 /* Workaround for cache instruction bug of VR4131 */
831                 if (c->processor_id == 0x0c80U || c->processor_id == 0x0c81U ||
832                     c->processor_id == 0x0c82U) {
833                         config &= ~0x00000030U;
834                         config |= 0x00410000U;
835                         write_c0_config(config);
836                 }
837                 icache_size = 1 << (10 + ((config & CONF_IC) >> 9));
838                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
839                 c->icache.ways = 2;
840                 c->icache.waybit = ffs(icache_size/2) - 1;
841
842                 dcache_size = 1 << (10 + ((config & CONF_DC) >> 6));
843                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
844                 c->dcache.ways = 2;
845                 c->dcache.waybit = ffs(dcache_size/2) - 1;
846
847                 c->options |= MIPS_CPU_CACHE_CDEX_P;
848                 break;
849
850         case CPU_VR41XX:
851         case CPU_VR4111:
852         case CPU_VR4121:
853         case CPU_VR4122:
854         case CPU_VR4181:
855         case CPU_VR4181A:
856                 icache_size = 1 << (10 + ((config & CONF_IC) >> 9));
857                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
858                 c->icache.ways = 1;
859                 c->icache.waybit = 0;   /* doesn't matter */
860
861                 dcache_size = 1 << (10 + ((config & CONF_DC) >> 6));
862                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
863                 c->dcache.ways = 1;
864                 c->dcache.waybit = 0;   /* does not matter */
865
866                 c->options |= MIPS_CPU_CACHE_CDEX_P;
867                 break;
868
869         case CPU_RM7000:
870                 rm7k_erratum31();
871
872         case CPU_RM9000:
873                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
874                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
875                 c->icache.ways = 4;
876                 c->icache.waybit = ffs(icache_size / c->icache.ways) - 1;
877
878                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
879                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
880                 c->dcache.ways = 4;
881                 c->dcache.waybit = ffs(dcache_size / c->dcache.ways) - 1;
882
883 #if !defined(CONFIG_SMP) || !defined(RM9000_CDEX_SMP_WAR)
884                 c->options |= MIPS_CPU_CACHE_CDEX_P;
885 #endif
886                 c->options |= MIPS_CPU_PREFETCH;
887                 break;
888
889         default:
890                 if (!(config & MIPS_CONF_M))
891                         panic("Don't know how to probe P-caches on this cpu.");
892
893                 /*
894                  * So we seem to be a MIPS32 or MIPS64 CPU
895                  * So let's probe the I-cache ...
896                  */
897                 config1 = read_c0_config1();
898
899                 if ((lsize = ((config1 >> 19) & 7)))
900                         c->icache.linesz = 2 << lsize;
901                 else
902                         c->icache.linesz = lsize;
903                 c->icache.sets = 64 << ((config1 >> 22) & 7);
904                 c->icache.ways = 1 + ((config1 >> 16) & 7);
905
906                 icache_size = c->icache.sets *
907                               c->icache.ways *
908                               c->icache.linesz;
909                 c->icache.waybit = ffs(icache_size/c->icache.ways) - 1;
910
911                 if (config & 0x8)               /* VI bit */
912                         c->icache.flags |= MIPS_CACHE_VTAG;
913
914                 /*
915                  * Now probe the MIPS32 / MIPS64 data cache.
916                  */
917                 c->dcache.flags = 0;
918
919                 if ((lsize = ((config1 >> 10) & 7)))
920                         c->dcache.linesz = 2 << lsize;
921                 else
922                         c->dcache.linesz= lsize;
923                 c->dcache.sets = 64 << ((config1 >> 13) & 7);
924                 c->dcache.ways = 1 + ((config1 >> 7) & 7);
925
926                 dcache_size = c->dcache.sets *
927                               c->dcache.ways *
928                               c->dcache.linesz;
929                 c->dcache.waybit = ffs(dcache_size/c->dcache.ways) - 1;
930
931                 c->options |= MIPS_CPU_PREFETCH;
932                 break;
933         }
934
935         /*
936          * Processor configuration sanity check for the R4000SC erratum
937          * #5.  With page sizes larger than 32kB there is no possibility
938          * to get a VCE exception anymore so we don't care about this
939          * misconfiguration.  The case is rather theoretical anyway;
940          * presumably no vendor is shipping his hardware in the "bad"
941          * configuration.
942          */
943         if ((prid & 0xff00) == PRID_IMP_R4000 && (prid & 0xff) < 0x40 &&
944             !(config & CONF_SC) && c->icache.linesz != 16 &&
945             PAGE_SIZE <= 0x8000)
946                 panic("Improper R4000SC processor configuration detected");
947
948         /* compute a couple of other cache variables */
949         c->icache.waysize = icache_size / c->icache.ways;
950         c->dcache.waysize = dcache_size / c->dcache.ways;
951
952         c->icache.sets = icache_size / (c->icache.linesz * c->icache.ways);
953         c->dcache.sets = dcache_size / (c->dcache.linesz * c->dcache.ways);
954
955         /*
956          * R10000 and R12000 P-caches are odd in a positive way.  They're 32kB
957          * 2-way virtually indexed so normally would suffer from aliases.  So
958          * normally they'd suffer from aliases but magic in the hardware deals
959          * with that for us so we don't need to take care ourselves.
960          */
961         switch (c->cputype) {
962         case CPU_20KC:
963         case CPU_25KF:
964                 c->dcache.flags |= MIPS_CACHE_PINDEX;
965         case CPU_R10000:
966         case CPU_R12000:
967         case CPU_SB1:
968                 break;
969         case CPU_24K:
970                 if (!(read_c0_config7() & (1 << 16)))
971         default:
972                         if (c->dcache.waysize > PAGE_SIZE)
973                                 c->dcache.flags |= MIPS_CACHE_ALIASES;
974         }
975
976         switch (c->cputype) {
977         case CPU_20KC:
978                 /*
979                  * Some older 20Kc chips doesn't have the 'VI' bit in
980                  * the config register.
981                  */
982                 c->icache.flags |= MIPS_CACHE_VTAG;
983                 break;
984
985         case CPU_AU1000:
986         case CPU_AU1500:
987         case CPU_AU1100:
988         case CPU_AU1550:
989         case CPU_AU1200:
990                 c->icache.flags |= MIPS_CACHE_IC_F_DC;
991                 break;
992         }
993
994         printk("Primary instruction cache %ldkB, %s, %s, linesize %d bytes.\n",
995                icache_size >> 10,
996                cpu_has_vtag_icache ? "virtually tagged" : "physically tagged",
997                way_string[c->icache.ways], c->icache.linesz);
998
999         printk("Primary data cache %ldkB, %s, linesize %d bytes.\n",
1000                dcache_size >> 10, way_string[c->dcache.ways], c->dcache.linesz);
1001 }
1002
1003 /*
1004  * If you even _breathe_ on this function, look at the gcc output and make sure
1005  * it does not pop things on and off the stack for the cache sizing loop that
1006  * executes in KSEG1 space or else you will crash and burn badly.  You have
1007  * been warned.
1008  */
1009 static int __init probe_scache(void)
1010 {
1011         extern unsigned long stext;
1012         unsigned long flags, addr, begin, end, pow2;
1013         unsigned int config = read_c0_config();
1014         struct cpuinfo_mips *c = &current_cpu_data;
1015         int tmp;
1016
1017         if (config & CONF_SC)
1018                 return 0;
1019
1020         begin = (unsigned long) &stext;
1021         begin &= ~((4 * 1024 * 1024) - 1);
1022         end = begin + (4 * 1024 * 1024);
1023
1024         /*
1025          * This is such a bitch, you'd think they would make it easy to do
1026          * this.  Away you daemons of stupidity!
1027          */
1028         local_irq_save(flags);
1029
1030         /* Fill each size-multiple cache line with a valid tag. */
1031         pow2 = (64 * 1024);
1032         for (addr = begin; addr < end; addr = (begin + pow2)) {
1033                 unsigned long *p = (unsigned long *) addr;
1034                 __asm__ __volatile__("nop" : : "r" (*p)); /* whee... */
1035                 pow2 <<= 1;
1036         }
1037
1038         /* Load first line with zero (therefore invalid) tag. */
1039         write_c0_taglo(0);
1040         write_c0_taghi(0);
1041         __asm__ __volatile__("nop; nop; nop; nop;"); /* avoid the hazard */
1042         cache_op(Index_Store_Tag_I, begin);
1043         cache_op(Index_Store_Tag_D, begin);
1044         cache_op(Index_Store_Tag_SD, begin);
1045
1046         /* Now search for the wrap around point. */
1047         pow2 = (128 * 1024);
1048         tmp = 0;
1049         for (addr = begin + (128 * 1024); addr < end; addr = begin + pow2) {
1050                 cache_op(Index_Load_Tag_SD, addr);
1051                 __asm__ __volatile__("nop; nop; nop; nop;"); /* hazard... */
1052                 if (!read_c0_taglo())
1053                         break;
1054                 pow2 <<= 1;
1055         }
1056         local_irq_restore(flags);
1057         addr -= begin;
1058
1059         scache_size = addr;
1060         c->scache.linesz = 16 << ((config & R4K_CONF_SB) >> 22);
1061         c->scache.ways = 1;
1062         c->dcache.waybit = 0;           /* does not matter */
1063
1064         return 1;
1065 }
1066
1067 extern int r5k_sc_init(void);
1068 extern int rm7k_sc_init(void);
1069
1070 static void __init setup_scache(void)
1071 {
1072         struct cpuinfo_mips *c = &current_cpu_data;
1073         unsigned int config = read_c0_config();
1074         int sc_present = 0;
1075
1076         /*
1077          * Do the probing thing on R4000SC and R4400SC processors.  Other
1078          * processors don't have a S-cache that would be relevant to the
1079          * Linux memory managment.
1080          */
1081         switch (c->cputype) {
1082         case CPU_R4000SC:
1083         case CPU_R4000MC:
1084         case CPU_R4400SC:
1085         case CPU_R4400MC:
1086                 sc_present = run_uncached(probe_scache);
1087                 if (sc_present)
1088                         c->options |= MIPS_CPU_CACHE_CDEX_S;
1089                 break;
1090
1091         case CPU_R10000:
1092         case CPU_R12000:
1093                 scache_size = 0x80000 << ((config & R10K_CONF_SS) >> 16);
1094                 c->scache.linesz = 64 << ((config >> 13) & 1);
1095                 c->scache.ways = 2;
1096                 c->scache.waybit= 0;
1097                 sc_present = 1;
1098                 break;
1099
1100         case CPU_R5000:
1101         case CPU_NEVADA:
1102 #ifdef CONFIG_R5000_CPU_SCACHE
1103                 r5k_sc_init();
1104 #endif
1105                 return;
1106
1107         case CPU_RM7000:
1108         case CPU_RM9000:
1109 #ifdef CONFIG_RM7000_CPU_SCACHE
1110                 rm7k_sc_init();
1111 #endif
1112                 return;
1113
1114         default:
1115                 sc_present = 0;
1116         }
1117
1118         if (!sc_present)
1119                 return;
1120
1121         if ((c->isa_level == MIPS_CPU_ISA_M32R1 ||
1122              c->isa_level == MIPS_CPU_ISA_M64R1) &&
1123             !(c->scache.flags & MIPS_CACHE_NOT_PRESENT))
1124                 panic("Dunno how to handle MIPS32 / MIPS64 second level cache");
1125
1126         /* compute a couple of other cache variables */
1127         c->scache.waysize = scache_size / c->scache.ways;
1128
1129         c->scache.sets = scache_size / (c->scache.linesz * c->scache.ways);
1130
1131         printk("Unified secondary cache %ldkB %s, linesize %d bytes.\n",
1132                scache_size >> 10, way_string[c->scache.ways], c->scache.linesz);
1133
1134         c->options |= MIPS_CPU_SUBSET_CACHES;
1135 }
1136
1137 static inline void coherency_setup(void)
1138 {
1139         change_c0_config(CONF_CM_CMASK, CONF_CM_DEFAULT);
1140
1141         /*
1142          * c0_status.cu=0 specifies that updates by the sc instruction use
1143          * the coherency mode specified by the TLB; 1 means cachable
1144          * coherent update on write will be used.  Not all processors have
1145          * this bit and; some wire it to zero, others like Toshiba had the
1146          * silly idea of putting something else there ...
1147          */
1148         switch (current_cpu_data.cputype) {
1149         case CPU_R4000PC:
1150         case CPU_R4000SC:
1151         case CPU_R4000MC:
1152         case CPU_R4400PC:
1153         case CPU_R4400SC:
1154         case CPU_R4400MC:
1155                 clear_c0_config(CONF_CU);
1156                 break;
1157         }
1158 }
1159
1160 void __init r4k_cache_init(void)
1161 {
1162         extern void build_clear_page(void);
1163         extern void build_copy_page(void);
1164         extern char except_vec2_generic;
1165         struct cpuinfo_mips *c = &current_cpu_data;
1166
1167         /* Default cache error handler for R4000 and R5000 family */
1168         set_uncached_handler (0x100, &except_vec2_generic, 0x80);
1169
1170         probe_pcache();
1171         setup_scache();
1172
1173         r4k_blast_dcache_page_setup();
1174         r4k_blast_dcache_page_indexed_setup();
1175         r4k_blast_dcache_setup();
1176         r4k_blast_icache_page_setup();
1177         r4k_blast_icache_page_indexed_setup();
1178         r4k_blast_icache_setup();
1179         r4k_blast_scache_page_setup();
1180         r4k_blast_scache_page_indexed_setup();
1181         r4k_blast_scache_setup();
1182
1183         /*
1184          * Some MIPS32 and MIPS64 processors have physically indexed caches.
1185          * This code supports virtually indexed processors and will be
1186          * unnecessarily inefficient on physically indexed processors.
1187          */
1188         shm_align_mask = max_t( unsigned long,
1189                                 c->dcache.sets * c->dcache.linesz - 1,
1190                                 PAGE_SIZE - 1);
1191
1192         flush_cache_all         = r4k_flush_cache_all;
1193         __flush_cache_all       = r4k___flush_cache_all;
1194         flush_cache_mm          = r4k_flush_cache_mm;
1195         flush_cache_page        = r4k_flush_cache_page;
1196         flush_icache_page       = r4k_flush_icache_page;
1197         flush_cache_range       = r4k_flush_cache_range;
1198
1199         flush_cache_sigtramp    = r4k_flush_cache_sigtramp;
1200         flush_icache_all        = r4k_flush_icache_all;
1201         flush_data_cache_page   = r4k_flush_data_cache_page;
1202         flush_icache_range      = r4k_flush_icache_range;
1203
1204 #ifdef CONFIG_DMA_NONCOHERENT
1205         _dma_cache_wback_inv    = r4k_dma_cache_wback_inv;
1206         _dma_cache_wback        = r4k_dma_cache_wback_inv;
1207         _dma_cache_inv          = r4k_dma_cache_inv;
1208 #endif
1209
1210         build_clear_page();
1211         build_copy_page();
1212         local_r4k___flush_cache_all(NULL);
1213         coherency_setup();
1214 }