Pull altix into release branch
[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/init.h>
11 #include <linux/kernel.h>
12 #include <linux/sched.h>
13 #include <linux/mm.h>
14 #include <linux/bitops.h>
15
16 #include <asm/bcache.h>
17 #include <asm/bootinfo.h>
18 #include <asm/cache.h>
19 #include <asm/cacheops.h>
20 #include <asm/cpu.h>
21 #include <asm/cpu-features.h>
22 #include <asm/io.h>
23 #include <asm/page.h>
24 #include <asm/pgtable.h>
25 #include <asm/r4kcache.h>
26 #include <asm/system.h>
27 #include <asm/mmu_context.h>
28 #include <asm/war.h>
29 #include <asm/cacheflush.h> /* for run_uncached() */
30
31
32 /*
33  * Special Variant of smp_call_function for use by cache functions:
34  *
35  *  o No return value
36  *  o collapses to normal function call on UP kernels
37  *  o collapses to normal function call on systems with a single shared
38  *    primary cache.
39  */
40 static inline void r4k_on_each_cpu(void (*func) (void *info), void *info,
41                                    int retry, int wait)
42 {
43         preempt_disable();
44
45 #if !defined(CONFIG_MIPS_MT_SMP) && !defined(CONFIG_MIPS_MT_SMTC)
46         smp_call_function(func, info, retry, wait);
47 #endif
48         func(info);
49         preempt_enable();
50 }
51
52 /*
53  * Must die.
54  */
55 static unsigned long icache_size __read_mostly;
56 static unsigned long dcache_size __read_mostly;
57 static unsigned long scache_size __read_mostly;
58
59 /*
60  * Dummy cache handling routines for machines without boardcaches
61  */
62 static void cache_noop(void) {}
63
64 static struct bcache_ops no_sc_ops = {
65         .bc_enable = (void *)cache_noop,
66         .bc_disable = (void *)cache_noop,
67         .bc_wback_inv = (void *)cache_noop,
68         .bc_inv = (void *)cache_noop
69 };
70
71 struct bcache_ops *bcops = &no_sc_ops;
72
73 #define cpu_is_r4600_v1_x()     ((read_c0_prid() & 0xfffffff0) == 0x00002010)
74 #define cpu_is_r4600_v2_x()     ((read_c0_prid() & 0xfffffff0) == 0x00002020)
75
76 #define R4600_HIT_CACHEOP_WAR_IMPL                                      \
77 do {                                                                    \
78         if (R4600_V2_HIT_CACHEOP_WAR && cpu_is_r4600_v2_x())            \
79                 *(volatile unsigned long *)CKSEG1;                      \
80         if (R4600_V1_HIT_CACHEOP_WAR)                                   \
81                 __asm__ __volatile__("nop;nop;nop;nop");                \
82 } while (0)
83
84 static void (*r4k_blast_dcache_page)(unsigned long addr);
85
86 static inline void r4k_blast_dcache_page_dc32(unsigned long addr)
87 {
88         R4600_HIT_CACHEOP_WAR_IMPL;
89         blast_dcache32_page(addr);
90 }
91
92 static void __init r4k_blast_dcache_page_setup(void)
93 {
94         unsigned long  dc_lsize = cpu_dcache_line_size();
95
96         if (dc_lsize == 0)
97                 r4k_blast_dcache_page = (void *)cache_noop;
98         else if (dc_lsize == 16)
99                 r4k_blast_dcache_page = blast_dcache16_page;
100         else if (dc_lsize == 32)
101                 r4k_blast_dcache_page = r4k_blast_dcache_page_dc32;
102 }
103
104 static void (* r4k_blast_dcache_page_indexed)(unsigned long addr);
105
106 static void __init r4k_blast_dcache_page_indexed_setup(void)
107 {
108         unsigned long dc_lsize = cpu_dcache_line_size();
109
110         if (dc_lsize == 0)
111                 r4k_blast_dcache_page_indexed = (void *)cache_noop;
112         else if (dc_lsize == 16)
113                 r4k_blast_dcache_page_indexed = blast_dcache16_page_indexed;
114         else if (dc_lsize == 32)
115                 r4k_blast_dcache_page_indexed = blast_dcache32_page_indexed;
116 }
117
118 static void (* r4k_blast_dcache)(void);
119
120 static void __init r4k_blast_dcache_setup(void)
121 {
122         unsigned long dc_lsize = cpu_dcache_line_size();
123
124         if (dc_lsize == 0)
125                 r4k_blast_dcache = (void *)cache_noop;
126         else if (dc_lsize == 16)
127                 r4k_blast_dcache = blast_dcache16;
128         else if (dc_lsize == 32)
129                 r4k_blast_dcache = blast_dcache32;
130 }
131
132 /* force code alignment (used for TX49XX_ICACHE_INDEX_INV_WAR) */
133 #define JUMP_TO_ALIGN(order) \
134         __asm__ __volatile__( \
135                 "b\t1f\n\t" \
136                 ".align\t" #order "\n\t" \
137                 "1:\n\t" \
138                 )
139 #define CACHE32_UNROLL32_ALIGN  JUMP_TO_ALIGN(10) /* 32 * 32 = 1024 */
140 #define CACHE32_UNROLL32_ALIGN2 JUMP_TO_ALIGN(11)
141
142 static inline void blast_r4600_v1_icache32(void)
143 {
144         unsigned long flags;
145
146         local_irq_save(flags);
147         blast_icache32();
148         local_irq_restore(flags);
149 }
150
151 static inline void tx49_blast_icache32(void)
152 {
153         unsigned long start = INDEX_BASE;
154         unsigned long end = start + current_cpu_data.icache.waysize;
155         unsigned long ws_inc = 1UL << current_cpu_data.icache.waybit;
156         unsigned long ws_end = current_cpu_data.icache.ways <<
157                                current_cpu_data.icache.waybit;
158         unsigned long ws, addr;
159
160         CACHE32_UNROLL32_ALIGN2;
161         /* I'm in even chunk.  blast odd chunks */
162         for (ws = 0; ws < ws_end; ws += ws_inc)
163                 for (addr = start + 0x400; addr < end; addr += 0x400 * 2)
164                         cache32_unroll32(addr|ws,Index_Invalidate_I);
165         CACHE32_UNROLL32_ALIGN;
166         /* I'm in odd chunk.  blast even chunks */
167         for (ws = 0; ws < ws_end; ws += ws_inc)
168                 for (addr = start; addr < end; addr += 0x400 * 2)
169                         cache32_unroll32(addr|ws,Index_Invalidate_I);
170 }
171
172 static inline void blast_icache32_r4600_v1_page_indexed(unsigned long page)
173 {
174         unsigned long flags;
175
176         local_irq_save(flags);
177         blast_icache32_page_indexed(page);
178         local_irq_restore(flags);
179 }
180
181 static inline void tx49_blast_icache32_page_indexed(unsigned long page)
182 {
183         unsigned long indexmask = current_cpu_data.icache.waysize - 1;
184         unsigned long start = INDEX_BASE + (page & indexmask);
185         unsigned long end = start + PAGE_SIZE;
186         unsigned long ws_inc = 1UL << current_cpu_data.icache.waybit;
187         unsigned long ws_end = current_cpu_data.icache.ways <<
188                                current_cpu_data.icache.waybit;
189         unsigned long ws, addr;
190
191         CACHE32_UNROLL32_ALIGN2;
192         /* I'm in even chunk.  blast odd chunks */
193         for (ws = 0; ws < ws_end; ws += ws_inc)
194                 for (addr = start + 0x400; addr < end; addr += 0x400 * 2)
195                         cache32_unroll32(addr|ws,Index_Invalidate_I);
196         CACHE32_UNROLL32_ALIGN;
197         /* I'm in odd chunk.  blast even chunks */
198         for (ws = 0; ws < ws_end; ws += ws_inc)
199                 for (addr = start; addr < end; addr += 0x400 * 2)
200                         cache32_unroll32(addr|ws,Index_Invalidate_I);
201 }
202
203 static void (* r4k_blast_icache_page)(unsigned long addr);
204
205 static void __init r4k_blast_icache_page_setup(void)
206 {
207         unsigned long ic_lsize = cpu_icache_line_size();
208
209         if (ic_lsize == 0)
210                 r4k_blast_icache_page = (void *)cache_noop;
211         else if (ic_lsize == 16)
212                 r4k_blast_icache_page = blast_icache16_page;
213         else if (ic_lsize == 32)
214                 r4k_blast_icache_page = blast_icache32_page;
215         else if (ic_lsize == 64)
216                 r4k_blast_icache_page = blast_icache64_page;
217 }
218
219
220 static void (* r4k_blast_icache_page_indexed)(unsigned long addr);
221
222 static void __init r4k_blast_icache_page_indexed_setup(void)
223 {
224         unsigned long ic_lsize = cpu_icache_line_size();
225
226         if (ic_lsize == 0)
227                 r4k_blast_icache_page_indexed = (void *)cache_noop;
228         else if (ic_lsize == 16)
229                 r4k_blast_icache_page_indexed = blast_icache16_page_indexed;
230         else if (ic_lsize == 32) {
231                 if (R4600_V1_INDEX_ICACHEOP_WAR && cpu_is_r4600_v1_x())
232                         r4k_blast_icache_page_indexed =
233                                 blast_icache32_r4600_v1_page_indexed;
234                 else if (TX49XX_ICACHE_INDEX_INV_WAR)
235                         r4k_blast_icache_page_indexed =
236                                 tx49_blast_icache32_page_indexed;
237                 else
238                         r4k_blast_icache_page_indexed =
239                                 blast_icache32_page_indexed;
240         } else if (ic_lsize == 64)
241                 r4k_blast_icache_page_indexed = blast_icache64_page_indexed;
242 }
243
244 static void (* r4k_blast_icache)(void);
245
246 static void __init r4k_blast_icache_setup(void)
247 {
248         unsigned long ic_lsize = cpu_icache_line_size();
249
250         if (ic_lsize == 0)
251                 r4k_blast_icache = (void *)cache_noop;
252         else if (ic_lsize == 16)
253                 r4k_blast_icache = blast_icache16;
254         else if (ic_lsize == 32) {
255                 if (R4600_V1_INDEX_ICACHEOP_WAR && cpu_is_r4600_v1_x())
256                         r4k_blast_icache = blast_r4600_v1_icache32;
257                 else if (TX49XX_ICACHE_INDEX_INV_WAR)
258                         r4k_blast_icache = tx49_blast_icache32;
259                 else
260                         r4k_blast_icache = blast_icache32;
261         } else if (ic_lsize == 64)
262                 r4k_blast_icache = blast_icache64;
263 }
264
265 static void (* r4k_blast_scache_page)(unsigned long addr);
266
267 static void __init r4k_blast_scache_page_setup(void)
268 {
269         unsigned long sc_lsize = cpu_scache_line_size();
270
271         if (scache_size == 0)
272                 r4k_blast_scache_page = (void *)cache_noop;
273         else if (sc_lsize == 16)
274                 r4k_blast_scache_page = blast_scache16_page;
275         else if (sc_lsize == 32)
276                 r4k_blast_scache_page = blast_scache32_page;
277         else if (sc_lsize == 64)
278                 r4k_blast_scache_page = blast_scache64_page;
279         else if (sc_lsize == 128)
280                 r4k_blast_scache_page = blast_scache128_page;
281 }
282
283 static void (* r4k_blast_scache_page_indexed)(unsigned long addr);
284
285 static void __init r4k_blast_scache_page_indexed_setup(void)
286 {
287         unsigned long sc_lsize = cpu_scache_line_size();
288
289         if (scache_size == 0)
290                 r4k_blast_scache_page_indexed = (void *)cache_noop;
291         else if (sc_lsize == 16)
292                 r4k_blast_scache_page_indexed = blast_scache16_page_indexed;
293         else if (sc_lsize == 32)
294                 r4k_blast_scache_page_indexed = blast_scache32_page_indexed;
295         else if (sc_lsize == 64)
296                 r4k_blast_scache_page_indexed = blast_scache64_page_indexed;
297         else if (sc_lsize == 128)
298                 r4k_blast_scache_page_indexed = blast_scache128_page_indexed;
299 }
300
301 static void (* r4k_blast_scache)(void);
302
303 static void __init r4k_blast_scache_setup(void)
304 {
305         unsigned long sc_lsize = cpu_scache_line_size();
306
307         if (scache_size == 0)
308                 r4k_blast_scache = (void *)cache_noop;
309         else if (sc_lsize == 16)
310                 r4k_blast_scache = blast_scache16;
311         else if (sc_lsize == 32)
312                 r4k_blast_scache = blast_scache32;
313         else if (sc_lsize == 64)
314                 r4k_blast_scache = blast_scache64;
315         else if (sc_lsize == 128)
316                 r4k_blast_scache = blast_scache128;
317 }
318
319 /*
320  * This is former mm's flush_cache_all() which really should be
321  * flush_cache_vunmap these days ...
322  */
323 static inline void local_r4k_flush_cache_all(void * args)
324 {
325         r4k_blast_dcache();
326 }
327
328 static void r4k_flush_cache_all(void)
329 {
330         if (!cpu_has_dc_aliases)
331                 return;
332
333         r4k_on_each_cpu(local_r4k_flush_cache_all, NULL, 1, 1);
334 }
335
336 static inline void local_r4k___flush_cache_all(void * args)
337 {
338         r4k_blast_dcache();
339         r4k_blast_icache();
340
341         switch (current_cpu_data.cputype) {
342         case CPU_R4000SC:
343         case CPU_R4000MC:
344         case CPU_R4400SC:
345         case CPU_R4400MC:
346         case CPU_R10000:
347         case CPU_R12000:
348         case CPU_R14000:
349                 r4k_blast_scache();
350         }
351 }
352
353 static void r4k___flush_cache_all(void)
354 {
355         r4k_on_each_cpu(local_r4k___flush_cache_all, NULL, 1, 1);
356 }
357
358 static inline void local_r4k_flush_cache_range(void * args)
359 {
360         struct vm_area_struct *vma = args;
361
362         if (!(cpu_context(smp_processor_id(), vma->vm_mm)))
363                 return;
364
365         r4k_blast_dcache();
366 }
367
368 static void r4k_flush_cache_range(struct vm_area_struct *vma,
369         unsigned long start, unsigned long end)
370 {
371         if (!cpu_has_dc_aliases)
372                 return;
373
374         r4k_on_each_cpu(local_r4k_flush_cache_range, vma, 1, 1);
375 }
376
377 static inline void local_r4k_flush_cache_mm(void * args)
378 {
379         struct mm_struct *mm = args;
380
381         if (!cpu_context(smp_processor_id(), mm))
382                 return;
383
384         /*
385          * Kludge alert.  For obscure reasons R4000SC and R4400SC go nuts if we
386          * only flush the primary caches but R10000 and R12000 behave sane ...
387          * R4000SC and R4400SC indexed S-cache ops also invalidate primary
388          * caches, so we can bail out early.
389          */
390         if (current_cpu_data.cputype == CPU_R4000SC ||
391             current_cpu_data.cputype == CPU_R4000MC ||
392             current_cpu_data.cputype == CPU_R4400SC ||
393             current_cpu_data.cputype == CPU_R4400MC) {
394                 r4k_blast_scache();
395                 return;
396         }
397
398         r4k_blast_dcache();
399 }
400
401 static void r4k_flush_cache_mm(struct mm_struct *mm)
402 {
403         if (!cpu_has_dc_aliases)
404                 return;
405
406         r4k_on_each_cpu(local_r4k_flush_cache_mm, mm, 1, 1);
407 }
408
409 struct flush_cache_page_args {
410         struct vm_area_struct *vma;
411         unsigned long addr;
412         unsigned long pfn;
413 };
414
415 static inline void local_r4k_flush_cache_page(void *args)
416 {
417         struct flush_cache_page_args *fcp_args = args;
418         struct vm_area_struct *vma = fcp_args->vma;
419         unsigned long addr = fcp_args->addr;
420         unsigned long paddr = fcp_args->pfn << PAGE_SHIFT;
421         int exec = vma->vm_flags & VM_EXEC;
422         struct mm_struct *mm = vma->vm_mm;
423         pgd_t *pgdp;
424         pud_t *pudp;
425         pmd_t *pmdp;
426         pte_t *ptep;
427
428         /*
429          * If ownes no valid ASID yet, cannot possibly have gotten
430          * this page into the cache.
431          */
432         if (cpu_context(smp_processor_id(), mm) == 0)
433                 return;
434
435         addr &= PAGE_MASK;
436         pgdp = pgd_offset(mm, addr);
437         pudp = pud_offset(pgdp, addr);
438         pmdp = pmd_offset(pudp, addr);
439         ptep = pte_offset(pmdp, addr);
440
441         /*
442          * If the page isn't marked valid, the page cannot possibly be
443          * in the cache.
444          */
445         if (!(pte_val(*ptep) & _PAGE_PRESENT))
446                 return;
447
448         /*
449          * Doing flushes for another ASID than the current one is
450          * too difficult since stupid R4k caches do a TLB translation
451          * for every cache flush operation.  So we do indexed flushes
452          * in that case, which doesn't overly flush the cache too much.
453          */
454         if ((mm == current->active_mm) && (pte_val(*ptep) & _PAGE_VALID)) {
455                 if (cpu_has_dc_aliases || (exec && !cpu_has_ic_fills_f_dc)) {
456                         r4k_blast_dcache_page(addr);
457                         if (exec && !cpu_icache_snoops_remote_store)
458                                 r4k_blast_scache_page(addr);
459                 }
460                 if (exec)
461                         r4k_blast_icache_page(addr);
462
463                 return;
464         }
465
466         /*
467          * Do indexed flush, too much work to get the (possible) TLB refills
468          * to work correctly.
469          */
470         if (cpu_has_dc_aliases || (exec && !cpu_has_ic_fills_f_dc)) {
471                 r4k_blast_dcache_page_indexed(cpu_has_pindexed_dcache ?
472                                               paddr : addr);
473                 if (exec && !cpu_icache_snoops_remote_store) {
474                         r4k_blast_scache_page_indexed(paddr);
475                 }
476         }
477         if (exec) {
478                 if (cpu_has_vtag_icache && mm == current->active_mm) {
479                         int cpu = smp_processor_id();
480
481                         if (cpu_context(cpu, mm) != 0)
482                                 drop_mmu_context(mm, cpu);
483                 } else
484                         r4k_blast_icache_page_indexed(addr);
485         }
486 }
487
488 static void r4k_flush_cache_page(struct vm_area_struct *vma,
489         unsigned long addr, unsigned long pfn)
490 {
491         struct flush_cache_page_args args;
492
493         args.vma = vma;
494         args.addr = addr;
495         args.pfn = pfn;
496
497         r4k_on_each_cpu(local_r4k_flush_cache_page, &args, 1, 1);
498 }
499
500 static inline void local_r4k_flush_data_cache_page(void * addr)
501 {
502         r4k_blast_dcache_page((unsigned long) addr);
503 }
504
505 static void r4k_flush_data_cache_page(unsigned long addr)
506 {
507         r4k_on_each_cpu(local_r4k_flush_data_cache_page, (void *) addr, 1, 1);
508 }
509
510 struct flush_icache_range_args {
511         unsigned long start;
512         unsigned long end;
513 };
514
515 static inline void local_r4k_flush_icache_range(void *args)
516 {
517         struct flush_icache_range_args *fir_args = args;
518         unsigned long start = fir_args->start;
519         unsigned long end = fir_args->end;
520
521         if (!cpu_has_ic_fills_f_dc) {
522                 if (end - start >= dcache_size) {
523                         r4k_blast_dcache();
524                 } else {
525                         R4600_HIT_CACHEOP_WAR_IMPL;
526                         protected_blast_dcache_range(start, end);
527                 }
528
529                 if (!cpu_icache_snoops_remote_store && scache_size) {
530                         if (end - start > scache_size)
531                                 r4k_blast_scache();
532                         else
533                                 protected_blast_scache_range(start, end);
534                 }
535         }
536
537         if (end - start > icache_size)
538                 r4k_blast_icache();
539         else
540                 protected_blast_icache_range(start, end);
541 }
542
543 static void r4k_flush_icache_range(unsigned long start, unsigned long end)
544 {
545         struct flush_icache_range_args args;
546
547         args.start = start;
548         args.end = end;
549
550         r4k_on_each_cpu(local_r4k_flush_icache_range, &args, 1, 1);
551         instruction_hazard();
552 }
553
554 #ifdef CONFIG_DMA_NONCOHERENT
555
556 static void r4k_dma_cache_wback_inv(unsigned long addr, unsigned long size)
557 {
558         /* Catch bad driver code */
559         BUG_ON(size == 0);
560
561         if (cpu_has_inclusive_pcaches) {
562                 if (size >= scache_size)
563                         r4k_blast_scache();
564                 else
565                         blast_scache_range(addr, addr + size);
566                 return;
567         }
568
569         /*
570          * Either no secondary cache or the available caches don't have the
571          * subset property so we have to flush the primary caches
572          * explicitly
573          */
574         if (size >= dcache_size) {
575                 r4k_blast_dcache();
576         } else {
577                 R4600_HIT_CACHEOP_WAR_IMPL;
578                 blast_dcache_range(addr, addr + size);
579         }
580
581         bc_wback_inv(addr, size);
582 }
583
584 static void r4k_dma_cache_inv(unsigned long addr, unsigned long size)
585 {
586         /* Catch bad driver code */
587         BUG_ON(size == 0);
588
589         if (cpu_has_inclusive_pcaches) {
590                 if (size >= scache_size)
591                         r4k_blast_scache();
592                 else
593                         blast_scache_range(addr, addr + size);
594                 return;
595         }
596
597         if (size >= dcache_size) {
598                 r4k_blast_dcache();
599         } else {
600                 R4600_HIT_CACHEOP_WAR_IMPL;
601                 blast_dcache_range(addr, addr + size);
602         }
603
604         bc_inv(addr, size);
605 }
606 #endif /* CONFIG_DMA_NONCOHERENT */
607
608 /*
609  * While we're protected against bad userland addresses we don't care
610  * very much about what happens in that case.  Usually a segmentation
611  * fault will dump the process later on anyway ...
612  */
613 static void local_r4k_flush_cache_sigtramp(void * arg)
614 {
615         unsigned long ic_lsize = cpu_icache_line_size();
616         unsigned long dc_lsize = cpu_dcache_line_size();
617         unsigned long sc_lsize = cpu_scache_line_size();
618         unsigned long addr = (unsigned long) arg;
619
620         R4600_HIT_CACHEOP_WAR_IMPL;
621         if (dc_lsize)
622                 protected_writeback_dcache_line(addr & ~(dc_lsize - 1));
623         if (!cpu_icache_snoops_remote_store && scache_size)
624                 protected_writeback_scache_line(addr & ~(sc_lsize - 1));
625         if (ic_lsize)
626                 protected_flush_icache_line(addr & ~(ic_lsize - 1));
627         if (MIPS4K_ICACHE_REFILL_WAR) {
628                 __asm__ __volatile__ (
629                         ".set push\n\t"
630                         ".set noat\n\t"
631                         ".set mips3\n\t"
632 #ifdef CONFIG_32BIT
633                         "la     $at,1f\n\t"
634 #endif
635 #ifdef CONFIG_64BIT
636                         "dla    $at,1f\n\t"
637 #endif
638                         "cache  %0,($at)\n\t"
639                         "nop; nop; nop\n"
640                         "1:\n\t"
641                         ".set pop"
642                         :
643                         : "i" (Hit_Invalidate_I));
644         }
645         if (MIPS_CACHE_SYNC_WAR)
646                 __asm__ __volatile__ ("sync");
647 }
648
649 static void r4k_flush_cache_sigtramp(unsigned long addr)
650 {
651         r4k_on_each_cpu(local_r4k_flush_cache_sigtramp, (void *) addr, 1, 1);
652 }
653
654 static void r4k_flush_icache_all(void)
655 {
656         if (cpu_has_vtag_icache)
657                 r4k_blast_icache();
658 }
659
660 static inline void rm7k_erratum31(void)
661 {
662         const unsigned long ic_lsize = 32;
663         unsigned long addr;
664
665         /* RM7000 erratum #31. The icache is screwed at startup. */
666         write_c0_taglo(0);
667         write_c0_taghi(0);
668
669         for (addr = INDEX_BASE; addr <= INDEX_BASE + 4096; addr += ic_lsize) {
670                 __asm__ __volatile__ (
671                         ".set push\n\t"
672                         ".set noreorder\n\t"
673                         ".set mips3\n\t"
674                         "cache\t%1, 0(%0)\n\t"
675                         "cache\t%1, 0x1000(%0)\n\t"
676                         "cache\t%1, 0x2000(%0)\n\t"
677                         "cache\t%1, 0x3000(%0)\n\t"
678                         "cache\t%2, 0(%0)\n\t"
679                         "cache\t%2, 0x1000(%0)\n\t"
680                         "cache\t%2, 0x2000(%0)\n\t"
681                         "cache\t%2, 0x3000(%0)\n\t"
682                         "cache\t%1, 0(%0)\n\t"
683                         "cache\t%1, 0x1000(%0)\n\t"
684                         "cache\t%1, 0x2000(%0)\n\t"
685                         "cache\t%1, 0x3000(%0)\n\t"
686                         ".set pop\n"
687                         :
688                         : "r" (addr), "i" (Index_Store_Tag_I), "i" (Fill));
689         }
690 }
691
692 static char *way_string[] __initdata = { NULL, "direct mapped", "2-way",
693         "3-way", "4-way", "5-way", "6-way", "7-way", "8-way"
694 };
695
696 static void __init probe_pcache(void)
697 {
698         struct cpuinfo_mips *c = &current_cpu_data;
699         unsigned int config = read_c0_config();
700         unsigned int prid = read_c0_prid();
701         unsigned long config1;
702         unsigned int lsize;
703
704         switch (c->cputype) {
705         case CPU_R4600:                 /* QED style two way caches? */
706         case CPU_R4700:
707         case CPU_R5000:
708         case CPU_NEVADA:
709                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
710                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
711                 c->icache.ways = 2;
712                 c->icache.waybit = __ffs(icache_size/2);
713
714                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
715                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
716                 c->dcache.ways = 2;
717                 c->dcache.waybit= __ffs(dcache_size/2);
718
719                 c->options |= MIPS_CPU_CACHE_CDEX_P;
720                 break;
721
722         case CPU_R5432:
723         case CPU_R5500:
724                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
725                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
726                 c->icache.ways = 2;
727                 c->icache.waybit= 0;
728
729                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
730                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
731                 c->dcache.ways = 2;
732                 c->dcache.waybit = 0;
733
734                 c->options |= MIPS_CPU_CACHE_CDEX_P;
735                 break;
736
737         case CPU_TX49XX:
738                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
739                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
740                 c->icache.ways = 4;
741                 c->icache.waybit= 0;
742
743                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
744                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
745                 c->dcache.ways = 4;
746                 c->dcache.waybit = 0;
747
748                 c->options |= MIPS_CPU_CACHE_CDEX_P;
749                 c->options |= MIPS_CPU_PREFETCH;
750                 break;
751
752         case CPU_R4000PC:
753         case CPU_R4000SC:
754         case CPU_R4000MC:
755         case CPU_R4400PC:
756         case CPU_R4400SC:
757         case CPU_R4400MC:
758         case CPU_R4300:
759                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
760                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
761                 c->icache.ways = 1;
762                 c->icache.waybit = 0;   /* doesn't matter */
763
764                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
765                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
766                 c->dcache.ways = 1;
767                 c->dcache.waybit = 0;   /* does not matter */
768
769                 c->options |= MIPS_CPU_CACHE_CDEX_P;
770                 break;
771
772         case CPU_R10000:
773         case CPU_R12000:
774         case CPU_R14000:
775                 icache_size = 1 << (12 + ((config & R10K_CONF_IC) >> 29));
776                 c->icache.linesz = 64;
777                 c->icache.ways = 2;
778                 c->icache.waybit = 0;
779
780                 dcache_size = 1 << (12 + ((config & R10K_CONF_DC) >> 26));
781                 c->dcache.linesz = 32;
782                 c->dcache.ways = 2;
783                 c->dcache.waybit = 0;
784
785                 c->options |= MIPS_CPU_PREFETCH;
786                 break;
787
788         case CPU_VR4133:
789                 write_c0_config(config & ~VR41_CONF_P4K);
790         case CPU_VR4131:
791                 /* Workaround for cache instruction bug of VR4131 */
792                 if (c->processor_id == 0x0c80U || c->processor_id == 0x0c81U ||
793                     c->processor_id == 0x0c82U) {
794                         config |= 0x00400000U;
795                         if (c->processor_id == 0x0c80U)
796                                 config |= VR41_CONF_BP;
797                         write_c0_config(config);
798                 } else
799                         c->options |= MIPS_CPU_CACHE_CDEX_P;
800
801                 icache_size = 1 << (10 + ((config & CONF_IC) >> 9));
802                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
803                 c->icache.ways = 2;
804                 c->icache.waybit = __ffs(icache_size/2);
805
806                 dcache_size = 1 << (10 + ((config & CONF_DC) >> 6));
807                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
808                 c->dcache.ways = 2;
809                 c->dcache.waybit = __ffs(dcache_size/2);
810                 break;
811
812         case CPU_VR41XX:
813         case CPU_VR4111:
814         case CPU_VR4121:
815         case CPU_VR4122:
816         case CPU_VR4181:
817         case CPU_VR4181A:
818                 icache_size = 1 << (10 + ((config & CONF_IC) >> 9));
819                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
820                 c->icache.ways = 1;
821                 c->icache.waybit = 0;   /* doesn't matter */
822
823                 dcache_size = 1 << (10 + ((config & CONF_DC) >> 6));
824                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
825                 c->dcache.ways = 1;
826                 c->dcache.waybit = 0;   /* does not matter */
827
828                 c->options |= MIPS_CPU_CACHE_CDEX_P;
829                 break;
830
831         case CPU_RM7000:
832                 rm7k_erratum31();
833
834         case CPU_RM9000:
835                 icache_size = 1 << (12 + ((config & CONF_IC) >> 9));
836                 c->icache.linesz = 16 << ((config & CONF_IB) >> 5);
837                 c->icache.ways = 4;
838                 c->icache.waybit = __ffs(icache_size / c->icache.ways);
839
840                 dcache_size = 1 << (12 + ((config & CONF_DC) >> 6));
841                 c->dcache.linesz = 16 << ((config & CONF_DB) >> 4);
842                 c->dcache.ways = 4;
843                 c->dcache.waybit = __ffs(dcache_size / c->dcache.ways);
844
845 #if !defined(CONFIG_SMP) || !defined(RM9000_CDEX_SMP_WAR)
846                 c->options |= MIPS_CPU_CACHE_CDEX_P;
847 #endif
848                 c->options |= MIPS_CPU_PREFETCH;
849                 break;
850
851         default:
852                 if (!(config & MIPS_CONF_M))
853                         panic("Don't know how to probe P-caches on this cpu.");
854
855                 /*
856                  * So we seem to be a MIPS32 or MIPS64 CPU
857                  * So let's probe the I-cache ...
858                  */
859                 config1 = read_c0_config1();
860
861                 if ((lsize = ((config1 >> 19) & 7)))
862                         c->icache.linesz = 2 << lsize;
863                 else
864                         c->icache.linesz = lsize;
865                 c->icache.sets = 64 << ((config1 >> 22) & 7);
866                 c->icache.ways = 1 + ((config1 >> 16) & 7);
867
868                 icache_size = c->icache.sets *
869                               c->icache.ways *
870                               c->icache.linesz;
871                 c->icache.waybit = __ffs(icache_size/c->icache.ways);
872
873                 if (config & 0x8)               /* VI bit */
874                         c->icache.flags |= MIPS_CACHE_VTAG;
875
876                 /*
877                  * Now probe the MIPS32 / MIPS64 data cache.
878                  */
879                 c->dcache.flags = 0;
880
881                 if ((lsize = ((config1 >> 10) & 7)))
882                         c->dcache.linesz = 2 << lsize;
883                 else
884                         c->dcache.linesz= lsize;
885                 c->dcache.sets = 64 << ((config1 >> 13) & 7);
886                 c->dcache.ways = 1 + ((config1 >> 7) & 7);
887
888                 dcache_size = c->dcache.sets *
889                               c->dcache.ways *
890                               c->dcache.linesz;
891                 c->dcache.waybit = __ffs(dcache_size/c->dcache.ways);
892
893                 c->options |= MIPS_CPU_PREFETCH;
894                 break;
895         }
896
897         /*
898          * Processor configuration sanity check for the R4000SC erratum
899          * #5.  With page sizes larger than 32kB there is no possibility
900          * to get a VCE exception anymore so we don't care about this
901          * misconfiguration.  The case is rather theoretical anyway;
902          * presumably no vendor is shipping his hardware in the "bad"
903          * configuration.
904          */
905         if ((prid & 0xff00) == PRID_IMP_R4000 && (prid & 0xff) < 0x40 &&
906             !(config & CONF_SC) && c->icache.linesz != 16 &&
907             PAGE_SIZE <= 0x8000)
908                 panic("Improper R4000SC processor configuration detected");
909
910         /* compute a couple of other cache variables */
911         c->icache.waysize = icache_size / c->icache.ways;
912         c->dcache.waysize = dcache_size / c->dcache.ways;
913
914         c->icache.sets = c->icache.linesz ?
915                 icache_size / (c->icache.linesz * c->icache.ways) : 0;
916         c->dcache.sets = c->dcache.linesz ?
917                 dcache_size / (c->dcache.linesz * c->dcache.ways) : 0;
918
919         /*
920          * R10000 and R12000 P-caches are odd in a positive way.  They're 32kB
921          * 2-way virtually indexed so normally would suffer from aliases.  So
922          * normally they'd suffer from aliases but magic in the hardware deals
923          * with that for us so we don't need to take care ourselves.
924          */
925         switch (c->cputype) {
926         case CPU_20KC:
927         case CPU_25KF:
928                 c->dcache.flags |= MIPS_CACHE_PINDEX;
929         case CPU_R10000:
930         case CPU_R12000:
931         case CPU_R14000:
932         case CPU_SB1:
933                 break;
934         case CPU_24K:
935         case CPU_34K:
936         case CPU_74K:
937                 if ((read_c0_config7() & (1 << 16))) {
938                         /* effectively physically indexed dcache,
939                            thus no virtual aliases. */
940                         c->dcache.flags |= MIPS_CACHE_PINDEX;
941                         break;
942                 }
943         default:
944                 if (c->dcache.waysize > PAGE_SIZE)
945                         c->dcache.flags |= MIPS_CACHE_ALIASES;
946         }
947
948         switch (c->cputype) {
949         case CPU_20KC:
950                 /*
951                  * Some older 20Kc chips doesn't have the 'VI' bit in
952                  * the config register.
953                  */
954                 c->icache.flags |= MIPS_CACHE_VTAG;
955                 break;
956
957         case CPU_AU1000:
958         case CPU_AU1500:
959         case CPU_AU1100:
960         case CPU_AU1550:
961         case CPU_AU1200:
962                 c->icache.flags |= MIPS_CACHE_IC_F_DC;
963                 break;
964         }
965
966         printk("Primary instruction cache %ldkB, %s, %s, linesize %d bytes.\n",
967                icache_size >> 10,
968                cpu_has_vtag_icache ? "virtually tagged" : "physically tagged",
969                way_string[c->icache.ways], c->icache.linesz);
970
971         printk("Primary data cache %ldkB, %s, linesize %d bytes.\n",
972                dcache_size >> 10, way_string[c->dcache.ways], c->dcache.linesz);
973 }
974
975 /*
976  * If you even _breathe_ on this function, look at the gcc output and make sure
977  * it does not pop things on and off the stack for the cache sizing loop that
978  * executes in KSEG1 space or else you will crash and burn badly.  You have
979  * been warned.
980  */
981 static int __init probe_scache(void)
982 {
983         extern unsigned long stext;
984         unsigned long flags, addr, begin, end, pow2;
985         unsigned int config = read_c0_config();
986         struct cpuinfo_mips *c = &current_cpu_data;
987         int tmp;
988
989         if (config & CONF_SC)
990                 return 0;
991
992         begin = (unsigned long) &stext;
993         begin &= ~((4 * 1024 * 1024) - 1);
994         end = begin + (4 * 1024 * 1024);
995
996         /*
997          * This is such a bitch, you'd think they would make it easy to do
998          * this.  Away you daemons of stupidity!
999          */
1000         local_irq_save(flags);
1001
1002         /* Fill each size-multiple cache line with a valid tag. */
1003         pow2 = (64 * 1024);
1004         for (addr = begin; addr < end; addr = (begin + pow2)) {
1005                 unsigned long *p = (unsigned long *) addr;
1006                 __asm__ __volatile__("nop" : : "r" (*p)); /* whee... */
1007                 pow2 <<= 1;
1008         }
1009
1010         /* Load first line with zero (therefore invalid) tag. */
1011         write_c0_taglo(0);
1012         write_c0_taghi(0);
1013         __asm__ __volatile__("nop; nop; nop; nop;"); /* avoid the hazard */
1014         cache_op(Index_Store_Tag_I, begin);
1015         cache_op(Index_Store_Tag_D, begin);
1016         cache_op(Index_Store_Tag_SD, begin);
1017
1018         /* Now search for the wrap around point. */
1019         pow2 = (128 * 1024);
1020         tmp = 0;
1021         for (addr = begin + (128 * 1024); addr < end; addr = begin + pow2) {
1022                 cache_op(Index_Load_Tag_SD, addr);
1023                 __asm__ __volatile__("nop; nop; nop; nop;"); /* hazard... */
1024                 if (!read_c0_taglo())
1025                         break;
1026                 pow2 <<= 1;
1027         }
1028         local_irq_restore(flags);
1029         addr -= begin;
1030
1031         scache_size = addr;
1032         c->scache.linesz = 16 << ((config & R4K_CONF_SB) >> 22);
1033         c->scache.ways = 1;
1034         c->dcache.waybit = 0;           /* does not matter */
1035
1036         return 1;
1037 }
1038
1039 extern int r5k_sc_init(void);
1040 extern int rm7k_sc_init(void);
1041 extern int mips_sc_init(void);
1042
1043 static void __init setup_scache(void)
1044 {
1045         struct cpuinfo_mips *c = &current_cpu_data;
1046         unsigned int config = read_c0_config();
1047         int sc_present = 0;
1048
1049         /*
1050          * Do the probing thing on R4000SC and R4400SC processors.  Other
1051          * processors don't have a S-cache that would be relevant to the
1052          * Linux memory managment.
1053          */
1054         switch (c->cputype) {
1055         case CPU_R4000SC:
1056         case CPU_R4000MC:
1057         case CPU_R4400SC:
1058         case CPU_R4400MC:
1059                 sc_present = run_uncached(probe_scache);
1060                 if (sc_present)
1061                         c->options |= MIPS_CPU_CACHE_CDEX_S;
1062                 break;
1063
1064         case CPU_R10000:
1065         case CPU_R12000:
1066         case CPU_R14000:
1067                 scache_size = 0x80000 << ((config & R10K_CONF_SS) >> 16);
1068                 c->scache.linesz = 64 << ((config >> 13) & 1);
1069                 c->scache.ways = 2;
1070                 c->scache.waybit= 0;
1071                 sc_present = 1;
1072                 break;
1073
1074         case CPU_R5000:
1075         case CPU_NEVADA:
1076 #ifdef CONFIG_R5000_CPU_SCACHE
1077                 r5k_sc_init();
1078 #endif
1079                 return;
1080
1081         case CPU_RM7000:
1082         case CPU_RM9000:
1083 #ifdef CONFIG_RM7000_CPU_SCACHE
1084                 rm7k_sc_init();
1085 #endif
1086                 return;
1087
1088         default:
1089                 if (c->isa_level == MIPS_CPU_ISA_M32R1 ||
1090                     c->isa_level == MIPS_CPU_ISA_M32R2 ||
1091                     c->isa_level == MIPS_CPU_ISA_M64R1 ||
1092                     c->isa_level == MIPS_CPU_ISA_M64R2) {
1093 #ifdef CONFIG_MIPS_CPU_SCACHE
1094                         if (mips_sc_init ()) {
1095                                 scache_size = c->scache.ways * c->scache.sets * c->scache.linesz;
1096                                 printk("MIPS secondary cache %ldkB, %s, linesize %d bytes.\n",
1097                                        scache_size >> 10,
1098                                        way_string[c->scache.ways], c->scache.linesz);
1099                         }
1100 #else
1101                         if (!(c->scache.flags & MIPS_CACHE_NOT_PRESENT))
1102                                 panic("Dunno how to handle MIPS32 / MIPS64 second level cache");
1103 #endif
1104                         return;
1105                 }
1106                 sc_present = 0;
1107         }
1108
1109         if (!sc_present)
1110                 return;
1111
1112         /* compute a couple of other cache variables */
1113         c->scache.waysize = scache_size / c->scache.ways;
1114
1115         c->scache.sets = scache_size / (c->scache.linesz * c->scache.ways);
1116
1117         printk("Unified secondary cache %ldkB %s, linesize %d bytes.\n",
1118                scache_size >> 10, way_string[c->scache.ways], c->scache.linesz);
1119
1120         c->options |= MIPS_CPU_INCLUSIVE_CACHES;
1121 }
1122
1123 void au1x00_fixup_config_od(void)
1124 {
1125         /*
1126          * c0_config.od (bit 19) was write only (and read as 0)
1127          * on the early revisions of Alchemy SOCs.  It disables the bus
1128          * transaction overlapping and needs to be set to fix various errata.
1129          */
1130         switch (read_c0_prid()) {
1131         case 0x00030100: /* Au1000 DA */
1132         case 0x00030201: /* Au1000 HA */
1133         case 0x00030202: /* Au1000 HB */
1134         case 0x01030200: /* Au1500 AB */
1135         /*
1136          * Au1100 errata actually keeps silence about this bit, so we set it
1137          * just in case for those revisions that require it to be set according
1138          * to arch/mips/au1000/common/cputable.c
1139          */
1140         case 0x02030200: /* Au1100 AB */
1141         case 0x02030201: /* Au1100 BA */
1142         case 0x02030202: /* Au1100 BC */
1143                 set_c0_config(1 << 19);
1144                 break;
1145         }
1146 }
1147
1148 static void __init coherency_setup(void)
1149 {
1150         change_c0_config(CONF_CM_CMASK, CONF_CM_DEFAULT);
1151
1152         /*
1153          * c0_status.cu=0 specifies that updates by the sc instruction use
1154          * the coherency mode specified by the TLB; 1 means cachable
1155          * coherent update on write will be used.  Not all processors have
1156          * this bit and; some wire it to zero, others like Toshiba had the
1157          * silly idea of putting something else there ...
1158          */
1159         switch (current_cpu_data.cputype) {
1160         case CPU_R4000PC:
1161         case CPU_R4000SC:
1162         case CPU_R4000MC:
1163         case CPU_R4400PC:
1164         case CPU_R4400SC:
1165         case CPU_R4400MC:
1166                 clear_c0_config(CONF_CU);
1167                 break;
1168         /*
1169          * We need to catch the early Alchemy SOCs with
1170          * the write-only co_config.od bit and set it back to one...
1171          */
1172         case CPU_AU1000: /* rev. DA, HA, HB */
1173         case CPU_AU1100: /* rev. AB, BA, BC ?? */
1174         case CPU_AU1500: /* rev. AB */
1175                 au1x00_fixup_config_od();
1176                 break;
1177         }
1178 }
1179
1180 void __init r4k_cache_init(void)
1181 {
1182         extern void build_clear_page(void);
1183         extern void build_copy_page(void);
1184         extern char except_vec2_generic;
1185         struct cpuinfo_mips *c = &current_cpu_data;
1186
1187         /* Default cache error handler for R4000 and R5000 family */
1188         set_uncached_handler (0x100, &except_vec2_generic, 0x80);
1189
1190         probe_pcache();
1191         setup_scache();
1192
1193         r4k_blast_dcache_page_setup();
1194         r4k_blast_dcache_page_indexed_setup();
1195         r4k_blast_dcache_setup();
1196         r4k_blast_icache_page_setup();
1197         r4k_blast_icache_page_indexed_setup();
1198         r4k_blast_icache_setup();
1199         r4k_blast_scache_page_setup();
1200         r4k_blast_scache_page_indexed_setup();
1201         r4k_blast_scache_setup();
1202
1203         /*
1204          * Some MIPS32 and MIPS64 processors have physically indexed caches.
1205          * This code supports virtually indexed processors and will be
1206          * unnecessarily inefficient on physically indexed processors.
1207          */
1208         if (c->dcache.linesz)
1209                 shm_align_mask = max_t( unsigned long,
1210                                         c->dcache.sets * c->dcache.linesz - 1,
1211                                         PAGE_SIZE - 1);
1212         else
1213                 shm_align_mask = PAGE_SIZE-1;
1214         flush_cache_all         = r4k_flush_cache_all;
1215         __flush_cache_all       = r4k___flush_cache_all;
1216         flush_cache_mm          = r4k_flush_cache_mm;
1217         flush_cache_page        = r4k_flush_cache_page;
1218         flush_cache_range       = r4k_flush_cache_range;
1219
1220         flush_cache_sigtramp    = r4k_flush_cache_sigtramp;
1221         flush_icache_all        = r4k_flush_icache_all;
1222         local_flush_data_cache_page     = local_r4k_flush_data_cache_page;
1223         flush_data_cache_page   = r4k_flush_data_cache_page;
1224         flush_icache_range      = r4k_flush_icache_range;
1225
1226 #ifdef CONFIG_DMA_NONCOHERENT
1227         _dma_cache_wback_inv    = r4k_dma_cache_wback_inv;
1228         _dma_cache_wback        = r4k_dma_cache_wback_inv;
1229         _dma_cache_inv          = r4k_dma_cache_inv;
1230 #endif
1231
1232         build_clear_page();
1233         build_copy_page();
1234         local_r4k___flush_cache_all(NULL);
1235         coherency_setup();
1236 }