Merge branch 'for-3.5-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj...
[firefly-linux-kernel-4.4.55.git] / arch / sparc / mm / srmmu.c
1 /*
2  * srmmu.c:  SRMMU specific routines for memory management.
3  *
4  * Copyright (C) 1995 David S. Miller  (davem@caip.rutgers.edu)
5  * Copyright (C) 1995,2002 Pete Zaitcev (zaitcev@yahoo.com)
6  * Copyright (C) 1996 Eddie C. Dost    (ecd@skynet.be)
7  * Copyright (C) 1997,1998 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
8  * Copyright (C) 1999,2000 Anton Blanchard (anton@samba.org)
9  */
10
11 #include <linux/kernel.h>
12 #include <linux/mm.h>
13 #include <linux/vmalloc.h>
14 #include <linux/pagemap.h>
15 #include <linux/init.h>
16 #include <linux/spinlock.h>
17 #include <linux/bootmem.h>
18 #include <linux/fs.h>
19 #include <linux/seq_file.h>
20 #include <linux/kdebug.h>
21 #include <linux/log2.h>
22 #include <linux/gfp.h>
23
24 #include <asm/bitext.h>
25 #include <asm/page.h>
26 #include <asm/pgalloc.h>
27 #include <asm/pgtable.h>
28 #include <asm/io.h>
29 #include <asm/vaddrs.h>
30 #include <asm/traps.h>
31 #include <asm/smp.h>
32 #include <asm/mbus.h>
33 #include <asm/cache.h>
34 #include <asm/oplib.h>
35 #include <asm/asi.h>
36 #include <asm/msi.h>
37 #include <asm/mmu_context.h>
38 #include <asm/io-unit.h>
39 #include <asm/cacheflush.h>
40 #include <asm/tlbflush.h>
41
42 /* Now the cpu specific definitions. */
43 #include <asm/viking.h>
44 #include <asm/mxcc.h>
45 #include <asm/ross.h>
46 #include <asm/tsunami.h>
47 #include <asm/swift.h>
48 #include <asm/turbosparc.h>
49 #include <asm/leon.h>
50
51 #include "srmmu.h"
52
53 enum mbus_module srmmu_modtype;
54 static unsigned int hwbug_bitmask;
55 int vac_cache_size;
56 int vac_line_size;
57
58 struct ctx_list *ctx_list_pool;
59 struct ctx_list ctx_free;
60 struct ctx_list ctx_used;
61
62 extern struct resource sparc_iomap;
63
64 extern unsigned long last_valid_pfn;
65
66 static pgd_t *srmmu_swapper_pg_dir;
67
68 const struct sparc32_cachetlb_ops *sparc32_cachetlb_ops;
69
70 #ifdef CONFIG_SMP
71 const struct sparc32_cachetlb_ops *local_ops;
72
73 #define FLUSH_BEGIN(mm)
74 #define FLUSH_END
75 #else
76 #define FLUSH_BEGIN(mm) if ((mm)->context != NO_CONTEXT) {
77 #define FLUSH_END       }
78 #endif
79
80 int flush_page_for_dma_global = 1;
81
82 char *srmmu_name;
83
84 ctxd_t *srmmu_ctx_table_phys;
85 static ctxd_t *srmmu_context_table;
86
87 int viking_mxcc_present;
88 static DEFINE_SPINLOCK(srmmu_context_spinlock);
89
90 static int is_hypersparc;
91
92 static int srmmu_cache_pagetables;
93
94 /* these will be initialized in srmmu_nocache_calcsize() */
95 static unsigned long srmmu_nocache_size;
96 static unsigned long srmmu_nocache_end;
97
98 /* 1 bit <=> 256 bytes of nocache <=> 64 PTEs */
99 #define SRMMU_NOCACHE_BITMAP_SHIFT (PAGE_SHIFT - 4)
100
101 /* The context table is a nocache user with the biggest alignment needs. */
102 #define SRMMU_NOCACHE_ALIGN_MAX (sizeof(ctxd_t)*SRMMU_MAX_CONTEXTS)
103
104 void *srmmu_nocache_pool;
105 void *srmmu_nocache_bitmap;
106 static struct bit_map srmmu_nocache_map;
107
108 static inline int srmmu_pmd_none(pmd_t pmd)
109 { return !(pmd_val(pmd) & 0xFFFFFFF); }
110
111 /* XXX should we hyper_flush_whole_icache here - Anton */
112 static inline void srmmu_ctxd_set(ctxd_t *ctxp, pgd_t *pgdp)
113 { set_pte((pte_t *)ctxp, (SRMMU_ET_PTD | (__nocache_pa((unsigned long) pgdp) >> 4))); }
114
115 void pmd_set(pmd_t *pmdp, pte_t *ptep)
116 {
117         unsigned long ptp;      /* Physical address, shifted right by 4 */
118         int i;
119
120         ptp = __nocache_pa((unsigned long) ptep) >> 4;
121         for (i = 0; i < PTRS_PER_PTE/SRMMU_REAL_PTRS_PER_PTE; i++) {
122                 set_pte((pte_t *)&pmdp->pmdv[i], SRMMU_ET_PTD | ptp);
123                 ptp += (SRMMU_REAL_PTRS_PER_PTE*sizeof(pte_t) >> 4);
124         }
125 }
126
127 void pmd_populate(struct mm_struct *mm, pmd_t *pmdp, struct page *ptep)
128 {
129         unsigned long ptp;      /* Physical address, shifted right by 4 */
130         int i;
131
132         ptp = page_to_pfn(ptep) << (PAGE_SHIFT-4);      /* watch for overflow */
133         for (i = 0; i < PTRS_PER_PTE/SRMMU_REAL_PTRS_PER_PTE; i++) {
134                 set_pte((pte_t *)&pmdp->pmdv[i], SRMMU_ET_PTD | ptp);
135                 ptp += (SRMMU_REAL_PTRS_PER_PTE*sizeof(pte_t) >> 4);
136         }
137 }
138
139 /* Find an entry in the third-level page table.. */ 
140 pte_t *pte_offset_kernel(pmd_t * dir, unsigned long address)
141 {
142         void *pte;
143
144         pte = __nocache_va((dir->pmdv[0] & SRMMU_PTD_PMASK) << 4);
145         return (pte_t *) pte +
146             ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
147 }
148
149 /*
150  * size: bytes to allocate in the nocache area.
151  * align: bytes, number to align at.
152  * Returns the virtual address of the allocated area.
153  */
154 static unsigned long __srmmu_get_nocache(int size, int align)
155 {
156         int offset;
157
158         if (size < SRMMU_NOCACHE_BITMAP_SHIFT) {
159                 printk("Size 0x%x too small for nocache request\n", size);
160                 size = SRMMU_NOCACHE_BITMAP_SHIFT;
161         }
162         if (size & (SRMMU_NOCACHE_BITMAP_SHIFT-1)) {
163                 printk("Size 0x%x unaligned int nocache request\n", size);
164                 size += SRMMU_NOCACHE_BITMAP_SHIFT-1;
165         }
166         BUG_ON(align > SRMMU_NOCACHE_ALIGN_MAX);
167
168         offset = bit_map_string_get(&srmmu_nocache_map,
169                                         size >> SRMMU_NOCACHE_BITMAP_SHIFT,
170                                         align >> SRMMU_NOCACHE_BITMAP_SHIFT);
171         if (offset == -1) {
172                 printk("srmmu: out of nocache %d: %d/%d\n",
173                     size, (int) srmmu_nocache_size,
174                     srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
175                 return 0;
176         }
177
178         return (SRMMU_NOCACHE_VADDR + (offset << SRMMU_NOCACHE_BITMAP_SHIFT));
179 }
180
181 unsigned long srmmu_get_nocache(int size, int align)
182 {
183         unsigned long tmp;
184
185         tmp = __srmmu_get_nocache(size, align);
186
187         if (tmp)
188                 memset((void *)tmp, 0, size);
189
190         return tmp;
191 }
192
193 void srmmu_free_nocache(unsigned long vaddr, int size)
194 {
195         int offset;
196
197         if (vaddr < SRMMU_NOCACHE_VADDR) {
198                 printk("Vaddr %lx is smaller than nocache base 0x%lx\n",
199                     vaddr, (unsigned long)SRMMU_NOCACHE_VADDR);
200                 BUG();
201         }
202         if (vaddr+size > srmmu_nocache_end) {
203                 printk("Vaddr %lx is bigger than nocache end 0x%lx\n",
204                     vaddr, srmmu_nocache_end);
205                 BUG();
206         }
207         if (!is_power_of_2(size)) {
208                 printk("Size 0x%x is not a power of 2\n", size);
209                 BUG();
210         }
211         if (size < SRMMU_NOCACHE_BITMAP_SHIFT) {
212                 printk("Size 0x%x is too small\n", size);
213                 BUG();
214         }
215         if (vaddr & (size-1)) {
216                 printk("Vaddr %lx is not aligned to size 0x%x\n", vaddr, size);
217                 BUG();
218         }
219
220         offset = (vaddr - SRMMU_NOCACHE_VADDR) >> SRMMU_NOCACHE_BITMAP_SHIFT;
221         size = size >> SRMMU_NOCACHE_BITMAP_SHIFT;
222
223         bit_map_clear(&srmmu_nocache_map, offset, size);
224 }
225
226 static void srmmu_early_allocate_ptable_skeleton(unsigned long start,
227                                                  unsigned long end);
228
229 extern unsigned long probe_memory(void);        /* in fault.c */
230
231 /*
232  * Reserve nocache dynamically proportionally to the amount of
233  * system RAM. -- Tomas Szepe <szepe@pinerecords.com>, June 2002
234  */
235 static void srmmu_nocache_calcsize(void)
236 {
237         unsigned long sysmemavail = probe_memory() / 1024;
238         int srmmu_nocache_npages;
239
240         srmmu_nocache_npages =
241                 sysmemavail / SRMMU_NOCACHE_ALCRATIO / 1024 * 256;
242
243  /* P3 XXX The 4x overuse: corroborated by /proc/meminfo. */
244         // if (srmmu_nocache_npages < 256) srmmu_nocache_npages = 256;
245         if (srmmu_nocache_npages < SRMMU_MIN_NOCACHE_PAGES)
246                 srmmu_nocache_npages = SRMMU_MIN_NOCACHE_PAGES;
247
248         /* anything above 1280 blows up */
249         if (srmmu_nocache_npages > SRMMU_MAX_NOCACHE_PAGES)
250                 srmmu_nocache_npages = SRMMU_MAX_NOCACHE_PAGES;
251
252         srmmu_nocache_size = srmmu_nocache_npages * PAGE_SIZE;
253         srmmu_nocache_end = SRMMU_NOCACHE_VADDR + srmmu_nocache_size;
254 }
255
256 static void __init srmmu_nocache_init(void)
257 {
258         unsigned int bitmap_bits;
259         pgd_t *pgd;
260         pmd_t *pmd;
261         pte_t *pte;
262         unsigned long paddr, vaddr;
263         unsigned long pteval;
264
265         bitmap_bits = srmmu_nocache_size >> SRMMU_NOCACHE_BITMAP_SHIFT;
266
267         srmmu_nocache_pool = __alloc_bootmem(srmmu_nocache_size,
268                 SRMMU_NOCACHE_ALIGN_MAX, 0UL);
269         memset(srmmu_nocache_pool, 0, srmmu_nocache_size);
270
271         srmmu_nocache_bitmap = __alloc_bootmem(bitmap_bits >> 3, SMP_CACHE_BYTES, 0UL);
272         bit_map_init(&srmmu_nocache_map, srmmu_nocache_bitmap, bitmap_bits);
273
274         srmmu_swapper_pg_dir = (pgd_t *)__srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
275         memset(__nocache_fix(srmmu_swapper_pg_dir), 0, SRMMU_PGD_TABLE_SIZE);
276         init_mm.pgd = srmmu_swapper_pg_dir;
277
278         srmmu_early_allocate_ptable_skeleton(SRMMU_NOCACHE_VADDR, srmmu_nocache_end);
279
280         paddr = __pa((unsigned long)srmmu_nocache_pool);
281         vaddr = SRMMU_NOCACHE_VADDR;
282
283         while (vaddr < srmmu_nocache_end) {
284                 pgd = pgd_offset_k(vaddr);
285                 pmd = pmd_offset(__nocache_fix(pgd), vaddr);
286                 pte = pte_offset_kernel(__nocache_fix(pmd), vaddr);
287
288                 pteval = ((paddr >> 4) | SRMMU_ET_PTE | SRMMU_PRIV);
289
290                 if (srmmu_cache_pagetables)
291                         pteval |= SRMMU_CACHE;
292
293                 set_pte(__nocache_fix(pte), __pte(pteval));
294
295                 vaddr += PAGE_SIZE;
296                 paddr += PAGE_SIZE;
297         }
298
299         flush_cache_all();
300         flush_tlb_all();
301 }
302
303 pgd_t *get_pgd_fast(void)
304 {
305         pgd_t *pgd = NULL;
306
307         pgd = (pgd_t *)__srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
308         if (pgd) {
309                 pgd_t *init = pgd_offset_k(0);
310                 memset(pgd, 0, USER_PTRS_PER_PGD * sizeof(pgd_t));
311                 memcpy(pgd + USER_PTRS_PER_PGD, init + USER_PTRS_PER_PGD,
312                                                 (PTRS_PER_PGD - USER_PTRS_PER_PGD) * sizeof(pgd_t));
313         }
314
315         return pgd;
316 }
317
318 /*
319  * Hardware needs alignment to 256 only, but we align to whole page size
320  * to reduce fragmentation problems due to the buddy principle.
321  * XXX Provide actual fragmentation statistics in /proc.
322  *
323  * Alignments up to the page size are the same for physical and virtual
324  * addresses of the nocache area.
325  */
326 pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
327 {
328         unsigned long pte;
329         struct page *page;
330
331         if ((pte = (unsigned long)pte_alloc_one_kernel(mm, address)) == 0)
332                 return NULL;
333         page = pfn_to_page( __nocache_pa(pte) >> PAGE_SHIFT );
334         pgtable_page_ctor(page);
335         return page;
336 }
337
338 void pte_free(struct mm_struct *mm, pgtable_t pte)
339 {
340         unsigned long p;
341
342         pgtable_page_dtor(pte);
343         p = (unsigned long)page_address(pte);   /* Cached address (for test) */
344         if (p == 0)
345                 BUG();
346         p = page_to_pfn(pte) << PAGE_SHIFT;     /* Physical address */
347         p = (unsigned long) __nocache_va(p);    /* Nocached virtual */
348         srmmu_free_nocache(p, PTE_SIZE);
349 }
350
351 /*
352  */
353 static inline void alloc_context(struct mm_struct *old_mm, struct mm_struct *mm)
354 {
355         struct ctx_list *ctxp;
356
357         ctxp = ctx_free.next;
358         if(ctxp != &ctx_free) {
359                 remove_from_ctx_list(ctxp);
360                 add_to_used_ctxlist(ctxp);
361                 mm->context = ctxp->ctx_number;
362                 ctxp->ctx_mm = mm;
363                 return;
364         }
365         ctxp = ctx_used.next;
366         if(ctxp->ctx_mm == old_mm)
367                 ctxp = ctxp->next;
368         if(ctxp == &ctx_used)
369                 panic("out of mmu contexts");
370         flush_cache_mm(ctxp->ctx_mm);
371         flush_tlb_mm(ctxp->ctx_mm);
372         remove_from_ctx_list(ctxp);
373         add_to_used_ctxlist(ctxp);
374         ctxp->ctx_mm->context = NO_CONTEXT;
375         ctxp->ctx_mm = mm;
376         mm->context = ctxp->ctx_number;
377 }
378
379 static inline void free_context(int context)
380 {
381         struct ctx_list *ctx_old;
382
383         ctx_old = ctx_list_pool + context;
384         remove_from_ctx_list(ctx_old);
385         add_to_free_ctxlist(ctx_old);
386 }
387
388
389 void switch_mm(struct mm_struct *old_mm, struct mm_struct *mm,
390                struct task_struct *tsk)
391 {
392         if(mm->context == NO_CONTEXT) {
393                 spin_lock(&srmmu_context_spinlock);
394                 alloc_context(old_mm, mm);
395                 spin_unlock(&srmmu_context_spinlock);
396                 srmmu_ctxd_set(&srmmu_context_table[mm->context], mm->pgd);
397         }
398
399         if (sparc_cpu_model == sparc_leon)
400                 leon_switch_mm();
401
402         if (is_hypersparc)
403                 hyper_flush_whole_icache();
404
405         srmmu_set_context(mm->context);
406 }
407
408 /* Low level IO area allocation on the SRMMU. */
409 static inline void srmmu_mapioaddr(unsigned long physaddr,
410     unsigned long virt_addr, int bus_type)
411 {
412         pgd_t *pgdp;
413         pmd_t *pmdp;
414         pte_t *ptep;
415         unsigned long tmp;
416
417         physaddr &= PAGE_MASK;
418         pgdp = pgd_offset_k(virt_addr);
419         pmdp = pmd_offset(pgdp, virt_addr);
420         ptep = pte_offset_kernel(pmdp, virt_addr);
421         tmp = (physaddr >> 4) | SRMMU_ET_PTE;
422
423         /*
424          * I need to test whether this is consistent over all
425          * sun4m's.  The bus_type represents the upper 4 bits of
426          * 36-bit physical address on the I/O space lines...
427          */
428         tmp |= (bus_type << 28);
429         tmp |= SRMMU_PRIV;
430         __flush_page_to_ram(virt_addr);
431         set_pte(ptep, __pte(tmp));
432 }
433
434 void srmmu_mapiorange(unsigned int bus, unsigned long xpa,
435                       unsigned long xva, unsigned int len)
436 {
437         while (len != 0) {
438                 len -= PAGE_SIZE;
439                 srmmu_mapioaddr(xpa, xva, bus);
440                 xva += PAGE_SIZE;
441                 xpa += PAGE_SIZE;
442         }
443         flush_tlb_all();
444 }
445
446 static inline void srmmu_unmapioaddr(unsigned long virt_addr)
447 {
448         pgd_t *pgdp;
449         pmd_t *pmdp;
450         pte_t *ptep;
451
452         pgdp = pgd_offset_k(virt_addr);
453         pmdp = pmd_offset(pgdp, virt_addr);
454         ptep = pte_offset_kernel(pmdp, virt_addr);
455
456         /* No need to flush uncacheable page. */
457         __pte_clear(ptep);
458 }
459
460 void srmmu_unmapiorange(unsigned long virt_addr, unsigned int len)
461 {
462         while (len != 0) {
463                 len -= PAGE_SIZE;
464                 srmmu_unmapioaddr(virt_addr);
465                 virt_addr += PAGE_SIZE;
466         }
467         flush_tlb_all();
468 }
469
470 /* tsunami.S */
471 extern void tsunami_flush_cache_all(void);
472 extern void tsunami_flush_cache_mm(struct mm_struct *mm);
473 extern void tsunami_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
474 extern void tsunami_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
475 extern void tsunami_flush_page_to_ram(unsigned long page);
476 extern void tsunami_flush_page_for_dma(unsigned long page);
477 extern void tsunami_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
478 extern void tsunami_flush_tlb_all(void);
479 extern void tsunami_flush_tlb_mm(struct mm_struct *mm);
480 extern void tsunami_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
481 extern void tsunami_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
482 extern void tsunami_setup_blockops(void);
483
484 /* swift.S */
485 extern void swift_flush_cache_all(void);
486 extern void swift_flush_cache_mm(struct mm_struct *mm);
487 extern void swift_flush_cache_range(struct vm_area_struct *vma,
488                                     unsigned long start, unsigned long end);
489 extern void swift_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
490 extern void swift_flush_page_to_ram(unsigned long page);
491 extern void swift_flush_page_for_dma(unsigned long page);
492 extern void swift_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
493 extern void swift_flush_tlb_all(void);
494 extern void swift_flush_tlb_mm(struct mm_struct *mm);
495 extern void swift_flush_tlb_range(struct vm_area_struct *vma,
496                                   unsigned long start, unsigned long end);
497 extern void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
498
499 #if 0  /* P3: deadwood to debug precise flushes on Swift. */
500 void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
501 {
502         int cctx, ctx1;
503
504         page &= PAGE_MASK;
505         if ((ctx1 = vma->vm_mm->context) != -1) {
506                 cctx = srmmu_get_context();
507 /* Is context # ever different from current context? P3 */
508                 if (cctx != ctx1) {
509                         printk("flush ctx %02x curr %02x\n", ctx1, cctx);
510                         srmmu_set_context(ctx1);
511                         swift_flush_page(page);
512                         __asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
513                                         "r" (page), "i" (ASI_M_FLUSH_PROBE));
514                         srmmu_set_context(cctx);
515                 } else {
516                          /* Rm. prot. bits from virt. c. */
517                         /* swift_flush_cache_all(); */
518                         /* swift_flush_cache_page(vma, page); */
519                         swift_flush_page(page);
520
521                         __asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
522                                 "r" (page), "i" (ASI_M_FLUSH_PROBE));
523                         /* same as above: srmmu_flush_tlb_page() */
524                 }
525         }
526 }
527 #endif
528
529 /*
530  * The following are all MBUS based SRMMU modules, and therefore could
531  * be found in a multiprocessor configuration.  On the whole, these
532  * chips seems to be much more touchy about DVMA and page tables
533  * with respect to cache coherency.
534  */
535
536 /* viking.S */
537 extern void viking_flush_cache_all(void);
538 extern void viking_flush_cache_mm(struct mm_struct *mm);
539 extern void viking_flush_cache_range(struct vm_area_struct *vma, unsigned long start,
540                                      unsigned long end);
541 extern void viking_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
542 extern void viking_flush_page_to_ram(unsigned long page);
543 extern void viking_flush_page_for_dma(unsigned long page);
544 extern void viking_flush_sig_insns(struct mm_struct *mm, unsigned long addr);
545 extern void viking_flush_page(unsigned long page);
546 extern void viking_mxcc_flush_page(unsigned long page);
547 extern void viking_flush_tlb_all(void);
548 extern void viking_flush_tlb_mm(struct mm_struct *mm);
549 extern void viking_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
550                                    unsigned long end);
551 extern void viking_flush_tlb_page(struct vm_area_struct *vma,
552                                   unsigned long page);
553 extern void sun4dsmp_flush_tlb_all(void);
554 extern void sun4dsmp_flush_tlb_mm(struct mm_struct *mm);
555 extern void sun4dsmp_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
556                                    unsigned long end);
557 extern void sun4dsmp_flush_tlb_page(struct vm_area_struct *vma,
558                                   unsigned long page);
559
560 /* hypersparc.S */
561 extern void hypersparc_flush_cache_all(void);
562 extern void hypersparc_flush_cache_mm(struct mm_struct *mm);
563 extern void hypersparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
564 extern void hypersparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
565 extern void hypersparc_flush_page_to_ram(unsigned long page);
566 extern void hypersparc_flush_page_for_dma(unsigned long page);
567 extern void hypersparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
568 extern void hypersparc_flush_tlb_all(void);
569 extern void hypersparc_flush_tlb_mm(struct mm_struct *mm);
570 extern void hypersparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
571 extern void hypersparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
572 extern void hypersparc_setup_blockops(void);
573
574 /*
575  * NOTE: All of this startup code assumes the low 16mb (approx.) of
576  *       kernel mappings are done with one single contiguous chunk of
577  *       ram.  On small ram machines (classics mainly) we only get
578  *       around 8mb mapped for us.
579  */
580
581 static void __init early_pgtable_allocfail(char *type)
582 {
583         prom_printf("inherit_prom_mappings: Cannot alloc kernel %s.\n", type);
584         prom_halt();
585 }
586
587 static void __init srmmu_early_allocate_ptable_skeleton(unsigned long start,
588                                                         unsigned long end)
589 {
590         pgd_t *pgdp;
591         pmd_t *pmdp;
592         pte_t *ptep;
593
594         while(start < end) {
595                 pgdp = pgd_offset_k(start);
596                 if (pgd_none(*(pgd_t *)__nocache_fix(pgdp))) {
597                         pmdp = (pmd_t *) __srmmu_get_nocache(
598                             SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
599                         if (pmdp == NULL)
600                                 early_pgtable_allocfail("pmd");
601                         memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
602                         pgd_set(__nocache_fix(pgdp), pmdp);
603                 }
604                 pmdp = pmd_offset(__nocache_fix(pgdp), start);
605                 if(srmmu_pmd_none(*(pmd_t *)__nocache_fix(pmdp))) {
606                         ptep = (pte_t *)__srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
607                         if (ptep == NULL)
608                                 early_pgtable_allocfail("pte");
609                         memset(__nocache_fix(ptep), 0, PTE_SIZE);
610                         pmd_set(__nocache_fix(pmdp), ptep);
611                 }
612                 if (start > (0xffffffffUL - PMD_SIZE))
613                         break;
614                 start = (start + PMD_SIZE) & PMD_MASK;
615         }
616 }
617
618 static void __init srmmu_allocate_ptable_skeleton(unsigned long start,
619                                                   unsigned long end)
620 {
621         pgd_t *pgdp;
622         pmd_t *pmdp;
623         pte_t *ptep;
624
625         while(start < end) {
626                 pgdp = pgd_offset_k(start);
627                 if (pgd_none(*pgdp)) {
628                         pmdp = (pmd_t *)__srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
629                         if (pmdp == NULL)
630                                 early_pgtable_allocfail("pmd");
631                         memset(pmdp, 0, SRMMU_PMD_TABLE_SIZE);
632                         pgd_set(pgdp, pmdp);
633                 }
634                 pmdp = pmd_offset(pgdp, start);
635                 if(srmmu_pmd_none(*pmdp)) {
636                         ptep = (pte_t *) __srmmu_get_nocache(PTE_SIZE,
637                                                              PTE_SIZE);
638                         if (ptep == NULL)
639                                 early_pgtable_allocfail("pte");
640                         memset(ptep, 0, PTE_SIZE);
641                         pmd_set(pmdp, ptep);
642                 }
643                 if (start > (0xffffffffUL - PMD_SIZE))
644                         break;
645                 start = (start + PMD_SIZE) & PMD_MASK;
646         }
647 }
648
649 /* These flush types are not available on all chips... */
650 static inline unsigned long srmmu_probe(unsigned long vaddr)
651 {
652         unsigned long retval;
653
654         if (sparc_cpu_model != sparc_leon) {
655
656                 vaddr &= PAGE_MASK;
657                 __asm__ __volatile__("lda [%1] %2, %0\n\t" :
658                                      "=r" (retval) :
659                                      "r" (vaddr | 0x400), "i" (ASI_M_FLUSH_PROBE));
660         } else {
661                 retval = leon_swprobe(vaddr, 0);
662         }
663         return retval;
664 }
665
666 /*
667  * This is much cleaner than poking around physical address space
668  * looking at the prom's page table directly which is what most
669  * other OS's do.  Yuck... this is much better.
670  */
671 static void __init srmmu_inherit_prom_mappings(unsigned long start,
672                                                unsigned long end)
673 {
674         pgd_t *pgdp;
675         pmd_t *pmdp;
676         pte_t *ptep;
677         int what = 0; /* 0 = normal-pte, 1 = pmd-level pte, 2 = pgd-level pte */
678         unsigned long prompte;
679
680         while(start <= end) {
681                 if (start == 0)
682                         break; /* probably wrap around */
683                 if(start == 0xfef00000)
684                         start = KADB_DEBUGGER_BEGVM;
685                 if(!(prompte = srmmu_probe(start))) {
686                         start += PAGE_SIZE;
687                         continue;
688                 }
689     
690                 /* A red snapper, see what it really is. */
691                 what = 0;
692     
693                 if(!(start & ~(SRMMU_REAL_PMD_MASK))) {
694                         if(srmmu_probe((start-PAGE_SIZE) + SRMMU_REAL_PMD_SIZE) == prompte)
695                                 what = 1;
696                 }
697     
698                 if(!(start & ~(SRMMU_PGDIR_MASK))) {
699                         if(srmmu_probe((start-PAGE_SIZE) + SRMMU_PGDIR_SIZE) ==
700                            prompte)
701                                 what = 2;
702                 }
703     
704                 pgdp = pgd_offset_k(start);
705                 if(what == 2) {
706                         *(pgd_t *)__nocache_fix(pgdp) = __pgd(prompte);
707                         start += SRMMU_PGDIR_SIZE;
708                         continue;
709                 }
710                 if (pgd_none(*(pgd_t *)__nocache_fix(pgdp))) {
711                         pmdp = (pmd_t *)__srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
712                         if (pmdp == NULL)
713                                 early_pgtable_allocfail("pmd");
714                         memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
715                         pgd_set(__nocache_fix(pgdp), pmdp);
716                 }
717                 pmdp = pmd_offset(__nocache_fix(pgdp), start);
718                 if(srmmu_pmd_none(*(pmd_t *)__nocache_fix(pmdp))) {
719                         ptep = (pte_t *) __srmmu_get_nocache(PTE_SIZE,
720                                                              PTE_SIZE);
721                         if (ptep == NULL)
722                                 early_pgtable_allocfail("pte");
723                         memset(__nocache_fix(ptep), 0, PTE_SIZE);
724                         pmd_set(__nocache_fix(pmdp), ptep);
725                 }
726                 if(what == 1) {
727                         /*
728                          * We bend the rule where all 16 PTPs in a pmd_t point
729                          * inside the same PTE page, and we leak a perfectly
730                          * good hardware PTE piece. Alternatives seem worse.
731                          */
732                         unsigned int x; /* Index of HW PMD in soft cluster */
733                         x = (start >> PMD_SHIFT) & 15;
734                         *(unsigned long *)__nocache_fix(&pmdp->pmdv[x]) = prompte;
735                         start += SRMMU_REAL_PMD_SIZE;
736                         continue;
737                 }
738                 ptep = pte_offset_kernel(__nocache_fix(pmdp), start);
739                 *(pte_t *)__nocache_fix(ptep) = __pte(prompte);
740                 start += PAGE_SIZE;
741         }
742 }
743
744 #define KERNEL_PTE(page_shifted) ((page_shifted)|SRMMU_CACHE|SRMMU_PRIV|SRMMU_VALID)
745
746 /* Create a third-level SRMMU 16MB page mapping. */
747 static void __init do_large_mapping(unsigned long vaddr, unsigned long phys_base)
748 {
749         pgd_t *pgdp = pgd_offset_k(vaddr);
750         unsigned long big_pte;
751
752         big_pte = KERNEL_PTE(phys_base >> 4);
753         *(pgd_t *)__nocache_fix(pgdp) = __pgd(big_pte);
754 }
755
756 /* Map sp_bank entry SP_ENTRY, starting at virtual address VBASE. */
757 static unsigned long __init map_spbank(unsigned long vbase, int sp_entry)
758 {
759         unsigned long pstart = (sp_banks[sp_entry].base_addr & SRMMU_PGDIR_MASK);
760         unsigned long vstart = (vbase & SRMMU_PGDIR_MASK);
761         unsigned long vend = SRMMU_PGDIR_ALIGN(vbase + sp_banks[sp_entry].num_bytes);
762         /* Map "low" memory only */
763         const unsigned long min_vaddr = PAGE_OFFSET;
764         const unsigned long max_vaddr = PAGE_OFFSET + SRMMU_MAXMEM;
765
766         if (vstart < min_vaddr || vstart >= max_vaddr)
767                 return vstart;
768         
769         if (vend > max_vaddr || vend < min_vaddr)
770                 vend = max_vaddr;
771
772         while(vstart < vend) {
773                 do_large_mapping(vstart, pstart);
774                 vstart += SRMMU_PGDIR_SIZE; pstart += SRMMU_PGDIR_SIZE;
775         }
776         return vstart;
777 }
778
779 static inline void map_kernel(void)
780 {
781         int i;
782
783         if (phys_base > 0) {
784                 do_large_mapping(PAGE_OFFSET, phys_base);
785         }
786
787         for (i = 0; sp_banks[i].num_bytes != 0; i++) {
788                 map_spbank((unsigned long)__va(sp_banks[i].base_addr), i);
789         }
790 }
791
792 /* Paging initialization on the Sparc Reference MMU. */
793 extern void sparc_context_init(int);
794
795 void (*poke_srmmu)(void) __cpuinitdata = NULL;
796
797 extern unsigned long bootmem_init(unsigned long *pages_avail);
798
799 void __init srmmu_paging_init(void)
800 {
801         int i;
802         phandle cpunode;
803         char node_str[128];
804         pgd_t *pgd;
805         pmd_t *pmd;
806         pte_t *pte;
807         unsigned long pages_avail;
808
809         sparc_iomap.start = SUN4M_IOBASE_VADDR; /* 16MB of IOSPACE on all sun4m's. */
810
811         if (sparc_cpu_model == sun4d)
812                 num_contexts = 65536; /* We know it is Viking */
813         else {
814                 /* Find the number of contexts on the srmmu. */
815                 cpunode = prom_getchild(prom_root_node);
816                 num_contexts = 0;
817                 while(cpunode != 0) {
818                         prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
819                         if(!strcmp(node_str, "cpu")) {
820                                 num_contexts = prom_getintdefault(cpunode, "mmu-nctx", 0x8);
821                                 break;
822                         }
823                         cpunode = prom_getsibling(cpunode);
824                 }
825         }
826
827         if(!num_contexts) {
828                 prom_printf("Something wrong, can't find cpu node in paging_init.\n");
829                 prom_halt();
830         }
831
832         pages_avail = 0;
833         last_valid_pfn = bootmem_init(&pages_avail);
834
835         srmmu_nocache_calcsize();
836         srmmu_nocache_init();
837         srmmu_inherit_prom_mappings(0xfe400000,(LINUX_OPPROM_ENDVM-PAGE_SIZE));
838         map_kernel();
839
840         /* ctx table has to be physically aligned to its size */
841         srmmu_context_table = (ctxd_t *)__srmmu_get_nocache(num_contexts*sizeof(ctxd_t), num_contexts*sizeof(ctxd_t));
842         srmmu_ctx_table_phys = (ctxd_t *)__nocache_pa((unsigned long)srmmu_context_table);
843
844         for(i = 0; i < num_contexts; i++)
845                 srmmu_ctxd_set((ctxd_t *)__nocache_fix(&srmmu_context_table[i]), srmmu_swapper_pg_dir);
846
847         flush_cache_all();
848         srmmu_set_ctable_ptr((unsigned long)srmmu_ctx_table_phys);
849 #ifdef CONFIG_SMP
850         /* Stop from hanging here... */
851         local_ops->tlb_all();
852 #else
853         flush_tlb_all();
854 #endif
855         poke_srmmu();
856
857         srmmu_allocate_ptable_skeleton(sparc_iomap.start, IOBASE_END);
858         srmmu_allocate_ptable_skeleton(DVMA_VADDR, DVMA_END);
859
860         srmmu_allocate_ptable_skeleton(
861                 __fix_to_virt(__end_of_fixed_addresses - 1), FIXADDR_TOP);
862         srmmu_allocate_ptable_skeleton(PKMAP_BASE, PKMAP_END);
863
864         pgd = pgd_offset_k(PKMAP_BASE);
865         pmd = pmd_offset(pgd, PKMAP_BASE);
866         pte = pte_offset_kernel(pmd, PKMAP_BASE);
867         pkmap_page_table = pte;
868
869         flush_cache_all();
870         flush_tlb_all();
871
872         sparc_context_init(num_contexts);
873
874         kmap_init();
875
876         {
877                 unsigned long zones_size[MAX_NR_ZONES];
878                 unsigned long zholes_size[MAX_NR_ZONES];
879                 unsigned long npages;
880                 int znum;
881
882                 for (znum = 0; znum < MAX_NR_ZONES; znum++)
883                         zones_size[znum] = zholes_size[znum] = 0;
884
885                 npages = max_low_pfn - pfn_base;
886
887                 zones_size[ZONE_DMA] = npages;
888                 zholes_size[ZONE_DMA] = npages - pages_avail;
889
890                 npages = highend_pfn - max_low_pfn;
891                 zones_size[ZONE_HIGHMEM] = npages;
892                 zholes_size[ZONE_HIGHMEM] = npages - calc_highpages();
893
894                 free_area_init_node(0, zones_size, pfn_base, zholes_size);
895         }
896 }
897
898 void mmu_info(struct seq_file *m)
899 {
900         seq_printf(m, 
901                    "MMU type\t: %s\n"
902                    "contexts\t: %d\n"
903                    "nocache total\t: %ld\n"
904                    "nocache used\t: %d\n",
905                    srmmu_name,
906                    num_contexts,
907                    srmmu_nocache_size,
908                    srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
909 }
910
911 void destroy_context(struct mm_struct *mm)
912 {
913
914         if(mm->context != NO_CONTEXT) {
915                 flush_cache_mm(mm);
916                 srmmu_ctxd_set(&srmmu_context_table[mm->context], srmmu_swapper_pg_dir);
917                 flush_tlb_mm(mm);
918                 spin_lock(&srmmu_context_spinlock);
919                 free_context(mm->context);
920                 spin_unlock(&srmmu_context_spinlock);
921                 mm->context = NO_CONTEXT;
922         }
923 }
924
925 /* Init various srmmu chip types. */
926 static void __init srmmu_is_bad(void)
927 {
928         prom_printf("Could not determine SRMMU chip type.\n");
929         prom_halt();
930 }
931
932 static void __init init_vac_layout(void)
933 {
934         phandle nd;
935         int cache_lines;
936         char node_str[128];
937 #ifdef CONFIG_SMP
938         int cpu = 0;
939         unsigned long max_size = 0;
940         unsigned long min_line_size = 0x10000000;
941 #endif
942
943         nd = prom_getchild(prom_root_node);
944         while((nd = prom_getsibling(nd)) != 0) {
945                 prom_getstring(nd, "device_type", node_str, sizeof(node_str));
946                 if(!strcmp(node_str, "cpu")) {
947                         vac_line_size = prom_getint(nd, "cache-line-size");
948                         if (vac_line_size == -1) {
949                                 prom_printf("can't determine cache-line-size, "
950                                             "halting.\n");
951                                 prom_halt();
952                         }
953                         cache_lines = prom_getint(nd, "cache-nlines");
954                         if (cache_lines == -1) {
955                                 prom_printf("can't determine cache-nlines, halting.\n");
956                                 prom_halt();
957                         }
958
959                         vac_cache_size = cache_lines * vac_line_size;
960 #ifdef CONFIG_SMP
961                         if(vac_cache_size > max_size)
962                                 max_size = vac_cache_size;
963                         if(vac_line_size < min_line_size)
964                                 min_line_size = vac_line_size;
965                         //FIXME: cpus not contiguous!!
966                         cpu++;
967                         if (cpu >= nr_cpu_ids || !cpu_online(cpu))
968                                 break;
969 #else
970                         break;
971 #endif
972                 }
973         }
974         if(nd == 0) {
975                 prom_printf("No CPU nodes found, halting.\n");
976                 prom_halt();
977         }
978 #ifdef CONFIG_SMP
979         vac_cache_size = max_size;
980         vac_line_size = min_line_size;
981 #endif
982         printk("SRMMU: Using VAC size of %d bytes, line size %d bytes.\n",
983                (int)vac_cache_size, (int)vac_line_size);
984 }
985
986 static void __cpuinit poke_hypersparc(void)
987 {
988         volatile unsigned long clear;
989         unsigned long mreg = srmmu_get_mmureg();
990
991         hyper_flush_unconditional_combined();
992
993         mreg &= ~(HYPERSPARC_CWENABLE);
994         mreg |= (HYPERSPARC_CENABLE | HYPERSPARC_WBENABLE);
995         mreg |= (HYPERSPARC_CMODE);
996
997         srmmu_set_mmureg(mreg);
998
999 #if 0 /* XXX I think this is bad news... -DaveM */
1000         hyper_clear_all_tags();
1001 #endif
1002
1003         put_ross_icr(HYPERSPARC_ICCR_FTD | HYPERSPARC_ICCR_ICE);
1004         hyper_flush_whole_icache();
1005         clear = srmmu_get_faddr();
1006         clear = srmmu_get_fstatus();
1007 }
1008
1009 static const struct sparc32_cachetlb_ops hypersparc_ops = {
1010         .cache_all      = hypersparc_flush_cache_all,
1011         .cache_mm       = hypersparc_flush_cache_mm,
1012         .cache_page     = hypersparc_flush_cache_page,
1013         .cache_range    = hypersparc_flush_cache_range,
1014         .tlb_all        = hypersparc_flush_tlb_all,
1015         .tlb_mm         = hypersparc_flush_tlb_mm,
1016         .tlb_page       = hypersparc_flush_tlb_page,
1017         .tlb_range      = hypersparc_flush_tlb_range,
1018         .page_to_ram    = hypersparc_flush_page_to_ram,
1019         .sig_insns      = hypersparc_flush_sig_insns,
1020         .page_for_dma   = hypersparc_flush_page_for_dma,
1021 };
1022
1023 static void __init init_hypersparc(void)
1024 {
1025         srmmu_name = "ROSS HyperSparc";
1026         srmmu_modtype = HyperSparc;
1027
1028         init_vac_layout();
1029
1030         is_hypersparc = 1;
1031         sparc32_cachetlb_ops = &hypersparc_ops;
1032
1033         poke_srmmu = poke_hypersparc;
1034
1035         hypersparc_setup_blockops();
1036 }
1037
1038 static void __cpuinit poke_swift(void)
1039 {
1040         unsigned long mreg;
1041
1042         /* Clear any crap from the cache or else... */
1043         swift_flush_cache_all();
1044
1045         /* Enable I & D caches */
1046         mreg = srmmu_get_mmureg();
1047         mreg |= (SWIFT_IE | SWIFT_DE);
1048         /*
1049          * The Swift branch folding logic is completely broken.  At
1050          * trap time, if things are just right, if can mistakenly
1051          * think that a trap is coming from kernel mode when in fact
1052          * it is coming from user mode (it mis-executes the branch in
1053          * the trap code).  So you see things like crashme completely
1054          * hosing your machine which is completely unacceptable.  Turn
1055          * this shit off... nice job Fujitsu.
1056          */
1057         mreg &= ~(SWIFT_BF);
1058         srmmu_set_mmureg(mreg);
1059 }
1060
1061 static const struct sparc32_cachetlb_ops swift_ops = {
1062         .cache_all      = swift_flush_cache_all,
1063         .cache_mm       = swift_flush_cache_mm,
1064         .cache_page     = swift_flush_cache_page,
1065         .cache_range    = swift_flush_cache_range,
1066         .tlb_all        = swift_flush_tlb_all,
1067         .tlb_mm         = swift_flush_tlb_mm,
1068         .tlb_page       = swift_flush_tlb_page,
1069         .tlb_range      = swift_flush_tlb_range,
1070         .page_to_ram    = swift_flush_page_to_ram,
1071         .sig_insns      = swift_flush_sig_insns,
1072         .page_for_dma   = swift_flush_page_for_dma,
1073 };
1074
1075 #define SWIFT_MASKID_ADDR  0x10003018
1076 static void __init init_swift(void)
1077 {
1078         unsigned long swift_rev;
1079
1080         __asm__ __volatile__("lda [%1] %2, %0\n\t"
1081                              "srl %0, 0x18, %0\n\t" :
1082                              "=r" (swift_rev) :
1083                              "r" (SWIFT_MASKID_ADDR), "i" (ASI_M_BYPASS));
1084         srmmu_name = "Fujitsu Swift";
1085         switch(swift_rev) {
1086         case 0x11:
1087         case 0x20:
1088         case 0x23:
1089         case 0x30:
1090                 srmmu_modtype = Swift_lots_o_bugs;
1091                 hwbug_bitmask |= (HWBUG_KERN_ACCBROKEN | HWBUG_KERN_CBITBROKEN);
1092                 /*
1093                  * Gee george, I wonder why Sun is so hush hush about
1094                  * this hardware bug... really braindamage stuff going
1095                  * on here.  However I think we can find a way to avoid
1096                  * all of the workaround overhead under Linux.  Basically,
1097                  * any page fault can cause kernel pages to become user
1098                  * accessible (the mmu gets confused and clears some of
1099                  * the ACC bits in kernel ptes).  Aha, sounds pretty
1100                  * horrible eh?  But wait, after extensive testing it appears
1101                  * that if you use pgd_t level large kernel pte's (like the
1102                  * 4MB pages on the Pentium) the bug does not get tripped
1103                  * at all.  This avoids almost all of the major overhead.
1104                  * Welcome to a world where your vendor tells you to,
1105                  * "apply this kernel patch" instead of "sorry for the
1106                  * broken hardware, send it back and we'll give you
1107                  * properly functioning parts"
1108                  */
1109                 break;
1110         case 0x25:
1111         case 0x31:
1112                 srmmu_modtype = Swift_bad_c;
1113                 hwbug_bitmask |= HWBUG_KERN_CBITBROKEN;
1114                 /*
1115                  * You see Sun allude to this hardware bug but never
1116                  * admit things directly, they'll say things like,
1117                  * "the Swift chip cache problems" or similar.
1118                  */
1119                 break;
1120         default:
1121                 srmmu_modtype = Swift_ok;
1122                 break;
1123         }
1124
1125         sparc32_cachetlb_ops = &swift_ops;
1126         flush_page_for_dma_global = 0;
1127
1128         /*
1129          * Are you now convinced that the Swift is one of the
1130          * biggest VLSI abortions of all time?  Bravo Fujitsu!
1131          * Fujitsu, the !#?!%$'d up processor people.  I bet if
1132          * you examined the microcode of the Swift you'd find
1133          * XXX's all over the place.
1134          */
1135         poke_srmmu = poke_swift;
1136 }
1137
1138 static void turbosparc_flush_cache_all(void)
1139 {
1140         flush_user_windows();
1141         turbosparc_idflash_clear();
1142 }
1143
1144 static void turbosparc_flush_cache_mm(struct mm_struct *mm)
1145 {
1146         FLUSH_BEGIN(mm)
1147         flush_user_windows();
1148         turbosparc_idflash_clear();
1149         FLUSH_END
1150 }
1151
1152 static void turbosparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1153 {
1154         FLUSH_BEGIN(vma->vm_mm)
1155         flush_user_windows();
1156         turbosparc_idflash_clear();
1157         FLUSH_END
1158 }
1159
1160 static void turbosparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1161 {
1162         FLUSH_BEGIN(vma->vm_mm)
1163         flush_user_windows();
1164         if (vma->vm_flags & VM_EXEC)
1165                 turbosparc_flush_icache();
1166         turbosparc_flush_dcache();
1167         FLUSH_END
1168 }
1169
1170 /* TurboSparc is copy-back, if we turn it on, but this does not work. */
1171 static void turbosparc_flush_page_to_ram(unsigned long page)
1172 {
1173 #ifdef TURBOSPARC_WRITEBACK
1174         volatile unsigned long clear;
1175
1176         if (srmmu_probe(page))
1177                 turbosparc_flush_page_cache(page);
1178         clear = srmmu_get_fstatus();
1179 #endif
1180 }
1181
1182 static void turbosparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1183 {
1184 }
1185
1186 static void turbosparc_flush_page_for_dma(unsigned long page)
1187 {
1188         turbosparc_flush_dcache();
1189 }
1190
1191 static void turbosparc_flush_tlb_all(void)
1192 {
1193         srmmu_flush_whole_tlb();
1194 }
1195
1196 static void turbosparc_flush_tlb_mm(struct mm_struct *mm)
1197 {
1198         FLUSH_BEGIN(mm)
1199         srmmu_flush_whole_tlb();
1200         FLUSH_END
1201 }
1202
1203 static void turbosparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1204 {
1205         FLUSH_BEGIN(vma->vm_mm)
1206         srmmu_flush_whole_tlb();
1207         FLUSH_END
1208 }
1209
1210 static void turbosparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1211 {
1212         FLUSH_BEGIN(vma->vm_mm)
1213         srmmu_flush_whole_tlb();
1214         FLUSH_END
1215 }
1216
1217
1218 static void __cpuinit poke_turbosparc(void)
1219 {
1220         unsigned long mreg = srmmu_get_mmureg();
1221         unsigned long ccreg;
1222
1223         /* Clear any crap from the cache or else... */
1224         turbosparc_flush_cache_all();
1225         mreg &= ~(TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE); /* Temporarily disable I & D caches */
1226         mreg &= ~(TURBOSPARC_PCENABLE);         /* Don't check parity */
1227         srmmu_set_mmureg(mreg);
1228         
1229         ccreg = turbosparc_get_ccreg();
1230
1231 #ifdef TURBOSPARC_WRITEBACK
1232         ccreg |= (TURBOSPARC_SNENABLE);         /* Do DVMA snooping in Dcache */
1233         ccreg &= ~(TURBOSPARC_uS2 | TURBOSPARC_WTENABLE);
1234                         /* Write-back D-cache, emulate VLSI
1235                          * abortion number three, not number one */
1236 #else
1237         /* For now let's play safe, optimize later */
1238         ccreg |= (TURBOSPARC_SNENABLE | TURBOSPARC_WTENABLE);
1239                         /* Do DVMA snooping in Dcache, Write-thru D-cache */
1240         ccreg &= ~(TURBOSPARC_uS2);
1241                         /* Emulate VLSI abortion number three, not number one */
1242 #endif
1243
1244         switch (ccreg & 7) {
1245         case 0: /* No SE cache */
1246         case 7: /* Test mode */
1247                 break;
1248         default:
1249                 ccreg |= (TURBOSPARC_SCENABLE);
1250         }
1251         turbosparc_set_ccreg (ccreg);
1252
1253         mreg |= (TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE); /* I & D caches on */
1254         mreg |= (TURBOSPARC_ICSNOOP);           /* Icache snooping on */
1255         srmmu_set_mmureg(mreg);
1256 }
1257
1258 static const struct sparc32_cachetlb_ops turbosparc_ops = {
1259         .cache_all      = turbosparc_flush_cache_all,
1260         .cache_mm       = turbosparc_flush_cache_mm,
1261         .cache_page     = turbosparc_flush_cache_page,
1262         .cache_range    = turbosparc_flush_cache_range,
1263         .tlb_all        = turbosparc_flush_tlb_all,
1264         .tlb_mm         = turbosparc_flush_tlb_mm,
1265         .tlb_page       = turbosparc_flush_tlb_page,
1266         .tlb_range      = turbosparc_flush_tlb_range,
1267         .page_to_ram    = turbosparc_flush_page_to_ram,
1268         .sig_insns      = turbosparc_flush_sig_insns,
1269         .page_for_dma   = turbosparc_flush_page_for_dma,
1270 };
1271
1272 static void __init init_turbosparc(void)
1273 {
1274         srmmu_name = "Fujitsu TurboSparc";
1275         srmmu_modtype = TurboSparc;
1276         sparc32_cachetlb_ops = &turbosparc_ops;
1277         poke_srmmu = poke_turbosparc;
1278 }
1279
1280 static void __cpuinit poke_tsunami(void)
1281 {
1282         unsigned long mreg = srmmu_get_mmureg();
1283
1284         tsunami_flush_icache();
1285         tsunami_flush_dcache();
1286         mreg &= ~TSUNAMI_ITD;
1287         mreg |= (TSUNAMI_IENAB | TSUNAMI_DENAB);
1288         srmmu_set_mmureg(mreg);
1289 }
1290
1291 static const struct sparc32_cachetlb_ops tsunami_ops = {
1292         .cache_all      = tsunami_flush_cache_all,
1293         .cache_mm       = tsunami_flush_cache_mm,
1294         .cache_page     = tsunami_flush_cache_page,
1295         .cache_range    = tsunami_flush_cache_range,
1296         .tlb_all        = tsunami_flush_tlb_all,
1297         .tlb_mm         = tsunami_flush_tlb_mm,
1298         .tlb_page       = tsunami_flush_tlb_page,
1299         .tlb_range      = tsunami_flush_tlb_range,
1300         .page_to_ram    = tsunami_flush_page_to_ram,
1301         .sig_insns      = tsunami_flush_sig_insns,
1302         .page_for_dma   = tsunami_flush_page_for_dma,
1303 };
1304
1305 static void __init init_tsunami(void)
1306 {
1307         /*
1308          * Tsunami's pretty sane, Sun and TI actually got it
1309          * somewhat right this time.  Fujitsu should have
1310          * taken some lessons from them.
1311          */
1312
1313         srmmu_name = "TI Tsunami";
1314         srmmu_modtype = Tsunami;
1315         sparc32_cachetlb_ops = &tsunami_ops;
1316         poke_srmmu = poke_tsunami;
1317
1318         tsunami_setup_blockops();
1319 }
1320
1321 static void __cpuinit poke_viking(void)
1322 {
1323         unsigned long mreg = srmmu_get_mmureg();
1324         static int smp_catch;
1325
1326         if (viking_mxcc_present) {
1327                 unsigned long mxcc_control = mxcc_get_creg();
1328
1329                 mxcc_control |= (MXCC_CTL_ECE | MXCC_CTL_PRE | MXCC_CTL_MCE);
1330                 mxcc_control &= ~(MXCC_CTL_RRC);
1331                 mxcc_set_creg(mxcc_control);
1332
1333                 /*
1334                  * We don't need memory parity checks.
1335                  * XXX This is a mess, have to dig out later. ecd.
1336                 viking_mxcc_turn_off_parity(&mreg, &mxcc_control);
1337                  */
1338
1339                 /* We do cache ptables on MXCC. */
1340                 mreg |= VIKING_TCENABLE;
1341         } else {
1342                 unsigned long bpreg;
1343
1344                 mreg &= ~(VIKING_TCENABLE);
1345                 if(smp_catch++) {
1346                         /* Must disable mixed-cmd mode here for other cpu's. */
1347                         bpreg = viking_get_bpreg();
1348                         bpreg &= ~(VIKING_ACTION_MIX);
1349                         viking_set_bpreg(bpreg);
1350
1351                         /* Just in case PROM does something funny. */
1352                         msi_set_sync();
1353                 }
1354         }
1355
1356         mreg |= VIKING_SPENABLE;
1357         mreg |= (VIKING_ICENABLE | VIKING_DCENABLE);
1358         mreg |= VIKING_SBENABLE;
1359         mreg &= ~(VIKING_ACENABLE);
1360         srmmu_set_mmureg(mreg);
1361 }
1362
1363 static struct sparc32_cachetlb_ops viking_ops = {
1364         .cache_all      = viking_flush_cache_all,
1365         .cache_mm       = viking_flush_cache_mm,
1366         .cache_page     = viking_flush_cache_page,
1367         .cache_range    = viking_flush_cache_range,
1368         .tlb_all        = viking_flush_tlb_all,
1369         .tlb_mm         = viking_flush_tlb_mm,
1370         .tlb_page       = viking_flush_tlb_page,
1371         .tlb_range      = viking_flush_tlb_range,
1372         .page_to_ram    = viking_flush_page_to_ram,
1373         .sig_insns      = viking_flush_sig_insns,
1374         .page_for_dma   = viking_flush_page_for_dma,
1375 };
1376
1377 #ifdef CONFIG_SMP
1378 /* On sun4d the cpu broadcasts local TLB flushes, so we can just
1379  * perform the local TLB flush and all the other cpus will see it.
1380  * But, unfortunately, there is a bug in the sun4d XBUS backplane
1381  * that requires that we add some synchronization to these flushes.
1382  *
1383  * The bug is that the fifo which keeps track of all the pending TLB
1384  * broadcasts in the system is an entry or two too small, so if we
1385  * have too many going at once we'll overflow that fifo and lose a TLB
1386  * flush resulting in corruption.
1387  *
1388  * Our workaround is to take a global spinlock around the TLB flushes,
1389  * which guarentees we won't ever have too many pending.  It's a big
1390  * hammer, but a semaphore like system to make sure we only have N TLB
1391  * flushes going at once will require SMP locking anyways so there's
1392  * no real value in trying any harder than this.
1393  */
1394 static struct sparc32_cachetlb_ops viking_sun4d_smp_ops = {
1395         .cache_all      = viking_flush_cache_all,
1396         .cache_mm       = viking_flush_cache_mm,
1397         .cache_page     = viking_flush_cache_page,
1398         .cache_range    = viking_flush_cache_range,
1399         .tlb_all        = sun4dsmp_flush_tlb_all,
1400         .tlb_mm         = sun4dsmp_flush_tlb_mm,
1401         .tlb_page       = sun4dsmp_flush_tlb_page,
1402         .tlb_range      = sun4dsmp_flush_tlb_range,
1403         .page_to_ram    = viking_flush_page_to_ram,
1404         .sig_insns      = viking_flush_sig_insns,
1405         .page_for_dma   = viking_flush_page_for_dma,
1406 };
1407 #endif
1408
1409 static void __init init_viking(void)
1410 {
1411         unsigned long mreg = srmmu_get_mmureg();
1412
1413         /* Ahhh, the viking.  SRMMU VLSI abortion number two... */
1414         if(mreg & VIKING_MMODE) {
1415                 srmmu_name = "TI Viking";
1416                 viking_mxcc_present = 0;
1417                 msi_set_sync();
1418
1419                 /*
1420                  * We need this to make sure old viking takes no hits
1421                  * on it's cache for dma snoops to workaround the
1422                  * "load from non-cacheable memory" interrupt bug.
1423                  * This is only necessary because of the new way in
1424                  * which we use the IOMMU.
1425                  */
1426                 viking_ops.page_for_dma = viking_flush_page;
1427 #ifdef CONFIG_SMP
1428                 viking_sun4d_smp_ops.page_for_dma = viking_flush_page;
1429 #endif
1430                 flush_page_for_dma_global = 0;
1431         } else {
1432                 srmmu_name = "TI Viking/MXCC";
1433                 viking_mxcc_present = 1;
1434                 srmmu_cache_pagetables = 1;
1435         }
1436
1437         sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1438                 &viking_ops;
1439 #ifdef CONFIG_SMP
1440         if (sparc_cpu_model == sun4d)
1441                 sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1442                         &viking_sun4d_smp_ops;
1443 #endif
1444
1445         poke_srmmu = poke_viking;
1446 }
1447
1448 /* Probe for the srmmu chip version. */
1449 static void __init get_srmmu_type(void)
1450 {
1451         unsigned long mreg, psr;
1452         unsigned long mod_typ, mod_rev, psr_typ, psr_vers;
1453
1454         srmmu_modtype = SRMMU_INVAL_MOD;
1455         hwbug_bitmask = 0;
1456
1457         mreg = srmmu_get_mmureg(); psr = get_psr();
1458         mod_typ = (mreg & 0xf0000000) >> 28;
1459         mod_rev = (mreg & 0x0f000000) >> 24;
1460         psr_typ = (psr >> 28) & 0xf;
1461         psr_vers = (psr >> 24) & 0xf;
1462
1463         /* First, check for sparc-leon. */
1464         if (sparc_cpu_model == sparc_leon) {
1465                 init_leon();
1466                 return;
1467         }
1468
1469         /* Second, check for HyperSparc or Cypress. */
1470         if(mod_typ == 1) {
1471                 switch(mod_rev) {
1472                 case 7:
1473                         /* UP or MP Hypersparc */
1474                         init_hypersparc();
1475                         break;
1476                 case 0:
1477                 case 2:
1478                 case 10:
1479                 case 11:
1480                 case 12:
1481                 case 13:
1482                 case 14:
1483                 case 15:
1484                 default:
1485                         prom_printf("Sparc-Linux Cypress support does not longer exit.\n");
1486                         prom_halt();
1487                         break;
1488                 }
1489                 return;
1490         }
1491         
1492         /*
1493          * Now Fujitsu TurboSparc. It might happen that it is
1494          * in Swift emulation mode, so we will check later...
1495          */
1496         if (psr_typ == 0 && psr_vers == 5) {
1497                 init_turbosparc();
1498                 return;
1499         }
1500
1501         /* Next check for Fujitsu Swift. */
1502         if(psr_typ == 0 && psr_vers == 4) {
1503                 phandle cpunode;
1504                 char node_str[128];
1505
1506                 /* Look if it is not a TurboSparc emulating Swift... */
1507                 cpunode = prom_getchild(prom_root_node);
1508                 while((cpunode = prom_getsibling(cpunode)) != 0) {
1509                         prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
1510                         if(!strcmp(node_str, "cpu")) {
1511                                 if (!prom_getintdefault(cpunode, "psr-implementation", 1) &&
1512                                     prom_getintdefault(cpunode, "psr-version", 1) == 5) {
1513                                         init_turbosparc();
1514                                         return;
1515                                 }
1516                                 break;
1517                         }
1518                 }
1519                 
1520                 init_swift();
1521                 return;
1522         }
1523
1524         /* Now the Viking family of srmmu. */
1525         if(psr_typ == 4 &&
1526            ((psr_vers == 0) ||
1527             ((psr_vers == 1) && (mod_typ == 0) && (mod_rev == 0)))) {
1528                 init_viking();
1529                 return;
1530         }
1531
1532         /* Finally the Tsunami. */
1533         if(psr_typ == 4 && psr_vers == 1 && (mod_typ || mod_rev)) {
1534                 init_tsunami();
1535                 return;
1536         }
1537
1538         /* Oh well */
1539         srmmu_is_bad();
1540 }
1541
1542 #ifdef CONFIG_SMP
1543 /* Local cross-calls. */
1544 static void smp_flush_page_for_dma(unsigned long page)
1545 {
1546         xc1((smpfunc_t) local_ops->page_for_dma, page);
1547         local_ops->page_for_dma(page);
1548 }
1549
1550 static void smp_flush_cache_all(void)
1551 {
1552         xc0((smpfunc_t) local_ops->cache_all);
1553         local_ops->cache_all();
1554 }
1555
1556 static void smp_flush_tlb_all(void)
1557 {
1558         xc0((smpfunc_t) local_ops->tlb_all);
1559         local_ops->tlb_all();
1560 }
1561
1562 static void smp_flush_cache_mm(struct mm_struct *mm)
1563 {
1564         if (mm->context != NO_CONTEXT) {
1565                 cpumask_t cpu_mask;
1566                 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1567                 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1568                 if (!cpumask_empty(&cpu_mask))
1569                         xc1((smpfunc_t) local_ops->cache_mm, (unsigned long) mm);
1570                 local_ops->cache_mm(mm);
1571         }
1572 }
1573
1574 static void smp_flush_tlb_mm(struct mm_struct *mm)
1575 {
1576         if (mm->context != NO_CONTEXT) {
1577                 cpumask_t cpu_mask;
1578                 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1579                 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1580                 if (!cpumask_empty(&cpu_mask)) {
1581                         xc1((smpfunc_t) local_ops->tlb_mm, (unsigned long) mm);
1582                         if (atomic_read(&mm->mm_users) == 1 && current->active_mm == mm)
1583                                 cpumask_copy(mm_cpumask(mm),
1584                                              cpumask_of(smp_processor_id()));
1585                 }
1586                 local_ops->tlb_mm(mm);
1587         }
1588 }
1589
1590 static void smp_flush_cache_range(struct vm_area_struct *vma,
1591                                   unsigned long start,
1592                                   unsigned long end)
1593 {
1594         struct mm_struct *mm = vma->vm_mm;
1595
1596         if (mm->context != NO_CONTEXT) {
1597                 cpumask_t cpu_mask;
1598                 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1599                 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1600                 if (!cpumask_empty(&cpu_mask))
1601                         xc3((smpfunc_t) local_ops->cache_range,
1602                             (unsigned long) vma, start, end);
1603                 local_ops->cache_range(vma, start, end);
1604         }
1605 }
1606
1607 static void smp_flush_tlb_range(struct vm_area_struct *vma,
1608                                 unsigned long start,
1609                                 unsigned long end)
1610 {
1611         struct mm_struct *mm = vma->vm_mm;
1612
1613         if (mm->context != NO_CONTEXT) {
1614                 cpumask_t cpu_mask;
1615                 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1616                 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1617                 if (!cpumask_empty(&cpu_mask))
1618                         xc3((smpfunc_t) local_ops->tlb_range,
1619                             (unsigned long) vma, start, end);
1620                 local_ops->tlb_range(vma, start, end);
1621         }
1622 }
1623
1624 static void smp_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1625 {
1626         struct mm_struct *mm = vma->vm_mm;
1627
1628         if (mm->context != NO_CONTEXT) {
1629                 cpumask_t cpu_mask;
1630                 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1631                 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1632                 if (!cpumask_empty(&cpu_mask))
1633                         xc2((smpfunc_t) local_ops->cache_page,
1634                             (unsigned long) vma, page);
1635                 local_ops->cache_page(vma, page);
1636         }
1637 }
1638
1639 static void smp_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1640 {
1641         struct mm_struct *mm = vma->vm_mm;
1642
1643         if (mm->context != NO_CONTEXT) {
1644                 cpumask_t cpu_mask;
1645                 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1646                 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1647                 if (!cpumask_empty(&cpu_mask))
1648                         xc2((smpfunc_t) local_ops->tlb_page,
1649                             (unsigned long) vma, page);
1650                 local_ops->tlb_page(vma, page);
1651         }
1652 }
1653
1654 static void smp_flush_page_to_ram(unsigned long page)
1655 {
1656         /* Current theory is that those who call this are the one's
1657          * who have just dirtied their cache with the pages contents
1658          * in kernel space, therefore we only run this on local cpu.
1659          *
1660          * XXX This experiment failed, research further... -DaveM
1661          */
1662 #if 1
1663         xc1((smpfunc_t) local_ops->page_to_ram, page);
1664 #endif
1665         local_ops->page_to_ram(page);
1666 }
1667
1668 static void smp_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1669 {
1670         cpumask_t cpu_mask;
1671         cpumask_copy(&cpu_mask, mm_cpumask(mm));
1672         cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1673         if (!cpumask_empty(&cpu_mask))
1674                 xc2((smpfunc_t) local_ops->sig_insns,
1675                     (unsigned long) mm, insn_addr);
1676         local_ops->sig_insns(mm, insn_addr);
1677 }
1678
1679 static struct sparc32_cachetlb_ops smp_cachetlb_ops = {
1680         .cache_all      = smp_flush_cache_all,
1681         .cache_mm       = smp_flush_cache_mm,
1682         .cache_page     = smp_flush_cache_page,
1683         .cache_range    = smp_flush_cache_range,
1684         .tlb_all        = smp_flush_tlb_all,
1685         .tlb_mm         = smp_flush_tlb_mm,
1686         .tlb_page       = smp_flush_tlb_page,
1687         .tlb_range      = smp_flush_tlb_range,
1688         .page_to_ram    = smp_flush_page_to_ram,
1689         .sig_insns      = smp_flush_sig_insns,
1690         .page_for_dma   = smp_flush_page_for_dma,
1691 };
1692 #endif
1693
1694 /* Load up routines and constants for sun4m and sun4d mmu */
1695 void __init load_mmu(void)
1696 {
1697         extern void ld_mmu_iommu(void);
1698         extern void ld_mmu_iounit(void);
1699
1700         /* Functions */
1701         get_srmmu_type();
1702
1703 #ifdef CONFIG_SMP
1704         /* El switcheroo... */
1705         local_ops = sparc32_cachetlb_ops;
1706
1707         if (sparc_cpu_model == sun4d || sparc_cpu_model == sparc_leon) {
1708                 smp_cachetlb_ops.tlb_all = local_ops->tlb_all;
1709                 smp_cachetlb_ops.tlb_mm = local_ops->tlb_mm;
1710                 smp_cachetlb_ops.tlb_range = local_ops->tlb_range;
1711                 smp_cachetlb_ops.tlb_page = local_ops->tlb_page;
1712         }
1713
1714         if (poke_srmmu == poke_viking) {
1715                 /* Avoid unnecessary cross calls. */
1716                 smp_cachetlb_ops.cache_all = local_ops->cache_all;
1717                 smp_cachetlb_ops.cache_mm = local_ops->cache_mm;
1718                 smp_cachetlb_ops.cache_range = local_ops->cache_range;
1719                 smp_cachetlb_ops.cache_page = local_ops->cache_page;
1720
1721                 smp_cachetlb_ops.page_to_ram = local_ops->page_to_ram;
1722                 smp_cachetlb_ops.sig_insns = local_ops->sig_insns;
1723                 smp_cachetlb_ops.page_for_dma = local_ops->page_for_dma;
1724         }
1725
1726         /* It really is const after this point. */
1727         sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1728                 &smp_cachetlb_ops;
1729 #endif
1730
1731         if (sparc_cpu_model == sun4d)
1732                 ld_mmu_iounit();
1733         else
1734                 ld_mmu_iommu();
1735 #ifdef CONFIG_SMP
1736         if (sparc_cpu_model == sun4d)
1737                 sun4d_init_smp();
1738         else if (sparc_cpu_model == sparc_leon)
1739                 leon_init_smp();
1740         else
1741                 sun4m_init_smp();
1742 #endif
1743 }