f8fb4911d360b1c68d10d6c043ac2d4ee467ad7d
[linux-block.git] / arch / sparc / mm / srmmu.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * srmmu.c:  SRMMU specific routines for memory management.
4  *
5  * Copyright (C) 1995 David S. Miller  (davem@caip.rutgers.edu)
6  * Copyright (C) 1995,2002 Pete Zaitcev (zaitcev@yahoo.com)
7  * Copyright (C) 1996 Eddie C. Dost    (ecd@skynet.be)
8  * Copyright (C) 1997,1998 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
9  * Copyright (C) 1999,2000 Anton Blanchard (anton@samba.org)
10  */
11
12 #include <linux/seq_file.h>
13 #include <linux/spinlock.h>
14 #include <linux/memblock.h>
15 #include <linux/pagemap.h>
16 #include <linux/vmalloc.h>
17 #include <linux/kdebug.h>
18 #include <linux/export.h>
19 #include <linux/kernel.h>
20 #include <linux/init.h>
21 #include <linux/log2.h>
22 #include <linux/gfp.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25
26 #include <asm/mmu_context.h>
27 #include <asm/cacheflush.h>
28 #include <asm/tlbflush.h>
29 #include <asm/io-unit.h>
30 #include <asm/pgalloc.h>
31 #include <asm/pgtable.h>
32 #include <asm/bitext.h>
33 #include <asm/vaddrs.h>
34 #include <asm/cache.h>
35 #include <asm/traps.h>
36 #include <asm/oplib.h>
37 #include <asm/mbus.h>
38 #include <asm/page.h>
39 #include <asm/asi.h>
40 #include <asm/smp.h>
41 #include <asm/io.h>
42
43 /* Now the cpu specific definitions. */
44 #include <asm/turbosparc.h>
45 #include <asm/tsunami.h>
46 #include <asm/viking.h>
47 #include <asm/swift.h>
48 #include <asm/leon.h>
49 #include <asm/mxcc.h>
50 #include <asm/ross.h>
51
52 #include "mm_32.h"
53
54 enum mbus_module srmmu_modtype;
55 static unsigned int hwbug_bitmask;
56 int vac_cache_size;
57 EXPORT_SYMBOL(vac_cache_size);
58 int vac_line_size;
59
60 extern struct resource sparc_iomap;
61
62 extern unsigned long last_valid_pfn;
63
64 static pgd_t *srmmu_swapper_pg_dir;
65
66 const struct sparc32_cachetlb_ops *sparc32_cachetlb_ops;
67 EXPORT_SYMBOL(sparc32_cachetlb_ops);
68
69 #ifdef CONFIG_SMP
70 const struct sparc32_cachetlb_ops *local_ops;
71
72 #define FLUSH_BEGIN(mm)
73 #define FLUSH_END
74 #else
75 #define FLUSH_BEGIN(mm) if ((mm)->context != NO_CONTEXT) {
76 #define FLUSH_END       }
77 #endif
78
79 int flush_page_for_dma_global = 1;
80
81 char *srmmu_name;
82
83 ctxd_t *srmmu_ctx_table_phys;
84 static ctxd_t *srmmu_context_table;
85
86 int viking_mxcc_present;
87 static DEFINE_SPINLOCK(srmmu_context_spinlock);
88
89 static int is_hypersparc;
90
91 static int srmmu_cache_pagetables;
92
93 /* these will be initialized in srmmu_nocache_calcsize() */
94 static unsigned long srmmu_nocache_size;
95 static unsigned long srmmu_nocache_end;
96
97 /* 1 bit <=> 256 bytes of nocache <=> 64 PTEs */
98 #define SRMMU_NOCACHE_BITMAP_SHIFT (PAGE_SHIFT - 4)
99
100 /* The context table is a nocache user with the biggest alignment needs. */
101 #define SRMMU_NOCACHE_ALIGN_MAX (sizeof(ctxd_t)*SRMMU_MAX_CONTEXTS)
102
103 void *srmmu_nocache_pool;
104 static struct bit_map srmmu_nocache_map;
105
106 static inline int srmmu_pmd_none(pmd_t pmd)
107 { return !(pmd_val(pmd) & 0xFFFFFFF); }
108
109 /* XXX should we hyper_flush_whole_icache here - Anton */
110 static inline void srmmu_ctxd_set(ctxd_t *ctxp, pgd_t *pgdp)
111 {
112         pte_t pte;
113
114         pte = __pte((SRMMU_ET_PTD | (__nocache_pa(pgdp) >> 4)));
115         set_pte((pte_t *)ctxp, pte);
116 }
117
118 /*
119  * Locations of MSI Registers.
120  */
121 #define MSI_MBUS_ARBEN  0xe0001008      /* MBus Arbiter Enable register */
122
123 /*
124  * Useful bits in the MSI Registers.
125  */
126 #define MSI_ASYNC_MODE  0x80000000      /* Operate the MSI asynchronously */
127
128 static void msi_set_sync(void)
129 {
130         __asm__ __volatile__ ("lda [%0] %1, %%g3\n\t"
131                               "andn %%g3, %2, %%g3\n\t"
132                               "sta %%g3, [%0] %1\n\t" : :
133                               "r" (MSI_MBUS_ARBEN),
134                               "i" (ASI_M_CTL), "r" (MSI_ASYNC_MODE) : "g3");
135 }
136
137 void pmd_set(pmd_t *pmdp, pte_t *ptep)
138 {
139         unsigned long ptp = __nocache_pa(ptep) >> 4;
140         set_pte((pte_t *)&pmd_val(*pmdp), __pte(SRMMU_ET_PTD | ptp));
141 }
142
143 /*
144  * size: bytes to allocate in the nocache area.
145  * align: bytes, number to align at.
146  * Returns the virtual address of the allocated area.
147  */
148 static void *__srmmu_get_nocache(int size, int align)
149 {
150         int offset, minsz = 1 << SRMMU_NOCACHE_BITMAP_SHIFT;
151         unsigned long addr;
152
153         if (size < minsz) {
154                 printk(KERN_ERR "Size 0x%x too small for nocache request\n",
155                        size);
156                 size = minsz;
157         }
158         if (size & (minsz - 1)) {
159                 printk(KERN_ERR "Size 0x%x unaligned in nocache request\n",
160                        size);
161                 size += minsz - 1;
162         }
163         BUG_ON(align > SRMMU_NOCACHE_ALIGN_MAX);
164
165         offset = bit_map_string_get(&srmmu_nocache_map,
166                                     size >> SRMMU_NOCACHE_BITMAP_SHIFT,
167                                     align >> SRMMU_NOCACHE_BITMAP_SHIFT);
168         if (offset == -1) {
169                 printk(KERN_ERR "srmmu: out of nocache %d: %d/%d\n",
170                        size, (int) srmmu_nocache_size,
171                        srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
172                 return NULL;
173         }
174
175         addr = SRMMU_NOCACHE_VADDR + (offset << SRMMU_NOCACHE_BITMAP_SHIFT);
176         return (void *)addr;
177 }
178
179 void *srmmu_get_nocache(int size, int align)
180 {
181         void *tmp;
182
183         tmp = __srmmu_get_nocache(size, align);
184
185         if (tmp)
186                 memset(tmp, 0, size);
187
188         return tmp;
189 }
190
191 void srmmu_free_nocache(void *addr, int size)
192 {
193         unsigned long vaddr;
194         int offset;
195
196         vaddr = (unsigned long)addr;
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 /* Return how much physical memory we have.  */
230 static unsigned long __init probe_memory(void)
231 {
232         unsigned long total = 0;
233         int i;
234
235         for (i = 0; sp_banks[i].num_bytes; i++)
236                 total += sp_banks[i].num_bytes;
237
238         return total;
239 }
240
241 /*
242  * Reserve nocache dynamically proportionally to the amount of
243  * system RAM. -- Tomas Szepe <szepe@pinerecords.com>, June 2002
244  */
245 static void __init srmmu_nocache_calcsize(void)
246 {
247         unsigned long sysmemavail = probe_memory() / 1024;
248         int srmmu_nocache_npages;
249
250         srmmu_nocache_npages =
251                 sysmemavail / SRMMU_NOCACHE_ALCRATIO / 1024 * 256;
252
253  /* P3 XXX The 4x overuse: corroborated by /proc/meminfo. */
254         // if (srmmu_nocache_npages < 256) srmmu_nocache_npages = 256;
255         if (srmmu_nocache_npages < SRMMU_MIN_NOCACHE_PAGES)
256                 srmmu_nocache_npages = SRMMU_MIN_NOCACHE_PAGES;
257
258         /* anything above 1280 blows up */
259         if (srmmu_nocache_npages > SRMMU_MAX_NOCACHE_PAGES)
260                 srmmu_nocache_npages = SRMMU_MAX_NOCACHE_PAGES;
261
262         srmmu_nocache_size = srmmu_nocache_npages * PAGE_SIZE;
263         srmmu_nocache_end = SRMMU_NOCACHE_VADDR + srmmu_nocache_size;
264 }
265
266 static void __init srmmu_nocache_init(void)
267 {
268         void *srmmu_nocache_bitmap;
269         unsigned int bitmap_bits;
270         pgd_t *pgd;
271         p4d_t *p4d;
272         pud_t *pud;
273         pmd_t *pmd;
274         pte_t *pte;
275         unsigned long paddr, vaddr;
276         unsigned long pteval;
277
278         bitmap_bits = srmmu_nocache_size >> SRMMU_NOCACHE_BITMAP_SHIFT;
279
280         srmmu_nocache_pool = memblock_alloc_or_panic(srmmu_nocache_size,
281                                             SRMMU_NOCACHE_ALIGN_MAX);
282         memset(srmmu_nocache_pool, 0, srmmu_nocache_size);
283
284         srmmu_nocache_bitmap =
285                 memblock_alloc_or_panic(BITS_TO_LONGS(bitmap_bits) * sizeof(long),
286                                SMP_CACHE_BYTES);
287         bit_map_init(&srmmu_nocache_map, srmmu_nocache_bitmap, bitmap_bits);
288
289         srmmu_swapper_pg_dir = __srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
290         memset(__nocache_fix(srmmu_swapper_pg_dir), 0, SRMMU_PGD_TABLE_SIZE);
291         init_mm.pgd = srmmu_swapper_pg_dir;
292
293         srmmu_early_allocate_ptable_skeleton(SRMMU_NOCACHE_VADDR, srmmu_nocache_end);
294
295         paddr = __pa((unsigned long)srmmu_nocache_pool);
296         vaddr = SRMMU_NOCACHE_VADDR;
297
298         while (vaddr < srmmu_nocache_end) {
299                 pgd = pgd_offset_k(vaddr);
300                 p4d = p4d_offset(pgd, vaddr);
301                 pud = pud_offset(p4d, vaddr);
302                 pmd = pmd_offset(__nocache_fix(pud), vaddr);
303                 pte = pte_offset_kernel(__nocache_fix(pmd), vaddr);
304
305                 pteval = ((paddr >> 4) | SRMMU_ET_PTE | SRMMU_PRIV);
306
307                 if (srmmu_cache_pagetables)
308                         pteval |= SRMMU_CACHE;
309
310                 set_pte(__nocache_fix(pte), __pte(pteval));
311
312                 vaddr += PAGE_SIZE;
313                 paddr += PAGE_SIZE;
314         }
315
316         flush_cache_all();
317         flush_tlb_all();
318 }
319
320 pgd_t *get_pgd_fast(void)
321 {
322         pgd_t *pgd = NULL;
323
324         pgd = __srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
325         if (pgd) {
326                 pgd_t *init = pgd_offset_k(0);
327                 memset(pgd, 0, USER_PTRS_PER_PGD * sizeof(pgd_t));
328                 memcpy(pgd + USER_PTRS_PER_PGD, init + USER_PTRS_PER_PGD,
329                                                 (PTRS_PER_PGD - USER_PTRS_PER_PGD) * sizeof(pgd_t));
330         }
331
332         return pgd;
333 }
334
335 /*
336  * Hardware needs alignment to 256 only, but we align to whole page size
337  * to reduce fragmentation problems due to the buddy principle.
338  * XXX Provide actual fragmentation statistics in /proc.
339  *
340  * Alignments up to the page size are the same for physical and virtual
341  * addresses of the nocache area.
342  */
343 pgtable_t pte_alloc_one(struct mm_struct *mm)
344 {
345         pte_t *ptep;
346         struct page *page;
347
348         if (!(ptep = pte_alloc_one_kernel(mm)))
349                 return NULL;
350         page = pfn_to_page(__nocache_pa((unsigned long)ptep) >> PAGE_SHIFT);
351         spin_lock(&mm->page_table_lock);
352         if (page_ref_inc_return(page) == 2 &&
353                         !pagetable_pte_ctor(mm, page_ptdesc(page))) {
354                 page_ref_dec(page);
355                 ptep = NULL;
356         }
357         spin_unlock(&mm->page_table_lock);
358
359         return ptep;
360 }
361
362 void pte_free(struct mm_struct *mm, pgtable_t ptep)
363 {
364         struct page *page;
365
366         page = pfn_to_page(__nocache_pa((unsigned long)ptep) >> PAGE_SHIFT);
367         spin_lock(&mm->page_table_lock);
368         if (page_ref_dec_return(page) == 1)
369                 pagetable_dtor(page_ptdesc(page));
370         spin_unlock(&mm->page_table_lock);
371
372         srmmu_free_nocache(ptep, SRMMU_PTE_TABLE_SIZE);
373 }
374
375 /* context handling - a dynamically sized pool is used */
376 #define NO_CONTEXT      -1
377
378 struct ctx_list {
379         struct ctx_list *next;
380         struct ctx_list *prev;
381         unsigned int ctx_number;
382         struct mm_struct *ctx_mm;
383 };
384
385 static struct ctx_list *ctx_list_pool;
386 static struct ctx_list ctx_free;
387 static struct ctx_list ctx_used;
388
389 /* At boot time we determine the number of contexts */
390 static int num_contexts;
391
392 static inline void remove_from_ctx_list(struct ctx_list *entry)
393 {
394         entry->next->prev = entry->prev;
395         entry->prev->next = entry->next;
396 }
397
398 static inline void add_to_ctx_list(struct ctx_list *head, struct ctx_list *entry)
399 {
400         entry->next = head;
401         (entry->prev = head->prev)->next = entry;
402         head->prev = entry;
403 }
404 #define add_to_free_ctxlist(entry) add_to_ctx_list(&ctx_free, entry)
405 #define add_to_used_ctxlist(entry) add_to_ctx_list(&ctx_used, entry)
406
407
408 static inline void alloc_context(struct mm_struct *old_mm, struct mm_struct *mm)
409 {
410         struct ctx_list *ctxp;
411
412         ctxp = ctx_free.next;
413         if (ctxp != &ctx_free) {
414                 remove_from_ctx_list(ctxp);
415                 add_to_used_ctxlist(ctxp);
416                 mm->context = ctxp->ctx_number;
417                 ctxp->ctx_mm = mm;
418                 return;
419         }
420         ctxp = ctx_used.next;
421         if (ctxp->ctx_mm == old_mm)
422                 ctxp = ctxp->next;
423         if (ctxp == &ctx_used)
424                 panic("out of mmu contexts");
425         flush_cache_mm(ctxp->ctx_mm);
426         flush_tlb_mm(ctxp->ctx_mm);
427         remove_from_ctx_list(ctxp);
428         add_to_used_ctxlist(ctxp);
429         ctxp->ctx_mm->context = NO_CONTEXT;
430         ctxp->ctx_mm = mm;
431         mm->context = ctxp->ctx_number;
432 }
433
434 static inline void free_context(int context)
435 {
436         struct ctx_list *ctx_old;
437
438         ctx_old = ctx_list_pool + context;
439         remove_from_ctx_list(ctx_old);
440         add_to_free_ctxlist(ctx_old);
441 }
442
443 static void __init sparc_context_init(int numctx)
444 {
445         int ctx;
446         unsigned long size;
447
448         size = numctx * sizeof(struct ctx_list);
449         ctx_list_pool = memblock_alloc_or_panic(size, SMP_CACHE_BYTES);
450
451         for (ctx = 0; ctx < numctx; ctx++) {
452                 struct ctx_list *clist;
453
454                 clist = (ctx_list_pool + ctx);
455                 clist->ctx_number = ctx;
456                 clist->ctx_mm = NULL;
457         }
458         ctx_free.next = ctx_free.prev = &ctx_free;
459         ctx_used.next = ctx_used.prev = &ctx_used;
460         for (ctx = 0; ctx < numctx; ctx++)
461                 add_to_free_ctxlist(ctx_list_pool + ctx);
462 }
463
464 void switch_mm(struct mm_struct *old_mm, struct mm_struct *mm,
465                struct task_struct *tsk)
466 {
467         unsigned long flags;
468
469         if (mm->context == NO_CONTEXT) {
470                 spin_lock_irqsave(&srmmu_context_spinlock, flags);
471                 alloc_context(old_mm, mm);
472                 spin_unlock_irqrestore(&srmmu_context_spinlock, flags);
473                 srmmu_ctxd_set(&srmmu_context_table[mm->context], mm->pgd);
474         }
475
476         if (sparc_cpu_model == sparc_leon)
477                 leon_switch_mm();
478
479         if (is_hypersparc)
480                 hyper_flush_whole_icache();
481
482         srmmu_set_context(mm->context);
483 }
484
485 /* Low level IO area allocation on the SRMMU. */
486 static inline void srmmu_mapioaddr(unsigned long physaddr,
487                                    unsigned long virt_addr, int bus_type)
488 {
489         pgd_t *pgdp;
490         p4d_t *p4dp;
491         pud_t *pudp;
492         pmd_t *pmdp;
493         pte_t *ptep;
494         unsigned long tmp;
495
496         physaddr &= PAGE_MASK;
497         pgdp = pgd_offset_k(virt_addr);
498         p4dp = p4d_offset(pgdp, virt_addr);
499         pudp = pud_offset(p4dp, virt_addr);
500         pmdp = pmd_offset(pudp, virt_addr);
501         ptep = pte_offset_kernel(pmdp, virt_addr);
502         tmp = (physaddr >> 4) | SRMMU_ET_PTE;
503
504         /* I need to test whether this is consistent over all
505          * sun4m's.  The bus_type represents the upper 4 bits of
506          * 36-bit physical address on the I/O space lines...
507          */
508         tmp |= (bus_type << 28);
509         tmp |= SRMMU_PRIV;
510         __flush_page_to_ram(virt_addr);
511         set_pte(ptep, __pte(tmp));
512 }
513
514 void srmmu_mapiorange(unsigned int bus, unsigned long xpa,
515                       unsigned long xva, unsigned int len)
516 {
517         while (len != 0) {
518                 len -= PAGE_SIZE;
519                 srmmu_mapioaddr(xpa, xva, bus);
520                 xva += PAGE_SIZE;
521                 xpa += PAGE_SIZE;
522         }
523         flush_tlb_all();
524 }
525
526 static inline void srmmu_unmapioaddr(unsigned long virt_addr)
527 {
528         pgd_t *pgdp;
529         p4d_t *p4dp;
530         pud_t *pudp;
531         pmd_t *pmdp;
532         pte_t *ptep;
533
534
535         pgdp = pgd_offset_k(virt_addr);
536         p4dp = p4d_offset(pgdp, virt_addr);
537         pudp = pud_offset(p4dp, virt_addr);
538         pmdp = pmd_offset(pudp, virt_addr);
539         ptep = pte_offset_kernel(pmdp, virt_addr);
540
541         /* No need to flush uncacheable page. */
542         __pte_clear(ptep);
543 }
544
545 void srmmu_unmapiorange(unsigned long virt_addr, unsigned int len)
546 {
547         while (len != 0) {
548                 len -= PAGE_SIZE;
549                 srmmu_unmapioaddr(virt_addr);
550                 virt_addr += PAGE_SIZE;
551         }
552         flush_tlb_all();
553 }
554
555 /* tsunami.S */
556 extern void tsunami_flush_cache_all(void);
557 extern void tsunami_flush_cache_mm(struct mm_struct *mm);
558 extern void tsunami_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
559 extern void tsunami_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
560 extern void tsunami_flush_page_to_ram(unsigned long page);
561 extern void tsunami_flush_page_for_dma(unsigned long page);
562 extern void tsunami_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
563 extern void tsunami_flush_tlb_all(void);
564 extern void tsunami_flush_tlb_mm(struct mm_struct *mm);
565 extern void tsunami_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
566 extern void tsunami_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
567 extern void tsunami_setup_blockops(void);
568
569 /* swift.S */
570 extern void swift_flush_cache_all(void);
571 extern void swift_flush_cache_mm(struct mm_struct *mm);
572 extern void swift_flush_cache_range(struct vm_area_struct *vma,
573                                     unsigned long start, unsigned long end);
574 extern void swift_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
575 extern void swift_flush_page_to_ram(unsigned long page);
576 extern void swift_flush_page_for_dma(unsigned long page);
577 extern void swift_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
578 extern void swift_flush_tlb_all(void);
579 extern void swift_flush_tlb_mm(struct mm_struct *mm);
580 extern void swift_flush_tlb_range(struct vm_area_struct *vma,
581                                   unsigned long start, unsigned long end);
582 extern void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
583
584 #if 0  /* P3: deadwood to debug precise flushes on Swift. */
585 void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
586 {
587         int cctx, ctx1;
588
589         page &= PAGE_MASK;
590         if ((ctx1 = vma->vm_mm->context) != -1) {
591                 cctx = srmmu_get_context();
592 /* Is context # ever different from current context? P3 */
593                 if (cctx != ctx1) {
594                         printk("flush ctx %02x curr %02x\n", ctx1, cctx);
595                         srmmu_set_context(ctx1);
596                         swift_flush_page(page);
597                         __asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
598                                         "r" (page), "i" (ASI_M_FLUSH_PROBE));
599                         srmmu_set_context(cctx);
600                 } else {
601                          /* Rm. prot. bits from virt. c. */
602                         /* swift_flush_cache_all(); */
603                         /* swift_flush_cache_page(vma, page); */
604                         swift_flush_page(page);
605
606                         __asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
607                                 "r" (page), "i" (ASI_M_FLUSH_PROBE));
608                         /* same as above: srmmu_flush_tlb_page() */
609                 }
610         }
611 }
612 #endif
613
614 /*
615  * The following are all MBUS based SRMMU modules, and therefore could
616  * be found in a multiprocessor configuration.  On the whole, these
617  * chips seems to be much more touchy about DVMA and page tables
618  * with respect to cache coherency.
619  */
620
621 /* viking.S */
622 extern void viking_flush_cache_all(void);
623 extern void viking_flush_cache_mm(struct mm_struct *mm);
624 extern void viking_flush_cache_range(struct vm_area_struct *vma, unsigned long start,
625                                      unsigned long end);
626 extern void viking_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
627 extern void viking_flush_page_to_ram(unsigned long page);
628 extern void viking_flush_page_for_dma(unsigned long page);
629 extern void viking_flush_sig_insns(struct mm_struct *mm, unsigned long addr);
630 extern void viking_flush_page(unsigned long page);
631 extern void viking_mxcc_flush_page(unsigned long page);
632 extern void viking_flush_tlb_all(void);
633 extern void viking_flush_tlb_mm(struct mm_struct *mm);
634 extern void viking_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
635                                    unsigned long end);
636 extern void viking_flush_tlb_page(struct vm_area_struct *vma,
637                                   unsigned long page);
638 extern void sun4dsmp_flush_tlb_all(void);
639 extern void sun4dsmp_flush_tlb_mm(struct mm_struct *mm);
640 extern void sun4dsmp_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
641                                    unsigned long end);
642 extern void sun4dsmp_flush_tlb_page(struct vm_area_struct *vma,
643                                   unsigned long page);
644
645 /* hypersparc.S */
646 extern void hypersparc_flush_cache_all(void);
647 extern void hypersparc_flush_cache_mm(struct mm_struct *mm);
648 extern void hypersparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
649 extern void hypersparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
650 extern void hypersparc_flush_page_to_ram(unsigned long page);
651 extern void hypersparc_flush_page_for_dma(unsigned long page);
652 extern void hypersparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
653 extern void hypersparc_flush_tlb_all(void);
654 extern void hypersparc_flush_tlb_mm(struct mm_struct *mm);
655 extern void hypersparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
656 extern void hypersparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
657 extern void hypersparc_setup_blockops(void);
658
659 /*
660  * NOTE: All of this startup code assumes the low 16mb (approx.) of
661  *       kernel mappings are done with one single contiguous chunk of
662  *       ram.  On small ram machines (classics mainly) we only get
663  *       around 8mb mapped for us.
664  */
665
666 static void __init early_pgtable_allocfail(char *type)
667 {
668         prom_printf("inherit_prom_mappings: Cannot alloc kernel %s.\n", type);
669         prom_halt();
670 }
671
672 static void __init srmmu_early_allocate_ptable_skeleton(unsigned long start,
673                                                         unsigned long end)
674 {
675         pgd_t *pgdp;
676         p4d_t *p4dp;
677         pud_t *pudp;
678         pmd_t *pmdp;
679         pte_t *ptep;
680
681         while (start < end) {
682                 pgdp = pgd_offset_k(start);
683                 p4dp = p4d_offset(pgdp, start);
684                 pudp = pud_offset(p4dp, start);
685                 if (pud_none(*__nocache_fix(pudp))) {
686                         pmdp = __srmmu_get_nocache(
687                             SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
688                         if (pmdp == NULL)
689                                 early_pgtable_allocfail("pmd");
690                         memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
691                         pud_set(__nocache_fix(pudp), pmdp);
692                 }
693                 pmdp = pmd_offset(__nocache_fix(pudp), start);
694                 if (srmmu_pmd_none(*__nocache_fix(pmdp))) {
695                         ptep = __srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
696                         if (ptep == NULL)
697                                 early_pgtable_allocfail("pte");
698                         memset(__nocache_fix(ptep), 0, PTE_SIZE);
699                         pmd_set(__nocache_fix(pmdp), ptep);
700                 }
701                 if (start > (0xffffffffUL - PMD_SIZE))
702                         break;
703                 start = (start + PMD_SIZE) & PMD_MASK;
704         }
705 }
706
707 static void __init srmmu_allocate_ptable_skeleton(unsigned long start,
708                                                   unsigned long end)
709 {
710         pgd_t *pgdp;
711         p4d_t *p4dp;
712         pud_t *pudp;
713         pmd_t *pmdp;
714         pte_t *ptep;
715
716         while (start < end) {
717                 pgdp = pgd_offset_k(start);
718                 p4dp = p4d_offset(pgdp, start);
719                 pudp = pud_offset(p4dp, start);
720                 if (pud_none(*pudp)) {
721                         pmdp = __srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
722                         if (pmdp == NULL)
723                                 early_pgtable_allocfail("pmd");
724                         memset(pmdp, 0, SRMMU_PMD_TABLE_SIZE);
725                         pud_set((pud_t *)pgdp, pmdp);
726                 }
727                 pmdp = pmd_offset(pudp, start);
728                 if (srmmu_pmd_none(*pmdp)) {
729                         ptep = __srmmu_get_nocache(PTE_SIZE,
730                                                              PTE_SIZE);
731                         if (ptep == NULL)
732                                 early_pgtable_allocfail("pte");
733                         memset(ptep, 0, PTE_SIZE);
734                         pmd_set(pmdp, ptep);
735                 }
736                 if (start > (0xffffffffUL - PMD_SIZE))
737                         break;
738                 start = (start + PMD_SIZE) & PMD_MASK;
739         }
740 }
741
742 /* These flush types are not available on all chips... */
743 static inline unsigned long srmmu_probe(unsigned long vaddr)
744 {
745         unsigned long retval;
746
747         if (sparc_cpu_model != sparc_leon) {
748
749                 vaddr &= PAGE_MASK;
750                 __asm__ __volatile__("lda [%1] %2, %0\n\t" :
751                                      "=r" (retval) :
752                                      "r" (vaddr | 0x400), "i" (ASI_M_FLUSH_PROBE));
753         } else {
754                 retval = leon_swprobe(vaddr, NULL);
755         }
756         return retval;
757 }
758
759 /*
760  * This is much cleaner than poking around physical address space
761  * looking at the prom's page table directly which is what most
762  * other OS's do.  Yuck... this is much better.
763  */
764 static void __init srmmu_inherit_prom_mappings(unsigned long start,
765                                                unsigned long end)
766 {
767         unsigned long probed;
768         unsigned long addr;
769         pgd_t *pgdp;
770         p4d_t *p4dp;
771         pud_t *pudp;
772         pmd_t *pmdp;
773         pte_t *ptep;
774         int what; /* 0 = normal-pte, 1 = pmd-level pte, 2 = pgd-level pte */
775
776         while (start <= end) {
777                 if (start == 0)
778                         break; /* probably wrap around */
779                 if (start == 0xfef00000)
780                         start = KADB_DEBUGGER_BEGVM;
781                 probed = srmmu_probe(start);
782                 if (!probed) {
783                         /* continue probing until we find an entry */
784                         start += PAGE_SIZE;
785                         continue;
786                 }
787
788                 /* A red snapper, see what it really is. */
789                 what = 0;
790                 addr = start - PAGE_SIZE;
791
792                 if (!(start & ~(PMD_MASK))) {
793                         if (srmmu_probe(addr + PMD_SIZE) == probed)
794                                 what = 1;
795                 }
796
797                 if (!(start & ~(PGDIR_MASK))) {
798                         if (srmmu_probe(addr + PGDIR_SIZE) == probed)
799                                 what = 2;
800                 }
801
802                 pgdp = pgd_offset_k(start);
803                 p4dp = p4d_offset(pgdp, start);
804                 pudp = pud_offset(p4dp, start);
805                 if (what == 2) {
806                         *__nocache_fix(pgdp) = __pgd(probed);
807                         start += PGDIR_SIZE;
808                         continue;
809                 }
810                 if (pud_none(*__nocache_fix(pudp))) {
811                         pmdp = __srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE,
812                                                    SRMMU_PMD_TABLE_SIZE);
813                         if (pmdp == NULL)
814                                 early_pgtable_allocfail("pmd");
815                         memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
816                         pud_set(__nocache_fix(pudp), pmdp);
817                 }
818                 pmdp = pmd_offset(__nocache_fix(pudp), start);
819                 if (what == 1) {
820                         *(pmd_t *)__nocache_fix(pmdp) = __pmd(probed);
821                         start += PMD_SIZE;
822                         continue;
823                 }
824                 if (srmmu_pmd_none(*__nocache_fix(pmdp))) {
825                         ptep = __srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
826                         if (ptep == NULL)
827                                 early_pgtable_allocfail("pte");
828                         memset(__nocache_fix(ptep), 0, PTE_SIZE);
829                         pmd_set(__nocache_fix(pmdp), ptep);
830                 }
831                 ptep = pte_offset_kernel(__nocache_fix(pmdp), start);
832                 *__nocache_fix(ptep) = __pte(probed);
833                 start += PAGE_SIZE;
834         }
835 }
836
837 #define KERNEL_PTE(page_shifted) ((page_shifted)|SRMMU_CACHE|SRMMU_PRIV|SRMMU_VALID)
838
839 /* Create a third-level SRMMU 16MB page mapping. */
840 static void __init do_large_mapping(unsigned long vaddr, unsigned long phys_base)
841 {
842         pgd_t *pgdp = pgd_offset_k(vaddr);
843         unsigned long big_pte;
844
845         big_pte = KERNEL_PTE(phys_base >> 4);
846         *__nocache_fix(pgdp) = __pgd(big_pte);
847 }
848
849 /* Map sp_bank entry SP_ENTRY, starting at virtual address VBASE. */
850 static unsigned long __init map_spbank(unsigned long vbase, int sp_entry)
851 {
852         unsigned long pstart = (sp_banks[sp_entry].base_addr & PGDIR_MASK);
853         unsigned long vstart = (vbase & PGDIR_MASK);
854         unsigned long vend = PGDIR_ALIGN(vbase + sp_banks[sp_entry].num_bytes);
855         /* Map "low" memory only */
856         const unsigned long min_vaddr = PAGE_OFFSET;
857         const unsigned long max_vaddr = PAGE_OFFSET + SRMMU_MAXMEM;
858
859         if (vstart < min_vaddr || vstart >= max_vaddr)
860                 return vstart;
861
862         if (vend > max_vaddr || vend < min_vaddr)
863                 vend = max_vaddr;
864
865         while (vstart < vend) {
866                 do_large_mapping(vstart, pstart);
867                 vstart += PGDIR_SIZE; pstart += PGDIR_SIZE;
868         }
869         return vstart;
870 }
871
872 static void __init map_kernel(void)
873 {
874         int i;
875
876         if (phys_base > 0) {
877                 do_large_mapping(PAGE_OFFSET, phys_base);
878         }
879
880         for (i = 0; sp_banks[i].num_bytes != 0; i++) {
881                 map_spbank((unsigned long)__va(sp_banks[i].base_addr), i);
882         }
883 }
884
885 void (*poke_srmmu)(void) = NULL;
886
887 void __init srmmu_paging_init(void)
888 {
889         int i;
890         phandle cpunode;
891         char node_str[128];
892         pgd_t *pgd;
893         p4d_t *p4d;
894         pud_t *pud;
895         pmd_t *pmd;
896         pte_t *pte;
897         unsigned long pages_avail;
898
899         init_mm.context = (unsigned long) NO_CONTEXT;
900         sparc_iomap.start = SUN4M_IOBASE_VADDR; /* 16MB of IOSPACE on all sun4m's. */
901
902         if (sparc_cpu_model == sun4d)
903                 num_contexts = 65536; /* We know it is Viking */
904         else {
905                 /* Find the number of contexts on the srmmu. */
906                 cpunode = prom_getchild(prom_root_node);
907                 num_contexts = 0;
908                 while (cpunode != 0) {
909                         prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
910                         if (!strcmp(node_str, "cpu")) {
911                                 num_contexts = prom_getintdefault(cpunode, "mmu-nctx", 0x8);
912                                 break;
913                         }
914                         cpunode = prom_getsibling(cpunode);
915                 }
916         }
917
918         if (!num_contexts) {
919                 prom_printf("Something wrong, can't find cpu node in paging_init.\n");
920                 prom_halt();
921         }
922
923         pages_avail = 0;
924         last_valid_pfn = bootmem_init(&pages_avail);
925
926         srmmu_nocache_calcsize();
927         srmmu_nocache_init();
928         srmmu_inherit_prom_mappings(0xfe400000, (LINUX_OPPROM_ENDVM - PAGE_SIZE));
929         map_kernel();
930
931         /* ctx table has to be physically aligned to its size */
932         srmmu_context_table = __srmmu_get_nocache(num_contexts * sizeof(ctxd_t), num_contexts * sizeof(ctxd_t));
933         srmmu_ctx_table_phys = (ctxd_t *)__nocache_pa(srmmu_context_table);
934
935         for (i = 0; i < num_contexts; i++)
936                 srmmu_ctxd_set(__nocache_fix(&srmmu_context_table[i]), srmmu_swapper_pg_dir);
937
938         flush_cache_all();
939         srmmu_set_ctable_ptr((unsigned long)srmmu_ctx_table_phys);
940 #ifdef CONFIG_SMP
941         /* Stop from hanging here... */
942         local_ops->tlb_all();
943 #else
944         flush_tlb_all();
945 #endif
946         poke_srmmu();
947
948         srmmu_allocate_ptable_skeleton(sparc_iomap.start, IOBASE_END);
949         srmmu_allocate_ptable_skeleton(DVMA_VADDR, DVMA_END);
950
951         srmmu_allocate_ptable_skeleton(
952                 __fix_to_virt(__end_of_fixed_addresses - 1), FIXADDR_TOP);
953         srmmu_allocate_ptable_skeleton(PKMAP_BASE, PKMAP_END);
954
955         pgd = pgd_offset_k(PKMAP_BASE);
956         p4d = p4d_offset(pgd, PKMAP_BASE);
957         pud = pud_offset(p4d, PKMAP_BASE);
958         pmd = pmd_offset(pud, PKMAP_BASE);
959         pte = pte_offset_kernel(pmd, PKMAP_BASE);
960         pkmap_page_table = pte;
961
962         flush_cache_all();
963         flush_tlb_all();
964
965         sparc_context_init(num_contexts);
966
967         {
968                 unsigned long max_zone_pfn[MAX_NR_ZONES] = { 0 };
969
970                 max_zone_pfn[ZONE_DMA] = max_low_pfn;
971                 max_zone_pfn[ZONE_NORMAL] = max_low_pfn;
972                 max_zone_pfn[ZONE_HIGHMEM] = highend_pfn;
973
974                 free_area_init(max_zone_pfn);
975         }
976 }
977
978 void mmu_info(struct seq_file *m)
979 {
980         seq_printf(m,
981                    "MMU type\t: %s\n"
982                    "contexts\t: %d\n"
983                    "nocache total\t: %ld\n"
984                    "nocache used\t: %d\n",
985                    srmmu_name,
986                    num_contexts,
987                    srmmu_nocache_size,
988                    srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
989 }
990
991 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
992 {
993         mm->context = NO_CONTEXT;
994         return 0;
995 }
996
997 void destroy_context(struct mm_struct *mm)
998 {
999         unsigned long flags;
1000
1001         if (mm->context != NO_CONTEXT) {
1002                 flush_cache_mm(mm);
1003                 srmmu_ctxd_set(&srmmu_context_table[mm->context], srmmu_swapper_pg_dir);
1004                 flush_tlb_mm(mm);
1005                 spin_lock_irqsave(&srmmu_context_spinlock, flags);
1006                 free_context(mm->context);
1007                 spin_unlock_irqrestore(&srmmu_context_spinlock, flags);
1008                 mm->context = NO_CONTEXT;
1009         }
1010 }
1011
1012 /* Init various srmmu chip types. */
1013 static void __init srmmu_is_bad(void)
1014 {
1015         prom_printf("Could not determine SRMMU chip type.\n");
1016         prom_halt();
1017 }
1018
1019 static void __init init_vac_layout(void)
1020 {
1021         phandle nd;
1022         int cache_lines;
1023         char node_str[128];
1024 #ifdef CONFIG_SMP
1025         int cpu = 0;
1026         unsigned long max_size = 0;
1027         unsigned long min_line_size = 0x10000000;
1028 #endif
1029
1030         nd = prom_getchild(prom_root_node);
1031         while ((nd = prom_getsibling(nd)) != 0) {
1032                 prom_getstring(nd, "device_type", node_str, sizeof(node_str));
1033                 if (!strcmp(node_str, "cpu")) {
1034                         vac_line_size = prom_getint(nd, "cache-line-size");
1035                         if (vac_line_size == -1) {
1036                                 prom_printf("can't determine cache-line-size, halting.\n");
1037                                 prom_halt();
1038                         }
1039                         cache_lines = prom_getint(nd, "cache-nlines");
1040                         if (cache_lines == -1) {
1041                                 prom_printf("can't determine cache-nlines, halting.\n");
1042                                 prom_halt();
1043                         }
1044
1045                         vac_cache_size = cache_lines * vac_line_size;
1046 #ifdef CONFIG_SMP
1047                         if (vac_cache_size > max_size)
1048                                 max_size = vac_cache_size;
1049                         if (vac_line_size < min_line_size)
1050                                 min_line_size = vac_line_size;
1051                         //FIXME: cpus not contiguous!!
1052                         cpu++;
1053                         if (cpu >= nr_cpu_ids || !cpu_online(cpu))
1054                                 break;
1055 #else
1056                         break;
1057 #endif
1058                 }
1059         }
1060         if (nd == 0) {
1061                 prom_printf("No CPU nodes found, halting.\n");
1062                 prom_halt();
1063         }
1064 #ifdef CONFIG_SMP
1065         vac_cache_size = max_size;
1066         vac_line_size = min_line_size;
1067 #endif
1068         printk("SRMMU: Using VAC size of %d bytes, line size %d bytes.\n",
1069                (int)vac_cache_size, (int)vac_line_size);
1070 }
1071
1072 static void poke_hypersparc(void)
1073 {
1074         volatile unsigned long clear;
1075         unsigned long mreg = srmmu_get_mmureg();
1076
1077         hyper_flush_unconditional_combined();
1078
1079         mreg &= ~(HYPERSPARC_CWENABLE);
1080         mreg |= (HYPERSPARC_CENABLE | HYPERSPARC_WBENABLE);
1081         mreg |= (HYPERSPARC_CMODE);
1082
1083         srmmu_set_mmureg(mreg);
1084
1085 #if 0 /* XXX I think this is bad news... -DaveM */
1086         hyper_clear_all_tags();
1087 #endif
1088
1089         put_ross_icr(HYPERSPARC_ICCR_FTD | HYPERSPARC_ICCR_ICE);
1090         hyper_flush_whole_icache();
1091         clear = srmmu_get_faddr();
1092         clear = srmmu_get_fstatus();
1093 }
1094
1095 static const struct sparc32_cachetlb_ops hypersparc_ops = {
1096         .cache_all      = hypersparc_flush_cache_all,
1097         .cache_mm       = hypersparc_flush_cache_mm,
1098         .cache_page     = hypersparc_flush_cache_page,
1099         .cache_range    = hypersparc_flush_cache_range,
1100         .tlb_all        = hypersparc_flush_tlb_all,
1101         .tlb_mm         = hypersparc_flush_tlb_mm,
1102         .tlb_page       = hypersparc_flush_tlb_page,
1103         .tlb_range      = hypersparc_flush_tlb_range,
1104         .page_to_ram    = hypersparc_flush_page_to_ram,
1105         .sig_insns      = hypersparc_flush_sig_insns,
1106         .page_for_dma   = hypersparc_flush_page_for_dma,
1107 };
1108
1109 static void __init init_hypersparc(void)
1110 {
1111         srmmu_name = "ROSS HyperSparc";
1112         srmmu_modtype = HyperSparc;
1113
1114         init_vac_layout();
1115
1116         is_hypersparc = 1;
1117         sparc32_cachetlb_ops = &hypersparc_ops;
1118
1119         poke_srmmu = poke_hypersparc;
1120
1121         hypersparc_setup_blockops();
1122 }
1123
1124 static void poke_swift(void)
1125 {
1126         unsigned long mreg;
1127
1128         /* Clear any crap from the cache or else... */
1129         swift_flush_cache_all();
1130
1131         /* Enable I & D caches */
1132         mreg = srmmu_get_mmureg();
1133         mreg |= (SWIFT_IE | SWIFT_DE);
1134         /*
1135          * The Swift branch folding logic is completely broken.  At
1136          * trap time, if things are just right, if can mistakenly
1137          * think that a trap is coming from kernel mode when in fact
1138          * it is coming from user mode (it mis-executes the branch in
1139          * the trap code).  So you see things like crashme completely
1140          * hosing your machine which is completely unacceptable.  Turn
1141          * this shit off... nice job Fujitsu.
1142          */
1143         mreg &= ~(SWIFT_BF);
1144         srmmu_set_mmureg(mreg);
1145 }
1146
1147 static const struct sparc32_cachetlb_ops swift_ops = {
1148         .cache_all      = swift_flush_cache_all,
1149         .cache_mm       = swift_flush_cache_mm,
1150         .cache_page     = swift_flush_cache_page,
1151         .cache_range    = swift_flush_cache_range,
1152         .tlb_all        = swift_flush_tlb_all,
1153         .tlb_mm         = swift_flush_tlb_mm,
1154         .tlb_page       = swift_flush_tlb_page,
1155         .tlb_range      = swift_flush_tlb_range,
1156         .page_to_ram    = swift_flush_page_to_ram,
1157         .sig_insns      = swift_flush_sig_insns,
1158         .page_for_dma   = swift_flush_page_for_dma,
1159 };
1160
1161 #define SWIFT_MASKID_ADDR  0x10003018
1162 static void __init init_swift(void)
1163 {
1164         unsigned long swift_rev;
1165
1166         __asm__ __volatile__("lda [%1] %2, %0\n\t"
1167                              "srl %0, 0x18, %0\n\t" :
1168                              "=r" (swift_rev) :
1169                              "r" (SWIFT_MASKID_ADDR), "i" (ASI_M_BYPASS));
1170         srmmu_name = "Fujitsu Swift";
1171         switch (swift_rev) {
1172         case 0x11:
1173         case 0x20:
1174         case 0x23:
1175         case 0x30:
1176                 srmmu_modtype = Swift_lots_o_bugs;
1177                 hwbug_bitmask |= (HWBUG_KERN_ACCBROKEN | HWBUG_KERN_CBITBROKEN);
1178                 /*
1179                  * Gee george, I wonder why Sun is so hush hush about
1180                  * this hardware bug... really braindamage stuff going
1181                  * on here.  However I think we can find a way to avoid
1182                  * all of the workaround overhead under Linux.  Basically,
1183                  * any page fault can cause kernel pages to become user
1184                  * accessible (the mmu gets confused and clears some of
1185                  * the ACC bits in kernel ptes).  Aha, sounds pretty
1186                  * horrible eh?  But wait, after extensive testing it appears
1187                  * that if you use pgd_t level large kernel pte's (like the
1188                  * 4MB pages on the Pentium) the bug does not get tripped
1189                  * at all.  This avoids almost all of the major overhead.
1190                  * Welcome to a world where your vendor tells you to,
1191                  * "apply this kernel patch" instead of "sorry for the
1192                  * broken hardware, send it back and we'll give you
1193                  * properly functioning parts"
1194                  */
1195                 break;
1196         case 0x25:
1197         case 0x31:
1198                 srmmu_modtype = Swift_bad_c;
1199                 hwbug_bitmask |= HWBUG_KERN_CBITBROKEN;
1200                 /*
1201                  * You see Sun allude to this hardware bug but never
1202                  * admit things directly, they'll say things like,
1203                  * "the Swift chip cache problems" or similar.
1204                  */
1205                 break;
1206         default:
1207                 srmmu_modtype = Swift_ok;
1208                 break;
1209         }
1210
1211         sparc32_cachetlb_ops = &swift_ops;
1212         flush_page_for_dma_global = 0;
1213
1214         /*
1215          * Are you now convinced that the Swift is one of the
1216          * biggest VLSI abortions of all time?  Bravo Fujitsu!
1217          * Fujitsu, the !#?!%$'d up processor people.  I bet if
1218          * you examined the microcode of the Swift you'd find
1219          * XXX's all over the place.
1220          */
1221         poke_srmmu = poke_swift;
1222 }
1223
1224 static void turbosparc_flush_cache_all(void)
1225 {
1226         flush_user_windows();
1227         turbosparc_idflash_clear();
1228 }
1229
1230 static void turbosparc_flush_cache_mm(struct mm_struct *mm)
1231 {
1232         FLUSH_BEGIN(mm)
1233         flush_user_windows();
1234         turbosparc_idflash_clear();
1235         FLUSH_END
1236 }
1237
1238 static void turbosparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1239 {
1240         FLUSH_BEGIN(vma->vm_mm)
1241         flush_user_windows();
1242         turbosparc_idflash_clear();
1243         FLUSH_END
1244 }
1245
1246 static void turbosparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1247 {
1248         FLUSH_BEGIN(vma->vm_mm)
1249         flush_user_windows();
1250         if (vma->vm_flags & VM_EXEC)
1251                 turbosparc_flush_icache();
1252         turbosparc_flush_dcache();
1253         FLUSH_END
1254 }
1255
1256 /* TurboSparc is copy-back, if we turn it on, but this does not work. */
1257 static void turbosparc_flush_page_to_ram(unsigned long page)
1258 {
1259 #ifdef TURBOSPARC_WRITEBACK
1260         volatile unsigned long clear;
1261
1262         if (srmmu_probe(page))
1263                 turbosparc_flush_page_cache(page);
1264         clear = srmmu_get_fstatus();
1265 #endif
1266 }
1267
1268 static void turbosparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1269 {
1270 }
1271
1272 static void turbosparc_flush_page_for_dma(unsigned long page)
1273 {
1274         turbosparc_flush_dcache();
1275 }
1276
1277 static void turbosparc_flush_tlb_all(void)
1278 {
1279         srmmu_flush_whole_tlb();
1280 }
1281
1282 static void turbosparc_flush_tlb_mm(struct mm_struct *mm)
1283 {
1284         FLUSH_BEGIN(mm)
1285         srmmu_flush_whole_tlb();
1286         FLUSH_END
1287 }
1288
1289 static void turbosparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1290 {
1291         FLUSH_BEGIN(vma->vm_mm)
1292         srmmu_flush_whole_tlb();
1293         FLUSH_END
1294 }
1295
1296 static void turbosparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1297 {
1298         FLUSH_BEGIN(vma->vm_mm)
1299         srmmu_flush_whole_tlb();
1300         FLUSH_END
1301 }
1302
1303
1304 static void poke_turbosparc(void)
1305 {
1306         unsigned long mreg = srmmu_get_mmureg();
1307         unsigned long ccreg;
1308
1309         /* Clear any crap from the cache or else... */
1310         turbosparc_flush_cache_all();
1311         /* Temporarily disable I & D caches */
1312         mreg &= ~(TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE);
1313         mreg &= ~(TURBOSPARC_PCENABLE);         /* Don't check parity */
1314         srmmu_set_mmureg(mreg);
1315
1316         ccreg = turbosparc_get_ccreg();
1317
1318 #ifdef TURBOSPARC_WRITEBACK
1319         ccreg |= (TURBOSPARC_SNENABLE);         /* Do DVMA snooping in Dcache */
1320         ccreg &= ~(TURBOSPARC_uS2 | TURBOSPARC_WTENABLE);
1321                         /* Write-back D-cache, emulate VLSI
1322                          * abortion number three, not number one */
1323 #else
1324         /* For now let's play safe, optimize later */
1325         ccreg |= (TURBOSPARC_SNENABLE | TURBOSPARC_WTENABLE);
1326                         /* Do DVMA snooping in Dcache, Write-thru D-cache */
1327         ccreg &= ~(TURBOSPARC_uS2);
1328                         /* Emulate VLSI abortion number three, not number one */
1329 #endif
1330
1331         switch (ccreg & 7) {
1332         case 0: /* No SE cache */
1333         case 7: /* Test mode */
1334                 break;
1335         default:
1336                 ccreg |= (TURBOSPARC_SCENABLE);
1337         }
1338         turbosparc_set_ccreg(ccreg);
1339
1340         mreg |= (TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE); /* I & D caches on */
1341         mreg |= (TURBOSPARC_ICSNOOP);           /* Icache snooping on */
1342         srmmu_set_mmureg(mreg);
1343 }
1344
1345 static const struct sparc32_cachetlb_ops turbosparc_ops = {
1346         .cache_all      = turbosparc_flush_cache_all,
1347         .cache_mm       = turbosparc_flush_cache_mm,
1348         .cache_page     = turbosparc_flush_cache_page,
1349         .cache_range    = turbosparc_flush_cache_range,
1350         .tlb_all        = turbosparc_flush_tlb_all,
1351         .tlb_mm         = turbosparc_flush_tlb_mm,
1352         .tlb_page       = turbosparc_flush_tlb_page,
1353         .tlb_range      = turbosparc_flush_tlb_range,
1354         .page_to_ram    = turbosparc_flush_page_to_ram,
1355         .sig_insns      = turbosparc_flush_sig_insns,
1356         .page_for_dma   = turbosparc_flush_page_for_dma,
1357 };
1358
1359 static void __init init_turbosparc(void)
1360 {
1361         srmmu_name = "Fujitsu TurboSparc";
1362         srmmu_modtype = TurboSparc;
1363         sparc32_cachetlb_ops = &turbosparc_ops;
1364         poke_srmmu = poke_turbosparc;
1365 }
1366
1367 static void poke_tsunami(void)
1368 {
1369         unsigned long mreg = srmmu_get_mmureg();
1370
1371         tsunami_flush_icache();
1372         tsunami_flush_dcache();
1373         mreg &= ~TSUNAMI_ITD;
1374         mreg |= (TSUNAMI_IENAB | TSUNAMI_DENAB);
1375         srmmu_set_mmureg(mreg);
1376 }
1377
1378 static const struct sparc32_cachetlb_ops tsunami_ops = {
1379         .cache_all      = tsunami_flush_cache_all,
1380         .cache_mm       = tsunami_flush_cache_mm,
1381         .cache_page     = tsunami_flush_cache_page,
1382         .cache_range    = tsunami_flush_cache_range,
1383         .tlb_all        = tsunami_flush_tlb_all,
1384         .tlb_mm         = tsunami_flush_tlb_mm,
1385         .tlb_page       = tsunami_flush_tlb_page,
1386         .tlb_range      = tsunami_flush_tlb_range,
1387         .page_to_ram    = tsunami_flush_page_to_ram,
1388         .sig_insns      = tsunami_flush_sig_insns,
1389         .page_for_dma   = tsunami_flush_page_for_dma,
1390 };
1391
1392 static void __init init_tsunami(void)
1393 {
1394         /*
1395          * Tsunami's pretty sane, Sun and TI actually got it
1396          * somewhat right this time.  Fujitsu should have
1397          * taken some lessons from them.
1398          */
1399
1400         srmmu_name = "TI Tsunami";
1401         srmmu_modtype = Tsunami;
1402         sparc32_cachetlb_ops = &tsunami_ops;
1403         poke_srmmu = poke_tsunami;
1404
1405         tsunami_setup_blockops();
1406 }
1407
1408 static void poke_viking(void)
1409 {
1410         unsigned long mreg = srmmu_get_mmureg();
1411         static int smp_catch;
1412
1413         if (viking_mxcc_present) {
1414                 unsigned long mxcc_control = mxcc_get_creg();
1415
1416                 mxcc_control |= (MXCC_CTL_ECE | MXCC_CTL_PRE | MXCC_CTL_MCE);
1417                 mxcc_control &= ~(MXCC_CTL_RRC);
1418                 mxcc_set_creg(mxcc_control);
1419
1420                 /*
1421                  * We don't need memory parity checks.
1422                  * XXX This is a mess, have to dig out later. ecd.
1423                 viking_mxcc_turn_off_parity(&mreg, &mxcc_control);
1424                  */
1425
1426                 /* We do cache ptables on MXCC. */
1427                 mreg |= VIKING_TCENABLE;
1428         } else {
1429                 unsigned long bpreg;
1430
1431                 mreg &= ~(VIKING_TCENABLE);
1432                 if (smp_catch++) {
1433                         /* Must disable mixed-cmd mode here for other cpu's. */
1434                         bpreg = viking_get_bpreg();
1435                         bpreg &= ~(VIKING_ACTION_MIX);
1436                         viking_set_bpreg(bpreg);
1437
1438                         /* Just in case PROM does something funny. */
1439                         msi_set_sync();
1440                 }
1441         }
1442
1443         mreg |= VIKING_SPENABLE;
1444         mreg |= (VIKING_ICENABLE | VIKING_DCENABLE);
1445         mreg |= VIKING_SBENABLE;
1446         mreg &= ~(VIKING_ACENABLE);
1447         srmmu_set_mmureg(mreg);
1448 }
1449
1450 static struct sparc32_cachetlb_ops viking_ops __ro_after_init = {
1451         .cache_all      = viking_flush_cache_all,
1452         .cache_mm       = viking_flush_cache_mm,
1453         .cache_page     = viking_flush_cache_page,
1454         .cache_range    = viking_flush_cache_range,
1455         .tlb_all        = viking_flush_tlb_all,
1456         .tlb_mm         = viking_flush_tlb_mm,
1457         .tlb_page       = viking_flush_tlb_page,
1458         .tlb_range      = viking_flush_tlb_range,
1459         .page_to_ram    = viking_flush_page_to_ram,
1460         .sig_insns      = viking_flush_sig_insns,
1461         .page_for_dma   = viking_flush_page_for_dma,
1462 };
1463
1464 #ifdef CONFIG_SMP
1465 /* On sun4d the cpu broadcasts local TLB flushes, so we can just
1466  * perform the local TLB flush and all the other cpus will see it.
1467  * But, unfortunately, there is a bug in the sun4d XBUS backplane
1468  * that requires that we add some synchronization to these flushes.
1469  *
1470  * The bug is that the fifo which keeps track of all the pending TLB
1471  * broadcasts in the system is an entry or two too small, so if we
1472  * have too many going at once we'll overflow that fifo and lose a TLB
1473  * flush resulting in corruption.
1474  *
1475  * Our workaround is to take a global spinlock around the TLB flushes,
1476  * which guarentees we won't ever have too many pending.  It's a big
1477  * hammer, but a semaphore like system to make sure we only have N TLB
1478  * flushes going at once will require SMP locking anyways so there's
1479  * no real value in trying any harder than this.
1480  */
1481 static struct sparc32_cachetlb_ops viking_sun4d_smp_ops __ro_after_init = {
1482         .cache_all      = viking_flush_cache_all,
1483         .cache_mm       = viking_flush_cache_mm,
1484         .cache_page     = viking_flush_cache_page,
1485         .cache_range    = viking_flush_cache_range,
1486         .tlb_all        = sun4dsmp_flush_tlb_all,
1487         .tlb_mm         = sun4dsmp_flush_tlb_mm,
1488         .tlb_page       = sun4dsmp_flush_tlb_page,
1489         .tlb_range      = sun4dsmp_flush_tlb_range,
1490         .page_to_ram    = viking_flush_page_to_ram,
1491         .sig_insns      = viking_flush_sig_insns,
1492         .page_for_dma   = viking_flush_page_for_dma,
1493 };
1494 #endif
1495
1496 static void __init init_viking(void)
1497 {
1498         unsigned long mreg = srmmu_get_mmureg();
1499
1500         /* Ahhh, the viking.  SRMMU VLSI abortion number two... */
1501         if (mreg & VIKING_MMODE) {
1502                 srmmu_name = "TI Viking";
1503                 viking_mxcc_present = 0;
1504                 msi_set_sync();
1505
1506                 /*
1507                  * We need this to make sure old viking takes no hits
1508                  * on its cache for dma snoops to workaround the
1509                  * "load from non-cacheable memory" interrupt bug.
1510                  * This is only necessary because of the new way in
1511                  * which we use the IOMMU.
1512                  */
1513                 viking_ops.page_for_dma = viking_flush_page;
1514 #ifdef CONFIG_SMP
1515                 viking_sun4d_smp_ops.page_for_dma = viking_flush_page;
1516 #endif
1517                 flush_page_for_dma_global = 0;
1518         } else {
1519                 srmmu_name = "TI Viking/MXCC";
1520                 viking_mxcc_present = 1;
1521                 srmmu_cache_pagetables = 1;
1522         }
1523
1524         sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1525                 &viking_ops;
1526 #ifdef CONFIG_SMP
1527         if (sparc_cpu_model == sun4d)
1528                 sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1529                         &viking_sun4d_smp_ops;
1530 #endif
1531
1532         poke_srmmu = poke_viking;
1533 }
1534
1535 /* Probe for the srmmu chip version. */
1536 static void __init get_srmmu_type(void)
1537 {
1538         unsigned long mreg, psr;
1539         unsigned long mod_typ, mod_rev, psr_typ, psr_vers;
1540
1541         srmmu_modtype = SRMMU_INVAL_MOD;
1542         hwbug_bitmask = 0;
1543
1544         mreg = srmmu_get_mmureg(); psr = get_psr();
1545         mod_typ = (mreg & 0xf0000000) >> 28;
1546         mod_rev = (mreg & 0x0f000000) >> 24;
1547         psr_typ = (psr >> 28) & 0xf;
1548         psr_vers = (psr >> 24) & 0xf;
1549
1550         /* First, check for sparc-leon. */
1551         if (sparc_cpu_model == sparc_leon) {
1552                 init_leon();
1553                 return;
1554         }
1555
1556         /* Second, check for HyperSparc or Cypress. */
1557         if (mod_typ == 1) {
1558                 switch (mod_rev) {
1559                 case 7:
1560                         /* UP or MP Hypersparc */
1561                         init_hypersparc();
1562                         break;
1563                 case 0:
1564                 case 2:
1565                 case 10:
1566                 case 11:
1567                 case 12:
1568                 case 13:
1569                 case 14:
1570                 case 15:
1571                 default:
1572                         prom_printf("Sparc-Linux Cypress support does not longer exit.\n");
1573                         prom_halt();
1574                         break;
1575                 }
1576                 return;
1577         }
1578
1579         /* Now Fujitsu TurboSparc. It might happen that it is
1580          * in Swift emulation mode, so we will check later...
1581          */
1582         if (psr_typ == 0 && psr_vers == 5) {
1583                 init_turbosparc();
1584                 return;
1585         }
1586
1587         /* Next check for Fujitsu Swift. */
1588         if (psr_typ == 0 && psr_vers == 4) {
1589                 phandle cpunode;
1590                 char node_str[128];
1591
1592                 /* Look if it is not a TurboSparc emulating Swift... */
1593                 cpunode = prom_getchild(prom_root_node);
1594                 while ((cpunode = prom_getsibling(cpunode)) != 0) {
1595                         prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
1596                         if (!strcmp(node_str, "cpu")) {
1597                                 if (!prom_getintdefault(cpunode, "psr-implementation", 1) &&
1598                                     prom_getintdefault(cpunode, "psr-version", 1) == 5) {
1599                                         init_turbosparc();
1600                                         return;
1601                                 }
1602                                 break;
1603                         }
1604                 }
1605
1606                 init_swift();
1607                 return;
1608         }
1609
1610         /* Now the Viking family of srmmu. */
1611         if (psr_typ == 4 &&
1612            ((psr_vers == 0) ||
1613             ((psr_vers == 1) && (mod_typ == 0) && (mod_rev == 0)))) {
1614                 init_viking();
1615                 return;
1616         }
1617
1618         /* Finally the Tsunami. */
1619         if (psr_typ == 4 && psr_vers == 1 && (mod_typ || mod_rev)) {
1620                 init_tsunami();
1621                 return;
1622         }
1623
1624         /* Oh well */
1625         srmmu_is_bad();
1626 }
1627
1628 #ifdef CONFIG_SMP
1629 /* Local cross-calls. */
1630 static void smp_flush_page_for_dma(unsigned long page)
1631 {
1632         xc1(local_ops->page_for_dma, page);
1633         local_ops->page_for_dma(page);
1634 }
1635
1636 static void smp_flush_cache_all(void)
1637 {
1638         xc0(local_ops->cache_all);
1639         local_ops->cache_all();
1640 }
1641
1642 static void smp_flush_tlb_all(void)
1643 {
1644         xc0(local_ops->tlb_all);
1645         local_ops->tlb_all();
1646 }
1647
1648 static bool any_other_mm_cpus(struct mm_struct *mm)
1649 {
1650         return cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids;
1651 }
1652
1653 static void smp_flush_cache_mm(struct mm_struct *mm)
1654 {
1655         if (mm->context != NO_CONTEXT) {
1656                 if (any_other_mm_cpus(mm))
1657                         xc1(local_ops->cache_mm, (unsigned long)mm);
1658                 local_ops->cache_mm(mm);
1659         }
1660 }
1661
1662 static void smp_flush_tlb_mm(struct mm_struct *mm)
1663 {
1664         if (mm->context != NO_CONTEXT) {
1665                 if (any_other_mm_cpus(mm)) {
1666                         xc1(local_ops->tlb_mm, (unsigned long)mm);
1667                         if (atomic_read(&mm->mm_users) == 1 && current->active_mm == mm)
1668                                 cpumask_copy(mm_cpumask(mm),
1669                                              cpumask_of(smp_processor_id()));
1670                 }
1671                 local_ops->tlb_mm(mm);
1672         }
1673 }
1674
1675 static void smp_flush_cache_range(struct vm_area_struct *vma,
1676                                   unsigned long start,
1677                                   unsigned long end)
1678 {
1679         struct mm_struct *mm = vma->vm_mm;
1680
1681         if (mm->context != NO_CONTEXT) {
1682                 if (any_other_mm_cpus(mm))
1683                         xc3(local_ops->cache_range, (unsigned long)vma, start,
1684                             end);
1685                 local_ops->cache_range(vma, start, end);
1686         }
1687 }
1688
1689 static void smp_flush_tlb_range(struct vm_area_struct *vma,
1690                                 unsigned long start,
1691                                 unsigned long end)
1692 {
1693         struct mm_struct *mm = vma->vm_mm;
1694
1695         if (mm->context != NO_CONTEXT) {
1696                 if (any_other_mm_cpus(mm))
1697                         xc3(local_ops->tlb_range, (unsigned long)vma, start,
1698                             end);
1699                 local_ops->tlb_range(vma, start, end);
1700         }
1701 }
1702
1703 static void smp_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1704 {
1705         struct mm_struct *mm = vma->vm_mm;
1706
1707         if (mm->context != NO_CONTEXT) {
1708                 if (any_other_mm_cpus(mm))
1709                         xc2(local_ops->cache_page, (unsigned long)vma, page);
1710                 local_ops->cache_page(vma, page);
1711         }
1712 }
1713
1714 static void smp_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1715 {
1716         struct mm_struct *mm = vma->vm_mm;
1717
1718         if (mm->context != NO_CONTEXT) {
1719                 if (any_other_mm_cpus(mm))
1720                         xc2(local_ops->tlb_page, (unsigned long)vma, page);
1721                 local_ops->tlb_page(vma, page);
1722         }
1723 }
1724
1725 static void smp_flush_page_to_ram(unsigned long page)
1726 {
1727         /* Current theory is that those who call this are the one's
1728          * who have just dirtied their cache with the pages contents
1729          * in kernel space, therefore we only run this on local cpu.
1730          *
1731          * XXX This experiment failed, research further... -DaveM
1732          */
1733 #if 1
1734         xc1(local_ops->page_to_ram, page);
1735 #endif
1736         local_ops->page_to_ram(page);
1737 }
1738
1739 static void smp_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1740 {
1741         if (any_other_mm_cpus(mm))
1742                 xc2(local_ops->sig_insns, (unsigned long)mm, insn_addr);
1743         local_ops->sig_insns(mm, insn_addr);
1744 }
1745
1746 static struct sparc32_cachetlb_ops smp_cachetlb_ops __ro_after_init = {
1747         .cache_all      = smp_flush_cache_all,
1748         .cache_mm       = smp_flush_cache_mm,
1749         .cache_page     = smp_flush_cache_page,
1750         .cache_range    = smp_flush_cache_range,
1751         .tlb_all        = smp_flush_tlb_all,
1752         .tlb_mm         = smp_flush_tlb_mm,
1753         .tlb_page       = smp_flush_tlb_page,
1754         .tlb_range      = smp_flush_tlb_range,
1755         .page_to_ram    = smp_flush_page_to_ram,
1756         .sig_insns      = smp_flush_sig_insns,
1757         .page_for_dma   = smp_flush_page_for_dma,
1758 };
1759 #endif
1760
1761 /* Load up routines and constants for sun4m and sun4d mmu */
1762 void __init load_mmu(void)
1763 {
1764         /* Functions */
1765         get_srmmu_type();
1766
1767 #ifdef CONFIG_SMP
1768         /* El switcheroo... */
1769         local_ops = sparc32_cachetlb_ops;
1770
1771         if (sparc_cpu_model == sun4d || sparc_cpu_model == sparc_leon) {
1772                 smp_cachetlb_ops.tlb_all = local_ops->tlb_all;
1773                 smp_cachetlb_ops.tlb_mm = local_ops->tlb_mm;
1774                 smp_cachetlb_ops.tlb_range = local_ops->tlb_range;
1775                 smp_cachetlb_ops.tlb_page = local_ops->tlb_page;
1776         }
1777
1778         if (poke_srmmu == poke_viking) {
1779                 /* Avoid unnecessary cross calls. */
1780                 smp_cachetlb_ops.cache_all = local_ops->cache_all;
1781                 smp_cachetlb_ops.cache_mm = local_ops->cache_mm;
1782                 smp_cachetlb_ops.cache_range = local_ops->cache_range;
1783                 smp_cachetlb_ops.cache_page = local_ops->cache_page;
1784
1785                 smp_cachetlb_ops.page_to_ram = local_ops->page_to_ram;
1786                 smp_cachetlb_ops.sig_insns = local_ops->sig_insns;
1787                 smp_cachetlb_ops.page_for_dma = local_ops->page_for_dma;
1788         }
1789
1790         /* It really is const after this point. */
1791         sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1792                 &smp_cachetlb_ops;
1793 #endif
1794
1795         if (sparc_cpu_model != sun4d)
1796                 ld_mmu_iommu();
1797 #ifdef CONFIG_SMP
1798         if (sparc_cpu_model == sun4d)
1799                 sun4d_init_smp();
1800         else if (sparc_cpu_model == sparc_leon)
1801                 leon_init_smp();
1802         else
1803                 sun4m_init_smp();
1804 #endif
1805 }