Merge master.kernel.org:/home/rmk/linux-2.6-serial
[linux-2.6] / 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 start;
468         unsigned long 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 start, unsigned long end)
532 {
533         struct flush_icache_range_args args;
534
535         args.start = start;
536         args.end = end;
537
538         on_each_cpu(local_r4k_flush_icache_range, &args, 1, 1);
539         instruction_hazard();
540 }
541
542 /*
543  * Ok, this seriously sucks.  We use them to flush a user page but don't
544  * know the virtual address, so we have to blast away the whole icache
545  * which is significantly more expensive than the real thing.  Otoh we at
546  * least know the kernel address of the page so we can flush it
547  * selectivly.
548  */
549
550 struct flush_icache_page_args {
551         struct vm_area_struct *vma;
552         struct page *page;
553 };
554
555 static inline void local_r4k_flush_icache_page(void *args)
556 {
557         struct flush_icache_page_args *fip_args = args;
558         struct vm_area_struct *vma = fip_args->vma;
559         struct page *page = fip_args->page;
560
561         /*
562          * Tricky ...  Because we don't know the virtual address we've got the
563          * choice of either invalidating the entire primary and secondary
564          * caches or invalidating the secondary caches also.  With the subset
565          * enforcment on R4000SC, R4400SC, R10000 and R12000 invalidating the
566          * secondary cache will result in any entries in the primary caches
567          * also getting invalidated which hopefully is a bit more economical.
568          */
569         if (cpu_has_subset_pcaches) {
570                 unsigned long addr = (unsigned long) page_address(page);
571
572                 r4k_blast_scache_page(addr);
573                 ClearPageDcacheDirty(page);
574
575                 return;
576         }
577
578         if (!cpu_has_ic_fills_f_dc) {
579                 unsigned long addr = (unsigned long) page_address(page);
580                 r4k_blast_dcache_page(addr);
581                 if (!cpu_icache_snoops_remote_store)
582                         r4k_blast_scache_page(addr);
583                 ClearPageDcacheDirty(page);
584         }
585
586         /*
587          * We're not sure of the virtual address(es) involved here, so
588          * we have to flush the entire I-cache.
589          */
590         if (cpu_has_vtag_icache) {
591                 int cpu = smp_processor_id();
592
593                 if (cpu_context(cpu, vma->vm_mm) != 0)
594                         drop_mmu_context(vma->vm_mm, cpu);
595         } else
596                 r4k_blast_icache();
597 }
598
599 static void r4k_flush_icache_page(struct vm_area_struct *vma,
600         struct page *page)
601 {
602         struct flush_icache_page_args args;
603
604         /*
605          * If there's no context yet, or the page isn't executable, no I-cache
606          * flush is needed.
607          */
608         if (!(vma->vm_flags & VM_EXEC))
609                 return;
610
611         args.vma = vma;
612         args.page = page;
613
614         on_each_cpu(local_r4k_flush_icache_page, &args, 1, 1);
615 }
616
617
618 #ifdef CONFIG_DMA_NONCOHERENT
619
620 static void r4k_dma_cache_wback_inv(unsigned long addr, unsigned long size)
621 {
622         unsigned long end, a;
623
624         /* Catch bad driver code */
625         BUG_ON(size == 0);
626
627         if (cpu_has_subset_pcaches) {
628                 unsigned long sc_lsize = cpu_scache_line_size();
629
630                 if (size >= scache_size) {
631                         r4k_blast_scache();
632                         return;
633                 }
634
635                 a = addr & ~(sc_lsize - 1);
636                 end = (addr + size - 1) & ~(sc_lsize - 1);
637                 while (1) {
638                         flush_scache_line(a);   /* Hit_Writeback_Inv_SD */
639                         if (a == end)
640                                 break;
641                         a += sc_lsize;
642                 }
643                 return;
644         }
645
646         /*
647          * Either no secondary cache or the available caches don't have the
648          * subset property so we have to flush the primary caches
649          * explicitly
650          */
651         if (size >= dcache_size) {
652                 r4k_blast_dcache();
653         } else {
654                 unsigned long dc_lsize = cpu_dcache_line_size();
655
656                 R4600_HIT_CACHEOP_WAR_IMPL;
657                 a = addr & ~(dc_lsize - 1);
658                 end = (addr + size - 1) & ~(dc_lsize - 1);
659                 while (1) {
660                         flush_dcache_line(a);   /* Hit_Writeback_Inv_D */
661                         if (a == end)
662                                 break;
663                         a += dc_lsize;
664                 }
665         }
666
667         bc_wback_inv(addr, size);
668 }
669
670 static void r4k_dma_cache_inv(unsigned long addr, unsigned long size)
671 {
672         unsigned long end, a;
673
674         /* Catch bad driver code */
675         BUG_ON(size == 0);
676
677         if (cpu_has_subset_pcaches) {
678                 unsigned long sc_lsize = cpu_scache_line_size();
679
680                 if (size >= scache_size) {
681                         r4k_blast_scache();
682                         return;
683                 }
684
685                 a = addr & ~(sc_lsize - 1);
686                 end = (addr + size - 1) & ~(sc_lsize - 1);
687                 while (1) {
688                         flush_scache_line(a);   /* Hit_Writeback_Inv_SD */
689                         if (a == end)
690                                 break;
691                         a += sc_lsize;
692                 }
693                 return;
694         }
695
696         if (size >= dcache_size) {
697                 r4k_blast_dcache();
698         } else {
699                 unsigned long dc_lsize = cpu_dcache_line_size();
700
701                 R4600_HIT_CACHEOP_WAR_IMPL;
702                 a = addr & ~(dc_lsize - 1);
703                 end = (addr + size - 1) & ~(dc_lsize - 1);
704                 while (1) {
705                         flush_dcache_line(a);   /* Hit_Writeback_Inv_D */
706                         if (a == end)
707                                 break;
708                         a += dc_lsize;
709                 }
710         }
711
712         bc_inv(addr, size);
713 }
714 #endif /* CONFIG_DMA_NONCOHERENT */
715
716 /*
717  * While we're protected against bad userland addresses we don't care
718  * very much about what happens in that case.  Usually a segmentation
719  * fault will dump the process later on anyway ...
720  */
721 static void local_r4k_flush_cache_sigtramp(void * arg)
722 {
723         unsigned long ic_lsize = cpu_icache_line_size();
724         unsigned long dc_lsize = cpu_dcache_line_size();
725         unsigned long sc_lsize = cpu_scache_line_size();
726         unsigned long addr = (unsigned long) arg;
727
728         R4600_HIT_CACHEOP_WAR_IMPL;
729         protected_writeback_dcache_line(addr & ~(dc_lsize - 1));
730         if (!cpu_icache_snoops_remote_store)
731                 protected_writeback_scache_line(addr & ~(sc_lsize - 1));
732         protected_flush_icache_line(addr & ~(ic_lsize - 1));
733         if (MIPS4K_ICACHE_REFILL_WAR) {
734                 __asm__ __volatile__ (
735                         ".set push\n\t"
736                         ".set noat\n\t"
737                         ".set mips3\n\t"
738 #ifdef CONFIG_32BIT
739                         "la     $at,1f\n\t"
740 #endif
741 #ifdef CONFIG_64BIT
742                         "dla    $at,1f\n\t"
743 #endif
744                         "cache  %0,($at)\n\t"
745                         "nop; nop; nop\n"
746                         "1:\n\t"
747                         ".set pop"
748                         :
749                         : "i" (Hit_Invalidate_I));
750         }
751         if (MIPS_CACHE_SYNC_WAR)
752                 __asm__ __volatile__ ("sync");
753 }
754
755 static void r4k_flush_cache_sigtramp(unsigned long addr)
756 {
757         on_each_cpu(local_r4k_flush_cache_sigtramp, (void *) addr, 1, 1);
758 }
759
760 static void r4k_flush_icache_all(void)
761 {
762         if (cpu_has_vtag_icache)
763                 r4k_blast_icache();
764 }
765
766 static inline void rm7k_erratum31(void)
767 {
768         const unsigned long ic_lsize = 32;
769         unsigned long addr;
770
771         /* RM7000 erratum #31. The icache is screwed at startup. */
772         write_c0_taglo(0);
773         write_c0_taghi(0);
774
775         for (addr = INDEX_BASE; addr <= INDEX_BASE + 4096; addr += ic_lsize) {
776                 __asm__ __volatile__ (
777                         ".set push\n\t"
778                         ".set noreorder\n\t"
779                         ".set mips3\n\t"
780                         "cache\t%1, 0(%0)\n\t"
781                         "cache\t%1, 0x1000(%0)\n\t"
782                         "cache\t%1, 0x2000(%0)\n\t"
783                         "cache\t%1, 0x3000(%0)\n\t"
784                         "cache\t%2, 0(%0)\n\t"
785                         "cache\t%2, 0x1000(%0)\n\t"
786                         "cache\t%2, 0x2000(%0)\n\t"
787                         "cache\t%2, 0x3000(%0)\n\t"
788                         "cache\t%1, 0(%0)\n\t"
789                         "cache\t%1, 0x1000(%0)\n\t"
790                         "cache\t%1, 0x2000(%0)\n\t"
791                         "cache\t%1, 0x3000(%0)\n\t"
792                         ".set pop\n"
793                         :
794                         : "r" (addr), "i" (Index_Store_Tag_I), "i" (Fill));
795         }
796 }
797
798 static char *way_string[] __initdata = { NULL, "direct mapped", "2-way",
799         "3-way", "4-way", "5-way", "6-way", "7-way", "8-way"
800 };
801
802 static void __init probe_pcache(void)
803 {
804         struct cpuinfo_mips *c = &current_cpu_data;
805         unsigned int config = read_c0_config();
806         unsigned int prid = read_c0_prid();
807         unsigned long config1;
808         unsigned int lsize;
809
810         switch (c->cputype) {
811         case CPU_R4600:                 /* QED style two way caches? */
812         case CPU_R4700:
813         case CPU_R5000:
814         case CPU_NEVADA:
815                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
816                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
817                 c->icache.ways = 2;
818                 c->icache.waybit = ffs(icache_size/2) - 1;
819
820                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
821                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
822                 c->dcache.ways = 2;
823                 c->dcache.waybit= ffs(dcache_size/2) - 1;
824
825                 c->options |= MIPS_CPU_CACHE_CDEX_P;
826                 break;
827
828         case CPU_R5432:
829         case CPU_R5500:
830                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
831                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
832                 c->icache.ways = 2;
833                 c->icache.waybit= 0;
834
835                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
836                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
837                 c->dcache.ways = 2;
838                 c->dcache.waybit = 0;
839
840                 c->options |= MIPS_CPU_CACHE_CDEX_P;
841                 break;
842
843         case CPU_TX49XX:
844                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
845                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
846                 c->icache.ways = 4;
847                 c->icache.waybit= 0;
848
849                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
850                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
851                 c->dcache.ways = 4;
852                 c->dcache.waybit = 0;
853
854                 c->options |= MIPS_CPU_CACHE_CDEX_P;
855                 break;
856
857         case CPU_R4000PC:
858         case CPU_R4000SC:
859         case CPU_R4000MC:
860         case CPU_R4400PC:
861         case CPU_R4400SC:
862         case CPU_R4400MC:
863         case CPU_R4300:
864                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
865                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
866                 c->icache.ways = 1;
867                 c->icache.waybit = 0;   /* doesn't matter */
868
869                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
870                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
871                 c->dcache.ways = 1;
872                 c->dcache.waybit = 0;   /* does not matter */
873
874                 c->options |= MIPS_CPU_CACHE_CDEX_P;
875                 break;
876
877         case CPU_R10000:
878         case CPU_R12000:
879                 icache_size = 1 << (12 + ((config & R10K_CONF_IC) >> 29));
880                 c->icache.linesz = 64;
881                 c->icache.ways = 2;
882                 c->icache.waybit = 0;
883
884                 dcache_size = 1 << (12 + ((config & R10K_CONF_DC) >> 26));
885                 c->dcache.linesz = 32;
886                 c->dcache.ways = 2;
887                 c->dcache.waybit = 0;
888
889                 c->options |= MIPS_CPU_PREFETCH;
890                 break;
891
892         case CPU_VR4133:
893                 write_c0_config(config & ~CONF_EB);
894         case CPU_VR4131:
895                 /* Workaround for cache instruction bug of VR4131 */
896                 if (c->processor_id == 0x0c80U || c->processor_id == 0x0c81U ||
897                     c->processor_id == 0x0c82U) {
898                         config &= ~0x00000030U;
899                         config |= 0x00410000U;
900                         write_c0_config(config);
901                 }
902                 icache_size = 1 << (10 + ((config & CONF_IC) >> 9));
903                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
904                 c->icache.ways = 2;
905                 c->icache.waybit = ffs(icache_size/2) - 1;
906
907                 dcache_size = 1 << (10 + ((config & CONF_DC) >> 6));
908                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
909                 c->dcache.ways = 2;
910                 c->dcache.waybit = ffs(dcache_size/2) - 1;
911
912                 c->options |= MIPS_CPU_CACHE_CDEX_P;
913                 break;
914
915         case CPU_VR41XX:
916         case CPU_VR4111:
917         case CPU_VR4121:
918         case CPU_VR4122:
919         case CPU_VR4181:
920         case CPU_VR4181A:
921                 icache_size = 1 << (10 + ((config & CONF_IC) >> 9));
922                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
923                 c->icache.ways = 1;
924                 c->icache.waybit = 0;   /* doesn't matter */
925
926                 dcache_size = 1 << (10 + ((config & CONF_DC) >> 6));
927                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
928                 c->dcache.ways = 1;
929                 c->dcache.waybit = 0;   /* does not matter */
930
931                 c->options |= MIPS_CPU_CACHE_CDEX_P;
932                 break;
933
934         case CPU_RM7000:
935                 rm7k_erratum31();
936
937         case CPU_RM9000:
938                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
939                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
940                 c->icache.ways = 4;
941                 c->icache.waybit = ffs(icache_size / c->icache.ways) - 1;
942
943                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
944                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
945                 c->dcache.ways = 4;
946                 c->dcache.waybit = ffs(dcache_size / c->dcache.ways) - 1;
947
948 #if !defined(CONFIG_SMP) || !defined(RM9000_CDEX_SMP_WAR)
949                 c->options |= MIPS_CPU_CACHE_CDEX_P;
950 #endif
951                 c->options |= MIPS_CPU_PREFETCH;
952                 break;
953
954         default:
955                 if (!(config & MIPS_CONF_M))
956                         panic("Don't know how to probe P-caches on this cpu.");
957
958                 /*
959                  * So we seem to be a MIPS32 or MIPS64 CPU
960                  * So let's probe the I-cache ...
961                  */
962                 config1 = read_c0_config1();
963
964                 if ((lsize = ((config1 >> 19) & 7)))
965                         c->icache.linesz = 2 << lsize;
966                 else
967                         c->icache.linesz = lsize;
968                 c->icache.sets = 64 << ((config1 >> 22) & 7);
969                 c->icache.ways = 1 + ((config1 >> 16) & 7);
970
971                 icache_size = c->icache.sets *
972                               c->icache.ways *
973                               c->icache.linesz;
974                 c->icache.waybit = ffs(icache_size/c->icache.ways) - 1;
975
976                 if (config & 0x8)               /* VI bit */
977                         c->icache.flags |= MIPS_CACHE_VTAG;
978
979                 /*
980                  * Now probe the MIPS32 / MIPS64 data cache.
981                  */
982                 c->dcache.flags = 0;
983
984                 if ((lsize = ((config1 >> 10) & 7)))
985                         c->dcache.linesz = 2 << lsize;
986                 else
987                         c->dcache.linesz= lsize;
988                 c->dcache.sets = 64 << ((config1 >> 13) & 7);
989                 c->dcache.ways = 1 + ((config1 >> 7) & 7);
990
991                 dcache_size = c->dcache.sets *
992                               c->dcache.ways *
993                               c->dcache.linesz;
994                 c->dcache.waybit = ffs(dcache_size/c->dcache.ways) - 1;
995
996                 c->options |= MIPS_CPU_PREFETCH;
997                 break;
998         }
999
1000         /*
1001          * Processor configuration sanity check for the R4000SC erratum
1002          * #5.  With page sizes larger than 32kB there is no possibility
1003          * to get a VCE exception anymore so we don't care about this
1004          * misconfiguration.  The case is rather theoretical anyway;
1005          * presumably no vendor is shipping his hardware in the "bad"
1006          * configuration.
1007          */
1008         if ((prid & 0xff00) == PRID_IMP_R4000 && (prid & 0xff) < 0x40 &&
1009             !(config & CONF_SC) && c->icache.linesz != 16 &&
1010             PAGE_SIZE <= 0x8000)
1011                 panic("Improper R4000SC processor configuration detected");
1012
1013         /* compute a couple of other cache variables */
1014         c->icache.waysize = icache_size / c->icache.ways;
1015         c->dcache.waysize = dcache_size / c->dcache.ways;
1016
1017         c->icache.sets = icache_size / (c->icache.linesz * c->icache.ways);
1018         c->dcache.sets = dcache_size / (c->dcache.linesz * c->dcache.ways);
1019
1020         /*
1021          * R10000 and R12000 P-caches are odd in a positive way.  They're 32kB
1022          * 2-way virtually indexed so normally would suffer from aliases.  So
1023          * normally they'd suffer from aliases but magic in the hardware deals
1024          * with that for us so we don't need to take care ourselves.
1025          */
1026         switch (c->cputype) {
1027         case CPU_20KC:
1028         case CPU_25KF:
1029         case CPU_R10000:
1030         case CPU_R12000:
1031         case CPU_SB1:
1032                 break;
1033         case CPU_24K:
1034                 if (!(read_c0_config7() & (1 << 16)))
1035         default:
1036                         if (c->dcache.waysize > PAGE_SIZE)
1037                                 c->dcache.flags |= MIPS_CACHE_ALIASES;
1038         }
1039
1040         switch (c->cputype) {
1041         case CPU_20KC:
1042                 /*
1043                  * Some older 20Kc chips doesn't have the 'VI' bit in
1044                  * the config register.
1045                  */
1046                 c->icache.flags |= MIPS_CACHE_VTAG;
1047                 break;
1048
1049         case CPU_AU1000:
1050         case CPU_AU1500:
1051         case CPU_AU1100:
1052         case CPU_AU1550:
1053         case CPU_AU1200:
1054                 c->icache.flags |= MIPS_CACHE_IC_F_DC;
1055                 break;
1056         }
1057
1058         printk("Primary instruction cache %ldkB, %s, %s, linesize %d bytes.\n",
1059                icache_size >> 10,
1060                cpu_has_vtag_icache ? "virtually tagged" : "physically tagged",
1061                way_string[c->icache.ways], c->icache.linesz);
1062
1063         printk("Primary data cache %ldkB, %s, linesize %d bytes.\n",
1064                dcache_size >> 10, way_string[c->dcache.ways], c->dcache.linesz);
1065 }
1066
1067 /*
1068  * If you even _breathe_ on this function, look at the gcc output and make sure
1069  * it does not pop things on and off the stack for the cache sizing loop that
1070  * executes in KSEG1 space or else you will crash and burn badly.  You have
1071  * been warned.
1072  */
1073 static int __init probe_scache(void)
1074 {
1075         extern unsigned long stext;
1076         unsigned long flags, addr, begin, end, pow2;
1077         unsigned int config = read_c0_config();
1078         struct cpuinfo_mips *c = &current_cpu_data;
1079         int tmp;
1080
1081         if (config & CONF_SC)
1082                 return 0;
1083
1084         begin = (unsigned long) &stext;
1085         begin &= ~((4 * 1024 * 1024) - 1);
1086         end = begin + (4 * 1024 * 1024);
1087
1088         /*
1089          * This is such a bitch, you'd think they would make it easy to do
1090          * this.  Away you daemons of stupidity!
1091          */
1092         local_irq_save(flags);
1093
1094         /* Fill each size-multiple cache line with a valid tag. */
1095         pow2 = (64 * 1024);
1096         for (addr = begin; addr < end; addr = (begin + pow2)) {
1097                 unsigned long *p = (unsigned long *) addr;
1098                 __asm__ __volatile__("nop" : : "r" (*p)); /* whee... */
1099                 pow2 <<= 1;
1100         }
1101
1102         /* Load first line with zero (therefore invalid) tag. */
1103         write_c0_taglo(0);
1104         write_c0_taghi(0);
1105         __asm__ __volatile__("nop; nop; nop; nop;"); /* avoid the hazard */
1106         cache_op(Index_Store_Tag_I, begin);
1107         cache_op(Index_Store_Tag_D, begin);
1108         cache_op(Index_Store_Tag_SD, begin);
1109
1110         /* Now search for the wrap around point. */
1111         pow2 = (128 * 1024);
1112         tmp = 0;
1113         for (addr = begin + (128 * 1024); addr < end; addr = begin + pow2) {
1114                 cache_op(Index_Load_Tag_SD, addr);
1115                 __asm__ __volatile__("nop; nop; nop; nop;"); /* hazard... */
1116                 if (!read_c0_taglo())
1117                         break;
1118                 pow2 <<= 1;
1119         }
1120         local_irq_restore(flags);
1121         addr -= begin;
1122
1123         scache_size = addr;
1124         c->scache.linesz = 16 << ((config & R4K_CONF_SB) >> 22);
1125         c->scache.ways = 1;
1126         c->dcache.waybit = 0;           /* does not matter */
1127
1128         return 1;
1129 }
1130
1131 extern int r5k_sc_init(void);
1132 extern int rm7k_sc_init(void);
1133
1134 static void __init setup_scache(void)
1135 {
1136         struct cpuinfo_mips *c = &current_cpu_data;
1137         unsigned int config = read_c0_config();
1138         int sc_present = 0;
1139
1140         /*
1141          * Do the probing thing on R4000SC and R4400SC processors.  Other
1142          * processors don't have a S-cache that would be relevant to the
1143          * Linux memory managment.
1144          */
1145         switch (c->cputype) {
1146         case CPU_R4000SC:
1147         case CPU_R4000MC:
1148         case CPU_R4400SC:
1149         case CPU_R4400MC:
1150                 sc_present = run_uncached(probe_scache);
1151                 if (sc_present)
1152                         c->options |= MIPS_CPU_CACHE_CDEX_S;
1153                 break;
1154
1155         case CPU_R10000:
1156         case CPU_R12000:
1157                 scache_size = 0x80000 << ((config & R10K_CONF_SS) >> 16);
1158                 c->scache.linesz = 64 << ((config >> 13) & 1);
1159                 c->scache.ways = 2;
1160                 c->scache.waybit= 0;
1161                 sc_present = 1;
1162                 break;
1163
1164         case CPU_R5000:
1165         case CPU_NEVADA:
1166 #ifdef CONFIG_R5000_CPU_SCACHE
1167                 r5k_sc_init();
1168 #endif
1169                 return;
1170
1171         case CPU_RM7000:
1172         case CPU_RM9000:
1173 #ifdef CONFIG_RM7000_CPU_SCACHE
1174                 rm7k_sc_init();
1175 #endif
1176                 return;
1177
1178         default:
1179                 sc_present = 0;
1180         }
1181
1182         if (!sc_present)
1183                 return;
1184
1185         if ((c->isa_level == MIPS_CPU_ISA_M32R1 ||
1186              c->isa_level == MIPS_CPU_ISA_M64R1) &&
1187             !(c->scache.flags & MIPS_CACHE_NOT_PRESENT))
1188                 panic("Dunno how to handle MIPS32 / MIPS64 second level cache");
1189
1190         /* compute a couple of other cache variables */
1191         c->scache.waysize = scache_size / c->scache.ways;
1192
1193         c->scache.sets = scache_size / (c->scache.linesz * c->scache.ways);
1194
1195         printk("Unified secondary cache %ldkB %s, linesize %d bytes.\n",
1196                scache_size >> 10, way_string[c->scache.ways], c->scache.linesz);
1197
1198         c->options |= MIPS_CPU_SUBSET_CACHES;
1199 }
1200
1201 static inline void coherency_setup(void)
1202 {
1203         change_c0_config(CONF_CM_CMASK, CONF_CM_DEFAULT);
1204
1205         /*
1206          * c0_status.cu=0 specifies that updates by the sc instruction use
1207          * the coherency mode specified by the TLB; 1 means cachable
1208          * coherent update on write will be used.  Not all processors have
1209          * this bit and; some wire it to zero, others like Toshiba had the
1210          * silly idea of putting something else there ...
1211          */
1212         switch (current_cpu_data.cputype) {
1213         case CPU_R4000PC:
1214         case CPU_R4000SC:
1215         case CPU_R4000MC:
1216         case CPU_R4400PC:
1217         case CPU_R4400SC:
1218         case CPU_R4400MC:
1219                 clear_c0_config(CONF_CU);
1220                 break;
1221         }
1222 }
1223
1224 void __init r4k_cache_init(void)
1225 {
1226         extern void build_clear_page(void);
1227         extern void build_copy_page(void);
1228         extern char except_vec2_generic;
1229         struct cpuinfo_mips *c = &current_cpu_data;
1230
1231         /* Default cache error handler for R4000 and R5000 family */
1232         set_uncached_handler (0x100, &except_vec2_generic, 0x80);
1233
1234         probe_pcache();
1235         setup_scache();
1236
1237         r4k_blast_dcache_page_setup();
1238         r4k_blast_dcache_page_indexed_setup();
1239         r4k_blast_dcache_setup();
1240         r4k_blast_icache_page_setup();
1241         r4k_blast_icache_page_indexed_setup();
1242         r4k_blast_icache_setup();
1243         r4k_blast_scache_page_setup();
1244         r4k_blast_scache_page_indexed_setup();
1245         r4k_blast_scache_setup();
1246
1247         /*
1248          * Some MIPS32 and MIPS64 processors have physically indexed caches.
1249          * This code supports virtually indexed processors and will be
1250          * unnecessarily inefficient on physically indexed processors.
1251          */
1252         shm_align_mask = max_t( unsigned long,
1253                                 c->dcache.sets * c->dcache.linesz - 1,
1254                                 PAGE_SIZE - 1);
1255
1256         flush_cache_all         = r4k_flush_cache_all;
1257         __flush_cache_all       = r4k___flush_cache_all;
1258         flush_cache_mm          = r4k_flush_cache_mm;
1259         flush_cache_page        = r4k_flush_cache_page;
1260         flush_icache_page       = r4k_flush_icache_page;
1261         flush_cache_range       = r4k_flush_cache_range;
1262
1263         flush_cache_sigtramp    = r4k_flush_cache_sigtramp;
1264         flush_icache_all        = r4k_flush_icache_all;
1265         flush_data_cache_page   = r4k_flush_data_cache_page;
1266         flush_icache_range      = r4k_flush_icache_range;
1267
1268 #ifdef CONFIG_DMA_NONCOHERENT
1269         _dma_cache_wback_inv    = r4k_dma_cache_wback_inv;
1270         _dma_cache_wback        = r4k_dma_cache_wback_inv;
1271         _dma_cache_inv          = r4k_dma_cache_inv;
1272 #endif
1273
1274         build_clear_page();
1275         build_copy_page();
1276         local_r4k___flush_cache_all(NULL);
1277         coherency_setup();
1278 }